Published by

World Scientific Publishing Co. Pte. Ltd.

5 Toh Tuck Link, Singapore 596224

USA office: 27 Warren Street, Suite 401-402, Hackensack, NJ 07601

UK office: 57 Shelton Street, Covent Garden, London WC2H 9HE

Library of Congress Cataloging-in-Publication Data

Names: Coleman, Sidney, 1937–2007. | Chen, Bryan Gin-ge, 1985– editor.

Title: Lectures of Sidney Coleman on quantum field theory / [edited by]

Bryan Gin-ge Chen (Leiden University, Netherlands) [and five others].

Description: New Jersey : World Scientific, 2018. | Includes bibliographical

references and index.

Identifiers: LCCN 2018041457| ISBN 9789814632539 (hardcover : alk. paper) |

ISBN 9789814635509 (pbk. : alk. paper)

Subjects: LCSH: Quantum field theory.

Classification: LCC QC174.46 .C65 2018 | DDC 530.14/3--dc23

LC record available at https://lccn.loc.gov/2018041457

British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library.

Copyright © 2019 by World Scientific Publishing Co. Pte. Ltd.

All rights reserved. This book, or parts thereof, may not be reproduced in any form or by any means, electronic or
mechanical, including photocopying, recording or any information storage and retrieval system now known or to
be invented, without written permission from the publisher.

For photocopying of material in this volume, please pay a copying fee through the Copyright Clearance Center,
Inc., 222 Rosewood Drive, Danvers, MA 01923, USA. In this case permission to photocopy is not required from
the publisher.

For any available supplementary material, please visit

https://www.worldscientific.com/worldscibooks/10.1142/9371#t=suppl

Printed in Singapore

to Diana Coleman

and for all of Sidney’s students—past, present, and future

Contents
Foreword

Preface

Frequently cited references

Index of useful formulae

A note on the problems

1Adding special relativity to quantum mechanics

1.1Introductory remarks

1.2Theory of a single free, spinless particle of mass µ

1.3Determination of the position operator X

2The simplest many-particle theory

2.1First steps in describing a many-particle state

2.2Occupation number representation

2.3Operator formalism and the harmonic oscillator

2.4The operator formalism applied to Fock space

3Constructing a scalar quantum field

3.1Ensuring relativistic causality

3.2Conditions to be satisfied by a scalar quantum field

3.3The explicit form of the scalar quantum field

3.4Turning the argument around: the free scalar field as the fundamental object

3.5A hint of things to come

Problems 1

Solutions 1

4The method of the missing box

4.1Classical particle mechanics

4.2Quantum particle mechanics

4.3Classical field theory

4.4Quantum field theory

4.5Normal ordering

5Symmetries and conservation laws I. Spacetime symmetries

5.1Symmetries and conservation laws in classical particle mechanics

5.2Extension to quantum particle mechanics

5.3Extension to field theory

5.4Conserved currents are not uniquely defined

5.5Calculation of currents from spacetime translations

5.6Lorentz transformations, angular momentum and something else

Problems 2

Solutions 2

6Symmetries and conservation laws II. Internal symmetries

6.1Continuous symmetries

6.2Lorentz transformation properties of the charges

6.3Discrete symmetries

7Introduction to perturbation theory and scattering

7.1The Schrödinger and Heisenberg pictures

7.2The interaction picture

7.3Dyson’s formula

7.4Scattering and the S-matrix

Problems 3

Solutions 3

8Perturbation theory I. Wick diagrams

8.1Three model field theories

8.2Wick’s theorem

8.3Dyson’s formula expressed in Wick diagrams

8.4Connected and disconnected Wick diagrams

8.5The exact solution of Model 1

Problems 4

Solutions 4

9Perturbation theory II. Divergences and counterterms

9.1The need for a counterterm in Model 2

9.2Evaluating the S matrix in Model 2

9.3Computing the Model 2 ground state energy

9.4The ground state wave function in Model 2

9.5An infrared divergence

Problems 5

Solutions 5

10Mass renormalization and Feynman diagrams

10.1Mass renormalization in Model 3

10.2Feynman rules in Model 3

10.3Feynman diagrams in Model 3 to order g2

10.4O(g2) nucleon–nucleon scattering in Model 3

11Scattering I. Mandelstam variables, CPT and phase space

11.1Nucleon–antinucleon scattering

11.2Nucleon–meson scattering and meson pair creation

11.3Crossing symmetry and CPT invariance

11.4Phase space and the S matrix

12Scattering II. Applications

12.1Decay processes

12.2Differential cross-section for a two-particle initial state

12.3The density of final states for two particles

12.4The Optical Theorem

12.5The density of final states for three particles

12.6A question and a preview

Problems 6

Solutions 6

13Green’s functions and Heisenberg fields

13.1The graphical definition of

13.2The generating functional Z[ρ] for G(n)(xi)

13.3Scattering without an adiabatic function

13.4Green’s functions in the Heisenberg picture

13.5Constructing in and out states

Problems 7

Solutions 7

14The LSZ formalism

14.1Two-particle states

14.2The proof of the LSZ formula

14.3Model 3 revisited

14.4Guessing the Feynman rules for a derivative interaction

Problems 8

Solutions 8

15Renormalization I. Determination of counterterms

15.1The perturbative determination of A

15.2The Källén-Lehmann spectral representation

15.3The renormalized meson propagator

15.4The meson self-energy to O(g2)

15.5A table of integrals for one loop

16Renormalization II. Generalization and extension

16.1The meson self-energy to O(g2), completed

16.2Feynman parametrization for multiloop graphs

16.3Coupling constant renormalization

16.4Are all quantum field theories renormalizable?

Problems 9

Solutions 9

17Unstable particles

17.1Calculating the propagator for µ > 2m

17.2The Breit–Wigner formula

17.3A first look at the exponential decay law

17.4Obtaining the decay law by stationary phase approximation

18Representations of the Lorentz Group

18.1Defining the problem: Lorentz transformations in general

18.2Irreducible representations of the rotation group

18.3Irreducible representations of the Lorentz group

18.4Properties of the SO(3) representations D (s)

18.5Properties of the SO(3,1) representations D (s +, s −)

Problems 10

Solutions 10

19The Dirac Equation I. Constructing a Lagrangian

19.1Building vectors out of spinors

19.2A Lagrangian for Weyl spinors

19.3The Weyl equation

19.4The Dirac equation

20The Dirac Equation II. Solutions

20.1The Dirac basis

20.2Plane wave solutions

20.3Pauli’s theorem

20.4The γ matrices

20.5Bilinear spinor products

20.6Orthogonality and completeness

Problems 11

Solutions 11

21The Dirac Equation III. Quantization and Feynman Rules

21.1Canonical quantization of the Dirac field

21.2Wick’s theorem for Fermi fields

21.3Calculating the Dirac propagator

21.4An example: Nucleon–meson scattering

21.5The Feynman rules for theories involving fermions

21.6Summing and averaging over spin states

Problems 12

Solutions 12

22CPT and Fermi fields

22.1Parity and Fermi fields

22.2The Majorana representation

22.3Charge conjugation and Fermi fields

22.4PT invariance and Fermi fields

22.5The CPT theorem and Fermi fields

23Renormalization of spin-½ theories

23.1Lessons from Model 3

23.2The renormalized Dirac propagator

23.3The spectral representation of

23.4The nucleon self-energy

23.5The renormalized coupling constant

Problems 13

Solutions 13

24Isospin

24.1Field theoretic constraints on coupling constants

24.2The nucleon and pion as isospin multiplets

24.3Experimental consequences of isospin conservation

24.4Hypercharge and G-parity

25Coping with infinities: regularization and renormalization

25.1Regularization

25.2The BPHZ algorithm

25.3Applying the algorithm

25.4Survey of renormalizable theories for spin 0 and spin ½

Problems 14

Solutions 14

26Vector fields

26.1The free real vector field

26.2The Proca equation and its solutions

26.3Canonical quantization of the Proca field

26.4The limit µ → 0: a simple physical consequence

26.5Feynman rules for a real massive vector field

27Electromagnetic interactions and minimal coupling

27.1Gauge invariance and conserved currents

27.2The minimal coupling prescription

27.3Technical problems

Problems 15

Solutions 15

28Functional integration and Feynman rules

28.1First steps with functional integrals

28.2Functional integrals in field theory

28.3The Euclidean Z0[J] for a free theory

28.4The Euclidean Z[J] for an interacting field theory

28.5Feynman rules from functional integrals

28.6The functional integral for massive vector mesons

29Extending the methods of functional integrals

29.1Functional integration for Fermi fields

29.2Derivative interactions via functional integrals

29.3Ghost fields

29.4The Hamiltonian form of the generating functional

29.5How to eliminate constrained variables

29.6Functional integrals for QED with massive photons

Problems 16

Solutions 16

30Electrodynamics with a massive photon

30.1Obtaining the Feynman rules for scalar electrodynamics

30.2The Feynman rules for massive photon electrodynamics

30.3Some low order computations in spinor electrodynamics

30.4Quantizing massless electrodynamics with functional integrals

31The Faddeev–Popov prescription

31.1The prescription in a finite number of dimensions

31.2Extending the prescription to a gauge field theory

31.3Applying the prescription to QED

31.4Equivalence of the Faddeev–Popov prescription and canonical quantization

31.5Revisiting the massive vector theory

31.6A first look at renormalization in QED

Problems 17

Solutions 17

32Generating functionals and Green’s functions

32.1The loop expansion

32.2The generating functional for 1PI Green’s functions

32.3Connecting statistical mechanics with quantum field theory

32.4Quantum electrodynamics in a covariant gauge

33The renormalization of QED

33.1Counterterms and gauge invariance

33.2Counterterms in QED with a massive photon

33.3Gauge-invariant cutoffs

33.4The Ward identity and Green’s functions

33.5The Ward identity and counterterms

Problems 18

Solutions 18

34Two famous results in QED

34.1Coulomb’s Law

34.2The electron’s anomalous magnetic moment in quantum mechanics

34.3The electron’s anomalous magnetic moment in QED

35Confronting experiment with QED

35.1Higher order contributions to the electron’s magnetic moment

35.2The anomalous magnetic moment of the muon

35.3A low-energy theorem

35.4Photon-induced corrections to strong interaction processes (via symmetries)

Problems 19

Solutions 19

36Introducing SU(3)

36.1Decays of the η

36.2An informal historical introduction to SU(3)

36.3Tensor methods for SU(n)

36.4Applying tensor methods in SU(2)

36.5Tensor representations of SU(3)

37Irreducible multiplets in SU(3)

37.1The irreducible representations q and q

37.2Matrix tricks with SU(3)

37.3Isospin and hypercharge decomposition

37.4Direct products in SU(3)

37.5Symmetry and antisymmetry in the Clebsch–Gordan coefficients

Problems 20

Solutions 20

38SU(3): Proofs and applications

38.1Irreducibility, inequivalence, and completeness of the IR’s

38.2The operators I, Y and Q in SU(3)

38.3Electromagnetic form factors of the baryon octet

38.4Electromagnetic mass splittings of the baryon octet

39Broken SU(3) and the naive quark model

39.1The Gell-Mann–Okubo mass formula derived

39.2The GMO formula applied

39.3The GMO formula challenged

39.4The naive quark model (and how it grew)

39.5What can you build out of three quarks?

39.6A sketch of quantum chromodynamics

Problems 21

Solutions 21

40Weak interactions and their currents

40.1The weak interactions circa 1965

40.2The conserved vector current hypothesis

40.3The Cabibbo angle

40.4The Goldberger–Treiman relation

41Current algebra and PCAC

41.1The PCAC hypothesis and its interpretation

41.2Two isotriplet currents

41.3The gradient-coupling model

41.4Adler’s Rule for the emission of a soft pion

41.5Equal-time current commutators

Problems 22

Solutions 22
42Current algebra and pion scattering

42.1Pion–hadron scattering without current algebra

42.2Pion–hadron scattering and current algebra

42.3Pion–pion scattering

42.4Some operators and their eigenvalues

43A first look at spontaneous symmetry breaking

43.1The man in a ferromagnet

43.2Spontaneous symmetry breaking in field theory: Examples

43.3Spontaneous symmetry breaking in field theory: The general case

43.4Goldstone’s Theorem

Problems 23

Solutions 23

44Perturbative spontaneous symmetry breaking

44.1One vacuum or many?

44.2Perturbative spontaneous symmetry breaking in the general case

44.3Calculating the effective potential

44.4The physical meaning of the effective potential

45Topics in spontaneous symmetry breaking

45.1Three heuristic aspects of the effective potential

45.2Fermions and the effective potential

45.3Spontaneous symmetry breaking and soft pions: the sigma model

45.4The physics of the sigma model

Problems 24

Solutions 24

46The Higgs mechanism and non-Abelian gauge fields

46.1The Abelian Higgs model

46.2Non-Abelian gauge field theories

46.3Yang–Mills fields and spontaneous symmetry breaking

47Quantizing non-Abelian gauge fields

47.1Quantization of gauge fields by the Faddeev–Popov method

47.2Feynman rules for a non-Abelian gauge theory

47.3Renormalization of pure gauge field theories

47.4The effective potential for a gauge theory

Problems 25

Solutions 25
 48The Glashow–Salam–Weinberg Model I. A theory of leptons

48.1Putting the pieces together

48.2The electron-neutrino weak interactions

48.3Electromagnetic interactions of the electron and neutrino

48.4Adding the other leptons

48.5Summary and outlook

49The Glashow–Salam–Weinberg Model II. Adding quarks

49.1A simplified quark model

49.2Charm and the GIM mechanism

49.3Lower bounds on scalar boson masses

50The Renormalization Group

50.1The renormalization group for ϕ 4 theory

50.2The renormalization group equation

50.3The solution to the renormalization group equation

50.4Applications of the renormalization group equation

Concordance of videos and chapters

Index

Foreword

Generations of theoretical physicists learned quantum field theory from Sidney Coleman. Hundreds attended his
famous lecture course at Harvard University — the lecture hall was usually packed with listeners well beyond
those registered for Physics 253 — while many more encountered photocopies of handwritten notes from the
course or saw videos of his lectures long after they had been recorded. Coleman’s special gift for exposition, and
his evident delight for the material, simply could not be matched. A Coleman lecture on quantum field theory
wasn’t merely part of a course; it was an adventure.

Sidney Coleman was born in 1937 and grew up in Chicago. He showed keen interest in science at an early
age, and won the Chicago Science Fair while in high school for his design of a rudimentary computer. He studied
physics as an undergraduate at the Illinois Institute of Technology, graduating in 1957, and then pursued his
doctorate in physics at the California Institute of Technology. At Caltech Coleman befriended Sheldon Glashow
(then a postdoc), took courses from Richard Feynman, and wrote his dissertation under the supervision of Murray
Gell-Mann. In 1961, as he was completing his dissertation, Coleman moved to Harvard as the Corning Lecturer
and Fellow. He joined the Physics Department faculty at Harvard soon after that, and remained a member of the
faculty until his retirement in 2006.1

For more than thirty years, Coleman led Harvard’s group in theoretical high-energy physics. Colleagues and
students alike came to consider him “the Oracle.”2 At one point Coleman’s colleague, Nobel laureate Steven
Weinberg, was giving a seminar in the department. Coleman had missed the talk and arrived during the question-
and-answer session. Just as Coleman entered the room, Weinberg replied to someone else, “I’m sorry, but I don’t
know the answer to that question.” “I do,” Coleman called out from the back. “What was the question?” Coleman
then listened to the question and answered without hesitation.3

Coleman had an off-scale personality, inspiring stories that colleagues and former students frequently still
share. He kept unusual hours, working late into the night; at one point he complained to a colleague that he had
been “dragged out of bed four hours before my usual time of rising (i.e., at 8 o’clock in the morning) to receive your
telegram.”4 Indeed, he had refused to teach a course at 9 a.m., explaining, “I can’t stay up that late.”5 Coleman’s
penchant for chain-smoking — even while lecturing — made at least one journalist marvel that Coleman never
mistook his chalk for his cigarette.6 In 1978, Harvard Magazine published a profile of Coleman, to which he took
exception. As he wrote to the editors:

Gentlemen:

In your September-October issue, I am described as “a wild-looking guy, with scraggly black hair
down to his shoulders and the worst slouch I’ve ever seen. He wears a purple polyester sports jacket.”

This allegation is both false to fact and damaging to my reputation; I must insist upon a retraction.

The jacket in question is wool. All my purple jackets are wool.7

Little wonder that Coleman was often described as a superposition of Albert Einstein and comedian Woody Allen.8

Coleman had an extraordinary talent for wordplay as well as physics, and a lively, spontaneous wit. Once,
while a journalist was preparing a feature article about him, Coleman received a telephone call in his office; the
caller had misdialed. “No, I’m not Frank,” the journalist captured Coleman replying, “but I’m not entirely
disingenuous either.”9 Coleman frequently sprinkled literary and historical allusions throughout his writings,
published articles and ephemeral correspondence alike. Writing to a colleague after a recent visit, for example,
Coleman noted that the reimbursement he had received did not cover some of his travel expenses: “Samuel
Gompers was once asked, ‘What does Labor want?’ He replied, ‘More.’”10 The lecture notes in this volume
likewise include passing nods to Pliny the Elder, the plays of Moli`ere, Sherlock Holmes stories, and more. He
took the craft of writing quite seriously, at one point advising a friend, “Literary investigation by bombarding a
manuscript with prepositions is as obsolete as electrical generation by beating a cat with an amber rod.”11

Early in Coleman’s career, colleagues began to admire his unusual skill in the lecture hall. Just a few years
after joining Harvard’s faculty, he was in such high demand for the summer-school lecture circuit that he had to
turn down more requests than he could accept.12 He became a regular lecturer at the annual Ettore Majorana
summer school in Erice, Italy, and developed a warm friendship with its organizer, Antonino Zichichi. Usually
Coleman volunteered topics on which he planned to lecture at Erice — not infrequently using the summer course
as an opportunity to teach himself material he felt he had not quite mastered yet — though sometimes Zichichi
assigned topics to Coleman. Preparing for the 1969 summer school, for example, Zichichi pressed him, “Please
stop refusing to lecture on the topic I have assigned to you. It is not my fault if you are among the few physicists
who can lecture [on] anything.”13

Coleman worked hard to keep his lectures fresh for his listeners, putting in significant effort ahead of time on
organization and balance. He described his method to a colleague in 1975: “The notes I produce while preparing a
lecture are skeletal in the extreme, nothing but equations without words.” That way he could be sure to hit his main
points while keeping most of his exposition fairly spontaneous.14 The light touch on his first pass-through meant
that Coleman needed to expend significant effort after the lectures were given, converting his sparse notes into
polished prose that could be published in the summer-school lecture-note volumes. He often confided to
colleagues that his “slothful” ways kept him from submitting manuscripts of his lecture notes on time.15 Likely for
that reason he shunned repeated invitations from publishers to write a textbook. “Not even the Symbionese
Liberation Army would be able to convert me to writing an elementary physics text,” he replied to one eager
editor.16

Luckily for him — indeed, luckily for us — Coleman’s lecture course on quantum field theory at Harvard was
videotaped during the 1975–76 academic year, with no need for him to write up his notes. Filming a lecture course
back then was quite novel, so much so that Coleman felt the need to explain why the large camera and associated
equipment were perched in the back of the lecture hall. “The apparatus you see around here is part of a CIA
surveillance project,” he joked at the start of his first lecture, drawing immediate laughter from the students. He
continued: “I fall within their domain because I read JETP Letters,” inciting further laughter.17 Hardly a spy caper,
the videotapes were actually part of an experiment in educational technology.18

News of the tapes spread, and soon they were in high demand well beyond Cambridge. Colleagues wrote to
Coleman, asking if they could acquire copies of the videotapes for their own use, from as far away as Edinburgh
and Haifa.19 Coleman’s administrative assistant explained to one interested colleague in 1983 that the tapes had
begun to “deteriorate badly” from overuse, yet they remained “in great demand even if those in use are in poor
condition.”20 Years later, in 2007, Harvard’s Physics Department arranged for the surviving videotapes to be
digitized, and they are now available, for free, on the Department’s website.21 As David Derbes explains in his
Preface, the editorial team made extensive use of the videos while preparing this volume.

Physicists knitted together what we now recognize as (nonrelativistic) quantum mechanics in a flurry of
papers during the mid-1920s. The pace was extraordinary. Within less than a year — between July 1925 and June
1926 — Werner Heisenberg submitted his first paper on what would become known as “matrix mechanics,” Erwin
Schrödinger independently developed “wave mechanics,” several physicists began to elucidate their mathematical
equivalence, and Max Born postulated that Schrödinger’s new wavefunction, ψ, could be interpreted as a
probability amplitude. A few months after that, in March 1927, Heisenberg submitted his now-famous paper on the
uncertainty principle.22

In hindsight, many physicists have tended to consider that brief burst of effort as a capstone, the end of a
longer story that stretched from Max Planck’s first intimations about blackbody radiation, through Albert Einstein’s
hypothesis about light quanta, to Niels Bohr’s model of the atom and Louis de Broglie’s suggestive insights about
matter waves. To leading physicists at the time, however, the drumbeat of activity during the mid-1920s seemed
to herald the start of a new endeavor, not the culmination of an old one. Already in 1926 and 1927, Werner
Heisenberg, Pascual Jordan, Wolfgang Pauli, Paul Dirac and others were hard at work trying to quantize the
electromagnetic field, and to reconcile quantum theory with special relativity. They had begun to craft quantum
field theory.23

Those early efforts quickly foundered, as a series of divergences bedeviled physicists’ calculations. By the
early 1930s, theorists had identified several types of divergences — infinite self-energies, infinite vacuum
polarization — which seemed to arise whenever they tried to incorporate the effects of “virtual particles” in a
systematic way. Some leaders, like Heisenberg, called for yet another grand, conceptual revolution, as sweeping
as the disruptions of 1925–27 had been, which would replace quantum field theory with some new, as-yet
unknown framework. No clear candidate emerged, and before long physicists around the world found their
attention absorbed by the rise of fascism and the outbreak of World War II.24

Soon after the war, a younger generation of physicists returned to the challenge of quantum field theory and
its divergences. Many had spent the war years working on various applied projects, such as radar and the
Manhattan Project, and had developed skills in wringing numerical predictions from seemingly intractable
equations — what physicist and historian of science Silvan (Sam) Schweber dubbed, “getting the numbers out.”
Some had gained crash-course experience in engineers’ effective-circuit approaches while working on radar;
others had tinkered with techniques akin to Green’s functions to estimate rates for processes like neutron diffusion
within a volume of fissile material.25 After the war, these younger physicists were further intrigued and inspired by
new experimental results, likewise made possible by the wartime projects. In the late 1940s, experimental
physicists like Willis Lamb and Isidor Rabi — using surplus equipment from the radar project and exploiting
newfound skills in manipulating microwave-frequency electronics — measured tiny but unmistakeable effects,
including a miniscule difference between the energy levels of an electron in the 2s versus 2p states of a hydrogen
atom, and an “anomalous” magnetic moment of the electron, ever-so-slightly larger than the value predicted by
Dirac’s equation.26

Prodded by what seemed like tantalizing evidence of the effects of virtual particles, young theorists like Julian
Schwinger and Richard Feynman worked out various “renormalization” techniques in 1947 and 1948, with which
to tame the infinities within quantum electrodynamics (QED). They soon learned that Schwinger’s approach was
remarkably similar to ideas that Sin-itiro Tomonaga and colleagues had developed independently in Tokyo, during
the war. Early in 1949, meanwhile, Freeman Dyson demonstrated a fundamental, underlying equivalence between
the Tomonaga–Schwinger approach and Feynman’s distinct-looking efforts, and further showed that
renormalization should work at arbitrary perturbative order in QED — a remarkable synthesis at least as potent,
and as surprising, as the earlier demonstrations had been, two decades earlier, that Heisenberg’s and
Schrödinger’s approaches to quantum theory were mathematically equivalent.27

Dyson became adept at teaching the new approach to quantum field theory. Hectographed copies of his
lecture notes from a 1951 course at Cornell University quickly began to circulate.28 The unpublished notes
provided a template for the first generation of textbooks on quantum field theory, written after the great
breakthroughs in renormalization: books like Josef Jauch and Fritz Rohrlich’s The Theory of Photons and
Electrons (1955) and Silvan Schweber’s massive Introduction to Relativistic Quantum Field Theory (1961),
culminating in the pair of textbooks by James Bjorken and Sidney Drell, Relativistic Quantum Mechanics (1964)
and Relativistic Quantum Fields (1965).29

Yet the field did not stand still; new puzzles soon demanded attention. In 1957, for example, experimentalist
Chien-Shiung Wu and her colleagues demonstrated that parity symmetry was violated in weak-force interactions,
such as the β-decay of cobalt-60 nuclei: nature really did seem to distinguish between right-handed and left-
handed orientations in space. Theorists T. D. Lee and C. N. Yang had hypothesized that the weak nuclear force
might violate parity, and, soon after Wu’s experiment, Murray Gell-Mann, Richard Feynman, and others published
models of such parity-violating interactions within a field-theory framework.30 Yet their models suffered from poor
behavior at high energies, which led others, including Sidney Bludman, Julian Schwinger, and Sheldon Glashow
to return to suggestive hints from Yang and Robert Mills: perhaps nuclear forces were mediated by sets of force-
carrying particles — and perhaps those particles obeyed a nontrivial gauge symmetry, with more complicated
structure than the simple U(1) gauge symmetry that seemed to govern electrodynamics.31

Mathematical physicists like Hermann Weyl had first explored gauge theories early in the 20th century, when
thinking about the structure of spacetime in the context of Einstein’s general theory of relativity. Decades later, in
the mid-1950s, Yang and Mills, Robert Shaw, Ryoyu Utiyama, and Schwinger suggested that nontrivial gauge
symmetries could help physicists parse the nuclear forces.32 Yet applying such ideas to nuclear forces remained
far from straightforward. For one thing, nuclear forces clearly had a finite range, which seemed to imply that the
corresponding force-carrying particles should have a large mass. But inserting such mass terms by hand within
the field-theoretic models violated the very gauge symmetries that those particles were meant to protect. These
challenges drove several theorists to investigate spontaneous symmetry breaking in gauge field theories during
the late 1950s through the mid-1960s, culminating in what has come to be known as the “Higgs mechanism.”33

Another major challenge concerned how to treat strongly coupled particles, including the flood of nuclear
particles — cousins of the familiar protons and neutrons — that physicists began to discover with their hulking
particle accelerators. Dyson observed in 1953 that hardly a month went by without the announcement that
physicists had discovered a new particle.34 Whereas electrons and photons interacted with a relatively small
coupling constant, e2 ~ 1/137 (in appropriate units), many of the new particles seemed to interact strongly with
each other, with coupling constants g2 ≫ 1. The small size of e2 had been critical to the perturbative approaches
of Tomonaga, Schwinger, Feynman, and Dyson; how could anyone perform a systematic calculation among
strongly coupled particles? For just this reason, Feynman himself cautioned Enrico Fermi in December, 1951,
“Don’t believe any calculation in meson theory which uses a Feynman diagram!”35

Some theorists, like Murray Gell-Mann and Yuval Ne’eman, sought to make headway by deploying symmetry
arguments and tools from group theory. Gell-Mann introduced his famous “Eightfold Way” in 1961, for example, to
try to understand certain regularities among nuclear particles by arranging them in various arrays, sorted by
quantum numbers like isospin and hypercharge.36 Others, such as Geoffrey Chew, embarked on an even more
ambitious program to replace quantum field theory altogether. Chew announced at a conference in June 1961 that
quantum field theory was “sterile with respect to the strong interactions” and therefore “destined not to die but just
to fade away.” He and his colleagues focused on an “autonomous S-matrix program,” eschewing all talk of
Lagrangians, virtual particles, and much of the apparatus that Dyson had so patiently assembled for making
calculations in QED.37

Amid the turmoil and uncertainty, quantum field theory was never quite as dead as theorists like Chew liked to
proclaim. Nonetheless, its status among high-energy physicists seemed far less settled in 1970 than it had been
in 1950. The tide turned back toward field theory’s advocates during the mid-1970s, driven by several important
developments. First was the construction of a unified model of the electromagnetic and weak interactions,
accomplished independently by Sheldon Glashow, Steven Weinberg, and Abdus Salam. Though they had
published their work in the mid-1960s, it only attracted sustained attention from the community after Gerard ’t
Hooft and Martinus Veltman demonstrated in 1971–72 that such gauge field theories could be renormalized in a
systematic way. (As Coleman observed, ’t Hooft’s work revealed “Weinberg and Salam’s frog to be an enchanted
prince.”) Soon after that, in 1973–74, teams of experimentalists at CERN and Fermilab independently found
evidence of weak neutral currents, as predicted by the Glashow–Weinberg–Salam theory.38

Meanwhile, other theorists, led by Yoichiro Nambu, Murray Gell-Mann, and Harald Fritzsch, developed
quantum chromodynamics: a well-defined scheme for treating strong interactions among quarks and gluons,
developed in analogy to QED but incorporating the kind of nontrivial gauge structure at the heart of the
Glashow–Weinberg–Salam electroweak theory. The demonstration in 1973 by Coleman’s student David Politzer
and independently by David Gross and Frank Wilczek that the effective coupling strength between quarks and
gluons in this model should decrease at short distances (or, correspondingly, high energies) — which came to be
known as “asymptotic freedom” — breathed new life into field-theoretic approaches to the strong interactions.39
Before long, the distinct threads of electroweak unification and quantum chromodynamics were knitted into a
single “Standard Model” of particle physics, a model built squarely within the framework of quantum field theory.40

Even as physicists’ pursuit of field-theoretic techniques outstripped the template of perturbative QED during
the 1960s, Dyson’s crisp pedagogical model, which had been honed in the era of QED’s great successes,
continued to dominate in the classroom. Nobel laureate David Gross, for example, recalled his first course on
quantum field theory at Berkeley in 1965, in which he and his fellow students were taught “that field theory equals
Feynman rules”: quantum field theory was still taught as if all that mattered were clever techniques for performing
perturbative calculations.41

Coleman plotted a different course when he began teaching quantum field theory at Harvard a few years
later. Early in the first semester, his students would practice drawing Feynman diagrams for perturbative
calculations, to be sure. But in Coleman’s classroom, quantum field theory would no longer be taught as a mere
grab-bag of perturbative techniques. Coleman’s course, in turn, helped to reinvigorate the study of quantum field
theory more generally, following a protracted period when its fate seemed far from clear.

One obvious distinction between Coleman’s pedagogical approach and Dyson’s was an emphasis upon
group theory and gauge symmetries. Coleman, after all, had written his dissertation at Caltech on “The Structure
of Strong Interaction Symmetries,” working closely with Gell-Mann just at the time that Gell-Mann introduced his
“Eightfold Way.” Coleman incorporated group-theoretic techniques into his teaching rather early, devoting his first
set of summer-school lectures at Erice to the topic in 1966 (drawing extensively from his dissertation); Howard
Georgi likewise recalls learning group theory from a course that Coleman taught at Harvard around the same time.
Coleman continued to refine his presentation over the years. By the mid-1970s, he devoted several weeks of his
course on quantum field theory to non-Abelian groups like SU(3) and their role in gauge field theories.42

Second was an emphasis on path-integral techniques. Although Feynman had developed path integrals in his
Ph. D. dissertation and published on them in the 1940s, they had garnered virtually no space in the textbooks on
quantum field theory published during the 1950s and 1960s. Nonetheless, several theorists began to recognize
the power and elegance of path-integral techniques over the course of the 1960s, especially for tackling models
with nontrivial gauge structure.43 When Coleman began teaching his course on quantum field theory, he featured
functional integration and path-integral methods prominently.

Third was an emphasis on spontaneous symmetry breaking. Coleman liked to joke with his students about his
curious inability to predict what would become the most important developments in the field. Indeed, handwritten
lecture notes from 1990 record him explaining:

At crucial moments in the history of physics, I have often said about a new idea that I think it must be
wrong. When the quark model was proposed, I thought it was wrong; likewise the Higgs mechanism.
That’s a good sign. If I say something isn’t worth paying attention to, it probably isn’t worth paying
attention to. If I say it’s wrong, then the idea merits careful examination — it may be important.44
Peter Higgs himself recalled that when he visited Harvard in the spring of 1966 to present his work, Coleman was
ready to pounce, so certain was he that Higgs’s work must be mistaken.45

For all the joking, however, Coleman rapidly became a leading expert on spontaneous symmetry breaking
and one of its best-known expositors. He lectured on the subject at Erice and incorporated extensive material on
symmetry breaking in his Harvard course on quantum field theory. Not only that: together with his graduate student
Erick Weinberg, Coleman extended the idea to symmetries of an effective potential that could be broken by
radiative corrections (known today as “Coleman–Weinberg” symmetry breaking), and later, with Curt Callan and
Frank De Luccia, Coleman explored “the fate of the false vacuum,” laying crucial groundwork for our modern
understanding of early-universe cosmology.46

Coleman’s lecture course on quantum field theory thus moved well beyond the earlier pedagogical tradition
modeled on Dyson’s notes and typified by Bjorken and Drell’s Relativistic Quantum Fields. The differences lay not
just in topics covered, but in underlying spirit. Coleman presented quantum field theory as a capacious framework,
with significant nonperturbative structure. His style was neither overly rigorous nor narrowly phenomenological,
offering an introduction to the Standard Model with an emphasis on general principles. In that way, his course
remained more accessible and less axiomatic than many of the books that began to appear in the 1980s, such as
Claude Itzykson and Jean-Bernard Zuber’s compendious Quantum Field Theory (1980) or Ta-Pei Cheng and
Ling-Fong Li’s more specialized Gauge Theory of Elementary Particle Physics (1984).47

In at least one significant way, however, Coleman’s pedagogical approach remained closer in spirit to
Dyson’s lectures than more recent developments. Renormalizability retained a special place for Coleman: models
were adjudicated at least in part on whether divergences could be systematically removed for processes involving
arbitrarily high energies. This, after all, was how (in Coleman’s telling) ’t Hooft’s results had ennobled the
Glashow–Weinberg–Salam model. Though Coleman was deeply impressed by Kenneth Wilson’s work on the
renormalization group — late in 1985, he remarked upon “Ken Wilson’s double triumph” of uniting the study of field
theory and critical phenomena — Coleman never fully adopted the viewpoint of effective field theory.48 In effective
field theories, physicists allow for an infinite tower of nonrenormalizable interaction terms — all terms consistent
with some underlying symmetries — and calculate resulting processes for energy scales below some threshold,
Λ. Though effective field theory techniques have become central to research in many areas of high-energy physics
over the past three decades, today’s popular textbooks on quantum field theory still rarely devote much space to
the topic — so Coleman’s lecture course continues to enjoy excellent company.49

I took the two-semester course Physics 253 with Sidney Coleman during the 1993–94 academic year, my first
year in graduate school. For each semester, a large percentage of students’ grades in the class derived from how
well they did on a final exam: a 72-hour take-home exam that Coleman distributed on a Friday afternoon. My
recollections of the weekend I spent working on the exam that fall semester are a bit hazy — much sweating, a bit
of cursing, and very little sleep — and in the end, I ran out of time before I could complete the last problem. Sleep
deprived, desperate, and more than a little inspired by Coleman’s own sense of humor, I decided to appeal to a
variant of the CPT theorem, on which we had focused so dutifully in class. In haste I scribbled down, “The final
result follows from CBT: the Coleman Benevolence Theorem.” A few days later I got back the marked exam. In
Coleman’s inimitable, blocky handwriting he had scrawled, “Be careful: This has not been experimentally
verified.”50

Pace Coleman’s warning, there is plenty of evidence of his benevolence. I was struck recently, for example,
by a questionnaire that he filled out in 1983, in preparation for his 30th high school reunion. By that time he had
been elevated to an endowed professorship at Harvard and elected a member of the U.S. National Academy of
Sciences. Yet in the space provided on the form for “occupation or profession,” Coleman wrote, simply, “teacher.”

And quite a teacher he was. Throughout his career, he supervised 40 Ph.D. students. He routinely shared his
home telephone number with students (undergraduates and graduate students alike), encouraging them to call
him at all hours. “Don’t worry about disturbing me if you call me at home,” he advised a group of undergraduates in
the early 1990s. “I’m home most nights and usually stay up until 4 a.m. or so.” That sort of dedication left an
impression. “I would like to thank you for providing me with the best academic course I have ever encountered
throughout my college career,” wrote one undergraduate upon completing a course on quantum mechanics with
Coleman. “I commend you on your excellent preparation for each and every lecture, your availability and
helpfulness to each student, and particularly, your concern that the student develop his understanding and interest
in the subject.” Another undergraduate, who had taken Physics 253, wrote to Coleman a few years later that he
was “one of the very best teachers I have had in my life.”51

That spirit infuses this volume. Producing these lecture notes has been an enormous labor of love, initiated
by David Derbes and brought to fruition thanks to the tireless efforts of a large editorial team, with special
contributions from Bryan Gin-ge Chen, David Derbes, David Griffiths, Brian Hill, Richard Sohn, and Yuan-Sen
Ting. This volume is proof that Sidney Coleman inspired benevolence enough to go around.52

David Kaiser

Germeshausen Professor of the History of Science

and Professor of Physics
Massachusetts Institute of Technology

1HowardGeorgi, “Sidney Coleman, March 7, 1937 – November 18, 2007,” Biographical Memoirs of the National
Academy of Sciences (2011), available at http://www.nasonline.org/publications/biographical-memoirs/memoir-pdfs/coleman-
sidney.pdf.
2Quoted in Roberta Gordon, “Sidney Coleman dies at 70,” Harvard Gazette (29 November 2007).
3David H. Freedman, “Maker of worlds,” Discover (July 1990): 46–52, on p. 48.
4Sidney Coleman to Antonino Zichichi, 5 November 1970. Coleman’s correspondence is in the possession of his
widow, Diana Coleman. Many of the letters quoted here will appear in the collection, Theoretical Physics in Your
Face: Selected Correspondence of Sidney Coleman, ed. Aaron Wright, Diana Coleman, and David Kaiser
(Singapore: World Scientific, forthcoming).
5Quoted in Gordon, “Sidney Coleman dies at 70.”
6Freedman, “Maker of worlds,” p. 48.

7Sidney Coleman to the editors of Harvard Magazine, 10 October 1978. The profile appeared in Timothy Noah,
“Four good teachers,” Harvard Magazine 80, no. 7 (September-October 1978): 96–97. My thanks to Tiffany
Nichols for retrieving a copy of the original article.
8Freedman, “Maker of worlds,” p. 48.

9Quoted in Freedman, “Maker of worlds,” p. 48.

10Sidney Coleman to Geoffrey West, 10 October 1978. Samuel Gompers, who founded the American Federation
of Labor (AFL), had been a major figure in the American labor movement during the late nineteenth and early
twentieth centuries.
11SidneyColeman to Avram Davidson, 6 June 1977. Davidson (1923–1993), a longtime friend of Coleman’s, was
an award-winning science fiction author.
12See, e.g., Sidney Coleman to Jack Steinberger, 6 October 1966.
13Antonino Zichichi to Sidney Coleman, 4 March 1969; cf. Coleman to Zichichi, 26 May 1967. Coleman
republished several of his Erice lectures in his book, Aspects of Symmetry: Selected Erice Lectures (New York:
Cambridge University Press, 1985).
14Sidney Coleman to Luis J. Boya, 18 April 1975.
15See, e.g., Sidney Coleman to Gian Carlo Wick, 8 October 1970; Coleman to Zichichi, 5 November 1970.
16Sidney Coleman to Gavin Borden, 12 March 1976. The “Symbionese Liberation Army” was a group of left-wing
radicals that perpetrated several high-profile acts between 1973–75 in the United States, most famously
kidnapping the wealthy publishing heiress Patty Hearst and allegedly “brainwashing” her into supporting their
cause. See Jeffrey Toobin, American Heiress: The Wild Saga of the Kidnapping, Crimes, and Trial of Patty Hearst
(New York: Doubleday, 2016).
17The American Institute of Physics began translating Soviet physics journals into English in 1955, including the
Journal of Experimental and Theoretical Physics (JETP), as part of a Cold War effort to stay current on Soviet
scientists’ advances. See David Kaiser, “The physics of spin: Sputnik politics and American physicists in the
1950s,” Social Research 73 (Winter 2006): 1225–1252.
18Coleman participated in other experiments involving videotaped lectures around the same time: Sheldon A.
Buckler (Vice President of Polaroid Corporation) to Sidney Coleman, 16 April 1974; Peter Wensberg (Senior Vice
President of Polaroid Corporation) to Sidney Coleman, 27 February 1975; and Sidney Coleman to Steven Abbott,
27 February 1976.
19David J. Wallace to Sidney Coleman, 12 June 1980; J. Avron to Sidney Coleman, 12 July 1983.
20Blanche F. Mabee to J. Avron, 19 August 1983; see also David J. Wallace to John B. Mather, 24 June 1980.
21https://www.physics.harvard.edu/events/videos/Phys253
22Many of the original articles are reprinted (in English translation) in B. L. van der Waerden, ed., Sources of
Quantum Mechanics (Amsterdam: North-Holland, 1967). See also Max Jammer, The Conceptual Development of
Quantum Mechanics (New York: McGraw-Hill, 1966); Olivier Darrigol, From c-Numbers to q-Numbers: The
Classical Analogy in the History of Quantum Theory (Berkeley: University of California Press, 1992); and Mara
Beller, Quantum Dialogue: The Making of a Revolution (Chicago: University of Chicago Press, 1999).
23Many important papers from this effort are reprinted (in English translation) in Arthur I. Miller, ed., Early
Quantum Electrodynamics: A Source Book (New York: Cambridge University Press, 1994). See also Silvan S.
Schweber, QED and the Men Who Made It: Dyson, Feynman, Schwinger, and Tomonaga (Princeton: Princeton
University Press, 1994), chap. 1; Tian Yu Cao, Conceptual Developments of 20th Century Field Theories (New
York: Cambridge University Press, 1997), chaps. 6–8; and the succinct historical introduction in Steven Weinberg,
The Quantum Theory of Fields, vol. 1 (New York: Cambridge University Press, 1995), chap. 1.
24See especially Schweber, QED and the Men Who Made It, chap. 2.
25Schweber, QED and the Men Who Made It, pp. xii, 452 and chaps. 7–8; see also Julian Schwinger, “Two
shakers of physics: Memorial lecture for Sin-itiro Tomonaga,” in The Birth of Particle Physics, ed. L. M. Brown and
L. Hoddeson (New York: Cambridge University Press, 1983), pp. 354–375; and Peter Galison, “Feynman’s war:
Modelling weapons, modelling nature,” Stud. Hist. Phil. Mod. Phys. 29 (1998): 391–434.
26See especially Schweber, QED and the Men Who Made It, chap. 5.
27See the articles reprinted in Julian Schwinger, ed., Selected Papers on Quantum Electrodynamics (New York:
Dover, 1958). See also Schweber, QED and the Men Who Made It, chaps. 6–9; and David Kaiser, Drawing
Theories Apart: The Dispersion of Feynman Diagrams in Postwar Physics (Chicago: University of Chicago Press,
2005), chaps. 2–3.
28Kaiser, Drawing Theories Apart, pp. 81–83.
29J. M. Jauch and F. Rohrlich, The Theory of Photons and Electrons (Reading, MA: Addison-Wesley, 1955); S. S.
Schweber, An Introduction to Relativistic Quantum Field Theory (Evanston, IL: Row, Peterson, 1961); J. D.
Bjorken and S. Drell, Relativistic Quantum Mechanics (New York: McGraw-Hill, 1964); Bjorken and Drell,
Relativistic Quantum Fields (New York: McGraw-Hill, 1965). Decades later, Dyson’s 1951 lecture notes were
beautifully typeset by David Derbes and published by World Scientific, so they are readily available today:
Freeman Dyson, Advanced Quantum Mechanics, ed. David Derbes, 2nd ed. (Singapore: World Scientific, 2011).
30See, e.g., Allan Franklin, The Neglect of Experiment (New York: Cambridge University Press, 1986), chap. 1.
31Sidney Coleman described some of this work in a magnificent, brief essay: Coleman, “The 1979 Nobel Prize in
Physics,” Science 206 (14 December 1979): 1290–1292. See also the helpful discussion in Peter Renton,
Electroweak Interactions: An Introduction to the Physics of Quarks and Leptons (New York: Cambridge University
Press, 1990), chap. 5.
32Several original papers are available in Lochlainn O’Raifeartaigh, ed., The Dawning of Gauge Theory
(Princeton: Princeton University Press, 1997).
33See,e.g., L. M. Brown, R. Brout, T. Y. Cao, P. Higgs, and Y. Nambu, “Panel discussion: Spontaneous breaking
of symmetry,” in The Rise of the Standard Model: Particle Physics in the 1960s and 1970s, ed. L. Hoddeson, L.
Brown, M. Riordan, and M. Dresden (New York: Cambridge University Press, 1997), pp. 478–522; and Cao,
Conceptual Developments of 20th Century Field Theories, chaps. 9–10.
34Freeman Dyson, “Field theory,” Scientific American 188 (April 1953): 57–64, on p. 57.
35Richard Feynman to Enrico Fermi, 19 December 1951, as quoted in Kaiser, Drawing Theories Apart, p. 201; see
also ibid., pp. 197–206; and L. Brown, M. Dresden, and L. Hoddeson, eds., Pions to Quarks: Particle Physics in
the 1950s (New York: Cambridge University Press, 1989).
36M. Gell-Mann and Y. Ne’eman, eds., The Eightfold Way (New York: W. A. Benjamin, 1964).
37Quoted in Kaiser,Drawing Theories Apart, p. 306. See also G. F. Chew, S-Matrix Theory of Strong Interactions
(New York: W. A. Benjamin, 1961); Chew, The Analytic S Matrix: A Basis for Nuclear Democracy (New York: W.
A. Benjamin, 1966); and Kaiser, Drawing Theories Apart, chaps. 8–9.
38Coleman, “The 1979 Nobel Prize in Physics,” 1291. See also Martinus Veltman, “The path to renormalizability,”
in Hoddeson et al., The Rise of the Standard Model, pp. 145–178; and Gerard ’t Hooft, “Renormalization of gauge
theories,” in ibid., pp. 179–198. On the experimental detection of weak neutral currents, see Peter Galison, How
Experiments End (Chicago: University of Chicago Press, 1983), chap. 4. Glashow, Weinberg, and Salam shared
the Nobel Prize in Physics in 1979; Veltman and ’t Hooft shared the Nobel Prize in Physics in 1999.
39David Gross, “Asymptotic freedom and the emergence of QCD,” in Hoddeson et al., The Rise of the Standard
Model, pp. 199–232. Politzer, Gross, and Wilczek shared the Nobel Prize in Physics in 2004.
40Laurie
Brown, Michael Riordan, Max Dresden, and Lillian Hoddeson, “The Rise of the Standard Model,
1964–1979,” in Hoddeson et al., The Rise of the Standard Model, pp. 3–35.
41Gross, “Asymptotic freedom and the emergence of QCD,” p. 202.
42Sidney Coleman, The Structure of Strong Interaction Symmetries (Ph.D. dissertation, Caltech, 1962); Coleman,
“An introduction to unitary symmetry” (1966), reprinted in Coleman, Aspects of Symmetry, chap. 1; Georgi,
“Sidney Coleman,” p. 4.
43RichardFeynman, “Spacetime approach to non-relativistic quantum mechanics,” Rev. Mod. Phys. 20 (1948):
367–387; L. D. Faddeev and V. N. Popov, “Feynman diagrams for the Yang–Mills field,” Phys. Lett. B 25 (1967):
29–30. See also Schweber, QED and the Men Who Made It, pp. 389–397; Veltmann, “The path to
renormalizability,” pp. 158–159; Gross, “Asymptotic freedom and the emergence of QCD,” pp. 201–202.
44Transcribed from p. 210 of the handwritten lecture notes for Physics 253B, spring 1990.
45Peter Higgs in “Panel discussion: Spontaneous breaking of symmetry,” p. 509. See also Higgs, “My life as a
boson,” available at http://inspirehep.net/record/1288273/files/MyLifeasaBoson.pdf.
46Coleman, “Secret symmetry: An introduction to spontaneous symmetry breakdown and gauge fields” (1973),
reprinted in Coleman, Aspects of Symmetry, chap. 5. See also S. Coleman and E. J. Weinberg, “Radiative
corrections as the origin of spontaneous symmetry breaking,” Phys. Rev. D 7 (1973): 1888–1910; Coleman, “The
fate of the false vacuum, I: Semiclassical theory,” Phys. Rev. D 15 (1977): 2929–2936; C. G. Callan, Jr. and S.
Coleman, “The fate of the false vacuum, II: First quantum corrections,” Phys. Rev. D 16 (1977): 1762–1768; and
S. Coleman and F. De Luccia, “Gravitational effects on and of vacuum decay,” Phys. Rev. D 21 (1980):
3305–3315.
47ClaudeItzykson and Jean-Bernard Zuber, Quantum Field Theory (New York: McGraw-Hill, 1980); Ta-Pei
Cheng and Ling-Fong Li, Gauge Theory of Elementary Particle Physics (New York: Oxford University Press,
1984).
48SidneyColeman to Mirdza E. Berzins, 19 December 1985. Other leading field theorists shared Coleman’s
emphasis on renormalizability at the time. See, e.g., Steven Weinberg, “The search for unity: Notes for a history of
quantum field theory,” Daedalus 106 (Fall 1977): 17–35; cf. Weinberg, “Effective field theory, past and future,”
Proceedings of Science (CD09): 001, arXiv:0908.1964 [hep-th].
49For interesting historical perspectives on the shift to (nonrenormalizable) effective field theory approaches, see
Tian Yu Cao, “New philosophy of renormalization: From the renormalization group equations to effective field
theories,” in Renormalization: From Lorentz to Landau (and Beyond), ed. L. M. Brown (New York: Springer, 1993),
pp. 87–133; and Silvan S. Schweber, “Changing conceptualization of renormalization theory” in ibid., pp.
135–166. For brief introductions to effective field theory, see Anthony Zee, Quantum Field Theory in a Nutshell,
2nd ed. (Princeton: Princeton University Press, 2010), chap. VIII.3; and Steven Weinberg, The Quantum Theory
of Fields, vol. 2 (New York: Cambridge University Press, 1996), chap. 19. See also Iain W. Stewart’s course on
“Effective Field Theory,” available for free on the edX platform at https://www.edx.org/course/effective-field-theory-mitx-8-eftx .
50Despite my flub on the exam that first semester, Coleman kindly agreed to serve on my dissertation committee.
51SidneyColeman, memo to undergraduate advisees, 6 September 1991; Robert L. Veal to Sidney Coleman, 23
June 1974; Mark Carter to Sidney Coleman, 3 July 1981.
52It
is a pleasure to thank David Derbes for inviting me to contribute this Foreword to the volume, and to Diana
Coleman for sharing copies of Professor Coleman’s correspondence. I am also grateful to Feraz Azhar, David
Derbes, David Griffiths, Matthew Headrick, Richard Sohn, Jesse Thaler, and Aaron Wright for helpful comments
on an earlier draft.

Preface

Sidney Coleman was not only a leader in theoretical particle physics, but also a hugely gifted and dedicated
teacher. In his courses he found just the right balance between rigor and intuition, enlivened by wit, humor and a
deep store of anecdotes about the history of physics. Very often these were first-hand accounts; if he wasn’t a
participant, he was an eyewitness. He made many important contributions to particle theory, but perhaps his most
lasting contribution will prove to be his teaching. For many years he gave a series of celebrated summer school
courses at the International Center for Scientific Culture “Ettore Majorana” in Erice, Sicily, under the directorship of
Antonino Zichichi. A collection of these was published as a book, Aspects of Symmetry, by Cambridge University
Press in 1985. This work, Prof. Coleman’s only previous book, is now recognized as a classic.

Over three decades, Prof. Coleman taught Physics 253, the foundation course on quantum field theory to
Harvard’s graduate students in physics.53 Many of the top American theoretical particle physicists learned
quantum field theory in this course. Alas, he died much too young, at the age of 70, before he took the time away
from his research to write the corresponding textbook. Brian Hill, one of Prof. Coleman’s graduate students at
Harvard (and the Teaching Fellow for the course for three years), had taken very careful notes of the course’s first
seven months from the fall of 1986, about one and a half semesters. He edited and rewrote these after every
class. Xeroxes of Brian’s handwritten notes were made available at Harvard for later classes, and served for
nearly two decades as a de facto textbook for the first part of the course. In 2006, Bryan Gin-ge Chen, a Harvard
undergraduate in physics, asked Brian if he could typeset his notes with the standard software . Bryan got
through Lecture 11. Yuan-Sen Ting, an undergraduate overseas, followed up in 2010, completing the typesetting
of Brian’s notes through Lecture 28. (Yuan-Sen completed a PhD at Harvard in astrophysics; Bryan moved to
Penn for his in physics.) These notes were posted in 2011, with Brian Hill’s introduction, at the arXiv, a free online
repository for papers in physics, mathematics and many other subjects. I found them in the summer of 2013.

Like many, I wrote to Yuan-Sen and Bryan to express my thanks, and asked: Might we see the second
semester some day? Yuan-Sen wrote back to say that unfortunately they did not have a copy of any second
semester notes. I am a high school teacher, and I have been privileged to teach some remarkably talented young
men and women, a few of whom later took Prof. Coleman’s course. One of these, Matthew Headrick at Brandeis,
had not only a set of second semester notes (from a second graduate student, who wishes to remain anonymous),
but also a complete set of homework problems and solutions. He kindly sent these to me. I got in touch with Yuan-
Sen and Bryan and suggested we now type up the second semester together with homework problems and
solutions. Yuan-Sen was trying to finish his thesis, and Bryan had taken a position in Belgium. While they couldn’t
add to the work they’d already done, they offered their encouragement. Yuan-Sen also suggested that I get in
touch with his colleague, Richard Sohn, who had been of significant help to them while typing up Brian’s notes.
Richard had gone through the notes carefully and corrected typos and a few minor glitches in note-taking or the
lectures themselves. Even more enticing, Harvard’s Physics Department, perhaps recognizing that something
special was taking place, had videotaped Prof. Coleman’s entire course in 1975–76. (The cameraman was Martin
Roček, now at Stony Brook.) This was an experiment, as Prof. Coleman himself remarks at the very beginning of
the first lecture. Other courses were videotaped, but it’s noteworthy that, according to Marina Werbeloff, Harvard’s
Physics Librarian, only Prof. Coleman’s tapes continued to circulate for thirty years. Aware that the frequently
borrowed VHS tapes were starting to deteriorate, she had them digitized. In 2007, Maggie McFee, then head of
the department’s Computer Services, set up a small server to post these online in 2008. Perhaps the two
semesters of notes and the videos could provide enough to put together something like the book that Prof.
Coleman might have written himself. Some years earlier I had stumbled onto copies of Freeman Dyson’s famous
Cornell notes (“Advanced Quantum Mechanics”, 1951) at MIT’s website, and with Prof. Dyson’s permission had
typeset these with for the arXiv. Soon thereafter World Scientific contacted Prof. Dyson and me to publish
the notes as a book. Yuan-Sen wondered if World Scientific would be interested in publishing Coleman’s lectures.
I emailed Lakshmi Narayanan, my liaison for Prof. Dyson’s notes. Indeed, World Scientific was very interested.

Now began a lengthy series of communications with all the interested parties. Neither Richard nor I sought
royalties. Prof. Coleman’s widow Diana Coleman is alive and the deserving party. She was happy to allow us to
proceed. I got in touch both with Brian Hill and with the author of the second semester notes; each graciously
agreed to our using their invaluable notes for this project. Through the kindness of Ms. Werbeloff, who responded
after I asked Harvard about using the videotapes, I got in touch with Masahiro Morii, the chair of Harvard’s Physics
Department, who obtained approval from Harvard’s Intellectual Property Department for us to use the videos. Ms.
Werbeloff arranged to have the digitized video files transferred to a hard drive I sent to her. I cloned the returned
drive and sent that to Richard. David Kaiser of MIT, a physicist and historian of physics, and also a former
graduate student of Prof. Coleman’s, generously agreed to write a foreword for these lectures. Additionally, Prof.
Kaiser carefully read the manuscript and provided many corrections. He and Richard visited Ms. Coleman in
Cambridge and got from her xeroxes of Prof. Coleman’s own class notes, a priceless resource, particularly as
these seem to be from the same year as the videotapes. Through Matt Headrick, I was able to contact the authors
of the 1997–98 homework solutions, two of Prof. Coleman’s graduate teaching assistants, David Lee and Nathan
Salwen. They not only gave their permission for their solutions to be used, but generously provided the
source. Finally, we obtained a second set of lecture notes from Peter Woit at Columbia. Richard and I set to work
from five separate records of the course: Brian Hill’s and the anonymous graduate student’s class notes; Prof.
Coleman’s own notes; class notes from Peter Woit; and our transcriptions of the videotaped lecture notes.
Richard, far more conversant with modern field theory than I, would tackle the second semester, while I would
start folding in my transcriptions of Prof. Coleman’s videotaped lectures into the Hill–Ting–Chen notes, together
with homework and solutions. To be sure, there are gaps in nearly all our accounts of the course (though Brian
Hill’s notes are complete, there are pages missing in Prof. Coleman’s notes, quite a few electronic glitches in the
forty-year old videotapes, and so on), but we seem to have a pretty complete record. The years are different
(1975–76 for the videotapes and for Prof. Coleman’s own lecture notes, 1978–79 for Peter Woit’s notes, 1986–87
for Brian Hill’s notes, and spring 1990 for the anonymous graduate student’s), but the correspondence between
these, particularly in the first semester, is remarkably close.54 All of the contributions have been strictly voluntary;
we have done this work out of respect and affection for Sidney Coleman.

Richard and I had been at work for about six months, when David Griffiths, who earned his PhD with Prof.
Coleman, found the Hill–Chen–Ting notes at the arXiv, and wrote Yuan-Sen and Bryan to ask about the second
semester. Yuan-Sen forwarded the email to me, and I wrote back. Prof. Griffiths, now emeritus at Reed College
and the author of several widely admired physics textbooks, also wanted to see Prof. Coleman’s course notes
turned into a book. He has been an unbelievably careful and valuable critic, catching many of our mistakes,
suggesting perhaps a hundred editorial changes per chapter, clarifications or alternative solutions in the
homework, and generally improving the work enormously. Many of the last chapters were read by Prof. Jonathan
L. Rosner, University of Chicago, who cleared up several misunderstandings. The responsibility for all errors, of
course, rests with the last two editors, Richard and me.

The editors are profoundly grateful to all who have so generously offered their time, their expertise and their
work to this project. We are particularly grateful to the talented staff at World Scientific for their hard work and their
immense patience. We hope that were Prof. Coleman alive today, he would be pleased with our second-order
efforts. They are only an approximation. This book can never be the equal of what he might have done, but we
hope we have captured at least a little of his magic. May later generations of physics students learn, as so many
before them have learned, from one of the best teachers our science has known.

David Derbes
The Laboratory Schools
The University of Chicago

53Harvard offered Physics 253, on relativistic quantum theory, for decades before it became Prof. Coleman’s
signature course. For example, in 1965–66 Julian Schwinger taught the class, then called “Advanced Quantum
Theory”. Prof. Coleman also taught the course, in 1968 (and perhaps earlier). In 1974 it was renamed “Quantum
Field Theory”. He first taught this course in 1975–76, repeated it on and off until 1986, and thereafter taught it
annually through the fall of 2002. He was to have taught the second semester in 2003, but his health did not permit
it; those duties fell to Nima Arkani-Hamed (now at IAS, Princeton). I am very grateful to Marina Werbeloff for this
information, which she gleaned at my request by searching through fifty years of course catalogs. I thank her for
this and for much other assistance, without which the project likely would not have been completed.
54Very late in the project, we obtained a set of class notes, problems and exams from 2000–01, courtesy of
another former student, Michael A. Levin of the University of Chicago. Through Michael we were able to get in
touch with Prof. Coleman’s last Teaching Fellow (1999–2002), Daniel Podolsky of the Israel Institute of
Technology (Technion), Haifa. Daniel had two sets of typed notes for the course; his own, beautifully ’ed,
which Michael had originally provided, and other notes (but missing many equations) from spring, 1999. A
Harvard student hired by Prof. Coleman recorded the lectures as she took notes, and typed them up at home. The
following summer Daniel worked with Prof. Coleman to edit the notes, but they did not get very far. Daniel’s notes
(both sets) were used primarily to check our completed work, but in a few places we have incorporated some very
valuable insights from them.
 Frequently cited references

Abers & Lee

Ernest S. Abers and Benjamin W. Lee, “Gauge Theories”, Phys. Lett. 9C (1973) 1–141.
GT
Arfken & George B. Arfken and Hans J. Weber, Mathematical Methods for Physicists, 6th ed., Elsevier,
Weber MMP 2005.
Bjorken &
James D. Bjorken and Sidney D. Drell, Relativistic Quantum Mechanics, McGraw-Hill, 1961.
Drell RQM
Bjorken &
James D. Bjorken and Sidney D. Drell, Relativistic Quantum Fields, McGraw-Hill, 1962.
Drell Fields
Cheng & Li Ta-Pei Cheng and Ling-Fong Li, Gauge Theory of Elementary Particle Physics, Oxford U.P.,
GT 1985.
Close IP Frank Close, The Infinity Puzzle, Basic Books, 2011, 2013.
Coleman
Sidney Coleman, Aspects of Symmetry: Selected Erice Lectures, Cambridge U.P., 1985.
Aspects
Crease & Robert P. Crease and Charles C. Mann, The Second Creation: Makers of the Revolution in 20th-
Mann SC Century Physics, Collier-Macmillan Publishing, 1986. Reprinted by Rutgers U.P., 1996.
Goldstein et Herbert Goldstein, Charles P. Poole, and John L. Safko, Classical Mechanics, 3rd. ed., Addison-
al. CM Wesley, 2001.
Gradshteyn & Izrael S. Gradshteyn and Iosif M. Ryzhik, Table of Integrals, Series and Products, 4th. ed.,
Ryzhik TISP Academic Press, 1965.
Greiner &
Walter Greiner and Berndt Müller, Gauge Theory of Weak Interactions, 4th ed., Springer, 2009.
Müller GTWI
Greiner &
Walter Greiner and Berndt Müller, Quantum Mechanics: Symmetries, 2nd ed., Springer, 1994.
Müller QMS
Greiner &
Walter Greiner and Joachim Reinhardt, Field Quantization, Springer, 1996.
Reinhardt FQ
Greiner &
Reinhardt Walter Greiner and Joachim Reinhardt, Quantum Electrodynamics, 4th ed., Springer, 2009.
QED
Greiner et.al W. Greiner, Stefan Schramm, and Eckart Stein, Quantum Chromodynamics, 2nd rev. ed.,
QCD Springer, 2002.
Griffiths EP David Griffiths, Introduction to Elementary Particles, 2nd rev. ed., Wiley-VCH, 2016.
Itzykson & Claude Itzykson and Jean-Bernard Zuber, Quantum Field Theory, McGraw-Hill, 1980. Reprinted
Zuber QFT by Dover Publications, 2006.
J. David Jackson, Classical Electrodynamics, 3rd ed., John Wiley, 1999.
Jackson CE

Landau & Lev D. Landau and Evgeniĭ M. Lifshitz, Quantum Mechanics: Non-Relativistic Theory, 2nd ed.,
Lifshitz QM Addison-Wesley, 1965.
Lurié P&F David Lurié, Particles and Fields, Interscience Publishers, 1968.
PDG 2016 C. Patrignani et al. (Particle Data Group), Chin. Phys. C 40 (2016) 100001, http://pdg.lbl.gov.
Peskin &
Michael E. Peskin and Daniel V. Schroeder, An Introduction to Quantum Field Theory, Westview
Schroeder
Press, 1995
QFT
Rosten Joys Leo Rosten, The New Joys of Yiddish, rev. Lawrence Bush, Harmony, 2003.
Ryder QFT Lewis H. Ryder, Quantum Field Theory, 2nd ed., Cambridge U.P., 1996.
Schweber
Silvan S. Schweber, QED and the Men Who Made It, Princeton U.P., 1994.
QED
Schweber Silvan S. Schweber, An Introduction to Relativistic Quantum Field Theory, Row, Peterson & Co.,
RQFD 1960. Reprinted by Dover Publications, 2005.
Schwinger
Julian Schwinger, ed., Selected Papers on Quantum Electrodynamics, Dover Publications, 1959.
QED
Weinberg
Steven Weinberg, The Quantum Theory of Fields I: Foundations, Cambridge U.P., 1995, 1995.
QTF1
Weinberg
Steven Weinberg, The Quantum Theory of Fields II: Modern Applications, Cambridge U.P., 1996.
QTF2
Zee GTN Anthony Zee, Group Theory in a Nutshell for Physicists, Princeton U.P., 2016.
Zee QFTN Anthony Zee, Quantum Field Theory in a Nutshell, 2nd ed., Princeton U. P., 2010.

Index of useful formulae

A note on the problems

The classes in Physics 253a (fall) and Physics 253b (spring) ran as two ninety-minute lectures per week. Students
were assigned problem sets (from one to four problems) nearly every week, and given solutions to them after the
due date. As this book has fifty chapters, it seemed reasonable to include twenty-five problem sets. These include
all the assigned problems from 1997–98 (with two exceptions),1 some additional problems from other years that
were not assigned in 1997–98, and a handful of final examination questions. In 1975–76, Coleman began the
second semester material a little early, in the last part of the last lecture of 253a, Chapter 25. This material was
moved forward into Chapter 26, which marks the beginning of the second semester, and an approximate dividing
line for the provenance of the problems. Usually, those from 253a are placed before Chapter 26; those from 253b,
after. Problems 14 are transitional: though assigned in the second semester, they involve first semester material.

The editors obtained complete sets of assigned problems and examinations (and their solutions) from the
year 1978–79, the years 1980–82, and the year 1986–87 from Diana Coleman (via David Kaiser); 1990–91, from
Matthew Headrick; 1997–98 from Matthew Headrick, Nathan Salwen and David Lee; 2000–01 from Michael
Levin; and examination questions from 1988–2000 from Daniel Podolsky. John LoSecco provided a problem cited
in the video of Lecture 50 (Problem 4 on the 1975a Final) and its solution, which appears here as Problem 15.4. In
fact, only a few assigned problems from these other years do not appear in this book. Most of the problems were
used over and over throughout the roughly thirty years that Coleman taught the course; sometimes a problem
used in an examination was assigned for homework in later years, or vice versa. The solutions were written up by
Teaching Fellows (notably by Brian Hill, but very probably some are due to Ian Affleck, John LoSecco, Bernard
Grossman, Katherine Benson, Vineer Bhansali, Nathan Salwen, David Lee, and Daniel Podolsky, among many
others unknown to us). Some solutions, particularly to the exam questions, are by Coleman himself. It’s hard to
know the authorship of many solutions—the same problems assigned ten years apart often have essentially
identical solutions, though in different handwriting, and we may not have the original author’s work. Now and then
the editors have added a little to a problem’s solution, but most of the solutions are presented just as they were
originally.

A century ago, it was customary in British mathematics textbooks to cite the provenance (if known) of
problems; e.g., an exercise taken from the Cambridge Mathematical Tripos was indicated by the abbreviation “MT”
and the year of the examination. Here, (1998a 2.3) indicates Problem 2.3 assigned in the fall of 1998. To aid the
reader in finding a particular problem, succinct statements of them are given below. (Incidentally, the Paracelsus
epigraph2 in Problems 1 comes from the first 253a assignment in 1978.)

1.1Show that the measure is Lorentz invariant.

1.2Show that á0|T(ϕ(x)ϕ(y))|0ñ is a Green’s function for the Klein–Gordon equation.

1.3Show that the variance of ϕ(x) over large regions is tiny, and roughly classical; over small regions, it
fluctuates as a typical quantum system.

2.1Find the dimensions of various Lagrangians in d spacetime dimensions.

2.2Rework Problem 1.3 from the point of view of dimensional analysis.

2.3Obtain the Maxwell equations from the Maxwell Lagrangian, L = − FµνFµν.

2.4Obtain both the canonical energy-momentum tensor, and an improved, symmetric version, for the
Maxwell field.

3.1Obtain Schrödinger’s equation from a given Lagrangian.

3.2Quantize the theory in 3.1.

3.3Show that only two of the symmetries {P, C, T} have corresponding unitary operators for this theory.

3.4Examine dilation invariance for the massless Klein–Gordon theory.

4.1Evaluate the real constant α in Model 1 in terms of its only Wick diagram, .

4.2Demonstrate various properties of the coherent states of a single harmonic oscillator.

4.3Obtain expectation values of an operator in terms of a generalized delta function.

5.1Evaluate áp|S − 1|p′ñ for the pair model, and show that its S-matrix is unitary, i.e.áp|S†S − 1|p′ñ = 0.

6.1Let Model 3 describe kaon decay into pions (K ~ ϕ, π ~ ψ), and determine the value of g/mK to one
significant digit.

6.2Compute dσ/dΩ (c.o.m. frame) for elastic NN scattering in Model 3 to lowest order in g.

6.3Compute dσ/dΩ (c.o.m. frame) for N + N → 2π in Model 3 to lowest order in g.

6.4Determine the behavior of the S-matrix in a free scalar field under an anti-unitary operator (as
required for CPT symmetry).

7.1Determine the two-particle density of states factor in an arbitrary frame of reference.

7.2Calculate the decay A → B + C + D in a theory of four scalar fields {A, B, C, D} if A is massive but
the other three are massless, with L′ = gABCD.

7.3Determine the density of states factor for particle decay if the universe is filled with a thermal
distribution of mesons at a temperature T.

8.1Replace a free Klein–Gordon field ϕ by ϕ = A + gA 2, and show that to O(g2) the sum of all A-A
scattering graphs vanishes.

9.1Calculate the imaginary part of the renormalized meson self-energy ′(p2) in Model 3 to O(g2).

9.2Compute the Model 3 vertex −i ′(p2, p′2, q2) to O(g3) as an integral over two Feynman parameters,
for p2 = p′2 = m2.

10.1Calculate the renormalized “nucleon” self-energy, ′(p2), in Model 3 to O(g2), expressing the
answer as an integral over a single Feynman parameter.

10.2Verify the Lie algebra of the Lorentz group’s generators using the defining representation of the
group.

11.1Find the positive energy helicity eigenstates of the Dirac equation.

11.2Work out trace identities for various products of Dirac gamma matrices.

12.1Attempt the canonical quantization of a free Klein–Gordon field ϕ(x) with anticommutators, and
show that the Hilbert space norm of áϕ|{θ, θ†}|ϕñ cannot be positive.

12.2Compute, to lowest nontrivial order, dσ/dΩ (c.o.m. frame) for the scattering N + ϕ → N + ϕ if L′ =
gψψϕ (the “scalar” theory).

12.3Compute, to lowest nontrivial order, dσ/dΩ (c.o.m. frame) for the scattering N + N → N + N if L′ =
gψiγ5ψϕ (the “pseudoscalar” theory).

13.1In the “pseudoscalar” theory, L′ = gψiγ5ψϕ, calculate the renormalized “nucleon” self-energy,
′(k2), to O(g2). Leave your answer in terms of an integral over a single Feynman parameter.

13.2In the same theory, compute the renormalized meson self-energy, ′(k2), to O(g2). Again, leave
your answer in terms of an integral over a single Feynman parameter. Check that the imaginary
part of this quantity has the correct (negative) sign.

14.1Given an interaction Hamiltonian of four-fermion interactions whose S-matrix is CPT-invariant,

show that the Hamiltonian is itself invariant. Investigate under what circumstances it is invariant
under the sub-symmetries of CPT, e.g., PT and P.

14.2Derive the superficial degree of divergence D for a general Feynman graph in d spacetime
dimensions.

14.3In the generalized “pseudoscalar” theory, L′ = gNiγ5τ • πN, calculate various N + π → N′ + π′

amplitudes, using isospin invariance.
14.4Show that to O(g2), the original “pseudoscalar” theory’s scattering amplitudes coincide with those
of a second Lagrangian, L′′ = µ–1 , for appropriate choices of a and b.

15.1Investigate the (four dimensionally) longitudinal solutions of the free Proca Lagrangian, and
construct its Hamiltonian. Show that for an appropriate identification of A0 and its conjugate
momentum with ϕ and π of the Klein–Gordon equation, the Hamiltonians of the two theories are
identical.

15.2Construct the Hamiltonian of a free, massive vector in terms of its creation and annihilation
operators.

15.3Let Aµ be a vector of mass µ be coupled to two Dirac fields ψ 1 and ψ 2 of mass m1 and m2,
respectively, according to the interaction Lagrangian L′ = gA µ(ψ 1γµψ 2 + ψ 2γµψ 1). Compute the
decay width Γ for ψ 1 → ψ 2 + γ if m1 > m2 + µ to lowest nonvanishing order.

15.4Compute elastic meson–meson scattering to O(g2) in a scalar theory with the interaction L′ =
−(1/4!)gϕ′4 − (1/4!)Cϕ′4 in terms of the Mandelstam variables s, t, and u. Define the counterterm C
by the requirement that i = −ig when all four mesons are on the mass shell, at the symmetry point
where the Mandelstam variables all equal 4µ2/3.

16.1A Dirac field is minimally coupled to a Proca field of mass µ. Compute, to lowest nontrivial order,
the amplitude for elastic fermion–antifermion scattering, and show that the part proportional to
kµkν/µ2 vanishes.

16.2In this same theory, compute the amplitude for elastic vector–spinor scattering to lowest nontrivial
order, and show that if the meson’s spin vector εµ is aligned with its four-momentum kµ (for either
the outgoing or incoming vector), the amplitude vanishes. Repeat the calculation, substituting a
scalar for the spinor.

16.3Two Dirac fields A and B of masses mA and mB interact with a complex charged scalar field C of
mass mC according to the Lagrangian L′ = g(Aiγ5BC + Biγ5AC*). Let the fields be minimally
coupled to a Proca field, and let their charges (in units of e) be qA, qB, and qC, such that qA = qB +
qC. Show that the amplitude for γ + A → B + C vanishes to lowest order (eg) if the Proca spin is
aligned with its four-momentum.

17.1A scalar field is quadratically coupled to a source J, i.e., with an interaction term Jϕ 2. From
Chapter 27, it can be shown that á0|S|0ñJ = (det[A − iϵ]/ det[K − iϵ])–1/2, where A = ( 2 + µ2 − J),
and K = ( 2 + µ2). Show that you obtain the same result by summing Feynman graphs.

17.2Using functional integrals, determine the photon propagator D Cµν in Coulomb gauge, ∇•A = 0.

17.3Compute, to O(e2), the invariant Feynman amplitude for electron–electron scattering in both the
Coulomb and Feynman gauge, and show that the final answers are the same.

18.1Compute, to O(e2), the renormalized photon self-energy ′µν(p2) in the theory of a charged Dirac
field minimally coupled to a massless photon. Write the answer as an integral over a single
Feynman parameter, and handle the divergences with Pauli–Villars regulator fields.

18.2Compute, to O(e2), the renormalized photon self-energy ′µν(p2) in the theory of a charged
spinless meson minimally coupled to a photon. Write the answer as an integral over a single
Feynman parameter, and handle the divergences with dimensional regularization.

18.3Add to the standard Maxwell Lagrangian the interaction term L′ = − λ(∂µAµ + σAµAµ)2 and a ghost
Lagrangian Lghost. Determine the latter, the ghost propagator, and the Feynman rules for the
ghost vertices.

19.1Carry out computations for a charged scalar particle minimally coupled to a massless photon
parallel to those earlier calculations for a charged Dirac particle: Ward identity and its verification
at tree level, identification of the normalized charge with the physical charge, determination of
 F1(q2) and F2(q2).

19.2Compute the decay width Γ for the process ψ 1 → ψ 2 + γ if m1 > m2, for the Lagrangian L′ =
gψ 2σµνψ 1Fµν + h.c.

20.1The errors on the anomalous magnetic moments, and hence on 1 + F2(0), of the electron and the
muon are 3 × 10–11 and 8 × 10−9, respectively. What bounds do these place on a hypothetical
massive photon whose mass M is much greater than the muon’s mass?

20.2Express the renormalized photon propagator (in Landau gauge) in terms of its spectral
representation with spectral function ρ(k2), and show that the hadronic contribution ρH(a2) is
proportional, to O(e4) and O(e2m2/a2), to the total cross-section σT for e+-e− → hadrons.

21.1Show that the two SU(3)-invariant quartic self-couplings of the pseudoscalar octet, Tr(ϕ 4) and
(Tr(ϕ 2))2, are proportional to each other.

21.2Show that the magnetic moments within the SU(3) decuplet are proportional to the charge.

21.3Assuming that the magnetic moments of quarks are proportional to their charges, µ = κqσ, where
σ is the vector of Pauli matrices, determine the ratio of the proton and neutron magnetic moments,
and compare with experiment.

22.1Consider the scattering of two distinct, spinless particles below inelastic threshold. Find the
relation between the s-wave scattering length a and the invariant Feynman amplitude, A,
evaluated at threshold.

22.2Consider a massless neutrino and an electron coupled to a Proca field W of mass M. For the
process ν + ν → W + W, there are nine independent amplitudes. Find them. Some are well-
behaved at high energy, but others grow without limit. Which are which? Show that all amplitudes
become well-behaved by the addition of new terms to the Lagrangian,
, if f is chosen proportional to g.

23.1A charged scalar ψ of mass m is minimally coupled to the photon. A second massless neutral
meson ϕ is coupled through the term L′ = gϕϵµνλσFµνFλσ. Determine dσ/dΩ to O(e2g2) for the
process γ + ψ → ϕ + ψ.

23.2Starting from the Goldstone model, find a solution ϕ(z) of the field equations, such that ϕ(±∞) =
±a. These solutions could represent “domain walls” in the early universe. Find the energy of these
domain walls in terms of the Goldstone parameters λ and a.

24.1Verify an approximation (44.51) used in the derivation of the scalar field’s effective potential.

24.2Consider the full Yukawa theory of a triplet of pions and the nucleon doublet, with isospin-invariant
interaction L′ = −igNγ5τ • ΦN and a quartic pion self-coupling λ(Φ • Φ)2 + LCT. Let the fields now
be minimally coupled to a massless photon, and determine the contributions to the proton and
neutron form factors F2(0).

24.3Consider the Goldstone model minimally coupled to a Proca field with mass µ0. What is the mass
of this “photon” after the symmetry breaks? Does the Goldstone boson survive, and if so, what is
its mass?

25.1A free Proca field of mass µ is coupled to a real scalar field ϕ of mass m by the interaction
Lagrangian L′ = gA µAµϕ. There are nine independent amplitudes for the process A + ϕ → A + ϕ,
some well-behaved at high energy, and some not. Find them. Which are which? Show that all
become well-behaved with the addition of a new term, hϕ 2AµAµ, for an appropriate choice of h.

25.2From the infinitesimal form of the non-Abelian gauge transformation, determine the finite
(integrated) form, and show that its corresponding unitary matrix U(s) satisfies a particular
differential equation.

25.3Compute k′µMµ for the elastic scattering of non-Abelian gauge bosons off Dirac particles in the
 tree approximation (i.e., to O(g2)) where Mµ is the matrix element of a conserved current, by
setting ε′µ* = k′µ.

25.4Compute, to O(g2), elastic vector–scalar scattering in the Abelian Higgs model, for the case in
which both the initial and final vector mesons have zero helicity, but at fixed scattering angle θ(≠ π,
0). Show that the amplitude approaches a limit at high energy, even though some individual
graphs grow with energy.

1[Eds.] Two questions were omitted, as the videotaped lectures of 1975–76, on which the text is based, include
their solutions: (1997a 2.3), on the form of the energy-momentum tensor for a scalar field; and (1998b 10.1), on
the mixing angle for the ρ and ω eigenstates of the mass-squared matrix for the JP = 1− meson octet. The first is
worked out in §5.5, (5.52)–(5.58); the second appears in §39.3, (39.19)–(39.35).
2[Eds.]Theophrastus von Hohenheim (1493–1541), known as Paracelsus, a pioneering Swiss physician,
alchemist, and astrologer.

1
Adding special relativity to quantum mechanics

1.1Introductory remarks

This is Physics 253, a course in relativistic quantum mechanics. This subject has a notorious reputation for
difficulty, and as this course progresses, you will see this reputation is well deserved. In non-relativistic quantum
mechanics, rotational invariance simplifies scattering problems. Why does adding in special relativity, to include
Lorentz invariance, complicate quantum mechanics?

The addition of relativity is necessary at energies E ≥ mc2. At these energies the reaction

is possible. At slightly higher energies, the reaction

can occur. The exact solution of a high energy scattering problem necessarily involves many-particle processes.

You might think that for a given E, only a finite number, maybe only a small number, of processes actually
contribute. But you already know from non-relativistic quantum mechanics that this isn’t true. For example, if a
perturbation δV is added to the Hamiltonian H, the ground state energy E0 changes according to the rule

Intermediate states of all energies contribute, suppressed by energy denominators.

For highly accurate calculations at low energy, it’s reasonable to include relativistic effects of order (v/c)2.
Intermediate states with extra particles will contribute corrections of the same order:

As a general conclusion, the corrections of relativistic kinematics and the corrections from multi-particle
intermediate states are comparable; relativity forces you to consider many-body problems.

There are however very special cases, due to the specific dynamics involved, where the kinematic effects of
relativity are considerably larger than the effects of pair states. One of these is the hydrogen atom. That’s why
Dirac’s theory1 gives excellent results to order (v/c)2 for the hydrogen atom, even without considering pair
production and multi-particle intermediate states. This is a fluke.2 Dirac’s success was a good thing because it
told people that the basic ideas were right, but it was a bad thing because it led people to spend a lot of time
worrying about one-particle, two-particle, and three-particle theories, because they didn’t realize the hydrogen
atom was a very special system. We will see that you cannot have a consistent relativistic picture without pair
production.

Units

Because we’re doing relativistic (c) quantum mechanics (ħ), we choose units such that

This leaves us with one unit free. Typically we will choose it in a given problem to be the mass of an interesting
particle, which we will then set equal to one. We’ll never get into any problems with that. Just remember that an
ordinary macroscopic motion like scratching your head has infinitesimal velocity and astronomical angular
momentum! Consequently, in terms of dimensions,

Also, it’s useful to know

We will say things like the inverse Compton wavelength of the proton is “1 GeV”.

Lorentz invariance

The arena for all the physics we’re going to do is Minkowski space, flat spacetime in which there are a bunch
of points labeled by four coordinates. We write these coordinates as a 4-vector:

Sometimes I will suppress the index µ when there’s no possibility of confusion and simply write xµ as x. This is not
the only four-component object we will deal with. In classical mechanics there is also the momentum of a particle,
which we can call pµ;

The zeroth component of this 4-vector, the time component, has a special name: the energy. The space
component p is of course called the momentum, and sometimes I will write pµ as p. I can indiscriminately write p
as k, because ħ = 1. The time component k0 of kµ is the frequency ω. Any contravariant 4-vector aµ can be written
as aµ = (a0, ai) = (a0, a); similarly the covariant 4-vector aµ = (a0, ai) = (a0, −a). The four-dimensional inner product
a · b between two 4-vectors aµ, bν is

where a • b is the usual 3-vector inner product. We will as above adopt the so-called Einstein summation
convention, I presume familiar to you, where sums over repeated indices are implied. This inner product is
invariant under Lorentz transformations. Please note I have adopted the “west coast” metric signature,3 (+ − − −).
The inner product of a 4-vector aµ with itself usually will be written a2;

The inner product can also be written as

where the metric tensor gµν is defined by

This object is used to lower indices;

It is convenient to have an object to raise indices as well. We define the metric tensor with upper indices as the
inverse matrix to the metric tensor with lower indices;
where δνµ is the conventional Kronecker delta,

This is an easy equation to solve; gµν is numerically equal to gµν if we have units such that c = 1.

Lorentz transformations on 4-vectors will be denoted by 4 × 4 matrices Λµν. These act on 4-vectors as
follows:

Because of the invariance of the inner product,

The Lorentz transformations form a group in the mathematical sense: The product of any two Lorentz
transformations is a Lorentz transformation, the inverse of a Lorentz transformation is a Lorentz transformation
and so on. This group has a name: O(3, 1). The O stands for the orthogonal group. The (3, 1) means that it’s not
quite an orthogonal group because three of the terms in an inner product have one sign and the fourth has the
other. This group is in fact a little too big for our purposes, because it includes transformations which are not
invariances of nature: parity and time reversal which as you probably know are broken by the weak interactions.
We will restrict ourselves in this course strictly to the connected Lorentz group, those Lorentz transformations
which can be obtained from the identity by continuous changes. Thus we exclude things like parity and time
reversal. Mathematicians call the connected4 Lorentz group, SO(3, 1), with the S meaning “special”, in the sense
that the determinant of the matrix equals 1. If we were talking about rotations, we would be looking not at all
orthogonal transformations, but rotations in the proper sense, excluding reflections. Every element of the full
Lorentz group can be written as a product of an element of the connected Lorentz group with one of the following:
{1, P, T, PT}. The parity operator P reflects all three-space components,

The time reversal operator T reflects the time t; T : t → −t, and PT is the product of these. By Lorentz invariance we
will mean invariance under SO(3, 1).

Under the action of the Lorentz group, 4-vectors fall into three classes: timelike, spacelike and null (or
lightlike). These terms describe the invariant square of a 4-vector aµ;

The same terms are applied to 4-vector inner products. Given two 4-vectors x and y, the invariant square of the
difference (x − y) between them, (xµ − yµ)(xµ − yµ) = (x − y)2, will be called the separation or the interval.

Actually the world is supposed to be invariant under a larger, though no more glamorous, group, which
contains the homogeneous Lorentz group as well as space-time translations; this is the Poincar´e group. Nobody
found that exciting because invariance under translations was known in Newton’s time. Nevertheless we will have
occasion to consider this larger group. Its elements are labeled by a Lorentz transformation Λ and a 4-vector a.
They act on space-time points by Lorentz transformation and translation through a.

Conventions on integration, differentiation and special functions

The fundamental differential operator is denoted ∂µ, defined to be

It acts on functions of space and time. Note that I have written the operator with a lower index, while I have written
xµ with an upper index. This is correct. The operator ∂µ does not transform like a contravariant vector aµ, but
instead like a covariant vector aµ. The easy way to remember this is to observe that
by definition. If we wrote both the operator and the coordinate with lower indices, we should have a g rather than a
δ on the right-hand side. An object almost as important as the Laplace operator ∇2 is the d’Alembert operator ∂2,
which we’ll write as 2,

This is a Lorentz invariant differential operator.5

Now for integration. When I don’t put any upper or lower limits on an integral, I mean that the integral is to run
from −∞ to ∞. In particular, a four-dimensional integral over the components of a 4-vector aµ;

Delta functions over more than one variable will be written as δ(3)(x) for three dimensions or δ(4)(x) for four
dimensions. If we define the Fourier transform (k) of a function F(x) as

where k and x are both 4-vectors, then

I will try to adopt the convention that every dk (or dp) has a denominator of 2π. This will unfortunately lead me to
writing down square roots of 2π at intermediate stages. But I will craftily arrange matters so that in the end all
factors of dk will carry denominators of 2π, and there will be no other place a 2π comes from. That’s important.
Sometimes we get sloppy, and act like 1 = −1 = 2π and 1/(2π) = = “one-bar” or something. Well, suppose you
predict a result from a beautiful theory. Someone asks if it is measurable, and you say, yes it is. You’re going to
feel pretty silly if they spend a million and a half dollars to do the measurement and can’t find it because you’ve
put a (2π)2 in a numerator when it should have been in the denominator. ..

There’s one last function I will occasionally use, θ(x), the theta function.6 The theta function is defined by

Its value at the jump, x = 0, will be irrelevant in every place we use the function. The derivative of the theta function
is a delta function;

We are now ready to investigate our very first example of a relativistic quantum system.

1.2Theory of a single free, spinless particle of mass µ

The state of a spinless particle is completely specified by its momentum, and the components of momentum form
a complete set of commuting variables:7

The states are normalized by the condition

The statement that these kets |pñ form a complete set of states, and that there are no others, is written

so that any state |ψñ can be expanded in terms of these;

If we were doing non-relativistic quantum mechanics, we’d finish describing the theory by giving the Hamiltonian
H, and thus the time evolution of the states; H|pñ = (|p|2/2µ)|pñ.

For relativistic quantum mechanics, we take instead

That’s it, the theory of a single free, spinless particle, made relativistic.

How do we know that this theory is Lorentz invariant? Just because it contains one relativistic formula does
not necessarily mean it is relativistic. The theory is not manifestly Lorentz invariant. The theory is however
manifestly rotationally and translationally invariant. Let’s be more precise about this.

Translation invariance

To any active translation specified by a given 4-vector aµ, there should be a linear operator U(a) satisfying
these conditions:

The operator U satisfying these conditions is U(a) = eiP·a where Pµ = (H, P).

Aside. I’ve laid out this material in pedagogical, not logical order. The logical order would be to state:

1.We want to set up a translationally invariant theory of a spinless particle. The theory will contain
unitary translation operators U(a).

2.Define Pi as

From (1.34), [Pi, Pj] = 0; and from (1.32), P = P†.

3.Declare Pi to be a complete set and classify the states by momentum.

4.Define H = , and thus give the time evolution.

Continuing with the pedagogical order:

States described by kets are transformed by U(a) = eiP·a as follows:

where |xñ means a state centered at xµ; |0ñ means a state centered at the origin. Operators O transform as

and expectation values transform as

Reducing the transformations to space translations,

Only operators localized in space transform according to this rule. The position operator does not:
which looks like the opposite of the operator transformation rule (1.39) given above. The is not an operator
localized at q, so there is no reason for these last two equations to look alike.

Rotational invariance

Given a rotation R ∈ SO(3), there should be a unitary operator U(R) satisfying these conditions:

Denote a transformed ket by |ψ′ñ = U(R)|ψñ, and require for any |ψñ the rule

so we get

A U(R) satisfying all these properties is given by

That (1.42) and (1.43) are satisfied is trivial. To prove (1.41), insert a complete set between U and U †:

Let p′ = Rp; the Jacobian is 1, so d3p′ = d3p, and

To prove (1.45), write

The proof of (1.46) is left to you.

Constructing Lorentz invariant kets

Our study of rotations provides a template for studying Lorentz invariance. Suppose a silly physicist took for
normalized three-momentum states the kets |pñS defined by

These kets are normalized by the condition

The completeness relation is

If our silly physicist now took U S(R)|pñS = |RpñS, his proofs of (1.41), (1.45) and (1.46) would break down, because
Let’s apply this lesson. The usual 3-space normalization, áp|p′ñ = δ(3)(p − p′), is a silly normalization for Lorentz
invariance; d3p is not a Lorentz invariant measure. We want a Lorentz invariant measure on the hyperboloid p2 =
(p0)2 − |p|2 = µ2, p0 > 0. The measure d4p is Lorentz invariant. To restrict it to the hyperboloid, multiply it by the
Lorentz invariant factor δ(p2 − µ2)θ(p0). That yields our relativistic measure on the hyperboloid8

Figure 1.1: Restricting |dp| to the invariant hyperboloid p2 = µ2

where

Later on, we’ll want factors of 2π to come out right in Feynman diagrams, so we’ll take for our relativistically
normalized kets |pñ

so that

From the graph of the hyperbola, it looks like the factor multiplying d3p ought to get larger as |p| gets large. This is
an illusion, caused by graphing on Euclidean paper. It’s the same illusion that occurs in the Twin Paradox: Though
the moving twin’s path appears longer, in fact that twin’s proper time is shorter.

Now let’s demonstrate Lorentz invariance. Given any Lorentz transformation Λ, define

The unitary operator U(Λ) satisfies these conditions:

The proofs of these are exactly like the proofs of rotational invariance, using the completeness relation

and the invariance of the measure,

1.3Determination of the position operator X

We have a fairly complete theory, except that we don’t know where anything is; a particle could be at the origin or
at the Andromeda galaxy. In non-relativistic quantum mechanics, if a particle is in an eigenstate of a position
operator X, its position x is its eigenvalue. Can we construct a position operator, X, for our system? Fortunately we
can write down some general conditions about such an operator, conditions we can all agree are perfectly
reasonable, which will be enough to specify this operator uniquely.9 And then there will be a surprise, because
we’ll find out that this uniquely specified operator is totally unsatisfactory! There will be a physical reason for that.

What conditions do we want our X operator to satisfy? These conditions will not involve Lorentz invariance,
but only invariance under rotations and translations in space:

We impose the first condition because x is an observable. The second condition is the rule (1.40). The third
condition says that X transforms as a 3-vector, so we might as well write it as X or its components as Xi. Then, by
taking ∂/∂ai of the second condition and evaluating at ai = 0, we get the usual commutator i[Pi, Xj] = δij. Now you
see a new origin for this familiar equation.

From the commutator, we can deduce something about Xi;

where R i is a remainder that must commute with Pj in order to give us the right result. We know that this
expression for Xi has the right commutation relations. Now to find R i.

We know something about our system. We know that the three components Pi are a complete set of
commuting operators. From non-relativistic quantum mechanics, we know that anything that commutes with a
complete set of commuting operators must be a function of those operators. Therefore R i must be some function
of the Pi’s. According to the third condition (1.68), Xi must transform as a 3-vector, and so must R i. That tells us R i
must be of the form

where F(|p|2) is an unknown function of |p|2. But any such function of this form is a gradient of some scalar
function G(|p|2); that is,

This specifies the position operator to be

We can do more. We can eliminate the remainder term entirely by changing the phase of the P states:

I’m perfectly free to make that reassignment. It does not affect the physics of theses states. These are still
eigenstates of Pi with eigenvalues pi, they are still eigenstates of H with eigenvalues , and they are still
normalized in the same way. This is a unitary transformation; call it U(G):

and so the operators change accordingly:

The only formula this transformation affects in all we have done so far is the expression for Xi on |pñ. Now we have

Thus the unique candidate for the Xi operator—providing one chooses the phase for the eigenstates
appropriately—is nothing more nor less than the good old-fashioned Xi operator in non-relativistic quantum
mechanics, the operator which in P space is i∂/∂pi. Let’s make this choice, and drop the G subscripts from now on.

Now that we have found our Xi operator, we know where our particle is. Or do we? Let’s do a thought
experiment. If we really have, in a relativistic theory, a well-defined position operator, we should be able to say of
our particle that it does not travel faster than light. That is, we can start out with a state where our particle is
sharply localized, say at the origin, 10 allow that state to evolve in time (according to the Schrödinger equation,
since we know the Hamiltonian), and see if at some later time there is a non-zero probability for the particle to
have moved faster than the speed of light. We have all the equipment; we need only do the computation. Let’s do
it.

We start out with a state |ψñ localized at the origin at time t = 0, i.e.,

Because the Xi operator is its usual self, we can make use of the usual relation11

and so at t = 0

We wish to compute the probability amplitude for the particle to be found at position x at time t, which by the
general rules of quantum mechanics is given by

The operators Xi and Pj are the same; only the Hamiltonian is novel. Thus we can do the computation, we just put
the pieces together:

because H|pñ = ωp|pñ, and so áx|e−iHt|pñ = e−iωptáx|pñ = e−iωpt+ip•x/(2π)3/2. Compute the integral in the usual way
by going over to polar coordinates,

(letting r = |x|, p = |p| and ωp = .) This is a messy integral, full of oscillations. It’s difficult to tell if it
vanishes outside the light cone, or not. Remember, in our units, the speed of light is 1. Since we started out with r =
0 at t = 0, if the particle is traveling faster than light, the probability amplitude for r > t will be non-zero.

To calculate the integral, we extend p to complex values, and let p → z = x + iy. We’ll take the x axis as part of
a contour C, and close the contour with a large semicircular arc above or below the x-axis. Our integrand is not
however an analytic function of p, because the function ωp = has branch points at p = ±iµ, and thus also a
branch line connecting these two points. I choose to write the branch cuts as extending from +iµ up along the
positive imaginary axis, and from −iµ down along the negative imaginary axis. If we distort the semicircular contour
C to avoid the branch cuts, the integrand is an analytic function within the region bounded by the distorted contour,
as shown below.

Figure 1.2: Contour for evaluating the integral

Since the integrand is analytic within C, the integral along C, say counter-clockwise, gives zero. The original
integral along the x axis then equals the rest of the integral in the opposite sense, going clockwise along the large
arcs, going down on the left side of the branch cut, and up on the right side. Along the upper branch cut, p is
parametrized by iy. The value of ωp is discontinuous across this branch cut; its value on either side of the branch
cut is 12

Along the large arcs, p is parametrized by Reiθ = R cos θ + iR sin θ, where π ≥ θ ≥ π/2 on the left-hand arc, and
π/2 ≥ θ ≥ 0 on the right-hand arc. The integrand involves eipr−iωpt which is bounded, since r > t, by e−Rr sin θ.
Consequently in the limit R → ∞, the large arcs contribute nothing to the integral. The small arc likewise contributes
nothing in the limit as the small circle’s radius goes to zero. Then

Though the ωp part of the exponential is damped on the left side of the cut, the exponential increases on the right
side. However since r > t, the strictly damped part of the exponential, −ry, dominates over the increasing part
. Changing the limits in the first term gives

This is bad news, boys and girls, because this integrand is a product of positive terms. Therefore the integral is not
zero, and our particle always has uncertain nonzero probability amplitude for traveling faster than the speed of
light. So the particle can move faster than light and thus backwards in time, with all the associated paradoxes. I
hope you understand the titanic meaning of that statement.

Things are not so bad, however, as you would think. The particle doesn’t have much of a probability of
traveling faster than light. It’s impossible to do the integral, which means the answer is a Bessel function.13 But it’s
rather trivial to bound the integral by keeping only the increasing exponential part of the sinh, and then replacing
by y. This will give us an overestimate. We have then

The chance that the particle is found outside of the forward light cone falls off exponentially as you get farther from
the light cone. This makes it extremely unlikely that, for example, I could go back in time and convince my mother
to have an abortion. But if it is at all possible, it is still unacceptable if you’re a purist. If you’re a purist, the Xi
operator we have defined is absolutely rotten, no good, and to be rejected. If instead you’re a slob, it’s not so bad,
because the amplitude of finding the particle outside of the forward light cone is rather small. It’s exponentially
damped, and if we go a few factors of 1/µ, a few of the particle’s Compton wavelengths away from the light cone,
the amplitude comes down quite a bit.

What we have discovered is that we cannot get a precise determination of where the particle is. But if we’re only
concerned with finding the particle to within a few of its own Compton wavelengths, in practice things are not so
bad. In principle, the inability to localize a single particle is a disaster. How does nature get out of this disaster? Is
there a physical basis for an escape? Yes, there is.

Suppose I attempt to localize a particle in the traditional gedanken experiment methods of Niels Bohr. (In fact,
this argument is due to Niels Bohr.14) I build an impermeable box with moveable sides. I put the particle inside it. I
turn the crank, like the Spanish Inquisition, and the sides of the box squeeze down. It appears that I can localize
the particle as sharply as I want. What goes wrong? What could relativity possibly have to do with this?
 Figure 1.3: Particle in a box with a movable wall

The point is this. If I try to localize the particle within a space of dimensions L on the order of its own Compton
wavelength, L ∼ O(1/µ), then not relativity, but our old reliable friend the Uncertainty Principle comes into play and
tells us

If the dispersion in p is on the order of µ, then so must p itself be at least the order of µ. Then we have enough
energy in the box to produce pairs.

Figure 1.4: Particle squeezed in the box

Like the worm Ouroboros,15 this section ends where it began, with pair production. If we squeeze the particle
down more and more, we must have more and more uncertainty in momentum. If we have a large spread in
momentum, there must be a probability for having a large energy inside the box. If we have a large energy inside
the box, we know there’s something inside the box, but we don’t know it’s a single particle. It could be three
particles, or five, or seven. The moral of the story is that we cannot satisfactorily localize the particle in a single-
particle theory.

So we can localize something, but what we’re localizing is not a single particle. Because of the phenomena of
pair production, not only is momentum complementary to position, but particle number is complementary to
position. If we make a very precise measurement of position, we’ll have a very big spread in momentum and
therefore, because pair production takes place, we do not know how many particles we have. Relativistic causality
is inconsistent with a single-particle quantum theory. The real world evades the conflict through pair production.
That’s the physical reason for the mathematics we’ve just gone through. This leads to our next topic, a discussion
of many free particles.

1 [Eds.]
P. A. M. Dirac, “The Quantum Theory of the Electron”, Proc. Roy. Soc. Lond. A 117 (1928) 610–624; “The
Quantum Theory of the Electron. Part II”, Proc. Roy. Soc. Lond. A 118 (1928) 351–361.
2 [Eds.] See H. Bethe and E. Salpeter, Quantum Mechanics of One- and Two-Electron Atoms, Plenum Publishing,
1977, and references therein; republished by Dover Publications, 2008; and M. E. Rose, Relativistic Electron
Theory, Wiley, 1961, pp. 193–196. Rose explicitly shows the suppression of positron density near the hydrogen
nucleus as |p| → 0, and ascribes this suppression to Coulomb repulsion acting on positrons.
3[Eds.] The official text for the course was the two-volume set Relativistic Quantum Mechanics and Relativistic
Quantum Fields (hereafter, RQM and Fields, respectively) by James D. Bjorken and Sidney D. Drell, McGraw-Hill,
1964 and 1965, respectively. Coleman said this (in 1975) about the books: “I will try to keep my notational
conventions close to those of Bjorken and Drell. It’s the best available. People like it by an objective test: it is the
book most frequently stolen from the Physics Research Library.”
4 [Eds.] Strictly speaking, the connected Lorentz group is the orthochronous Lorentz group, SO+(3, 1), the
subgroup of SO(3, 1) preserving the sign of the zeroth component of a 4-vector.
5 [Eds.]
Most authors write for the d’Alembertian, rather than 2. Coleman used 2, so that’s what is
used here.
6 [Eds.]
Also denoted H(x), and frequently called the Heaviside step function, after Oliver Heaviside (1850–1925)
who used it extensively. See H. Jeffreys, Operational Methods in Mathematical Physics, Cambridge Tracts in
Mathematics and Mathematical Physics No. 23, Cambridge U. P., 1927, p. 10.
7[Eds.] Because p = ħk, and in our units ħ = 1, we could equally well use kets |kñ; p and k both stand for
momentum.

8 [Eds.] The equality follows from the identity δ(f(x)) = where {ai} are the zeroes of f(x). Then
The θ(p0) factor kills the second delta function, and integrating over p0 gives just the factor (2ωp)−1, times the
remaining d3p. Similarly, one can show d3xδ(|x|2 − R 2) = R sin θ dθ dϕ.
9 [Eds.] See §22 “Schrödinger’s Representation”, in P. A. M. Dirac, The Principles of Quantum Mechanics, 4th ed.
revised, Oxford U. P., 1967.
10 By translational invariance and superposition, we could easily get the evolution of any configuration from this
calculation.
11 [Eds.] Consider áx|Xi|pñ. If we let the operator operate to the left, we get

but operating to the right,

so the quantity áx|pñ satisfies the differential equation i∂/∂pi áx|pñ = xi áx|pñ. Then áx|pñ = Ceip•x, where C may
depend on x (but not p). By considering áx|Pi|pñ, you can show C is a constant, and we set C = 1/(2π)3/2 for
convenience. See Dirac, op. cit., §23, “The momentum representation”.
12 [Eds.]
For the details, see the example on pp. 71–73 of Mathematics for Physicists, Phillipe Dennery and André
Krzywicki, Harper and Row, 1967, republished by Dover Publications, 1996, or Example 2 in Chap. 7, pp.
202–205 of Complex Variables and Applications, Ruel V. Churchill and James Ward Brown, McGraw-Hill, 1974.
13 [Eds.] Coleman is joking. Mathematica fails to find a closed form for this integral, so it isn’t really a Bessel
function.
14[Eds.] N. Bohr and L. Rosenfeld, “Field and Charge Measurements in Quantum Electro Phys. Rev. 78 (1950)
794–798.
15 [Eds.] The Worm Ouroboros is a fantasy novel by E. R. Eddison, published in 1922; J. R. R. Tolkien was an
admirer. The ouroboros (Greek , “tail” + , “devouring”) is the image of a snake or dragon eating its own
tail. Originally ancient Egyptian, it entered western tradition via Greece, and came to be associated with European
alchemy during the Middle Ages. It is often used as a symbol of the eternal cycle of death and rebirth. See A
Dictionary of Symbols, J. E. Cirlot, Dover Publications, 2002, pp. 15, 48, 87, 246–247. The German chemist
August Kekulé reported that a dream of the ouroboros led him to propose the structure of the benzene ring: O.
Theodor Benfey, trans., “August Kekulé and the Birth of the Structural Theory of Chemistry in 1858”, J. Chem. Ed.
35 (1958) 21–23.

2
The simplest many-particle theory

In the last section we fiddled around with the theory of a single, relativistic spinless particle. We found some things
that will be useful to us in the remainder of this course, like the Lorentz transformation properties of the particle,
and some things that served only to delineate dead ends, e.g., we could not define a satisfactory Xi operator.
When we tried to localize a particle, we found that the particle moved faster than the speed of light. At the end of
the lecture I pointed out that the problem of localizing a particle could be approached from another viewpoint.
Instead of staying within the theory of a single particle, we could imagine an idealized example of the real world in
which pair creation occurs. We discovered that we couldn’t localize a particle in a box. If the box was too small, it
wasn’t full of a single particle, it was full of pairs. This motivates us to investigate a slightly more complicated
system, a system consisting of an arbitrary number of free, relativistic spinless particles. The investigation of this
system will occupy this whole section. The problem of localization should be in the back of our minds but I won’t
say anything about it.

2.1First steps in describing a many-particle state

The general subject is called Fock space.1 That is the name for the Hilbert space, the space of states that
describes the system we’re going to talk about. We’ll discover that when we first write down Fock space it will be
extremely ugly and awkward. We will have to do a lot of work to find an efficient bookkeeping algorithm to enable
us to manipulate Fock space without going crazy. The bookkeeping will be managed though the algebra of
objects called annihilation and creation operators, which may be familiar to you from an earlier course in quantum
mechanics.

The devices I’m going to introduce here—although we will use them exclusively for the purposes of relativistic
quantum mechanics—are not exclusively applied to that. There are frequently systems in many-body theory and in
statistical mechanics where the number of particles is not fixed. In statistical physics we wish frequently to
consider the so-called grand canonical ensemble, where we average over states with different numbers of
particles in them, the number fluctuating around a value determined by the chemical potential. In solid state
physics, there’s typically a lot of electrons in a solid but we are usually interested only in the electrons that have
stuck their heads above the Fermi sea, the conduction electrons. The number of these electrons can change as
electrons drop in and out of the Fermi sea. So the methods are of wider applicability. In order to keep that clear, I
will use our non-relativistic normalization of states and non-relativistic notation and just switch to relativity at the
end when we want to talk about Lorentz transformation properties.

Let me remind you of the Hilbert space of non-relativistic one-particle states we had before: the momentum
kets |pñ labeled by a basis vector p, and normalized with a delta function,

which is standard for plane waves. These are simultaneous eigenstates of the Hamiltonian, H, with eigenvalues
(but that’s not going to be relevant here), and of the momentum operator P, with eigenvalues p;

They also of course have well defined Lorentz transformation properties, which we talked about last time, but I’m
not going to focus on that for the moment. In the last section this was a complete set of basis vectors for our
Hilbert space; a general state was a linear combination of these states. But now we are after a bigger Hilbert
space, so these will be just a subset of the basis vectors. I will call these “one-particle basis vectors”. We are
considering a situation where we look at the world and maybe we find one particle in it, but maybe we find two or
three or four, and maybe we find some linear combination of these situations. Therefore we need more basis
vectors. In particular we need two-particle basis vectors. I’ll write down the construction for them, and then I’ll just
write down an “et cetera” for the remainder (the three-particle states, the four-particle states, . . . )

A two-particle state describes two independent particles, and will be labeled by the momenta of the two
particles, which I will call p1 and p2, which can be any two 3-vectors. (Don’t confuse these subscripts with vector
indices, the index labels the particle.) We will assume that our spinless particles are identical bosons, and
therefore to incorporate Bose2 statistics we label the state |p1 p2ñ which in fact is the same state designated by
|p2, p1ñ. It doesn’t matter whether the first particle has one momentum and the second the other, or vice versa. We
will normalize the states again with traditional delta function normalization:

The states are orthogonal, unless the two momenta involved are equal, either for one permutation or the other.
We have to include both those terms or else we’ll have a contradiction with the normalization equation for a single
particle. These states are eigenstates of the Hamiltonian and their energies are of course the sum of the energies
associated with the two individual particles, and they are eigenstates of the momentum operator, and their
momentum is the sum of the two momenta of the two individual particles:

The extension to three particles is just “et cetera”. While “et cetera” is of course the end of the story, we have not
really started with the beginning of the story. There is one thing we have left out. It is possible that we look upon
the world and we find a probability for there being no particles. And therefore we need to add at least one basis
vector to this infinite string, to wit, a no-particle basis vector, a single state, to account for a possibility that there
are no particles around at all. We will denote this state |0ñ. It is called the vacuum state. We will assume that the
vacuum state is unique. This state |0ñ is of course an eigenstate of the energy with eigenvalue zero, and
simultaneously an eigenstate of momentum with eigenvalue zero:
The vacuum state is Lorentz invariant, U(Λ) |0ñ = |0ñ. All observers agree that the state with no particles is the
state with no particles. We will normalize it to 1,

This normalization is conventional for a discrete eigenstate of the Hamiltonian, one which is not part of the
continuum. Please do not confuse the vacuum state with the zero vector in Hilbert space, which is not a state at
all, having probability zero associated with it, nor with the state of a single particle with 3-momentum p equal to
zero, |0ñ. That ket is denoted with the vector 0.

We now have a complete catalog of basis vectors. A general state |Ψñ in Fock space will be some linear
combination of these basis vectors:

This is some number, some probability amplitude times the no-particle state, the vacuum, plus the integral of some
function ψ 1(p) times the ket |pñ plus a one over 2! inserted by convention—I’ll explain the reason for that
convention—times the integral over two momenta of a function of both momenta times the two-particle ket, with
dots indicating the three-particle, the four-particle et cetera states, going on forever.

I should explain the factor of . Since the state |p1, p2ñ is the same as the state |p2, p1ñ, without any loss of
generality we can choose

That is to say we can choose a Bose wave function for two bosons to be symmetric in the two arguments. I then
insert the to take account of the fact that I am counting the same state with the same coefficient twice when
I integrate once over p1 and once over p2 in one order, and then in the other order. Likewise, successive terms for
the three-particle or four-particle states will have corresponding factors of , et cetera.

The squared norm |Ψ|2 of the state |Ψñ is

The state |Ψñ exists and is normalizable as always only if |Ψ|2 < ∞, so we can multiply it by a constant and make
its norm 1, and speak about probabilities in a sensible way.

Well, in a sense we have solved our problem. We have described the space of states we want to talk about.
But we have described it in a singularly awkward and ugly way. To describe a state, we need an infinite string of
wave functions: the zero particle wave function, a function of a single variable, a function of two variables, a
function of three variables, a function of four variables et cetera, ad nauseam. Handling a system of this kind by
conventional Schrödinger equation techniques, describing the dynamics by an interaction operator made up out of
q’s and d/dp’s and some sort of incredible integral-differential operator that mixes up functions of two or three or
four or any number of variables with other such functions is a quick route to insanity. We need to find some
simpler way of describing the system.

2.2Occupation number representation

In order to minimize problems that arise when one is playing with delta functions and things like that I will brutally
mutilate the physics by putting the system in a periodic box. So we will have only discrete values of the momentum
to sum over instead of continuous values to integrate over. Of course this is a dreadful thing to do: It destroys
Lorentz invariance; in fact it even destroys rotational invariance. But it’s just a pedagogic device. In a little while I’ll
let the walls of the box go to infinity and we’ll be back in the continuum case.

With the system in a periodic box, we imagine the momenta restricted to a discrete set of values which are
labeled by

The box is a cube of length L; the numbers nx , ny and nz are integers. Instead of filling out 3-space, the momenta
span a cubic lattice. Since we have discrete states we can use ordinary normalization rather than delta function
normalization. For example, in the one-particle states

the Kronecker delta equaling 1 if p = p′, and zero otherwise. Integrations in the continuum case are replaced by
sums over the whole lattice of the allowed momenta. These will be discrete, infinite sums.

In this box we can label our basis states in a somewhat different way than we have labeled them up to now. In
our previous analysis we haven’t exploited Bose statistics much; it’s been rather ad hoc. We tell how many
particles there are, we imagine the particles are distinguishable, we give the momentum of the first particle, the
momentum of the second, the third, et cetera; and then we say as an afterthought that it’s the same as giving the
same set of momenta in a different permutation. Now as you all know from elementary quantum statistical
mechanics where you count states in a box, there is a much simpler way of describing the basis states. We can
describe our basis states by saying how many particles there are with this momentum, how many particles are
there with that momentum, how many with some other momentum. We can describe our states by giving them
occupation numbers N(p), a function from the lattice of allowed one-particle momenta into the integers which is
simply the number of particles with momentum p. Obviously this is exactly equivalent; it describes not only the
same Hilbert space but the same set of basis vectors as we’ve described before, providing of course we have the
condition that

No fair writing down a state where there is one particle with each momentum! That’s not in our counting, and
anyway it would be a state of infinite energy.

This is not a change of basis in the normal sense, but just a relabeling of the same basis in a different way.
We can label our states by an infinite string of integers {N(p)}, a sort of super-matrix3 which you can imagine as
this three-dimensional lattice whose axes are px , py , and pz , with integer numbers N(px , py , pz ) sitting on every
lattice point. Of course, most of the numbers will be zero. I’ll write such a labeling this way, with a curly bracket,
just to remind you that this is not a state labeled by a single number N(p) and a single vector p, but by this matrix
of integers.

The advantage of the occupation number labeling is of course that the Bose statistics is exploited, taken care
of automatically. When I say there is one particle with this momentum and one particle with that momentum, I
have described the state; I don’t have to say which is the first particle and which is the second. In terms of this
labeling the Hamiltonian has a very simple form:

The energy of the many-particle state is the sum of the energies of the individual particles. The momentum
likewise has a very simple form:

Staring at the expression (2.13) for the energy, we notice something that wasn’t obvious in the other way of writing
things: First, the energy is a sum of independent terms, one for each value of p, and second, within each
independent term we have a sequence of equally spaced energies separated by ωp. We can have zero times ωp,
1 times ωp, 2 times ωp, and so on. Such a structure of energy spacings is of course familiar to us: It’s what occurs
in the harmonic oscillator. In fact this is exactly like the summation we would get if we had an infinite assembly of
uncoupled harmonic oscillators, each with frequency ωp, except that the zero-point energy, the ωp, is missing.
But other than that, this looks, both in the numbers of states and their energies, exactly like an infinite assembly of
uncoupled harmonic oscillators. The two systems are completely different. In our many-particle theory, N(p) tells
us how many particles are present with momentum p. In a system of harmonic oscillators, N(p) gives the
excitation level of the oscillator labeled by p. Still, let us pursue this clue. And in order that we will all know the
same things about the harmonic oscillator, I will now digress into a brief review on this topic. Most people will have
seen this material in any previous quantum mechanics course. I apologize, but theoretical physics is defined as a
sequence of courses, each of which discusses the harmonic oscillator. 4

2.3Operator formalism and the harmonic oscillator

Consider a single harmonic oscillator. The momentum p and the position q are now not numbers, they are
quantum operators obeying the much-beloved commutation relations

The Hamiltonian is 5

I subtract 1 here to adjust the zero of my good old-fashioned harmonic oscillator so that the ground state has
energy zero. The Hamiltonian will then look exactly like one of the terms in the sum (2.13), not just qualitatively like
it.

Now the famous way of solving this system is to introduce the operator a and its adjoint a†,

It is easy to compute the commutator [a, a†]:

We get a contribution only from the cross-terms, both of which give equal contributions and cancel out the that
comes from squaring the . It is also easy to rewrite the Hamiltonian in terms of a and a†, since

Then

As promised, this expression for the harmonic oscillator Hamiltonian looks exactly like one of the terms in
(2.13); we need only confirm that N = a†a is a number operator. From these two equations, (2.18) and (2.20), plus
one additional assumption, one can reconstruct the entire state structure of the harmonic oscillator. I will assume
the ground state is unique. Without this assumption I would not know for example that I was dealing with a
spinless harmonic oscillator; it might be a particle of spin 17, where the spin never enters the Hamiltonian. Then I
would get twice 17 + 1 or 35 duplicates of a harmonic oscillator, corresponding to the various values of the z
components of the spin. The assumption of a unique ground state will take care of the possibility of other
dynamical variables being around that would give us a multiply degenerate ground state. Let’s now determine the
system. I presume you’ve all seen it before.

Compute the commutators of H with a and a†;

Let us consider some energy eigenstate of this system. I will assume it’s labeled by its energy, and denote it in the
following way, |Eñ, where of course

Now consider H acting on the state a† |Eñ. By the equation above,

Thus, given a state |Eñ of energy E, I can obtain a state of energy E + ω by applying a†. I can draw a
spectroscopic diagram, Figure 2.1.

Figure 2.1: Traveling up and down the ladder of energy states

And of course I can build this ladder up forever by successive applications of a†. By the same reasoning applied to
a, I obtain a similar equation:

By applying a I can go down the ladder. For this reason a† and a are called “raising” and “lowering” operators
because they raise and lower the energy. Can we go up and down forever?

I don’t know yet about going up, but about going down I can say something. The Hamiltonian is the product of
an operator and its adjoint, and therefore it always has nonnegative expectation values and non-negative
eigenvalues. So the energy must be bounded below. There must be a place where I can no longer continue going
down. Let me write the lowest energy eigenstate, the ground state, as |E0ñ, which by assumption is unique. Now
there’s no fighting this equation (2.25): Applying a to |E0ñ gives me a state which is an eigenstate with energy E0 −
ω;

On the other hand by assumption there is no eigenstate with energy lower than E0. The only way these apparent
contradictions can be reconciled is if

because a|E0ñ = 0 satisfies the equation (2.26) for any value of E0. This of course determines the energy of the
ground state, because H = a†a, and therefore

Therefore the ground state, by assumption unique, is a state of energy 0, and I will relabel the ground state:

meaning, the ground state is the state of zero energy. All the other states of the system have energies which are
integer multiples of ω because the ladder has integer spacing. We can label these |nñ, i.e.,

We obtain the states |nñ from systematic application of a† on the ground state:

Equation (2.30) follows from (2.31) and commuting H with (a†)n, confirming N = a†a as a number operator. Let’s
say

and obtain C n by normalizing the states with the usual convention,

If I compute the square of the norm of the state a†|nñ = C n |n + 1ñ, the inner product of this ket with the
corresponding bra on the right-hand side, I get

That determines C n up to a phase. I have not yet made any statement that determines the relative phases of the
various energy eigenstates, and I am free to choose the phase so that {C n} are real:

We then have the fundamental expression for the action of a† on an arbitrary state |nñ,

By similar reasoning or by direct application of the definition of the adjoint, we determine

I have snuck something over on you. I have talked as if the ladder of states, built out of successive
applications of a† on the ground state, is the entire space of energy states. You know that is true for the harmonic
oscillator, but I haven’t proved it using just the algebra of the a’s and a†’s. So let’s demonstrate that.

If we have an operator A which commutes with both p and q, then A must be a multiple of the identity:

where λ is some constant.6 Say there is a state |ψñ which has a component not on the ladder. Presumably there is
a projection operator, P which projects |ψñ onto the ladder. Since a† and a keep a state on the ladder, it must be
that

But as q and p can be written as linear combinations of a and a†, we can say

The projection operator is proportional to the identity, so there are no parts of any ket |ψñ not on the ladder; there
are no other states except those already found.

Two or three or four decoupled harmonic oscillators can be handled in exactly the same way. We simply have
two or three or four sets of raising and lowering operators. The Hamiltonian is the sum of expressions of this form
over the various sets. Conversely, if we have a system with the structure of a harmonic oscillator, with equally
spaced energy eigenstates, we can define operators a and a† for each set, and then regain the algebraic structure
and the expression for the Hamiltonian and complete the system in that way.

This completes the discussion of the harmonic oscillator. Its entire structure follows from these algebraic
statements (2.18)–(2.25) and the mild assumption of minimality, that there is only one ground state.

2.4The operator formalism applied to Fock space

Now let us turn to the particular system we have: An infinite assembly of harmonic oscillator-like objects, one for
every point in our momentum space lattice. The analogs, the mathematical equivalents to the harmonic oscillator
excitation numbers are the occupation numbers. Therefore we can define raising and lowering operators on the
system, a†P and ap, one for every lattice point, that is to say, for every value of p.

The lowering operators associated with different oscillators have nothing to do with each other:

The raising operators associated with different oscillators have nothing to do with each other:

The raising and lowering operators for two oscillators have the conventional commutators

equalling 1 if they describe the same oscillator, and commuting otherwise.

The Hamiltonian is the sum of the Hamiltonians for each of the individual oscillators,

The oscillators are labeled by the index p, a 3-vector on the lattice, and each has energy ωp. We haven’t talked
about the momentum operator in our discussion of a single oscillator, but of course it will be given by an
expression precisely similar to the Hamiltonian. The factor multiplying ωp, a†Pap, has eigenvalues N(p), so the
momentum operator is

This set of equations defines Fock space in the same way as the corresponding set of equations defines the single
oscillator. The only change is a change in nomenclature.
 As you’ll recall, the thing that corresponded to the excitation level of an oscillator was the number of particles
bearing momentum p. We will no longer call a†P and ap “raising” and “lowering” operators, respectively. We will
call them creation and annihilation operators because applying a†P raises an equivalent oscillator, that is to
say, adds one particle of momentum p; applying ap lowers an equivalent oscillator, i.e., removes one particle of
momentum p. Another term will be changed from that of the oscillator problem. We normally do not call the
simultaneous ground state of all the oscillators “the ground state”; we call it as I have told you, the vacuum state.
The vacuum state is defined by the equation that any annihilation operator applied to it gives zero:

The advantage of these algebraic equations over the original definition of Fock space is great. You see here they
take only a few lines. The original definition filled a page or so. As shown by the argument with the oscillators, they
give you the complete structure of the space: they tell you what the states are, they tell you what their
normalizations are, they tell you the energy and momentum of any desired state. So we have made progress, by
reducing many equations to a few.

I am now going to blow up the box, letting L → ∞, and attempt to go to the continuum limit. I will not attempt to
go to the continuum limit in the occupation number or equivalent oscillator formalism. That is certainly possible but
it involves refined mathematical concepts. Instead of a direct product of individual oscillators spaces, we would get
a sort of integral-direct product, a horrible mess. The point is that we can generalize these algebraic equations
directly. These contain the entire content of the system. We can generalize them simply by taking a step backward
from what we did to get to the box in the first place: we replaced all Dirac delta functions by Kronecker deltas, and
all integrals by sums. If we undo this, replacing sums with integrals and Kronecker deltas with Dirac deltas, we will
get a system that gives us continuum Fock space. I’ll check that it works. I won’t check every step because most
of it is pretty obvious, but I’ll check a few examples for you.

I’m going to define the system purely algebraically just as I defined the oscillator and Fock space for a box
purely algebraically. There are a fundamental set of operators, a†P and ap for any value of p now, not just integer
values defined on the lattice, and they obey these equations;

The Hamiltonian is

and the total momentum operator is

where

is the number operator. That’s it. These statements (2.47)–(2.50), together with the technical assumption that the
ground state of the system is unique, will define continuum Fock space in the same way the precisely parallel
statements defined Fock space for particles in a box. Let’s check that for a few simple states.

First, the ground state of the system, the vacuum, which is assumed to be unique, is defined by

for all p. Directly from the expressions for the energy and the momentum, this state is an eigenstate of the energy
with eigenvalue zero, and of the momentum, with eigenvalue zero. Of course the algebraic structure doesn’t tell us
how we normalize the vacuum. That’s a matter of convention, and we will choose that convention to be the same
as before,

the vacuum state has norm 1. To make one-particle states we apply creation operators to the vacuum; that is a
one-particle state of momentum p. If all we were working from were the previous algebraic equations for the
harmonic oscillators, according to (2.36) the one-particle state of momentum p would be obtained like this:
Let’s assume this is right for the continuum Fock space, and compute the norm of this state:

We have our fundamental commutation relations and so we will commute;

The first term is δ(3)(p − p′) times the norm of the vacuum which is one. The second term is zero because ap′
acting on the vacuum is zero; every annihilation operator acting on the vacuum is zero. Thus the state has the
right norm. This looks good.

What about the energy of the single-particle states? Well, it’s the same story:

We know the commutations required to compute this; all a†P commute with each other, all ap commute with each
other, the commutation of any a†P and any ap is a delta function:

and H on the vacuum is zero. Then

which also looks good. Et cetera for the momentum, et cetera for the two-particle states, the three-particle states
and so on. Here is an example of a two-particle state, just to write down its definition,

This state |p1, p2ñ is of course automatically equal to the state |p2, p1ñ because the two creation operators
commute. It doesn’t matter what order you put them in. The Bose statistics are taken account of automatically. I
leave it to you to go through the necessary commutators to show that the state has the same normalization as the
one we wrote down before, and that it has the right energy and the right momentum. The operations are exactly
parallel to the operations I’ve done explicitly for the single particle state.

To summarize where we have gone: The algebraic equations plus the technical assumption that there exists
a unique vacuum state, the ground state of the system, completely specify everything about Fock space we
initially wrote down formally. This is obviously a great advantage; it’s much simpler to manipulate these
annihilation and creation operators than it would be to manipulate a number plus a function of one variable plus a
function of two variables plus a function of three variables et cetera.

Now there are two further points I want to make before we leave the topic of Fock space and go on to our next
topic. One is a point for mathematical purists. Those of you who are not mathematical purists may snooze while I
make this point. In the technical sense of Hilbert space theories these a†P and ap we have introduced are not
operators because when applied to an arbitrary state they can give you a non-normalizable result. For instance,
a†P |0ñ is a plane wave |pñ, and a plane wave is not normalizable (áp|pñ = δ(3)(0) = ∞). Occasionally while browsing
through Physical Review, or more likely through Communications in Mathematical Physics, you may come across
people not talking about these things as operators—they are purists—but as “operator valued distributions.” A
“distribution”, to a mathematician, means something like a delta function, which is not itself a function, but it
becomes a function when it is smeared, integrated over in a product with some nice smoothing function. These
a†P and ap are operator-valued distributions labeled by indices p, and the things that are really sensible operators
are smeared combinations like ∫ d3p f(p)a†P where f(p) is some nice smooth function. That creates a particle in a
normalizable state, a wave packet state, with f(p) the momentum space wave function describing its shape. And
that’s the thing that people who are careful about their mathematics like to talk about.7 I am not a person who is
careful about his mathematics; I won’t use that language. But in case you run across it in some other course, you
should know that there are people who use this language and this is the reason why they use it. They prefer to talk
about the smeared combinations rather than the ap’s themselves.

Secondly—and this is not for purists, this is for real—since we’re back in infinite space, we can sensibly talk
about Lorentz transformations again. To complete this section, I should specify how Lorentz transformations are
defined in terms of the creation and annihilation operators. The essential trick is to observe that just as we defined
the relativistically normalized states, so we can define relativistically normalized creation operators that when
applied to the vacuum will create states with the correct relativistic normalization.

I will call these operators α(p) and α †(p). The momentum index p is now a 4-vector, but the fourth component
is constrained just as before: p0 = ωp. These creation operators are defined by

Operating on the vacuum, this operator makes the same state as a†P does. It creates the same particle, but with
relativistic normalization and not just the δ(3)(p − p′) normalization (see (1.57)):

Of course there is also a relativistic annihilation operator which we can write down just by taking the adjoint,

These operators, as you can convince yourself, transform simply under Lorentz transformations. We can
determine the Lorentz and translation properties of these operators α †(p) and α(p) from the assumed
transformations of the kets. First, consider the vacuum. It’s obvious that the vacuum is Lorentz invariant, since it is
the unique state in our whole Fock space of zero energy and zero momentum, and that’s a Lorentz invariant
statement. So U(Λ) acting on the vacuum must give us the vacuum, since U is unitary and does not change the
norm:

Then, for a single particle state |pñ, assume that

and for a multi-particle state,

From (2.64), we determine how α †(p) behaves under a Lorentz transformation:

So

That is to say, α †(Λp) is the creation operator of the transformed 4-momentum. And of course taking the adjoint
equation

Just to check that these are right, let’s compute the transformation acting on a multi-particle state, |p, p1, p2 , . . . ,
pnñ. We have

which is the desired result. The same argument would have worked for any pi, or for that matter any set of the pi’s,
since the kets are symmetric in the pi’s. Here’s another way to think about the transformation of a multi-particle
state:
So this system, defined by the operators α(p) and α(p)†, admits unitary Lorentz transformations, as of course it
should, because it’s the same system we were talking about before. The action of these Lorentz transformations
can be defined if we wish by these equations (2.63)–(2.67). That enables us to tell how every state Lorentz
transforms.

Likewise the translation properties of the creation and annihilation operators are easily found from the
transformations of the kets. The unitary operators of translations are U(a) = eiP·a. Because

and

we get

and

A derivation analogous to the Lorentz transformation leads to the translational properties of α †(p) and α(p),

In the next section I return to the question which inspired us. (Actually, what inspired us is the fact that
quantum electrodynamics predicts the right anomalous magnetic moment of the electron, but we won’t get to that
until the second half of this course!8) What inspired us in this elegant but historically false line of reasoning, to
consider an infinite, great big Hilbert space in the first place was the problem of localization. In the next section I
will talk about localization from another tack, not about localizing particles, but instead about localizing
observations. Incidentally, note that these operators α(p) and α †(p) depend on time as well as space. We are
working in the Heisenberg representation, in which the states are constant but the operators depend on time,
rather than the Schrödinger representation in which the operators are time-independent but the states evolve in
time.

1 [Eds.] V. Fock, “Konfigurationsraum und zweite Quantelung” (Configuration space and second quantization),
Zeits. f. Phys. 75 (1932) 622–627.
2I was recently informed by a colleague from subcontinental India that this name should be pronounced “Bōsh”,
and I’ll try to train myself to pronounce it correctly. But bosons are still “bōsäns”.
3 [Eds.] In other words an infinite, rank 3 array, with integer matrix elements 1, 2 , . . .
4 [Eds.] A variation of this remark attributed to Coleman is: “The career of a young theoretical physicist consists of
treating the harmonic oscillator at ever-increasing levels of abstraction.”
5 [Eds.] This form of the Hamiltonian is the result of a canonical transformation of the usual harmonic oscillator
Hamiltonian,

with [Q, P] = i, namely

This canonical transformation preserves the commutator, [q, p] = i and leads to the form

See Goldstein et al. CM, pp. 377–381.

6 [Eds.] If A commutes with q, then it is either a constant or a function of q. But if it is a function of q, it cannot
commute with p. So it must be a constant, i.e., a multiple of the identity. This is an application of Schur’s lemma.
See Thomas F. Jordan, Linear Operators for Quantum Mechanics, Dover Publications, 2006, pp. 69–70.
7 [Eds.]For an accessible, slim and inexpensive book about distributions, also known as “generalized functions”,
see M. J. Lighthill, An Introduction to Fourier Analysis and Generalised Functions, Cambridge U. P., 1958.
8 [Eds.] §34.3, pp. 743–749.

3
Constructing a scalar quantum field

3.1Ensuring relativistic causality

In ordinary, non-relativistic quantum mechanics, every Hermitian operator is an observable. Given anything
measurable, any physicist, if she is only crafty enough, can manage to think up some apparatus that measures it.
She measures the position x with a bubble chamber, she measures the momentum p with a bending magnet, she
may have to be a real genius to measure a symmetrized product of p4 times x8, but in principle there’s nothing to
keep her from measuring that. It’s only a matter of skill. I cannot measure the length of my own foot, but that’s only
due to my lack of skill. The words “Hermitian operator” and “observable” are synonymous, and we do not bother to
introduce distinctions about what an idealized observer can measure and a Hermitian operator.

In particular, every observer can measure every operator and therefore every observer can measure non-
commuting operators. One can measure σx and also measure σy . Now you know the measurement of non-
commuting observables does not commute. If I have an electron and I measure its σx , and I turn my back for a
moment, and Carlo Rubbia,1 having just come in on a jet plane, sneaks into the room—he does a lot of
experiments—and measures something that commutes with σx , like px , and then sneaks out again before I can
turn around, I won’t notice any difference when I measure σx a second time. If on the other hand when Carlo
comes in he measures σy , then I will notice a big difference. My system will no longer be in the eigenstate of σx in
which I had carefully prepared it; it will now be in an eigenstate of σy , and I will notice the change. I will know that
someone has made a measurement even if I keep my eyes closed.

If we say every observer can measure every observable, even in the most idealized sense of “observer” and
“observable”, then we encounter problems in a relativistic theory. I after all have a finite spatial extent, and my
travels, far and wide as they are, occupy only a finite spatial extent. And, alas, the human condition states that I
also have only a finite temporal extent. There is some region of space and time within which all the experiments I
can do are isolated. The earth goes around the sun, so my spatial extent is perhaps the diameter of the solar
system in width, and, if I give up smoking soon, maybe 75 years in length, but. . . that’s it. Now let us imagine
another observer similarly localized, say somewhere in the Andromeda galaxy, and he has a similar life
expectancy and spatial extent. If both he and I can measure all observables, then he can measure an observable
that doesn’t commute with an observable I can measure. Therefore, instead of Carlo Rubbia sneaking into the
room after I’ve done my experiment, he can just stay in Andromeda and do his experiment on a non-commuting
observable. Just as if Carlo had come sneaking into the room, I would notice that my results had changed, and
thus would deduce that he has made a measurement. That’s impossible because the only way to get information
from him there in Andromeda to me here on earth between the time of my two measurements is for information to
travel faster than the speed of light. Therefore it cannot be that I can measure everything, and it cannot be that he
can measure everything. There must be some things that I can measure and some things that he can measure,
and it must be that everything that I can measure commutes with everything that he can measure. Otherwise he
could send information faster than the speed of light, namely, the information that he has measured an observable
that does not commute with an observable that I have just measured. The reasoning is abstract, but I hope simple
and clear.

Even if we are going to be generous in our idealization, we have to realize that somehow in any sensible
relativistic quantum mechanical theory, there must be some things that can be measured by people who are
constrained to live in a certain spacetime region and some things they cannot measure. Within every region of
space and time, out of the whole set of Hermitian operators, there must be only some of them that those people
can measure. If this were not so we would run into contradictions between the most general principles of quantum
mechanics—the interpretation rules that tell us how Hermitian operators are connected with observations—and
the principle of Einstein causality, that information cannot travel faster than the speed of light. I have said a lot of
words. Let me try to make them precise.

Say we have two regions of space and time, R 1 and R 2, open sets of points if you want to be mathematically
precise. These regions are such that no information can get from R 1 to R 2 without traveling faster than the speed
of light. Mathematically the condition can be expressed like this. If x1 is any four-dimensional point in R 1, and x2 is
any four-dimensional point in R 2, then the square of the distance between these points is negative:

Recall (1.18) that two points satisfying this condition are said to be separated by a spacelike interval. Every point
in our region R 1 is spacelike separated from every point in our region R 2.

Figure 3.1: Spacelike separated regions

Let O1 be any observable that can be measured in R 1. Likewise, let O2 be any observable that can be measured
in R 2. Our theory must contain a rule for associating observables with spacetime regions:

Every Hermitian operator that is measurable in the region R 1, even according to our most abstract and
generalized sense of measurement, must commute with every Hermitian operator that is measurable in region R 2.
It’s got to be so. This conclusion comes just from the general conditions of quantum mechanics and the statement
that no information can travel faster than the speed of light. This is a very severe, very curious kind of restriction to
place on a theory, in addition to all the usual dynamics, imposing a rule associating observables with spacetime
regions. We’ve certainly never encountered it in quantum mechanics before; we’ve never encountered such a rule
in relativistic quantum mechanics, either. Have we encountered it in relativistic classical mechanics? Well, in the
relativistic theory of a point particle we haven’t; but we have in classical electrodynamics.

Maxwell’s equations have fields in them, say the electric field as a function of xµ, a spacetime point. You know
very well what you can measure if you’re stuck in the spacetime region R 1 and all you have is an electrometer.
You can measure E(x, t) in the region R 1, and that’s it. To phrase it perhaps more precisely, depending on the
shape and size of the pith balls in your electrometer, you can measure some sort of average of the field smeared
over some function f(x) which is to vanish if x is not in R 1. Of course, if you have more than just pith balls, if you
have dielectrics and other electromagnetic materials, you might be able to measure squares of the electric field or
more. For example, you might be able to measure

where f(x1, x2) = 0 if x1 and/or x2 is outside of R 1. That is to say there is an entity in familiar classical physics that
does enable us, in a natural way, to associate observations with definite spacetime regions: the electromagnetic
field. What you can measure are the electric and magnetic fields, or perhaps combinations of them, in that region,
but not outside that region. We can’t design an apparatus right here to measure the electric field right now over
there in Andromeda. This gives us a clue as to how to associate observables with spacetime regions in relativistic
quantum mechanics. What we need is the quantum analog of something like the electromagnetic field. We have to
find a field, ϕ(x), or maybe a bunch of them, ϕ a(x), operator-valued—because we’re now in quantum mechanics,
and observables are operators—functions of space and time. Then the observables—I’m just pretending to guess
now, but it’s a natural guess—the observables we can measure in a region R 1 are things that are built up out of
the field (or fields, if there are many fields involved), restricted to the spacetime region R 1. What is strongly
suggested is that quantum mechanics and relativistic causality force us to introduce quantum fields. In fact,
relativistic quantum mechanics is practically synonymous with quantum field theory.
 So one way (perhaps not the only way) of implementing Einstein causality—that nothing goes faster than the
speed of light—within the framework of a relativistic quantum theory is to construct a quantum field, the hero of
this course—one of the two heroes, I should say; the other is the S matrix, but we won’t get to that for a while. The
center of all our interest, the hero sans peur et sans reproche2 of this course, is the quantum field. That will give us
a definition of what it means to localize observations. Once we have that definition, we won’t have to worry about
what it means to localize a particle. Forget that, that’s irrelevant. If we know where the observations are, we don’t
have to know where the particles are. If we know where the Geiger counter is, and if we know what it means when
we say the Geiger counter responds, the implications of doing a measurement with the Geiger counter change.
Ultimately, what we want to describe are observations, not particles. The observables we build as functions of the
fields will commute for spacelike separations if the fields commute for spacelike separations. If we can construct a
quantum field, we will settle two problems at once. We will see how our theory can be made consistent with the
principle of causality, and we will make irrelevant the question of where the particles are. If we can’t do it with
fields, we’ll have to think again. But we will be able to do it with fields. I want to remind you of something I said at
the end of the last section. These fields will depend not only on space but on time. That is, we are working in the
Heisenberg picture.

3.2Conditions to be satisfied by a scalar quantum field

We will try to build our observables from a complete set of N commuting quantum fields ϕ a(x), a = 1, . . . , N, each
field an operator-valued function of points xµ in spacetime. We will construct our fields out of creation and
annihilation operators. What I am going to do is write down a set of conditions that we want our fields to satisfy, so
that they give us a definition of locality. Some of these conditions will be inevitable; any field we can imagine must
satisfy these conditions. Some of them will just be simplifying conditions. I’ll look for simple examples first, and
then if I fail in my search for simple examples—but in fact I won’t fail—we can imagine systematically loosening
those conditions and looking for more complicated examples.

These five conditions will determine the form of the fields:

1.[ϕ a(x), ϕ b(y)] = 0 if (x − y)2 < 0, to guarantee that observables in spacelike separated regions
commute.

2.ϕ a(x) = ϕ a(x)†. The fields are to be Hermitian (and so observable).

3.e−iP·y ϕ a(x)eiP·y = ϕ a(x − y). The fields transform properly under translations.

4.U(Λ)†ϕ a(x)U(Λ) = ϕ a(Λ−1x). The fields transform properly as scalars under Lorentz transformations.

5.The fields are assumed to be linear combinations of the operators,

I want to say more about these conditions before we apply them.

Let’s start with the first and second conditions. We will frequently have occasion to deal—not so much in this
lecture, but in subsequent lectures—with non-Hermitian fields. Of course, only a Hermitian operator can be an
observable but sometimes it’s convenient to sum together two Hermitian operators with an i in the summation to
make a non-Hermitian operator. So I might as well tell you now that if I talk about a non-Hermitian field I mean its
real and imaginary parts, or, this being quantum mechanics, its Hermitian and anti-Hermitian parts, are separately
observables, to take account of the possibility of non-Hermitian fields. In other words,

That’s just tantamount to saying the Hermitian and anti-Hermitian parts of the fields are all considered as a big set
of fields, and all obey the first condition.

Now let’s talk about the third and fourth conditions, on the transformation properties of the fields. We know in
the specific case we’re looking at, Fock space, how spacetime translations and Lorentz transformations act on the
states. This should tell us something about how these transformations act on the fields. Let’s try to figure out what
that something is, by considering the more limited transformations of space translations and ordinary rotations.
First, space translations.
 We can conceive of an operator valued field even in non-relativistic quantum mechanics. Suppose for
example we have an electron gas or the Thomas–Fermi model. We can think of the electron density at every
space point as being a field, an observable, an operator. It’s a rather trivial operator, of course, delta functions
summed over the individual electrons, but it’s an operator that’s a function of position. Let’s call this operator ρ(x).
The point x is not a quantum variable, it is just the point at which we are asking the question “What is the electron
density?” If we have some arbitrary state |ψñ, I’ll write the function f(x) for the expectation value of ρ(x) in that state
|ψñ:

Now suppose we spatially translate the state. I define

This is the state where I’ve picked up the whole boxful of electrons and moved it to the right by a distance a. Now
what of the expectation value? It becomes

If there’s any sense in the world at all, this must be f(x − a). Perhaps that minus sign requires a little explanation.

Let’s say I plot f(x) peaked near the origin. Now if I translate things by a distance a to the right, I get the
second plot:

Figure 3.2: Expectation values for a state |ψñ and the state |ψ′ñ translated by a

The value of f(x − a) is peaked at x = a if f(x) is peaked at the origin. That is the correct sign for moving the state
over from being centered at the origin to being centered at a. And that’s why there is a minus sign in this equation
and not a plus sign. Now of course

by the definition of |ψ′ñ. On the other hand, rewriting (3.6) for x → x − a, we can say

Since |ψñ is an arbitrary state, and a Hermitian operator is completely determined by its expectation values in an
arbitrary state, we can eliminate |ψñ and write

in agreement with (1.39). This is simply the statement that if you translate the fields as you translate the states, this
is what you find. We generalize this in the obvious way to translations in Minkowski space, where we have both
purely spatial translations and time translations, and that is our third condition;

This last equation is in fact four equations at once. Three of them correspond to the three space components of a
and are just the previous equation rewritten. The fourth I obtained by generalization, but it should not be unfamiliar
to you. The fourth equation, the one where a points purely in the time direction, is simply the integrated form of the
Heisenberg equation of motion, since P0 is the Hamiltonian.3

As far as condition 4 goes, let’s first consider how a set of fields in general transforms under an ordinary
rotation:

However, if the fields transform as scalars, R ab = δab. That is to say,

The R −1 appears here for the same reason that the −a appeared in the previous argument. If the expectation
value is peaked at a given point, the transformed expectation value will be peaked at the rotated point.

That’s the transformation for a scalar field, like ρ(x), but not every field is a scalar. If we were to consider, for
example, ∇ρ(x), the gradient of ρ, we would discover as an elementary exercise that

As you undoubtedly know from previous courses, the gradient of a scalar is a vector, and this is the transformation
rule for rotated vector fields, a set of three operators for every spatial point. Of course gradients of scalars are not
the only vectors. There are all sorts of three-dimensional vector fields one encounters in classical physics that are
not gradients of scalar fields, for example, the magnetic field. And there are more complicated objects with more
complicated transformation laws under rotations: tensor fields, spinor fields, etc.

From the behavior under rotations, we now generalize to Lorentz transformations, just as we generalized the
space translation behavior to spacetime translations. In general, a set of fields ϕ a(x) will transform as

Again, if the fields transform as scalars,

The Λ−1 appears here for the same reason that the R −1 appeared before. If the expectation value is peaked at a
given point, the transformed expectation value will be peaked at the Lorentz transformed point. One can consider
the Lorentz transformation of much more complicated objects: tensor fields, spinor fields, etc. (In particular, the
gradient ∂µϕ a transforms as a Lorentz 4-vector if ϕ a is a Lorentz scalar.) However the scalar field is certainly the
simplest possibility, and therefore for my fourth condition I will assume my fields transform like scalars under
Lorentz transformations. This is an assumption of pure simplicity. If this doesn’t lead to a viable theory, we’ll have
to consider fields with more complicated transformation laws.

So conditions 1 and 2 are universal, absolutely necessary, while conditions 3 and 4 are just simplifying
assumptions. We can think of unitary transformations acting in two separate ways: as transformations on the
states, such that |ψñ → U|ψñ, or as transformations on the operators, A → U †AU; but not both.4

Condition 5 is a super-simplifying condition. We have these a’s and a†’s floating around, so we’ll make a very
simplifying assumption that the ϕ a(x) are linear combinations of the ap’s and ap†’s. If the linear combinations
prove insufficient, we’ll consider quadratic and higher powers of the operators. But this won’t be necessary.

3.3The explicit form of the scalar quantum field

In order to exploit these five conditions, I’ll have to remind you of the properties of the ap and ap† operators. We
worked all these out in the last lecture, so I’ll just write them down. First, the algebra of the creation and
annihilation operators:

Then, for translations,

and finally for Lorentz transformations,

To construct the fields, I’ll now use these properties of the ap’s and ap†’s, and the five conditions in reverse order.
Condition 1 is the hardest to check.

First we’ll satisfy condition 5. I’ll simply try to find the most general ϕ without an index a. If I find several such
solutions, I’ll call them ϕ 1, ϕ 2, and so on. We can start with ϕ(0) and use condition 3 to shift ϕ(0) → ϕ(x);
It will be convenient to write the fields in terms of α(p) and α †(p). There is no harm and much to be gained by
putting in the Lorentz invariant measure. So the most general form looks like this:

where fp and gp are some unknown functions of p. As always these functions fp and gp depend not on four
independent variables pµ, but only three.

At this stage fp and gp are arbitrary functions, so there’s an infinite number of solutions to condition 5, which is
not surprising. They’re not restricted to be Hermitian, or to be complex conjugates of each other. We get more
information about fp and gp by examining Lorentz invariance.

A special case of condition 4 tells us

Applying the transformation to ϕ(0) I find

Note that U(Λ) goes right through fp and gp like beet through a baby. Now I define p′ = Λp, and I can write the
integration over p′. The Lorentz invariant measure is the same, so it doesn’t change at all. The only thing that
changes is

so that

Comparing this with the first expression above for ϕ(0), the two integrands must be equal. But the α(p)’s and
α †(p)’s are linearly independent operators, therefore the coefficients must be equal, and I deduce

The values of p are constrained to lie on the upper invariant mass hyperboloid I drew previously (Figure 1.1). It
follows from special relativity that I can get from any point on this hyperboloid to any other on it by a Lorentz
transformation. Because relativity can change the value of p without changing the values of fp and gp, they have
the same value for all values of p. That implies

where f and g are constants to be determined. So conditions 5 and 4 have taken us pretty far; we are down to two
unknown constants. What about ϕ(x)? Will that involve other constants? No, because if I use condition 3,
replacing x with 0 and a with −x, I obtain

which of course I can compute since I have an expression for ϕ(0), and I know how the operators α(p) and α †(p)
transform (the same as the ap and a†p, see (3.20)). Then
Here is ϕ(x), and I still only need the two arbitrary constants f and g. I haven’t used all of the content of conditions
3, 4 and 5; for example, I’ve only used condition 4 at the origin. But I leave it as a trivial exercise for you to show
that every expression of this form satisfies the conditions 3, 4 and 5 for all x and all a. You can almost read it off, I
think.

So let’s summarize the situation before we apply conditions 1 and 2. A general field satisfying 3, 4 and 5 can
be written as the sum of two independent fields, which I will call ϕ (+)(x) and ϕ (−)(x). The fields ϕ (+)(x) and ϕ (−)(x)
are the coefficients of f and g, but I’ll write them not in terms of the α’s but the a’s, since it’s easier to compute the
commutators using the a’s:

where

and as usual when the 4-vector p appears, its time component is ωp. Note that ϕ (−)(x)† = ϕ (+)(x). The assignment
of ϕ (−)(x) to the field involving a†p seems completely bananas but it was established by Heisenberg and Pauli,5 on
the basis that ϕ (−)(x) only involves p0’s with negative frequencies, i. e. with a sign of +ip0x0 in the exponential’s
argument, and similarly with ϕ (+)(x).6

Now to apply condition 2, hermiticity. Two independent Hermitian combinations are

These are two independent cases of the most general choice satisfying condition 2:

where θ could in principle be any real number.

Now to satisfy condition 1. There are three possible outcomes:

Possibility Two independent solutions, ϕ 1(x) and ϕ 2(x), which commute with themselves and with each other.
A: Any combination aϕ 1(x) + bϕ 2(x) is observable, with a and b real constants.
iθ (+) −iθ (−)
Possibility Only the single combination e ϕ (x) + e ϕ (x) is observable. The most general Hermitian
B: combination is, aside from an irrelevant multiplying factor, some complex number of magnitude 1
times ϕ (+)(x) plus that complex number’s conjugate times ϕ (−)(x).
Possibility
The program crashes. We’ll need to weaken condition 5 or think harder.
C:

So either we have two fields or we have one field, and if we have one field, it must be of this form (3.35) to be
Hermitian. Actually we can shorten our work a bit by realizing that we can get rid of the phase factor by redefining
ap and a†p:

If we make such a redefinition, that changes no prior equation before the definition of ϕ(x). Then I might as well
consider equivalently Possibility B′:

We really only have two independent possibilities to consider: Possibility A, in which we say both ϕ (+)(x) and
ϕ (−)(x) are local fields (i.e., they commute for spacetime separations), and Possibility B′ in which we say just the
sum of ϕ (+)(x) and ϕ (−)(x) is observable, and the difference is not observable.

Now we will look at these two possibilities systematically. Everything we have to compute to check A, we also
have to compute to check B′. So let’s start with A. We want to see that everything commutes with itself for
spacelike separation. If A is true, then ϕ 1(x) must commute with ϕ 2(y), and each must commute with itself. For
example, we must have the commutator [ϕ 1(x), ϕ 2(y)] equal to zero for spacelike separations. Is it?
This function is one of a series of similar functions that will turn up again and again in our investigations. This one
is actually a Neumann function or something, but its name in quantum field theory is Δ+. It’s a function of the
difference of the spacetime points, (x − y), as is obvious from the expression, and the mass, from the definition of
ωp and the value of p0. To keep things short, since we’re only worried about one mass, I’ll suppress the µ2 and
just call this Δ+(x − y). If we were worrying about several different types of particles with different masses we
would have to distinguish between the different Δ+’s.

You might expect that Δ+(x) is a Lorentz scalar function:

This is indeed true. The argument of the exponential is a Lorentz scalar, and the factors have come together to
make the Lorentz invariant measure. The Lorentz invariance of Δ+(x) will be a useful fact to us later. Another
useful relation is

which follows easily from the definition of Δ+(x − y).

The real question we want to ask now is: Does Δ+(x) = 0 if x2 < 0? If so, we’re home free. Otherwise we have
to look at Possibility B′. Well, it doesn’t. We know it doesn’t from Chapter 1. If I take the time derivative of Δ+(x),
that cancels out the ωp in the denominator,

and we get precisely the integral (1.81) we had to consider in Chapter 1 when I wondered whether particles could
travel faster than the speed of light. Now if a function vanishes for all spacelike x2, its time derivative surely
vanishes for all spacelike x2. By the explicit computation of Chapter 1, its time derivative doesn’t vanish, so the
function doesn’t vanish, either. The answer to the question “Is Δ+(x) = 0 for spacelike x2?” is “No”. (Never waste a
calculation!) Possibility A is thrown into the garbage pail, and we turn to the only remaining hope, Possibility B′. If
B′ also gets thrown into the garbage pail not only this lecture but this entire course will end in disaster!

Here we only have one field so we only have one commutator to check. Now fortunately since this ϕ(x) is
ϕ (+)(x) + ϕ (−)(x), the commutator is the sum of four terms we have already computed:

This iΔ(x − y) is a new Lorentz invariant function (using the notation of Bjorken and Drell;7 the conventions differ
from text to text). Like Δ+(x − y), iΔ(x − y) depends on the square of the mass µ2, but if there’s only a single type of
particle around, we don’t need to write it. Does this expression equal zero for spacelike separations, (x − y)2 < 0?
Yes, and we can see this without any calculation. A spacelike vector can be turned into its negative by a Lorentz
transformation,8 so

and so

Possibility B′ thus escapes the garbage pail, and we don’t have to consider Possibility C. Our single free scalar
quantum field of mass µ is then written
or in terms of the α(p)’s and α †(p)’s,

(Particularly important equations will be boxed.) We have constructed the scalar field. It is the object that
observables are built from. Now we take off in a new direction.

3.4Turning the argument around: the free scalar field as the fundamental object

Several times in the course of our development we have introduced auxiliary objects like the annihilation and
creation operators and then showed that the whole theory could be defined in terms of their properties. I would
now like to show that the whole theory can be reconstructed from certain properties of the free quantum field. I will
have to derive those properties. The structure we have built is rigid and strong enough to be inverted. We can
make the top story the foundation.

The first property is trivial to demonstrate, that ϕ(x) obeys this differential equation:9

This is just a statement that in momentum space p2 equals µ2. This is most easily shown by rewriting (3.46) in
terms of the explicitly Lorentz invariant measure:

If I differentiate twice with respect to xµ I obtain

so that

The product xδ(x) is identically zero, so the product (−p2 + µ2)δ(p2 − µ2) guarantees the integrand vanishes, and
ϕ(x) satisfies the differential equation.

The equation (3.47) has a famous name. It is called the Klein–Gordon equation.10 As you might guess from
the name, it was first written down by Schrödinger,11 but he didn’t know what to do with it. He wrote it down as a
relativistic analog of the free Schrödinger equation. Recall that Schrödinger’s original equation comes from the
replacement E → iħ∂/∂t, p → −iħ∇ into a Newtonian expression for the energy, now regarded as an operator
equation acting on a wave function. Schrödinger, no dummy, knew the relativistic expression for energy and made
the same substitutions into that. Then he said “Arrgh!” or the German equivalent, because he observed that the
solutions had both positive and negative frequencies. And he said, “If this is a one-particle wave equation we are
in the soup because we only want positive energies. We don’t want negative energies!” We have encountered this
equation not as a one-particle wave equation—that’s the wrong context, that’s garbage—but as an equation in
quantum field theory where particles may be created and annihilated.

We already have the second property:

These two equations (3.47) and (3.51), as I’ll sketch out, completely define the Hilbert–Fock space and everything
else. We postulate these two equations, together with the assumption of hermiticity (ϕ(x) = ϕ(x)†) and the scalar
field’s behavior under translations and Lorentz transformations (conditions 3 and 4 on p. 34.) In this way the scalar
field which we’ve introduced as an auxiliary variable can just as well be thought of as the object that defines the
theory.

We begin with the Klein–Gordon equation, (3.47). We can write the solution to it in its most general form (the
factors of (2π)−3/2(2ωp)−1/2 are included for later convenience),

This is the most general expression for a solution of the Klein–Gordon equation with unknown Fourier components
ap and bp. The condition of hermiticity requires that bp = a†p, so that the most general solution is just what we had
before, (3.45):

We could now deduce the commutators of ap and a†p uniquely by substituting this expression into the second
equation (3.51) and Fourier transforming the result. Once we have observed that the commutators are unique, we
don’t have to go through the whole calculation because we already know one commutator of ak with a†k that is
consistent with everything else, the delta function δ(3)(k − k′), as in (3.18).

Finally from condition 3,

we can deduce the commutators of the ak and a†k with the Pi’s and the Hamiltonian simply by differentiation. For
example, differentiating the previous equation gives the Heisenberg equation of motion

Plugging the expression (3.45) in gives

which gives, by Fourier transformation, the commutators of ap and a†p with the Hamiltonian, telling us that a†p is
an energy raising operator, identical to (2.22), and ap is an energy lowering operator, the same as (2.21). And off
we go! Just as in the middle of the last section, we can reconstruct all of Fock space on the basis of this operator
algebra.

So that was a sketch, not a proof. I’ve leapt from mountain peak to mountain peak without going through the
valleys but I hope the logic is clear. Of course this procedure does not give us a zero of energy, the energy of the
ground state, but that’s just a matter of convention. We can always define that by convention to be zero.

Now this is not all. We go on because (3.51) can be weakened. This commutator, our condition 1 (p. 34) can
be replaced by two separate equations, two new commutators, say 1′(a) and 1′(b). The first specifies the
commutator of ϕ(x, t) with ϕ(y, t). That is to say, the time components of the two points x and y are to be taken as
equal; this is the so-called equal time commutator. The result is a definite function which we can compute. The
second will be the equal time commutator of (x, t) with ϕ(y, t), where the dot always means time derivative. This
will equal something else, again a definite numerical function which we will shortly compute.12

Why do I say that condition 1 can be replaced by conditions 1′(a) and 1′(b)? Well, it’s because the
Klein–Gordon equation is a differential equation second-order in the time. I can operate on the commutator with
2 = ∂2x , considering the variable y as fixed. Therefore I can just bring the operator through and use the
x
Klein–Gordon equation. Consequently

We know the solution of the second-order differential equation for arbitrary values of the argument if we know its
value and its first time derivative at some fixed time: the initial value conditions. We need only compute the
equation 1′(a)

and equation 1′(b)

for some fixed time t, integrate away as in all books on differential equations, and we will know the solution
uniquely. That will be sufficient—because we know iΔ(x − y) obeys a differential equation, second-order in
time—to compute these commutators and iΔ(x − y) for all times. So let’s calculate. From (3.51),

because the integrand is an odd function. Equation 1′(b) is also easily computed:

because the integrand is an even function.

As I’ve argued, conditions 1′(a), 1′(b) and 2 are sufficient to reconstruct the whole theory. The field which we
introduced as an auxiliary entity not only gives us a definition of locality consistent with the dynamics we had
before, but in fact all the dynamics we had earlier can be expressed in terms of this field: it obeys the
Klein–Gordon equation, it is Hermitian, it satisfies these two equal time commutation relations. Your homework
problems ask you to play with these equations to develop certain identities that will be useful to us later on in the
course.

3.5A hint of things to come

I’m now going into mystic and visionary mode, to remind you that these equations look very similar to some
equations you might very well have encountered before, in mechanics and in non-relativistic quantum mechanics:
good old canonical commutation relations and canonical quantization. We have a set of equations for the
Heisenberg picture operators in non-relativistic quantum mechanics. There’s normally an ħ in these equations but
I’ve set ħ equal to 1. In fact, there is a third that comes with the first two:

Now the first two of these equations bear a certain structural similarity to the equations (3.60) and (3.61) if I
identify (x, t) with pa and ϕ(x, t) with qb. Instead of the discrete indices a and b labeling the various coordinates I
have continuous indices x and y, and as a consequence of that, instead of a Kronecker delta I have a Dirac delta
function, but otherwise they look very similar.

To test that vague similarity let me try to compute the analog of the third equation. If I identify pa with
(x, t) in the system I have to compute

which, if the analogy holds, should equal zero. This is nearly the same computation as before;

because the integrand is again an odd function. The commutator does equal zero, which looks awfully like the
third equation. To summarize,

Therefore, there seems to be some vague connection with the system we have developed without ever talking
about canonical equal time commutation relations and the canonical quantization method. Maybe. Or maybe I’m
just dribbling on at the mouth. But there seems to be a certain suggestive structural similarity. In the next section, I
will exploit that similarity in a much more systematic way. It’s going to take two or three minutes for me to explain
what that systematic way is.

With the new method I will develop this entire system by a completely different and independent line of
approach. This method will be the method of canonical quantization, or as I will describe it somewhat colorfully,
the “method of the missing box”.

At the start of the next section I will review, in my characteristic lightning fashion, the introductory parts arising
from material I assume you all know, the mechanics of Lagrange and Hamilton. You also may or may not know
that you can generalize classical particle theory consistent with an infinite number of degrees of freedom or a
continuous infinity of degrees of freedom and write down Hamiltonians and Lagrangians for classical field theory.
There is also a standard procedure for getting from classical particle theory to non-relativistic quantum mechanics
which I will review. What we will attempt to do in the second half of the next lecture is fill in the “missing box”, to get
to the thing you don’t know anything about, or I pretend you don’t know anything about, quantum field theory. We
will again arrive at the same system, but by another path.

1[Eds.] Rubbia shared the 1984 Physics Nobel Prize with Simon van der Meer, for experimental work leading to
the discoveries of the W± and Z0 (see §48.2). At the time of these lectures, Rubbia was commuting between
CERN and Harvard.
2[Eds.]Originally this French phrase described the “perfect knight” Pierre Terrail, Chevalier de Bayard
(1473–1524), “without fear and without flaw”.
3[Eds]. Let aµ = (dt, 0, 0, 0) be an infinitesimal translation in time. Then with P0 = H,

(expanding the right-hand side in a Taylor series) or i[H, ϕ(x)] = dϕ(x)/dt, which is just the Heisenberg equation of
motion.
4[Eds.]One can define a transformation U in terms of its action on the states, and then check that it acts correctly
on the operators, or one can define it in terms of its action on the operators, and check that it has the proper effect
on states. But one should not simply assume that it works both ways.
5[Eds.]Werner Heisenberg and Wolfgang Pauli, “Zur Quantendynamik der Wellenfelder” (On the quantum
dynamics of wave fields) Zeits. f. Phys. 56 (1929) 1–61, “Zur Quantendynamik der Wellenfelder II”, Zeits. f. Phys.
59 (1930) 168–190.
6[Eds.] Schweber RQFT, p. 167.
7[Eds.] Bjorken & Drell Fields, Appendix C.
8[Eds.] This would not be true for a timelike vector. Proper Lorentz transformations move a timelike vector xµ
satisfying x2 = κ2, inside the light cone, (xµ1 and xµ2 in the diagram) around the upper and lower hyperboloids t =
, respectively, but cannot change the sign of t, and so cannot transform a forward pointing vector like xµ1
into a backward pointing vector like xµ2. By contrast, proper Lorentz transformations move a spacelike vector xµ
satisfying x2 = −κ2, outside the light cone, (xµ3 and −xµ3) around on the hyperbolic sheet |r|2 − t2 = κ2. Since both
xµ3 and −xµ3 lie on the same sheet, a spacelike vector can always be Lorentz transformed into its negative.

9[Eds.] To remind the reader: Though most authors let ≡ ∂µ∂µ, Coleman writes 2.

10[Eds.]Walter Gordon, “Der Comptoneffekt nach der Schrödingerschen Theorie” (The Compton effect according
to Schrödinger’s theory), Zeits. f. Phys. 40 (1926) 117–133; Oskar Klein, “Elektrodynamik und Wellenmechanik
vom Standpunkt des Korrespondenzprizips” (Electrodynamics and wave mechanics from the standpoint of the
Correspondence Principles”), Zeits. f. Phys. 41 (1927) 407–422. According to Klein’s obituary (“Oskar Klein”,
Physics Today 30 (1977) 67–88, written by his son-in-law, Stanley Deser), Klein symmetrically anticipated
Schrödinger’s more familiar equation, but was prevented from publishing it by a long illness.
11[Eds.] Erwin Schrödinger, “Quantisierung als Eigenwertproblem (Viete Mitteilung)” (Quantization as an
eigenvalue problem, part 4.), Ann. Physik 81 (1926) 109–139. English translation in Collected Papers on Wave
Mechanics, E. Schrödinger, AMS Chelsea Publishing, 2003. See equation (36).
12[Eds.] The commutators can be restricted to equal times, because spacelike vectors can always be Lorentz
transformed to purely spatial vectors, with zero time components. With the 4-vector x − y transformed to a purely
spatial vector, x0 − y0 = 0. So the restriction of spacelike separation can be replaced by the weaker condition x0 =
y0. See note 8, p. 42.

Problems 1

He who can do nothing,

understands nothing.

PARACELSUS
1.1 Some people were not happy with the method I used in class to show

where p and p′ are single-particle 3-momenta connected by a Lorentz transformation, while ω and ω′ are the
associated energies. Show the equation is true directly, just by using the elementary calculus formula for the
change in a volume element under a change of coordinates. (HINT: The equation is obviously true for rotations, so
you need only to check it for Lorentz boosts (to frames of reference moving at different speeds). Indeed, you need
only check it for a boost in the z-direction.)
(1997a 1.1)

1.2 This problem and 1.3 deal with the time-ordered product, an object which will play a central role in our
development of diagrammatic perturbation theory later in the course.

The time-ordered product of two fields, A(x) and B(y), is defined by

Using only the field equation and the equal time commutation relations, show that, for a free scalar field of mass µ,

and find c, the constant of proportionality.

(1997a 1.2)

1.3 Show that

The limit symbol indicates that goes to zero from above, i.e., through positive values. (If were not
present, the integral would be ill-defined, because it would have poles in the domain of integration.) (HINTS: Do the
p0 integration first, and compare your result with the expression for the left-hand side obtained by inserting the
explicit form of the field (3.45). Treat the cases x0 > y0 and x0 < y0 separately.)
(1997a 1.3)

1.4 In a quantum theory, most observables do not have a definite value in the ground state of the theory. For a
general observable A, a reasonable measure of this quantum spread in the ground state value of A is given by the
ground state variance of A, defined by
where the brackets á· · ·ñ indicate the ground state expectation value.

In the theory of a free scalar field φ(x) of mass µ, define the observable

where a is some length. Note that the Gaussian has been normalized so that its space integral is 1; thus this is a
smoothed-out version of the field averaged over a region of size a. Express the ground state (vacuum) variance of
A(a) as an integral over a single variable. You are not required to evaluate this integral except in the limiting cases
of very small a and very large a. In both these limits you should find

where α and β are constants you are to find (with different values for the different limits), and the . . . denote terms
negligible in the limit compared to the term displayed. You should find that var A(a) goes to zero for large a while it
blows up for small a. Speaking somewhat loosely, on large scales the average field is almost a classical variable,
while on small scales quantum fluctuations are enormous.
(1997a 1.4)

Solutions 1

1.1 Consider Λ, a boost in the z-direction:

where tanh χ = v/c = v (in units where c = 1.) The change of volume element is given by the Jacobian determinant,

where

But

so

Using (S1.1) for ω′, and (S1.5) for the Jacobian gives

1.2 The Heaviside theta function, or step function, θ(x), is defined by

The extension to θ(x − a), where a is a constant, should be clear. Its derivative is a delta function:
Using theta functions, we can write the time-ordered product (P1.2) of two operators A(x), B(y) like this:

The d’Alembertian 2 (the 4-vector equivalent of ∇2, the Laplacian) is

Look at the first partial derivative with respect to x0:

Delta functions are even: δ(x − y) = δ(y − z). Also, as δ(x − a) = 0 unless x = a, f (x)δ(x − a) = f (a)δ(x − a). The two
terms involving delta functions can be written

because the equal time commutator of the two fields equals zero. (But see the Alternative solution below!) Then

The second derivative goes much the same way,

because at equal times. The Laplacian does not act on the θ functions, so

and consequently

because ϕ satisfies the Klein–Gordon equation, Then

in agreement with (P1.3), and the constant c = −i.

Alternative Solution. A purist might object to setting the quantity δ(x0 − y0)[ϕ(x), ϕ(y)] equal to zero. After all, the
differential equation is second-order, and maybe we should carry the second time derivative all the way through;
. Then

Let’s look carefully at the first line of the previous equation. We can write

Delta functions really only make sense in the context of being under an integral sign, multiplying some suitably
smooth function. If we integrate ∂0δ(x0 − y0)ϕ(x)ϕ(y) with respect to x0 and use integration by parts, assuming that
ϕ(x) → 0 as x0 → ±∞, then we can say

Using this identity, we have

Plug this (and a similar expression with ϕ(x) and ϕ(y) swapped, and an extra − sign) into the original equation to
obtain

which gives the same result.

1.3 We need to show that for x0 > y0,

and for y0 > x0,

The right-hand sides of (S1.10) and (S1.11) are the same. Swap x and y (so that now x0 > y0) and obtain

The only difference between (S1.10) and (S1.12) is the sign of the exponential’s argument. But if we take p → −p,
nothing changes except the sign of the argument: (S1.10) and (S1.12), and hence also (S1.11), are equivalent.
(There’s a second argument that’s worth seeing; it will be given at the end.) Let’s work on (S1.10) first. In the
product of ϕ(x)ϕ(y), there will be four separate products of creation and annihilation operators, aa, aa†, a†a and
a†a†. Sandwiched between vacuum states, only the second term survives, because Because
we can write

Then

The integrals sandwiched between the vacuum states are c-numbers, so the integrals merely multiply the inner
product á0|0ñ = 1. Either integral can be done quickly owing to the delta function. Performing the q integral gives

Now to work on the right-hand side of (S1.10), substituting in the value c = −i found in Problem 1.2:

Rewrite the p0 integral:

Because is small, we can write

where η is also a small quantity: η → 0 as → 0. Rewrite the p0 integral once again;

 Figure S1.1: Contours for the p0 integral in

Use Cauchy’s integral formula to evaluate this, by extending p0 to the complex plane. We’ll use a contour which
has a large semicircular arc of radius R and a diameter along the real axis; we need to choose the upper or the
lower contour. There are two poles, at ±(ωp − iη). For case (S1.10), x0 > y0, the quantity in the exponential will be
negative if Im p0 is negative, so that the semicircular arc of radius R will contribute nothing as R → ∞. That means
we take the contour below the real axis, enclosing only the root ωp − iη. Then by Cauchy’s formula

the extra factor of (−1) coming because the bottom contour is clockwise. We can now safely take the limit η → 0,
and the p0 integral gives

Put this back into the original integral (S1.14) to obtain

The right-hand side of (S1.15) is identical to the right-hand side of (S1.13), so the left-hand side of (S1.15) must
equal the left-hand side of (S1.13). That establishes case (S1.10) (and (S1.11) also, since they’re equivalent).

Case (S1.11) can also be done on its own. Now y0 > x0. By symmetry, we can write down at once the
equivalent of (S1.13):

The right side of (S1.10) is the same as the right side of (S1.11). The p0 integral is the same as before, but now y0
> x0. That means the imaginary part of p0 has to be positive in order to guarantee that the semicircular arc
contributes nothing. Now we take the upper contour, counter-clockwise, which encloses only the root −ωp + iη.
Then

as η → 0. Not surprisingly, the sign of the exponential’s argument changes from the previous calculation.
Substitute this back into (S1.14) to obtain

Unfortunately, the sign of the exponentials do not now match up to give the inner product of two 4-vectors; we’d
need the space parts to be negative. That’s easy to arrange: Let p → −p. Equation (S1.17) becomes

The right-hand side of (S1.16) is the same as the right-hand side of (S1.18), so the left-hand side of (S1.16) is the
same as the left-hand side of (S1.18). That establishes (S1.11).

1.4 We first notice that

since ϕ(x, 0) is linear in ap and So var A = áA2ñ. To calculate áA2ñ, notice that A2
involves . In the product of the two ϕ’s, the only non-zero term will be of the form , where q is a
dummy momentum variable. Then

The vacuum expectation value can be rewritten as

because Then integrating over q with the help of the delta function,

The integrals over x and y have the same form. Looking at the integral over x,

using the identity with c = 1/a2 and b = ipi. Then

Now to consider the limits. As a → 0,

As a → ∞,

For ϕ (+)(x) with ϕ (+)(y), spacelike, shmacelike; they all involve nothing but annihilation operators, and all
annihilation operators commute with each other no matter what we multiply them by, so that’s zero. By similar
reasoning, or by taking the adjoint, the same thing goes for ϕ (−)(x) with ϕ (−)(y).

Now we come to the crunch. Let’s compute

Just as claimed, as a → ∞, the variance tends to zero; as a → 0, when quantum fluctuations are expected to
be enormous, the variance tends to infinity.

4
The method of the missing box

In the last lecture I told you we would find the same object, the quantum field, we had found a few minutes earlier,
by a rather lengthy sequence of investigations, using a totally different method which I described in my
characteristic colorful way as the method of the missing box. The method may be illustrated by this diagram:
 Figure 4.1: The missing box

I presume that three of these boxes are familiar to you. I will give brief summaries of them, complete but fast,
in the first half of this lecture. We start out at the upper left corner with classical particle mechanics, summarize
that, and, moving down, summarize how that extends to quantum particle mechanics. If you’ve had a good course
in non-relativistic quantum mechanics, you know that there is a standard procedure for getting from classical
particle theory to quantum theory, which I will review, called canonical quantization. Just to remind you of what
that is I’ll say it in great detail: You write the system in Hamiltonian form and you set commutation relations
between the classical p’s and q’s. This leads to quantum particle theory. We also can move across, to the right,
and summarize how classical particle mechanics is extended to systems with infinite numbers of degrees of
freedom, indeed continuously infinite numbers of degrees of freedom, i. e., classical field theory: the classical
theory of Maxwell’s equations, of sound waves in a fluid, of elasticity in an elastic solid; classical continuum theory.
What we will attempt to do in the second half of this lecture is to fill in the missing box, quantum field theory. As the
arrows show it can either be viewed as the result of applying canonical quantization to classical field theory,
following the arrow down; or alternatively, by following the arrow across from quantum particle theory, generalizing
to systems with a continuous infinity of degrees of freedom. In the language of the algebraic topologists, this is a
commutative diagram.

4.1Classical particle mechanics

Classical particle mechanics deals with systems characterized by dynamical variables, ordinary real number
functions of time called generalized coordinates. I will denote these as

In the simplest system these may be the coordinates xi of an assembly of N particles moving in 3-space where i
goes from 1 to 3N. These could represent the three Cartesian coordinates or the three spherical coordinates of
each of the particles. Lagrangian systems are those whose dynamics are determined by a function L called the
Lagrangian. It depends on the qa’s and their time derivatives, which I indicate with a dot, a, and possibly explicitly
on the time:

We define a functional1 called the action, S—ahistorically by the way; it is not the action first introduced by
Maupertuis2—by the integral

The Lagrangian determines the equations of motion of the system via Hamilton’s Principle, which is the statement
that if I consider a small variation in the qa’s,

the resulting change in the action is zero:

I use δ to indicate an infinitesimal variation; the Weierstrass revolution in calculus has not yet reached this lecture:
we are Newtonians. The variations are subject to the restriction that they vanish at both endpoints of the
integration;
From Hamilton’s Principle one can derive equations of motion by the standard methods of the calculus of
variations. One simply computes δS for a general change δqa:

And here I have made a slight notational simplification by adopting the Einstein summation convention over the
index a, so I don’t have to write a couple of sigmas. As it will turn up again and again in our equations, I will define
pa, the canonical momentum conjugate to qa,

(By the way I’ve arranged my upper and lower indices so that things look like they do in relativity: Differentiation
with respect to an object with a lower index gives you an object with an upper index and vice versa. It’s just a

matter of definition.) From the definition of δqa, of course . By substitution and integration of the last
term by parts we obtain

Since δqa are supposed to be arbitrary infinitesimal functions which vanish at the boundaries, the last term above
equals zero and the quantity inside the square brackets must vanish everywhere. Thus we obtain the equations of
motion

These are the Euler–Lagrange equations. I will not do specific examples. I presume you’ve all seen how this
works out for particles and a system of particles with potentials and velocity dependent forces and all of those
things. This gets us halfway through the first box. I will now discuss the Hamiltonian formulation.

We consider the expression defined by the Legendre transformation3

H is called the Hamiltonian. It can be thought of as a function not of the qa’s and the a’s, which is the natural way
to write the right-hand side, but of the qa’s and the pa’s and possibly also of time. I will just tell you something
about the Hamiltonian which you may remember from classical mechanics, though in fact we will prove it in the
course of another investigation later on. If the Lagrangian is independent of time, then the Hamiltonian is identical
with the energy of the system, a conserved quantity whose conservation comes from invariance of the Lagrangian
under time translation.

Let us consider the change in the Hamiltonian when we vary the qa’s and the a’s (or equivalently, the
pa’s and the qa’s) at a fixed time:

the sum on a always implied. This is just the Chain Rule for differentiation. The second and fourth terms cancel,
and ∂L/∂qa = a. Because we can vary the pa’s and qa’s independently, we can now read off Hamilton’s equations,

I presume they are also familiar to you, and I shall not bother to give specific examples.

This is a standard derivation, but I should like to make a point that is sometimes not made in elementary texts.
We will have to confront it several times, not in this lecture but in subsequent lectures. In order to go from the
Lagrangian formulation to the Hamilton formulation there are certain conditions which the pa’s and qa’s must obey
as functions of the a’s and the a’s. The pa’s and qa’s must be complete and independent. Tacitly I’m assuming
that these functions, the qa’s and the a’s, have two properties. I assume, first, that it is possible to write the
Hamiltonian as a function of just the qa’s and the pa’s. Maybe that’s not so. In most simple cases it is so, but it’s
very hard to prove in general that it is always so, because I can write examples where it’s not so. So this is the
condition which we will call completeness. If the set of the qa’s and the a’s is complete, it is possible to
express the qa’s and the a’s as functions of the qa’s and the pa’s at least to such an extent that it is possible
to write the Hamiltonian as just functions of the qa’s and the pa’s. By independent I mean that I can make small
variations of the qa’s and the pa’s at any time by appropriately choosing the variations of the qa’s and the a’s

independently. If I couldn’t make such small variations, if there were some constraint coming from the definition of
the pa’s that kept me from varying them all independently, then I couldn’t get from (4.10) to (4.11), because I
couldn’t vary them one at a time.

To give a specific example where the qa’s and the pa’s are complete but not independent, consider a particle
of mass m constrained to move on the surface of a sphere of unit radius. If you know any classical mechanics at
all, you know there are two ways of doing this problem. You have three dynamical variables, three components of
the position vector x of the particle. You can of course go to some coordinates in which you have only two
variables, such as spherical coordinates. Then you don’t have any equation of constraint and off you go by the
standard methods. Alternatively you can keep all three coordinates and write things in terms of Lagrange
multipliers. That is to say you can write a Lagrangian,

by the method of Lagrange multipliers, which I hope you all know—if not, take five minutes to read the appropriate
section in Chapter 2 of Goldstein.4 If I stick this last term in a Lagrangian I get precisely equivalent equations to the
Lagrangian in two coordinates without the constraint. By varying with respect to λ, I obtain the equation of
constraint, and by varying with respect to the three other variables I obtain the equations of motion with the force of
constraint on the right-hand side. From the viewpoint of mechanics this constrained Lagrangian is just as good as
the other. However it does not allow passage to a Hamiltonian form by the usual procedure: pλ, the canonical
momentum associated with the variable λ, happens to be zero. There is no = dλ/dt in the Lagrangian, and λ is
not an independent variable. I cannot get the Hamilton equations of motion involving the three components of x
and their conjugate momenta and λ and its conjugate momentum because I cannot vary with respect to pλ, which
is zero by definition; zero is not an independently variable quantity. The equation (4.10) is true, but the equation
(4.11), which appears to be such an evident consequence of it, is false. Things break down because the
generalized coordinates aren’t independent. There is no method of Lagrange multipliers in the Hamiltonian
formulation of mechanics.

This is just something to keep in the back of your mind because all of the examples we will do in this
lecture—in fact, everything—will be complete and independent. But then in later lectures we’ll get to things where
they’re not. And if you have a Lagrangian system in such a form where you do not have a bunch of independent
variables, then you have to beat on it, in the same way as we beat on this example, by eliminating Lagrange
multipliers until you get it into shape where you can go to Hamiltonian form. This completes for the moment my
discussion of the first box, classical particle mechanics.

4.2Quantum particle mechanics

We go now to the second box, quantum particle mechanics and canonical quantization. I’m going to explain the
arrow leading from classical mechanics to quantum mechanics, and something about what lies at the end of the
arrow. Of course it will not be everything in quantum mechanics; that’s a course by itself!

Canonical quantization is a uniform procedure for obtaining a quantum mechanical system from a given
classical mechanical system in Hamiltonian form, by turning a crank. It is certainly not the only way of getting
quantum mechanical systems. For example, when you took quantum mechanics, you didn’t take care of the theory
of electron spin by starting out with the classical theory of spinning electrons and canonically quantizing it.
However it is a way and it has certain advantages. I will first explain the prescription and then the ambiguities that
inevitably plague canonical quantization. Finally I will explain its advantages.

The quantum mechanical system has a complete set of dynamical variables that are the q’s and the p’s of the
classical system. I will abuse notation by using the same letters for the quantum variables as the classical
variables, instead of writing them with capitals or writing them with a subscript “op” or something. The classical
dynamical variables obey the canonical Poisson brackets5

We replace these dynamical variables by time-dependent (Heisenberg picture) operator-valued functions, which
obey these universal commutators, independent of the system:

(Traditionally there is a factor of ħ on the right-hand side of the last equation, but we’re keeping ħ = 1.) The
commutators are trivial except for the (q, p) commutators, and for that matter the (q, p) commutators are also
pretty trivial. We assume that the set of qa and pa are Hermitian (and hence observable), and complete.

The Hamiltonian of the quantum system is the same as the classical Hamiltonian, but now it is a function of
the operators qa and pa;

Please notice that the prescription for constructing the Hamiltonian is inherently ambiguous. It doesn’t tell you
what order you are to put the qa’s and pa’s in, when you write out the expression for H. In the classical expression
it doesn’t matter if you write p2q2 + q2p2 or if you write 2pq2p, but in the quantum theory it does make a difference.
I choose this particular example because the ambiguity cannot be resolved just by saying a quantum Hamiltonian
should be Hermitian. This is just an ambiguity that we have to live with. The prescription of replacing the classical
p’s and q’s by their quantum counterparts does not define a unique theory except in especially simple cases. In
general there is no way to resolve ordering ambiguities. If we write the commutator with traditional units,

so there are no ordering ambiguities in the classical limit.

For this reason we always try and write our quantum systems in terms of the coordinates of our classical
system before canonical quantization, so that the ordering ambiguity causes the least damage for particles
moving in a potential. (We usually quantize the system directly in Cartesian coordinates. If we are then to do a
transformation to spherical coordinates, we do that after we have quantized the system, after we have written
down the Schrödinger equation.) Why do we do this? It is an ambiguous rule. Why on earth would any sane
person or even an inspired madman have written down this particular rule rather than some others? Well,
historically the only motivation for connecting a classical mechanical system with a quantum system was the
Correspondence Principle, the statement that the quantum system in some sense should reproduce the classical
system if, for some set of experiments concerning that system, classical physics gives a good description.

The operator that generates infinitesimal time evolutions of the quantum system is the classical Hamiltonian
function of the quantum qa and pa operators. For any operator A(t),

the last term appearing only if the operator has an explicit time dependence in addition to the implicit time-
dependence arising from the qa’s and pa’s. In particular, we can rederive the Heisenberg equations (4.11),

Let me explain the second step. Because of the canonical commutation relations, taking the commutator of qa
or pa with any function of the p’s and q’s amounts to differentiation with respect to the conjugate variable, times a
factor of ±i. For example, taking the commutator of a monomial such as qbpc pdqe with qa, we get

If there is a single pa in the expression, I get a 1 from the commutator, if there is a pa2 I get 2pa’s, if there’s a pa3 I
get 3pa2, etc. Thus the quantum mechanical definition of the Hamiltonian tells us that a equals i[H, qa], which is
∂H/∂pa. Since pa is just another operator, likewise a is just the commutator i[H, pa] which because of the minus
sign when I switch around the canonical commutator gives us −∂H/∂qa. Canonical quantization is a prescription
that guarantees that the Heisenberg equations of motion for the quantum mechanical system are identical in form
with the Hamilton equations for the corresponding classical system. This is an expression of the Correspondence
Principle.

Consider a state in which classical mechanics offers a good description, a state where at least for the
duration of our experiment, áqnñ, the expectation value of qn, equals áqñn, the nth power of the expectation value of
q, within our experimental accuracy—we don’t know that q is statistically distributed—and likewise the expectation
value ápnñ of the nth power of p is ápñn, the nth power of the expectation value of p. Then by taking the expectation
value of the quantum equations of motion, we observe that they equal, via the mean values of the particle position
and momentum, the classical equations of motion. Of course, if the state does not obey that classical condition, if it
is not (within our experimental accuracy) a sharp wave packet in both p and q, then quantum mechanics gives
different results from the classical physics.

This concludes my rather brief discussion of the arrow descending from the first box, classical mechanics to
quantum mechanics, the second box. We have taken care of classical mechanics and quantum mechanics in one
half hour. Well, of course there is a lot more to be said about these systems and we’ll return to them occasionally
to get clues to say some of those things. But that’s the only part of them I will need for this lecture.

4.3Classical field theory

Now we come to something that might be novel to some of you: the extension from classical particle mechanics to
classical field theory. In general the only difference between classical particle mechanics and classical field theory
is that in one case the variables are finite in number, and in the other case one has an infinite number of variables.
The infinite number of the dynamical variables, say in Maxwell’s electromagnetic theory, are labeled sometimes
by a continuum index. Instead of worrying about the position of the first particle, the position of the second particle,
the position of the nth particle, one worries about the value of the electromagnetic field at every spatial point and
the value of the magnetic field at every spatial point.

That is to say instead of having qa(t) one has a set of fields ϕ a(x, t). I make no assumptions about their
Lorentz transformation properties or even about the Lorentz invariance of the theory at this moment; I will shortly.
These fields may be components of vectors or scalars or tensors or spinors or whatever; I don’t care about that
right now. The important thing to remember in the analogy is that in going from the first box to the third, it is not
that t is analogous to the quadruplets xi and t, but that a, the index that labels the variables, is analogous to a and
x.

It is sometimes a handy mnemonic to think of x = (t, x) as a generalization of t, but that is not the right way to
think about it. For example we are used to giving initial value data at a fixed time t in classical particle mechanics.
In classical field theory, we need initial value data not at some fixed time t and some x, but at a fixed time t and all
x. That x is continuous is in fact irrelevant because if I wanted to—although I shan’t—I could just as well trade
these variables for their Fourier coefficients in terms of some orthonormal basis. And then I would have a discrete
set, say harmonic oscillator wave functions, and I would have a discrete variable replacing x. The big difference is
that the index is infinite in range, not finite. I will stay with x because I presume you all know that in manipulating
functions of variables it doesn’t matter whether you use a discrete basis or a continuum basis to describe them,
whether you use harmonic oscillator wave functions or delta functions. With a discrete basis you have a
Kronecker delta and with a continuum basis you have a Dirac delta. Otherwise the rules are exactly the same.

In classical particle mechanics you have a bunch of dynamical variables qa(t) which evolve in time. In
classical field theory you have a bunch of dynamical variables ϕ a(x, t) that evolve in time interacting with each
other. In classical particle mechanics the individual dynamical variables are labeled by the discrete index a; in
classical field theory the individual dynamical variables are labeled by both the discrete index a and the continuous
index x. We can summarize the correspondence like this:

In general I have some Lagrangian that is determined by some complicated functions of the ϕ a’s at every spatial
point and their time derivatives and I just go, carrying on with the system. However I will instantly make a
simplification.

In the final analysis we are interested only in Lorentz invariant theories. If we have an action S that is the
integral of something that is local in time, it seems that it should also be the integral of something that is local in
space, because space and time are on the same footing in Lorentz transformations. Likewise since the integrand
involves time derivatives of only the first order, it should only involve first order space derivatives. Therefore we’ll
instantly limit the general framework (which I have not even written down) to the special case in which the
Lagrangian—the ordinary Lagrangian L in the sense of the first box—will be the integral over 3-space of
something called a Lagrangian density, L:

This is in general some function of ϕ 1, . . . ϕ N, some function of ∂µϕ 1, . . . ∂µϕ N and possibly some function of the
spacetime position x. We will indeed consider Lagrangians that depend explicitly on the position x when we
consider systems subject to external forces.

This is a specialization. There are of course many non-Lorentz invariant theories that follow these criteria: first
order in space and time derivatives, integral over d3x, and so on. Most of the theories that describe
hydrodynamics, a continuum system, do. Most of the theories that describe elasticity are of this form. But there are
many that do not. For example, if we consider the vibrations of an electrically charged crystal, we have to insert
the Coulomb interaction between the different parts of the crystal which is not expressible as an integral of a single
spatial density of the crystal variables; it’s a double integral involving the Coulomb Green’s function. But we will
restrict our attention to this form.

When we write down an expression of this form—whenever we have an infinite number of degrees of
freedom—we have to worry a lot about questions of convergence. I will of course behave in typical physicist slob
fashion and avoid such worry simply by ignoring these questions. But it should be stipulated that this object, L, is
well defined. It is tacitly assumed that all the ϕ’s go to zero as x goes to infinity. We will only consider
configurations of that sort. Otherwise the Lagrangian would be a divergent quantity, and everything we do would
be evidently nonsense. So without saying more about it, I will establish a rule that we assume whenever possible,
whenever necessary, that not only the ϕ’s are sufficiently differentiable so that we can do all the derivatives we
want to do, but also that they go to zero sufficiently rapidly so we can do all the integration by parts we want to do.
I leave it to mathematicians to worry about how rapid “sufficiently” is.

We define the Lagrangian L as

and the action S as

(but the time integration is limited). We can derive the Euler–Lagrange equations from this expression for the
action.

It’s useful now to treat all four coordinates as analogous to t. If the Lagrangian density is a Lorentz invariant,
the Euler–Lagrange equations will be Lorentz covariant; Lorentz invariance is now manifest. Treating the four
coordinates equally is a bad thing to do for Hamiltonian dynamics but a good thing to do for this particular
problem; it will allow us to do all four (if necessary) integrations by parts in one fell swoop. So let’s do that and
derive the Euler–Lagrange equations:

Observe that δ(∂µϕ a) equals ∂µ(δϕ a). I can now perform integration by parts. In the space derivative, the space
boundary term vanishes by my assumption that everything goes to zero at spatial infinity. In the time derivative,
the time boundary term vanishes not from that assumption but from the universal condition attached to Hamilton’s
Principle, that I only consider variations that are zero at the initial and final times; δϕ a(x, t1) = δϕ a(x, t2) = 0. Then
Following closely upon my development in particle mechanics I will simply define an entity called πaµ;

Since δϕ a is an arbitrary function (aside from going to zero at spatial infinity and at the time boundaries) I deduce,
just as in the particle mechanics case, the Euler–Lagrange equations of motion,

These are the same Euler–Lagrange equations of motion derived from the same Hamilton’s Principle as in the
other case. All that we have changed is to have had an infinite number of variables, and to have specified that the
Lagrangian depended on these variables in a rather restricted way. The quantities πaµ should not be thought of as
a 4-vector generalization of pa. The correspondence is actually

Now you may not be as familiar with these equations as with their particle mechanics analogues. So let me here
pause from my general discussion to do a specific example. Once I do that specific example maybe there won’t be
as many questions about the general discussion as there would be if I asked for questions now.

EXAMPLE. A Lagrangian density L for a single real scalar field

I want to construct a simple example. Well, first, the simplest thing I can imagine is one real scalar field, ϕ(x)
= ϕ ∗(x), instead of a whole bunch of fields. Secondly, simple here really means that the equations of motion are
linear. That requires a Lagrangian density L quadratic in ϕ and ∂µϕ, because the equations of motion come from
differentiating the Lagrangian. I’ll assume a quadratic Lagrangian so I’ll get linear equations of motion. And, thirdly,
since I want the equations of motion eventually to be Lorentz invariant I want L to be a Lorentz scalar. That looks
like a good set of criteria for constructing a simple example. Here is the most general of the simple Lagrangians
we can construct:

Of course this determines the example completely. I’ve put a one half in front for later simplifications. There is
some unknown real coefficient a times ∂µϕ ∂µϕ. That’s the only Lorentz invariant term I can make that’s quadratic
in ∂µϕ. I can’t make anything Lorentz invariant out of ϕ and ∂µϕ. If I multiply them together I just get a vector. And
finally I can have some other coefficient b times ϕ squared, where a and b are arbitrary numbers. Now I hate to
work with more arbitrary coefficients than I need, so I will instantly make a simplification that comes from
redefining ϕ;

If we rewrite the Lagrangian in terms of ϕ′, the Lagrangian becomes

From now on I will drop the primes and just call this field ϕ. So in fact we have in this Lagrangian just two elements
of arbitrariness, an arbitrary real number (b/a), and the discrete choice about whether we choose the + sign or the
− sign. We’ll later see that this discrete choice is determined by the requirement that the energy must be positive.
That’s sort of obvious because the Hamiltonian is linearly related to the Lagrangian. So if I take minus the
Lagrangian I’ll get minus the Hamiltonian. If it’s positive in one case, it’s going to be negative in the other. And if it
is positive in no cases, if the energy cannot be bounded in either case, I wouldn’t have looked at this example!

Now let’s use our general machine. Defining

the Euler–Lagrange equations become, since ∂L/∂ϕ = ±(b/a)ϕ,

The one half is canceled by the fact that we’re differentiating squares. Or, plugging in the definition of πµ,

which is rather similar to the Klein–Gordon equation that materialized in the latter part of last lecture. This of
course is another reason why I chose this particular example.

Let us now go to the question of the Hamiltonian form. I’ll postpone the Hamilton equations of motion for a
while and just try and derive the Hamiltonian in its guise as the energy. The question is, what is the analog of p?
Well, it’s pretty obvious what the p is. You recall that one way of defining p was by a partial derivative. You could
say

the dots indicating the other term which contains no time derivative. That’s the definition of pa; it’s the thing that
multiplies d a. Now going over to functionals, there’s an unfortunate change in notation that really makes no
sense: we use a wiggly delta, δ, instead of a straight d, but of course it’s the same concept, the infinitesimal
change of the dependent variable under the infinitesimal change of the independent variable;

the dots representing terms with no time derivatives. What their explicit forms are, I don’t care. Some have
gradients and some have nothing differentiated, but they don’t have any time derivatives. Hence the thing that is
the analog of pa, in fact the thing that is pa for an infinite number of degrees of freedom, is

which is the canonical momentum. This expression is also equal to our previous πµa, (4.25), with µ set equal to
zero;

So it’s the time component of πµa which is the generalized version of the canonical momentum, sometimes called
the canonical momentum density. Parallel to this equation

is this equation

indeed, they are the same equation. In the former all the summations are absorbed into the summation
convention; in the latter half the summations are absorbed into the summation convention and the other half are
written as integrals. The expression

is called the Hamiltonian density; it’s the thing you have to integrate to get the Hamiltonian.

The fact that we obtain the Hamiltonian, the total energy in the time-independent case, as an integral over x
at fixed time is of course not surprising. To find out how much energy there is in the world, you add up the energy
in every little infinitesimal volume. Let’s apply these formulas to our simple example, (4.30);

the minus sign coming from our metric. The canonical momentum density π is the zero component of πµ =
∂L/∂(∂µϕ), so

i.e., ∂0ϕ(x, t). Therefore the Hamilton density H is

We choose the + sign, to ensure that the π2 cannot become arbitrarily large and negative; we want the energy to
be bounded below. And if we don’t want the ϕ 2 term to become arbitrarily large and negative, we had better
choose (b/a) to be less than zero, a fact that I will express by writing (b/a) as minus the square of a real number, µ;
(b/a) = −µ2. Thus our equations now have only one unknown quantity in them, the positive number µ2, if we’re to
have positive energies. Here is what we have in that case:

The equations of motion become

which is just the Klein–Gordon equation. Note that the Hamiltonian is the sum of three positive terms.

We could now go on and write down the classical Hamilton equations of motion in the general case and then
proceed to canonical quantization. However time is running on and I will do things in one fell swoop. I will describe
canonical quantization immediately. After all, this classical field is just the same as the classical particle system,
except that a runs over an infinite range symbolized by the two variables a and x. So that part about the
Correspondence Principle in the whole song and dance I gave about going from classical mechanics to quantum
mechanics should still be true. Therefore, I will now describe the “missing box”: quantum field theory.

4.4Quantum field theory

We simply write down the corresponding canonical commutators for the quantum field, just as we did to go from
classical mechanics to quantum mechanics:

We know that we should have [q, p] = iδ. Which delta? Well, for discrete indices, a Kronecker delta; for continuous
indices, a Dirac delta. The quantum Hamiltonian H is the integral

where H is a function of ϕ 1, ϕ 2, . . . (and their spatial derivatives); π1, π2, . . . , and possibly also explicitly of x and t,
though not in our simple example. But we might consider systems with external forces.

The set (4.47) is essentially the same set (4.13) we wrote down to find quantum particle mechanics. It’s not
even a generalization; the only generalization is to an infinite number of degrees of freedom. Since I never worried
about whether my sums on a were infinite or finite in all my formal manipulations, I don’t have to go through the
computations again. They are the same computations. The only change is notational. For continuous indices we
write a sum as an integral, but every operation is the same once you learn that transcription rule. The advantage
of this procedure is that it reproduces the classical field theory in the limit where classical mechanics is supposed
to be valid. There’s just a lot more p’s and q’s. Otherwise there is no difference.

Let us check this with our specific example by explicitly deriving the Heisenberg equations of motion and
seeing that they give us the Euler–Lagrange equations. I won’t bother to write down the equal time commutators
for our specific example because they are these equations (4.47) with the a’s and b’s erased, because there is
only one ϕ and there is only one π. Okay? So let’s do it with the example.

There is a universal rule (4.16) for computing the time derivative of any operator. We used that rule to compute
the Heisenberg equations of motion in the particle case. I will now use this rule to compute them for π and ϕ, just
as we computed them for p and q.
 I’ll start out with ϕ because that’s easier. I will do this in tedious detail to pay my dues so that every
subsequent such calculation I can do with lightning-like rapidity. The only thing in the Hamiltonian that ϕ does not
commute with is π. The rule says

just using the rule [a, b2] = b[a, b] + [a, b]b. This equation should be no surprise to you. It is one of the two Hamilton
equations;

Secondly I will compute (x, t) by the same universal Heisenberg equation of motion, (x, t) = i[H, π]. Now
there are two terms with which π does not commute: the gradient term and the ϕ 2 term. Let’s write things out.

We have a factor of −1 different from the previous equation, since we are now reversing the order of the
commutator of π with ϕ. The is again canceled because we’re always commuting with squares. We get

I have used the fact that the commutator of π with ∇ϕ is proportional to the gradient of the delta function, which
follows from differentiating the commutator with respect to y. The integral is also trivial, though not quite so trivial
as before, because we have to do an integration by parts. But it is one I think we can do by eye. This expression
should be (x, t). Plugging in from (4.50) to eliminate π and write a differential equation in terms of ϕ we obtain

which is of course the classical equation of motion, the Klein–Gordon equation.

Thus we have checked, in our specific example, the consistency of the procedure, and shown that the
Heisenberg equations of motion yield the classical Euler–Lagrange equations of motion, at least up to ordering
ambiguities which are rather trivial for linear equations of motion.6

Now we have obtained the Heisenberg equations of motion, the Klein–Gordon equation and the equal time
commutators for our free scalar field in two different ways. These two methods define the same system. As I said,
from here on in I could go through everything I did in the first three lectures running backwards and show that the
system defines an assembly of free, spinless Bose particles, Fock space, the whole routine. One way occupied
the first three lectures and the other took only one lecture. Actually if I had started out this way I would have had to
run over a lot of the material in the first three lectures in the opposite order so it might have taken me two and a
half lectures rather than one.

In any event we have two methods. One method is full of physical insight, I hope. I tried to put as much
physical insight into it as I could. We built the many-particle space out of the one-particle space. We knew why we
wanted to look for a field. It wasn’t because Heisenberg told us we had to look for a field. We had some physical
reasons for it. We constructed the field, we found it was unique under certain simplifying assumptions, we
deduced its properties and then we showed everything was characterized in terms of the field. The other method
is completely free of physical insight. We have this mechanical device like a pasta machine: the canonical
quantization procedure. You feed in the dough at one end, you feed in the classical theory, and the rigatoni, the
quantum theory, comes out at the other. It’s totally mechanical. When you’re done you have a set of equations
that you hope characterizes the system but you’ve got a lot of work to do to find their physical interpretation.

Well, since I’ve characterized these two methods praising the first so much and being so pejorative about the
second, you should not be surprised when I tell you that in the remainder of the course we will use the second
method almost exclusively. The reason is very simple. The first method we could go through because we already
understood everything. It was just a system of free particles in a box or on an infinite space. We already had
access to a complete solution to the physics; we already knew the whole spectrum of the theory. If we had tried to
apply the first method to an interacting system we wouldn’t be able to get off the ground, because we would have
to know in advance the exact spectrum of the theory. Here if we want to introduce interactions in the canonical
method, at least formally, we just write ’em down. For example, here’s an interaction:
We have a free theory, L(ϕ, ∂µϕ), equation (4.44), and I’ll throw in this interaction. Better give it a minus sign so
the classical energy at least will be positive:

There it is! There is an interaction between the system’s fields, okay? We could do canonical quantization at least
formally, if there are no problems with summing over infinite numbers of variables (and in fact we’ll see there are,
but that particular nightmare lies far in our future). We get a theory that looks like it has a nice energy bounded
below, it looks Lorentz invariant, everything commutes for spacelike separations because they commute for equal
times, and the whole thing is Lorentz invariant. So it’s got all the general features we want it to have. And it looks
like particles can scatter off of each other because if we do old-fashioned Born perturbation theory, the expansion
of the interaction term will involve two annihilation operators and two creation operators. At the first order in
perturbation theory, you can go from one two-particle state to another two-particle state, two into two scattering.
At the second order, we’ll get two-particle states into four-particle states and into six-particle states: pair
production! So there it is! We may not know what it means, but at least it’s a cheap way of constructing an
interacting field theory that obeys all of our general assumptions. Of course this means there’s a lot of work to be
done. Why did I write down this interaction with a power 4 and not the power ? Well, you’ll learn why I didn’t write
down ; there’s a reason for it. But you won’t learn that till later on.7 But at least we wound up with some
equations to play with that don’t look as if they have any evident inconsistencies with the general principles of
relativity and causality. So we can begin investigating the properties of such theories. It is just such an
investigation that will occupy the next several lectures or indeed, essentially the remainder of the first term of the
course.

4.5Normal ordering

I have one more thing I want to say about the free field. Let’s do another consistency check for our system. Since
we have ϕ’s and ’s = π’s that obey the canonical commutators and obey the Klein–Gordon equation we can, as
sketched out in the last lecture, express the field in terms of annihilation and creation operators. Just as a
consistency check, let us take such an expression, plug it into this expression (4.45) for the Hamiltonian density
and see if we get the same thing, equation (2.48), for the energy as a function of annihilation and creation
operators as we found before, for the Fock space of spinless particles. Here’s the Hamiltonian again,

and let’s write ϕ(x, t) once again in terms of its Fourier expansion, equation (3.45), separating out the space and
time parts,

This defines the operators ap and a†p. Our game is to plug this expression into the Hamiltonian (recalling that π =
), do the space integral, and see if we get a familiar result. This will lead to a triple integral, but we can do some
of the integrations in very short order. Look, for example, at the first term only,

We’ll get four terms in multiplying out the a’s and a†’s, all involving exponentials like e±ix·(p±p′). The space integral
is done easily, producing a delta function in momentum,8 which allows us to do the integral over p′ quickly,

because ωp = ω−p. The other two terms in the Hamiltonian can now be done by eye,
What will I get for the Hamiltonian? I will now do this in one fell swoop having so well organized my computation:

We observe that there is a certain simplification here. For example this first term is zero, because the factor (−ω2p
+ |p|2 + µ2) is zero. Of course we could’ve checked that out on a priori grounds. We know the equations of motion
should tell us the Hamiltonian is independent of the time. If it is independent of the time it is not going to have any
factors like these time dependent exponentials. The second term has this other factor, (ω2p + |p|2 + µ2). It doesn’t
simplify so drastically but it still simplifies to 2ω2p. Therefore, we have

This is almost but not quite what we expected, (2.48):

The expression (4.62) differs from what we wanted by a constant . . . and, surprise, that constant is infinite.
Because (4.62) is of course

The result of commuting [ap, a†p] gives δ(3)(p − p) = δ3(0). It’s only the first term we want. We don’t like that
second term.

Now what can we say about this aside from making expressions of disgust? This infinity is no big deal for two
reasons. First, you can’t measure absolute energies, only differences, so it’s stupid to worry about what the zero
point energy is. This occurs even in elementary physics. We usually put interaction energies equal to zero when
particles are infinitely far apart, but for some potentials you can’t do that, and you have to choose the zero of
energy somewhere else. There was some fast talking you let me get away with at the end of last lecture, probably
because you were tired. I said: “We’ve got the equal time commutators of the Hamiltonian with the canonical
variables, the equations of motion. Because these tell you the commutators of the annihilation and creation
operators with the Hamiltonian, they determine everything except for the zero point of the energy, which we don’t
care about.” Well, that’s still true. They have determined everything except for the zero point of the energy. And if
we still want to say we don’t care about it we can say “infinite, schminfinite”; it’s just a constant, so I can drop it. I
can always put the zero of the energy wherever I want.

In general relativity, the absolute value of the energy density does matter. Einstein’s equations,

couple directly to the energy density T00. Indeed, introducing a change in the vacuum energy density, in a
covariant way like this

is just a way of changing the cosmological constant Λ, a term introduced by Einstein and repudiated by him ten
years later. No astronomer has ever observed a non-zero cosmological constant. We won’t talk about why the
cosmological constant is zero in this course. They don’t explain it in any course given at Harvard because nobody
knows why it is zero. 9

Secondly, we can see physically why the second term comes in if we think of the analogy between this
system and a harmonic oscillator. We have an infinite assembly of harmonic oscillators here but we wrote things
just as if the individual Hamiltonians were p2 + q2; we haven’t got the extra term of −1 as in (2.16). Therefore, we
get the zero point energies in the expression for the individual oscillators. And since there is an infinite number of
oscillators we get a summed infinite zero point energy. It’s doubly infinite: infinite because of δ(3)(0) and infinite
because ∫ d3pωp is infinite. Generally there are two types of infinities: infrared infinities, which disappear if we put
the world in a box (the δ(3)(0) would be replaced by the volume of the box); and ultraviolet infinities, due to
arbitrarily high frequencies. The bad term here has both types of infinities.

An alternative way of saying the same thing is that canonical quantization gives you the right answers up to
ordering ambiguities, and the only problem here is the order. I will use my freedom to get rid of ordering
ambiguities by defining those terms ordered in another way. This idea, although it sounds silly and brings
universal ridicule, is in fact a profitable way to proceed. I will therefore define an unconventional way of ordering
expressions made only out of free fields which I will call normal ordering. I’ll write down that definition and then I’ll
show you that normal ordering defines the right ordering. By the way the most significant feature of this calculation
is I’m being very cavalier about the treatment of infinite quantities. And if you think it’s bad in this lecture, just wait!

Let {ϕ a1(x1), . . . , ϕ an(xn)} be a set of free scalar fields. There may be a whole bunch of them with different
masses and so on. The normal-ordered product of the fields, indicated by colons on either side,

means that this is not to be interpreted as the ordinary product, but instead is the expression reordered with all
annihilation operators on the right and a fortiori all creation operators on the left.

That is the definition of normal ordering, of this normal ordered product of a string of free fields. I don’t have to
tell you the order of the annihilation operators because they all commute with each other. Just break every field up
into its annihilation and creation parts, and you shove all the annihilation parts on the right. If the expression
involves a sum of products, each of those terms is redefined by sticking all the annihilation operators on the right.
This seems like a dumb definition. Nevertheless, take my word for it, this concept will be very useful to us in the
sequel. This enables us to write down the proper formula for the Hamiltonian in terms of local fields:

That just tells us that whenever we run across the product of an a and an a† we put the a on the right and therefore
the adjoint a† on the left. What could be simpler? To advance this elaborate definition just to take care of what I
said in words five minutes ago may seem extremely silly to you, but we will use the normal ordered product again
and again in this course. This is the first occasion we have had to use it and so I introduced it here. The name is a
little bit bad because “normal order product” causes some students to get confused and weak in the head. They
think you start out with the ordinary product and then you apply an operation to it called normal ordering. That is
not so. This whole symbol, the string of operators and the colons, define something just as AB defines the product
of the operator A and the operator B. In particular, “normal order” should not be interpreted as a verb, because it
leads to contradictions. Suppose, for example, you attempted to normal order an equation, like this:

We don’t “normal order” equations. Normal ordering is not derived from the ordinary product any more than the
cross product is derived from the dot product.

The divergent zero-point energy is the first infinity encountered in this course. We’ll encounter more ferocious
infinities later on. We ran into this one because we asked a dumb question, a physically uninteresting question,
about an unobservable quantity. Later on we’ll have to think harder about what we’ve done wrong to get rid of
troublesome infinities.

This concludes what I wish to say about canonical quantization of the free scalar field. If we wanted to get as
quickly as possible to applications of quantum field theory, we’d develop scattering theory and perturbation theory
next. But first we are going to get some more exact results from field theory.

Next lecture we’ll go through the connection between symmetries and conservation laws. We’ll talk about
energy and momentum and angular momentum and the friends of angular momentum that come when you have
Lorentz invariance. We’ll talk about parity and time reversal, all for scalar fields. We’ll talk about internal
symmetries like isospin, found in systems of π mesons (which are scalar particles, and so within our domain). And
we’ll talk about the discrete internal symmetries like charge conjugation and so on, all on the level of formal
classical field theory made quantum by canonical quantization.
1 [Eds.]In this book, a functional F[f] is a function F of a function f, mapping f to a number, real or complex, and will
be realized by an integral.
2 [Eds.] See Ch. IX, §100 in Edmund T. Whittaker, A Treatise on the Analytical Dynamics of Particles and Rigid
Bodies, Cambridge U. P., 1959. Maupertuis’ action, introduced in 1744, is the integral ∫ (∂L/∂ )dt. Whittaker
says Maupertuis’ Principle of Least Action was actually established by Euler. Lagrange’s equations were
introduced in his Mécanique analytique in 1788. For Hamilton’s introduction of his equations see W. R. Hamilton,
“On the application to dynamics of a general mathematical method previously applied to optics”, Report of the
British Association for the Advancement of Science, 4th meeting, 1834, pp. 513–518. Lanczos, citing Cayley, says
that Lagrange and Cauchy anticipated Hamilton; see Cornelius Lanczos, The Variational Principles of Mechanics,
4th ed., University of Toronto Press, 1970, p. 168. See also Whittaker, op. cit., Ch. X, §109 and Arthur Cayley,
“Report on the Recent Progress of Theoretical Dynamics”, in his Collected Papers, Cambridge U. P., 1890, v. III,
pp. 156–204 for further references. What is now universally called “the action” was originally called “Hamilton’s
first principal function”. See v. II, Lecture 19, p. 19-8 in Richard P. Feynman, Robert B. Leighton and Matthew
Sands, The Feynman Lectures on Physics (the New Millennium edition), Basic Books, 2010.
3 [Eds.] Goldstein et al. CM, Section 8.1, pp. 334–338.
4 [Eds.] Goldstein et al. CM, Section 2.4, pp. 45–50.
5 [Eds.] Goldstein et al. CM, Section 9.5, pp. 388–396.
6 By the way much of the material in this lecture is covered in Chapters 11 and 12 of Bjorken and Drell, the first
two chapters of volume II, in a somewhat different way so you might want to look at that. You don’t need to look at
it but you might want to.
7 [Eds.] See §16.4.
8 [Eds.] ∫ d3x e±ip·x = (2π)3δ(3)(p)
9 [Eds.] Applied to the universe as a whole, Einstein’s equations imply that its size is not static. Einstein found this
conclusion unacceptable, and introduced Λ to keep the size fixed. Edwin Hubble’s discovery in 1929, establishing
the universe’s expansion, apparently removed the need for Λ. In his posthumously published autobiography,
George Gamow wrote “Much later, when I was discussing cosmological problems with Einstein, he remarked that
the introduction of the cosmological constant was the biggest blunder of his life.” (G. Gamow, My World Line,
Viking, 1970, p. 44.) Gamow’s account seems to be the only record of Einstein’s repudiation of Λ. But things are
not so simple. In 1998, two teams measuring supernova distances discovered that the expansion of the universe
is accelerating, consistent with Λ > 0. For this discovery Saul Perlmutter, Adam Riess, and Brian P. Schmidt were
awarded the Nobel Prize in 2011. The observational value on Λ is, in “natural” units where G = h = c = 1, on the
order of (10−3eV)4: A. Zee, Einstein’s Gravity in a Nutshell, Princeton U. P. 2008, p. 359; PDG 2016, p. 349 quotes
a value for ρΛ = (2.3 × 10−3eV)4. (In natural units, 1 eV = 1.76 × 10−36 kg, 1 eV −1 = 1.97 × 10−7 m, and in
conventional units, Λ = (8πG/c4)ρΛ ∼ 10−69 s2/m4 ∼ 10−52 m−2.)

5
Symmetries and conservation laws I. Spacetime symmetries

Last lecture we discussed canonical quantization and how it established correspondences between classical field
theories and quantum field theories. We also talked about how those correspondences had to be taken cum grano
salis because they included ordering ambiguities. At the last moment we had to check to make sure that we could
order things in such a way that everything went through all right.

Today I would like to begin a sequence of lectures that will exploit that correspondence by studying the
connection in classical physics between symmetries and conservation laws, and extending that to quantum
physics by the canonical quantization procedure. We will thus obtain explicit expressions for objects like the
momentum or the angular momentum, et cetera, in field theory, even including for example interactions like λϕ 4.
Of course, these expressions we find will also have to be taken with a grain of salt. We always have to check that
we can make sense out of them by appropriately ordering things and we will do that check first. We will begin with
typical cases for the free field theory.
 Having cleared my conscience by telling you that nothing is to be trusted, I will now conduct the entire lecture
as if everything can be trusted, without worrying about fine points.

5.1Symmetries and conservation laws in classical particle mechanics

As always I will begin with classical particle mechanics and consider a general Lagrangian involving a set of
dynamical variables and their time derivatives, and perhaps explicitly the time,

I would like to consider some one-parameter family of continuous transformations on these dynamical variables. I
will assume for every real number λ I have defined some transformation

that turns the old motion of the system into some new motion parameterized by the number λ. I will always
assume we have chosen the zero of λ such that qa(t, 0) = qa(t).

As a specific example let’s consider a transformation for a particular class of systems, an assembly of point
particles, say. I’ll give them different masses. The Lagrangian is

That’s the conventional kinetic energy, plus some potential energy V(r,s) depending only on the differences
between the positions xr and xs of the rth and sth particles, respectively. The sort of transformation I want to
consider for this system is a spatial translation along some particular direction, to wit, the transformation

where e is some fixed vector. I translate all the particles by an amount λ along the direction e. Other examples of
one-parameter families of transformations which we frequently find it profitable to consider in classical mechanics
are time translations, rotations about a fixed axis and Lorentz transformations in a fixed direction. We will talk
about all of these, and others, in the course of time.

Now we return to the general case. It will be convenient to study infinitesimal transforma- tions,

qa goes into qa plus an object I will call Dqa times dλ, the infinitesimal change in the parameter λ, where

If I know how qa transforms I know how a transforms, since it is just the time derivative of qa. Thus D a, the
infinitesimal change of a, defined in the same
way, is d/dt of Dqa, as
we see just by differentiating (5.4) with
respect to t; λ is a constant and t-independent. We also know how the Lagrangian transforms:

We will always call the expression ∂L/∂ a the canonical momentum, pa. Similarly we know how any function of qa’s
and a’s transforms under either the finite or the infinitesimal version of the transformation.

Definition. We will call a transformation a symmetry if and only if

for some function F (qa, a, t). This equality must hold for arbitrary functions qa(t), which need not satisfy the
equations of motion.

Most transformations are not symmetries. Why do I adopt such a peculiar definition? Well, our intuitive idea of
a symmetry is that a symmetry is a transformation that does not affect the dynamics. When we say a theory is
invariant under, say, a space translation, we mean if we take a motion picture of the system, and if we then space
translate the initial conditions, we get the space translated motion picture. Certainly this would be true if the
Lagrangian were unchanged by this transformation. But it could also be true if the change DL in the Lagrangian
were of the form dF/dt, because a change of this form simply adds a boundary term to the action integral. And as
we saw in our derivation of the Euler–Lagrange equations we can add boundary terms to the action integral at will
without affecting the form of the Euler–Lagrange equations. Explicitly,1

Since S′ and S differ only by a quantity which equals zero on variation, the conditions δS′ = 0 and δS = 0 give
equations of motion with the same form.

Whenever one has such an infinitesimal symmetry (in a Lagrangian) one has a conservation law. This
amazing general theorem which I will now prove is called Noether’s Theorem.2 In fact the proof is practically
already done.

I will prove it by explicitly constructing a formula for the conserved quantity, a function of the qa’s and a’s
which as a consequence of the Euler–Lagrange equations is independent of time. I will call this conserved
quantity Q, in general; Q for “quantity”:

This is a universal definition (notice we are using the summation convention). I will now show that this quantity is
independent of time:

Now I will use the Euler–Lagrange equation, which tells us that a = ∂L/∂qa. We have two expressions for DL. The
first one, (5.6), tells us that the sum of the first two terms in (5.10) is DL. The definition of a symmetry, (5.7), tells us
that the last term in (5.10) is −DL. Therefore, the sum of the three terms is equal to zero, and dQ/dt = 0.

So this equation, (5.9), is the magic, universal formula. Given a one-parameter family of symmetries, (5.4),
first you extract an infinitesimal symmetry, (5.5), and then from the infinitesimal symmetry you extract a
conservation law. (There is no guarantee that Q ≠ 0, or that for each independent symmetry we’ll get another
independent Q. In fact, the construction fails to produce a Q for gauge symmetries.3 The rules are universal and of
general applicability. I will give three examples.)

EXAMPLE 1. For the Lagrangian (5.2), space translation of all the particles through a fixed vector, e:

The Lagrangian is unchanged under these translations because V depends only on the differences between
positions, and all are translated by the same amount, λe. F of course for this particular example is zero, because
the Lagrangian is unchanged under these translations, and therefore DL = dF/dt = 0. From (5.9), the conserved
quantity is

This quantity Q is the sum of the canonical momenta pr dotted with e, the change in the corresponding coordinate.
By this method we obtain an infinite number of conservation laws, for there are an infinite number of choices of e.
But in fact they can all be written as a linear combination of three linearly independent conservation laws which we
obtain by taking e to be the unit vector along each coordinate axis, and therefore we actually obtain only three
conservation laws,

where

This expression is not peculiar to the Lagrangian (5.2). Whenever we have a Lagrangian from which we get
conserved quantities from spatial translation invariance, whether or not the system looks anything like a collection
of point particles, we’ll call the conserved quantity the momentum, p. The expression (5.14) for the momentum
would not be so simple if the Lagrangian contained velocity dependent forces, but the conservation laws would
nevertheless exist.

EXAMPLE 2. A general Lagrangian L(qa, a) where I only assume that it is independent of the time: ∂L/∂t = 0. Look
at time translation:

The only time dependence in the Lagrangian is that through the qa’s and their time derivatives. Therefore F equals
L, because F is that which when differentiated with respect to time gives you the change in the Lagrangian. The
conserved quantity is (summing over a)

Whenever we get a conserved quantity from time translation invariance, we’ll call the conserved quantity the
energy, E. It is related to time translation as the momenta are related to space translations. It is also sometimes
called the Hamiltonian, H, when written as a function of the p’s and the q’s. I’m sure this is familiar material to
those of you who have taken a standard undergraduate mechanics course.

EXAMPLE 3. Again using the Lagrangian (5.2), consider a rotation about an axis e through an angle λ:

This Lagrangian is rotationally invariant, so DL = 0 and F = 0, as in Example 1. The conserved quantity is

Again, taking e to be a unit vector along a coordinate axis, we obtain three conservation laws, one for each
component of angular momentum, J. Whenever we get conserved quantities from rotational invariance, we’ll call
the conserved quantities the angular momentum.

There is nothing here that was not already in the Euler–Lagrange equations. What Noether’s theorem
provides us with is a “turn the crank” method for obtaining conservation laws from a variety of theories. Before this
theorem, the existence of conserved quantities, like the energy, had to be noticed from the equations of motion in
each new theory. Noether’s theorem organizes conservation laws. It explains, for example, why a variety of
theories, including ones with velocity-dependent potentials, all have a conserved Hamiltonian, or energy, as in
Example 2.

5.2Extension to quantum particle mechanics

Now when we quantize the theory, when we engage in canonical quantization, an amusing extra feature appears.
I will state a theorem which I will not prove, or more properly, will prove only for a restricted class of theories. Most
of the cases we will consider will belong to this class. When we come to one that does not fall under the restriction
we will check the theorem by explicit computation.

In the quantum theory there is a peculiar closing of the circle. In classical mechanics and in quantum
mechanics modulo4 ordering ambiguities, whenever we have an infinitesimal symmetry we have a conservation
law, a conserved quantity. In quantum theory the circle closes: We can use the conserved quantity to re-create the
infinitesimal symmetry. Specifically,

That is to say, the conserved quantity Q is the generator of the infinitesimal transformation, something in fact we
have already exploited in our general discussions for the components of the energy and momentum. This is
obviously true if both Dqa and F are independent of a, because in that case the only term in Q (defined in (5.9))
that does not commute with qa is pa and the commutator manifestly gives the desired result:

It is not so obvious that (5.19) holds if Dqa or F involve the a’s. It is nevertheless true but I don’t want to go

through the trouble of proving the general result. We have up to now seen one case where it is not obviously true.
That one case is time translation, where Dqa does involve the a’s and so does F. But the equation is
nevertheless true because in that case Q is the Hamiltonian, and (5.19) is the Heisenberg equation of motion:

I have gone fast because I presume this material is mainly familiar to you.5

5.3Extension to field theory

So much for classical particle mechanics and quantum particle mechanics. We now turn to classical field theory.
As with the special class of classical field theories I discussed last lecture, I have a Lagrangian density that
depends on a set of fields ϕ a, their derivatives ∂µϕ a, and perhaps explicitly on the spacetime location xµ. I will
construct my notation in such a way that when things become relativistic, the notation will be right for the
relativistic case, but I will not assume Lorentz invariance until I tell you we now assume Lorentz invariance. So do
not be misled by the appearance of upper and lower indices and things like that, into thinking I’m assuming
Lorentz invariance at a stage when I’m not.

Now in one sense there is no work to be done because our only general formula, (5.9), goes through without
alteration. It’s just that instead of a sum on a discrete index we have a sum on a discrete index and an integral on
a continuous index. In another sense however we get extra information because the dynamics are so very special,
because the Lagrangian is obtained by integrating a local density point by point in space. And we will see not only
a global conservation law that tells us the total quantity of Q is unchanged, we will also be able to localize the
amount of Q and see Q flowing from one part of space to another part of space in such a way that the total
quantity of Q is unchanged. That’s a feature of the special structure of the class of theories we are looking at, that
the Lagrangian is obtained by integrating a local function of the fields. We can see these extra features in
electromagnetism.

Electromagnetism possesses a conserved quantity Q, the charge, the integral of the charge density ρ:

There is also a current density, j, and a much stronger statement of charge conservation than . Local
charge conservation says

Integrate this equation over any volume V with boundary S to get

using Gauss’s theorem. This equation says that you can see the charge change in any volume by watching the
current flowing out of the volume. Imagine two stationary, opposite charges separated in space, suddenly winking
out of existence at some time t′ with nothing happening anywhere else, as in Figure 5.1. You can’t have this. This
picture satisfies global charge

Figure 5.1: Two charges winking out of existence

conservation, but violates local charge conservation. You have to be able to account for the change in charge in
any volume, and there would have to be a flow of current in between the two charges. Even if there were not a
current and a local conservation law, we could invoke Lorentz invariance to show this scenario is impossible. In
another frame the charges do not disappear simultaneously, and for at least a moment, global charge
conservation is violated. Field theory, which embodies the idea of local measurements, should have local
conservation laws.

Well, let’s try and just go through the same arguments in this case as we went through before. Our dynamical
variables are now a set of fields, ϕ a(x), and we consider a one-parameter set of transformations of them,

with ϕ a(x, 0) = ϕ a(x). We define as before

Definition. We consider an infinitesimal transformation a symmetry if and only if

That is to say, the change in the Lagrange density is the divergence of some four-component object Fµ(ϕ a, ∂µϕ a,
x). This equality must hold for arbitrary ϕ a(x), not necessarily satisfying the equations of motion.

This is an obvious generalization of our condition in the particle case, (5.7). The integral of the divergence
also vanishes from the action principle; the time derivative disappearing for the reasons I have stated and the
space derivative disappearing because we always assume everything goes to zero sufficiently rapidly in space so
we can integrate by parts. Of course, the F of the previous discussion can be obtained from this more general
expression. Consider the change in the Lagrangian, L,

The space derivatives disappear by integration by parts, and the time derivative can be pulled out of the integral.
So the F of our previous discussion, (5.7), exists in this case and it is simply the space integral of F0,

As in (5.8), the variation in the action results in boundary terms which can be discarded,

Thus a symmetry transformation does not affect the equations of motion (we consider only variations that vanish
at the endpoints when deriving the equations of motion). So the previous case, classical mechanics, is a special
case of the general theory. However we can do more, as I announced earlier. Let me do that “more” now by
following a path parallel to the earlier discussion leading up to (5.6).

I will compute DL for a field theory;6

(The quantities πaµ were defined in (4.25); the µ = 0 components are the canonical momenta.) Parallel to the
earlier discussion I will define a four-component object which I will call Jµ,

(This is not necessarily a Lorentz 4-vector, because I’m not making any assumptions about Lorentz transformation
properties.) There is an obvious parallelism between the definition (5.9) of a global object, Q, and this definition
(5.27) of the four local objects, the four components of Jµ.

I will now show that the Euler–Lagrange equations of motion imply something interesting about the
divergence ∂µJµ of this object, Jµ:
By the Euler–Lagrange equations of motion,

and everything else I will copy down unchanged,

Just as before we have two expressions for DL. One of them, (5.26), is the sum of the first two terms in (5.28). The
other one occurs in the definition of Fµ, (5.22). So we get

Thus we arrive at Noether’s Theorem applied to field theory: For every infinitesimal symmetry of this special type,
(5.22)—this is a specialization of our previous formalism, just as this formula, (5.27) is a specialization of our
previous formalism—we obtain something that we can call a conserved current. I will explain what that means in
a moment.

Now for the physical interpretation of this. I will define J0 as the density of stuff. What the stuff is depends on
what symmetry we are considering. I will call J, the space part of this, the current of stuff. I will now show that the
words I have attached to these objects, density for J0 and current for J, have a simple and direct physical
interpretation involving stuff flowing around through space in the course of time.

Let me take any ordinary volume V in space—not in spacetime—which has a surface S, as shown in Figure
5.2. The equation (5.29) we have derived tells us

Integrating this equation over the volume V, I find

Figure 5.2: A volume V, its surface S, and a unit normal

by Gauss’s theorem. The last term is the integral over the surface S. The (–) sign indicates the outward pointing
normal vector , the standard Gauss’s theorem notation, dotted into J. This equation verifies the
interpretation I have given you, because it says if I take any volume that’s got a certain amount of stuff in it, the net
amount of stuff changes with time depending on how much stuff is flowing out of the boundaries. Notice that the
signs are right: If J is pointing outwards that means stuff is flowing out, so this time derivative is negative and
indeed there in (5.30) is the minus sign.

Of course this means, since stuff only leaves one volume in order to appear in an adjacent volume, that the
total quantity of stuff is conserved, assuming of course that everything will go smoothly to zero at infinity so we
don’t have a current at infinity. Then

So Q is independent of time. This is in fact just our general result again. Remember our definition of J0. Then Q is
the integral of J0:

(Notice that πa0 is just the thing we previously called πa, the conjugate momentum density, and the integral of F0 is
the previous F.)

This is just our previous formula, (5.9). The total conserved quantity is pDq summed over everything which in
this case means both summed and integrated, minus the quantity F. So of course the general case contains all
the consequences of the special case, which is what you would expect for special cases and general cases. But it
contains more: Not only do we have a global conservation law that the total quantity of stuff is unchanged, we
have a local conservation law that tells us we can watch stuff floating around, J, and we have localized stuff, J0.
But there is a subtlety we need to address.

5.4Conserved currents are not uniquely defined

Let’s gather our basic equations,

Okay. There in summary is everything we’ve done until now.

There is, even in classical physics, a certain ambiguity present in the definition of the stuff Q, the current Jµ
and the object Fµ whose divergence is the change DL in the Lagrange density. The reason is this. Suppose I
redefine Fµ by adding to it the divergence of some object Aµν, where all I say about Aµν is that it is antisymmetric:

We defined Fµ through its divergence, ∂µFµ; we have not defined Fµ itself. Under (5.33) the divergence itself goes
as

Now ∂µ and ∂ν commute with each other, and Aµν is antisymmetric, so

So our new Fµ satisfies the defining equation just as well as our old Fµ. However this changes the definition (5.27)
of the current Jµ:

because we’ve added something to Fµ and therefore we’ve subtracted something from the current. So we have
another definition of the current that is just as good as our old definition, in terms of local density of stuff and the
flow of stuff. On the other hand, I didn’t call your attention to any such ambiguity in particle theory and indeed there
was none. So we would expect that the definition of the total charge is unchanged. Let’s verify that. Our charge
transforms under (5.33) like this:

Why did I only write ∂iA0i, instead of ∂νA0ν? Shouldn’t I have ∂0A00 in addition? Well, yes, but A00 is zero, because
Aµν is antisymmetric.

Now, the second term of (5.37) is a space integral of a space derivative and therefore it equals zero by
integration by parts, assuming, as we always do, that everything goes to zero rapidly enough at infinity to enable
us to integrate by parts as many times as we want to. Therefore, although we have an infinite family of possible
definitions of the local current, this ambiguity gets washed out when we integrate J0 to obtain the total quantity of
stuff.

Some textbooks try to avoid this point, or nervously rub one foot across the other leg and natter about the best
definition or the optimum definition, or what is it that unambiguously fixes the definition of a four-component
current, Jµ. And the right answer is, of course, there’s nothing to natter about, there’s nothing to be disturbed
about. It is something to be pleased about. If we have many objects that satisfy desirable general criteria, then
that’s better than having just one. And in a special case when we want to add some extra criteria, then we might
be able to pick one out of this large set that satisfies, in addition to the general criteria, the special criteria we want
for our immediate purposes. If we only had one object for the current, we would be stuck. We might not be able to
make it work. The more freedom you have, the better. So, there are many of them? Good! We live with many of
them. It doesn’t affect the definition of the globally conserved quantities. It’s like being passed a plate of cookies
and someone starts arguing about which is the best cookie. They’re all edible! And when we come to particular
purposes, we may well want to redefine our currents by adding the derivative of an antisymmetric tensor to make
things look especially nice for some special purpose we may have in mind.

5.5Calculation of currents from spacetime translations

I’m now going to apply this general machinery to the particular cases of spacetime translations and Lorentz
transformations. It will just be plug-in and crank, both long and tedious, because for the spatial translations I’ll
have a lot of indices floating around, and for Lorentz transformations I will have even more indices floating around.
So I will cover the board with indices, and you will all feel nauseous, but... I gotta do the computation.

We want to apply the general formula, (5.27), first to the case where our theory is translation invariant, that is
to say where the Lagrangian density L does not depend explicitly on x, and then to the case when our theory is
Lorentz invariant, that is to say when the Lagrangian density is a Lorentz scalar.

First, we will study spacetime translations. We’ve discussed these transformations earlier for particle
mechanics. We know the globally conserved quantities we will get out of this are the momentum and the energy.
Since in field theory we always get densities as well, we will actually recover the density of energy, which we found
last lecture (the Hamiltonian density), and the density of momentum, and also obtain a current of energy showing
how energy flows and a current of momentum. The sort of transformation we wish to consider is

where eρ is some constant four-component object. I put the index ρ in the lower position just to make some later
equations look simple. The infinitesimal transformation—no assumptions about the Lorentz transformation
properties of ϕ a at this stage, they could be the components of a vector—is of course obtained by differentiating
with respect to λ at λ = 0,

which gives an expression which I will write

What we expect to get from here is a set of conserved currents that depend linearly on eρ. We have to
compute the actual coefficients of eρ using the formula (5.27). Since this is an invariance of the Lagrangian, the
currents will be eρ dotted into some object, Tρµ,

I’m using Lorentz invariant notation but I’m not assuming anything. This is just the most general linear function of
eρ. We will of course find that we get an infinite number of conservation laws this way because we have an infinite
choice of eρ’s, but we only have four linearly independent ones. Therefore we will obtain actually four conservation
laws for the four values of the index ρ. They will be of the form

because we have four independent infinitesimal transformations. That’s just ∂µJµ with the eρ factored out. The
object we will obtain in this way has a name. It is called the canonical energy-momentum tensor. It is called
“canonical” because it is what we get by plugging into our general formula. It’s called a tensor because although
we haven’t talked about Lorentz invariance, it is sort of obvious by counting indices that in a Lorentz invariant
theory it will be a tensor field.

The energy-momentum tensor is not unique. Different energy-momentum tensors may be obtained by adding
the divergence of an antisymmetric object Aρµλ:

so that θρµ, like Tρµ, has zero divergence in its last index:

The second term vanishes because ∂µ∂λ is symmetric in µ and λ, and Aρµλ is antisymmetric in those indices.
There are many different energy-momentum tensors in the literature. There’s a tensor of Belinfante, 7 there is a
tensor which I had a hand in inventing8 that is very useful to consider if you were playing with conformal
transformations, but we won’t talk about any of that. We will just talk about this one, since this is not a lecture on
the 42 different energy-momentum tensors that occur in the literature. And of course they all unambiguously
define the same conserved quantities when you integrate. These conserved quantities are called Pρ,

They are called Pρ because for space translations one gets the conservation of momentum and for time
translations one gets the conservation of energy. Those are the objects, energy and momentum, which one
normally sticks together in a single four-component object. So this is the general outline of what has to happen.
The only thing we have to do is to compute Tρµ explicitly.

Now we have the general formulas. We have Dϕ a, (5.40). The only thing we need to compute is DL , (5.26).
Well, by assumption everything is translationally invariant. The only spacetime dependence of L is via the field, so

This is not as it stands the divergence of something, it’s the gradient of something. But it’s easy enough to make it
a divergence. One simply writes this as

That’s the rule for raising indices. Note that ∂µ commutes with eρ because e is a constant vector, and with gµν
which is a constant tensor. Thus we have the object we have called Fµ, (5.22),

We can use our general formula, (5.27) to construct the conserved current,

We obtain the tensor Tρµ by factoring out eρ, (5.41),

This is the general formula for the canonical energy-momentum tensor. Notice there is no reason for it to be a
symmetric tensor. It turns out to be a symmetric tensor for simple theories, but in general we should distinguish
between the indices ρ and µ. The first term doesn’t have any obvious symmetry between ρ and µ. There is this
symmetry for the free field theory we talked about, because πaµ was just ∂µϕ a. But in general Tρµ will not be
symmetric.

The index µ plays the role of the general index in our discussion of currents. If it is a time index, 0, you get a
density. If it is a space index, any of {1, 2, 3}, you get a current. The index ρ tells you what you get a density of and
what you get a current of in each particular case. When ρ is zero, you get the density of energy or the current of
energy, depending on the value of µ. When ρ is a space index, you get the density of the space component of
momentum or the current of that space component of momentum.

Just to check that we haven’t made any errors, let us look at T00 which should be the density of energy;
density because the second index µ is zero, energy because the first index ρ is zero:

This is simply the Hamiltonian density, (4.40), which we arrived at last lecture. So indeed this is the quantity which
when integrated over all space gives you the total energy.

To make another check, let’s compute the total momentum, for a case where we know what is going on, by
integrating the density of momentum over all space. The case where we know what is going on is that of a single
free quantum field of mass µ. There is only one πµ, which I remind you is ∂µϕ, equations (4.25) and (4.44). The
density of momentum is Ti0, ρ = i because we’re looking at momentum, µ = 0 because we’re looking at a density,
and is therefore
Just to check that this is right, the total momentum P should be obtained by integrating this quantity,

The minus sign is there because Ti0 has ∂i with an upper index, and ∇ is ∂i with a lower index. When we raise the
space index we get a minus sign from the metric.

Now let’s actually evaluate this component (5.53) for the free quantum field, plugging in our famous
expression (3.45) in terms of annihilation and creation operators,

Let’s see if we get our conventional momentum, up to possible ordering trouble such as we encountered with the
Hamiltonian. This is a consistency check. Well, the calculation is almost like the calculation of the Hamiltonian at
the end of the last lecture (p. 72), and therefore we can use the same shortcuts as there.

The x integral and the (2π)3 will be killed in making two delta functions, δ(3)(p − p′) and δ(3)(p + p′). That takes
care of one p integral, say p′, and we will end up with a single p integral. I gave you a general argument last time
why the terms with two creation operators and two annihilation operators should vanish, so I won’t even bother to
compute them this time. They’ll still have coefficients that oscillate in time and therefore must go out because of
the conservation equation.9 So I’ll just compute the coefficients of apa†p and a†pap, and I get

As before with the Hamiltonian (4.62), this is not the right expression; the first term is out of order for our
convention of having the annihilation operators to the right and therefore we will commute the first term. We get

Here if I’m willing to be especially cavalier with infinities I can simply say well, this second integral in (5.57) is the
integral of an odd function of p, albeit a divergent integral with a divergent coefficient, and therefore it gives me
zero. If I’m willing to be more precise I mumble something about ordering ambiguities and say that in the quantum
theory the proper result is not the expression (5.55), but this expression,

with normal ordering. In either case we certainly have no more troubles than we have with the Hamiltonian and we
have less if you’re willing to accept that dumb argument about the integral of an odd function being zero. And we
got the right answer with the right sign. So that suggests that the formulas we have derived in the general case are
not total nonsense.

5.6Lorentz transformations, angular momentum and something else

We’ve gone through the machine for spacetime translations. Obviously the next step is the other universal
conservation law, from Lorentz transformations (including both rotations and boosts). Here there is a technical
obstacle we have to surmount. We don’t have an explicit expression for a Lorentz transformation matrix as we do
for spatial translation. It’s some 4 × 4 matrix Λ that obeys some godawful constraint.10 Therefore we can’t directly
find the infinitesimal transformation by differentiating with respect to the parameters of the Lorentz transformation
because we don’t have a parameterized form of a Lorentz matrix. I will avoid this problem by writing down the
conditions that an infinitesimal transformation be an infinitesimal Lorentz transformation, and we’ll find the
infinitesimal Lorentz transformation directly from these conditions. In the first instance Lorentz transformations are
defined as acting on spacetime points, so let us consider an infinitesimal transformation acting on a spacetime
point, and see what conditions make it a Lorentz transformation.
 So we consider the infinitesimal form of (1.15),

Now I’ve got to be very careful how I put my upper and lower indices. That is certainly the most general linear
transformation on xµ. I could have put the index ν on the s downstairs and the second ν on the x upstairs and find
the same thing, of course, but I choose to do it this way because otherwise if I had one index upstairs and one
downstairs I would go batty trying to figure out which index on s was the first index and which was the second. By
keeping both of ϵ’s indices upstairs I don’t have that problem.

A second vector, yµ, under the same transformation but lowering all the indices, goes into

This infinitesimal transformation is a Lorentz transformation if xµyµ is unchanged (1.16) for general x and y.
Substituting,

and because the transformation is infinitesimal we only retain terms to first order in dλ.

In order to compare the second term to the third, I will lower the indices on ϵµν and raise them on x and y. But
of course when I raise the coordinate indices I get the ν on the x and the µ on the y. That’s not good for
comparison, so I’ll exchange µ with ν. They’re just summation indices and it doesn’t matter what we call them.
Then we get

Now for this to be a Lorentz transformation, the sum must equal xµyµ. That’s the definition of a Lorentz
transformation, it doesn’t affect the inner product. Therefore, since x and y are perfectly general and the
coefficient of the term bilinear in y and x is ϵµν + ϵνµ I find

That is to say, ϵµν is an antisymmetric matrix. You could write ϵ with both indices upper or with both lower; either
way ϵ is an antisymmetric matrix although a different antisymmetric matrix because of the intervention of the
metric tensor. If you write it with one upper and one lower index, it’s something horribly ugly; it’s not antisymmetric
at all.

So an infinitesimal Lorentz transformation is characterized by a 4 × 4 antisymmetric matrix. Let’s just check if

this makes sense when counting parameters. A 4 × 4 antisymmetric matrix has 4 × (4 − 1) = 6 independent
entries. That’s just right, because there are six parameters in Lorentz transformations: three parameters to
describe the three axes about which one can rotate, and three to describe each direction in which one can perform
pure Lorentz boosts.

Let’s consider the case where

all other matrix entries zero. In that case I find from the formula (5.59)

(Raising a space index gives a minus sign.) Only x1 and x2 get changed. Equation (5.65) is the infinitesimal form
of the rotation

Notice that (5.65) is what you get by differentiating (5.66) with respect to λ and setting λ to zero, in accordance
with equations (5.4) and (5.5),

So ϵµν with non-zero components only for the indices 1 and 2 corresponds to what one usually calls a rotation
about the z-axis; x3 and x0 are of course are unchanged.

To take another example, consider

all other entries zero. Here only x0 and x1 get changed;

In the first expression, we raise the x index 1, gaining a minus sign, but there’s also a minus sign in ϵ01. In the
second, there’s no minus sign in ϵ10, and there’s no minus sign from raising the index, so

This is the infinitesimal form of

which is a Lorentz boost along the x1 direction. Please notice how the signs of the metric tensor take care of the
sign differences between finite rotations and finite Lorentz transformations, one using trigonometric functions and
the other using hyperbolic functions, just by introducing minus signs at the appropriate moment. So it all works out;
it all takes care of itself.

Now we come to the dirty work of figuring out the implications of all this for a field theory, with scalar fields
only. I have not yet written down the Lorentz transformation properties of fields other than scalars. That’s the only
thing I know how to Lorentz transform. However, just as for the case of translations we can write down some
things in general. We know we will obtain a conserved current, Jµ. We know it must be linear in s and therefore I
will write
things as

The is there to prevent double counting. Since ϵ is antisymmetric, with no loss of generality I can define Mλρµ to
be antisymmetric in the indices ρ and λ. If it had a symmetric part, that would vanish in the summation on the
indices ρ and λ. And therefore I put a 1 in here because really I’m counting twice; I’m counting M01 once when I
sum it with ϵ01, counting it again when I sum it with ϵ10. Since ϵλρ is constant and perfectly general aside from the
antisymmetry condition, I know (from (5.29)) that

Therefore I will obtain six global conservation laws,

Remember it’s µ that plays our general role here, λ and ρ are just along for the ride to multiply the ϵ which I have
factored out.11 The 4 × 4 antisymmetric tensor Mλρ0 will give us six conservation laws. Three of these should be
old friends of ours, the conservation of angular momentum. We know for example that if we look at ϵ12 we get z
rotations which lead to the conservation of the z component of angular momentum. So J12, aside from a sign or
normalization factor, should be identical with the third component of angular momentum, J23 with the first, J31 with
the second, because those are the conservation laws you get from those rotations. On the other hand the (01),
(02), (03) components of J will be new objects that will give us new conservation laws to associate with Lorentz
invariance, laws we have not previously studied. We will see what those conservation laws are at the end of this
lecture. The computation will be hairy, because I’ve got three indices to keep track of. I hope I have organized it in
such a way that it will not be too bad. But now let’s compute.

We’re only considering scalar fields, so I will study12

(Λ−1x)ρ is to be an infinitesimal Lorentz transformation, to wit

(See (5.59).) Therefore Dϕ a is obtained by expanding out to first order in dλ and dividing by
dλ,

Since I chose to write (5.72) in terms of lower indices on ϵλρ I will drop my indices and raise them again:

I know this drives some people crazy. When I was in graduate school a friend of mine, Gerry Pollack, now a
distinguished worker on noble gas crystals, once said to me, “I’m so bad at tensor analysis that whenever I raise
an index I get a hernia.” Nevertheless, you will have to acquire facility with these things, although this is about as
many indices as you will ever have to manipulate in this course.

Now by the same token, since we are assuming Lorentz invariance, that is to say, we assume L is a Lorentz
scalar,

I will choose to write this as

That is straight substitution. Now we may notice that for the special case of scalar fields, this particular
combination is one we have seen before, aside from the x. It’s simply the definition (5.50) of the canonical energy-
momentum tensor, Tρµ:

So in terms of the energy-momentum tensor, the conserved current is

This is not the end of the story; xσTρµ is not antisymmetric in σ and ρ, and the symmetric part of it is irrelevant,
since ϵρσ is antisymmetric. To construct Mρσµ, (5.72), I should antisymmetrize the product xσTρµ in σ and ρ, and
write

and therefore

I want to talk about the meaning of this. The derivation may have put you to sleep but if it didn’t, it should have
been totally straightforward, step-by-step, plug-in and crank. A lot of indices to take care of but we took care of
them.

This tensor Mσρµ is a collection of six objects labeled by the antisymmetric pair of indices σ and ρ, each of
which has four components labeled by the index µ. Each of them is respectively, depending upon the value of µ, a
current of stuff or a density of stuff. Let us compute a typical component of this thing for various values of σ and ρ
to see if the expressions for these conserved quantities are physically reasonable or physically preposterous.

Let us compute J12, (5.74). This is

Now this is a very reasonable expression for the z component of the angular momentum, which was what this
object should be. I am simply saying I have a density of the two-component of momentum P2 distributed
throughout space given by T20, and also the density of the one-component of momentum P1 given by T10. To find
the total angular momentum I just take x in a cross product with the density of momentum and integrate it: x1 times
the density of the two-component of momentum minus x2 times the density of the one-component of momentum.
That’s the normal thing you would write down for the total angular momentum of a continuum system, a fluid or a
rigid body or something like that, where you have a momentum density. More properly, I should say it’s the orbital
angular momentum, if we think quantum mechanically for a moment. And the reason for that is because we’re
considering a set of spinless particles. If we had vector or tensor or spinor fields, we would have extra terms in
Dϕ a, (5.77), that would generate extra terms in Mρσµ. These could be identified as the spin contribution to the
angular momentum, that which does not come from x × p. However I won’t bother to do that out in detail, I just
wanted to show you a particular case.

Now what about the funny components—the ones we haven’t talked about before or perhaps haven’t seen
before in a non-relativistic theory—the conserved quantities like J10?

Well, that also has a definite meaning and it is not a surprise conservation law. You might think it’s some new law,
the conservation of zilch,13 never seen before! Not true. Notice that this is a very peculiar conservation law in
comparison to the others. It explicitly involves x0, the time. We’ve never seen a conservation law explicitly
involving the time before. That however has an advantage. It means we can bring the x0 out through the integral
sign and write J10 as

Now, what does d/dt of this thing say?

You have seen the non-relativistic analog of this formula. This is simply the law of steady motion of the center-of-
mass.

For a system of point particles or a continuum system, if you recall, you define the center- of-mass as the
integral of the mass density ρ(x, t) times the position x, divided by the total mass M,

The time derivative of the center-of-mass, the velocity of the center-of-mass, is a constant, equal to the total
momentum P divided by the total mass M,

Equation (5.89) is the relativistic analog of the x1 component of that law, (5.91), multiplied by the total mass. The
only change is precisely the change you would expect if you have seen Einstein’s headstone,14 E = mc2, and
remember we’re working in units where c = 1. Instead of the mass density and the law of steady motion of the
center-of-mass, we have the energy density, T00, and therefore we have the law of steady motion of the center of
energy. The center of energy of a relativistic continuum system moves in a straight line with velocity P/E, where E
is the total energy. The x component of that law is the same as (5.89) divided by E. Therefore the three
conservation laws which we get from Lorentz transformations are not new conservation laws at all, but simply the
relativistic generalization of the old non-relativistic law of steady motion of the center-of-mass, trivially generalized
to become the law of steady motion of the center of energy. The conserved quantities Ji0 corresponding to Lorentz
boosts can be written

The quantities R i are the components of the center of energy. The Ji0 are the Lorentz partners of the components
of angular momentum, and the law of steady motion of the center of energy is the Lorentz partner of the law of the
conservation of angular momentum.

You don’t normally think of the law of steady motion of the center-of-mass (or energy) as a conservation law
because you don’t normally think of conserved quantities as explicitly involving t, but these do, and this is a
conservation law. And that’s the end of this lecture.

Next lecture we will go on and talk about less familiar symmetries and less familiar conservation laws.

1 [Eds.] See L. D. Landau and E. M. Lifshitz, Mechanics, §2, “The principle of least action”, p. 4.
2 [Eds.] The theorem was stated and proved by Emmy Noether in 1915 while helping David Hilbert with general
relativity, and published by her in 1918. See E. Nöther, “Invariante Variationsprobleme”, Nachr. d. Königs.
Gesellsch. d. Wiss. zu Göttingen, Math-phys. Klasse (1918) 235–257. English translation by M. A. Tavel,
“Invariant Variation Problem”, Transport Theory and Statistical Physics 1 (1971) 183–207, and ’ed by Frank
Y. Wang at arXiv:physics/0503066v1. See also Dwight E. Neuenschwander, Emmy Noether’s Wonderful
Theorem, rev. ed., Johns Hopkins Press, 2017.
3 [Eds.]See note 2, p. 579. Coleman may have meant to say “gauge invariance” for “gauge symmetries”. As note 2
makes clear, Coleman did not regard gauge invariance as a symmetry. On the other hand, (global) phase
invariance does lead to conserved quantities.
4 [Eds.] Slang (American?) for the prepositional “except for”, just as 5 modulo 3 = 2 (5 equals a multiple of 3,
except for the remainder of 2). This usage occurs about a dozen times in the lectures.
5 [Eds.] A question was asked: “Can you extend your remarks in the more general case, [when Dqa and F involve
the a’s] that up to ordering...” Coleman responds: “No, if you don’t worry about ordering it is true. It can be proven
formally if you don’t worry about ordering.” The student follows up: “Are there cases where there simply doesn’t
exist any ordering?” Coleman replies: “Yeah, there are cases even where this breaks down, that dQ/dt is zero.
We won’t run into any such cases but they exist. Quantum field theorists call them by the pejorative name of
anomalies. There is a whole lore about when they exist and when they don’t, there’s an elaborate theory, but it’s
on a much greater level of sophistication. We’ll talk about that. I can’t tell you the conditions under which this
general formula (5.19) is true or false in ϕ 4 theory because we don’t even know how to make sense out of ϕ 4
theory yet. We don’t know how to order the ϕ 4 term. We’ll play with it formally as if we did; and then later on when
we learn more about it we’ll see that most of the formal playing can be redeemed. But at the moment I can’t say
anything.”
6 D(∂µϕ a) is of course equivalent to ∂µ(Dϕ a).
7 [Eds.] Often called the Belinfante-Rosenfeld tensor. See F. J. Belinfante, “On the current and density of the
electric charge, the energy, the linear momentum and the angular momentum of arbitrary fields”, Physica viii
(1940) 449–474, and L. Rosenfeld, “Sur la tenseur d’impulsion-´energie”, (On the momentum-energy tensor),
Méem. Acad. Roy. Belg. Soc. 18 (1940) 1–30. This tensor is symmetric, as required by general relativity.
8 [Eds.] C. G. Callan, S. Coleman and R. Jackiw, “A New, Improved Energy Momentum Tensor”, Ann. Phys. 59
(1970) 42–73. This tensor is traceless, as required in conformally invariant theories.
9 [Eds.] The term involving two annihilation operators is . The quantity multiplying p is
manifestly even, while p is odd, and so the integral vanishes. The same argument applies to the term involving
two creation operators.
10 [Eds.] In matrix terms, ΛT gΛ = g, or in components, .
11[Eds.] When Coleman says that a tensor Tλρ···µ is conserved, he means ∂µTλρ···µ = 0. He always puts a
conserved 4-tensor’s conserved index farthest to the right, in the last position, and always denotes this index as µ.
These conventions, particularly the first, are unusual.
12 [Eds.] This looks strange, but it’s correct. Under Λ: xµ → x′µ = Λµνxν, the transformation induced in a field ϕ is
ϕ(x) → ϕ′(x′) = S(Λ)ϕ(Λ−1x), where S(Λ) is a matrix depending on the tensorial character of ϕ. For a scalar, S(Λ)
equals 1. See (3.16).
13 [Eds.] Although it looks Yiddish, “zilch” (“nothing, zero”) apparently derives from a fictional insignificant person
(in Yiddish, a nebbish; see Rosten Joys, p. 387), “Mr. Zilch”, who appears in a 1920s-era comic magazine,
Ballyhoo. Coleman seems to be using it as a generic synonym for some unimportant quantity. This usage appears
a few times in this book.
14 [Eds.]Coleman is joking. Einstein has neither a grave nor a headstone. His body was cremated and the ashes
scattered in an unknown location, as he wished. On the other hand, Boltzmann’s headstone (in Vienna) has S = k
log W on it. See also note 33, p. 749.

Problems 2

2.1 Even though we have set ħ = c = 1, we can still do dimensional analysis, because we still have one unit left,
mass (or 1/length). In d space-time dimensions (1 time and d − 1 space), what is the dimension (in mass units) of
a canonical free scalar field, ϕ? (Work it out from the equal-time commutation relations.) Still in d dimensions, the
Lagrangian density for a scalar field with self-interactions might be of the form

What is the dimension (again in mass units) of the Lagrangian density? The action? The coefficients an? (As a
check, whatever the value of d, a2 had better have the dimensions of (mass)2.
(1997a 2.1)

2.2 Dimensional analysis can sometimes give us very quickly results that would otherwise require tedious
computations. In Problem 1.4, I defined the observable

where ϕ(x) was a free scalar field of mass µ, and a was some length. I defined the variance of A as var A = áA2ñ −
áAñ2, and I asked you to show that for small a,

and to find the coefficients α and β. (I also asked you to study things for large a, but that’s not relevant to this
problem.) If we’re working at very small distances, it’s reasonable to assume that the Compton wavelength h/µc
might as well be infinite, that is to say, we might as well replace µ by zero. In this case, the coefficient β is
completely determined by dimensional analysis.

(a) For a general dimension d (with a (d − 1) dimensional Gaussian replacing the three-dimensional one in the
definition of A(a)), find β. Check your result by showing that it reproduces the answer to Problem 1.4 for d = 4.

(b) What if instead of ϕ we had the energy density, T00? (Again, take µ = 0.)
(1997a 2.2)

2.3 In class thus far all my examples have involved scalar fields. Here’s a vector field theory for you to explore:
Consider the classical theory of a real vector field, Aµ, with dynamics defined by the Lagrangian density

Derive the Euler–Lagrange equations. Show that if we define 1

and further define two 3-vectors E and B by

then E and B obey the free (empty space) Maxwell’s equations in rationalized units (with neither 4π’s nor ϵ0’s.)
(1997a 2.4)

2.4 Use the procedure explained in Chapter 5 to construct Tµν, the energy-momentum tensor, for the theory of the
proceeding problem. This turns out to be a rather ugly object; Tµν is not equal to Tνµ and T00 is not the usual
electromagnetic energy density, (|E|2 + |B|2). However, as I explained in class, we can always construct a new
energy-momentum tensor that gives the same energy and momentum as the old one by adding the divergence of
an antisymmetric object.

Show that if we define

then, for an appropriate choice of the constant a, θνµ = θµν, and θ00 is the usual energy density, (|E|2 + |B|2).
Find this value of a.
(1997a 2.5)

1 [Eds.] This definition differs by a sign from that given in (14.1), p. 68 in Bjorken & Drell Fields.
 Solutions 2

2.1 As in Lecture 1, define

and let [A] denote the units of the quantity A. Then

We also have

Since

for any (integer) power n, and [dnx] = [L]n, it follows

Following the hint, consider the equal-time commutator (3.61),

It follows

The units of the Lagrangian density L can be deduced from the kinetic term, (∂µϕ ∂µϕ);

The action S is the integral over all space-time of the Lagrangian density, so

To find the units of an, note that all the terms of the Lagrangian density must have the same units, so

We were asked to check that the units of a2 should be equal to (mass)2, whatever the value of d. According to
(S2.7), [a2] = [M]2, independent of d. The interpretation of µ as a mass in the Klein–Gordon Lagrangian density
(4.44) is consistent with its units.

2.2 (a) The d − 1 dimensional Gaussian is just the d − 1 product of individual Gaussians, so ([a] has the units [L])

To normalize the observable in d − 1 dimensions, we have to redefine A(a) as

By definition,

and
If as before we take for small a

then

We know α is a constant and therefore independent of a, the only variable with dimensions. Consequently α has
to have no units, and so, because a has the units of [L] = [M]−1

In the solution to Problem 1.4, we found β0 = −2 for small a, which agrees with this result.

(b) The canonical energy-momentum tensor is defined by (5.50), and its component T00 is

(summation on a); this is also the Hamiltonian density H (see (4.40)). Then [T00] = [L] = [M]d = [H]. If we define

we get

If we set

then by the previous reasoning, since a has the units of [L] = [M]−1, we find βH = −2d. We note that the fluctuations
of the energy density grow more rapidly at small distances than those of the field itself.

2.3 We start with the Lagrangian density (P2.2)

The Euler–Lagrange equations are

The first term is identically zero. Using the identity

the Euler–Lagrange equations are

The quantity in the square brackets becomes, multiplying it all out, the antisymmetric Fλσ:

and so the Euler–Lagrange equations become

These represent four different equations. First, let’s look at σ = 0:

if, as the problem suggests, we call Fi0 = Ei. (Recall ∇i = ∂i = ∂/∂xi.) That is Gauss’s Law in empty space. Now
consider σ = i, in particular, let’s say σ = 1. Then
(the term ∂1 F11 is identically zero, since F11 = 0). Following the identification in the original problem, B1 = F32, B2
= F13, and B3 = F21, and using the antisymmetry of Fµν, we have F01 = −E1 = −Ex. Then this equation (S2.15)
becomes

which is the x component of Ampère’s Law. Similarly, i = 2 and i = 3 are the y and z components of Ampère’s Law.

The identification of the components of Fµν with the electric and magnetic fields is an easy consequence of
identifying the 4-vector Aµ with a four-component object (ϕ, A), the electric potential ϕ and the magnetic vector
potential A. Then

The Euler–Lagrange equations give half of Maxwell’s equations, Ampère’s Law and Gauss’s Law, but not the
other half. Those can be obtained from the Bianchi identities,

which follow easily from the definition of Fµν as a sort of four-dimensional “curl” of the 4-vector Aµ. The Bianchi
identities are non-zero only when {λ, µ, ν} are all different, so there are only four non-vanishing components. Let
one of the indices be zero, and the other two be {1, 2}. Then (recall ∂i = −∇i)

which is the z component of Faraday’s Law, ∇ × E = −∂B/∂t. The set {0, i, j} give all three components of Faraday’s
Law. If none of the indices are zero, there is only one non-vanishing component,

the last of Maxwell’s equations.

2.4 Using the results of Problem 2.3, we have from the definition of the canonical energy-momentum tensor (5.50)

The first term is not symmetric in {µ, ν}. Following the suggested prescription, we add the divergence of an
antisymmetric tensor,

We already know from (5.44) that ∂µ θνµ = 0. We need to determine the value for a so that θµν is symmetric.

Because of the boxed Euler–Lagrange equations (S2.13) above, ∂λ Fµλ = 0, so

and the new tensor becomes

The other problem with T00 is that it fails to give the correct energy density for Maxwell’s theory. What about θ00?
Let’s see:
If we choose a = 1, the term in the parentheses is just Fνσ, and the resulting tensor is symmetric:

as desired.
 6
Symmetries and conservation laws II. Internal symmetries

I would like to continue the discussion of symmetries and conservation laws that we began last lecture by
considering a new class of continuous transformations. From these we will extract the associated conserved
currents and the associated global conservation laws, like conservation of electric charge, conservation of baryon
number and conservation of lepton number which we have not yet considered in detail. This new class of
symmetries is not universal; they occur only in specific theories whose Lagrangians1 have special properties. We
believe on good experimental grounds that if we attempt to explain the world with a field theory, that theory had
better be translationally invariant and Lorentz invariant. Those symmetries led to the conservation of Pµ and Jµν.
However, some field theories which people have invented to understand the world turn out to have larger groups
of symmetries than just those associated with the Poincaré group. These symmetries commute with spacetime
translations and with Lorentz transformations, and so we expect that the conserved quantities Q associated with
them will be Lorentz scalars. These new symmetries are given the somewhat deceptive name of internal
symmetries. “Internal” historically meant that somehow you were doing something to the interior structure of the
particle; you were not moving it about in space or rotating it. The word is deceptive because, as you will see, it
applies to theories of structureless particles, in particular, to free field theories. Nevertheless the nomenclature is
standard, and I will use it. For us, “internal” will mean “non-geometrical”. These internal symmetries will not relate
fields at different spacetime points, but only transform fields at the same spacetime point into one another.
Conservation laws are the best guide for looking for theories that actually describe the world, because the
existence of a conservation law is a qualitative fact that greatly constrains the form of the Lagrangian.

6.1Continuous symmetries

EXAMPLE 1. SO(2)

As a simple example of a theory that possesses an internal symmetry, let me take a theory involving a set of
scalar fields, all of them free and all of them with the same mass and—this is the simplest nontrivial case—I will let
the index a range over only two values, a = {1, 2}. The Lagrangian is

So this is simply a sum of two free Lagrangians, each of them for a free scalar field of mass µ.

Now this Lagrangian possesses a rather obvious symmetry. Since everything involves the quadratic form
ϕ aϕ a, it
is invariant under a group that is isomorphic to the two-dimensional rotation group of Euclidean geometry,
SO(2). This will describe not two-dimensional rotations in the x-y plane, or in the y-z plane but in the 1-2 plane
between the fields ϕ 1 and ϕ 2. To be specific, for any λ, if I make the transformation

the Lagrangian is obviously unchanged.

This is a symmetry of this particular sample Lagrangian. It is not connected in any way with geometry; it’s not
a spatial translation and it’s not a Lorentz transformation. I could write more complicated Lagrangians which
possess the same symmetry. For example, I could add to this any power of ϕ aϕ a times some negative constant,
(negative, so it will come out with a positive sign in the energy), like the quadratic

or a term in ϕ aϕ acubed or to the fifth power. The new Lagrangian would still be invariant under this transformation
because ϕ aϕ a is invariant under this transformation: the sum of the squares is preserved by rotations.

Now let us extract the consequences of this symmetry. Let’s feed it into our general machinery, turn the crank
and see what happens. In terms of the general formula (5.21),

We need the derivatives (4.31),

We also need the four-component object I called Fµ last time, defined by (5.22),

This Lagrangian is unchanged, so Fµ = 0. We construct the current by our general formula, (5.27),

This is the formal classical expression. As we see, this current is conserved, ∂µJµ = 0, because both ϕ 1 and ϕ 2
satisfy the Klein–Gordon equation with the same mass µ. We will later investigate whether or not the formal
expression has to be normal ordered, in the case when g = 0. What we have to do to make sense of the theory
when g is not equal to zero is a subject we will investigate much later in the course. The associated conserved
quantity Q is the integral of the zero component of this current.

Let’s compute Q in the case where g = 0. And I remind you once again of our expression (3.45) for the free
fields,

I should really have a draftsman write this formula on a piece of cardboard which I could nail up above the
blackboard. The creation and annihilation operators satisfy the relations

We compute Q by our usual tricks. It’s exactly the same calculation as the others we have done (e.g., the
calculation of P, (5.53) through (5.58)),

Once again there’s no need to keep track of the product of two annihilation operators or two creation operators.
On a priori grounds these products must vanish because their coefficients involve oscillating factors that have no
hope of canceling, and Q is supposed to be time independent. I have written it already in normal ordered form, not
that it matters here. There’s no need to worry about the order of the operators because a type 1 operator and a
type 2 operator always commute.

The expression (6.11) for the charge is very nice. It has all the properties you would expect for an internal
symmetry. It commutes with the energy, (2.48); it commutes with the momentum, (2.49); and it annihilates the
vacuum:

And as we’ll see shortly (§6.2), it is also Lorentz invariant, because it is the space integral of the time component of
a conserved current (6.7). The expression is nice, however it is hardly transparent. On the other hand, the charge
Q is not diagonal with respect to the operators ap(a) and {ap(b)†}:

(where ϵ12 = −ϵ21 = 1, ϵ11 = ϵ22 = 0). The first term in the integrand (6.11) replaces a type 2 particle with a type 1
particle; the second term acts vice versa with 2 replacing 1.
(the is there for a reason that will become clear shortly). Likewise I will define cp and cp† as the other obvious
combinations,

One can make things much simpler by defining new annihilation and creation operators which are linear
combinations of our original ap(a) and ap(b)†. We will define bp and bp† as

These are also annihilation and creation operators. They create particles in states that are linear combinations of
state 1 and state 2. It is easy to check that they, too, obey the commutators for annihilation and creation
operators. All the commutators vanish except for

I inserted the in the denominators so that this would come out equal to δ(3)(p − p′) rather than twice that. If it is
not obvious to you that all the other commutators are zero, let me show you. Any annihilation operator, bp or cp,
commutes with any other annihilation operator, since both of these are linear combinations of commuting
operators. For the same reason, any creation operator, bp† or cp†, commutes with any other creation operator. So
let’s check the annihilation operator bp with the creation operator cp†,

The other combination also commutes:

The b’s and c’s obey the same algebra as the a(1)’s and a(2)’s because, for any given value of p, the b’s and c’s
are annihilation and creation operators for orthogonal single particle states. There are two states which we called,
arbitrarily, the type 1 meson and the type 2 meson. Whenever we have a degenerate subspace of states, we are
perfectly free to choose a different orthogonal linear combination to be our basis vectors. Here we have chosen a
linear combination of a type 1 meson with a type 2 meson to be a b-type meson, and the orthogonal linear
combination to be a c-type meson. If we have a Fock space with two degenerate kinds of particles–with the same
mass, µ–it doesn’t matter which two independent vectors we choose to be our fundamental mesons.

Why do I choose these combinations, (6.14) and (6.15)? I could just as well have chosen the coefficients of
ap and ap† to be sin θ and cos θ for the b’s, and cos θ and − sin θ for the b†’s and so on; the algebra would have
worked out the same. Well, I choose these combinations because both the expression of the charge Q and the
algebra of Q with the b’s and c’s work out particularly simply.

By substitution, you can see pretty easily that

where N b and N c are the number of b-type and c-type mesons, respectively, in a given state (see (2.50)). Then
you will be able to check by eyeball that unlike the type 1 and type 2 mesons, the b and c type mesons are
eigenstates of Q; Q is diagonal with respect to these mesons:

Thus we have a much simpler interpretation of Q. We have diagonalized Q by writing this expression (6.19),
and made it easy to see what basis vectors diagonalize Q. As a result, we have two kinds of particles with different
values of Q. The value of Q does not depend on the momentum of the particle. A b type, whatever its momentum,
carries a value Q = +1, and the other, the c type, carries a value of Q = −1. The two kinds are like particles and
antiparticles, with the same mass but opposite charge. We see in (6.19) that Q is simply N b minus N c . This is very
similar to electric charge. These particles for example could be π+ and π− mesons, and Q could be the electric
charge. The total charge of the system is obtained by counting the number of particles of one kind and subtracting
the number of particles of the other kind. For this reason I called this Q “charge”, but we haven’t deduced the
conservation of electric charge or anything like that. I have simply cooked up an arbitrary example with a
symmetry leading to a conservation law that has some structural resemblance to the conservation of electric
charge. I said “π+ and π− mesons”, but I could just as well have said “electrons and positrons”, aside from the fact
that electrons and positrons have spin. Q needn’t be electric charge. If we were considering electrons and
positrons, I could have let Q be lepton number instead of electric charge. Lepton number also has this kind of
structure.

In terms of the new operators, we can write the Hamiltonian as (see (4.63))

This expression is easily obtained from the sum of the two free field Hamiltonians for ϕ 1 and ϕ 2 by substitution.
I’ve introduced these combinations of the original a(1)’s and a(2)’s to simplify the representation of the charge Q in
terms of annihilation and creation operators.

Aside: Complex fields

I would like to digress now, in a direction that really has nothing to do with symmetries. I would like to talk
about putting together two real fields to make a complex field. The simple, diagonal expression of the charge
suggests that maybe we should make this complex combination not just on the level of the annihilation and
creation operators, but on the level of the fields themselves. That might make things look even simpler. Therefore
let me define a new field ψ, complex and non-Hermitian, and its adjoint, ψ*,

Properly I should write ψ † for the adjoint, but the star (∗) is traditionally used for this purpose in the literature. In
terms of creation and annihilation operators, ψ and ψ* are written

Like every free field, this ψ is an operator that can both annihilate and create. It has a definite charge changing
property. It always lowers the charge by 1, either by annihilating a b particle with charge +1 or by creating a c
particle with charge −1. Likewise ψ* always raises the charge, either annihilating a c particle of charge −1 or
creating a b particle with charge +1.

Our old fields ϕ 1 and ϕ 2 have rather messy commutators with Q. If you commute either with Q you get the other
with some coefficient:

Note that this equation follows the general rule for charges and symmetries, (5.19),

that the conserved charge generates the transformation. The new fields ψ and ψ* have neat commutators with Q:

The new fields ψ and ψ* have very interesting equal-time commutators. The fields ϕ 1 and ϕ 2 commute with
themselves and with each other at equal-times. Because they are linear combinations of ϕ 1 and ϕ 2, ψ and ψ* also
commute with themselves and with each other at equal-times:

More interesting is ψ(x, t) with ∂0ψ(y, t). That also happens to be zero, because it will involve the commutator of bp
with cp†, and from (6.17) these commute:
The adjoint of this commutator, [ψ*(x, t), *(y, t)], involves the commutator of bp† with cp′. But from (6.18), it also
equals zero,

Indeed the only non-zero equal-time commutators are ψ(x, t) with ∂0ψ*(y, t) and ψ*(x, t) with ∂0ψ(y, t),

Of course since they are linear combinations of ϕ 1 and ϕ 2, ψ and ψ* also obey the Klein–Gordon equation,

Now why did I bother to do all this, to rewrite the theory of two scalar fields in terms of a complex field and its
conjugate? Well, to recast the Lagrangian (6.1) in terms of ψ and ψ*. We can just as well write

If we look at this theory’s structure, equations (6.28)–(6.33), and read it backwards, it looks very much as if these
are equations we could have found by doing something that, at first glance, seems extremely silly. If we had
started out with this Lagrangian, (6.34), and treated ψ and ψ* as if they were independent variables, and not in fact
each other’s complex conjugate, it would have seemed the ultimate in dumb procedure. But let’s proceed anyway.

By varying the Lagrangian with respect to ψ*, we obtain the Klein–Gordon equation for ψ,

and by varying with respect to ψ we would obtain the Klein–Gordon equation for ψ*. Treating ψ and ψ* as
independent variables, we find that the canonical momentum to ψ is ∂0ψ*,

Likewise, the canonical momentum conjugate to ψ* is ∂0ψ, expressed in the adjoint equation which I won’t bother
to write down. Canonical quantization then leads to (6.31). For the other commutators, we would find that ψ and ψ*
commute at equal times because they are q type variables, and that ψ and ∂0ψ commute at equal times because
they are the q for one variable and the p for another variable. So had we been foolish enough to write the
Lagrangian in terms of complex fields to begin with, and to treat ψ and ψ* as if they were independent, we would
have obtained, in this particular instance at least, exactly the same results as we obtained by doing things
correctly, treating ϕ 1 and ϕ 2 as real independent variables. My motivation may have been baffling, but I went
through this sequence of computations to make this point.

So it turns out it is not dumb to treat ψ and ψ* as independent. I will begin—I will not complete it, because
once you’ve seen how the first part of it goes, the rest of it will be a trivial exercise—to show that you will always
get the right results if you have a Lagrangian expressed in terms of complex fields, and simply treat ψ and ψ* as if
they were independent. I sketch out why it is legitimate as far as the derivation of the Euler–Lagrange equations
goes. Once you’ve seen my method you will see that the same method can be carried through to obtain the
Hamiltonian form, the equal-time commutators, and so on.

Suppose I have a Lagrangian that depends on a set of fields ψ and ψ*, complex conjugates of each other,
and also on the gradients of ψ and ψ*,

For most practical purposes this Lagrangian is set up so that the action integral is real, guaranteeing that the
Hamiltonian will be Hermitian when we’re all done with quantization. That’s not going to be necessary to any of the
proofs I’m going to give, but I might as well point it out. (This restriction is not practical in all cases; a real
Lagrangian is not a completely general function of these variables.) If I were to go through the variational
procedure that leads to the Euler–Lagrange equations, varying both ψ and ψ*, I would obtain

This is the integral of some god-awful mess obtained by doing all my integration by parts, some coefficient I’ll just
call A to indicate I’m not concerned about its structure, times δψ, plus the conjugate god-awful mess, A* times
δψ*. Nobody can fault me on that.

Now if I were foolishly to treat δψ and δψ* as independent, that is, if I were to consider the variation

that would be obvious nonsense, because ψ* is the conjugate of ψ; I can’t vary them independently. I would obtain
an equation of motion which says

but saying nothing about A*. Likewise by making δψ* arbitrary and δψ = 0, I would obtain A* = 0 and those would
be my two Euler–Lagrange equations of motion. This is obviously illegitimate. I cannot vary ψ without
simultaneously varying ψ* because they’re conjugates. On the other hand, what I certainly can do is choose
matters such that

with δψ real. From this I deduce

That’s legitimate. Alternatively, I could just as well arrange things such that δψ is pure imaginary,

from which I deduce

The net result of the equations (6.42) and (6.44) is

The consequences of this correct procedure are exactly the same as the consequences of the manifestly silly
procedure. I leave it as an exercise to carry out all the other steps of the canonical program, the introduction of
canonical momenta and the Hamiltonian, and the working out of canonical commutation relations, to show that in
general it all comes out the same as if one had treated the ψ and ψ* as independent variables.

Just to show how this goes, I’ll work out the form of the current in the ψ-ψ* formalism. Let’s remember how our
transformations work in our original basis, (6.2). For every λ,

so that

The group defined by the symmetry (6.47) is called U(1), the unitary group in one dimension. It has the same
algebraic structure as SO(2); mathematicians call these two groups “isomorphic”. We have a very simple
expression for Dψ, (see (5.21))

Likewise by considering the transformation properties of ψ*, or by taking the conjugate of this equation,

These are two of the ingredients we need to construct the canonical current. The others are the respective
conjugate momenta:
The Lagrangian (6.34) is obviously invariant under the symmetry (6.47) and Fµ of course is still equal to zero
whether we express the Lagrangian in terms of ϕ 1 and ϕ 2 or in terms of ψ and ψ*:

By inspection, ∂µJµ = 0; the current is conserved.

On the classical level, it is an elementary substitution to show that (6.51) is the same current (6.7) as before;
work it out if you don’t believe me. On the quantum level for free fields it is, in this case, necessary to normal order
so that we get the same results as before. When we write things as (6.51), ψ and ∂µψ* do not commute, and we
have to normal order to make sure that all of our annihilation and creation operators are in the right place. This
concludes the discussion of complex fields.

EXAMPLE 2. SO(n)

We’ve discussed a very simple example involving internal symmetries in which there were only two fields.
There was a digression on the method of complex fields, which enabled us to simplify somewhat the
representations of things in that case. We can get much more complicated internal symmetry structures simply by
returning to our original expression (6.1) with real fields,

and possibly some interaction, say

but now a runs not from 1 to 2, but from 1 to some number n, your choice.

In the same way that the previous theory was invariant under SO(2), this Lagrangian (6.52) is invariant under
SO(n), the connected group of all orthogonal transformations on n real variables. We can imagine these fields as
being labeled by vectors in some n-dimensional space, an abstract, internal space. For every rotation in that n-
dimensional space, there is a transformation of the fields among themselves that would leave the Lagrangian
invariant. We can go through the elaborate procedure of constructing the currents, but I hope you have learned
enough about the n-dimensional rotation group to know that a complete and independent set of infinitesimal
transformations are rotations in the 1-2 plane, rotations in the 2-3 plane, rotations in the 1-3 plane, etc.,2 each of
which is something we have already done, with a slight relabeling. Therefore we will obtain n(n − 1) conserved
currents, because we have n choices for the first number that labels the plane, n − 1 choices for the second, and
as it doesn’t matter what we call first or second when we’re labeling, we divide by 2. The form of these currents,
say the current corresponding to rotation in the a-b plane, will be exactly the same as before,

That’s the same expression as (6.7), with 1 and 2 replaced by a and b. As you see, only n(n − 1) of these
currents are independent, because when a = b the current is zero, and if I interchange a and b, I just get minus the
same current.

There is no analog in this more complicated case for the trick of complex fields. The reason is very simple.
That trick was based on diagonalizing the charge. Here I can hardly expect to diagonalize all the charges
simultaneously, because the corresponding transformations do not commute; a 1-2 rotation does not commute
with a 2-3 rotation. Therefore the corresponding charges should not commute. Still, in some cases it is convenient
to pick one of the n(n − 1) charges as a “nice” charge, and arrange the fields to diagonalize this one charge, the
others remaining ugly and awkward. For example, for n = 3, when we have three degenerate particles, it is
frequently convenient to diagonalize arbitrarily the 1-2 charge. We introduce the fields

6.2Lorentz transformation properties of the charges

Before I leave the general discussion on continuous symmetries there is one gap in our arguments. I have not
discussed the Lorentz transformation properties of the conserved quantities associated with these currents. You
expect from the examples I have given you that the Lorentz transformation properties are the same as those of the
currents except that one index is absent. That is to say, if the current transforms like an nth-rank tensor, we want to
show that the conserved quantity transforms like an (n − 1)st-rank tensor. Where we have a current that transforms
like a 4-vector, the associated conserved quantity is charge, which is a scalar. When we have a two-index object,
for example the energy-momentum tensor, the associated conserved quantity is a one-index object, the total four-
momentum. When we have a three-index object, such as Mνλµ that I talked about before, the angular momentum
currents, the associated object Jνλ is a two-index object, which appears to be a tensor, but we haven’t proved that
it is a tensor. We’ve proved it in particular cases by writing explicit expressions for these objects and showing how
they transform, whereupon it is manifest that they transform in the desired way. But we haven’t shown it in general.
So let me now attack the general problem: If we know how the current transforms, how does the associated
charge transform? I will do in detail in the case of a 4-vector current, Jµ. Once I do it, you will visualize with your
mind’s eye, by adding extra indices to the equations, how the whole thing works out for the energy-momentum
tensor and the angular momentum current.

I have a conserved current, ∂µJµ = 0. I will assume it transforms, as in the case of an internal symmetry, like a
vector field. That is to say under a Lorentz transformation Λ,

The field ψ lowers the 1-2 charge Q12 by one unit, ψ* raises it by one unit and ϕ 0 doesn’t change it at all. The
fields change the charge or not because ψ either creates negatively charged particles or annihilates positively
charged particles, vice versa for ψ*, and ϕ 0 creates and annihilates neutral particles, whence the subscript
“nought”. This notation occurs most frequently when we consider the system of the three π mesons. In the
absence of electromagnetism and the weak interactions, as you probably know from other courses, all three pions
would be degenerate in mass. Indeed this is due to a group with precisely the structure of SO(3) called the
isospin group that acts in exactly the prescribed way on the three pions.3 We pluck out the 1-2 subgroup
because when we introduce electromagnetism the full SO(3) invariance is broken, but a subgroup SO(2) remains,
and this 1-2 invariance corresponds to the conservation of electric charge. The charged pions have the same
mass, but the neutral π0 has a different mass. Of course the π mesons interact with a lot of other particles, so
there’s plenty of “+ . . . ” in the Lagrangian.
Remember when we were discussing field transformation laws, we said that there’s always an inverse operator in
the argument, but the un-inverted thing outside. This will be my only input. This equation could be in classical
physics, in which we take some field configuration and Lorentz transform it, and the current transforms in this way.
Or it could be in quantum mechanics where this transformation is effected by a unitary transformation. Since all
the equations I manipulate will be linear in Jµ, it will be irrelevant whether Jµ is a c-number field or an operator field.

Now we define Q as

We can define Q at any time, since the charge is independent of time by the conservation equation. So just for
notational convenience I will choose t = 0. Because we know how Jµ transforms, we know how Q transforms. It will
go into some object which we’ll compute in a moment and which I will denote by Q′. We wish to ask, “Is Q′ = Q?”
That is to say, in this case, is the space integral of the time component of the current a scalar? To make this
demonstration work, I will have to rewrite the expression for the charge in a way that makes its Lorentz
transformation properties more evident:

turning the space integral (6.57) into a four dimensional integral with x0 = 0. The 4-vector nµ = (1, 0, 0, 0) is the unit
vector pointing in the time direction, so n · x = x0. The expression n · J(x) is simply a fancy way of writing J0. An
equivalent way of writing the same thing is

In this form it is easy to see how to make Q′ look more comparable to Q. We define

because the space derivative of this theta function is zero, and the time derivative gives us δ(x0);
This form (6.59) may make you feel a little nervous because it looks like we can make Q equal zero by integrating
by parts. But we do not have control over the time boundary conditions, and θ(x) = 1 for all positive x, so we can’t
get rid of the boundary term in the time integration by parts.

I will now write a corresponding expression for Q′;

We transform the fields, and then do the same experiment on the transformed field configuration. We’re taking an
active view of transformations. We do not change the integration surface at the same time. The experimenter is
not transformed; that would be a no-no. If we changed both the current and the integration surface, we would
obviously get the same answer. So we are measuring Q′, the same Q defined in exactly the same way for the
transformed current. We have n · J in (6.58), but in (6.61) we have n · (ΛJ), that’s Λ0νJν(Λ−1x), written out in
compressed notation. That is the same integral of the same component, the time component, for the current
corresponding to the transformed field configuration.

and so, by Lorentz invariance of the inner product,

We plug these into our integral (6.61), and we find

In the first step, we use the invariance of d4x under a Lorentz transformation, and in the second step, we simply
change the variable of integration, as is our privilege, we can call it what we please. The third step is just the same
reasoning as gets us from (6.58) to (6.59). Now the only difference between the expressions for Q and Q′ is that n
has been redefined. For Q′, the surface of integration is t′ = 0, and we take n′ · J in the t′ direction. Our active
transformation has had the exact same effect as if we had made a passive transformation, changing coordinates
to x′ = Λ−1x. It’s the same old story, the difference between an alias, another name, and an alibi, another place.
The former corresponds to a passive transformation and the latter to an active transformation.4

To show Q = Q′, we will compute Q − Q′, and see that it equals zero:

Now integration by parts is legitimate, because we can drop the surface terms, the integral over dS µ:

In the surface integral, the quantity in brackets, although not zero, certainly goes to zero at any fixed x as t → ∞,
because eventually n · x becomes positive and n′ · x also becomes positive, each θ function equals 1, and the
difference vanishes. Likewise as t → −∞, eventually both arguments become negative and each becomes zero.

Figure 6.1: The spacetime surfaces n · x = 0 and n′ · x = 0

 Here’s spacetime, showing the surface n · x = 0 and the surface n′ · x = 0, some Lorentz transformed plane.
Okay? The difference of the two θ functions is +1 in this shaded region on the right, where you’re above the n · x
surface but below the n′ · x surface; the difference is −1 in the shaded region on the left, where you’re above and
below in the opposite order; zero when you’re above both surfaces, so both θ’s equal +1, and zero when you’re
below both surfaces, so both θ’s equal zero. Therefore, I can integrate by parts in time without worrying about
boundary terms, as the surface integral goes to zero. So

which equals zero, as ∂µJµ = 0. Thus Q′ = Q. QED

I’ve constructed this argument 5 so that you can readily see that hardly anything is changed if I had had a
tensor current, say Tλµ, instead of a vector current.6 If we’d had a tensor current, the only difference would have
been an extra index with an extra Lorentz matrix on it which I would never ever have had to play with. This matrix
would simply have been carried through all of these equations, playing no role in any of my manipulations, except
to emerge at the end, to tell me that Pµ was a 4-vector.

6.3Discrete symmetries

Of course, there are all sorts of symmetries in nature that are not continuous, not part of some connected group
that contains the identity transformation. Among them are such old friends from non-relativistic quantum
mechanics as parity and time reversal. So we will now study discrete symmetries.

A discrete symmetry is a transformation where

but there’s no parameter in the transformation; it simply doesn’t appear. There’s no such thing as a parity
transformation by 7°; there is only parity: either there is space reflection or there is no space reflection. It’s not like
a rotation. We will assume these things are symmetries in the usual sense. That is to say that, at least for
appropriately chosen boundaries, the action is invariant:

Of course, there may be many fields, but I leave off the indices out of sheer laziness.

Now in a rough and heuristic way, we would expect such a transformation to be a symmetry of classical
physics. And in terms of classical physics this symmetry does what a symmetry always does: it enables you to
generate new solutions of the equations of motion out of old solutions. But in general it is not connected with a
conservation law, as continuous symmetries are. In quantum mechanics there will be no Hermitian operator
associated with these things, to generate the infinitesimal transformation, for the excellent reason that there is no
infinitesimal transformation. We would nevertheless expect that there would be a unitary operator that effects the
finite transformation. Indeed though the argument is rough and ready, everything is determined from the action by
appropriate variations in canonical quantization and so on. The action is the same for ϕ as it is for ϕ′. We should
find a one-to-one correspondence between the Hilbert space we get by doing things in terms of ϕ and the Hilbert
space we get by doing things in terms of ϕ′, since step by step, every step’s the same. If the transformation
doesn’t change the action it can’t change the quantum mechanics; and that means there’s a unitary
transformation that turns ϕ into ϕ′. You know this argument is rough because it’s a lie for time reversal, where
there is no unitary transformation, but we won’t get to that until the end of this lecture. This is just a rough
argument for the sake of orientation. Let’s do some particular cases where we can see simply what is going on
and tell whether or not there is a unitary transformation.

Charge conjugation

The first case I will turn to is our good old example of two free fields of the same mass, (6.1),

On a formal level everything I say will also be true if there’s an interaction, say of this form, for example:
I said that this system was SO(2) invariant but in fact it has a larger invariance group of internal symmetries,
including a discrete internal symmetry. It has full O(2) invariance. That is to say it is invariant not just under proper
rotations but under improper rotations; not just under rotations but also under reflections. We’ve already studied all
the consequences of the rotations. And since every reflection is the product of some standard reflection and a
rotation, we might as well just consider one standard reflection which I will choose to be the transformation

because bringing the ∂ρ through the constant ϵσρ does no harm, nor does bringing the ∂ρ through xσ. Since ∂ρxσ
is gρσ, symmetric in ρ and σ, the term vanishes upon summation with the antisymmetric ϵρσ. By the same trick as
before, used to go from (5.46) to (5.47), I can write this as

Now we have all we need to get the whole thing, the conserved current Jµ, (5.27). We have the change Dϕ a in
the field and we have the change in the Lagrangian written as the divergence of something, Fµ. We can put the
whole thing together and get Jµ,

At least in the free field case, where we can explore the Hilbert space completely, and even in the general case, if
we are willing to extract from non-relativistic quantum mechanics a statement that any operation that doesn’t
change the canonical commutators is unitarily implementable, we can see that there is a unitary transformation
that effects (6.70). In the free case, g = 0, we just read off from (6.70) that if there is a unitary transformation U
such that

then U operates on the annihilation operators like this:

and the same thing for the creation operators just by taking the adjoint.

A unitary transformation that does the job in the free case acts on states with a definite number of particles of
type 1 and a definite number particles of type 2 by multiplying the state by (−1), or equivalently eiπ, raised to the
number operator N 2 of 2-type particles, where

Then

That obviously has the desired property, and works just as well on the fields:

The first equation follows because N 2 commutes with ϕ 1. The second equation is true because ϕ 2 will either
create or annihilate a type 2 meson, and hence change their number by 1.

This unitary transformation is perhaps more simply expressed in terms of the b’s and the c’s. First, recall the
definition (6.23) of the complex field ψ and its conjugate ψ*. Then

Equally well you could say that this U acting on any state turns all the b-type particles into c-type particles and all
the c-type particles into b-type particles. From equations (6.14) and (6.15),
Such a transformation is called charge conjugation. “Conjugation” is a bad word; it sounds like it shouldn’t be
unitary. After all, complex conjugation is not a unitary operation. Perhaps it would better be called
“particle–antiparticle exchange”, because the transformation exchanges particles and antiparticles, π+ ’s and π−’s
for example. We normally call this symmetry C, and put a little subscript C on the unitary operator, U C, to tell you
that that’s the transformation it’s associated with. We can rewrite the transformations on bp and cp in a compact
form,

As I said before, in general a unitary operator is not an observable, and therefore we normally don’t get a
conserved quantity even though the unitary operator may commute with the Hamiltonian. However there is one
special case in which a unitary operator does give us a conserved quantity, and that is when the unitary operator
is itself Hermitian. This happens in the case of charge conjugation because operating twice with U C is just the
identity. Applying C once, you turn every b-type particle into a c-type particle, and then applying it a second time
you turn it back again into a b-type particle;

Because U C is also unitary, U †C = U C−1, and so

That is to say U C is both unitary and Hermitian. That is rather obvious in terms of the C operator’s action on type 1
and type 2 particles, where the eigenvalues were +1 and −1, numbers that are both of modulus one and real. Note
that from (6.19),

(this is the Q associated with the continuous group SO(2)). And so in this particular case, even though this
transformation is not associated with a continuous symmetry, we can divide states up into C-eigenstates, because
C is also a Hermitian operator. This is usually not done in practice except when you have equal numbers of
particles and antiparticles, considering states of a π+-π− system for example. The terminology we now use for
particles connected by this kind of transformation, to have equal numbers of particles and antiparticles, is even
and odd under charge conjugation, depending upon whether the wave function is symmetric or antisymmetric
under exchange of the π+ and π− variables. Since charge conjugation commutes with the Hamiltonian, the notion
of even or odd under charge conjugation can be used to deduce consequences for transition amplitudes. Actually
we won’t do that for π+’s and π−’s because you gain no information there that you don’t gain from parity, but we
will use it for electrons and positrons, where you do gain additional information.

I haven’t deduced particle–antiparticle symmetry. I have simply given an example of a theory which I cooked
up to possess a symmetry that is structurally similar to a symmetry I know exists in nature by experiment, just to
show you how such a symmetry could arise within the context of Lagrangian field theory.7

Parity

As my next example, I would like to discuss parity. Parity changes the signs of the spatial coordinates,
leaving the time coordinate untouched:

Parity is closely related to reflection (say, reflection in the x-y plane, which would take z → −z and leave x, y and t
unchanged). A parity transformation is the same thing as a reflection (in any plane) followed by a rotation about
the normal to that plane by 180º. So a theory with rotational symmetry is parity-invariant if and only if it is
reflection-invariant. But parity is an improper rotation (its determinant equals −1), and parity invariance is not
implied by rotational invariance alone. Nevertheless, until the discovery by Wu and her group8 that parity was
violated in beta decay, it was universally assumed that any realistic physical theory would be parity-symmetric.
 An ordinary scalar (mass m, for example) is invariant under parity, while an ordinary 3-vector, like velocity, v,
changes sign:

On the other hand, a cross-product of two vectors (the angular momentum L = r × p, say) picks up two minus
signs, and the scalar triple product w = a • (b × c) is a scalar that changes sign:

We call these axial vectors and pseudoscalars, respectively, because of their anomalous behavior under parity.

In a field theory we can have scalar fields, pseudoscalar fields, vector fields, axial vector fields, and so on.
Moreover, if there are several fields, they can be mixed by the parity transformation. In this sense the parity
transformation is intrinsically ambiguous: it takes x into −x (and t into t), but what else it does is a matter of
convention and convenience, though we will assume that its action is always linear:

(summing on repeated indices). Parity turns the fields at a point (x, t) into some linear combination of the fields at
the point (−x, t). A theory is parity invariant if the action is unchanged by some transformation of the form (6.85),
but it is not always obvious how we should choose the coefficients Mab. Parity can be very strange and I hope to
amuse you by cooking up a bunch of theories, some of which have no actual resemblance to nature, in which P
takes peculiar forms.

EXAMPLE 3. Scalar field with a quartic interaction

Let’s look at a scalar field with a quartic interaction:

(I am tired of writing ∂µϕ∂µϕ; you know what the first term means.) This obviously possesses a parity invariance,

This transformation changes the Lagrangian

but it doesn’t change the action. In the case of g = 0, it is implemented by the unitary transformation

The parity transformation turns either a creation or an annihilation operator with momentum p into a creation or
annihilation operator with momentum −p. The proof is simple: Apply (6.87) to the definition (6.8) of the free fields,
and then change the integration variable p into the integration variable −p. This turns x into −x and doesn’t
change t. Thus parity takes a particle going, say, this way, →, and turns it into particle going this way, ←, the usual
thing that parity does in non-relativistic particle physics. Acting on the basis states,

There is an alternative parity transformation, which I will call P′,

This transformation is also an invariance of our Lagrangian (6.86) if (6.87) is, because our Lagrangian is invariant
under ϕ → −ϕ (a trivial internal symmetry closely corresponding to what we did to ϕ 2, (6.70), in the discussion of
charge conjugation), and the product of two symmetries is a symmetry. The transformation law (6.87) is called the
scalar transformation law, and (6.91) is called the pseudoscalar transformation law. The unitary
transformation U P′ is given by

where N is the number of pseudoscalar fields being acted on. Likewise, on a basis state describing n
pseudoscalar particles,
The first important point of this example is that it is merely a matter of convention for a particular theory whether
you say ϕ is a scalar field or ϕ is a pseudoscalar field. Whenever there is an internal symmetry in a theory, I can
multiply one definition of parity by an element of the internal symmetry group, discrete or continuous, and get
another definition of parity. This theory has two symmetries, among others; one which is C-like and one which is
P-like. The product CP is a symmetry; and which you call parity and which you call charge conjugation or ϕ → −ϕ
times parity is a matter of taste; nobody can fault you. What is important is the total group of symmetries admitted
by Lagrangians, from which one draws all sorts of physical consequences, not what names one attaches to
individual members. As long as you have one possible definition of parity, and you have internal symmetries
around, you can always adopt a new convention and new nomenclature. You can take the product of one of those
internal symmetries and parity and call that parity, and call your original parity the product of your new parity and
the inverse internal symmetry. Nobody can stop you and nobody should, as long as when you are writing your
papers or giving your lectures, you are clear about what convention you are using.

Of course if the Lagrangian does not have the internal symmetry then you might end up with a unique
definition of parity because there will be no internal symmetries from which you can multiply parity.

EXAMPLE 4. Cubic and quartic scalar interactions together

Consider the Lagrangian

If I take L(1), the same Lagrangian as before, and add to it a term hϕ 3, then ϕ → −ϕ (in the sense of (6.91)) is no
longer a good definition of parity nor is it a symmetry. In this case the only sensible definition of parity is the scalar
law, without the minus sign; the pseudoscalar won’t work. You can call the pseudoscalar transformation “parity” if
you want, but then you have got yourself into the position of saying this theory is not parity conserving, which is a
silly thing to say. In nature, in the real world, sometimes there is no good definition of parity. There is no way of
defining parity so that the weak interactions preserve parity.

If you throw away the weak interactions you have a lot of internal symmetries: the commuting one parameter
groups corresponding to electron number, muon number, nucleon number, electric charge, and strangeness. The
relative parity of the electron and the muon is a matter of convention. You can always multiply muon number into
your definition of parity to change the parity of the muon and the muon neutrino and nothing else. The relative
parity of the electron and the proton is a matter of convention, as is that of the proton and the Λ hyperon; you can
multiply strangeness into that definition of parity. Usually these conventions are established to make all those
relative parities +1, but that’s just convention.

I have shown you an example where the scalar transformation is an okay definition of parity but the
pseudoscalar is not. I will now construct examples where it goes the other way. These examples are rather
unnatural, involving scalar fields. When finally we talk about fermions, we will find we can write very simple
interactions that have this property, but it can also be shown with scalar fields. To do so, I have to write down a
grotesque sort of interaction, using the four-dimensional Levi–Civita tensor, ϵµνρσ, which is completely
antisymmetric (like its three-dimensional cousin ϵijk), and ϵ0123 = 1. With this and with four 4-vectors, one can
form a Lorentz scalar but it will have funny parity properties. I will now give an example of how to make something
where the pseudoscalar law is forced on us if we hope to have the Lagrangian invariant.

EXAMPLE 5. Coupling via ϵµνρσ

If we were to declare all four fields to transform as scalars, then the Levi–Civita term breaks parity because, as you
will notice, every term involves one time index and three space indices. The space derivatives change sign under
parity, and the time derivatives do not. We pile up three minus signs when we parity transform this object, which is
a disaster since three minus signs change the sign of this term. We have to declare that one of the fields is
pseudoscalar and three of the fields are scalar, or vice-versa. Since we have total freedom to make the whole
large group of internal transformations, it’s a matter of taste which one (or three) of the four we call pseudoscalar.
That is just a matter of how we multiply an internal symmetry by a parity.

EXAMPLE 6. The last example, plus a sum of cubic terms

There is no good definition of parity for L(4). I have to have a minus sign in one (or three) of the fields to make L(3)
work out all right, but then the new term, in h, is disastrous, with a sign of −1. On the other hand if I choose all the
fields to be scalar, to get the new term to work out, it breaks the invariance of L(3). Whether I choose scalar or
pseudoscalar fields, it doesn’t matter; there is no symmetry that can be interpreted as parity for this Lagrangian.

Now this demonstration might lead you to think the only possible effect of a parity transformation is a plus sign
or a minus sign, where the particles have intrinsic positive parity or intrinsic negative parity. I will now give an
example where the only possible definition of parity has an i in it. This will be super-grotesque and will involve a
complex scalar field, ψ.

EXAMPLE 7. Modifying the last example by adding new fields

My free Lagrangian now has five fields in it, four real scalar fields ϕ a with some four masses µa, and a complex
scalar field ψ with some fifth mass µ5. We still have the h term that keeps us from letting the scalar fields be
pseudoscalar. Now, however, I’ve multiplied this last term by the sum of the squares of ψ and ψ*. The sesquilinear
form in the fields with an epsilon tensor is not one we will encounter in any of the theories we will take seriously,
but it’s still an amusing example. Though grotesque, it’s got all the properties we want: it’s Hermitian, and if we are
creative, it will have a legitimate parity. The four real fields can be taken as scalars, so all the terms except for the
last are all right. We need the last to go into +1 times itself. That will happen for the last term even with scalars,
provided

Since ψ(x, t) goes into iψ(−x, t), the square of ψ supplies the missing minus sign for the epsilon term from i
squared, and the same is true for ψ*. The other terms in ψ and ψ* are unchanged by (6.98).

This is just for fun, but it is an example where parity is so strange that, as you can readily convince yourself,
for this grotesque theory this is the only possible parity that will work. And in this case, things are so strange that
the square of parity is not even 1 on the complex field. If you ever read a paper in which someone says on general
a priori grounds the square of parity must be +1, send him a postcard telling him about this example. He may say,
“Oh, I wouldn’t call that parity,” but then you would say he was being pretty foolish, because if the world really
were like this, there would be this very useful symmetry that turns observables at the point x into observables at
the point −x, putting all sorts of restrictions on scattering cross-sections and energy levels and all the things a
symmetry usually does, and its square happens not to be one. If he doesn’t want to call that parity, what is he
going to call it?9

Time reversal

Now of the famous discrete symmetries known and loved by physicists, I have left one undiscussed: time
reversal. Time reversal is rather peculiar in that unlike all the other symmetries we have discussed until now, it is
not represented by a unitary operator; it is represented by an anti-unitary operator.
Consider a particle in one dimension moving in a potential. The classical theory is invariant under the time
reversal transformation

While q(t) goes into q(−t), p(t) goes into −p(−t) because p is proportional to . That is to say if you take a motion
picture of this classical system and run the reel for the projector backwards, you will obtain a motion perfectly
consistent with the classical equations of motion. One’s first guess is that there should be a unitary operator,
which I’ll call U T, that effects this transformation in the quantum theory:
This, however, leads one into a grinding contradiction almost immediately. We know from the canonical
commutators that, at equal times,

Apply U T to the right-hand side of the commutator and U T† to the left side, for the time t = 0:

which is unfortunately not i but −i. It looks like we would have to give up our canonical commutation relations to
implement time reversal. Thus we have obtained an immediate contradiction with our hypothesis, so the answer to
this is not “What is the operator?”, but instead, “There is no (unitary) operator.”

There is a second contradiction. We expect that U T, if it exists, should reverse time evolution, i.e.,

Take d/dt of both sides of this equation at t = 0 to obtain

Canceling the i’s, we see that H and −H are related by a unitary transformation. Operators so related must have
the same spectrum, and yet they cannot both be bounded from below! A unitary time reversal operator makes no
sense whatsoever. The resolution of these difficulties is well known: Time reversal is not a unitary operator but an
anti-unitary operator. As I will prove, anti-unitary operators are also anti-linear.

Before getting into anti-unitary operators, let’s review the properties of unitary operators. Unfortunately, one
reason the Dirac notation is so wonderful is that a lot of facts about linear operators are embedded in it
subliminally. Anti-linear operators are therefore difficult to describe in Dirac notation. So instead of using bras and
kets I will use an alternative notation. I will label states by lowercase Latin letters: a, b, . . . These are vectors in
Hilbert space. And instead of talking about the inner product áa|bñ I will write that as (a, b). Complex numbers will
be denoted by Greek letters, α, β, . . . and operators will be denoted by capital Latin letters, A, B, . . .

An operator U is unitary if two conditions are met: it is invertible, and for any two vectors a and b in Hilbert
space, the Hilbert space inner product (a, b) is preserved:

Thus U preserves the norm. (The simplest unitary operator is 1.) An operator U is linear if for any two complex
numbers α and β and any two vectors a and b in Hilbert space,

The condition (6.105) is sufficient to show that U is linear, by a variation on a theorem to be shown below. The
adjoint A† of a linear operator A is defined by

It’s easy to show that if U is unitary, then U † = U −1:

the first step following from U being unitary. A transformation of the states a → Ua can be thought of as a
transformation on the operators:

An anti-unitary operator is an invertible operator, traditionally represented by an omega, Ω (one of the few
instances of felicitous notation in theoretical physics, as an omega is a U upside down), defined by

The product of two anti-unitary operators is a unitary operator, the product of an anti-unitary object and a unitary
object is anti-unitary, and so on. The multiplication table is shown in Figure 6.2.
 Figure 6.2: Multiplication table for Ω and U

Such operators Ω certainly exist. A simple example which obeys all of these conditions in one-dimensional
quantum mechanics is complex conjugation K of the Schrödinger wave function. The complex conjugate of a
linear superposition of two wave functions is the superposition of the complex conjugate with complex conjugate
coefficients:

Likewise if I complex conjugate both factors in the inner product I complex conjugate the inner product:

A useful fact (especially conceptually) is that any anti-unitary operator Ω can be written as the product of a unitary
operator U and the complex conjugation operator K: Ω = UK. It’s easy to prove this by construction: take U = ΩK.

An operator A (not necessarily invertible) is called anti-linear if

To show that an anti-unitary operator must be also anti-linear, consider the inner product of

with itself. If this is equal to zero, the positive-definite inner product implies that the original state is zero, i.e.,

It suffices to multiply out the nine terms of the inner product (6.114) and apply the relation (6.110) to remove all
instances of Ω, e.g.,

and then expanding the five terms containing (αa + βb). Sure enough, you obtain zero, thus establishing (6.115).
(The analogous proof that unitary operators are necessarily linear only uses properties of the inner product, and is
even easier.)

The transformation of the states under an anti-unitary operator Ω, a → Ωa, can also be thought of as a
transformation of the Hermitian operators in the theory, though in a more limited sense. Consider the expectation
value of a Hermitian operator A acting on the state a. It transforms under Ω as

So the transformation may alternatively be thought of as A → Ω−1 AΩ. We do not write Ω†AΩ because Ω† is not
even defined for anti-unitary operators.

The resolution of the contradictions, (6.102) and (6.104), is that time reversal is effected by an anti-unitary
operator. For the first contradiction, (6.102),

because whenever we drag a complex number through an anti-unitary operator we complex conjugate it. Thus the
right- and left-hand sides of the equation match and the contradiction disappears. Indeed, for this particular
problem, it is easy to find the anti-unitary operator that effects time reversal: it is complex conjugation in the x
representation. That turns x into x and it turns p which is −i∂/∂x into −p because of the i.

As for the second contradiction, (6.104),

which resolves the second contradiction, provided H is invariant under time-reversal. So much for a lightning
summary of the situation in non-relativistic quantum mechanics.
You may have heard of Wigner’s beautiful theorem,10 which tells you that, up to phases, an operator that
preserves the norm of the inner product must be either unitary or anti-unitary. (It is not necessary to preserve inner
products; they aren’t measurable. Only the probabilities are measurable.) In the study of symmetries, as Wigner
pointed out, all one really has to consider on a priori grounds are unitary and anti-unitary operators; there is no
need worrying that someday we will find a symmetry that is implemented by a quasi-unitary operator or some
other entity not yet thought of by mathematicians. Simply put, Wigner’s theorem says that if F is a continuous
transformation mapping some Hilbert space H into itself, and if F preserves probabilities, then F must be the
product of a phase and a unitary or an anti-unitary operator. That is, if a, b ∈ H, then

where ϕ: H → R.

We now wish to take our standard field theoretic system, the free scalar field of mass µ, and find a time
reversal operator. So we are interested in the system defined by the Lagrange density

I pick this one because we can explicitly write the operators on the state space. What I said about parity also
applies to time reversal; I can multiply the time reversal operator by any internal symmetry and obtain an equally
good time reversal operator. Let’s try to figure out what ΩT must be, working directly with the states, the opposite
direction from which we worked before, and then show what ΩT does to the fields. In a relativistic theory, it is more
convenient to study ΩPT than ΩT, that is to say the product of parity and time reversal. The reason is very simple.
Acting on xµ, PT multiplies all four components by −1. This operation commutes with the Lorentz group. Time
reversal multiplies only t by −1, singling out one component of the 4-vector xµ, and does not mesh well with
Lorentz transformations.

Now, what do we expect the combined symmetry PT to do to a single-particle state? Well, if I have a particle
whose momentum vector is represented by an arrow, →, parity will reverse the sign, and make it ←; but time
reversal will reverse it again from ← to →. So I expect PT to do nothing to the momentum of the particle. Therefore I
define the anti-unitary operator ΩPT acting on a complete set of basis states (assuming that ΩPT |0ñ = |0ñ)

For either kind of operator, unitary or anti-unitary, if you specify its action on a complete orthonormal basis, you
have specified it everywhere. Notice that this does not imply (as it would for a unitary operator) that ΩPT = 1,
because it’s an anti-unitary operator, and therefore, although it turns these states into themselves, it doesn’t turn i
times these states into themselves; it turns them into −i times these states. Okay, that’s our guess. I’ve defined an
anti-unitary operator which is a symmetry if there ever was one; it commutes with Lorentz transformations, the
Hamiltonian, and the momentum; that’s surely good enough to be a symmetry..11 Let’s figure out what it does to
the fields, ϕ.

Well, let’s begin with the annihilation and creation operators. The formulas that define the annihilation and
creation operators only involve real numbers, and ΩPT does nothing to p, so one easily deduces that

Equivalently, multiplying from the left by Ω−1PT we get

It sure looks like ΩPT acts like 1 so far. By the same reasoning we have a similar equation with ap† replacing ap.
Now what about the field? Here comes the cute trick. The field, as you recall from (6.8), is

Now when I apply Ω−1PT and ΩPT to this, what happens? Well, nothing happens to the d3p, nothing happens to the
(2π)3/2, nothing happens to the , nothing happens to the ap. But ahh, the e−ip·x gets complex conjugated, and
likewise the eip·x , so I get ϕ(−x), which is exactly what I would want for a PT operation–it turns the field at the
spacetime point xµ into the field at the spacetime point −xµ:

The operator ΩPT is not acting like 1, the identity, because the operator is anti-unitary. Any equation defining an
operator in terms of the states where it only has real matrix elements will commute with ΩPT, but not if the
elements are complex or imaginary.

This concludes the discussion of time reversal. Because we were dealing with scalar particles, the discussion
was rather simple. Much later in this course when we deal with particles with spin, time reversal will be somewhat
more complicated, just as it is with spin in non-relativistic quantum mechanics. This also concludes the general
discussion of symmetry. Our next topic is the beginning of perturbation theory.

1We have left particle mechanics behind, and I’ll often use Lagrangian to mean the Lagrangian density, L.
2[Eds.]See Goldstein et al. CM, Section 4.8, “Infinitesimal Rotations”, pp. 163–171, or Greiner & Müller QMS,
Section 1.8, “Rotations and their Group Theoretical Properties”, pp. 35–37.
3[Eds.] Isospin is equally well described by the group SU(2), which is locally isomorphic to SO(3). See note 37, p.
791.
4[Eds.] See pp. 18–19 and p. 36 in Greiner & Müller QMS.
5[Eds.] For an extended version of this argument, see Eugene J. Saletan and Alan H. Cromer, Theoretical
Mechanics, J. Wiley & Sons, (1971), pp. 282–283. In the literature this argument is sometimes called “Laue’s
theorem”, after Max von Laue (Physics Nobel Prize 1914, x-ray diffraction). See M. Laue, “Zur Dynamik der
Relativitätstheorie” (On the dynamics of relativity theory), Ann. Phys. 35 (1911) 524–542. Von Laue (he gained the
“von” through his father, in 1913) was courageously public in his fierce opposition to the Nazis. Lanczos writes,
“Years after the Second World War an eminent physicist from Germany visited [Einstein] in Princeton. As he was
about to leave, he asked Einstein whether he wanted to send greetings to his old friends in Germany. ‘Grussen
Sie Laue’, was Einstein’s answer: ‘Greetings to Laue’. ‘Yes’, said the visitor, ‘l shall be happy to convey these
greetings. But you know very well, Professor Einstein, that you have many other friends in Germany’. Einstein
pondered for a moment, then he repeated: ‘Grussen Sie Laue’.” (C. Lanczos, The Einstein Decade 1905–1915,
Paul Elek Scientific Books, (1971), p. 23.)
6[Eds.] See note 11, p. 94.
7[Eds.]A student asks about the CPT Theorem. Coleman responds: “CPT is very different. That’s something we
won’t get to until very late in this course if we bother to do it all. Just from general assumptions of field
theory—Lorentz invariance, the positivity of the energy, and locality (the fact that fields commute at spacelike
separations)—without making any assumption about the form of the Lagrangian, or even whether things are
derived from a Lagrangian, you can show that there is CPT invariance. This is the famous CPT Theorem.
Although one after another—parity, time reversal, and charge conjugation—have fallen to experimenters, the
combined symmetry CPT remains unbroken, and we believe the reason is the CPT Theorem. Indeed one of the
most revolutionary experimental results conceivable—well, violation of conservation of energy would also be
pretty revolutionary—would be that CPT had been found to be violated. If that were so, we would not only have to
sacrifice all the particular theories with which we hope to explain the world; we would have to sacrifice the general
ideas, including the idea of a Lagrangian field theory and indeed the general idea of local fields. It would be back
to Lecture 1; this whole course would be canceled out!”
8[Eds.]
C. S. Wu, E. Ambler, R. W. Hayward, D. D. Hoppes, and R. P. Hudson, “Experimental Test of Parity
Conservation in Beta Decay”, Phys. Rev. 105 (1957) 1413–1415.
9[Eds.] A student asks: Why are we concentrating on linear transformations? Coleman replies: “The linear
functions come from the fact that in all our examples, the kinetic energy is a standard quadratic form. That means
nonlinear transformations will turn the kinetic energy from a quadratic function of the fields to a messy function of
the fields. We could rewrite things in an ordinary, perfectly symmetric theory with two fields in it, ϕ 1 and ϕ 2. Out of
sheer perversity I could introduce fields ϕ 1′ = ϕ 1 and, say, ϕ 2′ = (ϕ 2 + aϕ 1)2. And then my kinetic energy would
look rather disgusting and my ordinary isospin transformations discussed earlier that turned ϕ 1 into ϕ 2 would look
like horrible nonlinear transformations. That’s a silly thing to do but it is not absolutely forbidden. So there is
nothing sacred about linear transformation laws of fields. It’s the bilinear structure of the kinetic energy that makes
linear transformation laws of such interest to us.”
10[Eds.] Eugene Wigner, Group Theory and Its Application to the Quantum Mechanics of Atomic Spectra,
Academic Press, 1959, Appendix to Chap. 20, “Electron Spin”, pp. 233–236, and Chap. 26, “Time Inversion”, pp.
325–348; Weinberg QTF1, Chap. 2, Appendix A, pp. 91–96. Wigner, a childhood friend of John von Neumann and
trained as a chemical engineer, was instrumental in the construction of the Chicago nuclear pile (2 December
1942), and, with Alvin M. Weinberg, wrote the book on the design of subsequent reactors. Perhaps the leading
proponent of group theoretical methods in quantum mechanics, he shared the 1963 Physics Nobel with Maria
Goeppert-Mayer (until 2018, the only other woman Physics Laureate besides Marie Curie) and J. Hans D. Jensen.
Wigner was Dirac’s brother-in-law; Dirac married Wigner’s sister Margit in 1937.
11[Eds.] Coleman will state later (§22.4) that the Klein–Gordon equation is invariant under PT. He doesn’t prove
this, but it’s obvious. The KG operator is second order in both x and t, so it is invariant. We’ve seen that ϕ(x) →
ϕ(−x). By the Chain Rule, the Klein–Gordon equation is thus invariant under PT.

7
Introduction to perturbation theory and scattering

We are now going to turn to a topic that in one guise or other will occupy us for the rest of the semester, the topic
of perturbation theory and scattering. This will lead us to Feynman diagrams, complicated homework problems,
worries about renormalization, and everything else. But we begin at the beginning.

I want to divide the problem into two pieces: perturbation theory, and scattering, at least on our first go-
through. First I will discuss perturbation theory: How one solves quantum dynamics in perturbation theory, how
one finds the transition matrix or whatever you wish to discuss, between states at finite times in perturbation
theory. Next, I will discuss the asymptotic problem: Given such a solution, how does one extract from it scattering
matrix elements. So first I’ll discuss perturbative dynamics. After that I will discuss scattering.

7.1The Schrödinger and Heisenberg pictures

I begin by reminding you of the two pictures that play such a large role in ordinary quantum mechanics, the
Schrödinger and Heisenberg pictures. I will put little subscripts on things, S or H, to indicate whether we are in the
Schrödinger picture or the Heisenberg picture, respectively. First, the Schrödinger picture. In the Schrödinger
picture, the fundamental dynamical variables, the p’s and the q’s, are time-independent:

I’ll speak as if there’s only one p and one q, just to simplify equations, but everything I say will be true if there are a
million p’s and a million q’s. The states, on the other hand are time-dependent, and obey the Schrödinger
equation

The Hamiltonian H depends on pS, qS and perhaps also t.

The fundamental dynamical problem in the Schrödinger picture is this: Given the state |ψ(t′)ñ at any time t′,
determine the state at a later time t. We define an operator U(t, t′), called the time evolution operator, by this
equation,

That is to say, the U operator takes the state at time t′ and produces a state at time t. U(t, t′) is a linear operator
since the Schrödinger equation is a linear equation, and is a unitary operator,

because the Schrödinger equation conserves probability. The operator U(t, t′) obeys what we might call a sort of
group property, a composition law

That is to say if I go first from time t″ to time t′, and then from time t′ to time t, that’s the same as going from t″ to t in
one fell swoop. The U matrix also obeys a differential equation, the Schrödinger equation,

with the initial condition

This differential equation is a direct consequence of the Schrödinger equation (7.2). Notice that the initial condition
and the composition law imply

Solving dynamics in the Schrödinger picture is equivalent to finding this U operator. If H is, simply a function of pS
and qS, that is to say, if H does not depend explicitly on t, then we can at least write a formal expression for the U
matrix,

Things get more complicated if H is time-dependent. For the time being, we’ll assume H is time-independent.

The Heisenberg picture is the same as with a time-dependent unitary transformation. In the Heisenberg
picture, the states are defined to be time-independent. Just so we can compare the two pictures, we identify the
Heisenberg states with the Schrödinger states at the arbitrarily chosen time t = 0:

so that

In the Heisenberg picture, on the other hand, the fundamental p and q operators are defined to be time-
dependent. In particular,

because we identify qH(0) with qS(0),

and likewise for p. I won’t bother to write down the equation for p. The reason we define things this way is that a
Heisenberg picture operator AH(t) evaluated between Heisenberg picture states at any particular time is
equivalent to the corresponding Schrödinger picture operator AS(t) evaluated between Schrödinger picture states
at the same time:

It’s just in one case you’ve got the U operator on the states and in the other case you’ve got the U operator on the
operators, but the combined expression is the same. The correspondence between AS(t) and AH(t) follows:

From this, we find the time evolution of the fundamental operators pH(t) and qH(t) in the Heisenberg picture:

This is general quantum dynamics, independent of perturbation theory.

7.2The interaction picture

I would now like to turn to perturbation theory computations of the U operator. Notice please that solving the
dynamics in the Heisenberg picture is tantamount to solving it in the Schrödinger picture: they are both equivalent
to finding the U operator.

We will consider a class of problems where the Hamiltonian H(p, q, t) is the sum of a free Hamiltonian, H 0,
let’s say in the Schrödinger picture, and a Hamiltonian H′ that may or may not depend on the time,
Ultimately we are interested in real-world dynamics, where the total Hamiltonian is time-independent. But it’s
frequently useful, when we’re doing some approximations to the real world, to consider time-dependent
Hamiltonians. For example, if we have an electron in a synchrotron, we don’t normally want to have to solve the
quantum mechanics of the synchrotron. We could do it that way, but it’s inconvenient, and we normally consider
the synchrotron as a time-dependent pattern of classical external electric and magnetic fields acting on the
electron. And therefore I will consider time-dependent interaction Hamiltonians. We assume that we could solve
the problem exactly if it were not for H′. We wish to get a power series expansion for the dynamics in terms of H′.
That’s our problem. We can go first-order in H′, second-order in H′, etc. If you want, you can put a hypothetical
small coupling constant in front of H′, and say we are finding a power series expansion in that coupling constant,
but I won’t bother to do that.

This is most easily done by going to a special picture called the interaction picture (also known as the Dirac
picture), which is sort of halfway between the Schrödinger picture and the Heisenberg picture.1 We move from
the Schrödinger to the interaction picture with the same kind of transformation that takes us from the Schrödinger
picture to the Heisenberg picture, but now using only the free part of the Hamiltonian:

where qS(t) = qS(0), and pI(t) similarly. Of course, we must also change the states,

and the operators,

This ensures

The advantage of the interaction picture is this: If H′ were zero, H would equal H 0, and |ψ(t)ñI in the interaction
picture would be independent of time, because it would be the Heisenberg picture; (7.19) would reduce to (7.11).
Thus all the time dependence of |ψ(t)ñI comes from the presence of the interaction.

We can derive a differential equation for |ψ(t)ñI which we will then attempt to solve iteratively in perturbation
theory:

where

H′S(pS, qS, t) can be expanded as a power series in pS and qS, and factors of e−iH0teiH0t can be inserted
everywhere to turn pS and qS into pI and qI. H I(t) is the same function of the interaction picture p’s and q’s as the
Schrödinger interaction Hamiltonian H′S is of the Schrödinger picture p’s and q’s. As promised, the time evolution
of |ψ(t)ñI goes to zero when H′ goes to zero.

This equation, (7.22), is the key equation. By solving it iteratively, we will obtain the solution to the time
evolution problem as a power series in H I. We will always use perturbation theory for the case where H 0 is time-
independent. On the other hand, please notice that even if H′ is time-independent, H I might well be time-
dependent because of the time dependence of pI and qI. In all the cases we will treat, H I will be a polynomial, e.g.,
λϕ 4. This equation (7.23) is true modulo ordering ambiguities if H I is any function of the p’s and q’s.

We solve (7.22) by introducing the interaction picture operator U I(t, t′), defined by the equation

This is of course just like the ordinary U(t, t′) operator. You give me the state of the system at a time, oh, 100 BCE,
and, by operating on it with U I, I will tell you the state of the system now. It obeys equations similar to the ordinary
U. It’s unitary:
and, just as with the earlier U, one can get from t″ to t by going through an intermediate time t′,

From these two equations one can derive a third as in the earlier case:

The earlier equation (7.8) wasn’t useful to us, but this one will be.

U I is not an independent entity; it is given in terms of the ordinary U. Let’s look at t = 0 when all of our pictures
coincide:

just as in passing from the Schrödinger to the Heisenberg picture. From (7.3)

Moreover, from (7.19) and (7.24),

Then, from the identity of the kets at t = 0,

For other times, things can be reconstructed using the known properties of the U’s. For example,

Finally, (from (7.22) and (7.24)) U I obeys a differential equation

just as in the development in the Schrödinger picture.

7.3Dyson’s formula

Our task now will be to solve this differential equation, (7.33). That is, we want to find a formal power series
solution for it which is equivalent to solving dynamics in the interaction picture, and, by formula (7.31), to solving
the dynamics in any picture. If we were doing the very simplest kind of quantum mechanical system, with a one-
dimensional Hilbert space, then the solution would be simply

Unfortunately, H I is not a one-by-one matrix; it isn’t even an infinity-by-infinity matrix in most cases, and H I’s at
different times do not commute with each other. So this formula is false. If we attempt the differentiation to make
things work out, we’ll find, after we differentiate, that we get all sorts of factors inside other factors, which we can’t
drag out to the left. I will take care of this difficulty by introducing a new ordering, called time ordering, rather
parallel to the normal ordering we saw earlier.

Given a sequence of operators A1(t1), A2(t2), . . . , An(tn) labeled by the times, I define the time-ordered
product T(A1(t1)A2(t2) . . . An(tn)) of the string of operators as

the same string of operators rearranged, such that the operator with the latest time is on the far left, then the next
latest time, then the next, and so on. The convention, thank God, has a simple mnemonic, “later on the left”, easy
to remember. If two or more times are equal, then the time ordered product is in fact ambiguous. There are cases
where we have to worry about that ambiguity, if the two operators do not commute at equal times. In the
exponential for U I, however, we will apply the time ordering to factors of H I, and since H I commutes with itself at
equal times, there is no problem. You have seen this time ordering before, for two operators. I defined it in the first
homework assignment (Problems 1.2 and 1.3, p. 49). That earlier definition agrees with this one for the case when
there are only two operators.

The time-ordering symbol shares many features with the normal-ordering symbol. For example, the order in
which you write the operators down inside the brackets is completely irrelevant, since the actual order in which we
are to multiply them is determined by their times, not by the order in which they are written. As with the normal-
order product, I must warn you the time-ordering prescription is not, “Compute the ordinary product and then do
some mysterious operation to it, called time ordering”. It is a new way of interpreting those symbols as they are
written. I say this to keep you from getting into contradictions. Suppose you have two free fields, ϕ(t1) and ϕ(t2).
The time-ordered product of the commutator of these two is zero, but the commutator is a number, and how can
the time-ordered product of a number be zero? That’s false reasoning. Time ordering a product means:
“Rearrange the terms and then evaluate the product.”

I will now demonstrate that the correct solution to our problem is the following beautiful formula, due to
Dyson2

almost the same formally as (7.34), but (7.36) defines a completely different operator because everything is to be
time ordered. This is called Dyson’s formula. I will say a little about the meaning of the formula and then show
you that it solves the equation. This formula is valid only if t is greater than t′. It is not true otherwise. Fortunately
that presents no difficulties because if we know how to compute U I for one ordering, we know from (7.25) and
(7.27) how to compute it for the other ordering, by taking the adjoint.

This formula is only interpretable as a formal power series. It’s not saying, “Compute the integral, find out
what operator is exponentiated, and then do something.” I will write out the first three terms in the power series
just to emphasize that:

The first term is 1, and the time-ordering symbol does nothing to that. The second term involves only a single
operator, so again the time-ordering symbol carries no force. The third term involves two integrals from t′ to t, and
here I can’t drop the time-ordering symbol because I have two operators and two times. Over half the range of
integration where t1 is greater than t2, this symbol is to be written first H I(t1) then H I(t2). Over the other half of the
range of integration where t1 is less than t2 the two operators are to be flipped.

Now why is this time-ordered power series the solution to the differential equation (7.33)? It certainly obeys
the boundary conditions: it’s equal to one when t = t′ because the integrals are all zero, and the series reduces to
the first term only. Let’s evaluate its time derivative:

Inside the time-ordering symbol everything commutes, so in doing our differentiation we don’t have to worry about
the orders of the operators. We will get just what we would get by differentiating naïvely, to wit, everything inside
the time-ordering symbol: H I(t) times the time-ordered exponential; the time-ordering symbol takes care of all the
ordering for us. Now comes the beauty part: t is the absolute latest time that occurs anywhere in the expression
because the integral runs from t′ to t and t is greater than t′. Therefore the Hamiltonian H I(t) has the latest time of
any of the operators that occur within the time ordering, and latest is left-est! The Hamiltonian is always on the left
in every term in the power series expansion, so we can write

That is precisely the differential equation for which we sought a solution, so the argument is complete. If the
question is how do we do time-dependent perturbation theory, to find the dynamics as a formal power series in an
interaction, the answer is Dyson’s formula. Although perfectly valid, Dyson’s formula is rather schematic, and we
will beat on it quite a bit using all sorts of combinatorial tricks to find efficient computational rules. The entire
contents of time-dependent perturbation theory is in this formula, (7.36).
 7.4Scattering and the S-matrix

The next problem we will discuss is scattering theory. I presume you have taken a course in non-relativistic
quantum mechanics, and so you have a general idea of the shape of non-relativistic scattering theory. I would like
to review some features of that theory, just to emphasize certain points to see what the beau idéal of a scattering
theory should be, what criteria should we choose, and then we will try to construct a description of scattering in
relativistic quantum mechanics. We will emphasize features important for our purposes that might not have been
considered important in non-relativistic quantum mechanics, and so may be new to you.

What I mean by an ideal scattering theory is a description of scattering, of what information you have to drag
out of the dynamics to compute cross-sections or something, that makes no reference whatsoever to perturbation
theory. Then, if you could solve the problem exactly, if you could find the U matrix, you’d have a machine. You’d
feed a Lagrangian in, you’d turn the crank, and you would fill out the cross-sections. You cannot solve for the U
matrix exactly in typical cases. You might for example only be able to solve for it in perturbation theory. But that’s
all right. If you have an approximate U matrix, you put it into exactly the same machine, you turn the crank and out
comes an approximation for the cross-section.

So let’s consider a non-relativistic Hamiltonian,

This is really the simplest case, and I will assume that V(x) goes to zero, say, faster than 1/x2, so we don’t have to
worry about long-range forces. (I think it suffices to say it goes to zero faster than 1/(x log x), but forget that.)

Characteristic of a scattering problem is that a quite complicated motion at finite time interpolates between
simple motion, according to the free Schrödinger equation, in the far past and the far future. Say I have this
potential, V(x), localized say in the vicinity of my overflowing ashtray. In the very, very far past, far from this
potential, I prepare a wave packet. I allow this wave packet to move through space towards the potential. It goes
along as if it were a free wave packet, until it (or its fringes, since it’s spreading out) intersects the potential. Then it
goes bananas, it wiggles and bounces around in quite complicated ways. And then, after a while, fragments of the
wave packet fly out in various directions. If I then look in the very far future I have just a bunch of free wave
packets now all moving away from the potential. The problem with scattering is a problem of connecting the simple
motion in the far past with the simple motion in the far future. Let us frame these words in equations.

Since I talked about wave packets, I better look at the Schrödinger picture. Let |ψ(t)ñ be a solution of the free
Schrödinger equation

where

This ket |ψ(t)ñ represents the wave packet I have prepared in the far past. If there were no potential, it would just
evolve according to the free Schrödinger equation. In the very far past, because the wave packet is very far from
the potential, it evolves according to the free Schrödinger equation. The ket |ψ(t)ñ, and the other solutions to the
free Hamiltonian, H 0, belong to a Hilbert space, H 0. The solutions to H, the actual Hamiltonian of the world, belong
to a Hilbert space H. Somewhere in H there is a corresponding state that, in the distant past, looks very much like
|ψ(t)ñ. We will call this state |ψ(t)ñin. It is a solution of the exact Schrödinger equation that represents what the wave
packet really does:

The two states |ψ(t)ñ and |ψ(t)ñin are connected by the requirement that if I look in the very far past, I can’t tell the
difference between them. That is to say,

(where |ψñ = |ψ(0)ñ and |ψñin = |ψ(0)ñin). The norm of the difference of the states goes to zero as t → −∞. The
operation of associating the appropriate state |ψñin ∈ H to a given state |ψñ ∈ H 0 in the limit t → −∞ can be called
“in-ing”.
 I emphasize that |ψ(t)ñin is a genuinely normalizable wave packet state. I can’t make |ψ(t)ñin a plane wave,
because a plane wave state has no norm. And physically, it doesn’t make any sense to talk about a plane wave. It
doesn’t make any difference whether you go to the far past or the far future. Because a plane wave has infinite
spatial extent, it never gets away from the scattering center.

The distinction between past and future makes a great deal of difference to human beings, but not so much to
quantum dynamics, so we need to consider the far future as well. Given another state |ϕ(t)ñ ∈ H 0, there is another
state in H that looks a great deal like |ϕ(t)ñ in the far future, which we’ll call |ϕ(t)ñout. This ket is also a solution of the
exact Schrödinger equation:

In the far future, these two corresponding states cannot be distinguished:

The operation of associating the appropriate state |ϕñout ∈ H to a given state |ϕñ ∈ H 0 in the limit t → ∞ can be
called “out-ing”.

For every free motion there is a physical motion that looks like it in the far past and another physical motion
that looks like it in the far future.3 In-ing and out-ing connect the free solution to the exact solution in the far past
and the far future, respectively, turning free motions into physical motions. We use free particle states as
descriptors, to describe actual interacting particle states. We know how to associate a state with these descriptors
by these correspondences.

Think of classical scattering in a potential. The analog of a free motion would be a straight-line motion in
classical mechanics. Figure 7.1 shows some motion of the particle, when there is no potential. That’s the analog of
|ψ(t)ñ. If the potential is restricted to some finite space-time region, the real motion of the particle looks like Figure
7.2. The particle enters the potential and it deviates from that, and then it comes out, again moving freely. At the
lower right is ψ(t)ñin, the exact motion that looks like |ψ(t)ñ in the far past. At the upper left is |ϕ(t)ñout, the exact
motion that looks like |ϕ(t)ñ in the far future. The in and out states are exact solutions of the real Hamiltonian at all
times. In scattering theory, we are trying to find the probability, and hence the amplitude, that a given state looking
like |ψñ in the far past will look like |ϕñ in the far future, namely outáϕ|ψñin. (Notice that we don’t have to put a t in this
expression because both |ψñin and |ϕñout evolve according to the exact Schrödinger equation, and their inner
product is independent of time.) The correspondences between ψñin and |ψñ and between |ϕñout and |ϕñ allow us to
define a very important operator in H 0, the scattering matrix S, which acts between the descriptor states.4 We
define the S-matrix by the equation5

Figure 7.1: The free ψ (V = 0)

Figure 7.2: ψ in and ϕ out, asymptotic to the free ψ and ϕ

 The S-matrix obeys certain conditions. For example,

That is, the scattering matrix conserves probability. It also conserves energy, if this is a time-independent
problem:

Notice the H is H 0, because the energy operator acts on the descriptors, the states that move according to the free
equation. The operator S turns free states of a given energy into other free states of a given energy. You prove
this by computing the expectation value of the energy (or any power of the energy) in the far past, when you can’t
tell the in state from the free state, and computing it again in the far future, when you can’t tell the out state from
the free state, and requiring that these values be the same.

So much for the scattering theory of a single particle and a potential. I’ve gone through it in a dogmatic way
without proving any of these equations because I presume you have seen them before. We’re not going to use all
this formalism, by the way, in relativistic theory, at least not for the time being.
Now it should be emphasized that in this way of linking states, it looks like there’s a connection with
perturbation theory, with breaking up the Hamiltonian into an H 0 and a V. This is not so. And the easiest way to
demonstrate that is to consider another simple system.

Let’s consider three particles, all with the same mass, with central potentials between them:

where the arguments of the potentials are the usual differences between the centers of the particles. Let me
assume that V12 all by itself could make a bound state. The center of mass Schrödinger equation is

where µ is the reduced mass, and r = |x1 − x2|. It has one bound state, and none of the other potentials make
bound states, they’re all repulsive, and this one has only one. This could be, aside from the long-range nature of
the forces (and the hypothetical equality of the masses), a proton, an electron and a neutral π0 meson. There is no
binding between the proton and pion, nor between the electron and the pion, but there is binding between the
proton and the electron, to make a hydrogen atom. Of course, the hydrogen atom has an infinite number of bound
states.

Now if we seek for descriptors here, we find things fall into two channels, one in which, in the far past, the
states look like three free particles, and the other looking like one free particle and one bound state of the 1 and 2
particles. Both of those can happen since 1 and 2 can bind. Therefore we have two kinds of states, type I with
corresponding in and out states like this:

labeling these as type I states. These are solutions to the exact Schrödinger equation that in the far past and the
far future look like three widely separated particles. For them, H 0 is just

We also have states of type II: orthogonal, exact solutions of the Schrödinger equation for which a complete basis
could be specified by giving the momentum of the third particle which doesn’t bind, and the combined momentum
p of the 1-2 pair (with respect to the center of mass), which is in a bound state

For these, the Hamiltonian is

If V12(r) is not in the Hamiltonian, these type II free states will not time-evolve in the appropriate way. It’s V12 that
keeps them held together; without this potential, the 1 and 2 particles will fly away from each other. In this case,
there are two alternatives for the free Hamiltonian for the definition of the in and out states, depending on what
kind of states we look at. All the in states of type I are orthogonal to all the in states of type II, and the same is true
of the out states. If a state looks like three widely separated particles in the far past it is also not going to look like
one free particle and one bound state in the far past. Its probability for doing that is zero. On the other hand the in
states of type II are not orthogonal to the out states of type I or vice versa: ionization can occur. You can scatter a
free particle off of a bound state in the far past and get three free particles in the far future. So this shows the
situation to be more complicated, and the complication has nothing to do with perturbation theory.

Now, what are we looking for? What is the beau idéal, the grail of a quantum field theory, to describe
relativistic quantum scattering? What sort of in states and out states do we expect to have? Well, fortunately we
have locality and all that. We imagine that if we have a particle of type zilch,6 we can have two widely separated
particles of type zilch, or three widely separated particles of type zilch, so we would expect that our descriptor
states would belong to a Fock space for a bunch of free particles. That would correspond to 1, 2, 3 . . . particles of
various kinds, moving in toward each other or moving away from each other in the far past, all in appropriate wave
packets. What kind of particles should be there? Well, all the stable particles in the world, whatever they are!
That’s a big list of particles. There’s electrons, and neutrinos, and there are hydrogen atoms in their ground states,
and there are photons, and there are alpha particles and there are ashtrays. (That’s a stable system; I don’t think
I’ve ever seen an ashtray decay–it has a lot of excited states, you can put dimples in it and everything, but it’s a
stable system. Fortunately, we have to go to quite a high center-of-mass energy before we begin to worry about
ashtray–anti-ashtray production.) They should all be there, and there should be a great big Fock space that
describes states of, say, one electron, 17 photons, 14 protons, 4 alpha particles, and 6 ashtrays. And then there
would be some S-matrix that connects one to the other.7

To describe a scattering theory that is capable of handling the situation is a tall order. After setting up these
high hopes, I will make you all groan by describing the simple way we are going to do scattering theory for our first
run through. This first description of scattering theory will obviously be inadequate. We will eventually develop a
description that in principle will enable us to handle a situation of this complexity. In practice, of course, it’s a
different story, just as in practice it’s a very difficult thing to compute ionization in any sensible approximation. But
we will develop a description where, if we did know the time evolution exactly, we would be able to compute all
scattering matrix elements exactly. This description will however take quite a long time to develop.

There are many features of the general description that are rather obscure, if you are working with no specific
examples to think back on. And so I will begin with the crudest and simplest description of scattering, the most
ham-handed possible. Then, as we go along doing examples, we will find places where this description clearly has
to be fixed up. To make our Model A work,8 we will add a tail fin here, and change the carburetor there. After
we’ve gained a lot of experience with this jerry-built jalopy, I will go through a sequence of one or two lectures on a
very high level of abstraction, where I explain what the real scattering theory is like. I do things this way so that you
can get a lot of particular examples under your belt before we fly off into a Never Never Land, or really, an Ever
Ever Land, of abstraction.

I will now explain the incredibly crude approximation we will use. I will take H I(t) and bluntly multiply it by a
function of time f(t, T, Δ) depending on the time, t, and two numbers T and Δ:

This function f(t, T, Δ) will provide a so-called adiabatic turning on and off of the interaction. It will be equal to 1 for
a long time which I will call T, and then it will slowly make its way to zero over a time I will call Δ. This function is
illustrated by Figure 7.3. Why have I stuck this artificial function in my theory? Well, if we think of particle scattering
in a potential, this approximation makes the computation of the S-matrix rather simple. In the far past when f(t, T,
Δ) is zero, the theory is not in some sense asymptotically equal to a free theory, it is exactly equal to a free theory.
So we have a free wave packet going along on its way to the potential. While it’s on its way to the potential, we
turn on the interaction, but it doesn’t know that until it reaches the potential. And then it reaches the potential and
scatters, and goes off in fragments. After the fragments have all flown away, we carefully turn off the potential
again. Again, the wave packet fragments don’t notice that, because they’re away from the potential.
 Figure 7.3: The adiabatic function f (t, T, Δ)

For a scattering of particles in a potential, we have a very simple formula for the S-matrix. We don’t have to
worry about in states and out states, because the in states are the states in the far past, and the out states are the
states in the far future:

In the far past and the far future, f(t) = 0 and H = H 0, and the Hamiltonian that gives the evolution of the
asymptotically simple states, H 0, is the full Hamiltonian, H. So the S-matrix can be written

We want the limits T → ∞ and Δ → ∞. We keep the interaction on for a longer and longer time, and turn it on and
off more and more adiabatically. Δ/T goes to zero in the limit, so at the fringes, the transient terms we would expect
to get from the boundaries are trivial compared to the terms we get from U I(∞, −∞), while keeping the potential on.
The interaction picture is highly suitable to our purposes, because it takes out the factors of eiH0t that are in the
free evolution of the initial and final states. There is no harm to the physics in computing the S-matrix this way for
particle scattering in a potential. This approach may lack something in elegance. Instead of solving the real
problem, you solve a substitute problem with an adiabatic turning on and off function, and then you let the turning
on and off go away. But it certainly corresponds to all the physics we would think would be there.

Here’s why (7.58) is true.9 By the definition of the S-matrix, (7.47), in the Schrödinger picture,

There are two problems with the adiabatic approach. We’ve already talked about the problem with bound
states. The second problem is this: In what sense are the particles really non-interacting when they’re far from
each other? Haven’t we all heard about those virtual photons that surround charged particles, and stuff like that?
Well, we’ll eventually worry about that question in detail, but for now let me say this. In slightly racy language, the
electron without its cloud of photons is called a “bare” electron, and with its cloud of photons, a “dressed” electron.
The scattering process goes like this: In the far past, a bare electron moves freely along. A billion years before it is
to interact, it leisurely dresses itself. Then it moves along for a long time as a dressed electron, briefly interacts
with another (dressed) electron, and moves away for a long time, still dressed. Then it leisurely undresses. For the
time being, though, we will adopt this supremely simple-minded definition, (7.58), of the S-matrix, because it
enables us to make immediate contact with time-dependent perturbation theory, and start computing things.

As we compute things, we will find that indeed this method is too simple-minded. We will have to fix it up
systematically, but we will discover how to do that. Meanwhile, we will be doing lots of calculations and gaining
lots of experience, developing our intuition. And then finally we will junk the Model A altogether, and replace it with
the supreme model of scattering theory. So that is the outline of what we will be doing. Next time, we will begin
exploring our simple-minded model by developing a sequence of algorithms, starting from Dyson’s formula, to
evaluate the U I matrix in terms of diagrams.

1[Eds.]
See Schweber RQFT, Section 11.c, “The Dirac Picture”, pp. 316–325; J. J. Sakurai, S. F. Tuan, ed.,
Modern Quantum Mechanics, rev. ed., Addison-Wesley, 1994, p. 319.
2[Eds.] Freeman J. Dyson, “The Radiation Theories of Tomonaga, Schwinger, and Feynman”, Phys. Rev. 75
(1949) 486–502; see equation (32). (Dyson denotes time ordering in this article by the symbol P; see equation
(29).) Coleman adds, “Without the use of the time ordering notation, this formula for U I(t, t′) was written down by
Dirac 15 years before Dyson wrote it this way.” He is probably referring to P. A. M. Dirac, “The Lagrangian in
Quantum Mechanics”, Phys. Zeits. Sowjetunion 3 (1933) 64–72. Both Dyson’s and Dirac’s articles are reprinted in
Schwinger QED. For the historical background, see Schwinger’s preface to this collection, and Schweber QED. A
careful proof of (7.36) is given in Greiner & Reinhardt FQ, Section 8.3, pp. 215–219.
3[Eds.] See John R. Taylor, Scattering Theory: The Quantum Theory of Non-relativistic Collisions, Dover
Publications (2006), Section 2-c, “The Asymptotic Condition”, pp. 28–31.
4[Eds.]The S-matrix was introduced by John A. Wheeler: “On the Mathematical Description of Light Nuclei by the
Method of Resonating Group Structure”, Phys. Rev. 32 (1937) 1107–1122; see his equation (31). It was extended
and refined by Heisenberg: W. Heisenberg, “Die ‘beobachtbaren Größen’ in der Theorie der Elementarteilchen”
(The “observable sizes” in the theory of elementary particles), Zeits. f. Phys. 120 (1943) 513–538; part II, Zeits. f.
Phys. 120 (1943) 673–702.
5[Eds.]To distinguish between the action and the S-matrix, different fonts are used for these quantities: the action
by S, and the S-matrix by S.
6[Eds.] See note 13, p. 96.
7I didn’t list
the neutron. There’s a reason for that. Neutrons aren’t stable; they last 15 minutes on the average. We
never find, in the very far future, a neutron coming out; we find an electron, a proton and an anti-neutrino coming
out, but not a neutron.
8[Eds.] The Ford Model A, sold from 1927–1931, was the successor to Ford’s Model T automobile.
9[Eds.] See Schweber RQFT, Section 11.c, “The Dirac Picture”, pp. 316–318.

Problems 3

In the first three problems, you are asked to apply the methods of the last few weeks to a non-relativistic field
theory, defined by the Lagrangian

where b is some real number. As your investigation proceeds, you should discover an old friend hiding inside a
new formalism. (This Lagrange density is not real, but that’s all right: The action integral is real; the effect of
complex conjugation is undone by integration by parts.)

3.1 Consider L as defining a classical field theory.

(a) Find the Euler–Lagrange equations.

(b) Find the plane-wave solutions, those for which ψ = ei(p•x−ωt), and find ω as a function of p.

(c) Although this theory is not Lorentz-invariant, it is invariant under spacetime translations and the internal
symmetry transformation

Thus it possesses a conserved energy, a conserved linear momentum, and a conserved charge associated with
the internal symmetry. Find these quantities as integrals of the fields and their derivatives. Fix the sign of b by
demanding the energy be bounded below.

(As explained in class, in dealing with complex fields, you just turn the crank, ignoring the fact that ψ and ψ ∗ are
complex conjugates. Everything should turn out all right in the end: The equation of motion for ψ will be the
complex conjugate of that for ψ ∗, and the conserved quantities will be real. Warning: Even though this is a non-
relativistic problem, our formalism is set up with relativistic conventions. Don’t miss minus signs associated with
raising and lowering spatial indices.)

(1997a 3.1)

3.2 (a) Canonically quantize the theory. (Hint: You may be bothered by the fact that the momentum conjugate to
ψ ∗ vanishes. Don’t be. Because the equations of motion are first-order in time, a complete and independent set of
initial-value data consists of ψ and its conjugate momentum, iψ ∗, alone. It is only on these that you need to
impose the canonical quantization conditions.)

(b) Identify appropriately normalized coefficients in the expansion of the fields in terms of plane wave solutions
with annihilation and/or creation operators.

(c) Write the energy, linear momentum and internal-symmetry charge in terms of these operators. (Normal-order
freely.)

(1997a 3.2)

3.3 For a relativistic complex scalar field, I constructed in class a unitary charge-conjugation operator, U C, a
unitary parity operator, U P, and an anti-unitary time-reversal operator, ΩT, such that

For the theory at hand, only two of these three operators exist. Which two? Construct them (that is to say, define
them in terms of their action on the creation and annihilation operators).

(1997a 3.3)

3.4 The Lagrangian of a free, massless scalar field

possesses a one-parameter family of symmetry transformations, called scale transformations, or dilations,

defined by

(a) Show that the action is invariant under this transformation.

(b) Compute the associated conserved current and the conserved quantity, Q.

(c) Compute the commutator of Q with ϕ, and show that this obeys the assertion, following (5.19), that the
conserved quantity Q is the generator of the infinitesimal transformation:

(d) Compute the commutator of Q with the components of the four-momentum, Pµ, and show that

(You are not required to write things in terms of annihilation and creation operators, nor need you worry about
whether the formal expression for Q should be normal ordered.)1

(1991a 6)

1 [Eds.] In the context of quantum mechanics, the dilation operator Q is represented by xµPµ and is often denoted
D. It is easy to see that [D, Pµ] = iPµ. Because D does not commute with Pµ, it does not commute with P2, and so
only massless theories can have dilation invariance. The dilation operator, together with the Poincaré group
operators {Pµ, Mµν} and four “special conformal operators” {Kµ}, form the 15 parameter conformal group, the
group that leaves invariant the square of a lightlike 4-vector. In addition to the usual Poincaré commutators

we have

The conformal group was discovered in 1909 by Harry Bateman (“The conformal transformations of a space of
four dimensions and their applications to geometrical optics”, Proc. Lond. Math. Soc. 7, s.2, (1909), 70–89), and
later that year was shown by Ebenezer Cunningham to be the largest group of transformations leaving Maxwell’s
equations invariant (“The principle of relativity in electrodynamics and an extension thereof”, Proc. Lond. Math.
Soc. 8, s.2, (1909), 77–98).

Solutions 3

3.1 The Lagrange density L has the form

Treating ψ and ψ ∗ as independent fields, the Euler–Lagrange equations are, for ψ,

and for ψ ∗,

That answers (a). As expected, these equations are complex conjugates of each other. For b < 0 (and we will see
shortly that this condition is necessary) the first equation is nothing but the time-dependent Schrödinger equation
for a free particle of mass m = −1/(2b). To find plane wave solutions, set

and plug into the equations of motion. We find

That answers (b). To answer (c), recall the definition (5.50),

Note that π0ψ = iψ ∗, and π0ψ * = 0. For T00 we obtain

The space integral of T00 gives the Hamiltonian:

The integrand is positive definite. If the energy is to be bounded from below, we have to take b < 0. To make the
analogy with the Schrödinger equation explicit, set b = −1/2m. Then using integration by parts,

which should be familiar as the expectation value of the Hamiltonian for a free particle in the Schrödinger theory.
For ν = i, we find for the momentum density (recall ∇ = ∂i = −∂i)

and so the momentum is

which is the expectation value of the momentum for a free particle in the Schrödinger theory. For the internal
symmetry,

To construct the conserved current, use (5.27),

Here, Fµ = 0 because the Lagrange density is invariant under the symmetry. Then

so

In the usual single-particle Schrödinger equation, the integral of the square of the wave function is used to
determine its normalization. If the norm is constant, probability is conserved. In the language of quantum field
theory, as we will see in the next problem, Q is associated with the number of particles.

3.2 (a) The classical field theory of the L (S3.1) resembles the Schrödinger theory for a single free particle. What
happens in the context of quantum field theory? Since the Euler–Lagrange equations are first-order in time, ψ and
its conjugate momentum π = iψ ∗ form a complete set of initial-value data. Impose the canonical commutation
relations:

(b)Try a Fourier expansion, following (3.45),

where f(p) and g(p) are functions to be determined, and (from (S3.2)) ωp = |p|2/2m. If we assume the relations
(3.18) and (3.19),

then we find

This must equal δ3(x − y) to satisfy the canonical commutation relation (S3.8). In the original expression (3.45),
f(p) = g(p) = 1/(2π)(3/2) and the equal-time commutator of two fields vanishes (4.47), as required. That won’t
do here. There is a clue, however, in the original wording of the problem: “Identify appropriately normalized
coefficients in the expansion of the fields in terms of plane wave solutions with annihilation and/or creation
operators.” We can satisfy the canonical commutation relation by the choice of coefficients

This choice also ensures that (S3.7) holds, because ψ contains only annihilation operators, and ψ ∗ contains only
creation operators.

(c) Obtain the expressions for H, P and Q by plugging in the expressions for ψ and ψ ∗ into (S3.3), (S3.4) and
(S3.6), respectively:
That’s the Hamiltonian. The momentum P goes the same way:

Finally, the charge Q:

This is the theory of a set of free, non-relativistic, identical bosons, all with mass m. Each boson has momentum p
and energy E = ω = |p|2/2m. The conserved charge Q is the number of bosons. Note that all of the operators H, P
and Q are time-independent.

3.3 First, define a parity transformation as in (6.89):

Then

The measure d3p over all momenta is invariant under the reflection p → −p, the energy is a quadratic function of p
and so invariant under reflection, and the last integral is by definition ψ(−x, t). In exactly the same way

There’s no problem with a unitary parity operator, U P. Now for time reversal. Because T : p → −p, it follows that we
should have (recalling that ΩT is anti-unitary)

Then

as desired. The field ψ ∗(x, t) transforms in exactly the same way;

So there’s no problem with an anti-unitary operator ΩT. The problem is with charge conjugation. A unitary charge
conjugation operator U C, if it exists, would transform ψ into ψ ∗, and vice-versa:

The canonical commutation relation (S3.7) says, dividing out the i,

Then
But

This is a contradiction. There is no such operator U C for this theory.

This model is very much like the complex Klein–Gordon theory, with three exceptions: the energy is non-
relativistic, it lacks a charge conjugation operator, and there are no antiparticles. The charge Q counts simply the
number of particles, rather than the difference between particles and antiparticles.

3.4 (a) Let yµ ≡ eλxµ. Then the transformation on ∂µϕ(x) becomes

Thus D: L(x) → e4λL(y). Then the action becomes

Relabeling the dummy variable y to x, the action is manifestly invariant under dilations.

(b) From (5.21),

Then using (5.26) and (4.25),

But ∂µxα = δα µ, so ∂µxµ = 4. Then

where Fµ = xµL. The Noetherian current is defined by (5.27),

and the conserved charge Q is the space integral of the zeroth component,

where π ≡ π0 (see (4.27)).

(c) We need to show [ϕ(y), Q] = iDϕ(y). The charge Q is time-independent, so we can take its time to be the same
as y0, the time of ϕ(y). Then, using the equal-time commutators (3.60) and (3.61),

as required.

(d) We need to calculate [Pµ, Q]. The expression for Pµ, (5.45), is the component Tµ0 of the canonical energy-
momentum tensor density Tµρ given by (5.50), Tµρ = πρ∂µϕ − gρµL, so that
We will need the commutators of Q and the derivatives ∂µϕ. The easiest approach is to differentiate the
commutator [ϕ, Q] because Q is a constant. (Alternatively, we could proceed as in (b), but that is a lot more work.)
Then

and in particular,

Using these relations,

The second term can be written

because Pµ is time-independent, the second integral is a divergence to be transformed into a surface integral at
infinity, and ∂iyi = 3. Then

which was to be shown.

8
Perturbation theory I. Wick diagrams

We begin with the expression for the S-matrix introduced last time, (7.58), written in terms of Dyson’s formula,
(7.36),

We will use good old quantum field theory to evaluate this object, applied to three specific examples, model
theories, which we will discuss at various times throughout these lectures.

8.1Three model field theories

Here are our three models.

MODEL 1:
Model 1 is a scalar field, ϕ(x), interacting with some spacetime-dependent c-number function, ρ(x), which we may
vary experimentally as we wish. I will assume, to make everything simple, that ρ(x) goes to zero as x goes to
infinity in either a spacelike or timelike direction. The variable g is a free parameter called the coupling constant.
I could of course absorb g in ρ(x). But later on, I would like to study what happens if I increase g while keeping ρ(x)
fixed.

I choose this Lagrangian because the field obeys the equation of motion

This equation is very similar to the fundamental equation of electrodynamics in the Lorenz1 gauge,

where Jµ is the electromagnetic current. In the real world, the electromagnetic current is some complicated
function of the fields of charged particles. You’ve seen how to construct them, (5.27). It’s frequently convenient,
however, to consider a simpler problem, where Jµ is just some c-number function under our experimental control.
We could move large charged bodies around on tracks in some classical way, changing the current. This makes
light, photons in the quantum theory. Model 1 describes a theory analogous to the electromagnetic field (which we
don’t yet know how to quantize) in an external current, a scalar field for a meson in an external current.

We know from electromagnetic theory that this current Jµ makes light. We also know that light is photons, so
this current makes photons. We would expect, in the analogous case, that when we wiggle the source ρ(x)—turn it
on and off and shake it around—we should shake off mesons. We will try to compute exactly how many mesons
are shaken off and in what states. This will be our simplest model, because there is no need here to invoke an
adiabatic turning on and off function. The real honest-to-goodness physics of the problem with ρ(x) automatically
turns itself on and off by assumption.

MODEL 2:

Our second model is exactly the same as Model 1, except that we restrict ρ to be a function of x only, independent
of time. Analytically, Model 2 is somewhat simpler, but physically it requires a bit more thought. Again I’ll assume
ρ(x) goes to zero as rapidly as necessary to make any of our integrals converge as x → ∞.

Model 2 is analogous to good old electrostatics:: Given a static charge distribution or a constant current
distribution, compute the electromagnetic field it makes. In Model 2 we have a static source. We don’t know at this
stage what’s going to happen. Maybe mesons will scatter off this static source, and it will act like a potential in
which they move, we’ll see. This problem requires slightly more sophisticated thought, as we will see, because
here we will indeed have to put in a turning on and off function; the physics doesn’t turn itself off.

MODEL 3:

The third model involves two fields, one neutral, ϕ, and one charged, ψ, which is a linear combination of two other
scalar fields, ϕ 1 and ϕ 2, as in (6.23). As the coupling constant g goes to zero, we have three free particles: a
particle and its antiparticle from the terms in ψ ∗ψ, and a single neutral particle from the terms in ϕ 2. In the last
term, we have a coupling between them.

The equation of motion for the ϕ field is

This is beginning to look like the real thing. Aside from the fact that nothing has spin, and I haven’t put in any
derivatives or tensor indices, this is very similar in its algebraic structure to what we would expect for real
electrodynamics. In real electrodynamics, the current Jµ is not prescribed, but is due to the presence of charged
particles. Here the electromagnetic field mimicked by the ϕ field is coupled to a quadratic function in the fields of
the charged particles. If Model 2 can be described as quantum meso-statics, Model 3 is quantum meso-dynamics.

The equation of motion for the ψ field is

This model is also very similar to Yukawa’s theory of the interaction between mesons and nucleons.2 These fields
play an important role in the theory of nuclear forces. And so I will sometimes refer to the ψ and ψ ∗ particles as
nucleons and antinucleons respectively, and the quanta created by the ϕ as mesons. They are of course scalar
nucleons and scalar mesons. Actually we had better not push this theory too far (we’ll only do low orders in
perturbation theory with it). The Hamiltonian contains the term gϕψ ∗ψ, which is not bounded below for either sign
of g.

We will attempt to evaluate the U I matrix (and thus the scattering matrix) in all these cases by Dyson’s
formula, written as

The integral in the exponential is equivalent to ∫ dt H I(t). The interaction Hamiltonian density for Model 1 is

Note that since we are always working in the interaction representation, ϕ(x) = ϕ I(x). For Models 2 and 3, we must
put in the adiabatic function f(t),

For Models 1 and 2, we have to take ρ real in order that H I be Hermitian. In all three cases, we will attempt to
analyze the problem by interaction picture perturbation theory.

So these are the three models we’re going to play with. I should tell you in advance that it will turn out that for
Models 1 and 2, we will be able to sum our perturbation theory and solve the models exactly. That should not
surprise you because the Heisenberg equations of motion are linear, and anything that involves linear equations
of motion is an exactly soluble system.

8.2Wick’s theorem

Our general trick will be an algorithm for turning time ordered products into normal ordered products and some
extra terms. Time ordered products are not defined for every string of field operators; they are only defined for
strings of field operators that have time labels on them. Normal ordered products are not defined for any string of
operators; they are only defined for strings of free fields. Fortunately in Dyson’s formula, we have both things:
operators with time labels, and operators which are free fields. So it makes sense to talk about writing those
things alternatively, in terms of time ordered products and normal ordered products. This is a useful thing to do,
because it’s very easy to compute the matrix elements of normal ordered products once you have them.

For example, consider Model 1, with H I = gρ(x)ϕ(x). At the nth order of perturbation theory, we will have a
string of n ϕ’s. If we sandwich the normal ordering of this string between two-particle states,

the expression must equal zero for n > 4. In that case, each term will contain the product of five operators at least.
Each will have either too many annihilation operators, three or more, or too many creation operators. If the former,
then let the operators act on the state on the right. Two of these get you to the vacuum at best, but the third
annihilates the state. If the latter, the product has too many creation operators, whereupon acting on the state on
the left, the same arguments apply and again the state is annihilated. All the normal ordered products that involve
more than four field operators are of no interest to us.

What happens with (8.13) if n = 4? All that can happen is that the annihilation part of two of the field operators
must annihilate the two initial particles, taking you down to the vacuum, and then two others spit out the two final
particles bringing you back to the final two-particle state. If we can find an algorithm for turning a time ordered
product of operators into a normal ordered product of those operators, plus perhaps some c-number terms, we will
have gone a long way in making the successive terms of this perturbation expansion easier to compute, and
minimizing the amount of operator algebra we have to play with.

Fortunately there is such an algorithm, due to Wick.3 To explain it, I will have to give some definitions. Let A(x)
and B(y) be free fields. (We’re always dealing with free fields, since we’re always in the interaction picture.) Define
an object called the contraction of A(x) and B(y), as the difference between the time ordered product and the
normal ordered product of the two fields:

For free fields, I can prove that the contraction is a c-number. We will evaluate it for the cases we need, that is to
say for two ϕ’s, a ϕ and a ψ, a ψ and a ψ ∗, etc.

To prove the contraction is a c-number I will assume for the moment that x0 > y0. The corresponding formula
when x0 < y0 will follow by the same reasoning. In this case,

Break each field up into its creation and annihilation parts,

where A(−) and B(−) contain each field’s respective creation operators, while A(+) and B(+) contain the annihilation
operators (see the discussion following (3.33) on p. 39). Then

There are four terms in the product A(x)B(x), and three of them are already normal ordered. The only one that is
not normal ordered is the last. Therefore the right-hand side is the normal ordered product, plus a commutator:

The commutator is a c-number (see (3.38)):

A similar argument goes for x0 < y0, so that we can write

That tells us that the contraction is a c-number.

We can write another expression for the contraction simply by taking the ground state expectation value of
(8.14) above:

because, by design, a normal ordered product always has zero vacuum expectation value. Consequently,

By an amazing “coincidence”, the right-hand side of this equation is something which you computed in your first
homework (see Problem 1.3). I will save all of us the time to work it out again, and just remind you

Whenever I write an ϵ in the denominator in the future, you will need to remember that we are to take the
expression in the limit ϵ → 0+. Although you didn’t do this for ψ, it’s essentially the same calculation, and it’s very
easy to see that

You get two equal terms from the ϕ 1 and ϕ 2, but that 2 is canceled by the in the definitions of ψ and ψ ∗. All
other contractions equal zero:

That’s how it goes for two fields. Wick proved the same procedure works for a string of fields. We’ll want two
pieces of notation before diving into the proof. First, suppose we have a normal ordered string of fields, and want
to contract two which are not immediately adjacent. Then

And, just for short, write

With those two conventions established, let’s state Wick’s Theorem:

If n is even, the last terms contain n contractions, otherwise they contain (n − 1) contractions and a single
field. It’s perhaps not surprising that you obtain all the terms on the right-hand side of (8.28). A remarkable and
graceful feature of the theorem is that each term appears exactly once, with coefficient +1.

Proof. The proof proceeds by induction on n. Let W(ϕ 1ϕ 2 · · · ϕ n) denote the right-hand side of (8.28). It is
trivially true that T(ϕ 1) = W(ϕ 1), because both sides simply equal ϕ 1. We’ve already established

By the induction hypothesis, assume T(ϕ 1ϕ 2 · · · ϕ n−1) = W(ϕ 1ϕ 2 · · · ϕ n−1). If we can show T(ϕ 1ϕ 2 · · · ϕ n) =
W(ϕ 1ϕ 2 · · · ϕ n), we’re done. Without loss of generality, we can relabel the fields, such that x01 ≥ x02 ≥ x03 ≥ · · · ≥
x0n, and suppose we have then, by the induction hypothesis,

The job now is to show T(ϕ 1ϕ 2 · · · ϕ n) = W(ϕ 1ϕ 2 · · · ϕ n). Multiply both sides of (8.30) by ϕ 1:

The left-hand side of this equation is T(ϕ 1ϕ 2 · · · ϕ n), because x01 is larger than all the other times. The right-hand
side is

W contains two types of elements, normal ordered strings and contractions. All of the terms in (8.32) are normal
ordered. The first two terms on the right-hand side contain all contractions that do not involve ϕ 1, as well as the
remainder (if any) of uncontracted fields in normal order. Within the commutator, all the purely c-number terms, if
any, will commute with ϕ 1(+). The other terms will produce all the contractions that do involve ϕ 1. Either a
contraction involves ϕ 1, or it does not. Therefore, the right-hand side of (8.32) is a normal ordered series
containing all possible contractions of the n fields: it is equal to W(ϕ 1ϕ 2 · · · ϕ n). QED

I leave it as an exercise to show that Wick’s theorem can also be written in the form

8.3Dyson’s formula expressed in Wick diagrams

Wick’s theorem is very nice, but we are going to find something even better: We’re going to find a diagrammatic
rule for representing every term in the Wick expansion, i.e., the application of Wick’s theorem to the Taylor
expansion of Dyson’s formula (8.1). Instead of having to write complicated contractions, we can just write simple
looking diagrams. These are not yet the famous Feynman diagrams.4 I am introducing these objects ad hoc, to
make the eventual passage to Feynman diagrams (in Chapter 10) as painless as possible. I will call these objects
Wick diagrams. They differ from Feynman diagrams because Wick diagrams represent operators and Feynman
diagrams represent matrix elements. Most textbooks go directly to the Feynman diagrams. I find the
combinatorics gets too complicated that way. I will explain this diagrammatic rule using our third example, Model
3, (8.6), which has the most complicated interaction Hamiltonian, (8.12), of the three models we are considering:

We’ll keep the compressed notation, writing ϕ(x1) as ϕ 1, etc., to simplify things. Dyson’s formula (8.9), the thing we
have to study, is the time-ordered exponential of the expression

A typical term in Dyson’s formula arising in nth order of perturbation theory will involve a product of n copies of this
expression (8.12), integrated over points x1, x2, . . . , xn. Let’s look at the second-order term:

I will draw a diagram, starting with dots labeled 1, 2, and so on, indicating x1, x2 etc. The number of dots is the
order in perturbation theory to which you are going. Associated with each dot, we will draw an arrow going into the
dot, and an arrow going out from the dot, and a line without an arrow on it at all, one end attached to the dot. An
arrow going in corresponds to the factor ψ; an arrow going out is a ψ ∗, and the plain line corresponds to ϕ. We
draw these at point 1 to associate the fields at x1, and similarly for the second dot. In this way I can associate
various terms that occur in the expansion with a pattern of dots, with three lines coming out from each dot.

Figure 8.1: Two points, for the second-order term in Dyson’s formula

Figure 8.2: Two vertices, for the second-order term in Dyson’s formula

What is the prescription for contractions? Whenever two fields are contracted, for x1 and x2, say, we will join
the appropriate lines from those two dots. We can either join a straight line with a straight line if there’s a ϕ − ϕ
contraction, or we can join the head of an arrow with the tail of an arrow if there’s a ψ ∗ − ψ contraction. For
example, the Wick expansion of the second-order contribution (8.35) includes the term

Associated with this term is the diagram in Figure 8.3. The term (8.36) can contribute to a variety of physical
processes. The operator

Figure 8.3: Second-order diagram for Model 3, with ϕ − ϕ contraction

contains, within the ψ field, operators that can annihilate a “nucleon”, N, as well as operators that can create an
“antinucleon”, N, while the operators within the ψ ∗ field can create N and annihilate N (see (6.24).) Consequently
the amplitude

will not be zero, because there are two annihilation operators in the two ψ fields to destroy the two nucleons in the
initial state, and two creation operators in the two ψ ∗ fields to create two nucleons in the final state. The term
(8.36) thus contributes to these reactions:

It cannot contribute to N + N → N + N, which would require the ψ field to create N and the ψ ∗ field to annihilate N.
That’s a good thing, because such a process would break the U(1) symmetry and thus violate charge
conservation. On the other hand, it looks like the operator (8.37) could contribute to the process

which does not violate charge conservation, but it does violate energy-momentum conservation. That would be a
disaster. The coefficient of the term after integrating over x1 and x2 had better turn out to be zero.

As a second example, another term in the Wick expansion of the second-order contribution is

Figure 8.4: Second-order diagram for Model 3, with ψ ∗ − ψ contraction

For the diagram corresponding to this term, see Figure 8.4. Here we have an operator: ψ 1ϕ 1ψ ∗2ϕ 2: containing an
uncontracted ψ, an uncontracted ψ ∗ and two uncontracted ϕ’s. This particular operator could contribute, for
example, to the processes

That is, “nucleon” plus meson (remember, “nucleons” are what our ψ fields annihilate) go to “nucleon” plus meson,
because the operator ψ 1ϕ 1ψ ∗2ϕ 2 contains a nucleon annihilation operator, a meson annihilation operator, a
nucleon creation operator (in ψ ∗) and a meson creation operator. It could also make a contribution to the matrix
elements of the process “antinucleon” plus meson goes into “antinucleon” plus meson, because every term that
contains a nucleon creation operator also contains an antinucleon annihilation operator. Or, for example, it could
contribute to the process where ϕ + ϕ go into N plus N, or N plus N go into ϕ + ϕ, picking annihilation and creation
operators in the right way. Notice that we can’t have N → ϕ + ϕ + N, because of energy-momentum conservation.

Just as we can draw a diagram from the corresponding expression in the Wick expansion, so we can write
down the Wick expansion term from a given diagram. For example, consider the diagram in Figure 8.3. Reading
this diagram and remembering what the theory is, with the rules given earlier about drawing the diagrams, we can
write down what is going on. This is a second-order perturbation diagram because there are two vertices. Each
vertex contributes a term (−ig), and a term of 1/2! comes from the expansion of the exponential. We have d4x1
d4x2 because we’ve got two d4x’s; two vertices. The internal line corresponds to a contraction of the two ϕ
operators, and this is the only contraction. The external lines show two ψ fields (the inward going arrows) and two
ψ ∗ fields (the outward going arrows.) So the remainder of the operator must correspond to the normal ordered
product: ψ ∗1ψ 1ψ ∗2ψ 2: of the “nucleon” field operators. Therefore we recover the associated operator (8.36),

Given any term in the Wick expansion, we can find the corresponding Wick diagram, and vice-versa:

The Wick diagrams are in 1: 1 correspondence with the terms in the Wick expansion.

The entire Wick expansion may be represented by a series of diagrams, every possible diagram, though some
may evaluate to zero. For example, for Model 3, the terms in Wick’s theorem of 17th order consist of all diagrams
with 17 dots, with all lines connecting them drawn in all possible ways, ranging from the first term in Wick’s
theorem, the normal ordered product of 17 × 3 = 51 fields, with no lines connecting the dots, to the second term
with one contraction, diagrams with one line joining two dots, to the third term with two contractions, with two lines
joining the dots, etc. In first-order perturbation theory, the Wick expansion involves a product of three operators,
and has two terms,

and so two diagrams, but both turn out to vanish by energy-momentum conservation. This is a product of three
field operators. The first term has nothing contracted, and vanishes unless we’ve stupidly chosen our meson mass
to be so large it can decay into nucleon and antinucleon. The second term vanishes again by energy-momentum
conservation because you can’t build a one meson state that has the same energy and momentum as the
vacuum state.

Some terms in the Wick expansion contribute nothing. For example, this term

is zero, because the contraction is zero. So we will never write down a diagram like Figure 8.5: the arrows
must line up with the same orientation. That means we can shorten the middle two arrows in Figure 8.4, and
redraw this as in Figure 8.6. Notice only the topological character of these diagrams is important. If I could have
written a term twisted upside down or bent around upon itself, it wouldn’t matter; it would represent the same term.
It’s enough that we represent the three field operators associated with each integration point by an object as
shown in Figure 8.2, and when we contract two field operators, we join their corresponding lines. So we have a
one-to-one correspondence between these diagrams and the terms in Wick’s theorem. Because the terms

Figure 8.5: A forbidden process in Model 3

Figure 8.6: Second-order diagram (Figure 8.4) for Model 3, with ψ ∗ − ψ contraction, redrawn

are distinct, so are their diagrams, as shown in Figure 8.7. After integration over d4x1 and d4x2, however, the
operators corresponding to these diagrams are the same. Just to remind you, these Wick diagrams are not
Feynman diagrams, but they are most of the way to them. Feynman diagrams do not have labeled points, but they
will have labeled momenta on the external lines.

Figure 8.7: Each term in the Wick expansion gets its own diagram

Some Wick diagrams do not have any external lines. Those are the terms where everything is contracted. We
will discover what they mean in the course of time. For example, this term also occurs in second-order
perturbation theory for Model 3:

The appropriate diagram is given in Figure 8.8:

 Figure 8.8: The diagram from the operator : :

Here I have contracted the ϕ at 1 with the ϕ at 2. I can join an undirected line to an undirected line because there
is a non-zero ϕ – ϕ contraction. I can join the head of an arrow to a tail of the arrow because there’s a non-zero ψ
– ψ ∗ contraction. It would be incorrect to draw a diagram in which I connected the head of an arrow to the head of
an arrow because that would be a ψ ∗ – ψ ∗ contraction, which vanishes. You might think that there is a second
diagram, with the labels 1 and 2 switched. But that is exactly the same as Figure 8.8 rotated through 180° in the
plane of the page. There is only one way to contract all the fields. That’s what Wick’s theorem says: Make all
possible contractions. This means simply that we draw diagrams with all possible connections. Diagrams with no
external lines are perhaps a little unexpected, but they’re there because Wick’s theorem tells you they’re there.

8.4Connected and disconnected Wick diagrams

Having given you a headache over Wick’s theorem and then over the diagrammatic representation of Wick’s
theorem, I will now give you even more of a headache by manipulating these diagrams in certain ways. It’s
obvious that if we attempted to compute all these diagrams individually and then sum them up, we would do the
same computation several times. For example, in Figure 8.7, as I emphasized, the diagram on the right, with 1
and 2 interchanged, is not the same as the original diagram on the left: it represents a different term in the
integrand. However the integrals are identical because we end up integrating over x1 and x2, and that will give us
exactly the same answer for both diagrams once we’re done integrating. Remember, we apply Wick’s theorem
before we integrate. Indeed, any other diagram we obtain from a given diagram by merely permuting the indices
will give us the same result, because all that the indices on the vertices tell us is what we call x1 and what we call
x2, and we’re integrating over all of them in the end.

So I will introduce a little more combinatorics notation. Given some diagram D, let the number of vertices be
n(D). I will say that two diagrams D 1 and D 2 are “of the same pattern” if they differ from each other only by a
permutation of the indices on the vertices. The two diagrams in Figure 8.7 are of the same pattern. Within the Wick
expansion of (8.1) are various operators O(D) associated with a particular diagram D. For example, let D be the
diagram in Figure 8.3. The operator O(D) associated with it is (8.36), but multiplied by the factorial 2! for reasons
that will become clear:

For any diagram D of a given pattern and its associated operator O(D), introduce the operator

I’m going to pay special attention to the factor of n(D)!. Factorials are always important in combinatoric discussions
so I write it out in front. There are n(D)! ways of rearranging the indices. This does not mean however that there
are n(D)! different diagrams of the same pattern. It would be lovely if it were so, but it is not so. In the case of
Figure 8.3, there are n(D)! = 2! = 2 different diagrams, because when we exchange 2 and 1 we get a different
diagram. In the case of Figure 8.8 though, there ain’t! I need to introduce the symmetry number, S(D), equal to the
number of permutations of indices that do not change anything. For example, in the case of Figure 8.8,
exchanging the indices 1 and 2 doesn’t change a thing; S(D) = 2. For a second example, consider the diagrams in
Figure 8.9. Diagram (a) is not distinct from diagram (b), or from two other cyclic permutations. But these are
distinct from similar diagrams with non-cyclic permutations. For this diagram, S(D) = 4.

Figure 8.9: A fourth-order contribution in Model 3

 A more complicated example is shown in Figure 8.10. This diagram contributes to nucleon–nucleon scattering
in the sixth order of perturbation theory. This diagram has S(D) equal to 2. There are only two permutations of the
indices that don’t change anything, corresponding to switching all of the bottom indices with all the top indices, or
rotating the diagram about the horizontal dashed line. You see that vertex 1 plays exactly the same role as vertex
2, contract meson at 1 with meson on 2, 5 and 6 play exactly the same role as 4 and 3.

Once I have taken account of this, say by declaring 4 to be the top vertex of the nucleon loop, then all the others
are completely determined. Once I decide which of 4 and 5 is 4 and which is 5, then I have everything labeled
uniquely, and all other permutations of the indices will reproduce different terms in the Wick expansion. You can
play around, if you enjoy these sorts of combinatoric games, trying to invent diagrams with S(D) = 3, or 6, and so
on, for all sorts of things.

Figure 8.10: A sixth-order contribution in Model 3

How many distinct terms do we get with each pattern? There are n(D)!/S(D) terms. If we permute the indices
in all possible ways we get n(D)! different things, but we’re over-counting by S(D). Summing over a whole
pattern—everything of the same pattern as a given diagram—yields

Therefore the n(D)! gets knocked down into simply S(D). Well, it looks a bit complicated but we’ve saved
ourselves labor. If we were really going to compute this diagram, there are 6! different permutations and it would
really be rather stupid to compute all 720 different diagrams.

All the diagrams I’ve written down up to now are connected. “Connected” means (in any theory) that the
diagram is in one piece; all the parts of the diagram are contiguous to at least one other part at a vertex. People
sometimes confuse “connected” with contracted. You can have a connected diagram without a contraction, as
shown in Figure 8.11.

Figure 8.11: A first-order contribution in Model 1

But you can imagine a disconnected diagram. Here is one that arises in fourth order.

Figure 8.12: A fourth-order disconnected graph in Model 1

This is a perfectly reasonable Wick diagram. Anything I can draw, as long as I don’t connect the head of an arrow
with the head of an arrow (or tail to tail), is acceptable. Here is a more complicated diagram with three
disconnected components.

Figure 8.13: A sixth-order disconnected graph in Model 3

Now we come to a marvelous theorem involving Wick diagrams. I will state it first:

I have to define some variables. Let D r(c), r = 1, 2, 3, . . . , ∞ be a complete set of connected diagrams, one of each
pattern. A general diagram D will have some integer nr components of the pattern of D r(c), where the nr’s could be
any non-negative integer. For example, the diagram in Figure 8.13 has two of the nr’s not equal to zero. One of
them, the one corresponding to the connected diagram with vertices 1 and 2, is equal to 2, and the one
corresponding to the connected diagram with vertices 5 and 6 is equal to 1.

I’m going to try to write what a general diagram gives us, from its individual connected parts, in terms of the
operators associated with all the diagrams of each pattern. After all, it is pretty easy. (This will be the last of our
combinatoric exercises.) Consider the graph in Figure 8.13. The operator in the piece containing vertices 1 and 2
has an integral over d4x1 d4x2, and we’ve only got functions of x1 and x2 in the integrand. The next piece goes the
same way, with functions only of x3 and x4, and the final piece likewise with x5 and x6. The entire expression for
the diagram splits into three factors: the diagram yields an operator which is, apart from the combinatoric factor, a
product of other operators. From the first piece we get some operator from doing the x1-x2 integral, some operator
from doing the x3-x4 integral from the second, some operator from doing the x5-x6 integral from the third. So for
this one diagram, we get a single operator squared, and another operator once. That’s characteristic of
disconnected diagrams: the operators associated with them are simply the normal ordered products of the
operators associated with the individual connected components. The contribution for a disconnected diagram D (d)
with connected components D r(c) may be written as

In fact this holds not only for a disconnected diagram D (d), but for a general diagram D. If D is connected, the
product involves only a single term: nr = 1 for that single diagram, and nr = 0 for all other diagrams.

Now, what about the symmetry number S(D) for all the diagrams of a particular pattern? Consider the
combinatoric factor for this single diagram, Figure 8.13. How many permutations can I make that will not change
the diagram? Well, first I could permute the indices. Within each component I can certainly permute the indices
just as if that component were there all by itself. Therefore I get the product on r of 1/S(D r(c))nr. I can do it in the first
component, I can do it in the second, I can do it in the third, but now I can do one thing more. If I have two identical
components, I can bodily exchange the indices in the first component and those in the second, with 1 and 2 as a
block for 3 and 4. That’s an extra permutation. And if I have three identical components, I can do 3! extra
permutations. Therefore I have for the sum of all diagrams D of a particular pattern

We are now in a position to get a very simple expression for the matrix U I which is the sum of all diagrams.
Here in (8.51) I’ve got an expression for a diagram in terms of the operators attached to connected diagrams. The
final stroke, and the end of the combinatorics calisthenics for the moment, is to recognize that U I(∞, −∞) is the
sum over all possible patterns. That is to say,

Now we can commute the sum and the product, to obtain

The sum on each of the nr’s simply gives us the famous formula for the exponential:

Everything is inside the normal ordering symbols so I don’t have to worry about how the operators go. By another
easy manipulation we can write

and thus
 Now we can forget about all of our combinatorics. We have this one wonderful master theorem which is
obviously not special in any way to some particular theory, that the sum of all Wick diagrams, the matrix U I(∞, −∞),
is in fact simply the normal ordered exponential of the sum of the connected diagrams. It was a long journey, but it
was worth it. This is a very nice theorem to have, and it is important. Actually it is more important in statistical
mechanics and condensed matter physics than it is in our present study of quantum field theory. In statistical
mechanics, you study the operator e−βH, and in particular its trace, Tr e−βH, the partition function. The operator
e−βH is, after all, not that different in its algebraic structure from the operator e−iHt. Typically you compute the
partition function in perturbation theory, and then you take its logarithm to get the free energy, the quantity you
really want. This identity, (8.56), is the key to getting a direct perturbative expansion for the free energy, rather
than having to first compute the partition function in perturbation theory, and then compute its logarithm by a
horrible operation. The free energy is just the sum of the connected diagrams.

8.5The exact solution of Model 1

I will now use the formula (8.56) to solve Model 1, whose interaction Hamiltonian density is

where ρ is some spacetime function that goes to zero in all directions as rapidly as we please. In Model 1 there
are also diagrams. The vertices look much simpler. The primitive vertex out of which all diagrams are built is just a
single line with a single vertex because there is only one ϕ field with each H I. I’ll call this diagram D 1.

Figure 8.14: Diagram D 1 in Model 1

This still means we can make a lot of diagrams. For example, I could make a diagram of 42nd order by joining
forty-two of those vertices, one on top of another. The set of Wick diagrams is infinite, but there are only two
connected Wick diagrams. D 1 is the first.

A second diagram, D 2, looks like this:

Figure 8.15: Diagram D 2 in Model 1

If you have a pattern of vertices such that only one line can come out of any one of them, you can only draw two
connected diagrams, D 1 and D 2. Each of them is the only diagram of their pattern. D 1 has only the single figure, so
its symmetry number S(D 1) equals 1. All the diagrams D 1 correspond to the operator

The −ig comes from (8.9). D 2 has its symmetry number equal to 2. That is to say, if you exchange 2 and 1, you get
the same barbell, just flipped around. The operator corresponding to all of the diagrams D 2 is

There are no operators left in O2; it is equal to some complex number −α + iβ which you’ll compute in a homework
problem:

By our general theorem, (8.56), we have a closed form expression for U I(∞, −∞):

This is the complete expression for the S-matrix as a sum of normal ordered terms. The first factor is a complex
number we’ll call A, whose magnitude |A| is an overall normalization constant which I will determine later by a
consistency argument. (We won’t care about its phase.)

As I told you, Model 1 is exactly soluble. There may be fifty ways to solve it exactly. It has linear equations of
motion, and anything with linear equations of motion is essentially an assembly of harmonic oscillators. An
assembly of harmonic oscillators can always be solved by any method you wish. Few are the methods so
powerless that they cannot successfully treat an assembly of harmonic oscillators.

Now let’s evaluate the expression for: exp(O1):. After all, ϕ is a free field, so we know what ϕ is in terms of
annihilation and creation operators, namely our old formula (3.45),

Define the Fourier transform (p) of ρ(x) as (this is the same definition as (1.23))

That is, for a function f(t) of time and a function g(x) of space,

Then (note (−p) = (p)∗)

(Remember, the four components of pµ are not free; p0 equals ωp.) To keep from writing the complicated
expression (8.65) for O1 over and over again, I will write5

where h(p) is defined by

It’s important to observe that if ρ(x) is non-zero but its Fourier transform (p) vanishes on the mass shell,6
when p0 = ωp, then nothin’ happens. This is simply the law of conservation of energy-momentum, and the
diagrammatic observation that the operator O1 makes mesons one at a time.7 The amount of energy and
momentum drawn off from the source must be consistent with the meson energy-momentum relation. If (p,
ωp) is zero, even if O1 has a lot of other Fourier components that aren’t zero, off the mass shell, it’s not going to
be able make a meson. If h(p) is non-zero, O1 can make mesons.

Let’s examine the simplest case. We start out with the vacuum state, turn on our source, wiggle it around,
oscillate it, and mesons come flying out. How many mesons? To answer this, we’ve got to compute U I(∞, −∞) on
the ground state of the free field, because, by assumption, that’s the system’s starting condition. That’s the
experiment we wish to do. This gives us

The ap’s and the normal ordering symbol take care of each other. Because of normal ordering, the ap’s are on the
right where they meet the vacuum and get turned into zero. Only the first term in the sum, equal to 1, survives.
We’re left with just those terms that have nothing but a†p’s in them. The a†p’s all commute with each other, so I no
longer have to write the colon.
 In Chapter 2, I defined a two-particle wave function |p1, p2ñ (see (2.59)). The extension to an n-particle state is
straightforward:

We write the general state |ψñ as

where

Comparing (8.68) with (8.70) we have

But what happened to the factor of 2! in the second term? That disappeared because there are two possibilities for
the two-particle state. Either the first creation operator in the integral creates |p1ñ and the second creates |p2, p1ñ,
or vice-versa: the states |p2, p1ñ and |p1, p2ñ are the same. This symmetry cancels the 2! from the exponential. In
fact, the symmetry cancels the n! factor in the nth term.

The probability P(n) of finding n mesons in the final state is given by

(The divisor n! prevents over-counting.) Substituting in from (8.72),

It is now easy to sum up P(n). Of course, the sum of P(n) over all n must be one; that is the conservation of
probability. Put another way, we demand the unitarity of the S-matrix. Then |ψñ will have norm 1, as it is equal to
S|0ñ, the result of a unitary matrix acting on a ket of norm 1. Therefore

so that

That’s the consistency argument. In (8.62), we defined A = , and thus

Substituting into (8.74), P(n), the probability of finding n particles in the final state, is then given by

the famous Poisson distribution.

Thus we find, in this radiation process, the probability of finding n mesons—what a high-energy physicist
would call the “multiplicity distribution”—is a Poisson distribution. What is the average number of mesons
produced? That’s also an interesting question. Or as we say, what is the mean multiplicity? If you do the
experiment a billion times, what is the average number áNñ of mesons made each time?
That’s just standard fun and games with the Poisson distribution. So this quantity α = ∫ d3p|h(p)|2 is in fact the
mean multiplicity. Because α is proportional to g2, the square of the coupling constant, the probability P(n′) of any
particular number n′ of mesons decreases as g increases, but áNñ increases.

The n-particle states we make are very simple. Well, it is a very simple theory. The n-particle states are all
determined in terms of the one-particle state, and the wave function for the n mesons is just a product of the n
single meson wave functions. It’s as close to an uncorrelated state as you can get, modulo the conditions imposed
by Bose statistics. This kind of state occurs in quantum optics. In the corresponding optical problem, you have
some big piece of charged matter moving up and down. The photon state turns out to be this kind of state, and so
a peculiar optical terminology is used to describe such states: they are called “coherent states”. These are
characteristic not just of classical sources, but of all conditions where the source that is making the mesons or the
photons can be effectively treated as classical. For example, if we have a charged particle passing through
matter, it’s slowed down by the fact that it is ionizing atoms, and hence it gives off a lot of photons. In extreme
cases, these photons produce the so-called Cherenkov radiation. The very energetic photons know that the
charged particle is not just a classical source, because they give it a gigantic recoil whenever it emits one of those
very energetic photons. But from the viewpoint of not so energetic photons, what we call “soft” photons, the piece
of matter is enormously heavy, essentially a classical object that does not recoil. So the soft part of the photon
spectrum emitted in the passage of a charged particle through matter is a coherent state pattern. The bending of a
charged particle in a magnetic field also qualifies as a coherent state pattern.

Coherent states of the harmonic oscillator are

where a† and a are respectively the usual harmonic oscillator raising and lowering operators, (2.17). These states
diagonalize a:

The coherent states |λñ in Model 1 are

These states are also eigenvectors of ϕ +(x) with eigenvalue λϕ :

Except for a factor of 1/n!, the state |λñ has an n particle part which is just the product of n one-particle states. The
expectation values áxñ = áλ|x|λñ and ápñ = áλ|p|λñ oscillate sinusoidally like the classical variables.8

Let’s now compute the average energy, produced in the process where we start off with the vacuum state,
wiggle the scalar source around, turn it off, and then see how many mesons are left. The average energy, the
expectation value of the Hamiltonian in the final state, is

the n! because we don’t want to over-count states. Otherwise we would be counting the state ψ (2)(p1, p2) and the
state ψ (2)(p2, p1) separately. That is a bad thing to do, because they are the same state. The expression (8.84)
can be simplified because everything is symmetric. We can just as well write

in terms of one of the ωp’s, say ωp1, as the others give n − 1 equal contributions, Since the first term is zero, when
n equals zero, I can write the summation from n − 1 = 0 to ∞; the term with n = 0 does not contribute. The integral
is simple to do, because (n − 1) of the integrals give us α, (8.77). So we obtain
Of course, the summation is nothing but a fancy way of writing 1. So we have a simple expression for the mean
energy emitted in our process. It is simply

The mean momentum can be obtained by an identical computation with p’s replacing ωp’s, and that is equal to

This completes for the moment our analysis of Model 1. We’ll return to it later and find out some other things about
it.

In the next lecture I will go on to Model 2, which is also exactly soluble.

1[Eds.]Often rendered “Lorentz”, after Hendrik A. Lorentz (1853–1928), but in fact due to Ludvig V. Lorenz
(1829–1891). See J. D. Jackson and L. B. Okun, “Historical Roots of Gauge Invariance”, Rev. Mod. Phys. (2001)
73, 663–680.
2[Eds.] See note 11, p. 193.
3[Eds.] Gian-Carlo Wick, “The Evaluation of the Collision Matrix”, Phys. Rev. 80 (1950) 268–272.
4[Eds.]In an interview with Charles Weiner of the American Institute of Physics, Richard Feynman said, “I was
working on the self-energy of the electron, and I was making a lot of these pictures to visualize the various terms
and thinking about the various terms, that a moment occurred—I remember distinctly—when I looked at these,
and they looked very funny to me. They were funny-looking pictures. And I did think consciously: Wouldn’t it be
funny if this turns out to be useful, and the Physical Review would be all full of these funny-looking pictures? It
would be very amusing.” Quoted by Schweber QED, p. 434. For a history of the introduction and dispersion of
Feynman diagrams, see David Kaiser, Drawing Theories Apart, U. of Chicago Press, 2005.
5[Eds.] Coleman’s f(p) has been changed to h(p) to avoid confusion with the adiabatic function f(t).
6[Eds.] The mass shell is the four-dimensional hyperboloid p2 = µ2.
7[Eds.]At the beginning of the next lecture, a student asks about this remark. Coleman replies, “I said we could
understand that [the four-momentum restricted to the mass shell value] physically by looking at the structure of the
diagrams, which we could interpret as saying that mesons were made one at a time. If we had had an interaction
like ρϕ 2, then we would have diagrams like this:

with two ϕ’s coming to a single vertex. Then we would not have found the same mass shell constraint, because
you could add the two momenta of these produced mesons together to make practically anything in Fourier space.
It would not be only the value on the mass shell that would be relevant.”
8[Eds.] For more about coherent states, see Problem 4.2, p. 175, and the references at the end of its solution.

Problems 4

4.1 In class we studied

and found an operator for U I(∞, −∞) as the product of a constant, A (which I wrote as e (−α+iβ)), and a known
operator. In class we found α by a self-consistency argument. Find α by evaluating the real part of the relevant
diagram, Figure (8.15) on p. 168:
and show that this agrees with what we found in class. You may find the following formula useful:

(1997a 4.1)

4.2 In solving Model 1 in class, I mentioned the idea of a coherent state.1 Although we won’t use coherent states
much in this course, they do have applications in all sorts of odd corners of physics, and working out their
properties is an instructive exercise in manipulating annihilation and creation operators.

It suffices to study a single harmonic oscillator; the generalization to a free field (= many oscillators) is trivial.
Let cc

and, as usual, let us define

Define the coherent state |zñ by

where z is a complex number and N is a real, positive normalization factor (dependent on z), chosen such that
áz|zñ = 1.

(a) Find N.

(b) Compute áz|z′ñ.

(c) Show that |zñ is an eigenstate of the annihilation operator a, and find its eigenvalue. (Do not be disturbed by
finding non-orthogonal eigenvectors with complex eigenvalues: a is not a Hermitian operator.)

(d) The set of all coherent states for all values of z is obviously complete. Indeed, it is overcomplete: The energy
eigenstates can all be constructed by taking successive derivatives at z = 0, so the coherent states with z in some
small, real interval around the origin are already enough. Show that, despite this, there is an equation that looks
something like a completeness relation, namely

and find the real constants α and β.

(e) Show that if F(p, q) is any polynomial in the two canonical variables,

where and are real numbers. Find and in terms of z and z∗.

(f) The statement that |zñ is an eigenstate of a with known eigenvalue (part (c), above) is, in the q-representation, a
first-order differential equation for áq|zñ, the position-space wave function of |zñ. Solve this equation and find this
wave function. (Don’t bother with normalization factors here.)
(1997a 4.2)

4.3 Let K be a Hermitian operator, and |ψñ a state of norm 1. Given a function f(K) of K, its expectation value in the
state |ψñ is defined by

Suppose we introduce the function η(k) of a real variable k:

Then (as you can easily show)

This works in ordinary quantum mechanics as well as in quantum field theory. Find η(k) for the vacuum state of a
free scalar field of mass m, if
and g(x) is some infinitely differentiable c-number function that goes to zero rapidly at infinity. You should find that
η(k) a Gaussian whose width is proportional to the integral of the square of the Fourier transform of g(x).

H INTS:

(a) Express the delta function as a Fourier transform,

(b) The results of Problem 4.1, and the discussion from (8.62) to (8.77) may be helpful. You may assume β =
0 in (8.62).

Comment: That the answer is a Gaussian should be no surprise. After all, the theory is really just that of an
assembly of uncoupled harmonic oscillators.
(1986a 11)

1[Eds.] Roy J. Glauber, “Photon correlations”, Phys. Rev. Lett. 10 (1963) 83–86. Glauber won the 2005 Nobel
Prize in Physics for research in optical coherence.

Solutions 4

4.1 Recall how α was defined (see (8.59) and (8.60)):

Using the expression (8.23) for the contraction,

because for any complex number z = a + ib, Re(iz) = −Im(z). To make use of the hint (P4.1), note that

Substituting,

By definition, , so
because ρ(x) = ρ(x)∗ and ωp = ω−p. Substituting,

using (8.67), in agreement with (8.77).

4.2 We have to do (a) and (b) at the same time. Let a properly normalized oscillator energy eigenfunction be
denoted |nñ, n an integer. Recall (2.36):

so (a†)n |0ñ = |nñ. Then (we are told to take N to be real)

The inner product of two such states will be

Set the norm of the coherent state vectors áz|zñ equal to 1 to obtain

That answers (a). Then the inner product of two vectors gives

which answers (b).

(c) To show |zñ is an eigenvector of a, recall (2.37),

Then operating with a,

The kets |zñ are eigenvectors of a with eigenvalue z.

A more elegant approach is to recall that for canonically conjugate variables u and v, when [u, v] = 1, then

Since [a, a†] = 1, it follows

(d) The problem states that derivatives of the coherent states |zñ in the neighborhood of z = 0 generate the energy
eigenstates |nñ. Then the |zñ’s form a complete set, because the energy eigenstates are a complete set. In fact, the
|zñ’s are “overcomplete”, because áz|z′ñ ≠ 0 even when z ≠ z′. It isn’t clear that the problem asks us to demonstrate
this first statement, and indeed it’s not straightforward to do so.

2
The difficulty arises because the normalization constant N = e |z| depends on the function |z|2 which does
not have a derivative everywhere. However, its derivative does exist at the origin, and only at the origin, where it
equals zero. 1 If it is permissible to regard all the derivatives of N as equal to zero at the origin, the demonstration
proceeds like this:
 We are now asked to find α and β such that

Write z = x + iy, and use the form (S4.3) for the kets (and the appropriate bras):

Go to polar coordinates: x + iy = reiθ, and dx dy = r dr dθ. Then

The θ integral is

so

Let (β + 1)r2 = u. Then the r integral becomes

Plugging this result in, and using the standard equation 1 = ∑|nñán|, we find

That is, β = 0 and α = 1/π. That answers (d).

(e) Start with a general monomial, pm qn. From (2.19),

we have

for some undetermined coefficient matrix C ij; the exact values do not matter for this argument. Equation (S4.5) is
an identity for any c-number variables x and y;

Then

Any normal-ordered polynomial : F(p, q) : is simply a linear combination ∑ cmn : pm qn : of such monomials, and

That is, = Im z, and = Re z.

(f) The kets |zñ are eigenvectors of a: a |zñ = z |zñ, so

Try the solution áq|zñ = ef(q). Then

Divide out ef(q) to obtain

so

At q = 0, we have á0|zñ = eC, so

(For more about coherent states, see J. J. Sakurai, Modern Quantum Mechanics, Addison-Wesley, 1985, p.
97; Problem 2.18, p. 147, and references therein; D. J. Griffiths, Introduction to Quantum Mechanics, 2nd ed.,
Cambridge U. P., Problem 3.35, p. 127; and W. Greiner, Quantum Mechanics: Special Chapters, Springer, 1998,
Section 1.5, pp. 16–20.)

4.3 Following Hint (a),

which can be written suggestively as

Because there is actually no time-dependent operator in the expression, we can just as well write

Now it so happens that we have already worked out this matrix element, in Model 1 (see (7.59), (8.9), and (8.10)):

(only the zeroth term in the power series for the exponential survives). The form of A comes from (8.62):

and (S4.2) with (8.67),

substituting g → q and ρ(x) → G(x). Using (8.63) and (8.64),

so

Plugging this into (S4.7), and assuming, from Hint (b), that β = 0, we have

which is indeed a Gaussian, whose width σ is proportional to the integral of the square of the Fourier transform of
g(x).

Alternative solution. Let M and N be two operators. The Baker–Campbell–Hausdorff formula2 says
If [M, N] is a c-number, or otherwise commutes with M and N, the formula for Z truncates after three terms, and

The field ϕ (see (3.45)), written more conveniently in terms of ϕ ± (see (3.33)), is ϕ(x) = ϕ +(x) + ϕ −(x), so the
exponent in (S4.6) can be expressed as the sum of two operators:

From (3.38), the commutator is

From (S4.15)

because M includes only a†p, so á0| eM = á0|, and similarly eN |0ñ = |0ñ, and the rest of the problem goes as before.

1 [Eds.] R. V. Churchill and J. W. Brown, Complex Variables and Applications, 4th ed., Mc-Graw Hill, 1984, p. 40.
2[Eds.] Often invoked, rarely cited. See Example 1.2, pp. 20–27 and Exercise 1.3, pp. 27–29 in Greiner &
Reinhardt FQ. The formula predates quantum mechanics by a quarter century. John E. Campbell, “On a law of
combination of operators (second paper)”, Proc. Lond. Math. Soc. 29(1) (1897) 14–32; Henry F. Baker,
“Alternants and continuous groups”, Proc. Lond. Math. Soc. (Ser. 2) 3 (1905) 24–47; Felix Hausdorff, “Die
symbolische Exponentialformel in der Gruppentheorie” (The symbolic exponential formula in group theory), Ber.
Verh. Sächs. Akad. Wiss. Leipzig 58 (1906) 19–48. Reprinted in Hausdorff’s Gesammelte Werke, Band IV,
Analysis, Algebra und Zahlentheorie, Springer, 2002, pp. 431–460. Baker calls commutators “alternants”.

9
Perturbation theory II. Divergences and counterterms

Now I turn to our Model 2, whose Hamiltonian density

is exactly the same as in Model 1, except that ρ(x) is now time-independent. This interaction doesn’t actually turn
off in the far past and the far future. To fit it into our somewhat clumsy formulation of scattering theory, we have to
insert an adiabatic switching function f(t) that turns the interaction on and off by hand:1

The field ϕ(x) is the interaction picture ϕ I(x), but I won’t write the subscript I on the field. A plot of the adiabatic
function f(t) was given earlier, in Figure 7.3, but for convenience I’ll draw it again. The left dashed line occurs at t =
−T/2, and the right dashed line at t = T/2.
 Figure 9.1: The adiabatic function f (t, T, Δ)

The function slowly rises during a time interval Δ from 0 to the value 1 at t = −T/2, stays at 1 until t = T/2, then it
goes down to zero in a way that is supposed to be symmetric with its rise.

9.1The need for a counterterm in Model 2

Something peculiar occurs in this model, and it shows us that we have been a bit too sanguine about the
harmlessness of an adiabatic function’s turning the interaction on and off. If we compute the S-matrix using our
formula, we find, doing the naive calculation, that there are terms that depend on the time T in a nontrivial way,
terms which do not go to zero in the limit T → ∞. We should have

but that’s not what happens when we have the adiabatic function in our interaction Hamiltonian. Let me explain the
physics of why that happens. I will show you how to cure it, and then we will solve the model by summing up the
diagrams.

Let me first introduce some notation. We use |0ñ to represent the ground state of the non-interacting theory.
Therefore H 0 on |0ñ equals zero:

Of course the real physical theory also has a ground state, |0ñP, whose energy E0 is not likely to be zero:

This energy arises in the theory not from the adiabatic function f(t), but just from the extra term H I added to its
Hamiltonian. Here, |0ñP is the actual ground state of the interacting system without the adiabatic f(t), or with f(t) = 1,
if you prefer. Generally when we add an interaction term to a Hamiltonian, not only does the ground state wave
function change, but the ground state energy also changes. So the new ground state |0ñP will have some energy
which I will call E0.

Now let’s make a chart of how Model 2’s ground state evolves in the Schrödinger picture.

We start out with the ground state of the non-interacting theory, at the beginning of time. Up to the time −T/2 −
Δ, nothing has happened because the Hamiltonian H 0 is a non-interacting Hamiltonian, and

The ground state doesn’t even acquire a phase, because its energy is zero. We then slowly turn on the interaction
over a time Δ, to reach its full strength at the time −T/2. By the adiabatic theorem2 we expect the ground state |0ñ
of the non-interacting system to move smoothly from t = −(T/2) + Δ to t = −(T/2) into the ground state |0ñP of the
interacting system with probability 1. I haven’t established any phase conventions for the state, so we might get
the physical vacuum |0ñP with some phase, which I will write simply as e−iγ− where γ− is some real number:

Between −T/2 and T/2, the system evolves in time according to the full interacting Hamiltonian. The state |0ñP is
an eigenstate of the full interacting Hamiltonian, so it gains a new phase, winding up as e−iγ − e−iE 0T|0ñP:

Finally we reach the time T/2, and again the adiabatic hypothesis takes over from t = T/2 to t = T/2 + Δ. The
physical state |0ñP turns back into the state |0ñ associated with the free Hamiltonian, H 0, but with a new phase
factor which I’ll call γ+. The state becomes e−i(γ++γ−+E0T) |0ñ, an exponential factor times the non-interacting
vacuum state, the “bare” vacuum as we sometimes say. This is a straightforward computation in the Schrödinger
picture, using the adiabatic theorem of non-relativistic quantum mechanics, which, if we’re lucky, should be true in
this instance. Incidentally, according to time-reversal invariance, the phases γ− and γ+ should be equal.

The Schrödinger state at time t = −∞ is |0ñ. At time t = ∞, it is e−i(γ++γ−+E0T |0ñ. Writing the state at t = ∞ in
terms of the U matrix, the time-evolution matrix, we find

We have an equation, (7.31), that tells us that U I(t, 0) is e−iH0tU(t, 0). By taking the adjoint, U I(0, t) equals U(0,
t)eiH0t. We see, writing U(∞, −∞) as U(∞, 0)U(0, −∞) that

since |0ñ is an eigenstate of H 0 with eigenvalue zero. Consequently,

Now this is just dumb. In the theory without the artificially introduced f(t), this can’t possibly be the S-matrix
element between the initial ground state and the final ground state. In the real theory, without the f(t), T does not
appear, so you can hardly get an answer that depends on T. The sensible way to define this S-matrix element is
to say that its vacuum expectation value is 1. You start out with a static source with no mesons going in, it just lies
there like a lump. At the end of time, there are no mesons coming out. In this analysis, we have obtained a
spurious phase factor. The origin of that spurious phase factor is my hand-waving argument that when you turn on
the interaction, the system is going adiabatically from the free particle states to the corresponding in states. I
forgot about phases! The states can develop phases. And if we have a mismatch between the vacuum state
energy of the free theory, and the corresponding vacuum state energy of the interacting theory, then we will get a
spurious phase factor, as we have seen. If we can rid ourselves of the mismatch, we’ll eliminate the problem.

Now there’s a very simple way of getting rid of the unwanted phase factor and obtaining a correct theory, by
adding an extra term to our interaction Hamiltonian, called a counterterm. I will eliminate the phase factor for the
ground state and then worry about whether there are corresponding spurious phase factors for the other states of
the theory.3

I write

I have added to the Hamiltonian a new extra term, little a. It’s just a number. It is called a counterterm, because it
is designed to counteract our error. I will choose the value of a so that the phase factor we found in (9.11) is
completely canceled:

In other words, I choose a such as to force

This equation determines the counterterm. Thus a is not a free constant, and I do not have to go beyond the
scattering perturbation theory I have previously developed to compute it. I can just compute it self-consistently,
order by order, in perturbation theory for the U I matrix simply by imposing, in whatever order in computation or
whatever approximation I am doing, this condition (9.14), which fixes a.

Of course, we can also compute a as a by-product of our computation of the S-matrix. That’s interesting,
because in the limit as T goes to infinity,

and therefore
My counterterm a is identified with E0 in the limit of large T. If I happen to be interested in the numerical value of
the ground state energy, I can compute it, because in the limit of large T, a is equal to the ground state energy, E0.

So we’ve done two things with this counterterm. We have eliminated our error in mis-matching the phases,
i.e., mismatching the energies for the ground state, and we have found a way to use the U I matrix to compute the
ground state energy, if we want to do that. Adding the counterterm is a good thing to do. It cures our disease, and
also gives us a bonus, the ground state energy.

9.2Evaluating the S matrix in Model 2

In the case of Model 2, once we have matched the ground state energies for the interacting and non-interacting
systems, there should be no problems for the other states of the theory, because all the other states presumably
consist of 1, 2, 3, 4, etc., meson wave packets impinging on ρ. And if we go to the very far past, those states are
away from ρ, and therefore they should add exactly as much to the energy as they would in a free field theory. On
the other hand, we don’t expect this to happen in Model 3.

In Model 3, the particles are interacting even when they are far away from ρ—there is no ρ in fact, but instead
ψ ∗ψ—and even when they are far away from each other. In that case we should expect an energy mismatch for
the states with real mesons in them as well as for just the ground state. However in Model 2, knock on wood, we
anticipate that this counterterm will take care of all the phase factors caused by any energy mismatch. If it doesn’t,
we will discover that soon enough, as we explicitly compute the S-matrix. If it indeed involves terms that don’t go
to constants as T approaches infinity, I will know that my confident statement was wrong.

I know that many are uncomfortable when I give general arguments. You’ve been trained for years that if the
argument involves an equation, you just accept it, but if it involves words, you don’t understand it. But now we’re
going to do the computation. 4 We have our Hamiltonian, (9.12). We have the condition (9.14) that fixes a. We
remember that as T → ∞, a = E0.

We now have three connected Wick diagrams. Two are exactly the same as in the previous model: D 1, which
we talked about before,

Figure 9.2: Diagram D 1 in Model 2

and D 2, which is just a number. We will calculate it.

Figure 9.3: Diagram D 2 in Model 2

Now, because we have a new term in the Hamiltonian, a, we have a third diagram, D 3, which I’ll represent by a
cross:

Figure 9.4: Diagram D 3 in Model 2 for the counterterm +ia

It doesn’t have any lines on it. Its contribution as a connected diagram is simply +ia, and its symmetry number is 1.

As before, define the operators corresponding to these diagrams. For the first,

For the second diagram,

and finally, for the third,

 As before, the S-matrix can be written

or, somewhat symbolically,

The contributions of D 2 and D 3 are pure numbers, so normal ordering is unnecessary for them. Only these two
diagrams contribute to the vacuum-to-vacuum U I matrix element á0|S|0ñ, given by the exponential of their
contributions. This is, by the definition of a, equal to one, so the contributions of D 2 and D 3 sum to zero:

Therefore, if we are interested in calculating the ground state energy, we just have to calculate D 2. That will fix a,
and a is the ground state energy.

However, if we are not interested in computing the ground state energy, but only the S-matrix element, we
need compute neither D 2 nor D 3, since their sum is zero. This is in general what will happen even if we have a
more complicated theory with such a counterterm. The effect of the counterterm will be to cancel all Wick
diagrams with no external lines, because the sum of all those diagrams makes precisely a phase factor which by
assertion is to be canceled by a.

So to get the S-matrix we need only calculate D 1. Let’s go.

The argument of the exponential, O1, is exactly the same as in Model 1, (8.58), except for the time independence
of ρ(x) and the adiabatic function f(t). Putting in the explicit form (3.45) of ϕ(x), the previous four-dimensional
Fourier transform (8.63) for ρ(x, t) now factors into a three-dimensional Fourier transform and a one-dimensional
one:

There’s our old , now a product of two terms, times ap, just as before, (see (8.65)), plus the Hermitian
conjugate.5

Well, what does this tell us? Look again at the graph of f(t), Figure 9.1: It approaches a constant function
equal to 1 for large T. So for a large T, its Fourier transform, ,

approaches the Fourier transform of 1, or 2π times a delta function. If we plot against

Figure 9.5: The Fourier transform of f(t)

ωp, we’ll get some very highly peaked function with its spread on the order of 1/T, and a height on the order of
2πT, to make the total area equal to 2π. As T grows larger and larger, gets narrower and higher, and
eventually becomes 2π times a delta function concentrated at ωp = 0.

Now this has interesting implications. Since ωp is always greater than µ, goes to zero for any ωp of
interest, because it is concentrated at the origin, and has a spread only O(1/T). Eventually 1/T gets much less than
µ, so
Therefore

As T goes to infinity, the S-matrix goes to the exponential of zero, which is 1. This S-matrix is indeed a unitary
matrix and completely free of dependence on T, as required. Of course it’s physically rather uninteresting. It’s as if
we have this lump ρ(x) sitting there, and we send a meson to scatter off of it, the meson doesn’t scatter! It just
goes right on by . . .

That the Model 2 S-matrix turns out to be equal to 1 can be explained with much the same physical argument
we used to describe the production of mesons in Model 1. Following (8.67), I argued that the Model 1 operator O1
vanishes unless is non-zero on the mass shell . Additionally, in Model 1, mesons were
absorbed or emitted by the source one at a time, because of the corresponding Diagram 1. In Model 2 we have the
same Diagram 1, and we have an example of a non-zero function f(t) whose Fourier transform vanishes on
the mass shell. In fact vanishes everywhere except for a tiny neighborhood of ωp = 0, which does not
include any part of the mass hyperboloid. Again, the mesons are either absorbed or emitted by the source one at a
time. A time independent source like ρ(x) cannot transfer energy; it can only transfer momentum. That means it
can’t absorb or emit a meson, because those processes require energy transfer. A meson always has non-zero
energy. So the S-matrix is identically equal to 1, and there is no scattering in Model 2.

This theory is a complete washout as far as scattering is concerned. While this was easy to see in the
formalism we have built up, it was obscure when people were evaluating this same model theory in the Born
approximation. Not until the discovery of miraculous cancellations of all the fourth-order terms in the Born series
did people realize that they should try to prove the S-matrix for this model was identically equal to 1, to all orders.6

This result holds in the massless case as well. Since there is no scattering for all p ≠ 0, you have only to prove
that the non-vanishing of in the neighborhood of ωp = 0, a set of measure zero, does not screw up wave
packets centered about p = 0.

Even if the S-matrix is uninteresting, we can still compute the ground state energy. That may be interesting.
So let us now turn to that.

9.3Computing the Model 2 ground state energy

Let’s write down the condition that these two diagrams, D 2 and D 3, cancel:

where O2 and O3 are given by (9.18) and (9.19), respectively. Using the identity (see (8.22) and (8.23))

the contribution of D 2 can be written

Let us now go to the limit of large T, because that’s what we have to do to compute the energy. In this limit,
is sharply peaked about ωp = 0. That means in the second integral, we can simply replace ωp with the value
0. With this replacement, the denominator will never equal zero, so we no longer need the iϵ nor the limit, and we
can write

Now we invoke a famous relation, Parseval’s theorem7:

As f(t) has the value 1 for the interval (−T/2, T/2) we can say that its square is also equal to 1 in that region, and

From (9.20) and (9.22), we require O2/2! = −O3, which is, in the limit as T → ∞,

Setting O2/2! = −O3 we have

The T’s cancel, the i’s cancel, so I get a real energy, which is a relief. The ground state energy is given by the
formula

This is in a sense the final and complete answer to our problem. It tells us what the ground state energy is.
Note that the sign is negative, as we should expect. There’s a general theorem that if you add a term to the
Hamiltonian with zero expectation value in the unperturbed ground state, then that always lowers the energy.
That’s a trivial consequence of the variational principle. The term we have added is linear in ϕ, and therefore has
zero expectation value in the unperturbed ground state. If the sign had not come out negative I would have been
very disturbed.

It’s worth a little work to transform this formula (9.36) from momentum space into position space. It can be
written as8

where

I called this model “quantum meso-statics”, because ρ(x) is a sort of classical version of “nucleon density”,
just like the classical charge distributions that enter in electrostatics. So I’ve written the energy of the system in a
form that looks very much like the energy of an electrostatic system:9

The is also there in electrostatics. There is a minus sign in (9.37), whereas in electrostatics there’s a plus sign.
The potential (9.38) represents an attractive potential between our infinitesimal elements of “nucleonic charge”,
rather than the repulsive one as in electrostatics. Also, the integrand 1/(|p|2 + µ2) of (9.38) is not the Fourier
transform of the Coulomb potential 1/|x − y| of electrostatics, but something different, representing the interaction
between two infinitesimal elements of “nucleonic charge”, as opposed to electric charge.

The integral (9.38) for V(x) can be performed in the usual way.10 Let |p| = p, and |x| = r. Then

The last integral can be done by Cauchy’s theorem. The integrand has two poles, at p = ±iµ. Because r is always
positive, I can safely complete the contour of integration in the upper half p plane where the exponential
decreases unbearably rapidly; see Figure 9.6.
 Figure 9.6: Contour of integration for V(r) in Model 2

Now all I have within the contour of integration is a single pole, p = iµ. I can evaluate the integral by Cauchy’s
residue formula:

which is known as the Yukawa potential.11

Thus the infinitesimal elements of this quantity ρ(x), which we have called “nucleonic charge density”, have
an interaction energy proportional to a Yukawa potential. Notice that the singularity of the Yukawa potential at r = 0
is the same is as the singularity of the Coulomb potential at r = 0. Of course, the large r behavior is very different.
The Yukawa potential falls off rapidly with distance, being essentially negligible when r is several times greater
than 1/µ, that is to say, when r is several times greater than the Compton wavelength (h/µc, in conventional units)
of the meson, a meson which doesn’t scatter, but is still responsible for the force between the elements of nuclear
matter.

We could model a two-nucleon system like this:

where the “δ(3)(x)”s are similar to delta functions. They are highly-peaked functions which vanish outside of a
small interval around x. The nucleons are localized in neighborhoods of x1 and x2. Substituting (9.42) and (9.41)
into (9.37),

This force is attractive between like charges, and short-range, and so has some of the essential features of the
real nuclear force. That the force here is attractive turns out to be an example of a general rule: For forces
mediated by the exchange of even-spin particles, like particles attract; for forces mediated by the exchange of
odd-spin particles, like particles repel. This force is mediated by the exchange of zero-spin bosons, so it is
attractive.

Notice also that if we had specified ρ(x) as a point charge (or a collection of point charges),

then just as in electrostatics the energy would be infinite. That’s an important observation:

This divergence is called an ultraviolet divergence, because in p-space it corresponds to the integral blowing up
at high |p|. If ρ(x) is a delta function, then (p) is a constant, and the integral (9.36) blows up like d3p/|p|2.

This divergence, appearing in the term in (9.43) and not depending on the positions of the nucleons, is
nothing to worry about. If nuclear matter need not be an assembly of point particles, then ρ(x) need not be a delta
function. Even if there were some fixed number of point particles, say seven of them moving about on little tracks,
the terms coming from the self-energy of the particles are totally irrelevant. You cannot measure that energy. It
exerts no force. It doesn’t change as you move the particles apart. The only term that you actually measure is the
part that depends on the separation between the particles—the only thing you can adjust—and that part is of
course perfectly finite, if they’re at finite distances from each other.

I wanted to emphasize in this model that first, we get a Yukawa force, and second, we get an ultraviolet
divergence if we go towards the point-particle limit. This may cause us some troubles when we finally get to Model
3, where the interaction is ϕψ ∗ψ without any integrating functions to smear things out. Our nucleons there are not
like the nucleons here. They’re real particles that can recoil and be produced, but they still definitely interact with
the ϕ field at a single point, and therefore we might get an infinite energy shift which we would have to worry
about.

9.4The ground state wave function in Model 2

We can compute not only the ground state energy, but also the ground state wave function, an expansion of the
physical vacuum |0ñP into the basis states |p1, . . . , pnñ, eigenstates of the non-interacting Hamiltonian H 0. Thus
we want to calculate the quantities

This is just an exercise to show that restricting ourselves to time-dependent perturbation theory is not as restrictive
as you might think. We can do all the things we usually do in non-relativistic quantum mechanics with time-
independent perturbation theory. In particular we can construct the ground state wave function.

We use the interaction Hamiltonian

When we studied the interaction turning on and off adiabatically, we said that in the large T, large Δ limit,

That’s just the statement that the U operator up to t = 0, halfway along the way after the interaction has been
turned on, times the bare vacuum |0ñ, equals the physical vacuum |0ñP, times a phase factor. This phase factor is
of no physical interest, and I won’t bother writing it down. Because e−iH0t makes no difference to the ground state,
(9.48) is equivalent to (see (7.31))

Now let us consider, for anything that’s adiabatically turned on,

As usual we’ll consider the limit ϵ → 0+. If we extend f(t) for positive t in the following rather discontinuous way,

(this function is graphed in Figure 9.7) then we can write (9.49) as

and therefore

Figure 9.7: The extended adiabatic function f(t)

Now we know how to compute that. Indeed, we learned how to compute it last lecture, when we were looking at
Model 1. It’s just that now the space and time dependence of the source, ρ(x)f(t), are somewhat peculiar.

This expression (9.53) gives the expansion of the physical ground state in terms of appropriate wave
functions of the non-interacting Hamiltonian, or as we say in our somewhat colorful way, the amplitude for finding
n bare mesons in the physical ground state. That’s the confusing language people use to describe the expansion
of the interacting system’s ground state in energy eigenstates of the non-interacting system. The ground state just
lies there. There are no particles moving around in it.

We can apply the results of Model 1–(8.62), (8.71), and (8.72)–to write
where, analogous to (8.67),

The expression (9.54) is always a product, whatever the form of ρ(x). The Fourier transform of the adiabatic
function (9.51) is

Then the expression for α, analogous to (8.77), becomes

The probability of finding n bare mesons is the probability amplitude squared for the physical ground state
having a component in the n meson subspace of the non-interacting Hamiltonian,

the Poisson distribution we had before.

Something very interesting happens to the expansion (8.70) of the ground state wave function, if we consider
a point particle: ρ(x) goes to a delta function, and (p) becomes a constant. The expansion blows up! The reason
is that α diverges logarithmically (at large |p|):

This isn’t as bad a divergence as the energy, which, as you’ll recall, went at high |p| like d3p/|p|2 ∼ dp. Still, α → ∞
as ρ(x) approaches a delta function. So what do we make of that?

Recall from last time that we found

The average number áNñ of bare mesons in the theory gets very, very large as the source gets more and more
concentrated. On the other hand, the probability P(n) of finding any given number n of bare mesons goes to zero
as the source becomes a point. As ρ(x) goes to a point in position space, or (p) goes to a constant in Fourier
space, α and the peak of the Poisson distribution zoom out towards infinity. That’s disgusting behavior. It’s a good
thing that in the future we won’t worry about computing things like the difference between the ground state
energies of the interacting and non-interacting Hamiltonians for a single particle, or the amplitude for finding the
non-interacting ground state in the interacting ground state. Nobody really should worry about those questions,
because in real models with realistic theories, you don’t have the freedom to turn off the interaction. You don’t
have the freedom to find out what the energy of the one-electron state would be, if there were no electromagnetic
interaction, because, although we give ourselves considerable airs at times, we do not have the power to change
the electromagnetic interaction, say the fine-structure constant, by one jot or tittle.

Fortunately those things which are physically measurable in this theory—for example, the interaction energy
between two separated point charges—do not display such pathological ultraviolet divergences. So, (knock on
wood), maybe we’ll be lucky. Maybe we can get by with our theory of point particles even if it turns out to include
all sorts of disgusting infinities. Perhaps those infinities won’t enter into any physically observable quantities;
perhaps they will. Probably it depends on what the theory is. We’ll have to wait and see.

9.5An infrared divergence

There is another divergence implicit in the integral for α which has nothing to do with what I have been discussing.
It is perhaps best expressed not by thinking of this integral as an example of Model 2, expanding the ground state,
but as an example of Model 1, where we have just chosen a particular form of f(t), one where we turn things on
very slowly and then turn them off abruptly. In this case the formula (9.57) has another kind of divergence, not
dependent on how ρ(x) is distributed. It has to do with the mass of the meson. The formula for α blows up as µ
goes to zero, unless (p) vanishes at |p| = 0:

If µ = 0, then ωp = |p|, and at the low-energy end the integral blows up like d3p/|p|3. For obvious reasons, this is
called an infrared divergence. Since we will eventually have to confront theories of massless particles that are
indeed radiated in interaction processes—in particular we will have to confront the theory of photons—it is
perhaps worth saying a few words about this divergence.

This divergence is also unphysical. Let’s call our massless mesons “photons” for a moment, abusing
language. If we have a source which we build up slowly and turn off abruptly, on the average we will radiate an
infinite number of “photons”. That’s rather silly. This example is very far from being a real photon experiment, but
in a real photon experiment there is a detector, say a photomultiplier tube. You will never read a report that says,
“We observed an infinite number of counts. . . ” Although an infinite number of photons are radiated in this process,
only a finite amount of energy is radiated, because the formula for the expectation value áEñ of the energy

has an extra factor of ωp, as we showed at the end of last lecture. Putting in the factors we have

This integral does not diverge as µ goes to zero. At small |p| it behaves as d3p/|p|2, which is perfectly convergent.

What has happened recalls Zeno’s paradox of Achilles and the tortoise. You have a finite amount of energy to
distribute, but photons are massless. You can give smaller and smaller amounts of energy to each photon. You
could give half the energy to one photon, and a quarter of the energy to another photon, an eighth of the energy to
a third photon, a 16th of the energy to a fourth, and so distribute a finite amount of energy among an infinite
number of photons.

Most of the photons from this infinite number, indeed the overwhelming majority, had arbitrarily low energy.
That means they had very, very long wavelengths. The actual experimental apparatus, a photomultiplier tube or a
radar antenna or anything else at all, however you are detecting your photons, has a low-frequency cut-off. If the
photon is sufficiently soft that the electromagnetic radiation has a sufficiently large wavelength, then you cannot
detect it with any finite experimental apparatus. To detect a photon that has a wavelength of a thousand light
years, you need a radar antenna that is a thousand light years on a side. Those are not found in your average
high-energy physics laboratory! The reason we got an infinite answer again, in the extreme limit µ → 0, is because
we were asking a unphysical question, just as unphysical as asking about the energy of a point source if we
turned off the interaction. These are unphysical questions. How many photons would we detect if we had an
experimental apparatus that could detect any photon, no matter how long its wavelength? That is an impossible
question. If we asked a different question, what is the average number of photons we can detect if our
experimental apparatus can only detect photons of momentum greater than a certain threshold |p|min, then it is
easy to see that in the integral for α we would not integrate all the way down to zero, but just down to our low-
energy experimental cut-off. And then, even as µ went to zero, α would not go to infinity, but to a finite value.12
Once again, we’re saved! It’s a real Perils of Pauline story.13 If we’re sloppy, and ask questions that are
empirically unanswerable, we get, in extreme—but physically reasonable—limits, nonsense answers. If we’re
careful and restrict ourselves only to asking questions corresponding to experiments we can really do, then we get
finite answers, even in those extreme limits.

So far, in our simple theories, the divergences have restricted themselves to unobservable quantities, and
thus kept in quarantine. Such theories are called renormalizable. Whether that situation will prevail when we go to
more complicated theories than the ones we have at hand here, is a question that will be resolved only by future
investigation, which I will begin next lecture, when we start to tackle Model 3.

1[Eds.]
Localized particles are described by wave packets, but because scattering in terms of wave packets is
mathematically awkward, initial and final states are usually represented by plane waves. The use of plane waves
leads to mathematical ambiguities if the interaction does not go to zero sufficiently rapidly as t → ±∞. These
ambiguities are removed by introducing an adiabatic switching function f(t), often of the form e−ϵ|t|. See Lurié P&F,
pp. 213–214.
2[Eds.] Leonard I. Schiff, Quantum Mechanics, 3rd ed., McGraw-Hill, 1968. See Section 35, “Methods for time-
dependent problems”, pp. 279–292.
3[Eds.] There are in principle two reasons to add a counterterm in Models 2 and 3: to deal with the factor T arising
from the adiabatic function, and to ensure that the vacuum energy is the same with and without the interaction.
Neither of these motivations applies to Model 1. By assumption ρ(x, t) → 0 as t → ±∞ all by itself, so there is no
need to add the adiabatic function, and T does not appear. Then, as is evident from (8.85), á0|H 0|0ñ = á0|H|0ñ = 0;
the sum reduces to a single term proportional to n = 0. That is, the Model 1 interaction does not change the
vacuum energy, so a is not needed here, either.
4[Eds.] A student asks: Is the reason why you made a general argument because we’re going to do
renormalization? Coleman replies: “Yeah, we’re going to get there. We’re going to talk about renormalization, in a
little while, or at least part of it. We won’t get to wave function and charge renormalization for a few weeks. But
we’ll talk about mass renormalization.”
5[Eds.] Note that and .
6[Eds.] For an older approach to Models 1 and 2, see Gregor Wentzel, Quantum Theory of Fields, trans. C.
Houtermans and J. M. Jauch, Interscience, 1947, Chap. II, §7, “Real fields with sources”, pp. 37–48; republished
by Dover Publications, 2003.
7[Eds.]Some reserve the name “Parseval’s theorem”’ for the Fourier series version of this theorem, and call the
Fourier integral version “Plancherel’s theorem”. See Gilbert Strang, Introduction to Applied Mathematics,
Wellesley-Cambridge Press, 1986, p. 313; or Philippe Dennery and André Krzywicki, Mathematics for Physicists,
Harper & Row, 1967, p. 224, Theorem 2. Others make no distinction between the discrete and continuous cases,
and call both versions “Parseval’s theorem”, e.g., Philip M. Morse and Herman Feshbach, Methods of Theoretical
Physics, Part I, McGraw-Hill, 1953, p. 456, or Richard Courant and David Hilbert, Methods of Mathematical
Physics, vol. II, Interscience, 1962, p. 794.
8[Eds.] See note 5, p. 189.
9[Eds.] Jackson CE, p. 41, equation (1.52).
10[Eds.] In the video for Lecture 9, Coleman remarks that his students from Physics 251 (the Harvard graduate
course in non-relativistic quantum mechanics) could “probably wake up screaming while doing this integral. But for
the benefit of those of you who have missed that golden experience”, he goes through the calculation in detail,
adding, “this kind of integral is very useful in doing the hydrogen atom and all sorts of such things.”
11[Eds.] Hideki Yukawa (1907–1981), Nobel Prize in Physics 1949. See “On the Interaction of Elementary
Particles. I.”, Proc. Phys.-Math. Soc. Japan 17 (1935) 48–57. Reprinted in D. M. Brink, Nuclear Forces, Pergamon
Press, 1965.
12[Eds.] Coleman is referring to the infra-red divergence. This famous problem is not discussed in this book. The
classic treatment is due to Bloch and Nordsieck: F. Bloch and A. Nordsieck, “Note on the Radiation Field of the
Electron”, Phys. Rev. 50 (1937) 54–59. A fuller explanation is given in J. M. Jauch and F. Rohrlich, The Theory of
Photons and Electrons, 2nd expanded ed., Springer-Verlag, 1976, Section 16-1, pp. 390–405, or Bjorken & Drell
RQM, pp. 162–176, and Bjorken & Drell Fields, pp. 202–207. It should perhaps be mentioned that the first edition
of Jauch and Rohrlich (1955) was among the very first American textbooks to teach the use of Feynman
diagrams; see David Kaiser, Drawing Theories Apart: The dispersion of Feynman diagrams in postwar physics, U.
Chicago Press, 2005, Chapter 7, pp. 253–263.
13[Eds.]A series of short, silent World War I-era movies shown before a full-length feature, with the title heroine in
a succession of grave dangers from week to week, only to be rescued in the nick of time.

Problem 5

5.1 The pair model, invented by G. Wentzel1, is a variant on Model 2 in which there is a bilinear interaction of the
meson field with a time-independent c-number source, instead of a linear one. This is more complicated than
Model 2, but the theory is still exactly soluble, because it is still just a quadratic Hamiltonian. Unlike Model 2, in this
model scattering (but only elastic scattering) can occur.

The Hamiltonian for the theory is of the form H = H 0 + H I, where H 0 is the standard Hamiltonian for a free
scalar field of mass µ. The interaction Hamiltonian H I is

where g is a positive constant, and ρ(x) is some smooth, real function of space only that goes to zero rapidly at
infinity. (Note that the interaction here is not the integral of a local density, but the square of such an integral.)

(a) Compute áp|(S − 1)|p′ñ, the scattering matrix element between (non-relativistically normalized) one-meson
states, by summing up all the connected Wick diagrams (shown below). Start with Dyson’s formula (8.1), and use
Wick’s theorem (8.28) to evaluate the relevant terms. Don’t worry about f(t) or any counterterms.

Show that

where F(ωp) is a function you are to compute in terms of an integral over .

(b) The pair model has no non-vanishing Wick diagrams for one particle going into more than one particle; thus
the S matrix restricted to one-particle initial and final states should be unitary. Explicitly verify this. That is, show
explicitly that S†S = 1 for two one-particle states:

Many comments:

(1) In addition to the diagrams shown, there are also diagrams with no uncontracted fields, (i.e., no external
legs), but you don’t have to worry about them; they’re cancelled in the computation of the S matrix by the ground
state energy counterterm, just as in Model 2.

(2) Note that every vertex in the diagrams represents a seven-dimensional integral: two three-dimensional
spatial integrals and one time integral.

(3) I’ve only drawn one diagram of each pattern. There are others, obtained by permuting the labels 1, 2, . . . ,
n.

(4) Even after you assign the labels, there are still 2n identical terms, because there are two choices at each
vertex, of which field gets contracted which way. This cancels the 1/2n from the nth power of H I in Dyson’s formula.

(5) Don’t get involved with mathematical niceties. Assume that ρ(x) is sufficiently smooth and falls off
sufficiently rapidly as |x| → ±∞ to justify any manipulations you wish to make, that all power series converge, etc.

(6) The answer involves an integral over p defined in terms of , the Fourier transform of ρ(x). It’s not
possible to simplify this integral for general ρ(x); don’t waste your time by trying to do so. On the other hand, if you
have more complicated things than this (double integrals, unsummed infinite series, etc.), you have more to do.

(7) Don’t assume ρ(x) is spherically symmetric.

(1997a 5.1)

1[Eds.]
“Zur Paartheorie der Kernkräfte” (Towards a pair theory of nuclear forces), Helv. Phys. Acta 15 (1942)
111–126.
 Solution 5

5.1 (a) The interaction Hamiltonian is

The matrix element of interest is

From Wick’s Theorem (8.28), the relevant terms are

The contractions are c-numbers, and they can be moved outside the inner product, leaving

In the notation of (3.33),

(recall that the annihilation operators are in ϕ +, and the creation operators in ϕ −). Sandwiched between (p| and
|p′), the first and last terms give zero. We have already accounted for the two different orderings of x and y, so the
normal ordering simply replaces ϕ(x1,t1) by ϕ − (x1,t1), and ϕ(yn,tn) by ϕ + (yn,tn):

We deal with the permutations in two steps. First, from Coleman’s comment (4), we can cancel the factor of 1/2n
(because we can swap xi and yi.) Next, there are n pairings (n − 1 contractions plus one uncontracted pair). These
can be arranged in any order, so there are n! ways. This cancels the factor of 1/n!. Then

From (2.53) and (3.33),

Using (9.29) for the expression of the contractions,

Since

so

Next, we do the space integrals, using (8.64):

(for i = 1, . . . , n − 1). Then

Now do all the time integrals:

These time integrals yield a product of delta functions which simplifies enormously:

because δ(x − a) δ(x − b) = δ(x − a) δ(a − b). This leaves

Now do all the integrals. This sends every to ωp:

The remaining ki integrals are all identical. Define

Then we have

Summing the (assumed convergent!) geometric series gives

Comparing this with (P5.1), we conclude

with G(ωp) given by (S5.6).

Alternate solution (S. Coleman)

The problem can be solved graphically much more quickly. The fundamental vertex is

where

The propagator is, as before,

Then
in agreement with the earlier answer, (S5.8).

(b) We need to show

Some preliminary identities. First,

and

From (S5.8),

so

Using (P4.1) and footnote 8, p. 9 we can write

Because ωp and ωk are both positive, δ(ωp + ωk) = 0. Then

The remaining term of (S5.12) can be expressed as

The right-hand side of (S5.12) is equal to the sum of (S5.13) and (S5.14), but these cancel, so the left-hand side of
(S5.12) is zero. That establishes (P5.2). The pair model S matrix is unitary.

(For more about the pair model, see Schweber RQFT, Section 12c, “Other simple models”, p. 371, and
references therein.)

10
Mass renormalization and Feynman diagrams

I now want to consider our third model,

There are two new features that come up here. First, there are new problems arising in the same way that the
energy-shift problem arose last lecture. These problems are subsumed under the term mass renormalization.
Mass renormalization is unfortunately a term used in two senses, both for the phenomenon that occurs, and for
the prescription we follow to deal with the phenomenon, a prescription for adding counterterms to the Lagrangian.
It’s a peculiar linguistic situation in which the disease and the cure have the same name. The second new feature
is that our Wick graphs will not have the extraordinarily simple structure they had in the previous case. In Model 3
we no longer have a finite family of connected Wick graphs, but an infinite family. We therefore have no hope of
computing the S-matrix in closed form as we did for the previous two examples, at least not by these methods, nor
by any methods known to man or woman. I will not speak about alien life-forms; they may be cleverer.1 All we can
do is settle down with a specific matrix element for a particular scattering process and compute it, order by order in
perturbation theory, until we reach the limits of our computational abilities. It will prove convenient to use not Wick
diagrams, but another kind of diagram, a Feynman diagram (also called a Feynman graph), that represents the
contribution of a Wick operator to a particular matrix element. These two topics will occupy this lecture. I’d first like
to begin by discussing mass renormalization.

10.1Mass renormalization in Model 3

This subject has an interesting history. Let me begin with a rigid sphere immersed in a fluid, a problem considered
by George Green of Nottingham, he for whom Green’s functions are named. He published the results of his
investigation on the motion of a pendulum in an ideal fluid, in 1834, in the Transactions of the Royal Society of
Edinburgh.2 Green’s problem can be posed in the following way.

Suppose I have a rigid sphere of volume V, say a small, spherical zeppelin filled with hydrogen or some other
very light gas, immersed in a perfect fluid of density ρ, with zero viscosity. Let’s say that

a much more reasonable result than the 9g we obtained naïvely.

The physical explanation for this phenomenon was given a decade later in a review article by Stokes,3 well
known as the inventor of Stokes’ theorem. Stokes pointed out that if you imagine a rigid sphere moving through a
fluid with some velocity v, the fluid can’t just stand there, because it’s got to move to get around the sphere. As you
know, there is a pattern of flow set up in the fluid, which you might have looked at in earlier courses. If we were to
calculate the total momentum ptotal of this equilibrium configuration, we would find that the momentum is m0v plus
the fluid momentum, which is expressed in terms of the zeroth and first Legendre polynomials only, as I recall, to
get the velocity potential for the fluid. You integrate this velocity potential to get the momentum in the fluid. If you
do this integral, you obtain an answer mv, where m is defined in (10.3):

so that the sphere has a mass of one tenth of the volume it displaces. Now if we do an elementary statics
calculation on this object, there is a gravitational force m0g pulling downwards, and an Archimedian buoyancy
force 10m0g pushing upwards. If we let go of the sphere, we should observe an upward acceleration of the object,
the net force over its mass, equal to 9g.

Figure 10.1: Sphere in fluid

Now if you’ve ever let go of a ping-pong ball which you have held at the bottom of a swimming pool, or in a
sink, you will know that this is grossly in error. The ping-pong ball does not go up with an acceleration of 9g. You
might at first ascribe this effect to fluid friction. But that can’t be so during the early stage of the motion, because all
such frictional forces are proportional to the velocity. Until the system builds up some substantial velocity, friction
cannot be important. It’s important in the late stages of motion as the ping-pong ball approaches terminal velocity,
but not in the early stages.

Green discovered while he was doing the small vibration problem (which should be good enough for the early
stages of the motion) a remarkable result, which I will quote: “Hence in this last case [of a spherical mass] we shall
have the true time of the pendulum’s vibration, if we suppose it to move in vacuo, and then simply conceive its
mass augmented by half that of an equal volume of the fluid, whilst the moving force with which it is actuated is
diminished by the whole weight of the same volume of fluid.” Green’s result says that there was actually an
effective mass—what we might call the physical mass, the only mass we could measure if we couldn’t take the
ping-pong ball out of the water; if, say, the universe were filled with water. That effective mass m is

and equal to 6m0, if the fluid’s density is ten times the sphere’s. Consequently the ping-pong ball’s acceleration in
the perfect fluid should be
The response of the system to a small external force is the derivative of the total momentum with respect to
the velocity, and you obtain m, not m0. Thus what we have here is a system something like our ping-pong ball,
interacting with a continuum system, in this case an ideal fluid. And we find the mass of the system is changed, as
a result of its interactions with the continuum.4

The next time this idea enters physics is in the electron theory of Lorentz,5 much later in the 19th century, and
Abraham’s work6 on the electron theory of Lorentz. Lorentz thought of the electron as a rigid body of some
characteristic radius r carrying a charge, e. He observed that if you computed the momentum of such a body in
steady motion—nowadays we know about relativity, and we do it more easily just by computing the mass—you
would obtain not only the energy of the body at rest, but also the energy of the attached Coulomb field integrated
over all space. This contribution will be equal to some constant k, depending upon whether it’s a spherical shell of
charge or a uniformly distributed sphere of charge, times e2/r:

If you put an electron on a scale, you will not weigh the electron by itself, but the electron with its associated
electromagnetic field. Your scale tells you the combined mass of the two things:

Likewise, if you attempt to accelerate an electron, you are not only putting the electron into steady motion, you are
moving the associated Coulomb field. You don’t leave the Coulomb field behind when you give the electron a little
push; the field moves with it. Therefore you get not just the momentum of the moving m0, a rigid body, but also the
momentum of the electromagnetic field that moves with it.

Thus in general whenever we have a particle interacting with a continuum system, its mass is changed from
what it would be if the interaction with the continuum system were not present, whether the continuum system is
the classical hydrodynamic field or Maxwell’s electrodynamics. (We didn’t really need this historical introduction; I
just can’t resist talking about Green and Stokes.)

Now let’s consider the theory we have to worry about. I’ll just focus for the moment on the meson mass. We
have a Lagrangian

plus nucleon terms, denoted by the dots, which I won’t bother to write down at the moment. In honor of the
previous discussion, I’ve written the quantity that multiplies ϕ 2 as µ20, rather than µ2, because after all µ0 is the
mass in the absence of any interactions.

Now there is absolutely no reason to believe that in the presence of the interaction, the square of the meson
mass will be µ20. That is unquestionably what the meson mass would be, if the interaction were turned off. We
solved that theory,7 and we found out that the coefficient of ϕ 2 is the meson mass. However, just as interactions
with the hydrodynamic field and the electromagnetic field change the effective mass, respectively, of ping-pong
balls immersed in water, and charged shells within an electromagnetic field, so we would expect the interactions
with the nucleons to change the mass of the meson. If we were able to solve this theory exactly, and we looked at
the one-meson state with momentum zero, we’ve no reason to expect its mass to be µ0. It will be some
dynamically determined number, and it’s a complicated computation to figure out what it is. So the actual, physical
mass of the meson, µ2, is in general not equal to µ20:
 This is not only an interesting phenomenon, it is also a problem for a scattering theory in the same way the
energy mismatch for the vacuum state was a problem. If we arrange to turn on the interaction adiabatically, the
mass, and therefore the energy, of a single-meson state coming in will change. Even an isolated single-meson
state, even if it’s far from anything, will develop a phase just as the vacuum state developed a phase, in the
course of turning the interaction on and off. When we compute the one-meson-to-one-meson S-matrix element,
we should find it equal to 1 for the same reason the vacuum-to-vacuum S-matrix element is 1. If the universe is
empty except for a single meson, the meson is not going to scatter off of anything, it’s just going to go on. In fact
we will not get 1, but instead some preposterous phase factor involving the length of time T during which we’ve
kept the interaction on.

We will avoid that difficulty by introducing counterterms. Consider the following Lagrangian:

so that the interaction Hamiltonian density is

This condition determines b completely. I compute the phase I would get for the one-meson-to-one-meson
amplitude, and I force it to be one. If I have computed that phase to some order in perturbation theory, I’ve fixed b
to that order in perturbation theory. Likewise, this fixes c, again to any order in my expansion:

where |pñ and |p′ñ are one-“nucleon” states. Unfortunately my notation is not well enough developed so that you
can see at a glance whether a given ket is a one-nucleon state or a one-meson state. (I’ll use p’s for nucleon
momenta and q’s for meson momenta, as a visual aid.) Furthermore, not only do these conditions fix these
counterterms, as they are in principle computable quantities, but they allow me to answer questions. For example,
assuming this were a realistic theory of the world, I can ask: what is the bare mass of the meson if I know its
physical mass? How much of its mass is due to its interactions and how much of its mass was given to us by God
before the interactions are turned on? I can compute that bare mass. I see from the terms in the Lagrangian that

So if I want to compute the masses, I have a systematic way of computing them order by order in perturbation
theory.

Note that I’ve written µ2, not µ20, and m2 instead of m20. I still need my old-fashioned vacuum counterterm a (I’ll
come back and make further remarks about that). I added a counterterm bϕ 2, which will have to do with the mass
of the meson, and a counterterm cψ*ψ, to do with the mass of the “nucleon”. It doesn’t matter whether these are
taken to be positive or negative; there is no standard convention about their signs.

The functions of these b and c counterterms are to adjust matters so that the masses of the meson and the
“nucleon” stay the same as I turn on the interaction, just as the function of the a counterterm is to adjust matters so
the energy of the vacuum stays the same when I turn on the interaction. The mass of the meson begins to change,
because there’s an interaction. That won’t bother me. I’ll just turn on b at the same time with just the bare mass,
µ20, keeping in step, so that the physical mass always stays equal to µ2. It’s µ2 when the interaction is off, and µ2
when the interaction is on. I should say it’s µ2 in the average sense. It ranges so that the phase mismatch
integrates to zero, just as for the vacuum state I arrange matters so the phase mismatch integrates to zero. The
same procedure holds for the mass of the “nucleon”.8

I should make one technical remark. Please notice here we have added a to the Hamiltonian density, not to
the Hamiltonian as before. The reason is that the vacuum is a homogeneous system in space, of infinite spatial
extent, so we would expect not to find a finite total energy shift, but instead an energy shift per unit volume, just as
if we had an infinite crystal and changed the strength of the electromagnetic interactions a tiny bit. The energy of
the whole crystal would change by an infinite amount because it’s infinite! It’s the energy per cubic centimeter that
we hope to change by a finite amount, and since this is a spatially infinite system, I have added my counterterm to
the Hamiltonian density rather than to the Hamiltonian.

Now these three additional terms I have added, a, b and c, are of course not free parameters. They are
completely determined. The counterterm a is determined by the requirement that á0|S|0ñ equals one:

I can obtain a to any order in perturbation theory by computing the vacuum-to-vacuum matrix element to that order
in perturbation theory. The b counterterm is determined by the condition that there be no phase mismatch
between one-meson states, |qñ and |q′ñ,

You may wonder: Is this all? Have I gotten rid of all mismatches in phase, energy and mass? Well, of course
we can’t really tell until I do computations, or else put our scattering theory on a firmer foundation than we have
now, with this dumb f(t) function. But it looks plausible. I’ve arranged things so that there’s no energy mismatch
between the interacting vacuum and the bare vacuum; nor between a physical one-meson state and a bare one-
meson state, or between a physical one-nucleon state and a bare one-nucleon state. If I have a scattering state
that’s 32 mesons and 47 nucleons all coming in from the far past, all thousands of light years away from each
other, then the energy of the multiparticle state is simply the sum of the energies of the single particles. That’s an
empirical fact. With these counterterms I have arranged that the energies of the single particles are all coming out
right, so the energy of the multiparticle state should also come out right. It looks like these three counterterms are
sufficient to take care of all of our problems of mismatch. Later on we will discover if this is right or not, after we
have put our scattering theory on a firmer foundation. Then we will see just how many we need.

But for the moment things look good, so keeping my fingers crossed I will run with this Lagrangian. That takes
care of the first topic.

10.2Feynman rules in Model 3

Now for the second topic. We know what our Lagrangian looks like, and now I’m going to talk about the
diagrammatic representation of that Lagrangian. I will now explain what Feynman diagrams are.

I wish to compute matrix elements of the S-matrix between particular states. Remember, every term in the
Wick expansion will in general contribute to many independent scattering processes depending upon whether we
use the loose external lines to create a particle or annihilate a particle. I would now like to write down a different
sort of diagram for matrix elements. For example, let’s consider a process in which a nucleon with momentum p1
plus a nucleon with momentum p2 goes into a nucleon with momentum p′1 plus a nucleon with momentum p′2, to
order O(g2), just for simplicity:

Let’s consider the O(g2) contribution (the lowest non-zero order) to the S-matrix elements

There’s always a 1 term so I’ll just subtract that out. These kets are relativistically normalized states. There will be
a variety of Wick diagrams that may contribute to (10.17). I’ll write down the ones to order g2, neglecting the
effects of the counterterms for the moment. (I’ll talk about the counterterms later.)

Figure 10.2: O(g2) Wick diagram for Model 3 nucleon–nucleon scattering

Just to remind you, the arrow going into a vertex corresponds to a field ψ, the arrow coming out of a vertex
corresponds to a field ψ ∗, and the line at a vertex without an arrow corresponds to a field ϕ.

The term of O(g2) in S − 1 is

Notice that there is no sign of the adiabatic function, f(t). After all the hoopla about f(t), we will (knock heavily on
wood!) simply go to the limit f(t) → 1. I think I’ve taken account of all the residual effects that come from f(t) with my
renormalization counterterms. Later on we will worry a great deal about whether this is legitimate.9

In the original Wick diagrams, it didn’t matter how I had the external lines sticking out from the diagram. The
external lines of the new diagrams will be oriented, following particular conventions. All the fields that are going to
annihilate a particle in the initial state I will put on the right, where the initial state is.10 All the fields that are going to
create a particle in the final state, I’ll put on the left, where the final state is. Then I will label the external lines with
the momentum of the particles they are annihilating and creating. For example, I’ll write down two typical
diagrams, (a) and (b), for this process (there are actually four; the other two, (c) and (d), are obtained by
permuting the vertices in (a) and (b), respectively):

Figure 10.3: O(g2) Momentum diagrams for Model 3 nucleon–nucleon scattering

In Diagram (a), I can use the free nucleon field at 1 to annihilate a nucleon of momentum p1 and the free
nucleon field at 2 to annihilate a nucleon of momentum p2. I can use the free antinucleon field at 1 to create a
nucleon of momentum p′1 and the free antinucleon field at 2 to create a nucleon of momentum p′2. Thus the initial
state |p1, p2ñ goes into the final state |p′1, p′2ñ. There are of course 3 other ways of doing this, even with only this
single Wick diagram. I could for example produce an alternative Diagram (b) with p1 and p2 swapped, where I use
the field at 1 to annihilate the nucleon state with momentum p2 and the field at 2 to annihilate the nucleon state
with momentum p1, the reverse of the previous situation.

I would like to discuss first the combinatoric factors associated with these kinds of diagrams, and second, how
you actually evaluate the diagrams. The combinatoric factors are much simpler than they are for Wick diagrams,
at least for this theory. The reason is very simple. If we look at diagrams of this sort obtained from, say, Diagram
(a) by permuting the indices, we notice that all the vertices are uniquely labeled, assuming for the moment that p1
is not equal to p2, and p′1 is not equal to p′2. (I’m not excluding forward scattering; p1 could equal p′1. All I’m
excluding is scattering at threshold which is after all only a single point, and we can always get to it by continuity.
The two four-momenta are equal near a threshold at the center of mass, where the particles are mutually at rest
with respect to one another.)

For this theory, Model 3, the one over n! in Dyson’s formula is canceled for Feynman diagrams, except if you
are considering a diagram that contains a disconnected component with no external lines; these contribute to
vacuum-to-vacuum amplitudes. Within that disconnected component there’s nothing that’s absorbing anything or
emitting anything, and you may have trouble labeling vertices uniquely. However this isn’t going to trouble us,
mostly because those disconnected components are all summed together by the exponentiation theorem, (8.49),
to make a numerical factor multiplying the whole expression. That factor is supposed to be canceled by the a
counterterm, anyway. So it doesn’t matter if we calculate them correctly or not, as they sum up to zero and we
need never write them down in the first place. If however we want to calculate the energy per unit volume of the
ground state, following our calculation in the last lecture, we do have to keep the combinatoric structure straight.
But if we’re only interested in computing S-matrix elements, all of those things cancel among themselves, and we
don’t have to worry about them.

You can have residual combinatoric factors left over in other theories. For example, a ϕ 4 interaction spells
trouble because we would have identical lines emerging from each vertex, and we would have to think a little bit
more carefully. In ϕ 4 theory there would be four meson lines coming in or out of each vertex, and they all look the
same. You follow one in, and I say, follow the meson line out, and you say, which meson line? There are three
going out. You can get into trouble. But I chose a nice model, without that complication. Fortunately
meson–nucleon theory and quantum electrodynamics, which will occupy us for some time, are very similar in their
combinatoric structure, and in QED also the 1/n! factors disappear. You may, however, have leftover symmetry
numbers even in Feynman diagrams. It depends on the theory. That takes care of the combinatorics.

Now we come to the actual evaluation of these diagrams. If I do one of them, you will see how all of them go.
So let me do Diagram (a). The only term in the Wick expansion of (10.18) that can contribute to two nucleons
scattering into two nucleons (hereafter, NN → NN) is

Vertex 1 is uniquely labeled as the vertex where the nucleon of momentum p1 is absorbed. With a few
exceptions, once I have labeled one vertex uniquely in a diagram, all the other vertices are uniquely labeled. In the
present example, 2 is a vertex you get to by following the meson line from the vertex where p1 is absorbed. If it
were a much more complicated diagram you could just trace through it: follow the nucleon line along the arrow,
follow the nucleon line against the arrow, follow the meson line. As soon as you label one vertex uniquely every
other vertex is labeled uniquely by such a set of instructions. The corresponding diagram we’d get by permuting 1
and 2 would be a different term–still the same term in the Wick expansion, but a completely different contribution,
though numerically equal–and would precisely cancel out the 2! in Dyson’s formula, and in general cancel the n! in
a complicated diagram. Henceforth we erase the labeling on the vertices and just drop the factor of 1/n!. Diagrams
of this sort, without labeled vertices but with labeled ends, are called Feynman diagrams.

The nucleon annihilation operators in ψ(x1) and ψ(x2) have to be used to annihilate the two incoming nucleons
with momenta p1 and p2, and the corresponding creation operators in ψ(x1)* and ψ(x2)* have to be used to create
outgoing nucleons with momenta p′1 and p′2, so as not to get a zero inner product. We have

and so

First we have to compute the amplitude for uncontracted fields absorbing and emitting a meson and a
nucleon. For example (see (2.53) and (6.24)),

That amplitude is very simple, because we have relativistically normalized states, with a factor of (2π)3/2 in
their normalization (see (1.57)). Now you see why I originally put in that factor. I said then that we’d want factors of
2π to come out right in Feynman diagrams. This normalization guarantees that the free field matrix element to
annihilate a single nucleon is simply e−ip·x . The same holds for absorption of a meson, emission of a nucleon,
emission of an antinucleon, etc. Then

The term (x1 ↔ x2) means that two other terms appear in (10.24) that are exactly the same as the two shown, but
with these variables swapped. These terms correspond to Diagram (c) and Diagram (d) (not drawn), identical to
Diagram (a) and Diagram (b), respectively, but with permuted vertices (1 ↔ 2).

The contribution to S − 1 from Diagram (a) is an integral over x1 and x2. Because of this integration, Diagram
(c) is equivalent to Diagram (a) and makes the same contribution. We can thus regard these two diagrams, with
the vertex labels erased, as a single Feynman diagram, and simply drop the 1/2! factor in Dyson’s formula, as we
talked about earlier. That takes care of all the loose, uncontracted fields. I still have the contraction of ϕ(x1) and
ϕ(x2).

Earlier we found a rather simple expression for this contraction, (9.29):

If we insert this expression in (10.19), we notice that a great simplification occurs, because all of the x integrals are
trivial. They simply give us delta functions:
Because δ(x − a)δ(x − b) = δ(b − a)δ(x − b), we can write (10.25) as

All we’re left with is an easy integral over the momentum q of the internal meson line. If we define the invariant
Feynman amplitude Afi by

then the amplitude Afi for the O(g2) contribution to NN → NN scattering can be written

Given a Lagrangian, there is a set of rules for drawing diagrams and associating factors with elements of the
diagrams, to produce amplitudes for physical processes. These are the famous Feynman rules. So let me now
write down (in the box on p. 216) the Feynman rules for this theory, with initial states on the right and final states
on the left.12 Notice that the vertex and the counterterms have energy-momentum conserving delta functions. The
vertex contains a factor of (−ig) coming from the expansion, and the meson and nucleon counterterms contain
factors of ib and ic, respectively. You follow the flow of momentum around the diagram like current flowing around
an electrical circuit. The sum of momenta flowing into a vertex is the sum of momenta that flows out, much like
current into and out of a junction in a circuit. The difference between what flows in and what flows out–which
should be zero–is the argument of the delta function.

With these two terms go two pictures (see Figure 10.4) and two stories. The story that goes with Diagram (a)
is this: A nucleon with momentum p1 comes in and interacts at a point. Out comes a nucleon with momentum p′1
and a “virtual” meson with momentum q. This “virtual” meson then interacts with a nucleon p2, and out comes a
nucleon with momentum p′2. The interaction points x1 and x2 can occur anywhere, and so they are integrated over
all possible values. Furthermore, the virtual meson, unlike a real meson, can have any 4-momentum q, so q is to
be integrated over all possible values, though as you can see from the factor (q2 − µ2 + iϵ)−1, q likes to be on the
mass shell, with q0 = ± The story belonging to Diagram (b) is much the same, except that the roles of
the nucleons with momenta p1 and p2 are reversed.

Figure 10.4: O(g2) Feynman diagrams for Model 3 nucleon–nucleon scattering

Fairy tales like this helped Feynman think about quantum electrodynamics. In our formalism, they are little
more than stories, but in the path integral formulation of quantum mechanics, these fairy tales gain some
justification, as we will see in the second half of the course. The words not only match the pictures, they parallel
the mathematics.11 What we have done for Diagram (a) is completely general. We could have been working out a
much more complicated diagram, of arbitrary complexity.

The a counterterm is just a number. I don’t need a special diagrammatic rule for that. The a counterterm has
no momentum associated with it, and its delta function has an argument of zero. If the system were in a box, the
term (2π)4δ(4)(0) would turn into VT, the volume of spacetime in the box. This counterterm is designed to cancel
all the vacuum bubble diagrams, those without external lines, which you will see also have a factor of δ(4)(0). Like
the counterterm a, the counterterms b and c will be expressed as infinite power series in the coupling constant g. I
will explain to you shortly how we determine them order by order.

A minor technical point here. You might think there should be a factor of in the meson counterterm because
the term in the interaction Hamiltonian (10.11) is − bϕ 2. But you have two possible terms, depending on
which ϕ you are going to contract in the forward direction as you move along the internal line, and which ϕ you are
going to contract in the backward direction. There are always two choices, and those two choices cancel out the
.

That’s it! That’s every diagram for Model 3. To calculate things, you just take what you need from all of this
stuff, stick it together for the appropriate process, and you get a big integral. Everything is fixed by the momentum
on the external lines, which affects the momentum on the internal lines via the delta functions.
Feynman rules for Model 3

1.For external lines momenta are directed

2.Assign a directed momentum to every internal line.

3.

These rules are very simple. They are also cleverly arranged. They enable you directly to compute the S-
matrix element, to any arbitrary order in perturbation theory, between any set of relativistically normalized states,
with any number of incoming mesons and any number of nucleons, and any number of outgoing mesons and
nucleons. Just draw all possible diagrams with the appropriate number of vertices, write down for each of these
diagrams an expression given by these rules, and do the integrals, to the best of your ability. In many cases the
integrals are trivial, and in other cases they are complicated.

So for the time being we can forget Wick’s theorem, we can forget Dyson’s formula, we can forget fields. All
we have is a sequence of rules, like the rules of arithmetic, for computing any contribution to any order for any
scattering process in this field theory. Please notice these rules have been arranged to take care of one of the
important practical problems of theoretical physics: keeping track of the 2π’s. The only factors of 2π which appear
anywhere in these rules are due to a 1/(2π) for every momentum integral, and a 2π for every delta function. There
is no problem keeping track of the 2π’s (there may be some left over).

In most of the diagrams that we have written down so far, all of the internal momenta are fixed by the four-
momentum conserving delta functions. It is often trivial, as it is for Diagram (a), to get rid of all the internal
integrals, and just be left with one delta function expressing overall four-momentum conservation. We should
expect to see such a factor in an S-matrix element for a theory with spacetime translation invariance. On the other
hand, one can write down diagrams, say one of this structure,

Figure 10.5: A O(g4) contribution to N + N → N + N involving a virtual meson → virtual N + N

where the internal momenta are not fixed completely by energy-momentum conservation. The virtual meson splits
into a virtual nucleon and antinucleon, which recombine to make a virtual meson, which then hits the second
original nucleon. I can always add an extra momentum, +p to the right nucleon line and −p to the left antinucleon
line, and everything is still conserved. So diagrams that have internal closed loops will still have residual integrals.
Note that positive momentum flow will often be indicated by the lighter arrows off a line: down ↓ or up ↑, to the left
← or right →. The Feynman arrows are on the lines, and point inwards, toward a vertex for a fermion ψ, and
outwards, away from a vertex, for an anti-fermion ψ ∗.

For any such diagram we can imagine a metaphorical interpretation, and we sometimes attach words to it.
We say these virtual processes conserve energy and momentum because of the four-dimensional delta function
that appears at every vertex. The funny thing about virtual particles is that they need not be on the mass shell.
They can have any four-momentum, and you have to integrate over all possible four-momenta, given the factor13

4.Integrate over all internal momenta (lines belonging to “virtual particles”).

This interpretation is due to Feynman who, by fiddling around, by a miracle, got these rules before Dyson and
Wick. The miracle was genetic: Feynman was a genius. Factors like (10.29) give the probability amplitude, in this
metaphorical language, for the virtual particle going between two vertices, as in Diagram (a). They describe how a
virtual particle propagates from one vertex to another. For this reason, they are called Feynman propagators.
The language, I stress, is purely metaphorical. If you don’t want to use it, don’t use it. But then you’ll find 90% of
the physicists in the world will be unintelligible to you when they give seminars. It’s very convenient, but it should
not be taken too seriously.

We have derived these rules without any talk about virtual particles, or summing up probability amplitudes for
the propagation of virtual particles.14 We’ve derived them just from the standard operations of non-relativistic
quantum mechanics and a lot of combinatorics.15

10.3Feynman diagrams in Model 3 to order g2

I will begin going systematically through all the Feynman diagrams that arise in our model theory to order g2, or at
least those allowed by energy-momentum conservation, one at a time. They come together in families, so it’s not
so tedious. I won’t finish the survey in this lecture. With each there will be a little point of physics I would like to
discuss.

O(g) diagrams. There are only two:

I won’t bother to write the actual values of the external momenta until it is time to compute things. Diagram 1.1
above represents the decay of a meson into a nucleon–antinucleon pair. It is zero, unless we choose the physical
mass µ of the meson to be larger than twice the mass m of the nucleon. I will talk later about what happens when
we do make that choice, but for the moment I will assume µ < 2m. Then our meson is stable, and there is a
genuine asymptotic meson state, and that diagram vanishes by energy-momentum conservation. Nor can any of
the other processes of a real nucleon decaying into a real nucleon and a real meson, all on the mass shell, occur.
These processes are well known to be impossible, no matter how we choose the masses. Diagram 1.2 is even
less likely. It represents the vacuum spontaneously absorbing a meson. That is also equal to zero by energy-
momentum conservation, because of the energy-momentum conserving delta function that comes out in front.
Those are the two diagrams of order g.

O(g2) diagrams. There are twenty-three (or seventeen, accounting for symmetries):

Diagram 2.1 is Diagram 1.1 above, doubled. It’s two mesons decaying and is again equal to zero if µ < 2m. Of
all O(g2) diagrams, this one has the most external lines: six. I’ll now start counting at the other end, and go up from
zero external lines to four.

Diagram 2.2 consists of three separate diagrams, two contributions to the vacuum energy, and the
counterterm:

There are vacuum self-energy corrections from 2.2 (a) and 2.2 (b), but there’s also an a term. These are the
only contributions of order g2 to á0|S|0ñ. The counterterm is fixed by the requirement that these three contributions
sum to zero, so that there are no O(g2) terms in the vacuum-to-vacuum amplitude:

That is to say, the vacuum-to-vacuum amplitude has no corrections. Of course, the a term is an infinite power
series; this equation is just the O(g2) expression of (10.12), and determines the a term only to second order in g. If
I wanted to know the vacuum energy per unit volume, a, to O(g2) I could calculate it from these diagrams.

Onwards! Diagrams with two external lines (mesons first, then nucleons):

The first diagram is interesting if I want to compute the bare mass of the meson to order g2. If I don’t, the
condition that fixes the meson mass renormalization counterterm b is precisely the condition that these two
diagrams sum to zero, to O(g2). If they didn’t, there would be a nonzero correction of order g2 to the one-meson-
to-one-meson matrix element, and there shouldn’t be:

Again, this is just the O(g2) statement of (10.13).

It’s nearly the same story for the nucleon, and the same answer, except that now there are three diagrams.
The first two diagrams taken together will give the bare mass of the nucleon to O(g2). The condition that fixes the
nucleon mass counterterm c is that the contributions of these three terms sum to zero, to O(g2):

(the O(g2) statement of (10.14)).

We’ve gone through a large number of Feynman diagrams with hardly any labor. Now we are left only with
the diagrams of O(g2) with four external lines. Diagrams with four external lines describe seven separate
processes, but we’ll only look at four of these, because the other three can be obtained by a particular symmetry
from the first four. Two of the lines must be incoming and two outgoing, otherwise energy-momentum
conservation will make them vanish trivially. We cannot have a single particle go into three, or the vacuum go into
four. First I’ll write down these processes and then I will write down the diagrams. If you don’t yet have these
Feynman rules in your head, you soon will.

We could have nucleon–nucleon scattering, Figure 10.7:

2.5N + N → N + N

We could also have antinucleon–antinucleon scattering, but that’s connected to nucleon–nucleon scattering by C,
the charge-conjugation operator, since our theory does have charge conjugation invariance:

So I’m not going to bother to discuss antinucleon–antinucleon scattering, since it is diagram for diagram identical
with nucleon–nucleon scattering. We could have nucleon–antinucleon scattering,

2.6N + →N+N
C doesn’t help me much here. We could have nucleon–meson scattering,

2.7N + ϕ → N + ϕ

which is connected by C to antinucleon–antimeson scattering:

I’m only writing down the processes that conserve charge. Remember, the nucleons have charge one, the mesons
have charge zero. And finally, we could have nucleon–antinucleon annihilation into meson plus meson,

2.8 +N→ϕ+ϕ

That process is connected by time reversal, T, to meson plus meson makes a nucleon–antinucleon pair:

So I won’t bother to write that one down.

You may be wondering about the process ϕ + ϕ → ϕ + ϕ, meson-meson scattering. This process occurs in
Model 3, but it is O(g4):

Figure 10.6: Lowest order contribution in Model 3 to ϕ + ϕ → ϕ + ϕ scattering, O(g4)

so we won’t discuss it now.

Thus we have but four processes of O(g2) with four external legs to consider. For each of these, we will find
two Feynman diagrams which we will have to sum up. I would like to write down those two Feynman diagrams for
all four processes, eight Feynman diagrams in all. I will discuss the physical meaning of each term in the
perturbation expansion, because each is interesting.

In order to simplify matters, I will use the notation of (10.27), which I’ll repeat here:

The i is there by convention,16 so our relativistic scattering amplitude Afi will have the same phase of the
amplitude f(θ) as defined in non-relativistic quantum mechanics.17

10.4O(g2) nucleon–nucleon scattering in Model 3

Now let’s look at the diagrams corresponding to process 2.5, nucleon–nucleon scattering. We’ve already looked at
these in Fig. 10.4 and even found the invariant amplitude (10.28). I’ll draw them again, with a variation:

Figure 10.7: O(g2) Feynman diagrams for Model 3 NN → NN

There are the two diagrams I’ve written down before, (a) and (b), and a new one, (b′), you sometimes see.
People tired of writing p’s whenever they draw a diagram sometimes leave the p’s off of these diagrams. They
start with (a), and to let you know two of the p’s are exchanged in the other one, they sometimes write (b′) instead
of (b), stealing a notational device from electrical engineering. The drawing indicates that you’re to put the
momenta in by yourself at the same places on the two diagrams. Then the terms will take care of themselves. The
diagrams (b) and (b′) are the same diagram, just written with the lines twisted around.

Though we’ve already found the invariant amplitude for this process, it’s worth doing again quickly with the
Feynman rules. They give for Diagram (a) the contribution

Momenta are positive leaving a vertex and negative entering. All of our internal momenta are fixed, so there are
no leftover integrations, no leftover delta functions except the one delta function for overall energy-momentum
conservation. The internal momentum q in Diagram (a) is fixed by the delta function to be p1 − p′1 or equivalently
p′2 − p2 (they’re the same) and the internal momentum in Diagram (b) is fixed to be p1 − p′2. The term from
Diagram (b) is added to that of Diagram (a), to give for the amplitude Afi

Dividing out the common factors (except for the i), all that is left in this case is

Both of these diagrams are second order so I have a squared g, and all I have left is the Feynman propagator for
the meson, i/((p1 − p′1)2 − µ2 + iϵ) from the first diagram, and from the second, i/((p1 − p′2)2 − µ2 + iϵ). This
expression for Afi is exactly the same as we found before, (10.28). That’s it! Wasn’t it easy?

I would now like to discuss the meaning of these two terms. After all, relativistic quantum mechanics is, among
other things, supposed to approach non-relativistic quantum mechanics in the low-energy regime. We have all
done, I hope, many Born approximation computations in non-relativistic quantum mechanics. Have we ever seen
an expression like (10.38) before?

Well, it’s easiest to see the connection between our amplitude and the Born approximation in the center-of-
momentum frame. We have Lorentz invariance, so why not use the center-of-momentum frame?

Figure 10.8: NN → NN scattering in the center of momentum

In the center-of-momentum frame, the three-momenta p1 and p2 of the incoming particles are equal and
opposite, so we can write

The four-momenta are thus

The energies of the outgoing particles are the same, and the magnitudes of the outgoing momenta are the same
magnitude as p; they just have a different direction, with a different unit vector e′. That is, the new four-momenta
are

The denominator (p1 − p′1)2 works out like this:

The angle θ is the scattering angle, and Δ = |Δ|, where Δ is the non-relativistic momentum transfer,
The other denominator is

where the non-relativistic cross momentum transfer Δc is defined by

The cross momentum transfer is the momentum transfer that would arise if you considered the particle we have
arbitrarily labeled as 2 as the descendent of the particle we have labeled as 1, rather than 1 being the descendent
of 1.

Substituting this in, we find the invariant amplitude Afi in the center-of-momentum frame,

We can now drop the iϵ because the denominators are positive. All the i’s and minus signs cancel. This is the
same expression as we obtained earlier, (10.28), just written in a special coordinate frame, the center-of-
momentum frame. It should now look much more familiar to you.

People were scattering nucleons off nucleons long before quantum field theory was around, and at low
energies, they could describe scattering processes adequately with non-relativistic quantum mechanics. The non-
relativistic amplitude AfiNR for scattering is proportional to an integral of the potential,

in the lowest order of perturbation theory. This is the famous Born approximation, the lifeblood of non-relativistic
quantum scattering. As we found earlier (see the discussion leading to (9.41)),

What we have obtained as the first term in relativistic perturbation theory, the term of lowest nontrivial order, is
precisely what we would have obtained if we had used non-relativistic perturbation theory to compute the
scattering amplitude for a Yukawa potential, to lowest nontrivial order. This is in perfect agreement with what we
discovered last lecture. We found, to use Feynman’s language, the exchange of a virtual meson between two
static sources ρ(x) produced a Yukawa potential. Here the exchange of a virtual meson between two moving
sources, two actual particles, produces a scattering amplitude that would be produced in this order of perturbation
theory by a Yukawa potential.

What about the second term, where Δ is replaced by Δc ? It, too, has an analog in non-relativistic quantum
theory. In non-relativistic scattering theory involving two identical particles, where you have to take account of the
symmetry, it’s very convenient to introduce something called the exchange operator, , which when acting on a
two-particle wave function exchanges the two particles:

If we consider a non-relativistic scattering problem in which

then we will find

So the term with Δ would come from an ordinary Yukawa potential, and the term with Δc would come from an
exchange Yukawa potential. That we get both a Yukawa potential and an exchange Yukawa potential is not
surprising, because these are identical particles. The scattering amplitude must be invariant under the
interchange of p′1 and p′2. That is to say, if the first term in (10.46) is present, the second term also must be
present, because there is no way of telling apart the configurations in which you have exchanged the particles
from the configurations in which you have not. And since we are working in a formulation of many-particle theory,
quantum field theory of scalar particles, where Bose statistics is automatic, it must automatically come out having
the right symmetry properties. The presence of the first term demands the presence of the second; the interaction
must have the form
That takes care of process 2.5, nucleon–nucleon scattering.

Next lecture I will begin to discuss nucleon–antinucleon scattering with similar arguments. We will find some
things in common and some things different, and continue with the other processes. Then I will discuss some
mysterious connections that exist between these processes; in particular I will discuss crossing and CPT
invariance on the level of second-order perturbation theory. I will then go on to a dull but unfortunately necessary
kinematic exercise of how to connect S-matrix elements to cross-sections, which are what experimentalists
publish, after all...

1[Eds.] Coleman was a passionate fan of science fiction.

2[Eds.]George Green, “Researches on the vibrations of pendulums in fluid media”, Trans. Roy. Soc. Edin. 13
(1834) 54–63. Reprinted in Mathematical Papers of George Green, ed. N. M. Ferrers, AMS/Chelsea Publishing
Company, 1970. Green (1793–1841), by profession a miller and almost entirely self-taught in mathematics and
physics, was completely unknown when he self-published An essay on the application of mathematical analysis to
the theories of electricity and magnetism in 1828. Einstein declared the Essay twenty years ahead of its time. After
the success of his Essay, Green was urged to attend Cambridge University, and did so, entering as an
undergraduate at the age of 39. No portrait or other likeness of Green is known. See D. Mary Cannell, George
Green, Mathematician and Physicist 1793–1841, The Athlone Press, 1993; Julian Schwinger, “The Greening of
Quantum Field Theory: George and I”, https://arxiv.org/pdf/hep-ph/9310283.pdf.
3[Eds.] G. G. Stokes, “Memoir in some cases of fluid motion”, Trans. Camb. Phil. Soc. VIII (1849) 105–137. The
paper was presented on May 29, 1843. Reprinted in v. I of Stokes’ Mathematical and Physical Papers, Cambridge
U.P., 1880.
4[Eds.] See also Lev D. Landau and Evgeni M. Lifshitz, Fluid Mechanics, Pergamon Press, 1966, §11 and the
following Problem 1, pp. 31–36; Kerson Huang, Statistical Mechanics, 2nd ed., John Wiley & Sons, 1987, Section
5.9, “Examples in Hydrodynamics”, pp. 117–119.
5[Eds.] H. A. Lorentz, The Theory of Electrons and its Applications to the Phenomena of Light and Radiant Heat (a
course of lectures delivered at Columbia University, New York, March and April, 1906), 2nd ed., B. G. Teubner,
1916. Reprinted by Dover Publications, 2011.
6[Eds.]For some background on the electron theory of Max Abraham and H. A. Lorentz, see Intellectual Mastery
of Nature, v. 2, C. Jungnickel and R. McCormmach, U. of Chicago Press, 1990, pp. 231–241. Abraham’s revision
of A. Föppl’s influential text Theorie der Elektrizität is regarded as a classic, and was itself revised by R. Becker.
Though an expert on relativity, Abraham believed in the luminiferous aether.
7[Eds.] See §4.4.
8[Eds.] The reader may be wondering if counterterms are to be added at every difficulty, and if the addition of
these terms is going to have unwanted side-effects. In renormalizable theories, the number of counterterms is
finite, and their addition will not alter the physics. Much more will be said about renormalization later in the course,
in Chapters 15, 16, 25, and 33.
9[Eds].See note 3, p. 186. Even if f(t) has been set equal to 1, a is still needed in Model 3 to ensure that the
physical vacuum’s energy is equal to zero.
10[Eds.] Coleman puts initial states on the right and final states on the left, in effect choosing a time axis running
right to left. So his diagrams should be read right to left. Though unconventional, this choice aligns with matrix
elements áf|S|iñ.

Now let’s go on to our next process,

2.6N + N → N + N

the class of diagrams that contribute to NN scattering. As before, there are two, denoted (a) and (b); see Figure
11.1.

Figure 11.1: O(g2) Feynman diagrams for Model 3 NN → NN

 In both diagrams, the incoming nucleon line, drawn at upper right and pointing in, toward the vertex, has
momentum p1; the incoming antinucleon line, at lower right, pointing out, away from the vertex, has momentum p2.
(Remember, I draw a line pointing outward to indicate an incoming ψ ∗ field, which can either create a nucleon or
annihilate an antinucleon. Similarly, I draw a line pointing inward, towards the vertex, to indicate an outgoing ψ ∗
field.) The outgoing nucleon has momentum p′1, and the outgoing antinucleon has momentum p′2. Once again
writing down the graph is mechanical. The internal momentum q is completely determined by the conservation of
energy-momentum at each vertex. In Diagram (a) it is equal to p1 − p′1 or equivalently p′2 − p2. In Diagram (b), it is
equal to p1 + p2 since both p1 and p2 are coming in to that vertex. Thus the amplitude, by the same reasoning as
before, is
As you recall, at the end of last lecture, we began our study of four scattering processes. The first one we
considered was nucleon–nucleon scattering where “nucleon” should be imagined with invisible quotation marks
around it in our model. There we discovered two graphs, each of which had a clear non-relativistic analog. One
corresponded to the Born term for scattering in the direct Yukawa potential, and the other corresponded to the
Born term for scattering in the exchange Yukawa potential. Of course for nucleon–nucleon scattering, if one is
there, the other has to be there, just because of Bose statistics.
11[Eds.]
When Feynman presented his work at the 1948 Pocono conference, Bohr responded that if Feynman
would ascribe classical trajectories to electrons and photons, he had completely misunderstood quantum
mechanics: Schweber QED, pp. 344–345.
12[Eds.] A reminder: in Feynman diagrams, time conventionally flows from left to right. Coleman’s time runs from
right to left.
13[Eds.] Some define the Feynman propagator without the i in the numerator, cf. Bjorken & Drell Fields, p. 42,
equation (12.71). Caveat lector! Note: Coleman wrote (q2) for F(q
2); this notation is used in §15.3. See also
problem 1.3, p. 49.
14[Eds.] Again Coleman foreshadows Feynman’s sum over histories, the path integral formulation of quantum
mechanics; see the aside on p. 656.
15[Eds.] A student asks: “How do we know that all this is right?” Coleman replies: “Sure: experiment. You could
have asked the same question in classical mechanics: How do you know, Mr. Newton, that gravity is proportional
to 1/r2, rather than proportional to r, as Mr. Hooke suggests? It’s unambiguous to check a theory if the couplings
are weak, and you can do perturbation theory. If the couplings are strong, and perturbation theory is useless, then
it’s not at all unambiguous. It’s a very hard job that’s still in progress to try and figure out the theory that explains
the strong interactions. In electrodynamics, at least, we can make predictions, and we can check, by experiment,
that the Lagrangian and the rules we write down describe reality.”
16Iwould be just as happy if the convention were otherwise but I’m not going to change all the literature. I once
gave a course in which I adamantly refused to put in that dumb i by convention, and proved the Optical Theorem
for the real parts of scattering amplitudes. It got pretty silly at the end. I wound up putting the i back in.
17[Eds.] In non-relativistic quantum mechanics, one describes an incoming particle by a plane wave ψ = eikz. Far
from the scattering center, the scattered particle is described by a spherical wave, f(θ)eikr/r. See David J. Griffiths,
Introduction to Quantum Mechanics, 2nd ed., Prentice-Hall, 2004, p. 401, or Landau & Lifshitz, QM, p. 469.

11
Scattering I. Mandelstam variables, CPT and phase space

The four problems in the next assignment1 are all on material that you either already know or will know at the end
of this lecture, or perhaps at the very beginning of next lecture; I’m not quite sure how far we’ll get. The fourth one
has a little interest to it, and the other three are just dull, dumb computation. I encourage you to do them because
the only way you will learn to manipulate Feynman diagrams is by doing one Feynman calculation through from
beginning to end and keeping track of all the π’s and all the other factors we’re going to talk about. As Feynman
said in a lecture when I was a graduate student: “If you say ya understand the subject, but you don’t know where
to put the π’s and minus signs, then you don’t understand nuttin’.”2
 11.1Nucleon–antinucleon scattering

These are the Feynman graphs obtained just by following the Feynman rules. The delta functions all take care of
themselves except for one overall (2π)4δ(4)(p′2 + p′1 − p2 − p1) to conserve energy-momentum, which as you
recall from the end of last lecture I factored out when I defined Afi in (10.27). You should be able to write formulas
down by eye like this yourself. So much for the expression. Now for the interpretation.

The first term has the same interpretation as last time. It corresponds to the non-relativistic Born
approximation for a Yukawa potential of range µ−1. It’s exactly the same as the first term we had for NN scattering,
process 2.5 (see §10.4). Unlike the amplitude for that process, here the second term is not an exchange potential;
it’s not p1 − p′2 or anything like that, it is p1 + p2. Of course since a nucleon and an antinucleon are not identical
particles, there is no reason why a Yukawa potential should be accompanied in this process by an exchange
potential. We can understand its physical meaning if we observe that in the center-of-momentum frame, in which
the total three-momentum is zero,

ET is the total energy of the original two-particle system. So

The first term is the good old first Born approximation, and then there is the almost as good old second Born
approximation summed over a complete set of energy eigenstates. Now, if there is in our unperturbed problem an
isolated energy eigenstate |nñ lying below the threshold at Ep = Ei = En for the continuum states in the scattering,
such that án|V|iñ ≠ 0, then we will get a pole from the second-order formula. (This is an unusual situation in
potential scattering, but not in field theory.) Of course we don’t see this pole in physical scattering; we have to
analytically continue below the physical region. Furthermore, as is obvious from the structure of this expression,
the pole occurs in the partial wave with ℓ = ℓn, which has the same angular momentum as the state |nñ. If I expand
the second approximation out in terms of angular momentum eigenstates and if V is rotationally symmetric, I will
only get a nonzero matrix element if the angular momentum of this state is equal to that of the partial wave I am
studying. The second Born approximation thus reveals that the pole, or at least one of these two poles, at ET = µ,
corresponds to an energy eigenstate.

This contribution to the amplitude has a pole—in fact, two poles, ET = ±µ—which are presumably below the
threshold for creating a nucleon–antinucleon pair. As µ < 2m, the denominator can never equal zero, and we need
not worry about the iϵ.

We have not talked about partial-wave analysis. If I did a partial-wave decomposition of nucleon–antinucleon
scattering, these poles would occur in the s-wave amplitude. You don’t have to know much about Legendre
polynomials to understand why that is so: 1/ET2 is rotationally invariant, so the only factor of Pℓ(cos θ) which
occurs is for ℓ equals zero, the constant Legendre polynomial. The amplitude has no angular dependence at all,
and therefore it is pure s-wave.

Now, do we encounter such poles in non-relativistic perturbation theory? The answer is yes, we do; typically
not in lowest order but in second order. In the non-relativistic formula (10.47) I wrote Afi as proportional to the first
Born approximation, (“proportional”, because we hadn’t worked out the kinematic factors, yet). To the second
approximation,3

This is the non-relativistic analog of the pole in the relativistic scattering at ET = µ. It is exactly what we would
get in a non-relativistic problem in which there was an isolated energy eigenstate, in addition to continuum states,
before we turned on our potential V. The pole at ET = −µ that comes along with the pole ET = µ is without a non-
relativistic analog, but that’s not surprising. After all, if I make µ very close to 2m, the pole at ET = µ might well be
within the expected domain of validity of non-relativistic physics, but the other pole at ET = −µ is at least 2µ below
threshold, or in non-relativistic units, 2µc2 below threshold, quite a long distance out to trust non-relativistic
physics. Once again this second term is, aside from various kinematic factors, not a novel phenomenon of
relativistic theory, but simply a conventional energy-eigenstate pole, which has precisely the location, and
precisely the angular dependence, that one would expect from the non-relativistic formula.

Up to this stage, we have found nothing new. We’ve learned things, such as the right ways to generalize
some non-relativistic phenomena to a particular relativistic problem, but we have found no relativistic phenomena
that are without non-relativistic analogs. Now we go on to the next process. Again we will find nothing
fundamentally new—just Yukawa potential-like terms, exchange Yukawa potential-like terms, and energy
eigenstate pole-like terms. Let’s work it out.

11.2Nucleon–meson scattering and meson pair creation

There are two processes left, and we’re going to give the last, N + N → ϕ + ϕ, very little attention. Next to last is
nucleon–meson scattering:

2.7N + ϕ → N + ϕ

Here are the diagrams that contribute to nucleon–meson scattering. Antinucleon–meson scattering is trivially
related to this by charge conjugation, and the amplitudes are the same. There are two diagrams, though the
second can be written as either (b) or (b’):

Figure 11.2: O(g2) Feynman diagrams for Model 3 Nϕ → Nϕ

Once again the internal momentum is fixed, p1 + p2 for Diagram (a) and p1 − p′2 for Diagram (b). The invariant
amplitude is

Please notice that the propagator mass is m2, not µ2 this time, because it is an internal nucleon line, not an
internal meson line.

Now I need not belabor the first graph, Diagram (a), which describes an energy-eigenstate pole, just like the
earlier Diagram (b) in nucleon–antinucleon scattering (Figure 11.1). The only difference is that now the energy
eigenstate is a nucleon, appearing in the meson–nucleon channel, rather than the meson appearing in the
nucleon–antinucleon channel. The arguments are, except for replacing µ by m, word for word the same as those I
have just given.

The second graph, Diagram (b) in Figure 11.2, looks like an exchange Yukawa potential. It’s rather odd in
terms of non-relativistic scattering theory to see an exchange potential without a direct potential, but after all,
mesons and nucleons are not identical particles. If nucleon and antinucleon can have a direct potential without an
exchange potential, apparently meson and nucleon can have an exchange potential without a direct one. It is
slightly different kinematically from the exchange potentials we discussed in the cases of nucleon–nucleon and
nucleon–antinucleon scattering, because its range in the center-of-momentum frame is energy dependent. As I
will now demonstrate, this arises because the meson and nucleon have different masses.

In the center-of-momentum frame,

The four-momenta are

The denominator of the exchange term then becomes

 In nucleon–nucleon scattering, we had

That is,
where, as in (10.45), Δc = p1 − p′2 is the cross momentum transfer. Unlike the case of nucleon–nucleon
scattering, the energy terms in meson–nucleon scattering do not cancel, because µ ≠ m. This affects the range of
the potential, because it is dependent on the Fourier transform of the amplitude.

so that the reciprocal of the mass µ serves as a range parameter, R. In nucleon–meson scattering, however,

The reciprocal of R 2, formerly µ2, is now m2 − . Consider the limits:

The inverse of the range parameter, R −1, goes to m as p → ∞, as if two equal-mass particles were exchanging an
object of mass m, just as in the usual Yukawa potential. On the other hand,

R −1 becomes smaller than m as p → 0. (The right hand side is positive, because we have chosen 2m > µ. If this
were not the case, the meson would not be stable.) Thus we have an exchange potential with an energy-
dependent range, a novelty from the point of non-relativistic physics.

If we attempt to give some reality to this system, neglecting the fact that real nucleons have spin, by imagining
the nucleon–meson mass ratio is that of a real nucleon and a real π meson, that is to say something like 7:1, R −2
of the potential at low energies would be on the order of

The Yukawa potential for nucleon–π meson scattering at high energies, ignoring spin, would go roughly as

so the potential has much longer range at low energy than at high energy. At any given energy it’s like the Born
approximation for a Yukawa potential, but the value of this R parameter changes with the energy. It’s a purely
kinematic effect; there’s no mystery to it. In the spinless case, this might have a significant effect on low energy
meson–nucleon dynamics. In the spinless case, an exchange potential has the same effect as a direct potential in
even partial waves, but the opposite effect in odd partial waves. Therefore we have a potential of rather long range
in this problem which is attractive in sign for the even partial waves, and repulsive in sign for the odd partial
waves. If we wish to be a little ambitious, and imagine we could turn up the potential just a slight amount while still
using these ideas from perturbation theory—a dangerous step, but let’s take it—we would expect in this case,
because the potential is an exchange potential, to perhaps begin seeing bound states in the even partial waves,
but never in the odd. If it were a direct potential, we would of course see bound states in all partial waves.

This is in fact very close to the dynamics of actual meson–nucleon low energy scattering for complicated
reasons that we won’t get to until quite late in the course. There is a potential between meson and nucleon caused
by the exchange of a nucleon, rather like a Yukawa potential of quite long range at low energy, because µ is a very
small number compared to m. When we take account of all the spin and isospin factors, it turns out to be attractive
in odd partial waves and repulsive in even, and it isn’t quite strong enough to make a bound state, but it is strong
enough to make a nearly-bound state, or resonance, which is the famous Δ or N ∗ resonance: an unstable p-wave
state in the pion–nucleon system with a mass of 1232 GeV. We’ve now got about half the physics required to
establish that. The parts we don’t have are the kinematics involving spin, which we will get to in the course of time,
and a good reason why we should trust lowest-order perturbation theory. We’ll see not too much later why we
should trust it, at least for the long-range part of the potential, because it doesn’t get corrected. I have gotten
ahead of my systematic analysis of lowest-order Feynman graphs, but I thought I would give you a taste of future
physics.

The eighth process, and the last, of this class of O(g2) Feynman graphs, is nucleon plus antinucleon goes into
two mesons:

2.8N + N → ϕ + ϕ

I won’t bother to treat this in detail, since this process involves absolutely no novel features, and anyway I’ve given
this to you as a homework problem (Problem 6.3, p. 261).

There are two graphs, shown in Figure 11.3. The second is the same as the first, with p′2 and p′1
interchanged. We can now just stare at these from what we have learned already without even writing down the
expression and say, “Aha!” Diagram (a) is a direct Yukawa potential with energy-dependent range because the
mass of the nucleon is not the mass of the meson, and Diagram (b) is an exchange Yukawa potential with energy-
dependent range. And of course the direct and the exchange potentials must come in together because, even
though nucleon and antinucleon are not identical particles, meson and meson are. So if you have one graph, you
must have the other graph.

Figure 11.3: O(g2) Feynman diagrams for Model 3 NN → ϕϕ

That concludes the discussion, our first runthrough of the twenty-odd lowest non-vanishing Feynman
diagrams to O(g2) that arise in this theory.

11.3Crossing symmetry and CPT invariance

We have identified three kinds of phenomena that arise in lowest-order perturbation theory, and have labeled
them by names corresponding to the entities they become in the non-relativistic limit, to wit: direct Yukawa
potential, exchange Yukawa potential, and energy-eigenstate pole. Although no one of these things is in any
sense a relativistic novelty, there is a relativistic novelty, which I would now like to discuss. These three things are
in fact aspects of one thing. I would like to explain how they are connected. It goes under the name of crossing. It
is sometimes called crossing symmetry. You should put the word “symmetry” in quotes, because it has nothing
to do with symmetries and conservation laws in the sense we’ve discussed them.

In order to discuss crossing, I have to introduce a slightly different notation than the one I have been using
until now, and a slightly more general field theory. Just to keep things straight, consider a field theory in which
there are four different kinds of particles, call them 1, 2, 3 and 4, none of which are equal to their antiparticles.
That’s the most general case. These particles have various trilinear interactions, and can exchange various
charged and neutral mesons making the sort of Born approximation graphs we have been talking about. I would
like to consider a general graph involving one of each of these particles. I don’t know what’s going on inside, and I
don’t want to specify the process at the moment, so I draw a blob. By convention I will choose to arrange all my
lines so they all go inward.

Figure 11.4: The crossing diagram

 If we read it from right to left, Figure 11.4 represents the process

I will also arrange all my momenta {pr}, r = 1, 2, 3, 4 to be oriented inward, contrary to my previous convention. (I
do not recommend this change of convention in general, but it’s suitable for this particular discussion.) With this
convention,

We will tell which particle is incoming and which particle is outgoing not by saying whether it’s on the right or left
but simply by checking whether the zeroth component pr0 is positive or negative. If it is positive it is an incoming
particle; if negative it’s outgoing. If it’s really an outgoing particle, then the inward-oriented momentum is on the
bottom mass hyperboloid, not the top.

Thus this blob in my new compressed notation could describe a variety of processes:

I’m not telling you what the process is. You can deduce that only when you know the values of the p’s—which
ones have positive zeroth components and which ones have negative zeroth components. Of course, there are
three other processes this could describe, the charge conjugates of these processes, with 1 replacing 1, and so
on. I won’t write those down for reasons that will become clear shortly. We’ll get to them. I’m not assuming in any
way that the interactions between these particles conserve charge or parity or are time-reversal invariant or
anything like that. We’re going to be very general.

No matter which process I am describing, it is very convenient to introduce an over-complete set of three
kinematic variables to describe the system. Of course, for any given process we only need two, the energy and the
scattering angle (in the center of momentum reference frame). Nevertheless, for reasons that will become clear, I
want to introduce three:

Any two of the three constitute a complete set of relativistic invariants, and any invariant can be expressed in
terms of these. For the process (11.18a), drawn in Figure 11.4, s is the energy in the center-of-momentum frame,
while t and u are minus the squares of momentum transfer, one direct and one cross momentum transfer. Which
one I call direct and which I call crossed is a matter of taste, if the four particles are different. For this reason,
process (11.18a) is sometimes called “the s–channel process”, meaning it is the channel for which s is interpreted
as the energy in the center-of-momentum frame. For the same reason (11.18b) is called the t–channel process,
and (11.18c) is called the u–channel process. There is no physics in any of this. This is just a bunch of notations
that may seem to you to be over-complex until we get to the pay-off.

Suppose we read the crossing diagram, Figure 11.4, top to bottom. In this channel—I shouldn’t really call it a
channel, but people do, by an abuse of language—in this process (11.18b), the variable t is the energy in the
center-of-momentum frame, while s and u are momentum transfers, and vice versa with (11.18c) and u. The
variables s, t and u are called Mandelstam variables after Stanley Mandelstam. 4

Because only two relativistic invariants are needed, and here we have three, there must be some formula
relating s, t and u. Let’s derive this relationship. We have

However I have an additional piece of information. Because all the pi are oriented inward, the total momentum
flowing into the diagram is zero (11.17), and so is its square:
Subtracting (11.21) from (11.20) and dividing by two we find the rather pleasant and symmetric constraint

This expresses in a rather neat and simple way the dependence of the three variables.

We can represent the symmetric dependence graphically in a simple way. Let s, t and u be perpendicular
axes as shown. The relationship (11.22) describes a plane:

Figure 11.5: The Mandelstam plane s + t + u = ∑i mi2

We can indicate the values of s, t and u by a point in this plane, since there are really only two independent
variables. Take the origin of the plane to be the center of the equilateral triangle bounded by the lines s = 0, t = 0
and u = 0. I don’t want to destroy the symmetry between s, t and u by, say, picking s and t, and declaring u the
independent one. Instead, I introduce three unit vectors in the plane, ês , êt, and êu, as shown in Figure 11.6. The
unit vector ês is perpendicular to the line s = 0, êt is perpendicular to the line t = 0, and êu is perpendicular to the
line u = 0.

Figure 11.6: The Mandelstam plane and its unit vectors

The angle between any two of these unit vectors is 2π/3, and they have the property that

Each unit vector has a square of 1, and an inner product with the other two unit vectors equal to − .

As you can show for yourself, the vector r (Figure 11.7) from the origin to the point (s, t, u) can be written

If we dot r with any Mandelstam unit vector, say ês , we obtain the constraint

and likewise for t and u. So every point in the plane is associated with a triplet of numbers s, t and u which obey
(11.22),

Just to show you how this works, consider the line s = 0, the set of all points whose vectors r satisfy the relation

The lines t = 0 and u = 0 are similar. For a given point on the Mandelstam plane, s is the perpendicular distance
from the point to the line s = 0, t is the perpendicular distance to the line t = 0 and u is the perpendicular distance to
the line u = 0. Given a point r, if you want to know what s, t and u are, you just have to draw the three
perpendiculars to these three lines and measure the distances; see Figure 11.7.5

This is a very useful plot, not only in this problem but in problems like three-particle decays where you’d like
very much to express things in terms of a symmetric set of variables, especially if the decay involves three
identical particles, say three neutral pions. The energies of the three pions would be a useful set of variables, but
they’re constrained: the sum of those energies has to be the energy of the decaying particle. In the case of three-
particle decays, this is called a Dalitz plot,6 and we’ll say something about it next lecture. The case we’re
considering is called the Mandelstam–Kibble plot. It was no doubt invented by Euclid; it’s nothing but classical
geometry. So far, this plot is just a way of representing three constrained variables. Let’s get back to our three
scattering processes.

Figure 11.7: The variables s, t and u for a point r

Not every point in this plane corresponds to a physical scattering process. For example, if I pick a random
point where all three values s, t and u are positive, (say, near the origin), that would correspond to no physical
scattering process, because two of them have to be the squares of a momentum transfer, which is either zero or
negative. Let’s sketch out the regions of the Mandelstam–Kibble plot that correspond to our three physical
scattering processes, as in Figure 11.8. In general the boundaries are rather complicated and involve solving
quartic equations, so just to give you an idea of what they look like, I will restrict myself to the case where all of the
masses are equal. For convenience I will choose mr2 = 1, for each r. The four particles may still be distinct.

Figure 11.8: The Mandelstam–Kibble plot

In the center-of-momentum frame, p1 + p2 = (E1 + E2, 0). The physical region for the s channel is

The threshold center-of-momentum energy is s = 4, but s can be above threshold. The other variables t and u can
vary, but they both have to be less than or equal to zero: one is −Δ2 (10.43), and the other is −Δ2c (10.45). So the
inside of the upper triangle bounded by the lines t = 0 u = 0 is the physical region for s–channel scattering.
Likewise inside the lower right triangular region bounded by the lines s = 0 and u = 0 is the physical region for
t–channel scattering, and similarly for u–channel scattering. One benefit of this way of studying the scattering is
that we only have to do things once, not three times. If our particles have unequal masses, then the boundaries of
the physical regions look a little bit more complicated. They curve around, wiggle and bend. Of course, they
asymptotically approach this plot as the energy gets large compared to the masses. The shaded regions never
overlap, because there’s no possible way that you can have a process being in two channels at the same time,
say the physical s–channel and the physical t–channel.

Up until now we’ve been living in one or the other of these shaded regions. In our theory of mesons and
nucleons, the masses are different, and so our kinematics are not those of equal masses, and the regions aren’t
quite so simple in shape. But we’ve been living in one or the other of those shaded regions and have
systematically gone dancing, exploring things in each of these allowed regions. The actual amplitudes we have
obtained are however defined for all s, t and u. They are meromorphic functions of the invariants. In particular,
consider a process, say as shown in Figure 11.9, involving a fifth type of particle, different from the other four,
being exchanged between two others.

Figure 11.9: Scattering involving m5

Reading right to left, that’s an s–channel pole; it gives us a term proportional to

where m5 is the mass of the fifth particle that I have not talked about yet. I introduced it just to make that diagram,
Figure 11.9, possible. This equals

That’s a meromorphic function. The pole is located at s = m25 which had better be below the 1–2 threshold.

The line s = m25 is where that function has a pole. Unfortunately I cannot draw it as a point in the complex plane
because I would need two complex variables, which are hard to draw; it’s a four-dimensional graph. But
fortunately the location of the pole is on the real part of that plane and is everywhere along the line s = m25.

That amplitude (11.30) is analytically defined for all s, t and u, aside from right on top of the pole, of course.
What does that amplitude look like? In the s–channel, it looks like an energy-eigenstate pole. What does it look
like if I’m in the t–channel? Well, I just analytically continue the same amplitude. In the t–channel, reading the
crossing diagram, Figure 11.4, from top to bottom, the incoming particles are 1 and 3. Read this way, s is a
momentum transfer (squared) in the t–channel, and t is the center-of-momentum energy (squared). So exactly the
same meromorphic function down in the lower right shaded region, the t–channel, looks like a momentum-transfer
pole, i.e., a direct Yukawa potential. And over in the lower left shaded region, the u–channel, it looks like an
exchange Yukawa potential. That is, the three classes of phenomena we have been discussing—the Yukawa
potential, the exchange potential and the energy-eigenstate pole—are in fact simply three aspects of the same
meromorphic function restricted to three disconnected regions of the complex s–t–u plane. A direct Yukawa
potential in this sense is simply the analytic continuation of an energy-eigenstate pole, and so is an exchange
Yukawa potential. These processes are no more independent entities than are the functions sin z and i sinh z,
objects that look very different, but are the same meromorphic functions restricted to two different real
environments in the complex plane.

Figure 11.10: The Mandelstam–Kibble plot, showing s = m25

Therefore we have this property, unfortunately called crossing symmetry, but all the same a remarkable
feature of scattering theory. By reading the crossing diagram from top to bottom we’re crossing the line from the
past into the future; by reading it from bottom to top, we’re crossing from the future into the past. These three
processes appear to be completely different. The same process in our model could be nucleon–antinucleon
annihilation into meson–meson pair production, reading top to bottom; reading right to left, it could be
meson–nucleon scattering. Nevertheless, they’re connected by analytic continuation. In fact the three different
phenomena we have discussed are manifestations of a single meromorphic function in this s–t–u plane. This is
something we do not see in non-relativistic physics: the Yukawa potential in non-relativistic physics has no
connection with an energy-eigenstate pole for some other process. The regions are physically separated by
energies ∼ m2. When we go to the non-relativistic limit, (including a c2 in the mass terms), and the three regions
become very far apart: mc2 is large, and we can’t analytically continue from one region to another. They become
disconnected as c goes to infinity.

To what extent do we expect this crossing symmetry to be an artifact of our lowest-order theory? Well,
certainly we expect things to be much more complicated when we go to higher orders, because we know even in
non-relativistic physics the scattering amplitudes in general do not only have poles, they also have branch cuts. So
presumably there will be all sorts of cuts floating around this complex two-variable plane, and we’ll have to worry
when we analytically continue: Do we go above the cut or below the cut, and which way do we go around? I
assure you, the analysis can be carried out to all levels, with appropriate worrying, as we will do in the second
semester. These processes, which are apparently—I say it again because it’s so important—apparently totally
disconnected from each other, are in fact connected by a process of analytic continuation. The scattering
amplitude for one of them is the analytic continuation of the scattering amplitude for any one of the others. It’s a
remarkable fact.

What about the processes I have not written down? What about, for example, the process

Well, what is s for this process? It is the same as it was for 1 + 2 → 3 + 4. That’s just changing the signs of all the
momenta; it doesn’t affect these quadratic invariants. These processes correspond to the same point in the
Mandelstam plane. If I change the sign of all four momenta, changing all of my incoming particles into outgoing
antiparticles, and all of my outgoing particles into incoming antiparticles, I haven’t changed a single thing. In fact,
this equality has nothing to do with analytic continuation, and nothing to do with lowest-order perturbation theory.
Because my Feynman diagrams only involve quadratic functions of the p’s, all my Feynman rules to all orders in
perturbation theory for the theory in question are manifestly unchanged if I make the transformation

for all r. We could have all sorts of complex coupling constants that would break parity invariance, and we could
have terms that involve ψ + iψ ∗ floating around somewhere that would break charge conjugation invariance. We
could even have parity-violating terms involving derivative interactions. I haven’t told you yet how to derive the
Feynman rules for derivative interactions, but in momentum space the derivatives are replaced by momenta, and
we might expect the interactions to change sign. But if there’s any grace in the world, the rules should involve at
the vertices an epsilon tensor ϵµνρσ with four momenta. And since the epsilon tensor has four indices in four
dimensions, and therefore involves four momenta, when I change the sign of all momenta, interior as well as
exterior, that term in epsilon, multiplied by (−1)4, is not going to change sign, either. It’s special to four dimensions,
but it’s still true that the derivative coupling won’t change when p → −p. It looks like any Lorentz invariant
interaction I can write down will be invariant under changing the sign of all the momenta.

In general, then, for any Lorentz invariant interaction, to all orders of perturbation theory, amplitudes are
unchanged if I take all the momenta and change their signs, that is to say, if I take every incoming particle and turn
it into an outgoing antiparticle with exactly the same three-momentum. With our convention, this means
multiplying the four momenta by −1. This invariance is called CPT symmetry for reasons I will shortly make clear.
It says the scattering amplitude for a given process is exactly the same as the scattering amplitude for the reverse
process where all the incoming particles are turned into outgoing antiparticles and vice versa, to all orders of
perturbation theory. This is the CPT theorem.7 It is just a consequence of Lorentz invariance, but it is a
remarkable result.

The CPT theorem says that the world may violate parity—we’ve written down examples. It may violate
charge-conjugation invariance. It may violate time reversal. It’s trivial to write down examples to do that. But, if the
world is Lorentz invariant, it cannot violate CPT. Notice this is not like parity, for which there are phase
ambiguities, or charge conjugation or time reversal individually. There is no phase ambiguity, there are no minus
signs. This theorem not only tells you that there is a CPT symmetry, it tells you what it does, at least for theories
only involving scalar particles, which is all we can handle now. CPT symmetry turns an incoming nucleon into an
outgoing antinucleon with a plus sign. It turns an incoming K+ meson into an outgoing K− meson. As we will see,
the theorem can be generalized to spinor particles; CPT does something to their spin. It’s called CPT, because it
combines the operations of time reversal, charge conjugation and parity taken together. It changes incoming
particles to outgoing particles, which is what time reversal T does; it changes particles to antiparticles, which is
what charge conjugation C does; and it changes the sign of space variables, which is what parity P does. Notice it
does not change the sign of three-momentum, because of the combined action of TP.
 To give a specific example, consider once again the s–channel process

The amplitude for this process can be written as

where A is a function of the particles’ four momenta. If we charge conjugate this process with the operator C, we
get a related process with an amplitude ACfi,

If a theory is invariant under charge conjugation, then these amplitudes are the same, but in general they won’t be
the same. Now let’s consider the charge-conjugated s–channel process under T, the time-reversal operator. If you
run a movie backwards, the products of a reaction become the reactants, and vice versa. What once went north
now goes south, and what once went up now goes down: velocities are reversed. So time reversal does two
things: it switches the role of incoming and outgoing particles, and it reverses the direction of velocities. Finally, we
apply the parity operator P to the time-reversed, charge-conjugated s–channel process. What this does is undo
the reversal of velocities without swapping the roles of the incoming and outgoing particles:

The amplitude for this process can be written as

The original s–channel process and its CPT-transform occupy the same point on the Mandelstam–Kibble plot, and
so the change of pr → −pr cannot change anything. That is,

I want to emphasize the importance of CPT invariance. If an experiment were found to violate CPT, it would
not be like the downfall of parity in the original Wu-Ambler experiments, 8 nor like the violation of CP invariance
(and hence T individually) in the Fitch-Cronin experiments9 on K0 decays. All that happened then was that
someone said, “Well, that just means it’s not a CP-conserving theory.” So we write down our possible
Lagrangians, and all those terms we crossed out before, because they were CP-violating, we now leave in. That’s
not a revolution. We just go back and fix things up with a CP-violating interaction, and that’s it. But if CPT violation
is observed in the laboratory, that means Lagrangian field theory in general is cooked! Out the window! We’d have
to start afresh. That would be a revolution.

11.4Phase space and the S matrix

So much for grand abstract themes and powerful, beautiful general theorems. We now have to begin a bit of dirty
work. We now have a deep understanding of everything about S-matrix elements, except how to connect them to
cross-sections, which is what experimentalists measure. And we have to get that kinematics right if we want to
understand what we’re doing. So from a high plane of abstraction, we descend to a valley of practicality and go
through the manipulations required to find the formulas that turn S-matrix elements into differential cross-sections,
dσ/dΩ.

This is purely a kinematic problem, and there are two ways of approaching it. One is to be extra-careful and
consider a realistic scattering experiment with wave packets. That’s the right way to do it, but it takes a long time.
So I will do it fast and dirty by putting everything in a box, computing my scattering amplitude in the box, and then
letting the size of the box go to infinity. I’m also going to turn things on and off in time. If you use wave packets, the
box and turning things on and off in time are unnecessary and awkward. The box is there to make the kinematics
simple by replacing integrals by discrete sums. So I put the world in a cubical box of side L, and volume V = L3,
with momenta given by

and I put in an adiabatic function f(t), turned on for a time T. I will then choose the one-particle states in the
theory’s Hilbert space to be box normalized as we discussed earlier, in §2.2, treating Fock space in a box,
The momenta p and p′ run over the discrete set allowed by the box. This is also the commutator of the creation
and annihilation operators,

I have not described the expansion of the free field in a box before, but it’s easy to see what it is:

Instead of an integral over p, we have a sum on p, and in the denominator , instead of the square root of
2ωp we had before, in (3.45). Instead of the 1/(2π)3/2 which is appropriate for a delta-function normalization, we
have a 1/ appropriate for box normalization. It is easy to see that this is right by checking that it gives the right
equal time commutators (3.61) between ϕ(x, t) and (y, t).10

Now since everything is in a box and the interaction is only going on for a finite time, I can directly calculate
the transition probability between a given initial state and a given final state:

We subtract the one so if there is no interaction then there’s no transition. I will restrict myself in these lectures to
two kinds of initial states. I could consider a one-particle initial state:

That’s one of the nice advantages of turning the interaction on and off, because then I can get a crude but
serviceable theory of decay processes, to wit, I put the particle in the box, turn on the interaction, and watch it
decay into various final states. I might have an unstable particle in my model, and I will develop rules for
calculating the lifetime of such an unstable particle. I may also want to consider a scattering processing in which I
have two particles. Your first thought might be to write

But that’s not correct. Each particle has probability one of being somewhere in the box. If (11.44) were correct, as I
let the box get bigger and bigger the probability of finding the second particle in the neighborhood of the first
particle goes to zero, and therefore I should expect the transition amplitude to go to zero as the box expands,
which is not correct. The right way to normalize this initial state is to set

Then we can say one of the particles has probability 1 of being in the box and the second particle has probability 1
of being in any unit volume. As we let the box get bigger and bigger, the probability of the second particle being
near the first particle stays constant. We don’t want them not to scatter simply because there is no chance of
getting within any appreciable distance of each other, and we put in the to take care of that.

Now we are ready to go. I will write down the expression for the matrix element of S − 1 for any given final
state (some collection of particles with specified momenta). I’m not going to restrict the final states; they could be
two-particle states or they could be 32-particle states. What will I have? Well, I want to write this so it looks as
much as possible like (10.27), the formula we found in the relativistic case. I’ll write it like this:

I’ll have an invariant amplitude, not quite the same as before, because it involves a sum on p’s, rather than a
continuous integral. I’ll indicate that by putting in a superscript VT, indicating it represents factors from the box
volume and the time during which the interaction is turned on. The invariant amplitude AfiVT is so constructed that

There is also a factor that looks like a delta function but isn’t, quite. I’ll write down what that is in a moment.

The first three factors on the right of (11.46) look like the relativistic form. That would be all there was if the
states |pñ in the box were normalized the same way that relativistic states |pñ are normalized, but they ain’t! There
are extra factors you have never seen before, coming from the states’ normalization. Instead of (10.20), we have
I’ve got to put in this energy denominator and factor for each of the annihilation and creation operators, and
take the product over all the final particles. The product over all initial particles is a little different. We get 1/
for each particle’s energy, but we only have one factor of 1/ whether we have one or two particles in the
initial state. If there is only one particle, we get a 1/ factor. If there are two, we have (1/ )2, but one
factor will be canceled by the unconventional normalization (11.45) we used to define the initial two-particle state.
That is, there will be a single factor of 1/ whether the initial state is one particle or two particles.

Finally, let’s address the new delta function, δ(4)VT(pi − pf). We’ve got to be a little bit more careful about this,
because when we calculate the probability (11.42), we’re going to get its square. If we say δ(4)VT(pi − pf)
approaches a delta function, that doesn’t mean its square goes to the square of a delta function, because the
square of a delta function is garbage. Let me write down an explicit expression for this thing:

That’s how our energy-momentum-conserving delta function came out before, by doing an x integral. Here we’re
also doing an x integral, but we’re doing it only over a restricted volume of space and a restricted duration of time.
Sure enough, this is a highly peaked function that goes to an ordinary delta function as V and T go to infinity:

There’s no question about that. But that’s not what we’re interested in. As I said, we’re interested in its square.

Well, its square is also going to approximate a delta function, because if something is highly peaked, its
square is also highly peaked. We’ll get a delta function again, but with what coefficient? A very short computation
turning a Fourier transform into an integral over position space using Parseval’s theorem (9.32), shows us that this
is (2π)4VT:

Therefore we should write

Now we are ready to do the limit except for one thing. When we compute the square, we get the transition
probability to a fixed final state. This is of course a dumb thing to look at as the volume of the box goes to infinity,
because the allowed values of p are little dots lying on a cubic lattice.

Figure 11.11: Density of allowed momentum values

If we focus on some small volume of p space, we get more and more dots as the size L of the box increases: pi ∝
1/L, so the separation between dots inside the volume of p space decreases. The number of states in a small
volume d3p goes like

Therefore if we want something that has a smooth limit (as V and T → ∞), what we should look at is a differential
transition probability,

which is the probability for going to some fixed volume of final state space. We don’t want to compute the total
transition probability, we want to compute the transition probability integrated over some small region of final state
space.

We are now in a position to stick all of this together and allow V and T to go to infinity. Now comes the crunch:
Will V disappear? Putting all the factors together we have

Ta-daa! All the factors of V cancel, so there’s no problem in going to the limit V → ∞. It looks like we didn’t make
any errors. Well that’s rather nice, isn’t it? There’s our old Lorentz-invariant measure coming back again. We’ve
still got the factor of T, but of course we expect that if we keep the interaction on forever, and have particles
described by plane-wave states, they will go on interacting forever. So we divide by T, and taking the limits, we
can write the differential transition probability per time as

This is the master formula, sometimes written as

where D is an invariant phase space differential, an element of volume in final state space, called the relativistic
density of final states,

D should really be called the final state measure. It’s like the density of final states that you always have to play
with when you do time-dependent perturbation theory in non-relativistic physics. Of course it’s energy-momentum
conserving, so there’s (2π)4δ(4)(pi − pf) for the total incoming momentum which is determined by what the initial
state is, minus the total outgoing momentum which is the sum of the four-momenta for all the outgoing particles.
That tells you there’s no probability for making a final state which doesn’t conserve energy and momentum.

The master formula is easy to remember. The density of final state factors is the one thing that’s unnatural.
It’s there to make things have the right Lorentz transformation properties; for example, so a moving particle will
decay more slowly than a stationary particle. Please notice I have gone to great care to arrange these conventions
so there is no problem remembering where the 2π’s go. You may think this is a silly thing to be proud of, if you’ve
never tried to do a Feynman calculation in another convention. In the Feynman rules and in the density-of-states
factor, there is one and only one origin of a 2π in a denominator, and there is one and only one origin of a 2π in the
numerator: Every 2π in the denominator is associated with a differential dp in the numerator; every 2π in the
numerator is associated with a delta function of p.

At the beginning of the next lecture I will apply these rules to obtain specific formulas for scattering into two
particles, scattering into three particles, decay processes, etc. Once I’m done with that, you will be prepared to do
the homework.

1[Eds.] Problems 6, p. 261.

2[Eds.]
Coleman delivers this in a New York accent, adding, “I try to get his tone of voice, but it’s difficult. . . .”
(Coleman did his doctorate on unitary symmetry under Murray Gell-Mann at Caltech, and took graduate courses
there from Richard Feynman.) A moment later, a student asks: “Did Feynman refer to them as Feynman
diagrams?” Coleman replies: “No. Drawings I think he called them,” waving his hand as if to say, “Don’t get me
started...”, to much laughter. Then he adds: “When Feynman and Schwinger did much of this work in parallel, it
was said that Feynman’s papers were written in such a way as to make you believe anyone could have done the
computation, and Schwinger’s were written in such a way as to make you believe that only Schwinger could have
done it. Schwinger did not use diagrammatic methods.” Julian Schwinger (1918–1994) shared the 1965 Nobel
Prize with Feynman and Shin’ichiō Tomonaga (1906–1979) for advances in quantum electrodynamics. In 1980,
Schwinger described his reaction in 1947 to the introduction of Feynman diagrams “Like the silicon chip of more
recent years, Feynman was bringing computation to the masses.” J. Schwinger, “Renormalization theory of
quantum electrodynamics: an individual view”, in Laurie M. Brown and Lillian Hoddeson, eds., The Birth of Particle
Physics, Cambridge U. P. 1983, p. 343.
3[Eds.] See Marvin L. Goldberger and Kenneth M. Watson, Collision Theory, Dover Publications, 2004, p. 306,
equation (376.b) and Philip M. Morse and Herman Feshbach, Methods of Theoretical Physics, Part 2, Mc-Graw
Hill, 1953, p. 1077, equation (9.3.49).
4[Eds.] S. Mandelstam, “Determination of the Pion–Nucleon Scattering Amplitude from Dispersion Relations and
Unitarity”, Phys. Rev. 112 (1958) 1344–1360.
5[Eds.] Adapted from John M. Ziman, Elements of Advanced Quantum Theory, Cambridge U. P., 1969, p. 205.
6[Eds.]Richard H. Dalitz (1925–2006) was a particle physicist from Australia, and a student of Rudolf Peierls at
Birmingham. Soon after Peierls went to Oxford, he invited Dalitz to join him. There Dalitz taught Christopher
Llewellyn-Smith, a future director general of CERN, and many others. Dalitz was one of the early proponents of
quarks as physical entities, and not merely mathematical abstractions. Dalitz plots are introduced in R. H. Dalitz,
“On the analysis of τ-meson data and the nature of the τ meson”, Phil. Mag. 44 (1953) 1068–1080, and “Decay of
τ Mesons of Known Charge”, Phys. Rev. 94 (1954) 1046–1051.
7[Eds.]Usually, the CPT theorem states that if a local quantum field theory is Lorentz invariant and the usual
connection between spin and statistics holds, then the theory is invariant under the combination of operations
CPT. J. Schwinger, “Theory of Quantized Fields, I.”, Phys. Rev. 82 (1951) 914–92 (reprinted in Schwinger QED);
W. Pauli, “Exclusion Principle, Lorentz group and reflection of space-time and charge”, in Niels Bohr and the
Development of Physics, pp. 30–51, McGraw-Hill, 1955; Gerhart Lüders, “Proof of the CPT theorem”, Ann. Phys.
(NY) 2 (1957) 1–15.
8[Eds.] See footnote 8, p. 121.
9[Eds.]J. H. Christenson, J. W. Cronin, V. L. Fitch and R. Turlay, “Evidence for the 2π decay of the K20 meson”,
Phys. Rev. Lett. 13 (1964) 138–140. In 1980 James Cronin and Val Fitch won the Physics Nobel Prize for this
work.
10[Eds.] (1/L) ∑∞n=−∞ e−i(2πn/L)x = (1/L) ∑p e−ipx = δ(x) is the Fourier series expansion of the delta function.

12
Scattering II. Applications

I will devote this lecture to the systematic exploitation of the formulas (11.57) and (11.58). This will be a nice
change, because it will not involve any new ideas and therefore you don’t have to think too hard. I will apply these
formulas to five straightforward (or even pedestrian) topics. First I will discuss decay processes. After decay
processes, I’ll talk about cross-sections and I’ll explicitly evaluate D for two-particle final states in the center-of-
momentum frame. I’ll discuss the famous Optical Theorem that connects the imaginary part of the forward
scattering amplitude to the total cross-section. Finally I will discuss D for three particle final states and say a little
bit more about those Dalitz plots that are so useful.

12.1Decay processes

Let us begin with decay processes. We start out with some one-particle state that would be stable were it not for
its interactions, turn on the interactions and watch it decay. We know the rate at which it decays from our master
magic formula, (11.57). There’s only one particle in the initial state. The differential transition probability for
decaying into some n-particle final state is conventionally called dΓ. It is given by our master formula,

Ep is the energy of the initial particle, with momentum p. I don’t bother to put an index on it because there is only
one particle. The amplitude for the decay is Afi; this is to be multiplied by the invariant phase space differential, the
density of final states D. The differential transition property is clearly a differential of something, and that is
contained in the factor D. If the particle decays into a final state containing three particles, D will include a factor
like d3p1 d3p2 d3p3.

The total transition probability per unit time, Γ, is obtained by integrating dΓ over all final momenta and
summing over all possible final states, if there are many final particle states into which this thing can decay—three
mesons, two mesons, a nucleon and antinucleon etc.1 Γ is typically evaluated for the incoming particle at rest (its
momentum equal to zero). When you see a table of Γ’s, it doesn’t say, “This is Γ for the particle moving at one
third the speed of light”. Then the energy is that of the particle at rest, Ep = m or µ, whatever the mass of the
incoming particle is, so

This way of writing the formula makes it very clear what the decay amplitude is for a moving particle. The quantity
∫ |Afi|2 D is Lorentz invariant, because D is a Lorentz invariant measure, and Afi is a Lorentz invariant object. If we
evaluate this expression for incoming momentum equal to something else, p, then the only difference is the factor
in front, 2m being replaced by 2Ep. That is,

This is of course just what we would expect from time dilation. This equation expresses the fact that a moving π
meson decays more slowly than a stationary π meson. The faster it moves, the more slowly it decays. That helps
to explain the physical meaning, at least in this case, of that mysterious factor of 1/2Ep for the initial particle, the
only thing that is not Lorentz invariant in our expression. It damn well had better be there, otherwise we would
have predicted that a moving particle decayed at the same rate as a stationary particle, which would be bad news
both from the viewpoint of relativity theory and from the viewpoint of experiment.

12.2Differential cross-section for a two-particle initial state

If we have a beam of particles impinging on a stationary target, or if we have a target moving into a beam of
particles, the differential cross-section is defined as the transition probability per unit time per unit flux. By
definition, the differential element dσ of the cross-section is

That is to say, we divide the differential transition probability per unit time by the flux of particles impinging on the
target. The energies E1 and E2, and the velocities v1 and v2, are those of the initial particles. The unit flux, the
number of beam particles hitting the target particle per unit time per unit area, is the difference in the three-
velocities of the incoming particles:

Let’s understand this factor. Our normalization convention is such that we can think of one of the particles as
having probability one for being somewhere in the box, and the other particle having probability one for being in a
given unit volume,

Suppose the particle with momentum p1 presents some area, A, to the particle beam with momentum p2. The
normal to the area A is parallel to the direction v1 − v2. See Figure 12.1.

In a time t, the area normal to v1 − v2 sweeps out a volume |v1 − v2|At. The particle flux is defined to be
 Figure 12.1: Flux for a two-particle initial state

So the differential cross-section (12.4) is our general formula (11.57) divided by the flux, (12.5).

I can think of one of these particles as being the target. Which of these two p’s I associate the square root of
V with is a matter of taste, but let me consider the first one as being the target, somewhere in the box. The second
one is the beam. It has probability one for being someplace in the box. I have the target moving through the box
with velocity v1, while the beam moves through the box with velocity v2. The probability flux hitting the target is v1
− v2, the ordinary non-relativistic difference of velocities. I emphasize that. A friend of mine who once did a thesis
on neutrino–neutrino scattering, a rather abstract subject but of some cosmological interest, had a factor of 4 error
in his thesis because he said, “Oh, they’re relativistic particles; their relative velocity here must be c.” It is not. For
neutrino–neutrino scattering, it is 2c, if they head into each other. We’ll see that that’s consistent with relativity
also. If I turn on my stopwatch for one second and ask how much beam has passed the target in that one second,
the answer is one over v1 − v2 worth of beam. That’s unambiguous.

The total cross-section σ is obtained by summing and integrating over the final states:

We have an evidently non-Lorentz invariant factor in front, (4E1E2|v1 − v2|)−1, times a Lorentz invariant. I will now
discuss the Lorentz transformation properties of the factor in front to demonstrate that they are what one would
think they should be.

Consider a special form of (12.8) in a Lorentz frame in which the two particles are moving head on towards
each other, or one is catching up to the other, where the two three-momenta are aligned. I will consider

where I’ve chosen the coordinate system such that the x axis is aligned with v1, and so also with either v2 or −v2.
Because v = p/E,

For later computations it will be useful to have the expression for (12.10) in the center-of-momentum frame.
Then

and (12.10) simply becomes

where ET is the total energy of the two particles, ET = E1 + E2, and pi is the common magnitude of the incoming
three-momenta. We can rewrite (12.8) for σ in the center-of-momentum frame,

We all know from non-relativistic physics the geometrical picture of the total cross-section, and why it is called
a cross-section. We have a beam of particles heading in one direction, which we will call the negative x axis, and
some object, off of which the beam is scattering, heading in the direction of the positive x-axis:
 Figure 12.2: Object in beam

The total cross-section gives the total probability of scattering, because it is simply the geometrical cross-section
presented by the object. In the classical picture, if the beam hits the object, it scatters, and if it misses the object, it
doesn’t. This picture, Figure 12.2, would indicate that the total cross-section should be the same in any Lorentz
frame that preserves the expression (12.10). If I make a Lorentz transformation which is a boost along the x
direction, that preserves the expression. That is to say, a Lorentz transformation restricted to t and x, a rotation in
the (0, 1) plane, will change the appearance of the target in the x direction by Lorentz contraction, but it won’t
change its perpendicular dimensions, so it won’t change its cross-section. Of course, if I make another kind of
Lorentz transformation, then things are different; then I am distorting the particle in the direction that the beam
sees, and then I shouldn’t expect the total cross-section to be invariant.

Now let’s check that this is so. Is (12.10) invariant under Lorentz transformations restricted to the (0, 1) plane,
or is it not? Well, it’s obvious that it is, once I’ve got things in this form, because

where 2, 3 here are Lorentz indices, equal to y, z, respectively; ϵρλµν is the completely antisymmetric object we
talked about before. The right-hand side of (12.14) has only two non-vanishing terms if we fix ρ = 2 and λ = 3: a
term where µ = 0 and ν = 1, and another where µ = 1 and ν = 0. These terms have a minus sign between them,
and they give you the two terms on the left-hand side. Or maybe they give you the terms with the sign reversed; it
doesn’t matter, it’s the absolute value. The right-hand side of (12.14) is obviously invariant under Lorentz
transformations restricted to the (0, 1) plane, because the 2 and 3 indexed variables don’t change and the rest is a
Lorentz invariant sum. So this expression (12.10) is okay.

The total cross-section does what it should. If in any Lorentz frame where the 3-momenta are along the x
axis, you compute the total cross-section, you get the same result as in any other Lorentz frame in which the 3-
momenta are along the x axis. Thus we have established that cross-sections Lorentz contract as they should
Lorentz contract: not at all, if you make your Lorentz transformation along the direction of the beam. Please notice
once again that the mysterious factors of 1/E that come into our formula for the transition probability per unit time
are essential for the result to come out right.

12.3The density of final states for two particles

We now turn to the third topic. It would be very nice to have a more compact formula for the density of final states
than the awful expression (11.58). So let me now compute D for a two-particle final state, where the initial particles
are in the center-of-momentum frame:

I’ll do it in this frame because it’s the simplest case. (It’s also pretty easy to do it in some other frame; see Problem
7.1.) In the case of two final particles with momenta p3 and p4,

I’ll split the four-dimensional energy-momentum-conserving delta function into two factors,

where ET is the total incoming energy. We now want to cancel out some of the delta functions against some of the
differentials. The easy one to do first is the integration over p4, and we use the δ(3)(p3 + p4) to cancel it out. That
is to say, if we integrate D with any function, doing the integral is the same as replacing p4 by −p3, canceling the
d3p4 and canceling the delta function. So (12.16) becomes
where p4 is now constrained to be −p3. And of course that means E4 is also a function of p3.

I now go to angular variables, and write d3p3 = |p3|2d|p3|dΩ3. Then

Now I’ll cancel the delta function of the incoming energy by integrating over d|p3|, which fixes |p3| = |p4|. We’ll
need the important rule (note 8, p. 9), which I presume you recall,

where x0 is the root of f(x) = 0. (If there are several zeros, you get a sum of such terms, one from each zero, but
there won’t be in this case.) Performing the integration over |p3|, I get

I don’t need the absolute value here since these are positive quantities. We have

By differentiation, E3 dE 3 = |p3| d|p3| and E4 dE 4 = |p3| d|p3|, so

If we substitute (12.23) into (12.21), one factor of |p3| cancels and the product E3E4 cancels. Let’s identify E3 + E4
= ET, the total final energy, and |p3| as the magnitude |pf| of the momentum of either final particle in the center-of-
momentum frame. We obtain a formula important enough for me to put in a box:

Please notice that the magnitudes of the initial particles’ momenta |pi| and the final particles’ momenta |pf| in the
center-of-momentum frame are different if the final and the initial particles have different masses. The factor dΩf
describes the solid angle associated with d3pf.

EXAMPLE 1. Calculating dσ/dΩ in the center-of-momentum frame.

To compute dσ/dΩ in the center-of-momentum frame, we return to (12.4), and substitute in (12.12) and
(12.24):

so that

I should make two remarks. Please notice the factor of |pf| over |pi|. In an inelastic process, the masses of the
final particles are different from those of the initial particles, and so |pf| does not equal |pi|. Even though time
reversal may tell us that the amplitude for an inelastic process is the same as the amplitude for the time-reversed
process, this does not mean the cross-section for the process is the same as the cross-section for the time-
reversed process. We have |pf| over |pi| in one case and |pi| over |pf| in another, so that even if the amplitudes are
the same, the differential cross-sections will not be the same. This will be a familiar result from non-relativistic
physics, if you’ve ever studied things like the scattering of electrons off atoms. These collisions can excite the
atom, and occur as both exothermic and endothermic reactions. Thus for example in our model the total cross-
section for nucleon–antinucleon annihilation into meson plus meson is not the same as a total cross-section for
meson–meson production of a nucleon–antinucleon pair, even though the amplitudes are identical.

Note that for an exothermic reaction, we can have pi = 0 but pf ≠ 0. That means dσ/dΩ and hence σ can be
infinite even when the amplitude Afi is finite. It is simple to understand the meaning of the ratio with pi in the
denominator. As pi → 0, i.e., approaching threshold, the two particles spend more time near each other,
increasing the likelihood of interaction. Engineers slow down neutrons in atomic piles to minimize the denominator,
and maximize the chance of neutron capture in the pile.

Second remark: We can now compare this expression (12.26) for the relativistic differential cross-section to
what we find in non-relativistic physics, to make the correspondence between our non-relativistic and relativistic
notational conventions. That’s a useful thing to do if we ever want to check that things have the right non-
relativistic limit. In non-relativistic physics, we also define a scattering amplitude, using the normalization
conventions convenient for non-relativistic physics. In non-relativistic quantum mechanics, we always have elastic
scattering by a potential, with pf = pi = p. The scattering amplitude, usually called f(p, cos θ), is given by a very
simple formula,

In the elastic case, comparing (12.27) with (12.26), we see the connection between the relativistic and the non-
relativistic scattering amplitude,

(up to a phase). When I derive the Optical Theorem, we’ll see that the sign should be positive (and that there is no
phase factor). This is the scattering amplitude as conventionally defined in non-relativistic potential scattering, and
we now see how it is related to our relativistic scattering amplitude.

EXAMPLE 2. Calculating Γ for ϕ → NN in Model 3.

We have assumed that µ < 2m, so this process should not occur. For a moment, let’s relax this constraint. We
have, ignoring the adiabatic function,

The relevant Feynman diagram for the decay of a meson at rest into a nucleon–antinucleon pair is Diagram 3(c)
on p. 216:

Figure 12.3: Vertex for Model 3

The contribution of this graph is

so that, to first order in g, the invariant amplitude for meson decay into a nucleon–antinucleon pair is given by

That is,

It couldn’t be simpler. From (12.2) and (12.24), to O(g2), the decay rate Γ of the muon is

In the center-of-momentum frame, |p1| = |p2| = |pf|, ET = , and ET = µ. Then

and

Clearly, this is imaginary unless µ > 2m. That is, the meson is stable if µ < 2m, as we have assumed.
 12.4The Optical Theorem

For simplicity I will assume I am working in a theory in which there is only one kind of particle, say a meson. Then,
when I sum over all final states, I don’t have to complicate my notations unnecessarily to indicate summing over
mesons and nucleons and antinucleons. The generalization of my arguments to more complicated theories will be
trivial.

I start out with this famous equation, expressing the unitarity of the S-matrix:

I will deduce the Optical Theorem as a consequence of this equation. Our invariant amplitudes are defined in
terms of S − 1, but it is pretty easy to find an equation corresponding to (12.36) in terms of S − 1:

I now evaluate this identity between initial and final states, |iñ and |fñ, respectively:

I’ll begin with the right-hand side. Remembering (10.27),

(because of the adjoint, I complex conjugate the amplitude and swap its indices). So we have

The left-hand side of (12.38) evaluated between |fñ and |iñ I will write in terms of a complete set |mñ of
intermediate states. I’m assuming there is only one kind of particle, so the intermediate states |mñ are r-particle
states: |mñ = |q1, . . . , qrñ. Then, using (1.64),

I divide by r! to keep from over-counting the states. That’s simply the left-hand side of (12.38) written in terms of
the sum of a complete set of intermediate states.

We simplify this expression as before:

The left-hand side of (12.38) can then be written

where we replace δ(4)(pm − pi)δ(4)(pf − pm) by δ(4)(pf − pi)δ(4)(pf − pm), and take the common delta function factor
outside the sum. Comparing the right-hand side (12.40) of (12.38) with the left-hand side (12.4), I have (2π)4δ(4)(pf
− pi) on both sides of the equation, so I can divide it out:

Once I have divided out the delta function, I can safely set |iñ = |fñ, because I will no longer encounter an infinity.
Let’s do that. Then
 What do we get? The left-hand side of this equation is twice the imaginary part of Aii. On the right-hand side, I
get something very interesting. Comparison with (11.58) shows

In particular, say that the initial state has two particles:

Then the right-hand side of (12.45) becomes a statement about the cross-section:

using (12.13). That is to say, it is the total cross-section, except for that funny factor 4|pi|ET. Setting the two sides
equal, we can divide out a common factor of 2.

Thus after excruciatingly dull labor we have arrived, as a consequence of the unitarity of the S-matrix, at the
famous Optical Theorem,

This asserts that the imaginary part of the relativistic forward scattering amplitude equals twice the total energy
times the momentum of the particles in the center-of-momentum frame times the total cross-section. It doesn’t
matter if the particles are incoming or outgoing, because the elastic scattering amplitude in the forward direction is
the same for the initial states as the final states. It is just a consequence of the unitarity of the S-matrix, and
therefore it is a very general result.

Incidentally, since the right-hand side of (12.49) is zero for NN → NN scattering up till O(g4), it follows that Im
Aii = 0 up to O(g4) in Model 3. This proves that the forward scattering for NN → NN is real at O(g2) in Model 3. In
fact, Feynman amplitudes are real to O(g2) for scalar fields in all directions.

Just to check, let’s compare this with the equally famous Optical Theorem of non-relativistic scattering
theory.2 Recalling that |pf| = |pi| = |p| for elastic scattering, we have

That is, the imaginary part of f(|p|, cos θ) in the forward direction, is equal, by some elementary algebra, to |p|/4π
times σ. Comparing (12.49) and (12.50), we see that

Looking back to (12.28), and taking the imaginary part of both sides, we conclude that the + sign is appropriate,
and that there is no phase factor:

Indeed, we didn’t have to go through this argument for relativistic scattering, because the Optical Theorem of non-
relativistic scattering theory is true whether or not the theory is relativistically invariant. We’ve shown that it is just
a consequence of unitarity. It’s an amusing exercise, in case you had not seen it in non-relativistic quantum
mechanics, to see it come out here.

12.5The density of final states for three particles

I will now consider phase space for three-body final states, just to show you that these integrals are not particularly
difficult. I will again work in the center-of-momentum frame:

Since you’ve already seen me do one of these calculations in gory detail, I will skip immediately to something you
can see by eyeball. I’ll write down D from our general formula, (11.58):
 First, we will take care of the 2π’s. There are three factors of (2π)3 in the denominator, one for each of the final
particles, and there is a (2π)4 in the overall energy-momentum-conserving delta function, so the net is (2π)−5.
There is an energy denominator, 2E for each final particle, that gives us (8E1E2E3)−1. There are d3p1 and d3p2
which I will immediately write in polar form. There’s a d3p3 which I will cancel off against the space part of the delta
function, and need not write down. That gives

And there is a remaining delta function of E1 + E2 + E3 − ET. The hard part will be doing the integral to get rid of E3
which will cancel out one of our other variables.

Figure 12.4: The angles θ12 (latitude) and ϕ 12 (azimuth)

Let’s look at the angular integrals (Figure 12.4). Here’s p1, which I’ll take to be my z direction. Off someplace
is p2, and between them is the angle θ12. I’ll hold p1 fixed, and integrate over Ω2 first. The angular differentials can
be written as

where θ12 is the relative angle between p1 and p2, and ϕ 12 is the azimuthal angle of p2 relative to p1.

We write the angular integrals in this way because the only variable that depends on θ12 when we keep the
other variables fixed is E3. We can thus cancel the energy delta function against the integral in θ12:

Using the rule about delta functions of functions,

where ξ is the value of cos θ12 that ensures E1 + E2 + E3 = ET. The derivative is easy:

The integral over θ12 is trivial. Carrying that out, we obtain

This expression becomes especially simple once you recall the famous result3 that |p| d|p| = E dE, so

Even more of the denominators cancel out when I make those substitutions, and I find a remarkably simple
expression for the density factor in three-body phase space:

We have now run out of delta functions, and can’t carry out any more integrations. That’s as far as one can go in
the general case. To calculate lifetimes or cross-sections in a specific case, we need to know how the amplitude
for that process depends on the various 4-momenta.
 In truth, the situation is not quite as simple as it may appear: E1 and E2 are severely restricted because step
(12.58) is true only if cos θ12 goes from −1 to 1, so that the zero in the delta function occurs within the range of
integration. That is to say, E1 and E2 are not allowed to range freely. Indeed we can see what happens at one
extreme where the vectors p1 and p2 are aligned, and θ12 = 0. The sum of the energies is

The quantity |p1 + p2| is the biggest |p3| can get. And that upper limit in the integration had better be greater than
or equal to ET, which in turn must be greater than or equal to the lower limit of the integration,

These boundaries define an awful, hideous-looking region in (p1, p2) space, or equivalently (E1, E2) space. One
can work at it and beat on it and you will end up with a cubic equation involving E1 and E2 to determine the
boundaries of this region, but that is still pretty terrible.

So there is some ugly-looking region in (E1, E2) space, or in a Dalitz plot, something that looks, for example,
like Figure 12.5. The dots denote events. It’s not quite as monstrous as that; I believe it’s convex. But in general
there is some monstrous blob where kinematics allow the final particles to come out.

Figure 12.5: Dalitz plot for E1 and E2 for three-particle final states

Although you do have a simple thing to integrate, you have to integrate it over terrible boundaries of integration.
This prospect causes strong men to weep, brave women to quail and sensible people to go to their nearest digital
computer.

An especially interesting application of the formula (12.62) occurs in the decay of a spinless particle. If a
spinless particle decays into three spinless particles, or indeed if a particle with spin decays into three particles
with spin, but you average over all the spins, so there’s no preferred direction, then obviously the differential
decay amplitude dΓ doesn’t depend on the angular variables ϕ 12 and Ω1. Of course, that’s not true if you have a
particle with a definite spin, or if you have two particles coming in, because then there’s a specified direction in the
problem (the direction of the spin, or the direction along which the two particles approach one another). For the
decay of a spinless particle, you might as well go ahead and do the integral over Ω1 and ϕ 12 of D:

In such a case, you will frequently find people making plots in E1 and E2, or the symmetric E1E2E3 diagram,
analogous to the Mandelstam diagram. They will put little dots whenever they’ve observed a decay event, as
shown in Figure 12.5. This is very groovy,4 because you can directly read off the squares of the invariant matrix
elements without going through any kinematic computations. They’re proportional to the density of dots, with a
proportionality factor we know to be 1/(32π3). It’s very nice, when you’re trying to see if experiment and theory fit,
not to have to do any complicated phase space computations.

12.6A question and a preview

As you now know, this course is divided into two kinds of lectures: those that are totally understandable and
inexpressibly boring, and those that have exciting ideas in them—well, perhaps I’m giving myself airs—but are
absolutely incomprehensible. Next time we will turn to the second kind of lecture.

We’re going to try and redeem our scattering theory and all our Feynman graphs by re-establishing things in
such a way that the adiabatic function f(t) does not appear, and doing things straight, in the real world. This will
involve a long sequence of arguments. Just the beginning of this topic will occupy a couple of lectures. Working
out all the details that follow from it, which will involve us in all sorts of strange things with strange names like wave
function renormalization, will take another two lectures. So it’s going to take us a lot of time, and it’s going to begin
in a rather abstract way. We will start by investigating, in our old framework of scattering theory, what seems to be
a silly question. I’ll tell you what the silly question is now, although the investigation won’t proceed until the next
lecture.

The silly question is: What is the meaning of a Feynman diagram when the external lines are off the mass
shell? External lines represent real particles, whose 4-momenta satisfy pµpµ = m2; they are said to be “on the
mass shell”. Internal lines, by contrast, represent virtual particles, which do not lie on the mass shell. A Feynman
diagram gives us the scattering amplitude when the external lines are on the mass shell. However, if I take a
Feynman diagram, let’s say a particularly complicated and grotesque-looking diagram for meson–meson
scattering, as in Figure 12.6, the Feynman rules as I wrote them don’t say the external lines have to be on the
mass shell. If I were some sort of maniac, I could compute this diagram with the external lines not obeying pµpµ =
m2. Does a Feynman diagram with the external lines off the mass shell have any meaning? Well, it has a sort of
primitive meaning that one can see. It could be the internal part of some more complicated Feynman diagram, as
in Figure 12.7: Yeah, that’s a homework problem. (Ha! I’m just kidding.) If I’m to have any hope of evaluating this
diagram, I might try to put a dotted line around this inner part, evaluate it, get some function of the four 4-momenta
coming in at the four vertices of the inner square, plug that dotted inner part as a black box into the bigger
diagram, and then do the big integrals. So at least here is some sense of talking about Feynman diagrams with
the external lines off the mass shell: It might be the internal part of a more complicated Feynman diagram. In my
Feynman rules, although the outer lines of the larger diagram have to be on the mass shell, these lines on the
smaller diagram don’t, because they’re internal lines.

Figure 12.6: O(g4) ϕ + ϕ → ϕ + ϕ scattering

Figure 12.7: O(g12) ϕ + ϕ → ϕ + ϕ scattering

Next lecture I will show that within the framework of our old scattering theory, these Feynman diagrams with
lines off the mass shell can be given two other meanings, aside from the rather trivial meaning I have assigned to
them originally by drawing this dotted circle. The second meaning is that these Feynman graphs, or rather, the
sum of all Feynman graphs with a given number of external lines, can be related to objects called Green’s
functions that determine the response of the system to a particular kind of external source. In particular, if I take
the Hamiltonian density for a system, and combine together, say, Model 1 and Model 3, there’s a lot of
interactions in the Model 3 H I, but there’s also an external source,

Were we to compute the vacuum-to-vacuum matrix elements in the presence of ρ, we could make particles with
the source, as we did in Model 1. But now it won’t be so simple, because the particles are interacting. We have a
new interaction in the theory caused by ρ. I’ll write down its analytic form next lecture. And that new interaction in
turn makes new Feynman diagrams:5

Figure 12.8: Feynman graph for an interaction term ρ(x)ϕ(x)

 If I am to compute what happens in this combined theory, say to fourth order in ρ and some order in the
coupling constant, as shown in Figure 12.9, one of the diagrams I will encounter will be the Feynman diagram
shown in Figure 12.6, the thing with the circle around it in Figure 12.7. But now, because of the interaction with ρ,
denoted by the dots, the formerly external lines are internal lines, and therefore have to be integrated off the mass
shall. I will work out the details of that next lecture and develop a second meaning of these graphs, with external
lines off the mass shell, as Green’s functions, in the primitive sense of George Green, a function that tells you the
response of a system when you kick it, the system’s response to an external source.

Figure 12.9: ϕ + ϕ → ϕ + ϕ scattering to O(ρ4)

I will then give a third meaning. I will show that in fact these things express a certain property of the
Heisenberg fields, the exact solutions to the Heisenberg equations of motion. I will then assemble these three
things and write down a formula that is really just a statement that you get a scattering amplitude by taking a
Feynman graph with lines off the mass shell and putting the lines on the mass shell. I will connect that to a certain
expression constructed of the Heisenberg fields. That expression will turn out to have no reference to our original
adiabatic turning-on-and-off function f(t), and that will be the starting point of our new investigation of scattering
theory. I will then attempt, by going through considerable contortions and waving my hands at a ferocious rate, to
justify that expression without talking about the turning-on-and-off function, and thus getting a formulation of
scattering theory that has nothing to do with turning on and off. That is the outline of the next couple of lectures.

1[Eds.] In conventional units, Γ = ħ/τ, where τ is the particle’s mean lifetime. Γ has the units of energy.
2[Eds.] L. I. Schiff, Quantum Mechanics, 3rd ed., McGraw-Hill, 1968, p. 137; or Landau & Lifshitz, QM, p. 476.
3Just differentiate E2 = |p|2 + m2, to get 2E dE = 2|p| d|p|.
4[Eds.] “Groovy” is antiquated American slang from the 1960’s, meaning “excellent”; here, “welcome”, or “a good
thing”.
5[Eds.] Figure 12.8 does not appear in the video of Lecture 12, but it does in Coleman’s handwritten notes for
October 28, 1986. In the video of Lecture 13, Coleman says that this Feynman graph was calculated when talking
about Model 1. In fact, the Wick diagram was calculated in Chap. 8 (see (8.66)), but not the Feynman graph. The
O(g0) matrix element is

Problems 6

6.1 The most common decay mode of the short-lived neutral kaon (mass 498 MeV) is into two charged pions
(mass 140 MeV). For this process, Γ = 0.773 × 1010 s−1. Make the silly assumption that the only interactions
between pions and kaons are those of Model 3 of the lectures, with the kaon playing the role of the meson and the
pion of the “nucleon”, and compute, from these experimental data, the magnitude of the dimensionless quantity
g/mK, to one significant digit. Can you see why this is called a “weak interaction”?

Comments:

(1) Actually, the silly assumption is irrelevant: by Lorentz invariance, the matrix element for this process, a, is
just a number; the center-of-momentum energy is the kaon mass, and, by rotational invariance in the c.o.m.
frame, a cannot depend on the angle of the outgoing pions. What we are really doing is computing a, without any
dynamical assumptions at all.

(2) Take ħ = 6.58 × 10−22 MeV-s.

(1997a 6.1)

6.2 In Model 3, compute, to lowest non-vanishing order in g, the center-of-momentum differential cross-section
and the total cross section for “nucleon”–“antinucleon” elastic scattering.
(1997a 6.2)

6.3 Do the same for “nucleon–antinucleon” annihilation into two mesons. WARNING : Don’t double-count the final
states.
(1997a 6.3)

6.4 In class, I showed that to every order of perturbation theory, the invariant Feynman amplitude was unchanged
under multiplication of all 4-momenta by −1, and I claimed that this was equivalent to invariance of the S matrix
under an anti-unitary operator, ΩCPT. In this problem, you’re asked to work out explicitly what an anti-unitary
symmetry implies about the S matrix, to verify (or perhaps refute) my claim. For notational simplicity, we’ll work
with time reversal, ΩT; the extension to TCP is trivial.

Let us denote the action of time reversal on a state |añ by |aTñ:

Thus, in the theory of a free scalar field,

Assume that in the interacting theory,

and also assume a like equation with “in” and “out” interchanged.

(a) Show from the definition of the S matrix in terms of in and out states that this rule implies

(Note that this is a sensible equation, in that both the left- and right-hand sides are linear functions of |bñ and anti-
linear functions of |añ.)

(b) Get the same result as (P6.4), starting from the fundamental formula (7.59) of our adiabatic scattering theory,

assuming, of course, that the interaction is invariant under time reversal:

Solutions 6

6.1 The relevant Feynman graph is Figure 12.3 on p. 251. The relevant equation is (12.35). The coupling constant
g has units of [L]−1 or MeV, and so does Γ. The experimental value of Γ is given in s−1. To get the units right, we
have to put ħ in: Γ → ħΓ. Then

Solving for g/mK gives

The estimate helps to explain why this is a “weak” interaction, with g/mK on the order of 10−5 smaller than α =
e2/ħc.

6.2 The formula for the differential cross-section is given by (12.26),

Because the collision is elastic (between particles of identical mass), we have |pi| = |pf|, and the differential cross-
section becomes

The amplitude comes from the two Feynman graphs in Figure 11.1 and is given by (11.1), with p′1 → p3 and p′2 →
p4, respectively,

In the center-of-momentum frame we have

The total energy , so

Let θ be the angle between pi and pf, and ϕ be the azimuthal angle about the pi axis. Then

and similarly

Plugging these in to the amplitude,

We can safely drop the iϵ’s, because neither denominator will become zero. Then plugging into (S6.3),

To obtain the cross-section, we need to integrate this over the solid angle, dΩ = −dϕ d cos θ. There is no ϕ
dependence, so we can do that by inspection, to obtain 2π. Pulling out the constant terms, and writing cos θ = z,
the cross-section is

The integral is easily done with the substitution u = 2|pi|2 (1 − z) + µ2. Then

The cross-section has a finite limit as .

6.3 The relevant diagrams are the two Feynman graphs in Figure 11.3. Let’s redraw these, to let q, q′ stand for the
meson momenta, and p, p′ for the “nucleon” and “antinucleon” momenta, respectively:

The amplitude is

In the center-of-momentum frame we have

The total energy can be written in two equivalent forms, by energy conservation:

This means that |q| can be written as , but we’ll express our answers using both |p| and |q|. We can
also write

Let θ equal the angle between p and q. Then

Similarly,

Both these quantities are negative definite, so neither contribution to the amplitude gives a pole, and once again
the iϵ terms may be dropped. Then

Mindful of the warning, recognize that the two final states |q, q′ñ and |q′, qñ are the same, and divide by 2 to
prevent overcounting. Then

As before, to obtain σ, we integrate this over the solid angle. Once again there is no ϕ dependence, so the dϕ
integration gives 2π. Writing cos θ = z, the cross-section is (note that the limits have been halved, as the integrand
is even)

The integral is of the form

This identity may be obtained by differentiation of the standard integral

with respect to b. Using the expression (S6.21), we obtain, with a = 2|p||q| and b = (|p|2 + |q|2 + m2),

You can substitute , if you prefer. Note that σ → ∞ as |p| → 0.

6.4 (a) The S-matrix is defined by (7.47):

For clarity, introduce the inner product

Then

Using the anti-unitarity of ΩT, (see (6.110))

so in particular

The statement of the problem says that we are to assume

Then

which was to be shown.

(b) Using Dyson’s expansion (7.36) for the S matrix,

Change variables: let −ti = τi. We have to adjust the inequalities: if ti > tj, then −ti < −tj. There will be an extra −1
from each change of the variables of integration, and a second −1 from changing the implicit limits of integration,
e.g., , or an overall change of (−1)2 for each integral:

which was to be shown.

13
Green’s functions and Heisenberg fields
We will now consider diagrams with external lines off the mass shell. Although much of what we say will not be
restricted to our model theory, Model 3, I’ll use that continually as an example. Here is a typical Feynman graph
(the same one we looked at last time, Figure 12.6), which I choose to evaluate not just with lines on the mass shell
but with lines off the mass shell.

Figure 13.1: O(g4) ϕ + ϕ → ϕ + ϕ scattering

That is an interesting object because it could be an internal part of a more complicated Feynman graph, as I
explained at the end of the last lecture. For simplicity I will deal only with graphs with external meson lines. The
extension to graphs with both external mesons and nucleon lines, or more complicated kinds of theories when you
have 17 kinds of particles in 17 different kinds of external lines, is trivial. I will not assume that the only particles in
the theory are the mesons, just that the only graphs we’re going to look at are those with external meson lines.

13.1The graphical definition of (n)(ki)

I define the four-point function (4)(k

1, k2, k3, k4) to be the sum of all graphs with four external meson lines to
all orders of perturbation theory. I will indicate this sum graphically by a shaded blob, as in Figure 13.2.

All of the external momenta are labeled. As in our discussion of crossing, all k’s are oriented inward and, by
energy-momentum conservation, their sum must be zero. Because we’re off the mass shell, and dealing with
spacelike as well as timelike momenta, there’s no point in adopting any other orientation convention.

Figure 13.2: Graphical representation of (4)(k1 , . . . , k4)

I have some freedom about how to define these graphs. I define them to include: all connected graphs, all
disconnected graphs, all delta functions, including the overall energy-momentum conserving delta function which
we previously have been factoring out of our Feynman graphs, and all propagators (including those on the
external lines). The disconnected graphs are rather trivial for the four-point function (4)(k ) (see Figure
i
13.4), but of course we’re going to consider things with more than four lines shortly. Putting the propagators on the
external lines is just a convenience if the blob is a internal part of some more complicated graph, like this one:

Figure 13.3: The blob as an internal graph

I draw a dotted line about the internal part I’m studying. It’s a matter of convenience whether I put the propagators
on these lines inside the blob, within the dotted line, or outside the blob. I’ll put them inside the blob. That will turn
out later to be convenient.

To give a definite example, let me write down the first few graphs that contribute in our theory to (4)(k ,...,
1
k4), the first few that are inside the blob. You could have zeroth-order contributions in which all that happens is
that the four lines go right through the blob and don’t interact at all, plus two permutations depending on whether I
match up k1 with k2, k3 or k4. And there would be fourth-order corrections, including Figure 13.1 and its friends,
and higher-order corrections:

Figure 13.4: The series for (4)(k

1 , . . . , k4)

(Meson–meson scattering in our theory begins with O(g4), although nucleon scattering processes begin in O(g2).)
Analytically this equation is

the dots indicating two permutations corresponding to the two other ways I can pair up momenta with k3, plus
terms of order g4. The momenta in the delta functions are plus because all the momenta are oriented inward. The
k21 could just as well be k23, equally the k24 could be k22; it doesn’t matter because of the delta functions.

If you have an expression for (4) off the mass shell, you have it on the mass shell as well, simply by putting
the lines on the mass shell. We can, if we know (4), compute the corresponding S-matrix element. In the
particular case we’re studying at the moment, we have

with the momenta k1, k2, k3 and k4 now on the mass shell. The product of the factors k2r − µ2 is to cancel out the
four propagators we’ve put on the outer lines by convention; we now take them off. The argument of (4) is
symmetric; how I arrange the momenta doesn’t matter. I’ll say (4)(−k −k4, k1, k2). What results is just our
3,
old formula for the S-matrix element again. Please notice that the three disconnected graphs I wrote down that
arise in zeroth order make no contribution to the S-matrix, as indeed they should not, because they each have
only two propagators, as in (13.1), two pole factors, but we have four factors of zero in front of them, and therefore
they get completely canceled out.

So this is our rule. If you have (4)(k ), the Feynman diagrams on the mass shell are obtained by taking
i
the Feynman diagrams off the mass shell, canceling out the propagators we put in by convention, and putting the
lines on the mass shell. We define (n)(k ) in exactly the same way for n external lines (restricted here to
i
mesons), directed inwards:

As with (4), the functions (n) follow these conventions:

1.The momenta ki are oriented inward.

2.The external lines include propagators (k2i − µ2 + iϵ)−1.

3.All 4-momentum conserving delta functions are included.

4.All connected graphs are included.

5.All disconnected graphs are included.

As you might guess from the twiddle, we can define (n)(k ) as the Fourier transform of some object,
i
G(n)(x 1 , . . . , xn):
Since all of the (n)’s are even functions of the momenta, it hardly matters what signs I use for the exponents in
the Fourier transform, but I want to be consistent in my notation, defined in (8.63).

13.2The generating functional Z[ρ] for G(n)(xi)

We can attach a second meaning to (n)(k ) by changing the Hamiltonian of our theory to consider a
i
combined version of, for instance, Model 3, or some general theory involving a scalar field, and Model 1. That is to
say, we can take H and imagine changing it, by adding to it:

where ρ(x) as usual is some smooth function that vanishes as x goes to infinity. Then if we are to compute á0|S|0ñ
(or any S-matrix element) in the presence of ρ, we have a new diagram in our theory which I could indicate by a
dot, with a single line coming out of it: If I orient

Figure 13.5: Feynman graph for an interaction term ρ(x)ϕ(x)

the momentum k to move outwards, so it will fit onto other things where k is going inwards, it is easily seen to be −i
(−k): Or, since ρ is a real function, this could just as well be written

Figure 13.6: Feynman graph for an interaction term ρ(x)∗ϕ(x)

−i (k)∗. That is the value of that vertex we obtained in Model 1.1

If we now consider the matrix element á0|S|0ñ in the presence of this source, ρ, we can expand things in a
power series of our new vertex, imagining we have already summed up all powers and all of our old vertices. For
example, to fourth order in ρ, what we get is shown in Figure 13.7. This blob is precisely the same, Figure 13.4, as
we defined before. You have the four lines coming out, and they can do whatever they want with each other, so
long as there’s no ρ involved, because we’re only going to fourth order in ρ for this particular expression.

Figure 13.7: (4)(k

1 , . . . , k4) with (k)

We define á0|S|0ñ in the presence of the source ρ to be a functional of ρ, a numerical function of ρ, which we
call Z[ρ]:

We say “functional” rather than “function”, because of a dumb convention that a numerical function of a function is
called a functional. Often the convention is to use square brackets for the argument of a functional: Z[ρ].

There is a residual combinatoric factor of 1/n! because this is a vacuum-to-vacuum diagram, so our usual
arguments that all the n!’s cancel do not apply. Why this factor is 1/n! is easy to see. If we imagine restricting
ourselves to the case where the first ρ gives up momentum k1, the second gives a momentum k2, etc., then all of
our lines are well-defined, and we have no factor of 1/n!. On the other hand, when we integrate over all k’s in this
expression, we overcount each those terms n! times, corresponding to the n! permutations of a given set of k’s,
and therefore we need a 1/n! to cancel it out. I know combinatoric arguments are often not clear the first time you
hear them, but after a little thought, they become clear.

This formula (13.6) is so set up that it can also be written as a formula in position space simply by invoking a
generalization of Parseval’s Theorem, (9.32),

the last equality following if g(x) is a real function. Then, since ρ(x) is a real function,

The G(n)(x1 . . . xn)’s now reveal their second meaning, as Green’s functions, objects that give the response
of a system (in this case, the vacuum) to an external perturbation (here, ρ(x)ϕ(x)). George Green of Nottingham
introduced the concept in the early 19th century for a linear system, so he only had a single Green’s function. Now
we have a system that has a possible nonlinear response, and therefore we have an infinite power series in
powers of ρ. That’s why we denote these functions with G’s, in honor of Green.2

An amusing feature of the formula (13.8) is that all physical information about the system, at least concerning
experiments involving mesons, is embedded in the single functional Z[ρ]. If you know Z[ρ] for an arbitrary ρ, then
you know the G(n)’s. And if you know the G(n)’s, then you know the scattering amplitudes. It’s a fat chance you’ll
know Z[ρ] for an arbitrary ρ. Nevertheless, it’s sometimes formally very useful. Instead of manipulating the infinite
string of objects on the right-hand side of (13.8), it can be easier to work with the single object Z[ρ]. We’ll give
some examples of that.

Z[ρ] is sometimes called a generating functional. This terminology comes from the theory of special functions.
In working with, say, Legendre polynomials, it’s convenient to have a generating function, a single function of two
variables. When you do a power series expansion in one of the variables, you obtain the Legendre polynomials as
the coefficients of the powers of the variables. That’s useful in proving things about special functions. Z[ρ] is the
same sort of thing: If we expand Z[ρ] out in a power series of the ρ’s, the coefficients are the Green’s functions:

You can play cunning tricks with these generating functionals. Although that’s not really the point of this lecture, I
cannot resist a digression. It is easy to write down a generating functional that gives you not the full set of Green’s
functions but only the connected Green’s functions:

where the c means that the expression includes connected graphs only. That’s our old exponentiation theorem,
(8.49). Remember, the sum of all Feynman graphs for á0|S|0ñ is the exponential of the sum of the connected
Feynman graphs. This relation is often written as

where iW[ρ] = Zc [ρ] is the sum of the connected Feynman graphs. So if you want the generating functional for the
connected Green’s functions, the sums of the connected graphs, you just take the logarithm of Z[ρ]. We won’t
use this formula immediately, but it is so cute and its demonstration so easy, I could not resist putting it down.

13.3Scattering without an adiabatic function

Thus far our discussion of Green’s functions and the generating functional has been in the framework of our old
theory, where the interaction Hamiltonian is adiabatically turned on and off with the function f(t). The reason I’ve
gone through these manipulations is to get a formulation that I can extend to the case where f(t) is abolished, i.e.,
set equal to one. We forget about all of our old theories, and start afresh on the problem of computing the S-matrix.

We begin with

That is, we now set f(t) always and forever equal to one. No more will we talk about an adiabatic turning-on-and-off
function. I can however still take my Hamiltonian and add to it a source term involving ρ(x), a c-number space-time
function which I control:
I now redefine Z[ρ] as the amplitude for getting from the physical vacuum |0ñP to the physical vacuum, in the
presence of the source ρ:

where U I(∞, −∞) is the Schrödinger picture U I operator. This is different from (13.6) because, for the moment, I
don’t want to talk about bare vacua. The physical vacuum is the real vacuum, the ground state of the Hamiltonian.
I will assume I have normalized my theory so that the physical vacuum has energy zero:

(The Hamiltonian H in (13.15) does not include ρ(x). If it did, it would be a time-dependent Hamiltonian, and there
would be no well-defined ground state.) We will introduce a normalizing constant to give the physical vacuum
norm 1:

Equation (13.14) is our new definition of Z[ρ]. There is no bare vacuum |0ñ in the picture. I have this real, honest to
goodness theory, without the artifice of the adiabatic function. I make the theory even more complicated by adding
the term ρ(x)ϕ(x). I start with the vacuum state. I then wiggle my source ρ(x) and I ask, what is the amplitude that
I’m still in the vacuum state? I don’t write (13.14) in terms of the S-matrix because I don’t know yet what the S-
matrix is (remember, in Section 7.4, we introduced the function f(t) to facilitate the definition of the S-matrix). As
before, I define (n)(k ) and G(n)(xi) as successive terms in a power series expansion of Z[ρ] in powers of ρ.
i

I now want to ask two questions.

Question 1.Are the (n)(k )’s given by the formal sum of the Feynman graphs, as with our first
i
scattering theory?

Z[ρ] does not have the same definition as before, but of course it’s not the same theory: we no longer have an
adiabatic turning-on-and-off function. This is a question linked to perturbation theory. (Whether the sum converges
is not a question I strive to answer in this lecture or indeed in this course.) We will answer this question shortly,
and the answer is yes.

Question 2.Is (13.2) still true in the new theory, without the adiabatic function f(t)?

It is clear that the object on the right-hand side of (13.2) is well-defined without reference to perturbation
theory, without expansion in any coupling constant lurking inside H. Maybe we found this object by being a genius
in summing up perturbation theory; maybe an angel flying through the window gave it to us on a golden tablet. To
put the second question another way: Do we get the S-matrix element from (13.2)? This question has nothing to
do with perturbation theory. The full answer will have to wait till next time, but I’ll tell you now: it is almost true.
There is a correction.

This program will give us what I described in an earlier lecture as a real scattering theory: one where you
have a formula, (13.2), that tells you how to extract the S-matrix elements if you can solve the dynamics exactly: if
you can obtain the (n)’s. You could find them from perturbation theory (the answer to the first question) and

thus develop perturbation theory for S-matrix elements. However, if you have some other approximate method for
solving the dynamics—a variational method, Regge poles, dispersion relations, maybe some brand new method
from the latest issue of Physical Review Letters—it doesn’t matter; it just means you have a different
approximation for the right-hand side. This formula (13.2) is exact (apart from the correction), and you can feed in
the (n)’s from your preferred method to get the approximation for the S-matrix element.

We’ll actually construct in and out states, with a certain amount of hand-waving, to find the S-matrix element
as the inner product of an in state and an out state, as in (7.47), when I was sketching out non-relativistic
scattering theory. I will then show that the S-matrix element is, aside from a correction factor, given by the right-
hand side of (13.2). The correction factor is called wave function renormalization. We will have defined the S-
matrix without an adiabatic turning-on-and-off function.

13.4Green’s functions in the Heisenberg picture

I now turn to Question 1. I will first develop a formula, independent of perturbation theory, for these Green’s
functions that will be extremely useful for comparison with the corresponding series from perturbation theory. I
have a Hamiltonian, H + ρϕ. I will investigate this Hamiltonian not by Wick’s theorem, but by Dyson’s formula. I’ll
split it up in a rather peculiar way:

treating the source term ρ(x)ϕ(x) as the interaction Hamiltonian H I, and the original Model 3 Hamiltonian H as if it
were H 0. I’ve put quotes around it temporarily, because later I’m going to break H up into the original free
Hamiltonian H 0 plus the Model 3 interaction, which I’ll call H′;

We have the freedom to do this, because Dyson’s formula says we can divide things up into a free part and an
interacting part any way we please. In this way of doing things, when ρ = 0, the interaction picture field is the real,
honest to goodness Heisenberg field, because the interaction picture field ϕ I(x) is always the Heisenberg field
ϕ H(x) when you throw away the interaction Hamiltonian:

Thus we can apply Dyson’s formula to compute Z[ρ] in exactly the same way as we used it to obtain the S-
matrix in Model 3, (8.1) (though we will put Z[ρ] in terms of the U I matrix, as we haven’t talked about the S-matrix,
yet):

This is the time-ordered exponential of (−i) times the integral of the interaction Hamiltonian (with quotes
understood) of fields in the interaction picture (again, with quotes understood). It’s the same old Dyson formula;
I’ve just broken things up into a free part and an interacting part in a different way. Z[ρ] can be expanded as a
sequence of powers in ρ. I can’t use Wick’s theorem because the Heisenberg field doesn’t have c-number
commutators for arbitrary separations. But I can still expand the power series:

Comparing this formula (13.21) for Z[ρ] with the previous expression (13.8), we see that

So we have a third meaning for the blobs, the Green’s functions G(n)(x1 , . . . , xn): They are, in position space,
simply the physical vacuum expectation values of the time-ordered product of a string of Heisenberg fields ϕ H(x1),
. . . , ϕ H(xn). In (13.21), you’ve got G(n) defined as in (13.8), except that Z[ρ] is now given in terms of the physical
vacuum |0ñ and the U I operator, instead of the bare vacuum |0ñ and the S-matrix. The expressions are term by
term equal, if we make the identification (13.22). All the other factors, the minus i’s and the n!’s, come out right. Of
course, that is a consequence of choosing the right notational conventions originally.

This is one side of Question 1. We’ve defined G(n)(xi) in (13.22). The other side of that question is: What
corresponding quantities G(n)Feyn(x1 , . . . , xn) do we get by summing up Feynman graphs? Are they the same?

Let Z[ρ]Feyn denote Z[ρ] as we would compute it by summing the Feynman graphs, and the quantities
G(n)Feyn(x1 . . . xn) will be defined to be the coefficients of powers of ρ, as before. We will show that Z[ρ]Feyn is
equal to the original Z[ρ]. The expression for Z[ρ]Feyn is

where H I is the old Model 3 interaction Hamiltonian,

We only restrict the time limits of the integral; the space limits go from −∞ to +∞. The numerator approaches the
vacuum expectation value of U I(∞, −∞), the sum of all the Feynman graphs for the vacuum-to-vacuum transition in
the presence of ρ. The denominator is the same thing without the ρ term, the sum of all vacuum-to-vacuum
graphs in the absence of ρ. It cancels out the disconnected vacuum bubbles that may be in our graphs. You may
say, “Oh, there’s no need to do that because we’ve got our counterterm to normalize the energy properly, and the
disconnected vacuum bubbles are removed.” That’s what I said earlier. But that applies to a theory with an
adiabatic function. As we will see in a moment, this denominator indeed cancels out the disconnected vacuum
bubbles in the real theory, without an adiabatic function.3

Please notice it is the bare vacuum appearing in (13.23), and not the physical vacuum. In our derivation of the
Feynman rules, we used the interaction picture fields, free fields. We shoved all the free particle ap’s to the right,
and all the free particle a†P’s to the left where they vanished, because they encountered the bare vacuum. To
show that G(n)Feyn(x1 , . . . , xn) is the same as the real G(n)(x1 , . . . , xn), we will need to figure out what turns the
bare vacuum into the physical vacuum.

We expand Z[ρ]Feyn in powers of ρ, and obtain

where

Both sides of (13.25) are symmetric under interchange of the arguments x1 to xn. With no loss of generality I will
take these things to be time-ordered; to wit, t1, the time part of x1 to be greater than or equal to t2, the time part of
x2, all the way down to tn. Since t+ and t− are going to plus and minus infinity, I might as well begin evaluating my
limit when t+ is greater than all of the ti’s and t− is less than all of the ti’s. In this case the time ordering of the
objects within the numerator is rather trivial. We can write the numerator as (using the definition (7.36) of U I(t, t′))

The denominator is simply á0|U I(t+, t−)|0ñ, so the Feynman Green’s functions can be written

The group property (7.26) of the U I’s tells me that

We also know that we can use the U I to find the Heisenberg fields in terms of the interaction picture fields. The
correspondence (7.15) between Heisenberg and Schrödinger picture operators says

where, from (7.9), U(t, t′) = e−iH(t−t′). We also have a correspondence (7.20) between the Schrödinger and
interaction pictures,

Combining (13.30) and (13.31) we obtain

from (7.31).

We see now that we can get at least part of (13.22) in the Feynman expression (13.26), if we break up each of
the U I’s into going from one time ti to zero, and then from zero to the next time ti+1. We find associated with each
ϕ I exactly those operators required to turn it into a ϕ H, and we will obtain a string of Heisenberg fields:

We can thus write

There are no U I’s in between, it’s just a string of ϕ H’s, time-ordered because of our convention. I’ve broken up the
denominator in the same way.

We are halfway there. We have almost the same expression here in (13.34) as we have in our definition
(13.22). Things are automatically time-ordered by our convention on how we’ve arranged the x’s. We’ve regained
the Heisenberg fields. The only thing is, instead of the physical vacuum, we have this funny quantity, the bare
vacuum, and a leftover U I matrix. The algebra may be dull, but I hope it is not obscure. There are, it’s true,
technical difficulties when one has derivative interactions, when π’s as well as ϕ’s enter the interaction
Hamiltonian H I. I will ignore those technical difficulties for the moment. Much later on, when we encounter realistic
theories with derivative interactions, like the electrodynamics of spinless mesons, I will devote a lecture to
straightening everything out for derivative interactions.

We now have to worry about what happens as t+ and t− go to ±∞. Much as we hate to do it, there will be times
in this course when we have to think seriously about limits, and this is one of them.

We have two limits. We’ll take them one at a time. It will later turn out that it doesn’t matter what order we take
them in. We’ll hold t+ fixed, and consider the limit as t− → −∞. First, the numerator. Regard the bra á0|U I(t+,
0)ϕ H(x1) · · · ϕ H(xn) as a fixed state áψ| for the moment:

We can do the same thing for the denominator, letting á0|U I(t+, 0) be the fixed state áχ|. We have

because the bare vacuum is an eigenstate of the free Hamiltonian with eigenvalue 0. There is a complete set of
states, |nñ, of the Hamiltonian H;

In particular, the physical vacuum |0ñP is one of these states, with

Every state of this set except the physical vacuum |0ñP is a continuum state. I’ll separate that out. If we now insert
this complete set into (13.36), we obtain

The sum on n is really an integral, but I use standard quantum mechanics conventions and write it as a sum. Thus
our limit becomes

What do we have here, in the sum? We have a continuum integral of oscillating terms. There are one-particle, two-
particle, three-particle energy eigenstates, but they’re in the middle of a continuum. Now there’s a well known
theorem from Fourier analysis that says a continuum integral of oscillating terms goes to zero, as t goes to infinity:
all the oscillations cancel out. This is known as the Riemann–Lebesgue lemma.4 Physically, the
Riemann–Lebesgue lemma says that if you take the inner product of a state with a fixed state in some fixed region
and wait long enough, the only trace of that state that will remain is its true vacuum component. All the one-
particle states and multiparticle states will have run away.
 Consequently

This result states that the time limit makes the bare vacuum into the physical vacuum. The denominator goes the
same way. By exactly the same reasoning, the Riemann–Lebesgue lemma applies to the other limit, as t+ goes to
infinity, with the result

because the factors á0|0ñP and Pá0|0ñ cancel, and the norm of the physical vacuum is 1. The time ordering symbol
is of course irrelevant because we have arranged things so that xi is later than xi+1 for all values of i.

We have answered Question 1 in the positive. This thing we get by summing up all the Feynman graphs is
indeed the actual Green’s function as we have defined it. This result is tricky but it’s pretty. The tricky part is this:
by taking the time limit, we wash out everything except the real physical vacuum state.

13.5Constructing in and out states

We turn now to Question 2. How do we construct the S-matrix without the adiabatic function? Given these G(n)’s,
how do we compute the S-matrix in terms of them? This question has nothing to do with perturbation theory, and
nothing to do with breaking the Hamiltonian up into two parts. We won’t be able to answer it until next time. We
first have to figure out how to construct in and out states.

Since I will always be working in the Heisenberg picture, for the remainder of this lecture and the first part of
next lecture, I will denote ϕ H(x) just by ϕ(x), the Heisenberg picture field:

Also, as the physical vacuum |0ñP is the only vacuum we’ll be talking about, we will set

The physical vacuum satisfies these conditions:

The vacuum is an eigenstate of the energy and momentum operators, with eigenvalues zero, and it is normalized
to one. I assume we have physical one-meson states |pñ in our theory. (If the meson is unstable, there’s no point in
trying to compute meson–meson scattering matrix elements.) I will relativistically normalize them:

These states are eigenstates of the momentum operator:

where µ is the real, physical mass of a real meson. Those are just notational conventions. I won’t write down the
normalization for a two-meson state now, because a two-meson state could be an in state or an out state, and
they aren’t the same things; a state that looks like two mesons in the far past may look like a nucleon and an
antinucleon in the far future. One of the problems we’re going to confront is how to construct those states. We’ll
have troubles enough with just the vacuum and the one-particle states.

Because the computations we’re going to go through are long, I should give you an overview of what we’re
going to do. We’re going to be inspired by our previous limiting process. There we saw how we could pluck out the
vacuum state by taking some object involving finite times, and going to a limit. I’m going to do this same sort of
thing again. Since our field operators are interacting, they’re not going to make only one-particle states when they
hit the vacuum. They’ll make one-particle states, two-particle states, three-particle states, and 72-particle states;
they’re capable of doing a lot. We’re going to make several definitions to construct a limit in time that will enable
me to get, from the field operator hitting the vacuum, only the one-particle part. If I can do that, I will have crafted
something like a creation operator for a single particle. And then I will be able to use these “creation operators”
next time to create states that look like two-particle states, either in the far past or the far future, by making a time
limit go to −∞ or ∞, respectively. All that will be shown in detail. Our first job is to find a time limit that makes
exclusively a one-particle state.

We will need some conventions about the scale of our field. I’m going to work with these Heisenberg fields,
without using any details of the equations of motion, just the fact that there are equations of motion; ϕ(x) is a local
scalar field. I’m not even going to say this field obeys the canonical commutators. I will require two normalization
conditions.

The first condition concerns the (physical) vacuum expectation value of the Heisenberg field. By translational
invariance, (condition 3, just before (3.4)) this will be independent of x:

I require my field to have a vacuum expectation value of zero. If it is not zero, I will redefine the field, subtracting
from it the constant á0|ϕ(0)|0ñ:

Second, I need to specify the normalization of the one-particle matrix element. Because these one-particle
states are momentum eigenstates,

Since Lorentz transformations don’t change ϕ′(0), or change any one-meson state to any other one-meson state,
the coefficient ák|ϕ′(0)|0ñ of eik·x must be Lorentz invariant, and so can depend only on k2. Presumably, the one-
particle state is on its mass shell. Then k2 = µ2, and ák|ϕ′(0)|0ñ is a constant. By convention this constant is
denoted by

(The notation comes from one of Dyson’s classic papers5 on quantum electrodynamics, in which he defined three
quantities {Z1, Z2, Z3}. If we were to treat a one-nucleon state the same as a one-meson state, the equivalent
constant for a one-nucleon state would be called Z2. We won’t get to Z1 for weeks, so don’t worry about it.) Now
redefine ϕ′(x) by

where ϕ s (x) is the subtracted field. I will assume Z3 is not zero, so this definition makes sense. Then ϕ′(x) has the
property that it has the same matrix element between the physical vacuum and the renormalized one-particle
state as a free field has between the bare vacuum and the bare one-particle state:

These two conditions, (13.49) and (13.52), are just matters of definition. ϕ′(x) is called the renormalized field, if
ϕ(x) is the canonical field, obeying canonical commutators. It’s called “renormalized” for an obvious reason: we
have changed the normalization. Z3 is called, for reasons so obscure and so embedded in the early history of
quantum electrodynamics I don’t want to describe them, “the wave function renormalization constant”. It should be
called “the field renormalization constant”.

I can now tell you what the “almost” in the answer to Question 2 means. Even without the adiabatic function,
the naive formula (13.2) is almost right. The only correction is that the Green’s functions are those of the
renormalized fields, not those of the ordinary fields. These Green’s functions differ from the earlier versions by
powers of Z 3−1/2. In due course, we will establish the right formula for the renormalized fields.

The renormalized fields have been scaled in such a way that if all they did was create and annihilate single-
particle states when hitting the vacuum, they would do so in exactly the same way as a free field. They do more
than that, however, and therefore we’ve got to define a limiting procedure. It’s actually not so bad. Most of our
work will consist of writing down a bunch of definitions, and then investigating their implications.

Unfortunately, we would get into a lot of trouble if we were to try and do limits involving plane wave states, so I
would like to develop some notation for normalizable wave packet states:

Associated with each of these wave packets is a function f(x),

which is obtained with exactly the same integral as the ket |fñ, but whose integrand has e−ik·x instead of the ket |kñ.
For reasons that will become clear in a moment, I don’t want to denote F(k) by (k). This function f(x) is a positive
frequency solution to the free Klein-Gordon equation:

We also have

Furthermore, if the one-particle state |fñ goes to a plane wave state |kñ, F(k′) goes to (2π)32ωkδ(3)(k − k′), and f(x)
goes to the plane wave solution e−ik·x . I’ve arranged a one-to-one mapping such that our relativistically normalized
states correspond to plane waves with no factors in front of them.

I’m now going to define an operator that at first glance looks absolutely disgusting:

Remember, ϕ′(x) is a Heisenberg field, a function of x and t; this produces a function of t only. We can say some
things about this object. In particular, we know its vacuum-to-vacuum matrix element:

We can also work out its one-particle matrix elements:

(using (13.53) in the second step), so that, as part of an inner product with a one-particle bra, we can say

A calculation analogous to (13.60), but differing in one crucial minus sign, gives

Thus this operator ϕ′f(t) has time-independent matrix elements from vacuum to vacuum, and from vacuum to any
one-particle state; the time-dependent phases cancel in (13.60). In fact, if we just restrict ourselves to the one-
particle subspace at any given time, ϕ′f(t) is like a creation operator for the normalized state |fñ.

What about a multiparticle state? Suppose I take a state |nñ with two or more particles, and total momentum
pµn;

The matrix element of the state |nñ with our new creation operator ϕ′f(t) can be worked out in exactly the same
way. There is a small complication in that we don’t know the normalization of án|ϕ′(x)|0ñ:

and we don’t know what án|ϕ′(0)|0ñ is, yet. In terms of this quantity,
The real killer is in the exponential, e−i(ωpn − En)t. A multiparticle state always has energy En > ωpn, more energy
than a single particle state with momentum pn. For example, a two-meson state with pn = 0 can have any energy
from E = 2µ to infinity. The one-meson state with p = 0 has energy E = µ. So the exponential will provide the same
sort of oscillatory factor as we saw in (13.40). Thus we can use the same argument with the operator ϕ ′f(t) as we
did with the U I matrix, (13.41):

by the Riemann–Lebesgue lemma, provided |nñ is a multiparticle state.

Let áψ| be a fixed, normalizable state, and consider the limit as t → ±∞ of the matrix element áψ|ϕ′f(t)|0ñ:

For any fixed state |ψñ sitting on the left of the operator, the matrix element with the vacuum state will give us
nothing, by (13.59); the matrix elements with the one-particle states will give us F(k), independent of time, by
(13.60); and everything else in the whole wide world will give us oscillations which vanish, by (13.66). Thus

So this is exactly analogous to the formula (13.60) we found with the one-particle state |kñ sitting on the left. The
operator just projects out the part F(k) and gives you áψ|fñ. That is, we have something that can act either in the far
past or the far future as a creation operator for a normalizable state |fñ.

An analogous calculation gives

because the arguments of the exponentials add and never cancel, for every single particle or multiparticle
momentum eigenstate.

Now this procedure looks very tempting as a prescription for constructing two particle in states and two
particle out states, and to find S-matrix elements. I will yield to that temptation at the beginning of the next lecture.

1[Eds.] See note 5, p. 258.

2[Eds.] Feynman’s propagators were the first systematic use of Green’s functions in quantum field theory. R. P.
Feynman, “The Theory of Positrons”, Phys. Rev. 76 (1949) 749–759. The introduction of sources to obtain them
was pioneered by Schwinger in a series of papers: “On gauge invariance and vacuum polarization”, Phys. Rev. 82
(1951) 664–679; “The theory of quantized fields I.”, Phys. Rev. 82 (1951) 914–927; “The theory of quantized fields
II.”, Phys. Rev. 91 (1953) 713–728. All of these papers may be found in Schwinger QED. For accessible
introductions to Green’s functions, see J. W. Dettman, Mathematical Methods in Physics and Engineering, 2nd
ed., McGraw-Hill, 1969, Chap. 5; and F. W. Byron and R. W. Fuller, Mathematics of Classical and Quantum
Physics, Addison-Wesley, 1970, Chap. 7. Both of these texts have been reprinted by Dover Publications.
3[Eds.] Also, to agree with (13.6), we should have Z[ρ] = 1 when ρ = 0. The denominator ensures this.
4[Eds.] See M. Spivak, Calculus, 3rd ed., Publish or Perish, 1994, Problem 15.26, p. 317, or W. Rudin, Real and
Complex Analysis, McGraw-Hill, 1970, p, 103. The coefficients of e−iE nt do not need to be continuous, but only
integrable.
5[Eds.] Freeman J. Dyson, “The S matrix in quantum electrodynamics”, Phys. Rev. 75 (1949) 1736–1755, and
reprinted in Schwinger QED. The constants Z1, Z2 and Z3 are introduced on p. 1750.

Problems 7

7.1 In class we derived (12.24), the two-particle density of states factor, D, in the center-of-momentum frame, PT =
0,

where the notation is as explained in §12.2. Find the formula that replaces this one if PT ≠ 0. Comment: Although
the center-of-momentum frame is certainly the simplest one in which to work, sometimes we want to do
calculations in other frames, for example, the “lab frame”, in which one of the two initial particles is at rest.
(1997a 7.1)

7.2 Let A, B, C, and D be four real scalar fields, with dynamics determined by the Lagrangian density

where m and g are positive real numbers. Note that A is massive while B, C, and D are massless. Thus the decay
of the A into the other three is kinematically allowed. Compute, to the lowest non-vanishing order of perturbation
theory, the total decay width of the A. What would the answer be if the interaction were instead gAB3? H INT: The
trick here is to find the kinematically allowed region in the EB − EC plane. Some of the constraints are obvious: EB
and EC must be positive, as must ED = m − EB − EC. One is a little less obvious: cos θBC (called θ12 in class) must
be between −1 and 1.
(1997a 7.2)

7.3 In class I discussed how to compute the decay of a particle into a number of spinless mesons, assuming the
universe was empty of mesons before the decay. Sometimes (for example, in cosmology), we wish to compute the
decay of a particle (at rest), not into an empty universe, but into one that is filled with a thermal distribution of
mesons at a temperature T. This is not hard to do, if we treat the mesons in the final state as non-interacting
particles (frequently a very reasonable approximation), and assume there are no other particles of the same type
as the initial particle in the initial distribution of particles. (This frequently happens in cosmology. For example, the
initial particles could be very massive and were produced [in thermal equilibrium] at an early epoch when the
temperature is very high. The expansion of the universe rapidly brings these particles out of equilibrium and
reduces their density to a negligible value. They then decay in an environment consisting of a hot gas of much
less massive particles.) Show that in this case, the only change in the formalism presented in class is that the
density of states factor, D, has an additional multiplicative factor, f(E/kT), for each final meson, where E is the
meson energy, k is Boltzmann’s constant and f is a function you are to find.

Possibly useful information: (1) For any system in thermal equilibrium, the probability of finding the system in
its nth energy eigenstate is proportional to exp(−En/kT). (2) For a single harmonic oscillator, án|a†|n − 1ñ = (see
(2.36)).

Cultural note: There are problems in which one has to use the results of Problem 7.1 together with those of
this problem (extended to the case in which there is an initial-state thermal distribution, as well as a final-state
one). One famous example is the scattering of high energy cosmic ray protons off the 3 K cosmic microwave
background radiation.
(1997a 7.3)
 Solutions 7

7.1 Let the total energy be ET and the total momentum PT, and let the 3-momenta of the final particles be k and q.
Let the corresponding masses and energies be mk , Ek and mq, Eq. The two-particle density of states is, from
(12.16),

Integrate over q, using the final delta function:

Let θ be the angle between k and PT, and ϕ the azimuthal angle of k about the PT axis. Then

In the argument of the delta function only Eq depends on θ. Using the identity in Footnote 8 on p. 9,

where θk is the value of θ at which (Ek + Eq − ET) = 0. Now q = PT − k means

so

and hence

We can now do the cos θ integral:

(using |k| d|k| = Ek dE k in the final step). Everything (including whatever may multiply D) is to be evaluated at q =
PT − k, and θ = θk . Determine θk by putting Eq = ET − Ek into (S7.5),

and solving for cos θk :

For PT ≠ 0, the density of states is given by (S7.8), with the restrictions noted above, and θk given by (S7.10). (For
the decay of a particle of mass M, E2T − |PT|2 = M2.)

7.2 The decay width Γ is given by (12.2),

The amplitude Afi is given graphically and analytically by

The density of states factor is given, for a final state of three spinless particles, by (12.65),

so

The task now is to determine the kinematically allowed region. By conservation of momentum and energy,

Also, because B, C, and D are massless, we have

By the Triangle Inequality,

so that, substituting,

Add EB + EC to each, and divide by 2, to obtain

The allowed region in the EB − EC plane is triangular, with an area of m2:

Plugging this area into the decay width gives

Observe the large difference from the naive “dimensional analysis” guess of g2m; 512π3 ≈ 16,000.

If the interaction had been gAB3 instead of gABCD, there would have been no distinction between the three
fields B, C, and D. We would have had

and we might have naively expected the amplitude to increase by (3!)2. But integrating over all final states now
over-counts by a factor of 3!, since the outgoing particles are indistinguishable. We must divide by 3!. The new
answer is

This decay width is 3! or six times the earlier value.;

7.3 We will use the index i as a generic label to distinguish mesons according to their particle type as well as their
momentum. For simplicity, we can imagine that we are working in a box with discrete momenta. We will consider
the decay of a particle of type 0 into a set of particles of types 1, 2, . . . , j. By assumption, the original state has
only a single particle of type 0, but may have ni of type i. The relevant matrix elements of the Wick expression for
this process are of the form

Let us assume that the ‘background’ state has ni particles of type i. Recalling that

we see that the decay amplitude Afi is enhanced by a factor

when compared with the analogous process with no background states. So the probability of transition,
proportional to |Afi|2, will be enhanced over the vacuum probability of decay by the square of the factor (S7.26).

The probability that there are ni quanta of type i is, by a standard thermodynamic argument,

where as usual β = 1/kT. The overall probability of decay is

This shows that the decay width has an extra factor of (1 − e−βEi)−1 for each mode created by the decay process.
We would get the same result if we were to change the density of states, (11.58), by the substitution

where

14
The LSZ formalism

Let me summarize some of the things we said last time, and the question we are trying to answer. With every
normalizable one-particle state |fñ we have associated a function F(k)

Likewise I associated with the same state a function f(x)

which is a positive frequency solution of the Klein-Gordon equation,

So I have a one to one mapping between normalizable states and solutions of the Klein-Gordon equation. For my
renormalized field operator ϕ′(x), I defined a function of time ϕ′f(t) as

and I showed that, for any fixed, normalizable state |ψñ

It will also be important that

That is: as f(x) goes to a plane wave, the state |fñ goes to a relativistically normalized momentum eigenstate |kñ.
That was the conclusion of everything we investigated last lecture. We showed that this operator ϕ′f(t) in the limit
as the time t goes to either positive or negative infinity was, so to speak, a one-particle creation operator. Of
course, at intermediate times it is by no means a one-particle creation operator. It makes, as would any smeared
version of the field operators at fixed time, not just a single-particle state, but two-particle states, three-particle
states . . . , ad infinitum, at least if we investigate in higher and higher orders of perturbation theory. Only in the limit
t → ±∞ do we cancel out all the multiparticle terms that would in principle contribute to this matrix element at any
finite time, because of the non-cancellation of phases. So that’s where we wound up.

14.1Two-particle states

We can of course get some related formulas from the result (14.5). For example, if we put the vacuum on the other
side,

then as we found in (13.69), even for the one-particle states we have a phase mismatch: in this matrix element, all
of the phases have a positive frequency and never cancel. So this limit is zero. All the phases mismatch, which is
again what you would expect if this asymptotic limit is producing something like a creation operator. A creation
operator does indeed annihilate the vacuum on the left. Of course, we have certain trivial equations that follow
from (14.5) just by taking the adjoint;

The operator ϕ′f†(t) is not Hermitian, because there’s an explicit i in the definition (14.4); moreover, f(x) is not a
real function. The adjoint equation has a limit of zero:

Again, this is what you would expect if ϕ′f is a creation operator, because ϕ′f† should then be an annihilation
operator. This is just what an annihilation operator does: it makes a one-particle state from the vacuum on the left,
and kills the vacuum on the right.

Now we come to the great leap of faith. I assume I have two functions F1(k) and F2(k), which are associated
with nice, normalized, non-interacting wave packet states |f1ñ and |f2ñ, respectively, in the sense of (14.1). We
require that the functions F1(k) and F2(k) have no common support in momentum space. That is,

By making this statement, we are leaving out only a negligible region of phase space. When we eventually let the
kets |f1ñ and |f2ñ go to plane wave states, this restriction will exclude just the configurations with two collinear
momenta, which correspond to scattering at threshold in the center-of-mass frame, a case we excluded in our
other analysis also. Thus one of these kets is associated with a one-particle state which is going off in some
direction, and the other is associated with a one-particle state going off in another direction. I’ll call the functions
and states associated with F1(k) and F2(k), f1(x), f2(x) and |f1ñ, |f2ñ, respectively.

I now want to consider what happens if I take the limit

the operator ϕ′f2(t) acting on not the vacuum now, but on the state |f1ñ. Well, (14.11) is a matrix element, which we
can think about in either the Schrödinger or the Heisenberg picture: matrix elements are matrix elements, even
though these are all Heisenberg fields. Let’s think about this operation in the Schrödinger picture. I have a state
|f1ñ, described by some wave packet, say with the center of the wave packet traveling in some direction. I wait for
some very large future time, say, a billion years. If I wait long enough, that wave packet has gotten very very far
away, maybe several galaxies over in the original direction. Now I come into this room. I have an operator which if
I applied it to the vacuum would make a state |f2ñ. If I were now to go a billion light years in the opposite direction,
carrying this operator, and hit the vacuum with it there, it would make a single-particle state with distribution f2.
That’s the physics of what is going on.

So let me ask a question. What happens if I apply it not to the vacuum, but to the state that has that other
particle over there, way beyond the Andromeda galaxy, two million light years away? Well, if there’s any sense in
the world whatsoever, the fact that that other particle is on the other side of the Andromeda galaxy should be
completely irrelevant. I’d have to travel to the other side of Andromeda to see it’s there. As far as I’m concerned, I
don’t know in the whole region of spacetime in which I’m working that I haven’t got the vacuum state. The particle
that is really there, that is secretly there, I can hardly expect to see in any experiment I can do, because it is all the
way over on the other side of Andromeda. It can’t affect what I’m doing in this room, or what I’m doing two million
light years away in the other direction. I am making a state by this operation that is effectively a two-particle state,
with the two particles in the far future moving away from each other, one going in one direction and the other going
in another direction. Therefore, I assert this limit should exist and should give the definition of a two-particle out
state, a state that in the far future looks like two particles moving away from each other:

That’s an argument, not a proof. If you want a mathematical proof you have to read a long paper by Klaus
Hepp;1 but it is physically very reasonable. The only thing I am incorporating is that there is some rough idea of
localization in this theory, some sort of approximation to position. And if there’s a particle on the other side of the
Andromeda galaxy traveling away from me, I’ll never know it.

In fact the analysis can be extended to collinear momenta, but it requires much more complicated reasoning,
and the result is not even on the rigorous level of Hepp’s argument. The physics is clear, even if the momenta are
collinear, because wave packets tend to spread out. If I wait long enough, I’ll have a negligible probability for the
first particle to be anywhere near the second particle even though the centers of the wave packets are moving in
the same direction. So it turns out it’s also true for collinear momenta. The limit will be a little slower, because in
spreading out, they’re not moving away from each other as fast as they would if their motions were pointing in
different directions.

Of course, if our limit were for the time to approach minus infinity, all the arguments would be exactly the
same, but time reversed. Instead of an “out”, I would have an “in”:

Thus we have the prescription for constructing in states and out states, states that look like two-particle states in
the far past, and states that look like two-particle states in the far future. I use two-particle states only for
simplicity. After I go through all the agonies I will go through for two particles scattering into two particles, if you
wish you can extend the arguments to two into three or two into four or seven into eighteen. We can construct
states that are indeed asymptotic states.

14.2The proof of the LSZ formula

We’re now in a position to answer Question 2 (p. 273). As I told you, the answer to Question 2 is almost “Yes”: the
relation (13.2) needs to be modified by replacing the ϕ fields with the renormalized fields, the ϕ′ fields. Analogous
to G(n)(x1, . . . , xn) and (n)(k
1, . . . , kn) defined in (13.4) and (13.22), we now define

with

Let’s look at a specific example, the four-point function ′(4)(k1, . . . , k4):

with
That’s the renormalized Green’s function, G′(4)(x1, . . . , x4), the physical vacuum expectation value of a string of
renormalized Heisenberg fields, and G(4)(x1, . . . , x4) is the old Green’s function; there’s a factor of Z3−1/2 for each
renormalized field. The question we want to test is whether the renormalized version of (13.2) is true:

We want to prove this relation, due originally to Lehmann, Symanzik and Zimmerman,2 and known as the LSZ
reduction formula.

What I will actually prove is the analog of (14.18) for wave packet states of the form (14.12) and (14.13).
Scattering is physically defined only for wave packet states; a plane wave state never gets far away from the
interaction because it has uniform probability density over all space. Let the final states be characterized by two
non-overlapping wave packets |g1ñ and |g2ñ analogous to the non-overlapping initial wave packets |f1ñ and |f2ñ.
Then what I will prove is that

I’ve put in a question mark for the time being, to indicate we haven’t yet proved it at this stage.

Now this does reduce to the statement (14.18) when we allow the f’s and g’s to go to plane waves as stated in
(14.6). When I make the f’s and g’s plane waves, I simply get the definition of the Fourier transform in (14.19), with
the momenta associated with g1 and g2 replaced by minus their natural value because I’m complex conjugating.
Operating on a function of position space with ( 2 + µ2) is the same thing as multiplying that function in
momentum space by (−k2 + µ2), and that produces the propagator factors in (14.18) except for a minus sign,
which is taken care of by replacing the (−i) in (14.18) by the i in (14.19). The sign of the i’s doesn’t matter for the
four-point function, because i 4 is (−i)4; but I want to construct the arguments so you can see easily how trivial the
generalization is to n particles in and m particles out. So if I prove (14.19), I will be home. We’ll start with the left-
hand side, and transform it into the right-hand side.

Now, in order to study the limit of (14.19) as the wave packets turn into plane waves, I will establish a useful
lemma. Say we have a function f(x) which is a solution of the Klein-Gordon equation, and which goes to zero
rapidly as |x| → ∞. That is, the wave packet |fñ to which f(x) corresponds is a nice, normalizable wave function that
dies away at infinity sufficiently rapidly that integration by parts on spatial derivatives on f(x) is legitimate. We can’t
say the same for time derivatives, because this thing is evolving in time. Let A(x) be another quantity which could
be a single field, or maybe a string of operators, with the dependence on the other variables besides x
suppressed. If I define, in analogy with ϕ′f(t) in (14.4),

then the lemma says

The proof is straightforward, starting with the left-hand side of the lemma:

Now few things are easier to do than the time integral of a time derivative, and so we obtain

by the Fundamental Theorem of Calculus, QED.

 We can establish a similar equation for the conjugate function f*. We’ll now assume A is some Hermitian
operator, A = A†, and note that

Then as you can show easily

There is a sign flip for the adjoint.

Armed with this lemma we can now turn the formidable expression (14.19) into a grotesque sequence of
limits. Let’s do the x4 integration. Using the lemma, we have

You might think I have swindled you because I’ve slipped the ϕ′f2 in here, and in doing so, pushed a time
derivative ∂/∂t4 past the time ordering symbol. As you know from Problem 1.2, that may give me a term involving
an equal time delta function, but that’s irrelevant in this limit, because if I keep {t1, t2, t3} fixed and send t4 to either
plus or minus infinity, t4 is not the same time as any of the other three times, and therefore I can push the time
derivative through the time ordering symbol without losing (or gaining) anything.

Continuing in this way, we can turn (14.26) into

It doesn’t matter which one I integrate first. That’s easy to show, if you assume that the time ordered product has
any sort of reasonable large distance fall-off, if it’s a tempered distribution or something like that. If we had
reduced the integrals in some other order, we would have had a different order of limits. In fact, all 4! orderings
lead to the same result. As an exercise you can do the limits in any other order, and see that you get exactly the
same answer.

The successive limits are actually duck soup.3 Let’s do one of them and see what happens. Let’s do the t4
limit. We’ve got two terms, one for t4 goes to −∞ and the other for t4 goes to +∞, with the arguments {x1, x2, x3}
held fixed. I have ϕ′g1†(t1), ϕ′g2†(t2), ϕ′f1(t3), and ϕ′f2(t4). Now, what happens in the limit as t4 goes to −∞? Looking
only at this part, we have

Well, t4 = −∞ is certainly earlier than any finite times. The time ordering symbol says that as t4 goes to −∞, ϕ′f2(t4)
goes all the way over on the right, where it encounters the vacuum and, from (14.5), makes the state |f2ñ:

That takes care of the t4 → −∞ limit. What about t4 → +∞? Well, plus infinity is later than any finite time and
therefore ϕ′f2(t4) is situated all the way over on the left, where it hits the vacuum and, from (14.7), gives us zero:

I don’t bother to specify whether |f2ñ is an in state or an out state, because for a one-particle state, they’re the
same: one particle just sits there, or travels along; it doesn’t have anything to scatter off of. So far, so good. We’re
getting there.

Let’s look at the t3 limit. That’s much the same story. When t3 → ∞, it is the latest time, and thus the time
ordering puts ϕ′f1 on the extreme left, where it hits the vacuum and produces zero. When t3 → −∞, ϕ′f1 goes
against the state |f2ñ on the right, and according to (14.13) produces the state |f1, f2ñin. (It’s definitely an in state,
because both creation operators were at a time of minus infinity.) We wind up with

Now let’s look at the t2 limits. There’s a limit as t2 goes to +∞. Because of the time ordering symbol, the
operator ϕ′g2†(t2) ends up on the extreme left where, from (14.8), it makes a one-particle state, ág2|. Ignoring the t1
limits for a moment, the right-hand side of (14.31) becomes

Unfortunately, there’s a term left over. In the limit as t2 → −∞, the operator ϕ′g2†(t2) winds up against the state |f1,
f2ñ. We don’t know what that is, but I’ll denote it by |ψñ:

With only one operator left, there is no more need for the time ordering symbol. Putting in the last limits, the right-
hand side of (14.31) becomes

In the second step, recall (14.8):

Now, what have we got? Aside from an ordering of g1 and g2, irrelevant because these are Bose particles, we
have proved, at the cost of rather lengthy calculation, exactly what we set out to prove. Let’s summarize where we
are. We’ve addressed the two questions raised in the last chapter (p. 273). We have answered Question 1: we
correctly compute the Green’s functions for the unnormalized fields by summing up the Feynman diagrams. And
we have answered Question 2: we correctly get S-matrix elements by putting the Green’s function lines on the
mass shell and multiplying by factors of (k2 − µ2) to get rid of the extra propagators for the renormalized Green’s
functions. That’s why I said the answer to Question 2 was not “Yes” but “Almost”. We have to use the Fourier
transforms of the renormalized Green’s functions, the vacuum expectation values of the time-ordered product of
the renormalized fields, to get the right S-matrix elements.

Now if we just consider the answer to Question 2 in isolation, without worrying about how we compute things,
we also see that we have what I described in an earlier lecture as the beau idéal of a scattering theory: a way of
finding the S-matrix elements from the finite time dynamics, without resorting to any approximation procedure.
That is given by the LSZ formula (14.18), from which I can now erase the question mark: it is correct. I don’t know
why it’s called a reduction formula, maybe because you get some reduced information from a Green’s function by
only looking at its mass-shell value in Fourier space.

The mathematical expression (14.19) makes sense even with the f and g wave packets replaced by plane
waves. We’ll make that expression the definition of an (S − 1) matrix element for plane waves. Of course we only
get something physically measurable when we smear out the plane waves into wave packets. This situation is
analogous to the expression (9.39) for electrostatic energy. No one can build a point charge, and so no one can
make a charge distribution that directly measures the Coulomb potential. All you can do is measure E0 for various
charge distributions. Then you can abstract the notion of an interaction between two point charges. The formula
analogous to (9.39) is
 The only thing that was required in deriving the LSZ reduction is that somehow we could get our hands on a
local field with a non-zero vacuum to one-particle matrix element, that makes some kind of particle out of the
vacuum. It can make any other kind of junk it wants, as long as it has a non-zero matrix element. We don’t
demand that the field satisfy the canonical commutation relations. It could be something like ϕ 2. That might be a
good one to look at for a two-particle state. Or maybe if that doesn’t work, ∂µϕ ∂µϕ, or if that one doesn’t work,
maybe it’s a 72-particle state, maybe we want to look at ϕ 70 ∂µ ϕ∂µϕ. As long as we can find such a local field for
making the desired particle out of the vacuum, we know in principle how to calculate the S-matrix element. In
practice it’s just as much a mess as before. It’s very hard to find out that there is a 72-particle bound state in a
theory, let alone to compute its mass. That’s what we need, since we’ve got to get the exact mass in the LSZ
formula; the field has to obey the real Klein-Gordon equation with the real physical mass. Using different fields
would change the Green’s functions off mass shell, but would have no effect on S-matrix elements.

The assumptions I mentioned explicitly in deriving the LSZ reduction formula have no reference whatsoever
to whether the particle we’re talking about, the meson, is a fundamental particle or a composite particle. It does
have an explicit reference to the fact that it is a spinless particle. That’s because we’ve only set up the formalism
for scalar fields, but it is fairly obvious that if I have a particle of spin one I can play the same sort of game with the
vector field, etc. Therefore this formula (14.18) contains, in addition to the correct version of perturbation theory,
the answer to how we compute S-matrix elements for composite particles like hydrogen atoms or nuclei, or
blackboard erasers. To compute this Green’s function is no easy job: it’s a complicated mess. But in principle we
have solved the problem. We shifted the field to get rid of its vacuum-to-vacuum expectation value, scaled it to put
its vacuum to one-particle matrix element in standard form and off we went. There are no problems of principle.
There are, as usual, the enormously difficult problems of practice, which you know about if you’ve ever looked at
the scattering of molecules off molecules, or problems of that kind. There’s no problem in defining the S-matrix,
although there may be severe problems in computing it. So we have found a formulation of scattering theory that
in principle is capable of describing any conceivable situation.

Other formulas can be derived using methods of the same type as those used to derive the LSZ formula. For
example, one can stop “half way” in the reduction formula and obtain

This method is used to derive theorems about the production of “soft” (low energy) pions and photons. We can
also use LSZ methods to derive expressions for the matrix elements of fields between in and out states. For
example,

Of course, this is really just an abstraction of the relation

In the same way that we showed that the right-hand side of (14.19) is equal to the right-hand side of (14.27), we
can show that the right-hand side of (14.38) is equal to

and these limits evaluate to the left-hand side of (14.38).

I should say that at the moment there is in principle no need for counterterms, except for the trivial vacuum
energy counterterm. We do want to define our theory so the energy of the vacuum is zero. But aside from that,
there is no need to introduce any counterterm. If we could solve the theory exactly, we could write down the
Lagrangian in its original form in terms of bare masses and unrenormalized fields, compute all the Green’s
functions exactly, compute the physical mass of the particles so we know what mass to use in the reduction
formula, compute the vacuum to one-particle matrix elements as we want to know how to rescale the fields, crank
our answer into the reduction formula, and off we go! In practice the counterterms will come back again and I will
talk about that shortly. But they come back again as a matter of convenience, and not as a question of necessity.

14.3Model 3 revisited

Let me return to our highly unrealistic example, Model 3. By the way, none of what we’ve done so far has anything
specifically to do with our particular example; it’s completely general. In its full glory, our example looks like this:

The quantity µ0 is the bare mass of the meson, which may have absolutely no connection with the physical mass µ
of the meson. Similarly, m0 is the bare mass of the nucleon, and m is its physical mass. The constant g0 is
something I will call a bare coupling constant, some parameter that characterizes the theory. I’ll just stick a nought
on it, you’ll learn why in a moment. We’ll compute the conventionally defined coupling constant g from g0, and then
invert the equation to eliminate g0, which is not directly measured, from all other quantities of interest. The
Lagrangian may include a trivial constant, to adjust the zero of the energy to come out right. I won’t even bother to
give it a special name, at this stage.

In principle, we could compute everything in this theory. After we had managed to solve the theory by some
analytic tour de force, we could then determine the actual physical masses and the renormalized fields. In this
case, since we have two kinds of particles around, mesons and nucleons, we have two kinds of renormalized
fields. We have the renormalized meson field defined as before:

Here |0ñ is the real physical vacuum, the only vacuum we’re talking about. By the way, I have tacitly assumed in all
of this that Z3 was chosen to be a positive real number. We are always free to do this: it’s just a statement about
how we choose the phase of the one-meson states. Thinking back, I realize that I assumed ϕ′ was Hermitian if ϕ
was Hermitian, and that’s not true if Z3 is not real. The renormalized meson field is determined by the statements
(see (13.53))

where ák| is a one-meson state, the lightest state, other than the vacuum, of charge zero in the theory.

We have to renormalize the nucleon fields likewise with an independent renormalization constant, Z2:

This constant is determined by similar equations. First,

There’s no need to add a constant to the nucleon field, because the vacuum expectation value of ψ is
automatically zero as a consequence of electric charge conservation. The nucleon field carries electric charge
one and therefore can hardly connect the vacuum with the vacuum. So we don’t have to bother shifting to impose
this condition: the symmetries of the theory impose it for us. And of course a condition similar to the meson’s scale
(14.42) holds for the nucleon,

where áp| is a one-antinucleon state.

There’s no need to impose a similar condition for ψ′† making one nucleon because this matrix element is
guaranteed to be identical to (14.45) by charge conjugation invariance, indeed by TCP invariance. So it’s the
same story as we’ve talked about before, except that we have two wave function renormalization constants
because we have two kinds of fields in the theory.

We still have renormalization of various quantities, the masses and the coupling constant. The physical
meson mass µ is in general not equal to the bare meson mass µ0, the physical nucleon mass m is in general not
equal to the bare nucleon mass m0. We also have to renormalize the coupling constant. If this were a realistic
theory—which it ain—t-the physical coupling constant g would in general not be equal to the bare coupling
constant g0. If this were some real interaction like electrodynamics or the weak interaction, the coupling constant
in those little tables circulated by the Particle Data Group would be the coupling constant as defined by some
standard experiment set up by an IUPAP committee,4 say the Coulomb interaction between two distantly
separated charges, or perhaps pion–nucleon scattering at a certain point for the strong interactions, or beta decay
for the weak interactions. There’s no reason why the answer to that standard experiment should be g0. The
answer to the standard experiment might be something else. So whatever it is that appears in the tables as a
result of experimental measurement as the physical coupling constant is certainly not equal to g0, unless the
experiment has been incredibly cunningly chosen. Of course, these quantities are not entirely unrelated.

For example, if the theory is free, when g0 = 0, Z2 and Z3 are equal to one. In the interacting theory, they
might have had corrections of order g0 (in fact, as we will see, the corrections are of order g20). But they certainly
reduce to one as g0 goes to zero. Likewise m2 is m20 plus corrections of order g0. We surely want any sensible
definition of the coupling constant as physically defined to reduce to the coefficient in the Lagrangian for very weak
coupling, so we will normally accept from that IUPAP committee as a sensible definition only one such that g = g0
with perhaps higher order corrections.

Now in principle we could solve the theory in the following way. We could do perturbation theory, which would
give us the Green’s functions for unrenormalized ψ’s and ϕ’s, up to some finite order, as a power series
expansion in g0 with m0 and µ0 held fixed. We could then determine the physical masses as functions of m0 and µ0
and g0, all the wave function renormalization constants as functions of m0 and µ0 and g0, and the result of that
standard experiment defined by that IUPAP committee as functions of these parameters. We could then adjust the
values of these bare parameters to give the right answer as multiplied by the Z’s and compute their scattering
matrix elements. That’s possible, but it’s also an enormous pain in the neck. You are computing the wrong
Green’s functions in terms of the power series in the wrong coupling constant with the wrong masses held fixed.
For practical purposes we’d like to do an expansion in a realistic theory like quantum electrodynamics, not in the
bare charge, but in the actual charge that is measured, with the physical masses held fixed and in terms of the
renormalized Green’s functions, which are the things we’re after at the end. This is purely a practical question of
convenience. It has nothing to do with one of principle.

We can avoid the wrong expansion by rewriting the Lagrangian (14.40) in terms of these renormalized
quantities with things left over.

There is a lot of leftover stuff, because the Lagrangian (14.40) written in terms of the bare quantities isn’t equal to
this Lagrangian (14.46) written in terms of the physical quantities, without the leftover stuff. We take all the leftover
parts and sum them up into counterterms:

The expression LCT looks pretty horrible:

There’ll be some coefficient linear in ϕ′ that will come from shifting the quadratic term because (13.52) ϕ is
proportional to ϕ′ plus a constant. A, B, C, D, E and F and the new value of the constant are given simply by
requiring the two Lagrangians to be equal, although these formulas will be of absolutely no interest to us. If you
work things out, A is −Z31/2 µ02 á0|ϕ|0ñ, B is Z3 − 1, C is −µ2 + Z3 µ20, and so on.

Now the general strategy is this. Please notice that all of these coefficients are going to be things at least of
order g, and the coefficient F is going to be at least of order g2, because of how we’ve defined things. And
therefore our strategy will be to treat these as we treated the counterterms before, that is to say, to compute
everything treating the set {A, . . . , F} as free parameters and then to fix them by imposing our renormalization
conditions, the conditions that define the renormalized mass and renormalized scale of the fields. Notice we have
just enough conditions to do this. If we ignore the (constant) and (constant)′ we have six counterterms {A, . . . , F}
and we have six renormalization conditions:

Renormalization conditions for Model 3

1.á0|ϕ′|0ñ = 0 fixes A

2.áq|ϕ′(0)|0ñ = 1 fixes B
 3.The physical meson mass, µ, fixes C

4.áp|ψ′(0)|0ñ = 1 fixes D

5.The physical nucleon mass, m, fixes E

6.The definition of g fixes F

(The condition á0|ψ′|0ñ = 0 is automatic, so we don’t have to impose it.)

So we systematically go out order by order in perturbation theory treating the set of constants {A, . . . , F} as
free parameters. To any fixed order in perturbation theory we impose the values of the set of constants by
asserting our six renormalization conditions. We then determine {A, . . . , F} self-consistently as a power series in
the physical coupling constant, and we have achieved our desired end. We have turned a stupid, although in
principle valid, perturbation series for the wrong Green’s functions in terms of the wrong coupling constant with the
wrong masses held fixed, into a systematic perturbation expansion for the right Green’s functions in terms of the
right coupling constant with the right masses held fixed.

Now we’re almost ready to begin doing computations to see how this formalism works out. (“Almost ready”
means we’ll get to it next time.) But before we do that, there are two questions which we have to consider. The first
problem is this. Considering LCT (14.47) as part of the interaction, we have introduced derivative interactions (the
terms proportional to B and D), and our whole formalism is set up for non-derivative interactions. So there is an
awkward but necessary technical point we have to investigate: What is the effect of a derivative interaction? That
is, what does it do to the Feynman rules? We’ve got vertices in the theory corresponding to terms in the
Lagrangian that have derivatives in them, and that will give us all sorts of problems. It changes the definition of the
π’s (they’re no longer ∂0ϕ’s), and everything gets horribly messed up. So we’ve got to worry about those
derivative interactions. That’s a trivial problem which I’ll take care of this lecture.

The second problem confronting us is that our renormalization conditions are not well set up to be
systematically applied in perturbation theory: they’re not phrased in terms of Green’s functions. The second
condition doesn’t have to do with Green’s functions. The fifth condition, that m is the physical mass of the particle,
is not phrased in terms of Green’s functions. Our whole Feynman apparatus is set up for computing Green’s
functions. If we really want to build a smoothly running, well oiled machine where we can just grind out
calculations without any thought, or better yet, write a computer program that will do it for us, we would like to
phrase the renormalization conditions in terms of properties of certain Green’s functions. We’d like to say that the
sum of all Feynman graphs of a certain kind vanish, or equals one, or something. That’s equivalent to the equation
that gives us the scale of the field. We haven’t got that yet. So these are the two tasks before us that we have to
complete before we can automate this scheme, and be able to compute without thought. We begin with the first
task, derivative interactions.

14.4Guessing the Feynman rules for a derivative interaction

The general formalism for derivative interactions is an incredible combinatoric mess. Things really get awful. The
coupling constant enters into the definition of the canonical momentum:

The interaction Hamiltonian has the canonical momentum in it, and therefore you have all sorts of problems about
what the time ordered product of a string of interaction Hamiltonians and a string of fields means, because they no
longer commute at equal times:

In particular, it is no longer true that H I = −LI:

Things are just too horrible to contemplate, at least at this stage of our development. After we’ve hardened our
heads, karate fashion, by banging them on some difficult problems, we will return to the question of derivative
interactions and straighten everything out for them. But I really don’t want to get into that whole complicated
subject to handle such a simple derivative interaction as this sort, like the term proportional to D. After all, it’s not
really much of an interaction, it’s just a term quadratic in the fields. Therefore what I will do is guess the Feynman
rules appropriate to this derivative interaction, and then try and show you that my guess is okay by doing some
simple examples. When I’m done, it will be obvious how to generalize the results to this theory, and we will see
that the generalization gives the desired results.

The first theory I will look at will be the simplest of all theories, a free scalar field:

This doesn’t have a derivative interaction; indeed, it doesn’t have any interaction at all. But we can fake matters so
it looks like it has a derivative interaction by introducing a new variable ϕ′:

Of course ϕ′ is not a renormalized field in the standard sense here; perhaps I should not call this constant Z3. The
field ϕ is already perfectly adequately normalized. Nevertheless to remind you of the problem it is connected with,
I will call this quantity Z3. It’s some arbitrary constant, maybe one plus a squared coupling constant. If we rewrite
this theory in terms of ϕ′, we obtain

For a simple example, I’ll take (Z3 − 1) to be equal to g2, a coupling constant, and I will call the term in (Z3 − 1) an
interaction. I’d like to get a Feynman rule for that vertex. We had a similar term in the free Lagrangian of Model 3,
a counterterm which I wrote as bϕ 2. But this term was not in the interaction Lagrangian, and in any case it did not
have a derivative. When we were looking at Model 3, we were still in a state of primal innocence, unaware of wave
function renormalization; we only considered mass renormalization. In Model 3, this term had the Feynman graph
and Feynman rule (see p. 216, 3.(d))

Here I have an interaction

which I’ll indicate by this:

Figure 14.1: Meson derivative interaction

It’s the only interaction in the theory. I give the two lines momenta q and q′, both oriented inward. The interaction
has two parts, one coming from the µ2 term, and one coming from the (∂µϕ)2 term. We worked out the result of the
µ2 before, as written above. By analogy we get a Feynman rule that looks like this:

That’s how we treated our old mass counterterm, and it’s unquestionably what descends from the µ2 part of the
interaction.

But what do we get from the second term, (∂µϕ′)2? God only knows. Though we are not divine, we are allowed
to guess. Since we see µ2 here, I will guess q2 belongs where we now have (···). This is just a sheer, blind guess:

No derivatives, µ2; two derivatives, q2: total guesswork. We are guessing that a derivative ∂µ in the interaction
leads to a power of momentum qµ in the Feynman rules, to within factors of ±i:

Now I’ll check my guess by computing ′(2), the two-particle Green’s function, by summing up the
perturbation expansion in this object. Since I already know the exact form of ′(2), there should be no problem
in seeing whether the guess is right in this simple case.

First, the exact answer, from (14.14) and (14.15):

That’s the Green’s function in the original free theory multiplied by the inverse of Z3. (I omit the iϵ out of sheer
laziness.) This is exact. It doesn’t depend on my guess.

Do we get the same thing by summing up diagrams? The series of diagrams is very simple; see Figure 14.2.
There’s a zeroth order diagram, the first correction, second-order correction, three of ’em on a line, etc. Those are
the total collection of all Feynman diagrams with two external lines, with the interaction indicated by an × on the
line. Now what do we have? First I get (2π)4δ(4)(q + q′), they’re all energy conserving. That’s just my convention
that I’ll keep the δ(4)(q) in my Green’s functions. Let’s look in detail at the first two graphs:

Figure 14.2: Perturbation series for the free scalar field ′(2)

The first diagram in the series is just the free propagator times (2π)4 times the delta function. No question there;
that’s the same diagram that emerged in the free theory. That will be a common factor. It’s in all of these things if
we count from the left because all have a propagator on the left: our Green’s functions have propagators on the
external legs.

What happens when I consider one vertex and one propagator? Well, we notice something rather peculiar.
The factor of (2π)4 from the interaction cancels the (2π)4 in the denominator of the integration. The delta function
from the interaction term is canceled by the integration; the internal momenta are the same. The i from the
interaction term and the i in the propagator multiply to make a factor of (−1). The interaction term has a factor of
(q21 − µ2) that exactly cancels the same factor in the propagator. The only factors not canceled are (−1)(Z3 − 1) =
(1 − Z3). The result of adding both one more interaction and one more propagator is simply to add a factor of 1 −
Z3. The net result for the first two graphs is

Likewise the result of adding two more interactions and two more propagators is to add a factor of (1 − Z3)2. We
know how to sum a geometric series:

in agreement with the exact result. As usual we take physicist license and don’t worry about questions of
convergence. The agreement leads us to believe our guess (14.57) about the Feynman rule for the derivative
coupling is correct.

The guess also works for an interacting theory as I will now demonstrate. To avoid inessential complications
of notation, I will use an interacting theory which has only a single field,

I’ve included a factor of 4! to be in line with conventions. It’s a nice interacting theory that’s only got one field in it.
Once again I define

I also choose that ϕ has a vacuum expectation value of zero, because the theory is symmetric under ϕ → −ϕ. I
won’t bother to make the redefinitions that are responsible for mass and coupling constant renormalization, which
in principle we’d have to do in this theory, just as in our simple model. I’ll write exactly the same Lagrangian as

I just want to make the same comparison in this theory that I did in my simple model, and find out what is the effect
of making this sort of interaction, to see if the same guess is right.

Suppose I have a graph for (n) in this model. For example, this is a graph for the unrenormalized (4),

that is to say, computed using the first form of this Lagrangian:

I don’t even care what the Feynman rules are for this graph. This is the only kind of graph with just two vertices
that can possibly emerge if I use the first form of the Lagrangian.

Figure 14.3: Graph for (4) in gϕ 4 theory

I will introduce a little topological notation. Such a graph has n external lines, I internal lines and V vertices; n,
I and V are numbers associated with a particular graph. What are the connections between these quantities? They
are not three free parameters. They are connected by the law of conservation of ends of lines. Every external line
has one end that ends on a vertex. Every internal line has two ends that end on a vertex. Every vertex has four
lines ending on it (because the theory has a term in ϕ 4). Then

A slight variant of this formula will turn out to be useful:

Now let’s turn to the second form of the Lagrangian. From any one of these graphs can be generated an
infinite family of graphs which differ simply by any arbitrary number of my new interaction, which I represent just as
before, by a ×, on any one of these six lines. In Figure 14.4, I’ve added seven. I have to sum up this infinite family
in order to find out the graph that corresponds to this, which, if my guess was right, should be the graph for ′(4).

Figure 14.4: Graph for ′(4) in gϕ 4 theory

The family of graphs for ′(n) differs from the corresponding graphs for (n) in two ways. First, I have a
different coefficient of the four-particle vertex. For (n), each four-particle vertex is multiplied by g0; for
′(n), it’s multiplied by Z23g0.
So the net factor from the vertices is Z32V. The comparison is really graph by graph,
but we’ll shortly see the factors that depend on anything except n disappear. Second, every line, internal or
external, has all of these crosses sitting on it like crows on a telephone wire, any number, and we have to sum that
up. I can put any number on each propagator. In Figure 14.4, I have six geometric series, which are all
independent and which all sum up, because there are six lines internal or external in the diagram. I sum up all of
those things, all the possibilities, 17 insertions on the first line, none on the fourth, 42 on the second, etc. From
each I get an independent geometric series. Fortunately we did that summation earlier, and we discovered the
summation just multiplied each propagator by Z3−1. Therefore we have Z3−1 from every line, since we do it on both
internal and external lines, because the Green’s function has propagators on the external lines. That gives another
factor of Z3−(n+I). All in all, we have

using our topological statement, 2V − I = n. This is of course exactly the result we would’ve obtained by
substitution.

This argument obviously carries over into the general case. If I make this operation on a much more
complicated Lagrangian, I’ll induce an extra interaction proportional to Z3 − 1. In each vertex I’ll stick a power of
Z31/2 associated with the number of ϕ fields coming into that vertex. When I have an internal line, the result of
summing all the crows on the telephone wire will give me a factor of Z3−1, which will precisely cancel the factors at
the two vertices, between which the internal line goes. If I have an external line I still get a factor of Z3−1 but it only
goes into one vertex. Therefore it’s only half canceled, and I’m left with an overall factor of Z3−1, so for the n
vertices in ′(n), the factor is Z3−n/2. The argument is trivially generalized to any theory. Once you see how this
one works, you should be able to see how it works in any case. So we have indeed, without having to go through
all the work it would take us to develop the general theory of derivative interactions, successfully guessed the right
form (14.57) for this particular derivative interaction, at least, to within a sign. Next time we will deal with the
second problem, expressing things in terms of Green’s functions.

1[Eds.]Klaus Hepp, “On the connection between the LSZ and Wightman quantum field theory”, Comm. Math.
Phys. 1 (1965) 95–111.
2[Eds.] Harry E. Lehmann, Kurt Symanzik and Wolfhart Zimmerman, “Zur Formulierung quantisierter
Feldtheorien”, (Toward the formulation of quantized field theories) Nuovo Cim. ser. 10, 1 (1955) 205–225.
3[Eds.]“Duck soup” is idiomatic American English for “a task easily accomplished” (and also the title of a Marx
Brothers movie (1933)), synonymous with “a piece of cake” or “a snap”.
4[Eds.] The International Union of Pure and Applied Physics.

Problems 8

8.1 One consequence of our new formulation of scattering theory is that it doesn’t matter much what local field you
assign to a particle: any field that has a properly normalized vacuum to one-particle matrix element will give the
right S matrix element. (See the discussion following (14.35).)

Consider the theory of a free scalar field,

Let us define a new field, A, by

In terms of A, the Lagrangian becomes

If you had been presented with this Lagrangian, and didn’t know its origin, you would probably think it described a
highly nontrivial theory, with complicated non-zero scattering amplitudes. Of course, you do know its origin, and
thus you know that it must predict vanishing scattering. Verify this by actually summing up all the graphs that
contribute to meson–meson elastic scattering in the A-field formulation, to lowest nontrivial order in g, i.e., g2, and
showing that the sum vanishes.

Comments:

(1) Our general theory does not tell us that the A field Green’s functions are the same as the ϕ field Green’s
functions, so the amplitudes may not vanish if the external momenta are not on the mass shell.

(2) To the order in which we are working, we can completely ignore renormalization counterterms.

(3) This is a theory with derivative interactions. As discussed in class, this leads to potential problems: the
interaction Lagrangian is not the same as minus the interaction Hamiltonian, and we can’t pull time derivatives
through the time-ordering symbol in Dyson’s formula. Much later in this course, we shall study such theories using
the methods of functional integration, and discover that (to this order in perturbation theory) these problems
cancel. Take this on trust here; use the naive Feynman rules as explained in class. (That is to say, treat the theory
as if the interaction Hamiltonian were minus the interaction Lagrangian, and as if every derivative ∂µ became a
factor of −ipµ for an incoming momentum, and ipµ for an outgoing one.)

(4) If you have an interaction proportional to A4, there are 4! different ways of choosing which fields annihilate
and which create which mesons. If you don’t keep proper track of these (and similar) factors, you’ll never get the
right answer.

(5) A graph of order g2 may contain either one vertex proportional to g2 or two vertices each proportional to g.

(6) To get you started, here are the graphs you’ll have to study (with various momenta on the external lines):

(a)

(b)

(1997a 8.1)

Solutions 8

8.1 In terms of the field A, the Lagrangian may be written

where

Following the advice, we take H I = −H I. There is also a −i in Dyson’s formula. Corresponding to the vertex

we have this graph, and, using the naive rule about derivatives–∂µA → −ik µA for an incoming particle–this
Feynman rule:

The 2! arises from the symmetry of the two identical factors ∂µA, and the 3! from the symmetry of the three
identical factors A in the second term. The rule can be simplified by using the identity

Moreover, in this case, k1 + k2 + k3 = 0. Consequently the Feynman rule becomes

Consider a diagram like (6)(a) in the statement of the problem, for example,
This diagram makes a contribution iA12,34 to the total 2 → 2 amplitude equal to

We are only looking for contributions on the mass shell, when ki2 = µ2, in which case

There are two other diagrams of the same form, obtained by permutations, making contributions iA13,24 and
iA14,23. Adding these all together gives all contributions from diagrams of the form (6)(a),

This may not look very symmetric, but it is, because k1 + k2 + k3 + k4 = 0, by virtue of the conventions about inward
pointing momenta. That means

Making the equivalent substitution for the other two contributions gives, after a little algebra,

so that the total contribution from all of the diagrams of the form (6)(a) is

If we evaluate this on the mass shell, then ki2 = µ2, and so

Now for the graphs of the form (6)(b). Corresponding to the vertex

we have this graph, and this Feynman rule:

Once again, the factorials arise from symmetry. There are two identical terms A and two identical terms ∂µA, so
we have a factor of 2! from each. The 4! from the symmetry of the four identical factors A in the second term. The
sum of the (42) products of the two different momenta is easily seen to satisfy the identity

because the sum of the 4-momenta is zero. The contribution iA(b) from (6)(b) is then

On the mass shell, this is ig2µ2. The total amplitude is thus

as required. There is no scattering on the mass shell. In fact, there is no scattering off the mass shell, because the
contributions (S8.13) and (S8.17) cancel, whatever the values of the four 4-momenta.

15
Renormalization I. Determination of counterterms
At the end of the last lecture, we found ourselves with six counterterms in our Lagrangian, (14.47), and six
renormalization conditions (box, p. 302) for fixing them: that the expectation value of the renormalized meson field
in the vacuum state be zero, a wave function renormalization condition for the meson field, a wave function
renormalization condition for the nucleon field, two conditions that the mass parameters appearing in our
Lagrangian be the physical masses, and one condition to fix the coupling constant. These conditions determine
the six counterterms order by order in perturbation theory.

15.1The perturbative determination of A

Our perturbation theory is set up for the computation of Green’s functions, and so we would like to phrase our six
renormalization conditions in terms of Green’s functions. At the moment only one of them, involving A, is
immediately phrased in terms of Green’s functions:

The physical vacuum expectation value of the renormalized meson field should be zero. This is a Green’s
function: it can be thought of as the vacuum expectation value of the time-ordered product of one field.
Graphically, this condition is simply

Figure 15.1: The renormalization condition for A

This makes it very easy to determine iteratively the renormalization counterterm A order by order in perturbation
theory. In the Lagrangian there’s a bunch of stuff, then there is Aϕ′ plus a bunch of other stuff:

Let’s imagine I have A as a power series expansion in g. I will write this as

Graphically:

where the (n) over the vertex means you are taking the term proportional to gn.

Now let us suppose that we have computed everything up to order n – 1, all Feynman graphs for all Green’s
functions and, by some method which I have not yet explained, all counterterms; not only A, but {B, C, . . . , F} up to
order n − 1. I will show how that enables us by a computation to determine An.

The argument is very simple. We have our renormalization condition, Figure 15.1, which states that for all
values of g, the blob equals zero. We will now compute this blob to O(gn). There will be two terms:

The first term will be all sorts of complicated Feynman diagrams that may well involve, as internal parts, all of the
other counterterms {B, , F} in lower order. Those terms are known in principle, because they only involve at their
vertices counterterms of lower order. By assumption we know these counterterms: we are analytically muscular,
and we can compute any Feynman diagram. Then there is one unknown object that contributes, and it contributes
in only one way: the nth order of A. The nth order of A never appears as an internal part of some diagram of more
complicated structure, because if it did, that diagram would be of one order higher than n − 1. The whole thing
sums to zero. Therefore this relation (15.5) fixes the nth order of A. So this is how we could iteratively determine
the counterterm A and put up a table of its values: the first, the second, the third orders and so on, if we can do the
same sort of trick for the other counterterms.

We’ll later see that exactly the same thing will happen for all the other counterterms. We will phrase our
renormalization conditions in such a way that a certain sum of graphs, a Green’s function or an object defined in
terms of Green’s functions, is equal to zero. We will carefully choose a sum so that if we compute it to nth order,
the nth order counterterm will come in only in a simple form like this, plus known stuff, and then we’ll have a
systematic iterative procedure for computing the counterterms.

In fact, we hardly need A. There is a special feature for this rather simple counterterm, A, that means we don’t
even have to keep track of this table. This isn’t true for the other counterterms. Suppose I consider any graph of
the following structure:

Figure 15.2: A tadpole diagram

These types of graphs are sometimes called tadpole diagrams for obvious reasons.1 Here I’ve got absolutely
anything inside this left blob, and I have a large number of lines or a small number of lines, it matters not, coming
out. Then I have a line connecting the first blob to a second blob on the right, containing anything else—but with
no external lines. That is to say, the graph has a topological structure of two parts connected by a single internal
line such that, if I cut that line, the graph separates into two discrete pieces. Now if I sum over all the possible
things I can put in for anything else, to a given order, without changing this part on the left, then I obtain the
relation

Summing up “anything else” gives the shaded blob, which, by the renormalization condition, is zero. So the net
contribution of these tadpole diagrams is zero. Since it is only in graphs of the structure shown in Figure 15.2 that
this counterterm appears, in fact we need not worry about the counterterm or about the renormalization condition.
They’re going to cancel out. All the tadpole graphs sum up to zero and so you can ignore them, just as you can
ignore the graphs with disconnected vacuum components.

This demonstration was pretty trivial. That was a good thing because I was able to show the iterative
establishment of counterterms in a simple context. I now turn to something much more complicated, the phrasing
of the wave function renormalization and mass renormalization conditions in terms of Green’s functions. Despite
the added complications, we will be able to reach the end in a fairly short time.

15.2The Källén-Lehmann spectral representation

I will begin by making a general study, with hardly any assumptions, of the two-point function, á0|ϕ′(x)ϕ′(y)|0ñ. I will
derive some properties of this object. From this object of course I can reconstruct the Green’s function ′(2) just by
multiplying by theta functions. We use systematically the identity, true for any state |nñ,

and in the case of a one-particle state |pñ,

by our normalization condition, (14.42).

Now I will analyze this object by putting in a complete set of intermediate states and eliminating both the x and
the y dependence by using (15.7):
The vacuum state gives no contribution because ϕ′ has a vanishing vacuum expectation value. From the one-
particle states, I just get ones from the matrix elements, and obtain e−ip⋅(x−y). The sum over the multiparticle
intermediate states—of course it’s a sum and an integral—has a prime to indicate we are excluding the vacuum
and one-particle states. The first term is an object we have discussed before, (3.38), in connection with quantizing
a free field, with p0 = ωp. There we called it Δ+(x − y; µ2), where µ2 is the physical mass of the meson:

This big sum is going to give us some Lorentz invariant function of pn which vanishes unless p0 is on the upper
hyperboloid; by assumption, we only have positive energy states in our theory. Therefore I will write it in the
following way:

The (2π)3 is unfortunate, but I’ll run into a convention clash with standard notation if I put a (2π)4 there. (Our
convention, violated here, is that every momentum integral is accompanied by a 2π in the denominator.) The theta
function θ(q0) ensures that things vanish, except on the upper hyperboloid. The function σ(q2) is defined by this:

If I stick this expression into the previous equation (15.11) I obviously get equation (15.10), just by doing the
integral over q. We know that σ is a function of q2, rather than q, because of Lorentz invariance: the sum over
intermediate states is Lorentz invariant. Alternatively, the left-hand side of (15.11) must be a function only of (x −
y)2. So its Fourier transform should be a function only of q2.

We know other general features about σ(q2). In perturbation theory we would expect that the lightest
multiparticle states that can be made by a ϕ′ field hitting the vacuum are either two mesons or a
nucleon–antinucleon pair. So we would expect in perturbation theory that σ(q2) equals zero, for q2 less than the
minimum of 4m2 if the nucleon–antinucleon pair is lighter, or 4µ2 if the two meson state is lighter:

Of course this is just a perturbation statement. In the real theory, there might be bound states appearing which lie
below either the meson–meson or nucleon–antinucleon threshold. Say that the lightest particle in the theory is the
meson. Then in the real theory, in any event

The value of η depends on the energy of the bound state. If the bound state sinks below the one-meson state,
then we call the bound state the one-meson state, because by definition the one-meson state is the lightest state
with the quantum numbers of the meson. If they’re right on top of each other, then we were making the wrong
assumption about the spectrum: there are two one-meson states, and we have to rethink the whole thing.
Additionally, because σ is defined in (15.12) as an integral of squares times positive terms, we also know that
σ(q2) is always greater than or equal to zero:

These two facts, (15.14) and (15.15), will be very important to us in our subsequent development.

We can rewrite the expression (15.10) as follows:

Here a2 is a new dummy variable. Because of the delta function, this is just another way of writing σ(q2)θ(q0). The
advantage is that I can now do the q integral, because what I have here is, for each fixed value of a, the
expression that gives me Δ+(x − y; a2) for a free field of mass a2. We have the definition (3.38),

But we also have the relativistic measure (1.55):

so we can write an alternative definition,

Thus (15.16) can be written

That is to say, we’ve written the exact vacuum expectation value of the product of two fields as a superposition of
free field vacuum expectation values, integrated over the mass spectrum of the theory. This is sometimes called
the spectral representation for that reason. It is also called the Källén-Lehmann spectral representation.2
You may see it in the literature written in the form

where ρ(a2) is of course equal to δ(µ2 − a2) + σ(a2). We won’t use this form much.

I will now use the spectral representation, first, to get a representation of the commutator that will give us an
interesting inequality, and second, to get a representation of the Green’s function, the time-ordered product. Since
we have everything represented as a linear superposition of free field quantities, we can simply go through all of
our old free field manipulations appropriately superimposing them. Thus for example (see (3.42))

We can now compute the vacuum expectation value of the equal time commutator. This is amusing because we
know what the equal time commutator is, in terms of Z3, since we know ϕ′ in terms of canonical fields and Z3: ϕ′ is
Z3–1/2ϕ s (13.52), where the shifted field ϕ s = ϕ − á0|ϕ|0ñ. Since ϕ s differs from ϕ only by a subtracted c-number
(13.49), they have the same commutators, and so, from (3.61)

On the other hand we can get exactly the same thing from the spectral representation. We know from (3.59)
and (3.61) that

Differentiating (15.22) with respect to y0 and evaluating at equal times will give us

Comparing these two expressions, (15.25) and (15.23), we find Lehmann’s sum rule:

Or, since σ(a2) is guaranteed to be non-negative,

Most likely, we should expect Z3–1 to be greater than 1. It will only be equal to 1 if σ(a2) vanishes, which would be
a pretty trivial field theory.3 Equivalently, we can say

It is sometimes alleged that this statement (15.28) has a trivial explanation. After all, Z31/2 is defined to equal
ák|ϕ(0)|0ñ where ϕ is the unrenormalized field (see (13.51)). I will now tell you an argument that is a lie, but at least
it will help you remember which way the sign goes between Z3 and 1. People say, “Look, we know ϕ(0) hitting the
vacuum makes a single bare particle and therefore ák|ϕ(0)|0ñ is the amplitude for making a physical particle. So
it’s the inner product between a physical particle and a bare particle, which is less than or equal to one, like all
inner products between appropriately normalized states. Therefore Z3 is less than or equal to 1.” This argument is
a lie, of course, because ϕ(0) is scaled so that it has amplitude 1 for making a bare particle when applied to the
bare vacuum, and here we are applying it to the physical vacuum. So the argument is completely useless.
Nevertheless it’ll help you remember the sign. By the way, this stuff is treated at enormous length in all the
standard texts, including Bjorken and Drell.

15.3The renormalized meson propagator ′

However amusing we may have found our work thus far, we haven’t gotten very close to expressing our
renormalization conditions in terms of perturbation theory objects, i.e., in terms of Green’s functions. So we will go
on and compute the renormalized Green’s functions. Of course once we know the vacuum expectation value of
the unordered product of a pair of fields, we can obtain the two-particle Green’s function ′(2)(p, p′) by a linear
sequence of operations: permuting the arguments, and multiplying by theta functions, Fourier transforming etc. So
we can just write down the answer. It’s convenient to express things in terms of objects with the delta function
factored out. So I’ll write the expression for ′(2)(p, p′) as

where the entity ′(p2) is sometimes called the renormalized propagator. (Sometimes the prime indicates
renormalized fields.) So I take the ordinary Feynman propagator, and now we’ll put all sorts of corrections on it, to
get what really happens when one meson goes into a blob and one meson comes out. ′(p2) is, by the spectral
representation, a linear superposition of free propagators with the same weighting function as with Δ, (15.22). That
is to say,

This spectral representation of ′(p2) tells us something very interesting about the analytic properties of ′(p2)
considered as a function of complex p. As you see, for example, for all p2’s not on the positive real axis, this
integral defines an analytic function of p2. If p2 is not on the real axis, the denominator never vanishes, so the
function is well-defined and its derivative is also well-defined. Thus if I were to draw the complex p2 plane,

Figure 15.3: The analytic properties of ′(p2) in the complex p2 plane

′(p2) would be an analytic function in that plane, except for a pole at µ2 and the branch cut, a line of singularities
beginning from µ2 + η, where σ(p2) begins to be non-zero, extending out presumably to infinity. The actual
physical value of ′ for real p is of course totally unambiguous. Along the branch cut, though, we have to say
which side of the cut we’re on. Feynman’s iϵ prescription tells us µ2 = p2 + iϵ, which means we are above the cut
in this analytic continuation of ′. The original ′, defined only for real p2, is obtained by taking the analytic
function onto the cut from above. Those of you who have studied the analytic properties of partial wave amplitudes
in non-relativistic scattering theory will not find this analytic structure surprising.
 We could get into troubles if the σ(a2) integral doesn’t converge. That means the sum over intermediate
states doesn’t converge. The formula (15.12) is great if σ(a2) has any reasonable behavior. The swindle I’ve put
on you is that if σ(a2) grows too rapidly at infinity, say like a power of a2, then if you look at this function in position
space, it’s a well-defined distribution, but it’s not true that a theta function times a distribution is necessarily a
distribution. I will assume that in our case the time-ordered product is defined. That is the sort of thing purists have
to worry about. Maybe we’ll become purists when we get a deeper understanding of field theory, and we may go
back and worry about that.

In principle, if we ever reach any trouble we will be quite willing to act like slobs. If we cannot justify our
intermediate stages with what we’ve got, we will brutally truncate our theory by throwing away the high momentum
modes, therefore making σ(a2) vanish beyond a certain point, a cutoff, and guaranteeing the convergence of
everything. We’ll just cut them out of the theory, bam! We will then have a sick, nonsensical theory. That’s not real
physics, but we’ll just go ahead. When we finally get the S-matrix elements, if they have nice smooth limits as the
cutoff goes away, we’re happy. We will have reached a satisfactory result even if our intermediate stages are
garbage. If they don’t, there’s no point worrying about mathematical rigor, because nothing we can do will make
sense out of it. That’s our general attitude whenever we run into trouble because of the high energy behavior of
integrals. People untrained as carpenters who nevertheless build houses are called “wood butchers”. The attitude
I’m describing is that of a “physics butcher”.

We actually know a little bit more about −i ′(p2). It has what is called a Schwarz reflection principle in the
theory of functions of a complex variable. It’s easy to see from the spectral formula (15.30)

Once we’ve multiplied by −i to get rid of the i in the numerator, conjugating p2 is the same as conjugating the
function, in the domain of analyticity. If you’re above the cut, this is the value below the cut. The discontinuity over
the cut is therefore connected to the imaginary part of ′. By a formula we used previously in a homework
problem,4 the imaginary part is given by

That’s the difference between the value above the cut and the value below the cut.

The mass and wave function renormalization conditions are embedded in a statement about the Green’s
functions:

This equation contains our two renormalization conditions: for mass, that there is a pole at µ2, and for the wave
function, that the residue of that pole is i. If the field had been normalized differently, the residue of the pole would
be 17i or i or something. This gives us, in principal, a way of determining the mass of the particle and the
normalization of the field in terms of the properties of ′, which is defined in terms of a Green’s function.

This condition will have to be massaged a bit to put it into the best form for doing the computation we want to
do. To that end I will define a special kind of Green’s function, called a one-particle irreducible Green’s
function, denoted by 1PI, for “one-particle irreducible”, and indicated by a blob like this, with however many
external lines coming out of it:

Figure 15.4: One-particle irreducible diagram

This is the sum of all connected graphs that cannot be disconnected by cutting a single (one particle) internal line.
By convention, when we evaluate 1PI diagrams, we do not include the energy-momentum conserving delta
function, nor external line propagators. These conventions will turn out to simplify our algebra with these things.

To give an example of what is and what is not a 1PI graph, take two diagrams from nucleon–antinucleon into
nucleon–antinucleon scattering in Model 3, as shown in Figure 15.5. Diagram (a) is not 1PI because cutting (or
removing) the internal line divides it into two separate parts. On the other hand, Diagram (b) is 1PI, because there
is no way I can split it into two parts by cutting any one internal line: it still remains connected. I have to break at
least two internal lines in Diagram (b) to make it fall apart.

Figure 15.5: The difference between 1PI and not 1PI.

We can now define an object to express in simple terms our mass and wave function renormalization
conditions. Looking at the 1PI 2 meson function, we define i ′(p2), a function of p2 only, the sum over all 1PI
diagrams:

For reasons that will soon become clear, ′(p2) is called the meson self-energy operator.

Now let’s look at the renormalized two-particle Green’s function, ′, and write it in terms of ′. This is a lovely
process. We make drawings which are easy to manipulate, and then they turn into equations which are also easy
to manipulate. What is the perturbation series for this object? Well, first we could have a single unadorned line, in
zeroth order. Then we could have a one-particle irreducible diagram just sitting there, with the two external lines to
give the propagators that we’ve left off by convention. After that, we could have a diagram that’s actually one-
particle reducible; that is to say, which I can cut someplace and make fall into two parts. If there’s only one place
where I can cut it, then on the left of the cut there must be something one-particle irreducible, and likewise on the
right. Here I explicitly display the one line I can cut to make it fall into two parts; it’s got to be cut somewhere
between there as we go along. And then everything else by definition must be one-particle irreducible because
there’s only one place where I can cut it. If there are two places where I can cut it, . . . , well, you see where we’re
going:

Now what does this say in equations? Factoring out the overall delta functions that occur everywhere on the
left-hand side, and writing5

as in (10.29), we have

The geometric series sums up to a propagator with a mass term of µ2 + ′(p2). This is why ′(p2) is called the self-
energy operator, or sometimes the self-energy function, or the self-mass function, because it adds to the mass.

We can now phrase our renormalization conditions in terms of ′. ′ is an analytic function near p2 = µ2
except for a pole, and therefore ′–1 is an analytic function near p2 = µ2, period, since the inverse of a pole is a
zero, which does not affect analyticity. Therefore ′ has a power series expansion in terms of p2 at p2 = µ2:

The value of ′(µ2) must be zero. If it were not, from (15.36), we would not have a pole in ′(p2) for p2 = µ2. Thus
we have
from the statement that the physical mass of the meson is µ2. At the pole p2 = µ2, the residue of ′ must be i.
Expanding the denominator,

because ′(µ2) = 0. Therefore the first derivative of ′ must be zero. If it were not zero, the residue would not be i,
but instead i times the reciprocal of (1 plus the first derivative of ′). Consequently we must have

These two statements about ′, that it and its derivative vanish at p2 = µ2, are precisely equivalent to our two
renormalization conditions, that ′ has a pole at µ2, and that its residue at this pole equals i.

Now the nice thing about these conditions is that they enable us to determine iteratively the mass and wave
function renormalization constants in exactly the same way as we outlined for the A counterterm. Let’s focus on
the B and C counterterms. From (14.47),

These two terms lead together to an interaction which I’ll indicate diagrammatically by a single cross:

The interaction is determined in terms of both coefficients. There’s a B part which, because of the derivative
coupling, gives us iBp2, and the C part which gives us −iC, as demonstrated at the end of last lecture. As before,
we break this up into a power series in the coupling constant:

In diagrams, we indicate each term of order gn by (n):

We assume that we know everything to order n − 1, and are about to compute things to order n. We have

I assume we can compute the lower orders in perturbation theory (the “known stuff”); to determine Bn and C n we
impose the two constraints (15.38) and (15.40):

and

All this goes through, mutatis mutandis, for the nucleon field, since our nucleon is not really that different from
the meson, despite the name we’ve given it. It’s just another scalar field. I won’t bother to write down the whole
spectral representation for the nucleon, but for the self-energy term we can write, analogous to the renormalized
meson propagator,

and the appropriate 1PI graph,

The corresponding counterterms are, from (14.47),

The values of D and E, the counterterms associated with nucleon mass and wave function renormalization,
respectively, are fixed by the two conditions

It’s just the same thing written over again.

We found a very nice result when we were considering the A counterterm: we could simply ignore it. Indeed,
we could ignore all graphs that contain tadpoles. Nothing nearly as nice happens here, unfortunately. We cannot
ignore graphs that have these sorts of insertions on them if they occur in internal lines. But we can ignore these
kinds of insertions if we are dealing with external lines on the mass shell—in particular if we are computing S-
matrix elements. For an S-matrix element, with all external lines on the mass shell, we can ignore all corrections to
external lines. The reason is that in getting an on-shell S-matrix element, we multiply by (p2 − µ2) and then go on
to the mass shell, thus turning the external bare propagator into i. The result of all possible corrections to the
external lines is just to turn the propagator into ′, which has a pole at the same place and a residue at the
same place as the original propagator. So there’s no need to bother with these corrections.

15.4The meson self-energy to O(g2)

In principle, if I were to go on, I should now investigate the coupling constant renormalization, and therefore
complete our program of writing down all the equations to determine all the renormalization constants iteratively.
But just for variety, I’d like to do a simple computation, of the meson self-energy operator function ′(p2) to order
g2. This calculation doesn’t require the coupling constant renormalization. Here you can see how the
renormalizations work out. We’ll also learn some little tricks about how to do the integrals which occur in
renormalization.

Two graphs contribute to order g2 to ′:

One is the first Feynman graph containing a closed loop we are going to look at seriously, and the other is the g2
contribution to the counterterms B and C, which we determine iteratively in terms of the other entities. We’ll write
these contributions as6

Now B2 and C 2 are determined iteratively by the two conditions (15.38) and (15.40):

If I’m not interested in doing higher-order computations, I can eliminate the counterterms from (15.54) at once, and
write

That’s obviously right. I’ve added a term proportional to p2 and a constant term such that the total expression and
its first derivative vanish at p2 = µ2. If you want to compute the counterterms B2 and C 2 to O(g2), of course, you
can do so just by comparing (15.54) to (15.56). But if I’m only interested in computing ′(p2) to O(g2), this
expression (15.56) suffices.

Now let’s do the computation. The important thing is to compute is f (p2), the contribution from the closed
nucleon loop. Then we’ll plug it into (15.56) and get the real ′(p2). Well, to do that, let’s label our momenta:

Figure 15.6: Nucleon loop O(g2) contribution to ′(p2)

There is momentum p coming in. There’s an unknown internal loop momentum which I’ll call q. The
momentum at the top is q + p, and the momentum going out is p. The internal momenta are oriented in the
direction of the arrows (not that it matters, since all the propagators are even functions). The loop momentum q is
not determined by energy-momentum conservation. It runs counter-clockwise around the loop, and we have to
integrate over it.

So we have

There’s a (−ig) from each vertex, and two Feynman propagators, and we have to integrate over the unknown
momentum. The propagator of the antinucleon line carries momentum q and the propagator of the nucleon line
carries the momentum q + p. This is simply a straightforward application of the Feynman rules. As stated, this will
be a function of p only. You may be getting a little antsy. Although many of this integral’s properties are not
obvious, one of them leaps out: it’s divergent! (The denominator goes like q4, at large q, while d4q ~ q3dq, so the
integral is logarithmically divergent.) For the time being we’ll put on blinders. But we’ll worry about that very soon.

The next stage is to manipulate this integral (15.57) using a famous formula due to Feynman:7

The variable x is called a Feynman parameter. This assumes that a and b are such that there is no pole in the
domain of integration. The integral is simple to check, since the integrand is a rational function. We will apply this
formula to our integral for f (p2) using the two Feynman denominators as b and a. Because of the iϵ’s they indeed
satisfy the condition that the denominator never vanishes inside the domain of integration. Therefore I have

To complete the square, let

Thus I can write the integral as

Rewriting integrals in terms of Feynman parameters and completing the square is often convenient. I call this
rewriting the “parameter plus shift” trick. (In the next chapter I will show you how to do this when there are three or
four or five lines running around the loop; then we’ll need more than one Feynman parameter.) We will shortly
transform this integral from Minkowski space into Euclidean space. That will make the integral awfully easy to do,
because the Lorentz invariant integral will become rotationally invariant (in four Euclidean dimensions). You’ll
notice we’ve changed gears. From doing highbrow theory we’re now grubbing around with integrals, but it’s
important you learn how to do both.

Now we have to face the fact that this integral is divergent. Actually we don’t have to do that, because what
we are really interested in is not f, but the whole thing in the square brackets (15.56). If we just look at the first
two terms, the difference f (p2) − f (µ2) goes like 1/q6 (inside the integrand) at high q, and therefore the integral
converges. The last term, the derivative of f with respect to p2 is obviously convergent because the derivative
drags down another power of q2. Therefore ′(p2) is finite. What a surprise! And I mean it really is a surprise. We
embarked upon this renormalization just to turn the wrong perturbation theory for the wrong quantities in terms of
the wrong expansion parameter with the wrong masses held fixed, into the right perturbation theory for the right
quantities with the right expansion parameter and the right masses held fixed, without ever bothering our little
heads about the question of infinities. Renormalization turns out to reveal itself not as Clark Kent, but Superman,
come to rescue us when we are confronted with this otherwise insuperable problem of divergences. We would
have come to a screaming halt at this point if we had not renormalized our perturbation theory. As it turns out,
however, this means that C 2 is in fact given by a divergent integral. Whether that’s bad news or good is something
we still have to worry about. But the quantity ′(p2), the only thing that’s physically observable, is represented by a
perfectly convergent integral.8 Note that a second subtraction is not needed to render ′(p2) convergent in Model
3. Put another way, only the mass counterterm C is needed for the self-energy to be finite; in this theory, a finite
′(p2) does not require the wave-function counterterm B. We’ll come back to this.

Will this continue to all orders in perturbation theory? Does this happen only in this theory, or in all theories?
Well, those are interesting questions, but for the moment let us be thankful for what we have and continue with
this computation. We will turn to those questions later.

15.5A table of integrals for one loop

I will explain how one finishes doing the integral (15.61). This is one of a family of similar integrals that arise in
one-loop integration. It is useful to have a table of such integrals. We’ll derive this integral table now. That requires
a little side story, and then we can assemble the whole thing and get the answer to the integral (15.61) for f (p2).

Let me suppose I have an integral of this form:

where a is some real number. We will normally consider the case n ≥ 3, and n an integer, in which case the
integral is convergent. However we frequently run across expressions with lesser values of n as parts of sums of
terms such as here, such that the total thing is convergent, even though the individual terms are not. And
therefore we should also provide in the integral tables values of this integral for n = 1 or 2, but those are to be taken
cum grano salis, to be used only in convergent combinations.

To do this integral I am going to rotate the contour of the q0 integration. First I’ll write it out explicitly,

Let’s consider where the singularities arise in the complex q0 plane. We have two possibilities: |q|2 − a can be
greater than zero, or |q|2 − a can be less than zero. It could also be equal to zero but that’s trivial, as the two other
cases go continuously into each other. In either case, the contour can be rotated as shown, so that it runs up the
imaginary q0 axis, because the rotation does not cross any poles. This is called a Wick rotation.9 This rotation
translates our integral from Minkowski space into Euclidean space.

Figure 15.7: The Wick rotation for q0

We define
and therefore

Thus our integral becomes

(I may still have to hold on to the iϵ for n = 2.)

We now have a four-dimensional, spherically symmetric integral to do in Euclidean space. So we need

another little piece of lore. Everyone knows how to do spherically symmetric integrals in ordinary three-
dimensional space. How will we do spherically symmetric integrals in four-dimensional Euclidean space?
Consider

where f(qE2) is any function of qE2. If I introduce a variable z = qE2, then I expect

In other words, this integral should equal some constant α, arising from the angular integration, times the integral
of zdz—that’s from the r3dr—times f(z), integrated from zero to infinity.

But what is α? Since α is a universal constant, we could find its value by integrating any constant over
spherical coordinates in four space, but that’s a pain in the neck. We only have to evaluate this integral for a single
2
function to find out what α is. I will look at the function f = e−qE which is just the product of four Gaussians, one for
each component:

On the right-hand side I have

Therefore we have determined α without having to go to spherical coordinates in four-dimensional space, thank
heavens: α is π2. And we have the general rule:

That’s how you determine the volume of a sphere in 4-space without doing any work.10

Now we’re in a position to derive the integral table, given below. I’ll reserve for next time plugging the
appropriate formula into the expression for f(p2) and then doing the integral. Actually we only need to do one
integral from this table, the case n = 1. From (15.67),

From that we can get all the others, by differentiating with respect to a. It will appear in a convergent combination,
so we don’t lose anything by truncating the integration at some high q2, which I call Λ. I’ll assume Λ is much
greater than a. Then, using (15.72),

The integral is pretty simple because the numerator can be written as z − a + a. So I get
which can be approximated as

I can neglect the last term because Λ is supposed to be much larger than a.

Integral table for Feynman parametrized integrals

The Minkowski space integral,

with n integer and Im a > 0, is given by

for n ≥ 3. For n = 1, 2,

and

where the dots indicate divergent terms that cancel when two such terms are subtracted, provided the total
integrand vanishes for high q faster than q–4.

Now if this is part of a convergent combination of terms that in fact do not depend on Λ, so that the integral
doesn’t depend on Λ, that means all terms and the individual integrands that depend on Λ must vanish in such a
combination. That’s what “convergent combination” means. So the two terms with explicit factors of Λ vanish in
convergent combinations. The same is true however many such terms there are. If you now look at the entry in the
integral table for I1, you will see, with the appropriate insertions of i’s and Euclidean rotations, π2 from α and (2π)4
from the denominator of I1, what we have derived is just the I1 entry. You can get I2, I3, . . . by differentiating with
respect to a. I leave that to you as an exercise. You are now in a position to derive the integral table for yourself in
exactly the same way I did.11

Next time we will apply the integral table to complete our computation of ′ to second order. We will discuss
coupling constant renormalization, talk about the marvelous properties of realistic pion–nucleon scattering and
nucleon–nucleon scattering, and have a little more to say about renormalization in general.

1[Eds.]
S. Coleman and Sheldon L. Glashow, “Departures from the Eightfold Way”, Phys. Rev. 134 (1964)
B671–B681. The term was coined by Coleman. The Physical Review editors originally objected to this name, so
Coleman offered “lollipop diagram” or “sperm diagram” as alternatives. The editors accepted “tadpole diagram”.
See Peter Woit, Not Even Wrong, Perseus Books, New York, 2006, p. 54. A tadpole diagram is a blob with only
one line coming out:
2[Eds.] Gunnar Källén, “On the Definition of the Renormalization Constants in Quantum Electrodynamics,” Helv.
Phys. Acta 25 (1952) 417–434; Harry Lehmann, “Über Eigenschaften von Ausbreitungsfunktionen und
Renormierungskonstanten quantisierter Felder”, (On the characteristics of fields quantized by propagation
functions and renormalization constants), Nuovo Cim. 11(4) (1954) 342–357. Källén (pronounced “chal-LANE”)
was a prominent Swedish quantum field theorist, the author of highly regarded textbooks on QED and elementary
particle physics, and one of the first to join CERN’s staff. He died in 1968 when the plane he was piloting crashed
in Hannover, en route to Geneva from Malmö. Källén was 42.
3[Eds.] If σ(a2) = 0, the theory admits no states with p2 > µ2, so no particle creation.
4[Eds.] See (P4.1), p. 175. The relevant formula is = −iπδ(x) + .
5[Eds.] Reminder: Coleman used (p2) for what is usually denoted F(p2).
6[Eds.] In the video of Chapter 23, Coleman says that the superscript f stands for for “Feynman”, not “finite”.
7[Eds.] In a letter to Hans Bethe, Feynman called this identity “a swanky new scheme”. Quoted in Schweber QED,
p. 453.
8[Eds.] This calculation is duplicated, though with a somewhat different focus, in Lurié P&F, Section 6-4, pp.
266–274.
9[Eds.] Gian-Carlo Wick, “Properties of Bethe-Salpeter wave functions”, Phys. Rev. 95 (1954) 1124–1134.
10[Eds.] The volume of an n-dimensional sphere of radius R is

For n = 4, the volume is π2R 4. The surface areas are obtained by differentiation with respect to R, e.g., for n = 4,
the surface area is 2π2R 3. See D. M. Y. Sommerville, An Introduction to the Geometry of N Dimensions, Dover
Publications, 1958, pp. 135–6.
11[Eds.] See also Appendix A.4, pp. 806–808 in Peskin & Schroeder QFT. Copies of this integral table were
handed out in class over the years; handwritten at first but later typed.

16
Renormalization II. Generalization and extension

This lecture will be something of a smorgasbord, with a lot of little topics. First, I would like to complete the
computation of the meson self-energy ′(p2) to O(g2). We will check the calculation by looking at the analyticity
properties of our result, and comparing them with what we would expect on general grounds. Next I will explain
how you tackle graphs either with more lines on a single loop, or with more than one loop. I will show how we can
perform systematically the Feynman trick for the associated integrals, and reduce everything to an integral over
parameters. Then I will return to our renormalization program to consider coupling constant renormalization, the
one renormalization we have not yet discussed in detail. Finally I will make a few not very deep remarks about
whether renormalization gets rid of infinities for every theory, or only for certain special theories.

16.1The meson self-energy to O(g2), completed

The first topic I will just begin in media res. Using our integral table for I2 on the parametric integral we had,
(15.61),

we obtain

The dots represent irrelevant terms from the integral table, terms that vanish in a convergent combination, which
we have. Recalling (15.56) we have, ignoring the irrelevant terms,

We plug (16.2) into this and get

We need not retain the iϵ in f(µ2) or its derivative, i.e., in the denominators of (16.4), and we shouldn’t: f (µ2) had

better be a real number, otherwise something’s gone drastically wrong with our computation. The terms B and C
have to be real, because the Lagrangian is Hermitian. Indeed, since the maximum of x(1 − x) is , and we assume
as always that (µ < 2m) (otherwise the muon would be unstable, decaying into a nucleon–antinucleon pair), m2 −
µ2x(1 − x) is positive definite, and the −iϵ is never needed to avoid the singularity. So

This ugly expression is our final result for ′(p2) to O(g2). Things don’t get any prettier if you integrate. The x-
integral is in fact elementary and can be found in a table of integrals.1 I believe it gives you an inverse tangent, but
don’t take my word for it. I leave it for interested parties to carry out the integration.

As a consistency check, I would like to investigate the analytic properties of this integral. We want to be sure
that the remaining iϵ is unnecessary for p2 < 4m2, and that there is a cut at 4m2. After all, ′(p2) is linearly related
to the inverse of ′(p2) (see (15.36)), and therefore it should have the same analytic properties as ′(p2), except
of course ′(p2) doesn’t have a pole at µ2, where ′(p2) has a pole: it has a zero. Therefore ′(p2) should be
analytic, except for a cut along the positive real axis. I claim that in this order of perturbation theory, the cut begins
at 4m2 (corresponding to a virtual nucleon–antinucleon pair), and not (as you might suppose) at 4µ2
(corresponding to a pair of virtual mesons).

The argument that the cut begins at 4m2 just requires looking at a Feynman graph. From (15.31) the cut is
associated with the function σ(p2) (see Figure 15.3), the amplitude for the meson field to make a state when
applied to the vacuum. If we consider

the graph for that will consist of, to first order in g, simply Figure 16.1 (a):

Figure 16.1: O(g) and O(g3) Feynman graphs for án|ϕ′(0)|0ñ

The field ϕ′ applied to the vacuum can make a nucleon–antinucleon state, and that’s the only order g2 contribution
to the spectral representation, because the contribution from this graph gets squared to make the nucleon loop
(Figure 15.6). The field ϕ′ doesn’t make a meson pair until order g3, as in Figure 16.1 (b), so we won’t get
contributions from two-meson intermediate states in the spectral representation until we reach order g6. We won’t
see them in O(g2). Thus we expect ′(p2) to be an analytic function of p2 aside from a cut beginning at 4m2, as
asserted.

Now let’s work out the analytic properties of ′(p2). It’s an analytic function except for the branch cut
introduced by the cut in the logarithm; see Figure 16.2. This branch cut survives when we do the x integral.

Figure 16.2: Branch cut for ln z

The only part of the integral that we’ve got to study is the numerator of the logarithm, ignoring the iϵ. Troubles
arise if the function is evaluated at the cut,

If this numerator is negative or zero for x in the range of integration, 0 ≤ x ≤ 1, then we’re on the branch line of the
logarithm. When will the numerator be non-positive?
 For 0 < x < 1, x(1 − x) is always positive, so if the imaginary part of p2 is not equal to zero, then the numerator
has a non-zero imaginary part. At the boundary, when x = 0 or x = 1, the numerator is equal to m2, and that’s not a
negative number. So there’s no singularity if Im p2 ≠ 0. If p2 lies along the negative real axis (p2 ≤ 0), again there’s
no problem, because the numerator is positive. So the only case we have to worry about is p2 real, and greater
than 0.

Let’s graph the numerator. It is of course an upward pointing parabola; the coefficient of x2 is positive.

Figure 16.3: Graph of m2 − p2x(1 − x) for 0 ≤ x ≤ 1

At x = 0 and x = 1, the numerator is equal to m2. The numerator reaches its minimum value at x = . To check that
it is positive throughout the domain of integration, we need only check its value at x = :

If p2 < 4m2, we are away from the cut, because then the argument of the logarithm is always positive, and we can
drop the iϵ in the numerator. On the other hand, if p2 ≥ 4m2, there’s a cut, because then the argument becomes
non-positive, and we have to keep the iϵ in our prescription. In this case it matters whether we’ve approached the
real p2 axis from above or below. We should shout in triumph, because these are exactly the analyticity properties
we had anticipated on general grounds.

16.2Feynman parametrization for multiloop graphs

We now turn to the second topic, the machinery of putting together many denominators to generalize Feynman’s
trick, and carrying out an integral that may have more than one loop in it. We can call this “loop lore” or “higher
loopcraft”. As we will see, there is essentially nothing new.

The first thing is to put together many denominators, all the denominators that run around a loop. We know
how to do the parametrization when there are only two denominators. But of course there may be many of them,
more than two. Even with only a single loop, we may have more than two propagators. For example this graph,
Figure 16.4, which we’ve discussed before (see p. 267), has four propagators:

Figure 16.4: A single loop graph with four propagators

At the moment, we do not know how to do the integral associated with this graph.

Consider a product of n Feynman denominators:

The number i goes from 1 to the number n of internal lines around our loop. Each ai is some function of the various
external momenta and the loop momentum. I will derive a parametric expression for (16.9). First, write each of
these denominators as
so that

I will multiply (16.11) by an integral B that is equal to 1, by the rules for integrating a delta function:

where β is a positive constant. We choose β = βi, and rewrite (16.11) as follows:

Changing the integration variables from βi to α i ≡ βi/λ, the right side becomes

(Because of the delta function, there is no contribution if any of the α i’s are greater than 1, so we might as well
lower the upper limits of integration.) The λ integral is elementary:

and we conclude that

This formula tells us how to write a product of Feynman denominators as one big super-Feynman denominator
with parameters, raised to a power. The α’s are Feynman parameters. They are the generalizations of the
variable x in our previous formula, (15.58). The right-hand side of (16.16) looks like an integral over n parameters
if there are n denominators, but of course the delta function makes one of the integrals trivial. In the case n = 2,
you can let α 2 = y and α 1 = x. Once you use the delta function to perform the y-integration, y becomes 1 − x, and
you obtain the earlier formula.

So (16.16) is the generalization to more lines. Please notice it is not clear a priori that the parametrization is
always a good thing to do. It means that any one graph which starts out as an integral d4k can be reduced to an
integral essentially over n – 1 parameters, where n is the number of lines in the loop. This is obviously a good
thing to do if there are four or fewer lines in the loop, as 4 − 1 < 4. It is not obvious that parametrization is a good
thing to do if there are five lines or more.

With the aid of this formula, and the integral table on p. 330, you can reduce any graph with only one loop to
an integral over Feynman parameters. If we could do the remaining α i integrals, we would be very happy people.
But unfortunately it doesn’t turn out that way. These are usually messy integrals that cannot be done in terms of
elementary functions, except in simple cases. And that is why people who calculate things like the sixth order
correction to the anomalous magnetic moment of the electron spend a lot of time programming computers.

This parameter technique can be generalized to graphs with more than one loop. We’ve seen how more lines
are incorporated; I will now discuss more loops. That will complete the lore of doing Feynman integrals, at least for
theories that involve only non-derivative interactions of scalar particles. For more complicated theories, it’s pretty
much the same, except that those integrals have factors in the numerator as well as in the denominator when all
the dust settles.

As an example, suppose I take ϕ 4 theory:

In this case the lowest order nontrivial contribution to the meson self-energy would involve a graph that looks like
this:

Figure 16.5: The lowest order graph for the meson self-energy in ϕ 4 theory

If we call the external momentum p, we have two possible momenta, k1 say, running around the top loop and k2
running around the bottom loop. The momentum on the lowest arc is then p − k1 − k2, all oriented from right to left:

Figure 16.6: Momentum flow for a meson self-energy graph in ϕ 4 theory

Aside from the combinatorial factors (which are a great pain in the neck for ϕ 4 theories, since you have four
identical fields at each vertex), and the constant numerical factors—the g’s, the (2π)4’s, and the i’s—this graph is
associated with an integral of the general form

I’ve suppressed the iϵ’s.

Now let’s consider the general case where we have ℓ loops, and a ki for each loop: i in this case goes from 1
to ℓ. We also have a bunch of external momenta, pj, how many there are depends on how many external lines
there are; and we have in general, n internal lines. I will sketch how to do such an integral, using nothing but our
integral table and the Feynman formula (16.16) to reduce it to an integral over Feynman parameters.

The first part of the trick is to use the Feynman formula to reduce all the internal lines simultaneously to one
big denominator, as I’ve just sketched out. Thus we arrive at an integral of the following form (again I will suppress
numerical factors, including the (n − 1)!):

D is going to be some quadratic function of the k’s, obtained by combining all the denominators. Every internal
momentum is of course a linear function of the loop momenta, the k’s, and the external momenta, the p’s. In our
example, the internal momenta are k1, k2, and p − k1 − k2. So D will be of the following form:

Aij is a symmetric ℓ × ℓ matrix, linearly dependent on the Feynman parameters α i, and independent of the external
momenta pj. If all α i > 0 (as is the case, within the region of integration), it can be shown that Aij is invertible.2 The
Bi are a set of ℓ 4-vectors, linear in the α’s and the external momenta pi. In our example, one of the Bi ⋅ ki might be
α 3p ⋅ k1. C is a scalar depending linearly on the α’s as well as on the squares of the pi’s and the squares of the
masses appearing in the propagators. This is inevitably the general form that D will take.3

Further simplification can be made, because Aij is invertible. We can perform a shift on the loop momenta, to
remove the linear terms (involving the vectors B). That is to say, we can define

Substituting in for ki turns the denominator into

where the new scalar C′ is given by

We’re just doing in general what we did for the one-loop integral. We’ve eliminated all the terms linear in the loop
momentum in the denominator by defining new integration variables which are shifted versions of our old ones.

We can make the integral simpler yet by using the fact that Aij is a symmetric matrix, and therefore we can
diagonalize it. We can introduce new integration variables k″i, linear combinations of the k′’s corresponding to the
eigenvalues ai of Aij, and thereby make Aij diagonal. Since Aij is a symmetric matrix, the transformation k′i → k″i is
an orthogonal transformation with determinant 1, and hence Jacobian equal to 1. The integral becomes

C′, independent of the internal momenta k′i, is not changed by this transformation.

We can make one last transformation for one last simplification:

The ai’s are of course positive within the domain of integration, though there may be some places where they
vanish at the boundaries of the integration. Thus we find our integral becomes

We see now that we didn’t have to worry about analyzing the matrix A, because the product of the eigenvalues is
just the determinant of the matrix. That is, the integral becomes, finally,

So you don’t actually have to go through the diagonalization, you just have to be able to compute the determinant.
You knew the value of C′ before you ever diagonalized the matrix.

We now have the situation in the shape where we can systematically do all the k′′′ integrals, one right after
another, just using our integral table. Whenever we do one of them, we’ll knock n down by two (because k2 → z
after Wick rotation; see the table), and pick up a horrendous numerical factor, and we’ll just keep on going until we
do them all. By this algorithm, we can systematically reduce any integral arising from any Feynman graph,
providing always, of course, it is a convergent graph, or arises in a convergent combination of graphs, in an
integration over Feynman parameters equal in number to the number of internal lines. Thus for example for the
graph I sketched out, the integral would be eight-dimensional in the first instance, over d4k1d4k2. Our prescription
reduces it to a three-dimensional integral over three Feynman parameters, and one of those is trivial because of
the delta function. It’s not the world’s most exciting subject, but if you are ever confronted with the problem of
computing a multiloop graph like Figure 16.5, you will be happy that I have shown you this algorithm. I’ve arranged
it so there are no numerical factors to memorize, just a procedure to understand, which you can work out afresh
for every particular instance. In principle you can reduce any Feynman graph to an integral over Feynman
parameters. At that point, typically, you are stuck, but you can always work it out numerically with a computer.

16.3Coupling constant renormalization

I would now like to discuss briefly the condition that will determine our final renormalization constant. Remember
we were going through the renormalization program for this theory, Model 3, and we had left one thing to fix: the
condition that determines the physical value of g, a matter to be decided (on the basis of appropriate experiments)
by a IUPAP committee (see p. 301), and which would eventually set the value of our last renormalization constant,
F (p. 302):

I will first state the definition, then show you how it works in fixing F iteratively, and finally explain how it can be
connected, through a physically realizable experiment, to what looks at first glance like a totally unphysical object.
To determine A, we studied the one-point Green’s function. To determine B, C, D and E we studied a two-point
Green’s function. To study F, we have to study a three-point Green’s function, with one ψ, one ψ *, and one ϕ.

Define the object −i ′(p2, p′2, q2) as this one-particle irreducible (1PI) graph,

It is of course a Lorentz invariant function. (The −i is included so that ′ = g to lowest order.) Since the three
momenta are arranged so that p + p′ + q = 0, −i ′(p2, p′2, q2) is a function really of only two independent vectors
which we can take to be p and p′, and therefore a function of three inner products: p2, p′2 and p ⋅ p′. Actually, it will
be more convenient for us to write ′ as a function of p2, p′2 and q2 (q2 is linearly related to p ⋅ p′ and the other
two).

Up to third order in perturbation theory it’s easy to see that there are only a very few graphs that contribute to
this thing:

There is the first-order graph, with a contribution −ig(2π)4δ(4)(p + p′ + q). Then there is a genuine monster of a
third-order graph. And finally there may be a counterterm,

evaluated only to third order in perturbation theory. I’ve assigned the monster middle graph as a homework
problem (Problem 9.2), to check that you understand the algorithms for doing integrals for loop graphs like these.

To define the renormalized coupling constant, I impose this condition:

This is the definition of the physical g: We set ′ = g at the one point where all three lines are on the mass shell:

To find a trio of 4-vectors satisfying these conditions, as well as the conservation of momentum p + p′ + q = 0,
some of the components have to be complex. This point cannot be attained by any physical scattering processes,
as the meson is stable. It can be shown, however, that the domain of analyticity of ′, considered as a function of
three complex variables, is sufficiently large to define the analytic continuation of ′ from any of its physically
accessible regions to this point, (16.33), and ′ is real there. (The homework problem asks you to check the
reality of ′ to third order.) The choice (16.32) is totally arbitrary, but it is the one we make for reasons I will explain
shortly. This condition determines F iteratively, order by order in perturbation theory, in exactly the same way as
the other counterterms. For example, because of (16.32), the sum of the last two graphs in (16.30) must cancel at
p2 = p′2 = m2, which determines F3, the coupling constant counterterm F to third order. This completes our
specification of renormalization conditions.

In principle, since the definition of the coupling constant is completely arbitrary, anything that gives g to lowest
order is as good as anything else. That’s the one condition I want to maintain, so I can iteratively determine F.
Aside from that, any value of p2 would do—m2, (m2/µ)2, whatever—and the same goes for p′2 and q2. At this level
it’s just a matter of reparametrizing the theory, according to another IUPAP committee, defining the coupling
constant differently. Still, it is worth devoting a few minutes to explain why this particular definition (16.32) is
useful, and is therefore used by many workers in the field, not for this theory, which is only a model, but for the
corresponding real one, and other theories. The point is this. I’ll show that the square of ′ is a physically
observable quantity if you do the right experiment. That’s all we can hope for, because we can arbitrarily change
the sign of g just by changing the sign of ϕ.
 Consider the process of meson–nucleon scattering:

with everything on the mass shell. I’d like to divide the graphs that contribute to this process into two classes:
those that can be cut in two, and everything else.

The unshaded blobs are one-particle irreducible graphs. The parts that can be cut in two look broadly like s, t and
u–channel graphs, denoted (b), (c) and (d), respectively. Recalling that s is the center-of-mass energy for the
meson–nucleon system, all of the graphs that can be cut in two by dividing a nucleon propagator are in (b). This
graph on mass shell has the form

and has a pole in it, at s = m2. The full nucleon propagator is staring at us from the middle of the graph. As s → m2,
all the graphs in (b) will certainly have a pole. The graphs in (a) we don’t know anything about. However it seems
plausible that they will not have a pole, because they don’t have a propagator joining two parts of a graph. If
you’ve got two or three particles running across the graph, we’ll be integrating over all those propagators, and
we’ll not get poles, but cuts. I ask you to take on trust that only the graphs in (b) have poles at s = m2, while the
others, the graphs in (a), (c) and (d), are analytic at s = m2, although they may have terrible singularities
someplace else. We know that the graphs in (b) have poles at s = m2. That only these have poles at s = m2, is just
a flat assertion I’m asking you to swallow. The graphs in (c) and in (d) presumably have poles at t = µ2 and u = m2,
respectively, but we don’t expect these to have poles at s = m2, and it’s reasonable that they are analytic in s.

Every graph that can be cut in two by cutting a nucleon propagator is drawn as shown in (b), with incoming
meson and nucleon lines meeting in a one-particle irreducible blob, the full nucleon propagator, another one-
particle irreducible blob, and outgoing meson and nucleon lines. Why is this so? The incoming external lines are
S-matrix elements, so they do not get any decorations. So it is obviously one-particle irreducible when you cut
either external line. The line in the middle can be decorated as much as we please, so we decorate it in every
possible way, and get the full propagator. We then go on to the next vertex.

These graphs have a pole at s = m2. What is the residue of that pole? We happen to know it, because the blob
in the middle is the renormalized propagator D′(s) = i/(s − m2). The blob on the right is −i ′(s, m2, µ2), and on the
left the blob is −iΓ(m2, s, µ2). To find the residue at the pole we have to evaluate the coefficient in (16.36) of (s −
m2) at s = m2. The vertices are both simply −i ′(m2, m2, µ2) = −ig. The contribution from the propagator is just i.
Everything else by assumption is analytic near s = m2. That is,

Thus we know how to determine g, or more properly g2, physically. We look at meson–nucleon scattering. It is
some function of s, and of course also the momentum transfer. We extrapolate in s below threshold numerically, or
in principal by an analytic continuation, to the point s = m2. We find a pole there at s = m2, and we determine the
residue of the pole. That is g2, aside from the factor of −i. So that’s how we physically define g.

Why did I say meson–nucleon scattering and not, for example, nucleon–nucleon scattering? No reason in the
world. I can run through exactly the same reasoning for nucleon–nucleon scattering and I will now do it.

I do exactly the same thing for the meson pole that we know from lowest order occurs in nucleon–nucleon
scattering, the t–channel pole.
By exactly the same reasoning as before, with t replacing s, I get

So I could just as well do nucleon–nucleon scattering, and look at the extrapolation to the pole at t = µ2, which of
course is outside the physical region. In the physically accessible scattering region, t runs from zero to some
number depending on the energy. But extrapolate to the pole of t = µ2 and then again you’ll compute g2. Notice
these are two completely different experiments. It’s not that they’re related by crossing or anything; there’s no way
you can cross meson–nucleon scattering into nucleon–nucleon scattering. It’s two different extrapolations for two
completely different experiments. I claim that the two of them, when you massage them in two different ways, will
end up giving you the same number. Now no one has done this in nature, because there are no scalar particles
with these kinds of interactions in nature. But they have studied the pion–nucleon system. This is a real system,
which is similar in its combinatoric structure to Model 3, although there are lots of Dirac matrices floating around at
the vertices. I will tell you what happens. I should also emphasize that this is not a perturbation theory result.
Although we have obtained it in the context of perturbation theory, this is true in all orders, the whole summed up
theory.

Chew and Low4 analyzed pion–nucleon scattering in the forward direction, where the best experiment was,
analytically continued to the nucleon pole that exists in pion–nucleon scattering, and computed g2. They got it to
within two or three percent, because of the experimental inaccuracies. As I recall, for this system g2 is 13.7, so
you don’t want to use perturbation theory.

Several years later Mike Moravcsik said, “Gee, there’s a lot of data on nucleon–nucleon scattering. Wouldn’t
it be nice if we could extract out the effect of the pion pole?” In fact he knew this was possible. The longest range
part of the force, the Yukawa potential with a range of the inverse of the pion mass, should come just from this
pole, due to the pion exchange. And there will be cuts in t beginning someplace else which will give shorter range
potentials. Moravcsik knew that there were tremendous phase shift tables on nucleon–nucleon scattering. He and
his colleagues5 made the first few phase shifts completely free parameters, to take care of the short range part of
the potential, whatever it was, at low energies. The remaining scattering data were fit with the Born approximation.
Why the Born approximation? Because when you go out to large phase shifts, you’re very far from the center of
the potential, so even if the coefficient in front of the Yukawa potential is large, it’s still a weak force insofar as it
affects the large phase shifts. With g2 and the pion mass as free parameters, low energy nucleon–nucleon
scattering with arbitrary phase shifts were fit for the first four or five partial waves, and the higher partial waves
were fit with the Born approximation. And lo and behold, the actual pion mass came out, somewhere between 135
and 140 MeV, and the best fit coupling constant was, as I recall, within five or ten percent of that found in
pion–nucleon scattering. The experimental errors were a little worse for this system, compared with those found
by Chew and Low from looking at a completely different system, but the agreement between the values of g2 was
not bad. What we have done in Model 3 is to equate the coupling constant to residues arising from the poles in two
different systems, nucleon–nucleon scattering and nucleon–meson scattering. And when two similar real systems
were compared, pion–nucleon and nucleon–nucleon scattering, the values of the coupling constants were found
to agree within a few percent. So our procedure, though applied to a toy model, seems to be checked by a real
experiment.

That takes care of topic three, coupling constant renormalization. You have now seen how in principle to
compute an arbitrary graph in our model theory, including the effects of all renormalization counterterms, or at
least reduce it to an integral over Feynman parameters.

16.4Are all quantum field theories renormalizable?

The last topic I want to discuss is the relationship between renormalization and infinities. We have seen that in the
low order graphs of Model 3, the renormalization constants eat the infinities. In fact we have more renormalization
constants than we need to eat the infinities that occur in this theory. For example, we have a graph6 associated
with coupling constant renormalization, in ′(p2, p′2, q2):

Figure 16.7: O(g3) graph in Model 3

Its integral is finite; at high k it goes as d4k/k6 because there are three denominators around the loop. At least at
O(g3), the coupling constant counterterm, (16.31), is not needed to eat the infinities. It is required, of course, to
give the beautiful result of g2 measurable in two different experiments. (This extreme convergence is peculiar to
Model 3 and other models where all the coupling constants have positive mass dimension, as we will see later.)
Let’s look at a few low order graphs for a somewhat more complicated theory, using only crude estimates, to see if
the renormalization constants we have will eat the infinities or not.

As a first example let me take a single free scalar field, interacting with itself due to a quartic interaction:

I’ve already written down several graphs from that theory. This theory is much more divergent in low orders than
Model 3. For example, in order g02, we get this graph which I wrote down earlier, the so-called “lip graph”:7

Figure 16.8: O(g02) graph in ϕ 4 theory

At high k, this graph is quadratically divergent, not just logarithmically divergent, because you have d4k1 d4k2
in the numerator, eight powers of k, and only six powers of k in the denominator, from the three propagators.
Fortunately, in this theory we have more counterterms than we need. Remember when we were studying Model
3’s cubic interaction gψ *ψϕ, all we needed was a mass renormalization counterterm, C, , to make
the self-energy finite; the additional subtraction caused by the wave function renormalization counterterm B was
not needed.8 It is an easy check in this theory, which you can do on the back of an envelope, that the ϕ 4 theory
needs both renormalization counterterms to render things finite. But these two suffice. Before, in Model 3, the first
subtraction turned a logarithmically divergent integral into a convergent integral. With a quartic interaction, the first
subtraction turns a quadratic divergence into a logarithmic divergence, and the second subtraction turns the
logarithmic divergence into a convergent integral. To put it another way, all we really need to know is the second
derivative of this graph with respect to p2, since its value and its first derivative at the renormalization point are
fixed. Every time we differentiate with respect to p2, we put an extra power of p2 in the denominator. Do it twice,
and you’ve made the integral convergent. Recall (16.3), and note that

so that

We also have to the same order, g02, a correction to the four-point function, as shown in Figure 16.9, plus crossed
versions of this graph. We’ve seen this before (Figure 14.3). It’s only logarithmically divergent. That will be
canceled by the coupling constant renormalization counterterm shown in Figure 16.10, which one introduces in
this theory. The counterterm just makes a single subtraction and treats the correction to the four-point function just
like the treatment of ′(p2), and makes everything finite.
 Figure 16.9: O(g02) correction to 4 in ϕ 4 theory

Figure 16.10: O(g02) four-point counterterm in ϕ 4 theory

At O(g03), we have a graph like Figure 16.11. The integral associated with this graph is finite, because at high k it
goes as d4k/k6. Well, things look pretty good. But it also looks like we are approaching some sort of boundary.

Figure 16.11: O(g03) graph in ϕ 4 theory

Consider the fifth degree interaction

Here things are going to blow up in our faces. A Lagrangian is renormalizable only if all the counterterms required
to remove infinities from Green’s functions are terms of the same type as those present in the original Lagrangian.
But that isn’t going to happen here. Let’s look at the simple one-loop graph in Figure 16.12.

Figure 16.12: O(g2) correction to (5) in ϕ 5 theory

This graph is logarithmically divergent, our good old d4k/k4 integral. To cancel this graph’s divergence, we’d need
a term that would give rise to a graph like Figure 16.13. This is a ϕ 6 counterterm. But there was no ϕ 6 term in our
original Lagrangian, and so we’d have to add a term of higher degree to the Lagrangian than was originally
present. We’re stuck! This theory, even on the lowest level of renormalization, does not eliminate the infinities.
Someone says, “Okay, wiseguy, I guessed the wrong theory. I agree the ϕ 5 theory is no good. But I’ll put in a ϕ 6
term from the start!” Well, let’s see what happens with this theory.

Figure 16.13: O(g2) six-point counterterm in ϕ 5 theory

Now we’ve got a ϕ 6 term that will admit a counterterm to cancel the logarithmic divergence from the graph in
Figure 16.12. But with the ϕ 6 term, you also get these graphs, Figure 16.14 (a), arising at order h2, and a cross-
term graph, Figure 16.14 (b), of order gh. These would require new counterterms of ϕ 7 and ϕ 8 to cancel them.
Well then, I guess I need ϕ 7 and ϕ 8 interactions in the Lagrangian, too:

Figure 16.14: Graphs arising from gϕ 5 and hϕ 6 terms

But then you need counterterms to up to 12th order, which require new terms in the Lagrangian. It just keeps going
up and up, an unending escalation of ambiguities! In order to cancel all of the divergences that arise generated by
the ϕ 5 term, you need to add a ϕ 6 term. To cancel the divergences of the ϕ 6 term you have to introduce a ϕ 7 and
a ϕ 8 term. To cancel the divergences of the ϕ 7 and ϕ 8 terms, you need up to ϕ 12 terms, and it doesn’t stop there.
It never stops! As soon as we introduce the ϕ 5 term, it’s like the single bite of the apple in the Garden of Eden, the
whole thing collapses. The only way of making everything finite and eliminating all the divergences is to have an
infinite string of coupling constants. Theories that cannot be made finite without introducing an infinite number of
counterterms are called non-renormalizable. Theories where you need only a limited number of interactions to
get a finite theory are called renormalizable. We have not shown that theories with interactions involving only
three and four fields are renormalizable; we’ve just shown that nothing goes wrong in low order. There’s a
complicated theorem which I will talk about later to show that they are in fact renormalizable. But I’d like to
postpone discussing that until we can handle fermions, and do everything at once.

At least from the viewpoint of perturbation theory, as soon as we introduce a little bit of ϕ 5, everything goes
crazy. Notice that it’s the infinities that make the situation drastically more constrained than in non-relativistic
quantum mechanics. In non-relativistic quantum mechanics, you describe your dynamical degrees of freedom,
and then you can write down the interaction between them pretty much as you please: two body forces, three
body forces, four body forces, nothing goes wrong with any of those things as long as they aren’t too pathological.
Here, it’s not so. If I have a single ϕ field, I can have a cubic term, and I can have a quartic term. And that’s it.
Anything else, the whole thing goes bananas. Whether that is because of our ignorance, or because such theories
are really and truly nonsensical, at this time no one knows.

Next time we’ll discuss what happens if the meson becomes heavier than twice the nucleon, and the meson
becomes unstable.

1[Eds.] From Mathematica,

if Re ≠ 0
and ∉ Reals.
2[Eds.] See Noboru Nakanishi, Graph Theory and Feynman Integrals, Gordon and Breach, 1971, theorem 7-2, p.
58.
3[Eds.] In response to a question, Coleman adds that the energy-momentum conserving delta functions at the
vertices have been left out. In the example, the original graph had momenta as shown below.

The two delta functions, δ(4)(p − k1 − k2 − k3) and δ(4)(k1 + k2 + k3 − p′), at the vertices produce an overall energy-
momentum conserving delta function δ(4)(p′ − p), and allow a trivial integration over one of the three loop
momenta, k3. Then D is, from multiplying out (16.18),

Comparing with (16.20), one identifies

Thus Aij is linearly dependent on the α’s and independent of the external momenta, Bi are 4-vectors linearly
dependent on the α’s and the external momenta, and C is a scalar linearly dependent on the α’s, the squares of
the external momenta and the squares of the masses in the propagators, exactly as described. Note also that, as
expected, det A ≠ 0.
4[Eds.] G. F. Chew and F. E. Low, “Effective-Range Approach to the Low-Energy p-Wave Pion-Nucleon
Interaction”, Phys. Rev. 101 (1956) 1570–1579. The coupling constant f2 = (4m2/µ2)g2 was found to be 0.08,
giving g2 = 14.5 (with mN = 939 MeV, and µπ = 139.6 MeV).
5[Eds.] Peter Cziffra, Malcolm H. MacGregor, Michael J. Moravcsik and Henry M. Stapp, “Modified Analysis of
Nucleon-Nucleon Scattering. I. Theory and p–p Scattering at 310 MeV”, Phys. Rev. 114 (1959) 880–886.
6[Eds.] Though drawn differently, this is the same as the “monster” graph in (16.30), whose evaluation is Problem
9.2, p. 349.

7[Eds.] Often drawn as , and called the “sunset” graph.

8[Eds.] See the discussion on p. 327.

Problems 9

9.1 (a) Compute, using (16.5), the imaginary part of the meson self-energy .

(b) Use the spectral representation (15.30) for to show that in (15.36), to all orders,

and find the constant of proportionality. H INT: You may wish to use the hint in Problem 4.1, (P4.1).

(c) Compute σ(p2) to O(g2), and verify that your answer to (a) is consistent with your answer to (b).

Hint: The spectral density σ(p2) is defined (15.12) in terms of the matrix element of a renormalized field between
the vacuum and a multiparticle state. Use (14.37) and (14.15) to express this matrix element in terms of an
integral over an appropriate renormalized Green’s function.
(1986a 18)

9.2 (a) Compute the Model 3 vertex (16.29), to order g3, as an integral over (two) Feynman
parameters, for p2 = p′2 = m2.

(b) Show that this is an analytic function of q2 in the entire complex q2 plane except for a cut along the positive real
axis beginning at q2 = 4m2. (This function also has interesting analytic properties when all three arguments are
complex, but untangling the analytic structure of a function of three complex variables is a bit too much work for a
homework problem.)
(1986a 19)

Solutions 9

9.1 (a) From (16.5), discarding the iϵ’s in the denominators (dismissing the possibility that µ2 > 4m2—see p. 331),

For real p2, this expression has non-zero imaginary part only when the real part of the argument of the logarithm
becomes negative. For x ∈ [0, 1], this can happen if p2 > 4m2. The iϵ prescription tells us how to deal with the
branch cut when p2x(1 − x) > m2. Let ξ = p2x(1 − x) − m2. We know ξ is real, because p2 is real. Then
So

The function −ξ(x) is an upside-down parabola with roots x = {x1, x2} = (1 and positive between
those roots. If p2 > 4m2, to O(g2)

If p2 < 4m2, Im

Though the problem specified beginning with (16.5), the imaginary part of (p2) can also be calculated from
the integral in note 1, p. 332. Taking the limit ϵ → 0, we have

where the dots indicate other real terms. Again, if 4m2 > p2, this expression is real, and Im (p2) = 0. So look at p2
> 4m2:

exactly as before.1

(b) The spectral representation (15.30) for ′(p2) says

We also have, from (15.36),

Setting these two expressions for −i ′(p2) equal to each other,

Using the hint (P4.1), we can write (in the limit as ϵ → 0)

Take the imaginary part of both sides of (S9.9),

There are two cases: p2 ≠ µ2, and p2 = µ2. In the first case, we can drop the iϵ on the left-hand side, because there
will be no pole to avoid: µ is the physical mass of the meson, so by definition the pole is at p2 = µ2; as a
renormalization condition, (µ2) = 0 (see (15.38)). We can also drop the delta function on the right. Then

That is, for p2 ≠ µ2,

The constant of proportionality is −π. For the case p2 → µ2, the limit of the left-hand side (S9.11) is −πδ(p2 − µ2).
All will be well if

In perturbation theory, this is fine because we know (15.13) that σ(p2) = 0 if p2 < min(4µ2, 4m2). If µ < m, then p2 =
µ2 < 4µ2, and σ(µ2) = 0. If m < µ, we also know that µ2 < 4m2 (because the meson is stable), so again σ(µ2) = 0.
Consequently, within perturbation theory, whatever the value of p2, (S9.13) holds. (Another derivation will be given
in Chapter 17; see (17.4).)

(c) We need to evaluate, to O(g2), the spectral density σ(q2), (15.12):

(The prime means that |nñ is neither a single meson state nor the vacuum.) The kets |nñ can be taken as out
states; they are a complete set. Using (14.37) and (13.3), we can say

The equality between these two expressions comes from (14.15). We Fourier transform the Green’s functions with
the convention (13.4), and obtain

Substitute this into (S9.15) and obtain, after differentiation and integration,

(We actually only need this for y = 0.) The Green’s functions (13.3) contain the counterterms, an overall energy-
momentum conserving δ(4)(k1 + k2 + . . . − q) (so we can do the integration easily), and propagators for all n + 1
external lines. The n factors of (k2i − µ2)/i cancel out all the outgoing particle propagators. Because the terms
outák . . . kn|ϕ′(y)|0ñ appear squared in σ(p2), and we are only asked to calculate σ(p2) to O(g2), we need only
1
calculate the Green’s functions to O(g). We are freed from computing counterterms, which are all zero to O(g) in
Model 3. Nor need we worry about more than three legs on the Green’s functions. Consider ′(7). Say the meson
goes straight through the blob. One of the possible contributions to the other parts of ′(7) is the disconnected
graph that looks like this:

Each disconnected part has its own delta function, e.g., the graph at left has δ(4)(k1 + k2 + k3). The arguments of
these delta functions will never equal zero, because all the ki are outgoing, so these graphs do not contribute. The
only exception is if all the other parts look like a single line, but those are excluded by the prime on the summation:
we are not including single meson final states. So we needn’t consider the meson simply passing through a blob.
The only surviving possibilities include a meson with momentum q branching into a nucleon–antinucleon pair, plus
perhaps extra disconnected parts. If there are extra disconnected parts, then as before, each contains a delta
function whose argument will never equal zero. These parts, if they even appear, will contribute nothing. So we
are left with one contribution to exactly one Green’s function, ′(3)(−k, −k′, q) at order g, with this graph:

The outgoing nucleon has momentum k, the outgoing antinucleon has k′ and the incoming meson has q. After we
substitute this into (S9.17), the nucleon and antinucleon propagators cancel the momentum factors, leaving only a
q integration to be performed (with y = 0 in the exponential’s argument):
So to O(g2),

The summation in this case is a pair of integrations over k and k′ over all states |k, k′ñ with one nucleon and one
antinucleon:

Because of the delta function, we can take the meson propagator out, with k + k′ = p. Moreover, as k and k′ are
timelike vectors, their time components are positive, and so must be p0. We can then set θ(p0) = 1, and write

The extra factor of 1/2π allows us to write the coefficient of the delta function as (2π)4, and the two integrals are
now seen to be nothing but the density D, of final states for two particles, (12.16). Integrating over the angle we
obtain (for the center-of-momentum frame) from (12.24)

The delta function says that p = k + k′. The total energy ET is just In addition, we have k = (k0, k)
and k′ = (k′0, k′). Since both k and k′ are on the mass shell, and we’re in the center-of-momentum frame, we have
k′ = (k0, −k). Then

So

Putting all the factors together,

which gives the value of σ(p2) to O(g2), as asked for. Consequently, from (S9.13),

because, to O(g), D′(p2) = (p2 − µ2 + iϵ)−1. This is identical to (S9.4), as required. If p2 < 4m2, outák, k′|ϕ′(0)|0ñ = 0,
so in that case, (p2) = 0, as before.

9.2 (a) The Model 3 vertex to O(g3) is, for p2 = p′2 = m2,

The value of F3, the counterterm F to O(g3), is determined by the renormalization condition −ig to
be equal to iΓf(m2, m2, µ2). Redrawing the middle term,

Applying the Feynman rules (box, p. 216) to this graph,

First we follow the formula (16.16), and rewrite:

We can integrate over α 3 easily because of the delta function. Rewriting α 1 = x, α 2 = y and α 3 = 1 − x − y, the
integral becomes

where

Now we shift the variable of integration:

The momentum integral becomes

Using the integral table on p. 330, the equation (I.2) for the case n = 3 gives

We can simplify a a little bit. By momentum conservation q = p + p′. Squaring both sides we find 2p · p′ = q2 − (p2 +
p′2). Next, we are told to restrict our attention to p2 = p′2 = m2. Then

Finally, then,

and

(b)Where is (m2, m2, q2) analytic in q2? The only part of the vertex function that depends on q2 is Γf(m2, m2, q2).
This is by inspection an analytic function except where the denominator equals zero (hence the iϵ in the
denominator). The x, y integration runs over a triangular region as shown:

Because of the form of the denominator, it makes sense to reparametrize the integral as

Then we can rewrite the denominator as (w 2 − z2)q2 − 4(w 2m2 + (1 − w)µ2). The function will cease to be analytic if

Every pair (w, z) gives an excluded value of q2. It’s easy to see that for any value of w, z can be chosen arbitrarily
close to w, so that q2 can be made arbitrarily large. The minimum value of q occurs when the denominator takes
its maximum value, namely for z = 0. Then

and the least value of this in both w and z is clearly Consequently the function is analytic
except for a branch cut from q2 = 4m2 → ∞, along the real axis.

1 [Eds.] For real x, Re tan−1 (−ix) =

17
Unstable particles

I’d like to review briefly what we’ve done in the past few chapters. We’ve gone through a whole lot of complicated
technical reasoning, and you might not be able to see the forest for the trees or even see the trees for the leaves.
I’d just like to devote a few moments to recapitulating the arguments of the last four chapters.

First, we gave a description of scattering theory on a basis that had nothing to do with any approximation
scheme, nor with any artifactual features like adiabatic turning on and off. We proved, in my usual handwaving
sense, the Lehmann, Symanzik and Zimmerman reduction formula, (14.19). That formula tells us if you know the
Green’s functions exactly, then you know the S-matrix elements exactly, and a fortiori if you have an
approximation for the Green’s functions, you have an approximation for the S-matrix elements. Second, we gave
a new interpretation to our only approximation technique, Feynman perturbation theory. We showed that Feynman
perturbation theory is in fact a perturbation theory for the Green’s functions when the lines are off the mass shell,
and we took account of all the trivial kinematic factors in the way I described. However, Feynman perturbation
theory is actually many possible perturbation theories, because there are many possible ways of breaking a given
Hamiltonian up into a free part and an interacting part. With each such break-up you get a different expansion.
The most naive way of breaking up the Hamiltonian, gathering all the quadratic terms and calling them the free
Hamiltonian, and simply taking the cubic and higher terms and calling them the interaction, leads to an expansion
which, although perfectly valid (at least, as valid a priori as any other expansion) is not particularly useful. It is an
expansion for the wrong Green’s functions, those of the unrenormalized fields, in terms of the wrong coupling
constant, the impossible-to-observe bare coupling constant, with the wrong parameters held fixed, and the
unknowable bare masses. We corrected this difficulty by breaking up the Hamiltonian in a different way, where the
free Hamiltonian was given in terms of the renormalized fields, the physical masses and the physical coupling
constant. The difference between the free Hamiltonian in terms of the renormalized quantities and everything else
is now the interaction. Such an expansion necessarily generates counterterms. Of course, there is no way of
telling what the counterterms are until you insert into your computational scheme the definition of the parameters
you call µ and the quantity you call ϕ′, etc. Therefore we went through a long song and dance in which we found
out how to define those things consistently within our perturbative scheme, that is to say, as properties of Green’s
functions. That gave us the definition of the physical masses and physical coupling constants for the renormalized
fields, and allowed us to insert explicitly into our scheme the definitions of µ, m, ϕ′, g, etc.

In the course of these developments we went on many excursions, and found many interesting things that will
be useful outside of this computational scheme. Three of them in particular will be very important to us later. First
was the spectral representation (15.16) for the propagator, which comes up in other contexts of physics. Second
we learned the lore of loops, how to reduce any Feynman integral from a momentum integral to an integral over
Feynman parameters. Third, we encountered and conquered, at least in low orders, the infinities that occur in
Feynman perturbation theory. From this way of looking at things we found a surprise, that at least in low orders, for
certain simple theories, the infinities could be absorbed by the renormalization counterterms. On the other hand,
we found that it was easy to construct theories for which the infinities could not be absorbed by the
renormalization counterterms, as I discussed at the end of the last chapter. Of course we did not prove that our
theories with cubic and quartic interactions are renormalizable in the sense that all infinities that occur in all orders
are eaten by the counterterms. We showed only that this holds in low order, and we postpone until a future date
the question of whether it happens to all orders. That’s the summary of what we went through.
 17.1Calculating the propagator for µ > 2m

I would now like to turn to the remaining loose ends in Model 3. The one computation we have not redeemed in the
new formulation of scattering theory is the one1 where you computed the ratio g/mK from the decay rate Γ of the K
meson, represented by the field ϕ, when the mass µ = mK of the ϕ was greater than twice the mass, m = mπ, of the
“nucleon” field ψ:

In our formulation of scattering theory, the concept of an unstable particle never occurs. There’s an unstable
particle? Nu, there’s an unstable particle.2 You just forget about the field associated with that particle. It will never
appear in asymptotic states, because it decays long before the time gets to be plus or minus infinity, and you
compute scattering matrix elements between stable particle states. Nevertheless, it is amusing to ask the
question: What if someone with no physical intuition whatsoever decided to follow through our computational
scheme, and chose µ, the renormalized mass of the meson, to be greater than 2m?

Surely our hypothetical person must run into trouble: We can’t make a stable meson with mass greater than or
equal to 2m. But what is the specific trouble he will encounter? Well, before I answer that question, I’d like to
derive an identity (S9.13) for the imaginary part of ′, which appears in the solution to Problem 9.1. This identity
has nothing to do with whether or not the meson is stable, but it will be useful to us in our investigation.

From (15.36),

We can deduce a formula for the imaginary part of ′ using the fact that for any complex number z, the imaginary
part of z−1 is minus the imaginary part of z divided by the absolute value of z squared:

We have a formula (15.31) for the imaginary part of −i ,

So we find

in agreement with (S9.13). (We’re always talking about real p2 here, because (15.31) assumes real p2.) Recall the
definition, (15.12), of σ(p2):

(the prime indicating that neither a single meson state nor the vacuum state is included in the sum). So we can
write (17.4) as

Here of course not only is p2 real but p0 is greater than zero, otherwise Im ′(p2) would be zero. So we can drop
the θ(p0). We’ll use this formula very shortly.

A significant feature of this result, as we saw both in the discussion following Figure 16.2 and in Problem 9.1,
is that ′(p2) has an imaginary part for p2 > 4m2 in lowest nontrivial order in perturbation theory. Thus, if our
hypothetical dumb cluck, attempting to carry through the renormalization program for µ2 > 4m2, attempts to
impose the condition

he will arrive at a complex counterterm B2, because he has to subtract not only a real number but also an
imaginary number. He would have to be really dumb not to realize that something is going wrong in this case,
because he should realize that he is adding to his Hamiltonian a non-Hermitian operator by allowing B2 to be
complex. Even if he doesn’t know anything else, he should recognize that that’s bad news. So imposing the
condition (17.6) is not possible unless you want to play with theories with non-Hermitian Hamiltonians.

Of course this means that if µ2 > 4m2, our standard renormalization conditions are not applicable. (Everything
I say for ′ itself holds as well for its derivatives.) From the viewpoint of pure scattering theory, in fact, we don’t
need any renormalization conditions for this particle, because this particle is unstable, and therefore will never
appear on the outside of the scattering process. So how we renormalize it is totally irrelevant. However,
renormalization does also serve the purpose of eliminating infinities, and it would be nice to have some set of
conventions that accomplish this. The conventions I’ll choose to adopt in the remainder of this discussion are
these:

These conditions are arbitrary, but they enable us to fix our subtraction constant in a way that’s continuous as µ2
goes from below 4m2 to above 4m2. They’re certainly not forced on us by any general principle. I adopt them
because they will make the things we’re going to discuss somewhat simpler. The conditions have no physics in
them, but they are certainly allowed; I can fix those counterterms any way I want.

Let’s use these conditions to compute what the propagator looks like to order g2. We did it for the case when
things are stable (see (15.36) and (16.4)), when µ2 < 4m2. Let’s now do it for the case where µ2 > 4m2. I will
restrict myself to the neighborhood of the point µ2 on the real p2 axis because I don’t want to write down horrible
equations. In fact I will treat (p2 − µ2) itself as a quantity of O(g2). The obvious tool for looking at the propagator in
this region is a power series expansion of (17.2):

The first term by hypothesis is O(g2), so we keep that. The third term is in fact O(g4). The factor (p2 − µ2) is O(g2).
The derivative d ′/dp2 evaluated at p2 = µ2, like ′(p2), is also O(g2). We know from our renormalization
condition (17.7) that ′(µ2) has vanishing real part:

So the inverse propagator becomes

Now we happen to have a formula, (17.5), that will give us the imaginary part of ′(µ2). To order g2, the only
diagram that contributes to this formula looks like this:

The only thing the meson field can make to order g2 is a nucleon–antinucleon pair. Since the meson field is off the
mass shell, the value of this diagram is −ig from the coupling constant, times i/(p2 − µ2), the lowest-order
propagator. That makes the contribution to order g2. So (17.5) becomes

The (p2 − µ2)2 is canceled by the | (p2)|−2, which must also be taken to lowest order in g2. We wind up with the
same diagram and much the same matrix element we computed3 for the kaon decay. The contribution from
(17.12) differs from that of the matrix element we looked at in our old dumb theory of unstable particle decay
mainly by the factor of (p2 − µ2)−2, but we have, nicely enough, | |−2 in front to cancel that factor out. The
limit p2 → µ2 of Im ′(p2) gives us, apart from the factor of − , the rest frame amplitude for the decay. From (12.2),
so that

Then

and

These last two expressions are equivalent, because Γ here is also O(g2).

Thus we see that the effect of the interaction that enables this particle to decay is to displace the pole in the
propagator from p2 equals µ2 to p2 equals (µ − iΓ/2)2. Here’s what it looks like on the complex plane.

Figure 17.1: The pole µ2 shifted to (µ − iΓ)2

When you look at this drawing, you may think I have gone bananas. The dashed × on the cut is the pole µ2. Now
I’ve got a pole at (µ − iΓ/2)2 in the complex plane, even though I have proved to you that the propagator was
analytic in the complex plane. The point is, however, that we’re starting out above the cut, and doing a power
series expansion there. So we get a power series expansion that’s valid in some circle when we go above the cut,
and when we go down, we go through the cut into the second sheet of this function of a complex variable. I
indicate that by a dotted line. You should think of this as a circular disk of paper that is sitting on top of the cut, and
then goes through the cut and extends onto the second sheet. So this pole is there, all right, but it’s not in the
function on the cut plane, but in what we would get if we were to analytically continue the function from above the
cut, to the second sheet.

17.2The Breit–Wigner formula

I would now like to investigate the physical consequences of what we have discovered in lowest order perturbation
theory. In particular, I would like to discuss two features traditionally associated with unstable particles, the
Breit–Wigner formula4 and the exponential decay law, and see how they arise in this formalism. I will do this by
means of two thought experiments. I will completely forget the perturbation theory origin of these formulas. My only
assumption will be that I have some field with a propagator, of the form

for some real parameters µ and Γ, and for some range of the real axis. Of course there’s no need for the iϵ now
that I’ve put in the iµΓ. Certainly this expression for the propagator isn’t true everywhere; it doesn’t have the right
analytic properties. But I assume there is some tiny stretch of the real axis where this is an excellent
approximation for the propagator. That certainly happens in second-order perturbation theory, as we’ve seen, but
it may well happen in other circumstances where there is no trace of a fundamental ϕ′ particle. You could, for
example, arrive at an unstable state because of some mechanism analogous to those which arise in non-
relativistic quantum mechanics, where you have a bound state that’s not quite bound. Only if you turned up the
value of the coupling constant would this state become a bound state, but now it appears as an unstable state, or
resonance. So the form (17.18) of the propagator is the only assumption I will make here.

The first thought experiment involves the momentum analysis of an idealized production process. I add to the
Lagrangian, L, that describes this theory, an external perturbation of the sort we’ve discussed so frequently
before,

I’ll look at an extremely idealized case where ρ(x) is simply a delta function:

I’ve called the coupling constant λ so you won’t confuse it with any coupling constant that’s already in the
Lagrangian, like g. Now in this case it’s trivial to determine the amplitude for going into any given final state. At
space point 0, at time point 0, I bash the vacuum of the system with this field. I impulsively turn it on and turn it off,
and see what comes out of the vacuum. To lowest order in λ, the amplitude for going to any given state |nñ from
the vacuum is simply

ignoring any phase factor. I’ll write this as a proportion, since I’m not particularly interested in 2π’s or anything like
that. Let k be any four-momentum that’s allowed in the final states of the theory. Then the probability of producing
a final state with momentum k, which I’ll call P(k), can be written as

The delta function is to extract just the part that has momentum k. I can keep λ as small as I want, so I can
suppress those corrections (of order λ3) as much as I like. There may be kinematic factors in here, but that’s all
washed into the proportionality sign.

Near k2 = µ2, the sum in (17.22) has a very simple formula. It is precisely what appeared on the right-hand
side of our expression for the imaginary part of the propagator (see (15.12) and (15.31)). So P(k) is proportional to
λ2 times the imaginary part of (k2):

Now we have an ansatz (17.18) for (k2), and we can easily find its imaginary part. We just plug that in:

If we graph P(k) as a function of k2, we get the characteristic Lorentzian line shape, or Breit–Wigner shape, very
sharply peaked near the mass of the unstable particle with a width depending on the parameter Γ, as shown in
Figure 17.2. This is the same result you find for scattering amplitudes in non-relativistic theories near an unstable
energy level.

Figure 17.2: The Breit–Wigner distribution

You may be a bit disturbed if you remember that the full width at half maximum is supposed to be Γ, in the
ordinary non-relativistic analysis, and here it looks like it’s 2µΓ, but that’s simply because we’ve written things in
terms of the squared invariant mass, k2, of the state that’s produced. If we wrote things in terms of the center-of-
momentum energy of that state, that is to say, if we chose the four-vector k to be (E, 0), then

since E is supposed to be close to µ. Dropping the terms of O(λ3),

Aside from the factor of 4µ2 in front, this is now the conventional Breit–Wigner denominator, with the where it is
in the familiar formula. In terms of energy we have exactly the same kind of peak, and Γ is indeed the full width at
half maximum, just as it should be in the conventional Breit–Wigner formula.

You have to be careful when you compute Γ’s to higher orders, because as soon as you get diagrams that
involve internal meson lines, you get funny things going on. Those meson lines can be on the mass shell in the
range of integration, so pretty soon you will start computing amplitudes that have funny singularities in them. The
computation cannot be carried out to all orders. The best way, if you want to calculate Γ to all orders for some
exceptionally refined analysis, is to compute the propagator to all orders. That leads you to no ambiguities along
the real axis, and you can extrapolate the propagator into the complex plane and see where the pole is.

That takes care of the Breit–Wigner formula, one of the two hallmarks in non-relativistic scattering theory of
the occurrence of an unstable state. We have, I grant you, not done a general scattering experiment but a very
idealized production experiment: we get a ϕ′ hammer and hit the vacuum with it and see what comes out. But as
expected, what comes out, the probability P(E) as a function of the center-of-momentum energy, has the
characteristic Breit–Wigner shape.

17.3A first look at the exponential decay law

The second hallmark of unstable particles is the exponential decay law. This was the first principle used in the
studies of radioactive nuclei once it was discovered in the late 19th century. A second thought experiment will
enable us to see if the exponential decay law is in our model.

Suppose we conducted an experiment to measure the lifetime of some unstable particle, say a radioactive
nucleus, or if you want to be more glamorous, a K meson or something like that. How would we do it? Well, first
we’d make a K meson in some specific region, with some well-defined momentum. This is already rather tricky.
Typically you send a high-energy accelerator beam into a target a few centimeters across or maybe a bit larger,
and all sorts of junk comes out. And then you put in all sorts of magnets and various devices designed to filter the
momentum, and to make sure you’ve only got K mesons coming out at the other end. When the beam hits the
target, there will be all sorts of things coming out that you’re not interested in, like π mesons and fragments of the
target, several atoms, molecules boiling off the sides. You don’t want anything but the K mesons. Then you move
your K meson detector a certain distance down the beam, and see how the population of K mesons falls off as you
move the detector further and further down the beam. And if you are lucky, the curve you will plot will have a nice
exponential in it. By looking at the parameters of that exponential you will see the decay law.

Now that’s a rather complicated experiment. Let’s idealize it a bit. First we want something that makes things
localized in a certain region of space and time and also localized in a certain region of momentum space. We’ll use
the same sort of production process as in the last discussion to do that, writing

That’s not a very realistic production process; people do not have things coupled to scalar fields in general, but it
is one which we can manipulate very simply analytically. I want the function f(x) to have two properties. I want f(x)
to be reasonably well localized in space and time, at the origin, so I know that I’ve made the particle at the target
and not someplace else. And I also want its Fourier transform (k) to be reasonably well localized in k space, at
some mean value of k I’ll call k. These properties are represented symbolically in Figure 17.3.

We can imagine f(x) as a Gaussian in all four variables, although that won’t be relevant. I want to make sure
I’m looking at K mesons, and not two pion states or a fragment of the target or something like that coming out. So
I’ll arrange matters so that k2 is near µ2. In particular I’ll assume that (k) is sufficiently sharp in momentum
space that throughout the relevant region of momentum for this production process, I can use the approximation
(17.18) for the propagator. That’s my single assumption. That may mean I have to have a fast target in position
space, but we’ll see that’s no problem; I can certainly always do that. One last significant point: f(x) is real, which
of course implies that
 Figure 17.3: The function f(x) and its Fourier transform (k)

I’ll make my state by taking the vacuum and hitting it with some source that makes a bunch of stuff:

That’s the production apparatus. I have produced this state, maybe with some tiny coefficient, mainly vacuum, but
vacuum I won’t detect. Now I want to detect the state. What about the detection apparatus? As a theorist, I’m very
economical. Instead of inventing new detection apparatus, I’ll use the same thing as for production. After all, this
formula (17.28) tells me that if (k) produces a given amplitude for making K mesons of momentum k, it produces
an amplitude of the same magnitude for absorbing K mesons of the same momentum, k. So I’ll just move myself
down the beam to a spacetime point y, and set up a detection apparatus. The amplitude A(y) I wish to study is
therefore

Perhaps a spacetime diagram would be useful. Figure 17.4 shows a light cone and two spacetime regions (the
shaded circles). In the neighborhood of the origin, there is a region in which f(x) is concentrated and where the
kaons will be produced.

Figure 17.4: Spacetime diagram for production and detection of K mesons

Some huge, timelike distance away at a point y is a second region where f(y) is sufficiently different from zero,
and where we’ll detect the kaons. We want y to be so far away that these regions have no chance of overlapping.
That’s the experimental setup. I have only one free variable in the problem, y: how far down the beam in space
and time can I locate my detector. And I want to see the dependence of this amplitude (17.30) on y. Now of course
this is something that I can compute in a moment in terms of the Fourier transform of f(x) and the two-point
function á0|ϕ′(x′)ϕ′(x)|0ñ, which I also presumably know. Since y is far later in time than x, we can time-order this
expectation value with negligible error. The amplitude becomes

Writing

(I can insert this approximation for the propagator legitimately because I assume (k) is concentrated near k), the
evaluation of A(y) is just a Fourier transform operation. We obtain, as you might have guessed,

Suppose we knew quantum mechanics, but didn’t know anything about quantum field theory. If we were told
about this experiment—we had a production apparatus that produced particles in a restricted range of momentum
and a detection apparatus that only detected particles in the restricted range of momentum—what would be our
naive guess for the asymptotic properties of this expression as y2 goes to ∞? Let’s approximate the kaon’s proper
time s0 by

What would we expect the amplitude to look like as a function of s0? Well, as an experimenter I would say I’d put in
a big fat proportionality sign because I got this creation and detection apparatus from a theorist’s workshop, so I
don’t know its properties or its resolution or anything like that. Then, the particle is traveling at proper time s0 and it
has mass µ, so I’d expect there to be a phase factor from the good old Schrödinger equation. But it’s an unstable
particle and it decays. I expect an exponential decay from the square of the amplitude, so half that magnitude in
the amplitude itself. And if I’m a very sophisticated experimentalist I know from my studies of the Schrödinger
equation that wave packets tend to spread out for large time t such that the amplitude goes down like t−3/2. When
y2 is very far down beam, portions of the decay products miss the detector. So I would insert the factor of s0−3/2 in
the amplitude to account for the spreading out of that wave packet. Then our naive guess looks something like
this:

This is a dumb guess, based on the picture that I am painting of an unstable particle, traveling around like an
ordinary particle, developing a phase, and spreading out. But it’s got this little extra feature: it decays. But what is
the analytical result? I will show you that the asymptotic form for large y2 of the amplitude (17.33) has exactly the
form of our naive guess, (17.35). That requires some analysis. I will use one analytical tool and one trick. The
analytical tool is the method of stationary phase.5

17.4Obtaining the decay law by stationary phase approximation

If we have an integral

with a real function θ(t) that varies rapidly compared to the rate at which g(t) varies, then in general the value of
this integral will be zilch—nothing. Because θ(t) is oscillating rapidly, the exponential averages out to zero. The
main contribution to the integral comes at points of stationary phase, where the derivative of θ(t) is zero. At such
points θ(t) is not rapidly varying; it doesn’t vary at all. Therefore the integral is dominated by stationary phase
points. I will assume there is only one such point, t0;

If there are several such points, you get a sum. People normally like to phrase this principle by putting a
parameter λ in front of θ(t) and say “we’re studying large λ.” That’s neither here nor there. There may or may not
be an adjustable parameter in the theory. It’s just that if θ(t) is varying very rapidly, and g(t) is not, this is a good
approximation. We therefore approximate the integral by its value near the stationary phase point. By the
stationary phase approximation

The integral is trivial: it is a complex version of a Gaussian, and it gives us

If θ″(t0) = 0, we have to think again. We have to expand out to the quartic term, or cubic, whichever is the first non-
vanishing one. We will apply this method to our amplitude, (17.33).

Our integral is of stationary phase form because we have a complex exponential with argument k · y, and all
four components of y are getting huge. So we have four integrals we can do by stationary phase. But before we
can do that, we need to use a trick. We have a problem with the propagator in (17.33). We can use the
approximation (17.18) only because we’re also near k2 = µ2, and therefore the phase of the denominator is also
changing rapidly over the region of integration as we pass by the pole. It changes by 180°, very rapidly if Γ is very
small. We certainly don’t want to find an approximation that’s good only for y2 = s0 very large compared to Γ−1,
because then we’ll properly get zero for the value of the integral. We’d like to put the phase variation of the
propagator into a form where we can treat it also by stationary phase. That’s the trick. We write the propagator as
an exponential integral:
The reason for the scaling of the integration variable s by 2µ will become clear later. This turns my four-
dimensional integral into a five-dimensional integral, but what I gained from that is putting the propagator up in the
exponential, where I can treat its phase variation by the stationary phase formula.

Using this trick, the amplitude (17.33) becomes

Now the first step is to do the four k integrals by stationary phase. There are just two phase factors that involve k,
the product k · y and the quadratic term from the propagator, so θ(k) = −k · y + (sk2/2µ). One finds easily kα 0 =
(µ/s)yα , so θ(k0) = −(µy2/2s). Each of the four k integrals gives a factor

and turns one component of k into the corresponding component of k0= (µ/s)y, in | (k)|2 and in the exponent.
Carrying out all four integrals, we have

That does the first stationary phase integral. Note the interpretation of k = (µ/s)y. If you classically propagate a
stable particle with 4-velocity vα = kα /µ, in a proper time s, it will arrive at a point yα = vα s. Since vα vα = 1, it follows
s= . This is just classical kinematics, but you see we have recovered it in the limit of large y from quantum
field theory. Here, the conditions of stationary phase give an equation from classical mechanics.

Now we’re ready to do the s integral, also by stationary phase. The phase is rapidly varying because it has
this gigantic factor y2 in it. We find easily s0 = , θ(s0) = µ , and θ″(s0) = (µ/s0). Note that there is no
stationary phase point if y2 is spacelike, because as y → ∞, y2 → −∞, and there is no probability that a particle will
be detected. I plug into (17.42), and evaluate everything at s0 = . Note that now k0
= µ(y/s0), and

It looks like our dumb guess (17.35) was not so dumb after all. In the amplitude we see a number of factors
common to the dumb guess. The square of the Fourier transform represents the factors that depend on the details
of the experimental apparatus producing and detecting our unstable particles. There is a common factor, e−iµs 0,
giving the evolution of phase as this particle of mass µ marches along in time. There is the common exponential
decay factor, and there is the common factor of one over s03/2, the spreading of the wave packet. I hope you have
understood the physical import of what we have obtained. We have derived the exponential decay law. The
statement that the propagator in a certain region of k space can be approximated by the expression (17.18) is
completely equivalent to the statement that under the physical circumstances in which we would expect to observe
an exponential decay law, we do observe an exponential decay.

This is the cleanest derivation of the exponential decay law: the stationary phase approximation to the pole on
an imaginary sheet. There are other derivations in the literature that are wrong. If you just Fourier transform the
original expression for the amplitude A(y), you don’t take into account the momentum cuts in the detection
apparatus. There’s a famous false statement in the literature that the decay law is not strictly exponential. It is true
that there are satellite terms in the amplitude that are non-exponential. The interpretation of those satellite terms is
that they are experimental background. Experimentalists know about them, and they take account of them. The
exponential decay law is 100% valid.

Some people set up a thought experiment where they haven’t been as careful as I have been to put a good
momentum filter in at the beginning and at the end. If you do that kind of thought experiment then you get in fact a
very large contribution for making two π mesons, with the mass say half that of the K mesons. Then the
experimental apparatus you built up has a nice probability for detecting pairs of π mesons, as well as detecting K
mesons, because you haven’t got a sharp enough momentum filter, and then you get a mess. And doing the data
analysis, you may be led to say the exponential decay is just an approximation. To avoid that, you’ve got to do the
experiment so that you only get momentum near the Breit–Wigner peak. Then you suppress those unwanted
things enormously: uncorrelated π mesons are randomly distributed in phase space, more or less. If you don’t do
that, then you get something that looks very different, and there are papers in the literature by very bright people
many years ago, when this phenomenon was not so well understood, that said, “Hey, the exponential decay law
should not be true. There should be terms, for example, that go as inverse powers of the time.” That’s what
happens if you just Fourier transform the propagator without putting in these (k)’s. When you Fourier transform
the propagator, if you don’t put in this momentum spread, if you just have a sharp position experiment, then you
get an enormous contribution from the two π meson states because then you can make, in particular, two π
mesons on threshold. Two π mesons on threshold have small relative velocity and therefore do not spread very
much from each other. And if you work things out, you get a one over s to the sixth term coming in with a tiny
coefficient. And that’s just wrong. Even a physicist of the stature of Abdus Salam once thought, back in the 1950s,
that the decay was not purely exponential.6 That’s the threshold singularity he was seeing, not the pole on the
second sheet. So there’s an error in the literature.

This concludes everything we are going to say in a world in which there are only scalar particles. Next time we
will begin studying particles with spin.

1[Eds.] Problem 6.1, p. 261.

2[Eds.]
The Yiddish word nu has a bewilderingly large number of meanings. In this context, it means “So what?”
Rosten says that it is the word most frequently used in Yiddish, besides oy and the articles: Rosten Joys, pp.
271–272.
3[Eds.] Problem 6.1, p. 261. See also Figure 12.3, p. 251 and (12.30).
4[Eds.] Gregory Breit and Eugene Wigner, “Capture of Slow Neutrons”, Phys. Rev. 49 (1936) 519–531.
5[Eds.]See Section 8.2, pp. 229–234 of G. N. Watson, A Treatise on Bessel Functions, 2nd ed., Cambridge U.
Press, 1966, and Sections 17.03–17.05, pp. 471–474, of Harold & Bertha S. Jeffreys, Methods of Mathematical
Physics, Cambridge U. Press, 1946.
6[Eds.] P. T. Matthews and Abdus Salam, “Relativistic Theory of Unstable Particles. II.”, Phys. Rev. 115 (1959)
1079–1084. See Section 5, and equation (5.4).

18
Representations of the Lorentz Group

We will now put aside for a while those questions of Green’s functions and factors of i and k2 − µ2 that drove us
crazy. We’ll come back to them eventually, and generalize them to the case of spin one-half particles. Now we are
going to look at a topic that has nothing to do with quantum field theory, but a lot to do with Lorentz
transformations. We are going to construct the quantum theory of spin one-half particles and the Dirac equation. I
do not wish to start out by saying, “Well, you all know the Dirac equation”, and start covering the boards with
gamma matrices. Instead, I want to derive the Dirac equation as a classical field theory, and then canonically
quantize it by our standard machine to find out what’s going on. Part of this discussion can be done in some
generality. The general discussion will be useful for subsequent purposes and will also give us additional insight
into the structure of the Dirac equation.

18.1Defining the problem: Lorentz transformations in general

I will begin by asking what are the most general possible transformation properties of a finite number of fields
under the Lorentz group, assuming they transform linearly; they just shuffle among themselves. Let Λ be an
element of the Lorentz group, which is called by its friends SO(3, 1).1 Say I have a set of fields ϕ a(x), which
transform under the Lorentz group according to the rule

(a = 1, 2, . . . , n; the sum on b is implied). What are the possible choices for the matrices D ab? We know there are
many choices. The fields ϕ a could be Lorentz scalars, for which D ab is the identity matrix. There are vector fields
typified for us by the derivatives of scalar fields. For these fields, the matrices D ab are the 4 × 4 Lorentz matrices
Λµν themselves. There are tensor fields where D ab are products of a bunch of those Lorentz matrices, one for
each index, which are here all summed up in the super-index a, but what else is there? What are the possibilities?
I will explore the constraints placed on the matrix D ab by what we know about the matrix U(Λ). In order to keep
from writing indices when I don’t really need to, I’ll assemble ϕ into a big (column) vector and simply write (18.1)
as

where ϕ is some n component vector, and D is some n × n matrix.

The transformations U are constrained. If I have two Lorentz transformations, U(Λ1) and U(Λ2), then as I said
much earlier (see (1.62)), U of the product should be the product of the U’s for the individual ones:

Actually this isn’t quite right. It’s impossible to rule out in quantum mechanics that (18.3) might need to be
generalized to

We know it’s not quite right even if only rotations are considered, let alone the full Lorentz group. It turns out that
for the rotation group in three dimensions, SO(3), and for the Lorentz group, the phases can be removed except in
spinor representations, where a rotation by π about any axis followed by a second such rotation results in a
net multiplication by −1. I won’t bother to write down the general definition of a ray representation, as it is called.2
Spinor representations are used to describe spin-½ particles, and so we expect minus signs if there are spin-½
particles in the theory. Since God put spin-½ particles into the world, we must allow the occasional minus sign if
we want to describe reality. We’re going to be a little bit sloppier with spin-½ than we have been with Bose
particles.

From (18.2) and (18.3) we obtain a constraint on D. It goes like this:

because the inverse of the product is the product of the inverses in the reverse order. Now let’s write out the same
thing using the product equation (18.3):

D is just a numerical matrix, U is an operator in Hilbert space; they have nothing to do with each other, and they
commute. By inspection we obtain

And again I tell you that strictly speaking, we are working with a looser condition than this, and occasional minus
signs are also okay in this equation. It follows, if we let both Λ1 and Λ2 be the identity matrix, that

The representation of a group is a set of matrices, one associated with each group element, that obeys the
same algebra as the group elements they represent:

(If we were considering the ordinary rotation group or the 17-dimensional rotation group or the discrete group that
describes the symmetries of a crystal, we would have the same equation (18.7) with Λ replaced by the appropriate
symbol labeling the transformation in question.) It is also easy to demonstrate that

Thus the matrices D form a finite dimensional representation of the Lorentz group. The D matrices obey all the
properties of their corresponding group elements, and you might reasonably think that from any set of D’s you
could reconstruct the group. But that’s not necessarily so. Many of the group elements can be mapped into a
single D, so that D(Λ) = D(Λ′) even if Λ ≠ Λ′. The trivial prototypical example is to assign D(Λ) = 1 for all elements
Λ. On the other hand, if D(Λ) = D(Λ′) only when Λ = Λ′, the representation is said to be faithful.

Our problem of finding all possible linear field transformation laws, involving only a finite set of fields and
consistent with Lorentz invariance, is equivalent to finding all finite dimensional matrix representations of SO(3, 1)
satisfying the equations (18.7) and (18.8). That makes it sound like a very difficult problem. But as we’ll see, it’s a
very easy problem. Once we have found these D representations, we can use them to construct field
transformation laws, which, from now on, we’ll think of not as being laws of the quantum theory, but laws for the
transformation properties of a classical field before quantization. From these possible laws we’ll then select out
some particularly tasty looking fields with not too many components, capable of describing spin-½ particles. We’ll
attempt to construct quadratic Lagrangians out of them, so we’ll get free field theories, and then try to develop a
theory of free spin-½ particles. Please notice that I want the D matrices to be finite dimensional, but I’m not going
to impose the constraint3 that the D’s be unitary (D † = D –1). The U’s, of course, have to be unitary, but that doesn’t
necessarily mean the D’s are.

For example, the 3-vector representation of the (3-dimensional) rotation group, D(R) = R, is unitary. Consider
a rotation about the z axis through an angle θ:

On the other hand, the 4-vector representation of the Lorentz group, D(Λ) = Λ, is not unitary. Consider a boost
along the x-axis by a speed v (as usual, γ = (1 − β2)–1/2, and β = v):

so the representation D(Λ) = Λ is not unitary, even though U is a unitary operator down there in Hilbert space. So
do not assume that we are looking only for unitary representations.

To find all matrix representations D obeying these equations is to answer a big question. It can be replaced
by a smaller question, because there are two trivial ways of obtaining new representations from old. One is this. If
D(Λ) is a representation, so is

for any fixed matrix T, because it doesn’t affect anything in the multiplication. If D(Λ1)(DΛ2) = D(Λ1Λ2), the same is
true for the transformed representations D(Λ)′:

If we have two representations related in this way, we write

and say that D(Λ) is equivalent to D(Λ)′. This just corresponds to choosing different linear combinations of the
ϕ a’s as the fundamental fields. We can generate an infinite number of new, equivalent representations from old
ones. But it’s trivial. We will restrict our problem to finding all finite dimensional, inequivalent representations of
SO(3, 1).

There is a second, trivial way of making new representations from old. Suppose I have two representations,
D (1)(Λ) and D (2)(Λ), of dimensions n1 and n2, respectively. The dimension n describes both the number of fields
involved and the size of the matrices, n × n. I can build a new representation in the following way. I make a great
big matrix

This matrix is called the direct sum of D (1) and D (2), and denoted D (1)(Λ) ⊕ D (2)(Λ). This, too, is a representation.
When I multiply these things together, D (1) and D (2) never talk to each other; they just multiply independently. The
dimension of this representation is the sum of the dimensions of the component representations:

I’m not interested in representations that can be written as direct sums. If I tell you I have a field theory that’s
Lorentz invariant with a scalar field, and I have another Lorentz invariant field theory with a vector field, it would not
surprise you that I can build a Lorentz invariant field theory that has five fields in it, one scalar and the four
components of the vector. If a representation D can be written as a direct sum of two representations of smaller
dimensions, or is equivalent to a direct sum, we say D(Λ) is reducible. If it is not reducible, then we say, to no
one’s surprise, that it is irreducible. Our task of finding all possible Lorentz transformation laws of fields has thus
been reduced to the task of making a catalog of all inequivalent, irreducible finite dimensional representations of
SO(3, 1).

Now this is a problem that was solved for the rotation group SO(3) many years ago. It is part of the standard
lore of quantum mechanics, though perhaps not in this language. Every quantum mechanics course has a section
in it devoted to the subject of angular momentum, and there you saw this problem solved, although, like the man in
Molière’s play who didn’t know he was speaking prose all his life,4 you may not have known that you were in fact
finding the irreducible inequivalent representations of the rotation group. By a wonderful fluke peculiar to living in
(3 + 1) dimensions, the representations of SO(3, 1) can be obtained rapidly from the representations of SO(3).

18.2Irreducible representations of the rotation group

Let’s now consider the related problem, finding all inequivalent irreducible representations for the rotation group.5
SO(3) is the group of rotations in space (or, as the mathematicians would write, R 3) about some axis by some
angle. Every rotation matrix R can be labeled by an axis,6 , and an angle, θ:

Notice that the angle and the axis appear as a product. By convention, the angle is always chosen to be less than
or equal to π; if you rotate by more than π, that’s equivalent to rotating by the supplementary angle about the
opposite axis. We will use the multiplication rules of the rotation group to gain information about the
representations, D.

First, observe that if you have two rotations about the same axis with two different angles, the angles simply
add:

So any representation, not necessarily irreducible, must satisfy

Let’s differentiate this equation. (As usual I’m being a mathematical slob, and will assume everything is
differentiable.) I will define

The derivative of the representation evaluated at θ = 0 must be some linear function of . This defines a vector of
“angular momentum” matrices L, sometimes called the generators of infinitesimal rotations. If I differentiate
(18.20) with respect to θ′ and set θ′ equal to zero I obtain 7

This differential equation is trivial to solve,8 using the “initial condition” D(R(0)) = 1:

We’ve simplified our problem enormously. We don’t have to work out D(R( θ)) for all axes and all angles θ. We
just have to tabulate the three “angular momentum” matrices {Li}, i = {1, 2, 3}. (There are 3 generators because
SO(3) is a three parameter group. In general, the group SO(n) is described with n(n − 1) parameters.)
 Of course our concept of equivalence and reducibility also apply here. Two representations are equivalent if
and only if the two L’s are equivalent:

(in the sense of (18.13)). If the representation is a direct sum, so too are the L’s:

So as far as checking for reducibility and equivalence, we might as well work with the L’s as with the D’s.

Let’s work out the algebra of the matrices {Li}. The transformation of a vector v under an infinitesimal rotation
by θ about an axis is given by9

Moreover, the operators L transform as a vector:

so for an infinitesimal transformation

Equating terms of O(θ) gives

Letting be , or , we obtain the famous angular momentum commutation relations

sum on k implied. The generators L are said to form a representation of the Lie algebra of the rotation group; the
D’s form a representation of the Lie group.10 Any finite dimensional set of matrices that form a representation of
the rotation group necessarily lead to a triplet of finite dimensional matrices that obey the angular momentum
commutation rules. Thus if we can find, up to equivalence and direct sum, all matrices that obey these
commutation relations, we will have all representations of the rotation group. (We might find some things that
aren’t representations. I won’t take the time to show you that the process is reversible.)

This problem was solved by Pauli.11 Irreducible representations D (s)(R) are labeled by an index s called the
spin:

where L(s) is a triplet of matrices appropriate to the spin s. Let me recall a number of well-known facts about these
matrices. The index s equals {0, , 1, . . . }, etc. The dimension of the representation D (s)(R) is 2s + 1. The square
of L(s) is a multiple of the corresponding identity matrix, I:

It is convenient to label eigenstates by eigenvalues m of Lz = L3. I’ll now switch to Dirac notation even though I’ve
only got a finite dimensional space:

The eigenvalue m takes as many values as the representation’s dimension, 2s + 1. The first few matrices L(s) are:

(Note that the bold type for σ is doing double duty: for the vector nature of the triplet of sigmas, and also to remind
you that each sigma is a 2 × 2 matrix.) For larger values of m, you can find worked out, in nearly every quantum
mechanics textbook, the explicit matrix elements of Lx , Ly and Lz in this m basis. We can always choose our basis
such that these matrices are Hermitian:
This is not surprising, as the eigenvalues are observables. So the representation matrices D (s) are unitary. This is
a special feature of the rotation group, or indeed of any compact group, any group of finite “volume”. It is not true of
the Lorentz group, as we’ll see. Finally, the analog of (18.7) is true for the integer values of s, but true only to within
a phase for half-integer values of s: they are double-valued. For any rotation,

We should expect then that D (s)(R(2π )) = D (s)(R(0)) = 1. However, it turns out

The representation is only good to within a factor of −1 for half-integer values of s. The double-valued character of
the half-integer representations will not prevent our using them for physical purposes.

18.3Irreducible representations of the Lorentz group

I will now go through this whole routine for the Lorentz group. You might expect this will take a substantial
investment of time and effort. It is not so, by a fluke which we will soon discover. One subgroup of the Lorentz
group is of course the rotation group. By an abuse of notation, I will indicate these rotations by the symbol R even
though R is no longer a 3 × 3 matrix, but now a 4 × 4 matrix acting trivially on the time components of 4-vectors.
Another subgroup of the Lorentz group concerns pure accelerations, or boosts. A boost A( ϕ) along a given axis
and velocity parameter ϕ (called the rapidity) is a pure Lorentz transformation that takes a particle at rest and
changes its velocity to some new value along that axis.12 For example, a boost A( ϕ) along the z = x3 direction by
an amount ϕ is defined as

while x1 and x2 are unchanged. The hyperbolic tangent of the rapidity ϕ is the new speed;13

This is easy to see by considering x′3 to be the z component of the primed frame’s origin. Then

It’s standard special relativity lore that every Lorentz transformation can be written as a product of a rotation and
an acceleration. If we know the representation matrices for the rotations and the accelerations, we know them for
everything. As with the rotations, we have defined things with this angle ϕ so that two successive boosts by
different hyperbolic angles ϕ and ϕ′ along the same axis give a combined boost along the same axis:

Thus we can treat the rotations as we treated them before, and the accelerations in exactly the same way as the
rotations, simply replacing R’s by A’s at appropriate points.

As before (18.21), define

and analogously

The {Mi} will generate the boosts just as the {Li} generate the rotations. We find, with the initial conditions that
D(R(0)) = D(A(0)) = 1, that

The next step is to figure out all the commutators of L and M. If we know L and M we know the representation
matrix for an arbitrary rotation and an arbitrary boost, and by multiplication, we can find the representation matrix
for any general Lorentz transformation. I won’t compute the commutators for you, but I’ll write them down and try
to make them plausible. For the rotation generators,

That of course is no news; these commutators are the same as (18.30) because the rotations are a subgroup.

This is not a big surprise; it’s just the statement that M transforms like a vector under infinitesimal rotations, just
like L. Both (18.46) and (18.47) can be shown with the same method as (18.30). We also have

the minus sign from swapping i and j is compensated for by the minus sign from exchanging the two terms in the
commutators. The only one you have to work at is

Because {Mi} transform as a 3-vector, the method used to derive (18.30) fails to produce (18.49), and I leave this
as an exercise.14 The minus sign in (18.49) is important. If we were doing the four-dimensional rotation group
SO(4), rather than SO(3, 1), we could’ve made almost the same definitions with sinh’s and cosh’s replaced by
sines and cosines, and then we would have gotten a plus sign in this last commutator.

To show you that the commutators are at least self-consistent, let me remark that if the theory we are studying
has not only Lorentz invariance but also parity invariance—it need not, of course—then we can figure out how L
and M transform under parity. Parity commutes with rotations. And therefore

On the other hand, parity switches the sign of a boost, because it transforms a velocity to its opposite. So M goes
to minus M:

Please notice that these commutators are consistent with that, because they are unchanged by the replacements
L into L and M into −M: (18.46) is totally unchanged; (18.47) gets a − sign on both the right- and left-hand side;
and (18.49) gets two minus signs on the left-hand side and no change on the right-hand side.

I will now find, in a very few lines, all the irreducible representations of the Lorentz algebra. It’s based on a
special trick. If we were unfortunate enough to live in five-dimensional space, the trick wouldn’t work. But
fortunately we live in four-dimensional spacetime and the trick works. I define operators analogous to the raising
and lowering operators you’ll recall from quantum mechanics,

Let us compute the commutation rules for J(+) and J(–):

The same result is obtained with (−) instead of (+) in both places:

What about Ji(+) with Jj(–)? We find

Thus {Ji(+)} and {Ji(–)}commute. We have reduced this apparently formidable algebra into two commuting angular
momentum algebras. Exactly this problem arises in ordinary non-relativistic quantum mechanics, where we have
both orbital and spin angular momentum, each of which obey the commutation rules of the rotation group, but
which commute with each other.

It is now a snap to write down a complete set of irreducible, inequivalent representations of the Lorentz group.
They are characterized by two independent spin quantum numbers, s+ and s–, one each for J(+) and J(–),
respectively, and are written as

The squares of these operators J(+) and J(–) are multiples of the identity:

The complete set of basis states is described by two numbers, m+ and m–, eigenvalues of Jz (+) and Jz (–),
respectively:

The states |m+, m–ñ are simultaneous eigenstates of the commuting operators Jz (+) and Jz (–). The eigenvalues m+
run from −s+, −s+ + 1, . . . , s+, and the eigenvalues m– run from −s–, −s– + 1, . . . , s–. Hence

The dimension of D (s +, s –)(Λ) is also the number of basic vectors. To make things more explicit, consider the matrix
element

J(+) has nothing to do with m–, so I simply get δm–,m′– times the matrix element ám′+|J(+)|m+ñ, full of square roots,
which you will find in any elementary quantum mechanics book. The same equation holds if the plus and minus
signs are swapped. We have two commuting “angular momenta”, so there’s no problem in finding all the
irreducible, inequivalent (finite dimensional) representations of SO(3, 1).

We can always choose things so that J(+) and J(–) are Hermitian matrices. L, the sum of these, is indeed
Hermitian, so the representations D(R( θ)) are unitary:

The same is not true of M which is −i times the difference of J(+) and J(+). So M is an anti-Hermitian matrix, and
consequently the representations D(A( ϕ)) are not unitary:15

18.4Properties of the SO(3) representations D(s)

Now that we have all of the representations of SO(3, 1), we would like to know their properties. We can deduce a
list of properties just by knowing some elementary facts about the rotation group, SO(3). From these I will derive
properties of the representations of SO(3, 1).

Complex conjugation

If I complex conjugate (this is not to be confused with taking the Hermitian adjoint) a representation of SO(3),
or in fact of any group, I again obtain a representation

because the product of two complex conjugated matrices is the complex conjugate of the product. Since there’s
only one irreducible representation of a given dimension, the complex conjugate must be equivalent to D (s)(R):

That is, for some matrix T

and therefore we must have

This doesn’t necessarily mean we can write the J’s as imaginary matrices. It just means that there is some
transformation T such that

(the same T, of course, for all three Ji’s for a given s).

Direct product

If we have a set of fields that transform under a rotation as an irreducible representation D (s 1)(R), a vector, a
spinor or something, and if we have a second set of fields that transform as some other irreducible representation
D (s 2)(R), we can consider all products of components of the two fields. This defines a brand new representation of
the group called the direct product, denoted by

The dimension of the direct product is of course the product of the dimensions of the two representations.
Because you’re multiplying two things together, you have two indices to play with:

This product is certainly a representation. But it’s usually not an irreducible representation. There is a rule for
finding how it breaks up into irreducible representations. It’s equivalent to a direct sum which I will indicate this
way,16

The quantity on the right is a direct sum over s, as in (18.16), not a numerical sum. s goes from |s1 − s2| to s1 + s2
by unit integer steps. This is the so-called rule for addition of angular momentum written in slightly more
sophisticated language, and you should be familiar with it. Thus for example if I multiply together D ( ) times D ( ),
the product of two spinors gives four objects, and I obtain a D (0) ⊕ D (1), a scalar and a vector, a one-dimensional
object and a three-dimensional object.

Exchange symmetry

There’s also a sub-item we can add for the direct product. If s1 = s2, then it’s a sensible question to ask what
happens when you exchange them, since they transform in the same way. If s1 = s2, then D (2s 1), the irreducible
representation of highest spin, is symmetric under exchange. That is probably familiar to you, but if not, it can be
found in many quantum mechanics texts. Then D (2s 1−1) is antisymmetric under exchange, etc. These are three
facts about the rotation group. I presume you’ve seen them before, though perhaps in different language. If they
seem new, you may be suffering merely from linguistic impedance matching.

18.5Properties of the SO(3, 1) representations D(s+, s–)

I will now take what we know about the representations D (s)(R) to discuss seven questions about the properties of
the SO(3, 1) representations D (s +, s –)(Λ). The discussions of these questions will be very brief, because we know
the answers, we’ve just got to put things together and keep track of factors like i’s.17

Complex conjugation

The equivalence (18.67) between a representation of SO(3) and its complex conjugate doesn’t quite work for
SO(3, 1). J(+) and J(–) are ordinary rotation matrices, and for any particular value of s, they are equivalent to minus
their conjugates by (18.70). Therefore L, which is their sum, is equivalent to −L*:

But M, − i times the difference of J(+) and J(–), is equivalent to +M*, because of the intervening i:
D (s)(R) is equivalent to its complex conjugate because of the sign change of the generators J. Here, the disgusting
lack of sign change in M prevents D (s +, s –) from being equivalent to its conjugate. We can introduce a sign change
in the right place if we exchange J(+) and J(–). This will not change the sign of L, but it does change the sign of M.
Therefore we deduce

D (s +, s –)(Λ) is equivalent under complex conjugation to D (s –, s +)(Λ). The effects of complex conjugation can be
canceled out up to an equivalence transformation by exchanging J(+) and J(–). Notice that there is some funny
business going on. If I have a set of fields, and they transform in a certain way, their complex conjugates do not
transform in the same way unless s+ is equal to s–.

Parity

Recall that parity turns L into L and M into −M. The operation that turns M into −M again can be thought of as
exchanging J(+) and J(–). Equivalently we could say, “Parity exchanges J(+) and J(–).” Thus if we wish to have a
parity-conserving theory involving only fields that transform according to the representation ( , ) we have the
chance of the proverbial snowball in hell: Parity acting on a field that transforms like D (s +, s –)(Λ) must turn it into
field that transforms like D (s –, s +)(Λ):

On the other hand, parity plus complex conjugation turns a field into one that transforms in the same way. We will
see later on that this property will make it easy for us to construct theories that are CP invariant, but neither C
invariant nor P invariant, a nice thing for weak interaction theory. Onward!

Direct product

We’ve got two independent angular momenta. We add them together independently. There’s no problem:

where s+ goes by unit steps from |s+1 − s+2| to s+1 + s+2 and s– independently does the same, between |s–1 − s–2|
and s–1 + s–2. Here are two angular momenta that don’t talk with each other. Add them together, and they still
don’t talk with each other.

Exchange symmetry

Exchange symmetry is a reasonable topic only if two representations are of the same spin, just as before: s±1
= s±2. Well if you exchange ’em, you exchange both the s+ and the s– parts, so it’s symmetric if the two parts are
individually symmetric or if the two parts are individually antisymmetric, and antisymmetric otherwise. Thus

because it’s antisymmetric in the first variable and symmetric in the second; likewise s+ = 2s+1, s– = 2s–1 − 1 is
antisymmetric, etc; and

because it’s antisymmetric in both variables.

The rotation subgroup of the Lorentz group

What happens when I look at just the rotations, at the SO(3) subgroup of the Lorentz group SO(3, 1)? Any
representation of a big group is a representation of any subgroup, but if it’s an irreducible representation of the big
group, it might not be an irreducible representation of the subgroup. Well,
Thus if we just restrict ourselves to rotations, we can think of J(+) and J(–) as being like orbital angular momentum
and spin angular momentum—it’s as if we have coupled orbital angular momentum and spin angular momentum,
and only consider the combined rotation group: simultaneous spin and orbital rotations by the same angle. This is
just our direct product formula again, so I have

We’ll see some examples in a moment.

How are vectors represented?

Where in our representations will we find a vector field like Vµ? A vector field transforms according to some
representation of the Lorentz group. That representation is pretty obviously irreducible, so it must be somewhere
in our catalog. What do we know about a vector? First, we know it’s got four components, so the representation is
four-dimensional:

Since both of these factors are integers, there are not many solutions. To be precise we have three possible
solutions. First, we could have s+ = , s– = 0. That gives a product of 4 × 1. But it’s obviously no good because it is
not equivalent to its complex conjugate, whereas a vector representation is certainly equivalent to its complex
conjugate; we’d need to have s+ = s–. This representation also does not admit a parity (again, we’d need s+ = s–)
and a vector certainly does. So this representation fails on two counts. And the representation s+ = 0, s– = is also
ruled out.

Finally, we have s+ = s– = . This is the only possibility, and as Sherlock Holmes used to say, therefore it is the
right answer.18 So a vector field transforms according to the four-dimensional irreducible representation D ( , ).
Let’s check that, by using our previous result, (18.82):

The direct sum goes from | − | to ( + ) by integer steps, so there are only these two. The first, D (1)(R), is a
spatial 3-vector, and the second is a scalar, a single number that doesn’t transform under rotations. Is this indeed
what happens to a Lorentz 4-vector when we restrict ourselves to rotations? It certainly is: the time component is
unaffected by rotations, and the three space components transform like a 3-vector. So it all holds together; it
checks.

How are tensors represented?

Once we have vectors, we can construct tensors, because tensors are direct products of vectors. (I’ll only talk
about rank 2 tensors.) Where are the tensors in our classification of representations of the Lorentz group? We can
find them if we think about their properties. With our formula (18.78), we can figure out how rank two tensors like
Tµν transform. It doesn’t matter whether I write upper or lower indices, of course. That’s just an equivalence
transformation, with the metric gµν as the matrix that effects the equivalence transformation. There is a basis of all
two index tensors, Tµν, for a 16-dimensional representation of the Lorentz group. If I take such a tensor and
Lorentz transform it in the standard way I get 16 linearly independent objects that shuffle among themselves
according to the Lorentz transformation I have made. The transformation of Tµν defines some 16 × 16 matrix
representation D(Λ):

Its form depends on how I choose the basis for the 16-dimensional space of tensors. I want to find out what it is in
terms of irreducible representations. A tensor is an object that transforms like the product of two vectors. So

Let’s check our dimensions. The dimension of D ( , )(Λ) is 4, and 4 × 4 = 16, as required. The direct product is
given by our product algorithm (18.78). For the rotation group, one half and one half gives you zero and one. Here
we’re doing two such sums independently and getting all possible combinations.
 Now let’s check that this is right by adding up the dimensions. Using (18.62) we have

And indeed, 9 + 3 + 3 + 1 is 16. We also know how these things transform under permutation of the indices s±i, i =
{1, 2}. If we think of this D (s +, s –) as in (18.78), then from (18.79), D (1,1)(Λ) is symmetric under the exchange (1 ⇆
2), and the representations D (1,0)(Λ) and D (1,0)(Λ) are antisymmetric. Likewise, in agreement with (18.80), the
representation D (0,0)(Λ) is symmetric because it’s antisymmetric in both the indices. Thus the general theory of
representations of the Lorentz group says that we should be able to break the 16-dimensional space up into a
nine-dimensional subspace, two three-dimensional subspaces and a one-dimensional subspace. When we apply
the Lorentz transformation, a tensor constructed out of basis tensors in any one of these subspaces goes into a
tensor in the same subspace. Parts of the tensor in different subspaces don’t talk to each other under Lorentz
transformations; they each transform independently. That’s what the direct sum means.

Let’s try to figure out what this break-up is in traditional tensor language. Every rank 2 tensor Tµν can be
written unambiguously as the sum of a symmetric tensor, Sµν, and an antisymmetric tensor, Aµν:

with

Since the two indices {µ, ν} transform identically, symmetric tensors transform into symmetric tensors, and
antisymmetric tensors go into antisymmetric tensors under Lorentz transformations. So (18.89) is a Lorentz
invariant break-up. Thus I have written my representation as a direct sum, and the Lorentz transformation can be
written as a block diagonal matrix, with a part that acts on the space of symmetric tensors and a part that acts on
the space of antisymmetric tensors. How many linearly independent components does a symmetric tensor have?
For n × n matrices there are n(n + 1) symmetric elements. For n = 4, we have a ten-dimensional subspace. The
number of antisymmetric matrices fills a 16 − 10 = 6-dimensional subspace. Let’s check that with our algorithm.
We have two symmetric subspaces, the nine-dimensional representation D (1,1)(Λ) and the one-dimensional
representation D (0,0)(Λ). Then

The symmetric ten-dimensional subspace is written as a direct sum of a nine-dimensional subspace and a one-
dimensional subspace; the antisymmetric six-dimensional subspace is written as the direct sum of two three-
dimensional subspaces. So far, things are checking out.

Let’s now consider a symmetric tensor, Sµν. If we think of Sµν as a matrix, we can break it up into a traceless
part and a part proportional to the metric tensor gµν:

µν ≡ S µν − gµνSλλ is traceless, as you can quickly verify:

Thus we have broken up the ten-dimensional subspace of symmetric tensors into a nine-dimensional subspace of
traceless, symmetric tensors, and a one-dimensional subspace of symmetric tensors proportional to gµν. A tensor
proportional to gµν stays proportional to gµν under a Lorentz transformation, and if it’s traceless, it remains
traceless after the transformation, because these are Lorentz invariant equations. So we have block diagonalized
the representation.

The break-up of the antisymmetric tensor Aµν is a little trickier, because we normally don’t think of an
antisymmetric tensor as being the sum of two three-component objects. You’ve played with antisymmetric tensors
in electromagnetic theory, where the field vectors E and B combine19 to form an antisymmetric tensor Fµν. You
don’t think of Fµν as being broken up into the sum of two 3-component objects, each of which transforms only into
itself under the action of the Lorentz group, because that’s not true of E and B: they transform into each other. The
mathematical reason you don’t think of this division of Fµν is that the representations D (1,0)(Λ) and D (0,1)(Λ) are
not real; they’re complex conjugates of each other, as in (18.76). The breakup of the six-dimensional subspace
into two three-dimensional subspaces will in fact involve complex combinations of the components of the
antisymmetric tensor. I’ll demonstrate how that goes.

For any antisymmetric tensor Aµν, define its dual, *Aµν:

I’ve put in a factor because in such a sum over two antisymmetric tensors, there is always double counting. This
is a Lorentz invariant way of associating one antisymmetric tensor in a linear way with another. Just to see what
the dual looks like, consider a particular element of Aµν, say A01:

λ = 2 and σ = 3 and vice versa give the only non-zero combination; these are equal and you get A23. Lowering the
indices, A01 = − A01, so

Let’s do it again. What is the double dual? Find the double dual of A23:

because raising a pair of spatial indices does not change the sign of the tensor, and 2301 is an even permutation
of 0123, so ϵ2301 = +1. There is nothing special about the set of indices (0, 1) and (2, 3), so we find

Now the operation of forming a dual of a tensor obviously commutes with all Lorentz transformations, since ϵµναβ
does,20 and certainly lowering indices does. Therefore I have a linear operation, *, defined on the six-dimensional
space, with the property that its square is −1. I can form eigentensors of this operation, and the eigenvalues λ
must have the values ±i, since λ2 = −1. That is to say, I can write any Aµν as a linear combination of Aµν(+) and
Aµν(–), where

The tensors Aµν(±) are eigentensors of the dual operation:

Therefore we have these two kinds of objects, Aµν(+) and Aµν(–), each of which form a three-dimensional subspace
of the six-dimensional space of antisymmetric rank 2 tensors. They are course the representations D (1,0)(Λ) and
D (0,1)(Λ). I will not bother to work out which is which.

To summarize, a vector transforms according to representation ( , ); a scalar according to representation (0,

0); a traceless, symmetric tensor according to the representation (1, 1), an antisymmetric tensor according to the
reducible representation (1, 0) ⊕ (0, 1), which we can reduce if we are willing to form complex combinations.

Next time we will start building field theories from some of the simple representations that we have found
here, in particular D ( , 0) and D (0, ), which we need for the Dirac equation.

1 [Eds.] SO(3, 1) is the group of orthogonal transformations with determinant 1 which preserve the square of the
Minkowski norm,

2 Theonly good reference I know is Valentine Bargmann’s “On unitary ray representations of continuous groups”,
Ann. Math. 59 (1954) 1–46. I am not recommending that you study this article. We will get all the right results with
much less effort, by being cavalier and lucky.
3 [Eds.] Coleman could not impose this constraint even if he wanted to: the Lorentz group is non-compact, and
there are no faithful, finite-dimensional, unitary irreducible representations of non-compact groups. See Ashok Das
and Susumu Okubo, Lie Groups and Lie Algebras for Physicists, World Scientific, 2014, p. 47.
4 [Eds.] Coleman is referring to Monsieur Jourdain, the title character of Molière’s Le Bourgeois Gentilhomme,
1670. Jean-Baptiste Poquelin (1622–1673), known by his stage name Molière, is widely regarded as one of the
greatest French writers.
5 These are carefully constructed in a few pages, in a way that generalizes to other groups, beginning on p. 16 of
the first edition of Howard Georgi’s Lie Algebras in Particle Physics, Benjamin-Cummings (1982). Actually what
are constructed there are the representations of the Lie algebra of SO(3) rather than the Lie group, but you’ll see
that is what we want. ([Eds.] See also Chapter 3 of the second edition, Perseus Press, 1999.)
6 [Eds.] Coleman uses e for the axis. The notation was changed to avoid confusion with e, the base of natural logs.
7 [Eds.] = .
8 [Eds.] Different components of L do not commute, but i •L does commute with e−i •Lθ because [i •L, i •L] = 0.
9 [Eds.] Equation (18.26) is the limiting case for infinitesimal θ of Rodrigues’ formula ,

Alternatively, consider a rotation of x = (x, y, z) about the z axis though an infinitesimal angle θ:

which is the same as x → x + θ × x + (θ2).

10 [Eds.]The reader has likely encountered the concepts of Lie groups and Lie algebras in earlier courses. Briefly,
Lie groups are groups whose elements are analytic functions of one or more continuous parameters; every Lie
group thus contains an infinite number of elements. The most familiar is probably SO(2), the group of rotations in a
plane; each element is characterized by a single parameter, the angle through which the rotation is carried out. A
Lie group can be constructed by the exponentiation of a set of parameters multiplying a finite set of generators,
which among themselves satisfy the group’s Lie algebra. See §36.2, note 8, p. 782 and note 16, p. 784.
11 [Eds.]
Coleman is probably referring to Pauli’s paper, “Zur quantenmechanik des magnetischen elektrons”, (On
the quantum mechanics of magnetic electrons) Zeits. f. Phys. 43 (1927) 601–623, which introduces the Pauli
matrices. Reprinted in L. C. Biedenharn and H. van Dam, Quantum Theory of Angular Momentum, Academic
Press, 1965. English translation by David H. Delphenich online at http://neo-classical-physics.info/electromagnetism.html.
12 [Eds.]
In his lectures, Coleman used the same symbol, e, for both the axis of rotations and the axis of boosts. To
avoid possible confusion, the axis for a boost will be denoted by the unit vector .
13 [Eds.] The Lorentz group is non-compact because the parameter ϕ in (18.39) is unbounded, and the matrix
elements sinh and cosh increase monotonically with ϕ.
14 [Eds.] The boost equivalent to Rodrigues’ formula is

Let Mi generate a boost along the xi axis. Under an infinitesimal boost along , x′µ = xµ + iϕ µ xν +
ν (ϕ 2):

The matrices M2 and M3 are found in the same way. It follows easily that, e.g., [M1, M2] = −iL3. See Problem 10.2,
p. 387.
15 [Eds.] Because the Lorentz group is non-compact, the faithful, finite dimensional representations D(A( ϕ)) = e−i
•Mϕ cannot be unitary. See note 3 on p. 371.
16 [Eds.] Sometimes called “the Clebsch–Gordan series” in the literature.
17 [Eds.]Near this point in the videotape of Lecture 18, a student yawns loudly. Coleman responds: “Come on, you
can’t say it’s boring. It’s not boring. As Dr. Johnson said, in another context, ‘A man who is tired of group theory is
tired of life, sir.’ I made a killing. I get a good salary. It’s all done with group theory! We were even thinking of
advertising on matchbooks: ‘Learn how to make $20,000 a year through group theory!’ But then the job market
collapsed, so the whole scheme fell apart...” (Samuel Johnson (1709–1784), to his friend and biographer James
Boswell: “Sir, when a man is tired of London, he is tired of life.” Entry for September 20, 1777 in J. Boswell, The
Life of Samuel Johnson, LL.D., 1791.)
18 [Eds.] “How often have I said to you that when you have eliminated the impossible, whatever remains, however
improbable, must be the truth?” Sherlock Holmes to Dr. John Watson. Sir Arthur Conan Doyle, The Sign of Four,
Smith, Elder & Co, 1908, Chapter 6, “Sherlock Holmes gives a demonstration”, p. 94.
19 [Eds.] See Problem 2.3, p. 99.
20 [Eds.] Strictly speaking, the Levi–Civita symbol ϵλµαβ is a tensor density, and under Lorentz transformations

Under proper Lorentz transformations (SO(3, 1)), the determinant equals 1, so there’s no problem.

Problems 10

10.1 In Chapter 16, we computed ′(p2), the renormalized meson self-energy operator, to O(g2), in Model 3. We
expressed in (16.5) the answer as an integral over a single Feynman parameter, x, and we saw that ′(p2) was an
analytic function of p2, except for a cut running from 4m2 to ∞. In the same theory, compute the renormalized
“nucleon” self-energy, ′(p2), again to O(g2). Express the answer as an integral over a single Feynman parameter,
and show that this too is an analytic function of p2, except for a cut running from a location you are to find, to ∞.
(1997a 9.1)

10.2 Verify the commutation relations (18.46)–(18.49), using the defining representation of the group, D(Λ) = Λ.
For example,

and

Expressions for rotations and boosts along the and directions can be found from these by cyclic permutation of
x, y and z. Check by explicit computation that

(The other relations follow from these by cyclic permutation.)

(1997a 9.2)

Solutions 10

10.1 The renormalized “nucleon” self-energy is, analogous to (15.56),

where −i ′(p2) is the sum of all two-point 1PI diagrams. At O(g2), the only two-point 1PI diagram is shown below:

(This is diagram 2.4 (a), following (10.31).) The Model 3 Feynman rules (p. 216) give for this diagram

Combining the denominators with a Feynman parameter x, we have

Shift the integration by setting k = q + xp:

Using the integral table on p. 330, (I.4) gives us

From (S10.1)

The question to be answered now concerns the branch cut. The shared denominator of the expressions between
the curly brackets can be rewritten:

so we need not worry about the denominator. Then ′(p2) has a branch cut discontinuity only should the
numerator f(x) of the fraction in the logarithm equal zero for some x ∈ [0, 1]:

This function is a quadratic in x, so will either be concave up or down. It’s easy to see that

If f(x) is concave down, there will never be a value x ∈ [0, 1] where f(x) = 0. So we need worry only about concave
up, i.e., p2 > 0. And in fact we know p2 ≥ m2. We will have f(x) ≤ 0 only if the minimum value of f(x) is less than or
equal to zero. So we need to find this minimum value:

so that

This minimum value will be less than or equal to zero only if

This is a quadratic in p2, and the roots are

The root p2 = (m − µ)2 < m2 is impossible (if µ > 2m, the meson would be unstable), so p2 ≥ (m + µ)2 is the start of
the branch cut. This is what we expect from the spectral representation. To O(g2), the only particle state a nucleon
field can make when applied to the vacuum is a state containing one meson and one nucleon.

10.2 For the defining representation of SO(3, 1), we have the generators of rotations,

and similarly (it’s just the cyclic permutations; (Li)jk = −iϵijk, with ϵ123 = 1, and (Li)0k = (Li)k0 = 0)

The generators of the boosts are

and My , Mz similarly; (Mi)νµ = i(δiµg0ν − δ0µgiν):

Then

As the problem states, the other commutators can be found in the same way, or by cyclic permutation, in
agreement with (18.46)–(18.49).

19
The Dirac Equation I. Constructing a Lagrangian

We are now in a position to take the simplest of the Lorentz group representations that have a chance of
representing particles of spin-½ and making field theories with them. In the first instance we will consider field
theories with linear equations of motion, so we’ll have theories of free particles. After we quantize them, we’ll start
adding interaction terms, following the path of the first part of this course, and develop theories of interacting
particles.

19.1Building vectors out of spinors

We will want to construct a Lagrangian, and of course we want this Lagrangian to be a Lorentz scalar. The
Lagrangian will have derivatives in it, which transform as vectors. So it would be good to see how we might
construct a vector out of whatever we use to represent a spin-½ particle, in order to build Lorentz scalars as inner
products between derivatives and these new vectors.

We know how to represent spin-½ as representations of the rotation group, with Pauli spinors. The simplest
Lorentz group representations that could describe particles of spin-½ are the representations D ( ,0)(Λ) and D (0,
)(Λ), which reduce to Pauli spinors when we restrict ourselves to the rotation group. The generators L of rotations
for both these representations are

The generators M of boosts differ for the two representations:

the plus sign applying to D ( ,0)(Λ) and the minus sign to D (0, )(Λ).

Thus for example, consider the two component objects u+ or u–, belonging to D ( ,0)(Λ) or D (0, )(Λ),
respectively, and transforming accordingly under the Lorentz group. (For the moment we’ll ignore the space and
time dependence of the u’s.) Under rotation about an axis through an angle θ, these transform just like a Pauli
spinor:

It doesn’t matter which case we’re looking at, u+ or u–, the generator L is always σ. On the other hand, under a
boost along an axis with a speed v = tanh ϕ,

Please notice that the two objects u± transform differently under boosts. These are two component objects, each
of which transforms according to some irreducible representation of the Lorentz group. They are called Weyl
spinors.1 And because parity exchanges fields belonging to D ( ,0)(Λ) and D (0, )(Λ),

Let’s see what we can build out of u+ and u†+ by putting together bilinear forms in u+ and u†+. Everything I say will
go for the minus case, within trivial sign changes. Four linearly independent bilinear forms can be built out of u+
and u†+. Because u+ transforms like D ( ,0), its conjugate u†+ transforms like D (0, ). Then the bilinear forms
transform like

Whether we use u+ or u–, it doesn’t matter: one is the conjugate and the other is not. And the product is of course
simply D ( , ), which is the representation for a vector, as we’ve seen earlier. Therefore if I put together bilinear
forms in u+ and u†+, the four independent bilinear forms should transform like the four components of a vector.
Let’s work out precisely what that vector is. I’ll write it as the contravariant vector Vµ. There’s only one possible
choice for the time component:

That’s certainly the only bilinear form which is a scalar under rotations, from the ordinary theory of spin-½
particles. Likewise, up to a multiplicative factor which I’ll call η, there is only one possible choice for the three
space components: 2

Our general formalism hasn’t led us astray, at least so far. The four bilinear forms we can make can indeed be
arranged into an SO(3) scalar and an SO(3) vector, which is what we want for a 4-vector.

Let’s try to figure out how these bilinear forms transform under boosts by applying an acceleration, say about
the z axis by a hyperbolic angle ϕ. Of course they must behave as a 4-vector for the appropriate choice of η, but
it’s amusing to work it out. First, we need the transformations of u+ and u†+:
The argument iMz = ± σz in the exponential is now a Hermitian matrix. That’s the difference between Lorentz
transformations and rotations, which have an i in the exponential:

Now let’s work out what happens to the four components of our putative vector and see if they indeed transform
as components of a vector should transform. Well, we know how u+ and u†+ transform, so we just stick the
transformed u’s into (19.7):

The in the exponent disappears. This is a very easy matrix to compute, since the even powers in the power
series expansion are proportional to one, σ2 being one, and the odd powers are proportional to σz :

The even powers give us cosh ϕ, the odd powers give us sinh ϕ. This is

which is the statement

which is just what we want, if we choose η = 1. That is, we can identify a set of bilinear terms with a Lorentz 4-
vector:

Let’s check the other components, starting with V3:

Here σz commutes with σz , so I can use the same expansion again:

and

so that

which is again the right answer.

What about V1 or V2? Those are supposed to be unchanged under an acceleration in the z direction. Well, V1
or V2 goes into

Now σz anticommutes with either σx or σy ;

and therefore, when I bring a σz through a σx or a σy , it gets turned into −σz . So

because the combination e− σzϕ e σzϕ is known to its friends as 1. The result is
in other words,

Thus everything works out just the way it should. Still, it is reassuring to see the marvelous algebra of the Pauli
matrices doing our job for us, enabling us to construct, out of these two component objects u+ and u†+, a vector
which has a sensible transformation law not only under rotations but under Lorentz transformations as well. The
Weyl spinors u+ and u†+ don’t transform like vectors, but more like square roots of vectors, as it were, because it
is bilinear combinations of Weyl spinors that act like Lorentz vectors.

Exactly the same reasoning applies for u–, except there is a minus sign in the σ matrix. If we were working
with u–, the corresponding vector object Wµ, a completely different vector from Vµ, would be

The vectors Vµ and Wµ are products of a Weyl spinor and its adjoint. Which of the two different kinds of Weyl
spinors you are working with affects only the sign of the space component of the vector.

Incidentally, the complex conjugate u+* is equivalent to u–. Our starting point is (19.3). Complex conjugate this
equation:

The σi are not all real, and the −i goes into i. Now use an identity:

The identity is easy to prove, because σy is the only imaginary σi, and also the only σ matrix that commutes
rather than anticommutes with σy . We can make a similarity transformation using T = σy ,

because σy 2 = 1, and I can insert it in between every factor in the power series expansion of the exponential. This
is a formal proof of the assertion made in (18.67). But we still have to look at the boosts. It’s the same
manipulation, starting with (19.4):

Then making the same similarity transformation,

which is the appropriate matrix (19.4) for u–. So u+* transforms in a way equivalent to the way u– transforms, after
a change of basis, with T = σy .

19.2A Lagrangian for Weyl spinors

Now let’s try to build a free field theory using a u+ object only. Let’s promote these things from two component
objects to two component fields, functions in space and time, and attempt to build a free classical field theory. I’ll
do the u+ case in detail, and then I’ll just tell you how the answers change if you have the u– field instead. This will
be our first stab at making a Lagrangian L for a spin-½ particle. Guided by our experience with scalar fields, we
begin with some criteria for the theory:3

Criteria for a Weyl Lagrangian

1.L must be a Lorentz scalar, to guarantee Lorentz invariance.

2.L must have an internal symmetry, to give a conserved charge.

3.L must be bilinear in the fields, to give linear equations of motion.

 4.L should have no more than two derivatives, for simplicity.

5.The action S = d4xL should be real; S = S*.

The first requirement needs no discussion. What about the second? Every known spin- particle is associated
with some conservation law; the conservation of baryon number or the conservation of lepton number, hence I
might as well only consider free Lagrangians that obey that conservation law. So I will demand invariance under a
phase transformation

with arbitrary α, since we know from our previous experience that it’s phase transformations like these that give us
conservation laws in scalar field theories. Third, I want to obtain linear equations of motion, so I want my
Lagrangian to be bilinear in the fields. Since I also want it to be invariant under phase transformations, I want each
term in the Lagrangian to contain one factor of u+ (or its derivative) and one factor of its adjoint (or its derivative).
That’ll simultaneously give me linear equations of motion and guarantee invariance under the phase
transformation. We can say more about the derivatives. In the scalar case I was able to get by with no more than
two derivatives in the equations of motion, so to keep things simple, and following our general formalism, as a
fourth condition I’ll demand no more than two powers of derivatives in any term in the Lagrangian. Thus we can in
principle have three kinds of terms in the Lagrangian:

(With integration by parts, terms of type (c) could be replaced by terms with one derivative on u†+ and one on u+.)
These are just generic. We don’t know however whether these will obey the first condition, that L be a Lorentz
scalar. What is consistent with constructing a scalar?

We’ve already shown there are four linear combinations of type (a). None of them transforms like a scalar;
they transform like the four components of a 4-vector. And there isn’t any way of putting together the four
components of the vector to make a scalar that would be only bilinear in the u’s. We can make a scalar, but it
would have the form

which is quartic in the u’s. So: no bilinear terms of type (a). That’s pretty grim. We would expect from our previous
experience that the mass term would show up as a quadratic term of type (a). It looks like we will only be able to
construct a theory of massless particles. We also know that this theory will not conserve parity, because to get
parity we need both a u+ and a u–. So we’ll get a theory of massless particles that is incapable of expressing parity
invariance. Well, after all, neutrinos exist.4 Let’s see where we can go with this, and later, we may try more
complicated theories that have a chance of working for spin-½ particles other than neutrinos, like electrons or
protons.

By the same token we can’t include a term of type (c). The derivative operator is a vector, the bilinear forms
all transform like vectors, and out of three vectors there is no way of building a scalar. You can build a vector, or
some crazy kind of three index tensor, but you can’t build a scalar.

Fortunately there are possible terms of type (b), because we can put together the vector index of a derivative
with the index of the vector Vµ, (19.15), that we found before. An invariant Lorentz product of these two vectors
can be written as

Here I’ve put together the index of ∂µ with the index of Vµ and had the derivative act only on u+. We could of course
also put the derivative on u†+, but if we’re constructing an action out of this expression (19.34),

I can move the derivative with integration by parts:

That’s the same thing, aside from the minus sign. So in fact I have only one invariant, this object (19.34), which I
can use to make a Lagrangian. Everything else is either not Lorentz invariant, or equivalent to (19.34) under
integration by parts.

At the end of all this messing around, we find we have essentially a unique Lagrangian, aside from a scale
factor in front:

The magnitude of the proportionality constant can be absorbed by rescaling the u’s, just as in the scalar case we
analyzed so long ago. The adjoint of the integrand of (19.35) is the (positive) integrand of (19.36), but integration
by parts of the Lagrangian (19.35) turns it into −L. To satisfy the fifth criterion, that the action be real, the
coefficient in front has to be purely imaginary. So we are left with just two choices, as in our earlier analysis of the
scalar case:

We won’t be able to fix the plus or minus sign until we finally quantize this theory, put it in canonical form, compute
the energy and see whether it is positive or negative. For the u– case, the only difference would be a different sign
for the gradient term:

As we’ll see, this has a profound effect on the particles we finally get out of this theory. I now propose to explore
this Lagrangian (19.38) first on the classical level, and next time on the quantum level.

19.3The Weyl equation

Our first step is to derive the equations of motion which we get by varying the fields. The easiest variation to do is
that with respect to u†+ since we don’t even have to do any integration by parts. We obtain the Weyl equation

By varying with respect to u+ and integrating by parts we just get the adjoint of this equation, as is usual for
complex fields. That gives us no new information. Equation (19.40) is our equation of motion. It may not look
Lorentz covariant, but it is. Of course for u– we would get

We can gain some insight into the meaning of (19.40) by multiplying it on the left by the operator ∂0 − σ•∇:

The product is simple to work out. The cross terms cancel because ∂0 commutes with σ•∇, and

so we obtain

which is of course just what we we expect for a massless particle. All plane wave solutions of this equation are of
the form

The spinor up is constant (independent of x), and since p2 = 0,

These plane waves are like those for massless scalar particles. I should make a remark, although this is
anticipating things a bit. We should expect in the quantum theory that when we expand out the general solutions to
the field equation in terms of these linearly independent plane wave solutions, the coefficients of the e−ip⋅x terms
will be annihilation operators, and the coefficients of e+ip⋅x terms will be creation operators, both for mass zero
particles. That’s just a guess based on what we discovered for scalar theories.

Let’s determine up. For simplicity I will consider the case

So we have a particle moving in the +z direction. The sign of p0 is irrelevant; it factors out of the equation. And
indeed the magnitude of p0 factors out of the equation. Plugging the plane wave solution (19.45) into the Weyl
equation (19.40) for u+, we obtain, dividing out ±ip0,

That’s a pretty easy equation to solve. If we use the standard representation of the Pauli matrices,

we find

This means that the Weyl equation has one linearly independent solution for each value of the 4-momentum on
the light cone. Thus we would expect, when we quantize this theory, that we have one kind of particle for each
momentum, described by one annihilation operator and one creation operator.

Let’s make a guess about the quantum theory of this particle. In particular I’m interested in its spin. Well, I
shouldn’t really say “spin”, because spin is a concept that applies only to particles with mass, because only for a
particle of non-zero mass can we Lorentz transform to its rest frame and there compute its angular momentum,
which is its spin. For a massless particle, there is no rest frame, so we can’t talk about the spin. We can however
talk about its helicity, the component of angular momentum along the direction of motion. That’s perfectly
reasonable and doesn’t involve the rest frame. So let’s try to compute Jz , for a one-particle state, |pñ, with
momentum p, associated with this equation of motion, (19.40).

By comparison with what we found in the scalar theory, we’d expect to write the quantum field as a
superposition of these solutions, some with annihilation operators, and some with creation operators. Therefore
we should expect, aside from inessential normalization factors, that if we put the quantum field between the
vacuum and this one-particle state, we would obtain something proportional to upe−ip⋅x :

That’s a straightforward transposition of what we discovered in the scalar theory. I’ll be interested in this equation
only at the point x = 0. That will suffice to allow us to determine the helicity:

Now we would expect that a particle moving in the z direction can always be chosen to be an eigenstate of helicity
Jz . There’s only one particle, so it must automatically turn out to be an eigenstate of Jz :

where λ, the eigenvalue of Jz , is the helicity of the particle. So the unitary transformation

that effects a rotation about the z axis by the angle θ in the Hilbert space of the theory, applied to the state |kñ,
results in the eigenvalue equation

Then

(I assume the vacuum is rotationally invariant, so applying eiJzθ to |0ñ shouldn’t have any effect.) But we know from
(18.1) and (19.3) that
and so, from (19.52),

because up is an eigenstate of σz with eigenvalue +1. Comparing the two sides of (19.56), we see λ = + . Thus
the particles in this theory—if there are particles, if we can successfully quantize it—annihilated by u+ have helicity
+ , but this theory does not have particles of helicity − . Such a theory is only possible when the particles are
massless, and when parity is not conserved. If the particles had a mass, you could always transform to a
reference frame traveling faster than the particle. In that frame, the particle would be going in the opposite
direction, and hence with reversed helicity.

Of course there are two kinds of particles in this theory, because this is a charged field.5 And therefore the
field should not only annihilate particles of charge +1, but create antiparticles of charge −1, just as a charged
scalar field does. That is to say, there will also be terms proportional to eip⋅x , and they will be creation operators for
different particles. These won’t be the same as the original particles, because the field isn’t real. To investigate the
antiparticles, I have to put the antiparticle state on the left,

so it has a chance of being made. This will of course be proportional to the same up since it doesn’t matter which
sign of e±ip⋅x I look at. For the antiparticle, however, the equation corresponding to (19.55) is

The antiparticle helicity is λ′. From this point on, the whole argument goes through in exactly the same way as
before, except that e−iλθ is replaced by e+iλ′θ. Therefore we find

The antiparticle has helicity − . Thus our guess is that this theory, if we can successfully quantize it, will describe
massless particles and their antiparticles. The massless particles will carry one charge, and the antiparticles will
carry the opposite charge. The particles, by definition those objects annihilated by u+, will have helicity , and the
antiparticles, created by u+, will have helicity − . Similarly, particles created by u†+ will have helicity + , and
particles destroyed by u†+ will have helicity − . Conventionally, a particle with helicity + is called “right-handed”. If
you could see a right-handed particle spinning as it came toward you, it would appear to spin in a counter-
clockwise fashion.

For u–, of course, because of the minus sign in the equations of motion everything gets switched around; σz
gets replaced by −σz , but otherwise nothing is changed. If we were to consider the theory of a u– field, we would
find the particles’ helicity to be − , and the antiparticles have helicity + . Of course that’s what you would expect,
because the complex conjugate of a u+ field is a u– field. When we complex conjugate the fields we simply change
the definition of what we call “particle” and what we call “antiparticle”. This structure is no longer alien to physics,
although it was when Hermann Weyl first proposed it.6 Physicists dismissed this theory as the work of a dumb
mathematician: there was no parity invariance, the particles were massless, this was nothing like our world. But in
fact Weyl’s theory describes precisely the massless neutrino with but one helicity (left-handed); the antineutrino
has the opposite helicity. We haven’t quantized this u+ theory yet, but what I’ve described is what we would expect
if quantization were to go like the quantization of scalar theories. It’s possible for a massless particle to have only
one helicity, if the theory is not invariant under parity. That’s perfectly Lorentz invariant. If you introduce parity
invariance, then the helicities have to come in pairs; if a particle can be in a given helicity state, it must be able to
occupy a state with the opposite helicity. (A massive particle has to have every helicity between its maximum
helicity and minus its maximum helicity by integer steps. But massless particles, if they are parity invariant, have
only two helicity states.)

For example, photons have helicity +1 and −1. That’s because electromagnetism is parity invariant, and if a
photon has helicity +1, helicity −1 also has to exist. If electromagnetism were not parity invariant, it would be
possible to conceive of a photon that has only one helicity, say +1. Because they’re massless, they’re allowed not
to have helicity zero. You could always add to the electromagnetic field a massless scalar field, and call the three
states you get this way “the photon”. Then there would be helicity +1, −1 and 0. You might think it perverse to add
a such a scalar field, and I would agree with you, but it is certainly possible. As Pauli said in a very similar context
about only using irreducible representations in building a theory, “What God has put asunder, let no man join
together.”

19.4The Dirac equation

This Weyl theory does a nice job with the free neutrino, but of course there are a lot of spin − particles in the world
that are not massless. To get beyond massless particles and get something that has a chance of being a
reasonable theory of the free electron or the free proton, we have to complicate our theory somewhat. We’ve
explored everything we can reasonably do with a just a u+ field or a u– field. We also know that the interactions of
the proton and the electron are parity-conserving up to an excellent approximation. So we will now try and make a
parity-conserving theory. To make a parity-conserving theory, we need both a u+ and a u– field because parity
turns u+ into u–, and vice versa. We can list some criteria for a theory of massive spinors:

Criteria for a Lagrangian with massive spinors

1.L must be a Lorentz scalar, to guarantee Lorentz invariance.

2.L must be bilinear in the fields, to give linear equations of motion.

3.L must have an internal symmetry, to give a conserved charge.

4.L must be invariant under parity.

5.L should have no more than one derivative.

6.The action S = d4xL should be real; S = S*.

I want L bilinear in the fields so I have linear equations of motion for the free field theory. I still want to preserve an
invariance that corresponds to charge conjugation,

I don’t want two conserved charges, so I don’t want to say there’s an independent phase transformation for u+ and
u–. Then I want the Lagrangian to be invariant under a parity transformation and I’ll assume in the first instance the
most general form

I know parity interchanges u+ and u– but it might multiply them by a phase. I’m going to be as general as I can be.
As we’ll soon see, this generality is spurious, and we can pin ϕ 1 and ϕ 2 to be fixed numbers. I don’t care if the
square of the operation is not one. Any sort of transformation of this form I’ll call “parity”. In the fifth criterion, I will
be a little more restrictive than I was in the preceding case, and assume no more than one derivative. I could
assume two derivatives, but after all, in the previous case I got along just fine with one derivative, so let’s try one
derivative here.

Now let’s write down the most general Lagrangian. Because of condition one, we have several kinds of terms.
We could have u†+u+, either with or without a derivative in there somewhere, and likewise u–†u–:

And we could have these terms,

Now we’ve already classified all the u†+u+ and u–†u– terms. The only Lorentz invariants involving these terms,
(19.38) and (19.39), have derivatives. How do these change under parity? Because
we have

The phase factors (19.64) are irrelevant, because they cancel out in this combination. Parity transforms the two
Lagrangians, (19.38) and (19.39), into each other. Consequently their sum is invariant. The real benefit for the
relative minus sign between Lorentz transformation laws for u+ and u–, arising from (19.4) and leading to the
different forms of 4-vectors in (19.15) and (19.25), is that we can build a parity invariant theory.

What about u†+u–? Whichever of u+ or u– transforms like D ( ,0)(Λ), the other transforms in the other way. But
the adjoint takes care of that, and switches it back again. So we have to deal with things like this:

for which we get D (0,0), that’s a scalar. That means we can build the scalar without any worries. But the second
part, D (1,0)(Λ), is half of an antisymmetric tensor. This means that with this form we can’t build anything with
derivatives in it, because the derivative operator is a vector, and there’s no vector here to dot things into. However
we can build a non-derivative term, like this:

where m is an arbitrary complex number. This is the only combination that’s a scalar under rotations, so it must be
the combination that’s a scalar under the full Lorentz group. Because we want the Lagrangian to be Hermitian, we
add the other possibility, with the conjugate coefficient:

Thus the most general Lagrangian satisfying our five conditions takes the form

It involves a single arbitrary complex number m which I will shortly trade for a positive real number.

Now let’s get rid of the phase in m. I can always redefine

I have the freedom to change variables when writing down my Lagrangian. If I substitute that in, and drop the
primes, the terms in the derivatives aren’t affected, but the terms in m and m* are affected; their phases are
changed. I can always absorb the phase of m in such a transformation into my definition of u+, to make m real, and
greater than or equal to zero. This changes the definition of parity, of course; ϕ 1 and ϕ 2 in (19.64) are changed.
With m chosen to be real, the new definition of parity is

Once I’ve chosen the phase of u+ so that m is real, I no longer have the freedom to assign phases differently to u+
and u–. Of course I still have an infinite set of possible choices for the phase ϕ 1, because I can always multiply
parity by an internal symmetry and declare that to be parity. That is my privilege. It’s only the total group of
symmetries of the Lagrangian that counts, not what names we give to any individual member. I will define a
standard parity which is simply the natural choice:

This is purely a convention of nomenclature. If I were perverse, I could’ve chosen any one of (19.74) to call parity.
That, too, would be a symmetry of the Lagrangian.

We now have an ugly-looking Lagrangian, characterized by a single real number, m, and an overall sign
choice, exactly as many free parameters as we had in the corresponding case of the free scalar field:

The equations of motion are scarcely more complicated than the Weyl equations:
Now it’s pretty easy to see what to do with these equations. I multiply (19.77) on the left by i(∂0 − σ•∇), and find

That is, u+ obeys the Klein-Gordon equation appropriate for a particle of mass m. Likewise we can start with
(19.78), multiply by i(∂0 + σ•∇) and in exactly the same way show that

So I have not lied to you in my choice of the symbol “m” for this free field theory: m is indeed the mass of the
quantum of the field. Further implications of the theory, what the spins of the particles are and so on, are topics for
next time.

Our Lagrangian is the sum of a bunch of grotesque-looking terms. Let us simplify our notation somewhat by
incorporating u+ and u– together into a single four-component object, ψ:

The top two components of ψ are the two components of u+, and the bottom two are the components of u–. I will
define three 4 × 4 matrices, α, which are block diagonal with the Pauli sigma matrices σ, −σ and zeros elsewhere:

I will define a fourth 4 × 4 matrix, β, which is block off diagonal:

These matrices α and β are chosen so that the ugly Lagrangian (19.76) has a rather simple form in terms of α, β
and ψ:

The equations of motion can be obtained by writing (19.77) and (19.78) in terms of α, β and ψ, but we can get
them directly by varying the Lagrangian with respect to ψ †:

This equation is called the Dirac equation, though expressed here in a slightly different basis than that written
down by Dirac in 1929. The forms of ψ, α and β used here are called the Weyl basis, and the matrices α and β
are called Dirac matrices. Next time we will begin exploring the properties of the Dirac equation—finding out what
the solutions are, making guesses about the properties of the particles represented by those solutions, and so on.
Since we’ll be spending a lot of time with the Dirac equation, we will want to develop a sequence of algorithms for
handling its solutions as effectively as possible. Finally, we will quantize the Dirac theory, looking at the energy
and establishing the correct sign of the Lagrangian.

1 [Eds.] Pronounced “vile”. Noticing the curious reaction of his students, Coleman adds: “Not that they are
disgusting, but that they were first explored by Hermann Weyl.” (Hermann Weyl (1885–1955), among the great
mathematicians of the twentieth century, also contributed to relativity and quantum theory. He was a colleague
and friend of Schrödinger, Einstein and Emmy Noether, whose funeral oration he gave at Bryn Mawr College, 17
April 1935.)
2[Eds.] In the video of Lecture 19, Coleman uses α for the factor written here as η, to avoid confusion with the
Dirac matrices, α i, and the fine-structure constant, α.
3 [Eds.] We remind the reader that the action is denoted , and the S-matrix is denoted S.
4 [Eds.] In 1975, neutrinos were thought to be strictly massless.
5 See the second criterion in the box on p. 397. The current associated with this symmetry is just Vµ, given by
(19.15), as you can check by our general formula (5.27) applied to the symmetry (19.31). I will not work out the
current for the Weyl particles, but will do it for the massive particles described by spinor fields.
6 [Eds.] H. Weyl, “Elektron und Gravitation”, Zeits. f. Phys. 56 (1929) 330-352; English translation in L.
O’Raifeartaigh, The Dawning of Gauge Theory, Princeton U. P., 1997, pp. 121–144; the Weyl equation appears
on p. 142 (setting fp = 0 in the absence of gravitational coupling).

20
The Dirac Equation II. Solutions

Having derived the Dirac equation, we’re now going to manipulate it, study its solutions, and write it in different
bases with new notation, all with the aim of making the free Dirac equation easy to work with. Later on, when we
have to do complicated things with it and make it part of an interesting quantum field theory, we’ll be able to do it
efficiently.

20.1The Dirac basis

The Dirac equation in the form we found it last time had a Lagrangian, (19.84):

where m was a positive real number, and

Notice that these matrices are Hermitian,

They also obey the Dirac algebra (also called the “Clifford algebra”1)

where the curly brackets indicate the anticommutator, the sum of the products in the two different orders; by
definition

Finally, the square of each α i, and of β, is 1.

We can write the generators of Lorentz boosts and rotations in terms of α and β. The Lorentz boost generator
M is (19.2)

because the generator is iσ/2 for u+ and −iσ/2 for u–. The rotation generators are

because, from (19.3), u+ and u– transform the same way under rotations. Finally, from (19.5), parity exchanges u+
and u–, so we can write

This formulation is specific to a particular basis: arranging the two components of u+ and the two components
of u– into a four-component object ψ in a certain way, (19.81). The equations (20.6), (20.7) and (20.8) depend on
the explicit form of α, β and ψ. The Dirac algebra (20.4), the boost commutators and the other parts of the Lorentz
algebra, do not. If we had chosen our basis differently, if we had put u+ and u– together to make ψ in a different
way, the explicit forms of α, β, L and M would be changed, but the equations expressing the Dirac and Lorentz
algebras would not be. They would simply be expressed in terms of the α, β and L in the new basis. This particular
basis is called the Weyl representation of the Dirac equation. (The word “representation” is a bit strange, since its
usage here has little to do with group theory.) It is not the representation in which Dirac first wrote down the
equation. He chose to write

In such a basis, the 1/ is inserted, so that the Lagrangian’s term

will be unchanged. In the standard representation, with this form of ψ,

The matrix β is block diagonal, α is block off-diagonal, and L still has the form (20.7), because the sum and
difference of u+ and u– transform in the same way under rotations as ordinary Pauli spinors. The Dirac
representation is called the “standard representation” for historical reasons; it was the one first written down by
Dirac. Aside from the explicit forms of ψ, α and β, equations (20.2) versus (20.9) and (20.11), everything is the
same in both representations. I don’t expect you to see offhand what α and β look like in the standard
representation, but you can check it in a moment just by plugging them into the Lagrangian (20.1) and seeing that
you get the same quadratic functions2 of u+ and u–.

20.2Plane wave solutions

The standard representation is especially convenient for finding explicit solutions to the Dirac equation in the limit
of small p. In that limit, the term in β dominates, and we have diagonalized β in this representation. Let me work
out the plane wave solutions to the Dirac equation, (19.85):

I will look first at positive frequency solutions, so-called because they have a negative term in the exponential:3

(The four-component coefficient up is not to be confused with the two-component objects u+ or u–.) These are the
solutions you would expect in the quantum field theory to multiply particle annihilation operators. (I’ll talk about the
solutions that would be associated with antiparticle creation operators later.) Since we know all solutions of the
Dirac equation obey the Klein–Gordon equation with mass m, p0 can be chosen to be Ep;

Plugging (20.12) into the Dirac equation, the i and the −i cancel, and I obtain

When p = 0, this equation is particularly easy to solve in the standard representation. Then E becomes m, and we
obtain

This equation has two linearly independent solutions, u0(r), r = {1, 2}. Explicitly, they are

These solutions are normalized so that

The reason for this peculiar normalization convention will become clear shortly. Also,

This results from α being a block off-diagonal matrix. These two results can be put together into a suggestive form,

Maybe this normalization, which looks like it might be Lorentz invariant, will hold even for solutions with non-zero
p.

By guesswork identical to that used for the Weyl equation (§19.3), u(1) should be associated with the
annihilation operator for a particle at rest with zero momentum and Jz = + . The particle in this equation resembles
an electron, so I’ll call it an electron. The solution u(2) is the same thing with Jz = − . Of course, that’s the real
reason we have two solutions: We have a theory of massive particles with spin one half, and we cannot get by
with fewer than two solutions for each value of the momentum. I will not always use these two solutions called u(1)
and u(2), because I may not be interested in the z axis; maybe I want to look at the x-axis. However I will always
normalize my solutions so that this normalization convention (20.19) holds. Any linear combination of these u(r) will
be just as good. So much for solutions at rest.

What about moving solutions, solutions associated with nonzero p? It might be thought that we have to solve
a more complicated equation to construct them. Actually, we don’t. The theory has been constructed to be Lorentz
invariant, so let’s exploit this Lorentz invariance and obtain a solution associated with a nonzero p by applying a
Lorentz boost to a solution associated with the zero p. Thus I define (see (18.45))

The operator in front of u0(r) is a Lorentz boost along an axis by a hyperbolic angle ϕ. The axis is chosen to
boost the particle at rest in the direction of the desired momentum, p,

and ϕ is chosen to boost it by the right amount,

so that ϕ = 0 when Ep = m. The normalization conditions obeyed by these solutions are simple, since (20.19) form
the space and time components of a 4-vector:

Just to be clear about this, let’s work out the explicit case for the particle moving in the positive z direction.
Then = and the relevant matrix is α z ,

Because α z 2 is 1, it’s easy to compute

where

Now you see the reason for the in the normalization (20.16): it’s to cancel out those ugly denominators. Thus
we find
 A nice thing about (20.20) and the normalization (20.16) is that the solution (20.27) has a smooth limit as m
goes to zero (though the method we have used to define these functions does not):

Thus we can smoothly take the limit as the particle mass becomes negligible, either because we’re studying the
physics of a massless fermion, or because we’re doing a process where an electron (or something similar) gets
produced at such a high energy that its mass is negligible.4

Using (20.27), we have

so that, in agreement with (20.23),

Similarly we can construct

and we can work out that

in agreement with (20.23).

Of course, everything I have said about the positive frequency solutions goes through, mutatis mutandis, for
the negative frequency solutions. Writing vp for these spinors,

Once again

When we finally quantize this theory, we expect these solutions to multiply creation operators for positrons, the
antiparticles of electrons. We plug (20.36) into the Dirac equation and get an equation almost identical to (20.14):

The minus sign on the β term is a reflection of the different sign of the exponential’s argument. Once again the
solution is most easily done in the case when p = 0:

The negative frequency solutions are, like those of positive frequency, two in number. In the standard
representation,

Because we expect these to be the coefficients of creation operators rather than annihilation operators, the same
sign switch in the eigenvalues that occurred when we were discussing helicity5 occurs here, and we expect v(1) to
multiply the creation operator for a positron with Jz = − , while v(2) creates a positron with Jz = + . These states are
supposed to multiply a creation operator, and therefore the phase of the helicity gets switched, because the state
is on the left, rather than the right, of the creation operator.

We define moving solutions in exactly the same way as before (see (20.20)):

This gives

We have the same normalization condition among the v’s as for the u’s:

So much for the Dirac equation and its plane wave solutions.

20.3Pauli’s theorem

So far we have discussed two different representations of the Dirac matrices, due to Weyl and Dirac. Of course
there are an infinite number; any invertible 4 × 4 matrix can be used to transform one representation of the Dirac
matrices into another (by a similarity transformation). But there are some properties of the Dirac matrices that are
independent of one’s choice of basis. In particular, the set of four matrices {α i, β} obey the Dirac algebra

It follows that the square of any α i equals 1, and we require the same of β. I will prove a theorem6 due to Pauli:

Theorem 20.1. Any set of 4 × 4 matrices with unit squares obeying the Dirac algebra is equivalent to the Weyl
representation.

Actually the entire structure of the theory is embedded in these algebraic equations: Any equivalent set of 4 ×
4 matrices defines physically the same Dirac equation as any other equivalent set. The set of Dirac matrices used
is merely a matter of choice of basis for the four components of the Dirac field. So, implicit in this proof is a more
significant result: All irreducible representations of the Lorentz group symmetric under parity and containing only
spin-1/2 particles are equivalent. In effect, there is only one such representation, and all the others are related by
similarity transformations. The proof I’ll use is not the one given by Pauli, but one based on our analysis of the
representations of the Lorentz group, since we already have a lot of useful theory from our earlier work.

The proof goes as follows. Define Mi and Li by

I will now prove that the M’s and the L’s so defined obey the commutation relations for the set of generators of a
representation of the Lorentz group, as a consequence of the Dirac algebra (20.4). The [Mi, Mj] commutator is of
course true by definition. Let’s use it to determine one of the components of L, say the x component:

because α y and α z anticommute. So we have

Ly and Lz are found by cyclic permutation.

Now let’s check a typical [Li, Mj] commutator, say Lx with My . If we get this one right, we get the others by
cyclic permutation.
This is of course the right result for Lx with My (18.47). Let’s check a typical [Li, Lj] commutator, say Lx with Ly :

because α z 2 is 1, and in the second term, the α z at the left can be moved to the right through both α y and α x ,
which gives two sign changes. Thus all the commutators check. If I start out with four 4 × 4 matrices with unit
squares obeying the Dirac algebra, I generate from them a four-dimensional representation of the Lorentz group.
But which one?

There are many four-dimensional representations of the Lorentz group. For example, there are D ( ,0) and D ( ,
), but neither of these can be the right one, because

So the representation we’ve generated with the α’s contains only the eigenvalues ± of Lz . The representation D (
,0) contains Lz = ± , and D ( , ) contains only integer values of Lz . There are thus only three possibilities:

These are the only four-dimensional representations of the Lorentz group with the right eigenvalues of Lz .

We now use β to select from these three the unique representation generated by the α’s. From the identity

we have (see (20.44)),

and

since L is bilinear in the α’s (see (20.47)). The similarity transform of both L and M with β as the matrix T are
exactly how these generators transform under parity, as in (18.50) and (18.51), respectively. Therefore β can be
used to define a parity operation. The representation we’ve generated thus must be invariant under parity, and so
equivalent to its parity transform. But under parity, as we have seen in (18.77), the two indices of D (s 1,s 2) are
swapped. Of the three candidates in (20.51), only the last, D (0, )(Λ) ⊕ D ( ,0)(Λ), is equivalent to its parity transform.

So suppose I have some matrices α and β, not necessarily the Dirac representation, which satisfy the Dirac
algebra, and I’ve found a nonsingular matrix T which takes the three α i to the Weyl basis for this representation:

Since β must anticommute with α, and its square is one, whatever form it had before, its form β′ after the similarity
transform T must be7

The unknown λ ≠ 0 multiplies a 2 × 2 identity matrix 1. I’m not assuming αW and β′ are Hermitian. Now I make a
second similarity transformation:

where

By elementary multiplication, this transformation doesn’t do anything to αW;

but
This similarity transform turns β′ into its Weyl form, as desired. Therefore, one and the same transformation ST
turns a given set of unit square 4 × 4 matrices satisfying the Dirac algebra into the Weyl representation, QED.

So what is the point of the theorem? If we want to write down Dirac matrices or the Dirac equation in some
crazy basis, we don’t have to construct this matrix S′ = ST. We are guaranteed that any four unit square matrices
satisfying the Dirac algebra will be connected to the standard matrices by some S′. Secondly, and more
importantly, it distinguishes what is important from what is not. A lot of talented people, some of them Nobel
laureates, did complicated computations in the early 1930s involving spin-½ particles. Typically this work was
done by writing down explicit solutions of the Dirac equation. But they were doing things the hard way. There is no
need to write down these solutions, because any desired calculation can be performed using only the algebra of
the Dirac matrices. Messy as the anticommutation relations are, they are a lot less trouble than working with
explicit 4 × 4 matrices. The whole structure of the theory lies in these anticommutation relations.

20.4The γ matrices

The manipulation of Dirac matrices is facilitated by a formalism introduced by Pauli (and automated further by
Feynman). Pauli’s scheme assumes the four Dirac matrices are Hermitian:

Both the Weyl representation and the standard representation satisfy this criterion. We will assume our Dirac
matrices are all Hermitian in the sequel. We now define a somewhat peculiar “adjoint” of ψ:

The quantity ψ is called the Dirac adjoint, though in fact it was introduced by Pauli. The motivation for this
definition comes from the mass term, mψ †βψ, in the Dirac Lagrangian, which transforms as a scalar. In terms of
the Dirac adjoint,

The term ψ †βψ may have appeared a little awkward for a scalar; ψψ looks much more natural. With

in the Dirac basis (20.11),

(This follows even more simply in the Weyl basis.) This expression, the sum of two quantities minus the sum of
two others, is a Lorentz scalar.

We would like to define a new adjoint A for any 4 × 4 matrix A, such that (Aψ) equals ψ A, just as for the
ordinary adjoint. The obvious answer is

The Lorentz matrices D(Λ) that effect Lorentz transformations

play well with the Dirac adjoint operation. Here,

Taking the Dirac adjoint of D(Λ)ψ, we have

Because ψψ is a Lorentz scalar,

Since ψ is arbitrary, we deduce that

Although the D(Λ) are not unitary, they are “Dirac unitary”. While they do not preserve the conventional quadratic
form, the sum of the squares, as a unitary matrix does, they do preserve this unconventional quadratic form
(20.65), the sum of two squares and the difference of two others.8

We saw earlier that two 4-vectors, Vµ and Wµ, could be constructed from bilinear products:

The sum of these is also a 4-vector, and can be written (in either basis) as

If we insert β2 = after ψ †, we can make the 4-vector nature of the bilinear product explicit:

where we define the four Dirac gamma matrices:

In the Dirac basis (20.11),

Under a Lorentz transformation,

but for a vector,

so we have to have

With a slight abuse of language, we say that the gamma matrices “transform as a vector”. Actually, the matrices
themselves don’t transform at all; but sandwiched between two Dirac spinors, the quantity ψγµψ transforms as a
vector.

From their definition,

All of these relations can be summarized in one line:

where is a 4 × 4 identity matrix (which we usually don’t bother to write explicitly). Other properties of the gamma
matrices are:

(The statements about (γµ)† hold only for Hermitian β and α i.)
 We can rewrite the Dirac Lagrangian (20.1) in terms of them:

Products of 4-vectors and gamma matrices occur frequently. Feynman introduced a useful shorthand for these
products:

(pronounced “a slash”). Then

and similarly

The Dirac Lagrangian can be rewritten in the slash notation,

and the equation of motion (from varying ψ)

If we multiply on the left with (i + m), we obtain the Klein–Gordon equation:

That is, each of the four components of ψ satisfies the Klein–Gordon equation.

20.5Bilinear spinor products

We’ve seen that ψψ is a Lorentz scalar, and ψγµψ is a Lorentz vector. It is worthwhile to investigate the
transformation character of other bilinear spinor products, with two or more gamma matrices sandwiched between
them. As we’ll see, there are sixteen9 linearly independent bilinear forms: scalar (1 component), vector (4),
antisymmetric tensor (6), axial vector (4), and pseudoscalar (1). First, though, a brief detour to consider the
behavior of ψψ under parity.

Earlier we argued, in the paragraph following (20.54), that for Dirac spinors, β = γ0 effects a parity
transformation:

Taking the Dirac adjoint of this equation gives

so we have

and consequently

That is to say, ψψ(x, t) transforms as a scalar under parity. By the same token,

which is exactly how you’d expect a vector to transform.

 The next most complicated expression is

The symmetric part of the bilinear expression is nothing new (it’s just the old scalar ψψ); but the antisymmetric
part is. It’s conventional to define

(The factor of i is included so that σµν = σµν.) It’s easy to verify that the bilinear expression ψσµνψ transforms as a
tensor. Consider the transformation of ψγµγνψ:

The bilinear ψγµγνψ transforms as a tensor, so the commutator ψ[γµ, γν]ψ = 2iψσµνψ must as well—it’s an
antisymmetric second-rank tensor.

What about the three gammas? Only if all three are different will the product differ from a single gamma matrix
(to within a sign). Disallowing the products equivalent to a single gamma, the products of three gammas produce
only four independent matrices. These Lorentz transform as the components of a 4-vector. For example, consider
γ1γ2γ3. We have

Each of the four independent matrices γλγµγν can be multiplied by the square of the “missing” gamma. That is, the
four independent matrices can be written collectively as

where the “fifth γ matrix”, γ5, the unique product of four gammas that does not reduce to the product of two
gammas or the identity matrix, is defined10 to be (with the convention ϵ0123 = +1)

with properties

We’ll sometimes work with iγ5 in preference to γ5, because

The quantity iψγ5ψ transforms as a scalar under proper Lorentz transformations (with detΛ = 1), but under parity,

That is, iψγ5ψ is a pseudoscalar. (It is also Hermitian, and can appear in a Lagrangian with a real coefficient.)

Finally, we have the bilinear product

Under proper Lorentz transformations, it behaves as a vector. But under parity,

The quantity ψγµγ5ψ is thus an axial vector.

We have now found five bilinear spinor products transforming in distinct ways under parity and Lorentz
transformations:
 Any 4 × 4 matrix can be expressed as a linear combination of 16 basis elements. These 16 bilinear products
thus form a basis for any bilinear product. Ultimately we will build interactions out of these bilinear products.

20.6Orthogonality and completeness

We can express the normalization conditions (20.23) and (20.43) in terms of the gamma matrices:

The taking the product of these with pµ we get

If we substitute the plane wave solutions (20.12) and (20.36) into the equations of motion (20.90), we obtain

so that

Comparing (20.109a) with (20.111a), and (20.109b) with (20.111b), we find

So the solutions up(1) and up(2) are Dirac orthogonal to each other, as are vp(1) and vp(2).

What about the mixed expressions vp(r)up(s) and up(r)vp(s)? Taking the Dirac adjoint of (20.110a) gives

and multiplying on the right with vp(s), we have

However, (20.110b) says

Multiplying this equation by up(r) on the left gives

Comparing (20.114) with (20.116), it follows, as m ≠ 0, that

and returning to (20.116), we have also

The positive and negative frequency spinors are Dirac orthogonal, and a vector made of positive and negative
frequency spinors vanishes. So much for orthogonality.

In calculations to come, we will frequently need to evaluate expressions involving sums of spinors. Suppose
we apply the operator ( − m) to up(r), and likewise ( + m) to vp(r). Then

Consider the sum (note the “backwards” order, with the column vector first and the row vector second)

This 4 × 4 matrix A acts like a projection operator. If we apply A to a linear combination of up(r) and vp(r), only the
up(r) component survives:

Well, we already have a 4 × 4 matrix with these properties:

so we have two completeness relations,11

(the second following by the same reasoning as for the first) and two complementary spinor projection
operators Pu and Pv :

The operators have the expected properties:

The sum is the identity:

where is the 4 × 4 identity matrix. Acting on either up(s) or vp(s), they are orthogonal

and idempotent:

These relations will be very helpful in computing processes with spin-½ particles.

Next time we will tackle the canonical quantization of the Dirac field, the calculation of the appropriate
propagators and the Feynman rules.

1 [Eds.] W. K. Clifford, “Applications of Grassmann’s Extensive Algebra”, Amer. Jour. Math. v. 1 (1878) 350–358;
reprinted (along with all of Clifford’s papers) in Mathematical Papers by William Kingdon Clifford, ed. Robert
Tucker, Macmillan, 1882. See also Appendix E, pp. 675–677 of Bernard de Wit and Jack Smith, Field Theory in
Particle Physics v. 1, North-Holland, 1986. William Kingdon Clifford (1845–1879) was a British geometer who
translated Bernhard Riemann’s inaugural lecture (June 10, 1854) “ Über die Hypothesen, welche der Geometrie
zu Grunde liegen” (On the hypotheses which lie at the base of geometry), Abhand. König. Gesell. Wiss. Gött. 13
(1868) 133–150, into English: Nature VIII (1873), No. 183, pp. 14–17; No. 184, pp. 36–37. This work led to
Clifford’s brief speculative note anticipating Einstein’s general relativity, “On the space-theory of matter”, Camb.
Phil. Soc. Proc. v.2, 1866–1876, Feb. 21, 1870, pp. 157–158, suggesting that matter curved space, which action
might be the basis of gravity.
2 [Eds.] The Dirac and Weyl representations are related by a similarity transformation,

3 [Eds.] From the Schrödinger equation, Hψ = iħ(∂ψ/∂t) = ħωψ has a positive eigenvalue ħω if ψ has a time
dependence of the form exp(−iωt).
4 This normalization convention is not that of Bjorken and Drell. ([Eds.] See Bjorken & Drell RQM, p. 31, equation
(3.11).)
5 [Eds.] See pp. 400–401.
6 [Eds.]W. Pauli, “Contributions mathématiques à la théorie des matrices de Dirac” (Mathematical contributions to
the theory of Dirac’s matrices), Annales de l’Institut Henri Poincaré 6, n. 2 (1936) 109–136.
7 [Eds.] Let

where {a, . . . , d} are all 2 × 2 matrices. Because {β′, αW} = 0, it follows

The set {1, σ} ≡ σµ is a complete basis for 2 × 2 matrices, so we can write a = aµσµ, and similarly for {b, c, d}.
Then it is easy to show that aµ = dµ ≡ 0, and bi = ci = 0. That is, only b0 and c0 are non-zero, so b = b01 and c =
c01. Finally, in order that β′2 = , we have to have c0 = b0−1, so β′ has to have the form (20.56).
8 [Eds.] Video 20 ends here, at 1:02:18. Typically classes ran for 90 minutes, and occasionally for as long as 115
minutes.
9 [Eds.] By “bilinear product” we mean a linear combination of terms of the form (ψ (i))*ψ (j), i, j = 1, . . . , 4 over the
components of the spinor. There are (obviously) 16 such terms in all, and what we are doing here is collecting
together the linear combinations with specific tensorial behavior.
10 [Eds.] Coleman uses γ5 and γ5 interchangeably. Here only the lower index γ5 will be used.
11 [Eds.] The relations (20.123a) and (20.123b) can be checked explicitly, in the special case of p = (0, 0, pz ), from
(20.27), (20.33) and (20.42), but it’s tedious. For example, letting the indices on Dirac spinors and matrices run
from 1 to 4,

and
 Problems 11

11.1 For any p, find two independent positive frequency solutions (i.e., u’s, not v’s) of the Dirac equation that are
eigenstates of helicity, angular momentum along the direction of motion. (The solutions displayed in class are not
helicity eigenstates unless p points along the z-axis.) Express the four components of u as explicit functions of θ
and ϕ, the polar angles of the direction of motion. H INT: The helicity operator commutes with rotations.
(1997a 10.1)

11.2 The following identities are easy to see. For the last two, use the cyclic property of the trace, Tr(ABC) =
Tr(CAB):

Carry on. Compute Tr( ), Tr( ), and the trace of up to four slashed vectors and one factor of γ5. The last
computation, of Tr( γ5), will involve ϵαβµν. Just to make sure we are all working with the same sign
conventions, choose ϵ0123 = +1. (You can find the answers to these in any relativistic quantum theory text, but it’s
more fun, as well as more instructive, to work them out yourself.)
(1997a 10.2)

Solutions 11

11.1 Start with the helicity eigenstates, (20.27) and (20.33), for momenta in the z direction,

For momentum in an arbitrary direction (θ, ϕ), the unit vector is given by

To get from = to = (sin θ cos ϕ, sin θ sin ϕ, cos θ), first we rotate about the y axis by θ, and then we rotate
about the z axis by ϕ. From (20.7):

So the rotation operators (18.23) we need are

Then
and

The spinors up(1) and up(2) have positive and negative helicities, respectively, because the original upz(1) and
upz(2) had those helicities, and the helicity operator commutes with the rotation operators.

11.2 First, it is easy to see that the trace of the product of three gamma matrices is zero: 1

the third equality following from the cyclic property of the trace. By the same argument, the trace of the product of
an odd number of gamma matrices is zero. Thus

For an even number of gamma matrices, we use repeatedly the identity

The algorithm is simple: Using this identity, start with the rightmost gamma, work it through to the leftmost
position, then use the cyclic property to return it to its original position. For example,

so that

and hence

Let’s do this for four gammas:

The last term is the same as the original. Move it to the left-hand side of the equation, and divide by 2, to obtain

Then

Now we come to γ5:

We have just shown, though, that the trace of four gammas must have two indices in common to have a non-
vanishing trace. But there are no repeated indices in γ5. So its trace is zero:

(This is also evident from its explicit form in the Dirac basis (20.102).) Notice that γ5 is itself the product of an even
number of gammas, so the trace of γ5 times an odd number of gammas vanishes.

What about the trace of γ5 with two gammas? If in γ5γµγν, µ = ν, then the product reduces to ±γ5, and that
trace vanishes. Say that µ ≠ ν. Pick a value of α different from both µ and ν. Then (no sum on α)
So the trace of γ5γµγν vanishes for all choices of µ and ν:

Finally, what about four gammas with γ5? Clearly, we need to have all four gammas be different. If any two are the
same, then we have again two gammas with γ5, whose trace vanishes. And if all are different, then their product is
nothing but ±iϵµνρσγ5. That is,

We can determine the sign by looking at γ5γ0γ1γ2γ3 = iγ5γ5 = i, so

To sum up, the trace of γ5 with fewer than 4 gammas is zero; and with four, it’s an expression proportional to
ϵαβµν.

1 [Eds.] Call this the “gamma-5 trick”.

21
The Dirac Equation III. Quantization and Feynman Rules

We will now canonically quantize the free Dirac Lagrangian. After that, we’ll consider simple interacting models,
and construct the Feynman rules for a theory involving fermions. We will refer to these fermions for the time being
as nucleons, though the formalism will hold good for any m ≠ 0, spin- particle, including electrons. Our interacting
theories will involve ψ, , and a scalar field ϕ. We considered a similar theory in Model 3, but with spinless
nucleons. In that theory, we had only three fields to worry about: ϕ, ψ and ψ ∗. We now have nine: 4 components
for ψ, 4 for , and ϕ. So we’ll have more combinatorics to juggle, and we’ll worry about some minus signs
due to Fermi statistics.1

21.1Canonical quantization of the Dirac field

From the Dirac Lagrangian,

we obtain the canonical momentum

The Dirac spinor ψ has four components {ψ a}, a = 1, . . . , 4, and each has its canonical momentum {πψ a}. These
four pairs form a complete and independent set of initial data, since the Dirac equation is linear. Following our
usual program, we should impose the equal time commutation relations (4.47)

We will see, however, that Fermi statistics require that these conditions be modified. The Hamiltonian is (4.40),
the last equality following from the Euler–Lagrange equations (4.26). We express the Fermi fields in a manner
analogous to the expression (6.24) for complex scalar fields,

Each Fermi field ψ contains two spinor components up(r) for positive frequency solutions, and two for negative
frequency solutions, vp(r). The operators bp(r) multiplying the positive frequency solutions, in analogy with the
charged scalar field expression, annihilate nucleons. (See §6.1.) Then the operators cp(r)†, multiplying negative
frequency solutions, must create antinucleons. Of course ψ †(y) is nearly the same, except that the operators and
spinors appear as their adjoints, the signs of the exponentials are reversed, and x is replaced by y:

The operators bp(r)† create nucleons, and the operators cp(r) annihilate antinucleons. It’s hard to keep these
operators straight, so here’s a chart to summarize them:

What are the commutation relations for the creation and annihilation operators? To avoid having to play
around with Fourier analysis, use the ansatz

The functions B(p) and C(p) are to be determined. We’ll assume all other commutators between the b’s and c’s
are zero. Then automatically (21.2) and (21.3) are satisfied, and

From the completeness relations (20.123a) and (20.123b) (multiplied by γ0 to convert and to u† and v†,
respectively),

Changing p → −p in the second integrand of (21.10), we obtain the canonical commutation relations (21.4) if we
set B(p) = C(−p) = ±1; the terms proportional to p and m cancel, and

But not all is well. Let’s compute the energy:

where we’ve used the spinor relations (20.108a), (20.108b) and (20.118). Then
No matter which sign we choose, this expression is not positive-definite, and hence not bounded below. If we
choose the plus sign, the antinucleons carry negative energy. This is a mess. We didn’t run into this problem in
either the charged (6.22) or uncharged (4.63) scalar cases, where the bilinear combinations of creation and
annihilation operators appear with positive signs, because the Hamiltonian for scalar fields is quadratic in the
derivatives (4.56), while that for Dirac fields is linear in the derivatives (21.14).

The way out of these troubles was found by Jordan and Wigner.2 We know that in ordinary quantum
mechanics, multi-particle wave functions describing fermions have to change sign upon interchange of two
particles, to enforce Pauli’s exclusion principle. This antisymmetry suggests that we adopt the following scheme:
We divide all quantities into two classes, Bose and Fermi. Bose fields are to be quantized as usual, with
commutators. Fermi fields, on the other hand, are now to be quantized with anticommutators (20.5). We will
assume that a Bose operator and a Fermi operator always commute with each other. That is, let Bose-type
variables and their conjugate variables be denoted {qa} and {pa}, and let Fermi-type variables and their conjugate
variables be denoted {θa} and {πa}. The new commutator rules are

Now you may be concerned that changing commutators to anticommutator is going to produce unwanted side
effects somewhere. In fact, I have arranged things so that all our previous manipulations go through just as well
for anticommutators as for commutators. Though we’ve determined B(p) = C(p) = ±1, we don’t yet know what sign
to choose. Consider

for some test functions fa(x). Then for a state |ϕñ we have

but we also have (using (21.4), but with the anticommutator)

Only the + sign is consistent. So we must take B = C = +1. The revised equal time commutation relations for ψ and
ψ † are

From (21.22b) we find unambiguously

and from (21.22a), we find that all other anticommutators vanish; in particular,

Returning to the energy calculation, we now have, choosing the plus sign,

As we’ve done before (see the discussion following (4.64)), we can redefine the zero of the energy, and discard
the infinite constant. The energy of the Dirac field becomes

This is positive definite, and the problem of energy unbounded from below is solved.

Now, what about the states of the Dirac theory? Are they like those in the Fock space of the scalar theory (see
Chapter 2), with a 0-particle subspace, a 1-particle subspace, and so on? For pedagogical simplicity, suppose
that the theory only has particles, created by the Fermi operators bp† and annihilated by bp, without antiparticles
(and their associated operators, cp† and cp). We’ll also forget about spin, so our b operators will not carry the
spinor indices. The only non-zero anticommutator is

and the Hamiltonian is

Using the identity

we find bq is an energy-lowering operator:

and by the same argument bq† is an energy-raising operator:

The equations for a single Dirac particle state are the same as for the scalar field (§2.4):

The multi-particle Dirac states look the same formally as multi-particle scalar field states:

but they differ in an important respect. A two-particle boson state (2.59) is symmetric under interchange

because the boson creation operators commute. But a two-particle fermion state is antisymmetric, because the
creation operators anticommute:

This antisymmetry enforces the exclusion principle: if q1 = q2, then

The square of any Fermi creation or annihilation operator is zero. The energies of the multi-particle states work
out as we expect:

The properties of Fermi fields affect the extent to which they can be observed. Observables made from Bose
fields commute at equal times, and so by Lorentz invariance commute for all spacelike separations. On the other
hand, for Fermi fields ψ a at equal times,

Observables that do not commute at spacelike separations are unphysical, so ψ(x) is not an observable.
Observables can only be made from products of an even number of Fermi fields, which do commute at spacelike
intervals. Moreover, consider the behavior of Fermi fields under rotation. Under a rotation by 2π, a Fermi field
changes sign. This is not the behavior of an observable. No meter on any experimental apparatus has ever given
a different reading when the experiment was rotated by 2π. In some sense, a Fermi field ψ is the “square root” of
an observable.

Finally, let’s look at the classical limit of a Fermi field, as ħ → 0. There are in fact several limits. First, consider
N bosons, all having the same energy ħω, in a box. The system has an energy E = Nħω. If we keep N, E and |p| =
ħk fixed, and let ħ → 0, then ω → ∞ and k → ∞. The wavelength λ = 2π/k of the quanta goes to zero, and there will
be no diffraction. This limit corresponds to the quanta acting like classical particles. Alternatively, we could keep all
the variables (E, ω, k, and λ) except N fixed. The limit ħ → 0 corresponds to N → ∞. In this limit, quantum
granularity is lost, but the quanta still exhibit wave behavior. We could repeat the first limit with fermions, but not
the second. In classical electromagnetic theory, we can have many photons in each mode. We can’t do this with
fermions, because of the Pauli exclusion principle. There is no classical wave behavior for them.

Formally, the limit ħ → 0 in the commutator algebra leads to a classical theory of commuting boson fields. But
for fermion fields, in order to have agreement even at O(ħ0), we need c-numbers that anticommute. In the
literature, anticommuting quantities (even if they are not c-numbers) are called Grassmann variables, or more
formally, “elements of a Grassmann algebra”.3 Without such numbers, we can’t even preserve the Heisenberg
equations of motion in the limit ħ → 0.

21.2Wick’s theorem for Fermi fields

We’re going to modify Model 3, our scalar meson–nucleon theory (8.6), treating the nucleons as Dirac fields. The
Lagrangian takes the form

where the dots indicate counterterms, which we’ll neglect for now. The matrix Γ will be 1 if the field ϕ is a scalar,
and iγ5 if ϕ is a pseudoscalar. We’ll use perturbation theory to study the interactions, so we’ll need Dyson’s
formula, Wick’s theorem and all that.

The first hurdle is time ordering. Suppose a point P at coordinates x is outside the origin’s light cone, so that
x2 < 0. In the solid coordinates as shown in Figure 21.1, x0 > 0, and for a scalar field ϕ,

Figure 21.1: A point P in two coordinate systems

In a different frame of reference, indicated by the dashed coordinates, we have x0 < 0, and so

There is no problem when x2 < 0, because for spacelike separations, [ϕ(x), ϕ(0)] = 0, and there’s no problem for
x2 > 0, because then there’s no ambiguity about which is earlier. On the other hand, suppose we were considering
Fermi fields, ψ a(x). Then the time ordering is not Lorentz invariant. In the solid frame,

but in the dashed frame,

because {ψ a(0), ψ b(x)} = 0 for spacelike separations.4

The way to patch this up is to put an extra minus sign into the definition of the time ordered product whenever
the number p of permutations required to turn a product of Fermi fields into a time ordered product is odd. Define
time ordering (7.35) on Fermi fields as

For example,

Dyson’s formula (7.36) will not be a problem, because H I will be quadratic in Fermi fields, and so all permutations
will involve even powers. Normal ordering needs the same prescription. For products involving Fermi fields,

where again p is the number of permutations needed to put all the Fermi creation operators to the left of all the
Fermi annihilation operators. For example, if ψ 1 and ψ 2 are Fermi fields,

where ψ (+) and ψ (–) are defined analogously to ϕ (+) and ϕ (–) in (3.33):

As with time-ordered products, Fermi fields anticommute within normal-ordered products:

With the generalized time-ordered and normal-ordered products, Wick’s theorem (8.28) can be proved for
Fermi fields in exactly the same way as it was for Bose fields. We won’t do that here, but we will show that the
theorem holds for two Fermi fields ψ and χ. If, following (8.20), we define

making due allowance for anticommutation, then Wick’s theorem says

We’ll postpone the calculation of the contraction itself. We’ll prove the theorem by cases. Suppose x0 > y0. Then,
from (21.51),

For x0 > y0, T(ψ a(x) b(y)) = ψ a(x) b(y), so

and, from (21.48),

Subtracting : ψ a(x) b(y) : from T(ψ a(x) b(y)) gives

So (21.52) is true if x0 > y0. If y0 > x0, all three expressions pick up an overall minus sign. From (21.51),
but from (21.46) and (21.50), respectively,

and the demonstration goes through exactly as before. That establishes the Fermi field version of Wick’s theorem,
at least for two fields.

21.3Calculating the Dirac propagator

In the sort of theories we’re looking at, such as (21.40), the free meson propagator is unaltered by the presence of
Fermi fields. It’s the contraction of two Fermi fields we need to compute:

Three remarks need to be made. First, by reasoning parallel to that in the scalar case (pp. 156–157), the
contraction of two Fermi fields is a c-number. The difference between the two orderings is just an anticommutator,
which is always a c-number, because of the minus signs we’ve inserted into our definition of time ordering and
normal ordering. Second, in order to keep from cluttering equations with indices, I will adopt this convention:
Whatever the order of the operators, the order of the indices in this expression will be such that the index is
always on the right. That way we won’t get a silly expression that is a 4 × 4 matrix for one ordering and a 1 × 1
matrix for another ordering. Third, because we’ve defined both the time-ordered product and the normal-ordered
product to be antisymmetric in interchange of the two operators,

exchanging the two operators in the contraction gives us a minus sign.

We have to write down once again the expression (21.6) for the free field, ψ(x):

And of course (y) is almost the same thing:

In such expressions, p0 is always equal to Ep.

Taking the vacuum expectation value of (21.60), the contraction is equal to the vacuum expectation value of
the time ordered product alone, because the vacuum expectation value of the normal-ordered product vanishes:

We will compute this, for x0 > y0 and for x0 < y0, and then join the two results together. For x0 > y0 the is on the
right, where only the creation part, proportional to bp′ (r′)†, is relevant; the ψ is on the left where only the annihilation
part, proportional to bp (r), contributes. So I obtain

But from (21.23a),

so
We have a wonderful identity, (20.123a), for the spinor sum: it is + m. And therefore we can write this whole
expression (for x0 > y0) as

(I put an x on the derivative to show differentiation with respect to x and not y.) This integral is the same one we
encountered while evaluating the scalar propagator (S1.13). If we imagine a hypothetical scalar field φ of mass m
(which is not the scalar field coupled with the Fermi field in this theory), we can write, if x0 > y0,

because the second integral in (21.68) is just equal to the expectation value of the time-ordered product of some
scalar field, φ(x). Equivalently

I now turn to the case y0 > x0. The order of the operators is reversed:

Of course the order of the matrix indices is not reversed; otherwise the integral is exactly the same, though the
integrand looks different:

Swapping x and y makes the creation and annihilation parts change places, so I get the cp(r)’s exchanged for the
bp(r)’s, and I get the sum on our vp(r)’s instead of the up(r)’s, and the sign of the exponent changes. The spinor sum
is, by (20.123b), ( − m), and thus

because the integral is unchanged under p → −p. That is, the expression á0|T(ψ(x) (y))|0ñ is exactly the same for
y0 > x0 (21.73) as it is for x0 > y0 (21.68); for all times,

From (8.23), we can write down immediately the Fourier transform of the scalar field contraction:

The effect of the (i x + m) is to put ( + m) in the numerator.

Thus the analog of the scalar field propagator i/(p2 − m2 + iϵ) is, for a fermion field,

Though we’re dealing with a four-component field, it has only twice as many physical degrees of freedom as a
charged boson field. The field has four components, but there are only two spin states for the particle, and two for
the antiparticle. There are actually only two kinds of particles we can exchange, so we should have some kind of
projection operator for the exchange of those particles, at least as p2 approaches m2, and we pick up the one-
particle states. And as can be seen from (20.124a) and (20.123a), we’ve got the projection operator on the
positive frequency states in the numerator.

Another way of understanding the propagator is to write it in an alternative form which you will frequently find
in the literature. Since the only function of the iϵ is to tell us how to control the pole, we can put a minus iϵ in the
numerator with no loss of generality:

Because m is a positive number, (m − iϵ)2 puts the pole in the same place m2 − iϵ does. We can thus rewrite the
denominator as

Then5

(We can be a little cavalier here about matrix manipulations, because − m + iϵ commutes with + m − iϵ.) In this
form, the Feynman propagator for the Dirac theory very closely parallels the Feynman propagator for a scalar
theory. In a scalar theory, the free Klein–Gordon equation in momentum space involves the operator p2 − m2. The
scalar Feynman propagator is i over this operator with the pole difficulty resolved by giving the mass a small
(negative) imaginary part. The Dirac equation in momentum space involves the operator ( − m). The fermion
Feynman propagator is i over this operator with the pole ambiguity resolved by giving the mass a small (negative)
imaginary part. In short,

We shall see later on, when I talk about quantization through functional integration, that the propagator is always,
in a sense, the inverse of the operator D that appears in the free Lagrangian ϕDϕ (with i∂ → p).

21.4An example: Nucleon–meson scattering

Before writing down the Feynman rules in such a theory it’s probably best to see how things work out by
evaluating a particular diagram and watching how all the various factors fit together. We’ll consider the Lagrangian
(21.40), which describes a free Fermi field, a free meson field and an interaction between them:

where Γ is either 1 or iγ5. In the former case the theory is parity invariant if ϕ(x) is a scalar; in the latter case it is
parity invariant if ϕ(x) is a pseudoscalar. For what I am going to do now, the choice of Γ is irrelevant. Let’s consider
a typical scattering process, for example nucleon plus meson goes into nucleon plus meson:

A Feynman diagram (drawn in two ways) that contributes to this process to lowest order is shown in Figure 21.2.
We adopt the same diagrammatic conventions as in the scalar model, with the spinor charged nucleon field
replacing the scalar charged nucleon field we had before (§8.3). The incoming nucleon and meson are
characterized by the momenta p and q, respectively; the momenta p′ and q′ denote the respective outgoing
momenta. And of course the nucleon is in some spin state. We’re constructing S-matrix elements between states
of definite spin, so I give an index r for p, and an index s for p′ where {r, s} equals 1 or 2, telling you whether the
4
nucleon is spin up or spin down. Let’s look at the order g2 term in Dyson’s formula, S = Te−i d x HI :
 Figure 21.2: A diagram for lowest order nucleon–meson scattering

The relevant terms in the Wick expansion corresponding to this diagram are

where the subscripts 1 and 2 indicate that the functions depend on x1 and x2, respectively. The picture for the
second term looks identical to the first picture, except for an interchange 1 ↔ 2 of the dummy variables. The
second operator is the same as the first, and serves only to cancel the 2!; the two pictures are the same diagram
written twice.

Let’s write down the S-matrix element between the final state and the initial state coming from this term in the
Wick expansion:

Just as in the scalar case, we use relativistically normalized states (1.57), so we don’t have to keep track of the
factors of (2π)3/2 and 1/ . Those are automatically taken care of as part of the density of states (11.58) in our
rules for turning S-matrix elements into cross-sections. We then get the integral over d4x1 d4x2 and a bunch of
exponential factors. First, ϕ 1 is annihilating a meson in the initial state, so I obtain e−iq⋅x 1, and ψ 1 is annihilating the
initial nucleon, so I have e−ip⋅x 1. Likewise everything is being created at x2, so I have positive exponential factors
for x2: eiq′⋅x 2eip′⋅x 2. We can drag ϕ 2 as we please inside the normal-ordered product, since it commutes with the
ψ’s. Then I will have an integral over some momentum k that occurs in the Fourier expansion of the propagator,
the contraction of 2 and ψ 1.

Now we have to deal with the matrices and the spinors. Let’s go in the order in which the integral is set up,
from right to left. The annihilation of a meson carries nothing besides the exponential. Annihilation of the nucleon,
however, carries the factor here of up(r). That takes care of the first factors and the initial state. Then as we go
along, there’s a Γ. Then we’ve got the contraction, i/( − m + iϵ), followed by another Γ, and (s) from the final
p′
state:

Figure 21.3: Antinucleon–meson scattering to lowest order

The x integrals are trivial, and as usual give us a (2π)4 times an overall energy momentum conserving delta
function, just as in the scalar case. The k integral is also trivial here, because of the delta functions:

or, using the notation of (10.27),

We have to be careful about the sign of the intermediate momentum, k. Should k = p + q or k = −(p + q)? The
fermion propagator is not invariant under change of sign of p, as the scalar propagator is. But it is clear from
(21.86) that the plus sign is correct here.

This amplitude (21.88) is the sort of generalization you would expect even if you hadn’t gone through the
derivation. When you have a set of four fields, you also have a bunch of 4 × 4 matrices as you pass through a
vertex and propagate something. Of course a matrix element is not a matrix, it is a number. So to make it a
number, you need a column vector like up(r) for the initial nucleon, and a row vector like p′(s) for the final nucleon.

Before I write down what happens in general, let me consider a second process nearly the same as
nucleon–meson scattering, antinucleon–meson scattering:

The diagram is shown in Figure 21.3. Many things are the same. The principal change is this. In the left diagram in
Figure 21.2, I think of the vertex on the right, where the nucleon and the meson are annihilated, as point 1, and the
vertex on the left as point 2. The diagram in Figure 21.3 comes from exactly the same term (21.84) in the Wick
expansion. Now however the operator needed to annihilate the initial antinucleon is found in 2, and the field ψ 1
creates the final antinucleon. So I think of these vertices, read right to left, as 2, 1. Matrix multiplication still goes
from 1 to 2 along the line. We obtain

At the left-hand side of the matrix we have pr, the factor associated with the annihilation of the initial antinucleon,
and then we have a Γ. The propagator’s denominator is −( + ) − m + iϵ, because we switched x1 and x2, and so
we’ve changed k into −k in the energy–momentum conserving delta function. There’s another Γ, and then there is
the final antinucleon being created, which gives us vp′(s). But there is a new feature, a factor of −1 coming from
Fermi statistics. In the Wick term (21.84), the field that annihilates the initial antinucleon, 2, is over on the left,
where it shouldn’t be. To put things in the right order, I have to switch the ψ 1 and 2. Within a normal-
ordered product, the switching throws in a minus sign.

These two examples contain practically all the novel features we encounter in the Fermi theory. So I can now
write down the Feynman rules for theories involving fermions.

21.5The Feynman rules for theories involving fermions

I’ll list the rules in three sections. First, I’ll tell you what the factors are. Then I’ll give the rules for handling the
matrices. Finally I’ll tell you what to do about the terrible Fermi minus signs.

Feynman rules for theories with fermions I. Factors

1.For every . . .Write . . .

(a)internal meson line

(b)internal nucleon line

(c)vertex −igΓ

2.Ensure momentum conservation at each vertex: (2π)4δ(4)( pout − pin)

3.Multiply by and integrate over each internal momentum.

4.Spinor factors:

•For every incoming nucleon, write a u.

•For every outgoing nucleon, write a .

•For every incoming antinucleon, write a .

•For every outgoing antinucleon, write a v.

 We still have to take care of all the matrix and spinor factors. First I’ll state the rules, and then I’ll explain them
with examples.

Feynman rules for theories with fermions II. Assembling the pieces

1.Along a fermion line:

•Starting with the arrowhead, follow each fermion line backwards

through the diagram, assembling factors as you go.

2.For a closed fermion loop:

•Include a factor of (−1).

•Take the trace of the product of Dirac factors.

Leaving aside the counterterms for the moment, the factors are pretty much the same as in Model 3 (box, p.
216.). We assume that the initial state is on the right. The momentum orientation does not affect the meson
propagator, because q2 is (−q)2, but it does affect the fermion propagator. We’ll orient the momentum in the same
direction as the arrow on the line for nucleons, and in the opposite direction for antinucleons. If you happen to find
it convenient in a particular graph to orient a nucleon’s momentum q the other way, that’s fine by me, but then you
must write the propagator as i over (− − m + iϵ). Every vertex, for example with q and p coming in, p′ going out,
gives us a factor −igΓ(2π)4δ(4)(p + q − p′), exactly the same as in the scalar theory, except that it’s now a matrix
because of the presence of Γ.

Some row vectors and column vectors are associated with initial and final fermions. For every incoming
nucleon, I have the u appropriate to the nucleon’s state. If the nucleon happens to be in one of our standard states,
then it is that up(r). If it’s not, the state will be some linear combination of the u’s. With every incoming antinucleon I
have associated a , as shown in the last example. With every outgoing nucleon I have a and with every
outgoing antinucleon I associate a v. This is nothing more than a reflection of the fact that ψ annihilates nucleons
and creates antinucleons, but annihilates antinucleons and creates nucleons.

Because fermions appear bilinearly in the Lagrangian (we’ll soon see that quartics are ruled out), a fermion
line either goes all the way through a graph, or it appears in a loop. We have matrices associated with fermion
lines, and row or column vectors associated with incoming or outgoing fermions. It doesn’t matter which way the
fermion line is going through the diagram. As we habitually write from left to write, we’ll start at the head of the
arrow and work against the arrows. At an incoming antinucleon, we write down a ; at an outgoing nucleon, we
write a . The next thing you encounter is a vertex. Write down the matrix Γ for the vertex. Then you get a
propagator associated with the internal line, followed by another vertex with another Γ. This may repeat a number
of times. When you get to the tail of the line, you arrive at either an incoming nucleon, and write a u, or an outgoing
antinucleon, and a v. That’s the order in which things come out in Wick’s theorem.

Now, to assemble the pieces.

EXAMPLE. A linear fermion graph

Figure 21.4: A linear fermion graph

Consider the graph shown in Figure 21.4. Numbering the various factors, the amplitude for this graph is

EXAMPLE. A closed fermion loop

Consider a completely closed fermion loop, as in Figure 21.5. The term in Wick’s theorem that gives us this
closed loop is, ignoring the factors of (−ig),
 Figure 21.5: A closed fermion loop

This is an O(g4) diagram describing two mesons → two mesons scattering. The factor of interest in this
contribution to the process is

To put the contraction between ψ 4 and 1 in Fermi fields (21.75), I have to move ψ 4 past seven other Fermi fields,
so there’s an extra minus sign. But this isn’t all. Before taking the contractions, look at the Dirac indices
(summation over repeated indices implied):

The term between the colons is Mhh = Tr(M) for the matrix M = (ψ 4 1Γψ 1 2Γψ 2 3Γψ 3 4Γ).
Thus the contribution is

With a closed loop it doesn’t matter where you start multiplying the matrices: The trace is invariant under
cyclic permutations of the matrix factors. Start anywhere, and working against the arrows, write down the vertex
and propagator matrices until you get back to where you started. In the product, you will have the makings of a
contraction, but in the wrong order: the on the left and the ψ on the right. And as I remarked (21.61), that is
minus the contraction in the standard order. So our first minus sign rule is: For every closed fermion loop, include
a factor of (−1). We will check this rule for consistency when we compute the meson self energy in this theory.
There the factor of (−1) from Fermi statistics will be very important for the closed nucleon loop that occurs. As you
will see, this factor is needed to guarantee that the imaginary part of the self energy has the proper sign,
consistent with the spectral representation. If it weren’t there, we would obtain an insane answer for the meson
self energy.

In general, though, it can be tricky to get the signs right. We found an extra minus sign in antinucleon–meson
scattering as compared with nucleon–meson scattering, because we had to switch around the operators. It is
possible to give a sequence of rules for the result of switching around operators in the general case, but it’s
awkward. Say we have an initial state of 32 nucleons and 47 antinucleons, and a final state of six nucleons and
seven antinucleons. You’ve got to establish rules about what you mean by a six-nucleon, seven-antinucleon state;
you have to specify in what order they are created (you’ll see why this is relevant when we work through an
example). I will just make the simple statement that, as we see already from the string (21.92), the normal-ordered
operators (for each particle) always come in the order : ψ:, where the ψ is associated with the tail of a line and
the with the head of a line. That’s the only fact you have to remember. Whether you’re going to use it to
create or annihilate, the operator ψ is always associated with the beginning of the line; the operator
following is associated with the end of the line. However many lines traverse the diagrams in different directions,
making hairpin turns, it doesn’t matter in what order I put strings of ψ’s, because pairs of ψ’s always
commute with each other. You do have to look to see if the annihilation operators and creation operators are in the
right places, or if you have to switch them around, depending upon whatever ordering you have adopted for the
creation of the initial state. Once you get the knack of it, these rules are not difficult to work with. Instead of saying r
and s, I’ll specify the spinors as u and u′, which are linear combinations of u(1) and u(2). The internal momentum is
fixed by energy momentum conservation. On the left it is p + q running along the arrow, and on the right it is p′ − q.
To make things definite, I will choose Γ = iγ5.

Figure 21.6: Meson–nucleon scattering to lowest order

 Let’s write down the invariant amplitude for these two diagrams.

The order of matrix multiplication is with the head of the arrow on the left, the tail of the arrow on the right, a for
every outgoing particle, a u for every incoming particle. The second diagram, on the right, contributes nearly the
same as the first. I have no Fermi minus signs; the expression (21.84) I get from Wick’s theorem is ψ, with ψ
and in the right positions to annihilate the initial nucleon and to create the final nucleon, respectively.

We can simplify this. In this kinematic region we don’t need to keep track of the ϵ’s, and we can rationalize
the denominators:

I can get rid of the γ5’s in a flash because γ5 anticommutes with and . So I just drag it through, and use γ52 = 1:

This expression in fact simplifies enormously. This is typically what happens in Feynman calculations. The
calculations with spinors are horrible, but not so horrible as one would think naively, because of the spinors’
properties. Here in the first term we have acting on a free particle spinor on the right, which I remind you carries
momentum p. And therefore u = mu, (20.110a). Likewise in the second term ′ ′ equals ′m, (20.113).
Then the ’s cancel the m’s and we’re left with

It’s rather pleasant once you get the knack of it, like doing a crossword puzzle. You just move things around to
eliminate some factors when they’re hitting solutions of the free Dirac equation.

Here’s a second example, nucleon–nucleon scattering:

Figure 21.7: Nucleon–nucleon scattering to lowest order

The Fermi minus signs are a bit more complicated, though the Γ algebra is considerably simpler. I will write down
the expression for the amplitude without determining the Fermi minus sign factors, which I’ll just write as “(sign 1)”
and “(sign 2)”, which are going to be equal to +1 or −1. Indeed, I don’t know what the factors are before I specify
the initial state.

Again there’s no need to include the iϵ factors.

First, let’s talk through the graph on the left. From the top line, (−ig) ′1(iγ5)u1. That’s all there is to it; there’s
just a vertex, there are no internal propagators. From the bottom line, (−ig) ′2iγ5u2. The vertical line
represents the meson propagator, i/((p1 − p′1)2 − µ2). The second graph gives a similar expression. To determine
the Fermi signs (1) and (2), we have to use the magic ψ rule. I will label creation operators for the initial state
simply as b1† and b2† to avoid writing a lot of p and (r) indices. Let’s take the initial state to be

That is, nucleon number 2 is created first, and then nucleon number 1. The final state should be
 We can now work out what happens using the ψ rule. We always have the operator associated with the
head of the line to the left of the operator associated with the tail of the line. Let’s do the left graph in Figure 21.7
first. The tail of the top line is annihilating particle 1. The head of the line is creating particle 1′ and, from left to
right, head goes before tail. From the Wick expansion, we have

Likewise on the bottom line, head goes before tail, and we have
If 1 equals 1′ and 2 equals 2′, the final state is the same as the initial state, not minus the initial state. Then the
adjoint gives us

That’s the ψ rule: heads before tails. The net result for the S-matrix element involves the factor

It doesn’t matter in which order I write the two pairs of operators;

Permuting the operators gives an overall plus sign.

But the operators in (21.106) are not in the right order to annihilate and create the initial state (21.101) and the
final state (21.103). The operator b1 is in a great position to kill the incoming nucleon 1, but b1′† is not all the way
over on the left to create the outgoing nucleon 1′. Therefore I rearrange it by bringing b1′† over to the left. That
requires two permutations, so it’s an overall plus sign:

Now everything is in great position: b1 can knock off b1†, b2 can then eliminate b2†; b′1† can take care of b′1, and
similarly b′2† can cancel off b′2. In this case, the Fermi sign, (sign 1), equals +1. (It is not really as tedious as this.
After you’ve gone through it two or three times, you can do it by eyeball.)

What about the second case, the rightmost diagram in (21.106)? Here we have on one line nucleon 1 being
annihilated and nucleon 2′ being created. On the other line we have 2 being annihilated and 1′ being created.
Using the ψ, head-before-tail rule, the corresponding Wick term is

Again I want to put the operators in the correct position,

The b1 is still in the right place, and this time, so is the b′1†. But the b2 and b′2† need to switch places. When we
permute the operators, we get a minus sign: (sign 2) = (−1).

Unless one were extraordinarily clever, one could not have guessed the absolute sign of either of these two
terms. One can, however, easily guess that the relative sign had to be −1. The statement that the relative sign is
−1 is simply the statement that if one interchanges all the “1” and “2” labels, the total amplitude changes by a sign,
just as we would expect for a scattering process involving Fermi particles. This is frequently a useful rule.
Sometimes you don’t have to work out the absolute sign if all you’re going to do is square the amplitude at the end
of the computation. Frequently Fermi statistics are good enough to tell you the relative signs between the various
graphs. This is not always true. It doesn’t work for example in meson–meson into nucleon–antinucleon scattering.
But sometimes it’s enough.

21.6Summing and averaging over spin states

I have now told you all there is to say about the actual computation of S-matrix elements between particles in
definite spin states. If you are interested in particular spin states, say states of definite helicity for the initial and
final fermions, all you have to do is plug in the appropriate u’s and ’s for the initial and final particles and evaluate
the matrix element. However there is a large class of experiments in which one is either uninterested in, or unable
to measure, the spin of the initial or final states. One frequently does experiments with unpolarized beams of
particles, and in which we choose not to measure the spin of the final nucleon. In such cases, one is frequently
interested in cross-sections which are summed over final spins (since your apparatus responds whatever the final
spin is), and averaged over initial spins (because you have a statistical distribution of initial spins in the incoming
beam).

As a specific example, let’s return to nucleon–meson scattering,

We showed in (21.98) that the scattering amplitude was some function F(E, θ) times a bilinear spinor expression,
′ u:

where E is the center-of-mass energy and θ is the scattering angle. We want to compute |Afi|2, say between a
definite polarization state r and a final one s. The initial spinor is characterized by momentum p, and the final
spinor by p′. For these particular states,

What we want to do is square the amplitude, sum on r and divide by 2 (because we’re averaging over the initial
spins, two in number), and sum over s (the final spin). We use these facts: the function F(E, θ) is independent of
the spins r and s; is “self-bar”, because qµ is real and the gamma matrices are self-bar (20.84); and the Hermitian
adjoint of the bilinear spinor product is the same as its Dirac adjoint (20.66):

Then

Now we borrow a cunning idea due to Feynman,6 which has saved generations of physicists from having to
compute sixteen 4 × 4 matrix elements with explicit spinors and sum them all up. That’s what they used to do,
when they were doing this sort of computation back in the 1930’s. He observed that a number can be thought of
as a 1 × 1 matrix, and that a 1 × 1 matrix is equal to its trace. Thus

The trace is invariant under cyclic permutation of factors, so I can write (21.112) as

I moved the factor up(r) from the rightmost position to the leftmost to make use of the wonderful completeness
relation (20.123a):

Here is the redemption of that homework problem7 on traces of Dirac matrices. You might have wondered why
you were working out all those dumb trace identities. Recall that the trace of an odd number of γ matrices always
vanishes. The traces of a product of two and of four slashed quantities are given by the identities (S11.13) and
(S11.16), respectively:

So we’re all set up for completing the computation:

This can be simplified somewhat. The meson is on its mass shell, and therefore q2 = µ2. So

This expression, by trivial kinematic exercises that I won’t bother to go through, can be reduced to functions of the
only two invariants, the center-of-mass energy, E and the center-of-mass scattering angle, θ.

Now that we have our general formalism, we can discuss charge conjugation, time reversal, and TCP
invariance. We’ll do that next time, and then begin renormalization for theories involving fermion fields.

1 [Eds.] The videotape of Lecture 21 does not begin until §21.3. To make matters worse, Coleman’s notes
covering the first two sections of this chapter are also missing. Thus these first two sections are based entirely on
Brian Hill’s and Peter Woit’s reliable notes, with some guessed-at interpolation by the editors.Caveat lector!
2 [Eds.]P. Jordan and E. Wigner, “Über das Paulische Äquivalenzverbot” (On the Pauli exclusion principle), Zeits.
f. Phys. 47 (1928) 631–651; reprinted in Schwinger QED. The anticommutator is given in their equation (36), p.
639; see also the chart on p. 640.

with canonical commutation relations, we encountered a disastrous contradiction (21.15) with the positivity of
energy. We succeeded when we used canonical anticommutators (if we chose (±) to be +). Much earlier we were
able to quantize the free charged scalar field,

with canonical commutators (if we chose (±) to be +). Attempt to quantize the free charged scalar field with (nearly)
canonical anticommutators:

where λ is a (possibly complex) constant.

Show that one reaches a disastrous contradiction with the positivity of the norm in Hilbert space; that is to say,
with (21.20):

for any operator θ and any state |ϕñ.

H INTS: (1) Canonical anticommutation implies that, even on the classical level, ϕ and ϕ * are Grassmann variables.
If you don’t take proper account of this (especially in ordering terms when deriving the canonical momenta), you’ll
get hopelessly confused. (2) Dirac theory is successfully quantized using anticommutators; the sign of the
Lagrangian is fixed by appealing to the positivity of the inner product in Hilbert space. If we attempt to quantize the
theory using commutators, we get into trouble with the positivity of the energy. The Klein–Gordon theory is
successfully quantized using commutators; the sign of the Lagrangian is fixed by appealing to the positivity of
energy. So it’s to be expected that we’d get into trouble, if we attempted to quantize the Klein–Gordon theory with
anticommutators, with the positivity of the inner product.
(1997a 11.1)

12.2. Compute the differential cross-section dσ/dΩ in the center-of-mass frame, to lowest non-trivial order in
perturbation theory, averaged over initial spins and summed over final spins, for meson–nucleon scattering in the
“scalar” theory discussed in §21.4,

Note: You are required only to compute dσ/dΩ, not the total cross-section, for this problem and the next.
(1997a 11.2)
4 [Eds.] Because the anticommutators (21.23a) and (21.23b) are the only ones involving the b’s and the c’s that
do not vanish, the anticommutators {ψ a(x), ψ b(y)} and {ψ a†(x), ψ b†(y)} vanish for all values of x and y, not only for
(x − y)2 < 0. The case {ψ a(x), ψ b†(y)} can be computed with the help of (21.11) and (21.12):

But we know (3.51) that iΔ(x − y) = 0 for (x − y)2 < 0, so the same is true for {ψ a(x), ψ b†(y)}.
5 [Eds.] In the literature, the Feynman propagator i/( − m + iϵ) for the Dirac field is often written as F( ) or SF( ).
Bjorken and Drell define this propagator as iSF( ); see Bjorken & Drell RQM, p. 93, equation (6.42).
6 [Eds.]
R. P. Feynman, “The theory of positrons”, Phys. Rev. 76 (1949) 749–759. See equation (36); “Sp” = spur,
German for trace. See also R. P. Feynman, Quantum Electrodynamics, W. A. Benjamin, 1962, Lecture 23, “A
method of summing matrix elements over spin states”, pp. 112–114. The technique seems to have been first used
by the Dutch theorist Hendrik B. G. Casimir (1909–2000), and is sometimes called “Casimir’s trick”: H. Casimir,
“Über die Intensität der Streustrahlung gebundener Electronen” (On the intensity of radiation scattered by bound
electrons), Helv. Phys. Acta 6 (1933) 287–305. See §4, p. 293; Griffiths EP, p. 251. Casimir’s autobiography
(Haphazard Reality: Half a Century of Science, Harper & Row, 1984) draws its title from a Bohr quote: “When
telling a true story, one should not be overly influenced by the haphazard occurrences of reality.”
7 [Eds.] Problem 11.2, p. 425.

Problems 12

12.1 When we attempted to quantize the free Dirac theory

12.3 The same for nucleon–antinucleon scattering in the “pseudoscalar” theory,

Note: Since you are only interested in cross-sections, all you need to know is the relative sign between the two
graphs; the absolute sign is irrelevant.
(1997a 11.3)

Solutions 12

12.1 From the Lagrangian

we derive the canonical momentum to ϕ(x),

Note that as ϕ(x) and ϕ *(x) are regarded as Grassmann variables, when we move the derivative ∂/∂(∂0ϕ) past
∂µϕ *, we pick up an extra minus sign. The question asks that we impose the (nearly) canonical anticommutation
relations

As usual, expand ϕ(x) in terms of annihilation and creation operators,

Then
We can invert these relations to solve for bp and b†p′:

Then

Similarly,

and

Consequently, á0|{bp, bp′†}|0ñ and á0|{c†p, cp′}|0ñ cannot both be positive, so the positive definite norm does not
hold if we attempt to canonically quantize a scalar field with anticommutators.

12.2 Using the Feynman rules for fermions (box, p. 443), we see that the vertex involves two fermion lines and a
meson line. The relevant Feynman diagrams are shown in Figure S12.1.

Figure S12.1 Graphs for lowest order meson–nucleon scattering

These two contributions add, because only bosons have been swapped;

Using (12.26) (averaging over the initial spins, and summing over the final spins),

because the scattering is elastic, so |pf| = |pi|. Using the Feynman rules, we find

For convenience, let’s define

Then

The differential cross-section becomes

Calculating the sum,

(In the fourth step, we use the result that the trace of an odd number of γ’s is zero.) We therefore have

In the center-of-momentum frame (with the incident nucleon in the x direction and the outgoing nucleon in the xy
plane),

and so

If desired, we could put these factors into (S12.17), to get dσ/dΩ in terms of p and the scattering angle θ. ■

12.3 The relevant Feynman diagrams are shown in Figure S12.2.

Figure S12.2 Graphs for lowest order nucleon–antinucleon scattering

 The total amplitude for this scattering process can be written as

the relative minus sign coming from the exchange of the fermion lines for p′1 and p2 between the diagrams. We
can write down the amplitudes as (letting up1(r) ≡ u1, etc.)

The overall sign of doesn’t matter here, because it’s going to be squared. So we can take

where

so

We need to average over both pairs of initial spins and sum over the final spins, which we do with the trace
theorems:

Let’s call this quantity . We can move the γ5’s past the gammas in the slashed momenta, and we find

The two middle terms are equal. (This is not obvious, but it’s so.) Using the trace identities,

As in the solution to 12.2, we have

because the scattering is elastic. In the center-of-momentum frame,

and so

These factors go into (S12.28), to get dσ/dΩ in terms of p and the scattering angle θ.

22
CPT and Fermi fields

We are going to discuss for a Dirac theory the famous discrete symmetries of nature: parity P, charge conjugation
C, and time reversal T. We’ve already talked a lot about parity (§6.3, §18.5, and §20.1; (18.50), (18.51), and
(19.5)), but I will say more. As always in a relativistic theory, it’s more convenient to discuss the product of parity
and time reversal, PT. I will also prove, in the same diagrammatic way I proved it for purely scalar theories (§11.3),
the CPT Theorem.
3 [Eds.]Hermann Grassmann (1809–1877), German schoolteacher and polymath, invented exterior algebra in his
Die Lineale Ausdehnungslehre (“The linear theory of extended magnitudes”) (1844; second edition 1862). The
term “Grassmann variable” for an anticommuting quantity may derive from F. A. Berezin, The Method of Second
Quantization, Academic Press, 1966. Berezin describes (p. 5) anticommuting variables as “elements of a
Grassmann algebra”. For a brief description of Grassmann’s work, see J. L. Coolidge, A History of Geometrical
Methods, Dover Publications, 1963 and 2003, Ch. VI, §4, pp. 252–257, and D. Fearnley-Sander, “Hermann
Grassmann and the Creation of Linear Algebra”, Amer. Math. Monthly, 86 (1979) 809–817.

22.1Parity and Fermi fields

In the theory of a single scalar field, we defined parity simply as x going into −x in the argument of the field:

However we realized that when we had a theory of more fields {ϕ a(x, t)}, we could have a more complicated
definition. It is possible that the fields mix up among themselves, in addition to the space point changing:

I gave several examples in Chapter 6 (pp. 122–125) where the matrix Mab was diagonal: some of the fields were
multiplied by +1, others by −1. The same is true for spinor fields. If we have a set {ψ a(x, t)} of spinor fields, we may
not have the freedom to make individual phase transformations on each of them, but perhaps only on all of them
collectively; there may be only one internal symmetry, not one for each field. In that case, we may have to mix the
fields up among themselves to define parity, and multiply the different fields by different phase factors:

As an example, let’s consider the theory of two spinor fields, ψ A and ψ B, interacting with a spinless field ϕ.
The interaction

is parity invariant if ψ A transforms in the standard way:

and ϕ is a pseudoscalar:
Likewise a second interaction is invariant under parity if ψ B transforms the same as ψ A:

Now I throw in a third interaction without a γ5,

The theory described by

is not parity invariant, if ψ B transforms the same as ψ A. But there are other possible definitions of parity. Among
others, this one works:

While ψ A keeps its standard transformation, and ϕ remains a pseudoscalar (6.91), we’ve changed ψ B’s
transformation to include a minus sign. The Lagrangian (22.5) is invariant under this definition of parity.

Of course a definition of parity could also include a phase factor in the transformations of A and B:

That would still be all right, but here comes the usual ambiguity.1 Whenever we have both an internal symmetry
and one good definition of parity, we can just as well define parity anew, by multiplying the original parity by an
internal symmetry. Which of these we choose to call parity is merely a matter of convention. We could describe the
situation by saying perhaps that the B particle has opposite intrinsic parity to the A particle. A state of two A
particles in an s-wave state would be even in parity, an eigenstate of parity with eigenvalue (+1); the same would
hold true for a state of two B particles in an s-wave state. But a state of an A particle and a B particle would be an
eigenstate of parity with eigenvalue (−1).2

What about the parity of antiparticles? For a charged Bose field, the particle and the antiparticle have the
same parity: whatever happens to the particle, whether it gets a (+) sign or a (−) sign, the antiparticle gets the
same; ϕ and ϕ * have the same parity transformation laws. What happens to the antiparticles in the theory of a
Fermi field, ψ(x, t)?

The particles are associated with the fields via their annihilation and creation operators. We assume that
there is a unitary operator U P in Hilbert space (with no spinor indices) effecting this change of the field:

That tells us how the unitary transformation associated with parity acts on the creation and annihilation operators,
and therefore on the states. We have

and

To work out the effects of U P on the creation and annihilation operators, we need to know how the other factors of
ψ, the spinors up(r) and vp(r), transform. For spinors describing a particle or antiparticle at rest,
The spinors up(r) and vp(r) are related to their rest frame versions by the same Lorentz boost:

with = p/|p|, and ϕ = sinh–1(|p|/m). By the known anticommutation properties (20.4) of β and α i, we can write

and by the same argument

Rewriting (22.9),

Thus the unitary operator that effects parity acting on free particle states (or on in and out states, if we’re talking
about an interacting theory) will transform a one-nucleon state |p, r; Nñ into |−p, r; Nñ with eigenvalue (+1); and a
one-antinucleon state with eigenvalue (−1). That is to say, parity has the same properties in
quantum field theory that it has in non-relativistic quantum mechanics: it changes the sign of the momentum, but
does not affect the spin. However it gives opposite signs to a nucleon and an antinucleon; a nucleon and an
antinucleon have opposite parity. Thus, for example, a nucleon and an antinucleon in an s-wave state has parity
−1, while a nucleon and nucleon, or an antinucleon and an antinucleon in an s-wave has parity +1.

This has important experimental implications if you are dealing with a parity-conserving theory. As an
example, let me consider the processes

at rest. (That is to say, the proton and antiproton are “at rest”—for example, slow antiprotons which we’re sending
into a block of ordinary matter.) We know from non-relativistic quantum mechanics that such exothermic reactions
at small velocities of the incoming particle are dominated by the s-waves. This is also true in relativistic quantum
mechanics. If there is no spatial momentum, then there is no spatial angular momentum: if p vanishes, r × p
vanishes. That argument has nothing to do with relativity. At rest the process is dominated by s-waves, and
therefore there are two relevant states:

The total angular momentum J is conserved. Both of these states are parity eigenstates with eigenvalue −1. On
the other hand, in the final state π+ + π–, the particle and antiparticle are bosons, and therefore they have the
same intrinsic parity, whatever that may be. If the final state is π0 + π0, they obviously have the same parity. It
turns out the pion is a pseudoscalar, with parity (−1). That’s irrelevant, because the square in any case will be +1.
So the parity of the final two pion states is determined by the value of ℓ:

(Note that the orbital angular momentum ℓ contributes a factor of (−1)ℓ to the parity.) If one were to do this
experiment with a polarized target and a polarized beam of antinucleons with J = 0, two pions would not be
produced, because both angular momentum and parity must be conserved. The J = 0 state for the two pions is
forbidden by conservation of parity:

The J = 0 state is forbidden from creating two pions, or indeed, any two particles of the same intrinsic parity;
typically it goes into three pions. Only the J = 1 state for is allowed to make π+ + π−, or π0 + π0. This
example demonstrates that what we have derived is not merely some formal convention, but something that
actually carries experimental consequences.

22.2The Majorana representation

The choice of the right coordinate system often simplifies a particular problem. So too with representations of the
gamma matrices. Our discussion of charge conjugation will be facilitated by choosing to work in a representation
in which all the gamma matrices are imaginary. Let me review what we found in the theory of a charged (complex)
scalar field. There our starting point was the Klein–Gordon equation. The Klein–Gordon operator is real.
Therefore if ϕ is a solution, so too is ϕ *:

We saw in (6.27) that there is a close connection between charge conjugation and complex conjugation: the
complex conjugate of a complex field has the opposite charge. For a complex field ϕ, we can identify these two
operations:

or in terms of creation and annihilation operators, from (6.78),

(and similarly for bp† and cp†).

Now we have to deal with the Dirac equation. Is there a similar connection between complex conjugation and
charge conjugation here? Let’s look at complex conjugation.

I write ψ ∗ rather than ψ †, meaning I will take the complex conjugate of each of the four components of the Dirac
field, but I will not turn a column vector into a row vector. Likewise when I discuss the quantum theory I will use an
asterisk (*) to mean the operator adjoint of each of the four operators. If you like, you can think of ψ ∗ as (ψ †)T. I’m
sorry for that notation, but I have no other symbol to use for just obtaining the adjoint of operators without turning
column vectors into row vectors.

Is the charge conjugated Dirac equation true? It depends on the representation of the gamma matrices. If we
can find a representation of the gamma matrices in which they are all imaginary,

then the answer is “yes”, because then the Dirac equation would be real, just like the Klein–Gordon equation.
Given the symmetry in this representation—ψ ∗ is a solution if ψ is—we would be able to find a similar symmetry in
any other representation just by making the right transformation. Do such representations (22.26) exist? They do.
Their utility was first pointed out by Ettore Majorana3 and they are called Majorana representations.

I will demonstrate the existence of a Majorana representation by constructing a set of four purely imaginary 4
× 4 matrices that obey the Dirac algebra. The trick is to write down our original standard representation (20.11) of
β and α i and shuffle them around (perhaps putting i’s in certain places) so that the gamma matrices

are all imaginary and everything obeys the right algebra. By Pauli’s theorem (§20.3), we can find a similarity
transformation T to swap α 2, the only imaginary matrix in the Dirac representation of {β, α i}, with β. This exchange
makes all the gammas imaginary and preserves the algebra. Of course, this set of imaginary gammas is just one
of an infinite number.4 Such a similarity transformation is given by the unitary matrix

With this transformation, we have

With this set {βM, α iM} of matrices, the definition (20.74) leads to this Majorana representation:

and for completeness,

As you can check, the Majorana gammas satisfy the Dirac algebra:

or, multiplying both sides by β,

Therefore

and taking the Hermitian conjugate of both sides,

The |p, r; Nñ be a nucleon state. Then

On the other hand, for an antinucleon state |p, r; Nñ

They have the right squares, they anticommute with each other, and they are manifestly imaginary. Therefore
there is a charge conjugation invariance, at least on a classical level (providing we treat the components of ψ as
Grassmann variables).

Before I turn to charge conjugation, let me write down some general conclusions that follow from the choice of
a Majorana representation. These will be useful not only for charge conjugation, but also in our discussion of time
reversal, which looks simpler in a Majorana representation than in any other. I should emphasize that results
derived in this particular representation will hold for all representation-invariant objects such as . But the
Majorana representation is advantageous even when we look at the properties of ψ and individually. In this
representation, the Lorentz transformations have the nice property that the matrices D(Λ) (18.1) are real:

Just to convince you of this, I’ll write down explicit expressions for the L’s and the M’s. Let’s start with Mz , which, I
remind you, is

Since γ0 and γ3 are imaginary matrices, their product is real, and the i makes things imaginary:

which holds, mutatis mutandis, for the other components of M. The representation of a boost along the z axis by a
hyperbolic angle ϕ is, from (18.45),
This is a real matrix. As another example, let me take a rotation about the z axis. From (20.46),

so again

and likewise for the other components. A representation matrix for rotation about the z axis by an angle θ is, from
(18.44),

which is again a real matrix. Therefore, the Lorentz matrices in the Majorana representation are real, as
advertised.

22.3Charge conjugation and Fermi fields

Now let’s work out what charge conjugation does to the plane wave solutions of the free Dirac equation, the u’s
and the v’s. The positive frequency solutions satisfy

We are working now in a basis where the gamma matrices are imaginary. If I take the complex conjugate of this
equation, the first term changes sign:

Therefore the complex conjugate solution up(r)* is a v-type solution (20.119b), and (22.37) invites the tentative
identification

Because the Lorentz transformations are real in this basis, complex conjugation commutes with Lorentz
transformations. If we can show

we can with a clear conscience remove the question mark in (22.38).

(these can also be obtained from u0(r)M = Tu0(r)). Using the explicit form of (22.34) in the Majorana basis,

you can quickly check the eigenvalues of u0(1)M and u0(2)M:

In exactly the same way we obtain the Majorana versions of the v solutions by working them out from (20.39),
v0 = −βv0 and βM:

(You can also obtain −iv 0(1)M = Tv0(1), iv 0(2)M = Tv0(2).) By inspection, we see that indeed

Let’s take a particle at rest, and look at u0(1), which is supposed to be an eigenstate of Lz with eigenvalue + .
We can’t just quote (20.16), because we’re using a different set of gammas. We return to (20.15), u0 = βu0, using
βM instead of the standard β, and find two solutions:

which establishes (22.39), and hence also (22.38). The v’s are also eigenstates of Lz :

That is, the eigenvalues of u’s and v’s take opposite signs for their complex conjugates. There’s a simpler way to
see this. If we take the complex conjugates of the first of (22.42), we obtain

because Lz is imaginary. This makes sense physically. The spinor associated with the annihilation of a particle
with z component of angular momentum + is u(1), with eigenvalue of L z = + . But it’s the v(1) with eigenvalue of Lz
= − that’s associated with the creation of a particle with Lz = + , the sign flip coming because one’s got a creation
operator and the other’s got an annihilation operator.

In the Majorana representation we can rewrite the usual expression (21.6) as

replacing vp(r) with up(r)*. Likewise we can rewrite ψ ∗ ((21.7), but without the transpose):

Notice the similarity between the two expressions. If I define a unitary charge conjugation operator U C such that

then

Requiring that (22.48) be the same as (22.50), and comparing terms, I instantly deduce that

These rules applied to a many-particle state define a unitary operator, if we also require the reasonable condition
that the vacuum is invariant under its action:

You might have been worried about complex conjugation turning a positive frequency solution into a negative
frequency solution. You may have thought “Uh oh, we’re going to get something that exchanges annihilation and
creation operators.” It doesn’t happen that way: annihilation operators stay annihilation operators, and creation
operators stay creation operators. The operator C, as you can easily demonstrate, commutes with the free field
Hamiltonian H, which is written as an integral of the sum of normal ordered products of annihilation and creation
operators, bp(r)†bp(r) and cp(r)†cp(r); C merely exchanges the b’s and c’s, so a particle state and an antiparticle
state have the same energy. Thanks to the way we’ve set up the correspondence, complex conjugation does not
mix spin up and spin down states. It’s exactly as if the spin up electron were a boson whose antiparticle was a spin
up positron.

What does charge conjugation look like if we’re not using a Majorana representation? So as not to jettison
completely the approach that most books take to charge conjugation, I’ll show you what this looks like in a general
basis. Let ψ S be a Dirac spinor in some other basis. Then there is a transformation S such that

Then

Writing ψ M = S–1ψ S, and multiplying both sides by S gives

By taking the adjoint of both sides, we see that the rules (22.51) apply equally to the creation operators,

The matrix appears explicitly if we’re not working in a Majorana representation. This is all we’ll have
to say about , because the Majorana representation calculations we’ll do are vastly simpler.

So much for the free field. To discuss an interacting theory we have to consider the charge conjugation
properties of the various combinations of fields that describe the interactions, to see whether or not they commute
with charge conjugation. All of the interactions we will deal with can be written in terms of the sixteen fundamental
quadratic forms (chart, p. 420) built out of pairs of some spinor field and some Dirac adjoint field: ψ AMψ B, where
ψ A and ψ B are two (perhaps different) Dirac fields and M is some 4 × 4 matrix, either 1 or γ5 or one of the four γµ,
etc. I will assume that ψ A and ψ B have the charge conjugation properties (22.49)

It follows that

(where the superscript T denotes the transpose). Of course when I’m dealing with more than one field, everything I
said about parity in this context (see the discussion following (22.7)) also applies to charge conjugation. There are
cases in which theories do not look charge conjugation invariant if you give every field the same phase factor
under charge conjugation, but by putting in an extra minus sign in front of one field or another, you can save
things. I won’t bother to show that here. I’ll work out the charged conjugation properties of the quadratic forms
assuming everything has this charge conjugation property. It’s not hard to figure out what happens if I put a minus
sign in one of these transformations.

I want to study charge conjugation in a quantum field theory, so I will assume the object ψ AMψ B occurs in the
interaction picture Hamiltonian, in the normal-ordered form : ψ AMψ B : so I don’t have to worry about delta
functions appearing. Equivalently I can say that this is an object built out of those funny anticommuting objects
that appear in the classical theory of Fermi fields. Products of Grassmann variables have the same combinatorial
structure as normal-ordered Dirac fields: when I switch the two fields I get a minus sign. That’s going to be
important later on. Then

That has the structure you would expect for charge conjugation: it takes an operator that annihilates a B and
creates an A, and turns it into an operator that annihilates an anti−A and creates an anti−B. Thus we can work out
the 16 quadratic forms, just by using this rule.

Let’s make a table of the bilinear forms and their behavior under charge conjugation:

First consider the scalar and pseudoscalar bilinears. In ψψ, the matrix M is 1 which is both self-bar and real.
Therefore ψψ goes into ψψ. In other words, ψψ is even under charge conjugation, as you would expect, because
the free Hamiltonian’s mass term is mψψ, and that’s certainly invariant under C. For the pseudoscalar, M = iγ5.
The matrix γ5 is the product of four gamma matrices times i (20.102), so it is imaginary. It is also anti-self-bar,
because γ5 = γ0γ5†γ0 = γ0γ5γ0 = −γ5. Thus iγ5 is real and self-bar, so ψiγ5ψ is also even under C.

This object is not in a particularly nice form to express as ψ A(stuff)ψ B, but it is in a nice form to write as
ψ B(stuff)ψ A, if we realize that every one-by-one matrix is equal to its transpose. Therefore I can write

When rearranging these things, because of the definition of the transpose of a product, I have to move ψ B to the
left of ψ A. That gives me a minus sign. At the moment M is any 4 × 4 matrix, and its Lorentz transformation
properties are irrelevant. On the other hand, γ0 is a Hermitian imaginary matrix in the Majorana basis, so

Then

The M term is starred because (20.66) γ0M†γ0 = M; complex conjugation changes M† to MT. I “star” M† to undo the
complex conjugate of the adjoint. Thus I have the general “bar–star rule”:

For the vector bilinear, M = γµ is self-bar, but it is imaginary. So ψγµψ goes into −ψγµψ, and the vector
bilinear is odd under C. This should be no surprise. When ψ describes an electron, the bilinear ψγµψ is the
electromagnetic current. The charge changes sign under charge conjugation if anything does, and so must the
current. For the axial vector, things are different. The matrix γµγ5 is self-bar, but γµ is imaginary, and so is γ5: γµγ5
is real. Then neither starring nor barring do anything to these matrices, and ψγµγ5ψ is the axial vector current,
which is even under charge conjugation. Finally, the tensor bilinear product σµν is i times the commutator of two
gamma matrices. Therefore it is both self-bar and imaginary, and so it is odd under charge conjugation. The
derivations, I remind you, are specific to the Majorana basis, because of course the properties of M∗ depend on
what basis you are in. But the results are basis-independent. Should you forget what’s even and what’s odd, and
you want to rederive it, I recommend that you work in the Majorana basis.

Thus the two model theories we’re looking at, with scalar and pseudoscalar interactions, ψψϕ and ψiγ5ψϕ,
respectively, are charge conjugation invariant, providing we define the scalar field to transform appropriately under
C. On the other hand, we have a different sort of interaction that arises in classical electromagnetism: the
interaction JµAµ, where Jµ is the electric current and Aµ is the 4-vector potential. In quantum electrodynamics
(which we will later discuss in detail) the coupling takes the form

where Aµ is a vector field. This is charge conjugation invariant only if Aµ goes into −Aµ:

The electromagnetic field is a real quantum field, and so its quanta (photons) are neutral particles. If you wish to
define charge conjugation in such a way that electromagnetism is invariant under C, then the photon has to be
odd under C; a one-photon state is multiplied by −1.5

This has interesting consequences for the properties of states, particularly those built up out of one particle
and one antiparticle. (States built up out of two particles are of course turned into states built out of two
antiparticles by charge conjugation, and who cares what the relative phase is—it’s a completely different process.
In any case, neither a two-particle state nor a two-antiparticle state can be a charge conjugation eigenstate.)
Suppose I have such a particle/antiparticle state, let me call it |ψñ:

The function frs(p, p′) is some smearing-out function to make ψ a nice, normalizable two-particle state. When I
apply charge conjugation to ψ, I don’t change anything in this expression except that b ↔ c:
 Positronium is a bound state of e+ and e− in an s-wave, in the ground state; it’s like the hydrogen atom, with a
positron in place of the proton. As in our earlier discussion on nucleon–antinucleon annihilation (p. 462), there are
two s-wave states available depending on the two spin states: J = 1, called ortho-positronium, and J = 0, called
para-positronium; strangely, “para” means they’re anti-parallel. If you have an electron captured by an proton, it
quickly cascades down to a normal hydrogen atom in an s-wave state. Unlike the ground state of the hydrogen
atom, the s-wave state of positronium is not stable, because it can decay into photons; the electron and positron
can annihilate each other.

These operators are in the wrong order for comparing the final state with the initial state, and I have to switch them
around, which gives me the famous Fermi minus sign:

Now r and s are just summation variables, p and p′ are integration variables, so I can exchange r with s and p with
p′. Thus

as one would have guessed, the sign depending upon whether the smearing function frs(p′, p) is symmetric or
antisymmetric in its arguments. However, somewhat surprisingly,

it is the anti-symmetric smearing function that goes with the even state under charge conjugation, the eigenstate
with eigenvalue +1, and the symmetric smearing function that creates the odd state. This is just a result of having
to reorder the two creation operators.

EXAMPLE. The decay of positronium

The J = 1 s-state is a totally symmetric wave function, and it is symmetric in spin and symmetric in space, and as
we have just seen, odd under charge conjugation: C = −1.

The other s-state, with J = 0, is symmetric in space but antisymmetric in spin and therefore it is even under charge
conjugation:

Photons have C = −1, and so the two photons have a net C = +1. Therefore the decay into two photons is allowed
for the J = 0 state, but forbidden for the J = 1 state. The electromagnetic coupling is not strong. As you probably
know, you get an extra factor in the amplitude of around 1/137 (typically more like 1/2π times 1/137) whenever you
emit another photon; the probability is the square of the amplitude. Even without considerations of charge
conjugation, the decay into two photons is much more probable than the decay into more than two photons. But
because of C invariance, the J = 1 state must go into three photons, a final state of C = −1, but much more slowly
than the J = 0 state going into two photons. So although both ground states of positronium are unstable, the J = 0
state is considerably less stable than the J = 1 state.6

The commutation relations of C and P are peculiar, even for a free Dirac theory. If, for example, I take a one-
particle state |ψñ, and apply first P and then C, I get minus the result of applying first C and then P:

(we assume the vacuum is parity and charge conjugation invariant). In the top line parity acts first on the nucleon
creation operator, and reverses the direction of p (22.15); then the charge conjugation turns it into an antinucleon
(22.51). In the second line charge conjugation first turns the nucleon into an antinucleon, and then parity reverses
p and introduces a minus sign (22.15). The same thing happens with one-antinucleon states. In general, if I act on
a state with an odd number of fermions in it, PC is −CP; if I act on a state with an even number (including zero), PC
is +CP. This can be summed up by saying

since the rotation by 360∘ about any axis multiplies every individual fermion wave function by −1, and therefore is
+1 acting on states with an even number of fermions, and −1 acting on a state with an odd number of fermions.
That’s a perfectly legitimate symmetry operator. Any rotation, including one by 360∘, is a legitimate symmetry of
the theory, so it’s not surprising that it should turn up in the product of P and C. In other words,

where N f is the number of fermions in the state that CP is acting on.

22.4PT invariance and Fermi fields

I will next discuss PT, which is always easier in a relativistic theory than T by itself, since the product of parity and
time reversal commutes with Lorentz transformations. I will continue using a Majorana basis. This is convenient
because there is a connection between C and PT via the CPT theorem, and what’s sauce for C is sauce for PT.7

Again I will begin by looking at the scalar case. I remind you that the Klein–Gordon equation is invariant8
under the combined actions of parity and time reversal:

We want to ask if there’s a similar symmetry in free Dirac theory,

I’ve put in a matrix M here which I’ll try to figure out later. The answer is “yes”, if we can cancel out the sign
reversal caused by changing x to −x with the matrix M, such that γµM is −Mγµ. Of course there is such a matrix: it’s
γ5 or some scalar multiple of γ5. And to make the CPT theorem come out right in the end (and not become the
iCPT theorem) I will define the matrix most people choose conventionally to effect PT:

For later purposes, I note that we’re going to realize this symmetry with an anti-unitary operator. We suspect that
there is an anti-unitary operator, ΩPT, such that

Let me work out some of the properties of this hypothetical anti-unitary operator before I actually show you that it
exists. Perhaps its most interesting property is its square:

Remember, in the Majorana representation iγ5 is real; it slips neatly through the ΩPT. Thus (ΩPT)2 is not equal to 1
though it must be unitary, since it is the square of an anti-unitary operator. It turns a Fermi field into minus itself. Of
course, we know such an operator:

it’s the old friend introduced last section, a rotation about any axis by 2π. You might think I have pulled a swindle,
because that minus sign is only there because I chose M to be iγ5 rather than γ5. Might we have obtained (ΩPT)2 =
1 with no minus sign, by choosing γ5? No, and I will demonstrate that.

Suppose I consider an alternative definition of PT,

where eiθ is an arbitrary phase factor. I can certainly do this for the free theory; that’s an internal symmetry. Apply
it twice:
 At this point in our discussions of C and P, I worked through the transformations of all the bilinear forms. I
won’t do that with PT, because a homework problem9 asks you to do some of that. But I will work out what
happens to ψψ, so the homework problem won’t be too difficult.

It follows from the definition of an anti-unitary operator (6.110), that if Ω–1AΩ is A′, where A is some ordinary
linear operator, then Ω–1A†Ω is A′†. Let’s apply this general rule to PT (22.80):

Now let’s look at ψ:

The γ0 is just sitting there, so we can drag it through the ΩPT. It’s a numerical matrix, but it’s imaginary, because
we’re working in a Majorana representation, so we get an additional minus sign. Therefore

This means (combining (22.80) and (22.87))

Now when I apply ΩPT a second time, I can slip the iγ5 out through the external ΩPT with no problem, because it’s
real. But when I bring the phase factor through an anti-unitary operator, it gets complex conjugated, and I get
exactly the same thing as before. Thus ΩPT has a square of −1 and there’s no fighting it by putting in a phase
factor or anything like that. You will still have a square of an operator that produces this U operator, the rotation by
2π.

This seems reasonable. After all, the term ψψ occurs in the free Lagrangian multiplying the mass, and you would
expect it to have nice PT transformation properties. How about the kinetic term in the free Lagrangian?

In comparison with ψψ, I get three extra minus signs: one from the i, one from the imaginary γµ, and one from
changing the sign of ∂µ. There is a further factor of −iγ5 from the transformation of ψ, and a factor of iγ5 from
transforming ψ. Finally, moving the leftmost γ5 through the γµ gives a fourth minus sign. So the end result is
simply

As expected, both terms in the free Lagrangian have the same transformation properties; otherwise we could
hardly expect the free Dirac theory to be PT invariant. Please notice the critical role of the i. Because ΩPT is an
anti-unitary operator, the transformation properties of i times an operator are opposite to those of the operator
without the i. Bringing the i through the ΩPT has the nontrivial effect of introducing an extra minus sign. Thus if one
is writing down interaction Hamiltonians and wishes to check that they conserve PT, or P or C, typically the
restrictions implied by P or C are that certain coupling constants vanish, or are equal to others. The restriction
implied by time reversal or PT invariance is usually that certain coupling constants are complex conjugates of
other coupling constants.10 Let me convince you briefly that I can construct an anti-unitary operator that does this
job. I will rewrite (22.80) and put the iγ5 over onto the other side, and for convenience I will replace x by −x:

I’ll check this equation by writing down the expression for a free field and seeing if ΩPT does sensible things to
creation and annihilation operators. From (21.6)
The c terms will turn out to follow the b terms with no alteration. Now let’s transform the right-hand side of the
expression using (22.91). I replace x by −x, which sends −ip ⋅ x to ip ⋅ x. Then I apply the anti-unitary
transformation, which turns ip ⋅ x back again to −ip ⋅ x. And finally I multiply by −iγ5. We obtain

Note that up(r) is complex conjugated, because ΩPT is an anti-unitary operator. We now have the quantity,
ΩPT−1bp(r)ΩPT, we wish to compute.

Let’s check that this is sensible. It’s only going to be sensible if the objects −iγ5u(r)* are also u’s. Well, they
are, because

This follows because Lz as you’ll recall is purely imaginary in the Majorana basis (22.35). Lz , a product of two
gamma matrices, also commutes with γ5, and therefore

If I start out with a state with spin up at rest, PT transforms it into a state with spin down at rest. There is a sign
change. This is just what we would expect. It’s reasonable that the combined result of parity and time reversal not
change the momentum of a particle. Consider a particle moving in some direction. I change the direction of time,
and the velocity is reversed. I make a parity transformation which changes x to −x, and the particle is back to its
original velocity. What about the spin? In non-relativistic theory we know that parity does not affect the angular
momentum: σ and L both commute with parity. On the other hand, time reversal of course changes the sign of the
spin. If it’s rotating one way, and I run the motion picture backwards, it’s whirling around the other way. And
therefore the combined operation PT should change the sign of the spin, and we found that it does.

22.5The CPT theorem and Fermi fields

I would now like to discuss the proof of the CPT theorem when spin-½ particles are involved. It’s not going to be
too difficult because we already did the hard part when we discussed this theorem for scalar particles (§11.3).
Earlier we showed the CPT theorem is trivial, at least order by order in perturbation theory. (It’s not at all trivial to
prove it rigorously.) In perturbation theory, the CPT theorem is just the statement that if we reverse the sign of all
the momenta in any Feynman graph, so that every incoming particle becomes an outgoing antiparticle, then the
Feynman graph is unchanged. And that result was the CPT theorem:

We found, in our discussion of charge conjugation, that complex conjugating turns a u into a v (22.38)—that is, it
changes the sign of the term. But dragging γ5 through the Dirac operator changes its sign again, and the two
minus signs cancel. So −iγ5up(r)* is, to within some phase factor, some other u. What other u it is depends on how
you’ve chosen your bases and the phases of the u’s, which I don’t want to go through in detail.11 It implies some
reshuffling of the b’s, which defines the anti-unitary operator on the one-particle states, and thus, a fortiori on the
many-particle states. I won’t bother working it out in detail, except to make one remark. I will show by example that
PT reverses the direction of spin. We’ve already seen ((22.42) and (22.46)) that the rest state u0(1) and its
complex conjugates have opposite spins:
The would like to do the same thing here, but more cleverly than the last time, in a way that will also indicate how
this proof can be generalized to particles of arbitrary spin. For simplicity, I will consider a graph involving a theory
of nucleons and mesons, perhaps in our scalar or pseudoscalar theory, or maybe in some grotesque theory that
doesn’t obey parity, charge conjugation, time reversal, PC, CT, or TP. It could be some messy, horrible theory full
of ϵµνρσ’s and derivative couplings and God knows what. For simplicity I will restrict myself to a graph which has
only one fermion line going through it:

(We’ll see at the end what happens if there are 17 fermion lines.) I have this one fermion line,
 Figure 22.1 N + ϕ 1 + ϕ 2 + → N′ + ϕ 1′ + ϕ 2′ + …

from which a bunch of meson lines come out. These mesons may have grotesquely complicated self interactions,
but I don’t care. There is some incoming spinor u, and some outgoing spinor u′. The amplitude for this graph is
going to be a mess, but I can certainly write it in the following form:

where

The numerator N will be a function of p, q1, q2, . . . , p′, full of gamma matrices. And then there will be a nasty
Feynman denominator D full of iϵ’s and inner products of momenta—a function of practically everything in the
whole graph, including what’s in the shaded blob. Then this whole thing is going to be integrated over all the
internal momenta. That’s the general form of a Feynman amplitude.

No matter how grotesque the theory is, whatever the Feynman rules may be, we know that a Feynman
amplitude (10.27) has got to be Lorentz invariant. Therefore this Feynman amplitude has to be equal to the same
thing with everything Lorentz transformed (all the internal momenta, all the external momenta and all the spins):

I should really write D(Λ) in place of D –1(Λ) here, but for reasons that will become clear I’ll write the inverse, which
is equivalent to D(Λ) = γ0D(Λ)†γ0. I should mention that in M, both the numerator and the denominator are Lorentz
transformed, but only a purist bothers with the denominator, since it is expressed in terms of invariant inner
products.

Now comes the cunning part. This expression can be analytically continued to a Lorentz transformation with
complex parameters θ and ϕ. “Analytic continuation?” you say, “Ugh! That’s always a big job; you have to show
that you don’t encounter cuts, poles, essential singularities…” None of that matters here! The denominator is
manifestly analytically continuable, since it’s a function of inner products that need not be Lorentz transformed
(though you may put the Λ’s in if you wish). The numerator has two separate types of factors, the D(Λ)’s and
polynomials in the p’s and q’s. There’s never any problem in analytically continuing a polynomial. D(Λ) is an
exponential function of gamma matrices times the parameters ϕ and θ, and exponential functions are analytic.
That’s why I wrote D –1(Λ) rather than D(Λ). The complex conjugate of an analytic function is not an analytic
function, but its inverse is. As the D(Λ) and D –1(Λ) are analytic, there is no problem with poles. Therefore D(Λ) can
be analytically continued to

This equation is true for real ϕ’s and θ’s, and it’s evidently true by the principle of analytic continuation for
complex ϕ’s and θ’s.

What do we get for our analytic continuation? Complex ϕ’s and θ’s give us something totally unphysical:
complex momentum, disgusting. That may be useful in some other context, but it’s not obvious that it’s of much
help here. There is however a particular complex Lorentz transformation that is extremely useful for proving the
CPT theorem. And incidentally, by doing the proof this way, you will see how to generalize it to arbitrary spins.

Let’s consider the complex Lorentz transformation

That’s a product of a boost and a rotation (see the discussion following (18.40)). The rotation through the real
angle 180° about the z axis is certainly reasonable. That changes the sign of x and y, but does nothing to z and t
(note 9 on p. 374.):
In order to make the signs come out right, I’ll set the parameter ϕ equal to iπ, which makes a 180° “rotation” in the
zt plane. Just to show you what this particular boost does, let’s work out what the usual z axis boost does with a
real rapidity ϕ (18.38):

Now if I take ϕ to be iπ,

and I get

Thus

So this Λ is in fact −1. That’s not a physical Lorentz transformation, but we can obtain it by analytic continuation
from a real Lorentz transformation.

I have now done exactly what I did in the scalar theory: I have changed the sign of each and every momentum
with this particular Lorentz transformation. At the same time, I’m also doing something to the spinors, because
they are being transformed by the D(Λ)’s. Of course, I have to change the spinors, the u for an incoming nucleon
must become a v for an outgoing antinucleon. How does D(Λ) do that?

Earlier I wrote down the form of D(Λ) for a boost along the z axis (22.33) and for a rotation about the z axis
(22.36), in terms of gamma matrices. For the rotation I have

For the acceleration, I have

because the bracketed even and odd terms equal cos π and sin π, numbers known to their friends as 0 and 1,
respectively. Likewise iγ0γ3 is a matrix whose square is −1 so

With a little rearrangement

and for the Lorentz matrix,

Now let’s work out what D(Λ) is. Note that γ1γ2 is a matrix whose square is −1. Therefore

The spinor transformation D(Λ) for this particular Lorentz matrix Λ = −1 is nothing but our old friend γ5, whose
inverse, by the way, is also γ5. Therefore what we have shown is that our original amplitude (22.99) (which by
Lorentz invariance is exactly the same as (22.101)) can be written as

Notice that because γ5 anticommutes with , γ5u can be interpreted as the Dirac wave function v associated with
an outgoing antinucleon: if u is a positive frequency solution of the Dirac equation,

then γ5u is a negative frequency solution:

and the negative frequency solutions are v’s. Likewise u was formerly associated with an outgoing nucleon, but
uγ5 is just right to be associated with the incoming antinucleon; it acts like v:

The things are not quite right. The amplitude (22.115) is the Lorentz transform of the original amplitude, and
therefore equal to it, but it is not yet the amplitude for the CPT reversed process, because of the Fermi minus sign.
Remember, we’ve done nucleon scattering (example, p. 447) and antinucleon scattering (Problem 12.3).
Antinucleon scattering is the CPT transformed version of nucleon scattering. And we found nucleon scattering and
antinucleon scattering differed by a minus sign. That is, to make the minus sign come out right, we have to have

But there is an easy fix for this. To patch up the missing minus sign, we simply require

Then, finally,

The u for an incoming nucleon is replaced by iγ5u which is a v, the appropriate object for an outgoing antinucleon,
and, since iγ5 is self-bar, likewise u for an outgoing nucleon is replaced by uiγ5, which is the same as v, the
appropriate object for an incoming antinucleon (see box, p. 443).

This whole argument generalizes if 72 nucleon lines go through the blob, or 35 antinucleon lines, or a line
with a hairpin turn. Relative to the original process, you get a minus sign in the amplitude for the CPT reversed
process for each external fermion line. In any Lorentz invariant theory of Dirac fields and scalar fields, you get
exactly the same amplitude for a given process and the CPT reversed process if you follow this prescription:

How to obtain ACPT from A

•Every incoming particle becomes an antiparticle of the same momentum

•Every u for an incoming nucleon is replaced by iγ5u

•Every v for an outgoing antinucleon is replaced by iγ5v

•Every u for an outgoing nucleon is replaced by uiγ5

•Every v for an incoming antinucleon is replaced by viγ5

The reason it’s iγ5 everywhere is because there is only one complex Lorentz transformation that effects all of
these changes, so it’s always the same matrix. The secret of the CPT theorem is analytic continuation to complex
Lorentz transformations.

Now, if you happen to have a field theory of spin-3/2 particles on hand, you can see how to derive the CPT
theorem for them. Just compute D(Λ) for this complex Lorentz transformation, and insert an i because they’re
fermions. To prove the theorem for a spin 2 particle, on the other hand, we don’t insert an i into D(Λ), because
they’re bosons. That’s all there is to it. You tell me how a particle transforms under the Lorentz group, and I will tell
you the right matrix that goes into CPT. So, it’s much simpler than either C or PT, and it’s universal. If someone
comes to you and says, “I have written a Lagrangian that is invariant under parity with the conventional phase,
charge conjugation with the conventional phase, but in time reversal I have to insert an unconventional phase”,
you know that either this person has made an error, or else the theory as written is not Lorentz invariant.

Next time, I will sketch out how to do renormalization for a spinor theory.

1 [Eds.] See the discussion following (6.93), page 123 and Example 2, page 123.
2 [Eds.] The parity P of a state is the product of the intrinsic parities Pi of the i constituent particles times (-1) raised
to the power of the angular momentum ℓ of the state: P = × (−1)ℓ. See Bjorken & Drell Fields, Section 15.11,
pp. 108–113.
3 [Eds.] Ettore Majorana (pronounced “Mah-yore-AHN-a”) (1906–?), a brilliant Sicilian student of Fermi’s, first
postulated the existence of the neutron, but Fermi could not convince him to write the paper. In 1938 he boarded a
ship from Naples to Palermo and was never seen again. The Erice summer school in Sicily where Coleman gave
so many celebrated courses is named in Majorana’s honor. Jo~ao Magueijo has written a biography of Majorana,
A Brilliant Darkness, Basic Books, 2009.
4 [Eds.] In the Hill–Ting–Chen notes, Coleman chose a different Majorana representation,

5 [Eds.] This is perfectly reasonable: classically, the vector potential A is proportional to the source charge, so of
course it changes sign under C.
6 [Eds.] The rates are:

See A. Czarnecki and S. Karshenboim, “Decays of Positronium”, 14th International Workshop on High Energy
Physics and Quantum Field Theory, 1999, Moscow; https://arxiv.org/pdf/hep-ph/9911410.pdf.
7 [Eds.] For those readers whose first language is not English, this refers to an old English proverb, “What’s sauce
for the goose is sauce for the gander” (a gander is a male goose): What applies to one case applies to the other.
8 [Eds.] See note 11, p. 130.
9 [Eds.] Problem 14.1, p.545.
10 [Eds.] Again, see Problem 14.1, p.545.
11 [Eds.] For completeness, here are the details for the Majorana representation (22.29):

An easy calculation with the spinors (22.40) shows −iγ5u0(1)* = u0(2), −iγ5u0(2)* = −u0(1), and the same holds for
the v’s. A boost along the z direction involves α 3 (20.44) which commutes with iγ5, so we can also say −iγ5up(1)* =
up(2), etc. If (22.93) is to equal (22.92), we require

The same relations hold for the cp(r)’s and the respective adjoints, the bp(r)†’s and the cp(r)†’s. Once again, PT
does nothing to the momentum but it does reverse the spin.

23
Renormalization of spin- theories

This lecture is devoted to the subject for which you have no doubt all been waiting, the renormalization program
for a theory involving spinor fields.
 23.1Lessons from Model 3

In Model 3, we discussed scalar nucleons. Here I will also work with a specific example, our meson-nucleon theory
with γ5 interactions, referred to in the antique literature as the “ps-ps theory”, pseudoscalar mesons with
pseudoscalar couplings:

Since we’re about to do renormalization, I warn you that this m0 is not equal to the observed mass of the one
nucleon state, µ0 is not the observed meson mass and the coupling constant g0 may not be the coupling constant
as defined by some hypothetical experiment, say by looking for the t-channel pole in nucleon–nucleon scattering.
Not only are these not the physically defined masses and coupling constants, but ψ and ϕ are not the most
convenient fields for scattering theory. In the case of the meson field, we saw in Model 3 (13.52) that we had to
introduce a wave function renormalization counterterm to get the right S-matrix elements. That goes through in
exactly the same way as before (§14.2), because none of our general formulas depended on the presence of the
nucleon field in the theory. All those expressions we found for the meson propagator and so on were true whether
or not there were other fields in the theory. So we look at the vacuum expectation value of the meson field, which
is of course independent of position by translational invariance.

(Since the vacuum is parity even, the pseudoscalar meson field has vanishing vacuum expectation value, so we
don’t have to introduce a shift, as we did in Model 3 (13.52). With ϕ a scalar field rather than a pseudoscalar, with
an interaction the field ϕ’s vacuum expectation value would not necessarily vanish. If it did not, we would
have to shift the field. This complication cannot arise in ps-ps theory, and so we avoid the complication.)

The next thing we did (13.51) was to look at the matrix element á0|ϕ(0)|pñ (where |pñ describes a one meson
state). We defined

By Lorentz invariance cannot depend on p, when the states are relativistically normalized (1.57) (as we did
before and will do here). We introduced a field ϕ′(x):

that had “good” matrix elements—the same amplitude as a free field for making a one particle state out of the
vacuum or annihilating a one particle state. By convention we defined the phase of the one particle states such
that that Z3 is a positive real number. This is the right field to use for the mesons in the reduction formula, and the
one that gives us the correct S-matrix elements.

We want to do exactly the same thing for the nucleon field. The first step is to study the amplitude for
annihilating a nucleon state or creating an antinucleon state. As before we need only study one matrix element,
say the annihilation amplitude

and again I will assume the states are relativistically normalized. The associated matrix element with the nucleon
state on the left is connected to this one by complex conjugation. The corresponding expressions involving are
connected to these by charge conjugation invariance, which is preserved by this theory. If you’re working with a
theory that does not respect change conjugation invariance, the operators are connected by CPT. So in general
this matrix element is the only one I have to study; the others can be found by symmetries. Because of Lorentz
invariance, I can construct the state |r, pñ by boosting a state at rest,

Here p(0) is not the time component of pµ; it is a four-vector that corresponds to a particle with momentum 0:
where m is the actual physical mass of the nucleon.

Let me first write down the assumed properties of the states. I assume that the spectrum of states will
resemble that in the free theory. There may be bound states or something like that, but there is a stable physical
meson which has the properties of the meson state in the free theory as far as parity and Lorentz transformation
properties are concerned. It’s a pseudoscalar: odd under parity. I assume there is a physical nucleon. Its mass is
not equal to m0, but it is a spin-½ particle carrying charge one and transforming under parity as the free nucleon
does. So the states of the nucleon can all be obtained by applying appropriate Lorentz boosts to a nucleon at rest.
The states at rest are two in number. By convention I’ll choose |1, p(0)ñ to represent a particle with spin up,

J is the operator in Hilbert space that corresponds to total angular momentum. This state has the parity
transformation property

Parity does nothing to spin, and since the particle’s at rest, it does nothing to the momentum. I will denote by |2,
p(0)ñ the state obtained by applying the lowering operator in the usual way for a spin-½ particle:

That defines its phase and everything else. Thus by applying symmetry operators, I can generate everything from
a single particle state at rest spinning up. That is, I need only study the matrix elements of the nucleon field at
position zero in the spin up state and at rest:

Once I’ve studied that, by Lorentz transformations and rotations I know the general matrix elements. The state |1,
p(0)ñ is the “descendent”, in some sense, of the free nucleon. For the free field theory and the nucleon at rest we
have (21.6)

Let’s compute these matrix elements (23.11). Will they look like (23.12)? The field ψ is a four component
object, so in principle there could be four matrix elements here, one for each of the four components of ψ.
However we have two conditions which we can use to define the state: (23.8) tells us it’s spin-½ and (23.9) says
it’s parity plus. Consider a rotation acting on ψ(0):

where θ is an arbitrary angle. We can evaluate this expression in two ways. We can first apply the operator eiθJ z
to the vacuum which is of course rotationally invariant, then apply e−iθJ z to the one nucleon state |1, p(0)ñ and
obtain by assumption e− iθ times the original matrix element:

On the other hand, from the known transformation properties (19.3) of the field, this object is also equal to

where Lz is the 4 × 4 matrix (20.7) that effects rotations while acting on the four components of ψ. (It’s not the z
component of the orbital angular momentum L in non-relativistic quantum mechanics. The distinction between
spin and orbital angular momentum is not Lorentz invariant.) In order to match (23.14) with (23.15), the matrix
element must be composed only of eigenstates of Lz with eigenvalue + . Thus if we work in the standard
representation,

then the matrix element must have the form

(I use the standard representation of the Dirac matrices because it is most convenient for discussing states at
rest.)

At first glance it looks like we have two independent numbers in contrast to the scalar case, where we had
only one number to characterize this matrix element (23.3). But of course we have not yet used parity. Let’s apply
that. Both the state on the right and the state on the left are by assumption eigenstates of parity with eigenvalue
+1; the vacuum is certainly parity invariant, and we have said (23.9) the nucleon state has positive parity. So

On the other hand, by the known parity transformation properties (20.8) of the nucleon field,

(Because we’re at the origin, the parity change x → −x is irrelevant.) Using the explicit standard representation
(20.75) of β = γ0,

Consistency ((23.17), (23.18), (23.19), and (23.20)) requires that

and there is in fact only one unknown number, a, which (following (13.51)) I will define to be

By choosing the phase appropriately for the state |1, p(0)ñ, I can arrange that Z2 is real and positive.

Thus we can arrange that the matrix element of the fully interacting field between the real physical one
particle eigenstate and the vacuum is the same as the matrix element of the free field between a one particle state
and the vacuum, if we rescale the field:

Then

simply by applying rotations and Lorentz transformations. That is to say, ψ′ has the same matrix elements as the
free field.

I will not go through the reduction formula afresh for the Fermi case because it’s exactly the same as the
derivation in the pure scalar case (§14.2); there are just more indices floating around and Fermi minus signs
taking care of the conventions in our time ordered product. Once we had gone through this, we could then write
things in terms of the rescaled fields, compute Feynman diagrams, put the external lines on the mass shell and
obtain the correct S-matrix elements just as when we only had Bose fields to play with. None of the arguments we
went through in our derivation of the LSZ formula is sensitive to spin.

Thus our task is the same as in Model 3 (§14.3): to rewrite the Lagrangian in terms of the physical masses,
the physical coupling constants and the renormalized fields, thus generating a bunch of counterterms which we
need to determine. In Model 3 (14.47), there were six counterterms (including one for a shift); here there are five:

We have to determine the counterterms {A, B, C, D, E} iteratively in perturbation theory and then we can do
calculations. After that, all we’ve got to do is multiplicatively renormalize ψ. That will produce matrix elements
between the vacuum and one particle states identical to those of a free field, as they should be. The rest of the
development is largely a repeat of what we did for the scalar nucleons in Model 3.

A digression on theories that do not conserve parity

 Up till now I have used universal principles to construct our model theories: Lorentz invariance, translation
invariance, rotational invariance (well, that’s just a subgroup of Lorentz invariance), CPT to connect the matrix
elements of ψ to those of . We learned last time that CPT was universal. But the one thing I have used that
doesn’t have the flavor of universality is parity. We might want to study theories in which parity is not conserved.

When I first taught this course back in the early Neolithic era, I didn’t bother to discuss renormalization in a
non-parity conserving theory, because the only such theories concerned weak interactions; it’s only the weak
interactions that violate parity. But at the time, there were no renormalizable theories of the weak interaction. So
why should we bother to write down the prescription? The answers would be infinite and thus uninteresting. Well,
times have changed, due to work done in this very building. 1 There are now renormalizable weak interaction
theories rather more complicated than this,2 but still I would like to make a digression about how this part of the
analysis changes. It’s really very simple and it will just take a few lines.

What if there is no parity invariance, and you can’t use parity conservation to eliminate b? How then do we
construct a field that has the same matrix elements as the free field? We are rescued by γ5. In the standard
representation

As you’ll recall (from (22.116) and (22.117)), γ5 turns positive frequency solutions at rest into negative frequency
solutions at rest. It anticommutes with β which determines the difference between these solutions, and it
commutes with all the Lorentz generators. Therefore if I apply γ5 to (23.17), I obtain an equation that has exactly
same Lorentz transformation properties as the original equation. Doing that gives

We can construct ψ′ as a Lorentz covariant linear combination of ψ and γ5ψ to produces the desired matrix
element:

The combination aψ − bγ5ψ knocks out the lower non-zero entry, and to make the other entry equal to 1, I have to
divide by a2 − b2. Then the matrix element with ψ′ will have just the in the first entry and zeros everywhere
else:

as we’d like. Of course, when I make that substitution in the Lagrangian, I get all sorts of parity non-conserving
terms involving γ5 and things like that, but I should, because I started out with a parity non-conserving
Lagrangian. I’ll have a lot more counterterms. I won’t carry through the parity non-conserving case any further than
this, but that’s the general story. It just means life gets a little bit uglier; the fewer symmetries you have, the more
kinds of counterterms you have to worry about.

Let’s return to the main topic. We need to construct five renormalization conditions that will fix the five
counterterms. These conditions will express the fact that the physical charge is whatever we decide to define it as,
presumably by some nice experiment that’s directly connected to physically observable quantities. We also
require that the meson field is properly normalized, that the nucleon field is properly normalized, that the physical
mass of the meson is µ, that the physical mass of the nucleon is m. Of course we’ve already gone through the
analysis for the meson. Nothing we said about the meson field alone, and in particular about the corrected meson
propagator (15.36) in Model 3, is altered by the fact that some of our intermediate states might have spin-½
particles in them. So that gives us two of the five terms we are looking at, from our previous analysis.3 Let me
remind you of what we found for the meson in Model 3.

As you recall (15.29), we began by studying the full Green’s function for one meson in and one meson out
(the Fourier transform of the time ordered product of two meson fields), which I defined in terms of the
renormalized propagator ′:

There is always an energy-momentum conserving delta function, with both p and p′ considered positive if they
point inward. By Lorentz invariance, what’s left has to be a function of p2 only. I then (15.30) derived a spectral
representation for ′(p2) by putting in a bunch of intermediate states:

and summing over states. I got the first term from the contribution from the one-particle intermediate states. Then
there was an unknown continuum which could be thought of as a continuous superposition of free particle
propagators. The integral goes from wherever the lowest threshold is, µ2 + η, to infinity. As far as this field knows,
the only difference between creating a discrete one particle state and a many particle state is that the masses are
smeared out, rather than taking their values at a fixed point. We also found that σ(a2) was greater than or equal to
zero. This meant that if we drew the complex p2 plane, extending p2 to complex values, ′(p2) was an analytic
function except for a cut beginning at the lowest threshold, and the pole at µ2. The first term of (23.30) gives us a
pole at p2 = µ2, and the second a function analytic apart from the cut.

Figure 23.1: The analytic properties of ′(p2) in the complex p2 plane

We then folded this in with a purely diagrammatic analysis to look at the one-particle irreducible (1PI) part of
′(p2), defined (15.33) to be −i ′(p2):

I then derived the equation (15.34)

which is expressed analytically as (see (15.38) and (15.40))

And from this, I deduced two renormalization conditions that fix what I am now calling the A and B counterterms,

This is just a sketch of what we went through before, because I want to parallel each of these steps for the case of
the Fermi field.

23.2The renormalized Dirac propagator ′

There are really no particularly grave complications except those caused by the fact that the Fermi field has four
components. The first step is to study the one nucleon Green’s function, the sum of all Feynman graphs with a
one nucleon field and one antinucleon field. This is the spin-½ analog of the boson propagator ′(p2), (23.29).
(For the time being, we’ll write the argument of ′ as p, rather than p2.) Like every Green’s function, ′(p) will have
an energy–momentum conserving delta function in front of it, but now it is a 4 × 4 matrix, which we can write as a
linear combination of a basis of 4 × 4 matrices, the sixteen combinations of Dirac matrices. The most general
expression looks like this:

We could have a multiple of the identity matrix which by Lorentz invariance could be an arbitrary function a(p2).
We could have γ5 times a function b(p2) (I don’t know if b(p2) is real, so I won’t bother to put the i in front of γ5).
There could be a term γµ which for Lorentz invariance must be multiplied by pµ, the only vector in the problem,
times a function c(p2). Maybe there’s a function d(p2) times γ5. Finally, I could have some function to e(p2) times
σµν, but Lorentz invariance requires pµ and pν dotted into it, and that drops out immediately, because σµν is
antisymmetric: σµνpµpν = 0.

Now, if the theory did not conserve parity, I could have all four of the surviving terms. But parity gives us an
enormous simplification: the terms with the γ5’s have the wrong parity transformation properties. I don’t have to
work it out. It’s obvious that however they transform under parity, they’ll transform opposite to the two without the
γ5, a(p2) and c(p2) . These I know are right because they have to be there in zeroth order perturbation theory,
when ′ is just the propagator F, (21.73):

The terms in γ5 must be wrong! So I’ll just set them to zero because of parity. For the parity conserving case, I
have just two unknown scalar functions,

Given these two functions a(p2) and c(p2), let me consider them for the moment as functions of some scalar
variable z, a complex number:

Written in this way, ′(z) has an even part and an odd part. But if we recall the identity

we can write

Instead of considering this as two scalar coefficient functions, both of which are unchanged as p goes into −p, we
can consider it as a single function, with no particular oddness or evenness properties, of the variable . There’s
no ambiguity because there’s only one matrix in the problem, and one matrix always commutes with itself. You
don’t have to worry about orderings. This is the place where things become much harder in the parity non-
conserving case, because then we would have matrices around that wouldn’t commute with each other. In
particular, we would have γ5 in the mix, and it commutes with almost nothing.

23.3The spectral representation of ′

We’ve now expressed the fermion Green’s function in terms of ′( ), the matrix function analogous to the scalar
function ′(p2) (15.30). The wrinkles are that ′( ) is a function of the matrix variable rather than the scalar
variable p2, and is itself a matrix. That takes care of part one. Now part two: deriving the spectral representation of
′( ).

First I will simply write down the spectral representation. It will look totally mysterious to you. And then I will
explain where it came from. You’ll be able to go through it in your head—if not now in front of me, when I’m looking
at you with my beady eyes, later on. Here is the full expression:

The integrals go from a lower bound in perturbation theory of (m + µ)2 to infinity. In comparison with (23.30), the
third term may be a bit of a surprise. I’ll explain where it comes from.

Remember the derivation (§15.2) of the old spectral representation (15.30). We began with the unordered
product, no time ordering symbol, and put in the complete set of intermediate states. From the one particle
intermediate states, we got the same result as in the free theory, because as far as one particle states go, you
can’t tell the renormalized field from the free field. From the higher states we had all sorts of states that could be
made by hitting the vacuum with ϕ(0). All those states had the same quantum numbers as a single particle. In the
frame in which their spatial momentum p was zero, they were states of zero angular momentum. The only thing
was that there was a continuous distribution of them rather than the single isolated point, and therefore we got a
continuous smeared-out integral of one particle things. That’s the fortune cookie size description of what we did
there.

What happens here? We’ll have a contribution from the one particle intermediate states when we expand the
product of two field operators, a ψ and a , in terms of one particle intermediate states. That’s the first term in
(23.42). It’ll give us the same result, ( ), as in the free theory. Now we consider all the continuum states (whose
mass begins, in perturbation theory at least, at (m + µ)2, the meson–nucleon threshold) that we can make by
hitting the vacuum with (0). We can make two kinds in the rest frame of the states. The two upper
components of ψ which are even under parity can make states of spin-½ in their rest frame, and parity plus, or,
using the notation JP for angular momentum J and parity P, states of ½+. Those states are just like one nucleon
states except they’re smeared out in mass, and therefore we get a smeared out distribution, with an appropriate
smearing function σ+(a2), of one nucleon propagators, where, in analogy with (15.12),

(The prime on the sum means that we are not including single particle states.)

On the other hand in the continuum there are states besides ½+. Even in perturbation theory we not only have
a nucleon and a meson in a p wave state with JP = ½+, but also nucleon–meson s wave states which are ½−.
These ½− states are only connected to the vacuum in their rest frame by the two lower components of ψ, which
are odd under parity. They’re eigenstates of β with eigenvalue −1, and they contribute the third term, with σ–(a2),
to ′( ):

The distribution functions σ+ and σ– are both non-negative by the positivity of the norm on Hilbert space. Note the
presence of the projection operators in the definitions. You might worry that the projection operator − m (acting
on the two lower components of ψ) is negative in a frame where p = 0. But that’s as things should be, because
has a minus sign in its definition for the two lower components in comparison with ψ †. So both projection operators
produce a positive definite result. On the other hand, the functions can be zero:

That is, of course, if the momentum is below threshold to make a meson.

Equation (23.42) is sometimes written in an even more suggestive form obtained by un-rationalizing the
denominators:

(Note that we always add a negative imaginary part to the mass.) In this form, we can discuss the analytic property
of ′ as a function of —or rather, the analytic properties of ′(z), the function of a single variable obtained by
replacing the 4 × 4 matrix by a complex number z. If we draw the complex z plane, as in Figure 23.2, we see that
′(z), no longer an even function, is an analytic function of z, except for a pole at z = m, the physical mass of the
nucleon, and two branch cuts: from the σ+ term, a cut that gives you singularities when z ≥ (m + µ), and from the
σ– term a corresponding cut going off on the left-hand axis. This term becomes singular when z ≤−(m + µ). This
looks a little bit different than Figure 23.1, the corresponding drawing for the meson. There’s a left hand cut as well
as a right hand cut, but the general features remain the same. The statement that the renormalized mass of the
particle is m gives us the location of the pole, and the residue of the pole is given to us by the scale of the fields. Of
course, Figure 23.2 is just the structure in perturbation theory. If bound states develop, there may be further poles
around here someplace due to the bound states; on the right side if they’re positive parity, and on the left side if
they’re negative parity. And if the bound states are sufficiently low you might move the location of the cuts, you
might have a bound state that’s lighter than the nucleon or the meson, in which case the cut will move down from
(m + µ), etc. But the general features are as sketched.

Figure23.2: The analytic properties of ′(z) in the complex z plane

23.4The nucleon self-energy ′

We are still following the road map provided by Model 3. We have renormalized the nucleon wave function. Then
we wrote the Lagrangian in terms of the renormalized wave functions and the new counterterms. The goal is to
obtain equations allowing us to compute the counterterms in perturbation theory. As in Model 3, we have obtained
expressions for the renormalized nucleon propagator, both as a definition and in terms of a spectral
representation. Finally, we have looked at the analyticity of the renormalized nucleon propagator from its spectral
representation.

To keep the parallelism going, the next step is to define the one-particle irreducible (1PI) diagram occurring in
′( ) and sum up the sequence of graphs, analogous to the role of ′(p2) in ′(p2). I define a function of which is
traditionally denoted ′( ):

The 1PI graphs have the same transformation properties—Lorentz, parity, and so on—as the full sets of graphs.
This definition for Fermi fields is parallel to the definition (23.31) of ′(p2), for Bose fields, except that ′( ), like ′(
), is a function of p2 plus a function of p2 times , and will be written as a single function of . Here, ′( ) is the
nucleon self-energy. We know that ′( ) can be written as a geometric series in −i ′( ). I won’t even bother to
write it down. It’s exactly the same as (23.32), you just replace ′ with ′ and ′ with ′. Everything is a function
of matrices, and there’s matrix multiplication but they’re all functions of the single matrix so they all commute. So
I obtain (15.36)

As in (23.33), the requirement that the renormalized propagator ′( ) must have a pole at = m sets the
condition

The condition parallel to (23.34) that the residue of the pole be i is

(This notation—differentiation with respect to a matrix—is standard, but it affects some people like fingernails
scraping over a blackboard.) Just as (23.33) and (23.34) enable us to determine A and B iteratively, order by order
in perturbation theory, these equations (23.49) and (23.50) do the same for C and D in (23.25).
EXAMPLE. ′( ) to order g2

In our discussion of Model 3 (§15.4; §16.1), we computed the meson self-energy ′(p2) to order g2. Following
what I did then, I will now compute ′( ) to order g2, the first nontrivial order. This calculation will involve the
manipulation of Dirac matrices as well as internal loops. I will not carry it out all the way. When I get things down to
integrals that can be found in our integral table I will stop. But I will want to get to its matrix structure.

To order g2, we have two kinds of graphs:

We have a closed loop, and we have the counterterm, evaluated to second order in g, using the same notation
(15.54) as in Model 3. Thus we have two contributions:

Invoking the same guess as before,4 the result of the derivative in the counterterm C ′i ψ′ is just a power of
momentum.

The term −iΣf( ) is the contribution from the Feynman graph with the loop (the superscript f stands for
“Feynman”, not “finite”), and C 2 and D 2 are the counterterms to O(g2). These can be eliminated from (23.52) using
the two renormalization conditions (23.49) and (23.50):

This is the same reasoning that took us from (15.54) to (15.56), except that instead of p2 and p2 − µ2, I have and
( − m), respectively. Of course the computation is a bit different because it’s a dynamically different theory and
we have to manipulate matrices.

Just to remind you of the relevant part of the interaction Lagrangian (23.25), here it is:

The graph for −i ′f( ) gives us

We just have the boson propagator, and then, head before tail, iγ5, the fermion propagator, iγ5. No u and no
because this is not a scattering matrix element, nor a number; it is a part of a propagator, which is a 4 × 4 matrix.

Well, of course this is a divergent integral. We’ve got counterterms that are going to take care of that,
although it’s not going to be quite as easy as it was before. I can gather together all of the i’s. I’ll bring the γ5
through, changing the sign of and , multiplying the other γ5, and becoming 1. This doesn’t change the sign of m.
(Of course, γ5 commutes with 1.) Then

The next step is always the same. I use the Feynman parametrization (15.58) with

to write the integral in a parametric form:

Now I shift the momentum, exactly as before (15.60):

We are hoping that the counterterm subtractions will be enough to make the integral convergent. Writing Σf( ) in
terms of k′,

I should say that it’s very difficult to avoid making mistakes. You write down a lot of mysterious equations, you
erase a lot and curse a lot, that’s how you do it. (Every course in Feynman diagrams is also a course in foul
language.) Now we notice something useful: the denominator is an even function of k′—that’s what the shift buys
us—and we are integrating over all values of each component of k′µ, from −∞ to ∞. The ′ in the numerator is an
odd function and thus irrelevant; it vanishes upon integration:

Thus we have explicitly displayed the result as some function of p2 times plus some function of p2 times the
identity matrix, which is what we anticipated; it’s the only form that can result if the theory is Lorentz invariant.

The complete ′( ) is given by (23.53). It’s only out of a misguided puritanism that I write down the whole
thing, but I want to show you how it works out in full detail at least once:

where the terms in the curly brackets are (remember: the last two terms are evaluated at = m, so a fortiori p2 =
m2)

In our calculation of the Model 3 ′(p2), we discovered (p. 327) that the single subtraction of f (p2) − f (µ2)

was enough to render ′(p2) finite; the derivative term proportional to (p2 − µ2) was separately finite. Model 3 is an
example of a “super-renormalizable” theory, where we don’t need all the subtractions. In the present case, the ps-
ps theory, one subtraction isn’t sufficient. If we look at each of the first two terms in (23.64), we see that the
coefficient of is logarithmically divergent, as is the coefficient of m. And even after the subtraction, at large k′, the
sum of the first two terms goes as

which is still logarithmically divergent. There’s no cancellation in the numerator, as there was in Model 3.

I can hardly praise this expression (23.64) for its beauty, but I can at least ask: Is it finite, or have I been
leading you down the garden path and doing my computations in a non-renormalizable theory? I would not be so
nasty. The last term, thank God, has no problem with divergences:

So this is finite. How do all the other terms go? They all have denominators that go at high k′ as k′4, so they’re all
proportional to d4k′/k′4. That’s the only divergent part at high k′; everything else converges at high energy.
Combining all the numerators gives

Now you will please notice that this is in fact zero. That is to say, ′( ) at high k′ is convergent. I went through all
the details to show you that it really works out.

Actually there is a quicker way to show that ′( ) is convergent. From (23.53), we see that ′( ) differs from
Σf( ) by a constant term and a term linear in . Consequently

The second derivative of Σf( ) completely determines ′, just by integration. So if we know the second derivative
of Σf( ), we can get ′( ):

We can quickly compute the second derivative. If it is finite, which we will show in a moment, ′( ) is certainly
going to be finite. Written in the crudest possible way,

I’ve suppressed the iϵ’s. Whenever I differentiate Σf( ) with respect to , I drag an extra into the denominator. I
differentiate it twice and I get d4k over something that goes like one over the fifth power of k, which is obviously a
convergent integral:

That is the crude, slovenly argument that the integral is convergent. However, differentiating (23.70) is not the
simplest way to compute ′( ); if you really want to determine ′( ), it’s best to use (23.64) and the integral table in
§15.5. That the second derivative (but not the first) is finite also tells us that we need two counterterms, and thus
two subtractions, to make ′( ) finite.

EXAMPLE. A second look at ′(p2) to order g2

Let’s look at the meson self-energy in the ps-ps theory. A similar remark about the divergence of the nucleon
self-energy operator ′( ) applies to the meson self-energy operator, ′(p2): the second derivative of f (p2)

determines ′(p2). The O(g2) diagram for ′(p2) is just Figure 15.6 again, plus the counterterm graph:

In Model 3, the “nucleons” were scalar particles, and the contribution (15.57) to f(p2) involved only scalar
propagators. Now, of course, we get a different contribution to the fermion loop, following the Feynman rules:

Rationalizing the denominators, moving the iγ5 through the numerators, and taking the trace, we get

This integral looks like it is quadratically divergent. But whenever I differentiate with respect to p2, I drag an extra
k2 into the denominator.5 It is the second derivative of f (p2) with respect to p2 that is relevant (compare (15.56)
with (23.53)):

because this is a boson expression. It’s the value at p2 = µ2 and the first derivative with respect to p2 at p2 = µ2 that
enter into the renormalization equations. Once I differentiate with respect to p2, I change the integral from
quadratically divergent into logarithmically divergent because I put a k2 into the denominator. A second
differentiation with respect to p2 turns it from logarithmically divergent to convergent. These two derivatives turn it
from d4k over k2 to d4k over k6. The meson self-energy ′(p2) is right on the borderline of being renormalizable.
Two subtractions are needed and two subtractions are what the renormalization prescription gives us. One
derivative would turn this from being quadratically divergent to logarithmically divergent; not enough. Notice the
marvelous way in which the renormalization program just scrapes by and saves us! Here where we’re
differentiating with respect to p2, the first graph is quadratically divergent. We need those four powers of k in the
denominator to make it convergent. The fermion self-energy is just linearly divergent, but we’re only differentiating
with respect to . There we need those two powers of k in the denominator, not two powers of k2, to save us.

23.5The renormalized coupling constant

Finally, I will discuss the renormalization condition that fixes the last counterterm, E: the definition of the
renormalized coupling constant g, order by order in perturbation theory. There are some cunning things here. If
we’re just interested in getting rid of infinities, we could define the three point function at any combination of
momenta. But we want an elegant definition that will connect it to something that we can actually measure. That
requires a little care.

You’ll recall in Model 3 we defined the renormalized coupling constant −i ′(p2, p′2, q2) as a one-particle-
irreducible (1PI) graph:

And we found (Problem 9.3, (S9.28), p. 352), to lowest order in perturbation theory

We fixed E by choosing to evaluate ′ on the mass shell:

where

but these two equations cannot be satisfied unless some momentum components are complex.6 We can’t reach
this point by a physical process. The only way to get there is by analytic continuation, and you had to accept on
trust that we could do this. The great advantage of this definition is that when we discussed processes like
meson–nucleon scattering with scalar mesons and scalar nucleons and all sorts of corrections to the vertices and
the internal nucleon line, for example a process like the one shown in (23.79), the coefficient of the pole in the s
channel below threshold (or the t channel unphysical pole in nucleon–nucleon scattering) was given directly in
terms of g:

The contribution of this graph on mass shell is

with a pole at s = (p + q)2 = m2. To get the residue of the pole, all the external lines had to be on the mass shell;
and there

(with the scalar “nucleon” propagator for ′(s)). That made for a very nice definition, giving us something that was
physically measurable in this hypothetical theory (see 16.37)).

Now I will try to do the same thing in our current theory, in which the nucleons are fermions described by Dirac
spinors with complicated matrix structure. (I can’t call it the “true” theory; the true theory is quarks and non-Abelian
gauge mesons.) It’s exactly the same diagram as (16.29), but now the nucleon lines are fermion lines and
therefore −i ′(p′, p, q) is a 4 × 4 matrix:

In perturbation theory ′ starts out very simple:

However a technical obstacle arises which must be surmounted. In general ′ will contain all sorts of god-awful
matrices because you have two vectors to play with: any two of pµ, p′µ and qµ to dot into any of the sixteen Dirac
matrices. We could have γ5 or σµνpµp′ν, and so on. Some of them may be thrown out by parity and other
considerations, but most of them survive. ′ could be a horrible object, requiring up to 16 different conditions for
its determination. So we can’t say that for some particular set of momenta ′ is equal to iγ5 times g, because the
equations would be overdetermined. We can certainly adjust matters so the coefficient of γ5 is what we want, by
picking our counterterm as we please. But we’ve got all those other coefficients. Therefore I will look at a simpler
object:

I’ll multiply ′(p′, p, q), restricted for the moment to p2 = p′2 = m2, on the left and the right by these projection
operators. For the moment, I’ll keep the meson off the mass shell. I will demonstrate presently that

That is to say, the object (23.84) is determined by one scalar function G(q2) of the remaining variable q, once I’ve
set p and p′ on the mass shell.7

Before I show that (23.85) is true, I should ask: Have we really lost any useful information by looking at
(23.84) rather than (16.29)? The answer is “no”. If we look at the right hand graph of (23.79), its contribution is

At the pole,

So the combination ( + + m) ′(p + q, p, q) comes in automatically from the product ′(p + q) ′(p + q, p, q). No
such projection operator appears automatically on the right of ′(p + q, p, q), but we can slip it in without loss of
generality, because

That is,

The bracketed quantity has just the form of (23.84). The same argument goes through for the remaining factors in
(23.79), ′Γ(p′, p + q). So I may have restricted myself in what I can look at, destroying this marvelous, rich
structure of 42 different combinations of Dirac matrices that could be in ′(p, p′, q), by putting projection operators
on either side of it. But I saved all the parts that are important when I’m computing the residue of the pole in the on
mass shell scattering process.

Now let me give a quick demonstration that (23.85) is true. The right hand graph of (23.79) with the nucleons
on the mass shell, but the meson off, can be thought of as a variety of processes, depending upon whether the
nucleon lines are on the upper or lower mass hyperboloid. It could be the sort of matrix element we used when we
were discussing decay processes, with ϕ(0) between the vacuum and a nucleon antinucleon out state:
When some of the fields are on the mass shell, they become in states or out states, depending on where they are,
and the others stay as fields. Or it could be that both are on the upper hyperboloid:

There I don’t have to say in or out, depending on how I put them. Or, you know, the nucleon and the antinucleon
on the right, in an in state. It doesn’t matter.

Now all these processes are of course connected by analytic continuation. I say “of course”, because it takes
a month to prove it. But they’re all obtained from the same function and it’s just a matter of whether q2 is timelike or
spacelike, positive or negative, to say which process you’re describing. So as far as counting invariants, I might as
well count them with the process (23.90), to see how many numbers I need to describe this process, or any other
related to it:

By Lorentz invariance I might as well look at the frame in which q carries a timelike momentum to make a real pair:
the total momentum of the pair equals the momentum carried by the meson. So I’ll make

You hit the vacuum with the meson field; the field makes a pair. I choose to look in the Lorentz frame in which the
pair is made at rest.

Now what do I know about the field? ϕ(0) is Lorentz invariant, in particular rotationally invariant, so the final
state must have total angular momentum zero, and ϕ(0) is a pseudoscalar field. The state you make by hitting the
vacuum with ϕ(0) is odd under parity: its parity equals −1, and its JP equals 0−. The states of the
nucleon–antinucleon pairs come out with the given value of q2. The two spin ½ particles in their final state can
make S = 1 or S = 0. To conserve angular momentum, their J value must equal 0. Their possible values of JP are:

But the second violates parity. Thus there’s just one invariant amplitude, the amplitude for making this s-wave
state, which is completely determined once I’ve given its center of mass energy q2. If we didn’t have parity
invariance we’d have two. Thus this process is indeed described by a single function of q2. All we’re doing is
counting states here. If there were 72 partial wave states it could go into, it would be described by 72 functions of
q2. We have only one. Therefore we are free to impose our one renormalization condition, to wit: at q2 = µ2, G(µ2)
= g:

That is the definition of the physical coupling constant that corresponds to what we did in the boson case. The task
is done.

Next time I will talk about isospin and how it fits in with field theory.

1 [Eds.] Sheldon L. Glashow, “The Renormalizability of Vector Meson Interactions”, Nuc. Phys. 10 (1958)
107–117; and “Partial-Symmetries of Weak Interactions”, Nuc. Phys. 22 (1960) 579–588; Steven Weinberg, “A
Model of Leptons”, Phys. Rev. Lett. 19 (1967) 1264–66; J. Schwinger, “A Theory of the Fundamental Interactions”,
Ann. Phys. 2 (1957) 407–434. Glashow earned his doctorate under Schwinger. Harvard’s Physics Department is
housed in the Lyman Laboratory building, 17 Oxford St., Cambridge, MA.
2 [Eds.] Abdus Salam, in Elementary Particle Physics, ed. N. Svartholm, Almqvist and Wiksell, Stockholm 1968;
Weinberg, ibid.; Glashow, ibid. Glashow, Salam and Weinberg shared the 1979 Physics Nobel for their
electroweak theory. A little later this theory was indeed shown to be renormalizable by Gerard ’t Hooft and
Martinus Veltman, whose work was recognized by another Physics Nobel, in 1999.
3 [Eds].See (15.46)–(15.48), and (16.5). Note that the terms B and C in (14.47) correspond to the terms A and B
in (23.25), respectively.
4 [Eds.] See §14.4, (14.57) and Problem 8.1, Comment (3), p.309.
5 [Eds.] After canceling the iγ5 and taking the trace,

Using the Feynman parameter trick and shifting k = k′− px exactly as before,

As before, the term linear in k′ in the numerator is odd, and can be dropped. The net result is

This is a function of p2, as required. Every differentiation with respect to p2 reduces the integrand’s power of k′ by
2.
6 [Eds.] In its rest frame, a real (on mass shell) nucleon cannot emit a real meson. Where would the energy come
from? It could emit a virtual meson, but that meson’s momentum would be complex. This is easy to show
algebraically: p′2 = p2 + q2 + 2p ⋅ q, or q2 = −2p ⋅ q = −2mEq. This is impossible if q2 = µ2.
7 [Eds.]In modern language, G(q2) is called a running coupling constant, dependent on momentum (or in position
space, separation); see §50.4. It is impossible to put all three particles on their mass shell in the (nonphysical)
process N → N + π, so Coleman wants to look at the value of G(q2) in experimentally accessible processes like N
+ N → N + N.

Problems 13

13.1 In the discussion (§23.4) of renormalization of the “pseudoscalar” theory,

we sketched a computation of the renormalized nucleon self-energy ′(p2) to O(g2). Complete the calculation.
Again, leave the integral over the Feynman parameter undone.
(1991a 11.2)

13.2 In the same theory,

(a) Compute the renormalized meson self-energy, ′(k2), to lowest nonvanishing order in perturbation theory,
O(g2). Leave the integral over the Feynman parameter undone, just as it was left undone in our discussion of the
same object (16.5) in Model 3. H INT: All you need for this problem are the conditions that fix B and C, the ∂µϕ∂µϕ
and ϕ 2 counterterms in (14.47). These conditions are the same as in Model 3, (15.38) and (15.40):

(b) We derived a formula for the imaginary part of [−i ′(k2)]–1, for real k2, in terms of the spectral function σ(k2),

(For this equation in Model 3, see (S9.13) and (17.4).) Because we want σ(k2) > 0 above the two-particle
threshold, k > 2m, it follows directly that the imaginary part of the self-energy ′(k2) is negative for k > 2m. Check
that in your calculation the imaginary part of the self-energy has the right (negative) sign, confirming the
correctness of the rule, “a minus sign for every closed fermion loop” (or possibly only confirming that you’ve made
an even number of sign errors.) N OTE: The minus sign rule (item 2 in the table on page 443) for closed fermion
loops is essential in getting this sign right.
(1997a 11.4; 1991a 11.3)

Solutions 13

13.1 The renormalization conditions give us the self-energy (23.53)

To O(g2), −iΣf( ) is the amplitude for this diagram: . Then (23.55)

The denominator can be simplified:

Shifting the momentum k → q = k + xp gives

The linear term is odd, so it contributes nothing and may be discarded. Then using the integral table ((I.4), p. 330),
we get

Finally, from (S13.1),

We can simplify no further without integrating over the Feynman parameter, x.

13.2 (a) The renormalization conditions give us the renormalized self-energy (16.3)
To O(g2), −i f(k2) is the amplitude for this diagram:

Using the Feynman rules in §21.5, being careful to put in a minus sign for the closed Fermi loop, we have

Now we use Feynman’s trick (16.16) to combine the denominators:

Plugging this expression into (S13.8) we have

Now shift the momentum q;

and after algebra,

The denominator is even, and so the terms odd in q′ in the numerator can be discarded:

The integrand can be rewritten:

and the integral becomes

Now we consult the integral table (box, p. 330). From (I.4),

and from (I.3),

Then

(We can drop this iϵ; it doesn’t affect the analytic properties of the expression.) Then from (S13.7)

We dropped the iϵ from the denominators, because below the two-particle threshold, k2 = µ2 < 4m2, and the
denominators will never equal zero:

That’s as far as we can take this expression without integrating over the Feynman parameter, x.

(b) From (S13.18)

We need to investigate the sign of the imaginary part of this expression for k2 > 4m2. From (S9.2) we have

The imaginary part of f(k2) will be zero unless

In the region of integration, x(1 − x) takes its greatest value at x = , and so we must have

which is exactly the region of interest. Then 3k2x(1 − x) > m2, so the coefficient of the logarithm is positive, and
consequently

which was to be shown.

24
Isospin

Much of what I’ve talked about so far has a rather indirect connection with experiment. I’d like to discuss a set of
experimental phenomena, though not in our main line of development, in which we can use the field theoretic
ideas we have been developing, not in a precise numerical way but generally, as clues to construct a theory that
describes experiment. This will not involve rigorous logical development, or even what passes for rigorous logical
development by the standards of modern theoretical physics, but rather guesswork.1

24.1Field theoretic constraints on coupling constants

The subject I would like to discuss is one that you have probably encountered previously, isotopic spin. The
standard development of isotopic spin begins with the study of nuclear energy levels.2 Consider a sequence of
light nuclei containing the same total number of nucleons, but differing from one another by their charge, for
example boron-12 (five protons and seven neutrons), carbon-12 (six of each), and nitrogen-12 (seven protons,
five neutrons). (Nuclei with the same number of nucleons but different charges are called mirror nuclei, or
isobars.) Inspecting Figure 24.1 we notice that3 these states differ from each other merely by the exchange of
protons and neutrons. The near-equality of energy levels suggests that the force between two protons is
approximately equal to that between two neutrons; the nuclear force is independent of the identity of the nucleons.

Of course there are small differences in the energy levels, but these can be accounted for at least qualitatively
by Coulomb corrections: the nitrogen nucleus has a charge of seven while the boron has a charge of five. If one
looks at carbon-12, with six protons and six neutrons, one finds a similar energy level spectrum, except that there
are states (indicated by dashed lines) that are not present in the other cases. This can be viewed as an effect of
the exclusion principle, which is less restrictive than with seven protons and five neutrons or vice versa. If we
imagine these as particles interacting in some sort of collective potential, there are energy levels that can be
occupied in the case of carbon, but are forbidden by the exclusion principle in the other cases. This implies that
the proton–neutron nuclear force is the same as the proton–proton or neutron–neutron force, except that there are
states in which a proton and neutron can occupy that a proton and proton or neutron and neutron cannot. So I
should compare energy levels in antisymmetric neutron–proton states only; there are no symmetric
neutron–neutron or proton–proton states. This is all in neglect of electromagnetism and of course the weak
interactions. (To determine the effects of the weak interactions on nuclear energy level spectra would require more
sophisticated experiments.)

Figure 24.1: Nuclear energy levels for the isobars 12B, 12C and 12N

I said that the differences in energy levels between these mirror nuclei were qualitatively what would be
expected from Coulomb corrections. Actually there are two effects that make the energy levels slightly different.
Besides the Coulomb corrections there is the fact that the mass of the proton is not quite equal to the mass of the
neutron. It’s very tempting (and we will yield to that temptation here) to assume that the proton and neutron mass
difference is itself an electromagnetic effect, and that the mass of the proton would be equal to that of the neutron,
if we had the magical power to turn off electromagnetism (and the weak interactions). There is no real evidence
for this, but the order of magnitude of the energy difference is roughly what you would expect on the basis of
dimensional analysis, for a sphere of charge q and the approximate radius r of a nuclear particle,

It is a plausible idea.4 No one has been able to calculate the proton and neutron mass difference.5 We will see
later what consequences can be drawn from this hypothesis, and how well it is supported by experiment.

So far I’ve made no use of field theory. Indeed this argument could be constructed by someone who only
knew non-relativistic quantum mechanics. From field theory we know that at least part of the force between two
nucleons, certainly the long-range part of the force, is due to the exchange of a π meson. Depending on what the
scattering process is, this may be a π0 meson or a charged π meson. I will assume, in the same spirit as before,
that the pions all weigh the same in the absence of electromagnetism. (In reality there is of course a 4 or 5 MeV
difference between the pion masses.) I will attempt to construct a field theory consistent with what is known about
the pions and nucleons that would give a force due to the exchange of pions, and then investigate the constraints
placed on the coupling constants. The field theory will involve a proton field and a neutron field, which I will denote
simply by “p” and “n”. These are four-component Dirac spinor fields. The p field is, by my usual convention, the
field that annihilates the proton and creates the antiproton, and likewise for the n field. This theory will also involve
a charged field for the pions, which I will call ϕ +. This field annihilates π+. It has a conjugate ϕ +†, which I’ll call ϕ –.
There will also be ϕ 0, the field of the neutral pion, which is equal to its adjoint, ϕ 0 = ϕ 0†. I will assume that the
pions interact with the nucleons through interactions of the sort we have been discussing. (Later I will indicate that
this assumption can be relaxed.) The pions are empirically known to be pseudoscalar particles: they require γ5
interactions:

That’s the most general Lagrangian consistent with Lorentz invariance, parity and electric charge conservation
that does not involve any derivative couplings and allows for trilinear pion–nucleon interaction. It involves three
unknown parameters, the two real parameters gP and gN and the complex parameter gC. (The free Lagrangian is
assumed to be of standard form.) We can simplify matters somewhat by using our freedom to redefine the phases
of the complex pion fields: we can absorb the phase of gC into ϕ +, and stipulate that gC is greater than or equal to
zero, and we can use our freedom to reverse the sign of ϕ 0 to arrange that gP is greater than zero. The coupling
constants gP and gN have to be real if the Lagrangian is to be Hermitian. The dots indicate renormalization
counterterms, but we won’t bother with them here.
 One thing we know about these forces is that they are strong. Otherwise nuclei would fall apart because of
the electromagnetic repulsion. In fact if we do scattering experiments, and define the renormalized “charge” in the
manner I explained last time, then the coupling constants turn out to be absolutely gigantic. We’ll see that they’re
all the same order of magnitude. A typical coupling constant g, from the analysis of pion–nucleon scattering gives

This is a very large number, and therefore perturbation theory, the only analytic tool we have now, is completely
useless for analyzing this problem. (In QED, the coupling constant e2/4π ~ 1/137.) On the other hand we have one
result that is independent of perturbation theory: the pole in the pion propagator lies at the physical mass of the
pion. That’s one of our renormalization conditions, and it’s true to all orders in perturbation theory. If the total force
is to be symmetric, i.e., the same between any two nucleons, we should at least have the part of it caused by the
pion pole be symmetric. That is the force we get by doing ordinary perturbation theory to second order in any of
these coupling constants. One condition that is certainly implied by this is that pp scattering equals nn scattering
equals pn scattering in an antisymmetric state to lowest order in perturbation theory:

This is not an assumption that the strong interactions are weak: it is using what we know, that the lowest order
graph in the neighborhood of the pole suffers no corrections because of our renormalization condition, and
therefore gives us the exact scattering amplitude at that point. And if the exact scattering amplitude obeys these
equalities, then in particular the residue at the poles should obey these equalities.

The next task will be to actually compute the lowest order scattering graphs. Of course we’ve done them in a
simpler theory, one that has only one nucleon and one meson. We need to compute them as functions of the
unknown parameters gP, gN and gC, and see what restriction the assumption of nucleon symmetry places on
those coupling constants.

We’ll begin with pp scattering. We have an initial state |iñ which is characterized by some spinor u1 and some
4-momentum p1, some spinor u2 and some 4-momentum p2 which for shorthand I will simply write as 1, 2:

and similarly for the final state |fñ

We wish to compute the scattering amplitude. There are two graphs: the direct graph, and the exchange graph:

Figure 24.2: O(g2) Feynman diagrams for pp scattering

The existence of the neutron and the charged pion are irrelevant here. Because these are two protons, the
only particle that can enter is a π0. These graphs will both be proportional to gP2 times some function f, a thing
we’ve computed before (21.100). I will write it out explicitly in a moment (though we won’t need its actual form) for
1, 2 going into 1′, 2′, and then we have the exchange graph, which is the same function with 1 and 2 interchanged:

(see the example on p. 447 for an explanation of the minus sign) where

Next we do neutron–neutron scattering. There again we have exactly the same two graphs with the π0 exchanged
and the amplitude is gN2 times the same thing. Thus from the statement that the pp force equals the nn force we
derive the equation

Now we have to construct the scattering amplitude for the antisymmetric state, properly normalized; the initial
state is
The first particle is still a proton, the second particle is still a neutron, we’ve just changed the momentum and spin
labels. Similarly, the final state is

There are four terms in the scattering amplitude. There will be direct terms coming from |1, 2ñ with |1′, 2′ñ, and from
|2, 1ñ with |2′, 1′ñ. These will give me identical expressions that will cancel out the in the denominator. So from
these terms I will get just the same result (24.11) as before:

Then there will be the exchange terms, from |2, 1ñ with |1′, 2′ñ and from |1, 2ñ with |2′, 1′ñ. This contribution will have
an overall minus sign because there’s an explicit minus sign there. Once again they will give identical
contributions canceling out the , and the contribution aE is obtained by making the appropriate exchange, and
adding the minus sign:

To no one’s surprise, we have an expression of similar form to (24.14). Adding the two gives

This admits of two possibilities:

(we chose gP ≥ 0 but gN could still be negative).

The case of proton–neutron scattering is a bit more complicated because we have to construct an
antisymmetric state. Let me first do it for a non-antisymmetrized state. So again I’ll take |iñ = |1, 2ñ. It doesn’t matter
whether the proton or the neutron creation operator comes first, as long as I adopt the same convention for the
final state. Later I will construct the amplitude for the antisymmetric state. There are two possible graphs. In the
graph on the left, the proton labeled by 1 and 1′ and the neutron, labeled by 2 and 2′, fly past each other with the
exchange of a π0. In the one on the right, the proton labeled by 1 comes in, emits a π+, turning into a neutron, 2′
and the neutron coming in, 2, absorbs the π+ and turns into the proton, 1′. (Of course I could just as well say that
the original neutron emits a π– and turns into a proton—it’s the same graph.)

Figure 24.3: O(g2) Feynman diagrams for pn scattering

These graphs are the same as those in Figure 24.2, except that the coupling constants enter differently:

Comparing (24.16) to (24.7) and (24.9), we find

This equation asserts that (at least in the neighborhood of the pole) proton–neutron scattering in the antisymmetric
state is the same as proton–proton scattering and neutron–neutron scattering.

Depending upon whether we adopt Possibility A or Possibility B we obtain two solutions to (24.17)
If we adopt Possibility A, then gC = 0. This is no good experimentally. It would mean that there is no π+p
scattering, which is untrue. In fact the most direct experimental evidence is that you have an obvious nucleon pole
in pion photoproduction off nucleons, but I didn’t want to rest the argument on that because we haven’t yet
discussed electrodynamics. In any event, gC = 0 is in flat contradiction with experiment and must be rejected.

With Possibility B, gP = −gN, we deduce that gC = gP. Remember, we have chosen our phases so that both
gC and gP are positive, so there is no sign ambiguity in taking the square root. Therefore we have found
essentially the unique possibility. Our three unknown coupling constants have been reduced to one overall
unknown coupling constant. It is therefore useful to change the notation slightly, and introduce a new coupling
constant:

The form of our interaction Lagrangian then becomes

The minus sign arises because, just as before, the annihilation and creation operators have to be rearranged.

24.2The nucleon and pion as isospin multiplets

This Lagrangian (24.20) has more symmetries than you might think. The square roots of 2 are a little ugly, but
remember that we had a similar when we defined a complex field (6.23) as a sum of two real fields. That
suggests we should define fields ϕ 1 and ϕ 2:

In the same spirit I will relabel ϕ 0 and call it ϕ 3. The form of the interaction then becomes

Define a vector Φ by

I will also define an eight-component nucleon spinor N:

N consists of a four-component Dirac spinor p sitting on top of the four-component Dirac spinor n. Likewise

I can write out my Lagrangian in terms of these objects. I’ll first do the free Lagrangian. Note that

so the free Lagrangian for the mesons can be written as

Similarly, the free Lagrangian for the nucleons can be written as the sum of the proton and neutron terms:

which I will write, by an abuse of notation, as

I will write the interaction similarly as

where τ is a set of three 8 × 8 matrices, block diagonal with respect to the four Dirac components, chosen to
reproduce the couplings in (24.22):

where is the 4 × 4 matrix whose elements are all zero, and is the 4 × 4 identity matrix.

These three matrices are not strangers to us. They are precisely the three Pauli matrices, but 8-dimensional.
Indeed this whole Lagrangian is revealed to be symmetric under a group that is isomorphic to the three-
dimensional rotation group, SO(3), or equivalently SU(2), the group of unitary 2 × 2 matrices with determinant
equal to 1. This group has nothing to do with spacetime geometry; it’s a purely internal symmetry, acting on ϕ 1,
ϕ 2, and ϕ 3 as well as on the neutron and proton field. The internal space in which the transformations are carried
out will be called isospace. The transformation of the triplet of fields Φ is

That is, R( θ) is a 3 × 3 rotation matrix characterized by the axis and the angle θ which acts on Φ. The trio of
fields Φ transforms like a vector under this internal group, so we’ll call it an isovector. The eight-component
object N transforms as an isospinor:

under the same group. (Note that the generators of isospin rotations are τi, not τi, just as the generators of
rotations for spin-½ are σi.) Just to remind you that these transformations have nothing to do with space
rotations, I’ve written the unchanged variable x as the argument of the fields. This group is called the isospin
group, or sometimes the group of isotopic rotations; “spin”, to emphasize that it is group-theoretically identical
(“isomorphic”, in mathematical language) to ordinary three-dimensional spin, and “iso-” to indicate that it connects
together nuclear isobars.6 We haven’t done anything that might be called rigorous, but we’ve actually got a lot of
symmetry out of a simple assumption, that the proton–proton force and the neutron–neutron force are the same as
the neutron–proton force, in antisymmetric states. Using the field-theoretic idea that the long-range part of these
forces is caused by the exchange of a pion, we have found that our theory is symmetric under a three-parameter
continuous group of internal symmetries.

This is our third encounter with the three-dimensional rotation group. We came across it in its proper guise as
SO(3) in §5.6. We met it again analyzing the representations of the Lorentz group (§18.3), when we were able to
reduce the Lorentz group into two SO(3) factors. And now we see it a third time, as a purely internal symmetry
group. This is very convenient for us, as we do not have to develop a new group theory for these different
problems; we just have to continually apply the theory of the three-dimensional rotation group. Unfortunately this is
the end of our luck: the next group we will encounter if we continue this line of development is SU(3), the group
associated with the Eightfold Way of Gell-Mann and Ne’eman. And for that you have to learn some additional
group theory and representation theory. It is not expressible in terms of SO(3).

Since this is an internal symmetry, we can apply our machinery (§5.3). I’ll just sketch out the results. Since we
have a three-parameter continuous group, we can deduce three conserved isospin currents, {J1µ, J2µ, J3µ}. I’ll
label them with the index i = 1, 2, 3. Do not confuse this with the spatial vector index. It is easy to figure out that the
pion field contributes a term

That’s just the isospin analogy of our old friend r × p in a new guise. The nucleon field contributes

That’s the same form as the electromagnetic current of a charged fermion field, except there’s a τ to account for
three isospin components. So the isospin current is

By integrating the time components Ji0 we obtain three generators:

These are all conserved quantities, neglecting as always electromagnetism and the weak interactions:
The three generators of isotopic spin, by the usual arguments, obey the algebra of the rotation group:

(sum on repeated indices implied).

For later purposes it will be convenient to introduce the raising and lowering operators I+ and I–:

and write the algebra associated with the three-dimensional rotation group:

This tells us that I+ and I– are I3 raising operators and lowering operators, and

These are just formulas which I copied out of the section of the non-relativistic quantum mechanics book that I
happened to have on hand, changing J’s to I’s. I presume you are familiar with all of them.

The Φ field transforms like a vector under isospin rotations. A field that transforms under rotations as a vector
has J = 1; here we say Φ carries I = 1. The nucleon field, transforming as an isospinor, has I = . Once we have
given the transformation properties of the fields, we know the transformation properties of the particles they create
and annihilate. I will simply write down the table giving the total isospin I and the value of I3:

The vacuum is of course an isoscalar with I = 0. The proton and neutron are both I = since they’re created by
hitting the vacuum with the I = nucleon field. Their values of I3 are +½ and −½ as you can see simply by reading
off the spinor:

The π+, π– and π0 form an isotriplet. They dance among themselves under the action of the isospin group. The π0
obviously has I3 = 0, slightly less obviously the π+ has I3 = 1 and the π– has I3 = −1. To make sure things are right,
we simply observe that we can turn a proton into a neutron plus a π+ virtually. That’s one of our couplings. For the
isospin to add up, I3 has to be conserved. The proton has I3 = +½, the neutron has I3 = −½, the π+ had damn well
better have I3 = +1. I’ve gone through this rather briefly because it’s essentially a direct copy of what I presume
you have done many times for ordinary spin. The only novelty is replacing J, angular momentum or S, spin by I at
appropriate places.

24.3Experimental consequences of isospin conservation

Isospin is very restrictive and very easy to test experimentally beyond the simple NN scattering we started with.
For example, let’s consider pion–nucleon scattering.

There are three possible pion states and two possible nucleon states, so there are six possible initial states
and six possible final states. In principle this would give us 36 scattering amplitudes. Of course many of them are
zero because of the conservation of electric charge. So just to see how much additional information isospin gives
us, let’s first count how many amplitudes there are if we only insist on charge conservation. There is a unique
initial state with q = 2, π+p, and the only thing it can scatter into is π+p. So there is only one amplitude, one
function of space and spin variables, for the q = 2 channel. There are two states of charge 1, π0p and π+n, and
there are four scattering amplitudes if we consider scattering in the q = 1 channel: the elastic scattering of π0p →
π0p and π+n → π+n, and what is called charge exchange scattering, π0p → π+n, and vice versa. Likewise for
charge 0 we have two states, π–p and π0n, so again there are four amplitudes. And finally for charge −1 we have
the unique possibility π–n → π–n. This gives us 10 scattering amplitudes. Two pairs of these are connected if we
also insist on time reversal invariance, because π0p → π+n is connected by time reversal to π+n → π0p, and
likewise π–p → π0n is the reverse of π0n → π–p. This reduces the 10 amplitudes to 8. (Time reversal is a good
symmetry for everything except the weak interactions.7) This classification can be summarized with a chart:

On the other hand, if we do an isospin analysis, a pion is isospin 1, and the nucleon is isospin ½. Combining
these gives only two possible total isospins, I = and I = . We have in fact only two possible kinds of final states,
ignoring space and spin degrees of freedom, and only two amplitudes. All these 10 independent amplitudes, or 8
independent amplitudes, are some linear combinations of these two amplitudes, one for the isospin-½ channel
and one for the isospin- channel, A1/2 and A3/2, respectively. Therefore isospin, even for the simple problem of
pion–nucleon scattering (a very well measured process experimentally) produces enormous restrictions. It
enables us to predict these eight independent scattering amplitudes in terms of two unknown functions of
momentum and spin. It is a very restrictive assumption—modulo electromagnetic corrections, of course. In actual
fact, especially for low momentum transfer scattering, electromagnetic corrections can be quite important.
Although electromagnetism is weak, it acts over long ranges and therefore dominates the small momentum
transfer part of the scattering amplitude. To experimentally check that all the amplitudes are linear combinations of
A1/2 and A3/2, you must either restrict yourself to large momentum transfers or make explicit numerical corrections
for electromagnetic effects. This just makes life harder for someone who wants to design an experiment to check
isospin invariance; it does not affect the conclusion.

Let’s explore a very simple application. There is a famous resonance in one of the easiest of scattering
experiments to do, π+p scattering, that occurs in the total cross-section. If you look at a plot of the total cross-
section for π+p as a function of the center-of-momentum energy, there’s an enormous bump centered around
1232 MeV, with a width of around 100 MeV, an obvious resonance. It’s the famous Delta resonance, Δ++, with
charge +2e. Aside from kinematic factors which are slowly varying over the width of this resonance, σ(π+p) is
proportional to the imaginary part of the forward scattering amplitude for π+p → π+p, by the Optical Theorem
(12.49). But this amplitude in turn must be written completely in terms of the isospin- scattering amplitude
A3/2, since π+p is an isospin- state:

We can compare this with pion–nucleon scattering in which the initial state is not purely I = . Looking at our
last chart, there are four possibilities: π0p, π+n, π–n, or π–p. It’s always easier to do experiments with proton
targets than with neutron targets, since hydrogen is easily available. There is no corresponding substance made
out of neutrons. And it’s always easier to do experiments with charged pions, because charged beams can be
guided with magnets. Neutral pion beams are much harder to manipulate. So I will consider the π–p amplitude.
The initial state π–p as we see from the multiplet table has I3 = −½, and therefore we know from standard
Clebsch–Gordan rules (just replacing J’s by I’s) that a π–p state will be some linear combination of the state with I
= and I3 = −½ and the state with I = ½ and I3 = −½. I looked things up in a table, and I found the coefficients are
1/ and − :

Thus if I compute the forward scattering amplitude for π–p → π–p, I obtain

Now I know there is a big resonance, the Δ++, with isospin- . I don’t know of course whether there might
be, by some incredible fluke, a second resonance with isospin-½ sitting at exactly the same point. Certainly the
simplest assumption is that there isn’t any such resonance, and therefore that the imaginary part of the isospin-½
amplitude will be relatively small. Then

That is, σ(π–p) should look the same as σ(π+p) but diminished in height. If I let h be the height of σ(π+p), then the
height of σ(π–p) should be h. And indeed if you actually look at the experimental data, which are available in all
sorts of tables,8 there is a corresponding peak in the π–p amplitude, and this height is one third of the peak of
σ(π+p). See Figure 24.4. Isospin is vindicated! The peak in the π+p amplitude is a little more than 200 millibarns,
and that of σ(π–p) is about 70 millibarns. It’s a beautiful check.

Figure 24.4: σT for π+p and π–p scattering, with Δ++ resonance

A second application of isospin follows if we fold together isospin with some earlier concepts. We know that
composite states of spinless identical particles must be symmetric in the spatial variables. Once we introduce
spin, the state must be either symmetric or antisymmetric in the product of space and spin variables, depending
on whether the particles are bosons or fermions. It’s interesting to ask what happens if we introduce isospin. The
reasoning is very simple. If I have a two-particle state made out of bosons, say, I make that state by hitting the
vacuum with two creation operators

The subscript (i) tells me what the isospin is, or more precisely what the value of I3 is. Now since these are bosons,
these creation operators commute and therefore a fortiori just as a consequence of our general formalism for
handling multi-particle states, the state is symmetric under the interchange of all variables. So for a multi-particle
state built of identical bosons, the state must be symmetric in space variables times spin variables times isospin
variables. By the same reasoning, a multi-particle state built of identical fermions must be antisymmetric in space
variables times spin variables times isospin variables. In the Fermi case this is sometimes called the generalized
Pauli principle. It is simply a consequence of the algebra of creation and annihilation operators for independent
particles, that they commute for bosons and anticommute for fermions. It is in fact totally free of dynamical
assumptions. It is simply a consequence of our bookkeeping rules, which have no physical content. But it does
have consequences; it makes some implications of isospin easy to see.

For example, suppose we have a particle called X, some unstable particle that decays into two pions in some
charged combinations, I won’t specify what:

If X decays to some state with two pions, it must have some definite isospin, if we assume the interactions are
isospin symmetric. If X has even J, the two pions, since they have no spin, must be in a state of even L, which is
symmetric in space. And therefore since the overall state must be symmetric, it must also be even in isospin. That
is to say, Itot must be an even number. Thus even J must be associated with I = {0, 2} and odd J must be
associated with the antisymmetric isospin wave function, and the antisymmetric combination of two spin-1’s is I =
1. For instance, a particle with angular momentum 1 and isospin 2 or 0 is forbidden from decaying into two pions,
if its decay goes through the strong interactions. Conversely a particle with angular momentum 1 and isospin 1
(such as the ρ meson) is allowed to decay into two pions. As an example of the other case, there is the ω meson
which has angular momentum 1 and isospin 0. Its principal decay mode is into three pions. There is a very tiny
admixture of two pions, but that’s believed to be due to the intervention of electromagnetism, which of course
does not respect isospin invariance.
 You may have noticed that I have started talking about particles other than pions and nucleons. We know that
there’s a host of strongly interacting particles. Indeed most particles participate in the strong interactions. Only a
very few do not: electrons and muons, their neutrinos, the photon and the graviton, the intermediate vector bosons
of the weak interactions.9 Strongly interacting particles are called hadrons. Although we have done our analysis
just with pions and nucleons, we know that any system of particles that participates in the strong interactions must
observe isospin invariance. We know that from field theory and from the fact that the strong interactions are
strong. After all, even if the particles are not pions and nucleons, even if we’re only discussing pion–nucleon
scattering, these other particles can occur as internal lines. For example, we could scatter a π meson off a proton
and exchange a ρ meson, as in Figure 24.5, a particle we had barely mentioned until now. Or we could build
complicated internal loops with Λ’s and Σ’s and what have you running around inside the loops. Now in general we
can’t compute these effects perturbatively, because these are strong interactions (though we can, for example,
compute the residue of the ρ meson pole). Remember, the characteristic strength of the coupling constant is on
the order of magnitude of 10. On the other hand, unless something miraculous is going on, there is no reason for
believing these effects are small. Thus if there were other strongly interacting particles which did not respect
isospin invariance, we would expect them to corrupt the isospin invariance of the pion–nucleon system, and we
have seen nothing that suggests that. That’s not a proof. Maybe some crazy dynamics comes along and makes
all these effects, individually large in perturbation theory, sum up to be something small. If that were true, it would
be very exciting. It would tell us something about very strong interactions, but it does not seem likely. And putting
aside that possibility, we know that everything that interacts strongly with the pion–nucleon system must interact in
a way that conserves isospin, so the isospin invariance of pion–nucleon interactions will not be corrupted.
Remember, it’s the total pion–nucleon force that is known to be isospin invariant, and therefore if the ρ interacted
in an isospin-violating way, this would produce a force between pions and nucleons that did not obey the
assumption we originally put down, an assumption that seems well supported by the data. That’s a pretty vague
argument, but it’s powerful. Of course, this doesn’t apply to particles that interact electromagnetically; the
electromagnetic force is not isospin invariant. The strong interactions are isospin invariant because pion–nucleon
scattering is isospin invariant. We could say the same using nucleon–nucleon scattering instead.

Figure 24.5: Scattering a π meson off a proton with ρ meson exchange

24.4Hypercharge and G-parity

Are there any other quantities that we know are conserved exactly for the strong interactions? I don’t mean
angular momentum or linear momentum or things like that; I’m talking about internal symmetries. There are. We
know of two conservation laws, internal symmetries, that hold good for all interactions, not just the strong
interactions. These are baryon number, B, sometimes called “nucleon number”, and Q, electric charge. There’s
also a discrete internal symmetry, charge conjugation, and we’ll talk about that later. For the moment I just want to
talk about these things that are associated with an infinitesimal phase transformation.

Baryon number is pretty trivial; it just goes along for the ride. As far as anyone knows, baryon number
commutes with isospin:

No one has ever observed an isotopic multiplet that has baryons and mesons mixed together, alternating in
isospin or something like that. It’s just an extra conservation law which we’ll add in a little footnote at the end. On
the other hand,

electric charge certainly doesn’t commute with isospin, because electric charge is not constant within an isotopic
multiplet, as we can see just from the proton–neutron case; they have different electric charges. Indeed if we
restrict ourselves to a system of nucleons and pions, we can write down the commutators. The states we wrote,
which were I3 eigenstates, were also Q eigenstates:

and raising I3, going from a π0 to a π+ or from a π– to a π0 or from a neutron to a proton, also raises the electric
charge. So the commutator of Q with I± equals ±I±:

(This is just for the system of pions and nucleons. We don’t know that it’s true in general.) That suggests we
define an object equal to Q minus I3, which is denoted Y, the factor of being added to make Y an integer:

Y is called the hypercharge.10

Conservation of isospin and conservation of electric charge trivially imply conservation of hypercharge. As we
have defined it, Y commutes with isospin, at least for the system of nucleons and pions:

Of course that doesn’t prove that Y commutes with isospin in general. It could be that this commutator of two
conserved quantities, which must itself be a conserved quantity, is some conserved quantity X which vanishes for
the particles we have discussed until now, but

If that were so, we would have not five conservation laws—three components of isospin, baryon number and
electric charge, or equivalently hypercharge—but six! There would also be the conservation of X which is
something altogether different. That would be very nice; it would introduce an extra simplification into the theory of
the strong interactions. In fact we would get three new conserved quantities. Because of the three components of
isospin, X had better carry an isovector index. But it ain’t so. We will accept the simplest possibility that this
commutator is zero for all particles, not just evaluated between states made up out of pions and nucleons. We
therefore deduce that hypercharge, like baryon number, commutes with isospin and is just something along for
the ride.

The computation of hypercharge for a given particle is made easy by observing that the average value of I3
over an isotopic multiplet is zero: we always have equal numbers of plus and minus factors. Therefore

The average hypercharge áYñ equals twice the average charge áQñ over the isotopic multiplet. But this average
hypercharge is in fact the value for each member of the isotopic multiplet, because hypercharge is constant over
the multiplet. For example, the nucleon multiplet, the proton and neutron, have charge 1 and charge 0
respectively, and thus both proton and neutron have a hypercharge of 1. The pion multiplet has average charge
zero, so each pion has hypercharge of zero. This means that for the system of nucleons and pions, the
conservation of hypercharge is really trivial, since this is precisely the value of the baryon number assigned to
these particles. But this is not true if you look at other strongly interacting particles. For example, the Λ hyperon11
is a particle with baryon number 1 and it is an isotopic singlet and is electrically neutral, so it has Y = 0. Likewise
the K meson is an isodoublet with charges +1 and 0, so it has hypercharge 1 even though it has a baryon number
zero. So in general if we consider strongly interacting particles beyond the proton and neutron, the conservation of
hypercharge is independent from the conservation of baryon number and gives us useful constraints. Once again,
let’s summarize these results with a table:

So far then, the continuous part of the internal symmetry group of the strong interactions is a five-parameter
group generated by exponentiating the three components of isospin, baryon number and hypercharge. Isospin
obeys angular momentum commutation rules. Baryon number and hypercharge commute with each other and
with isospin:
 We also have a well-known discrete internal symmetry which we have discussed in some detail (§6.3, §11.3,
and §22.3), charge conjugation. We have to work out the commutators of charge conjugation with isospin. It is
perhaps easiest to start with the nucleon system. The J3 isospin current is, from (24.34),

This is the difference of two currents of the sort we have discussed before. We can work out how this changes
under C from the chart on p. 469. The key equation is

Then

and therefore the integral of its zeroth component, the generator I3 also changes sign. The J1 current is much the
same,

It also turns into minus itself under charge conjugation:

and so does the generator I1. But J2µ is different:

The difference does not come from the overall factor of i. If you work in the Majorana basis, you will find,
surprisingly,

What’s going on is that only J2µ is antisymmetric under the interchange of the spinors ψ A and ψ B on either side of
the Dirac matrix; the others are symmetric under this interchange. So J2µ is unchanged under C:

We should check that the same rules are obeyed by the pion part of the current, but in the interest of time I will
assume that they are obeyed in general:

(Again, there might be some additional term on the right-hand side that happens to vanish for pions and nucleons.
That would be groovy: it would give us an additional conserved quantity. But unfortunately it is not so.)

This complicated set of rules, different for the three components Ii, makes charge conjugation a somewhat
awkward object to work with if we are considering the strong interactions. It’s convenient to define a new operation
G, sometimes misleadingly called “G-parity” (it’s got nothing to do with space reflection). It is defined as the
product of charge conjugation times a rotation through 180° about the 2-axis in isospin space:

Note that the order of the factors is irrelevant, since U C commutes with I2. The motivation for this definition is that
the 180° rotation about the isospin 2-axis changes the sign of both I3 and I1, but does nothing to I2, and cancels
out the effects of charge conjugation. So G, beautifully enough, commutes with all three components of isospin:

Thus we can assign to isotopic multiplets, provided they contain both particles and antiparticles (like the pion
triplet), definite values of G.12 Of course, G turns nucleons into antinucleons because it has charge conjugation in
it. For the pions it’s easy to see what happens. The π0 is even under charge conjugation since among other
things, it couples to iγ5p, a bilinear well known to be even under charge conjugation. The 180° rotation turns π0
into −π0. Since G commutes with isospin, what I say for the π0 must be true for both the π+ and the π–. Therefore
the pion field transforms as

Thus the pion, the particle created by the pion field acting on the vacuum, is G odd, and therefore we obtain a
useful selection rule from G-parity conservation: it tells us, for example, that the process 2π → 3π, pion production
in π−π scattering, must be forbidden because the initial state has G = +1 and the final state has G = −1. This
process is not forbidden by anything else. It is not forbidden by isospin conservation; you can indeed put together
five vectors to make an isoscalar: stick three of them together with ϵijk, and dot the last two to make a scalar. It is
not forbidden by parity, even though the pion is pseudoscalar: this is a five-particle vertex, and therefore involves
four independent 4-vectors, which we can stick together with ϵµνλρ. It is not forbidden, to my knowledge, by
anything except the combination of charge conjugation and isospin, that is to say by G-parity.

1[Eds.] This lecture does not appear either in the Hill–Ting–Chen notes, or in Coleman’s notes. This chapter is
based on the videotape of Lecture 24 and Peter Woit’s notes.
2 [Eds.] Werner Heisenberg, “Über den Bau der Atomkerne”, (On the structure of atomic nuclei), Zeits. f. Phys. 77
(1932) 1–11, English translation in D. M. Brink, ed. Nuclear Forces, Pergamon Press, 1965.
3 [Eds.] Figure 24.1 is based on Figure 9, p. 104 of Fay Ajzenberg-Selove, “Energy Levels of Light Nuclei, A =
11–12”, Nuc. Phys. A506 (1990) 1–158. The ground states of 12B and 12N correspond to an excited state of 12C,
about 15.1 MeV higher than its ground state. This has been set as the zero for the whole diagram, as in the
original figure. See Steven Weinberg, Lectures on Quantum Mechanics, Cambridge U. P., 2013, p. 133, and
Brink, op. cit., Figure 7a, p. 59 for a similar (and quantitative) comparison between the energy levels of 11C and
11B.

4 [Eds.]But it suggests a mass difference with the wrong sign: the proton is less massive than the neutron, and its
electromagnetic energy should make the proton more massive.
5[Eds.] No one had managed to do it at the time of Coleman’s lecture (1976), but a recent calculation of the n-p
mass difference using QCD lattice gauge theory together with perturbative QED gives pretty good results: mn − mp
= +1.512 MeV, about 17% larger than the empirical value of +1.293 MeV. See Sz. Borsanyi et al., “Ab initio
calculation of the neutron–proton mass difference”, Science 347 (2015) 1452–1455. For a discussion of this
result, see Frank Wilczek, “A weighty mass difference”, Nature 520 (2015) 303–304.
6 It should really be called “isobaric spin”, rather than “isotopic spin”, but it isn’t; it’s called isotopic spin by a
historically well-embedded slip of the tongue. ([Eds.] Isobars are nuclei with the same number of nucleons;
isotopes are those with the same number of protons; isospin transformations conserve the number of nucleons,
but not necessarily the number of protons.)
7 [Eds.] The combined symmetry CPT is always good. Since CP is violated by the weak interactions (note 9, p.
240), T must be as well.
8 [Eds.] Figure 24.4 is based largely on the graph on p. 229 in Murray Gell-Mann and Kenneth M. Watson, “The
Interactions Between π-Mesons and Nucleons”, Ann. Rev. Nuc. Sci. 4 (1954) 219–270. Coleman earned his PhD
under Gell-Mann at Caltech in 1962. The horizontal axis is the kinetic energy Kπ of the pion in the laboratory frame
(proton at rest); it is related to the total energy (proton plus pion) in the CM frame by Ecm = .
During the lecture, Coleman explained that he had drawn the figures from memory. John LoSecco, now a
professor at Notre Dame and the Teaching Fellow for the course in 1975–76, pulled out his copy of the Particle
Data Group booklet, looked up the relevant graph, and held it aloft so that Coleman could consult it. Coleman
joked that he’d lost his own copy, but they refused to send him a duplicate until there was a new printing.
9 [Eds.] The gauge bosons of the weak force, formerly “intermediate vector bosons” or IVB’s, are now universally
known by W+, W− and Z0. All were found in a series of experiments at CERN in the period 1981–83. This work
was recognized by the 1984 Physics Nobel Prize, awarded to Simon van der Meer and Carlo Rubbia, the
experimental team’s leaders.
10[Eds.] Coleman has not yet discussed the quantum number S, strangeness, associated with strange quarks.
Hypercharge was introduced independently by Murray Gell-Mann, and by Kazuhiko Nishijima and Tadao Nakano
as baryon number plus strangeness. The relation Q = I3 + Y is often called the Gell-Mann–Nishijima relation,
and will be introduced in (35.52). See M. Gell-Mann, “The Interpretation of the New Particles as Displaced
Charged Multiplets”, Nuov. Cim. 4 (Supplement) (1956) 848–866; T. Nakano and K. Nishijima, “Charge
Independence for V-Particles”, Prog. Theo. Phys. 10 (1953) 581–582; K. Nishijima, “Some Remarks on the Even-
Odd Rule”, Prog. Theo. Phys. 12 (1954) 107–108; K. Nishijima, “Charge Independence Theory of V Particles,”
Prog. Theo. Phys. 13 (3) (1955) 285–304. See also Ryder QFT, pp. 14–15. For the historical background, see
Crease & Mann SC pp. 177–179.
11 [Eds.] “Hyperon” is an old-fashioned term for a baryon containing one or more strange quarks, i.e., with
strangeness S ≠ 0, but not any charm, top or bottom quarks. The term was in use before those other quarks were
postulated. See note 10, p. 520.
12[Eds.] Time may have prevented Coleman from making the usefulness of G-parity as clear as he might have
wished. The point is that charge conjugation invariance is of limited direct utility, because for a particle to be an
eigenstate of C it must be electrically neutral. But if you combine C with an isospin rotation that turns, e.g., a π+
into a π−, then charged particles can be eigenstates of this joint operation. As it is conserved, you can immediately
read off implications for physical processes.G-parity will be revisited in §35.4.

25
Coping with infinities: regularization and renormalization

Thus far we’ve been slovenly with renormalization, but we’ve gotten away with it. For example, returning to the
ps–ps theory (23.1), we had for the vertex function

The two graphs contributing to ′ at O(g3) are shown in Figure 25.1. At high k, the integral

Figure 25.1: O(g3) graphs for ′ in ps-ps theory

for the left-hand diagram goes as

This is divergent, but it’s only logarithmically divergent. The right-hand graph contributes the O(g3) counterterm
−E3γ5, (see (23.25)) which cancels the divergence.

So far, this is about as far as we’ve taken renormalization. Throwing around ill-defined quantities and finding
that they always end up in convergent combinations isn’t really enough. It’s time to answer the fundamental
question. Is renormalization necessary and sufficient to get rid of infinities?1

25.1Regularization

No one knows of a quantum field theory that is nontrivial and finite. It was realized long ago that we should, as an
intermediate step, render finite those quantities that are formally infinite before carrying out calculations, to avoid
making ad hoc cancellations of infinite quantities. This process is called regularization. Typically it involves
introducing a parameter, often denoted Λ, with Λ < ∞. At the end of the calculation, we restore the quantities to
their original values (usually by taking the limit Λ → ∞). If all goes well, there will be no trace of the intermediate
regularization, so we can be reasonably confident of the results. Several regularization schemes have been
introduced.

Method 1: Brute force

 We simply throw away (k) and (k) for |k| > Λ. Alternatively, we can put the entire theory in a box of finite
size, reducing the degrees of freedom to a finite number. This procedure is admired by the mathematically
inclined. The disadvantages are serious: we lose both Lorentz invariance and gauge invariance.

Method 2: Propagator modification

This procedure was pioneered by Feynman, and by Stueckelberg and Rivier2 and developed extensively by
Pauli and Villars,3 by whose names it is often known. The idea is to replace propagators in the Feynman integrals
by expressions that fall off fast enough at high momentum so that loop integrals will be finite. In the simplest
version, we make the replacement

which is O(1/k4). Similarly,

Changing a propagator’s momentum dependence from inverse square to inverse fourth power may not be enough
to make some diagrams finite. More generally,

We can look at the behavior of this for high k2 by expanding each term; for instance:

The choice of the coefficients {ci} determines the high k behavior of the regularized propagator (25.4):

and the pattern should be clear. The case N = 1, c1 = −1 and M1 = Λ reproduces (25.2). With this procedure we
can make the propagator fall off like any inverse polynomial in k2. Notice that the Mi’s have to have different values
to solve for the ci’s.

There is also an operator form of the Pauli–Villars procedure. Suppose the original theory has the form

Introduce a Pauli–Villars regulator field, ϕ 1. The new Lagrangian is

so that the contraction is given by (see (9.29))

The field ϕ 1 is a little strange, appearing as the imaginary part of Φ. Because Φ is not Hermitian, neither is it
Lagrangian. Define N 1 as the number operator for ϕ 1 particles. Then

That is, ϕ 1 anticommutes with (–1)N1. We can gain some insight into what is going on by defining a new inner
product,
This metric is not positive definite. If the state |añ has an odd number of ϕ 1 particles in it,

For states without ϕ 1 particles, the inner product is its old self. The purpose of the new inner product is to make Φ
Hermitian. With it,

so

In the old metric, which was positive definite, the Hamiltonian wasn’t Hermitian, and thus didn’t conserve
probability. In the new metric, not positive definite, the Hamiltonian is Hermitian, probability is conserved, and the
S-matrix is unitary. At the end of our calculations, we won’t be interested in amplitudes that contain the phony ϕ 1
particles in initial or final states. When we take M → ∞, we remove them, because energetically they cannot be
produced. So we have every expectation that in this limit the resulting theory will be sensible.

Regulator fields have many desirable properties. They preserve Lorentz invariance and internal symmetries
in theories with massive particles (though if the original theory is massless they may spoil some symmetries).
They conserve probability in processes low in energy compared with the cut-off mass, and with some
modification, they even preserve gauge invariance in QED. Finally, they are easy to introduce.

Method 3: Dimensional regularization

The idea here is to modify the number of spacetime dimensions from 4 to a continuous variable, d, chosen to
make integrals (or sums) convergent. At the end, one takes the limit d → 4. This procedure was proposed
independently by several physicists4 in 1972, but is usually associated with ’t Hooft and Veltman.

Consider the integral

Here, k is taken to be a vector in d-dimensional Euclidean space:

whereas for a vector in d-dimensional Minkowski space

The integral (25.14) is convergent if n > d/2. (We computed a similar integral (I.1) in d = 4 Minkowski space for the
integral table on p. 330.) We go from Minkowski space to Euclidean space via a Wick rotation (15.7), to simplify
the calculation and remove the poles. Later on, in §28.2, we’ll use this trick to turn an oscillating exponential into a
damped exponential, and thus guarantee convergence of certain integrals. To evaluate this integral, we use a trick
to convert a denominator into an exponential. Recall the gamma function,

If we change variables, letting t = αλ, with α real and positive, then

In particular, letting α = (k2 + a2),

so (25.14):
But

so

(using (25.16) again, this time with α = a2 and n → n − (d/2)). Or, returning to the original expression,

(This expression is identical with (I.1) on p. 330 when d = 4, modulo factors of (2π)4 and an i, from (15.66).) The
idea is now to adopt this formula for complex and continuous values of d. If you stay away from even integers d ≥
2n, the expression is well-defined. You do renormalization with regularized quantities in terms of an arbitrary d,
and only after obtaining expressions for the graphs plus counterterms in convergent combinations (with poles in (d
− 4) cancelling) do you take the limit d → 4.

Technical issues arise in changing the number of spacetime dimensions, of course. For example, you can’t
maintain

as a dimensionless constant, because that is true only for d = 4. (Sometimes, “dimension” will be used as
shorthand for the powers of mass, [M], or inverse powers of [L], length, of an object.) And how should we define a
set of Dirac γ matrices in a different number of spacetime dimensions? This is a particular problem for γ5. But
even with a simpler theory like

there are complications. The quantity ddxL must be dimensionless, so [L ] = [M]d. Because

(see Problem 2.1, p. 99 and (S2.4), p. 101), we know µ has dimension 1, just as mass does in four dimensions.
However, we must have [λϕ′4] = [M]d, so

Only in four dimensions is λ dimensionless. To keep λ dimensionless as we change d, we introduce a parameter ν

with the dimension of mass, [M]1, and rewrite the interaction as

You might think that after we take the limit d → 4, all ν dependence would go away. But that is not so, as the next
example shows.

Figure 25.2: O(λ2) correction to (4) in ϕ 4 theory

EXAMPLE. An O(λ2) contribution to the four-point function (4)

Consider the diagram below:

The contribution A from this diagram is proportional to

where a contains masses, external momenta and perhaps Feynman parameters. I have suppressed the
Feynman parameter integral (15.58). According to (25.18),

Let ϵ ≡ 2 – (d/2) be very small; then5

Perhaps substituting (25.26) into (25.25) and evaluating all but the pole term at ϵ = 0 would give the finite part of
this expression. In fact, more care is required. We pull out a factor of ν4−d, the dimension of this Green’s function;
the remainder is dimensionless.6

Substituting (25.26) into (25.25),

Rewriting,

Putting it all together,

Had we set ϵ = 0 (or equivalently, d = 4) prematurely, we would have lost the ln(ν2/πa2) term.

A companion to dimensional regularization is called minimal subtraction, a method of determining

counterterms. It makes no reference to the physical mass and physical coupling constants, so it is not good for
comparison with experiment. Theorists like it because it makes no comparison with experiment, and because it is
easy. It amounts to just throwing away the pole terms in the dimensionally regularized integrals.

Continuing with our example, we found (25.27) that (suppressing the Feynman parameter integral) the four
point function gave a contribution proportional to

The coefficient of this pole is unambiguous. Minimal subtraction says we introduce a counterterm in LCT to cancel
it:

There’s a systematic way to add counterterms, to which we now turn.

25.2The BPHZ algorithm

To make things simple, we’ll restrict our attention to theories describing only spin-0 and spin-½. To explain this
iterative algorithm we need to introduce some useful terminology. We’ll look at Lagrangians of the form

and L0 is a sum of free Lagrangians. Here is a table of some typical Lagrangians with their fi, bi, and di values:
 We’ll need these numbers shortly.

The superficial degree of divergence D

In the integral associated with any Feynman diagram, let PN be the power of the momenta in the numerator
and PD the power in the denominator. For instance, every loop integral puts 4 powers of momenta into the
numerator (in the form of d4p); every boson propagator puts 2 powers into the denominator, every fermion
propagator puts 1 power into the denominator, and every derivative puts a factor of p into the numerator. Then
define D, the superficial degree of divergence, as

and so on. For example, consider the following diagrams. In ϕ 4 theory,

which is superficially quadratically divergent. In pion–nucleon theory, we have

superficially linearly divergent;

superficially logarithmically divergent; and

which is superficially convergent. Well, why do I say “superficially”? Consider this diagram:

The rule says it’s superficially convergent, but in fact it’s divergent!7 Despite its inability to predict a diagram’s
divergence accurately, D will be very useful, as we’ll see.

Taylor expansion about the point p = 0

For non-massless particle theory, Feynman diagrams are analytic functions of external momenta around pi =
0. We will henceforth assume that there are no massless particles in our theory, so we can Taylor expand the
expressions associated with our diagrams:

The first diagram lacks a linear term (all the p terms come from propagators, p2, or loop integrals, p4); the zeroth
and second-order terms are the first terms in the Taylor expansion about p = 0. In the second diagram, there are
zeroth and first-order terms in both p and p′.

With these preliminaries out of the way, I can now describe the algorithm, originally applied to theories
renormalized with cut-off parameters:

The BPHZ algorithm

 1.Compute in perturbation theory to all orders until you reach a 1PI diagram with D ≥ 0.

2.Add to L counterterms to cancel the terms in the graph’s Taylor expansion (about zero) of order ≤ D.

3.Return to 1, continuing to compute with the new, corrected L′ = L + LCT.

The algorithm also explains exactly what form these terms take, as we’ll see shortly. The algorithm appears
in an article by Bogoliubov and Parasiuk.8 The power of the algorithm derives from a theorem by Hepp.9
Zimmerman10 showed that all ultraviolet divergences are removed by the algorithm, so the procedure is known as
BPH or BPHZ renormalization.

Bogoliubov’s algorithm removes all divergences (if the theory does not involve massless fields). The
Hepp’s
Green’s functions resulting from the algorithm are independent of the cut-off Λ as Λ → ∞, to all orders in
theorem:
perturbation theory, no matter what the regularization procedure.

The algorithm solves the problem of renormalization, since the counterterms are built up correctly. At each order
of perturbation theory, the only new problems arise from new divergences connected to superficially divergent
diagrams. Other divergences are taken care of automatically by earlier counterterms. We’ll see how this works
with specific examples.

25.3Applying the algorithm

Instead of just stating theorems in a loud voice, I will now compute in a simple way the superficial degree of
divergence of a particular Feynman graph in a theory of this kind. This will enable us to see the difference between
a renormalizable and a non-renormalizable theory.11 In addition, I will state a rule for constructing the
counterterms based on the superficial degree of divergence.

I will need to define some terms. FE is the number of external Fermi line in a graph. FI is the number of
internal Fermi lines. Likewise BE is the number of external Bose lines, and BI is the number of internal Bose lines.
Let ni be the number of vertices of type i; that is to say, coming from an interaction of the i th type in our effective
Lagrangian (25.32), and as before di is the number of derivatives in that interaction.

I will write a formula for the superficial degree of divergence D of such a graph. This will simplify things
enormously. First I’ll just count powers. Every internal Bose line gives a factor of one over p2 from the propagator
and a factor of d4p from our integration, or two powers of p in the numerator. Some of those will be reduced by
delta functions at the vertices, but I’ll take care of that later. Every internal Fermi line gives us one d4p in the
numerator and one power of p in the denominator, a total of three powers of p in the numerator. Every derivative
interaction will give us one power of p in the numerator.

I’ve overcounted the internal momenta because not all of their integration variables are independent: every
vertex has a delta function, and that knocks out four integration variables. These are all 1PI graphs and therefore
a fortiori connected. However, there is one overall delta function left over for energy momentum conservation. So
I’ve overcounted the number of delta function restraints by 4. This is simply a general formula for what I did before
when I was counting numerators and denominators. Four powers of the internal momenta for each internal line of
any kind, cut down by two for a Bose propagator, reduced by one for a Fermi propagator, knocked down by four
for each delta function at a vertex, except for one delta function left over for overall energy momentum
conservation. That one doesn’t restrain the loop momentum. Then

In this form the expression for D is a mess. Fortunately we can simplify it by using the laws of conservation of
boson and fermion ends. Every external boson line has one end that winds up on a vertex and one end that is left
hanging. Every internal boson line has two ends, each tied to a vertex. Every vertex of i th type has bi boson ends
tied to it. Then

This is the law of conservation of boson ends. Likewise there is a law of conservation of fermion line ends:
By elementary algebra we may eliminate the factors involving internal lines and only be left with factors involving
the vertices and the external lines:

Substituting these into (25.41) gives

This formula is extremely nice because it tells us how much more divergent a graph becomes when we add an
interaction of a given type. We can simplify it further if we define the index of divergence, δi, of an interaction
Lagrangian Li:

Then

I won’t prove it, but this formula contains the explicit prescription for constructing counterterms:

I am obviously going to have to introduce a lot of counter terms if I have an interaction in my theory with δi positive.
Whenever I add an extra internal vertex of that type, I increase the superficial degree of divergence by one, I have
to make more subtractions in my Taylor expansion, and I have to add therefore a counterterm with more
derivatives. I want to give two specific examples to show you what’s going on.

EXAMPLE. A ϕ 4 interaction

I will figure out the counterterms iteratively by using the formula (25.47). There is only one interaction, the term in
ϕ 4. For this interaction, the number b1 of boson lines is 4, the number f1 of fermion fields is zero, and the number
d1 of derivatives is zero. The index of divergence δ of this interaction is zero:

Now let us compute the superficial degree of divergence, D, from (25.47). No matter how many internal ϕ 4
vertices we have, even if I drew a complicated diagram that would cover a couple of blackboards, elementary
algebraic counting shows this term δ1 = 0, so it contributes nothing. There are no fermions in the theory, so FE = 0.
The superficial divergence D is determined just by the number of external boson lines:

Graphs with more than four external boson lines will always be superficially convergent, and by the Bogoliubov
prescription will require no counterterm. I need consider only graphs with BE ≤ 4. Because the theory is invariant
under ϕ → −ϕ, we don’t have to consider graphs with odd numbers of boson lines: they vanish. Thus we have to
look only at three cases: BE = 0, BE = 2, and BE = 4. The possibility BE = 0, D = 4 is irrelevant: graphs with no
external lines are vacuum to vacuum graphs, and we throw those away. The next case is BE = 2, D = 2. This is of
the form

According to (25.48) we need a counterterm with two ϕ’s and two derivatives. We will have to subtract out the first
two terms in the Taylor series expansion. Therefore to any order in perturbation theory this graph will introduce
counterterms proportional to ϕ 2, which will cancel the zeroth order term in the Taylor expansion with some
coefficient depending on how far we’ve brought in perturbation theory and what the cutoff is, and some terms
proportional to (∂µϕ)2:

Then I have BE = 4, D = 0, corresponding to this graph:

The only counter term introduced here is

There are no derivatives because D = 0, and we only go to zeroth order in the Taylor expansion, so we have no
powers of momentum. The counterterms are new interactions, and for consistency I should check that they don’t
change my divergence counting. The C counterterm is proportional to ϕ 4, so it, too has δ = 0. The B term has two
ϕ’s and two derivatives so again has δ = 0. The A term has only two ϕ’s and no derivatives so has δ = −2, which
is groovy. They don’t change the divergence counting of the original Lagrangian.

Thus in this theory, the counterterm Lagrangian is a sum, with some coefficients I have to compute, of a ϕ 2, a
(∂µϕ)2 and a ϕ 4:

These counterterms can be interpreted in our usual way by rescaling the field, to make the (∂µϕ)2 coefficient its
usual self, . We assemble the other two terms to define the bare mass and the bare coupling constants. The
result therefore of Hepp’s theorem applied to this example is to make all observable quantities in this theory, to
any finite order in perturbation theory, independent of the cut-off, in the large cut-off limit. This can be done if the
field is appropriately rescaled and if the bare parameters are chosen in an appropriate cut-off independent way,
because the terms we have added are of the same form as the terms that were there in the first place. That is what
we mean when we say a theory is renormalizable. The ϕ 4 interaction is a renormalizable theory. You choose the
bare coupling constant in the appropriate cut-off independent way, the bare mass in an appropriate cut-off
independent way, rescale the field in an appropriate cut-off independent way, and all the divergences will cancel
to any finite order in perturbation theory. I want that point firmly in your head.

EXAMPLE. A ϕ 5 interaction

This interaction has five boson lines, and δ = 1. When we consider a graph containing more and more of these ϕ 5
interactions, the superficial degree of divergence will get larger and larger in the graph’s Taylor expansion about
zero (see the example on p. 344). We’ll have to make more subtractions and we need more and more different
kinds of counter terms. Not only do the coefficients change order by order in perturbation theory, but their
qualitative character changes as well. A graph that goes to sufficiently high order in the ϕ 5 interaction, a graph
with two external boson lines, will have D equal to 1 million—that happens to millionth order in the ϕ 5
interaction—and therefore we would have to subtract something with two ϕ’s and a million derivatives. This theory
is non-renormalizable. We are off on the unending escalation of ambiguities that characterizes such theories. As
we go to higher and higher orders in perturbation theory, we need more and more different kinds of counter terms
that cannot be interpreted as simply a rescaling of ϕ and a redefinition of the parameters that occur in our original
Lagrangian.

People sometimes say, “Well, so what? You’ve got a prescription that fits most things uniquely, you know the
Bogoliubov prescription is unambiguous and tells you what those counter terms are.” But the Bogoliubov
prescription is arbitrary; Bogoliubov invented it to make the theorem easy to prove. You don’t have to subtract at
zero, you could subtract at some randomly chosen point of momentum space if you want to avoid all those
thresholds. You could subtract different Green’s functions at different points, you could subtract the second-order
term in the Taylor expansion about the point zero, the third-order term in the Taylor expansion about some other
point. The whole thing is just an ad hoc prescription to make the algorithm run simply. If you get a ϕ 17 interaction,
or (∂µϕ)42 term that comes out as part of the counterterm prescription, there’s no reason why it shouldn’t have
been there in the original Lagrangian. A non-renormalizable theory involves an unlimited number of terms and free
parameters.

So this is the dividing line between renormalizable and non-renormalizable theories: either all the interactions
have δ less than or equal to zero, or some of the interactions have δ greater than zero. If you have positive δ’s
you are cooked; it is a non-renormalizable theory. It is bad news. I don’t know how to make sense out of them, and
nobody else does, either. Every few years someone has an idea about how to deal with them and every few years
he’s shot down.12

I should say that renormalization makes a lot of people nervous, dealing with a theory that involves infinite
quantities, the bare charge and the bare mass, in its Heisenberg equations of motion. Suppose at some future
date the constructive field theorists conquer quantum electrodynamics in the sense of establishing a rigorous proof
that shows if you put in a cut-off, the equations of motion have unique well-defined solutions, and those solutions
have a definite limit as the cut-off goes to infinity, presuming you adjust the bare coupling constants and the bare
masses appropriately as functions of the cut-off: they prove non-perturbatively what has been proved in
perturbation theory. In that sense they construct a mathematically well-defined theory, albeit through a limiting
procedure. Now there it is, a mathematically well-defined theory that obeys all the general assumptions you’d
want a quantum field theory to obey: it’s got local fields that commute for spacelike separations, it’s Lorentz
invariant, it has a particle spectrum, et cetera. Are you going to reject it out of hand just because you don’t like the
fact that it’s defined through a limiting procedure? That was Bishop Berkeley’s objection to the calculus. He said
infinitesimals didn’t exist. But calculus was later reformulated in terms of a limiting procedure, and you can
formulate renormalization in terms of a limiting procedure, through regularization. Maybe God did things that way,
with limiting procedures. And perhaps if there is a physical cut-off, it may be at some distance so small that it might
as well not be there for all practical purposes. It might be that gravity in its mysterious way does something
strange, although nobody knows how it could. But we can do dimensional analysis and see that the characteristic
(“Planck”) length of gravity is 10–33 centimeters, which is at least 10 orders of magnitude shorter than the current
experimentally accessible range of distances. And if there is a cut-off at that distance, who cares?

25.4Survey of renormalizable theories for spin 0 and spin ½

For the type of theories we are considering, scalar fields and Dirac spinor fields, the degree of divergence is
connected with the dimensionality of the interaction (or equivalently, with the dimensionality of the coupling
constant that multiplies the interaction) in a relatively simple way. We can see that by elementary dimensional
analysis. The derivative operator has dimensions of length to the inverse first power, or, in the units we are using,
where mass and length have inverse dimensions, dimensions of mass to the first power:

The action has the dimensions of Planck’s constant; that is to say, it is dimensionless:

Since d4x has the dimensions of L4, the Lagrangian must have dimensions of length to the inverse fourth power,
or equivalently, mass to the fourth power:

The Lagrangian for a scalar field contains a kinetic term (∂µϕ)2 with two derivatives and two ϕ’s. This term must
have dimensions of M4, so

By the same argument the spinor field has dimensions of mass to the because its Lagrangian is iψ ψ:

Counting only the dimensions of the fields and the derivatives (ignoring whatever dimension any coupling constant
has), the dimension (the power of M) of an interaction Lagrangian is

That is, not including the coupling constant,

(Let the raw dimension of an interaction be its dimension without including the coupling constant.) A check: if δi is
zero, the dimension is 4. If you remember the rules for dimensions you also remember the rules for computing the
index of divergence δi in powers of mass. Equivalently if you include the coupling constant and arrange matters so
the whole Lagrange density has dimensions [M]4, the dimension of the coupling constant is δi in units of inverse
mass.

An interaction is said to be of renormalizable type if the index of divergence δi is less than or equal to zero. As
we include more and more of these interactions going to higher order in perturbation theory, we do not increase
the superficial degree of divergence D; we will not need to add more and more counterterms. It is possible to make
a complete list of these renormalizable interactions in four dimensions. The minimum case of δ is −3. The case δ
= −4 is in principle possible with no derivatives, no fermions, and no bosons, but that’s not much of an interaction;
that’s just adding a constant to the Lagrangian. Therefore we’ll start with −3. Here the only possibility is a term
linear in a scalar field, ϕ:

(We’ll use ϕ and ψ generically. When I write ϕ in a theory with 21 different scalar fields in the Lagrangian, it could
be any linear combination of 21 such terms.)

We can get by with one scalar field and one derivative, which is not particularly interesting since that’s not Lorentz
invariant, and it also vanishes by integration by parts; or with two scalar fields, ϕ 2. Again it could be ϕ 1ϕ 2, for
instance, if there are two scalar fields.

At δ = −1, things are a bit richer. We could have ϕ 3, or ϕ∂µϕ (which is not Lorentz invariant), we could have ψψ (or
if our theory is not parity conserving, ψiγ5ψ). These three kinds of interactions with δ strictly less than zero are
sometimes called super-renormalizable. Although they require counterterms, they only require a finite number of
them in perturbation theory. When you put in enough of these interactions D becomes negative and no new
counterterms are required. Super-renormalizable theories are of course much nicer than merely renormalizable
theories, because the divergent part of the perturbation series terminates. Unfortunately, at least in four
dimensions the only super-renormalizable theories we can get are either trivial, in which the spinor products are
the only interactions, or the energy is unbounded below, if we allow the ϕ 3 interaction without a ϕ 4 term to
compensate for it. In fewer dimensions than four, of course, the counting is rather different, and you can find
theories that are super-renormalizable with sensible energy spectra. These are nice models to look at if you want
to do some rigorous mathematics and prove that a quantum field theory exists. There are still divergences to
handle, but it’s much easier than in four dimensions.

Finally we have δ = 0:

the genuinely renormalizable types of interactions. Here we can have ϕ 4; (∂µϕ)2, two derivatives and two scalar
fields; ψ ψ, the normal term that arises in the free Lagrangian; ψ iγ5ψ which you might encounter as a
counterterm in a theory with parity-violating interactions; and finally the two kinds of Yukawa coupling, to a scalar
or pseudoscalar field: ψψϕ and ψiγ5ψϕ. That’s it. That completes the list as far as the fields we have talked about.
In a little while we will talk about what happens when you allow for vector fields and how this formalism is
extended.

As you see, renormalizability is a very severe restriction. It’s one of the striking differences between
relativistic local quantum mechanics (i.e., quantum field theory), and non-relativistic quantum mechanics. In non-
relativistic quantum mechanics, there is no a priori constraint of any sort on the interactions. There may be two
body forces with arbitrary potentials, there may be three body forces, four body forces, etc. As far as anyone
knows there is no general criterion that restricts in any significant way the interactions between the particles. You
don’t want them to be so singular that the energy is unbounded below and so on, but aside from that anything
goes. In quantum field theory, if you accept renormalizability as a criterion that distinguishes sensible theories
from nonsensical theories, or at least those theories about which we can say something significant beyond lowest
order in perturbation theory from those we cannot, things are much more restricted. Once you have told me the
number of spinless fields and the number of Dirac bispinor fields in the theory, I have only a finite number of free
parameters which I can adjust, the coefficients of the renormalizable couplings.

Interactions of renormalizable type do not generate an infinite sequence of counterterms. However in normal
parlance we may use the word “renormalizable” in a slightly stronger sense: we not only want the number of
counterterms generated to be finite, but we want them all to be interpretable as redefinitions of parameters
multiplying terms that already occur in our initial Lagrangian; all counterterms are of the same form as terms in the
original Lagrangian. In the literature such a theory is called strictly renormalizable. For example, if we take a
theory of scalar fields, as well as theories that can be generated from it by rescaling the fields (such as by a wave
function renormalization counterterm), all the counterterms that arise in every order of the Bogoliubov iterative
procedure can be reinterpreted as corrections to the “bare” parameters of the theory.

Thus for example in the strict sense of renormalizability, our good old friend, the Yukawa interaction with a
pseudoscalar meson,

with δ = 0 (25.46), is not a renormalizable theory, because as we can see from our formula or just by counting, this
graph for meson–meson scattering

is logarithmically divergent: d4k over over k4. Equivalently, BE = 4, FE = 0, so (25.47) D = 0. On the other hand if I
add a ϕ 4 interaction to the Yukawa interaction,

I have a ϕ 4 counterterm that can be used to cancel out this divergence, and it is easy to check that the theory is
strictly renormalizable. The only divergent graphs are those that can be interpreted as redefining the parameters
in the Lagrangian. In the sense of strict renormalization, there is no point in talking about Yukawa theory as a one
parameter theory; it is a two parameter theory. You have to specify independently the Yukawa coupling g and the
ϕ 4 coupling λ.

It’s possible to give some general theorems that characterize large classes of strictly renormalizable theories,
involving only a set of spin zero and spin ½ fields (you’ll have to specify how many of each there are). I’ll give three
such theorems.

Theorem 1. The most general Lagrangian involving all interactions of raw dimension less than or equal to
four, or equivalently, with δ less than or equal to zero, is strictly renormalizable.

How do I prove that? I start with (25.47), moving some of the terms over to the other side:

Let a given diagram contain a divergence with D ≥ 0 and index δ. According to step 2 of the Bogoliubov algorithm,
I add counterterms to cancel the Taylor expansion of the divergence about p = 0 up to order D. FE tells me the
number of Fermi fields that I have to put into my counterterm, BE tells me the number of boson fields, and D tells
me the maximum number of derivatives I have to include to subtract the appropriate terms in the Taylor expansion.
(We might not need to go as high as D, because it’s possible that we already have counterterms to cancel that
order from earlier in the algorithm.) That is, for any diagram
But then the left-hand side of (25.72) is just the formula that enters into the definition (25.46) of δi; it’s the same
combination. So the δ of this diagram is

Thus the δ of the counterterms for a diagram is always less than or equal to the sum of the δ’s of the interactions
in the diagram:

It’s elementary algebra. I say less than or equal because I have to subtract all the terms in the Taylor series up to
order D. Thus if my original Lagrangian contains all monomials with δ less than zero, every counterterm I
introduce will be a monomial with δ less than zero, and therefore it can be reinterpreted as a renormalization of
the coefficient of one of those monomials.

Theorem 2. The most general Lagrangian involving all interactions consistent with some internal symmetry or
parity, of (raw) dimension less than or equal to four, or equivalently, with δ less than or equal to zero, is strictly
renormalizable.

Unless I am so perverse as to choose a cut-off procedure that all by itself violates the internal symmetry or
parity, parity-violating graphs or internal symmetry-violating graphs will not occur. Even though they may have a
superficial degree of divergence (D) greater than or equal to zero, I will not have to make any subtractions for
them because they are zero, and therefore all terms in their Taylor expansion are zero. Thus for example Yukawa
theory with a ϕ 4 interaction

is, by this criterion, strictly renormalizable, because it represents the most general interaction between these kinds
of fields consistent with parity. In principle, if it weren’t for parity, I could have a ψψϕ counterterm and a ϕ 3
counterterm or a term linear in ϕ as a counterterm, but those would all violate parity. Likewise the corresponding
isospin and parity invariant Yukawa theory:

The first term is our old Yukawa interaction (24.29) for a triplet of pions. Now just as before we have to add the
possibility of a ϕ 4 interaction, but the only one that is consistent with isospin invariance for ϕ—the only way I can
make an isoscalar without introducing derivatives—is (Φ • Φ)2. This is the most general interaction of this form
only involving terms with dimension less than or equal to 4, δ less than or equal to zero, which is invariant under
both parity and under isospin rotations. It is therefore strictly renormalizable. The only kinds of counterterms we
will encounter are terms of the same sort we had to begin with in the Lagrangian. It’s really very simple. Of course,
the reason it’s very simple is because I cheated on you: I told you Hepp’s theorem without telling you the proof. If I
had gone through the proof of Hepp’s theorem you wouldn’t think it was so simple. But once you have that big
theorem, everything else falls out.

Theorem 3. The conclusions of Theorem 2 remain true if any symmetry-breaking interaction is added to the
Lagrangian, provided the symmetry-breaking interaction’s (raw) dimension equals 1, 2 or 3. This result was
discovered in 1970 by Symanzik. We will call this Symanzik’s rule.13

The point here is that if you have an asymmetric interaction as well as a symmetric one but the asymmetric
interaction is of low dimension, with a negative δ, by (25.75) it will only introduce asymmetric counterterms that
also have a negative δ. For example if the only interaction in your theory that breaks the symmetry has δ ≤ –2, you
will only get counterterms that violate your symmetry considerations of δ ≤ –2. And therefore you will never
generate a higher value of δ than that of the original interaction. Such symmetry breaking is sometimes called
“super-renormalizable symmetry breaking”, or “soft symmetry breaking”. For example we could break the
symmetry in (25.77) by adding an unequal mass term for the π0,

to give the π0 a different mass than the π+ and the π–. That is a symmetry breaking term with δ = −2, of dimension
2, and indeed it is the only possible symmetry breaking term of dimension 2 or less consistent with parity and
charge conjugation, etc. This symmetry-breaking term will never generate any counterterms except those of the
same form, also consistent with parity and charge conjugation. Thus for example it is perfectly consistent with
renormalization within the framework of meson–nucleon theory to say that the theory is completely isospin
symmetric except for a difference between the bare mass of the charged pions and the neutral pion.

The bare masses of the nucleons are the same because that counterterm is never forced on you. The
counterterm has dimension 3, δ = −1. All the bare couplings, of dimension 4, remain symmetric. Although this is a
cute result, it unfortunately does not help us explain mass differences in nature (for example, between the neutron
and the proton) on the basis of electromagnetism. That interaction, ψγµψAµ, is of dimension 4. This theorem is just
what we don’t want; we want something that goes the other way, in which the bare masses are the same, and it’s
the coupling that becomes asymmetric. That requires much more straining, and is not an easy result like
Symanzik’s theorem. It requires setting up a theory much more complicated than electromagnetism, called a
spontaneously broken gauge field theory.

The Symanzik rule is useful in other cases. There are many established models of chiral symmetry, notably
the sigma model of pion–nucleon interactions,14 in which the symmetry is broken by a term linear in one of the
scalar fields only. That is of course consistent with the Symanzik rule; that term has dimension one, and as we’ve
seen the only possible term of dimension 1, δ = −3 is a term linear in a scalar field.

This concludes my discussion for the moment. Of course we will have to return to the topic when we discuss
electrodynamics.

1[Eds]. Unfortunately the videotape of Lecture 25 starts about 70 minutes into the lecture. The first part of this
chapter is interpolated from the Hill–Ting–Chen and Woit notes, and Coleman’s own notes. See also Coleman’s
1971 Erice lecture “Renormalization and symmetry: a review for non-specialists”, Chapter 4 in Coleman Aspects.
In 1976, copies of this chapter were handed out in class.
2 [Eds.] Richard P. Feynman, “Relativistic Cut-Off for Quantum Electrodynamics”, Phys. Rev. 74 (1948)
1430–1438; Dominique Rivier and Ernst C. G. Stueckelberg, “A Convergent Expression for the Magnetic Moment
of the Neutron”, Phys. Rev. 74 (1948) 218; Erratum, 986.
3 [Eds.] Wolfgang Pauli and Felix Villars, “On Invariant Regularization in Relativistic Quantum Theory”, Rev. Mod.
Phys. 21 (1949) 434–444.
4 [Eds.] J. F. Ashmore, “A Method of Gauge-Invariant Regularization”, Lett. Nuovo Cim. 4 (1972) 289–90; C. G.
Bollini & J. J. Giambiagi, “Dimensional Renormalization: The Number of Dimensions as a Regularizing
Parameter”, Nuovo Cim. 12B (1972) 20–26; G. M. Cicuta & E. Montaldi, “Analytic Renormalization Via Continuous
Space Dimension”, Lett. Nuovo Cim. 4 (1972) 329–32; Gerard ’t Hooft & Martinus Veltman, “Regularization and
Renormalization of Gauge Fields”, Nuc. Phys. B44 (1972) 189–213. The name ’t Hooft is pronounced
(approximately) as “ɘt HOAFT”, to rhyme with “(u)t loaf(ed)”.
5 [Eds.] The singularities of Γ(z) occur at z = 0 and negative integers. Writing z = −s + ϵ with s an integer and ϵ
small,

See equation (3.17), p. 152 in Pierre Ramond, Field Theory: A Modern Primer, Benjamin, 1981. Derivations are
given in Appendix 8D of Hagen Kleinert and Verena Schulte-Frohlinde, Critical Properties of ϕ 4 Theories, World
Scientific, 2001, pp. 126–129; the formula is equation (8D.24); and in Ryder QFT, Appendix 9B, pp. 385–387. For
our purposes, s = 0. The digamma function ψ(s) is the derivative of the logarithm of Γ(s):

where γ is the Euler–Mascheroni constant), equal to 0.57721… and H s = 1 + + + is the harmonic series of
order s; H 0 = 0, and so ψ(1) = −γ. See Julian Havil, Gamma, Princeton U. Press, 2003, p. 58.
6 [Eds.] To lowest non-trivial order, (4) = , proportional to λ in four dimensions, and in d dimensions
proportional to λν4−d. That sets the dimensions of all the terms in (4).
7 [Eds.] This is an application of Weinberg’s theorem; the graph contains the divergent (25.35) as a subgraph, and
so it too is divergent. See Bjorken & Drell Fields, p. 324.
8 [Eds.] Nikolai N. Bogoliubov and Ostap S. Parasiuk, “Über die Multiplikation der Kauselfunktionen in der
Quantentheorie der Felder” (On the multiplication of causal functions in the quantum theory of fields), Acta Math.
97 (1957) 227–266.
9 [Eds.] Klaus Hepp, “Proof of the Bogoliubov–Parasiuk Theorem on Renormalization”, Comm. Math. Phys. 2
(1966) 301–326.
10 [Eds.] Wolfhart Zimmerman, “Local Operator Products and Renormalization in Quantum Field Theory”, pp.
399–589 in Lectures on Elementary Particles and Quantum Field Theory (1970 Brandeis University Summer
Institute in Theoretical Physics), v. 1, eds. Stanley Deser, Marc Grisaru, and Hugh Pendleton, MIT Press, 1970;
“Convergence of Bogoliubov’s method of renormalization in momentum space”, Comm. Math. Phys. 15 (1969)
208–234.
11 [Eds.] The videotape of Lecture 25 begins here.
12 [Eds.]The video of Lecture 25 ends here. The remainder of this chapter comes from the first 36 minutes of the
video of Lecture 26.
13 [Eds.] Kurt Symanzik, “Renormalization of models with broken symmetry”, pp. 263–278, in Fundamental
Interactions at High Energies (Coral Gables Conference on High Energy Physics II), eds. A. Perlmutter, G. J.
Iverson & R. M. Williams, Gordon and Breach, 1970.
14 [Eds.]
Benjamin W. Lee, Chiral Dynamics, Gordon and Breach, 1972. See also B. W. Lee, “Renormalization of
the σ-Model”, Nuc. Phys. B9 (1969) 649–672.

Problems 14

14.1 Let ψ A, ψ B, ψ C and ψ D be four Dirac spinor fields. These fields interact with each other (and possibly with
unspecified scalar and pseudoscalar fields) in some way that is invariant under P, C, and T, where these
operations are defined in the “standard way” as discussed in Chapter 22:

Likewise,

in a Majorana basis (one in which γµ = −γµ*). Finally,

again in a Majorana basis. Now let us consider adding a term to the Hamiltonian density,

where the gi’s are (possibly complex) numbers.

(a) In class, we proved the CPT theorem for S-matrix elements. It would be really weird if the S-matrix were
CPT invariant but the Hamiltonian density were not. Show that ′(0) is CPT -invariant regardless of what the g’s
are.

(b) Under what conditions on the g’s is ′(0) invariant under C? Under P? Under T? PC ? CT ? TP ?
Reminder: ΩPT is anti-unitary.

(1998b 1.1)

14.2 In class I computed, in four dimensions, the superficial degree of divergence, D, for a general Feynman
graph with FE external Fermi line and BE external Bose lines, in a theory where the Lagrangian was the sum of
monomials in scalar fields, Dirac fields and their derivatives,

The result (25.47) was

 where ni is the number of vertices of ith type and δi, the index of divergence, is

dim i is the dimension of i in units of mass, not counting any dimensions attached to the coupling
constants.

Derive the corresponding formulae in d dimensions for arbitrary positive integer d. For arbitrary d, what is the
largest value of n for which ϕ n is of renormalizable type? For what values of d is (ψψ)2 of renormalizable type?

Comments: In any number of dimensions the action must be dimensionless, or the Lagrangian must
have dimension d (in mass units). Thus the mass dimension of both scalar and Dirac fields depend on d. Also,
Dirac fields in d dimensions are just like Dirac fields in 4 dimensions, except for the number of components, which
is irrelevant to our interest here.

(1998b 1.2)

14.3 In §21.4 we spent some time computing things for the theory of a Dirac bispinor Yukawa-coupled to a
neutral pseudoscalar meson, described by the interaction Lagrangian

(This interaction was also the subject of Problems 12.3, 13.1, and 13.2.) To order g2 the Feynman amplitude for
the process ϕ + ψ → ϕ + ψ is given by the sum of two graphs:

In equations

where g2M1 and g2M2 are the contributions of the first and second graphs, respectively. (These are functions of
momentum and spins, but we won’t need their explicit forms for this problem.)

Now let us consider the isospin-invariant theory of pions and nucleons discussed in class (24.29),

Compute to order g2, in terms of g, M1 and M2, the amplitudes for the following processes:

1. p + π+ → p + π+

2. n + π+ → n + π+

3. n + π+ → p + π0

Also compute a1/2 and a3/2, the scattering amplitudes for the pure I = and I = initial (and therefore final)
states.

(1998b 5.3)

14.4 In this problem, you are to compare two theories of the interactions of mesons and nucleons. In both
theories the free Lagrangian is the same:

The first theory was discussed in class, with a pseudoscalar Yukawa coupling:

The second theory is defined by “gradient-coupling” and a quadratic coupling to the meson,

Here a and b are real dimensionless constants; they are assumed to be independent of g, but may depend on the
dimensionless ratio µ/m. Show that to lowest nontrivial order in perturbation theory—order g2—the two theories
predict the same scattering amplitudes for both meson–nucleon scattering and nucleon–nucleon scattering, if a
and b are properly chosen. Find the proper choices. (Note that since we are free to redefine the sign of the meson
field in the two theories independently, we can always by convention take both g and a to be positive.)

Remark. I have not yet derived the Feynman rules for derivative couplings in class, and I do not expect you to
derive them from first principles for this problem. (But see Problem 8.1, comment (3), p. 309, and §14.4, (14.57).)
Take the following on trust: An interaction of the form

generates a vertex of the form

with which there is associated a factor

where all momenta are directed inward.

(1980 253a Final, Problem 3; 2000 253a Final, Problem 1)

Solutions 14

14.1 (a) The x dependence is not at issue here, so we suppress the arguments of the fields. To begin with, let’s
study the transformation properties of the individual bilinear forms. From Chapter 20 (box, p. 420)

and from Chapter 22, (box, p. 469 and (22.63))

Under PT (22.87),

so (in the Majorana basis, where γµ and γ5 are imaginary)

and

From these results we can determine the effect of CPT. Define

Then
The last equality follows because

By construction, H′† = H′. On the other hand, under CPT all the bilinears are turned into −1 times their adjoints.
The Hamiltonian is built of pairs of bilinears, so the signs cancel, and

Thus the Hamiltonian is invariant under CPT, without conditions on the gi. (If any of the gi’s are complex, the
operator ΩCPT has to be anti-unitary, to turn gi into gi*.)

(b) Because of CPT invariance, the Hamiltonian is invariant under C if and only if it is invariant under PT, and
similarly for the others. So there are really only three other cases to check:

Under P, or equivalently, under CT, the first term of the Hamiltonian transforms as follows:

So the first term is unchanged. The last term is likewise unchanged. However, the second and third terms pick up
an overall minus sign. Thus

The only way this can equal the original Hamiltonian is if g2 = g3 = 0.

Under C, or equivalently, under PT, the first term of the Hamiltonian transforms as follows:

The full Hamiltonian includes, as part of its Hermitian conjugate, the term g1* †. The only way
this can equal the transform of the first term is if g1* = g1, i.e., g1 is real. The same argument holds for g4. The
second and third terms pick up an extra minus sign under C. Consequently the Hamiltonian will be invariant under
C, or PT, if g1 and g4 are real, and if g2 and g3 are imaginary.

Finally, let’s consider T, or equivalently PC. We need to work out what happens to the bilinears under PC:

We see that both the axial vector and the vector terms transform as vectors (not axial vectors) under PC—except
for the switch in ordering, which (as noted in (S14.7)) amounts to Hermitian conjugation. Therefore the Hamiltonian
will be invariant under T, or under PC, if all the gi’s are real, because the Hamiltonian is built up of products of
vectors and axial vectors.

To summarize,
(Most of the homework solutions in this book were generated by graduate students. Problems assigned as
homework one year often became exam problems another year, and vice versa. In addition to being used as the
first homework problem in Physics 253b in 1998, 14.1 appeared in the Physics 253a final in 1981. This solution is
Coleman’s, with a few extra steps.)

14.2. In d dimensions, the superficial degree of divergence of a Feynman diagram is

where

L is the number of loops, each putting ddp into the integrand of the Feynman diagram

BI is the number of internal scalar lines, each bringing 1/p2 at high p

FI is the number of internal fermion lines, each bringing 1/p at high p

ni is the number of interaction vertices of type i

di is the number of derivatives at vertices of type i, each bringing a factor of p

The only difference between this formula and the four-dimensional formula is that the coefficient of L is now d. The
number of loops is

The ni is for the d-momentum conserving δ functions at each vertex, and the 1 is for the overall d-momentum
conserving δ function. Inserting (S14.17) into (S14.16) we get

As discussed in lecture, we can count the number of scalar and fermion line-ends in two different ways and find
the constraints

where BE is the number of external scalar lines, FE is the number of external fermion lines, bi is the number of
scalar fields in interaction i, and fi is the number of fermion fields in interaction i. Combining (S14.18) and (S14.19),
we find

where the index of divergence δi is

From the equal-time commutators or anticommutators, it follows that in d dimensions, a scalar field has mass
dimensions (d – 2), and a spinor field has mass dimensions (d – 1). Therefore δi + d equals the mass
dimensions of the interaction i (not including the dimension of the coupling parameter). The interaction is of
renormalizable type when the index of divergence δi ≤ 0, i.e., when the dimension of the interaction is less than or
equal to d.

The index of divergence δi is more fundamental for this analysis than the superficial degree of divergence D
because it focuses on individual interaction vertices and is not concerned with the number of loops, integration
momenta, etc. We just look at what’s going on at a vertex, and that tells us if the interaction is renormalizable or
not.

The interaction ϕ n has dimensions n(d − 2) and is of renormalizable type if

For d ≤ 2, ϕ n is of renormalizable type for all n. For d ≥ 3, we must have

Thus there are no nontrivial interactions of this kind for d ≥ 6. As a check, for d = 4, we get n ≤ 4, which we know is
true from §25.3.

Finally, the interaction (ψψ)2 has dimensions 2(d − 1) and is of renormalizable type when

14.3 Let the isovector of pion fields be denoted

(see (24.21)). Writing N = , and with τ the Pauli matrices, the interaction Lagrangian (P14.4) becomes (see
24.20))

For reaction 1, p + π+ → p + π+, the first graph cannot contribute, because there is no intermediate state with a
charge of +2, but the second can:

That is, the amplitude a(pπ+ → pπ+) is given by

Similarly, the second graph cannot contribute to reaction 2, n + π+ → n + π+, and so

On the other hand, both graphs contribute to reaction 3, n + π+ → p + π0:

The amplitude now comes from two cross-terms,

Using the Clebsch–Gordan tables in the Particle Data Group’s Review of Particle Properties,1 we have
(writing |I, Iz ñ for the isospin eigenstates)

So

That answers the original question. Note that we can solve for the amplitudes:

In §24.3 we compared pπ+ → pπ+ with pπ– → pπ–. In this second process, only the first graph contributes:
 That is,

From (24.44) we have

Then, as we found earlier,

just as in (S14.27). Thus the results (S14.26) and (S14.27) are consistent with the arguments and results in §24.3.

14.4 (a) N + N → N + N

Theory I:

Theory II:

But the fermions are on the mass shell, so

and likewise u2′( 2′ − 1)γ5u2 = 2mu2′γ5u2. Thus we obtain equality if

(b) N + ϕ → N + ϕ

Theory I—as in the lecture (§21.4):

Theory II:

where both q and −q′ are inwards; the last term has a factor of 2 for symmetry (there’s a choice which ϕ emits,
and which ϕ absorbs, mesons). Once again moving the γ5’s, and substituting a = µ/2m,

Now u′( ′− m) = 0, and so we can substitute u′( ′− m + ′) for u′ ′, and similarly u = ( + − m)u. Then
because the overall delta function ensures that p + q = p′ + q′. Expanding the cubic,

because 2 = A2. Then, because u = mu, we have finally

Similarly (first changing the signs of both and ′, which does not affect the product)

Substituting (S14.39) and (S14.41) into (S14.37), we have

Comparing the Theory I amplitude (S14.35) with the Theory II amplitude (S14.42), we see that the terms in square
brackets are identical. If the two amplitudes are to agree, the terms in the curly brackets must vanish:

In short, the interaction Lagrangian

(with m the nucleon mass) is completely equivalent (for first-order processes) to the (pseudoscalar Yukawa)
interaction Lagrangian LI = igψγ5ψϕ.

Remark. Dyson’s solution2 is much less work (and characteristically elegant). Start with LI, and change the
nucleon field:

Then ψψ → ψ exp{2iαγ5ϕ}ψ, and ψγµψ is unchanged. Split L = L0 + LII into kinetic and potential terms, to find

Multiplying everything out and gathering terms gives

If we choose α = (g/2m), the linear term in ϕ goes away. A second change of variables ϕ → −ϕ gives

exactly as before.

1 [Eds.] PDG 2016, http://pdg.lbl.gov/2016/reviews/rpp2016-rev-clebsch-gordan-coefs.pdf

2 [Eds.] F. J. Dyson, “The Interactions of Nucleons with Meson Fields”, Phys. Rev. 73 (1948) 929–930.
 26
Vector fields

After scalar and spinor fields, the next case is the vector field. As the spin gets higher and higher, and the degrees
of freedom increase, we have more and more indices to keep track of, but apart from that it will be pretty much a
rerun of what we did for the scalar and spinor fields.

26.1The free real vector field

I’ll call the vector field Aµ(x) in honor of the most famous example: electrodynamics. I’ll do the real case, Aµ = Aµ*,
because, as in the scalar case, the extension to complex fields is trivial. We begin as always by writing down the
possible terms in the Lagrangian that are Lorentz scalars. Fortunately we can short circuit a lot of the stuff we did
for spinors—I presume you know the Lorentz transformation properties of a vector.1

The first step is to write down the most general Lagrangian, quadratic in Aµ, with no more than two
derivatives, to define the classical field theory. I will then restrict the parameters by requiring that the energy be
positive, and quantize canonically. Here are the possible terms.

1.No derivatives: There is only one Lorentz invariant form:

2.One derivative: I can build nothing, because with three vector indices, one from the derivative and
two from the field, there’s no possible way to make a scalar. We’ll always have an uncontracted
index somewhere.

3.Two derivatives: Things get more complicated. At first glance (and certainly with integration by parts)
I can always arrange matters so that one field is differentiated once and the other field is
differentiated once. There are apparently three possibilities:

As far as the Lagrangian goes, the third is the same as the second: using integration by parts, I can turn them into
each other by switching the derivatives around (ignoring surface terms):

So in fact there are only three possible terms: (26.1), (26.2a) and (26.2b). That’s slightly more complicated than
the scalar field, where we had only two possible terms, but not much. I can rescale the fields to turn the coefficient
of one of these terms into whatever I please, up to a sign. So I will write the most general form of the Lagrangian
as

The factor will turn out later to be a convenient choice for the first term; there’s some unknown real coefficient a
in the second term, and some other real coefficient b in the third term, and that defines our Lagrangian. Higher
order terms could be added, but then the Lagrangian would not describe a free field: free field Lagrangians mean
linear equations of motion.

The next step is to vary the Lagrangian and derive the equations of motion:

This is a messy equation and it’s rather hard to see what particles of what mass are being described here. Let’s
just blithely go ahead and look for plane wave solutions:
The four-vector εν is called the polarization vector. Plugging this into the equations of motion gives

We could write this as a 4 × 4 matrix in k acting on the vectors εµ, and find the eigenvectors in the usual way.
Instead, we’ll just read them off. This equation has two kinds of solutions: longitudinal solutions, where εµ is
aligned along kµ; and transverse solutions, with εµ perpendicular to kµ.

(a) Longitudinal: εν ∝ kν.

where µL is the mass of the longitudinal mode;

(b) Transverse: ε ⋅ k = 0.

where µT is the mass of the transverse mode.

So this theory is capable of describing two types of oscillations—longitudinal, with εµ parallel to kµ, and
transverse, with εµ perpendicular to kµ. It’s rather like the three dimensional theory of an elastic solid.2 The
longitudinal oscillations have one mass, and the transverse have another. Under a Lorentz transformation a
longitudinal oscillation remains longitudinal and a transverse oscillation remains transverse. One would expect
upon quantization that the longitudinal oscillations would correspond to scalar particles (one degree of freedom),
and the transverse oscillations would correspond to spin one, (vector particles with three degrees of freedom).
There are three independent vectors perpendicular (in a four dimensional sense) to a given four-vector like kµ.3
This should really be no surprise. We already know we could describe an ordinary scalar meson in terms of a
vector field; to wit, its gradient ∂µϕ. We can do it, but we don’t particularly want to: we’d like to get a theory that
when we quantize it describes vector particles only, without longitudinal oscillations.

Is it possible to arrange the parameters in our Lagrangian to suppress these longitudinal oscillations? Well,
the answer is obvious: if we choose a = −1, so long as b ≠ 0, our free wave equation has no longitudinal solutions;
there are only transverse modes, with a mass b ≡ −µ2. The transverse solutions are the only things around; there
is only one mass in the theory. Notice that this trick for suppressing the longitudinal solutions does not work when
−b = µ2 = 0. In that case, if I set a = −1, then in fact I can have longitudinal oscillations of any mass. Instead of
getting no solutions I have simply no restricting equation, because (26.7) becomes

which is unquestionably true, but it doesn’t limit the longitudinal motions very much! (It’s also possible to construct
a theory with only longitudinal waves and no transverse waves. In this case there are both scalar and vector
mesons with independent masses. This is not the easy way to do that. Why should you join together what God
hath put asunder?)4

With these choices, a = −1 and b = −µ2 ≠ 0, the Lagrangian becomes

Remember: the middle term is equivalent to −(∂µAν)(∂νAµ), plus surface terms. This Lagrangian is so simple I can’t
resist introducing notation to make it look obscure. Define the field strength tensor

This is the convention in modern high energy physics literature.5,6 Then

The µ2 → 0 limit of (26.12) describes free electromagnetism, written in relativistic form.7 If I interpret A0 as the
scalar potential and Ai as the vector potential, (26.11) tells me that Fij is the magnetic field, B, the curl of the vector
potential. Likewise F0i is the time derivative of the vector potential plus the gradient of the scalar potential, (−1)
times the familiar formula for the electric field, E. As µ2 goes to zero, the Lagrangian (26.12) becomes E2 − B 2
times a factor, the familiar Lagrangian for free electromagnetic theory. We now see why we have no restraint on
the longitudinal oscillations when µ2 = 0, because what we’ve been calling a longitudinal oscillation is equivalent in
conventional electromagnetic theory to a gauge transformation. It is well known that you can add the four-
gradient of a function λ(x) to the four-potential Aµ, with no change to the physics:

In the limit µ2 → 0, the gauge invariance of electromagnetism pops into the theory described by (26.12), leading to
some funny problems in quantizing the theory, as we will see. That doesn’t happen in this theory for any other
value of µ2.

26.2The Proca equation and its solutions

Just as an exercise, let’s rederive the equations of motion (26.4) for the choices a = −1, b = −µ2

from the Lagrangian (26.12). The field Aµ enters the Lagrangian four times: twice in each field tensor Fµν, and thus
four times in its square. A variation of the Langrangian with respect to Aµ will produce a factor of 4 to cancel out the
overall factor of multiplying FµνFµν. Written in terms of Fµν, the Euler–Lagrange equations are

This equation of motion, equivalent to (26.14), is called the Proca equation.8 As the Klein–Gordon equation is to
spin 0 and the Dirac equation is to spin , so the Proca equation is to spin 1. Note: If µ = 0, you get

which are two of the (empty space) Maxwell equations.9

Taking the divergence of the Proca equation, we get

The first term is zero because of the antisymmetry of Fµν. Assuming µ2 ≠ 0, we see that the divergence of Aν
vanishes:

This equation is called the Lorenz condition. It ensures the suppression of the longitudinal waves. We’ve done
this computation before in momentum space, (26.8), kµAµ = 0. Now we see it again in position space.10

Returning to the Proca equation in terms of the A’s,

The first term is equal to ∂ν∂µAµ which is zero. What remains is the Klein–Gordon equation:

The waves are transverse, with mass µ2. There are four solutions, but only three are linearly independent because
of the constraint (26.8). We now know in position space what we found earlier in momentum space.

As we did for the solutions of the Dirac equation (20.112a) and (20.112b), I’d like to establish some
normalization conventions for the three independent solutions (labeled by r) to the Klein–Gordon equation,

with the three polarization vectors εµ(r) orthogonal to kµ, (26.18):

In the rest frame,

and we choose the solutions to be orthonormal to each other, in this frame:

By Lorentz invariance, (26.24) is true in any inertial frame. These are three spacelike vectors, for instance, the
three unit vectors in (x, y, z) perpendicular to timelike k. We get the (−) sign in (26.24) from the metric. We have

and we take k0 > 0. In the rest frame, for example, we can choose the usual orthonormal space basis (for linear
polarization):

or, for spin along the z axis (circular polarization):

Here, the vectors ε(1) and ε(2) pick up the phase e±iθ if rotated through θ about the z axis; they are eigenstates of
Jz with m = ±1.

For the Dirac equation, normalization conditions (20.112a) and (20.112b) led to the completeness relations:

We also have a completeness relation for the vector field, easily derived. The analogous expression for the vector
field is

What does this sum equal? By Lorentz invariance, it must be that

(where A and B are constants) because there are no other quantities that transform as rank 2 Lorentz tensors. We
know from (26.22) that

If we multiply Pµν by εν(s) we obtain from (26.24)

Consequently we have a completeness relation for the massive vector theory:

To check this expression, consider the rest frame. If µ and ν are both space indices, i and j respectively, we have

which can be confirmed by inspection from either (26.26), where εi(r) = δir, or (26.27). If either µ or ν is 0, the
definition (26.28) gives Pµν = 0, since the time components of all the ε vectors from (26.26) or (26.27) are zero.
And that’s just what we get:

The projection operator Pµν serves the same purpose as the projection operators (20.123a) and (20.123b), and
will be just as useful when we need to sum over spins of vector mesons.

26.3Canonical quantization of the Proca field

So much for the classical Proca equation and its plane wave solutions. We now turn to canonical quantization.
There will be some complications quite apart from juggling all those indices. As usual it’s convenient to break
things up into space and time components: the Latin indices like i and j will run over the space indices only. This
split will separate the p’s and q’s; the q’s are the space components Ai of the vector field Aµ.

The first turns into a because F0i = Fi0. The only part with a time derivative is F0i = ∂0Ai − ∂iA0; Fij involves only
space derivatives. The p’s, the canonical momentum densities conjugate to Ai, are

but—surprise!?—the momentum conjugate to A0 is zero:

Is this a disaster or not? Well, that depends on whether or not the three quantities Ai and the three conjugate
momenta F0i are a complete and independent set of initial value data (IVD). I will demonstrate that in fact the
entire set of initial value data is given in terms of the set {Ai, F0i} at a fixed time: we won’t need π0. That there is no
momentum conjugate to A0 is totally irrelevant. Even if there were a non-zero π0, we’d have to throw it away; we
already have a complete set.

To prove this, note that each component of Aµ obeys the Klein–Gordon equation (26.20). The field ϕ(x, t0)
and its time derivative ∂0ϕ(x, t0) at a fixed time t0 provide a complete set of IVD for a field ϕ(x, t) satisfying the
Klein–Gordon equation, and we have four decoupled Klein–Gordon equations here. Our task is to show that at
any given time, the quantities {Ai, ∂0Ai, A0, ∂0A0} can be given in terms of {Ai} and {F0i}. If we can do that, we will
have demonstrated that the six fields {Ai, F0i} suffice; we don’t need eight fields, as one might have thought for the
four decoupled Klein–Gordon equations. In particular, we don’t need π0.

We have {Ai} and {F0i}. That certainly includes Ai. From the constraint equation (26.18) we obtain ∂0A0:

If I know Ai at a fixed time, I can certainly find its divergence at that time, and therefore I know ∂0A0. Next, let’s look
at the full form of the equations of motion:

From the ν = 0 component

which determines A0 in terms of space derivatives of the known quantities F0i. Finally, from F0i and A0, we obtain
∂0Ai:

If I already know A0, I can find its space derivatives, and I had F0i from the start, so I can determine ∂0Ai. That
completes the argument.
 In short, once we have Ai(x, t0) and F0i(x, t0) at some initial time t0, that’s all we need to compute Aµ(x, t). The
eight quantities {A0, Ai, ∂0A0, ∂0Ai} can be found from the six quantities {Ai, F0i}, and the equations of motion. This
is really just transversality, expressed by the funny condition, a = −1. Had we not imposed transversality, we would
have had four degrees of freedom. With that condition, we have only three independent Klein–Gordon equations,
not four. We should need only six functions, two for each Klein–Gordon equation, not eight. And that’s just what
we found. This isn’t a surprising result. Consider other field theories:

In every case,

Now let’s compute the Hamiltonian density. As always,

This is an awkward expression for our purposes, because we want to write everything in terms of p’s and q’s, i.e.,
in terms of F0i and Ai. Using the identity

the Hamiltonian density can be written

All we’re really interested in is the Hamiltonian, the space integral of the Hamiltonian density H. We can integrate
(26.43) by parts, to rewrite the second term on the right:

the last equality following from (26.39). Then

Of course if I really wanted to express the Hamiltonian as a function of p’s (Fi0) and q’s (Ai), I should write for A0
the expression (1/µ2)∂iF0i.

Now we come to the question of positivity of the energy. Each of these terms is individually negative. The first
and fourth terms inside the brackets are each negative sums of squares, because F0i = −F0i and Ai = −Ai. The
second term inside the brackets is positive, because A0 = A0. The third term inside the brackets is a positive sum
of squares, because Fij = Fij. Consequently the quantity in the square brackets is a sum of four negative terms.
The Hamiltonian must be bounded from below, so the overall sign must be (−). With our sign ambiguity resolved,
we can write for the Hamiltonian

and for the Lagrangian,

This is pretty much like what we got for the scalar field, (4.44), except that the sign of the mass term looks wrong.
Remember that the true dynamical variables, the q’s and the p’s, are {Ai} and {F0i}, respectively. Rewriting the
Lagrangian we have

So it’s just like the conventional expression in Lagrangian mechanics,

The term with time derivatives in T has a positive coefficient, as does the corresponding term in (26.48), + (F0i)2.
Likewise one of the terms in V, the mass term − µ2(Ai)2, has a negative coefficient just as it should, introduced by
the metric rather than by an explicit minus sign in front. The canonical momentum is given by (26.36), with the
minus sign:

Let’s canonically quantize this theory (4.47):

To shorten a lengthy calculation, I will write the field at any spacetime point in terms of a sequence of Fourier
coefficients, with the usual measure:

This is much like what we did for the Dirac field (21.6), except that this field Aµ is Hermitian. In place of the Dirac
spinors we have the analogous polarization vectors εµ(r) multiplying the operators ak(r)† and ak(r). It’s a long
calculation to determine the commutation relations of the operators ak(r) and ak(s)† which ensure the canonical
commutation relations for Ai and F0j. Rather than plug and chug, let’s guess the commutation relations are the
same as (2.47), for the scalar case:

and all others zero. In other words, the operators ak(s)† and ak(r) are creation and annihilation operators. From
these we will check the commutation relations, (26.51).

To confirm that these commutators (26.55) indeed give the canonical commutation relations (26.51) is a
horrendous computation, because of all the indices. And therefore I’ll be a little sneaky, and recall that when we
did a free scalar field, we found the commutators for arbitrary times. Check that these commutation relations work
by recalling the scalar theory:

The commutation relation [ϕ(x), ϕ(y)] for arbitrary time is

where iΔ(x − y) can be written, from (3.38) and (3.42), as

Let ∂µx = ∂/∂xµ. Then Δ(x − y) satisfies the following conditions for x0 = y0 (see (3.60), (3.61) and (3.66),
respectively):

(These correspond to [qi, qj] = 0, [pi, qj] = −iδij, and [pi, pj] = 0.) The computation we have to do here is basically the
same, except that after we’ve commuted everything, we also have a polarization sum on r. For each value of r, it’s
the same computation we did for the scalar field. By analogy

using (26.32). So we’ll get exactly the same Fourier transform as for the scalar field. The only thing is that in
momentum space, the completeness relation will be stuck inside the integral. But it’s easy to see what the Fourier
transform produces, since multiplication by kµ is Fourier-equivalent to differentiation: kµ → −i∂µ:

Without doing any computation at all, but just by being sneaky, we obtain the expression above. That saves a lot
of labor. There’s no reason to do a calculation twice when you’ve already done it once.

But does it give the correct commutation relations? Start with the easiest:

because both Δ(x − y) and its gradient ∂ix Δ(x − y) vanish at x0 = y0. So that one checks: [qi, qj] = 0. (Note that A0
does not commute with Ai at equal times. Because of (26.39), A0 is some awful divergence of the canonical
momentum.) Next, we have

Then for x0 = y0,

as required for the equal-time canonical commutation relations for Fj0 and Ai; the [pi, qj] commutator checks.
Finally, for the last set (with x0 = y0)

The first term is zero because two time derivatives become (∂0x ∂0y ) acting on Δ(x − y), which equals zero at equal
times. The second term vanishes because Δ(x − y) = 0 for x0 = y0, and so a fortiori does its gradient. That makes
[pi, pj] = 0. We have verified the equal time canonical commutation relations. As an exercise, you should be able to
show that, analogous to (2.48)

(plus a divergent constant, usually dropped, disposed of by normal ordering).11

If however we try to canonically quantize the vector meson in the limit as µ2 → 0, we come to a screeching
halt. In the limit as µ2 → 0, the Proca equation reduces to the Maxwell equations. We take A0 to be the scalar
potential, Ai to be the vector potential, and

just the usual definition of the electric field. Then ∂iFi0 = ∇ • E = 0, according to Maxwell’s equations (in empty
space), so the commutator of ∂iFi0 with anything should be zero. But
What should be zero is not zero. We will solve this problem soon enough, but you should be aware of it. The
problem does not arise if µ2 ≠ 0, because the Proca equation component equivalent to Gauss’s Law in empty
space is

The commutator (26.66) that gave us trouble with µ2 = 0 is no longer a problem. Using (26.59),

Unlike (26.66), this result is perfectly consistent with the canonical commutation relations.

26.4The limit µ → 0: a simple physical consequence

I want to discuss a topic more interesting than this dumb commutation computation (which, once you’ve done, you
never have to think about again). Let’s consider something with a little more physics in it: how an actual physical
process is affected as the mass of this real vector meson goes to zero. I can’t talk about a very complicated
theory, because we have yet to compute the vector meson’s propagator. However there is one theory I can
discuss, the analog of our earlier scalar theory, Model 1.12 That interaction (8.57) was between the field ϕ and a c-
number source, ρ(x):

Here we’ll have a c-number source Jµ, some arbitrary vector function of x vanishing at infinity, coupled to Aµ. That
is, we write the Lagrangian as

I can discuss the process where that c-number source emits one meson. This is the analog of the Feynman graph,
Figure 8.14, that we talked about in Model 1. The wavy line in Figure 26.1 indicates the vector meson. Let me write
down the equation of motion,

Figure 26.1: Emission of one vector meson by Jµ

Does this have a sensible zero mass limit? Let’s take the divergence of this equation, and obtain

Now the first term is always zero, by the antisymmetry of Fµν. If this is to make any sense at all when I set µ2 = 0, I
had better have ∂µJµ = 0, otherwise everything will go bananas. So let’s consider not a general external current
but a conserved current, for which ∂µJµ = 0. That’s my one condition.

What is the amplitude for the emission of a meson of helicity type r (p. 400) and momentum k? We sort of
know what happens from our old scalar model. There, the amplitude for one meson emission was proportional to
(k), (8.67). Here, the amplitude Afi must look something like this:

Within kinematic factors, it has to look like this. To lowest nontrivial order, the amplitude must be linear in µ(k),

and it has to be a Lorentz scalar. There are only two other four-vectors available with which to build an invariant
amplitude, εµ(r)* and kµ. But by our conservation condition we must also have

so the expression (26.73) is the only Lorentz scalar available. Let’s consider a meson of specified momentum, and
a nice, smooth function µ(k). Take that momentum to point in the z direction with some magnitude |k|, and of
course it has energy .

If the amplitude for this process goes as the corresponding scalar field’s (8.67), it has a factor 1/ . Let’s
assume that is so. For small values of µ,

In the rest frame of the particle, this amplitude does not have a smooth limit as µ → 0. Instead let’s consider the
limit (µ/|k|) → 0, for each of three independent kinds of mesons that can be emitted. Remember that the εµ’s are
restricted to be orthonormal spacelike vectors orthogonal to kµ. Two are

Linear combinations of these, (1/ )(ε(1)µ ± iε(2)µ), have helicity ±1. Then there’s one unchanged by rotations
about the z axis, the third vector ε(3)µ with helicity zero, which must be orthogonal to kµ and to the other two εµ’s,
so it cannot have x or y components. It must look like this:

(To satisfy (26.24) I have to divide by µ.) We’re now ready to go. I’m going to show you that something very
interesting happens to the amplitude for emitting a meson of this kind in the limit as µ → 0.

The amplitude for emitting a meson of type 3 goes as

From the conservation rule (26.74), with k parallel to , we have

Putting 3 in terms of 0, we notice an amazing cancellation:

No matter how small µ is, this is a system with three degrees of freedom. Everyone says the photon is massless.
But suppose the photon had a mass of 10−23 of the electron’s. This would be a very hard thing to determine
experimentally. Some people say, “No, absolutely not! It would be trivial to detect experimentally because we
know the real massless photon has only two degrees of freedom; polarized light and so on. If we took a hot oven
and let things come to thermal equilibrium, because the walls are emitting and absorbing photons, we wouldn’t get
the Planck Law, but instead times the Planck Law.” This is garbage. The amplitude for the oven walls to radiate
a helicity zero photon, according to this current, goes to zero in the limit as µ/|k| → 0. At every stage in the limiting
process there are indeed three degrees of freedom just as you’d expect from a theory of massive vector mesons.
But as µ/|k| → 0, the amplitude for emitting the third photon goes to zero. If the photon mass is small enough, it will
require twenty trillion years for that oven to reach thermal equilibrium!13

Thus we see something very interesting. Our whole formalism collapses completely as µ → 0. But if we go to
the end and compute the amplitude for a physically reasonable, very simple process—the emission of a single
meson by an external source—we find it goes to what you would expect if you knew anything about
electrodynamics, namely, the electrodynamic answer: A vector meson of mass 10−23 times the electron mass
looks just like a photon. Shake your source, and despite the three degrees of freedom, except for a negligible
factor, only helicity ±1 mesons are emitted; there is no amplitude to emit helicity zero.

The black body law turns out not to be the best way to determine an upper limit for the photon mass. The best
way is the Coulomb law which would be modified to a Yukawa law, or the analog for a dipole. Although there are
magnetic fields over cosmic distances, these are not good because the Compton wavelength is messed up by
interstellar plasma. The best measurement14 was made from the magnetic dipole field of the earth, by the
Explorer 12 satellite, at night, because in the daytime Explorer 12 was in the solar wind. In the night time the earth
shields Explorer 12 from the solar wind, and the plasma effects are much smaller. The satellite was 10,000 km
from the earth’s center, and measured the dipole field with something like 10 or 15 percent accuracy. Therefore a
number on the order of 107 m is the current best lower bound on the Compton wavelength of the photon. The
Compton wavelength of the electron is 10–12 m, so this provides a upper bound to the photon’s mass of about
10–19 of the electron’s mass, or about 10−47 g.

26.5Feynman rules for a real massive vector field

Let’s briefly consider interactions of the massive vector field with either an electron (equivalent to quantum
electrodynamics with a massive photon, “massive QED”) or a charged scalar field (“massive charged scalar
electrodynamics”).15 This is a good background for actual QED, which we will begin to tackle next time. We’ll add
an interaction between the real massive vector field with a Dirac field,

where Γ = 1, with Aµ a vector; or Γ = iγ5, with Aµ an axial vector; and e is a coupling constant. (Interactions of this
type occur in discussions of the Z0 meson in electroweak theory, which we’ll discuss near the end of this course.)
The first thing we have to do is to work out the propagator for the massive vector meson.

As usual, we define the Wick contraction as the time-ordered vacuum expectation value:

Consider first a scalar field of the same mass. Recall (3.38) the vacuum expectation value,

By analogy,

The time-ordered vacuum expectation value of the scalar field is (P1.4)

and for the vector fields,

We’d now be ecstatically happy (well, let’s just say mildly cheerful) if the time-ordered product were equal to this
obvious guess:

The indicates that it’s not obvious we can bring the derivatives through the θ functions, which we’d need to do to
express the time-ordered product as an integral. This is all right so long as µ and ν are not both equal to zero. If µ
and ν are both space indices, there’s no problem because the θ functions depend only on time. When one of the
indices is a space index and the other a 0, then pulling the ∂µ∂ν in front of the θ in (26.88) will give, apart from the
right-hand side of (26.87), the term

Because of the delta function in front, y − x has a vanishing time component, so it is a spacelike vector. Therefore
by making a rotation by π, we can turn x − y into y − x, while the function Δ+ is itself rotation invariant. The two
parts of the extra term cancel, so there’s no problem here, either.
 When both indices equal zero, it is not all right. If we could drag the time derivatives through the θ functions,
we could write, for example,

On the left-hand side of this equation, we have the Klein–Gordon operator acting on the scalar Feynman
propagator ΔF(x − y), the Fourier transform of (10.29); that gives us −iδ(4)(x − y). On the right-hand side we have
the Klein–Gordon operator acting on Δ+(x − y), that is, on the vacuum expectation value of the product of two
fields—one at x and one at y—each of which obeys the Klein–Gordon equation. Therefore the right-hand side is
zero, and (26.90) is equivalent to the statement that −iδ(4)(x − y) = 0. Thus, in the case of ν = µ = 0, we have a
problem. It looks like the propagator is not Lorentz invariant, because its 00 component has an extra piece coming
from the time derivatives on the θ functions. Well, this is certainly a Lorentz invariant theory; is there somewhere
else where there could possibly be another extra piece to cancel this one?

You may remember (§26.3) that when we were doing the initial value problem for a free vector field, we
discovered that A0 was not an independent dynamical variable. If we break the interaction into two pieces,

then the space term is like a “q” in the language of p’s and q’s, since the Ai’s are independent dynamical variables,
each with its canonical momentum. But A0 is proportional to ∂iFi0 by the Proca equation, and Fi0 is a p-type
variable; thus a term involving A0, like γ0ΓψA0, is in fact a derivative interaction involving the p’s. When we write
the theory in Hamiltonian form, in terms of p’s and q’s, then this term must be involved in the Hamiltonian, but the
other terms in ψ are not; H I ≠ −LI. Then the Hamiltonian, the thing that appears in Dyson’s formula, is not going to
look Lorentz invariant, either. Maybe God is on our side, and these two difficulties cure each other. In fact during
the heroic period of the late Forties it was shown that this desperate prayer is indeed answered: if you treat H I
naïvely as if it were equal to −LI, and the propagator as if it were equal to the right-hand side of (26.88), the
troubles cancel.16 Later on, we’ll see how the troubles cancel when we develop more efficient methods for
handing these kinds of theories (§29.4). For the moment accept on trust that everything works if you treat H I = −LI
and the propagator as

The Feynman rules for the theory with the interaction Lagrangian (26.82) are set out in the box below.

Feynman rules for Massive QED

1.For every …

Write …

(a)

(b)

(c)

2.Ensure momentum conservation at each vertex: (2π)4δ(4)( pout − pin)

3.Multiply by and integrate over all internal momenta q.

4.Spinor factors:

For every electron, a factor

for every positron, a factor

 5.Polarization factors:

For every vector meson, a factor , with ε ⋅ k = 0, ε′⋅ k′ = 0.

In doing spin sums the completeness relation

is as useful as the Dirac sum rules (20.123a) and (20.123b).

As an example of these rules, consider the Compton effect with massive photons, as shown in Figure 26.2.
(Here, Γ = 1.)

There are two graphs, with amplitude given by

Figure 26.2: Compton scattering with massive photons

We don’t need the pole-avoiding iϵ’s in the denominators. (Beware the notational ambiguity: ′* ≡ εµ*γµ, not
(εµγµ)*.) To see that this makes sense, we should expect that the amplitude for longitudinal massive photons
would be zero: in the Compton effect, the initial and final states involve physical photons, and there are no
physical longitudinal photons.17 That is, if εµ is parallel to kµ, we need to have the amplitude equal zero. It is zero,
because of current conservation. Let’s see how that goes.

From the equations of motion we have (26.72):

First, a consistency check. Using the LSZ formalism (14.37),

With

(dividing by µ as in (26.78), for type 3 massive photons),

(Note: this assumes the states |k, pñ and |k′, p′ñ describe particles that are on the mass shell, not virtual particles.)
So that checks. What about the amplitude (26.94) itself? Substituting (26.97),

The spinor u satisfies (20.110a) ( − m)u = 0; likewise (20.113) ′( ′ − m) = 0. Adding these zero terms, we can
write
Thus we have the physically reasonable result that the Compton amplitude for longitudinal photons is zero. This
result also verifies that the current is conserved; otherwise, the LSZ computation would be inconsistent.

Next time we will talk about the interactions of vectors, massive and massless, as an introduction to quantum
electrodynamics.

1[Eds.] See §18.3.

2[Eds.] Phonon vibrations likewise have transverse and longitudinal modes. See, e.g., Charles Kittel, Introduction
to Solid State Physics, 7th ed., J. Wiley, 1996, Chapter 4.
3[Eds.]For general results on the orthogonality of four vectors, see A. O. Barut, Electrodynamics and the Classical
Theory of Fields and Particles, Macmillan, 1964, Chapter 1. Reprinted by Dover Publications, 1980.
4[Eds.]Oral tradition ascribes this quip (“What God hath put asunder, let no man join together.”) to Pauli, in
response to attempts by Weyl and Einstein (see footnote 3, p. 583) to unite electromagnetism with gravity.
5[Eds.] Warning! Bjorken and Drell define Fµν as Fµν ≡ ∂νAµ − ∂µAν, which differs by a minus sign; see Bjorken &
Drell Fields, p. 68, equation (14.1).
6The tensor Fµν is a differential form, so you can write it using the exterior derivative d as F = dA; but we won’t.
([Eds.] See Ryder QFT, Section 2.9.)
7[Eds] See Problem 2.3, p. 99 and Jackson CE, Chap. 12; Ryder QFT, Sect. 3.3; Lev D. Landau and Evgeniĭ M.
Lifshitz, The Classical Theory of Fields, 3rd rev. ed., Pergamon, 1971, §23 and §27.
8[Eds.] Alexandru Proca, “Sur la théorie ondulaitoire des électrons positifs et négatifs” (On the wave theory of
positive and negative electrons), J. Phys. Radium 7 (1936) 347–353. Proca, a Romanian-French physicist, was a
student of de Broglie. See Y. Takahashi, An Introduction to Field Quantization, Pergamon, 1969 for a discussion
of the Proca field as well as other less familiar fields, e.g., the Duffin–Kemmer–Petiau field and the
Rarita–Schwinger spin- field, which comes up in supergravity theories. See also Ryder QFT Sections 2.8 and
4.5.
9[Eds.]
The ν = 0 component corresponds to Gauss’s Law, ∇ • E = 0, and the ν = i components to the three
components of Ampère’s Law, (∇ × B)i = (∂E/∂t)i.
10[Eds.] In the videotaped lectures, Coleman frequently calls it “Fourier space” instead of “momentum space”.
11[Eds.] See §4.5 for the scalar case. The calculation of the massive vector’s Hamiltonian is the subject of
Problem 15.2, p. 591.
12[Eds.] See §8.5.
13[Eds.] See also L. de Broglie, Mécanique Ondulaire du Photon et Théorie Quantique des Champs (Wave
Mechanics of the Photon and the Quantum Theory of Fields), 2nd ed., Gauthier-Villars 1957, Chapter V, §5; and
L. Bass and E. Schrödinger, “Must the Photon Mass Be Zero?”, Proc. Roy. Soc. Lond. A 232 (1955) 1–6.
14[Eds.] As of 1975! See V. L. Patel, “Structure of the equations of cosmic electrodynamics and the photon rest
mass”, Phys. Lett. 14 (1965) 105-106. Current bounds are about 105 times more stringent; see A. S. Goldhaber
and M. M. Nieto, “Photon and graviton mass limits”, Rev. Mod. Phys. 82 (2010) 939–979.
15[Eds.] Parts of this section are based on class notes from 1999, provided by Daniel Podolsky.
16[Eds.] See also the example beginning with (27.75), p. 587, and note 7, p. 589.
17[Eds.] Coleman is using the term “longitudinal photon” to mean that its polarization 4-vector εµ is parallel to its
four-momentum: εµ ∝ kµ. Usually this term describes a photon whose polarization 3-vector is parallel to its
direction of motion; ε ∝ k.

27
Electromagnetic interactions and minimal coupling
Last time I got started on one of the main theories of this course, quantum electrodynamics, by talking about the
theory of a free massive vector meson. I will now address three topics, all dealing with the interactions of a vector
field. I won’t do more than introduce the last topic, which will be the subject of the next lecture.

First, I will talk about the classical Lagrangian theory of a vector meson field interacting with other fields. If the
vector meson is either massless or has a very small mass, we can think of it as the photon, whose interactions
with matter fields constitute electrodynamics.1 To describe the interactions of vector fields, we need to discuss
three things that turn out to be intimately related: Gauge invariance for the massless case, a conserved current for
the massive case, and the minimal coupling prescription. At the end of this discussion we’ll be in a position to write
down the interaction of photons with an arbitrary system: a free meson field, a free fermion field, or with interacting
meson and fermion fields. We won’t yet be able to write down the Feynman rules for such a theory, because we
will encounter a large number of purely technical problems. These problems make up the second topic of this
lecture. They are certainly soluble by methods we already have. But if I attempted to solve them in that way, we
would be led into a sequence of extremely narrow and complicated arguments which would involve us in a large
number of combinatorial calisthenics that I would just as soon avoid. Therefore I will stop the discussion at that
stage and move on to the third topic, to introduce a new technique of great generality, the method of functional
integrals.

27.1Gauge invariance and conserved currents

To remind you of the system we studied last time, we had a Lagrangian density of the Proca form for a massive
vector meson

where

plus possibly an interaction with an external current Jµ, of the form −JµAµ, where Jµ was a c-number function of
space and time. The equation of motion you get from varying this Lagrangian is the Proca equation with a current
source:

(this is the same as (26.71), with the indices on Fµν reversed).

To describe electromagnetism, we’ll need to talk about massless vectors. In the limit µ → 0, the Proca
equation reduces to the Maxwell equations. If we take the four-divergence of both sides, the first term vanishes
because of the antisymmetry of Fµν:

so that

Trouble ensues in the limit µ → 0. We discovered in the massive case that we could get a smooth limit of this
theory as µ → 0 only if ∂µJµ = 0, i.e., only if the vector meson is coupled to a conserved current. So we learned it
was a good thing to have a conserved current. And considering the emission of a single meson by this conserved
current acting on the vacuum, we discovered a very interesting fact: of the three helicity states of a massive vector
meson, one of them completely decoupled as the mass of the vector meson µ goes to zero; the amplitude for the
helicity = 0 state goes to zero linearly with the mass.

We’ll take the working hypothesis that we can have a sensible limit µ → 0 only when ∂νJν = 0. If this is so, the
limit µ → 0 of (27.2) is Maxwell’s equations, with the field components

Maxwell’s equations emerge in rationalized Heaviside-Lorentz units with c = 1. These are the units God uses, so
we’ll use them, too.

I would like to consider a more general kind of theory in which the Lagrangian L has the same free form
(26.47) as before, plus a contribution from the matter fields, with Lagrangian L′, which may include an interaction
of the vector meson with something else; scalar mesons, nucleons, what have you. I will represent the fields here
generically by a big column vector ϕ,

The components ϕ i will be scalars, components of spinors ψ and ψ *, the pion fields and the proton fields and
whatever else, perhaps even Aµ itself, and conceivably derivatives of Aµ. The Lagrangian is

When I vary this Lagrangian, I know what we get from the first two terms, the Proca Lagrangian LP, but I don’t
know what we get from the third, because I don’t know what L′ is. Whatever we get, though, I can write like this:

The stuff I get from varying L′ with respect to Aµ I’ll call −Jµ:

so that the Euler–Lagrange equation for Aµ becomes

This is the same equation (27.2) that we had before, but now Jµ is some complicated function of ϕ and ∂µϕ and
maybe even Aµ.

We discovered that it was a good thing in the massive vector theory that the external source current was
conserved. I would like to arrange matters so that our new current, (27.9), is also conserved:

It’s a four-vector, after all, and it is the electromagnetic current in the massless case. A priori I could couple the
vector meson to the other fields in many possible ways, but only some of them will yield current conservation as a
consequence of the equations of motion. I would like to find ways of coupling Aµ to the fields ϕ so that (27.11) is
true. So that’s one problem: find the right coupling.

We have a second and apparently unrelated problem: gauge invariance. On the classical level, we can
discuss the theory of a vector field with µ2 = 0, but in the massless case, we had trouble with canonical
quantization. Now, in electromagnetism, one of the standard dogmas is that the electric and magnetic fields are all
there is; the potentials do not make any difference. You can take the scalar and vector potentials, assembled into
the four-vector Aµ, and you can add to them the gradient of any function χ(x) of space and time, but this
transformation does not affect Fµν, which is the only real thing (see (26.11)):

Phrased in terms of an infinitesimal transformation δχ, if I make an infinitesimal variation in Aµ such as

then the free Lagrangian, − FµνFµν is unchanged:

This is the so-called gauge invariance of the theory. In Maxwell’s theory, you can choose this function δχ as you
wish. You may choose the Coulomb gauge, or the Lorenz gauge, or some other gauge; it don’t matter. No one
ever reads a paper describing an experiment involving photons with a footnote that says, “This experiment was
done in Coulomb gauge.”
 Because of this, one would like to arrange matters so that L′ has the property that all the matter fields
transform in such a way that δL′ = 0. That is, the Aµ’s and the Fµν’s transform according to (27.12), and the matter
fields ϕ transform in some other way dependent on χ:

We’re trying to discover the transformation of ϕ(x) that will preserve, in the interacting theory, the desirable
property of gauge invariance.

These two questions are ostensibly unrelated: the problem of getting a conserved current in the case of a
massive photon, and the problem of preserving gauge invariance in the case of a massless photon. Although they
look like they are unrelated, in fact they have the same solution. In particular I will show that if L′ preserves gauge
invariance, it will also generate a conserved current.

Suppose I have assigned some transformation properties (27.15) to ϕ such that L′ is gauge invariant. I then
transform the fields: δAµ = ∂µδχ, and δϕ equals something—I’ll have to figure out what that is—and compute δLP.
I don’t have to know how ϕ transforms explicitly. All I need to know is that δL ′ = 0. The only term in LP not gauge
invariant is the vector meson’s mass term, which transforms as

Hamilton’s Principle tells me that the change in the action is zero for arbitrary variations of the fields which vanish
on the boundaries of the region of integration. In particular, I can choose δχ to vanish on these boundaries.
Integrating by parts we have

If the action is to be invariant for any choice of δχ, for example a four-dimensional delta function, then we must
have

Then as a consequence of the equations of motion,

Now we return to the definition of Jµ. I know the equations of motion imply ∂µAµ = 0. I also know (27.9)

I take the divergence of both sides:

The first term on the right-hand side is zero because Fµν is antisymmetric. The second term on the right-hand side
is also zero, because I just proved it for µ2 ≠ 0. Therefore I know the current is conserved, ∂νJν = 0, as a
consequence of the equations of motion. Thus my two problems have a single solution. If I can arrange my matter
Lagrangian L′ such that it is gauge invariant, and then break gauge invariance only by giving the photon a mass, I
will obtain a conserved current. To solve the gauge invariance problem for the massless photon is to solve the
conserved current problem for the massive photon. It’s straightforward algebra.

I should emphasize that gauge invariance (δL′ = 0 under (27.12) and (27.15)) is not a symmetry. The
conserved current (5.27) associated with it (not to be confused with Jν in (27.20)!) is zero, simply because the
Lagrangian doesn’t change at all.2

There is a big difference between gauge invariance and an internal symmetry, such as an isospin rotation in
the theory of pions and nucleons. There is a difference between a proton and a neutron, and they are turned into
each other by an isospin rotation. But there is no physical difference between the state of the electromagnetic field
in Lorenz gauge or in Coulomb gauge or in any other gauge. Gauge invariance is like general coordinate
invariance in general relativity, or like the statement that the contents of a physics paper are unchanged if it is
translated from English into French. These are different descriptions of the same system, not different systems
with symmetric dynamics. That’s not clear from the analytic structure of the theory, but it’s the meaning we attach
to the physics in both cases. We could test isospin invariance in the real world (in situations where the
electromagnetic force can be neglected), by doing proton–proton scattering and then doing neutron–neutron
scattering. You would get two papers, one describing proton–proton scattering at 300 GeV, and the other
neutron–neutron scattering at 300 GeV, and they would produce the same cross-section. That would be a test.
But you will never read a paper saying photon–electron scattering in the Coulomb gauge and photon–electron
scattering in the Lorenz gauge give you the same cross-section, and claiming to have verified gauge invariance.

So far I’ve shown that if L′ is chosen to be gauge invariant, that has the desirable effect of preserving the
gauge invariance of massless electrodynamics, and if we break the gauge invariance just by adding a photon
mass term, we obtain a massive vector meson coupled to a conserved current. Now I will tackle the problem of
constructing a Lagrangian invariant under the gauge transformation (27.12). The key to the construction is a
prescription, called minimal coupling, that will generate extra terms, such that the resulting Lagrangian will
automatically be gauge invariant. These terms will not involve the derivatives of Aµ, but only Aµ itself. I’ll explain at
the end where the word “minimal" comes from. The prescription is a machine, but it is not universally applicable.
You give me a Lagrangian for matter (or whatever you want to call it) without electromagnetism, and, provided one
condition is met, I’ll generate the Lagrangian including the interactions with electromagnetism.

The necessary condition is that the matter Lagrangian Lm(ϕ, ∂µϕ) describing a set of fields—scalar, spinor,
vector, whatever, but not including Aµ itself—has a one-parameter group of internal symmetries of the sort we
talked about earlier (in Chapter 6). Under this group, an infinitesimal transformation of ϕ is given by

or, for a finite transformation,

If this transformation leaves Lm unchanged, under normal circumstances it would enable us to deduce the
existence of a conserved current, Jµ, (5.27):

This seems a reasonable and necessary condition to get an interaction Lagrangian coupled to the photon. If the
photon is coupled to the matter fields with a constant, say the electromagnetic coupling constant e, and if you
have a conserved current when the photon is around, you can imagine it should stay conserved as e → 0. So you
want to start out with a theory that has a conserved current even before you include the photon. I’ll give two
examples.

EXAMPLE 1. The Lagrangian for a free Dirac field is

If we assemble ψ and into a two-component vector ϕ,

and take Q to be a matrix with eigenvalues ±1,

then the spinor ψ has eigenvalue +1, and its Dirac adjoint has eigenvalue −1:

Following the formula (5.27) for the construction of the conserved current,

We already know that this current is conserved:

EXAMPLE 2. The Lagrangian for a free charged scalar field is

Again we’ll say

The current associated with this symmetry is

This current is also conserved;

Now to couple the vector meson. Our first try is to add the term −eJµAµ, where e is a coupling constant:

where LP is the Proca Lagrangian (26.47) and Lm (either (27.25) or (27.31)) is the free matter Lagrangian. Note
that this requires a slight redefinition of the current (compare (27.9)):

Is the current still conserved? It is in Example 1, but not in Example 2. In Example 1, the equations of motion
become

Multiply the top equation by on the left, the bottom by ψ on the right, and subtract these from each other to find

so the current remains conserved. In Example 2, however, the equations of motion become

Multiply the first by ϕ ∗ and the second by ϕ, and subtract the second from the first:

which is not necessarily zero. If we take the divergence of the vector meson equation of motion, we get, discarding
the zero term ∂µ∂νFµν,

We can substitute for the divergence of Aµ into the previous equation to obtain

or rewriting,

We could try to iterate the Lagrangian to see how to add higher powers of e, but there’s a better way.

27.2The minimal coupling prescription

To motivate the minimal coupling prescription, consider this set of transformations on the matter fields, of exactly
the same form (27.22) as before, but now with δλ an arbitrary function of space and time:
That is not, of course, an invariance of our Lagrangian, because if I compute δ(∂µϕ), I’ll get two terms;

The first term is hunky dory, no problem there. But the second term is a disaster. However we’ve also got the
electromagnetic field involved, which under an infinitesimal gauge transformation obeys (27.13). Consider a
combination of the ordinary derivative ∂µ acting on ϕ with a product of the vector field Aµ times ϕ:

This expression is called the covariant derivative. Its transformation δ(D µϕ) looks like this:

We can make the second and third terms cancel if we choose

in which case

The covariant derivative of ϕ transforms in exactly the same way as ϕ, which is why it’s called “covariant”. The
combined transformations are

or, for finite transformations,

This set of transformations is called collectively a “gauge transformation”, though historically that term was
applied, as in (27.12), only to the potentials. The infinitesimal parameter δλ(x) is “local”; it is a function of space
and time. These transformations emerged from Hermann Weyl’s generalization of general relativity in 1918. He
suggested that not only could you rotate local coordinate systems, but that there was no absolute standard of
length. Weyl’s theory used a real function in the exponent. He was wrong; but this idea has resurfaced as
conformal invariance on a string’s world sheet. Weyl called his theory eichinvarianz, or “scale invariance”, but it
was translated into English as “gauge invariance”. It was reintroduced with an imaginary exponential, by F. London
in 1927, who named it after Weyl’s theory.3 Gauge transformations have no “active” interpretation, only a
“passive” one; it’s just a change of coordinates.

The minimal coupling prescription is simple:

That is,

This Lagrangian is invariant under the transformations

so the Lagrangian has a conserved current,

The equations of motion become

No matter how complicated L is, the right-hand side of the Proca equation is a conserved current.

EXAMPLE 1, revisited

which gives us the same conserved current as before (27.29).

EXAMPLE 2, revisited

Something more interesting happens in the scalar case. Minimal coupling

leads to the interaction

including both a linear coupling of Aµ to the derivatives of ϕ and ϕ *, and a quadratic coupling of Aµ to ϕ, and to a
current (27.36)

Notice that the terms in the electromagnetic current linear in the vector field Aµ do not cancel, as they did with the
Dirac Lagrangian. Of course, you could say that the current must have an Aµ in it to be gauge invariant. The
current (27.61) could also be written

Aside from the corrections required to make the electromagnetic current gauge invariant, (27.61) has the same
form as the conserved current (27.33) in the non-interacting case.

Is this current conserved? The scalar equations of motion become

Using the same trick as before, we find

If we now take the divergence of the minimal coupling current (27.61), we get

The right-hand side is just i times the left-hand side of the previous equation. Thus

Although the original current (27.33), without the Aµ term, was not conserved, the current from minimal
coupling is conserved. The minimal coupling prescription gives the right result, with minimal effort! The new
current reproduces the derivative coupling term obtained before, but it also includes a term directly proportional to
the vector meson Aµ. It’s got to have an Aµ in it to make the derivative in the current covariant, as in (27.62). This
latter term produces in the Lagrangian a quadratic non-derivative interaction,

In its detailed structure the electrodynamics of charged scalar particles is very different from that of charged
spinor particles, and the Feynman rules we’ll get eventually in these two cases will also be quite different. In
particular, the quadratic term will give rise to “seagull” diagrams, as shown in Figure 27.1. In both cases, the
currents obtained with minimal coupling also have the desirable property that they remain conserved in the limit as
the electromagnetic coupling constant goes to zero. Nevertheless, it’s the same prescription for both. You give me
a Lagrangian that has a conserved current in the absence of electromagnetism, and I will generate the gauge
invariant Lagrangian that describes the coupling with electromagnetism, with the minimal coupling prescription.
The electromagnetic current Jµ is given by (27.36).

Figure 27.1: Seagull diagram in scalar electrodynamics

It should not be thought that minimal coupling between a vector meson and a “matter” field is the only way to
obtain a conserved current. For example, consider this Lagrangian, coupling a Dirac field to a vector meson:

where (20.98) σµν = i[γµ, γν]. The extra term is sometimes called a “Pauli term”.4 That’s perfectly gauge
invariant: Fµν is gauge invariant, and σµνψ is gauge invariant without any funny business with Aµ; it doesn’t
involve any derivatives. That’s why minimal coupling (the vector field-matter field coupling arising only as part of
the covariant derivative) is called “minimal” coupling.5 You could always complicate matters by including additional
gauge invariant terms to the Lagrangian, which would therefore still yield conserved currents. This Lagrangian has
a conserved current,

The second term comes from shifting the derivative in the Pauli term from Aµ to the product σµνψ with an
integration by parts;

How might such a term arise? Consider a different Lagrangian:

Because σµν is antisymmetric and ∂µ∂ν is symmetric, this vanishes trivially. Suppose instead that we first apply
minimal coupling:

But D µD νψ is not symmetric, and this term does not vanish. In fact,

In four-dimensional theories, we usually will not have to worry about these non-minimal interactions because
typically they, and in particular this one, turn out to be non-renormalizable. Since we have not discussed the
renormalization problem for electrodynamics—we don’t even have the propagator for the free photon—this is a
premature remark. I make it anyway.

27.3Technical problems

I would now like to turn to the second topic: technical problems in quantizing electrodynamics. In principle
everything is all set up. We know what our interactions are. We have this gigantic machine, canonical quantization
with canonical commutators for bosons and anticommutators for fermions. All we have to do is grind through:
develop a Hamiltonian, write down Dyson’s formula, apply Wick’s theorem, get the Feynman rules, use the
renormalization prescription, pull out the S-matrix and happily start computing, say, the anomalous magnetic
moment of the electron to order e8.

Now why am I not going to launch into that progression? Well, things get pretty complicated. There are three
reasons why they get complicated. They’re all technical complications; none of them is insuperable and none of
them really requires the panacea that will emerge a little later, but they all lead us into horrendous complications.
One is the problem of gauge invariance, which arises when µ2 = 0. You remember that for µ2 ≠ 0, we had no
problem canonically quantizing the massive photon, nor any problem quantizing the free field theory. However for
µ2 = 0, the canonical quantization program came (26.66) to a screeching halt; we couldn’t do it. I can now reveal
that the reason is gauge invariance.

The canonical quantization program, indeed the Hamiltonian formulation of classical mechanics, depends on
having a complete and independent set of initial value data. Take, for example, massive vector meson theory (p.
562). Say I give you the fields, three Ai(x, t)’s and the three F0i(x, t)’s, at time t = 0, and their first thirty-two
derivatives at time t = 0, and you propose to tell me, by solving the equations of motion, what they are at any future
time. In the Proca theory there’s no problem. But in a gauge invariant theory this is impossible. Suppose you’ve
determined {Ai(x, t)} for all subsequent times. I say these solutions cannot be unique. For I have the freedom to
gauge transform this set of fields {Ai(x, t)} with a function χ(x, t) whose derivatives ∂iχ(x, t) vanish at t = 0 and within
a little slice of width ϵ around t = 0, but are nonzero for t > ϵ. Then the transformed fields {A′i(x, t)} (and their time
derivatives) are the same as the original set at t = 0, and different at later times; in particular, say, at t = 1.
Therefore I have two sets of fields, {Ai(x, t)} and {A′i(x, t)}, with exactly the same initial value data, but they are
different at t = 1. You can’t possibly get a unique solution for the initial value problem, and you can’t possibly write
electromagnetism in Hamiltonian form. It’s just a consequence of gauge invariance. Were we to attempt canonical
quantization of electromagnetism, we’d first have to impose a gauge condition, for example the Lorenz gauge
∂µAµ = 0, or the Coulomb gauge ∇ • A = 0, or some other. But the Coulomb gauge destroys manifest Lorentz
invariance, while the Lorenz gauge does not specify the fields completely.6 Is the resulting theory Lorentz
invariant, and are the results independent of the choice of gauge, as they must be? These questions are much
easier to address if we quantize using a different method, that of functional integrals, to which we will soon turn.

The second problem I have alluded to many times: the problem of derivative couplings. I have always been
sort of antsy whenever questions about derivative couplings came up in these lectures, and I mumbled “Hrmph,
hrmph, we’ll talk about that later.” The standard tricks can be applied to theories with derivative couplings, and
we’ll have to confront such theories (such as scalar electrodynamics), but they lead to a terrible mess. I’ll work out
some details so you can stand on the precipice and look into this canyon full of garbage, to see exactly what
would happen and what sort of problems we would run into if we tried to take seriously a theory with derivative
couplings. (Before, in §14.4, we just guessed.) But we’ll pull back, and not plunge into that pit.

EXAMPLE. Pseudoscalar-spinor derivative coupling

Let’s consider a simple example, a spinor field ψ interacting with a pseudoscalar meson field ϕ via this
interaction,

with some coupling constant f. I’ll just let Kµ stand for γµγ5ψ. Let’s try to write this in Hamiltonian form. The
Lagrangian is

It has none of the special problems of electromagnetism. There’s no momentum conjugate to , but we
won’t worry about that. The canonical momentum conjugate to ψ is i γ0; no problem there. The canonical
momentum conjugate to ϕ, however, looks a little funny:

(Note the subscript f, to distinguish this quantity from the free canonical momentum π = ∂0ϕ.) This means the
interaction Hamiltonian H I will not equal minus the interaction Lagrangian, LI,

because of the presence of this extra term, fK0.

To be more explicit, if we focus on the terms in the Hamiltonian that involve time derivatives of ϕ, we’ll find first

which is not simply (πf)2, as it would have been if we’d had a non-derivative coupling. Second, we can write the
term in the Lagrangian involving time derivatives of ϕ like this:

When we assemble the Hamiltonian using the standard formula, all the terms that do not involve the derivatives of
ϕ will come together to give us the free Hamiltonian for the scalar and spinor fields, the original interaction
Lagrangian, and one extra piece:

In Dyson’s formula you’re interested in H I, the interaction Hamiltonian density in the interaction picture, written as
a function of interaction picture fields. Here,

The term −fKµ∂µϕ is a nice Lorentz invariant object to go up there in the exponential. But the last term, the little
bastard, doesn’t cancel out! It’s as disgustingly non-Lorentz invariant as an object can be:

and it’s sitting there in the interaction Hamiltonian, in the exponential in Dyson’s formula, giving us what looks like
a terrible, non-Lorentz invariant four point vertex involving four Fermi fields in our Feynman rules.

This is not the only difficulty that would come up in this theory. In addition to this four Fermi, non-covariant
term in our interaction Hamiltonian in the interaction picture, we have also a non-covariant contraction term, as I
will demonstrate. I remind you of the definition of a time-ordered product of two Bose operators, A(x) and B(y),

Suppose I wanted to compute the time derivative (with respect to x0) of this thing. From differentiating the field
operators I obtain simply the time-ordered product of the derivative ∂x 0A(x)B(y)—couldn’t be nicer. But when I
differentiate the theta functions, I get a delta function. So I get three terms,

or equivalently

In the theory we wish to consider here, our interaction Hamiltonian not only involves ϕ’s, but time derivatives of
ϕ’s. So we’ll have to compute the contraction function for two time derivatives of ϕ’s which we’ll do using this
identity. The contraction function is the vacuum expectation value of the time-ordered product, so I might as well
work with the time-ordered product.

Let’s consider two derivatives on the time-ordered product of two ϕ’s (in the interaction picture),

First bring the y time derivative through. We can do that with no problems because the equal time commutator that
develops is the equal time commutator of ϕ(x) with ϕ(y), which is zero;

I’ve still got the second derivative to worry about, and when I bring the x time derivative through, the extra term
(the equal time commutator) is not zero, and I get

or,

Thus we have a rather peculiar equation that we’ll have to also feed into our Feynman rules in computing the
contraction functions of ∂µϕ I(x) with ∂νϕ I(y). There’s no problem with the space terms. Pitching terms onto the
other side,

We have a non-covariant interaction Hamiltonian, and a non-covariant contraction function. The first term on the
right is a nice covariant object, with the same Fourier transform as the Feynman propagator, with an extra couple
of k’s in the numerator, because the derivatives are outside, but then we’ve got this extra term which is disgusting.

I have led you to the edge of the precipice, but we’re not going to plunge into that pit of garbage. Of course
these two diseases turn out to be each other’s cure. The theory is after all Lorentz invariant, and you must get
Lorentz invariant answers finally. It turns out that the disgusting term in the interaction Hamiltonian cancels the
disgusting term in the contraction function, after a horrendous amount of combinatorics that I’m not going to do.
You can now see the sort of problems that we would have to deal with if we attempted to treat this theory in a
straightforward way. I promise that you get all the correct answers this way, but to redeem that promise would
require a lot of work.7

The third technical difficulty is the same problem we encountered attempting to compute the propagator or the
Hamiltonian for even the massive photon field. Here it comes again. If you recall, doing the free theory, (see
(26.45) and the sentence following) we had to eliminate A0 from the Lagrangian before we could write down the
Hamiltonian. The equation that eliminated it was the field equation evaluated for µ = 0. In an interacting theory
we’ll have the modified field equation where this will now be a function of all those charged particle fields in the
theory and therefore the equation we will have is

In the course of eliminating A0 we’ll introduce terms in the interaction Hamiltonian of the form (J0)2, from squaring
A0, just like those ugly (K0)2 terms we have here. Likewise when computing the A0 propagator, because A0 is
related to a canonical momentum, we’ll have exactly the same sort of problems we had when we were computing
[∂0ϕ(x), ∂0ϕ(y)]. A0 is related to the time derivative of Ai, and if we attempt to compute the A0 propagator we’ll get
non-covariant terms there for the same reason we got them in the previous example. So even in spinor
electrodynamics with a massive photon, with none of the other problems of derivative interactions, in the course of
eliminating A0 to set up the theory in Hamiltonian form, we will find almost as many troubles as in the previous
theory. This is because A0 is related to a momentum density just like ∂0ϕ, and squaring this term will give us a
non-covariant interaction.

So it looks like we have a lot of problems. None are insuperable, in principle. If we just kept the faith, plugged
along, did all our combinatorics right and brushed our teeth every day, we would no doubt arrive at the right
answer. It just gets messier and messier. This is a good point for us to break off the discussion of electrodynamics
and begin to discuss a method that allows us to organize some of the mess. Things will still be hairy once we have
learned this method, but they will be considerably less hairy than if we’d attempted to solve the same problem by
straightforward means. Thus I will begin next time the topic of a method which is very useful in doing complicated
problems of this kind, the method of functional integrals.

1[Eds.] Readers may wish to know in advance Coleman’s strategy for discussing features of the massless vector
field: to consider those features one after another in relation to a massive vector field, and then take the limit as its
mass goes to zero.
2[Eds.]In response to a student’s question, Coleman reiterates, “No, it is not a symmetry. The conserved current
associated with it is zero.” For a somewhat different viewpoint but much the same conclusion, see S. Weinberg,
“Dynamic and Algebraic Symmetries”, pp. 290–393, in Lectures on Elementary Particles and Quantum Field
Theory (1970 Brandeis University Summer Institute in Theoretical Physics), v. 1, eds. Stanley Deser, Marc
Grisaru, and Hugh Pendleton, MIT Press, 1970. Weinberg finds (equation (2.B.10)) that there is a non-zero
conserved current,

but that it does not lead to a new charge independent of that found from global phase invariance with χ a constant.
It is perhaps worth noting that Coleman’s prescription requires setting χ = 0. If that is done here, the current
vanishes.
3[Eds.] H. Weyl, “Gravitation und Elektrizität”, Sitzungs. Pruess. Akad. iss. Berlin (1918) 465–480. English
translation in L. O’Raifeartaigh, The Dawning of Gauge Theory, Princeton U. P., 1997. O’Raifeartaigh also
includes an English translation of London’s paper, F. London, “Quantenmechanische Deutung der Theorie von
Weyl” (Quantum mechanical interpretation of Weyl’s theory), Zeits. f. Phys. 42 (1927) 375, and likewise of Weyl’s
independent article, “Elektron und Gravitation”, Zeits. f. Phys. 56 (1929) 330–352. See also J. D. Jackson and L.
B. Okun, “Historical roots of gauge invariance”, Rev. Mod. Phys. 73 (2001) 663–680, and H. Weyl, Space-Time-
Matter, translated by Henry L. Brose, Dover Publications, 1952; in §16, eichinvarianz is translated as “calibration
invariance”. In typed notes for Feb. 11, 1999, Coleman suggested consulting Pauli’s famous review article,
“Relativitätstheorie” (1921), for a discussion of Weyl’s theory: W. Pauli, Theory of Relativity, Pergamon, 1958,
§65, pp. 192–202; note 20, p. 223; republished by Dover Publications, 1981. Pauli was all of 21 when he wrote the
article.
4[Eds.] W. Pauli, “Relativistic Field Theories of Elementary Particles”, Rev. Mod. Phys. 13 (1941) 203–232. The
Pauli term appears in equation (91) and is defined in the previous (unnumbered) equation.
5[Eds.]
The name is due to M. Gell-Mann, “The interpretation of new particles as displaced charge multiplets”,
Nuovo Cimento 4, Supplement 2, (1956) 848–866.
6[Eds.]For canonical quantization of the Maxwell field in the Coulomb gauge, see Bjorken & Drell Fields, Chap.
14, pp. 68–80. For canonical quantization in the Lorenz gauge, see, e.g., Franz Mandl and Graham Shaw,
Quantum Field Theory, John Wiley and Sons, 1984, Chap. 5, pp. 86–90. The original canonical quantization of
the electromagnetic field is due to Enrico Fermi, “Sopra l’Elettrodinamica Quantistica” (On quantum
electrodynamics), Rend. Lincei 5 (1929) 881–887, reprinted as paper 50 in v. I of Fermi’s Collected Papers, ed. E.
Segrè et al., U of Chicago Press, 1962. A much expanded version of this work, in English, is E. Fermi, “Quantum
Theory of Radiation”, Rev. Mod. Phys. 4 (1932) 87–132, reprinted as paper 67 in Fermi’s Collected Papers, v. I
and in Schwinger, QED.
7[Eds.] A similar example was given in §26.5. See also Section 4.4 of Chapter 4, “Secret Symmetry”, pp. 154–156
in Coleman Aspects, and the paper cited there, Kuo-Shung Cheng, “Quantization of a General Dynamical System
by Feynman’s Path Integration Formulation”, J. Math. Phys. 13 (1972) 1723–1726. (But see also note 7, p. 623.)
For an explicit example showing the combinatorics and cancellation of two non-covariant pieces, see the
discussion of scalar electrodynamics in Itzykson & Zuber QFT, pp. 282–285.

Problems 15

15.1 In class (§26.1) I constructed a vector field theory for which the solutions were four-dimensionally transverse
waves, and I quantized it to construct a theory of free vector mesons. In this problem you are asked to carry out
the same program for the complementary theory, one for which the only solutions are four-dimensionally
longitudinal.

Consider

where µ2 is a positive number. Derive the field equations. Show that the solutions are longitudinal waves of mass
µ. Show that A0 and its conjugate momentum are a complete set of initial value data. Construct the Hamiltonian in
terms of these, and determine the overall sign of the Lagrangian by demanding that the Hamiltonian be bounded
below. Show that if you make an appropriate identification of A0 and its conjugate momentum with ϕ and π of
Klein-Gordon theory, the Hamiltonians of the two theories are identical. (Thus there is no need to go any farther in
the quantization program.)
(1998b 2.1)

15.2 In a theory of a free vector meson, compute the Hamiltonian as a function of annihilation and creation
operators. Normal order freely.
Comment: I doubt if there is a single person who will be surprised by the answer to this problem. Nevertheless, it’s
fun to see how it comes out of that mess of F’s and A’s.
(1998b 2.2)

15.3 Let ψ 1 and ψ 2 be two Dirac fields, of mass m1 and m2 respectively, and let Aν be a real vector meson field of
mass µ. Let the interaction of these fields be

with g a real number. If m1 > m2 + µ, this interaction will cause ψ 1 to decay into ψ 2 and Aν. Compute the decay
width Γ (12.33) for this process to lowest non-vanishing order in perturbation theory.

Comment: The vector meson in this problem is coupled to a non-conserved current, because m1 ≠ m2. This leads
to disaster when µ goes to zero. You should see this disaster in your answer.
(1998b 2.3)

15.4 Consider the theory of a Hermitian scalar field ϕ defined by

Here µ is the renormalized mass, g is the positive renormalized coupling constant, C is the O(g2) coupling
constant renormalization counterterm, and the dots indicate the mass and wave-function counterterms. (I have not
bothered to write these out explicitly because they are not needed for the problem at hand.)

To lowest order, the amplitude for two-particle elastic meson–meson scattering

is associated with the single graph

The amplitude is given by

(the 1/4! is cancelled by the 4! ways in which the four meson fields can annihilate and create the incoming and
outgoing mesons). We define the renormalized coupling constant g (and determine, order by order, the
counterterm C in perturbation theory) by insisting that the above equation (P15.4) be exact (and so, no
contributions from higher-order graphs) when all four mesons are on the mass shell, at the symmetry point s = t =
u = 4µ2/3. I remind you that for the scattering process

the Mandelstam variables (11.19) are

Compute the meson–meson elastic scattering amplitude to order g2. Express the answer as a function of s, t,
and u. Although it is possible to do all integrals in the problem explicitly, it suffices to express your answer as a
sum of terms, each of which is written as an integral over a single Feynman parameter. Check your answer by
verifying that, to O(g2), the forward scattering amplitude obeys the Optical Theorem (12.49). Note that to get the
total cross-section to O(g2), you only need the scattering amplitude to O(g), which I have given you. (HINTS. Be
careful not to double-count the final states when computing the total cross-section.)
(1975 253a Final, Problem 4; 1986 253a Final, Problem 2)
 Solutions 15

15.1 We consider the Lagrangian density

where s = ±1, to be determined. Independent of s, the Euler-Lagrange equations give

As in (26.5), we look for plane wave solutions Aµ = εµe−ik⋅x , and find

If k ⋅ ε = 0, the only solutions are trivial. Since the polarization is parallel to the momentum, the plane-wave
solutions we have found are longitudinal waves. Take the inner product of (S15.2) with kν to find

The longitudinal waves have mass µ.

We now show that initial value data for A0 and its conjugate momentum,

are enough to determine Aν at any time. (The momenta πi conjugate to the Ai are all zero.) Taking the divergence
of (S15.1), we obtain the Klein–Gordon equation for π0:

It follows that π0 is completely determined by initial value data for π0 and ∂0π0. Setting ν = 0 in (S15.1), we have

and thus initial value data for A0 and π0 completely determine π0 at any time. Again from (S15.1),

and therefore Aν is completely determined as well.

Taking ∂0A0 from (S15.3), the Hamiltonian density is given by

Substituting the right-hand side of (S15.5) for Ai, we find

where we’ve used s2 = 1, and the dots indicate a three-divergence (s/µ2)∂i(π0∂iπ0) which we can convert to a
surface integral at infinity. Each term between the square brackets in (S15.6) is non-negative (note the lowered
superscript i), and so is the Hamiltonian if we choose s = +1. Making that choice,

If we define
then, in terms of these variables, the Hamiltonian becomes

which is indeed identical to the free Klein–Gordon theory, as was to be shown. The extra minus sign in (S15.7)
attached to ϕ (or alternatively, to π) is needed to get the correct equations of motion„ e.g., ∂0ϕ = π, as follows from
(S15.4).

15.2 We start with the Hamiltonian density (26.46) for a free massive vector meson, slightly rewritten:

Writing Aµ in terms of creation and annihilation operators, we have

We recall that the polarization vectors are orthonormal:

If we define

then we can write

Because

it follows that

To save space and writing, let

From (S15.12) and the judicious use of (S15.13) we find

There are eighteen terms in the Hamiltonian density. We know that H must be time-independent, and we know
that the a and a† operators are independent. So just to see how this goes, let’s look only at the four integrands
0
which are proportional to ak0a−k, 0e−2ik t. The coefficients of this factor in the Hamiltonian density are

The other time-dependent terms cancel similarly. The time-independent terms are

From the definitions (S15.11) of aµ and aµ† and the normalization (26.24), it follows

and so finally

This result is not a surprise; it’s surely what we all expected, in agreement with (26.64).

15.3 Applying the Feynman rules in the box on p. 571 to the diagram in the problem, we can write the amplitude
for the decay process as

where u1s and u2s′ are the polarization spinors corresponding to ψ 1, ψ 2 with momenta p1 and p2, and εν(r)(k) is the
polarization vector of Aν, with p1 = p2 + k. Summing over initial spins and averaging over final spins, we then have
(21.115)
letting A and B stand for the trace terms. Now we use the trace identities worked out in Problem 11.2, dropping
terms with an odd number of γ’s:

B can be simplified a little, because γµ( 2 + m2)γµ = −2 2 + 4m2, so

That gives

In the rest frame of the decaying ψ 1, p1 = (m1, 0), and p1 = p2 + k, so eliminating k and Eγ,

(Note that |p2| is imaginary for m1 < m2 + µ, as it should be.) The decay amplitude in ψ 1’s rest frame becomes

For the decay probability per unit time, Γ, we have

where D is the density of final states, given by

Finally, from (12.33) (see also Example 2 on p. 251)

(The factor comes from averaging over final spins.) We see in this answer the foretold disaster as µ → 0, that Γ
diverges. This is because the current j µ = g(ψ 1γµψ 2 + ψ 2γµψ 1), to which the vector is coupled, is not conserved:

The proximate cause of the non-conservation of the current is that m1 ≠ m2, just as the problem stated. Indeed, if
m1 = m2, not only is the current conserved, but the troublesome term in (S15.25), divergent as µ → 0, vanishes.
The moral of this story is that coupling a current to a massless vector does not work unless the current is
conserved. See the discussion on p. 579 about the necessity of a conserved current for the principle of minimal
coupling to work.

15.4 To order g, the only scattering graph is

As stated, this leads to

To O(g2), there are three scattering graphs, and the counterterm:

Graph (4) is the charge constant renormalization diagram to O(g2); its value is

The value C 2 is fixed by the renormalization condition

Thus, to the desired order,

All we need to do is to evaluate graphs (1), (2), and (3). Note that the subtraction eliminates the (logarithmic)
divergence, so all the integrals are finite.

Let’s look at graph (1):

The factor of (1/2) arises because the (1/4!) at the vertices are incompletely cancelled; there is no way of
distinguishing the two internal lines.

Note that A1 is a function of (p + q)2 = s only. Likewise, A2 is the same function of (p − q)2 = t, and A3 of (p −
q′)2 = u. Furthermore, the integral is identical to one we did in class, the meson self-energy −i f(p2) in Model 3
(15.57), with p replaced by p + q (remember, the “nucleons” in Model 3 were scalars). Using the result (16.2), we
have

Then

Now to check the Optical Theorem,

Using the standard formula (12.26) in the center of momentum frame,

and thus

the extra factor of coming from the scattering of identical particles. Since s = ET2 = (2 )2 = 4(|pi|2 + µ2),
we have
For forward scattering, s ≥ 4µ2, t = 0, and u = 4µ2 −s ≤ 0. That means f(t) and f(u) are real, and the imaginary part
of A can come only from f(s):

To investigate the imaginary part of f(s), recall the nearly identical integral (S9.1) in Problem 9.1, p. 349. The
integrand has an imaginary part only when

In this region of x, the imaginary part of the logarithm is −π (see (S9.2)). By exactly the same analysis used in
Problem 9.1,

where

and so

in agreement with (S15.36). The Optical Theorem is valid, to O(g2). Note that

That is, to second order in g, Im A approaches a positive, finite constant.

28
Functional integration and Feynman rules

The first version of functional integrals (integrals over an infinite number of dimensions), called “path integrals”,
was introduced into physics in the late 1940s by Feynman,1 but these methods were not fully appreciated until the
early 1960s. This method will give us enormous advantages, and enable us to settle with derivative coupling,
superfluous variables, gauge invariance and so on. Functional integrals are sometimes called “integration over
function spaces” or “integration over infinite dimensional spaces”.2 We all know how to do integrals over a one-
dimensional space, or over n-dimensional space; I am now going to take n to infinity.

28.1First steps with functional integrals

For the moment, we’ll put aside the vector fields and their associated problems, and talk about what will in the first
instance be purely a topic in mathematics (butcher grade—the way physicists do it), and then we’ll eventually
come back and develop a bunch of techniques using this mathematical method that will help us unravel things.

I begin with a simple one-dimensional integral, the Gaussian,

where a is a positive, real number to ensure damping of the integral at infinity. By analytic continuation, the identity
is true whenever the integral converges. That is to say, a can be a complex number, so long as its real part is
greater than zero. We can also do the n-dimensional version of this integral. Let x be a vector in this n-
dimensional space. For some symmetric matrix A, I define

By making an orthogonal transformation to diagonalize A, I instantly find that

so that

provided all of the eigenvalues λi are positive—or, by analytical continuation, if Re(x, Ax) > 0. That is enough to
make the integral converge. These factors of 2π are irritating, so I will introduce a notational convention. I will write

and write the previous integral as

Of course, if we can do a Gaussian integral, we can do a general quadratic form by completing the square.
Consider the quadratic form

where b is some n-vector. Q(x) is minimized at x = x;

Then

and so

Thus I find, with y = x − x,

where e−Q(x) = exp is a constant.

Once we can do a general quadratic form, we can do a polynomial times a generalized Gaussian. If I have
any polynomial P(x) in x, an expression of the form

can be computed by taking derivatives;

That is, whenever I see a component of x (x1 or x2 or x17), I differentiate with respect to the same component of
b—∂/∂b1 or whatever—this drags that component of x down from e−Q(x). Now I take the derivative outside the
integral
For example,

So I have told you something you no doubt already know, although perhaps in a somewhat more compressed
notation than you are used to: how to integrate Gaussians, generalized Gaussians, and polynomials times
generalized Gaussians.

It will turn out for later purposes to be convenient to integrate over functions not only of real n-vectors but
complex n-vectors. I don’t have something fancy in mind involving contour integrals or anything like that, I just
mean integrating over the real part and then integrating over the imaginary part. In particular, I’ll take a complex
vector and break it up into real and imaginary parts like this:

and similarly for z*. I’ve included the , for reasons that will become clear shortly. 3 I define

whence it follows for example that

Well, it’s pretty trivial. You diagonalize A, you write z in terms of x and y, and the comes in automatically as I’ve
arranged matters. I simply have one integral for the x and one integral for the y. So each eigenvalue occurs twice,
and I get the exponent now equal to −1 rather than − . And similar formulas follow for generalized quadratics and
polynomials times general quadratics.

Now comes the big leap of faith. I have arranged all the formulas so that the dimension n of the vector space
over which I am integrating never appears explicitly. Therefore I am going to simply extend these formulas to an
infinite-dimensional space! This is a functional integral. We’re just going to say these formulas define integrals of
Gaussians, and polynomials times Gaussians, which will turn out to be practically everything we will need to do,
over an infinite-dimensional space. Everything is exactly the same, except that the sums in (28.2) and (28.3) run
not from 1 to n, but from 1 to ∞. Obviously this involves deep and subtle mathematical questions, about which I will
say nothing.

More generally, you start out with an infinite-dimensional space, a Hilbert space of some sort. You have a
quadratic form on it, defined by some infinite matrix, some positive definite operator. That’s completely legitimate.
Then you take that infinite-dimensional space, and you look at a finite-dimensional subspace. You compute the
integral in that finite-dimensional subspace, and just restricting yourself to that subspace, you can compute the
determinant. Absolutely no problem there. Then you let the finite-dimensional space get larger and larger, until it
fills out the whole space, adding basis vector after basis vector one at a time. If there’s a limit of the integral, it will
be the limit of the determinant. If there isn’t, that’s our bad luck.

It’s a deep question, which we leave for the mathematicians, to determine for which quadratic forms the limits
of the integral and the determinant exist. Another deep question which we leave to the mathematicians is, if the
limits exist for one way of filling up the Hilbert space, do they exist for another choice of basis vectors, where you
fill the space out in another order? We’ll just leave these questions alone, and blithely manipulate equations,
assuming that everything will be okay unless something goes wrong. If we can compute the integral, no doubt it
can be rigorously shown to exist. If we get zero or infinity or something like that, then we’re going to be in trouble.
We’ll try to avoid that sort of thing. Since we’re going to apply these things to field theory, of course we will get zero
or infinity an awful lot of the time, but those will just be our old friends the ultraviolet divergences coming up again,
and we can get rid of them by cutting off the theory in any one of the standard ways.

You can also do this for continuous spaces: the set of all functions in 4-space, for example, or all functions on
a line. These can be turned into a discrete space by expanding the functions in terms of, say, harmonic oscillator
wave functions. Likewise for the set of all functions in n dimensions. So there’s no difference between a discretely
infinite Hilbert space and a continuously infinite Hilbert space; that’s just the difference between a discrete basis
and a continuous basis.

Now there are two points I want to make. First, the sort of space over which these integrals are defined is a
very big space. In fact, precisely how big it is doesn’t matter. It could be a Hilbert space, it could be a bigger space,
the space of all continuous functions. It hardly matters when we do the integral because of this exponential
damping. If you throw in some finite number of basis vectors that are badly behaved in one way or another, the
exponential damping will cut them out; they’ll make a zero contribution to the integral.

Just to emphasize how big it is, I will consider an infinite-dimensional space, and the simplest possible
Gaussian integral, for A = 1:

Let’s now consider a function which is not a polynomial, the step function θL localized on a gigantic box. Define

That is to say, this is a step function equal to 1 inside an infinite-dimensional hypercube of edge length 2L, and
equal to zero elsewhere. Then4

Now what is this quantity? Well, each of the infinite number of terms is identical, and each of them is a little bit less
than 1, no matter how big L is, so long as L < ∞. Take an infinite product of terms, each of them or or , you
always get the same answer: zero! So this integral is completely well defined, and is equal to zero. Function
space is so big that if you use θL(x) to define a measure on function space—it’s a positive functional, and so
defines a measure, by giving a volume to every set—then the measure of an infinite-dimensional box of side 2L is
zero! There’s a lot more outside than inside. That’s a set of measure zero, like the rational numbers inside the real
numbers with ordinary Lebesgue measure. Function space is a very big space. There’s more outside the box than
inside. If I take a slice on a straight line I’ve got a lot outside and little inside; if I take a box in a plane I have a lot
more outside compared to inside; if I take a cube I’ve got even more outside than inside. When I go to infinite
dimensional space I’ve got hardly any inside at all compared to outside. You can get lost in it if you try to be
careful, so we won’t be. You should be warned about that.

The second point has to do with the choice of a basis. Here I have used an infinite-dimensional space
described in terms of a discrete basis. But I could equally well define an infinite-dimensional space in terms of a
continuum basis. For example, the space could be the space of all real, integrable functions ϕ(x) defined on 4-
space, as a Hilbert space. The inner product is

I could define a quadratic form Q[ϕ] in terms of a c-number valued function ϕ(x) as

A(x, y) is called an integral kernel. That’s the same sort of thing as (28.7), only now defined in function space. This
is a functional, a number-valued function that depends on ϕ(x).5 If I take these functions ϕ(x) and expand them in
a discrete basis, I will end up with an infinite-dimensional matrix, an infinite-dimensional vector, and an ordinary
number. I could go through all my integration formulas, and at least formally, they would make sense. Whether
they actually make sense would depend upon how astute I am in choosing the object A(x, y).

28.2Functional integrals in field theory

I will now explain a fundamental formula, which I will give in a form only partially defined, and which I will prove
later on. Take a theory of a single classical scalar field ϕ with non-derivative interactions L ′(ϕ), and a linear
coupling Jϕ with an external current J. Define the Lagrangian

and the classical action, Sc

This functional depends on two c-number valued functions, ϕ(x) and J(x). You give me a ϕ(x), you give me a J(x), I
can compute, with perhaps considerable labor, depending on how strange their forms are, the number Sc [ϕ, J]. In
the quantum theory, obtained from this classical theory by canonical quantization, there is an object Z[J] we have
seen before, (13.6),

where S is the S-matrix. It’s the generating functional for the Green’s functions. I will demonstrate, in a certain
sense which I will make precise,

N is a normalization constant independent of J, adjusted such that

N is closely related to the disconnected vacuum-to-vacuum graphs which we divide out in Dyson’s formula (see
(13.23) and the discussion following). The precise sense will be, for our purposes, to every order in perturbation
theory. When I expand out (28.27) in powers of L ′(ϕ), I will have a quadratic form in Sc (a Gaussian integral) times
polynomials in ϕ. I know how to do Gaussians times polynomials, and I will prove, order by order in perturbation
theory, that the right-hand side is equal to the left-hand side. I will have to make some subsidiary definitions to
prove it, but I will prove it.

The advantage of doing things this way is that on the left-hand side we have an object, Z[J], with all those
commutators that were giving us so much trouble. But on the right-hand side, the object has no quantum objects,
just classical fields which all commute with each other. They’re just ordinary c-numbers, and I’m integrating over
’em. This will turn out to be an enormous advantage, and enable us to settle with a single stroke all the problems
associated with derivative interactions, superfluous variables, gauge invariance and so on.

As it stands, it doesn’t look like the action Sc [ϕ, J], with the Lagrangian (28.24) is the sort of functional integral
we can do safely, even without worrying about the infinite dimensions of function space, and even for a free field
theory, L′ = 0. Instead of a nice positive definite quadratic form in the exponential, or at least a form with a positive
definite real part, we have an overall factor of i multiplying Sc : the exponential oscillates. Put simply, to make
sense of (28.27), we have to continue both sides into Euclidean space: to obtain four-vectors with real space
components, imaginary time components, and a quadratic form of definite sign. We discussed this earlier (§15.5)
in the context of analytic continuation of Feynman integrals. I first have to demonstrate that the Green’s functions
of a quantum field theory can be continued into Euclidean space. Then I will show that the functional integral is
well-defined in Euclidean space, (or as well-defined as our other functional integrals have been), and then I’ll be
able to show that the two sides are equal.

When we were doing loop integrations, we studied the propagator in the q0 plane and found that it had poles
in a typical case as shown in Figure 15.7; either (|q|2 −a) > 0 or (|q|2 −a) < 0. In either case, we did not cross any
poles if we rotated our contour of integration onto the imaginary axis:

leading to the Wick-rotated values

With these rotated time components, the Lorentz square k2 of a momentum four-vector kµ turned into a negative
Euclidean square kE2:

(We made this rotation after we had performed all the momentum shifts.) If we do this simultaneously to all the
external momenta in the problem, and at the same time that we rotate the external momenta in the complex plane
we rotate the internal momenta in the complex plane, to preserve energy-momentum conservation at every vertex,
this obviously goes through. We end up with a Feynman integral that has no zeros in its denominators anywhere.
Everything is the square of a Euclidean vector plus a positive mass squared. The function is not only well-defined,
it is an analytic function of the external momenta. When we rotate our external energies in the complex plane this
defines an analytic continuation in k-space.

On the left-hand side of (28.27), we have an expression in x-space. It’s pretty easy to see what we have to do
in x-space to keep things going right in k-space; we have to rotate x in the opposite direction:

The phase factor in x0 cancels the phase factor in k0; otherwise the Fourier transform would develop an
exponential blowup. The minus sign is going to be important in making our formulas come out right. So, to all
orders in perturbation theory, we can define our Green’s functions for Euclidean spacetime separations, where the
formal connections between the complex variable x0 and the real variable x4 are given above.

It’s also possible to give a direct position-space argument to demonstrate that everything can be continued to
imaginary time, without recourse to perturbation theory. That’s sufficiently amusing that I will give it. (See (13.25).)

We want to study a position space Green’s function, (13.22):

For convenience, I will assume

so that I can drop the time-ordering symbol; things are already time-ordered. Now explicitly pull out the time
dependence using the Heisenberg equations of motion:

I can now investigate what happens when I attempt to analytically continue these to imaginary times. It’s
convenient to introduce a complete set of energy eigenstates

and insert a complete set

between every pair of field operators; we get, for instance

Then

By assumption

so that, as we rotate the x0’s downward onto the lower imaginary axis, we get (−i) × (−i) in the exponent, a damped
exponential. If the sum converged when the exponential oscillated, with the factor of i, it will converge even better
as a damped exponential, with a factor of −1. In fact, this is not just a well-defined function of x but, because of the
marvelous exponential, an analytic function of x. No matter how many times we differentiate with respect to some
x4, although we get more powers of E, that terrific damped exponential keeps things from blowing up. Notice that if
we had tried to rotate in the other way, up to the positive imaginary x-axis, we would have gotten an increasing
exponential and would have become extremely nervous at this point.

So, the left-hand side of (28.27) is a Euclidean generating functional, a completely well-defined object. To get
an idea of what it looks like, let’s compute it first for a free field theory.

28.3The Euclidean Z0[J] for a free theory

We know how to compute the free theory generating functional Z0[J] for (28.24) when L′ = 0 in Minkowski space.
This theory is nothing but (8.57), our old Model 1 of the three models we considered in the early part of the course
(with the replacement ρ(x) → J(x) and setting g = 1). We found in §8.5 that

and so to within a phase

with α given by (S4.1) (see the solution to Problem 4.1, p. 177):

(recall *(k) = (−k)). Z0[J] is the exponential of the sum of connected graphs of the form

In the argument of the exponential, one i is from the propagator, two i’s come from the ’s and from the
combinatoric factor. It will be convenient to write this in position space

where ΔF(x − y) is the Feynman propagator (10.29) in position space,6

We go to Euclidean space by rotating x0 → −ix 4. The d’Alembertian operator □2 and the four-dimensional
measure d4x are transformed to their Euclidean forms (with a subscript E):

Continuing Z0[J] into Euclidean space, we get

in our compact notation, treating J(x) as a vector in an infinite-dimensional Hilbert space. (Note that there’s no
need for the iϵ in the Euclidean propagator; −□E2 is like +kE2, so that (−□E2 + µ2) is a positive-definite operator.)
This looks like what we get from a Gaussian integral; let’s check that it is.

Start by writing the argument of the exponential in the right-hand side of (28.27) in Minkowski space. We
rotate to Euclidean space and perform an integration by parts:

Because (−□E2 + µ2) is a positive-definite operator in Euclidean space, everything damps out nicely. We have a
formula for doing Gaussian integrals with a positive-definite matrix or operator, the extension of (28.11):
The argument (28.49) in the exponential of the functional integral (28.27) of is of exactly this form with

Then

The normalization constant N is chosen so that Z[0] = 1:

The constant N = det(−□E2+µ2) is divergent,7 but so what? The determinant’s divergence is a reflection of the
infinite zero-point energy of the free theory if we don’t normal order things. Then

the same as (28.48).

We have learned how to continue things into Euclidean space in a simple example, how to do a functional
integral of an interesting sort, and we have verified the assertion that (28.27) is valid in the simple case of a free
field. Although these integrals are really defined in Euclidean space, we will adopt a construction that treats them
as if they are defined in Minkowski space.

28.4The Euclidean Z[J] for an interacting field theory

Having verified (28.27) for the case of a free field, it’s trivial to verify the formula for the case of an interacting field.
We will show that our formula is precisely equivalent to Dyson’s formula (8.9):

where |0ñB denotes the bare vacuum and ϕ I is the field in the interaction picture. This is ordinary perturbation
theory before the application of Wick’s theorem. Using the trick (28.13), recast for functional integrals, we can
write this as

where

is the generating functional for noninteracting Green’s functions. (We needn’t worry about ordering of operators
within the differentiation; the time ordering takes care of that.) This is the left-hand side of the functional integral,
(28.27). Now for the right-hand side. Writing S0 for the free part plus the source term in the action,

the functional integral on the right-hand side can be written as8

 4
It’s the same trick as we used on the left-hand side. In one case, the exponential ei d xJϕ I is interpreted as an
operator inside a time ordering symbol; in the other case, the exponential eiS0[ϕ, J] is a function of c-number fields
inside an integral. But the trick works the same in either case. Equations (28.56) and (28.59) are the same if we
choose

Thus I have proved the startling assertion that I made earlier: that you can, for this particular kind of interaction,
represent the generating functional, the thing that tells you everything you want to know about the theory, in terms
of a functional integral.

We have been doing integrals blithely in Minkowski space. It is easy to rotate from Minkowski to Euclidean
space and back again. If there’s ever an ambiguity in the Minkowski space integral, we have to be more careful.
Such ambiguities can occur. If we had tried to do the free-field case directly in Minkowski space, we would have
encountered (□2 + µ2)–1, an ill-defined object. We would have to stick in an iϵ to make it a well-defined object.
There is nothing in our functional integral formulas to tell us if it’s +iϵ or −iϵ. The iϵ is introduced automatically by
continuing back from Euclidean to Minkowski space. We know which way we have to continue back on general
principles, and that puts in the +iϵ. The functional integral is properly done in Euclidean space where there is no
ambiguity in finding the inverse of the operator.

These formulas generalize to any theory of non-derivative interactions. Here are some simple extensions:

1. Set of scalar fields

(summation implied over repeated indices). U contains both mass terms and a non-derivative interaction. The
action is

and the generating functional is

2. Complex fields

As usual, we assemble real fields pairwise into complex fields:

This is not a big generalization, but it is convenient if L can be written in terms of complex fields.

3. Beyond Minkowski space

The formalism of functional integrals is not restricted only to four dimensions; we can have any integer
number of dimensions. In particular, it works for (quantum) particle mechanics in one dimension, say, an assembly
of harmonic oscillators with perhaps anharmonic interactions:

We can define the ground state to ground state Green’s function in the usual way. The S-matrix is

That’s the restriction from integrating over the four dimensions of a field theory to integrating over no space
dimensions and one time dimension for particles. The form (28.68) of the functional integral is in fact more general
than (28.27), because if a runs over an infinite set, say the Fourier components of the field, we can always think of
a field theory as a special case of a particle theory. The particle form only requires a restriction on how the time
derivatives of the fields enter L; how the space derivatives of the fields enter L is irrelevant. V may be a
complicated interaction between the Fourier components but it doesn’t involve time derivatives. We will use the
particle language when discussing derivative interactions next time, because it is more general.

28.5Feynman rules from functional integrals

We have found a functional integral representation for the generating functional, Z[J], which we have shown is
parallel to the original development from Dyson’s formula. Originally we went through a long journey from Dyson’s
formula to derive the Feynman rules: Wick’s theorem, diagrammatic representation of the terms in the Wick
expansion, we danced, we stood on our heads, finally we found the Feynman rules. I would like to demonstrate
that we can also get the Feynman rules by directly manipulating the functional integral, just to show you the utility
of functional integrals.

Recall our functional integral formula for Z[J], (28.56), in the case of a single scalar field for the generating
functional Z[J]:

I’d like to write out Z0[J] a little more explicitly than before:

The Feynman propagator ΔF(x − y) is just shorthand for what we called the contraction in our earlier discussion.
It’s the Fourier transform of the propagator in momentum space:

Let’s expand Z0[J]. We get zero propagators, one propagator with two J’s, two propagators with four J’s, and
so on:

As an example, graphically the fourth-order term of (28.71) is represented as

where there is a J attached to each end point and a Feynman propagator between end points.

Now for the other part of (28.56). The first exponential involves the expression

Every −iδ/δJ(y) will knock off an iJ(xk ), using

the generalization of . If for example the interaction is a simple cubic interaction,

then the first nontrivial term in (28.59) will involve

acting on the diagrams in (28.72). When this term hits the set of graphs above, three of the end points will have
the J’s removed. That results in two different sets of graphs. The first looks like this:

and those three free end points will join together into a three-particle vertex:

The whole thing will be multiplied by ig, and there will also be a combinatorial factor arising from the freedom of
deciding which end points are differentiated and joined together. The second set is

The three free ends are joined together to make a tadpole diagram:

The higher orders in the expansion of (28.56), the functional integral recreates the Feynman rules without the
need for any explicit normal ordering. For instance, in second order of L ′ we obtain diagrams including these:

and

So we get a string of terms that we can represent as propagators with J’s, and a second string that we can
represent as vertices. Differentiating with respect to the J’s will connect the free propagators in all possible ways to
make the diagrams. In higher orders in the expansion of (28.59), the functional integral recreates the Feynman
rules; this expression is Wick’s theorem, in a very compact notation. The operation of contraction becomes here
the operation of differentiation. Whatever the Gaussian is, it defines a propagator; the polynomial L ′(ϕ) defines
the vertices. Then we just stick things together according to the Feynman rules.

Suppose that I have, by dint of hard work, an expression of the following form (it generalizes trivially from one
field to many):

where A is any differential operator, S′ is any function of ϕ, with 47 derivatives in it and a non-local integral kernel
of 17 variables, just some horrible mess. I write down naive interactions corresponding to S′; there’s no time
ordering, because these are just classical c-number fields, and everything commutes with everything else. The
propagator D F is the inverse of A with the appropriate factor of i. It is the solution of this equation

the inverse operator written as an integral kernel of two variables, x and y. Any ambiguities that arise in inverting
the differential equation are to be resolved by continuing into Euclidean space where the functional integral is
really supposed to be defined; that will tell us where to put in the iϵ.

This equation (28.83) doesn’t mean much yet, because the only classical field theory for which we know how
to write the functional integral is the one for which we already know the Feynman rules. We will soon encounter
more complicated field theories where we do not know the Feynman rules, but for which we will nevertheless be
able to write the generating functional as a functional integral. Once we do that we can use (28.84) to find D F as
the inverse of A. We can forget the explicit formulas about integrating a Gaussian, or a polynomial times a
Gaussian, etc. We will naively read off the interactions from the functional integral: a derivative gives a factor of
momentum, and so on. This is one of the utilities of functional integrals. They will allow us to handle all the terrible
problems with derivative interactions and anything else. If we can only get the theory in the form of (28.83), then
we can just read off the Feynman rules. Never again will we have to worry about Dyson’s formula, Wick’s theorem
or anything else. This will be the magic method.

28.6The functional integral for massive vector mesons

As an example of how this method works its magic, I will use it in a case where we have not yet justified it. Let me
suppose it is true for the theory of massive vector mesons, a theory for which we can write the generating
functional as a functional integral. (I will prove that later on.) To give an example of this algorithm at work let me
assume I have proved it and I will attempt to construct the propagator. To avoid confusion between the operator A
and the vector field Aµ, I will write the vector field as Bµ. The propagator is given by the free action (26.47), with
Fµν given by (26.11):

I have used the antisymmetry of Fµν in going to the second line. The operator Aµν, which we have to invert as a
matrix differential operator, is

The matrix Green’s function for the differential operator Aµν is defined by

(The superscript P is for Proca.) If the Green’s function is ambiguous, if the problem does not have a unique
solution, we settle the ambiguity by adding an iϵ so we can rotate into Euclidean space. This is the prescription
generalized to a field with many components.

The solution to (28.87), like all differential equations with constant coefficients, is most easily found in
momentum space:

Recalling that

we get, applying µν to (28.88),

We have a 4 × 4 matrix equation that we have to invert. We break it up into the sum of two projection operators,
the transverse and longitudinal projection operators, PTµν and PLµν, respectively:

These have the usual properties of projection operators—their sum equals the identity, they’re idempotent, and
they’re orthogonal to each other:
Since µν is a linear combination of gµν and kµkν, it can be written as a linear combination of the transverse and
longitudinal projection operators:

Then it’s very easy to solve (28.90): we just invert the coefficients of the two projection operators, and multiply by i:

There is an ambiguity in the first term which I will resolve shortly by going into Euclidean space. We can easily
check that P
µν(k) is the inverse by multiplying by µν:

The propagator is

This is the correct expression for a massive vector field’s propagator. (I had to resolve the ambiguity when k2 = µ2
with an iϵ. This iϵ is also necessary if I want to rotate the propagator into Euclidean space: I have to avoid any
poles.)

Note that (28.97) has the same general form as we found in the case of a free fermion (21.73), or a free
spinless particle (10.29), a fraction formed from a projection operator divided by the particle’s momentum squared
minus its mass squared. For free spinless particles, the numerator was 1, which we can think of as the projection
operator onto the one physically allowed state. For the fermion we had + m, the projection operator onto the
physically allowed states. Here we have

which is nothing but the projection operator PTµν onto the three allowed transverse polarization vectors. All these
expressions have the same form.

Even after we derive the Feynman rules using this magic method, we will still have troubles. For one thing,
(kµkν)/µ2 does not appear to have a smooth limit as µ → 0. For another, (kµkν)/µ2 looks badly behaved as k → ∞;
maybe it spoils the renormalizability of the theory. It doesn’t, but we will have to check that.

Next time I will redeem this computation by showing that I can indeed write the generating functional for
electrodynamics as a functional integral, even when we have to eliminate degrees of freedom from L before we
can write the integral in Hamiltonian form. We will also consider scalar electrodynamics, so we’re going to have to
worry about theories with derivative couplings. And I haven’t said anything at all about using the functional integral
formalism for fermions. That will lead us into rather peculiar waters. For bosons, we write the functional integral
over Bose fields, that is, classical c-number fields, the limit of quantum Bose fields as ℏ → 0. This would lead you
to suspect, if you were bold at guessing, that for a theory with fermions, you write things as a functional integral
over classical Fermi fields. But what are classical Fermi fields? We will model these with Grassmann variables,
anti-commuting c-numbers,9 and we will learn how to do calculus with them.

1[Eds.]R. P. Feynman, “Space-Time Approach to Non-Relativistic Quantum Mechanics”, Rev. Mod. Phys. 20
(1948) 367–387, also reprinted in Feynman’s Thesis: A New Approach to Quantum Theory, ed. Laurie M. Brown,
World Scientific Press, 2005. See also Richard P. Feynman and Albert R. Hibbs, Quantum Mechanics and Path
Integrals, McGraw–Hill, 1965; edited and corrected by Daniel F. Styer, Dover Publications, 2010. Feynman
devised the technique to reformulate quantum mechanics. Much of this chapter restates, in different words,
Section 4 of Coleman’s 1973 Erice lecture, “Secret Symmetry”, reprinted as Chapter 5 in Coleman Aspects.
Copies of this lecture were handed out during this class.
2[Eds.]
The American mathematician Norbert Wiener (1894–1964), investigating Brownian motion, had developed
methods similar to Feynman’s about a decade earlier. For an illuminating article about his work and its
connections to Feynman’s, see Mark Kac, “Wiener and integration in function spaces”, Bull. Amer. Math. Soc. 72
(1966) 52–68. A brief introduction to function space, Lebesgue measure and generalized functions is given in
Chap. III, pp. 179–255 of Mathematics for Physicists, Phillipe Dennery and André Krzywicki, Harper and Row,
1967, republished by Dover Publications, 1996.
3[Eds.]
Square roots of 2 have already appeared in field theory, for similar reasons, when we decomposed a
complex field. See the digression in §6.1, p. 109. Note that for ordinary two dimensional integrals, the Jacobian
determinant J = ∂(z, z*)/∂(x, y), with z and z* defined as in (28.16), equals −i:

so that, to within a phase constant of norm 1, the identification (dz*)(dz) ≡ (dx)(dy) is unobjectionable.
2
4[Eds.]The function (1/ ) dt e− t is the error function, erf(x), with erf(x) = 1. See H. Margenau and
G. M. Murphy, The Mathematics of Physics and Chemistry, Van Nostrand Co., 1952, pp. 487–489.
5[Eds.]In the literature the dependence of a functional on its arguments is often in square brackets; see the
sentences following (13.6).
6[Eds.] ΔF(x − y) can be thought of as a matrix element of an operator which is formally the inverse of the
Klein–Gordon operator,

Making the usual replacement pµ → i∂µ, and inserting a complete set of momentum eigenstates |kñ, we find

the usual expression. ΔF(x) is sometimes written symbolically as −i(□x 2 + µ2 −iϵ)−1, as in the right-hand side of
(28.48). This notation really means the middle or right-hand side of (28.44).
7[Eds.] The calculation of det A = det(−□E2 + µ2) is given in Greiner & Reinhardt FQ, pp. 377–378, equation
(12.57). Using the identity det A = exp(Tr[ln A]), (see Arfken & Weber MMP, Chapter 3, “Determinants and
Matrices”, p. 224, equation (3.171) for a proof of this identity)

See also Problem 17.1, p. 679.

8[Eds.] Remember that everything here is a classical quantity so we don’t need to worry about commutators.
9[Eds.] See note 3, p. 434.

29
Extending the methods of functional integrals

We’re going to discuss four cute topics in functional integration that extend the method from Bose fields and non-
derivative interactions to Fermi fields, derivative interactions, and constrained variables.

29.1Functional integration for Fermi fields

So far, we’ve only dealt with functional integrals for bosonic systems, where all the dynamical variables, whether
quantum mechanical or classical, have commutation properties, and are represented by c-number fields, the
classical limits of Bose fields. To define functional integrals for fermionic systems, whose dynamical variables
have anticommutation properties, we need to introduce anticommuting c-numbers, also known as Grassmann
variables.1
 Introduce the anticommuting quantities

with the properties that any two of them anticommute:2

These imply that the square of any Grassmann variable vanishes:

As a result, the Taylor expansion of any function of Grassmann variables will have only a finite number of terms. In
particular,

because the square of any Grassmann variable is zero.

I will impose some conditions on the integrals of these quantities.3

1. Linearity:

The quantities α and β can be any constants (independent of η and ), even Grassmann numbers; the minus
signs are appropriate in the first integral if α and β are Grassmann variables. When we take α and β out of the
double integral, there is no change of sign in either case, since they are going past both dη and d .4

2. Translation invariance:

where ξ is another Grassmann variable. This is analogous to the translation invariance of the integral dxf(x);
all Grassmann integrals are definite, with limits unchanged by translation.

3. Normalization: To make things simple later, I impose the normalization condition

You may be surprised that I’ve chosen a normalization condition which, with ordinary numbers, would be an
increasing exponential rather than a decreasing exponential. But of course that’s irrelevant. Remember

so who knows what’s increasing and what’s decreasing?

I will now show that these three requirements taken together uniquely determine an integral of the form

The key point is that I can get a complete integral table for any function of η and defined by a power series. The
Taylor expansion of g(η) can only involve two terms:

because all higher powers of η vanish. Likewise, the Taylor expansion of f(η, ) can have only four terms,
each proportional to one of these four expressions:

Any other string of η’s and ’s will be zero if it has more than one η or more than one in it. So the
Grassmann integral table will have only four entries. Once I know how to integrate the set (29.10), I can integrate
any function f(η, ) defined by its Taylor series. So let’s determine the table.
 By condition 2,

Expanding g as a Taylor series,

By condition 1,

because ξ and B are arbitrary. By the same reasoning,

It follows

because the quantities in the brackets are independent of the integration variables. Finally, by condition 3,

Assuming dη η > 0, we get in fact two integral tables. For a single Grassmann variable,

and for a Grassmann variable and its conjugate,

The rule for integration of Grassmann variables is simple: it’s the same operation as differentiation.

We can now do more complicated integrals, such as a Gaussian. In this, a is an ordinary commuting number:

Contrast this with the result we would have with ordinary complex numbers:

In (29.19) we get an a; in (29.20) we get an a–1. This can be viewed as a consequence of the Taylor series
terminating after a finite number of terms. We will shortly give another explanation for this difference. The 2π is
irrelevant; that’s just because we’ve normalized the two integrals differently.

We are now ready to do a general Gaussian in many anticommuting variables, η1, , ηn. Consider the
quadratic form

The matrix elements of A are commuting quantities, and the set {ai} are A’s eigenvalues. (A may involve bilinear
forms of anticommuting quantities that are not variables of integration.) Now define the n-dimensional Grassmann
measure
so that

Here we get the determinant, detA. Recall (28.18),

The only difference between integrating Gaussians over anticommuting numbers and integrating Gaussians over
commuting numbers is that in the first case the determinant appears and in the second case the inverse of the
determinant appears. From this we can derive rules for integrating generalized quadratic forms, polynomials of
anticommuting c-numbers times exponentials, etc.

Now we take the same bold, slovenly step we made in the commuting case to go from a finite-dimensional
vector space over anticommuting numbers, where we have proved everything, and extend it blithely to an infinite-
dimensional space, i.e., a function space. With the integral defined, does the same formula (28.59) give us the
sum of all the Feynman diagrams for theories with Fermi fields? I will restrict myself to theories that are no higher
than second order in Fermi fields, just for simplicity; this will be sufficient for our purposes.

Consider a Fermi field; it could be a single field ψ or a multi-component field

For notational simplicity I will just call it ψ(x). Let ϕ(x) be a bosonic field. Assume

SB is the purely bosonic part. A(ϕ) is a combination of terms. It may involve the kinetic energy, ψ(i − m)ψ, and
functions of scalar or pseudoscalar or vector fields in the theory, ϕ or Aµ or ϕ 15 or whatever. As an example, we
could have

I want to consider the functional integral

to within some normalization constant. For the moment, we are only interested in the Fermi part of the integral,
and we can treat the Bose fields as fixed. (We can do the Bose integrals after we’ve integrated over the Fermi
fields.) This is the most general case. Does (29.27) equal what we would obtain, aside from a normalization
constant, from conventional Feynman perturbation theory? With that approach, we know how to get the vacuum-
to-vacuum amplitude: we would compute the connected Feynman diagrams, including the effects of whatever
Bose fields are in A, treated as external fields. The big dot represents whatever vertex is generated by A:

The overall minus sign arises because every one of these diagrams involves one and only one Fermi loop and
therefore we get exactly one minus sign from each loop. That is the sum of the connected Feynman diagrams.
The exponential of the sum of the connected Feynman diagrams is certainly the right answer for the vacuum-to-
vacuum amplitude, á0|S|0ñ. Is this the same as the fermionic functional integral?

It would be very nice if the question mark could be removed. It would make life simple in summing up all those
loop diagrams, provided you can compute the determinant of an operator in an infinite-dimensional space. Can we
do it without actually summing up the diagrams? Well, yes, we can.
 Consider the case where ψ is a Bose field; a field which has exactly the same dynamics, exactly as many
components, except we quantize it according to Bose statistics rather than Fermi statistics. That gives us a sick
field theory, by the spin-statistics theorem.5 Nevertheless, the Feynman rules for such a theory would be well-
defined, and the sum of the Feynman diagrams would be well-defined. In the Bose case, we don’t have the Fermi
minus sign so we would get ((28.50) and (28.18))

That’s on the functional integral side. On the diagrammatic side we would get exactly the same exponential of
exactly the same sum of diagrams, except we wouldn’t have the Fermi minus sign. Otherwise everything is the
same:

There are arrows on the Bose fields because they are still charged. The functional integral gives the right answer
for bosons, assuming there are no derivative interactions, which we are assuming for the moment; we’ll take care
of that case later. Therefore (detA)–1 must equal the sum of the diagrams. But if (detA)–1 is equal to the sum of the
diagrams, then if I stick a minus sign in the sum we get

in the Fermi case, QED. The fact that an A appears in the Fermi case and an A–1 appears in the Bose case is
merely a reflection of the Fermi minus sign for single fermion loops, which we obtained earlier from the
complicated combinatorics of anticommuting Fermi fields.6

So far we’ve only shown that one horrendous expression is equal to another equally horrendous expression,
but not what it’s good for. When we start manipulating things we will see how useful it is.

29.2Derivative interactions via functional integrals

We now enter the darkest part of these lectures. Our proofs will be both complicated and inadequate; all the
fiddling detail of pure mathematics, but with none of its generality and rigor. I will restrict myself to the one case
where the results I’m about to state have been carefully proven. I’ll give a proof shortly but the proof should be
considered to be between two very large quotation marks. Its combinatoric complexity will be matched only by its
lack of rigor. I will consider a classical Lagrangian in particle, not field, language (sum on repeated indices):

It is no more than quadratic in derivatives so we can get the Hamiltonian without having to solve a quadratic or
higher order equation to determine the p’s:

Therefore the Hamiltonian is

I will now state a result that I will first exploit and then come back and prove: The Lagrangian form of the
generating functional is given by the expression7

That peculiar factor, 1/2, is a surprise. It is not going to give us any problems if the only derivative interaction
is linear in the derivative, but it will give us problems if there are terms quadratic in the derivatives. This factor is
less of a surprise if we write the action integral in Hamiltonian form:
These two are equal for solutions to the equations of motion. I would like to consider the right-hand term as a
function of p and q, regarded as independent quantities, defined over a larger space of functions than the left-
hand side. The term on the left is defined for arbitrary motions in q-space; the term on the right is defined for
arbitrary motions in phase space, with twice as many dimensions. With this form of the action,

This is the Hamiltonian form of the generating functional.8

I claim that the Lagrangian and Hamiltonian forms of the generating functional are equal. H is at most a
quadratic function of the p’s, so with q held fixed the p integral is simply a Gaussian. We know how to evaluate a
Gaussian: find the minimum of the expression and plug the answer back in;

Putting this back in recreates the Lagrangian form of the action, by putting in the Hamilton equation that gives us
pa as a function of the a’s. However, because the coefficient of the quadratic term in (29.35) is not a constant, we
have to insert the determinant, to the 1/2 power, of the coefficients of the quadratic form, and that introduces the
1/2 coming from the A–1 in H. The equation (29.36) is just the Gaussian integral in p-space coupled with

So the Lagrangian and Hamiltonian forms of Z[J] are equal. I still haven’t “proved” that the Hamiltonian form
actually gives us the generating functional, and so for the moment I ask you to take (29.36) on faith. We will
manipulate (29.36) and show some of its consequences, make some comments about it, and after we’ve done a
few examples we’ll go back and prove it.

Two points should be made. First, the standard canonical methods of constructing a quantum theory from a
classical theory suffer from ordering ambiguities: there are many ways we can order the p’s and q’s if H contains
forms like (29.40). But the expressions above have no ordering ambiguities whatsoever. That is, the functional
integral procedure defines one among many possible ways of ordering any given Hamiltonian of the stated type.
We could figure out what way that is, but we won’t in this course. But there is a way of ordering.9 Second, the
Hamiltonian form of the functional integral, although very useful (as you’ll see), must be taken cum grano salis,
because, unlike the Lagrangian form, it does not become well-defined when we rotate into Euclidean space. The
pa a term integrated dt picks up an i when dt → idt, but we get a compensating i from a, so it continues to oscillate
even in Euclidean space, instead of becoming a damping factor. Indeed, it is possible to show that for a sufficiently
complicated H, the answer we get depends on whether we integrate over the p’s first or over the q’s first. It is not a
well-defined, uniformly convergent integral. It converges, even in Euclidean space, only because of cancellations
of phases, and therefore the order in which we integrate over phase space may be important. We will ignore this
point, which concerns only purists. If you are a purist, when you write it in Hamiltonian form (29.38), you must add
a little footnote that says “integrate over the p’s first.” Once we integrate over the p’s, we obtain the Lagrangian
form of the path integral, which does become well-defined after we rotate into Euclidean space.

Integrating over the q’s first instead of the p’s first gives a different ordering for the Hamiltonian; the two
orderings are not consistent. Eventually, when we go to a high enough order in perturbation theory we’ll find
different expressions for Green’s functions, depending on the order of integration. As a simple example, consider
the Hamiltonian

This is a quadratic form in the q’s with fixed p’s and a quadratic form in the p’s with fixed q’s. So we can do either
the p integral first or the q integral first explicitly and then be left with a mess, which we can evaluate perturbatively.
The two prescriptions begin to differ at higher orders in λ.

In fact, most of our work will be done with the Lagrangian form of the functional integral, (29.36), with the
explicit determinant sitting out there in front. That’s a very nice form. But it has one deficiency: we can’t evaluate it
using the Feynman rules, because it’s not the integral of an exponential. This problem can be eliminated by
introducing extra fields that exponentiate the determinant. I will now explain how this is done.
 29.3Ghost fields

We have the expression

We need to get (detA)1/2 up into the exponential. To do this I will introduce new fields, so-called ghost fields,10
Fermi fields, ηa and a, that put the determinant into the exponent, as in (29.27). They do not correspond to any
dynamical degrees of freedom that are actually in the system; hence we give them the pejorative name ghosts.
They have no physical interpretation. Given the action S, we define the effective action

Then the functional integral becomes

The ghost variables (they are not fields here, but we’ll soon look at ghost fields) are just things we have stuck in to
move the determinant from out in front, where we can’t do anything with it, up into the exponential where we can
evaluate it by the ordinary Feynman rules.

EXAMPLE. Change of variables for a scalar field

Let’s do an example, cooked up so that we know the right answer in advance. Apart from the source term, it’s
a free field theory. That way we can check to see if all these prescriptions are correct.11

The source term is there because we’ll be talking about Green’s functions, not just S-matrix elements. Change the
variables from ϕ to A:

The quantity g is a free parameter, which I will treat as a coupling constant. The physics is completely unaffected
by this substitution. It was chosen for its algebraic simplicity. This transformation is not invertible, but that’s
irrelevant. We will be doing perturbation theory and the perturbation series is formally invertible. The Lagrangian
becomes

This is the same theory we started with, though it looks like a horrible interacting theory. All the Green’s functions
obtained by computing Z[J] and functionally differentiating with respect to J should be the same for the two
different forms of L. It is not obvious how this will work out; it looks as though there are cubic interactions, quartic
interactions, and derivative interactions, all governed by g. All that must cancel. But we won’t see it cancel if we
naïvely read the Feynman rules off the second form of L.

To get the Feynman rules we must look at an effective Lagrangian which involves a single ghost field, η(x):

Following (29.42), (1 + gA) is the square root of the coefficient of the derivative term in the Lagrangian (29.47).
You see the unphysical ghostly nature of η and if you look at their propagators, using our standard rules. The
propagator is always i divided by the coefficient of the quadratic term. So we have, in addition to our normal fields,
a propagator for the ghost field, denoted by a dotted line:
The ghost field η(x) is unphysical in two ways: its propagator has no momentum dependence, and it is a spinless
field obeying Fermi statistics. But it’s got to be there to make everything come out right.

We can show that everything comes out right if we include the ghost fields, but not otherwise. The simplest
possible Feynman calculation that shows this is the computation of Z[J] to first order in J and to first order in g.
This is the one-point (tadpole) function, which should vanish since it vanishes in the original Lagrangian.12 (It
should vanish to all orders in g, but that’s beyond my computational abilities, and I suspect beyond your patience.)
There are five terms of O(g) and O(J) in the effective Lagrangian:

These give rise to four graphs of first order in g and J:

All four are quadratically divergent—they go as d4k/k2—but we will sum them up nonetheless. The sum had better
come out to be zero.

There are some factors common to all four diagrams. There is an overall i from the interaction Lagrangian.
There is a factor of g for the interaction. There is a (0), evaluated at k = 0 since the propagator lines carry zero
momentum: the momentum at the loop vertices is always zero, and momentum is conserved at the vertices. And
there is an integration d4k/(2π)4 over loop momentum for each of the loops. After factoring out the common
elements, we have:

1.In the first graph, one of the three A’s at the vertex contracts with the A coupled to the source, to
produce a zero-momentum propagator, i/(−µ2); the other two contract to give the loop, and a meson
propagator of i/(k2 − µ2 + iϵ). There is also a vertex which gives a factor of ig, but we’ve already
factored out the g. Because the three A’s at the vertex are indistinguishable, there are three choices
of which of them is contracted with the source term A. This gives a net contribution of

2.In the second graph, the derivatives on A give terms of ±ik µ times a creation or annihilation operator.
Since the meson line has momentum 0, the undifferentiated A at the vertex must be the one
contracted with the A from JA, giving once again i/(−µ2). If the differentiated A were contracted with
it, we would get 0. The two (∂µA)’s are contracted together, giving (ik µ)(−ik µ) = k2. And there’s a
factor of i from the vertex. The second diagram contributes

3.The third graph is pretty simple. It has only one interaction, which we’ve already taken out. So all we
have is the propagator for the internal loop times an explicit factor of . We just have to contract the
two fields with each other, and we get

4.The contribution from the fourth graph looks peculiar. There is a zero-q A propagator i/(−µ2); there is
an interaction, i, and the combinatorics for the vertex gives 1, there’s no choice there. There is a
ghost propagator around the loop, i. And finally there is a minus one for the loop because the ghosts
are fermions. That gives

The sum of the four graphs is , where

The expression in the curly brackets can be simplified:

combining the last two terms with common denominators. In the limit ϵ → 0, this is zero, the right answer! Please
notice the absolute necessity of the ghosts, without which we would not have gotten this to work out. It’s the fourth
term, with its fermion loop minus sign, that makes everything cancel. We are saved by the friendly ghosts.

29.4The Hamiltonian form of the generating functional

Now let’s “prove” the Hamiltonian form of the functional integral, (29.38). We’ll do the Hamiltonian form rather than
the Lagrangian, so we don’t have to worry about constructing the Hamiltonian out of the Lagrangian. We want to
show that it’s equal to Dyson’s formula (7.36) and Dyson’s formula is given in terms of the Hamiltonian.

For simplicity consider a single scalar field.

(The argument goes through without alteration, aside from a proliferation of indices, if there are many fields of
various spins.) The field K, a source coupled to the canonical momentum π, will eventually be set to zero, but the
term Kπ will be useful at intermediate stages. Dyson’s formula, universally valid, tells us that

where the subscript I indicates the interaction picture. We use the proportionality symbol to avoid complications
from N, N′, etc. Now we see the advantage of introducing the source K: swapping the fields for the derivatives

we can take the H ′ outside and write this as

(My conscience tells me I should tell you when I’m cheating. This is a swindle: δ/δJ and δ/δK are commuting
operators, but ϕ I and πI are not. Thus we have chosen some (unspecified) ordering. We will ignore this problem.)
This takes care of Dyson’s formula.

Now let’s look at the functional integral, which I want to show is equal to Dyson’s formula.

with H given by (29.53). Again we pull out H ′, and obtain

Comparing (29.56) and (29.58), we see the same operator H ′(−iδ/δJ(y), −iδ/δK(y)) acting on two different
expressions. We will prove the theorem in general if we can prove it for a free field theory, augmented with J and K
sources. So, are these two expressions equal?

The π integral on the right-hand side is a pure Gaussian (28.50), with determinant 1. The quadratic form is
with minimum given by

The quadratic form at the minimum is

so the Gaussian π integral gives

Note that the last term in the exponential has no dependence on ϕ, so it can be taken out of the functional integral.

Now let’s evaluate Dyson’s formula:

Both ϕ I and πI are free fields, each with its own propagator. The only kind of diagrams we will get will be ϕ I–ϕ I
contractions, πI–πI contractions as well as the joint propagators, ϕ I–πI contractions. Thus we get

The ’s are because we can’t tell the vertices ϕ I(x)ϕ I(y) apart from ϕ I(y)ϕ I(x), and the same holds true for the
quadratic term in πI. No such symmetry factor is needed for the ϕ IπI term. Both ϕ I and πI are free fields, linear in a
and a†, so by Wick’s theorem, their contractions are all c-numbers. The ϕ contraction is straightforward: it is the
Feynman propagator.

In the interaction picture,

To evaluate the contraction involving ϕ I and πI, use the identity (27.86)

Because (3.60) [ϕ I(x, t), ϕ I(y, t)] = 0,

so

To evaluate the contraction involving two πI’s, use the identity (27.86) again,

from (3.61). Then

We encountered a similar expression in (27.91). There, the first term was nicely covariant, but the second was
disgustingly non-covariant, and something we didn’t want. Here, though, it will be welcome.

Now the moment of truth:

which agrees with (29.63). It works! This completes the demonstration that the functional integral really does give
us everything, including the messy exp{ d4x(K(x))2} term that came out of the Hamiltonian form of the functional
integral. It gives us the right result for a free field theory and therefore, if we accept this sloppy proof, it gives us the
right results for an interacting theory. QED

There is one other useful rule for handling functional integrals. It’s quite simple, and does not involve anything
like these hairy complications. It will enable us to describe electrodynamics for a massive photon, including scalar
electrodynamics. (We need new physics to take care of gauge invariance for real photons.)

29.5How to eliminate constrained variables

Sometimes we encounter in Lagrangian systems dynamical variables for which the Euler–Lagrangian equations
are not equations of motion but are simply equations of constraint. They tell us something about the initial value
data but not how things develop in time. An example is the time component A0(x) in vector field theory. There are
no terms involving ∂0A0 in L, so the A0 equation of motion is simply an equation of constraint:

We frequently encounter Lagrangians of the form (written in terms of a single particle, for simplicity)

The variable y appears quadratically with a coefficient a independent of q. Its time derivative ẏ does not appear in
L. Thus y is a constrained variable, not a dynamical variable. We must solve for y to eliminate it from the theory
before we can obtain the Hamiltonian. The Euler–Lagrange equation for y is

Thus we obtain for the Lagrangian

This is the Lagrangian we have to use when working with the Hamiltonian form. The (almost trivial) point I want to
make is this: Aside from a normalization factor,

This is just our old rule (28.11) for doing a Gaussian. We evaluate the Gaussian at its minimum, which is precisely
this prescription. (We pick up an irrelevant constant because a is independent of q.)

So if we have a constrained variable that enters the Lagrangian L at most quadratically and the coefficient of
the quadratic term is independent of the other dynamical variables, we might as well integrate over it, which is
equivalent to eliminating it. The two prescriptions are the same. If a does involve the other dynamical variables,
then integrating over it isn’t the same as eliminating it: a determinant will appear and we can handle that
determinant in the usual way, with ghost fields. That won’t occur in the cases we have at hand.

This procedure is easily extended to many variables, or a field. For example, suppose the Lagrangian for a
system of particles is

The conditions

fix the yi as functions of the q’s and the ’s, allowing us to eliminate the y’s. If the Lagrangian is of the form

with Aab independent of the q’s and the ’s, then functionally integrating over the ya is equivalent to eliminating
them. If the Aab depend on the q’s and the ’s then we introduce a set of ghost fields {ηa, a}:

29.6Functional integrals for QED with massive photons

Now we can apply these two rules (the simple rule for elimination of constrained variables and the rule for the
Hamiltonian form of the functional integral) to extract and derive the complete Feynman rules for, first, spinor
electrodynamics with charged spinor particles and, second, scalar electrodynamics with charged scalar particles,
in both cases with a massive photon. The trick is just writing the same damned action in 14 different ways
(actually, only in two different ways).

Let’s begin with spinor electrodynamics. We already know a form for the action

This is called the second-order action because it is quadratic in the Aµ derivatives. There is an equivalent form
with many more constrained variables. This is called the first-order action because it only involves derivatives to
the first power.

Fµν and Aµ in this form are to be considered completely independent dynamical variables in the Lagrangian sense.
The equation of motion for Fµν is

Plugging this into L, a necessary step to get H, we arrive back at the original form, (29.82).

Now the game goes like this. By our rule for eliminating constrained variables

just by eliminating the six components of Fµν, since the quadratic term has constant coefficients. On the other
hand, we could equally well choose to eliminate some other variables from the theory. In particular, we could
choose to eliminate Fij and A0, leaving F0i and Ai. The Fij2 and A02 terms have constant coefficients, and the Fij
∂iA0 cross-terms involve no time derivatives. So these four variables {Fij, A0} follow the rule for constrained
variables just as the other six, {Fµν}, and we could write the same integral, just by choosing to integrate over a
different bunch of variables first, as

But what is Sother? Well, we’ve done this step before, in §26.3, and we found (26.48). If we eliminate Fij and A0 to
write the action in terms of Ai and F0i only, we get the Hamiltonian form SH of the action, because {Ai, F0i} are the
canonical (q, p) pairs of the massive vector field:

and so we can write

Now the argument is complete: two things equal to the same thing are equal to each other. What S1st is good for
is to allow us to go from SH (the Hamiltonian form we get from canonical quantization but which is awful for
Feynman rules) to S2nd (which is not what we get from canonical quantization but which produces covariant-
looking Feynman rules quite easily). The functional integral makes the change of variables easy. Each of these is
the generating functional. From the second-order form

with appropriate source terms, we can derive the Feynman rules by just reading off the propagators and the
interactions from S. The task is done; I’ve derived the Feynman rules for massive electrodynamics with a massive
photon.

We begin to see the utility of the functional integral formalism. The point is not that it’s particularly easy to
evaluate a functional integral; the rules for evaluating a functional integral are just the Feynman rules. The point is
that the functional integral is particularly easy to manipulate. Here we’ve gone from a formalism with 10
independent variables (six F’s, four A)’s to a formalism with four independent variables (four A’s) to a formalism
with six independent variables (three A’s and their conjugate momenta), and we do it just by writing down the
equations. No fancy unitary operators, canonical transformations, etc. The functional integral with the Hamiltonian
form (SH ) of the action is always right; that’s our general theorem. The functional integral with the second-order
action S2nd gives us the naive Feynman rules. SH is trustworthy; S2nd is useful; and they’re equivalent. Therefore
the naïve Feynman rules are right.

I will complete the analysis next time by using the exact same trick to handle the case where there are
charged scalar fields. We could do it by elementary methods, but it would be a terrible mess. Not only do we have
there the problem of the A0–A0 contraction, which we’ve taken care of without ever worrying about it, we’ve got
derivative interactions, momentum-momentum contractions on top of the A–A contractions, A’s with π’s, and it
would just look awful. It can be done; it was originally done that way, without functional integrals. It would take us
a week to do that one problem right. Next time I will do it in two minutes. And then I will do some sample
computations dealing with massive photon electrodynamics.

1[Eds.]See footnote 3 on p. 434. Here, “c-number” means “classical number”, as opposed to “q-number”, or
“quantum number”, typically an operator of some sort. The nomenclature is due to Dirac: P. A. M. Dirac, “Quantum
Mechanics and a Preliminary Investigation of the Hydrogen Atom”, Proc. Roy. Soc. 110 (1926) 561–579; the terms
appear on p. 562. Initially “c-number” was meant to imply a commuting number, so the concept of an
anticommuting c-number seems a contradiction in terms; as Coleman remarks (Aspects, p. 156), “Anticommuting
c-numbers are notoriously objects that make strong men quail.”
2[Eds.] The bar, suggesting a Dirac adjoint (20.62), is a little misleading; these objects are eventually going to be
scalar fields that obey Fermi statistics! It would be less confusing to write them as η and η*, but that is not
Coleman’s notation, which we follow here. (The bar may help remind the reader that η and , like Fermi fields,
anti-commute.)
3[Eds.] For an accessible overview of the algebra and calculus of Grassmann variables, see L. D. Faddeev and A.
A. Slavnov, Gauge Fields: Introduction to Quantum Theory, Benjamin/Cummings, 1982, Section 2.4, pp. 49–55.
For more about integration of Grassmann variables, see Peskin & Schroeder QFT, pp. 299–300. The definitions
dη = 0, dη η = 1 are due to Berezin: F. A. Berezin, The Method of Second Quantization, Academic Press, 1966,
p. 52, equation (3.10).
4[Eds.]Coleman defines the integration measure in the opposite order, d dη, which differs by a minus sign from
this choice. This will affect how some functions of the Grassmann variables are represented. One reason for
reversing the order is that, for Grassmann variables, integration is the same operation as differentiation. With the
order dη d , we have

Using Coleman’s order, the derivatives give −1. In a moment he will use this integral (29.7) as a normalization
condition.
5[Eds.]
W. Pauli, “The Connection Between Spin and Statistics”, Phys. Rev. 58 (1940) 716–722; reprinted in
Schwinger QED.
6[Eds.] See §21.5.
7[Eds.]
On p. 184 of Aspects, in Note 21 Coleman cites K. S. Cheng, “Quantization of a General Dynamical
System by Feynman’s Path Integration Formulation”, J. Math. Phys. 13 (1972) 1723–26 for the derivation of
(29.36). Starting from the Lagrangian gij i j, the measure dqi, reminiscent of the general relativistic
measure d4x, is simply written down, to keep the Lagrangian invariant under general coordinate
transformations on the qi; from it, a different result is obtained.
8[Eds.] See Appendix B of Richard P. Feynman, “An Operator Calculus Having Applications in Quantum
Electrodynamics”, Phys. Rev. 84 (1951) 108–128; and L. D. Faddeev, “The Feynman Integral for Singular
Lagrangians”, Theo. Math. Phys. 1 (1969) 1–13. Feynman’s equation (16-a) is Faddeev’s equation (1).
9[Eds.]For an extensive review of the problem and some solutions, see S. Twareque Ali and Miroslav Engliš,
“Quantization methods: a guide for physicists and analysts”, Rev. Math. Phys. 17 (2005) 391–490.
10[Eds.] Ghost particles, or fictitious particles, the quanta of ghost fields, were introduced by Feynman to solve
some problems encountered in quantizing gravitational and Yang–Mills fields (see footnote 5, p. 646) at the one
loop level: R. P. Feynman, “The Quantum Theory of Gravitation”, Acta Phys. Polon. 24 (1963) 697–722. Reprinted
in Selected Papers of Richard Feynman, Laurie M. Brown, ed., World Scientific, 2000. DeWitt, who extended
Feynman’s idea to multiple loops, describes ghost particles’ role this way: “The fictitious particles play a
compensating role, canceling the effects around the closed loops of the non-transverse [gravitational and
Yang–Mills] field modes … Their presence is central to the preservation of the unitarity of the S-matrix and to the
complete invariance of the theory under group transformations, as well as changes in supplementary conditions.
In principle, they are needed even in electrodynamics. However, in that special case, the vertex [between the
ghost particles and the photon] vanishes, owing to the Abelian character of the gauge group.” (See the box on p.
1042, item (h); for an Abelian theory, cabc = 0.) B. S. DeWitt, Dynamical Theory of Groups and Fields, Gordon and
Breach, 1965, p. 227. (Also published as part of Relativity, Groups and Topology (Les Houches 1963), Gordon
and Breach, 1964; p. 812.) See also B. S. DeWitt, “Quantum Theory of Gravity II. The Manifestly Covariant
Theory”, Phys. Rev. 162 (1967) 1195–1239 and “Quantum Theory of Gravity III. Applications of the Covariant
Theory”, Phys. Rev. 162 (1967) 1239–1256. Ghost fields are usually associated with the names of Faddeev and
Popov; see Chapter 31, and note 8, p. 1036.
11[Eds.] See Problem 8.1, p. 309. The field A is not to be confused with the matrix A in (29.42).
12[Eds.] See also Coleman Aspects, Ch. 4, “Secret Symmetry”, Section 4.6, pp. 158–159.

Problems 16

16.1 A massive vector meson (with the standard Proca free Lagrangian) is minimally coupled to a Dirac particle,
with coupling constant e. Compute, to lowest nontrivial order in perturbation theory, the amplitude for elastic
fermion-antifermion scattering and explicitly verify that the contribution of the term in the vector meson propagator
proportional to kµkν/µ2 vanishes. (You are not asked to do spin sums or compute cross sections, or even to
simplify the amplitude any more than is needed to demonstrate the desired result.)
(1998b 3.1)

16.2 (a) In the theory of the previous problem, compute the amplitude for elastic vector-spinor scattering, again to
lowest nontrivial order. Verify that if the vector meson spin vector, εµ, is aligned with its four momentum kµ, for
either the incoming or the outgoing meson, the amplitude vanishes, even when the meson in question is off mass
shell (but the other particles are on mass shell). (Parenthetical remark as above.) Of course, what you are
verifying is that Aµ has vanishing divergence between initial and final states defined by the on-shell particles.

(b) The same problem, but this time with a scalar particle rather than a Dirac particle. Use the naïve Feynman
rules (see the table on p. 644). But be sure to include the “seagull” diagram; otherwise you won’t get the right
answer.
(1998b 3.2)

16.3 Consider a theory of two charged Dirac fields A and B with masses mA and mB, and a complex charged
scalar field C with mass mC. These interact with a Yukawa-like coupling

where g′ = geiϕ , and g is a positive (real) number. (The phase of g has been absorbed into the definition of C.)
Now consider this theory minimally coupled to a massive “photon”, with the three fields having charges (in units of
e) qA, qB and qC with qA = qB + qC.

Scalar meson photoproduction, γ + A → B + C, first occurs in order eg. As in the preceding problem, show that the
amplitude for this process in this order vanishes if the “photon” spin is aligned with its 4-momentum. (Note that you
have to sum three graphs.)
(1998b 3.3)

Solutions 16

16.1 The lowest-order diagrams for fermion–anti-fermion scattering are shown below. The amplitude is

The relative minus sign between the two terms is due to the exchange of external fermion lines. We note that

Thus any term proportional to (p − p′)µ(p − p′)ν or (p + q)µ(p + q)ν vanishes. The amplitude simplifies to
This problem is very similar to the EXAMPLE on Coulomb scattering in §30.3, p. 646.

16.2 (a) The lowest-order diagrams for vector–spinor scattering are shown below. The amplitude is

Let the incoming meson’s polarization vector be aligned with its momentum: εµ(3) = λkµ, for some constant λ.
Then

We use the same trick as in (30.40). Since ( − m)u(1) = (2)( ′− m) = 0, we can rewrite (S16.6) as

By similar arguments, we can also show that iA = 0 when εν*(4) = λk′ν.

(b) The Feynman rules for vector-charged scalar interactions are given in the box on p. 644. The three relevant
diagrams are shown below:

The amplitude is

(the seagull term has an extra factor of 2 due to the combinatorics of two identical A’s). When εaµ = λkµ,

Though the vector isn’t on the mass shell, the scalars are: p2 = (p′)2 = m2. Using these constraints,

Plugging these expressions into the appropriate denominators,

By similar arguments, we can also show that iA = 0 when

16.3 There are three ways the reaction A + γ → B + C can occur, as shown below.

In graph (a), A can absorb a photon, and then decay into B and C. This process contributes to the amplitude iA a
term

Alternatively, as shown in graph (b), A can decay into B and C, and B can absorb the photon, contributing a term

Finally, as shown in graph (c), A can decay into B and C, and C can absorb the photon. This gives

The amplitude is the sum of these three terms. When εaµ = λkµ, we obtain

We use the same trick as before: since we can add zero terms to the factors of in the
amplitude to rewrite the factor in the brackets as

Because ℓ + q = k + p, k = ℓ + q − p, and

so that

We have to have qA = qB + qC, to conserve charge at the ABC vertex.

30
Electrodynamics with a massive photon

Last time we used the technique of eliminating constrained variables to prove an important theorem in
electrodynamics: that all three forms of the functional integral, the first-order form, the second-order form and the
Hamiltonian form, are equal. From this we showed that the naive Feynman rules are valid for spinor
electrodynamics with a massive photon. The theorem is also true for the electrodynamics of charged scalar
particles interacting with a massive photon; let’s demonstrate that.1

30.1Obtaining the Feynman rules for scalar electrodynamics

If we attempted to treat scalar electrodynamics canonically we’d be in a terrible mess. We’d have all the problems
associated with the derivative interactions of the scalar field2 and the problems associated with the elimination of
A0 from the vector equations of motion3 They are problems that can be handled, but they are messy. Here the
functional methods pay off: we have only to write down three horrible equations, instead of many more.

We’ll use the same trick as we used for spinors. We begin with the second-order form of the Proca
Lagrangian (26.10), and the minimally coupled charged Klein–Gordon Lagrangian4 that follows from (27.59):

where, from (27.48),

We can also write the theory in first-order form (with an action that involves no more than first derivatives):

This is a mess but it’s a simple generalization of what we did with spinors, going from (29.82) to (29.83). The
Euler–Lagrange equations of motion for Fµν, πµ, and πµ* are trivial:

S1st has been cooked up to generate these equations. When we substitute them back into (30.3), we get the
conventional form (30.1) of the Lagrangian. The second-order form has second derivatives, with no trace of Fµν,
and πµ as independent dynamical variables. We have discussed this process before (§29.5). It involves searching
for the minimum of a quadratic form. Indeed, we see that the quadratic terms in the variables to be
eliminated—FµνFµν and πµπµ*—have only constant coefficients. Thus, just as before,

where S2nd is the action in Lagrangian form, a function only of the variables displayed in the integration measure.

We could however choose to eliminate a different set of constrained variables: A0, Fij, πi and πi*. We can do it
by our trick if these terms enter the Lagrangian no more than quadratically, and if the coefficients of the quadratic
terms are constants. None of the terms gives us problems since these are quadratic and a fortiori quadratic or
linear or constant in the variables we wish to eliminate. The only terms that could give us problems are πµ*(∂µ +
ieAµ)ϕ or πµ(∂µ − ieAµ)ϕ *. There are no cross terms of any kind between any of the constrained variables. These
terms do involve A0 but only multiplied by π0* or π0 which we are not intending to eliminate, and likewise πi or πi*,
but only multiplied by Ai, which we are keeping. So ieπµ*Aµϕ and −ieπµ*Aµϕ * are both linear in the terms we are
going to lose upon substitution. We are left with F0i, π0, π0*, ϕ, ϕ * and Ai. These are precisely the q’s and p’s, the
fields and their conjugate momenta, of the Hamiltonian formulation. Writing the action in terms of these variables is
what it means to write the action in Hamiltonian form, as SH, in terms of the independent variables of Hamiltonian
dynamics:
 Table 30.1: Hamiltonian variables for scalar electrodynamics

The last form is the functional integral in the Hamiltonian form, so it is guaranteed to give us results equivalent to
canonical quantization and Dyson’s formula. All three forms are equal. The middle form gives us the naïve
Feynman rules. Therefore, just as in charged spinor-massive photon theory, the naive Feynman rules are true.
Every derivative interaction is simply a factor of pµ, etc. We have redeemed the guess (14.57), as promised at the
beginning of Chapter 28. We just read the Feynman rules off the Lagrangian, and the problem is solved.

Once we get the basic trick, the demonstration goes very fast. There are no worries about pulling time
derivatives through time-ordered products. We solved that problem once for a free field (see (29.66) through
(29.71)), but that was the only time we had to solve it, by dint of proving our general theorems when we talked
about J’s and K’s. From now on, it’s going to happen automatically; this formula is going to take care of everything.
It is the same formula that works in quantum electrodynamics.

We will go through this line of reasoning again for gauge fields. We’ll write the functional integral in two forms,
obtained from each other by the most trivial of manipulations in functional integration language, although rather
difficult manipulations if we attempt to do it in operator language. One form manifestly gives the right generating
functional because it is the Hamiltonian form, but it is difficult to try to derive Feynman rules in this language—it
doesn’t even look covariant. We have π0 but not πi, F0i but not Fij. The other form looks nice and covariant and
has nice simple Feynman rules. The two functional integrals are equal. In the Hamiltonian form it’s easy to show
everything is OK; in the Lagrangian form it’s easy to derive Feynman rules. This has been a rather abstract
stretch, so perhaps it’s time to do some specific computations.

30.2The Feynman rules for massive photon electrodynamics

I’ll give the Feynman rules for massive electrodynamics with both scalar and spinor interactions, and then some
low order computations with spinor electrodynamics, a somewhat simpler theory than scalar electrodynamics.
There will be problems on scalar electrodynamics in the next homework (Problems 16).

Scalar electrodynamics with a massive photon

We’ve established that we can read off the Feynman rules from the second-order form of the Lagrangian, so
let’s start with that. Rewriting (30.1) and integrating twice by parts,

The massive vector propagator D Fµν(k) is given by (28.97), and the scalar propagator by the Feynman propagator
F(q), (10.29). I’ll write down the complete set of Feynman rules for scalar-massive photon interactions (see also
the box on p. 571).
Feynman rules for massive photon scalar electrodynamics
You should imagine that there are quotes around the word “photon” in these rules. The factors are easily read off
the Lagrangian. For example, the three-point vertex arises from the term −ieAµ(ϕ *∂µϕ − (∂µϕ *)ϕ). In momentum
space, we pick up an i from the vertex, a factor of −ie from the coefficient of the fields, and a factor of (−iqµ − iq′µ)
from the derivatives acting on ϕ and −ϕ *, respectively. That gives an overall factor of −ie(qµ + q′µ). Similarly the
four-point vertex has a factor of i from the vertex, e2 from the coefficient, gµν from the two fields, and a factor of 2
from the combinatorics (AνAµ gives the same contribution as AµAν).

Spinor electrodynamics with a massive photon

In the same way, we can read off the Feynman rules from the Lagrangian for a Fermi field coupled to a
massive photon. This is the sum of the free Proca Lagrangian (26.47), written in terms of the Aµ’s, and the
minimally coupled (27.58) Dirac Lagrangian:

This is a somewhat simpler theory than scalar electrodynamics, where we have to worry about the balance
between the e2Aµ2 interaction and the derivative interaction. The Feynman propagator SF( ) for a fermion is given
by (21.76). Here are the Feynman rules for spinor electrodynamics with a massive photon:
Feynman rules for fermions and massive photons

There is a factor in both these sets of Feynman rules that at first glance should make you a bit nervous about
a smooth passage to a zero-mass limit, to actual photons. In both the scalar and the spinor cases, the term
in the vector propagator looks like bad news, in two ways. This term will give us trouble not only in going to the
zero-mass limit, but also in keeping the theory renormalizable. At high energies, the propagator goes not like 1/k2,
as with scalar propagators, but like k2/k2, which is simply 1. That certainly is going to make Feynman integrals
much more divergent than they would be in a theory with scalar ‘photons’. We’re going to have to worry about that.
In the low order computations, I will demonstrate that this term could have been crossed out without changing
anything. In a future lecture, I will demonstrate that you can get rid of this part of the propagator altogether for a
massive Abelian theory, with only one vector particle, like massive QED. For massive Yang–Mills theories, with
more than one vector particle, you can’t get rid of it. For non-Abelian gauge theories the situation is different.
There the massless theory is not the limit of the massive theory.5

30.3Some low order computations in spinor electrodynamics

We’ll look at two processes, Coulomb scattering and Compton scattering.

EXAMPLE. Coulomb scattering

Let’s consider the elastic scattering of two electrons, e(p1) + e(p2) → e(p′1) + e(p′2) to O(e2). The topological
structure of this process is given in Figure 30.1 (cf. Figure 21.7).

Figure 30.1: Coulomb scattering

The internal momenta in the two graphs are

The invariant amplitude Afi is a sum of the contributions A1 and A2 of the first and second graphs, respectively:

where the relative minus sign is due to the exchange of the two incoming identical fermions (see the discussion of
nucleon-nucleon scattering in §21.5, pp. 447–448).

Let’s focus on the apparently disastrous terms, with kµkν/µ2 or qµqν/µ2. These are free fermions, so

Therefore the kµkν/µ2 term in A1 actually drops out, and the same thing happens to the qµqν/µ2 term in A2. (We will
see later the general reason why these terms are always absent.) Then

Aside from the spin factors this is much like the exchange of a scalar meson, (21.100): the only difference is that
we have a γµ instead of a γ5. This is exactly the same as if we had exchanged four scalar mesons, one coupled to
each of γ0, γ1, γ2 and γ3. It looks like there are four kinds of photons and that one of them, is very peculiar: it has a
negative propagator, proportional to g00. That’s an illusion. From the orthogonality condition, k ⋅ ε = 0, we see that
there are only three transverse photons being exchanged; the projection operator doesn’t make any difference. To
obtain the differential cross-section, I would need to do the spin sum. I won’t do that here. If you want to see that,
it’s in Bjorken and Drell.6 I would instead like to talk about the zero-mass limit, which we can now take in a smooth
way since we’ve gotten rid of the kµkν/µ2 terms in the vector propagator.

The zero-mass limit: µ2 → 0.

There is first the fact that the forward peak, which typically occurs in lowest order scattering, has now moved
onto the verge of the physical region.7 The forward peak is infinite: k2 = 0 in the forward direction and the
denominator blows up. So the differential cross-section dσ/dΩ has an infinite peak in the forward direction. This is
no surprise: exactly the same thing occurs in non-relativistic Coulomb scattering, under the Rutherford formula:8

It is due to the long range nature of the Coulomb force. There is nothing particularly field theoretic about it; it’s just
what happens when we scatter charged particles.

Something else stands out in the µ → 0 limit. If you’ve had some previous exposure to QED, the limit of
(30.12) or (30.13) may look a little strange. It looks like we are exchanging four photons, one of which has a
negative sign in its propagator. This isn’t the quantum electrodynamics you may have seen before, quantized in
the Coulomb gauge, in which there are only two kinds of photons, the two three-dimensional transverse photons,
with polarization vectors ε perpendicular to k. In addition, something arises in the Coulomb gauge quantization
that doesn’t seem to appear in this formulation at all, an instantaneous action-at-a-distance Coulomb interaction.9
Where does that come from?

Well, that Coulomb interaction is actually in (30.10); we’ll see this by rewriting and reinterpreting the
amplitude by purely algebraic means. To make life simple, so that we don’t need to continually write and u, we
define the “currents” jµi as

From (30.11), each current is conserved:

To see the non-relativistic Coulomb contribution, separate out the space and time parts:

Write (30.12) as

(the same reasoning applies to A2, with a different set of j’s). Separate the j(r)’s (r = 1, 2) into their spatially
transverse and longitudinal parts:

By construction,

and by current conservation,

Rewriting A1 in terms of j(r),

Calling the four photons “apples”, we see that we have exchanged four apples for two apples and an orange.10
The first term on the right side of (30.22) may be interpreted as the exchange of two transverse photons. There is
the typical massless photon propagator in the denominator, and in the numerator the interaction between the
traverse parts of the current. So there is only an interaction between two types of photons, because there are only
two independent components to a transverse 3-vector field. In the second term (the orange), the coefficient of
j 0(1)j 0(2) is simply −e2/|k|2, with no k0 in the denominator. That is to say, it does not correspond to a time-
dependent interaction, which would have k0 in its Fourier transform, but to an instantaneous interaction. Notice
also the appropriate sign change has taken place; previously the propagator had the wrong sign. And indeed, the
Fourier transform of 1/|k|2 is the Coulomb interaction ∝ 1/|r| between the currents. Thus this amplitude, which
appears to correspond to the exchange of four kinds of photons, one with the wrong sign, is indeed equivalent to
the exchange of two transverse photons plus the instantaneous Coulomb interaction between the charge
densities, just like standard QED in the Coulomb gauge.

It is worthwhile to compare this result with what we found in Model 2 (§9.3). If we let the interaction
Lagrangian be

(without the counterterm a) then as we found earlier (9.36) for a time-independent ρ(x) = ρ(x),

The quantity in the square brackets is the Yukawa potential V (|x − y|); scalar exchange is attractive between
identical particles. In the massive vector case,

we have

Because the current is conserved, ∂µJµ = 0, it follows that

so that (in agreement with the general result) the second term in the propagator can be discarded, and we can
safely set µ = 0 to study the photon interaction:

If we’re looking at electrostatics,

Analogous to (30.24), (30.28) leads to

That is, the Coulomb potential

is repulsive between identical charges, the extra minus sign coming from the g00 in the propagator.11 Returning to
(30.22), it’s easy to see that the Coulomb interaction is repulsive if j 0(1) and j 0(2) have the same sign and attractive
if j 0(1) and j 0(2) have opposite signs:
As the currents have the same time t in their arguments, the interaction is instantaneous.

Of course we’ve only been playing with low order diagrams. If we want to show that similar things are true in
general we either have to crank up an enormous amount of combinatoric machinery, or establish some general
formalism which enables us to short-circuit the combinatoric machinery, i.e., functional integration. I will do that
later on, but I thought you should see how these things work out in particular diagrams before I show you the
general argument.

EXAMPLE. Compton scattering

The next process I would like to discuss, although not in nearly as much detail as Coulomb scattering, is
Compton scattering, e(p) + γ(k, ε) → e(p′) + γ(k′, ε′). (See Bjorken and Drell for a fuller discussion.12) Aside from
the extra indices this is just the same sort of thing as meson-nucleon scattering (see Figure 11.2, p. 228). There
are two diagrams with the same topological structure. The invariant Feynman amplitude is

Figure 30.2: Compton scattering

(Note that − ′ = ′ − .) We don’t need the iϵ in the denominator because the pole is not in the physical region.
We could rationalize the propagator denominators, commute γ matrices around, and use the fact that u and u′ are
free solutions to the Dirac equation ( u = mu) to simplify things. I shall not bore you with that; it’s a standard
computation that you can look up in Bjorken and Drell. Instead, I would again like to focus on the zero mass limit.
We will find some interesting properties of this as µ2 → 0.

The zero-mass limit: µ2 → 0

Let’s recall something from an earlier lecture (§26.4): the emission or absorption of a photon by an external
current distribution j µ. We found (26.73) the amplitude Afi for the emission of a single photon was proportional to
ε*⋅ :

We knew, because the external current was conserved, that

We also showed13 that for helicity 0,

From this we found the amplitude A3 for emission of a helicity 0 photon,

The point of this exercise was to demonstrate how the theory of a massive photon with three helicity states goes
over to the theory of a massless photon with two helicity states as the mass goes to zero. Is a corresponding thing
true in our more complicated theory, full of interactions? The key equation is the analog of current conservation,
(30.16), kµ µ(k) = 0.

How do we compute the amplitude for emission or absorption of a photon a → b + γ in a fully interacting field
theory? We know how to compute that amplitude in the general theory: we do it via the LSZ reduction formula,
(14.18). We reduce and reduce and reduce. Imagine that we have reduced everything until only the last photon is
unreduced. Then the reduction formula gives the amplitude Afi as

The photon contributes its polarization vector ε*µ, and the Klein–Gordon operator takes care of the pole in the
propagator. The unreduced field Aµ(x) is the exact Heisenberg field, the states are the exact physical states. The
time-ordering symbol T is unnecessary here; we have only one field left. If we’re taking account of
renormalization, we should use the renormalized A′µ here, with amplitude 1 for making a photon, but we can
absorb the constant into the proportionality. The Heisenberg equation of motion,

follows from the Euler–Lagrange equations and current conservation. The derivative ∂µ commutes with the
Klein–Gordon operator, and therefore we find

This is precisely the same statement, in position space rather than in momentum space, that we used earlier (kµ
µ(k) =0) to prove the suppression of helicity zero photons. Therefore the argument for the suppression of helicity
zero photons as µ → 0 should be as true in the full field theory as it was in the theory with a c-number source.

General arguments are always nice, but one sleeps better at night if one has made particular checks in
simple cases. So just to make sure nothing is going wrong, let me attempt to check this formula, that Afi = 0 when
εµ ∝ kµ, which is equivalent to checking the conservation equation. The conservation equation is exact whether or
nor µ = 0. The suppression of helicity zero states is a kinematic consequence of it as µ → 0. So let’s take the
expression for the amplitude, (30.33), plug in εµ = kµ/µ and see if the amplitude vanishes or not. Looking at the
numerators, we can say

since ( − m)u = ′( ′− m) = 0. We can substitute these expressions in the numerators, and the extra factors
cancel the Feynman denominators. Thus, just as in (26.100), we find that the full term is

Thus the amplitude vanishes for any non-zero mass when εµ = kµ/µ and therefore vanishes for helicity 0 states
when µ → 0.

Just because we’ve proven that something is true in general does not mean it is true at the level of individual
Feynman diagrams. When we know something is true in general for all values of e, it is true for all derivatives with
respect to e at e = 0; that is, for all orders of perturbation theory. However, each order of perturbation theory is a
sum of Feynman diagrams. It means that it must be true for that whole sum; it does not mean that it must be true
for individual Feynman diagrams. In this case we must take account of both diagrams to get the proper
cancellation. We can use the general formula (30.37); we don’t have to go through the complicated combinatorics
for diagrams of order e4, e6, etc., to prove the theorem. The exact statement is that if εµ = kµ/µ then Afi = 0. The
second part of the argument is that, for a helicity zero state, we have (30.36):

That’s just kinematics. Write down the properly normalized helicity zero polarization. Then dotting kµ/µ into the
amplitude gives zero, plus a term proportional to µ/k0, the mass over the energy. The only part of the argument
that needs to be checked in the full field theory is the statement that with εµ = kµ/µ that Afi = 0. We’ve given a
general proof and a specific example. We can write the amplitude for emitting a photon in a particular spin state r
as

It doesn’t matter where this amplitude Mµ came from, as long as it obeys the condition that kµMµ = 0. That shows
the suppression of helicity zero photons, because those photons have εµ ∝ kµ.

Summing and averaging over photon spins

Before I depart from specific Feynman calculations I would like to make a few comments about summing over
photon spins.14 We typically have to do a spin sum over final states,

the second equality following from (26.93); −gµν + (kµkν/µ2) is the projection operator onto the three four-
dimensional transverse vectors. It’s just like the spinor sum + m, the projection operator onto the positive energy
spinors. The vector spin sums are considerably easier than the spinor sums; for one thing kµaµ = 0, so the second
term in the projection operator gives zero. Also we don’t have the analog of all those ugly extra γ matrices to
commute around. So we simply obtain −Mµ*Mµ.

Likewise for averaging over initial photon spins, if one has an unpolarized beam. This polarization sum
(26.93) is true whether the mass is large or small. If the mass is small we may think that we want to sum over only
two spin states. But we might as well sum over the third, because the amplitude for emitting the third is negligible.
If we have an unpolarized beam and we wish to average over initial spins, we get the above result (30.44)
multiplied by 1/3, for the three photon states if µ ≈ k0, or by 1/2 for the two photon states if µ ≪ k0.15 If µ ≪ k0 (the
current experimental bound has λCompton ≫ 104 km for the photon),16 even the most imperfect light bulb will not
emit all three helicity states with indifference. It will in fact emit no helicity zero photons. We may think it’s an
unpolarized beam, but in fact it’s just a random mixture of two polarizations, not three; we would make an error
inserting a . In practice, we don’t have to worry about intermediate ranges of mass. Either we are talking about
something like ρ mesons, which, if unpolarized, really have three polarization states, or photons, which have, for
all practical purposes, two, even if the photon mass is not strictly zero but only, say, 10−30 melectron.

30.4Quantizing massless electrodynamics with functional integrals

We are now going back to the wonderland of functional integrals, where I will adopt the lecture style of the Delphic
Oracle. 17 We’re going to begin a discussion of massless electrodynamics. Not, as we have treated it until now, as
the limit of the massive theory as µ → 0, but sui generis, as a theory by itself, without embedding it in another family
of Lagrangians, and try to quantize it. The Lagrangian is nearly the same as (30.8), but now there is no mass term
for the photon:

This theory cannot be directly quantized by naive canonical methods. We cannot eliminate A0 when µ = 0 and the
whole canonical quantization program falls apart. That’s because this theory has gauge invariance. Let’s review.

Gauge transformations are not like ordinary internal symmetries. They do not turn one physical situation into
a distinguishable physical situation with an identical scattering amplitude, like isospin transformations turn a
proton into a neutron. Rather, they simply represent changes in the description, not actual changes of the state.
More conventional transformations can be sensibly interpreted both actively, as changing the state, or passively,
as changing the description. But a gauge transformation is only sensibly interpreted in the passive sense; it is a
change in the description, like translating a physics paper from English into French.18

In order to canonically quantize the theory we must pick a gauge. If we don’t have a condition telling us what
gauge we are in then we don’t have a well-defined initial value problem. No matter how many derivatives of fields
we specify on the initial value surface, we can always make a gauge transformation; that is, the identity on the
initial surface and not the identity at some future time. Therefore, we must adopt a condition that firmly and forever
fixes the gauge, such as the Coulomb gauge condition ∇⋅ A = 0. Then we have eliminated, by convention and by
fiat, the gauge degrees of freedom, and we have a well-defined initial value problem. If we are lucky in our choice
of gauge, we can crank out the entire canonical machinery, eliminate the constrained variables, impose canonical
commutation rules and be able to compute everything. We hope that the choice of gauge doesn’t matter as far as
actually observable quantities go, i.e., we predict gauge invariant results. The computations may be simpler in one
gauge than in another, but the final answers should be the same in all gauges. Right now this statement is an act
of faith. It is hoped that gauge invariance of the classical theory which we are about to quantize will carry over at
least that much into the canonically quantized theory, but this remains to be shown; we will show it next time.19
That’s the canonical viewpoint of quantization of theories with gauge invariance as expressed by Fermi and Dirac,
circa 1929–1930.20

We could also take a functional integral viewpoint. We have not been thinking of the functional integral as a
primary object but as something we derived from canonical quantization. But some young revolutionaries could
just as well take functional integration as fundamental, and forget about canonical quantization: we have these
magic formulas, let’s just apply them. However, if they were to try that in this case, they would run into trouble.
Let’s press our luck and see how far we get.

We would try to find the Feynman propagator by inverting the quadratic part of L. Following the development
from (28.86) to (28.94), we break things up into transverse and longitudinal projection operators

Only the transverse part of the field enters L because only the antisymmetric derivative of the field appears in the
Lagrangian,21 and the longitudinal part doesn’t enter into the quadratic part of the Lagrangian at all. That is,
setting µ2 = 0 in (28.95),

Inverting the coefficients as in (28.94) to get the Feynman propagator, we obtain

This is garbage; I can’t do any computations with this disastrous propagator.

Many years ago, two young Russians, Faddeev and Popov, looked at this problem from the functional integral
point of view and made a guess about what to do.22 With that guess they were able to verify canonical
quantization. However, in order to explain their guess I will have to tell you a little bit more about Feynman’s
original formulation of functional integrals.

Aside. A brief historical digression on Feynman’s sum over histories

Feynman called his functional integrals path integrals, and the specific operation of calculating them the
sum over histories.23 He didn’t do it in a source formalism with a generating functional. He wanted to compute
actual transition matrix elements. We’ll just write down his formula without proof, for the simplest example of a
particle in a potential

Feynman wanted to compute the transition amplitude áq2|e−iH(t2−t1)|q1ñ for the state where the particle was at
position q1 at time t1, to the state where it was at position q2 at time t2. He showed that it could be written as24

where the integration doesn’t go over arbitrary functions in the range t1 to t2 but is restricted to run over functions
that are held fixed at the end points, q(t1) = q1, q(t2) = q2, just as in Hamilton’s formulation of Lagrangian
mechanics. Feynman described this as a “sum over histories”. He said this was a neat formulation of quantum
mechanics, and indeed it was. We imagine the particle goes over all possible classical paths from the desired
initial state q1 at t1 to the desired final state q2 at t2. We sum ei Ldt over all possible paths to get the transition
matrix element. The functional integral gives a precise meaning to the concept of summation. Briefly, we divide the
time between t1 and t2 into N equal intervals of width Δt, with t2 = t1 + NΔt, and approximate a classical path as a
connected set of linear segments. Two such are shown25 in Figure 30.3. We can add in a source term and let t1 →
−∞, t2 → ∞, and then we see how to get our formulation. If we kept the end points q1 and q2 fixed, we would get the
transition matrix element from some initial state |q1ñ to some final state |q2ñ over an infinite stretch of time.

We aren’t keeping the end points fixed, but that doesn’t matter. Our discussion (§7.3) of how Dyson’s formula
gave us the S-matrix elements included an argument, for field theories, that as we go to the far past and the far
future, no matter what states we have on the right and the left, all that survives is the vacuum-to-vacuum
transition; all the other parts are canceled by contributions of oscillating phases.26 So aside from some
normalization factor, we could do it with a particular q1 and q2, and get the vacuum-to-vacuum transition. Or, we
could let q1 and q2 be free, integrate over all possible q1’s and all possible q2’s, and then just change the
normalization factor. So our form comes from Feynman’s form. In fact, it’s even better. We really work in
Euclidean space where the non-vacuum states aren’t canceled by some Riemann–Lebesgue argument, but by a
decreasing exponential; it knocks them out even more forcefully.27

Figure 30.3: Two approximate paths from (q1, t1) to (q2, t2)

There’s a second thing that we can see from this formulation that’s not obvious from our formulation. We see
why classical mechanics is important in the small ħ limit. From Feynman’s formulation, we can see where
Hamilton’s Principle comes from by restoring the ħ. The dimensions of ħ, J-s, are those of an action, so
Feynman’s formulation should really read

As ħ → 0, the phase factor on the right-hand side oscillates more and more rapidly. Rapidly oscillating integrals are
dominated by points of stationary phase,28 points where the phase is stationary when we vary the integration
variables. In our case, the phase is the action S and the integration variable is the coordinate q, so we must vary it
so that

sticking to Feynman’s boundary conditions q(t1) = q1, q(t2) = q2. This is nothing but Hamilton’s Principle that picks
out the classical motions. The reason that classical motions are important in the small ħ limit, according to
Feynman, is because of the principle of stationary phase. They are the points where the phase, the action, is
stationary.

We can now explain Faddeev and Popov’s central idea. They said that putting the Lagrangian (30.45) into the
functional integral was a very dumb thing to do, because Feynman says “sum over histories”. If you have a gauge
theory, then the same history, exactly the same motion for all observable quantities, may be represented by an
infinite number of different fields, all of them connected to each other by a gauge transformation. So we’re not
summing over the histories in the right way. If we just stuck the Lagrangian, (30.45) into the functional integral and
tried to sum over histories, we’d be summing over the same histories many, many times—in fact, an infinite
number of times—for each history. No wonder, said Faddeev and Popov, we get infinity when we attempt to
evaluate the integral! That’s where the , the infinity, comes from. The Russian dogma is that you must change
the functional integral formula to sum over each history only once. Not once over it in one gauge, once over it in
another gauge, and on and on—but once and only once. How do we arrange for the functional integral to do that?
How do we fix up our formula so that we sum over each history only once? Well, it involves putting in something
like a delta function, and what I mean by that I will explain in a more precise manner next time. I will implement the
Faddeev–Popov idea in equations, then apply it, and finally justify it by recourse to canonical quantization.
1[Eds.] The Feynman rules for scalar electrodynamics are treated in Greiner & Reinhardt QED, Section 8.4, pp.
434–435; H. Kleinert, Particles and Quantum Fields, World Scientific, 2015, Chapter 17.
2 [Eds.] See the example on p. 587.
3 [Eds.] See §29.5.
4 [Eds.] We use m for the scalar meson mass because we are writing µ for the vector meson mass.
5 [Eds.] The reader is probably familiar with the terms “Yang–Mills” and “non-Abelian gauge theories”, just as
Coleman’s students were in 1976. In 1954, Yang and Mills wrote a landmark paper generalizing Maxwell’s theory
of a single vector field to a theory of three vector fields transforming among themselves under the Lie group SU(2).
The gauge invariance of electrodynamics is based upon the Lie group U(1). This group has only one generator,
and so it is trivially Abelian: its generator commutes with itself. SU(2), with three non-commuting generators, is
non-Abelian. The terms “Yang–Mills theory” and “non-Abelian gauge theory” are effectively synonymous, even
though electrodynamics is a Yang–Mills theory, and general relativity, though a gauge theory, is not usually
regarded as a Yang–Mills theory: C. N. Yang and R. Mills, “Conservation of Isotopic Spin and Isotopic Gauge
Invariance”, Phys. Rev. 96 (1954) 191–195; see also §46.2 and §47.3. For the different zero-mass limits of
Abelian vs. non-Abelian gauge theories, see note 22, p. 1044.
6 [Eds.]See Bjorken & Drell RQM, pp. 102–106 for electron scattering in a Coulomb potential; they obtain the Mott
cross-section for Rutherford scattering in equation (7.22):

The Mott formula is the relativistic generalization of the Rutherford formula. For electron–electron scattering,
called Møller scattering, see section 7.9, pp. 135–140. That cross-section, in the high-energy limit, is given in
equation (7.84), p. 138.
7 [Eds.]See §11.3. In terms of the Mandelstam variables, k2 = t and q2 = u. In the limit that both t and u approach
zero, the process becomes unphysical.
8[Eds.] See Problem 11.1, p. 397, in D. J. Griffiths, Introduction to Quantum Mechanics, 2nd ed., Cambridge U.
P., 2016; Landau & Lifshitz, QM, §133, pp. 516–519.
9 [Eds.] See Bjorken & Drell Fields, Chap. 14, pp. 68–81; the instantaneous Coulomb interaction part of the
propagator is given in equation (14.55), p. 80; or Appendix A, p. 301 in J. J. Sakurai, Advanced Quantum
Mechanics, Addison–Wesley, 1967. This instantaneous Coulomb interaction is also found in the classical solution
of Maxwell’s equations in Coulomb gauge. See the paragraph following equation (10.10), p. 441 in David J.
Griffiths, Introduction to Electrodynamics, 4th ed., Pearson, 2013.
10 [Eds.] What follows from here through (30.32) is based on class notes from 1999, supplied by Daniel Podolsky.
11 [Eds.]The identical argument is given by Zee QFTN, pp. 32–33, with explicit citation of Coleman’s QFT course.
Incidentally, Zee earned his PhD under Coleman. The first demonstration that photon exchange leads to the
Coulomb potential seems to have been given by V. A. Fock and Boris Podolsky, “On the quantization of electro-
magnetic waves and the interaction of charges in Dirac’s theory”, Phys. Zeits. Sowjetunion 1 (1932) 801–817,
following Dirac’s demonstration of the one-dimensional (attractive!) result: P. A. M. Dirac, “Relativistic Quantum
Mechanics”, Proc. Roy. Soc. Ser. A 136 (1932) 453–464. Schweber writes, “In 1932, Fermi and Bethe set out to
derive the interaction potential between two charged particles, including magnetic and retardation effects (H.
Bethe und E. Fermi, “Über die Wechselwirkung von zwei Elektronen” (On the interaction of two electrons), Zeits.
Phys. 77 (1932) 296-306; reprinted in Enrico Fermi: Collected Papers, v.1, ed. E. Segrè et al., U Chicago Press,
1962). . . Bethe and Fermi’s aim was to reveal the relation between Møller’s and Breit’s approaches, and more
important, to demonstrate how perturbation theory could be used to generate transparent results. It is clear from
their derivation that Bethe and Fermi considered the force between the charged particles as arising form the
exchange of the photons between them.” Silvan S. Schweber, “Enrico Fermi and Quantum Electrodynamics,
1929–32”, Phys. Today 55 (2002) 31–36. An expression equivalent to (30.30) is given in Gregor Wentzel,
Quantum Theory of Fields, trans. C. Houtermans and J. M. Jauch, Interscience, 1949, p. 132, equation (17.38);
the original German text was published in 1943.
12 [Eds.] Bjorken & Drell RQM, pp. 127–132. See also the example in §26.5, p. 571.
13 [Eds.] This argument is a little incomplete. For massive vector fields, the polarization vectors εµ are orthogonal
to the 4-momentum: kµεµ = 0. With kµ = (ωk, 0, 0, |k|), the polarization vector ε(3)µ for a helicity 0-vector was given
as

However, as usual,
so that we can write

Though ε(3)µ and kµ must be orthogonal for a massive vector, ε(3) ⋅ k = 0, ε(3)µ and kµ/µ are also parallel to within
(µ/|k|).
14 [Eds.] Bjorken & Drell RQM, p. 125.
15 [Eds.] Bjorken & Drell RQM, p. 131; Jackson CE, pp. 694–5; V. B. Berestetskiĭ, E. M. Lifshitz, and L. P.
Pitaevskiĭ, Quantum Electrodynamics, 2nd ed., Pergamon, 1982, pp. 354–364.
16 [Eds.] See §26.4.
17[Eds.] Coleman jokes: “Like her, I speak while breathing in a gas—not a natural gas, as she did, but tobacco
smoke.” In the videotaped 1975–6 lectures, Coleman typically smoked eight cigarettes during a ninety-minute
lecture. The Oracle of Delphi, (700 BCE–400 CE), known as the Pythia, was the high priestess of Apollo, thought
to be able to foretell the future.
18 [Eds.] Coleman is reiterating a point he made originally on p. 579.
19 [Eds.] For a full discussion of gauge transformations at the operator level, including indefinite-metric state space
and the appearance of the Coulomb interaction, see K. Haller, “Gauge Problems in Spinor Quantum
Electrodynamics”, Acta Phys. Austr. 42 (1975) 163–215.
20 [Eds.] E. Fermi, “Quantum Theory of Radiation”, Rev. Mod. Phys., 4 (1932) 87–132; P. A. M. Dirac, “The
Quantum Theory of the Emission and Absorption of Radiation”, Proc. Roy. Soc. London, A114 (1927) 243–265.
Both papers are reprinted in Schwinger QED.
21 [Eds.] Since Fµν is gauge invariant, it can only be transverse.
22 [Eds.] Ludvig D. Faddeev and Victor N. Popov, “Feynman diagrams for the Yang–Mills Field”, Phys. Lett. 25B
(1967) 29–30; V. N. Popov and L. D. Faddeev, “Perturbation Theory for Gauge-Invariant Fields”, Fermilab report
NAL-THY-57 (1972); reprinted in G. ’t Hooft, ed., 50 Years of Yang–Mills Theory, World Scientific, 2005; and L. D.
Faddeev, “Introduction to Functional Methods”, pp. 1–40, in Methods in Field Theory (Les Houches 1975), R.
Balian and J. Zinn–Justin, eds., North-Holland, 1976. The Faddeev–Popov methods are discussed in Chapter 31.
23[Eds.] See Richard P. Feynman and Albert R. Hibbs, Quantum Mechanics and Path Integrals, McGraw-Hill,
1965; edited and corrected by Daniel F. Styer, Dover Publications, 2010. Feynman had been looking for a way to
base quantum mechanics not on the Hamiltonian, but the Lagrangian. A colleague, Herbert Jehle, told Feynman
about Dirac’s paper: P. A. M. Dirac, “The Lagrangian in Quantum Mechanics”, Phys. Zeits. Sowjetunion 3 (1933)
64–72; reprinted in Schwinger QED. See also Feynman’s Thesis: A New Approach to Quantum Mechanics, ed.
Laurie M. Brown, World Scientific, 2006. For the development of the path integral method, including many
historical references, see D. Derbes, “Feynman’s Derivation of the Schrödinger Equation”, Am. J. Phys. 64 (1996)
881–884.
24[Eds.] The development of (30.50) is given in pp. 60–62 of Ernest S. Abers and Benjamin W. Lee, “Gauge
Theories”, Phys. Lett. C9 (1973), 1–145. (Physics Letters C subsequently became Physics Reports.)
25 [Eds.] Figure 30.3 is based on Figure 5.3, p. 157 in Ryder QFT.
26 [Eds.] See (13.40) and the discussion following.
27 [Eds.] See (28.39) and the discussion following.
28 [Eds.] See §17.4.

31
The Faddeev–Popov prescription

At the end of the last lecture I was on the verge of describing the bright idea of Faddeev and Popov.1 Their
essential insight was this. Feynman tells us to sum over histories. However, if we just blindly perform the functional
integral in a gauge theory, we are not summing over all histories once and only once; we are summing over each
history many times, in all of its various gauge-transformed versions. But we should count each history exactly
once. That was their idea. It was just a guess. We’re going to formulate this guess in a precise form, explore its
consequences and then prove it is true by showing that it is equivalent to canonical quantization for the theories of
interest.

31.1The prescription in a finite number of dimensions

In order to write down the guess of Faddeev and Popov, it’s easier to start with a finite-dimensional analog where
we integrate over only a finite-dimensional space, then generalize, in my usual brutal way, to a function space, by
simply copying down some of the equations and changing some of the symbols. And then I will have arrived at
their guess. The finite-dimensional model of what we’re doing in a gauge invariant field theory is this: we have a
function that depends on n + m real variables

That’s the analog, in some sense, of the gauge invariant eiS. (I’ve called them zi, but they are in fact real
variables.) We divide the z’s up as follows:

The idea is that F depends only the xr’s:

The y’s are like the gauge degrees of freedom: we can change the y’s without changing the value of F, just as a
gauge transformation does not change the value of exp{iS}.

Perhaps a picture will be helpful; see Figure 31.1. As I am restricted to a two-dimensional blackboard, there
will be only one x and one y.

Figure 31.1: Gauge freedom described as motion along a line

Along any of these lines parallel to the y axis, the function F is a constant. They are the finite-dimensional analogs
in the gauge system of the various motions that are connected together by gauge transformations. Gauge
transformations are like translations in the y direction.2

Now we want to define an integral. We’ll obviously get a divergent integral if we try to integrate this thing over
all the z’s. So we define the integral by integrating only over the x’s:

That’s certainly an integral that cuts each set of these equivalent points, each vertical line, only once, because I’m
just integrating along the surface y = 0. Equivalently, we could write this as an integral over all the z’s:

where the delta function restricts us to yb = 0. Of course we don’t have to restrict ourselves to the flat surface y = 0.
We could restrict the integral to some curved surface defined by

which cuts the lines as shown in Figure 31.2, and integrate restricting the yb’s to that surface:

Figure 31.2: Fixing the values of the y’s

That’s the same integral as (31.5). It may not be convenient to parameterize the surface by yb = fb(x1, x2...). It
might be better to give the y’s implicitly as functions of the x’s; i.e., by a set of equations

such that when we solve for the y’s we get yb = fb(x1, xn). A completely equivalent way to write I is

The factor Δ is the Jacobian determinant to take account that we are integrating with respect to different variables;

(∂Gb/∂yc is an m × m matrix.) That just reproduces the same m-dimensional delta function as in (31.7). We’ve
rewritten a very simple integral, where we just integrate over the x’s, in a much more complicated form. But it is the
same integral.3

31.2Extending the prescription to a gauge field theory

Armed with this finite-dimensional knowledge, we can now describe the Faddeev–Popov prescription for a gauge
field theory; in particular, for quantum electrodynamics.4 In QED, we have a set of fields transforming in various
ways under a gauge transformation parameterized by χ:

These transformations describe physically equivalent situations: given any history (Aµ, ψ, etc) as a function of
space and time, if I apply a gauge transformation I get a new set of functions which describe the same history. For
notational convenience we will assemble all the fields into a single big field Φ.

We have a gauge invariant action S[Φ] which is unchanged by this transformation

corresponding to (31.3).

The Faddeev–Popov prescription is this. First, we pick a gauge. That is to say, we adopt some condition that,
out of this infinite family of gauge-equivalent motions, picks out one and only one. This is equivalent to picking out
an integration surface that passes through each of the lines in Figure 31.2 only once. Recall that a gauge is some
condition G(Φ) = 0, analogous to the earlier Gb(zi) = 0, that you choose to eliminate the freedom to make gauge
transformations.5

Figure 31.3: Fixing the gauge

Some standard gauges:

Coulomb gauge:

(The Coulomb gauge is also known as the radiation gauge.) As you all know from your experience of classical
electrodynamics, once you have adopted this condition, you have no further freedom to make gauge
transformations, assuming that you impose the usual boundary condition that A falls off at infinity. (We’ll review
this argument in a moment.)

Axial gauge:

This is a gauge we will find convenient for proving certain theorems, although it’s terrible for computations; it
destroys the manifest rotational invariance of the theory. It’s called “axial gauge” because it picks out a certain
coordinate, i.e., a certain axis.6

Lorenz gauge:7

You may be a little worried about this choice, because it does not completely fix the gauge: we can always add to
Aµ a quantity ∂µχ, where χ is a solution to the homogeneous wave equation 2χ = 0. But this is not a problem.
Remember that we’re secretly doing all functional integrals in Euclidean space (which is hidden within our
perverse notation). In Euclidean space, the homogeneous wave equation8 is the (4-dimensional) Laplace
equation. With our usual assumption that everything goes to zero at infinity, the Laplace equation has no non-
trivial solutions. It’s worth spending a few moments on this.

Aside: Why the Lorenz gauge condition determines Aµ uniquely

We use much the same argument with the Coulomb gauge, (31.14). That choice is also supposed to pick out
a unique potential, and keep it from changing under gauge transformations. We can still make a gauge
transformation

but if A′ is to stay in the Coulomb gauge, we must have

The usual boundary condition that A should vanish at infinity implies that χ ≡ 0, because the operator ∇2 has a
unique inverse with sensible boundary conditions. The corresponding equation in the Lorenz gauge, (31.16), is

If A′µ is to stay in the Lorenz gauge, we must have

Even if we impose boundary conditions at infinity, the homogeneous wave equation has many solutions, to wit, all
the free motions of a massless scalar particle. Were we working in Minkowski space, this would be a problem; Aµ
is not unique in the Lorenz gauge. But we are working in Euclidean space, and (31.20) is actually

The Euclidean 2 (28.46) has the same properties as ∇2; it is just the 4-dimensional analog of the Laplacian. It
E
has a unique inverse with reasonable boundary conditions, which implies that χ ≡ 0 is the only solution: the
apparent freedom expressed by (31.19) and (31.20) is illusory. We’re not really doing our functional integrals in
Minkowski space, we’re doing them in Euclidean space where we lose information. In particular, we lose the +iϵ
prescription and the pole in the propagator when we work in Euclidean space. That pole is in Minkowski space; the
pole is where the free solutions lie.

These three gauges—Coulomb, Lorenz, axial—are popular choices, but one could obviously write down
many more. We could have any other gauge condition, as long it removes the freedom to make further gauge
transformations.

I can now state the Faddeev–Popov prescription (originally an educated guess, or ansatz): it is the direct
generalization of the prescription (31.9) to function space. There’s a normalization factor, a functional integral over
all the fields, there’s an eiS as always, a delta function of whatever gauge function G(Φ) you have chosen, and
finally there is a Jacobian determinant:

The variable χ is the analog of the yb variables above; changing χ moves us along the vertical lines in Figure 31.1,
from one configuration to a gauge-equivalent configuration. The quantity δ[G] is a delta function in function space,
a delta functional, the analog of the finite-dimensional product of delta functions, δ(Gb), such that

I want to make four remarks before going on to apply this prescription to the three gauges described above, and to
show that it is equivalent to canonical quantization.

Remarks

First, whether or not the prescription is right, one thing is assured: the expression (31.22) is guaranteed to be
gauge invariant, in the sense that it does not depend on the choice of G. We get the same value of the integral for
one G as for any other, from the previous argument for the finite-dimensional case. The integrand in the Coulomb
gauge will look very different from the integrand in the axial gauge, but the two integrals must give the same
answer. So, if we can prove that the prescription is right, i.e., equivalent to canonical quantization, in any one
gauge—just one—then it has to be right in all gauges.

Second, I have pulled a small swindle on you. We have assumed that S is gauge invariant: S(Φ) = S(Φ′). But
typically the sort of actions S we’ve been talking about involve source terms like JµAµ or ψ which break the
gauge invariance. So, we must say that S has sources in it coupled only to gauge invariant operators like F2µν,
ψ, γµψ, and so on. That is, we should expect only those combinations formally gauge invariant in the
classical theory to be independent of which gauge we do our computations in. Since we also firmly believe that the
only physical observables are gauge invariant quantities, that should be sufficient to characterize the theory. If we
can show that the Faddeev–Popov ansatz gives the same results for Green’s functions of strings of gauge
invariant operators, no matter what the gauge—how we choose the G—then we’ve shown that it defines the same
physics in any gauge. Once we settle down in a particular gauge to do our computations, then in order to evaluate
the functional integral perturbatively, it might be convenient to introduce sources coupled to Aµ and ψ, and other
things like that, as an intermediate stage. But we shouldn’t expect either the AA Green’s function or the ψ
Green’s function to be independent of the choice of G; we do expect this independence from the gauge invariant
operators we construct from the A’s and ψ’s.
 Third, (a point I should have emphasized repeatedly during this entire discussion), everything here must be
taken with a grain of salt. All of this is just the formal manipulation of canonical field theory, and at the end we will
have to worry about ultraviolet divergences and whether they mess up our manipulations. Functional integration is
more compact than the manipulation of the canonical equations of motion, but it’s no more rigorous. When we’re
done with all this, we’ll have to see if we can put in a cutoff that preserves all the formal properties that we wanted
to preserve. In particular, we will have to worry about cutting off this theory in such a way that gauge invariance is
maintained. We can do it and we will do it.

And finally, functional integration does not prove anything. For proofs, we have to come down in the end to
canonical quantization. The power of functional integration is that it allows us to change variables with
unparalleled facility. The more changes of variables we have to worry about, the more useful the functional integral
is. In gauge theories, where we have an enormous family of variable changes to worry about, the functional
integral is practically the essential way of doing things.

Now let’s talk about the determinant. In quantum electrodynamics we have only one gauge condition G and
one gauge parameter χ. In each of the three gauges we have considered, the operator δG/δχ is a constant with
respect to the fields Φ, and det(δG/δχ) = Δ is independent of Φ, so we can absorb the determinant Δ into the
normalization constant N. For the Coulomb gauge, the determinant involves the Laplace operator, for the Lorenz
gauge the d’Alembert operator, and for the axial gauge, the derivative of a delta function. In any of these three
cases, the determinant is a constant. If we were using a more perverse choice of gauge for the Abelian case, or if
we were doing a theory with a more complicated group of gauge transformations like a non-Abelian gauge field
theory, in which we gauge not just electric charge conservation but say isospin conservation, then in general we
would not be able to get rid of the determinant factor. We would have to treat it in the usual way, by putting
det(δG/δχ) into the exponential via ghost fields. I shan’t do that here, because I don’t need to.

31.3Applying the prescription to QED

Let’s assume the Faddeev–Popov prescription is correct, and apply it to quantum electrodynamics. Later we’ll
prove that it is correct. For this application we’ll work in the Lorenz gauge, and choose

where f(x) is some fixed function. This is a generalization of the Lorenz gauge, (31.16), but the determinant Δ is
still a constant, because f doesn’t enter into δG/δχ.

We have

As shown in §29.3, Δ = det(δG/δχ) can be handled with ghost fields:

where the Faddeev–Popov ghost action SFP is given by

This is a constant with respect to the fields in Φ, and so it can also be absorbed into the normalization constant.
We say that the ghosts decouple.

The Faddeev–Popov prescription says that the generating functional is

independent of f; we get the same Z no matter what f is. (The determinant det(δG/δχ) is included in the
normalization constant N.) Since Z doesn’t depend on f we can also write
where N′ is a new normalization factor and F is a functional of f; any functional will do. Because the integrand in
(31.29) is independent of f, we can integrate it over f with any weighting functional F[f] and get the same answer,
modulo the proportionality factor. To get nice Feynman rules, we choose F to be equal to the exponential of a
quadratic form:

with ξ real, to be chosen at our convenience.9 Following the integration over (df) the generating functional is

where the effective action is

with the effective Lagrangian the sum of the original Lagrangian plus a gauge-fixing term,

By this device I have taken care of the earlier problem that bothered us, (30.48), the lack of a contribution from the
four-dimensional longitudinal component of Aµ in S; when we tried to find the longitudinal part of the photon
propagator, we got = ∞. Now we can just read the propagator off the quadratic terms of the effective Lagrangian:

That’s just the universal rule for functional integrals: invert the quadratic part of the Lagrangian to get the
propagator.10 The parameter ξ can be anything we want. This looks very much like the propagator in (30.48), but
without the disastrous garbage in the last term. The gauge-fixing term has removed it and rendered the
propagator finite.

We still haven’t shown that the Faddeev–Popov ansatz is right; that it agrees with canonical quantization. But
whether it’s right or wrong, any gauge invariant quantity computed using one value of ξ will have the same answer
using any other value of ξ. We’ve shown that they are all equivalent to each other.

In a slightly different use of the word “gauge” in the literature, the family of propagators (31.35) represent what
are called covariant gauges because the propagators look covariant. If we had done the same thing with the
Coulomb gauge, we would have had an extra term just involving the space part of k, because we’ve only got
space derivatives, and it wouldn’t look covariant. The covariant gauges have various names for particular choices
of ξ. The two most popular choices in the literature are ξ = 1, called the Feynman gauge,11

and the limiting case ξ → 0, called the Landau gauge,12

The Feynman gauge is useful for evaluating low order Feynman graphs. Additionally, it’s nice to have only gµν
instead of keeping track of all the kµ’s. We obtain the Landau gauge in the limit ξ → 0. This looks like a singular
limit in (31.31) but this limit really just restores the delta function in the original form of (31.29). Its utility for certain
general arguments was pointed out by Lev Landau. It is four-dimensionally transverse, kµ µν(k) = 0; in a formal
sense this is like setting ∂µAµ = 0. A third choice, ξ = 3, is sometimes used. This is the Yennie–Fried gauge13 with
propagator
We will say nothing more about it, other than to note that it has useful infrared properties.

All of these gauges give the same answer for any gauge invariant quantity. However, non-gauge invariant
quantities will look different in the different gauges. The photon propagator is one such, since Aµ looks different in
different gauges.14

31.4Equivalence of the Faddeev–Popov prescription and canonical quantization

I’ve shown that all these gauges, and indeed billions of other gauges, are equivalent. I will now show that the
Faddeev–Popov ansatz is equivalent to canonical quantization in a particular gauge, the axial gauge A3(x) = 0. I
choose this gauge because canonical quantization in the axial gauge is super-easy. Once I’ve shown that it’s true
in one gauge, I know it’s true in all other gauges.

Our first step to canonically quantize the theory in the axial gauge is just to impose the condition A3(x) = 0, to
fix the gauge. We find a set of p’s and q’s, and we write everything in terms of them. Finally we show that the
resulting expression is equivalent to the Faddeev–Popov prescription in the axial gauge. In the following, the Latin
indices i, j, etc., take the values 1 and 2 only; I will explicitly write the terms with 0 and 3.

As usual, the clue is the first-order form of the Lagrangian

This is messy, but canonically quantizing the first-order Lagrangian is like shooting fish in a barrel. The
independent variables are the two components Ai, their canonical momenta F0i, and the Fermi field ψ and its
canonical momentum . All the other variables—Fij, Fi3, F03, and A0—are constrained, defined in terms of the
independent variables and their space derivatives only. Fij and Fi3 are trivially constrained, given in terms of the
space derivatives of Ai:

Next, the component F03 is determined from the Euler–Lagrange equation obtained from (31.39),

For ν = 0 we have

which determines F03 in terms of known quantities. Finally

determines A0. The generating functional we get from canonical quantization is

On the other hand, by our general rule (29.77), the constrained variables disappear from the functional integral

since their coefficients are just constants.15 To within a new normalization constant, this is equal to the
Hamiltonian form, and therefore it is another expression for Z:

The part of the Lagrangian that remains in the action after integrating over the constrained variables is just p − H,
the action written in terms of Ai and F0i; this is the same argument as before. We obtain a third expression for Z by
integrating the first-order version (31.45) over all the F’s. As always, that eliminates the F’s and brings the
Lagrangian back into the second-order form, written in terms of the A’s and ψ’s with second derivatives:

where the delta function allows us to integrate over all four components of Aµ. But this is precisely the
Faddeev–Popov ansatz (31.22) for the axial gauge. Comparing (31.22) with the last line of (31.46), you may be
troubled by the apparent lack of the Δ factor. But the determinant is a constant for the axial gauge and can be
absorbed into N′′. And since the Faddeev–Popov prescription is independent of which gauge we choose, if it is
right in the axial gauge, it is right in any other gauge, and we are done. QED

This is a good place to write down the Feynman rules for electrodynamics with a massless photon. We’ll use
Feynman gauge:

Feynman rules for QED (Feynman gauge)

For every …

Write …

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

As an exercise, you can try to show the equivalence between canonical quantization and the Faddeev–Popov
ansatz in the Coulomb gauge, ∇ • A = 0. That gauge is a little harder, because we have to split A into transverse
and longitudinal components, A T and A L, and play the same game with A L replacing A3. But the conclusion is the
same, 16 as you would expect. Because it’s true for the axial gauge, it must be true for the Coulomb gauge, and
for all other gauges.

The axial gauge is a terrible gauge for doing any kind of actual computation: Lorentz invariance and even
rotational invariance of the theory are not manifest. But canonical quantization in the Lorenz gauge(s), with a nice
propagator, is much more involved. It requires subsidiary fields and then showing they don’t matter, conditions on
physically allowed states, etc.17 Canonical quantization in the axial gauge is trivial. The power of the functional
integral method is that it enables us to prove that a given formula is right in a gauge where canonical quantization
is simple but the Feynman rules are complicated, and then to change instantly to a different gauge, in which
canonical quantization may be complicated but the Feynman rules are simple.

31.5Revisiting the massive vector theory

All the computations in massless electrodynamics are the same ones we’ve done in massive electrodynamics.
They are right not only for massless electrodynamics as the massless limit of massive electrodynamics, but also if
we approach massless QED ab initio by canonical quantization. We need to address one final thing about
massive electrodynamics, the kµkν/µ2 term in the propagator: it doesn’t make any difference. To get rid of that
term in massive electrodynamics, consider the Lagrangian interacting only with a Fermi field, augmented by a new
scalar field, ϕ:

(the interactions with a Bose field can be handled similarly).18 The parameters a and b will be adjusted later. We
just add these terms in. They don’t change the physics in the slightest, because there is no coupling of ϕ to
anything else; it is a free field. Because the sources to be added to the Lagrangian will couple only to ψ, and Aµ,
we can write the theory as a functional integral, now including dϕ, over the additional terms. This extra functional
integration only changes the normalization factor. Since ϕ is completely unphysical, we don’t have to apply any
positivity constraints on a or b. If they have the wrong signs then ϕ will represent a field with negative energy or
negative probability because it has the wrong sign in the propagator. But it doesn’t couple to anything, so who
cares?

Now make a change of variables: define ψ′ by

Trade ψ for ψ′:

We have introduced an illusory coupling between ψ′ and ϕ, illusory because it doesn’t affect S-matrix elements.
We have cunningly arranged matters so that the only thing that comes into the fermion vertex is + ϕ. We
simplify the Feynman rules by considering not the separate propagators for Aµ and ϕ but the propagator for the
combination, the only thing that enters the coupling to ψ′. What is the propagator for Aµ + ∂µϕ?

The parameters a and b can be anything. Different choices give different Feynman rules but they all give the same
S-matrix. By choosing a and b in two different ways we can make this propagator look considerably simpler.

For example, if we choose

then the last term in the propagator is

exactly canceling the second term in the Aµ contraction. Then the propagator is

(the “Feynman gauge” (31.36); the quotes are because the “photon” has a mass; the superscript P is for Proca).
Of course this massive theory has no actual gauge invariance. Nevertheless, it looks like the Feynman gauge
propagator in genuine gauge invariant electrodynamics with a massless photon, except that its denominator is k2
− µ2, instead of k2. Alternatively, we could choose
That gives

which is the Proca propagator in the “Landau gauge” (31.37). We could even get the Proca propagator in
something like the “covariant gauges” by the substitution

which gives

As ξ → 1, we get the Proca propagator in the “Feynman gauge”; ξ → 0 gives it in the “Landau gauge”.

So we can perform exactly the same transformations in massive electrodynamics as in the massless theory.
We can make the kµkν part of the propagator look like whatever we want by this trick of adding a non-dynamical
field, or by an appropriate generalization of it: we can always get rid of the troublesome kµkν/µ2 term in the case of
a single massive vector.19 This conclusion does not generalize to non-Abelian theories, involving more than one
vector: massive Yang–Mills theories do not smoothly turn into massless Yang–Mills theories as the mass goes to
zero. 20

We’ve simulated a gauge transformation by introducing a new dynamical variable ϕ, but we haven’t changed
the physics, because ϕ is decoupled from everything else. It is only the mass term for Aµ that breaks gauge
invariance. Otherwise we’d get lots of extra terms from the transformation (31.48). If there were not a conserved
current—if the interactions with the Fermi fields broke gauge invariance—we couldn’t make this trick work.

Were ϕ a dynamical variable, we’d have to have both a and b positive to get a physically sensible theory.
Then we could absorb a into the normalization of the fields, and we’d be left with b. If a and b have opposite signs,
the theory contains tachyons; if they are both negative, the propagator has the wrong sign, and we’ve destroyed
the positivity of the inner product in the Hilbert space. But since the ϕ terms are completely decoupled, off in a
world by themselves, these constants a and b can be whatever we want. (We know that to get the theory to look
“Feynman gauge”-like, there must be some pathological degree of freedom in the theory somewhere, because of
the signs of the poles in the propagator: three of them are right, but one is wrong.)

31.6A first look at renormalization in QED

We’ve said practically all that needs to be said about the theory of a single vector field, on a formal level and in low
orders in perturbation theory. We’ll now start tackling the question of renormalization. Let’s restrict ourselves to
genuine electrodynamics, the theory without a photon mass term.

You might think that we had taken care of all the problems associated with renormalization because we’ve
now put our propagators, in the Landau gauge or the Feynman gauge, into a form with the same high-k behavior
as ordinary spinless boson propagators. They go like 1/k2, whether you’re in the massless theory or the massive
theory. Since Lorentz invariance was not involved in the statement of the BPHZ theorem,21 we could apply it and
treat the four components Aµ as four scalar mesons with 1/k2 propagators and one funny sign: gµν/k2 has four
components with 1/k2. The individual components would couple to γµ’s in non-Lorentz invariant ways, but who
cares about that?

We could simply try the straight BPHZ renormalization procedure, starting out with renormalized (primed)
fields:
and the Lagrangian written in terms of these fields,

and generate counterterms. (As usual, e is the physical charge, however we choose to define it—say à la BPHZ at
zero momentum transfer. Likewise m is the physical mass, defined perhaps according to BPHZ.) The interaction
is of renormalizable type, of dimension 4, exactly like a scalar meson Yukawa coupling. The only difference from
scalar renormalization is that because Aµ has an index, Lorentz invariance will allow us to write down some more
counterterms. We’ll just write them down to see what the possible counterterms are, as given by the BPHZ
procedure. And then we’ll discover something disgusting.

Table 31.1: Possible counterterms in massless electrodynamics

Table 31.1 is the complete set of possible counterterms with dimension ≤ 4 (and consistent with the invariances of
the theory) according to the BPHZ procedure; {A, B, . . . , G} are constants. (QED is parity invariant, so we’ve
excluded ′γ5ψ′ and ′γµγ5ψ′Aµ′.) These terms are just what conventional reasoning gives you, similar to
meson-nucleon theory (§23.1; the constants are introduced in (23.25)). Let’s talk about these terms. Some of them
are disastrous.

If we generate A(A′µ)2 or G(A′µA′µ)2 terms, we’ve destroyed the gauge invariance of the theory! But it was only
the gauge invariance of the theory that told us both that the generating functional is independent of ξ, and that the
Z obtained from the Faddeev–Popov theory was equivalent to the Z found with canonical quantization in the axial
gauge. If we don’t have gauge invariance, we don’t have ξ-independence and we don’t have equivalence to
canonical quantization. So all hell can break loose if these counterterms are present. We don’t know if our theory
is physically sensible if we have non-gauge invariant terms proportional to A and G. It may be some theory with
ghosts that are physically observable or lack of conservation of probability, or some other incredible nonsense. So
it is absolutely essential that

because those terms break gauge invariance. If those terms are generated the whole thing goes into the trash;
well, we had a nice formal structure, but if it’s not preserved by renormalization, forget about it…

It would be very nice, although not absolutely essential, if we had

because the ratio C/D determines what the bare charge is, aside from factors that depend on A′µ, and A′µ is the
same for the electron and the proton. Of course we know empirically that

to many decimal places. In QED, with electrons only, it wouldn’t matter. On the other hand, suppose we have a
more complicated theory, describing not only electrons but also protons and π mesons and all their strong
interactions. We know that the interactions of the proton and those of the electron are very different. If we didn’t
have a rule like (31.62), then we’d be in a peculiar situation. We would find the electric charge renormalization for
the proton and that for the electron would be different, and we would have to say that the empirical equality of the
electric charges is a coincidence. The bare charges would be completely different, by amounts that depend on the
cutoff in the strong interaction coupling constant, and God knows what else, but apparently they have been so
cunningly adjusted by God while creating the universe as to make the size of the physical charges exactly equal.
Who would believe that? Well, it’s a possibility; maybe God is that nice. Although at present we cannot insist that
C = −D, let’s remember that this equality would be very helpful if we are to explain the universality of electric
charge, that the equality of the physical charges for two different particles should imply equality of the bare
charges.

Finally, we also have, though it’s not worth much, that

That is, there is no renormalization of the gauge parameter ξ; we don’t need to introduce a counterterm for the ξ
term.

Well, are these conditions (31.61)–(31.63) on {A, C, D, F, G} true, or not? In fact, we will be able to prove,
order by order in renormalized perturbation theory, that if our conditions

are true to a given order, then they are true in the next order up. Thus we will show the consistency of the
renormalization program with gauge invariance. The tools we will use to prove these are the Ward identities.22 I’ll
briefly tell you what these things are.

Let’s focus on (31.62). If it is true, it implies that there is some connection between the graph in Figure 31.4,
which tells us the D-type counterterm, and the graph in Figure 31.5, which tells us the C-type counterterm. We will
show, in the old-fashioned way, that there is some kind of connection between those things. Define

Figure 31.4: The electron-photon vertex

Figure 31.5: The electron propagator

At equal times

Consider the divergence of the time-ordered product

Recall the rule (27.86) for finding the time derivative of a time-ordered product: we get equal-time commutators
plus a differentiated term. The differentiated term is irrelevant because of current conservation,

so we only pick up the equal-time commutators, when x0 = y0 and x0 = z0. We get

The space δ function comes from the commutator and the time δ function from differentiating the θ function.23 In
the next lecture we will do things in a quite different way.

But we can begin to see how we can establish a connection between some part of Figure 31.4 and some part
of Figure 31.5. In some sense jµ is the thing the photon is going to couple to when it burrows into that diagram and
hits its first Fermi line. The first thing it’s going to do is hit a Fermi line and then it’s coupled to j µ. So in some sense
Figure 31.4 is connected with the left hand side of (31.69). On the other side of the equation we have known
quantities times the fermion two-point function, which is Figure 31.5. But it’s written in a terribly messy form and it’s
sort of awful. I haven’t really got a Green’s function because we need to take off the photon line to get to the jµ.
We’ve derived an equation for the full Green’s functions but renormalization is expressed in terms of 1PI Green’s
functions. We would have to manipulate and manipulate and manipulate this thing to prove what we eventually
want to prove, which is C = −D. And then we’d have to write down a whole bunch of other messy equations and
manipulate them to show the other things we want to prove, A = F = G = 0. So we won’t do it this way. It’s also a
mess because we’ve written (31.69) in terms of unrenormalized fields, and we have to figure out how to write it in
terms of renormalized fields, which is ugly. Instead, we will use a method based on functional methods; we will
essentially read off these equations and all the consequences of things we get by the above manipulations without
going through any combinatoric work.

As a preliminary for that, we must further develop the functional integral method. In particular, we need the
generating functional for 1PI diagrams in terms of the generating functional for full Green’s functions. We will do
that next time.

1[Eds.]
See note 22, p. 655, and note 10, p. 1037. The Faddeev–Popov technique is discussed in every modern
QFT textbook. See Peskin & Schroeder QFT, Section 9.4, pp. 294–298.
2[Eds.]Coleman is using some concepts from differential geometry: fiber bundles. Connections over vector
bundles (analogous to Christoffel symbols in general relativity) are another way of viewing Yang–Mills fields. See
Sections 18.1c–18.2c, pp. 479–488 in Theodore Frankel, The Geometry of Physics: An Introduction, 3rd ed.,
Cambridge U.P., 2012.
3[Eds.] The logic and the associated procedure in going from (31.9) to (31.10) is much the same as that used to
connect both sides of (1.55). See footnote 8, p. 9.
4[Eds.] For a complete justification of the Faddeev–Popov ansatz, including the important demonstration that the
determinant Δ is itself gauge invariant, see Ryder QFT, Section 7.2, pp. 245–255.
5[Eds.] See the discussion on p. 586 about the collision between canonical quantization and gauge invariance.
“Choosing a gauge” is really shorthand for “choosing a gauge condition”, but this is the language used.
6[Eds.]
Axial gauge is also called the Arnowitt–Fickler gauge. R. L. Arnowitt and S. I. Fickler, “Quantization of the
Yang–Mills Field”, Phys. Rev. 127 (1962) 1821–1829.
7[Eds.] Often called the “Lorentz” gauge in the literature, but this is a misnomer. See footnote 1, p. 153.
8[Eds.] Coleman calls the homogeneous wave equation “the d’Alembert equation”.
9[Eds.] The variable ξ, commonly used for this purpose in the literature, has been substituted for Coleman’s
original α, which is easily confused with the fine-structure constant. See “Generalized Renomalizable Gauge
Formulation of Spontaneously Broken Gauge Theories”, Kazuo Fujikawa, Benjamin W. Lee, and A. I. Sanda,
Phys. Rev. D6 (1972) 2923–2943.
10[Eds.] After Fourier transforming the operator between the A’s in (31.33), it can be written as −k2PTµν −
(k2/ξ)PLµν, with the projection operators in (30.46). The inverse of this operator is –(1/k2)PTµν − (ξ/k2)PLµν. By
convention, this inverse is multiplied by i, and +iϵ is added to the k2 denominator.
11[Eds.]R. P. Feynman, “Space-Time Approach to Quantum Electrodynamics”, Phys. Rev. 76 (1949) 769–789.
Reprinted in Schwinger QED.
12[Eds.]
L. D. Landau, A. A. Abrikosov and I. M. Khalatnikov, “Асимптотическое Выражение для Гриновской
Функции Электрона в Квантовой Электродинамике”, (An Asymptotic Expression for the Electron Green’s
Function in Quantum Electrodynamics) Dokl. Akad. Nauk SSSR 95 (1954) 773–776; L. D. Landau and I. M.
Khalatnikov, “The Gauge Transformation of the Green Function for Charged Particles”, J. Exper. Theor. Phys.
USSR 29 (1955) 89–93; English trans. Soviet Physics JETP 2 (1956) 69–72. Republished in Collected Papers of
L. D. Landau, ed. Dirk ter Haar, Pergamon Press, 1965, pp. 659–664.
13[Eds.]
H. M. Fried and D. R. Yennie, “New Techniques in the Lamb Shift Calculation”, Phys. Rev. 112 (1958)
1391–1404.
14[Eds.]
These authors did not, of course, name these gauges after themselves. Walter Heitler first introduced the
name “Lorentz gauge” (not “Lorenz gauge”, though it should have been) and “Coulomb gauge” in his influential
text, The Quantum Theory of Radiation, 3rd ed., Oxford U. P., 1954, p. 3; reprinted by Dover Publications, 1984.
Bruno Zumino gave the names “Feynman”, “Landau” and “Yennie” to the respective gauges: B. Zumino, “Gauge
Properties of Propagators in Quantum Electrodynamics”, J. Math. Phys. 1 (1960) 1–7. For more about these
gauges, their properties and relations to each other, see J. D. Jackson and L. B. Okun, “Historical Roots of Gauge
Invariance”, Rev. Mod. Phys. (2001) 73, 663–680, and N. Nakanishi, “Indefinite-Metric Quantum Field Theory”,
Suppl. Prog. Theor. Phys. 51 (1972) 1–95. Incidentally, Okun introduced (1962) the term “hadron” (from the Greek
 , “stout, fat, strong”) as the antonym to “lepton” (Greek , “slight, thin, small”).
15[Eds.] See §29.5, and Section 5.4 of “Secret Symmetry”, in Coleman Aspects.
16[Eds.] See P. H. Frampton, Gauge Field Theories, 3rd ed., Wiley-VCH Verlag BmbH, 2008, Section 2.3, pp.
59–65.
17[Eds.] A good summary of the problems encountered in the canonical quantization of QED can be found in K.
Haller and E. Lim–Lombridas, “Quantum Gauge Equivalence in QED”, Found. of Phys. 24 (1994) 217–247.
18[Eds.] In Brian Hill’s notes for Feb. 12, 1987, Coleman credits this trick to Stueckelberg, and following him, the
field ϕ is denoted B: E. C. G. Stueckelberg, “Die Wechselwirkungskräfte in der Elektrodynamik und in der
Feldtheorie der Kernkräfte (Teil II und III)” (Forces of interaction in electrodynamics and in the field theory of
nuclear forces (parts II and III)), Helv. Phys. Acta 11 (1938) 299–328; reprinted in E. C. G. Stueckelberg: An
Unconventional Figure of Twentieth Century Physics, J. Lacki, H. Ruegg, G. Wanders, eds., Birkhäuser, 2009;
English translation by D. H. Delphenich, online at http://neo-classical-physics.info/electromagnetism.html.
19[Eds.]See Problems 16 and their solutions, pp. 635–639; the kµkν part of the propagator does not contribute to
the amplitudes involving a massive vector.
20[Eds.] See note 22, p. 1044.
21[Eds.]
See §25.2, and “Renormalization and symmetry: a review for non-specialists”, Ch. 4, pp. 99–124 in
Coleman Aspects.
22[Eds.] What Coleman calls “the Ward identities” are conventionally known as “the Ward–Takahashi identities”: J.
C. Ward, “An Identity in Quantum Electrodynamics”, Phys. Rev. 78 (1950) 182; Y. Takahashi, “On the Generalized
Ward Identity”, Nuovo Cim. 6 (1957) 371–375; Coleman Aspects, “Symmetry and symmetry-breaking: currents”,
Section 5.4, pp. 108–111; Peskin & Schroeder QFT, “The Ward–Takahashi Identity”, Section 7.4, pp. 238–244.
23[Eds.] Bjorken & Drell Fields, Problem 19.1, p. 376.

Problems 17

17.1 In Model 1 (§8.5), we were able to calculate (8.61), the vacuum-to-vacuum S-matrix element for a free scalar
field linearly coupled to an external source J(x). We now know that the reason we were able to do this so easily
was that we were evaluating a Gaussian functional integral. But we also have a Gaussian functional integral if the
source is coupled quadratically to the field.

Consider

With this Lagrangian, there is no linear term in the exponential, and so the functional integral for Z[J] produces only
the inverse square root of the determinant of the quadratic operator A,

From (13.6), (28.27) and (28.50), we can derive that

(Here I’ve used the fact that the matrix element is 1 when J = 0 to fix the normalization factor; I’ve also dropped
some factors of i that cancel between numerator and denominator.) Show that this is the same as the answer you
get by summing Feynman graphs.

If we had a complex scalar field (with coupling Jϕ *ϕ), functional integration would give us the square of
(P17.3). Can you see where the difference is in the Feynman graphs?

H INTS:
 (a) I think you’ll find it more convenient to write out the Feynman graphs as integrals in position space rather
than in momentum space.

(b) You’ll never get the right answer if you don’t get the symmetry factors right.

(c) For any diagonalizable matrix A,

This can be extended to appropriate complex matrices by analytic continuation.1

(1998b 4.1)

17.2 As discussed in §31.2, the Coulomb gauge (or “the radiation gauge”) is defined by the gauge-fixing condition

Just as for the covariant gauges discussed in §31.3, the Faddeev–Popov determinant Δ is a constant which we
can absorb into the normalization of the functional integral. Show that in this gauge, the i–j part of the photon
propagator is

where, as usual, i and j are spatial indices. Compute also the i–0 and 0–0 parts of the propagator.
(1998b 5.1)

17.3 A Dirac electron is minimally coupled to a massless photon with coupling constant e. Compute to O(e2) the
invariant Feynman amplitude for electron–electron scattering in both the Coulomb gauge and the Feynman
gauge, and show that the final answers are the same. H INTS: Use the fact that the spinors obey the Dirac
equation.

Remark: This is an extension of a computation that will be done in the lectures (§34.1) for the vacuum-to-vacuum
amplitude in the presence of an external c-number source.
(1998b 5.2)

1[Eds.] See note 7, p. 608.

Solutions 17

17.1 We start with the functional integral result,

Using the identity (P17.4), we have

Since

we find
Writing this out in x-space, we obtain

because (−□x 2 − µ2)(−iΔF(x)) = δ(4)(x).

Now consider the Feynman graph calculation with

Recall (13.11),

where W[J] is the sum of connected vacuum graphs. Any connected graph for this theory can be described as a
simple loop of n alternating J interactions and ϕ propagators. The diagram for n = 5 is shown in Figure S17.1. At
order Jn, the corresponding amplitude is

where N D is the number of distinct diagrams.

We can generate all diagrams of this pattern by permuting the n vertices and/or switching the two ϕ fields at
each vertex. If each of these operations were to yield a distinct diagram, we would obtain 2nn! diagrams,
canceling the 2nn! in the denominator. But our diagrams have an n-fold cyclic rotational symmetry and mirror
symmetry (switching all pairs of ϕ’s at each vertex simultaneously and arranging the vertices in reverse cyclic
order). The number of distinct diagrams at order Jn is therefore

We conclude that once again,

identical to the functional integral result.

If we now consider complex scalar mesons, the only difference in the Feynman graph calculation is that each
propagator carries a definite direction (one vertex contributes the ϕ and the other contributes the ϕ *). This
eliminates the mirror symmetry of the real scalar theory, and the factor of disappears from the coefficient of the
integral:

Expressed as determinants,

the exponent of − is replaced by −1.

Figure S17.1 Graph for scalar field LI = Jϕ 2, n = 5

17.2 We begin with the generalized Coulomb gauge constraint

Under an infinitesimal gauge transformation

Then for the gauge-fixing function

and

This determinant is a constant (that is, it is independent of the fields over which the integration is performed) so it
can be absorbed into the normalization of the functional integral.

As in §31.3, we integrate (31.31) over all possible f(x) with weight

This generates an effective action of the form

after an integration by parts, where

In momentum space we find

Using rotational invariance, we can assume that k points in the x-direction: k = (|k|, 0, 0). We then have

where N and R are 2 × 2 matrices, and 0 is a 2 × 2 zero matrix;

The propagator is

Because of the block structure of µν, it’s easy to invert:

and

where ΔN is the determinant of N:

Then
 In analogy with the passage to the Landau gauge (31.37), we now obtain the Coulomb gauge propagator by
taking the limit ξ → 0. In this limit

For a general momentum k, we have

The 0–0 component is the Fourier transform of the static Coulomb potential (see §9.3).

Alternatively, we can use 3-space projection operators, analogous to (28.91) and (28.92), to find the
propagator:

Expressed in terms of these,

and so

Taking the limit ξ → 0, we again obtain (S17.7c).

We can express the propagators in a slightly more covariant-looking way. We replace the Coulomb gauge
condition (S17.1) with

where

The matrix µν is

Inverting this and setting ξ to 0 again we find the compact form

The individual matrix elements of this reproduce (S17.7).2

17.3 The diagrams for lowest nontrivial electron–electron scattering are shown below:
 The amplitude for these together is

where the currents are defined as

and the momenta are

The minus sign between the two contributions comes from the fermion interchange rule (see §21.4). The
propagator Cµν in Coulomb gauge is given by (S17.7). In Feynman gauge,

The currents are conserved because the spinors obey the Dirac equation. For example, consider J11µ:

and the same is true for the others:

Then

Using these relations, we have

and similarly

Then the amplitude for electron–electron scattering in Feynman gauge is given by

While we should expect that the photon propagator is gauge-dependent (it is, after all, the Fourier transform of
á0|T(Aµ(x)Aν(y))|0ñ, and Aµ is gauge-dependent), this gauge dependence does not affect a scattering amplitude in
a concrete calculation.

2[Eds.] This form is mentioned in S. Pokorski, Gauge Field Theories, 2nd ed., Cambridge U. P., 2000. Bjorken and
Drell have a form similar to (S17.12), but with an additional k2nµnν term in the numerator of the second part:
Fields, equation (14.54), p. 79. This cancels the static Coulomb interaction, rendering D 00 = 0. They continue with
a discussion about why only the covariant part of the propagator matters.
 32
Generating functionals and Green’s functions

We are going to use functional methods to analyze the structure of Green’s functions. In particular we will discuss
three different generating functionals.1 These will give us

1.Full Green’s functions (given by our old friend, Z[J])

2.Connected Green’s functions

3.One-particle irreducible (1PI) Green’s functions.

The immediate reason for going through this analysis is to derive the Ward identities directly for one-particle
irreducible Green’s functions, and thus to complete the renormalization program for quantum electrodynamics.
The methods are of general utility, however, and applicable in a wide variety of circumstances. There is nothing
especially quantum electrodynamical about them; we could have talked about them earlier. It’s simply that this is
the first time that we’ve encountered a problem of sufficient combinatoric complexity to make it worthwhile to go
through this general derivation.

32.1The loop expansion

To simplify the notation as much as possible, we will conduct the discussion for a theory of a single scalar field ϕ.
In a more general theory the only alteration is appropriately sprinkling the equations with indices and being careful
of the order of terms if one is dealing with Fermi fields, and so has anti-commuting sources instead of commuting
sources. The field ϕ could be the renormalized field or the unrenormalized field; as far as the combinatorics are
concerned, it doesn’t matter. Its dynamics are determined by some action

where J is a c-number function of x. The generating functional for full Green’s functions G(n)(x1, , xn) associated
with this action is 2

N is a normalization factor chosen so that Z[0] = 1:

Finding the generating functional for connected Green’s functions is trivial because of our powerful theorem of
general utility, (8.49), that the sum of all Feynman graphs is the exponential of the sum of all connected Feynman
graphs.3 There are no vacuum-to-vacuum graphs; they are canceled by the normalization factor N. Define the
functional W[J] by

The i is put in by convention, so that all the i’s disappear when we rotate into Euclidean space. W[J] is the
generating functional for connected Green’s functions, Gc (n)(x1, , xn):

where Gc (n) is the sum of all connected Feynman diagram with n external legs. I made this remark much earlier,
when we first introduced Z[J].

Before we go on to constructing the generating functional for 1PI Green’s functions, let’s discuss an amusing
property of W[J]. This property, which can be described as counting loops and counting ħ’s, is known as the loop
expansion4 or the semi-classical expansion. Throughout this course we have been setting ħ = 1 except on
occasions when it is convenient to restore it; this is one of those occasions. To see how to count loops, re-
introduce ħ into the equations, writing

Putting ħ into the action puts ħ into all our propagators and vertices, and therefore puts ħ into every Feynman
graph. Let us count the powers of ħ associated with a given Feynman graph.

Every propagator or internal line will yield an ħ because the propagator is given by the inverse of the quadratic
part of the action, and every quadratic part of the action has a 1/ħ in it. Every vertex yields a 1/ħ because the
vertices are just read off from the non-quadratic part of the action. Therefore the contribution G of an arbitrary
graph with I propagators (internal lines) and V vertices is proportional to ħ(I−V) :

On the other hand, for the contribution Gc of a connected graph, the number of loops L is

This is an old result, but I’ll remind you of the derivation.5 There is an integration for every internal line and an
energy–momentum conserving delta function for every vertex. The delta functions kill all the internal integrations
except for one left over for overall energy–momentum conservation. So the number of free integration variables,
or equivalently the number of loops L, is given by I − V + 1, and the power of ħ in the contribution Gc is

Notice I emphasize “connected”. If the graph were disconnected, for example if there were two components, it
would have an overall energy–momentum conserving delta function for each of its component pieces, and the
formula (32.8) would not be true. Expanding W in powers of ħ is equivalent to expanding W in the number of loops
in the graph. We find

The is a rather peculiar power series. The “no loop” term is called the tree approximation. The name is
borrowed from topological network theory.6 These graphs without loops are called tree graphs, and they are
O(1/ħ). The graphs with one loop are O(1), the graphs with two loops are O(ħ), and so on. Actually, expansion in ħ
is garbage. Although ħ is a quantity with dimensions, we can always choose our units so that ħ = 1. So there’s no
particular reason to believe in an expansion in ħ; the truncated series is not necessarily a good approximation. But
I wanted to point out that if we did expand in ħ, that would be equivalent to counting the number of loops.

Just to give you something definite to look at, let’s look at our old friend ϕ 4 theory. In Figure 32.1 every
external vertex has a J; that’s indicated by the dot. Notice that there are many graphs at a given order in ħ. The
tree approximation itself goes on forever; there are infinitely many vacuum-to-vacuum connected graphs in the
presence of J with no internal loops. Of course, there are only a few 1PI graphs. At the tree level only the graphs in
Figure 32.2 are 1PI. The graph can be cut on the internal line joining the two 4-point vertices. 7 So it would
be even nicer to have a generating functional for 1PI graphs.

Figure 32.1: Expansion in loops is equivalent to expansion in powers of ħ

Figure 32.2: One particle irreducible tree graphs in ϕ 4 theory

 32.2The generating functional for 1PI Green’s functions

Recall that 1PI graphs can’t be split into two distinct pieces by cutting a single internal line. We will define the 1PI
diagrams with certain cunning conventions so that everything comes out right in the end. Then we will show how to
construct the generating functional for 1PI Green’s functions, Γ[ϕ], in terms of objects we already know, Z[J] or
(32.5) iW[J] = lnZ[J]. We will write Γ[ϕ] as a functional of a classical variable ϕ(x), for reasons that will become
clear shortly, rather than in terms of the c-number current J. We will call Γ[ϕ] the effective action. As I will show, in
the tree approximation

This assertion is true, and I’ll back it up with two or three diagrams. That’s not enough, of course. (Let me not
make invidious comparisons with my distinguished colleagues who teach this subject. One calculation done in
detail is worth a hundred general arguments.)

This generating functional has a functional Taylor expansion

just like the functional Taylor expansion of Z (32.2) or W (32.5) in terms of J. The Fourier transforms of the (n)(p1,
, pn) (what we really compute in momentum space when we compute 1PI graphs) are

We include the energy–momentum delta function so that in later manipulations we don’t have to divide it out. For
n ≠ 2 the (n) are defined as:

with all of our usual conventions: no energy–momentum conserving δ functions, no propagators on the external
lines, just as before.8 (The external legs are said to be amputated.) These 1PI Green’s functions are only well-
defined if the sum of the momenta is zero: pi = 0.

For n = 2 we will define (n) in a somewhat peculiar way. Recall that we had defined the full propagator
(15.29) by saying

giving the renormalized or unrenormalized propagator for the renormalized or unrenormalized field, respectively.
We define (2)(p, −p) by

(If the theory involves many fields instead of one, this formula will be expressed in terms of a matrix inverse.) This
is almost our usual definition. In our old definition, i (2) would have been −i ′(p2), the self-energy operator, which
was defined (15.33) as the sum of all 1PI graphs with two external lines. With the old definition, we found (15.36)

obtained from summing the geometric series of 1PI graphs. Thus

The definition of (2)(p, −p) differs from the definition of the other (n)(pi)’s by the addition of the tree-level term (p2
− µ2), a trivial term determined by the free Lagrangian, to the old definition of the sum of the 1PI graphs with two
external legs.
 Now why on earth have we chosen this peculiar definition of Γ(2)? The reason is simple. With these
definitions, if we treat Γ[ϕ] as an honest to goodness action, S[ϕ], for the field ϕ—that would seem to be
dumb—and compute W or equivalently, Z using only tree graphs, deriving Feynman rules in the tree
approximation, forgetting about higher order diagrams—double dumb—then we get the exact W or Z; our errors
cancel each other out! Why does this happen? That’s easy: I’ve cooked it up so that it should happen.

For example, how do we get the propagator in the tree approximation? From the action, we take the
coefficient of the quadratic term in ϕ, Fourier transform it, invert that and multiply it by i, and that’s our answer.
That’s why we defined (2)(p, −p) (32.16) as we did. Using Γ[ϕ] as the action, this procedure results in

which is exactly right.

What about the higher order terms?9 Let’s look at, for example, the exact three-point function (3) in a theory
with a ϕ3 interaction, Figure 32.3. I wrote down a formula like that earlier (16.29). We follow a line through the
graph. There has got to be a place where we can cut each line, as close to the central blob as we can. Everything
between the dot (the J) and the cut becomes the exact propagator. Everything left over is by definition one-particle
irreducible, because it’s the last place we can cut the line. This is the simplest (indeed, it is the only) tree graph in
such a theory: source J—propagator—vertex for all three lines. Likewise, the exact four-point function in ϕ 4 theory
is shown in Figure 32.3. Once again, these are precisely the graphs that would appear in the tree approximation,
except that all the propagators are the exact propagators and all the vertices are 1PI blobs. I’ve got some
disconnected parts, plus permutations, and those are of course what I would get in the tree approximation in a
theory in which a shaded blob with only two legs is the exact propagator. Then there will be graphs where I can cut
to put two lines on one side and one on the other, so there’s an intermediate line. And there will be graphs where I
cannot cut to produce two lines on one side and one on the other. If we treat the 1PI graphs as giving us effective
interaction vertices, then to find the full Green’s functions we only have to sum up tree graphs, never any loops,
because all the loops have been stuffed inside the definition of the propagators and the 1PI graphs. This
marvelous property of the 1PI graphs is important. Taking the 1PI graph generating functional for a quantum
action enables us to turn the combinatorics of building up full Green’s functions from 1PI Green’s functions into an
analytic statement, and we end up with the correct expressions for the full Green’s functions. We’re turning a
topological statement of one-particle irreducibility into an analytic statement that we will find easy to handle. From
this point of view it’s very simple to see the rule that connects Γ[ϕ] to W[J]. From the loop expansion (32.10)

Figure 32.3: Exact three-point function (3) in ϕ 3 theory

Figure 32.4: Exact four-point function (4) in ϕ 4 theory

Comparing (32.6) and (32.20), we see that Γ[ϕ] has the same general structure as S[ϕ]:

In the tree approximation, we drop the second term on the right, and the assertion (32.11) is proved.

Let’s imagine a reader who has read this far and dutifully worked all the problems, but who is nonetheless
slightly confused at this point. This is precisely how we would compute the generating functional if we treated Γ[ϕ]
as a quantum action. The term with no loops, the tree graphs, gives the correct W[J]. This is simply the analytic
prescription for the statement about the marvelous property of 1PI Green’s functions. The terms of O(1), O(ħ) and
so on, give us the graphs with loops, which don’t appear in the statement; it contains only tree graphs. We’ve
converted a topological statement, that we have the generating functional for 1PI graphs, into a analytic
statement, and thence, using our previous lore about counting loops and ħ’s, into an equation.

We only want the term in (32.10) which is O(1/ħ). How do we evaluate the leading term in the limit of small ħ?
By the method of stationary phase.10 Moreover, we don’t even need to worry about the determinant in finding the
stationary phase, because that is a term of O(1). All we have to do is look for the point of stationary phase, and
solve the equation11

That is the functional equivalent of (17.37), the variational derivative of the phase in the integral with respect to ϕ.
This determines ϕ as a (nonlocal) function of J: ϕ J . Once we’ve found the point ϕ J of stationary phase, we plug
that value back into the integral and the leading term in the evaluation of the integral is its value at the point of
stationary phase:

This is going backwards; it’s not the result we want. This is the procedure for constructing W from Γ, while we want
to construct Γ from W. However, (32.23) has the form of a Legendre transformation, going from Γ as a functional of
ϕ to W as a functional of J. It is very easy to invert a Legendre transformation.12 Starting out with W, we get J as a
functional of ϕ:

so that

which determines J(ϕ) ≡ Jϕ . Thus we have two Legendre transform pairs:

The whole procedure for finding the generating functional for 1PI graphs from the generating functional for
connected graphs (I should say Green’s functions, the sum of all graphs) is simply a Legendre transformation. We
differentiate W with respect to J to define the new variable ϕ, then use the equations above. We’ve turned a
complicated combinatoric operation involving getting into the insides of graphs, slicing them this way and slicing
them that way and throwing away graphs if we can cut them in two by cutting a single internal line, into an analytic
operation, simply by doing a Legendre transformation.

There’s one point that will be useful later on. Because of our functional integral, it’s rather trivial to
differentiate W with respect to J. Differentiating (32.4) with respect to J

But from (32.25), it follows that

and so

which could be taken as a definition of ϕ(x). Put another way,

independent of the normalization N. From this equation, ϕ(x) can be thought of as the mean value of ϕ averaged
over function space, with J-dependent measure in function space, d(ϕ)eiS[ϕ, J]. That’s a good reason for calling ϕ
the classical field.13

32.3Connecting statistical mechanics with quantum field theory

There is an amazing parallelism between what we’ve been doing in quantum field theory, with functional integrals,
and statistical mechanics.14 This parallelism is much exploited in mathematical physics, to steal theorems from
one discipline and bring them to the other. In statistical mechanics we have a single variable, β = 1/T, in units
natural for statistical mechanics (that is, the Boltzmann constant kB is set equal to 1). We begin by defining a
partition function, Z(β):

The partition function, the trace over the entire Hilbert space of e−βH, is a nice thing to calculate but it’s not
particularly what we’re interested in. We typically begin by computing the logarithm of the partition function, the
Helmholtz free energy, F:

We differentiate ln Z with respect to β to obtain the first useful property of a statistical mechanical theory, the
average value of the energy, called the internal energy:

We haven’t included effects of the volume or the chemical potential. Now we make a Legendre transformation,
turning from a function of β to a function of E. Define the entropy, S:

There are obvious parallels between these thermodynamic equations and the equations we’ve just derived for the
generating functionals, drawn in Table 32.1.

Instead of a statistical mechanical sum, the trace, we have a sum over function space. And now you can see
why people call it Z. In one case we have an infinite number of variables characterizing our system, J(x) at every
spacetime point x, and in the other case we have a single inverse temperature β. Apart from this, there is a nearly
perfect analogy between the operations in statistical mechanics and those in quantum field theory, which is pretty
much Helmholtz’s statistical mechanics transformed into quantum language. Helmholtz wrote down these
equations except that instead of taking a trace, he integrated over momentum and position, dpdq; he was
summing classical systems over classical phase space. Gibbs’s version includes chemical potentials
characterizing the volume of the system, external magnetic fields, etc. His version has a bunch of parameters in
addition to β, so it looks even more like the quantum field theory version. We can make the analogy look very
close. Once we rotate our fields into Euclidean space, the analogy is perfect, because then we really are summing
over all possible configurations of a classical field of a theory in four space dimensions, just as in classical
statistical mechanics. Then quantum field theory is identical to the classical statistical mechanics of a classical
field in four space dimensions, with ħ playing the role of T and J playing the role of an external field, such as a
magnetic field B, coupled to the system. This is in fact not very useful. All it tells us is that classical statistical
mechanics of a classical field theory in four spatial dimensions is a very complicated problem which we are not
going to solve. However, it is a very useful analogy for people who want to prove things about field theories in
fewer than four spacetime dimensions where the classical system is not so complicated, and rigorous theorems
have been proved. Then they use precisely this analogy to prove rigorous theorems about the corresponding
quantum field theory in 1 + 1 or 1 + 2 dimensions.15

Table 32.1: Parallels between statistical mechanics and quantum field theory

32.4Quantum electrodynamics in a covariant gauge

What are the consequences of this functional formalism for quantum electrodynamics? We can derive important
relations from gauge invariance. All the fields in quantum electrodynamics have well-defined gauge transformation
properties. For an infinitesimal gauge transformation

where δχ is an arbitrary (infinitesimal) function of spacetime. We’ll write this in generic form by assembling all the
fields, including Aµ, into a single big column vector Φ

and under the action of the infinitesimal gauge transformation (32.36),

where A(Φ) is some 3 × 3 diagonal matrix of operators acting on δχ, determined by (32.36). For our gauge
transformation, A(Φ) is no more than first order in Φ, and hence first or zeroth order in the fields Aµ, ψ and ψ. (If we
didn’t have the explicit transformation in hand, we would need to assume that A(Φ) was first order in Φ.) For Aµ it
is the differential operator ∂µ, and for ψ and ψ, it is multiplication by −ieψ or ieψ, respectively. Since we will be
using this for renormalization theory, I should tell you whether the fields and charges are renormalized or
unrenormalized. In fact, it doesn’t matter. These can be the unrenormalized fields and the bare charges. Or, they
could be the renormalized fields to some finite order in perturbation theory. In that case, we are studying the effect
of the gauge transformation on the action to get constraints on the divergent parts of graphs, which we will have to
cancel off with counterterms in the next order.

I will assume that the action S consists of a gauge invariant piece SGI and a non-invariant, gauge-fixing piece
SGF. In a covariant gauge, SGF equals the integral of the four-divergence of Aµ squared, times −1/(2ξ):

When we make this infinitesimal gauge transformation (32.36) on the fields, the gauge invariant part of the action
doesn’t change, and the gauge-fixing term changes according to

where B(Φ) is also first order or less in Φ. Here

The only thing we will use in what follows is that A and B obey the first order conditions as stated, and that the
gauge transformation doesn’t change the measure in function space:

The gauge transformation has determinant 1, since it just adds a constant (that is, independent of Φ) to Aµ, while ψ
and ψ experience a rotation proportional to δχ. The argument will be generalizable to any case obeying these
three conditions: neither A nor B is more than first order in the fields, and the integration measure is invariant.

Why do we want these three things? We want them because it means that if we make this infinitesimal
transformation then the functional integral is not going to change. Using (32.38), (32.40) and (32.41), and
expanding out to first order in δχ,

since the value of an integral is not changed by a change of variables. Then

for arbitrary δχ. That simply says we know how our integrand changes under a gauge transformation. Since we
could interpret a gauge transformation as just a redefinition of our integration variables, it doesn’t change at all.

Now we come to the key point. A and B are linear in Φ. When we have an integral like (32.43), its value is the
mean value of a linear function. Then we can replace Φ by its mean value Φ, and write16

There will be a multiplicative factor of eiW[J], but so what? The whole thing equals zero, so we’ll just take out this
factor. Then (32.43) becomes

This follows from the linearity of A and B. (If either A or B were other than linear in Φ, e.g., quadratic, we could not
substitute Φ for Φ; (Φ 2) ≠ (Φ)2.) We know what J is. From (32.22),

Therefore we have from (32.45)

This equation tells us how Γ[Φ] transforms under gauge transformations. If we write down gauge transformations
for the mean fields Φ that are exactly the same as the gauge transformations (32.38) that we originally had for
quantum fields,

then the change in Γ[Φ], which is a functional of Φ alone, is

the chain rule for differentiation.17 That is, under the gauge transformation (32.47), the effective action Γ[Φ]
transforms as

using (32.46).

Comparing (32.40) and (32.49), we see that

(the last equality following because SGI does not change under a gauge transformation). Under a gauge
transformation, at most linear in the fields, the change in the effective action equals the change in the classical
action.18 From this general relation we will derive many identities. 19 These Ward identities will help establish the
relations (31.64), and hence that quantum electrodynamics is renormalizable. We will call (32.50) the generic
Ward identity.

Let’s review the argument briefly. Given

then we obtain the amazing result (32.50). We can rewrite that result as

By definition, Γ[Φ] − SGF[Φ] must be the gauge invariant part of Γ[Φ]; it doesn’t change under the gauge
transformation. That is,

So the generating functional is gauge invariant except for a gauge-fixing term of the same form as that in the
original Lagrangian. This is not true for the Green’s functions; for example, both the photon propagator and the
electron propagator are ξ-dependent.

Let’s apply (32.46) to spinor electrodynamics in the Lorenz gauge.20 The action consists of a gauge invariant
part SGI, and a gauge-fixing (non-gauge invariant) part SGF:

Under the infinitesimal gauge transformation

we obtain from (32.46), after an integration by parts,

or, since δχ is an arbitrary function,

This equation applies to the entire generating functional Γ[Φ]. It encompasses a large number of equations for the
1PI Green’s functions, which can be derived from the series expansion (32.12).

Next time we will apply the generic Ward identity to the renormalization of quantum electrodynamics.
1 [Eds.] Peskin & Schroeder QFT, Section 11.5, pp. 379–383; Ryder QFT, Sections 6.4–6.5, pp. 196–207.
2 [Eds.] See (13.8). Note from (13.5) that ρ in (13.8) differs from J in (32.2) by a sign.
3 [Eds.] The theorem in (8.49) is written in terms of Wick diagrams, but it holds for Feynman diagrams as well.
4 [Eds.] Coleman Aspects, pp. 135–136; Ryder QFT, pp. 317–8.
5 [Eds.]
This derivation may have been in the lost part of the videotape of Lecture 25. It does not seem to appear
elsewhere in the videos, though it is in Coleman Aspects, Section 3.4, pp. 135–6. See also Peskin & Schroeder
QFT, equation (10.2), p. 316.
6 [Eds.] See, e.g., Figure 1.5.1, p. 12 of R. Diestel, Graph Theory, 3rd ed., Springer-Verlag, 2005.
7 [Eds.] The reader may be wondering about the definition of “one-particle irreducible” (p. 321). The topology of
the first tree graph, , seems to be the same as that of , with one internal line, and yet the former is 1PI,
and the latter is not. Perhaps a better way to think about what makes a graph 1PI is this: if an internal line is
removed, what remains? If there are separate pieces, the graph is not 1PI. In the second diagram, after the
internal line is removed, two pieces are left: it is not 1PI. In the case of the first diagram, nothing is left: it is 1PI.
The 1PI graphs are also called proper diagrams. See, e.g., M. Kaku, Quantum Field Theory: A Modern
Introduction, Oxford U. P., 1993, p. 219.
8 [Eds.] We assume that (0) = 0.
9 [Eds.] In the following we will use these graphical symbols:

10 [Eds.] See (17.36)–(17.39).

11 [Eds.] Using (28.74),

In general, though it may seem counterintuitive, for a given functional F[ϕ],

Consider the differential of a scalar-valued function F(v) with a vector argument:

We have to “sum” over all the “components” of ϕ(x). This requires an integral.
12 The way they invert Legendre transformations in the books would drive you crazy.
13 [Eds.] It will be useful to note that if f(ϕ) is a linear function of ϕ, f(x) = αx + β with α and β some constants, then

14 [Eds.]
F. Reif, Fundamentals of Statistical and Thermal Physics, McGraw-Hill, 1965. In the video of Lecture 32,
Coleman states that the discussion of the analogy between statistical mechanics and quantum field theory was
unplanned, and a few of his equations are erroneous in their factors of β. Those have been corrected here.
15 [Eds.] For examples of the relation between quantum field theory and statistical mechanics, see, e.g., John B.
Kogut, “An introduction to lattice gauge theory and spin systems”, Rev. Mod. Phys. 51 (1979) 659–713; and B. M.
McCoy, “The Connection Between Statistical Mechanics and Quantum Field Theory”, in Statistical Mechanics and
Field Theory, V. V. Bazhanov and C. J. Burden, eds., World Scientific, (1995); pp. 26–128.
16 [Eds.] See footnote 13, p. 694.
17 [Eds.]The videotape of Lecture 32 ends prematurely at this point, at 1:06:19. The lecture may have continued
for another 24 minutes. Judging from the start of the next lecture, however, it appears that little more was added.
The remainder of this chapter is based on notes from Coleman, Woit and the anonymous graduate student.
18 [Eds.] John Preskill, Notes for Caltech’s Physics 205 (1986–7), Ch. 4, pp. 4.55–4.57. On line at
http://www.theory.caltech.edu/~preskill/notes.html.
19 [Eds.] See footnote 22, p. 675.
20 [Eds.] We omit the overbars on the fields here as they would cause confusion: ψ would be ψ, ψ would be ψ, etc.
 33
The renormalization of QED

Last time, in a frenzy of enthusiasm I derived the generic Ward identity, (32.50). To review, suppose S, the
classical action for a gauge field theory describing a set of fields Φ, can be written in the form

Here SGI[Φ] is invariant under the gauge transformation

and SGF[Φ] is a (non-gauge invariant) gauge-fixing term, at most quadratic in Φ. Then the effective action Γ[Φ],
the generating functional for 1PI diagrams, has the same structure as the classical action:

with the same gauge-fixing term, but written in terms of the mean, or “classical”, field Φ (32.31). The fruit of the last
lecture is the statement that, if Φ is subjected to an infinitesimal gauge transformation of the form which leaves
SGI[Φ] invariant,

then we obtain the generic Ward identity,

provided that δSGF[Φ] is at most linear in the fields. This allows us to replace its argument Φ with Φ. If you’ve seen
the conventional Ward identity, you may not recognize it in this formalism. I’ll explain the connection, and derive
some other consequences of this result.

Equation (32.50) will enable us to complete the renormalization program for quantum electrodynamics, with
or without a massive photon, by showing that all the required counterterms are gauge invariant. I will now prove
this in detail for QED (or a general theory of the same form) following the BPHZ program (§25.2). Such a theory
includes only gauge invariant and renormalizable interactions (with dimension ≤ 4, as described in §25.4, p. 538),
apart from the gauge-fixing term which is restricted to be no more than quadratic in the dynamical variables.

33.1Counterterms and gauge invariance

The proof proceeds inductively, or more accurately, iteratively, in a sequence of five statements.

1.We assume that we need only gauge invariant counterterms to render the theory finite to O(en),
where e is some coupling constant. (If there are multiple coupling constants, g1, g2, . . . , we assume
that we have only gauge invariant counterterms of O(g1n), O(g2n), …) We wish to prove that only
gauge invariant counterterms are needed to make the theory finite to O(en+1). The assertion is
obviously true for O(e0): no counterterms are needed for n = 0, and so these zero terms are trivially
gauge invariant.

2.For a theory of this type with only renormalizable interactions, the BPHZ algorithm says that, if we’ve
made everything finite to O(en), all divergences to O(en+1) can be canceled by adding an additional
term to the interaction

is a sum of counterterms computed to O(en+1) with divergent coefficients depending on the

cutoff (assuming that there is a suitable gauge invariant cutoff). The functionals are expressed in terms
of the renormalized fields, Φ′. (It doesn’t matter whether we’re using renormalized fields or not in this
argument; how the fields scale has nothing to do with this proof.) According to BPHZ, if we have only
interactions of renormalizable type, then is a polynomial in Φ′ and ∂µΦ′ of dimension ≤ 4. That’s
 simply the general BPHZ result for working in four dimensions. We add the appropriate counterterms
to get rid of all divergences to the next order, and then we iterate.1

3.To O(en+1), adding the counterterms gives a new term to the effective action,

a term of O(en+1) coming from the terms we’ve added to the Lagrangian: they generate 1PI diagrams
just by themselves. And in we simply replace Φ′ by Φ′, just as if it were a term in the free
Lagrangian. However, the counterterms will also appear as internal parts of other complicated
Feynman diagrams once we’ve added these interactions. But since there’s an explicit en+1 in front, the
new term produces an effect of at least O(en+2) in those other graphs, because any complicated
Feynman diagram has at least one vertex in addition to these new vertices we’ve added. Thus (33.6)
should really be written

4.Staring at this formula (33.7), we see that, before we add the counterterms, to O(en+1),

where Γfinite is independent of the cutoff, and so unaffected in the limit as the cutoff → ∞; BPHZ says
that the sum of the terms in Γ[Φ′] is supposed to be finite to O(en). Since we can get rid of all the
divergences by adding [Φ′], which has cutoff-dependent coefficients, all of the divergences must
be of the form [Φ′].

5.Now we use the Ward Identity. We have a gauge transformation

We assume, at the nth step of the iteration, that we need only gauge invariant counterterms to O(en), so
the Ward identity is valid. The Ward identity tells us that the gauge transformation leaves everything in
Γ[Φ′] invariant except for the quadratic term SGF[Φ′]. This term is certainly not divergent; it’s something
like

It doesn’t have any power series expansion in e; it’s a fixed, known quantity. So there is no need to
introduce counterterms for the gauge-fixing part of Γ[Φ′]. If, apart from SGF, Γ[Φ′] is gauge invariant,
then the cutoff-dependent part, the counterterms, must be gauge invariant:2

6. Therefore, only gauge invariant counterterms are needed to O(en+1), and by induction, to all orders.

We see that, to each order in perturbation theory, the generating functional of 1PI graphs is gauge invariant, aside
from the non-gauge invariant terms that are exactly the same as in the classical action. This means that to each
order all of the cutoff-dependent terms are gauge invariant; the need to introduce non-gauge invariant
counterterms never arises. We conclude that all divergences can be removed with gauge invariant counterterms.
In particular, the gauge-fixing term is unrenormalized. As we will see, gauge invariance via the Ward identity
imposes relations among the counterterms.

33.2Counterterms in QED with a massive photon

Let’s do a couple of examples, spinor electrodynamics and scalar electrodynamics.

EXAMPLE. Spinor electrodynamics

From the BPHZ prescription (see §25.2), we have

The first three terms are each gauge invariant; for the last two terms, there is no correction, as we’ll see. This
Lagrangian includes all possible gauge invariant counterterms of dimension ≤ 4 and is therefore strictly
renormalizable. It gives finite answers for appropriate cutoff-dependent choices of the counterterms A, B and C to
every order in a power series expansion in e.

Once we have this Lagrangian, either because it has come down to us from heaven on a golden tablet, or by
the results of tedious labor to some finite order in perturbation theory, we can write it in terms of unrenormalized
fields and define the values of the bare charge (and any other coupling constants, such as λ in ϕ 4 theory), the
bare masses and the gauge parameter ξ. The Lagrangian in terms of unrenormalized fields is scaled so that the
kinetic energies, the derivative terms, are of standard form:

These are the bare masses and charge. As we’ll see shortly, this is also a bare ξ. We also define the quantities
that give us the scale between ψ and ψ′, and between Aµ and Aµ′:

There is a Z1 that occurs in traditional treatments of electrodynamics; we won’t talk about it. We can show Z1 is
equal to Z2 because of gauge invariance.3 These Zi are called the renormalization constants.

I should say something about the gauge-dependence—or, since we’re working in a theory with a massive
photon, the ξ dependence—of these things. As you’ll recall (§31.5), changing gauges (or equivalently, switching
values of ξ, as in §31.3) by introducing an auxiliary field does not require a redefinition of Aµ, though it does
require a redefinition of ψ. So we expect that Z3 likewise would not get redefined, and hence would be ξ-
independent (gauge-independent in the massless case). Z2 on the other hand might well be ξ-dependent; we
don’t know.4

We also expect the bare masses of the particles to be ξ-independent. After all, they’re masses; how can they
depend on the gauge? An interpolating field may depend on the gauge; it may be a different operator in one gauge
from another, but a mass is a mass, a physically observable quantity. By the same argument, the charge should
be ξ-independent. To summarize:

Comment. There’s a curious consequence to the possible dependence of Z2 on ξ. We won’t go through the
derivation of the spectral representation for vector fields; it follows essentially the same line of reasoning as for
scalar fields.5 We can use the positivity of the weight function to show that Z3 ≤ 1, our usual result. On the other
hand, Z2 is defined in terms of the ψ′ field, which is the original ψ field that we started out with in a nice theory with
positive definite metric in Hilbert space, and the usual good properties, times an exponential of this preposterous
field with minus signs in the propagator, negative metric intermediate particles, etc.6 That may spoil the positivity
of the spectral formula for the spinor field. So don’t be surprised if during a computation in one of these gauges
with ξ, Z2 > 1, while for some other value of ξ, Z2 < 1. That’s just because once we’ve made the change of field,
we’ve mixed it up with this auxiliary field so there may be negative weights appearing in the spectral weight
function. (I thought I should mention this, but it’s just a side comment.)

Now let’s work out the relations between the renormalization constants and the counterterms,
renormalization constants by straight algebraic substitution of (33.14) into (33.13). Comparing the transformed
(33.13) with the Lagrangian (33.12), we find
 We have identified two of the equations as being interesting. First,

This implies that as µ2 → 0, µ02 → 0 and vice versa, unless Z3 develops a pole, a rather unlikely possibility. We
may get some logarithms because of those intermediate photons, but a pole is rather strong. If we start with a zero
bare mass for the photon we get a zero renormalized mass, or if we set the renormalized mass to zero we get a
zero bare mass. So the zero mass of the photon is preserved by renormalization.

The second interesting equation is

This is important because it represents the universality of charge renormalization. It was only laziness that kept
us from writing down a theory with many more fermionic fields. However many we might have started with, we
would have discovered that those with the same e would also have the same e0. Some of the subatomic particles
have, besides electromagnetism, strong interactions, and some of them don’t. Consider a theory of electrons and
protons: only the protons interact with mesons via the strong force. But the only renormalization constant that
appears in the charge is Z3, which has nothing to do with the Fermi fields and their renormalization; it concerns the
photon. Though other fields may all have different interactions and therefore different Z2’s, they’re all going to
have the same Z3, because there’s only one photon. Thus if particles have the same bare charge then they have
the same renormalized charge. This is reassuring. I raised this as a question earlier (see p. 675): Why is it that the
proton and the electron seem to have exactly the same physical charge? If their renormalizations were different, if
that of the proton were dependent on the strong interactions, then perhaps God had to have been incredibly kind
to adjust the bare charges in such a way that the physical charges came out to be equal. He didn’t have to be kind,
but merely uncomplicated. If God decreed that the bare charges are equal, then automatically the renormalized
charges are equal. Indeed, if the decree had been that the proton’s bare charge were three times the electron’s,
then the proton’s renormalized charge would come out three times the electron’s; the ratio of the renormalized
charges is the same as the ratio of the bare charges for all n of the fields:

We can imagine God deciding that there will be two charges, the proton’s, qp = e0 and the up quark’s, qu = e0. He
fixes the bare parameters at the Planck length. He lets the renormalization run to our scale, and voilà, the ratio of
up quark charge to proton charge is still 2:3! More dramatically, the electron and the antiproton have the same
charge, even though the electron does not participate in the strong interactions. This is a deeply satisfying result.

We’ve obtained this by an elegant but rather abstract argument, but we can understand it physically. Suppose
we have a particle, say a proton, in a box and we wish to compute the expectation value of the electric charge. We
know how to compute the expectation values of operators from non-relativistic quantum mechanics: we expand
the actual state of the system in terms of the eigenstates of the non-interacting system and evaluate the operators
in that expansion. At first glance that looks very complicated because the proton could be a bare proton with
charge 1 or a bare neutron and a bare π+ meson, also charge 1, etc. But, we really don’t need to know this
expansion because anything that the proton can virtually become is also a system of charge 1, a consequence of
charge conservation. Electric charge differs from pseudoscalar coupling constants, for which the analogous result
is not true. As a consequence of charge conservation anything the proton goes into must have charge 1. It can
never be found as a bare neutron plus a bare π−; therefore the expansion doesn’t matter. The total charge in the
box is 1 no matter what the proton’s wave function is, because it’s just a superposition of things with charge 1, or
more precisely, with charge e0. Though this argument is very comforting, on this level it makes us a bit nervous
because it might indicate that there is no charge renormalization at all. And here we definitely see one, e = Z31/2e0.
Why is that so? It is because we measure the physical charge (we’ll give such a gedanken measurement shortly)
by going far away from the box and looking at the long distance behavior of the electric field. We don’t really
construct a J0 measuring operator and stick it inside the box; we look at the electric field at large distances.

Now in a field theory the vacuum is a dielectric. 7 A dielectric is a material from which, when we impose a
constant electric field, there arises a correction to the expression for the ground state energy of the system, from
the good old Maxwell expression |E|2 to |E|2 plus corrections. That’s the defining characteristic of a dielectric. In
a quantum field theory, if we impose a constant electric field, there are lots of complicated bubble graphs and so
on contributing to the energy, and the vacuum will have dielectric properties. If we put a charge +Q in a dielectric,
as Faraday knew, it is shielded; the amount of the shielding depends on its dielectric constant.8 Imagine a tiny
observer within the dielectric, looking at the electric field some distance away from the charge. That observer does
not see the charge Q that we put into the dielectric, but instead Q′ < Q, because the charge polarizes the dielectric
which in turn shields the charge, as in Figure 33.1.

Figure 33.1: Dielectric screening of a charge, and as seen by an observer in the dielectric

Of course if we are outside the dielectric, the missing charge appears on the surface of the dielectric. But we
are not outside the dielectric; we are not outside the vacuum: we swim in the vacuum as fish swim in the sea.
Therefore we are in the Faraday situation, inside the dielectric, and we see the charge as shielded. This does not
depend on the constitution of the charge we put into the dielectric. It is a universal result that only depends on the
dielectric constant of the medium. We can now interpret Z3 as the dielectric constant of the vacuum. Notice that it
is charge shielding: Z3 < 1 and therefore e < e0. The dipoles don’t align themselves the other way in the dielectric:
Z3 is not greater than 1.

This description of course is just a metaphor. No one would accept that as a convincing argument. But it is
easier to hold in our heads than the long argument we have been running through, which should be convincing. All
of the answers are perfectly standard; you can find them in Bjorken and Drell or Schweber or Lurié or any other
reference.9 The methods, however, are my own.

EXAMPLE. Scalar electrodynamics

We’ve been concentrating on the electrodynamics of charged spinors. We should say a few words about
charged scalars. The whole story is pretty much the same except for a technical detail. There is one additional
gauge invariant counterterm of dimension 4 which might be needed in a charged scalar theory, but is not required
in a spinor theory: the quartic interaction (ϕ *ϕ)2, just as we found in our pseudoscalar Yukawa theory, (25.71). To
generate such a counterterm we need to add a quartic interaction of the scalar particle with itself

to the Lagrangian (30.7) to render scalar electrodynamics renormalizable. What results is

The electrodynamics of charged spinor particles has just one coupling constant: e0. To be renormalizable, the
electrodynamics of charged scalar particles has to include two coupling constants: e0 and λ0; we must include a
scalar self-interaction. We can easily see where it comes from if we consider the scattering diagram in Figure 33.2.
In the Lagrangian (33.18)

Figure 33.2: A squared seagull diagram in scalar electrodynamics

the scalars have a direct interaction with the photons, ϕ ?ϕAµ2. This is an exceptionally simple graph because it
has no derivative couplings. The integral is obviously proportional to

at high k: 1/k2 comes from each photon propagator, and there are no derivatives. This graph produces a
logarithmic divergence and we need the renormalization of λ0 in order to cancel it. That’s not a golden argument
because there are lots of other graphs in the same order, and it requires a little checking to show that they don’t
cancel among themselves; they don’t. For example, we have the graph shown in Figure 33.3 with derivative
coupling. Each vertex

Figure 33.3: A derivative coupling diagram in scalar electrodynamics

contributes a momentum factor, and there are four propagators. At high k, this diagram is proportional to

which is also logarithmically divergent. But these two graphs do not give equal and opposite contributions, and do
not cancel each other.

The need for an additional quartic interaction in scalar electrodynamics is exactly parallel to the phenomenon
we ran up against in our discussion of the renormalization of the pseudoscalar Yukawa theory, where gψγ5ψϕ
was not strictly renormalizable; there we likewise had to add a quartic interaction.10

33.3Gauge-invariant cutoffs

All of these arguments about gauge invariant counterterms make sense only if we have a way of introducing a
high-energy cutoff on the Feynman integrals to regularize the counterterms in such a way that gauge invariance is
preserved. We’ll revisit two methods we looked at in §25.1, one of them only briefly: dimensional regularization, ÿ
la ’t Hooft–Veltman, and Sirlin et al.,11 and an earlier method, the regulator fields of Pauli and Villars.12

Dimensional regularization

As in §25.1, the basic idea of dimensional regularization is to extend the dimensionality of space from four
dimensions to some unspecified number n, not necessarily an integer. The ultraviolet divergences which we
encounter by integrating over all momenta are replaced by singularities related to the number of dimensions
through Γ(z), the gamma function.13 Earlier we established (25.18) in Euclidean space, which we now rewrite as

With n = 4 and α = 2, for instance, the left-hand integral is logarithmically divergent. But with n less than 4 and α ≥
2, the integral is convergent. This suggests that we take

where ϵ is a small positive quantity. (Some authors let n = 4 − 2ϵ.) The logarithmic divergence becomes a pole
arising from the gamma function. Previously we found

(γ = 0. 5772 … is the Euler–Mascheroni constant). From the definition (33.20), we have ϵ = 2 − n, so from
(25.26) we get

Moreover, we have the functional equation

and so
Similarly,

Following ’t Hooft and Veltman, we adopt (33.19) for arbitrary complex n and analytically continue in n. Γ(α −
) has singularities at n = 2α, 2α + 2, 2α + 4, etc. So as long as we stay away from even integers n > 2α, this
expression is well-defined. Instead of letting an auxiliary mass go to infinity (as in the Pauli–Villars method, p. 526)
for n = 4, we manipulate the pole in Γ, doing our renormalization in arbitrary n. Only after we have the expressions
for the graphs in a convergent form (with the poles in (n – 4) canceling) do we let n → 4. A function defined on the
integers can be analytically continued (almost) uniquely in such a way that the analytic continuation is also gauge
invariant.

It’s fairly obvious that dimensional regularization preserves gauge invariance. Dimensional regularization
starts out with formal Feynman integrals for integer dimensions and unambiguously continues them to complex
dimensions, whereupon the divergences become poles. Any property which is true for integer dimensions will
evidently be true for the unambiguous analytic continuation. In particular, gauge invariance does not depend on
the dimension of spacetime; we could write it down in 72 dimensions. The fields may have more indices but we will
still have gauge invariance. There is nothing special about four dimensions. That’s more of a swindle than an
argument, but nevertheless it turns out to be right. Those who want a detailed proof looking into the guts of
Feynman diagrams can go to ’t Hooft’s lectures where he talks about matters of this kind.14 When it comes to
dimensional regularization, either we get arguments that we don’t believe or we get arguments that we don’t
understand. That’s the nature of the subject, I fear. All arguments in this area fall into two classes: those that are
incredible and those that are incomprehensible. These classes are not mutually exclusive.

In general we must keep everything in n-dependent form and take the limit n → 4 only after all the necessary
computations have been performed. For instance, in n dimensions the metric tensor is

From (33.19) we can easily work out other integrals. We have

By symmetry

To craft the Dirac gamma matrices appropriate to an n-dimensional spacetime, we first create Euclidean
metric Dirac matrices (later we will sprinkle in enough i’s to make them obey the Minkowski metric):

We know in four dimensions, . Though the dimension of spacetime is n, we don’t a priori know the
dimension of the Dirac matrices. What is the trace of the unit matrix, Tr(1), in this algebra? Of course,

where γµ is any of the Dirac matrices. We have to consider even n and odd n separately. For even n, define the
set of matrices {ai} by

with the adjoints

and so on. Then the algebra of these matrices is just that for fermionic simple harmonic oscillators:

The union of the sets {ai} and {ai†} forms a Clifford algebra15 of dimension 2(n/2) , in 1-1 correspondence with the
Dirac matrices. The Dirac matrices themselves form a Clifford algebra of dimension 4 × 4. So it seems reasonable
to assign 2(n/2) as the dimension of the Dirac identity in n dimensions:

Note that this reduces to the usual result when n = 4. In fact, we could just as well take16

where f(n) is a smooth function of n, and f(4) = 4.

Nearly all the usual trace theorems generalize readily. For example,

There is one problem in extending the gamma matrices to n dimensions: γ5. Recall that we found ((S11.22), p.
428)

but there is no obvious extension to higher dimensions; the Levi-Civita symbol (with four indices) is specifically 4-
dimensional. While γn+1 can be defined, in analogy with the definition (20.102) of γ5 as i times the product of all
the n gamma matrices, the presence of γ5 in some currents—but fortunately, not in QED’s—leads to anomalies in
those field theories.

In the case of odd n, you treat γ1, . . . , γn-1 as before, with

and the equivalent of γ5 is γn, with

There are two inequivalent choices, connected by parity. For parity conservation, you have to add them together.
Then for theories conserving parity,

Regulator Fields

The method of regulator fields is somewhat more old-fashioned than dimensional regularization. But it’s worth
talking about because it’s cute and easy to show that it works fairly well. Recall that this method was basically very
simple.17 We took a Lagrangian and added to it terms in an extra field ϕ 1:

The term L0(ϕ) is the free Lagrangian (including the gauge-fixing term), L′(ϕ) is the interaction Lagrangian with
counterterms, and ϕ 1 is a very heavy field with mass M; the i in L′ is to give a relative minus sign in ϕ 1’s
propagator. This ϕ 1 is not a physical field but instead another kind of ghost. (If the divergences are very bad, there
may be a need for more than one regulator field.) The result of these additions is to change the propagators from
their usual form, by subtracting a new propagator from the original propagator:

Every time we have a graph with a propagator we subtract an extra propagator with a very heavy mass, and then
all the integrals become convergent. If one heavy mass is not sufficient, we subtract more of them, appropriately
weighted, enough to make all of our integrals convergent:

The cr need not all be real, which gives the relative minus sign in (33.43). The cr can be chosen so that the
propagators vanish as quickly as we want at large k2.

We can use exactly the same trick for the photon, massive or massless, with no problem. I’ll write down the
spinor electrodynamic Lagrangian with the new terms for a massless photon:

That doesn’t affect the gauge invariance in any way; it just adds something to Aµ. L is still gauge invariant under
the transformation

When however, we try the same trick for the charged particles, say the fermions in the theory, we run into
trouble, as Pauli and Villars pointed out. We encounter a lot of graphs that are divergent and only contain fermion
propagators around the loops that are responsible for the divergence, as in Figure 33.4. This one is awful; it’s
quadratically divergent. No matter how nice we make the photon propagator, that’s not going to do anything to the
divergence of that graph. Here’s another example, Delbrück scattering, shown in Figure 33.5. This one is only
logarithmically divergent.

Figure 33.4: Vacuum polarization

Figure 33.5: Delbrück scattering

If we try to handle these divergent fermion loops in the same way as those for the photon or the scalar fields,
say by changing the term ψγµψ that appears in the interaction,

then we’ve made everything finite, but we have also destroyed the gauge invariance and, worse yet, broken
current conservation. The divergence of the cross terms in this object is not zero, as it should be if the current is
conserved; it’s proportional to the difference of the masses of the two fields:

Not only have we broken current conservation, we’ve done it in a disgusting way: the larger we make our cutoff
mass m1, the worse we break it. The derivative of the current is supposed to be zero, but here it is proportional to
the difference between the cutoff mass and the physical mass.

Pauli and Villars thought up a clever trick to take care of this. In their method, we don’t subtract the individual
fields; we subtract, with appropriate coefficients to take care of all divergences, the currents, which certainly
preserves current conservation and gauge invariance. The subtracted terms are just a bunch of fields minimally
coupled. We introduce three regulator spinor fields ψ i,
There are no i’s in the coefficients, as you might have expected from (33.48), and you may be wondering what’s
going on. We choose to give the ψ i both heavy masses and strange statistics: while ψ 2 obeys Fermi statistics, ψ 1
and ψ 3 are required to obey Bose statistics; they are unphysical ghost fields. These ghosts are similar to those
introduced earlier (§29.3 and §31.3), except that here we have spinor fields obeying Bose statistics. Before, we
had scalar fields obeying Fermi statistics. (We’re mad with power; we can do what we want!) The result of these
regulator fields ψ 1 and ψ 3 obeying Bose statistics is that we don’t have the usual minus sign for closed loops in
which they appear. Their loops have a sign opposite to the “real” Fermi field ψ, so that by appropriately adjusting
the ci’s we can make all the divergences cancel. Of course, another result is that this theory is completely
unphysical, with negative energies, but it’s just a cutoff procedure. It may be crazy, but it preserves both Lorentz
invariance and gauge invariance.

The full regulator field prescription for putting in a gauge invariant cutoff goes like this:

(a)For photons, subtract the heavy propagators as described before.

(b)For fermions, subtract the loops of the heavy particles (don’t do a thing with the propagators).

(c)For charged bosons, likewise subtract the loops of the heavy particles (these regulator fields obey
traditional Bose statistics), and do nothing with the propagators.

This is a little bit less clean than subtracting propagators, but it has the great advantage of preserving gauge
invariance. I will ask you to compute the photon self-energy using both of these regularization methods, once for
the spinor case (Problem 18.1) and once for the scalar case (Problem 18.2), for homework. These computations
are actually simple. You will learn things by doing them, and they’re historically important.

Here’s what happens (though the details are left to you). For the box diagram, Figure 33.5, because of the
statistics of the regulator Fermi fields, the divergence is proportional to

which is zero. For the photon self-energy, Figure 33.4, the divergence is proportional to

(note that the summation now goes from zero to 3) where a is a function of the external momenta and various
parameters, and m0 = m, the mass of the actual fermion. Expanding the denominator out in inverse powers of k2,
and taking account of the statistics of the regulator Fermi fields, we get

The first two terms vanish, and we can choose the heavy masses mr2 so that the third sum also vanishes.

The history of the calculations of vacuum polarization is amusing.18 In the late 1940’s, there was a great
problem with the self-mass of the photon. People didn’t have any deep understanding of gauge invariance at all.
Ward had not yet written down the very first Ward identity. Schwinger and Feynman were the only two people who
were able to renormalize quantum electrodynamics.19 They had the great secret. Schwinger was plugging along in
Coulomb gauge because he knew what he was doing there, and Feynman was working in Feynman gauge
because he didn’t care if he knew what he was doing, as long as his answers were consistent. They both knew
however that renormalization shouldn’t require putting in a mass for the photon because that would break gauge
invariance, even if they weren’t quite sure how to formulate gauge invariance precisely. They both found, when
they computed the vacuum polarization graph (Figure 33.4), that they needed a photon mass counterterm. That
caused a great deal of irritation. Both of them made nervous remarks and swept it under the rug and said, “We
have to set the photon self-energy to zero by gauge invariance,” and then quickly went on to computing something
observable. Well, I don’t really know what Feynman said, but Schwinger actually says in one of his early papers,20
“We just set this to zero by gauge invariance. It’s divergent and therefore ambiguous and we set it equal to zero.”
The trouble was that neither Feynman nor Schwinger were using gauge-independent cutoffs. Pauli and Villars
clarified everything by their realization that we could systematically introduce a cutoff procedure in a gauge
invariant way, and then explicitly and unambiguously compute the photon self-energy, to show that it is zero. 21
 33.4The Ward identity and Green’s functions

As noted in (32.56) at the end of the previous chapter, the Ward Identity applies to the entire generating functional
Γ[Φ]. It encompasses a large number of equations for the 1PI Green’s functions, which can be derived from the
generic series expansion (32.12). For convenience, we will (for now) stick to spinor electrodynamics, with a
massive photon:

I have dropped the primes and the bars on the fields. Just remember that we are talking about renormalized mean
fields. We’ll need the infinitesimal gauge transformation, so I’ll write it down again explicitly:

Γ will be gauge invariant except for the integral of the last two terms in (33.54). There are lots of Green’s functions
to worry about, so I will introduce a systematic notation. I will refer to

where n is the number of ψ’s and the number of ψ’s (these are equal unless Γ ≡ 0) and m is the number of Aµ’s.
These Γ(n,n,m) objects depend on a bunch of position variables, and their Fourier transforms (n,n,m) depend on a
bunch of momentum variables. If the Green’s functions involve photons, they will also have tensor indices
associated with the Aµ fields.

For example, the photon propagator (to lowest order, O(e0), and dropping the iϵ) is a generalization of
(31.35):

The quantities in the square brackets are the projection operators PTµν and PµνL defined earlier,

This makes it easy to compute µν(0,0,2) to lowest order. Recalling the definition of (2)

we have

Both the transverse and longitudinal parts have mass µ (though the coefficient of k2 in the longitudinal part is
1/ξ).22 For massive photons the most convenient gauge is the Feynman gauge, ξ = 1, because the pole at k2 = 0,
which would otherwise lead to some mild technical problems, cancels between the two terms. (The Landau gauge
is also nice for certain purposes.)

We wish to study the corrections of O(e) and higher to µν(0,0,2) (k). The way we study those corrections is by
expanding out Γ in terms of the fields, according to (32.12) and (32.13). The term in Γ involving Γ(0,0,2) is bilinear in
Aµ:

We want to see what happens to Γ under a gauge transformation. In particular, when Aµ → Aµ + ∂µδχ(x), because
δAµ is independent of e, the associated change δΓ must be as well:
Applying the infinitesimal gauge transformation to (33.59), we find, after an integration by parts,

from which it follows

or, Fourier transforming both sides,

When we add a divergence to Aµ, δΓ acquires only a contribution from the gauge fixing and mass terms in (33.54).
In position space we add a gradient to Aµ, we integrate by parts, and pick up a term proportional to the divergence
of Γ, which is like kµ in momentum space. Since there’s no change beyond the change in the zeroth order term, the
kernel in (33.59), µν(0,0,2) , must obey (33.63), and

must be just the zeroth order term. Gauge invariance thus forces kµ µν(0,0,2) (k) to be whatever it is at zeroth order
in e, regardless of what else is going on.

Let’s look at this in detail. ′

µν(k) gets modified by the 1PI vacuum polarization graphs µν in exactly the
same way as the scalar propagator ′(p2) (15.36) was modified by ′(p2) in §15.3.23 We expect a similar situation
here, with the following modification: both µν and

Figure 33.6: 1PI vacuum polarization

′ will in general have both transverse and longitudinal parts:

µν

Because the projection operators are idempotent and orthogonal, when we string the ′
µν’s and the µν ’s
together, the transverse parts combine with the transverse parts, the longitudinal parts combine with the
longitudinal parts and there are no transverse-longitudinal cross terms:

Putting these all together we get for the full propagator and its inverse

The Ward identity applies to the full 1PI Green’s function. Substituting (33.58) and (33.68) into (33.63) we find

because kµPTµν = 0. Only the terms proportional to PLµν in (33.58) and (33.68) survive contraction with kµ. We
conclude that

That is, all the possible 1PI graphs beyond zeroth order in e must contribute only to the part that is proportional to
the transverse projection operator, PTµν. This is an important consequence of gauge invariance. Lorentz
invariance by itself tells us that the propagator or its inverse (0,0,2) is the sum of two terms, a complicated function
proportional to PTµν and another such function proportional to PLµν. The Ward identity, on the other hand, tells us
that only the transverse term is corrected; the zeroth order longitudinal term just sits around. Whatever the 1PI
graph is, it will be proportional to PTµν:

We can do even better, because this expression has a pole at k2 = 0, and we shouldn’t expect to find a pole
because we haven’t got any massless particles in this theory of massive electrodynamics. Even if the photon is
massless, we shouldn’t expect to find a pole, because this is a 1PI graph and we’ve taken out the one photon
pole, so we can say

to avoid the spurious pole.

The result (33.72) has the typical form of an equation that we would get if we deduced the Ward identities
from current conservation. To emphasize the result, all the corrections to the 1PI photon self-energy are purely
transverse. From this we can deduce that there is no photon self-mass term: a photon self-mass would require a
correction24 proportional to gµν.

We can also obtain the right-hand side of (33.69), namely

and hence the result (33.70), from our general formula (32.56), modified to handle a massive photon:

We apply (33.74) to (33.59). The first two terms contribute nothing; we are not looking at the fermions. The rest of
the equation gives

Now we take δ/δAν(y) of this expression, and get

which in momentum space becomes (33.73):

Let’s go on and study a more complicated expression. This Ward identity will describe the fundamental three-
point vertex, Figure 33.7.

Figure 33.7: 1PI three-point function

The relevant terms in Γ are

Under a gauge transformation, these two terms mix up among themselves; the terms don’t.

Now we apply an infinitesimal gauge transformation to these two parts together. The first part will produce a
term in ψψ, and so will the second, because Aµ just picks up a term equal to ∂µδχ. The coefficient of ψψ must be
zero, because I know from (32.50) that the result of the gauge transformation on Γ just gives a term linear in Aµ,
with no ψψ term in it:

(33.78)

These are the only terms in ψψ, and as Γ is gauge invariant apart from the terms quadratic in Aµ, this expression
must be zero. 25 Extracting the coefficient of ψ(x)ψ(y)δχ(z) (integrating by parts in the last term)26 we obtain

We have simply applied a gauge transformation to the action, invoked the Ward identity, and said that the gauge
transformation can have no effect on this term.

This is, by the way, strikingly similar in structure to (31.69), an equation we found by manipulating the
canonical commutation rules when we differentiated a current and two fields, despite the differences: Γµ(1,1,1) is
not a current; it’s not a full Green’s function, but a 1PI Green’s function; it doesn’t involve unrenormalized entities,
but rather renormalized entities. Aside from these differences, it’s the same equation.

Traditionally, (33.79) is derived in terms of currents;

(We talked about this earlier; see the discussion on p. 675.) The equations of motion (27.10) give

and the current is conserved:

From the canonical commutation relations we can find

and a similar equation follows for ψ with the sign changed. Then

This looks a lot like equation (33.79). The traditional derivation, even when done carefully, is shorter than the
modern version, starting from (32.50). But it has disadvantages. First, it is couched in terms of unrenormalized
fields. Next, the Green’s functions are neither full Green’s functions, nor 1PI functions, and it’s hard to keep
straight charged scalars and charged spinors.

Again, we can get (33.79) from the general relation (33.74). Take the (left) ψ(x) derivative,27 the (right) ψ(y)
derivative and set all the remaining fields to 0:

This equation is true in general. Applying it to (33.77), we obtain (33.79) once again:

In momentum space, µ(1,1,1) ≡ µ is the Fourier transform of the 1PI three-point function, Figure 33.7. Assign
the momenta more conventionally (instead of having all the momenta going in): p′ = p + k. Then

We already have a notation for the function (1, 1, 0) , the inverse of the renormalized electron propagator:
We are going to Fourier transform (33.79). It’s tedious to do this by hand, but it’s easy to see what the general
structure will be, so I’ll just write down the answer. The only interesting question is what happens with the delta
functions in (33.79). We know that

Therefore the delta functions will give terms where the momentum carried by the propagator is either p, equal to p′
− k; or p′, equal to p + k, depending on which delta function we’re integrating over. Instead of performing the
Fourier transforms, we can just guess the result, though we might get factors of i wrong:

We’ll check our guess by demanding that the equation be right to first order in e:

If we substitute these values into (33.88), they give a correct equation linking p, p′ and k:

So as it stands, (33.88) is correct; it is in fact the original Ward identity.28 Diagrammatically (33.88) is

We can immediately write down two consequences of this relation.

We obtain the first consequence by differentiating (33.88) with respect to kµ at kµ = 0, with p fixed. Then ∂/∂kµ
= ∂/∂p′µ, and

so that29

That is, inserting a very soft (zero-momentum) photon into an electron line is equivalent to differentiating the
inverse of the electron propagator. Thus Γµ for a zero-momentum photon is known completely in terms of the
electron propagator. This is an amazing result. What a surprise! It just comes out of gauge invariance, out of the
Ward identity.

33.5The Ward identity and counterterms

The second consequence answers the question “How are the BPHZ-renormalized quantities e, ψ′ and A′µ, related
to the renormalized and experimentally measurable physical quantities?” We can use the original Ward identity
(33.88) to derive the remarkable result that

This is only true when the photon’s mass is zero, i.e., µ2 = 0. The physical charge is defined by the condition

where p, p′ and k are on the mass shell,

We usually need the four-momenta to be complex to satisfy these four conditions, but when µ2 = 0 we need only
these conditions:
We expand ′−1( ) in powers of ( − m)
F

The Ward identity says

Substituting (33.96) into the Ward identity, we get

Sandwiching this equation between u′ and u, we get, using (33.95),

because ( − m)u = 0. Therefore

Neat! This would not be true for massive vector boson theory. For µ2 ≠ 0

(It can be shown that terms O(e3BPHZ) are also O(µ2/m2).)

We can see how (33.93) also leads to an earlier result (31.62) that we got (much more cheaply) about the
conspiracy of counterterms. If we wish to compute the counterterms to some order in perturbation theory, we write
′−1( ) in a power series
F

Likewise, (33.89)

B fixes the ψi ψ counterterm, and Ceγµ fixes the eψ ψ counterterm.

Applying the identity (33.93) to these expansions, we have at p = 0

Thus B and C are connected in exactly the way our earlier argument 30 said they would be: B = −C. It’s much more
complicated to establish this through the Ward identity than with our earlier method. Nevertheless, it’s reassuring
to see it done another way, as here. It also tells us something else that the previous derivation did not. That is, we
could obtain the exact same relationship between the two counterterms if we didn’t renormalize at 0 for the
electron but put the electron on the mass shell. If we put the electron on the mass shell and defined our coupling
constants that way, the corresponding equation would not be (33.102), but instead

and a corresponding equation for Γµ as a power series in − m. We would of course find exactly the same thing
by applying the Ward identity, now not at = 0 but at = m. The differentiations are step-by-step the same.

So we can preserve our subtractive procedure even if we put the electron on the mass shell, rather than
putting it at the BPHZ point of zero momentum transfer. We still have a perfect matching between the
counterterms required of the charge renormalization type, like C, and those required of the electron wave function
renormalization type, like B. On the other hand, as a very important point, even if the photon has a mass, we still
have to keep the photon at zero momentum transfer, because (33.93) is true only when kµ = 0. We can, with no
loss and perfect matching of the counterterms, keep the electron on the mass shell instead of at the BPHZ point.
But the photon, whether it’s massive or massless, has to be kept at zero momentum transfer to get that perfect
matching.

Of course the divergent parts of the counterterms will still match, since the question of what counterterms we
have to add to the Lagrangian to purge the answer of divergent quantities is independent of what subtraction point
we use, and how we parameterize the theory after we’ve gotten rid of the infinities. But in general, the finite parts of
the counterterms, e.g., the coupling constants, will be different if we have a subtractive renormalization scheme
where all the particles are on the mass shell, unless the photon has mass zero.

This has definite physical consequences. Some people, for example J. J. Sakurai, said that the ρ meson was
just like a photon except that it was heavy, and it coupled to the isospin current instead of the coupling to the
electromagnetic current.31 Sakurai’s theory of strong interactions was a minimally coupled theory, with certain
complications caused by the non-Abelian nature of the isospin group. In this theory, the ρ0 meson couples to the I3
current instead of the electromagnetic current, and it has a strong coupling constant, of order 1, rather than of
order 1/ . Otherwise, he said, everything was exactly the same. Similarly he wanted the ω meson to be
coupled to the hypercharge current, in effect having two photons, one which is strongly interacting and massive
called the ω, and one which is weakly interacting and massless, the real photon. Therefore we get universal ω
coupling, just as we get universal photon coupling. But we only get universal ω coupling when the ω is
extrapolated to zero momentum transfer. As the mass of the ω is 782 MeV, that’s an extrapolation of nearly 0.8
GeV, which is a long way off the mass shell, especially for the strong interactions. So the idea was hard to check
even if we could compute the ω − NN coupling constant fωNN, which is not easy. (The ω particles are not
particularly stable, though they’re more stable than most.) Even if we could have compared fωNN with, say, the ω −
ππ coupling constant fωππ, and we found that they were 40% off, one from the other, Sakurai would still have been
happy. He would just say, well, that’s the error we make because we’re extrapolating from a physical ω on the
mass shell down to zero momentum transfer. If we have a real massive photon, the consequences on physically
observable quantities are hard to check, unless the coupling constant is weak, where we can check everything by
doing perturbative computations. And that’s what we’ll do next time, when I finally get to the anomalous magnetic
moment of the electron.

1 There is a little technical sticking point here. We add the counterterms by doing a power series expansion about
the point 0. If we are considering electrodynamics with a massless photon, there is the possibility of singularities at
the point 0: that value sits on top of the photon mass shell. We won’t worry about that here, but assume that we
give the photon a small mass and afterwards consider the limit as the mass goes to zero.
2 [Eds.] The audio of Lecture 33’s videotape is unintelligible from 17:45 to 23:55. The argument has been filled in
from John Preskill’s “Notes for Caltech’s Physics 205 (1986–7)”, Ch. 5, pp. 5.60–5.61, at
http://www.theory.caltech.edu/~preskill/notes.html, and from the anonymous graduate student’s notes. The relevant
sections of Coleman’s own notes are missing, and Woit’s are a little elliptical.
3 [Eds.] Ward’s original goal was to prove the equality Z1 = Z2 conjectured by Dyson, and this is the identity Ward
refers to in his article’s title: J. C. Ward, “An Identity in Quantum Electrodynamics”, Phys. Rev. 78 (1950) 182.
Today the term “Ward identity” usually refers to a preliminary result Ward obtained, from which Z1 = Z2 follows;
see note 29, p. 721. See also Peskin & Schroeder QFT, Section 7.4, pp. 238–244; the original identity and its
consequence are on p. 243.
4 [Eds.] In fact, Z2 is gauge-dependent. See Kenneth Johnson and Bruno Zumino, “Gauge Dependence of the
Wave-Function Renormalization Constant in Quantum Electrodynamics”, Phys. Rev. Lett. 3 (1959) 351–352;
Herbert M. Fried, Modern Functional Quantum Field Theory: Summing Feynman Graphs, World Scientific, 2014,
p. 80; Greiner & Reinhardt QED, p. 298.
5 [Eds.] Bjorken & Drell Fields, Section 16.11. But see §34.1.
6 [Eds.] Crudely speaking, the function χ in the exponent in (31.11) is the line integral of a combination of the Aµ
and A′µ fields. When gauge transformations are evaluated carefully at the operator level, the ψ field picks up an
exponential factor that depends on the non-transverse photon modes. See K. Haller, “Operator Gauge
Transformations in Quantum Electrodynamics”, Nuc. Phys. B57 (1973) 589–603 and “Gauge Problems in Spinor
Quantum Electrodynamics”, Acta Phys. Austr. 42 (1975) 163–214.
7 [Eds.] Peskin & Schroeder, QFT, Section 7.5, p. 255.
8 [Eds.] N. W. Ashcroft and N. D. Mermin, Solid State Physics, Harcourt Publishers, 1976.
9[Eds.] Bjorken & Drell Fields, p. 303, equation (19.33); Schweber, RQFT, p. 634, equation (114) and p. 635,
equation (126); Lurié, P&F, p. 300, equation 6(365).
10 [Eds.] See the discussion following (25.69), on p.540.
11 [Eds.]
See footnote 4 on p. 528. Also see W. J. Marciano and A. Sirlin, “Dimensional Regularization of Infrared
Divergences”, Nucl. Phys. B88 (1975) 86–98; Peskin & Schroeder QFT, Section 7.5, pp. 249–251; and Ryder
QFT, Section 9.2, pp. 313–318. For a detailed review, see G. Leibbrandt, “Introduction to the Technique of
Dimensional Regularization”, Rev. Mod. Phys. 47 (1975) 849–876.
12 [Eds.]
W. Pauli and F. Villars, “On Invariant Regularization in Relativistic Quantum Theory”, Rev. Mod. Phys. 21
(1949) 434–444.
13 [Eds.] Arfken & Weber MMP, Chapter 8, “The Gamma Function”, pp. 495–533.Warning! Don’t confuse the
gamma function Γ(z) with Γ[ϕ], the generating functional of 1PI graphs. The context should make it clear which is
which.
14 [Eds.]See, for instance, G. ’t Hooft and M. J. G. Veltman, Diagrammar, CERN publication 73-9, 1973 (available
at the CERN Document Server, cds.cern.ch) or M. Veltman, Diagrammatica, Cambridge U. P., 1994.
15 [Eds.] See note 1, p. 407.
16 [Eds.] Ryder QFT, p. 333.
17 [Eds.] See p.526, in particular (25.7).
18[Eds.] Schweber QED, Chapter 7, pp. 335–340; Chapter 10, pp. 443–444; Crease & Mann, Chapter 6, pp.
102–108.
19 [Eds.] Tomonaga had also figured it out, but only his colleagues in Japan knew that he had: D. Ito and K.
Nishijima, “Japanese Researchers Reveal Tomonaga’s Path to QED Renormalization”, Letter to the Editors,
Physics Today (51), 7 (1998) 15–16.
20 [Eds.] “If the electromagnetic field is that of a light quantum, the vacuum polarization effects are equivalent to
ascribing a proper mass to the photon. Previous calculations have yielded non-vanishing, divergent expressions
for the light quantum proper mass. However, the latter quantity must be zero in a proper gauge invariant theory.”
Julian Schwinger, “Quantum Electrodynamics. I. A Covariant Formulation”, Phys. Rev. 74 (1948) 1439–1461; see
p. 1440. Wentzel found Schwinger’s claim “highly objectionable”: Gregor Wentzel, “New Aspects of the Photon
Self-Energy Problem”, Phys. Rev. 74 (1948) 1070–1075.
21 [Eds.] See Problem 18.1, p.725.
22 [Eds.] In the video of Lecture 33, at 1:07:33, Coleman writes ξ where (33.57) and (33.58) have 1/ξ. As a
consistency check, the propagator for the case µ = 0 is consistent with 1/ξ; see Coleman Aspects, Chap. 4,
“Secret Symmetry”, p. 164, equation (5.26) or Peskin & Schroeder QFT, p. 297, equation (9.58). (There was also a
sign error in the second term in (33.58).)
23 [Eds.] The 1PI graph in (15.33) is defined as −i ′(k2). The sign difference in the definitions of ′ and ′µν is to
balance a corresponding sign difference between the definitions of ′ and ′µν.

24 [Eds.] Greiner & Reinhardt QED, Section 5.2, pp. 257–258.

25 [Eds.] In fact, the gauge transformation also produces terms proportional to ψ(x)ψ(y)Aµ(z). The point is that the
terms which are proportional to only ψψ, with no Aµ factors, must vanish separately.
26 [Eds.] Use the identity δχ(x) = d4zδ(4)(x − z)δχ(z).
27[Eds.] A note for purists: To keep the signs consistent, all ψ derivatives must be taken from the left and all ψ
derivatives must be taken from the right.
28 [Eds.] See footnote 3, p. 704.
29 [Eds.] This is Ward’s original identity; Ward, op. cit., note 22, p. 675.
30 [Eds.] See Table 31.1, p. 674. The renormalization constants were labeled differently there, Cψ ψ and eDψ ψ
in place of Bψ ψ and eCψ ψ, but their roles within the Lagrangians and their relationship to each other are the
same in both places.
31 J. J. Sakurai, “Theory of Strong Interactions”, Ann. Phys. 11 (1960) 1–48. In the end, of course, Sakurai’s
influential ideas did not provide the framework for a gauge theory of the strong interactions, which was instead
realized in quantum chromodynamics, with massless vectors.

Problems 18

18.1 In the theory of a charged Dirac field minimally coupled to a massless photon, compute the renormalized
photon self-energy, ′µν(k2), to lowest nontrivial order in perturbation theory, O(e2). Write the answer as an integral
over a single Feynman parameter. Handle divergences by the Pauli–Villars method of regulator fields as
explained in §33.3, pp. 712–715. From the Fermi loop integral, subtract identical loop integrals with heavy Fermi
masses times coefficients chosen to cancel both the quadratic and the logarithmic divergences in the integral. Use
the BPHZ procedure to fix the counterterm: choose it to cancel the second-order term in the expansion about k2 =
0. Verify that even before you send the masses to infinity, the Green’s function is proportional to

(kµ is the photon momentum) as the Ward identity tells us it should be; see (33.67) through (33.71).

H ISTORICAL N OTE: As mentioned in Chapter 33, this problem was a famous technical pain-in-the-neck in the late
1940’s. If you just blithely manipulate divergent integrals, it looks like a photon self-mass counterterm is needed.
Pauli and Villars invented their gauge invariant cutoff to show that this apparent contradiction of gauge invariance
is just a consequence of slovenliness, not a sign of deep sickness in the theory. See note 18, p. 715.
(1998b 7.1); historical note from (1987b 9)

18.2 Perform the same computation for a charged, spinless meson, but this time use dimensional regularization
instead of the Pauli–Villars method. Warning: In n dimensions, .
(1998b 7.2)

18.3 Even in quantum electrodynamics, it is possible (though not usual) to work in a gauge where ghost fields are
needed. For example, this is a valid form of the electrodynamic Lagrangian:

Here Lem is the standard Lagrangian, with neither gauge-fixing nor ghost terms, and λ and σ are arbitrary real
numbers.

(a) What is Lghost?

(b) What is the ghost propagator?

(c) What are the vertices involving ghost fields?

(1998b 6.1)

Solutions 18

18.1 To lowest nontrivial order, and ignoring for the moment the contributions of the counterterm and the regulator
fields, we have

and to the same order,

The unrenormalized self-energy is, following the Feynman rules (described in the boxes on p. 443, p. 443, and p.
645),

(the first minus sign comes from the rule for fermion loops). We can ease the evaluation of the integral by the
substitution
This substitution will make the denominator even in q, so that we will be able to discard terms odd in q in the
numerator (at least when part of a convergent combination, which we assume):

The product of the numerators gives nine terms, but four (the terms linear in m) contain an odd number of gamma
matrices, and so have zero trace. Of the remaining five, two are linear in q and so are odd functions, and will
vanish upon integration (in a convergent combination). That leaves in the numerator

using the identities Tr[γµγν] = 4gµν, Tr[γµ γν ] = 4(2aµaν − gµνa2). Then

Let’s combine the denominators with the Feynman parametrization:

with

Then

and

To get rid of the cross-terms, let

Then, dropping the prime on the q’s, as well as terms linear in q (which will integrate to zero),

As in (34.62), the quantity qµqν can be replaced by q2gµν. We now have

where

To investigate the divergences of this expression, consider its large q behavior. It’s helpful to rewrite it as

Integrating fµν over d4q, the first term is quadratically divergent, and the second is logarithmically divergent. From
the chart on p. 527, it follows that we need three heavy masses. Adding the contribution from the regulator fields,
we get

choosing the coefficients bi and the masses Mi in accord with (33.53),

In the following we choose b1 = −b2 = b3 = −1. With these choices, the integrand in (S18.11) goes as O(q−6) as q →
∞, and so the integral is convergent. Applying the integral formulae (I.4) and (I.3) from the box on p. 330,

(the dots indicating divergent terms that cancel when two such terms are subtracted, provided the total integrand
vanishes for high q faster than q−4), we find

so that, with a = x(1 − x)k2 − m2 + iϵ,

There are two notable features of this result. The quantity is transverse, even before we include the regulator
fields, in agreement with (33.70). Also, this result has no mass dependence except in the argument of the
logarithm; the mass in the coefficient cancels. Including the regulator terms, we have

The counterterm is determined by the BPHZ prescription. The superficial degree of divergence1 of the
original graph (S18.1) is D = 2, indicating that it is quadratically divergent. We must, à la BPHZ, add counterterms
to L to cancel the terms in the Taylor expansion of µν up to order D in k. Because the integral (S18.14) is well-
defined for k = 0 and is multiplied by an explicitly O(k2) factor, there are no zeroth or first-order terms in k. Thus the
counterterm diagram makes a contribution2

which is to be added to (S18.14). Our renormalization conditions require (33.72) that the renormalized self-energy
′µν satisfies

with

The constraint (S18.16) follows automatically from gauge invariance of the 1PI diagrams. We choose E as

The renormalized self-energy then becomes

Taking the limits M1, M2, M3 → ∞, we find

which is both finite and transverse, as expected. Note that the counterterm is added into the Lagrangian via the
term
and so preserves gauge invariance. There is no need of a gauge-breaking photon mass term to renormalize the
self-energy, which remains gauge invariant:

Additionally, we see that to this order,

so that

as required.

18.2 The O(e2) diagrams contributing to the photon self-energy in the scalar case are shown below:

and to the same order,

Using the Feynman rules in the box on p. 644 (extended to d spacetime dimensions),

(Note that we are not writing eν4−d for the coupling constant, as in (25.23). This is not necessary since we are
using mass-shell renormalization conditions, and all dependence on ν drops out.)

As before, we combine the denominators in the first term with Feynman parametrization:

We get rid of the cross-term by shifting the momentum variable: let p = p′ − xk. Then

Linear terms in p′ vanish upon integration, leaving

As in the previous problem, we can replace p′µp′ν by a constant times gµνp′2; the constant is 1/d instead of . Then

From the Euclidean integral (33.19), we have the Minkowski space integral formula

and similarly from the Euclidean integral (33.26), we get the Minkowski version

Using these formulae, we get

The first term can be transformed with an integration by parts:

Substituting this last expression into the first term of (S18.26), we obtain

using Γ(n + 1) = nΓ(n). The terms independent of k cancel, and the rest once again combine to make a transverse
expression.

As in the previous problem, the counterterm diagram makes a contribution

Writing

the renormalization condition

fixes E:

so that

and

We now set d = 4 − δ, and use the expansions

(the value of the Euler–Mascheroni constant, γ, doesn’t matter here, because it is multiplied by δ). To O(δ0),

which again is both finite and transverse. Again, we see that ′T(k2) → 0 as k2 → 0, as required.

18.3 The effective Lagrangian is the result of averaging over gauge-fixing constraints of the form
The variation of the vector field under an infinitesimal gauge transformation

produces a variation in F(A):

We then have

We turn the determinant, det[δF/δχ], into a functional integral over complex ghost fields:

(up to a phase that can be absorbed into the functional integral normalization), where

(we can absorb the coupling constant 1/e into the normalization of the ghost fields; cf. Peskin & Schroeder QFT, p.
514). Then

The ghost Lagrangian is

From this, we can read off the ghost propagator,

and the ghost-photon vertex,

The first i comes from Dyson’s formula; the second from the derivative on η, a field that annihilates the incoming
ghost. The ghost field couples to the photon with polarization εµ(k −k′). The asymmetric appearance of the k’s is
odd, but correct.3

1[Eds.] See §25.4.

2[Eds.] See the table on p. 674; E is the photon wave function renormalization constant.
3[Eds.] Cheng & Li GT, pp. 262–263: “One should preserve a consistent convention of entering the momentum of
either the left or the right ghost line at every vertex. The ghost only enters in closed loops.” Peskin & Schroeder
QFT, p. 515 write only the outward k on the ghost lines.

34
Two famous results in QED

We now begin to discuss a sequence of problems involving interactions of quantum electrodynamics with an
external conserved, c-number current distribution Jµ:
We will restrict ourselves to the case of the real photon, with zero mass: µ2 = 0. Problems of a quantum
electrodynamic system subject to an external charge distribution are quite common. They are not realistic: there
are no external classical charges in the world, as far as we know, controlled by God and not by the motion of
particles. But in a typical problem in which we have an electron whirling around a synchrotron, it’s quite
reasonable to take the distribution of currents inside the synchrotron magnets as external and given, and not
worry about solving for the motions of all those electrons. We will obtain two famous results.

34.1Coulomb’s Law

The first thing we want to check is the vacuum-to-vacuum amplitude to second order in J. That’s an experiment
where we take two external charge or current distributions, and see what the interaction energy is between them.
We want to confirm that we’ve calibrated everything correctly and that e really is the e measured by Monsieur
Coulomb in his famous experiment. 1 The vacuum-to-vacuum amplitude to order J2 (assuming the charges are
weak, so that’s the leading order) is very simple:

The current makes a photon, the photon goes into electron-positron pairs (or whatever), and then comes back
again, reassembling into the renormalized propagator ′µν. The black dot is the interaction with the external
current. Both momenta are directed inward, and

The photon is massless by our renormalization conditions; the BPHZ prescription (at zero mass) and the mass
shell renormalization agree. The exact photon propagator can be written as

We will systematically suppress the iϵ from the one-photon intermediate states and our normalization is such that
the residue at the photon pole is 1. The first term in (34.4) is the contribution from the one-photon state. The
continuous contribution in the second term, à la the Lehmann spectral representation,2 contains contributions
from multi-particle intermediate states with (mass)2 = a2. Then there is the final gauge-dependent term (§31.3),
which we know suffers no radiative corrections.3 This is the same derivation as for the scalar case (§15.3) with
just a couple of extra indices floating around.

The Fourier transform µ(k) of the external current distribution is defined by

In lowest order, the amplitude for emitting a photon of momentum k is µ(k); that for absorbing a photon is µ(−k).
Current conservation implies

The graph in question is easy to compute:

The kµ µ terms drop out by current conservation, (34.6).

The significant feature is the coefficient of the 1/k2 term. We know from the study of one photon exchange4
that this term gives the standard Coulomb force. If the µ(k)’s correspond to static charge distributions, the
coefficient is 1, just as it would be in the free theory. If we take two charges that are far apart, the Yukawa
potentials (coming from the integral over a2 in (34.4)) fall off with distance, and the surviving force is the Coulomb
force. This reproduces (albeit with blackboard and chalk) M. Coulomb’s experiment, that the force between two
charges widely separated (on the atomic scale) is e2/r2. And it is e2; this is the proper way to couple in an external
current distribution if e is to be what M. Coulomb measured.

What about the Lehmann spectral distribution? In §10.4 we studied two spinless particles exchanging a
spinless particle. Then, as shown earlier5 we found that the associated propagator

corresponds to a Yukawa force. It gives a potential

Likewise, 1/k2, the Yukawa potential with µ = 0, corresponds to a Coulomb potential, the first term in (34.7).
Incidentally, the sign here is different from the case of scalar exchange; in (9.29) we have

whereas here we have

And that’s right. Scalar exchange is attractive between identical sources, but we know that the Coulomb force is
repulsive between identical sources. The forces are different because the residues of the poles in the propagator
have different signs. Take two large, external macroscopic charge distributions, idealized as classical. Put them at
different locations and measure the force between them to order J2, in the limit of weak charges (so we don’t have
to worry about nonlinear effects). For a time-independent external source, the sum of the vacuum graphs is
related to the energy shift in the ground state of the theory,6 so we can determine how the ground state energy is
changed by the force between these external charge distributions. It’s exactly how we determine an internuclear
force in molecular theory, by measuring the ground state energy of the electron system with fixed nuclear
positions. Here the analog of the electron system is the entire quantum electrodynamical vacuum state.

In the spectral distribution, physically a2 is the squared mass of the intermediate state. Contributions from
states with a2 > 0 drop off faster than the Coulomb force, due to their Yukawa form. I should mention that the
lowest value of a2 is not the squared mass of the lightest charged particle intermediate state (that would be (2m)2),
because we could have an intermediate state of γ → 3γ. (By charge conjugation, we can’t have a two-photon
state; the photon is odd under charge conjugation; Aµ |0ñ produces an odd charge conjugation state which cannot
be two photons.) This process only occurs through charged particles, as in Figure 34.1. But that doesn’t matter.
The fact that the transition matrix element from the vacuum to that state of Aµ involves graphs with internal
electron lines is not the point. The lowest value of a2 is the mass of the three-photon state, which is 0. I don’t know
how that three-photon state contributes, if it gives 1/r5 or 1/r10 or something else, but it goes to zero more quickly
than 1/r2.

Figure 34.1: Intermediate state with three photons

34.2The electron’s anomalous magnetic moment in quantum mechanics

We’ve calibrated the one-photon exchange so that the resulting Coulomb potential’s e2 has the right size. Now
let’s turn to something more complicated, and consider a current distribution scattering an electron. This will lead
to a theoretical determination of the anomalous magnetic moment of the electron. This famous calculation was
carried out independently by Feynman and Schwinger,7 and it made everyone believe in quantum field theory.8
Feynman and Schwinger did this to one loop. We’ll look at this first in the context of quantum mechanics.
Analogous to the one-loop calculation in QED, we’ll consider a first-order quantum mechanical calculation with a
weak external current distribution.

Figure 34.2: The electron-photon vertex

We have an incoming spinor u with momentum p, an outgoing spinor u′ with momentum p′, and a photon of
momentum k, with (see Figure 34.2)

To lowest order, the amplitude for this scattering process is

which can be written as the (generic) form (suppressing the iϵ of the photon propagator)

Fµ(k) is a current matrix element. It is a function with the spinors u and , some Dirac matrices and some
functions of the momenta. To lowest order, Fµ(k) is just the factor within the curly brackets in (34.11):

Let’s try to count the number of (Lorentz) invariant amplitudes there can be in Fµ(k), and then try to construct them.

The easiest way to count them is to go into the cross channel, reading the diagram straight down from the dot,
and consider the process

We can use standard angular momentum arguments to compute how many terms there are. We have a current
making an e+e− pair. In the center of momentum frame, kµ = (k0, 0) is timelike. The current has the same
properties as a JP = 1− particle, described by a spin 1 spatial vector with parity minus.9 So we wish to build a 1−
state out of an electron and a positron. Since e+ and e− are particle and antiparticle and fermions, they have
opposite intrinsic parities, and therefore the 1− state must have even ℓ.10 There are two possibilities:

Those can each be put together in a unique way to make a state of total angular momentum 1 and parity minus.

Charge conjugation is also a restriction but it happens to give us no further constraint. The current of course
makes charge conjugation odd states, but since the two constituents are fermion and anti-fermion, both the ℓ = 0,
s = 1 and ℓ = 2, s = 1 states are symmetric in both space and spin, and so are odd under charge conjugation.

Therefore there are no more than two invariant functions of k2 required to describe this process: the invariant
functions which, when we (analytically) continue to timelike kµ, represent the amplitudes for making the ℓ = 0 state
and the ℓ = 2 state. With that knowledge, we will write down two functions, F1(k2) and F2(k2), which satisfy all the
constraints of parity, charge conjugation, Lorentz invariance, etc. These two functions11 completely characterize
Fµ(k2):
Both the bilinear products γµu and σµνu are charge conjugation odd.12 Both of the terms in (34.15) are
vectors, and both F1 and F2 are real functions. They are called form factors. This is a subject where things bear
the names of distinguished physicists, and these functions, F1 and F2, are known respectively as the Dirac form
factor and the Pauli form factor. The factors of i and e/2m in the second term are for convenience, as you’ll see.

From our renormalization conventions and (34.13), we have one piece of information:13

This is the condition on the renormalization of the electric charge: that e is the 1PI function at k2 = 0. We know
nothing about F1 at any other value of k2, except to lowest order in e, and we know nothing at all about F2 (short of
actually calculating them). This analysis is special to a spin-½ particle but not to the electron; it could be any
spin-½ particle. It could be a proton, in which case there would be all sorts of strong interactions14 inside the 1PI
graph.

Aside.

This is a side remark, but it’s so important physically that I am not ashamed to devote a minute to it. Even
though the proton has these strong interactions, the fact is that we can analyze, for example, electron–proton
scattering,

at O(e2), Figure 34.3, in terms of two functions like these. The blob is unknown; it sums up the effects of the strong
interactions.15 We may not know what those functions F1 and F2 for the proton are until we can calculate with the
strong interactions, but we know that there are just two of them. The electron only interacts with electromagnetism
(and the weak interaction, but that’s just bubkes 16). Therefore we have a great simplification in studying
electron–proton scattering. In principle the scattering amplitude would be a bunch of bispinor covariant products
times arbitrary functions of two variables, say the energy and the angle. We have turned it into an expression
involving just two unknown functions, F1 and F2, of a single variable, k, and that is progress no matter how you
slice it. We could also turn this argument on its head, and say that by doing electron–proton scattering we get
information on F1 and F2 and perhaps learn something about the strong interactions, because these make F1 and
F2 for the proton what they are. It’s good both ways. We can say that we’ve reduced our ignorance of
electron–proton scattering even in the absence of understanding the strong interactions. Or we can say that we
use electrons as a probe to investigate the strong interactions of the proton in a very simple situation, instead of
looking at, say, a nucleus with 42 protons interacting with each other.

Figure 34.3: Electron–proton scattering

I have described the utility of these form factors. Now I will discuss their physical interpretation. Here’s why
we’ve singled out k2 in the formula to define Fµ. Let’s suppose that we are going to solve Maxwell’s equations for
the given current distribution Jµ. In the Lorenz gauge, for example, we would have to solve the equation

Here Aµc is a classical solution. In Fourier space,

Likewise, the classical electromagnetic field Fc µν associated with this classical potential

becomes, in Fourier space,

This enables us to give a new meaning to our interaction amplitude. Substituting (34.15) and (34.18) into (34.12),
and using (34.20) we find

In the third step we can write an antisymmetric product because of the summation with the antisymmetric matrix
σµν, and divide by 2 to avoid the double counting.

The first term, with the Dirac form factor, is exactly the same interaction with the classical field as would be
produced if we had a fundamental coupling . It differs in that it has a k2 dependence through F1(k2),
rather than a simple factor of e alone. Because of its k2 dependence, it’s not a constant, as it would be for coupling
to a point charge: a constant in momentum space is a delta function in position space. So we can say that the
effect of the interactions of the electron with the electromagnetic field (or the effect of the strong interactions with
the proton) is to “spread out” the particle. That is why this is called a form factor: it tells us the form of the electron,
the way in which the interaction is spread out.

The second term, the Pauli form factor, is an interaction of a new type, a spin-dependent interaction, the kind
that would arise if we had a Pauli term17

in the Lagrangian. That sort of interaction is gauge invariant and consistent with charge conjugation, though it’s
non-minimal; it would lead to something like the second term in (34.21), with a constant F2. Of course it can’t be
there as a fundamental interaction because it’s nonrenormalizable. It’s of dimension 5, in four dimensions; the
Dirac bilinear has dimension 3, while the derivative and the field Aµ each have dimension 1. Nevertheless the
effects that would be made by such an interaction can arise, not as a point coupling but in a spread out way as a
result of the quantum electrodynamic correction that make F2.

The only sure result so far is (34.15). We’re just playing with these objects F1 and F2 to try to get some idea of
their physical meaning. We can go farther by using a cute identity due to Walter Gordon (of Klein–Gordon fame),
the Gordon decomposition.18 We’ll start by doing something very stupid: we’ll complicate the simple expression

Using the free particle Dirac equations

the simple expression becomes

Now

Likewise

Assembling everything we get the Gordon decomposition

This decomposition has amusing consequences and gives us further insight into the physical meaning of the form
factors. Recalling kν = p′ν − pν, we can write (34.21) as

The first term, with (p + p′)µ µc , looks like the coupling of a spinless charged particle ϕ to an external electric field
c in lowest order, through its antisymmetrized current ϕ *∂µϕ − (∂µϕ *)ϕ. This is as close to a spin-independent
µ
coupling as a relativistic spin-½ particle can get: with no γ matrices inside. So the first term looks spin-
independent (though of course there are spin-dependent factors inside the bispinor product). Spin-independent
terms cannot contribute to the magnetic moment, so this first term is irrelevant to its calculation. The second term,
however, is spin-dependent, at least in the non-relativistic limit, and will contribute to the magnetic moment.

Let’s go immediately to the non-relativistic limit to learn about the spin-dependent coupling in the low velocity
regime. The free particle Dirac spinor in the non-relativistic limit is (see (20.27) and (20.33))

U is a two-component spinor with the non-relativistic normalization

For k2 ≈ 0,

but we don’t know the value of F2(0). In the extreme non-relativistic limit, for spinors at rest,19

The Lorentz generators, σ0i, are pure off-diagonal; they mix up large components with small components. 20 On
the other hand, the rotation generators σij are simply the Pauli spin matrices written as a vector, σ; ϵijk turns a
vector into an antisymmetric tensor: σij ≡ ϵijkσk . From the field strength tensor Fµν ((34.19); see also (S2.17) on p.
103) we have

Now we can very easily study the spin-dependent term in (34.29) in the non-relativistic limit. The only terms
that contribute are where µ and ν are i and j. We get a factor of 2 because we sum over everything twice, once in
the order ij and once in the order ji. The spin-dependent term becomes in the non-relativistic limit21

In non-relativistic quantum mechanics, the analog of what we call iA is given by a standard formula in first-order
perturbation theory:22

That is, the effective non-relativistic interaction is

Recall that the magnetic moment operator23 µ is defined by a charged particle’s coupling to an external
magnetic field:

(the non-relativistic spin operator is σ). Comparing (34.37) and (34.38) we see that the magnetic moment is 24
where µB is the Bohr magneton,

and g is the ratio of a particle’s magnetic moment µ to a Bohr magneton. It was introduced25 by Alfred Landé in
1921.

The first term in µ, e/m, is called the Dirac moment. It is an excellent approximation to the magnetic moment
of the electron, which asserts that for the electron, g = 2. It is found in all the books on non-relativistic quantum
mechanics, where the explanation is postponed to relativistic quantum mechanics. Well, here it is. For the
electron there is a small correction because F2 is non-zero. To lowest order, the correction F2(0) is O(e2). This is
the anomalous moment. That’s all we can say using only quantum mechanics. Now we proceed to its quantum
electrodynamic calculation.

34.3The electron’s anomalous magnetic moment in QED

Our task is to compute F2(0) to O(e2) in an orgy of Feynman computations. This is a Nobel Prize-class calculation.
At the New York meeting of the American Physical Society in 1948 (the community of physicists was so small at
that time that the sessions could be held in the classrooms of Columbia University), Julian Schwinger got a
standing ovation for the computation. 26 (This is not to be considered a hint. He did it for the first time. I’m doing it
for perhaps the 700th time, and indeed, the fourth time in the past twenty-four hours, until I got the signs right!)

The lowest order contribution, Figure 34.4, has amplitude

This tells us which factors we have to subtract out. Next, to order e3 (the graphs are only of O(e2), but there is an
overall factor of e) we have four diagrams, as shown in Figure 34.5. Graph (a) is the vacuum polarization on the
external photon. Graph (b) is the photon wave function counterterm; the × indicates the counterterm computed to
O(e2) so that the whole graph is O(e3). Graph (c) is the vertex correction. Graph (d) is the charge renormalization
counterterm; here, the × means the counterterm must be computed to O(e3). Though we went through all that
work with renormalization theory, in fact we don’t have to calculate a counterterm for this process. Graphs (a), (b)
and (d) are all proportional to γµu. They all have only a γµ at the vertex regardless of what happens upstairs;
they contribute only to F1. So we only need to worry about graph (c); that gives troubles enough. F2(k2) should
come out to be finite without any worries about counterterms or subtractions.27 The relevant part of Fµ, (34.12), is,
in the Feynman gauge

Figure 34.4: Lowest order contribution to the magnetic moment

Figure 34.5: O(e3) contributions to Fµ

 Figure 34.6: O(e3) contribution to the anomalous magnetic moment

where

and

(there is no (−i) at the external field vertex, because it was factored out in the definition of Fµ). We’ll evaluate N µ
and D separately, taking the numerator first. We won’t need the iϵ’s, so we’ll drop them.

Begin by moving ′ through γλ, recalling ′=m :

Do the same for γλu in the second fermion propagator:

The −m terms cancel the +m terms, and

Multiply this out:

The first term on the right we can simply drop; it is proportional to γµ and so contributes to F1, not F2. We’re
looking only for terms proportional to σµν. Leave the numerator for now, and let’s turn our attention to the
denominator, which is rather simple.

The electrons are on-shell:

so

and the denominator simplifies to

We also have the marvelous denominator-combining formula using Feynman parameters (16.16),

The integration is over the triangular region Δ defined in Figure 34.7. Thus
 Figure 34.7: Region of integration

Complete the square:

Shift the integration variable q to q′:

so

We can simplify this, using k = p′ − p:

We are not interested in terms of O(k2); they just give corrections to F2 as k2 moves away from 0. Therefore,

(and the same goes for all subsequent terms in p ⋅ p′). We are left with

That completes the simplification of the denominator. It’s important that the denominator is even in q′. It may seem
dull to you now but that’s because the standards of drama have changed; in 1948 it drew cheers.

Now back to the numerator, N µ. We must substitute the expression for q in terms of q′, (34.55), into (34.48).
We will get terms with no power in q′, terms linear in q′ and terms quadratic in q′. The terms linear in q′ are odd
functions (the denominator is even) and so vanish upon integration; we can go ahead and drop them. What’s left
is

This is its most horrendous form; it will simplify drastically.

In the first and second terms we drop ′2 and 2, respectively,28 since they both equal m2 by (34.49) and
hence these terms are proportional to γµ (and contribute only to F1). Rewriting,

The only term quadratic in q′ will involve an integrand ( γλγργµγσγλu)q′ρq′σ. But

because the integration is Lorentz invariant. We can check that the is correct by taking the trace of both sides
and seeing that they are equal. Therefore we can make the replacement
The second identity we will use is

This is easy to check.29 For any given γµ, one of the four γλ’s commutes with it, and three anti-commute, (while
γ0γ0 = γ1γ1 = γ2γ2 = γ3γ3 = 1), so we are left with 1 − 3 = −2 factors of γµ. Using (34.64) twice in the right-hand
side of (34.63), we see that (the two 2’s cancel the )

We drop this term since it’s proportional to γµ.

We have now reached an important point. Because we have eliminated all the q′2 terms from the numerator,
the integral is manifestly convergent. It goes like

The q′2 in the numerator, had it not been proportional to γµ, could have given a logarithmically divergent integral
and hence the wrong anomalous magnetic moment in a very drastic way, to wit, a divergent one. The answer we
get may be right or it may be wrong, but it will certainly be finite.

We are still left with the remaining two God-awful terms in (34.61). In the first term, anti-commute through ′:

The second equation follows from (34.57); in the third equation we drop terms O(k2) and proportional to γµ.
Likewise

In both of these expressions we drop the (p ⋅ p′) term because it is proportional to γµ.

The remaining term can be reduced from five γ matrices to three, with the aid of this little wonder of an
identity, stated without proof:30

It’s proved along the same lines as the previous identity. So

Use (34.10) to turn the p′’s on the right over to p’s and the p’s on the left to p′’s:

leaving us with ’s still floating around on the right or the left. For instance,

dropping the m term because that’s proportional to γµ. For the last term in (34.70), we find

once again dropping the term proportional to γµ. We’re looking for F2(0), so we ignore O(k2) terms. Putting all the
pieces together, we have

We can simplify things by the following observation. We are integrating over a region symmetric under the
exchange x ↔ y. If there are parts of the integrand antisymmetric under this exchange, they will vanish.
Consequently, we can go ahead and make the replacements

This has no effect on the denominator (34.59), but the numerator becomes (using [γµ, ] = −2iσµνkν)

At last we’re done with spinor algebra. We have from (34.15), (34.43), (34.44), and (34.59), and dropping the
prime on q′,

We can now extract F2(0), the coefficient of (ie/2m)( σµνu)kν:

The table from Chapter 15 tells us how to do any integral of this form.31 The relevant formula is (I.2):

Then

Therefore the magnetic moment of the electron, to the order we are working in, is (see 34.39))

In rationalized units it’s

that appears in Coulomb’s Law. The quantity α is called the fine-structure constant,

From (34.80),

The current experimental value32 is

The agreement between the experimental and theoretical result, 33 calculated only to first order, is to within 0.16%.

Next time we’ll look at higher order corrections to the electron’s magnetic moment, and also that of the muon.
These comparisons with experiment will lead us to consider the electromagnetic interactions of hadrons.
1C. A. Coulomb, “Premier Mémoir sur l’Electricité & le Magnétism” (First memoir on electricity and magnetism),
Histoire de l’Académie Royale des Sciences (1785) 569–578; the second and third memoirs follow immediately at
pp. 578–612 and pp. 612–640. These are freely available online at gallica.bnf.fr.
2 [Eds.] This is the spectral representation for scalars with the photon polarization sum in the numerator; see
(15.30); Schweber RQFT, Section 17b, pp. 659–677, in particular equation (66); and Bjorken & Drell Fields,
Section 16.11, pp. 166–170, and the (unnumbered) equation following equation (16.173).
3 [Eds.] Given a process in lowest order of e2, “radiative corrections” are higher order contributions to that process,
typically described by diagrams with loops or the emission of extra photons in the final state (bremsstrahlung, or
“braking radiation”). See Peskin & Schroeder QFT, Ch. 6, p. 175. The Ward Identity guarantees that any radiative
corrections arising from the gauge-fixing term in the Lagrangian will vanish when contracted with kµ; see (31.63).
(M. Headrick, private communication).
4 [Eds.] See the discussion following (30.22), pp. 648–650, and note 11, p. 650.
5 [Eds.] See (9.38) through (9.41).
6 [Eds.] See Section 3.7 in Chap. 5, “Secret Symmetry” in Coleman Aspects; Peskin & Schroeder QFT, pp. 96–98.
7 [Eds.] The result was first discussed at a small Washington D.C. conference in November 1947, (Schweber
QED, p. 317) and submitted as a letter to Physical Review a month later, though the explicit calculation was not
carried out in the letter: Julian Schwinger, “On Quantum-Electrodynamics and the Magnetic Moment of the
Electron”, Phys. Rev. 73 (1948) 416–7. Incidentally, the result is printed erroneously, as ( π)e2/ħc; the correct
value is (1/2π)(e2/ħc) = α/(2π). For Schwinger’s calculation, see equation (1.122) and footnote 3 in Julian
Schwinger, “Quantum Electrodynamics III. The Electromagnetic Properties of the Electron–Radiative Corrections
to Scattering”, Phys. Rev. 76 (1949) 790–817. Though Feynman does not seem to have published his calculation
at the time, it follows easily from equation (24) in R. P. Feynman, “Space-Time Approach to Quantum
Electrodynamics”, Phys. Rev. 76 (1949) 769–789 (reprinted in Schwinger QED); see R. P. Feynman, Quantum
Electrodynamics, W. A. Benjamin, Inc., 1962, p. 145, where Feynman derives the result from equation (28-10),
identical to his article’s equation (24).
8 [Eds.]“Unlike Bethe, Weisskopf and most of the other people at the Shelter Island conference [June 2–4, 1947],
Schwinger’s imagination was captured not by the Lamb shift but by the discrepancy in the magnetic behavior of
the electron... ‘That was much more shocking,’ Schwinger said. The Lamb effect, as Bethe showed, could be
accounted for almost entirely without the use of relativity. ‘The magnetic moment of the electron, which came from
Dirac’s relativistic theory, was something that no non-relativistic theory could describe correctly... To be told (a)
that the physical answer was not what Dirac’s theory gave; and (b) that there was no simpleminded way of
thinking about it, that was the real challenge. That’s the one I jumped on.’”, Crease & Mann SC, p. 132. Schweber
(QED, p. 318) writes, “The importance of Schwinger’s calculation [worked through during a five hour (!) talk at the
Pocono Conference, March 30, 1948] cannot be underestimated [sic]. In the course of theoretical developments
there sometimes occur important calculations that alter the way the community thinks about particular
approaches. Schwinger’s calculation is one such instance. By indicating, as Feynman had noted [in a letter to his
friend Herbert Corben, after the Washington conference] that the ‘discrepancy in the hyperfine-structure of the
hydrogen atom noted by Rabi can be explained on the same basis as that of the electromagnetic self-energy, as
can the shift of Lamb’ [emphasis added by Schweber], Schwinger had transformed the perception of quantum
electrodynamics. He had made it into an effective, coherent, and consistent computational scheme to order e2.”
Coleman did not discuss the Lamb shift—a tiny difference (~ 1060 MHz, with Δω/ω ~ 1 × 10−6) between the 2s1/2
and 2p1/2 energy levels of hydrogen, degenerate in the Dirac theory—in the course. The measurement was
carried out (1946–47) by Willis Lamb (a student of Oppenheimer) and Robert C. Retherford at Columbia
University: Willis E. Lamb, Jr. and Robert C. Retherford, “Fine Structure of the Hydrogen Atom by a Microwave
Method”, Phys. Rev. 72 (1947) 241-243. Lamb shared the 1955 Nobel Prize (with his Columbia colleague,
Polykarp Kusch, who’d measured the electron’s magnetic moment to high precision) for this work. Bethe’s non-
relativistic derivation of this result, famously carried out while traveling by train to Schenectady from New York
after the Shelter Island conference (2–4 June 1947), and based on Hendrik Kramers’ idea of mass
renormalization presented there, pointed the way to further progress: H. A. Bethe, “The Electromagnetic Shift of
Energy Levels”, Phys. Rev. 72 339-341. This was the first successful application of renormalization theory to QED:
Laurie M. Brown, Renormalization: From Lorentz to Landau (and Beyond), Springer, 1993, p. 4. See also J. J.
Sakurai, Advanced Quantum Mechanics, Addison-Wesley, 1967, Section 2.8, pp. 64–72 for a beautifully clear
treatment of Bethe’s calculation.
9 [Eds.] See §6.3 and §22.1.
10 [Eds.] See note 2, p. 460.
11 [Eds.] See Peskin & Schroeder QFT Section 6.2, pp. 185–6, or Greiner & Reinhardt QED, Exercise 3.5,
“Rosenbluth’s Formula”, pp. 113–114, for the derivation of the general form in (34.15).
12 [Eds.] See the chart on p. 469. Reminder: σµν = [γµ, γν]; (20.98), p. 419.
13 [Eds.]S. D. Drell and F. Zachariasen, Electromagnetic Structure of Nucleons, Oxford U. Press, 1961, End Note
6, pp. 105–106.
14 [Eds.] Drell and Zachariasen, op. cit.
15 [Eds.] In the video of Lecture 34, Coleman adds: “Well, the effects are unknown, unless you can solve the
strong interaction problem. In that case, what are you doing sitting in this class?”
16 [Eds.] Yiddish, bubkes (various pronunciations; often “BOOPkiss”, to rhyme with “put this” or “BUPkiss”, to
rhyme with “up this” and sometimes spelled bupkes), “a contemptibly insignificant quantity”. See Rosten Joys, p.
44.
17 [Eds.] This sort of interaction appeared previously. See (27.69), p. 585, and the discussion following.
18[Eds.] W. Gordon, “Der Strom der Diracschen Elektronentheorie” (The current in Dirac electron theory), Zeit.
Phys. 50 (1928) 630–632; Peskin & Schroeder QFT, Problem 3.2, p. 72.
19 [Eds.] Antisymmetric 4-tensors like Fµν and σµν can be described simply as a 3-vector and an axial 3-vector.
For example,

and Fµν = (F0i, Fij) = (E, B). Likewise it can be shown that

where α is given in (20.11) and Σ is given by (20.7). Consequently, σµνFµν = iα•E − Σ•B. See V. B. Berestetskiĭ,
E. M. Lifshitz, and L. P. Pitaevskiĭ, Relativistic Quantum Theory Part I, Pergamon Press, 1971, Problem 1, p. 67
and p. 100.
20 [Eds.] In the standard representation, the bottom two-component spinor is, in the non-relativistic regime,
proportional to (v/c)2 times the top two-component spinor. The lower spinor consists of the “small” components
(going to zero as (v/c) → 0) and the upper spinor contains the “large” components. See Bjorken & Drell RQM, p.
12.
21 [Eds.] Recall the identity ϵijkϵilm = δjlδkm − δjmδkl, so ϵijkϵijm = 2δk m. See also (37.47), p. 812
22 [Eds.] See Ch. 6, “Perturbation Theory”, pp. 129–157, equation’s (38.2), (38.5), and (40.5) in Landau & Lifshitz
QM.
23 [Eds.] D. J. Griffiths, Introduction to Quantum Mechanics, Second Edition, Prentice Hall, 2005, Section 4.4.2,
pp. 181–182.
24 [Eds.] The classical, non-relativistic treatment gives (in the units used here) the electron’s magnetic moment
equal to one Bohr magneton. The empirical value of g for an electron is very close to 2. It was a great success of
the Dirac equation that it predicted g = 2 exactly. See also Jackson CE, Section 11.8, “Thomas Precession”, pp.
548–553.
25 [Eds.] A. Landé, “Über den anomalen Zeemaneffekt (Teil I)”, (On the anomalous Zeeman effect (Part I)), Zeits.
f. Phys. 5 (1921) 231–240.
26[Eds.] Freeman Dyson was in the audience. He later wrote his parents, “The great event came on Saturday
morning [Jan. 31], and was an hour’s talk by Schwinger, in which he gave a masterly survey of the new theory...
There were tremendous cheers when he announced that the crucial experiment had supported his theory: the
magnetic splitting of two of the spectral lines of gallium... were found to be in the ratio 2 times 1.00114 to 1; the old
theory gave for this ratio exactly 2 to 1, while the Schwinger theory gave 2 times 1.00116 to 1.” Schweber QED, p.
320.
27 There are no corrections on the external legs. The electrons are on-shell, so any corrections to the external legs
are simply canceled out by the counterterms, and therefore we don’t bother to write either of them down.
28 [Eds.] = pα pβγα γβ = pα pβ{γα , γβ} = pα pβgαβ = p2.
29 [Eds.] We have γλγµγλ = (2gλµ − γµγλ)γλ = 2γµ − 4γµ = −2γµ.
30 [Eds.] Bjorken & Drell Fields, Appendix A, p. 284.
31 [Eds.] See the box on p. 330.
32 [Eds.] PDG 2016; http://pdg.lbl.gov/2016/tables/rpp2016-sum-leptons.pdf .
33 [Eds.] Julian Schwinger is buried in Mt. Auburn Cemetery, about a mile west of Harvard Square. Above his
name, his tombstone bears the inscription .

35
Confronting experiment with QED

In the last lecture, to general jubilation and relief on my part, I derived the correct O(α) formula for the magnetic
moment of the electron

where α is the fine-structure constant

Today I will discuss other experiments involving this formula and, in a qualitative but not quantitative way (except
by quoting other people’s results) the higher order corrections to the electron magnetic moment.1

35.1Higher order contributions to the electron’s magnetic moment

To get an idea of the size of these effects we have to know what α is. The current experimental value is 2

That is, with a standard deviation of 31 in the last two digits displayed. The fact that α is not known exactly means
that we don’t know the first-order correction to µ exactly. The uncertainty in α –1, 31 × 10−9, leads to an uncertainty
in µ,

We can make a rough guess about the size of the higher order corrections, which we have not computed but of
course are in the literature:

Plugging this number (35.3) in, we find to O(e2)

I’ve only carried it out this far, even though the error in α is known much better than that, because the anticipated
size of the (α/π)2 correction would be perhaps 5 in the sixth digit. The current experimental value is 3

The uncertainty is 26 in the last two displayed digits. As we see, to our expected level of agreement with
experiment, i.e., up to the point where we expect the (α/π)2 corrections to come in, the agreement is perfect. This
is impressive. This is already five decimal places of agreement. The calculation took 45 minutes; the experiment
took 20 years. The theoretical calculation (including higher order corrections) gives4

(the uncertainties coming from the eighth-order QED term, an estimate of the tenth-order term, the hadronic and
electroweak contributions, and the uncertainty in α, respectively). To within a couple of standard deviations, the
agreement between theory and experiment is perfect, a few parts in 1013. The most significant source of error is in
the uncertainty in α. There is also a theoretical uncertainty in the last term because there are so many graphs to
compute.5 Instead of bothering to compute some of them, we just estimate that they are less than a certain
amount. But we don’t know if they are going to cancel or add together, so we have a purely theoretical uncertainty,
caused by lack of strength on the part of theoreticians, in the O(α/π)3 terms.

35.2The anomalous magnetic moment of the muon

There is another number for which we can use exactly the same formula, the magnetic moment of the muon. The
muon is just a heavy electron:

But as far as we know, in all of its interactions, the muon is exactly the same as the electron. This is one of the
great unsolved mysteries. In the elegant formulation of I. I. Rabi, “Who ordered that?”6 It’s a good question, and
still unanswered. In any event, there it is, so we can test our theory against the experimental results for the muon.
The magnetic moment of the muon is known experimentally to surprising accuracy:7

(the parenthetical numbers are the uncertainty in the µ+ and µ− moments, respectively). The theoretical QED
value is 8

This is agreement to within 1 part in 108. That is the situation comparing theory and experiment, and we all agree
that this is heartening. At the end of this lecture we’ll be doing a computation in which we’ll be overjoyed to get
agreement to within 20%.

There are some questions that should be asked. Why are the two magnetic moments different? Why is the
muon moment larger than the electron moment? Of course, out to the fifth decimal place, the computation of the
muon is identical to the computation for the electron. But obviously something is happening in higher orders that is
different for the muon and for the electron. What is that something? Also, why is the theoretical error figure for the
muon different from the one for the electron? Is it just that the theoretical physicists who work on computing the
muon moment are less energetic than those who work on computing the electron moment? These questions turn
out to have the same answer.

Many of the graphs that contribute to the anomalous magnetic moment of the electron are exactly the same
as, and in one-to-one correspondence with, the graphs that contribute to the anomalous magnetic moment of the
muon. As far as these graphs go, the number

(µ is the lepton’s magnetic moment, m is the relevant lepton’s mass) is exactly the same for the electron and the
muon: it’s a dimensionless number, and the mass of the lepton is irrelevant. At the fourth order however, we begin
to encounter one and only one graph that is different for the electron and the muon. This graph involves two
leptons (and it’s the first such we’ve seen), one on the external line and another on the internal loop. Here we can
have a difference between muon and electron graphs, because one type of lepton can be inside, and the other
can be outside.

These are qualitatively different graphs. For the muon moment, we have a heavy particle going through the
external line and a light particle running around the internal line. In the other case, the electron moment, the roles
of heavy and light are reversed. (If we have the same particle on both fermion lines, electron–electron or
muon–muon, the computations would be identical by dimensional analysis.) In fact, although fourth order
corrections are in general difficult to work out, it’s not especially difficult to compute this particular graph. We won’t
do it explicitly, but it’s sort of simple in its structure because it breaks up into two computations which we’ve
already done. In the internal photon loop we have the correction to the photon propagator (last week’s set of
homework problems9); once we’ve put in that corrected propagator, we have the graph we did last time, for the
electron moment, with one slight modification, which I will now describe.
 Figure 35.1 Magnetic moment diagram with two leptons

Figure 35.2 Higher order contributions to the magnetic moment

To O(e4), we might as well replace the graph in Figure 35.1 (and all of its friends) by the one in Figure 35.2,
with the corrected photon propagator in place of the fermion loop. The changes caused by the corrected photon
propagator’s including all sorts of other things besides the lepton loop are going to come in at O(e6). (There will
also be a host of counterterms.) The actual relationship may be written

We know the corrected photon propagator can be written in the spectral form (34.4):

The first term on the right reproduces the first graph on the right hand side in (35.12), which we calculated in
§34.3. Then we have a continuous superposition of the same graphs with heavy photons, whose squared mass
equals a2; σ(a2) is the photon spectral function.10 Finally we have gauge-dependent terms, which are irrelevant
if we’re working in Landau gauge.

If we knew how to compute the photon spectral function to the relevant order (and indeed we do, because we
have the photon self-energy, or, by a trivial manipulation of the homework problem you just did, the muon
intermediate state contribution to the photon spectral function), and if we knew how to compute the anomalous
magnetic moment to the lowest order, with a heavy photon instead of a zero-mass photon as we did before, then
we would just put these two things together. Thus we would be able to compute (perhaps in terms of an integral
we’d have to do numerically) the contribution of the graphs in Figure 35.1 to the anomalous moment, without
having to worry about any complications from renormalization or overlapping integrals or anything fancy.

Contribution to F2(0) from a “photon” with µ2 = a2

This is a trivial generalization of what we did last time. Most of that work involved manipulating the numerator.
We don’t have to do that again, because the numerator doesn’t give a damn about what’s going on in the
denominator, which is the only place that the photon mass appears. Recall that last time we ended up with an
expression for F2(0), the Pauli form factor, as in the first line of (34.79):

That was our old denominator. It came from the fermion mass factors when we applied Feynman’s formula. If the
photon carries a squared mass equal to a2, the only difference comes in the denominator. When we parametrize
the integral with Feynman’s trick, there will be a term proportional to the photon mass squared. Formerly we had
the integral with x for one electron propagator, y on the other and 1 − x − y for the photon propagator. So the
change to massive photons leads to

That’s the answer. This integral is elementary. Switching to the independent variables x − y and x + y, the x − y
integral is trivial. The remainder becomes a polynomial over a quadratic form which we can look up in an integral
table, so it’s not particularly difficult; I won’t bother.

I will however make some remarks about this expression (35.15). First, F2(0) is positive. Second, it is a
monotonic decreasing function of a2; the heavier the photon, the less the contribution it makes. Third, it becomes
the standard result as a2 → 0:

Fourth, as a2 → ∞, the m2(x + y)2 term becomes negligible, so we drop it. Then the (1 − x − y) in the numerator
cancels the (1 − x − y) in the denominator, so

To summarize, the integral for F2(0)

1.is positive

2.is a monotonic, decreasing function of a2

3.goes to the earlier result, α/2π, as a2 → 0

4.goes as O(m2/a2) for a2 ≫ m2

From these conclusions, we can see that the muon contribution to the electron moment is going to be very
small. The contribution of the lepton intermediate states to the spectral function, that is, from the graph

will be zero if a2 < 4mℓ2, where mℓ is the mass of the lepton in the loop. That is, σ(a2) = 0 for a2 < 4mℓ2. So

as (me/2mµ) ≈ 1/400 ≈ (α/π). Because of this suppression factor, the muon contribution is on the order of (α/π)4;
it’s negligible, and simply not worth computing.

On the other hand, the electron contribution to the muon moment is going to be very much larger. The
electron is very light, so the loop makes a large contribution to the spectral function, as if the equivalent photon,
with a2 = 4me2, were massless; it’s effectively massless on the scale of the muon. We can get an estimate of this
function very quickly, forgetting about all the numerical coefficients, simply by asking, “What if the electron were
massless?” Just by dimensional analysis, the spectral weight function σ(a2) in (35.13) would have to go like 1/a2,
since there is no mass in the problem; it’s being integrated over da2, and σ(a2)da2 has dimensions of 1. If we’re
integrating 1/a2 over da2 we get a logarithm. That is,

This expression has a logarithmic divergence as the electron mass goes to zero, though it’s certainly not divergent
if the electron has a non-zero mass. Since this is a number on the order of 2 ln(200) ≈ 10 times (α/π)2, it will give a
rather large contribution, much larger than the muon contributes to the electron’s moment. The difference between
these will make the muon moment larger than the electron moment, because the contribution is positive. When we
look at the experimental numbers, we find that they differ precisely at order (α/π)2 ≈ 5 × 10−6. To this order

They start differing in the fifth decimal place, at O((α/π)2), with a rather large coefficient, 5 or 6, in line with the
theoretical estimate.
The hadronic contribution to the leptons’ magnetic moments

Where do the hadrons come into the game? To a first approximation our quantitative knowledge of the strong
interactions is zero. Therefore when we consider higher order terms, we’ll eventually have graphs with strongly
interacting particles appearing, produced by the photon. The hadrons first appear at O(e4), where the internal
muon or electron loop in Figure 35.1 could be replaced by, for example, a pion pair, π+π−. Once a strongly
interacting particle gets into our graph, we’re cooked, because the strong interactions can’t be analyzed by
perturbation theory: the coupling constant is too large, g ≈ 15. We can complicate the graph enormously by
inserting any number of strong interactions, without adding any power of e. For example, the π+π− loop could
generate a proton-neutron loop, as shown in Figure 35.3. That doesn’t give us a higher power of e, it gives us
powers of g. Well, perhaps not g = 15, but we know it will not be negligible compared to what we have. Once we
open the door to a hadron pair, we’re going to have to introduce all the effects of the strong interactions.
Fortunately, at least to O(e4), the same rules apply as in Figure 35.2. The effects of all of these things up to O(e4)
is simply to give account of the hadrons’ strong interaction corrections to the photon propagator. Now we can
already see, unless something very funny is going on, that for the electrons the strong interaction corrections are
going to be negligible, because the lightest hadron, the pion, is roughly as heavy as the muon.11 And therefore
we’ll get the same suppression factor. The effects of the hadron intermediate states are not small, but the hadrons
are heavy. So for the electron, the effects of the strong interactions are O(e8), negligible.

Figure 35.3 Pion loop with proton-neutron loop

But the situation for the muon is different. The muon belongs to the lepton family, a “lightweight” particle, but
of course it’s almost as heavy as the pion, which is a hadron. So in principle we have to take the strong interaction
corrections seriously for the muons. They will be O((α/π)2) surely, but whether the factor in front is large or small is
something that requires computation to determine. (There’s a homework problem12 for you to see how a specific
computation is done. You won’t have to do the integral; you won’t need to. But you will see how one determines
experimentally the strong interaction corrections to the spectral density function σ(k2).)

The result is that in fact these corrections are negligible. The hadronic contributions to the muon moment13
are about 7 × 10–11. If we look into the guts of the computation, we can see what’s happening. The real reason is
that those two pions essentially have no effect until they’re resonant with a ρ state, and the ρ is about six times
more massive than the pion: 770 MeV vs. 135 MeV. That brings in a suppression factor of (2mπ/mρ)2 ≈ , which
helps a great deal. The effects, in any event, are quite small, just on the verge of being experimentally
measurable.

This concludes the discussion of experiment and the anomalous magnetic moments of the muon and the
electron. I want you to be impressed by something other than what people are usually impressed by in this
discussion, namely, that nine decimal place accuracy between experiment and theory. And indeed, that’s very
impressive. But after all, it requires a lot of hard work to get that accuracy. I want to demonstrate in this lecture how
we can understand the qualitative nature of some of those further decimal places without doing much work at all,
just by thinking about the general structure of the theory. That’s just as important. If you can’t do that, you’re liable
to launch on a computation that gives you a number that’s meaningless because you’ve neglected effects you
should have taken into account. You have to learn how to make qualitative estimates before you begin to do
quantitative computations. 14

We have already begun talking about the interplay of the strong interactions and electromagnetism. This is an
interesting subject because if we have processes that we can consider as purely electromagnetic, like the
anomalous moment of the electron up to order e6, then in principle we know everything, if we are willing to work
hard enough. And if we have processes that are purely strong, we know nothing. Well, we know quite a bit more
than we knew once. The interesting half world is where we have strong interaction corrections to electromagnetic
process or equivalently electromagnetic corrections to strong interaction processes, where we know half of what’s
going on. Does that enable us to tell anything about experiment? Or does the ignorance of the strong interactions
corrupt everything so that we know nothing about nothing?

This is not a systematic subject, but rather a subject where people have one clever idea after another, and
each of them gives us a little bit of knowledge. We will discuss two such topics: a low-energy theorem (due to
Francis Low15 for elastic photon–hadron scattering, principally Compton scattering off a proton or a neutron, and
selection rules following from the quantum numbers conserved under the strong interactions: isospin I,
hypercharge Y, G-parity G = eiπIyC for hadrons emitting one photon, two photons, etc. We will obtain selection
rules close in spirit to those in atomic spectroscopy arising from the effects of spin-orbit terms in the LS coupling
model. In that case it’s hard to compute those terms, but we can make lots of statements about the rotational
transformation properties. A similar statement will hold for photon–hadron processes.

35.3A low-energy theorem

Let ω be the photon energy in the center of momentum frame (in which the energies of the incoming photon and
outgoing photon are equal). There are two processes of interest. The first is

(here p can be any charged spin-½ particle; the proton, for example). It has a differential cross-section which in
the center of momentum frame can be written as (12.26)

where A is the amplitude, not averaged over anything. As ω, the energy of the photon, goes to zero, ET2 goes to
4m2. We’ll see what A looks like shortly. The result, which we’ll derive, is Low’s theorem on soft photons:

The first two terms are known in terms of e and µ, the charge and static magnetic moment of the proton, including
their full angular dependence and everything else. The third O(ω2) term is unknown; it involves the inner working
of the strong interactions. In fact, if we don’t consider the question of infrared divergences (caused by internal
photons), the expression (35.22) can be shown to hold to all orders in e with, possibly, the second terms having
some logs of ω in it. However, we will just work to lowest nontrivial order in e; that is, e2. The second process is

(n can be any spin-½, electrically neutral particle, for instance, a neutron), has a differential cross-section that can
be expressed as

The ω term is a charge-magnetic moment cross term, which vanishes because n is uncharged. There will be
higher powers in ω, and we don’t know what the O(ω3) term is. I don’t claim that (35.23) is a convergent power
series in ω; just that the terms vanish as ω → 0 at least as fast as ω2.

We begin by considering photon–proton scattering, as shown in Figure 35.4.

The ε and ε′ are the photon polarizations, k and k′ are the photon momenta, u and u′ are the proton spinors and p
and p′ are the proton momenta. Energy–momentum conservation says

Figure 35.4 Photon-nucleon scattering

We put the protons on the mass shell:

We’ll keep the photon masses completely free, but at a later stage in the computation we will let them go to zero:

We might imagine a world in which there are heavy photons, or even two different kinds of heavy photons with two
different masses. We could scatter a photon of the first kind and pull out a photon of the second kind. Or the
incident photon could be virtual, emitted from an e+ – e− pair.

The amplitude A for this process is going to be

where Aµν is constructed out of u and ′ and p’s and k’s, etc. We break Aµν up into two terms

where ABµν is the pole term (or the Born term) and AAµν is analytic. I’ll explain what we mean by ABµν, and then
you’ll know what we mean by AAµν. Graphically,

The blob is the term that gives us the residue at the pole that we know is present in these two graphs, at s = m2
and u = m2, respectively. That is, it is the value of this three-point function with everything on the mass shell16

for each blob. The pole comes from the nucleon–antinucleon–photon vertex, when the nucleon is on the mass
shell. So the pole term in Aµν is

The analytic terms are the remainder. They are analytic in the sense that we have extracted out the total residue of
all the graphs that have poles in them. So AAµν has a Taylor expansion in k and k′ near k = 0 and k′ = 0. Of course
we have not gone through a discussion of the analytic properties of the Feynman graphs, but it’s plausible that
those other graphs should be analytic at this point. After all, this was the same reasoning we used in our earlier
discussion of strong interactions, when we discussed how to compute the πNN coupling constant g in Model 3 by
looking for the pole in pion-nucleon scattering.17 There is a potentially singular term, ABµν with poles in it; all the
rest is non-singular at that point.

So far this is nothing new; we haven’t introduced anything special about electrodynamics. The specifically
electrodynamic part comes from the conservation of current, whatever the photon mass is: if we replace the
photon polarization vector ε′ν by k′ν, the photon momentum, or εµ by kµ, we get 0:18

Our not having k2 = 0 and k′2 = 0 means that we can treat k and k′ as independent variables; the only constraint is
(35.25)

together with the requirement that the protons must be on their mass shell. Written another way,

As long as k − k′ is kept spacelike (which will still allow us to vary k and k′ in four independent directions), we can
keep p and p′ on the mass shell with no problems.19 Therefore, we can differentiate AAµν independently with
respect to k and to k′, treating k and k′ each as four independent variables. From (35.33),

By construction, AAµν is analytic at k = 0, and so is its derivative. Multiplying its derivative by k and sending k → 0
causes the second term to vanish. Therefore,

It has no term of zeroth order in k. By exactly the same reasoning

Now if AAµν is O(k) and O(k′), then the first term in the power series must be O(kk′):

There can’t be a term with k but not k′, because that wouldn’t be zero when k′ → 0; there can’t be a k′ with no k
because that wouldn’t be zero when k → 0; there can be a kk′ term. Thus the conservation of charge, which
implies (35.33), plus the analysis of singularities in terms of extracting out the poles, has given us something more
powerful than either would have given us independently. If we just did the singularity analysis all we would know
would be that AAµν is analytic (non-singular). Now we know much more. We know it vanishes as kk′ as k and k′
independently go to zero. This is the low-energy theorem of F. E. Low:

If we did a similar singularity analysis for a process involving 72 photons we would get something vanishing like
the product of all 72 photon momenta.

Armed with this knowledge, we return to the case where k2 = k′2 = 0. Now both k and k′ are of order ω, the
photon energy: the photon is on the mass shell, and the space parts of k and k′ are the same magnitude as their
time parts. Therefore, in this particular case, the low-energy theorem becomes

where the Born term is given by

The diagram in (35.31) is a typical Feynman graph except that at the vertex, in place of the bare coupling, we’ve
put the effect of all the renormalization corrections, which as far the residue of the pole goes is just summed up in
this expression for iAB. The O(ω2) comes into (35.40) because it doesn’t matter which components of k and k′ we
have in (35.38); the product will be O(ω2).

We want now to count powers of ω, so as to establish the results in (35.22) and (35.23). Since both p and k
are now on the mass shell, the denominator is

In the proton’s rest frame, p = (m, 0), and

which goes to zero as the photon energy goes to zero. It looks like we get a factor of ω in the denominator of
(35.41), coming from the cross term F1(0) with F1(0), charge with charge:

All the other terms have explicit powers of k in the numerator and are not O(ω–1). So we have to look at possible
terms of O(ω–1) to count how many powers of ω–1 we get from the Born term; we will have to look at the cross
terms between the Born term and the analytic term to see what we don’t know about the total cross section.

Possible terms of O(ω–1)

To O(ω), (35.41) is

In general,

We can always choose ε to have only two components, perpendicular not only to the four-dimensional k but
perpendicular to the space part of k. The space part of k is aligned with the space part of p in the center of
momentum frame. Therefore we also have

since the time part of k is a fortiori aligned with the time part of p. So

and (35.44) becomes

Therefore in the actual scattering amplitude the term of O(ω–1) vanishes; the scattering amplitude begins at /ω ~
O(1).

We can now extract the results stated in (35.22) and (35.23). Staring at (35.41) and the corresponding
expression for the cross-term, we have a term of O(1) proportional to e2 coming from the term in the middle;
we’re not going to get rid of that. We’ll also have terms of O(ω) and O(ω2) coming from the F1F2 and F2F2 terms,
respectively, which we could compute (but it’s tedious). Hence we have established (35.22):

What if, instead of a proton, we were considering a neutron (or indeed any electrically neutral spin-½
particle)? For a neutron,

In that case the e ′∗ and the e terms in (35.41) are completely missing from the amplitude. We see that the
amplitude is O(k) because there are two k’s in the two F2(0) terms and a k in the denominator. The denominator
pole survives in this case, but of course it’s killed by the powers of k in the numerator:

Therefore the amplitude itself is a known term of O(ω) plus an unknown term of O(ω2). We square that to get the
cross section, and obtain the result stated in (35.23). The leading term is the Born approximation, and gives the
famous Thomson formula for low-energy scattering.20

I left something out of the argument. I said that kνAAµν = 0, but actually we only know kνAµν = 0 for the full Aµν.
Well, it’s easy to check that the Born (pole) amplitudes by themselves satisfy this equation, kνABµν = 0, because
they are exactly the Born amplitudes that would arise in a completely gauge invariant theory with anomalous
magnetic moment couplings; and by conservation of charge, kνABµν = 0. Since kνAµν = 0 and kνABµν = 0, we must
have kνAAµν = 0 also.

35.4Photon-induced corrections to strong interaction processes (via

symmetries)
We now turn to our second topic. The first class of processes we will consider are of O(e) in amplitude: they
involve one interaction of the photon with strongly interacting particles. For instance

or (see Figure 35.5)

where i is the initial hadronic state, f is the final hadronic state, γ is a photon and e+ and e− form an
electron–positron pair. (The second process is of O(e2) but as far as the strong interaction end of the graph goes,
there is only one photon involved.) Equivalently, we look at hadron electromagnetic form factors, all to lowest
nontrivial order in e.

Figure 35.5 Second order electromagnetic correction to strong interaction process

All of these processes are governed by the matrix element of the electromagnetic current between the initial
and final hadronic states:

We will assume that the electromagnetic current is constructed by the minimal coupling prescription, whatever the
theory of hadrons may be. We know the Gell-Mann–Nishijima relation21

so that

Iz and Y are commuting quantities. The currents j µIz and j µY are the currents we would get if the photon were
coupled exclusively to Iz or Y, respectively. The strong interactions22 strictly conserve both isospin and
hypercharge, and the electromagnetic interactions conserve Iz and Y (but not I), so we have

Now Iz , the integrated time component of the isotopic current j µIz, is part of an isotriplet. If we count the quantum
numbers of the initial and final hadron states, from the j µIz term in the electromagnetic current we have ΔI = 1. But
Y, the integrated time component of the hypercharge current j µY, is an isosinglet, so that gives us ΔI = 0. That is, in
electromagnetic processes,

The G-parities are different for these two cases.23 G is the product of charge conjugation and a 180° rotation
about the y (or 2) axis in isospin space:

Thus, for example, using the usual conventions,24 with

we have
That is, under G-parity the isovector π = (π1, π2, π3) transforms very simply:

The Lagrangian describing hadrons’ electromagnetic interactions is invariant under charge conjugation, and the
electromagnetic current shows up in it as the product j µemAµ. Since the photon is charge-conjugation odd (the
electromagnetic field changes sign when positive and negative charges are interchanged), so is j µem.25 From
(35.53), both jµY and j µIz must be odd under charge conjugation:

Charge conjugation changes the hypercharge of the hadrons. But these currents have opposite G-parities. Since
j µY is an isosinglet, Iy (or indeed any rotation in isospin space) has no effect on it, so its G-parity is the same as its
charge conjugation: odd. On the other hand, while the current j µIz has the same charge conjugation properties as
the hypercharge current, when we rotate jµIz by 180° about the 2 axis in isospin space, it changes sign once more.
So jµIz is G even, and we have

Thus we get the selection rules for two types of hadronic reactions that can be induced by the emission of a single
photon

Of course the ΔG rule is only useful if the initial and final states are G eigenstates.26 For example, it doesn’t tell us
anything about p → (something) except to connect it to → (something), where is an antiproton. Let’s look
at some examples.

EXAMPLE 1: γ decays

The famous decay27

is allowed. The Σ0 is the Iz = 0 component of an isotriplet, and the neutral Λ is an isosinglet. Both have Y = 0. So

The ΔI = 1 is permitted. (G-parity is irrelevant here, because it can only be checked if we know the relative phase
of the process Σ0 → Λ + γ.) The decay (35.63) is indeed an allowed process, coming from the ΔI = 1, ΔG = 0 part
of the electromagnetic current, jµIz.

For the emission of two photons, apply the same rule twice; it is a second-order process. Consider the decay

Again, π0 is the Iz = 0, Y = 0 component of an isotriplet:

so again we have

The total change in G is 1, because the final hadronic state is the vacuum, which has G = 0, and the initial state is
a single pion, which has G = −1. The (absolute value of the) total change in isospin is also 1. Therefore one photon
must come from each of the currents in (35.62). Notice that if these rules are correct, a neutral G-odd isosinglet
(such as the ω) would not be allowed to decay into two γ’s.

EXAMPLE 2: Magnetic moments within an isomultiplet

 Say we have an isomultiplet, and all of its members have magnetic moments. Well, are they all independent,
or are they connected in some way? The magnetic moment is connected by kinematic factors (which we have
worked out for one electron) to the matrix element of j µem:

Suppose the states |a, ñ and |b, ñ are members of an isomultiplet. In fact the current only has diagonal matrix
elements, but it’s useful to consider anything on the right-hand side and anything on the left-hand side. The
electromagnetic current is the sum (35.53) of two parts, one of which transforms like an isoscalar and whose
matrix elements must therefore be proportional to δab, and one of which transforms like the z-component (or the 3
component) of an isovector and therefore, by the Wigner–Eckart theorem28, must be proportional to Iz . For
example,

Therefore we have the following rule for the magnetic moments:

α and β are constants, α coming from the isovector part of the current, β from the isoscalar part. We must be able
to solve the strong interaction problem to actually compute them.

Unfortunately, this is a useless formula for practical purposes, because the only isomultiplet for which we
have measured all the magnetic moments is the neutron and proton, which has two magnetic moments. It’s no
great feat to fit two experimental numbers with two adjustable parameters! The Σ moments would be a test, but
unfortunately the Σ0 moment is hard to measure because it is so damned unstable, decaying rapidly into γ + Λ in a
time of ≈ 7.4 × 10−20 seconds.29 (The Σ± lifetimes are much longer, ≈ 1.48 × 10–10 seconds.) So at the moment
this is a beautiful formula which everyone believes, but which is totally untested. However, in a later lecture30 we
will obtain a similar formula by identical reasoning based on SU(3), which relates the magnetic moments of a
larger group of particles. That formula has more constraints and has been tested; it’s pretty good.31

EXAMPLE 3: Second-order processes and O(e2) corrections in i → f.

These are processes where some initial hadronic state goes into some final hadronic state via
electromagnetic interactions. They are processes of O(e2), although they may not involve any explicit photons or
currents. Typically, the only cases in which these can be measured experimentally are those in which the process
is forbidden by the selection rules for the strong interactions, so they arise only because of the electromagnetic
corrections. Examples of such processes include these two decay modes of the η into three pions:

Because the η is even under G-parity it should not decay (via the strong interactions) into three pions, which have
odd G-parity. However, the decay does occur, presumably due to electromagnetic corrections. The η is
electrically neutral and the decay

is allowed. But if there were a G-odd version of the η, a scalar particle with IG = 0−, it could not decay into two γ’s.

Another example of such a process is the electromagnetic mass-splitting within an isomultiplet. In that case i
and f are just a single particle each. In these processes, if we understand quantum electrodynamics but not
necessarily the strong interactions, we have a blob with strong interaction mysteries going on within. From the
inside of this blob we pluck, as if pulling the string on a violin, two pairs of charged fermions (quarks, perhaps)
connected by a photon, as in Figure 35.6. If there are fundamental charged bosons lurking inside the blob, then
we can also have graphs like Figure 35.7, where the charged boson comes out and the photon forms a loop,
because a charged boson couples directly to two photons. This second kind of graph can be eliminated by
choosing a photon propagator which is a bit different from the ones we’ve considered before:
 Figure 35.6 Second order electromagnetic correction to strong interaction process: fermions

Figure 35.7 Second order electromagnetic correction to strong interaction process: bosons

It’s just another choice of α; it’s as good a gauge as any other. It has the advantage that

and therefore Figure 35.7, proportional to , vanishes. The remaining graph, Figure 35.6, can be thought of
(aside from kinematic factors) as proportional to

Therefore, the transformation properties of the first graph are the transformation properties of the products of two
currents. Remember that the individual currents are the sum (35.53) of two parts; for the product of two currents
we simply apply the reasoning leading to (35.62) twice. From the product of the two j µY currents, we combine ΔI =
0 and ΔG = 1 with ΔI = 0 and ΔG = 1 to get overall ΔI = 0 and ΔG = 0. Continuing in this way, and mindful of the
way isospin I adds, we obtain the following:

The last of these, ΔI = 1 and ΔG = 0, is out. If we combine two objects that transform as Iz ’s, we can’t make a
product that behaves as Iz in an isospin 1 state: the isospin 1 state is antisymmetric, so that Clebsch–Gordan
coefficient vanishes. It’s just like the cross-product of two identical vectors; you get zero. That’s the complete list
of transformation properties. The isospin can change by 0, 1 or 2. If it changes by an even amount, it must be ΔG
= 0; if it changes by an odd amount, it must be ΔG = 1.

Let’s return to η decay. The η is G-even:

Its mass is 547.86 MeV. The masses of the π0 and π± are 134.98 MeV and 139.57 MeV, respectively, each less
than ⅓ of the η’s mass. So the decay of the η into three or fewer pions is energetically possible, though some
processes may be otherwise forbidden:

Note the non-conservation of isospin in these last reactions.

In the three-pion decays, G changes by 1 so this decay must have

Since the initial hadronic state, the η, has isospin 0, the final state of three π’s, must be an I = 1 state with Iz = 0.
Thus, although at the moment we know nothing about the momentum distribution of the three final π’s, we know
quite a bit about the isospin dependence of the three-π wave function: it must be a state of total isospin 1. This
should enable us, with a little work, to calculate such things as the ratio of η decay rates:

Next time, I will begin to establish such a connection and test it with experiment. I will then begin a new topic
(and pursue that for several more lectures), one that is qualitatively different from what we have been doing in the
last few weeks. Instead of fancy field theory, we’ll start fancy group theory: I will talk about SU(3).

1[Eds.] In the video of Lecture 35, Coleman quotes many experimental numbers, some as “the best available.”
These numbers were the best available during the years of the videos, 1975–1976. In these lectures the editors
have endeavored to quote the best experimental numbers available to them, circa 2016.
2 [Eds.] PDG 2016, http://pdg.lbl.gov/2016/reviews/rpp2016-rev-phys-constants.pdf.
3 [Eds.] Ibid., http://pdg.lbl.gov/2016/reviews/rpp2016-list-electron.pdf.
4 [Eds.] T. Aoyama et al., “Tenth-Order QED Lepton Anomalous Magnetic Moment: Eighth-Order Vertices
Containing a Second-Order Vacuum Polarization”, Phys. Rev. D 85 (2012) 033007. The theoretical value is cited
in equation (3).
5 [Eds.] Over 12,000 in (e10). T. Aoyama, op. cit.; D. Styer, “Calculation of the anomalous magnetic moment of
the electron”, www.oberlin.edu/physics/dstyer/StrangeQM/Moment.pdf.
6[Eds.] Crease & Mann SC, p. 169, endnote 112 (text on p. 440): “Neither Rabi nor anyone else can remember
where this now-famous remark was first made, but Rabi thinks it was an American Physical Society meeting in
New York City.” Isidor Isaac Rabi, at Columbia from 1929 until his death in 1988, was the winner of the 1944 Nobel
Prize in physics for his discovery of nuclear magnetic resonance, and Julian Schwinger’s thesis advisor. Though
discovered in 1935, the muon’s leptonic nature was not recognized until 1947.
7 [Eds.] A. Hoecker and W. J. Marciano, “The Muon Anomalous Magnetic Moment”, pp. 583–587 in J. Beringer et
al. (Particle Data Group), “Review of Particle Physics”, Phys. Rev. D86 (2012) 010001; equation (3). This is the
average for µ+ and µ−.
8 [Eds.] Ibid., equation (6). Including all standard model contributions (from electroweak and hadronic
interactions), the theoretical value of the muon magnetic moment is given in equation (14) as 1.001 165 918 02 (2)
(42)(26), differing by 2 parts in 1010 from the experimental result (35.10).
9 [Eds.] Problem 18.1, p. 725.
10 [Eds.] Bjorken & Drell Fields, Section 16–11, pp. 166–170. See the equation following (16.173); σ(a2) is
denoted (M2).
11 [Eds.] In fact, it’s about a third again as heavy: mµ = 105.66 MeV; mπ = 139.57 MeV. PDG 2016, p. 32; p. 37.
12 [Eds.] Problem 20.2, p. 817. Note that in this problem, σ(k2) is denoted ρ(k2) to avoid (some) confusion between
the spectral function and the total cross-section σT for e+-e− → hadrons.
13 [Eds.] Hoecker and Marciano, op. cit. See p. 584, equation (13).
14 [Eds.] This admonition recalls John A. Wheeler’s “First Moral Principle”: Never start a calculation before you
know the answer. Wheeler (1911–2008) was a postdoc and colleague of Bohr’s, and the research supervisor of at
least 46 Princeton PhD students, including Kip S. Thorne and Richard Feynman. He is credited for reviving the
study of general relativity in the US after World War II, and popularizing the term “black hole.”
15[Eds.] F. E. Low, “Scattering of Light of Very Low Frequency by Systems of Spin ½”, Phys. Rev. 96 (1954)
1428–1432. Note that this is a “low-energy theorem”, not a “Low energy theorem”! See also Bjorken & Drell Fields
Sect. 19–13, pp. 357–362.
16 [Eds.] See note 11 on p. 738.
17 [Eds.] §16.3, particularly (16.37).
18 [Eds.]
This is yet another statement of the Ward identity; see Peskin & Schroeder QFT, equation (5.79), p. 160
and Section 7.4, pp. 238–244.
19 [Eds.] 0 > (k − k′)2 = (p′ − p)2 = 2m2 − 2p ⋅ p′

⇒
p ⋅ p′ > m2.

20 [Eds.] Bjorken & Drell Fields, pp. 361–362: their equation (19.137) is = (ε ⋅ ε′)2 for k → 0.
Thomson
21 [Eds.] See note 10, p. 520. For notational convenience, we’re using Iz in place of I3.
22 [Eds.]
Here, at 1:12:50 in the video of Lecture 35, there is hissing for about 15 seconds so Coleman cannot be
heard. The end of this sentence comes from Coleman’s own notes.
23[Eds.] T. D. Lee and C. N. Yang, “Charge Conjugation, a New Quantum Number G, and Selection Rules
Concerning a Nucleon-Antinucleon System”, Nuovo Cim. 3 (4) (1956) 749–753; T. D. Lee, Particle Physics and
Introduction to Field Theory, Harwood Academic Publisher, New York, 1981, Section 11.2, “G-Parity”, pp.
225–230; Section 11.3, “Applications to Mesons and Baryons”, pp. 230–240. See also §24.4.
24 [Eds.] See (S14.20), p. 551.
25 [Eds.] M. Gell-Mann and A. Pais, “Behavior of Neutral Particles under Charge Conjugation”, Phys. Rev. 97
(1955) 1387–1389.
26 [Eds.] See note 12 on p. 523.
27 [Eds.] PDG 2016, p. 94.
28 [Eds.] J. J. Sakurai, Modern Quantum Mechanics, Addison-Wesley, 1994, pp. 238–242.
29[Eds.] PDG (2016), p. 94. The quark model predicts that the Σ0 moment is the average of the Σ+ and Σ−
moments: P. Pal, An Introductory Course of Particle Physics, Taylor and Francis, 2015, Section 10.8.2, pp.
283–288.
30 [Eds.] §38.3, pp. 835–839.
31 [Eds.] Coleman adds: “Good enough to get me my first job at Harvard, and tenure. That, and my charm.” He is
referring to his first publication, written with Sheldon L. Glashow; see note 40, p. 841.

Problems 19

19.1 We’ve worked out some general properties of a charged Dirac field minimally coupled to a massless photon.
We

(a) derived the Ward identity for the photon–spinor–antispinor 1PI vertex (33.79);

(b) verified the identity in the tree approximation (33.90);

(c) used the identity to prove that the physical charge, defined as the quantity that appears in the gauge
transformation of the physically renormalized photon field, is the same as the physical charge defined as the
quantity that appears in the vertex with everything on the mass shell (33.100).

We then went on to analyze the kinematic structure of the vertex with the two spinors on the mass shell, but
with the photon carrying arbitrary momentum q. We

(d) showed that there were only two independent form factors F1 and F2;

(e) constructed the explicit expressions (34.15) defining F1(q2) and F2(q2) ;

(f) observed that our result (c) implied (34.16) that F1(0) = 1.

Do the parallel constructions for the case where the charged particle is a scalar.

Comment : This is an easy problem. Step (a) is trivial, since I never used the spin of the charged field in the
derivation. The only change is a notational one, replacing ′ by ′. All the other steps, though non-trivial, are much
easier when you don’t have to worry about spin and γ matrices.
(1998b 8.1)
19.2 Two electrically neutral Dirac fields, ψ 1 and ψ 2, of masses m1 and m2, respectively, interact with a massless
photon through the coupling

where g is a real number and “h. c.” stands for Hermitian conjugate. These fields have no other interactions. If m1
is greater than m2, the decay

is kinematically allowed. Compute the decay width Γ for this process (summed over final spins and averaged over
initial spins) to lowest nontrivial order in perturbation theory.

Comments:

(a) This theory is nonrenormalizable, but for this problem, it doesn’t matter, since we’re only working in tree
approximation.

(b) You are actually computing the decay

to lowest order in electromagnetism, but to all orders in the strong interactions. Just as in the class discussion of
electron–proton scattering (§34.2, in particular the aside starting on p. 738) all the effects of the strong interactions
can be summed up in terms of two form factors, F1 and F2. (Because the incoming and outgoing hadrons have
different masses, the detailed definition of F1 is a little different than in class, but this is not important here.) At q2 =
0 one can show that F1(q2) = 0; otherwise one would have a very strange inverse-square force between neutral
particles. So all the difficult-to-compute effects of the strong interactions are summed up in a single number, F2(0).
(1991b 6.2)

Solutions 19

19.1 (a) Following the definition (33.56), let ′(n1,n2,m) stand for the full Green’s functions, with n1 initial scalar
particles, n2 final scalar particles, and m photons. As in (33.88), the Ward identity (actually the Ward–Takahashi
identity) is

where p′ = p + q. For scalar particles we have (32.16)

and so

(b) In the tree approximation, we have

as well as

and the identity is verified at the tree level; cf. (33.90).

(c) The physical charge ephys is defined by the condition

with everything on the mass shell: p2 = p′2 = m2, q2 = 0. Differentiating (S19.3) with respect to qµ, we get

In the limits p2 → m2, p′2 → m2, q2 → 0, we find

In these same limits, (S19.6) becomes

Comparing (S19.8) with (S19.6), we obtain as in (33.100)

(d) In §34.1, we used crossing symmetry and angular momentum arguments to consider the photon as decaying
into a charged particle–antiparticle pair. Since the photon carries quantum numbers JPC = 1−−, and the charged
particles are scalars, the final state must have j = ℓ = 1. Thus there is only one invariant amplitude, and hence only
one vertex function, unlike in spinor electrodynamics.

(e) Analogous to rule (c) in the box of Feynman rules for scalar electrodynamics on p. 644, define the sole
invariant vertex function by

for p2 = p′2 = m2. From (S19.6) and (S19.10), we know

Taking the limit of (S19.11) as q → 0 gives

which was to be shown.

■

19.2 The relevant Feynman diagram is:

We can always choose εµ such that

With this choice,

Also,

Then

Averaging over initial spins and summing over final spins, as well as the two polarization states gives
The trace of an odd number of γ’s is zero, and = 0:

using (S11.16), p. 428 for the trace of four γ’s. We can reduce the dot products because p = p′ + k, and so

Then

The decay width is given by (12.33),

The integration over the solid angle gives 4π. In the center of momentum frame, ET = m1, p ⋅ k = m1k0, and

so that finally

is the decay width.

■

36
Introducing SU(3)

At the end of the last lecture, we were discussing the phenomenology of η decays. I have a little more to say about
that topic, and then we’ll begin a discussion of the approximate symmetry group SU(3).

36.1Decays of the η

Recall that the η is a spin-0 meson with negative parity and positive G-parity:

The pion has IG = 1− (35.59). The η decay

must be into a state with I = 1, Iz = 0, because the only part of the second-order electromagnetic interaction that
can change G carries ΔI = 1; see (35.73). (Remember, Iz is conserved by electromagnetic interactions, but I itself
is not.) The final state must be the Iz = 0 member of an isotriplet. This enables us to connect the decays η → 3π0
and η → π+π−π0 by Clebsch–Gordan considerations of the isospin. We’ll look at that process in some detail to get
an idea of what sort of restrictions the Clebsch–Gordan arguments impose. We’ll have to introduce a set of
variables for the three-pion system. We talked about the kinematics of three-particle decays in §12.5. We’ll also
make use of Dalitz plots (see the discussions following Figure 11.7 on p. 235, and Figure 12.5 on p. 256).1

Convenient variables are the energies of the three pions. Those are not all independent, of course, because

Thus the allowed points in the diagram can either be in the E1, E2 or E3, etc., plane, which form the Dalitz plot
(here, {1, 2, 3} label the three pions). For our purposes it will be useful to introduce new variables

The center of the Dalitz plot, assuming we ignore the electromagnetic mass differences between the pions, is the
point where all the ϵ’s vanish. Treating the three pions for the moment as distinguishable particles (which we
certainly can, except for a set of measure zero in the Dalitz plot) we will introduce three isospin unit vectors

just like the polarization vectors we had for photons, except that in this case they measure the directions of the
three one-particle states in isospin space. For the different pion states

How many amplitudes can we construct that have I = 1? This is pretty easy. Label a representation of the
rotation group by its spin. If we put together two pions, without worrying about statistics or anything, we can
construct a state of isospin 0, 1, or 2:

(This is the Clebsch–Gordan series (18.73) applied to isospin. That is, the combined states have I1 + I2, I1 + I2 − 1,
|I1 − I2|). If we put in a third pion, we get isospin 1 three times, one from each of the factors:

We don’t care about the non-1 part. Thus we should be able to construct three functions of the ϵi’s, linear in the
three e’s, that transform like an isovector under isotopic rotations. It is easy to see what those functions are:

where F, G and H are functions of the ϵi’s. The amplitude A is the sum of three linearly independent amplitudes
that transform like isovectors. The generalized exclusion principle tells us that the total amplitude must be fully
symmetric when we interchange space, spin (not relevant here), and isospin variables. So the first thing we note
is

because 1( 2 ⋅ 3) is already symmetric under the interchange of 2 and 3. Likewise G and H must be the same
function as F: the amplitude must have the form

The first entry in each F is the preferred position for that case. We’ve now constructed the most general decay
amplitude that has the proper isospin transformation properties consistent with Bose statistics. Of course, the
actual decay final state is the Iz = 0 member of this triplet, because electromagnetism preserves Iz . We project it
out by dotting the amplitude A into a unit vector pointing in the z-direction:

The next stage in the analysis is to make an approximation, which will introduce a small error. The ϵ’s are all
fairly small. Even in the case of the 3π0 decay, the mass of the neutral pion is 135 MeV, so the rest energy of three
neutral pions is 405 MeV. Thus we only have 143 MeV available for the decay, which has to be split somehow
among the three pions. In the case of charged pion decay, π+π−π0, the situation is even worse, with even less
energy: the charged pions are each 4.6 MeV heavier than the neutral pion, so we have 9.2 MeV less than the 3π0
decay. This means that the effects of electromagnetic mass differences, usually negligible, are in fact significant
for this process, because they cut down the available amount of phase space by 10 to 15%. They are large effects
if we have very little energy available, and even a small mass difference can be a large effect. But we ignore that
and live with this possible error.

What we will not ignore is this: because there is a small amount of energy available for the decay, the ϵ’s tend
to be small. Therefore we will make a linear approximation for the function F. That is, in phase space Γ, we will
ignore

where D is the region of the Dalitz plot. This could have been |ϵi|2; it doesn’t matter because it’s all symmetric. We
will expand F only to first order in ϵ and then ignore, in computing the total cross-section, quadratic effects,
because we expect ϵ to be small compared to a typical strong interaction mass over the entirety of the Dalitz plot;
therefore effects of order ϵ2 should be negligible.

Thus we will approximate F as

where B′ = B − C since ϵ1 + ϵ2 + ϵ3 = 0. A is a constant, the value at the center of the Dalitz plot. The energies ϵ2
and ϵ3 must have the same coefficient, because F(ϵ1, ϵ2, ϵ3) must be symmetric under interchange of ϵ2 and ϵ3.
We’ve certainly made some error by this approximation, and we expect that error to be about 10 to 15%.

Now we’re in business. We’ll first consider

In this case

The • and i • j factors in all three terms contribute 1, and we have a decay amplitude

In the linear approximation, the distribution of points in the Dalitz plot is completely flat for 3π0 decays,
independent of position in the Dalitz plot.2

For the other decay,

it doesn’t matter how we choose the ’s. Let

From (36.10) only the first term contributes since 1 is orthogonal to 2 and 3, so we obtain for the decay
amplitude just the first term

Thus the distribution in the Dalitz plot in the linear approximation is symmetric, independent of the π+ and π−
energies, and linearly dependent on how much energy is given to the π0.

Now let’s compute the total decay rates. At first glance it looks like we can’t compare the total decay rates,
because one of them depends only on A and the other one depends on both A and B′:

Ignoring the mass differences between the pions, the Dalitz plot is completely symmetric under the interchange ϵ1
↔ ϵ2. Therefore the integral over the Dalitz plot of ϵ1 is the same as the integral over the Dalitz plot of ϵ2 or ϵ3 or
equivalently ⅓ the integral of the sum. But the sum is zero and therefore its integral is zero, so the ϵ1 term in
(36.22) integrates to 0. Therefore, although in this approximation we see a linear term in the distribution of points in
the Dalitz plot, the actual total number of points in the Dalitz plot is unaffected except by quadratic terms. Points
are shifted to one side or the other of the line where ϵ1 = 0, but the same number go to one side as to the other.

The 3π0 decay is much more straightforward. From (36.18),

The (1/3!) is to ensure that we don’t count the same experimental event 6 times. The total decay rate is difficult to
measure but the branching ratio is well-known. The experimental numbers are3

which is in agreement within the expected theoretical error caused by neglecting the three pion mass differences.
It’s about 5% off from the theoretical value of = 1.5.

It could have been that the linear approximation was poor, but it’s not. If we actually look at the density of
points in the Dalitz plot, the linear approximation fits them pretty well. To improve the approximation, we would
have to look at how the size of the Dalitz plot changes with available energy, and make a correction for that. We’d
just have less phase space to decay into. That would enhance the 3π0 decay because there’s more phase space
around; the π0’s are lighter.

The big problem in η decay is not phase space but determining the absolute rate. This is very tricky, involving
something called the Primakoff effect. 4 It involves η production from a photon interacting with the Coulomb field of
a nucleus. This is a difficult experiment and the results keep fluctuating. The big theoretical problem is that unless
the experiments are very badly wrong, the amplitude is embarrassingly larger than it should be on the basis of
crude estimates of the size of an electromagnetic effect. With that, we leave η decays, and begin a new topic.

36.2An informal historical introduction to SU(3)

We’ve seen that we can get a lot of results about properties that have the strong interactions entering into them in
a complicated way, even in almost total ignorance of the strong interactions, just by exploiting the known
symmetries of the strong interactions and the known transformation properties of the other interactions under the
strong interaction symmetry group (which, however, was not known for a long time). Here is a short historical
survey of how we came to SU(3).5

In the mid-to-late 1950’s some very smart people, including Murray Gell-Mann and Julian Schwinger, began
thinking that maybe one could play this game even more daringly. What principally made them think this was, first,
the introduction by Gell-Mann and Nishijima a few years earlier, of strangeness or hypercharge,6 which indicated
that isospin seemed to be a good quantum number for all the new particles that had been coming out of the new
generation of high-energy machines (low-energy machines by today’s standards); exotic particles like Λ’s and Σ’s,
K mesons, ρ’s and ω’s, all fit into isospin multiplets. And they were beginning to be assembled into even larger
families. In particular there seemed to be eight so-called baryons, strongly interacting particles with spin ½, that
were pretty much like the nucleons. All eight were assigned baryon number +1, and parity plus,

though they had different isospins and hypercharges. They were rather close together in mass by the scale of
these things, as shown in Table 36.1. These particles seemed vaguely similar in their properties. As far as the
experimental evidence went, they seemed to have the same parity, they unquestionably had spin ½ and baryon
number 1 and they were all relatively close together in mass, the mass splittings between the heaviest and the
lightest being of the order of 15 to 20% of the mean mass of this collection of particles.
 Table 36.1 The eight baryons: JP = +, B = 1

This led them to an idea. Back in the 1930’s, when the only strongly interacting particles known were the
proton and neutron, Heisenberg and others had suggested that if we neglected electromagnetism, then because
of the strong interactions that remained (“nuclear forces”, as they were then called), the world would be much
more symmetric than it was in reality.7 In particular it would possess isospin symmetry. Therefore, Schwinger and
Gell-Mann, at around the same time, said maybe the same thing could be done with the strong interactions.
Maybe they split into two families, very strong and medium strong, which we will ignore:

Guided by the principle that ignoring electromagnetism leads to the neutron and proton having the same mass,
they hypothesized that if the medium strong interactions were ignored—a much bolder step than ignoring
electromagnetism—then all eight of these particles would have the same mass and would be part of a degenerate
multiplet of a larger symmetry group than isospin. This hypothetical larger symmetry wouldn’t be as good as
isospin; a 10 to 20% error is much worse than a 1% error or a 0.1% error. But it would still better than nothing. It’s
a lot easier to try this group theory idea than to attempt to solve the dynamics of the strong interactions.

Criteria on G

It was clear what the problem was. In mathematical language, we want some internal symmetry group, G, that
first of all contains a product of the SU(2) of isospin and the U(1) of hypercharge:

We want the new symmetry group to include the old symmetries when we don’t ignore the medium strong
interactions. Next, G must have an 8-dimensional, unitary, irreducible representation,8 to accommodate the eight
observed baryons in a single representation of G. And because that representation is irreducible they will all have
the same mass. Otherwise SU(2) ⊗ U(1) (which lacks such a representation) would solve the problem. Finally,
when we reduce G to this subgroup, G → SU(2) ⊗ U(1), we don’t want just any 8-dimensional irreducible
representation, but one that decomposes into these observed particles in Table 36.1 exactly:

(that is, (I = , Y = −1), (I = 1, Y = 0), (I = 0, Y = 0), and (I = , Y = 1), respectively).

At that time nobody knew anything about group theory except Wigner, and he wasn’t talking.9 They were sort
of desperate, so they played around, they guessed at it; Schwinger10 and Gell-Mann11 made the same guess,
called global symmetry. There wasn’t much data at the time so it wasn’t immediately obvious, but after a year or
two it became clear that this idea was dead wrong. People tried other things. There was a whole school of
thought. Saul Barshay said the Λ had opposite parity from the Σ; they were just bad experiments and we should
look for a seven-dimensional representation; the Λ was coincidentally sitting there in the middle. 12 There were
people who fiddled around with the idea that maybe there were nine particles and we just hadn’t found the ninth
one yet because it was little bit heavier than the Ξ; perhaps we should have been looking for a group with a nine-
dimensional representation. People played this game for a while. But it became clear, and it’s certainly known
now in retrospect, that there aren’t any other particles around in the same mass range. Insofar as the weak
interactions allow us to define the relative parities, the parities of these eight baryons are all the same. Therefore,
the problem as originally framed, with the three criteria above, is correctly posed.

A few years later, around 1960, Gell-Mann guessed the right group. It’s a very interesting anecdote to me,
because I was present at the time as a graduate student at Caltech. At that time Gell-Mann and Shelly Glashow,
who was a post-doctoral fellow at Caltech, were working on a Yang–Mills theory, not knowing what it was good
for.13 (Steve Weinberg discovered that a few years later.) They wanted to learn something about Lie groups
(which are involved in Yang–Mills theories). At that time Lie group theory was considered recherché, like fiber
bundles, not something a respectable physicist knew about. I then had a totally undeserved reputation for
mathematical sophistication. They asked me, “Do you know anything about Lie group theory?” I replied, “Who,
me? I can tell an ϵ from a δ but that doesn’t mean I’m André Weil; of course I don’t.”14 Well, fortunately at that
time the Caltech mathematics department was in the same building as the physics department. Murray went
upstairs and found a mathematician, Richard Block,15 who was willing to talk to him. Block told him to consult a
book16 in French which was a précis of the main results on Lie group theory.17 Murray’s fluent in French so he did
just that. And two days later, he returned in a state of great excitement and said, “There is a group with an 8-
dimensional representation called SU(3).” And we said, “SU(3)? 8-dimensional representation? Ah, go away.” It
turned out to work, as we all discovered very shortly. That was how it in fact happened historically.

A second physicist, Yuval Ne’eman18 made the discovery independently.19 Yuval had been trained as an
engineer and he had wanted to become a physicist since 1947, but the military situation kept him from going to
graduate school. And finally, after having served as acting head of Israeli military intelligence during the Sinai
campaign (besides other wartime experience), he was able to talk the Israeli general staff into getting him a half-
time job as a military attaché in London. The other half of his time he spent at graduate school at Imperial College
under Abdus Salam. At first, the British Foreign Office was reluctant about giving him permission to do this,
because they confused high energy physics with building bombs, as if he wanted to be a spy. But he finally got the
head of the Israeli general staff, Moshe Dayan,20 to write a letter explaining that Ne’eman’s objective was
education, not espionage. And Yuval subsequently described himself as the only graduate student accepted by
Salam on the strength of a letter of recommendation from Moshe Dayan.21 Salam put him on this problem, and he
also came up with SU(3) at around the same time as Gell-Mann.22 I remember Murray rushing into the department
with a preprint, exclaiming “Some Israeli colonel has made the same discovery!” Enough of stories from my youth.
(The golden age of a physicist is 23...)

Later, people began to wonder if something had slipped through the net, and started to look at the problem in
a systematic way, too late as always, and to investigate the mathematics to reach the solution. That’s what we’ll
do in lieu of the historical order; we’ll turn to the mathematical answer. I won’t prove this is the answer, because it’s
not a sophisticated proof, it’s just tedious. Then we’ll systematically explore the possibilities given by this answer.
We’ll see that nothing works except SU(3), and then we’ll spend a lot of time on SU(3).

The answer is as follows.23 Every group G satisfying these three criteria

•It has an 8 dimensional irreducible representation, D(G)

•It contains SU(2) ⊗ U(1) as a subgroup

•D(G) → {Ξ, N, Σ, Λ}exactly as G → SU(2) ⊗ U(1)

contains as a subgroup either

A.a group G0 satisfying minimal global symmetry, equating some meson–baryon coupling constants.
(Recall that the original erroneous suggestion of Schwinger and Gell-Mann, “global symmetry”,
suggested all pion–baryon coupling constants were equal.) Later Lee and Yang24 pointed out that
the Gell-Mann–Schwinger group contained a subgroup (to be described below) that was a better fit
with experiments: it gave almost the same results of the original group, but with fewer wrong
predictions.

B.SU(3), the group of all 3 × 3 unitary unimodular (determinant = 1) matrices.

Of course there are many groups containing SU(3). One answer to the problem is obviously SU(8), the group of all
8 × 8 unitary unimodular matrices. That contains everything that satisfies the problem. But these groups either
contain minimal global symmetry, or SU(3). We will systematically investigate these two possibilities.

First let’s consider option A, a group G0 that satisfies minimal global symmetry.25 G0 is the product of three
SU(2) factors:

People thought to try this, because at the time the only Lie group they knew about was isospin’s SU(2) Thus its
generators can be written as the three commuting triplets of isospin generators:
How are the isospin and hypercharge embedded as subgroups? I is the simultaneous rotation in the 1–2 isospin
space

and Y is twice the z-component of I(3):

These commute and obviously obey the SU(2) ⊗ U(1) algebra.

The representation D(G0) to which the eight baryons are assigned is in fact a reducible representation of this
group. A general representation is labeled by three spins, s1, s2 and s3, for the three isospins, D (s 1,s 2,s 3), just as
two spins came into our analysis of the Lorentz group.26 The representation in question is the direct sum

G0 is therefore not a solution to the problem because this is not irreducible. But the theorem says G0 is contained
in many solutions to the problem, not that it is itself the solution. If we reduce to isospin and hypercharge alone

then the first factor in (36.33), D ( , ,0), has the sum of two isospin-½’s added together and we obtain isospin 1 and
isospin 0:

both with Y = 0, since D ( , ,0) is a singlet under the third isospin group and therefore carries zero hypercharge.
This is the Σ and the Λ in Table 36.1. In the second factor, D ( ,0, ) we only have isospin + 0 = , so both factors
have I = , but because Iz can have the two values and − , we have Y = 2Iz (3) = ±1; that’s the N and the Ξ.

Indeed, the first thing people said when they looked at this pattern was, “Well, obviously there is some sort of
reflection or something at work here, some symmetry operation that exchanges the N and the Ξ, and changes the
sign of the hypercharge. They look the same, they’re at opposite ends of the chart, flip it around.” In fact, that
guess turned out to be dead wrong, as we will demonstrate immediately. It was however very plausible, and it led
to the group (36.29).

Well, this group is in direct contradiction to many experimental results, although of course none of them was
known at the time it was proposed. The easiest way27 to establish a contradiction is to define a group element
called R:

It is a rotation by 180° about the y-axis in the third isospin group. R, belonging to I(3), commutes with isospin,
which is constructed from the first and second isospin groups, (36.31). But it changes the sign of hypercharge,
because that’s proportional to Iz (3):

From the existence of the group element R, one can derive an almost endless string of contradictions with
experiment. For example, for every hadron (stable or unstable) with a given non-zero hypercharge, there must be
another hadron of the same mass (within 10 to 20%) with the opposite hypercharge, obtained by applying R, which
changes the sign of the hypercharge while leaving isospin alone. And of course R commutes with baryon number,
so this second hadron is not the antiparticle of the first; it’s something else.

Now there’s the well-known Δ(1232), a big, fat28 resonance in pion–nucleon scattering with I = and Y = 1, so
there should be another big, fat resonance around the same mass with I = and Y = −1. You can search the
Rosenfeld tables29 to your heart’s content, but you will not find such an object; it is not there. Of course at the time
they didn’t know it wasn’t there.

Another argument that (36.29) is the wrong group comes from the analysis of magnetic moments that I
described last time (Example 2, p. 766). The Λ is a singlet under the third isospin, so R doesn’t do anything to the
Λ. If we have a one-Λ intermediate state, we get the same state after operating on it with R:

On the other hand, if we look at the Y-component of the electromagnetic current, the part that comes from jYµ, R
changes its sign, because it changes the sign of Y :

If we look at the electromagnetic current between two Λ states, from isospin considerations, only j Yµ contributes:
j Izµ is a component of an isovector, and cannot have a non-vanishing matrix element between two isospin-zero

states. But by R this matrix element is zero. Apply R †R on the two sides of j Yµ; the Λ doesn’t change sign, but the
current does, so the matrix element vanishes:

by (36.38) and (36.39). Thus we are led to the conclusion that the magnetic moment of the Λ must be zero:

The magnetic moment of the Λ wasn’t measured until 1963, so the prediction (36.41) didn’t bother anyone at the
time. But we know now the Λ moment is about a third of the neutron moment,30 a typical hadron magnetic
moment, not particularly small in any reasonable sense.

I could go on. We could run through the Particle Data Group tables, demonstrating that any symmetry
including a hypercharge reflection operator of this kind is guaranteed to be wrong. The universe is not symmetric
under a change of sign of hypercharge, while keeping the sign of isospin and the baryon number unchanged.
Thus, G0 is out.31

Therefore, if we accept this theorem (stated without proof), the last best guess is SU(3): either that works or
the game is up. Well, of course it does work, otherwise I wouldn’t be giving this lecture. Rather than doing things
directly for SU(3), I’d like to devote some time to mathematical preliminaries in which we construct the
representations of SU(3). After we have our mathematical machinery set up, we’ll apply it to a variety of physics
problems. It’s worth the effort, because we know in advance that it’s going to be good for physics.

36.3Tensor methods for SU(n)

We start with the hadrons in Table 36.1 as an eight-dimensional representation of some group and an embedding
of SU(2) ⊗ U(1) in that group. It’s extremely tedious. You look in books on Lie group theory and count up all the Lie
groups with 8-dimensional irreducible representations, (or 4-dimensional representations, because it could be two
4’s connected by a discrete element). You write them down and figure out all possible ways of fitting in isospin and
hypercharge, and you are left with, I believe, 13 possible groups. And then you figure out which contains which
one, you make a big diagram with boxes and trees, and you find that they all end up either in G0 or in SU(3). I did it
for my thesis;32 believe me, it was dull. At the end of it, though, I knew Lie group theory. If I had known Lie group
theory a year before... Ah, well.

So we will begin by making some preliminary remarks about the representations of SU(n), for arbitrary n.
Then we will specialize, first all the way down to SU(2), just so we can check that the methods work in a case
where we already know the answer, and then to SU(3). This is in fact the method introduced by Hermann Weyl in
his book on the classical groups.33

SU(n) is the group of all n × n unitary matrices U with determinant 1:

We already know one representation of SU(n), to wit, an n-dimensional representation, where the group is
represented by the matrices themselves. A complex n-vector x transforms under the action of the group according
to the rule
It is convenient to write these transformations out in index form as if they were Lorentz tensors, except they’re not,
they’re SU(n) vectors:

For the moment the reason that I put one of those indices downstairs is just perverse, but I am adopting the
summation convention.

Another representation that we know off-hand is the complex conjugate representation. The complex
conjugate vectors, (yi)∗, form the basis for the conjugate representation. We indicate the components of a
conjugate vector by a subscript:

Given the representation (36.44), then the complex conjugate is also a representation:

using the complex conjugate matrix. It may or may not be equivalent to (36.44). These two representations are
equivalent in SU(2) but not in SU(3) or higher SU(n). We use a notation that mimics that of ordinary four-
dimensional tensor analysis. The mimicry is introduced for a reason: it is to remind us that if we take one vector
that transforms as the first representation and the second according to the conjugate representation and sum
them up, then the summation of upper and lower indices is an invariant operation; that’s just the definition of a
unitary matrix. It’s precisely the object that preserves the quadratic form:

Thus we have two kinds of vectors, upper-index vectors and lower-index vectors, just as in ordinary tensor
analysis. What we don’t have is a metric tensor that allows us to raise and lower indices. At this level they’re just
two different kinds of objects that transform in two different ways. We may form tensors with arbitrary numbers of
upper and lower indices by taking direct products of vectors and conjugate vectors:

Upper indices transform as if they’re upstairs vectors, lower indices as downstairs, conjugate vectors. I won’t take
the time to write it out, because you can see how it goes: there’s a U for every upstairs index and a U ∗ for every
downstairs index. These tensors form a bunch of representations of our group SU(n). Of course they’re not
guaranteed to be irreducible, but they are guaranteed to be representations.

There are all sorts of manipulations we can do on these tensors that are invariant operations. For one thing,
symmetrizing or antisymmetrizing on a pair of upper or lower indices is an invariant operation. That’s because any
two upper indices transform in the same way. So if we break a tensor up into a part that’s symmetric on its first two
upper indices and a part that’s antisymmetric on the first two upper indices, then if we make the transformation the
symmetric part goes into something symmetric and the antisymmetric part goes into something antisymmetric.
Likewise for lower indices. On the other hand, it’s pointless to symmetrize a tensor between an upper index and a
lower index; it’s allowed, but if we make a transformation on that tensor, the symmetry won’t be preserved: the
upper index transforms differently than the lower index. Another invariant operation is contraction: summing an
upper and a lower index (this is the trace of the matrix over the two summed indices):

We cannot, however, sum two lower indices together, nor two upper indices, because there’s no metric tensor.

There are also some invariant tensors around; tensors, which when transformed according to the rules, don’t
transform at all. One is the Kronecker delta, :

That of course is just the statement that our matrices are unitary or equivalently, that summing an upper and a
lower index is an invariant operation. Another invariant tensor is the Levi–Civita ϵ. Using the identity

for an n × n matrix , we see that under the action of SU(n),

because det U † = 1. Likewise, ϵijk p is invariant:

I will now explain Hermann Weyl’s program for SU(n), which for n > 3 requires exquisite knowledge of the
representations of the permutation group.34 Thankfully, no such expertise is needed for either SU(2) or SU(3).

Weyl’s program attempts to find all the representations in the following way. Take all these tensors;
symmetrize and antisymmetrize and multiply by epsilon tensors like crazy, until we’ve gone as far as we can,
constructing invariant subspaces from the set of n-index tensors that are irreducible (we hope). We then prove that
these are in fact irreducible representations, crossing our fingers that some clever graduate student won’t come
along and say “Ha! You forgot about contracting this index with that index,” so we can get an even smaller
subspace. If we can prove that you can’t reduce the subspace, and hence that these representations are
irreducible, then we will have constructed a complete set of irreducible representations of SU(n). That’s the
program. The amazing thing is that the program works (if you’re Hermann Weyl). Even if you’re not, it’s pretty
easy to make the program work for SU(2) and SU(3).

Weyl’s program for SU(2)

We will work through the process for SU(2). In this case each of the epsilon tensors has only two indices. That
means we can use them to raise or lower indices, just like the metric tensor:

Of course this is not like the (symmetric) metric tensor, gµν; the ϵ’s are antisymmetric, but they are still invariant
objects with two indices. Given any tensor with a bunch of upper and lower indices, we can always convert it into a
tensor with only upper indices by raising indices with the aid of the epsilon tensor, or only lower indices by lowering
with the epsilon. Thus we need only look at tensors with all indices either upper or lower.35

Next, we can always write tensors with more than one index as a sum of a symmetric and an antisymmetric
part. Let’s take a tensor xij of rank 2. We can break that up into its symmetric and antisymmetric parts:

Since it only has two indices, the antisymmetric part must be proportional to the epsilon tensor (two indices,
antisymmetric):

Here x is a scalar. (In the analysis of the Lorentz group, we similarly (18.91) split up a symmetric tensor into a part
equal to its trace times gµν, and a traceless, symmetric part.)

Thus if we have a tensor of mixed symmetry, we can always symmetrize it. The antisymmetric part can be
written in terms of tensors of lower rank (with fewer indices), with the ϵ’s taking up the missing indices. As we
systematically examine bigger and bigger tensors looking for new representations, the only tensors we have to
consider are those which have only upper indices and are fully symmetric.

What can we do to simplify those tensors? Let’s guess. Our guess, which we will try to verify, is that the
irreducible representations are generated by transformations on the space of these tensors. We will, with the
benefit of hindsight, describe the number of their indices by the integer 2s:

(adopted so we can see the connection with standard notation for SU(2)). Call the representation D (s)(U), a unitary
unimodular matrix or, for short, (s):

What is the dimension of (s)? How many independent tensors there are with the desired symmetry
properties? Since the tensor is completely symmetric, the only significant feature about a component is how many
1’s it has and how many 2’s it has. How they’re distributed is completely irrelevant; that’s what complete symmetry
means. So the question is: if we have 2s objects and we want to put them into two boxes, how many ways are
there of doing it? One box will hold the 1 indices, the other holds the 2’s.

This is easy.36 We imagine the 2s objects are written out on a line. Imagine we have a wall which we put
down somewhere in between the dots. Everything to the left will be a 1, everything to the right will be a 2; that’s the
two boxes. For a tensor with s = 3 and r, the number of 1’s, equal to 2, the diagram looks like this:

There are only (2s + 1) places we can put the wall, starting to the left of all the dots and ending to the right of all the
dots. So for SU(2)

We’ve seen that factor (2s + 1) before, in the context of angular momentum, and we know it’s right for SU(2).37
(For SU(3) there will be three boxes in the corresponding computation.) Writing

and with Iz = σz (the usual Pauli matrix), the eigenvalues are the familiar

For the subgroup of pure Iz rotations,38 with Iz = σz , for any angle θ,

That’s how we embed Iz rotations in SU(2). It’s a unitary matrix of determinant 1. (Note that in spinor space, the
identity corresponds to an Iz rotation through an angle of 4π.) Then

If we make an isospin rotation for this particular U on a tensor component with r 1’s and (2s − r) 2’s,

where r = 0, 1, 2, . . . , 2s. Then

so that the eigenvalues of Iz are −s, −s + 1, . . . , s − 1, s, which is of course the correct Iz content of a
representation of SU(2). Getting them was no great triumph, but we wanted to check that this method works.

36.4Applying tensor methods in SU(2)

Let’s work out how the field theory of pions and nucleons would look in this notation. The pions have isospin 1 and
the nucleons have isospin ½. We begin with the representation D ( ) = ( ), which is supposed to be the nucleon
field. Instead of calling it x, we’ll call it N: it’s the nucleon field with space and spin dependence suppressed:

Under a transformation U ∈ SU(2) we expect N → UN, i.e.,

The conjugate fields transform according to the conjugate representation:

The definition of the conjugate representation gives us the complex conjugate. We write the vector on the left
rather than on the right; this switches the indices and gives us the transpose. Taken together, we get the Hermitian
adjoint, U †. For example, the bispinor product that we use to make an invariant Lagrangian is

which is manifestly SU(2)-invariant; no surprise.

Aside. You can make a representation that transforms like a row vector, from one that transforms as a column
vector, by using the ϵ tensor:

Although that doesn’t involve conjugate fields, it would transform like a row vector, just like the conjugate fields.
And because

we would have

That’s why when we build an isospin singlet two-nucleon state, it is

The minus sign from the ϵ tensor is doing the job. That’s the rule for putting two spin-½ objects together to make a
spin-0 object (the singlet configuration, if we were talking about ordinary spin, rather than isospin). The spin-0
combination is antisymmetric; the spin-1 combination is symmetric. We either complex conjugate to lower the
index (which makes antinucleons) or multiply by the ϵ tensor to lower the index.

Now let’s turn to the pions (or the Σ’s, or any other system with isospin 1) with the aim of finding invariant
quantities suitable for a Lagrangian. The pion corresponds to a symmetric tensor with two indices:

An equivalent expression (and as it turns out, a more convenient choice) is obtained by lowering one of the
indices with the ϵ tensor:

Of course, the mixed object is no longer symmetric but it still has a constraint in it; it is traceless:

because ϕ ik is symmetric. Because we’ve lowered an index this transforms as the outer product xi ⊗ yj of one row
vector yj and one column vector xi. Under the action of U we can write things in matrix form

This transforms not like the inner product of row times column, which is a scalar, but as the outer product of
column times row, which is an object with two indices. If we do things this way, it is obvious that it’s consistent with
the group to impose the condition that

We’ll do that for the pions, because we want three real fields. That corresponds to a traceless 2 × 2 Hermitian
matrix. If we were dealing with the Σ’s, where we’d want three complex fields, we wouldn’t impose (36.78).
(Physically, the antiparticle of the π+ is the π−, but the antiparticle of the Σ+ is not the Σ−, but an entirely different
particle with baryon number −1.)

We can see how the 2 × 2 matrix ϕ transforms by multiplying

and working out the properties of the individual components. This will help us identify the components of the ϕ
tensor. NN is a typical two-index object, one upper and one lower, that transforms like . In the 1-1 spot we have
p , an object which has zero charge and carries Iz = 0. Therefore must be some number α, times the neutral
pion field ϕ 0. In the 1-2 spot, p carries charge 1 and has Iz = 1, so must be some multiple of the positive pion
field. We’ll just call it ϕ + since how we scale our fields is a matter of taste. By conjugation we must have ϕ – in the
or n spot, with Iz = −1, and by the requirement that is traceless, we have −αϕ 0 in the or n
spot, also with zero charge and Iz = 0:

How we fix α is again a matter of taste. We can normalize any one of our independent dynamical variables any
way we please. But it’s convenient to take (as a possible term in a Lagrangian) the expression

which is invariant under U:

Now

Earlier, with

and ϕ 0 = ϕ 3, we found (24.27) the invariant mass term contained the expression

So we’ll choose

to ensure that the mass term will come out right. Then39

Therefore, the free Lagrangian for the pion can be written as

After taking the traces, this is the standard expression (24.27), but here written in a form that is manifestly SU(2)-
invariant.

The nice thing about doing things this way is that it’s easy to write down couplings that are manifestly SU(2)-
invariant. Well, it was easy to write them before, but don’t forget that last time, in §24.1, we spent some time, from
(24.11) to (24.14), working out the Clebsch–Gordan coefficients for the rotation group. But now we can do it in one
line. We have a matrix ϕ. We have a row vector and a column vector, the nucleon fields and the antinucleon
fields. We want to form an object that is an SU(2) scalar. If we’re talking about Yukawa coupling, it’s very easy to
see that it must be

our old friend. This product—row vector, matrix, column vector—is a scalar: the U’s cancel the U †’s and the whole
thing is obviously invariant.

We are using our prior knowledge of SU(2), in particular that the representations are irreducible, which we
showed in §18.2. We don’t have any prior knowledge of SU(3), so when we begin studying it, we will have to work
to show that our representations are irreducible. The form on the left-hand side of (36.87) is equivalent to dotting
the generators τi with the vector ϕ i. It’s exactly the same form. All those 1/ ’s that we got last time in the isospin-
invariant interaction come out automatically here. It’s just (24.20), to within a factor of :
There is only one SU(2)-invariant Yukawa interaction. Since both methods are right, both methods must give the
same result. Instead of fiddling with raising and lowering operators, it now comes from the condition that we want
the trace of the square of the matrix to be properly normalized. As before (25.77), we’ll need a ϕ 4 interaction for
renormalizability:

That’s all I want to say about SU(2), in this matrix and vector notation. It’s inferior to the other way of doing
things. It’s nice if we’re working with isospin-½ and isospin-1, but when we go to an object with higher isospin, say
to isospin- , we can certainly write it as a 3-index tensor, but there’s no nice way of writing a 3-index object as a
matrix. Still, the tensor notation is useful. First, it is completely general, with all those indices. It’s like van der
Waerden’s method of treating the Lorentz group as a product of SU(2) ⊗ SU(2), where he has two kinds of
indices, called “dotted” and “undotted”, much used in the literature.40 Second, the matrix trick is nice if we’re only
worried about certain selected representations of low dimensionality, which will be our situation in SU(3).

36.5Tensor representations of SU(3)

In SU(3) everything is the same as in SU(2), except that the epsilon tensor ϵijk has three indices. This makes all
the difference in the world. For one thing, we can’t make the epsilon tensor act as an ersatz metric tensor to raise
and lower indices. We can lower an index with it but at the same time we raise the rank of the tensor by 1, which is
bad. On the other hand, we can still write a given antisymmetric tensor in terms of a vector:

a familiar trick from the three-dimensional rotation group, where we write an antisymmetric 3 × 3 matrix as an axial
vector. We can thus get rid of antisymmetric parts, blithely reducing them as we move to tensors of higher rank to
objects of lower rank which we have presumably already investigated in our iterative procedure. But we cannot get
rid of either the upper or the lower indices; we’re going to have to live with both of them. Therefore the sort of
object we arrive at is a tensor with n upper indices, m lower indices, completely symmetric in both sets because we
can always get rid of the antisymmetric parts using ϵ:

And, since summing on indices is an invariant operation, we can arrange that our tensors are fully traceless: if we
sum any upper index with any lower index we get 0:

by subtracting out a bunch of terms proportional to .

That’s all we can do, generalizing what we did for SU(2). This set of tensors defines a representation. It may
or may not be irreducible, but there they are, these symmetric, traceless tensors. We’ll assume for now that they
are irreducible. We know how the group acts on all tensors with n upstairs indices, m downstairs indices and no
trace anywhere in between. The technique of decomposing representations into symmetric and antisymmetric
parts gets really messy for SU(4), SU(5), etc., and is not the method of choice. For groups of higher dimension
than SU(3), we have to use the permutation group and Young tableaux,41 which were introduced into this subject
by Hermann Weyl, cursed be his name.42

Thus we have defined representations which we will call

In the next lecture, we will call these representations IR’s for short, an acronym for irreducible representation. Next
time we will deduce their properties, their dimensions, their isospin content, and what happens when we multiply
them together in analogy to the matrix tricks we saw here. And finally we will prove that they are in fact a complete
and inequivalent set of irreducible representations, thus putting our knowledge of SU(3) on the same solid footing
as our knowledge of SU(2). In the lecture after that, I’ll show you four applications.
1 [Eds.] The detailed analysis of Dalitz plots is a highly technical subject. See J. D. Jackson and D. R. Tovey,
Section 46, “Kinematics”, in PDG 2016, http://pdg.lbl.gov/2016/reviews/rpp2016-rev-kinematics.pdf.
2 [Eds.] G. Barton and S. P. Rosen, “Dalitz Plot for the Decay η → π+ π− π0”, Phys. Rev. Lett. 8 (1962) 414–416.
3 [Eds.] Ibid. See also PDG 2106, p. 37.
4[Eds.] H. Primakoff, “Photo-Production of Neutral Mesons in Nuclear Electric Fields and the Mean Life of the
Neutral Meson”, Phys. Rev. 81 (1951) 899; A. Halprin, C. M. Andersen and H. Primakoff, “Photonic Decay Rates
and Nuclear-Coulomb-Field Coherent Production Processes”, Phys. Rev. 152 (1966) 1295–1303.
5[Eds.] “An introduction to unitary symmetry”, Chapter 1 in Coleman Aspects; S. Coleman, “Fun with SU(3)” in
High-Energy Physics and Elementary Particles, ed. C. Fronsdal, IAEA, Vienna, 1965; M. Gell-Mann and Y.
Ne’eman, The Eightfold Way, W. A. Benjamin Publishers, 1964; reprinted by Westview Press, 2000. Harry Lipkin
suggests that SU(3) was called the “Eightfold Way” because it took people eight years (1953–1961) to figure
things out; H. Lipkin, “Quark Models and Quark Phenomenology”, invited talk at the Third Symposium on the
History of Particle Physics, Stanford Linear Accelerator Center, June 24–27, 1992.
6[Eds.] M. Gell-Mann, “The Interpretation of the New Particles as Displaced Charged Multiplets,” Nuovo Cim. 4
Suppl., (1956) 848–866; K. Nishijima, “Charge Independence Theory of V Particles,” Prog. Theor. Phys. 13 (1955)
285–304. See also §24.4.
7 [Eds.] See note 2, p. 507.
8 [Eds.]
H. M. Georgi, Lie Algebras in Particle Physics, Addison-Wesley Publishing Co., 1982, pp. 5–6; R. N. Cahn,
Semi-Simple Lie Algebras and Their Representations, Benjamin/Cummings Publishing Co., 1984; Zee GTN, pp.
122–123.
9 [Eds.]E. P. Wigner, Group Theory and its Application to the Quantum Mechanics of Atomic Spectra, Academic
Press, 1959.
10 [Eds.] J. Schwinger, “A Theory of the Fundamental Interactions”, Ann. Phys. 2 (1957) 407–434.
11 [Eds.] M. Gell-Mann, “Model of the Strong Couplings”, Phys. Rev. 106 (1957) 1296–1300. See equation (10)
and the discussion following equation (12); “global symmetry” was the name given to the hypothesis that the
pion–baryon coupling constant was the same for all the baryons. “Global” symmetry today means something very
different. It is used in the context of gauge theories. If the group parameters depend on xµ, the symmetry is called
local; if not, the symmetry is called global.
12 [Eds.] S. Barshay, “Hyperon–Antihyperon Production in Nucleon–Antinucleon Collisions and the Relative Σ − Λ
Parity”, Phys. Rev. 113 (1959) 349–351.
13 [Eds.] S. L. Glashow and M. Gell-Mann, “Gauge Theories of Vector Particles”, Ann. Phys. 15 (1961) 437–460.
14 [Eds.] André Weil (1906–1998), influential French mathematician and brother of the philosopher and mystic
Simone Weil (1909–1943). He was one of the founders of the team writing mathematics under the group
pseudonym “Nicolas Bourbaki”.
15 [Eds.] Crease & Mann SC, pp. 266–268.
16 [Eds.] Séminaire “Sophus Lie”, École Normale Supérieure, Paris. Volume 1: 1954–1955; Volume 2: 1955–1956.
17 [Eds.] Sophus Lie, Norwegian mathematician 1842–1899. For background on Lie’s life and work, see D. J.
Struik, A Concise History of Mathematics, G. Bell and Sons, London, 1954; reprinted by Dover Publications, New
York, 1987; B. Fritzsche, “Sophus Lie: A Sketch of his Life and Work”, J.Lie Theory, 9 (1) (1999) 1–38. Many
physicists of the time learned Lie group theory from a little book by Harry J. Lipkin, Lie Groups for Pedestrians,
North-Holland, 1965, reprinted by Dover Publications, 2002.
18 [Eds.]
Yuval Ne’eman (1925–2006), Israeli physicist, soldier and politician. Ne’eman and Gell-Mann shared the
1969 Nobel Prize in Physics for their work on SU(3).
19 [Eds.] Crease & Mann SC, pp. 269–272.
20 [Eds.] Moshe Dayan (1915–1981), Israeli military leader and politician.
21 [Eds.]“Salam laughed at the recommendation from Dayan, and told Ne’eman to bring a recommendation from
a physicist. Ne’eman never did, but Salam accepted him anyway—partly, he has said, to repay a debt incurred by
Islamic science, which in its medieval heyday owed much to Jewish scholars.” Crease & Mann SC, p. 270.
22[Eds.] Y. Ne’eman, “Derivation of Strong Interactions from a Gauge Invariance”, Nucl. Phys. 26 (1961)
222–229.
23 [Eds.] This theorem is due to Coleman, proved by him in his PhD thesis: S. Coleman, The Structure of Strong
Interaction Symmetries, Caltech, 1962, http://thesis.library.caltech.edu/2386/1/Coleman_sr_1962.pdf.
24[Eds.] T. D. Lee and C. N. Yang, “Some Considerations on Global Symmetry”, Phys. Rev. 122 (1961)
1954–1961. Lee and Yang cite A. Pais, “Note on Relations between Baryon–Meson Coupling Constants”, Phys.
Rev. 110 (1958) 1480–1481, which suggested that the Gell-Mann–Schwinger group corresponded to the direct
product of three unitary unimodular groups; see footnote 2 in Pais. In fact the Lee–Yang group included also a
discrete symmetry R (their equation (10)) to make the representations irreducible.
25 [Eds.] “An introduction to unitary symmetry”, Ch. 1 in Coleman Aspects, pp. 3–5.
26 [Eds.] §18.5.
27 [Eds.] Coleman op. cit., p. 4.
28 [Eds.] The Δ has a width ≈ 117 MeV. PDG 2016, p. 91.
29[Eds.] Now the Particle Data Group tables. Named for Arthur H. Rosenfeld (1926–2017), American physicist.
The tables started as unpublished data tables to support a long review article with Gell-Mann: M. Gell-Mann and
A. H. Rosenfeld, “Hyperons and Heavy Mesons (Systematics and Decay)”, Ann. Rev. Nucl. Sci., 7 (1957)
407–478.
30 [Eds.] In the videotape of Lecture 36, Coleman says “half” instead of “a third”. The current values are µΛ =
(−0.613 ± 0.004)µN, µn = (−1.913)µN; PDG 2016, p. 92.
31 [Eds.] Another candidate for G0 was the exceptional group G2: R. E. Behrends, J. Dreitlein, C. Fronsdal, and W.
Lee, “Simple Groups and Strong Interaction Symmetries”, Rev. Mod. Phys. 34 (1962) 1–38, and references
therein. (J. L. Rosner, private communication.) Incidentally, the author “W. Lee" seems to be Benjamin W. Lee.
32 [Eds.] See note 23, p. 784.
33[Eds.] H. Weyl, The Classical Groups, Princeton U. P., 1953. See also S. Coleman, “Fun with SU(3)”, op. cit.,
and J. Mathews and R. Walker, Mathematical Methods of Physics, 2nd ed., Addison-Wesley, 1969, Chapter 16,
pp. 424–470. In the preface the authors state, “Much of Chapter 16 grew out of fruitful conversations with Dr.
Sidney Coleman.”
34[Eds.] Chapter III, §7, pp. 136–140; Chapter V, §12–14, pp. 347–369 in H. Weyl, The Theory of Groups and
Quantum Mechanics, trans. H. P. Robertson (of the Robertson-Walker metric in general relativity), E. P. Dutton,
New York, 1931; reprinted by Dover Publications, 1950. Originally published in German as Gruppentheorie und
Quantenmechanik, Verlag von S. Hirzel, Leipzig, 1928.
35 [Eds.] In the videos, and in Aspects, Chapter 1, “An introduction to unitary symmetry”, Coleman uses lower
indices.
36 [Eds.]This is an illustration of the “sticks and stones” or the “balls and urns” or the “stars and bars” method. W.
Feller, An Introduction to Probability Theory and Its Applications, 3rd ed., v. 1, John Wiley and Sons, 1950, Chap.
II, Section 5, “Application to Occupancy Problems”, pp. 38–40.
37 [Eds.] Perhaps a reminder of the relationship between SU(2) and SO(3) would be helpful. Let X = x•σ = xiσi
(sum on i), and let U(R) = exp{−i(θ/2) •σ} where is a unit vector; U(R) ∈ SU(2) because •σ is traceless and
Hermitian. Then under the transformation X → X′ = UXU†, it is easy to show X′ = x′•σ with x′ given by Rodrigues’
formula (see note 9, p. 374)

using the identities σiσj = δij + iϵijkσk , and U(R) = cos(θ/2) − i sin(θ/2) •σ (here, is a 2 × 2 identity matrix).
That is, there is a double-valued (±U) homomorphism between SU(2) and SO(3), and a rotation of a 3-vector x
about a unit axis through θ is a rotation through (θ/2) of the associated operator X in spinor space. In more
careful language, SU(2) is the covering group of SO(3): the two groups share the same Lie algebra [Li, Lj] = iϵijkLk ,
and SU(2) is locally isomorphic to SO(3).
38 [Eds.]
Greiner & Müller, Quantum Mechanics—Symmetries, Chapter 5, “The Isospin Group (Isobaric Spin)”, pp.
95–98. See also note 37, p. 791; in particular, U(R z (θ)) = cos(θ/2) − i sin(θ/2)σz .
39 [Eds.] It’s traditional to describe isospin in terms of the matrices τi, even though in the case of I = , these are
exactly the Pauli matrices σi. Here the τi imbed the pion isovector into the isospinor space of the nucleons.
40[Eds.] This notation is not so common today, though it sometimes appears in supersymmetric theories; it was
more frequently used forty years ago. See B. L. van der Waerden, “Spinoranalyse” (“Spinor Analysis”),
Nachrichten-Akad. der Wiss. Göttingen, Math.-Phys. Kl. (1929) 100–109; B. L. van der Waerden, Group Theory
and Quantum Mechanics, Springer, 1974, §23, “The Representations of the Lorentz-Group”, pp. 114–117; Ryder
QFT, pp. 433–439.
41 [Eds.] Arfken & Weber MMP, Section 4.4, pp. 274–276.
42 [Eds.] Weyl, op. cit., §13, pp. 358–362.
 37
Irreducible multiplets in SU(3)

Last time, we made a guess about the irreducible representations of SU(3).1 In SU(3), as in SU(2), we used
invariant tensors (the Kronecker delta and the antisymmetric epsilons, ϵijk and ϵijk) to reduce the rank of
tensors. With the former, we could reduce a tensor’s rank by two, by summing over an upper and a lower index to
form the trace; and with the latter, we could reduce its rank by one, trading two upper indices for a lower, or vice
versa, by summing over two upper or two lower indices. The guess was that the irreducible representations should
have as their basis the set of all tensors with n upper indices and m lower indices

which are totally symmetric in both the upper and lower indices, and traceless in every pair of an upper and a
lower index. Otherwise, contraction with either an epsilon or a delta could lower the tensor’s rank. We call these
representations

where g ∈ SU(3).

I’ve certainly defined a representation: a tensor of this kind does go into a linear combination of other tensors
of this kind under the action of the group. In the course of this lecture we will answer the following questions:

•Are these representations irreducible?

•Are they inequivalent for different n and m?

•Are these all of the irreducible representations?

I’m going to answer these in counter-mathematical order, but in perfect rigor, insofar as I can achieve it. First I will
deduce all sorts of useful properties of these representations. Then I will apply these properties to prove that these
representations are in fact inequivalent, irreducible and complete. We will call these representations “IR”’s (for
irreducible representations).

37.1The irreducible representations q and

The properties of the IR’s that I want to examine are the usual things to consider when one encounters a new
group and its representations, the same properties we investigated for the Lorentz group in §18.3, or for the
rotation group in quantum mechanics.

Conjugate representations

From (36.44), (36.45) and (36.46), it’s easy to see that

The conjugates of upper indices transform like lower indices, and vice versa; that’s the way we’ve defined things.
So the representations and their complex conjugates (with m ↔ n) are equivalent. The matrices may not be the
same but they can be turned into each other by a change of basis.

Dimensions

How many independent components do these tensors (37.1) have? That requires a little more work. The key
formula is

so that
If you know the meaning of the symbols, it’s easy to see why (37.4) is true. Recall that (n, 0) describes a
completely symmetric tensor with only upper indices, and (0, m) a completely symmetric tensor with only lower
indices. We cannot impose the traceless condition on either of these varieties of tensors; they have only one type
of index. When we take the product of one of each sort, we obtain completely symmetric tensors with a bunch of
upper and lower indices, but there’s no guarantee of tracelessness. Equation (37.4) is simply the mathematical
statement that a general tensor, fully symmetric in both upper and lower indices, can be written as the sum of a
fully symmetric, traceless tensor plus a general, fully symmetric tensor with one less upper index and one less
lower index (times a string of invariant Kronecker delta’s). For example,

If we take the trace of both sides (by setting j = k and summing), we find

(the same is true if we set i = k). We can determine the dimensions of (n, m) if we can find the dimensions of (n, 0)
and (0, m).

Let’s begin by computing the dimension of (n, 0). This is just combinatorics. As (n, 0) has only one kind of
index and it’s completely symmetric, the problem may be restated: given n objects, how many ways can we put
them into three boxes labeled 1, 2 and 3? That will tell us how many 1’s, how many 2’s, and how many 3’s there
are. Put the n objects in a line and draw two barriers, creating three boxes. I’ve chosen n = 6 for convenience. I’ll
put them in boxes by drawing two lines somewhere between them (Figure 37.1). Everything to the left of the
leftmost line will be 1’s. Everything to the right of the rightmost line will 3’s. The 2’s will be those between the lines.
There are (n + 1) places for the first barrier, just as for SU(2). There are (n + 2) places for the second barrier,
because we have the choice of putting the second barrier in front of or behind the first barrier (but which barrier we
call the first and which the second is irrelevant). Therefore2 the number of completely symmetric tensors with n
upper indices or n lower indices (the combinatorics are indifferent to type)—the number of ways of putting n
objects into three boxes—is

Figure 37.1: One arrangement of six objects in three boxes

The comes because the two barriers are indistinguishable.

Figure 37.2: One object in three boxes

As a check, how many ways can we put one object into three boxes? Three, of course; and that’s what this
formula gives (see Figure 37.2):

For another example, consider two objects into three boxes, as in Figure 37.3. That also works out:

Figure 37.3: Two objects in three boxes

Solving (37.5) for dim(n, m) we find

This is a more complicated formula than 2s + 1, the corresponding formula for SU(2), because SU(3) is a more
complicated group. But it is still relatively straightforward.

For example, let’s work out the dimensions of some low-lying representations. See Table 37.1:

Table 37.1: SU(3) representations and their dimensions

Aha! The 8’s include the baryons; the 10 will describe the representation in which appear the Δ and its friends,
among them the Ω−, famed in song and story. There is no particle as far as I know that has been assigned to a 27-
plet, but the 27 will come into our theory for certain operators.

The convention for the vulgar 3 name is to label the IR by its dimensions (I’ll use bold type to distinguish
representations from ordinary numbers), adding a bar if the second index is greater than the first, to distinguish
between complex conjugate pairs of representations. This labeling is unfortunately not unique. For instance, as
one can check,

but neither of these representations occurs frequently in the literature, and most people call a representation 3 or 3
etc., rather than (1, 0) or (0, 1).

Isospin and hypercharge

How should we embed the SU(2) ⊗ U(1) subgroup inside SU(3)? We can decide that once we’ve made a
decision about the isospin and hypercharge of the three-dimensional representations, 3 and 3. That will be
determined by the interplay of mathematics and physics. The mathematics will tell us the possible ways to do it;
the physics will tell us the right way to do it so that the baryon octet comes out as it should.

For the moment I will worry about embedding SU(2), the isospin subgroup inside SU(3), and take care of
hypercharge shortly. Let’s begin with the fundamental triplet representation, (1, 0) = 3. That’s a three-dimensional
representation and therefore when we restrict SU(3) to isospin there are three possibilities:

Those are the only distinct ways of partitioning 3 into a sum of positive integers: 3, 2 + 1, 1 + 1 + 1.

Possibility (c) is no good for both mathematical and physical reasons. If everything in the triplet transforms
like an isosinglet, everything in the triplet bar also transforms like an isosinglet and everything in all the IR’s
transforms like an isosinglet. That’s a trivial embedding of SU(2) inside SU(3). We want a non-trivial embedding.

Possibility (a) is also no good. It says that the triplet 3 has isospin 1, the conjugate 3 has isospin 1, and
therefore everything made by taking direct products would contain only integer isospins. This leaves no room for
such friendly particles as the nucleons, with half-integer isospin. So possibility (a) is also no good. Mathematically
it’s fine, but physically it fails to accomodate the baryon octet.

The only remaining possibility is (b), an isodoublet plus an isosinglet. We will represent this graphically by
writing the fundamental three-dimensional representation as a column vector. I have the freedom to make
similarity transformations, so I will arrange things such that the first two entries are the isodoublet and the third
entry is the isosinglet. Thus, SU(2) would consist of that subgroup of SU(3) which leaves the third unit vector
unchanged. These are conventionally labeled

The u, conventionally called “up”, and d, called “down”, form the isodoublet with its isospin up and down; s is the
isosinglet; s originally stood for “singlet”, or according to some wags, “sideways”. For historical reasons it is called
“strange”.4 We can consider this triplet as hypothetical states of some unknown particles called quarks (we love to
give names to the unknowns).5 All other hadrons can be built out of 3’s and 3’s, i.e., out of quarks and antiquarks.
Alternatively, we can consider the 3 representation as triplet fields for some particles that have never been
observed.6 In any case, we’ll denote the 3 representation by q (in honor of the quark model, though I won’t discuss
that theory until later), and 3 by q.

How are we going to assign hypercharge to the q representation? Hypercharge commutes with isotopic spin
and therefore the two elements of the doublet must have the same hypercharge, which I will call α. The matrix Y
whose eigenvalues are the hypercharges is also the generator of the U(1) symmetry. That is, the matrices g ∈
U(1) are exponentials of Y :

where χ is some real parameter; if g is to be unitary, Y must be Hermitian. We want these matrices g to be
elements of SU(3) as well, and so we need the exponential of the hypercharge matrix to have determinant 1.
Using the formula7

it follows that

which means the hypercharge matrix Y itself must have trace 0. The trace of the matrix is the sum of its
eigenvalues. As the doublet elements u and d each have hypercharge α, the singlet must have hypercharge −2α.
We will fix α by physical considerations. Once we have assigned isospin and hypercharge values to the
representation (1, 0) = q, then we also know the isospin and hypercharge content of the representation
. These values are summarized in Table 37.2.

Table 37.2: Isospin and hypercharge for the q and representations

Our table of the dimensions of the representations led us to suspect that the baryons must be put into the 8 =
(1, 1) representation, the product of a row vector and a column vector with the trace subtracted out. We will label
representations of the subgroup by their isospin and hypercharge like this (with the hypercharge as a superscript):

In this notation, we write (37.9)

From (37.4)
because (0, 0) is a singlet. In the vulgar notation,

which makes sense: nine states on either side. On the other hand, we have

(using the Clebsch–Gordan series (18.73 for SU(2)) so that, from (37.16)

The trivial representation of the group, (0) 0, is an isosinglet with hypercharge 0. It is the same as the
representation (0, 0) and can be dropped on both sides. That is,

We observe that this octet matches up nicely with the eight JP = + baryons8 of Table 36.1: (0) 0 is in a jolly
position to be the Λ, the (1) 0 is well-suited to be the Σ, while the ( )3α and ( )−3α can be identified with the nucleon
and the cascade (Ξ), respectively, if we choose

With this identification,

There is of course the option to choose α = − and change the identification of the nucleon and the cascade, but
this is just a matter of convention. If we choose α = − we would be switching the left and right entries in the
representation:

We would have the same hypercharge assignments for the s that we originally had for the quarks. All that would
change would be the convention about what we call a quark and an antiquark. Our equations would look different
but the physics would be the same. For the benefit of mathematical purists, this corresponds to the existence of an
outer automorphism of the group SU(3) induced by complex conjugation.9 However, if the physics is clear then
which sign we choose is just a matter of convention.

We have finally for the quarks:

Table 37.3: The quarks and their properties

The quarks have fractional charges.10

37.2Matrix tricks with SU(3)

Baryons

Following the SU(2) tricks of the last lecture, we will work out a matrix representation for the octet
representation (1, 1) which we had identified with the JP = + baryons. This is a traceless tensor with one upper
index and one lower index:

Recall that under a transformation g ∈ SU(3)

We can consider as a matrix which transforms as if it were the outer product of a q and a , minus the trace
times the identity matrix:

This is similar to what we did earlier, when we considered (36.79) the pions transforming as an outer product of an
N and an N. That’s the way that outer products of two vectors typically transform, like a rank 2 tensor. Remember
that our q consists of a 2 × 1 block, an isodoublet with Y = + , and a 1 × 1 isosinglet block with Y = − ; likewise
has a 1 × 2 isodoublet block with Y = − and a 1 × 1 isosinglet block with Y = :

The outer product q ⊗ gives

The diagonal elements are connected by the traceless condition:

When we restrict ourselves to the SU(2) subgroup, those transformations leave the third component of the vector
unchanged. We also see that the entries have the given hypercharge assignments, (37.20). Thus we can write
down the matrix

The Σ terms in the upper 2 × 2 block follow the pattern of (36.85). The scale factor β is to be determined; the −2β
in the lower right is required by the trace condition. The third row has −Ξ0, not +Ξ0, in order to have the
conventional (Condon and Shortley)11 phase relations between the members of an isodoublet: we use ϵij to lay
the 2 × 1 block (upper right) on its side.

The expression (37.28) for ψ is the representation of the eight baryons as a 3 × 3 matrix. Choosing a value for
β is a matter of taste, but we normally arrange things to avoid SU(3)-violating wave function renormalizations.
Taking a hint from the trick (36.86) we used with SU(2), we want to look at Tr , which appears in the free
Lagrangian’s mass term; is the (Dirac) adjoint matrix of ψ, transforming under SU(3) as

Then Tr is invariant:

The fields in ψ are Dirac fields. When we form , the fields appear in it as their Dirac adjoints, e.g., instead of p.
The trace is a sum of squares:

from which we determine

We could choose β = − but that’s just a matter of convention on the phase of the Λ, which thus far is
undetermined. The final result for ψ is

The free Lagrangian involving this mass degenerate octet12 can now be written

Mesons

As it happens, there are also eight pseudoscalar mesons, JP = 0−, of low mass; see Table 37.4.

In fact there’s a ninth pseudoscalar meson, the η′, with (I)Y = (0) − and mass 960 MeV, that most people think is a
singlet. It doesn’t mix up much with these eight even with SU(3)-violating interactions. It is possible that medium
strong interactions mix members of different multiplets. Note that the absolute value of the pseudoscalar meson
mass splitting (410 MeV between the π’s and the η) is comparable to that of the baryon octet (380 MeV between
the nucleons and the Ξ’s; see Table 36.1 on p.782). That suggests that medium strong splittings may all be of the
same order, but at different mass levels (ignoring the η′).

Table 37.4: The eight pseudoscalar mesons: JP = 0−

We can represent the eight pseudoscalar mesons by a matrix ϕ of exactly the same form as that for the
baryon matrix ψ. Since the (1, 1) representation is self-conjugate, (1, 1)* ~ (1, 1), so is the meson octet
representation. We can, if we wish, impose the condition

for the 0− mesons. The choice that ϕ be Hermitian is invariant under SU(3) transformations. That will give us eight
real fields instead of eight complex fields. We don’t need to do that; we certainly don’t want to do it for the baryons
unless we want to write things in terms of Majorana fields13, which is a bad move: the baryons are not their own
antiparticles. But we can do it for the mesons and we choose to do so, because there are only eight
pseudoscalars, not 16. The ϕ matrix looks, with one small difference, exactly the same as ψ:

There is no minus sign in front of the K0 (as there is before the Ξ0 in the baryon matrix) because of a clash of
conventions. We have two conventions we want to follow. One is that the phase relations between an isodoublet
should be as found in Condon and Shortley14; the other is that K0 should be the conjugate field to K0. For the K0
we adopt the second convention, and disobey the Condon–Shortley phase convention. When the first papers
were written, this was the choice everyone made, and it’s now standard.

We can now write down an SU(3)-invariant (indeed, it’s SO(8)-invariant) meson Lagrangian. Adding it to
(37.33),

That’s just the sum of the squares of the fields, all normalized the same way. We can also make a guess about
invariant Yukawa interactions. Both ψ and ϕ transform according to (37.23), so the trace of any product of ψ’s and
ϕ’s will be SU(3)-invariant. We have one ψ, one and one ϕ. By the cyclic invariance of the trace we can always
put the in the first position, but then there are two possibilities, whether the ϕ precedes the ψ or follows the
ψ. In fact in the literature the symmetric and antisymmetric combinations of these two possibilities are used. We
write15

Of course nothing we have said so far demonstrates that these are the only couplings, but we will shortly find that
they are. For the moment they are two Lorentz-invariant, SU(3)-invariant couplings which one can write down.

37.3Isospin and hypercharge decomposition

Earlier, we worked out the isospin and hypercharge values of the octet (1, 1). To understand the isospin and
hypercharge decomposition of a general representation (n, m), we need a digression on some other matrix tricks.
Return to (37.4):

If we can find out the isospin and hypercharge decompositions of (n, 0), we can get those of (0, m) by complex
conjugation. Then we’ll know what isospins and hypercharges are in (n, 0) ⊗ (0, m) and in (n − 1, 0) ⊗ (0, m − 1),
and by subtraction we can find the isospins and hypercharges in (n, m). So our immediate goal is to work out the
isospin and hypercharge decomposition of (n, 0).

This is fairly easy: (n, 0) is the completely symmetric product of n (1, 0) (or q, or 3) representations. The
quarks contain a doublet with hypercharge and a singlet with hypercharge − , Table 37.3. When we take the
direct product, we obtain only a certain number of representations, according to the Clebsch–Gordan series for
SU(2):

This will be a sum of terms. For the first term we can take the completely symmetric product of n isospin ’s, with
all the isospins aligned and their hypercharges added, so that gives us

The next term we can have has (n − 1) isospin ’s and one isospin 0:

The hypercharge is down by 1 since we’ve traded a doublet with Y = + for a singlet with Y = − . Continuing along,
at the end we have simply the n isospin 0 piece with hypercharge − n:

Thus the isospin-hypercharge content we get when reducing SU(3) down to SU(2) ⊗ U(1) is

For instance,

If we want to find out what’s in (n, 0) ⊗ (0, m) we make a large rectangular array, n by m blocks, as in Figure 37.4.
In each block of the table we multiply together the two representations, using the conventional isospin combining
rules. Then we fill up the entire array, add all the blocks together, and that’s the product (n, 0) ⊗ (0, m). However,
to get the representation (n, m), we have to subtract out what’s in (n − 1, 0) ⊗ (0, m − 1). That’s in fact exactly the
same series except that it begins one stage further on. The hypercharge is a little bit off in (n − 1, 0) but is a little bit
off in the opposite direction in (0, m − 1) so it doesn’t matter in the product. The content of (n − 1, 0) ⊗ (0, m − 1) is
the entire array aside from the border—the shaded area. Therefore the content of (n, m) is simply the entries on
the top and left border of the rectangular array. That is our algorithm.

Figure 37.4: Graphical decomposition of (n, m)

Let’s check this by working out the answer to something we already know, (1, 1). This is a 2 × 2 table, shown
in Figure 37.5.

This is just the Σ and the Λ, the nucleon N, and the Ξ, the same answer (37.20) we obtained before.

Figure 37.5: Graphical decomposition of (1, 1)

Figure 37.6: Graphical decomposition of (2, 2)

Now let’s go after bigger game. I will work out (2, 2), the 27-plet. That’s a little bit more ambitious, a 3 × 3 box,
Figure 37.6. Check the dimensions, summing along the row and the column:

Aha! It’s exactly what we should have found.

The only way to learn these algorithms is to pick a representation and work things out for it. I’ve not assigned
a homework problem on this topic, but you should work out the isospin content of (3, 3) or (3, 4).

In the literature the isospin-hypercharge contents of these representations are frequently depicted on weight
diagrams. A dot is placed on the weight diagram for every particle with a given Iz and Y. The weight diagrams for
the representations 3, 3 and 8 are shown in Figure 37.7. A dot in a circle indicates two particles with the same
values of Iz and Y ; a dot in two circles indicates three. The term “weight diagram” comes from Cartan’s general
theory of the representations of semi-simple Lie groups.16

There is a deep reason why these things come out to be beautiful geometric figures rather than random arrays of
dots, but that has to do with the structure theory of Lie groups and we won’t go into it here. As another drill, work
out the weight diagrams for some other representations, such as (3, 0). For those of you of Pythagorean
inclinations, I leave it as an exercise to demonstrate that (n, n) always makes a hexagon and (n, 0) or (0, n) always
makes a triangle. As far as I’m concerned, these pretty diagrams are not useful.17

Figure 37.7: The weight diagrams for the representations 3, 3 and 8

Figure 37.8: The weight diagram for the representation 27

37.4Direct products in SU(3)

How do we decompose the direct product (n, m) ⊗ (n′, m′)? This is itself certainly a representation. It should be
expressible as a direct sum of IR’s. Which ones? What are the SU(3) analogs of the familiar vector addition
algorithm and Clebsch–Gordan series of SU(2)? To write it as an equation,

We know the answer for a product of (1, 0)’s and (0, 1)’s, but not for the general expression.

I will show you a non-standard algorithm for computing these direct products. It’s one I invented18 around
1964. It seems wonderfully simple and elegant to me; everyone else just looks things up in tables.19 These tables
are typically produced using a much more complicated algorithm (invented by Hermann Weyl) that is good for a
general Lie group. The algorithm I’m going to show you is a special trick that works for SU(3) only. It will enable us
to compute the direct product of anything with anything in SU(3), using only the back of an envelope and
elementary arithmetic.

The algorithm has two stages. First, we reduce the product (n, m) ⊗ (n′, m′) to the sum of certain special
reducible representations (which I will define) by removing all remaining traces. Second, we reduce these special
reducible representations to a sum of IR’s by getting rid of all antisymmetric parts.

The first stage in the algorithm involves turning a direct product (n, m) ⊗ (n′, m′) into a sum of tensors of the
form

This tensor is completely symmetric in four sets of indices, but not otherwise: in the first upper n indices, in the last
upper n′ indices, in the first lower m indices, and in the last lower m′ indices. It is completely traceless. Roughly
speaking it is what we would get if we took the direct product (n, m) ⊗ (n′, m′) with all traces removed, but without
any symmetrization among either the n and n′ indices or the m and m′ indices. We go from the direct product to
the form (37.43) by removing traces.

We start out with two traceless tensors, (n, m) and (n′, m′). In the direct product (n, m) ⊗ (n′, m′) we needn’t
bother with traces formed from the “first” indices, the n’s with the n′’s, or the “last” indices, the m’s with the m′’s;
that’s already done. We need separate out only those tensors that can be obtained by contracting, in all possible
ways, indices from the “outside” sets, the n’s with the m′’s, and from the “inside” sets, the n′’s with the m’s. So we
get a double direct sum

The process terminates whenever we run out of indices to contract; i.e., whenever a zero appears in the series on
the right. In a more compact form we have

We can peel indices off of either the two “outside” indices n, m′ or the two “inside” indices n′, m, but we have to
strip off the same number in either case, because we’re contracting indices. That takes care of all the traces.

In the second stage of the algorithm, we remove the antisymmetric parts of the terms (n, n′; m, m′) in (37.44).
How do we do that? We can express any tensor as the sum of a symmetric and an antisymmetric tensor, the latter
in terms of the three index ϵ’s. Consider the pair of indices i 1 and i n+1 in the tensor

We can write the antisymmetric part in these indices as a tensor s of lower rank:

Which il’s we pick doesn’t matter because those pairs are completely symmetric. We’ve picked two upper indices
just for simplicity; we could just as well have picked two lower indices.

Now an amazingly helpful fact keeps us from getting involved in a lengthy calculation: the tensor s in
(37.46) is already symmetric in all of its lower indices k, j 1, j m+m′. This is not obvious, but I will prove it. That
means in our systematic splitting into symmetric and antisymmetric parts, we can only turn two upper indices into
a lower index, or four upper indices into two lower indices, etc. And once we start turning pairs of upper indices
into lower indices, we can’t turn around and start contracting two lower indices into an upper index, because any
such contraction will vanish.

The proof is in fact quite straightforward. We prove it by contracting s with an ϵ tensor in any chosen pair of
lower indices, say j1 and jm+1, and showing that the result is zero. The product of two ϵ tensors can be written in
terms of the products of three Kronecker deltas:

There are six terms altogether, positive or negative depending on the permutation of the indices. Then

Notice that the first term involves a Kronecker delta on i n+1 and j m+1. But we started with a tensor that’s
completely traceless, and so

Indeed, every term in the permutation (37.47) involves some Kronecker delta that sums over two of our original
indices, because there are only two free indices, k and r; k could be on one of the Kronecker δ’s, r could be on
another, but the third has to involve a pair of our original indices. So every one of the six terms will take a trace of
an object that is traceless, and thus vanish. Therefore the s tensor is already symmetric in the lower indices once
we begin contraction of the uppers with ϵ’s. QED
 So, here is the second stage of our algorithm (amusin’ and confusin’, but all the same very practical),

That’s the end of the story. Start with a tensor and choose which indices, upper or lower, to contract. Once you’ve
finished contracting ϵ’s with pairs of those indices, you’re done with that tensor, because it will already be
symmetric in its other indices.

The algorithm as stated here is incredibly simple. We just multiply two representations (n, m) and (n′, m′)
together. Then we do these operations as many times as we can. When any term gets to 0 we stop:

1.Take indices from the inside, m with n′, and from the outside, n with m′;

2.Take two indices off the left (the uppers) and make one on the right (the lowers),
or vice versa.

Let me give some examples of this algorithm in action. First let’s look at one we’ve already done, to check that it
works.

EXAMPLE 1: (1, 0) ⊗ (0, 1) (or 3 ⊗ 3)

We already know (37.4) that (1, 0) ⊗ (0, 1) ~ (1, 1) ⊕ (0, 0). What does the algorithm say?

Stage 1: Following (37.45), we have

Stage 2: According to (37.50),

The ϵ contractions require two upper or lower indices, and these terms have only one: the sums contribute
nothing. Likewise

as required. In terms of dimensions the result is

as we already knew. Of course the algorithm is trivial in this case; nevertheless we got the right answer. Armed
with this confidence, let’s go on to a more interesting case.

EXAMPLE 2: Scattering amplitudes for the octet, 8 ⊗ 8 (or (1, 1) ⊗ (1, 1))

We know the critical importance of isospin invariance in pion-nucleon scattering. Isospin tells us that our
amplitudes are either I = or I = and that connects a lot of data.20 SU(3) is a much more powerful group than
SU(2). We’ve got an octet of mesons and an octet of baryons. What is the corresponding statement for
meson–baryon scattering in SU(3)? Do we have two amplitudes, as with isospin? 17? 121? And what are their
transformation properties?

Well, we know how to compute the answer: apply the algorithm. The first stage is:
This is (37.45): r runs from 0 to 1, s runs from 0 to 1, giving four terms. For the second stage:

Therefore

To check that we haven’t double counted or under counted in developing our algorithms, let’s write this out in the
vulgar notation, where we label the representations by their dimensions and see if the dimensions come out right.
Recall dim(n, m) = (n + 1)(m + 1)(n + m + 2), so dim(2, 2) = 27, dim(3, 0) = dim(0, 3) = 10, and

Unlike in SU(2), the same representation can occur twice in a direct product. In the SU(2)-invariant theory of pion-
nucleon scattering we have two amplitudes, but in the SU(3) theory of meson–baryon scattering—and that
includes anything: K’s off of Ξ−’s, not that anyone has done that experiment, η’s off of Λ’s, you name it—there are
more.

There is a tricky point in finding how many amplitudes there are. Because if the initial meson–baryon state is
in the 8 ⊗ 8 representation, then the final meson–baryon state is represented by the same formula (37.58):

The arrows mean non-zero S-matrix elements. So a 27 can scatter only into a 27, a 10 can scatter only into a 10,
etc. There are six vertical arrows, six possible amplitudes. There are two ways for each of the 8’s to scatter. The
two crossed arrows for 8 → 8 are trivially related to each other by time-reversal. So, assuming time-reversal
invariance, the two crossed arrows count as only one, giving a total of seven independent amplitudes for
meson–baryon scattering. When we can build two octet states for the final states of the baryon and the meson, we
could have in principle (not counting time reversal) four amplitudes connecting octet with octet (two vertical
arrows, two crossed). Just as in the theory of scattering of particles with spin by a spin-dependent force, if we can
build several J = states, say ℓ = 1 and ℓ = 2, we can have ℓ = 1 and ℓ = 2 crossed matrix elements.21

37.5Symmetry and antisymmetry in the Clebsch–Gordan coefficients

In our discussion of both isospin and the rotation group (group theoretically the same thing), and of the Lorentz
group, we had to deal with the question of the symmetry and antisymmetry of elements of the direct product under
exchange of the two objects, if the two representations are identical. Put another way, we had to worry about the
symmetry and antisymmetry of the Clebsch–Gordan coefficients.

For example, if we are putting together two pions in an s-wave state, in general the two pions can have I = 0,
1, or 2. I = 1 is excluded because that is antisymmetric in isospin, and it won’t match with the s-wave state, which
is symmetric in space and (trivially) in spin (the complete state would be antisymmetric—illegal for bosons). Given
two identical representations, which terms in the series generated by this algorithm are symmetric, and which are
antisymmetric? We can think we’re putting together two particles in an s-wave or a p-wave or a J = or a triplet 1
state or whatever. Or, we’re multiplying two fields at the same spacetime point, and we’re only going to get either
the symmetric or the antisymmetric combinations, depending on whether the fields satisfy Bose or Fermi statistics,
respectively.

The direct product (n, m) ⊗ (n, m) generates a set of terms. The whole algorithm is set up in such a way that
the symmetry is manifest. The algorithm begins with stage 1 (37.45):
If we exchange the two n’s and the two m’s on the left-hand side, that exchanges r and s in the sum:

That is, if r = s the terms are symmetric. If r ≠ s the two terms change places, and by forming sums and differences
we can form both a symmetric and an antisymmetric combination:

In stage 2, every time we peel off a pair of indices from one side to put one on the other, we use an ϵ
tensor—an antisymmetric object. Therefore whenever we use an ϵ tensor we change symmetric to antisymmetric
and vice versa; if we use it twice we restore the status quo, and so on. So, in stage 2, successive terms change
symmetry. Of course we don’t have to worry about these terms in stage 1. If we’re just counting the number of
symmetric and antisymmetric objects, the fact that the signs keep flipping (so we have to take the sums once and
the differences the next time to get a symmetric object) is irrelevant; if r ≠ s, the numbers of symmetric and
antisymmetric objects stay the same. It’s only the terms where r = s that we have to keep track of the sign
changes. Thus we see in (37.55) that under exchange of two objects, (1, 1; 1, 1) and (0, 0; 0, 0) are symmetric,
and that (0, 1; 1, 0) and (1, 0; 0, 1) exchange places.

EXAMPLE 3: Coupling two identical octets

To see what sort of representations we’d get in this case, apply the algorithm to (37.57):

Thus, in coupling two 8’s to make a sequence of representations of SU(3), the 27, one of the octets and the singlet
are symmetric; the 10, the 10 and the other octet are antisymmetric.

If we were considering meson-meson scattering in the s-wave, then Bose statistics would tell us that the
overall wave function has to be symmetric. In the s-wave, the space part is already symmetric. The only states
that the two mesons can occupy are the 27, the symmetric octet (not two octets as in the meson–baryon case)
and the singlet. Therefore, for meson-meson scattering in the s-wave, although there are eight kinds of initial
mesons and eight kinds of final mesons, there are only three SU(3)-invariant amplitudes:

Next time, we will take care of the last part of the program announced at the beginning of this lecture and
prove the irreducibility, inequivalency and completeness of the IR’s. Instead of losing you all at the end, I will lose
you all at the beginning. (Those of you who are not group theory mavens22 may come to the lecture fifteen
minutes late; you will miss nothing.) The remainder of that lecture will be the beginning of applications: the Gell-
Mann–Okubo mass formula, electromagnetic moments, electromagnetic mass differences, and some other
related things.

1 [Eds.] This chapter is largely a reworking of two prior talks: Trieste, 1965 (S. Coleman, “Fun with SU(3)” in High-
Energy Physics and Elementary Particles, C. Fronsdal, ed., IAEA, Vienna, 1965) and Erice, 1966 (reprinted in
Coleman, Aspects, Chapter 1, “An introduction to unitary symmetry”).
2 [Eds.] See note 36, p. 791.
3[Eds.] Coleman adds: “In the sense of botany, with no pejorative connotation.” See D. Gledhill, The Names of
Plants, 4th ed., Cambridge U. P., 2008.
4 [Eds.] Crease & Mann SC pp. 171-177.
5 [Eds.] F. Halzen and A. D. Martin, Quarks and Leptons: An Introductory Course in Modern Particle Physics, John
Wiley and Sons, 1984; Crease & Mann SC, Chapter 15, pp. 280-285. This paragraph is taken nearly verbatim
from the 1976 video of Coleman’s lecture 37 (starting at 0:26:50), for its historical interest.
6 [Eds.] Physical quarks were introduced independently by Gell-Mann and George Zweig a couple of years after
people started playing with SU(3). They were widely disbelieved until around 1968, when deep inelastic scattering
of electrons off protons at the Stanford Linear Accelerator (SLAC) revealed structure inside the proton. See note
28, p. 859 and note 13, p. 1096.
7 [Eds.] See note 7, p. 608, and Problem 17.1, (P17.4), p. 679.
8 [Eds.] Table 36.1 is on p. 782.
9 [Eds.] See Art. 64, p. 48 in Allan Clark, Elements of Abstract Algebra, Wadsworth, 1971; reprinted by Dover
Publications, 1984. Inner automorphisms are the similarity transformations familiar to every physicist.
10 [Eds.] Evidently it was the fractional charges that discouraged Gell-Mann from initially suggesting that quarks
were physical entities; Crease & Mann SC, p. 281.
11 [Eds.] E. U. Condon and G. H. Shortley, The Theory of Atomic Spectra, Cambridge U. P. 1935; reprinted with
corrections 1951; reprinted 1991. The phases are discussed in Chapter 3, section 3, pp. 48–49.
12 [Eds.] If SU(3) invariance were to hold exactly, the masses of a multiplet like the baryon octet would all be the
same. The different masses indicate that nature breaks SU(3) invariance.
13 [Eds.] See §22.2, and note 3, p. 464. Majorana fermions are their own antiparticles, ψ C = ψ, as opposed to Dirac
fermions, which are not; Ryder, QFT, p. 429; Palash B. Pal, “Dirac, Majorana, and Weyl Fermions”, Am. J. Phys.
79 (2011) 485–498. Majorana fermions are used in supersymmetric theories.
14 [Eds.] See note 11, p. 805.
15 [Eds.] Coleman adds, “The subscripts D and F on the coupling constants have no particular meaning. It’s one
of those things Murray Gell-Mann found in the Séminaire “Sophus Lie”, so perhaps they stand for famous French
politicians.” Another possibility comes from the algebra of SU(3), whose eight generators are typically written λa
(called the Gell-Mann matrices):

where σi are the Pauli matrices, i = {1, 2, 3}, and j = {1, 2}. These λa are to SU(3) what the Pauli matrices are to
SU(2); like the σi, the λa matrices are traceless and Hermitian. Analogous to the Pauli matrix algebra, the {λa}
satisfy

One may then define the matrices (Fa)bc ≡ (Fa)bc ≡ −ifabc, in which case [Fa, Fb] = ifabcFc . Similarly one defines
the matrices (D a)bc ≡ (D a)bc ≡ dabc. As in the coupling, the matrices F are antisymmetric, and the D’s are
symmetric; see P. Carruthers, Introduction to Unitary Symmetry, Interscience Publishers, 1966, Sections 2.2, p.
30, and 2.6, pp. 50–52; or Table I, in M. Gell-Mann, “The Eightfold Way”, Caltech Synchrotron Radiation
Laboratory report CTSL-20, 1961 (unpublished); reprinted in The Eightfold Way, M. Gell-Mann and Y. Ne’eman,
Benjamin, 1964. Coleman never writes down the λa explicitly.
16 [Eds.] H. Georgi, Lie Algebras in Particle Physics, Addison-Wesley, 1982; 2nd ed., Perseus Books, 1999; for
Élie Cartan, see note 11, p. 1017.
17 [Eds.] This opinion is not widely shared. The editors have inserted weight diagrams where they were thought to
be helpful.
18 [Eds.] Coleman, “Fun with SU(3)”, op. cit.; S. Coleman, “The Clebsch–Gordan Series for SU(3)”, J. Math. Phys.
5 (1964) 1343–1344.
19 [Eds.] J.J. de Swart, “The Octet Model and its Clebsch–Gordan Coefficients”, Rev. Mod. Phys. 35 (1963)
916–939; reprinted in M. Gell-Mann and Y. Ne’eman, The Eightfold Way, Benjamin, 1964.
20 [Eds.] See §24.3, and Problem 14.3, p. 1082.
21 [Eds.]J. R. Taylor, Scattering Theory: The Quantum Theory of Nonrelativistic Collisions, J. Wiley & Sons, 1972,
Section 6-g.
22 [Eds.] A maven (rhymes with “raven”) is a connoisseur or expert; Rosten Joys, p. 221.
 Problems 20

20.1 One day someone suggests to you that, in addition to the ordinary photon, there is a second, heavy photon,
with exactly the same interactions but with a mass M, very much larger than the muon mass and, a fortiori, the
electron mass. You decide to investigate the possible existence of this particle by seeing whether its effects on the
anomalous magnetic moments of the electron and the muon are detectable. What lower bound on M do you
deduce from the fact that conventional theory fits 1 + F2(0) for the electron with an error of no more than 3 ×
10–11? For the muon with an error of no more than 8 × 10–9?

Comment : If you carry the answer out to more than 10% accuracy, you don’t understand the meaning of
experimental error.
(1998b 8.2)
20.2 In class I said (see the paragraph following (35.13), and also note 12, p. 757) that we could calculate the first
O(e4) effects of strong interactions on lepton magnetic moments if we knew σ(q2), (here written as ρ(q2) to avoid
confusion with the cross-section σ, which plays a leading role in this problem) in the spectral representation (34.4)
of the renormalized photon propagator:

By arguments identical to those given for a scalar field theory (§15.2) leading to (15.12), ρ(k2) is given by

where the sum runs over all states except the one-photon state, and pn is the total momentum of the n-particle
state. From this formula it is clear that ρ(k2) is O(e2).

Let σT(a2) be the total cross-section for electron-positron annihilation into hadrons (see Figure P20.1),
averaged over initial spins, with total center of momentum energy a. Show that

and find the constant K. The subscript H on the ρ indicates the contribution of hadronic intermediate states only.

Figure P20.1 Amplitude for e+ + e- → hadrons (figure from B. Grossman’s solution (1979b 12))

(1979b 12; 1998b 8.3)

Solutions 20
20.1 The heavy photon would make an additional contribution δF2(0), much like (34.42), except that the heavy
photon propagator has a denominator of (q2 − M2). The penultimate result is given in (35.15); all we have to do is
substitute M for a. But let’s go through the steps. There is no change to the numerator, but the new propagator
results in a denominator D′:

because the electrons are on their mass shells; p2 = p′2 = m2. As in (34.53) we combine the denominators with
Feynman’s trick, and the denominator becomes

Following (34.54) and (34.55), we effect the same completion of the square and exactly the same shift q = q′ − xp′
− yp, and obtain

As in the original calculation, we use the relation k = p′ − p (k is the momentum of the external photon), so k2 =
2m2 − 2p ⋅ p′, and

We drop the second-order term in k. Following the same steps as the original calculation, we extract the
contribution δF2(0) as

We evaluate the q′ integral using the table in the box on p. 330 and get

which is exactly (35.15), with a → M. We are told M ≫ m, so we can approximate

Let ϵ be the disagreement between experiment and the usual theory in the measurement of 1 + F2(0). Then

For the electron (m ≈ 0.5MeV), M ≥ 3GeV; for the muon (m ≈ 100MeV), M ≥ 30GeV.

20.2 The expression cited, (P20.2), is the vector version of the Kallén–Lehmann spectral representation (15.12).
The hadronic contribution ρH(k2) to ρ(k2) is

where the sum is over all intermediate states with at least one hadron. We could use the LSZ reduction formula to
relate the matrix elements án|A′µ(0)|0ñ to the Green’s function á0|T[ϕ′a(x1) ϕ′n(xn)A′µ(0)]|0ñ, but that’s not
necessary for this problem.

We want to relate ρH(k2) to the total cross-section σT(k2) for e+-e− annihilation into hadrons, averaged over
initial fermion spins. First, we calculate the amplitude for a given hadronic final state (see Figure P20.1):

The spin-averaged hadronic cross-section in the center of momentum frame of the e+-e− pair is, from (12.13) and
(21.111),
Let kµ = pµ + p′µ be the 4-momentum of the photon in the center of momentum frame:

The spin sum is worked out using Casimir’s trick, (21.112)

Putting the pieces together, we have

where

We are told to drop terms of O , and

so we can replace the square root by 1, and (S20.12) becomes (recalling that Mµν is O(e2))

We turn now to the hadronic spectral density, ρH. It satisfies the constraint (P20.2)

Contracting this constraint with p′µpν + p′νpµ − gµνa2 gives

Recalling ρH is O(e2), we have

and finally we obtain what was to be shown,

with K = 1/π.

38
 SU(3): Proofs and applications

Last time I said I would prove that the representations we constructed are irreducible, inequivalent and complete. I
will redeem that pledge now.1 After the proofs, I will move on to applications.

38.1Irreducibility, inequivalence, and completeness of the IR’s

In order to prove these properties of the IR’s, I will have to steal two general theorems from group theory; the
proofs are in Tinkham’s book or Wigner’s book.2 At least the first of them should be obvious.

Let G be some compact3 Lie group and let g ∈ G be an element of G. Representations of compact groups are
always equivalent to unitary representations,4 which are always completely reducible5 into direct sums of finite-
dimensional, inequivalent irreducible representations; we never need worry about infinite-dimensional ones.6
Thus given a representation D(g) of a compact group G, we can write

where {D (r)(g)} is a complete set of inequivalent irreducible representations, r is some index (or perhaps a
multiplet of indices), and the integers nr are the number of times the D (r)(g)’s appear in the decomposition. For any
group we have the one-dimensional trivial representation D (0)(g) = 1 for every element g ∈ G. If we consider the
direct product of any two representations

this is itself a (unitary) representation and so it is equivalent to a sum of irreducible representations:

ntrs is the number of times the representation D (t) occurs. 7

Theorem 38.1.

That is, the number of times the trivial representation occurs is once if D (r) = D (s), and not at all if D (r) ≠ D (s).8
This is in a sense a fact we all know, sometimes called Schur’s lemma.9 In field theoretic language it is the
statement that if we have a set of fields that transforms according to an irreducible representation of the group, we
can make one and only one mass term from the field and its conjugate. If you foolishly tried to make an invariant
mass term from a field that transforms one way and the conjugate of a field that transforms the other way, say from
an isovector and an isotensor, you couldn’t make it at all: there is no such invariant. Equivalently we could
consider D (s) as labeling a set of states on the left of the S-matrix and D (r) as labeling a set of states on the right of
the S-matrix. Then the statement is that if r and s are different, there is no invariant S-matrix element: they cannot
scatter into each other; and if the states do transform the same way, r = s, there is only one invariant S-matrix
element. You can take Theorem 38.1 on trust or look it up in the books; we are going to exploit it. This theorem has
a corollary that gives us a trivial test for irreducibility.

Corollary 1. D(g) is irreducible if and only if D ⊗ D contains D (0) once and only once.

Proof:

If D is reducible, then when I multiply it by its conjugate, I’ll get a sum of terms as in (38.2). Every irreducible
component D (i) in D will be multiplied by its conjugate D (i) in D, and I’ll obtain as many D (0)’s as there are
irreducible components. The only way that D (0) could appear once and only once is if D contains only one
component, i.e., it is irreducible. QED
 Thus all I have to do to confirm that the (n, m)’s are irreducible is to check how many times a direct product of
an IR with its conjugate contains a trivial representation of the group. Since I don’t know yet that the putative IR’s
(n, m) are irreducible, I first have to find out how many times each representation (n, m) contains D (0). If I’m lucky,
the answer will be that only (0, 0) contains D (0). But I haven’t proved that.

So the first step is to determine which (n, m)’s of SU(3) contain D (0), i.e., which ones contain an object that is
invariant under all group representations. If an (n, m) contains such an invariant object, it must have zero isospin
and zero hypercharge:

We happen to have a handy algorithm (§37.3) for determining the isospin-hypercharge content of any (n, m).
From that algorithm it’s clear that only (n, n) is a possibility. Let’s look at this in more detail. From the block
diagrams discussed in §37.3, if we have different n’s and different m’s, when the isospin adds up to zero the
hypercharge will not. And when the hypercharge adds up to zero the isospins will be different; see Figure 37.4.
The only time10 both I = 0 and Y = 0 is when n = m. So we’ve only got to look at (n, n), which contains only one
state with I = 0, Y = 0. For example, if we look at (1, 1), the thing with I = 0, Y = 0 is , the 3-3 component of the
tensor . That’s obvious: isospin acts only on indices with value 1 or 2, and a tensor with an equal number of
upper and lower indices, all of which have only the value 3, has Y = 0. Likewise, if we look at (2, 2), the I = 0, Y = 0
piece would be the component of the tensor ; that part is unchanged under isospin and hypercharge
transformations. Can this component be invariant under all group operations? No, because SU(3) contains, in
particular, a transformation which switches the third basis vector with the second. So there is a group element g
such that

The component is not invariant under the isospin-hypercharge subgroup. The only possibility, then, is (0, 0),
which contains only a single element, and nothing changes under isospin and hypercharge transformations. To
our question “Which IR’s contain D (0)?”, we now have an answer: only (0, 0). That comes simply from the isospin-
hypercharge block algorithm. So the algorithm is not only useful for computing things, it’s useful for proving
general theorems.

Therefore, to check for irreducibility, we have only to compute how many times the direct product of a given
representation and its conjugate contains (0, 0). If we know how many times it contains (0, 0), we know how many
times it contains D (0) and then we’ll know whether or not it’s irreducible. We are trying to show that the
representations (n, m) are irreducible, so we consider the direct product (n, m) ⊗ (n, m).

Theorem 38.2. In (n, m) ⊗ (n, m), the representation (0, 0) appears exactly once.

Proof: We will use our algorithm to count representations in the decomposition of the direct product. Let’s
begin at the end: we want (0, 0) to come out when we’re done. From (37.50) and (37.4), the only four-index symbol
that leads to (0, 0) in stage 2 of the algorithm is (0, 0;0, 0):

In our algorithm for reducing the four-index symbols we always take two off of one set of indices (upper or lower)
and add one to the other set as in (37.50). Well, we’re never going to get zeros by adding ones to some positive
number. The only way we’re going to get zeros is from (0, 0; 0, 0), produced in stage 1 of the algorithm. How many
times is (0, 0; 0, 0) produced from (m, n) ⊗ (n, m)? Recall from (37.45) that in stage 1, we take a single index from
the outside pair (in this case, reducing the m’s to m − 1’s) and from the inside pair (turning the n’s to n − 1’s). We
can do the former operation m times and the latter n times. And that’s the one and only time that the four-symbol
(0, 0; 0, 0) will be produced, when we take m indices off the m’s and n indices off the n’s:

Consequently, the representation D (0) appears once and once only in the direct product (n, m) ⊗ (n, m). QED

Theorem 38.3. The IR’s (n, m) are irreducible.

Proof: By the corollary, (n, m) is irreducible: D (0) appears but once in (n, m) ⊗ (n, m). QED
 Up to now I hadn’t proven that the representations (n, m), which I have cavalierly referred to as IR’s, are in
fact irreducible. One of them might have been the direct sum of 17 irreducible representations, including D (0). We
had to prove that only (0, 0) contains D (0) before we could establish that the (n, m)’s really are irreducible
representations, and thus deserving of the label “IR”.

Next, I will demonstrate that the IR’s (n, m) are inequivalent (for different n’s and m’s). That comes easily from
the Theorem 38.3. Say that the IR (n′, m′) is equivalent to (n, m). Consider

then how may times does this contain (0, 0)? It will contain (0, 0) after the second stage of our algorithm only if it
contains a term (0, 0; 0, 0) with four zeros after the first stage. Because we subtract equal numbers of indices from
the outer and inner indices and stop when we reach a zero, the only way to reach four zeros is if n = n′ and m = m′.
Thus we have:

Theorem 38.4. The representations (n, m) and (n′, m′) are equivalent only if n = n′ and m = m′.

Earlier we were concerned that as (4, 0) and (2, 1) were both 15-dimensional, they might secretly be
equivalent. But it’s not so: multiplying (0, 4) ⊗ (2, 1) as in (38.2), we’d have to get nrs0 = 1 for them to be
equivalent. But (0, 0) does not appear in the direct product; nrs0 ain’t one, it’s zero. So (4, 0) and (2, 1) are
inequivalent, despite having the same dimension. So far we have a set of representations that are guaranteed to
be both irreducible and inequivalent. Have we found all of the irreducible representations? That is, are they
complete? We know that when we used the tensor trick for SU(2) in §36.3 we got all of them. On the other hand if
we had tried the same trick for SO(3), we would have missed the spin-½ representations.11 So have we found
them all, or are we missing some?

We’ll now steal another theorem from group theory, the so-called orthogonality theorem12

Theorem 38.5. Let G be a compact group and as before, let D (r)(g) be a complete set of inequivalent irreducible
representations. Then

The subscripts on the D’s indicate matrix elements. We put coordinates on the group and we have a little
Jacobian determinant there; we integrate over the whole group. (We also know what the integral is when r = s but
we don’t need that for the theorem.13) It’s the statement that, for U(1) or SO(2), for example, where the irreducible
representations are all one-dimensional, einθ, that

It happens to be true in general.14

Consider the representation (1, 0), which has15 eight independent matrix elements D αij(1, 0) , α = 1, 2, . . . , 8.
For that representation we’ll consider all the matrix elements together and write them (and their conjugates, D α (0,
1) (g)) as

The set {yα } are coordinates in group space. If we know the yα ’s and the yα ’s, we know what the group element is.
When I take the direct product of two representations I get matrix elements which are simply the ordinary
numerical products of the matrix elements of the original representations. So, direct products have matrix
elements that are monomials in the yα ’s and the yα ’s.

Let us now prove by contradiction16 that we have all the representations of SU(3). Assume that there is some
irreducible representation, D (?)(g), which we have missed. By the orthogonality theorem, (38.10), its matrix
elements are orthogonal to those of all the representations that are in our list. In particular, we have
Then D ij(?)(g) must be orthogonal to all linear combinations of the yα and yα . But it must be also orthogonal to the
matrix elements of (1, 0) ⊗ (1, 0), (1, 0) ⊗ (0, 1), and (0, 1) ⊗ (0, 1), i.e., to yα yβ, yα yβ, and yα yβ. All the
representations in our list contain everything that can be made out of direct products, and so D ij(?)(g) is orthogonal
to every polynomial P(yα , yα ) in the yα ’s and the yα ’s:

This is because D ij(?)(g) is not in our list, and so it has to be orthogonal to the others. Now by the approximation
theorem of Weierstrass, given a complete set of coordinates on any space, anything that is orthogonal to all the
polynomials in the coordinates has to be zero.17 Therefore,

That is the unique function orthogonal to all the polynomials. Then D ij(?)(g) can’t be a representation because
every representation must equal 1 when g equals the identity element. QED This orthogonality proof is a very nice
method. It’s practically the only trick in the past few lectures which has not been stolen from Hermann Weyl or
Claude Chevalley.18

That is the end of the mathematics part of this lecture. If you didn’t follow the math, don’t worry. It’s pretty, it’s
fun, and you’ll understand things more deeply if you understand these arguments, but this discussion was not
particularly about quantum field theory. Suffice it to say that in order to make myself an honest man I have
explicitly proved to you that the representations (n, m) are indeed what I have been acting as if they were: a
complete set of inequivalent, irreducible representations of SU(3).

38.2The operators I, Y and Q in SU(3)

You’ll recall that in recent lectures (§24.3, §§35.3–35.4, §36.1) we derived all sorts of electromagnetic relations
between form factors and magnetic moments, assuming that the SU(2) of isospin was perfect, and treating
electromagnetism only to lowest order.

Now we will do the same thing with SU(3): I will neglect the effects of the medium-strong interactions and
assume that SU(3) is perfect. It’s a bigger group so we should get more relations, perhaps something a bit more
useful than the one (35.68) that connected the magnetic moment of the Σ+ to that of the Σ- (sadly, far beyond the
reach of current experiment). We hope to connect theory with something that we can actually measure. The first
thing I’ll look at are electromagnetic formulas in the limit of perfect SU(3). Or, going to all orders of the medium-
strong interactions, to first order in electromagnetism and zeroth order in the cross terms between
electromagnetism and the medium-strong interactions.

We expect to make errors on the order of 10–20% by assuming that SU(3) is perfect. After all, in the particular
case of the baryon octet, the individual baryons lie within 15 or 20% of the mean mass of the octet. That’s what
we’ve got to live with, until we have a complete dynamical theory of the SU(3) symmetry breakdown. Then we
could take care of the medium-strong interactions with Feynman diagrams.

You will recall that in our isospin analysis, the key point was that the electromagnetic charge was a generator
of the group, one of the symmetries of the strong interactions. And therefore the electromagnetic current, by the
minimal coupling prescription, transformed like this generator, the charge, which by the Gell-Mann–Nishijima
relationship was (35.53) a linear combination of the z-component of the isospin, an isovector, and hypercharge, an
isosinglet. We don’t know yet how the generators of SU(3) transform under the action of SU(3). Before we can
start applying the same techniques, we’ll have to figure out how they transform. Are they a (3, 0), a (1, 1) or
something else? We need some preliminary work to determine the SU(3) generators.

Let me just remind you how we deduce the angular momentum transformation properties under the action of
the rotation group.19 In classical mechanics we have a three-dimensional vector x which undergoes an
infinitesimal rotation defined20 by a three-dimensional rotation matrix R:

The rotation R is specified by an axis and an angle θ. It is an element of the group SO(3):
An infinitesimal rotation can be represented as an operator D specified by R:

J is a vector whose components are operators. Under the action of D(R)

Since D(R) is unitary, J is Hermitian:

We can apply the same analysis to SU(3). First we have a Hilbert space in which we have our quantum
mechanical theory. We have isospin I, a three-vector composed of operators; we have a unitary operator U(R) in
the Hilbert space associated with isospin transformations; R is an isospin rotation matrix acting on the operators I.
Here the operator is an element of SU(3):

The analog of (38.18) is

That’s the statement that the three generators of isospin transform like an isovector. It comes out as a vector
because isospin is an SU(2) subgroup of SU(3), and SU(2) is the covering group21 of SO(3). Again we have an
infinitesimal rotation, now in isospin space, labeled by an axis and an infinitesimal angle δθ. It’s close to the
identity and it goes about some axis by an infinitesimal angle. The corresponding U for the infinitesimal rotation is

linear in the components of I since we’re only going to first order in δθ. That’s the primary definition of I: I is an
isovector because it’s dotted into , which is an isovector, and δθ is what labels the rotation. Very shortly I will
go through the same analysis for SU(3). We’ll see how to label an infinitesimal SU(3) transformation and then we’ll
know how the generators of SU(3) must transform by the same reasoning.

Before we do that we need to know how to represent isospin generators as matrices. The three-dimensional
rotation group is the same, at least locally, as the two-dimensional unitary group and we know that the isospin
generators transform like the three pions. We went through considerable labor to find out how to write the three
pions as a 2 × 2 traceless matrix. Recall (36.85) that we found

For convenience, I’ll multiply ϕ by ; everything will still transform the right way.

From this we can read off the 2 × 2 matrix that corresponds to the three isospin generators. I’ll write it down and it
will look like there are a pair of algebraic errors but in fact it’s the right answer:

Looking at (38.24), apparently I’ve made two algebraic errors: I’ve transposed the plus and minus components
and I’ve left out the .

In fact, I’ve done neither. The raising and lowering operators for isospin are

whereas the charged pion fields are defined with in the denominators. Therefore, I± is what corresponds to
ϕ ∓, not what corresponds to ϕ ±. That’s where the went.

Why have the plus and minus components of I switched places from ϕ? Because ϕ + is the field that
annihilates positively charged pions, that is, lowers the isospin, and I– is the operator that lowers the isospin, just
as p annihilates the |pñ state. I’m sorry, but that’s the way life is! Minus and plus are used in different ways when
defining isospin raising and lowering operators and charged pion fields. We have to live with that convention
clash. That’s why the minus and plus components have changed places (as in (24.21)) and why the has
disappeared: I– is the isospin lowering operator while ϕ + is the isospin lowering field, because it annihilates a π+
which has positive Iz .

Now we’ll work out the generators of SU(3). If g is an SU(3) matrix it obeys two equations:

We want to consider an infinitesimal SU(3) transformation, which has the form

with ϵ a 3 × 3 matrix and δθ an infinitesimal angle. All the “direction” part of the transformation is in ϵ, just as all
the direction in SU(2) lies in the choice of .

We deduce that ϵ is a 3 × 3 Hermitian matrix:

To find detg, look in a frame where ϵ is diagonal:

Then (38.28) becomes

The determinant of this matrix is the product of the diagonal terms. We ignore the terms of order (δθ)2 and higher.
So

Though we have computed the determinant in a coordinate frame in which ϵ is diagonal, the trace is independent
of what coordinate basis we use. Thus

The group SU(2) has 22 − 1 = 3 generators, and so needs three parameters, characterized by a vector, to
describe a group element; SU(3) has 32 − 1 = 8 generators, so its parameters are conveniently characterized by a
traceless 3 × 3 matrix.

Parallel to (38.22), we write

and Tr[ϵG] is a linear function of ϵ; ϵ is a matrix, G is a matrix of operators just as I is a vector with operator
components. In order that U(g) be unitary, G must be traceless and Hermitian:

Just as we deduced from (38.21) that I transforms as a vector, so we deduce here that G is a 3 × 3 matrix of
operators that transforms as an octet, to wit,

where g ∈ SU(3). This equation has exactly the same form as (38.21). In the center of the left-hand side we have
a 3 × 3 matrix of operators; U(g) is a unitary operator that implements SU(3) on some Hilbert space, likewise
U †(g); and g and g† are 3 × 3 matrices of numbers. We have the correspondence.

Let’s try to figure out the explicit form of G. Here’s what we know so far:
 Table 38.1 Correspondence between SU(2) and SU(3)

1.In the upper left 2 × 2 block where the pions sat in the pseudoscalar octet, (37.35), the isospin
generators must sit, as in (38.25). Those are the things that transform as an isovector, just like the
pions.

2.On the diagonal, where we would have the η in the pseudoscalar mesons, we must have the
isosinglet symmetry generator, the hypercharge Y (with a multiplicative constant α that we’ll have to
determine).

3.In the other spots we will have some generators that transform like the kaons: strangeness-changing,
hypercharge-changing generators. We won’t study them here. They are however very important in
weak interaction theory where they have names like λ5 and λ6, again because of historical
conventions.22 We’ll fill parts of the matrix we aren’t considering with an asterisk, *.

Therefore the matrix G looks like this:

To check the normalization, consider the 3 × 3 defining (“fundamental”) representation of the group, (1, 0) or
3; the quarks. After all, this is a matrix of generators for any representation; in particular, it should be true for the
quarks. (For convenience, Table 37.3 of quark properties is reprinted below.) For Iz , the ϵ that corresponds to an
infinitesimal Iz rotation is determined by the condition

Table 37.3 The quarks and their properties

That is what an infinitesimal Iz rotation does to the defining (1, 0) representation. Multiplying (38.39) by (38.38) we
find

That’s jolly good; we didn’t make some dumb mistake with the normalization.

Now let’s determine α. An infinitesimal rotation in the Y direction is determined as in (38.39), and we find

Multiplying (38.41) by (38.38) we obtain

But we want the trace to equal Y, the generator of infinitesimal hypercharge rotations, not 17 times the
hypercharge. So we set

Therefore

We’ll be studying electromagnetism shortly so we need to find the ϵ corresponding to a Q rotation:

This ϵQ is a linear combination of the corresponding ϵ’s that we already have, (38.39) and (38.41):

Naturally, the charges of the three quarks are the eigenvalues of ϵQ. Finally, since we will be doing three problems
involving the baryon octet, we’ll again write down that matrix, (37.32), with B taking the place of ψ:

We now have all the machinery we will need. We have the ϵ matrix that corresponds to the electric charge, we
have the baryon octet matrix and we have the matrix of generators, which we won’t need for a while.

38.3Electromagnetic form factors of the baryon octet

We are first going to study the electromagnetic form factors of the baryon octet.23 Aside from the neutron and
proton, the only form factors that have been measured (for five of the other six baryons)24 are the magnetic
moments, and the electric form factors F1(0).

Consider the matrix element of a general current in the octet j µ (a matrix made out of currents just like the
generator matrix G) between some final baryon state described by B′ and an initial baryon state in B,

(all space and spin indices have been suppressed). These currents will be involved in a variety of reactions.
We might want to look at the hypercharge form factor or the isospin form factor; they might be different. We could
look at strangeness-changing currents; those turn out to play an important role in weak interaction theory,
although they’re not the only currents. For the moment, we are concerned with the electromagnetic current, and
therefore we will be interested in

where ϵQ is the charge matrix, (38.46); that choice of ϵ picks out the electromagnetic current:

For a general matrix element made from an octet current between two octet baryons, how many independent
matrix elements are there, apart from the functions25 F1 and F2? That is, out of an 8 (|Bñ) and an 8 (j µ), how many
8’s (áB′|) can we make? We know the answer. From (37.58), we see that we can make two 8 multiplets:
Thus, the general matrix element, for any baryon on the right, any baryon on the left and any ϵ, is given in terms of
just two quantities, neglecting space and spin dependence: two F1’s and two F2’s. In particular, this means that if
we know, in the idealized limit of perfect SU(3) symmetry, the electromagnetic form factors F1 and F2 of the proton
and the neutron, then we know F1 and F2 for every baryon in the octet. And furthermore we know the matrix
elements of the strangeness-changing currents and the hypercharge currents and any other linear combination
we want.

It’s exactly like meson–nucleon coupling, (37.37): an octet coupled to two octets. We can write the most
general form

where α and β are scalar coefficients, B is the 3 × 3 matrix of incoming baryons and B′ is the 3 × 3 matrix of
outgoing baryons; the 3 is to tidy up the denominators in the charge matrix (38.46). The coefficients α and β are of
course actually functions of space and spin, they’re F1(q2) and F2(q2) with all sorts of spinor factors; if we’re just
looking at the magnetic moment they’re simply numbers. The B and B′ describe which baryon we’re looking at.
For example, if the initial and final states are both protons,

The particular ϵ that we pick, (38.46), (38.41) or (38.39), determines which form factor we get. These forms occur
because we have employed minimal coupling, which goes through for SU(3) just as it does for isospin; an 8
operator acting on an 8 state going into an 8 state. With the choice of ϵ = ϵQ, (38.50) will tell you all the magnetic
moments. I will evaluate this formula shortly for the specific case of interest, the electromagnetic current matrix
element.

In principle, how many observable form factors are there? There are eight baryons and they can all have
different magnetic moments, even in the limit of perfect SU(3) symmetry. From (38.50) we get eight objects,
arising from matrix elements of the form áb|jµem|bñ, without cross terms áb′|j µem|bñ with b ≠ b′. But there is one
possible cross term, between the Σ0 and the Λ: since both lie on the diagonal in the baryon octet, (38.47), the
electromagnetic current can have a non-zero matrix element between them. Moreover, the selection rules (35.62)
for isospin and charge conjugation allow it (35.63). That’s a good thing, because the principal decay mode of the
Σ0 is just this:

and the reaction is extremely fast, ~ 10−20s. (Even after studying this decay for 20 years all we have is an upper
bound for the lifetime26.) That means it’s as low order in electromagnetism as it can be, to wit, first. And therefore
there had better be a non-vanishing electromagnetic current matrix element between the Σ0 and the Λ.

So there are nine observable quantities here: the eight magnetic moments of the baryons and the Σ0 → Λ + γ
transition matrix element. (In the limit of pure SU(3) symmetry, the latter is F2-dominated, since F1 between Σ0
and Λ vanishes: F1 is the matrix element of the charge operator and both have zero charge.) Our formula enables
us to deduce these things in terms of two parameters, α and β. We can then solve for α and β in terms of the
proton and neutron moments and predict the other moments. In principle, we will find in the literature all nine of
these things computed in terms of the magnetic moments of the proton and the neutron. But I will be somewhat
less ambitious and just compute the ones that I can find in the Particle Data Group tables.27 If you want to
compute others and make prophecies about future tables you are encouraged to do so. The current table has the
measurements of seven baryon magnetic moments in it.28 The proton and the neutron are known precisely; the
others have relatively large uncertainties.

We’ll start with the proton. The important thing to remember is that when we multiply a matrix by a diagonal
matrix, life is very simple. If the diagonal matrix is on the left, every row of the other matrix gets multiplied by the
diagonal entry; if we multiply it on the right every column of the other matrix gets multiplied by the diagonal entry.
For the proton, from the 3α term we get
And from the 3β term we get −β. Continuing with the other baryons, we obtain 29 Table 38.2. The theoretical values
for µp and µn are not listed, since they were used30 to fix α and β.The predictions are in an ideal world where there
are no mass differences between the baryons. In reality we expect them to be off by 20-30%.

Table 38.2 The baryon magnetic moments expressed in nuclear magnetons, µN =

These coefficients, α and β, could be chosen to fit the magnetic moments, the charges, or indeed any linear
combination of the two form factors. Just to confirm that we haven’t made any algebraic errors let’s check that
things are right if we use the charge F1(0) instead of the magnetic moment. If we set

then we replace the µ’s in the above chart by Q’s. We find

Everything checks out.

The proton and neutron moments, ever since they were first measured by Rabi, have been determined to a
fare-thee-well. Compared to the hyperon moments31 they have practically no experimental uncertainties at all.32
In Table 38.2 I wrote down the theoretical answers with no experimental errors; p and n errors—and hence those
in α and β—are negligible.

The units are nuclear magnetons, using the proton mass for the magnetic moment:

Shouldn’t we use the hyperon’s mass instead of the proton’s? Who knows? There’s a 20% difference. This is after
all a computation for perfect SU(3); it would be cheating to make a decision on that. We’ll use the proton mass
because that’s how they’re expressed in the literature. Over time the numbers in Table 38.2. have gone up and
down like the Dow Jones average. We have to be genuinely sophisticated to know the meaning of the
experimental uncertainty; that’s the secret wisdom of the theorist. The standard deviation in modern high energy
experiments is a unit like the ‘league’ in medieval Europe: a German league was three and a half times as long as
an English league.33 There is just as vast a difference in standard deviations. Even if we take the standard
deviations dead seriously, which I would not advise doing, the error bars stay narrow but the number leaps up and
down from year to year. This is good.34 The only thing you can decide, if you make a ten-year average, is that it’s
something with a 1% error, but that ‘something’ we know only to within 50%. These are very hard experiments,
because these particles don’t live long. Rabi got the Nobel Prize (1944) for measuring the proton’s magnetic
moment. The Σ+ is a lot trickier. Though we haven’t succeeded in measuring the Σ0’s magnetic moment, we do
have a measurement35 of the size (but not the sign) of the transition moment in Σ0 → Λ + γ:

We can compute the transition moment, with B′ given by the Λ’s matrix and B by the Σ0’s, and we find

The magnitude is quite close, even if the sign is not yet established.

The Λ moment was measured in a precession experiment. 36 You don’t have to polarize it; you can tell its spin
by a fluke. It decays very asymmetrically into a nucleon and a pion

There’s a large correlation between the spin of the Λ and the direction of the decay products, so you can tell how
it’s spinning from their trajectories; it’s also produced preferentially in a certain spin state. You make a beam of
Λ’s, send them through a magnetic field, watch them decay and measure the precession of the magnetic moment.
This is not easy. Even so, the agreement is within 20%, rather good even if you take the experimental errors
seriously. Improving these results is difficult. But getting them from symmetry arguments is easy. This comparison
looks very promising.

38.4Electromagnetic mass splittings of the baryon octet

We can also use SU(3) to study second-order electromagnetic effects, just as we did37 for η decay in SU(2). As
an example of that, we will study electromagnetic mass differences between members of the same octet; for
instance, between Σ+, Σ0 and Σ−. That is a second-order electromagnetic effect, we believe, although nobody can
compute it because it diverges. However, since all we’re going to get are linear relations between things, we don’t
care what makes it finite. And in fact this will offer us not only a cute way of testing SU(3) but a cute way of testing
the idea that the mass difference is purely electromagnetic, since we can’t test it by computing the proton/neutron
mass difference.38

The first thing we’ve got to do is count the number of invariants we have, in order to see if we can make any
prediction. The second-order electromagnetic effect transforms like the product of two currents, as far as its
internal symmetry properties go:39

The product of the two currents thus can be regarded as the known direct product (37.58) of an octet with an octet:

This is the product of any two currents: any current from the octet with any other current from the octet. Of course
we’re interested in the case where both of them are electromagnetic currents or, more to the point, both currents
are the same. That means that the antisymmetric (under parity) combinations cannot appear. So one of the 8’s,
the 10, and the 10 are out by antisymmetry (37.63). We have the initial and final baryons, B and B′, respectively,
that have to be hooked together with this product in an SU(3)-invariant way. That’s also 8 ⊗8, but with no
particular symmetry or antisymmetry. Here are the states and the possible coupling to the current product:

(I’ve crossed out the antisymmetric representations). We can make an invariant by hooking a 27 to a 27, the 8 in
j µem ⊗jνem to either of the 8’s in B′⊗B, and the 1 to the 1. However, the singlet to the single is irrelevant: it just
shifts all the masses by the same amount. It’s an electromagnetic mass shift, but it doesn’t produce an
electromagnetic mass difference. Therefore we have three unknown constants and there are four observed
electromagnetic differences: one within the neutron–proton, one within the cascade and two within the Σ’s (Σ+ −
Σ0 and Σ-- Σ0) so the computation is worth doing. With four experimental quantities and three free parameters we
can make one prediction.
 Now we have to write down the three invariants. They will involve ϵQ twice because they involve the current
twice. We will just write down three linearly independent SU(3)-invariant terms for the electromagnetic contribution
Δmem to the mass splitting:

This expression has the right properties: it’s linear in B, antilinear in B′, involves two ϵ’s for the two currents, and is
an SU(3) invariant.

I could now begin to calculate but it’s useful to simplify the matrix algebra by observing that

where I is a 3 × 3 identity matrix, and P is the projection operator

Then

Therefore we could write Δmem as

The last term, dTr(B′B), is a mass shift that affects all the baryons equally, and we can drop it. It comes from all
the I’s in the product. We went through this little trick because it’s easier to compute the terms for a matrix that’s
mainly zeros than for a matrix that’s full of ’s and ’s. We would get the same result using (38.60).

Let’s write down all the things we’re going to worry about. Multiplying by P on the left just multiplies the first
row by 1 and annihilates all the other rows; multiplying by P on the right multiplies the first column by 1 and
annihilates all the other columns. So we get an a for the p, an a for the Σ+ and a a for the Σ0. In this way we get
Table 38.3..

Table 38.3 Electromagnetic contributions to the mass shifts, Δmem

Now we can form linear combinations of the observable differences that are independent of a, b and c and are
therefore zero. The difference that is least well measured is mΞ-− mΞ0:

We don’t want to introduce the Σ0; we have no other information on c. Instead, write b = (b − a) + a:

This is the desired formula.40 How does it compare with the experimental data? The observed mass splittings41
are (in MeV):

From this we compute the mass difference between Ξ- and Ξ0:

The prediction is pretty good. Actually the agreement is surprisingly good, when we recall the differences between
the predicted and experimental values of the magnetic moments.

Aside.

Some day there will be one other result deduced from (38.64). The electromagnetic corrections to the
Hamiltonian have an allowed off-diagonal term that can connect Σ0 to Λ. Therefore, although I didn’t compute it
here, there is a small amount of mixing between the Σ0 and the Λ induced by this allowed off-diagonal term.
Equivalently, there is a tiny transition vertex, which is computable from the off-diagonal term and is of the same
order as all these other things, 4 or 5 MeV. A Σ0 comes in, something electromagnetic goes on, and a Λ comes
out.42

Figure 38.1 Electromagnetic correction to Σ0 → Λ decay

This will yield a correction to the 8 ⊗8 baryon propagator. It is a second-order electromagnetic interaction,
just like the mass shift. It is not analogous to the magnetic moment: the photon is only a virtual photon. One may
ask: How can a neutral particle have an electromagnetic mass shift? The answer is that the neutral particle
decomposes virtually into into charged particles which then recombine. There are all these charged baryons and
mesons in the theory: Σ0 can become (for instance) a Σ+ and a π−. That’s why the neutron has a magnetic
moment. The blob has all sorts of things inside; we don’t know what. We don’t understand the details of how
electromagnetism combines with the strong interactions. If we did, we could compute the proton–neutron mass
difference, which we can’t.43 But we can explore this hypothesis. This marvelous agreement with experiment
(38.68) not only tests SU(3), it checks the idea that the mass differences are electromagnetic.

Now, how do we measure this? It’s not easy. It introduces a small amount of mixing of Σ0 and Λ or,
equivalently, by time reversal, of Λ and Σ0. So the Σ0’s we see coming out are not 100% Σ0 (the neutral member of
the isotriplet); they have a tiny admixture of the isosinglet Λ and vice versa. But that’s a hard thing to look for.

Some years ago Richard Dalitz made a suggestion for measuring this quantity. But it doesn’t give a good
check, so I didn’t bother to look up the numbers in the literature; the uncertainties are still too large. It has to do
with things called hypernuclei.44 Every once in a while, when a Λ goes into a detector, it gets captured by a proton
and forms something like a deuteron, but made out of a proton and a Λ. And then, because the Λ is unstable, it
decays and we see this hypernucleus exploding. It could also happen with heavy nuclei. So if we know something
about nuclear forces, we obtain some idea of the nature of the force between the Λ and a nucleon.

Now because the Λ is an isosinglet, pion exchange (the usual mechanism for the proton–neutron interaction),
cannot occur: there is no πΛΛ vertex—it wouldn’t conserve isospin. The only thing that can happen, in fact, if we
don’t take account of electromagnetic effects, is this process:

Figure 38.2 Λ capture by nucleons

That’s allowed, and that process leads to the principal force between nucleons and Λ’s. But it’s a force of
somewhat shorter range than the normal nuclear force, because instead of exchanging one pion we’re
exchanging two. The range is bounded by the mass of the exchanged particles:
The lightest thing we can possibly exchange has a mass of 2mπ. However, with the electromagnetic vertex (the
blob), we can get another process, shown in Figure 38.3. The Λ comes along and becomes a Σ0. Equivalently the
Λ coming out of the beam is not pure Λ; it’s a mixture of Λ and Σ0. From then on it’s the same story: Σ0 emits a π
and turns into a Λ and the π interacts with the nucleon (the vertex can occur in either location). We would normally
think this would be a very small correction to the Λ-nucleon interaction (it’s electromagnetic, not strong). But this is
in fact not so, for two reasons. First, it is a longer range force than the first one,

Figure 38.3 Λ–Σ electromagnetic vertex

So it can catch Λ’s that make glancing collisions. Second, it’s not as small as you might suppose. By a fluke,
the Λ and the Σ are very close together in mass, about 75 MeV apart. Thus the denominator of the Σ propagator is
rather small, for small momentum transfer. So it’s not typically electromagnetic in size; that is, not down by , but
just down by something like 5/75 ~ , because we have this small denominator amplifying a small vertex.

Thus, as Dalitz suggested in the mid-1960’s, by a close study of hypernuclei we should be able to detect this
force and estimate its coefficient. Since we know everything in the diagram, we should be able to define the
correction due to this force, deduce the Σ0–Λ mixing matrix element, and thereby get another check of SU(3).45

There is another process that runs by mixing matrix elements where the mixing matrix element is not
electromagnetic but medium-strong. But before we discuss that we will have to discuss the medium-strong
interactions and the famous Gell-Mann–Okubo formula. We’ll begin with that next time.

1 [Eds.] §38.1, from the video of Lecture 38, is again largely a reworking of the references cited in note 1, p. 797.
2 [Eds.] E. P. Wigner, Group Theory and its Application to the Quantum Mechanics of Atomic Spectra, Academic
Press, 1959. M. Tinkham, Group Theory and Quantum Mechanics, McGraw-Hill, 1964. Two more recent books on
group theory for physicists are: Howard Georgi, Lie Algebras in Particle Physics: From Isospin to Unified Theories,
2nd ed., Westview Press, 1999; and A. Zee, Group Theory in a Nutshell for Physicists, Princeton U. P., 2016,
hereafter Zee GTN.
3 [Eds.] “Compact” means a finite volume or parameter space; mathematically, a compact set is one which is
closed and bounded. The rotation group is compact; the Lorentz group is not. See note 15, p. 379.
4 [Eds.] Wigner, op. cit., Theorem 1, p. 74.
5 [Eds.] See Chapter III, Section 4, p. 123 in H. Weyl, Group Theory and Quantum Mechanics, trans. H. P.
Robertson, reprinted by Dover Publications, 1953; Tinkham, op. cit., Section 3–5, pp. 29–30.
6 [Eds.] This is a consequence of the Peter–Weyl theorem: A. Barut and R. Raczka, Theory of Group
Representations and Applications, World Scientific, 1986; A. Wawrzyńczyk, Group Representations and Special
Functions, D. Reidel, 1984; reprinted by Springer, 1986.
7[Eds.] See Section 16-3, equation (16-22), p. 436 in J. Mathews and R. Walker, Mathematical Methods of
Physics, Addison-Wesley, 1969.
8 [Eds.] Statements about groups are often easily grasped if you consider them in terms of the quantum theory of
angular momentum. Recall that a direct product of two different angular momentum states (the irreducible
representations of SU(2)) with ℓ1 and ℓ2 will give new states with ℓ bounded by ℓ1 + ℓ2 ≥ ℓ ≥|ℓ1 − ℓ2|. Note that ℓ will
not equal 0 unless ℓ1 = ℓ2. (In SU(2) there is no distinction between D(g) and D(g); all the tensors can be written
with either upper or lower indices only.)
9 [Eds.] Wigner, op. cit., Theorem 2, pp. 75–76. Coleman states in “Fun with SU(3)” (op. cit., footnote 3, p. 342):
“Actually, this is not in [Wigner] in precisely this form; however it is a trivial corollary of Schur’s lemma, and the fact
that every representation of a compact group is equivalent to a unitary representation. (It can also be derived
simply from the orthogonality relations.)” See note 12, p. 827 for a proof.
10 [Eds.] In the lecture, this statement is not proved. Here is a proof. In the decomposition of (n, m) into (I)Y IR’s
(Figure 37.4) the general term along the top edge has

while the general term along the left edge has

The (I)Y IR’s come out of the direct product (It)Yt ⊗ (Il)Yl. Those have Y = (n−m) + (k −j), and values of I given by
the Clebsch–Gordan series,

where ℓmax is determined by the requirement that Imin be non-negative:

Set both Y and I equal to zero, and solve for n and m,

Both n and m have to be non-negative, and so

Multiply the top equation by j, the bottom by k, and subtract:

Subtract this from the equation ℓ ≥ j − 2k to obtain 0 ≥ 3j. But j is a non-negative integer, so j = 0. Similarly k = 0,
and consequently n = ℓ = m.QED
11 [Eds.] See Zee GTN, Section IV.1, pp. 185–195, for the application of these tensor methods to SO(3): only
those IR’s with dimension equal to 2j + 1 (with j a non-negative integer) are found.
12 [Eds.] See Wigner, op. cit., Theorem 4, equation (9.31), p. 79 for discrete groups; for continuous groups, see
equation (10.12), p. 101. Incidentally this theorem affords a quick proof of Theorem 38.1. The orthogonality
relations (Wigner’s equation (9.31)) say

where h is the order of the group and ℓ is the dimension of the matrices D (i) (and also of the Kronecker delta). But
from (38.2) we have

because D (0) = D (0) = 1. Using the orthogonality relations on both sides gives

Canceling the common factors, we obtain n0ij = δij.QED SU(3) is continuous, and the sum over g ∈ G should be
an integral, but the theorem goes through with integrals just the same.
13 [Eds.] For completeness,

where dim r is the dimension of the representations. Wigner, op. cit., p. 101, equation (10.12).
14 [Eds.] Compare also dΩYℓm*Yℓ′m′ = δℓℓ′δmm′; the spherical harmonics Yℓm(θ, ϕ) form an irreducible
representation of SO(3) on the unit sphere.
15 [Eds.] This is the representation generated by the eight traceless, Hermitian 3 × 3 Gell-Mann matrices {λα }; see
note 15, p. 807. The representation D (1,0)(g) = exp{(i/2)θα λα } where θα are eight parameters. Similarly D (0,1)(g) =
exp{-(i/2)θα λα }. The determination of λαij as the matrix elements 2 áqi|Fα |qjñ where áqi|qjñ = δij and {Fα } are the
elements of the Lie algebra of SU(3) is worked out in Greiner & M¨uller QMS, Exercise 8.1, pp. 221–224.
16 [Eds.] Coleman Aspects, “An introduction to unitary symmetry”, pp. 16–17; Coleman, “Fun with SU(3)”, op. cit.,
pp. 343–344.
17 [Eds.] D (?) can be approximated by a sequence of polynomials Pn(yα , yα ) which converge uniformly to D (?), and
so uniform convergence gives dgPn(yα , yα )D (?)(g) → dg|D (?)(g)|2 = 0, and thus D (?)(g) = 0. See Harold and
Bertha S. Jeffreys, Methods of Mathematical Physics, Cambridge U. P., 1946, Sections 14.08–14.081, pp.
417–418. In the context of integration over group spaces, an identical argument applied to the spherical
harmonics (for SO(3)) is given by Charles Loewner, Theory of Continuous Groups, MIT Press, 1971; see Lecture
VIII, p. 62. Republished by Dover Publications, 2008.
18 [Eds.] H. Weyl, The Classical Groups, Princeton U. P., 1939, 1973 (paperback ed., 1997); C. Chevalley, Theory
of Lie Groups, Princeton U. P., 1946, 1999.
19 [Eds.] See Goldstein et al. CM, Section 4.8, pp. 163–171; Mathews and Walker, op. cit., Section 16-7, pp.
461–466.
20 [Eds.] See §18.2. Again, Coleman uses to denote the axis of rotation, and again we have changed his notation
to to avoid confusion with the base of the natural logarithms.
21 [Eds.] See note 37, p. 791.
22 [Eds.] See note 15, p. 807. Note that the operator for Y is proportional to λ8, and Ii to λi, i = {1, 2, 3}.
23 [Eds.] Sidney Coleman and Sheldon Lee Glashow, “Electrodynamic Properties of Baryons in the Unitary
Symmetry Scheme”, Phys. Rev. Lett. 6 (1961) 423–425. See also note 40, p. 841.
24 [Eds.] PDG 2016, pp. 88–95. Conspicuous by its absence is the Σ0. The theoretical value is (α + β) = − µn
(see Table 38.2. on p. 837). The Σ0’s lifetime, ~ 10−20s, has thus far precluded the measurement of its magnetic
moment.
25 [Eds.] See (34.15).
26 [Eds.] In the four decades since this lecture was given, our knowledge of the Σ0 has gotten better. It’s been
established that the Σ0 has a mean life of 7.4 ± 0.7 × 10−20 s, and this decay mode is responsible for ~ 100% of Σ0
decays: PDG 2016, p. 94.
27 [Eds.] PDG 2016.
28 [Eds.] Coleman said “five” in 1976, and listed only p, n, Σ+, Λ and Ξ−.
29 [Eds.] The experimental values in this table, taken from PDG 2016 and rounded to four places, differ from the
1976 values.
30 [Eds.]
In the more modern approach, the magnetic moments of the baryons are given in terms of the magnetic
moments of the up, down and strange quarks. The agreement between theory and experiment is much better. See
Griffiths EP, Table 5.5, p. 190; D. Perkins, Introduction to High Energy Physics, 4th ed., Cambridge U. P., 2000;
and PDG 2016.
31 [Eds.] The term “hyperon” is defined in note 11, p. 521.
32 [Eds.] G. Breit and I. I. Rabi, “On the Interpretation of Present Values of Nuclear Moments”, Phys. Rev. 46
(1934) 230–231; I. I. Rabi, J. M. B. Kellogg and J. R. Zacharias, “The Magnetic Moment of the Proton”, Phys. Rev.
46 (1934) 157–163; “The Magnetic Moment of the Deuton”, [sic ; nowadays “deuteron”], Phys. Rev. 46 (1934)
163–165.
33 [Eds.] J. B. Friedman, K. M. Figg, S. D. Westrem, and G. G. Guzman, Trade, Travel and Exploration in the
Middle Ages, Garland, 2000.
34 [Eds.] Perhaps in the sense that the fluctuations keep us honest? Your guess is as good as ours.
35 [Eds.] P. C. Petersen et al., “Measurement of the Σ − Λ Transition Magnetic Moment”, Phys. Rev. Lett., 57
(1986) 949–952.
36 [Eds.] R. L. Cool, E. W. Jenkins, T. F. Kycia, D. A. Hill, L. Marshall and R. A. Schluter, “Measurement of the
Magnetic Moment of the Λ0 Hyperon”, Phys. Rev. 127 (1962) 2223–2230; L. Schachinger et al., “Precise
Measurement of the Λ0 Magnetic Moment”, Phys. Rev. Lett. 41 (1978) 1348–1351.
37 [Eds.] See §35.4 and §36.1.
38 [Eds.] But see note 5, p. 508.
39 [Eds.] See Example 3 in §35.4, p. 767.
40[Eds.] The relation (38.66) is known in the literature as the Coleman–Glashow mass formula: Sidney
Coleman and Sheldon Lee Glashow, “Electrodynamic Properties of Baryons in the Unitary Symmetry Scheme”,
Phys. Rev. Lett. 6 (1961) 423–425. The paper covers both the magnetic moments and the mass splittings of the
baryon octet. In the 1990 lectures, Coleman said at this point, “Modesty forbids me, but honesty compels me to tell
you that this was my first published paper.”
41 [Eds.] PDG 2016.
42 [Eds.]
This effect is second order, and not to be confused with the first order, diagonal mixing between Σ0 and Λ
which was discussed in the paragraph preceding (38.52).
43 [Eds.] See note 5, p. 508.
44 [Eds.] R. H. Dalitz, ”The ΛΛ-Hypernucleus and the Λ–Λ Interaction”, Physics Letters 5 (1963) 53–56; R. H.
Dalitz and G. Rajasekaran, “The Binding of ΛΛ-Hypernuclei”, Nucl. Phys. 50 (1964) 450–464; A. Gal, “The
Hypernuclear Physics Heritage of Dick Dalitz” in J. Pochodzalla and T. Walcher, Proceedings of the IXth
International Conference on Hypernuclear and Strange Particle Physics, Springer, 2007.
45 [Eds.] H. Mueller and J. Shepard, “Λ–Σ0 Mixing in Finite Nuclei”, J.Phys. G 26 (2000) 1049–1064.

39
Broken SU(3) and the naive quark model

Thus far we’ve talked about a world with perfect SU(3) symmetry, broken only by the effects of electromagnetism,
which we treated perturbatively. The mysterious medium-strong interactions (stronger than electromagnetism)
which break SU(3) symmetry were ignored. Now we’ll turn the tables, treating those medium-strong interactions
as a perturbative effect on top of the strong interactions while ignoring electromagnetism.1

39.1The Gell-Mann–Okubo mass formula derived

We have to start farther back. By hypothesis we assume that the Hamiltonian for the strong interactions is the sum
of a very strong part, invariant under SU(3), and a medium-strong part which breaks SU(3):

We don’t know nearly as much about the medium-strong interactions as we do about electromagnetism: the latter
is mediated by photons, and the requirement of renormalizability selected out minimal coupling. That enabled us
to show that the current transformed the same way as the charges, and much else. We have no such handle here.
If we’re to make any progress along similar lines, we have to make guesses, either about the dynamical theory of
the medium-strong interactions or about the predictions of such a dynamical theory insofar as pure symmetry
arguments go. We’d have to guess how the medium-strong interactions transform under SU(3). Such a venture is
not a priori guaranteed to be successful. It’s possible that SU(3) is a good symmetry of nature, but the medium-
strong interactions are very complicated. Perhaps these medium-strong interactions transform under SU(3) as
sums with more or less equal weights of pieces that behave like every conceivable SU(3) representation. In that
case we would not get a sum rule analogous to our electromagnetic ones, e.g., (38.66).

However, it turns out a postiori that the simplest guess anyone (and in particular, Gell-Mann) would have
been tempted to make, that the medium-strong interactions have simple transformation properties under SU(3),
fits experiment very well. Gell-Mann guessed2 that

H MS transforms under SU(3) like a member of (1, 1), the octet representation.

But which member? That is uniquely determined by the fact that the medium-strong interactions preserve isospin
and hypercharge. So it must be that

(the Λ-like member, if you will).3 I stress once again that this is pure hypothesis. If this guess did not work, it
wouldn’t necessarily mean that SU(3) is wrong. We’d try something else; maybe it transforms like a member of (2,
2) or the sum of members from (1, 1) and a (2, 2). However, this is certainly the simplest guess one could make
that would give baryon mass differences to lowest order in perturbation theory, and therefore it is the thing to try. If
it works, it is evidence for both Gell-Mann’s guess and the general idea of SU(3) symmetry. Historically, Gell-
Mann made this guess (39.2) before the η was discovered (by a couple of weeks).4 So it wasn’t a guess made
after the fact, but before.

We’re going to use this hypothesis and first order perturbation theory to compute the medium-strong mass
differences within SU(3) multiplets. Well, there’s a little finagle that’s traditionally used: We compute δm for
fermions but δm2 for bosons. It’s not really that important; it doesn’t matter whether we use δm or δm2 in first
order, because δm2 = 2m δm. If the splitting for m is small it wouldn’t matter if we obtain the relationships for the
shifts in m2 or the shifts in m. The splittings are not really that small. There seems to be some small improvement
obtained by using the rule of δm2 for bosons, so that’s what we’ll do. This finagle was inspired by field theoretic
ideas in which we think of corrections to the self-energy operator of a fermion making a shift in the mass, while
those to a boson’s self-energy shift the square of the mass.5

An amusing thing is that once you’ve made this hypothesis, you can write down the formulas for the mass
shifts within a general SU(3) multiplet (an octet, a decuplet, and so on) in closed form. We don’t have to multiply
matrices and tensors tediously for a given multiplet. I will count how many unknown constants there are for the
masses within any given multiplet. and construct that many operators.

Our task is to find out how many times (n, m) ⊗ (m, n) contains (1, 1). (Remember, (m, n) is the conjugate of
(n, m).) That will tell us how many ways there are to couple an octet to a and a ψ, and therefore how many terms
will be in our mass formula. I will prove to you that

The proof can exploit our algorithm for the reduction of a direct product, but it can also be shown directly, by
manipulating tensors. I’ll use the latter method.

Suppose I have some traceless symmetric tensor ψ j1 jmi1 in and its conjugate l1 ln
k 1 k m, the former
representing an object that transforms according to the representation (n, m), and the latter according to (m, n).
We want to sum up the indices on ψ in such a way that we are left with an octet; i.e., a traceless symmetric
tensor with one upper index and one lower index, . There are only two ways we can do it in general without
using the ϵ tensor. (We will come back later to use the ϵ tensor.)

We sum all of set A with all of set b, and all but one of set B with all but one of set a, leaving alone one upper index
and one lower index. Then I can take out the trace to make the octet. Alternatively, I could sum all of set B with all
of set a, and all but one of set A with all but one of set b, and take out the trace. Those are the two ways of making
an octet.

Using the ϵ tensor does me no good in this case. If I use only one ϵ tensor I’ll get two more lower indices than
upper indices, or vice versa, and there’s no way of summing that will leave us with one upper index and one lower
index. If we use two ϵ tensors that’s the same as the string of three δ’s in various permutations, and therefore the
same as the original thing we explored. Thus there are, at most, two ways of constructing an octet. If I have only
upper indices on one side and only lower indices on the other (i. e., if n or m is zero), one of the possibilities is no
possibility at all; there would only be one.

So, if I can construct two operators such that in any representation they are octets, or equivalently, octets plus
singlets (an additional singlet term is not going to bother us; that can be absorbed into the overall mass before I
turn on the medium-strong interactions), then I say that the mass splitting is proportional to the matrix elements of
those operators. I will now construct two such operators.

One of them is trivial. Remember our generator matrix, (38.44):

The ∗’s indicate strangeness-changing objects, which we don’t care about. This is a traceless symmetric matrix
that transforms like an octet. Therefore one operator whose matrix element would follow the octet rule would be
the I = Y = 0 component:

This is similar to the reasoning that establishes the Wigner–Eckart theorem.6 For example, the matrix element of
every vector operator is proportional to the matrix elements of the angular momentum: there’s only one way to
couple angular momentum 1 to a representation times its conjugate; therefore the same Clebsch–Gordan
coefficients must occur in the matrix elements of the total angular momentum as occur in the matrix elements of
the operator we are studying, and thus the two are proportional.

Likewise here there are two possible independent Clebsch–Gordan coefficients, so I need to find the two
possible octet operators. I have found one of them; I will now find the other. Given any matrix A, its cofactor matrix
cof[A] is also a matrix.7 The cofactor matrix is the matrix made up of the determinants of the minors that enter into
the expression for the inverse:

Since the A’s are operators we had better take the symmetric combination; the curly brackets indicate the
anticommutator. The object cof[A] transforms like a matrix if A does; in this case, like a combination of an octet
and a singlet as it’s not necessarily traceless. Therefore we will construct , the determinant on the minor of
the 33 element:

That is the second operator that transforms like the 33 element of an octet. All the other cofactors carry nonzero
isospin and/or hypercharge, and are ruled out by Gell-Mann’s guess.

We have arrived at the Gell-Mann–Okubo (GMO) mass formula8, a linear combination of these two
operators and a constant term:

The additive constant a is the mass present before the medium-strong interactions are turned on; a, b and c have
to be fitted to experiment within each supermultiplet. For representations where either n or m is zero, the b and c
terms are proportional; in that case, the GMO formula reduces to the first two terms alone:9

The original proposal was by Gell-Mann. He worked out the Clebsch–Gordanry only for the octet. Later Okubo
showed, by a different argument than the one here, how to write it for any representation. The cute argument I’ve
given is due to a Russian named Smorodinski ,10 who showed it to me at the Dubna conference in 1964.

39.2The Gell-Mann–Okubo mass formula applied

There are a lot of complete SU(3) representations around where all the particles have been discovered. Let’s
begin with the famous one, the baryon octet: the 8 with JP = +. The results are shown in Table 39.1 (masses in
MeV).

Table 39.1 Gell-Mann–Okubo mass splitting in the baryon octet

Thus we obtain the formula originally written down by Gell-Mann:

It gives one relationship among these baryon masses.

Now, what about experiment? Given the things we are neglecting, we expect the accuracy of this formula to
be second order in the medium-strong interactions. And although the medium-strong interactions are medium
strong, even their second order contributions are larger than electromagnetic effects, so we won’t bother to take
account of the electromagnetic mass shifts within a multiplet. We take the average over all the electromagnetic
masses. The uncertainties in these masses are negligible compared to the errors expected in this formula.11 With
the masses in MeV, we have

This is pretty good; the agreement is better than we would expect. Even the most enthusiastic SU(3) fan has to
admit that this agreement is fortuitous because, for heaven’s sake, the error is less than or on the order of a typical
electromagnetic mass splitting, and we cannot expect it to be that good.

The next multiplet we’ll look at among the fermions is the well-known JP = + resonance decuplet in
pion–nucleon scattering. In a perfectly SU(3)-symmetric world, that must occur as part of some representation of
the direct product of an octet times an octet, 8 ⊗ 8:

Now if all we knew about was the Δ, with I = , Y = 1, we could still say that there are only a small number of
possibilities here. It certainly can’t be in a singlet because that doesn’t contain any object with I = , nor do the two
octets; they contain (37.20) only I = {1, , 0}. It can’t be in the 10 because that contains an object with I = but it
has Y = −1, not +1. So the only possibilities are that the Δ is part of the 10 or part of the 27.

In the early days of SU(3), some people thought the Δ was in the 27. Glashow and Sakurai wrote a long paper
called “The 27-Fold Way”12 in which they gave a variety of convincing arguments about why it should be part of a
27. But it turns out in fact to be part of a 10. Its relevant energy range in all hypercharge channels has been
extensively explored. None of the other objects that would be there in a 27 fits the things that would be there in a
10.

There was a famous incident at the 1962 Rochester Conference.13 The Δ had been known, the Σ* (or
Σ(1385), in modern usage) had been discovered, and the discovery of the Ξ* (Ξ(1530)) was announced at the
conference. (Today we name the baryon resonances according to their mass, isospin, and hypercharge, instead
of giving them individual names like Σ* and K*, as we used to.)14 Gell-Mann looked at it and said “That’s a
decuplet.” He then predicted the Ω−, and gave its mass. We will see how well he predicted it.

Figure 39.1 The weight diagram for the baryon decuplet, with the predicted Ω−

The decuplet is given in Table 39.2; the masses are in MeV. The Δ (Δ(1232)) is a broad bump; where we put
the mass is a matter of taste. The Σ(1385) (formerly Σ*) was named after the familiar particle with the same I and
Y. The Ξ(1530) (or the Ξ*) is an excited state of the Ξ. As this multiplet is a decuplet, (3, 0), the GMO formula
simplifies to (39.10),
That means that the differences in mass should be proportional to Y, and a graph of M vs. Y should give a straight
line; see Figure 39.2.

According to (39.10), the mass splitting in the decuplet is proportional to the hypercharge difference. The first
three baryons are separated by ΔY = 1, and so we expect equal spacing in the mass splitting:

Table 39.2 The + baryon decuplet

Figure 39.2 Gell-Mann–Okubo mass splitting for the + decuplet

Thus Gell-Mann was able to predict the mass of the then-unknown tenth baryon as

The measured mass15 is 1672.45 ±0.29 MeV.

Actually we have two predictions, unlike the single prediction we had for the baryon octet: once we have one
difference, say between the masses of the Σ* and the Δ, we can predict the other two. You see that the agreement
is very good, considering that we are only going to the lowest order in the medium-strong interactions. Gell-Mann’s
guess seems to be holding up remarkably well, but it’s always dangerous policy to attempt to deduce deep
physics from something like that. We’ll look at another application which also works well. Then we’ll come to one
that doesn’t work, and there’s some interesting history attached to that.

Next, the pseudoscalar meson octet, JP = 0− (see Table 39.3). In this case we can just copy down the
baryon formula, except now the squared masses appear in the GMO formula because they’re bosons.16 The
analog of both the nucleon and the cascade (the Ξ) is the kaon. So we obtain, analogous to (39.11),
 Table 39.3 The pseudoscalar meson octet (plus one, the η′)

Writing it in its original form, predicting the η mass,

With mK = 496 MeV and mπ = 137 MeV, the η should be a little heavier than the kaon, and it is:17

Again this is better than we would expect: not the typical 20% error but about 5% error or less. (The η′ mixes only
a little bit with the η, because of their very different masses.) This is really surprising because the pion mass is so
far out of line with the other masses that it’s amazing the formula works at all. Indeed, one of the reasons people
spent four years exploring the dead end of global symmetry (§36.2) was because, as the product of three SU(2)’s,
it enabled us to put the pions in a representation of the product of the three SU(2)’s all by themselves, D (0,1,0) in
the notation of (36.33). Everyone said, “Well, obviously the eight baryons must be part of a multiplet. But the pions
and the kaons are so different in mass that we should be able to put the pions in a multiplet by themselves. It’s just
too preposterous to imagine there is a symmetry which puts the pions and the kaons together. They look too
different.”

Figure 39.3 The weight diagram for the pseudoscalar meson octet

I also remember, with pain at the time, and delight in retrospect, not long after Glashow and I published our
electromagnetic mass formula,18 there was a paper19 by Sakurai proposing a different theory of electromagnetic
masses. It contained the scathing footnote, “Formulas in the SU(3) symmetry scheme have recently been derived
by Coleman and Glashow. The reader should understand that these formulas are valid only in the approximation
(mK/mπ)2 = 10 = 1.” That was a well-taken criticism. Nevertheless, the formula works.

39.3The Gell-Mann–Okubo mass formula challenged

Let’s go on to the next octet discovered in those golden days of 1961, the vector bosons, JP = 1−. The ρ had just
been discovered. Those of you who think of the ρ as a permanent part of our universe should remember that the
original Rosenfeld tables contained a subtable labeled “Resonances” that had only one entry in it. There was a
time when the ρ was a great discovery, and it was followed soon after by the ω and the K*, in January of 1961 or
so. They were obviously an octet, although that turned out to be a little bit wrong, for reasons I’ll explain shortly.
We have the corresponding GMO formula for these mesons:
and we get the results shown in Table 39.4. The K* has the same quantum numbers as the kaon, except that it is a
vector particle.

Table 39.4 The vector meson octet

We can combine the results in the table and obtain the same formula (39.16) as for the η mass, (using the
average K* mass of 894 MeV) and we find (in GeV2)

This is not good. It’s a disaster, especially when compared with the results in the earlier cases. Well, Murray Gell-
Mann is not a man to give up an idea lightly. He said, at the time: Suppose that, just by chance, in the absence of
SU(3) breaking there happens to be a vector SU(3) singlet very close in mass to the ω. This is what we would
compute: 0.865 GeV2 would be the squared mass m82 of the octet if the singlet weren’t around. But because the
singlet is around, things are going to be different.

Let’s focus attention on the subspace of the big Hilbert space of the world which is a two-dimensional space
spanned by a state of the octet vector meson ω8, and the hypothetical singlet vector meson ω1. I’ll write down the
2 × 2 mass squared matrix for that. Gell-Mann said the ω1 will acquire some mass, both because it has some
mass in the absence of the medium-strong interactions, and because the medium-strong interactions may give us
some correction. The ω8 will acquire a mass, which we’ve just computed. But an octet operator can not only
connect octet with octet, it can connect octet with a singlet. And exactly the same interaction that puts in the
diagonal elements m12 and m82 could put in a cross-term, x:

(The phase between ω1 and ω8 is at our disposal so we can make x real; ordinarily we’d have x in one corner and
x* in the other.) The off-diagonal matrix element should be on the same order of magnitude as the diagonal matrix
elements. Therefore, if m1 and m8 were very widely separated, by say 0.5 or 0.6 GeV2, and x was on the order of
the things we’ve been computing, around 0.1 GeV2, then the effects of the off-diagonal term would be negligible.
On the other hand, if by some fluke m1 and m8 happen to be fairly close together, then this off-diagonal term can
have a very large effect. If we have two nearly degenerate levels in atomic physics, and we introduce a mixing
matrix element, a symmetry-breaking Hamiltonian that connects the two, then one of them is pushed up and the
other is pushed down, and the amount of the pushing can be comparable to the size of the mixing matrix element,
i.e., around 0.2 GeV or so on the scale we’re working with; see Figure 39.4. Gell-Mann hypothesized this was the
situation here, an instance of nearly degenerate perturbation theory. In (39.19) we see the ω8 mass lower than the
predicted result. The other one, ω1, presumably is higher. We can’t predict the exact amount of pushing up and
down unless we know both m12 and m82. But if Gell-Mann’s idea is right, we can say there must be an isosinglet
meson, JP = 1−, with a squared mass somewhere between 1.1 and 1.2 times greater than 0.865 GeV2.

Figure 39.4 Level shift of the ω1 and ω8

That was his prediction; not a precise number, but that there should be such a particle in nature. Among the
graduate students then at Caltech, this particle was known as the fudge-on.20And we were uniformly surprised
when, not long afterwards, the ϕ was discovered, with exactly the predicted properties, JP = 1−, and a squared
mass in the right ballpark:

Gell-Mann’s prediction was not quantitative; it could have differed by 10% one way or the other and it would still
have been all right. And in any case, there was the particle.

Now you might say there is no predictive power (other than qualitative) in this scheme, because we don’t
obtain a precise number from it. We have for this system three unknown quantities, m12, m82 and x. One of them
we know from the GMO formula, m82, but we don’t know a priori either m12 or x. We have two experimental
numbers, mω and mϕ , and therefore if we wanted to, we could deduce m12 and x. But that’s hardly a prediction;
we can’t go to God and ask, “What values did You assign to m12 and x?” to check it.

However, we can predict the so called mixing angle, which is, in a sense, experimentally measurable. The
mixing angle is the angle that tells us how much of the physical ω is ω1 and ω8 and how much of the physical ϕ is
ω1 and ω8. I will first go through the computation of the mixing angle, then I’ll explain how to measure it
experimentally.

If we stick to the two-dimensional Hilbert space (39.20), consisting of ω and ϕ at rest (with the spin degrees of
freedom suppressed), we would diagonalize the matrix and the eigenstates of M2 would represent the actual
physical particles, according to conventional, nearly degenerate perturbation theory. Since the M2 matrix is real,
its eigenvectors and eigenvalues must be real combinations of ω1 and ω8. Therefore we can define the angle θ by

(It turns out that ϕ is mostly ω8, so this choice minimizes the mixing angle.) That’s conventional perturbation
theory. We could determine the angle θ in terms of physically observable quantities, the masses. We could do it
by first finding m12 and x, but we can do it much more directly. Any matrix can be written in terms of its
eigenvectors and its eigenvalues:

Computing from the GMO as if there were no ω1,

We see that θ is closer to 0° than to 90°, so our choice was suitable: according to this the ϕ is mainly ω8 and the
ω is mainly ω1. In a moment I will explain why this number is useful.

Let’s apply this mixing theory to leptonic decays of the vector bosons.

The vectors do not often decay into e+e−, but such decays do occur. We understand electromagnetism very well,
so we know the kind of diagram that is responsible for this process:

A vector meson comes into a mysterious strong interaction blob (about which we can say nothing), a virtual
photon comes out and decays into an e+-e− pair. The amplitude is proportional to the matrix element of the
electromagnetic current between the vacuum and the one vector meson state, V. Once we peel off the
electron–positron pair the matrix element is what’s left.
 This is a very simple example of a one (virtual) photon process, in which we have to take the matrix element
of an octet operator between an octet state (if this is an octet vector meson), or a singlet state (if it is a singlet
vector meson). An octet operator cannot, in the limit of strict SU(3) symmetry, connect a singlet state to a singlet
state: both this singlet vector meson and the vacuum are SU(3) singlets. Therefore, in the SU(3)-symmetric world,
in which the eigenstates are ω1 and ω8, the ω1 decay amplitude vanishes:

To investigate the octet decays, we’ll introduce a matrix (as in (38.47) for the baryon octet) which we’ll call V
for vector meson. The amplitude should be proportional to

where ϵQ is the 3 × 3 matrix (38.46) for the electromagnetic current,

Of course the amplitude A is trivially zero for other than the two neutral members of the octet,21 ω8 and ρ0. Let’s
compute it for these two neutral mesons.

The ω8, assumed to be the I = 0 member of an octet, should behave like a Λ, so (38.47) its matrix Vω is

Multiplying the matrices and taking the trace,

The ρ0, the I = 1, Iz = 0 member of the vector meson octet, acts like its opposite number, π0, of the pseudoscalar
octet, so (37.35) its matrix is given by

and

Thus in a world with perfect SU(3) symmetry, we’d have

There would be a ρ0, an ω1 and an ω8. The ω1 would not decay into e+e− at all. The ω8 would decay into e+e−
and the ρ0 would decay into e+e− three times faster, because the decay rate is the square of the amplitude.

In the real world the SU(3) symmetry is not perfect. In this case because of the small energy denominators,
the mixing effects are much larger than other effects of symmetry breaking. The mass eigenstates are not ω1 and
ω8, but ϕ and ω. Taking account only of the mixing effect (not of the other supposedly small effects of SU(3)
symmetry breaking), which is anomalously large because ω1 and ω8 are close in mass, we find the prediction for
the branching ratios, compared here with the experimental numbers:22

I have not taken any account of phase space. We are not close to a threshold; the e+e− threshold is very low
compared to the masses of the ϕ, ω, and ρ, so phase space correction factors will be on the order of other effects
of the medium-strong interactions which we are systematically neglecting. And if we do include phase space, we
apparently make things worse. For example, we’d make the first prediction a little less than 0.3.
 The important point is that that these ratios are in the right ballpark, mostly due to the 3’s in the denominators.
That 3 is a pure SU(3) Clebsch–Gordan coefficient. If we didn’t have SU(3) it would be a priori as plausible to put
the 3 in the numerator as in the denominator, which would change our result by a factor of 9. Then the predictions
would be very different. So this is a non-trivial result. It’s not simply that they come out right qualitatively.

I’ve gone through many more predictions in the past two lectures than throughout the rest of the course. Just
because these arguments haven’t involved deep thoughts and field theoretic concepts that make your head feel
funny, don’t think this isn’t the real stuff. We’ve seen how all these were predicted:

•two magnetic moments in the baryon octet (µΣ and µΞ, Table 38.2, p. 837.)

•one electromagnetic mass splitting in the baryon octet (Coleman–Glashow, (38.66), p. 840.)

•one baryon octuplet medium-strong mass splitting (GMO, (39.12), p. 849.)

•two baryon decuplet medium-strong mass splittings (GMO, (39.13), p. 850.)

•one previously unknown member of the baryon decuplet (the Ω− from GMO, (39.14), p. 851.)

•one pseudoscalar meson medium-strong mass splitting (GMO, (39.17), p. 852.)

•one previously unknown vector meson singlet (the ϕ, to explain a deviation from GMO, (39.21), p.
854.)

•two ratios of electromagnetic decay rates (for the vector bosons, (39.35), p. 857.)

and they’ve all been borne out by experiment with good agreement. This concludes the numerical application and
the actual experimental checks that we are going to make of SU(3), although there are many others. The literature
is chock-a-block with them.

39.4The naive quark model (and how it grew)

I’m now going to talk about some SU(3)-related ideas. I’d like to say a few words, although I’m not going to make
any predictions or do any numerical work, about the famous quark model.23 It’s a real rags-to-riches story. The
quark model started out as universally scorned, but it was gradually accepted by even the most snooty of us. All
rags-to-riches tales establish a trajectory, sometimes with the hero ascending from poverty smoothly to success,
as in Horatio Alger’s novels, and sometimes his path is Dickensian, with reversals of fortune. And the quark model
may fall from grace once more. I will describe the basic ideas of the naive quark model from the viewpoint of SU(3)
symmetry, without talking about quark dynamics, which is a much more disputed subject. This is going to be more
informal and semi-popular in structure than what has come before. I’ll give you a sort of historical outline and
make comments occasionally about how SU(3) ideas come in. After discussing the naive quark model I will make
a few remarks about the gauge field theory of the strong interactions.

From the very first days of SU(3), and indeed even before the quark papers were written, people realized that
there might be some physics in the group theoretical statement that all SU(3) representations could be built out of
3’s and 3’s. The bosons, after all, came in octets and singlets and the fermions came in octets and decuplets, with
some singlets at higher energy. This suggested a composite model. In particular, there was something extremely
attractive in the formula

If there is a fundamental triplet and an anti-triplet, and if particles are bound states of these triplets, we get bosons
that come in octets and singlets only. Some people suggested that maybe those triplets were not just
mathematical figments of the imagination. Perhaps there really are particles that transform like SU(3) triplets:
funny particles with charges like and − and fractional strangeness and so on.24 And then the real mesons and
baryons, the octets and decuplets that we see, are bound states of these triplets. There was a wave when people
would investigate alternative triplet models. They said, “Well, if you give the triplet some sort of baryon number,
then of course this octet and singlet are going to have baryon number 0, a particle and an anti-particle. It’s fine for
the mesons but it’s no good for the baryons.” Other people said, “Well, maybe you need two kinds of triplets, one
that carries baryon number and one that doesn’t. Then you’d make a baryon by binding together a baryon
number-carrying triplet with a non-baryon number-carrying triplet.” And some people said, “Maybe you need a
particle around that’s a fundamental singlet that carries baryon number and you’d make a baryon by putting
together three of these things.”

The big discovery, which is rather trivial but important at the time, was made simultaneously by Gell-Mann25
and Zweig,26 who said, “Suppose we put together three (1, 0)’s. What do you get?” In an equation, the question is

Let’s work out what this is with our algorithm (§37.4). The first step (37.45) gives

and that is the only term we obtain; there is no possibility of contracting an inner or outer pair of indices. In the
second step (37.50), we get

or in terms of their dimensions,

We have to carry out the last product:

We already know (37.4) that (0, 1) ⊗ (1, 0) is (1, 1) ⊕ (0, 0). We have to compute the first product:

Again there is no possibility of peeling indices off the inner or outer pairs. The second step gives

so that finally

In common language,

Now things get very intriguing if we take this idea seriously—if we say that there are fundamental objects called
aces by Zweig and quarks by Gell-Mann.27 Gell-Mann’s name caught on and Zweig left the field.28 Hadrons are
made up of bound states of quarks, and if we say quarks have baryon number (why not, if they have charge
and hypercharge ), then we’re in pretty good shape. Things begin to look a little less artificial: mesons are
supposed to be quark–antiquark bound states, 3 ⊗ 3, octets and singlets. And sure enough, all the mesons that
have ever been discovered are octets and singlets. We’ve also got three quark bound states; they’re singlets and
octets and decuplets. Sure enough, all those objects have baryon number 1; their antiparticles are three bound
states with baryon number −1; those also go into singlets and octets and decuplets. No 27-plets or 35-plets or
other exotic objects around, and no 10’s with baryon number 1, although there are 10’s with baryon number −1,
their antiparticles.

So it looks as if there’s a glimmer of truth, maybe, in this idea. It seems to agree with the phenomenology of
the observed hadrons. On the other hand it also looks sort of silly. We say a nucleus in made up of neutrons and
protons. The way we establish that is not by studying nuclear structure, but by bashing a nucleus and watching
neutrons and protons fly out. At that time, and indeed ever since, no one has ever observed a free quark. No
matter how hard we hit a hadron, more hadrons fly out (mostly pions), but not any quarks. The reason for that is
perhaps the quarks are very heavy. Since the hadrons are not very heavy, the quarks must be very tightly bound;
then it’s an extreme relativistic system. The binding energy is comparable to the E = mc2 energy of the quarks.
Therefore we would expect that there would be a large probability for virtual quark–antiquark pairs. Why should it
be three quarks any more than five quarks, or seven, or 10? If we have all that energy around, why should it look
like a simple non-relativistic system? We’d expect the wave function to have a large amplitude for having all sorts
of pairs in it.

Also, what kind of force is it that can bind together a quark and an antiquark, or three quarks, but doesn’t bind
together two quarks? Why can’t we have a two-quark bound state? That would be a sextuplet plus an anti-triplet,
6 ⊕ 3, particles with fractional charge. Those multiplets would be every bit as easy to see as a quark itself. Why
aren’t they around?

So there was all this argument back and forth; nothing ever got anywhere. There were those who believed in
the quark model and those who didn’t. Those who believed in it used essentially non-relativistic reasoning to work
things out because they didn’t care what the kinematics were, it looked non-relativistic. And the results were sort
of good and sort of bad, and it was a big mess.29

The people pursuing the naive quark model went even further, saying, “If we can treat this as a non-
relativistic system, we can use good old non-relativistic quantum mechanics, including the ideas of spin-
independent forces, even though we know that’s ludicrous from the viewpoint of relativistic quantum mechanics.”
“We’ll just be very bold, ” they said, “and we’ll use very naïve reasoning, including the picture of things in an
attractive potential with spin-independent forces, to figure out not just the SU(3) assignments for these things, but
their spins.”

We’ll describe this work, beginning with the q system, the mesons. That’s sort of easy because there’s no
Pauli antisymmetrization to worry about. We’ll assume the quarks are SU(3) triplets with baryon number and spin
. That’s the simplest assumption if we’re going to wind up with baryon states with half-integral angular
momentum. The most tightly bound state will be an s-wave, as always for a central potential. Of course, as
previously established by SU(3) analysis, the product of a quark and an antiquark has to produce 8 ⊕ 1. These
mesons have to have parity −1, because they are made from fermions and antifermions.30 The spins can be
aligned to make spin 1 or spin 0: JP = 1− or 0−. That is, we should obtain vector bosons and pseudoscalar
mesons:

Of course the 0− and 1− multiplets don’t both have the same mass, but that’s presumably due to the spin-orbit
interaction. On a qualitative level it looks good; it looks like we’ve explained not only why the lightest bosons are
octets and singlets, but also why they’re pseudoscalars and vectors, not scalars and axial vectors, for example.31
Of course you would expect scalars and axial vectors to come up eventually; they would be p-states or d-states in
this imagined potential. But the lightest ones should be pseudoscalar and vector, and indeed they are. It looks
good. It also looks crazy, but by God it’s organizing the data correctly! So we go on. We don’t ask critical
questions: how can this happen or that happen? Because I don’t know the answer; I’m just trying to explain things.

39.5What can you build out of three quarks?

Let’s go on to the system of three bound quarks, qqq. Here again we expect some sort of interactions, so the
lightest bound state we would expect with all three quarks as close to each other as can be with no centrifugal
barrier, any pair in an s-wave, the space part of the wave function being totally symmetric. This means the
SU(2)spin ⊗ SU(3) part must be antisymmetric, because the quarks are supposed to be fermions. But this doesn’t
work. Therefore the quark modelers, every last one of them utterly mad, said: “Well, let’s try the alternative
hypothesis: treat the quarks as if they were bosons.” In fact they aren’t bosons, if they exist at all. A little later on
we’ll see how that difficulty is actually resolved, with a much better solution than bosonic quarks.

Let us work out what happens if one assumes that the SU(2)spin ⊗ SU(3) part of the qqq wave function is
totally symmetric. (That’s a nice little group theory exercise, whether or not you believe this nonsense about
bosonic quarks.) Given a three-particle system of fermions, we want to work out how we put spin and SU(3) wave
functions together to make something that is symmetric under permutations of the q’s. (For the moment, we are
ignoring the Pauli principle. We’ll come back to it.) This requires a small group theoretical digression on the
permutation group on three objects, called S3 by its friends.32

A brief digression on the group S3

S3 is a finite group with 3! = 6 elements, because there are 3! permutations on three objects:
The top row is the initial arrangement of the three objects, and the bottom is the final arrangement. The first group
element is the identity (that is, 1 → 1, 2 → 2, and 3 → 3), which we can write as I. The second element leaves 1
alone and swaps 2 with 3: it is more conveniently written as (23). Similarly the other elements are written, in order,
as (1 2), (1 3), (1 2 3) (that is, 1 → 2, 2 → 3, and 3 → 1) and finally (1 3 2):

The permutation (2 3) is called odd, because the number of steps (moving a bottom number to the left or right by
one position) needed to bring 132 back to the arrangement 123 is one, an odd number. The second, third and
fourth elements are all odd; the first, fifth and last, requiring zero and two steps, respectively, are even. If we only
have two objects and we were flipping them around, then we would know there would be two irreducible
representations of the group, either symmetric (the trivial representation), or antisymmetric. With three objects we
can get only one other irreducible representation that is called the mixed representation,33 and that has dimension
2. I summarize these in Table 39.5. Note that the number of the elements of the IR is given by the square of its
dimension. The action of the representation (s) is to multiply a group element by 1. The representation (a)
multiplies the group element by +1 if the permutation is even, and by −1 if odd. And then there is a two
dimensional “mixed” representation, (m). An elementary result in the theory of finite groups states34

where dr is the number of rows or columns in the rth irreducible representation’s (square) matrix, N(G) is the order
of the group G (how many elements it has), and the sum is over all the irreducible representations of G. It’s easy to
check this equation for S3:

I can’t take the time to work out the theory of finite groups, but I will try to show at least that there is a two-
dimensional irreducible representation of S3.

Let us consider a three-dimensional vector space with axes labeled x, y and z, Figure 39.5. This space forms
a representation in particular of the permutation group on three objects, just permuting x, y and z. There is
obviously an invariant one-dimensional subspace, not under the full group of rotations but under the group of
permutations, which is

Table 39.5 The type and dimensions of the IR’s of S3

Figure 39.5 Axes for representations of S3

That’s a one-dimensional invariant subspace of the three-dimensional space. It forms a basis for the trivial
symmetric representation of the group.
 Figure 39.6 Invariant two-dimensional subspace for the mixed representations of S3

Now let us consider all vectors orthogonal to this. That’s a two-dimensional subspace, the plane passing
through the origin. It’s hard to draw, so I’ll sketch Figure 39.6 instead; this is a plane parallel to the one we mean
and that intersects the x, y and z axes at 1; it is displaced from the one we want. This two-dimensional subspace
cannot be split into two invariant one-dimensional subspaces; we can’t reduce things any further. How could we
have a vector in this plane that was invariant under permutations? To be invariant under permutations it has to
point just as much in the x-direction and the y-direction as in the z-direction. There’s no way to do that here: it’s like
a Mandelstam plot35—in what direction will such a vector point? The two-dimensional vectors on this subspace
form an irreducible representation of the group: we can’t break it down further into two one-dimensional
subspaces. That is the mixed representation.

To build states out of three quarks, we have to consider products of the representations of SU(3) and the
representations of SU(2)spin. Because we are proceeding under the foolish assumption that the wave functions
are symmetric, we are really only interested in the symmetric products. We can work out what these
representations give when multiplied together by considering the products of representations of S3. For example,

and so on. We know that (s) times anything will give us that same anything, and (a) times (a) is (s). We get that (a)
⊗ (m) has to give us a two-dimensional irreducible representation, and there’s only one around, namely (m) itself.
There are six independent products (because the ⊗ is symmetric), and we can represent the products neatly in a
product table, Table 39.6.

Table 39.6 The ⊗ table of the IR’s of S3

The last product, (m) ⊗ (m), you have to take on trust. But I can make it plausible. You can see that the
dimensions check out: 2 × 2 = 1 + 1 + 2. We can work out (m) ⊗ (m) from the fact that if (a) ⊗ (m) contains (m) then
(m) ⊗ (m) must contain (a). (This is the famous permutation symmetry of Clebsch–Gordan coefficients.36)
Similarly, (m) ⊗ (m) must contain (s) because (s) ⊗ (m) contains (m); and it’s got to have something left over; it
can’t be an (s) and it can’t be an (a), there’s only one of each, so it has to be (m). That’s all the background we
need on S3.

If we are going to make baryons out of three quarks, we’ve got to put together their three spins and their three
SU(3) quantities. (I remind you that we are proceeding under the ridiculous assumption that the SU(2)spin ⊗ SU(3)
part of the qqq wave function is symmetric.) I’ll start with the spins.37

The state where all the spins are aligned is symmetric under the permutation group. On the other hand, the
states and get mixed up with each other when we permute the three spins, so together they form a mixed
representation of the permutation group, the sum of two two-dimensional representations of the rotation group.
This direct sum forms a single irreducible representation of the direct product of the rotation group and the
permutation group, SU(2)spin ⊗ S3.

Now for the SU(3) contents of the products of the three quarks:

What are their symmetries under S3? If we think of each of the 3’s as a vector with one index (3 = (1, 0)) then the
three-index tensor that transforms like a 10 is the one that is totally symmetric under permutations of the three
indices. That’s our irreducible representation (3, 0), totally symmetric under permutation of the three upper
indices.38 So the 10 is guaranteed to be symmetric, by construction. What is the singlet? There is an object that is
SU(3)-invariant that has three indices, ϵijk. Is there an object totally antisymmetric under permutation of the three
indices? Yes, the same ϵijk. So the singlet (ϵijkqiqjqk ) is antisymmetric. The 8’s get flipped among themselves.
They’re mixed; you can take it on trust or construct the wave functions.

Let’s see what we can put together from (39.52) and (39.53) to make a totally symmetric wave function.
There’s nothing in (39.52) that we can combine with the 1 (the (a) representation in (39.53)) to make a symmetric
wave function: (a)SU(3) with either (s)spin or (m)spin makes an antisymmetric wave function according to our times
table, and we’re imagining that only the symmetric ones work. The antisymmetric singlet in (39.53) is ruled out.
There are no singlets. And that’s right! There are no low-lying baryon singlets.

We can put together a mixed object with a mixed object (i.e., two elements from the (m) representation) with
appropriate Clebsch–Gordan coefficients to make a symmetric object. Therefore we can have an octet with JP =
+:

It’s assumed to be + because it’s three identical particles in a totally symmetric spatial wave function. That’s right!
We have such an octet, our old friends {N, Σ, Ξ, Λ} in Table 36.1, p. 782. And we can put together the symmetric
spin states (39.52), J = , with the symmetric SU(3) states (39.52) in 10, to make a decuplet with JP = +:

That fits experiment also! It’s Gell-Mann’s decuplet (Table 39.2, p. 851). You can’t make a decuplet with JP = +

nor an octet with JP = + , if the wave functions are to be symmetric. All you can make is a decuplet with JP = +

and an octet with JP = + . Isn’t that peculiar? This is why, as I said a little earlier, the SU(2)spin ⊗ SU(3) part can’t
be antisymmetric: it simply doesn’t fit the observed spectrum. (Of course we could make them in this model in
excited states where the spatial wave function is not totally symmetric; one of the two quark pairs is in a p-wave or
something. We’d expect that to have higher energy.) But good heavens, this is wonderful and nutty at the same
time. It’s wonderful because when we put together wave functions for states describing three fermionic quarks, we
get just the right spectrum. And it’s nutty because these three-quark wave functions have to be totally symmetric,
in violation of the Pauli principle, if (as we believe) the quarks are fermions. That’s impossible with only the quark
characteristics we’ve considered thus far, at least for the lowest energy, s-wave states. (And if these are not the
lowest energy states, then where are they? Thus far none lower have been observed.) Somehow a way has to be
found around the Pauli principle and the notion of fermionic quarks to allow us these symmetric wave functions.39

Now this is an exciting moment so I will tell you the answer we think we have. Here “we” means everyone; it’s
a universally accepted idea. It began as a mad speculation and became established dogma without ever passing
through the test of experiment. The idea is a new quantum number called color. We use it to explain two curious
facts—not the fact that a non-relativistic model works well, but two other facts: Why should the three quark wave
function be totally symmetric in SU(3) ⊗ SU(2)spin; and why should there be quark–antiquark and three quark
bound states, but nothing else, in particular, no quark–quark bound states? And maybe it could explain a third
curious fact: if non-relativistic reasoning is good, then why can’t we knock quarks out of the hadrons? Why do we
never see a free quark? The idea is this. We’ve going to increase the number of quarks by giving them this new
quantum number, color. Formerly there were three quarks, now there will be nine: each of {u, d, s} will come in
three colors. We’ll see that more quarks can solve a lot of problems. I will sketch out what color is, and some of its
consequences.40
 39.6A sketch of quantum chromodynamics

The actual symmetry group GS of the strong interactions is going to be SU(3) ⊗ SU(3). The quarks are going to be
triplets under both of these groups: qiα . The i = {1, 2, 3} is for the first SU(3) that we know and love, and α = {1, 2,
3} is for the mysterious second SU(3). The first SU(3), the up, down and strange, is called flavor (a joke of
Nambu’s)41; this is our old friend, the quark type: {u, d, s}. The second SU(3) goes with the new quantum number,
color: red, green, and blue.42

Now, particles with color have never been seen. That’s led to a hypothesis that there are very, very strong
interactions between colored particles so that the only physically observable states are color singlets. That is, the
observable baryons and mesons are “colorless”, and transform trivially, as scalars under SU(3)color. This is
sometimes called color confinement. We’ve learned how to do this, though there is some dispute.

We come to a second idea, based on color, bearing the elegant name of quantum chromodynamics, or
QCD for short.43 It’s another wonderful name from Caltech. There is always something new out of Pasadena, to
paraphrase Pliny,44 and always a joke. The concept of a quantum field theory arising from color came from
Nambu and Gell-Mann originally, but many others, notably Ken Wilson, Lenny Susskind, David Politzer, Frank
Wilczek, and David Gross, made significant contributions to the theory early on.45

The general idea is to treat the force associated with color SU(3) just as we treated electromagnetism. In
QED, electromagnetic force is mediated between charged particles by the exchange of photons coupled to
charge. In much the same way, in QCD a force will be mediated between colored particles—in particular, between
quarks—by the exchange of “color photons” coupled to color. As we will see (Chapter 47), there is a procedure for
obtaining the field theory corresponding to a given symmetry group by making that group a local symmetry. The
local symmetry group is called a gauge group; SU(3)color is the unbroken gauge group of QCD.46 Each generator
of the gauge group is associated with one massless vector boson, called a gauge boson. That means we have
eight massless vector bosons for the eight generators of SU(3)color. But it’s not like electromagnetism in that the
force is very strong. A necessary consequence of this is that color non-singlet states will be much heavier than
color singlet states, because color non-singlet states, having a non-zero color, will have non-zero values of these
gauge fields at infinity. Therefore when we add up the energy that’s stored in the gauge fields, outside the particle,
we get a positive answer. On the other hand, if we have a color singlet, the gauge fields vanish at infinity; then at
least we don’t have that positive contribution. Whether you can make color non-singlet states infinitely heavier
than color singlet states is an open question. It’s much easier to answer the question in two dimensions, and there
the answer is “yes”.47

In QCD, the eight gauge bosons—the “color photons”—are also colored, because SU(3)color is a non-Abelian
group. These gauge bosons have zero bare mass but they might acquire a mass through the color force; they are
coupled to each other. Everything has a funny name here.48 The gauge bosons are called gluons because they
glue things together. The gluons transform like the octet representation of the group, like its generators. And just
as we can’t observe free quarks because they are not color singlets, we cannot observe gluons. We could of
course observe bound states made up of gluons just like we could observer bound states make up of quarks;
these are called glueballs. But at the moment there is no definite meson that has been identified as a glueball;
quarks seem to be sufficient. 49

Let’s see how SU(3)color solves our problems. First, although we can make a color singlet out of a quark and
an antiquark, 3color ⊗ 3color (37.17), we can’t make a color singlet out of two quarks (39.40):

so it explains why there are no two quark bound states. Indeed, no bound states of two quarks and an antiquark,
or two antiquarks and a quark, by similar computations. We simply can’t make a color singlet this way. But (39.45)

So we can put together three quarks to make a singlet, and it is antisymmetric in color (39.53). The generalized
Pauli principle is that the wave function has to be antisymmetric in the product of all the degrees of freedom:50
By our rule that it has to be a color singlet, it is forced to be antisymmetric in color: the 1 is antisymmetric. By the
ideas of the naive quark model, the space part of the wave function is symmetric, and therefore SU(2)spin ⊗
SU(3)flavor is symmetric also. So the whole thing will be antisymmetric, and there’s no violation of the Pauli
principle, as there would be without the new quantum number, color. This new quantum number answers two
questions, but as yet it has not really answered the third: why can’t we knock quarks out of hadrons? People
believe that quark confinement can be shown from QCD (it has been shown51 in lattice QCD), but it’s an open
question.52

Here are some facts about quantum chromodynamics, stated without proof.53

•Color is a gauge symmetry. There is an octet of massless vector bosons called gluons, each a
Yang–Mills field associated with color.

•Color is an unbroken symmetry, because if a gauge symmetry is broken all hell breaks loose.54

•Color singlets are favored, for the same reason that electromagnetic charge neutrality is favored: we
get electric fields if positive and negative electric charges are separated.

•q and are connected by a flux tube. The quark–antiquark potential grows linearly with distance.
(Any non-Abelian gauge theory of vector fields has this property.)

Figure 39.7 Electric dipole and quark–antiquark flux tube

Figure 39.8 Pulling a quark and antiquark apart makes another meson

Let’s contrast an electric dipole with a quark–antiquark system, as in Figure 39.7. Look at a π meson, made of a
q pair. We go to the hardware store and buy a couple of quark hooks to pull the quark and the antiquark apart.
This puts in energy just like separating charges puts energy into the electric field. Eventually there is enough
energy stored between the quark and the antiquark to make a new quark–antiquark pair ′q′ between the
original pair; we wind up with two mesons, q′ and ′q. It’s like breaking a string: if we want a one-ended
piece of rope, we could tie one end of a two-ended piece of rope to something that won’t move, and pull the other
end until it snaps, as in Figure 39.8. But we do not get a one-ended piece of rope; we get two ropes each with two
ends. It doesn’t work.

Next time we will begin the investigation of current algebra.

1 [Eds.] At the beginning of his lectures, Coleman always asked if there were any questions. In the video of Lecture
39, a student asks about the field theory involved in these SU(3) predictions. Coleman gives a lengthy answer, in
the end admitting that there isn’t much field theory involved. With a smile, he asks the student, “Why did you ask
that question? Was it an implicit criticism, ‘What are these lectures doing in a course on quantum field theory?’
This is a course on relativistic quantum mechanics, ” echoing his first sentence in the first lecture.
2 [Eds.] M. Gell-Mann, “Model of the Strong Couplings”, Phys. Rev. 106 (1957) 1296–1299; M. Gell-Mann, “The
Eightfold Way”, Caltech Synchrotron Radiation Laboratory report CTSL-20, 1961 (unpublished); reprinted in The
Eightfold Way, M. Gell-Mann and Y. Ne’eman, Benjamin, 1964.
3 [Eds.] See (37.22).
4 [Eds.] A. Pevsner et al., “Evidence for a Three-Pion Resonance Near 550 MeV”, Phys. Rev. Lett. 7 (1961)
421–423.
5 [Eds.] In 1990, Coleman gave Feynman credit for the idea.
6 [Eds.] E. Merzbacher, Quantum Mechanics, 3rd ed., John Wiley, 1998, Chap. 17, pp. 432–437; Arfken & Weber
MMP, p. 273.
7 [Eds.] Arfken & Weber MMP, p. 168.
8[Eds.] M. Gell-Mann, op. cit. (the formula appears as equation (4.8) in the Caltech report); S. Okubo, “Note on
Unitary Symmetry in Strong Interactions”, Prog. Theo. Phys. 27 (1962) 949–966; (reprinted in Gell-Mann and
Ne’eman, op. cit. ); S. Okubo, “Note on Unitary Symmetry in Strong Interaction II: Excited States of Baryons”,
Prog. Theo. Phys. 28 (1962) 24–32.
9 [Eds.] For representations for which n = 0 or m = 0, I(I + 1) − Y2 = AY + B. For m = 0, A = + n, and B = n( n +
1). For the 10 = (3, 0), we have (I + )2 = (Y + 3)2, or I = Y + 1. For n = 0, A → −A and n → m. See J. McL.
Emmerson, Symmetry Principles in Particle Physics, Oxford U. P., 1970, pp. 121–123. Note that there is an
overall sign error in equation (5.12).
10 [Eds.] Yakov A. Smorodinski (1917–1992), Russian mathematical physicist coauthor with Lev Landau of
Lectures on Nuclear Theory, Dover Publications, 2011. See M. Shifman, ed., Under the Spell of Landau, World
Scientific, 2013, Chapter 4.
11 [Eds.] PDG 2016.
12 [Eds.] S. Glashow and J. Sakurai, “The 27-Fold Way and Other Ways: Symmetries of Meson-Baryon
Resonances”, Nuovo Cim. 25 (1962) 337-354.
13 [Eds.] The 11th International Conference on High Energy Physics (Rochester Conference), July 1962, was held
at CERN. According to Crease & Mann SC, pp. 273–274, Gell-Mann predicted the Ω− would have a mass of 1685
MeV. It was discovered in early 1964, at Brookhaven, within 1% of Gell-Mann’s estimate: V. E. Barnes et al.,
“Observation of a Hyperon with Strangeness Minus Three”, Phys. Rev. Lett. 12 (1964) 204–206. After this
discovery, “there was no doubt, SU(3) was in.” A. Pais, Inward Bound, Oxford U. P. 1986, p. 557.
14 [Eds.] PDG 2016.
15 [Eds.] PDG 2016, p. 96.
16 [Eds.] The coefficient b in the general form (39.9) is 0, because the K has both Y = 1 and Y = −1. The same is
true for the vector meson octet in Table 39.4; see (39.18).
17 [Eds.] PDG 2016.
18 [Eds.] See note 40, p. 841.
19 [Eds.] J. J. Sakurai, “New Resonances and Strong Interaction Symmetry”, Phys. Rev. Lett. 7 (1961) 428–428.
Note 15 reads in part, “It is easy to see, however, that most statements made in [Coleman and Glashow’s] paper
are expected to be accurate only up to a factor (mK/mπ)2 ≈ 13.”
20 [Eds.] For readers unfamiliar with the term, a fudge factor is an ad hoc quantity introduced into a calculation or
measurement, ostensibly to account for error, to bring the number obtained closer to a desired value.
21 [Eds.] The octet meson decays into an e+-e− pair, so it must be neutral.
22 [Eds.]F. Nichitiu, “Introduction to the Vector Meson”, Laboratori Nazionali de Frascati publication, LNF-95/056
(1995); available from www.iaea.org; PDG 2016.
23 [Eds.] F. Halzen and A. D. Martin, Quarks and Leptons: An Introductory Course in Modern Particle Physics,
John Wiley, 1984; Griffiths EP ; Harry J. Lipkin, “Quarks for Pedestrians”, Phys. Lett. C8 (Physics Reports ) (1973)
173–268.
24 [Eds.] Crease & Mann SC, Chapter 15, “The King and his Quarks”, pp. 280–308.
25 [Eds.] M. Gell-Mann, “A Schematic Model of Baryons and Mesons”, Phys. Lett. 8, 214-215 (1964). This is the
first use of the term “quark” in the literature (see note 6 in the article).
26 [Eds.] G. Zweig, “An SU(3) Model for Strong Interaction Symmetry and its Breaking”, CERN Report 8182/Th
401 (1964), unpublished. Version 2 reprinted in Developments in the Quark Theory of Hadrons, eds. D.
Lichtenberg and S. Rosen, Hadronic Press, Nonantum, MA, 1981, Vol. 1, pp. 22–101; also available as CERN-
TH-412 from the CERN document server, cds.cern.ch.
27 [Eds.] The term “quark” comes from James Joyce’s infamously difficult Finnegans Wake, a novel unlikely to
have been read by many other physicists besides Gell-Mann when he appropriated the term. See the reference in
note 25, p. 858.
28 [Eds.] According to Crease & Mann SC, p. 285, both men submitted their epochal articles to the same journal
(CERN’s Physics Letters), but Zweig’s was rejected: “It was all right for someone of Gell-Mann’s stature to
advocate the lunatic notion that most of matter was made up of ineffable entities that were invisible to experiment;
having no reputation to protect him, Zweig was denied an appointment at a major university because the head of
the department thought he was a ‘charlatan’.”
29 [Eds.] Coleman adds, “It still is a big mess in many respects. We’re starting to get an inkling of how to
incorporate these ideas into a relativistic quantum field theory. That we must save for your post-Physics 253
studies. I’m just giving you an overview at this moment.” He was speaking in 1976, already three years after the
pioneering work of David Politzer, David Gross, and Frank Wilczek established that quantum chromodynamics is
asymptotically free: H. D. Politzer, “Reliable Perturbative Results for Strong Interactions?”, Phys. Rev. Lett. 30
(1973) 1346–1349; D. J. Gross, and F. Wilczek, “Ultraviolet Behavior of Non-Abelian Gauge Theories”, Phys. Rev.
Lett. 30 (1973) 1343–1346. The three men shared the 2004 Nobel Prize for this work; Politzer cited “Sidney
Coleman, my beloved teacher from graduate school”, in his acceptance speech: H. David Politzer, “The Dilemma
of Attribution”, on-line at https://www.nobelprize.org/nobel_prizes/physics/laureates/2004/politzer-lecture.pdf; the citation is the
first sentence of the second paragraph.
30 [Eds.] The quark q and the antiquark have opposite parities, and the parity of the meson is the product of
these, which has to be −1: Pmeson = PqP = −1.See note 2, p. 460.
31 [Eds.] If the quarks had spin 0, the lightest particles would have JP = 0+ and the next lightest would be JP = 1−,
i.e., scalars and axial vectors.
32[Eds.] J. Mathews and R. Walker, Mathematical Methods of Physics, 2nd ed., Addison-Wesley, 1969, pp.
425–433.
33 [Eds.] For finite groups, there are only as many irreducible representations as conjugacy classes (two elements
a, b are conjugate under a similarity transformation; a → gag−1 = b). There are three such for S3: under
conjugation the identity turns into the identity; a two-cycle turns into a two-cycle, and a three-cycle turns into a
three-cycle. So there are only three irreducible representations: (s) (every element represented by 1), (a) (even
elements represented by 1, and odd elements by −1), and (m), (each element is represented by a 2 × 2 matrix).
34 [Eds.]K. Riley and M. Hobson, Mathematical Methods for Physics and Engineering, 2nd ed., Cambridge U. P.,
Section 25.7.1; Zee GTN, pp. 104–108.
35 [Eds.] See Figure 11.5, p. 233.
36 [Eds.]
A. Messiah, Quantum Mechanics, Volume 2, North Holland Publishing, 1962, Appendix C; reprinted (in a
combined edition of both volumes together) by Dover Publications, 2017.
37 [Eds.] L. Schiff, Quantum Mechanics, McGraw-Hill, New York, 1968, pp. 218–225.
38 [Eds.] Recall that our tensors are totally symmetric within both the upper indices and the lower indices. See
(36.57).
39[Eds.] The construction of the properly antisymmetrized quark wave functions, including color, is intricate and
nontrivial for baryons: Griffiths EP, Section 5.6.1, pp. 181–188; see also Problem 21.3, p. 871.
40 [Eds.] The idea of a new quantum number for quarks is due independently to Greenberg (who called it
“parastatistics”), and Han and Nambu: O. W. Greenberg, “Spin and Unitary-Spin Independence in a Paraquark
Model of Baryons and Mesons”, Phys. Rev. Lett. 13 (1964) 598–602; M. Y. Han and Y. Nambu, “Three-Triplet
Model with Double SU(3) Symmetry”, Phys. Rev. 139 (1965) B1006–1010. The new quantum number was given
the name “color” by Gell-Mann; M. Gell-Mann, “Quarks”, Acta Phys. Austriaca Suppl. 9 (1972) 733–761. The term
is introduced on p. 736.
41 [Eds.] Yōichirō Nambu (1921-2015) was a Japanese-American physicist at the University of Chicago for sixty
years. He had been a student of Tomonaga’s in Tokyo during and immediately after World War II. He made many
groundbreaking contributions to condensed matter and particle theory, work recognized by the Physics Nobel
Prize in 2008 for his research into spontaneous symmetry breaking. Bruno Zumino famously joked that he once
had the idea to talk to Nambu to get ten years ahead of the crowd, “but by the time I figured out what he said, ten
years had passed.” M. Mukerjee, “Profile: Yoichiro Nambu”, Sci. Amer. Feb. 1995, pp. 37–39.
42 [Eds.] Coleman used red, white and blue in this lecture. Perhaps to avoid nationalist controversies, the
community switched to the primary colors of light not long after QCD became well-known.
43 [Eds.]
Peskin & Schroeder QFT; P. H. Frampton, Gauge Field Theories, 3rd enlarged and improved ed., Wiley-
VCH, 2008; Greiner et.al QCD; Chris Quigg, “Gauge Theories of the Strong, Weak and Electromagnetic
Interactions”, Benjamin-Cummings, 1983, Chapter 8, pp. 193–268. The Greek word for “color” is , “chr ma”.
According to David Gross, the first appearance of the term “quantum chromodynamics” in the literature was in the
review article by W. Marciano and H. Pagels, “Quantum Chromodynamics”, Phys. Reps. 36C (1978) 137–276; the
authors ascribe the name to Murray Gell-Mann: David Gross, “Asymptotic Freedom and the Emergence of QCD”,
in The Rise of the Standard Model: Particle Physics in the 1960s and 1970s, eds. Lillian Hoddeson, Laurie M.
Brown, Michael Riordan, and Max Dresden, Cambridge U. P., 1997, pp. 199–233. Note however that Susskind’s
1976 Les Houches lectures (see note 45, p. 867) antedate the Marciano and Pagels review article by two years.
The Oxford English Dictionary credits Dietrick E. Thomsen, “Chromodynamics”, Science News 109, June 26,
1976, 408–409 with the first print citation.
44 [Eds.] Pliny’s text reads unde etiam vulgare Graciae dictum semper aliquid novi Africam adferre (whence the
proverb, even common in Greece, that “Africa is always bringing forth something new”), Gaius Plinius Secundus
(CE 23–79), Historia Naturalis, book 8, section 42; Natural History, v. III: Books 8–11, ed. H. Rackham, (dual
language edition) Loeb Classical Library, Harvard U. P., 1940. Often quoted from the Adagia (1500) of Erasmus of
Rotterdam, Ex Africa semper aliquid novi (Out of Africa, always something new). Naturalist, scholar and
statesman, Pliny the Elder died while successfully rescuing a friend and her family during the Vesuvius eruption
that destroyed Pompeii and Herculaneum, according to his nephew Pliny the Younger, who witnessed the events.
For a history of the phrase, see Harvey M. Feinberg and Joseph B. Sodolow, “Out of Africa”, Jour. Afric. Hist. 43
(2002) 255–261.
45 [Eds.] Two useful references are Crease & Mann SC, pp. 296–299 and pp. 327–336, and Close IP, pp.
257–279. Historically, the development of SU(3)color seems to have had three separate phases. If quarks were
physical entities and fermionic, some way had to be found to deal with the Pauli principle, which seemed at odds
with parts of the baryon spectrum: Greenberg, op. cit.; Han and Nambu, op. cit. Then, once color’s value had been
established (e.g., multiplying some theoretical expressions by factors of three, vastly improving agreement with
experiment), it was quickly realized that these color charges could serve as a source of a new interaction, which
might in turn bind the quarks together: Y. Nambu, “A Systematics of Hadrons in Subnuclear Physics”, in Preludes
in Theoretical Physics in Honor of V. F. Weisskopf, eds. A. de-Shalit, H. Feshbach, and L. Van Hove, North-
Holland, 1966, p. 133–142; M. Gell-Mann, “Quarks”, Acta Phys. Austriaca Suppl. 9 (1972) 733–761; H. Fritzsch
and M. Gell-Mann, Proc. XVI Intern. Conf. on High Energy Physics, Chicago, 1972, v. 2, p. 135–165, available as
hep-ph/0208010v1 at https://arXiv.org; W. A. Bardeen, H. Fritzsch, and M. Gell-Mann, in Scale and Conformal Symmetry
in Hadron Physics, ed. R. Gatto, Wiley, 1973, p. 139–151; H. Fritzsch, M. Gell-Mann and H. Leutwyler,
“Advantages of the Color Octet Gluon Picture”, Phys. Lett. 47B (1973) 365–368. Finally, the consequences of the
Yang–Mills nature of QCD began to be investigated: the work of Politzer and Gross & Wilczek on asymptotic
freedom cited earlier (note 29, p. 860) and H. D. Politzer, “Asymptotic Freedom: An Approach to Strong
Interactions”, Phys. Lett. 14C (1974) 129–180; K. G. Wilson, “Confinement of Quarks”, Phys. Rev. D10 (1974)
2445–2459; L. Susskind, “Coarse Grained Quantum Chromodynamics”, pp. 208–308 in Weak and
Electromagnetic Interactions at High Energies: Les Houches 1976, R. Balian and C. H. Llewellyn Smith, eds.,
North-Holland Press, 1977. Quantum chromodynamics is now a firmly established theory, supported by an
immense body of experimental and theoretical results.
46 [Eds.] See note 5, p. 646, and note 54, p. 869.
47 [Eds.] “Charge Shielding and Quark Confinement in the Massive Schwinger Model”, Sidney Coleman, R.
Jackiw, and Leonard Susskind, Ann. Phys. 93 (1975) 267–275.
48 [Eds.] Coleman adds, “I love everything about this subject except its nomenclature, which makes me feel like
Bozo the Clown whenever I deliver a lecture on it.”
49[Eds.] Coleman adds, “This is a guided tour of the land of mists and fogs. The castle in the distance may turn
out to be a mere fata morgana. We should have at least one lecture like that in this course, off the high road and
into the swamps…” That was a fair description in 1976. Forty years later, the castle of QCD is no longer distant. It
is no mirage, but a stately fortress, beautifully built from the finest granite and marble.
50 [Eds.] See note 39, p. 866.
51 [Eds.] Wilson, op. cit.
52 [Eds.] The Clay Mathematics Institute has offered a substantial cash prize for the demonstration of a closely
related result: http://www.claymath.org/millennium-problems/yangвҐҮmills-and-mass-gap.
53 [Eds.]In 1990, Coleman told the class: “Believe them because Sidney is glamorous, not because his arguments
are convincing.”
54 [Eds.] Gauge symmetry does break, and then provides a renormalizable mechanism (due to Higgs, Englert and
Brout, and others) for the generation of gauge boson masses, as in the electro-weak theory of Glashow, Salam
and Weinberg. But thus far the gluons appear not to have a mass generated by this symmetry breaking. The
mechanism will be described in Chapter 46.
 Problems 21

21.1 (The first part of this problem is solved in old lecture notes1 handed out in class, but you might have fun
working it out yourself.)

(a) At first glance, it looks as if there are two possible SU(3)-invariant quartic self-couplings of the pseudoscalar
octet, Tr(ϕ 4) and (Tr(ϕ 2))2. Show that these are in fact proportional to each other, and find the constant of
proportionality. (The most straightforward way to do this is to write ϕ in diagonal form. Don’t forget to use the fact
that Tr(ϕ) = 0.)

(b) A true story: Some years ago the theory group at Saclay was investigating bound-state approximations in
quantum field theory, and decided to use as a theoretical laboratory the theory of a pseudoscalar meson octet with
the most general renormalizable SU(3)-invariant parity-invariant quartic self-couplings. (There were no baryons or
other fields in the theory.) Of course, they found that the lightest two-meson bound states were s-waves and, since
their theory was SU(3)-invariant and since an s-wave is symmetric is space, these were singlets, octets and 27-
plets. However, to their surprise, the octets and 27-plets were degenerate in mass. Can you explain why?

Possibly useful information: (1) In the lectures I explained how our method of finding representations of SU(n)
became much more complicated for n > 3. However, the complications don’t emerge until we start considering
tensors with rank greater than 2. (2) Our methods work as well for SO(n) as for SU(n); the only difference is that
for SO(n) there’s no distinction between upper and lower indices. (3) For SO(n),

For n > 4, the three representations on the right-hand side are irreducible and inequivalent.
(1998b 9.1)

21.2 In class I worked out the magnetic moments (Table 38.2, p. 837) and electromagnetic mass splittings (Table
38.3, p. 841) of the baryon octet. Things turn out to be even simpler for the decuplet. As in class, assume perfect
SU(3) symmetry, broken only by electromagnetism. Show that all magnetic moments within the decuplet are
proportional to the charge, and in particular, that neutral elements have vanishing magnetic moments. Find a
correspondingly simple statement for the electromagnetic mass splittings.
(1998b 9.2)

21.3 SU(3) allows only one possible coupling of the electromagnetic current to a quark and an antiquark. Thus (by
the same reasoning used for the decuplet in the previous problem), in the limit of perfect SU(3) symmetry, if quarks
are observable, their magnetic moments would be proportional to their charges. In the non-relativistic limit,

where κ is an unknown constant, q is the electric charge of the quark in question, and σ is the vector of Pauli spin
matrices.

In the naive quark model discussed in class, the baryons are considered as non-relativistic three-quark bound
states with no spin-dependent interactions. Thus, as in atomic physics, we can compute the baryon moments in
terms of the quark moments, that is, in terms of the single unknown constant κ, if we know the baryon wave
function. For the lightest baryon octet, the one that contains the proton and the neutron, there is no orbital
contribution to the magnetic moments because each quark has zero orbital angular momentum. Thus all we need
is the spin-flavor-color part of the wave function. Of course, since the assumption of perfect SU(3) already gives all
the baryon moments in terms of the proton and neutron moments, the only new information we get from this
analysis is the ratio of these moments. Compute the ratio and compare it to experiment.

Remark. It’s clear from the way the calculation is set up that it’s the total moment you will be computing, not
the anomalous moment. Be careful that you don’t use the anomalous moments when you make the computation.

H INT: You will need the spin-flavor part of the wave functions for both the proton and the neutron to do this
problem. Here is an easy way to construct them without resorting to tables of 3j symbols. It is trivial to construct
the wave function for the Iz = Jz = state of the Δ; it is |u ↑, u ↑, u ↑ñ, with all three quarks being up quarks, and all
three spinning up. If we apply both an isospin lowering operator and a spin lowering operator to this, we obtain the
Iz = Jz = state of the Δ. The Jz = state of the proton must be orthogonal to this. The Jz = state of the neutron
(up to an irrelevant phase) is obtained from the proton state by exchanging u and d.
(1998b 9.3)

1 [Eds.] “An Introduction to Unitary Symmetry”, the Erice notes from the summer of 1966, originally published in
Strong and Weak Interactions – Present Problems, Academic Press, 1966, and reprinted in Coleman Aspects.

Solutions 21

21.1 (a) There are at least two ways to solve this problem. First, the straightforward method. The 3 × 3 matrix ϕ
describing the pseudoscalar octet (37.35) is hermitian and traceless. Let the eigenvalues of ϕ be λ1, λ2, and λ3.
Because the matrix is traceless and can be diagonalized, it follows

The trace of a matrix is invariant under similarity transformations, so Tr[ϕ 4] and (Tr[ϕ 2])2 are unchanged, and we
can take ϕ to be diagonal. We then have

and

The two expressions are proportional, and the constant of proportionality is 2.

A slicker solution is found in Section 2.13 in Coleman Aspects. One way to write the characteristic equation of
the matrix ϕ is

Expanding this out, we find

which can be written as

Because ϕ is traceless, this becomes

Multiply by ϕ and take the trace to obtain

which is what we found earlier.

(b) The degeneracy of the 8 and the 27 of SU(3) is presumably due to a larger hidden symmetry G of the
Lagrangian for which the 8 and the 27 are parts of one multiplet. But what is the larger group? Since there are 8
pseudoscalar mesons, an obvious guess (reinforced by the “possibly useful information”) is SO(8). Let ϕ denote
the matrix of fields
where Ta are the 8 matrix generators2 of SU(3), normalized to satisfy

From part (a) there is only one invariant quartic coupling,

For a pseudoscalar meson theory, the most general renormalizable Lagrange density which is Lorentz invariant,
parity invariant, and SU(3) invariant is thus

This L does have a larger symmetry group, SO(8), under which ϕ transforms as a real 8-dimensional vector.
Physical states of this theory must therefore form representations of SO(8), and the s-wave bound states lie in a
representation contained in the symmetric product of two 8-dimensional vectors.

We’ve already considered ((37.64), (38.58), and (38.59)) the symmetry of the direct product of two 8-
dimensional vectors. We found that the symmetric parts of this product were

In the decomposition of 8 ⊗ 8, we had another 8, a 10 and an 10, but those were all antisymmetric.

Let’s now come at this product from SO(8). From the “possibly useful information” we have that the product of
two vectors Aa ⊗ Bb in SO(8) gives a scalar, an antisymmetric tensor and a traceless symmetric tensor. There are
8(8 + 1) = 36 symmetric products, and 8(8 − 1) = 28 antisymmetric ones. Taking the trace from the symmetric
tensor with 36 elements, AiBj + AjBi, we decompose it into a 1 and a 35, so

We know the 35 is a symmetric, irreducible multiplet of SO(8), into which can fit snugly the 8 and the 27 of SU(3).
The octet and 27-plet are degenerate, because they belong to the same multiplet, the 35, in SO(8).

21.2 (a). Electromagnetic form factors are associated (§38.3) with the matrix element

where D is a member of the decuplet and jµem is the electromagnetic current. We are looking for the SU(3)-
invariant couplings between D, D and jµem. D and D transform as 10 and 10, respectively; the electromagnetic
current is a member of an octet. The number of singlets in (3, 0) ⊗ (0, 3) ⊗ (1, 1) is the number of unknowns
needed to determine the magnetic moments of the decuplet. That in turn is determined by the number of 8’s in 10
⊗ 10, or the number of 10’s in 10 ⊗ 8, etc.; a tensor product of two representations includes an invariant (a singlet)
if the two representations are identical, and then the product includes exactly one singlet. Using the algorithm in
§37.4 we have

Since the tensor product 10 ⊗ 10 contains exactly one 8, there is exactly one invariant term in áD|j µem|Dñ. By the
Wigner–Eckart theorem, matrix elements of tensor operators in the same representation are proportional to each
other. As both jµem and the charge operator, Q transform as members of an octet, the current is proportional to Q,
so we can write

Then F1 = a1q, F2 = a2q.

(b) Electromagnetic mass splitting is associated (§38.4) with the matrix element
where the superscript S denotes the symmetric product. We already know

Among the terms of the product (8 ⊗ 8|symmetric) ⊗ (10 ⊗ 10) are three invariant couplings arising from the
products 1 ⊗ 1, 8 ⊗ 8 and 27 ⊗ 27.

For the product 1 ⊗ 1, the operator is a tensor operator, and we get a contribution

This term gives equal contributions to the masses of all the decuplet; it’s a shift, not a splitting.

Now we need, for the other two invariants, an operator which transforms like the symmetric product of two
currents. One candidate is Q2. It has the right component of (2, 2) in it. It probably also has some (1, 1) and (0, 0)
components. The latter doesn’t matter, but the (1, 1) components could screw things up if it were not the right
member of the octet. Fortunately, it is the right member, as an easy argument shows. A full SU(2) ⊗ U(1)
subgroup of SU(3) commutes with Q2, and hence with any (1, 1) piece of Q2. This determines, up to a factor,
which member of the octet this piece is. Since the same argument holds for the current–current Hamiltonian, it
and Q2 both contain the same member of the octet. In fact, this argument applies to Q itself: it must be the same
member of an octet as the octet piece of the current–current electromagnetic Hamiltonian. So the matrix element
of the current–current Hamiltonian must be a linear combination of Q and Q2:

The EM contributions to the masses of the baryon decuplet should be fit by the formula mD = a + bq + cq2.

20.3 We are asked to compute the ratio of the magnetic moments of the proton and the neutron. The magnetic
moment of a baryon in state |Bñ is computed from the magnetic moments of the three constituent quarks:

where, according to the problem,

Following the advice given in the problem, we start with the wave function for the Δ++:

Applying I– and J– operators to this, we obtain the wave function for the Δ+:

The proton wave function must be orthogonal to |Δ+ñ. Start with the piece proportional to |uudñ. It is symmetric in
flavor for the first two quarks and so must also be symmetric in spin for them:

The pieces proportional to |uduñ and |duuñ can be found in the same way. The normalized wave function is then

The neutron wave function is just the same, but with u ↔ d:

The proton’s magnetic moment is given by

and the neutron’s similarly. Looking only at the first part of the proton wave function,
Applying the operator κqσz to this wave function gives

which, when cleaned up, is

so that

the extra factor of 3 arising because the other two kets are just permutations of the first one, and all the inner
products give the same contribution (we are taking the basic kets |q1s1q2s2q3s3ñ to be orthonormal). In exactly the
same way,

and

Then we predict

From the 2016 Particle Data Group tables, in units of the nuclear magneton µN,

so that

in excellent agreement with the prediction.

(Incidentally, this problem is worked out as Example 5.3 in Griffiths EP, pp. 189–190.)

2 [Eds.] These {Ta} are half the Gell-Mann λ matrices; Ta = λa; see note 15, p. 807. Unfortunately, if we used λa
for the generators, it would be very easy to confuse them with the eigenvalues λi.

40
Weak interactions and their currents

We’re going to begin a completely new topic.1 The subject goes under the general name of current algebra,2
although the actual algebra will not emerge until the next lecture. I will try to keep the conventions consistent within
the lectures on this topic, although not necessarily in agreement with the literature (or even with previous lectures).
In order to explain the subject I will begin by giving a lightning summary of the weak interactions.3 Not the weak
interactions as we know them today, with CP violation, neutral currents, renormalizable spontaneously broken
gauge theories of the weak and electromagnetic interactions, and all that, but the weak interactions as they were
in my childhood: the low-energy phenomenology of the Fermi theory4 as improved by Feynman and Gell-Mann,
circa 1965, when nobody had yet seen an intermediate vector boson. That’s still a pretty good theory for
practically all weak interaction processes below a few GeV.

40.1The weak interactions circa 1965

In the mid 1960’s, the interaction Lagrangian responsible for the weak interactions took the form of a universal
Fermi constant GF, governing the strength of all weak interactions, divided by the square root of 2 by convention,
times the product of some currents Jλ:

The current Jλ will be described in detail below. Briefly, it is made up of all the fields that describe the weakly
interacting particles under consideration. GF is small in comparison to the scale of α = e2/4π ≈ 1/137:

It is determined from measurement of the µ lifetime.5 GF has units; it has to. The space integral of a conventionally
normalized current, a “charge”, is dimensionless, so a current has dimensions L−3 or M3, and a product of two
currents as in (40.1) has dimensions of M6. In order to make a Lagrange density of dimension M4 the Fermi
constant must have dimensions M−2. The interaction is therefore non-renormalizable,6 and that was one of the
great problems with weak interaction theories in the mid-1960’s. The theory enabled people to compute everything
at low energy with dazzling accuracy. But whenever they tried to compute higher-order corrections they got
divergences and infinities that could not be absorbed into a renormalization. It was frustrating that the weak
interactions were so weak. If only the weak interactions had been a little bit stronger, we could have seen the
second-order effects easily in feasible experiments. And then we might have gotten some idea about what’s going
on. But we couldn’t, and so we had to rely upon genius to figure out the weak interactions. We think genius came
through and delivered the answers. But we still aren’t sure, because the second-order effects remain hard to
measure.

The weak current Jλ is the sum of a hadronic part and a leptonic part:

While the theory is charge conserving, both of the currents carry charge; Jλ is a positively charged (+1) current
that creates positively charged particles when acting on the vacuum, and its Hermitian adjoint is negatively
charged. Thus the current’s matrix elements áf|Jλ|iñ produce a change of charge

and their adjoints also change charge:

The interaction is set up to be CP-conserving. (It does not include the small CP-violating effects observed in the
neutral kaon decays.7)

This form (40.1) of the Lagrangian can describe three kinds of interactions. Purely leptonic weak interactions,
such as muon decay or high-energy neutrino scattering off electrons to make muons, come from the product of
lepton currents. From the study of muon decay,

we know that the proper form of the leptonic weak current is given by

(We use the notation in which the labels of the fermions stand for the four-component Dirac spinors associated
with them.) The current Jℓλ annihilates an electron or a muon, negatively charged particles. This current fits muon
decay, a beautiful process, and it sets the scale of Jλ as well as the size (40.2) of the Fermi constant, by providing
the scale of the leptonic part.8 Muon decay proceeds through weak interactions so the lowest-order perturbation
theory should be absolutely reliable. The particles involved in muon decay have no interactions worth worrying
about aside from the weak interactions. Electromagnetism is present, but typically it does not make important
corrections to muon decay, and when it does, we know how to compute them. From experiments we had learned
enough essentially to read off the Lagrangian, and (40.1), with (40.3) and (40.7) express what we had determined
up to 1967, when things changed dramatically. Note that the current (40.7) is neither pure vector, ψγµψ, nor pure
axial vector, ψγµγ5ψ, but the difference of these, a “V − A” form.

We can also have the product of hadronic currents which give us purely hadronic weak interactions, such as

And for that, the evidence for the current-current form of the interaction was zero, since we knew nothing about the
strong interactions. They contaminate the process in complicated ways, making it essentially impossible to read
off the interaction. The hadronic part Jhλ of the current was conjectured (based on symmetry) in the 1960s to have
a form like (40.7). About the hadronic part we will say no more now, except that it is made up only of hadronic
fields. We will shortly describe it more fully.

The more interesting things are the so-called semi-leptonic decays, where a hadron h goes into another
hadron (or hadrons), h′, or possibly the vacuum, plus a pair of leptons:

The matrix element governing this process is

if the lepton combination is negatively charged. (If they’re positively charged we study the complex conjugate of
this matrix element.) This factors into hadronic and leptonic parts because leptons and hadrons interact only
weakly (ignoring electromagnetism), so in lowest order we simply get

Thus the situation is rather like that of an electron scattering off a proton, where the matrix element factors into a
known part and a mysterious part.9

We know all the dependence on the leptonic part, so we can write the entire matrix element in terms of a one-
hadron matrix element of Jhλ, if h and h′ are both single hadrons. We can parameterize this process in terms of
weak interaction form factors, analogs of electromagnetic form factors. This is as big an improvement as what
we get by writing electron scattering off a proton in terms of the proton form factors.10 Instead of a function of many
kinematic variables we have a function of only one kinematic variable, the momentum transfer to the current. In an
atypical case like pion decay where the π+ goes into leptons (with no final hadrons), things are even simpler. We
have a matrix element of the current between a one pion state and the vacuum. Instead of form factors we just
have a number, since there are no free kinematic variables. I will discuss these in more detail shortly.

40.2The conserved vector current hypothesis

By studying these semi-leptonic decay processes we have learned a lot about Jhλ. The leptonic current (40.7) is
parity-violating.11 It’s the difference of a vector current Vℓλ and an axial vector current Aℓλ; following Feynman and
Gell-Mann12 we assume the same is true of the complete current Jλ:

Under parity, Vλ transforms as a true vector and Aλ as an axial vector, so the sum or difference of these cannot
conserve parity. On the other hand, both parts taken together (in the combination (40.11)) are charge-conserving;
Jhλ has ΔQ = +1 (40.4), but Jℓλ† has ΔQ = −1.

Both the vector and axial vector parts of the hadronic current act alike, though there are two separate cases.
In the first case, neither the vector nor the axial vector part changes the hypercharge, and both obey the selection
rule ΔI = 1. The currents in the first case transform under isospin like the positively-charged component of an
isovector; they have the isospin transformation properties of the π+ state. In the second case, the two parts of the
hadronic current change the hypercharge by +1 and the total isospin by . This is the famous ΔI = rule for semi-
leptonic decays.13 That is, this current has the same transformation properties as the K+ state:

We knew something else, even in the late 1950s, about the vector part of this current with ΔY = 0. This is the
famous conserved vector current hypothesis of Feynman and Gell-Mann, or CVC for short.14

Consider JhλΔY=0, the part of the hadronic current that doesn’t change the strangeness; it contributes for
example to neutron β decay. It had been known for a long time that the vector part of that, VhλΔY=0, obeyed
universality: the β decay constant seemed to be about the same as the coupling constant in muon decay. The
matrix element of the vector current at small momentum transfers (the proton is so close to the neutron that only
small momentum transfers are relevant) seemed to be pretty close to 1, in the scale at which the vector current,
Vℓλ, had matrix element 1 between electron and neutrino. Feynman and Gell-Mann argued that this couldn’t be a
coincidence. Even if VhλΔY=0 started out initially with matrix element 1, the strong interactions were going to get
into our computation and change things from 1 to or or something. How can it stay 1? Well, they continued,
there is only one case we know in which the matrix element of a current at small momentum transfer is not
affected by the strong interactions: the current must be conserved. The premier example of this is
electromagnetism, where F1(0) stays firmly fixed at 1. It has no strong interaction corrections. The proton has an
anomalous magnetic moment but it doesn’t have an anomalous charge. That argument was true for the lowest
order in electromagnetism and to all orders in the strong interactions. Perhaps there would be a parallel between
the electromagnetic current and any other conserved current. So they guessed that this current VhλΔY=0 has got
to be a conserved current.

Gell-Mann and Feynman knew of only one conserved current, the positively-charged isospin current. So they
guessed, in accordance with (40.13), that VhλΔY=0 was proportional to the charge-raising (and hence isospin
raising) isospin current, whose integral is I+:

with α w a constant to be determined. In this way they explained the so-called universality of the weak
interactions: the vector part of the matrix element for neutron β decay, for small momentum transfers, was
precisely15 the same as that for muon decay, 1. (It was actually 1 within 1 or 2% but they said the difference could
be due to an electromagnetic vertex correction or something.) The vector part of the ΔY = 0 current was to be
proportional to just the ΔQ = 1 conserved isospin current, not with any funny Pauli-type terms, e.g., ∂µσµλ times
some other factor, but exactly that. This was a bold guess. They could have included many other conserved
currents, by adding divergences of antisymmetric tensors, but they thought that made the interaction too ugly.

The physicists of the time could check this guess. The strong interactions are isospin invariant, so the form
factors for these decays, F1 and F2, should be related just by an isospin rotation to the form factors for the Iz
current. Those we know from the Gell-Mann–Nishijima relation (35.52): Iz is part of the electromagnetic current,
which we get by taking the difference between the proton and the neutron form factors. Measuring these weak
interaction form factors is difficult. The least difficult to measure is the analog of F2 at zero, and that is not easy.
You have to look at a cunningly chosen nuclear β decay so that the F1 form factor obeys the wrong selection rules
and can’t play a role. Then only the F2 form factor is involved. After you’ve divided out the nuclear physics matrix
elements, you end up with, in principle, a measurement of F2(0) for this weak interaction current. That should be
related by an isospin rotation to the electromagnetic F2(0), the difference between the proton and neutron. This is
called weak magnetism16. Because this is not a course in the weak interactions I’m not going to give all the
details. This is a testable hypothesis which has been tested, and it works. (We’re not however going to use it in our
current algebra discussions.)

40.3The Cabibbo angle

Since we have devoted a lot of time to talking about SU(3), and since this is a lightning summary of the weak
interactions, I should tell you about Cabibbo’s work in 1963; he fit the weak interactions together with SU(3).17 He
said, “Ha! Feynman and Gell-Mann have told us that VhλΔY=0, the ΔI = 1, strangeness-conserving part of the
vector current is the isospin current JλI+. What about the other part, with ΔI = , the part that changes
strangeness?” We know the isospin current is part of an SU(3) octet; Iz is one of the eight generators of SU(3). We
know there is another positively-charged current in the same octet, the pseudoscalar mesons.18 If the top entry in
(40.13) is π+-like, there’s also the bottom entry, K+-like. That’s the only other positively-charged object in the octet.
We also know that the bottom entry in (40.13) has the same isospin and strangeness properties as the K+.
Therefore if you label SU(3) octet currents by the transformation properties of the appropriate meson, so we have
a vector current that transforms like the π+, the positively-charged isospin current, and a current that transforms
like the K+, the positively-charged strangeness-changing current, it seems very natural to imagine that the total
vector weak interaction current is simply the sum of these two things:19

After all, in the world of perfect SU(3), who can tell the difference between a π+ and a K+? Indeed, Cabibbo
suggested that the combinations are weighted together in such a way that the sum of the squares of the
coefficients is 1:

The angle θC is called the Cabibbo angle. The Cabibbo angle θC must be close to zero, so that cos θC will be 1
within a few percent, to agree with our earlier discussion. The things that Feynman and Gell-Mann thought were
electromagnetic corrections were really an electromagnetic correction plus terms of order (θC)2.

Let that be for a moment. Cabibbo’s general idea was that God said, “Let there be weak interactions and let
there be medium-strong interactions that break SU(3)”, apparently without looking to make sure they were in the
same direction. If there were no medium-strong interactions you could, with an SU(3) rotation, turn VK+λ into Vπ+λ
without affecting electromagnetism. And then there would be no strangeness-changing at all, by definition, since
you can define strangeness as you wish if there are no SU(3)-violating interactions. The angle θC is a measure of
mismatch of the directions in SU(3) space chosen by the medium-strong interactions and chosen by the weak
interactions. It just happened that the directions didn’t quite match.

That was Cabibbo’s idea. And being a bold man, he said exactly the same thing should be true for the axial
vector currents, with the same angle:

(Some people tried a different angle θ for the axial vectors, but the best fit to experiment is θ = θC.) Cabibbo
postulated another octet of axial vector operators that formed the axial, nonconserved currents. The positively
charged members were to be put together with the same angle. People looked at (40.16) (Shelly Glashow and I
among others) and said, “What a random guess. What’s the experimental evidence for that?” At the time, nobody
appreciated how attractive a guess this was.

Cabibbo’s guess gives us a lot of information about semi-leptonic decays of the baryon octet. There are a lot
of these,20 nine or ten. How many unknown constants do we have in this matrix element? All of these decays
proceed at relatively small momentum transfer so we really have to know only the various form factors in the
vector and axial vector currents at zero momentum transfer. We know the Fermi constant GF from muon decay.
That is not a free parameter; I’ll put it in parentheses. We need to determine the Cabibbo angle θC. We have the
matrix elements of the vector and axial vector currents but we know those at zero momentum transfer in the SU(3)
limit because they are SU(3) conserved currents. And we have the axial vector currents, assumed to be an octet.
(We call them “currents”, even though they aren’t associated with any conservation laws; they’re just vector
operators.) They can couple, octet to octet, to the baryon octet with some unknown constants d and f.21 Thus we
have four parameters with which we can fit all semi-leptonic baryon decays, three of them free:

There are a lot of decays that we can fit with these. We know the vector matrix elements in terms of the parameter
θC, and the axial vector matrix elements in terms of the parameters d and f. And it fits; it’s the right theory.
 40.4The Goldberger–Treiman relation

I want to say a little more about the various semi-leptonic matrix elements and the matrix elements of the vector
and axial vector hadronic currents, in particular for the processes (40.9) of nuclear β decay

and pion decay. For neutron decay we need the matrix element of the hadronic current (at point x) between a
proton and a neutron:

Define the momentum k as the difference of the neutron and proton momenta:

For n → p this should be quite small. The only term that survives from the vector current at really small momentum
transfers is the analog of an F1 form factor, called gV (k2). There also is a σµν form factor and other stuff like that,
but that’s got powers of k in it.22 From the axial vector current, because of the V − A definition, there is also a
gA(k2) term. And then there is some other junk, which I will write down in more detail shortly:

These are the dominant elements in low energy neutron decay, where the momentum transfer is very small (a few
MeV). The other terms are all proportional to powers of k, and are killed off in the limit k → 0.

At zero momentum transfer, the value of gV (k2), which we’ll call gV , should be cos θC, if we accept Feynman
and Gell-Mann as modified by Cabibbo. The Cabibbo angle is rather small,23 about 13°. So to the order in which
we’re working, we’ll just ignore strangeness-changing weak interactions and set cos θC to 1:

That introduces an error of a few percent, but we’re not going to get any formulas accurate even to a few percent
in the remainder of this lecture. The other term, gA(0), has been measured in neutron decay,24 and this value will
become significant to us. In the literature typically one finds the ratio gA/gV but as gV is essentially 1, the reported
value25 can be taken for gA:

The other process to consider is π− decay:26

The π− is a pseudoscalar particle and it’s easy to see using parity that the vector current matrix element must
vanish between a pseudoscalar particle and the vacuum:

Only the axial vector current has a nonzero matrix element:

The form of the right-hand side is not hard to explain. The only vector around is the pion momentum. There must
be a factor of e−ip⋅x . The remainder is a well-known number (divided by ) called the pion decay constant and
denoted Fπ; it is measured from the pion decay rate. It doesn’t depend on k2 (by analogy with (40.20)) because
there is only one momentum here. The is there to make subsequent equations look simple.27 We put in an i by
convention.

We can connect Fπ to gA(0) (i.e., pion decay to neutron decay). The error in the pion constant is considerably
less than the error in gA(0). Without any error bars28 it is:
This is straight phenomenology. The relation (40.26) is just the statement that charged pion decay occurs. 29 The
form of the matrix element is completely determined by parity and other constraints. You measure the rate of pion
decay in the process π+ → µ+ + νµ to establish the value of Fπ.

Now comes a deep insight. Take the divergence of (40.26):

It isn’t zero. The divergence of the axial vector current, a pseudoscalar field, has nonzero matrix elements
between the one-pion state and the vacuum. The only new piece of information is that pions decay. Back in §14.2,
when we went around in circles with the LSZ reduction formula, I said that one of its consequences was that any
local field was as good as any other for computing S-matrix elements. It didn’t matter what we used for a local
field, so long as it had nonzero amplitude for connecting the particle in question to the vacuum. Therefore we have
the freedom to define the π− field as

That is a definition. I do not say this ϕ π−(x) is a canonical field; God forbid! But it is a perfectly legitimate local
operator that will serve as a pion field. It may or may not be equal to the canonical pion field that appears in the
Lagrangian of a theory with fundamental pions in it.

Whether the pion is a fundamental object or not, the matrix element can connect a neutron and a proton:
there is a strong interaction. I’ll factor out the pion pole; there’s an i from the pion propagator and an i from the
pion–nucleon vertex that gives a −1:

The is to take care of the convention in the pion–nucleon coupling constant. The factor g(k2) is a form
factor. If we chose a different candidate for the pion field we would get a different form factor. Whatever we
choose, the residue at the pion pole is always going to be the same: g(mπ2) will be the conventionally defined
strong interaction constant g, and the experimental value of g is known, again with a negligible error on the scale
in which we are working: 30

The best way of determining g is from the forward pion–nucleon scattering.

Now let’s look a bit more closely at the axial vector matrix elements between proton and neutron. There are
three possible invariants.31

The first term we wrote down before (40.21); in the second, gM(k2) is the analog of the magnetic form factor; it will
disappear in the computation we are going to do. The gp(k2) form factor is in the axial vector current but not in the
vector current; the axial vector current isn’t conserved (as is obvious from (40.28)). It’s easy to see that all these
form factors have to be there. The first two terms are just the analog of the computation we did for the vector
current with γ5 inserted. The last term contributes to the longitudinal part of Aµ. It’s like the divergence of a scalar
which produces a factor kµ.

Making the sensible kinematic approximation that

one can trivially compute the divergence of (40.32). The ’s and the γ5’s go together and as usual they hit the
spinors on the right and the left. One gets (recall k = pn − pp)

The first and third terms in (40.32) are trivial; the middle term drops out because
Comparing (40.34) with (40.30) and the definition of ϕ π− we obtain

This is nothing but a definition. It is completely empty of any physical content. It simply connects two ways of
parameterizing the same matrix element, the matrix element for the divergence of the axial vector current.

Now we introduce physics into this equation with the following hypothesis: The function g(k2) is free of
singularities up to the three-pion threshold. We have extracted out the one-pion pole. Furthermore, if our
experience with the electromagnetic form factors, with which this object is closely analogous, is any guide, we
don’t expect a lot of variation even at the three-pion threshold. Experimentally, the electromagnetic form factors
don’t exhibit big changes at twice the mass squared of the pion; it’s only when you get up to the ρ mass that they
have gigantic bumps in them. So even at the three-pion threshold, if you say something like the ρ mass or the Ω
mass or the mass of some axial vector meson is the thing to look at, you don’t expect g(k2) to vary enormously
over the region between k2 = 0 and k2 = mπ2. That’s a nice analytic region. The threshold at 9mπ2 probably has a
small discontinuity, if the electromagnetic form factors are a guide. If so, mπ2 is a small fraction of the way, about
10%, to the nearest singularity. Therefore we make the hypothesis that in a region of analyticity, small compared to
the distance to the nearest singularity and small compared to the characteristic length involved in the problem, we
would expect

This is a physical hypothesis: that this matrix element varies in the same way as every other matrix element we
can measure; i.e., not much over a distance of mπ2 once we’ve extracted out the pion singularity. The first part of
the next lecture we will explain this hypothesis in four different ways, because it is so critical.

Evaluating both sides of (40.35) at k2 = 0, and using the hypothesis of (40.36), we see that the mπ2’s and
factors of cancel. Multiplying both sides by Fπ, we find

This is the famous (and at one time, notorious) Goldberger–Treiman relation.32 Please notice that the only thing
that has gone into this is kinematics and the single hypothesis about the rate of variation of g(0); there was no
added physics besides that.

The result (40.37) is remarkable. It connects the pion decay constant Fπ, the strong interaction pion–nucleon
coupling constant g and the nucleon axial vector decay constant, gA. It was very strange, because in those days
people thought nucleons were fundamental, so maybe pions are bound states (nucleon plus antinucleon). They
had a notion that pion decay was caused by nucleon decay:33 a pion comes along, becomes a nucleon-
antinucleon pair and these β decay, as shown in Figure 40.1.

Figure 40.1 An old and erroneous view of pion decay

Notice that this picture leads to a relationship which is the wrong way around: you’re connecting the decay
constant Fπ to the product of g at the πNN vertex and gA at the ANN vertex; the g would wind up on the right side
of (40.37), and give the erroneous

I also emphasize that the factors are of completely different magnitude. On the left-hand side, we have this
enormous number g, 13.3, and nothing else is so large. If this works out, we have a right to be proud.

Now the experimental situation. Well, what is the answer? Putting numbers into (40.37) we get
The left-hand side is 2.61, and the right-hand side, is 2.54, in units of mp. The agreement is, by any standard,
excellent. We will have much more to say about this next time.

1 [Eds.]Lectures 40–45 in the videotapes concern dispersion relations. Owing to length and time constraints, the
editors have decided not to include these six lectures. (The most serious casualty of this omission is the Adler-
Weisberger relation.) This (short) chapter represents the last 48 minutes of (a very long) Lecture 45, starting at
1:00:15. This book’s last nine chapters coincide with the last nine videotaped lectures.
2 [Eds.]
“Soft Pions”, Chapter 2, pp. 36–66 in Coleman Aspects; Stephen L. Adler and Roger F. Dashen, Current
Algebras and Applications to Particle Physics, W. A. Benjamin, 1968; Jeremy Bernstein, Elementary Particles and
Their Currents, W. H. Freeman, 1968.
3[Eds.] E. D. Commins and P. H. Bucksbaum, Weak Interactions of Leptons and Quarks, Cambridge University
Press, Cambridge, 1983; Greiner & Müller GTWI; J. C. Taylor, Gauge Theories of Weak Interactions, rev. ed.,
Cambridge U. P., 1979; Abers & Lee GT.
4 [Eds.] E. Fermi, “Tentativo di una teoria dei raggi β” (A provisional theory of beta radiation), Nuovo Cim. 11
(1934) 1–19; reprinted in The Collected Papers of Enrico Fermi, v.1: Italy, 1921–1938, ed. Emilio Segré, U.
Chicago P., 1962. English translation in C. Strachan, The Theory of Beta Decay, Pergamon, 1969.
5[Eds.] D. B. Chitwood et al., “Improved Measurement of the Positive-Muon Lifetime and Determination of the
Fermi Constant”, Phys. Rev. Lett. 99 (2007) 032001; PDG 2016, p. 627, “Gauge & Higgs Boson Particle Listings”.
6 [Eds.] The dimensionality of the coupling constant is a quick indicator of whether the interaction is
renormalizable or not. If the coupling constant has negative mass dimension or positive length dimension, as G
does, the interaction is non-renormalizable; see §16.4.
7 [Eds.] Greiner & Müller GTWI, Section 8.2; also see note 9, p. 240.
8 [Eds.] Bjorken & Drell RQM, Section 10.11, pp. 247–257; Greiner & Müller GTWI, Section 6.2, pp. 208–211.
9 [Eds.] Commins and Bucksbaum, op. cit., Section 4.7, pp. 156–159.
10 [Eds.] Commins and Bucksbaum, op. cit., Bjorken & Drell RQM, pp. 242–246; §34.2, p. 738.
11 [Eds.]
T. D. Lee and C. N. Yang, “Question of Parity Conservation in Weak Interactions”, Phys. Rev. 104 (1956)
254–258; C. S. Wu, E. Ambler, R. W. Hayward, D. D. Hoppes, and R. P. Hudson, “Experimental Test of Parity
Conservation in Beta Decay”, Phys. Rev. 105 (1957) 1413–1415.
12 [Eds.] R. P. Feynman and M. Gell-Mann, “Theory of the Fermi Interaction”, Phys. Rev. 109 (1958) 193–198.
13 [Eds.] For instance, Λ → p + e− + νe. Quang Ho-Kim and Pham Xuan Uem, Elementary Particles and Their
Interactions, Springer, 1998; §6.6.2 and §16.1.4.
14 [Eds.] Feynman and Gell-Mann, op. cit.; S. S. Gershtein and Y. B. Zel’dovich, “Meson Corrections in the Theory
of Beta Decay”, Zh. Eksp. Teor. Fiz. 29 (1955) 698–699. [Sov. Phys. JETP 2 (1956) 576–578.] E. C. G. Sudarshan
and R. E. Marshak, “Chirality Invariance and the Universal Fermi Interaction”, Phys. Rev. 109 (1958) 1860-1861;
Greiner & Müller GTWI, p. 209.
15 [Eds.] Greiner & Müller GTWI, Section 6.2, pp. 208–209.
16 [Eds.] Commins and Bucksbaum, op. cit., pp. 166–167; pp. 189–200.
17 [Eds.] Nicola Cabibbo, “Unitary Symmetry and Leptonic Decays”, Phys. Rev. Lett. 10 (1963) 531–533; Greiner
& Müller GTWI, §6.4.
18 [Eds.] Table 39.3, p. 852.
19 [Eds.]The currents are easily described in terms of the eight 3 × 3 Gell-Mann λ matrices (note 15, p. 807), and
(37.35). The first three λ’s are

where σi are the Pauli matrices, generate the isospin subgroup of SU(3). In particular, I+ = λ1 + iλ2 has exactly the
one non-zero matrix element occupied by π+ in (37.35). Similarly,
for j = {1, 2}. And sure enough, λ4 + iλ5 has exactly one non-zero matrix element, the same as that occupied by K+
in (37.35). If the hypercharge matrix Y is given by (38.41) ≡ λ8, it’s easy to show that λ4 + iλ5 = Y+, i.e., [Y, Y+]
= Y+. Given an octet of vector currents, {Vλa}, a = {1, . . . 8} transforming like a (1, 1) or 8 representation of SU(3),
we can make the assignments

See D. H. Lyth, An introduction to current algebra, Oxford U. P., (1970), p. 6.

20 [Eds.] See Table 10 in A. Faessler, T. Gutsche, Barry R. Holstein, Mikhail A. Ivanov, Jürgen G. Körner, and
Valery E. Lyubovitskij, “Semileptonic decays of the light JP = 1/2+ ground state baryon octet”, Phys. Rev. D78
(2008) 094005.
21 [Eds.] See note 15, p. 807.
22 [Eds.] For the general form of the axial current matrix element, see Exercise 3.3, p. 88–91 in Greiner et. al QCD.
23 [Eds.] The Cabibbo angle is expressed in terms of the Wolfenstein parameter λ; λ = Vus = sin θC. Various
experiments have established Vus ≈ 0.225, giving θC ≈ 13°; PDG 2016, “Vud, Vus, the Cabibbo Angle and CKM
Unitarity”, pp. 1011–1013. In 1975 Coleman quoted a value of about 15°. The Cabibbo angle now finds a home in
the CKM matrix, from the work of Makoto Kobayashi and Toshihide Maskawa, who extended Cabibbo’s ideas to
include a third generation of quarks (t, top and b, bottom). M. Kobayashi and T. Maskawa, “CP-Violation in the
Renormalizable Theory of Weak Interaction”, Prog. Theo. Phys. 49 (1973) 652–657. Kobayashi and Maskawa
shared the 2008 Physics Nobel Prize for this work (Nambu was also honored in 2008, but for spontaneous
symmetry breaking; see Chapter 43).
24[Eds.] J. Liu et al., “Determination of the Axial-Vector Weak Coupling Constant with Ultracold Neutrons”, Phys.
Rev. Lett. 105 (2010) 181803.
25 [Eds.] PDG 2016, “Baryon Particle Listings”, p. 1516.
26 [Eds.] PDG 2016, p. 37.
27 [Eds.] Many authors define this matrix element without the .
28 [Eds.] PDG 2016, p. 1112.
29 [Eds.] Neutral pions decay primarily into two γ’s. There is no lighter hadron to decay into.
30 [Eds.] T. Ericson et al., “Determination of the Pion–Nucleon Coupling Constant and Scattering Lengths”, Phys.
Rev. C66 (2002) 014005; V. A. Babenko and N. M. Petrov, “Study of the Charge Dependence of the
Pion–Nucleon Coupling Constant on the Basis of Data on Low-Energy Nucleon-Nucleon Interactions”, Phys.
Atom. Nucl. 79 (2016) 67–71. The latter give g2π0 = 13.55(13), and g2π± = 14.55(13).
31 [Eds.] Coleman Aspects, “Soft Pions”, §3, p. 41. The subscript M is for “magnetic”, p for “pseudoscalar”.
32 [Eds.] M. Goldberger and S. Treiman, “Decay of the Pi Meson”, Phys. Rev. 110 (1958) 1178–1184. The original
relation is their equation (24).
33 [Eds.] Bernstein, op. cit., p. 171 has a close match to Figure 40.1, for the decay of a π+. He draws the axial
lepton current Aµ asthe lepton vertex with antimuon and muon neutrino, as shown:

41
Current algebra and PCAC

Last time we derived the Goldberger–Treiman relation:1

a nontrivial equality between two things of quite different orders of magnitude, the large dimensionless constant g
(~ 13.3) from pion–nucleon strong interactions, and the comparatively small coupling constant gA (~−1.25) from
nuclear beta decay. It has been confirmed within experimental error.

It will be helpful for what is to come2 to summarize how we came to this result. We started with the matrix
element for neutron beta decay (40.21), and looked specifically at the axial vector part:

Then we considered the weak decay of the pion, and wrote down (40.26) the matrix element of the axial current
between a π− state and the vacuum, in terms of the pion momentum and Fπ (~ 0.196mp), the pion decay constant.
We took the divergence of that equation to obtain

from which we defined

On the other hand, we know that the pion field can connect a proton and a neutron. Factoring out the pion pole, we
wrote the matrix element for this process in terms of a form factor, g(k2):

Taking the divergence of (40.32) and comparing it with (40.30), we found

This equation is really nothing but a definition. We put physics into it by making the hypothesis that the value of the
form factor g(k2) at k2 = mπ2 was the strong coupling constant g, and that g(mπ2) ≈ g(0):

The assumption g(mπ2) ≈ g(0) was plausible because we have extracted out the only singularity that we come
near when extrapolating from zero momentum to mπ2, the pion pole, in the definition of g(k2). Using this
hypothesis and taking the limit as k2 → 0 of both sides of (40.35), we arrived at the Goldberger–Treiman relation,
(40.37).

When it was first derived, it was said to be good only to within 10%, but no one was particularly disturbed by
that. You’d expect errors on the order of (mπ/mρ)2 or something, taking the ρ as a typical hadron, and we’re
making an extrapolation over a distance of one mπ2, perhaps 5%, within experimental error. The 10% error was
due to a bad measurement of neutron beta decay. It’s rather tricky to extract the axial vector contribution to
neutron beta decay from angular correlations, but with modern measurements the relation (40.37) is right on the
nose (40.39). It is so good that it’s mysterious. It’s an astonishing result.

The equation (40.29) is a key step in the derivation of the Goldberger–Treiman relation. It’s important to
understand what this equation is saying. There are many interpretations floating around in the literature. To these
we now turn.

41.1The PCAC hypothesis and its interpretation

Equation (40.29) is known by a peculiar acronym, PCAC. (I will explain what PCAC stands for in a moment.)
There are four interpretations that one comes across, in the literature or in conversations. Two are perfectly
acceptable, one is silly, and one is wrong.

One way of looking at the equation is to say that certain matrix elements áf|ϕ π−(0)|iñ involving off-mass shell
pions are slowly varying, and so can be successfully approximated by constants as the momentum transferred
goes from 0 to the pion mass shell. This is sometimes phrased as “the slow variation of the matrix elements as a
function of momentum transfer”. I would prefer to say normal rather than slow variation, because the momentum
dependence we need to justify the Goldberger–Treiman relation is pretty much the same as we find in the F1 and
F2 form factors, and other off-shell matrix elements of local operators which we can measure. This meaning is
acceptable. (I say it’s acceptable because it’s the one I choose to adopt!)

There was far more confusion about the equation in its early days than today. Another statement which you
will find in some otherwise quite profound papers goes like this: “We assume that the derivative of the axial current
is proportional to the pion field, ” or

Adler’s classic paper3 begins this way. This is one of the reasons the equation is called PCAC; the acronym
stands for partially conserved axial current. The idea is almost like the conserved vector current hypothesis,
except that we assume that the divergence of the axial current is not zero, but is instead the canonical pion field,
and call this “partial conservation”. This interpretation is silly, because from the viewpoint of the LSZ reduction
formula, any field that has a non-zero matrix element between the pion state and the vacuum is as good a pion
field as any other. There is absolutely no reason to assume that the canonical pion field in a strongly interacting
field theory has matrix elements any more slowly or any more rapidly varying than any other randomly selected
local operator. So this is an assumption without any apparent content. Nevertheless, we will have occasion very
shortly to worry a bit about inventing models that obey this relation, despite the fact that I have identified it as
silly.4

Figure 41.1 Pion pole dominance

Sometimes in the mid-1960’s, when I was going around lecturing on these topics, people in the audience,
typically otherwise intelligent experimenters, would suggest another way to think about PCAC ((though this point
of view never found its way into the literature). They’d say, “Aren’t you making a mystery out of something very
simple? Isn’t it just pion pole dominance?” (The process is shown in Figure 41.1.) “After all, ” they said, “here’s a
neutron coming in, a proton going out, here are two leptons coming out. You could say the total energy of the
leptons is very small, so if we imagine calculating the right dispersion relation in some cross channel in terms of
the lepton energy or something like that, it would seem very reasonable to dominate this by the pole diagram for
the pion.” It might seem very reasonable to just throw in the one-pion pole and say that’s what dominates at low
energies. But whether it’s reasonable or not, it’s dead wrong! It’s not merely a silly explanation; it’s worse than
that. The pion couples derivatively to the leptons; it’s the only way it can couple so its matrix element has a total
momentum. The process is proportional to kµ because the pion matrix element in the diagram is proportional to kµ.
The gp term in (40.32) has a big fat factor of momentum transfer kµ sticking out in front, but the gA term does not.
That is, if we were to write down the contribution of the pion pole diagram we would have to say

We would also predict, if you work it out,

(The definitions come together with our conventions to give a minus sign.5) Now this is neither the
Goldberger–Treiman relation nor experimentally verified. So this interpretation is worse than silly; it’s simply
wrong.

The fourth interpretation is based on an idea that was the great discovery of Nambu.6 At the time the idea
seemed rather peculiar, but its growth led to the whole theory of spontaneous symmetry breakdown in field theory.
If we imagine a world almost the same as ours, except the pion mass in that world is zero, mπ2 → 0, then the axial
vector current (in that world) would be conserved,

in the sense that from (40.29) we have

If you assume this holds, then as mπ2 → 0, ∂µAµ → 0. This is another meaning of partially conserved axial current:
it’s partially conserved in that its conservation is broken only by a very small parameter, one of the smallest
parameters in hadron physics, perhaps the only small parameter: the ratio of the pion mass to a typical hadron
mass. This looks like a very different assumption than what we have been talking about; it certainly doesn’t look
as if there’s any conceivable way, in a world with massless pions, of connecting pion decay matrix elements to
nucleon decay matrix elements. Nevertheless, using nothing but the equations above, one can show that this
leads to exactly the same conclusion.

Since the axial vector current is supposed to be strictly conserved in a world with zero pion mass, we get,
instead of something being proportional to a pion matrix element,

It looks like we’ll run into trouble if we send k2 → 0 in this expression, as we did before. The second term would
appear to vanish. Then we’d deduce gA = 0, which is hardly the Goldberger–Treiman relation. However we’ve left
something out. Now we’re working in a hypothetical world in which there is a massless pion. Thus (41.3) gp has a
pion pole in it. Setting the pion mass to 0 in that expression gives

The residue at the pion pole is unambiguous. If the pion is assumed to be the only massless particle in this
hypothetical world, there will also be the non-pole terms. (There may be a three-pion term that produces a cut near
k2 = 0.) Thus, when we take the limit as k2 → 0, we get, from the first term in (41.6), 2mgA(0) and from the second
term we get, not zero, but the residue of the pole at k2 = 0, which is −gFπ, by the preceding calculation. This is
nothing but the Goldberger–Treiman relation again:

So Nambu’s interpretation looks good, although how it connects with our other ways of reasoning is at the moment
a trifle mysterious. The connection will not be revealed until we discuss other topics, many lectures from now.

41.2Two isotriplet currents

Before I go on I would like to make a tiny notational change, since the weak interactions are going to recede into
the background for a while. I’ll write things in an isospin-symmetric form. If we restrict our attention to the
hypercharge-conserving current, the plus and minus components of an isotriplet are the only ones that contribute
to the weak interactions.7 But there is nothing to prevent us from considering a complete isotriplet of axial vector
currents, Aµa, a = {1, 2, 3}, and likewise an isotriplet of vectors, Vµa, to go with the isotriplet of pions, ϕ a. If we scale
Aµa properly, the in the original definition (40.29) of ϕ − disappears. As usual (24.21),

so if we now set

then

We can thus write an isospin-covariant version of the PCAC equation, and define the isotriplet pion field simply by

To maintain the “V − A” form of the currents

we write, parallel to (41.10)8

We extend the CVC hypothesis of Feynman and Gell-Mann (40.14) to the whole isotriplet of vector fields:

Letting ψ N be the adjoint isospinor (p, n), and ψ N the corresponding column isospinor, we have for beta decay

and in particular,

Finally we can determine α w (40.14). For small momentum transfers, (40.21) says

and so

Then Vµ3 equals 2gV times the current associated with the third component of isospin:

Since gV , according to Cabibbo, is

the normalization (41.20) is very convenient. We have scaled our current so that the third component of the vector
partner of the axial vector triplet has matrix elements +1 in a proton state and −1 in a neutron state, as Iz has
eigenvalues ± for the proton and neutron. That’s suitable for our purposes. I put the into the definition (40.29)
originally to get this normalization at the end.

Figure 41.2 Pion decay in the Fermi–Yang model

Let’s return for a moment to the second definition of PCAC. Before people really understood pion β decay,
way back in the early Paleolithic era, they had an idea: neutrons and protons were fundamental and everything
else was somehow made up of neutrons and protons, the so-called Fermi–Yang model.9 Pions were bound
states of a nucleon and an antinucleon; the π− was really a neutron and an antiproton. It was the right idea except
they used neutrons and protons instead of quarks. They had a picture of decay, of say the pion, in which the pion
comes in, makes a proton–neutron pair somehow, and then the proton–neutron pair β-decays into leptons: only
the axial current contributes. That picture leads to the conclusion that Fπ equals, aside from some proportionality
constant, the strong interaction coupling constant times the axial vector β decay coupling constant

exactly the opposite of the Goldberger–Treiman relation.

It was originally thought that the Goldberger–Treiman relation depended on the strength of the strong
interactions. Goldberger and Treiman first derived their relation in 1958, by a method not nearly as simple as the
one given here,10 but by an incredibly complicated method that involved deriving 42 dispersion relations in a row
and making all sorts of unreliable approximations about what π-π scattering was like at low energy in order to
estimate certain terms. I remember Fred Zachariasen came to Caltech in 1960 to give a sequence of lectures on
dispersion relations. The great triumph of dispersion theory was then considered to be the Goldberger–Treiman
relation. It took him two full hours of lectures in order to derive it, assuming the audience already knew all the stuff
about weak interaction phenomenology. People looked at that and said “Boy, this is really an impressive result,
because we’ve somehow cracked a strong interaction problem. Look, instead of the g you get from lowest order
perturbation theory [i.e., the mistaken (41.22)], there’s a 1/g [correct]. That result must be telling us something
important about the strength of the strong interactions, to turn this g into 1/g.” In fact, the assumptions we have
made—slow (or normal) variation, PCAC—and the Goldberger–Treiman relation itself, have nothing whatsoever
to do with the strong interactions being strong. One way of demonstrating that would be to create a model that
obeys all of our assumptions except that the strong interaction coupling constant is not strong but weak, and
seeing that we draw the same conclusions. Let’s do just that.

41.3The gradient-coupling model

If we can find a model that satisfies all of our assumptions then it should obey all of our conclusions, unless we’ve
made an error in the argument. We can then see what happens in lowest order perturbation theory. The easiest
way to construct such a model is to arrange matters so that (41.12) is true for canonical pion fields in our theory. In
lowest order perturbation theory a canonical field is sharply distinguished from any other operator, and the form
factors for canonical fields are typically trivial; they’re constants or perhaps powers of the momenta. If we arrange
our theory in a sensible way we can make them constants. In that case we will not only have slow variation, or
normal variation, we will have no variation, in lowest order perturbation theory. It will be true manifestly that all of
our assumptions hold.

To check that this result does not depend on mysterious facts about the strong interactions being strong, I
propose to construct a model field theory, not meant to be realistic in any way, save that it obeys our key
assumptions. First, the theory has an axial current. This could become the axial vector current for β decay. If the
argument is sound, the theory is guaranteed to obey the Goldberger–Treiman relation, to lowest order in the
strong interaction coupling constant.

We think that the correct model of hadrons is probably the quark model. Of course, if the coupling constants
are weak, the quarks will have no bound states; the theory will describe only free quarks. But then we can hardly
define quantities like gA, gp, Fπ and other quantities of that sort, because we won’t have any nucleons or pions. If I
want to build a model to test the logic of this argument, I need it to have pions and nucleons in it at lowest order in
perturbation theory. That means a model with fundamental pion and nucleon fields. The model will provide a
theoretical laboratory where we can see if all of our assumptions work out, for things we can explicitly compute. In
particular, we want to see if both PCAC and the Goldberger–Treiman relation emerge in our model,
notwithstanding the weakness of our “strong” interactions.

We know how to construct conserved currents in a Lagrangian field theory. Let me review the procedure.11
Suppose we have a set of fields ϕ a(x) that transform (under some operation) into ϕ a(x, λ), specified by a single
parameter λ

with the condition that for λ = 0,

Define the first-order change in the field ϕ a as usual (5.21):

We assume this is an internal symmetry, so we don’t have to worry about adding total divergences to the
Lagrangian.12 Define a canonical momentum vector as

The change in the Lagrange density is

because differentiation with respect to x and differentiation with respect to λ commute. Using the Euler–Lagrange
equations of motion, we can rewrite the first term, and

where the current Jµ associated with this symmetry is

Then

and the Lagrangian is invariant if the current is conserved:

On the other hand, if the Lagrangian is not invariant (perhaps because we’ve added to it some small term that
breaks the invariance), then we can deduce the divergence of the current from (41.30). This is the formula we will
use to create a model that displays naïve PCAC, and therefore yields the Goldberger–Treiman relation to lowest
order in perturbation theory, once we’ve chosen an appropriate symmetry group, G.

The theory we will consider is the gradient-coupling model of pions and nucleons:13

N is the nucleon field, an isospinor composed of two Dirac 4-spinors, and ϕ = {ϕ a} is an isotriplet of pseudoscalar
pions. Since we’re only working to lowest order in perturbation theory, we won’t bother with bare masses and
unrenormalized fields; in fact, the interaction term has dimension 5 (§25.4) and is thus nonrenormalizable. I’ve
written the coefficient of the interaction term as the coupling constant g divided by twice the mass mN of the
isospinor. If I were to compute the matrix element of a pion field off the mass shell between two nucleons on their
mass shell, the γµ and the derivative operators would act on the nucleon spinors and give us a factor of 2mN, one
from the spinor on the right and one from the spinor on the left, leaving us with just a g(k2) as defined before,

There will be higher order corrections. But we’re going to ignore them, since we are assuming the strong
interactions are weak. It’s clear that this is a theory that has in lowest order perturbation theory a slowly-
varying—indeed, constant—coupling g(k2).

The question is, does it contain naïve PCAC? The trick is to find the right group G and the right transformation
T ∈ G. Since we want an isotriplet of conserved currents, the transformation in this case is going to have three
components; an isotriplet of transformations:

where a is now the isospin index; it runs over the three pions, a = 1, 2, 3. The transformations just add a constant
isovector λ to the pion field ϕ and do nothing to the isospinor N, so that part of the Lagrangian is invariant, while
both the kinetic term of the ϕ field and the interaction term (which depend on ϕ only through its derivative) are
unchanged, unlike a pseudoscalar Yukawa interaction. Only one term in the Lagrangian is not invariant under this
transformation: the pion mass term.

It is easy to determine the change in the Lagrangian, which is just the infinitesimal change in the pion mass
term. D is written as a vector D a, because it is one of three objects:

The change in the Lagrangian is just the pion field, times minus the square of its mass. We’ve deduced the
divergence of the current before we’ve found the current itself, but the current is easy enough to compute from
(41.29). The canonical momentum of the nucleon does not contribute because the nucleon field does not change.
The canonical momentum of the pion is the current, because D aϕ b is just the Kronecker delta:

and

so the current is
I write the current as Aµa rather than as the usual axial current Aµa for a reason that will become clear in a
moment. And just to write it down again,

which certainly satisfies Nambu’s formulation of PCAC: if mπ = 0, the axial current is conserved.

Why did I call it Aµa rather than Aµa? Because we have no normalization condition on this current. We are
considering λa here, but we could just as well have written 7λa, in which case we would have obtained 7 times the
current. If I’m going to make this current mimic, in the limit that kµ → 0, the ordinary axial vector current (40.32), I
will have to scale Aµa so that Aµa’s one-nucleon matrix element áp|Aµa|nñ agrees with (40.32). With this in mind,
define

so that

With this definition, the nucleon term −gANγµγ5τaN has the right one-nucleon matrix elements (40.32) in lowest
order perturbation theory; the second term −(2mNgA/g)∂µϕ a gives us the pion contribution (40.28).

Now it is easy to see that the Goldberger–Treiman relation works. If we take the matrix element of this axial
vector current (normalized to have the right nucleon matrix element) between the vacuum and a one-pion state, in
lowest order perturbation theory, the only thing that contributes is the pion field. The nucleon term in lowest order
makes a nucleon–antinucleon pair, but to turn that into a pion requires higher powers of the strong interaction
coupling constant g.14 The matrix element of the axial vector, that is, the matrix element of its pion term, −(
mNgA/g)∂µϕ π−, between the vacuum and π− gives you ik µ times Fπe−ik⋅x divided by :

Taking the divergence of each side and using the lowest order equations of motion for the ϕ π− field gives, in the
limit as k → 0,

which is the Goldberger–Treiman relation again. So the Goldberger–Treiman relation has nothing to do with the
strong interactions being strong. We can construct a model in which the strong interactions are weak, and we get
exactly the same result.

Let’s compare the gradient-coupling model to the Fermi–Yang model. The Feynman diagram coming out of
the latter in Figure 41.2 implies Fπ ∝ gAg, but that’s wrong. The physics ain’t like that: the pion doesn’t really go
into a proton and a neutron for pion beta decay.15 In the gradient-coupling model, the pion does contribute to
nucleon beta decay; the axial vector current comes out having a nucleon part and a pion part. They’re linked
together by the PCAC condition. If we changed the ratio of the two terms in (41.40) we would no longer have an
equation that guarantees (in lowest order perturbation theory) that the matrix element is a constant. And then the
process wouldn’t run. That ratio is linked together in this model by having, in Nambu’s way of looking at things, an
almost conserved (or, to use the jargon, a partially conserved) axial vector current that would be conserved if the
pion mass were zero. 16 That’s how we’ve set up this model, so that the current would be conserved if the pion
mass vanished. That tells us there must be a certain relation between the pion and the nucleon parts of the
current. The process embodied in the diagram Figure 41.2 from the 1949 theory is just so much garbage.

41.4Adler’s Rule for the emission of a soft pion

Actually, this model has another use, beyond being an instructive example of a theory that embodies both good
definitions of PCAC (the one following from the hypothesis of slowly varying pion–nucleon matrix elements, or the
Nambu definition that says the current must be conserved when the pion mass goes to zero). Since it satisfies all
of our assumptions in lowest order perturbation theory, it must yield all of our conclusions. It plays a role in a
famous rule due to Adler.17

We consider some hadronic scattering process in which an initial state goes into a final state plus a pion
carrying momentum k:

(where k = pf − pi). We wish to consider this process in the limit when all four components of k go to zero, that is, to
relate it to the process i → f. This is of course a totally unphysical limit; the pion is off the mass shell if kµ → 0.
Nevertheless we want to develop a rule analogous to the Goldberger–Treiman relation, for studying such a
process in this limit. Depending on the case at hand and what kinematic regime we are in, we may or may not be
able to extrapolate back to a physically observable region and obtain an interesting result. I won’t go into the
details here; this is just supposed to be a survey of current algebra methods. I will use it later on in this lecture, but
for now I only want to show Adler’s method.

The essential idea is that the amplitude for the one-pion emission matrix element, by the LSZ reduction
formula (§14.2), is

Since these are momentum eigenstates we might as well consider the amplitude at x = 0, and forget about the
e−ik⋅x ; the factor of (k2 − mπ2) comes from the reduction formula. On the other hand, PCAC says (41.5)

Since ∂µ is −ik µ, with k the pion momentum, we can write

It looks at first glance as if this matrix element goes to 0 as k → 0. This is in general not the case; sometimes it is,
and sometimes it isn’t. The reason why it does not necessarily go to zero can be best shown by a specific
example.

Figure 41.3 Pole graph, showing line ℓ

Suppose we are considering nucleon–nucleon scattering. Some graphs that arise in perturbation theory
illustrate the argument. Here is one, shown in Figure 41.3, of sufficient complexity to make the point. We have the
axial vector current on one nucleon vertex although there’s no propagator associated with it; I’ll draw a little wiggly
line with a terminal dot for the axial vector current, Aµ. Consider that graph, where the axial vector current attaches
to an external line. Let ℓ be the line from the lower vertex of the axial current to the upper vertex of the leftmost
pion. As the momentum transferred by the axial vector current goes to 0, the line ℓ goes onto the mass shell
because it has the same momentum as the external line. Thus this graph will produce a pole from the ℓ propagator
as k → 0 in the matrix element of the axial vector current. I will call these graphs, where the axial vector current
attaches to one of the external lines, pole graphs.

Pole graphs are not the complete set of graphs one can obtain by decorating this process. We could also
have the axial vector current connecting somewhere in the middle of the diagram. I’ll just call these guts graphs.
 Figure 41.4 Guts graph

The important point about the guts graphs is that, except at special kinematic configurations, they will in general
not develop singularities as k → 0. The presence of singularities is governed by the Landau rules.18 The particles
on external lines are real, physical particles, but the others, on internal lines, aren’t. The Landau rules are
connected to how you assign internal momenta to a physical process in a diagram. If you have a vertex where a
current carrying zero momentum meets an internal line, that doesn’t change anything; it has absolutely no effect
on the process. After all, the virtual particles are not on-shell, and so cannot produce a pole as k → 0. A particle
goes along and at some point absorbs momentum zero. Nu, it keeps on going. Even at the places where in fact
there are singularities, such as thresholds, one would expect from this analysis that the singularities that develop
would be perhaps square root or logarithmic singularities, the sort associated with the beginnings of a cut. They
wouldn’t be of large enough power to kill this linear factor of kµ out in front. So certainly at every point except
thresholds, and possibly even there, the guts graphs contribute nothing to the emission of the soft pion (the pion
with zero momentum). On the other hand, the pole graphs may or may not contribute; that depends on the
particular process.

The pole graphs are exactly computable in terms of strong interaction processes not involving the emission of
a pion because at this vertex we simply have the matrix element of the axial vector current at zero momentum
transfer and everything to the right of that point is simply nucleon–nucleon scattering with no pion emitted. So the
guts graphs contribute nothing and the pole graphs contribute something that’s computable in terms of the strong
interaction process without the soft pion.

There is a simple rule that summarizes this result. From Figure 41.3 we can find the residue of the pole, the
axial vector current at zero momentum transfer: γµγ5. We multiply that by kµ and use the pion reduction, (41.47).
Explicitly,

This is precisely the Feynman diagram contribution we would get in that preposterous nonrenormalizable
gradient-coupling theory (41.32). We don’t have to keep track of all the factors; they’ve got to come out the same
as in the gradient-coupling theory, because that theory, though no one takes it seriously, obeys all of our
assumptions. The upshot of this reasoning is Adler’s rule: to lowest order on external lines,

Gradient-coupling theory is exact for the emission of a soft pion.

That’s a compact way of stating it. In greater detail, Adler’s rule says: to calculate a strong interaction matrix
element for the emission of a soft pion, find the matrix element without the pion emission, and using gradient-
coupling, sum all the terms obtained by attaching a pion to each external line.19 You need only apply the gradient-
coupling rule to the external lines; the contributions from the guts graphs vanish automatically. You still have to
worry about the extrapolation. There may be singularities other than these that you have to consider. But as the
pion momentum goes to zero, if you define the pion field to be the divergence of the axial vector current, then this
is an exact statement. The typical applications are therefore near threshold, so that when you get to a physical
pion, all of the invariants involving the pion momentum will be small; not just k2 but also k ⋅ p where p is any other
momentum in the problem.20

The application of this was first derived in another context by Nambu and Lurié. They applied it to pion
production in nucleon–nucleon scattering near the pion production threshold, and related that to nucleon–nucleon
scattering.21 We’re going to give an application of this shortly in which we won’t have to worry about threshold
singularities. Threshold singularities present special problems and there’s a little song and dance of dubious
plausibility to take care of them.

Now, if you can do it with one soft pion, why can’t you do it with two? That would be better, and give us even
more information. Or perhaps three or four? That would enable us, maybe, to discuss pion–nucleon scattering. We
know what happens with one pion when the four-momentum goes to zero. If we also know what happens when
two pion four-momenta simultaneously go to zero then we know an awful lot about pion–nucleon scattering.
However, if we are studying two soft pions this way, it is the time-ordered product

that goes into the reduction formula. These 4-divergences inside the time-ordered product are no help at all;
they’ve got to be outside the time-ordered product where they can act on the Fourier transform factors through
integration by parts and turn into momenta. Therefore we’ve got to pull the derivatives out of the time-ordered
product. But as you may remember from our discussion of gradient-coupling theories (when we were talking about
Feynman rules for derivative interactions, or indeed from an early homework problem22), when you pull a gradient
operator out of a time-ordered product, you wind up with an extra equal-time commutator, due to the gradient
acting on the θ function needed for time-ordering. Written symbolically,

Therefore, if we hope to discuss processes involving two soft pions, we have to know something about the
commutation relations of the vector and axial vector currents. The vector currents don’t look like they’ll come in
here, but in fact they must; they’re required to close the algebraic structure. This is why we now begin to study
current commutators and why the whole set of methods that I’m describing now is called current algebra.23

41.5Equal-time current commutators

I will focus on the commutators of V0a, the vector current, and A0a, the axial vector current. It’s actually only the
temporal components that we will need because it’s only the zero components that have time derivatives hooked
on them, in (41.50).

First, the vector currents. As you’ll recall, these are proportional (41.15) to the isospin currents (that’s just the
CVC hypothesis of Feynman and Gell-Mann), with a constant of proportionality equal to 2gV (41.19). At equal
times

This is the isospin algebra scaled up by 2gV . Integration over all space and division by 2gV gives the isospin
charges,24 which are the isospin generators:

So the isospin charges obey the right commutators (24.38):

The δ function in (41.51) ensures that the currents {V0a} commute for spacelike separations. Actually the δ
function is a bit of a swindle. We could have made the same statement if we put a lot of ugly terms on the right-
hand side involving derivatives of δ functions. They would all go away when we integrate the commutators to
obtain the isospin algebra. But (41.51) is certainly the structure one gets in the simplest models, where you just
get δ functions from commuting the canonical fields and canonical momenta—the currents are proportional to
ϕ πa, so we’re going to get δ functions at equal times. In more complicated models there might be terms
proportional to the gradient of a δ function; we will ignore them here. In fact when we consider a specific
application, we will see the possible presence of such terms is irrelevant. Anyway, these are the commutators that
would hold in a simple model such as any isospin-symmetric theory of pions and nucleons with renormalizable
interactions, or the quark model. So I’ll just write down these simplest forms:

This is just the statement that the axial currents transform like an isovector, and therefore the isospin generators
applied to them rotate them as an isovector should be rotated.

Fortunately neither of these two is what we really want. The one we want is

This commutator has nothing to do with the isospin group, and it’s dependent on the model.. For example, in the
gradient-coupling model the axial currents commute with each other, and the commutator is zero. They commute
with each other because they’re canonical momentum densities for the three pion fields, up to a proportionality
factor, and the canonical momenta commute at equal times. On the other hand, in the quark model, this
commutator is not zero. So the question of what we should put for the commutator depends on what model we
pick.

This is very nice, for the following reason. Up to now we’ve been doing things that were almost totally model-
independent. In the strong interactions we always worry about model-dependent predictions because we can’t
compute anything in zeroth approximation in the strong interactions. We don’t know whether one theory makes a
given prediction or another theory makes a different prediction. We can’t do perturbation theory. But (41.55) is
something which we can compute in any given model without having to solve all the equations of motion, if we
know what the axial currents are. Coupling constants are irrelevant: strong, weak, small; it doesn’t matter. So this
offers us something we can abstract from a model and write down. And then, if we are able to carry through the
same sorts of computations for two soft pions that we’ve just done for one soft pion, we can check that prediction
without having to solve a problem of strong interaction dynamics. This is one of the rare instances in the theory of
the strong interactions where you can extract something from the particular form of a Lagrangian that is true for
some theories and not true for others, and use it to make experimentally verifiable predictions. Well, we had one
other instance of that: the symmetries of the Lagrangian could be used that way. But this is something that goes
beyond simple symmetry. It is not an experimentum crucis in the sense of scientific method as described by 19th
century philosophers of science. There may be 40 billion models that give the same answer for the question
posed by the commutator (41.55). And there may be 40 billion that give different answers. But still, if one answer
gives us the right predictions, we can reject all the others.

It’s fairly straightforward to compute this commutator, (41.55), in any given model. The currents are
expressed as functions of the canonical fields. You know the commutators of the canonical fields and use the
Jacobi identity.25 There may be technical difficulties, because even in a renormalizable theory the product of two
fields at exactly the same space-time point, which is what goes into a current, is a divergent and ill-defined object,
and therefore one may require some special care. Normally the way to do that is to split the points, give them
slightly different values but all at the same time, compute the commutators, which is perfectly kosher, and then let
the splitting go to zero. 26 That’s a very clean way of doing it, and sometimes it will reveal terms, called Schwinger
terms, proportional to derivatives of δ functions, which you would have missed if you had been naïve.27 If you are
worried about ultraviolet divergences screwing you up, you might try to prove these commutators order-by-order in
perturbation theory. That’s a job for a very nervous person.

The terms discovered by Schwinger are proportional to a derivative of a δ function in the V0−Vi commutator.
He found them originally in electrodynamics, where it looks, naïvely, as if there should not be such terms, but if
you’re a little bit more careful, in particular, if you split points in the way described, they appear. The phrase
Schwinger term is sometimes used in the literature to describe “anomalous” terms and commutators generically,
terms that are not there if you’re sloppy but are there if you make a more careful investigation. In fact the
Schwinger term in the V0−Vi commutator is a bit mysterious in that it has a divergent coefficient. When you get to
the point of deriving Ward identities, it washes out in the end, so that term is irrelevant. They’re not really
mysterious, they’re very well understood; they’re just troublesome. It’s just one of those things like remembering to
zip up your fly if you’re a man; you’ve got to remember to check for possible Schwinger terms. They’re not going
to be relevant to what I’m going to do, so I won’t talk about them here. If you were to take a course on current
algebra, you’d hear a lecture on how to deal with Schwinger terms.

To do calculations with axial currents, we have to make a guess for the question mark in (41.55); we have to
have some model of the strong interactions from which we can abstract that commutator. I’ll take a simple model,
the quark model I described earlier.28 Nobody knows how to compute anything with the quark model because in
lowest order perturbation theory none of the particles we know are present, there are just the damned quarks.29
Nevertheless it’s a well-defined Lagrangian field theory and we can compute the equal-time commutators. And if
they don’t involve any brand new objects that we need special dynamics to compute, so much the better.

In the quark model, hadrons are made up out of quarks, which are Fermi–Dirac fields. The quarks are held
together by some forces, mediated by gluons. But gluons won’t contribute to the isospin or hypercharge currents,
because they couple to color, which commutes with ordinary (flavor) SU(3). And therefore the gluons carry no
charge, no hypercharge, no isospin, no nothing, except color. In particular, they aren’t going to contribute to the
weak interaction currents. The currents in the quark model will just be quark bilinears, qMγµ(γ5)q: q is the quark
field, and M is some isospin matrix depending on what current you’re looking at. A quark current Jµ has the form

Which current you’re looking at will tell you what matrix M acts on the (here suppressed) isospin or SU(3) indices
of the quark. It’s very easy to compute the commutators of these currents just by using the equal-time
anticommutators of the quark fields, but it’s even easier to do the computation using a little trick which I will now
describe.

Suppose the quarks were massless and non-interacting. That’s a preposterous assumption, but since we’re
only computing an equal-time commutator and those statements don’t affect the equal-time commutators, it
doesn’t matter. The 1 ± γ5 are helicity projection operators,30 akin to (20.124a). In a theory of free massless
fermions, a Dirac spinor splits into two uncoupled Weyl spinors (§19.1) which are eigenstates of γ5 with
eigenvalues ±1, because γ5 anti-commutes with . So 1 ± γ5 are the projection operators on two uncoupled Weyl
spinors; + for one and − for the other. The same thing works on the other side. If we drag 1 ± γ5 through the γµ the
γ5 changes sign. But since γ5 is anti-self bar (20.103), when it acts to the left, on q, its sign changes again. If the
quarks were massless and non-interacting, one of these currents with the + sign would couple left-handed quarks
only to left-handed quarks, or positive helicity to positive helicity. The other would couple negative helicity only to
negative helicity. Thus the currents with the + and − signs would deal with two completely decoupled dynamical
systems having nothing to do with each other. In particular, at equal times they would have to commute. Therefore
we have for any quark currents

This would be a trivial statement if the quarks were massless and non-interacting. If that were the case, the
commutators would be zero not only at equal times but at all times: you would have two separate worlds of right-
handed quarks and left-handed quarks, and they would never see each other. On the other hand we also know
these commutators can be computed just from what we already have, from the equal-time Dirac algebra of the
quark fields, and so they have nothing to do with whether or not the quarks are massless and non-interacting. The
point is that this must be true regardless of the existence of the quark interactions. In the gradient-coupling model,
V + A and V − A don’t commute, so the commutator [A0, Ai] you derive from that assumption is not correct. The
reason is that the axial vector current has a term in it linear in the gradient of the pion field. The vector current, the
conventional isospin current, does not have a gradient term. Work it out and see for yourself.

We can compute the axial vector commutator in five minutes by using the equal-time commutators of q with q,
which aren’t affected by either the quarks’ masses or their interactions. We will call this principle (abstracted from
the quark model, but in fact true in a much larger class of models) the chirality principle. The word chirality is
often associated with current algebras. (“Kheir” is the Greek word for “hand”, as in chiromancy which is what
Madam Selena practices down in Harvard Square, reading palms.31) This is because chiral symmetry is
associated with left-handed and right-handed spin-½ particles. An algebraic statement like (41.57) is called a
statement about chiral algebra.32

We can now fill in the question mark, for the quark model, by elementary algebra:

because the cross-terms cancel (the commutators are symmetric in x − y, but antisymmetric in {a, b}.) This
completes our current algebra:

We have an algebraically closed system. We can now compute arbitrary numbers of commutators of these
objects, and therefore the sorts of things that would involve an arbitrary numbers of soft pion emissions.

The associated charges (the space integrals of the currents) obey a similar algebra without the δ functions. In
fact it’s easy to see what that algebra is. Going back to (41.57), if V + A and V − A commute with each other, then
V + A obeys an isospin algebra all by itself and V − A obeys an isospin algebra all by itself. It’s rather like the
decomposition we made (§18.3, p. 376) for the Lorentz group into two rotation groups, except there’s no need to
insert an i into half of the operators. This is the algebra of SU(2) ⊗ charges, so we don’t have to worry about the δ
function. The charges obey an SU(2) ⊗ SU(2) algebra. This is sometimes called the SU(2) ⊗ SU(2) chiral algebra
and, along with PCAC, is one of the two pillars of all current algebra computations.

A nice way of getting the same thing was pointed out by Gell-Mann and Ne’eman.33 There is a vague idea
that the weak interactions are pretty much the same for leptons as they are for hadrons. This vague idea has a
dignified name, called lepton-hadron universality. It was an idea that was floating around for 20 years, that all
weak interactions have the same strength. The question is, how can you compare leptons and hadrons? Leptons
are electrons, muons, and taus, and their neutrinos. Hadrons are nucleons and pions and all the others in those
big fat Particle Data Group tables. Or, if you take the other attitude, hadrons are quarks and gluons. Those still
don’t look much like leptons.34 They’ve got strong interactions and the leptons don’t. If you consider color, there
are a lot more quarks than there are leptons. What does it mean that the weak interactions are pretty much the
same for the quarks as for leptons? (I’ll ignore for the moment the strangeness-changing currents, but I’ll make a
remark about them later.) Here we have obtained an algebraic structure that we could generate completely from
the weak interaction currents. We take the positively charged weak interaction current and its adjoint. These obey
the same algebra as the isospin-raising and lowering operators (as is to be expected from the CVC hypothesis).
We commute the charged weak currents to get the Iz -like operator, giving us the three components of the
isotriplet. We can break the whole thing up into parity-conserving and parity-violating parts and generate this
entire algebraic structure by successive commutations just from the weak interaction current. Or if we wanted to,
we wouldn’t have to do the parity breakup; we could just look at the V − A part of the current and say that’s the
whole thing. That will give us an SU(2) algebra.

Now if we did the same thing for leptons, it would be the same computation. And it would give the same
answer as is shown from this quark model example. Because instead of building the currents from quarks in
(41.56), we could use lepton pairs: the electron and its neutrino, the muon and its neutrino, or the tau and its
neutrino. If we consider the electron and electron neutrino to be some sort of isodoublet, then the weak interaction
current is the matrix element of an isospin-raising current. It’s the same sort of structure:

is just like

Therefore Gell-Mann and Ne’eman suggested that the right way to make the comparison was to state: you have a
lepton current and a hadron current. From the lepton current you generate an algebra. You take the current and
its adjoint and commute and commute until the whole thing closes. From the hadron current you generate an
algebra; you take the whole thing and commute and commute until the whole thing closes. The precise statement
of lepton-hadron universality is that these two algebras are the same; they’re isomorphic algebraic structures.35
As we have demonstrated, this is a statement that is consistent with leptons being different from hadrons. There
doesn’t have to be a precise parallelism, as many quarks as there are leptons or anything like that. You could
have fundamental scalar mesons among the hadrons, and so on. It wouldn’t matter. The right statement of
universality is that the algebraic structures are the same. I state without proof that this is also true if you take
account of strangeness-changing currents, as in the Cabibbo theory. As far as the commutators go, the direction
chosen by the medium-strong interactions is irrelevant. You can make an SU(3) rotation that will turn the Cabibbo
currents into pure isospin-raising and lowering currents, whence the algebraic structure is the same. So the
Cabibbo theory is also consistent with this form of universality; the algebra is exactly the same.

Now, we will apply these current commutators to study a pion with momentum k and isospin a scattering off of
some initial hadronic state h with momentum p, going into a final pion state with momentum q and isospin b plus a
final state h′ with momentum p′, as in Figure 41.5. In this particular example I will assume h is a nucleon. That’s
the case for which we have most experimental information. Sometimes I will use facts that depend on a specific
feature, such as its spin. But in fact the arguments will be general and h could be any hadronic target.

Figure 41.5 Pion–hadron scattering

The general technique is as follows. I will obtain constraints on the form of the amplitude Aba for the process
(41.62) at (or near) the point q = k = 0. (We will actually get a derivative at that point.) I will then make a power
series expansion, keeping track of only the terms I know, and extrapolate it to the physical point where the pion
has the smallest possible energy, that is, to threshold. One might be a little nervous about that because a
threshold is the beginning of a cut, but I’m sorry, it’s the best I can do, and the best anyone can do. We will just
have to cross our fingers and hope it works out. So the game is to make a power series expansion about the point
q = k = 0, extrapolate to threshold and get the threshold amplitudes, which are proportional to the scattering
lengths,36 and compare them with the experimentally observed scattering lengths. I will systematically neglect
terms of O(mπ2), because it will turn out that I can’t go beyond first order in this expansion, and mπ2 is second-
order in certain invariants.

In order to do the extrapolation we had better count how many independent invariants we have, so we will
know how many terms to write down in the power series. There’s p2 and p′2, which are both equal to the square of
the target mass, mT:

We will not take the target off the mass shell; there is no need to do that. If the pions had a fixed mass there would
be two invariants we would have to deal with. One would be for example k ⋅ p, related to the pion mass in the
center of energy frame.37 The other would be k ⋅ q, related to momentum transfer from the pions to the target.
Finally, there are the masses of the two pions, k2 and q2. That’s a complete list of independent invariants. In our
entire range of extrapolation, k2 and q2 are of O(mπ2), so we don’t even have to worry about first-order terms in the
power series expansion in them. So is k ⋅ q = O(mπ2); when the pion is at threshold, k = q and then the product is
k2. Therefore we have in fact only one coefficient to expand our power series in, k ⋅ p.

The next step to construct the power series for our amplitude. There will be a constant term, the value when k
and q are 0. There will be a first-order term times k ⋅ p. We’ll have to keep that because when we extrapolate to
threshold, it’s of order mπ. There will be all sorts of terms of order mπ2, which we are going to neglect because we
don’t know what else to do with them (but at least the neglect is systematic; benign neglect, if you will). And there
will be pole terms which, just as in the discussion of Adler’s rule, will have to be taken account of separately.
These terms can vary rapidly since they have poles in their denominators. Thus

Here is an advanced glance at what is going to occupy us for the first half of the next lecture. We’re going to
look systematically at these terms, one after the other. We will show that the pole terms are also of order mπ2, so
they are in fact negligible. We will show that A0ba is likewise negligible; it is of order mπ2. These steps will follow
without any specific assumption about the form of the axial-axial commutators. They will be in that sense model-
independent. We will just use the assumptions that we had before we filled in the question mark. Finally we will get
the A1ba term from the current commutators. It, thank God, will not be of order mπ2 (although it would be
interesting if the pion–nucleon scattering lengths were zero, up to order mπ2). We’ll actually get a precise
expression for that term from the current commutator. I will then assemble the whole thing and compare it with
experiment.

1 [Eds.] M. Goldberger and S. Treiman, “Decay of the Pi Meson”, Phys. Rev. 110 (1958) 1178–1184.
2[Eds.] Much of this lecture duplicates material in Coleman’s 1967 Erice lecture, “Soft Pions”, republished in
Coleman Aspects, Chapter 2, pp. 36–66.
3 [Eds.] Stephen L. Adler, “Consistency Conditions on the Strong Interactions Implied by a Partially Conserved
Axial-Vector Current”, Phys. Rev. 137 (1965) B1022–B1033. See Adler’s equation (3).
4[Eds.] In response to a student’s question about what this interpretation means, Coleman responds: “I wouldn’t
have brought it up but … It’s embedded in the literature. It is silly. There’s nothing to understand … Ten years ago
I remember, in this very seminar room, Francis Low, Steve Adler and I screaming at each other about whether this
was meaningless or not. You’ve been brainwashed by me, and so you see it’s meaningless. But if you look up all
those papers from the mid-1960s, the golden age of current algebra, you will find people saying the key is that the
divergence of the axial vector current is the canonical pion field. And that’s just a silly statement. You can’t derive
anything from that.”
5 [Eds.] Drawing neutron β decay as in Figure 41.1 amounts to replacing áp|Jµh|nñ with áp|ϕ –|nñá0|Aµ|π–ñ.
6 [Eds.] Y. Nambu, “Axial Vector Current Conservation in Weak Interactions”, Phys. Rev. Lett. 4 (1960) 380–383.
7 [Eds.]Coleman is discounting neutral weak interactions, which had been seen at CERN two years earlier: F. J.
Hasert et al., “Search for Elastic Muon-Neutrino Electron Scattering”, Phys. Lett. 46B (1973) 121–124; F. J. Hasert
et al., “Observation of Neutrino-like Interactions without Muon or Electron in the Gargamelle Neutrino Experiment”,
Phys. Lett. 46B (1973) 138–140.
8 [Eds.] Remember, all this is for the ΔY = 0 currents.
9 [Eds.] E. Fermi and C. N. Yang, “Are Mesons Elementary Particles?”, Phys. Rev. 76 (1949) 1739–1743. See
also note 33, p. 887, and the discussion following Figure 40.1, p. 888.
10[Eds.] The simplification is due to Gell-Mann and collaborators: M. Gell-Mann and M. Lévy, “The Axial Vector
Current in Beta Decay”, Nuovo Cim. 16 (1960) 705–726; J. Bernstein, S. Fubini, M. Gell-Mann, and W. Thirring,
“On the Decay Rate of the Charged Pion”, Nuovo Cim. 17 (1960) 757–766. Gell-Mann and Lévy describe the
Goldberger–Treiman approximations as “violent …[and] not really justified.” See the discussion following their
equation (12).
11 [Eds.] See §5.3, p. 82.
12 [Eds.] Divergences of antisymmetric tensors are occasionally added to the quantity Fµ to obtain conserved
currents with particularly desirable features, as in §5.4. In the introduction to Chapter 6, p. 105, Coleman defines
an internal symmetry as one which does not relate fields at different space-time points, but only transforms fields
at the same point. Thus derivatives do not arise in the change D of the Lagrangian for an internal symmetry: D
is zero, and so is Fµ.
13[Eds.] In the video of Lecture 46, Coleman reminds the class that they’ve seen the one meson, one nucleon
version of this model as a final examination question in the fall of 1975. It appears in this book as Problem 14.4, p.
546. see (P14.7).
14 [Eds.] See note 15, p. 898.
15 [Eds.] Fermi and Yang’s 1949 article (op. cit. ) neither discusses beta decay nor includes any diagrams, but the
process π+ → n + p in Figure 41.2 is implicit in their Lagrangian. The same diagram also arises in the gradient-
coupling model, but its contribution is only part of other processes second order in g, and so has nothing to do with
the Goldberger–Treiman relation.
16 [Eds.] Nambu, op. cit.
17 [Eds.] S. L. Adler, “Consistency Conditions on the Strong Interactions Implied by a Partially Conserved Axial-
Vector Current. II”, Phys. Rev. 139B (1965) 1638–1642; Y. Nambu and D. Lurié, “Chirality Conservation and Soft
Pion Production”, Phys. Rev. 125 (1962) 1429–1436; Coleman Aspects, p. 50; Cheng& Li GT, p. 155.
18 [Eds.]
Bjorken & Drell, Fields, Section 18.6, pp. 231–242; L. D. Landau, “On Analytic Properties of Vertex Parts
in Quantum Field Theory”, Nuc. Phys. 13 (1959) 181–192; James D. Bjorken, “Experimental tests of quantum
electrodynamics and spectral representations of Green’s functions in perturbation theory”, thesis, Stanford
University, 1959. Coleman discussed the Landau rules in the lectures on dispersion relations, regrettably not
included in this book.
19 [Eds.] Coleman Aspects, p. 50.
20[Eds.] In Woit’s notes, Coleman remarks that the Goldberger–Treiman relation follows as a special case of
Adler’s rule for a one nucleon initial and final state. This is shown explicitly in Weinberg QTF2, p. 190.
21 [Eds.] Nambu and Lurié, op. cit.
22 [Eds.] Problem 1.2, p. 49.
23 [Eds.] Coleman Aspects, pp. 50–52; S. Treiman, R. Jackiw, and D. J. Gross, Lectures on Current Algebra and
its Applications, Princeton U. P., 1972; S. L. Adler and R. F. Dashen, Current Algebras and Applications to Particle
Physics, W. A. Benjamin, 1968; S. Treiman, R. Jackiw, B. Zumino, and E. Witten, Current Algebra and Anomalies,
Princeton U. P., 1985.
24 [Eds.] Despite appearances, the integral is independent of time. See §6.2, and in particular (6.57) and the
discussion following.
25 [Eds.] [A, [B, C]] + [B, [C, A]] + [C, [A, B]] = 0.
26 [Eds.] For a brief discussion of the point-splitting technique, see Peskin & Schroeder QFT, Section 19.1.
27[Eds.] Julian Schwinger, “Field Theory Commutators”, Phys. Rev. Lett. 3 (1959) 296–297; Itzykson & Zuber,
QFT, p. 224 and p. 530.
28 [Eds.] Cheng & Li GT, Section 4.4, pp. 113–124; Griffiths EP, Section 1.8, pp. 37–44.
29 [Eds.] Happily, this is no longer true. It was not really true in 1976, but we have learned a great deal more since
then.
30 [Eds.] Peskin & Schroeder QFT, p. 142.
31 [Eds.] Greek , “kheir”, (hand). The adjective “chiral” was evidently introduced into science by Lord Kelvin in
his Robert Boyle Lecture, Oxford University, May 16, 1893: “I call any geometrical figure, or group of points,
‘chiral’, and say that it has ‘chirality’ if its image in a plane mirror, ideally realized, cannot be brought to coincide
with itself. Two equal and similar right hands are homochirally similar. Equal and similar right and left hands are
heterochirally similar or ‘allochirally’ similar (but heterochirally is better). These are also called ‘enantiomorphs,’
after a usage introduced, I believe, by German writers. Any chiral object and its image in a plane mirror are
heterochirally similar.” Lord Kelvin (Sir William Thomson), The molecular tactics of a crystal, Oxford U. P., 1894,
§22, note [8]. Available at http://www.gutenberg.org, book 54976.
32 [Eds.] Cheng & Li GT, pp. 132–136.
33 [Eds.] Murray Gell-Mann and Yuval Ne’eman, “Current-Generated Algebras”, Ann. Phys. 30 (1964) 360–369.
34 [Eds.] This is a matter of taste. Many others see a striking resemblance between quarks and leptons.
35 [Eds.] D. Burton, An Introduction to Abstract Mathematical Structures, Addison–Wesley, 1965, p. 57; Michael
Artin, Algebra, Prentice-Hall, 1991, Section 2.3, pp. 48–51.
36 [Eds.] Landau & Lifshitz, QM, p. 502; A. Messiah, Quantum Mechanics, North Holland Publishing, 1962, p. 408
and p. 861; M. L. Goldberger and K. M. Watson, Collision Theory, John Wiley, 1964, pp. 287–298. The scattering
length a is closely related to the s-wave phase shift; see Problem 22.1. Note that different authors define the
scattering length with different signs; see note 1, p. 920.
37 [Eds.] See the paragraph following (5.91), p. 96.

Problems 22

22.1 Consider the scattering (below inelastic threshold) of two distinct spinless particles. A two-particle state with
definite total momentum and vanishing center-of-momentum angular momentum is necessarily an eigenstate of
the S-matrix; the associated eigenvalue is defined to be

where δ is the s-wave phase shift. The s-wave scattering length, a, is defined to be the leading term in the
expansion of δ near threshold:

where k is the center-of-momentum momentum of either particle. Find the relation between a and the invariant
Feynman amplitude, , evaluated at threshold. How (if at all) does this relation change if the two particles are
identical? (There is no loss of generality in considering spinless particles; near threshold, s-wave amplitudes
dominate all others, and thus spin angular momentum and (vanishing) orbital angular momentum are
independently conserved.)

Comment: You could have done this exercise at any time within the past few months. It appears now because we
shall shortly be computing invariant Feynman amplitudes at threshold and comparing them to experimental
measurements of s-wave scattering lengths.
(1982b 14)

22.2 Consider the following theory of the interactions of a massless Dirac field, ν (the neutrino), a charged
massive Dirac field, e (the electron), and a massive, charged (i.e., complex) vector field, Wµ:

where Fµν = ∂µWν − ∂νWµ, and g, µ and m are positive numbers.

For the process

there are several independent amplitudes at fixed energy and angle. We don’t have to worry about the helicity of
the neutrino, because the factor of (1 + γ5) guarantees that only one helicity state participates in dynamics;
however, each W can have helicity (spin along the direction of motion) of {1, 0, −1}, and thus there are nine
amplitudes. An interesting limit in which to consider these amplitudes is that of high center-of-momentum energy,
with center-of-momentum scattering angle θ fixed, but with θ ≠ 0 and θ ≠ π. (This last restriction guarantees that
all three Mandelstam invariants—s, t, and u—grow with energy.)

(a) To lowest non-trivial order of perturbation theory, (g2), some of the nine helicity amplitudes approach (angle-
dependent) constants in the high-energy fixed-angle limit described above; we will call these amplitudes “nice”.
Others, however, grow as a power of the energy; these amplitudes are “nasty”. Which are the nasty amplitudes?
Find the explicit high-energy forms of the nasty amplitudes, retaining the leading power of the energy only. (Don’t
worry about getting the phase or the sign right; in any case I haven’t defined the phase of helicity eigenstates.)

(b) Now consider adding another term to the Lagrangian, involving a second massive Dirac field, e′, of opposite
charge (possibly a positron, but its mass M need not be the same as the electron’s):

where M and f are positive numbers. A traveller once told me that if f were chosen proportional to g, some of the
nasty amplitudes in this process would become nice. But I’ve forgotten which ones, and what the constant of
proportionality is. Find out for me.

Possibly useful information: At some stage in this computation, you may want the Dirac matrices in the
standard representation ( is a 2 × 2 identity matrix, is a 2 × 2 zero matrix, and σi are the three Pauli matrices):

(1981 253b Final, Problem 3)

Solutions 22

22.1 (a) We work in the center of momentum (CM) frame. Let |kñ be a two particle state with momentum k;
asymptotically this ket describes two plane waves. Near the threshold for inelastic scattering, the scattering of
spinless particles is isotropic; above threshold a bound state knows no distinguished direction and decays
isotropically. Thus

where is the invariant Feynman amplitude. Now let |kñ be a s-wave state, and thus rotationally invariant, with
definite linear momentum k = |k|. Then

and ák′|kñ = δ(k − k′). Also,

The ket |kñ is obviously an eigenstate of S: momentum is conserved and the final state must be in an s-state as
well. The eigenvalue is e2iδ since S is unitary. That is,

for small k. Now below threshold,

where µ is the reduced mass. But (note 8, p. 9)

Substituting this expression into (S22.3),

Taking the inner product of both sides of (S22.3) with |k′ñ,

Comparing this equation with the previous equation we obtain a = −2πiµ .

22.2 The process to be considered is

described by this Feynman diagram:

Note that there is no crossed graph, because there is no W*e−ν vertex. The invariant amplitude is given by

We write down everything in the CM frame:

where |q| = → E as E → ∞. Similarly

So the mass term in the Feynman denominator is negligible in this limit if θ ≠ 0.

Now for the spinors. The neutrino spinor u obeys these conditions:

The solution to these is

These spinors obey a useful relation (which you can check easily):

The polarization vectors, when θ = 0, are given by (see (26.74) and (26.75), p. 567)

For θ ≠ 0, we rotate by θ for W,

the superscripts denoting the W’s helicity, h. For W’s polarization vectors when θ ≠ 0, we rotate by θ + π:

the superscripts denoting the W’s helicity, h.

Just by counting powers, we see that the amplitudes are nice unless hh = 0, i.e., at least one of the vectors
has zero helicity. Even in that case, we get nasty amplitudes only from the leading term in the high-E expression
from ε(0) or ε′(0). From (S22.10) the computation gives, for h = 0 and h ≠ 0,

where we’ve used u = 0 in the second step, and (1 + γ5)u = 2u in the third. For h = 0,

(The sign changes, because p′ − q′ = −(p − q).) Using the useful relation (S22.15), we find that there are five nasty
amplitudes:
(b) Now we introduce e′, a massive Dirac fermion of opposite charge. This interaction leads to the crossed graph

Table S22.1: The five nasty amplitudes in ν–ν → WW

Everything goes as before, except g → f, m → M, e ↔ e′, q ↔ q′. For h = 0, in addition to the amplitude (S22.19), we
get a new term:

Squaring this we find

in the region of interest. Neglecting the (electromagnetic) mass differences between m2T′ and mT2 (presumed
small in comparison with mπ2),

This has several consequences for the expansion of the amplitude. Recall that we took the invariant
amplitude to be a constant plus something times k ⋅ p, plus O(mπ2), plus pole terms:

The process is invariant under crossing symmetry, which exchanges pion isospins and momenta; the incoming
pion of isospin a and momentum k becomes an outgoing pion of isospin b and momentum −q:

At the level of this expansion, k ⋅ p is odd under crossing of k and q. This means A1ba must be isospin odd:

These are the essential kinematic facts, so the whole amplitude will be invariant under crossing. Now the pole
term, which I’ll evaluate by exploiting Adler’s rule (p. 901). Applied to this process, Adler’s rule says that in the limit
of one pion’s momentum going to zero, the complete amplitude should be given exactly by the gradient-coupling
theory. I’m going to do this backwards from the usual way it’s done in the literature, and start out at the end. We’ll
see what we can discover about Aba from the information we already have, before making explicit use of the
information gained from current algebra, i.e., the commutators of the axial vector currents. For this part of the
discussion, I will assume we have a JP = 1/2+ target, the nucleon or maybe the Σ or Λ. The pole terms, although
rapidly varying, will turn out to be O(mπ2) at threshold, and therefore irrelevant. The easiest way to see that is to
inspect the Feynman diagram. At threshold (again, ignoring the mass difference between T and T′), the incoming
and outgoing pions have the same momentum, q:

The pions are derivatively coupled. These are the pole terms where all the derivatives have been pulled outside of
the time-ordering symbol. At threshold

because, from the point of view of the target, the kinetic energy of the pion is zero. That means there is a Lorentz
frame in which

and hence

Since this is a covariant equation, it must be true in every frame. Because the target is on the mass shell,

so commutes with . We don’t even have to bother rationalizing the denominator to evaluate (42.9). It’s just a
bunch of commuting matrices acting on the eigenvectors u and , and we get

Adding this amplitude ′ to the result for , the nasty terms cancel if f2 = g2. A similar result holds for h = 0.
That is, all the nasty amplitudes become nice if f2 = g2.

42
Current algebra and pion scattering

We continue our investigation of pion scattering. In principle the target could be anything, although at a later stage
we will make it a proton.

42.1Pion–hadron scattering without current algebra

Consider the process

in which a pion plus some target hadron T goes into another pion plus T′, drawn from the same isospin multiplet
as T; for example, these reactions:
A pion comes in with momentum k and isospin a and a pion goes out with momentum q and isospin b. The target
T comes in with momentum p and a product hadron T′ goes out with momentum p′. (I’ll explicitly display the
isospin indices only for the pions.) We are interested in expanding the amplitude about the point where both pions
are soft, k = q = 0, and extrapolating to the closest physically accessible point we can get to: threshold. We are
systematically going to neglect the O(mπ2) terms. In this region, the invariant products of pion momenta are all
O(mπ2),

Figure 42.1: Pion–hadron scattering

We will neglect them because we’re only going to keep terms of O(1) and O(mπ). Thus the only invariant we have
to play with, other than p2 and p′2 (which are kept on the mass shell) is k ⋅ p. You may ask, “Why are you keeping k
⋅ p, but ignoring q ⋅ p?” Because they’re the same, to O(mπ2). From the conservation equation

which is indeed O(mπ2). So the first lesson we learn from Adler’s rule is that the pole terms in this particular
process are totally irrelevant at threshold. (They may be large somewhere else in the region of extrapolation.)

I will now take care of the A0ba by showing that it is also O(mπ2), leaving us with just the A1ba term to compute.
Consider the special case where one pion has zero four-momentum and the other is on the mass shell:

In this case

(aside from pole terms, which we’ve already taken account of). That’s Adler’s rule: only the pole terms are
important when one pion is soft. Because k = 0 means k ⋅ p = 0, (41.64) becomes

and we conclude

If A1ba were also O(mπ2), we would be uncomfortable. It will turn out to be just O(mπ) because of the explicit k ⋅ p.
We have simplified our analysis enormously.

What about scattering off of other targets? Pions are the one hadronic target to which this kinematic analysis
does not apply. In that case everything is of O(mπ2), and we have to keep track of these terms. (We will look at
π–π scattering shortly.) For π−e± scattering we just compute one-photon exchange; there’s no need to consider
the strong interactions.

From (42.18) we can evaluate the amplitude at threshold, neglecting the terms of O(mπ2):

The amplitude at threshold is equal1 to the scattering length aba, times a constant2 proportional to the sum of the
masses of the two particles:

Everything else is suppressed; consider it a matrix on the initial and final states.

We know from (42.8) that A1ba is antisymmetric in b and a. Therefore in particular it must be proportional to
ϵbac times something that depends on the index c. If I look at its matrix element between some initial target and
some final target state, which I’ll indicate just by brackets, it must be an isovector operator, something to take up
the c index, evaluated between the initial target state and the final target state. Since the initial and final target
states are states of an irreducible isospin multiplet, by the Wigner–Eckart theorem,3

That’s forced on us by antisymmetry and isospin invariance: it must be some vector operator by isospin
invariance, and it must be proportional to the matrix element of the isospin operator by the Wigner–Eckart
theorem. So we are left with just one unknown numerical constant to evaluate. I will call4 this constant 2iB.

The expression (42.21) can be made even simpler by observing that the matrix element of the pion’s isospin
is proportional to exactly the same ϵ symbol that appears above; by the right-hand rule for isotriplet states, it’s

Let

where the initial state and the final state, each being two-particle pion-target states, are given by

The matrix element áf|a|iñ can be found by some straightforward algebra. Combining (42.19), (42.20) and (42.21),
we have

Substituting (42.22) for the Levi–Civita symbol, we obtain

The total isospin of the two-particle state is I(I + 1):

(A similar argument is used in calculating spin-orbit coupling.) Assuming we’re looking at the expectation value in
a state of specific total isospin I, I can write Iπ • IT in terms of the total isospin:

(the isospin Iπ(Iπ + 1) contributes the 2, independent of the pion identity). Assembling all of this together we find the
scattering length in a state of specific I as

We still need to determine the constant B, defined in (42.21).

Kinematics and isospin analysis, though pedestrian, have provided us with a great deal of information.
Independent of the current commutators, we have found that the scattering length associated with a pion hitting a
general baryonic target is given in terms of a single coefficient, B. (There may be several scattering lengths
because there may be several possible initial isospins of the two-particle state.)

As a check on our work, let’s compute (42.29) for pion-nucleon scattering, and see if it’s right before we go
any further. After all, this equation makes a definite prediction. If that turns out to be wrong, there’s no point in
trying to compute the coefficient B. There are two possible isospins, I = and I = ; the measured scattering
lengths are5

Their ratio is

The isospin factor in our formula evaluates in these two cases to

Therefore the theory predicts

which is in good agreement with experiment. We hope to do better, but this is encouraging.

42.2Pion–hadron scattering and current algebra

To predict not merely the ratio of the scattering lengths but their magnitudes, we’ve got to determine the value of
that coefficient B. It could still be a disaster. If we compute that B = 0, we would absolutely have no reason to trust
(42.29), because then the observed scattering lengths would be entirely due to terms we have neglected. That the
ratio of the two terms we have retained is −1 : 2 is irrelevant if they’re both zero. Fortunately, we’ll find that B is not
zero.

The computation of B is tedious, so let’s organize it carefully. The amplitude iA is given by the reduction
formula (14.36),

(42.34)

We are interested only in A1ba, the coefficient of the p ⋅ k term; we can extract that coefficient within our
approximation by studying only forward scattering.6 That will simplify the kinematics a bit. Near the end of the
calculation I will say

I’ll keep the relativistic notation, but I want to point out that we can do this calculation in a frame in which the target
is at rest, with the pion also at rest, although possibly not on the mass-shell:

Although we’re going to use those commutators (41.59) from the last lecture, with the δ function in them, in fact all
we need are the commutators of the integrated axial charges. As the pion is carrying off zero three-momentum,
the reduction formula will involve only a trivial space integral.

Now we’ll use PCAC (41.12) to write the pion fields in terms of the divergence of the axial current. Then

where

and a, b are the incoming and outgoing pions, respectively. We now have to pull out the derivatives. Our general
rule for extracting a derivative from a time-ordered product of two fields A, B is always the same:

We’ve got to apply this identity twice. We first take out the y derivative. The order within the product is irrelevant
because of the time ordering.

The δ(y0 − x0) term is not in our current algebra; while we know [A0a(y, 0), A0b(x, 0)], we don’t know what the
commutator of A0a with ∂µAµb is. So we’re going to have to worry about that. However, we don’t really have to
worry very much, because I’ll show you that we can throw away the second term.

In the second term, the spatial part ∂x iAib(x) of ∂x µAµb(x) can be converted through integration by parts to k,
and dropped, because the pion is ultimately at rest; kµ = (k0, 0). The relevant part inside the integral (42.38) is the
spatial integral of the commutator with the time derivative at equal times:

because of the δ(y0 − x0). We don’t know this commutator, but we do know

Thus

The CVC hypothesis (40.2) is that the vector Vµc is the isospin current. Taking that as true, ∂0 of this integral is
proportional to ∂0 of the total isospin, which is zero. Even without CVC, the time derivative of the commutator is
zero:

Thus the antisymmetric part of the second term of (42.39) vanishes. Therefore if the second term is not zero, it
must be symmetric:

If it is zero, there’s nothing more to say. If it is symmetric, it contributes only to A0ba, which we have already
demonstrated is O(mπ2), and therefore irrelevant to the antisymmetric A1ba. In either case we can ignore it.

Onwards to the first term in (42.39). Once more we bring a derivative outside, now with respect to x:

We can neglect the first term: it will have no contribution except from pole terms. All the k’s are on the outside, so
we just get a term of O(mπ2). The second term in (42.44) is going to be the whole package. That’s of course the
commutator we know. More precisely, we know it for ν = 0, but that’s all we really need; the ∂y ν derivative is going
to be turned into a momentum and, as before, the spatial parts of all the pion momenta are zero. Just to keep
things looking covariant, though, we’ll write it as

simply for notational convenience; the space components of this will not contribute to the expression we’re going
to feed it into. The ∂ν acting on the Vν(y) gives us 0 because the vector current is conserved. Only its action on the
δ function is relevant:

Now we’re in business. We substitute (42.46), the only term that survives in the double divergence (42.39), into
the integral (42.38) and compute it:

(42.47)

Integrating the right-hand side by parts,

 (42.48)

If we now set k = q, the integral in the reduction formula, normally a mess, becomes trivial:7

because the kets |pñ are relativistically normalized (1.57), and the integral of the currents gives 2gV times the
isospin operators (41.52). By overall momentum conservation, k0 − q0 = p′0 − p0, so

We now substitute (42.50) into (42.37) via (42.38), and set k = q:

If we’re only interested in expanding in the range near k = 0, keeping linear terms, we can replace k2 by 0, whence
the mπ2 in the numerator cancels the mπ2 in the denominator. The i 2 gives me a minus sign. We’re going near k =
0 to get the coefficient of the k ⋅ p term. Canceling the common factors, we arrive at the amplitude:

From (42.18) and (42.21) we have

And so8

A constant useful in these calculations is

(the second equality coming from the Goldberger–Treiman relation (40.37).) In terms of L,

We can now plug this expression for B into (42.29) and get an expression for the scattering of a pion off of
any hadronic target with the exception of anther pion; that one we can’t do. Thus we obtain the universal
expression for the scattering length, called the Weinberg–Tomozawa formula,9

Note that the coefficient of the isospin factor is a universal number depending only on the mass of the target.10

The actual predictions are useful only for the pion–nucleon case, where we have good measurements of the
scattering lengths. Plugging in the numbers, we find for pion–nucleon scattering

This is excellent agreement,11 and a triumph of current algebra. We have verified the A0 with A0 commutators;
indeed we have used them to explain experiment.

The physics is everywhere: in crossing symmetry, in the assumption of Goldberger–Treiman-like smoothness

(which enables us to assume the amplitude has a smooth extrapolation, a polynomial fit, up to threshold, once the
two-pion poles are extracted), and finally in the current algebra commutators. The only new information above and
beyond the sort of reasoning we used in the Goldberger–Treiman relation is the current commutators. As we tried
to emphasize by the way we’ve organized this lecture, or disorganized it, even without the commutators we get
(42.33) the −2 : 1 ratio between the two scattering lengths. The current commutators just serve to fix the scale of
the scattering length.

42.3Pion–pion scattering

Let’s go on and consider the one case which has slipped through our net, π–π scattering, following a famous
analysis of Weinberg.12 The analysis is beautiful because it involves no new physics and no new integrals to be
computed. But it requires practically everything we know about the π–π system. It uses our extrapolation
techniques where we assume there is a polynomial fit. It uses current algebra every inch of the way. It uses some
new information about a commutator that we’re going to have to extract from the quark model. And it uses
crossing and analyticity and everything else. It’s a beautiful calculation.

We will end up predicting two scattering lengths. The two pions can be in an I = 0, an I = 1 or an I = 2 state. By
Bose statistics the I = 1 state is p-wave, so it has no defined scattering length.13 We will end up deducing a0 and
a2. The trick will be to consider the process in which a pion comes in with momentum k1 and isospin a and goes
out with momentum k′1 and isospin b, and a pion with momentum k2 and isospin c comes in and goes out with k′2
and isospin d. We will investigate this process near the point where all the momenta vanish and retain all the
terms of O(mπ2). We will then extrapolate to threshold as before. The nice thing about this problem is that there
are no pole terms. A pole term—a three-pion pole, or more properly an axial vector current π–π pole, as in Figure
42.3—is forbidden by parity. 14 Since everything is going to go off mass shell in this computation, we have a
straight expansion in terms of six invariants. It’s convenient to begin with the over-complete set of seven
invariants, consisting of the individual momenta squared, ki2 and k′2i, i = {1, 2}, and our usual Mandelstam
variables s, t and u, which, just to remind you, are

Figure 42.2: Pion–pion scattering

Figure 42.3: Pion pole in pion–pion scattering, forbidden by parity conservation

This set is linearly related by the kinematic identity

Now let’s write down the isospin-invariant amplitude. It has to carry the isospin indices of the four pions, and it
has to be invariant under the interchanges a ↔ c, b ↔ d, a ↔ b, and c ↔ d. Thus it must take the form

The brackets are to be filled in. The things in them are all connected to each other by crossing. When we know the
coefficient that appears in the first one we’ll know the coefficient that appears in the other two. We’ll go to first
order in our seven invariants.

Let’s first worry about whether we can have possible coefficients of k12 in the first term. By Bose statistics,
since this process is symmetric under interchange of a and c, if we have a k12 then we must also have a k22. By
time reversal (or equivalently, by crossing ab into cd), if we have k12 and k22, we’ve got to have a k1′2 and a k2′2
with the same coefficient:

That can be written in terms of s + t + u.

So rather than write the bracket contents in terms of the k’s, I’ll start over and use all possible linear combination of
s, t and u.

What can we have? We can certainly have a constant term which, to make it have the same dimensions as
everything else, we’ll call Amπ2. Then we can have a term proportional to s in the term where a and c are
symmetric, interchanging particle a and particle c, and keeping everything else fixed; this exchanges t and u, so
we must have a coefficient times t + u. All the other terms will involve st or momentum transferred squared or
momentum squared times s or t or u, or s2 or t2 or u2. But those terms are O(mπ4), so we can neglect them. In the
amplitude (42.61), when we apply all the constraints imposed by Bose statistics, crossing, and isospin we have
only three unknown numerical coefficients to determine:

The other terms are connected to this by crossing. In the δab term we permute15 the Mandelstam variables as (s t
u), and in the δad term we permute them as (s u t). Thus we have

Thus the entire amplitude, excluding terms of O(mπ4), is known in terms of three coefficients. The task now is to
compute them. We’re going to need three equations to solve for these coefficients.

The first comes from Adler’s rule. Imagine that one incident pion, say with momentum k1, is soft, so

According to Adler’s rule, the amplitude for soft pions vanishes except for pole terms. But there are no poles in this
process, so the amplitude at this point must vanish:

as in the previous analysis. This is also the point where

from the definitions of the variables (42.59). All of the quantities in square brackets are equal, so we obtain our first
equation:

Next consider what happens when we look in the forward direction, with pion 1 (with momentum k1) soft, and
pion 2 (with momentum k2) on mass shell:

We know in that case we get terms of O(k2) which come from that symmetric commutator, (42.42). Let’s write it
down. There will be a constant term about which we can say nothing. There is a term which we explicitly
computed from the current commutator,
where the target in ITf is now just the second pion. That’s simply our earlier computation (42.52). If we are allowed
to make k and k′ as small as we want, that part of the computation works whatever the target particle is, even a
pion. A little isospin algebra is necessary. By the previous analysis (42.22)

where c and d are the initial and final pions, respectively. The product of two ϵ’s is easy to compute (37.47), and A
takes the following form:

We now compare this to (42.65). In forward scattering the amplitude is easy to compute. We have for the
Mandelstam variables with k1 = k1′

and retaining terms of first order in k1 and k′1 (and using k2 = k′2),

In this régime the amplitude (42.65) is

Is this consistent with (42.73)? Sure enough, it is. There’s a constant part, and there is a term proportional to (k1 ⋅
k2) multiplying (δacδbd − δadδbc), exactly as predicted. Comparing the coefficients in (42.73) and (42.76), we
obtain our second equation:

Our third equation comes from a constraint on the constant term in (42.71). In our earlier analysis, we found
three terms in the amplitude A1. One had two derivatives explicitly on the outside, which, in the absence of pole
terms, vanishes when two of the moments vanish. Another was the commutator term, which we’ve just taken care
of. Finally, we had that funny term (42.40), which we argued was symmetric in a and b (42.43),

We’re going to have to study this term in the context of a particular model and see what happens. We have one
obvious model at hand. Let’s see what we can determine about this thing from the quark model, (§39.4 and
§41.5).

The currents are bilinear in the quarks (41.56), which carry isospin and color. Color interacts with the gluons,
about which I’ve said next to nothing. Gluons are supposed to carry color only; they are presumed to be isospin
neutral and flavor neutral. The divergence of Aµ is not going to be zero, but it surely is going to be a bilinear form
in the quark fields; one derivative on a (Dirac) quark field equals the quark field multiplied by something. So
∂µAµb(y) is going to be bilinear in the quark fields. A0a(x) is likewise bilinear in the quark fields; it’s just the quark
current. When you commute two bilinear fields you obtain another bilinear. That’s the way it has to be, because
the commutator produces one field out of the difference of two products of two fields. In any version of the quark
model we can think of where the gluons have only color, (42.78) must be bilinear in the quark fields.16 The up and
down quarks carry isospin 1/2, while the strange quark is isospin 0. So the bilinear resulting from (42.78), built from
u and d quarks only, must be I = 0 or I = 1. But I = 1 is ruled out; that would give an ϵ symbol17 and we proved
earlier that (42.78) is symmetric under the interchange of a and b. Therefore within the context of any sensible
quark model the constant in the power series expansion (42.73) must be pure I = 0 in the indices a and b:

That’s the only piece of information we will extract from the quark model: the constant term in the power series
expansion (42.76) for the forward scattering amplitude A, with one pion soft, another pion on the mass shell, must
be proportional to δab. That means that the δacδbd and the δadδbc terms which aren’t proportional to δab have got
to vanish. Thus we find our third equation,

Putting together all three equations (42.69), (42.77), and (42.80), we quickly solve for the constants A, B, and C:

We have determined everything.18

Our final expression for the scattering amplitude (42.65), which we know is valid near zero and hope is valid
all the way up to threshold, is

with the coefficients given by (42.81). Notice we needed the extra assumption (42.79) which we have extracted as
an unambiguous consequence from the quark model.19 Equal time commutators are among the few things we can
compute without solving strong interaction dynamics.20

Let’s find the presumed actual π–π scattering lengths, to compare with experiment. We need to evaluate the
amplitude at threshold. Which term we choose to evaluate at threshold is irrelevant because the amplitude has
crossing symmetry, but we’ll choose the usual one,

This gives the threshold amplitude

That’s the final answer for the amplitude! To write this in terms of scattering lengths, all that remains is some easy
kinematics and a little isospin analysis to find out what this obviously isospin-invariant 9 × 9 operator, turning initial
states into final states at threshold, is in terms of its eigenvalues; that is, in terms of its I = 2 eigenvalue and its I = 0
eigenvalue. First I’ll do the isospin analysis.

42.4Some operators and their eigenvalues

We already know one of the operators. Recall that

Other than a minus sign, this is the product of two isospin matrices. We may call the top line (with isospin indices a
and b) in Figure 42.2 “the π” and the bottom line (indices c and d) “the target”, as is our privilege. Hence we have
from (42.28)

Note that the last term inside the brackets is −4, because now both the pion and the target have isospin 1; I(I + 1)
is 2 for each of them. Had the target been a proton, that −4 would have been − ( + 1) − 2 = −2 . We can write
down the eigenvalues of the operator in (42.85) from (42.86), where I is now the total s-channel isospin. This
operator will have eigenvalues depending on I. The possible s-channel values for I (from the sum of the two pions’
isospin) are I = {0, 2}; I = 1 is ruled out by parity invariance. Then

That’s one operator that occurs in this decomposition (42.84). It’s the difference of one of the three δacδbd and
δadδbc. It happens our states are s-wave states at threshold, so they are symmetric under the exchange a ↔ c. If
we swap a and c we get the operator
which, acting on states symmetric in a and c, is of course the same thing as (42.87), and therefore it has exactly
the same eigenvalues; it is the same operator on these states. That takes care of two of the three operators in
(42.84).

The last operator we have to deal with is the remaining δacδbd. This is obviously a pure I = 0 operator: if you
rotate just the isospins of the original a and c and do nothing to the final states, it doesn’t change: it is isospin-
invariant. Therefore its I = 2 eigenvalue is 0. And its I = 0 eigenvalue is easily obtained. Take an isospin 0 initial
state. It has some isospin wave function

If we apply this last operator δacδbd to the matrix element (42.89), and sum over the initial indices

because δacδac summed over a and c is 3. So it reproduces that isospin 0 state in the final state as it should,
being an isospin-invariant operator, with a coefficient that is 3 times the original coefficient of δac in (42.89); the
eigenvalue here is 3.

We can now write the isospin-dependent part of (42.84) as

I summarize the eigenvalues of these operators, and the amplitude, in Table 42.1. The amplitude at threshold is
just the sum f(I) of these three operators, times the constant Bmπ2:

Table 42.1: Eigenvalues of all isospin operators

These are not numbers21 of which we can say, “Well, obviously they have to be −2 and 7!”

The function f(I) does not have the

structure (42.29) that we found for the scattering lengths for any hadronic target other than a pion. In (42.86), the
(−Iπ • IT) term, equal to − [I(I + 1) − 4], does follow this pattern (with I = 1 for a pion target), but the δacδbd term in
Table 42.1, which makes a large contribution to the I = 0 amplitude, does not. That’s because their origins are very
different. The first two entries in Table 42.1 came from the current algebra commutators. The third entry in the
table comes from the symmetric term (42.43) we neglected in our previous analysis, because it was of O(mπ2).

There is a small correction22 in the relation (42.20) connecting the amplitudes A with the scattering lengths a
in the situation where the target and the incoming particle are the same kind. Now the constant 8π(mπ + mT) is
increased by a factor of 2, so that

With this correction,

where L ≈ 0.1mπ-1 is given by (42.55). So the theoretical predictions for the I = 0 and I = 2 scattering lengths are
Are the experimental numbers in the ratio of −7 : 2? The old values of 0.20mπ-1 and −0.06mπ-1 certainly were, to
within roundoff errors; the current values, though consistent with the theoretical predictions, are not so nice.23

Unfortunately, π–π scattering is not an easy process to study experimentally. The best handle on it is
obtained by considering pion production in pion–nucleon scattering,24 as shown in Figure 42.4. A π comes in, a
nucleon comes in and you look at the pion pole. You try to extrapolate to that pion pole, which is not in the physical
region, and from that deduce something about low-energy π–π scattering. The fairest thing to say is that what we
obtain is consistent with what we deduce on theoretical grounds, although there is lots of room for error and even
more room for arguments about the validity of the method.

Figure 42.4: Pion production

This terminates our discussion of current algebra. There are many more things that could be said, and many
more good things you can do with current algebra, many more soft processes you can analyze involving soft
pions. Some of the nicest are analyses of β decay processes involving the emission of one pion, where you can
obtain again soft pion rules. They are a bit different in structure from what we have done here because you
explicitly want to use the structure of the weak interactions. So you study a matrix element

Here H w is the non-leptonic Hamiltonian if it’s a non-leptonic process, or the weak current if it’s a leptonic process.
And you could relate it to scattering lengths by exactly the same tricks we have been using here. So in a typical
analysis of this kind, instead of one pion and a commutator of two axial vector currents, you have to worry about
the commutator of an axial vector current and a weak interaction Hamiltonian. But otherwise the tricks are much
the same and the analysis is much the same. In this manner it is possible to learn quite a lot about leptonic decays
of kaons25 which go into leptons plus pions, either one π or two; both are observed. And it is possible to learn a
great deal about s-wave nonleptonic hyperon decays since the process is parity-violating. In a process like

the pion can appear in either the s-wave or the p-wave. About the p-wave process we learn nothing by these
methods because that automatically vanishes when the pion momentum goes to zero. Despite statements in the
early literature to the contrary, we know as little from current algebra about p-wave nonleptonic hyperon decays as
we do about p-wave scattering lengths in π–π scattering, to wit: nothing. But the s-wave ones can be computed by
these methods, and the results are in good agreement with experiment. But we won’t have time to go into that in
these lectures. See the literature for these results; there are numerous good review articles widely available in
addition to various books.26

1 [Eds.] The scattering length a is closely related to the s-wave phase shift δ0. Two definitions of the scattering
length, differing in sign, occur in the literature. The first, δ0 = −ka, appears in many quantum mechanics texts: K.
Gottfried, Quantum Mechanics, Volume 1: Fundamentals, W. Benjamin, 1966, equation (40), p. 393; Landau &
Lifshitz, QM, p. 501, equation (130.9); and note (‡), p. 502; A. Messiah, Quantum Mechanics, v. 1, North Holland
Publishing, 1962; reprinted by Dover Publications, 2014, equation (X.47), p. 392. The second, δ0 = +ka, is widely
used in the phenomenological analysis of high energy π-N scattering data: M. L. Goldberger and K. M. Watson,
Collision Theory, John Wiley, 1964; reprinted by Dover Publications, 2004, equation (296), p. 287. This latter
definition is used in (42.20).
2 [Eds.] The derivation of this constant is given in Appendix 3 of “Soft Pions”, Chapter 2 in Coleman Aspects, pp.
64–65.
3 [Eds.] See note 28, p. 766 and note 6, p. 847.
4 [Eds.] In the video of Lecture 47 and in Aspects, this constant B is defined by ba = ϵbacáITc ñB. However, later
1
on a second constant B is introduced. With the definition (42.21), the two B’s are one and the same.
5 [Eds.] J. Hamilton and W. S. Woolcock, “Determination of Pion–Nucleon Parameters and Phase Shifts by
Dispersion Relations”, Rev. Mod. Phys., 35 (1963) 737–787; D. V. Bugg, A. A. Carter, and J. R. Carter, “New
Values of Pion–Nucleon Scattering Lengths and f2”, Phys. Lett. B44 (1973) 278–280.
6 [Eds.] Forward scattering occurs when the initial and final states are the same; Itzykson & Zuber QFT, §5.3.1.
7 [Eds.] There are a couple of non-obvious steps in (42.49) going from the first expression on the right to the last.
First, Vµc (x) can be replaced by Vµc (x) because it is a conserved current (see note 24, p. 903). This allows the dx0
integration to be done only over the remaining exponential, giving the delta function 2πδ(x0 − y0). Next, the
relativistic normalization says that áp|p′ñ = (2π)32p0δ(3)(p − p′). That means

One then argues by Lorentz invariance that if the index 0 is replaced by ν on both sides, the equation remains
true.
8 [Eds.] In the video of Lecture 47, Coleman now spends several minutes trying to repair what he erroneously
believes to be a sign error. The confusion, due to different sign conventions in the scattering lengths a, was
resolved in the next lecture.
9 [Eds.]
S. Weinberg, “Pion Scattering Lengths”, Phys. Rev. Lett. 17 (1966) 616–621; Y. Tomozawa, “Axial-Vector
Coupling Constant Renormalization and the Meson–Baryon Scattering Lengths”, Nuovo Cim. 46A (1966)
707–717.
10[Eds.] The value for a proton target is about 0.085mπ–1, using Fπ = 0.197mp, mπ = 140 MeV, mp = 939 MeV,
and gV ~ 1.
11 [Eds.] Bugg et al., op. cit.
12 [Eds.]
Weinberg, op. cit. (note 9, p. 925); Coleman Aspects, “Soft Pions”, pp. 57–59. Much of this section follows
Weinberg’s argument very closely. See also Weinberg QTF2, pp. 197–202.
13 [Eds.] The scattering length is defined in terms of the s-wave phase shift. See note 1, p. 920.
14 [Eds.] Figure 42.3 does not appear in the video of Lecture 47 (upon which this chapter is based) nor in any
lecture notes. This is only a guess at what Coleman meant by the term “three-pion pole”.
15 [Eds.] Recall that the notation (a b c) means the cyclic permutation a → b, b → c, c → a. See the paragraph
before (39.48), p. 862.
16 [Eds.] In response to a student’s question, Coleman replies, “This hypothesis was originally made in the context
of the sigma model, which we have not discussed yet. Nobody knew about quarks with flavor and color then. They
knew about quarks but they didn’t understand why they had funny statistics. And they didn’t take the quark model
seriously.” (The sigma model is discussed in §§45.3–45.4, pp. 993–1002.)
17 [Eds.]The argument can perhaps be fleshed out a little. Weinberg (see note 9, p. 925) writes his equation (5) in
the context of both the σ and free-quark models,

where σab(x) is “some scalar field which may or may not have something to do with a real 0+ π–π resonance or
enhancement, and ‘S.T.’ means possible Schwinger terms.” The only two two-index isospin tensors are δab and
ϵab, so σab(x) ~ δabf(x) + ϵabg(x). Under parity, f(x) has to have I = 0 and g(x) has to have I = 1. As the commutator
was earlier shown to be symmetric, we take g(x) = 0. (In Aspects, p. 59, Coleman states that without the
assumption that the commutator is proportional to δab, it “could be any combination of I = 0 and I = 2”.)
18 [Eds.] The signs of A and B here are reversed from those in the video of Lecture 47, apparently due to the
differences in the definitions of the scattering length. They agree however with the signs in Coleman Aspects, “Soft
Pions”, with the exchange B ↔ C. Note that this value (42.81) of B agrees with that in (42.54).
19 [Eds.] In the original derivation, Weinberg op. cit. says (before his equation (4)) that the commutator is
“suggested by the σ model, and the free-quark model”. See note 17, p. 930.
20 [Eds.]The video of Lecture 47 ends here. The rest of this chapter is taken from the first 21 minutes of Lecture
48’s video, in order to keep this material on soft pions in one chapter, and begin Chapter 43 with the Lecture 48
material on spontaneously broken symmetry.
21 [Eds.] Weinberg, op. cit.; Lee, op. cit., Section 10b, pp. 76–77.
22 [Eds.] Coleman Aspects, “Soft Pions”, Appendix 3.
23 [Eds.] These values, especially a2, have changed markedly over the years. For more recent determinations of
the scattering lengths, see M. V. Olsson, “Rigorous Pion–Pion Scattering Lengths from Threshold πN → ππN
Data”, Phys. Lett. B410 (1997) 311–314; S. Gevorkyan, “Pion–Pion Scattering Lengths Determination from Kaon
Decays” in 7th International Workshop on Chiral Dynamics, Newport News, Virginia, 2012; I. Caprini, “Theoretical
Aspects of Pion–Pion Interaction” in International Conference on QCD and Hadronic Physics, Beijing, 2005. For a
history to pion scattering lengths, from Yukawa’s original hypothesis up through 2008, see J. Gasser, “On the
History of Pion–Pion Scattering” in International Workshop on Effective Field Theories: from the pion to the
upsilon, Valencia, Spain, 2009. Weinberg QTF2, p. 202 gives a0 = (0.26 ± 0.05)mπ, a2 = (−0.028 ± 0.012)mπ.
24 [Eds.] Olsson, op. cit.
25 [Eds.] Coleman Aspects, “Soft Pions”, Section 10, pp. 60–62.
26 [Eds.] For a brief introduction to current algebra, see D. H. Lyth, An Introduction to Current Algebra, Oxford U.
P., 1970. See also note 23, p. 902.

43
A first look at spontaneous symmetry breaking

I would now like to begin a brand new subject, one that will come back to current algebra, as well as leading us in
new directions towards renormalizable theories of the weak interactions. The subject (which will occupy us for
several lectures) is spontaneous symmetry breaking.1 I’ll begin with a few general remarks, and a parable.

43.1The man in a ferromagnet

It is a truism in non-relativistic quantum mechanics that the ground state of a system may not be invariant under
the symmetry group of the Hamiltonian of the system. In the case of the hydrogen atom, the potential is rotationally
invariant and the ground state is also; it’s an s-state. But for more complicated systems involving many particles,
this need not be so. For example, nuclear forces are rotationally invariant, but it is not true that all nuclear ground
states are rotationally invariant.2

For a nucleus this is no problem. But for a system of infinite spatial extent it can lead to strange things. The
standard example is the Heisenberg ferromagnet. In an idealized sense, a ferromagnet is an infinite crystalline
array of little dipoles.

Figure 43.1: Ground state in a Heisenberg ferromagnet

The law of interaction between the dipoles (the Heisenberg exchange force) says that neighboring dipoles like
to line up; the Hamiltonian is given by3

(the sum is only between nearest neighbors and the coefficients Jij are exchange coupling constants); the energy
is minimized when the spins all align. This interaction is rotationally invariant: it depends only on the relative angle
between the spins and so is unchanged if we rotate all the spins together. At zero temperature the ground state of
the ferromagnet looks like Figure 43.1—going on forever, if we imagine the ferromagnet to be infinite in
extent—with the net magnetization4 pointing in some direction. The ground state is a state of maximum spin. If
there are N spin- dipoles in the ferromagnet (with N typically on the order of Avogadro’s number), the spin of the
ground state is ℓ = N ⋅ , and so the state is 2ℓ + 1 = 2(N ⋅ ) + 1 = (N + 1)-fold degenerate. In the limit of an infinite
ferromagnet we can think of the ground state as infinitely degenerate, labeled by a continuous vector that can
point in any spatial direction, not a quantum vector with only a finite number of directions.5 Although the interaction
(43.1) is rotationally invariant the ground state is not; the total spin points in some preferred direction. The
symmetry of the interaction is SO(3); once the direction of the magnetization is set, the symmetry of the ground
state is reduced to SO(2), for rotations about the direction of the ground state:

The parable6 involves a man who lives inside the ferromagnet. He’s considerably larger than an individual dipole
so he can’t see the granular structure; it looks like a continuum to him. But he’s considerably smaller than the
ferromagnet, which may be a million light years on a side. We are in telephone communication with this man. We
tell him, “Physicists in the outside world have made a sensational discovery: the laws of nature are rotationally
invariant.” And he says: “You’re crazy! There’s this enormous magnetic field pointing north; it tries to pull the
fillings7 out of my teeth! The laws of nature are definitely not rotationally invariant; there’s a preferred
direction—north—and nature could not be more asymmetric.” And we say: “No, no, no! You think that because
you’re living in a big ferromagnet. The laws of nature are really rotationally invariant, even the laws of nature for a
ferromagnet. It’s just that neighboring dipoles like to point in the same direction. Your experiences are influenced
by the resulting strong magnetic field. If you had been living in a different ferromagnet, all the dipoles might have
chosen to line up in another direction, east for example. Then the force acting on your fillings would point east
instead of north.”

The man thinks for a moment, and replies. “All right, I’ll try to verify that experimentally.” So he decides to
change the orientations of the neighboring dipoles. If there is no preferred direction, then having all the dipoles in a
new direction should be just as good a ground state with just the same energy, and the total cost of moving to that
ground state should be zero. Figure 43.2 shows a few rotated dipoles.

Figure 43.2: The ground state perturbed by rotating some dipoles

Of course it takes some energy to rotate each dipole. The magnet is in three dimensions so there are many
boundaries in which some of the dipoles are changed in direction, and therefore the state gets a new energy. He
keeps on working. His altered domain keeps growing in surface area, which is where he’s losing energy, until he
reaches the boundaries of the ferromagnet. Things don’t look too good. If the ferromagnet is infinite, he can never
reverse all of the dipoles and get his invested energy back. If the ferromagnet is finite, say with periodic boundary
conditions so we don’t have to worry about sharp boundary effects, he begins gaining energy again when he’s
reversed half the dipoles in the ferromagnet but that’s still quite a lot of dipoles to reverse.

To the man in the magnet, the universe doesn’t look at all as though it is rotationally invariant. There’s no
easy experiment he can do that will reveal this rotational invariance to him. If he understands a lot of deep physics
and the laws of ferromagnets, then he can say “Oh yes, the universe is rotationally invariant. It just happens that
I’m living in a ferromagnet which has settled down in a particular direction.” But if he doesn’t understand the
physics he’ll never believe it.8 The rotational invariance of the Hamiltonian that governs the magnet in which he
lives is completely hidden from its occupant. As far as he is concerned, it’s just as if there were a large rotation-
violating term in his Hamiltonian. That is how symmetry is normally broken. But that’s not what happens here: the
Hamiltonian is perfectly symmetric; it’s the dynamics that causes the ground state to be asymmetric. The standard
terminology is unfortunate. We don’t describe the symmetry as hidden; instead we call this situation spontaneous
symmetry breaking. In these circumstances we say, “The symmetry is spontaneously broken”.

As you may remember, in the beginning of this course I said we always need an assumption when quantizing
a theory: that the ground state, the vacuum, is symmetric under whatever group the Hamiltonian is.9 The
assumption is easy enough to confirm in the case of a free field, but rather difficult to check otherwise. Now we
see why the man in the ferromagnet is a parable. The little man is us and the ferromagnet is the universe. It could
be, if the dynamics turned out right, that there is some symmetry that is possessed by the Lagrangian but not by
the ground state. It would be no easier for us to determine that such a symmetry held for the laws of nature than
for the man in the ferromagnet to discover his physical laws are rotationally invariant. That symmetry would be
completely hidden from us. Of course the symmetry in question is not one we’re familiar with. The symmetries we
know about are all manifest; they’re not hidden. The hidden ones are not rotational invariance or Lorentz
invariance or isospin invariance. Maybe there’s something else that nobody has ever thought of, at least until now,
because, despite not being symmetries of the ground state, they are in fact symmetries of nature, though hidden.

Whether the man in the ferromagnet does not see rotational invariance at all, or perceives it as a weakly
broken symmetry, depends entirely on how much he interacts with the magnetic field. If he and his apparatus and
everything he measures are made of polyethylene, for example, then for all practical purposes his world is
rotationally invariant. On the other hand, if he and everything else are made of iron, he will probably never suspect
the rotational invariance of his world. So spontaneous symmetry breaking can simulate a Hamiltonian with
everything from total asymmetry to weakly broken symmetry.

How might we discover these hidden symmetries? Three avenues come to mind. We could consider high
energy, ω. At high momentum transfer, we might see the symmetries we don’t see now. Or we could appeal to
high temperature, T. For example, the magnetization of the ferromagnet is lost above its Curie temperature (when
the spins are randomized) and its rotational symmetry is restored. We might not be able to do it in the laboratory,
but we could look at the early universe. Finally, we could apply high IQ. This is the route chosen historically by
Weinberg, Salam, and Glashow, among many others.

43.2Spontaneous symmetry breaking in field theory: Examples

We’ll begin investigating the possibility of a symmetric Lagrangian with a non-symmetric ground state, in the most
primitive way. We will look at some classical field theories involving only scalar fields, and discuss the classical
analogs to the phenomena we’ve been alluding to here. The examples we will discuss will be classical field
theories involving a set of real scalar fields, which we will assemble into a big n-vector

with a non-derivative interaction U(Φ):

Although these will be classical theories I will use quantum language: I will call the state of lowest energy “the
vacuum” and the parameters that govern the spectrum of small oscillations about the vacuum as “particle
masses”, etc. I will later redeem this abuse of language by showing that for this class of theories, these classical
descriptions can be thought of as the first term in a systematic quantum perturbative expansion.

The energy density of this theory is

where the (∂0Φ)2 and the (∇Φ)2 terms are summed over all the fields in Φ, of which there can be many. If the
theory has a state of lowest energy, U must be bounded below. I will add a constant to U so that it’s always
greater than or equal to zero, and attains zero for some value of Φ:

From (43.5) we see that the minimum energy state will occur when the non-negative terms (∂0Φ)2 and (∇Φ)2 are
as small as they can be, i.e., the ground state Φ is a constant, independent of both time and space. We will denote
this state by

and refer to it as the vacuum expectation value, or VEV for short, of the scalar field Φ, even though we are
dealing with a classical theory. The ground state áΦñ must be a minimum of U, and hence a zero of U because of
the way we’ve defined the potential:

If U has a unique minimum or zero then the ground state is unique. If it has several minima there are several
possible ground states.

EXAMPLE 1: Discrete symmetry

 Let’s consider an extremely simple example, where Φ is a single field ϕ.

(λ is a coupling constant; the unspecified constant will be determined so that U(ϕ) ≥ 0). The symmetry group10 is
just the cyclic group Z2; the symmetry is reflection, ϕ → −ϕ. The ground state is invariant under this symmetry and
the symmetry is manifest; it is not spontaneously broken. In order that the potential be bounded below, λ must be
positive:

Despite the choice of symbol, µ, which suggests a real, positive mass, we will put no restriction on µ2 and consider
either positive or negative µ2:

Case 1: µ2 > 0. The potential is strictly concave up as shown in Figure 43.3 and the ground state is unique;

If we do small oscillations about the ground state we find the energy-momentum relationship, or rather (since we
are doing the classical theory) the frequency-wave vector relationship, characteristic of a particle of mass µ. Here,
µ actually is the mass of the “meson”:

Figure 43.3: Single well

Figure 43.4: Double well

Case 2: µ2 < 0. The situation is dramatically different. As shown in Figure 43.4, in this case the potential points
down at the origin, because µ2 is negative, but eventually it turns up. There are two points, ±a (which we will
compute shortly), where the minima occur. These points occur symmetrically because of the invariance of U
under ϕ → −ϕ. We can thus write

The constant in (43.9) has been used to shift U up so that U min = 0;

The value of a can easily be determined by comparing the quadratic terms of (43.14) and (43.9):

so that
Remember, µ2 is a negative number so this is perfectly reasonable. We have two degenerate minima. Which of
these we choose to be the ground state (about which we do perturbations) is arbitrary. Any statement that we can
make about the physics as seen by a man living at +a or as seen by a man living at −a are easily transposed, one
into the other, by symmetry. But whichever one we choose, the symmetry is spontaneously broken. The ground
state of the theory is not invariant under the symmetry group of the Hamiltonian.11

We can explore the consequences of spontaneous symmetry breaking by shifting the field. Then we can read
off what happens for small oscillations about the ground state. Which minimum we choose is irrelevant. I’ll shift to
áϕñ = +a. Define a new field, ϕ′ (not to be confused with the renormalized field in the sense of quantum field
theory):

Then

When we expand this we have no terms linear in ϕ′, so we are indeed at a minimum of U. The actual mass
squared of the meson is the coefficient of ϕ′2 (the term that governs the energy-momentum spectrum of small
oscillations). It is obtained by expanding (43.18):

So µ2 is not the square of a mass, although the mass that eventually emerges is connected to it. Also we see that
someone living in the minimum of this potential would see nothing like a ϕ → −ϕ symmetry; this symmetry is
hidden from him. He would say, “I live in a very simple world in which there is only one meson, one kind of particle.
It doesn’t appear to obey any particular symmetry; ϕ → −ϕ is certainly not a symmetry here, because there’s a
cubic coupling in (43.19) of the meson with itself.”

EXAMPLE 2: Continuous symmetry and Goldstone bosons

Let’s consider a theory based on exactly the same principles but with a continuous internal SO(2) symmetry.
We begin with two scalar fields, ϕ 1 and ϕ 2. We’ll look at something we have already cooked up so that
spontaneous breakdown is guaranteed to occur (again, µ2 < 0)

The surface of revolution of this potential is a “Mexican hat” potential, sometimes called the “wine bottle” (or
“champagne bottle”) potential.

Figure 43.5: “Mexican hat” potential

This is invariant not just under ϕ i → −ϕ i but under any rotation in the ϕ 1 − ϕ 2 space:
The minimum occurs for any ϕ 1 and ϕ 2 that lie on a circle of radius a:

See Figure 43.6, where the possible ground states, the points in the trough, are indicated on the dotted circle.
Which point on the circle we choose to be the ground state of the theory is arbitrary. The ground states are
degenerate; they all have the same energy. But once we make our choice, the ground state loses its rotational
invariance: we have spontaneous breakdown of the symmetry.

For convenience, we will choose the ground state to be

and define shifted fields

The algebra is the same as before:

The symmetry has been completely obscured. Someone arriving in this world would have no idea that it has a
hidden internal SO(2) invariance. If we compute the masses, m12 is the same as before; m22 however is a
surprise: there is no term quadratic in ϕ′2 in (43.25), because there’s no constant for ϕ′2 to be multiplied by when
we take the square:

That is, ϕ 2 describes a massless particle!

Figure 43.6: Top view of the Mexican hat potential; ground states (heavy circle) at a =

Zero smells like a sacred number. Whenever it occurs we should sit up and pay attention, unless we’ve put it
in to begin with. Something like 2a2λ is just a number. But zero carries the scent of something general going on
here. Indeed there is. We will demonstrate that this happens not just for this particular form of the potential but for
any rotationally invariant potential involving ϕ 1 and ϕ 2. And then we will generalize it even further.

If ϕ did not point along one of the axes, we would obtain ϕ 1ϕ 2 cross terms in the Hamiltonian’s quadratic
form. We’d need to diagonalize that for the small oscillations in ϕ′1 and ϕ′2, but we’d obtain the same results: one
would be massive and one would be massless. That’s guaranteed, since the fields are connected by the
symmetry. All the convention-independent physics must be the same. It’s just the labels, what we call the massive
particle and what the massless, that may differ.

The easiest way to see that there is something general going on is to make a change to angular variables.
Define new fields, ρ and θ:
This is a lousy choice of coordinates were we interested in expanding things about ρ = 0: it’s singular at ρ = 0. But
we’re not; and for studying small vibrations about any other point, it’s as good as any other choice.

The SO(2) symmetry is characterized by some angle α. In these variables, the symmetry is realized as

The Lagrangian looks rather non-canonical:

The potential U is now some general rotationally invariant function; it depends only on ρ and not on θ. We’ll
assume

where a is some number determined by the shape of the potential. The ground state is

It lies anywhere on the circle ρ2 = a2 just as before.

Define new variables

Then the Lagrangian is written in terms of the ρ′ and θ′ fields, each describing a scalar meson:

We have cunningly chosen our coordinates so that the quadratic parts of turn out to be diagonal. We see that
the ρ′ meson has some mass that depends on the second derivative of U at the minimum of the potential, about
which we can’t say anything until we know what the potential is. The θ′ meson is guaranteed to be massless. If we
only keep the quadratic term a2(∂ µθ′)2, it’s a perfectly normal meson. It has a funny a2 factor, but that’s a
constant and we can always absorb it into θ′ by the classical analog of wave function renormalization. But it has
no mass for the excellent reason that it never appeared in the potential in the first place, so it doesn’t reappear
there after we’ve done the shift:

We do not need to know that the symmetry was spontaneously broken to write that down. But if there were no
shift, if a were zero, we wouldn’t be able to interpret the physics of the theory easily because there would be no
kinetic term at all in the θ′ variable. It’s only because of the shift that we get an a2(∂µθ′)2. Without the shift we would
say “Whoops! We’re working in a terrible coordinate system that’s singular at the point we’re investigating. We’d
better change to some other coordinate system, ” and then we’d wind up back at ϕ 1 and ϕ 2. This coordinate
system is good for investigating small perturbations around spontaneous symmetry breakdown but bad in
general, because the shift from Cartesian to angular coordinates is singular at the origin. But we’re investigating
physics somewhere on the circle, which is far from the origin.

We can in fact arrive at this result—a massless meson when a continuous symmetry is spontaneously
broken—without using any specific model, just by waving our hands, now that we see what’s going on. The
masses of the mesons in this classical language (I keep saying masses but I really mean the things that govern
the spectrum of small oscillations) are determined by the second derivatives of the potential U at its minimum.
There’s alway going to be a massless meson if the second derivative matrix has a vanishing eigenvalue. The
appearance of massless particles is a generic property for rotationally symmetric potentials where the potential’s
minimum is not at the center. Because things are rotationally invariant there’s one direction along which the
potential is guaranteed to be constant, to wit the direction along the circle. It takes no energy to create tangential
excitations in the trough.12 All derivatives along that direction and in particular the second derivatives evaluated at
the point we have chosen to be a ground state, will be zero, so we are guaranteed to have a massless meson. It is
a consequence simply of the spontaneous breakdown of a continuous symmetry.

This phenomenon was discovered in a particular model by Nambu and Jona-Lasinio,13 but they did not
recognize that massless mesons were a universal feature of spontaneously broken symmetry. That realization is
due to Jeffrey Goldstone.14 The massless mesons that inevitably appear as a result of spontaneous symmetry
breaking are therefore called Goldstone bosons. They are the signature of the spontaneous breakdown of a
continuous symmetry.

The geometric argument we gave says there’s a direction in which the potential is flat so the second
derivative in that direction is zero. Using that as a starting point we can investigate what happens in the general
case, where there’s a general continuous internal symmetry that is spontaneously broken.

EXAMPLE 3: Fermions and Yukawa coupling with spontaneous symmetry breakdown

Write a Fermi field ψ as (19.81)

where u+ and u– are the two-component Weyl spinors that transform according to the D (0, ) and D ( ,0)
representations of the Lorentz group (§19.1). They have helicity 1/2 and −1/2, respectively;15 they are also known
as “right-handed” and “left-handed” spinors.16 Recall also (19.4) that u+ and u– transform in the same way under
rotations but with opposite signs under boosts. In the Weyl representation,

so that the projections operators

project ψ onto u±.

The ordinary Dirac Lagrangian

is invariant under a constant phase transformation Tα :

However, because γ5 anti-commutes with γµ (20.103), the Dirac Lagrangian is not invariant under the chiral
transformation,17 Tχ:

Specifically, while the kinetic term ψi ψ is invariant under Tχ,

the mass term is not:

The mass term breaks chiral invariance.Of course that’s to be expected, because the bilinear ψψ mixes helicity
states. Only the massless Dirac Lagrangian is invariant under (43.40). We have

and

so that, under (43.40)

This is an example of chiral symmetry; it acts differently on right- and left-handed fermions.

To construct an invariant Lagrangian with massless spinors coupled to two charged scalars, the scalars must
transform chirally as

The Lagrangian is

The λ term is the only possible chirally invariant quartic coupling. The Yukawa coupling can be written as

This looks like a very restrictive symmetry.

But if µ2 < 0 the scalar potential becomes as before

Again, the minima lie on a circle of radius a and exhibit U(1) symmetry; see Figure 43.6. Choose as the ground
state

and write

The Yukawa coupling written in terms of ϕ i′ is

The mesons wind up with the same masses as in Example 2. However, the fermion also acquires a mass!

The vacuum state is not chirally symmetric, but the rest of the theory is.

43.3Spontaneous symmetry breaking in field theory: The general case

Assume that we have a set of n real scalar fields {ϕ i}, for the moment the only fields in the world, assembled into a
big vector Φ. We have an N-parameter group G with elements g ∈ G, characterized by real parameters λk , k = 1, .
. . , N (and not to be confused with the Gell-Mann matrices λa, the generators of SU(3)). Under the action of the
group18 (Φ * is the adjoint of Φ)

where Tk are Hermitian matrices that generate the group, perhaps the isospin matrices or some generalization of
them. With g near the identity (with λk small)

and the kinetic term ∂µΦ *⋅∂µΦ is unchanged under (43.54), because the matrices Tk are Hermitian.

The value of N is equal to the rank of the group, i.e., the number of generators: 3 if it’s isospin, 6 if it’s SU(2) ⊗
SU(2), etc. The group generators form a closed Lie algebra: the commutator of two generators must be another
generator:19

where ckjℓ are the structure constants of the group G.

The infinitesimal change D k Φ in the field Φ is, from (5.21), obtained by differentiating with respect to the kth
parameter:
I’ll assume the Lagrangian is of the general form

The potential function U(Φ) is assumed to be invariant under the group. U has minima at some points. We pick
one of them to be our vacuum, which we will call áΦñ. In general, as the above examples show, there is no reason
to believe that áΦñ is invariant under the group; maybe it is, maybe not. Let’s consider the case where it is invariant
only under a subgroup of the group. Define the subgroup H, the unbroken group

as the set of all transformations that leave áΦñ unchanged: h ∈ H if and only if

The remaining generators are the spontaneously broken generators.

For instance if we had a theory with SO(3) invariance, the SO(3) generalization of the SO(2) theory we
developed above, then áΦñ would be a vector that has a fixed length but points in an arbitrary direction. If we
chose it to point in the 3-direction, then the group H would consist of rotations about the 3-axis.

We can arrange the generators of G so that the first ones we come across are the generators of H:

G is the largest symmetry of the theory; H is the largest subgroup that leaves áΦñ unchanged:

All the other generators of G, by definition, do not leave áΦñ unchanged. That is to say, if a sum of the
spontaneously broken generators acting on the minimum áΦñ equals zero,

then we must have

No linear combination of the broken generators can leave áΦñ invariant. If there were such a linear combination, it
would be in H and we’ve already got all of H. Thus, the generators in (43.63) generate an (N − m)-dimensional
manifold, (N − m) independent directions, starting from a point áΦñ which is somewhere in our big space of field
strength values. There is a hyperplane of dimension (N − m) tangent to that point obtained by applying the
generators (m + 1), , N, in which the potential is a constant, because U is supposed to be invariant. We can
count the Goldstone bosons on this multi-dimensional Mexican hat.20 By exactly the same arguments as before,
there are

massless scalar Goldstone bosons, for which the second derivative matrix of U has to be zero. That is, there is
one Goldstone boson for every linearly independent spontaneously broken symmetry generator. The number of
Goldstone bosons is equal to the dimension of the group minus the dimension of the unbroken subgroup.21

There are N − m Goldstone bosons because (43.63) and (43.64) define a manifold going through the point of
the minimum, which is N − m dimensional. We have a tangent direction for every α k on which the potential is a
constant, because the potential is supposed to be invariant under the group. Therefore when we take the second
derivative in those directions we will get zero; the second derivative matrix projected onto that manifold will just be
bunch of zeros, for the same reason that it was zero along the circle in Figure 43.6. The symmetries that are
unbroken don’t give us any new directions; we just apply them to áΦñ and it sits there. It’s the other ones, the
broken symmetries, that sweep out a part of the space. It’s in those directions that we’re guaranteed that the
second derivative matrix about áΦñ will be zero.

There could of course be other massless bosons in the theory. Even in an ordinary theory with no
spontaneous symmetry breaking we can pick the parameters so that the mass happens to be zero. But these
would not be Goldstone bosons. In our simple case with a single field, with U(ϕ) = λϕ 4 + µ2ϕ 2, we could have
chosen µ2 to be zero. Then áϕñ = 0 would be the minimum of the potential because of the ϕ 4, and we would have a
massless boson, but not as a consequence of the Goldstone phenomenon. The theorem doesn’t say this is the
only way to get mass zero particles. That would be obviously false, as the example I’ve just given shows. But
spontaneous symmetry breakdown in these models is the way we inevitably get zero mass particles.

EXAMPLE 4: A multiplet of scalar fields

Let Φ be an n-vector of scalar fields, in a potential which is a multi-dimensional Mexican hat

and the group G is

The dimension of G, i.e., the number of independent planes in N-space, is

As before, the ground state satisfies

I can pick a generalized North Pole,

I’ll denote by Φ ⊥ the n − 1 dimensional vector (ϕ 1, ϕ 2, . . . , ϕ n-1), and then define Φ′ by

Using (43.70) the potential becomes

The masses of the mesons are

The subgroup H is SO(n − 1) with dimension

There are n − 1 Goldstone bosons:

43.4Goldstone’s Theorem

I will now engage in a bit of hopscotch. I’m going to skip over some other models for the moment to approach the
subject from another viewpoint, using the sort of general field theoretic arguments which arise in axiomatic field
theory.22 We’ll find that consequences similar to the classical phenomena emerge. I’ll then return to the classical
fields to look at two other examples of spontaneous symmetry breaking, the sigma model,23 which is connected
with current algebra, and the famous Abelian Higgs model.24 Once we know what to expect from the general
arguments, we will bridge this enormous gap between classical fields and axiomatic quantum fields by looking at
perturbation theory to verify that the conclusions hold to all orders. The classical field will also serve as the zeroth
order of a systematic approximation scheme.

Let’s dive into the general arguments.25 I’ll assume I know practically nothing except the most general things
about quantum fields, and see what we can prove in the way of rigorous theorems. Suppose I have a local scalar
field ϕ(x) and some local, conserved current jµ(x):
If I have a theory with a conserved current, I would normally say I have a symmetry: if the integral of j 0(x) over all
space exists and is non-zero, that integral is a charge Q, which generates a symmetry.

For reasons that will become clear shortly, I want to be particularly cautious. I hesitate to integrate j 0(x) over
all space because I’m not sure that the integral will converge. So I will define a rotationally-invariant function f(x) of
compact support:

The graph of this function is shown in Figure 43.7. Define the quantity

(If I wanted to be a real purist, I could also smear the integral of j 0(x, t) out in time, but I won’t do that.) Because the
integrand goes to zero outside a bounded region, there’s no question of the integral blowing up at infinity. As R
gets bigger and bigger I integrate over a larger and larger region. Formally the charge Q is defined as the limit as
R → ∞ of QR. You will see shortly why I’m being careful here.

Figure 43.7: The function f(x)

Consider a generalization of the commutator (6.26),

I claim that the limit of the commutator with QR always exists, whether or not QR exists. The reason is
very simple. The fields are supposed to be local, and must commute for spacelike separations. As R increases
into the region where QR might grow without limit, more and more of the QR integral will be spacelike separated
from the point y, and in the limit, the commutator is finite. By the standard arguments you use to prove that the
integral of a conserved current is independent of the time, you can show that this object is independent of the
time. It doesn’t matter what time you integrate over. Just do the usual integration by parts; you don’t have to worry
about the boundary terms because they’re all spacelike separated with respect to y, and therefore they all vanish.
The limit of the commutator exists even if the limit of QR(t) did not exist; that’s guaranteed by the fact that things
commute for spacelike separations. For example, if this were the ϕ 1−ϕ 2 rotation current in the two-field model,
(43.20), then ϕ(y) would be the field ϕ 1 and Dϕ(y) would be the field ϕ 2.

The unmistakable hallmark of spontaneous symmetry breaking is that, for some vacuum state |0ñ,

(there may be several vacua in the theory). If the symmetry were manifest, the field Dϕ(y) would have vanishing
vacuum expectation value. We sandwich (43.79) between vacua on the right and the left, the charge annihilates
the vacuum (Q|0ñ = 0) and the commutator is zero. This is a rigorous definition of what we mean (in the particular
case of scalar fields) by spontaneous symmetry breaking, using only objects we are sure exist, assuming we have
a local field and a local current. That is, the characteristic sign of spontaneous symmetry breaking (43.80) can be
restated as

If the ground state were symmetric, Q|0ñ = 0, the symmetry would be manifest and the field’s vacuum expectation
value would be zero.

Of course, spontaneous symmetry breaking could occur without this particular hallmark, which has emerged
in our simple models with only scalar fields. We could imagine a more complicated theory without scalar fields, but
with some other (perhaps non-local) object in the theory with a non-vanishing VEV (which would vanish if the
symmetry were manifest), but Q|0ñ = 0. We will however look only at the case in which the charge fails to annihilate
the vacuum, as in (43.80). Certainly if this happens then there is spontaneous symmetry breaking.

I will now prove the following

Theorem 43.1 (Goldstone26). If for a given continuous symmetry

then there is a zero mass particle in the theory.

I’ll prove the contrapositive proposition,27 which is logically equivalent: if, except for vacua, every state has
PµPµ ≥ ϵ > 0, that is, the theory’s particle spectrum has no massless particles, then á0|Dϕ|0ñ = 0. (The quantity ϵ
is called a mass gap.) I assume all the usual things: the theory doesn’t have tachyons, all P2 and all energies are
positive, Lorentz invariance, and so on.

The proof is extremely simple. Consider

We can write a spectral representation for this matrix element by the usual tricks, which we have used several
times:28

The θ(k0) means we include only positive energy intermediate states in the integral. The vacua do not contribute
to the sum of states because by Lorentz invariance,

(there is no Lorentz-covariant vector in the theory).29 So there are only non-vacuum states in the complete sum of
intermediate states within the spectral density σ:

modulo factors of 2π, etc. By assumption we have

because the vacuum does not contribute, and all non-vacuum states have energy larger than ϵ. Taking the
derivative of (43.84), by current conservation we find

because the Fourier transform of zero is zero. Within σ, k2 > ϵ. So we can divide by k2 with confidence:

Then

Notice that this argument would not work if there were mass zero particles in the theory; i.e., if we could not be
sure that PµPµ > 0. Then σ(k2) could have a delta function δ(k2) in it, and we’d have k2δ(k2) = 0 even though δ(k2)
≠ 0.

By exactly the same reasoning, putting the two fields in the other order

The order has nothing to do with the proof. We find σ*(k2) = 0 instead of σ(k2) = 0, but that doesn’t matter. If two
things are zero, their difference is zero, so

Dϕ is defined from this commutator through linear operations (6.26) and (6.57):
so we have

Heuristically, we say

Given a continuous symmetry,

we can derive many equations when the symmetry is manifest. Will they still be true if the symmetry breaks
spontaneously? I summarize the results with a table.

I’ve made a number of assumptions in this proof, and I should say a little bit about them. I have assumed that
the VEVs of the local fields are what the mathematicians call tempered distributions because their Fourier
transforms are tempered distributions. For purists I should say they are Schwartz distributions with test functions
defined only over compact sets; from that and the positivity of the energy we can prove that they are tempered
distributions.30 That involves fancy mathematics. I assumed only that these fields have the property that if we
integrate them with an infinitely differentiable function that vanishes outside some finite region, then the VEV we
get is a finite quantity. Otherwise we have no business talking about the VEV’s of j µ(x), ϕ(y), or their product; they
might not exist. With that assumption, we can prove that (43.84) is a tempered distribution, its Fourier transform is
a tempered distribution, and the proof goes through.

I’ve also assumed that Lorentz invariance is not spontaneously broken. It is possible to build perfectly
reasonable models in which Lorentz invariance is spontaneously broken, but that has nothing to do with the real
world. We already assumed that when we wrote (43.84); that is a Lorentz-invariant expression. One of the hardest
tasks for people like Arthur Jaffe and Konrad Osterwalder and their friends (who want to prove things rigorously)
is to show that their ground state is Lorentz invariant (after they’ve gone through a sequence of limiting operations
to construct the ground state in the first place). If we don’t have Lorentz invariance to begin with, in general the
theorem is not true. I know of no general theorem that says that if the Lagrangian is Lorentz invariant then the
ground state has to be Lorentz invariant. It seems to be true in all the models we’ve looked at; it’s certainly true to
all orders in perturbation theory. It’s also true in all the models that have been studied rigorously, like ϕ 4 in 2 and
in 3 dimensions and the Yukawa model in 3 dimensions. Nothing is known rigorously about 4-dimensional
theories. But it is known in the real world that Lorentz invariance is not spontaneously broken.

I should also say that the theorem doesn’t show that there is a massless field here, only that there’s a
massless particle. What makes the particle from the vacuum is presumably the field ϕ, because it has to come into
the set of intermediate states in á0|jµ(x)ϕ(y)|0ñ, and ϕ by assumption is a scalar field that makes that particle from
the vacuum. By refinement of the analysis we can show it has to be a spin-0 particle, but in this case it’s obvious:
the particle is made from the vacuum by ϕ. It has to be, otherwise it wouldn’t come into the sum over intermediate
states. It’s connected with the vector nature of the current, which is spontaneously broken. If we were dealing with
something where we had a spontaneously broken current with two indices, the massless Goldstone particle would
end up being a vector.

EXAMPLE 5: A simple model displaying Goldstone’s theorem

Here is a very simple example that will help you understand the general theorem much better. It’s a rigorously
solvable field theory in which the massless Goldstone particle and the non-zero vacuum expectation value
emerge naturally. Consider the Lagrangian

which possesses neither a potential nor even a mass term. The theory possess a symmetry

with infinitesimal transformation

The associated conserved current is

The current conservation equation is obvious; it happens to be the equation of motion for a free massless scalar
field:

The j0 component of the current is simply the canonical momentum,

As expected, at equal times

as our general theorems assert; integration of (43.102) would, in the absence of spontaneous symmetry breaking,
reproduce (6.26). Following (43.79),

Therefore we have both a massless particle and a non-vanishing VEV in the theory: the Goldstone phenomenon.
Yes, it’s a silly model, but with it we can compute what QR(t) does to the vacuum, because it’s a linear function of a
free field. We can get a good idea of what’s going on with these fancy general arguments by explicit calculation,
and we can see whether or not the charge really exists.31

Acting on the vacuum, the charge QR makes only one-particle states, |kñ, describing Goldstone bosons.
Consider ák|QR|0ñ:

The ∂0 in (43.99) gives us the i|k|. If we now scale x → R x, this expression can be written as

As R gets bigger and bigger, the function (kR) (the Fourier transform of f(x/R)) gets more and more sharply
peaked around |k| = 0; it gets most of its support from smaller and smaller |k|. For huge R it behaves much like a
delta function;32 see Figure 43.8.
 Figure 43.8: The function (kR)

The norm of QR |0ñ can be written

Rescale kR → k to get rid of four powers of R:

The integral no longer has any dependence33 on R. That is,

the norm of the state QR |0ñ is proportional to R. The dependence (43.109) can also be shown easily by
dimensional analysis. We have

so [QR |0ñ] = L, and the only length available is R itself.

Therefore

In the limit of large R the norm of the state QR |0ñ blows up; the state Q|0ñ is non-normalizable. The care above
was justified, even in this simple example: the fourth component of the current integrated over all space does
diverge (quadratically, like R 2) when applied to the vacuum state, so it’s a good thing I was a purist.

Naive arguments were once made that the charge Q associated with a continuous symmetry must annihilate
the vacuum, and as the charge in spontaneous symmetry breaking fails to do that, it must be that the space
integral of j 0(x, 0) doesn’t even exist; computations with a nonexistent charge are meaningless.34 We get around
this argument because of the Goldstone boson; the massless particle gives rise (43.106) to what amounts to a
delta function at the origin, a gigantic peak at |k| = 0. As we have seen, the space integral of the density j 0(x, 0)
applied to the vacuum indeed blows up. We can nevertheless compute with the charge Q itself if we are careful.
For example, we write the commutator of the charge and a field as the well-defined space integral of the
commutator of j 0(x, 0) with the field, as in (43.103) and (43.104), where we found the commutator [Q, ϕ] painlessly.
If you don’t understand the Goldstone theorem, look at this example and see how it works.

Next time we will go on with this general analysis to deal with one more question: Is there really only one
vacuum when there is spontaneous symmetry breakdown, or are there many vacua? After all, part of the time
we’re saying there’s a unique vacuum, part of the time we’re saying there are lots of vacuum states connected by
the symmetry group. Which is the right way of thinking about things? We will demonstrate that either way is the
right way. We will then return to the classical analysis and discuss how we can look at it as the zeroth order in a
systematic quantum expansion.

1 [Eds.]
“Secret Symmetry”, (Erice, 1973), Chapter 5 in Coleman Aspects, pp. 113–121; Peskin & Schroeder QFT,
pp. 347–352; Ryder QFT, pp. 282–293; Zee QFTN, pp. 223–230; Abers & Lee GT; Jeremy Bernstein,
“Spontaneous Symmetry Breaking, Gauge Theories, the Higgs Mechanism and All That”, Rev. Mod. Phys. 46
(1974) 7–259.
2 [Eds.] The ground state of the deuteron, for instance, has an non-zero quadrupole moment due to higher
angular momentum states. J. M. B. Kellogg, I. I. Rabi, N. F. Ramsey, Jr., and J. R. Zacharias, “An Electric
Quadrupole Moment of the Deuteron”, Phys. Rev. 57 (1940) 677–695.
3 [Eds.]Neil W. Ashcroft and N. David Mermin, Solid State Physics, Harcourt College Publishers, 1976, equation
(32.20), p. 681.
4 [Eds.] E. P. O’Reilly, Quantum Theory of Solids, Taylor and Francis, 2003, p. 128.
5 [Eds.] By the way, the Ising model also displays spontaneous symmetry breaking. See Robert Brout, Phase
Transitions, W. A. Benjamin, 1965, Chapter 2, pp. 7–29. Had he not died in 2011, Brout would surely have shared
the 2013 Physics Nobel with his colleague François Englert.
6 [Eds.] Coleman included this story in a brilliant article describing the work behind the 1979 Physics Nobel Prize,
won by Steven Weinberg, Abdus Salam and Sheldon L. Glashow: Sidney Coleman, “The 1979 Nobel Prize in
Physics”, Science 20 (1979) 1290–1292.
7 [Eds.] Coleman adds: “They make dental fillings out of iron in the ferromagnet world; it’s an easily available
material.”
8[Eds.] Other examples of spontaneous symmetry breakdown are a thin metal rod bending under pressure, in
Ryder QFT, pp. 282–283; and Salam’s banquet table, in M. Kaku, Hyperspace, Oxford U. P., 1994, p. 211.
9 [Eds.] See the discussion of (2.63), p. 29.
10[Eds.] The cyclic group Z2 = {1, −1} is introduced on p. 41 of Zee GTN, and used frequently in the rest of
Chapter I.
11 [Eds.] The double-humped potential was used to illustrate spontaneous symmetry breaking in Goldstone’s
landmark paper: J. Goldstone, “Field Theories with ‘Superconductor’ Solutions”, Nuovo Cim. 19 (1961) 154–164;
see Figure 7, p. 162. Coleman remarks that Russian physicists know the graph in the context of I. M. Lifshitz’s
work on disordered semiconductors, and sometimes refer to it rudely as “Lifshitz’s buttocks”. Ilya M. Lifshitz
(1917–1982), the brother of Lev Landau’s co-author Evgeniĭ M. Lifshitz, was an outstanding condensed matter
theorist. For the Lifshitz diagram, see, e.g., I. Z. Kostadinov and B. Alexandrov, “Lifshitz correlation in the hopping
conductivity of high-temperature superconductors in the localized state”, Physica C 201 (1992) 126–130; cf. their
Figure 1, p. 127.
12 [Eds.] Zee QFTN, p. 226, contrasts the mode of oscillation in the θ direction, “rolling along the gutter”, with that
in the ρ direction, “climbing the wall”; the latter requires more energy.
13[Eds.] Y. Nambu and G. Jona-Lasinio, “Dynamical Model of Elementary Particles Based on an Analogy with
Superconductivity I”, Phys. Rev. 122 (1961) 345–358; “Dynamical Model of Elementary Particles Based on an
Analogy with Superconductivity II”, Phys. Rev. 124 (1961) 246–264.
14 [Eds.] Goldstone, op. cit.; J. Goldstone, A. Salam, and S. Weinberg, “Broken Symmetries”, Phys. Rev. 127
(1962) 965–970.
15 [Eds.] See the discussion on p. 400. Peskin & Schroeder QFT, p. 47.
16 [Eds.] See the paragraph following (19.62), p. 401.
17 [Eds.] See note 31, p. 906.
18 [Eds.] In the video of Lecture 48, Coleman uses real matrices Tk , since Φ is real. To keep a single notation
throughout the book, we use Hermitian matrices for the Tk .
19 [Eds.]
Zee GTN, Chapter VI.3, “Lie Algebras in General”, pp. 364–375. Incidentally, the structure constants ckjℓ
themselves form a representation of the Lie algebra of the group G called the adjoint representation, defined by
(Tk )jℓ = −ic kjℓ; see Zee GTN, p. 365.
20 [Eds.] Coleman quips that such a potential would be “fashionable attire” for a multi-dimensional caballero.
21 [Eds.] The demonstration of (43.65), as well as the patterns of symmetry breaking for general rotation and
unitary groups, is given in Ling-Fong Li, “Group Theory of the Spontaneously Broken Gauge Symmetries”, Phys.
Rev. D9 (1974) 1723–1738.
22 [Eds.] R. Streater and A. Wightman, PCT, Spin and Statistics and All That, W. A. Benjamin Publishers, 1964;
republished by Princeton U. P., 2000.
23 [Eds.] The sigma model is discussed in §§45.3–45.4, pp. 993–1002.
24 [Eds.] The Abelian Higgs model is discussed in §46.1.
25 [Eds.] Coleman’s treatment follows very closely the presentation in G. S. Guralnik, C. R. Hagen and T. W. B.
Kibble, “Broken Symmetries and the Goldstone Theorem”, pp. 567–708 in Advances in Particle Physics v. 2, R.
Cool and R. Marshak, eds., John Wiley, 1968. The function fR(x), corresponding to Coleman’s f(x/R), is defined in
their footnote 4, p. 706, and graphed in their Figure 1, p. 580.
26 [Eds.] Goldstone, op. cit.; Goldstone, Salam, and Weinberg op. cit.; Guralnik, Hagen and Kibble, op. cit.
27 [Eds.] The contrapositive of “if A then B” is “if (not B) then (not A)”.
28 [Eds]. As in the Källen–Lehmann representation in §15.2; while discussing the full propagators and for
the spin-§ field (23.42) and the photon (34.4), respectively; and Problem 19.3, p. 1014.
29 [Eds.] It’s here that gauge theories provided a loophole for Higgs et al. to evade the Goldstone theorem. See
note 6, p. 187.
30 [Eds.] W. Appel, Mathematics for Physics and Physicists, Princeton U. P., 2007, p. 300.
31 [Eds.] On general grounds, it can be shown that the state Q|0ñ is not normalizable if Q|0ñ ≠ 0. Following Guralnik
et al., op. cit., p. 573 and Bernstein op. cit., p. 11, the argument goes like this. A vacuum state |0ñ is translation
invariant, and so is Q|0ñ. But then

If Q|0ñ ≠ 0, this diverges. Consequently, for spontaneous symmetry breaking, Q|0ñ must have a divergent norm. In
what follows, Coleman demonstrates Q|0ñ’s infinite norm in this simple model.
32 [Eds.] It’s easy to see that the Fourier transform . Approximating the function by

it follows that f(x/R) → 1 as R → ∞; and the (three-space) Fourier transform of 1 is (2π)3δ(3)(k). More directly, an
elementary integration (with µ ≡ cos θ) gives

Figure 43.8 shows the graph of for R = 2. The height of the peak is given by the limit of the function as |k| →
0:

as expected; we know δ(3)(k) ~ 1/|k|3, and R ~ 1/|k|. (The figure’s peak is 32π/3 units tall.) Like a delta function, its
area (divided by (2π)3) equals 1:

independent of R. (The cosine integral doesn’t really converge, but the Riemann–Lebesgue lemma suggests we
can ignore it.) Alternatively,

33 [Eds.] If one makes the replacement kR → k in the explicit form (*) (note 32, p. 956) of the approximate (kR),
one finds that the R dependence of (k) does not go away; instead, it depends on R 3. Take instead (k), after
the change of variables in the integral, as the special case of (kR) with R = 1.
34 [Eds.] See Bernstein, op. cit., Section II., pp. 10–11: “[T]he state Q|0ñ is not normalizable. This is difficult to live
with but not impossible, since in all applications we will consider commutators involving J0(x, t) and then integrate
safely later… Clearly this is a subject in which common sense will have to guide the passage between the Scylla
of mathematical Talmudism and the Charybdis of mathematical nonsense.” (Scylla and Charybdis were two sea
monsters, the first multiheaded (deadly to some of the crew) and the other generating a whirlpool (fatal for the ship
and all aboard). Odysseus and his companions had to navigate between them. For the Talmud, see Rosten Joys,
pp. 565–576.)

Problems 23

23.1 A scalar meson ψ of mass m and charge e is minimally coupled to electromagnetism. In addition there is a
massless neutral pseudoscalar meson, ϕ, with a nonminimal electromagnetic coupling:
Here g is a positive number, Fµν = ∂µAν − ∂νAµ, and ϵ0123 = +1. Compute, to lowest nonvanishing order in
perturbation theory (this is (e2g2)) the differential cross-section dσ/dΩ, averaged over initial photon polarizations,
for the process

Work in the center of momentum frame, and express your result in terms of the center-of-mass total energy and
the center-of-mass scattering angle.

Comments: (1) This theory isn’t renormalizable, but that doesn’t matter here, since you’re only working in tree
approximation. (2) Massless pseudoscalar mesons with this peculiar coupling appear in some of the extensions of
the standard model. In these models, an experimental upper bound on g comes from studying the conversion of
photons to ϕ’s deep inside a star. After they are produced, the (weakly-coupled) ϕ’s escape the star, stealing
away energy with potentially drastic effects on stellar dynamics. This problem is a simplified version of this
calculation, with the spin-½ charged particles inside the star (electrons and protons) replaced by scalar mesons.
(3) I didn’t ask you to compute the total cross-section because the integral that defines σ diverges at θ = 0. (In
case you’re curious, inside the star the divergence is cut off by the shielding of the Coulomb field by the
electron–proton plasma; this has an effect roughly similar to giving the exchanged photon a small rest mass.)
(1998 253b Final, Problem 1)

23.2 Example 1 of Chapter 43,

was a theory with two ground states, ϕ = ±a, connected by a discrete symmetry. Such theories are in bad repute,
for reasons linked to cosmology. Very early on, when the temperature of the universe is very high, the discrete
symmetry in such a theory is unbroken, just as in the ferromagnet discussed (briefly) in class. As the temperature
falls, the symmetry suffers spontaneous breakdown, and ϕ goes to one of its two allowed values. However, there
can be no correlations between regions of space that are causally disconnected, that is to say, that are so far
apart that a light signal could not have gone from one region to another in the time since the Big Bang.
(Cosmological sophisticates may substitute “the end of the inflationary epoch” for “the Big Bang” in the preceding
sentence.) Therefore, if ϕ is a in one of these regions, it is equally likely to be a or −a in another. We thus have a
picture of alternating regions of positive and negative ϕ, separated by transition zones, “domain walls”. As you
shall see when I finally get around to stating the problem, these domain walls typically have microphysical
thicknesses and energy densities; if they’re around, stretching across the universe, they mess up all sorts of
things in cosmology. (Since by pointing our telescopes in different directions, we can see causally disconnected
regions even now, there should be at least one domain wall currently stretching across the visible universe,
causing problems not just for cosmology but for observational astronomy.)

In this problem you are asked to work out the explicit form of the simplest domain wall, one that is time-
independent and flat, in Example 1. Find a solution of the field equations, ϕ(z), depending on the z coordinate,
such that ϕ(±∞) = ±a. (You may have to resort to an integral table.) Find the energy per unit area of this domain
wall, as a function of a and λ. (Note: This is a problem in classical physics.)

H INTS: The differential equation you’ll encounter will closely resemble the Newtonian equation of motion for a
point particle, with ϕ replacing the particle position and z the time. You should be able to go a long way towards
solving the equation by writing down the analog of the conservation of energy.

Something to think about but not to hand in: Why isn’t Example 2 in similar bad repute?
(1998b 10.2)

Solutions 23
23.1 The diagram for ψ + γ → ψ + ϕ is (arrows denote the scalar ψ fields; the plain line denotes the pseudoscalar
ϕ):

The Feynman rules for QED are given in the box on p. 670, but we have to work out the vertex corresponding to
the term in the Lagrangian coupling the photon to the pseudoscalar:

Take both photons as incoming, and Fourier transform the term. Every derivative becomes −ipµ for an incoming
momentum, and +ipµ for an outgoing momentum (see comment (3), Problem 8, p. 309). This term becomes

By functional differentiation δ3/δ δ α δ β, and including the usual factor +i from Dyson’s formula, this expression
leads to a vertex

using momentum conservation and the antisymmetry of the ϵ. The squared amplitude for ψ + γ → ψ + ϕ is given
by

using k2 = q2 = 0. It’s convenient to define

Then p + p′ = 2pt − k − q, and

The advantage of writing things in terms of pt is that these vectors have inner products that are simply expressed
in terms of center of momentum variables:

To obtain the differential cross-section, we need to sum over the final spins. Writing

we have from (30.44)

In this sum we have to calculate the square of the four-dimensional Levi–Civita tensor. In analogy with (37.47) we
have

(recalling ϵµαβγ = −ϵµαβγ). This sum would give us six terms, but three vanish because k2 = q2 = 0, and two of the
others are identical. Then averaging over the initial spins and summing over the final spins,

Finally, from (12.26),

As claimed, the integral that defines σT diverges at θ = 0.

■
23.2 We start with the general form of the Lagrangian

where = dϕ/dz. Using the suggestion to consider the “conservation of energy”,

Solving for gives

Solve for dz/dϕ:

Integrating with respect to ϕ,

The problem states that → as → ∞. Thus the integral in (S23.17) must diverge as → . We conclude
that C = 0, and

(Gradshteyn & Ryzhik TISP, integrals 2.143.2 and 2.143.3.) Inverting,

The energy density is

From (S23.14) we have − 2 + V(ϕ) = C = 0, so

The energy per unit area is

(Gradshteyn & Ryzhik TISP, integral 2.423.12.)

■

44
Perturbative spontaneous symmetry breaking

In this chapter we are going to look at two questions concerning spontaneous symmetry breaking. First, as we
have seen, the vacuum is not unique in theories with spontaneous symmetry breaking. Well, is this a problem or
not? Second, how does perturbation theory affect spontaneous symmetry breaking? Thus far I have considered
spontaneous symmetry breaking only in the context of classical fields. But (by an abuse of language) I have
described the results with quantum terminology and notation, e.g., writing the value of the classical field ϕ that
minimizes the potential U(ϕ) as a vacuum expectation value, áϕñ. I’ve done this to smooth the connection between
the classical results and what happens in quantum field theory. As I’ll show, the classical results are the lowest
order in a perturbation theory expansion. Do the conclusions we found in the lowest order of perturbation theory
survive in higher order? Might there be corrections to a Mexican hat potential? Or maybe there’s some other
potential, corrections to which cause the symmetry to break spontaneously? I’ll address these two questions in
turn.

44.1One vacuum or many?

We have seen several examples of field theories where a symmetry is spontaneously broken and a single
vacuum state develops into several equivalent, equally valid degenerate vacuum states. The Ising model1 is
another such theory. It is a model of ferromagnetism similar to the Heisenberg model I talked about last time.

In the Heisenberg model the Hamiltonian (43.1) is rotationally invariant. The net magnetization in the ground
state can point in any direction, breaking the symmetry from SO(3) of the Hamiltonian down to SO(2) of the ground
state:

In the simplest form of the Ising model the spins can only point up or down along one direction, say the z-direction.
The interaction Hamiltonian

(B is the magnetic field, vij is the exchange interaction) is not rotationally invariant. There are only two possible
vacua: all the spins pointing up or all the spins pointing down:

The cousin of the man in the Heisenberg ferromagnet lives in an Ising ferromagnet. He cannot change all of the
spins in an infinite system by any finite set of local operations. These vacua are not invariant under the symmetry
operation

but they are orthogonal:

What about linear combinations of these two vacua? They are also vacua:

These linear combinations are symmetric under the interchange (44.4) (modulo the overall sign in |oddñ), but they
are not orthogonal:

We will focus on the orthogonal vacua and the different Hilbert spaces built upon them, and we will elevate the
above discussion to the status of a theorem.

Assume that there is a finite number N of vacuum states |0, αñ with zero momenta:

These states are normalized so that

where α, β = 1, 2, , N. There are no other normalizable momentum eigenstates. That distinguishes them from
other P = 0 states, such as two-particle states in the center-of-momentum frame; the other states are in the
continuum and so are not normalizable. From (44.8), for a translation by some displacement a,
 The the algebra of quasilocal Hermitian operators

Figure 44.1: Bounded region R

with f = 0 if any of the xi is outside some bounded region R. There is a basis where all the vacua are independent
and cannot be connected by only quasilocal operators. These are called good vacua. That is, all the A ∈ are
diagonal. And the converse is true: If the A’s are diagonal, then there will always be good vacua; these are the
states that diagonalize the A’s. Formally we have the following theorem:
Theorem 44.1. There exists a basis for the vacuum states |0, αñ such that if A ∈ (R) then

The theorem says that it doesn’t matter how many vacua there are. We can consider all of them or only one
particular vacuum, the one we happen to live in, and just worry about that one. We never have to worry about all
the other vacua because nothing we can ever do with local operators can ever get us to any of the other vacua. In
our Ising example, we can’t change all the spins in all of space. The good vacua are globally distinct.

After this big song and dance, the proof of the theorem turns out to be fairly simple. It depends only on
translation invariance and causality, and follows easily from two lemmas:

Lemma 1: Let A, B be any two elements in (R). Then

The reason is that eiP • af(x)e−iP • a = f(x + a), since e−iP • a is a spatial translation operator. A is associated with
some region R A ⊂ R, B is associated with some region R B ⊂ R. When we apply the spatial translation operator we
translate B by some finite value of a, eventually to a position where it’s separated from A by a spacelike interval.
At that separation, they can no longer influence each other: the commutator must be zero.

Figure 44.2: Two bounded regions, spacelike separated

Lemma 2:

where P0 is the projection operator onto the vacua,

Proof: Evaluate the right-hand side of (44.14) by inserting a complete set of intermediate momentum eigenstates:

using (44.10). The only normalizable momentum eigenstates are the vacua, with zero momentum eigenvalues. All
the other states have continuous momentum eigenvalues. In the limit a → ∞, the continuum contributes nothing;
the phases cancel out by the Riemann–Lebesgue lemma. (That’s the same argument we used when discussing
the reduction formula.2) The only contribution comes from the vacuum states:
That proves the second lemma.

Obviously the order doesn’t matter: with B in front of A it’s the same argument. Putting these two things
together we get, using the first lemma,

one term with A and B in one order and another with the order reversed. But

The summation is nothing but the product of two matrices (summation over γ). Consequently the right-hand side of
(44.18) says

That is, for any A and any B

are commuting Hermitian matrices (A and B are supposed to be observables, and hence Hermitian).

Thus, with every observable A within we associate a matrix consisting of its matrix elements between the
different vacuum states. These matrices all commute with each other. If we have a family of commuting Hermitian
matrices, they can all be simultaneously diagonalized by one and the same unitary transformation.3 In this basis
none of them have off-diagonal matrix elements. QED

I’ve proved the theorem for a finite number of vacua so that these are finite-dimensional matrices, but it’s also
true if the number of vacua is infinite. If α is a continuous index, then the generalization of (44.9) is

and the theorem generalizes to

To speak a little in sophisticated mathematical talk, it could be that the big Hilbert space is not a direct sum of little
Hilbert spaces but a direct integral. But that hardly matters. It’s similar to what we do in going to the center of
momentum frame: the big Hilbert space spanned by eigenfunctions of all total momenta is a direct integral, not a
direct sum, of spaces with fixed momenta, but who cares?

It’s a cunning theorem. I don’t know who first proved it, nor where to find it in the literature. Arthur Wightman4
showed it to me in 1973. The significance of the theorem is this: it doesn’t matter if you say there’s one vacuum or
many; there are always good vacua. It shows that, even if we don’t know anything about spontaneous symmetry
breaking, and we’ve chosen a bad set of vacua, by a systematic constructive procedure we can always find a
good choice of bases for the vacuum subspace, such that no local operator can connect one vacuum to another.

I don’t know whether this theorem was motivated by spontaneous symmetry breaking or not; it may predate
the Goldstone–Nambu ideas. Its origin may lie in statistical mechanics, where similar things occur. In my
experience, when this sort of argument appears in statistical mechanics, it’s usually the product of a similar
argument in field theory. But there has also been a flow in the other direction, from statistical mechanics into field
theory, involving people like David Ruelle.5 I would guess this argument originated in axiomatic field theory.
Maybe somebody asked, “What happens if we assume there are a lot of vacua?” And that person, or someone
else, worked hard and showed that it didn’t make any difference, without necessarily thinking about the
application to spontaneous symmetry breaking.

44.2Perturbative spontaneous symmetry breaking in the general case

I want to return to making a bridge between simple classical arguments and rigorous quantum arguments. (The
bridge doesn’t go all the way; the constructive field theorists are trying to finish the job.) Now the classical analysis
of Chapter 43 will be redeemed as the leading term in a systematic perturbation theory expansion. Does the
spontaneous breakdown survive to all orders in perturbation theory, for appropriate choices of the parameters
(e.g. a negative mass squared term in our simple ϕ 4 theory)?6 Unfortunately, as always when we’re making
general perturbation theory arguments, we have to use the fearsome generating functionals7 (which I love but
many students hate). Bear with them; you’ll see how helpful they are. Later I’ll do some specific calculations to put
tangible flesh on bare and abstract bones. But first you’ll have to suffer through some unavoidable (and
unrelieved) formalism.

Let’s recall a few facts about generating functionals. (This will be just an aide-mémoire; a recapitulation of
earlier statements, without proofs.) To keep the notation simple I assume that I have a Lagrangian describing a
single scalar field, ϕ, with some mass term and self-interaction, and, if these are really renormalized fields, also a
counterterm:

(The argument is trivially generalizable.) If there is a mass term, it appears in U(ϕ). Define the action in the
presence of an arbitrary external c-number function of space and time, J(x):

and define Z[J] and W[J], the generating functionals for full and connected Green’s functions, respectively, by

(cf.(13.14), (28.27), and (32.4)), where N and N′ are normalization factors. Define as the vacuum expectation
value of ϕ in the presence of J:

if J is time independent.8 The state |0ñin is the vacuum in the far past, and |0ñout is the vacuum in the far future; the
vacuum is not the free field vacuum when there’s a J around, so we have to do it this way. I now make a Legendre
transformation (32.27) and define Γ[ ] by

with the equation

Recall (§32.2) that iΓ[ ] is the generating functional for one-particle irreducible (1PI) graphs. We exploited this fact
repeatedly in our investigations (§33.4) of the Ward identities in quantum electrodynamics.

There is a systematic way of expanding Γ[ ] that corresponds to an expansion in powers of ħ if we stick

an ħ back into . This is the semi-classical or loop expansion: expanding in no-loop graphs, one-loop graphs, etc.,
a natural kind of perturbation theory (32.21) for Γ[ ]. In the tree (no-loop) approximation Γ[ ] is just the
classical action, [ ]:

(Remember that the tree approximation is what we get when we sum up the tree graphs, with no-loop
corrections.9 If it’s a ϕ 4 theory, the 1PI graph with four external lines gives the ϕ 4 term and the inverse propagator
gives the (∂µϕ)2 − µ2ϕ 2 term, etc.)

We are now in a position to use this formalism. You thought it was set up to facilitate the study of the Ward
identities and renormalization theory in quantum electrodynamics. That’s true, but it was also designed to be used
in spontaneous symmetry breaking.

To take a definite example, let’s consider our ϕ 4 model. The condition that the symmetry breaks
spontaneously is that (44.25) is non-zero even though J is zero. This equation,
if it has solutions, will tell us whether or not spontaneous symmetry breaking occurs. If the theory wants to have
= 0, this equation will tell us that the solution will be = 0; likewise, if the theory wants = ±a. In that
case (44.29) is the statement that a nonzero expectation value of ϕ is tolerable with J = 0. As before I’ll denote
by áϕñ, now with much more justification than in the classical theory. I can shift the field, just as in the
classical analysis, and define a new quantum field ϕ′:

I can re-express Γ as

Now ′ = 0 (in the ground state of the theory) because of the way we’ve constructed it. I expand Γ[ ′ + áϕñ]
about ′ = 0 to obtain the 1PI Green’s functions when the symmetry is spontaneously broken.

Everything I did in the classical theory of spontaneous symmetry breaking earlier goes through without
alteration in the quantum theory (i.e., using perturbation theory), but with the effective action Γ[ ] substituted
for the classical action [ϕ]. Instead of trying to find minima by finding the stationary points of the classical action, I
find ground states by looking at the stationary points of the effective action; instead of finding effective coupling
constants and masses by expanding about the minima of the classical action, I find 1PI Green’s functions by
expanding about the minima of the effective action. It’s exactly the same game in the quantum and classical
theories (see Table 44.1; the penultimate pair of equations in the table will be explained presently).

Table 44.1 Classical and Quantum (Perturbative) SSB

Note that I have set

because I’m not interested in the spontaneous breakdown of translational symmetry. There are certain kinematic
simplifications coming from the fact that in theories we’re interested in, the ground state is spatially homogeneous:
is translationally invariant, a constant. I’ll come back to this later.

There’s no reason why translation invariance should not be spontaneously broken in a theory that describes
the real world. It occurs in statistical mechanics, for example, where the phenomenon is called crystallization.
There, instead of changing the square of the mass to cause the manifest symmetry to break spontaneously, one
changes the temperature. Let’s take a typical material such as iron, and imagine an iron universe, spatially infinite.
If the temperature is above a certain point, the ground state (in the sense of statistical mechanics) is spatially
homogeneous; it’s iron vapor. We lower the temperature below the freezing point of iron, and the ground state
becomes an infinite iron crystal, which does not have spatial homogeneity. If we now consider the rotation of a
crystal somewhere in the frozen iron, how it rotates depends on its position relative to a central lattice point. That’s
an example of spontaneous symmetry breakdown of translational invariance.
 I want to make three points that are absolutely critical. First, the shift ′ = − áϕñ commutes with the loop
expansion. That’s because the loop expansion is an expansion in powers of a parameter that multiplies the total
action. It is therefore completely indifferent as to how we break up the action into a free and an interacting part.
One way may be natural before we make the shift, and another way may be natural after we make the shift, but
that’s irrelevant. One-loop diagrams are not shifted into two-loop diagrams. Second, the analysis of the quantum
theory in the tree approximation recreates that of the classical theory. That is because in the tree approximation Γ
is the classical action. What we did with classical fields was not simply pedagogically useful, but in fact is the
zeroth stage of a systematic quantum expansion. All the words get changed, but the equations are exactly the
same. And we know how to compute the quantum corrections to this zeroth order term. We just compute the one-
loop corrections to Γ and then go through the same algebra as before. Finally, and most critically, spontaneous
symmetry breaking does not affect renormalization. The renormalization counterterms in a theory with
spontaneous symmetry breaking are of exactly the same type as if there were no spontaneous symmetry
breakdown. For example, if we consider a single scalar field with quartic self-interactions, the only counterterms
we need to compute in Γ are a ϕ 2, a ϕ 4 and a (∂µϕ)2 counterterm. We could make all of our renormalizations in the
computation of Γ before we do any shift (though sometimes it’s not the most expedient approach). Then we
certainly won’t need a ϕ 3 counterterm because, before we do the shift, there are no ϕ-ϕ-ϕ 1PI diagrams in ϕ 4
theory to be canceled out. After we do the shift, of course, a ϕ ′3 interaction will appear in the effective action, but
we still don’t need a ϕ ′3 counterterm, because we’ve already gotten rid of all infinities in computing Γ before we’ve
made the shift. The shift is a purely algebraic operation without a single integration over internal momenta, and
therefore cannot possibly introduce new ultraviolet infinities.

The value of is a function of the masses and the coupling constants (or whatever renormalized parameters
we choose) in the original process. If we choose to make our renormalizations before we’ve made our shift, we
probably won’t choose to renormalize on the mass shell, because that’s the wrong mass; the mass squared is a
negative number. After we make the shift and get to the physical theory, the one we really see, we might choose to
make a further finite renormalization to turn things into physical parameters for the shifted theory.

The renormalization of ϕ itself is basically the wave function renormalization. In ϕ 4 theory, you fix three finite
parameters (in any manner you choose: BPHZ, or maybe some fancy renormalization convention of your own, or .
. .) and you’ve fixed the theory. Those three parameters—conditions on the renormalization of the field (wave
function), on the two-point function (mass), and on the four-point function (coupling constant)—are enough to
absorb the infinities. The vacuum expectation value of the field depends on the choice of the three renormalized
parameters. If what you call the mass and the coupling constant are not what I call the mass and the coupling
constant, we’ll get different analytic expressions, but we’ll be describing the same physics. Any way we
renormalize is as good as any other. It’s just a matter of convention, something like the medieval disputes over the
length of a standard foot: was it to be based on the foot of England’s king or Belgium’s? It doesn’t matter, so long
as we stick with our conventions.

The program we have set out is beautiful in its conceptual simplicity. But it’s rather complicated to carry out,
because we’ve got to compute the effective action. That’s a messy thing to compute to all powers of ϕ, even in
one loop, because of the arbitrary external momenta on all the lines.10 Considerable simplification is made if we
use the fact that, in most of the cases we are interested in, (x) is independent of x, just a constant:

Investigating the effective action for constant ’s is much simpler than for general ’s. A constant field in
position space has a zero derivative, and so in momentum space we need consider only graphs with zero external
momenta, because that corresponds to a constant field in position space. In the classical case the Lagrangian
for a constant field has a vanishing kinetic term and a constant potential:

U(áϕñ) is the energy density of the ground state. Usually it’s set to 0, but if there are two local minima (as in Figure
44.4), U(áϕñ) will tell you which is which. The factor d4x, formally infinite and equal to the volume L3T of the
space-time box, takes care of translational invariance.11

In the same way from the effective action Γ[ ] we define a quantity V( ) called the effective potential:
In tree approximation

and

CT includes the loop corrections. Then

We defined V( ) so that it corresponds to the ordinary potential in the tree approximation and then has corrections.
V( ) is called the effective potential for the same reason that Γ[ ] is called the effective action. It is a
generating function, not a generating functional; it doesn’t depend on a variable field (x) but on a single
number, . Since iΓ is the generating functional of 1PI graphs, −iV( ) is the generating function of 1PI
graphs with all the external momenta equal to zero and with the (2π)4δ(4)(0) from overall energy-momentum
conservation divided out. That’s just the Fourier space equivalent12 of the integral d4x.

The rule for computing V( ) is very simple. You don’t have to worry about any external momentum. You just
have external lines each carrying zero momentum. Sum up all those graphs to one loop or two loops or however
many loops you’re going to do. I will do that summation in front of your very eyes for a general U(ϕ). We will get
the effective potential V( ). The condition that determines whether the symmetry is spontaneously broken or
not is then

—an ordinary derivative, not a functional derivative, because it’s just a function of a number; we treat Γ as if it
were , the action, and V( ) as if it were U(áϕñ), the potential.

If is not a constant and you imagine expanding in a Fourier series, any terms with non-zero
momenta have to enter at least quadratically for the momenta to cancel out: Γ is translationally invariant.
Therefore if we’re interested in derivatives near a constant field, we only need to know the value of the function for
the constant field. The variational derivatives with respect to the non-zero Fourier components of will
automatically be zero if evaluated at a constant . Further expanding Γ we get something like13

where W( ) would take all graphs and evaluate them to second order in the external momenta, picking up terms of
order k2, either second order in one momentum or first order in one and first order in another, as well as terms of
order 0. Think of V( ) not as the first term in an expansion but as a general functional evaluated for a constant
field. It’s defined for arbitrary fields, so in particular it’s defined for a constant field. Never mind whether there’s an
expansion about that point or not.

44.3Calculating the effective potential

Let’s now work out V( ) in a particular case. Actually, if we’re only interested in qualitative information there’s
hardly any point in computing it, because the loop expansion is the expansion in powers of a coupling constant if
there’s only one coupling constant in the theory—loop graphs have more powers of the coupling constant than
tree graphs. Therefore if we’re interested only in the qualitative behavior of the theory, and if the coupling constant
is small (the only case in which we have any right to use a diagrammatic expansion), there’s hardly any point to
the calculation. The moral has already come through: nothing qualitative will change. There will just be a small
correction to the picture we’ve already developed. Nor will there be any problem with renormalization. So we’re not
going to learn anything qualitatively new in this sample calculation. (There are special cases where the tree
approximation does not give an unambiguous answer. In such cases we do learn something new, and I’ll talk
about those later.) But we’ll get some feeling for the structure of the argument by doing this calculation.
 We begin with (44.22):

where U(ϕ) is or dimension 4 or less; otherwise the theory is not renormalizable. CT already contains whatever
counterterms we need. (Here, these will be quadratic and quartic.) To one-loop order the full expression for −iV(
) is

There will be an infinite sum of graphs, for which we adopt a special notation. The heavy black dot means the
following: I’m going to expand in powers of U (or equivalently, in powers of the coupling constant); I’m not even
going to put a mass term into the propagator. In the loop expansion it doesn’t matter how we split things up, so the
propagators are all going to to be14

The heavy dot is going to consist of everything that can go on a vertex:

diagram (a) for the mass term; diagram (b) for a ϕ 3 term if present, with one external line and a number, ,
multiplying it (this is supposed to be the generating function that, when differentiated with respect to , gives
us 1PI graphs); and diagram (c), corresponding to a ϕ 4 term, with two external lines carrying zero momentum and
multiplied by . For example, given

then the diagrams (a), (b) and (c) are equal to −2ia2, −3 ⋅ 2ia3 , and −4 ⋅ 3ia4 2, respectively. There could be
other terms if I were foolish enough to consider a non-renormalizable theory. These are all the possible
interactions that can go on the dot, either with no external lines, one external line, two external lines, or what have
you. All the external lines (multiplied by ) carry momentum zero; the other two lines are going around the
loop. We don’t need to do any fancy summation for ; we just look at the divergent graphs. There will be
counterterms at the one-loop level, but not at the tree level; terms linear in do not contribute to one-loop 1PI
diagrams.15

It’s very easy to get a rule for the heavy dot. Let’s compute for example the value of the vertex for the
potential

Only one term will appear in (44.44), the term with (n − 2) external lines coming off an internal line. The internal
line is made of contracting two fields; there are n possible choices about the first field, and (n − 1) choices for the
second. All the other fields are supposed to carry zero momentum. They are indistinguishable from each other so
we don’t have to worry about which is which. Each dot, however, carries a factor of λ to the power (n − 2),
because there are (n − 2) of the fields left. Finally there’s a −i because the Feynman rules are derived from exp(i
) and has −U in it. Thus if the potential has the form (44.45), the vertex is

If this is the value of the heavy dot for ϕ n then the value for a general U is the second derivative with respect to the
argument:
This reproduces (44.46). That’s the heavy dot vertex with all of those lines summed up, no matter how much “hair”
is sticking out of the heavy dot.

It is trivial to sum up the loops: it’s an infinite power series. To one loop order

where n is the number of heavy dots. Here’s where the factors come from: Each of the n propagators carries the
same momentum k because all of the external lines carry zero momentum. Each vertex contributes the same
amount, −iU″( ). The combinatoric factor, 1/(2n), arises because where we start, and the order in which we
go around, are unimportant—if we take an n-legged polygon, we get exactly the same graph if we rotate it by
(2π/n), and also if we reflect it; neither operation leads to a new term in the Wick expansion. So the factor (1/n!)
from Dyson’s formula is not completely canceled.16 (The ultraviolet divergences will be soaked up in the −i (
) term.)

Please note that we are not normal ordering our interactions. In general when discussing questions of
symmetry, and in particular complicated invariances like gauge invariance in quantum electrodynamics, it’s a very
bad idea to normal order things, despite what it says in elementary books. That leads to confusion, because
normal ordering does not commute with gauge transformations or with shifts. There are some places where it
won’t hurt you; you just generate new counterterms which you soak up in the old counterterms. But there are
many situations where the un-normal ordered expression is symmetric and the normal ordered form is not
symmetric. In those cases you certainly don’t want to normal order carelessly. If you are worried about the infrared
problem here because of a lack of m2 in the propagator, stop worrying. It will disappear when we sum the series. I
know it will disappear, and you do, too: I told you that it doesn’t matter how I split things up, so I could always add
an m2 to k2 in the propagators, and subtract it from U″( ).

Let’s evaluate (44.48). The in and the (−i)n cancel, and we multiply both sides by i. The sum in (44.48) is just
the logarithmic series for −ln(1 − x) with a ½ in front. We rotate to Euclidean space, which gives us another i on the
right-hand side:

You’ll notice that I’m keeping the iϵ even in Euclidean space. That’s just for safety’s sake. We’ll see later on that
it’s a good thing to do. If I write the integrand as

the second term integrates to a constant (i.e., independent of U″( )). It’s quadratically divergent, but to hell with
that; it can be absorbed into the renormalization. What remains is an elementary integral (you can find it in the
standard tables).17 I’ll put in a brutal cut-off and integrate from kE = 0 to kE = Λ. The factor U″( ) corresponds
to all those lines carrying zero momenta. Though a function of , it’s a constant, because is a constant
field. Making the substitution (44.50), the integral becomes18

(the term (Λ2) is a constant).

We get the expected divergent terms, with the counterterm ( ) evaluated to one loop order, but if U is of
quartic order or less, these are already accounted for: the divergent terms are of the same form as terms in the
original Lagrangian—the U″ term is at most a quadratic function, which tells us we need a quadratically divergent
counterterm, proportional to 2; and (U″)2 is a quartic function which tells us we need a logarithmically
divergent counterterm proportional to 4. We already have precisely those counterterms in our original CT

and therefore we can absorb all the Λ-dependent terms into . If I had been so foolish as to investigate a non-
renormalizable theory, say one with a ϕ 5 term in U, then (U″)2 ln Λ2 would give a term proportional to ϕ 6, and I
would be stuck: I have no counterterm to absorb it.19 Non-renormalizable theories are sick no matter how you look
at them; they’re no healthier from this vantage point.

We can absorb the divergent constants into the counterterms, leaving perhaps a residual finite part of the
counterterm (depending on what the renormalization conditions are). Thus we are left with

where is the finite part of the counterterm (to first order). I can’t specify it without knowing what the
renormalization conditions are.

When you look at this formula (44.52) you’ll say, “That’s lnU″( ) over what?” After all, U″( ) is something
with the dimensions of a mass squared. One term in it is the mass squared, for example. Well, it doesn’t matter
what we choose as a denominator for U″( ). If we change the denominator in the argument of the logarithm,
we merely pick up a finite term proportional to [U″( )]2 and that’s absorbed in the finite counterterms .
You tell me the renormalization conditions and I’ll tell you the denominator in ln[U″( )/(what)]. Putting in an
unspecified M2 for the denominator, we can write this as

The concludes our sample computation. I hope it has put some flesh on the idea of the effective potential. I
wanted to show you how to compute it, and to emphasize the point that the only renormalization constants needed
are just those that would be present if the symmetry were not spontaneously broken (the third of the three points
made earlier).

This formula can be immediately generalized to the case of many scalar fields, ϕ a, a = 1, 2, n. In this case,
the heavy black dot in (44.44) is labeled by the two indices on the two ϕ fields going in and out of the dot. It can
also be extended to the case of spinor and vector fields, but we will postpone that for a future lecture.

Suppose there were many fields ϕ a, as in the case of our model with Goldstone bosons. Then we define a
matrix

Figure 44.3: Multi-scalar loop

I would have to consider the same loop diagrams as in (44.42), except now each internal line could be of a
different kind. For example, we would get for the loop shown in Figure 44.3 the amplitude for a1 going into a2 in the
presence of the external field, followed by the amplitude for a2 going into a3, etc., summed over all the fields,
summed over repeated indices:

I did it for four lines, just as an example. I could have done it for n lines; I’d get exactly the same result. With the U″
matrix defined by (44.54), the formula (44.53) generalizes to
As before, we can’t specify ( ) until we know the renormalization conventions. The last term here is the trace
of the product of the indicated matrices. As U″ab( ) is a symmetric matrix of real quantities, it is Hermitian,
and so is ln U″: the logarithm of a matrix is a matrix. There’s no problem defining it. Every loop integral is exactly
the same as before.

This formula was first derived by Coleman and Weinberg: this Coleman and the other Weinberg, Erick
Weinberg.20 Steve Weinberg refers to this work as “that paper with pseudo-Goldstone bosons and a pseudo-
Weinberg”. We’ll learn what a pseudo-Goldstone boson is in the next lecture.21 The generalization to fermions will
turn out to be almost exactly the same, as will the inclusion of gauge fields—even non-Abelian gauge fields, which
I have not yet talked about.22

44.4The physical meaning of the effective potential

I’ve introduced the effective potential V( ) as, in some sense, the quantum generalization of the classical field
theory potential U(ϕ). It is U(ϕ) to lowest order, and then it gets quantum corrections. The potential U(ϕ) has the
mathematical meaning that its stationary points determine the ground states of the theory.23 But it has a physical
meaning as well: its value at the stationary points {áϕñ} is the energy per unit volume, the energy density, of the
ground state (or states) for which the field takes the value áϕñ. I want to show that V( ) has exactly the same
meaning:24 that V( ) is the energy density (43.5), 0 = E/L3, for a state of lowest energy with áϕñ restricted to
be . We normally consider the true ground state of the theory as the state of lowest energy without any
restriction. Suppose we put a restriction on it, that the expectation value of ϕ is to be fixed at some number .
I will demonstrate that the answer to the question “What is the lowest energy the system can have with the
restriction that áϕñ must equal ?” is V( ).

The question becomes important in the case when V( ) has two local minima, only one of which is an
absolute minimum. One can imagine that happening even in tree approximation. If I wrote down a theory with both
a ϕ 4 and a ϕ 3 coupling, then instead of those nice symmetric Russian buttocks,25 I would find one cheek higher
than the other. As before (when there was no ϕ 3 term) there would be two local minima, but now only one would
be an absolute minimum. From the viewpoint of perturbation theory it looks like I could expand about either
minimum equally well. Are they both vacua? That V( ) is the energy density of the ground state says “No”:
the higher one is a false vacuum; it has a higher energy than the lower state.26 If we attempted to put the system
into the higher state, we would expect it to eventually decay to the lower state. It’s not something we’d see in any
finite order in perturbation theory because it’s a barrier penetration problem. Such problems involve the
exponentials of terms proportional to (−1/ħ) and are therefore not seen in any order in a perturbation expansion in
powers of ħ, to wit, the loop expansion. Nevertheless, on simple energetic grounds, the higher minimum is an
imposter: a false vacuum.

Figure 44.4: Tilted double well

Another way of talking about the false vacuum is to consider the stationary points. With U(ϕ) we had to look
for minima. For V( ) we just have to look for stationary points, not necessarily minima. Well, what’s wrong
with the maximum between the two minima? Its derivative certainly vanishes there. What’s wrong with it is that it’s
unstable, and not just through barrier penetration.

I’ve claimed that V( ) is an energy density; in particular,

where 0 is the energy density of the ground state. Greater insight is to be gained by demonstrating why this is so.
It follows from minimizing the effective action, Γ[ ], which will amount to minimizing the effective potential, V(
). The argument is simple. I will look at the corresponding problem in ordinary quantum mechanics—determining
the minimum of a perturbed Hamiltonian—find the answer, and then generalize it to field theory, by inserting
integrals at appropriate places and replacing energies by energy densities.

Let’s consider the related problem in quantum mechanics, to find the state ψ such that áψ|H|ψñ is a minimum,
subject to the constraint áψ|ψñ = 1. This problem is often solved using the Rayleigh–Ritz method.27

How do we solve a variational problem with a constraint? We can either deal with it directly, in this case by
using only normalized trial states; or we can introduce a Lagrange multiplier. That is what I shall do here. Instead
of minimizing áψ|H|ψñ I will introduce a Lagrange multiplier, E, and minimize the quantity

I call the Lagrange multiplier E for the obvious reason: vary this quantity and you find that E is nothing but the
energy eigenvalue for H:

Now our field theory problem has a different constraint, namely that á0|ϕ|0ñ is to equal a fixed value, . To find
a quantum mechanical problem corresponding to the minimization of the effective potential (44.35), it’s necessary
to impose a second condition in addition to áψ|ψñ = 1. Let A be some operator (it doesn’t matter what it is). Then
impose

where A is a fixed value. The quantity to be minimized now becomes

With malice aforethought I call the second Lagrange multiplier J. We solve this variational problem with arbitrary J,
and then eliminate J from the problem to satisfy the constraint condition.

Define

at the minimum. The notation is beginning to make this quantum mechanical expression − [J] look a lot like the
corresponding field theoretic expression for W[J] in (44.26):

(The sign difference has to do with U(x) appearing with a positive sign in the Hamiltonian H, and V( ) with a
negative sign in Γ( ), derived from the Lagrangian.) Notice that − [J] is the ground state energy for the
altered Hamiltonian H − JA, because (44.61) is the Rayleigh–Ritz variational problem for H − JA. By a standard
theorem,

We know from non-relativistic quantum mechanics that if you vary the expectation value áψ|H|ψñ of a Hamiltonian
H with a parameter in it with respect to that parameter, you get the expectation value of the parameter’s
coefficient. (The term that comes from varying ψ in áψ|H − JA|ψñ is zero, because [J] is a minimum.) We
have to solve (44.64) to eliminate the Lagrange multiplier J. The energy is obtained from

You compute the function [J], you differentiate it to obtain A, you solve the resulting equation to obtain J in
terms of A, and finally you compute the quantity E(A), the desired result. This is an elementary exercise in non-
relativistic quantum mechanics.

You will notice a certain similarity, stressed by the notation, between E(A) in (44.65) and −Γ[ ] defined by a
Legendre transformation,
Here, W[J] corresponds to the generating functional for a constant external J, since we’re only dealing with
constant fields; the sum of all vacuum-to-vacuum diagrams where we changed the Hamiltonian (44.23) by adding
to it a term

much as we changed H by adding the term −JA. This W[J] is the generating functional for connected vacuum-to-
vacuum graphs (13.11). That is, W[J] evaluated for a particular J is proportional to the sum of all the connected
vacuum-to-vacuum graphs for this altered Hamiltonian, whose expectation value is the ground state energy
density28 in the presence of the source term J. So this quantum mechanical − [J] is exactly analogous to our field
theoretic W[J] (44.24): it is the ground state energy in the perturbed Hamiltonian, just as the field theoretic W[J] is
the ground state energy density. We differentiate this [J] with respect to J to define A, just as we
differentiated the earlier W[J] with respect to J to define . Then we make a Legendre transformation, which will
give us in this case −Γ[ ] evaluated for a constant field, or (44.35) +V( ).

Working out the quantum mechanical problem of determining the ground state energy with a restriction
reproduces every step, including an equivalent Legendre transformation, used to define Γ[ ], and hence V(
). It is the same argument, aside from the substitutions of A for ϕ and energy for energy density, (because
the connected vacuum-to-vacuum graphs give an energy density). Therefore we have proved that, if ϕ is the field
which minimizes V and for which áϕñ = , then

The quantity 0 is the lowest energy density subject to the constraint. In principle, would be the state of lowest
energy density if we obtained from W[J] the state of lowest energy density in the presence of the external source J.
We may not, since we’re computing W[J] perturbatively; we may run into trouble if level crossing takes place.
When the coupling constants are weak, another state that is not the ground state may come up and cross that
energy level, and we may find ourselves following the wrong state as we sum up our Feynman graphs. If
perturbation theory cannot tell us the true ground state energy, then we won’t get the true ground state energy for
the constrained problem, either. On the other hand if perturbation theory serves to give the true ground state
energy without constraint, it will also give us the true ground state energy with constraints.

There is a more direct way to establish (44.66). From (44.24), in the presence of a time-independent J,

because for the ground state, (44.34)

On the other hand,

because for ϕ = , J = 0, and from (44.40) and (44.35),

That is,

so that, with N′′ = 1 for the ground state, (44.66) follows.

Next time I will use this result to interpret V in another way, to explain why V sometimes develops an
imaginary part and therefore why it was a good idea to keep the −iϵ in (44.52). I’ll show that the −iϵ gives that
imaginary part the right sign. I will also discuss V in terms of something we threw away around the third lecture, the
zero-point energy of the ground state. I’ll show that V is just another way of writing down the zero-point energy in
an external field. Then I will discuss, on a much more lowbrow level, a particular model in tree approximation. We
won’t be missing anything, because we have learned the one-loop approximation won’t make any changes. This is
the famous sigma model.29 It will serve as a laboratory for some of the current algebra ideas we were discussing
earlier, in Chapter 41.

1[Eds.] See note 5, p. 936: the Hamiltonian in (44.2) appears as Brout’s equation (33.52), p. 713; and John
Preskill, Notes for Caltech’s Physics 205 (1986–7), Ch. 6, pp. 6.9–6.12; online at
http://www.theory.caltech.edu/~preskill/notes.html.
2[Eds.] See §13.4, in particular note 4, p. 278.
3[Eds.] See Arfken & Weber, MMP, Section 3.5, pp. 215–231, and Problem 3.5.8, p. 227.
4[Eds.] Arthur S. Wightman (1922–2013) was an American mathematical physicist, a founder of axiomatic
quantum field theory and originator of the Wightman axioms. A student of John A. Wheeler’s, Wightman spent
most of his career at Princeton. He is perhaps best known for his book PCT, Spin and Statistics, and All That,
written with R. F. Streater, Addison-Wesley, 1964, republished by Princeton U. P., 2000.
5[Eds.] David Ruelle is a Belgian-French mathematical physicist, well known for his work on statistical mechanics
and dynamical systems. Many others, notably Yōichirō Nambu, Philip Anderson, and Kenneth Wilson—Nobel
winners all—have made major contributions to field theory using ideas from statistical mechanics and condensed
matter theory.
6[Eds.]Much of the rest of this chapter comes from Sidney Coleman and Erick Weinberg, “Radiative Corrections
as the Origin of Spontaneous Symmetry Breaking”, Phys. Rev. D7 (1973) 1888–1910, and “Secret Symmetry”
(Erice 1973) in Coleman Aspects, pp. 113–184.
7[Eds.] See Chapter 28.
8[Eds.] Earlier, in §32.2, was simply a classical field, the argument of the effective action Γ[ ], (32.11), and
not necessarily the value that minimized a potential. In Coleman Aspects, this value is written as ϕ c : “Secret
Symmetry”, p. 312, equation (3.12), likewise in Coleman and Weinberg, op. cit., equation (2.4), p. 1890.
9[Eds] See note 6, p. 689.
10[Eds.] “Secret Symmetry”, Sections 3.4 and 3.5, pp. 135–138, in Coleman Aspects.
11[Eds.] The integral d4x is invariant under the translation x → x + a.
12[Eds.] d4x = d4xe−ip⋅x = (2π)4δ(4)(p) = (2π)4δ(4)(0).
13[Eds.] See equation (2.8), p. 1890 in Coleman and Weinberg, op. cit.
14[Eds.] In Coleman and Weinberg, op. cit., the scalar field was taken to be massless; see equation (3.1), p. 1892.
15[Eds.] Coleman Aspects, “Secret Symmetry”, Section 3.5, p. 136.
16[Eds.] Coleman Aspects, p. 137.
17[Eds.] Gradshteyn & Ryzhik TISP. The relevant integral is number 2.729.2.
18[Eds.]
Coleman’s value in the video of Lecture 49 (at 0:58:42) is incorrect. The value (44.51) agrees with
Coleman Aspects, “Secret Symmetry”, equation (3.33), p. 138 and with the anonymous graduate student’s notes,
as well as with equation (3.4) in Coleman and Weinberg, op. cit. The evaluation of (44.51) is a bit tricky; see
Problem 24.1, p.1003.
19[Eds.] §16.4; “Renormalization and Symmetry: A Review for Non-Specialists” in Coleman Aspects, Section 4,
pp. 104–106.
20[Eds.] Coleman and Weinberg, op. cit., equation (6.3), p. 1900. Erick Weinberg was Coleman’s student.
21[Eds.]Coleman adds that after hearing about this description, Jeffrey Goldstone asked Steven Weinberg, “Who
is pseudo-Goldstone?”
22[Eds.] In Woit’s notes, Coleman remarks that the calculation of the effective potential can be done via functional
integrals and the method of steepest descent, as in his Erice 1977 lectures, reprinted as “The Uses of Instantons”,
pp. 265–350 in Coleman Aspects. For an explicit calculation with functional integrals, see R. Jackiw, “Functional
evaluation of the effective potential”, Phys. Rev. D9 (1974) 1686–1701; Jackiw’s equation (3.5a) coincides with
(44.51) for the case n = 1.
23[Eds.] See Section 3.7 in “Secret Symmetry”, Coleman Aspects.
24[Eds.] In “Secret Symmetry”, Aspects, note 16, p. 139, Coleman states that this result is due to Symanzik: K.
Symanzik, “Renormalizable Models with Simple Symmetry Breaking”, Comm. Math. Phys. 16 (1970) 48–80. As in
“Secret Symmetry”, p. 140, Coleman uses L3 instead of V for a volume, to avoid confusion with V( ).
25[Eds.] See note 11, p. 940.
26[Eds.] S. Coleman, “Fate of the False Vacuum: Semiclassical Theory”, Phys. Rev. D15 (1977) 2929–2936; “The
Uses of Instantons”, Section 6, pp. 327–340 in Coleman Aspects.
27[Eds.] Arfken & Weber MMP, Section 17.8, pp. 1072–1074; E. Butkov, Mathematical Physics, Addison-Wesley,
1968, Section 13.5, pp. 565–567; F. Mandl, Quantum Mechanics, J. Wiley and Sons, 1982, Chapter 8, pp.
186–193. In quantum mechanics, the Rayleigh–Ritz method is also known as “the variational method”. The
method is sometimes posed as the variation of the ratio áψ|H|ψñ/áψ|ψñ for a trial function ψ(x).
28[Eds.] §32.2.
29[Eds.] See note 23, p. 994; the sigma model is covered in §§45.3–45.4.

45
Topics in spontaneous symmetry breaking

This chapter is a miscellany of three topics. First I’ll discuss the role of the negative imaginary part of the energy
(which came from the Feynman prescription for the propagators in (44.49)) in the effective potential. Next, I’ll
extend the effective potential to theories containing fermion fields. Finally, I’ll construct the famous sigma model of
four scalar fields: an isospin singlet, the sigma, and the pion triplet. It incorporates two kinds of symmetry
breaking, both spontaneous and explicit “soft” symmetry breaking. The model is constructed so that PCAC is
satisfied and gives the Goldberger–Treiman relation. More importantly, it provides a mechanism for the observed
small mass of the pions.

45.1Three heuristic aspects of the effective potential

Before proceeding let’s review some things.1 To construct the effective potential, we add a constant source term
to the Lagrangian. This changes the Hamiltonian:

That Hamiltonian is well-defined. It’s time-independent and it has a ground state |0ñ, and á0|ϕ(x)|0ñ = ϕ. We don’t
have to put an “in” or an “out” on the vacua, because if J is independent of time, |0ñout is the same as |0ñin. The
ground state just lies there; it doesn’t scatter. The energy without the source term is the volume of space (if we put
everything in a box) times V(ϕ). The prescription for the source term is: add Jϕ to the Lagrangian such that we
obtain a ground state in which ϕ(x) has the desired expectation value ϕ. In that state there will be a certain energy,
V(ϕ):

At the minimum of V we have

So at the actual minimum of V(ϕ) we don’t need a source to produce that ground state. That’s the general picture.

I’m going to put a little more flesh on this general picture of the effective potential by telling you how the
interpretation of the effective potential as an energy density explains something peculiar that could happen. Recall
the master formula we got last time for the case of a single field:

(I’m proud of that 1/(64π2).) But there’s something peculiar about this formula:

When the real part of the argument of the logarithm is negative, the −iϵ (which I’ve carefully retained for just this
purpose) gives you a negative imaginary part.2 (By the way, U′′(ϕ) in our standard model, pictured in Figure 43.4,
becomes negative for an interval near the origin.) But how can you have an energy with an imaginary part,
whatever its sign? Well, to make ϕ have the desired vacuum expectation value ϕ, we apply an external
perturbation, J(x). But it may not be possible for áϕñ to equal ϕ.

Consider classical electrodynamics (this involves vector fields rather than scalars, but the principle is the
same). We can apply an external charge distribution such that the electric field in some region has a given desired
value. (This is analogous to adding a J in (45.1) to make áϕñ a given value.) In particular, we can arrange that the
electric field has absolutely any value we want within that region, independent of space and time, by bracketing
the region between the charged plates of a large condenser.3 That gives us a constant electric field. In quantum
electrodynamics, however, we cannot fill a region with a field this way: the vacuum suffers dielectric breakdown.4 If
we have erected these giant condenser plates, even at opposite ends of the galaxy, and applied external charges
on them such that a constant electric field arises over the whole extent of the galaxy, it will be energetically
favorable for an electron–positron pair to materialize from the vacuum. Though that costs 2mc2, the system gains
energy when the electron files to the positively charged plate and the positron flies to the negatively charged plate.
That’s the product of the electron’s charge, the size of the electric field and the distance between the condenser
plates. The kinetic energy gained can be much greater than the 2mc2 lost in creating the pair. In that case the
vacuum boils off pairs until the charge on the condenser plates is neutralized, just as an ordinary dielectric in a
real condenser breaks down because of the atoms in it ionizing. You end up with zero electric field, no matter what
charge you try to put on the condenser plates. After all, the vacuum is a dielectric and can be polarized; that’s the
statement that the photon self-energy operator is not zero. If the region of space in which an electric field exists is
large enough, as long as the field magnitude is non-zero, this will happen. This is an example of the famous
totalitarian selection principle: “Everything that is not forbidden is compulsory.”5 If it’s energetically allowed, it’s
going to happen.

Something similar might happen with the scalar field. If we try to maintain a given value of the scalar field, it
may well be that some phenomenon akin to the boiling off of electron–positron pairs could occur to neutralize the
scalar field. If so, a configuration with a fixed expectation value of the scalar field will not be stable; it will decay.
For example, consider a theory of protons and neutrons and π mesons, all with arbitrary masses. We define the
energy of the neutron and expect its mass to be the pole in the neutron propagator. Suppose we alter the
parameters of the theory, and say that it’s not required to be isospin invariant. The neutron’s mass could become
larger than the sum of the proton’s mass and the π− meson’s mass. Should we reach that point, we’d find that our
nice energy formula had developed an imaginary part because the pole would move onto the second sheet.6 This
argument strongly suggests that an imaginary part in the energy density is a sign of instability, just as when we
follow real energy in the neutron example up to a certain point; it develops an imaginary part which is connected to
the neutron lifetime.7 I’ll now demonstrate that this negative imaginary part is equal to half the probability Γ of the
neutron’s decay per unit time.8

Recall that we found (44.67) for a time-independent source J within a box of volume L3 over a time T

where 0 is the energy density of the ground state of the perturbed system. If ϕ is the field which minimizes V and
for which áϕñ = ϕ, then

If V has a negative imaginary part (the −iϵ as in (45.2)), the amplitude develops a certain probability of
disappearing by boiling off pairs in that box. Thus it is important that the imaginary part is negative. That’s a
consistency check on this picture, because just as the neutron energy moves onto the second sheet, so the
energy density here should move onto the second sheet:

The probability Γ/2 of decay of this state per unit volume, the imaginary part of the energy, is the imaginary part of
V(ϕ) when U′′(ϕ) < 0:
plus higher loop corrections. The bigger the box the larger the chance it can decay, because there are more
places for it to decay into. This decay can occur, but it is exponentially damped.9 It’s not like barrier penetration,
because it occurs at the one loop level; it’s like a particle balanced on top of the peak in Figure 43.4. On the other
hand, if the potential were asymmetric, as in Figure 44.4, the higher minimum is unstable because of tunneling,
described by the factor exp{−(Γ/2)t}. We couldn’t tell at the one loop order that the higher potential was the wrong
choice. In such an instance we expect the maximum to reveal itself at (ħ), two loops (32.10). If this were a
potential in classical mechanics I could balance a particle on the peak. In quantum mechanics, as Heisenberg
pointed out, I can’t do that, because quantum fluctuations will cause it to move laterally away from the peak. And
as soon as a quantum fluctuation brings it to one side, off it falls. The behavior of V(ϕ) when U′′(ϕ) is negative
provides a fuller picture of the effective potential as an energy density and explains what happens when it
acquires an imaginary part.

It’s also instructive to go from a field theory in four dimensions to a particle system to one dimension. Instead
of ϕ we’ll call the dynamical variable x, a function of a single variable which we’ll call t. The Lagrangian is:

(Note that we are using a unit mass.) There’s no need for renormalizations here; nevertheless we could define a
V( ) (with the average value of x), and the formula would be exactly the same:

We threw away constants rather blithely last time (44.51) because they were absorbed in the renormalizations in
our four-dimensional theory. In a one-dimensional theory there is no need for any infinite renormalizations, but
unfortunately we get a constant here.10 The integral is easy to do using the boundary condition that V = 0 if U = 0,
a reasonable assumption: a free particle should not have any potential energy. Let

Then

and so

which gives, choosing the constant appropriately,

This result is very satisfying. Remember that the calculation is to one loop order. The loop expansion can be
thought of as a systematic expansion in powers of ħ. You have a particle of unit mass sitting in a potential U(x),
assumed symmetric so we know the ground state is at the origin; see Figure 43.3. In classical mechanics, its
energy would be U(0) at x = 0, the origin of the potential. What is the first quantum correction to the energy? The
particle moves as if it were a harmonic oscillator, in a quadratic potential whose frequency is given by the square
root of U″; we approximate the potential at its minimum:

You all know this game from the study of molecules: you add the zero point energy of the harmonic oscillator with
an ħ which has been suppressed here (we’re using units where ħ = 1):

This is the classical energy plus the zero point energy of a harmonic oscillator about the classical minimum. That’s
exactly right, just what you would expect for the first quantum correction to the energy. We also see from this
formalism that by combining these two analyses, we could solve Heisenberg’s problem: What is the probability per
unit time that a particle sitting at the origin of Figure 43.4 falls off the peak? That is left as an exercise.

So our expression for the one loop effective potential looks good. We understand the imaginary part when we
go to a system that we know well, particle quantum mechanics. It gives an intuitively right answer. In fact we can
get an idea of where (45.11) comes from by going back to four dimensions and doing the integral in a different
way. Take the four dimensional expression

and break it up into a time part and a space part.

Let’s do the time part using the result (45.10):

What is this equation saying? Remember way back: I said a free field theory was like a system of harmonic
oscillators. We blithely threw away a contribution to the energy, namely the zero point energy of the oscillators
summed over all the oscillators. 11 What we’ve got here is the analog of that expression. How did we get (45.11)?
We said

We neglected the higher order terms, solved the harmonic oscillator problem and got a zero-point energy. Here we
are doing something analogous for ϕ:

In (45.16) we’re computing the zero-point energy of the system, the energy of each oscillator in a free field theory
with squared mass U″(ϕ), summed over all the oscillators. Our mysterious one-loop computation (44.51), which
involved all those fancy summations and Feynman diagrams, is revealed to be exactly the same in every factor,
including the for the oscillator, as simply computing the zero-point energy of a free scalar field.

These three points are only heuristic: the −iϵ helps explain what happens when U′′ is negative; in one
dimension, the formula recapitulates the first quantum correction to the energy; and the result (44.51) is equivalent
to the calculation of the zero-point energy of a free field. They’re not essential to doing any computations—the
essential formula for calculations is (44.56)—but they help to provide physical meaning to the effective potential.

45.2Fermions and the effective potential

I turn now to a different technical problem that will lead to some new physics. I’d like to compute the effects of
fermions on the effective potential, because we will eventually have to deal with field theories that contain both
bosons and fermions. I don’t intend to generalize the effective potential to be a function of Fermi fields. That’s silly:
a Fermi field never gets a vacuum expectation value, by Lorentz invariance. Nevertheless there are 1PI graphs
with external lines restricted to bosons that have fermions running around the loop, for instance Figure 45.1. I
want to compute these fermion loops to get the one-loop corrections to the effective potential from any fermions in
the theory. I know how to take care of all the spinless bosons; we have that master formula. I now want to include
the fermions.

Figure 45.1: Fermion loop with external boson lines

We’ll have a Lagrangian that’s as before, plus a bunch of Fermi fields which we’ll indicate by an index a:
The ϕ Sa in the gabc term are scalar fields, the ϕ Pc in the fabc term are pseudoscalar fields; collectively we’ll refer to
the fields {ϕ Sa, ϕ Pb} as ϕ. Typically we choose the mass matrix mab to be diagonal but I’ll work in an arbitrary
frame; it doesn’t matter. The relevant counterterms for these kinds of graphs will be those which are functions of
Bose fields only. The Fermi counterterms will not come in since those correspond to graphs with external Fermi
line, and do not appear in this calculation. This is the most general possible renormalizable Lagrangian involving
spinless bosons and fermions.

There are some constraints on the matrix mab. A Lagrangian should be real, so mab must be Hermitian, and
since the ϕ’s are real fields, the coefficients g and f must also be Hermitian with respect to a and b:

This makes it convenient to write by writing the fermions in matrix form, assembling them into a big vector:

where m(ϕ) is the matrix

This matrix mab(ϕ) is not Hermitian because of the i in the γ5 term: γ5 is Hermitian, iγ5 is not (20.103). However, it
does obey a very nice relation, which we’ll exploit shortly:

Commuting γµ with γ5 gives us a minus sign that takes care of the minus sign introduced by the i. The matrix m(ϕ)
is well-named because it is the mass the fermions would have if you replaced the quantum fields ϕ by constant c-
numbers. That would be the fermion mass in tree level, if the particular values of ϕ happen to give the tree value
minimum of the potential.

We can sum up the contribution to the effective potential from the fermion loops. It’s exactly the same
computation as before (44.42). I draw the fermion loops with heavy black dots, each representing a factor of
−im(ϕ), −m(ϕ) from and i from the vertex:

That tells you how the fermions couple with the external field ϕ. There will be the usual terms in the sum of the
graphs coming from the boson loops and the counterterms. From the fermion loops, we’ll have an integral over all
momenta,

and a factor of the fermion propagator (once again splitting off the mass, as with the previous calculation)12

times −im(ϕ), both raised to the nth power. Then there will be combinatoric factors, in this case 1/n rather than
1/(2n), because if we reflect the diagram we get a different graph in which the line runs around the other way. We
sum over n and take the trace over everything, the Dirac indices as well as the things that label the fermions, and
add a minus sign for the fermion loop:

Only the even terms in the series contribute, because the trace of the product of an odd number of gamma
matrices is always zero, whether or not there are γ5’s infiltrating.13 We take account of that by replacing n by 2n,
indicating we’re only going to sum over the even terms of the series.

Let’s examine one of those terms by itself. Looking at the quantity in the square brackets, the i cancels the −i
so we don’t have to worry about them. We will have to raise to the nth power the quantity

where we have used (45.23) and canceled the k2 in the numerator against one in the denominator. Aside from the
minus sign in front, the series is the same as the boson series (44.49), with m†m in place of U″. So we just write
down the answer. (It will turn out to be just as easy for a very complicated theory with non-Abelian gauge fields
when we get to that.) This fermion contribution has a factor of (−1) for the loops, the same factor of 1/(64π2), and
the trace over {(m†(ϕ)m(ϕ))2 ln[m†(ϕ)m(ϕ) − iϵ]}, but otherwise it’s identical to the boson contribution:

(If there are n species of fermions, then m is a 4n × 4n matrix.) As with the boson series (44.53), it doesn’t matter
what M we choose; it will affect the size of the counterterm, but it won’t affect the sum.

We can understand this form in exactly the same way that we understood the scalar case—as a shift in the
zero-point energy. I haven’t talked about it, but you’ve probably all read about Dirac’s old electron theory, in which
he had a bunch of negative energy levels, the Dirac sea14, which he filled up. He said the real vacuum is the state
in which the negative energy levels are all occupied, and we tacitly agreed. We threw away a constant when we
normal ordered the Hamiltonian (21.26)—that constant was the sum of the energies over all the negative energy
levels.

Let’s suppose we have only a single type of Fermi field and m is a constant; then m†m is m2. We’ll get a
minus sign and four times the corresponding results for bosons. In the Dirac theory you don’t have zero-point
energies, you have all the negative energy levels; you get a minus sign from filling up those levels. When you
increase the mass, the negative energy levels for a given k get more negative. The 4 comes from the trace, but
there’s a more physical way of thinking about it. Before you had ω from the zero-point energy. But when you fill a
negative energy level, you need ω, not ω. That’s one factor of 2. The extra factor of 2 arises because the Dirac
particle has two directions of spin; you’ve got twice as many energy levels for Dirac particles of a given mass as
you do for Bose particles. That’s why you get (−4) times the earlier result.

This is not only pretty physics, it has a practical application. I told you that there was really no point in
computing the effective potential because, after all, we’re doing a one-loop correction, which is just a small
coupling constant approximation. It is not quite perturbation theory, but like perturbation theory, if it’s good for
anything, it can be useful only for small coupling constants. Unless we’re computing something like the
anomalous moment of the electron, where we have both high precision experiments and a theory in which we are
confident, why should we bother with anything beyond tree approximation? The one-loop correction just shifts
things around a little bit; it doesn’t change the qualitative picture which is all we’re interested in. There are in fact
two cases in which the one-loop corrections are important.

The first case is a theory with both bosons and fermions, with two coupling constants: λ, a quartic coupling
constant, and g, a Yukawa coupling constant. The small loop expansion is good only if both of these
dimensionless quantities are small compared to 1:

These are constraints when we do perturbation theory, including loop expansions. On the other hand, g never
appears on the level of the tree approximation for the effective potential; we are not considering external fermion
lines. There’s a perfectly reasonable possibility that somewhere in the range of coupling constants that we’re
interested in, both of these conditions are true but g is much greater than λ:

If that’s so, the dominant terms in the effective potential, even though all the coupling constants are small, would
be the term we just computed,

despite the fact that it occurs on the one-loop level. So what? It’s the first term that has a g in it. The qualitative
features of spontaneous symmetry breakdown will be dominated not by the tree approximation but by this monster
(45.28). Though that possibility is recherché, it’s nevertheless within our abilities to investigate it. One thing we
surely know how to do is perturbation theory, so we can investigate the domain of coupling constants in (45.30).
To do that we have to look at (45.28), because this is the first time we see g.

The second case is something proposed by Steve Weinberg, called accidental symmetry. It is best
explained by an example.15 Let’s go back to our old friend, the SU(3)-invariant meson–nucleon theory. SU(3) is
not spontaneously broken but it will make a good example. We have ϕ and ψ, boson and fermion SU(3) octets,
respectively. The Lagrangian is a free Lagrangian plus a quartic term. It may appear at first glance that you could
write down two quartic couplings, Tr(ϕ 4) and (Tr(ϕ 2))2. In fact there is only one; the first equals half the second.16
There’s some coupling constant λ, a D-type coupling constant gD and an F-type coupling constant17 gF:

Now if we do tree approximation this theory is not just SU(3) symmetric, it’s SO(8) symmetric: (Tr(ϕ 2))2 involves
the sum of the squares of the eight real meson fields and is invariant under orthogonal transformations of the
fields. The real theory is by no means SO(8) invariant, but only SU(3) invariant, as real theories tend to be.
There’s no way of defining an SO(8) transformation on the ψ’s. Even so, in tree approximation we’ll find an
effective potential which is SO(8) invariant. If we introduce a negative mass squared term to induce spontaneous
symmetry breaking, we’ll be completely stuck: we’ll have an SO(8)-invariant family of minima rather than an
SU(3)-invariant family of minima. In that situation there’s no way of telling which among the minima is the true
vacuum. If it weren’t for the Fermi terms, this wouldn’t matter; the theory really would be SO(8)-invariant, and
whichever vacuum you take would be as good as any other. But the Fermi terms are going to make a difference.
They’re not SO(8)-invariant, and in this vast smooth field of possible vacua they’re going to introduce little hills and
valleys no matter how small the Fermi coupling constants are. Those little hills and valleys are going to determine
the true vacuum. So even if the Fermi coupling constants gF and gD are much, much smaller than the quartic
coupling constant λ, you have to include the term in (45.28) in order to find out qualitatively what the true vacuum
is.

This is called accidental symmetry because by an accident the effective potential in tree approximation
admits a large symmetry group that has nothing to do with the real symmetry of the theory, or even the symmetry
of the effective potential beyond the tree approximation. You have to go beyond the tree approximation to gain
even qualitative information about the vacuum state. Once you’ve determined the true vacuum, you’ll find that
small oscillations about it are peculiar. There will be some directions that break the SO(8) symmetry, in which the
potential is curved up rather sharply by a λ-dependent amount. There will be other directions where the fermion
terms Tr({ , ϕ}iγ5ψ) and Tr([ , ϕ]iγ5ψ) put a small dimple in the potential, curving it up only a little if the
Yukawa couplings are small. Therefore when you explore the second derivatives of the effective potential to get
the masses of the particles, or more precisely the inverse propagators evaluated at zero momentum transfer,
you’ll find that some particles, those that correspond to directions where only the fermion terms keep the potential
from being flat, will be much, much less massive than others. These small mass particles are called pseudo-
Goldstone bosons because, in comparison with every other particle in the theory, their masses are not so far
from zero (genuine Goldstone bosons have zero mass).18 Pseudo-Goldstone bosons would appear as massless
Goldstone bosons in tree approximation, except that—only because of the Fermi loop corrections—they acquire a
small mass. I will not present any actual examples where fermions do either of these things, but I will shortly give
an example where vector mesons do both of these things. I will discuss the famous lower bound on the mass of
the Higgs meson in Weinberg’s theory of the weak interactions which comes from precisely this phenomenon.

45.3Spontaneous symmetry breaking and soft pions: the sigma model

So far I’ve discussed spontaneous symmetry breaking in a general way. I’m now going to turn to specific
examples. I will first investigate a particular model where I’ll apply the ideas of spontaneous symmetry breaking to
our old current algebra soft pion computation. Recall that I described four different meanings that one assigns to
PCAC.19 One was the statement that ultimately derives from Nambu.20 He said that you could derive the
Goldberger–Treiman relation by asserting that in the limit that the pion mass goes to zero, the axial vector current
is conserved:

Nobody knew how to make sense out of this; at the time it seemed like an orphic statement. The computations
based on this limit agreed with experiment, but the meaning of the limit was mysterious. Later on, Jeffrey
Goldstone and Nambu himself unraveled it. There is in fact an instance where the vanishing of a particle’s mass
is associated with conservation of the current: spontaneous symmetry breaking, with the particle a Goldstone
boson. This suggests interpreting Nambu’s statement as follows: chirality, the symmetry generated by the axial
vector current, is an approximate symmetry. But this approximate symmetry is not associated with explicit
symmetry-breaking terms in the Lagrangian; instead, the approximate symmetry would be realized in the
Nambu–Goldstone mode (43.95) associated with spontaneous symmetry breaking.21 If that mode of symmetry
breaking occurs, the limit runs the other way:

In this case, the three pions, which have precisely the right quantum numbers to be the Goldstone bosons
associated with the axial vector mesons, would become massless. That is, the pions are almost Goldstone
bosons.

Consider a Lagrangian which has a chiral-invariant part and a chiral-breaking part. When the chiral-breaking
part goes to zero you don’t have to wind up with perfect manifest symmetry; you could get unbroken symmetry for
the isospin subgroup of SU(2) ⊗ SU(2), while the symmetry of the chiral generators (from the axial vector currents)
could be spontaneously broken. That is, for the six-parameter group SU(2) ⊗ SU(2) (or equivalently, SO(4)), we
will have a three parameter subgroup manifest and unbroken, and another three parameter subgroup’s symmetry
spontaneously broken, leading to three Goldstone bosons. We’ll construct a model with a small, explicit chiral
symmetry-breaking term in the Lagrangian, small because the pion mass is small compared to the other hadron
masses. The idea is that if that term were to vanish then the symmetry would be exact, but it wouldn’t be manifest;
it would be hidden—spontaneously broken.

Another clue in that direction is our old gradient-coupling model (41.32), with the axial vector current (41.40)
given by22

though we’re not interested in the nucleon terms. Remember that this current would be conserved (its 4-
divergence would equal zero) if the pion mass were zero. It is however an instance of spontaneous symmetry
breaking. We have

because that’s ∂0ϕ with ϕ. It is in fact very close to our simplest example (43.102) of spontaneous symmetry
breaking and the Goldstone theorem, the free scalar field where the current is ∂µϕ. In hindsight, we were
investigating spontaneous symmetry breaking in the gradient-coupling model, although we didn’t know it. Its
conserved current (41.40) is like ∂µϕ in the free scalar theory. The gradient-coupling current has a correction
coming from the nucleon, but otherwise it provided an example of spontaneous symmetry breaking: as the pion
mass went to zero we had a conserved current, whose commutator with the pion field had a non-vanishing
vacuum expectation value. The gradient-coupling model is unsatisfactory in many ways. Although it’s perfectly
fine in tree approximation, it is non-renormalizable so we can’t use it to investigate higher-order corrections. In
addition, its current algebra is not the one we believe is true in the real world: the axial vector currents commute
with themselves.

I will now use the ideas of spontaneous symmetry breaking to create a renormalizable model of the interaction
of nucleons and pions which has both the right algebra and a symmetry in tree approximation (which is all we can
really investigate) realized in the Nambu–Goldstone mode, if the pion mass were zero. (This won’t involve strange
particles, just pions and nucleons.) We’ll presume we know how the axial vector currents act on the nucleons. I’ll
construct the simplest renormalizable model with these characteristics, including some scalar fields to break the
symmetry spontaneously. This will end up being the sigma model. The model will involve nucleons, isodoublets
of Dirac spinors; N, in the standard notation of Gell-Mann and Lévy:23

and an isotriplet of pion fields πa; and some other scalar or pseudoscalar fields. With the latter we can make the
symmetry break down spontaneously; we know how to do that by choosing the quadratic term in the Lagrangian
appropriately. And we want the axial vector currents Aµa to obey the correct current algebra.
 Let’s begin by examining the transformation properties of the nucleon. We know the contribution to the vector
current, normalized to be the isospin current:

the appearing because the fermions have isospin (the Ta are the three isospin analogs of the Pauli matrices).
That’s the CVC hypothesis. There may be other terms involving the other fields, but since we don’t know what they
are at the moment, we will keep them loose. Under an infinitesimal vector transformation

This group of transformations may be denoted SU(2)V . To keep from filling up the equations with 2gV ’s, I will
assume

(You can rescale the currents (45.37) by 2gV if you like; then it will obey the isospin relations and fulfill our current
algebra prescriptions.)

We know how to make an axial vector current that has the right commutators (we get to these not with
nucleons, but with quarks):

which corresponds to an infinitesimal transformation

(Both of these have minus signs because the Ta’s are Hermitian and (20.103) γ5 = −γ5.) This group of
transformations acts like SU(2), but unlike SU(2)V, the signs of these transformations on N and N are the same;
because of the γ5’s, this is an axial version of SU(2), which may be denoted SU(2)A. The two currents, vector and
axial vector, can be obtained from the Lagrangian

if the transformations are symmetries of this Lagrangian.

Let’s check that these are symmetries in the case where the nucleon is massless. The vector symmetry is
trivial, because is isospin invariant:

so all we really have to check is the axial symmetry:

The anticommutation of γ5 and γµ supplies the relative minus sign. Likewise the derivatives are invariant:

So the (rather trivial) theory of free massless nucleons does indeed possess these symmetries. We haven’t put
any scalar fields in the picture so we don’t yet have the possibility of spontaneous symmetry breakdown.

But we run into trouble if we try to add a nucleon mass term. There’s no problem with the isospin, (45.38):

The axial vector transformation (45.41) is another story:

We don’t get the cancellation in the axial transformation, we simply get the two terms adding together. That’s to be
expected; we’ve seen this before.24 Once again, a fermion mass term breaks the chiral invariance: we can’t have
an invariance with γ5 in it if we’ve got a fermion mass in the theory. Because I want to keep this chiral symmetry,
the model Lagrangian cannot include an explicit nucleon mass term.
 Perhaps you’ve recognized this set of transformation laws. Recall that there’s a local 2-to-1 isomorphism
between the SU(2) ⊗ SU(2) current algebra group and SO(4), the four-dimensional rotation group.25

It’s exactly the same analysis we went through for the Lorentz group, SO(3, 1), except there are no minus signs.
So we get SO(4) instead. The infinitesimal vector (isospin) transformations D Va correspond to rotations in the i-j
plane, where i, j = 1, 2, 3. The axial vector transformations D Aa are the analogs of the Lorentz boosts; they’re
rotations in the i-4 planes. These transformation laws are those of an infinitesimal four-vector (in R 4, not
Minkowski space) with NN being the fourth component and Nγ5TaN being the three standard space components
(in R 3). By a trivial computation

exactly what we would expect26 for an analog of Lorentz boosts with 4-vectors: the fourth component goes into
one of the three space components (45.47); the three space components go into the fourth component (45.49)
with a relative minus sign to keep the sum of the squares fixed:

This suggests that we can very simply add in Yukawa-type couplings by introducing a quartet of mesons (σ,
π) that have the same transformation properties as the nucleon bilinear products; that is, as a Euclidean four-
vector under the group. We’ll have a singlet σ that will transform as the fourth component of a vector, and three
vector components πa that go into the singlet under the axial transformations:

The three pion fields πa we will regard as having the properties of real pions, and so they will end up being
pseudoscalar; the σ is an unobserved particle. These transformations are summarized in Table 45.1.

Table 45.1: Axial transformations of the sigma model’s fields

We can now write down a much more complicated Lagrangian. The nucleons are still massless but we can
add an invariant Yukawa interaction, a coupling constant times the fourth component of one vector times the
fourth component of another plus (not minus!) the dot product of the respective vectors’ three-vector components:

We’ll write down the most general renormalizable field theory consistent with all this. We have the sum of the four
components of the four-vectors in SO(4), a quartic coupling and a mass term:

That’s the most general renormalizable, SO(4)-invariant Lagrangian of πa and σ: its interactions are the sum of the
most general quartic coupling, the most general quadratic coupling and the most general Yukawa coupling. The
vector current is the ordinary isospin current,

The axial vector current is

 45.4The physics of the sigma model

Let’s make the value of µ2 negative, and add a constant so that the SO(4) symmetry breaks spontaneously.
Making this choice, the Lagrangian can be written as

The possible vacua (the minima of the potential) lie on the sphere

Because the theory is completely SO(4) invariant, it doesn’t matter which vacuum we choose. We pick

(With any other choice for |0ñ, we’d have to redefine parity.) Thus the axial symmetries are broken. The vector
symmetry (isospin) remains; (45.58) are isospin-invariant statements. We define as usual

Expressed in terms of the shifted fields the Lagrangian becomes

Note that the nucleon acquires a mass, mN = ga, as a consequence of spontaneous symmetry breaking, despite
the fact that the Lagrangian is chirally invariant; it comes from the NNσ term. The σ also gets a mass, while the
three components πa remain massless; they are Goldstone bosons:

This is just a model; the σ does not correspond to an observed particle. (Some had suggested that the σ could be
identified with a broad resonance in the 0+ channel of π–π scattering.27)

Though the theory has all these different particles with different interactions, it is renormalizable. It has only
three parameters after spontaneous symmetry breaking, because it had only those to begin with: g, λ and a.
Among the particles is a massive nucleon. But the nucleon mass is not a free parameter, and that is preserved in
the renormalization. The nucleon mass will get corrections under renormalizations of higher order in g and λ, of
course, but they will be finite corrections that will not require any new counterterms: this is still a three-parameter
theory.

Let’s check the Goldberger–Treiman relation (40.37). We know how to do that in a world which has massless
pions.28 We’ll just take Aµa (45.55) which has contributions from both σ and πa, and make the shift σ → σ′ = σ − a.
We’ll get quadratic terms as before, and a ∂µπ term as a consequence of the shift, from the term linear in σ (the
term linear in ∂µσ is unaffected by the shift):

The PCAC statement in the form involving the divergence of the axial vector current is not applicable in a
massless theory. But in the form (40.26)

it can be studied in the massless theory. That’s the primary definition of Fπ. In tree approximation there is only one
term that contributes:

We also see by comparison of the currents with

that
The assumption 2gV = 1 works out. But we cannot get around the equality between gV and −gA (empirically they
are in the ratio ~ −4 : 5); that comes from terms unaffected by the shift. There are some things we can’t do no
matter how cunning our model is.

The Goldberger–Treiman relation says that

Well, does it work? Of course it works! It has to work, because of (45.62); that’s not PCAC, but it’s close enough.
From (45.61) we have mN = ga, from (45.64) Fπ = a, and from (45.65), gA = − . Then

This model gives us conserved vector currents, but it’s not a good model for the real world. The dynamics of this
model are not trustworthy: it has a σ and it doesn’t have strange particles in it. There are no fundamental fields for
σ, π or N; there are only fundamental fields for the quarks. The real world is nearly chiral SU(2) ⊗ SU(2)
invariant:29 the up and down quark masses are very small. Goldstone bosons are an inevitable consequence of
the breakdown of a non-gauge symmetry. The physical pions are so much less massive than the other hadrons
because they are almost Goldstone bosons.

Still, the sigma model provides a nice world to explore. Let’s add a term to that will give us PCAC:

where ′ is going to break the symmetry. To obtain PCAC we must choose

The divergence of the axial current is this change;

Is there an object around in our theory which has a change ∝ πa? Yes, the σ field, with transformation (45.51).
Therefore we choose

where c is a constant. The potential is then

This is guaranteed to put PCAC into the model. A term linear in σ has dimension 1; it is the most general such
term that breaks SO(4) symmetry and preserves isospin symmetry. Does the new term spoil the renormalizability
of the original model? No, because of Symanzik’s rule for “soft” symmetry breaking: terms of raw dimension ≤ 3
added to a renormalizable theory do not alter its renormalizability.30 Therefore the original sigma model with this
term added is still renormalizable. But now we no longer have so many vacua. Instead of the sphere described by
(45.57), we have an asymmetric surface of solutions; there is now a unique ground state because the axial
symmetry has been broken. The minimum on the right of Figure 45.2 is at áσñ, and the one on the left is now a
false vacuum, because of the linear term. (To make it easier to draw, I’ll sketch the potential with π • π = 0.) The
ground state is unique because the axial symmetry has been broken.

Figure 45.2: Tilted double well potential in the sigma model

There’s no point in writing the Lagrangian with c as an independent parameter; I may adjust it as I want.
Formerly we set áσñ = a. I wish to replace (45.64) by

I want the shift to result in this equality. We still have

in tree approximation. The nucleon mass becomes

All that will be much the same as before, with áσñ now replacing a. The amount we have to shift σ will be different,
which we determine by an appropriate choice of c. The only interesting nontrivial parts of the Lagrangian are the
quartic and linear terms.

We’re going to use c to fix áσñ, which will no longer equal a. Shifting the fields

the Lagrangian becomes

Let’s expand the λ term:

The quartic term is still symmetric in tree approximation, and has no reference to the shift parameters c and a.
The cubic term comes from the cross term and is completely determined in terms of λ and áσñ; no new parameters
are involved. The quadratic term comes from two places, cross terms and the square of σ′. We’ll leave the
constant terms alone, but we’ll eliminate the linear terms by a particular choice for c:

Let’s eliminate the parameter a in terms of something more physical (as we did before, when we set a = áσñ). The
pion mass comes from the coefficient of π′ • π′:

Likewise, the coefficient of σ′2 gives the mass of the σ′:

We already have the mass of the nucleon, (45.75). Rewriting the potential,

plus a constant, which we can drop. Rewriting the Lagrangian one last time,

That’s it. Even after the shift, this is still a four-parameter theory. The original version of this theory had
parameters a, c, λ and g. We’ll keep λ and g, but we have traded a and c for áσñ and mπ′2:

The assignments to c and áσñ assure that PCAC in the form (41.5) is satisfied.

To look at this in a slightly different way, the potential U can be written as

where U 0 is the σ model potential before the addition of the c term. The previous condition that determined the
expectation value of σ

is replaced by

This equation says “You give me a áσñ, and I’ll give you a c that will satisfy that equation.” It agrees with (45.79). It
is true to all orders in perturbation theory because we’re adding only a linear term to the Lagrangian. That just
adds a corresponding linear term to the effective action and therefore to the effective potential:

That’s the only c-dependent term in the effective action. V0 is the effective potential for the σ-model without an
explicit symmetry-breaking term, evaluated to whatever order in perturbation theory you wish. We can always
forget about c and say that áσñ is a free parameter.

This theory looks absurdly complicated. It’s got a quartic interaction, a cubic interaction, mass terms, the σ
mass is connected in a fancy way to Fπ and to λ, the quartic coupling constant, and the nucleon mass is
connected to Fπ and to g, the Yukawa coupling constant. What a funny Lagrangian! If I had just written it down to
start with and said “Here is a model that obeys all the current algebra constraints”, you would have laughed me
out of the room. Not only does it obey all the current algebra constraints, it’s renormalizable, needing only four
renormalization constants plus wave function renormalization. To check things out, we would want to take this
grotesque model and compute (in tree approximation) some of the things we computed with our general
assumptions to make sure that it agrees with the earlier results. Then we could be confident that we know what is
going on.

Next time we come to the Higgs mechanism, and you’ll see how to make Goldstone bosons disappear.

1[Eds.] “Secret Symmetry”, Section 3.7, pp. 139–142 in Coleman Aspects.

2[Eds.] For x > 0, ln(−x) = ln(x) + iπ (choosing the principle branch), but if both x and ϵ are positive, ln(−x −
iϵ) = ln(x) − iπ.
3[Eds.] In British English, “condenser”; more typically in American English, “capacitor”.
4[Eds.] For reference, the field strength required for the dielectric breakdown of air is 3 × 106 V/m. J. S. Rigden,
Macmillan Encyclopedia of Physics, Simon and Schuster, 1996.
5[Eds.]
Technically, the “Principle of Compulsory Strong Interactions”: M. Gell-Mann, “The Interpretation of the
New Particles as Displaced Charge Multiplets”, Nuovo Cim. 4, Supp. 2 (1956) 848–866, footnote on p. 859; and T.
H. White, The Once and Future King, Ace Books, 1966, p. 121.
6[Eds.]R. J. Eden, P. V. Landshoff, D. I. Olive, and J. C. Polkinghorne, The Analytic S-Matrix, Cambridge U. P.,
1966, Section 4.4, pp. 205–211; §§17.1–17.2, pp. 356-361.
7[Eds.] The video for Lecture 50 has a minute-long gap at 13:50, and the lecture notes describing this topic are
either incomplete or refer to “Secret Symmetry” in Coleman Aspects. What follows from here to the words “a
certain probability of disappearing” is an interpolation based on “Secret Symmetry”, p. 142.
8[Eds.] Adding a negative imaginary term to the potential is a standard device for representing unstable particles.
See, e.g., David J. Griffiths, Introduction to Quantum Mechanics, 2nd ed., Cambridge U. P., 2016, Problem 1.15,
p. 22. Don’t confuse the decay probability Γ with the effective action Γ.
9[Eds.] J. Schwinger, “On Gauge Invariance and Vacuum Polarization”, Phys. Rev. 82 (1951) 664–679; see the
discussion following equation (6.30).
10[Eds.] Gradshteyn & Ryzhik TISP. The relevant integral is number 2.733.1:

The integral’s limits are −∞ to ∞. Discarding the infinite constants, the result is 2πa, in agreement with (45.10). In
Woit’s notes, Coleman carries out the integral in (45.9) with contour integration; in the video of Lecture 50 he
describes this calculation.
11[Eds.] See the discussion following (2.13), p. 21.
12[Eds.] See note 14, p. 973.
13[Eds.] This is easy to show using the “gamma-5 trick”. See the solution to Problem 11.2, (S11.8), p. 428. Note
(20.102) that γ5 ≡ iγ0γ1γ2γ3 itself counts as an even product.
14[Eds.]Ryder QFT, Section 2.4, pp. 44–45; M. Kaku, Quantum Field Theory, Oxford U. P., 1993, p. 90; M.
Srednicki, Quantum Field Theory, Cambridge U. P., 2007, p. 9. Weinberg describes the shortcomings of this idea,
and quotes Schwinger: “The picture of an infinite sea of negative energy electrons is now best regarded as a
historical curiosity, and forgotten,” Weinberg QTF1, pp. 11–14. For a recent, more positive view of the sea and its
sound mathematical foundation, see J. Dimock “The Dirac Sea”, Lett. Math. Phys. 98 (2011) 157–166,
https://arxiv.org/pdf/1011.5865.pdf.
15[Eds.]
S. Weinberg, “Approximate Symmetries and Pseudo-Goldstone Bosons”, Phys. Rev. Lett. 29 (1972)
1698–1701; Coleman Aspects, “Secret Symmetry”, Section 3.8, pp. 142–144.
16[Eds.] See Problem 20.1 and its solution, pp. 871–873, in particular (S21.3); Ling-Fong Li, “Group Theory of the
Spontaneously Broken Gauge Symmetries”, Phys. Rev. D9 (1974) 1723–1738.
17[Eds.] See note 15, p. 807.
18[Eds.] Coleman adds: “As Jeffrey Goldstone asked Weinberg, ‘Who is Pseudo-Goldstone?’ He sounds like
someone who turns up in epigraphy, like Pseudo-Dionysus, who wrote that nice book on angels.” (Pseudo-
Dionysus the Areopagite, Christian Neoplatonist theologian, late fifth – early sixth century CE.) See also note 21,
p. 977.
19[Eds.] §41.1, pp. 890–892.
20[Eds.] Y. Nambu, “Axial Vector Current Conservation in Weak Interactions”, Phys. Rev. Lett. 4 (1960) 380–383.
21[Eds.] See notes 13 and 14, p. 944.
22[Eds.] See “Soft Pions”, pp. 36–56 in Coleman Aspects, and note 13, p. 896;.
23[Eds.] M. Gell-Mann and M. Lévy, “The Axial Vector Current in Beta Decay”, Nuovo Cim. 16 (1960) 705–726; J.
Schwinger, “A Theory of the Fundamental Interactions”, Ann. Phys. 2 (1957) 407–434; Peskin & Schroeder QFT,
pp. 347–363; Cheng & Li GT, pp. 149–151; B. W. Lee, “Chiral Dynamics”, in Cargèse Lectures in Physics, Volume
5, D. Bessis (editor), Gordon and Breach, 1972, pp. 119–178; printed in a separate volume as Chiral Dynamics,
Gordon and Breach, 1972. Ben Lee, a distinguished Korean-American physicist, was professor at SUNY Stony
Brook 1966–1973, and thereafter until his death both head of the theory group at Fermilab and a professor at the
University of Chicago. He was killed in June 1977 while driving from Chicago to Colorado with his family. A driver
in the opposite lane lost control of his truck as a result of a blown tire, crossed the divider, and smashed into Lee’s
car. He was 42. His wife and two children survived. Lee made many contributions to particle theory and taught the
community about the GSW model, non-Abelian gauge theories, the Faddeev-Popov method, and all that with his
review article with Ernest Abers (Abers & Lee GT).
24[Eds.] See Example 3 in Chapter 43, p. 944.
25[Eds.] See §18.3, particularly the paragraph following (18.49).
26[Eds.] See §5.6, in particular (5.70), p. 93.
27[Eds.] V. E. Markushin and M. P. Locher, “Structure of the Light Scalar Mesons from a Coupled Channel
Analysis of the s-wave ππ → KK Scattering”, in Workshop on Hadron Spectroscopy, T. Bressani et al., eds.,
Frascati, Italy, 1999; N. N. Achasov and G. N. Shestakov, “Phenomenological σ Models”, Phys. Rev. D49 (1994)
5779–5784.
28[Eds.] See the discussion following (41.5), pp. 892–892.
29[Eds] See the paragraph following (41.59), p. 906. Note also that only in the limit of massless quarks would chiral
invariance hold.
30[Eds.] See Theorem 3 in §25.4, p. 452, and “Renormalization and Symmetry: A Review for Non-Specialists” in
Coleman Aspects, p. 107.

Problems 24

24.1 Verify (44.51)

by carrying out the integration and making the appropriate approximations.
(Eds.)

24.2 In Chapter 24 we discussed the isospin-invariant Yukawa theory of mesons and nucleons,

(This Lagrangian is the sum of terms (24.27), (24.28), (24.29), and (25.77).) Here N is the nucleon isodoublet, Φ
is the pion isotriplet, m, µ, g, and λ are positive numbers, with µ < 2m, τ is the vector of Pauli matrices, and LCT is
the usual counterterm Lagrange density.

Now consider the same theory minimally coupled to electromagnetism (with massless photons). It is easy to
see that there is a contribution to F2(0) proportional to g2. Compute this contribution, for both the proton and the
neutron. It suffices to present the answers in terms of integrals over Feynman parameters.

A note on the definition of F2: In §34.2 we considered the scattering of an electron off a weak external current, Jµ;
L = −eJµA′µ, where e is the electron charge. The incoming electron has spinor u and four-momentum p, the
outgoing electron, u′ and p′ = p + k. Then

where

and (20.98) σµν = [γµ, γν]. (The coefficients have been chosen such that F1(0) = 1.) Use the same definitions for
the proton, with e the proton charge and m the nucleon mass. For the neutron, it clearly won’t do to use the
neutron charge, so let e be the proton charge here also. (Of course, for the neutron F1(0) = 0.)

Comment: Electromagnetism breaks isospin invariance, so you might be worried about the presence of isospin-
violating renormalization counterterms, like different wave-function renormalization constants for the neutron and
proton. This is not a problem here. The asymmetric counterterms are all at least of order e2 (times powers of g),
and finding F2(0) to order g2 only involves computing the nucleon–nucleon-photon vertex to order eg2.
(1998 253b Final, Problem 2)

24.3 Example 2 of Chapter 43 (p. 941) was a theory with spontaneous breakdown of U(1) internal symmetry. The
particle spectrum of the theory consisted of a massless Goldstone boson and a massive neutral scalar.
Furthermore, although I did not discuss it in class, this term in the Lagrangian

gives rise to the decay of the massive meson into two Goldstone bosons, with an invariant Feynman amplitude
proportional to a−1. (This is not a misprint: before reading the decay amplitude from the Lagrangian, we must first
rescale θ to put the free Lagrangian in standard form.) Now consider the theory minimally coupled instead to a
massive photon with mass µ0 (before symmetry breaking). What is the photon mass after the symmetry breaks?
Does the Goldstone boson survive? If it does, what is its mass? What about the decay amplitude discussed
above?

Comment: The Abelian Higgs model is the same theory minimally coupled to a massless photon. As we will see in
Chapter 46, the Goldstone boson disappears, and the photon acquires a mass, which we will compute.
(1998b 11.1)

Solutions 24
24.1 We need to show

where α = U″(ϕ) − iϵ, and the Euclidean subscript E has been suppressed. Using the general rule

the integral becomes

This integral can be found in Gradshteyn & Ryzhik TISP (2.729.2), but it’s not hard to do. By parts

The last integral is elementary:

We drop the constant last term, and find

which agrees with Gradshteyn & Ryzhik TISP 2.729.2. Then

The first term can be approximated:

Then

and the stuff in the curly brackets becomes

which was to be shown.

24.2 The interaction Lagrangian is

with

Let’s do p and n in order; each will closely resemble the class computation for the electron (§34.3), from which we
can steal some formulae.

For the proton:

Figure S24.1 O(eg2) contributions to the proton form factor Fµ

 From graph A:

where

The denominator is simplified as in class (34.57). Rewriting in terms of the photon momentum k = p′ − p, dropping
terms of O(k2), and shifting the momentum using q = q′ − px − p′y, we obtain for the denominator

If we substitute this expression for q into N µ, the terms linear in q′ are odd, and vanish upon integration. Those
terms in the numerator quadratic in q′ contribute only to F1. Thus for the determination of F2, we can take

so that

where Δ = {x ≥ 0, y ≥ 0, x + y ≤ 1} (see Figure 34.7). As in (34.71), we write p on the left, and p′ on the right, in
terms of k, and use , u = mu:

Keeping only terms linear in k, and symmetrizing (x, y → (x + y), due to the region of integration), we obtain

Finally, using (I.2) in the box on p. 330, (note the sign of a2)

and

we find for graph A,

From graph B:

and now (recalling (20.103) γ52 = 1)

The same shift as before eliminates D’s first-order term in q′, but the zeroth order term is different:

Treating the numerator as before, (x, y → (x + y))

Using the Gordon decomposition (34.28),

N µ becomes, throwing away terms linear in γµ which contribute only to F1,

Carrying out the q′ integration, and extracting the expression for F2(0) as before, we have from graph B

For the neutron, things are almost the same:

Figure S24.2 O(eg2) contributions to the neutron form factor Fµ

The contribution of graph A to the neutron is twice that of the contribution to the proton, due to the in the
coupling; and graph B’s contribution to the neutron is (–1) times that of the proton’s graph B, because the
interaction term has the opposite sign for e. We can summarize the results as follows:

Note: It’s easy to show that

This can be used to reduce the double integrals to single integrals. ■

24.3 The Lagrangian of Example 2 is given in its first form by

Perhaps the easiest way to couple this theory minimally to a vector field is to rewrite it as

where

and

Then the minimal coupling procedure gives, separating real and complex parts,

To study the spontaneous symmetry breaking, we parametrize the fields as in (43.27):

Adding the vector’s kinetic and mass terms, we obtain

(At first, it may seem strange that the ρ field is not coupled to the vector as the θ field is. Consider a U(1) gauge
theory for Φ, or equivalently, a SO(2) symmetry for the real fields {ϕ 1, ϕ 2}. If the symmetry is a rotation, i.e., θ → θ
+ α, then ρ does not change, and so there’s no need for it to have any gauge compensation through Aµ. Though
we are considering here a massive vector theory which is not gauge invariant, the argument still applies.)

The symmetry breaks, and we replace

The Lagrangian becomes

where

The ρ′ has a mass term

but to obtain the rest of this theory’s particle spectrum, we need to disentangle the terms quadratic in Aµ and in θ′.
Define

Writing L ′ in terms of Bµ, the tensor Fµν is unchanged:

A judicious choice of λ will eliminate the Bµ∂µθ cross-terms,

namely

The Lagrangian becomes

so that

The Goldstone boson θ′ does survive. Its kinetic term is a little strange, but we can redefine the field:

Then

where

(the dots indicate terms cubic or quadratic in ρ′). The term that governs the decay of a ρ′ into two Goldstone
bosons ϑ is

The amplitude for the decay ρ′ → 2ϑ is proportional to

As µ0 → 0, the decay amplitude goes to zero; as µ0 → ∞, the decay amplitude becomes proportional to a−1, as
claimed, and there is no dependence on the gauge coupling constant e.

46
The Higgs mechanism and non-Abelian gauge fields
Nambu and Jona-Lasinio’s investigations into spontaneous symmetry breaking were motivated by a desire to
understand the nucleon’s mass.1 While the value of the nucleon mass could be obtained by this mechanism, there
was an unfortunate byproduct: massless hadrons, which are not realized in nature. We’ve learned quite a lot
about Goldstone bosons. They seem to be inevitably associated with spontaneous breakdown of a continuous
symmetry: if we have spontaneously broken continuous symmetries we’ll have Goldstone bosons. We have a
very powerful and general theorem of this result due to Goldstone (§43.4), the proof of which does not depend on
perturbation theory. That would seem to suggest that there is no escape; no spontaneous symmetry breaking, at
least of continuous symmetries, without Goldstone bosons. But what we’ve learned about them so far makes it
seem that their only physical application, at least in high-energy physics, is to chiral dynamics. (They have
applications elsewhere in physics; things like spin waves in the ferromagnet2 and superconductivity 3 are closely
related to Goldstone bosons.) If a Lagrangian has a small symmetry-breaking term, there will be light mass
spinless particles. The only light mass spinless particles in the Particle Data Group tables are the pions, and only
they are possible candidates for Goldstone bosons or near-Goldstone bosons; there are no massless mesons.
This is sad, because the idea that the universe is more symmetric than it seems to be is very attractive, as is the
corollary that the apparent asymmetries in the universe are all the consequence of spontaneous symmetry
breaking. It would be unfortunate if there were no loophole, no way to get a spontaneous breakdown of a
continuous symmetry without a Goldstone boson.

However, there is a loophole. We haven’t thought about the possibilities of gauge invariant field theories. So
far we have discussed only one: quantum electrodynamics. The proof of the Goldstone theorem had two
assumptions: one was relativity, that the theory was Lorentz invariant; the other was positivity of the norm, that
poles in Green’s functions corresponded to physically observable states. These assumptions are false for at least
quantum electrodynamics, and maybe for other, similar theories. In QED, if we go to a covariant gauge like the
Feynman gauge or the Landau gauge, we have poles in Green’s functions corresponding to longitudinally
polarized photons which are definitely not physical states; they are gauge phantoms. We can instead go to a
gauge in which these gauge phantoms do not appear, for example Coulomb gauge, but the Coulomb gauge is not
Lorentz invariant. So the assumptions underlying Goldstone’s theorem do not hold for at least one gauge invariant
theory, QED. Here could be a way out.4

46.1The Abelian Higgs model

There is at least a possibility that Goldstone bosons in a theory with gauge invariance might turn out to be gauge
phantoms, objects that disappear if we choose a gauge in which only physically observable states occur. One
simple computation is worth two hours of argument. Let’s build a simple model that has both the Goldstone
phenomenon and gauge invariance, and see what happens in tree approximation (or equivalently, by doing a
classical analysis). Since the only symmetry we know how to turn into a gauge symmetry is the one-parameter Lie
group U(1) of QED, we’ll consider a theory invariant under U(1) (or equivalently SO(2)), the model 5 we discussed
in §43.2:

The infinitesimal transformations are

i.e., a rotation in the ϕ 1-ϕ 2 plane. The easiest way to see what is going on is to assemble ϕ 1 and ϕ 2 into a single
complex field with angular coordinates ρ and θ:

In this form the infinitesimal transformations become

If we define ρ′ by

the Lagrangian becomes

Whether or not this theory displays spontaneous symmetry breakdown, it has a U(1) invariance, and
therefore we can turn it into a gauge theory: we can couple electromagnetism to the scalars if we identify the
conserved current with the electromagnetic current. The minimal coupling prescription is:

where

This definition of the covariant derivative is equivalent, to within a relative sign, to the earlier (27.48). We normally
apply this formula only when DΦ is a linear homogeneous function of Φ, but as it stands it’s invariant under
redefinitions of the fields, so long as we appropriately change DΦ. We could first obtain the covariant derivatives in
the ϕ 1, ϕ 2 language and then transform to ρ and θ. To short-circuit a little algebra, we apply the prescription
directly to the transformed fields ρ and θ (it doesn’t matter in which order we do it). In the case at hand

The field ρ′ doesn’t change at all because it’s equal to ρ plus a constant, and ρ doesn’t change. The only field that
involves electromagnetic coupling is θ: Dθ = 1. We have from the minimal coupling prescription

The λ term has no derivative and is not altered by the minimal coupling prescription. The Lagrangian becomes

This is just ordinary quantum electrodynamics of charged scalar particles with a quartic self-interaction, written in
angular coordinates.

It’s not obvious how to obtain the particle content of the theory from this Lagrangian. Usually one can read off
the spectrum of small oscillations from the classical theory. But that’s difficult here because there’s a cross term
between Aµ and ∂µθ coming from the third term in . This difficulty can be circumvented by defining a new field,
Bµ:

This definition looks like—and is—a gauge transformation of Aµ. Since Fµν is invariant under a gauge
transformation,

(the two terms in ∂µ∂νθ cancel). We now have

The algebra is impeccable but the result is surprising. We now have in tree approximation, from the quadratic
terms in the Lagrangian, a massive scalar meson ρ′, which has the same squared mass (43.26) we computed
before for the non-Goldstone boson, ϕ 1,

and a vector Bµ, with the standard Proca Lagrangian for a vector boson of squared mass

The field θ has disappeared; there is no massless meson! The squared masses for the other mesons are positive
numbers, thank God. This looks rather surprising but at least the degrees of freedom have been conserved. Let’s
count them.

When the symmetry does not break spontaneously, when the a2 term and hence the bare mass are positive,
we obtain two scalar mesons ϕ 1 and ϕ 2, each with one degree of freedom, and one massless vector Aµ, with two
degrees of freedom (two polarization states); four in all. With spontaneous symmetry breaking we have a massive
scalar with one degree of freedom, and we have a massive vector which has three degrees of freedom
corresponding to the three polarization states: again four degrees of freedom. That’s the difference between a
massless and a massive vector boson. The photon only has the two transverse excitations but no longitudinal
excitation. No mysterious particles have appeared or disappeared. Although the whole thing happens in one
“swell foop”, one way of thinking about what is going on is to say that the Goldstone boson appears and it is
promptly eaten by the gauge field, which becomes heavy; the two degrees of freedom of the photon and the one
degree of freedom of the Goldstone boson combine together to give the three degrees of freedom of a massive
vector boson.

This fact was discovered independently by many people around the same time, but the one who understood
and explained it most clearly was Peter Higgs, and therefore it is usually called the Higgs mechanism (or Higgs
phenomenon). If I were really fair I would call it the Anderson-Brout-Englert-Guralnik-Hagen-Higgs-Kibble
mechanism, but I’m not going to do that.6 The Feynman rules for the Abelian Higgs model are summarized in the
box on p. 1015.

We can get a deeper insight into what is going on by remembering the reason for introducing the minimal
coupling prescription in the first place: gauge invariance. The rotation in the ϕ 1-ϕ 2 plane is a gauge
transformation:

Only θ, the phase of the field, transforms: we change only that, in an arbitrary way at every spacetime point. This
is an absolute invariance of the classical theory that has no effect on any physically observable quantities. The
photon field Aµ also transforms, but we’re not going to worry about that for the moment. There is nothing to stop us
from choosing our gauge, given any field configuration, so that, following (46.17), we can make the field θ(x)
disappear completely:

We can always gauge transform the phase of the field at any given point in spacetime so that the field is real and
positive.7 Once we have chosen such a gauge we see that the reason the Goldstone bosons don’t appear in the
final theory is that the degrees of freedom which they represent, to wit θ oscillations, can simply be gauged away,
as in (46.18). The Goldstone bosons are θ oscillations and we can choose the gauge so there ain’t no θ. In this
way we can get the second form of the Lagrangian (46.14), the one involving the B field, from the first form, simply
by applying the rather unconventional but perfectly legitimate gauge condition θ = 0.

Feynman rules for the Abelian Higgs model

1.For every …

Write …

(a)internal vector line

(b)internal scalar line

(c)Three point scalar vertex

(d)Four point scalar vertex

(e)Scalar-bivector vertex

(f)Seagull vertex

2.Ensure momentum conservation at each vertex: (2π)4δ(4)(∑pout − ∑pin)

3.Multiply by and integrate over all internal momenta q.

4.Polarization factors:
 For every vector boson, a factor with ε ⋅ k = 0, ε′ ⋅ k′ = 0.

We normally like to apply the gauge conditions on the divergence of Aµ or something like that, but that’s just
prejudice and habit. We could just as well fix our gauge by declaring θ = 0, whereupon the ∂µθ term in (46.11)
drops out and the original form of the Lagrangian becomes the form (46.14) with A replaced by B. The reason the
Goldstone bosons do not appear when spontaneous symmetry breaking occurs is that they were never there in
the first place. The degrees of freedom of the system which they represent are pure gauge degrees of freedom
and can always be gauged away. That’s why the world is not full of massless Goldstone bosons and massless
vector bosons; the photon is the only one.

This gives us a hint on how to generalize this phenomenon to more complicated theories involving non-
Abelian, noncommutative groups of symmetries, suffering spontaneous symmetry breakdown. All we have to do is
figure out how to promote them to gauge symmetries, just as we promoted the U(1) symmetry here from a global
symmetry with constant χ to a local symmetry with spacetime dependent χ(x). Remember from our earlier general
analysis the Goldstone bosons always corresponded to degrees of freedom associated with applying infinitesimal
symmetry transformations to the vacuum state. They just move us around the minimum of the Mexican hat
potential.8 If we have gauge invariance for those infinitesimal symmetry degrees of freedom, we can gauge away
the Goldstone bosons and prevent them from appearing at the end.

46.2Non-Abelian gauge field theories

We are going to generalize the gauge invariance of electrodynamics from a single gauge field, the photon, with a
simple U(1) or SO(2) group to a set of gauge fields and a general non-Abelian group. This sort of theory was first
written down by Yang and Mills in 1954, and is consequently called a Yang–Mills field theory.9 This problem has
nothing to do with spontaneous symmetry breaking. We will in due course employ these non-Abelian gauge fields
in spontaneous symmetry breaking, but first we’ll need to discover how to make a continuous internal symmetry
into a gauge symmetry.

Let’s begin with the general situation. We have a Lagrangian

depending on a set of fields Φ and their derivatives. For the moment we will assume the Φ’s are real scalar fields;
that’s just for notational simplicity. It’s trivial to generalize to the case where there are spinors or complex fields.
We have a group of infinitesimal transformations

The δωa are infinitesimal parameters independent of position and time. The Ta’s are Hermitian matrices,10 one for
each a:

These could, for example, be the matrices that generate SU(2) or SU(3) or SU(2) ⊗ SU(3) or whatever group you
have in mind. This transformation is supposed to be an invariance of :

Because the Ta generate a symmetry group, they are closed under commutation

Later on we will have to exploit certain symmetry properties of the structure constants cabc and therefore it’s
useful, before we even start talking about Lagrangian field theory, to make a few remarks about a nice way to
normalize these things.

I define a positive-definite quadratic form on the a-b space through the trace of a product of two generators,
normalized according to11

I will adopt different constants c for different groups. That is I will form linear combinations of the Ta’s so they’re
orthogonal in terms of this so-called trace norm. Then cabc is revealed to be

The orthogonality property of the trace extracts out the coefficient of Tc . This is very useful because it tells us that
the cabc’s have essentially the same symmetry properties as the ϵabc’s do in the particular case of SU(2), being
even under cyclic permutations and odd under anti-cyclic permutations:

That’s very convenient because it means we don’t have to really worry about whether we get abc or bca in an
expression; we may need to make a change of sign, but that’s all.

I now turn to making the Lagrangian (46.19) gauge invariant. That is, I wish to alter the theory in such a way
that the infinitesimal transformation of the fields Φ involves local—spacetime-dependent—parameters δωa(x):

This is to be an invariance of the Lagrangian. The transformation (46.27) is spacetime-dependent so that we can
impose gauge invariance to kill the Goldstone bosons. The problem comes up in δ(∂µΦ), which consists of a term
that’s jolly plus a term that’s just disgusting:

We want to arrange matters so that

The Lagrangian as it stands is not invariant under (46.27) because of the disgusting term. We’re going to have to
change just as we did in the Abelian case by introducing new fields, one photon-like field for each δωa. We’re
going to change the physics so that this extra term cancels out.

Use electrodynamics (27.48) as a model. From the minimal coupling prescription of electrodynamics there’s a
strong suggestion that we introduce a vector boson gauge field, a photon-like field, for each a, to sop things up.
Let’s investigate that possibility, and define a new covariant derivative appropriate to a non-Abelian gauge group:

where Aµa are some vector fields. I will concoct infinitesimal transformation laws for the fields Aµa so that

i.e., so that (D µΦ) transforms in the same way that Φ does. That’s our goal. It’s not obvious that we can arrange
that. If we can, then (Φ, D µΦ) will be gauge invariant because this Lagrangian will transform under local,
spacetime-dependent transformations in the same way as it did under global, spacetime-independent
transformations:

We’ll still have to worry about the free Lagrangian for the Aµa field, but we’ll get to that later. At least we’ll have
taken care of the minimal coupling terms. I have not put a free parameter or coupling constant in (46.30); any such
are included in the Aµa. Later I will make these explicit, but at this stage I’ll just absorb them into the Aµa to keep
from cluttering the equations. I don’t yet have a free Lagrangian for the Aµa fields, so I have no natural scale for
them.

It’s elementary (though tedious) to see what needs to be done to make (46.31) true. First we’ve got the term
we already know, (46.28). Then we’ll have the term coming from the change in Aµa; that’s what we want to
compute. This is supposed to satisfy our criterion, (46.31):

Expanding the various terms, we have

It only takes a little algebra to see what this implies about δAµa. The first term on the left-hand side of (46.34)
cancels the first term on the right-hand side. We move the remaining term on the left-hand side of (46.34) to the
right and bring the δAµa term to the left-hand side. Finally, we swap the dummy indices a and b in the last term in
(46.34), and obtain

The last two terms are just the commutator [Ta, Tb] times δωaAµbΦ. The right-hand side becomes

The last term can be rewritten. We have, swapping the dummy indices,

the last step following from the invariance of the structure constants under cyclic permutation (46.26). Thus all
sums involve Ta on an arbitrary Φ:

We can obtain the desired result if we choose

This is the key equation.12

The transformation law (46.39) for the gauge fields Aµa is rather complicated. Before going on to finding the
possible forms of the pure gauge field Lagrangian, it might be nice to understand their physical meaning, by
looking at these transformation laws in a particular special case, say SU(2), isospin, in which {a, b, c} run over 1, 2
and 3, and the structure constants cabc are ϵabc. Choose a gauge transformation associated with rotations about
the third axis in isospin space:

so the only thing left is δω3. What does (46.39) give for δAµ1 and δAµ2? For a = 1,

The derivative term is irrelevant because that’s non-zero only for a = 3. In ϵ13c , the only non-zero term has c = 2,
and ϵ132 = −1. By the same reasoning for a = 2 and a = 3,

In the last line the ϵ carries two 3’s and is therefore zero. These transformations look a great deal simpler. These
three gauge bosons in a sense transform like an isotriplet, like the generators of the group. If I restrict myself to
gauge transformations corresponding to rotations about the three axis, Aµ1 and Aµ2 transform like an ordinary pair
of charged particles, with I3 = ±1; they just rotate. As far as gauge transformations along the I3 axis goes, you can’t
tell that Aµ1 and Aµ2 are gauge particles, and not simply some particles that transform linearly and
homogeneously along the group. On the other hand, you can’t tell that Aµ3 carries any internal quantum numbers.
It acts just like an uncharged photon, transforming exactly as a photon would, with a gradient.

In the general case, the gauge boson corresponding to an infinitesimal transformation associated with only
one symmetry generator transforms like a photon: it gets only an added gradient. The other gauge bosons
transform like ordinary non-gauge fields; they just rotate among themselves according to the transformation of the
particular symmetry group we’re looking at (because of the cabc term, they transform like the group’s generators),
without any gradients. That’s why you need both terms. To put it another way, the first term is here so that if δωa
is constant, (a possibility, after all) the second term would drop out, and the gauge bosons would mix among
themselves just as the group generators do. That’s obviously necessary if the expression AµaTa is not to break the
symmetry. The second term is there so that if I consider gauge transformations along a particular direction in
internal symmetry space, the gauge boson associated with that direction transforms like a photon. That’s why we
have this funny transformation law with two terms in it.

The next stage is to find the analog of (Fµν)2. That is what gave us a perfectly satisfactory Lagrangian in the
case of electromagnetism. If we just had (Φ, D µΦ) we’d have a pretty dumb theory. We’d have the Aµ’s in the
theory but not their derivatives and we wouldn’t expect much dynamics to come out of that; you’d have introduced
new vector fields with no associated free Lagrangian. I’ll construct the analog to Fµν by the following chain of
reasoning. Consider the commutator of two covariant derivatives

I will demonstrate that one can find a function of the gauge fields and their derivatives, Faµν, such that

Both sides of this equation must transform under gauge transformations the same way as Φ does, because that’s
the property of covariant differentiation. The Faµν’s will have to transform in such a way that this equation doesn’t
break the invariance: δFaµν must transform homogeneously and indeed like the group generators, like the Aµa
transforms except without the derivative term:

An extra term in it would screw things up. The tensors Faµν are nice objects that transform linearly and
homogeneously, like the group generators. They will be the non-Abelian analogs of Fµν. By playing around with
them we will be able to find invariant quadratic Lagrangians. The transformation of the tensor (46.45) follows
trivially from the transformation of Aµa (46.39); it’s just the statement that the tensor should be covariant.

Before I demonstrate (46.44), again a tedious but straightforward computation, let me write the finite versions
of these expressions. The transformation for finite ωa is

The various quantities transform according to:

The thing we want to compute is the commutator of two covariant derivatives:

Most of the terms will disappear when we antisymmetrize. For example the first term, ∂µ∂νΦ, vanishes because
ordinary derivatives commute. There are two terms involving one derivative of Φ:

They differ only by the exchange of µ and ν; whether I sum over a or sum on b is a matter of notation. So the two
∂Φ terms together are symmetric in µ and ν and so will cancel when we antisymmetrize. There are two terms

Exchanging µ and ν is just commuting the order of those two matrices since I can relabel the summation indices.
So this equals

by using the algebra of the generators and the symmetry of the structure constants (46.26). Thus I have a function
of the gauge fields and their derivatives times Ta acting on Φ; referring to (46.44),
This field tensor is a sum of three terms. The first two, ∂µAνa − ∂νAµa, will look familiar to anyone who has studied
electromagnetism. The third nonlinear term, absent from electrodynamics, is the glory, the joy, and the nightmare
of non-Abelian gauge field theory, cabcAµbAνc .

Having obtained the field strength tensors Faµν we have to find ways to make gauge invariant objects out of
them. We can square them and then sum on a; that’s guaranteed to be invariant. Therefore we can write,
following electrodynamics

The reason for the constant g will become clear shortly; we don’t know a priori what that coefficient should be.
This structure is obviously invariant: it’s the sum of squares of all these objects that transform according to some
representation of the group, plus the original Lagrangian. That’s certainly the simplest possible generalization of
electrodynamics. The free constant is there because I have no way of knowing the relative scale of these two
terms.

We’ll now get rid of that constant; otherwise it leads to a quadratic term that is rather dumb, with a constant
1/g2 in it instead of the 1 that we want for the free theory of vector bosons. We will do it by rescaling: define

Then Faµν becomes

The g2 disappears from the term quadratic in the derivatives of the A fields and the Lagrangian becomes

The scale factor g is now sensibly located to act as a coupling constant. The quadratic terms in the Lagrangian
are totally free of g’s and the non-quadratic terms all have g’s in them. By convention we will drop the primes from
now on and always use these fields we’ve defined as primed fields. In particular, the covariant derivative and the
Yang–Mills field tensor become

This looks like a reasonable theory of a bunch of fields, if we can handle the problem of quantization of the
gauge fields Aµa, which after all caused a lot of trouble for electrodynamics and may cause us troubles here.
Some of the fields—the Aµa—are massless; the others, the Φ, are massive or massless depending on what is
like, the fields all interacting together in a complicated way governed by this coupling constant g. Even if there
were no Φ fields, the non-Abelian gauge fields would have inherent self-interactions. This is a striking contrast
between non-Abelian gauge field theory and Abelian gauge field theory (electromagnetism). In the absence of
charged particles pure electromagnetism is a free field theory; nothing could be more trivial. The pure gauge
theory of Yang–Mills fields13 has complicated interactions even if there are no electrons or π mesons or anything
else in the world, except the gauge bosons.

There’s a reason for this. Remember the old argument, which you were no doubt all told by your elders, about
why gravity is necessarily a nonlinear field theory. The argument goes as follows. The source of gravity is energy.
The graviton carries energy. Therefore the graviton must be a source of gravity. Therefore there would be a
nonlinear coupling even if gravity were all that the universe contained. The same thing is true here. For simplicity
let us think of the isospin example. We have three gauge fields, Aµ1, Aµ2 and Aµ3; Aµ3 is the gauge field for the T3
rotations. It couples to everything that carries the third component of isospin. Among things that carry the third
component of isospin are Aµ1 and Aµ2. We have just seen that they are rotated under an T3 rotation. Therefore
Aµ3 must couple to Aµ1 and Aµ2 even if there were no other particles in the world. The amazing thing is not that a
nonlinear coupling is necessary but that we can get by with such a simple nonlinear coupling that has only cubic
and quartic terms. In gravity we have to go on forever. Here we can stop after the fourth order. That’s amazing but
that’s the way things work. I have done this in a way that makes the Yang–Mills theory look as close to gravity as I
can without lying to you. If you’ve taken Steve Weinberg’s course in general relativity, or another’s, you’ll
recognize that (46.44) is very close to the definition of the Riemann–Christoffel tensor as the commutator of
covariant derivatives;14 and that writing the Lagrangian in the form (46.55) with the coupling constant in front of the
free Lagrangian (rather than inside the interaction) is precisely the way it’s written in gravity. There you have
1/(8πG) times R, the Ricci scalar, plus a matter-energy term that’s free of the gravitational coupling constant.15

One coupling constant, or many?

Now we have to address a tiny technical point: do we always have only one gauge coupling constant or can
there be several? This is really the question of whether the only invariant we can form is FaµνFµνa. Maybe we could
form different ones. For example, if our group was SU(2) then that would be the only invariant we could form; the
only way we could take two isovectors and put them together as a scalar is a dot product. On the other hand if our
group is chiral SU(2) ⊗ SU(2) the generators fall into two sets, the generators of the left-handed isospin and the
generators of the right-handed isospin. Now we can form two scalars: left dot left and right dot right and they can
have independent coefficients. This is obviously the general situation if you make appropriate definitions, which
we will now do.

If the generators Ta of the gauge group G transform according to an irreducible representation of G, then there
is only one invariant that can be constructed, FaµνFµνa, and therefore there is only one gauge coupling constant g.
In this case, for those of you who have been reading group theory books, we say that G is simple.16 On the other
hand, if the Ta’s transform according to a reducible representation of the group, as in for example chiral SU(2) ⊗
SU(2), then it’s easy to see from the antisymmetry of the structure constants that the algebra falls into a sum of a
bunch of subalgebras, none of which talk to each other; cabc vanishes unless a, b and c are associated with the
same factor (i.e., the same subgroup). G is a product, at least locally, of simple groups generated by the various
irreducible representations. And then we can have one gauge coupling constant for each factor, one for the right-
handed SU(2) and a completely different one for the left-handed SU(2). For example, if we were to consider the
product SU(2) ⊗ U(1), and call the quantities coupled to the associated gauge fields “isospin” and “hypercharge”,
there would be a hypercharge gauge boson and there would be an isospin triplet of isospin gauge bosons and
they could have different gauge coupling constants. The hypercharge is a SU(2) invariant and so is the square of
isospin. That’s in fact what happens in the Glashow–Salam–Weinberg model of weak interactions.17 Our formulas
for non-simple groups would be somewhat generalized. If we use the prime fields for example, D µΦ in (46.56)
would be ∂µΦ plus a g that may depend on a, still summing on repeated indices even though the index appears
three times:

Each ga would be a constant, for a given factor in a product of groups, but they might be different for different
group factors (in the example above, one constant for U(1) and another for SU(2)). Likewise

Now we have the complete theory, though we don’t yet know if we can quantize it. And if we can quantize it,
we don’t know if it’s renormalizable. But at least at a classical level we have the complete theory of non-Abelian
gauge fields. We can certainly write more complicated Lagrangians just as we could in electrodynamics. We could
have analogs of the Pauli coupling Fµνσµν; we could have terms involving the fourth power of Fµν but that would
add in more derivatives. Since we eventually hope to get a renormalizable field theory we should try to get away
with as few derivatives and as few powers of the fields as possible. This is the minimum number.

When Yang and Mills first proposed this theory for the case of isospin in the mid-fifties, nobody knew anything
about spontaneous symmetry breaking. It was thought that maybe these things did exist; perhaps there were
three massless vector bosons in addition to the photon. But they had to be very weakly coupled; otherwise they
would have been observed. People went through a long series of arguments involving cosmological experiments,
protons, neutrons, virtual pions, long-range forces that would compete with gravity and got all sorts of bounds. It
gradually became clear that this theory was hopeless; there was no chance that these massless vector gauge
bosons existed in the real world. The only massless gauge boson in the real world was the photon. Around 1960
and 1961, everyone was very excited about spontaneous symmetry breaking: Nambu, Jona-Lasinio, Goldstone
and the Goldstone phenomenon.18 But it was soon discovered that Goldstone bosons were inevitable, and
Goldstone’s theorem was proved.19 Then everyone said, well, that’s the end of that, because there aren’t any
massless scalar mesons in the world except maybe the pions, with approximately zero mass. We had two things
floating around that were more or less theorist’s toys: Yang–Mills theories of zero-mass vectors, and spontaneous
symmetry breaking with zero-mass Goldstone scalars.

The Higgs phenomenon reconciles these two problems; the two diseases turn out to be each other’s cure. 20
If the group that is spontaneously broken is a gauge group then the Goldstone bosons are eaten by the gauge
bosons, which become massive vector bosons. You wind up with neither unobserved massless gauge bosons nor
unobserved massless Goldstone bosons. Let’s see how this works out in the general case of spontaneous
symmetry breaking, again only in the tree approximation. In this case we can’t go farther because we can’t
quantize the theory yet. We will quantize it as soon as we are done with this discussion. It turns out to be just what
we did for electromagnetism over again. (We went through electromagnetism that way to serve as a warm-up for
the non-Abelian case.)

46.3Yang–Mills fields and spontaneous symmetry breaking

We want to consider the same kind of Lagrangian we had before, now promoted to a gauge theory.

The field Φ is an N-component vector; N is the total number of generators of the group. We have the covariant
derivative of Φ:

(remembering that we sum over repeated indices even when there are three of them). Just as before the state of
lowest energy is found by taking Φ to be a constant and setting the minimum of U(Φ) to 0. Again we call that state
the vacuum:

The whole group is now promoted to a gauge group.

We divide our group generators into two classes: those that annihilate the vacuum

where n < N, and the remaining orthogonal set of linearly independent generators that don’t:

Equivalently,

The generators (46.66) that annihilate the vacuum are the unbroken generators, while the others (46.68) are the
spontaneously broken generators.21 The unbroken generators define a subgroup H of G,

The group H remains a manifest symmetry.

We now have a gauge theory and therefore we must pick a gauge. I will impose a set of gauge conditions that
will eliminate the Goldstone bosons immediately. The Goldstone bosons correspond to oscillations determined by
Tb áΦñ; modes generated from áΦñ by infinitesimal transformations about the minimum:

They are what the spontaneously broken generators generate (in two senses of the word “generate”). By applying
a gauge transformation that counteracts just that group transformation (46.70) I can arrange that

I will choose as the gauge condition22

where Φ′ is the gauge-transformed Φ (46.47), áΦñ is a constant N vector in Φ space which minimizes U(Φ), Tb
are N × N matrices for the broken generators, and (46.72) is their dot product. This condition (which can be
imposed on any compact group) was first published in a paper by Weinberg.23 It is a set of linear conditions on the
fields, equivalent to the earlier condition (46.18) θ = 0, which immediately eliminates the Goldstone bosons.
Unphysical motions away from the ground state of the theory (corresponding to oscillations along the bottom of the
trough of the potential in Figure 43.5) are canceled by this equation. The relation (46.72) expresses only (N − n)
gauge conditions; there are still n conditions left. For the remaining n gauge conditions we choose whatever gauge
pleases us: axial gauge, Coulomb gauge, covariant gauge.

Weinberg’s “proof” tells us we can always fix the gauge according to this condition (46.72). But the gauge is
not uniquely determined when Φ has a zero, and this general argument breaks down. If you imagine classical
field configurations very far from the configuration of minimum energy, then

That doesn’t bother us when we’re doing perturbation theory in tree approximation, because then we’re always
expanding about the configuration of minimum energy; the deviations from that configuration are small. It is
evident however from the way we have found the Goldstone modes that it is always possible to choose a gauge in
which the Goldstone bosons are gauged away. They are precisely those modes (spanning a hypersurface in Φ
space) that are swept out by the action of the group generators on the ground state. And since they are made by
group generators, they can be unmade by group generators. That’s just what a gauge transformation does for
you.

This particular choice of gauge is called the U gauge;24 “U” stands for “unitary” and expresses the fact that
we have eliminated the unphysical degrees of freedom, the Goldstone bosons, which we know won’t be there
because they will be eaten by the gauge bosons. It has the additional property that certain cross terms in D µΦ •
D µΦ disappear after the shift is made. (We will see this when we discuss the Glashow–Salam–Weinberg model.) It
is in general a terrible gauge for anything except exploring the tree approximation. You can quantize in it, but you
get awful Feynman rules that lead to ostensibly non-renormalizable theories. When you sum up all the graphs, the
horrible divergences cancel. But only a madman would work in a formalism where every individual piece of the
computation is hideously divergent, and it’s only when you sum them all up that you get convergence.25 (There
are people who have worked in this gauge.) Let’s look at a couple of SO(n) examples in detail to see how this
gauge condition is applied.26

EXAMPLE. SO(2): Rotations in a plane

Write the fields ϕ 1−ϕ 2 in our SO(2) example, p. 941, as

Earlier we chose (43.23)

There is one generator, T, associated with rotations, and one Goldstone boson. An element of SO(2) has the form

and from it we obtain the generator T,

In agreement with (46.67), T does not annihilate the vacuum:

and the gauge condition (46.72) is

Because of the definition (46.3), the gauge condition ϕ′2 = 0 is the same as the θ = 0 condition (46.18) we applied
before.

EXAMPLE. SO(3): Rotations in three-space

There are three generators associated with the rotations in three-space (about each axis in turn):

Taking the vacuum to be

we see that

and hence there are two broken generators and two Goldstone bosons. There is one unbroken generator, so the
remaining symmetry is SO(2), rotations in the plane about the 1-axis. (The vacuum was chosen to have its non-
zero expectation value along the 1-axis rather than the 3-axis to simplify notation in what is to come.)

With the gauge condition (46.72) we don’t have to worry about the Goldstone bosons. We do have to worry
about the gauge bosons. We’d expect that some of them, those associated with the spontaneously broken
generators, are going to get a mass, and others, corresponding to the subgroup H of the unbroken generators, will
remain massless; for them the theory is much like ordinary electrodynamics. In tree approximation we only have to
worry about the quadratic terms that have an effect of making the shift

The only terms that do that in the Lagrangian are

The other terms are totally irrelevant; they’re not going to involve the vector bosons and won’t be affected by
making a shift of the scalar fields. After we make this shift, the only terms that will be both quadratic in the vector
bosons, and decoupled from the scalar bosons, are those involving the áΦñ part of Φ, an N-vector for N boson
fields. This equals, keeping only terms quadratic in the Aµa’s,

M2 is a vector boson mass matrix. The tree approximation is

(there is no summation on a and b). That’s an obviously symmetric matrix. It’s positive definite within the space of
spontaneously broken generators (for a, b > n) because of (46.68), which asserts that if I put any vector ca on the
left and any vector cb on the right I get a non-zero matrix element. So it is a positive definite symmetric matrix. Its
eigenvalues, which you have to compute for any particular given Φ, given the generators, determine what the
masses of the vector bosons are. If we choose a and b to correspond to one of the generators of H, then this
matrix is zero, so indeed the vector bosons corresponding to the unbroken symmetries remain massless;

Writing M2ab as a block diagonal matrix we get

The 0 blocks are from the unbroken generators, (46.66). For our SO(3) example above, n = 1, N − n = 2, and there
is only one unbroken generator, T1. The mass matrix is
The gauge bosons associated with the broken symmetry generators, T2 and T3, become massive by eating the
corresponding Goldstone bosons. The gauge boson associated with the unbroken generator, T1, remains
massless.

We have seen how in the tree approximation the Higgs mechanism generalizes to the non-Abelian case. If we
can successfully quantize the theory and prove it is renormalizable in any gauge, though perhaps not this one
(which is convenient for seeing what’s happening in this simple case), then all of this apparatus will, mutatis
mutandis, carry through to the quantum theory, for exactly the same reasons that were given for a pure scalar
theory. (Regrettably, I won’t be showing you the proof of renormalizability, because it’s a bit too complicated for
the time remaining in the course.) We have a Γ, a generating functional, a function of classical Φ fields and
classical A fields. We go to the minimum, we expand around the minimum, we work things out. It all goes through
smooth as silk. If it goes through for scalars it goes through for vectors also. Therefore our only remaining problem
of principle is to quantize this theory and find out precisely what are the Feynman rules for non-Abelian gauge
theories. That I will attack next time.

1[Eds.] See note 13, p. 944.

2[Eds.] Neil W. Ashcroft and N. David Mermin, Solid State Physics, Saunders College Publishing, 1976, pp.
705–709.
3[Eds.]Ryder QFT, Section 8.4, pp. 296–298; I. J. R. Aitchison and A. J. G. Hey, Gauge Theories in Particle
Physics vol. II: QCD and the Electroweak Theory, Institute of Physics Publishing, 2004, Section 17.7, pp.
218–224.
4[Eds.]
Coleman Aspects, “Secret Symmetry”, Section 2.4, pp. 121–124; Ryder QFT, Section 8.3, pp. 293–296;
Cheng & Li GT, Section 8.3, pp. 240–247; Peskin & Schroeder QFT, Section 20.1, pp. 690–700.
5[Eds.] Example 2, p. 941; see also Coleman Aspects, “Secret Symmetry”, Section 2.2, pp. 118–119.
6[Eds.] P. W. Anderson, “Plasmons, Gauge Invariance, and Mass”, Phys. Rev. 130 (1963) 439–442; F. Englert
and R. Brout, “Broken Symmetry and the Mass of Gauge Vector Mesons”, Phys. Rev. Lett. 13 (1964) 321–323; P.
W. Higgs, “Broken Symmetries, Massless Particles and Gauge Fields”, Phys. Lett. 12 (1964) 132–133; “Broken
Symmetries and the Masses of Gauge Bosons”, Phys. Rev. Lett. 13 (1964) 508–509; “Spontaneous Symmetry
Breakdown without Massless Bosons”, Phys. Rev. 145 (1966) 1156-1163; G. S. Guralnik, C. R. Hagen, and T. W.
B. Kibble, “Global Conservation Laws and Massless Particles”, Phys. Rev. Lett. 13 (1964) 585–587; Peskin &
Schroeder, QFT, pp. 690–692, 731–739; Cheng & Li, GT, pp. 241–243; Ryder QFT, 301–303. Englert and Brout,
Higgs, and Guralnik, Hagen, and Kibble shared the 2010 Sakurai Prize for contributions to the electroweak theory;
Peter Higgs and Fran¸cois Englert shared the 2013 Physics Nobel Prize for the work leading to the 2012
discovery of the Higgs boson. Sadly, Englert’s colleague Robert Brout died in 2011; the Nobel Prize is not
awarded posthumously.
7I should add: This operation would be singular if we were working near the point ρ = 0, because the phase θ of
the field is not well-defined when the magnitude ρ of the field vanishes. As we are doing a perturbation theory
expansion about ρ = a we needn’t worry; but in the absence of spontaneous symmetry breaking it could be a
problem.
8[Eds.] Figure 43.5, p. 941.
9[Eds.]Coleman Aspects, Section 2.3. See note 5, p. 646. The modern development of gauge theories began with
the epochal paper of Yang and Mills, which generalized the U(1) group to isospin and SU(2): C. N. Yang and R. L.
Mills, “Conservation of isotopic spin and isotopic gauge invariance”, Phys. Rev. 96 (1954) 191–195. Ronald Shaw,
a doctoral student of Salam’s at Cambridge, independently found an essentially identical theory in January 1954
(six months after Yang and Mills) and presented it in the last part of his PhD dissertation (1955): “Invariance under
general isotopic gauge transformations”, Part II, Chapter III. In retrospect, the first extension of the gauge
invariance in electromagnetism was Einstein’s general relativity (1915), accomplished before Weyl elucidated the
modern view of gauge invariance: H. Weyl, “Elektron und Gravitation”, Zeits. f. Phys. 56 (1929) 330–352. This
point of view was promoted by Ryoyu Utiyama, who extended the gauge prescription to an arbitrary group and
explicitly drew attention to the close relation between general relativity and Yang–Mills theory: R. Utiyama,
“Invariant theoretical interpretation of interaction”, Phys. Rev. 101 (1956) 1597–1607. For a collection of the
fundamental articles on gauge theories, including the last part of Shaw’s dissertation, Utiyama’s paper, and an
English translation of Weyl’s article, together with a valuable historical survey, see L. O’Raifeartaigh, The
Dawning of Gauge Theory, Princeton U. P., 1997. Gauge theories provide the framework for theories of all the
fundamental forces, and the literature on them is enormous. Every modern quantum field theory book discusses
the topic, e.g., Peskin & Schroeder QFT Chapters 14–22; Itzykson & Zuber QFT, Chapter 12; Ryder QFT,
Sections 3.5, 3.6, and Chapter 7. Entire books are devoted to gauge theories, e.g., Cheng & Li GT, and Chris
Quigg, Gauge Theories of the Strong, Weak, and Electromagnetic Interactions, 2nd ed., Princeton U. P., 2013.
Two of the earliest surveys remain very useful: J. C. Taylor, Gauge theory of weak interactions, Cambridge U. P.,
1976, 1979, and Abers & Lee GT.
10[Eds.] In the videotape of Lecture 51 (on which this chapter is based), Coleman uses a real representation since
these are real fields. However, in order to keep a single notation throughout these lectures, we use the Hermitian
representation.
11[Eds.]
This construction, standard in Lie algebras, is called the Cartan–Killing metric: Élie Cartan (1869–1951),
French, widely regarded one of the greatest mathematicians of the twentieth century; Wilhelm Killing
(1847–1923), German geometer and algebraist. See Zee GTN, Section VI.3, pp. 365-366.
12[Eds.] The sign of δωa (46.20) and the sign of the term in Aµa in (46.30), can be chosen in four different ways,
or, if you prefer, two different classes: both the same or both different. Each of these choices yields a unique set of
relative signs in (46.39). Note however that these choices differ from the Abelian (27.13) and (27.48).
13[Eds.] See note 9, p. 1016.
14[Eds.] S. Weinberg, Gravitation and Cosmology: Principles and applications of the general theory of relativity,
John Wiley and Sons, 1972, equation (6.5.1), p. 140; Charles W. Misner, Kip S. Thorne, and John Archibald
Wheeler, Gravitation, W. H. Freeman, 1970, 1971; reissued by Princeton U. P., 2017, Exercise 16.3, p. 389; A.
Zee, Einstein’s Gravity in a Nutshell, Princeton U. P., 2013, equation (4)–equation (9), pp. 341–342.
15[Eds.] See Weinberg op. cit., Section 12.4, pp. 364–365, in particular equation (12.4.2); Misner et al., op. cit.,
§21.2, pp. 491–492, specifically equation (21.18); Zee op. cit., equation (9), p. 390. For a discussion of gravity as a
field theory, see Zee QFTN, Section VIII.1, pp. 433–447. Both Feynman and DeWitt investigated Yang–Mills
theories as a warm-up to a quantum theory of gravity; see the references in note 10, p. 625, and once again, note
9, p. 1016.
16[Eds.] F. W. Byron and R. W. Fuller, Mathematics of Classical and Quantum Physics, Dover Publications, New
York, 1970, p. 596; Cheng and Li, GT, p. 87; Zee GTN, pp. 63–64. A subgroup S of a group G is called normal if it
is turned into itself under the action of the elements of G: gSg−1 = S. (This doesn’t mean that the elements are
individually invariant; the subgroup Z of G whose elements each commute with all the elements of G is called the
center of the group.) A normal subgroup is thus invariant under the group. A simple group is a group with no
normal subgroups (excluding the identity and the full group itself).
17[Eds.] The GSW model is the subject of Chapters 48 and 49.
18[Eds.] See note 13, p. 944.
19[Eds.] See note 14, p. 944.
20[Eds.] As Higgs himself has emphasized, this idea was first suggested by Philip Anderson, who conjectured that
“the Goldstone zero-mass difficulty is not a serious one, because one can probably cancel it off against an equal
Yang–Mills zero-mass problem,” though Anderson did not give any example of how this might happen. Peter
Higgs, “My Life as a Boson: The Story of ‘The Higgs’”, Int. J. Mod. Phys. A 17 (supplement 01), (2002) 86–88; P.
W. Anderson, “Plasmons, Gauge Invariance, and Mass”, Phys. Rev. 130 (1963) 439–441. Anderson’s remark is in
his penultimate paragraph, p. 441.
21[Eds.] Equations (46.66) and (46.67) are sometimes written Ta|0ñ = 0 and Tb|0ñ ≠ 0, respectively.
22[Eds.] See Abers and Lee, op. cit. (note 9, p. 1016), p. 28, equation (3.20).
23[Eds.] S. Weinberg, “General Theory of Broken Local Symmetries”, Phys. Rev. D7 (1973) 1068–1082.
24[Eds.] S. Weinberg, “Physical Processes in a Convergent Theory of the Weak and Electromagnetic
Interactions”, Phys. Rev. Lett. 27 (1971) 1688–1691; Abers and Lee, op. cit. (note 9, p. 1016), Section 3, pp.
20–25. A different class of gauges, the R ξ gauges of G. ’t Hooft and B. Lee, makes the renormalizability of the
theory manifest, though its unitarity is obscured: G. ’t Hooft, “Renormalizable Lagrangians for Massive Yang–Mills
Fields”, Nucl. Phys. B35 (1971) 167–188; Benjamin W. Lee, “Renormalizable Massive Vector-Meson
Theory—Perturbation Theory of the Higgs Phenomenon”, Phys. Rev. D5 (1972) 823–834.
25[Eds.] Weinberg calls such theories “cryptorenormalizable”; Weinberg, “General Theory of Broken Local
Symmetries”, op. cit.
26[Eds.] For reference, the number of generators of SO(n) is n(n – 1), and the number of generators of SU(n) is
n2 − 1: Zee GTN, p. 80 and p. 237, respectively.
 47
Quantizing non-Abelian gauge fields

We’ve investigated gauge field theories as classical field theories, including the effects of spontaneous symmetry
breaking. From our previous experience with field theories of scalars and spinors we know that if we can construct
a quantum theory, the classical theory can be reinterpreted as the first term—the tree approximation—in a
systematic perturbative expansion. If we ever want to obtain higher order corrections, however, we have to
construct the corresponding quantum theory. So we will now turn to the problem of quantizing gauge field
theories.1

47.1Quantization of gauge fields by the Faddeev–Popov method

We have already quantized the simplest gauge theory, electromagnetism.2 Let me provide a brief review of that
earlier treatment of electromagnetism, an aide méemoire for what is to come.

In the quantization of electromagnetism we used a magic method due to Faddeev and Popov. It may have
seemed complicated when we were struggling through it, but it is simple compared to what we now have to
confront. Recall how Faddeev–Popov quantization worked in electromagnetism. We started with a gauge invariant
action GI. We grouped all the dynamical variables—scalar, spinor, vector, whatever—together under a single
symbol, Φ (31.12). The first step was to choose an appropriate equation

to fix the gauge. We had several choices. One that proved especially useful was axial gauge

In it we were able prove directly that the functional in the Faddeev–Popov prescription was equivalent to canonical
quantization. Other gauges, specified by the condition

(where f(x) is some specified function of x) were also useful, because they gave us simple Feynman rules. By
subsequent functional integration over f we were able to get ∂µAµ up into an exponential; that led to Landau gauge
(31.37), Feynman gauge (31.36), etc., the so-called covariant gauges (31.35).3 Once we’ve picked a gauge, the
Faddeev–Popov prescription says that the generating functional Z is given by (31.22),

where GI is the original gauge invariant action, δ[F(Φ)] is a functional delta function of F(Φ), and the gauge
invariant determinant Δ is

Because of the delta function, we only needed to evaluate the determinant at the point F = 0 (ω is the parameter in
the original gauge transformation). Although it does not tell us the right thing to do for a given theory, the
Faddeev–Popov ansatz has the great advantage that it gives an expression that is manifestly gauge invariant and
independent of the choice of the function F, so long as F is a well-posed gauge fixing term. Once we proved the
Faddeev–Popov method was valid in one gauge (this task was easiest in axial gauge), we knew it was valid in any
other gauge.

It is essentially trivial to generalize this way of doing field theory to the case in which there is a larger gauge
group than in QED. Instead of gauging the U(1) group of electrodynamics, we gauge the SU(2) group or SU(3) or
whatever it is, with its several gauge parameters. We have many gauge fields, so we need to impose one gauge
condition for each of these fields, or, what is the same thing, one for each group generator:
A typical gauge condition could be the obvious generalization of (47.2)

or of (47.3),

We have to integrate over the surface where all the gauge conditions are held fixed, so we have a delta function
for each gauge condition. Finally, we have the functional determinant Δ that we need to cancel out the changes in
the delta function from gauge transformations; it’s the Jacobian factor arising from integrating the delta function.
This will be a determinant not only in function space but in a−b space as well. By trivial extension of the argument
given before, this generalized Faddeev–Popov ansatz with n gauge conditions will be independent of the choice of
gauge; we don’t have to prove that all over again:

with

The question is, can we prove that quantization according to (47.9) is equivalent to canonical quantization, along
the same lines as before? What does quantization look like in the covariant gauge (47.8), which we know gives us
nice Feynman rules in electrodynamics? The first step will be to show that Faddeev–Popov quantization is
equivalent to canonical quantization. I will not go through the whole argument again, but simply point out the
places where there might be differences, and show that everything goes through all right, unaltered from the
Abelian case.

Here we go, in the axial gauge (47.7). Recall for the unscaled vector field,

which becomes, after and dropping the primes on the vector,

Using (47.11), the change in A3a is

This is a mess. Fortunately, since we evaluate the determinant at the zero of the delta function δ[Fa(Φ)] = δ[A3a],
the first term can be dropped:

It’s just the same as in the Abelian case;4 the determinant is a constant:

Δ is independent of the fields over which the functional integration will be performed, and so it’s a constant; we
don’t have to worry about it. Thus Δ is irrelevant and can be absorbed into the normalization N.

The next step in showing the equivalence between the axial gauge generating functional and canonical
quantization was to write the theory in so-called first-order form, where we treat the F’s and the A’s as independent
variables. The gauge fields are the only part of the Lagrangian that’s relevant.

The other fields in the theory are not going to change much; it’s going to be a bunch of normal electrodynamic-type
couplings, which I won’t worry about. It’s only the possible nonlinear terms in (47.15) that may give us trouble. If
you treat as an independent variable and vary it, you obtain the defining equation for , the monstrous
expression in parentheses:
This is exactly the field strength tensor that we had computed before, (46.54). Plug that back into the Lagrangian
and obtain the second-order form,

So the first- and second-order forms give the same theory.

In the Abelian case we showed (§31.4) that functional integration in the axial gauge is equivalent to canonical
quantization in the first-order form, by dividing the variables into three sets (i = {1, 2}):

The “coordinates” Ai and “momenta” F0i were the only independent variables; all the other components of Aµ and
Fµν were superfluous variables given in terms of the coordinates and momenta at a fixed time by solving the
Euler–Lagrange equations. We then calculated with the functional integral. The constrained variables entered at
most quadratically and the coefficients of quadratic terms were constants independent of the fields. Thus we just
integrated over the constrained variables and eliminated them. We were left with the Hamiltonian form ((30.1) and
(31.45)) of the functional integral, which we know is equivalent to canonical quantization. Here we divide the
variables up in exactly the same way, putting in an extra index a on everything, and then checking that the
constrained variables enter the Lagrangian at most quadratically with constant coefficients (i = {1, 2}):

Looking at (47.15), we only have to worry about the extra trilinear terms. Let’s go through the constrained
variables one at a time and see what we get.

A0a: This term appears in the first-order Lagrangian in the combination gFµνacabcAµbAνc , i.e.,

Because gcabcAµbAνc is an antisymmetric tensor (from the symmetry properties of the structure constants), if Aµb
= A0b, then Aνc will have to be a space index, i. So Fµνa will be F0ia, a momentum. That is, we have

This term is a “momentum” times a “coordinate” times the dependent variable A0b; that’s linear in the constrained
variables, so it’s not a problem.

Fija: That can appear as the term

Both Aib and Ajc are “coordinates”, so this term is also linear in the constrained variable Fija.

Fi3a: The trilinear term is

The term vanishes by the gauge condition A3c = 0; there’s no trilinear term to worry about, so that’s trivially all
right.

: The comments about apply here as well; the trilinear term proportional to A3c vanishes.

It’s the same machine as before, and it runs just the same. I won’t go through the proof step-by-step. I just
wanted to point out that the extra terms that distinguish the non-Abelian case from the Abelian case have
absolutely no effect on the argument. Thus the Faddeev–Popov ansatz works just as well in the non-Abelian
theory. It is equivalent to canonical quantization by exactly the same proof we gave in the Abelian case.

47.2Feynman rules for a non-Abelian gauge theory

The next task is to find the Feynman rules, which we’ll derive from the functional (47.9). The axial gauge is
wonderful for proving the theory can be canonically quantized, but it’s terrible for deriving the Feynman rules; it’s
not even covariant. We know a better gauge for the Feynman rules:

where fa is an arbitrary function of x. We’ll explore the consequences of the ansatz in one of these Lorenz-like
gauges. We have to compute the change in this object Fa; fa is a c-number function that doesn’t change under a
gauge transformation. Using (47.11),

In the adjoint representation the structure constants themselves form a representation of the group generators:5

Then (47.21) becomes

This is exactly the covariant derivative operator (46.30) acting on a field that transforms according to the adjoint
representation of the group, i.e., like the gauge fields themselves or like the field strength tensor. Then

We know how to write a determinant as an integral over ghost fields,6 and here we have to do that, because Δ is
not a constant with respect to the fields. Introducing a set of ghost fields ηa, the determinant can be written

where the ghost action g is

(We first exploited this famous trick while studying derivative couplings.7) We don’t care about the overall
constant. This is the only difference between the derivation of the Feynman rules for the Abelian and non-Abelian
gauge field theories. In the Abelian case the structure constants cabc are zero (there is only one generator, which
trivially commutes with itself) and therefore we had to find the determinant of a constant; we didn’t have to
introduce any ghost fields. Here the determinant does depend on the dynamical variables Aµa so we have to
introduce ghost fields. The rest of the argument proceeds as before.

We exponentiate the argument of the delta function (31.29) by integrating with an appropriate function of f to
put a term proportional to

into the exponent as in the steps from (31.29) to (31.31). We arrive at the following effective Lagrangian—the thing
that has to be put into the functional integral (maybe we should call it the “Feynmanian”): the original gauge
invariant Lagrangian, together with the (∂µAµa)2 to adjust the transverse parts of the propagator as we please, and
the ghost part:

(The free parameter ξ which determines the gauge could be different for different fields.) The ghosts now have
real dynamics. In the tree approximation they are massless charged ghost fields (massless because there is no
mass term), interacting trilinearly with the vector boson. They have a very funny looking interaction; it doesn’t look
gauge invariant. Well of course it doesn’t look gauge invariant! The whole point of the Faddeev–Popov ansatz is to
pick a gauge which destroys gauge invariance. The gauge-fixing term (∂µAµa)2 doesn’t look gauge invariant either,
and with good reason; it’s not invariant. The ghosts are just to be treated like normal particles except that,
strangely enough, they have a Fermi minus sign for every closed loop. That’s their peculiar feature; it’s what
makes them ghosts. The whole thing’s got to work out; you’ll never get a negative residue in a Green’s function
from going around a ghost loop, because every gauge invariant quantity could just as well be computed in axial
gauge where there are no ghosts. So the ghosts always have to cancel out against longitudinal photons or what
have you in any specific calculation of a gauge invariant quantity. But they’re there. You can’t get away from them.
It was a great discovery that they are necessary.

The history of the ghost fields is interesting. Feynman and Bryce DeWitt, independently and around the same
time, started trying to quantize gravity by guess work; a messy problem.8 They realized that Yang–Mills fields
have many of the same features as gravity. (Recall my earlier discussion of gravity as to why nonlinear self-
coupling of the fields is necessary.9) They saw that Yang–Mills theory is simpler than gravity, and must have said
to themselves, “We only have to go up to quartic terms instead of an infinite series, so let’s try to quantize
Yang–Mills fields.” Independently they tried to guess the right Feynman rules and then computed away. Both men
discovered that their first guesses didn’t work; they found a breakdown of unitarity and gauge invariance and
everything else, e.g., at the one-loop level the imaginary part of the forward scattering amplitude was not given by
the Optical Theorem. The reason it wasn’t is that they had left out a term, which they eventually realized was a box
with a ghost going around the box. Then they gave up, because they didn’t know how to go beyond the one-loop
level. There the matter sat for eight or nine years, until Faddeev and Popov came along and invented this method.
They showed that the ghosts not only cured the problems at the one-loop level but at all levels.10

The effective Lagrangian (47.27) is horrendous, and we’re not going to do any non-tree level computations in
this theory—except for one fairly trivial calculation. One can in principle read off the Feynman rules just by
standard methods: every derivative becomes a momentum, etc. But they are very cumbersome because
everything is carrying so many indices, especially if you look at vertices connecting multiple vector bosons. Every
one is carrying a momentum, a Lorentz index and an internal symmetry index. Since things get God-awful looking,
I’ll write down a few simple things and work out one complicated one. I won’t even try to write down the complete
set of Feynman rules; no one would remember them.11

For pure gauge theories there is the gauge boson which carries indices µ and ν and indices a and b. It has a
propagator

That is the conventional propagator: the bosons are independent (δab) and they always have zero mass.

We have the ghosts which are charged particles (η ≠ η), so we indicate them by directed, dotted lines with
indices a and b, and their propagator is simply the conventional propagator for a massless scalar field:

The only non-conventional thing comes in when one considers ghost loops and then one has to add an extra
minus sign; these ghost fields are scalars, but they obey Fermi statistics.

Then there are all sorts of interactions coming from GI. There is in particular a tri-vector interaction from the
term. has (47.16) a term linear in the derivatives of the fields plus a term quadratic in the fields. So from
the cross term there will be an interaction like Figure 47.1. We will shortly compute it because that’s the messiest
one: it’s got derivatives and internal indices and space-time indices.

Figure 47.1: Tri-vector vertex

There is a quad-vector interaction, Figure 47.2, which comes from the square of the non-derivative term in
2. That one is actually not quite so horrendous because it doesn’t have any factors of momentum to keep
straight.
 Figure 47.2: Quad-vector vertex

There is an interaction of the gauge fields with the ghosts, coming from the Aµa factor hiding in the covariant
derivative (47.26) of the ghost field, as shown in Figure 47.3.

All three of these things have Feynman rules that are nasty to work out; I’ll do just the first of them, the most
dreadful of the lot, shown in Figure 47.1. Then I will make some remarks about its physical meaning. People
normally scream when they see this rule in a paper; they say “Ugh, where does that come from?” I’ll try to
convince you that it’s really very sensible physically.

Figure 47.3: Ghost-vector vertex

The term in the Lagrangian that’s going to do the dirty work for the trilinear interaction is

The original coefficient was − but the cross term doubles it. We can simplify this a little bit by first observing that
cabcAµbAνc is automatically an antisymmetric tensor in µ and ν so there’s no need to keep both terms in (∂µAνa −
∂νAµa); one of them gives the same thing in the summation as the other so the coefficient is just −1:

I used α and β in because I want to save µ and ν for a different purpose.

Let’s look at Figure 47.1. On one leg we have a vector boson carrying indices a, µ with momentum p; another
has b, ν, momentum q; the third, c, λ, momentum r. The momenta are not independent, of course:

I’ll choose them all directed inward. Figuring out this vertex is difficult because there are so many possibilities:
which field absorbs which of the three bosons. We’ve got to worry about all of them. There are 3! possibilities
coming up. Some factors are common: there’s a −ig in front; since we’re always differentiating an incoming field to
get a momentum, we pick up −i from that; it’s like an annihilation operator. There’ll always be a cabc in some
permutation or other, and aside from minus signs one permutation is the same as another because of our cunning
symmetry condition, (46.26). So we can just write it as cabc and worry about whether we get plus or minus signs in
various combinations.

Now comes the mess. Let’s take the case where one boson, (µ, p), is absorbed by the first factor—the one
carrying the a; the other bosons are absorbed by the b and c, no permutations, just as stated. The first boson is
being differentiated so we get a p because the first boson carries momentum p. But p with what index? The same
index as the derivative which is the index of the second boson, so it’s pν. The other two indices are being summed
together so I get gµλ. There are two other terms that are trivially obtained from these by cyclic permutations; those
I can just cycle around clockwise. I get q with the next index over, λ. Then I have two remaining indices to sum,
gµν. Then r with the next index over, µ, times gνλ. Then there are the terms where I go anti-cyclic. Instead of
summing each of these with the index attached to each of the momenta in the clockwise direction, I attach the
index in the anti-clockwise direction, and thus cabc changes sign. So we have pλgµν, minus q with the next index
up, qµgνλ, minus rνgµλ. What a mess! If you keep your wits about you you can derive it.12 The final expression for
the trilinear vector boson vertex is:
God have mercy on anyone who tries to do a two-loop computation with these things appearing at every vertex
and having to be summed over. Tini Veltman wrote a computer program, schoonschip, to do these complicated
computations in Yang–Mills theory.13 But there are some simple calculations. I’ll show you the explicit one-loop
computation of the effective potential, which is fairly easy. There, by being cunning in our choice of gauge, we can
make the indices take care of themselves. Similar calisthenics, although slightly less strenuous, will give you the
other two kinds of fundamental vertices, but I won’t bother to derive them.

Though the tri-vector boson vertex looks complicated, in fact it has a very simple physical meaning. Suppose
we consider the theory of an ordinary photon coupled to charged scalar bosons. Remember from scalar
electrodynamics what that vertex looks like.14 I’ll redraw it to ease comparison, with all the momenta going inward;
see Figure 47.4. The photon carries an index µ and the scalars have no indices.

Figure 47.4: Scalar-scalar-vector vertex

As we found earlier, this diagram makes a contribution proportional to the sum of the incoming and outgoing
momenta (due to the derivative coupling of the scalar and the photon). Here one is incoming and one is outgoing,
so

We’ve got a term in (47.33) that is just like that; the coefficient of gνλ is (rµ − qµ). So these two terms can be read
as saying that the particle (a, µ, p) acts like a photon and the other two act like spinless charged particles. It has a
coefficient which is given by the group structure constants, and it doesn’t matter what polarization state they’re in,
what λ or ν is, so this is just a gλν. Each of the two massless gauge bosons has two polarization states. Then
these four polarization states act just like four independent charged scalars. We can look at the vertex in different
ways. We can say (a, µ, p) is like a photon and the others are like a charged particle, or we could say (c, λ, r) is like
a photon and the others like charged particles, or we could say the third one is like a photon and the other two are
like charged particles. In the case of SU(2), the 1 and 2 bosons act as charged particles when seen by the 3,
which acts like the photon; they’re the source of the 3. In the same way, the 1 and the 3 are the source of the 2,
etc. That’s the wonder of Yang–Mills theory; amusing but confusing. So the contribution of the diagram in Figure
47.1 can be thought as the sum of charged scalar-photon interactions over three permutations, depending on
which one we think of as the photon.

In the same way, the four-gauge boson coupling, which also is a mess of permutations, can be thought of as
just summing over permutations of the seagull diagram, Figure 47.5,

Figure 47.5: Scalar-vector seagull vertex

which is also present in scalar electrodynamics,15 changing which pair you think of as the photons and which pair
you think of as the charged particles. Although the vertex in Figure 47.2 looks more complicated than charged
scalar electrodynamics, it’s really not, though certainly the computations are more complicated. It’s only that you
have many choices as to which boson you think of as the photon. Therefore you have sums over many
permutations. This point of view is something you won’t find in the literature, but I think it helps you understand, at
least in a semi-physical way, why these complicated structures necessarily arise.

The full Lagrangian, including coupling to fermions and scalars, is

where (46.59) D µ = ∂µ − igTaAµa. The Feynman rules for this Lagrangian are given in the box on p. 1042
 47.3Renormalization of pure gauge field theories

Let’s consider a pure gauge vector theory, without scalar or spinor fields. All of our vector propagators, no matter
what the gauge, have a denominator k2. At every vertex, we have either four undifferentiated vector fields, or three
fields, one of which is differentiated. So we immediately have the first half of a proof of renormalization. The only
counterterms we will need will be monomials of the same form as the monomials appearing in the original
Lagrangian, with no more fields and no more derivatives. We do not have the proof of the second part of
renormalization, showing that these monomials come through with exactly the right coefficients to correspond to a
redefinition of the parameters in the original Lagrangian. In electrodynamics, we obtain the connection between
the various counterterms by lengthy arguments, systematically exploiting gauge invariance and its consequence,
the Ward identities.16 That argument depends on the Lagrangian’s gauge-fixing terms being no more than
quadratic in the fields, and does not hold in the non-Abelian case: the ξ term is quadratic, but the ghost term is
trilinear in the fields. Therefore another proof is required.
Feynman rules for a non-Abelian gauge theory

The original proofs were unbelievably horrible.17 They were gradually improved and simplified,18 until finally
Becchi, Rouet, and Stora at Marseille found a proof that was only believably horrible. I refer you to those papers if
you want to see the proof.19 Take my word for it, it is possible to prove that Yang–Mills theory is strictly
renormalizable.

Spontaneous symmetry breaking is irrelevant for renormalization; we’ve already established that. We can
compute the generating functional without asking whether or not the symmetry breaks spontaneously. If it does,
we go and look at the generating functional to see if we have to shift the fields. If you’ve proved a theory is
renormalizable, that proof holds regardless of whether or not there is symmetry breaking. That conclusion follows
from the same argument we gave for our scalar field theories. You may have to be concerned if you’re proving
things on a fine technical level. If you do your subtractions `a la BPHZ, have you proved Hepp’s theorem with
subtraction at a Euclidean point, for example. If you don’t have symmetry breaking you have massless vector
bosons. These may give you bad infrared divergences that make things blow up at the BPHZ point. But on the
level at which we’re working it doesn’t matter. In principle, in the generating functional formalism, the question of
renormalizability and the occurrence of spontaneous symmetry breaking are completely separated. First we
renormalize; then we look through the renormalized theory to see if the symmetry breaks.

We should also add that in certain cases, especially theories involving boson coupling to γ5-type currents
(axial vector currents), the naïve proof of the Ward identities can break down due to the occurrence of
anomalies,20 unless great care is taken with the cutoff. Occasionally you have to worry about anomalies. But they
have been cataloged; it’s an exercise in nosology, the categorization of diseases. You only have to check that
they cancel among the various axial vector currents. There is another long song and dance to deal with them.21

Until the work of Faddeev and Popov, the renormalizability of Yang–Mills fields was only conjectured; the
Feynman rules for non-Abelian gauge theories were not known and so renormalizability could not be proven. How
could you possibly tell if the divergences from all the graphs canceled before you knew what the graphs were? It
gets pretty complicated. You’ve got to look at it the right way to make it look simple. Otherwise you will calculate to
a certain level and then you’ll vomit up a bunch of indices and decide to look at another problem. It’s just a matter
of organizing the details. Faddeev and Popov found this neat way using the functional integral to organize the
theory. With their prescription, you can see everything happening at once. You don’t have to drive yourself crazy
computing things like (47.33).

Veltman was probably the only person concentrating on Yang–Mills theory throughout the middle 1960’s. And
he, being a man of great taste, must have said to himself: “If I’m going to tackle this problem I’m going to have to
learn how to manipulate all those indices. It will be hard to avoid mistakes; I’m not a machine.” So he wrote that
computer program, schoonschip, to do the work. But his premise was faulty. He thought, “Well, obviously the
massless theory is the limit of the massive theory. I know how to quantize the massive theory. I’ll just go ahead
with that, so I won’t have to worry about gauge invariance; there isn’t any.” He put in a mass term, intending to
compute everything and at the end to let the mass go to zero. Unfortunately, unlike electrodynamics, the massless
theory of Yang–Mills fields is not the smooth limit of the massive theory. That’s now a well-known result, 22 but it
was not known at the time that Veltman was looking at the massive Yang–Mills theory. For example, if you have
an SU(2) theory you have three photons coming together at a vertex. If you work things out you can always get rid
of one of the three helicity zero photons, but you can’t get rid of all three of them simultaneously. There was just no
way of making it work. So he was, unknown to himself and unknown to everyone else, pursuing a dead end. The
way forward with the massless theory was blocked until Faddeev and Popov’s paper appeared.

The Faddeev–Popov paper was not widely understood at first. It was obscurely written and it took about a
year or two for its results to sink in. In fact it took ’t Hooft’s discovery of the theory’s renormalizability for the
significance of the paper to be appreciated.23 Faddeev and Popov had obtained the Feynman rules in a very
strange way. But Yang–Mills theory finally came together. Essentially independently, ’t Hooft found the right
Feynman rules. Even more, he discovered the crucial point: spontaneous symmetry breaking enables the
construction of theories involving massive vector bosons which might provide, in a way I will describe later, a
renormalizable weak interaction theory. He didn’t know that Weinberg had proposed a similar model four years
earlier. Though he had conjectured that his model would be renormalizable, Weinberg had been unable to see
why it would be so. In his first papers ’t Hooft presented renormalization as a formal argument, manipulating
infinite quantities and imposing gauge invariance when necessary. I remember that time very well. I met Tini
Veltman in Marseille24 and he said, “A graduate student of mine has a renormalizable theory of massive charged
vector bosons,” and I said, “I don’t believe it.” He said, “It’s true,” and we were about to make a bet. It’s a good
thing we didn’t because I would have lost a lot of money. That’s why I’m one of the few people in the world who
doesn’t mispronounce ’t Hooft’s name, because Veltman told me his name, Gerard ’t Hooft. When I got the
preprint and saw how he spelled his name, my reaction was, “What a funny way to spell ‘et Hoaft’ instead of the
reaction of everyone else, which is “What a funny way to pronounce ‘tooft’ ”.

47.4The effective potential for a gauge theory

Let’s calculate something in this theory of non-Abelian gauge fields. We’ll choose a quantity that we’ve been
computing in stages, the effective potential.25 We’re generalizing the effective potential to add gauge fields to the
scalar and fermion fields we’ve considered previously. It’s a very simple object because it only involves external
scalar lines, and they all carry zero momentum. Even though we’re summing up a huge number of graphs, they’re
very simple graphs.

The effective potential V (ϕ) which depends upon the classical scalar fields is going to be the sum of several
terms. There’s the zeroth order contribution U(ϕ); the contribution VS of the scalars themselves,

(adding the trace over the internal group indices on the scalar fields), the contribution VF from the fermions,

where m(ϕ) is the fermion mass matrix, (45.22); the contribution VG from the gauge fields which we are about to
compute; and finally the contribution from the counterterms. These are finite terms of the same form as other
quantities that occur in the classical potential U, and are determined by the other quantities once we fix our
renormalization conditions. All together we have for the one-loop contribution

Now for the contribution from the gauge fields. At first glance we get a horrible mess. We can have graphs
like Figure 47.6, where there are scalar fields on the outside and gauge fields running around inside. Those we
can handle by our usual techniques.

But we also have the trilinear scalar-scalar-vector interactions, and thus the possibility of a graph in which
something like Figure 47.7 happens. However, by an astute choice of gauge we can make these disappear.

Figure 47.6: Loop with A2-ϕ 2 couplings

Figure 47.7: Loop with A-ϕ-∂ϕ couplings

I should have emphasized this important point earlier: if a theory includes a gauge field, the effective potential
is not a gauge invariant object because the ϕ fields are not gauge invariant. Then again, neither is the propagator.
Nothing is gauge invariant until you put it all together and assemble physically observable quantities; those are
gauge invariant. Any gauge should be as good as any other, so long as you don’t change gauges in the middle of
a computation.

We will use Landau gauge, in which the propagator is 26

Why is that a good gauge? Let’s focus on what’s happening at the ϕ-∂ϕ-A vertex.
 Figure 47.8: ϕ-∂ϕ-A vertex

Figure 47.8 shows a scalar boson coming out with momentum zero, a gauge boson coming in with momentum k
and a scalar boson emerging with the same momentum k. Therefore, depending upon how you orient things, the
sum or difference of the momenta carried by the internal scalar boson (in Figure 47.7) is k. (In this case the
orientation doesn’t matter.) At the vertex we have a factor kµAµ because it’s always the sum of the momenta that
occurs in item (i) in the list of vertices. In the Landau gauge k hits the propagator and kills it:

So as long as a boson line has external momentum equal to zero, the vertex vanishes! We don’t have to worry
about the ϕ-∂ϕ-A; we just have to worry about graphs of the kind in Figure 47.6.

What are those graphs? They’re graphs like Figure 47.9:

Figure 47.9: Vector boson loop

using a little black dot to indicate that four-boson interaction, just the same way as before (see (44.44) and
(45.24)). That black dot comes from the term in the Lagrangian

the D µϕ ⋅ D µϕ term expanded out to second order in A. This is just the mass term of the vector boson in the
presence of an external ϕ field, and therefore the form of this vertex, the form of the black dot, is simply

The vector boson mass matrix is defined in (47.39); it’s the mass of the vector boson in a given ϕ field. That
acts just like the fermion case (45.22) where we had m(ϕ) appearing at each vertex.

We still have to deal with all those propagators running around the loops. The propagators are δ functions in
ab space but they do have µν indices:

Fortunately the −i from the propagator cancels the i from the vertex, and you always have one propagator for each
vertex. Also fortunately, the factor in the square brackets of the propagator (47.41) is a projection operator in µν
space, and thus idempotent:

Whether you multiply three lines or 17 lines or 121 lines, you just get the same thing. At the end you have to take
the trace:

Otherwise the computation is exactly the same as the scalar or the spinor or any other computation we’ve done.
We’ve got a string of matrices running around, we have a

for every internal line. All the index structure collapses and it becomes the scalar computation all over again,
because of the choice of gauge. The only difference in the vector contribution to the effective potential from the
scalar contribution is a factor of 3 coming from the trace of the Landau gauge propagator:
 Note that this has a definite physical meaning, just as the other contributions had. Remember that VS was the
zero point energy of a sum of independent harmonic oscillators. So is VG. Why does it have a factor of 3? A
massive vector boson has three degrees of freedom! A massive scalar has one. So we’ve got three times as
much zero point energy. There are three virtual oscillators for every momentum state of a massive vector boson.
Thus the factor of 3 is easy to see on physical grounds.

These formulas will be important to us later. At the moment we’ve just been accumulating them for when we
finally discuss things like Weinberg’s famous lower bound on the mass of the Higgs boson.27 But they’re very
simple, aside from the 64π2 which one has to memorize; you figure out what they are just by counting up zero
point energies. The minus sign for fermions, (45.28), is because the zero point energy goes the other way. Instead
of subtracting the zero point energies of the individual oscillators you’re adding the energies of the negative energy
states, filling up the holes in the Dirac sea.

Aside from our discussion of the sigma model as a model of current algebra, all the stuff about gauge fields
and Higgs phenomena and so on, admittedly beautiful (and also elegantly and wittily presented), are nevertheless
just theoretician’s toys, with no apparent connection to the real world. Next time I’ll show you how all these ideas
were put to work, and turn to the theory that makes all this important: the famous Glashow–Salam–Weinberg
model of weak and electromagnetic interactions.

1 [Eds.] Ryder QFT, Section 7.2, pp. 250–260; E. S. Abers and B. W. Lee, Phys. Lett. 9C (1973) 1–141, Sections
12 and 13; Cheng & Li GT, Chap. 9, pp. 248–278.
2 [Eds.] See §31.3, pp. 665.–668.
3 [Eds.] Ryder QFT, Section 7.1, pp. 245–250.
4 [Eds.] See the paragraph before §31.3, p. 665, and the solution to Problem 17.2, p. 682.
5 [Eds.] See note 19, p. 947; Howard Georgi, Lie Algebras in Particle Physics: From Isospin to Unified Theories,
2nd ed., Perseus Books, 1999, Section 2.4, pp. 48–50.
6 [Eds.] See §29.3 and §31.3.
7 [Eds.] See §29.3, pp. 625–628.
8 [Eds.] See note 10, p. 625.
9 [Eds.] See p. 1022.
10 [Eds.] See note 10, p. 625. Feynman presented ghosts at the one loop level in a talk at the 1962 Warsaw
(Jabłonna) conference on gravity (known as “GR3” in the relativity community). Responding to persistent
questioning by DeWitt, Feynman went into detail about the one-loop result; the transcribed talk (and the question
period, following) were published (see note 8, p. 1037): The Feynman Lectures on Gravitation, Richard Feynman,
Fernando B. Morinigo and William G. Wagner, ed. Brian Hatfield, Addison-Wesley, 1995; pp. xxviii–xxix. DeWitt
extended the idea to two loops in 1964, and via a functional integral, to all orders, in the last weeks of 1965. For a
variety of reasons (page charges, a dispute with a reviewer, and other work) DeWitt didn’t publish this last result
until 1967, about two weeks prior to the first appearance of Faddeev and Popov’s results (but not their method): B.
S. DeWitt, “Theory of Radiative Corrections for Non-Abelian Gauge Fields”, Phys. Rev. Lett. 12 (1964) 742–746;
B. S. DeWitt, “Quantum Theory of Gravity II. The Manifestly Covariant Theory”, Phys. Rev. 162 (1967)
1195–1239; L. D. Faddeev and V. N. Popov, “Feynman Diagrams for the Yang–Mills Field”, Phys. Lett. B25
(1967) 29–30; C. DeWitt-Morette, The Pursuit of Quantum Gravity: Memoirs of Bryce DeWitt, Springer, 2011, pp.
20–22; p. 52; pp. 126–127. At the end of 1966, Faddeev was visiting the IHES near Paris at the same time as
Stanley Deser, who introduced him to DeWitt’s work. Spurred by this, Faddeev and Popov wrote up their Physics
Letters article, and shortly thereafter produced a much longer preprint (“Теория возмущений для
калибровочно-инвариантных полей”, Kiev 1967, ITP-67-36). But it was never published—quantum field theory
was doctrina non grata in the former Soviet Union: L. D. Faddeev, “Quantizing the Yang–Mills Fields”, in At the
Frontier of Particle Physics: Handbook of QCD (Boris Ioffe Festscrift), M. Shifman, ed., World Scientific,
2001–2002. An English translation, “Perturbation Theory for Gauge-Invariant Fields”, appeared only in 1972 as a
Fermilab preprint (NAL-THY-57), nine years after Feynman’s Acta Physica Polonica article. Though frequently
xeroxed and passed hand to hand, this prized translation likewise was never published before it appeared in
anthologies decades later: C. H. Lai, ed., Gauge Theory of Weak and Electromagnetic Interactions, World
Scientific, 1981, pp. 213–233; L. D. Faddeev, Forty Years in Mathematical Physics, World Scientific, 1995, pp.
31–51; G. ’t Hooft, ed., Fifty Years of Yang–Mills Theory, World Scientific, 2005, pp. 40–50. The Fermilab preprint
is available online: http://lss.fnal.gov/archive/1972/pub/Pub-72-057-T.pdf.
11 [Eds.] Though not stated in the lectures, the complete set of Feynman rules for a Yang–Mills theory is given in
the box on p. 1042.
12 [Eds.] A more formal but straightforward way to obtain (47.33) is to consider the effective action Γ = d4x for
the interaction (47.31) and take the three functional derivatives (δ/δAµa(x))(δ/δAνb(y))(δ/δAλc (z))Γ. This yields a
series of terms that reproduce (47.33).
13 [Eds.] “Schoonschip” is Dutch for “clean ship”, or loosely, “shipshape”, everything neat and tidy. M. Veltman,
“An IBM-7090 Program for Symbolic Evaluation of Algebraic Expressions, Especially Feynman Diagrams”, CERN
PRINT-65-879 (1965); M. Veltman, “schoonschip”, CERN preprint, July 1967; M. Veltman and D. Williams,
“Schoonschip ’91” (University of Michigan preprint UM-TH-91-18, June 9, 1991); available at
https://arxiv.org/abs/hep-ph/9306228. Written in assembly language, Veltman’s program was designed to
automate the calculation of roughly 50,000 terms in radiative corrections to a process of photons interacting with a
charged vector boson. It is arguably the first computer program written to perform symbolic algebra.
14 [Eds.] See rule (h) and its diagram in the box on p. 670.
15 [Eds.] See Figure 27.1, p. 585, and rule (i), p. 670.
16 [Eds.] §32.4, specifically (32.50); §33.4.
17 [Eds.] G. ’t Hooft, “Renormalization of Massless Yang–Mills Fields”, Nuc. Phys. B33 (1971) 173–199;
“Renormalizable Lagrangians for Massive Yang–Mills Fields”, Nuc. Phys. B35 (1971) 167–188; G. ’t Hooft and M.
Veltman, “Regularization and Renormalization of Gauge Fields”, Nuc. Phys. B44 (1972) 189–213; “Combinatorics
of Gauge Fields”, Nuc. Phys. B50 (1972) 318–353. Gerard ’t Hooft and Martinus Veltman shared the 1999
Physics Nobel Prize for the proof of the renormalizability of massless and massive (via the Higgs mechanism)
Yang–Mills fields.
18 [Eds.] A. A. Slavnov, “Ward Identities in Gauge Theories”, Theo. Math. Phys. 10 (1972) 99–104; B. W. Lee,
“Renormalizable Massive Vector-Meson Theory–Perturbation Theory of the Higgs Phenomenon”, Phys. Rev. D5
(1972) 823–835; J. C. Taylor, “Ward Identities and Charge Renormalization of the Yang–Mills Field”, Nuc. Phys.
B33 (1971) 436–444; J. C. Taylor, Gauge Theories of Weak Interactions, Cambridge U. P., 1976, 1978, Chapters
12, 13, and 14, pp. 94–127.
19 [Eds.] “The generalized Ward–Takahashi identities for non-Abelian gauge theories were first formulated in a
rather complicated way . . . Fortunately the formulation of these identities has been simplified by a device due to
Becchi, Rouet, and Stora.” Taylor, op. cit., p. 94: C. Becchi, A. Rouet and R. Stora, “The Abelian Higgs–Kibble
Model. Unitarity of the S Operator”, Phys. Lett. B52 (1974) 344–346; “Renormalization of the Abelian Higgs–Kibble
Model”, Commun. Math. Phys. 42 (1975) 127–162; “Renormalization of Gauge Theories”, Ann. Phys. 98 (1976)
287–321; also in Renormalization Theory, G. Velo and A. S. Wightman eds., Reidel, 1976; I. V. Tyutin, Lebedev
Institute preprint FIAN n. 39 (1975) (unpublished); M. Z. Iofa and I. V. Tyutin, “Gauge Invariance of Spontaneously
Broken Non-Abelian Theories in the Bogolyubov–Parasyuk–Hepp–Zimmerman Method”, Theo. Math. Phys. 27
(1976) 316–322; Ryder QFT, Sections 7.5 and 7.6, pp. 277–282; Cheng & Li GT, Section 9.7, pp. 267–278;
Peskin & Schroeder QFT, Section 16.4, pp. 517–521. The generalized Ward–Takahashi identities are often called
“Slavnov–Taylor” identities, and the “device” referred to by Taylor is usually described as the “BRST
transformation”, after Becchi, Rouet, Stora, and Tyutin, who shared the 2009 Dannie Heineman Prize for its
discovery; Itzykson & Zuber QFT, Section 12-4, pp. 594–606. See also Gerard ’t Hooft, “Reflections on the
renormalization procedure for gauge theories”, Nuc. Phys. B912 (2016) 4–14, a memorial issue to Raymond Stora
(1930–2015).
20 [Eds.] J. S. Bell and R. Jackiw, “A PCAC Puzzle: π0 → γγ in the σ Model”, Nuovo Cim. A60 (1969) 47–61;
Steven L. Adler, “Axial-Vector Vertex in Spinor Electrodynamics”, Phys. Rev. 177 (1966) 2426–2438; Barry R.
Holstein, “Anomalies for Pedestrians”, Am. J. Phys. 61 (1993) 142–147; Cheng & Li GT, Section 6.2, pp. 173–182.
See also note 5, p. 82.
21 [Eds.] Weinberg QTF2, Chapter 22, pp. 359–420.
22 [Eds.] A. A. Slavnov and L. D. Faddeev, “Massless and Massive Yang–Mills Fields”, Theo. Math. Phys. 3
(1971) 312–316.
23 [Eds.] See note 17, p. 1042.
24 [Eds.] Presumably the two met at the Colloquium on Renormalization Theory, CNRS, Marseille, June 1971. At
Veltman’s invitation, ’t Hooft gave a brief report of his results at the Amsterdam International Conference on
Elementary Particles, 30 June–6 July 1971. See Frank Close, The Infinity Puzzle, Basic Books, 2011, Chapter 11,
“And Now I Introduce Mr. ’t Hooft”.
25 [Eds.] The effective potential is defined by (44.38), calculated for the scalar field in §44.3 and for the fermion
field in §45.2; “Secret Symmetry” in Coleman Aspects, Section 3.5, pp. 136–138; Appendix, pp. 180–182.
26 [Eds.] See note 12, p. 667.
27 [Eds.] See §49.3.

Problems 25

25.1 A real vector field of mass µ is coupled to a real scalar field of mass m in an unconventional way:

where g is a real number. The vector field is not coupled to a conserved current, and thus we might expect the
theory to suffer from various ailments.

We will choose to study vector-scalar elastic scattering,

For this process there are nine independent amplitudes at fixed energy and angle, because both the incoming and
outgoing vectors may have helicity (spin along the direction of motion) equal to any of {1, 0, −1}. An interesting
limit in which to consider these amplitudes is that of high center-of-momentum energy, with CM scattering angle
fixed, but equal to neither 0 nor π. (This restriction guarantees that all Mandelstam invariants—s, t, and u—grow
with energy.)

(a) To lowest nontrivial order in perturbation theory, (g2), some of the nine helicity amplitudes approach (possibly
angle-dependent, possibly vanishing) constants in the high-energy, fixed-angle limit described above. We will call
those amplitudes “nice”. Others, however, grow as a power of the energy; these we will call “nasty”. Which are the
nasty amplitudes? Find the explicit high-energy forms of the nasty amplitudes, retaining only terms that grow as
positive powers of the energy. (Since I haven’t defined the phase of helicity eigenstates, don’t worry about getting
the phase (let alone the sign) of the answer right.)

(b) Now let us add another term to the Lagrangian:

If we add the contribution of this term (in tree approximation) to our previous computation, then, for an appropriate
choice of h, some of the nasty amplitudes become nice. Which ones? What is the appropriate choice of h? (Cf.
Problem 22.2 and its solution.)
(1987 253b Final, Problem 3)

25.2 In class discussions of gauge field theories (§46.2), I described how the matter fields transformed under a
finite gauge transformation,

and also under an infinitesimal one, g = 1 + δω,

where δω ≡−iωaTa (46.27), g ∈ G, and Ta are the generators of some representation of the Lie group G. For the
fields Fµν, I only described infinitesimal transformations,

(the matrix form of (46.45)). It’s easy to see however that this infinitesimal transformation implies that under finite
transformations,

(the matrix form of (46.49)). The argument runs as follows: (1) Every finite transformation can be built up as a
product of infinitesimal ones. (2) The stated transformation law under finite transformations has the group
property: the result of first applying the transformation g1 and then applying the transformation g2 is the same as
that of applying the transformation g2g1. (3) The stated finite transformation agrees with the known infinitesimal
transformation for g = 1 + δω. (If you’re disturbed by taking the infinite product of infinitesimal transformations to
get a finite transformation, you can rephrase the whole argument in terms of integrating differential equations, but
really, it’s not worth the bother.)

(a) Use similar reasoning to show that the matrix form of (46.39)

(where Aµ ≡ iAµaTa) implies

(b) Let x(s) be some path in spacetime, where the path parameter s runs from 0 to ∞. Suppose you have a unitary
matrix, U(s), which solves the differential equation

with the boundary condition U(0) = 1. (Note the similarity to interaction-picture perturbation theory.) Show that the
solution to the differential equation

with the boundary condition U (g)(0) = 1, is

(1987b 17)

25.3 In the lectures on quantum electrodynamics, we studied processes where some initial state i went into some
final state f, plus a photon of momentum k′ and polarization vector ε′. (Both i and f could be multiparticle states.)
The invariant amplitude for this process was (see (26.70) and (35.28))

for some Mµ, the matrix element of a conserved current. Thus, even when the photon was off the mass shell (k′2 ≠
0), k′µMµ = 0. Furthermore, we showed that this remained true even if the initial state contained an off-mass-shell
photon, so long as all the other particles in the initial and final states were on the mass shell. (You may remember
that this was important in our derivation of the low energy theorem for photon-nucleon scattering in §35.3.) For the
purposes of this problem, “on the mass shell” means, for a Dirac particle, not only p2 = m2, but also ( − m)u = 0;
for a gauge boson, not only k2 = 0, but also ε ⋅ k = 0.

As you have seen, non-Abelian gauge field theories are in many ways generalizations of electrodynamics.
Consider a non-Abelian gauge theory with some gauge group G, with a coupling constant g and a set of N Dirac
fields of mass m transforming according to some representation of G with generators Ta. The defining equations of
this theory were given in §46.2, but here they are, summarized:

The T’s are a set of N × N matrices, acting on the internal symmetry indices of ψ only; the coefficients cabc are the
“structure constants” for the Lie algebra of G’s generators {Ta} with cabc = −cbac, Latin indices run from 1 to dim G,
and the sum over repeated indices is implied. (Incidentally, the sign of g differs in the literature.)

Compute k′µMµ for the elastic scattering of gauge bosons off Dirac particles in the tree approximation, i.e., to
order g2. Let all the particles except the final gauge boson be on the mass shell, and investigate the circumstances
when k′µMµ vanishes. Set to compute k′µMµ.

Comment: In this problem I found it convenient to use explicit group indices for the gauge fields, but to treat the
Dirac fields as one big vector with 4N components. Thus the diagram shown below yields the amplitude
This is one of three diagrams you have to consider; the other two are the cross of this (as in electrodynamics) and
t-channel gauge-boson exchange.

(1998b 6.2)

25.4 In the Abelian Higgs model, compute, in tree approximation, vector-scalar elastic scattering for the case in
which both the initial and the final vector mesons have helicity zero, in the limit of high center-of-momentum
energy, with center-of-momentum scattering angle θ fixed, but equal to neither π nor 0. (This guarantees that all
three Mandelstam invariants—s, t, and u—grow with energy.) Show that in this limit, the amplitude approaches a
(possibly angle-dependent) constant, even though some of the individual graphs that contribute to the amplitude
grow as powers of energy. (This is the overt version of the Abelian Higgs model, as opposed to the covert version
in Problem 25.1, above.)
(1998b 11.2)

Solutions 25

25.1 To lowest order in g2, the elastic scattering of a massive vector and a scalar looks like this diagram:

The relevant Feynman rules are for the vector-vector-scalar vertex and the vector propagator:

In the CM frame,

The primed quantities are obtained by rotating by θ;

In the graphs, the gµν term in the propagator makes only nice amplitudes; even in the worst case, helicity 0 to
helicity 0, ε′*⋅ ε grows like E2, but that’s canceled by the denominator. Thus we need only keep track of the kµkν
term:
where I’ve used the orthogonality of the vector’s momenta and its polarization vectors: ε ⋅ p = ε′⋅ p′ = 0. For initial
and final helicities not equal to 0, the numerator grows no faster than the denominator:

The other amplitudes grow with energy:

Since in the worst case ( 00) the amplitude grows like E2, we can safely expand everything for high E and
discard terms that are down by at least two powers of E compared to the leading term. Thus,

The less nasty amplitudes become in the regime of large E

The worst becomes

(b) Now add in the new interaction. The relevant Feynman rule for the vertex is simple:

This results in a new term added to the amplitude (S25.3):

The nasty amplitudes become

If we choose

all of the nasty amplitudes become nice!

Comment: This Lagrangian is, in disguise, just the Abelian Higgs model, after the symmetry breaks
spontaneously:

This “miraculously” mild high-energy behavior is a reflection of the secret renormalizability of the theory. If we
considered simple scalar field theories, this is what we would find for a renormalizable interaction like ϕ 4, but not
for a nonrenormalizable one like ϕ 2(∂µϕ)2. We look at helicity zero states because we know from our study of
vector mesons coupled to non-conserved currents (as this appears to be, if your eyes cannot pierce the veil of
spontaneous symmetry breaking) that these are the states most likely to display pathological behavior.

25.2 (a) First, the given finite transformation can be built up as a succession of n infinitesimal transformations. Let

The infinitesimal transformation

can be written as (keeping only terms up to the first order in Δω)

Applying this twice gives

using the identity (1 + Δω)(1 − Δω) = 1 to first order in Δω. Repeating the operation n times gives

Next, the stated finite transformation obeys the group property:

Finally, the finite transformation agrees with the infinitesimal transformation for g = 1 + δω:

(b) First, note that U (g)(s) satisfies the boundary condition U (g)(0):

Now plug (P25.12) into (P25.11) to see if it works:

using (∂µg(x(s))g−1(x(s)) = −g(x(s))∂µg−1(x(s)). So it does work.

25.3 The scattering can be described by this diagram:

In the tree approximation, the diagram includes three graphs:

The Feynman rules can be obtained from the Lagrangian. From the term we have the vertex (see
the box on p. 1042, item (g))

and from the term gcabc(∂µAνa)AbµAcν we have the vertex (box on p. 1042, item (e))
 We set ε′*ν = k′ν, and look at diagrams (1) and (2) together (remembering that the fermions are on their mass
shell):

Anticommuting ′ and through ′ and using ( − m)u = 0 = ( ′ − m)u′ yields

Now for diagram (3), which includes the vertex (S25.21) and the vector boson propagator, D µν(q), where q =
p′ − p = k − k′. The general covariant gauge propagator

is the sum of the Feynman gauge propagator plus terms in qµ. In (3), this propagator will be contracted with the γµ
in the fermion-meson vertex (S25.20) and sandwiched between ′ and u. But these qµ terms are irrelevant,
because the fermions are on their mass shell:

Thus we might as well use Feynman gauge,

From the general form (S25.21), the upper vertex is (reversing the directions of q = k − k′ and k′)

so that (with ε′*ν = k′ν)

where the dots indicate the contributions from (1) and (2). Using (S25.24) once again,

Adding the contributions (S25.23) from (1) and (2), we find

This expression does not vanish for off-shell incoming mesons, but it does for those on-shell. This is
consistent with QED: in QED, kµMµ ≠ 0 if there are off-shell charged particles (whether bosons or fermions is
irrelevant). With respect to the charge to which Aµb couples, Aµa is charged, unless cabc = 0 for all c.

25.4 The Feynman rules for the Abelian Higgs model are given in the box on p. 1015. The diagrams responsible
for scalar-vector elastic scattering at tree level are shown below:
 The corresponding amplitudes are (note: M2 = a2e2 is the mass of the vector; m2 = 2λa2 is the mass of the
Higgs boson):

Polarization vectors for helicity 0 are given in (26.78) for motion in the direction. Viewed in the center of
momentum frame, let the initial vector be traveling in the direction, and the final vector in the ′ direction, with
is the center of momentum scattering angle. Then

These obey the following relations:

The squares of the propagators’ momenta are

Using these relations, in the limit of large ω we have

The terms without any ω dependence are all (1). Adding the amplitudes,

As expected, the terms that grow with energy cancel, and the total amplitude is (1).

48
The Glashow–Salam–Weinberg Model I. A theory of leptons

Recall when I wrote down the weak interaction Lagrangian in the current-current form (40.1), I told you it was
nonrenormalizable: we couldn’t compute higher order corrections. In practice that didn’t matter for most
experiments, because the coupling constant is weak. Nature seemed to work in such a way that even the square
of the Fermi constant times infinity was effectively a small number; it’s very hard to find any conflict with
experiments. For many years, it was nevertheless a beau idéal of theoretical physicists to concoct a
renormalizable weak interaction theory. Finally, Glashow, Salam, and Weinberg did it, by constructing a gauge
field theory with spontaneous symmetry breaking.1 Because it was a gauge theory, with only renormalizable
interactions and small coupling constants, it was guaranteed to be renormalizable. As the dust of spontaneous
symmetry breaking settles, the interactions become very complicated, and it doesn’t look like a gauge field theory
at all. It’s got massive vector bosons, as well as a massless one that is identified with the photon. The whole thing
looks grotesque and disgustingly non-renormalizable, but that’s an illusion. Just as in our discussion of the sigma
model (§45.4), there are all sorts of secret relations among the coupling constants, which are preserved by
renormalization because it is secretly a symmetric theory. These secret relations guarantee that when you work
everything out, the theory remains renormalizable. That’s the importance of our earlier comment, that
renormalizability and spontaneous symmetry breaking are separable phenomena.2

48.1Putting the pieces together

The Glashow—Salam—Weinberg model (hereafter GSW model) is supposed to describe the real world, when
sufficiently generalized. There are many variants: the Georgi–Glashow model,3 the Pati–Salam model,4 there’s
this model and that model. The GSW model was the first one proposed, and it is still the simplest. These are all
models that are cooked up to yield a renormalizable theory of the weak interactions.

What would a model describing the real world have to include? For spontaneous symmetry breaking to occur
in perturbation theory, it has to have fundamental scalars. We don’t want any Goldstone bosons around at the
end, because they certainly aren’t there in the real world. So there will have to be gauge fields present to eat the
Goldstone bosons and become massive vector bosons; the only massless gauge field around is the photon. The
real world also has leptons and hadrons, and possibly quarks. And although we’re not going to expect perturbation
theory to offer much insight into the strong interactions, we’ll eventually have to extend the model to contain either
fundamental baryons and mesons or colored quarks. The first version of the model we’ll discuss will include only
scalars, gauge fields and leptons—for simplicity, only the electron and its neutrino. Later on we’ll see what
happens if we put in other leptons. It’s a very simple weak interaction theory, one in which there’s only an effective
current-current interaction between electrons and their neutrinos. (I’m leaving the muons out for the
moment—we’ll soon get to a theory that involves them.)

The first thing to decide on is the symmetry group of the theory. There will be a gauge group G we choose to
be U(2), that is to say, SU(2) plus phase transformations.5

This is very much like the isospin and hypercharge of the strong interactions. We don’t need to invent a new
terminology; we’ll just call these generators IW and YW, the weak isospin and weak hypercharge, respectively.
(These are not, of course, the familiar generators I and Y which occur in SU(3).) The weak charge is, by analogy
with the Gell-Mann–Nishijima relation,6

Because these symmetries break spontaneously, they don’t correspond to any manifest invariances of the real
world.7 There will also be an additional global U(1) symmetry, having nothing to do with gauge transformations,
which we’ll just impose on the Lagrangian as a phase transformation on the Fermi fields. The conserved charge
associated with this symmetry will be lepton number.8

Since we have a four-parameter gauge group (three from SU(2), one from U(1)) we will have four vector
bosons, one that we will call Vµ, corresponding to the weak hypercharge, and a family of three that we will call
Wµa, a = {1, 2, 3}, corresponding to the isospin generators. As these are two independent groups, they are allowed
independent gauge coupling constants.9 Following Weinberg we will call them g′ and g:

(the unconventional factor of will simplify later expressions). Once we have introduced the scalar field and Fermi
field content of the theory, the interactions of the vector bosons are completely determined: they follow the minimal
coupling principle. What are the scalar fields and what are the Fermi fields? There is only going to be one multiplet
of scalar fields ϕ. Its eigenvalues are

If this were the original I and Y we’d be describing the kaons. We’ll write the four real scalar fields {ϕ i}, i = 1, . . . , 4,
as a two-component, complex isospinor (I = ) ϕ. We will call these complex fields ϕ + and ϕ 0, just like the kaons
(K+ and K0):

This is an abuse of language since we don’t know what the electric charge is; the symmetry isn’t broken yet. The
scale of the generators is defined10 so that subsequent expressions are simple, once we write down the covariant
derivative of ϕ:

The in the Vµ term has to do with how we scale the generators τa of the weak isospin SU(2), the ordinary Pauli
matrices,

The matrix y is the generator of the Abelian weak hypercharge. As ϕ has Y = 1, y can be replaced here by the
identity matrix:

Since ϕ is a column vector, ϕ † is a row vector, with covariant derivative

(the derivative acting to the left). These are the only scalar fields in the model.

The most general Lagrangian invariant under the group G allowing for the possibility of spontaneous
symmetry breaking is

YM is the pure gauge field part, just the Abelian electrodynamic part for Vµ (that is, (26.47) with µ2 = 0) and the
standard form (the first term of (46.58)) for Wµa, the triplet. (D µϕ)† • D µϕ is the gauge invariant kinetic energy and
interaction. There can’t be any derivative interactions—they’re not renormalizable—but we’re allowed quartic and
quadratic non-derivative interactions between ϕ and ϕ †. Nor are there linear nor trilinear interactions, because you
can’t make a scalar with one isospinor or three isospinors. The only symmetric interaction is the one in square
brackets. We’ve summed things together in the conventional way giving us two parameters so that the symmetry
breaks spontaneously. (If the a2 term had the opposite sign, the Lagrangian would still be invariant, but it would
not lead to spontaneous symmetry breaking.) This is the most general renormalizable Lagrangian we can build
from these fields. The fermions are of course very important, but let’s take a preliminary look at what we have so
far.

We’re going to investigate this model in tree approximation, where we have the minimum value of ϕ. Because
ϕ is a two-component complex vector, at the minimum that sum of the four squares of the (real) fields must be a2:

With the full U(2) group at our disposal we can take any two-component vector and make it one of our basis
vectors. Which one we choose doesn’t matter; they’re all connected by the symmetries.

We will choose the symmetry breaking so that the expectation value of ϕ is

(with a real). The advantage of this is that áϕñ does not break electric charge conservation since it is ϕ 0 that
develops an expectation value. On the other hand, the other three of the four generators of the group are broken.
You can make a phase transformation along the “0” axis and one along the “a” axis and that’s all you can do.
Therefore we know already that we expect to find one massive scalar, and three Goldstone bosons which are
eaten by three of the four gauge bosons to make three massive vector bosons. We also know that two of these will
carry charges, plus and minus; they will be the isospin raising and lowering vector bosons, since electric charge
conservation is not violated. One of the massive vectors will be neutral; it will be some linear combination which
we have yet to compute, of the I3 vector boson and the hypercharge vector boson, since there are two electrically
neutral generators. The electric charge is the single remaining symmetry. This is supposed to be a realistic model,
and the only massless gauge boson we know about is the photon. The other three (massive) bosons will end up
being the intermediate vector bosons,11 the exchange of which simulates the current-current interaction. But we
haven’t gotten to that yet because we haven’t gotten to the leptons which source their currents.

We see already on this level that we have a model which at least meets the minimum criteria for a realistic
model of the weak and electromagnetic interactions: the symmetry breaking is such that there is only one
massless vector boson remaining. Notice that the way this model was cooked up is perfectly general. Once we
have stated the symmetry transformation properties of the fields and require that the interactions have to be
renormalizable, spontaneous symmetry breaking occurs in such a way that only one generator is unbroken.
There’s only one massless vector boson left at the end of the game.

For the fermions we do the same thing: We can have them transform under this group any way we like. Once
we stipulate their transformations, we can write down the most general interaction involving them. Then we will
examine the effects of spontaneous symmetry breaking on the fermions.

This is a theory which knows nothing about parity. When you first hear about spontaneous symmetry
breaking you might say “Oh, that’s marvelous. Parity non-conservation is going to arise as a consequence of
spontaneous symmetry breaking. That’s how we’re going to get parity non-conservation into the weak
interactions.” In fact, the GSW model goes exactly the other way. It says that the original dynamics which God
created before spontaneous symmetry breaking occurred is so ignorant of parity that it’s not written in terms of
Dirac four-component fields, but in terms of Weyl fields, two-component spinors.12 It’s not parity non-conservation
in the weak interactions that’s a result of dynamics; rather, it’s parity conservation in the electromagnetic
interactions.

Now let’s introduce the Fermi fields and their transformation properties. First I have to show you a little
notation to write left-handed and right-handed Weyl fields. For convenience, so we don’t have to go back to that
crazy σ notation and I can still use γµ’s, we’ll just take the four-component field and break it up into what we will
call left and right fields; these are the γ5 eigenstates.

Since (20.103) γ5 is anti-self-bar, the corresponding expression for ψ has a minus sign in it:

Of course

In a basis where γ5 is block diagonal,

then

where the dots (⋅) indicate some non-zero entries. Even though they are written as four-component spinors they
really have only two non-zero components; two are zero by the equations (48.13) that define them. A trivial
computation shows that
If we had only the kinetic energy term, the two helicity states would be dynamically independent. The mass term,
however, mixes them:

Next we define the (weak) isospin and hypercharge of the Fermi fields. We will have two, L and R, each
carrying lepton number. The field L is an isodoublet which is made up entirely of left-handed fields. Its
eigenvalues are

It’s like the {K0, K−} isodoublet.13 Its covariant derivative is:

The last term changes sign (as compared with (48.8)) because yL = −L. There is also a right-handed field R which
is an isosinglet. Its eigenvalues are

It’s a little peculiar, like the Ω−. Its covariant derivative is given by

There’s no − in the last term because yR = −2R.

The Lagrangian has the other terms as before, (48.10): the gauge invariant kinetic energy and Yukawa
couplings (the only renormalizable interaction the scalar fields and other fields can have). Notice that the
hypercharge of L minus the hypercharge of ϕ equals the hypercharge of R so we can have a hypercharge-
conserving Yukawa interaction by coupling L, ϕ, and R, with a real coupling constant, f:

where “h. c.” is the Hermitian conjugate. By a proper choice of phase we can always make f positive. If we hadn’t
chosen the hypercharges to allow an invariant Yukawa coupling, we would have gotten a rather trivial theory, as
we would have no interaction between the fermions and the scalar bosons. This is the most general
renormalizable Yukawa interaction. You might say “Couldn’t I put a γ5 in LϕR?” No, because R and L are γ5
eigenstates, so if we put in a γ5 that’s just putting in a factor of 1 or −1; it’s not an independent coupling. These
terms (48.24) are all there are. There are many free parameters, but we’ve written down every one. The full
Lagrangian is

What are the implications of this Lagrangian? First, we’re going to have electric charge left as a symmetry.
We can take L and R, break them up into components, and figure out what their electric charges are. The left-
handed fields L will have a negatively charged field in the bottom component, because

With malice aforethought we will call that field the left-handed electron, eL (more accurately, we are using only the
non-zero parts of the spinor, its two lower components, for eL). In the top component there will be a neutral field:

We’ll call this field (or more accurately, its non-zero components) νL, the left-handed electron neutrino. Then

(This is a four-component object.) The right-handed field R is an isosinglet, and it has charge

We’ll call that field the right-handed electron:

Let’s summarize the scalar fields {ϕ} and the left and right lepton fields {L, R} and their properties:

Table 48.1 The scalar and lepton fields’ properties in the GSW model

The result of spontaneous symmetry breaking is to give some particles masses. It will also tell us the
interactions with that remaining scalar boson. The Yukawa coupling gives the fermion masses; a gives the scale
of the breakdown.

And there it is, a mass term for the fermions, (48.19). We started out with these three massless Weyl fields that
had absolutely nothing to do with each other. One of them is a weak isodoublet, one of them is a weak isosinglet.
We write down the most general renormalizable interaction Lagrangian, we make the shift, and miraculously a
mass term appears! The mass of the electron is

The neutrino remains massless. We’ve done the most general case and the neutrino mass comes out to be zero.
We’ll always be left with one massless particle in a theory of this kind. We start out with an odd number of Weyl
fields. We can pair two of them together to make a mass term, but the third one is just left there. This is a
consequence of there being fewer right-handed fields than left-handed fields, so somebody has to be the odd man
out; we call him the neutrino. There’s no way we can give the neutrino a mass with this scheme.14 Of course, we
don’t know f and we don’t know a, so we can’t actually calculate the electron’s mass. But we’ve seen how the
electron gets a mass and the neutrino doesn’t; other fermions get a mass and their neutrinos don’t, by the same
automatic mechanism, no matter what the coupling constants are.

What can we say about the vector boson masses? Three of them are massive and one of them is massless.
The Lagrangian has a term

Expanding the covariant derivatives and shifting the fields ϕ → ϕ′ + áϕñ, we get a large number of terms, including
cross terms of the form (Ta is a generic generator)

However, since we are working in the U gauge, (46.72), all these cross terms vanish, leaving us with terms
involving only ϕ † and ϕ or only áϕ †ñ and áϕñ. The masses arise as a result of the shift when the ϕ and ϕ † are
replaced by their vacuum expectation values.

The ordinary derivative part of D µ will give nothing. The other part will give a term linear in the vector fields which,
when squared, will give the tree approximation masses. There are two kinds of terms obviously. There’s W1 and
W2 which involve τ1 and τ2 and turn the lower vector in (48.12) into an upper vector, which then gets squared.
And there are the two neutral ones, W3 and V which involve τ3 and the identity matrix, and turn the lower vector
into itself. Let’s write down those two terms separately.

From (48.6) and (48.9)

(with yϕ replaced by ϕ, because ϕ has y = 1). Then

That’s the vector boson mass matrix; it’s pretty easy to diagonalize. I’ll call the new fields Wµ± and Zµ:

The fields Wµ± describe charged vector bosons W± made from Wµ1 and Wµ2. They have the same mass:

The field Zµ describes a massive neutral vector boson Z0 with mass squared greater than that of the W±:

Finally there is a remaining orthogonal neutral vector boson

That orthogonal combination has no mass term. That’s reasonable: we have an unbroken symmetry, so we’ve got
to have a remaining massless vector boson, the photon:

The the expressions for the neutral vectors Aµ and Zµ we get

Three of the four real components of ϕ have been eaten by the gauge fields to give us a charged vector doublet
(of unknown mass, until we determine the parameters of the theory); a neutral massive vector boson, also of
unknown mass (except that it is guaranteed to be heavier); and a massless vector boson. The last part of ϕ,
corresponding to the real part of ϕ 0, remains. It is referred to in the literature as the Higgs boson.15

48.2The electron-neutrino weak interactions

How are the weak interactions described in this theory? Let’s look at the charged part of the current after we’ve
made the shift in the scalar field. That comes just from the τa in the covariant derivative (48.21), in the term Li L.
Those are the only charged terms. Here comes the real wonder. From D µ I have a , and writing W1 ∓ iW2 as
W±, I get a . Using (48.35),

because of the in the definition (48.13) of L. That’s the unique coupling of the charged vector bosons to the
fermions. Please notice: this automatically has the (V − A) form γµ(1 − γ5), because L = (1 − γ5)ψ. So the
interactions are automatically maximally parity violating. You might say “Ha, that’s nice but the weak interaction is
current times current.” Well, this Lagrangian leads to a four fermion interaction as a result of vector boson
exchange, a plus at one end and a minus on the other, as shown in Figure 48.1. In fact, this interaction looks a
great deal like Fermi’s theory (40.1), particularly at low momentum transfer. Recall that the W± are massive vector
bosons, and their propagators are

Figure 48.1: W-vector mediated four fermion interaction

What happens to this if we imagine the boson is very massive compared to the mass of the electron, so k is much
less than M: k ¿ M? Then kµkν/M2 is bubkes, M2 is a constant, and we simply get

That is, for small momentum transfer (small compared to the mass of the vector boson), I get effectively a point
coupling just like the Fermi coupling, with the identification

using the definition (48.40) for M2W±.

Please notice that the weakness of the weak interactions is revealed to be an illusion. The W vector coupling
constants in these theories are g and g′. We will shortly extract the coupling of the photon, and we will see that
both g and g′ are roughly the order of magnitude of the electromagnetic coupling constant, e. The smallness of the
Fermi constant has absolutely nothing to do with the presence of weak dimensionless parameters; the weakness
of the weak interactions is not due to weak couplings of these vector bosons. It is a consequence of the size of the
parameter a, a mass that entered the original Lagrangian: a is very large compared with the electron mass, which
sets the mass scale. Recall we found (48.32) the electron mass is the Yukawa coupling constant f times a. So the
weak interactions are weak not because of tiny dimensionless coupling constants like 10−5, but because there is a
superweak Yukawa coupling f that makes the electron mass much smaller than the characteristic mass scale, a,
the only parameter in the theory with the dimensions of mass. They’re weak due to the fact that the intermediate
vector bosons have large masses compared to the masses of the leptons, because the Yukawa couplings are
weak. It’s the weakness of the Yukawa couplings that makes the weak interactions look weaker than the
electromagnetic interactions.

48.3Electromagnetic interactions of the electron and neutrino

Instead of writing out the leptons explicitly we’ll simply write

where J3µ is the leptonic weak I3 current,

and Yµ is the leptonic weak hypercharge current,

We can substitute the formulas (48.44) for Wµ3 and Vµ into the Lagrangian and find the couplings of the two mass
eigenstates, Aµ and Zµ, to the leptons, at least in tree approximation, which is all we’re considering:

The combination that appears in the first term is nothing but the electromagnetic current, the third component of
weak isospin plus half the weak hypercharge, just as in the Gell-Mann–Nishijima relation:16

That is unsurprising. Electromagnetism has a manifest gauge symmetry so the massless particle, the photon,
must couple to the electromagnetic current. Electric charge conservation is not spontaneously broken. As a check,
we can calculate explicitly what Jµem is:
which is just what it should be.

The new information from the GSW theory is that its coupling constant e, what we normally call the electron’s
charge, is given in terms of g and g′ by

This equation can be made a bit more transparent by squaring and inverting:

The solutions of this equation can be parameterized in terms of an angle assuming, as is indeed the fact, that e is
a known quantity and that g and g′ are unknown quantities:

As promised, we see that g and g′ are indeed (e). The parameter θW is called the Weinberg angle.17 It was
introduced by Weinberg who, with commendable modesty, called it the weak interaction angle.18

When Weinberg and Salam first proposed this model, there were no observed weak neutral currents. All they
knew about were the Fermi constant and the electromagnetic charge. Aside from the quartic coupling constant λ
(which gives (46.15) the Higgs boson mass in terms of a), the one quantity not predicted in terms of known
quantities is the Weinberg angle. We can substitute this expression (48.57) for g into the formula (48.48) for the
mass of the W± to get

This formula means that it’s going to be a long time19 before anyone directly observes a W±, because it gives a
lower bound when θW is a multiple of π/2, and using a2 = /(4GF), and GF ≈ 10−5/mp2, this lower bound is

That’s a large number. Things are even worse with the neutral vector boson, the Z0. Just plugging into the formula
for the Z0 mass and doing a little algebra

There’s a 2 instead of a , and we get a lower bound whose square is twice as large as the lower bound of M2W±:

So it’s even harder to see the Z0 than the W±. Of course these two lower bounds cannot be attained
simultaneously; the first occurs when θW = π/2, and the second when θW = π/4.

When you go beyond tree approximation, the bounds change only by terms of order e2 or order λ2 and so
on—small corrections if all these parameters are small. That means, by the way we defined θW, that θW cannot be
close to an integer multiple of π/2; if it were, then g or g′ would be large. But so long as g and g′ are small enough
to justify perturbation theory, so long as all the dimensionless parameters of the theory—g, g′, λ and f—are much
less that 1, as they seem to be, this determines our scale of mass; it’s irrelevant how large a is. The tree
approximation should then be reliable, because it is simply the lowest order in perturbation theory. All the formulas
we are writing down will obtain corrections of higher powers of the various coupling constants. This is a
renormalizable theory that should be finite and computable and the corrections will be small. We will later worry
about a case where these coupling constants are all small but differ from each other by many orders of magnitude.
Then we have to worry about one-loop corrections in the large coupling constants affecting formulas that only
involve the small coupling constants in zeroth order; we will see that in more detail as we go on. They have
corrections which have been computed by people who wanted to check the renormalizability of these schemes. In
general, corrections are small provided that you don’t choose θW perversely, so that either g or g′ is enormous.
 Now let’s turn to the coupling of the Z boson. Because it is necessarily very heavy, the Z has got to have a
Fermi-type interaction, at least at acceptable energies. Let’s write out the coupling (48.52), in a form where we can
see what the interaction is. We’ll split this into an electromagnetic current and a remainder. If the parameters turn
out such that the Z boson is coupled only to the electromagnetic current then we get a short-range interaction
obeying exactly the same selection rules as for electromagnetism and very weak to boot. At low energies, we’ll
never be able to distinguish the short-range interaction from higher-order electromagnetic corrections. So it’s
important, if the effects of the Z boson are in any way observable, that it be coupled to something other than just
the electromagnetic current. This combination will be electromagnetism plus something else, and it will be the
amount of the something else that will tell us the observable effects. We have the formula for the electromagnetic
current, so we can eliminate Yµ in terms of Jµem and Jµ3. The result can be written in the following form (I will skip
a few lines of trivial algebra):

I’ve used the formula (48.41) for the mass of the Z to simplify the expression. Note also the identity

We see two things. First, while the term Jµem is parity-conserving, the term Jµ3 is maximally parity-violating; it
contributes only left-handed things. Therefore the parity-violating effects which would be the signature of the
presence of this object, which would be due to the cross terms between Jµem and Jµ3 in one-Z exchange, are
proportional to sin2θW. In this sense θW is very much like the Cabibbo angle.20

Putting the MZ in front is a good idea because you get a 1/MZ2 from the Z propagator in the exchange. You
get an effective Fermi-type interaction with, aside from Clebsch–Gordan factors, a Fermi-scale strength,
proportional to 1/a2, with the cross term between the two neutral currents. Therefore the theory inevitably predicts
a parity-violating neutral current-current type interaction of calculable magnitude at low energies, at least in this
very simplified model in which all you have in the world are electrons and their neutrinos.

48.4Adding in the other leptons

Let’s generalize the model to include the muons, the taus and their neutrinos. The obvious thing is just to put in
additional left-handed doublets and additional right-handed singlets:

However, saying things this way is a bit backwards. Presumably the particle structure should emerge as a
consequence of spontaneous symmetry breaking. I’ll start out with a model that involves three left-handed SU(2)
doublets Lα , α = {1, 2, 3} with exactly the same transformation properties under SU(2) ⊗ U(1), and three singlets
R α , right-handed Weyl fields:

Let’s ask the following interesting question: If we write the most general renormalizable Lagrangian consistent with
these fields, what will happen as a result of spontaneous symmetry breaking? Will it inevitably turn out to be a tau
and a massless neutrino, a muon and a massless neutrino and an electron and a massless neutrino, or is there a
possibility that other things could happen if we choose the coupling constant properly? The only undetermined
coupling constants are the Yukawa couplings (other than the coupling constants involving the ϕ fields, its self-
interaction and the gauge field couplings; these are completely determined by the stated transformation properties
and the minimal coupling prescription). Thus the new term in the Lagrangian is

where fαβ is in general a 3 × 3 matrix. If we were considering generalizations of this model with other, newly
discovered, kinds of leptons, then we would have to run α and β over a larger range. Everything else in the
Lagrangian is as before, completely determined.

We will demonstrate the following diagonalization theorem:21 Given any n × n matrix f we can always write f
in the following form.

U 1 and U 2 are unitary, Δ is diagonal and positive; it’s a diagonal matrix with only positive entries. This is a pure
matrix theorem; it does not depend on f being Hermitian or anything like that. We will first give the application of
this theorem to the GSW model and then prove it.

All of our left-handed doublets transform in exactly the same way under the gauge group, as do all of our right-
handed singlets. They have the same weak hypercharge and the same weak isospin. Therefore we are perfectly
free in a Lagrangian of this kind, without changing anything else, to redefine our doublets and our singlets by
unitary transformation, mixing them up any way we want (to keep the kinetic energy unchanged the transformation
must be unitary). In particular we can define

By the theorem, we can write f = U 1†ΔU 2, with Δ diagonal, as follows:

Therefore, in terms of these transformed fields our Lagrangian involves separate Yukawa couplings summed on
α:

That is to say, we can diagonalize the Yukawa couplings by independently shuffling around the right-handed fields
and the left-handed fields. We get a sum of Yukawa systems, decoupled into mass eigenstates, each of which has
exactly the same structure as the electron-neutrino system we have considered before.

The proof of the theorem goes as follows. Given any non-singular matrix f, f f† is a positive definite Hermitian
matrix:

H is the unique positive square root. Because H = H †, it can be diagonalized by a unitary matrix U 1:

Then

Define the matrix U 2 in the obvious way, from (48.67):

Showing that U 2 is unitary will complete the proof of the theorem. But this is easy:

On the right-hand side, there is a product U 1U 1† = 1 and a Δ−1 on either side of the Δ2. All the terms collapse,
and22

What is the significance of this result? No matter how we try to arrange the model, as long as it is consistent
with the constraints of renormalizability and gauge invariance, and as long as it doesn’t involve any fields other
than the ones we’ve itemized, we automatically get separate electron, muon and tau systems, as far as the
Yukawa coupling is concerned, independently coupled to the ϕ field. Thus we have
The lepton masses (in MeV) are23

The coupling constants fe, fµ and fτ are diagonal matrix elements of Δ, about which we can say nothing a priori.
Since the left-handed doublets transform in exactly the same way, all of the currents are exactly the same as
before: all the things that go into electromagnetism or the weak interactions or the Z-mediated interactions, are
sums of separate electron, muon and tau parts with identical coefficients as in (48.52). Because all of the algebra
is the same this is automatic.

48.5Summary and outlook

Let’s summarize what we have found in the GSW theory. In the course of the summary we’ll introduce some new
language that is frequently used in discussing these things. So far we have a theory only of leptons. We have not
yet put in the quarks. The theory has many good features:

• It provides a renormalizable theory of weak interactions (modulo the pesky question of anomalies,24 which
we will not discuss). That enables us to compute higher order weak corrections. Of course once we compute them,
we find they’re tiny. But they give the experimentalists a good excuse to ask for lots of money from their
governments so they can measure them. That’s a good thing.

• It unifies electromagnetism and the weak interactions, which is aesthetically very pleasing. These forces are
two aspects of the same force. It is not that we have two independent field theories; we have one. All the coupling
constants are of the same order of magnitude provided we don’t choose θW to be exceptionally small. It’s not true
that there’s a large vector boson coupling constant, for electromagnetism, and a small one for the weak
interactions. They’re all about the same size. It is spontaneous symmetry breaking that causes this completely
symmetric theory to put on a false beard and appear to be two grotesquely different things, electromagnetism and
the weak interactions.

• The theory exudes naturalness. When we have a spontaneously broken gauge theory, there are some
things that are generally true no matter what values we assign to the coupling constants, provided that we choose
values such that spontaneous symmetry breaking occurs. Features are said to be natural if they do not depend
on some perverse choice of coupling constants, but are true over a wide range of values. (I use the word “natural”
in a technical sense.) We write down what the fields are and how they transform under the gauge group; those are
the rules of the game. We write down the most general renormalizable interaction Lagrangian satisfying those
rules. Then we find certain general consequences that match experimental results, without having to fine tune the
coupling constants. This is a very pleasing theory. Here are six examples of naturalness in the GSW theory, good
features that emerge automatically, independent of the values of parameters:

1.Electromagnetism conserves parity, and

2.the mass of the photon is zero. It’s not possible to arrange the parameters in the theory so that the
interaction is renormalizable in any other way than what we have written down. The symmetry
always breaks down leaving electromagnetism with only a remaining U(1) symmetry unbroken.

3.The form of the weak interactions takes the (V − A) form of the Fermi interaction. This comes about
no matter what values the parameters have; we derived that without any assumptions about them.

4.Each of the leptons has a separately conserved lepton number. That is a consequence of the
diagonalization theorem. The only terms that can possibly mix the different leptons and their
neutrinos are off-diagonal terms in the Yukawa coupling. We did not require a diagonal Yukawa
matrix a priori; we allowed for the possibility of an arbitrary fαβ and then showed that we can always
choose the fields so that the off-diagonal terms disappear.

5.The leptons display universality in the Lagrangian: all currents are made up of electron, muon or tau
parts with exactly the same coefficients. No matter how we choose the initial parameters, after
spontaneous symmetry breakdown, at least to lowest order, all the leptons couple under the weak
interactions with the same coupling constants, and they have the same charges.
 6.The neutrinos are massless, no matter how we choose the parameters. (If we introduced an
electrically neutral right-handed field for the leptons before spontaneous symmetry breaking, we
could get a neutrino mass.)

There are two things as yet unexplained, which we would like to see addressed in the ultimate theory. They
are both associated with thus far unnatural features.

• Why is GF small? The reason the weak interactions look weak is because the W± and Z0 are heavy in
comparison to the other particles. As I pointed out earlier (see the paragraph following (48.48)), that is connected
to the reasons, as yet unknown, for the inequalities

The size of the Yukawa couplings determines the masses of the leptons in terms of the sole parameter with
dimensions of mass, a. Because we have the masses of the vector bosons MV ~ a, and GF ~ 1/a2, the size of a
has to be large, and so to obtain the observed lepton masses, the Yukawa coupling constants f must be very
small. Nobody knows why the f’s are so small.25 The theory could still be analyzed perturbatively if the ratios of the
lepton masses to the vector masses were on the order of 1/10; but then the weak interactions would not be so
weak. That is unnatural. We can certainly choose these Yukawa constants to be small, but we need not do so.
The theory doesn’t explain that.

• Why are the lepton masses so different? We have

We can arrange matters so that these conditions are met. We simply have to choose the Yukawa coupling
constant fe to be 100 or 150 times smaller than the Yukawa coupling constant fµ, but there is no reason why we
have to choose it that way. We could choose them to be equal and then we’d get a theory with equal lepton
masses. That isn’t the real world, but the theory doesn’t explain why it isn’t.

These are not serious shortcomings. It would be nice to have a theory in which the electron-muon mass ratio
was a computable quantity not associated with the ratio of free parameters.26 It would be very nice to have a
theory in which the weakness of the weak interactions was not merely consistent with the theory but inevitable in it.
We do not yet have such a theory.

Next time we will expand the GSW model to include quarks, and thereby all strongly interacting particles
made from them.

1[Eds.] S. L. Glashow, “Partial Symmetries of Weak Interactions”, Nucl. Phys. 22 (1961) 579–588; S. Weinberg, “A
Model of Leptons”, Phys. Rev. Lett. 19 (1967) 1264–1266; A. Salam, “Weak and Electromagnetic Interactions”, in
Elementary Particle Theory: Relativistic Groups and Analyticity. (Eighth Nobel Symposium), N. Svartholm, ed.,
Almqvist and Wiksell, Stockholm, 1968. See also Cheng & Li GT, Chapters 11 and 12, pp. 336–400.
2[Eds.] See §44.2, p. 970.
3[Eds.]
H. M. Georgi and S. L. Glashow, “Unity of All Elementary-Particle Forces”, Phys. Rev. Lett. 32 (1974)
438–441.
4[Eds.] J. Pati and A. Salam, “Lepton Number as the Fourth ‘Color’”, Phys. Rev. D10 (1974) 275–289.
5[Eds.] In 1990, Coleman added: “This comes from God. If you ask why, you will be fried by a lightning bolt.” In
fact, Gell-Mann and Lévy in 1960 already had the weak charge-changing current as inducing transitions between
members of an SU(2) doublet (private communication, Jonathan L. Rosner): M. Gell-Mann and M. Lévy, “The
Axial Vector Current in Beta Decay”, Nuovo Cim. 16 (1960) 705–726. Schwinger had earlier considered vectors
mediating the weak interactions as members of a family including the photon: Julian Schwinger, “A Theory of the
Fundamental Interactions”, Ann. Phys. 2 (1957) 407–434. Schwinger writes (p. 424): “The exceptional position of
the electromagnetic field in our scheme, and the formal suggestion that this field is the third component of a three-
dimensional isotopic vector, encourage an affirmative answer. We are thus led to the concept of a spin one family
of bosons, comprising the massless, neutral, photon and a pair of electrically charged particles that presumably
carry mass...” Glashow, a student of Schwinger’s, had taken up Schwinger’s idea of the weak interactions
mediated by massive vectors in his thesis (1959). In its appendix he states, “It is of little value to have a potentially
renormalizable theory of beta processes without the possibility of a renormalizable electrodynamics. We should
care to suggest that a fully acceptable theory of these interactions may only be achieved if they are treated
together.” Sheldon Lee Glashow, “The Vector Meson in Elementary Particle Decay”, thesis, Harvard University,
1959. In an article published the same year, Glashow extended the ideas of his thesis and considered the group
SU(2) ⊗ U(1): Sheldon L. Glashow, “The Renormalizability of Vector Meson Interactions”, Nucl. Phys. 10 (1959)
107–117; Crease & Mann SC, pp. 222–223.
6[Eds.] See note 10, p. 520, and note 21, p. 764.
7[Eds.] Henceforth we drop the subscript W.
8[Eds.] Griffiths EP, pp. 28–29.
9[Eds.] See p. 1023.
10[Eds.] The notation used here differs from that in the videotape of Lecture 52. Following Aitchison, we include
the hypercharge generator y (the generator is often omitted for Abelian gauge groups): I. J. R. Aitchison, An
Informal Introduction to Gauge Theories, Cambridge U. P., 1984, p. 108, equation (7.13). The editors have found
this practice helpful in avoiding (some) sign errors. Otherwise we use Weinberg’s original notation, as Coleman
did in later years teaching Physics 253b. Neither Coleman nor Weinberg wrote y explicitly.
11[Eds.] Cheng & Li GT, pp. 342–345.
12[Eds.] See §19.1.
13[Eds.] See Table 37.4, p. 806, and Figure 39.3, p. 852.
14[Eds.] In 1976, neutrinos were believed to be massless, but the 1988 discovery of neutrino oscillations requires
the neutrinos to have a non-zero mass. The current bound is mν < 2eV; PDG 2016, p. 758. The 2015 Nobel Prize
in Physics was awarded to Takaaki Kajita of the Super-Kamiokande Collaboration and Arthur B. McDonald of the
Sudbury Neutrino Observatory Collaboration for establishing that these oscillations occur. Various extensions of
the standard model have been proposed to incorporate massive neutrinos. See Vernon Barger, Danny Marfatia
and Kerry Whisnant, The Physics of Neutrinos, Princeton U. P., 2012, Chapter 9, “Model Building”, pp. 99–114.
15[Eds.] Many physicists independently considered the Goldstone model coupled to a massless vector, and found
the mechanism whereby the vector became massive and the Goldstone boson disappeared (see note 6, p. 1014).
Only Higgs predicted (1964) that there would be an observable massive scalar left over: “Broken Symmetries and
the Masses of Gauge Bosons”, Phys. Rev. Lett. 13 (1964) 508–509. Its properties were described in his
subsequent paper (1966): “Spontaneous Symmetry Breakdown without Massless Bosons”, Phys. Rev. 145 (1966)
1156–1163. Its discovery, confirming the mechanism, was announced at CERN on July 4, 2012. The current mass
of the scalar—the Higgs boson—is 125.09 ± 0.24 GeV: PDG 2016, p. 30. The 2013 Nobel Prize in Physics was
awarded to Peter Higgs and Fran¸cois Englert for their elucidation of the mechanism leading to the scalar’s
prediction.
16[Eds.] See note 10, p. 520, and (35.52)–(35.53).
17[Eds.] Cheng & Li GT, pp. 351–352. The current values are sin2θW ≈ 0.23129(5) or θW ≈ 28.746°: PDG 2016,
p. 119.
18[Eds.] In fact the idea of a mixing angle was introduced by Glashow: S. Glashow, “Partial Symmetries of Weak
Interactions”, Nuc. Phys. 22 (1961) 579–588 (the angle is introduced on p. 585); Crease & Mann SC, p. 226;
Close IP, p. 118, pp. 292–293.
19[Eds.] Coleman made this statement in 1976. The W± and Z0 were discovered at CERN in 1983; see note 9, p.
519. The current values of their masses are MW±: 80.385 ± 0.015 GeV; MZ0: 91.1876 ± 0.0026 GeV. See PDG
2016, p. 29.
20[Eds.] Peskin & Schroeder QFT, p. 605.
21[Eds.] The form (48.67) is called the singular-value decomposition. That f can be written in this form is a well-
known theorem, originating in differential geometry. See Gilbert Strang, Introduction to Linear Algebra, 5th ed.,
Wellesley-Cambridge Press, 1998, Chapter 7, pp. 364–400.
22[Eds.] In 1976, Coleman claimed that the theorem is true even when f is singular, though care has to be taken
because of possible zero eigenvalues of Δ.
23[Eds.] PDG 2016, p. 32.
24[Eds.] See note 21, p. 1044.
25[Eds.] In 1990, Coleman added: “Steve Weinberg says that the quarks would have masses similar to the W± and
Z0 were the Yukawa couplings comparable to the gauge couplings g and g′. The real question is: why are the
quarks we’re made of—the u and the d—so anomalously light? The top quark for example has a mass ~ 180
GeV.” The current value of mt is 173.21 ± 0.51 GeV; PDG 2016, p. 36.
26[Eds.] H. Sato, “Muon-Electron Mass Ratio and CP Violation as a Quantum Effect”, Nucl. Phys. B148 (1979)
433–444; K. Nishijima and H. Sato, “Higgs-Kibble Mechanism and the Electron-Muon Mass Ratio”, Prog. Theor.
Phys. 59 (1978) 571–578.
 49
The Glashow–Salam–Weinberg Model II. Adding quarks

Now we’re going to assimilate the strongly-interacting particles into our scheme. The rules of the game will be
exactly the same as before; we’ll just throw in some more fields. We have perfect freedom to choose these fields
as we wish, and specify how they transform under the gauge group and what the Yukawa couplings are. We will
choose them cunningly (or rather, we will employ other people’s cunning choices), so that after spontaneous
symmetry breaking the model begins to resemble the real world. We will construct a renormalizable theory that
unifies electromagnetism with the weak interactions. After we’re done, we will see which properties of the semi-
leptonic weak interactions emerge naturally and which do not.

49.1A simplified quark model

We will start with a simplified quark model in which there are no strange particles. In this case we don’t need the
strange quark, s, and we can get by with just two quarks, the up quark, u, and the down quark, d. Instead of SU(3),
we just have SU(2), isotopic spin. The up and down quarks have charges Q = and Q = − respectively:

Table 49.1: Quarks and their charges

We will build our hadrons out of these two quarks. Quarks also carry a color index (§39.6), having to do with
color SU(3) that couples to the gluons, but we’re not going to have to worry about that. Color factors out of this part
of the analysis. We simply want the quark form of the currents. Those gluons are there, but they and the SU(2) ⊗
U(1) gauge group don’t talk to each other, by assumption. This means that the formulas we get for quark masses,
etc. will have large corrections due to the strong interactions between the quarks, even though the formulas we get
for the currents will not be affected.1

By exactly the same trick we used with the leptons (see (48.28) and (48.30)), we can build a left-handed
doublet and two right-handed singlets out of the quark fields:

(The subscript on L1 is to distinguish it from the leptonic doublet L, and also from a second left-handed quark
doublet we are shortly going to introduce.) L1 is a weak isodoublet, the R’s are two weak isosinglets.2 Their weak
hypercharges are determined (48.2) by their charge assignments.3 The only thing that is going to be different is
the Yukawa couplings; everything else is exactly the same in its transformation properties.

Table 49.2: The up and down quark fields divided into left and right fields

We can make two invariant Yukawa couplings. We can have the original coupling

That’s down-right simple. It’s perfectly consistent with isospin and hypercharge conservation. (Isospin is obvious.
To check hypercharge, we need only check the electric charge where the ϕ 0 couples dL and dR which is obviously
charge conserving.) However, there is another possible Yukawa coupling, because we can put in two right-handed
quarks. Recalling the definition of ϕ,

we can introduce the charge conjugate4 field, ϕ C:

The vacuum expectation values of ϕ and ϕ C are

Because uR has been assigned exactly the right transformation properties, we can put together a hypercharge
invariant coupling of the following form (see (48.24)):

When spontaneous symmetry breaking occurs the first term will give a mass to the down quark and the second
term will give a mass to the up quark:

This extra term does not occur in the purely leptonic sector because that would require the presence of a right-
handed field carrying the same charges as the upper element of L, (48.28). The upper element is a neutrino and
there is no right-handed field carrying neutrino charge or with the proper assignment of weak hypercharge.5

Everything here goes just as before: parity-conserving electromagnetism, Fermi theory of the weak
interactions, universality, because everything transforms the same way under the gauge group. So we just get the
sum of a quark term plus a non-quark term. We still have three massless neutrinos, now joined by two massive
quarks. We get independent conservation of individual lepton number and quark (i.e., baryon) number. Here we
don’t have to use the diagonalization theorem because of the way we have defined the left and right fields and
arranged the coupling. There is no possible way of writing a hypercharge-invariant quark-lepton Yukawa coupling,
because the quarks carry fractional weak hypercharge (Table 49.2), while the leptons carry integral weak
hypercharge, as do the ϕ’s (Table 48.1). There is no way of adding two integers to make a fraction. To say it
plainly, we have independent conservation of quark number (or, if you prefer, baryon number), and of lepton
number.

What we don’t yet have are the strange particles.6 There is something else in this theory that is grossly
unnatural: the approximate conservation of isotopic spin (strong, not weak). That depends on f1 and f2 being
approximately equal. There is no symmetry principle in this theory that would require f1 and f2 to be approximately
equal; it is just a coincidence that they are. As far as I know, no one has found a model with all these other
satisfactory features in which the approximate conservation of isotopic spin is a natural result. It’s like the
weakness of the weak interactions. We can make isotopic spin approximately conserved. We just have to do it “by
hand”, and choose f1 approximately equal to f2. But there is no symmetry principle forcing us to do so.

49.2Charm and the GIM mechanism

Onward! Now we come to the bright idea7 of Glashow, Iliopoulos and Maiani, the GIM mechanism. It wasn’t
phrased in the context of this form of the theory, but in quite another, where it was much less obvious what the
right thing to do was—namely, suppressing strangeness—changing neutral currents. The bright idea is this. The
reason we keep getting universality is because everything is a left-handed doublet. The charged fields Wµ± couple
to purely left-handed currents. So everything couples universally because we’ve simply got the same damn Pauli
matrix all the time. If we’re going to have some form of universality after we introduce strange quarks, then they’ve
got to be put into a left-handed doublet also. Unfortunately if we only have three flavors of quarks, there’s no way
we can make two doublets. We can make a doublet and a singlet, but that’s the end of the game. To carry this
scheme on in any natural (in the vernacular sense) way, we’re going to have to have an even number of quarks.
The smallest even number greater than three is four, so we’ll need (at least) four quarks. The new quark is called
charm.8 This new quark has a charge of + , and a new quantum number, charm, with C = 1, just as the strange
quark has S = −1.

We will arrange the two quarks {c, s} exactly the same as {u, d}, as shown in Table 49.3 (compare Table
49.2). What results in two left-handed weak isodoublets, with the upper field having Q = + and the lower having Q
= − , and four right-handed weak isosinglets, two with Q = + and two with Q = − . Then we have all sorts of
possible Yukawa couplings. It’s now not a 2 × 2 matrix but a 4 × 4 matrix, since there are four things on the right
and four things on the left of the Yukawa coupling. Rather than choose the Yukawa couplings and see what mass
eigenstates result, it’s perhaps better to tackle the problem in reverse: we’ll choose the mass eigenstates, and see
what Yukawa couplings, and hence what doublets, come out of those choices. (The singlets we can mix up as we
wish; they’ll have to be chosen to be the mass eigenstates.)

Table 49.3: The strange and charm quark fields divided into left and right fields

We’ll choose our mass eigenstates to consist of a down quark and a strange quark, which have Q = − , and
an up quark and a charmed quark, which have Q = . They are determined only up to a phase, and we’ll take
advantage of that freedom.

These are some linear orthogonal combinations of the original entries in our left-handed doublets. The problem is
to determine what orthogonal combinations they are or, equivalently, how the doublets are made out of the mass
eigenstates, the inverse of the transformation we did before for the case of the leptons.

Things are pretty much constrained. We have two identical doublets, which we can mix up as we wish by a
unitary transformation. We can always choose the first to have for its upper spot the left-handed up quark, uL. If
we pick them randomly, one will have some linear combination of uL and the left-handed charm quark, cL, and the
other will have another combination. Then we’ll form a mixture so one is pure uL. In the lower spot we must have
some combination with norm 1 of the two possible things that can go in the lower spot, the left-handed down quark
field dL and the left-handed strange quark field, sL. We’ll have some phase times the cosine of some angle (not
θW; an independent angle) times dL plus some other phase times the sine of that angle times sL:

That’s the most general thing we can build with charge − and norm 1. We can absorb the phases eiδ1 and eiδ2
into the dL and the sL. So we can always choose one of our left-handed doublets to look like this:

Now this is obtained by diagonalizing a Hermitian mass matrix, so the other doublet must be orthogonal to it. In the
top slot we must have some phase and the only orthogonal isodoublet, charmed left, cL. In the bottom slot we
must have the vector that is orthogonal to the vector in (49.9), so

This is forced on us by orthogonality. We can always choose the phase of cL so that eiγ is the same as eiδ,
upstairs and downstairs, because we haven’t talked at all about the phase of cL. Once we do that, we have a
common phase factor, and we can send them both to 1 simply by changing the phase of the doublet L2. That
doesn’t change its gauge transformation properties or anything else. The upshot of this is,9 that there are no
phases that correspond to any physically observable quantities. The only unknown thing we have in the result is
this angle θ. That’s the result of putting in perfectly arbitrary Yukawa couplings consistent with the symmetries of
the theory. We’ve just systematically pushed out arbitrary phases that are just matters of convention.10 This angle
θ is nothing but (40.16) the Cabibbo angle, θC.11 Remember, the charged weak currents always involve Pauli τ
matrices. They are left-handed currents which take the upper part of one of these doublets and mix it up with the
lower part of one of these doublets. We haven’t needed the charmed quark in our low-energy phenomenology, so
it must be very heavy, not present in any of the observed particles.12 So we don’t have to worry about L2, which
will involve the so-called charmed current. For instance, from L1τ1γµL1, we get both a strangeness-conserving
current with magnitude cos θC,

and a strangeness-changing current with amplitude sin θC,

Their coefficients are cos θC and sin θC, exactly as predicted by the Cabibbo theory (40.16). The extension to
include four quarks should be clear. In addition to the terms in (48.25), there will be the additional kinetic terms Lα
Lα and R α R α , and perhaps additional Yukawa couplings of the form Lα ϕR α , where α = {1, 2}.

What features does this model have, and are they natural? The following features in this theory are natural:

•The Fermi theory with Cabibbo-expressed universality: cos θC times the strangeness-conserving
current, sin θC times the strangeness-changing current, with only left-handed (V − A) currents.

•Discounting the weak interactions, the electromagnetic interactions conserve all the independent
quark numbers, once we’ve defined them in terms of these mass eigenstates: four independent
quantities which count the kinds of quarks, or simply the four currents {uγµu, dγµd, γµs, γµc}; the z-
component of isospin, Iz ; and hypercharge, Y.

•There is no ΔY ≠ 0 neutral current. The Z0 boson does not contribute to strangeness-changing

decays. This is very important if the theory is to match experiment. If the Z0 did contribute to ΔS ≠ 0
decays, we would instantly get the decay process

which no one has observed. Why is this ruled out? If we look at the neutral currents, they act on fields
either in the upper or the lower entry of L2. With cL there is no problem, you have charm with charm
and nothing happens. In the bottom of L2 we could get a cross term cos θC × (−sin θC). Remember
they’re universal, they’re always added together from all the doublets. In the bottom part of L1 we get
the same term with the opposite sign. So these two terms cancel automatically, i.e., naturally. You
don’t have to fudge the parameters. Automatically the cross term in the bottom of L1 is canceled by the
cross term in the bottom of L2 when we construct the currents. There is no strangeness-changing
neutral current.13

Aside from continual fights with whether it agrees with detailed experiments, the model fails to explain two
results naturally, and that makes this model slightly unsatisfactory:

•The mass of the up quark and the mass of the down quark are equal to (e2):

One would expect that to be so in order that we could be deluded into believing that all isotopic spin
violation is electromagnetic, the standard dogma.14 This is equivalent to saying that the appropriate
Yukawa coupling constants are equal up to terms of order e2. That is not natural. We can choose them
to be equal up to (e2), but there is no reason why they should be approximately equal. That is
the big unnatural feature of this model. The riddle is: why is isotopic spin approximately good?

•CP violation is missing. In this theory, CP symmetry is unbroken. That’s the bad thing about our
wonderful ability to eliminate all of these arbitrary phases by choosing our conventions properly. If
we had some phases left around that we could not eliminate, we might have a chance of a CP-
violating current. We don’t have one in this model. We have written down all the renormalizable
 interactions there are, and none gives CP violation after the symmetry spontaneously breaks. You
couldn’t find a more CP-conserving model than the GSW theory.15

There are a thousand and one models, variations on the themes of the GSW model. The one I have
described here is known as the standard model.16 This is the best of a bad lot, in that most of the desirable things
are natural, and the fewest desirable things are unnatural. Once you see how it is done, you too can construct a
model. You just fiddle around putting in a bunch of left-handed quarks and right-handed quarks to make a larger or
smaller gauge group, you throw in lots of unobserved particles, let the machine rip and deduce what happens.
Most of these models either involve huge numbers of unobserved particles or involve some things that are now
natural coming out as unnatural. For instance, the Georgi–Glashow model 17 has no neutral currents in it, which
was thought to be an advantage at one time. Not anymore. But it had the unfortunate feature that the
masslessness of the neutrino was unnatural. Certainly e − µ universality was unnatural. It required genius to invent
this game but unfortunately it only requires persistence to continue playing it indefinitely. The literature is
chockablock with models, for example there’s the Pati–Salam model.18 You have to be a real expert to be familiar
with them all. To the extent that the experiments verify any model, however, they seem to support the original.
That may change, of course, with a new generation of accelerators and physicists, or if someone playing this
game comes up with an idea no one has thought of before. Perhaps someone will introduce a little extra twist, just
as spontaneous symmetry breaking and the Higgs phenomenon by themselves were little extra twists, before they
were incorporated into a new theory. But for the moment this model is the standard. 4f

49.3Lower bounds on scalar boson masses

I want to show you at least one nontrivial calculation that involves a higher-order correction. This is the simplest I
know. It uses the effective potential. Originally the calculation was carried out in the GSW model. But the algebra
is complex, so I will discuss it in a simplified model. Trust me when I say that similar arguments, keeping track of
all of the coupling constants, can be made in the full model.

In 1976 Steve Weinberg worked out the dynamics for his model and got a lower bound19 on the mass of the
single remaining scalar boson, the Higgs boson, corresponding to the real part of what we called the ϕ 0 field.20 I
will discuss it in a much simpler model where the essential physics is the same. The model is our old friend, the
purely Abelian Higgs model: one gauge field, the photon, plus two real fields that form an SO(2) doublet:

We worked out the covariant derivatives some time ago:

From (46.2)

When spontaneous symmetry breaking occurs in the tree approximation we get a vector boson of mass

and a scalar boson of mass

(If you don’t believe me, differentiate the λ term twice about ϕ 1 = a, ϕ 2 = 0.) We can’t say what’s a big mass and
what’s a small mass. That depends on what our scale is. But we can talk about the ratio

It looks like mS could be as large or as small as we want. We can keep λ and e2 small, so perturbation theory is
good, and make λ either much larger or much smaller than e2. A similar remark can be applied in the case of
practical interest, the GSW model, where perhaps we can hope to calculate this ratio. We’d be interested in
knowing how large the scalar’s mass is.21
 Now I will demonstrate that in fact this is wrong, using only the physics we have: that there is a lower bound
on the ratio

The critical point is this: if

then U, the scalar potential, is much less than the contribution from the gauge field loops:22

Therefore we have no right to compute the mass just in tree approximation, without including the effects of the
gauge field loops. Their effects, as we will see shortly, are (e4), but if λ ¿ e4, which is possible if λ and e are both
small, we have no right to neglect them. They are the first terms involving e that appear in the spontaneous
symmetry breaking problem. Therefore we will approximate the effective potential to investigate the situation,
always assuming (49.18). That’s the region we’re interested in. From (47.36),

U CT is the finite part, the counterterms. I won’t bother about the higher-order correction involving loops with scalar
bosons running around them, because those are higher powers in λ, which is supposed to be negligible and we’ve
already included the leading term of λ, in U.

Let’s investigate this object, V. I’ll first write down the gauge field part, (47.44):

The trace is trivial because there is only one vector boson in the game. It is convenient to define a parameter ρ,
the “length” of ϕ 2, by

and to simply write µ2(ϕ) as

the tree approximation mass. The gauge field part is

The rest of the argument is just algebra. The combination of U and the counterterms will come together and give
us some coefficient which we’ll figure out later. It’s going to be some function of λ determined by our
renormalization, but at this moment we don’t care what it is. So U + U CT will be some coefficient α times ϕ 2, plus
some other coefficient A times ϕ 4 plus a possible constant, and V will be all this, plus (49.24). We’ll just subtract
the constant, that’s not going to make a difference, and obtain

We’ve used the identity

Because ρ4 ln e2 is proportional to ρ4, we’ve just included that term in β.

Now we have to determine these constants α and β. We could go through renormalization systematically and
determine them in terms of the coupling constants and renormalization conventions. But we might as well get
them directly in terms of quantities we want, and avoid a lot of algebra. I will impose two conditions. First, there is
going to be spontaneous symmetry breaking, so (44.39) V(ρ) has a minimum:
That value of ρ determines the scale of mass. At the end, we’re just going to be computing a dimensionless ratio.
To avoid a lot of complicated algebra, I’ll simply choose ρ = 1; that sets the mass scale:

That will eliminate one of the two unknown quantities α and β. The second condition will be determined by the
statement that at the minimum ρ = 1, the second derivative of V gives the mass of the scalar boson:

More precisely, this gives the inverse of the scalar boson propagator at zero momentum transfer. But that’s
equivalent to the scalar boson mass, except for higher-order corrections, of order e4 and so on. These are the
equations that will determine α and β in terms of the quantities of interest. Since the symmetry breaking occurs at
ρ = 1, the vector boson mass in these units is simply

(again, plus higher-order corrections that we’re not interested in; we’re just looking at the leading terms).

Starting from (49.25), the first step is to consider

We can automatically eliminate one of the coupling constants, β, from (49.27) at ρ = 1:

This determines the ρ4 coefficient and we write

We’ve split the ρ4 term up into two parts so that they individually have vanishing derivatives at ρ = 1. There
remains one unknown constant α, which we will determine in terms of the scalar boson mass, by differentiating
twice.

Differentiating the first term twice at ρ = 1 is no hard job. We get 2 from the first term, −6 from the second
term, so that gives −4α. In the other term the ρ3 terms cancel out in the first derivative. The only non-zero term will
come when we differentiate the logarithm:

At first glance it looks as though we could make the mass anything we want, by an appropriate choice of α.
However, to make the mass go to zero we would have to choose α to be

But what is α? Let’s look back at V in (49.33). All the terms except the first have vanishing second derivatives at
the origin:

Therefore if we choose the mass very small, that is if α is positive, we have a potential that is concave upward at
the origin and also at the minimum we are exploring, near ρ = 1: there is a stable point at the origin. It doesn’t look
like the tree approximation (Figure 43.4), but nevertheless that’s what we’ve got. It looks like Figure 49.1. In that
case we have to worry: does spontaneous symmetry breaking occur? It does, if the minimum at ρ = 1 is less than
the minimum at the origin, which then is a false vacuum. On the other hand if the minimum at ρ = 1 is greater than
the minimum at the origin, then ρ = 1 corresponds to a phony vacuum. In that case we would be exploring the
second derivative of the potential at a place that has absolutely nothing to do with real physics, because in fact the
theory does not experience spontaneous symmetry breaking. So we have a criterion for spontaneous symmetry
breaking:

Figure 49.1: The false vacuum at ρ = 0

What is V(1)? That’s fairly easy to calculate:

Since we require V(1) ≤ 0, we now have an upper bound on α, which means the mass can’t get too small:

This generates (49.34) a lower bound on the scalar boson mass:

Or, writing things in terms of dimensionless quantities (using (49.30)) so our peculiar scale conventions are not
relevant:

Therefore you cannot make the scalar mass over the vector mass ratio as small as you please. It is obvious that
similar reasoning will work in the Weinberg model. We get functions of exactly the same form. It’s just that
everything will be much more complicated because we’ve got a lot of gauge bosons to use in computing the
effective potential, and therefore there’s a lot more algebra involving sin θW and cos θW. But the physics is
identical.23 In the Weinberg model, it gives an absolute lower bound on the Higgs mass on the order of 3.72 GeV.

The next lecture will be the last. In response to popular request, it will be on the renormalization group and its
uses, and its connection with non-Abelian gauge field theory. It will be somewhat different in character from the
other lectures, because I cannot cover the whole subject in ninety minutes. It will be structured more like a
colloquium. I will ask you to take certain things on trust and show you other things in detail.

1[Eds.] Cheng & Li GT, Chapters 11 and 12, pp. 336–400.

2[Eds.] To distinguish between weak isospin and the old, strong isospin, the latter will be described in this chapter
as “isotopic spin”.
3[Eds.] In parallel with the strong Gell-Mann–Nishijima relation linking isotopic spin, the usual hypercharge, and
charge, Q = Iz + Y. See note 10, p. 520, and (35.52). At this point in 1976, Coleman added, “This goes to show
that God has less imagination than the high energy theorists, who have thought of many possibilities more
baroque than this.”
4[Eds.] See p. 119.
5[Eds.] As always, Coleman is assuming the neutrinos are massless. See note 14, p. 1006.
6[Eds.]F. Halzen and A. D. Martin, Quarks and Leptons, John Wiley, 1984, Section 1.7, pp. 26–27; Section 2.9,
pp. 44–46.
7[Eds.] S. Glashow, J. Iliopoulos and L. Maiani, “Weak Interactions with Lepton-Hadron Symmetry”, Phys. Rev. D2
(1970) 1285–1291; L. Maiani, “The GIM Mechanism: Origin, Predictions and Recent Uses”, in Rencontre de
Moriond: Electroweak Interactions and Unified Theories, La Thuile, Valle d’Aosta, Italy, 2013, avaliable on-line at
https://arxiv.org/abs/1303.6154.
8[Eds.] “Aesthetic arguments led J. D. Bjorken and me to conjecture a fourth quark, more than a decade ago.
Since leptons and quarks are most fundamental, and since there are four kinds of leptons, should there not also
be four kinds of quarks? We called our construct the charmed quark, for we were fascinated and pleased by the
symmetry it brought to the sub-nuclear world. The case for charm—or the fourth quark—became much firmer
when it was realized that there was a serious flaw in the familiar three-quark theory, which predicted that strange
particles would sometimes decay in ways that they did not. In an almost magical way, the existence of the
charmed quark prohibits these unwanted and unseen decays, and brings the theory into agreement with
experiment. Thus did my recent collaborators John Iliopoulos, Luciano Maiani, and I justify another definition of
charm, as a magical device to avert evil.” Sheldon L. Glashow, “The Hunting of the Quark”, The New York Times
Magazine, July 18, 1976, pp. 154, 159, 161; reprinted in The Charm of Physics, Sheldon L. Glashow, Copernicus
Books, 1991; B. J. Bjørken and S. L. Glashow, “Elementary Particles and SU(4)”, Phys. Lett. 11 (1964) 255–257;
Glashow, Iliopoulos, and Maiani, op. cit. Note the way Bjorken signed the article, in “disguise”, due to the
whimsical character of the proposal; Crease & Mann SC, p. 291.
9[Eds.] In Glashow, Iliopoulos, and Maiani op. cit., p. 1287, the hadronic weak current is written as (their equation
(3))

where q is the quark column vector (c, u, d, s) (the authors use ( ′, , , λ), respectively). The matrix C H
must have the form (their equation (4))

where is the 2 × 2 zero matrix, and U is a 2 × 2 matrix, if JµH is to carry unit charge. After asserting that “The
strong-interaction Lagrangian is supposed to be invariant under chiral SU(4), except for a symmetry-breaking term
transforming, like the quark masses, according to the (4, 4) ⊕ (4, 4) representation. This term may always be put
into real diagonal form by a transformation of SU(4) ⊗ SU(4), so that [baryon number], Q, Y, C and parity are
necessarily conserved by these strong interactions,” the authors state, “Nevertheless, suitable redefinitions of the
relative phases of the quarks may be performed in order to make U real and orthogonal...” If, however, the two
families of quarks are joined by a third generation, as was pointed out by Kobayashi and Maskawa, CP-violation
can be incorporated into a theory of quarks: Makoto Kobayashi and Toshihide Maskawa, “CP-Violation in the
Renormalizable Theory of Weak Interaction”, Prog. Theo. Phys. 49 (1973) 652–657. This idea is called the CKM
mechanism (after their initials, and Cabibbo’s). Two more quarks, {b, t}, “bottom” and “top”—a third
generation—were observed in 1977 and 1995, respectively, at Fermilab. Kobayashi and Maskawa were awarded
half of the 2008 Nobel Prize in Physics for their explanation of CP-violation via a mixing of the three generations of
quarks; the mixing matrix is called the CKM matrix. (The other half went to Nambu for spontaneous symmetry
breaking.) The new quarks each have a new quantum number: bottom has B, equal to −1, and the top has T = +1.
The generalized Gell-Mann–Nishijima relation is (Iz corresponds to strong isotopic spin)

where the baryon number = for all quarks (− for antiquarks); PDG 2016, p. 279, equation (15.1). In the video
of Lecture 53, Coleman mentioned a theorem by Maiani, probably referring to the results reported in L. Maiani,
“CP Violation in Purely Lefthanded Weak Interactions”, Phys. Lett. 62B (1976) 183–186, where the
Kobayashi–Maskawa model of three doublets is analyzed. There it is shown that three mixed doublets reduce to
two mixed doublets and an unmixed one, and hence no CP-violation results, when one real angle vanishes, or
two quarks of the same charge are degenerate in mass (L. Maiani, private communication).
10[Eds.]
Following Halzen and Martin, op. cit., p. 283, equation (12.110), the doublets L1 and L2 can be written
compactly as

where the weak (primed) eigenstates are related to the mass or physical (unprimed) eigenstates by

(This gives the same result as in the GIM paper, though their U appears different, due to a different ordering of the
four quark fields.) The weak currents are then of the form
11[Eds.] See note 17, p. 882.
12[Eds.] “Observed particles” as of May, 1976, when this lecture was given. The current estimate of mc is 1.27 ±
0.03GeV; PDG 2016, p. 36. The meson was found in November, 1974, by simultaneous discoveries at
Brookhaven National Lab (headed by Sam Ting, MIT; the resonance was called “J”) and SLAC (headed by Burton
Richter; “ψ”); the meson is denoted J/ψ today. These two men shared the 1976 Nobel Prize for its discovery. The
J/ψ’s current mass is 3096.900 ± 0.006 MeV, PDG 2016, p. 1371. The various states of are collectively
called “charmonium”.
13[Eds.] Greiner & Müller GTWI, p. 230, footnote 15, echo Glashow’s second meaning of “charm”, stating that
charm equates to magic, since it helps remove the unwanted currents.
14[Eds.] Forty years ago, masses within an isotopic multiplet were believed to be equal to within a few percent,
i.e., to within (e2): all mass splitting was thought to be due to electromagnetic effects. This is no longer the case;
strong interactions appear to be responsible for much of the difference, and isotopic spin invariance is now
regarded as only approximately exact: Griffiths EP, pp. 135–136, and footnote on p. 135. The current values are:
mu = 2.2−0.4+0.6 MeV, md = 4.7+0.5−0.4 MeV; PDG 2016, p. 36. Lattice QCD calculations agree very well with these
numbers; “Precise Charm to Strange Mass Ratio and Light Quark Masses from Full Lattice QCD”, C. T. H. Davies,
C. McNeile, K. Y. Wong, E. Follana, R. Horgan, K. Hornbostel, G. P. Lepage, J. Shigemitsu, and H. Trottier, Phys.
Rev. Lett. 104 (2010) 132003 find, at an energy of 2 GeV, mu = 2.01(14) MeV and md = 4.79(16) MeV.
15[Eds.] D. Chang, X-G. He, and B. McKellar, “Ruling out the Weinberg Model of Spontaneous CP Violation”,
Phys. Rev. D63 (2001) 096005. But see note 9, p. 1081 for the CKM mechanism, which offers a way to explain CP
violation with a third generation of quarks, a new pair observed only in 1977 and 1995, respectively.
16[Eds.] Griffiths EP, pp. 49–52. Today, “the standard model” usually means the GSW model of the electroweak
forces, plus quantum chromodynamics with three generations of quarks {u, d; c, s; t, b} and leptons {e, νe; µ, νµ; τ,
ντ}.
17[Eds.]
Howard Georgi and S. L. Glashow, “Unity of All Elementary-Particle Forces”, Phys. Rev. Lett. 32 (1974)
438–441.
18[Eds.] Jogesh C. Pati and Abdus Salam, “Lepton Number as the Fourth ‘color’”, Phys. Rev. D10 (1974)
275–289.
19[Eds.] Steven Weinberg, “Mass of the Higgs Boson”, Phys. Rev. Lett. 36 (1976) 294–296.
20[Eds.] Peskin & Schroeder QFT, pp. 715–717.
21[Eds.] In 1976, Coleman said, “If its mass could be something like 10 eV, for example, it might be worth looking
for.” He probably meant “10 GeV”. In 1990, he said that “mH can’t be too small, or we would have seen it; it can’t
be too large, or we’d have no right to do perturbation theory. We expect to see it at the SSC.” In the anonymous
graduate student’s notes, Coleman has penciled in that mH > 58 GeV. Alas, the American large accelerator, the
Superconducting Super Collider, was canceled in 1993. As the world knows, we did see the Higgs, at CERN’s
Large Hadron Collider in 2012.
22[Eds.] See Figure 47.9, p. 1047.
23[Eds.] Weinberg op. cit., note 19, p. 1085. The bound Weinberg obtained is (see his equation (8))

Weinberg also considered a theory in which all bare masses are zero, and obtained a lower bound of about 7 GeV.
The current value of the Higgs mass is 125.09 ± 0.24 GeV, well above Weinberg’s lower bounds; PDG 2016, p.
30.

50
The Renormalization Group
We’re going to take up a totally new subject and dispose of it in the course of a single lecture. The subject goes
under the name of the renormalization group, often abbreviated as RG.1 As you will see, this is a pretentious
name for a rather simple set of ideas. It’s an old idea that must have occurred to physicists many, many times: if
you do experiments at sufficiently high energies, in some sense of the word energy, the masses of the elementary
particles should be irrelevant. The questions are: is this true, and in what sense do we mean high energy?
Obviously we don’t mean it in the naїve sense. We don’t mean something like total cross-section, because total
cross-sections involve the imaginary parts of forward scattering amplitudes.2 No matter how high s is, t is fixed at
zero for forward scattering amplitudes,3 and therefore we would expect at least the masses of the particles
exchanged in the cross channel—the t channel—to be of critical importance, no matter how high the energy.

50.1The renormalization group for ϕ4 theory

Let’s take a definite theory and make this idea a bit more precise. For simplicity I’ll use our old friend, ϕ 4 theory
with bare mass µ:

I’ll call the coupling constant g instead of λ. Let’s try to find a region in which we have a better chance of masses
becoming unimportant at high energies than we would, for example, for fixed-t scattering processes or total cross-
sections. I’ll pick a thoroughly unphysical region.

I’ll consider an n-point function of n momenta, as defined in (32.14):

The (n) are 1PI functions, although we could just as well worry about full Green’s functions. I will consider the
region where the pi are all Euclidean. That is already very unphysical and rather difficult to measure, but it makes
sure that I’m not going to encounter any singularities when I move around. I will also assume no partial sum of the
pi is zero. The sum of all the pi’s has to be zero, but I want no subset of them to add up to zero: 4

Then no matter how I cut the graphs into two parts, I won’t get momentum summing to zero; there will always be
some non-zero momentum flowing around. I’ll define an overall energy scale E (really, a pseudo-energy since
we’re in the Euclidean region) in terms of the sum of the pi2, with a minus to take care of my Euclidean metric
convention:

I’ll define some angular variables Ωij, not all independent:

I will consider the deep Euclidean region, very far from the mass shell,5 with the angular variables held constant:

What I’m really doing is scaling up all the momenta with some overall scale E, so that any momentum passing
through any part of the graph is scaled up. In that limit we would expect the mass to be unimportant. That’s a
guess. I won’t go through complicated combinatorics of showing whether or not that guess is true order by order in
perturbation theory; we haven’t the time.

We can make some preliminary progress by dimensional analysis. Using only the fact that the field ϕ(x) has
dimensions of mass, we find

(n) is a dimensionless function of µ/E, the coupling constant g and all the Ω’s; I’ll leave out the indices.
(Incidentally, the equation (50.7) is an application of Weinberg’s bound.6) Let’s check that: the two-point function
(2) goes like an energy squared (32.18), and the four-point function (4) is dimensionless (25.29), in agreement
with (50.7). Those checks suffice to fix the powers of E, times the dimensionless function (n).

The question we want to ask about the behavior of (n) is: Does it have a limit as µ → 0 for fixed values of E
and the Ω’s? By dimensional analysis, that’s the same as the limit E → ∞ with n, Ω and g fixed. Asking whether or
not the n-point functions are independent of the masses in the deep Euclidean region is equivalent to asking if a
zero mass theory exists as a nice smooth limit. This is a complicated question for a general graph, and a trivial one
for a tree graph. Let me take the case where it’s interesting enough to have structure, but not too complicated to
make this lecture infinitely long: a one-loop graph, as shown in Figure 50.1, with all momenta directed inward. It’s
like the graphs we considered when we were doing the effective potential (see (44.42) and (44.44)), except now
the external momenta are non-zero. This one-loop graph will involve an integral over the single loop momentum ℓ,
times a product of four propagators over all the internal lines

Figure 50.1: One-loop diagram

(where qi2 = ℓ2 + other stuff). I want to know if there’s a limit as µ2 → 0, which is the same as asking if there’s a limit
as E → ∞. Naїvely I would argue as follows: there is a limit as long as two q’s don’t vanish simultaneously. After all,
there are four powers of ℓ2 in the numerator, and therefore if only one qi vanishes at some particular point in the
region of ℓ integration, that’s not going to bother me. I can call that point the center of my ℓ integration. Then I will
have

which is still convergent in the ultraviolet regime. But if two q’s vanish at the same point, I get

and possible troubles from a logarithmic divergence (or worse, if more than two q’s vanish).

Now, is it possible to assign the loop momentum such that two internal momenta, say q1 and q3 in Figure
50.1, vanish simultaneously? No, it’s not. If two of the internal momenta are both equal to zero, then the sum of the
four external momenta p′i must be zero by energy-momentum conservation, and by hypothesis I’m in the region
where no partial sum of the external momenta vanishes.7 Provisionally it looks like, if the one-loop diagrams aren’t
lying to us, and more complicated things don’t happen on the multi-loop level, then in the deep Euclidean region

This argument looks pretty good. The only problem is that it’s wrong. We actually know what the first corrections
to the four-point function (4) are in this theory:

where the dots indicate terms (g3); the × in the last graph indicates a counterterm. Consider the bubble graph in
the series above. We know a lot about this graph, which we have encountered many times in the course of these
lectures.8 In particular, we know (25.29) that it grows logarithmically with the energy:

where c is some constant, and (S15.41) its imaginary part goes to a positive, finite constant. The crossed graphs
don’t cancel out; they all make the same sort of contribution in the deep Euclidean region. In that region, according
to (50.9), (4) is supposed to be a constant, but the expression in (50.11) scales like lnE.

What’s gone wrong? Well, I’ve been a little bit careless about the problem of renormalization subtractions.
I’ve been treating these graphs as if they were convergent. We have two conventions at hand. Both conventions
get into trouble when the mass goes to zero, because they collide into each other. One convention is to do our
renormalizations on the mass shell. That’s a disaster if the mass is zero, because the mass shell is on top of all
the singularities: the one-particle pole, the two-particle cut, the three-particle cut, etc., all on top of each other. The
other is to do all our subtractions with all external momenta equal to zero—the BPHZ prescription.9 That’s also a
disaster because if all the external momenta are zero, we can get hideous infrared divergences when the mass
goes to zero. While this sort of argument is golden before we do our renormalization subtractions, the very fact
that we have to make the renormalization subtractions keeps it from working. The renormalization subtractions
themselves, although they cancel the ultraviolet divergences, introduce new infrared divergences. One could
argue that in the deep Euclidean region, the particle loses all knowledge of what its mass is. However, we are
expressing Green’s functions for renormalized fields as functions of a parameter g. What is g and what is the
renormalized field? The renormalized field, and g, associated with a four-point function in this theory, are
renormalized on the mass shell. Those two quantities remember the mass shell in their definition, no matter how
far into the deep Euclidean region we go. This is a disease but it is very easy to cure. We won’t have any problems
as long as we make all of our subtractions at some fixed point in the Euclidean region that has absolutely nothing
to do with the masses.

Define renormalized quantities—the renormalized wave function and charge, etc.—at some point M2,
characterized by a mass M in the Euclidean region. Then our Green’s functions look more complicated because
there is now another mass in the problem. They’re functions of µ/E, E/M, Ω and g:

This g is now a completely different g for the same physical theory, because we’ve defined g differently. Then the
(n) should have a limit as µ → 0. Therefore we should be able to define a massless theory, and that should be
equivalent to the massive theory in the deep Euclidean region

Not only are our unrenormalized Green’s functions defined far away from the mass shell, but so are the
renormalization prescriptions themselves. So all of our counterterms are insensitive to the mass.

Is the trick clear? In order to avoid the renormalization conventions bringing the mass shell back in again,
we’ve got to pick the renormalization point, where we define both the scale of the renormalized fields and the
renormalized coupling constants, to be some point very far off the mass shell, so that they will not see the mass
term. At this moment this is just a hope: that if we do things this way there will be a smooth zero-mass limit. I will
now turn it from a hope to a flat assertion and tell you that it can be proven order by order in perturbation theory. I
won’t do that here.10 The argument for one-loop graphs is given above; there are many complexities in analyzing
multi-loop graphs.

We don’t have any predictions at this stage. Before we had a beautiful prediction that everything would just
go like a power of E by dimensional analysis. Even if we set µ/E to zero, we still have a dimensionless parameter
E/M. So it looks like we’ve solved our problem in principle, but gained no practical information. In fact one gains an
enormous amount of practical information by doing this. We’re going to study massless theories, and in particular,
develop general techniques for analysing the energy behavior of Green’s functions in massless field theories,
which I will write down in a more general form than this simple example.

The renormalization group is a technique for studying fully massless renormalizable field theories. It doesn’t
work for nonrenormalizable theories. (Nothing works for nonrenormalizable theories, to our knowledge; we don’t
even know if they exist in any real sense.) We want to study them because the behavior of such a fully massless
theory will mimic the behavior of a real theory with masses when we go to the deep Euclidean region. Indeed there
are other cases where we can study certain properties of a real theory with masses by studying corresponding
properties of a fully massless theory. There is a long, famous, and very important analysis that shows that certain
quantities associated with deep inelastic electroproduction, the so-called “moments” of the Bjorken structure
functions, Fi(k2, x) (i = 1, 2), behave as they would behave in a fully massless theory of quarks (the variable x is
defined below). 11 Let me describe the physics very briefly.

Deep inelastic electroproduction

 Figure 50.2: Deep inelastic electroproduction

The process in deep inelastic electroproduction, shown in Figure 50.2, is

where N is a nucleon, typically a proton, and X is any multiparticle state. Let k equal the momentum of the virtual
photon, and p the momentum of the target nucleon. If we know the momentum transfer k, we know everything,
because k2 and k + p are invariants; k is spacelike, because k = ℓ′ − ℓ, where ℓ and ℓ′ are the electron momenta,
and k2 = k ⋅ (ℓ′ − ℓ) < 0 in an inelastic collision.12 It is useful to introduce the Bjorken scaling variable, x;

where E is the energy of the virtual photon in the lab frame, and mN is the mass of the target nucleon. Because (k
+ p)2 > p2, elementary kinematics give

The moments of the structure functions are defined by

as −k2 → ∞. In fact, as −k2 → ∞, the Fi’s appear to depend only on x. This phenomenon is called Bjorken scaling.
Shortly after Bjorken’s discovery, it was realized that this behavior implied that the electrons scattered off pointlike
particles inside the nucleon; moreover, these particles behaved as if they were essentially free.13

So there are lots of things we know, beyond the fact that these simple Green’s functions, in the deep
Euclidean region, behave as they would behave in a fully massless theory. And then there are things that we can
actually measure that are insensitive to the masses in the theory. That’s an important but secondary matter.
Someone else does some hard work and says “Look, the quantity zilch is insensitive to the masses as the masses
go to zero,” and then you say “I can use the renormalization group to study zilch.” These are two separate issues.

I want to explain what I mean by a fully massless theory. A fully massless theory is one which has no masses
in it and no parameters with the dimensions of mass. No ϕ couplings, no ϕ 3 couplings, just a set of dimensionless
coupling constants. In fact we know in what sort of interactions such dimensionless coupling constants appear:
quartic interactions between scalar mesons, Yukawa interactions, and gauge field interactions. That’s it. Their
values are not important. I’ll call them

The theory will involve a set of fields

These fields may be the fundamental fields of the theory, whose Green’s functions we want to study:

Perhaps we want to look at something like the Green’s functions for a string of currents in electrodynamics. Or
maybe we want to investigate something peculiar, like the Green’s function for seven of the ϕ’s. It doesn’t matter
whether these are fundamental fields or not, nor what their Lorentz transformation properties are. These
properties will not be relevant in our analysis.

Finally, despite the fact that it is a fully massless theory, it has one mass, M, which determines the mass
scale at which we define all of our renormalization conventions. That mass cannot be zero. If it were, then as we
make subtractions at zero, we’re subtracting infrared divergent quantities. This sole mass M will define the
renormalization point, the place where we subtract our propagators, and the place where we set four-point
functions equal to a certain value to get the physical coupling constant.

I’ll apply the method to a Green’s function, but once we see how it works, we’ll see that it applies to practically
anything. A general Green’s function in such a theory will be14

The f(r) are scalar functions, which I will choose to be dimensionless by pulling out sufficient powers of E. It’s
important that the f(r) are scalars, not the components of a 3-vector. That they are Lorentz scalars is going to be
irrelevant. The range of r depends on what the Green’s function is (how many covariants we can make). If it’s a
two-point function for a spinor field there’ll be two of them; if it’s for a scalar field there’ll be one; if it involves 17
fields of very high spin there’ll be all sorts of things with Dirac γ matrices and tensor indices, and then they’ll break
up into a bunch of scalar invariants.

I stress that it’s the physical masses that are zero, not the bare masses. That is the one renormalization
convention that must be imposed at the point 0, rather than at the point M. For vector theories, gauge invariance
imposes zero physical mass if you have zero bare mass (modulo questions of spontaneous symmetry breaking,
which aren’t relevant for this kind of analysis). And for spinors, in most of the theories we’re interested in, γ5
invariance requires zero physical mass if the bare mass is zero. It’s only for scalar fields that zero bare mass does
not imply zero physical mass; in this case you have to make a subtraction. Then you may worry about whether
that subtraction will give you new infrared divergences. It doesn’t, but you’ll have to take my word for it. (A
subsidiary argument needs to be made; the wave function renormalization prevents it.)

You should really not trust me in some matters. If I tell you that something has been proved, that doesn’t
mean that I’ve actually gone through the paper and read all the details. It means that someone has sent me a
preprint.15 I look at it, and if it looks too horrible to wade through, I read the abstract which says a theorem has
been proved. I tell my students it’s been proved. Then two years later somebody comes by and says that this proof
was no good, and I say “It’s not been proved?” It’s like that joke: “Life is not a fountain?”16

We want to study the behavior of this Green’s function (50.21) as we change E. I’ll suppress both Ω, since
that’s going to be held fixed throughout the entire argument, and also the index r since we’ll just look at these
things one at a time; which one is not particularly relevant. One feature will turn out to give us powerful
information: M has an arbitrary value, so long as it is somewhere in the Euclidean region. That is, if I change M by
a dimensionless infinitesimal amount ϵ,

then I can keep the same physical theory; I have just changed my renormalization convention. I will also have to
change all the g’s by an amount of order ϵ because I’ve changed the renormalization point. There will be some
function, βa, a function of all the g’s but not of M by dimensional analysis:

When I write a g inside βa(g) I mean the set {g1, …, gm}; the functions {βa} are functions of all the g’s. The {βa} are
called, not surprisingly, beta functions. Each of the coupling constants may be related to all the others in a
complicated nonlinear way. I may have to rescale my fields:

The γA are functions, not Dirac matrices. For reasons to be explained later they are often called anomalous
dimensions. Under these changes (50.22)–(50.24), f will remain fixed:

because it’s the same theory.

Please note that it does not matter what the momenta are as long as the renormalization point is in the
Euclidean region. (I don’t want to make my renormalization subtractions on top of singularities.) Of course the
massless theory is comparable to the massive theory only in the deep Euclidean region. But if I separate the
question into two parts—how do I study the massless theory, and when can I use the massless theory to study the
massive theory—the question of being in the deep Euclidean region is relevant to the second part, but not to the
first.

I should say that (50.24) may be a matrix equation. We have occasionally talked about cases where we have
to mix together several fields which have the same dimension and the same Lorentz transformation properties as
a consequence of the renormalization, and we might get much worse things, if we’re looking at objects like ϕ 4,
which might get mixed up with (∂µϕ)2. So really the γ in (50.24) should be thought of as a matrix. For algebraic
simplicity and subsequent equations, however, I will assume it’s diagonal; that the fields we are studying do not
mix up with each other under renormalization. But I will make a little point here. It might be that

That is, some of our fields may mix up with each other in the course of renormalization. I just won’t worry about
that here. The generalization to the case where these matrices are present is fairly trivial.

I am assuming that in the deep Euclidean region, the Green’s functions are continuous. If you’ve got the same
physical theory, you may suddenly wake up and find yourself in the world of fully massless particles. You want to
parametrize it. You say, “I suspect this is fully massless ϕ 4 theory (or Yang–Mills theory). I’m going to do a bunch
of experiments. I’ll measure some Green’s functions in the deep Euclidean region,” (you have terrific experimental
apparatus) “and define the coupling constants.”

So you find the coupling constant g for gϕ 4 theory and say to the outside world “I found myself in a universe of
gϕ 4 with coupling constant equal to 0.1, for the massless theory.” And they say “How did you define g?” And you
say, “Oh, I defined it as a four-point function at me, the mass of an electron, with s = t = u, and with all external p2 =
−me2.” And they come back and say “That’s not the standard way. We want you to define it at the mass of a
Coleman,” or some other arbitrary different mass. But it’s the same theory. So you say, “Oh, all right, I’ll make that
measurement.” And you say it’s the theory with g = or or whatever. But it’s still the same theory. All M tells you
is how you label the Green’s function, and how you scale your field. So, suppose they want the coupling constant
for four fields at the mass of an electron. Well, you ask, “Which fields?” They say, “Scalar fields.” You ask “How do
you want it renormalized? Renormalized so that the two-point function has first derivative equal to one at −me2?”
And they reply, “No, we want it to be at −m2Coleman.” And so on. But it’s the same theory. The amazing thing is that
by keeping our wits about us we can use this trivial fact to get nontrivial information.

50.2The renormalization group equation

These Green’s functions, or these dimensionless scalar quantities f that characterize the Green’s functions, are
unchanged by this trivial group of transformations, which is just the reparametrization of the theory. The set of
transformations (50.22)–(50.24) is called the renormalization group. Rarely has there been a more pretentious
name in the history of physics. It’s like calling classical dynamics “the study of the Hamiltonian group of time
translations”. Nevertheless, that’s what it’s called. I’ve written this in terms of infinitesimals, but everything I can
write in terms of infinitesimals I can of course write in terms of a differential equation, and I will now do so. This
differential equation says that f does not change under these combined things. I’ll first write it down for a particular
function f:

where

is the sum of the little γ’s associated with whatever fields occur in the definition of the Green’s function; each of the
fields ϕ A is getting rescaled by an amount 1 + γA, and that just multiplies the whole Green’s function by the
number 1 + γ. This is known variously as the renormalization group equation (or RGE), or as the
Callan—Symanzik (or CS) equation.17 This differential equation follows from the trivial statement that it doesn’t
matter what the mass M is: γ depends on the particular f you are studying and in this way depends on how many
fields of which kind it has in it. Everything else is totally independent of what the particular function f is.

In any fully massless field theory there are, by ordinary dimensional analysis, functions of the coupling
constant only (not depending on the renormalization point), the β’s; one for each coupling constant. There are
other functions of the coupling constants, one for each field you happen to be studying, the γ’s, such that each
and every Green’s function will obey this differential equation (50.27). From this fact we could compute the β’s and
the γ’s iteratively in perturbation theory, because there is no problem computing the Green’s functions iteratively
in perturbation theory. Thus we could fix the β’s and the γ’s as those coefficients that make this equation true. It
looks much more complicated but it’s still the same trivial statement that the value of M is irrelevant, and that the
effects of infinitesimally changing M can be compensated for by effects of changing the coupling constants, with
the {βa}, and changing the scale of the fields, with the {γA}. There is no profound input into it, but it yields
surprisingly profound output.

A nice exercise (that I leave to you) is to find the β’s and the γ’s by applying this equation to Green’s functions
at the point where the renormalization constants are defined. That makes life particularly simple, but it doesn’t
matter. If you know the Green’s functions for any specified value of the coupling constants, you know how you
have to change the definition of the coupling constants, and rescale the fields, to keep the physics the same. Then
you can compute β and γ. Thus the β’s and the γ’s have well-defined perturbation theory expansions. Whether
they’re convergent or not is of course an open question, just as it is for the Green’s functions. Here are two
examples to give you an idea of how these things go.

EXAMPLE 1. ′ = gϕ 4

The single constant β arises because we have to redefine the one coupling constant g when we go to a new
renormalization mass. The coupling constant is defined in terms of the four-point function, and therefore we will
have to redefine it only if the four-point function has a nontrivial momentum dependence. Such dependence arises
at the one-loop level as shown in Figure 50.3—that’s a term in the four-point function that does depend on
momentum—and therefore β will first appear in order g2 in this theory. There’s only one coupling constant so

Figure 50.3: The one-loop contribution to β in ϕ 4 theory

where c is a constant. The function γA appears if we have to change the scale of the field when we change the
renormalization point. That happens only if the propagator has a nontrivial momentum dependence. In this theory
that also happens in order g2, as shown in Figure 50.4. There’s only one field, so there’s only one γ:

That’s obvious. What’s not obvious, and will be important to us, is the sign of c (or more precisely, the sign of β);
the sign of d will turn out to be completely irrelevant. In the interest of time, I ask you to take on trust that c is
positive:

Figure 50.4: The sunset diagram contribution to the propagator in ϕ 4 theory

(It’s trivial to verify. Just compute these graphs, which is pretty easy in a fully massless theory. You won’t obtain
any complicated functions at all, just ln(E/M).) That sign will be important to us later, although it’s not yet clear why.

EXAMPLE 2. QED: ′ = −g γµψAµ

It’s not always true that β and γ first appear in the same order of perturbation theory. For example, take
quantum electrodynamics, with a massless electron as well as a massless photon. The coupling constant is
defined in terms of the three-point function and a famous diagram: The first diagram that gives the three-point
function momentum dependence is shown in Figure 50.5, which is (g3) (or (e3), as we said earlier18):

Figure 50.5: The vertex correction in QED to (e3)

The amplitude for an off-shell electron to go into a physical electron and a photon has momentum dependence. It
therefore will introduce momentum dependence on the renormalization point, and how you define the scale of the
field given in the fixed theory. These things are defined on the mass shell, not off. It’s only odd orders in this case
because we have to stick on a photon in two places. It also will turn out—and be important to us later—that c′ is
positive,

That’s what we have to do a computation for. We actually have that computation in hand.19 In QED there are two
separate γ functions, γψ for the electron’s field and γA for the photon’s. The first time the electron propagator
starts getting momentum dependence is in Figure 50.6, and

We’ve always got to add an even number of powers. For the photon vacuum polarization, the relevant graph is
shown in Figure 50.7. We have

Figure 50.6: The electron self-energy to (e2)

This is how you get the powers of g. It should be clear how we actually compute the coefficients. We just stick in
the graphs at the appropriate order and then fix the coefficients so the renormalization group equations are true.20

Why did I bother to go through all this? I wanted to show you that M is an irrelevant parameter. It’s necessary,
but its value isn’t important; we always get the same physics. Therefore no matter what Green’s function I study, I
know its M dependence from (50.27) if I know the β’s and γ’s. In fact I don’t know what they are, but presumably I
could calculate them in perturbation theory. There are certainly far fewer quantities than the possible number of
Green’s functions. If I do know them, I know the M dependence. By dimensional analysis f(r) depends on M only in
the combination E/M. So if I know the M dependence, I know the E dependence. It’s almost as good as the old
case where we were being very naïve, not worrying about renormalization effects. In that case we assumed the E
dependence of all these dimensionless functions was trivial: they were E-independent. Here we say: Well, we
don’t know them trivially. But if we know the β’s and the γ’s, a finite set of functions, then we’ll know the E
dependence of everything. I will now go through a little exercise using the method of characteristics.21 I will write
down the general solution of the renormalization group equation in terms of initial value data at a fixed M, show
how we get the solution at a general value of M, study its properties and apply it.

Figure 50.7: The vacuum polarization to (e2)

50.3The solution to the renormalization group equation

How do I solve this equation using physical intuition, assuming that I know the β’s and γ’s exactly (which in
general I don’t)? Actually I hardly have to do any work. Although it may not look like it, this has a similar structure
to an equation whose solution we can almost obtain by inspection. Let me show you that second equation,22 then
I’ll write down the solution. Let ρ(x, t) be a scalar function for the population of bacteria in a fluid, at the position x at
a given time t. The bacteria are carried with a known velocity v(x) down a transparent tube, subject to a position-
dependent illumination L(x), also known, which determines their rate of reproduction. The bacteria move down the
pipe only because the fluid is moving. They have a limitless amount of sugar to eat; all they need is light. Then
they grow exponentially, depending on how much light they’re exposed to. Under these conditions, ρ(x, t) obeys
the following differential equation:

You can see a family resemblance to the renormalization group equation (50.27). The motion of a fluid element in
a given velocity field v(x) is often described by a device well known in hydrodynamics, the convective or total
derivative D/Dt:

I’ll solve the equation (50.36), and then, after making a transcription between the two equations, it will be very
easy to obtain the solution of the renormalization group equation. With one g they are the same equation, with v
playing the role of β and −γ playing the role of L. When you’ve got several g’s, it’s much the same story, except
that instead of moving down a pipe in a given velocity field, the bacteria are moving in an n-dimensional space.

The solution is pretty simple; it requires two steps. First we find out how to describe the motion of an element
of fluid, and then we work out the history of the bacteria. So step one is to solve an ordinary differential equation.
One defines the function as the solution of the equation

with the boundary condition

That’s an ordinary differential equation, not a partial differential equation. It determines x as a function of the
single variable t and also of the boundary values (x, 0). Physically, (x, t) is the position at time t of a
fluid element which was at x at a time t = 0. The differential equation (50.38) tells us how an element of fluid moves
in the given velocity field. In particular, (x, t1 − t2) is the position at time t1 of the element of fluid which
reaches x at a time t2.

In step two we study the bacteria. At time t = 0, the function ρ(x, t) is equal to some function P = P(x)
of how many bacteria were then at location x—the initial value of ρ. The bacteria multiply exponentially depending
on the value of L at the point where they are now. The whole thing is time-translation invariant, so the solution to
the bacteriological problem (50.36) is 23

To check this solution, note that at t = 0, the exponential disappears, (x, 0) is simply x, and the solution becomes

as it should.

Let’s forget the charming bacteria and return to the renormalization group equation. We may have no
comparable intuition for the renormalization group equation, but surely we have enough wit to make the
translation from one equation to the other. The solution of the RG equation in terms of the β’s and γ’s is going to
be the pivot point for the rest of the lecture.
 Table 50.1: Translation between bacteria and renormalization group variables

Following the solution to the bacteria problem, define as a set of functions that solve the simultaneous
ordinary differential equations

The function βa depends on all the ’s, with the boundary conditions

This leads to functions a(g, t) depending on the initial values of all the g’s and t, called running coupling
constants.

Now to make the substitution.24 We identify t with

Then the solution to the RG equation (transcribed from (50.40)) is

where F = F(g) = f(1, g).

This tells us exactly what we would expect the first-order differential equation to tell us: that if we know the
value of the function at any fixed E for all values of the coupling constants, and if we know the β’s and γ’s, then we
know the function at all E’s and for all coupling constants. That’s a very powerful statement, but of course, it’s also
a trivial statement. It’s the statement that M is an irrelevant parameter, exploited by straightforward calculus. We
have found the general solution (50.45) to the first-order partial differential RG equation (50.27) in terms of the
solution for a system of first-order ordinary differential equations (which, by the way, is the “method of
characteristics” referred to earlier).

I should make a small point. In order to keep the parallelism I had to substitute –ln(E/M) for t, but frequently in
the literature, because of the way all of the minus signs come into the solution of the hydrodynamic equation, the
parameter that enters is in fact ln(E/M). But it’s the same prescription.

50.4Applications of the renormalization group equation

I will show you three out of a host of applications of this equation and its general solution. The three applications
will be:

•The zeros of β. In a one-coupling constant theory, β(g) always has a zero at the origin (at g = 0). The
question is: if there are zeros elsewhere, does that tell us anything about the high-energy behavior of
the function? That’s not a question we can answer from perturbation theory. We have to invent some
non-perturbative method of analysis to see what happens when β has a zero. Nevertheless, the
consequences of β having a zero are so interesting that it is worth pursuing them, even if we don’t
know whether or not it does.

•Study of powers of ln(E/M) in perturbation theory. This is also sometimes known as summing the
leading logarithms.25 When we did the four-point function in perturbation theory, we found to lowest
order there was no logarithm. To (g2) there was one power of a logarithm. Does this go on? To
(g3) are there two powers of logarithms? To (g4) are there three? Maybe things
sometimes go bananas. When we go to (g18) do we get 22 powers of a logarithm, or does
each order introduce a single power of the logarithm? Who knows? But with the aid of this little
wonder, the RG equation, we’ll be able to answer that question, and without doing any work.

•*Asymptotic Freedom*. I have saved the best for last, and, in the manner of Hyman Kaplan,26 I put
asterisks around it, for reasons you will soon appreciate.
The effects of zeros of β

For simplicity I will assume we are working in a one coupling constant theory, like gϕ 4, so I can draw a graph
of β. Figure 50.8 shows a hypothetical function for β. The only thing we know (50.11) is that this begins with a
positive coefficient times g2, so it starts out like Figure 50.8, with a quadratic root at the origin. After that we are in
a state of total ignorance. Maybe some people who work with lattice quantum field theory can compute the strong
coupling limit and tell us something about it, but I can’t. So let’s just make a guess about β and see its
consequences. I have no particular reason for assuming this form; it’s just a nice example for describing the
zoölogy of the zeros of β. To make life interesting, I’ll assume it has a zero at a point g1 and a second zero at g2, a
third zero at g3, and then it stays negative. Who knows? That’s just a guess. If you want 14 zeros or a double zero
or a triple zero, you can work out the consequences. I’ll work out the consequences of this one.

Figure 50.8: Hypothetical β function in ϕ 4 theory

Now something very interesting occurs if we study what happens to (50.42). For a theory with only one
coupling constant,

I’ll assume I start out with my initial condition

I know the solution to this equation. If t is anywhere in region I, β is positive and therefore increases as t
increases, because the slope stays positive until g1. But the curve β( (g)) can’t cross at g1 into negative β,
because at g1, β = 0 and d /dt = 0. In terms of our hydrodynamic analogy, this is a stagnation point; it’s a
sink. The velocities pour into it. Therefore, in region I,

No matter where we start in region I, we end up at the same place. It doesn’t matter where we are on the river, we
will eventually go over Niagara Falls. I’ve drawn it so that it’s got a nice derivative at = g1. We can also find
the approach to the limit. In region I,

where a is a positive constant, a > 0. Near the limit we can drop the (g − g1)2 terms; those will be second-order in
a small quantity:

the solution to which is

In region I, reaches g1 pretty quickly, like an exponential in t once it is in the neighborhood of g1:

What effect does this have on our general Green’s function? Well, is going to g1 no matter where in region I
it started from. This means that f (50.45) goes to a function F(g1) (because is going to g1), times an
exponential which I can break up into two parts:27
In the first integral, γ is changing as I go through all the intermediate values of t in (50.45). As → g1 that part
converges and is just going to be some multiplicative constant; I don’t particularly care about that. In the second,
the integrand is a constant, and the integral is trivial. Writing the constant value of the first integral as lnK for later
convenience,

So we have

I get simple power behavior. No matter what Green’s function I start out with, no matter which coupling constant I
choose, the f goes like a simple power. It doesn’t matter what the initial value of the coupling constant is; the
asymptotic form is totally independent of the initial value g < g1 of the coupling constant, within the range I’m
studying. Only in the scaling constant K, which involves all the coupling constants, do I have any information about
where I started from. And that K is trivial. I can always get rid of it by changing the normalization of my fields.

Remember that γ has the following structure:

The A’s label the various fields that go into f, which is defined for the Green’s function

in (50.21). The function f depends on what fields are there, but that’s all it depends on: how many times each field
comes in. So the rule for the power is indeed very simple: we have appropriate powers of (E/M), the powers
determined by the value of γ at g1 for each field that’s in the Green’s function. This is very similar to the sort of
behavior we found (50.9) when we were using simple dimensional analysis, and not worrying about
renormalization effects. There we also just got a simple power. In that case the simple power was 0 for a
dimensionless function, in particular (4). But that’s not what (50.55) says. For this reason these quantities γ are
sometimes called anomalous dimensions.28 We obtain the same sort of scaling behavior we would get from
dimensional analysis if the dimensions of the fields were something different than what we naïvely expect—if
instead of the scalar field having dimension 1 it had dimension 1 + γ. We’ve shown that if 0 < g < g1 then as t → ∞,
→ g1. Who knows if β has a zero or not? (It doesn’t actually seem to have a zero in λϕ 4 theory.) This is
marvelous stuff, isn’t it?

Let’s go on to region II, g1 ≤ g < g2. If g = g1, β = 0 and stays g1 forever. If however g2 ≥ g > g1, because β is
negative, we’re pushed again to g1. That’s exactly the same story in the whole of region II:

The asymptotic behavior of the theory at high s is determined by the behavior at g1, which is sometimes called an
ultraviolet stable fixed point,29 “ultraviolet” because we are going to high energy, ln(E/M) going to ∞, and “stable”
because on either side of it, g is inexorably drawn into that value, and will not budge once it gets there.

What about g2? Well, g = g2 is peculiar, an exceptional point; β vanishes at g = g2, so = g2 forever, no
limit required:

On the other hand if we’re a little way to the left or right of g2 we get drawn into either g1 or g3, respectively. So g2
is called a UV unstable, or sometimes an infrared stable, fixed point. It’s the same story with g3 as we had with
g1. If g is up in region III, g > g2 as I’ve drawn it, it increases to g3; if it’s down in region IV it decreases to g3, so

We summarize these results in Table 50.2. (The infrared stable fixed points are sometimes of physical interest in
statistical mechanics, but that’s a long story that I don’t want to go into.30)

Table 50.2: Possible values of from the hypothetical β in Figure 50.8

Notice the simplicity of the asymptotic structure we get in this hypothetical model. This is a theory with an
apparent free parameter g in it, a coupling constant that I can vary any way I like. But no matter what initial value I
give the coupling constant, I get only three possible asymptotic values: g1, g2 (for the single choice g = g2), or g3.
The asymptotic form is a discontinuous function of the initial value of the coupling constant, governed by the
values of the single β in this theory: we have only three different theories here, not a continuous infinity. That’s the
fundamental result of the renormalization group: what the coupling constant is depends on what you choose for
your scale of mass. In this model, we can get to any coupling constant between 0 and g1 just by changing the
mass; likewise between g1 and g2, and between g2 and g3. (And if we pick the mass so we start out at an IR stable
fixed point, if we start out at g2, then it will stay that way no matter where we choose the mass M.) The way we
choose M doesn’t depend on where we start. It’s arbitrary. If I say, for example, g1 = 1, there is no difference
between the theory with g = and the theory with g = 100. We have a theory with g = 100 if we choose the
renormalization point to be the mass of an electron. The theory with g = is obtained if we use, instead of the
mass of an electron, the mass of a Coleman. It’s the same theory, though with two different renormalization
conventions, so of course it has the same asymptotic behavior. The two versions differ only in the mass scale, but
that’s trivial; we get rid of M by dimensional analysis.

We started with a naïve viewpoint. We thought that the theory depended trivially on the mass (the
dimensionless functions were mass-independent), and nontrivially (in some complicated way) on the coupling
constant. We were deluding ourselves. This is precisely wrong! The f’s depend nontrivially on the mass (through
the γ’s), and trivially on the coupling constant. These results are found through simple deduction, starting from the
observation that we need some mass in a naturally massless theory, but the value of that mass is arbitrary; the
mass and coupling constant are continuously varying quantities. We’ve turned our naïve viewpoint upside down.

For hypothetical purposes, we can choose the shape of the curve β( ) any way we want; nothing is known in
general. All I need to assume is that the graph is continuous and, I suppose, for this little estimate, differentiable at
the point where β changes sign. If the curve stays above the g-axis as t → ∞ then we will not get smooth
asymptotic behavior. All those theories would be the same because we could turn one into another by changing
the renormalization point, but we wouldn’t get simple power behavior. The high-energy behavior would be some
awful mess, depending on precisely at what rate things went to infinity, and how γ( ) grew with the coupling
constant. We wouldn’t be able to simplify the integral (50.53) as we did.

Summing the leading logs

I’ll put aside making guesses about the graph of β, and turn next to the study of the structure of perturbation
theory and the logarithms that appear in it. A particular Green’s function, some f , will typically be a
mess. If we compute things out to large order, we find some numerical coefficients, powers of g, and powers of
ln(E/M):

It will in general be that complicated, but not more so. You might think, “Hmm, why isn’t there a ln(ln(E/M)), or
powers of (E/M)—say, ?” I’ll now show that f is indeed of the form (50.59), by using the
renormalization group equations, again for a single coupling constant theory. It’s easy enough to generalize the
argument.

My starting point will be the differential equation (50.46) for the running coupling constant :
with the boundary condition = g at t = 0 (50.43). That’s certainly right. This is the β which we compute order
by order in perturbation theory. The series starts with m = 2, because we know that in ϕ 4 theory, to lowest order β
∝ g2 (50.29). I will show that this equation admits a power series solution of the form

That is, with every power n of g we get a power of t no higher than (n – 1). Once I have proved this, you will easily
see how to organize the logarithms in an arbitrary Green’s function, because t gets replaced by ln(E/M) and the γ
gives us various powers from this power series (50.28). All the f’s have a power series in , and γ has a power
series in , and we just plug it all into the RG equation. If has the form (50.61), then all the coefficients
of the RG equation do as well, and so f will have the form (50.59). This will tell us in particular that we
never get more than one more power of the same logarithm for each extra power of g. We’ll never get log log or
or anything else like that.

I’ll prove (50.59) in an absolutely trivial way. If I take a series of the form (50.61), with the minimum value of r
equal to some integer, k, and plug it into the right-hand side of (50.60), its mth power is a series of the same form,
except it has a larger minimum value of r. For example, consider n = 3. The least value of r in (50.61) is 1, and the
term corresponding to n = 3 is

Now consider the term m = 2 in the series on the right-hand side of (50.60). When I square (50.62) I get

That is, the minimum r equals 2. When I cube it I’ll get things like g9t6 which has r = 3, three fewer powers of t. So
the nth power of the term with minimum r = k winds up with r = nk. On the left-hand side of (50.59), the derivative d
/dt (50.60) knocks off one t but doesn’t do anything to the g’s, so the corresponding term in the derivative
has minimum r = k + 1.

It’s easy to see what happens next. I plug (50.61) into (50.60). On the right-hand side, there will be no terms
with r = 1 (i.e., no term of the form gntn−1) because β( ) begins with g2. There will be terms with r = 2. They’ll only
come from the r = 1 term in the expansion (50.61) plugged into the 2 term. All the other terms will have r = 3.
On the left-hand side, the t derivative will also have no r = 1 terms, and r = 2 terms only from the r = 1 term in the
original expansion. I’ll match the terms of (g2) on either side with r = 2:

That tells me the terms with r = 1 in the original expansion completely determine β to order g2. All the terms in the
original expansion with r > 1 have r ≥ 2 when I plug them in, and will give terms with r ≥ 4. I don’t have to worry
about them. And the g3 terms, even from my original term, will give me terms with r = 4. So I know all the terms
with r = 1 in the function if I know c2. Likewise if I know both c2 and c3 I know all the terms with r = 2. If I know c2, c3,
and c4, I know all the terms with r = 1, 2, and 3. You see what is happening. I keep building up the power of r
whenever I raise the power of the series expansion for . I keep raising the power of g relative to the power of t =
ln(E/M). The two series clearly can be made equal to each other iteratively, so I’ve shown that has a power
series solution of the form (50.61). Second, the iterative solution demonstrates an amazing fact: if I want to know
the highest power of t in any given power of g, the so-called leading logarithms, I need only know c2. If I want to
know next-to-leading logarithms I need only compute c3.

So I’ve learned two remarkable things. First, I do have a power series expansion of the form (50.59) with m
bounded by n for any given Green’s function, bounded in a rather trivial way by how many powers of g emerge in
the lowest order. Second, I can easily find the coefficients of gm m−1. If I’m studying, for example, a
four-point function, and I want to look at order g128, I know that it occurs as the product g128 127. What is
the coefficient of that term? I only need to know the terms of ( 2) in this expansion, and β to (g2). That is, I
only need to do a one-loop computation and plug it into the renormalization group equation to get, with 100%
accuracy, the coefficient of g128 127. There are few more efficient ways of finding that coefficient. Writing
down all diagrams of 128th order and studying their asymptotic form is not the right way to do it. If I want to know
the coefficient of 126 then I have to do a two-loop calculation.
 Again it’s just a consequence of the earlier arguments. The only input needed for this argument is that M is an
irrelevant parameter. When I change M, all the terms I generate from the power series expansion (50.59) in terms
of logarithms have got to come together, and be absorbed in some way into the redefinitions of g and the overall
scale. That’s the secret of the magic. It means there are complicated tight relations among those coefficients, as
we’ve seen.

*Asymptotic Freedom*

I will now discuss the hero of the hour (actually the hero of the lustrum31), asymptotic freedom. From our
previous analysis, we know that at high energies perturbation theory is liable to be unreliable, even if we start out
with a small coupling constant. The reason is that we not only get powers of g but powers of ln(E/M). As the
validity of perturbation theory, in the most naïve sense, requires that the things that multiply your various
coefficients should be small, we need not only that |g| < 1, but also that g|ln(E/M)|¿ 1:

I will show how we can use the renormalization group to improve perturbation theory, to replace (50.65) with a
single condition

which may sometimes be valid when the two separate conditions are not met. We’ll get an idea of how we can do
this by summing the logarithms, as we talked about above. I’ll again take as an example a simple theory with only
one coupling constant, quantum electrodynamics:

I’m going to solve the renormalization group equations approximately. Everything in (50.45) is given automatically
in a power series in . If is small, that is groovy.

Now let’s see about , by solving the equation (50.60) approximately:

(we know from (50.32) that to lowest order, d /dt in QED is ( 3)). I’m going to assume that I’m working in a
range where is small, so I can neglect the higher orders. I will later check that for self-consistency. By
solving the equation I’ll know when is large and when it’s small. Solving this first-order differential equation
equation is trivial, as easy as doing your income tax:

with the boundary condition (50.43), = g at t = 0, has the solution

The whole approximation is based on the idea that 2 stays small. Certainly if g2 is small at our starting point
that’s true for small t. But the question is, can we get beyond that?

Now we notice something marvelous. Recall (50.33): c′ is positive. If t → −∞, 2 stays small. If t → +∞,
we’re out of luck: becomes imaginary. Therefore, we can indeed extend perturbation theory, but only to
arbitrarily large negative t = ln(E/M). We can’t extend it to arbitrarily large positive ln(E/M), because the
approximation becomes inconsistent. But to arbitrarily large negative ln(E/M) we can improve perturbation theory
and replace, as stated, (50.65) by the single condition (50.66). This is wonderful. Unfortunately, it’s also
absolutely useless. The reason is that large negative ln(E/M) means very small E. We’re not interested in the
behavior of the massless theory at very small E. It’s supposed to simulate the behavior of the massive theory only
for very large E. When we go to very small E, the deep infrared region, we’ll again see those masses we threw
away at the very beginning of this lecture, unless we’re really living in a world with fully massless electrodynamics.
And in that case, we’d be able to sum up this infrared structure exactly. That’s true but not particularly interesting.
The problem is the positive sign of c′.

I come finally to the sensational discovery made independently by Politzer, ’t Hooft, and Gross and Wilczek,
the last two working collaboratively.32 Though ’t Hooft did not realize its consequences, Politzer, and Gross and
Wilczek, did, and went crazy. They made a very simple computation (which you yourself are capable of doing with
the methods I’ve shown you) of β for a non-Abelian Yang–Mills theory, with a multiplet of fermions, or without, it
doesn’t matter. It’s still a theory with a single coupling constant, the gauge coupling constant; the graphs look the
same. They discovered that c′, and hence β(g), is negative if there are not too many fermions. “Not too many” is a
technical issue. For an SU(N) gauge theory with nf species of fermions in the fundamental representation

In the gauge group we associate with color, SU(3), 17 triplets of fermions in the fundamental representation are
too many, but 16 are not.33 That’s the cross-over point.

Now what does this mean? Well, it means that everything I have said before is still true, except ultraviolet
replaces infrared, because the sign of c′ has changed, from positive to negative. We can now sum up the
improved perturbation theory in a region that is of interest to us, the high-energy region where our massless theory
is supposed to simulate a massive theory. The high-energy behavior of a theory of Yang–Mills particles and
fermions, is computable at arbitrarily high Euclidean energies. We can probe these theories with electroproduction
experiments. The high-energy behavior is computable, so we can predict in a Lagrangian field theory what is
going on at high energies. Furthermore, it doesn’t matter how large the coupling constant is initially as long as it’s
not too large. We now know β points downward near the origin, as shown in Figure 50.9. For all we know it may
keep on going down forever. Maybe it has a zero someplace. If it does have a zero, we know that it’s not a small
number. If it were, we could compute it in perturbation theory. But in perturbation theory it doesn’t have a zero. For
any value of the coupling constant between this possible zero and the origin, by the arguments given before as E
→ ∞, we are forced into the origin. We’re not able to study what happens all the way along the g-axis, but that
doesn’t matter. Eventually we’re coming down to the origin. We don’t care how the renormalization group equation
drove us there. When we’re near the origin we can compute what happens using perturbation theory fixed up by
the renormalization group. If this zero does exist at all, in the theory in question, it doesn’t matter how large the
coupling constant is. If the zero is set at g = 17, for any g < 17 the method will work.

Figure 50.9: Beta function for asymptotic freedom in Yang–Mills theories

This is called asymptotic freedom, because instead of being pushed towards one of those points g1, we are
pushed towards a free field theory, g = 0. Writing b = −c′, where b > 0, the running coupling constant (50.70)
becomes

Asymptotically the theory is free, aside from corrections we know how to compute. They turn out to be powers of
logarithms, as I’ll now show. From (50.45), asymptotically f goes to F(0) times the exponential of the integral of γ.
Analogous to (50.35) we can say for a gluon, G,

using (50.72). The integral (50.45) becomes

so that
The correction involves a power of ln(E/M), as claimed. 34 We can compute these corrections with what we have.
Everything is predictable and everything looks like a free field theory, with tiny corrections that get tinier and tinier
as the energy gets larger and larger, because → 0. Of course what the coupling constant is depends on what the
renormalization mass is.

Let me tell you an anecdote. Asymptotic freedom was discovered by my graduate student David Politzer.35 I
was off at Princeton on a sabbatical, and he came down from Harvard to visit me. We were working on some other
(totally uninteresting) problem, trying to solve dynamical symmetry breakdown, to get the Nambu–Goldstone
phenomenon without fundamental scalar fields. We thought it would be easier than it turned out to be after a year
of labor. I said, “You’re getting nowhere with your thesis. It would be nice to know the renormalization group
functions for the Yang–Mills theory. Nobody’s worked them out yet. Why don’t you compute them? That’s not
going to be a lot of work, but it’s something to do.” Actually ’t Hooft had computed them the summer before, but
hadn’t published them. He announced them at a seminar in Marseille.36 I added, “Nobody expects them to come
out negative.” No one had thought in advance what the consequences would be if the beta function turned out to
be negative.

Politzer went back to Harvard, and here’s where you see the sign of genius. Not only did he follow my orders,
he knew what to do with the result. He called me up one night and said, “I’ve computed them, and they’re
negative.” And I said “Oh, that’s interesting. This is telling us something important about the strong interactions.”
He was very smart; he realized what it meant. Not only did he get the right sign, he drew the right conclusion,
which even someone as smart as ’t Hooft didn’t do. This result would explain why you apparently see free quarks
inside the nucleon when you do deep inelastic scattering. Then you’re probing this region of high energy, and in
that region the effective coupling constant , the quantity that governs the interactions among the quarks, is small.
In fact shortly thereafter, David Gross and I showed that in four dimensions, the only renormalizable field theories
that allow for asymptotic freedom are Yang–Mills theories.37 For everything else β is positive. The color
interaction between colored quarks is due to a non-Abelian Yang–Mills theory. They look freer and freer at higher
and higher energy because of this phenomenon, asymptotic freedom: is getting smaller and smaller with
higher energies.

The big test is deep inelastic electroproduction,38 as mentioned earlier:

It’s rather complicated to fit deep inelastic electroproduction because the data changes a lot. It usually turns out
that to get an accurate fit, you have to know something that the experimentalists haven’t quite measured yet, the
value of . It is known that 2/(4π) ≈ 0.5 near 1 GeV.39 So it’s falling off fairly rapidly. It works the other way
around. That’s the opposite side of asymptotic freedom. As you go the other direction, towards lower energies,
gets bigger and bigger, and faster and faster. That’s presumably why the quarks don’t get out of the
hadrons containing them: the force is getting stronger and stronger as they’re getting farther and farther apart at
larger distances. That’s infrared slavery; the quarks are confined.

I would like to say one or two sentences about the content of this course. There are many topics I have not
covered. I’ve said nothing about Regge poles, and for that I feel guilty.40 I’ve said nothing about many strong
interaction processes, like inclusive pion production. There’s a lot of important physics which you haven’t learned
from this course that involve, in one way or another, field theoretical ideas. I think it is pleasant, however, that in
the last few weeks, I’ve been able to deliberately contradict two things I previously taught as received dogma, the
last time I taught the second half of this course, five years ago. One was that weak interactions are much weaker
than electromagnetism. That’s false. The GSW model tells us they are exactly the same strength. We were
worried about non-renormalizable theories because we thought the weak interactions got stronger at higher and
higher energy with the piling up of all those powers of energy. That’s also false. The GSW model of the weak
intereactions is a renormalizable theory. They get to electromagnetic strength and stay there. And the other thing
is the marvelous reversal. Instead of believing the weak interactions get strong at high energies, we now believe
the strong interactions get weak at high energies, as I’ve demonstrated. That is the end of the course, and I hope
you’ve enjoyed it.41

1[Eds.] E. C. G. Stueckelberg and A. Petermann, “La normalisation des constantes dans la theorie des quanta”
(The normalization of constants in quantum theory), Helv. Phys. Acta 26 (1953) 499–520; M. Gell-Mann and F. E.
Low, “Quantum Electrodynamics at Small Distances”, Phys. Rev. 95 (1954) 1300–1311; T. D. Lee, Particle
Physics and Introduction to Field Theory, Harwood Academic Publishers, 1981, pp. 458–462; “Dilatations”, pp.
79–96 and “Secret Symmetry”, Sections 6.2 to 6.4, pp. 171–178 in Coleman Aspects; Ryder QFT, Section 9.4, pp.
334–339; Cheng & Li, GT, Chapter 3, pp. 67–85.
2[Eds.] The Optical Theorem, (12.49).
3[Eds.] See §11.3, p. 231. The Mandelstam variables s, t, and u are defined in (11.19a)–(11.19c).
4[Eds.] In the literature, a set of Euclidean {pi} is called unexceptional if no proper subset of them sums to zero.
5[Eds.] Cheng & Li GT, pp. 73–74.
6[Eds.]
Steven Weinberg, “High-Energy Behavior in Quantum Field Theory”, Phys. Rev. 118 (1960) 838–849;
Cheng & Li GT, pp. 73–74. A slightly different version is given in Bjorken & Drell Fields, pp. 322–324. See also
Coleman Aspects, p. 80, equation (3.6); Weinberg’s bound, that (n) grows no faster than E4−n times a polynomial
in ln(E/µ), is given on p. 81. Briefly, the dimensional analysis goes like this: the dimension of ϕ(x) is L−1 = E. From
the definition (13.22) of G(n)(x1, …, xn), it follows G(n)(xi) ~ En. Taking the Fourier transform (13.4) leads to (n)(pi) ~
E−4nEn ~ E−3n. Finally, from the definition (32.14) we can write

so that (n)(pi) ~ E4E2nE−3n ~ E(4−n).

7[Eds.] To flesh out this argument, let q2 = ℓ. Then

But if q3 = 0, the partial sum p′1 + p′2 + p′3 + p′4 = 0, contrary to hypothesis.
8[Eds.] The diagram was introduced in Figure 14.3, p. 307, made a second appearance in Figure 16.9, p. 344, and
was the subject of an example starting on p. 530, where the logarithmic dependence (25.29) was obtained, p. 531.
It reappeared briefly on p. 536. It was also the subject of Problem 4 on the 1975 253a final examination (see
Problem 15.4, p. 591, and its solution).
9[Eds.]
See §25.2; Chapter 4, “Renormalization and Symmetry: A Review for Non-Specialists”, pp. 103–104 in
Coleman Aspects; Peskin & Schroeder, QFT, pp. 337–344 describe the BPHZ procedure with the four-point
function in ϕ 4 theory.
10[Eds.] Coleman Aspects, “Dilatations”, Section 4.3, “Scaling and the Operator Product Expansion”, pp. 93–96;
Itzykson & Zuber, QFT, pp. 654–656.
11[Eds.] J. D. Bjorken, “Asymptotic Sum Rules at Infinite Momentum”, Phys. Rev. 179 (1969) 1547–1553;
Coleman Aspects, “Secret Symmetry”, Section 6.1, pp. 169–171. Be careful not to confuse the Bjorken functions
Fi’s with the form factors F1 and F2 from earlier lectures.
12[Eds.] Another way to see that k = ℓ′ − ℓ is spacelike: square both sides. Then k2 = 2m2 − 2ℓ ⋅ ℓ′. Go to the center
of momentum frame of the electrons, ℓ = (ℓ0, ℓ), ℓ′ = (ℓ0, −ℓ), and k2 = 2m2 − 2(ℓ02 + |ℓ|2) = 4(m2 − ℓ02) = −4|ℓ|2 < 0.
13[Eds.] The first direct evidence of quarks came from deep inelastic scattering experiments of electrons off
protons, carried out at the Stanford Linear Accelerator (SLAC) in 1967–1970, headed by Jerome I. Friedman,
Henry W. Kendall and Richard E. Taylor. These three shared the 1990 Nobel Prize in Physics for this work. See
Crease & Mann SC, pp. 299–308.
14[Eds.] F.T. = Fourier transform.
15[Eds.] This was in the days before the arXiv, when preprints arrived in an envelope. Occasionally the same result
simultaneously derived by two rival groups was sent by each to the other, and crossed in the mail.
16[Eds.]
An old joke, with many variants. A young man seeking enlightenment travels to a distant land to ask a
famous wise man the meaning of life. “Life,” the ancient sage tells him, “is like a fountain.” The young man thanks
the sage, and goes off to make his fortune. Many years later, he decides to revisit the old master in his last days.
He says to him, “Master, I thank you for your wonderful advice. It has served me well through many trials. But I
must confess to you that I really don’t understand it.” The sage reflects for a few moments, and asks the younger
man, “Life is not like a fountain?” For another version, see Jimmy Pritchard, The New York City Bartender’s Joke
Book, Warner Books, 2002.
17[Eds.] K. Symanzik, “Small Distance Behavior in Field Theory and Power Counting”, Commun. Math. Phys. 18
(1970) 227–246; Curtis G. Callan, Jr., “Broken Scale Invariance in Scalar Field Theory”, Phys. Rev. D2 (1970)
1541–1546; “Introduction to Renormalization Theory”, pp. 42–77 in Methods in Field Theory (Les Houches 1975),
eds. R. Balian and J. Zinn-Justin, North-Holland, 1976; “Dilatations”, p. 86 in Coleman Aspects; Ryder QFT,
Section 9.4, pp. 334–339. See also the closely related article by Wilson and the review article by Huang: Kenneth
G. Wilson, “Anomalous Dimensions and the Breakdown of Scale Invariance in Perturbation Theory”, Phys. Rev.
D2 (1970) 1478–1493; K. Huang, “A Critical History of Renormalization”, Int. J. Mod. Phys. A 28 (2013) 11330050;
available online at https://arxiv.org/abs/1310.5533. It is perhaps worth quoting the acknowledgement in Callan’s
1970 article: “It is a pleasure to acknowledge many discussions with Sidney Coleman, without which this paper
could not have been written.” Technically the CS equation is an inhomogeneous equation with a mass-related
term on the right-hand side. It becomes the RGE in the deep Euclidean region. For a discussion of the differences
between the two equations, see M. Kaku, Quantum Field Theory: A Modern Introduction, Oxford U. P., 1993,
Section 14.7, pp. 485–488.
18[Eds.] See Figure 34.6, p. 744.
19[Eds.] See the discussion of the electron’s anomalous magnetic moment, §34.3. The β function in QED requires
input from the electron self-energy, the photon self-energy and the Ward identity in the simple form Z1 = Z2, as
well as the vertex diagram. The value for c′ in (50.32) is c′ = 1/(12π2); Peskin & Schroeder QFT, pp. 415–416.
20[Eds.] Peskin & Schroeder QFT, Section 12.2, “The Callan–Symanzik Equation”, pp. 406–418.
21[Eds.] In 1990, Coleman said, “You can read about this [topic] in Courant and Hilbert, which nobody younger
than me has ever held in their hands.” He was referring to Richard Courant and David Hilbert, Methods of
Mathematical Physics, v.2, John Wiley Interscience Publishers, 1962, pp. 450–463. (The editors proudly serve as
counterexamples.) The work deserves to be better known by later generations. Long ago, the immense
importance of Courant–Hilbert to the development of quantum mechanics was famous: “In retrospect, it seems
almost uncanny how mathematics now prepared itself for its future service to quantum mechanics... [In May,
1924] Courant, utilizing Hilbert’s lectures, finished in Göttingen the first volume... Published at the end of 1924, it
contained precisely those parts of algebra and analysis on which the later development of quantum mechanics
had to be based; its merits for the subsequent rapid growth of our theory can hardly be exaggerated. One of
Courant’s assistants in the preparation of this work was Pascual Jordan...” Max Jammer, The Conceptual
Development of Quantum Mechanics, MIT Press, 1966, p. 207. The two volumes were deemed so crucial to the
war effort that the U.S. government (which had seized the copyright on all German works) had Interscience
publish an edition (in the original German, which nearly all American physicists of the era read) in 1943; 7000
copies were sold: Constance Reid, Courant in Göttingen and New York: The Story of an Improbable
Mathematician, Springer-Verlag, 1976, p. 465; also published with Reid’s earlier biography Hilbert (Springer-
Verlag, 1970) in a single volume, Hilbert–Courant, Springer-Verlag, 1986.
22[Eds.] Coleman Aspects, Chapter 3, “Dilatations”, pp. 88–90; Peskin & Schroeder QFT, pp. 418–420.
23[Eds.] Perhaps because the time was short, Coleman skipped a few steps in this derivation. At time t = 0, the
differential equation (50.36) can be written, with the given definitions, as

The solution to the associated equation for all times is

Shift t backwards to 0, and (50.40) follows. See Peskin & Schroeder QFT, pp. 418–420; Coleman Aspects,
“Dilatations”, pp. 88–89.
24[Eds.] M∂/∂M = (M/E)∂/∂(M/E) = ∂/∂t if t = ln(M/E).
25[Eds.] V. V. Sudakov, “Vertex Parts at Very High Energies in Quantum Electrodynamics”, Sov. Phys. JETP 3
(1956) 65–71; Cheng & Li, GT, pp. 316–320.
26[Eds.] Leonard Ross (pseudonym of Leo Rosten), The Education of H*Y*M*A*N K*A*P*L*A*N, Harcourt, Brace,
1937; The Return of H*Y*M*A*N K*A*P*L*A*N, Harper, New York, 1959. Combined as O K*A*P*L*A*N! My
K*A*P*L*A*N!, Harper and Row, 1976.
27[Eds.] Coleman Aspects, “Dilatations”, Section 4.2, pp. 90–92.
28[Eds.] Kenneth G. Wilson, “Renormalization Group and Strong Interactions”, Phys. Rev. D3 (1971) 1818–1846;
Ryder QFT, p. 326. Incidentally, Figure 50.8 bears a strong resemblance to Wilson’s Figure 1, p. 1826.
29[Eds.] Peskin & Schroeder QFT, p. 427; Ryder QFT, p. 327.
30[Eds.] Kerson Huang, Statistical Mechanics, 2nd ed., John Wiley & Sons, 1987, Chapter 18, pp. 441–467.
31[Eds.] In ancient Rome, the census was held every five years. At the end of the census, there was a period of
penitence and public expiation ceremonies, typically involving animal sacrifice, called the lustrum; the word
derives from the Greek verb , “luo”, to loosen, release, undo, or repent (it is a root of the word “analysis”, and of
the name of Aristophanes’ heroine Lysistrata, “undoing the army”; the Greek upsilon υ is often transliterated “y”);
N. G. L. Hammond and H. H. Scullard, eds., The Oxford Classical Dictionary, 2nd ed., Oxford U. P., 1970,
“Lustration”, p. 626. Coleman is using the word here in its sense of a five-year period. Asymptotic freedom was
discovered in 1973, within five years of the videotaped 1976 lectures.
32[Eds.]
H. David Politzer, “Reliable Perturbative Results for Strong Interactions?”, Phys. Rev. Lett. 30 (1973)
1346–1349; “Asymptotic Freedom: An Approach to Strong Interactions”, Phys. Reps. C14 (1974) 129–180; David
J. Gross and Frank Wilczek, “Ultraviolet Behavior of Non-Abelian Gauge Theories”, Phys. Rev. Lett. 30 (1973)
1343–1346; “Asymptotically Free Gauge Theories I”, Phys. Rev. D8 (1973) 3633–3652; “Asymptotically Free
Gauge Theories II”, Phys. Rev. D9 (1974) 980–992; G. ’t Hooft, unpublished remarks, Marseille Conference on
Renormalization of Yang–Mills Fields and Applications in Particle Physics, June, 1972; Gerard ’t Hooft, “When
was Asymptotic Freedom Discovered? or, The Rehabilitation of Quantum Field Theory”, Nuc. Phys. B (Proc.
Suppl.) 74 (1999) 413–425; David J. Gross, “Twenty-Five Years of Asymptotic Freedom” Nuc. Phys. B (Proc.
Suppl.) 74 (1999) 426–446; Crease & Mann SC, 329-335; CloseIP, pp. 258–276. See also note 29, p. 860 and
note 45, p. 867; Politzer, Gross, and Wilczek shared the 2004 Nobel Prize in Physics for this work. In his Nobel
speech, Wilczek refers to Coleman, who while visiting Princeton had been very helpful to him and Gross, as
“uniquely brilliant”: Frank A. Wilczek, “Asymptotic Freedom: From Paradox to Paradigm”, on-line at
https://www.nobelprize.org/nobel_prizes/physics/laureates/2004/wilczek-lecture.html.
33[Eds.] Peskin & Schroeder QFT, p. 541, equation (16.135). There appear to be 6 triplets (in three generations):
{d, u; s, c; b, t}.
34[Eds.] See also “Secret Symmetry” in Coleman Aspects, pp. 174–178.
35[Eds.] Politzer’s own account of this period is described in his Nobel lecture: H. David Politzer, “The Dilemma of
Attribution”, on-line at https://www.nobelprize.org/nobel_prizes/physics/laureates/2004/politzer-lecture.html.
36[Eds.] Close IP, pp. 261–264; ’t Hooft, op. cit., pp. 416–417. See note 32, p. 113.
37[Eds.]Sidney Coleman and David J. Gross, “Price of Asymptotic Freedom”, Phys. Rev. Lett. 31 (1973) 851–854.
Other field theories are asymptotically free in a different number of space-time dimensions, e.g., in two
dimensions, the Gross–Neveu model of Dirac fermions: David J. Gross and André Neveu, “Dynamical Symmetry
Breaking in Asymptotically Free Field Theories”, Phys. Rev. D10 (1974) 3235–3252.
38[Eds.] Peskin & Schroeder QFT, pp. 475–479.
39[Eds.] PDG 2016, Section 9.3.4, “Measurements of the strong coupling constant”, pp. 128–131, in particular, the
graph in Figure 9.4 on p. 131. Peskin & Schroeder QFT, p. 552, cite α s = 2/(4π) ≈ 0.4 at 1 GeV.
40[Eds.] Because of space and time constraints, Coleman’s six lectures on dispersion relations were not included
in this book.
41[Eds.] At the end of the last lecture, the students honor an old academic tradition: they applaud their professor.

Concordance of videos and chapters

Nearly all of the text in the chapters comes from the editors’ (RS and DD) transcriptions of the videotapes at the
Harvard Physics Department’s site https://www.physics.harvard.edu/events/videos/Phys253, with additional text from Sidney
Coleman’s original notes (1975–76), or from the sources named in the Preface. Occasionally the editors
interpolated text from these to fill out an argument or provide an insight from later versions of the course. These
interpolations are usually only a sentence or two, often relegated to the footnotes. The exceptions occur when a
lecture is fragmentary, e.g., Chapter 25. Below is a concordance to aid those who might want to watch the
Coleman videotapes as they read.
 Index

Page numbers for entries occurring in a footnote are followed by an n and the footnote number. Bold page
numbers indicate a term’s definition or an individual’s biographical sketch.

SC = Sidney Coleman.

Fµν, see field strength tensor

(2)(p, p′), two particle Green’s function, 319

(n), n-point Green’s functions, 269

GF, Fermi constant, 877

S3, symmetric group of three objects, 861

F( ), spinor field propagator, 439

U(t, t′), time evolution operator, 131

Z[ρ], see generating functional

{Z1, Z2, Z3}, see renormalization constants

2), scalar field propagator, 217

F(p

′(p2, p′2, q2), renormalized coupling constant, 496

Γ[ϕ], effective action for 1PI Green’s functions, 690

(2)(p, −p), sum of all 1PI graphs with 2 external lines, 691

(n)(p
1, · · ·, pn), sum of all 1PI graphs with n external lines, 690

′(p2), meson self-energy, 321

( ), see nucleon self-energy

α, Dirac matrices, 405, 407

α, fine-structure constant, 529, 749

β, Dirac matrix, 405, 407

βa(g), renormalization group, 1098

ϵµναβ, see Levi–Civita symbol

ϵijk, see Levi–Civita symbol

η, , ghost fields, 625

γ, Euler–Mascheroni constant, 530n5, 709, 731

γ(g), anomalous dimensions, 1100

γA(g), anomalous dimension, 1098

γµ, Dirac matrices, 417

γ5, Dirac matrix, 419

σµν, Dirac matrices, 419

1PI (one-particle irreducible), see Green’s function

4-vectors, classification, 4

Abelian gauge theory, 665, 1033, 1035, 1062, 1085

canonical quantization, 1035

covariant derivative

comparison with Yang–Mills, 1018

Faddeev–Popov quantization, 1033

Feynman rules, 625n10, 1036

functional integration, 1034

generator, 1061

Higgs model, 949, 1004, 1012–1016, 1051, 1054

BRST transformation, 1043

Feynman rules, 1014

renormalization, 1043

Weinberg’s bound on Higgs mass, 1085

Lagrangian, 674

Lie group, 646n5

massive

kµkν/µ2 in propagator contributes nothing, 645

not self-sourced, 1022

structure constants vanish, 625n10, 1036

weak hypercharge, 1061

Abers, Ernest S., xxxiii

Abraham, Max, 207

Abrikosov, Alexei A., 667n12

accidental symmetry, 991–993

aces, see Zweig, George

Achasov, Nikolay N., 998n27

Achilles (Zeno paradox), 197

action, 578, 604, 609, 610, 623, 625, 632

and discrete symmetry, 118–130

and Ward identity, 719

boundary terms, 79
classical, 604

gauge field, 701

complex fields, 111, 145

definition, 58

dimensions, 538, 546, 657

Dirac, 403

effective, 682, 691

ghost variables, 625

effects of gauge transformation on, 697–699

Faddeev–Popov ghost, 665

first-order form, 632, 642

gauge invariant, 662, 664, 697, 1031, 1032

gauge-fixing, 697

generating functional, 687

Hamilton’s Principle, 58

Hamiltonian form, 623, 633, 642, 669

Lagrangian form, 623, 642

loop expansion, 687, 688

Lorentz invariance, 64

Proca, 614, 633

propagator from, 691

quantum, 693

scalar fields, 65

scale invariance, 146

second-order form, 632, 634

spinor, 397, 632

spinor electrodynamics, 699

stationary, 657

unchanged by divergence, 84

units, 101

Weyl, 398

with source term, 968

adiabatic function, 143, 275

and scattering, 240

behavior and counterterm, 184–186

constructing S-matrix without, 278

extended, 195

Fourier transform, 195

tends to a delta function, 189

in exponential, 188

in Model 2, 183

mostly harmless, 184

not needed in Model 1, 154

removed via counterterm, 211

removing, 257

required in Models 2 and 3, 155

adiabatic theorem, 184, 185

adjoint representation, 947n19, 1035

Adler’s rule on soft pions, 899–902, 918, 919, 928

definition, 901

guts graphs, 900

pole graphs, 900

Adler, Stephen L., 890

anomalies, 1043n20

argument with Low and SC about PCAC, 891

current algebra, 902n23

soft pions, 899

Adler, Steven L.

current algebra, 877n2

Adler-Weisberger relation, 877n1

Affleck, Ian, xxxvii

Aitchison, Ian J. R., 1011n3, 1061

Ajzenberg-Selove, Fay, 507n3

Alexandrov, B., 940n11

Alger, Horatio, 858

Ali, S. Twareque, 624n9

Ambler, Ernest, 121n8, 239, 880n11

Ampère’s Law, 103, 558

amputated external legs, 690

Amsterdam (1971): ’t Hooft reports Yang–Mills renormalizable, 1045

analytic term, 760

Andersen, Carl M., 781n4

Anderson, Philip W., 967n5, 1014, 1024n20

Andromeda galaxy, 10, 32, 33, 293

angular momentum

commutation relations, 374

angular momentum, conservation of, 81

annihilation and creation operators

charged scalar fields, 107

charge conjugation, 119

commutation relations, 108

Fermi fields, 430

scalar field, 17, 26, 34, 42

action on vacuum, 27

commutation relations, 26, 37

commutator, 240

contraction, 156

in field expansion, 42

Lorentz transformation, 29–30

multi-particle states, 292

normal ordering, 74–75

parity, 122

translation invariance, 30

under combined PT, 130

scalar fields

pair creation, 71

scalar fields, under SO(2)

commutation relations, 107

spinor field

anticommutation relations, 432

charge conjugation, 467

in field expansion, 399

parity, 461–462

under PT, 474–475

Weyl field, 400

anomalies, 82

Adler-Bell-Jackiw, 1043

anomalous dimension, see γA(g), anomalous dimension

anomalous magnetic moment

electron, in QED, 30, 743–749, 751

higher order corrections, 751–752

Schwinger’s interest in, 736n8

small contribution from muon, 756

electron, in quantum mechanics, 736–743

leptons, hadron contribution to, 756–757

muon, in QED, 752–756

contribution from a “heavy photon”, 755–756

larger contribution from electron, 756

anti-linear operator, 128

anti-unitary operator, 127

multiplication table, 127

anticommutator, 408

antinucleon–meson scattering, 442

Aoyama, Tatsumi, 752n4

Appel, Walter, 954n30

Arkani-Hamed, Nima, xxix

Arnowitt, Richard L., 662n6

Arnowitt–Fickler gauge, see gauge, axial

Arrgh!, 43

Artin, Michael, 907n35

Ashcroft, Neil W., 707n8, 936n3, 1011n2

Ashmore, Jonathan F., 528n4

asymptotic freedom, 1112, 1114

averaging over spin states

photons, 653

spinors, 449

Avogadro, Amadeo, 936

axial gauge, see gauge, axial, 1031

and canonical quantization ⇔ F-P ansatz, 668

good for canonical quantization, 668

poor for calculation, 670

axial vector, 550, 569, 796, 861, 889

contribution to neutron beta decay, 890

decay constant, 887

Dirac bilinear, 420, 469

matrix elements, 886

meson, 887

partially conserved current (PCAC), 892

axial vector current, see also PCAC, 880, 885, 886, 891n4, 900, 1043

π–π pole, 926

and Cabibbo angle, 883

and Goldberger–Treiman relation, 898

charge-changing, 1060

commutator with vector current, 902

commutator with weak interaction Hamiltonian, 934

commutators, 906, 918

conserved, if pion mass were zero, 892

construction, 995

divergence, 887

divergence in massless theory, 998

in gradient-coupling model, 895

in sigma model, 993–994, 997

matrix element, 885

scale, 898

transformation, 996

triplet of, 893, 894

zero momentum transfer, 901

Babenko, Victor A., 886n30

bacteria model

and renormalization group equation, 1103–1105

Baker, Henry F., 181

Baker-Campbell-Hausdorff formula, 181

Balian, Roger, 655n22, 867n45, 1100n17

bar–star rule, 469

Bardeen, William A.
QCD, 867n45

Barger, Vernon, 1066

Bargmann, Valentine, 370n2

Barnes, Virgil E., 850n13

Barshay, Saul, 783

Barton, Gabriel, 779n2

Barut, Asim O., 557n3, 823n6

baryon number, 520

commutes with isospin, 520

baryon spin- octet

irreducible representation of some group G?, 782–783

masses equal, in the absence of EM?, 781–782

Bass, Ludvik, 568n13

Bateman, Harry, 146

Bazhanov, Vladimir V., 696n15

Becchi, Carlo M., 1043

Becker, Richard, 207n6

Behrends, Ralph E., 787n31

Belinfante, F. J.

symmetric energy-momentum tensor, 89n7

Bell, John S.

anomalies, 1043n20

Benfey, O. Theodor, 16n15

Benson, Katherine, xxxvii

Berestetski , Vladimir B., 654n15, 742n19

Berezin, Felix A., 434n3

integrals of Grassmann variables, 618n3

Beringer, Juerg, 753n7

Berkeley, George (Bishop of Cloyne), 537

Bernstein, Jeremy

current algebra, 877n2

Goldstone theorem, 955n31, 957n34

π decay, 887

simplified Goldberger–Treiman derivation, 894

spontaneous symmetry breaking, 935n1

Bessel function, 14

Bessis, Daniel, 994n23

β(g), beta function

effects of zeros in β(g), 1106–1110

ϕ 4 theory example, 1101

β > 0, 1101

QCD example

β < 0, 1113

consequences of β < 0, 1113–1115

QED example, 1102

β > 0, 1102

consequences of β > 0, 1113

zeros in β(g), 1106

β decay, see nuclear β decay

beta decay, see nuclear β decay

Bethe, Hans A., 2n2, 326n7, 736n8

Coulomb interaction from photon exchange, 650n11

Bhansali, Vineer, xxxvii

Bianchi identities, 103

Biedenharn, Lawrence C., 375n11

Big Bang, 959

Bjorken scaling, 1095, 1096

Bjorken scaling variable x, 1096

Bjorken structure functions Fi(k2, x), 1095

Bjorken, James D., xxxiii, 3n3, 1080n8

alias B. J. Bjørken, 1080n8

proposes charm (with Glashow), 1080n8

thesis (1959), 900n18

Bloch, Felix, 197n12

Block, Richard, 783

Bogoliubov algorithm, see BPHZ algorithm

Bogoliubov, Nikolai N., 533, 535, 537

Bohr magneton, 743

Bohr, Niels, 15n14

haphazard reality, 449n6

opposition to Feynman diagrams, 215n11

Bollini, Carlos Guido, 528n4

Boltzmann constant, 285, 695

Boltzmann, Ludwig, 96n14

Born approximation, 190, 222, 223, 225–227, 230, 231, 342, 762

electron scattering off neutron, 763

pole term, 760

Born perturbation theory, 71

Borsanyi, Szabolcs, 508n5

Bose

lines, 533

operators, 588

particles, 70, 298, 370, 991

pronunciation, 18

propagator, 534

Bose fields, 671

BPHZ algorithm, 531

charged, in Feynman diagrams, 622

classical quantities in functional integrals, 617

effective potential, 988

functional integral, 616, 617, 621

det A−1/2, 622

observables commute at equal times, 434

parity, 460

quantized with commutators, 431

Wick’s theorem, 436

Bose statistics, 18

automatic in many-particle scalar theory, 224

automatic with Bose quantization, 28

exploited in occupation number labeling, 21

makes Yukawa scattering symmetric, 225

Pauli–Villars regularization, 714

pion scattering, 927

pion–pion scattering, 926

Poisson distribution, uncorrelated states, 171

SU(3) Clebsch–Gordon series, 815

wave function symmetric, 19

Boswell, James, 381n17

bottom quark, 1082n9

bound states, 141

“Bourbaki”, “Nicolas”, 783n14

BPHZ algorithm, 531–533, 537, 540, 541, 673, 1094

step by step, 533

BPHZ renormalization, see also BPHZ algorithm, 533

Breit, Gregory, 360n4, 838n32

Breit–Wigner

formula, 360–362

peak, 367

bremsstrahlung, 734n3

Bressani, Tullio, 998n27

Brink, David M., 193n11, 507n2

Brose, Henry L., 583n3

Brout, Robert, 869n54, 1014

Ising model, 963n1

phase transitions, 936n5

Brown, James Ward, 14n12, 179

Brown, Laurie M., 225n2, 599n1, 625n10, 736n8, 867n43

BRST transformation, 1043n19

bubkes, 739, 1068

Bucksbaum, Philip H., 877n3, 879n9

Bugg, David V., 921n5

Burden, Conrad J., 696n15

Burton, David M., 907n35

Butkov, Eugene

variational method, 979n27

Byron, Frederick W., 271n2, 1023n16

Cabibbo

angle, 882, 894, 1071, 1082

currents, 907

currents in GSW model, 1083

Cabibbo theory

consistent with lepton-hadron universality, 907

Cabibbo, Nicola, 882, 1082n9

theory of weak currents, 882–883

Cahn, Robert N., 782n8

Callan, Curtis G.

thanks SC for help with Callan–Symanzik equation, 1100n17

traceless energy-momentum tensor, 89n8

Callan–Symanzik equation, 1100

Campbell, John E., 181

Cannell, D. Mary, 206n2

canonical commutation relations

massive vector field, 563

scalar quantum field, 69

canonical momentum

classical mechanics, 59

in Noether currents, 84

scalar field, 65

canonical quantization, 57, 577, 586

massive vector field, 561–565

massless vector field

complications due to gauge invariance, 586–587

quantum mechanics, 61–63

scalar field, 69–70

vector field, 555

Caprini, Irinel, 933n23

Carruthers, Peter A., 807n15

Cartan, Élie, 809n16, 1017n11

Carter, Antony A., 921n5

Carter, Janet R., 921n5

Casimir’s trick, 449, 820

Casimir, Hendrik B. G., 449n6

Cauchy’s theorem, 54, 192

Cauchy, Augustin-Louis, 58n2

Cayley, Arthur, 58n2

Chang, Darwin, 1084n15

charge conjugation

and Dirac bilinear products, 468–470

and Fermi fields, 465–468

and multiparticle eigenstates as even or odd, 121

and photon, 735

and scalar fields, 119–121

unitary operator U C = U †C for scalar fields, 120, 146

charge, electric

as generator of transformations, 110

conservation, 83–86

ensures áψñ = 0, 301

local vs. global, 83

defined as an integral, 83

of a Noether current, 86, 115

does not commute with isospin, 520

from SO(2) Noether current, 107

under Lorentz transformations, 115–117

universality, 675

physical interpretation, 706–707

preserved by renormalization, 705–706

charged scalars

quartic self-interaction, 707

charged vector bosons, 1067

charm, 1080

needed to suppress strangeness-changing neutral currents, 1080

no observable phases in GIM charm isodoublet, 1081

proposed by Bjorken and Glashow, 1080n8

charmed current, 1082

Charybdis, see also Scylla, 957n34

chemical potential, 17

Chen, Bryan Gin-ge, xxix, xxx, 525n1

Cheng, Kuo-Shung, 589n7, 623n7

Cheng, Ta-Pei, xxxiii

Cherenkov radiation, 171

Chevalley, Claude, 829

Chew, Geoffrey F., 342

chiral symmetry, 543

algebra, 906

broken by mass, 945

chirality principle, 906

fermions, 945

generated by axial vector current, 993

chiromancy, 906

Chitwood, Daniel B., 878n5

Christenson, J. H., 240n9

Christoffel symbol

analogous to connections over vector bundles, 660n2

Churchill, Ruel V., 14n12, 179

Cirlot, J. E., 16n15

CKM mechanism, 1082n9

and CP violation in GSW model, 1082n9

Clark, Allan, 803n9

classical field, 694, 701

classical scalar field, (x)

from functional integral, 694

Clebsch–Gordan

coefficients

η decays, 777

Fermi interaction from GSW model, 1071

GMO formula, 848

permutation symmetry and direct products, 864

pion–nucleon scattering, 517, 552

selection rule, hadronic EM processes, 768

SU(3), branching ratios, 857

SU(3), representations, 865

Yukawa interaction, 795

series, 380, 810

isospin, 778, 802, 807, 825n10

SU(3), 848
SU(3), Coleman’s algorithm, 810–813

Clifford algebra, 407, 711

Clifford, William Kingdon, 407

Close, Frank, xxxiii

’t Hooft announces QCD β > 0, 1115n36

cofactor, 848

coherent states

harmonic oscillator, 172

in Model 1, 171

Coleman, Diana, v, xxx, xxxvii

Coleman, Sidney, v, xxxiii, 315n1, 517n8, 868n47

“A man tired of group theory. . . ”, 381n17

argument with Adler and Low about PCAC, 891

article on GSW Nobel, 936n6

bereft of PDG booklet, 517n8

Bjorken and Drell “the best available”, 3n3

Callan thanks SC for help with CS equation, 1100n17

Clebsch–Gordan series for SU(3), 810n18

Coleman-Glashow formula, 835n23

course content remarks, 845n1, 1116

Delphic Oracle comparison, 654n17

dimensional regularization, 710

effective potential for massless scalar field, 973n14

experimental validation of theory, 217n15

false vacuum, 978n26

Feynman vs. Schwinger, 225n2

Feynman credited with m2 for bosons in GMO formula, 846n5

first paper: “Good enough. . . ”, 767n31

“Fourier space” for “momentum space”, 558n10

“Fun with SU(3)”, 788n33, 797n1, 828n16

God’s units for EM, 576

golden age of a physicist, 784

“If you can solve QCD, why are you here?”, 738n15

local gauge invariance not a symmetry, 579n2

man in the magnet, parable of the, 936–938

Mathews and Walker textbook, SC thanked, 788n33

“Modesty forbids, but honesty compels. . . ”, 841n40

Pauli’s Relativity recommended, 583n3

perturbative spontaneous symmetry breaking, 968n6

Politzer and asymptotic freedom, 1115

Politzer, in Nobel lecture: SC “my beloved teacher”, 860n29

QCD nomenclature and Bozo the Clown, 868n48

students’ golden experience in Physics 251, 192n10

“’t Hooft”, how to pronounce, 1045

theorist’s career and harmonic oscillator, 21n4

thesis (Caltech, 1962) under Gell-Mann, 784n23

traceless energy-momentum tensor, 89n8

“Weyl”, how to pronounce, 394n1

Wilczek, in Nobel lecture: SC “uniquely brilliant”, 1113n32

color, see quarks, 905

“color photons”, see gluons

color confinement, see infrared slavery

Columbia University

and I. I. Rabi, 752n6

J. Schwinger’s 1948 talk on QED, 743

Lorentz lectures on electrons, 207n5

Commins, Eugene D., 877n3, 879n9

commutation

with p, q and differentiation by conjugate, 63

complete, 60

completeness

of p’s and q’s, 60

completeness relation

Dirac equation, 422, 560

massive vector field, 560, 564, 571

complex fields, 109–113

variations δψ and δψ ∗ treated as independent, 111

Compton scattering, 43n10, 573, 646, 652

e± off p or n, 758

massive photon, 650–651

Compton wavelength, 99, 568

electron, 568

experimental bound on photon mass, 654

proton, inverse, 2

unit of distance for interaction, 15

Yukawa meson, 193

Condon, Edward U., 805

Condon–Shortley phase convention, 805

disobeyed by the K0, 806

conformal group, 146

connected graph, 165

number of loops, 688

connected Green’s functions, 272

great theorem, 688, 906

conserved current, 580

and gauge invariance, 583

coupling to massive vector field, 575

generated by gauge transformation if L is gauge invariant, 578

massive vector field

required for limit m → 0, 567

related to gauge invariance, 578

conserved quantity

and Noether’s Theorem, 79

as generator of transformations, 81–82, 110

conserved vector current, see weak interactions, CVC hypothesis

constancy of center of energy motion

partner to angular momentum conservation, 97

continuous transformations

in classical mechanics, 77

scalar field, 105–109

contraction of two fields, 156

cookies, 87

Cool, Rodney L., 838n36, 950n25

Coolidge, Julian Lowell, 434n3

Corben, Herbert, 736n8

Correspondence Principle, 62, 63, 68

cosmological constant, Λ, 73

de Coulomb, Charles-Augustin, 733n1

Coulomb scattering

massive photon, 646–650

zero-mass limit, 647–650

Coulomb’s Law, 568

counterterm, 186

in pseudoscalar-nucleon theory, 485

diagram in Model 2, 187

in Model 2, to remove spurious phase, 186

in Model 3, 208

Model 2’s as ground state energy, 186

coupling constant renormalization, 306

in pseudoscalar-nucleon theory, 496–500

Model 3, 338–342

Courant, Richard, 191n7, 1103n21

covariant derivative, 1013

Abelian gauge group, 582

GSW model

scalar field, 1061

non-Abelian gauge group, 1018

covariant gauge

Feynman gauge, 1032

Landau gauge, 1032

covering group, 791

CP violation

and the CKM mechanism, 1082n9, 1084n15

in K decays, 240

not in 4 quark GSW model, 1084

CPT invariance, 238

consequences of violation, 121

S-matrix inv’t ⇒ H inv’t, 545

CPT theorem, 239

and Fermi fields, 476

and spinor Hamiltonian, 545

violation, consequences of, 240

Crease, Robert P., xxxiii

creation operator, see annihilation and creation operators

Cromer, Alan H., 117n5

Cronin, James W., 240n9

crossing symmetry, 231, 237

cryptorenormalizable theories, 1026n25

crystallization

and spontaneous symmetry breaking, 970

cum grano salis, see salt, grain of

Cunningham, Ebenezer, 146

Curie temperature, 938

current

electromagnetic, 83

Dirac field, 514, 580

electron, 469

hadronic, 764

Model 1, 153

Proca theory, 577

scalar field, 584

scalar field, improved, 585

Y-component, 787

hypercharge jµY, 764

under C, 765

under G-parity, 765

isospin

nucleon, 514

isospin, z-component jµIz, 764

under C, 765

under G-parity, 765

Noether, 84

conserved, 85

not uniquely defined, 86–87

current algebra, 877, 902

and pion–hadron scattering, 908

equal-time commutators, 902

CVC hypothesis, 995

Czarnecki, Andrzej, 471n6

Cziffra, Peter, 342n5

d’Alembert equation, 663n8

d’Alembertian, 5, 51, 607, 665

Dalitz plot, 235, 245, 256, 777n1, 779, 780

η → 3π’s, 779

η decay, 778–781

Dalitz, Richard H., 235n6, 842, 843

Das, Ashok, 371n3

Dashen, Roger F., 877n2

current algebra, 902n23

Davies, Christine T. H., 1083n14

Dayan, Moshe, 784

de Broglie, Louis, 558n8, 568n13

de Swart, Johan J., 810n19

de Wit, Bernard, 407n1

de-Shalit, Amos, 867n45

decay processes, 245

deep Euclidean region, 1092

deep inelastic electroproduction, 1096

degenerate vacua, 942

Delbrück scattering, 713

Delphenich, David H., 375n11, 671n18

Delphi, Oracle of, 654

Dennery, Phillipe, 14n12, 191n7, 599n2

density of final states, 244

three particles, derivation, 254–256

two particles, 249, 250

Derbes, David, xxxi, 656n23

derivative coupling, 303, 304, 308, 546

Feynman rule, 902

guess, 305
naïve rule justified, 643

stated, 309

verified for ′(2) in ϕ 2 theory, 304–306

verified for ′(4) in ϕ 4 theory, 306–307

Yang–Mills fields, 1040

generating functional

Hamiltonian form, 623

Lagrangian form, 623

ghost field example, 626–628

in scalar electrodynamics, 615

invariant under p → −p, 238

minimal coupling prescription, 584

no problem with linear, 623

non-renormalizable between scalars, 1062

parity violation, 238

particle language advantageous, 611

Proca Lagrangian, 570

pseudoscalar-spinor example, 587

quadratic has problems, 623

renormalization constants, 323

renormalization needs quartic interaction, 708

scalar electrodynamics, 641

superficial degree of divergence D, 534

technical issues, 277, 587

via functional integrals, 617, 622

Deser, Stanley, 43n10, 533n10, 579n2, 1037n10

determinant of a differential operator, calculation of, 608n7

Dettman, John W., 271n2

DeWitt, Bryce S., 625n10, 1023n15, 1037

and ghost particles, 625n10

DeWitt-Morette, Cécile, 1037n10

diagonalization theorem, lepton terms in GSW model, 1072

proof, 1073

Dickens, Charles, 858

dielectric breakdown, 984

Diestel, Reinhard, 689n6

differential cross-section, dσ/dΩ, 240, 246

relativistic vs. non-relativistic, 251, 254

two particles, 250

differential form, 557n6

differential transition probability, 243

digamma function, 530n5

dilations, 146

dimension

powers of mass, 529

dimensional analysis, 288

anomalous dimensions, 1108

beta functions, 1098, 1100

energy dependence of 1PI graphs, 1092, 1095

Goldstone model, 957

M dependence of 1PI’s, 1103

M value irrelevant, 1109

mp-mn difference, 508

muon anomalous magnetic moment, 753, 756

Planck length, 537

scalar field theory in d dimensions, 99

variance in d dimensions, 99

dimensional regularization, 528–531

Dimock, Jonathan

Dirac sea, 990n14

Dirac adjoint, 416

bilinear forms, 468

building IR’s in SU(3), 805

ghost field, 617n2

Lorentz transformation, 416

of (s) u(r) equals ( (s) u(r))†, 449

ψ, have opposite charge, 580

spinor solutions of D. equation, 421

under parity, 419

Dirac algebra, 407, 408

Dirac basis for α and β, 407–409

Dirac bilinear products, 418

Gordon decomposition, 740–741

under charge conjugation, 468–470

bar–star rule, 469

under parity and Lorentz transformations, 420

Dirac delta function, 26, 46, 64, 69

Dirac equation, 369, 386, 402–406, 558, 685

and Klein–Gordon equation, 418

and electron magnetic moment, 736n8, 743

connection with Klein-Gordon equation, 405

Dirac basis, 408

Gordon decomposition, 1007

helicity eigenstates, 425

invariance under PT, 472–474

plane wave solutions, 409–412

propagator, 439

standard form, 418

statement, 406

Weyl basis, 406

Dirac field, 413

canonical anticommutators, 432

canonical quantization, 429–434

Dyson’s formula, 436

electromagnetic current, 580

Fourier integral expansion, 430

Gross–Neveu dynamical symmetry breaking, 1115n37

interaction with Proca field, 569, 911

Lagrangian, 580

minimally coupled to photon, 583

Pauli term, 585

time ordering, 435

Wick’s theorem, 434–437

Dirac γ matrices, 342, 417

algebra independent of basis, 415

algebra, summary, 418

γ5 trick for traces, 428

in n dimensions, 711–712

Majorana representation, 464

properties, 417–418

slashed notation, 418

trace identities, 425

trace of an odd number of γ’s is zero, 428, 990n13

Dirac Lagrangian

and chiral symmetry, 945–946

and nonconserved current, 591

construction, 402–406

criteria, 403

spinor electrodynamics, 644

standard form, 418

Dirac notation, 375

Dirac picture, see also interaction picture, 133, 144n9

Dirac sea, 990, 1048

Weinberg quotes Schwinger, 990n14

Dirac spinors, 409–412

completeness, 422

“large” and “small” components, 742

normalization, 410

orthogonality, 421

projection operators, 423

transformation under CPT, 479–480

transformation under PT, 475–476

Dirac, Paul A. M., 2n1, 10n9, 12n11, 617n1, 655n20

and Feynman’s path integral, 136n2, 656n23

Coulomb interaction from photon exchange, 650n11

dispersion relations, 1116n40

distributions, 28

divergence

index δi of, 534, 538, 551

infrared, 74
superficial degree D of, 532, 534, 545, 551

d dimensions, 545

formula, 535

ultraviolet, 74

divergence, infrared

in Models 1 and 2, 196

divergence, ultraviolet

in Model 2, 193

Dow Jones average, 838

Doyle, Arthur Conan, 383n18

Dreitlein, Joseph, 787n31

Drell, Sidney D., xxxiii, 3n3, 738n13

Dresden, Max, 867n43

dual of a tensor, 385

Dubna (1964): Smorodinski , SU(3), and SC, 848

duck soup, 296

Duffin, Richard J., 558n8

Duffin–Kemmer–Petiau field, 558n8

Dyson’s formula, 136, 212, 214, 216, 586, 588

and disconnected graphs, 274, 604

and effective potential, 975

and Fermi fields, 434

and generating functional, 628

and Green’s functions, 274

and S-matrix elements, 153, 656

and Wick diagrams, 158

and Wick’s theorem, 155

equivalence to functional integral, 609

equivalent to Hamiltonian form of functional integral, 628–631, 643

meson–nucleon scattering, 441

Dyson, Freeman, xxx, 136n2

and renormalization constants, 280n5

and Ward’s identity, 704n3

meson–nucleon interactions, 554

on Schwinger’s Columbia talk, 743n26

Eckart, Carl

Wigner–Eckart theorem, 920

Eddison, E. R., 15n15

Eden, Richard J., 985n6

effective action, 968, 1002, 1039

and renormalization, 970

classical action in tree approximation, 969

counterterm, 702

gauge field, 701

gauge transformation, 699

in spontaneous symmetry breaking, 969

loop expansion, 969

semi-classical expansion, 969

substituted for classical action, 969

effective action and gauge transformation, 699

effective action, Γ[ ], 690

effective potential, 972, 1002

agreement with quantum correction to ground state energy, 987

and Dyson’s formula, 975

and Wick’s theorem, 975

calculation for scalar field, 973–978

generalized to many fields, 977

calculation for Yang–Mills fields, 1045–1048

physical interpretation for factor of 3, 1048

effects of fermions, 988–993

determines true vacuum, 992

equivalent to free scalar field’s zero-point energy, 988

heuristic aspects, 983–988

not gauge invariant, 1046

one-loop correction

importance to accidental symmetry, 991–993

importance to Yukawa coupling, 991

physical meaning, 978–982

V( ) = E0, 979

eichinvarianz, 582
Einstein summation convention, 3, 59

Einstein’s equations, 73

Einstein, Albert, 557n4

and causality, 32

and cosmological constant, 73

and nonexistent headstone, 96

general relativity, 407n1

greetings to Laue, 118n5

on G. Green, 206n2

electromagnetic current, see current, electromagnetic

and magnetic moments, 766

electromagnetic form factor, 738

baryons, 835

Dirac, 738

interpretation, 740

Pauli, 738

interpretation, 740

electromagnetic interactions, 575

Coulomb’s Law, 733–735

coupling matter to photons, see also minimal coupling, 582–586

second-order hadronic processes, 767

Emmerson, John McL., 848n9

energy

conservation of, 80

imaginary part as sign of instability, 985

positivity of, massive vector field, 563

energy-momentum tensor

Belinfante-Rosenfeld, 89

Callan, Coleman, and Jackiw, 89

canonical, 88, 89

Englert, François, 869n54, 936n5, 1014, 1067n15

Engliš, Miroslav, 624n9

entropy, connection with 1PI generating function Γ[ϕ], 695

equal time anticommutator, Fermi field, 432

equal time commutator, 45

Bose field, 46

Erasmus, Desiderius, 867n44

Erice, Sicily, xxix, 464n3

Ericson, Torleif, 886n30

eta meson η, 846

decays into 3π’s, 767–769, 777–781

part of JP = 0− octet, 805

ratio of decay modes, 780

Euclidean generating functional

free field theory, 606

Euclidean space, 604

Euler, Leonhard, 58n2

Euler–Lagrange equations

as constraints, 631

classical mechanics, 59, 79, 81

complex scalar fields, 111, 112

Dirac field, 430

massless vector field, 1034

Proca field, 558, 577, 652

first order L, 642

QED, first order L, 668

scalar field, 66, 67, 69, 84

set of scalar fields, 896

exchange operator, 224

exchange potential, 224, 227, 228, 231

with energy-dependent range, 229

exclusion principle

enforced by antisymmetry of Dirac operators, 433

experimentum crucis, 904

Explorer 12 satellite, 568

Faddeev, Ludvig D., 618n3, 623n8, 625n10, 655, 657, 659, 1037

massless vs. massive Yang–Mills theories, 1044n22

Faddeev–Popov ansatz, see Faddeev–Popov prescription

Faddeev–Popov prescription, 659, 664, 1031

and Yang–Mills theory, 1044

applied to QED, 665–668

in axial gauge, 668–669

applied to Yang–Mills fields, 1031–1035

determinant, 680

effective action, 666

equivalent to canonical quantization, 668–669

finite-dimensional version, 659–661

in axial gauge, 669

non-Abelian gauge theory, 1033

non-gauge-invariance of, 1036

summary, 657

Faessler, A., 883n20

fairy tales, 215

false vacuum, 978

in sigma model, 1000

Faraday’s Law, 103

Faraday, Michael, 707

Fearnley-Sander, Desmond, 434n3

Feinberg, Harvey M., 867n44

Feller, William, 791n36

Fermi

constant GF, 877, 883

in GSW model, 1068

fields, 430, 434, 436, 445, 468, 541, 588, 617, 621, 673, 687, 706, 988, 1061, 1063, 1064

and charge conjugation, 465–468

and combined PT, 472–476

and Wick’s theorem, 434–437

bp(r) and cp(r) under PT, 475

commuting P and C, 472

fields, and parity, 459–463

fields, and time ordering, 435

fields, regulator, 714

lines, 533, 545, 988

operators, 431, 433

particles, 448
propagator, 534

statistics, 429, 430, 442, 445, 448, 626

theory of weak interactions, 877

unit of distance as an energy, 2

Fermi field, see also Dirac field

unobservable, because of anticommutation, 434

wrong statistics, 622

Fermi fields, 431

1PI graphs, 491

classical, as Grassmann variables, 616

functional integral, 616, 621

det A1/2, 622

functional integrals, 617

LSZ formula, 484

Wick’s theorem, 436

Fermi sea, 18

Fermi statistics

Pauli–Villars regularization, 714

SU(3) Clebsch–Gordon series, 815

Fermi, Enrico, 587n6, 655n20, 877n4

and canonical quantization of electrodynamics, 586n6

Coulomb interaction from photon exchange, 650n11

PCAC, 894, 898

Fermi–Yang model, 894

compared with gradient-coupling model, 898

fermion propagator F( ), 439

renormalized, in terms of two functions, 489

spectral representation, 489

Feshbach, Herman, 191n7, 227n3, 867n45

Feynman diagram, 212

also called Feynman graph, 205

analytic function of p if m ≠ 0, 532

called “drawings” by Feynman, 225n2

catalog of F.d.’s in Model 3, 218–220, 225–231

checking derivative rule, 305

connected, 621, 688

counterterm, 702

distinct from Wick diagrams, 158, 162

early use in textbooks, 198n12

easy to make mistakes, 493

external lines off mass-shell, 258

external source, 258

factors of 2π, 10

from (4)(k ), 269

i

history, 159n4

LSZ for Fermi fields, 484

medium-strong interaction, 830

meson–nucleon scattering, 440, 454

NN scattering, 455

perturbation theory origin, 131

perturbative determination of counterterm, 314

proof of dim. reg. not spoiling gauge inv., 710

quadratic in p ⇒ inv’t if p → −p, 238

representation of matrix elements, 159

SC’s time convention, 215n12

Schwinger on, 225n2

sum from functional integral for Bose fields, 622

sum from functional integral for Fermi fields, 621

sum of all = exp(sum of connected), 688n3

sums give Green’s functions, 298

true for sum ⇏ true for individual diagram, 653

zero point energy, 988

Feynman gauge, see gauge, Feynman

Feynman graph, see Feynman diagram

Feynman invariant amplitude A, 214

and s-wave scattering length, 911, 913

at threshold, 911

Compton scattering, massive photon, 651

meson decay into Goldstone bosons, 1004

QED, e-e scattering

Coulomb vs. Feynman gauge, 680

unchanged by p → −p, 261

Feynman parameters, 326

for many denominators, 334–338, 745

integration over, 327–330

more than one denominator, 335

more than one loop, 335–338

Feynman propagator

massive vector field, 570

“Feynman gauge”, 672

“Landau gauge”, 672

photon

Feynman gauge, 667

Landau gauge, 667

R ξ gauges, 666

spectral representation, 754

scalar field, 217

position space, 607

renormalized, 319

spectral representation, 319

spinor field, 439

renormalized, Model 3, 324

renormalized, pseudoscalar-nucleon, 488–491

spectral representation, 489–491

Yang–Mills field

R ξ gauges, 1038

Feynman rules, 586, 588

Abelian Higgs model, 1015

electromagnetism

technical problems, 575

fermions, 443–446

heads before tails ψ, 446–448

minus sign for closed fermion loops, 445

massive vector meson, 568–571, 613–615

Model 3, 215
QED, massive photon, 641–646

QED, massive photon, introduction, 632–634

QED, massless photon (Feynman gauge), 669

scalar electrodynamics, massive photon, 644

spinor electrodynamics, massive photon, 644

Yang–Mills fields, 1037–1041, 1042

Feynman slash notation, 418

Feynman, Richard P., 58n2, 623n8, 656n23, 667n11, 736n7

and anomalous magnetic moment, 736

and CVC, 880

and Gell-Mann–Okubo formula, 846n5

and ghost particles in quantum gravity, 625n10

and quantum gravity, 1037

and renormalization, 715

and “swanky new scheme” for integrals, 326n7

and V − A form of weak current, 880

at Pocono conference, 215n11

on Feynman diagrams, 158n4

on Schwinger’s Washington talk, 736n8

path integrals, 599

propagator modification, 526

propagators as Green’s functions, 271n2

quoted by SC: “you don’t understand nuttin’”, 225

sum over histories, 656

weak interactions, 877

Yang–Mills as a prelude to gravity, 1023n15

Feynmanian, 1036

fiber bundles, 783

analogous to Christoffel symbols in GR, 660n2

Fickler, Stuart I., 662n6

field strength tensor, vector field

Abelian, 557, 742

and Maxwell’s equations, 99, 102–103, 557–558

non-Abelian, 1021, 1034, 1035

Figg, Kristen M., 838n33

fine-structure constant, 749, 751, 878

uncertainty in α −1, 751n2

Finnegans Wake, see Joyce, James

first-order form, 1033

first-order Lagrangian, 668

Fitch, Val L., 240n9

Fock space, 17

analogy with harmonic oscillator, 25–28

and Lorentz transformations, 29

continuum, 27

defined by canonical commutation relations, 43

degenerate, 108

Hamiltonian, 72

in a box, 240–243

kets, 19

Lorentz transformations, 35

occupation number representation, 20–21

occupation numbers, 20

of all particles, 142

operator algebra, 44

Fock, Vladimir A., 17n1, 650n11

Follana, Eduardo, 1083n14

Föppl, August, 207n6

form factor, see electromagnetic f.f. or weak interaction f.f.

four-potential Aµ, see vector field

Fourier space, 169, 196, 298, 739, 972

synonymous with momentum space, 558n10

Fourier transform, 5, 169, 229, 269, 295, 298

Parseval’s theorem, 191, 242

Fπ, pion decay constant, 885

Frampton, Paul H., 670n16, 867n43

Frankel, Theodore, 660n2

Fried, Herbert M., 667n13, 704n4

Friedman, Jerome I., 1096n13

Friedman, John B., 838n33

Fritzsch, Harald

QCD, 867n45

Fritzsche, B., 784n17

Fronsdal, Christian, 781n5, 787n31, 797n1

Fubini, Sergio

simplified Goldberger–Treiman derivation, 894

fudge factor, 854n20

Fujikawa, Kazuo, 666n9

full Green’s functions, 687

Fuller, Robert W., 271n2, 1023n16

fully massless theory, 1097

functional, 58

functional integral, 575, 599, 602

equality of forms, 641

Feynman rules, 611

Feynman rules for massive vector mesons, 613–615

ghost variables, 625

Hamiltonian form, 623

functional integrals

and Feynman rules, 611–613

functional methods in quantum field theory

comparison with statistical mechanics, 695–696

(2)(p, p′), two particle Green’s function in terms of (p), 319, 691

(n), n-point Green’s functions, 269

G-parity, 523

and EM contributions to hadron processes, 764–766

violation, 764n23

Γ[ ], effective action for 1PI Green’s functions, 690

“marvelous property” = sum over tree graphs only, 691–692

generating functional, 690

(2)(p, −p), sum of all 1PI graphs with 2 external lines

in terms of (p), 691

(n)(p
1, · · ·, pn), sum of all 1PI graphs with n external lines, 690

γ(g), anomalous dimensions, 1100

in solution to RG equation, 1107–1108

reason for name, 1108

γA(g), anomalous dimension, 1098

γA in QED example, 1102

γψ in QED example, 1102

ϕ 4 theory example, 1101

gamma function Γ(z), 528

Gamow, George, 74n9

gander, see goose

Garden of Eden, 345

Gasser, Jürg, 933n23

Gatto, Raoul, 867n45

gauge

Arnowitt–Fickler, see gauge, axial

axial, 662

Coulomb, 577, 579, 662

covariant R ξ, 666

Feynman, 667

Landau, 667

and effective potential for a gauge theory, 1046

and photon spectral representation, 754

Lorenz, 577, 579, 662

canonical quantization of QED in, 586n6

radiation, see gauge, Coulomb

Yennie–Fried, 667

gauge boson, 868

gauge condition, 586

axial gauge, 662

Coulomb gauge, 586, 662

Faddeev–Popov prescription, 662

Lorenz gauge, 586, 662

gauge field theory

loophole to Goldstone’s theorem, 1011

gauge fields, non-Abelian, see Yang–Mills fields

gauge group, 867

simple, 1023

gauge invariance, 575, 577–579

and QED renormalization, 675

constructing an invariant Lagrangian, 579

distinguished from internal symmetry, 579

electromagnetism, 577

if broken only by mγ ≠ 0 then ∂µJµ = 0, 578

Lagrangian, 579

local, 1017

gauge phantom, 1012

gauge transformation, 558, 696

does not commute

with normal ordering, 975

electromagnetism, 577

including matter fields, 582

passive interpretation only, 583

gauge-invariant cutoff, 709

Gauss’s Law, 558

gedanken experiment, 15

Gell-Mann λa matrices, 807

and weak interaction currents, 882

Gell-Mann, Murray, 517n8, 585n5, 765n25, 781n5, 807n15, 810, 1060

and CVC, 880

and Eightfold Way, 514

and hypercharge, 520n10

and Rosenfeld tables, 786n29

and strangeness, 781

and V − A form of weak current, 880

coins term “quantum chromodynamics”, 867n43

Coleman’s advisor, 225n2

compulsory strong interactions, 985n5

current algebra, 906

first print appearance of “quark”, 858n25

Gell-Mann–Okubo mass formula, 848n8

names “color”, 866n40

proposes quarks are physical entities, 801n6

QCD, 867n45

renormalization group, 1091n1

search for G, 783–784

sigma model, 994

simplified Goldberger–Treiman derivation, 894

symmetries of the strong interactions, 781

weak interactions, 877

Gell-Mann–Nishijima relation, 521, 1060, 1078

and electromagnetic interactions of hadrons, 764

and weak form factors, 881

generalized, 1082

GSW model, 1069

Gell-Mann–Okubo mass formula, 843, 848

and JP = 0− pseudoscalar meson octet, 851–853

and JP = 1− vector meson octet, 853–857

singlet-octet mixing, 853–857

and JP = ½+ baryon octet, 849

and JP = + baryon decuplet, 849–851

derivation, 845–848

general relativity, 579, 582

compared with Yang–Mills, 1022–1023, 1037

Einstein’s equations, 73

energy, 73

gauge invariance, 1016n9

speculative anticipation by Clifford, 407n1

generalized Pauli principle, 518

generating function, 972

generating functional

defined in terms of the effective potential, 692

perturbative spontaneous symmetry breaking, 968

generating functional Z[ρ], 270

1PI graphs, 968

and Green’s functions, 687

as a functional integral, 604

equivalent to Dyson’s formula, 609

in quantum mechanics, 610–611

sample calculation, 606–608

constrained variables, 631–632

derivative interaction

Hamiltonian form, 623

Lagrangian form, 623

Euclidean, 606

Hamiltonian form

verified, 628–631

naïve Feynman rules, 632–634

Z[S2nd] ≡ Z[SH], 634

generating functional, Γ[ ] for 1PI Green’s functions

defined in terms of functional Taylor expansion, 690

Taylor coefficient , 691

Taylor coefficients , 690

generating functional, Γ[ϕ]

quantum action, 692

generating functional, iW[J]

functional Taylor expansion, 688

generating functional, Z[J]

functional Taylor expansion, 687

generator of transformation

and conserved quantity, 82

Georgi, Howard M., 373n5, 782n8, 809n16, 823n2, 1035n5, 1084n17

Georgi–Glashow model, 1059n3

Georgi–Glashow model, 1059, 1084

Gershtein, Semën S., 880n14

Gevorkyan, Sergey R., 933n23

ghost field, 625, 632, 665, 731, 1036

action, 1036

decouples in QED, 666

derivative interaction

example, 626–628

history, 625n10, 1037

in QED, 725, 732

lift determinants into exponentials, 625

Pauli–Villars regulator fields

scalars with odd signs, 712

spinor fields obeying Bose statistics, 714

propagator, 626, 732, 1038, 1042

ghost particles, 625

history, 625n10

Giambiagi, Juan José, 528n4

Gibbs, Josiah Willard, 695

GIM mechanism, 1079

Glashow, Sheldon Lee, 315n1, 485n1, 1060, 1084

27-fold way, 850n12

and charm, 1079

and Coleman-Glashow formula, 835n23

Georgi–Glashow model, 1059n3

GSW model, 1023, 1059

SC article, 936n6

introduces mixing angle, 1070n18

Lie groups, 783

on charm, 1080n8

proposes charm (with Bjorken), 1080n8

Glashow–Salam–Weinberg model, see GSW model

Glashow-Iliopoulos-Maiani mechanism, see GIM mechanism

Glauber, Roy J., 175n1

Gledhill, David, 800n3

global symmetry group

precursor to SU(3), 783

properties, 785–787

ruled out by hypercharge reflection, 787

glueballs, 868

gluons, 868

as “color photons”, 867

quarks bound by exchange of, 905

Goldberger, Marvin L., 227n3, 887, 889, 920n1

scattering length, 908n36

Goldberger–Treiman relation, see also weak interactions, Goldberger–Treiman relation, 887, 889, 898

and sigma model, 993

and strength of interactions, 894–895

in sigma model, 998–999

Goldhaber, Alfred S., 568n14

Goldstein, Herbert, xxxiii

Goldstone boson, 944

eaten by gauge boson in Higgs mechanism, 1014, 1024

signature of spontaneously breaking continuous symmetry, 944

Goldstone bosons

pions as approximate G. b. in sigma model, 999

Goldstone theorem

gauge theory loophole, 952

proof, 951–953

Goldstone’s theorem, 949

Goldstone, Jeffrey, 940n11, 944, 993

asks S. Weinberg about “pseudo-Goldstone”, 977n21

Goldstone theorem, 944

goose, sauce for, 472n7

Gordon decomposition, 740–741

Gordon, Walter, 43n10, 740

Gottfried, Kurt, 920n1

gradient-coupling model

and meson–nucleon scattering, 546

and nucleon–nucleon scattering, 546

and PCAC, 895–899

and spontaneous symmetry breaking, 994

Gradshteyn, Izrael S., xxxiii

grain of salt, see salt, grain of

grand canonical ensemble, 17

Grassmann variable, 434, 465, 617

algebra, 617–618

calculus, 618–620

combine like normal-ordered Dirac fields, 468

functional integral (Gaussian), 621–622

Gaussian integrals, 620

integral table, 619

integration ≡ differentiation, 620

model Fermi fields, 616

n-dimensional measure, 620

Grassmann, Hermann, 407n1, 434n3

gravity, see general relativity

Green’s function

and Feynman diagrams, 258, 259

connected, 687

one-particle irreducible (1PI), 321

generating functional Γ[ ], 690

topological definition, 689

Green’s function,

conventions, 269

Green’s function, , 267

Green’s functions, 271

as expectation values of Heisenberg fields, 274

as sums of Feynman graphs, 278n4

in Heisenberg picture, 274

Green, George, 205, 259, 271

Greenberg, Oscar W.

parastatistics, 866n40

Greiner, Walter, xxxiii, 180, 884n22

universality of weak interactions, 881n15

Griffiths, David J., xxxi, xxxiv, 180, 221n17, 742n23, 985n8

Grisaru, Marc, 533n10, 579n2

groovy, 257, 536, 1112

Gross, David J., 860n29, 867

and Coleman on asymptotically free theories, 1115

asymptotic freedom, 1113

current algebra, 902n23

first print appearance of “QCD”, 867n43

Grossman, Bernard, xxxvii, 817

group generators

spontaneously broken, 1025

unbroken, 1025

GSW model, 1059–1084

and Fermi constant GF, 1068

baryon, lepton number independently conserved, 1079

covariant derivative

scalar field, 1061

spinor fields, 1064

CP-violation via CKM mechanism, 1082n9

ΔY ≠ 0 neutral current naturally suppressed, 1083

electromagnetic interactions naturally conserve all quantities as observed, 1083

Fermi theory, Cabibbo universality is natural, 1083

fermion fields L, R., 1064

Gell-Mann–Nishijima relation, 1069

including u, d quarks, 1077–1079

including other lepton generations, 1072–1073

isotopic spin approximately conserved ⇐ (f1 ≈ f2), 1079

large a and small f make interactions weak, 1068

left-handed spinor field, 1063

lepton charges, 1065

lepton electromagnetic current has standard expression, 1069

lepton number as conserved charge of global U(1) symmetry, 1061

leptonic electromagnetic interactions, 1069–1070

leptonic weak interactions, 1067–1069

md = mu to O(e2), 1083

mass of leptons, 1065–1066

natural features, 1083

naturalness, 1074

no possible quark-lepton Yukawa coupling, 1079

parity conservation in EM, not in weak interactions, 1063

predictions of W±, Z0 masses, 1070

predicts parity-violating neutral current-current interactions, 1071

quark currents unaffected by SU(3)color, 1078

relations between coupling constants, 1069

renormalizable, 1074

right-handed spinor field, 1063

secretly symmetric, thus secretly renormalizable, 1059

unites electromagnetism and weak interactions, 1074

unnatural features, 1083

V − A form automatically (thus maximal parity violation), 1068

vector boson masses, 1066–1067

weak charge, 1060

weak hypercharge, 1060

weak isospin, 1060

Weinberg angle, 1070

why are the lepton masses so small?, 1075

why is GF small?, 1075

Guralnik, Gerald S., 950n25, 1014

Goldstone theorem, 951n26, 955n31

guts graph, 900

Gutsche, Thomas, 883n20

Guzman, Gregory G., 838n33

hadron, 519

Hagen, Carl R., 950n25, 1014

Goldstone theorem, 951n26

Haller, Kurt, 655n19, 671n17, 705n6

Halprin, Arthur, 781n4

Halzen, Francis, 801n5, 858n23, 1079n6, 1082n10

Hamilton’s Principle, 58, 59

and Feynman’s sum over histories, 657

Hamilton, James A., 921n5

Hamilton, William Rowan, 58n2

Hamiltonian

classical mechanics, 59

generator of infinitesimal time translations in quantum mechanics, 62

Hamiltonian density, 68

Han, Moo-Young, 866n40

harmonic oscillator, 22–25

coherent states, 172

operator formalism, 25

Hasert, Franz Josef, 893n7

Hatfield, Brian, 1037n10

Hausdorff, Felix, 181

Havil, Julian, 530n5

Hayward, Raymond W., 121n8, 880n11

He, Xiao-Gang, 1084n15

Headrick, Matthew, xxx, xxxvii, 734n3

Heaviside

θ function, 51

rationalized units, 576

Heaviside, Oliver, 5n6

Heisenberg

equations of motion, classical mechanics, 62

Heisenberg equation of motion

momentum conjugate to scalar field, 70

scalar quantum field, 69

Heisenberg equations of motion, 606

vector field, 577

Heisenberg exchange force, 936

Heisenberg ferromagnet, 935

Heisenberg picture, 61, 131–133

Heisenberg representation, 30

Heisenberg, Werner, 39n5, 140n4, 507n2

and isospin, 782

Heitler, Walter, 668n14

helicity, 400, 944

massless particles and parity, 402

helicity projection operators, 905

Helmholtz free energy, connection with generating functional iW, 695

Helmholtz, Hermann von, 695

Hepp’s theorem, 533, 536, 542, 1043

Hepp, Klaus, 293n1, 533n8

Hey, Anthony J. G., 1011n3

Hibbs, Albert R., 599n1, 656n23

Higgs boson, 1067

lower mass bound

role of false vacuum, 1088

Weinberg, via effective potential, 1084–1088

Higgs mechanism, 949, 1014, 1084

Abelian model

degrees of freedom, 1013

Goldstone boson eaten by gauge boson, 1014

Goldstone boson eaten by gauge boson, 1024

solves two massless problems, 1024

Higgs model, 1012–1016

Feynman rules, 1015, 1056

Higgs phenomenon, see Higgs mechanism

Higgs, Peter W., 1014

and Higgs boson, 1067n15

Anderson’s conjecture, 1024n20

loophole to Goldstone theorem, 952n29

higher loopcraft, see loop lore

Hilbert space, 17, 28, 118, 126, 129, 135, 138, 240, 372, 401, 451, 460, 602, 603, 607, 673, 695, 705,
830, 832, 854, 855, 964, 967

positive norm, 490

Hilbert, David, 79n2, 191n7, 1103n21

Hill, Brian, xxix, xxxi, xxxvii, 429n1, 525n1, 671n18

Hill, Daniel A., 838n36

Ho-Kim, Quang, 880n13

Hobson, Michael P., 862n34

Hoddeson, Lillian, 225n2, 867n43

Hoecker, Andreas, 753n7

Holmes, Sherlock, 383n17

Holstein, Barry R., 883n20

anomalies, 1043n20

Hooke, Robert, 218n15

Hoppes, Dale D., 121n8, 880n11

Horgan, Ronald R., 1083n14

Hornbostel, Kent, 1083n14

Houtermans, Charlotte Riefenstahl, 190n6, 650n11

Huang, Kerson, 207n4, 1109n30

renormalization group, 1100n17

Hudson, Ralph P., 121n8, 880n11

hypercharge Y, 520, 764

commutes with isospin, 521

embedded within SU(3), 785, 801

hypernuclei, 842

hyperon, 521

Iliopoulos, John, 1079

in and out states, 138–140

construction of two-particle states, 293

construction without an adiabatic function, 278

independent, 60

infrared slavery, 1116

initial value data

massive vector field, 561

massless vector field, 586

summary of theories, 562

integral table, Feynman parametrized denominators, 330

interaction

(−) scalar vs. (+) vector exchange, 650, 735

of renormalizable type, 538

catalog, 538–539

super-renormalizable, 539

Model 3, 327, 494

interaction picture, see also Dirac picture, 133

intermediate vector bosons, see weak interactions, vector bosons

internal energy, connection with classical field , 695

interval, 4

Iofa, Mikhail Z., 1043

irreducible representation, see respresentation, irreducible

Ising model, 963

isobars, see also mirror nuclei

isospace, 513
isospin, 75, 514, 764, 781

algebra, 515

with hypercharge and baryon number, 522

and charge conservation, 523

and nuclear energy levels, 507

and scattering, 516–520, 546, 551–553

commutes with hypercharge, 521

conserved in strong interactions, 519

current, 522–523

does not commute with electric charge, 520

embedded within SU(3), 785, 800

group, SO(3), 513–514

group, SO(3) (≅ SU(2), locally), 114

matrices, 513

multiplet

isospinor: nucleon, 512–515

isovector: pion, 512–515

raising and lowering operators, 515, 831

violated in electromagnetic interactions, 520

isospin symmetry, 782

isospinor, 513

isovector, 513

Ito, Daisuke, 715n19

Itzykson, Claude, xxxiv

Ivanov, Mikhail A., 883n20

Iverson, Geoffrey J., 542n13

Jackiw, Roman, 868n47

anomalies, 1043n20

current algebra, 902n23

evaluation of effective potential via functional integral, 978n22

traceless energy-momentum tensor, 89n8

Jackson, J. David, xxxiv, 153n1, 583n3, 668n14, 777n1

Jacobi identity, 904

Jaffe, Arthur, 954

Jammer, Max, 1103

Jauch, Josef-Maria, 190n6, 197n12, 650n11

Jeffreys, Bertha Swirles, 364n5, 829n17

Jeffreys, Harold, 5n6, 364n5, 829n17

Jehle, Herbert, 656n23

Jenkins, E. W., 838n36

Johnson, Dr. Samuel, 381n17

Johnson, Kenneth, 704n4

Jona-Lasinio, Giovanni, 1011

Goldstone boson, 944

Jordan, Pascual, 431n2

and Courant–Hilbert, 1103n21

Jordan, Thomas F., 25n6

Joyce, James

Finnegans Wake, origin of “quark”, 859n27

Jungnickel, Christa, 207n6

Kac, Mark, 599n2

Kaiser, David I., xxx, xxxvii, 159n4, 198n12

Kajita, Takaaki, 1066

Kaku, Michio, 690n7, 937n8

Dirac sea, 990n14

RGE vs. CS, 1100n17

Källén, Gunnar, 317n2

Källén-Lehmann spectral representation, see spectral representation

Kaplan, Hyman, 1106

Karshenboim, Savely G., 471n6

Kekulé, August, 16n15

Kellogg, Jerome M. B., 838n32, 935n2

Kelvin (William Thomson)

introduces chirality, 906n31

Kemmer, Nicholas, 558n8

Kendall, Henry W., 1096n13

Kent, Clark, see Superman

Khalatnikov, Isaak M., 667n12

Kibble, Thomas W. B., 950n25, 1014n5

Goldstone theorem, 951n26

Killing, Wilhelm, 1017n11

Kittel, Charles, 556n2

Klein, Oskar, 43n10

Klein–Gordon equation, 43, 558, 559, 561

from Heisenberg equations of motion, 70

invariant under PT, 130

Lagrangian defined, 68

trial Lagrangian, 67

Kleinert, Hagen, 530n5, 641n1

Kobayashi, Makoto, 884n23, 1081n9

Kogut, John B., 696n15

Körner, Jürgen G., 883n20

Kostadinov, Ivan Z., 940n11

Kramers, Hendrik A., 736n8

Krzywicki, André, 14n12, 191n7, 599n2

Kusch, Polykarp, 736n8

Kycia, Thaddeus F., 838n36

Lacki, Jan, 671n18

Lagrange multipliers, 60, 979

Lagrange, Joseph Louis, 58n2

Lagrangian

classical mechanics, 58

Lagrangian density, 64

Lai, C. H., 1037n10

Lamb shift, 736n8

Lamb, Willis E., 736n8

Lanczos, Cornelius, 58n2, 118n5

Landé g-factor, 743

Landé, Alfred, 743n25

Landau gauge, see gauge, Landau

Landau, Lev D., xxxiv, 79n1, 207n4, 557n7, 667n12

Landau rules, 900

Landshoff, Peter V., 985n6

Large Hadron Collider (CERN)

Higgs boson discovery, 1085n21

Laue, Max, 117n5

Lee, Benjamin W., xxxiii, 543n14, 666n9, 787n31, 994, 1043n18

sigma model, 994n23

Lee, David, xxx, xxxvii

Lee, Tsung-Dao, 764n23, 1091n1

and global symmetry, 785n24

left-handed spinors, 944

Legendre transformation, 59, 968

evaluation of Wħ, 693

Lehmann’s sum rule, 318

Lehmann, Harry E., 294n2, 317n2

Leibbrandt, George, 709n11

Leighton, Robert B., 58n2

Lepage, G. Peter, 1083n14

lepton number, 109

as conserved charge of global U(1) symmetry in GSW model, 1061

each family of l.’s has separately conserved l.n. in GSW model, 1075

leptonic decays of vector bosons, 855

Leutwyler, Heinrich

QCD, 867n45

Levi–Civita symbol, 124, 386, 712, 921, 962

invariant under SU(n), 789

Levin, Michael A., xxxin54, xxxvii

Lévy, Maurice, 1060

sigma model, 994

simplified Goldberger–Treiman derivation, 894

Li, Ling-Fong, xxxiii, 948n21, 992n16

Lie algebra, 375

adjoint representation, 947n19, 1035

Cartan–Killing metric, 1017n11

elements are algebraically closed, 947

SO(3), SU(2) share the same L. a., 791n37

structure constants, 1017

SU(3), 828

trace norm, 1017

Lie group, 375

and Yang–Mills theory, 646n5

as Yang–Mills gauge group, 1012–1013

SC on how they came into particle physics, 783–784

weight diagrams, 809

Lie, Sophus, 784n17

Lifshitz, Evgeniĭ M., 79n1, 207n4, 557n7, 654n15, 742n19

Lifshitz, Evgeniĭ M., xxxiv

Lifshitz, Ilya M., 940n11

Lighthill, Michael James (Sir James), 28n7

Lim–Lombridas, Edwin, 671n17

Lipkin, Harry J., 781n5, 858n23

Lie groups, 784n17

Liu, Jianglai, 884n24

Llewellyn Smith, Christopher H. (Sir Christopher), 867n45

local symmetry, 867

Locher, Milan P., 998n27

Loewner, Charles, 829n17

London, Fritz, 583

loop expansion, 688, 969

loop lore, 334

Lorentz gauge, see gauge, Lorenz

Lorentz group, 3, 4, 369

boost, 376

complex conjugate of D (s +,s −) ∼ D (s −,s +), 381

exchange symmetry, 382

group property, 370

group property with phase, 370

irreducible representations D (s +,s −)(Λ), 378–379

Lie algebra, 377

non-compact, 371, 376

parity turns D (s +,s −) into D (s −,s +), 382

raising and lowering operators, 378

rapidity ϕ = tanh−1(v), 376

rotation subgroup, 382

tensor representations, 383

vector representations, 383

Lorentz invariance, 2–4

measure, 9

spin zero, 9–10

Lorentz transformations, 42

complex parameters, 477

faithful representation, 371

finite dimensional representation, 371

general field, 369

infinitesimal, 92–97

spinor representations, 370

Lorentz, Hendrik A., 153n1, 207n5

Lorenz condition, 558

Lorenz gauge, see gauge, Lorenz

Lorenz, Ludvig V., 153n1

LoSecco, John, xxxvii, 517n8

Low, Francis E., 342, 758

argument with Adler and SC about PCAC, 891

low-energy theorem, 759

renormalization group, 1091n1

low-energy theorem, photon scattering

derivation, 758–763

lowering operator, 23

LSZ reduction formula, 294, 885

and Adler’s rule, 899

for Fermi fields, 484

proof, 294–298

Lüders, Gerhart, 238n7

Lurié, David, xxxiv, 899n17

soft pions, 901

Lyth, David H., 882n19, 934n26

Lyubovitskiĭ, Valery E., 883n20

MacGregor, Malcolm H., 342n5

Madam Selena, see chiromancy

magnet, SC’s parable of the man in, 936–938

magnetic moment, see also anomalous magnetic moment

magnetic moment operator, 742

Maiani, Luciano, 1079

theorem about CP-violation, 1082n9

Majorana fields, 806

Majorana representation

gamma matrices, 464

Lorentz transformations real, 465

Majorana, Ettore, 464, 806n13

Mandelstam variables, 233, 1091n3

Mandelstam, Stanley, 233n4

Mandelstam–Kibble plot, 235

Mandl, Franz, 586n6

variational method, 979n27

Mann, Charles C., xxxiii

Marciano, William J., 709n11, 753n7, 867n43

Marfatia, Danny, 1066

Margenau, Henry, 603n4

Markushin, Valeri E., 998n27

Marseille (June 1971): Veltman tells SC that ’t Hooft has a renormalizable theory of masive charged
fields, 1044

Marseille (June, 1972): ’t Hooft announces QCD β positive, 1113

Marshak, Robert E., 880n14, 950n25

Marshall, Lauriston C., 838n36

Martin, Alan D., 801n5, 858n23, 1079n6, 1082n10

Maskawa, Toshihide, 884n23, 1081n9

mass matrix, 988, 1028

mass renormalization, 205

electron theory, analysis of Lorentz and Abraham, 207

in fluid dynamics, Green’s analysis, 206

in fluid dynamics, Stokes’s analysis, 206

Model 3, 208

Mathews, Jon, 788n33, 830n19, 861n32

Matthews, Paul T., 367n6

Maupertuis, Pierre-Louis, 58n2

Maxwell’s equations, 100, 103, 558, 576

McCormmach, Russell, 207n6

McCoy, Barry M., 696n15

McDonald, Arthur B., 1066

McFee, Maggie, xxx

McKellar, Bruce H. J., 1084n15

McNeile, Craig C., 1083n14

mean field, 701

medium-strong interactions, 845

Gell-Mann’s guess, 846

Mermin, N. David, 707n8, 936n3, 1011n2

Merzbacher, Eugen, 847n6

meson self-energy, 321

calculation to O(g2), 495–496

meson–nucleon scattering, see also nucleon–meson scattering

Dyson’s formula, 441

Feynman diagram, 440

gradient-coupling model, 546

in Model 3, 341–342, 451

Messiah, Albert, 864n36, 920n1

scattering length, 908n36

method of characteristics, 1103

renormalization group equation, 1103, 1106

method of stationary phase, 364

applied to radioactive decay, 365–366

evaluation of Wħ, 693

metric tensor, 3

n dimensions, 710

“Mexican hat” potential, 941

Mills, Robert L., 646n5, 1016

minimal coupling, 575, 579, 582–585

Abelian Higgs model, 1012

prescription, 583

succeeds ⇐ Lm has int. symmetry/conserved Jµ, 579

why “minimal”, 585

minimal subtraction, 531

mirror nuclei, 507, 508

Misner, Charles W., 1022n14

missing box, method of, 47, 57

mixing angle, 855

Model 1, scalar field with c-number source ρ(x), 153

“electrodynamics with scalar current”, 153

and massless vector field, 566

average energy, 172–173

average momentum, 173

coherent states, 171

exact solution, 167–168

generating functional as Z[ρ], 606–608

probability of n mesons as Poisson distribution, 171

Model 2, scalar field with c-number source ρ(x), 154

“quantum meso-statics”, 154, 192

ground state energy, 190–191

and Yukawa potential, 193

ground state wave function, 194–195

divergent, for a point charge, 196

P(n) → 0 as ρ(x) → δ(3)(x), 196

S matrix equals 1, 187–189

ultraviolet divergence, 193

Model 3, Yukawa coupling to a complex scalar field, 154

“quantum meso-dynamics”, 154

and pseudoscalar-nucleon theory, 481–482

and Wick diagrams, 159

coupling constant and counterterm F, 339–340

coupling constant defined as 1PI, 339

definition of physical coupling constant g, 339

comparison with real nucleon–meson coupling, 340–342

Feynman rules, 215

list of counterterms, 302

meson self-energy, (k2)

analytic properties, 332–334

power series expansion, 322

to O(g2), 324–326, 331–332

meson self-energy, (p2), 321

meson–nucleon scattering, 341–342, 451

nucleon self-energy (p2), 324

nucleon–meson scattering, 228–231

nucleon–nucleon scattering, 210–214, 221–224, 341–342

sixth-order diagram, 164

perturbative determination of counterterms, 323–324

renormalization, 300

scalar field propagator for µ < 2m, 356–360

and Breit–Wigner formula, 361

and radioactive decay, 364–366

“super-renormalizable”, 327, 494

modulo, 81

Molière (Jean-Baptiste Poquelin), 373n4

Møller scattering, 647n6

momentum

conservation of, 80

momentum space

constant becomes δ(x), 740

derivatives replaced by momenta, 238

f(p) wave packet, 28

factors in Feynman rules, 644

Klein–Gordon operator in, 440

non-interacting wave packets have no common support, 292

p2 = m2, 43

solutions of differential equations, 614

waves transverse in position space are transverse in m.s., 559

Moravcsik, Michael J., 342n5

Morii, Masahiro, xxx

Morinigo, Fernando B., 1037n10

Morse, Philip M., 191n7, 227n3

Mott cross-section, 647n6

Mueller, Holger S. P., 843n45

Mukerjee, Madhusree, 866n41

Müller, Berndt, xxxiii

universality of weak interactions, 881n15

Murphy, George Moseley, 603n4

n-dimensional sphere, volume and integral, 329n10

Nakanishi, Noboru, 337n2, 668n14

Nakano, Tadao, 520n10

Nambu, Yōichirō, 866, 884n23, 899n17, 967n5, 993, 1011

Goldstone boson, 944

on color, 866n40

PCAC definition, 899

PCAC interpretation, 892

QCD, 867n45

soft pions, 901

Nambu–Goldstone mode, 993

Narayanan, Lakshmi, xxx

naturalness, 1074

Ne’eman, Yuval, 781n5, 807n15, 810, 848n8

and Eightfold Way, 514

current algebra, 906

search for G, 784

net magnetization, 936

Neuenschwander, Dwight E., 79n2

neutral vector boson, 1067

neutron β decay

as pion pole dominance, 892

Neveu, André, 1115n37

New York (January 1948): Schwinger and electron anomalous magnetic moment, 743

Newton, Isaac, 218n15

Niagara Falls and , 1107

Nichitiu, Florian G., 857n22

Nieto, Michael M., 568n14

Nishijima, Kazuhiko, 715n19, 1076n26

and Gell-Mann–Nishijima relation, 520n10, 764

and hypercharge, 520n10, 781n6

Noether current, 84

Noether’s Theorem, 79, 85

Noether, Emmy, 79n2

non-Abelian gauge field, see Yang–Mills field

non-renormalizable theories, 345

Nordsieck, Arnold, 197n12

normal order, 91

does not commute

with field shifts, 975

with gauge transformations, 975

normal ordering, 74

normal subgroup, 1023n16

nosology, 1044

nu, 356, 901

nuclear β decay, 881, 884

pion pole dominance, 891

nucleon self-energy

( ), ps-ps theory, 491

calculation to O(g2), 492–495

(p2), Model 3, 324

nucleon–meson scattering, see also meson–nucleon scattering, 440–442

in Model 3, 228–231

coupling constants compared, 342

Feynman diagram, 220

nucleon–nucleon scattering

Adler’s rule, 900–902

and π-N coupling constants compared, 342

Feynman diagram example, 447–448

gradient-coupling model, 546

guts graphs, 900

in Model 3, 210–214, 221–224, 341–342

sixth-order diagram, 164

pole graphs, 900

O’Raifeartaigh, Lochlainn, 402n6, 583n3, 1016n9

O’Reilly, Eoin P., 936n4

O(2) invariance

and charge conjugation, 119

Ōkubo, Susumu, 371n3

Gell-Mann–Okubo mass formula, 848n8

Okun’, Lev B., see Okun, Lev B.

Okun, Lev B., 153n1, 583n3, 668n14

Olive, David I., 985n6

Olsson, Martin V., 933n23

one virtual photon process, 856

one-particle irreducible (1PI) functions, see Green’s function

operator

annihilation, Fock space, 26

annihilation, on the vacuum is zero, 26

creation, Fock space, 26

Fock space, relativistic, 29–30

operator commutation relations

Dirac field, 432

massive vector field, 564

scalar field, 37

Optical Theorem, 221n16, 517, 1091n2

derivation, 252–254

ordering ambiguities, 624

orthogonality theorem for groups, 827

Osterwalder, Konrad, 954

ouroboros, 15

Pagels, Heinz R., 867n43

pair model, 199

pair production, 2, 15

Pais, Abraham, 765n25, 850n13

and global symmetry, 785n24

Pal, Palash B., 767n29, 806n13

Paracelsus, xxxviii, 49

Parasiuk, Ostap S., 533

parity, 4, 121
and Bose fields, 460

and Fermi fields, 459–463

and unitary operator U P for scalar fields, 122, 146

scalars vs. pseudoscalars, 122

theories that do not conserve p., 485–486

vectors vs. axial vectors, 122

parity and time reversal, combined

and anti-unitary operator ΩPT for scalar fields, 130

Parseval’s theorem, see also Plancherel’s theorem, 191, 271

partial wave analysis, 227–230

partially conserved axial current, see PCAC

particles

mass, 938

partition function in statistical mechanics

connection with S-matrix, 167

partition function, connection with generating functional Z, 695

Patel, V. L., 568n14

path integrals, 599

Pati, Jogesh C., 1059n4, 1084n18

Pati–Salam model, 1059, 1084

Patrignani, Claudia, xxxiv

Pauli form factor, 738, 740, 755

Pauli principle

generalized to include color, 868

violation in naïve quark model, 866–867

Pauli σ matrices, standard representation, 400

Pauli term, 585, 740

Pauli’s theorem on Dirac matrices, 413

Pauli, Wolfgang, 39n5, 238n7, 375n11, 413n6, 526n3, 583n3, 585n4, 622n5, 709n12, 715

on what God hath put asunder, 557n4

Pauli–Villars regularization, see regularization, Pauli–Villars

Pauline, Perils of, 198n12

PCAC, 890

and gradient-coupling model, 895–899

in the sigma model, 998–1001

Nambu interpretation, 892

slow variation of pion matrix element, 890

Peierls, Rudolf, 235n6

Pendleton, Hugh, 533n10, 579n2

Perkins, Donald H., 837n30

Perlmutter, Arnold, 542n13

Perlmutter, Saul, 74n9

Peskin, Michael E., xxxiv

Peter-Weyl theorem, 823n6

Petermann, André, 1091n1

Petersen, Priscilla C., 838n35

Petiau, Gérard, 558n8

Petrov, Nikolaĭ M., 886n30

Pevsner, Aihud, 846n4

phase space, 244

phonons, 556n2

photon

mass, 566–568

and black body radiation, 568

experimental evidence, 568

odd under charge conjugation, 735

remains massless post renormalization, 705

ρ meson as heavy photon, 723

photon-induced strong interaction corrections, 763

pion pole dominance

as an explanation for PCAC, 891

pion–hadron scattering

amplitudes and scattering lengths, 918–921

and Adler’s rule, 918–928

and current algebra, 908

Weinberg–Tomozawa formula, 925

without current algebra, 917–921

pion–nucleon scattering, see also pseudoscalar-nucleon theory

amplitudes, 516–520

coupling constants, 509–512

pion–pion scattering

Weinberg’s analysis, 926–933

pions

as approximate Goldstone bosons, 993

decay

Fermi-Yang model, 888

Goldberger–Treiman relation, 885

decay constant Fπ, 885

form an isotriplet, 515

in sigma model, approximate Goldstone bosons, 999

π0 → γγ allowed, 766

under C, 765

under G-parity, 523, 765

Pitaevskiĭ, Lev P., 654n15, 742n19

Plancherel’s theorem, see also Parseval’s theorem, 191n7

Planck Law, 568

Planck length, 537

Pliny the Elder, 867n44

Pochodzalla, Josef, 842n44

Pocono (March 1948)

Bohr opposes Feynman, 215n11

importance of Schwinger’s talk, 736n8

Podolsky, Boris, 650n11

Podolsky, Daniel, xxxin54, xxxvii, 569n15, 648n10

Poincaré group, 4, 105, 146

point-splitting, 904

Poisson brackets, 61

Poisson distribution, 171, 195

Pokorski, Stefan, 684

polarization vectors, 556

orthonormal, for Proca field, 559

pole graph, 900

pole term

from Born approximation, 760

Politzer, H. David, 860n29

and SC re QCD β, 1115

asymptotic freedom, 1113

cites SC in Nobel speech, 860n29

Nobel lecture, 1115n35

QCD, 867n45

Polkinghorne, Rev. Dr. John C., 985n6

Pollack, Gerry, 94

Poole, Charles P., Jr., xxxiii

Popov, Victor N., 625n10, 655, 657, 659, 1037

Poquelin, Jean-Baptiste, see Molière

position operator, unsatisfactory in relativistic quantum mechanics, 10–15

positronium

decay of, 471

Preskill, John, 699n18, 703n2

Ising model, 963n1

Primakoff effect, 781

Primakoff, Henry, 781n4

Pritchard, Jimmy, 1098n16

Proca equation, 558, 559, 576, 583

and Klein–Gordon equation, 559

Proca field, see vector field (massive) Proca, Alexandru, 558n8

projection operator

Dirac spinors, 423

harmonic oscillator, 25

role in propagators, 615, 655, 667

vector, longitudinal, 614

vector, transverse, 614

propagator, see Feynman propagator

proper diagram, see Green’s function, (1PI)

ps-ps theory, see pseudoscalar-nucleon theory

Pseudo-Dionysus, 992n18

pseudo-Goldstone bosons, 992

pseudoscalar

Dirac bilinear, 420

pseudoscalar-nucleon theory, 481–500

coupling constant renormalization, 496–500

interaction via SU(2) tensor methods, 792–795

PT invariance

and Fermi fields, 472–476

and scalar fields, 129–130

compatibility with Lorentz invariance, 129–130

Pythia, see Delphi, Oracle of

q-numbers and c-numbers, 617n1

QCD, see quantum chromodynamics

quantum chromodynamics, 867

features in brief, 869

introduction, 866–870

lattice QCD, 1083n14

quantum electrodynamics, 696

and external c-number current, 733

failure of naïve canonical quantization, 654–655

gauge requiring ghost field, 725

low order computations, 646

proof of renormalization, 701

scalars, 641

via functional integrals, introduction, 654–658

quantum mechanics, 61

quark model, 895, 904

three-quark bound states, 861

quarks, 801

color, 866

confinement, 1116

evidence of pointlike particles from deep inelastic scattering at SLAC, 1096

first print appearance of, 858n25

flavor, 866

naïve quark model, 858–866

correctly predicts baryons as singlets, octets, decuplets, 859

correctly predicts vector bosons and pseudoscalar mesons, 861

named by Gell-Mann, 859

properties
chosen to fit the JP = ½+ baryon octet, 803

table of JP = 0− baryon octet, 805

table of JP = ½+ baryon octet, 781

table of properties, 803

weak currents

unaffected by color, 1078

quasilocal operators, 964

Quigg, Chris, 867n43

Rabi, Isidor I., 736n8, 838n32, 935n2

Schwinger’s thesis advisor, 752n6

“Who ordered that?” re: muon, 752

Rackham, Harris, 867n44

Raczka, Ryszard, 823n6

radiative corrections, 734

raising operator, 23

Rajasekaran, G., 842n44

Ramond, Pierre, 530n5

Ramsey, Norman F., 935n2

rapidity ϕ = tanh−1(v), 376

Rarita, William, 558n8

Rarita–Schwinger field, 558n8

raw dimension, 538

Rayleigh–Ritz method, see variational method

reactions, exothermic vs. endothermic, 250

Reed, Michael C., 953n29

regularization, 526

dimensional regularization, 528–531, 709–712

gauge invariant, 710

minimal subtraction, 531

Pauli–Villars, 526

applied to fermions, 713

prescription, 714–715

regulator fields, 527, 712–715

Reid, Constance, 1103n21

Reif, Frederick, 695n14

Reinhardt, Joachim, xxxiiin54

relativistic causality, inconsistent with single-particle quantum theory, 16

relativistic scalar fields

conditions to be met by, 34

translations, 36

renormalizable, 198

renormalizable vs. non-renormalizable theories, 343–346

distinction between QFT and NRQM, 540

renormalizable theories, 346

spin-0 and spin-½, 538

strictly renormalizable, 540

renormalization

QED

counterterms, 674

gauge invariance, 675

introduction, 673–676

spinor fields, 482–488

unaffected by spontaneous symmetry breaking, 971

renormalization constants

Z1, charge renormalization constant, 280

and Ward’s identity, 704

Z2, spinor wave function r.c., 280, 300

Z2 real and positive, 484

and Ward’s identity, 704

gauge-dependent, 704

in pseudoscalar-nucleon theory, 484

Z3, meson wave function r.c., 280

and derivative coupling, 304

as vacuum dielectric constant, 707

in pseudoscalar-nucleon theory, 482

QED, 704–707

counterterms, 705–706

scalar electrodynamics

additional quartic counterterm, 707–708

renormalization group, 1091, 1100

anomalous dimensions, γA, 1098

asymptotic freedom, 1106, 1112–1115

definition, 1114

beta functions, βa, 1098

E dependence of all quantities determined by {βa, γA}, 1103

effects of zeros in β(g), 1106–1110

f ∼ (E/M)γ in ϕ 4 theory, 1108

leading logs, 1110–1112

ϕ 4 theory, 1091–1094

ϕ 4 theory example, 1101

physical masses must be zero, 1098

powers of ln(E/M), 1106

QED example, 1101–1102

renormalization point at some M ≠ 0, 1097

running coupling constant , 1105

IR stable fixed point, 1109

UV stable fixed point, 1108

UV unstable fixed point, 1109

summing the leading logs, 1106

renormalization group equation, 1100

solution, 1103–1105

renormalized field, 280

representation, 371

direct product, 380

direct sum, 372

reducible, 372

spinor, 370

representation, irreducible, 372

abbreviated IR, 796

Retherford, Robert C., 736n8

Richter, Burton, 1082n12

Riemann–Christoffel tensor, 1022

Riemann–Lebesgue lemma, 278, 282, 966

Riess, Adam, 74n9

Rigden, Johh S., 984n4

right-handed spinors, 944

Riley, Kenneth F., 862n34

Riordan, Michael, 867n43

Rivier, Dominique, 526

Robertson, Howard P., 790n34, 823n5

Roček, Martin, xxx

Rochester/CERN (July, 1962)

Gell-Mann predicts Ω−, 850

Rodrigues’ formula, 374n9

boost equivalent, 377n14

Rohrlich, Fritz, 197n12

Rose, Morris E., 2n2

Rosen, Simon Peter, 779n2

Rosenbluth’s formula, 738n11

Rosenfeld, Arthur H., 786n29

Rosenfeld, Léon, 15n14

symmetric energy-momentum tensor, 89n7

Rosner, Jonathan L., xxxi, 787n31, 1060n5

Ross, Leonard (Leo Rosten), 1106n26

Rosten, Leo, xxxiv

rotation group, 370, 372

and angular momentum transformation, 830n19

exchange symmetry, 381

generators, 373

irreducible representations, 373

irreducible representations and angular momentum, 373

Lie algebra, 375

representations

double-valued for half-integer s, 376

unitary representation, 371

rotational invariance, 8

rotations, 81

Rouet, Alain, 1043

Rubbia, Carlo, 31, 519n9

Rudin, Walter, 278n4

Ruegg, Henri, 671n18

Ruelle, David, 967n5

running coupling constant , 1105

IR stable fixed point, 1109

UV stable fixed point, 1108

UV unstable fixed point, 1109

Yukawa theory, 498

Rutherford scattering, 647n6

Ryder, Lewis H., xxxiv

Ryzhik, Iosif M., xxxiii

S3, see symmetric group S3

S-matrix, see scattering matrix

Saclay Institut de Physique Théorique

degenerate mass spectrum, 871

Safko, John L., xxxiii

Sakurai, Jun John, 134n1, 180, 723n31, 736n8, 766n28

27-fold way, 850n12

criticises Coleman–Glashow mass formula, 853

ρ meson as heavy photon, 723

Salam, Abdus, 1084

advisor of Ronald Shaw, 1016n9

advisor of Yuval Ne’eman, 784

and radioactive decay, 367n6

banquet table, 937n8

Goldstone theorem, 944n14, 951n26

GSW model, 485n2, 869n54, 938, 1023, 1026, 1059

SC article, 936n6

Pati–Salam model, 1059n4

Saletan, Eugene J., 117n5

Salpeter, Edwin E., 2n2

salt, grain of, 77, 327, 624, 664

Salwen, Nathan, xxx, xxxvii

Sanda, Anthony I., 666n9

Sands, Matthew, 58n2

Sato, Hiroyuki, 1076n26

scalar field

λϕ 4 interaction, 343–344

gϕ 5 interaction, 344–345

conditions required of a local relativistic, 34

effective potential, 967–978

Fourier integral expansion, 42

Goldstone model, 939–941

Higgs model, 1012–1016

Lagrangian, 64–68, 580

propagator, 49, 157

properties, 42–46

scalar potential, 557, 577

scattering

adiabatic approximation, 143–144

without an adiabatic function, 272–273

scattering length, 908, 920

scattering matrix (S-matrix), 586

defined in terms of Dyson’s formula, 155

in terms of U I(∞, −∞), 143–144, 155

non-relativistic quantum mechanics, 140

scattering matrix elements

averaging initial, summing final spin states, see also Casimir’s trick, 448

scattering theory

non-relativistic quantum mechanics, 138–140

Schachinger, Lindsay Carol, 838n36

Schiff, Leonard I., 184n2, 254n2, 864n37

Schluter, Robert A., 838n36

Schmidt, Brian P., 74n9

SCHOONSCHIP, 1040, 1044

Schramm, Stefan, xxxiiin54, 884n22

Schrödinger picture, 131

Schrödinger representation, 30

Schrödinger, Erwin, 43, 568n13

Schrödinger’s equation, 12, 20, 43, 62, 131–132, 138–142, 147–148, 364, 656n23

Schroeder, Daniel V., xxxiv

Schulte-Frohlinde, Verena, 530n5

Schur’s lemma, 25n6, 824

Schweber, Silvan S., xxxiv, 650n11

Dirac picture, 134n1

Schwinger terms, 904

Schwinger, Julian, xxxiv, 206n2, 485n1, 558n8, 736n7, 743, 749n33, 986n9, 1060

and anomalous magnetic moment, 736

and renormalization, 715

and sources, 271n2

CPT, 238n7

discards vacuum polarization, 715

Feynman diagrams “bringing computation to the masses”, 225n2

Pocono talk (1948) and anomalous magnetic moment, 736n8

sigma model, 994n23

symmetries of the strong interactions, 781

Scylla, see also Charybdis, 957n34

seagull diagram, 585

interpretation of quad-vector vertex in Yang–Mills theories, 1041

massive photon scalar QED, 644

squared, 708

Yang–Mills Feynman rules, 1042

Segrè, Emilio, 650n11, 877n4

selection rules

hadron-single photon emission, 765

hadronic processes, ΔG and ΔI, 768

hadronic weak interactions, 880

ω → γγ not allowed, 766

π0 → γγ allowed, 766

quantities conserved by strong interactions, 758

semi-leptonic weak interactions, ΔI = , 880

∑0 → Λ + γ allowed, 765

self-energy operator, 321

and , 691

corrections to, boson vs. fermion, 846

geometric series, 322

meson, calculated to O(g2), 324

meson, divergence of, 495

nucleon, divergence of, 495

photon, as vacuum polarization, 984

semi-classical expansion, see loop expansion, 969

Shaw, Graham, 586n6

Shaw, Ronald, 1016n9

Shelter Island (June 1947)

Bethe inspired to solve the Lamb shift, 736n8

Kramers suggests mass renormalization, 736n8

Shepard, James R., 843n45

Shestakov, Georgii N., 998n27

Shifman, Mikhail, 848n10, 1037n10

Shigemitsu, Junko, 1083n14

Shortley, George H., 805

sigma model, 982, 993–1002

and PCAC, 998–1001

and pions as approximate Goldstone bosons, 999

axial infinitesimal transformations of spin zero fields, 996

axial transformations, table, 997

axial vector current, fermions, 995

explicit mass term breaks chiral SU(2), 996

false vacuum, 1000

full axial vector current, 997

full vector current, 997

Goldberger–Treiman relation, 998–999

infinitesimal axial transformation, 995

infinitesimal isospin transformation, 995

invariant under chiral SU(2) ⊗ SU(2) if massless, 996

Lagrangian SO(4) invariant, 997

Lagrangian with Goldstone-Nambu potential, 997

particle spectrum, 998

PCAC term and Symanzik’s rule, 999

vector (isospin) current, fermions, 995

sign of potential and spin of exchanged quanta, 193

silly physicist, 9

Simon, Barry, 953n29

Sirlin, Alberto, 709n11

Slavnov, Andrei A., 618n3

massless vs. massive Yang–Mills theories, 1044n22

Slavnov–Taylor identities, 1043n18

Slavnov–Taylor identities, 1043n19

Smith, Jack, 407n1

Smorodinskiĭ, Yakov A., 848

SO(2)

and conservation of charge, 114

isomorphic to U(1), 113

symmetry group of electromagnetism, 114

SO(3), see also rotation group, 370

compared with SU(3), 514

isospin group, 114, 513–514

locally isomorphic to SU(2), 513, 791n37

SO(3,1), see also Lorentz group, 4

Lie algebra, 373n5

SO(4), 378

SO(n)

n(n − 1) generators, 374

Sodolow, Joseph B., 867n44

soft pions

and spontaneous symmetry breaking, 993

soft symmetry breaking, 1000

Sohn, Richard, xxx

Sommerville, D. M. Y., 329n10

spacelike separation

and commutation of operators, 31–35

spacelike vector, 42

Spanish Inquisition, 15

spectral representation, 734

photons, 754

scalar field, 317

spinor field, 489–491

spin

label for irreducible representations of the rotation group, 375

spin operator, 743

spin states, summing over

spinors, via Casimir’s trick, 449–450

vectors, 653–654

spin-statistics theorem, 622

spinless particle, 6

spinor electrodynamics, 632

Spivak, Michael, 278n4

spontaneous symmetry breaking, 935, 937

and soft pions, 993

broken generators, 947

characteristic sign, 1st version, 951

characteristic sign, 2nd version, 951

characteristic sign, sufficient but not necessary, 951

classical vs. quantum theories, 969

continuous symmetry example, 941–944

discrete symmetry example, 939–941

does not affect renormalization, 971

general case, 946–948

multiplet of scalar fields, 948–949

perturbative origins (Coleman-Weinberg), 967–978

SC and Politzer, absent fundamental scalars, 1115

simple model with no potential, 954–957

unbroken subgroup, 947

vs. manifest symmetry, 953

Yukawa coupling and fermion mass, 944–946

Srednicki, Mark, 990n14

standard basis for α and β, see Dirac basis for . . .

standard model, 1084

Stapp, Henry M., 342n5

Stein, Eckart, xxxiiin54, 884n22

Stokes, George Gabriel, 206n3

Stora, Raymond, 1043

Strachan, Charles, 877n4

Strang, Gilbert, 191n7, 1072n21

strange particles, 1077, 1079

strangeness, 520n10

strangeness-changing current, 1082

strangeness-preserving current, 1082

Streater, Ray F., 949n22

PCT, Spin and Statistics, and All That, 967n4

strong interactions

conserve isospin, 519

structure constants, 947, 1050

adjoint representation, 947n19

antisymmetric on two indices, 1023

invariant under cyclic permutation, 1019

SU(2), ϵijk, 1019

vanish for Abelian group, 625n10, 1036

Struik, Dirk J., 784n17

Stueckelberg’s trick, 671

Stueckelberg, Ernst C. G., 526, 671n18

renormalization group, 1091n1

Styer, Daniel F., 599n1, 656n23, 752n5

SU(2), 513

and G0, 785

and isospin, 114n3, 513

covering group of SO(3), 791

D (s) are irreducible, 795

dim D (s) = 2s + 1, 791

label s of D (s) ≡ (number of indices), 791

representations D (s) ≡ (s) irreducible?, 791

SU(3), 797

as group of strong interactions, 785

Clebsch–Gordan series, Coleman’s algorithm, 810–813

examples, 813–815

color, 867
and Fermi statistics, 868–869

source of strong force, 867

confirmed by Ω−, 850

conjugate representations, 798, 800

decay Σ0 → Λ + γ, 836, 841–843

determining the generators, 831–834

dim (n, m) = (n + 1)(m + 1)(n + m + 2), 799

dim (n, m) as a table, 799

electromagnetism and, 829–843

EM form factors of JP = ½+ baryon octet, 835–839

EM mass splittings of JP = ½+ baryon octet, 839–841

embedding SU(2) (isospin), 800

embedding U(1) (hypercharge), 801

Gell-Mann matrices λa, 807

Gell-Mann–Okubo mass formula, see also Gell-Mann–Okubo mass formula

Gell-Mann–Okubo mass formula applications, 849–853

Gell-Mann–Okubo mass formula derivation, 845–848

I and Y decomposition of (n, m), 807–810

graphical algorithm, 807–809

IR , 806

irreducible representations D (n,m) ≡ (n, m), 797

irreducible representations, guess, 796–797

list of predictions, 857

matrix tricks for baryons, 804

matrix tricks for mesons, 805

827

(n, m) are complete, proof, 827–829

(n, m) are irreducible, proof, 823–826

symmetry of Clebsch–Gordan coefficients, 815–816

weight diagrams, 809

27, 810

3, 3, and 8, 810

SU(n), 788

complex conjugate representation, 788

contraction (trace), 789

invariant tensors , 789

symmetric and antisymmetric representations, 789

Sudakov, V. V., 1106n25

Sudarshan, E. C. George, 880n14

Superman, 327

Susskind, Leonard, 867, 868n47

QCD, 867n45

Svartholm, Nils, 1059n1

“swell foop” = “fell swoop”, 1014

Symanzik’s rule, 542, 543

applied to sigma model, 1000

Symanzik, Kurt, 294n2, 542n13

, ground state energy density, 978n24

Callan–Symanzik equation, 1100

symmetric group S3, 861–864

even, odd, and mixed representations, 862

symmetric vacuum, 937

symmetry

intuitive concept, 78

SC definition, classical mechanics, 78

SC definition, field theory, 83

symmetry spontaneously broken

Nambu–Goldstone realization, 953

symmetry unbroken

Wigner–Weyl realization, 953

symmetry, discrete, 118

symmetry, internal, 105

SO(2), 105

SO(3), 114

SO(n), 113

symmetry, spacetime, 77–97

’t Hooft, Gerard, 485n2, 655n22, 710, 1037n10

and QCD β function, 1113n32, 1115

announces Yang–Mills renormalizable, 1045

asymptotic freedom, 1113

pronunciation, 528n4

and SC, 1045

renormalization of Yang–Mills theories, 1043n17

tadpole diagram, 315, 612

ghost field example, 626–628

Takahashi, Yasushi, 558n8, 675

Ward–Takahashi identities, 675n22

Tavel, Max A., 79n2

Taylor, John C., 877n3, 1016n9

Slavnov–Taylor identities, 1043n18

Taylor, John R., 139n3, 815n21

Taylor, Richard E., 1096n13

tensor

Dirac bilinear, 419

ter Haar, Dirk, 667n12

Terrail, Pierre, Chevalier de Bayard, 33n2

Thirring, Walter

simplified Goldberger–Treiman derivation, 894

Thomas precession, 743n24

Thomson formula, electron scattering, 763

Thorne, Kip S., 1022n14

time evolution operator, U(t, t′), 131

time-ordered product, 49, 136

time ordering, 136

time reversal, 4

and anti-unitary operator ΩPT for scalar fields, 130

anti-unitary operator, 125

time-reversal

and anti-unitary operator ΩT for scalar fields, 146

timelike vector, 42n8

Ting, Samuel C. C., 1082n12

Ting, Yuan-Sen, xxix, xxx, 525n1

Tinkham, Michael, 823n2

Tolkien, J. R. R., 15n15

Tomonaga, Shin’ichiro, 225n2, 866n41

and renormalization, 715n19

Tomozawa, Yukio, 925n9

top quark, 1082n9

totalitarian selection principle, 985

Tovey, Dan R., 777n1

transformation

SC’s requirement to be a symmetry, 78, 83

translation

space, 80

spacetime, 88

time, 80

translation invariance, 6–7

tree approximation, 689

tree graph, 689

Treiman, Sam, 887, 889

current algebra, 902n23

Trottier, Howard D., 1083n14

Tuan, S. F., 134n1

Tucker, Robert, 407n1

Turlay, René, 240n9

Tyutin, Igor V., 1043n19

U gauge, 1026

U(1), unitary group in one dimension

isomorphic to SO(2), 113

Uem, Pham Xuan, 880n13

Uncertainty Principle, 15

unit flux, 246

unitary gauge, 1026

unitary operator

sufficient conditions, 127

universality

charge renormalization, 705

electric charge, 675

physical interpretation, 706–707

weak interactions, 881

Utiyama, Ryoyu, 1016n9

vacuum as dielectric, 707n7

vacuum expectation value, 939

vacuum polarization, 715

all corrections transverse, 718

calculated via dimensional regularization, 725

calculated via Pauli–Villars, 725

history, 715

vacuum state, 19, 938

degenerate in theories with spontaneous symmetry breaking, 942, 963–967

good vacua, 965

good vacua are globally distinct, proof, 965–966

van Dam, Hendrik, 375n11

van der Meer, Simon, 31n1, 519n9

van der Waerden, Bartel L., 795

Van Hove, Léon, 867n45

variational method, 979n27

vector current

in sigma model, 997

vector field (massive)

canonical quantization, 561–565

completeness relation, 560, 564, 571

Compton scattering, 571

vanishing amplitude for longitudinal photons, 572–573

conserved current, 575

Fourier integral expansion, 563

irrelevance of (kµ kν/µ2) in propagator, 671–673

Lagrangian, 555–557

lim m → 0 exists ⇐ coupled to conserved current, 567, 576

longitudinal (zero helicity) A → 0 as m → 0, 568, 576, 651–653

vector field (massless)

and electromagnetism, 99, 102–103, 557–558

and gauge invariance, 558

and Model 1, 566

canonical quantization

technical problems, 565–566, 586–587

gauge invariance, 575

vector potential, 470, 557, 577

Velo, Giorgio, 1043n19

Veltman, Martinus, 485n2, 710

renormalization of Yang–Mills theories, 1043n17

vertex correction, 743

very-strong interactions, 845

VEV, see vacuum expectation value

Villars, Felix, 526n3, 709n12, 715

virtual particle, 214

virtual process, 217

von Laue, Max, 117n5

{W+, W−, Z0}, weak force vector bosons, 519, 1061–1063, 1066–1067

Wagner, William G., 1037n10

Walcher, Thomas, 842n44

Walker, Robert L., 788n33, 830n19, 861n32

Wanders, Gérard, 671n18

Wang, Frank Y., 79n2

Ward identity, 675n22, 704, 721

Ward, John Clive, 675, 704n3, 715

Ward–Takahashi identities, 675n22

Ward–Takahashi identities

and gauge invariance of counterterms, 701

sketch of proof, 702–703

and QED counterterms, 675

and radiative corrections from gauge-fixing term, 734n3

counterterms, 722

derivation, 696–700

derived, 687

individual Green’s functions, 715–721

original Ward identity, 675n22, 704, 721

relations among counterterms, 703

role in renormalization, 699

Warsaw/Jabłonna (1962) (GR3)

DeWitt, Feynman and ghosts, 1037n10

Watson, George Neville, 364n5

Watson, John, M.D., see Holmes, Sherlock

Watson, Kenneth M., 227n3, 517n8, 920n1

scattering length, 908n36

Wawrzyńczyk, Antoni, 823n6

weak charge

in GSW model, 1060

weak interactions, see also GSW model

V − A leptonic current, 879

axial vector current, 884

closed chiral algebra of equal time current commutators, 906

CVC hypothesis, 880–882

and universality, 881

CVC hypothesis stated explicitly, 881

equal time axial current commutators are model dependent, 903

Fermi (current-current) theory, 877

Fermi–Yang model, 894

compared with gradient-coupling model, 898

form factors, 880

gA(k2), 884

Goldberger–Treiman relation, 887, 889

gV(0), and cos θC., 884

gV(k2), analog of F1 in β decay, 884

hadronic decays, 879

lepton-hadron universality, 906

lepton and hadron algebras are the same, 907

leptonic decays, 878

π− field defined in terms of ∂µ Aµ, 885

pion β decay, 894

pion decay constant Fπ, 885

quark currents, 905

semi-leptonic decays, 879

lepton current violates parity, 880

semi-leptonic decays of the baryon octet, 884

universality, 881

vector bosons, 519, 1061–1063, 1066–1067

weak current as sum of hadronic and leptonic currents, 878

Wolfenstein parameter, 884n23

weak isodoublet, 1078

weak isosinglet, 1078

weak magnetism, 881

weak neutral current, 1070

Weierstrass approximation theorem, 829

Weierstrass, Karl, 59

Weil, André, 783n14

Weil, Simone, 783n14

Weinberg’s angle, 1070

Weinberg’s theorem on divergences, 532n7, 1092

Weinberg, Erick, 968n8

effective potential for massless scalar field, 973n14

jokingly described as “pseudo-Weinberg” by S. Weinberg, 977

perturbative spontaneous symmetry breaking, 968n6

student of SC, 977n20

Weinberg, Steven, xxxiv, 485n1, 507n3, 579n2

accidental symmetry, 991

Goldstone theorem, 944n14, 951

GSW model, 1023, 1059

SC article, 936n6

Higgs boson lower mass bound, 1085

pion scattering lengths, 925n9

“pseudo-Weinberg and pseudo-Goldstone bosons”, 977

quotes Schwinger on Dirac sea, 990n14

utility of Yang–Mills theories, 783

Weinberg–Tomozawa formula, 925

Weiner, Charles, 158n4

Weisskopf, Victor, 736n8

Wentzel, Gregor, 190n6, 199, 650n11

skeptical response to Schwinger, 715n20

Werbeloff, Marina D., xxix, xxx

Westrem, Scott D., 838n33

Weyl basis for α and β, 405–407

Weyl fields, 1063

Weyl spinors, 394

bilinear forms, 394–396

equation, 399

helicity, 400–402

Lagrangian, 399

criteria, 397

Weyl, Hermann, 557n4, 583n3, 790n34, 823n5, 829, 1016n9

algorithm for SU(3) Clebsch–Gordan series, 810

and SU(n), 788

and Emmy Noether, 394n1

proposes Weyl equation, 402

Wħ, loop expansion of functional W[J], 689

Wheeler, John A., 758, 1022n14

advisor of Feynman and Kip Thorne, 758n14

advisor of Wightman, 967n4

“First Moral Principle”, 758n14

S-matrix, 140n4

Whisnant, Kerry, 1066

White, Terence H., 985n5

Whittaker, Edmund T., 58n2

Wick contraction, 569

Wick diagram, 159

1-1 correspondence with Wick expansion terms, 161

combinatorics, 164

connected, 165

representation of operators, 159

sum of Wick diagrams as exponential of connected Wick diagrams, 165

Wick expansion, 158

algorithm for contractions, 160–161

and effective potential, 975

and statistical mechanics, 167

Wick rotation, 328, 528, 605

Wick’s theorem, 155, 157, 586, 613

and Fermi fields, 434

proof, 158

Wick, Gian-Carlo, 156, 328n9

Wiener, Norbert, 599n2

Wightman, Arthur S., 949n22, 1043n19

PCT, Spin and Statistics, and All That, 967n4

proof that good vacua are distinct, 967

student of Wheeler, 967n4

Wigner’s theorem, 129

Wigner, Eugene P., 129n10, 360n4, 431n2, 783, 823n2, 827n12, 828n13

Wigner–Eckart theorem, 920

Wigner–Eckart theorem, 847, 920

and electromagnetic current, 766

Wilczek, Frank A., 508n5, 860n29, 867

asymptotic freedom, 1113

Nobel lecture, 1113n32

describes Coleman as “uniquely brilliant”, 1113n32

Williams, R. M., 542n13

Wilson, Kenneth G., 867, 967n5

QCD, 867n45

renormalization group, 1100n17, 1108n28

Witten, Edward

current algebra, 902n23

Woit, Peter, xxxi, 315n1, 429n1, 507n1, 525n1, 699n17, 703n2, 901n20, 978n22, 986n10

Wong, Kit Yan, 1083n14

Woolcock, William S., 921n5

Wu, Chien-Shiung, 121n8, 239, 880n11

Yang, Chen Ning, 646n5, 764n23, 1016

and global symmetry, 785n24

PCAC, 894, 898

Yang–Mills fields, 625n10, 646n5, 660n2, 869

and effective potential, 1045–1048

broken vs. unbroken generators, 1025

compared with general relativity, 1022–1023, 1037

contrasted with electrodynamics

absent sources, Yang–Mills are self-coupled, 1022

coupling constant(s), 1023–1024

covariant derivative, 1018

expression for Faµν, 1020–1021

Faddeev–Popov method of quantization, 1031–1035

Feynman rules, 1037–1041, 1042

quad-vector vertex physical interpretation, 1041

tri-vector vertex physical interpretation, 1040–1041

tri-vector vertex worked out, 1038–1040

gauge transformation of Aaµ, 1019

physical interpretation, 1019–1020

Gell-Mann and Glashow, 783

Higgs mechanism provides mass, 1024

limit as m → 0, 673

quantization

electromagnetism as warm-up for, 1024

renormalization

BRST transformation, 1043

history, 1044–1045

only conjectured prior to Faddeev–Popov, 1044

significance of Faddeev–Popov, 1044

Slavnov–Taylor identities, 1043

’t Hooft-Veltman proof, 1043

Ward identities broken due to anomalies, 1043

Ward identities insufficient, 1041

R ξ gauges (’t Hooft-Lee), 1026n24

U gauge (Weinberg), 1025–1029

Weinberg discovers utility, 783

Yang-Mills fields

U gauge (Weinberg), 1066

Yennie, Donald R., 667n13

Yennie–Fried gauge, see gauge, Yennie–Fried

Yukawa coupling
SU(3) invariant

D and F terms, 807

Yukawa potential, 193, 226

via Born approximation, 223

Yukawa, Hideki, 193n11

Zacharias, Jerrold R., 838n32, 935n2

Zachariasen, Fredrik, 738n13

Goldberger–Treiman relation, 894

Zee, Anthony, xxxiv, 74n9, 650n11, 823n2, 1022n14

Zeeman effect, anomalous, 743n25

Zel’dovich, Yakov B., 880n14

Zichichi, Antonino, xxix

zilch, 96, 142, 365, 1097

Zimmerman, Wolfhart, 294n2, 533n10

Zinn–Justin, Jean, 655n22

Zinn-Justin, Jean, 1100n17

Zuber, Jean-Bernard, xxxiv

Zumino, Bruno

and Feynman, Landau and Yennie gauge names, 668n14

and gauge dependence of Z2, 704n4

current algebra, 902n23

on Nambu’s ten year lead, 866n41

Zweig, George

aces model (quark equivalent) couldn’t get published, 859n28

proposal of physical quarks (“aces”), 858n26

proposes quarks are physical entities, 801n6