Graduate Texts in Mathematics 39

Editorial Board
S. Axler EW. Gehring K.A. Ribet

Springer
New York
Berlin
Heidelberg
Barcelona
Budapest
Hong Kong
London
Milan
Paris
Santa Clara
Singapore
Tokyo
Graduate Texts in Mathematics

TAKEUTJIZARING. Introduction to 33 HIRSCH. Differential Topology.

Axiomatic Set Theory. 2nd ed. 34 SPITZER. Principles of Random Walk.
2 OXTOBY. Measure and Category. 2nd ed. 2nd ed.
3 SCHABFIlR. Topological Vector Spaces. 35 Au!xANDERlWBRMER. Several Complex
4 Hn.roNlSTAMMBACH. A Course in Variables and Banach Algebras. 3rd ed.
Homological Algebra. 2nd ed. 36 KEu.i!Y/NAMIOKA et al. Linear
5 MAc LANE. Categories for the Working Topological Spaces.
Mathematician. 37 MONK. Mathematical Logic.
6 HUGlms/P!PBR. Projective Planes. 38 GRAUERTIFRrrzsCHB. Several Complex
7 SERRE. A Course in Arithmetic. Variables.
8 TAKEUTJIZARING. Axiomatic Set Theory. 39 ARVESON. An Invitation to c*-Algebras.
9 HUMPHREYS. Introduction to Lie Algebras 40 KBMENY/SNEUlKNAPP. Denumerable
and Representation Theory. Markov Chains. 2nd ed.
10 COHEN. A Course in Simple Homotopy 41 APosTOL. Modular Functions and
Theory. Dirichlet Series in Number Theory.
11 CONWAY. Functions of One Complex 2nd ed.
Variable I. 2nd ed. 42 SERRE. Linear Representations of Finite
12 BEALS. Advanced Mathematical Analysis. Groups.
13 ANoBRSONIFuu.BR. Rings and Categories 43 Gll.LMAN/JBRlSON. Rings of Continuous
of Modules. 2nd ed. Functions.
14 GoLUBITSKY/GUILLBMIN. Stable Mappings 44 KENDIG. Elementary Algebraic Geometry.
and Their Singularities. 45 LoaVE. Probability Theory I. 4th ed.
15 BERBBRIAN. Lectures in Functional 46 LoaVE. Probability Theory n. 4th ed.
Analysis and Operator Theory. 47 MOISE. Geometric Topology in
16 WINTBR. The Structure of Fields. Dimensions 2 and 3.
17 RosBNBLATI. Random Processes. 2nd ed. 48 SACHSIWu. General Relativity for
18 HALMos. Measure Theory. Mathematicians.
19 HALMOS. A Hilbert Space Problem Book. 49 GRUBNBBRGlWBJR. Linear Geometry.
2nd ed. 2nd ed.
20 HUSBMOLLBR. Fibre Bundles. 3rd ed. 50 EDWARDS. Fermat's Last Theorem.
21 HUMPHREYS. Linear Algebraic Groups. 51 KUNGBNBBRG. A Course in Differential
22 BARNBSIMACK. An Algebraic Introduction Geometry.
to Mathematical Logic. 52 HARTSHORNE. Algebraic Geometry.
23 GRBUB. Linear Algebra. 4th ed. 53 MANIN. A Course in Mathematical Logic.
24 HoLMES. Geometric Functional Analysis 54 GRAVERlWATKlNS. Combinatorics with
and Its Applications. Emphasis on the Theory of Graphs.
25 HeWITT/STROMBBRG. Real and Abstract 55 BROwNIPBARcy. Introduction to Operator
Analysis. Theory I: Elements of Functional
26 MANes. Algebraic Theories. Analysis.
27 KEu.i!Y. General Topology. 56 MAsSEY. Algebraic Topology: An
28 ZARlsKIlSAMUEL. Commutative Algebra. Introduction.
Vou. 57 CRoWBLLIFox. Introduction to Knot
29 ZARlsKIlSAMUEL. Commutative Algebra. Theory.
VoW. 58 KOBLITL. p-adic Numbers, p-adic
30 JACOBSON. Lectures in Abstract Algebra I. Analysis, and Zeta-Functions. 2nd ed.
Basic Concepts. 59 LANG. Cyclotomic Fields.
31 JACOBSON. Lectures in Abstract Algebra 60 ARNOLD. Mathematical Methods in
n. Linear Algebra. Classical Mechanics. 2nd ed.
32 JACOBSON. Lectures in Abstract Algebra
m. Theory of Fields and Galois Theory. continued after index
 William Arveson

An Invitation to
C* -Algebras

, Springer
William Arveson
Department of Mathematics
University of California
Berkeley, California 94720
USA

Editorial Board
S. Axler EW. Gehring K.A. Ribet
Mathematics Department Mathematics Department Department of
San Francisco State University East Hall Mathematics
San Francisco, CA 94132 University of Michigan University of California
USA Ann Arbor, MI 48109 at Berkeley
USA Berkeley, CA 94720
USA

AMS Subject Classifications Primary: 46L05, 46LIO, 46KIO, 47CI0

Secondary: 81 A 17, 81 A54

Library of Congress Cataloging-in-Publication Data

Arveson, William.
An invitation to C*-algebras.
(Graduate texts in mathematics; 39)
Bibliography.
Includes index.
I. C*-algebras. 2. Representations of algebras. 3. Hilbert space. I. Title. II. Series.
QA326.A 78 512'.55 76-3656

Printed on acid-free paper.

© 1976 Springer-Verlag New York, Inc.

Softcover reprint of the hardcover Ist edition 2007
All rights reserved. This work may not be translated or copied in whole or in part without the
written permission of the publisher (Springer-Verlag New York, Inc., 175 Fifth Avenue, New
York, NY 10010, USA), except for brief excerpts in connection with reviews or scholarly
analysis. Use in connection with any form of information storage and retrieval, electronic
adaptation, computer software, or by similar or dissimilar methodology now known or hereafter
developed is forbidden.
The use of general descriptive names, trade names, trademarks, etc., in this publication, even
if the former are not especially identified, is not to be taken as a sign that such names, as
understood by the Trade Marks and Merchandise Marks Act, may accordingly be used freely
by anyone.

Production managed by Allan Abrams; manufacturing supervised by Jeffrey Taub.

9 8 7 6 5 4 3 2 (Corrected second printing, 1998)

ISBN-13: 978-1-4612-6373-9 e-ISBN-13: 978-1-4612-6371-5
001: 10.1007/978-1-4612-6371-5
 Preface

This book gives an introduction to C*-algebras and their representations on

Hilbert spaces. We have tried to present only what we believe are the most basic
ideas, as simply and concretely as we could. So whenever it is convenient (and it
usually is), Hilbert spaces become separable and C*-algebras become GCR. This
practice probably creates an impression that nothing of value is known about other
C*-algebras. Of course that is not true. But insofar as representations are con-
cerned, we can point to the empirical fact that to this day no one has given a
concrete parametric description of even the irreducible representations of any
C*-algebra which is not GCR. Indeed, there is metamathematical evidence which
strongly suggests that no one ever will (see the discussion at the end of Section
3.4). Occasionally, when the idea behind the proof of a general theorem is exposed
very clearly in a special case, we prove only the special case and relegate
generalizations to the exercises.
In effect, we have systematically eschewed the Bourbaki tradition.
We have also tried to take into account the interests of a variety of readers. For
example, the multiplicity theory for normal operators is contained in Sections 2.1
and 2.2. (it would be desirable but not necessary to include Section 1.1 as well),
whereas someone interested in Borel structures could read Chapter 3 separately.
Chapter I could be used as a bare-bones introduction to C*-algebras. Sections 2.1
and 2.3 together contain the basic structure theory for type I von Neumann
algebras, and are also largely independent of the rest of the book.
The level of exposition should be appropriate for a second year graduate student
who is familiar with the basic results of functional analysis, measure theory, and
Hilbert space. For example, we assume the reader knows the Hahn - Banach
theorem, Alaoglu's theorem, the Krein- Milman theorem, the spectral theorem
for normal operators, and the elementary theory of commutative Banach algebras.
On the other hand, we have avoided making use of dimension theory and most of

v
Preface

the more elaborate machinery of reduction theory (though we do use the notation
for direct integrals in Sections 2.4 and 4.3). More regrettably, some topics have
been left out merely to keep down the size of the book; for example, applications to
the theory of unitary representations of locally compact groups are barely men-
tioned. To fill in these many gaps, the reader should consult the comprehensive
monographs of Dixmier [6, 7].
A preliminary version of this manuscript was finished in 1971, and during the
subsequent years was widely circulated in preprint form under the title
Representations ofC*-algebras. The present book has been reorganized, and new
material has been added to correct what we felt were serious omissions in the
earlier version. It has been used as the basis for lectures in Berkeley and in Aarhus.
We are indebted to many colleagues and students who read the manuscript,
pointed out errors, and offered constructive criticism. Special thanks go to Cecelia
Bleecker, Larry Brown, Paul Chernoff, Ron Douglas, Dick Loebl, Donal
O'Donovan, Joan Plastiras, and Erling St0rmer.
This subject has more than its share of colorless and obscure terminology. In
particular, one always has to choose between calling a C*-algebra GCR, type I, or
postliminal. The situation is no better in French: does postliminaire mean post-
preliminary? In this book we have reverted to Kaplansky's original acronym,
simply because it takes less space to write. More sensibly, we have made use of
Halmos' symbol 0 to signal the end of a proof.

vi
 Contents

Chapter 1
Fundamentals
1.1. Operators and C*-algebras
1.2. Two density theorems 5
1.3. Ideals, quotients, and representations 10
1.4. C*-algebras of compact operators 17
I .5. CCR and GCR algebras 22
1.6. States and the GNS construction 27
1.7. The existence of representations 31
1.8. Order and approximate units 35

Chapter 2
Multiplicity Theory 40
2.1. From type I to multiplicity-free 41
2.2. Commutative C*-algebras and nonnal operators 49
2.3. An application: type I von Neumann algebras 56
2.4. GCR algebras are type I 59

Chapter 3
Borel Structures 61
3.1 . Polish spaces 61
3.2. Borel sets and analytic sets 64
3.3. Borel spaces 69
3.4. Cross sections 74

vii
Chapter 4
From Commutative Algebras to GCR Algebras 81
4.1. The spectrum of a C*-algebra 81
4.2. Decomposable operator algebras 88
4.3. Representations ofGCR algebras 94

Bibliography 103

Index 105

viii
An Invitation to
C* -Algebras
 1
Fundamentals

This chapter contains what we consider to be the essentials of non-

commutative C*-algebra theory. This is the material that anyone who wants
to work seriously with C*-algebras needs to know. The most tractable C*-
algebras are those that can be related to compact operators in a certain
specific way. These are the so-called GCR algebras, and they are introduced
in Section 1.5, after a rather extensive discussion of C*-algebras of compact
operators in Section 1.4.
Representations are first encountered in Section 1.3; they remain near
the center of discussion throughout the chapter, and indeed throughout the
remainder of the book (excepting Chapter 3).

1.1. Operators and C*-algebras

A C*-algebra of operators is a subset d ofthe algebra 2(£') of all bounded
operators on a Hilbert space £', which is closed under all of the algebraic
operations on 2(£') (addition, multiplication, multiplication by complex
scalars), is closed in the norm topology of 2(£'), and most importantly is
closed under the adjoint operation T 1-+ T* in 2(£'). Every operator T on
£' determines a C*-algebra c*(T), namely the smallest C*-algebra con-
taining both T and the identity. It is more or less evident that C*(T) is the
norm closure of all polynomials p(T, T*), where p(x, y) ranges over all
polynomials in the two free (i.e., noncommuting) variables x and y having
complex coefficients. However since T and T* do not generally commute,
these polynomials in T and T* are of little use in answering questions, and
in particular the above remark sheds no light on the structure of c*(T).
Nevertheless, C*(T) contains much information about T, and one could
1. Fundamentals

view this book as a description of what that information is and how one goes
about extracting it.
We will say that two operators Sand T (acting perhaps, on different
Hilbert spaces) are algebraically equivalent if there is a *-isomorphism (that
is, an isometric *-preserving isomorphism) of C*(S) onto C*(T) which carries
S into T. Note that this is more stringent than simply requiring that C*(S)
and C*(T) be *-isomorphic. We will see presently that two normal operators
are algebraically equivalent if and only if they have the same spectrum; thus
one may think of algebraically equivalent nonnormal operators as having
the same "spectrum" in some generalized sense, which will be made more
precise in Chapter 4.
We now collect a few generalities. A (general) C* -algebra is a Banach
algebra A having an involution * (that is, a conjugate-linear map of A into
itself satisfying x** = x and (xy)* = y*x*, x, YEA) which satisfies Ilx*xll =
IIxll2 for all x E A. It is very easy to see that a C*-algebra of operators on a
Hilbert space is a C*-algebra, and we will eventually prove a theorem of
Gelfand and Naimark which asserts the converse: every C*-algebra is iso-
metrically *-isomorphic with a C*-algebra of operators on a Hilbert space
(Theorem 1.7.3).
Let A be a commutative C*-algebra. Then in particular A is a commutative
Banach algebra, and therefore the set of all nonzero complex homomor-
phisms of A is a locally compact Hausdorff space in its usual topology. This
space will be called the spectrum of A, and it is written A. A standard result
asserts that A is compact iff A contains a multiplicative identity. Now the
Gelfand map is generally a homomorphism of A into the Banach algebra
C(A) of all continuous complex valued functions on A vanishing at 00. In
this case, however, much more is true.

Theorem 1.1.1. The Gelfand map is an isometric *-isomorphism of A onto C(A).

Here, the term *-isomorphism means that, in addition to the usual pro-
perties of an isomorphism, x* E A gets mapped into the complex conjugate
of the image of x. We will give the proof of this theorem for the case where
A contains an identity 1; the general case follows readily from this by the
process of adjoining an identity (Exercise 1.1.H).
First, let WE A. Then we claim w(x*) = w(x) for all x EA. This reduces
to proving that w(x) is real for all x = x* in A (since every x E A can be
written x = Xl + iX2' with Xi = xi E A). Therefore choose x = x* E A, and
for every real number t define Ut = eitx (for any element z E A, ez is defined
by the convergent power series L:'=o z"/n!, and the usual manipulations
show that ez+ w = ezeW since z and w commute). By examining the power
series we see that ut = e-itX, and hence utu t = e-itx+itx = 1. Thus lIu tl1 2 =
IlutUtl1 = 11111 = 1, and since the complex homomorphism w has norm 1 we
conclude exp t &/e iw(x) = Ieco(itx) I = Iw(ut)1 ~ 1, for all t E ~. This can only
mean &/e iw(x) = 0, and hence w(x) is real.

2
 1.1. Operators and C·-Algebras

Now let y(x) denote the image of x in C(A), i.e., y(x)(w) = w(x), WE A.
Then we have just proved y(x·) = y(x), and we now claim Ily(x)11 = Ilxll.
Indeed the left side is the spectral radius of x which, by the Gelfand-Mazur
theorem, is limn Iixnlil/n. But if x = x· then we have IIxl12 = Ilx·xll = Ilx 211;
replacing x with x 2 gives IIxl14 = IIx 2112 = Ilx411, and so on inductively, giving
IlxW" = Ilx2"11, n ~ 1. This proves Ilxll = limn IlxnWln if x = x·, and the
case of general x reduces to this by the trick Ily(x)11 2 = Ily(x)y(x)1I =
Ily(x·x)11 = Ilx·xll = Ilx 11 2, applying the above to the self-adjoint element x·x.
Thus y is an isometric ·-isomorphism of A onto a closed self-adjoint
subalgebra of C(A) containing 1; since y(A) always separates points, the
proof is completed by an application of the Stone-Weierstrass theorem. D

An element x ofa C·-algebra is called normal ifx·x = xx·. Note that this is
equivalent to saying that the sub C*-algebra generated by x is commutative.

Corollary. If x is a normal element of a C·-algebra with identity, then the

norm of x equals its spectral radius.
PROOF. Consider x to be an element of the commutative C*-algebra it
generates (together with the identity). Then the assertion follows from the
fact that the Gelfand map is an isometry. D

Theorem 1.1.1 is sometimes called the abstract spectral theorem, since it

provides the basis for a powerful functional calculus in C·-algebras. In order
to discuss this, let us first recall that if B is a Banach subalgebra of a Banach
algebra A with identity 1, such that 1 E B, then an element x in B has a spec-
trum SPA (x) relative to A as well as a spectrum SPB(X) relative to B, and in
general one has SPA (x) £; SPB(X). Of course, the inclusion is often proper.
But if A is a C·-algebra and B is a C·-subalgebra, then the two spectra must
be the same. To indicate why this is so, we will show that if x E B is invertible
in A, then X-I belongs to B (a moment's thought shows that the assertion
reduces to this). For that, note that x· is invertible, and since the element
(X·X)-IX· is clearly a left inverse for x, we must have X-I = (X·X)-IX·. SO
to prove that X-I E B, it suffices to show that (X·X)-l E B. Actually, we will
show that x·x is invertible in the still smaller C·-algebra Bo generated by
x·x and e. For since Bo is commutative, 1.1.1 implies that the spectrum
(relative to Bo) of the self-adjoint element x·x is real, and in particular this
relative spectrum is its own boundary, considered as a subset of the complex
plane. By the spectral permanence theorem ([23], p. 33), the latter coincides
with SPA (x· x). Because 0 ¢ SPA(X·X), we conclude that x·x is invertible in Bo.
These remarks show in particular that it is unambiguous to speak of the
spectrum of an operator T on a Hilbert space ./(~ so long as it is taken
relative to a C* -algebra. Thus, the spectrum of T in the traditional sense
(i.e., relative to ! (.In is the same as the spectrum of T relative to the
subalgebra c*(T). They also show that the spectrum of a self-adjoint element
of an arbitrary C* -algebra (commutative or not) is always real.

3
1. Fundamentals

We can now deduce the functional calculus for normal elements of C*-
algebras. Fix such an element x in a CO-algebra with identity, and let B be
the CO-algebra generated by x and e. Define a map of B into C as follows:
w -+ w(x). This is continuous and 1-1, thus since B is compact it is a
homeomorphism of B onto its range. By the preceding discussion the range
of this map is SPA.(X) = sp(x). So this map induces, by composition, an
isometric *-isomorphism of C(sp(x» onto B. It is customary to write the
image of a function f E C(sp(x» under this isomorphism as f(x). Note that
the formula suggested by this notation reduces to the expected thing when f
is a polynomial in Cand~; for example, if f(C) = C2~ then f(x) = x 2x*. This
process of "applying" continuous functions on sp(x) to x is called the func-
tional calculus.
In particular, when T is a normal operator on a Hilbert space we have
defined expressions of the form f(T), f E C(sp(T». In this concrete setting
one can even extend the functional calculus to arbitrary bounded (or even
unbounded) Borel functions defined on sp(T), but we shall have no particular
need for that in this book. It is now a simple matter to prove:

Theorem 1.1.2. Let 8 and T be normal operators. Then 8 and T are algebra-
ically equivalent if, and only if, they have the same spectrum.
PROOF. Assume first that sp(8) = sp(T). Then by the above we have
Ilf(8)1I = sup{lf(z)l:z E sp(8)} = IIf(T)II, for every continuous function f
on sp(8). This shows that the map 4> :f(8) -+ f(T), f E C(sp(8», is an iso-
metric *-isomorphism of C*(8) on C*(T) which carries 8 to T. Conversely,
if such a 4> exists, then the spectrum of 8 relative to C*(8) must equal the
spectrum of 4>(8) = T relative to C*(T). By the preceding remarks, this
implies sp(8) = sp(T). 0

EXERCISES

1.I.A. Let e be an element of a CO-algebra which satisfies ex = x for every x E A. Show

that e is a unit, e = e·, and lIell = 1.
1.1.B. Let A be a Banach algebra having an involution x ..... x· which satisfies IIxll2 :e;;
IIx·xll for every x. Show that A is a C·-algebra.
1.I.C. (Mapping theorem.) Let x be a self-adjoint element of a C·-algebra with unit
and let f E C(sp(x)). Show that the spectrum of f(x) is f(sp(x)).
I.I.D. Let A be the algebra of all continuous complex-valued functions, defined on the
closed disc D = {Izl :e;; 1} in the complex plane, which are analytic in the
interior of D.
a. Show that A is a commutative Banach algebra with unit, relative to the
norm IIfll = sUPI_I" 1 If(z)l·
b. Show thatf*(z) = Jfz'j defines an isometric involution in A.
c. Show that not every complex homomorphism w of A satisfies wlf') =
ro(f).

4
 1.2. Two Density Theorems

1.I.E. Let A be a C·-algebra without unit. Show that, for every x in A:

Ilxll = sup
IIYII.; I
IIxYII·
1.I.F. Let Sand T be normal operators on Hilbert spaces Jf and %. Show that C*(S)
is .-isomorphic to C·(T) iff speS) is homeomorphic to sp(T).
1.1.G. Let f:1R -+ C be a continuous function and let A be a C·-algebra with unit.
Show that the mapping x f-+ f(x) is a continuous function from {x E A:x = x·}
into A.
1.1.H. (Exercise on adjoining a unit.) Let A be a C*-algebra without unit, and for each x
in A let Lx be the linear operator on A defined by Y f-+ xy. Let B be the set of all
operators on A of the form A.l + Lx, A. E C, X E A.
Show that B is a C*-algebra with unit relative to the operator norm and the
involution (A.l + Lx)" = Al + Lx., and that x f-+ Lx is an isometric .-isomor-
phism of A onto a closed ideal in B of codimension 1. [Hint: use 1.l.B.]
1.1.1. Discuss briefly how the functional calculus (for self-adjoint elements) must be
modified for C·-algebras with no unit. In particular, explain why sin x makes
sense for every self-adjoint element x but cos x does not. [Hint: use 1.1.H to
define the spectrum of an element in a non-unital C·-algebra.]

1.2. Two Density Theorems

There are two technical results which are extremely useful in dealing with
*-algebras of operators. We will discuss these theorems in this section and
draw out a few applications.
The null space of a set Y s;;; 'p(.ne') of operators is the closed subspace of
all vectors e E .ne' such that Se = 0 for all S E Y. The commutant of Y (written
Y') is the family of operators which commute with each element of Y. Note
that Y' is always closed under the algebraic operations, contains the identity
operator, and is closed in the weak operator topology. Moreover, if Y is
self-adjoint, that is Y = Y* is closed under the *-operation, then so is Y'.
Now it is evident that Y is always contained in Y", but even when Y is a
weakly closed algebra containing the identity the inclusion may be proper.
According to the following celebrated theorem of von Neumann, however,
one has Y = Y" if in addition Y is self-adjoint.

Theorem 1.2.1 Double commutant theorem. Let d be a self-adjoint algebra

of operators which has trivial null space. Then d is dense in d" in both
the strong and the weak operator topologies.

PROOF. Let d wand d. denote the weak and strong closures of d. Then
clearly d. s;;; d w s;;; d", and it suffices to show that each operator TEd"
can be strongly approximated by operators in d; that is, for every 6 > 0,
every n = 1,2, ... , and every choice of n vectors eb e2, ... ,en
E.ne', there
is an operator SEd such that L~=l IITek - Sekl1 2 < 6 2 •

5
1. Fundamentals

Consider first the case n = 1, and let P be the projection onto the closed
subspace [d~l]. Note first that P commutes with d. Indeed the range of
P is invariant under d; since d = dO, so is the range of pl- = I - P, and
this implies P Ed'. Next observe that ~1 E [d ~1]' or equivalently, Pl-~l = O.
For if SEd then SPl-~l = Pl-S~l = 0 (because S~l E [d~l] and pl- is zero
on [d~l])' Since d has trivial null space we conclude Pl-~l = O. Finally,
since T must commute with P E d' it must leave the range of P invariant,
and thus T~l E range P = [d~l]. This means we can find SEd such that
IIT~l - S~tli < e, as required.
Now the case of general n ~ 2 is reduced to the above by the following
device. Fix n, and let Jf n = Jf E9 ... E9 Jf be the direct sum of n copies of
the underlying Hilbert space Jf. Choose ~1' ... , ~n E Jf and define 11 E Jf n
by 11 = ~ 1 E9 ~2 E9 ... E9 ~n' Let d n £ !l'(Jf n) be the "-algebra of all opera-
tors of the form {S E9 S E9 ... E9 S:S Ed}. Thus each element of d n can
be expressed as a diagonal n x n operator matrix

SEd. The reader can see by a straightforward calculation that an n x n

operator matrix (T i ), Tij E !l'(Jf), commutes with d n iff each entry Tij
belongs to d'. This gives a representation for d~ as operator matrices, and
now a similar calculation shows that (Tij) commutes with d~ iff (Tij) has
the form

with TEd". Thus we have a representation for d~. Now choose TEd"
and let Tn = T E9 T E9 ... E9 T. Then Tn E d~ so that the argument already
given shows that T n11 E [dnl1], thus we can find SEd such that Snl1 is within
e of Tnl1 in the norm of Jf n. In other words, Lk=l IIT~k - S~kW < e2 , as
required. D

By definition, a von Neumann algebra is a self-adjoint subalgebra fJl of

!l'(Jf) which contains the identity and is closed in the weak operator
topology. Note that 1.2.1. asserts that such an ~ satisfies ~ = ~", and this
gives a convenient criterion for an operator T to belong to ~: one simply
checks to see if T commutes with fJl'. As an illustration ofthis, let us consider
I
the polar decomposition. That is, let T E !l'(Jf), and let TI denote the
positive square root ofthe positive operator T" T (via the functional calculus).
Then ITIE C*(T*T), and in particular ITI belongs to the von Neumann

6
 1.2. Two Density Theorems

algebra generated by T. We want to define a certain operator V such that

e
VITI = T. Note first that for all = .Yt' we have IIITlell 2 = (ITle, ITle) =
(ITI 2 e, e) = (T*Te, e) = (Te, Te) = II Tell 2 • Therefore the map V:ITle -+ Te,
e E .Yt', extends uniquely to a linear isometry of the closed range of ITI onto
the closed range of T. Extend V to a bounded operator on .Yt' by putting
V = 0 on the orthogonal complement of ITI.Yt'. Then V is a partial isometry
(i.e., V*V is a projection) whose initial space is [ITI.Yt'] and which satisfies
ViTI = T. It is easy to see that these properties determine V uniquely, and
the above formula relating V and ITI to T is called the polar decomposition
of T. Now we want to show that V belongs to the von Neumann algebra
generated by T. By 1.2.1, it suffices to show that V commutes with every
operator Z which commutes with both T and T*. Now in particular Z com-
mutes with the self-adjoint operator ITI, and therefore Z leaves both ITI.Yt'
and (ITI.Yt').L invariant. In particular Z leaves the null space of V( = (I TI.Yt').L)
invariant and so ZV = VZ = 0 on the null space of V. Thus it suffices to
e
show that ZV = VZ on every vector of the form ITle, E.Yt'. But ZVITle =
ZTe = TZe, while VZITle = VITIZe = TZe, and we are done. This
proves the following
Corollary. Let T = Vi TI be the polar decomposition of an operator T E !l'(.Yt').
Then both factors V and ITI belong to the von Neumann algebra generated
by T.
The following density theorem is a special case of a theorem of Kaplansky
[16]. For a set of operators [/ we will write ball [/ for the closed unit ball
in [/, ball [/ = {S E [/: IISII ~ 1}.

Theorem 1.2.2. Let .511 be a self-adjoint algebra of operators and let d. be

the closure of .511 in the strong operator topology. Then every self-adjoint
element in ball d. can be strongly approximated by self-adjoint elements
in ball d.
PROOF. Note first that every self-adjoint element in the unit ball of the norm
closure ofd can be norm-approximated by self-adjoint elements in ball d.
Thus we can assumed is norm closed.
Now since the *-operation is not strongly continuous, we cannot im-
mediately assert that the strong closure of the convex set [/ of self-adjoint
elements ofd contains {T E d.: T = T*}. But its weak closure does (because
if a net Sft converges to T = T* strongly, then the real parts of Sft converge
weakly to T), and moreover since the weak and strong operator topologies
have the same continuous linear functionals (Exercise 1.2.E) they must also
have the same closed convex sets. Thus we see in this way that the strong
closure of [/ contains the self-adjoint elements of d •.
Now consider the continuous functions f:1R -+ [ -1, + 1] and g:[ -1,
+1] -+ IR defined by f(x) = 2x(1 + X 2 )-l and g(y) = y(1 + ,J1 _ y2)-1.
Then we have f 0 g(y) = y, for all y E [ -1, + 1], and clearly If(x) I ~ 1 for

7
1. Fundamentals

all x E IR. We claim that the map S ~ f(S) is strongly continuous on the set
of all self-adjoint operators on Yf. Granting that for a moment, note that
1.2.2 follows. For if T = TO E d sis such that IITII ~ 1, then So = g(T) is
a self-adjoint element of .91., so that by the preceding paragraph there is a
net Sn of self-adjoint elements ofd which converges strongly to So. Hence
f(Sn) --+ f(So) strongly. Now f(Sn) is self-adjoint, belongs to .91 (because .91
is norm closed), and has norm ~ 1 since If I ~ 1 on IR. On the other hand,
f g(y) = y on [ -1,1] implies f(So) = f g(T) = T, because IITII ~ 1,
0 0

and this proves T is the strong limit of self-adjoint elements of ball d.

Finally, the fact that f(S) = 2S(l + S2) - 1 is strongly continuous follows
after a moments reflection upon the operator identity
2[j(S) - f(So)] = 4(1 + S2)-1(S - So)(l + S~)-l - f(S)(S - So)f(So),
considering S tending strongly to So o
Kaplansky also proved that ball .91 is strongly dense in ball d s • That is
not obvious from what we have said, but a simple trick using 2 x 2 operator
matrices allows one to deduce that from 1.2.2 (Exercise 1.2.D).

Corollary. Let .91 be a self-adjoint algebra of operators on a separable Hilbert

space Yf. Then for every operator T in the strong closure of .91, there is a
sequence Tn E .91 such that Tn --+ T in the strong operator topology.
PROOF. We can assume IITII ~ 1, and since we can argue separately with
the real and imaginary parts of T, we can assume T = TO. Let eb
e2,' .. be
a countable dense set in Yf. By 1.2.2, for each n ~ 1, we can find a self-adjoint
element Tn in .91 such that IITnll ~ 1 and IITnek - Tekll < lin for k =
1,2, ... , n. Thus Tn --+ T strongly on the dense set gb e2,' .. } of Yf, and
since IITnll ~ 1, the corollary follows. 0

This corollary shows that in the separable case, the strong closure of a
CO-algebra of operators can be achieved by adjoining to the algebra all
limits of its strongly convergent sequences.
A CO-algebra is separable if it has countable norm-dense subset. A sepa-
rable CO-algebra is obviously countably generated (a countable dense set
clearly generates), and the reader can easily verify the converse: every
countably generated CO-algebra is separable. We conclude this section by
pointing out a useful relation between separably-acting von Neumann
algebras and separable CO-algebras.
Let Yf be a Hilbert space. Then it is well known that the closed unit ball
in 5l'(Yf) is compact in the relative weak operator topology ([7], p. 34).
Moreover, note that if Yf is separable then ball 5l'(Yf) is a compact metric
space. Indeed, if Ub U2, ... is a countable dense set of unit vectors in Yf then
the function
L r;- jl(Su; -
00

d(S, T) = Tu;, Uj)1

;,j= 1

8
 1.2. Two Density Theorems

defines a metric on ball !l'(ff) x ball !l'(ff) which, as is easily seen, gives
rise to the weak operator topology on ball !l'(ff). Thus we see in particular
that ball !l'(ff) is a separable complete metric space.
Now if fJI/ is any von Neumann algebra acting on a separable Hilbert
space then ball fJI/ is a weakly closed subset of ball !l'(ff) and therefore has
a countable weakly dense subset T 10 T 2, •.•• Thus the CO-algebra generated
by {Tl' T 2 , • •• } is a separable CO-algebra contained in fJI/ whose weak
closure coincides with fJI/, and we deduce the following result.

Proposition 1.2.3. For every von Neumann algebra fJI/ acting on a separable
Hilbert space there is a separable CO-algebra d c fJI/ which is weakly
dense in fJI/.

EXERCISES

1.2.A. An operator A E 9'(Jt") is said to be bounded below if there is an e > 0 such that
IIAxll ~ ellxll for every x E Jt". Prove that for such an operator A, C*(A) contains
both factors of the polar decomposition of A.
1.2.B. (Exercise on the polar decomposition.) Let T E 9'(Jt") have the polar decom-
position T = UITI.
a. Show that if R is a positive operator on Jt" and V is a partial isometry
whose initial space is [RJt"], and which satisfy T = VR, then V = U and
R = In
b. Show that U maps the subspace ker TJ. onto [TJt"].
I.2.C. Let f:1t be a von Neumann algebra acting on Jt", let..Hbe a subspace of Jt" whose
projection P belongs to f:1t, and let T E f:1t.
a. Show that the projection Q on [T..H] belongs to f:1t.
b. Show that there is a partial isometry U E f:1t satisfying U· U = P and
UU· = Q. (This shows that, in any von Neumann algebra f:1t, the partial iso-
metries do as good a job of moving subspaces around as arbitrary operators
in f:1t.)
1.2.D. (Exercise on Kaplansky's density theorem.) Let f:1t be a von Neumann algebra
on Jt".
a. Show that the set of all operators on Jt" E9 Jt" which admit a 2 x 2
operator matrix representation of the form

with A, B, C, D in f:1t, is a von Neumann algebra on Jt" EB Jt".

b. Let A be a ·-algebra of operators on Jt", and let T be an operator in the
unit ball of the strong closure of d. Use 1.2.2 to show that T can be strongly
approximated by operators in the unit ball of d. [Hint: consider the operator

(~. ~)
on Jt" EB Jt".]

9
1. Fundamentals

1.2.E. (Exercise on the weak and strong operator topologies.) Let Jff be a Hilbert
space and let f be a linear functional on Sf(Jff) which is continuous in the
strong operator topology.
a. Show that there exist vectors ~ 1, ... , ~. E Jff such that
If(T)1 ~ (1IT~1112 + ... + IIn.112)1/2, T E Sf(Jff).
b. With ~ I> ••• , ~. as in part a, show that there exist vectors '11> ... , '1. in Jff
such that f has a representation
f(T) = (T~l' '1d + ... + (T~., '1.),
T E Sf(Jff). [Hint: apply the Riesz lemma to a certain linear functional on the
Hilbert sum of n copies of Jff.]
c. Deduce that the weak and strong operator topologies have the same
continuous linear functionals and the same closed convex sets.

1.3. Ideals, Quotients, and Representations

In this section we will discuss a few basic properties of C -algebras and
introduce some terminology. By an ideal in a CO-algebra we will always mean
a closed two sided ideal. It is sometimes useful to consider left or right ideals,
but we shall not have to do so here.
Many C-algebras do not have identities. This is particularly true of ideals
in a given C-algebra, considered as CO-algebras in their own right. Very
often this lack is merely an annoyance, and the difficulty it presents can be
circumvented by simply adjoining an identity. For instance, the spectrum
of an element of a CO-algebra is a concept which obviously requires a unit.
The most direct way of defining the spectrum of an element x of a non-unital
C-algebra A is to consider x to be an element of the CO-algebra At obtained
from A by adjoining a unit, where sp(x) has an obvious meaning.
But frequently, and especially when dealing with ideals, it is necessary
to make use of a more powerful device. An approximate identity (or approxi-
mate unit) for a Banach algebra A is a net {e ..} of elements of A satisfying
(i) Ile;.11 = I, for every A.
(ii) lim;. IIxe;. - xII = lim;. IIe;.x - xII = 0, for every x E A.
When A is a C-algebra, one usually makes additional requirements of the
net {e;.} (see Section 1.8). It is a basic property of CO-algebras, as opposed to
more general Banach algebras, that approximate identities always exist. We
will eventually prove that assertion (cf. 1.8.2); but for our purposes in this
section, all we shall require is a simple result which implies that one-sided
approximate units exist "locally."

Proposition 1.3.1. Let A be a CO-algebra and let J be an ideal in A. Then

for every x in J there is a sequence et, e2, ... of self-adjoint elements of
J satisfying
(i) sp(en ) ~ [0,1] for every n;
(ii) limllxen - xII = O.

10
 1.3 Ideals, Quotients, and Representations

PROOF. Consider first the case where A has an identity e and x = x*. Define
en E A (via the functional calculus) by
en = nx 2(e + nx2)-l.
The function fn:1R --+ IR defined by f,.(t) = nt 2(1 + nt 2)-l vanishes at the
orgin, and is therefore uniformly approximable on sp(x) by polynomials of
the form a l t + a2t2 + ... + aktk. It follows that en belongs to the closed
linear span of x, x 2 , x 3 , ••• , and in particular, en E J.
en is clearly self-adjoint, and since the range of each f,. is contained in
the unit interval, it follows that sp(en) ~ [0, 1].
For (ii), notice that the spectrum of e - en is also contained in the unit
interval, and so lie - enll :::; 1 because e - en is self-adjoint. Moreover, the
nonnegative function
t 2(1 - f,.(t)) = t 2(1 + nt2 )-l
is bounded by lin, so that
Ilxen - xl1 2 = Ilx(e - en)112 = II(e - en )x 2 (e - en)11
:::; Ilx2(e - en)11 :::; lin.
Therefore limllxen - xii = O.
If x "# x*, then we may apply the above to x*x, obtaining en such that
Ilx*xen - x*xll --+ O. It follows that
Ilxen - xl1 2 = Ilx(en - e)11 2
= II(en - e)x*x(en - e)11 2 :::; Ilx*x(en - e)11 --+ 0
as n --+ 00, as required.
The case where A does not contain an identity is easily dealt with by
adjoining an identity, and is left for the reader. 0

Corollary 1. Every ideal in a C*-algebra is self-adjoint (i.e., is closed under

the *-operation).

PROOF. Let J be an ideal in a C*-algebra A, and let x be an element of J.

By 1.3.1 we can find a sequence en = e; in J so that x = limn xen. By taking
adjoints we have x* = lim enx*, and clearly the right side of that expression
belongs to J. 0

Now if J is an ideal in a C*-algebra A, then the quotient AjJ becomes a

Banach algebra in the usual way; for example, the norm of the coset repre-
sentative x ofx is defined as Ilxll = inf{ Ilx + zll; z E J}. Because of Corollary
1 above, we may introduce a natural involution in AjJ by taking x* to be
the coset representative of x*. It is significant that the norm in AjJ is a
C*-norm relative to this involution.

11
1. Fundamentals

Corollary 2. AIJ is a C* -algebra.

PROOF. By Exercise 1.1.B, it suffices to show that IIxl12 ~ IIx*xll, for every
x in A.
For that, fix x, and let E denote the set of all self-adjoint elements U of J
satisfying sp(u) ~ [0,1]' We claim first that
Ilxll = inf Ilx - xull·
ueE

Indeed, the inequality ~ is obvious because xu E J for each U E E, and it

suffices to show that for each k E J, there is a sequence Un E E satisfying
Ilx + kll ~ infn Ilx - xunll· Fix k, and choose Un for k as Proposition 1.3.1.
We have already seen that lie - unll ~ 1, so that
Ilx + kll ~ lim inf lI(x + k)(e - un)1I
n

= lim inf IIx(e - un) + k(e - un)11

n

= lim inf IIx - xunll ~ inf IIx - xunll,

n n

because k(e - un) = k - kUn tends to zero as n -+ 00.

To complete the proof, we apply the preceding formula twice to obtain
IIxll2 = inf Ilx(e - u)11 2 = inf II(e - u)x*x(e - u)1I
ueE u

~ inf Ilx*x(e - u)11 = Ilx*xll. 0

u

Now let A and B be C*-algebras and let n be a *-homomorphism of A into

B, that is, n preserves the algebraic operations and n(x*) = n(x)*. Note that
we do not assume that n is bounded, but nevertheless that turns out to be
true. To see why, consider first the case where both A and B have identities
and n maps eli to eB' Then clearly n must map invertible elements of A to
invertible elements of B, and this implies that n must shrink spectra. More-
over, since the norm of a self-adjoint element of a C*-algebra must equal its
spectral radius, we have
IIn(x*x)11 = r(n(x*x» ~ r(x*x) = Ilx*xll.

Since the left side of this inequality is IIn(x)*n(x)11 = IIn(x)11 2 and the right
side is Ilxll2, we conclude that n has norm at most 1.
Now suppose, in addition to the above hypotheses, that n has trivial
kernel. Then we claim that n is isometric. By reasoning as above, this will
follow if we show that, for every self-adjoint element z of A, n(z) and z have
the same spectrum. We know that sp n(z) ~ sp(z), so if these sets are different
then one can find a continuous function J:sp(z) -+ IR such that J :F 0 but
J = 0 on sp n(z). Now if J is a polynomial then we have J(n(z» = n(J(z».
In general, J is the norm limit on sp(z) of a sequence of polynomials (by the
Weierstrass theorem), and so by 1.1.1 we conclude that J(z) and J(n(z» are

12
 1.3 Ideals, Quotients, and Representations

the corresponding limits of polynomials. This proves the formula f(n(z)) =

n(f(z)) for arbitrary continuous f. But f(n(z)) is 0 because f = 0 on sp n(z),
so by the formula we have n(f(z)) = O. Since n is assumed injective we
conclude that f(z) = 0, and by 1.1.1 it follows that f = 0 on sp(z), a con-
tradiction.
Note finally that the preceding implies that the range of n is closed even
when n is not injective. For if we let J be the kernel of n, then n lifts in the
obvious way to an injective *-homomorphism ir. of AjJ into the range of n,
and of course ir. also preserves identities. Thus ir. is isometric by the above,
and in particular its range is closed.
To summarize, we have proved the following theorem, at least in the
presence of certain assumptions about units.

Theorem 1.3.2. Let A and B be C* -algebras and let n be a *-homomorphism

of A into B. Then n is continuous and n(A) is a C*-subalgebra of B. n induces
an isometric *-isomorphism of the quotient A/ker n onto n(A).

PROOF. We shall merely indicate how the general case can be reduced to
the above situation where both A and B have units and n(eA) = eB'
Assume first that A has a unit eA- By passing from B to the closure of
n(A) if necessary, we may assume that n(A) is dense in B. Since n(eA) is a
unit for n(A), it is therefore a unit for B, and we are now in the case already
discussed.
So assume A has no unit, and let At :2 A be the C*-algebra obtained
from A by adjoining a unit e. By adjoining a unit to B if necessary, we may
also assume that B has a unit eB' Define a map it:A t --+ B by
it(Ae + x) = AeB + n(x), X E A, A E C.

It is a simple matter to check that it is a *-homormorphism, and it is clearly

an extension ofn to At. Moreover, it(e) = eB, and thus we are again reduced
to the preceding situation. 0

Corollary. Let A be a C* -algebra and let Ixl be another Banach algebra norm
on A satisfying Ix*xl = Ix12, x E A. Then Ixl = Ilxll for every x E A.

PROOF. Let B = A, regarded as a C*-algebra in the norm Ixl. Then the

identity map is an injective *-homomorphism of A on B. Now apply 1.3.2. 0

The corollary shows that there is at most one way of making a complex
algebra with involution into a C*-algebra.
We come now to the central concept of this book.

Definition 1.3.3. A representation of a C*-algebra A is a *-homomorphism

of A into the C*-algebra .P(~) of all bounded operators on some Hilbert
space ~.

13
1. Fundamentals

It is customary to refer to n:A ~ 2(.no) as a representation of A on .no.

n is called nondegenerate if the C·-algebra of operators n(A) has trivial
null space. We leave it for the reader to show that, since n(A) is self-adjoint,
this is equivalent to the assertion that the closed linear span [n( A ).no] of
e
all vectors of the form n(x)e, x E A, E .no, is all of .no.
An invariant subspace vi{ for the C*-algebra n(A) is called a cyclic sub-
e
space if it contains a vector such that the vectors of the form n(x)e, x E A,
are dense in vI{: this is written vi{ = [n(A )e]. n is called a cyclic representa-
tion if .no itself is a cyclic subspace for n. It is clear from the preceding
paragraph that a cyclic representation is nondegenerate. More generally, a
representation of A on .no is non degenerate if, and only if, .no can be de-
composed into a mutually orthogonal family of cyclic subspaces (Exercise
l.3.F).
Let nand u be two representations of A, perhaps acting on different
spaces .no and %. nand u are said to be equivalent ifthere is a unitary operator
U:.no ~ % such that u(x) = Un(x)U· for all x in A; this relation is written
n ,.... u. Equivalent representations are indistinguishable in the sense that any
geometric property of one must also be shared by the other, and it is correct
to think of the unitary operator U as representing nothing more than a
change of "coordinates."
Finally, a nonzero representation n of A is called irreducible if n(A) is an
irreducible operator algebra, i.e., commutes with no nontrivial (self-adjoint)
projections. Because n(A) is a C·-algebra, this is the same as saying n(A)
has no nontrivial closed invariant subspaces (Exercise 1.3.D).
Now if A is commutative then so is every image of A under a repre-
sentation, and it is a simple application of the spectral theorem to see that
comm utative C* -algebras of operators on Hilbert spaces of dimension greater
than 1 cannot be irreducible (Exercise 1.3.E). So the only irreducible repre-
sentations of A are those of the form n(x) = w(x)I, where I is the identity
operator on a one-dimensional space and w is a nonzero homomorphism
of A into the complex numbers. This shows that we can identify the
equivalence classes of irreducible representations of a commutative C·-
algebra in a bijective way with its set of nonzero complex homomorphisms.
Moreover, it suggests that one should view an irreducible representation
(more precisely, an equivalence class of them) of a noncommutative C*-
algebra as filling a role similar to that of complex homomorphisms. This
analogy will be pursued to considerable lengths throughout the book. We
will find that while the generalization achieves some remarkable successes
within the class of GCR algebras (defined in Section 1.5), it also leads to
unexpected and profound difficulties in all other cases.
Returning now to the present discussion, we want to consider a useful
connection between representations and ideals. In general, a representation
of a C*-subalgebra of A on a Hilbert space .no cannot be extended to a
representation of A on .no. But if the subalgebra is an ideal then it can. To
see why, let J be an ideal in A and let n be a nondegenerate representation

14
 1.3. Ideals, Quotients, and Representations

of J on a Hilbert space £l. We claim first that for each x in A, there is a

unique bounded linear operator n(x) on £l satisfying n(x)n(y) = n(xy) for
every y E J. Indeed, uniqueness is clear from the fact that vectors of the
form n(y)~, y E J, ~ E £l, span £l. For existence, consider first the case
where n is cyclic, and let ~o E £l be such that [n(J)~oJ = £l. We claim
that IIn(xY)~oll ~ IIxll'lln(Y)~oll, for each y E J. To see this fix y, and choose
a sequence en = e~ E J with sp(en ) ~ [0, IJ and y*e n ~ y* (by 1.3.1). By
taking adjoints we see that eny ~ y, so that

Iln(xY)~oll = lim II n (xeny)eoll

n

= lim IIn(xen)n(Y)~oll ~ sup Ilxenll'lln(Y)~oll ~ Ilxll'lln(Y)~oll,

n n

as asserted. It follows that the map n(y)~o ~ n(xy)~o(Y E J) extends uniquely

to an operator n(x) on [n(I)~oJ = £l having norm at most Ilxll. The reader
can easily check that the required relation n(x)n(y) = n(xy) holds on all
vectors of the form n(z)~o, z E J, so it holds throughout £l.
In the general (noncyclic) case, one may apply Exercise 1.3.F to express
£l as an orthogonal sum of cyclic subspaces, define n(x) as above on each
cyclic summand and then add up the pieces in the obvious way to obtain
an operator on all of £l.
Thus we have established that there is a unique mapping n of A into
2(£l) which satisfies n(x)n(y) = n(xy) for x E A, y E J. This formula itself
implies that npreserves the algebraic operations, the involution, and restricts
to n on J (these routine verifications are left for the reader). Finally, if a
is any other representation of A on £l such that alJ = n, then for every
x E A, y E J we have
a(x)n(y) = a(x)a(y) = a(xy) = n(xy),
so that a = n by the uniqueness assertion of the preceding paragraph.
If n is a degenerate representation of J on £l, then we can make n non-
degenerate by passing from £l to the subspace [n( J)£l]. So we can still
obtain a unique extension n of n to A such that n(A) acts on [n(J)£l].
Now suppose, on the other hand, that we start with a representation n
of the full algebra A on £l. Choose any ideal J in A and let £lJ = [n(J)£l].
Since J is an ideal we have n(A)£lJ ~ £lJ> and thus £l = £lJ EB £ly gives
a decomposition of £l into reducing subspaces for n(A). Define representa-
tions nJ and aJ of A on £lJ and £ly respectively by nAx) = n(x)IJI"J and
aAx) = n(x)IJI";' Then in an obvious sense we have a decomposition n(x) =
nAx) EB aAx) of n, where on the one hand nJ is determined uniquely by
the action of n on the ideal J in the sense of the preceding paragraph, and
where aJ annihilates J (recall that [n(J)£lJ is the null space of the C*-
algebra n(J» and can therefore be regarded as a representation of the quo-
tient AjJ. These remarks show that once we know all of the representa-
tions of both an ideal J and its quotient AjJ, then we can reconstruct the

15
1. Fundamentals

representations of A. This procedure is particularly useful when dealing with

irreducible representations.

Theorem 1.3.4. Let n be an irreducible representation of A and let J be an

ideal in A such that n(J) # O. Then nl J is an irreducible representation
of J. Every irreducible representation of J extends uniquely to an irreducible
representation of A, and if two such representations of J are equivalent
then so are their extensions.
PROOF. Let n:A -+ .!e(JIl') be irreducible, such that n(J) # o. We claim that
n(J) is irreducible. For that it suffices to show that [n(J)~] = JIl' for every
~ # 0 in JIl'. Since [n(1)~] is invariant under the irreducible C'-algebra
n(A), [n(J)~] must be JIl' or {O}. If it is {O} then ~ belongs to the null space
of the C'-algebra n(J) and therefore ~ 1. [n(J)JIl']. Thus [n(J)JIl'] # JIl'is
n(A)-invariant, and therefore [n(J)JIl'] = O. This means n(J) = 0, a
contradiction.
If, conversely, n is an irreducible representation of J, then its extension
if must be irreducible because if(A) contains the irreducible subalgebra n(1).
Finally, if nand (J are irreducible representations of J and U is a unitary
operator between their respective spaces such that (J = UnU', then note
by uniqueness the extensions if and (j must also satisfy (j = UifU·. That
completes the proof. 0

EXERCISES

1.3.A. Let A be a CO-algebra, let J be an ideal in A, and let B be a sub CO-algebra of

A. Show that B + J is a sub CO-algebra of A and that B + JjJ and BIB n J
are canonically ,-isomorphic.
1.3.B. Let J be an ideal in a CO-algebra A and let x be a self-adjoint element of A.
Show that there is an element k E J satisfying

Ilx + kll = inf

tE)
Ilx + lll·
[Hint: use the functional calculus to "truncate" x.]
I.3.e. (Hahn decomposition.) Let x be a self-adjoint element of a CO-algebra A. Show
that there exist self-adjoint elements Yt> Y2 in A satisfying SP(Yi) ~ 0,
Y1YZ = 0, and x = Yl - Yz·
1.3.D. Let 1t be a nonzero representation of a CO-algebra A. Show that 1t is irreducible
iff the operator algebra 1t(A) has no nontrivial closed invariant subspaces.
1.3.E. Show that if 1t is an irreducible representation of a commutative CO-algebra on
a nonzero Hilbert space .Yf, then .Yf is one-dimensional. [Hint: use the spectral
theorem.]
1.3.F. Let 1t be a representation ofa CO-algebra A on.Yf. Show that 1t is nondegenerate
iff .Yf can be decomposed into an orthogonal sum of cyclic invariant subspaces.

16
 1.4 C·-Algebras of Compact Operators

1.4. C*-algebras of Compact Operators

One usually regards commutative C·-algebras (equivalently, C-algebras
of normal operators) as being the easiest to deal with. However, for many
purposes, C·-algebras of compact operators are more tractable even though
they may present a high degree of noncommutativity. We will see in this
section, for example, that their representation theory can be completely
worked out by more or less elementary methods. That result foreshadows
the general theory developed in Chapter 4 for GCR algebras. At the end of
this section we prove a Wedderburn-type theorem which makes the structure
of such algebras quite transparent.
Let Yf be a Hilbert space. An operator T E 2(Yf) is compact if the image
of the unit ball of Yf under T has compact closure in the norm topology of
Yf. The set ~(Yf) of all compact operators on Yf is a (closed two-sided)
ideal in 2(Yf) (see Exercise 1.4.A) and is therefore a C -algebra in its own
right. Let us first recall one or two facts about compact operators. For
every compact operator T, the spectrum of T is countable and has no
nonzero accumulation point ([27], p. 233). Thus if T is self-adjoint and
sp(T) = {A.l> A2, ... }, Ai E IR, then Ai -+ 0 as i -+ 00 and the spectral theorem
for T takes the form T = I:,= 1 AnEn where the {En} are mutually orthogonal
spectral projections for T. If An # 0 then the characteristic function of {An}
can be uniformly approximated on sp(T) by polynomials of the form p(x) =
al x + a2x2 + ... + akxk, and therefore En can be approximated in the
operator norm by polynomials al T + a2 T2 + ... + akTk. In particular
each projection En is compact, and therefore has finite rank (for An # 0).
Since IIIk"=n AkEkll = SUPk;:'n IAkl -+ 0 as n -+ 00, it follows that the finite
sums Ik= 1 AkEk converge in norm to T, and in particular T is a norm limit
of self-adjoint operators having finite rank.
Now let d be a C-subalgebra of~(Yf), fixed throughout the remainder
of this section. By cutting down to an d-invariant subspace, if necessary,
we may assume that d has trivial null space, and thus [d Yf] = Yf. Note
that while d does not contain the identity if Yf is infinite-dimensional, the
preceding paragraph shows that it does contain many finite-rank projections,
which can be obtained as spectral projections of self-adjoint operators in d.
A projection E in d is called minimal if E¥-O and the only subprojections
of E in dare 0 and E.

Lemma 1.4.1. Let E be a nonzero projection in d. Then E is minfmal iff

EdE = {A.E:A E IC}. Every nonzero projection in d is finite-dimensional,
and is a finite sum of orthogonal minimal projections.
PROOF. If Ed E consists of scalar multiples of E, then E is minimal. Con-
versely, assume E is minimal. It suffices to show that ETE is a scalar multiple
of E, for every self-adjoint TEd. Considering the spectral formula for

17
1. Fundamentals

ETE we have ETE = Ln AnFn where the Fn are mutually orthogonal

spectral projections of ETE. Since ETE annihilates E1. JIt' so does each Fn,
and hence Fn ~ E. Thus each nonzero Fn must be E, and hence ETE = AE
has the required form.
It is plain that every projection in .s;I is finite-dimensional, by compactness,
and the last phrase follows from the usual sort of finite induction. 0

Theorem 1.4.2. If .s;I is irreducible then .s;I = ~(JIt').

PROOF. Note first that .s;I contains a projection ofrank 1. For if F is a non-
zero spectral projection of any self-adjoint operator in .s;I then by the lemma
F contains a minimal projection E, and it suffices to show that E has rank 1.
Choose ~, 11 E EJIt', with ~ =1= 0 and 11 1. ~. If T is any operator in .s;I then
ETE has the form AE, by 1.4.1, and thus (11, T~) = (11, ETE~) = (11, A~) =
1(11, ~) = 0, thus 11 1. (.s;I~] = JIt' and so 11 = 0, which proves EJIt' = [~].
Next observe that .s;I contains every projection F of rank 1. Indeed, such
an F has the form F~ = (~, f)f where f is a unit vector in JIt'; so if E is any
rank 1 projection in.s;l, say E~ = (~, e)e, then we can find a sequence Tn E .s;I
such that Tne -+ f and II Tnell = 1 for every n. Thus TnET: E.s;I, and we have
IITnET:~ - F~II = II(~, Tne)Tn e - (~,f)fll ~ 211~1I'IITne - fll,
for every ~ E JIt'. It follows that IITnET: - FII-+ 0, proving the assertion.
We may now conclude that .s;I contains every finite rank projection, there-
fore every self-adjoint compact operator (spectral theorem), and thus .s;I =
~JIt'~ 0

Corollary 1. ~(JIt') contains no ideals other than 0 and ~(JIt').

PROOF. Let.s;l =1= 0 be an ideal in ~(JIt'). Applying 1.3.4 to the identity rep-
resentation of ~(JIt') we see that .s;I is irreducible, and 1.4.2 shows that
.s;I = ~(JIt'). 0

Corollary 2. Let PA be an irreducible C*-algebra of operators on JIt' which

contains a nonzero compact operator. The PA contains ~(JIt').
PROOF. PA n ~(JIt') is a nonzero ideal in PA. Arguing as in Corollary 1 we see
that PA n ~(JIt') is irreducible, and by 1.4.2 PA n ~(JIt') = ~(JIt'), as as-
serted. 0

Proposition 1.4.3. Let E be a minimal projection in .s;{, let ~ be a unit vector

in EJIt', and let JIt' 0 = (.s;I ~]. Then .s;I1Jf'o is irreducible, and infact .s;I1Jf'o =
~(JIt' 0)'

PROOF. The map T -+ TIJf'o is a *-homomorphism of.s;l into ~(JIt' 0) whose

range is .s;I1Jf'o' By 1.3.2 .s;I1Jf'o is a C*-subalgebra of ~(JIt' 0), and by 1.4.2 the
entire conclusion will follow provided we show that .s;I1Jf'o is irreducible.
For that, choose any R E !l'(JIt' 0) which commutes with .s;I1Jf'o' We will
show that R is a scalar. By replacing R with R - (R~, ~)I we can assume

18
 1.4 CO-Algebras of Compact Operators

(Re, e) = 0, and in this case we claim R = O. Now if S, TEd then by 1.4.1.

ET*SE has the form AE for some A E C. Thus (RSe, Te) = (RSEe, TEe) =
(ET* RSEe, 0 = (RET*SEe, e) = A(REe, e) = A(Re, e) = O. R = 0 follows
beca use de is dense in Jt' 0. 0
Let n be a nondegenerate representation of d on a Hilbert space .%, and
let .%0 be a subspace of .% invariant under n(d). Then no(T) = n(T)I.% 0
defines a nondegenerate representation of d on .%0. Such a no is called a
subrepresentation of n (this is written no :;:;; n). If {n;} is any family of sub-
representations of n whose respective subspaces are mutually orthogonal,
then Lni has an obvious meaning (its value at TEd is Lni(T», and defines
another sub representation of n. Now 1.4.3, along with a simple argument,
implies that the identity representation of d is the sum of mutually orthog-
onal irreducible subrepresentations. The following result is somewhat
stronger.

Theorem 1.4.4. Let n be any nondegenerate representation of d. Then there

is an orthogonal family {ni} of irreducible subrepresentations of n such
that n = Li ni' and each ni is equivalent to a subrepresentation of the
identity representation of d.
PROOF. First note that there is a minimal projection E E d such that n(E) #-
O. To see that, choose T = T* E d such that n(T) #- O. By our initial remarks
about the spectral theorem there is a spectral projection F of T such that
n(F) #- O. The assertion now follows from the second part of 1.4.1.
Choose such an E. By 1.4.1 there is a linear functional f on d such that
ET E = f(T)E, for all TEd. Let.% be the space on which n acts, and choose
a unit vector '1(resp. e) in n(E).%(resp. EJt'). Then .%0 = [n(d)'1] defines
a sub representation no of n. We will show that no is equivalent to the ir-
reducible subrepresentation of the identity representation of d defined by
the subspace [de] (1.4.3). Indeed, for TEd we have
Iln(T)'111 2 = Iln(T)n(E)'111 2 = Iln(T E)'111 2
= (n(ET*TE)'1, '1) = f(T*T)(n(E)'1, '1)
= f(T*T) = (ET*TEe, e) = II T e11 2 .
This shows that map U: Te 1---+ n(T)'1 extends uniquely to a unitary map of
[de] onto [n(d)'1], and the formula UT = no(T)U is immediate from the
definition of n. That proves the assertion about no.
Thus we have proved that every nondegenerate representation of d
contains an irreducible nonzero sub representation equivalent to a sub-
representation of id. The proof of the theorem can now be completed by an
exhaustion argument (i.e., by Zorn, choose a maximal family {ni} of orthog-
onal subrepresentations of n each of which is equivalent to an irreducible
subrepresentation ofid, and note that Lni must be n by maximality). 0

Let n be a representation of d on .%, and let n be a positive cardinal. Let

.%' be the direct sum of n copies of .%. Then we can define a representation

19
1. Fundamentals

n· of d on .Yt' by setting (n . TC)(T) =

TC I al TC(T), the direct sum extended
over n copies of TC( T). n . TC is called a multiple of TC.

Corollary 1. Every representation of ~(£) is equivalent to a multiple of the

identity representation.
PROOF. Let TC be a representation of d on .Yt. Since the identity representa-
tion id of ~(£) is itself irreducible, it follows that TC has a decomposition
Li TCi into orthogonal sub representations each of which is equivalent to id.
Choosing, for each i, a unitary operator Vi from .Yt' to the range space of £i
such that id = V;TCiVi and letting n be the cardinal of {TCi}, we see that
V = Ir
Vi implements the equivalence n . id = V·TCV. 0

Corollary 2. Every irreducible representation of ~(£) is equivalent to the

identity representation.

Corollary 3. Let £ and .Yt be Hilbert spaces. Then every ·-isomorphism IX

of ~(£) (resp. 2(£)) onto ~(.Yt) (resp. 2(.Yt)) is implemented by a
unitary operator V; IX(T) = VTV·.
PROOF. Consider first the case IX:~(£) ~ ~(.Yt). Then IX is an irreducible
representation of ~(£), and the conclusion follows from Corollary 2.
Now suppose IX is a .-isomorphism of 2(£) on 2(.Yt) and let lXo :~(£) ~
2(.Yt) be the restriction of IX to ~(£). Since IX is an irreducible representation
and ~(£) is an ideal in 2(£), it follows by 1.3.4 that lXo is irreducible. By
Corollary 2 lXo has the form lXo(T) = VTV·, where V is a unitary operator
from £ to .Yt. Since P(T) = VTV·, T E 2(£), is an irreducible repre-
sentation of 2(£) extending lXo, we conclude from 1.3.4. that P = IX, as
required. 0

We now wish to restate 1.4.4 in a more systematic way. The set of all
irreducible sub representations of the identity representation of d is parti-
tioned by the unitary equivalence relation -. Let sJ denote the set of all
these equivalence classes. sJ is called the spectrum of d, and here .sJ is to be
regarded as a set with no additional structure. For each ( E .sJ choose, once
and for all, an element TC, E (. It will be convenient to write ((T) for the
operator TC,(T), TEd. Now if {TCd is a family of irreducible representations
of d such that each TCi is equivalent to (, then as we have seen in the proof of
Corollary 1, the direct sum I? TCi is equivalent to n . (where n is the cardinal
number of {TCd. Thus 1.4.4 can be paraphrased as follows: every nonde-
generate representation of d is equivalent to a representation of the form
Lal n(O . (, the sum extended over all classes (E sJ, where (H n(') is a
function from sJ into the class of all nonnegative cardinal numbers. Now it
can be shown that if , H m(O is another cardinal-valued function, then
Lal IE!)
all
d
'E sJ
m(O· , and

£;
n(O· , are equivalent if, and only if, m(O = n(O for
(while one can give an elementary proof in this case because
~(£) we shall not do so, because a more general theorem will be

20
 1.4. CO-Algebras of Compact Operators

proved later, in Chapters 2 and 4). In particular, the function n(' ) occurring
in the expression n = I$ n(O' ( is well-defined by n; n(O is called the
multiplicity of ( in n, and n itself is called the multiplicity function of n. Thus,
two representations of d are equivalent iff they have the same multiplicity
functions, and we have here an effective classification of the representations
ofd.
Let us apply this to single operators. Choose a compact operator T E C£(£),
and let d be the C*-algebra generated by T. For every ( E .sJ define T~ as
(T). T~ is clearly compact and, since it generates the irreducible CO-algebra
(d), it is also irreducible. Moreover, T~ and T~, are not unitarily equivalent
for ( #- C because any unitary operator U satisfying UT~U* = T~, would
implement an equivalence between the inequivalent representation ( and
(' of d. Applying the above decomposition to the identity representation
of d and evaluating everything at T, we obtain the decomposition T =
I$ n«()' T~, where m' T, denotes the direct sum of m copies of T~. Now
since each operator n(O . T~ is compact and nonzero, n(') must take on finite
positive values 1,2, .... Moreover, the compactness of I$ n(O . T~ implies
that lim~~oo /lTdl = 0 in the sense that for every e > 0 there is a finite subset
E s; .sJ such that /I Tdl ~ e for ( ¢ E. Conversely, given any family {Ti: i E I}
of mutually inequivalent irreducible compact operators such that limi-+oo
/lTdl = 0, and given any set {ni:i E I} of positive integers, T' = Ir ni' Ti
defines a compact operator. As it turns out, the uniqueness assertion of the
preceding paragraph takes the following form: T' is unitarily equivalent to
T = L:$ n(O . T~ iff there is a bijection i E [ ~ (i E .sJ such that n«(;) = ni
and T~i is unitarily equivalent to T i, for all i E [.
This gives the precise sense in which the theory of C*-algebras reduces
the problem of classifying general compact operators to the problem of
classifying irreducible compact operators. But that is as far as the self-
adjoint theory takes us; the classification problem for irreducible compact
operators is a separate problem requiring its own techniques [1,2, 5J.
Let us now examine the algebraic structure of a general C* -algebra of
compact operators. An abstract CO-algebra A is called elementary if A is
*-isomorphic to the CO-algebra C£(£) of all compact operators on some
Hilbert space £. Now for any family {An:n E I} of C*-algebras (the index
set [ is perhaps uncountable), define the direct sum IAn as the set of all
functions n E [ ...... an E An which satisfy the condition limn-+ 00 /lanll = 0 in
the sense described above. We can make IAn into a CO-algebra by giving
it the "pointwise" operations (for example, {an} + {b n} = {an + bn}) and
the norm I {an} II = supn lIanll· The direct product nAn has a similar definition
except that one takes for its elements all functions {an} satisfying lI{a n}!! =
supn lIanll < 00. Thus nAn is much larger than IAn, except in the case
when the index set is finite.

Theorem 1.4.5. Every CO-algebra of compact operators is *-isomorphic to a

direct sum of elementary C* -algebras.

21
1. Fundamentals

PROOF. Let.91 be a C*-subalgebra of ~(K). As before, choose a representa-

tive , for each equivalence class of irreducible subrepresentations of the
identity representation id of .91. For each " let Kc be the subspace of K
on which, acts. We wi!) show that .91 is *-isomorphic with L~(Kc)' the sum
extended over all , E .91.
Now by the preceding discussion we can write id = L(j) n(O' " where
n(O is a positive cardinal for every' E d. Thus T = L(j) n(O' ,(T) for each
TEd. Clearly II Til = sUPc Iln(o . ,(T)II = sUPc II,(T)II, and since T is com-
pact we must have limc .... oo ,(T)II = O. Thus the map which associates to
each TEd the element {Td ofL~(Kc) defined by Tc = ,(T) is an isometric
*-homomorphic imbedding of.91 in L~(Kc)' We have to show that it maps
onto L~(Kc).
Choose an element {Td E L~(Kc). For each " '(.91) is an irreducible
C*-subalgebra of~(Kc) and so ,(.91) = ~(Jf'c) by 1.4.2. Thus there is an op-
erator Sc Ed such that ,(Sc) = Tc. By 1.3.2 we can choose Sc so that IISdl is
arbitrarily close to II Tdl, and in particular we can arrange that limc .... oo IISdl =
O. Let Ec be the projection of Jf' onto [.9I'Jf'c]. Since [.9I'K c] is invariant
under both .91 and .91' it follows that Ec belongs to the center of .91". We
claim first, that ScEc E .91, and second, that Ec .1 Ec- if' i= ". Granting that,
it follows that S = LScEc is a well-defined element of.91 (because IISdl tends
to 0), and we have ,(S) = ,(ScEc) = ,(Sc) = Tc (because Ec' is orthogonal
to K c' the range of the subrepresentation " for " i= O. Thus the theorem
will be proved.
The first claim follows if we show that .9IT £;; .91 for every self-adjoint
TEd"; or equivalently, that T.9I £;; .91 (since .91 is self-adjoint). For such
a T, Kaplansky's density theorem provides a bounded net Tn Ed such that
Tn ~ T strongly. Thus IITnK - TKII ~ 0 for every finite rank operator K,
and since Tn - T is bounded this persists for arbitrary compact K. Because
.91 is norm-closed we conclude that T.9I £;; .91.
It remains to prove that Ec .1 Ec' if , i= ". By definition of Ec this is
equivalent to .9I'Jf'c .1 Kc' if " i= C. Assume, to the contrary, that there is
an element T' E .91' such that Fc' T' Fc i= 0, Fc and Fc' denoting the projections
on Jf'c and Jf'c" From the polar decomposition of the operator Fc.TFc Ed'
we obtain a nonzero partial isometry U Ed' whose initial space (resp. final
space) is contained in Jf'c (resp. Kd Since these spaces are invariant under
the respective irreducible C*-algebras '(.91) and ('(d) we see that V = UIKc
is a unitary map of Jf'c on Kc" Since V,(T) = ,'(T)V (because U Ed') we
have a contradiction of the fact that Cand " are inequivalent. D

1.5. CCR and GCR Algebras

In the preceding section we saw how one can obtain rather complete
information about noncommutative C*-algebras of compact operators. We
are now going to introduce a much broader class of C*-algebras. The ideas,
and most ofthe results ofthis section, originated in a paper of Kaplansky [15].

22
 1.5. CCR and GCR Algebras

Definition 1.5.1. A CCR algebra is a CO-algebra A such that, for every ir-
reducible representation n of A, n(A) consists of compact operators.

The acronym CCR is supposed to suggest the tortured phrase "completely

continuous representations." 1.4.4 implies that every C-algebra of compact
operators is CCR. At the other extreme, every commutative CO-algebra is
CCR since each of its irreducible representations is one-dimensional. More
generally, A is CCR if each of its irreducible representations is finite-
dimensional (it is not necessary that they have the same dimension, or that
the dimensions are even bounded).
We shall only need one or two results about CCR algebras. Note first
that for every irreducible representation n of a CCR algebra A on £, we
see by 1.4.2 that n(A) must in fact coincide with the algebra ~(£') of all
compact operators on £'. Since n induces a *-isomorphism of A/ker n on
~(£) and since ~(£') has no nontrivial closed ideals, we conclude that 0
is a maximal ideal in A/ker n. Equivalently, the kernel of an irreducible
representation of a CCR algebra A is a maximal ideal in A. The reader is
cautioned that this conclusion is false for more general C-algebras.

Proposition 1.5.2. Let A be a CCR algebra and let nand (J be irreducible

representations of A such that ker n s; ker (J. Then nand (J are equivalent.
PROOF. Suppose nand (J act on Hilbert spaces £' and %. By the preceding
remarks n(A) = ~(£'), and by the hypothesis ker n s; ker (J, the mapping
A.: n(x) --+ (J(x) (x E A) defines an irreducible representation of ~(£') on %.
By Corollary 2 of 1.4.4, A. is unitarily equivalent to the identity representation,
and note that the latter implies n '" (J. 0

Note in particular that an irreducible representation of a CCR algebra

is determined (to equivalence) by its kernel. As we will see, this useful fact
is true in somewhat greater generality (1.5.4).
There are many common CO-algebras which are not CCR. Here is a
simple example. Let T be an irreducible operator on a Hilbert space £
which is not compact but whose imaginary part is compact. Then C(T) is
not CCR, for the identity representation of C(T) is irreducible but its range
contains the noncom pact operators T and I. Nevertheless, the structure of
C*(T) is not at all pathological. Note for example that C*(T) contains ~(£'),
by Corollary 2 of 1.4.2. Since the quotient map of C(T) on C*(T)/~(£)
annihilates the imaginary part of T, C*(T)/~(£') is generated by a single
self-adjoint element and the identity, and is therefore commutative. Thus
both the ideal ~(£') and its quotient C(T)/~(£) are CCR algebras of the
most tractable kind.
Now let A be a general C-algebra, and let n be an irreducible representa-
tion of A on £'. Since ~(£') is an ideal in 2'(£) it follows that the set
~1t = {x E A :n(x) E ~(£')} is an ideal in A, which contains ker n. Of course

23
1. Fundamentals

it is possible that ~1t = ker n. But in any event the intersection CCR(A) of
all these ideals ~1t as n runs over all irreducible representations is an ideal
in A, consisting ofthose elements of A which are compact in every irreducible
representation. Thus by 1.3.4 CCR(A) is a CCR algebra in its own right,
and moreover CCR(A) contains every other CCR ideal in A. Thus CCR(A)
is the largest CCR ideal in A. Needless to say, CCR(A) might be o.

Definition 1.5.3. A GCR algebra is a C*-algebra A such that CCR(AIl) 1= 0

for every ideal 1 1= A.

Recalling 1.3.2 we see that GCR algebras are characterized by the fact
that all of their *-homomorphic images contain a nonzero CCR ideal. Now,
given any quotient All of a C-algebra A, then we may compose an ir-
reducible representation of All with the quotient map of A on All to obtain
an irreducible representation of A. It follows that every quotient of a CCR
algebra is CCR. In particular, every CCR algebra is GCR (so that GCR
algebras are "generalized CCR algebras").

Proposition 1.5.4. Let A be a GCR algebra. Then for every irreducible rep-
resentation n of A on Yf, n(A) contains ~(Yf). Two irreducible representa-
tions of A which have the same kernel are equivalent.
PROOF. Now n(A) is *-isomorphic with Alker n and therefore it contains a
nonzero CCR ideal. Since the identity representation of the ideal is irreducible
(1.3.4) this implies in particular that n(A) contains a nonzero compact
operator. Therefore the first conclusion follows from Corollary 2 of 1.4.2.
Now let nand (J be two irreducible representations of A on spaces Yf
and %, respectively, such that ker n = ker (J. Thus the map Je: n(x) f-+ (J(x),
x E A, defines an irreducible representation of n(A). Since n(A) contains
~(Yf), the restriction JeI'C(JI") defines a nonzero representation of~(Yf) which
by 1.3.4 is irreducible. By Corollary 2 of 1.4.4 we conclude that JeI'C(JI") is
equivalent to the identity representation of ~(Yf). Finally, by 1.3.4. again,
it follows that Je itself is equivalent to the identity representation of n(A).
Note that this implies n is equivalent to (J. 0

The definition of GCR algebras we have given is not a convenient one to

check, for one first has to find all the ideals 1 in a given algebra A before he
can examine the quotients All. The definition of CCR algebras is easier to
deal with; one simply finds all irreducible representations n of A and verifies
that n(A) consists of compact operators. The preceding result implies that
GCR algebras have an analogous property, namely that for every irreducible
representation n, n(A) contains at least one nonzero compact operator (and
therefore all of them). Thus it is natural to ask if the latter property implies
that A is GCR. The answer is yes, but the proof is hard. The conclusion
follows (in the separable case) from a deep theorem of Glimm [13], or by a
more direct argument given independently by Dixmier [11]. The proof has

24
 1.4. C·-Algebras of Compact Operators

been extended to the inseparable case by Sakai [24, 25]. It turns out that
the second property asserted in 1.5.4 also characterizes separable GCR
algebras [13], [11], but that is also hard. In Theorem 1.5.5 below we will
give another structural condition which is equivalent to the GCR property.
Consider first the operator T in the discussion preceding 1.5.3. Then the
chain of ideals {O, ~(JIl'), C(T)} in C(T) has the property that the successive
quotients ~(JIl')/O, C(T)/~(JIl'), are CCR algebras. This property can be
generalized in the following way. A composition series in a C-algebra A is
a family of ideals {J~; 0 ~ rx ~ rxo} indexed by the ordinals rx, 0 ~ rx ~ rxo,
having the following properties:

(i) for all rx < rxo, J~ is contained properly in J d 1;

(ii) J o = 0, J~o = A;
(iii) if Pis a limit ordinal then J p is the norm closure of U~<p J~.

Thus in the example we have a composition series of length 3, though in

general of course a composition series can be infinite.

Theorem 1.5.5. Every GCR algebra A has exactly one composition series
{J~:O ~ rx ~ rxo} with the property that J~+1/l~ is the largest CCR ideal in
A/l~ for every rx, 0 ~ rx < rxo. Conversely, if A admits a composition series
{J ~: 0 ~ rx ~ rxo} such that each quotient J d 11J ~ is CCR, then A is GCR.
PROOF. Let A be a GCR algebra. We define the series {J~} by transfinite
induction. Put J 0 = O. Note that A = A/l 0 contains a nonzero CCR ideal.
Inductively, let P be an ordinal such that J a has been defined for all rx < P
and satisfies (i), (ii), (iii) above as well as the property stated in the theorem,
whenever they make sense. Now if Phas an immediate predecessor P_, and if
J p_ = A, then the construction ends. If J p_ =F A, then define J p to be the
set of all x E A which are mapped into CCR(A/lpJ under the quotient map
of A on A/lp_. Clearly J p is an ideal containing J p_ for which J p/l p_ =
CCR(A/l pJ is the largest CCR ideal in A/lp_. If, on the other hand, P is a
limit ordinal, then define J p as the norm closure of Ua</l J a • This defines a
composition series {J~:O ~ rx ~ rxo} having the required properties. Note
for example that the induction must terminate at some ordinal whose
cardinal number does not exceed 2card A.
For uniqueness, let {Ka:O ~ rx ~ Po} be another such composition
series, different from {J~:O ~ rx ~ rxo}, and assume for definiteness that
rxo ~ Po· Now if J~ = Ka for all rx ~ rxo then we have a contradiction, for
J~o = K~o = A implies that Po = rxo and the two series are identical. Thus
there is a first ordinal y ~ rxo such that J y =F Ky. Y > 0 because J o = Ko =
O. Note also that by property (iii) in the definition of composition series Y
cannot be a limit ordinal. Thus y has an immediate predecessor y _, and of
course we have J y_ = K y_. Since both quotients Jr/Jy_ and Ky/ly_ equal
the largest CCR ideal in A/ly_ we conclude that J y = Ky (mod JyJ and
hence J y = K y , a contradiction.

25
1. Fundamentals

Finally, let {J ~: 0 ~ rx ~ rxo} be a composition series for A such that

each quotient J d l/J~ is CCR. We have to prove that for every ideal K i= A,
AI K contains a nonzero CCR ideal. Now since U~ J ~ = A it follows that
there is a first y > 0 such that the ideal JylK in AIK is nonzero. We will
show that J yl K is a CCR algebra. Again, we may argue as above to see that
y is not a limit ordinal. Thus y has an immediate predecessor y _, for which
Jy)K = 0, i.e., J y_ S;; K. Thus the mapping x + J y_ 1--+ X + K, x E J y,
defines a *-homomorphism of the CCR algebra JiJy_ onto JyIK, so by
1.3.2 JiK is *-isomorphic to a quotient of the CCR algebra Jy/Jy_. Thus
JiK~CCR. 0

We remark that if A is a separable GCR algebra then its composition

series {J~:O ~ rx ~ rxo} must be countable. Indeed we can choose, for each
rx < 0(0' an element x~ E J~+l such that Ilx~ - zll ~ I for all zE J~. (this is
because J ~ + l/J ~ is not 0). Because {J~} is a well-ordered set of ideals it
follows easily that Ilx~ - xpll ~ I ifrx i= p. Because A is separable, {x~:O ~
rx < rxo} must be a countable set, and therefore rxo is a countable ordinal.
Note finally that the last statement of 1.5.5 shows that each operator T
of the type discussed above is a GCR operator in the following sense.

Definition 1.5.6. A Hilbert space operator T is called a GCR operator if

C(T) is a GCR algebra.

Since finitely generated C*-algebras are always separable, the composition

series associated with a GCR operator is countable. The class of all GCR
operators contains the normal operators, the compact operators, and a
great variety of others as well. Moreover, we will see in the sequel that it is
the GCR operators (and they alone) which lend themselves to the possibility
of analysis in terms of "generalized" spectral theory and multiplicity theory.
As a concluding remark, it is not hard to see that a general C-algebra
A contains a unique ideal K such that K is a GCR algebra in its own right
and AI K has no nonzero CCR ideals (see Exercise 1.5.B). C* -algebras with the
latter property are called NGCR. We shall have little, and certainly nothing
good, to say about NGCR algebras here.

EXERCISES

I.S.A. Show that if J is an ideal in a GCR algebra A, then both J and AjJ are GCR
algebras
I.S.B. A C*-algebra having no nonzero CCR ideals is called an NGCR algebra. Show
that every CO-algebra A contains a unique ideal K such that K is GCR and AjK
is NGCR.
I.S.c. Let d '= 2'(£') be an irreducible C*-algebra of operators. Show that A is an
NGCR algebra iff d contains no nonzero compact operators.

26
 1.6. States and the GNS Construction

1.S.D. Let S be the unilateral shift (cf. Exercise 1.4.0). Show that C*(S) is a GCR algebra
and describe its canonical composition series.
1.S.E. A unilateral weighted shift is an operator defined on an orthonormal base
e!, ez, ... by the condition

where {w n } is a bounded sequence of nonnegative real numbers.

a. Show that if Wn > 0 for every n, then C(A) is irreducible.
b. Show that if Wn :::::: b > 0 for every n, then C(A) cannot be an NGCR
algebra. [Hint: use 1.2.A and 1.5.C.]

1.6. States and the GNS Construction

We now want to discuss certain matters relating to the existence of
representations of C*-algebras, and how one goes about constructing them.
Ifwe are given a C*-algebra d of operators on a Hilbert space then there are
always certain representations in evidence, namely the identity representation
of d and its subrepresentations. If d contains only compact operators, then
the results of Section 1.4 provide a very explicit method for constructing all
possible representations of d from the irreducible subrepresentations of id
(to be perfectly accurate, we only obtain a representative from each unitary
equivalence class, but of course that is good enough). On the other hand, if
d contains some noncom pact operators, then this method does not exhaust
the possibilities; it can be shown, for example, that in this case there will
always exist irreducible representations of d which are not equivalent to
subrepresentations of the identity representation. Indeed, the identity re-
presentation of d might contain no irreducible subrepresentations whatso-
ever. In the extreme case, the given C*-algebra may be defined in some
abstract fashion which does not put into evidence even a single nontrivial
representation.
The purpose of this section is to show how representations of an abstract
C*-algebra can be constructed from certain linear functionals, and in
Section 1.7 we will show that these functionals (and therefore representa-
tions) always exist in abundance.
Let A be an abstract C-algebra with unit e, which will be fixed throughout
the discussion. A linear functional f on A is said to be positive if f(z*z) ~ 0
for every z in A; if f is normalized so that f(e) = 1, then it is called a state.
As an example, suppose we are given a representation n of A on a Hilbert
space Yf and a vector ~ in Yf. Then f(x) = (n(x)~, ~) defines a linear func-
tional on A, and it is very easy to see that, in fact, f is positive. If n is non-
degenerate, then f is a state if and only if ~ is a unit vector. We will first
describe a very useful procedure, due to Gelfand, Naimark, and Segal,
whereby one starts with a positive functional f and constructs n and ~ so
that the above relation is satisfied. We then characterize those states f which
give rise to irreducible representations n.

27
1. Fundamentals

With every linear functional f on A there is an associated sesquilinear

form L .] on A x A, defined by [x, y] = f( y* x). Notice that when f is a
positive linear functional its associated form is positive semidefinite, and
therefore satisfies the Schwarz inequality. In terms of f, this is simply
(1.6.1) If(y*xW ~ f(x*x)f(y*y),
for every x, y in A. 1.6.1 is called the Schwarz inequality for positive linear
functionals. We shall also require the following fact, which implies that
positive linear functionals are automatically continuous.

Proposition 1.6.2. Every positive linear functional f has norm f(e).

PROOF. We claim first that if x is an element of A satisfying x = x* and

Ilxll ~ 1, then there is a self-adjoint element YEA such that e - x = y2.
Indeed, the sub C*-algebra generated by x and e is abelian, and the image
of x under the Gelfand map is a real-valued continuous function taking
values in the interval [ -1, + 1]. Thus e - x corresponds to a continuous
function taking values in [0, + 2], which therefore has a continuous real-
valued square root. y may be taken as the inverse image of the latter under
the Gelfand map.
Now let z E A, Ilzll ~ 1. Then by the Schwarz inequality 1.6.1 we have
If(zW = If(e*zW ~ f(z*z)f(e*e) = f(z*z)f(e). So to prove Ilfll ~ f(e) it
suffices to show that f(z*z) ~ f(e), or equivalently, f(e - z*z) ~ 0. But by
the preceding paragraph we know that there is an element y = y* in A such
that e - z*z = y2, from which the assertion is evident. The opposite in-
equality Ilfll ~ f(e) is apparent from the fact that e has unit norm. 0

So in particular, every state of A has norm 1. It is significant that, while

the proof we have given works only for C* -algebras, 1.6.2 itself is true for a
much broader class of Banach *-algebras ([6], p. 22). We come now to the
main discussion, the result of which is summarized as follows:

Theorem 1.6.3. For every positive linear functional f on A, there is a re-

presentation 1t of A and a vector ~ such that
f(x) = (1t(x)~, ~)
for every x in A.

To construct 1t, one first considers the left regular representation 1to,
which represents A as an algebra of linear transformations. Specifically, for
each x in A, the linear transformation 1to(x) is defined on A by 1to(x):y -+ xy.
Clearly 1to preserves the ring operations of A and respects scalar multiplica-
tion. Moreover, [x, y] = f(y*x) defines a positive semidefinite sesquilinear
form on A x A which satisfies
(1.6.4) [1to(x)y, z] = [y, 1to(x*)z]

28
 1.6. States and the GNS Construction

for all x, y, z in A. For z fixed, the usual manipulations with the Schwarz
inequality show that the condition [z, z] = 0 is equivalent to the condition
[x, z] = 0 for every x in A. Thus the set N = {z E A:f(z*z) = O} is a linear
subspace of A. Moreover, 1.6.4 implies that [y, xz] = [x*y, z], from which
it follows that N is a left ideal in A. In particular, N is an invariant subspace
for every operator no(x), x E A.
This allows us to lift each operator no(x) in the natural way to a linear
transformation n(x) in the quotient space A/N. Similarly, we can define an
inner product (', . ) in A/N by
(x + N, y + N) = [x, y].
The relation 1.6.4 persists in the form
(1.6.5) (n(x)'l, 0 = ('1, n(x*m,
X E A, '1,' E A/N. Now suppose, for the moment, that we already know each
operator n{x) is bounded in the Hilbert norm on A/N. Then we may com-
plete A/N to obtain a Hilbert space :Yt', and every operator n(x) extends
uniquely to a bounded operator on :Yt', which we denote by the same symbol
n(x). The map x ---? n(x) preserves the algebraic operations of A, and 1.6.5
implies that n(x*) = n(x)*. Thus, n is a representation of A. Noting finally
that the vector ~ = e + N in :Yt' satisfies (n(xR, ~) = [xe, e] = f(e*xe) =
f(x), the desired representation of f is achieved.
It remains to show that each linear transformation n(x) is bounded.
Actually, we will show that Iln(x)'l11 ~ Ilxll'II'l11 for every '1 E A/N, x EA.
Since '1 must have the form y + N for some YEA, and since Iln(x)'l112 =
Ilxy + NI12 = f( (xy)*xy) = f(y*x*xy), this reduces to the inequality
f(y*x*xy) ~ IlxI12f(y*y). To prove the latter, fix y and consider the linear
functional g(z) = f(y*zy). Since f is positive, so is g, and by 1.6.2 we know 9
has norm g(e) = f(y*y).1t follows that g(x*x) ~ Ilx*xllf(y*y) ~ IlxI12f(y*y),
and the proof is complete. D

Thus every positive linear functional f can be represented in the form

f(x) = (n(x)~, ~), where n and ~ are as above. We remark that one may
always assume that ~ is a cyclic vector for n in the sense that n(A)~ is dense
in the underlying space. For if this is not so, simply replace ~ with its pro-
jection on the subspace [n(A)~] and n with the corresponding subrepresen-
tation;' these new quantities serve equally to represent f, and it is very
easy to check that they have the desired property.
Recall that a positive linear functional f is called a state if it is normalized
so that f(e) = 1. Evidently, the states form a convex set of linear functionals
on A, and an extreme point of this convex set is called a pure state. Thus,
a state f is pure iff for any two states gl and g2 and every real number t,
o < t < 1, the condition f = tg 1 + (1 - t)g2 implies gl = g2 = f. We
can now describe an important connection between pure states and ir-
reducible representations.

29
1. Fundamentals

Theorem 1.6.6. Let n be a representation of A having a unit cyclic vector e,

and let f be the state f(x) = (n(x)e, e). Then f is pure if, and only if,
n is irreducible.

PROOF. Assume first that f is pure, and let E be a projection in the com-
mutant of n(A). We will prove that E = 0 or 1.
Assume, to the contrary, that E =1= 0 and E =1= 1. Then we claim Ee =1= 0
and E1.e =1= O. For if, say, Ee = 0, then for every x E A we have En(x)e =
n(x)Ee = 0, and E = 0 follows from the fact that n(A)e is dense. The same
reasoning shows E1.e =1= O.
It follows that the real number t = IIEel12 = (Ee, e) is positive and less
than 1. Define linear functionals gl> g2 on A by gl(x) = t-I(n(x)e, Ee)
and g2(X) = (1 - t)-I(n(x)e, E1.e). Clearly f = tg l + (1 - t)g2' and we
claim now that each gi is a state. Indeed, since we can write (n(x)e, Ee) =
(n(x)e, E 2e) = (En(x)e, Ee) = (n(x)Ee, Ee), it is apparent that gl(x) =
t- 1 (n(x)Ee, Ee) is positive, and satisfies gl(e) = t-111Ee112 = 1. Similarly,
g2 is a state. Since f is an extreme point, we conclude, in particular, that
gl = f or, what is the same, (n(x)~, E~) = t(n(x)e, ~) for every x in A. This
implies (n(x)e, Ee - tel = 0, and since n(A)~ is dense, we see that E~ - t~ =
e,
O. Since E - tl commutes with n(A) and annihilates we may repeat an
argument given before to conclude that E - tl = O. But this is absurd
because E is a projection and 0 < t < 1.
Conversely, assume n is irreducible, and suppose we are given states
gi and t E (0, 1) such that f = tg l + (1 - t)g2' We will show that gl = f.
To that end, define a sesquilinear form [', .] on the dense linear manifold
n(A)e as follows:
[n(x)e, n(y)~] = tgl(Y*x).

Because 0 ~ tgl(x*x) = f(x*x) - (1 - t)g2(X*X) ~ f(x*x), it follows that

o~ [n(x)~, n(x)~] ~ Iln(x)~112, and in particular L']
is bounded. By a
familiar lemma of Riesz, there is an operator H on the underlying Hilbert
space satisfying 0 ~ H ~ 1, and [11, n
= (11, HO for all 11,' in n(A)~.
Taking 11 = n(y)e and, = n(z)~ we obtain tgl(z*y) = (n(y)e, Hn(z)e).
We claim now that H commutes with n(A). Since n(A)~ is dense, this
amounts to showing that

(n(y)~, Hn(x)n(z)~) = (n(y)e, n(x)Hn(z)e)

for every x, y, z in A. But the left side is (n(y)~, Hn(xz)~) = tgl«xz)*y), and
the right side is (n(x)*n(y)e, Hn(z)e) = (n(x*y)~, Hn(z)~) = tgl(z*x*y), from
which the assertion is evident.
Since n is irreducible, the commutant of n(A) consists of scalar operators,
and we conclude that there is a (real) scalar r such that H = r1. Hence,
tg 1(y) = (n(y)~, H~) = r(n(y)~, e) = rf(y). Setting y = e and noting that
f(e) = gl(e) = 1, we see that t = r, and finally we conclude that gl = f·
The proof that g2 = f is the same. 0

30
 1.7. The Existance of Representations

EXERCISES

1.6.A. For every unit vector, in a Hilbert space Jr, let w~ be the linear functional on
£'(Jr) defined by
w~(T) = (n, ,).

a. Show that each w~ is a pure state.

b. Show that the representations of £'(Jr) associated to the w~'s via the GNS
construction are all equivalent to the identity representation.
1.6.B. Let n, (1 be two representations of a C*-algebra A having respective cyclic vectors
" 1/. Assuming that (n(x)" ,) = ((1(x)I/,,,) for each x E A, show that nand (1 are
equivalent.
1.6.C. Let .91 be the CO-algebra C>(S) generated by the unilateral shift (see l.4.D and
1.5.0).
a. Find all irreducible representations of .91 (up to unitary equivalence).
b. Let p be a pure state of .91 which does not vanish on CC(Jr). Show that
there is a unit vector, in Jr such that p(T) = (T" ,), TEd.
c. Find all the remaining pure states of d.
1.6.D. Let A be the algebra of all polynomials in a single real variable having complex
coefficients. Give A the norm
IIfil = sup If(t)1
O"rE;;1

and the involution f*(t) = f(t).

a. Show that A has all the properties of a commutative CO-algebra with
identity except completeness.
b. Show that pU) = f(2) defines a linear functional on A which satisfies
pU>f) ;?; 0 for all J, but which is not continuous.
1.6.E. Let A be a C*-algebra without unit and let f be a linear functional on A satis-
fying f(x>x) ;?; 0 for all x. Let Al be the CO-algebra obtained from A be adjoining
a unit e, let M be a positive constant, and define J:AI ..... C by
f(Ae + x) = AM + f(x),
A E C, X E A. Prove that in order for 1 to be positive it is necessary for f to be
bounded and that M should satisfy M ;?; IIfil (note that 1.6.2 does not apply
to nonunital CO-algebras).

1.7. The Existence of Representations

The purpose of this section is to show that an abstract C*-algebra has
"sufficiently many" irreducible representations, and to present a famous
theorem of Gelfand and Naimark on the concrete representation of such
algebras. These results rest on the fact that self-adjoint elements of the form
z*z must have nonnegative spectra. This basic property of C*-algebras is
proved in 1.7.1, after one or two preliminaries.
As in the preceding section, A will denote an abstract C*-algebra with unit
e. If z is an element of A, we will use the notation sp(z) ~ 0 to express the fact

31
1. Fundamentals

that the spectrum of z is contained in the nonnegative real axis. sp(z) ~ 0

has the obvious meaning. We first want to point out that the condition
sp(z*z) ~ 0 is more or less obvious when z = z*. Indeed, we know from
Section 1.1 that sp(z) is real, and so the spectral mapping theorem implies
thatsp(z2) = p.2:.A.ESp(Z)} ~ O.

Lemma I. ffx and yare self-adjoint elements of A with sp(x) ~ 0 and sp(y) ~
0, then sp(x + y) ~ O.
PROOF. By multiplying the sum x + y by a sufficiently small positive scalar,
if necessary, we may assume Ilxll ~ 1 and Ilyll ~ 1. Now since the spectral
radius does not exceed the norm, the spectrum of x must be contained in the
real interval [0, + 1]. Thus, sp(e - x) 5; [0, + 1]. Since e - x is self-adjoint,
its norm agrees with its spectral radius, and we have lie - xii = r(e - x) ~ 1.
Similarly, lie - yll ~ 1. It follows from the triangle inequality that
e - t(x + y) = He - x) + He - y) has norm at most 1, and so the spec-
trum of the self-adjoint element Hx + y) must lie in the interval [0, + 2]. The
desired conclusion is now immediate. 0

In the following lemma, we shall have to make use of the fact that if x and
yare elements of an arbitrary Banach algebra with unit, then the nonzero
points in sp(xy) and sp(yx) are the same (Exercise 1.7.A).

Lemma 2. Let z E A be such that sp(z*z) ~ O. Then z = O.

PROOF. Assume that sp(z*z) ~ O. Then by the preceding remark we also

have sp(zz*) ~ O. Lemma 1, applied to the elements -z*z and -zz*, implies
that sp(z*z + zz*) ~ O.
On the other hand, let z = x + iy be the Cartesian decomposition of z,
where x and yare self-adjoint. A direct computation shows that z*z + zz* =
x 2 + y2. We have already remarked that Sp(X2) ~ 0 and Sp(y2) ~ 0; so by
Lemma 1 once again, we conclude that sp(z*z + zz*) = Sp(X2 + y2) ~ O.
Combining these two assertions, we see that the self-adjoint element
z*z + zz* = x 2 + y2 has spectrum {O}, and hence x 2 + i = O. Note that
this implies x = y = O. For Sp(X2) ~ 0, and since x 2 = - y2, we also have
Sp(X2) ~ O. As before, we conclude that x 2 = 0, and therefore x = O. Simi-
larly, y = O. That proves z = x + iy = O. 0

Theorem 1.7.1. For every z in A, one has sp(z*z) ~ o.

PROOF. We claim first that there exist self-adjoint elements u, v in A for which
uv = vu = 0 and z*z = u 2 - v 2 • Indeed, if we consider the real-valued con-
tinuous functions f, 9 defined on IR by

f(t) = {f t ~ 0
t < 0,

32
 1.7. The Existance of Representations

and

g(t) = {p t
t
~ °
< 0,
thenfg = ° and f(t)2 - g(tf = t. It follows that u = f(z*z) and v = g(z*z)
have the asserted properties.
Thus, we can write z*z = u2 - v2 as above. Because uv = 0, it follows
that vz*zv = vu 2v - v4 = - v4 ; and since sp( - v4) = - Sp(V4) ~ 0, we con-

°
clude that the element w = zv satisfies sp(w*w) ~ 0. By Lemma 2, - v4 =
w*w = 0, hence v4 = 0, and so v = because v is self-adjoint. All of this
shows that z*z = u2 is the square of a self-adjoint element of A, and the con-
°
clusion sp(z*z) ~ now follows from the remark preceding Lemma 1. 0

We have seen in the preceding section that a state must have norm 1. The
following provides a useful converse to this assertion.

Corollary. Let f be a bounded linear functional on A such that Ilfll = f(e) = 1.

Then f is a state.
°
PROOF.
°
We have to show that f(z*z) ~ for every z EA. By 1.7.1, it suffices
to show thatf(y) ~ for every self-adjoint element YEA having nonnegative
spectrum. Actually, we will prove this: for every YEA satisfying y*y = yy*,
f(y) belongs to the closed convex hull K of sp(y). For that, suppose, to the
contrary, that f(y) does not belong to K. Since K is the intersection of all

°
closed discs which contain sp(y), we infer that there is a complex number a
and an R > such that sp( y) is contained in the disc {A. E C: 1,1. - al ~ R},
but If(y) - al > R. The first assertion implies that the spectral radius of
y - ae does not exceed R, and since y - ae generates a commutative C*-
subalgebra, we know from 1.1.1 that Ily - aell = r(y - ae) ~ R. On the
other hand, If(y - ae)1 = If(y) - al > R, and these two assertions con-
tradict the fact that Ilfll = 1. 0

We are now in a position to prove a key result on the existence of states.

Theorem 1.7.2. Let x be a self-adjoint element of A. Then there is a pure state

f such that If(x)1 = Ilxll·
PROOF. We claim first that there is a state f with the property If(x)1 = Ilxll.
Consider the C*-subalgebra B generated by x and e. Since B is commutative,
it follows from 1.1.1 that there is a complex homomorphism w of B such that
Iw(x)1 = Ilxll (simply choose w corresponding to some point of the maximal
ideal space of B at which the Gelfand transform of x achieves its maximum
modulus). We already know that such an w satisfies Ilwll = w(e) = 1. By the
Hahn-Banach theorem, there is an extension f of w to A which satisfies
Ilfll = 1. The preceding corollary implies thatf is a state, and by construction
we have If(x)1 = Iw(x)1 = Ilxll·

33
1. Fundamentals

Now fix such a state f, and let I be the set of all states 9 of A satisfying
g(x) = f(x). I is clearly closed in the weak· -topology of the unit ball of the
dual of A, and therefore, by Alaoglu's theorem, it is compact. I is also
convex, and the preceding paragraph shows I is not void. Thus, the Krein-
Milman theorem implies that I has an extreme point g.
To complete the proof, we have to show that 9 is an extreme state. But if
g1 and g2 are states and 0 < t < 1 are such that 9 = tg 1 + (1 - t)g2, then
notice that g1 and g2 must belong to I.
For the conditions Igi(X)1 ~ Ilxll =
If(x)l, together with tg 1(x) + (1 - t)g2(X) = g(x) = f(x), imply that g1(X) =
g2(X) = f(x), i.e., gi E I.Because 9 is extreme in I,
we have the desired
conclusion g1 = g2 = g. 0

Corollary. For every nonzero element z in A, there is an irreducible represen-

tation 1t of A on a Hilbert space .Yf and a unit vector ~ in .Yf such that

111t(z)~11 = Ilzll > O.

PROOF. Taking x = z·z in the preceding theorem, we obtain a pure state f
of A such that f(z·z) = Ilz·zll = Ilz112. Let 1t and ~ be the representation and
unit vector associated with f via the GNS construction (1.6.3). Then

111t(z)~112 = (1t(z)~, 1t(z)~) = (1t(z·z)~, ~)

= f(z·z) = Ilz112,
and hence 111t(z)~11 = Ilzll. By 1.6.6, 1t is irreducible, and the proof is
complete. 0

This corollary can be used to prove that every locally compact group has
"sufficiently many" irreducible unitary representations. Unfortunately, an
adequate discussion ofthis important application would take us too far afield,
and instead we refer the reader to [6J.
We can now deduce the theorem of Gelfand and Naimark mentioned at
the beginning of the section.

Theorem 1.7.3. Gelfand-Naimark theorem. Every abstract C·-algebra with

identity is isometrically ·-isomorphic to a C*-algebra of operators.

PROOF. We have to exhibit an isometric representation 1t of A. By the pre-

ceding corollary we may choose, for each x "# 0 in A, a representation 1tx
such that II1tAx)11 = Ilxll. Now simply let 1t be the direct sum of all these
~ 0

This theorem and the corollary of 1.7.2 are valid for C*-algebras without
identities, and in fact can be deduced from what we have done by adjoining
an identity (see Exercise 1.7.C).

34
 1.8 Order and Approximate Units

EXERCISES

1.7.A. Let A be a Banach algebra with unit e, and let x, y belong to A. Show that e - xy
is invertible if and only if e - yx is invertible. Deduce that sp(xy) and sp(yx)
agree except possibly for the point O.
1.7.B. Let A be a CO-algebra with unit e and let x be a self-adjoint element of A satis-
fying Ilxll ~ 1. Show that sp(x) ~ 0 if and only if lie - xii ~ 1.
1.7.C. Generalize the corollary of 1.7.2 and Theorem 1.7.3 to CO-algebras without
identities.
1.7.D. Show that the CO-algebra 2'(Yt') has pure states other than the states w~, ~ E H,
of Exercise 1.6.A.

1.8. Order and Approximate Units

Approximate units have already been discussed rather briefly in Section
1.3. We are now going to indicate how approximate units can be constructed
in general C*-algebras, and we will complete the discussion of positive linear
functionals. While these results are important for some applications of
C*-algebra theory we will not require them here, and the reader may omit
this section without essential loss.
Let A be a C'-algebra, with or without unit. Given a self-adjoint element
x in A, it is customary to write x ~ 0 if the spectrum of x is nonnegative (if A
has no unit then it is understood that sp(x) is computed in the C'-algebra
obtained by adjoining a unit to A). Such elements are called positive. If
x and yare two self-adjoint elements, then x ~ Y (or Y ~ x) will mean that
y - x is positive. To say that this relation is transitive means that the sum of
two positive elements is positive, a fact that has already been established in
Lemma 1 of the preceding section. For the same reason, we see that if Xi ~ Yi
for i = 1, 2, ... , n, then Xl + ... + Xn ~ Yl + ... + Yn' Finally, notice
that x ~ Y ~ x implies Y = x (because Y - x is then a self-adjoint element
of A having spectral radius zero).
Thus ~ is a partial ordering of the set of self-adjoint elements of A. If X
is a locally compact Hausdorff space and we take A to be the concrete C'-
algebra Co(X) of all complex-valued continuous fUllctions vanishing at
infinity, then ~ reduces to the usual partial order f ~ 9 iff f(p) ~ g(p) for
every p in X. Carrying this a bit farther, the reader can easily see that the
Gelfand map of an abstract commutative C'-algebra A is also an order
isomorphism in the sense that x ~ Y in A iff x ~ y as functions in Co(A).
We remind the reader that some care must be exercised when utilizing the
functional calculus in C*-algebras A which have no identity: one may only
apply continuous functions which vanish at the origin to self-adjoint elements
of A (see Exercise 1.1.1). After this qualification, the preceding remarks show
that the functional calculus preserves order even in the case of noncommuta-
tive C*-algebras. For instance, if f and 9 are two (real-valued) continuous

35
1. Fundamentals

functions of a real variable such that f(t) ~ g(t) for every real t, then f(x) ~
g(x) for every self-adjoint element x in A.
There are two algebraic characterizations of this partial ordering which
are frequently useful, and are summarized as follows.

Proposition 1.8.1. Let x be a self-adjoint element of A. Then the following

are equivalent:
(i) x ~ 0.
(ii) There is an element y = y. in A such that x = y2.
(iii) There is an element z in A such that x = z·z.
PROOF. That (ii) implies (iii) is a triviality, (iii) implies (i) is 1.7.1, and (i) implies
(ii) is a simple exercise with the functional calculus which we leave for the
reader. 0

This result implies an invariance property of the partial order which is

often useful: ifx ~ y then z·xz ~ z·yzfor every element z in A. To see why,
use (ii) to find u = u· in A with y - x = u 2 and note that z·yz - z·xz has
the form w·w with w = uz.
An approximate unit (or approximate identity) for a C*-algebra A is a net
e). of self-adjoint elements of A satisfying

(i) e). ~ 0, Ile).11 ~ 1;

(ii) A ~ JJ. implies e). ~ e,,;
(iii) lim).llxe). - xii = lim).lle).x - xii = 0, for every x in A.
Before giving the construction of approximate units, we need a preliminary
result.

Lemma. Assume A has a unit and let x, y be invertible positive elements of

A satisfying x ~ y. Then y-l ~ X-I.
PROOF. Notice first that for every z E A the condition z·z ~ e (e denoting
the unit of A) is equivalent to Ilzll ~ 1. For example, if Ilzll ~ 1 then Ilz·zll ~ 1,
and by 1.7.1 we see that sp(z·z) is contained in the unit interval [0, 1]. Hence
sp(e - z·z) ~ 0. The converse is similar, and we omit the proof.
Since both x and yare positive, we may find positive square roots for each
of them (using the functional calculus), which are denoted by X l/2 and 12 . i
Now multiply x ~ y on left and right by y-1 /2 to obtain y-1 /2xy -1 /2 ~ e.
Theprecedingparagraph,appliedtoz = xI/2y-1/2,impliesllxI/2y-1/211 ~ 1,
hence Ily- I /2x l/2 11 ~ 1 (by taking the adjoint), and so zz· ~ e. That is simply
x1/2y-Ix1/2 ~ e, and now the conclusion y-l ~ X-I follows after another
double multiplication by x- 1/2 . 0

Theorem 1.8.2. Every C·-algebra has an approximate unit.

PROOF. We may as well assume that the given algebra A has no unit, and let
Al be the C*-algebra obtained from A by adjoining one.

36
 1.8 Order and Approximate Units

Let A be the directed set of all finite sets of self-adjoint elements of A

(ordered in the usual way by set inclusion), and for each integer n ~ 1 let
In be the function defined on the interval [0, + OCJ) by
f,.(t) = nt(1 + nt)-l.
Because f,. is continuous and vanishes at the origin, it operates on the set
of all positive elements of A. So for each A. E A, say A. = {x b . . • , x n }, we
may define a self-adjoint element e A in A by
eA = In(xi + ... + x;).
°
Because xi + ... + x; ~ and f,. maps [0, + OCJ) into the unit interval,
we have sp(eA ) s;: [0, 1], so that property (i) in the definition of approximate
units is satisfied.
For (ii), let A. ~ J1, say A. = {Xl' ... , x m } and J1 = {Xb ... , x n } where
m ~ n. To prove eA ~ eJL it suffices to show that e - eJL ~ e - eA (where
both sides of the latter inequality are interpreted as elements of AI). But
since 1 - f,,(t) = (1 + kt) -1, the latter inequality is simply
[e + n(xi + ... + X;)]-l ~ [e + m(xi + ... + x;,)]-r,
which follows from the lemma because
e + m(xf + ... + x;') ~ e + n(xi + ... + x;).
To prove (iii), it suffices to show that
x in A; for this condition implies IIx -
Ilx -
° °
xeAli ~ for every self-adjoint
eAxil ~ (simply by taking adjoints),
and so (iii) follows from these two conditions because every element of A
can be decomposed into the form Xl + iX2 with Xi = xi.
So fix x = x* in A, let m be a positive integer, and let A. = {Xb ... ,xn }
be any finite set of self-adjoint elements which contains x and has at least
m elements. Then we claim that IIx - xe AII 2 ~ 114m. Indeed, since x2 ~
xi + ... + x; we have
(e - eA)x 2(e - eA) ~ (e - eA)(xi + ... + x;)(e - eA).

Now the right side has the form gixi + ... + x;), where git) = (1 - f,.(t))-
t(1 - f,.(t)) = t(1 + nt)-2. Since gn is bounded above on [0, +OCJ) by 1/4n,
we conclude that (e - eA)x 2(e - eA) ~ (4n)-le. This asserts that the element
z = x - xe A = x(e - eA) satisfies z*z ~ (4n)-le, and an argument already
given in the proof of the lemma shows that IIzll2 ~ (4n)-1 ~ (4m)-I. (iii) now
follows because m was arbitrary. 0

It is significant that if A is separable, then the net e A can be chosen to be

a sequence el' e2' .... For if Xb ... is a sequence dense in the set of self-
adjoint elements of A then, in the notation of the preceding proof, one may
define en by

37
1. Fundamentals

Properties (i) and (ii) are clearly satisfied for this new sequence, and the
proof of (iii) is a minor variation of the above which is left for the reader.
It was shown in Section 1.6 that a positive linear functional on a C·-
algebra A with unit e is bounded, and in fact achieves its norm at e. Suppose
now that f is a positive linear functional on a nonunital C·-algebra A.
It is tempting to try to prove f is bounded by writing down a positive ex-
tension J of f to the C·-algebra Ai obtained by adjoining a unit e to A,
and then utilizing the preceding result. However, f must have the form
J(te + x) = tM + f(x) (t E IR, x E A), where M is some appropriately
chosen positive constant, and it is not hard to see that in order for J to be
positive it is necessary to choose M at least as large as
sup{f(x):x = x·, sp(x) ~ [-1, +1]} = sup{f(x):x = x·, IIxll :::; I}.
So it is necessary that the latter supremum be finite, and that requires f to
be bounded already.
What is required is the following result, which generalizes Proposition
1.6.2.

Theorem 1.8.3. Let f be a positive linear functional on a C* -algebra A and let

e). be an approximate unitfor A. Thenfis bounded, and Ilfll = lim).f(e).).
PROOF. We first claim that f is bounded on the positive part B+ of the
unit ball of A (here, of course, B+ denotes the set of self-adjoint elements x
satisfying x ~ 0 and Ilxll :::; 1). Notice that f(x) ~ 0 for every x in B+, so
if f is unbounded on B+ then we may find a sequence Xn E B+ such that
f(x n ) > l/n. Consider the infinite series In- 2 x n. This converges to a self-

adjoint element x E A, and since each ofthe partial sums is positive it follows
that x ~ 0 (see Exercise 1.8.B). Similarly, letting Yk denote the k-th partial
sum of the series, we have x ~ Yk and so f(x) ~ f(Yk). This implies that
f(x) ~ In - 2 f(x n ) ~In -
1 = + 00, a contradiction.
So let M = sup{f(x):x E B+}. Then for each x = x· E A with Ilxll :::; 1
we can write x = Xl - X2 with Xi E B+ (Exercise l.3.C.), so that If(x) I :::; 2M.
For an arbitrary Z E A with Ilzll :::; 1, write z = Yl + iYl where Yi = Yi and
IIYil1 :::; 1, and note that If(z)1 :::; If(Yl)1 + If(Yl)1 :::; 4M. That shows f is
bounded.
Now since A :::; Il implies e). :::; ell and f is positive, f(e).) is a (bounded)
increasing net; hence lim f(e).) exists and is at most Ilfll. For the opposite
inequality, fix x E A, Ilxll :::; 1. Then xe). --+ x as Ai, so by the Schwarz
inequality 1.6.1 (which does not depend on the existence of a unit) we have
li(x)j2 = lim If(xe).)j2 :::; lim sup f(d)f(x·x).
). ).

Since Ilx·xll :::; 1 we have f(x·x) :::; IIfll, and since d :::; e). (because of the
fact that t 2 :::; t over the unit interval and the order properties of the func-
tional calculus) it follows that f(eI) :::; f(e).). So the displayed inequality
implies that If(xW :::; Ilflllim). f(e).) from which the desired assertion
Ilfll :::; lim). f(e).) is immediate. 0

38
 1.8 Order and Approximate Units

We have already indicated at the end of Section 1.7 how the results of
that section could be derived for nonunital C·-algebras. With the help of
the preceding theorem it is now a simple matter to imitate the GNS con-
struction and the main results of Section 1.6 for such C·-algebras. The
explicit details are left for the reader.

EXERCISES

1.S.A. Let x, y be self-adjoint elements of a CO-algebra A satisfying 0 ~ x ~ y. Prove

that Ilxll ~ Ilyll·
1.S.B. Let A be a C-algebra and let P = {x E A:x ~ O}. Show that P is a closed cone
in A, dual to the cone of all positive linear functionals on A [Hint: see 1.7.B.]
1.S.C. Let J be an ideal in a C-algebra A, let {el} be an approximate unit for J, and
let x -+ x be the projection of A on AIJ. Prove that, for each element x in A,
lim IIx - xelll = 11xu.
1

1.S.D. Let n be a nondegenerate representation of a C-algebra A on a Hilbert space

JIf, and let {el} be an approximate identity for A. Show that the net of operators
n(el) converges in the strong operator topology to the identity operator on JIf.

39
2
Multiplicity Theory

The objects in a number of mathematical categories (groups, C -algebras,

a-fields of sets are but three examples) can be represented in a variety of
ways by means of operators on Hilbert spaces. For example, the permutation
group on n letters acts in an obvious way to permute the elements of an
orthonormal basis for an n-dimensional Hilbert space, and this gives rise
to a representation of that group as a group of unitary operators. Multi-
plicity theory has to do with the classification of these representations. This
classification problem amounts to finding, for a given object, a suitable
collection of representation invariants which will allow one to determine
when any two of its representations are, or are not, geometrically the
"same"-that is, unitarily equivalent. It goes without saying that in order
to classify representations one first has to find them (all of them), so an
adequate solution to this vague problem should contain a procedure for
constructing all possible representations of the object (more precisely, at
least one representative from every unitary equivalence class) in usable and
concrete terms.
Unfortunately, even for the case of separable C-algebras, such a detailed
and general classification appears today to be hopelessly out of reach. If
one restricts attention to an appropriate subcategory however, namely
C-algebras having only type I representations, then a quite satisfactory
theory is available, and that is the subject of this chapter and Chapter 4.
These results contain the Hahn-Hellinger classification of self-adjoint
operators, and more generally they give useful information about the struc-
ture of operators which generate GCR C-algebras.
The work behind this classification divides naturally into two distinct
parts. The first step is to analyze the set of operators in the range of a given
representation, and is very general in the sense that the analysis for unitary

40
 2.1. From Type I to Multiplicity-Free

representations of groups is identical with the analysis for representations

of C*-algebras, and so on. The work amounts to a reduction of the problem
for general representations to that for "multiplicity-free" representations,
and is taken up in 2.1 below. The second step is to classify the multiplicity-
free representations of a given object in terms of a suitable set of invariants.
The details here vary from one category to another, but generally the in-
variants turn out to be measures (more precisely, measure classes) defined
on a canonical a-field of subsets of a certain set (the dual object) associated
with the given object. This part of the analysis is carried out for commutative
C*-algebras in Section 2.2, and for the larger class of GCR algebras in
Chapter 4.

2.1. From Type I to Multiplicity-Free

For definiteness, we shall consider representations of a C*-algebra A,
fixed throughout this section, and in this section all nonzero representations
of A are assumed to be non degenerate. Let us first recall one or two simple
facts about representations of A. If a is a subrepresentation of a representation
n of A and ml is the range space of a, then since n(A) is a self-adjoint family
of operators ml1. is also n(A )-invariant and thus defines a subrepresentation
0"1.. 0" and a1. are orthogonal in an obvious sense and we have a + 0"1. = n.
More generally, if {ad is any family of mutually orthogonal subrepresen-
tations of n then Li ai has an obvious meaning and defines another sub-
representation of n. Given two such families {O"d and {at} for which ai is
equivalent to at for all i, then a simple "piecing together" argument shows
that the subrepresentations Li a i and Li at are equivalent. The notation
0"1 1- 0"2 means that a 1 and a2 are orthogonal subrepresentations of n.
Finally, it is a simple consequence of the definitions that two (not necessarily
orthogonal) subrepresentations a1> a2 of n are equivalent if, and only if,
there is a partial isometry in the commutant n(A)' with initial and final spaces
equal, respectively, to the range spaces of 0"1 and a2'
The term "multiplicity" is supposed to suggest the presence of multiple
copies of the same representation. Thus, if a is a given representation of A
on a Hilbert space Yf then n(x) = a(x) EB a(x) EB a(x), x E A, defines a
representation of A on Yf EB Yf EB Yf which apparently has multiplicity.
It would be too naive to say that n has multiplicity 3 however, since, for
example, a itself might admit such a multiple decomposition. Our starting
point, then, will be to isolate those representations which are multiplicity-
free in the sense that this kind of multiple decomposition cannot occur.
Definition 2.1.1. A representation n of A is said to be multiplicity-free if n
does not have two nonzero orthogonal equivalent subrepresentations.
In other words, n is multiplicity-free iff the only representation a of A for
which 0" EB 0" is equivalent to a subrepresentation of n is the zero represen-
tation. It is apparent from the definition that every subrepresentation of a

41
2. Multiplicity Theory

multiplicity-free representation is multiplicity-free. We shall make repeated

use of this observation. Usually, one finds a different definition than 2.1.1,
namely the following: n is multiplicity-free if the von Neumann algebra
n(A)' is abelian. To see, briefly, that the two are equivalent, suppose first
that n has two nonzero orthogonal equivalent subrepresentations a 1 and
a2' By the preceding remarks there is a partial isometry U E n(A)' having
initial and final spaces respectively the range spaces of a 1 and a 2' Thus U· U
and U U· are orthogonal nonzero projections and in particular U and U·
do not commute, proving that n(A)' is not abelian. Conversely, if n(A)' is
not abelian then there is a self-adjoint operator T E n(A)' which does not
commute with n(A)'; an application of the spectral theorem yields a pro-
jection E with the same properties. Thus the range of E cannot be invariant
under every operator in n(A)' and so there exists S in n(A)' such that
(l - E)SE #- O. Now the initial and final spaces of (l - E)SE are clearly
orthogonal, and so the polar decomposition applied to this operator yields
a nonzero partial isometry V E n(A)' such that V·V and VV' are orthogonal
projections. Thus, the ranges of V· V and V V· give rise to two nonzero
orthogonal equivalent subrepresentations of n, proving our assertion.
From here on in this section, we assume all representations act on sep-
arable Hilbert spaces. The assumption is not essential for this part of the
theory, but it allows the development to run more smoothly, and moreover
the overwhelming majority of representations that arise in practice are on
separable spaces. Later on, in Chapter 4, this assumption will become more
essential.
We now make the following tentative definition.

Definition 2.1.2. Let n be a representation of A and let n be a cardinal, 1 ~

n ~ ~o. n is said to have multiplicity n if there is a set of n mutually ortho-
gonal, mutually equivalent, multiplicity-free subrepresentations {aJ of n
such that n = Lj a j.
According to the following proposition, the multiplicity of a representa-
tion is well-defined. First, some notation. If n is a representation of A on Ye
and n is a cardinal, 1 ~ n ~ ~o, define n' n to be the representation
n . n(x) = n(x) EB n(x) EB . .. on Ye EB Ye EB .. " where the number of
summands is n. Observe that a representation n has multiplicity n iff there is
a multiplicity-free representation a such that n - n' a.

Proposition 2.1.3. Let {J..li:i E I} and h:j E J} be two countable families of

subrepresentations of n, each satisfying the three conditions of 2.1.2, such
that n = L~i = LVj.
Then card I = card J.

PROOF. Let m = card 1, n = card J, and suppose, to the contrary, that

1 ~ m < n ~ ~o. Choose io E I andjo E J, and let.A and JV be the respec-
tive commutants of m . ~io(A) and n' vjo(A). From the hypothesis it follows

42
 2.1. From Type I to Multiplicity-Free

that m' Ilio and n' Vjo are each equivalent to 1t (cf. the remarks preceding
2.1.1), hence .A and % are unitarily equivalent to 1t(A)' and they are in
particular *-isomorphic as C*-algebras.
Now as in the proof of 1.2.1 the elements of.A can be realized as m x m
matrices over the abelian von Neumann algebra llio(A)' and similarly %
consists of all n x n matrices over vjo(A)' (assume n is finite; the infinite case
will be dealt with separately). Note first that .A admits a *-homomorphism
onto M m, the algebra of all m x m matrices over the complex numbers;
simply choose a nonzero homomorphism (J) of the commutative C*-algebra
llio(A)' into C and define t/lw:.A ~ Mn by applying (J) element by element to
each matrix over llio(A)'. Thus % too admits a *-homomorphism ({Jw onto
Mm. Let %0 be the subalgebra of % consisting of all n x n operator matrices
having scalar entries. Then %0 is clearly *-isomorphic with M n , and
({Jw(% 0) #- 0 (since %0 contains the identity), thus ({Jwl.vo has trivial kernel
because Mn has no nonzero proper ideals. In particular ({Jw(% 0) is an n2 _
dimensional linear subspace ofthe m2 -dimensional space M m, contradicting
m < n.
In the case where n is infinite, the argument can be adapted as follows.
Note first that % contains an isometry U (i.e., U*U = I) which is not
unitary (i.e., UU* #- J). Indeed, letting Jt be the range space of Vjo' then
n . Vjo = Vjo EB Vjo EB ... acts on Jt EB Jt EB .. " an infinite direct sum.
Thus the "unilateral shift" U:el EB e2 ...
1-+ 0 EB el e2
EB E
EB ... (ei Jt) is
a nonunitary isometry which commutes with n' vjo(A). Thus .A contains
an operator V such that V*V = J #- VV*. Now some entry of the nonzero
m x m operator matrix J - V V* is nonzero, hence there is acorn plex
homomorphism (J) of llio(A)' which is nonzero at that entry. Defining t/I w:.A ~
M m as before and putting W = t/I w( V) E M m' we see that W* W = J but
WW* #- 1. This, however, is impossible in Mno a fact easily seen by comparing
the trace of J - W*W to the trace of J - WW* and noting that a positive
matrix with zero trace must be zero. Thus n could not be infinite, and the
proof is complete. D

There are two more concepts which are basic to this point of view. The
first is Mackey's notion of disjointness; the second is the central cover of a
subrepresentation. Two representations 1tl and 1t2 are said to be disjoint
(1tl d 1t2) if no nonzero subrepresentation of 1tl is equivalent to any sub-
representation of 1t2' The term has its origin in set theory, and in Section
2.2 we will see that this is precisely a subtle refinement of the usual idea of
disjoint sets. It is clear that if 1tl d 1t2 and 1tl and 1t2 are two other represen-
tations equivalent, respectively, to 1tl and 1t2, then 1tl d 1t2' Let 1t be a
multiplicity-free representation and let 0"1 and 0"2 be two orthogonal sub-
representations of 1t. Then a brief consideration of the definitions involved
shows that 0"1 d 0"2' While this is no longer true if 1t is not multiplicity-free,
we will see presently (2.1.5) that the converse is: two disjoint subrepresen-
tations of a third are necessarily orthogonal. There are two characterizations

43
2. Multiplicity Theory

of disjointness which are frequently very useful, and we pause now to present
them.

Proposition 2.1.4. Let Trl and Tr2 be two representations of A. Then the
following are equivalent:
(i) Trl d Tr2'
(ii) The only bounded operator T between the respective Hilbert spaces
which satisfies TTrt(x) = Tr2(x)T, x E A, is T = O.
(iii) The projection 1 E9 0 belongs to the weak closure of Trl E9 Tr2(A).
PROOF. Let.Jff i be the space on which Tri acts, i = 1,2.
(i) => (ii). Assume Trl d Tr2, and let T:.Jff t --+ .Jff 2 satisfy the condition
TTrI(x) = Tr2(x)T, x E A. By taking adjoints, we conclude that the operator
T*:.Jff 2 --+.Jff l satisfies T*Tr2(X) = Trl(X)T*, XEA. These two conditions
imply that the positive operator T* T commutes with Tr I (A) and hence
ITI = (T*T)1/2 belongs to TrI(A)'. So if T = Vi TI is the polar decomposition,
we see that the partial isometry V implements an equivalence between the
subrepresentations of Trl and Tr2 defined by the initial and final spaces of
V. Since Trl d Tr2 we must therefore have V = 0, hence T = O.
(ii) => (iii). Assume (ii). By the double commutant theorem, it suffices to
show that 1 E9 0 commutes with every operator T on .Jff I E9 .Jff 2 which
commutes with all operators Tr 1(X) E9 Tr2(X), x E A. The obvious computations
with 2 x 2 operator matrices show that such a T has a representation

T = (Til T12)
T21 T22
where, for each i and j, Tij:.Jff j --+ .Jff i satisfies TijTrix) = Tri(x)Tij for every
x EA. Since Trl d Tr2, the off-diagonal terms must vanish (by (ii)), so that T
is a direct sum T = Til E9 T 22' The conclusion is now evident.
The implication (iii) => (i) is a simple variation of what was just done,
and will be left for the reader. 0

Now let (T be a subrepresentation of Tr and let vi! be the range space of (T.
It is easy to see that [Tr(A)'vI!] is an invariant subspace for both Tr(A)' and
Tr(A)" (and, in particular, for Tr(A)), so that its projection belongs to the center
of Tr(A)'. Let (j be the sub representation of Tr corresponding to [Tr(A)' vI!]; (j
is called the central cover of (T. If (T = (j (equivalently, ifthe projection on the
range space of (T already belongs to the center of Tr(A)') then (T is called a
central subrepresentation of Tr. Central subrepresentations are important
because, on the one hand, the projections on their range spaces commute
with everything in sight and, on the other, these projections belong to the
center of the weak operator closure of Tr(A) (indeed the latter is, by the double
commutant theorem, the center of Tr(A)" which in turn coincides with the
center of Tr(A)'). The relation between disjointness, equivalence, and central
covers is expressed in the following result.

44
 2.1. From Type I to Multiplicity-Free

Proposition 2.1.5. Let 0"1 and 0"2 be subrepresentations of a representation n.

(i) If 0" I is equivalent to a subrepresentation of 0" 2, then (j I ~ (j 2.
(ii) If 0"1 is multiplicity-free and (jl ~ (j2, then 0"1 is equivalent to a sub-
representation of 0"2.
(iii) If 0"1 and 0"2 are multiplicity-free and (jl = (j2' then 0"1 and 0"2 are
equivalent.
(iv) 0"1 d 0"2 if, and only if, (jl 1- (j2·
PROOF
(i) Let 9Ri be the range space of O"i' i = 1,2. Then by hypothesis, there is
a partial isometry U E n(A)' such that U· U is the projection on 9R 1 and
U9R 1 <;; 9R 2. Thus, n(A)'9R 1 = n(A)'U·U9R 1 <;; n(A)'U9R 1 <;; n(A)'9R 2, so
that [n(A)'9R I ] <;; [n(A)'9R 2], and it now follows from the definition of
central cover that (j I ~ (j 2.
(iv) Suppose first that (jl and (j2 are not orthogonal, and let Ei be the
projection on the range space of 0";, i = 1,2. Note that there is an operator
T E n(A)' such that E2 TEl =f O. For if E2n(A)'El = {O} then for all T 1,
T2 E n(A)' and ¢i E EiJlt' (JIt' being the underlying space) we have
(T 1¢1, T 2¢2) = (E 2TiT 1E 1¢1, ¢2) = 0,
hence [n(A)'E 1J1t'] 1- [n(A)'E 2J1t'], or (jl 1- (j2' a contradiction. Applying
2.1.4 (ii), we conclude that 0" 1 and 0"2 are not disjoint.
Conversely, assume (jl 1- (j2' and let Ili ~ O"i be such that 111 and 112 are
equivalent. By part (i) we have lil = li2. But since Ili ~ O"i clearly implies
lii ~ (ji, then by hypothesis lil 1- li2. Thus the only conclusion is that
lil = li2 = 0 = 111 = 112, and this proves 0"1 and 0"2 are disjoint.
(ii) Assume 0"1 is multiplicity-free and (jl ~ (j2. First, we claim that every
nonzero subrepresentation 11 of 0"1 contains a nonzero subrepresentation
equivalent to a subrepresentation of 0"2. Indeed, regarding 11 as a subrepre-
sentation of n, we have li ~ (j, ~ (j 2 and in particular li is not orthogonal
to (j2. Utilizing property (iv) already proved, the claim follows.
Observe next that, by an exhaustion argument, we may express 0" 1 = Lj Ilj
as an orthogonal sum of nonzero subrepresentations Ilj' each equivalent to
some subrepresentation III of (J2. (In more detail, Zorn's lemma provides a
maximal family {Il j} of nonzero orthogonal subrepresentations of 0" I, each
of which is equivalent to a subrepresentation of 0"2, and maximality plus the
preceding paragraph show that Lllj must be all of 0" ,.)
We claim now that III 1- Il~ if j =f k, which is the crucial step in the argu-
ment. Indeed, Ilj and 11k are orthogonal subrepresentations of the multiplicity-
free representation 0"1' so in fact we have Ilj d 11k. Therefore III d Il~ and by
part (iv) again we conclude III 1- Il~, and in particular III 1- Il~. Thus we may
form the subrepresentation Lj III of 0"2, which is clearly equivalent to
Lj Ilj = 0"" and (ii) follows.
(iii) Suppose both 0" 1 and 0"2 are multiplicty-free, and (j, = (j 2. By part (ii),
0" 1 is equivalent to a subrepresentation 11 of 0"2. If 11 =f 0"2, then by considering

45
2. Multiplicity Theory

the orthogonal complement of the range space of Jl in that of (J2 we obtain

a nonzero sub representation v ~ (J2, V 1.. Jl. Since (J2 is multiplicity free we
have in fact v d Jl, and thus v d (J 1, since Jl is equivalent to (J l' By (iv), v 1.. 0'1'
and since 0'1 = 0'2 we have v 1.. 0'2' Since v ~ 0'2 (because v ~ (J2) we con-
clude v = 0, hence v = 0, a contradiction. 0

In general, as one would expect, the process of forming direct sums tends
to increase multiplicity, while the multiplicity of subrepresentations tends to
get smaller. Nevertheless the next result asserts that, under certain conditions,
multiplicity is preserved in passing to subrepresentations or to direct sums.

Proposition 2.1.6
(i) Let n be a representation of A having multiplicity n. Then every nonzero
central subrepresentation of n has multiplicity n.
(ii) Let {(Jj:j E J} be a mutually disjoint family of subrepresentations of n
such that each (Jj has multiplicity n. Then Li (Jj has multiplicity n.
PROOF
(i) We may as well assume that n has the form n . Jl, where Jl is a multi-
plicity-free representation on a Hilbert space .Yt'. Therefore, each operator
n(x) can be represented as an n x n diagonal operator matrix of the form

Jl(X) 0 ... )
( .
o Jl(x)
n(x) = .
. .
A typical operator in the weak closure of n(A) has a similar form, with Jl(x)
replaced with an element T in the weak closure of Jl(A). Thus it is clear that
a central projection C in n(A)" has a matrix representation

where E belongs to Jl(A)" the center of Jl(A)". So if nc (resp. JlE) denotes the
subrepresentation of n (resp. Jl) defined by C (resp. E), then the preceding
shows that nc = n . JlE' The conclusion follows because JlE is multiplicity-free.
(ii) Assume first that n = 1, i.e., each (Jj is multiplicity-free, and let Jl be
the subrepresentation Lj (Jj. Let k = card J. Then in an obvious sense every
operator Jl(x), x E A, has a representation as a k x k operator matrix (Tij),
where Tij = 0 if i "# j and T jj = (Jj(x), A routine computation now shows
that every operator in the commutant Jl(A), has a representation (Sij), where
for each i, j, SiPj(X) = (Ji(X)Sij for every x E A. Because (Ji d (Jj if i "# j, we
see from 2.1.4 that Sij = 0 ifi "# j; moreover, each "diagonal" term Sjj must

46
 2.1. From Type I to Multiplicity-Free

belong to the commutative von Neumann algebra O"j(A),. This makes it

apparent that Il(A), is abelian, as required.
In general, assume each O"j has multiplicty n > 1. Then for each j we
may write O"j = Lk O"jb where for each j {O"jk: k E K} is an orthogonal family
of equivalent multiplicity-free subrepresentations of O"j' and card K = n.
For each k E K define 11k = Lj O"jk' By the preceding paragraph (and the
obvious fact that subrepresentations of disjoint representations are disjoint)
we conclude that each 11k is multiplicity-free, and clearly
Lk 11k = Lk Lj O"jk = Lj O"j.
Since the Il'S are all equivalent, L O"j has multiplicity card K = n. 0

We now define the class of representations to which multiplicity theory

applies.

Definition 2.1.7. A representation n is called type I if every nonzero central

subrepresentation of n contains a nonzero multiplicity-free subrepresen-
tation.

Thus, n is type I in case every central subrepresentation of n contains an

irreducible subrepresentation (this is always the case when n acts on a
finite-dimensional space); and if n is a factor representation (i.e., n(A)' has
trivial center) then it is type I if it contains an irreducible subrepresentation.
In the event A is abelian, then every representation of A is type I. For every
representation of A clearly contains a cyclic subrepresentation (a representa-
tion n of A on Jt' is cyclic if there is a vector ~ E Jt' for which [n(A)~] = Jt'),
and as we will see in the next section (corollary of2.2.4), a cyclic representation
ofan abelian C*-algebra is mutliplicity-free. More generally, we will see later
on that every GCR algebra has only type I representations (cf. Theorem
2.4.1).
We now come to a main result.

Theorem 2.1.8 Decomposition theorem. Let n be a type I representation

of a C* -algebra on a separable Hilbert space, and let N be the set of all
positive cardinals not exceeding ~o. Then there is an orthogonal family
{nn: n EN} of central subrepresentations of n such that
(i) nn is either 0 or has multiplicity n;
(ii) n = Ln nn'
The family {nn} is uniquely determined by these conditions. Moreover,
for each n E N, the central subrepresentations of n having multiplicity n
are precisely the central subrepresentations of nn'
PROOF. Fix n E N. If there is no central sub representation of n having
multiplicity n, set nn = O. Otherwise, by Zorn, we may choose a maximal
family {Ili} of nonzero orthogonal central subrepresentations of n such that

47
2. Multiplicity Theory

each Ili has multiplicity n. Put nn = Li Ili' nn is clearly a central subrepresen-

tation, and by 2.1.5 (iv) and 2.1.6 (ii) nn has multiplicity n.
Now if 0" is a central subrepresentation of nn then 2.1.6 (i) shows that 0"
has multiplicity n. On the other hand, if 0" is a central subrepresentation of n
which has multiplicity n, then we claim that 0" ::::; nn. For if the range space
of 0" is not contained in that of nn, then since their respective projections
commute, the range space of 0" has nontrivial intersection with the orthogonal
complement of the range space of nn' That intersection determines a nonzero
central sub representation of n, orthogonal to nn = Li Ili' which is also a
subrepresentation of 0" and therefore by 2.1.6 (i) has multiplicity n. But this
contradicts the maximality of {Il;}. Thus 0" ::::; nn as asserted. This proves the
last sentence ofthe theorem, and a moment's thought shows how the unique-
ness statement follows from it.
It remains to prove that Ln nn = n. Assume not, and let no be the central
subrepresentation of n defined by the orthogonal complement of the range
space of Ln nn· Because n is type I, no contains a nonzero multiplicity-free
subrepresentation 0". Utilizing Zorn's lemma once again, choose a maximal
family {O"/j E J} of mutually orthogonal subrepresentations of no, each
equivalent to 0". Note that J is countable, by the separability hypothesis.
We will produce a nonzero subrepresentation A. of LO"j which is central
(in n and therefore in ~)). This will complete the proof, for by 2.1.6 (i) A. has
multiplicity n = card J and so A. ::::; nn by the preceding paragraph, con-
tradicting the fact that by construction A. is orthogonal to every nn. To define
A., let Il be the subrepresentation of n defined by the orthogonal complement
of the range space of LO"j in no, and let A. be the central subrepresentation
defined by the orthogonal complement of the range space ofJi in no. Clearly
A. ::::; LO"j. If A. = 0, then Ji = no and hence (j ::::; Ji. By 2.1.5 (ii) 0" is equivalent
to a subrepresentation 0"' of Il, but then 0"' contradicts maximality of {O"j}.
Thus, A. "# 0, and the proof is complete. 0

We are now in a position to give the promised reduction from type I

representations to multiplicity-free representations. First, note that a repre-
sentation having multiplicity n uniquely determines its multiplicity free
summand to equivalence. That is, let n = Li Ili = Lj Vj' where the {Ili}
(resp. {Vj}) are mutually orthogonal, mutually equivalent multiplicity-free
subrepresentations. Then we claim: each Ili is equivalent to each Vj. For by
2.1.5 (i) and the fact that the Il'S are all equivalent we have Jii = Jib and in
particular Ili ::::; Jik for all i, k. Hence n = Llli ::::; Jik so that n = Jik for all k.
Similarly n = Vj for allj. By 2.1.5 (iii) it follows that Ilk and Vj are equivalent.
The most general type I represention (on a separable space) can be
constructed as follows. Let N be the set of all cardinal numbers n, 1 ::::; n ::::; ~o,
and let S = {O"n: n EN} be a sequence of separable representations of A
having the following properties:
(i) each
(ii) O"rn d
O"n
O"n
°
is or multiplicity-free;
if m "# n.

48
 2.2. Commutative C'-Algebras and Normal Operators

Of course, the Hilbert space Yfn of an may vary with n. For each n we form
the separable space n' Yfn = Yfn Et> Yfn Et> ... , the direct sum of n copies
of Yfn' and we consider the representation n . an on n' Yf n. Note that n' an
has multiplicity n or is the 0 representation. Finally, define a representation
ns on 1 . Yf 1 Et> 2 . Yf 2 Et> 3 . Yf 3 Et> ... Et> ~o . Yf ~o by
ns(x) = 1 . al(x) Et> 2· a2(x) Et> ... Et> ~o . a~o(x),
x E A. Because of the disjointness condition (ii) it follows (2.1.5) that the
subrepresentations nn = n . an of ns form the canonical central decomposi-
tion of ns described in Theorem 2.1.8.
Now let Sf = {a~:n E N} be another family satisfying (i) and (ii). Let us
say S and Sf are equivalent if an and a~ are equivalent for each n.1t is clear that
if S and Sf are equivalent then so are the representations ns and ns,. Con-
versely, suppose ns and ns, are equivalent. By replacing Sf with an equivalent
family we can assume ns, = ns. By the uniqueness statement in 2.1.8 we have
n' an = n' a~ for all n, and by the discussion preceding the last paragraph
this entails the equivalence of an and a~.1t follows that S and Sf are equivalent.
Finally, it is an obvious consequence of2.1.8 that every type I representation
of A has the form ns, for some sequence S satisfying (i) and (ii). To summarize,
we have proved the following
Corollary 2.1.9. S +-+ ns defines a bijective correspondence between equiva-
lence classes of sequences S satisfying (i) and (ii) and the equivalence classes
of type I representations of A on separable Hilbert spaces.

Thus the classification problem for type I separable representations is

reduced to the problem of classifying multiplicity-free representations. We
remark that with only minor variations the entire analysis ofthis section may
be carried out for representations on Hilbert spaces which are inseparable.

2.2. Commutative C*-Algebras and

Normal Operators
Let N be a normal operator on a finite dimensional Hilbert space Yf. A
familiar theorem from linear algebra asserts that there is an orthonormal
base for Yf with respect to which the matrix of N is diagonal, that is
n

N = I AjPj,
j= 1

where P 1, ... , P n are mutually orthogonal one-dimensional projections and

A1> ... ,An range over sp(N) (perhaps with repetitions). Observe that the
projections P j occuring in this representation are not unique (for example,
consider the identity operator as N). However, if we collect the P/s corre-
sponding to a single value of A E sp(N), then the expression becomes
N = I AE;.,
;'Esp(A)

49
2. Multiplicity Theory

where {E A:). E sp(N) } is now merely a mutually orthogonal family of nonzero

projections having sum 1. In this form the expression on the right is uniquely
determined by N (note that the E;.'s are the minimal spectral projections of
N), and we may define the multiplicity of), E sp(N) by mN().) = dim E;.- The
function mN: sp(N) -+ {I, 2, ... , dim JIf} is called the multiplicity function
of N.
Now let M be another normal operator on some finite dimensional space,
say
M = 2: AF;.-
Aesp(M)

Then it is not hard to see that Nand M are unitarily equivalent if, and only if,
sp(N) = sp(M) and mN(A) = mM().) for all A. Thus, the finite set sp(N) to-
gether with the multiplicity function mN form a complete set of unitary
invariants for finite dimensional normal operators N.
Now let N be a normal operator on an infinite dimensional Hilbert space.
Corresponding to the above expression we have the spectral theorem, which
asserts that N has a decomposition

N = f AE(dA),
JsP(N)

where E(') is a projection-valued Borel measure on sp(N). Notice however,

that it is no longer clear how to define the "multiplicity" of a point AE sp(N)
in the very frequent case where the range of E(') contains no minimal nonzero
elements; indeed this is precisely the point at which the technical difficulty
appears. Nevertheless, we will see that it is possible to define a function
analogous to mN above, which is uniquely determined almost everywhere
(E(d)')), such that this function, together with the set sp(N) and the measure
class determined by E('), form a complete set of unitary invariants for N.
This analysis is carried out in the somewhat more general context of
representations of commutative C*-algebras; thus, rather than analyze the
normal operator N, we analyze the identity representation of c*(N).
Let A be a separable commutative C*-algebra, and let X be the spectrum
of A. Then X is a locally compact Hausdorff space and the Gelfand map
identifies A isometrically and .-isomorphically with C(X), the C*-algebra of
all continuous complex-valued functions on X vanishing at 00 (1.1.1). By
separability of C(X), X is second countable, and a theorem of Urysohn [18]
implies that, in fact, X is metrizable as a separable complete metric space.
We begin by describing a construction giving multiplicity-free representa-
tions of C(X). Let J.l be any finite Borel measure on X. Since C(X) is separable,
it follows easily that the Hilbert space L 2(X, J.l) is separable. Let 8l(X) denote
the set of all bounded complex valued Borel functions defined on X. 8l(X)
becomes a commutative C*-algebra with the pointwise operations and the
norm Ilfll = supx If(x)l, containing C(X) as a C*-subalgebra. Each function
f E 8l(X) gives rise to a multiplication operator L f on L2(X, J.l), defined by
(Lf~)(x) = f(x)~(x), and the map f -+ L f is a representation of 8l(X) on

50
 2.2. Commutative C*-Algebras and Normal Operators

L 2(X, J.1.). Let us write nil. for its restriction to C(X). We will see that nil. is
multiplicity-free, and we will study the way in which nil. depends on J.1.
Let :!E denote the set of all multiplications L I' f E .?I(X). :!E is clearly a
commutative *-algebra of normal operators.

Theorem 2.2.1. :!E, :!E', and the strong closure of nll.( C(X», are identical.
PROOF. It is clear that ni C(X» ~ :!E ~ :!E'; we will prove that :!E' ~ :!E,
and :!E is contained in the strong closure of nll.( C(X». Choose T E :!E', and let
1 be the obvious constant function, regarded as an element of L 2(X, J.l). Then
Tl E L 2(X, J.1.), and we may choose a representative g in the equivalence class
of functions Tl. We claim: g is essentially bounded and T = L g • Indeed if
~,'1 E L2(X, J.1.) and ~ is bounded, then IJg(x)~(x)if(x) dJ.1.(x) I = I(L~T1, '1)1 =

I(TL~l' f/)I = I(T~, f/)I ~ IITII·II~112 ·11f/112. This inequality implies that
g(x)1 ~ IITII a.e. (J.1.) so by modifying g on a set of measure zero we can
assume Ig(x)1 ~ ilTl1 for all x E X. Thus g E .?I(X). Since Lg and Tare
bounded, Lg = T will follow if we show that Lg = T on the dense set of
all bounded functions in L 2(X, J.1.). But if ~ E .?I(X), then Lg~ = ~ . g = L~g =
L~Tl = TL~l = T~,asrequired. Thus:!E' ~:!E.
To see that each multiplication L g , g E .?I(X), is in the strong closure of
nll.( C(X)), choose a bounded sequence of functions f n E C(X) such that
fn(x) --+ g(x) a.e. (J.1.). The dominated convergence theorem implies that
fn(x)~(x) --+ g(x)~(x) in L2, for every ~ E L 2, and hence LIn --+ Lg strongly.

As an immediate consequence we see that the commutant of nll.( C(X))

is abelian (namely :!E" = :!E), and therefore nil. is multiplicity-free.
Recall that two finite Borel measures J.1., v are singular (written J.1. 1. v) if
there is a Borel set E ~ X so that J.1.(E) = v(X\E) = 0, i.e., J.1. is concentrated
on X\E and v is concentrated on E. J.1. ~ v means as usual that v(E) = 0
implies J.1.(E) = 0 for every Borel set E ~ X, and finally J.1. and v are equivalent
if they have the same null sets (i.e., J.1. ~ v and v ~ J.1.). It is evident that
for two singular measures J.1. and v, the only possible measure (J satisfying
(J ~ J.1. and (J ~ v is (J = O. Conversely, by applying the Legesgue de-

composition in a straightforward manner, the reader can easily see that if

J.1. and v are not mutually singular then there is a nonzero measure (J satisfying
(J ~ J.1. and (J ~ v. We come now to an elegant characterization of unitary

equivalence and disjointness in the class of representations {nll.}'

Theorem 2.2.2. Let J.1. and v be two finite Borel measures on X. Then nil. is
equivalent to nv iff J.1. and v are equivalent measures. nil. is disjoint from nv iff
J.1. 1. v.
PROOF. Assume first that J.1. and v are equivalent. We will find a unitary
operator U:L2(X, J.1.) --+ L2(X, v) such that Unll. = nvU, By the Radon-
Nikodym theorem and the fact that J.1. ~ v, there is a nonnegative function
hE Ll(V) such that dJ.1. = h dv. Moreover, v ~ J.1. implies that h > 0 a.e. (v).

51
2. Multiplicity Theory

For ~ E L2(X, 11), define (U~)(x) = (h(X))1/2~(X). Note that U maps L2(X, 11)
isometrically into L2(X, v); indeed for ~ E L2(X, 11) we have JIU~12 dv =
JI~12h dv = JI~12 dll = 11~112. The property U1I:1l = 1I:.U is evident from the
definition of U, and the easiest way to see that U is onto is to note by a
symmetric argument that the map V:'1 -+ h- 1/2'1 is an isometry of L2(X, v)
into U(X,Il) which is both a right and left inverse of U. Thus 1I:1l and 11:.
are equivalent.
Now assume 11 .1 v. We will prove that 1I:1l d 11:•• First, choose a Borel
set E such that Il(X\E) = v(E) = O. Let f be the characteristic function
of E, and consider the measure 11 + v. Then 4.1.1 shows that L f E 1I:1l +.(A)".
Now since 11 is concentrated on E and v is concentrated on X\E, we may
regard L 2(X, 11) and L 2(X, v) as L 2(E, 11) and L 2(X\E, v) respectively. Thus
L 2(X, 11 + v) can be identified with L 2(E, 11) E!3 U(X\E, v) = L 2(X, 11) E!3
L2(X, v), and this process identifies 1I:1l +. with 1I:1l E!3 11:. in such a way that
L f becomes the projection I E!3 0 of L2(X,Il) E!3 L2(X, v) onto its first
coordinate space. In short, we have shown that I E!3 0 belongs to (1I:1l E!3 1I:.)(A)".
By 2.1.4 (iii) we conclude that 1I:1l d 11:••
Conversely, suppose 1I:1l d 11:•• Now if 11 and v are not singular, then there
is a nonzero measure (1 such that (1 ~ 11 and (1 ~ v. We will show that 11:"
is equivalent to a subrepresentation of 1I:1l (and, by symmetry, 11:" is also
equivalent to a subrepresentation of 11:.). This will clearly contradict 1I:1l d 11:••
By the Radon-Nikodym theorem there is a nonnegative Borel function
hE Ll(X, 11) such that d(1 = h dll. Let E be the Borel set {x E X:h(x) > O}.
Now L 2(E,Il) is a closed subspace of L2(X, 11) invariant under 1I:1l(A), and
thus defines a sub representation of 1I:1l" We claim that this subrepresentation
is equivalent to 11:". Indeed for each function ~ E L2(X, (1), define a function
U~ on E by (U~)(x) = h(X)1/2~(X). Then we have

IE iU~(xW dll(x) = Ix 1~(xWh(x) dll(X) = Ix 1~(xW d(1(x),

and it follows that U is a unitary map of L 2(X, (1) onto L 2(E, 11)' Evidently
U1I:,,(f) = 1I:1l(f)U for all f E A, and the claim is proved.
It remains to prove that 1I:1l "" 11:. implies 11 and v are equivalent measures.
Contrapositively, let us assume for definiteness that 11 is not absolutely
continuous with respect to v. Then by the Lebesgue decomposition we can
write 11 = 110 + (1, where (1 .1 v, (1 =1= 0, and 110 « v. Of course (1 ~ 11. Now
by the preceding paragraph 11:" is a nonzero representation equivalent to a
subrepresentation of 1I:1l" On the other hand (1 .1 v implies 11:" d 11:., by an
argument we have already gone through. Thus 1I:1l contains a nonzero
subrepresentation disjoint from 11:., and thus 1I:1l and 11:. could not be
equivalent. 0

Thus the unitary equivalence class of a particular 1I:1l depends not on

the particular measure 11 but on the equivalence class of all finite measures
which are equivalent to 11. Let us call such an equivalence class of measures
a measure class. As a simple example, let 11 be an atomic measure; i.e., there
52
 2.2. Commutative C·-Algebras and Normal Operators

is a countable (or finite) subset {Xb X2, ... } of X such that jl({xJ) > 0 for
all i, and jl(X\{Xl> X2, .. .}) = O. In this case a measure v belongs to the
measure class determined by jl iff v shares these two properties with jl.
Thus the measure classes determined by atomic measures correspond 1-1
with the countable subsets of X. Two such measures classes are mutually
singular iff their corresponding countable subsets are disjoint. This explains
why, even in the nonatomic case, it is convenient to think of measure classes
as "generalized" subsets of X. Indeed the reader may wish to discover for
himself the precise sense in which the measure classes can be made into a
Boolean a-ring (without identity).
We will now show that the above family {nil} of multiplicity-free repre-
sentations exhausts all ofthe (separable) possibilities. First, a general lemma.

Lemma 2.2.3. Let A be a C·-algebra and let n be a nondegenerate multi-

plicity-free representation of A on a separable Hilbert space Yf'. Then
there is a vector ~ E Yf' such that [n(A)~] = Yf'.
Note that every nonzero n(A)-invariant subspace .A contains a
nonzero vector ~ such that [n(A)~] ~ .A (any ~ "# 0 in .A will do). By
Zorn's lemma, then, we may find a set gi: i E I} of unit vectors in Yf' which
is maximal with respect to the property n(A)~i .In(A)~j; for i "# j in I.
Note that Li [n(AK] = Yf' (for otherwise any unit vector in (Li [n(A)~;]).L
would contradict maximality of {~i}). Since Yf' is separable, I is countable,
and we may enumerate the gi} as a sequence ~b ~2' .... Let ~ = Ln rn~n·
To see that ~ is cyclic, it suffices to show that each ~n belongs to [n(A)~]
(because Yf' = Ln [n(A)~n]). But the projection Pn on [n(A)~n] belongs to
n(A)', and since n(A)' ~ n(A)" because n(A)' is abelian, we conclude from
the double commutant theorem that Pn belongs to the strong closure of n(A).
Thus Pn~ E [n(A)~], and since Pn~ = 2-n~n' the proof is complete. D

We remark that the converse of this lemma is valid for commutative

C-algebras, but false in general. See the corollary following 2.2.4.

Theorem 2.2.4. Let A be a commutative separable C -algebra and let n be a

nondegenerate multiplicity-free representation of A on a separable Hilbert
space. Then there is afinite Borel measure jl on X, the spectrum of A, such
that n is equivalent to nil"
PROOF. We may identify A with C(X) via the Gelfand map. Let Yf' be the
representation space of n. Then by 2.2.3 there is a nonzero vector ~ E Yf' such
that Yf' = [n(A)n Define a bounded linear functional p on C(X) by
p(f) = (n(f)~, ~). p is clearly positive [i.e., p(lfI2) = (n(f)·n(f)~, ~) =
Iln(f)~W ~ 0] so by the Riesz-Markov theorem there is a positive Borel
measure jl on X such that p(f) = fi(x) djl(x), f E C(X). Now regarding
C(X) as a dense linear subset of L 2(X, jl), we define a map U: C(X) --+ Yf'
by Uf = n(fK Note that II Ufl12 = Iln(f)~112 = (n(fK n(f)~) = (n(ifI2)~,
~) = p(lfI2) = SIfl 2 djl, so that U is an isometry. Since its domain and
53
2. Multiplicity Theory

range are dense it extends to a unitary map of L2(X, J1.) onto Yf. To see that
Unp' = nU, choose f, 9 E C(X). Then regarding 9 as an element of U(X, J1.)
we see that Unp.(f)g = U(f' g) = n(fg)~ = n(f)n(g)~ = n(f)Ug. The re-
quired formula follows because C(X) is dense in L2(X, J1.). 0

We remark that the above results used separability of A in only a super-

ficial way, and in fact the results are valid, in only slightly modified form,
for separable multiplicity-free representations of arbitrary commutative
C*-algebras. In place of Borel measures on X, for example, one deals with
finite regular Borel measures J1. which are "separable" in the sense that L 2(X,
J1.) is separable. One may also extend the results to representations on in-
separable Hilbert spaces, at the expense of finiteness of the measure J1.. For
the somewhat fussy details, see ([7], Chapter I, Section 7).

Corollary. Let Yf be a separable Hilbert space and let n be a representation

of a separable abelian C*-algebra on Yf. Then n is multiplicity-free if and
only if n has a cyclic vector.
An abelian von Neumann algebra f?t on Yf is maximal abelian (i.e.,
f?t' = f?t) if and only if f?t has a cyclic vector.

PROOF. Note that the second paragraph follows from the first. For by
Proposition 1.2.3, every abelian von Neumann algebra (Jl on Yf is the strong
closure of an abelian separable C* -subalgebra sI; and in this event fJt' = (Jl
iff f?t' = sI' is commutative iff the identity representation of sI is multiplicity-
free.
For the first paragraph, the "only if" part is Lemma 2.2.3, and the converse
follows from 2.2.4 and the remark following 2.2.1. 0

Let us now summarize these classification results, and apply them to

obtain a complete set of unitary invariants for normal operators on separable
Hilbert spaces.
Let A be a separable commutative C*-algebra, and let X be the spectrum
of A. Identify A with C(X). Let us construct a representation of C(X) as
follows. Choose a sequence J1.oo, J1.l> J1.2, ••• of finite Borel measures on X such
that J1.i 1.. J1.j if i '# j (some, but not all, of the J1.i are allowed to be 0). Owing
to 2.2.1 and 2.2.2, np'"" np'" nP.2' ... forms a sequence of multiplicity-free
disjoint separable representations of C(X), and thus n = 00 . np'", EB 1 . np', EB
2 . nP.2 EB ... defines a separable representation of C(X) which is exhibited
in its unique canonical decomposition as in 2.1.9. Moreover, by 2.1.9 and the
preceding results of this section, every separable representation of C(X) has
this form. Finally, given two sequences J1.oo, J1.l>' .. , VOO, VI" •• of finite mea-
sures as above, then the associated representations are equivalent if, and only
if, J1.n and Vn are equivalent measures, for every n = 00, 1,2, ... (this, by
2.1.9 and 2.2.2). Thus, the separable representations of C(X) are completely

54
 2.2. Commutative C·-Algebras and Normal Operators

classified to unitary equivalence by sequences Coo, C l , C 2 , ••• of (finite)

measure classes on X satisfying Ci 1- Cj if i =F j.
Now fix J1.00' J1.l, J1.2, ... as above. We want to sketch an alternate descrip-
tion of the representation n = 00 . n lloo EB 1 . n lll EB ... which is somewhat
more revealing. There is clearly no essential change if we renormalize each
J1.j so that Lj J1.j(X) < 00, and in this case J1.(E) = J1.oo(E) + J1.l (E) + J1.2(E) + ...
defines a finite Borel measure on X. Now since J1.i 1- J1.j if i =F j, we can find
a disjoint sequence E oo , E10 E2, ... of Borel sets having union X, such that
J1.i(F) = J1.(F n E;) for every Borel set F and every i = 1, 2, .... It follows
from the singularity properties of {J1.i} that J1.i(Ej) = 0 if i =F j, and moreover
the partition {E;} is uniquely determined by these properties modulo sets
of J1.-measure zero. If one thinks of En roughly as being the set on which the
multiplicity of n equals n, then the following construction should become
transparent. For each n, 1 :::; n :::; 00, choose an n-dimensional Hilbert space
fn (where of course 00 is taken as ~o). Let L2(En' J1.; fn) denote the vector
space of all functions ~: En --+ f n' which are Borel functions in the sense that
x 1--+ (~(x), u) is a Borel function for each u E fn' and which satisfy II~W =
SEn 11~(X)112 dJ1.(x) < 00. After identifying functions which agree almost every-
where (J1.), L2(En' J1.; fn) becomes a Hilbert space. As one would expect,
the direct sum Yf of these Hilbert spaces can be identified with the Hilbert
space of all Borel functions ~ from X into the disjoint union f 00 u flU
f 2 U ... ,which satisfy ~(x) E f n for x E En and for which the norm
1I~1I2 = Ix 1I~(x)1I2 dJ1.(x) is finite. Then for each f E C(X) we can define a
multiplication operator n(f) on Yf by (n(f)~)(x) = f(x)~(x), x E X, and
thus n is a representation of C(X) on Yf. (For the details, and related results,
see Section 4.2.) It is more or less obvious that n is equivalent to 00 . n lloo +
1 . n lll + 2 . n ll2 + .. " which is to say that the restriction of n to each
subspace L 2(En' J1.; f n) is equivalent to n . nlln , n = 00, 1, 2, ....
Now for each x E X, let us define the "multiplicity" m(x) of n at x as
follows, m(x) = k if x E Ek , k = 00, 1,2, .... Thus m(') is a Borel function
from X into the set {!X), 1,2, ... } of denumerable cardinals, for which
Ek = {x EX: m(x) = k}. Since the partition {Ed is unique modulo sets of
J1.-measure zero, the multiplicity function m(') is uniquely defined to a
J1.-equivalence. Conversely, given a positive finite measure J1. on X and a Borel
function m(- ):X --+ {!X), 1,2, ... }, we obtain a partition {Eoo, E10 E 2, ... } of
XbytakingE k = {xEX:m(x) = k},andthefamilY{J1.dofmutuallysingular
measures is recovered by taking J1.k(S) = J1.(Sn Ed, S s;: X, k = 00, 1,2, ....
This description of n is conveniently expressed in terms of direct integral
notation:
n = LEe m(x) . x dJ1.(x),

where the symbol n' x (n = 00, 1,2, ... , X E X) suggests the direct sum of
n copies of the one-dimensional representation x:f E C(X) --+ f(x) E C. Of
course we have not given a formal definition of direct integrals; the right side
ofthis equation suggests nothing more than the construction described above,

55
2. Multiplicity Theory

given a pair (J,l, m) consisting of a measure J,l and a multiplicity function m.

The results ofthis section can therefore be restated as follows. Every separable
representation of C(X) is equivalent to a direct integral ftl m(x) . x dJ,l(x), for
some measure J,l on X and some Borel multiplicity function m:X --+ {oo, 1,
2, ... }. Given two such pairs (J,l, m) and (v, n), then JEll m(x)' x dJ,l(x) and
JEll n(x) . x dv(x) are equivalent if, and only if, J,l and v are equivalent measures
and m(x) = n(x) almost everywhere.
We will now interpret these results so as to obtain unitary invariants for
normal operators. Let X be a fixed compact set in the complex plane and let
us classify all normal operators N having spectrum X which act on separable
Hilbert spaces. For such an N let A be the C*-algebra generated by N and the
identity. Let ( E C(X) be the function ((z) = z, Z E X. Then the functional
calculus provides us with a unique representation 1tN of C(X) satisfying
1tN(O = N, and if M is another such normal operator then M and N are
unitarily equivalent iff 1tM and 1tN are equivalent representations of C(X).
Applying the above to 1tN, we obtain a finite measure v on X and a multi-
plicity function n: X --+ {oo, 1, 2, ... } such that 1tN = J~ n(z) . z dv(z). In par-
ticular, note that this provides a representation for N as "multiplication
by C in a certain L 2-space of vector-valued functions on X. If we denote this
symbolically by
N = fxEll n(z) . ((z) dv(z),

then we may conclude: N is unitarily equivalent to M = J~ m(z)' ((z) dJ,l(z)

if, and only if, J,l and v define the same measure class (J,l ""' v) and their mul-
tiplicity functions agree almost everywhere [m(x) = n(x) a.e. (J,l)]. In the
special case where X is contained in the real axis, the above is equivalent to
the Hahn-Hellinger classification of bounded self-adjoint operators.

2.3. An Application: Type I von Neumann Algebras

Let ~ be a von Neumann algebra acting on a separable Hilbert space Yf.
Assume that ~ is a factor, i.e., the center ~ n ~' consists only of scalars.
A projection E E [Jl is called minimal if the only projections F E ~ satisfying
o ~ F ~ E are F = 0 or F = E. Note that since ~ is a factor this is equiv-
alent to the requirement that every subprojection of E have the form CE
for some central projection C E ~ n ~'. When [Jl is not a factor, the ap-
propriate counterparts of minimal projections turn out to be projections E
all of whose subprojections (in ~) are "central slices" of E, that is, of the
form CE with C a central projection. Now fix E and consider the family of
operators E~E = {ETE:TE~}. E~E is clearly a --algebra, and it turns
out to be weakly closed as well. The process of passing from ~ to E[JlE is
called localization. Now it is not hard to see (though we do not require this
fact) that the above condition on E is satisfied if, and only if, the local von
Neumann algebra E~E is abelian. When E has the latter property it is
called an abelian projection.

56
 2.3. An Application: Type I von Neumann Algebras

Here is a simple example of an abelian projection. Choose any abelian

von Neumann algebra sf on a Hilbert space Yf and let fJt be the von
Neumann algebra of all n x n matrices over sf, n = 1, 2, ... (if one wishes,
he may cause fJt to act in the obvious way as operators on the direct sum of
n copies of Yf). Then for 'every projection P E sf, the operator matrix E
defined by

is an abelian projection in fJt.

Definition 2.3.1. A von Neumann algebra fJt is called type I if every nonzero
central projection in fJt dominates a nonzero abelian subprojection in fJt.

Thus a factor is type I iff it contains a nonzero minimal projection. At

the other extreme, an abelian von Neumann algebra is always type I, for
everyone of its projections is abelian. We also remark that if one deletes
the term "central" from the above definition the resulting definition turns
out to be equivalent [7]. We will deduce the structure theory for type I
von Neumann algebras from the results of Section 2.1. The proofs are only
sketched, but we mention all the steps and the reader should be able to
fill in the details easily.
There is a simple relation between abelian projections in fJt and multi-
plicity-free representations of fJt'. For if E is an abelian projection in fJt,
then 11:E(T') = T'IEX defines a representation of fJt' on the range of E (note
that 11:E is a subrepresentation of the identity representation of fJt'), and we
claim that 11:E is multiplicity-free. This will follow if we show that 11:E(fJt')'
is contained in EfJtIEX, for the latter is clearly isomorphic to the abelian
algebra EfJtE. But if T E Sf(EYf) commutes with 11:E(fJt') = fJt'IEX then TE
is an operator on Yf whose range is contained in EYf and which commutes
with fJt' (thus TE E EfJt), and therefore T = TElEX E EfJtIEX, as required.
Though we shall not require it, the converse is also true, so that EfJtlEX equals
the commutant of fJt'IEX ([7], p. 18).
The following theorem implies that the commutant of a type I von
Neumann algebra is type I.

Theorem 2.3.2. Let fJt be a von Neumann algebra on a separable Hilbert space.
Then the following are equivalent:
(i) The identity representation of fJt is type I.
(ii) fJt is a type I von Neumann algebra.
(iii) The identity representation of fJt' is type I.
(iv) fJt' is a type I von Neumann algebra.

57
2. Multiplicity Theory

PROOF. By symmetry and the double commutant theorem it suffices to

prove (i) = (ii) = (iii). And since (ii) = (iii) follows from the preceding
remarks, we are left with (i) = (ii).
Assume (i), and let C be a nonzero central projection of [it. Applying
2.1.8 to the identity representation of [it, we find orthogonal central pro-
jections Coo, C b C 2 , •.. having sum I such that the subrepresentation
T E f7t f-+ TlcnJl" has multiplicity n whenever Cn #- O. Choose an n such that
CCn #- O. By passing from C to CC n and from Yf to CnYf we may assume
that the identity representation of f7t has multiplicity n. Thus there is a
Hilbert space Yf 0 #- 0 and a von Neumann algebra f7to on Yf 0 such that
f7to is abelian and such that Yf = Yf 0 Et> Yf 0 Et> .•. (n times) and f7t consists
of all operators of the form To Et> To Et> ... (n times) with ToE f7t o. Since
the map To E [ito f-+ To Et> To Et> ... E f7t is clearly a --isomorphism, the
problem has been reduced to finding a nonzero abelian projection in f7t o.
But that is easy. Choose any nonzero vector ~ E Yf 0, and let E E .!£'(Yf)
be the projection onto [[ito~]. Because the range of E is [ito-invariant, E
belongs to [ito = [ito. Moreover, since EYf 0 is a cyclic subspace for the
abelian von Neumann algebra [itOIEJI"o it follows from the corollary of 2.2.4
that the commutant of [itijlEJI"o is abelian (and in fact equals f7tijIEJI"o)' Thus,
to prove that E is an abelian projection in [ito it suffices to show that E[itoIEJI"
commutes with [itOIEJI'" But a moment's thought shows that is obvious,
and the proof is complete. 0

Now let us see how this leads to a structure theorem for type I von
Neumann algebras on separable Hilbert spaces. Fix f7t, a type I von Neumann
algebra. By 2.3.2 above, the identity representation of [it' is a type I repre-
sentation, and thus by 2.1.8 there exist central projections Coo, C b C 2 , ..•
in [it such that if [itn (resp. f7t~) denotes the restriction of [it (resp. f7t') to the
range of Cn, then the identity representation of [it~ is 0 or has multiplicity n.
Now [it is clearly the direct product
TIn[itn = {Too Et> Tl Et> T2 Et> ... :TnEf7tn,suPnIITnll < oo},
and each nonzero factor [itn has the property that its com mutant [it~ has
uniform multiplicity n. Such a von Neumann algebra is called homogeneous
of degree n, and we have just proved that every type I von Neumann algebra
is a direct product of homogeneous von Neumann algebras of distinct degrees.
The uniqueness assertion of 2.1.8 implies that this decomposition is unique.
Let us now consider a typical nonzero homogeneous factor [itn' As in the
proof of the preceding theorem, we may represent [it~ as the set of all operators
having the form To Et> To Et> ... (n times) where To ranges over the com-
mutant of an abelian von Neumann algebra !r n' Therefore [itn = f7t~ is
represented as the algebra of all n x n operator matrices whose entries
range over the abelian von Neumann algebra !rn (if n = 00 then of course
one considers only those matrices that represent bounded operators). Now
while different decompositions of the space will lead to different abelian

58
 2.4. GCR Algebras Are Type I

algebras in place of fZ n, 2.1.5 (iii) implies that they are all unitarily equivalent.
Conclusion: every homogeneous von Neumann algebra of degree n "is" the
algebra of all (bounded) n x n matrices over an abelian von Neumann algebra
which is unique to unitary equivalence. Thus we have a complete description
of separably acting type I von Neumann algebras. As we have already
suggested, there are analogous results in the inseparable case, and all the
results of this section can be generalized in that direction. For an alternate
and more general treatment, along with much additional information about
type I von Neumann algebras, see [7]. The results of this section are due
to Kaplansky [17].

2.4. GCR Algebras Are Type I

It is customary to call a C-algebra A type [ if every (non degenerate)
representation of A is a type I representation. Unfortunately' this usage is
in direct conflict with the terminology of the preceding section: a type I
von Neumann algebra is almost never a type I C*-algebra. For example,
2(Jf') is a prototypical type I von Neumann algebra in the sense of 2.3.1.
While on the other hand, if Jf' is infinite dimensional then 2(Jf') is known
to have non type I representations, and so does not qualify as a type I C-
algebra in the above sense. Nevertheless, both usages of the term have
become so widespread that there is now no alternative but to learn to live
with them. As it happens, it is almost always obvious in context just what
is intended.
In any case, if one wants to use the multiplicity theory of Section 2.1 to
analyze representations, it is important to know that the underlying C*-
algebra one has is type I. We conclude this chapter by sketching the proof
of a theorem of Kaplansky [15J which implies that this is true for all the
C*-algebras that were encountered in Chapter 1.

Theorem 2.4.1. Every (nondegenerate) representation of a GCR algebra is

type [.
PROOF. Let A be a GCR algebra. Now a non degenerate representation n
of A is type I iff the identity representation of n(A)" is type I. According to
2.3.2 then it suffices to show that n(A)" is a type I von Neumann algebra;
and a moment's thought shows that this reduces to proving that n(A)"
contains a nonzero abelian projection. Since n(A) is isomorphic to the
quotient A/ker n, it is GCR, and thus we are reduced to proving the following
assertion: every nonzero separable GCR algebra .91 acting on a Hilbert
space has the property that .91" contains a nonzero abelian projection E.
To sketch the construction of E, note that there is a nonzero element
FEd such that a(F) is compact for every irreducible representation a of .91
(any nonzero element in the CCR ideal ofd will do). Actually, more is true.
One can even find F =I 0 in .91 such that a(F) has rank 0 or 1 for every

59
2. Multiplicity Theory

irreducible representation a (for the details, see [6], Lemma 4.4.4). Choose
such an F; by replacing F with F* F we can also assume F is positive. Consider
the subalgebra F d F of d. Since a(F) has rank 1 or 0 for every such a, all
operators of the form a(F)a(X)a(F) (X E d) commute with one another,
and therefore a(X Y - Y X) = 0 for all X, Y E F d F and every a. Since
the intersection of the kernels of all irreducible representations of d is
trivial (corollary of 1.7.2) we see that FdF is in fact abelian, and therefore
F d" F is abelian since it is contained in the weak closure of F d F.
Now if F is a projection then we are done. If it is not, then choose B > 0
so that the spectrum of F meets the interval [2B, + 00], and define a Borel
function g:[O, + (0) --> IR as follows: g(x) = I/x if x ~ B, g(x) = 0 if
o ~ x < B. Then g(F) belongs to .91", it commutes with F, and moreover
E = F g(F) is a nonzero projection such that £.91 E is abelian. 0

There is a converse to this theorem of Kaplansky's, namely, if a separable

C*-algebra A has only type I representations (or more simply if the factor
representations of A are all type I) then A is GCR. This converse is a deep
theorem of Glimm (see [13], or Section 9 of [6]).

60
 3
Borel Structures

In this chapter we have assembled certain nontrivial results on Borel

structures which are needed in the analysis of representations of noncom-
mutative C·-algebras. Since no exposition of this material exists in a form
appropriate for our purposes and since the results are useful in other areas
of mathematics, we have arranged that this chapter can be read independently
of the others.

3.1. Polish Spaces

A Polish space is a topological space which is homeomorphic to a separable
complete metric space. In this section we collect a few properties of Polish
spaces for later use. Note first that a countable direct product Pi of Polish 0
spaces Ph P 2, . . . is Polish (indeed, if di is a metric on Pi defining its topology,
then
00

d({Xi}, {y;}) = I
i~ 1
r i d(x;, Yi)/(l + d(Xi' Yi)

is a metric which defines the topology on Oi PJ Similarly, a countable direct

sum of Polish spaces is Polish. A closed subspace of a Polish space is ob-
viously Polish, but it is less clear that the same is true of open subspaces:

Lemma 3.1.1. An open subspace of a Polish space is Polish.

PROOF. Let G be an open set in P, and let d be a metric on P. Define a function
f:G --+ IR by f(x) = l/d(x, P\G) [d(x, S) of course denotes the distance from
x to the set S]. Since f is continuous on G, x 1-+ (x, f(x)) is a homeomorphism
of G onto the graph of f, a subspace of the Polish space P x IR. Thus it
suffices to show that the graph of f is closed in P x IR, by the preceding

61
3. Borel Structures

remarks. But if Xn E G, x E P, t E IR are such that Xn --+ x and f(xn) --+ t, then
f(xn) is bounded so that x cannot belong to p\G. Thus x E G, and by con-
tinuity t = limn f(xn) = f(x). 0

Recall that a subset of a topological space is called a GlJ if it is a countable

intersection of open sets.

Theorem 3.1.2. Let P be a Polish space and let S £: P be a GlJ. Then S is a

Polish space in its relative topology.
SKETCH OF PROOF. Let Gj £: P be a sequence of open sets such that S =
nGj. By the lemma, we can find Polish spaces P j and homeomorphisms
kP j --+ Gj • Define Q to be the set of all sequences (Pl> P2," .) E flP j such
that fl (Pl) = f2 (P2) = .... Q is clearly a closed subset of the Polish space
fl Pi> and is therefore Polish. Define a function f: Q --+ P as follows:
f(Pl, P2,' .. ) = fl(Pl)' A moment's thought shows that f maps Q onto
nG j = S, and it is easy to verify that f is 1-1, continuous, and its inverse
is continuous (because every /; has each of these properties). Thus S is
Polish. 0

A familiar result from the theory of metric spaces asserts that every closed
set is a GlJ. Thus the preceding theorem includes closed subspaces as well as
open ones. But these two cases of course fall far short of exhausting the
possibilities; for example the set S of irrationals forms a GlJ in the set of real
numbers, and therefore S is a Polish space. Finally, it is an interesting fact
that the converse of 3.1.2 is true: if a subspace S of a Polish space is Polish,
then S is a GlJ ([4], p. 197).
We come now to an important technical concept. Let X be a topological
space. For every k ~ 1 and every k-tuple of positive integers nl> ... , nk, let
Anln2 ... nk be a subset of X. The family {Anln2 ... n.} is called a sieve for X if the
following properties are satisfied:
(i) U:: = 1 AnI = X.
(ii) Uf = 1 AnI'" nk( = AnI'" nk,for every k ~ 1 and every nl> ... , nk ~ 1.
A sieve is called open if each AnI'" nk is an open set. The best way to under-
stand what sieves are is to construct one:

Proposition 3.1.3. Let P be a Polish space and let d be a metric defining the
topology of P. Then P admits an open sieve {AnI'" nk} with the following
properties:
(iii) diam(A nl ... nk) ~ 1/ k,for all k, nl> ... , nk;
(iv) AnI' "nk t £: AnI" 'nk,fo r all n l , . . . , nk,t.
(v) Each AnI' .. nk is nonempty.
PROOF. Since P is separable we can find a sequence A l , A 2 , ••• of nonempty
open balls of diameter ~ 1 whose union is P (for example, let {Xl' X2, ... }

62
 3.l. Polish Spaces

t
be a countable dense set and let Ai be the open ball of radius centered at xJ
This gives us {AnJ. Now fix n1 and consider the subspace An" In a similar
manner we can find nonvoid open balls An,!> An,2' ... in An, whose union
is An" whose diameters are all ~ t, and whose closures are all contained in
An" Thus we now have {A n,n2 }, and the proof is completed by an obvious
induction. 0

Now let {AnI" .nJ be an open sieve with the properties (iii), (iv), and (v)
above. Note first that the elements of the sieve form a base for the topology
of P. Indeed if x E P, 6 > 0, and B is the open ball about x of radius 6 (the
metric is of course the same as the one used in 3.1.3), then for any k we have
P = U {An, ... nk :n1 ~ 1, ... ,nk ~ I}, and hence there exist n1" .. , nk such
that x E An, ... nk; since diam(An, ... nJ ~ 11k we also have An"" nk £;; B so
long as we choose k larger than 116, proving the assertion. Second, note that
for every sequence n1' n2,' .. of integers, the sequence of non void closed sets
An" Anln2 , An,n2n3' ... is decreasing, and by (iii) their diameters shrink to
zero. So by the Cantor nested sets theorem for complete metric spaces, they
intersect in a unique point. By (iv) this intersection is nk'= 1 An", 'nk' and
thus the latter intersection contains one and only one point.

EXAMPLE. Let N = {t, 2, ... } be the set of positive integers, endowed with
the discrete topology. Then N is a Polish space (consider the metric
(j(m, n) = 0 or 1 according as m = n or m '" n), and thus the countable
Cartesian product N°O = N x N x ... is also a Polish space. Here is one
good metric on N°O: for two sequences {md, {nd E N°O let d( {md. {nd) =
Lk=1 2- k (j(mk' nk)' For each k-tuple of integers n b " " nk define Un" " nk =
{{mi} E N oo :m1 = n1"'" mk = nk}' Then {Un, ... nJ is a basis for the
topology on N°O, and note that it also forms an open sieve.

The following result implies that N°O is a "universal" Polish space in the
sense that every Polish space is homeomorphic to a quotient space of N°O.

Theorem 3.1.4. For every Polish space P, there is a continuous open mapping
of N°O onto P.
PROOF. Let {An, ... nJ be an open sieve for P having the properties of 3.1.3.
We will define a functionf:N°O -.. P as follows. Choose v = (nb n2,"') E N°O.
By the remarks following 3.1.3, we may define f(v) as the unique element of
nk'Now
= An""let nk'{U n, ... nJ be the system of open sets in N°O described above.
1

Since {U n"" nJ (resp. {AnI'" nJ) forms a base for the topology of N°O
(resp. P) and since the diameter of each An", 'nk is ~ 11k, a few moments
thought shows that all the desired properties of f (i.e., continuity openness,
surjectivity) will follow provided that we prove f(U n, ...nJ = An"" nk' for
every k ~ 1, n1, ... , nk ~ 1. For that, let v E Un, .. 'nk' Then v has the form
v = (n1' ... ,nk, nk+ 1•••. ) and so f(v) E n.i= 1 An"" nj £;; An", 'nk' proving

63
3. Borel Structures

one inclusion. Conversely, if x E An, ... nk' then An, ... nk = U

':'= 1 An, ... nkl
implies that there is an nk+ 1 so that x E An, ... nk+ " Repeating the argument
on An, ... nk+' gives nk+2 so that x E An, ... nk+2' and so on inductively. Thus
we can find a sequence nk + 1, nk + 2, ... so that x E An, ... nj for j ~ k. This
implies x E nk"= 1 An, ... nk (recall that the sequence Ek = An,.·. nk is de-
creasing), and therefore x = f(v) where v = (n1' n2," .). Since v E Un""nk'
we are done. 0

3.2. Borel Sets and Analytic Sets

Let P be a Polish space. A Borel set in P is a member of the a-field gen-
erated by the closed subsets of P. If f is a continuous map of P into another
Polish space Q, then while the inverse images under f of closed sets are
always closed, this is of course not so for direct images, even if f is one-to-
one. Indeed f(P) is in general not even a Borel set in Q. A remarkable theorem
of Souslin, however, implies that a one-to-one continuous function does map
closed sets to Borel sets (and therefore Borel sets to Borel sets). Thus it is
not too surprising that this result has great significance in the theory of
Borel structures. We will give a proof of Souslin's theorem in this section,
and in the next we will extract numerous consequences from it.
A subset of a Polish space P is called analytic if it has the form f(Q),
where Q is a Polish space and f is a continuous map of Q into P. First, we
want to show that Borel sets are analytic sets of a special kind.

Lemma. Let $' be a family of subsets of a set S, containing Sand 0, and

which is closed under countable intersections and countable disjoint unions.
Let $'0 = {E E $' :S\E E $'}. Then $'0 is a a-field.
PROOF. Clearly 0 E $'0, and $'0 is closed under complementation. So $'0
will be a a-field provided we show that it is closed under countable inter-
sections. But if En E$'o then both En and S\En belong to $', so that nn En E $',
and S\nn En = Un (S\E n) = Un
(S\En n En- 1 n ... n E2 n Ed is a count-
able disjoint union of elements of $'. Thus S\ nn
En E $', and hence nn
En E
~. 0

Theorem 3.2.1. Let P be a Polish space. Then for every Borel set E s; P,
there is a Polish space Q and a 1-1 continuous function f:Q --+ P such
that E = f(Q).
PROOF. Let $' be the class of all subsets of P having the stated property.
We will show that family of all subsets E, such that both E and P\E belong
to $', contains the Borel sets. For that, according to the preceding lemma,
it suffices to show that $' contains the open sets, the closed sets, and is
closed under countable intersections and countable disjoint unions. It is
clear that $' contains the open sets and closed sets since they are themselves
Polish subspaces.

64
 3.2. Borel Sets and Analytic Sets

Now let Eb E z , ... belong to ff. We claim: nnEn E ff. Choose Polish
spaces Pn and continuous 1-1 functions fn:Pn -+ P such that fn(P n) = En"
Define a function f:On Pn -+ P X P X •.. by f(P1, Pz,···) = (f1(pd,
fz(pz), ... ). Then f is obviously continuous and 1-1. Since the set ,1 =
{(XbX2, ... )EP X P X ···:x 1 = Xz = X3 = ... } is closed, Q = f-1(,1)
is a closed subspace of the domain of f, and is therefore Polish. Note that
f is an injection of Q onto {(x, x, x, .. .):x E nn En}, so that if PI is the
projection of P x P x ... onto its first coordinate space, then the compo-
sition PI fl.1 is the required continuous 1-1 map of Q onto
0 nn
En E ff.
Finally, if En E ff and Em (\ En = 0 for m #- n, then choosing Pn and
fn: Pn -+ P as above, form the Polish space Q = Ln Pn and define f: Q -+ P
by f(x) = f,,(x) if x E Pn. f is continuous, it is 1-1 because the En are dis-
joint, and f(Q) = Un En· That shows Un En E ff. D

This theorem is a slight weakening of a result of Souslin's, which asserts

that every Borel set is the continuous 1-1 image of a zero-dimensional
Polish space ([19], p. 447).

Corollary. Every Borel set in a Polish space is analytic.

The converse ofthis corollary is false (c.f. [19], p. 479, for an example).
The next result, due to Lusin ([19], p. 485), provides a key step in the
proof of 3.2.3 below, and is of considerable interest by itself. First, some
terminology. Two disjoint subsets A, B of a Polish space are said to be
separated if there are disjoint Borel sets E, F such that A ~ E, B ~ F. If
An, Bn are two sequences of subsets such that every Am is separated from
every B n , then some elementary manipulations with countable intersections
and unions (which we leave for the reader) show that Un An and Un Bn are
separated.

Theorem 3.2.2 Separation theorem. Two disjoint analytic sets in a Polish

space are separated.
PROOF. Let P be a Polish space and let A and B be disjoint analytic subsets
of P. By 3.1.4, every analytic set is the continuous image of N°O, so there
are continuous functions f, g: N°O -+ P such that A = f(N°O), B = h(N°O).
Assume now that A and B are not separated. Let {V nl ... nJ be the basis
for N°O described in the discussion preceding 3.1.4. Since N°O = Un Vn we
have A = Un f(V n) and B = Un g(V n), so by the preceding remarks there
exist integers mb ni such thatf(VmJ and g(VnJ are not separated. Similarly
V rnl = Ut V mll and V nl = Ul V nlf , so there exist mz, nz so that f(V mlrn2 )
and g(Vnln2 ) are not separated. Continuing inductively, we find mb mz, ... ,
nb nz, ... such that f(V ml ... mJ and g(V nl .. . nJ are not separated for every
k ~ 1. Define points fJ., v E N°O by fJ. = (mb m2, ... ), v = (n1' nz, . .. ). Now
f(fJ.) #- g(v) because A and B are disjoint, so there exist disjoint open sets

65
3. Borel Structures

U, V such that f(Jl) E U and g(v) E V. Now for each k, Um, ... mk is a neigh-
borhood of Jl for which diam Um , .. . mk --+ 0 as k --+ 00. Thus by continuity
we have f(Um, ... mJ £; U for large enough k. Similarly g(Un, . .. nJ £; V for
large k. But this contradicts the fact that by construction f(U m , . .. mJ is not
separated from g(U n , .. .•J. D

Note the close parallel between this argument and the customary proof
of the Bolzano - Weierstrass theorem.

Corollary 1. Let S be a subset of a Polish space P. If both Sand P\S are

analytic, then S is a Borel set.

We can already deduce a weak form of Souslin's theorem (3.2.3 below):

Corollary 2. Let P and Q be Polish spaces, and let f be a continuous 1-/

function of Ponto Q. Then f maps closed sets to Borel sets (and therefore
Borel sets to Borel sets).
PROOF. Let F £; P be closed. Because f is 1-1 and onto, f(F) and f(P\F)
are complementary sets in Q. Since F and P\F are Polish subspaces (cf. 3.1.1)
and since f is continuous, f(F) and f(P\F) are both analytic sets. An appli-
cation of Corollary 1 finishes the proof. D

The next corollary follows from 3.2.2 by routine manipulations with

countable unions and intersections, which we omit.

Corollary 3. Let An be a disjoint sequence of analytic subsets of a Polish

space. Then there are disjoint Borel sets En such that An £; En, for every n.

Before proceeding with the main theorem, we need some preliminaries.

A partition of a set S is a family & of disjoint subsets of S for which U& =
S. A partition &2 is said to be a refinement of a partition & I (this is written
& 2 :s;; &1) if each set in &2 is contained in some set in &1. Note that this
is equivalent to requiring that every set in &1 be a union of sets in &2. We
want to point out now that every Polish space P contains a decreasing
sequence &1 ~ &2 ~ ... of countable partitions such that each element of
&n is a G., whose diameter (relative to a fixed metric on P) is at most lin.
Indeed, as in the proof of 3.1.3, we can cover P with a sequence B I , B 2 , •••
of open balls having diameter :S;;1.Let&1 = {A b A2, ... },whereA I = Bb
An = Bn\(BI U ... U B n- 1 ), n > 1. Each An is a G." being an intersection of
an open set with a closed set, and thus we have & I. Inductively, given & I, ... ,
&n as above, choose a sequence C b C 2 , ••• of open sets having diameter
:s;; lin + 1 such that Un Cn = P, and put DI = Cl> Dn = Cn\(C I U ... U
Cn-I), n ~ 1. Then &n+1 = {A 11 Dk:A E &n, k = 1,2, ... } has all the
stated properties (note for example that A 11 D is a G., when both A and
Dare).

66
 3.2. Borel Sets and Analytic Sets

Theorem 3.2.3. Let P, Q be Polish spaces and let f: P --+ Q be a 1-1 con-
tinuous function. Then f(P) is a Borel set in Q.

PROOF. Choose a sequence f!> 1 ~ f!> 2 ~ . . . of countable partitions of P as

in the preceding remark, say f!>n = {Anlo A n2 , . .. }, n ~ 1. Now since each
set Ank is a Gd it is a Polish subspace of P. Thus for each n, U(A n1 ), f(A n2 ), ... }
forms a disjoint sequence of analytic sets in Q, and by Corollary 3 above there
are disjoint Borel sets Bn!> Bn2 , ... such that f(Ank) £; Bnk , k = 1, 2, ....
By replacing Bnk with Bnk n f(A nk )- we can assume f(Ank) £; Bnk £; f(Ank) -.
Moreover, by intersecting each element of {Bn+llo Bn+12' ... } with an
appropriate element of Bnlo Bn2 , . .. } (utilizing finite induction on n) we can
assume that each nonvoid Bn+ 1 k is contained in a unique member of {Bnl'
Bn2 , ... }, n ~ 1. Note that this implies An+ 1 k £; Ani if, and only if, Bn+ 1 k £;
Bni , for all n, j, k ~ 1.
We will show that f(P) = n:= Uk'=
1 1 B nk . This will complete the proof,
for the right side is clearly a Borel set. The inclusion £; is obvious, so choose
a point q in the right hand member. Then for each n there is a kn so that
q E Bnk n. Thus Bnk n n Bn+ 1 kn+' =I 0 and so Bn+ 1 kn+' £; Bnkn . By the pre-
ceding paragraph we have An+ 1 kn+' £; A nkn · Also q E Bnk n £; f(AnkJ-. Thus
A lk " A2k2 , ••• forms a decreasing sequence of non void closed sets in P whose
diameters shrink to O. By the Cantor nested sets property they intersect in a
unique point p. To see that q = f(p), let V be any closed neighborhood of
f(p). By continuity at p and the fact that diam Ankn --+ 0 we will have f(AnkJ £;
V for large enough n, and therefore q E Bnk n £; f(AnkJ- £; V for large n.
Since V was arbitrary this means q = f(p). 0

Corollary. Let f:P --+ Q satisfy the hypotheses of the preceding theorem.
Then f maps Borel sets to Borel sets.
PROOF. 3.2.3 implies that f maps closed sets to Borel sets; since f is 1-1
and the closed sets in P generate its Borel structure, the conclusion is im-
mediate. 0

The final result in this section is a measurability theorem for analytic

sets, adapted from ([19], p. 482). Afinite Borel measure on a Polish space P
is a mapping J1 from the Borel sets of Pinto [0, (0) such that J1(0) = 0 and
J1(Un En) = Ln J1(E n) for every sequence Elo E 2, ... of disjoint Borel sets.
A subset S £; P is J1-measurable if there exist Borel sets Bl £; S £; B2 for
which J1(B 2\Bd = O. This is equivalent to the existence of Borel sets B, N
such that S £; Band B\S £; N, where J1(N) = o.

Theorem 3.2.4. Let A be an analytic set in a Polish space P. Then A is J1-

measurable for every finite Borel measure J1 on P.

PROOF. Fix J1, a finite Borel measure, and let S be an arbitrary subset of P.
A Borel set E containing S will be called minimal if for every other Borel

67
3. Borel Structures

set F :::2 S one has f-L(E\F) = O. Note that every subset S has a minimal Borel
cover; for if we choose Borel sets En :::2 S so that f-L(En) tends to (1. = inf{f-L(F):
F Borel, F :::2 S}, then E = nn
En is a Borel cover of S for which f-L(E) = (1.,
and thus it is minimal because for every Borel set F :::2 S, En F :::2 S implies
f-L(E n F) ~ (1., hence f-L(E\F) = f-L(E) - f-L(E n F) ~ f-L(E) - (1. = O.
Now let A be an analytic set in P. We will produce Borel sets B, N such
that A £; B, B\A £; N, and f-L(N) = O. As in the proof of 3.2.2 there is a
continuous function f: N OO --+ P such that A = f(N OO ). Let {Un, ... nJ be the
basis for N OO described in the discussion preceding 3.1.4. For each k ~ 1
and nb"" nk ~ 1, choose a minimal Borel cover En""n. for f(Un""n.)'
By intersecting with f(Un""n.)- if necessary, we can assume

f(U n," 'n.) £; En, .. ·n. £; f(U n," 'n.)-·

Define B = U~ = 1 En" and define

It is clear that both Band N are Borel sets, and A £; B is immediate from
A = f(N OO ) = Un, f(U n,) £; B. Thus we must prove f-L(N) = OandB\A £; N.
Now since Un,··· n. = Uy;"
1 Un, ... n.f we have

00 00

f(Un""n.) = U f(Un""n.f)
(=1
£; U En""n.("
(=1

By minimality we conclude that f-L(E n, .. 'n.\U?= 1 En, .. 'n.f) = 0 for every

k ~ 1, n1' ... , nk ~ 1, and thus N, being the countable union of sets of
measure zero, has measure zero. To prove B\A £; N, choose P E B\A. Then
there exists n1 ~ 1 such that P E En,\A. Assume, contrapositively, that P ~ N.
Then in particular p ~ En,\U?= 1 En,!, and since PEEn, we can only have
P E U?= 1 En,f' Thus there is an n2 such that P E En,n2' Similarly P ~ En,n2\Uy;" 1
En,n2f (because P ~ N) and since P E En,n2 there must be an n3 so that p E En,n2 n3'
Continuing in this way we obtain a sequence n 1, n2, ... of positive integers
such that PEEn,. "n. for every k ~ 1. Define v E N OO by v = (nb n2, ... ).
n.
Now for each k Un, ... is a neighborhood of v for which diam Un, ... --+ 0 n.
as k --+ 00. Thus by continuity at v we have nk'=l f(Un""n.)- = {.f(v)}.
Since PEEn, .. 'n. £; f(U n, .. 'n.)- for every k, it follows that P = f(v) E A,
and this contradiction completes the proof. 0

Thus while analytic sets are not necessarily Borelian, at worst they differ
from Borel sets in a metrically trivial way. Note also that the preceding
theorem implies that a good many nonanalytic sets are measurable for
every finite measure f-L. Indeed since the wmeasurable subsets of P form a
a-field, it follows that every set in the a-field generated by the analytic sets is
measurable. In particular the complement of an analytic set A is measurable,
while by Corollary 1 of 3.2.2 P\A will not be analytic if A is not a Borel set.

68
 3.3. Borel Spaces

Finally, observe that this theorem is also true for a-finite measures, since
every a-finite measure is mutually absolutely continuous with a finite
measure.

3.3. Borel Spaces

In this section we deduce some rather striking consequences of the
preceding material. Most of these results are due to Mackey [20]. A Borel
space is a pair (X, BiJ) where X is a set and BiJ is a a-field of subsets of X,
whose elements we will call Borel sets. Thus every topological space X
gives rise to a natural Borel structure BiJ, the a-field generated by the closed
subsets of X. Unless there is cause for confusion, we will suppress the letter
BiJ in the notation for Borel spaces. A subspace of a Borel space X is a subset
S £; X endowed with the relative Borel structure, namely the a-field of all
subsets of S of the form S n E where E is a Borel subset of X. A mapping
f:X ~ Y between Borel spaces X and Y is a Borel function if f- 1 (E) is a
Borel set in X for every Borel set E £; Y. f is a Borel isomorphism if it is
1-1, onto, and both f and f- 1 are Borel functions (we have resisted an
impulse to use the term Borelomorphism). There are natural definitions for
the Cartesian product and direct sum offamilies of Borel spaces; for example
if Xl, X 2, ..• is a sequence of Borel spaces then the Borel structure in
[1n Xn is the a-field generated by all finite "rectangles" of the form {(Xl'
X2, .. ,):Xl EEl, ... , Xn E En}, where Ei is an arbitrary Borel subset of Xi'
1 ::;:;; i ::;:;; n, n = 1,2, .... Finally, a Borel space X is countably separated if
there is a sequence E1> E 2 , ••• of Borel subsets of X which separates points
of X (i.e., if X i= y then there is an n such that xdx) i= Xdy), XE denoting
the characteristic function of E).
We begin with a simple and useful lemma.

Proposition 3.3.1. Let X, Y be Borel spaces, with Y countably separated, and

let f:X -+ Y be a Borel function. Then the graph of f is a Borel subset
of X x Y.
PROOF. Let E 1 , E 2 , ••. be a separating family of Borel sets in Y. Let .,1 =
{(y, y):y E Y} be the diagonal in Y x Y. Then (Y x Y)\.,1 is the union of
the two sets Un En X (y\En) and Un (y\En) X En> so that .,1 is a Borel
subset of Y x Y. Now the function q>: X x Y -+ Y x Y defined by q>(x, y) =
(f(x), y) is Borel (since each coordinate is a Borel function), and since the
graph off is simply q> -1 (.,1), the proof is complete. 0

A Borel space X is called standard if X is isomorphic to a Borel subset of

a Polish space. A noteworthy fact, which we do not need, is that every
uncountable standard Borel space is isomorphic to the unit interval [0, 1]
with its usual Borel structure ([19], remark (i), p. 451). Thus a great variety
of mathematical constructions lead to identical Borel spaces. Nevertheless,

69
3. Borel Structures

as we will see presently, standard Borel spaces do not quite suffice for dealing
with problems in analysis involving equivalence relations.
The following generalizes Souslin's theorem (Theorem 3.2.3).

Theorem 3.3.2. Let X be a standard Borel space, let Q be a Polish space, and
let f be a I-I Borel map of X into Q. Then f(X) is a Borel set in Q and
f is an isomorphism of X onto f(X).
PROOF. We will show that f(X) is a Borel set; the rest follows from this and
the obvious fact that a Borel subset of the standard Borel space X defines
a standard subspace.
By definition, we may assume that X is a Borel subset of a Polish space
Po; and by 3.2.1. there is a I-I continuous map g of a Polish space Ponto
X. Thus f(X) = f g(P) and fog is I-I Borel function from Pinto Q.
0

Hence we may even assume that X = P is a Polish space.

Now Q is countably separated (take E 1 , E 2 , • •• to be a countable base
for the topology of Q) so that by 3.3.1. the graph of f is a Borel set in the
Polish space P x Q. Utilizing 3.2.1. again, we can find a Polish space R
and a I-I continuous function h: R -> P x Q such that h(R) = graph f.
The projection P2 of P x Q onto Q is continuous, its restriction to graph
f is 1-1 (since f is 1-1), and thus P2 h is a continuous I-I function from
0

R into Q whose range is f(P). By Souslin's theorem 3.2.3 f(P) is a Borel

sct. 0

Let X be a Borel space and let", be an equivalence relation in X. Let

XI'" be the set of all equivalence classes. Then the function q: X -> XI'"
which assigns to each point of X the equivalence class containing it maps
onto XI"', and thus determines a Borel structure on X I'" : by definition a
subset E of XI'" is a Borel set if q-I(E) is a Borel set in X. As an alternate
description, let us say a Borel subset B 5; X is saturated if B is a union of
equivalence classes, i.e., x E Band y '" x implies y E B. Then E +-+ q-I(E) is
a bijective correspondence between the Borel sets in XI'" and the saturated
Borel sets in X. Note that q is a Borel map of X onto XI"'. Here is a simple
example illustrating the kind of equivalence relation one encounters.

Example 3.3.3. Consider the classical problem of classifying complex

irreducible n x n matrices with respect to unitary equivalence. Fix n, a
positive integer, let Mn denote the set of all complex n x n matrices, and
let ~n be the set of irreducible members of Mn (a matrix is irreducible if it
commutes with no self-adjoint projections other than 0 and the identity).
Mn is a Polish space in its natural topology (for example, the distance between
two matrices (aij) and (bij) can be taken as Li.i laij - bijl and it is not very
hard to see that ~n is a Gb in Mn. Indeed, x E Mn is irreducible if the C*-
algebra generated by x is all of M n and, choosing Xo to be a fixed irreducible
element of M n , this is true iff the C· -algebra generated by x contains Xo'

70
 3.3. Borel Spaces

If we let P 1 (e, 11), P 2 (e, 11), ... be the countable set of all noncommutative
polynomials in the two free variables e,11 having rational coefficients, then
for each i ~ 1 the function x ~ Pi(X, x*) is continuous, hence each set
Gj = Ur; 1 {x E Mn:dist(xo, Pi(X, x*)) < Ijj} is open, and sO..?n = nf=
1 Gj
is a Go. By 3.1.2,..?n is a Polish space. Define - in ..?n by x - y iff x and y
are unitarily equivalent. Thus the quotient Borel space ..?nl - can be regarded
as a classification space for irreducible n x n matrices. Now the classification
problem, that is, the problem of finding a complete set of unitary invariants
for..? n' amounts to finding a convenient parametrization of ..?nl - in terms
of the familiar spaces of analysis. More precisely, one wants a countable
set of "well-defined" (i.e., Borel) functions from ..?nl - into, say, the real or
complex numbers (or in general some Polish space), which separates points
in ..?nl -. The existence of such a family of functions is easily seen to be
equivalent to saying ..?nl - is countably separated, and thus a knowledge of
the Borel structure of ..?nl - is of key importance. We shall return to this
example after Corollary 2 of 3.3.5 below.

It often happens that the quotient of a standard Borel space by a very

regular equivalence relation fails to be standard. The relevant definition is
the following. A Borel space X is called analytic if there is an analytic set A
in a Polish space such that X is isomorphic to A with its relative Borel
structure. Note that by the corollary of 3.2.1 each standard Borel space is
analytic. The converse is, however, false; see ([ 19], p. 479) for some interesting
examples. Thus while there is only one uncountable standard Borel space
modulo isomorphism, this is not true of uncountable analytic Borel spaces.
So far as we know, the latter have not been classified. We now prove the key
invariance property of analytic Borel spaces.

Theorem 3.3.4. Let X be an analytic Borel space, let Q be a Polish space,

and let f be a Borel map of X into Q. Then f(X) is an analytic set in Q.
PROOF. The argument is similar to that of 3.3.2. We can assume X is an
analytic subspace of a Polish space Po. Then by definition there is a con-
tinuous map of a Polish space P into Po having range X. Thus we may
assume (by considering the composition of the two functions) that X = P
is itself Polish.
By 3.3.1 the graph off is a Borel set in P x Q, and so there is a continuous
(in fact, 1-1) function 9 from a Polish space R into P x Q having range
graph f. Letting P2 be the projection of P x Q onto Q, we see that P2 0 g:
R .... Q is continuous and has range f(P). Thus f(P) is analytic. 0

The next result shows that the analytic subspaces of Polish spaces are
what you think they are.

Corollary 1. Let X be a subspace of a Polish space P. Then X is an analytic

Borel space in its relative structure if, and only if, X is an analytic set in P.

71
3. Borel Structures

PROOF. The if part is obvious. Conversely, suppose there exists an analytic

set Y in a Polish space Q and a Borel isomorphism f of X on Y. Now apply
3.3.4 to f - 1 to conclude that X is an analytic set. 0

We remark that Theorem 3.3.2 has a corollary analogous to Corollary 1;

the proof is the same.

Corollary 2. Let X and Y be analytic Borel spaces and let f be a /-/ Borel
map of X onto Y. Then f is a Borel isomorphism.
PROOF. Again we may assume X and Yare analytic subspaces of Polish
spaces P and Q. Let E be a Borel set in X; we have to show that f(E) is a
Borel set in Y. E has the form X (\ B for some Borel set B ~ P. Thus E is
analytic (here we use the fact that a countable intersection of analytic sets
is analytic, the proof of which is a trivial adaptation of an argument in the
proof of 3.2.1), so by 3.3.4 f(E) is an analytic set in Q. Similarly f(X\E) is
analytic, and f(E) (\ f(X\E) = 0 because f is 1-1. By the separation
theorem (Theorem 3.2.2) there is a Borel set C ~ Q such that f(E) ~ C
and f(X\E) ~ Q\C, which shows that f(E) = f(X) (\ C = Y (\ C is a Borel
~~~ 0

We remark that Corollary 2 remains valid if we replace the hypothesis

that Ybe analytic with the weaker requirement that Ybe countably separated.
See Corollary 1 of 3.3.5. The next result is an elegant Borel space analogue
of the Stone-Weierstrass theorem. First, we need a lemma. A Borel space
(X, 81) is said to be countably generated if there is a countable subset of (fI
which separates points and generates 81 as a a-field. It is easy to check that
every generating subfamily of (fI must separate points of X.

Lemma. A Borel space is isomorphic to a subspace of a Polish space if, and

only if, it is countably generated.
PROOF. The only if part is trivial, since every Polish space is countably
generated (consider a countable base for the topology) and hence so is every
subspace of it. Conversely suppose E l , E 2 , . . . are a generating family for
the Borel structure on X. Define P to be the cartesian product On P n , where
P n is the discrete two-element set {O, I}, n = 1,2, .... Then P is a Polish
space and we have a map cp:X --+ P defined by cp(x) = (XEJX), XE2(X), ... ),
XE denoting the characteristic function of E. cp is 1-1 because {En} separates
points. Moreover if we define Fn ~ P by Fn = {(Xl' X2,"') E P:xn = I},
then {Fb F2""} generates the Borel structure in P, so that {cp(X) (\ Fn}
generates the Borel structure in cp(X). Finally, since cp(E) = cp(X) (\ Fj , it
follows that cp is a Borel isomorphism of X on cp(X). 0

Theorem 3.3.5. Unique structure theorem. Let (X, 81) be an analytic Borel
space and let 81 0 be a countably generated sub-a-field of 81 which separates
points in X. Then 81 0 = (fl.

72
 3.3. Borel Spaces

PROOF. By the lemma, there is a Borel isomorphism f of (X, gBo) onto a

subspace Y of a Polish space P. Note that the Borel structure on Y is f(gBo).
Now f is also a 1-1 Borel map of (X, gB) into (Y, f(gBo)) (because fJU contains
gBo), so by Theorem 3.3.4 Y is an analytic set in P, and by Corollary 2 f
is an isomorphism of (X, fJU) on (Y,f(gBo)). In particular, gB = f -1(f(gBO)) =
fJU o, completing the proof. 0

We remark that the preceding result is false without the hypothesis that
gBo be countably generated. Here is a simple example. Let (X, gB) be the
standard Borel space consisting of the unit interval [0, 1] with its usual
Borel structure and let gBo be the sub a-field of all countable subsets of
[0, 1] together with their complements. Clearly gBo separates points, but
of course gBo #- gB. Evidently, gBo is not countably generated and in fact
no countable subfamily of gBo can separate points.
The next result implies that while the range of a standard Borel space
under a Borel map may fail to be standard, it is often no worse than analytic.

Corollary 1. Let X be an analytic Borel space, let Y be a countably separated

Borel space, and let f be a Borel map of X onto Y. Then Y is an analytic
Borel space.
PROOF. Let {El' E2' ... } be a separating family of Borel sets in Y. We
claim: {El' E 2, . .. } generates the Borel structure in Y. For this choose an
arbitrary Borel set F in Y, and let gB °be the a-field generated by {F, E 1,
E 2 , • •• }. Since (Y, gBo) is countably generated we may by the preceding
lemma regard (Y, fJU o) as a subspace of a Polish space P. By 3.3.4 Y is an
analytic set in P, thus (Y, gBo) is an analytic Borel space, thus by the unique
structure theorem gBo is generated by {Et. E 2 , • •• }. Since F was arbitrary,
this proves the claim.
Thus Y is countably generated, and so the above argument may be re-
peated to show that Y is analytic. 0

We may apply this to equivalence relations as follows.

Corollary 2. Let X be an analytic Borel space and let ~ be an equivalence

relation in X. Assume there is a sequence fl' f2' ... of real valued Borel
functions on X such that for any pair of points x, y in X one has x ~ y
iff f,,(x) = f,,(y) for all n. Then X/~ is an analytic Borel space.
PROOF. Let q:X --+ X/~ be the natural projection. Define gn:X/~ --+ ~
by gn(q(x)) = fn(x). Then each gn is a Borel function and the g's separate
points in XI ~. So if we let rt. r2, ... be the rational numbers in ~, then
the family of sets Emn = {z E X I ~ :gn(z) < rm} separate, and the conclusion
follows from Corollary 1. 0

Note, finally, that the converse of Corollary 2 is valid, since every analytic
Borel space is countably separated.

73
3. Borel Structures

Let us return to Example 3.3.3 and examine the Borel structure of the
classification space Jnl"'. As an application of the preceding corollary we
will show that J nl '" is analytic (in fact, it is standard, as 3.4.1 and 3.4.2
in the next section will show). Let 1(1 be the set of all finite words w(e, 1])
e,
in the two free variables 1]; thus w(e, 1]) = e2 1] 3 e1] is a typical element of
1(1. Let tr denote the canonical trace on Mn. For each w(e, 1]) E 1(1 define a
complex valued continuous function fw:Jn -+ e by fw(x) = tr(w(x, x*».
Clearly {fw: WE 1(1} is countable, and each fw is constant on equivalence
classes because the trace is unitarily invariant. So the set of all real and
imaginary parts of the functions fw will satisfy the requirements of the
corollary provided we show that fw('x) = fw(y) for all WE 1(1 implies x '" y,
for all x, y E I n • Fix x, y and assume the condition is satisfied. Then by
taking linear combinations we have tr(p(x, x*» == tr(p(y, y*» for every
e
noncommutative polynomial p(e, 1]) in the two free variables and 1]. We
claim that the map n:p(x, x*) ~ p(y, yO) is well-defined. For if p is a poly-
nomial for which p(x, x*) = 0, then p(x, x*)* p(x, x*) = 0, and since
p(x, x*)*p(x, x*) is another polynomial in x, x* we have tr(p(y, y*)*p(y, y*» =
tr(p(x, x*)*p(x, x*» = tr(O) = 0, and hence p(y, yO) = 0 (because tr(z*z) = 0
implies z = 0). By symmetry n is also 1-1. Note also that n preserves the
algebraic operations and the *-operation. Thus n is a *-isomorphism between
the CO-algebras generated by x and by y. Since both x and y generate Mn
as a C*-algebra (by irreducibility) n is in fact a *-automorphism of Mn.
Therefore (cf. Corollary 3 of 1.4.4) there is a unitary operator u E Mn such
that n(z) = uzu* for all z, and in particular y = n(x) = uxu*, proving x '" y.
By Corollary 2 above, Ynl'" is an analytic Borel space.
Note in particular that the functions fw(x) = tr(w(x, x*» form a complete
set of unitary invariants for J n' a fact first proved by Specht [26]. c. Pearcy
has subsequently shown that for each n one can find a finite subset of
{fw: W E ff} which separates J nl '" [22]. There is of course nothing that
distinguishes this particular set of invariants; for example if we let Pl(e, 1]),
P2(e, 1]), ... be the countable set of all noncommutative polynomials in the
e,
two free variables 1] having rational coefficients, then an argument similar
to the above shows that the sequence of continuous functions gn(x) =
IIPn(x, x*)11 (lIzll denoting the operator norm of z, regarding z as an operator
on the Hilbert space en) provides another countable separating family for
Jnl"'. Debates over the relative merits of different sets of invariants in the
general case have little substance; in practice one chooses the most con-
venient invariant parameters one can find in context.
.3.4. Cross Sections
Let X be a set and let f be a function from X onto another set Y. A cross
section for f is a function g: Y -+ X such that fog is the identity map id y .
Note that a cross section assigns to each point y E Ya unique element g(y)
in the set f -1( {y}); thus the existence of cross sections in general follows from

74
 3.4. Cross Sections

(and in fact is equivalent to) the axiom of choice. For most purposes, however,
this is not good enough for dealing with problems in analysis. For example,
if X and Yare topological spaces and f is continuous, one might want to find
a continuous cross section. Unfortunately, this is usually impossible (consider
the map f(x) = eix of the real line onto the unit circle), and the best one can
hope for is a Borel cross section. While even that is not always possible, there
is an effective theorem, due to Dixmier. which provides the Borel cross
sections required for the analysis of representations of GCR C*-algebras.
After giving a proof of Dixmier's theorem, we conclude this chapter with a
discussion of the general problem. It goes without saying that none of these
results depend on the axiom of choice.
We begin by taking another look at the space NCO (see the discussion
preceding 3.1.4). Define an order relation < in NCO by (mb m 2 , ••• ) < (nb
n2' ... ) iff the first k for which mk #- nk satifies mk < nk (note that this is the
"dictionary" order). If we define J1. ~ v to mean J1. < v or J1. = v, then ~ is
clearly a linear ordering of Nco. We claim that in fact this is a well-ordering in
the sense that every nonvoid closed set in NCO has a smallest element. Indeed
if F #- 0 is closed, define nl to be the smallest first coordinate of all points
in F, define n2 to be the smallest second coordinate of all points in F having
the form (nb *, *, ... ) (note that this set is nonvoid), n3 is the smallest third
coordinate of all points in F of the form (nb n2, *, *, ... ), and so on indefi-
nitely. Clearly v = (nb n2, ... ) is smaller than or equals any point in F, and
v is clearly a limit point of F; i.e., v E F. Thus NCO is well-ordered by ~.
For vEN co , let us write (-00, v) for the interval {J1.EN co :J1. < v} and
[v, + (0) for its complement {J1. E N co :J1. ~ v}. We claim first that each interval
(- 00, v) is open. Indeed if J1. = (ml, m2,' .. ) and v = (nb n2,' .. ) satisfy
J1. < v, then there is a k ~ 1 for which ml = n b · · · , mk-l = nk-b mk <
nk. Thus all points in the open set Um""m. = {(ii' i2, ... ):il = m b •.. ,
ik = md are smaller than v, proving (- 00, v) is open. Second we claim that
the intervals ( - 00, v) generate the Borel structure. Since the sets of the form
n.
Un, ... (as above) form a countable base for the topology, it suffices to show
that each Un, .. 'n. has the form (- 00, J1.) n (N°O\( - 00, v)) = ( - 00, J1.) n
[v, + (0). But for any integers n l , . . . , nk' it is clear that J1. = (nb ... , nk-l'
nk + 1, 1, 1, ... ) and v = (nb ... , nk- b nk, 1, 1, ... ) have the required
property, proving the claim.
Following is a variation of a theorem of Dixmier [9]. Its proof, however,
is more related to a selection theorem of von Neumann ([21], Section 15,
lemma 5).

Theorem 3.4.1. Let P be a Polish space, let Y be a Borel space, and let f be
a function from Ponto Y satisfying
(i) f maps open sets to Borel sets;
(ii) the inverse image of each point of Y is a closed subset of P.
Then f has a Borel cross section.

75
3. Borel Structures

PROOF. Assume first that the theorem has been proved for the special case
P = N°O. Then for a general P we can find a continuous open map g of N°O
onto P (3.1.4). Now the function f 0 g:N°O ----> Y satisfies (i) because g is an
open map, and it satisfies (ii) because g is continuous. Thus by assumption
there is a Borel cross section h for fog, and so go h provides the required
Borel cross section for f.
Let us now prove the theorem for P = N°O. For each y E Y, f - I( {y}) is
closed in N°O, thus by the well-ordering property f-I({y}) has a smallest
element denoted h(y). Clearly h: Y ----> N°O is a cross section for f, and we
must prove that h is Borelian. Since the open intervals (- 00, v) generate
the Borel structure it suffices to prove that h - 1 ( - 00, v) is a Borel set in y,
for each v E N°O, and by property (i) this will be true if we show h- I ( - 00, v) =
f( - 00, v). For s, note that h -I (E) s f(E) for every set E S N°O (Indeed
z E h-I(E) implies h(z) E E implies z = f 0 h(z) E f(E), because h is a cross
section). For :2, h 0 f(Jl) is the smallest element in f - 1 ({f(Jl)}), and in par-
ticular h 0 f(Jl) ~ Jl, Jl E N°O. Thus h 0 f( - 00, v) s (- 00, v) and hence
f(-oo, v) S h-I(-oo, v). D

Let S be a set and let G be a group of transformations of S onto itself

(thus we assume G contains the identity map, that each y EGis invertible,
and that G is a group under composition). Then we may define an equivalence
relation in S by p - q iff there exists Y E G such that y(p) = q, and the
quotient SI - is called the orbit space of S (mod G). The following cross
section theorem is the one we need for the discussion of C-algebras.

Corollary. Let P be a Polish space, let G be a group of homeomorphisms of

P, and let - be the equivalence relation determined by G. Assume that each
orbit G(p) = {y(p):y E G} is closed, pEP. If PI - has the quotient Borel
structure, then the canonical projection q: P ----> PI - has a Borel cross
section.
PROOF. Taking Y = PI - in 3.4.1, we see that q: P ----> PI - satisfies condi-
tion (ii), by hypothesis. For (i), let U be open. To see that q(U) is a Borel
set, we must show that q-l(q(U)) is a Borel set in P. But the latter is simply
UYEG y(U), an open set. The conclusion is now immediate. D

The next result implies that if a Borel map f:X ----> Y has a Borel cross
section, then Y can be no more pathological than X.

Proposition 3.4.2. Let X and Y be Borel spaces, with X countably separated,

and let f be a Borel map of X onto Y. Assume that f has a Borel cross
section g: Y ----> X. Then g( Y) is a Borel set in X and g is a Borel isomorphism
of Y onto g(Y). In particular, if X is a standard Borel space then so is Y.
PROOF. Let F S X be the set of all fixed points of the Borel function
go f:X ----> X. One sees easily that F = g(Y), and we claim that F is a Borel

76
 3.4. Cross Sections

set in X. Indeed, because X is countably separated, the diagonal L1 =

{(x, x):x E X} is a Borel set in X x X (see the proof of 3.3.1). Therefore
since x --+ (x, g f(x)) is a Borel map of X into X x X we conclude that
0

F = {x E X:(x, g f(x)) E L1} is a Borel set.

0

Finally, g: Y --+ F and fIF:F --+ Yare Borel maps inverse to one another,
so in particular g is an isomorphism of Y onto F. 0

Note that the set g(Y) in 3.4.2 is a Borel transversal for the equivalence
relation x "" y iff f(x) = f(y); i.e., it is a Borel set in X which meets each
equivalence class in a single point.
These are the only cross section results required for our present purposes,
and the reader may omit the rest of this section without essential loss.
We want to discuss "usable" cross sections in general, and to draw a sig-
nificant conclusion about why they may fail to exist. In the preceding section
we saw that if a Borel space Y is the range of, say, a standard Borel space X
under a Borel map f: X --+ Y, then Y is analytic if and only if it is countably
separated. Thus one might expect to have a Borel cross section for f when Y
is countably separated. However, if Y is merely analytic but not standard,
then 3.4.2 shows that f has no Borel cross section. One might then conjecture
that Borel cross sections exist when both Y and X are standard. Unfor-
tunately, even that is not true; D. Blackwell has given an example of a Borel
subset X of the unit square [0, 1] x [0, 1] such that if f(x, y) = x for
(x, y) E X, then f(X) = [0, 1] (a Polish space), while f does not have a Borel
cross section [3].
In spite of all this, one can do almost as well in the case where Y is countably
separated. The concept we need has to do with measurability. Let (X,~)
be a Borel space. A subset A £ X is called absolutely measurable if A is
J.l-measurable for every finite Borel measure J.l on X; equivalently, for every
finite Borel measure J.l there exist Borel sets Ell' Fil such that Ell £ A £ Fil
and J.l(F11\EIl ) = 0. The family d of all absolutely measurable sets is a a-field
containing ~, indeed it is the intersection (Ill d 11' where d 11 denotes the
a-field of all J.l-measurable subsets of X, and J.l runs over all finite Borel
measures. Because analytic sets in Polish spaces are absolutely measurable
(3.2.4) it follows that d is generally larger than ~. A function g from one
Borel space into another is called absolutely measurable if the inverse image
under g of every Borel set is absolutely measurable; equivalently, g is J.l-
measurable for every finite Borel measure J.l on the domain of g. Thus, abso-
lutely measurable functions are amenable to measure theoretic analysis,
and from that point of view are just as good as Borel functions. We now prove
that in all "reasonable" cases, absolutely measurable cross sections do exist.

Theorem 3.4.3. Let X be an analytic Borel space, let Y be a countably sep-

arated Borel space, and let f be a Borel map of X onto Y. Then f has an
absolutely measurable cross section.

77
3. Borel Structures

PROOF. By Corollary 1 of 3.3.5 Y is an analytic Borel space, so there is no

loss if we assume that X and Yare analytic subspaces of Polish spaces P
and Q. The map x ~ (x, f(x) ) E P x Q is a Borel map of X onto the graph
of f, so by 3.3.4 graph f is an analytic subset of P x Q. Thus graph f is the
continuous image of a Polish space and by 3.1.4 we may take the latter to be
N°O; so let cp:N°O -. P x Q be continuous, cp(N°O) = graph f. Finally, let
P2 be the projection of P x Q onto Q. Then P2 0 cp: N°O -. Q is continuous
and has range Y. We will find an absolutely measurable cross section
g: Y -. N°O for P2 0 cpo Granting that for a moment, let PI be the projection
of P x Q onto P. Then note that the function PI 0 cp 0 9 maps Y into X,
it is absolutely measurable (because PI 0 cp is Borel and 9 is absolutely
measurable), and it is a cross section for f because f 0 PI = P2 on graph f
and P2 0 cp 0 9 = id y implies f 0 (PI 0 cp 0 g) = id y • Thus the theorem will
follow if we can find such a function g.
Now by continuity (P2 0 cp)-I({y}) is closed in N°O, for each yE Y, and
as in 3.4.1 we define g(y) to be the smallest element in that set. Then 9 is a
cross section for P2 0 cp and, exactly as in 3.4.1, 9 - I( - 00, v) = P2 0 cp( - 00, v)
for every v E N°O, and we need only prove that P2 0 cp( - 00, v) is an absolutely
measurable set in Y. Since ( - 00, v) is open and therefore a Polish subspace,
P2 0 cp( - 00, v) is an analytic set in Q and thus measurable for every finite
measure on Q (3.2.4). So if p. is any finite Borel measure on Y, then
P2 0 cp( - 00, v) is measurable with respect to the Borel measure a(E) =
p.(E n Y) on Q, and therefore it is p.-measurable as well. 0

We conclude that if "" is an equivalence relation in an analytic Borel

space X for which XI"" is countably separated, then the canonical projection
q: X -. X I"" has an absolutely measurable cross section.
We now take up a common situation in which measurable cross sections
do not exist. Some terminology will be convenient. Let"" be an equivalence
relation in a Borel space X; though it is not essential we will assume for
simplicity that each equivalence class c(x) = {y E X:y "" x} is a Borel set.
Let 9' be the sub a-field of all "" -saturated Borel subsets of X (see the discus-

°
sion preceding Example 3.3.3). A Borel probability measure p. on X is called
ergodic (with respect to "") if p.(E) = or 1 for every saturated Borel set E.
The terminology derives from the case where"" is the equivalence relation
induced in X by a group of transformations. Thus p. is ergodic iff the restric-
tion of p. to 9' is {O, l}-valued. One example of this is any measure p.
concentrated on a single equivalence class (p.(c(x)) = 1 for some x EX).
Such an ergodic measure is called trivial. The following lemma implies that,
in the familiar Borel spaces, the trivial equivalence relation x "" y iff x = y
has only trivial ergodic measures.

Lemma. Let X be a countably generated Borel space, and let p. be a {O, 1}-
valued Borel probability measure on X. Then p. is a point mass.

78
 3.4. Cross Sections

PROOF. By the lemma preceding 3.3.5 we may assume X is a subspace of

a Polish space P. Let JJ. be a {O, 1}-valued measure on X, and define a Borel
measure v on P by v(E) = JJ.(X n E). It suffices to show that there is a point
PEP for which v({p}) = 1, and that will follow if we show that the closed
support of v (i.e., the complement of the union of all open sets V ~ P for
which v( V) = 0) is a singleton (here we use the fact that nonzero Borel
measures in Polish spaces have nonvoid closed supports). Assume not, say
p #- q are in the closed support of v. Then there are disjoint open neighbor-
hoods U, V of p and q, and hence v(U) > 0 and v(V) > O. Since v(U) +
v(V) = v(U u V) ~ 1, this contradicts the fact that v (like JJ.) is {O, 1}-
valued. D

Proposition. Let X be a countably generated Borel space, let ~ be an equiva-

lence relation in X, and let q: X -+ X I ~ be the canonical projection. Assume
that for every finite Borel measure v on X I ~ there is a v-measurable cross
section hv for q. Then every ergodic Borel measure on X is trivial.

PROOF. Choose an ergodic Borel probability measure JJ. on X. Then we may

define a Borel measure v on X/~ by v(E) = JJ.(q-l(E)). Since JJ. is ergodic
v is {O, 1}-valued, and thus so is the completion v of v. By hypothesis there
is a v-measurable cross section hv for q, and thus we may define a third Borel
measure JJ.o on X via JJ.o(F) = V(h; l(F)). Now JJ.o is {O, 1}-valued so by the
lemma there is a point p E X for which JJ.o({p}) = 1. Let y = q(p). Note that
{y} = h; l( {p}) (because hv is a cross section for q) and {y} is a Borel set
in XI ~ because c(p) = q-l({ y}) is a Borel set in X. Thus v( {y}) = v( {y}) =
JJ.o({p}) = 1,andit follows thatJJ.(c(p)) = 1. Thus JJ. is trivial. D

In particular, we conclude that if an equivalence relation ~ in a Polish

space P admits a nontrivial ergodic measure, then the projection q: P -+ PI ~
does not have an absolutely measurable cross section. Here is one example
of this. Let P be the unit circle{Izl
= 1} with its usual Borel structure, let t
be any real number not a rational multiple of TC, and define Zl ~ Z2 in P if
there is an integer n such that Z2 = eintz 1 • Then each equivalence class is a
countable subset of P (therefore an Fq). So if we let JJ. be normalized Lebesgue
measure on P then every equivalence class has measure zero. However, JJ.
is well-known to be an ergodic measure for ~ (the easiest way to see that a
saturated Borel set E satisfies JJ.(E) = 0 or 1 is to look at the Fourier coeffi-
cients of the function XE, noting that XE(eitz) = XE(Z) for all Z E P). Therefore
the projection q of P onto the orbit space PI ~ has no absolutely measurable
section.
Of course the axiom of choice implies that cross sections exist for q, but
in the precise sense above they are all too pathological to be of any use in
analysis (note, incidentally, the close parallel between such functions and
Kolmogorov's example of a non Lebesgue measurable set). This suggests

79
3. Borel Structures

that quotient spaces as in the example are inaccessible in an essential way:

their properties may be undecidable within the usual processes of analysis.
The precise formulation and proof of that conjecture is, however, a problem
in metamathematics, not analysis.
In Section 4.1 we will discuss the spectrum of an arbitrary (separable)
C* -algebra. If the C* -algebra is not GeR, it is known that its spectrum is a
quotient space of the above pathological type. So the preceding discussion
strongly indicates that there is no hope of giving a concrete parametric
description of the irreducible representations of such a C*-algebra.

80
 4
From Commutative Algebras to
GCR Algebras

In this chapter we will work out the representation theory of separable

GCR algebras. The basic results, appearing at the end of Section 4.3, are
presented so as to obviously generalize the results of Section 2.2 which clas-
sify normal operators and representations of abelian C-algebras. But in
contrast with the commutative case, it is necessary here to make essential
use of the material on Borel structures from Chapter 3.
The spectrum of an abstract CO-algebra is defined in Section 4.1, and we
develop its properties as a Borel space. The main result asserts that the
spectrum of a separable GCR algebra is a standard Borel space.
Section 4.2 contains a discussion of some results relating to reduction
theory that are needed in the last section. This material has been kept to
the barest minimum. Nevertheless, as the expert will notice, it is still possible
to deduce, say, the central decomposition for separably-acting von Neumann
algebras from these results. We have resisted the temptation to do that here.

4.1. The Spectrum of a C*-algebra

We have seen two ways by which the spectrum of a commutative C"-
algebra contributes to the analysis of the algebra. On the one hand it provides
a concrete representation for elements in the algebra as complex-valued
continuous functions, and on the other it forms a conveniently topolo-
gized "parameter" space for the irreducible representations of the algebra
(with each point p we associate the one-dimensional representation "evalua-
tion at p"). We are now going to generalize the notion of spectrum to non-
commutative CO-algebras.
The first thing to notice is that there are a number of possible candidates.
Even in the commutative case one may view the spectrum as a space of

81
4. From Commutative Algebras to GCR Algebras

maximal ideals (endowed with the hull-kernel topology) or as a space of

complex homomorphisms (with the relative weak· -topology). In the non-
commutative case the proper counterparts of maximal ideals are not maximal
ideals but primitive ideals (a closed two sided ideal K in a C·-algebra A is
primitive if there is a nonzero irreducible representation n of A on a Hilbert
space such that K = kerneln), and the counterparts of complex homomor-
phisms are irreducible representations. In place of the maximal ideal space
one has the set prim A of all primitive ideals of A endowed with the hull-kernel
topology: by definition a subset S £ prim A is closed iff every primitive ideal
which contains the intersection nK, K E S, is already in S. The topological
space prim A is called the structure space of A, and was introduced in general
ring theory, in a somewhat different way, by N. Jacobson ([6], Section 3).
We shall not make use of the structure space, but we do want to point out
one or two of its properties.
Let us first consider an example. Let A be the family of all continuous
functions F defined on the unit interval [0, 1], taking values in the C· -algebra
M 2 of all complex 2 x 2 matrices, such that F(O) is diagonal. Such an F can
be regarded as a 2 x 2 matrix (fij) of continuous complex valued functions
on [0, 1] satisfying fdO) = f21 (0) = O. A becomes a C· -algebra with respect
to the pointwise operations and the norm IIF!! = SUPt IIF(t)!!. If we cause M 2
to act in the obvious way on a two dimensional Hilbert space Yf, then for
each t, 0 < t ~ 1, we obtain an irreducible representation n, of A on Yf by
evaluation at t:nt(F) = F(t). Associated with the point t = 0 we have two
one-dimensional irreducible representations, namely IX(F) = fll (0) and
{3(F) = fdO). It can be shown that these are the only irreducible represen-
tations of A (to within unitary equivalence) and therefore prim A is identified
with the union {IX, {3} U (0, 1] of the interval (0, 1] with two additional points
IX, {3 in place of O. The hull-kernel topology restricts to the usual topology
on (0, 1], whereas basic neighborhoods of IX (resp. {3) look like {IX} U (0, 6)
(resp. {{3} u (0, s)), for 0 < S < 1. Thus every neighborhood of IX must inter-
sect every neighborhood of {3, and so prim A is not a Hausdorff space. Note,
incidentally, that A is even a CCR algebra.
In the general case prim A is not even a T 1 space (the algebraic reason
being that one primitive ideal can be contained properly in another), though
it is always at least To ([6], Section 3). But there are even worse difficulties.
Suppose for example that A is a simple C-algebra, i.e., A has no closed ideals
other than 0 and A. In particular there are no primitive ideals other than 0,
and therefore prim A consists of a single point. Since there exist simple C-
algebras having many inequivalent irreducible representations (all having
kernel 0), it is apparent that prim A will not be adequate to serve as a param-
eter space for the irreducible representations of A. For this reason the
spectrum is defined in terms of irreducible representations rather than ideals.
Roughly speaking, the spectrum A of a C·-algebra A is the set of all
equivalence classes of irreducible representations of A, endowed with the
"natural" Borel structure. When A is commutative the set A will turn out to

82
 4.1. The Spectrum of a CO-Algebra

be set of all complex homomorphisms of A with the Borel structure generated

by the weak"-closed subsets. It is also possible in general to define a topology
on A. But as the preceding paragraphs suggest, this topology generally has
such poor separation properties that it loses much (though by no means all)
of its usefulness. Moreover, since the construction and classification of rep-
resentations involves measures rather than continuous functions, it is the
Borel structure rather than the topology that we shall have to deal with. We
remark that this view of the spectrum ofa C*-algebra is due to Mackey [20].
Let us now make all ofthis more precise. Let A be a separable CO-algebra,
fixed throughout the remainder of this section. Let Yt' be a separable Hilbert
space, and consider the set rep(A, Yt') of all representations of A on Yt'; the
e,
elements ofrep(A, Yt') are allowed to be degenerate. For each x E A, '1 E Yt',
we have a complex-valued function n -+ (n(x)e, '1) defined on rep (A, Yt'), and
we may topologize rep(A, Yt') by taking the weakest topology making all of
e,
these functions continuous. For x E A, '1 E Yt', we have Iln(x)e - '1112 =
(n(x"x)e, e) - 29le(n(x)e, '1) + 11'111 2, so that n -+ n(x)e defines a continuous
e,
function from rep(A, Yt') into Yt' for each x, and evidently we could have
used these functions to define the topology on rep(A, Yt'). Now choose a
dense sequence Xl' X2, . .• (resp. el, e2, •.• ) in the unit ball of A (resp. Yt')
and put
d(n, n') = L
m,n=l
2- m- nlln(xm)en - n'(xm)enll·

Then d defines a metric on rep(A, Yt') which makes rep (A, Yt') into a metric
space. It is not hard to see that rep(A, Yt') is complete and this topology has a
countable base (we omit these details; see [9]), so that rep(A, Yt') becomes a
Polish space. The set irr(A, Yt') of all irreducible members of rep(A, Yt') thus
becomes a topological space in its relative topology.
Now for each n = 00,1,2, ... (where as usual 00 stands for ~o) choose a
fixed Hilbert space Yt'n of dimension n. For each n we have a natural equi-
valence relation", in irr(A, Yt'n), namely n '" (J iff nand (J are unitarily
equivalent. Give irr(A, Yt' n) the Borel structure generated by its topology,
and then define the "n-dimensional" part An of the spectrum of A as the Borel
quotient space irr(A, Yt' n)/ "'. Finally, the spectrum A of A is defined as the
direct sum of the Borel spaces Aoo , A1> A2 , • •••
Note first that every nonzero irreducible representation of A is represented
in A. More precisely, if n is an irreducible representation of A on some Hilbert
e
space.Yt, then for each nonzero vector E .Yt n(A)e is dense in .Yt, and since
A has a countable dense subset it follows that .Yt is separable. So if n is the
dimension of Yt' then n is equivalent to some element of irr(A, Yt'n), as as-
serted.
We shall now collect a few general properties of A.

Theorem 4.1.1. Let A be a C*-algebra, let Yt' be a separable Hilbert space,

and let T be a bounded self-adjoint operator on Yt'. Then the set !/ T of all

83
4. From Commutative Algebras to GCR Algebras

representations n E rep(A, £) for which T belongs to the weak closure of

n(A) is a Go in rep(A, £).
PROOF. Choose any real number r > IITII and choose n E rep(A, £). We
claim: n E [ / Tiff T belongs to the weak closure of n(Ar), Ar denoting the ball
of radius r in A. Indeed the "if" part is trivial, so assume n E [/ T' By the double
commutant theorem T belongs to the weak closure of the --algebra n(A),
so by Kaplansky's density theorem (1.2.2) there is a net Xn E A such that
Iln(xn)11 ~ IITII for all n and limn n(xn) = T in the weak operator topology.
Now if K is the kernel of n then by 1.3.2, Iln(x)11 = infllx + kll, k E K, so we
may choose the Xn above in such a way that Ilxnll is as close to IITII as we please.
In particular we can choose Ilxnll ~ r, and the claim follows.
Thus [/ T may be written as {n E rep(A, £); T E n(Ar)-W}, and it suffices
to show that the latter is a Go.
Now choose a countable dense subset ~1' ~2" .. in £. For each n ~ 1,
x E A, define a real function In,x on rep(A, £) by
n

In,x(n) = L
i,j= 1
I(T~i - n(x)~i> ~)I·

Now Iln(Ar)il ~ r for each n, so that T E n(Ar)-W if and only if for every n,
k ~ 1 there is an x E Ar such that In, x(n) < 11k. Symbolically,

[/T = n n U {n
00 00

E rep(A, £):fn,x(n) < 11k},

and since each function fn, x is continuous, this formula exhibits [/ T as a

Go. 0

Note that the postulated separability of A did not enter the above proof,
and in fact 4.1.1. and its corollaries are valid for separable representations
of inseparable C-algebras.

Corollary 1. The set of nondegenerate elements of rep(A, £) is a Go.

PROOF. This is [/1 where I is the identity operator on £. o
Corollary 2. irr(A, £) is a Go in rep(A, £).
PROOF. Let Sand T be the real and imaginary parts of any particular
irreducible operator on £. Then a C-algebra on £ is irreducible iff its
weak closure contains both Sand T. Thus irr(A, £) = [/s n [/ T is an
intersection of two Go's in rep(A, £), and the conclusion follows. 0

Corollary 3. For every no E rep(A, £), {O' E rep (A, £):0' C1no} is a Go in
rep(A, £).
PROOF. Let I EE> 0 be the projection of £ EE> £ onto its first coordinate
space. It follows from 2.1.4 that 0' C1no iff I EE> 0 belongs to the weak closure

84
 4.1. The Spectrum of a CO-Algebra

of 0' Et> 7to(A)(in !l'(Jff Et> Jff)). Now by 4.1.1, {,1, E rep(A, Jff Et> Jff):I Et> 0 E
,1,(A)-} isa G,j in rep (A, Jff Et> Jff);andsinceO' -+ 0' Et> 7to definesacontinuous
map ofrep(A, Jff) into rep(A, Jff Et> Jff), we conclude that {O' E rep(A, Jff):O'
7to}, being the inverse image of a G,j under a continuous map, is a G,j. D

Corollary 4. For every 7to E irr(A, Jff), the equivalence class {7t E irr(A, Jff):
7t '" 7to} is an Fa subset of the topological space irr(A, Jff).
PROOF. Let C be the complement of {7t:7t '" 7to} in irr(A, Jff). It suffices to
show that C is a G,j in irr(A, Jff). Now two irreducible representations of A
on Jff are inequivalent if and only if they are disjoint. Thus we have C =
irr(A, Jff) n {7t E rep(A, Jff):7t d 7to}, and the conclusion follows from
Corollary 3. D

Let us now examine An for a separable CO-algebra A. Corollary 2 and

the remarks preceding 4.1.1 show that irr(A, Jffn) is a G,j in the Polish space
rep(A, Jffn). By 3.1.2 irr(A, Jffn) is therefore a Polish space in its relative
topology; and in particular the Borel structure on irr(A, Jffn) is standard.
The inverse image of each point in An is a single equivalence class of elements
in irr(A, Jffn), which by Corollary 4 is a Borel (in fact Fa) set in irr(A, Jffn).
These statements remain valid of course for the disjoint union A = Aoo U
Al U A2 U .. " and this proves the following description, due to Mackey
[20], of the dual of an arbitrary separable C* -algebra A.

Theorem 4.1.2. A is a quotient space of a standard Borel space, in which each

singleton is a Borel set.

We have already indicated in a vague way that A can be made into a

topological space, but one whose topology is bad in that points are not
necessarily closed. 4.1.2 asserts that points are at least "definable."
It is natural to ask at this point if the Borel space A is a good one or a
pathological one. Since A is a quotient of a standard Borel space, the results
of Chapter 3 show that the best one would hope for in general is that A be
analytic, and Corollary 2 of 3.3.5 shows that this is equivalent to having A
countably separated. It turns out that if A is a GCR algebra then A is in
fact standard, and if A is not GCR then A is not even countably separated.
The second assertion follows from a deep theorem of Glimm, and the
interested reader should consult [13] for the details. The first one, however,
we will deduce as a corollary to the following theorem on cross sections.

Theorem 4.1.3. Let A be a separable GCR algebra, fix n = 00,1,2, ... , a

denumerable cardinal, and let q:irr(A, Jffn) -+ An be the canonical quotient
map. Then q has a Borel cross section.
PROOF. Since irr(A, Jffn) is a Polish space and since the equivalence relation
in irr(A, Jffn) is defined by a group of homeomorphisms (the unitary group

85
4. From Commutative Algebras to GCR Algebras

q{ of Yf n acts on irr(A, Yf n} in the obvious way: for U E q{ and n E irr(A, Yf},

U . n is the representation x f--+ Un(x}U*}, the corollary of 3.4.1 would yield
the required cross section provided that for each n E irr(A, Yf n} the equiva-
lence class C n = {a E irr(A, Yfn}:O' '" n} is closed. Unfortunately this is not
true, even when A is a GCR algebra (though by Corollary 4 above Cn is in
general at worst a countable union of closed sets). What we will do instead
is produce a countable partition {Eb E 2 , • •• } ofirr(A, Yf n } into disjoint G6's
such that each Ej is saturated with respect to equivalence (i.e., n E Ej , 0' '" n
implies 0' E Ej ) and such that C n is closed in the relative topology of E j , for
every n E Ej • Since each Ej is a Polish space (3.1.2) we may apply the corollary
of 3.4.1 to obtain a Borel cross section hj for qlEj' Because of the disjointness
and saturation of {E j } it will therefore follow that the union of the functions
hi, h2' ... is the desired Borel cross section for q.
Let us now construct {E j }. Since A is GCR it has a composition series
{Jp:O ~ p ~ ex} of ideals as in 1.5.5, where the ordinal ex is countable by
separability of A (cf. remark following 1.5.5). For each ordinal p, 0 ~ p ~ ex,
let Fp = {n E irr(A, Yfn}:n(Jp} ¥ {O} }. It is clear that Fp is saturated with
respect to equivalence, and a routine check (which we leave for the reader)
shows that F p is open in irr(A, Yf n}. Therefore each set F p+ I \FP' 0 ~ P < ~,
being the intersection of a saturated open set with a saturated closed set,
is a saturated G6 • The family {Fp+I\Fp:O ~ p < ex} is countable (since ex is
a countable ordinal), and its elements are mutually disjoint. Thus it suffices
to show that Uo",p<", Fp+I\Fp = irr(A, Yf n}, and that for each p < ex and
n E Fp+ I \Fp, C n is closed in Fp+ I\Fp.
For the first conclusion, choose n E irr(A, Yf n }. Since n(J",} = n(A} ¥ 0
and n( J o} = n(O} = 0, there is a first ordinal A. > 0 such that n( J).} ¥ O.
Note that A. cannot be a limit ordinal: for if it were then J). would be the norm
closure of Ull <). J Il , and thus n(JIl } = 0 for J.l < A. would imply n(J).} = 0,
contradicting the choice of A.. Thus A. has an immediate predecessor A. - 1,
and so nEF).\F)'_1 £ Uo",p<", Fp+I\Fp. This proves that irr(A, Yf n} =
Uo",p<", Fp+I\Fp.
It remains to prove that each equivalence class is closed in Fp+I\Fp.
Therefore choose n E Fp+ I \Fp, and let 0' be an element of (Fp+ I \Fp) n C;.
We claim that 0' is equivalent to n (this will complete the proof). Since
O'(Jp+d ¥ 0 and n(Jp+d ¥ 0 (both 0' and n are in Fp+d and since Jp+ 1
is an ideal, it suffices to show that the restrictions of 0' and n to J p + I are
equivalent (1.3.4). Since both 0' and n annihilate J p (neither belongs to Fp)
we may regard these restrictions as representations a', n', respectively, of the
CCR quotient Jp+dJ p. Note that a' and n' are still irredicible, by (1.3.4),
and we are now reduced to proving that a' and n' are equivalent. Now
since 0' belongs to the closure of C n , it is apparent that a' belongs to the
closure of Cn ' = {A. E irr(Jp+ l \1p, Yfn}:A. '" n'} (just choose a sequence Un
of unitary operators such that UnnU: converges to 0', then restrict everything
to Jp+ 1 and pass to the quotient Jp+dJp). Finally, we claim that ker n' £
ker a'. Indeed if x E J p+ 1/1p is such that n'(x} = 0, then A.'(x} = 0 for every

86
 4.1. The Spectrum of a C*-Algebra

A' E en" and since a' is a limit of such A's, we see that A'(X) = O. Since
J p + 1 jJ p is a CCR algebra, we can conclude from 1.5.2 that n' is equivalent
to ai, and the proof is now complete. 0

Corollary. If A is a separable GCR algebra, then A is a standard Borel space.

PROOF. For each n = 00,1,2, ... , the canonical quotient map ofirr(A, £'n)
onto An has a Borel cross section; by 3.4.2 An is standard, and therefore
A, being a countable direct sum of standard Borel spaces, is standard. 0
Theorem 4.1.3 was first obtained by Dixmier [9], in a slightly different
form. Its corollary is due to Dixmier [10], Fell [12], and Glimm [13]'
Theorem 4.1.3 above asserts that Borel cross sections exist (at least for
separable GCR algebras), and the question arises as to what extent they are
unique. We conclude this section with a discussion of uniqueness. The result
is a consequence of the following lemma, which is of some interest in its
own right. First, we need some preliminaries. Let IlIJ be the group of unitary
operators on a separable Hilbert space £'. It is a familiar fact that the strong
and weak operator topologies coincide on 1lIJ, therefore IlIJ is unambiguously
a topological space in either of these topologies, with respect to which it is
a subspace of the Polish space {T E .P(£'):IITII ~ 1} (the weak operator
topology on the latter). We want to point out that IlIJ is a Polish space.
For that choose a countable dense set of unit vectors Wi' W2, ••. in £' and
define a metric d on IlIJ by

L rn(IIUwn -
00

d(U, V) = Vwnll + IIU*wn - V*wnll)·

n=i

A simple argument which we omit shows that d is a complete metric on

1lIJ, and thus dIJ is Polish.
Now dIJ is also a group, and its center is the compact subgroup lf of all
scalar operators AI, IAI = 1. Consider the quotient group dIJIll', with the Borel
structure generated by its quotient topology. It turns out that IlIJIll' is Polish
as a topological space, but all we shall require is that dlJllf is a standard
Borel space. One way to see this is to note that the quotient map p:1lIJ -+ IlIJ/lf
satisfies the hypotheses of Theorem 3.4.1 and therefore has a Borel cross
section. By Proposition 3.4.2 IlIJ/lf is Borel isomorphic to a Borel subset of
the Polish space dIJ, which shows dlJllf is standard.

Lemma 4.1.4. Let A be a separable C*-algebra, let £' be a separable Hilbert

space, and let r be the graph of the unitary equivalence relation in
irr(A, £'), regarded as a subset of the Polish space irr(A, £') x irr(A, £').
Then r is a Borel set, and there is a Borel map y1--+ Yy of r into the
unitary group dIJ of £' such that l-(n,a)nV('n,a) = a for all (n, a) E r.
PROOF. Define a map ¢ of the Polish space irr(A, £') x IlIJ into the Polish
spaceirr(A,£') x irr(A,£')by¢(n, V) = (n, VnV*).Clearly¢iscontinuous,

87
4. From Commutative Algebras to GCR Algebras

maps onto r, and the inverse image of each point in irr(A,.Yf) x irr(A,.Yf)
is closed. We want to show now that 4> maps open sets to Borel sets. Note
first that 4>(n, V) = 4>(n', V') iff n = n' and V = V' (mod T). Thus 4> lifts
to a continuous map 4> of the topological space irr(A, .Yf) x O/t/T onto r
in the usual way. Now 4> is clearly 1-1 and it is a Borel map (by continuity)
if we give irr(A, .Yf) x O/t/T the product Borel structure. By the remarks
preceding the lemma, the domain of 4> is a standard Borel space. Thus by
3.3.2, r is a Borel set and 4> is a Borel isomorphism. In particular 4> maps
open sets to Borel sets. Since the quotient map p:O/t --+ O/t/T is open, the
map id x p:(n, V) E irr(A,.Yf) x O/t f--+ (n, p(V)) E irr(A,.Yf) x O/t/T is open,
and it is therefore apparent that 4> = 4> 0 (id x p) maps open sets to Borel
sets.
By Theorem 3.4.1, there is a Borel map h:r --+ irr(A, .Yf) x O/t such that
4> 0 h is the identity on r. Writing h(y) in coordinates (ny, ~), it follows
from an inspection of the formula 4> 0 h = id that the Borel map y - ~
has the stated properties. 0

Corollary Uniqueness of cross sections. Let A be a separable C·-algebra, let

X be a Borel space, and let .Yf be a separable Hilbert space. Let x f--+ nx,
x f--+ ax be any two Borel maps of X into irr(A, .Yf) such that nx '" ax for
all x E X. Then there is a Borel map x f--+ Ux of X into the unitary group
O/t of .Yf such that ax = UxnxU; for all x E X.

PROOF. Let r ~ irr(A,.Yf) x irr(A,.Yf) be the graph of the relation "'. By

the lemma there is a Borel map (n, a) E r --+ JI(", (1) E O/t having the asserted
properties of the lemma. Define the function U:X --+ O/t by Ux = JI("x, <1 x )'
U is a Borel map since it is the composition of the Borel map V with the
Borel map x f--+ (nx, ax), and the relation ax = UxnxU; is immediate from
the properties of V. 0

Note that if A is a separable GCR algebra and if we take X = An in

the above corollary, we obtain a uniqueness statement for the cross sections
whose existence was established in 4.1.3.

4.2. Decomposable Operator Algebras

Let (X, /1) be a finite measure space and let g be a separable Hilbert space,
all of which will be fixed throughout this section. In order that certain Hilbert
spaces associated with X, /1, g be separable, we also require that X be
countably generated as a Borel space. A vector-valued function ~:X --+ g
is called a Borel function if (~(x), u) defines a Borel function on X, for every
vector u E g. The set of all such Borel functions ~ satisfying the condition
11~112 = SIl~(x)1I2 d/1(x) < 00 forms a separable Hilbert space, after we identify
functions that agree a.e. (/1), which we denote by L 2(X, /1; g). Note, for
example, that x f--+ 1I~(x)1I is a Borel function since it can be expressed as a

88
 4.2. Decomposable Operator Algebras

countable supremum of Borel functions supn I(~(x), vn)l, Vi> V2,' .. being any
countable dense set of unit vectors in .Yt". Let &\J(X) denote the set of all
complex-valued Borel functionsf:X -+ C which are bounded (suPx If(x)1 <
(0). It is easy to see that each function f E &\J(X) gives rise to a bounded
multiplication operator L f on L2(X, j.I.;.Yt") (namely (Lf~)(x) = f(x)~(x),
x E X, ~ E L2(X, j.I.; .Yt"»; such operators are called diagonal, and the set 1Z
of all diagonal operators on L2(X, j.I.;.Yt") forms an abelian *-algebra which
contains the identity.
In this section we are going to study the family of all separable C* -algebras
d acting on L2(X, j.I.;.Yt") whose strong closures lie between 1Z and 1Z'. Such
a C*-algebra is called decomposable. Note, incidentally, that by 1.2.3, every
von Neumann algebra fF/ such that 1Z s;;; fF/ s;;; 1Z' is the strong closure of
such a C*-algebra. Considerable information will be obtained about the
structure of these algebras, for use later in 4.3. The main results are sum-
marized after Corollary 4 at the end of this section.
An operator valued function F:X -+ 2(.Yt") is called a Borel function if
F becomes measurable when 2(.Yt") is endowed with the Borel structure
generated by the weak operator topology; equivalently, x 1-+ (F(x)u, v) is a
scalar valued Borel function for every pair of vectors u, v E .Yt". Let &\J(X, .Yt")
denote the set of all such functions F which are bounded: supx IIF(x)1I < 00.
For each FE &\J(X, .Yt") the real-valued function IIF(x)11 can be written as
sUPm,n I(F(x)vm , vn)l, where VI' V2, ... is a countable dense set in the unit ball
of .Yt", and therefore IIF(x)11 is a Borel function. Similar arguments, along
with some juggling with inner products, show that &\J(X, .Yt") is closed under
(pointwise) sums, products, and adjoints, and that for each FE &\J(X, .Yt")
and ~ E L2(X, j.I.; .Yt"), F(x)~(x) defines another element of L2(X, j.I.; .Yt").
In particular, the pointwise operations and the norm IIFII = supx IIF(x)11
make &\J(X, .Yt") into a normed *-algebra, and in fact &\J(X, .Yt") is a C*-algebra
(i.e., is complete and satisfies IIF* FII = IIFI12) with identity. Now let us define,
for each FE &\J(X, .Yt"), a multiplication operator LF on L2(X, j.I.;.Yt") as
follows: (LF~)(X) = F(x)~(x), for ~ E L2(X, j.I.; .Yt"). The above remarks show
that LF~ belongs to L2, and the following properties are evident: IILFII ~ ess
sup IIF(x)11 ~ IIFII, L FG = LFLG, L aF + bG = aLF + bLG for scalars a, b, and
L F• = (LF)*' Thus F 1-+ LF is a representation of &\J(X,.Yt") on L2(X, j.I.; .Yt").
Moreover, it is clear that each operator LF commutes with the algebra 1Z
of diagonal operators. The following result shows that in fact these multiplica-
tion operators exhaust 1Z'.

Theorem 4.2.1. For every operator T E 1Z' there is an F E &\J(X, .Yt") such that
T = LF and IIF(x)11 ~ IITII for all x E x.

PROOF. Fix T E 1Z'. Now each vector u E.Yt" can be regarded as a constant
function in L2(X, j.I.; .Yt"), and so gives rise to a (square integrable) .Yt"-valued
function x 1-+ (Tu)(x) (of course, Tu is only an equivalence class of functions;
what we mean is, choose an element in that class). We claim: for each pair

89
4. From Commutative Algebras to GCR Algebras

u, V E %, I( (Tu)(x), v)1 ~ II Til '1Iull' Ilvll a.e. (Jl). Since ((Tu)(x), v) is a scalar-
valued function in L 2 (Jl) this will follow if we prove that

IJ( (Tu)(x), v)f(x)g(x) dJl(x)1 ~ II Til' Iluli' Ilvll

for all bounded functions f, g satisfying Ilflb IIgl12 ~ 1. For such functions
f, g put ~(x) = f(x)u and 11(X) = g(x)v. Then ~ = Lfu and since L f belongs
to ;!r and T commutes with ;!r, we have
J( (Tu)(x), v)f(x)g(x) dJl(x) = J(f(x)(Tu)(x), g(x)v) dJl(x)
= (LfTu, Lgv) = (TLfu, 11)
= (T~, 11).

Since I(T~, 11)1 ~ IITII'II~II'II11II, and since II~II ~ Ilull and 111111 ~ Ilvll (because
Ilf112' IIgl12 ~ 1), the desired inequality follows.
Now let Ub U2, ••• be a countable dense set in % and let % 0 be the count-
able set of all finite sums Li riui, where the r;'s range over the field I!) + il!) of
all complex rational numbers. %0 is a vector space over I!) + il!) which is
dense in %. Since %0 is countable, we can find a Borel set N £:: X such that
Jl(N) = 0 and I( (Tu)(x), v)1 ~ II Til' Iluli' Ilvll for all u, v E %0 and all x E
X\N. Moreover, since T is linear, we must have T(au + bv)(x) = a(Tu)(x) +
b(Tv)(x) a.e. (Jl), for each u, v E %0 and each a, bEl!) + il!); because there
are only a countable number of such relations, we may assume by throwing
out a larger null set M :2 N that T(au + bv)(x) = a(Tu)(x) + b(Tv)(x),
identically on X\M and simultaneously for all u, v E %0, a, bEl!) + il!). So
for each x E X\M we have a function [u, v]x = ((Tu)(x), v) defined on %0 x
% 0 having all the properties of a bounded bilinear form which make sense
on %0 x %0' Thus [u, v]x may be uniquely extended by continuity to a
bona fide bounded bilinear form on % x %, so that a familiar lemma of
Riesz provides us with an operator F(x) E 2(%) such that (F(x)u, v) =
((Tu)(x), v), u, v E %0' Now extend the function F to all of X by putting
F(x) = 0 on M. Clearly IIF(x)11 ~ II Til for all x E X, and the fact that LF = T
is a routine verification which will be left for the reader. 0

Corollary 1. The map F --+ LF is a --homomorphism of !J4(X, %) onto ;!r'.

Note that LF = 0 if and only if F(x) = 0 a.e. (Jl); thus these functions
form a closed ideal % in !J4(X, %) and the map F --+ LF induces an isometric
--isomorphism of !J4(X, %)/% onto ;!r'. Moreover, the quotient is identi-
fiable with an obvious L 00 space of operator-valued functions, thereby giving
a concrete representation of elements of ;!r'.
In particular, every C*-algebra contained in ;!r' has a replica in !J4(X,
%)/%, but it will turn out later that this representation is not convenient
for our purposes. What we actually need is a generalization of 4.2.1 which
provides a cross section for the map F 1-+ L F , which preserves all the algebraic
relations. Corollary 2 provides the required "lifting" theorem.

90
 4.2. Decomposable Operator Algebras

Corollary 2. Let d be a separable C* -subalgebra of fL'. Then there is an

isometric *-homomorphism n:d --+ PJ(X, %) such that L,,(T) = T for all
TEd.
PROOF. Let!/' be a countable subset of d which generates d as a C*-
algebra. We may assume !/'* = !/', and that!/' contains the identity if d
does. Let do be the countable family of all finite sume of operators of the
form rA1A2 ... An, where Aj E!/', r E Q + iQ. do is clearly dense in d
and is a *-algebra over the field Q + iQ of rational complex numbers. For
each element T E do, choose a function FT E PJ(X, %) as in 4.2.1. Now for
a, bE Q + iQ and S, TEd 0, we must have Fas + bT(X) = aFs(x) + bF T(X)
a.e. (J1) and (F T')(X) = (F T(X»* a.e. (J1) [because G f---+ LG is a*-homomor-
phism and LG = 0 iff G(x) = 0 a.e. (J1)]' Since there are only a countable
number of such relations there is a Borel set N such that J1(N) = 0, and all
of the above relations are satisfied on X\N. Define n: do f---+ PJ(X, %) by
n(T)(x) = F T(X) if x E X\N, and n(T)(x) = 0 if x E N. Then n is a *-
homomorphism of do into PJ(X, %) satisfying L,,(T) = T and Iln(T)(x)11 ~
II Til, T E do, X E X. Note that II Til ~ ess suplln(T)(x)11 ~ suplln(T)(x)11
follows from the first equation, so that n is in fact isometric. Thus the unique
extension of 7t to d has all the right properties. D

Let d be a separable C*-subalgebra of fL', and choose a homomorphism

n:d --+ PJ(X, %) as in Corollary 2. Then for each x E X, we obtain a rep-
resentation 7tx of d on % by evaluating n at x:nx(T) = n(T)(x), TEd.
We shall require some information about this family of representations of
d.

Proposition 4.2.2. Let F, F b F 2,

. . . be a sequence in PJ(X, %), such that the
sequence LFn of operators converges strongly to L F. Then there is a sub-
sequence n' £: n such that Fn,(x) converges strongly to F(x) for J1-almost
every x E X.
PROOF. The idea is simple, and we merely sketch the details. Note first that
the norms IILFJ are bounded (Banach-Steinhaus theorem). Letting Vb
V2, ... be a dense sequence in %, we see that
lim JIIFn(x)vj - F(x)vjW dJ1(x) = 0 for each j = 1, 2, ....
n

Thus for eachj there is a subsequence nlj < n2j < ... so that
Ii~ IIFnkix)Vj - F(x)Vjll = 0
for all x in the complement ofa Wnull set Nj £: X. By a simple induction we
can even arrange that the (k + l)st sequence is a subsequence of the kth.
The required subsequence is obtained by diagonalization (its kth term is
the kth term of the kth subsequence), and the exceptional null set can be
taken as N 1 U N 2 U . . . . By discarding another null set we may assume
IIFn(X) II ~ IILFnll for all x and all n, and now the boundedness condition and

91
4. From Commutative Algebras to GCR Algebras

the density of {Vj} in Yl insures strong convergence almost everywhere of

the subsequence. D

Now let d ~ fZ' be a separable C*-algebra with trivial nullspace, and

let rr: d --+ PJ(X, Yl) be as in Corollary 2 above.

Corollary 1. If FE PJ(X, Yl) is such that LF Ed", then F(x) E rrAd)" for
almost every x E X.
PROOF. It suffices to prove the assertion separately for the real and imaginary
parts of F, so we can assume F (and therefore L F ) is self-adjoint.
By the double commutant theorem and the corollary of Kaplansky's
density theorem (1.2.2), we can find a sequence Tn = T~ E d such that
Tn --+ LF strongly. By Lemma 1 (and by passing to a subsequence if necessary)
we may assume rr(Tn)(x) --+ F(x) strongly, for almost every x E X. Thus
F(x) belongs to the strong closure of rrAd) a.e. (f.1), as required. D

Recall that two representations rr, (1 of a C*-algebra are called disjoint

if no nonzero subrepresentation of rr is equivalent to any subrepresentation
of (1.

Corollary 2. If d is strongly dense in fZ', then rrx is irreducible for almost

every x EX.
PROOF. Choose any irreducible operator T E 2'(Yl), and let F be the
constant function F(x) = T, x E X. Then LF E fZ', so by hypothesis LF Ed".
By Corollary 1 we conclude that T E rrAd)" a.e. (f.1), from which the con-
clusion is evident. D

Corollary 3. Assume d has trivial nullspace. If fZ c;; d" c;; fZ', then there
is a Borel set N C;; X of measure zero such that rrx and rry are disjoint for
all pairs x "# y in X\!V.
PROOF. Since X is countably separated (in fact, countably generated), we
can find a sequence fi> f2, ... of bounded real valued Borel functions on X
which separates points. Each multiplication operator LIn belongs to fZ, and
since fZ is contained in the strong closure of d by the double commutant
theorem, we may find, for each n, a sequence Ani> A n2 , ... in d such that
limkAnk = LIn strongly, for every n = 1,2, ... (imitate the proof of Corollary
1). Since Ank = L,,(Ank) we may assume (utilizing 4.2.2 and passing to sub-
sequences if necessary) that for each n, limk rrAAnd = f,.(x)I strongly for all
x in the complement of some Wnull set N n • Put N = N J U N2 u···.
Now choose x "# y in X\N. To see that rrx and rry are disjoint, let T be a
bounded operator on Yl such that Trrx(A) = rry(A)T, A E d. This remains
true for A = Ank as above, and thus by taking strong limits on k (which
exist on X\N) we conclude Tfn(x) = fn(y)T = Tfn(Y), for all n = 1,2, ....
Since x "# y there is an n for which fn(x) "# f,.(y), and this implies T = O.
From 2.1.4 (ii) we conclude that rrx d rry. D

92
 4.2. Decomposable Operator Algebras

Corollary 3 has a plausible converse. Namely, let .511 be a separable

C*-subalgebra of ~' and suppose that we have a map n:d -+ fJI(X, %) as
in Corollary 1, which satisfies nx d ny for all x # y in the complement of a
Wnull set. Then it is quite reasonable to conjecture that .511" contains the set
~ of diagonal operators. This can be proved, in fact, if J1. is an atomic measure.
But in general, and surprisingly perhaps, it is false. It turns out that even
when X is a standard Borel space, there is a finite Borel measure J1. on X, a
separable C*-algebra .511 ~ ~', and a map n:d -+ fJI(X, %) as in Corollary
1 all of which satisfy

(i) nx d ny for all x # y in X, and

(ii) .511" is afactor (not of type I)

(cf. [8], Corollary 4 et seq.). Because of(ii) there is of course no hope of having
~ ~ .511"; for ~ ~ .511" ~ ~' implies that ~ is contained in the center of
.511", which by (ii) is only one-dimensional. This behavior illustrates only one
kind of pathology that occurs in the representation theory of C*-algebras
which are not GCR. We will describe others in the next section.
The next result gives a useful and concrete description of operators in
.511", when ~ ~ .511" ~ ~'. We assume, of course, that .511 has trivial null-
space. In terms of direct integrals (which we have deliberately avoided until
now) it asserts that .511" is the direct integral JEll nAd)" dJ1.(x). More generally,
it leads directly to the existence of the central decomposition for von
Neumann algebras acting on a separable Hilbert space. However, we will
not go into that here.

Corollary 4. Assume ~ ~ .511" ~ ~'. Then .511" coincides with the set of all
multiplications L F , F E fJI(X, %), satisfying F(x) E nx(d)" a.e. (J1.).

PROOF. The inclusion ~ is Corollary 1 above. For the other inclusion choose
FE fJI(X, %) satisfying F(x) E nAd)" a.e. We will show that LF commutes
with .511'. Choose TEd'. Since ~ ~ .511" we have .511' = .511'" ~ ~' and so
by 4.2.1 T has the form LG for some G E fJI(X, %). Let At, A 2 , ••• be a
generating sequence for d. Then since LG commutes with each An = L,,(An)
we have G(x)nx(An) = nx(An)G(x) a.e. (J1.) for every n = 1, 2, .... Thus there
is a Borel set N of measure zero such that G(x)nx(An) = nx(An)G(x) for every
x E X\N and every n = 1, 2, .... Hence G(x) E nAd)' for all x E X\N, and
since F(x) E nAd)" a.e. (J1.) we conclude that F(x) commutes with G(x) a.e.
(J1.). It follows that LF commutes with LG = T. 0

Let us now summarize the main points. For every separable C*-algebra
.511 ~ ~', there is an isometric *-homomorphism n:d -+ fJI(X, %) satis-
fying A = L,,(A), A E d. The representations nAA) = n(A)(x) have the
properties:

(i) If .511" = ~' then nx is irreducible for almost every x(J1.).

93
4. From Commutative Algebras to GCR Algebras

(ii) If fL £;; .JiI" £;; fL' then the {n,,} are mutually disjoint for all x in the
complement of a Jl-null set.
(iii) If fL £;; .JiI" £;; fL' then fL" consists precisely of those multiplications
LF satisfying F(x) E n,,(.JiI)" a.e. (Jl).

4.3. Representations of GCR Algebras

In Section 2.2 we classified the separable representations of separable
abelian C*-algebras in terms of measure classes and multiplicity functions.
We are now going to imitate that program for general separable GCR
algebras. While the details are somewhat more complicated in the GCR case,
the basic steps are the same. First, we will choose a finite Borel measure Jl on
the spectrum of the algebra and construct a multiplicity-free representation
n/J' which will depend only on the measure class of Jl. Second, we will show
that every separable multiplicity-free representation is equivalent to one of
these n/J's. Third, we will utilize the results of Chapter 2 to classify general
type I representations in terms of the multiplicity-free representations.
Let A be a separable GCR algebra, which will be fixed throughout this
section. Choose a finite Borel measure Jl on the spectrum A of A. Since A is
the direct sum Aoo U A1 U A2 ... , it is convenient to deal with each piece
An separately; thus we will assume for the moment that Jl is concentrated on

n, n,
some An, intending to reassemble the pieces later. By 4.1.3 there is a Borel
mapping 'E An 1--+ E irr(A, Jf n) such that belongs to the class " for
every' E An. Now form the "direct integral" n/J = JED n, dJlW. More precisely,
consider the Hilbert space L2(An, Jl; Jfn) of all Jfn-valued Borel functions
e:An -+ Jf n such that IIel1 2= JI~(OI12 dJl(O < 00. For every x E A, we may
define an operator n/J(x) on L 2(An' Jl; Jf n) by
(n/J(x)e)(,) = n,(x)e(C).

n,
, E An, eEL 2(Ano Jl; Jf n).1t is not hard to see that everything in sight is Borel
measurable, because' 1--+ is a Borel map of An into irr(A, Jf n). Thus n/J is a
representation of A on L 2(An' Jl; Jf n). It is clear that n/J depends on Jl, and
conceivable that it also depends on the particular cross section, 1--+ n,. We
first want to show that the unitary equivalence class of n/J is independent of
the particular cross section. Indeed, if' 1--+ (1, is another Borel map of An into
irr(A, Jfn) such that (1, E' for all, E An, then we may form (1/J = Je (1, dJl(O
in the same way we formed n/J" We claim:(1/J and n/J are equivalent. Indeed
by the corollary of 4.1.4 there is a Borel map' 1--+ Vc of An into the unitary
group of Jf n such that (1,(x)Vc = Vcn,(x), x E A, 'E An. Define an operator
Von L2(An, Jl; Jfn) by (Ve)(O = Vce(O, eE L2" E An. Since Vc is unitary for
all , it follows that V is unitary, and the relation (1, Vc = Vcn, becomes (1/J V =
Vn/J' proving the claim.
Therefore n/J is well-defined by Jl up to equivalence. Moreover, it will be
convenient to write, in place of n" where it is understood always that the
irreducible representations x E A 1--+ '(x), , E A", in fact come from a particu-

94
 4.3. Representations of GCR Algebras

lar Borel cross section for the quotient map. Thus the direct integral notation
for n" becomes n" = J(j) ( dj1.«().
The following theorem is the key step leading to the fact that n" is mul-
tiplicity-free. While the corresponding result for abelian C-algebras (2.2.1)
is straightforward, the argument here has to be made with some care.
Theorem 4.3.1. n"(A),, contains the algebra !L of all diagonal operators.

PROOF. It suffices to show that n"(A),, contains all diagonal operators of the
form L J where f is the characteristic function of a Borel set in An. Let 91 0
be the class of all Borel sets in An having this property. Clearly 91 0 is closed
under complementation, and a simple application of the bounded conver-
gence theorem shows that 91 0 is closed under countable unions. Thus 91 0 is
a sub a-field of the Borel field 91 on An. Since An is a standard Borel space
(corollary to 4.1.3.), the unique structure theorem (3.3.5) shows that 91 0 = 91
provided we can find a countable subfamily of 91 0 which separates points.
This countable subfamily is constructed as follows. Let r 1 , r2, ... be an
enumeration of the nonnegative rational real numbers, let at> a2, ... be a
countable dense set in A, and define subsets Skt ~ irr(A, Yf n) by
k, [ ~ 1.
Letting q be the natural map of irr(A, Yfn) onto An, we obtain a countable
family of sets q(Skt) in An. Now each Skt is clearly closed in irr(A, Yf n), and is
saturated with respect to the equivalence relation -. Thus each image q(Skt)
is a Borel set in An. We claim that {q(Skt)} separates points in An; equivalently,
if n, a E irr(A, Yf n) are such that n E Ski iff a E Ski for all k, [ = 1,2, ... , then
n - a. To see this, note that the hypothesis means Iln(adll ~ riff Ila(adll ~ r
°
for every rational r ~ and every k = 1,2, ... , and hence Iln(ak)11 = Ila(ak)11
for all k. Since {at> a2,' .. } is dense in A we conclude Iln(x)11 = Ila(x)1 for all
x E A and in particular n and a have the same kernel. By 1.5.4, n and a are
equivalent, as asserted.
It remains to prove that each set q(Skt) belongs to f!4 0 . For that, fix k and [,
and let J be the ideal n {ker a: a E Skt}. First, we claim Skt = {n E irr(A, Yf):
n(J) = O}. Let r be the (possibly inseparable) representation of A defined by
r = $0', a E Skt. Since J is the kernel of rand n(J) = 0, it follows that the map
r(x) t-+ n(x), x E A, is a well-defined representation of the C·-algebra r(A),
and so by 1.3.2 we have Iln(x)11 ~ Ilr(x)11 = SUPUESkt Ila(x)11 for all x EA. In
particular this inequality shows that Iln(ak)11 ~ r t , hence n E Skt' proving
the claim.

°
Now let et> e2, ... be an approximate identity for the ideal J (Section 1.3).
It is clear that if a E Skt then a(ej) = for all j (because a(J) = 0). Con-
versely, suppose a E irr(A, Yfn)\Skt. Then we claim: a(ej) -+ I in the strong

°
operator topology of 2(Yf n), as j -+ 00. Indeed, since Ila(ej)11 ~ 1 it suffices
to show that Ila(ej)17 - 1711 -+ for all 17 in a dense set of Yf n. Now by the
preceding paragraph a(J) '# 0, and since J is an ideal a(J) must therefore be
an irreducible C-algebra in 2(Yfn) (1.3.4). Thus a(J)~ is dense for every

95
4. From Commutative Algebras to GCR Algebras

"#
°since
~ O.Butifl]hastheforma(z)~withz E Jthena(ej)I] - I] = a(ejz - z)I] --+
Ilejz - zil 0, proving the claim.
--+
Finally, let f be the characteristic function of An\q(Sk(). We want to show
that L f belongs to 7tJl(A)". Now for each j, we obtain a bounded 9'(£'n)-
valued Borel function Ej on An defined by Ej(O = (ej), ( E An. It is clear
that the multiplication operator LEi belongs to 7tiA) (LEi is just 7tJl(ej)).
Moreover, the preceding paragraph shows that Ej «() --+ f(OI strongly, for
each (E An. An application of the bounded convergence theorem shows that
LEi --+ L f strongly in L 2(An' Jl; £' n), and in particular we conclude that
L f E7tJl(A)". 0

Corollary 1. 7tJl(A)" = ~'.

PROOF. It is obvious that 7tJl(A) commutes with ~, so by the above theorem

we have ~ ~ 7tJl(A)" ~ ~'. Moreover, since the mapping which associates
with each operator 7tJl(x) in 7tJl(A) the Borel function F(O = (x) satisfies the
hypothesis of Corollary 4 of 4.2.2., we see that 7tJl(A)" coincides with the family
of all multiplications L F , where FE 8I(An' Jl; £'n) satisfies F(O E (A)" a.e.
(Jl). Since (A)" = 9'(£'n) for every (E An (by irreducibility) we conclude
that 7tJl(A)" contains all multiplication operators. The conclusion now follows
from 4.2.1.

Corollary 2. 7tJl is multiplicity-free.

PROOF. By Corollary 1, 7tJl(A), is the abelian von Neumann algebra ~". 0

We remark that Theorem 4.3.1 and its corollaries are false if A is not GCR.
Indeed, for every separable non-GCR algebra A one can find a standard
Borel space X (which can be taken as a Borel subset of Aoo), a measure Jl on
X, and a Borel map x f-+ 7t x of X into irr(A, £'n), such that 7t x and 7t y are dis-
joint for x "# y, and which has the property that the range of the represen-
tation JEll 7t x dJl(x) generates a factor not of type 1 (cf. [8] and the discussion
following Corollary 3 of 4.2.2.). In particular JEll 7tx dJl(x) is not multiplicity-
free. We have already remarked in 4.1 that the spectrum of a non-GCR
algebra is "bad" as a Borel space. The above phenomenon illustrates quite
a different source of pathology associated with those C*-algebras.
Note also that Corollary 1 implies that the center of the von Neumann
algebra 7t(A)" coincides with the center of ~', which is in turn the von
Neumann algebra ~ of all diagonal operators. Since Jl is concentrated on
A.,~' is identified with the set of all multiplication operators L F , F E
8I(An' Jl; £'n); and since dim £'n = n, we see that the identity representation
of ~ has multiplicity n. Conclusion: if Jl is concentrated on An, then the center
of 7tJl(A)" has multiplicity n. This tells us how to find n, given 7tJl and the
knowledge that Jl is concentrated on some particular piece An of A.
Now, still assuming that Jl is concentrated on An, we want to investigate
how the multiplicity-free representation 7tJl depends on Jl. The answer is the
same as in the abelian case.
96
 4.3. Representations of GCR Algebras

Theorem 4.3.2. Let p. and v be two finite Borel measures concentrated on An.
Then 11:" '" 11:. iff P. and v are equivalent measures, and 11:" d 11:. iff P. and v
are mutually singular.
PROOF. We have arranged the proof of 2.2.2, so that it translates verbatim
to a proof of this theorem, provided one uses 4.3.1. in place of 2.2.1. For
example, to see that p. '" v implies 11:" '" 11:., let h = dp.ldv as in 2.2.2., and
define U:L2(Am p.; Jif"n) ~ L2(An, v; Jif"n) by the same formula as in 2.2.2
~ 2 ~

( E An, eEL (An' p.; Jif"n)

(this makes perfect sense even though e is vector-valued). Then we see that

showing that U is an isometry, and the arguments for UU* = I and U1I:" =
1I:.U are completed in the same way. 0

Now let us put the pieces together. Namely, choose a finite Borel measure
p. on A, where we do not assume that p. is concentrated on some An. If P.n
is the restriction of p. to An, then we can form 1I:"n as above, obtaining a separa-
ble multiplicity-free representation. Define 11:" = 11:"00 E9 11:", E9 11:"2 E9 ....
One can also think of 11:" as acting on a space of vector-valued functions.
Briefly, let L2(A, p.; Jif" 00 U Jif"l U Jif"2 u···) denote the space of all Borel
functions e from A into the direct sum U" Jif"" of the Borel spaces Jif"m such
thate(OE Jif""forall( E A"andforwhich the norm Ilell = (fA Ile(x)11 2 dp.(xW/2
is finite. Then for each x E A, 1I:,,(x) acts as follows: (1I:,,(x)e)(() = ((x)e(O, for
every (E A" and every n = 00,1,2, ... (the reader may wish to examine this
realization of 11:" in more detail). In order to suggest this interpretation, we
will write 11:" =J' ( dp.(o. In any case, 11:" is clearly a separable representation
(because each 1I:"n is), and we want to show now that 11:" is multiplicity-free.

Proposition 4.3.3. Let p. and 11:" = 11:"00 E9 11:", E9 11:"2 E9 ... be as above, and
let Pal, P 1> P 2 , ••• be the projections of the Hilbert space Jif" of 11:" onto the
respective coordinate subspaces. Let ce be the center of 1I:".(A)". Then each
P" E ce, 11:" is multiplicity1ree, and the abelian von Neumann algebra
celpnJt" is 0 or has multiplicity n,for every n = 00, 1, 2, ....
PROOF. We claim first than 1I:"m and 1I:"n are disjoint if m 1= n. To prove this,
it suffices to show that for every nonzero subrepresentation a of 11:,,", a(A)'
is an abelian von Neumann algebra having multiplicity n (this clearly implies
that no nonzero subrepresentation of 1I:"m can be equivalent to a subrepre-
sentation of 1I:"J Choose such a a ~ 1I:"n. By Corollary 1 of 4.3.1, the range
projection of a belongs to ce, and is therefore mUltiplication by the charac-
teristic function of a Borel set E ~ A". Letting v be the measure on A"
defined by v(S) = p.,,(S (") E), we have a = 11:., and the assertion now follows
from 4.3.1 and the remark preceding 4.3.2.

97
4. From Commutative Algebras to GCR Algebras

Since each subrepresentation 'ltlln of 'ltll is multiplicity-free, we can now

deduce from 2.1.6 (ii) that 'ltll is multiplicity-free. Therefore 'lt1l(A)' = ~, and
since each Pn clearly commutes with 'lt1l(A), we see that Pn belongs to ~.
To prove the last assertion ofthe theorem we claim that ~IPnJl" is the center
of 'ltIlJA)"; the remark preceding 4.3.2 will then give the desired conclusion.
Clearly ~IPnJl" is contained in the commutant of 'ltIlJA) (which is, of course,
the center of its weak closure). Conversely, if T is an operator which commutes
with 'ltIlJA), then TP n commutes with 'ltiA). Hence TPn belongs to ~, and
finally T E ~IPnJl'" 0

It is now a simple matter to see that the results of 4.3.2 persist for general
measures J.I. on A, not necessarily concentrated on a single piece An:

Corollary. Let J.I. and v be two finite Borel measures on A. Then 'ltll '" 'ltv iff
J.I. '" v, and 'ltll d 'ltv iff J.I. .1 V.

PROOF. Assume first that 'ltll '" 'It., and let U be a unitary operator such that
'ltv = U'ltIlU*. Let P 00' Ph P 2, .. · (resp. Qoo, Ql, Q2," .) be the projections
associated with the decomposition 'ltll = 'ltlloo EE> 'ltlll EE> ••• (resp. 'ltv = 'ltv., EE>
'ltv, EE> ••• ) as in 4.3.3. Then T 1-+ UTU* maps the center of'ltIl(A)" onto the
center of 'ltv(A)" in a mUltiplicity-preserving fashion. By 4.3.3 and the unique-
ness assertion of 2.1.8 we conclude that UPnU* = Qn, n = 00, 1,2, ....
In particular, the restriction of U to the range of Pn implements an equivalence
between 'ltlln and 'ltv n, n = 00, 1, 2, .... By 4.3.2 if follows that J.l.n '" Vn, and
hence J.I. = Ln J.l.n is equivalent to v = L Vn·
Conversely, if J.I. '" v and J.l.n(E) = J.I.(A n rl E) (resp. vn(E) = v(An rl E)),
then J.l.n '" vn; by 4.3.2 we have 'lt lln '" 'lt Vn and hence 'ltll = 'lt ll ., EE> 'ltll' EE> ••• is
equivalent to 'ltv = 'ltv., EB 'ltv, EB ....
If 'ltll d'ltv and we define J.l.n and Vn as in the preceding paragraph, then
'ltlln and 'lt vn are subrepresentations of'ltll and 'ltv respectively, so that 'ltlln d 'lt Vn '
4.3.2 implies J.l.n .1 Vn, and since J.l.n .1 Vm for m :F n (because An rl Am = 0)
we see that J.l.n .1 v = Ln Vn, hence J.I. =Ln J.l.n .1 v.
Finally, assume J.I. .1 v, and let J.l.n' Vn be as above. Then J.l.n .1 Vn for each
n = 00, 1, 2, ... , so by 4.3.2 we have 'ltlln d 'lt vn ' If m :F n then we also have
'ltllm d 'ltv"; for 4.3.3 applied to the measure J.l.n + Vm implies that I EE> 0 belongs
to the center of 'ltlln EE> 'ltvJA)", and as in the proof of 2.2.2 this implies
'ltlln d 'lt vm ' It follows that each 'ltlln is disjoint from 'ltv = 'ltv., EB 'ltv, EE> ••• (a
fact easily seen by considering the intertwining space of 'ltv., EE> 'ltv, EE> ••• and
'ltIlJ and similarly 'ltll = 'lt ll ., EE> 'ltll' EE> ••• is disjoint from 'ltv' 0

Thus we conclude, as in the commutative case, that the unitary equivalence

classes of the multiplicity-free representation 'ltll = f~ , dJ.l.(O correspond
1-1 with the (finite) measure classes on A.
We will now prove that every multiplicity-free representation "is" one
of these 'ltll"

98
 4.3. Representations of GCR Algebras

Theorem 4.3.4. For every nondegenerate multiplicity-:free representation n of

A, there is a finite Borel measure Jl. on A such that n is equivalent to nil"

PROOF. Let n be a multiplicity-free representation of A on a Hilbert space Yr.

Note that since A is separable, 2.2.3 implies that Yr is a separable space.
Let us first make a reduction. Decompose the center 'fl of n(A)" as in 2.1.8.
That is, find orthogonal projections Coo, C b C 2 , ••• in 'fl such that Coo +
C 1 + C 2 + ... = I and such that 'flic" Yr has multiplicity n. Let nn be the
restriction ofn to CnYr. Then n = noo El1nl El1n2 El1 ... , and if we can find
measures Jl.n concentrated on An such that nn is equivalent to nil"' then by
the discussion preceding 4.3.3 n itself will have the required form.
Therefore we can assume 'fl has multiplicity n. Now by Proposition 1.2.3
'fl contains a separable weakly dense C*-subalgebra 'flo which we may as
well assume contains the identity. Then the spectrum Z of 'flo is a compact
metric space (therefore Polish), and the inverse Gelfand map gives an
isometric representation </> of C(Z) whose range is 'flo. Now since 'flo =
</>( C(Z» is weakly dense in 'fl and since 'fl has multiplicity n, it follows that
the representation </> has multiplicity n. Thus if we let Yrn be the canonical
n-dimensional Hilbert space chosen for the construction of An in Sec-
tion 4.1, then the concluding discussion in Section 4.1 shows that there
is a finite Borel measure v on Z such that </> is unitarily equivalent to the
"multiplication" representation ft--+ L J of C(Z) on the space of vector-
valued functions L2(Z, v; Yrn). Thus we may as well take </> to be of this form.

As in Section 4.2, let aJ(Z) (resp. &I(Z, Yr n» denote the C* -algebra of all
bounded Borel functions (resp. bounded2(Yr n)-valuedBorel functions) on Z,
and let ~ = {L J:f E &I(Z)} be the algebra of all diagonal operators on
L2(Z, v; Yrn). Note that ~ is the weak closure of </>(C(Z» (by 4.1.1.) and
therefore ~ = 'fl is the commutant of n(A).
We want to associate Z with a Borel subset of An in a suitable way. This is
done as follows. By Corollary 2 of 4.2.1. we may choose an isometric "-
homomorphism W ofthe separable CO-algebra n(A) into &I(Z, Yrn) such that,
for each x E A, n(x) is "multiplication by" the function w 0 n(x). Thus for
each Z E Z we obtain a representation W z of n(A) on Yrn by wz(n(x» =
(w 0 n(x) )(z). Since n(A) is strongly dense in ~' we see by Corollary 2 of
4.2.2. that v-almost every W z is irreducible. Thus by discarding a Borel set of
v-measure zero, we can assume every W z is irreducible (note that after this
deletion Z is no longer a Polish space, but it does remain a standard Borel
space). Also, ~ C;; ~' = n(Z)" and Corollary 3 of 4.2.2. shows that by
discarding another Borel null set from Z we may assume w z , and W Z2 are
inequivalent for all Zl #- Z2 in Z. Now consider the function z t--+ W z 0 n.
This maps Z into irr(A, Yr n), different points go into inequivalent representa-
tions, and a simple check (which we omit) shows that it is in fact a Borel map.
Thus, letting q be the canonical map of irr(A, Yr n) into An, we see that
z t--+ q 0 W z 0 n is a 1-1 Borel map of the standard Borel space Z into the

99
4. From Commutative Algebras to GCR Algebras

standard Borel space An. By 3.3.2. (regarding An as a Borel subset of a Polish

space, by definition of standard Borel spaces) it follows that the range of this
map is a Borel set in An and that the function itself is a Borel isomorphism
of Z with a Borel set in An. Note that in this identification, Z r--. W z 0 1t becomes
a Borel cross section (restricted to Z) for the quotient map q: irr(A, Jt" n) -+
An. Finally, define a Borel measure Iln on An by Iln(E) = v(Z (\ E). Therefore,
this change of variables has converted each operator 1t(x), x E A, into the
operator 1tIl Jx) on _L 2(An' Iln; Jt" n) defined by i1tIlJx)~)(z) = W z 0 1t(x)~(z),
for Z E Z, 0 for Z E An\Z, where ~ runs over L 2(An' Iln; Jt" n). In other words,
1t is equivalent to 1t1l " = IT" ( dlli(). D

Let us recall now that every nondegenerate representation of a GCR

algebra is type I (2.4.1). So if that result is combined with the decomposition
theorem (2.1.8) and the preceding analysis, we see that the most general
(separable, nondegenerate) representation of A is obtained as follows. Choose
a sequence 1l0C!, Ill> 112, ... of finite Borel measure on A such that Ili 1. Ilj
if i =/; j. We do not assume Iln is concentrated on An, and in fact there is no
connection between these Iln and the Iln of the discussion preceding 4.3.3.

I: (
For each n, let 1tn be the corresponding multiplicity-free representation
dll n ((), and then define 1t by
1t = 00 . 1t E9 1 . 1t 1 E9 2 . 1t2 E9 ....
00

Then 1t is a typical representation of A. Moreover, ifvoo, Vb v2 , • •• is another

such choice of measures on An giving rise in the same way to a representation
(J, then 1t is equivalent to (J iff Iln ~ Vn for every n = 00, 1, 2, ... , and 1t d (J

iff each Iln is singular with respect to each vm .

As in the commutative case, there is a more revealing description of the
representation 00' 1t 00 E9 1 . 1t2 E9 2 . 1t2 E9 . .. obtained from the set of
measures {Iloo, Ill> 112," .}. As in 2.2, we may renormalize each Iln so that
Il = Iloo + III + 112 + ... is a finite measure, and we can find a partition
{Eoo, E 1 , E 2, . .. } of A into disjoint Borel sets so that Ilj(S) = Il(S (\ Ej),
j = 00,1,2, .... Thus we can define a Borel measurable multiplicity function
m: A -+ {oo, 1, 2, ... } exactly as in 2.2: m(O = k if ( E Ek . m is unique to a
Il-equivalence, and conversely the pair (Il, m) determines uniquely the set of
measures {Iloo, Ilb 1l2,· .. } as in 2.2. Let us rewrite the representation asso-
ciated with (11, m) symbolically as

1t = I: m(() . ( dll(().
I: I:
Conclusion. Every separable, nondegenerate representation of A is equiva-
lent to one these, and moreover m(() . ( dll(() is equivalent to n(() . ( dv(()
if and only if Il and V are equivalent measures and the multiplicity functions m
and n agree almost everywhere. These representations are disjoint iff Il and
v are mutually singular. Now of course the concrete realization of 1t suggested
by the direct integral notation has not been established, but it is a simple

100
 4.3. Representations of GCR Algebras

matter to do so. In the functional realization that we have described above,

one simply has to increase the multiplicity of ( by a factor of k uniformly
over the set Ek = g E A:m(O = k}, by taking a direct sum of k copies of
the Hilbert space associated with (. The details are an unexciting variation
on what we did at the end of Section 2.2, and are left for the interested reader.
In the case of commutative algebras A, the classification of the representa-
tions of A in Section 2.2 led directly to a complete set of unitary invariants
for separably acting normal operators in terms of measure classes on the
spectrum of the operator. Thus it is natural to expect that the results of this
section will lead to a similar classification of separably acting GCR operators
T in terms of measure classes on some sort of "generalized" spectrum of T
(recall that T is a GCR operator if C*(T) is a GCR algebra). All of this is true,
and readers who have come this far with us may wish to try their hand at for-
mulating the appropriate theorem for themselves.

101
Bibliography

I. Arveson, W. B. Unitary invariants for compact operators. Bull. Am. Math. Soc.,
76: 88-91, 1970.
2. - - Subalgebras of C"-algebras. Acta Math., 123: 141-224, 1969.
3. Blackwell, D. Proc. Nat. Acad. Sci. U.S.A., 58: 1836-1837, 1967.
4. Bourbaki, N. Elements of Mathematics, General Topology, Part 2. Reading, Mass.:
Addison-Wesley, 1966.
5. Deckard, D. and Pearcy, C. On unitary equivalence of Hilbert-Schmidt operators.
Proc. Am. Math. Soc., 16: 671-675,1965.
6. Dixmier, J. Les C*-algebres et leurs representations. Paris: Gauthier-Villars, 1964.
7. - - Les algebres ffoperateurs dans fespace Hi/bertien. Paris: Gauthier-Villars,
1957.
8. - - - Utilisation des facteurs hyperfinis dan la theorie des C"-algebres. C.R.
A cad. Sci., Paris, 258: 4184-4187, 1964.
9. - - - Dual et quasi-dual d'une algebre de Banach involutive. Trans. Am. Math.
Soc., 104: 278-283, 1962.
10. - - - Sur les structures Bon!liennes du spectre dune C*-algebre. Publ. [nst.
Hautes Etudes Sc., No.6: 297-303,1960.
II. - - Sur les C*-algebres. Bull. Soc. Math. France, 88: 95-112, 1960.
12. Fell, J. M. G. C"-algebras with smooth dual. 11/. J. Math., : 221-230, 1960.
13. Glimm, J. Type I C*-algebras, Ann. Math., 73: 572-612, 1961.
14. Halmos, P. R. A Hilbert Space Problem Book. Princeton: Van Nostrand; New
York: Springer, 1967.
15. Kaplansky, I. The structure of certain operator algebras. Trans. Am. Math. Soc.,
70: 219-255, 1951.
16. ---A theorem on rings of operators. Pac. J. Math., 1: 227-232, 1951.
17. ---Projections in Banach algebras. Ann. Math., 53: 235-249, 1951.
18. Kelley, J. L. General Topology. Princeton Van Nostrand; New York: Springer,
1955.

102
 Bibliography

19. Kuratowski, K. Topology, Vol. I. New York: Academic Press, 1966.

20. Mackey, G. Borel structures in groups and their duals. Trans. Am. Math. Soc., 85:
134-165, 1957.
21. von Neumann, J. On rings of operators: Reduction theory. Ann. Math., 50: 401-
485,1949.
22. Pearcy, C. A complete set of unitary invariants for operators generating finite
W*-algebras of type I. Pac. J. Math., 12: 1405-1416, 1962.
23. Rickari, C. E. Banach Algebras. Princeton: Van Nostrand, 1960.
24. Sakai, S. On type I C'-algebras. Proc. Am. Math. Soc., 18: 861-863, 1967.
25. - - O n a characterization of type I C-algebras. Bull. Am. Math. Soc., 72:
508-512, 1966.
26. Specht, W. Zur Theorie der MatrizenlI. Jahresbericht der Deutschen Mathematiker
Vereinigung, 50: 19-23,1940.
27. Riesz, F. and Sz-Nagy, B. Functional Analysis. New York: Frederick Ungar, 1955.

103
 Index

Abelian projection, 56 Elementary C*-a1gebra, 21

Absolutely measurable, 77 Equivalent measures, 51
Algebraically equivalent operators, 2 Equivalent representations, 14
Analytic set, 64 Equivalent sets of representations, 49
Analytic Borel space, 71 Ergodic equivalence relation, 78
Approximate identity, approximate unit, 36
Axiom of choice, 75, 79 Factor, 56
Functional calculus, 4
Borel function, 69
Borel isomorphism, 69 GCR algebra, 24
Borel set, 64 GCR operator, 26
Borel space, 69 G/},62
Borel transversal, 77 Gelfand- Naimark theorem, 34
GNS construction, 28
C*-algebra,2
C*-algebra of operators, I Homogeneous von Neumann algebra, 58
CCR algebra, 23
Central cover, 44 Ideal, 10
Central subrepresentation, 44 Involution, 2
Commutant, 5 Irreducible representation, 14
Compact operator, 17 *-isomorphism, 2
Composition series of a C*-algebra, 25
Countably generated Borel space, 72 Kaplansky's density theorem, 7
Countably separated Borel space, 69
Cross section for a function, 74 Localization, 56

Decomposable operator algebra, 89 Measure class, 53

Diagonal operator, 89 Minimal projection, 17
Direct product of C*-algebras, 21 Multiple of a representation, 20
Direct sum of C*-a1gebras, 21 Multiplicity-free representation, 41
Disjoint representations, 43 Multiplicity function, 21, 50, 55, 100
Double commutant theorem, 5 ",-measurable, 77

105
Index

Nondegenerate representation, 14 Saturated Borel set, 70

Normal operator, 3, 4 Schwarz inequality for states, 28
Null space of a set of operators, 5 Self-adjoint set of operators, 5
Separable C*-algebra, 8
Open sieve, 62 Separated sets, 65
Oribt space, 76 Sieve, 62
Orthogonal subrepresentations, 41 Singular measures, 51
Spectrum of a C*-algebra, 2, 83
Partition of a space, 66 Standard Borel space, 69
Polar decomposition, 6 State, 27
Polish space, 61 Structure space, 82
Positive element, 35 Subrepresentation, 19
Positive linear functional, 27 Subspace of a Borel space, 69
Primitive ideal, 82
Pure state, 29 Transversal for an equivalence relation, 77
Trivial ergodic measure, 78
Ouotient of a C*-algebra, 11 Type I C*-algebra, 59
Ouotient of a Borel space, 70 Type I representation, 47
Type I von Neumann algebra, 57
Refinement of a partition, 66
Representation, 13 von Neumann algebra, 6

Well-ordered Polish space, 75

106
Graduate Texts in Mathematics
conti~d from page ii

61 WIllTEHEAD. Elements of Homotopy 92 DIESTEL. Sequences and Series in Banach

Theory. Spaces.
62 KARGAPOLOv/MERLZJAKOV. Fundamentals 93 DUBROVINlFOMENKoINoVIKOV. Modern
of the Theory of Groups. Geometry-Methods and Applications.
63 BOl.LOBAS. Graph Theory. Part I. 2nd ed.
64 EDWARDS. Fourier Series. Vol. I 2nd ed. 94 WARNER. Foundations of Differentiable
65 WELLS. Differential Analysis on Complex Manifolds and Lie Groups.
Manifolds. 2nd ed. 95 SIllRYAEV. Probability. 2nd ed.
66 WATERHOUSE. Introduction to Affine 96 CONWAY. A Course in Functional
Group Schemes. Analysis. 2nd ed.
67 SERRE. Local Fields. 97 KOBLITZ. Introduction to Elliptic Curves
68 WEIDMANN. Linear Operators in Hilbert and Modular Forms. 2nd ed.
Spaces. 98 BR6cKERlToM DIECK. Representations of
69 LANG. Cyclotomic Fields ll. Compact Lie Groups.
70 MASSEY. Singular Homology Theory. 99 GRoVE/BENSON. Finite Reflection Groups.
71 FARKAslKRA. Riemann Surfaces. 2nd ed. 2nd ed.
72 STILLWELL. Classical Topology and 100 BERoICHRISTENsEN/REsSEL. Harmonic
Combinatorial Group Theory. 2nd ed. Analysis on Semigroups: Theory of
73 HUNGERFORD. Algebra. Positive Definite and Related Functions.
74 DAVENPORT. Multiplicative Number 101 EDWARDS. Galois Theory.
Theory. 2nd ed. 102 V ARADARAJAN. Lie Groups, Lie Algebras
75 HOCHSCIllLD. Basic Theory of Algebraic and Their Representations.
Groups and Lie Algebras. 103 LANG. Complex Analysis. 3rd ed.
76 IrrAKA. Algebraic Geometry. 104 DUBROVINlFoMENKoINoVIKOV. Modern
77 HEeKE. Lectures on the Theory of Geometry-Methods and Applications.
Algebraic Numbers. Part ll.
78 BURRIslSANKAPPANAVAR. A Course in 105 LANG. SL,(R).
Universal Algebra. 106 Sn..VERMAN. The Arithmetic of Elliptic
79 WALTERS. An Introduction to Ergodic Curves.
Theory. 107 OLVER. Applications of Lie Groups to
80 ROBINSON. A Course in the Theory of Differential Equations. 2nd ed.
Groups. 2nd ed. 108 RANGE. Holomorphic Functions and
81 FORSTER. Lectures on Riemann Surfaces. Integral Representations in Several
82 BOTTfru. Differential Forms in Algebraic Complex Variables.
Topology. 109 LEHTO. Univalent Functions and
83 WASHINGTON. Introduction to Cyclotomic Teicluniiller Spaces.
Fields. 2nd ed. 110 LANG. Algebraic Number Theory.
84 IRELAND/ROSEN. A Classical Introduction 111 HUSEMOLLER. Elliptic Curves.
to Modern Number Theory. 2nd ed. 112 LANG. Elliptic Functions.
85 EDWARDS. Fourier Series. Vol. ll. 2nd ed. 113 KARATZAslSHREVE. Brownian Motion and
86 VAN LINT. Introduction to Coding Theory. Stochastic Calculus. 2nd ed.
2nd ed. 114 KOBLITZ. A Course in Number Theory
87 BROWN. Cohomology of Groups. and Cryptography. 2nd ed.
88 PIERCE. Associative Algebras. 115 BERGERlGoSTIAUX. Differential Geometry:
89 LANG. Introduction to Algebraic and Manifolds, Curves, and Surfaces.
Abelian Functions. 2nd ed. 116 KELLEy/SRINIVASAN. Measure and
90 BR0NDSTED. An Introduction to Convex Integral. Vol. I.
Polytopes. 117 SERRE. Algebraic Groups and Class
91 BEARDON. On the Geometry of Qiscrete Fields.
Groups. 118 PEDERSEN. Analysis Now.
119 ROTMAN. An Introduction to Algebraic 148 ROTMAN. An Introduction to the
Topology. Theory of Groups. 4th ed.
120 ZIEMER. Weakly Differentiable Functions: 149 RATCLIFFE. Foundations of
Sobolev Spaces and Functions of Hyperbolic Manifolds.
Bounded Variation. 150 EISENBUD. Commutative Algebra
121 LANG. Cyclotomic Fields I and II. with a View Toward Algebraic
Combined 2nd ed. Geometry.
122 REMMERT. Theory of Complex Functions. 151 SILVERMAN. Advanced Topics in
Readings in Mathematics the Arithmetic of Elliptic Curves.
123 EBBINGHAus/HERMEs et al. Numbers. 152 ZiEGLER. Lectures on Polytopes.
Readings in Mathematics 153 FULTON. Algebraic Topology: A
124 DUBRovINIFoMENKoINoVIJwv. Modem First Course.
Geometry-Methods and Applications. 154 BROWN/PEARCY. An Introduction to
Part III. Analysis.
125 BERENSTEIN/GAY. Complex Variables: An 155 KASSEL. Quantum Groups.
Introduction. 156 KEcHRIs. Classical Descriptive Set
126 BOREL. Linear Algebraic Groups. 2nd ed. Theory.
127 MASSEY. A Basic Course in Algebraic 157 MALLIAVIN. Integration and
Topology. Probability .
128 RAUCH. Partial Differential Equations. 158 ROMAN. Field Theory.
129 fuLTON/HARRIS. Representation Theory: 159 CONWAY. Functions of One
A First Course. Complex Variable II.
Readings in Mathematics 160 LANG. Differential and Riemannian
130 DODSON/POSTON. Tensor Geometry. Manifolds.
131 LAM. A First Course in Noncommutative 161 BORWEIN/ERDELVI. Polynomials and
Rings. Polynomial Inequalities.
132 BEARDON. Iteration of Rational Functions. 162 ALPERIN/BELL. Groups and
133 HARRIS. Algebraic Geometry: A First Representations.
Course. 163 DIXON/MORTIMER. Pennutation
134 ROMAN. Coding and Infonnation Theory. Groups.
135 ROMAN. Advanced Linear Algebra. 164 NATIiANSON. Additive Number Theory:
136 ADKINslWEINTRAUB. Algebra: An The Classical Bases.
Approach via Module Theory. 165 NATIiANSON. Additive Number Theory:
137 AxLERIBoURDONIRAMEY. Harmonic Inverse Problems and the Geometry of
Function Theory. Sumsets.
138 COHEN. A Course in Computational 166 SHARPE. Differential Geometry: Cartan's
Algebraic Number Theory. Generalization of Klein's Erlangen
139 BREDON. Topology and Geometry. Program.
140 AUBIN. Optima and Equilibria. An 167 MORANDI. Field and Galois Theory.
Introduction to Nonlinear Analysis. 168 EWALD. Combinatorial Convexity and
141 BECKERIWEISPFENNINGlKREoEL. Grobner Algebraic Geometry.
Bases. A Computational Approach to 169 BHATIA. Matrix Analysis.
Commutative Algebra. 170 BREDON. Sheaf Theory. 2nd ed.
142 LANG. Real and Functional Analysis. 171 PETERsEN. Riemannian Geometry.
3rd ed. 172 REMMERT. Classical Topics in Complex
143 DOOB. Measure Theory. Function Theory.
144 DENNIsIFARB. Noncommutative 173 DIESTEL. Graph Theory.
Algebra. 174 BRIDGES. Foundations of Real and
145 VICK. Homology Theory. An Abstract Analysis.
Introduction to Algebraic Topology. 175 LICKORISH. An Introduction to Knot
2nd ed. Theory.
146 BRIDGES. Computability: A 176 LEE. Riemannian Manifolds.
Mathematical Sketchbook. 177 NEWMAN. Analytic Number Theory.
147 ROSENBERG. Algebraic K-Theory 178 CLARKEILEDYAEV/STERNlWoLENSKI.
and Its Applications. Nonsmooth Analysis and Control Theory.