Introduction to QCD∗

P. Z. Skands†

Theoretical Physics, CERN, 1211 Geneva 23, Switzerland

School of Physics & Astronomy, Monash University, Clayton VIC 3800, Australia

Updated: July, 2017

arXiv:1207.2389v5 [hep-ph] 7 Jul 2017

Abstract
These lectures were originally given at TASI 2012 and are directed at a level suitable
for graduate students in High Energy Physics. They are intended to give an introduction to
the theory and phenomenology of quantum chromodynamics (QCD), focusing on collider
physics applications. The aim is to bring the reader to a level where informed decisions can
be made concerning different approaches and their uncertainties. The material is divided
into five main areas: 1) fundamentals, 2) fixed-order perturbative QCD, 3) Monte Carlo
event generators and parton showers, 4) Matching at Leading and Next-to-Leading Order,
and 5) Hadronization and soft hadron-hadron physics.

Useful Complementary References

• Basic Quantum Field Theory: [1]
• Textbooks on QCD: [2–4]
• Jets and Jet Algorithms: [5]
• General-Purpose Event Generators: [6]
• The String Model: [7, 8]
• Underlying Event and Multiple Parton Interactions: [9, 10]
• Step-by-step PYTHIA tutorial: see “worksheet” available on the PYTHIA homepage.
• Monte Carlo methods and Random Numbers: [11, 12]
• The veto algorithm, trial showers, the Sudakov algorithm [13–15]

∗
Lectures presented at TASI 2012: Theoretical Advanced Study Institute in Elementary Particle Physics —
Searching for New Physics at Small and Large Scales. University of Colorado, Boulder, CO, June 4 – 29, 2012.
Based on lectures first given at ESHEP 2010 (Raseborg, Finland); updated for AEPSHEP 2014 (Puri, India), pre-
SUSY 2016 (Melbourne, Australia), and for MCNET 2017 (Lund, Sweden).
†
[email protected]

1
P. Skands Introduction to QCD

Contents
1 Introduction 3
1.1 A First Hint of Colour . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 The Lagrangian of QCD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Colour Factors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.4 The Strong Coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.5 Colour States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2 Hard Processes 17
2.1 Factorization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2 Parton Densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 Fixed-Order QCD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.4 The Subtraction Idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.5 Infrared Safety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3 Monte Carlo Event Generators 32

3.1 The Monte Carlo Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Theoretical Basis of Parton Showers . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Perturbation Theory with Markov Chains . . . . . . . . . . . . . . . . . . . . . 42
3.4 Decays of Unstable Particles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4 Matching at LO and NLO 45

4.1 Slicing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2 Subtraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.3 Unitarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49

5 Hadronisation and Soft Hadron-Hadron Physics 51

5.1 The String Model of Hadronisation . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.2 Soft Hadron-Hadron Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.3 Underlying Event and Multiple Parton Interactions . . . . . . . . . . . . . . . . 61
5.4 Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

References 67

Index 83

—2—
P. Skands Introduction to QCD

1 Introduction
When probed at very short wavelengths, QCD is essentially a theory of free ‘partons’ — quarks
and gluons — which only scatter off one another through relatively small quantum corrections,
that can be systematically calculated. But at longer wavelengths, of order the size of the proton
∼ 1fm = 10−15 m, we see strongly bound towers of hadron resonances emerge, with string-like
potentials building up if we try to separate their partonic constituents. Due to our inability
to perform analytic calculations in strongly coupled field theories, QCD is therefore still only
partially solved. Nonetheless, all its features, across all distance scales, are believed to be
encoded in a single one-line formula of alluring simplicity; the Lagrangian1 of QCD.
The consequence for collider physics is that some parts of QCD can be calculated in terms
of the fundamental parameters of the Lagrangian, whereas others must be expressed through
models or functions whose effective parameters are not a priori calculable but which can be
constrained by fits to data. However, even in the absence of a perturbative expansion, there
are still several strong theorems which hold, and which can be used to give relations between
seemingly different processes. (This is, e.g., the reason it makes sense to measure the partonic
substructure of the proton in ep collisions and then re-use the same parametrisations for pp col-
lisions.) Thus, in the chapters dealing with phenomenological models we shall emphasise that
the loss of a factorised perturbative expansion is not equivalent to a total loss of predictivity.
An alternative approach would be to give up on calculating QCD and use leptons instead.
Formally, this amounts to summing inclusively over strong-interaction phenomena, when such
are present. While such a strategy might succeed in replacing what we do know about QCD by
“unity”, however, even the most adamant chromophobe would acknowledge the following ba-
sic facts of collider physics for the next decade(s): 1) At the LHC, the initial states are hadrons,
and hence, at the very least, well-understood and precise parton distribution functions (PDFs)
will be required; 2) high precision will mandate calculations to higher orders in perturbation
theory, which in turn will involve more QCD; 3) the requirement of lepton isolation makes
the very definition of a lepton depend implicitly on QCD and 4) the rate of jets that are mis-
reconstructed as leptons in the experiment depends explicitly on it. And, 5) though many
new-physics signals do give observable signals in the lepton sector, this is far from guaranteed,
nor is it exclusive when it occurs. It would therefore be unwise not to attempt to solve QCD
to the best of our ability, the better to prepare ourselves for both the largest possible discovery
reach and the highest attainable subsequent precision.
Furthermore, QCD is the richest gauge theory we have so far encountered. Its emergent
phenomena, unitarity properties, colour structure, non-perturbative dynamics, quantum vs.
classical limits, interplay between scale-invariant and scale-dependent properties, and its wide
range of phenomenological applications, are still very much topics of active investigation,
about which we continue to learn.
In addition, or perhaps as a consequence, the field of QCD is currently experiencing some-
thing of a revolution. On the perturbative side, new methods to compute scattering am-
plitudes with very high particle multiplicities are being developed, together with advanced
techniques for combining such amplitudes with all-orders resummation frameworks. On the
non-perturbative side, the wealth of data on soft-physics processes from the LHC is forcing us
1
Throughout these notes we let it be implicit that “Lagrangian” really refers to Lagrangian density, L, the
four-dimensional space-time integral of which is the action.

—3—
P. Skands Introduction to QCD

“[...] It is concluded that the apparently anomalous features of the scattering can be
interpreted to be an indication of a resonant meson-nucleon interaction corresponding to
a nucleon isobar with spin 32 , isotopic spin 23 , and with an excitation energy of 277 MeV.”

Figure 1: The title and part of the abstract of the 1951 paper [16] (published in 1952) in which
the existence of the ∆++ baryon was deduced, based on data from Sachs and Steinberger at
Columbia [17] and from Anderson, Fermi, Nagle, et al. at Chicago [18]. Further studies at
Chicago were quickly performed in [19, 20]. See also the memoir by Nagle [21].

to reconsider the reliability of the standard fragmentation models, and heavy-ion collisions are
providing new insights into the collective behavior of hadronic matter. The study of cosmic
rays impinging on the Earth’s atmosphere challenges our ability to extrapolate fragmentation
models from collider energy scales to the region of ultra-high energy cosmic rays. And finally,
dark-matter annihilation processes in space may produce hadrons, whose spectra are sensitive
to the modeling of fragmentation.
In the following, we shall focus on QCD for mainstream collider physics. This includes the
basics of SU(3), colour factors, the running of αs , factorisation, hard processes, infrared safety,
parton showers and matching, event generators, hadronisation, and the so-called underlying
event. While not covering everything, hopefully these topics can also serve at least as stepping
stones to more specialised issues that have been left out, such as twistor-inspired techniques,
heavy flavours, polarisation, or forward physics, or to topics more tangential to other fields,
such as axions, lattice QCD, or heavy-ion physics.

1.1 A First Hint of Colour

Looking for new physics, as we do now at the LHC, it is instructive to consider the story of
the discovery of colour. The first hint was arguably the ∆++ baryon, discovered in 1951 [16].
The title and part of the abstract from this historical paper are reproduced in figure 1. In the
context of the quark model — which first had to be developed, successively joining together
the notions of spin, isospin, strangeness, and the eightfold way2 — the flavour and spin content
of the ∆++ baryon is:
∆++ = | u↑ u↑ u↑ i , (1)
clearly a highly symmetric configuration. However, since the ∆++ is a fermion, it must have
an overall antisymmetric wave function. In 1965, fourteen years after its discovery, this was
finally understood by the introduction of colour as a new quantum number associated with the
group SU(3) [22, 23]. The ∆++ wave function can now be made antisymmetric by arranging
2
In physics, the “eightfold way” refers to the classification of the lowest-lying pseudoscalar mesons and spin-1/2
baryons within octets in SU(3)-flavour space (u, d, s). The ∆++ is part of a spin-3/2 baryon decuplet, a “tenfold
way” in this terminology.

—4—
P. Skands Introduction to QCD

its three quarks antisymmetrically in this new degree of freedom,

∆++ = ijk | ui↑ uj↑ uk↑ i , (2)

hence solving the mystery.

More direct experimental tests of the number of colours were provided first by measure-
ments of the decay width of π 0 → γγ decays, which is proportional to NC2 , and later by the
famous “R” ratio in e+ e− collisions (R = σ(e+ e− → q q̄)/σ(e+ e− → µ+ µ− )), which is propor-
tional to NC , see e.g. [4]. Below, in Section 1.2 we shall see how to calculate such colour
factors.

1.2 The Lagrangian of QCD

Quantum Chromodynamics is based on the gauge group SU(3), the Special Unitary group in 3
(complex) dimensions, whose elements are the set of unitary 3 × 3 matrices with determinant
one. Since there are 9 linearly independent unitary complex matrices3 , one of which has de-
terminant −1, there are a total of 8 independent directions in this matrix space, corresponding
to eight different generators as compared with the single one of QED. In the context of QCD,
we normally represent this group using the so-called fundamental, or defining, representation,
in which the generators of SU(3) appear as a set of eight traceless and hermitean matrices, to
which we return below. We shall refer to indices enumerating the rows and columns of these
matrices (from 1 to 3) as fundamental indices, and we use the letters i, j, k, . . . , to denote
them. We refer to indices enumerating the generators (from 1 to 8), as adjoint indices4 , and
we use the first letters of the alphabet (a, b, c, . . . ) to denote them. These matrices can operate
both on each other (representing combinations of successive gauge transformations) and on
a set of 3-vectors, the latter of which represent quarks in colour space; the quarks are triplets
under SU(3). The matrices can be thought of as representing gluons in colour space (or, more
precisely, the gauge transformations carried out by gluons), hence there are eight different
gluons; the gluons are octets under SU(3).
The Lagrangian density of QCD is
1 a aµν
L = ψ̄qi (iγ µ )(Dµ )ij ψqj − mq ψ̄qi ψqi − Fµν F , (3)
4
where ψqi denotes a quark field with (fundamental) colour index i, ψq = (ψqR , ψqG , ψqB )T ,
γ µ is a Dirac matrix that expresses the vector nature of the strong interaction, with µ being a
Lorentz vector index, mq allows for the possibility of non-zero quark masses (induced by the
standard Higgs mechanism or similar), Fµν a is the gluon field strength tensor for a gluon5 with

(adjoint) colour index a (i.e., a ∈ [1, . . . , 8]), and Dµ is the covariant derivative in QCD,

(Dµ )ij = δij ∂µ − igs taij Aaµ , (4)

with gs the strong coupling (related to αs by gs2 = 4παs ; we return to the strong coupling
in more detail below), Aaµ the gluon field with colour index a, and taij proportional to the
3
A complex N × N matrix has 2N 2 degrees of freedom, on which unitarity provides N 2 constraints.
4
The dimension of the adjoint, or vector, representation is equal to the number of generators, N 2 − 1 = 8 for
SU(3), while the dimension of the fundamental representation is the degree of the group, N = 3 for SU(3).
5
The definition of the gluon field strength tensor will be given below in equation (11).

—5—
P. Skands Introduction to QCD

A1µ

QCD lecture 1 (p. 5) ∝ − 2i gs ψ̄qR λ1

ψqG
What is QCD Lagrangian
 + colour

 

0 1 0 0
  i  1 0 0   1 
ψ1 = − 2 gs 1 0 0
Quarks — 3 colours: ψa =  ψ2  0 0 0 0
ψ
ψqG ψqR 3
Quark part of Lagrangian:
Figure 2: Illustration of a qqg vertex in QCD, before summing/averaging over colours: a gluon
µ µ C C
in a state represented by λ1Lqinteracts
= ψ̄a (i γwith
∂µ δab − gs γin
quarks ab Aµ
tthe − m)ψ
states ψqR
b and ψqG .

SU(3) local gauge symmetry ↔ 8 (= 32 − 1) generators tab

1 . . . t8
ab
1 8
corresponding to 8 gluons Aµ . . . Aµ .
hermitean and traceless Gell-Mann matrices of SU(3),
A representation is: t A = 21 λA ,
0 1 0 1 0 1 0 1
0 1 0 0 −i 0 1 0 0 0 0 1
1 2 3 4
λ =@ 1 0 0 A, λ = @ i 0 0 A, λ = @ 0 −1 0 A, λ = @ 0 0 0 A,
0 0 0 0 0 0 0 0 0 1 0 0

√1
0 1
0 0
0 1 0 1 0 1
0 0 −i 0 0 0 0 0 0 3
5 6 7
λ =@ 0 0 0 A, λ = @ 0 0 1 A, λ = @ 0 0 −i A , λ8 = @ 0 √13 0 A ,
B C
−2
i 0 0 0 1 0 0 i 0 0 0 √3
. (5)

These generators are just the SU(3) analogs of the Pauli matrices in SU(2). By convention, the
constant of proportionality is normally taken to be
1
taij = λaij . (6)
2
This choice in turn determines the normalisation of the coupling gs , via equation (4), and
fixes the values of the SU(3) Casimirs and structure constants, to which we return below.
An example of the colour flow for a quark-gluon interaction in colour space is given in
figure 2. Normally, of course, we sum over all the colour indices, so this example merely gives
a pictorial representation of what one particular (non-zero) term in the colour sum looks like.

1.3 Colour Factors

Typically, we do not measure colour in the final state — instead we average over all possible
incoming colours and sum over all possible outgoing ones, wherefore QCD scattering ampli-
tudes (squared) in practice always contain sums over quark fields contracted with Gell-Mann
matrices. These contractions in turn produce traces which yield the colour factors that are as-
sociated to each QCD process, and which basically count the number of “paths through colour
space” that the process at hand can take6 .
6
The convention choice represented by equation (6) introduces a “spurious” factor of 2 for each power of the
coupling αs . Although one could in principle absorb that factor into a redefinition of the coupling, effectively
redefining the normalisation of “unit colour charge”, the standard definition of αs is now so entrenched that
alternative choices would be counter-productive, at least in the context of a pedagogical review.

—6—
P. Skands Introduction to QCD

Hadronic Z decay Drell-Yan DIS

γ ∗ /Z ∗
e− e+ → γ ∗ /Z 0 → q q̄ q q̄ → γ ∗ /Z 0 → `+ `− `q̄ → `q̄
∝ NC ∝ 1/NC ∝1

Figure 3: Illustration of the three crossings of the interaction of a lepton current (black) with a
quark current (red) via an intermediate photon or Z boson, with corresponding colour factors.

A very simple example of a colour factor is given by the decay process Z → q q̄. This
vertex contains a simple δij in colour space; the outgoing quark and antiquark must have
identical (anti-)colours. Squaring the corresponding matrix element and summing over final-
state colours yields a colour factor of
X
e+ e− → Z → q q̄ : |M |2 ∝ δij δji = Tr{δ} = NC = 3 , (7)
colours

since i and j are quark (i.e., 3-dimensional fundamental) indices. This factor corresponds
directly to the 3 different “paths through colour space” that the process at hand can take; the
produced quarks can be red, green, or blue.
A next-to-simplest example is given by q q̄ → γ ∗ /Z → `+ `− (usually referred to as the
Drell-Yan process [24]), which is just a crossing of the previous one. By crossing symmetry, the
squared matrix element, including the colour factor, is exactly the same as before, but since
the quarks are here incoming, we must average rather than sum over their colours, leading to
1 X 1 1 1
q q̄ → Z → e+ e− : |M |2 ∝ δij δji = Tr{δ} = , (8)
9 9 9 3
colours

where the colour factor now expresses a suppression which can be interpreted as due to the fact
that only quarks of matching colours are able to collide and produce a Z boson. The chance
that a quark and an antiquark picked at random from the colliding hadrons have matching
colours is 1/NC .
Similarly, `q → `q via t-channel photon exchange (usually called Deep Inelastic Scattering
— DIS — with “deep” referring to a large virtuality of the exchanged photon), constitutes yet
another crossing of the same basic process, see figure 3. The colour factor in this case comes
out as unity.
To illustrate what happens when we insert (and sum over) quark-gluon vertices, such as the
one depicted in figure 2, we take the process Z → 3 jets. The colour factor for this process can

—7—
P. Skands Introduction to QCD

Trace Relation Indices Occurs in Diagram Squared

a b
Tr{ta tb } = TR δ ab a, b ∈ [1, . . . , 8]

P a a a ∈ [1, . . . , 8] i j k
a tij tjk = CF δik
i, j, k ∈ [1, . . . , 3] a

P acd f bcd
a b
c,d f = CA δ ab a, b, c, d ∈ [1, . . . , 8]

  j i
taij tak` = TR δjk δi` − 1
NC δij δk` i, j, k, ` ∈ [1, . . . , 3] ∝ −1
NC (Fierz)
k ℓ

Table 1: Trace relations for t matrices (convention-independent). More relations can be found
in [3, Section 1.2] and in [1, Appendix A.3].

be computed as follows, with the accompanying illustration showing a corresponding diagram

(squared) with explicit colour-space indices on each vertex:

Z → qg q̄ : qk qk
X
|M |2 ∝ δij tajk tak` δ`i qj tajk takℓ qℓ
colours
= Tr{ta ta } δij ga ga δℓi (9)

1 qi qi
= Tr{δ} = 4 ,
2

where the last Tr{δ} = 8, since the trace runs over the 8-dimensional adjoint indices. If we
want to “count the paths through colour space”, we should leave out the factor 12 which comes
from the normalisation convention for the t matrices, equation (6), hence this process can take
8 different paths through colour space, one for each gluon basis state.
The tedious task of taking traces over t matrices can be greatly alleviated by use of the
relations given in Table 1. In the standard normalisation convention for the SU(3) generators,
equation (6), the Casimirs of SU(3) appearing in Table 1 are7
1 4
TR = CF = CA = NC = 3 . (10)
2 3
In addition, the gluon self-coupling on the third line in Table 1 involves factors of f abc . These
are called the structure constants of QCD and they enter via the non-Abelian term in the gluon
7
See, e.g., [1, Appendix
√ A.3] for how to obtain the Casimirs in other normalisation conventions. As an example,
choosing taij = λaij / 2 would yield TR = 1, CF = TR (NC2 − 1)/NC = 8/3, CA = 3.

—8—
P. Skands Introduction to QCD

A4ν (k2 )

∝ −gs f 246 [(k3 − k2 )ρ g µν

+(k2 − k1 )µ g νρ
+(k1 − k3 )ν g ρµ ]

A6ρ (k1 ) A2µ (k3 )

Figure 4: Illustration of a ggg vertex in QCD, before summing/averaging over colours: interac-
tion between gluons in the states λ2 , λ4 , and λ6 is represented by the structure constant f 246 .

field strength tensor appearing in equation (3),

a
Fµν = ∂µ Aaν − ∂ν Aaµ + gs f abc Abµ Acν . (11)
| {z } | {z }
Abelian non−Abelian

The structure constants of SU(3) are listed

in the table to the right. They define Structure Constants of SU(3)
the adjoint, or vector, representation of
SU(3) and are related to the fundamental- f123 = 1 (14)
representation generators via the commu-
1
tator relations f147 = f246 = f257 = f345 = (15)
2
ta tb − tb ta = [ta , tb ] = if abc tc , (12) 1
f156 = f367 = − (16)
2
or equivalently, √
3
f458 = f678 = (17)
if abc = 2Tr{tc [ta , tb ]} . (13) 2
Antisymmetric in all indices
Thus, it is a matter of choice whether one
prefers to express colour space on a basis of All other fabc = 0
fundamental-representation t matrices, or
via the structure constants f , and one can
go back and forth between the two.
Expanding the Fµν F µν term of the Lagrangian using equation (11), we see that there is a
3-gluon and a 4-gluon vertex that involve f abc , the latter of which has two powers of f and
two powers of the coupling.
Finally, the last line of Table 1 is not really a trace relation but instead a useful so-called
Fierz transformation, which expresses products of t matrices in terms of Kronecker δ functions.
It is often used, for instance, in shower Monte Carlo applications, to assist in mapping between
colour flows in NC = 3, in which cross sections and splitting probabilities are calculated, and
those in NC → ∞ (“leading colour”), used to represent colour flow in the MC “event record”.
A gluon self-interaction vertex is illustrated in figure 4, to be compared with the quark-
gluon one in figure 2. We remind the reader that gauge boson self-interactions are a hallmark

—9—
P. Skands Introduction to QCD

of non-Abelian theories and that their presence leads to some of the main differences between
QED and QCD. One should also keep in mind that the colour factor for the vertex in figure 4,
CA , is roughly twice as large as that for a quark, CF .

1.4 The Strong Coupling

To first approximation, QCD is scale invariant. That is, if one “zooms in” on a QCD jet, one
will find a repeated self-similar pattern of jets within jets within jets, reminiscent of fractals.
In the context of QCD, this property was originally called light-cone scaling, or Bjørken scal-
ing. This type of scaling is closely related to the class of angle-preserving symmetries, called
conformal symmetries. In physics today, the terms “conformal” and “scale invariant” are used
interchangeably8 . Conformal invariance is a mathematical property of several QCD-“like” the-
ories which are now being studied (such as N = 4 supersymmetric relatives of QCD). It is also
related to the physics of so-called “unparticles”, though that is a relation that goes beyond the
scope of these lectures.
Regardless of the labelling, if the strong coupling did not run (we shall return to the run-
ning of the coupling below), Bjørken scaling would be absolutely true. QCD would be a theory
with a fixed coupling, the same at all scales. This simplified picture already captures some of
the most important properties of QCD, as we shall discuss presently.
In the limit of exact Bjørken scaling — QCD at fixed coupling — properties of high-energy
interactions are determined only by dimensionless kinematic quantities, such as scattering
angles (pseudorapidities) and ratios of energy scales9 . For applications of QCD to high-
energy collider physics, an important consequence of Bjørken scaling is thus that the rate
of bremsstrahlung jets, with a given transverse momentum, scales in direct proportion to the
hardness of the fundamental partonic scattering process they are produced in association with.
This agrees well with our intuition about accelerated charges; the harder you “kick” them, the
harder the radiation they produce.
For instance, in the limit of exact scaling, a measurement of the rate of 10-GeV jets pro-
duced in association with an ordinary Z boson could be used as a direct prediction of the
rate of 100-GeV jets that would be produced in association with a 900-GeV Z 0 boson, and so
forth. Our intuition about how many bremsstrahlung jets a given type of process is likely to
have should therefore be governed first and foremost by the ratios of scales that appear in that
particular process, as has been highlighted in a number of studies focusing on the mass and
p⊥ scales appearing, e.g., in Beyond-the-Standard-Model (BSM) physics processes [25–28].
Bjørken scaling is also fundamental to the understanding of jet substructure in QCD, see,
e.g., [29, 30].
On top of the underlying scaling behavior, the running coupling will introduce a depen-
dence on the absolute scale, implying more radiation at low scales than at high ones. The
8
Strictly speaking, conformal symmetry is more restrictive than just scale invariance, but examples of scale-
invariant field theories that are not conformal are rare.
9
Originally, the observed approximate agreement with this was used as a powerful argument for pointlike
substructure in hadrons; since measurements at different energies are sensitive to different resolution scales,
independence of the absolute energy scale is indicative of the absence of other fundamental scales in the problem
and hence of pointlike constituents.

— 10 —
P. Skands Introduction to QCD

running is logarithmic with energy, and is governed by the so-called beta function,
∂αs ∂αs
Q2 2
= = β(αs ) , (18)
∂Q ∂ ln Q2
where the function driving the energy dependence, the beta function, is defined as

β(αs ) = −αs2 (b0 + b1 αs + b2 αs2 + . . .) , (19)

with LO (1-loop) and NLO (2-loop) coefficients

11CA − 4TR nf
b0 = , (20)
12π
2 − 10T C n − 6T C n
17CA 153 − 19nf
R A f R F f
b1 = 2
= . (21)
24π 24π 2
In the b0 coefficient, the first term is due to gluon loops while the second is due to quark ones.
Similarly, the first term of the b1 coefficient arises from double gluon loops, while the second
and third represent mixed quark-gluon ones. At higher loop orders, the bi coefficients depend
explicitly on the renormalisation scheme that is used. A brief discussion can be found in the
PDG review on QCD [31], with more elaborate ones contained in [3,4]. Note that, if there are
additional coloured particles beyond the Standard-Model ones, loops involving those particles
enter at energy scales above the masses of the new particles, thus modifying the running of
the coupling at high scales. This is discussed, e.g., for supersymmetric models in [32]. For the
running of other SM couplings, see e.g., [33].
Numerically, the value of the strong coupling is usually specified by giving its value at the
specific reference scale Q2 = MZ2 , from which we can obtain its value at any other scale by
solving equation (18),
1
αs (Q2 ) = αs (MZ2 ) Q2 , (22)
1 + b0 αs (MZ2 ) ln M 2
2 + O(αs )
Z

with relations including the O(αs2 ) terms available, e.g., in [3]. Relations between scales
not involving MZ2 can obviously be obtained by just replacing MZ2 by some other scale Q02
everywhere in equation (22). A comparison of running at one- and two-loop order, in both
cases starting from αs (MZ ) = 0.12, is given in figure 5. As is evident from the figure, the
2-loop running is somewhat faster than the 1-loop one.
As an application, let us prove that the logarithmic running of the coupling implies that
an intrinsically multi-scale problem can be converted to a single-scale one, up to corrections
suppressed by two powers of αs , by taking the geometric mean of the scales involved. This
follows from expanding an arbitrary product of individual αs factors around an arbitrary scale
µ, using equation (22),
n
Y   2 
µ 2
αs (µ1 )αs (µ2 ) · · · αs (µn ) = αs (µ) 1 + b0 αs ln + O(αs )
i=1
µ2i
   
µ2n
= αsn (µ) 1 + b0 αs ln + O(αs2 ) , (23)
µ1 µ22 · · · µ2n
2

— 11 —
P. Skands Introduction to QCD

0.6

α s (Q)
αs (Q)
0.5
2-loop
1-loop
0.4

0.3

0.2

VINCIARO OT
0.1

0
0 0.5 1 1.5 2
log (Q/GeV)
10

Figure 5: Illustration of the running of αs at 1- (open circles) and 2-loop order (filled circles),
starting from the same value of αs (MZ ) = 0.12.

whereby the specific single-scale choice µn = µ1 µ2 · · · µn (the geometric mean) can be seen to
push the difference between the two sides of the equation one order higher than would be the
case for any other combination of scales10 .
The appearance of the number of flavours, nf , in b0 implies that the slope of the running
depends on the number of contributing flavours. Since full QCD is best approximated by
nf = 3 below the charm threshold, by nf = 4 and 5 from there to the b and t thresholds,
respectively, and then by nf = 6 at scales higher than mt , it is therefore important to be aware
that the running changes slope across quark flavour thresholds. Likewise, it would change
across the threshold for any coloured new-physics particles that might exist, with a magnitude
depending on the particles’ colour and spin quantum numbers.
The negative overall sign of equation (19), combined with the fact that b0 > 0 (for nf ≤
16), leads to the famous result11 that the QCD coupling effectively decreases with energy,
called asymptotic freedom, for the discovery of which the Nobel prize in physics was awarded
to D. Gross, H. Politzer, and F. Wilczek in 2004. An extract of the prize announcement runs as
follows:
10
In a fixed-order calculation, the individual scales µi , would correspond, e.g., to the n hardest scales appearing
in an infrared safe sequential clustering algorithm applied to the given momentum configuration.
11
Perhaps the highest pinnacle of fame for equation (19) was reached when the sign of it featured in an episode
of the TV series “Big Bang Theory”.

— 12 —
P. Skands Introduction to QCD

What this year’s Laureates discovered was something that, at first sight, seemed
completely contradictory. The interpretation of their mathematical result was
that the closer the quarks are to each other, the weaker is the “colour charge”.
When the quarks are really close to each other, the force is so weak that they
behave almost as free particlesa . This phenomenon is called “asymptotic free-
dom”. The converse is true when the quarks move apart: the force becomes
stronger when the distance increasesb .
a
More correctly, it is the coupling rather than the force which becomes weak as the distance
decreases. The 1/r2 Coulomb singularity of the force is only dampened, not removed, by the
diminishing coupling.
b
More correctly, it is the potential which grows, linearly, while the force becomes constant.

Among the consequences of asymptotic freedom is that perturbation theory becomes better
behaved at higher absolute energies, due to the effectively decreasing coupling. Perturbative
calculations for our 900-GeV Z 0 boson from before should therefore be slightly faster con-
verging than equivalent calculations for the 90-GeV one. Furthermore, since the running of
αs explicitly breaks Bjørken scaling, we also expect to see small changes in jet shapes and in
jet production ratios as we vary the energy. For instance, since high-p⊥ jets start out with
a smaller effective coupling, their intrinsic shape (irrespective of boost effects) is somewhat
narrower than for low-p⊥ jets, an issue which can be important for jet calibration. Our cur-
rent understanding of the running of the QCD coupling is summarised by the plot in figure 6,
taken from a recent comprehensive review by S. Bethke [31,34]. A complementary up-to-date
overview of αs determinations can be found in [35].
As a final remark on asymptotic freedom, note that the decreasing value of the strong
coupling with energy must eventually cause it to become comparable to the electromagnetic
and weak ones, at some energy scale. Beyond that point, which may lie at energies of order
1015 − 1017 GeV (though it may be lower if as yet undiscovered particles generate large correc-
tions to the running), we do not know what the further evolution of the combined theory will
actually look like, or whether it will continue to exhibit asymptotic freedom.
Now consider what happens when we run the coupling in the other direction, towards
smaller energies. Taken at face value, the numerical value of the coupling diverges rapidly at
scales below 1 GeV, as illustrated by the curves disappearing off the left-hand edge of the plot
in figure 6. To make this divergence explicit, one can rewrite equation (22) in the following
form,
1
αs (Q2 ) = 2 , (24)
b0 ln Q
Λ 2

where
Λ ∼ 200 MeV (25)
specifies the energy scale at which the perturbative coupling would nominally become infinite,
called the Landau pole. (Note, however, that this only parametrises the purely perturbative
result, which is not reliable at strong coupling, so equation (24) should not be taken to imply
that the physical behavior of full QCD should exhibit a divergence for Q → Λ.)
Finally, one should be aware that there is a multitude of different ways of defining both Λ
and αs (MZ ). At the very least, the numerical value one obtains depends both on the renor-
malisation scheme used (with the dimensional-regularisation-based “modified minimal sub-
traction” scheme, MS, being the most common one) and on the perturbative order of the

— 13 —
P. Skands Introduction to QCD

0.5
April 2012

α s(Q) τ decays (N3LO)

Lattice QCD (NNLO)
0.4 DIS jets (NLO)
Heavy Quarkonia (NLO)
e+e– jets & shapes (res. NNLO)
Z pole fit (N3LO)
pp –> jets (NLO)
0.3
Landau Pole Asymptotic Freedom

& Confinement & Grand Unification?

0.2

0.1
QCD α s (Μ Z) = 0.1184 ± 0.0007
1 10 100
Q [GeV]
Figure 6: Illustration of the running of αs in a theoretical calculation (band) and in physical
processes at different characteristic scales, from [31, 34]. The little kinks at Q = mc and
Q = mb are caused by discontinuities in the running across the flavour thresholds.

calculations used to extract them. As a rule of thumb, fits to experimental data typically
yield smaller values for αs (MZ ) the higher the order of the calculation used to extract it (see,
e.g., [31, 34, 36, 37]), with αs (MZ )|LO >
∼ αs (MZ )|NLO > ∼ αs (MZ )|NNLO . Further, since the
number of flavours changes the slope of the running, the location of the Landau pole for
fixed αs (MZ ) depends explicitly on the number of flavours used in the running. Thus each
value of nf is associated with its own value of Λ, with the following matching relations across
thresholds guaranteeing continuity of the coupling at one loop,
 2  2
Λ5 21 mt 23
nf = 5 ↔ 6 : Λ6 = Λ5 Λ5 = Λ6 , (26)
mt Λ6
 2  2
Λ4 23 mb 25
nf = 4 ↔ 5 : Λ5 = Λ4 Λ4 = Λ5 , (27)
mb Λ5
 2  2
Λ3 25 mc 27
nf = 3 ↔ 4 : Λ4 = Λ3 Λ3 = Λ4 . (28)
mc Λ4

It is sometimes stated that QCD only has a single free parameter, the strong coupling.
However, even in the perturbative region, the beta function depends explicitly on the number
of quark flavours, as we have seen, and thereby also on the quark masses. Furthermore, in
the non-perturbative region around or below ΛQCD , the value of the perturbative coupling, as
obtained, e.g., from equation (24), gives little or no insight into the behavior of the full theory.
Instead, universal functions (such as parton densities, form factors, fragmentation functions,
etc), effective theories (such as the Operator Product Expansion, Chiral Perturbation Theory,

— 14 —
P. Skands Introduction to QCD

or Heavy Quark Effective Theory), or phenomenological models (such as Regge Theory or the
String and Cluster Hadronisation Models) must be used, which in turn depend on additional
non-perturbative parameters whose relation to, e.g., αs (MZ ), is not a priori known.
For some of these questions, such as hadron masses, lattice QCD can furnish important
additional insight, but for multi-scale and/or time-evolution problems, the applicability of
lattice methods is still severely restricted; the lattice formulation of QCD requires a Wick
rotation to Euclidean space. The time-coordinate can then be treated on an equal footing with
the other dimensions, but intrinsically Minkowskian problems, such as the time evolution of a
system, are inaccessible. The limited size of current lattices also severely constrain the scale
hierarchies that it is possible to “fit” between the lattice spacing and the lattice size.

1.5 Colour States

A final example of the application of the underlying SU(3) group theory to QCD is given by
considering which colour states we can obtain by combinations of quarks and gluons. The
simplest example of this is the combination of a quark and antiquark. We can form a total of
nine different colour-anticolour combinations, which fall into two irreducible representations
of SU(3):
3⊗3=8⊕1. (29)


The singlet corresponds to the symmetric wave function √13 RR̄ + GḠ + B B̄ , which
is invariant under SU(3) transformations (the definition of a singlet). The other eight lin-
early independent combinations (which can be represented by one for each Gell-Mann matrix,
with the singlet corresponding to the identity matrix) transform into each other under SU(3).
Thus, although we sometimes talk about colour-singlet states as being made up, e.g., of “red-
antired”, that is not quite precise language. The actual state RR̄ is not a pure colour singlet.
Although it does have a non-zero projection onto the singlet wave function above, it also has
non-zero projections onto the two members of the octet that correspond to the diagonal Gell-
Mann matrices. Intuitively, one can also easily realise this by noting that an SU(3) rotation of
RR̄ would in general turn it into a different state, say B B̄ , whereas a true colour singlet
would be invariant. Finally, we can also realise from equation (29) that a random (colour-
uncorrelated) quark-antiquark pair has a 1/N 2 = 1/9 chance to be in an overall colour-singlet
state; otherwise it is in an octet.
Similarly, there are also nine possible quark-quark (or antiquark-antiquark) combinations,
six of which are symmetric under interchange of the two quarks and three of which are anti-
symmetric:
 
|RRi
 |GGi   1 
  √ (|RGi − |GRi)
 |BBi  2
   
6 =  √1 (|RGi + |GRi)  3̄ =  √12 (|GBi − |BGi)  . (30)
 2 
 √1  √1 (|BRi − |RBi)
 2 (|GBi + |BGi)  2
√1 (|BRi + |RBi)
2

The members of the sextet transform into (linear combinations of) each other under SU(3)
transformations, and similarly for the members of the antitriplet, hence neither of these can

— 15 —
P. Skands Introduction to QCD

be reduced further. The breakdown into irreducible SU(3) multiplets is therefore

3⊗3=6⊕3. (31)

Thus, an uncorrelated pair of quarks has a 1/3 chance to add to an overall anti-triplet state
(corresponding to coherent superpositions like “red + green = antiblue”12 ); otherwise it is in
an overall sextet state.
Note that the emphasis on the quark-(anti)quark pair being uncorrelated is important;
production processes that correlate the produced partons, like Z → q q̄ or g → q q̄, will project
out specific components (here the singlet and octet, respectively). Note also that, if the quark
and (anti)quark are on opposite sides of the universe (i.e., living in two different hadrons),
the QCD dynamics will not care what overall colour state they are in, so for the formation of
multi-partonic states in QCD, obviously the spatial part of the wave functions (causality at the
very least) will also play a role. Here, we are considering only the colour part of the wave
functions. Some additional examples are

8 ⊗ 8 = 27 ⊕ 10 ⊕ 10 ⊕ 8 ⊕ 8 ⊕ 1 , (32)
3 ⊗ 8 = 15 ⊕ 6 ⊕ 3 , (33)
3 ⊗ 6 = 10 ⊕ 8 , (34)
3 ⊗ 3 ⊗ 3 = (6 ⊕ 3) ⊗ 3 = 10 ⊕ 8 ⊕ 8 ⊕ 1 . (35)

Physically, the 27 in the first line corresponds to a completely incoherent addition of the colour
charges of two gluons; the decuplets are slightly more coherent (with a lower total colour
charge), the octets yet more, and the singlet corresponds to the combination of two gluons that
have precisely equal and opposite colour charges, so that their total charge is zero. Further
extensions and generalisations of these combination rules can be obtained, e.g., using the
method of Young tableaux [38, 39].

12
In the context of hadronisation models, this coherent superposition of two quarks in an overall antitriplet state
is sometimes called a “diquark” (at low mqq ) or a “string junction” (at high mqq ), see section 5.1; it corresponds
to the antisymmatric “red + green = antiblue” combination needed to create a baryon wavefunction.

— 16 —
P. Skands Introduction to QCD

p1 p3

High  transverse-­‐
momentum  
interac2on  

p2 p4

Figure 7: Left: Rutherford scattering of quarks in QCD, exemplifying the type of process that
dominates the short-distance interaction cross section at hadron colliders. Right: an example
of what such a reaction looks like in a detector, in this case the ATLAS experiment.

2 Hard Processes
Our main tool for solving QCD at high energy scales, Q  ΛQCD , is perturbative quantum
field theory, the starting point for which is Matrix Elements (MEs) which can be calculated
systematically at fixed orders (FO) in the strong coupling αs . At least at lowest order (LO), the
procedure is standard textbook material [1] and it has also by now been highly automated,
by the advent of tools like M ADGRAPH [40], C ALCHEP [41] / C OMPHEP [42], and several
others [43–49]. Here, we require only that the reader has a basic familiarity with the methods
involved from graduate-level particle physics courses based, e.g., on [1, 4]. Our main concern
are the uses to which these calculations are put, their limitations, and ways to improve on the
results obtained with them.
For illustration, take one of the most commonly occurring processes in hadron collisions:
Rutherford scattering of two quarks via a t-channel gluon exchange — figure 7 — which at
leading order has the differential cross section

dσ π 4 ŝ2 + û2
qq 0 → qq 0 : = 2 αs2 , (36)
dt̂ ŝ 9 t̂2
with the 2 → 2 Mandelstam variables (“hatted” to emphasize that they refer to a partonic
2 → 2 scattering rather than the full pp → jets process)

ŝ = (p1 + p2 )2 , (37)
(1 − cos θ̂)
t̂ = (p3 − p1 )2 = −ŝ , (38)
2
(1 + cos θ̂)
û = (p4 − p1 )2 = −ŝ . (39)
2
Reality, however, is more complicated; the picture on the right-hand pane of figure 7 shows
a real dijet event, as recorded by the ATLAS experiment. The complications to be addressed
when going from left to right in figure 7 are: firstly, additional jets, a.k.a. real-emission cor-
rections, which can significantly change the topology of the final state, potentially shifting jets

— 17 —
P. Skands
Parton Distribution Functions Introduction to QCD

Hadrons are composite, with time-dependent structure:

u
p d
g
u

Figure 8: Illustration of partonic fluctuations inside a proton beam (from [50]).

fi (x, Q2) = number density of partons i
at momentum fraction x and probing scale Q2.
Linguistics
in or out of an experimentally defined (example):
acceptance region. Secondly, loop factors, a.k.a. virtual
!
2 e2 Q2)
corrections, change the number Fof2(x, Q ) =
available quantum i xfi (x,through
paths phase space, and hence
modify the normalization of the cross section (totali and differential). And finally, additional
structure
corrections are generated functionand byparton
by confinement distributions
the so-called underlying event. These cor-
rections must be taken into account to complete our understanding of QCD and connect the
short-distance physics with macroscopic experiments. Apart from the perturbative expansion
itself, the most powerful tool we have to organize this vast calculation, is factorization.

2.1 Factorization
In high-energy scattering problems involving hadrons in the initial state, we immediately face
the complication that hadrons are composite, with a time-dependent structure illustrated in
figure 8; there are partons within clouds of further partons, constantly being emitted and ab-
sorbed. Thus, before we can use perturbatively calculated partonic scattering matrix elements,
we must first address the partonic structure of the colliding hadron(s).
For the hadron to remain intact, the fluctuations inside it must involve momentum transfers
smaller than the confinement scale. Indeed, high-virtuality fluctuations are suppressed by
powers of
αs Λ 2
, (40)
|k|2
with Λ the confinement scale (∼ 200 MeV, see section 1.4) and |k| the virtuality of the fluctu-
ation. Thus, most fluctuations occur over timescales ∼ 1/Λ.
A hard perturbative probe, on the other hand, such as the exchanged photon in DIS (fig-
ure 3), interacts over a much shorter timescale 1/Q  1/Λ, during which the partonic fluctu-
ations in the struck hadron appear almost frozen. The hard probe effectively takes an instan-
taneous snapshot of the hadron structure, at a characteristic resolution given by ∼ 1/Q.
This is formalized by the factorization theorem [51] (see also the TASI lectures by George
Sterman [52]), which expresses the independence of long-wavelength (soft) structure on the
nature of the hard (short-distance) process. Originally formulated for DIS, factorization allows
us to write the cross section for lepton-hadron scattering as a convolution of a non-perturbative
but universal (i.e., process-independent) parton density function (PDF) and a perturbatively
calculable partonic scattering cross section. Denoting the fraction of the hadron momentum

— 18 —
P. Skands Introduction to QCD

carried by parton i by xi ,
p~i = xi p~h , (41)
we may write the lepton-hadron cross section on factorized form (see, e.g., [4, 53]),
XZ 1 Z
dσ̂`i→f (xi , Φf , µ2F )
2
σ`h = dxi dΦf fi/h (xi , µF ) , (42)
0 dxi dΦf
i

with i an index running over all possible parton types13 in the incoming hadron and f enu-
merating all possible (partonic) final states, with Lorentz-invariant phase space, Φf .
The parton density functions (PDFs), fi/h , parametrize the distribution of partons inside the
target hadron. They are not a priori calculable and must be constrained by fits to data. This is
discussed in section 2.2.
The partonic cross section, dσ̂, knows nothing of the target hadron apart from the fact that
it contained the struck parton. It is calculable within perturbation theory, as will be discussed
in section 2.3.
The dividing line between the two is drawn at an arbitrary (“user-defined”) scale µF , called
the factorization scale. There is some arbitrariness involved in choosing a value for µF . Some
heuristic arguments to guide in the choice of factorization scale are the following. On the
long-distance side, the PDFs include a (re)summation of fluctuations inside fluctuations up
to virtualities of order µF . It would therefore not make much sense to take µF significantly
larger than the scales characterizing resolved particles on the short-distance side of the calcu-
lation (i.e., the particles appearing explicitly in Φf ); otherwise the PDFs would be including
sums over fluctuations that happen on timescales shorter than those probed by the physical
process. Similarly, µF should also not be taken much lower than the scale(s) appearing in the
hard process. For matrix elements characterized by a single well-defined scale, such as the
Q2 scale in DIS or the invariant-mass scale ŝ in Drell-Yan production (q q̄ → √Z/γ ∗ → `+ `− ),
such arguments essentially fix the preferred scale choice to µF = Q or µF = ŝ, respectively,
which may then be varied by a factor of 2 (or larger) around the nominal value in order to
estimate uncertainties. For multi-scale problems, however, such as pp → Z/W + n jets, there
are several a priori equally good choices available, from the lowest to the highest QCD scales
that can be constructed from the final-state momenta, usually with several dissenting groups
of theorists arguing over which particular choice is best. Suggesting that one might simply
measure the scale would not really be an improvement, as the factorization scale is fundamen-
tally unphysical and therefore unobservable (similarly to gauge or convention choices). One
plausible strategy is to look at higher-order (NLO or NNLO) calculations, in which correction
terms appear that cancel the dependence on the scale choice, stabilizing the final result. From
such comparisons, a “most stable” initial scale choice can in principle be determined, which
then furnishes a reasonable starting point, but we emphasize that the question is intrinsically
ambiguous, and no golden recipe is likely to magically give all the right answers. The best we
can do is to vary the value of µF not only by an overall factor, but also by exploring different
possible forms for its functional dependence on the momenta appearing in Φf . A complemen-
tary useful discussion of the pros and cons of different factorization scale choices can be found
in the TASI lectures by Tilman Plehn [54].
13
Typically, only quarks and gluons are included in this sum, but also photons and even leptons can in principle
be included. Similarly, parton density functions are normally used to describe hadrons, but can also be defined,
e.g., to describe the cloud of virtual photons (and fermion pairs) surrounding an electron.

— 19 —
P. Skands Introduction to QCD

Secondly, and more technically, at NLO and beyond one also has to settle on a factorization
scheme in which to do the calculations. For all practical purposes, students focusing on LHC
physics are only likely to encounter one such scheme, the modified minimal subtraction (MS)
one already mentioned in the discussion of the definition of the strong coupling in Section 1.4.
At the level of these lectures, we shall therefore not elaborate further on this choice here.
We note that factorization can also be applied multiple times, to break up a complicated
calculation into simpler pieces that can be treated as approximately independent. This will be
very useful when dealing with successive emissions in a parton shower, section 3.2, or when
factoring off decays of long-lived particles from a hard production process, section 3.4.
We round off the discussion of factorization by mentioning a few caveats the reader should
be aware of. (See [52] for a more technical treatment.)
Firstly, the proof only applies to the first term in an operator product expansion in “twist”
= mass dimension - spin. Since operators with higher mass dimensions are suppressed by the
hard scale to some power, this leading twist approximation becomes exact in the limit Q → ∞,
while at finite Q it neglects corrections of order
[ln(Q2 /Λ2 )]m<2n
Higher Twist : (n = 2 for DIS) . (43)
Q2n
In section 5, we shall discuss some corrections that go beyond this approximation, in the
context of multiple parton-parton interactions.
Secondly, the proof only really applies to inclusive cross sections in DIS [51] and in Drell-
Yan [55]. For all other hadron-initiated processes, factorization is an ansatz. For a general
hadron-hadron process, we write the assumed factorizable cross section as:
XZ 1 Z 1 XZ dσ̂ij→f
dσh1 h2 = dxi dxj dΦf fi/h1 (xi , µ2F ) fj/h2 (xj , µ2F ) . (44)
0 0 dxi dxj dΦf
i,j f

Note that, if dσ̂ is divergent (as, e.g., Rutherford scattering is) then the integral over dΦf
must be regulated, e.g. by imposing some explicit minimal transverse-momentum cut and/or
other phase-space restrictions.

2.2 Parton Densities

The parton density function, fi/h (x, µ2F ), represents the effective density of partons of type/flavor
i, as a function of the momentum fraction14 xi , when a hadron of type h is probed at the fac-
torization scale µF . The PDFs are non-perturbative functions which are not a priori calculable,
but a perturbative differential equation governing their evolution with µF can be obtained by
requiring that physical scattering cross sections, such as the one for DIS in equation (42), be
independent of µF to the calculated orders [56]. The resulting renormalization group equation
(RGE) is called the DGLAP15 equation and can be used to “run” the PDFs from one pertur-
bative resolution scale to another (its evolution kernels are the same as those used in parton
showers, to which we return in section 3.2).
This means that we only need to determine the form of the PDF as a function of x a single
(arbitrary) scale, µ0 . We can then get its form at any other scale µF by simple RGE evolution.
14
Recall: the x fraction is defined in equation (41).
15
DGLAP: Dokshitzer-Gribov-Lipatov-Altarelli-Parisi [56–58].

— 20 —
P. Skands Introduction to QCD

2 8

xf(x,Q2)
xf(x,Q2)

1.8 7

1.6
6
1.4
5
1.2

1 4

0.8 3

0.6
2
0.4
1
0.2

0 -4 0 -4 -3 -2 -1
-3 -2 -1
10 10 10 10 10 10 10 10 x
x

Figure 9: Illustration of the change of the u (black), ū (red, dashed), c (green, dotted), and
g (blue, dot-dashed) distributions, from Q = µF = 2 GeV (left) to Q = µF = 100 GeV
(right). Note that a factor 0.1 has been applied to the gluon distribution. Plots made using the
HEPDATA online tool [64].

In the context of PDF fits (constraining the form of the PDF functions by fitting cross sections
to experimental data, e.g., from DIS [59, 60], Drell-Yan [61, 62], and pp → jets [63]), the
reference scale µ0 is usually taken to be relatively low, of order one or a few GeV.
The behavior of the PDFs as we evolve µF from a low scale, 2 GeV, to a high one, 100 GeV,
is illustrated in figure 9, for the MSTW16 2008 LO17 PDF set [65]. At low Q = µF = 2 GeV
(left), the proton structure is dominated by a few hard quarks (a “valence bump” is clearly
visible around x ∼ 0.2), while at higher scales Q = 100 GeV (right) we predominantly resolve
fluctuations within fluctuations, yielding increasingly large gluon- and sea-quark distributions
with rather small x values, while the valence quarks play a progressively smaller role.
We note that different collaborations, like CTEQ, MSTW, NNPDF, etc., use different ansätze
for the form of f (x, µ20 ). They may also include different data in the fits, and/or treat or weight
the data differently. Thus, results from different groups may not always be mutually compati-
ble. An example is given in figure 10, which shows the difference between the CTEQ6L1 gluon
PDF [66] (red dashed) and the MSTW 2008 LO one [65], normalized to MSTW (which would
thus be a flat line at zero), at µF = 10 GeV. The y axis shows the relative difference between
the sets, in per cent. Also shown are the 90% CL contours computed from the uncertainty vari-
ations included in the MSTW 2008 LO set (black). Using only the MSTW uncertainty band,
one would arrive at an estimated ∼ 5% uncertainty over most of the x range, while including
the CTEQ6L1 set would increase that to ∼ 10%. At NLO, this discrepancy is reduced, but not
removed. A significant effort is currently being undertaken within the PDF community to agree
on common, and more comprehensive, ways of defining PDF uncertainty bands [63, 67]. This
16
MSTW: Martin-Stirling-Thorne-Watt.
17
The “LO” means that the fit was performed using LO matrix elements in the cross section formulae.

— 21 —
P. Skands Introduction to QCD

Figure 10: Illustration of the difference between the MSTW 2008 and CTEQ6 LO gluon PDFs
at µF = 10 GeV. All curves are normalized to the central MSTW 2008 prediction. The black
solid lines show the 90% CL MSTW variations, while the dashed red line shows the CTEQ6L1
distribution.

is complicated due to the different ways of defining f (x, µ20 ) and due to the experimental data
sets not always being fully compatible with one another. For the time being, it is recommended
to try at least sets from two different groups, for a comprehensive uncertainty estimate.
Occasionally, the words structure functions and parton densities are used interchangeably.
However, there is an important distinction between the two, which we find often in (quantum)
physics: the former is a physical observable used to parametrize the DIS cross sections (see
e.g. [4]), while the latter is a “fundamental” quantity extracted from it. In particular, since the
parton densities are not, themselves, physically observable, they can only be defined within
a specific factorization scheme, order by order in perturbation theory. The only exception
is at leading order, at which they have the simple physical interpretation of parton number
densities. When going to higher orders, we tend to keep the simple intuitive picture from
LO in mind, but one should be aware that the fundamental relationship between PDFs and
measured quantities is now more complicated (due to the interplay between the PDFs and the
real and virtual corrections to the LO cross section), and that the parton densities no longer
have a clear probabilistic interpretation starting from NLO.
The reader should also be aware that there is some ambiguity whether NLO PDFs should
be used for LO calculations. In principle, the higher-order PDFs are better constrained and
the difference between, e.g., an NLO and an LO set should formally be beyond LO precision,
so that one might be tempted to simply use the highest-order available PDFs for any calcula-
tion. However, higher-order terms can sometimes be absorbed, at least partially, into effective
lower-order coefficients. In the context of PDFs, the fit parameters of lower-order PDFs will
attempt to compensate for missing higher-order contributions in the matrix elements. To the
extent those higher-order contributions are universal, this is both desirable and self-consistent.
This leads to some typical qualitative differences between LO and NLO PDFs, illustrated in
figure 11: NLO PDFs tend to be smaller at low x and slightly larger at high x, than LO ones.
Thus, it is quite possible that using an NLO PDF in conjunction with LO matrix elements can
give a worse agreement with data than LO PDFs do.

— 22 —
P. Skands Introduction to QCD

xf(x,Q2)
2
10

10

-1
10

-2
10 -4 -3 -2 -1
10 10 10 10 x

Figure 11: Illustration of the change between PDF fits using LO and NLO matrix elements:
the g distribution at LO (black) and NLO (red, dashed), and the u distribution at LO (green,
dotted) and NLO (blue, dot-dashed), for the MSTW 2008 PDF sets [65], at Q = µF = 10 GeV.
Plots made using the HEPDATA online tool [64].

Finally, another oft-raised question concerns which PDF sets to use for the parton-shower
evolution in Monte Carlo generators. Importantly, the equations driving the initial-state show-
ers in Monte Carlo models are only sensitive to ratios of PDFs [68]. Since the shower evolution
typically only has leading-logarithmic (LL) precision, it should be theoretically consistent to
use any (LO or better) PDF set to drive the evolution. However, similarly to above, there will
be subleading differences between different choices, and one is justified in worrying about
the level of physical effects that could be generated. Unfortunately, there is currently no way
to ensure 100% self-consistency. Since PDF fits are not done with MC codes, but instead use
analytical resummation models (see, e.g., the TASI lectures by Sterman [52]), which are not
identical to their MC counterparts, the PDF fits are essentially “tuned” to a slightly different
resummation than that incorporated in a given MC model. In practice, not much is known
about the size and impact of this ambiguity [69]. Known differences include: the size of
phase space (purely collinear massless PDF evolution vs. the finite-transverse-momentum mas-
sive MC phase space), the treatment of momentum conservation and recoil effects, additional
higher-order effects explicitly or implicitly included in the MC evolution, choice of renormal-
ization scheme and scale, and, for those MC algorithms that do not rely on collinear (DGLAP,
see [4]) splitting kernels (e.g., the various kinds of dipole evolution algorithms, see [70]),
differences in the effective factorization scheme.
As a baseline, we recommend simply using whatever PDF set the given MC model was
originally tuned with, since this should de facto (by fitting the available data) reabsorb as
much of the inconsistency as possible. Furthermore, it should be emphasized that underlying-
event and minimum-bias models based on multi-parton interactions (see section 5.2) usually
make the explicit assumption that the PDFs can be interpreted as physical number densities
even down to very low Q and x, a property which is generally only true for LO PDFs. It must

— 23 —
P. Skands Introduction to QCD

therefore be strongly discouraged to use (N)NLO PDF sets in this context.

2.3 Fixed-Order QCD

Consider the production of an arbitrary final state, F (e.g., a Higgs boson, a tt̄ pair, etc).
Schematically, we may express the (perturbative) all-orders differential cross section for an
observable O, in the following way:
∞ Z
X ∞
X
dσF (`) 2

= dΦF +k MF +k δ O − O(ΦF +k ) , (45)
dO ME k=0 `=0
| {z } | {z }
Σ legs Σ loops

where, for compactness, we have suppressed all PDF and luminosity normalization factors.
(`)
MF +k is the amplitude for producing F in association with k additional final-state partons,
“legs”, and with ` additional loops. The sums start at k = 0 and ` = 0, corresponding to
the Leading Order for producing F , while higher terms represent real and virtual corrections,
respectively.
The purpose of the δ function is to project out hypersurfaces of constant value of O in the
full dΦF +k phase space, with O(ΦF +k ) a function that defines O evaluated on each specific
momentum configuration, ΦF +k . (Without the δ function, the formula would give the total
integrated cross section, instead of the cross section differentially in O.)
We recover the various fixed-order truncations of perturbative QCD (pQCD) by limiting the
nested sums in equation (45) to include only specific values of k + `. Thus,

k = 0, ` = 0 =⇒ Leading Order (usually tree-level) for F production

k = n, ` = 0 =⇒ Leading Order for F + n jets
k + ` ≤ n, =⇒ Nn LO for F (includes Nn−1 LO for F + 1 jet, Nn−2 LO for F +
2 jets, and so on up to LO for F + n jets) .

For k ≥ 1, we are not considering inclusive F production; instead, we are considering the
process F + k jets. If we simply integrate over all momenta, as implied by the integration over
dΦF +k in equation (45), we would be including configurations in which one or more of the
k partons are collinear or soft. Such configurations are infrared divergent in QCD and hence
must be regulated. Since we talk about collinear and soft divergences (the origins of which will
be discussed in more detail in sections 2.4 and 3.2), cuts on angles and energies and/or cuts
on combinations, like transverse momenta, can be used to cut away the problematic regions of
phase space.
Recall, however, that pQCD is approximately scale invariant. This implies that a regu-
larization cut on a dimensionful quantity, like energy or transverse momentum, should be
formulated as a ratio of scales, rather than as an absolute number. For example, a jet with
p⊥ = 50 GeV would be considered hard and well-separated if produced in association with
an ordinary Z boson (with hard scale MZ = 91.2 GeV), while the same jet would be consid-
ered soft if produced in association with a 900-GeV Z 0 boson (see [25, 26] for more explicit
examples).

— 24 —
P. Skands Introduction to QCD

F @ LO F + 2 @ LO
LO for F + 2
(2) (2) → ∞ for F + 1
` (loops) 2 σ0 σ1 ... 2 (2)
σ0
(2)
σ1 ... → ∞ for F + 0

` (loops)
(1) (1) (1)
1 σ0 σ1 σ2 ... Max Born, 1 (1)
σ0
(1)
σ1
(1)
σ2 ...
1882-1970
Nobel 1954
(0) (0) (0) (0)
0 σ0 σ1 σ2 σ3 ... 0 (0)
σ0
(0)
σ1
(0)
σ2 σ3
(0)
...
0 1 2 3 ... 0 1 2 3 ...
k (legs) k (legs)

Figure 12: Coefficients of the perturbative series covered by LO calculations. Left: F pro-
duction at lowest order. Right: F + 2 jets at LO, with the half-shaded box illustrating the
restriction to the region of phase space with exactly 2 resolved jets. The total power of αs for
each coefficient is n = k + `. (Photo of Max Born from nobelprize.org).

The essence of the point is that, if the regularization scale is taken too low, logarithmic
enhancements of the type  2
n m≤2n QF
αs ln (46)
Q2k
will generate progressively larger corrections, order by order, which will spoil any fixed-order
truncation of the perturbative series. Here, QF is the hard scale associated with the process
under consideration, while Qk is the scale associated with an additional parton, k.
A good rule of thumb is that if σk+1 ≈ σk (at whatever order you are calculating), then the
perturbative series is converging too slowly for a fixed-order truncation of it to be reliable. For
fixed-order perturbation theory to be applicable, you must place your cuts on the hard process
such that σk+1  σk . In the discussion of parton showers in Section 3.2, we shall see how the
region of applicability of perturbation theory can be extended.
The virtual amplitudes, for ` ≥ 1, are divergent for any point in phase space. However,
as encapsulated by the famous KLN theorem [71, 72], unitarity (which essentially expresses
probability conservation) puts a powerful constraint on the IR divergences18 , forcing them to
cancel exactly against those coming from the unresolved real emissions that we had to cut out
above, order by order, making the complete answer for fixed k + ` = n finite19 Nonetheless,
since this cancellation happens between contributions that formally live in different phase
spaces, a main aspect of loop-level higher-order calculations is how to arrange for this cancel-
lation in practice, either analytically or numerically, with many different methods currently on
the market. We shall discuss the idea behind subtraction approaches in section 2.4.
A convenient way of illustrating the terms of the perturbative series that a given matrix-
element-based calculation includes is given in figure 12. In the left-hand pane, the shaded
box corresponds to the lowest-order “Born-level” matrix element squared. This coefficient
is non-singular and hence can be integrated over all of phase space, which we illustrate by
letting the shaded area fill all of the relevant box. A different kind of leading-order calculation
18
The loop integrals also exhibit UV divergences, but these are dealt with by renormalization.
19
Formally, the KLN theorem states that the sum over degenerate quantum states is finite. In context of fixed-
order perturbation theory, this is exemplified by states with infinitely collinear and/or soft radiation being degen-
erate with the corresponding states with loop corrections; they cannot be told apart by any physical observable.

— 25 —
P. Skands Introduction to QCD

F @ NLO F+1 @ NLO

NLO for F + 1
(2) (2) NLO for F + 0 → LO for F + 2
2 σ0 σ1 ... → LO for F + 1 2 (2)
σ0
(2)
σ1 ... → ∞ for F + 0
` (loops)

` (loops)
(1) (1) (1)
1 σ0 σ1 σ2 ... 1 (1)
σ0
(1)
σ1
(1)
σ2 ...
(0) (0) (0) (0)
0 σ0 σ1 σ2 σ3 ... 0 (0)
σ0
(0)
σ1
(0)
σ2 σ3
(0)
...
0 1 2 3 ... 0 1 2 3 ...
k (legs) k (legs)

Figure 13: Coefficients of the perturbative series covered by NLO calculations. Left: F produc-
tion at NLO. Right: F + 1 jet at NLO, with half-shaded boxes illustrating the restriction to the
region of phase space with exactly 1 resolved jet. The total power of αs for each coefficient is
n = k + `.

is illustrated in the right-hand pane of figure 12, where the shaded box corresponds to the
lowest-order matrix element squared for F + 2 jets. This coefficient diverges in the part of
phase space where one or both of the jets are unresolved (i.e., soft or collinear), and hence
integrations can only cover the hard part of phase space, which we reflect by only shading the
upper half of the relevant box.
Figure 13 illustrates the inclusion of NLO virtual corrections. To prevent confusion, first a
(1)
point on notation: by σ0 , we intend
Z
(1) (1) (0)∗
σ0 = dΦ0 2Re[M0 M0 ] , (47)

which is of order αs relative to the Born level. Compare, e.g., with the expansion of equa-
(1)
tion (45) to order k + ` = 1. In particular, σ0 should not be confused with the integral over
the 1-loop matrix element squared (which would be of relative order αs2 and hence forms part
(2)
of the NNLO coefficient σ0 ). Returning to figure 13, the unitary cancellations between real
and virtual singularities imply that we can now extend the integration of the real correction in
the left-hand pane over all of phase space, while retaining a finite total cross section,
Z Z Z
NLO (0) 2 (0) 2 (1) (0)∗
σ0 = dΦ0 |M0 | + dΦ1 |M1 | + dΦ0 2Re[M0 M0 ]
(48)
(0) (0) (1)
= σ0 + σ1 + σ0 ,

(0)
with σ0 the finite Born-level cross section, and the positive divergence caused by integrating
the second term over all of phase space is canceled by a negative one coming from the inte-
gration over loop momenta in the third term. One method for arranging the cancellation of
singularities — subtraction — is discussed in section 2.4.
However, if our starting point for the NLO calculation is a process which already has a
non-zero number of hard jets, we must continue to impose that at least that number of jets

— 26 —
P. Skands Introduction to QCD

F @ NNLO
NNLO for F + 0
(2) (2) → NLO for F + 1
2 σ0 σ1 ... → LO for F + 2

` (loops)
(1) (1) (1)
1 σ0 σ1 σ2 ...

(0) (0) (0) (0)

0 σ0 σ1 σ2 σ3 ...
0 1 2 3 ...
k (legs)

Figure 14: Coefficients of the perturbative series covered by an NNLO calculation. The total
power of αs for each coefficient is n = k + `. Green shading represents the full perturbative
coefficient at the respective k and `.

must still be resolved in the final-state integrations,

Z Z Z
NLO (0) 2 (0) (1) (0)∗
σ1 (p⊥min ) = dΦ1 |M1 | + dΦ2 |M2 |2 + dΦ1 2Re[M1 M1 ]
p⊥ >p⊥min p⊥1 >p⊥min p⊥ >p⊥min

(0) (0) (1)

= σ1 (p⊥ > p⊥min ) + σ2 (p⊥1 > p⊥min ) + σ1 (p⊥ > p⊥min ) ,
(49)
where the restriction to at least one jet having p⊥ > p⊥min has been illustrated in the right-
hand pane of figure 13 by shading only the upper part of the relevant boxes. In the second
term in equation (49), the notation p⊥1 is used to denote that the integral runs over the phase
space in which at least one “jet” (which may consist of one or two partons) must be resolved
with respect to p⊥min . Here, therefore, an explicit dependence on the algorithm used to define
“a jet” enters for the first time. This is discussed in more detail in the 2009 ESHEP lectures by
Salam [5].
To extend the integration to cover also the case of 2 unresolved jets, we must combine the
left- and right-hand parts of figure 13 and add the new coefficient
(2) (1) (2) (0)∗
σ0 = |M0 |2 + 2Re[M0 M0 ], (50)

as illustrated by the diagram in figure 14.

2.4 The Subtraction Idea

According to the KLN theorem, the IR singularities coming from integrating over collinear and
soft real-emission configurations should cancel, order by order, by those coming from the IR
divergent loop integrals. This implies that we should be able to rewrite e.g. the NLO cross
section, equation (48), as
Z  Z 
NLO Born (0) 2 (1) (0)∗
σ = σ + Finite dΦF +1 |MF +1 | + Finite dΦF 2Re[MF MF ] (51) ,

with the second and third terms having had their common (but opposite-sign) singularities
canceled out and some explicitly finite quantities remaining.

— 27 —
P. Skands Introduction to QCD

The first step towards this goal is to classify all IR singularities that could appear in the
amplitudes. We know that the IR limits are universal, so they can be classified using a set of
process-independent functions that only has to be worked out once and for all. A widely used
such set of functions are the Catani-Seymour (CS) dipole ones [73, 74], a method which by
now has even been partially automated [75,76]. Here, we shall instead use a formalism based
on antennae [78–80]. The distinction between the two is basically that one antenna is made up
of two dipole “ends”, hence the antenna formalism tends to generate somewhat fewer terms.
At NLO, however, there is no fundamental incompatibility — the antennae we use here can
always be partitioned into two dipole ends, if so desired. (Note: only the antenna method
has been successfully generalized to NNLO [81, 82]. Other NNLO techniques, not covered
here, are sector decomposition, see [83, 84], and the generic formalism for hadroproduction of
colorless states presented in [85].)
At NLO, the idea with subtraction is thus to rewrite the NLO cross section by adding and
subtracting a simple function, dσS , that encapsulates all the IR limits,
Z  
NLO Born (0)
σ = σ + dΦF +1 |MF +1 |2 − dσSNLO
| {z }
Finite by Universality
Z Z
(1) (0)∗
+ dΦF 2Re[MF MF ] + dΦF +1 dσSNLO . (52)
| {z }
Finite by KLN
The task now is to construct a suitable form for dσS . A main requirement is that it should be
sufficiently simple that the integral in the last term can be done analytically, in dimensional
regularization, so that the IR poles it generates can be canceled against those from the loop
term.
To build a set of universal terms that parametrize the IR singularities of any amplitude, we
start from the observation that gauge theory amplitudes factorize in the soft limit, as follows:
!
2 2m 2
j →0 2s ik 2m
|MF +1 (. . . , i, j, k, . . .)|2 → gs2 NC − 2 i − 2 k |MF (. . . , i, k, . . .)|2 ,(53)
g

sij sjk sij sjk

where parton j is a soft gluon, partons i, j, and k form a chain of color-space index contractions
(we say they are color-connected), gs is the strong coupling, and the terms in parenthesis are
called the soft eikonal factor. We here show it including mass corrections, which appear if i
and k have non-zero rest masses, with the invariants sab then defined as
sab ≡ 2pa · pb = (pa + pb )2 − m2a − m2b . (54)
The color factor, NC , is valid for the leading-color contribution, regardless of whether the
i and k partons are quarks or gluons. At subleading color, an additional soft-eikonal factor
identical to the one above but with a color factor proportional to −1/NC arises for each q q̄
pair combination. This, e.g., modifies the effective color factor for q q̄ → qg q̄ from NC to
NC (1 − 1/NC ) = 2CF , in agreement with the color factor for quarks being CF rather than CA .
Similarly, amplitudes also factorize in the collinear limit (partons i and j parallel, so
sij → 0), in which the eikonal factor above is replaced by the famous Dokshitzer-Gribov-
Lipatov-Altarelli-Parisi (DGLAP) splitting kernels [56–58], which were already mentioned in

— 28 —
P. Skands Introduction to QCD

section 2.2, in the context of PDF evolution. They are also the basis of conventional parton-
shower models, to which we return in section 3.2.
Essentially, what antenna functions, CS dipoles, and the like, all do, is to combine the soft
(eikonal) and collinear (Altarelli-Parisi) limits into one universal set of functions that achieve
the correct limiting behavior for both soft and collinear radiation. To give an explicit example,
the antenna function for gluon emission from a color-connected q q̄ pair can be derived from
the matrix elements squared for the process Z 0 → q q̄ → qg q̄ [86],
 
 2s  
sij 
2
M(Z 0 → qi gj q̄k ) 2  ik 1 sjk 
= gs 2CF  + +  , (55)
|M(Z 0 → qI q̄K )|2  sij sjk sIK sij sjk 
| {z } | {z }
eikonal collinear

where we have neglected mass corrections (see [87, 88] for massive expressions) and we
recognize the universal eikonal soft factor from equation (53) in the first term. The two
additional terms are less singular, and are required to obtain the correct collinear (Altarelli-
Parisi) limits as sij → 0 or sjk → 0.
However, since the singularity structure is universal, we could equally well have used the
process H 0 → q q̄ → qg q̄ to derive the antenna function. Our antenna function would then
have come out as [88],
 
 2s  
2 
2
M(H 0 → qi gj q̄k )  ik 1 sjk sij 
= gs2 2CF  + + +  , (56)
|M(H 0 → qI q̄K )|2  sij sjk sIK sij sjk sIK 
| {z } | {z } |{z}
eikonal collinear finite

where the additional term 2/sIK is non-singular (“finite”) over all of phase space. Thus, we
here see an explicit example that the singularities are process-independent while the non-
singular terms are process-dependent. Since we add and subtract the same term in equa-
tion (52), the final answer does not depend on the choice of finite terms. We say that they
correspond to different subtraction schemes. One standard antenna subtraction scheme, which
uses the antenna function defined in equation (55) rather than the one in equation (56), is
the Gehrmann-Gehrmann-de Ridder-Glover (GGG) one, given in [80].
If there is more than one color antenna in the Born-level process, the form of dσS is
obtained as a sum over terms, each of which captures one specific soft limit and either all or
“half” of a collinear one, depending on the specific scheme and the type of parton,
X
dσS = AIK→ijk |MF (. . . , I, K, . . .)|2 , (57)
j

with the sum running over all singular 3 → 2 “clusterings” of the (F + 1)-parton state to F
partons. An analysis of the different ways of partitioning the collinear singularity of gluons
among neighboring antenna is beyond the scope of this introduction, but useful discussions
can be found in [89–91].

2.5 Infrared Safety

A further requirement for being able to perform calculations within perturbative QCD is that
the observable be infrared safe. Note: by “infrared”, we here mean any limit that involves a

— 29 —
P. Skands Introduction to QCD

low scale (i.e., any non-UV limit), without regard to whether it is collinear or soft.
The property of infrared safety defines a special class of observables which have minimal
sensitivity to long-distance physics, and which can be consistently computed in pQCD. An
observable is infrared safe if:

1. (Safety against soft radiation): Adding any number of infinitely soft particles should not
change the value of the observable.

2. (Safety against collinear radiation): Splitting an existing particle up into two comoving
particles, with arbitrary fractions z and 1 − z, respectively, of the original momentum,
should not change the value of the observable.

If both of these conditions are satisfied, any long-distance non-perturbative corrections will be
suppressed by the ratio of the long-distance scale to the short-distance one to some (observable-
dependent) power, typically

Q2IR
IR Safe Observables: IR corrections ∝ (58)
Q2UV

where QUV denotes a generic hard scale in the problem, and QIR ∼ ΛQCD ∼ O(1 GeV).
Due to this power suppression, IR safe observables are not so sensitive to our lack of ability
to solve the strongly coupled IR physics, unless of course we go to processes for which the
relevant hard scale, QUV , is small (such as minimum-bias, soft jets, or small-scale jet substruc-
ture). Even when a high scale is present, however, as in resonance decays, jet fragmentation,
or underlying-event-type studies, infrared safety only guarantees us that infrared corrections
are small, not that they are zero. Thus, ultimately, we run into a precision barrier even for IR
safe observables, which only a reliable understanding of the long-distance physics itself can
address.
To constrain models of long-distance physics, one needs infrared sensitive observables.
Hence it is not always the case that infrared safe observables are preferable — the purpose
decides the tool. Instead of the suppressed corrections above, the perturbative prediction for
such observables contains logarithms of the type already encountered in equation (46),
 2 
n m≤2n QUV
IR Sensitive Observables: IR Corrections ∝ αs log , (59)
Q2IR

which grow increasingly large as QIR /QUV → 0. As an example, consider such a fundamental
quantity as particle multiplicities (= number of particles); in the absence of nontrivial infrared
effects, the number of partons tends logarithmically to infinity as the IR cutoff is lowered.
Similarly, the distinction between a charged and a neutral pion only occurs in the very last
phase of hadronization, and hence observables that only include charged tracks, for instance,
are always IR sensitive20 .
Two important categories of infrared safe observables that are widely used are event shapes
and jet algorithms. Jet algorithms are perhaps nowhere as pedagogically described as in the
2009 ESHEP lectures by Salam [5, Chapter 5]. Event shapes in the context of hadron colliders
20
This remains true in principle even if the tracks are clustered into jets, although the energy clustered in this
way does provide a lower bound on QUV in the given event, since “charged + neutral > charged-only”.

— 30 —
P. Skands Introduction to QCD

have not yet been as widely explored, but the basic phenomenology is introduced also by
Salam and collaborators in [93], with first measurements reported by CMS and ATLAS [94,95]
and a proposal to use them also for the characterization of soft-QCD (“minimum-bias”) events
put forth in [96].
Let us here merely emphasize that the real reason to prefer infrared safe jet algorithms
over unsafe ones is not that they necessarily give very different or “better” answers in the ex-
periment — experiments are infrared safe by definition, and the difference between infrared
safe and unsafe algorithms may not even be visible when running the algorithm on experi-
mental data — but that it is only possible to compute perturbative QCD predictions for the
infrared safe ones. Any measurement performed with an infrared unsafe algorithm can only
be compared to calculations that include a detailed hadronization model. This both limits the
number of calculations that can be compared to and also adds an a priori unknown sensitivity
to the details of the hadronization description, details which one would rather investigate and
constrain separately, in the framework of more dedicated fragmentation studies.
For LHC phenomenology, the preferred IR safe algorithm for jet reconstruction is currently
the anti-kT one [97], with size parameter R varying between 0.4 and 0.7, though larger sizes
can be motivated in certain contexts, e.g., to look for highly energetic jets and/or the boosted
decay products of high-mass objects [30, 98]. This algorithm generates circular-looking jets,
so subtracting off the energy believed to be associated with the underlying event (UE, see
section 5.2) is particularly simple.
For jet substructure, typically either the “kT” or “Cambridge/Aachen” algorithms are used,
see e.g. [30, 98]. The clustering measures used in these algorithms more closely mimic the
singularity structure of QCD bremsstrahlung and they are therefore particularly well suited
to “unravel” a tree of QCD branchings [5], such as a parton shower generates. The Cam-
bridge/Aachen algorithm may also be used to characterize the underlying event, see [99].

— 31 —
P. Skands Introduction to QCD

3 Monte Carlo Event Generators

In this section, we discuss the physics of Monte Carlo event generators and their mathematical
foundations, at an introductory level. We shall attempt to convey the main ideas as clearly
as possible without burying them in an avalanche of technical details. References to more
detailed discussions are included where applicable. We assume the reader is already familiar
with the contents of the preceding section on hard processes.
The task of a Monte Carlo event generator is to calculate everything that happens in a high-
energy collision, from the hard short-distance physics to the long wavelengths of hadronisation
and hadron decays. Obviously, this requires some compromises to be made. General-purpose
generators like H ERWIG [48, 100, 101], P YTHIA [102, 103], and S HERPA [104], start from
low-order (LO or NLO) descriptions of the perturbative hard physics and then attempt to
include the “most significant” corrections, such as higher-order matrix-element corrections and
parton showers, resonance decays and finite-width effects, underlying event, beam remnants,
hadronisation, and hadron decays. Each of them had slightly different origins, which carries
through to the emphasis placed on various physics aspects today:
• P YTHIA. Successor to J ETSET (begun in 1978). Originated in hadronisation studies.
Main feature: the Lund string fragmentation model.
• H ERWIG. Successor to E ARWIG (begun in 1984). Originated in perturbative coherence
studies. Main feature: angular-ordered parton showers.
• S HERPA. Begun in 2000. Originated in studies of the matching of hard-emission matrix
elements with parton showers. Main feature: CKKW matching.
There is also a large number of more specialised generators, mainly for hard processes within
and beyond the SM, a few that offer alternative shower models, and ones specializing in soft-
inclusive and/or heavy-ion physics.
An important aspect of contemporary generators is the ability to combine specialised ones
with general-purpose ones, via interfaces. The most common interface between partonic hard-
process and parton-shower generators is the Les Houches Event File (LHEF) standard, defined
in [105, 106] and “spoken” by most modern generator tools. For interfaces to experimental
analysis packages (like R IVET [107]) and detector simulations (like G EANT [108]), typically
the HepMC standard is used [109].
Hard processes were the topic of section 2. In this section, we shall focus mainly on parton
showers, with some brief comments on resonance decays at the end. Section 4 then concerns
the matching of matrix elements and parton showers. Finally, models of hadronisation and
the underlying event are the topic of section 5.
Several of the discussions below rely on material from the section on Monte Carlo Event
Generators in the PDG Review of Particle Physics [31] and on the more comprehensive review
by the MCnet collaboration in [6]. The latter also contains brief descriptions of the physics
implementations of each of the main general-purpose event generators on the market, together
with a guide on how to use (and not use) them in various connections, and a collection of
comparisons to important experimental distributions. We highly recommend readers to obtain
a copy of that review, as it is the most comprehensive and up-to-date review of event generators
currently available. Another useful and pedagogical review on event generators is contained
in the 2006 ESHEP lectures by Torbjörn Sjöstrand [50], with a more recent update in [110].

— 32 —
P. Skands Introduction to QCD

Relative uncertainty with n points 1-Dim d-Dim neval /point

Trapezoidal Rule 1/n2 1/n2/d 2d
Simpson’s Rule 1/n4 1/n4/d 3d
√ √
Monte Carlo 1/ n 1/ n 1

Table 2: Relative uncertainty after n evaluations, in 1 and d dimensions, for two traditional
numerical integration methods and stochastic Monte Carlo. The last column shows the number
of function evaluations that are required per point, in d dimensions.

3.1 The Monte Carlo Method

A ubiquitous problem in fundamental physics is the following: given a source located some
distance from a detector, predict the number of counts that should be observed within the solid
angle spanned by the detector (or within a bin of its phase-space acceptance), as a function
of the properties of the source, the intervening medium, and the efficiency of the detector.
Essentially, the task is to compute integrals of the form
Z
dσ
NCount (∆Ω) = dΩ , (60)
∆Ω dΩ

with dσ a differential cross section for the process of interest.

In particle physics, phase space has three dimensions per final-state particle (minus four
for overall four-momentum-conservation). Thus, for problems with more than a few outgoing
particles, the dimensionality of phase space increases rapidly. At LEP, for instance, the total
multiplicity of neutral + charged hadrons (before weak decays) was typically ∼ 30 particles,
for about 86 dimensions.
The standard 1D numerical-integration methods give very slow convergence rates for
higher-dimensional problems. For illustration, a table of convergence rates in 1 and d dimen-
sions is given in table 2, comparing the Trapezoidal (2-point) rule and Simpson’s (3-point) rule
to random-number-based Monte Carlo. In 1D, the 1/n2 convergence rate of the Trapezoidal
√
rule is much faster than the stochastic 1/ n of random-number Monte Carlo, and Simpson’s
rule converges even faster. However, as we go to d dimensions, the convergence rate of the
n-point rules all degrade with d (while the number of function evaluations required for each
“point” simultaneously increases). The MC convergence rate, on the other hand, remains
√
the simple stochastic 1/ n, independent of d, and each point still only requires one function
evaluation. These are some of the main reasons that MC is the preferred numerical integra-
tion technique for high-dimensional problems. In addition, the random phase-space vectors
it generates can be re-used in many ways, for instance as input to iterative solutions, to com-
pute many different observables simultaneously, and/or to hand “events” to propagation and
detector-simulation codes.
Therefore, virtually all numerical cross section calculations are based on Monte Carlo tech-
niques in one form or another, the simplest being the R AMBO algorithm [111] which can be
expressed in about half a page of code and generates a flat scan over n-body phase space21 .
21
Strictly speaking, R AMBO is only truly uniform for massless particles. Its massive variant makes up for phase-
space biases by returning weighted momentum configurations.

— 33 —
P. Skands Introduction to QCD

“This risk, that convergence is only given

with a certain probability, is inherent in
Monte Carlo calculations and is the rea-
son why this technique was named after the
world’s most famous gambling casino. In-
deed, the name is doubly appropriate because
the style of gambling in the Monte Carlo
casino, not to be confused with the noisy and
tasteless gambling houses of Las Vegas and
Reno, is serious and sophisticated.”

F. James, “Monte Carlo theory and practice”,

Rept. Prog. Phys. 43 (1980) 1145

Figure 15: Left: The casino in Monaco. Right: extract from [11] concerning the nature of
Monte Carlo techniques.

However, due to the infrared singularities in perturbative QCD, and due to the presence of
short-lived resonances, the functions to be integrated, |MF +k |2 , can be highly non-uniform,
especially for large k. This implies that we will have to be clever in the way we sample
phase space if we want the integration to converge in any reasonable amount of time — sim-
ple algorithms like R AMBO quickly become inefficient for k greater than a few. To address
this bottleneck, the simplest step up from R AMBO is to introduce generic (i.e., automated)
importance-sampling methods, such as offered by the V EGAS algorithm [112, 113]. This is
still the dominant basic technique, although most modern codes do employ several additional
refinements, such as several different copies of V EGAS running in parallel (multi-channel in-
tegration), to further optimise the sampling. Alternatively, a few algorithms incorporate the
singularity structure of QCD explicitly in their phase-space sampling, either by directly gen-
erating momenta distributed according to the leading-order QCD singularities, in a sort of
“QCD-preweighted” analog of R AMBO, called S ARGE [114], or by using all-orders Markovian
parton showers to generate them (V INCIA [90, 91, 115]).
The price of using random numbers is that we must generalise our notion of convergence.
In calculus, we say that a sequence {A} converges to B if an n exists for which the difference
|Ai>n − B| <  for any  > 0. In random-number-based techniques, we cannot completely rule
out the possibility of very pathological sequences of “dice rolls” leading to large deviations
from the true goal, hence we are restricted to say that {A} converges to B if an n exists for
which the probability for |Ai>n −B| < , for any  > 0, is greater than P , for any P ∈ [0, 1] [11].
This risk, that convergence is only given with a certain probability, is the reason why Monte
Carlo techniques were named after the famous casino in Monaco, illustrated in figure 15.

3.2 Theoretical Basis of Parton Showers

In section 2, we noted two conditions that had to be valid for fixed-order truncations of the
perturbative series to be valid: firstly, the strong coupling αs must be small for perturbation
theory to be valid at all. This restricts us to the region in which all scales Qi  ΛQCD .
We shall maintain this restriction in this section, i.e., we are still considering perturbative QCD.

— 34 —
P. Skands Introduction to QCD

Secondly, however, in order to be allowed to truncate the perturbative series, we had to require
σk+1  σk , i.e., the corrections at successive orders must become successively smaller, which
— due to the enhancements from soft/collinear singular (conformal) dynamics — effectively
restricted us to consider only the phase-space region in which all jets are “hard and well-
separated”, equivalent to requiring all Qi /Qj ≈ 1. In this section, we shall see how to lift this
restriction, extending the applicability of perturbation theory into regions that include scale
hierarchies, Qi  Qj  ΛQCD , such as occur for soft jets, jet substructure, etc.
In fact, the simultaneous restriction to all resolved scales being larger than ΛQCD and
no large hierarchies is extremely severe, if taken at face value. Since we collide and observe
hadrons (→ low scales) while simultaneously wishing to study short-distance physics processes
(→ high scales), it would appear trivial to conclude that fixed-order pQCD is not applicable to
collider physics at all. So why do we still use it?
The answer lies in the fact that we actually never truly perform a fixed-order calculation in
QCD. Let us repeat the factorised formula for the cross section, equation (??), now inserting
also a function, D, to represent the fragmentation of the final-state partons into observable
hadrons,

dσ XZ 1 XZ dσ̂ij→f
= dxi dxj dΦf fi/h1 (xi , µ2F )fj/h2 (xj , µ2F ) Df (Ô → O, µ2F ) , (61)
dO d Ô
i,j 0 f

with Ô denoting the observable evaluated on the partonic final state, and O the observable
evaluated on the hadronic final state, after fragmentation. Although the partonic cross sec-
tion, dσ̂ij→f , does represent a fixed-order calculation, the parton densities, fi/h1 and fj/h2 ,
include so-called resummations of perturbative corrections to all orders from the initial scale
of order the mass of the proton, up to the factorisation scale, µF (see section 2.2 and/or the
TASI lectures by Sterman [52]). Note that the oft-stated mantra that the PDFs are purely
non-perturbative functions is therefore misleading. True, they are defined as essentially non-
perturbative functions at some very low scale, µ0 ∼ a few GeV, but, if µF is taken large, they
necessarily incorporate a significant amount of perturbative physics as well. On the “fixed-
order side”, all we have left to ensure in dσij→f is then that there are no large hierarchies
remaining between µF and the QCD scales appearing in Φf . Likewise, in the final state, the
fragmentation functions, Df , include infinite-order resummations of perturbative corrections
all the way from µF down to some low scale, with similar caveats concerning mantras about
their non-perturbative nature as for the PDFs.

3.2.1 Step One: Infinite Legs

The infinite-order resummations that are included in objects such as the PDFs and FFs in
equation (61) (and in their parton-shower equivalents) rely on some very simple and powerful
properties of gauge field theories that were already touched on in section 2. In particular, we
saw in section 2.4 that we can represent all the infrared (IR) limits of any NLO amplitude with
a set of simple universal functions, based solely on knowing which partons are color-connected
(i.e., have color-space index contractions) with one another.
The diagrams in figure 16 show the basic origin of the universal IR singularities of gauge
theory amplitudes. On the left is shown a diagram (squared) in which an emission with small
sij interferes with itself. In the collinear limit, sij → 0, the propagator of the parent parton,

— 35 —
P. Skands Introduction to QCD

i i

j
I I

j
K K

k k

Figure 16: Diagrams (squared) giving rise to collinear (left) and soft (right) singularities.

I, goes on shell; the singularity of the associated propagator factor is the origin of the 1/sij
collinear singularities. On the right is shown the interference between a diagram with emission
from parton I and one with emission from parton K. The resulting term has propagator
singularities when both partons I and K go on shell, which can happen simultaneously if
parton j is soft. This generates the 2sik /(sij sjk ) soft singularity, also called the soft eikonal
factor or the dipole factor.
We now understand the fundamental origin of the IR singularities, why they are universal,
and why amplitudes factorise in the soft and collinear limits — the singularities are simply
generated by intermediate parton propagators going on shell, which is independent of the
nature of the hard process, and hence can be factorised from it.
Thus, for each pair of (massless) color-connected partons I and K in F , the squared am-
plitude for F + 1 gluon, |MF +1 |2 , will include a factor
 
2 2sik
|MF +1 | = gs NC2
+ collinear terms |MF |2 , (62)
sij sjk
| {z }
Antenna Function
where gs2 = 4παs is the strong coupling, i and k represent partons I and K after the branching
(i.e., they include possible recoil effects) and sij is the invariant between parton i and the
emitted parton, j.
The branching phase space of a color dipole (i.e., a pair of partons connected by a color-
index contraction) is illustrated in figure 17. Expressed in the branching invariants, sij and sjk ,
the phase space has a characteristic triangular shape, imposed by the relation s = sij +sjk +sik
(assuming massless partons). Sketchings of the post-branching parton momenta have been
inserted in various places in the figure, for illustration. The soft singularity is located at the
origin of the plot and the collinear regions lie along the axes.
The collinear terms for a q q̄ → qg q̄ “antenna” are unambiguous and are given in section 2.4.
Since gluons are in the adjoint representation, they carry both a color and an anticolor index
(one corresponding to the rows and the other to the columns of the Gell-Mann matrices),
and there is therefore some ambiguity concerning how to partition collinear radiation among
the two antennae they participate in. This is discussed in more detail in [90]. Differences
are subleading, however, and for our purposes here we shall consider gluon antenna ends as
radiating just like quark ones. The difference between quark and gluon radiation then arise
mainly because gluons participate in two antennae, while quarks only participate in one. This

— 36 —
P. Skands Introduction to QCD

sjk
P HASE S PACE FOR 2 → 3
KINEMATICS INCLUDING (E, P ) CONS
Original Dipole-Antenna:
I K

Collinear with K

Collinear with I sij

Soft

Figure 17: Illustration of the branching phase space for q q̄ → qg q̄, with the original dipole-
antenna oriented horizontally, the two parents sharing the transverse component of recoil, and
the azimuthal angle φ (representing rotations of the emitted parton around the dipole axis)
chosen such that the gluon is radiated upwards. From [90].

is related to the difference between the color factors, CA ∼ 2CF .

The problem that plagued the fixed-order truncations in section 2 is clearly visible in equa-
tion (62): if we integrate over the entire phase space including the region sij → 0, sjk → 0,
we end up with a double pole. If we instead regulate the divergence by cutting off the inte-
gration at some minimal perturbative cutoff scale µ2IR , we end up with a logarithm squared of
that scale. This is a typical example of “large logarithms” being generated by the presence of
scale hierarchies. Also note that the precise definition of µIR is not unique. Any scale choice
that properly isolates the singularities from the rest of phase space will do, with some typical
choices being, for example, invariant-mass and/or transverse-momentum scales.
Before we continue, it is worth noting that equation (62) is often rewritten in other forms
to emphasise specific aspects of it. One such rewriting is thus to reformulate the invariants sij
appearing in it in terms of energies and angles,

sij = 2Ei Ej (1 − cos θij ) . (63)

Rewritten in this way, the differentials can be partial-fractioned,

dsij dsjk dEj dθij dEj dθjk

∝ + . (64)
sij sjk Ej θij Ej θjk

This kind of rewriting enables an intuitively appealing categorisation of the singularities as

related to vanishing energies and angles, explaining why they are called soft and collinear,
respectively. Arguments based on this rewriting have led to important insights in QCD. For
instance, within the framework of conventional parton showers, it was shown in a sequence
of publications (see [116,117] and references therein) that the destructive interference effects
between two colour-connected partons (coherence) can be described, on average22 , by using
22
Averaged over azimuthal emission angles.

— 37 —
P. Skands Introduction to QCD

Dipole p⊥ Dipole Mass Angle Pythia p⊥

2
Angular Ordering pTevol -Ordering: side a
1.0 1.0 1.0 1.0
0.2
0.8 0.6 0.8 0.8 0.8

yrb = srb sarb = 1-xa

yrb = srb sarb = 1-xa

0.25
0.6 0.6 0.4 0.8 0.6 0.6
yjk

yjk
0.4 0.8 0.4 0.6 0.8 0.4 0.4
0.6
0.2 0.5
0.2 0.2 0.4 0.2 0.2
0.4 0.75
0.2 0.25 0.5
0.0 0.0 0.75
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
yij yij yar = sar sarb = 1-xb yar = sar sarb = 1-xb

Figure 18: A selection of parton-shower evolution variables, represented as contours over the
dipole phase space. Note: the right-most variable corresponds to evolution of only one of the
parents, the one with no collinear singularity along the bottom of the plot.

the opening angle of the emissions as the shower ordering variable. One should still keep
in mind, however, that Lorentz non-invariant formulations come with similar caveats and
warnings as do gauge non-invariant formulations of quantum field theory: while they can be
practical to work with at intermediate stages of a calculation, one should be careful with any
physical conclusions that rely explicitly on them.
We shall therefore here restrict ourselves to a Lorentz-invariant formalism based directly
on equation (62), pioneered by the dipole formulation of QCD cascades [86]. The collinear
limit is then replaced by a more general single-pole limit in which a single parton-parton in-
variant vanishes (as, for instance, when a pair of partons become collinear), while the soft limit
is replaced by one in which two (or more) invariants involving the same parton vanish simul-
taneously (as, for instance by that parton becoming soft in a frame defined by two or more
hard partons). This avoids frame-dependent ambiguities from entering into the language, at
the price of a slight reinterpretation of what is meant by collinear and soft.
In the generator landscape, angular ordering is used by the H ERWIG [117] and H ERWIG++ [118]
programs, and an angular veto is imposed on the virtuality-ordered evolution in P YTHIA 6 [119].
Variants of the dipole/antenna approach is used by the A RIADNE [120], S HERPA [121, 122],
and V INCIA [115] programs, while the p⊥ -ordered showers in P YTHIA 6 and 8 represent a
hybrid, combining collinear splitting kernels with dipole kinematics [123]. Phase-space con-
tours of equal value of some of these choices are illustrated in figure 18. During the shower
evolution, each model effectively “sweeps” over phase space in the order implied by these con-
tours. E.g., a p⊥ -ordered dipole shower (leftmost plot in figure 18) will treat a hard-collinear
branching as occurring “earlier” than a soft one, while a mass-ordered dipole shower (second
plot) will tend to do the opposite. This affects the tower of virtual corrections generated by
each shower model via the so-called Sudakov factor, discussed below. Experimental tests of
the subleading aspects of shower models can therefore quite important, see e.g. [124] for a
recent example.
Independently of rewritings and philosophy, the real power of equation (62) lies in the fact
that it is universal. Thus, for any process F , we can apply equation (62) in order to get an
approximation for dσF +1 . We may then, for instance, take our newly obtained expression for
F + 1 as our arbitrary process and crank equation (62) again, to obtain an approximation for
dσF +2 , and so forth. What we have here is therefore a very simple recursion relation that can
be used to generate approximations to leading-order cross sections with arbitrary numbers of

— 38 —
P. Skands Introduction to QCD

additional legs. The quality of this approximation is governed by how many terms besides the
leading one shown in equation (53) are included in the game. Including all possible terms,
the most general form for the cross section at F + n jets, restricted to the phase-space region
above some infrared cutoff scale µIR , has the following algebraic structure,
(0)

σF +n = αsn ln2n + ln2n−1 + ln2n−2 + . . . + ln + F (65)

where we use the notation lnλ without an argument to denote generic functions of transcen-
dentality λ (the logarithmic function to the power λ being a “typical” example of a function
with transcendentality λ appearing in cross section expressions, but also dilogarithms and
higher logarithmic functions23 of transcendentality > 1 should be implicitly understood to be-
long to our notation lnλ ). The last term, F, represents a rational function of transcendentality
0. We shall also use the nomenclature singular and finite for the lnλ and F terms, respectively,
a terminology which reflects their respective behavior in the limit µIR → 0.
The simplest approximation one can build on equation (65), dropping all but the leading
2n
ln term in the parenthesis, is thus the leading-transcendentality approximation. This approx-
imation is better known as the DLA (double logarithmic approximation), since it generates the
correct coefficient for terms which have two powers of logarithms for each power of αs , while
terms of lower transcendentalities are not guaranteed to have the correct coefficients. In so-
called LL (leading-logarithmic) parton shower algorithms, one generally expects to reproduce
the correct coefficients for the ln2n and ln2n−1 terms. In addition, several formally subleading
improvements are normally also introduced in such algorithms (such as explicit momentum
conservation, gluon polarisation and other spin-correlation effects [125–127], higher-order
coherence effects [116], renormalisation scale choices [128], finite-width effects [129], etc),
as a means to improve the agreement with some of the more subleading coefficients as well,
if not in every phase-space point then at least on average. Though LL showers do not mag-
ically acquire NLL (next-to-leading-log) precision from such procedures, one therefore still
expects a significantly better average performance from them than from corresponding “strict”
LL analytical resummations. A side effect of this is that it is often possible to “tune” shower
algorithms to give better-than-nominal agreement with experimental distributions, by adjust-
ing the parameters controlling the treatment of subleading effects. One should remember,
however, that there is a limit to how much can be accomplished in this way — at some point,
agreement with one process will only come at the price of disagreement with another, and at
this point further tuning would be meaningless.
Applying such an iterative process on a Born-level cross section, one obtains the description
of the full perturbative series illustrated in figure 19. The yellow (lighter) shades used here for
k ≥ 1 indicate that the coefficient obtained is not the exact one, but rather an approximation to
it that only gets its leading singularities right. However, since this is still only an approximation
to infinite-order tree-level cross sections (we have not yet included any virtual corrections), we
cannot yet integrate this approximation over all of phase space, as illustrated by the yellow
boxes being only half filled in figure 19; otherwise, the summed total cross section would still
be infinite. This particular approximation would therefore still appear to be very useless indeed
— on one hand, it is only guaranteed to get the singular terms right, but on the other, it does
23
Note: due to the theorems that allow us, for instance, to rewrite dilogarithms in different ways with logarithmic
and lower “spillover” terms, the coefficients at each λ are only well-defined up to reparametrisation ambiguities
involving the terms with transcendentality greater than λ.

— 39 —
P. Skands Introduction to QCD

F @ LO×LL(non-unitary)
(2) (2)
2 σ0 σ1 ...

` (loops)
(1) (1) (1)
1 σ0 σ1 σ2 ...

(0) (0) (0) (0)

0 σ0 σ1 σ2 σ3 ...
0 1 2 3 ...
k (legs)

Figure 19: Coefficients of the perturbative series covered by LO + LL approximations to higher-

multiplicity tree-level matrix elements. Green (darker) shading represents the full perturbative
coefficient at the respective k and `. Yellow (lighter) shading represents an LL approximation
to it. Half-shaded boxes indicate phase spaces in which we are prohibited from integrating
over the IR singular region, as discussed in sections 2.3 and 4.

not actually allow us to integrate over the singular region. In order to obtain a truly all-orders
calculation, the constraint of unitarity must also be explicitly imposed, which furnishes an
approximation to all-orders loop corrections as well. Let us therefore emphasise that figure 19
is included for pedagogical purposes only; all resummation calculations, whether analytical
or parton-shower based, include virtual corrections as well and consequently yield finite total
cross sections, as will now be described.

3.2.2 Step Two: Infinite Loops

Order-by-order unitarity, such as used in the KLN theorem, implies that the singularities caused
by integration over unresolved radiation in the tree-level matrix elements must be canceled,
order by order, by equal but opposite-sign singularities in the virtual corrections at the same
order. That is, from equation (62), we immediately know that the 1-loop correction to dσF
must contain a term,
Z  
(0) (1)∗ 2 (0) 2 dsij dsjk 2sik
2Re[MF MF ] ⊃ −gs NC MF + less singular terms , (66)
16π 2 sijk sij sjk
that cancels the divergence coming from equation (62) itself. Further, since this is universally
true, we may apply equation (66) again to get an approximation to the corrections generated
by equation (62) at the next order and so on. By adding such terms explicitly, order by order,
we may now bootstrap our way around the entire perturbative series, using equation (62) to
move horizontally and equation (66) to move along diagonals of constant n = k + `. Since
real-virtual cancellations are now explicitly restored, we may finally extend the integrations
over all of phase space, resulting in the picture shown in figure 20.
The picture shown in figure 20, not the one in figure 19, corresponds to what is actually
done in resummation calculations, both of the analytic and parton-shower types24 . Physically,
there is a significant and intuitive meaning to the imposition of unitarity, as follows.
24
In the way these calculations are formulated in practice, they in fact rely on one additional property, called
exponentiation, that allows us to move along straight vertical lines in the loops-and-legs diagrams. However, since
the two different directions furnished by equations (62) and (66) are already sufficient to move freely in the full
2D coefficient space, we shall use exponentiation without extensively justifying it here.

— 40 —
P. Skands Introduction to QCD

F @ LO×LL(unitary)
(2) (2)
2 σ0 σ1 ...

` (loops)
(1) (1) (1)
1 σ0 σ1 σ2 ...

(0) (0) (0) (0)

0 σ0 σ1 σ2 σ3 ...
0 1 2 3 ...
k (legs)

Figure 20: Coefficients of the perturbative series covered by LO + LL calculations, impos-

ing unitarity order by order for each n = k + `. Green (darker) shading represents the full
perturbative coefficient at the respective k and `. Yellow (lighter) shading represents an LL
approximation to it.

Take a jet algorithm, with some measure of jet resolution, Q, and apply it to an arbitrary
sample of events, say dijets. At a very crude resolution scale, corresponding to a high value
for Q, you find that everything is clustered back to a dijet configuration, and the 2-jet cross
section is equal to the total inclusive cross section,

σtot = σF ;incl . (67)

At finer resolutions, decreasing Q, you see that some events that were previously classified as
2-jet events contain additional, lower-scale jets, that you can now resolve, and hence those
events now migrate to the 3-jet bin, while the total inclusive cross section of course remains
unchanged,
σtot = σF ;incl = σF ;excl (Q) + σF +1;incl (Q) , (68)
where “incl” and “excl” stands for inclusive and exclusive cross sections25 , respectively, and the
Q-dependence in the two terms on the right-hand side must cancel so that the total inclusive
cross section is independent of Q. Later, some 3-jet events now migrate further, to 4 and higher
jets, while still more 2-jet events migrate into the 3-jet bin, etc. For arbitrary n and Q, we have
n−1
X
σF +n;incl (Q) = σF ;incl − σF +m;excl (Q) . (69)
m=0

This equation expresses the trivial fact that the cross section for n or more jets can be computed
as the total inclusive cross section for F minus a sum over the cross sections for F + exactly
m jets including all m < n. On the theoretical side, it is these negative terms which must
be included in the calculation, for each order n = k + `, to restore unitarity. Physically, they
express that, at a given scale Q, each event will be classified as having either 0, 1, 2, or
whatever jets. Or, equivalently, for each event we gain in the 3-jet bin as Q is lowered, we
must loose one event in the 2-jet one; the negative contribution to the 2-jet bin is exactly
minus the integral of the positive contribution to the 3-jet one, and so on. We may perceive
of this detailed balance as an evolution of the event structure with Q, for each event, which is
effectively what is done in parton-shower algorithms, to which we shall return in Section 3.3.
25 P∞
F inclusive = F plus anything. F exclusive = F and only F . Thus, σF ;incl = k=0 σF +k;excl

— 41 —
P. Skands Introduction to QCD

3.3 Perturbation Theory with Markov Chains

Consider again the Born-level cross section for an arbitrary hard process, F , differentially in
an arbitrary infrared-safe observable O, as obtained from equation (45):
(0) Z
dσF (0)
= dΦF |MF |2 δ(O − O(ΦF )) , (70)
dO
Born
where the integration runs over the full final-state on-shell phase space of F (this expression
and those below would also apply to hadron collisions were we to include integrations over
the parton distribution functions in the initial state), and the δ function projects out a 1-
dimensional slice defined by O evaluated on the set of final-state momenta which we denote
ΦF .
To make the connection to parton showers, we insert an operator, S, that acts on the
Born-level final state before the observable is evaluated, i.e.,
Z
dσF (0)
= dΦF |MF |2 S(ΦF , O) . (71)
dO S

Formally, this operator — the evolution operator — will be responsible for generating all
(real and virtual) higher-order corrections to the Born-level expression. The measurement δ
function appearing explicitly in equation (70) is now implicit in S.
Algorithmically, parton showers cast S as an iterative Markov (i.e., history-independent)
chain, with an evolution parameter, QE , that formally represents the factorisation scale of
the event, below which all structure is summed over inclusively. Depending on the partic-
ular choice of shower algorithm, QE may be defined as a parton virtuality (virtuality-order
showers), as a transverse-momentum scale (p⊥ -ordered showers), or as a combination of en-
ergies times angles (angular ordering). Regardless of the specific form of QE , the evolution
parameter will go towards zero as the Markov chain develops, and the event structure will
become more and more exclusively resolved. A transition from a perturbative evolution to a
non-perturbative one can also be inserted, when the evolution reaches an appropriate scale,
typically around 1 GeV. This scale, called the hadronisation scale, thus represents the lowest
perturbative scale that can appear in the calculations, with all perturbative corrections below
it summed over inclusively.
Working out the precise form that S must have in order to give the correct expansions
discussed in section 3.2 takes a bit of algebra, and is beyond the scope we aim to cover in these
lectures. Heuristically, the procedure is as follows. We noted that the singularity structure of
QCD is universal and that at least its first few terms are known to us. We also saw that we
could iterate that singularity structure, using universality and unitarity, thereby bootstrapping
our way around the entire perturbative series. This was illustrated by figure 20 in section 3.2.
Skipping intermediate steps, the form of the all-orders pure-shower Markov chain, for the

— 42 —
P. Skands Introduction to QCD

evolution of an event between two scales Q1 > QE > Q2 , is,

S(ΦF , Q1 , Q2 , O) = ∆(ΦF , Q1 , Q2 ) δ (O − O(ΦF ))

| {z }
F + 0 exclusive above Q2
X Z QE1 dΦrF +1
+ Sr (ΦF +1 ) ∆(ΦF , Q1 , QF +1 ) S(ΦF +1 , QF +1 , Q2 , O) ,
r QE2 dΦF
| {z }
F + 1 inclusive above Q2
(72)
with the so-called Sudakov factor,
" #
X Z Q1 dΦrF +1
∆(ΦF , Q1 , Q2 ) = exp − Sr (ΦF +1 ) , (73)
r Q2 dΦF

defining the probability that there is no evolution (i.e., no emissions) between the scales Q1 and
Q2 , according to the radiation functions Sr to which we shall return below. The term on the first
line of equation (72) thus represents all events that did not evolve as the resolution scale was
lowered from Q1 to Q2 , while the second line contains a sum and phase-space integral over
those events that did evolve — including the insertion of S(ΦF +1 ) representing the possible
further evolution of the event and completing the iterative definition of the Markov chain.
The factor dΦrF +1 /dΦF defines the chosen phase space factorisation. Our favorite is the
so-called dipole-antenna factorisation, whose principal virtue is that it is the simplest Lorentz
invariant factorisation which is simultaneously exact over all of phase space while only involv-
ing on-shell momenta. For completeness, its form is

dΦrF +1 dΦr3 dφ 1
= = dsa1 ds1b , (74)
dΦF dΦ2 2π 16π 2 sr

which involves just one color-anticolor pair for each r, with invariant mass squared sr =
(pa + p1 + pb )2 . Other choices, such as purely collinear ones (only exact in the collinear
limit or involving explicitly off-shell momenta), more global ones involving all partons in the
event (more complicated, in our opinion), or less global ones with a single parton playing
the dominant role as emitter, are also possible, again depending on the specific algorithm
considered.
The radiation functions Sr obviously play a crucial role in these equations, driving the
emission probabilities. For example, if Sr → 0, then ∆ → exp(0) = 1 and all events stay in the
top line of equation (72). Thus, in regions of phase space where Sr is small, there is little or
no evolution. Conversely, for Sr → ∞, we have ∆ → 0, implying that all events evolve. One
possible choice for the radiation functions Sr was implicit in equation (62), in which we took
them to include only the leading (double) singularities, with r representing color-anticolor
pairs. In general, the shower may exponentiate the entire set of universal singular terms, or
only a subset of them (for example, the terms leading in the number of colors NC ), which is
why we here let the explicit form of Sr be unspecified. Suffice it to say that in traditional parton
showers, Sr would simply be the DGLAP splitting kernels (see, e.g., [4]), while they would be
so-called dipole or antenna radiation functions in the various dipole-based approaches to QCD
(see, e.g., [73, 77, 80, 86, 90, 91, 122, 130]).

— 43 —
P. Skands Introduction to QCD

The procedure for how to technically “construct” a shower algorithm of this kind, using
random numbers to generate scales distributed according to equation (72), is described more
fully in [90], using a notation that closely parallels the one used here. The procedure is also
described at a more technical level in the review [6], though using a slightly different notation.
Various aspects of the Sudakov veto algorithm are discussed in [13–15]. Finally, pedagogical
introduction to Monte Carlo methods in general can be found in [11, 12].

3.4 Decays of Unstable Particles

In most BSM processes and some SM ones, an important aspect of the event simulation is how
decays of short-lived particles, such as top quarks, EW and Higgs bosons, and new BSM reso-
nances, are handled. We here briefly summarise the spectrum of possibilities, but emphasise
that there is no universal standard. Users are advised to check whether the treatment of a
given code is adequate for the physics study at hand.
The appearance of an unstable resonance as a physical particle at some intermediate stage
of the event generation implies that its production and decay processes are treated as being
factorised. This is called the narrow width approximation and is valid up to corrections of
order Γ/m0 , with Γ the width and m0 the pole mass of the particle. States whose widths are a
substantial fraction of their mass should not be treated in this way, but rather as intrinsically
virtual internal propagator lines.
For states treated as physical particles, two aspects are relevant: the mass distribution of
the decaying particle itself and the distributions of its decay products. For the mass distribu-
tion, the simplest is to use a δ function at m0 . The next level up, typically used in general-
purpose Monte Carlos, is to use a Breit-Wigner distribution (relativistic or non-relativistic),
which formally resums higher-order virtual corrections to the mass distribution. Note, how-
ever, that this still only generates an improved picture for moderate fluctuations away from
m0 . Similarly to above, particles that are significantly off-shell (in units of Γ) should not be
treated as resonant, but rather as internal off-shell propagator lines. In most Monte Carlo
codes, some further refinements are also included, for instance by letting Γ be a function of m
(“running widths”) and by limiting the magnitude of the allowed fluctuations away from m0 .
See also [131] for an elaborate discussion of the Higgs boson lineshape.
For the distributions of the decay products, the simplest treatment is again to assign them
their respective m0 values, with a uniform (i.e., isotropic, or “flat”) phase-space distribution.
A more sophisticated treatment distributes the decay products according to the differential
decay matrix elements, capturing at least the internal dynamics and helicity structure of the
decay process, including EPR-like correlations. Further refinements include polarisations of
the external states [125–127] (see also [132–134] for phenomenological studies) and assign-
ing the decay products their own Breit-Wigner distributions, the latter of which opens the
possibility to include also intrinsically off-shell decay channels, like H → W W ∗ . Please refer
to the physics manual of the code you are using and/or make simple cross checks like plotting
the distribution of phase-space invariants it produces.
During subsequent showering of the decay products, most parton-shower models will pre-
serve the total invariant mass of each resonance-decay system separately, so as not to skew the
original resonance shape.

— 44 —
P. Skands Introduction to QCD

F @ LO×LL F+1 @ LO×LL F & F+1 @ LO×LL

(2) (2) (2) (2) (2) (2)
2 σ0 σ1 ... 2 σ0 σ1 ... 2 σ0 σ1 ...

+ =
` (loops)

` (loops)

` (loops)
(1) (1) (1) (1) (1) (1) (1) (1) (1)
1 σ0 σ1 σ2 ... 1 σ0 σ1 σ2 ... 1 σ0 σ1 σ2 ...

(0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0)
0 σ0 σ1 σ2 σ3 ... 0 σ0 σ1 σ2 σ3 ... 0 σ0 σ1 σ2 σ3 ...
0 1 2 3 ... 0 1 2 3 ... 0 1 2 3 ...
k (legs) k (legs) k (legs)

Figure 21: The double-counting problem caused by naively adding cross sections involving
matrix elements with different numbers of legs.

4 Matching at LO and NLO

The essential problem that leads to matrix-element/parton-shower matching can be illustrated
in a very simple way. Assume we have computed the LO cross section for some process, F ,
and that we have added an LL shower to it, as in the left-hand pane of figure 21. We know
that this only gives us an LL description of F + 1. We now wish to improve this from LL to LO
by adding the actual LO matrix element for F + 1. Since we also want to be able to hadronize
these events, etc, we again add an LL shower off them. However, since the matrix element for
F + 1 is divergent, we must restrict it to cover only the phase-space region with at least one
hard resolved jet, illustrated by the half-shaded boxes in the middle pane of figure 21.
Adding these two samples, however, we end up counting the LL terms of the inclusive cross
section for F + 1 twice, since we are now getting them once from the shower off F and once
from the matrix element for F + 1, illustrated by the dark shaded (red) areas of the right-
hand pane of figure 21. This double-counting problem would grow worse if we attempted to
add more matrix elements, with more legs. The cause is very simple. Each such calculation
corresponds to an inclusive cross section, and hence naive addition would give

σtot = σ0;incl + σ1;incl = σ0;excl + 2 σ1;incl . (75)

P∞equation (68): F inclusive = F

Recall the definition of inclusive and exclusive cross sections,
plus anything. F exclusive = F and only F . Thus, σF ;incl = k=0 σF +k;excl .
Instead, we must match the coefficients calculated by the two parts of the full calculation
— showers and matrix elements — more systematically, for each order in perturbation theory,
so that the nesting of inclusive and exclusive cross sections is respected without overcounting.
Given a parton shower and a matrix-element generator, there are fundamentally three
different ways in which we can consider matching the two [90]: slicing, subtraction, and
unitarity. The following subsections will briefly introduce each of these.

4.1 Slicing
The most commonly encountered matching type is currently based on separating (slicing)
phase space into two regions, one of which is supposed to be mainly described by hard ma-
trix elements and the other of which is supposed to be described by the shower. This type of
approach was first used in H ERWIG [100], to include matrix-element corrections for one emis-
sion beyond the basic hard process [135, 136]. This is illustrated in figure 22. The method

— 45 —
P. Skands Introduction to QCD

F @ LO×LL-Soft (H ERWIG Shower) F+1 @ LO×LL (H ERWIG Corrections) F @ LO1 ×LL (H ERWIG Matched)
(2) (2) (2) (2) (2) (2)
2 σ0 σ1 ... 2 σ0 σ1 ... 2 σ0 σ1 ...

+ =
` (loops)

` (loops)

` (loops)
(1) (1) (1) (1) (1) (1) (1) (1) (1)
1 σ0 σ1 σ2 ... 1 σ0 σ1 σ2 ... 1 σ0 σ1 σ2 ...

(0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0)
0 σ0 σ1 σ2 σ3 ... 0 σ0 σ1 σ2 σ3 ... 0 σ0 σ1 σ2 σ3 ...
0 1 2 3 ... 0 1 2 3 ... 0 1 2 3 ...
k (legs) k (legs) k (legs)

Figure 22: H ERWIG’s original matching scheme [135, 136], in which the dead zone of the
H ERWIG shower was used as an effective “matching scale” for one emission beyond a basic
hard process.

has since been generalized by several independent groups to include arbitrary numbers of ad-
ditional legs, the most well-known of these being the CKKW [137], CKKW-L [138, 139], and
MLM [140, 141] approaches.
Effectively, the shower approximation is set to zero above some scale, either due to the
presence of explicit dead zones in the shower, as in H ERWIG, or by vetoing any emissions
above a certain matching scale, as in the (L)-CKKW and MLM approaches. The empty part of
phase space can then be filled by separate events generated according to higher-multiplicity
tree-level matrix elements (MEs). In the (L)-CKKW and MLM schemes, this process can be
iterated to include arbitrary numbers of additional hard legs (the practical limit being around
3 or 4, due to computational complexity).
In order to match smoothly with the shower calculation, the higher-multiplicity matrix ele-
ments must be associated with Sudakov form factors (representing the virtual corrections that
would have been generated if a shower had produced the same phase-space configuration),
and their αs factors must be chosen so that, at least at the matching scale, they become identi-
cal to the choices made on the shower side [142]. The CKKW and MLM approaches do this by
constructing “fake parton-shower histories” for the higher-multiplicity matrix elements. By ap-
plying a sequential jet clustering algorithm, a tree-like branching structure can be created that
at least has the same dominant structure as that of a parton shower. Given the fake shower
tree, αs factors can be chosen for each vertex with argument αs (p⊥ ) and Sudakov factors can
be computed for each internal line in the tree. In the CKKW method, these Sudakov factors
are estimated analytically, while the MLM and CKKW-L methods compute them numerically,
from the actual shower evolution.
Thus, the matched result is identical to the matrix element (ME) in the region above the
matching scale, modulo higher-order (Sudakov and αs ) corrections. We may sketch this as

ME corrections
z }| { z }| {
Matched (above matching scale) = Exact × (1 + O(αs )) , (76)

where the “shower-corrections” include the approximate Sudakov factors and αs reweighting
factors applied to the matrix elements in order to obtain a smooth transition to the shower-
dominated region.
Below the matching scale, the small difference between the matrix elements and the
shower approximation can be dropped (since their leading singularities are identical and this

— 46 —
P. Skands Introduction to QCD

F @ LO×LL-Soft (excl) F+1 @ LO×LL-Soft (excl) F+2 @ LO×LL (incl) F @ LO2 ×LL (MLM & (L)-CKKW)
(2) (2) (2) (2)
2 σ0 ... 2 σ0 ... 2 σ0 ... 2 σ0 ...

+ + =
` (loops)

` (loops)

` (loops)

` (loops)
(1) (1) (1) (1) (1) (1) (1) (1)
1 σ0 σ1 ... 1 σ0 σ1 ... 1 σ0 σ1 ... 1 σ0 σ1 ...

(0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0)
0 σ0 σ1 σ2 0 σ0 σ1 σ2 0 σ0 σ1 σ2 0 σ0 σ1 σ2
0 1 2 0 1 2 0 1 2 0 1 2
k (legs) k (legs) k (legs) k (legs)

Figure 23: Slicing, with up to two additional emissions beyond the basic process. The showers
off F and F + 1 are set to zero above a specific “matching scale”. (The number of coefficients
shown was reduced a bit in these plots to make them fit in one row.)

region by definition includes no hard jets), yielding the pure shower answer in that region,

shower correction
z }| { z }| {
Matched (below matching scale) = Approximate + (Exact − Approximate)
= Approximate + non-singular
→ Approximate . (77)

This type of strategy is illustrated in figure 23.

As emphasized above, since this strategy is discontinuous across phase space, a main point
here is to ensure that the behavior across the matching scale be as smooth as possible. CKKW
showed [137] that it is possible to remove any dependence on the matching scale through
NLL precision by careful choices of all ingredients in the matching; technical details of the
implementation (affecting the O(αs ) terms in eq. (76)) are important, and the dependence
on the unphysical matching scale may be larger than NLL unless the implementation matches
the theoretical algorithm precisely [138, 139, 143]. Furthermore, since the Sudakov factors
are generally computed using showers (MLM, L-CKKW) or a shower-like formalism (CKKW),
while the real corrections are computed using matrix elements, care must be taken not to (re-
)introduce differences that could break the detailed real-virtual balance that ensures unitarity
among the singular parts, see e.g., [142].
It is advisable not to choose the matching scale too low. This is again essentially due
to the approximate scale invariance of QCD imploring us to write the matching scale as a
ratio, rather than as an absolute number. If one uses a very low matching scale, the higher-
multiplicity matrix elements will already be quite singular, leading to very large LO cross
sections before matching. After matching, these large cross sections are tamed by the Sudakov
factors produced by the matching scheme, and hence the final cross sections may still look
reasonable. But the higher-multiplicity matrix elements in general contain subleading singu-
larity structures, beyond those accounted for by the shower, and hence the delicate line of
detailed balance that ensures unitarity has most assuredly been overstepped. We therefore
recommend not to take the matching scale lower than about an order of magnitude below the
characteristic scale of the hard process.
One should also be aware that all strategies of this type are quite computing intensive.
This is basically due to the fact that a separate phase-space generator is required for each of
the n-parton correction terms, with each such sample a priori consisting of weighted events

— 47 —
P. Skands Introduction to QCD

F @ LO×LL (F @ NLO×LL) - (F @ LO×LL) F @ NLO×LL (M C @ NLO)

(2) (2) (2) (2) (2) (2)
2 σ0 σ1 ... 2 σ0 σ1 ... 2 σ0 σ1 ...

+ =
` (loops)

` (loops)

` (loops)
(1) (1) (1) (1) (1) (1) (1) (1) (1)
1 σ0 σ1 σ2 ... 1 σ0 σ1 σ2 ... 1 σ0 σ1 σ2 ...

(0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0)
0 σ0 σ1 σ2 σ3 ... 0 σ0 σ1 σ2 σ3 ... 0 σ0 σ1 σ2 σ3 ...
0 1 2 3 ... 0 1 2 3 ... 0 1 2 3 ...
k (legs) k (legs) k (legs)

Figure 24: M C @ NLO. In the middle pane, cyan boxes denote non-singular correction terms,
while the egg-colored ones denote showers off such corrections, which cannot lead to double-
counting at the LL level.

such that a separate unweighting step (often with quite low efficiency) is needed before an
unweighted sample can be produced.

4.2 Subtraction
Another way of matching two calculations is by subtracting one from the other and correcting
by the difference, schematically

shower correction
z }| { z }| {
Matched = Approximate + (Exact − Approximate) . (78)

This looks very much like the structure of a subtraction-based NLO fixed-order calculation,
section 2.4, in which the shower approximation here plays the role of subtraction terms, and
indeed this is what is used in strategies like M C @ NLO [144–146], illustrated in figure 24. In
this type of approach, however, negative-weight events will generally occur, for instance in
phase-space points where the approximation is larger than the exact answer.
Negative weights are not in principle an insurmountable problem. Histograms can be filled
with each event counted according to its weight, as usual. However, negative weights do affect
efficiency. Imagine a worst-case scenario in which 1000 positive-weight events have been
generated, along with 999 negative-weight ones (assuming each event weight has the same
absolute value, the closest one can get to an unweighted sample in the presence of negative
weights). The statistical precision of the MC answer would be equivalent to one event, for
2000 generated, i.e., a big loss in convergence rate.
In practice, generators like MC@NLO “only” produce around 10% or less events with nega-
tive weights, so the convergence rate should not be severely affected for ordinary applications.
Nevertheless, the problem of negative weights motivated the development of the so-called
P OWHEG approach [147], illustrated in figure 25, which is constructed specifically to prevent
negative-weight events from occurring and simultaneously to be more independent of which
parton-shower algorithm it is used with. In the P OWHEG method, one effectively modifies the
real-emission probability for the first emission to agree with the F + 1 matrix element (this is
covered under unitarity, below). One is then left with a purely virtual correction, which will
typically be positive, at least for processes for which the NLO cross section is larger than the
LO one.

— 48 —
P. Skands Introduction to QCD

F @ LO1 ×LL (F @ NLO×LL) - (F @ LO1 ×LL) F @ NLO×LL (P OWHEG)

(2) (2) (2) (2) (2) (2)
2 σ0 σ1 ... 2 σ0 σ1 ... 2 σ0 σ1 ...

+ =
` (loops)

` (loops)

` (loops)
(1) (1) (1) (1) (1) (1) (1) (1) (1)
1 σ0 σ1 σ2 ... 1 σ0 σ1 σ2 ... 1 σ0 σ1 σ2 ...

(0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0)
0 σ0 σ1 σ2 σ3 ... 0 σ0 σ1 σ2 σ3 ... 0 σ0 σ1 σ2 σ3 ...
0 1 2 3 ... 0 1 2 3 ... 0 1 2 3 ...
k (legs) k (legs) k (legs)

Figure 25: P OWHEG. In the middle pane, cyan boxes denote non-singular correction terms,
while the egg-colored ones denote showers off such corrections, which cannot lead to double-
counting at the LL level.

F @ NLO×LL-Soft (∼ P OWHEG) F +2 @ LOn ×LL (∼ CKKW for F+2) F @ NLO×LL (M ENLOPS)

(2) (2) (2) (2) (2) (2)
2 σ0 σ1 ... 2 σ0 σ1 ... 2 σ0 σ1 ...

+ =
` (loops)

` (loops)

` (loops)
(1) (1) (1) (1) (1) (1) (1) (1) (1)
1 σ0 σ1 σ2 ... 1 σ0 σ1 σ2 ... 1 σ0 σ1 σ2 ...

(0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0) (0)
0 σ0 σ1 σ2 σ3 ... 0 σ0 σ1 σ2 σ3 ... 0 σ0 σ1 σ2 σ3 ...
0 1 2 3 ... 0 1 2 3 ... 0 1 2 3 ...
k (legs) k (legs) k (legs)

Figure 26: M ENLOPS. Note that each of the P OWHEG and CKKW samples are composed of
separate sub-samples, as illustrated in figures 23 and 25.

The advantage of these methods is obviously that NLO corrections to the Born level can
be systematically incorporated. However, a systematic way of extending this strategy beyond
the first additional emission is not available, save for combining them with a slicing-based
strategy for the additional legs, as in M ENLOPS [148], illustrated in figure 26. These issues
are, however, no more severe than in ordinary fixed-order NLO approaches, and hence they
are not viewed as disadvantages if the point of reference is an NLO computation.

4.3 Unitarity
The oldest, and in my view most attractive, approach [119, 149] consists of working out the
shower approximation to a given fixed order, and correcting the shower splitting functions at
that order by a multiplicative factor given by the ratio of the matrix element to the shower
approximation, phase-space point by phase-space point. We may sketch this as

correction
shower z }| {
z }| { Exact
Matched = Approximate × . (79)
Approximate

When these correction factors are inserted back into the shower evolution, they guarantee
that the shower evolution off n − 1 partons correctly reproduces the n-parton matrix elements,
without the need to generate a separate n-parton sample. That is, the shower approximation
is essentially used as a pre-weighted (stratified) all-orders phase-space generator, on which a
more exact answer can subsequently be imprinted order by order in perturbation theory. Since

— 49 —
P. Skands Introduction to QCD

F @ LO1 +LL (P YTHIA) F @ NLO+LOn +LL (V INCIA)

(2) (2) (2) (2)
2 σ0 σ1 ... 2 σ0 σ1 ...

` (loops)

` (loops)
(1) (1) (1) (1) (1) (1)
1 σ0 σ1 σ2 ... 1 σ0 σ1 σ2 ...

(0) (0) (0) (0) (0) (0) (0) (0)

0 σ0 σ1 σ2 σ3 ... 0 σ0 σ1 σ2 σ3 ...
0 1 2 3 ... 0 1 2 3 ...
k (legs) k (legs)

Figure 27: P YTHIA (left) and V INCIA (right). Unitarity-based. Only one event sample is pro-
duced by each of these methods, and hence no sub-components are shown.

the shower is already optimized for exactly the kind of singular structures that occur in QCD,
very fast computational speeds can therefore be obtained with this method [91].
In the original approach [119, 149], used by P YTHIA [102, 103], this was only worked
out for one additional emission beyond the basic hard process. In P OWHEG [147, 150], it
was extended to include also virtual corrections to the Born-level matrix element. Finally, in
V INCIA, it has been extended to include arbitrary numbers of emissions at tree level [90, 91]
and one emission at loop level [151], though that method has so far only been applied to
final-state showers.
An illustration of the perturbative coefficients that can be included in each of these ap-
proaches is given in figure 27, as usual with green (darker shaded) boxes representing exact
coefficients and yellow (light shaded) boxes representing logarithmic approximations.
Finally, two more properties unique to this method deserve mention. Firstly, since the
corrections modify the actual shower evolution kernels, the corrections are automatically re-
summed in the Sudakov exponential, which should improve the logarithmic precision once
k ≥ 2 is included, and secondly, since the shower is unitary, an initially unweighted sam-
ple of (n − 1)-parton configurations remains unweighted, with no need for a separate event-
unweighting or event-rejection step.

— 50 —
P. Skands Introduction to QCD

5 Hadronisation and Soft Hadron-Hadron Physics

We here give a very brief overview of the main aspects of soft QCD that are relevant for hadron-
hadron collisions, such as hadronisation, minimum-bias and soft-inclusive physics, and the
so-called underlying event. This will be kept at a pedestrian level and is largely based on the
reviews in [6, 31, 152].
In the context of event generators, hadronisation denotes the process by which a set of
coloured partons (after showering) is transformed into a set of colour-singlet primary hadrons,
which may then subsequently decay further. This non-perturbative transition takes place at the
hadronisation scale Qhad , which by construction is identical to the infrared cutoff of the parton
shower. In the absence of a first-principles solution to the relevant dynamics, event generators
use QCD-inspired phenomenological models to describe this transition.
The problem can be stated as follows: given a set of partons resolved at a scale of Qhad ∼
1 GeV, we need a “mapping” from this set onto a set of on-shell colour-singlet (i.e., confined)
hadronic states. MC models do this in three steps:

1. Map the partonic system onto a continuum of high-mass hadronic states (called “strings”
or “clusters”).

2. Iteratively map strings/clusters onto discrete set of primary hadrons (via string breaks /
cluster splittings / cluster decays).

3. Sequential decays into secondaries (ρ → ππ, Λ → nπ, π 0 → γγ, ...).

The physics governing this mapping is non-perturbative. However, we do have some knowl-
edge of the properties that such a solution must have. For instance, Poincaré invariance,
unitarity, and causality are all concepts that apply beyond perturbation theory. In addition,
lattice QCD provides us a means of making explicit quantitative studies in a genuinely non-
perturbative setting (albeit only of certain questions).
An important result in “quenched” lattice QCD26 is that the potential of the colour-dipole
field between a charge and an anticharge appears to grow linearly with the separation of the
charges, at distances greater than about 0.5 femtometers; this behavior is illustrated by the
plot shown in figure 28, from [153]. (Note that the axes are scaled by units of the string
√
tension κ ∼ 420 MeV. Additional labels corresponding to 1 GeV and 2 GeV are also provided,
on the y axis, and to 1 fm and 2 fm, on the x axis.) This is known as “linear confinement”, and
it forms the starting point for the string model of hadronisation, discussed below in section 5.1.
Alternatively, a property of perturbative QCD called “preconfinement” [154] is the basis of the
cluster model of hadronisation, described in [6, 31].
In the generator landscape, P YTHIA [102,103], Q GSJET [155], and S IBYLL [156] use string
fragmentation models (as do A RIADNE [120], D PMJET [157], P HOJET [158], and V INCIA [115]
via interfaces to P YTHIA), while H ERWIG [101] and S HERPA [104] use cluster fragmentation.
E POS [159] uses a combination of string hadronisation and a hydrodynamics-inspired statisti-
cal hadronisation model.
It should be emphasised that the so-called parton level that can be obtained by switching
off hadronisation in an MC generator, is not a universal concept, since each model defines
26
Quenched QCD implies no “dynamical” quarks, i.e., no g → q q̄ splittings allowed.

— 51 —
P. Skands Introduction to QCD
46 STATIC QUARK-ANTIQUARK POTENTIAL: SCALING. . . 2641

Scaling plot

2GeV-

1 GeV—
2

B = 6.0, L=16 ~ 'V

B = 6.0, L=32 ~
I
~
~

B = 6.2, L=24
I B = 6.4, L-24
-2 k, B = 6.4, L=32 I
A

2'
t

0.5 1.5 1 fm 2.5 3.5 4

l~
RK
FIG. 4. All potential data of the five lattices have been scaled to a universal curve by subtracting Vo and measuring energies and
Figure
distances28:
in appropriate units of &E. The dashed
Static quark-antiquark potential,
curve correspond V(R) = R —
as a tofunction of~/12R.
separation distance,
Physical units in quenched
are calculated by exploit-
ing the relation &cr =420 MeV. √
lattice QCD, from [153]. Note that the axes are scaled by units of the string tension κ ∼ 420
MeV. Additional labels corresponding to 1 GeV and 2 GeV are also provided, on the y axis, and
to 1 fm and 2
AM~a=46. 1A~fm,&235(2)(13)
on the xMeVaxis. . A constant term,turbative V0 , hasresults.
beenAlthough
subtractedwe arefrom
aware all
thatthe
our results.
lattice
resolution is not yet really suScient, we might dare to
The dashed
Needless line
to say, thiscorresponds
value does not to V (R) =apply
necessarily R −toπ/(12R).
previe~ the continuum behavior of the Coulomb-like
full QCD. term from our results. In Fig. 6(a) [6(b)] we visualize the
In addition to the long-range behavior of the confining confidence regions in the K-e plane from fits to various
potential it is of considerable interest to investigate its ul- on- and off-axis scale 32 latticesbyat aP=6. 0
the hadronisation
traviolet structure. Asscale differently. E.g., the
we proceed into the weak cou-
hadronisation canonbe
potentials the defined cutoff
[6.4]. We observe that the impact of lattice discretization
inpling
invariant mass,
regime lattice transverse
simulations momentum,
are expected or
to meet per- some
on e other
decreasesquantity,
by a factor with different
2, as we P = 6.using
tunes
step up from 0 to
different values for the cutoff. The former is equivalent to using different effective factori-
sation schemes, and the latter corresponds to different factorisation scales, for the soft non-
perturbative component150of the modelling. Comparisons to distributions at this level (i.e., with
Barkai '84 o
hadronisation switched140 off) may therefore be used to provide an idea MTC of '90 the overall impact of
Our results:---
hadronisation corrections within a given model, but should be avoided in the context of phys-
ical observables. Note also
130-that the corresponding MC fragmentation functions are intrinsically
defined at the hadronisation scale. They can therefore not be compared directly to those that
are used in fixed-order120-
/ analytical-resummation contexts, which are typically defined at a
factorisation scale of order
110- the scale of the hard process.
We use the term “soft hadron-hadron physics” to comprise all scattering processes for which
100- is not required to be present27 . This includes elastic, diffractive,
a hard, perturbative scale
minimum-bias, and pile-up processes, as well as the physics contributing to the so-called un-
derlying event. We give a brief introduction to such processes in section 5.2.
We round off with a80— discussion of the data constraints that enter in the tuning of Monte
Carlo models in section 5.4, and 5.6
give an5.8outline of a procedure
6.2
that could6.4
be followed in a
realistic set-up.
27
FIG.
Note,5. however,
The on-axisthat
stringwhile
tensiona hard
[in unitsscale
of theis quantity c = &Eto
not required /(abeAL present, it is not
) ] as a function of P. explicitly
Our results required to be
are combined withabsent
pre-
either. Thus,obtained
vious values by the MTc minimum-bias,
both diffractive, collaboration [10] and Barkai,
pile-up andMoriarty, and Rebbi [11].
underlying-event processes will have tails towards high-
p⊥ physics as well. For example, even tt̄ pair production can be viewed as a tail of minimum-bias interactions, and
there is a tail of diffractive processes in which hard dijets can be produced diffractively (see, e.g., [160]).

— 52 —
P. Skands Introduction to QCD

1 fm
Figure 29: Illustration of the transition between a Coulomb potential at short distances to the
string-like one of equation (80) at large q q̄ separations.

5.1 The String Model of Hadronisation

Starting from early concepts developed by Artru and Mennessier [161], several hadronisation
models based on strings were proposed in the late 1970’ies and early 80’ies. Of these, the
most widely used today is the so-called Lund model, implemented in the P YTHIA code. We
shall therefore concentrate on that particular model here, though many of the overall concepts
would be shared by any string-inspired method. (More extended discussions can be found in
Andersson’s book [8] and in an older comprehensive Physics Reports review [7].)
Consider the production of a q q̄ pair from vacuum, for instance in the process e+ e− →
∗
γ /Z → q q̄ → hadrons. As the quarks move apart, linear confinement implies that a potential

V (R) = κ R (80)

is asymptotically reached for large distances, R. At short distances, there is a Coulomb term
proportional to 1/R as well, cf. figure 28, but this is neglected in the Lund model. Such a
potential describes a string with tension (energy per unit length) κ, with the value [153]

κ ∼ (420 MeV)2 ∼ 0.18 GeV2 ∼ 0.9 GeV/fm , (81)

which, for comparison with the world of macroscopic objects, would be sufficient to lift a
16-ton truck [162].
The string can be thought of as parameterizing the position of the axis of a cylindrically
symmetric flux tube, illustrated in figure 29. Such simple q − q̄ strings form the starting point
for the string model. More complicated multi-parton topologies are treated by representing
gluons as transverse “kinks”, e.g., q − g − q̄. The space-time evolution is then slightly more
involved [8], and modifications to the fragmentation model to handle stepping across gluon
corners have to be included, but the main point is that there are no separate free parameters
for gluon jets. Differences with respect to quark fragmentation arise simply because quarks
are only connected to a single string piece, while gluons have one on either side, increasing
the energy loss per unit (invariant) time from a gluon to the string by a factor of 2 relative
to quarks, which can be compared to the ratio of colour Casimirs CA /CF = 2.25. Another
appealing feature of the model is that low-energy gluons are absorbed smoothly into the string,
without leading to modifications. This improves the stability of the model with respect to
variations of the infrared behavior of the parton shower.
As the partonic string endpoints move apart, their kinetic energy is gradually converted
to potential energy, stored in the growing string spanned between them. In the “quenched”
approximation, in which g → q q̄ splittings are not allowed, this process would continue until
the endpoint quarks have lost all their momentum, at which point they would reverse direction
and be accelerated by the now shrinking string.

— 53 —
P. Skands String Break
Introduction to QCD

leftover string,
further breaks

time
q
space
a) b)

Figure 30: a) Illustration of string breaking by quark pair creation in the string field. b)
Illustration of the algorithmic choice to process the fragmentation from the outside-in, splitting
off a single on-shell hadron in each step.

26
In the real world, quark-antiquark fluctuations inside the string field can make the tran-
sition to become real particles by absorbing energy from the string, thereby screening the
original endpoint charges from each other and breaking the string into two separate colour-
singlet pieces, (q q̄) → (q q̄ 0 ) + (q 0 q̄), illustrated in figure 30 a. This process then continues
until only ordinary hadrons remain. (We will give more details on the individual string breaks
below.)
Since the string breaks are causally disconnected (as can easily be realised from space-
time diagrams like the one in figure 30, see also [8]), they do not have to be considered in any
specific time-ordered sequence. In the Lund model, the string breaks are instead generated
starting with the leading (“outermost”) hadrons, containing the endpoint quarks, and iterating
inwards towards the center of the string, alternating randomly between fragmentation off the
left- and right-hand sides, respectively, figure 30b. One can thereby split off a single well-
defined hadron in each step, with a mass that, for unstable hadrons, is selected according to a
Breit-Wigner distribution.
The details of the individual string breaks are not known from first principles. The Lund
model invokes the idea of quantum mechanical tunneling, with a Gaussian suppression of the
transverse momenta and masses imparted to the produced quarks28 ,
" !# ∞
!
−π(m 2 + p2 )
q ⊥q
X 1 −nπ(m 2 + p2 )
q ⊥q
Prob(m2q , p2⊥q ) ∝ − ln 1 − exp = exp
κ n κ
n=1
! !
m2q κ/π −πm2q −πp2⊥q
→ exp exp , (82)
κ κ

where mq is the mass of the produced quark (the relevant range of values for the constituent
up, down, and strange quark masses is typically taken to be ∼ 300 – 500 MeV) and p⊥ is the
transverse momentum imparted to it by the breakup process (with the q̄ having the opposite
p⊥ ).
28
The full form of the series essentially accounts for the probability that no string-decay event already happened
and incorporates the exclusion principle that no more than one event can occur for each given quantum state of
the produced pair, while the limiting Gaussian form is the naı̈ve probability per phase-space volume; see [163].

— 54 —
P. Skands Introduction to QCD

This form of the suppression factor was originally derived in 1951 by Schwinger in QED,
for creation of electron-positron pairs in a strong constant electric field [164]29 . Its generali-
sation to QCD was first treated using WKB tunneling approximations [163, 166], and there is
also a more recent claim of an exact path-integral solution [167], in all cases finding the same
basic form as in the QED case30 . Since the Schwinger tunneling probability is derived for an
infinitely extended constant field, while QCD flux tubes in nature have a finite size, modifica-
tions to the above baseline picture may be considered. E.g., a flux tube with radius rs could
imply a cutoff for momenta p⊥max ∼ 2rs κ/π [163]. It is not straightforward to define the
relevant radius unambiguously, however, and we note that, within the effective description in
an MC program, tails to higher p⊥ values may be justified anyway due to unresolved perturba-
tive effects below the shower cutoff. A high-p⊥ cutoff is therefore typically not imposed in the
MC implementation. See [168] for a phenomenological study that explores some alternative
assumptions (e.g., a thermal suppression) in an event-generator context.
Due to the factorisation of the p⊥ and m dependence implied by equation (82), the p⊥ spectrum
of produced quarks in this model is independent of the quark flavour, with a universal average
value of

σ2 ≡ p2⊥q = p2x + p2y (83)

Z Z ∞ !
π ∞ −π(p2x + p2y )
= dpx dpy (p2x + p2y ) exp (84)
κ −∞ −∞ κ
Z ∞
π

= dp2⊥ p2⊥ exp −πp2⊥ /κ (85)
κ 0
= κ/π ∼ (240 MeV)2 . (86)

Bear in mind that “transverse” is here defined with respect to the string axis. Thus, the p⊥ in
a frame where the string is moving is modified by a Lorentz boost factor. Also bear in mind
that σ 2 is here a purely non-perturbative parameter. In a Monte Carlo model with a fixed
shower cutoff Qhad , the effective amount of “non-perturbative” p⊥ may be larger than this, due
to effects of additional unresolved soft-gluon radiation below Qhad (for example, the “Monash
Tune” of P YTHIA 8 [169] has a final-state shower cutoff at p⊥min = 500 MeV and uses a value
of σ = 335 MeV). In principle, the magnitude of this additional component should scale with
the cutoff, but in practice it is up to the user to enforce this by retuning (see section 5.4)
the effective σ parameter when changing the hadronisation scale. Since hadrons receive p⊥
contributions from two breakups, one on either side, their average transverse momentum
squared will be twice as large,
p2⊥h = 2σ 2 . (87)
Finally, we note that the assumption m2q  κ/π which is used to justify neglecting the sub-
leading terms equation (82) may be open to question and could lead to deviations of order
10% from the universal Gaussian spectrum, for realistic constituent masses.
The mass suppression implied by equation (82) is however less straightforward to inter-
pret. Integrated over p⊥ , the Schwinger prediction is that the differential string decay proba-
29
Note that extremely strong electric fields are required, so the effect has not been observed in the lab yet,
although attempts are being made at large laser facilities, see e.g. [165].
30
We note that the form in [167] also contains a term proportional to |dabc E a E b E c |2 with dabc the symmetric
structure constants of QCD and Ea the chromoelectric field with index a ∈ [1, 8].

— 55 —
P. Skands Introduction to QCD

bility (per unit space-time volume) via pair production of quarks with constituent mass mq , is
proportional to
∞
! !
X 1 nπm 2
q m2 κ/π
q πm 2
q
P (m2q ) ∝ exp − → exp − . (88)
n2 κ κ
n=1

However, since quark masses are notoriously difficult to define for light quarks, the value of the
strangeness suppression must effectively be extracted from experimental measurements, e.g.,
of the K/π ratio, with a resulting suppression of roughly s/u ∼ s/d ∼ 0.2 – 0.3. Inserting even
comparatively low values for the charm quark mass in equation (82), however, one obtains
a relative suppression of charm of the order of 10−11 . Heavy quarks can therefore safely be
considered to be produced only in the perturbative stages and not by the soft fragmentation.
Baryon production can be incorporated in the same basic picture [170], by allowing string
breaks to occur also by the production of pairs of so-called diquarks, loosely bound states of
two quarks in an overall 3̄ representation (e.g., “red + blue ∼ antigreen”, cf. the rules for
colour combinations in section 1.2). Again, the relative rate of diquark-to-quark production is
not known a priori and must be extracted from experimental measurements, e.g., of the p/π
ratio. More advanced scenarios for baryon production have also been proposed, in particular
the so-called popcorn model [171, 172], which is normally used in addition to the diquark
picture and then acts to decrease the correlations among neighboring baryon-antibaryon pairs
by allowing mesons to be formed inbetween them. Within the P YTHIA framework, a fragmen-
tation model including explicit string junctions [173] has so far only been applied to baryon-
number-violating new-physics processes and to the description of beam remnants (and then
acts to increase baryon stopping [174]).
This brings us to the next step of the algorithm: assignment of the produced quarks within
hadron multiplets. Using a nonrelativistic classification of spin states, the fragmenting q (q̄)
may combine with the q̄ 0 (q 0 ) from a newly created breakup to produce either a vector or a
pseudoscalar meson, or, if diquarks are involved, either a spin-1/2 or spin-3/2 baryon. Unfor-
tunately, the string model is entirely unpredictive in this respect, and this is therefore the sector
that contains the largest amount of free parameters. From spin counting alone, one would ex-
pect the ratio V /S of vectors to pseudoscalars to be 3, but this is modified by the V –S mass
splittings, which implies a phase-space suppression of vector production, with corresponding
suppression parameters to be extracted from data.
Especially for the light flavours, the substantial difference in phase space caused by the
V –S mass splittings implies a rather large suppression of vector production. Thus, for D∗ /D,
the effective ratio is already reduced to about ∼ 1.0 – 2.0, while for K ∗ /K and ρ/π, extracted
values range from 0.3 – 1.0. (Recall, as always, that these are production ratios of primary
hadrons, hence feed-down from secondary decays of heavier hadrons complicates the extrac-
tion of these parameters from experimental data, in particular for the lighter hadron species.)
The production of higher meson resonances is assumed to be low in a string framework31 .
For diquarks, separate parameters control the relative rates of spin-1 diquarks vs. spin-0 ones
and, likewise, have to extracted from data, with resulting values of order (qq)1 /(qq)0 ∼ 0.075 –
0.15.
31
The four L = 1 multiplets are implemented in P YTHIA, but are disabled by default, largely because several
states are poorly known and thus may result in a worse overall description when included.

— 56 —
P. Skands Introduction to QCD

The a parameter The b parameter

a = 0.9 a = 0.1 b = 0.5 b = 2.0
2.0
1.5
1.5
1.0
1.0
0.5 0.5

0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0
b = 1, m⊥ = 1 GeV a = 0.5, m⊥ = 1 GeV

Figure 31: Normalised Lund symmetric fragmentation function, for fixed m⊥ = 1 GeV. Left:
variation of the a parameter, from 0.1 (blue) to 0.9 (red), with fixed b = 1 GeV−2 . Right:
variation of the b parameter, from 0.5 (red) to 2 (blue) GeV−2 , with fixed a = 0.5.

With p2⊥ and m2 now fixed, the final step is to select the fraction, z, of the fragmenting
endpoint quark’s longitudinal momentum that is carried by the created hadron. In this re-
spect, the string picture is substantially more predictive than for the flavour selection. Firstly,
the requirement that the fragmentation be independent of the sequence in which breakups are
considered (causality) imposes a “left-right symmetry” on the possible form of the fragmenta-
tion function, f (z), with the solution [175]
 
1 a b (m2h + p2⊥h )
f (z) ∝ (1 − z) exp − , (89)
z z
which is known as the Lund symmetric fragmentation function (normalised to unit integral).
The a and b parameters, illustrated in figure 31, are the only free parameters of the fragmen-
tation function, though a may in principle be flavour-dependent. Note that the explicit mass
dependence in f (z) implies a harder fragmentation function for heavier hadrons (in the rest
frame of the string).
For massive endpoints (e.g., c and b quarks), which do not move along straight lightcone
sections, the exponential suppression with string area leads to modifications of the form [176],
2
f (z) → f (z)/z b mQ , (90)
with mQ the heavy-quark mass. Strictly speaking, this is the only fragmentation function that
is consistent with causality in the string model, though a few alternative forms are typically
provided as well.
As a by-product, the probability distribution in invariant time τ of q 0 q̄ 0 breakup vertices,
or equivalently Γ = (κτ )2 , is also obtained, with dP/dΓ ∝ Γa exp(−bΓ) implying an area law
for the colour flux [177], and the average breakup time lying along a hyperbola of constant
invariant time τ0 ∼ 10−23 s [8].
We may also ask, e.g., how many units of rapidity does the particle production from a
string span? Measuring pz along the string direction and defining rapidity by
 
1 E + pz
y = ln , (91)
2 E − pz

— 57 —
P. Skands Introduction to QCD

shower u(~p⊥0, p+)

QUV
··· QIR π + (~p⊥0 − p~⊥1, z1 p+)

dd¯

K 0(~p⊥1 − p~⊥2, z2(1 − z1 )p+)

ss̄

...

Figure 32: Iterative selection of flavours and momenta in the Lund string-fragmentation
model.

the absolute highest rapidity that can be reached, by a pion traveling exactly along the string
direction and taking all of the endpoint quark’s energy, is ymax = ln(2Eq /mπ ). I.e., the rapidity
region covered by a fragmenting string scales logarithmically with the energy, and since the
density of hadrons produced per unit rapidity is roughly constant (modulo endpoint effects),
the average number of hadrons produced by string fragmentation likewise scales logarithmi-
cally with energy.
The iterative selection of flavours, p⊥ , and z values is illustrated in figure 32. A parton
produced in a hard process at some high scale QUV emerges from the parton shower, at the
hadronisation scale QIR , with 3-momentum p~ = (~ p⊥0 , p+ ), where the “+” on the third com-
ponent denotes “light-cone” momentum, p± = E ± pz . Next, an adjacent dd¯ pair from the
vacuum is created, with relative transverse momenta ±p⊥1 . The fragmenting quark combines
with the d¯ from the breakup to form a π + , which carries off a fraction z1 of the total lightcone
momentum p+ . The next hadron carries off a fraction z2 of the remaining momentum, etc.

5.2 Soft Hadron-Hadron Processes

The total hadron-hadron (hh) cross section is around 100 mb at LHC energies [178], growing
slowly with the CM energy, σtot (s) ∝ s0.096 [179]. There are essentially four types of physics
processes, which together make up σtot :

1. Elastic scattering: hh → hh;

2. Single diffractive dissociation: hh → h + gap + X, with “gap” denoting an empty rapidity

region, and X anything that is not the original beam particle;

3. Double diffractive dissociation: hh → X + gap + X (both hadrons “blow up”);

4. Inelastic non-diffractive scattering: everything else.

In principle, higher “multi-gap” diffractive components may be defined as well, the most im-
portant one being central diffraction: hh → h + gap + X + gap + h, see the discussion of
diffraction in section 5.2.1 below.

— 58 —
P. Skands Introduction to QCD

Some important differences exist between theoretical and experimental terminology [180].
In the experimental setting, diffraction is defined by an observable rapidity gap, with |∆y|gap >
∼
3 typically giving clean diffractive samples. In the MC context, however, each diffractive pro-
cess type produces a whole spectrum of gaps, with small ones suppressed but not excluded.
Likewise, events of non-diffractive origin may produce accidental rapidity gaps, now with
large ones suppressed (exponentially) but not excluded, and in the transition region there
could even be quantum mechanical interference between the two. Due to this unphysical
model dependence of theoretical definitions of diffraction, we strongly advise to phrase mea-
surements first and foremost in terms of physical observables, and only seek to connect with
theory models as a second, separate, step.
The distinction between elastic and inelastic events is, however, unambiguous (modulo
pp → ppγ events); the final state either contains just two protons, or not. The total hadron-
hadron cross section can therefore be written as a sum of these two physically distinguishable
components,
σtot (s) = σel (s) + σinel (s) , (92)
where s = (pA + pB )2 is the beam-beam center-of-mass energy squared.
Another potentially confusing term is “minimum bias” (MB). This originates from the
experimental requirement of a minimal energy or number of hits in a given (experiment-
dependent) instrumented region near the beam, used to determine whether there was any
non-trivial activity in the event, or not. This represents the smallest possible “trigger bias”
that the corresponding experiment is capable of. There is thus no universal definition of “min-
bias”; each experiment has its own. We give a brief discussion of minimum bias in section 5.2.2
below.
Finally, in events containing a hard parton-parton interaction, the underlying event (UE)
can be roughly conceived of as the difference between QCD with and without including the
remnants of the original beam hadrons. Without such “beam remnants”, only the hard inter-
action itself, and its associated parton showers and hadronisation, would contribute to the ob-
served particle production. In reality, after the partons that participate in the hard interaction
have been taken out, the remnants still contain whatever is left of the incoming beam hadrons,
including also a partonic substructure, which leads to the possibility of multiple parton interac-
tions (MPI). We discuss MPI-based models of the UE in section 5.3 below. Other useful reviews
of MPI-based MC models can be found in [6, 152]. Analytical models are mostly formulated
only for double parton scattering, see e.g., [181–184].

5.2.1 Diffractive Scattering

As mentioned above, if the beam particles A and/or B are not elementary, the inelastic final
states may be divided into “diffractive” and “non-diffractive” topologies. This is a qualitative
classification, usually based on whether the final state looks like the decay of an excitation
of the beam particles (diffractive32 ), or not (non-diffractive), or upon the presence of a large
rapidity gap somewhere in the final state which would separate such excitations.
32
An example of a process that would be labeled as diffractive would be if one the protons is excited to a ∆+
which then decays back to p+ + π 0 , without anything else happening in the event. In general, a whole tower of
possible diffractive excitations are available, which in the continuum limit can be described by a mass spectrum
falling roughly as dM 2 /M 2 .

— 59 —
P. Skands Introduction to QCD

Given that an event has been labeled as diffractive, either within the context of a theoreti-
cal model, or by a final-state observable, we may distinguish between three different classes of
diffractive topologies, which it is possible to distinguish between physically, at least in princi-
ple. In double-diffractive (DD) events, both of the beam particles are diffractively excited and
hence none of them survive the collision intact. In single-diffractive (SD) events, only one of
the beam particles gets excited and the other survives intact. The last diffractive topology is
central diffraction (CD), in which both of the beam particles survive intact, leaving an excited
system in the central region between them. (This latter topology includes “central exclusive
production” where a single particle is produced in the central region.) That is,

σinel (s) = σSD (s) + σDD (s) + σCD (s) + σND (s) , (93)

where “ND” (non-diffractive, here understood not to include elastic scattering) contains no
gaps in the event consistent with the chosen definition of diffraction. Further, each of the
diffractively excited systems in the events labeled SD, DD, and CD, respectively, may in princi-
ple consist of several subsystems with gaps between them. Eq. (93) may thus be defined to be
exact, within a specific definition of diffraction, even in the presence of multi-gap events. Note,
however, that different theoretical models almost always use different (model-dependent) def-
initions of diffraction, and therefore the individual components in one model are in general
not directly comparable to those of another. It is therefore important that data be presented
at the level of physical observables if unambiguous conclusions are to be drawn from them.

5.2.2 Minimum Bias

In principle, everything that produces a hit in a given experiment’s minimum-bias trigger, is a

subset of minimum-bias (MB). In particular, since there is no explicit veto on hard activity, it
is useful to keep in mind that minimum bias includes a diverse mixture of both soft and hard
processes, though the fraction that is made up of hard high-p⊥ processes is only a small tail
compared to the total minimum-bias cross section33 .
In theoretical contexts, the term “minimum-bias” is often used with a slightly different
meaning; to denote specific (classes of) inclusive soft-QCD subprocesses in a given model.
Since these two usages are not exactly identical, in these lectures we have chosen to reserve
the term “minimum bias” to pertain strictly to definitions of experimental measurements, and
instead use the term “soft inclusive” physics as a generic descriptor for the class of processes
which generally dominate the various experimental “minimum-bias” measurements in theo-
retical models. This parallels the terminology used in the review [6], from which most of the
discussion here has been adapted. See equation (93) above for a compact overview of the
types of physical processes that contribute to minimum-bias data samples. For a more detailed
description of Monte Carlo models of this physics, see [6].
33
Furthermore, since only a tiny fraction of the total minimum-bias rate can normally be stored, the minimum-
bias sample would give quite poor statistics if used for hard physics studies. Instead, separate dedicated hard-
process triggers are typically included in addition to the minimum-bias one, in order to ensure maximal statistics
also for hard physics processes.

— 60 —
P. Skands Introduction to QCD

5.3 Underlying Event and Multiple Parton Interactions

In this subsection, we focus on the physics of multiple parton interactions (MPI) as a theoret-
ical basis for understanding both inelastic, non-diffractive processes (minimum-bias), as well
as the so-called underlying event (a.k.a. the jet pedestal effect). Keep in mind, however, that
especially at low multiplicities, and when rapidity gaps are present, the contributions from
diffractive processes should not be ignored.
Due to the simple fact that hadrons are composite, multi-parton interactions (several dis-
tinct parton-parton interactions in one and the same hadron-hadron collision) will always be
there — but how many, and how much additional energy and tracks do they deposit in a given
measurement region? The first detailed Monte Carlo model for perturbative MPI was pro-
posed by Sjöstrand in [9], and with some variation this still forms the basis for most modern
implementations [6].
The first crucial observation is that the t-channel propagators appearing in perturbative
QCD 2 → 2 scattering almost go on shell at low p⊥ , causing the differential cross sections to
become very large, behaving roughly as

dt dp2⊥
dσ2→2 ∝ ∼ . (94)
t2 p4⊥

At LHC energies, this parton-parton cross section becomes larger than the total hadron-hadron
cross section at p⊥ scales of order 4 − 5 GeV. This is illustrated in figure 33, in which we com-
pare the integrated QCD parton-parton cross section (with two different αs and PDF choices)
√
to the total inelastic cross section measured by TOTEM [178], for pp collisions at s = 8 TeV.
In the context of MPI models, this is interpreted straightforwardly to mean that each hadron-
hadron collision contains several few-GeV parton-parton collisions.
In the limit that all the partonic interactions are independent and equivalent, one would
simply have a Poisson distribution in the number of MPI, with average

σ2→2 (p⊥min )
hni(p⊥min ) = , (95)
σtot
with p⊥min a lower cutoff scale which we shall return to below, and σtot a measure of the
inelastic hadron-hadron cross section. This simple reinterpretation in fact expresses unitarity;
instead of the total interaction cross section diverging as p⊥min → 0 (which would violate
unitarity), we have restated the problem so that it is now the number of MPI per collision that
diverges, with the total cross section remaining finite.
Two important ingredients remain to fully regulate the remaining divergence. Firstly, the
interactions cannot use up more momentum than is available in the parent hadron. This
suppresses the large-n tail of the estimate above. In P YTHIA-based models, the MPI are or-
dered in p⊥ [9, 123], and the parton densities for each successive interaction are explicitly
constructed so that the sum of x fractions can never be greater than unity. In the H ERWIG
models [185, 186], instead the uncorrelated estimate of hni above is used as an initial guess,
but the generation of actual MPI is stopped once the energy-momentum conservation limit is
reached.
The second ingredient invoked to suppress the number of interactions, at low p⊥ and
x, is colour screening; if the wavelength ∼ 1/p⊥ of an exchanged coloured parton becomes

— 61 —
P. Skands Introduction to QCD

8 TeV pp
104

Integrated cross section [mb]

σ2→2(p ≥ p ) vs p
T Tmin Tmin
TOTEM σ INEL
3
10 α s=0.130 NNPDF2.3LO
α s=0.135 CTEQ6L1

102

10

VINCIAROOT
1

Pythia 8.183
10-1

1.5
Ratio

0.5

00 5 10 15 20
p
Tmin

Figure 33: Proton-proton collisions at 8 TeV. LO QCD parton-parton cross section (integrated
above pT min , for two different αs and PDF choices) compared to the total inelastic hadron-
hadron cross section. Towards the right of the plot, we see, as expected, that hard dijet events
is only a tiny fraction of the total cross section. The fact that the curves cross at a scale of
order 5 GeV is interpreted to mean that this is a characteristic scale relevant for MPI. [169].
0 5 10 15 20

larger than a typical colour-anticolour separation distance, it will only see an average colour
charge that vanishes in the limit p⊥ → 0, hence leading to suppressed interactions. This
provides an infrared cutoff for MPI similar to that provided by the hadronisation scale for
parton showers. A first estimate of the colour-screening cutoff would be the proton size,
p⊥min ≈ ~/rp ≈ 0.3 GeV ≈ ΛQCD , but empirically this appears to be far too low. In current
models, one replaces the proton radius rp in the above formula by a “typical colour screening
distance,” i.e., an average size of a region within which the net compensation of a given
colour charge occurs. This number is not known from first principles, though it may be related
to saturation [187]. In current MPI models, it is perceived of simply as an effective cutoff
parameter, to be determined from data.
Note that the partonic cross sections depend upon the PDF set used, and therefore the
optimal value to use for the cutoff will also depend on this choice [188]. Note also that the
cutoff does not have to be energy-independent. Higher energies imply that parton densities
can be probed at smaller x values, where the number of partons rapidly increases. Partons
then become closer packed and the colour-screening distance d decreases. The uncertainty
on the scaling of the cutoff is a major concern when extrapolating between different collider
energies [188–190].
We now turn to the origin of the observational fact that hard jets appear to sit on top of
a “pedestal” of underlying activity, which on average appears to be distributed evenly at all

— 62 —
P. Skands Introduction to QCD

7000 GeV pp Underlying Event

7000 GeV pp Underlying Event

〈d2 Nchg/dηdφ〉

Rivet 2.4.0, ≥ 400k events

〈d2 sum p /dηdφ〉 [GeV]

Rivet 2.4.0, ≥ 400k events

3 Average Charged Particle Density (TRNS) (|η| < 2.5, p > 0.1 GeV)
Sum(p ) Density (TRNS) (|η| < 2.5, p > 0.1 GeV/c) T
T T
ATLAS
ATLAS
2 Herwig++ (UE-EE-4)
Herwig++ (UE-EE-4)
Pythia 6 (370:P2012)
Pythia 6 (370:P2012) 2.5
T

Pythia 8 (Def)
Pythia 8 (Def)
Sherpa (Def)
Sherpa (Def)

1.5
2

1.5
1

mcplots.cern.ch
mcplots.cern.ch
0.5

ATLAS_2010_S8894728 ATLAS_2010_S8894728
0.5
Herwig++ 2.7.1, Pythia 6.428, Pythia 8.212, Sherpa 2.2.0 Herwig++ 2.7.1, Pythia 6.428, Pythia 8.212, Sherpa 2.2.0

0
2 ×2 2 ×2

+50% +50%

Ratio to ATLAS
Ratio to ATLAS

+20% +20%
+10% +10%

1 1
-10% -10%

-20% -20%

0.5 -50% 0.5 -50%

0 5 10 15 20 0 5 10 15 20
p (leading track) [GeV] p (leading track) [GeV]
T T

Figure 34: The rise of the Underlying Event, as a function of leading-track p⊥ , as measured in
the “Transverse Region” defined by azimuthal angles 60◦ < ∆φ < 120◦ from the leading track,
averaged over the available rapidity region. Left: average summed p⊥ (per unit ∆η∆φ) for
all tracks with p⊥ > 0.1 GeV. Right: average number of charged tracks (per unit ∆η∆φ) with
p⊥ > 0.1 GeV. Plots from mcplots.cern.ch [192].

rapidities (i.e., also far from the jet cores).

“Outside the [jet], a constant ET plateau is observed, whose height is independent

of the jet ET. Its value is substantially higher than the one observed for minimum
bias events.” (UA1 Collaboration (1983) [191])

That is, the so-called “underlying event” (UE) is substantially more active, with larger fluctua-
tions, than the average min-bias event. In the MPI context, this is interpreted as a consequence
of impact-parameter-dependence: in peripheral collisions, only a small fraction of events con-
tain any high-p⊥ activity, whereas central collisions are more likely to contain at least one
hard scattering; a high-p⊥ triggered sample will therefore be biased towards small impact
parameters, b, with a larger number of MPI (and associated increased activity).
The rise of the pedestal level from low to high trigger-object34 p⊥ is illustrated in figure 34.
The leveling off of the distributions above leading-track p⊥ values ∼ 5 GeV can be interpreted
as an effect of “maximum bias”; when the trigger p⊥ is high enough, the selected events are
34
Different measurements and detectors employ different types of trigger objects, such as the leading track,
leading track-jet, leading calorimeter jet, etc. In principle, the more inclusive (IR safe) the trigger is defined, the
cleaner the measurement of the pedestal rise will be.

— 63 —
P. Skands Introduction to QCD

essentially already maximally biased to small impact parameters, and from that point on-
wards the UE essentially becomes “independent of the jet ET” [191] (modulo spillover of
bremsstrahlung radiation from the hard-jet regions, which will still increase with the trigger-
jet p⊥ ; this component can be partially suppressed e.g. by vetoing events with bremsstrahlung;
by considering only the least active of the two UE regions; or by systematically classifying all
of the activity in the event as either “jet-like” or “plateu-like” [99]). The ability of a theory
model to describe the rise of the pedestal from the min-bias level to the UE plateau therefore
depends upon its modeling of the b-dependence, and correspondingly the impact-parameter
distribution (or, equivalently, the assumed mass distributions of the proton in b-space) consti-
tutes another main tuning parameter. A detailed discussion of impact-parameter dependent
models goes beyond the scope of these lectures; see [9, 174, 193].
For hard processes at the LHC atp13 TeV, the transverse energy, ET , in the UE is expected
to be about 3.3 GeV per unit ∆R = ∆φ2 + ∆η 2 area [190] (for a reference case of 100-GeV
dijets, including both charged and neutral particles, with no cut on p⊥ ), though with large
event-to-event fluctuations of order ±2 GeV [194]. Thus, for example, the total ET originating
from the UE, in a cone with radius 0.4 can be estimated to be ET UE (R = 0.4) ∼ 1.6 ± 1 GeV,
while the ET in a cone with radius 1.0 would be ET UE (R = 1.0) ∼ 10 ± 6 GeV. Note that the
magnetic field in realistic detectors will deflect some fraction of the soft charged component
of the underlying event into helix trajectories that will hence not contribute to the energy
deposition in the calorimeters.

5.4 Tuning
A main virtue of general-purpose Monte Carlo event generators is their ability to provide a
complete and fully differential picture of collider final states, down to the level of individual
particles. As has been emphasised in these lectures, the achievable accuracy depends both
on the inclusiveness of the chosen observable and on the sophistication of the simulation it-
self. An important driver for the latter is obviously the development of improved theoretical
models, e.g., by including matching to higher-order matrix elements, more accurate resum-
mations, or better non-perturbative models, as discussed in the previous sections; but it also
depends crucially on the available constraints on the remaining free parameters of the model.
Using existing data (or more precise calculations) to constrain these is referred to as generator
tuning.
Keep in mind that generators attempt to deliver a global description of the data; a tune
is no good if it fits one distribution perfectly, but not any others. It is therefore crucial to
study the simultaneous degree of agreement or disagreement over many, mutually comple-
mentary, distributions. A useful online resource for making such comparisons can be found at
the MCPLOTS web site [192] (which relies on computing power donated by volunteers, via
the LHC@home project [195]). The analyses come from the comprehensive R IVET analysis
toolkit [107], which can also be run stand-alone to make your own MC tests and comparisons.
Although MC models may appear to have a bewildering number of independently ad-
justable parameters, it is worth noting that most of these only control relatively small (exclu-
sive) details of the event generation. The majority of the (inclusive) physics is determined
by only a few, very important ones, such as the value of the strong coupling, in the per-
turbative domain, and the form of the fragmentation function for massless partons, in the
non-perturbative one.

— 64 —
P. Skands Introduction to QCD

Armed with a good understanding of the underlying model, an expert would therefore
normally take a highly factorised approach to constraining the parameters, first constraining
the perturbative ones (using IR safe observables and/or more precise theory calculations) and
thereafter the non-perturbative ones, each ordered in a measure of their relative significance
to the overall modeling. This allows one to concentrate on just a few parameters and a few
carefully chosen distributions at a time, reducing the full parameter space to manageable-
sized chunks. Still, each step will often involve more than one single parameter, and non-
factorisable correlations may still necessitate additional iterations from the beginning before a
fully satisfactory set of parameters is obtained.
Recent years have seen the emergence of automated tools that attempt to reduce the
amount of both computer and manpower required for this task, for instance by making full
generator runs only for a limited set of parameter points, and then interpolating between
these to obtain approximations to what the true generator result would have been for any
intermediate parameter point, as, e.g., in P ROFESSOR [196]. Automating the human expert
input is more difficult. Currently, this is addressed by a combination of input solicited from
the generator authors (e.g., which parameters and ranges to consider, which observables con-
stitute a complete set, etc) and the elaborate construction of non-trivial weighting functions
that determine how much weight is assigned to each individual bin in each distribution. The
field is still burgeoning, and future sophistications are to be expected. Nevertheless, at this
point the overall quality of the tunes obtained with automated methods appear to at least be
competitive with the manual ones.
However, though we have very good LHC tunes for essentially all the general-purpose gen-
erators by now, there are two important aspects which have so far been neglected, and which
it is becoming increasingly urgent to address. The first is that a central tune is not really worth
much, unless you know what the uncertainty on it is. A few individual proposals for system-
atic tuning variations have been made [189, 197], but so far there is no general approach for
establishing MC uncertainties by tune variations. (Note: in 2016 all of the general-purpose
MC collaborations published strategies for automated evaluations of perturbative shower un-
certainties, which we highly recommend the community to start using, see [15, 198, 199].)
The second issue is that virtually all generator tuning is done at the “pure” LL shower level,
and not much is known about what happens to the tuning when matrix-element matching is
subsequently included35 .
Finally, rather than performing one global tune to all the data, as is usually done, a more
systematic check on the validity of the underlying physics model could be obtained by instead
performing several independent optimisations of the model parameters for a range of different
phase-space windows and/or collider environments. In regions in which consistent parameter
sets are obtained (with reasonable ∆χ2 values), the underlying model can be considered as
interpolating well, i.e., it is universal. If not, a breakdown in the ability of the model to span
different physical regimes has been identified, and can be addressed, with the nature of the
deviations giving clues as to the nature of the breakdown. With the advent of automated tools,
such systematic studies are now becoming feasible, with a first example given in [188].
We round off by giving a sketch of a reasonably complete tuning procedure, without going
into details about the parameters that control each of these sectors in individual Monte Carlo
models (a recent detailed discussion in the context of PYTHIA 8 can be found in [169]):
35
One aspect, consistent αs variations, is discussed in [142].

— 65 —
P. Skands Introduction to QCD

1) Keep in mind that inabilities of models to describe data is a vital part of the feedback cy-
cle between theory and experiment. Also keep in mind that perturbation theory at (N)LO+LL
is doing very well if it gets within 10% of a given IR safe measurement. An agreement of 5%
should be considered the absolute sanity limit, beyond which it does not make any sense to
tune further. For some quantities, e.g., ones for which the underlying modeling is known to be
poor, an order-of-magnitude agreement or worse may have to be accepted.
2) Final-State Radiation and Hadronisation: mainly using LEP and other e+ e− collider
data. On the IR safe side, there are event shapes and jet observables. On the IR sensitive
side, multiplicities and particle spectra. Pay attention to the high-z tail of the fragmentation,
where a single hadron carries a large fraction of an entire jet’s momentum (most likely to give
“fakes”). Depending on the focus of the tuning, attention should also be paid to identified-
particle rates and ratios (perhaps with a nod to hadron-collider measurements), and to frag-
mentation in events containing heavy quarks and/or gluon jets. Usually, more weight is given
to those particles that are most copiously produced. The scaling properties of IR safe vs.
IR sensitive contributions can be tested by comparing data at several different e+ e− collider
energies.
3) Initial-State Radiation, and “Primordial36 kT ”: the main constraining distribution
is the dilepton p⊥ distribution in Drell-Yan events in hadron-hadron collisions. Ideally, one
would like to use several different Q2 values, and/or complementary processes, like p⊥ (dijet)
or p⊥ (tt̄). For any observables containing explicit jets, be aware that the UE can produce
small horisontal shifts in jet p⊥ distributions, which may in turn result in larger-than-expected
vertical changes if the distributions are falling sharply. Also note that the ISR evolution is
sensitive to the choice of PDFs.
4) Initial-Final Connections: (radiation from colour lines connecting the initial and final
states): DIS and jet broadening in hadron collisions. This is one of the most poorly controlled
parts of most MC models, highly sensitive to the treatment of coherence (see e.g. [200] for
illustrations). Keep in mind that it is not directly constrained by pure final-state observables,
such as LEP fragmentation, nor by pure initial-state observables, such as the Drell-Yan p⊥
spectrum, which is why we list it as a separate item here. The modeling of this aspect can have
important effects on specific observables, a recent example being the tt̄ forward-backward
asymmetry at the Tevatron [201].
5) Underlying Event: Good constraints on the overall level of the underlying event can
be obtained by counting the summed transverse energy (more IR safe) and/or particle multi-
plicities and average transverse momenta (more IR sensitive) in regions transverse to a hard
trigger jet (more IR safe) or particle (more IR sensitive), see e.g. [202]. Constraints on the
fluctuations of the underlying event are also important, and can be obtained, e.g., by compar-
ing to measurements of the RMS of such distributions [194] or by plotting salient quantities
along an axis from low to high UE activity [203]. Again, note that the UE modelling can be
very sensitive to the choice of PDFs [169, 188]. Finally, the modeling of the UE should be
universal, hence the same UE model should — ideally — be able to describe UE distributions
not only in inclusive-jet events, but also the UE in processes like Drell-Yan [204–207] or tt̄
production [208, 209].
6) Colour (Re-)Connections and other Final-State Interactions: By Final-State Inter-
36
Primordial kT : an additional soft p⊥ component that is injected on top of the p⊥ generated by the initial-state
shower itself, see [6, Section 7.1].

— 66 —
P. Skands Introduction to QCD

actions, we intend a broad spectrum of possible collective effects that may be included to a
greater or lesser extent in various models. These effects include: Bose-Einstein correlations
(see, e.g., [210]), rescattering (see, e.g., [211]), colour reconnections / string interactions
(see, e.g., [212–216]), hydrodynamics (see, e.g., [217]), etc. As a rule, these effects are
soft and/or non-perturbative and hence should not modify hard IR safe observables appre-
ciably. They can, however, have drastic effects on IR sensitive ones, such as particle multi-
plicities, momentum distributions, and correlations, wherefore useful constraints are typically
furnished by measurements of spectra and correlations as functions of quantities believed to
serve as indicators of the strength of these phenomena (such as event multiplicity [218–220]
or underlying-event activity [203]), and/or by collective-flow-type measurements. Finally,
if the model includes a universal description of underlying-event and soft-inclusive QCD, as
many MPI-based models do, then minimum-bias data can also be used as a control sample,
though one must then be careful either to address diffractive contributions properly or to in-
clude only gap-suppressed data samples. A complete MB and UE model should also be able
to describe the rise of the pedestal from MB to UE, e.g., in the transverse UE observables (see
above).
7) Beam Remnants: Constraints on beam-remnant fragmentation (see, e.g., [174]) are
most directly obtained in the forward region [221–227]. Especially for soft-inclusive triggers,
cuts designed to isolate diffractive topologies are then required to distinguish effectively be-
tween the fragmentation of a (diffractive) colour-singlet excitation of the beam particle and
that of a (non-diffractive) colour-charged remnant having exchanged some (non-zero) net
amount of colour charge with the other colliding beam hadron. To lowest order, the remnant
should be in a triplet (octet) colour state after having exchanged a quark (gluon) with the other
beam particle, but taking MPI into account much larger amounts of colour may be transferred,
with correspondingly larger possible colour representations of the remnant (see, e.g., [216]).
The amount of baryon transport from the remnant to a given rapidity region [228–230] can
also be used to probe how much the colour structure of the remnant was effectively disturbed,
with more baryon transport indicating a larger amount of “beam baryon blowup”. Ideally one
would also correlate the forward activity with a (central) measure of how much total colour
charge is scattered out of the remnant(s), such as an observable sensitive to the number of
MPI.

Acknowledgments
Thanks to C. Brust, M. Cacciari, Y. Dokshitzer, L. Hartgring, A. Kronfeld, A. Larkoski, J. Lopez-
Villarejo, C. Murphy, P. Nason, P. Pigard, S. Prestel, G. Salam, and T. Sjöstrand whose valuable
comments and sharing of insight contributed to these lectures. In addition, I have used ma-
terial from my 2010 ESHEP lectures [152] and 2014 AEPSHEP lectures, and from the ESHEP
lectures by Mangano [231], by Salam [5, 232], by Sjöstrand [50], and by Stirling [233], as
well as the recent review on Monte Carlo event generators by the MCnet collaboration [6].

— 67 —
P. Skands Introduction to QCD

References
[1] M. E. Peskin and D. V. Schroeder, An Introduction to quantum field theory.
Addison-Wesley, 1995.

[2] R. D. Field, “Applications of Perturbative QCD,” Front. Phys. 77 (1989) 1–366.

[3] R. Ellis, W. Stirling, and B. Webber, QCD and collider physics. Camb. Monogr. Part.
Phys. Nucl. Phys. Cosmol., 1996.

[4] G. Dissertori, I. Knowles, and M. Schmelling, Quantum Chromodynamics — High

energy experiments and theory. Oxford Science Publications, 2003.

[5] G. P. Salam, “Elements of QCD for hadron colliders,” 1011.5131. In Proceedings of the
17th European school on high-energy physics (ESHEP), Bautzen, Germany, 14 - 27
June, 2009, arXiv:1012.4643.

[6] MCnet Collaboration, A. Buckley, J. Butterworth, S. Gieseke, D. Grellscheid, S. Höche,

et al., “General-purpose event generators for LHC physics,” Phys.Rept. 504 (2011)
145–233, 1101.2599.

[7] B. Andersson, G. Gustafson, G. Ingelman, and T. Sjöstrand, “Parton Fragmentation and

String Dynamics,” Phys. Rept. 97 (1983) 31–145.

[8] B. Andersson, The Lund model. Camb. Monogr. Part. Phys. Nucl. Phys. Cosmol., 1997.

[9] T. Sjöstrand and M. van Zijl, “A multiple-interaction model for the event structure in
hadron collisions,” Phys. Rev. D36 (1987) 2019.

[10] T. Sjöstrand, “The Development of MPI Modelling in PYTHIA,” 1706.02166.

[11] F. James, “Monte Carlo theory and practice,” Rept. Prog. Phys. 43 (1980) 1145.

[12] S. Weinzierl, “Introduction to Monte Carlo methods,” hep-ph/0006269.

[13] S. Plätzer and M. Sjödahl, “The Sudakov Veto Algorithm Reloaded,” Eur. Phys. J. Plus
127 (2012) 26, 1108.6180.

[14] L. Lönnblad, “Fooling Around with the Sudakov Veto Algorithm,” Eur. Phys. J. C73
(2013), no. 3, 2350, 1211.7204.

[15] S. Mrenna and P. Skands, “Automated Parton-Shower Variations in Pythia 8,”

1605.08352.

[16] K. A. Brueckner, “Meson-Nucleon Scattering and Nucleon Isobars,” Phys.Rev. 86

(1952) 106–109.

[17] C. Chedester, P. Isaacs, A. Sachs, and J. Steinberger, “Total cross sections of pi-mesons
on protons and several other nuclei,” Phys.Rev. 82 (1951) 958–959.

[18] E. Fermi, H. Anderson, A. Lundby, D. Nagle, and G. Yodh, “Ordinary and Exchange
Scattering of Negative Pions by Hydrogen,” Phys.Rev. 85 (1952) 935–936.

— 68 —
P. Skands Introduction to QCD

[19] H. Anderson, E. Fermi, E. Long, and D. Nagle, “Total Cross-sections of Positive Pions in
Hydrogen,” Phys.Rev. 85 (1952) 936.

[20] H. Anderson, E. Fermi, D. Nagle, and G. Yodh, “Deuterium Total Cross Sections for
Positive and Negative Pions,” Phys.Rev. 86 (1952) 413–413.

[21] D. E. Nagle, The Delta: The First Pion Nucleon Resonance: Its Discovery and Applications.
1984. LALP–84-27, DE84 017107, based on a lecture given at the University of
Chicago Symposium in honor of H.L. Anderson, May 11, 1982.

[22] O. Greenberg, “Spin and Unitary Spin Independence in a Paraquark Model of Baryons
and Mesons,” Phys.Rev.Lett. 13 (1964) 598–602.

[23] M. Han and Y. Nambu, “Three Triplet Model with Double SU(3) Symmetry,” Phys.Rev.
139 (1965) B1006–B1010.

[24] S. Drell and T.-M. Yan, “Massive Lepton Pair Production in Hadron-Hadron Collisions
at High-Energies,” Phys.Rev.Lett. 25 (1970) 316–320.

[25] T. Plehn, D. Rainwater, and P. Z. Skands, “Squark and gluino production with jets,”
Phys.Lett. B645 (2007) 217–221, hep-ph/0510144.

[26] J. Alwall, S. de Visscher, and F. Maltoni, “QCD radiation in the production of heavy
colored particles at the LHC,” JHEP 0902 (2009) 017, 0810.5350.

[27] A. Papaefstathiou and B. Webber, “Effects of QCD radiation on inclusive variables for
determining the scale of new physics at hadron colliders,” JHEP 06 (2009) 069,
0903.2013.

[28] D. Krohn, L. Randall, and L.-T. Wang, “On the Feasibility and Utility of ISR Tagging,”
1101.0810.

[29] C. K. Vermilion, “Jet Substructure at the Large Hadron Collider: Harder, Better, Faster,
Stronger,” 1101.1335.

[30] A. Altheimer, S. Arora, L. Asquith, G. Brooijmans, J. Butterworth, et al., “Jet

Substructure at the Tevatron and LHC: New results, new tools, new benchmarks,”
J.Phys.G G39 (2012) 063001, 1201.0008.

[31] Particle Data Group Collaboration, J. Beringer et al., “Review of particle physics,”
Phys. Rev. D86 (2012) 010001.

[32] S. P. Martin, “A Supersymmetry primer,” hep-ph/9709356.

[33] P. Langacker, The standard model and beyond. 2010.

http:www.sns.ias.edu/~pgl/SMB.

[34] S. Bethke, “World Summary of αs (2012),” Nucl.Phys.Proc.Suppl. 234 (2013) 229–234,

1210.0325.

[35] D. d’Enterria and P. Z. Skands, eds., Proceedings, High-Precision αs Measurements from

LHC to FCC-ee, CERN. CERN, Geneva, 2015. arXiv:1512.05194.

— 69 —
P. Skands Introduction to QCD

[36] S. Bethke, “The 2009 World Average of alpha(s),” Eur.Phys.J. C64 (2009) 689–703,
0908.1135.

[37] G. Dissertori, A. Gehrmann-De Ridder, T. Gehrmann, E. Glover, G. Heinrich, et al.,

“Determination of the strong coupling constant using matched NNLO+NLLA
predictions for hadronic event shapes in e+e- annihilations,” JHEP 0908 (2009) 036,
0906.3436.

[38] A. Young, “Quantitative substitutional analysis I,” Proc. London Math. Soc. 33 (1901)
97.

[39] B. Sagan, “The ubiquitous young tableaux.” in Invariant Theory and Tableaux,
ed. D. Stanton, Springer Verlag, 1990, p.262.
http://www.mth.msu.edu/users/sagan/Papers/Old/uyt.pdf.

[40] J. Alwall, M. Herquet, F. Maltoni, O. Mattelaer, and T. Stelzer, “MadGraph 5 : Going

Beyond,” JHEP 1106 (2011) 128, 1106.0522.

[41] A. Pukhov, “Calchep 2.3: MSSM, structure functions, event generation, 1, and
generation of matrix elements for other packages,” hep-ph/0412191.

[42] CompHEP Collaboration, E. Boos et al., “CompHEP 4.4: Automatic computations from
Lagrangians to events,” Nucl.Instrum.Meth. A534 (2004) 250–259, hep-ph/0403113.

[43] A. Kanaki and C. G. Papadopoulos, “HELAC: A Package to compute electroweak

helicity amplitudes,” Comput.Phys.Commun. 132 (2000) 306–315, hep-ph/0002082.

[44] F. Krauss, R. Kuhn, and G. Soff, “AMEGIC++ 1.0: A Matrix element generator in
C++,” JHEP 0202 (2002) 044, hep-ph/0109036.

[45] M. Moretti, T. Ohl, and J. Reuter, “O’Mega: An Optimizing matrix element generator,”
hep-ph/0102195.

[46] W. Kilian, T. Ohl, and J. Reuter, “WHIZARD: Simulating Multi-Particle Processes at

LHC and ILC,” Eur.Phys.J. C71 (2011) 1742, 0708.4233.

[47] A. Cafarella, C. G. Papadopoulos, and M. Worek, “Helac-Phegas: A Generator for all

parton level processes,” Comput.Phys.Commun. 180 (2009) 1941–1955, 0710.2427.

[48] M. Bähr, S. Gieseke, M. Gigg, D. Grellscheid, K. Hamilton, et al., “Herwig++ Physics

and Manual,” Eur.Phys.J. C58 (2008) 639–707, 0803.0883.

[49] T. Gleisberg and S. Hoeche, “Comix, a new matrix element generator,” JHEP 0812
(2008) 039, 0808.3674.

[50] T. Sjöstrand, “Monte Carlo Generators,” hep-ph/0611247.

[51] J. C. Collins and D. E. Soper, “Parton Distribution and Decay Functions,” Nucl.Phys.
B194 (1982) 445.

[52] G. F. Sterman, “Partons, factorization and resummation, TASI 95,” hep-ph/9606312.

— 70 —
P. Skands Introduction to QCD

[53] CTEQ Collaboration, R. Brock et al., “Handbook of perturbative QCD: Version 1.1,”
Rev.Mod.Phys. 67 (1995) 157–248.

[54] T. Plehn, “LHC Phenomenology for Physics Hunters,” 0810.2281.

[55] J. C. Collins, D. E. Soper, and G. F. Sterman, “Transverse Momentum Distribution in

Drell-Yan Pair and W and Z Boson Production,” Nucl.Phys. B250 (1985) 199.

[56] G. Altarelli and G. Parisi, “Asymptotic freedom in parton language,” Nucl. Phys. B126
(1977) 298–318.

[57] V. N. Gribov and L. N. Lipatov, “Deep inelastic e-p scattering in perturbation theory,”
Sov. J. Nucl. Phys. 15 (1972) 438.

[58] Y. L. Dokshitzer, “Calculation of the structure functions for deep inelastic scattering
and e+ e− annihilation by perturbation theory in quantum chromodynamics,” Sov.
Phys. JETP 46 (1977) 641–653.

[59] NuTeV Collaboration, D. Mason et al., “Measurement of the Nucleon

Strange-Antistrange Asymmetry at Next-to-Leading Order in QCD from NuTeV Dimuon
Data,” Phys.Rev.Lett. 99 (2007) 192001.

[60] A. Cooper-Sarkar, “What did HERA teach us about the structure of the proton?,”
1206.0894.

[61] S. Alekhin, K. Melnikov, and F. Petriello, “Fixed target Drell-Yan data and NNLO QCD
fits of parton distribution functions,” Phys.Rev. D74 (2006) 054033, hep-ph/0606237.

[62] E. de Oliveira, A. Martin, and M. Ryskin, “Drell-Yan as a probe of small x partons at the
LHC,” 1205.6108.

[63] S. Alekhin, S. Alioli, R. D. Ball, V. Bertone, J. Blumlein, et al., “The PDF4LHC Working
Group Interim Report,” 1101.0536.

[64] A. Buckley and M. Whalley, “HepData reloaded: reinventing the HEP data archive,”
1006.0517. http://hepdata.cedar.ac.uk.

[65] A. Martin, W. Stirling, R. Thorne, and G. Watt, “Parton distributions for the LHC,”
Eur.Phys.J. C63 (2009) 189–285, 0901.0002.

[66] J. Pumplin, D. Stump, J. Huston, H. Lai, P. M. Nadolsky, et al., “New generation of

parton distributions with uncertainties from global QCD analysis,” JHEP 0207 (2002)
012, hep-ph/0201195.

[67] G. Watt and R. Thorne, “Study of Monte Carlo approach to experimental uncertainty
propagation with MSTW 2008 PDFs,” JHEP 1208 (2012) 052, 1205.4024.

[68] M. Bengtsson, T. Sjöstrand, and M. van Zijl, “Initial State Radiation Effects on W and
Jet Production,” Z. Phys. C32 (1986) 67.

[69] S. Gieseke, “Uncertainties of Sudakov form-factors,” JHEP 0501 (2005) 058,

hep-ph/0412342.

— 71 —
P. Skands Introduction to QCD

[70] NLO Multileg Working Group Collaboration, Z. Bern et al., “The NLO multileg
working group: Summary report,” 0803.0494.

[71] T. Kinoshita, “Mass singularities of Feynman amplitudes,” J. Math. Phys. 3 (1962)

650–677.

[72] T. D. Lee and M. Nauenberg, “Degenerate Systems and Mass Singularities,” Phys. Rev.
133 (1964) B1549–B1562.

[73] S. Catani and M. H. Seymour, “A general algorithm for calculating jet cross sections in
NLO QCD,” Nucl. Phys. B485 (1997) 291–419, hep-ph/9605323.

[74] S. Catani and M. H. Seymour, “The Dipole Formalism for the Calculation of QCD Jet
Cross Sections at Next-to-Leading Order,” Phys. Lett. B378 (1996) 287–301,
hep-ph/9602277.

[75] Z. Nagy, “Next-to-leading order calculation of three jet observables in hadron hadron
collision,” Phys.Rev. D68 (2003) 094002, hep-ph/0307268.

[76] R. Frederix, T. Gehrmann, and N. Greiner, “Automation of the Dipole Subtraction

Method in MadGraph/MadEvent,” JHEP 09 (2008) 122, 0808.2128.

[77] Y. I. Azimov, Y. L. Dokshitzer, V. A. Khoze, and S. I. Troian, “The String Effect and QCD
Coherence,” Phys. Lett. B 165 (1985) 147–150.

[78] D. A. Kosower, “Antenna factorization of gauge-theory amplitudes,” Phys. Rev. D57

(1998) 5410–5416, hep-ph/9710213.

[79] D. A. Kosower, “Antenna factorization in strongly-ordered limits,” Phys. Rev. D71

(2005) 045016, hep-ph/0311272.

[80] A. Gehrmann-De Ridder, T. Gehrmann, and E. W. N. Glover, “Antenna subtraction at

NNLO,” JHEP 09 (2005) 056, hep-ph/0505111.

[81] A. Gehrmann-De Ridder, T. Gehrmann, E. Glover, and G. Heinrich, “NNLO corrections

to event shapes in e+ e- annihilation,” JHEP 0712 (2007) 094, 0711.4711.

[82] S. Weinzierl, “NNLO corrections to 3-jet observables in electron-positron annihilation,”

Phys.Rev.Lett. 101 (2008) 162001, 0807.3241.

[83] G. Heinrich, “Sector Decomposition,” Int.J.Mod.Phys. A23 (2008) 1457–1486,

0803.4177.

[84] R. Boughezal, K. Melnikov, and F. Petriello, “A subtraction scheme for NNLO

computations,” Phys.Rev. D85 (2012) 034025, 1111.7041.

[85] S. Catani and M. Grazzini, “An NNLO subtraction formalism in hadron collisions and
its application to Higgs boson production at the LHC,” Phys.Rev.Lett. 98 (2007)
222002, hep-ph/0703012.

[86] G. Gustafson and U. Pettersson, “Dipole formulation of QCD cascades,” Nucl. Phys.
B306 (1988) 746.

— 72 —
P. Skands Introduction to QCD

[87] A. Gehrmann-De Ridder and M. Ritzmann, “NLO Antenna Subtraction with Massive
Fermions,” JHEP 0907 (2009) 041, 0904.3297.

[88] A. Gehrmann-De Ridder, M. Ritzmann, and P. Skands, “Timelike Dipole-Antenna

Showers with Massive Fermions,” Phys.Rev. D85 (2012) 014013, 1108.6172.

[89] A. J. Larkoski and M. E. Peskin, “Spin-Dependent Antenna Splitting Functions,”

Phys.Rev. D81 (2010) 054010, 0908.2450.

[90] W. Giele, D. Kosower, and P. Skands, “Higher-Order Corrections to Timelike Jets,”

Phys.Rev. D84 (2011) 054003, 1102.2126.

[91] J. Lopez-Villarejo and P. Skands, “Efficient Matrix-Element Matching with Sector

Showers,” JHEP 1111 (2011) 150, 1109.3608.

[92] G. F. Sterman, “Resummation, power corrections and prediction in perturbative QCD,”

in QCD and high energy hadronic interactions. Proceedings, 33rd Rencontres de Moriond,
Les Arcs, France, March 21-28, 1998, pp. 511–518. 1998. hep-ph/9806533.

[93] A. Banfi, G. P. Salam, and G. Zanderighi, “Phenomenology of event shapes at hadron

colliders,” JHEP 1006 (2010) 038, 1001.4082.

[94] CMS Collaboration, V. Khachatryan

p et al., “First Measurement of Hadronic Event
Shapes in pp Collisions at (s) = 7 TeV,” Phys.Lett. B699 (2011) 48–67, 1102.0068.

[95] ATLAS Collaboration, G. Aad et al., “Measurement of event shapes at large momentum
√
transfer with the ATLAS detector in pp collisions at s = 7 TeV,” Eur.Phys.J. C72
(2012) 2211, 1206.2135.

[96] K. Wraight and P. Skands, “Forward-Backward Correlations and Event Shapes as

probes of Minimum-Bias Event Properties,” Eur.Phys.J. C71 (2011) 1628, 1101.5215.

[97] M. Cacciari, G. P. Salam, and G. Soyez, “The Anti-k(t) jet clustering algorithm,” JHEP
0804 (2008) 063, 0802.1189.

[98] A. Abdesselam, E. B. Kuutmann, U. Bitenc, G. Brooijmans, J. Butterworth, et al.,

“Boosted objects: A Probe of beyond the Standard Model physics,” Eur.Phys.J. C71
(2011) 1661, 1012.5412.

[99] M. Cacciari, G. P. Salam, and S. Sapeta, “On the characterisation of the underlying
event,” JHEP 04 (2010) 065, 0912.4926.

[100] G. Corcella et al., “HERWIG 6: an event generator for hadron emission reactions with
interfering gluons (including supersymmetric processes),” JHEP 01 (2001) 010,
hep-ph/0011363.

[101] J. Bellm et al., “Herwig 7.0/Herwig++ 3.0 release note,” Eur. Phys. J. C76 (2016),
no. 4, 196, 1512.01178.

[102] T. Sjöstrand, S. Mrenna, and P. Skands, “PYTHIA 6.4 physics and manual,” JHEP 05
(2006) 026, hep-ph/0603175.

— 73 —
P. Skands Introduction to QCD

[103] T. Sjöstrand, S. Ask, J. R. Christiansen, R. Corke, N. Desai, et al., “An Introduction to

PYTHIA 8.2,” 1410.3012.

[104] T. Gleisberg, S. Höche, F. Krauss, M. Schönherr, S. Schumann, F. Siegert, and

J. Winter, “Event generation with Sherpa 1.1,” JHEP 02 (2009) 007, 0811.4622.

[105] E. Boos et al., “Generic user process interface for event generators,” hep-ph/0109068.

[106] J. Alwall et al., “A standard format for Les Houches Event Files,” Comput. Phys.
Commun. 176 (2007) 300–304, hep-ph/0609017.

[107] A. Buckley et al., “Rivet user manual,” 1003.0694.

[108] GEANT4 Collaboration, S. Agostinelli et al., “GEANT4: A Simulation toolkit,”

Nucl.Instrum.Meth. A506 (2003) 250.

[109] M. Dobbs and J. B. Hansen, “The HepMC C++ Monte Carlo event record for High
Energy Physics,” Comput. Phys. Commun. 134 (2001) 41–46.

[110] T. Sjöstrand, “Monte Carlo Tools,” 0911.5286.

[111] R. Kleiss, W. Stirling, and S. Ellis, “A NEW MONTE CARLO TREATMENT OF

MULTIPARTICLE PHASE SPACE AT HIGH-ENERGIES,” Comput.Phys.Commun. 40
(1986) 359.

[112] G. Lepage, “A New Algorithm for Adaptive Multidimensional Integration,”

J.Comput.Phys. 27 (1978) 192. Revised version.

[113] G. P. Lepage, “VEGAS - An Adaptive Multi-dimensional Integration Program,”.

CLNS-80/447.

[114] P. D. Draggiotis, A. van Hameren, and R. Kleiss, “SARGE: An Algorithm for generating
QCD antennas,” Phys.Lett. B483 (2000) 124–130, hep-ph/0004047.

[115] N. Fischer, S. Prestel, M. Ritzmann, and P. Skands, “Vincia for Hadron Colliders,”
1605.06142.

[116] G. Marchesini and B. Webber, “Simulation of QCD Jets Including Soft Gluon
Interference,” Nucl.Phys. B238 (1984) 1.

[117] G. Marchesini and B. R. Webber, “Monte Carlo Simulation of General Hard Processes
with Coherent QCD Radiation,” Nucl. Phys. B310 (1988) 461.

[118] S. Gieseke, P. Stephens, and B. Webber, “New formalism for QCD parton showers,”
JHEP 12 (2003) 045, hep-ph/0310083.

[119] M. Bengtsson and T. Sjöstrand, “A comparative study of coherent and non-coherent

parton shower evolution,” Nucl. Phys. B289 (1987) 810.

[120] L. Lönnblad, “Ariadne version 4: A program for simulation of QCD cascades

implementing the colour dipole model,” Comput. Phys. Commun. 71 (1992) 15–31.

— 74 —
P. Skands Introduction to QCD

[121] Z. Nagy and D. E. Soper, “Matching parton showers to NLO computations,” JHEP 10
(2005) 024, hep-ph/0503053.

[122] S. Schumann and F. Krauss, “A Parton shower algorithm based on Catani-Seymour

dipole factorisation,” JHEP 0803 (2008) 038, 0709.1027.

[123] T. Sjöstrand and P. Z. Skands, “Transverse-momentum-ordered showers and

interleaved multiple interactions,” Eur. Phys. J. C39 (2005) 129–154,
hep-ph/0408302.

[124] N. Fischer, S. Gieseke, S. Plätzer, and P. Skands, “Revisiting radiation patterns in e+ e−

collisions,” Eur.Phys.J. C74 (2014) 2831, 1402.3186.

[125] J. C. Collins, “SPIN CORRELATIONS IN MONTE CARLO EVENT GENERATORS,”

Nucl.Phys. B304 (1988) 794.

[126] I. Knowles, “Spin Correlations in Parton - Parton Scattering,” Nucl.Phys. B310 (1988)
571.

[127] P. Richardson, “Spin correlations in Monte Carlo simulations,” JHEP 11 (2001) 029,
hep-ph/0110108.

[128] S. Catani, B. R. Webber, and G. Marchesini, “QCD coherent branching and

semiinclusive processes at large x,” Nucl. Phys. B349 (1991) 635–654.

[129] M. A. Gigg and P. Richardson, “Simulation of Finite Width Effects in Physics Beyond
the Standard Model,” 0805.3037.

[130] W. T. Giele, D. A. Kosower, and P. Z. Skands, “A Simple shower and matching

algorithm,” Phys. Rev. D78 (2008) 014026, 0707.3652.

[131] M. H. Seymour, “The Higgs boson line shape and perturbative unitarity,” Phys.Lett.
B354 (1995) 409–414, hep-ph/9505211.

[132] T. Stelzer and S. Willenbrock, “Spin correlation in top quark production at hadron
colliders,” Phys.Lett. B374 (1996) 169–172, hep-ph/9512292.

[133] S. J. Parke and Y. Shadmi, “Spin correlations in top quark pair production at e+ e−
colliders,” Phys.Lett. B387 (1996) 199–206, hep-ph/9606419.

[134] J. M. Smillie and B. R. Webber, “Distinguishing spins in supersymmetric and universal

extra dimension models at the large hadron collider,” JHEP 0510 (2005) 069,
hep-ph/0507170.

[135] M. H. Seymour, “A simple prescription for first-order corrections to quark scattering

and annihilation processes,” Nucl. Phys. B436 (1995) 443–460, hep-ph/9410244.

[136] M. H. Seymour, “Matrix-element corrections to parton shower algorithms,” Comp.

Phys. Commun. 90 (1995) 95–101, hep-ph/9410414.

[137] S. Catani, F. Krauss, R. Kuhn, and B. R. Webber, “QCD matrix elements + parton
showers,” JHEP 11 (2001) 063, hep-ph/0109231.

— 75 —
P. Skands Introduction to QCD

[138] L. Lönnblad, “Correcting the colour-dipole cascade model with fixed order matrix
elements,” JHEP 05 (2002) 046, hep-ph/0112284.

[139] N. Lavesson and L. Lönnblad, “W +jets matrix elements and the dipole cascade,” JHEP
07 (2005) 054, hep-ph/0503293.

[140] M. L. Mangano, M. Moretti, F. Piccinini, and M. Treccani, “Matching matrix elements

and shower evolution for top-pair production in hadronic collisions,” JHEP 01 (2007)
013, hep-ph/0611129.

[141] S. Mrenna and P. Richardson, “Matching matrix elements and parton showers with
H ERWIG and P YTHIA,” JHEP 05 (2004) 040, hep-ph/0312274.

[142] B. Cooper, J. Katzy, M. Mangano, A. Messina, L. Mijovic, et al., “Importance of a

consistent choice of alpha(s) in the matching of AlpGen and Pythia,” Eur.Phys.J. C72
(2012) 2078, 1109.5295.

[143] N. Lavesson and L. Lönnblad, “Extending CKKW-merging to one-loop matrix

elements,” JHEP 12 (2008) 070, 0811.2912.

[144] S. Frixione and B. R. Webber, “Matching NLO QCD computations and parton shower
simulations,” JHEP 06 (2002) 029, hep-ph/0204244.

[145] S. Frixione, P. Nason, and B. R. Webber, “Matching NLO QCD and parton showers in
heavy flavour production,” JHEP 08 (2003) 007, hep-ph/0305252.

[146] S. Frixione and B. R. Webber, “The MC@NLO 3.4 Event Generator,” 0812.0770.

[147] S. Frixione, P. Nason, and C. Oleari, “Matching NLO QCD computations with parton
shower simulations: the POWHEG method,” JHEP 11 (2007) 070, 0709.2092.

[148] K. Hamilton and P. Nason, “Improving NLO-parton shower matched simulations with
higher order matrix elements,” JHEP 06 (2010) 039, 1004.1764.

[149] M. Bengtsson and T. Sjöstrand, “Coherent Parton Showers Versus Matrix Elements:
Implications of PETRA - PEP Data,” Phys. Lett. B185 (1987) 435.

[150] S. Alioli, P. Nason, C. Oleari, and E. Re, “A general framework for implementing NLO
calculations in shower Monte Carlo programs: the POWHEG BOX,” JHEP 06 (2010)
043, 1002.2581.

[151] L. Hartgring, E. Laenen, and P. Skands, “Antenna Showers with One-Loop Matrix
Elements,” JHEP 1310 (2013) 127, 1303.4974.

[152] P. Z. Skands, “QCD for Collider Physics,” 1104.2863. In 18th European school on
high-energy physics (ESHEP), Raseborg, Finland, 20 June - 3 July, 2010,
arXiv:1202.1629.

[153] G. Bali and K. Schilling, “Static quark - anti-quark potential: Scaling behavior and
finite size effects in SU(3) lattice gauge theory,” Phys.Rev. D46 (1992) 2636–2646.

— 76 —
P. Skands Introduction to QCD

[154] D. Amati and G. Veneziano, “Preconfinement as a Property of Perturbative QCD,” Phys.

Lett. B83 (1979) 87.

[155] S. Ostapchenko, “Status of QGSJET,” AIP Conf. Proc. 928 (2007) 118–125, 0706.3784.

[156] F. Riehn, R. Engel, A. Fedynitch, T. K. Gaisser, and T. Stanev, “A new version of the
event generator Sibyll,” 1510.00568.

[157] F. W. Bopp, J. Ranft, R. Engel, and S. Roesler, “Antiparticle to Particle Production

Ratios in Hadron-Hadron and d-Au Collisions in the DPMJET-III Monte Carlo,” Phys.
Rev. C77 (2008) 014904, hep-ph/0505035.

[158] F. W. Bopp, R. Engel, and J. Ranft, “Rapidity gaps and the PHOJET Monte Carlo,” in
High energy physics. Proceedings, LAFEX International School, Session C, Workshop on
Diffractive Physics, LISHEP’98, Rio de Janeiro, Brazil, February 16-20, 1998,
pp. 729–741. 1998. hep-ph/9803437.

[159] T. Pierog, I. Karpenko, J. M. Katzy, E. Yatsenko, and K. Werner, “EPOS LHC: Test of
collective hadronization with data measured at the CERN Large Hadron Collider,”
Phys. Rev. C92 (2015), no. 3, 034906, 1306.0121.

[160] S. Navin, “Diffraction in Pythia,” 1005.3894.

[161] X. Artru and G. Mennessier, “String model and multiproduction,” Nucl. Phys. B70
(1974) 93–115.

[162] M. Travis, T. E. Ford, F. Laine, E. Presley, J. Cash, et al., “Sixteen Tons.” Originally
recorded for Folk Songs of the Hills, Capitol Records, 1946. Capitol T-891.

[163] A. Casher, H. Neuberger, and S. Nussinov, “Chromoelectric Flux Tube Model of Particle
Production,” Phys. Rev. D20 (1979) 179–188.

[164] J. S. Schwinger, “On gauge invariance and vacuum polarization,” Phys. Rev. 82 (1951)
664–679.

[165] G. V. Dunne, “New Strong-Field QED Effects at ELI: Nonperturbative Vacuum Pair
Production,” Eur. Phys. J. D55 (2009) 327–340, 0812.3163.

[166] N. K. Glendenning and T. Matsui, “CREATION OF ANTI-Q Q PAIR IN A

CHROMOELECTRIC FLUX TUBE,” Phys. Rev. D28 (1983) 2890–2891.

[167] G. C. Nayak, “Non-perturbative quark-antiquark production from a constant

chromo-electric field via the Schwinger mechanism,” Phys. Rev. D72 (2005) 125010,
hep-ph/0510052.

[168] N. Fischer and T. Sjöstrand, “Thermodynamical String Fragmentation,” JHEP 01

(2017) 140, 1610.09818.

[169] P. Skands, S. Carrazza, and J. Rojo, “Tuning PYTHIA 8.1: the Monash 2013 Tune,”
Eur.Phys.J. C74 (2014), no. 8, 3024, 1404.5630.

— 77 —
P. Skands Introduction to QCD

[170] B. Andersson, G. Gustafson, and T. Sjöstrand, “A Model for Baryon Production in

Quark and Gluon Jets,” Nucl. Phys. B197 (1982) 45.

[171] B. Andersson, G. Gustafson, and T. Sjöstrand, “BARYON PRODUCTION IN JET

FRAGMENTATION AND UPSILON DECAY,” Phys.Scripta 32 (1985) 574.

[172] P. Eden and G. Gustafson, “Baryon production in the string fragmentation picture,”
Z.Phys. C75 (1997) 41–49, hep-ph/9606454.

[173] T. Sjöstrand and P. Z. Skands, “Baryon number violation and string topologies,” Nucl.
Phys. B659 (2003) 243, hep-ph/0212264.

[174] T. Sjöstrand and P. Z. Skands, “Multiple interactions and the structure of beam
remnants,” JHEP 03 (2004) 053, hep-ph/0402078.

[175] B. Andersson, G. Gustafson, and B. Söderberg, “A General Model for Jet

Fragmentation,” Z.Phys. C20 (1983) 317.

[176] M. G. Bowler, “e+ e− production of heavy quarks in the string model,” Z. Phys. C11
(1981) 169.

[177] K. G. Wilson, “Confinement of Quarks,” Phys.Rev. D10 (1974) 2445–2459.

[178] TOTEM Collaboration, G. Antchev et al., “Luminosity-Independent Measurement of

√
the Proton-Proton Total Cross Section at s = 8 TeV,” Phys.Rev.Lett. 111 (2013), no. 1,
012001.

[179] A. Donnachie and P. Landshoff, “pp and p̄p total cross sections and elastic scattering,”
Phys.Lett. B727 (2013) 500–505, 1309.1292.

[180] V. Khoze, F. Krauss, A. Martin, M. Ryskin, and K. Zapp, “Diffraction and correlations at
the LHC: Definitions and observables,” Eur.Phys.J. C69 (2010) 85–93, 1005.4839.

[181] B. Blok, Y. Dokshitzer, L. Frankfurt, and M. Strikman, “The Four jet production at LHC
and Tevatron in QCD,” Phys.Rev. D83 (2011) 071501, 1009.2714.

[182] B. Blok, Y. Dokshitser, L. Frankfurt, and M. Strikman, “pQCD physics of multiparton

interactions,” Eur.Phys.J. C72 (2012) 1963, 1106.5533.

[183] J. R. Gaunt and W. J. Stirling, “Single and Double Perturbative Splitting Diagrams in
Double Parton Scattering,” 1202.3056.

[184] A. V. Manohar and W. J. Waalewijn, “A QCD Analysis of Double Parton Scattering:

Color Correlations, Interference Effects and Evolution,” Phys.Rev. D85 (2012) 114009,
1202.3794.

[185] J. M. Butterworth, J. R. Forshaw, and M. H. Seymour, “Multiparton Interactions in

Photoproduction at HERA,” Z. Phys. C72 (1996) 637–646, hep-ph/9601371.

[186] M. Bähr, J. M. Butterworth, S. Gieseke, and M. H. Seymour, “Soft interactions in

Herwig++,” 0905.4671.

— 78 —
P. Skands Introduction to QCD

[187] M. Ryskin, A. Martin, and V. Khoze, “High-energy strong interactions: from ‘hard’ to
‘soft’,” Eur.Phys.J. C71 (2011) 1617, 1102.2844.

[188] H. Schulz and P. Skands, “Energy Scaling of Minimum-Bias Tunes,” Eur.Phys.J. C71
(2011) 1644, 1103.3649.

[189] P. Z. Skands, “Tuning Monte Carlo Generators: The Perugia Tunes,” Phys.Rev. D82
(2010) 074018, 1005.3457.

[190] P. Z. Skands, “Soft-QCD and UE spectra in pp collisions at very high CM energies (a

Snowmass white paper),” 1308.2813.

[191] UA1 Collaboration, G. Arnison et al., “Hadronic Jet Production at the CERN Proton -
anti-Proton Collider,” Phys. Lett. 132B (1983) 214.

[192] A. Karneyeu, L. Mijovic, S. Prestel, and P. Skands, “MCPLOTS: a particle physics

resource based on volunteer computing,” Eur.Phys.J. C74 (2014) 2714, 1306.3436.

[193] R. Corke and T. Sjöstrand, “Multiparton Interactions with an x-dependent Proton

Size,” JHEP 1105 (2011) 009, 1101.5953.

[194] ATLAS Collaboration, G. Aad et al., “Measurement of underlying event characteristics

√
using charged particles in pp collisions at s = 900GeV and 7 TeV with the ATLAS
detector,” Phys.Rev. D83 (2011) 112001, 1012.0791.

[195] D. Lombrana Gonzalez, L. Rivkin, M. Zacharov, I. an Giovannozzi, E. McIntosh, et al.,

“LHC@home: a Volunteer Computing System for Massive Numerical Simulations of
Beam Dynamics and High Energy Physics Events,” Conf.Proc. C1205201 (2012)
505–507.

[196] A. Buckley, H. Hoeth, H. Lacker, H. Schulz, and J. E. von Seggern, “Systematic event
generator tuning for the LHC,” Eur. Phys. J. C65 (2010) 331–357, 0907.2973.

[197] P. Richardson and D. Winn, “Investigation of Monte Carlo Uncertainties on Higgs

Boson searches using Jet Substructure,” 1207.0380.

[198] J. R. Andersen et al., “Les Houches 2015: Physics at TeV Colliders Standard Model
Working Group Report,” in 9th Les Houches Workshop on Physics at TeV Colliders
(PhysTeV 2015) Les Houches, France, June 1-19, 2015. 2016. 1605.04692.

[199] J. Bellm, S. Platzer, P. Richardson, A. Siodmok, and S. Webster, “Reweighting Parton

Showers,” 1605.08256.

[200] M. Ritzmann, D. Kosower, and P. Skands, “Antenna Showers with Hadronic Initial
States,” Phys.Lett. B718 (2013) 1345–1350, 1210.6345.

[201] P. Skands, B. Webber, and J. Winter, “QCD Coherence and the Top Quark Asymmetry,”
JHEP 1207 (2012) 151, 1205.1466.

[202] R. Field, “Min-Bias and the Underlying Event at the LHC,” Acta Phys.Polon. B42 (2011)
2631–2656, 1110.5530.

— 79 —
P. Skands Introduction to QCD

[203] T. Martin, P. Skands, and S. Farrington, “Probing Collective Effects in Hadronisation

with the Extremes of the Underlying Event,” Eur. Phys. J. C76 (2016), no. 5, 299,
1603.05298.

[204] CDF Collaboration, T. Aaltonen et al., “Studying the Underlying Event in Drell-Yan and
High Transverse Momentum Jet Production at the Tevatron,” Phys. Rev. D82 (2010)
034001, 1003.3146.

[205] CMS Collaboration, S. Chatrchyan et al., “Measurement of the underlying event in the
√
Drell-Yan process in proton-proton collisions at s = 7 TeV,” Eur. Phys. J. C72 (2012)
2080, 1204.1411.

[206] ATLAS Collaboration, G. Aad et al., “Measurement of distributions sensitive to the

√
underlying event in inclusive Z-boson production in pp collisions at s = 7 TeV with
the ATLAS detector,” Eur. Phys. J. C74 (2014), no. 12, 3195, 1409.3433.

[207] S. Alioli, C. W. Bauer, S. Guns, and F. J. Tackmann, “Underlying event sensitive

observables in Drell-Yan production using GENEVA,” 1605.07192.

[208] CMS Collaboration, “Study of the underlying event, b-quark fragmentation and
hadronization properties in tbart events,”. CMS-PAS-TOP-13-007.

[209] CMS Collaboration, “Underlying event measurement with ttbar+X events with p-p
collision data at sqrt(s)=13 TeV,”. CMS-PAS-TOP-15-017.

[210] L. Lönnblad and T. Sjöstrand, “Modelling Bose-Einstein correlations at LEP 2,” Eur.
Phys. J. C2 (1998) 165–180, hep-ph/9711460.

[211] R. Corke and T. Sjöstrand, “Multiparton interactions and rescattering,” JHEP 01

(2009) 035, 0911.1909.

[212] J. Rathsman, “A generalised area law for hadronic string reinteractions,” Phys. Lett.
B452 (1999) 364–371, hep-ph/9812423.

[213] P. Z. Skands and D. Wicke, “Non-perturbative QCD effects and the top mass at the
Tevatron,” Eur.Phys.J. C52 (2007) 133–140, hep-ph/0703081.

[214] F. Bopp and J. Ranft, “Inclusive distributions in p-p collisions at LHC energies
compared with an adjusted DPMJET-III model with chain fusion,” in Proceedings, 3rd
International Workshop on Multiple Partonic Interactions at the LHC (MPI@LHC 2011),
pp. 41–49. 2011. 1110.6403.

[215] S. Gieseke, C. Rohr, and A. Siodmok, “Colour reconnections in Herwig++,” Eur.Phys.J.

C72 (2012) 2225, 1206.0041.

[216] J. R. Christiansen and P. Z. Skands, “String Formation Beyond Leading Colour,” JHEP
08 (2015) 003, 1505.01681.

[217] K. Werner, I. Karpenko, and T. Pierog, “Hydrodynamical evolution in heavy ion

collisions and p p scatterings at the LHC: Ridges in A A and p p scattering,” Acta
Phys.Polon.Supp. 4 (2011) 629–634.

— 80 —
P. Skands Introduction to QCD

[218] P. R. Petrov, Strangeness enhancement in high multiplicity proton-proton collisions at

√
s = 7 TeV with the ALICE detector at the LHC. PhD thesis, Birmingham U.,
2012-11-28.

[219] C. Bierlich and J. R. Christiansen, “Effects of color reconnection on hadron flavor

observables,” Phys. Rev. D92 (2015), no. 9, 094010, 1507.02091.

[220] ALICE Collaboration, J. Adam et al., “Multiplicity-dependent enhancement of strange

√
and multi-strange hadron production in proton-proton collisions at s = 7 TeV,”
1606.07424.

[221] CMS Collaboration, S. Chatrchyan et al., “Measurement of energy flow at large

√
pseudorapidities in pp collisions at s = 0.9 and 7 TeV,” JHEP 11 (2011) 148,
1110.0211. [Erratum: JHEP02,055(2012)].

[222] CMS Collaboration, S. Chatrchyan et al., “Forward Energy Flow, Central

Charged-Particle Multiplicities, and Pseudorapidity Gaps in W and Z Boson Events
√
from pp Collisions at s = 7 TeV,” Eur. Phys. J. C72 (2012) 1839, 1110.0181.

[223] TOTEM Collaboration, G. Antchev et al., “Measurement of the forward charged

√
particle pseudorapidity density in pp collisions at s = 7 TeV with the TOTEM
experiment,” Europhys. Lett. 98 (2012) 31002, 1205.4105.

[224] CMS Collaboration, S. Chatrchyan et al., “Study of the underlying event at forward
√
rapidity in pp collisions at s = 0.9, 2.76, and 7 TeV,” JHEP 04 (2013) 072,
1302.2394.

[225] LHCb Collaboration, R. Aaij et al., “Measurement of charged particle multiplicities and
√
densities in pp collisions at s = 7 TeV in the forward region,” Eur. Phys. J. C74
(2014), no. 5, 2888, 1402.4430.

[226] CMS, TOTEM Collaboration, S. Chatrchyan et al., “Measurement of pseudorapidity

√
distributions of charged particles in proton-proton collisions at s = 8 TeV by the
CMS and TOTEM experiments,” Eur. Phys. J. C74 (2014), no. 10, 3053, 1405.0722.

[227] TOTEM Collaboration, G. Antchev et al., “Measurement of the forward charged

√
particle pseudorapidity density in pp collisions at s = 8 TeV using a displaced
interaction point,” Eur. Phys. J. C75 (2015), no. 3, 126, 1411.4963.

[228] S. Erhan, W. S. Lockman, T. Meyer, J. Rander, P. Schlein, R. Webb, and J. Zsembery,

“HYPERON PRODUCTION IN P P INTERACTIONS AT S**(1/2) = 53-GEV AND
62-GEV,” Phys. Lett. B85 (1979) 447–451.

[229] LHCb Collaboration, R. Aaij et al., “Measurement of V 0 production ratios in pp

√
collisions at s = 0.9 and 7 TeV,” JHEP 08 (2011) 034, 1107.0882.

[230] D0 Collaboration, V. M. Abazov et al., “Measurement of the forward-backward

asymmetry of Λ and Λ̄ production in pp̄ collisions,” Phys. Rev. D93 (2016), no. 3,
032002, 1511.05113.

— 81 —
P. Skands Introduction to QCD

[231] N. Ellis, R. Fleischer, et al., “Proceedings of the 2008 European School of High-Energy
Physics (ESHEP), Herbeumont-sur-Semois, Belgium, 8 - 21 June 2008,”.
CERN-2009-002.

[232] C. Grojean, M. Spiropulu, et al., “Proceedings of the 2009 European School of

High-Energy Physics (ESHEP), Bautzen, Germany, 14 - 27 June 2009,” 1012.4643.

[233] N. Ellis, R. Fleischer, et al., “2007 European School of High-Energy Physics, Trest,
Czech Republic, 19 August - 1 September 2007,”. CERN-2008-007.

— 82 —
Index
Adjoint representation, 5 Diquarks, 16, 56
αs , 5, 10–14 DIS, 7, 18, 20, 66
beta function, 11 Double counting, 45
ΛQCD , 13, 14 Drell-Yan, 7, 20
Landau Pole, see ΛQCD
Running coupling, 10–14 Eightfold way, 4
Altarelli-Parisi, see DGLAP Eikonal, 28, 29, 36
Angular ordering, 32, 37, 42 Elastic scattering, 58
Antenna functions, see Antennae Event evolution, see Evolution
Antennae, 28, 29, 36, 43 Event generators, see Monte Carlo
Anti-kT algorithm, see Jets Event shapes, 30
Area law, 57 Evolution, 40–42
Asymptotic freedom, 12, 13 Exclusive cross sections, 40, 41, 45

Baryons, 4, 56 Factorisation, 18, 35

Beta function, 11 Caveats, 20
Bjorken scaling, see QCD Scale invariance Factorisation scale, 19, 35, 42
Breit-Wigner distribution, 44 Flavour, 4, 12, 14
Bremsstrahlung, see Parton showers In string breaks, 55, 56
Fragmentation functions, 35, 52, 56
CA , 8, 10, 28 Fundamental representation, 5
Casimirs, 6, 8
Catani-Seymour dipoles, 28 Gell-Mann matrices, see SU(3)
Causality, 54 Gluons, 6, 8, 9
CF , 8, 10, 28 As kinks on strings, 53
CKKW, see Matching, 47 Contribution to beta function, 11
Cluster model, 51 PDFs, 21
Coherence, 15, 37
Hadronisation, 31, 51
Collinear limit, 24, 26, 28, 30, 35, 36
Hadronisation scale, 42, 51, 55
Color connections, 28
HERWIG, 32, 45, 46, 51
Color factors, 28
Colour connections, 8, 66 Inclusive cross sections, 20, 40, 41, 45
Colour factors, 6, 7, 10 Inelastic scattering, 58
Colour-space indices, see Colour connections Infrared divergences, 24, 25, 33
Confinement, 51, 53 Infrared safety, 29–31
Conformal invariance, 10
Convergence, 33, 34 Jets, 10, 24, 27, 35, 40
Algorithms, see [5]
Decuplet, 4, 16 Anti-kT algorithm, 31
Deep inelastic scattering, see DIS Cambridge/Aachen algorithm, 31
Detailed balance, 41 Definitions, see [5]
DGLAP equation, 20 Infrared safety, 29, 30
DGLAP kernels, 20, 28, 43 kT algorithm, 31
Diffractive scattering, 58, 59

83
P. Skands Introduction to QCD

Kinoshita-Lee-Nauenberg, see KLN theorem Narrow width approximation, 44

KLN theorem, 25, 27, 28, 40 NLO, 19, 24–28, 48
Infrared divergences, 24, 25
ΛQCD , see αs NNLO, 24, 26–28
Landau pole, see αs Nobel prize, 12, 25
Lattice QCD, 15, 51 Numerical integration, 33
Les Houches Event Files, see LHEF
LHEF, 32 Octet, 4
Lightcone scaling, see QCD Scale invariance
Linear confinement, see Confinement Parton Distributions, see PDFs
LO, 24, 25, 45 Parton level, 51
Lund symmetric fragmentation function, 56 Parton showers, 10, 22, 31, 34, 42
Bowler modification, 57 Evolution variable, 42
Infrared cutoff, see Hadronisation scale
Markov chains, 34, 42, 43 Initial-state evolution, 22
Matching, 45–50 Sudakov factor, 43
CKKW, 45, 47 Partons, 3, 20
L-CKKW, 45, 47 PDFs, 19, 20, 35
Matching scale, 47 POWHEG, see Matching, 50
MCatNLO, 48 pQCD, 24, 29, 34
MENLOPS, 49 Preconfinement, 51
MLM, 45, 47 Primary hadrons, 51
POWHEG, 48 p⊥ -ordering, 42
Slicing, 45 PYTHIA, 32, 50, 51, 53, 56, 61
Subtraction, 48
Unitarity, 49 QCD
VINCIA, 50 Colour factors, 6
Matching scale, see Matching Coupling, 5, 6, 10
Matrix elements, 6, 17, 24, 25 Lagrangian, 3, 5
MCatNLO, see Matching Landau Pole, see αs
MCnet review, 32 Renormalisation, see αs
MENLOPS, see Matching Scale invariance, 10, 24, 47
Mesons, 4, 55, 56 Structure constants, see SU(3)
Minimum-bias events, 58, 60 Trace relations, see SU(3)
MLM, see Matching Quarks, 4–7
Monte Carlo As string endpoints, 53
Convergence, 34 Contribution to beta function, 11
Decay distributions, 44 PDFs, 21
Event generators, 22, 32, 51, 64, 65 RAMBO, 33, 34
Integration, 32–34 Renormalisation, see αs
Monaco, 34 Resonance decays, 44
Pedagogical introduction, 44
RAMBO, 33 SARGE, 34
Statistics, 33 Scale invariance, see QCD
Tuning, 39, 55, 64, 65 Schwinger Mechanism, 54, 56
Uncertainties, 33, 64 Sector decomposition, 28
VEGAS, 34 Sextet, 15

— 84 —
P. Skands Introduction to QCD

SHERPA, 32, 51
Soft eikonal, see Eikonal
Soft limit, 24, 26, 28, 30, 36
Spin correlations, 39, 44
String breaks, 54
String junctions, 16, 56
String model, 51, 53, 55, 66
Strong coupling, 13
Structure functions, 22
SU(3), 4, 5, 8
Adjoint representation, 5
Fundamental representation, 5
Generators, 6
Of Colour, 4
Of Flavour, 4
Structure constants, 8
Trace relations, 8
Subtraction, 27–29
In the context of matching, 48
Subtraction schemes, 29
Sudakov factor, 43, 45, 50

TR , 8
Traces in SU(3), see SU(3)
Transcendentality, 39
Transverse-momentum-ordering, see p⊥ -ordering
Tuning, see Monte Carlo
Tunneling, 54
Twist, 20

Uncertainties
Factorisation, 19, 20
Monte Carlo statistics, 33
PDF sets, 21
Underlying event, 31, 51, 58
Unitarity, 25, 27, 40

VEGAS, see Monte Carlo

VINCIA, 34, 50
Matching, 50

Young tableaux, 16

— 85 —