Fundamentals of Functional Analysis

Universitext
DouglasFarenick
Fundamentals
of Functional
Analysis
Universitext
Universitext
Series Editors
Sheldon Axler
San Francisco State University
Carles Casacuberta
Universitat de Barcelona
Angus MacIntyre
Queen Mary University of London
Kenneth Ribet
University of California, Berkeley
Claude Sabbah
cole polytechnique, CNRS, Universit Paris-Saclay, Palaiseau
Endre Sli
University of Oxford
Wojbor A. Woyczynski
Case Western Reserve University Cleveland, OH
Universitext is a series of textbooks that presents material from a wide variety of

mathematical disciplines at masters level and beyond. The books, often well classtested by their author, may have an informal, personal even experimental approach
to their subject matter. Some of the most successful and established books in the
series have evolved through several editions, always following the evolution of
teaching curricula, to very polished texts.
Thus as research topics trickle down into graduate-level teaching, first textbooks
written for new, cutting-edge courses may make their way into Universitext.
More information about this series at http://www.springer.com/series/223
Douglas Farenick
Fundamentals of Functional
Analysis
123
Douglas Farenick
Department of Mathematics and Statistics
University of Regina
Regina, Saskatchewan, Canada
ISSN 0172-5939
ISSN 2191-6675 (electronic)
Universitext
ISBN 978-3-319-45631-7
ISBN 978-3-319-45633-1 (eBook)
DOI 10.1007/978-3-319-45633-1
Library of Congress Control Number: 2016954644
Mathematics Subject Classification (2010): 28-01, 46-01, 47-0, 54-01
Springer International Publishing Switzerland 2016
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made.
Printed on acid-free paper
This Springer imprint is published by Springer Nature
The registered company is Springer International Publishing AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
To Katherine
Preface
Functional analysis tightly interweaves concepts from analysis, topology, geometry,

and algebra. This book explores the fabric of functional analysis by providing a careful presentation of the fundamental results that are essential for the contemporary
mathematician. Organised into four thematic sections of (i) topology, (ii) measure
theory and integration, (iii) Banach spaces, and (iv) operator theory, the treatment is
entirely self-contained with the exception of a brief introductory section in Chapter 1
that draws attention to, but does not develop, the foundations upon which functional
analysis rests (e.g., the axiom of choice and the completeness of the real number
system).
The reader is assumed to have knowledge of basic real and complex analysis
at the level of the classic monographs of Rudin [49] and Brown and Churchill [8]
and to possess an understanding of algebra and linear algebra at the level of, say,
Hersteins well-known book [33]. However, as it is virtually impossible to study the
fundamental results of functional analysis without a working facility with topology
and measure theory, these two subjects form the books starting point. Both topics
are introduced on the assumption that the reader has no prior knowledge of them.
The reader who is familiar with these subjects can select a later entry point to
the book, knowing that he or she may refer back to the topological and measuretheoretic results as needed.
The topology treated herein is essentially of the point-set variety, the first
chapter of which covers the generic features in topology (subspaces, products,
continuity, quotients) while examining some specific and important examples in
detail (such as the Cantor set and the Cantor ternary function). The second half of the
study of topology concentrates on those structures and results that are required by
functional analysis: compactness and local compactness, Urysohns Lemma, second
countability, and the Stone-Cech

compactification, for example. A novel feature of
Chapter 2 is the inclusion of a purely set-theoretic proof of the Tietze extension
theorem.
The primary examples of Banach spaces are drawn from vector spaces of
integrable functions, and among the most important of such spaces are the ones
in which the underlying measure space is a topological space. Therefore, the second
vii
viii
Preface
thematic section of the book is devoted to the basics of measure theory, with
particular emphasis on regular Borel measures and Lebesgue measure. Because
spaces of complex measures also form a Banach space of interest, these are
examined as well. The Lebesgue integral is also developed and the essential
inequalities of analysis (Jensen, Hlder, Minkowski) are established.
The third part of the book is devoted to Banach and Hilbert spaces, duality, and
convexity. Equipped with topology and measure theory, the important examples in
functional analysis are studied, with Lp spaces and spaces of continuous complexvalued functions on a compact Hausdorff space being among these. The Riesz
representation theorem for the dual of C.X/, where X is compact Hausdorff, is
proved only in the case where the topology on X is metrisable. As it turns out,
this is the only case that is required for subsequent results in the book, including
Choquets integral representations for elements of compact convex sets.
The final part of the book concerns operator theory, beginning with a study
of bounded linear operators acting on Banach spaces and the spectral features
of operators. Because spectral theory is complex-analytic in nature, it seems
appropriate at this stage of the book to introduce and study spectral theory in the
context of Banach algebras. Although the treatment is relatively brief, the main
results about abelian Banach algebras are established, and a few applications to
classical analysis are made. The final two chapters deal with operators acting on
Hilbert spaces and self-adjoint algebras of such operators. The first of these two
chapters provides the basic theory of Hilbert space operators and includes the
spectral theory of compact normal operators, as well as a detailed analysis of
the Banach space of trace-class operators. Matrices of operators and some of the
most widely used operator inequalities are developed. The study of von Neumann
algebras and C -algebras in the final chapter is approached as a natural continuation
of basic Hilbert space operator theory and brings the book to a close by touching
upon some of its topological and measure-theoretic beginnings.
What is not in this book? I did not touch upon any of the multivariable aspects
of functional analysis, which explains the absence of product measures, Fubinis
theorem, and tensor products. And while the core topics of functional analysis
do come very close to those of harmonic analysis, I felt that the treatment of
Haar measure and the Fourier transform, for example, would be better suited to
books devoted to harmonic analysis. Some of the material that one would need
for applications of functional analysis to partial differential equations or physics,
such as unbounded operators and distributions, is also omitted. Although complex
analysis has a major role in the theory of Banach algebras and non-self-adjoint
operator algebras, my approach here has been to view the study of functional
analysis as a continuation of real analysis, which accounts somewhat for the
inclusion of measure theory and self-adjoint algebras of Hilbert space operators
at the cost of excluding Banach spaces of analytic functions and, for example,
holomorphic functional calculus.
This book can be read for self-study or be used as a text for a course on one of
the thematic sections or for a sequence of courses. It may also be used as a reference
work, keeping in mind the omission of certain significant topics as mentioned above.
Preface
ix
I have discussed many aspects of functional analysis with numerous colleagues,

friends, and students over the years, and I am grateful for all that I have learned from
them. I wish to acknowledge, in particular, Angshuman Bhattacharya and Sarah
Plosker, who patiently worked through some of the material on convexity with me,
and Martn Argerami, who used early versions of the measure theory and integration
chapters in his courses on real analysis.
In writing this book, I am especially indebted to Gwyneth Bergman for her
editorial work on initial drafts of the manuscript and to anonymous reviewers
who provided thoughtful criticism and suggestions. Lastly, I wish to thank Donna
Chernyk and Mahmood Alaghmandan for getting me started on this project and for
their gentle, regular reminders to get it completed.
Regina, Saskatchewan, Canada
June 2016
Douglas Farenick
Contents
Part I Topology
1
Topological Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1
Sets and Partial Orderings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2
Completeness of the Real Number System . . . . . . . . . . . . . . . . . . . . . . . . .
1.3
Topological Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4
Metric Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5
Subspaces and Product Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6
Closures, Interiors, and Limit Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.7
Continuous Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.8
The Cantor Ternary Set and Ternary Function . . . . . . . . . . . . . . . . . . . . . .
1.9
Weak Topologies and Continuous Maps of Product Spaces . . . . . . .
1.10 Quotient Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.11 Topological Equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
3
7
8
13
17
22
25
27
30
32
33
34
Topological Spaces with Special Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

2.1
Compact Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.2
Topological Properties Described by Nets . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3
Normal Spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.4
Properties of Metric Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5
Connected Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6
Stone-Cech
Compactification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
39
39
48
51
57
63
66
71
Part II Measure Theory and Integration

3
Measure Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1
Measurable Spaces and Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2
Measure Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3
Outer Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4
Lebesgue Measure. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
77
77
83
88
95
xi
xii
Contents
3.5
Atomic and Non-Atomic Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.6
Measures on Locally Compact Hausdorff Spaces . . . . . . . . . . . . . . . . . .
3.7
Signed and Complex Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
106
107
110
115
Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1
Integration of Nonnegative Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2
Integrable Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3
Complex-Valued Functions and Measures . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4
Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5
Fundamental Theorem of Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.6
Series. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.7
Integral Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Appendix: The Issue of How to Integrate. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
119
119
126
133
136
139
142
144
152
157
Part III Banach Spaces

5
Banach Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1
Normed Vector Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2
Subspaces, Quotients, and Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3
Banach Spaces of Continuous Functions. . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4
Banach Spaces of p-Integrable Functions . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5
Essentially-Bounded Measurable Functions . . . . . . . . . . . . . . . . . . . . . . . .
5.6
Banach Spaces of Complex Measures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.7
Separable Banach Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.8
Hilbert Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.9
Orthonormal Bases of Hilbert Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.10 Sums of Vectors and Hilbert Spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
165
165
173
175
181
187
189
191
195
202
208
210
Dual Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1
Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2
Linear Functionals. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.3
Hahn-Banach Extension Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.4
The Second Dual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.5
Weak Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.6
Linear Functionals on Lp and L1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.7
Linear Functionals on C.X/. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
215
215
220
223
226
227
232
239
245
Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.1
Convex Sets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.2
Separation Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3
Extreme Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.4
Extremal Regular Probability Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . .
249
249
251
255
259
Contents
xiii
7.5
Integral Representations of Compact Convex Sets . . . . . . . . . . . . . . . . . 261
7.6
The Range of Non-Atomic Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
Part IV Operator Theory
8
Banach Space Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

8.1
Examples of Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.1
Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.2
Operators on Finite-Dimensional Banach Spaces . . . . . . . . .
8.1.3
Finite-Rank Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.4
Integral Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.5
Multiplication Operators on C.X/ . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.6
Integral Operators on C.0; 1/ . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.7
Multiplication Operators on Lp .X; ; /, 1 p < 1 . . . .
8.1.8
Weighted Unilateral Shift Operators . . . . . . . . . . . . . . . . . . . . . . .
8.1.9
Adjoint Operators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2
Mapping Properties of Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3
Inversion of Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.4
Idempotents and Complemented Subspaces . . . . . . . . . . . . . . . . . . . . . . . .
8.5
Compact Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.6
Operator Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.7
The Spectrum of an Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.8
Eigenvalues and Approximate Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . .
8.9
The Spectra and Invariant Subspaces of Compact Operators . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
271
271
271
272
273
273
273
274
274
275
275
276
279
281
284
289
292
298
302
305
Spectral Theory in Banach Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

9.1
Banach Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2
Ideals and Quotients. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.3
Abelian Banach Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.4
Absolutely Convergent Fourier Series. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.5
The Exponential Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
309
309
314
316
323
325
326
10
Hilbert Space Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10.1 Hilbert Space Duality and Adjoint Operators . . . . . . . . . . . . . . . . . . . . . . .
10.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.2.1 Matrix Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.2.2 Multiplication Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.2.3 The Bilateral Shift Operator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.2.4 Toeplitz Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.2.5 Weighted Unilateral Shift Operators . . . . . . . . . . . . . . . . . . . . . . .
10.2.6 Rank-1 Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.2.7 Direct Sums of Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
329
329
335
335
335
336
336
338
338
339
xiv
11
Contents
10.3 Hermitian Operators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10.4 Normal Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.5 Positive Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.6 Polar Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.7 Strong and Weak Operator Topologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.8 Matrices of Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
10.9 Singular Values and Trace-Class Operators . . . . . . . . . . . . . . . . . . . . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
339
346
352
358
364
369
373
388
Algebras of Hilbert Space Operators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.2 von Neumann Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.3 Irreducible Operator Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.4 Abelian Operator Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.5 C -Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.6 Positive Elements and Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.7 Ideals and Quotients. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.8 Representations and Primitive Ideals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.9 Traces and Factors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
393
393
397
401
406
409
416
422
425
433
438
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
Part I
Topology
Chapter 1
Topological Spaces
At its most intuitive level, topology is the study of geometric objects and their
continuous deformations. Such objects might be concrete and easy to visualise, such
as a disc or an annulus in the plane R2 , or a sphere or torus in R3 , while other such
geometric objects might be a great deal more abstract in nature, as is the case with
higher-dimensional spheres and tori, groups of unitary matrices, and the unit balls of
the dual spaces of normed vector spaces. Fortunately, the concepts and methods of
topology apply at a very general level, thereby providing us with rather deep insight
into the nature of a variety of interesting mathematical structures.
This chapter is devoted to the introduction and study of generic features and
constructions of spaces in topology. The study of topological spaces that have
certain special properties is taken up in the second chapter. A topological space
is defined by set-theoretic axioms; for this reason, we begin with a very brief word
on set theory.
1.1 Sets and Partial Orderings

It is difficult to think of a branch of mathematics that does not depend crucially
on the formalism and language of set theory. Elementary uses of set theory are so
common as to be unnoticed: set union and intersection, subsets, Cartesian products,
relations and functions, and so forth. However, certain subtle aspects of set theory
(the Axiom of Choice, for example) lead to powerful tools (such as Zorns Lemma)
that are fundamental features of many areas of modern mathematics.
If one is not precise about what is meant by the term set, then it is possible to
introduce undesired paradoxes. In a non-rigorous discussion, it is common to refer to
a set as a collection of objects in which elements of the collection are determined or

D. Farenick, Fundamentals of Functional Analysis, Universitext,
DOI 10.1007/978-3-319-45633-1_1
1 Topological Spaces
described in some definitive manner. The set H of all hydrogen atoms, for example,
contains each and every hydrogen atom and no other elements; that is, x 2 H if and
only if x is a hydrogen atom.
A paradox put forward by the philosopher Bertrand Russell shows that a more
rigorous definition of set is required. Specifically, consider the set
R D fx j x is a set and x 62 xg:
By definition, R 2 R if and only if R 62 R, and therein lies the paradox.
To avoid contradictory assertions such as Russells paradox, various axioms
for set theory have been developed to formally describe what sets are and what
operations on sets may be carried out. The most widely used (and accepted) of these
axiomatic systems are the Zermelo-Fraenkel Axioms. The ZF Axioms assert, for
example, the existence of the empty set and the existence of an infinite set. These
axioms also define precisely the concept of subset and they confirm that the familiar
operations of set union and set intersection as valid operations in set theory.
If X and Y are sets, then the Cartesian product of X and Y is denoted by X Y,
and consists of all ordered pairs .x; y/ in which x 2 X and y 2 Y. More generally, if
X1 ; : : : ; Xn are sets, then X1 Xn is the set of all n-tuples .x1 ; : : : ; xn / such each
xj 2 Xj for each j D 1; : : : ; n.
If Y is a subset of a set X, then the notation Y c or X n Y will be used to denote the
complement of Y in X: namely, the set of all x 2 X for which x 62 Y.
Recall the following basic fact regarding unions and intersections.
Proposition 1.1. If fY g2 and fZ g2 are families of subsets of a set X, then
[
!
Y
0
@
2
1
Z A D
!
Y
2
0
@
Y \ Z ; and
.;/2
1
Z A D
Y [ Z :
.;/2
With respect to Cartesian products, union and intersection satisfy

[
1
! 0
[
Y @
Z A D
2
.;/2
1
! 0
\
Y @
Z A D
2
.;/2
Y Z ; and
Y Z :
1.1 Sets and Partial Orderings
With respect to complements, DeMogans Laws hold:

[
2
!c
X
\
2
Xc
and
\
2
!c
X
Xc :
2
A more extensive review of the ZF Axioms will not be undertaken here because
it is unnecessary for the subjects addressed in this book. However, it is worth
noting that quite recent research in functional analysis has drawn substantially upon
sophisticated techniques and results in abstract set theory, showing that set theory
has a role in analysis which reaches far beyond the foundational role that we are
focusing upon here.
In addition to the ZF Axioms, it is essential in functional analysis to use one
additional axiom: the Axiom of Choice. In informal language, the Axiom of Choice
asserts that if X is a set of pairwise disjoint nonempty sets, then there exists a set C
that contains exactly one element from each set in X. (Note, here, that X is a set of
sets.) Conceptually, C is constructed by choosing one element from each set in X.
The Zermelo-Fraenkel Axioms together with the Axiom of Choice are referred to
as the ZFC Axioms.
The existence of the set N D f1; 2; 3; : : : g of natural numbers results from an
axiomatic system developed by G. Peano. The : : : in f1; 2; 3; : : : g is not, of course,
rigorously defined; making formal sense of this unspecified portion of N is exactly
what the Peano axioms were created to do. Without pursuing these axioms in more
detail, let us accept as given that the Peano axioms establish N as a set in accordance
with the ZF axioms. (See, for example, [21, Chapter 1] for a detailed treatment of
these facts.)
In contrast to the subtle nature of the axioms that define a set, the definitions
of relations and functions are simple. A relation R from a set X to a set Y is a
subset of X Y, whereas a function is a relation F from X to Y such that, for any
x 2 X, if y1 ; y2 2 Y satisfy .x; y1 / 2 F and .x; y2 / 2 F, then necessarily y1 D y2 . It is
cumbersome to express functions as subsets of Cartesian products, and we normally
write y D f .x/ if .x; y/ 2 F (and so the symbol f depends on the set F).
Recall that if X and Y are sets, then a function f W X ! Y is injective if
f .x1 / D f .x2 / holds only for x1 D x2 , and is surjective if for every y 2 Y there is
at least one x 2 X such that y D f .x/. A function that is both injective and surjective
is said to be bijective.
Definition 1.2. Two sets, X and Y, are in bijective correspondence if there exists a
bijection f W X ! Y.
The following theorem is a useful criterion for bijective correspondence.
Theorem 1.3 (Schroeder-Bernstein). If, for sets X and Y, there exist injective
functions g W X ! Y and h W Y ! X, then X and Y are in bijective correspondence.
In comparing sets X to the set N of natural numbers, the Schroeder-Bernstein
Theorem has the following simpler form.
Theorem 1.4. The following conditions are equivalent for an infinite set X:
1. there exists an injection g W X ! N;
2. there exists a surjection h W N ! X;
3. there exists a bijection f W X ! N.
Recall the concept of countability, which is an essential idea in topology and
analysis.
Definition 1.5. A set X is countable if X is in bijective correspondence with some
subset of N.
Theorem 1.4 allows one to deduce, for example, that subsets of countable sets
are countable and that the union of a countable collection of countable sets is a
countable set.
The formal notion of partial order allows for a qualitative comparison of certain
elements in a set, and it also leads to the important concept of a maximal element.
Definition 1.6. A partial order on a set S is a relation denoted by that has the
following properties:
1. (reflexivity) for all 2 S, ;
2. (antisymmetry) for all ; 2 S, and implies that D ;
3. (transitivity) for all ; ; 2 S, b and implies that .
If, in addition, a partial order satisfies property (iv) below, then is called a linear
order.
(iv) (comparability) for all ; 2 S, either or .
A partial order on a set S can be used to assert that is in some sense larger
than , if it happens to be true that . This manner of thinking leads to a natural
notion of largest element.
Definition 1.7. In a partially ordered set S, an element 2 S is said to be a
maximal element if for 2 S the relation implies that D . If E S,
then an element 2 S is an upper bound for E if for every 2 E.
Because a partially ordered set need not be linearly ordered, if a partially ordered
set S has a maximal element, then this maximal element may or may not be unique.
The existence of a maximal element is established by Zorns Lemma.
Theorem 1.8 (Zorns Lemma). If S is a nonempty partially ordered set such that
every linearly ordered subset E S (where the linear order on E is inherited from
the partial order on S) has an upper bound in S, then S has a maximal element.
A detailed discussion of the facts reviewed above can be found in [21, 41, 49].
1.2 Completeness of the Real Number System
1.2 Completeness of the Real Number System

In the foundations of arithmetic, one begins with the Peano axioms that define the
set N of natural numbers and then proceeds to construct the set of Z of integers
by considering formal differences n m of natural numbers n; m 2 N. The set Z of
integers is more than just a setit is an arithmetic system as well. Specifically, in the
language of modern algebra, Z is a commutative unital ring and, therefore, it has a
field of fractions, which is the countable set Q of rational numbers. Constructing the
real numbers system R from the rational field Q is a fairly elaborate process due to
R. Dedekind; see [49] for the details of the construction. The end result, namely the
uncountable set R of real numbers, is once again a field, but it is no ordinary field, as
R is distinguished by its completeness as both a metric space and an ordered space.
By metric is meant the usual distance between real numbers; that is, if x; y 2 R,
then the distance between x and y is the absolute value jx yj of x y. By order is
meant the usual order on the real numbers whereby, if x; y 2 R, then exactly one of
the following three statements holds true: (i) x < y, (ii) y < x, or (iii) y D x.
Definition 1.9. If S is a nonempty set of real numbers, then
1. S is bounded above if there exists a real number z such that s z for every s 2 S,
and
2. S is bounded below if there exists a real number y such that y s for every s 2 S.
A real number z that satisfies condition (1) in Definition 1.9 is called an upper
bound for S, while the real number y that satisfies condition (2) is called a lower
bound for S.
Definition 1.10. A real number z is the supremum (or least upper bound) of a
nonempty set S of real numbers if
1. z is an upper bound for S, and
2. z g for every upper bound g of S.
The definition of infimum (or greatest lower bound) of S is similarly defined.
The completeness of R as an ordered space (which is to say that suprema exist)
is one of the most important and extraordinary features of the real number system.
Theorem 1.11 (Least Upper Bound Principle). If S R is a nonempty subset
that is bounded above, then S has a supremum in R.
To discuss the completeness of R as a metric space, recall the formal definition
of sequence.
Definition 1.12. An infinite sequence in a set X is a function f W N ! X.
Definition 1.12 makes no claims about whether the function f W N ! X is injective
or surjective, as neither need hold. It is typical to denote the element f .n/ 2 X by xn ,
for each n 2 N, and to write the infinite sequence f as fxn gn2N .
Throughout this book, we shall drop the adjective infinite and refer to infinite
sequences as sequences.
Definition 1.13. A sequence fxn gn2N is:
bounded above if there is a real number z such that xn z for every n 2 N;
bounded below if there is a real number y such that y xn for every n 2 N;
bounded if it is bounded above and below;
convergent if there is a real number x with the property that for each " > 0 there
exists N 2 N such that jx xn j < " for all n N; and
5. a Cauchy sequence for each " > 0 there exists N 2 N such that jxm xn j < " for
all m; n N.
1.
2.
3.
4.
The following theorem establishes the completeness of R as a metric space.

Theorem 1.14. Every Cauchy sequence of real numbers is convergent.
Theorems 1.11 and 1.14, the proofs of which can be found in [49], underly all of
real and functional analysis.
Definition 1.15. A set X has the cardinality of the continuum if X is in bijective
correspondence with R.
The cardinal number c shall henceforth denote the cardinality of the continuum.
1.3 Topological Spaces

Although the notion of a topology is abstract, the definition itself is motivated by
our experience with the real numbers system. For example, an intersection of a finite
number of open intervals results in the empty set or another open interval; and while
the union of a family of open intervals need not be an open interval, it is clear that
such a set retains features exhibited by open intervals. Such basic observations lead
to the following definition.
Definition 1.16. A topology on a set X is a collection T of subsets of X such that:
1. T contains both the empty set ; and the set X itself;
2. T is closed under arbitrary unions:
[
U 2 T ;
2
for every family fU g2 of sets U 2 T ;

3. T is closed under finite intersections:
n
\
kD1
Uk 2 T ;
for every finite collections fUk gnkD1 of sets Uk 2 T .

The pair .X; T /, but more often just X itself, is called a topological space.
Elements U 2 T are called open sets.
Example 1.17 (Examples of Topologies). The following collections of sets form a
topology on a nonempty set X:
1. the indiscrete topology Tindiscr D f;; Xg;
2. the discrete topology Tdiscr D fU j U Xg, which is the power set P.X/ of X;
3. the co-finite topology Tcofin D f;g [ fF c j F X; F is finiteg, which consists of
the empty set and the complements of all finite subsets of X; and
4. the co-countable topology Tcocntbl D f;g [fF c j F X; F is countableg, consisting of the empty set and the complements of all countable subsets of X.
Proof. The indiscrete and discrete topologies trivially satisfy the axioms for a
topology on X.
To verify that Tcocntbl is indeed a topology on X, note that ; 2 Tcocntbl by
construction and that X D ;c 2 Tcocntbl because the empty set Tcocntbl is countable.
If we now suppose that fU g is a family of sets U 2 Tcocntbl , then for each there
is a countable subset F X such that U D Fc . Therefore,
[
U D
because
Fc
!c
F
2 Tcocntbl ;
F is a subset of each of the sets F and is, therefore, countable.
This shows that Tcocntbl is closed under arbitrary unions. Finally, if U1 ; : : : ; Un 2

Tcocntbl , and if F1 ; : : : ; Fn X are countable subsets for which each Uj D Fjc , then
n
\
jD1
Uj D
n
\
jD1
1c
0
n
[
Fjc D @ Fj A 2 Tcocntbl
jD1
because a finite union of countable sets is countable. Hence, Tcocntbl is closed under
finite unions.
The proof that Tcofin is a topology on X is similar and simpler.
t
u
Definition 1.18. If a set X is endowed with two topologies, say T and T 0 , and if
T T 0 , then we say that T is coarser than T 0 , and that T 0 is finer than T .
Using the terminology above, the indiscrete topology is the coarsest topology
that a set X admits, whereas the discrete topology is the finest topology on X.
To arrive at further examples of topological spaces, it is useful to have tools that
offer simple ways to prescribe what is meant by an open set. Two such tools are to
be found in the notions of basis and subbasis for a topology.
10
Definition 1.19. A collection B of subsets of a set X is called a basis of subsets of

X if:
1. for every x 2 X, there is a B 2 B with x 2 B; and
2. for all B1 ; B2 2 B, and every x 2 B1 \ B2 , there is a B3 2 B such that x 2 B3
B1 \ B2 .
The relevance of the definition of basis is revealed by the following proposition.
Proposition 1.20. If B is a basis of subsets of X and if T is the collection of all
subsets U X with the property that for each x 2 U there is a B 2 B with x 2 B U,
then T is a topology on X.
Proof. It is clear that both the empty set ; and X satisfy the requirements for
membership in T .
[
Suppose that fU g2 is a family of sets U 2 T and let U D
U . Choose
any x 2 U; thus, x 2 U for some 2 . Because U 2 T , there is a B 2 B such that

x 2 B U . As U U, we conclude that U 2 T and, therefore, that T is closed
under arbitrary unions.
To show that T is closed under finite intersections, it is enough to show, for
all U1 ; U2 2 T , that U1 \ U2 2 T , and to then proceed inductively. If U1 ; U2 2 T
satisfy U1 \ U2 D ;, then U1 \ U2 2 T trivially. Therefore, suppose that there exists
an x 2 U1 \ U2 . Because B is a basis, there are B1 ; B2 2 B such that x 2 Bk Uk ,
for k D 1; 2, and there is a B3 2 B such that x 2 B3 B1 \ B2 . Hence, x 2 B3
B1 \ B2 U1 \ U2 , which proves that U1 \ U2 2 T . Proceeding by induction, we
conclude that T is closed under finite intersections.
t
u
Proposition 1.20 illustrates that to specify a topology on a set X, it is sufficient to
specify a basis B of subsets of X for the topology.
Definition 1.21. If B is a basis of subsets of X, and if T is the collection of all
subsets U X for which each x 2 U is contained in at least one B 2 B satisfying
B U, then T is called the topology generated by B.
To this point we have demonstrated that a basis of subsets determine a topology.
Conversely, every topology T on X admits a basis B of subsets of X such that T
is the topology generated by B.
Proposition 1.22. If T is a topology on X, and if B is the collection of subsets with
the property that, for every U 2 T and each x 2 U, there is a B 2 B with x 2 B U,
then B is a basis of subsets and T is the topology generated by B.
Proof. Exercise 1.103.
t
u
In light of Propositions 1.20 and 1.22, the following definition is made.

Definition 1.23. If T is a topology on X, then a basis for the topology T is a
collection B of subsets of X such that (i) B is a basis of subsets of X (in the sense
of Definition 1.19) and (ii) T is the topology generated by B (in the sense of
Definition 1.22). The elements of a basis B for a topology T on X are called basic
open sets.
11
Proposition 1.24. If T is a topology on X, and if B is a basis for the topology T ,

then the following statements are equivalent for a subset U X:
1. U 2 T ;
[
2. there is a family fB g of subsets B 2 B such that U D
B .
t
u
Proposition 1.20 and our knowledge of the real number system lead us to endow
the set R of real numbers with a topology.
Definition 1.25. Assume that a; b 2 R with a < b.
1. The subset of R denoted by .a; b/, and defined by .a; b/ D fx 2 R j a < x < bg, is
called an open interval.
2. The subset of R denoted by a; b, and defined by a; b D fx 2 R j a x bg, is
called a closed interval.
3. The subsets of R denoted by a; b/ and .a; b, and defined by a; b/ D fx 2 R j a
x < bg and .a; b D fx 2 R j a < x bg, are called half-open intervals.
The term open in the definition above is the traditional terminology of calculus
and real analysis and does not, a priori, refer to an open set in the sense of topology.
However, in an appropriate topology on R, these open intervals will in fact be open
sets.
Proposition 1.26. Let B be the set of all finite open intervals with rational end
points. That is, assume that
B D f.p; q/ R j p; q 2 Q; p < qg :
Then B is a basis of subsets of R.
t
u
Definition 1.27. The topology on the real number system R generated by the basis
B D f.p; q/ j p; q 2 Q; p < qg
of subsets of R is called the standard topology of R.
In elementary real analysis, open sets in R are defined in a different fashion. The
next proposition reconciles these two definitions, showing that the standard topology
on R yields open sets that are familiar from calculus and real analysis.
Proposition 1.28. If R is endowed with the standard topology, then the following
statements are equivalent for a subset U R:
1. U is a open set;
2. for every x 2 U there is a " > 0 such that
12
.x "; x C "/ U:
Proof. Suppose that assertion (1) holds. Suppose that U R is an open set, and let
B denote by B the basis of Proposition 1.26 for the standard topology T on R.
Choose x 2 U. As B is a basis for T , there exists B 2 B such that x 2 B U.
By definition, there are p; q 2 Q such that B D .p; q/. If " D minfq x; x pg, then
.x "; x C "/ .p; q/. Hence,
.x "; x C "/ .p; q/ U:
Conversely, suppose that statement (2) holds, and that U R satisfies hypothesis
(2). Thus, if x 2 U, then there is a " > 0 such that .x "; x C "/ U. Between any
two real numbers there is a rational number, and so let px ; qx 2 Q be such that x " <
px < x and x < qx < x C ". Then .px ; qx / 2 B and x 2 .px ; qx / .x "; x C "/ U.
Continuing this procedure for each x 2 U leads to:
U
.px ; qx / U:
x2U
The inclusions above show that U is a union of the family f.px ; qx /gx2U of basic
open sets; hence, U is an open set in the standard topology of R.
t
u
Corollary 1.29. Open intervals are open sets in the standard topology of R.
The order completeness of the real number system allows for a description of
open sets in R in terms of pairwise disjoint open intervals.
Proposition 1.30 (Cantors Lemma). If U R is open in the standard topology
of R, then there is a countable family fJn gn2N such that:
1. Jn T
is an open interval, for all n;
Jm D ;, if m 6D n; and
2. J[
n
3.
Jn D U.
n
Proof. For each x 2 U there exists " > 0 such that .x "; x C "/ U, by Proposition 1.28. Therefore, if x 2 U, then
ax D inf fu 2 R j .u; x/ Ug and bx D sup fz 2 R j .x; z/ Ug
are well defined. It may be that ax D 1 or bx D C1, or both. If bx is finite, then
bx 62 U; likewise, if ax is finite, then ax 62 U. Hence, each x 2 U determines an open
interval .ax ; bx / U, which shall be denoted by J.x/. Observe that
UD
[
x2U
J.x/
1.4 Metric Topologies
13
and that J.x/ D J.y/, for all y 2 J.x/. Indeed, if x1 ; x2 2 U, then either J.x1 / D J.x2 /
or J.x1 / \ J.x2 / D ;. Thus, U is the union of a family of pairwise disjoint open
intervals. In each open interval there is a rational number, and so each of these
intervals J.x/ can be labeled by a rational number. Because there are countably many
rational numbers, there are also countably many distinct intervals of the type J.x/.
t
u
While the use of a basis to specify a topology on X is convenient, it is often even
more convenient to define a topology by way of a much smaller collection of sets.
Definition
1.31. A collection S of subsets of X is a subbasis of subsets of X if
[
S D X.
S2S
The concept of subbasis is important in topology because of the next result.

Proposition 1.32. If S is a subbasis of subsets of X, then
BD
8
n
<\
:
Sj j n 2 N; S1 ; : : : ; Sn 2 S
jD1
9
=
;
is a basis of subsets of X.
t
u

Metrics provide a measure of the distance between any two points in a set; as a
consequence, metrics can be used to specify a topology on the set.
Definition 1.33. A metric on a set X is a function d W X X ! 0; 1/ such that, for
all x; y; z 2 X,
1.
2.
3.
4.
d.x; x/ D 0,
d.x; y/ D 0 only if y D x,
d.x; y/ D d.y; x/, and
d.x; z/ d.x; y/ C d.y; z/ .
The pair .X; d/ is called a metric space.

The inequality d.x; z/ d.x; y/ C d.y; z/ in Definition 1.33 is called the triangle
inequality.
Less formally, one can refer to X itself as a metric space, rather than the pair
.X; d/, if it is understood that d is the underlying metric on X.
14
The Euclidean metric on R2 is derived from the Pythagorean theorem, and it

admits a formulation in higher dimensions as well.
Definition 1.34. The Euclidean metric on Rn is the function d2 W Rn Rn ! 0; 1/
defined by
v
uX
u n
d2 .x; y/ D t .xj yj /2 ;
8 x; y 2 Rn :
(1.1)
jD1
Proposition 1.35. The Euclidean metric on Rn is a metric in the sense of Definition 1.33.
Proof. The function d2 in (1.1) plainly satisfies the first three conditions of
Definition 1.33. All that remains is to prove the triangle inequality. To this end,
an intermediate inequality is required.
Suppose that u; v 2 Rn and u 6D 0. Let fu;v denote the quadratic polynomial
fu;v .t/ D
n
X
.tuj C wj /2 D t2
jD1
n
X
u2j C 2t
n
X
jD1
uj wj C
jD1
2
X
w2j :
jD1
Thus, fu;v .t/ 0 for all t 2 R and fu;v .t0 / D 0 if and only if to uj D wj for every
j D 1; : : : ; n. Therefore, because u 6D 0, there is at most one t0 2 R that satisfies
fu;v .t0 / D 0. On the other hand, the quadratic formula yields
t0 D
2
and so the discriminant .2

Hence,
j uj wj
q P
P
P
.2 j uj wj /2 4. j u2j /. j w2j /
P
;
2 j u2j
j uj wj /
n
X
jD1
4.
uj wj / .
2
j uj /.
n
X
jD1
u2j /.
2
j wj / is a nonpositive real number.
n
X
w2j /;
jD1
implying that
v
0
10
1
X
u
n
n
n
X
X
u
u
uj vj t@
u2j A @
vj2 A:
jD1
jD1
jD1
15
If x; y; z 2 Rn , then
d2 .x; z/2 D
n
n
X
X
.xi zi /2
.xi yi C yi zi /2
iD1
iD1
n
n
n
X
X
X
.xi yi /2 C 2
.xi yi /.yi zi / C
.yi zi /2
iD1
iD1
iD1
v
u n
n
n
n
X
X
X
uX
2

.xi yi / C 2t .xi yi /2
.yi zi /2 C
.yi zi /2
iD1
iD1
iD1
iD1
12
0v
v
uX
uX
n
u n
u
D @t .xi yi /2 C t .yi zi /2 A
iD1
iD1
D .d2 .x; y/ C d2 .y; z//2 :

Hence, the d2 is indeed a metric in the sense of Definition 1.33.
t
u
Corollary 1.36. The formula d.x; y/ D jx yj, for x; y 2 R, defines a metric on R.

Proof. The function d is simply the Euclidean metric d2 on Rn in the case where
n D 1.
t
u
Definition 1.37. The inequality
v
0
10
1
X
u
n
n
n
X
X
u
u
uj vj t@
u2j A @
vj2 A;
jD1
jD1
jD1
(1.2)
for real numbers u1 ; : : : ; un ; v1 ; : : : ; vn is called the Cauchy-Schwarz inequality.

The Euclidean metric is not the only metric of interest.
Proposition 1.38. The following functions d1 and d1 are metrics on Rn :
d1 .x; y/ D
n
X
jxj yj j
(1.3)
d1 .x; y/ D max jxj yj j:
(1.4)
jD1
1jn
t
u
The reason for labelling these metrics as d1 , d2 , and d1 will be apparent in

subsequent chapters, particularly in connection with Banach space theory.
16
Proposition 1.39. If d is a metric on a set X and if

Bd D fBr .x/ j x 2 X; r 2 R ; r > 0g ;
(1.5)
where Br .x/ D fy 2 X j d.x; y/ < rg, then B is a basis.

Proof. We are to verify that Bd has the following properties:
1. for each x 2 X there is a B 2 Bd such that x 2 B;
2. for every B1 ; B2 2 Bd and each x 2 B1 \ B2 , there is a set B3 2 Bd such that
x 2 B3 B1 \ B2 .
The first item is obvious, as x 2 Br .x/ for every x 2 X and every r > 0. Assume now
that B1 ; B2 2 Bd and x 2 B1 \ B2 . There are xj 2 X and rj 2 R such that Bj D Brj .xj /.
Thus, d.x; xj / < rj for each j, and so 0 < sj D rj d.x; xj /. Let r D minfs1 ; s2 g. Then
x 2 Br .x/ Br1 .x1 / \ Br2 .x2 /.
t
u
Definition 1.40. The topology Td on X induced by the basis Bd of Proposition 1.39
is called a metric topology.
Thus, a metric space .X; d/ is a topological space .X; Td /.
The proof of Proposition 1.28 is easily adapted to the case of metric spaces to
yield that following characterisation of open sets.
Proposition 1.41. In a metric space .X; d/, a subset U X is an open set if and
only if for each x 2 U there is an " > 0 such that B" .x/ U.
Corollary 1.42. The metric topology on R induced by the metric d.x; y/ D jx yj,
for x; y 2 R, coincides with the standard topology on R.
Proof. Propositions 1.28 and 1.41 indicate that the metric topology and the standard
topology have exactly the same open sets.
t
u
It is important to note that different metrics on a set X can induce the same
topology.
Example 1.43. The metrics d1 and d2 on Rn induce the same topology; that is,
Td1 D Td2 .
Proof. First notice that if 1 ; : : : ; n 2 R are nonnegative, then by a binomial
expansion we see that
.1 C C n /2 12 C : : : n2 :
Taking square roots yields
n
X
jD1
11=2
0
n
X
j @
j2 A ;
jD1
which implies that d2 .x; y/ d1 .x; y/ for all x; y 2 Rn .
1.5 Subspaces and Product Spaces
17
On the other hand, the Cauchy-Schwarz inequality (1.2) yields

n
X
j D
jD1
n
X
0
j .1/ @
jD1
n
X
jD1
11=2 0
11=2
11=2
0
n
n
X
X
p
j2 A @
12 A D n @
j2 A :
jD1
jD1
p
Hence, d1 .x; y/ nd2 .x; y/, for all x; y 2 Rn .
dj
Let Br .x/ denote a basic open set in Tdj . To show that Td1 D Td2 , it is enough, by
Proposition 1.24, to show that Bdr 1 .x/ 2 Td2 and Bdr 2 .x/ 2 Td1 for all x 2 X and r > 0.
d2
To this end, if y 2 Bdr 1 .x/, then d2 .x; y/ d1 .x; y/ < r implies that
py 2 Br .x/; thus,
p
d1
d2
d2
Br .x/ Br .x/ 2 Td2 . Similarly, if y 2 Br .x/, then d1 .x; y/ nd2 .x; y/ < nr
t
u
implies that y 2 Bdp1 nr .x/; that is, Bdr 2 .x/ Bdp1 nr .x/ 2 Td1 .
One can also alter a given metric d on a space X, without changing the topology,
in such a way that the distances between points measured by the new metric are no
greater than 1.
Proposition 1.44. If .X; d/ is a metric space, then the function db W X X ! R,
defined by
db .x; y/ D min fd.x; y/; 1g ;
(1.6)
is a metric X. Furthermore, the metric topologies Td and Tdb on X coincide.

t
u
The standard topology on R was first introduced here by specifying a certain

basis for the topology, and it was noted subsequently that the topology is actually
induced by a metric on R. This represents one example of a metrisable space.
Definition 1.45. A topological space .X; T / is metrisable if there exists a metric d
on X such that T D Td , where Td is the metric topology induced by the metric d.
There are several instances throughout this book where we will need to determine
whether a certain topological space is a metrisable space.

Definition 1.46. The subspace topology TY on a subset Y X in a topological
space .X; T / is the collection
TY D fU \ Y j U 2 T g
of subsets of Y.
18
It is straightforward to verify that TY satisfies the axioms for a topology. Hence,

.Y; TY / is indeed a topological space. A subset W Y is said to be open in Y if
W 2 TY . It need not be true that an open set of Y is an open set of X. For example,
with respect to the standard topology of R, the set W D fx 2 R j 0 x < 1=2g is open
in Y D fx 2 R j 0 x < 1g, but not in R.
Proposition 1.47. If Y is an open set in a topological space .X; T /, then W Y is
open in Y if and only if W is open in X.
Proof. If W 2 TY , then there is a U 2 T such that W D Y \ U. As Y 2 T also and
T is closed under the intersection of two of its elements, W 2 T . Conversely, if
W 2 T , then W D W \ Y implies that W 2 TY .
t
u
Proposition 1.48. If B is a basis for a topology T on X and if Y X, then the
collection BY defined by BY D fY \ B j B 2 Bg is a basis for the subspace topology
TY on Y.
t
u
The space R is a metric space via the Euclidean metric. However, R is also the
Cartesian product of n copies of R. Given that R itself is a topological space, one
expects that there is a way in which these identical copies of R induce a topology
on Rn . This is indeed the case, and the resulting topology is called the product
topology. Throughout this book we will make use of Cartesian products of both
finite and infinite numbers of spaces, and so we consider the most general case here.
Let be a set, and suppose that, for each 2 , .X ; T / is a topological space.
The Cartesian product of the family fX g2 is defined by
n
X D f.x /2 j x 2 X g ;
2
where .x /2 denotes the function !
X , whose value at is x . More
2
informally, we consider .x /2 Q

to be a tuple of elements, one fromQeach X . If
Y X for every 2 , then Y is considered as a subset of X in the
natural way.
Proposition 1.49. Assume that f.X ; T /g2 isQ
a family of topological spaces and
define the following subcollections of subsets of X :
Q
B D f U j U 2 T ; 8 2 g
Q
B D f U j U 2 T ; 8 2 ; and
U D X for all but at most a finite number of g:
Then B and B are bases of subsets of
X .
19
The proof of Proposition 1.49 is a direct application of the definition. If is a

finite set, then B D B ; however, this is no longer the case when is an infinite
set.
Definition 1.50. If f.X ; T /g2 is a family of topological spaces, then
Q

1. the box topology on X
by the basis B , and
Q is the topology T induced

2. the product topology on X is the topology T induced by the basis B .
Because B B , the box topology is finer than the product topology; but if
is a finite set, then these two topologies coincide.
The following technical result is another consequence of the definitions.
Proposition 1.51. Suppose that f.X ; T /g2 is a family of topological spaces,
and, for each 2 , that B is a basis for T , Y is a subspace of X , and TY; is
the subspace topology of Y .
Q
Q
1. If B D f B j B 2 B g, then B is a basis for the
Q box topology of X .
Moreover, the box topology
of the product space Y coincides
with the
Q
Q
subspace topology
of
Y
inherited
from
the
box
topology
of
X
.

Q
2. If B D f B j B 2 B and B Q
D X for all but finitely many g, then B is
a basis for the product
topology of X . Moreover, the productQtopology of the
Q
product space Y coincides
Q with the subspace topology of Y inherited
from the product topology of X .
Another way to induce a topology on a Cartesian product of topological spaces
is to take advantage of special features that the family might have, as in the case of
products of metric spaces.
Proposition 1.52. If f.X; d /g2 is a family of metric spaces, and if d;b denotes
the metric on X as given by equation (1.6) of Proposition 1.44, then the formula
d ..x / ; .y / / D sup .d;b .x ; y //
defines a metric on
(1.7)
X .
t
u
Definition 1.53. The metric d of Proposition 1.52 is called the uniform metric on
Q
X and the resulting topology Td is called the uniform metric topology.
Proposition 1.54. If
Q f.X; d /g2 is a family of metric spaces,
Qthen the uniform
metric topology on X is finer than the product topology on X .
Q
Q
Proof. Let U D U X be a basic open set in the product topology. Thus,
there is a finite set F D f1 ; : : : ; n g such that U D X for all 62 F. Choose
x D .x / 2 U. For each j D 1; : : : ; n there is a "j > 0 such that B"j .xj / Uj . Let
of these "j and consider the basic open set B"x .x/ in the uniform
"x be the minimum Q
metric topology of X . If y 2 B"x .x/, then y 2 X D U for 62 F; and, for each
j D 1; : : : ; n, dj ;b .xj ; yj / < "x "j , which implies that yj 2 Uj . Hence, B"x .x/ U.
20
That is,
UD
B"x .x/ 2 T :
x2U
Hence, the uniform metric topology Td is finer than the product topology T .
t
u
Definition 1.55. If f.X ; T /g2 is a family of topological spaces such that each
X D X and T D T for some topological space .X; T /, then the Cartesian product
of the family f.X ; T /g2 is denoted by X and the T and T denote the box
and product topologies on X , respectively. For D f1; : : : ; ng N, the notation X n
is used for X .
Of special interest are the product spaces Rn and RN .
Proposition 1.56. The metric topology on Rn induced by the Euclidean metric
coincides with the product topology of Rn .
Proof. Let Td2 and T denote the metric and product topologies on Rn , and let
basic open sets of Td2 be denoted by Br .x/.
Suppose that U Rn is an open set in the product topology of Rn , and let x D
.x1 ; : : : ; xn / 2 U. Thus, there is a basic open set Bx 2 T such that x 2 Bx Rn . Such
m
Y
a basic open set Bx has the form Bx D
.pxj ; qxj / for some pxj ; qxj 2 Q, j D 1; : : : ; n.
jD1
Now if rx D min fjxj pxj j; jxj qxj jg, then for every y D .y1 ; : : : ; yn / 2 Brx .x/ we
1jn
have that
0
jxi yi j @
n
X
11=2
jxj yj j2 A
jD1
< r D min fjxj pxj j; jxj qxj jg

1jn
for each i D 1; : : : ; n. Hence, yi 2 .pxi ; qxi / for every i, implying that y 2 Bx . Thus,
UD
Brx .x/ 2 Td2 :
x2U
Conversely, if U is an open set in the metric topology of Rn and if x D

.x1 ; : : : ; xn / 2 U, then there exists "x > 0 such that B"x .x/ U (Proposition 1.41). For
each j D 1; : : : ; n, choose pxj ; qxj 2 Q such that pxj 2 .xj p"xn ; xj / and qxj 2 .xj ; xj C p"xn /.
Q
Thus, if y D .y1 ; : : : ; yn / 2 j .pxj ; qxj /, then
11=2 0
1
0
2 1=2
n
n
X
X
"
x
A D "x :
d2 .x; y/ D @
jxj yj j2 A < @
p
n
jD1
jD1
21
Hence,
0
1
[
[ Y
x
x
@ .pj ; qj /A D
B"x .x/ 2 T :
UD
x2U
x2U
Thus, the product topology and metric topology of Rn coincide.
t
u
Proposition 1.56 asserts that the product topology on Rn is metrisable. Because

different metrics on a given space can yield the same topologies, it can be difficult to
tell from topological properties alone whether the topology on a given topological
space .X; T / is induced by some metric. Nevertheless, we have the following
information about RN in its product topology.
Proposition 1.57. The topological space .RN ; T / is metrisable.
Proof. By Propositions 1.28 and 1.44, the standard topology on R is induced by
the metric d.x; y/ D min.jx yj; 1/. Basic open sets in this metric topology will be
denoted by Bd" .x/.
Define W RN RN ! R by

jxn yn j 1
1
;
d.xn ; yn / ;
D sup
..xn /n ; .yn /n / D sup min
n
n
n
n
n
and note that is a metric on RN . Basic open sets in this metric topology will be

denoted by BQ
" .x/.
Let U D n Un be an open set in the product topology of RN . Thus, there is a
finite subset F D fn1 ; : : : ; nk g N such that n1 < < nk and Un D R for all n 62 F.
Select x D .xn /n 2 U and let "j > 0 be such that Bd"j .xnj / Unj , for j D 1; : : : ; k. Now
"

let "x be the minimum of njj , where j D 1; : : : ; k, and note that if y 2 B"x .x/, then

y 2 U. Hence, U is a union of sets of the form B"x .x/, for x 2 U, which implies that
U is open in the metric topology T of RN .

Conversely, suppose that Br .x/, where x D .xn /n , is a basic open set in the metric
N
topology T of R . Choose k 2 N such that 1k < 2r , and for each n D 1; : : : ; k define
Un D Bdr=2 .xn /, which is an open set in the standard topology of R. Let Un D R for
Q
every n > k. Consider Ux D n Un , which is a basic open set in the product topology
of RN and which contains x. If y D .yn /n 2 Ux , then d.xn ; yn / < 2r if 1 n k, and
d.xn ; yn / 1 for all n > k. Thus, .x; y/ D supn n1 d.xn ; yn / < 2r , which implies that

Ux Br .x/. Hence,
Br .x/ D
Ux
x2U

implies that Br .x/ is open in the product topology T of RN .
t
u
22
1.6 Closures, Interiors, and Limit Points

Definition 1.58. A subset F X in a topological space .X; T / is a closed set if
there is an open set U 2 T such that F D U c .
Because the complements of closed sets are open sets, ; and X are both open
and closed. Furthermore, arbitrary intersections of closed sets are closed, and finite
unions of closed sets are closed.
Example 1.59. In the standard topology of R, closed intervals are closed sets.
Proof.
with a < b, then a; b is the complement of the union of the open
[If a; b 2 R,[
sets .x; a/ and .b; y/, proving that a; b is a closed set.
t
u
x<a
b<y
The next class of closed sets is used extensively in the study of topological
spaces.
Example 1.60. In the Euclidean metric space .RnC1 ; d2 /, where n 1, the n-sphere
8
9
n
<
=
X

Sn D x 2 RnC1 j d2 .x; 0/ D 1 D x 2 RnC1 j
xj2 D 1
:
;
jD1
is closed.
Proof. The set
[
U D B1 .0/ [
Bd2 .x;0/1 .x/
x2RnC1 ; d2 .x;0/>1
is open in RnC1 , and Sn is its complement.
t
u
Proposition 1.61. If Y is a subspace of a topological space X, then F Y is closed

in Y if and only if there is a closed set C X in X such that F D Y \ C.
Proof. If F Y is closed in Y, then F c \ Y is open in Y, and thus F c \ Y D Y \ U for
some open set U in X. Let C D U c , which is closed in X, and note that F D Y \ C.
Conversely, if there is a closed set C X in X such that F D Y \ C, then U D Cc
is open in X, and so Y \ U is open in Y. Thus, F c \ Y D Y \ .Y \ U c / D Y \ U, which
is open in Y; hence, F is closed in Y.
t
u
Definition 1.62. Assume that Y X is a subset of a topological space X.
1. The closure of Y is the subset Y of X defined by
\
YD
CY
and C is closed
C:
1.6 Closures, Interiors, and Limit Points
23
2. The interior of Y is the subset int Y of X defined by

[
int Y D
U:
UY and U is open
3. If Y D X, then Y is said to be dense in X.
The closure operation has the following properties.
Proposition 1.63 (Closures of Unions and Intersections). Assume that Y and Z
are subsets of X, and that fY g2 is a collection of subsets of X. The following
statements hold:
1. if Y Z, then Y Z;
2. Y [ Z D Y [ Z;
[
[
3.
Y
Y , and equality need not be attained;
4.
Y
Y , and equality need not be attained;
5. if Y Z, then Z n Y Z n Y, and equality need not be attained.

Proof. For (1), the definition of closure yields A A for every subset A of X. Thus,
Y Z Z implies that Z is a closed set containing Y. Hence, Y Z, by definition.
To prove (2), note that Y [ Z Y [ Z, which is closed. Hence, Y [ Z Y [ Z,
by definition of closure. Conversely, Y .Y [ Z/ implies, by (1), that Y Y [ Z.
Likewise, Z Y [ Z, and so Y [ Z Y [ Z.
The proofs of the remaining assertions are left to the reader (Exercise 1.116). u
t
If one has a nested inclusion Z Y X, then it is possible to consider the closure
of Z relative to the topology of X or to the subspace topology of Y. The following
proposition indicates the relationship between these closures.
Proposition 1.64. Assume that Z Y X, and denote the closure in Y of a subset
A Y by AY and the closure in X of a subset B X by BX . Then Z Y D Y \ Z X .
Proof. By Proposition 1.61, Y \ Z X is closed in Y; as this set also contains Z, it must
contain the closure Z Y of Z in Y. Furthermore, Z Y has the form Y \ C, for some
closed set C in X, again by Proposition 1.61. Therefore, Z X C, as C is closed in X
and contains Z. Thus, Y \ Z X Y \ C D Z Y and, hence, Z Y D Y \ Z X .
t
u
The main proposition relating closures and interiors is as follows.
Proposition 1.65. If Y X is a subset of a topological space X, then
c
1. Y D int .Y c / and
2. .int Y/c D Y c .
t
u
To obtain a clearer understanding of what is added to a set Y when passing from

Y to its closure Y, the notion of (topological) limit point is introduced.
24
Definition 1.66. A neighbourhood of a point x in a topological space X is an open

subset U of X such that U contains x.
Proposition 1.67. The following statements are equivalent for a subset Y of a
topological space X:
1. x 2 Y;
2. U \ Y 6D ; for every neighbourhood U of x.
Proof. Assume that x 2 Y, and that U is a neighbourhood of x. If it were true that
U \ Y D ;, then U would be a subset of Y c and, thus, would be contained in the
c
c
interior of Y c . However, int .Y c / D Y (Proposition 1.65); therefore, x 2 U Y ,
which is in contradiction to x 2 Y.
Conversely, assume that x 2 X, and that U \ Y 6D ; for every neighbourhood U of
c
x. Also assume, contrary to what we aim to prove, that x 62 Y. Thus, x 2 Y D int .Y c /,
c
which implies that int .Y / is a neighbourhood of x. Therefore, by hypothesis,
int .Y c / \ Y 6D ;. This is, however, a contradiction of the fact that int .Y c / \ Y
Y c \ Y D ;. Hence, it must be that x 2 Y.
t
u
The passage from Y to Y is a matter of adding the limit points of Y.
Definition 1.68. An element x 2 X is a limit point of a subset Y X if, for every
neighbourhood U of x, there is an element y 2 Y such that y 2 U and y 6D x. The set
of limit points of Y is denoted by L.Y/.
Proposition 1.69. Y D Y [ L.Y/.
Proof. If x 2 L.Y/, then U \Y 6D ; for every neighbourhood U of x. Thus, L.Y/ Y,
and therefore Y [ L.Y/ Y.
Conversely, suppose that x 2 Y and x 62 Y. By virtue of x 2 Y, U \ Y 6D ; for
every neighbourhood U of x. Because x 62 Y, for each neighbourhood U there must
be some y 2 Y with y 2 U. Hence, y 2 L.Y/, which proves that Y Y [ L.Y/.
t
u
The topological boundary is another important closed set associated with an
arbitrary set.
Definition 1.70. The boundary of a subset Y X in a topological space X is the set
@ Y defined by
@ Y D Y \ Y c:
Proposition 1.71. If Y X, then:
1.
2.
3.
4.
5.
@ Y D Y n int Y;
Y D int Y [ @ Y;
int Y D Y n @ Y;
Y is closed if and only if @ Y Y; and
Y is open if and only if Y \ @ Y D ;.
t
u
1.7 Continuous Functions
25
1.7 Continuous Functions

The interesting functions between topological spaces are the continuous ones. (We
shall use the terms function and map interchangeably.) To define what is meant
by a continuous function, we first recall some notation.
If X and Y are sets, if U X and V Y are subsets, and if f W X ! Y is a function,
then f .U/ and f 1 .V/ denote the subsets of Y and X, respectively, defined by
f .U/ D ff .x/ j x 2 Ug
f 1 .V/ D fx 2 X j f .x/ 2 Vg:

Observe that f f 1 .V/ V and U f 1 .f .U//.
Definition 1.72. If X and Y are topological spaces, then a function f W X ! Y is
continuous if f 1 .V/ is an open set in X for every open subset V of Y.
The definition of continuity is global, but it is convenient to be able to discuss
continuity as a local property.
Definition 1.73. A function f W X ! Y between topological spaces X and Y is
continuous at a point x 2 X if, for every neighbourhood V Y of f .x/, there is a
neighbourhood U X of x such that f .U/ V.
Global continuity is the same as local continuity at each point of the space.
Proposition 1.74. The following statements are equivalent for a function f W X ! Y
between topological spaces X and Y:
1. f is continuous;
2. f is continuous at every point x 2 X.
Proof. Suppose that f is continuous, that x 2 X, and that V Y is any neighbourhood
1
of f .x/. The continuity
1 of f implies that U D f .V/ is an open neighbourhood of x.
Further, f .U/ D f f .V/ V, and so f is continuous at x.
Conversely, suppose that f is continuous at each x 2 X. Choose any open set
V Y, and consider U D f 1 .V/. For each x 2 U, the set V is a neighbourhood
of x. Because f is continuous at each x, there is a neighbourhood Ux of x such that
f .Ux / V. Thus, Ux f 1 .V/ D U, and so
UD
[
x2U
fxg
Ux U:
x2U
Hence, U is the union of a family of open sets and is, therefore, open. This proves
that f is a continuous function.
t
u
In the context of metric spaces, the continuity of a function is given by the
following familiar criterion:
26
Proposition 1.75. If X and Y are metric spaces, with metrics dX and dY , respectively, then a function f W X ! Y is continuous if and only if for each x0 2 X and for
every " > 0 there is a > 0 such that
dY . f .x0 /; f .x/ / < " for all x 2 X that satisfy dX .x; x0 / < :
t
u
For each fixed y 2 X in a metric space .X; d/, one can ask whether the function
dy W X ! R defined by dy .x/ D d.x; y/, for x 2 X, is continuous. In other words, is the
metric d continuous in each of its variables? This is indeed true, but a more general
result, Proposition 1.77 below, will be necessary for our study of operator theory.
Definition 1.76. If S is a nonempty subset of a metric space .X; d/ and if x 2 X,
then the distance from x to S is the real number denoted by dist.x; S/ and defined by
dist.x; S/ D inffd.x; s/ j s 2 Sg:
(1.8)
Proposition 1.77. If S is a nonempty subset of a metric space .X; d/, then the
function dS W X ! R defined by dS .x/ D dist.x; S/ is continuous.
Proof. By definition, if x 2 X, then dist.x; S/ d.x; s/ for every s 2 S. Thus, by the
triangle inequality, dS .x/ d.x; s/ d.x; y/ C d.y; s/ for all s 2 S and y 2 X. That is,
by varying s through S,
dist.x; S/ d.x; y/ inf d.y; S/ D dist.y; S/:
s2S
Hence, dS .x/ dS .y/ d.x; y/. By interchanging the roles of x and y we obtain
dS .y/dS .x/ d.x; y/, and so jdS .x/dS .y/j d.x; y/ for all x; y 2 X. An application
of Proposition 1.75 now yields the continuity of dS .
t
u
Another useful criterion for continuity is given by the following proposition,
which involves closed sets and closures of sets.
Proposition 1.78. The following statements are equivalent for a map f W X ! Y of
topological spaces X and Y:
1. f is continuous;
2. f .A/ f .A/ for every subset A X; and
3. f 1 .C/ is closed in X for every closed set C in Y.
Proof. (1) ) (2). Assume that f is continuous and that A X. If x 2 A and if V is
a neighbourhood of f .x/, then f 1 .V/ is a neighbourhood of x, and so A \ f 1 .V/
is nonempty. Now, if z 2 A \ f 1 .V/, then f .z/ 2 V \ f .A/, which is to say that the
neighbourhood V of f .x/ has nonempty intersection with f .A/. Hence, f .x/ 2 f .A/.
(2) ) (3). Suppose that f .A/ f .A/ for every subset A X. Select a closed set
C Y and let A D f 1 .C/. Thus, f .A/ C and,
f .A/ C. By hypothesis,
therefore,

f .A/ f .A/ and so f .A/ C. Hence, A f 1 f .A/ f 1 .C/ D A A implies that
A D A. That is, f 1 .C/ is closed.
1.8 The Cantor Ternary Set and Ternary Function
27
(3) ) (1). If V Y is open, then f 1 .V c / is closed, by hypothesis. Because

f .V c / D f 1 .V/c , we deduce that f 1 .V/ is open, and so f is continuous.
t
u
1
Continuous maps possess the following basic properties, each of which is readily
verified.
Proposition 1.79. The following functions are continuous:
1. every constant map, which is to say that if X and Y are topological spaces, and if
y0 2 Y, then the function f W X ! Y given by f .x/ D y0 , for all x 2 X, is continuous;
2. the composition of continuous functions;
3. restrictions of continuous functions to subspaces; and
4. the inclusion map Y W Y ! X, Y .x/ D x for all x 2 Y, for every subspace Y X.

The middle third of the interval 0; 1 is the open interval .1=3; 2=3/. If one
removes the middle third from 0; 1, then the closed set
C1 D 0; 1=3 [ 2=3; 1
remains. Note that C1 is a union of two closed subintervals, each of length 1=3.
The middle thirds of C1 are the open intervals .1=9; 2=9/, which is the middle
third of 0; 1=3, and .7=9; 8=9/, which is the middle third of 2=3; 1. If these middle
thirds are removed from C1 , then one is left with
C2 D 0; 1=9 [ 2=9; 1=3 [ 2=3; 7=9 [ 8=9; 1 ;
which is a union of four subintervals, each of length 1=.32 /.
Proceed by induction. Once Cn1 has been constructed as a union of 2n1 closed
subintervals Fk , remove the middle third from each Fk to obtain Cn , a union of 2n
closed subintervals, each of length 1=.3n /.
Definition 1.80. The Cantor ternary set is the subset C 0; 1 defined by
\
CD
Cn :
n2N
Proposition 1.81. The Cantor set C is a nonempty closed set with no interior.
Proof. Each Cn is closed, and therefore so is C . Moreover, C contains the endpoints
of each subinterval in each Cn ; thus, C is nonempty.
Suppose that U is an open subset of C . Then U contains an open interval J, and
J is a subset of each Cn . For fixed n, J must lie in one of the closed subintervals
that form Cn , and such intervals have length .1=3/n . Thus, the length of J is at most
.1=3/n . But this length of J holds for all n, which implies that J is length zero; that
is, J D ;. Hence, U D ;.
t
u
28
Another way to characterise C is by considering the ternary expansion of each

2 0; 1. Recall that every real number admits a decimal representation. A very
similar argument show that every 2 0; 1 can be expressed in ternary form as
D
1
X
ck
;
3k
kD1
where ck 2 f0; 1; 2g
8 k 2 N:
(1.9)
The representation in (1.9) is not unique, because the real number 1=3, for
example, can be expressed in two ways: (i) c1 D 1 and ck D 0 for all k 2; and
(ii) c1 D 0 and ck D 2 for all k 2. This is the only kind of ambiguity that can arise,
and this ambiguity does not have any bearing on the description of C that follows.
It is also convenient to adopt base-3 notation for 2 0; 1:
D .0:c1 c2 c3 /3
means
D
1
X
ck
:
3k
kD1
In this notation, for example,

1
3
D .0:10000 /3
1
3
D .0:02222 /3
2
9
D .0:02000 /3
2
9
D .0:01222 /3
7
9
D .0:21000 /3
8
9
D .0:22000 /3
and so forth.
Proposition 1.82. 2 C if and only if D .0:c1 c2 c3 /3 , where ck 2 f0; 2g for
every k 2 N.
Proof. The left and right end points of any one of the closed intervals that make up
Cn will have ternary form
.0:g1 g2 g3 gn1 1000 /3
and
.0:g1 g2 g3 gn1 2000 /3
respectively, where gj 2 f0; 2g for all 1 j .n 1/. Thus,

D .0:c1 c2 c3 /3
is in this closed interval if and only if gj D cj for all 1 j .n 1/.
t
u
Proposition 1.83. C and R have the same cardinality.

Proof. On the one hand, the map that sends a binary sequence b D .b1 b2 : : : / to the
1
X
2bk
is an injection of the set B of all binary sequences into the
real number
3k
kD1
29
Cantor set C . On the other hand, the cardinality of B is that of the power set P.N/
of N, which in turn has the cardinality c of the continuum R. Hence, there is an
injection of R into the Cantor set C , and so C and R are in bijective correspondence
by the Schroeder-Bernstein Theorem (Theorem 1.3).
t
u
Another application of Proposition 1.82 leads to a very interesting continuous
function on 0; 1 that has quite remarkable features. Let W C ! 0; 1 be the
function that sends the ternary form of 2 C to a number in binary form: namely,
. .0:c1 c2 /3 / D
1
X
ck =2
kD1
2k
(1.10)
Definition 1.84. The function W 0; 1 ! 0; 1 defined by

.x/ D supf .y/ j y 2 C and y xg;
for x 2 0; 1, and where W C ! 0; 1 is the function defined by equation (1.10), is
called the Cantor ternary function.
Definition 1.85. If J R is an interval, then a function f W J ! R is monotone
increasing if f .x1 / f .x2 /, for all x1 ; x2 2 J such that x1 x2 . If f .x1 / < f .x2 /, for
all x1 ; x2 2 J such that x1 < x2 , then f is strictly monotone increasing.
Proposition 1.86. The Cantor ternary function is a monotone increasing continuous function and maps the Cantor set C onto 0; 1.
Proof. The map W C ! 0; 1 is clearly surjective and monotone increasing;
therefore, the same is true of W 0; 1 ! 0; 1.
To verify that is continuous, select y0 2 0; 1 and let " > 0. If y0 62 C , then
y0 62 Cn for some n 2 N. Thus, y0 is contained in one of the open intervals that have
been removed from Cn1 ; denote such an interval by .a; b/. The endpoints a and b
are the left and right endpoints of a closed subinterval of Cn , and so .x/ D .a/ for
all x 2 .a; b/. This shows that is constant on .a; b/, implying that is continuous
at the point y0 2 .a; b/.
However, if y0 2 C , then choose m 2 N such that .1=2/m < ". Let D .1=3/m .
Suppose that x 2 0; 1 is such that jx y0 j < . Without loss of generality, we may
suppose that x < y0 ; thus,
y0 x D
1
X
!k
;
3k
kDmC1
for some !k 2 f0; 1; 2g:
That is, in any ternary expansion of x, the first m ternary digits of x coincide with
those of y0 . Let 1 ; : : : ; m 2 f0; 2g be such that y0 D .0:1 ; m /3 and consider
z0 D .0:1 ; m 00 /3 2 C . Therefore,
z0 x y0 ) .z0 / .x/ .y0 /:
30
The inequality
.y0 / .z0 / D .0:
m
m
1
1
;
/2 .0: ; 00 /2 < .1=2/m D "
2
2
2
2
implies j.y0 / .x/j < ", proving the continuity of at y0 .
t
u
The proof of Proposition 1.86 shows that the Cantor ternary function is constant
on open intervals in the complement of C . We shall have need of this fact later, and
so this feature is recorded below for future reference.
Corollary 1.87. If J is any open interval in 0; 1nC , then is constant on J.
1.9 Weak Topologies and Continuous Maps of Product

Spaces
Traditionally one begins with topological spaces and considers continuous maps
between them. In many instances, especially in functional analysis, it is beneficial
to reverse this process to allow functions to determine the topology.
Proposition 1.88. If f.Y ; T /g2 is a family of topological spaces, X is a set, and
g W X ! Y is a function, for each 2 , then
1. there is a coarsest topology on X in which each function g W X ! Y is
continuous, and
2. for every topological space Z, a map f W Z ! X is continuous if and only if g f W
Z ! Y is continuous for all .
[
Proof. For each 2 let S D fg1
S . The
.U / j U 2 T g and let S D
collection S is plainly a subbasis, and thus induces a topology on X which

we denote by Twk . Observe that each g W .X; Twk / ! .Y ; T / is continuous.
Furthermore, if T 0 is any other topology on X in which every g is continuous, then
0
T 0 must contain every set of the form g1
.U /, where U 2 T . Hence, Twk T .
Next, fix a topological space Z and assume that f W Z ! X is a function. If
f is continuous, then so is every g f W Z ! Y , because the composition of
continuous maps is continuous. Conversely, if it is assumed now that g f W Z ! Y
is continuous

for all , then for every and every open set U Y , the set
f 1 g1 .U / is open in Z. Thus, f 1 .S/ is open in Z for every S in the subbasis
S , and so f 1 .B/ is open in Z for every B in the basis B for Twk generated by S .
By Exercise 1.126, this implies that f is continuous.
t
u
A result related to Proposition 1.88 is as follows.
Proposition 1.89. If f.Y ; T /g2 is a family of topological spaces, Y is a set, and
g W X ! Y is a function, for each 2 , then
1.9 Weak Topologies and Continuous Maps of Product Spaces
31
1. there is a finest topology on Y in which each function g W X ! Y is continuous,

and
2. for every topological space Z, a map f W Y ! Z is continuous if and only if f g W
X ! Z is continuous for all .
t
u
Definition 1.90. Each of the topologies introduced in Propositions 1.88 and 1.89 is
called the weak topology induced by the family of functions fg g2 .
Turning now to product spaces, we begin with a definition.
Definition 1.91. AssumeQ
that f.X ; T /g2 is a family of topological spaces. For
each 2 , the map p W 2 X ! X defined by
p ..x / / D x ;
is called a projection map.
Q
Observe that p maps each x 2 2 X onto the -th
Q coordinate of x.
Furthermore, if U is an open set in X , then p1
.U
/
D
A , where A D X
for all , and A D U . Hence, the

maps
p
are
continuous
with respect to both
Q
the product and the box topology of 2 X .
Q
Q
Proposition 1.92. In the product topology of 2 X , a map f W Z ! 2 X is
continuous if and only if p f W Z ! X is continuous for every 2 .
Q
Proof. If f W Z ! 2 X is continuous, then so is p f because the composition
of continuous maps is continuous.
Conversely, suppose that p f W Z ! X is continuous for every 2 . Fix
2 and suppose that U X is open. Let W D p1
.U /, which is open in
Q
1
1
2
Q X ; thus, .p f / .U / D f .W / is open in Z. Now if B is a basic open set
in 2 X , then there are 1 ; : : : ; n 2 and open sets Uj Xj , for j D 1; : : : ; n,
n
n
\
\
1
such that B D
Wj , where Wj D p1
.U
/.
Hence,
f
.B/
D
f 1 .Wj / is open
j
j
jD1
jD1
in Z. By Exercise 1.126, this implies that f is continuous.
t
u
One implication in Proposition 1.92 fails for the box topology.

Example 1.93. If RN has the box topology, and if pn W RN ! R denotes the
projection onto the n-th coordinate, then there exists a function f W R ! RN such
that pn f W R ! R is continuous for all n 2 N, yet f itself is not continuous.
Proof. Let f W R ! RN be defined by f .x/ D .x; x; x; : : : /. Thus,
Q pn f1.x/ D x,which
is plainly continuous. However, if V RN is the open set n . 1
; / 2 T , then
n n
1
f 1 .V/ D fx 2 R j x 2 . 1
;
/
8
n
2
Ng
D
f0g,
which
is
not
open.
Thus, f is not
n n
continuous.
t
u
32
1.10 Quotient Spaces

Suppose that
is an equivalence relation on a set X. For each x 2 X, the equivalence
class of x is denoted by xP ; thus,
xP D fy 2 X j y
xg:
If X=
denotes the set fPx j x 2 Xg of equivalence classes of X, and if .X; T / is
a topological space, then it is possible to endow X=
with the structure of a
topological space.
is an equivalence relation on a topological space .X; T /,
and if q W X ! X=
is the canonical quotient map q.x/ D xP , then
Tquo D fV .X=
/ j q1 .V/ 2 T g
is a topology on X=
and the function q W X ! X=
is continuous.
The proof of Proposition 1.94 is a matter of verifying the definitions.
Definition 1.95. The topology Tquo in Proposition 1.94 is called the quotient
topology on X=
.
Two common quotient structures in algebra arise from subspaces of vector spaces
and normal subgroups of groups. First, if L V is a linear subspace of a vector space
V, then the relation v
w if and only if v w 2 L is an equivalence relation on V
and the space of equivalence classes, which is denoted by V=L, has the structure
P and vP C w
P w/. Thus, if it
of a vector space under the operations vP D .v/
P D .v C
assumed that R has the standard topology, and if L is a subspace of Rn , then the
vector space Rn =L is, as a topological space, a quotient space.
A second familiar quotient structure occurs in group theory. If H is a normal
subgroup of a group G with binary operation , then the relation a
b if and only
if b1 a 2 H is an equivalence relation on G and the space of equivalence classes,
which is denoted by G=H, has the structure of a group under the binary operation P
given by aP P bP D a P b. Of particular interest is the additive abelian group .R; C/ and
its (normal) subgroup .Z; C/, which results in the quotient group R=Z.
Example 1.96. The function f W R=Z ! R2 , defined by f .Pt/ D .cos 2
t; sin 2
t/, for
t 2 R, is continuous.
Proof. Let F W R ! R2 be the function F.t/ D .cos 2
t; sin 2
t/. If p1 and p2 denote
the projections onto the first and second coordinates, respectively, then pj F is a
trigonometric function R ! R, and is therefore continuous. By Proposition 1.92,
the continuity of each pj F implies the continuity of F. Hence, F 1 .V/ is open in
R for every open subset V R2 .
Observe that F.t/ D f .Pt/ for every t 2 R; hence, if V is open in R2 , then

q1 f 1 .V/ D q1 .fPt j f .Pt/ 2 Vg/ D ft j F.t/ 2 Vg D F 1 .V/;
1.11 Topological Equivalence
33
which, by definition of the quotient topology, implies that f 1 .V/ is open in R=Z.
Hence, f is a continuous map.
t
u
A general method for analysing the continuity of functions defined on quotient
spaces is given by the next result.
Proposition 1.97. If X and Y are topological spaces, and if
is an equivalence
relation on the space X, then the following statements are equivalent for a function
g W .X=
/ ! Y:
1. g is continuous, considered as a map .X=
/ ! Y;
2. g q is continuous, considered as a map X ! Y.
Proof. By definition of quotient topology, the canonical projection q W X ! X=
is
continuous. Hence, if g is continuous, then so is g q.
Conversely, assume that g q is continuous. Select an open set V in Y; thus,
U D .g q/1 .V/ is open in X. Because

U D .g q/1 .V/ D q1 .fPx j g.Px/ 2 Vg/ D q1 g1 .V/ ;
the set g1 .V/ is open in X=
, by definition of the quotient topology.
t
u
1.11 Topological Equivalence

Definition 1.98. Assume that X and Y are topological spaces.
1. A bijective function f W X ! Y in which both f and f 1 are continuous is called a
homeomorphism.
2. The topological spaces X and Y are said to be homeomorphic if there exists a
homeomorphism f W X ! Y (or g W Y ! X).
The notation X ' Y is used to denote that X and Y are homeomorphic spaces.
In some sense, the goal of topology is to identify topological spaces up to
homeomorphism. This objective, however, is hugely unrealistic, and therefore
topologists have introduced other invariants of topological spaces that are less
stringent than that of topological equivalence. Even a basic question such as Are
the topological spaces Rn and Rm homeomorphic if n m? requires sophisticated
tools to resolve. (The answer to the question is no.)
Example 1.99. If a; b 2 R are such that a < b, then .a; b/ ' R.
Proof. First note that g W .
2 ;
2 / ! R, where g. / D tan. /, is a continuous
bijection with continuous inverse g1 .s/ D tan1 .s/. Hence, .
2 ;
2 / ' R. Now
consider the straight line L in R2 that passes through the points .a; b/ 2 R2 and
.
2 ;
2 / 2 R2 , and let F W R ! R be the equation of this linethat is, F is
the function whose graph is L. The function F is a homeomorphism; therefore,
34
the restriction f of F to the interval .a; b/ is a homeomorphism of .a; b/ and

.
2 ;
2 /. Hence, .a; b/ ' .
2 ;
2 / and, by transitivity of topological equivalence
(Exercise 1.130), we deduce that .a; b/ ' R.
t
u
Recall from Example 1.60 that the 1-sphere is the unit circle in R2 , which
we assume to have the subspace topology. The following example shows that the
quotient space R=Z is topologically equivalent to S1 .
Example 1.100. R=Z ' S1 .
Proof. Example 1.96 already shows that the map f W R=Z ! S1 defined by f .Pt/ D
.cos 2
t; sin 2
t/, for t 2 R, is continuous. As f is plainly bijective, all that remains
is to show that f 1 is continuous. As in Example 1.96, let F W R ! R2 denote the
function F.t/ D .cos 2
t; sin 2
t/.
1

Now, if U 2 R=Z is open, then f 1
.U/ D f .U/ D F q1 .U/ ; thus, the aim
is to prove that F.W/ is open in R2 , where W R is the open set W D q1 .U/.
Select t0 2 W and let " > 0 be such that V D .t0 "; t0 C "/ W and " < 1=2. By
Exercise 1.112, d2 .F.t/; F.t0 // D 2j sin.
.t t0 //j. Therefore, if t 2 V, then
jt
t0 j <
<
=2, and the fact that the sine function is strictly monotone increasing on
.0; 1=2/, leads to d2 .F.t/; F.t0 // < sin.
"/. Hence, F maps V into the basic open set
Bsin.
"/ .F.t0 //. Conversely, if .s; r/ 2 Bsin.
"/ .F.t0 //, then
" <
=2 implies that the
Euclidean distance between .s; r/ and F.t0 / is less than sin.
=2/ D 1. Hence, there is
semicircular arc in S1 that contains both .s; r/ and F.t0 /, and so there is a t 2 R with
jt t0 j 12 and F.t/ D .s; r/. Therefore, 2 sin
" > .2j sin.
.t t0 //j D 2 sin.
jt
t0 j/ implies, by the monotonicity of the sine function on the interval 0;
=2, that
jt t0 j <
", and so t 2 V. This proves that F maps V onto S1 \ Bsin.
"/ .F.t0 //.
Carrying this procedure out for every t0 2 W shows that F.W/ is open in S1 , which
completes the proof of the continuity of f 1 .
t
u
Problems
1.101. Let X be the set of rational numbers q for which 0 < q < 1 and be the set
of irrational numbers such that 0 < < 1. For each 2 , let X be the set of all
sequences of elements in X with limit . Prove the following assertions.
1. Each X is an infinite set.
2. X \ X 0 is a finite set, for every pair of distinct irrationals ; 0 2 .
1.102. Let T be the collection of all subsets U N with the property that a natural
number n belongs to U only if every divisor k 2 N of n belongs to U.
1. Prove that T is a topology on N.
2. Determine whether T coincides with the discrete topology on N.
Problems
35
1.103. Prove that if T is a topology on X, and if B is the collection of subsets with

the property that, for every U 2 T and each x 2 U, there is a B 2 B with x 2 B U,
then B is a basis of subsets and T is the topology generated by B.
1.104. Prove that if T is a topology on X and if B is a basis for the topology T ,
then the following statements are equivalent for a subset U X:
1. U 2 T ;
[
2. there is a family fB g of subsets B 2 B such that U D
B .
1.105. Prove that B D f.p; q/ R j p; q 2 Q; p < qg is a basis of subsets of R.

1.106. Consider Z, the set of integers. Fix a prime number p. For every natural
k 2 N and integer a 2 Z, let

Bk;a D a C bpk j b 2 Z :
Show that the collection B D fBk;a g.k;a/2NZ is a basis for a topology on Z.
1.107. Let
B D fa; b/ j a; b 2 R; a bg :
1. Prove that B is a basis.
2. The topology on R induced by this basis is called the lower-limit topology. Prove
that the lower-limit topology on R is strictly finer than the standard topology
on R.
1.108. Prove that if S is a subbasis of subsets of X, then
8
9
n
<\
=
BD
j n 2 N; S1 ; : : : ; Sn 2 S
:
;
jD1
is a basis of subsets of X.
1.109. Let .X; d/ be a metric space. Prove that a subset U X is an open set if and
only if, for each x 2 U, there is an " > 0 such that B" .x/ U.
v
0
10
1
n
u
n
n
X
X
X
u
u
1.110. Show that if
uj vj D t@
u2j A @
vj2 A, for real numbers u1 ; : : : ; un
jD1
jD1
jD1
and v1 ; : : : ; vn , then there is a 2 R such that vj D uj for every j D 1; : : : ; n.
36
1.111. Consider the functions Rn Rn ! R defined by

P
d1 .x; y/ D njD1 jxj yj j
d1 .x; y/ D max1jn jxj yj j;
for x; y 2 Rn .
1. Prove that d1 and d1 are metrics on Rn .
2. Prove that d1 and d1 determine the same topology on Rn
1.112. If t; t0 2 R, and if ut D .cos 2
t; sin 2
t/ and ut0 D .cos 2
t0 ; sin 2
t0 /, then
show that d2 .ut ; ut0 / D 2j sin.
.t t0 //j in the Euclidean metric space .R2 ; d2 /.
1.113. Assume that .X; d/ is a metric space.
1. Prove that the formula db .x; y/ D min fd.x; y/; 1g, for x; y 2 X, defines a metric
on X.
2. Prove that the metric topologies Td and Tdb on X coincide.
1.114. If f.X; d /g2 is a family of metric spaces, and if d;b denotes the metric
on X as given by equation (1.6) of Proposition 1.44, then prove that the formula
d ..x / ; .y / / D sup .d;b .x ; y //
defines a metric on
X .
1.115. Prove that if B is a basis for a topology T on X and if Y X, then the

collection BY defined by BY D fY \ B j B 2 Bg is a basis for the subspace topology
TY on Y.
1.116. Assume that fY g2 is a collection of subsets of a set X.
1. Prove that the following statements hold:
[
[
a.
Y
Y ;
b.
Y
Y;
c. if Y Z, then Z n Y Z n Y.
2. For each of the statements above, find an example to show that equality in the
inclusion is not achieved.
c
1.117. If Y X is a subset of a topological space X, then prove that Y D int .Y c /

and .int Y/c D Y c .
1.118. Let A X and B Y, and prove the following assertions for the product
topology on X Y:
1. A B D A B;
2. int.A B/ D int.A/ int.B/.
Problems
37
1.119. Suppose that X is a topological space and that U X. Prove that the
following statements are equivalent:
1. U is an open set in X;
2. for every subset Y X, U \ Y D U \ Y.
1.120. Assume that R has the standard topology and consider the following subsets:

AD

1 1
C j m; n 2 N
m n
and

1
B D f0g [
jn 2 N :
n
1. Prove that B D L.A/, the set of limit points of A.

2. Determine L.B/.
1.121. Let Y R2 be the set
Y D .x; sin.1=x// 2 R2 j 0 < x 1 :

Determine the closure of Y in the standard topology of R2 .
1.122. If Y X, then prove that int Y D Y n L.Y c /, where L.Y c / is the set of limit
points of Y c .
1.123. Prove the following statements for a subset Y of a topological space X.
1.
2.
3.
4.
5.
@ Y D Y n int Y.
Y D int Y [ @ Y.
int Y D Y n @ Y.
Y is closed if and only if @ Y Y.
Y is open if and only if Y \ @ Y D ;.
1.124. Determine the closure and boundary of Br .x/ D fy 2 X j d.x; y/ < rg in a

metric space .X; d/.
1.125. Assume that X and Y are metric spaces and denote their metrics by dX and
dY , respectively. Prove that a function f W X ! Y is continuous at a point x0 2 X if
and only if for every " > 0 there is a > 0 such that
dY . f .x0 /; f .x/ / < " for all x 2 X that satisfy dX .x; x0 / < :
1.126. Let X and Y be topological spaces and let BY be a basis for the topology on
Y. Prove that a map f W X ! Y is continuous if and only if f 1 .B/ is open in X for
every B 2 BY .
1.127. Assume that A and B are closed subsets of a topological space X and that
A [ B D X. Prove that if g W A ! Y and h W B ! Y are continuous maps for which
g.x/ D h.x/ for every x 2 A \ B, then the map f W X ! Y in which f .x/ D g.x/ for
x 2 A and f .x/ D h.x/ for x 2 B is continuous.
38
1.128. Assume that fX g2 is a family of topological spaces and that Y is a set.
Suppose that g W X ! Y is a function, for each 2 . Prove that
1. there is a finest topology on Y in which each function g W X ! Y is continuous,
and
2. for every topological space Z, a map f W Y ! Z is continuous if and only if f g W
X ! Z is continuous for all .
1.129. Prove that a quotient space X=
is compact, if X is compact.
1.130. Prove that if X ' Y and Y ' Z, then X ' Z.
1.131. Consider Rn as a metric space with respect to the Euclidean metric, and
suppose that x; y 2 Rn and that r; s 2 R are positive. Prove that Br .x/ ' Bs .y/.
1.132. Suppose that 1 k < n and that L is a k-dimensional vector space of Rn .
Prove that Rn =L ' Rnk .
1.133. Assume that f0; 1gN has the product topology and consider the Cantor
ternary set C .
1. Prove that the topological spaces f0; 1gN and C are homeomorphic.
2. Prove that if C N has the product topology, then C and C N are homeomorphic.
3. Assuming that 0; 1N has the product topology, prove that there exists a
continuous surjection f W C ! 0; 1N .
Chapter 2
Topological Spaces with Special Properties
Generic features of topological spaces and their continuous maps were considered
in the previous chapter. This chapter investigates certain qualitative features of
topological spaces: compactness (how small is a space?), normality (how separated
can disjoint closed sets in a space be?), second countability (what is the smallest
cardinality of the basis for the topology?), and connectedness (how disperse or
disjoint is a space?). This chapter will also introduce the notion of a net, which
is a natural extension of the concept of a sequence, and show how properties of nets
capture some of the topological features of prime interest in abstract analysis.
2.1 Compact Spaces

Definition 2.1. Suppose that fU g2 is a family of open subsets of a topological
space X and that Y X is a subset.
[
U Y.
1. The family fU g2 is an open cover of Y if
2. If fU g2 is an open cover of Y and if , then the family fU! g!2 is a

subcover of fU g2 if fU! g!2 is also cover of Y.
Definition 2.2. A subset K of a topological space X is a compact set if every open
cover of K admits a finite subcover.
Thus, K X is a compact subset if, for every family fU g2 of open sets U X
n
[
[
U K, there is a finite set F D f1 ; : : : ; n g such that
Uj K.
for which

DOI 10.1007/978-3-319-45633-1_2
jD1
39
40
2 Topological Spaces with Special Properties
The compactness of a space K indicates that K is, in some sense, rather small.
For example, the space R, which may be viewed as a line of infinite length, is not
compact because the open cover fB" .q/gq2Q of R, for some fixed " > 0, does not
admit a finite subcover.
Example 2.3. Every closed interval a; b is a compact subset of R.
Proof. Suppose that fU g2 is an open cover of a; b in R, and let
(
KD
x 2 .a; b j there is a finite subset F such that a; x
)
U :
2F
Choose 0 2 such that a 2 U0 . Hence, there is a "0 > 0 such that a; aC"0 / U0 ,
and therefore x 2 K for every x 2 .a; a C "0 /. Because K is nonempty and is bounded
above by b 2 R, the supremum of K, c D sup K, exists.
Select 1 2 such that c 2 U1 . Because U1 is open, there is a "1 > 0 such that
.c "1 ; c U1 . And because c is the least upper bound for K, there exists z 2 K
such that
[z 2 .c "1 ; c/. Because z 2 K, there is a finite subset F for which
a; z
U . Therefore,
2F
a; c D a; z
.c "1 ; c
[
2F
!
U
U1 D
U ;
2F 0
where F 0 D F [ f1 g, implying that c 2 K.

Now if it were true that c b, then we would have .c; c C "/ U1 for some
" > 0. By selecting
[ any w 2 .c; c C "/ we would obtain c; w U1 , and so a; w D
U , implying that w 2 K. But w 2 K would be in contradiction
a; c [ c; w
2F 0
to c D sup K. Hence, it must be that b 2 K, which proves that a; b is compact.
t
u
A convenient characterisation of compactness is given by Proposition 2.5 below.

Definition 2.4.
\ A family fE g2 of subsets E X has the finite intersection
property if
E ; for every finite subset F .
2F
Proposition 2.5. The following statements are equivalent for a topological

space X:
1. X\is compact;
F ;, for every family fF g2 of closed sets F X with the finite
2.
2
intersection property.
t
u
X, and if 0 < " 12 d.x1 ; x2 /, then B" .x1 / \
B" .x1 / and B" .x2 / of x1 and x2 , respectively,
In a metric space .X; d/, if x1 ; x2 2

B" .x2 / D ;. Thus, the neighbourhoods
separate the points x1 and x2 . Topological spaces with this separation property are
called T2 -spaces or Hausdorff spaces.
2.1 Compact Spaces
41
Definition 2.6. A topological space X is a Hausdorff space if, for every pair of
distinct points x; y 2 X, there are neighbourhoods U and V of x and y, respectively,
such that U \ V D ;.
The definition of topology is so general that the axioms on their own are
insufficient to settle the question of whether point sets (that is, sets of the form
fxg, for x 2 X) are closed. By adding the Hausdorff separation axiom, this fact about
closedness can be deduced.
Proposition 2.7. If Y is a finite set in a Hausdorff space X, then Y is closed.
t
u
The Hausdorff property also has a role in determining which subsets of a space
are compact.
Proposition 2.8. Assume that K X.
1. If X is compact and K is closed, then K is compact.
2. If X is Hausdorff and K is compact, then K is closed.
Proof. Suppose that X is compact and K is closed, and suppose that fF g2 is
a family of closed sets F K with the finite intersection property. Because K is
closed, the\
family fF g2 is also closed in X (Proposition 1.61); and because X is
compact,
F ;, by Proposition 2.5. Hence, K is compact.
2
Next, suppose that X is Hausdorff and K is compact. Let U D K c and choose x 2

U. Because X is Hausdorff, for every y 2 K there are disjoint neighbourhoods Uy and
Vy of x and y, respectively. By the compactness of K, the open cover fVy gy2K admits
a finite subcover fVyj gnjD1 . The corresponding sets Uy1 ; : : : ; Uyn are neighbourhoods
n
n
\
[
Uyj is a neighbourhood of x disjoint from
Vyj and, hence,
of x, and so Wx D
jD1
disjoint from K. Thus, K D

c
jD1
Wx is open, which implies that K is closed.
t
u
x2K c
Compactness is a topological property that is preserved under continuous maps

of spaces.
Proposition 2.9. Suppose that f W X ! Y is a continuous maps of topological
spaces.
1. If X is compact, then f .X/ is a compact subset of Y.
2. If X is compact, Y is Hausdorff, and if f is a bijection, then f 1 is continuous.
Proof. Assuming that X is compact, let fV g2 be an open cover of f .X/.
Thus, ff 1 .V /g2 is an open cover of X and therefore admits a finite subcover
ff 1 .V /g2F for some finite subset F . The inclusion
42
f .X/ f
[
2F
!
f
1
.V /
.V /;
2F
implies that f.V /g2F is a finite subcover of f .X/, and so f .X/ is compact.
If, in addition to X being compact, Y is Hausdorff and f is bijective, then consider
the function g D f 1 W Y ! X. If K X is closed, then K is compact; and, by what we
just proved, g1 .K/ D f .K/ is compact in Y. Because Y is Hausdorff, f .K/ is closed.
Hence, g1 .K/ is closed in Y for every closed set K in X, which proves that g D f 1
is continuous.
t
u
The second assertion of Proposition 2.9 says that if f W X ! Y is a continuous
bijection of a compact space X to a Hausdorff space Y, then f is a homeomorphism,
which simplifies considerably the proof in Example 1.100 that the quotient space
R=Z and the unit circle S1 are homeomorphic.
Example 2.10. R=Z ' S1 .
Proof. Example 1.96 already shows that the map f W R=Z ! S1 defined by f .Pt/ D
.cos 2
t; sin 2
t/, for t 2 R, is continuous. As f is plainly bijective, and because
S1 is Hausdorff (as a subspace of the metric space R2 ), all that remains is to
show that R=Z is compact. To this end, let fV g2 be an open cover of R=Z.
Thus, fq1 .V /g2 is an open cover of R and, in particular, of the closed interval
0; 1. By the compactness of 0; 1, there exists a finite subset F
S such that
fq1 .V /g2F is a cover of 0; 1. The inclusion R=Z D q .0; 1/ 2F .V / shows
that R=Z is compact. Hence, by Proposition 2.9, f is a homeomorphism.
t
u
Another application of Proposition 2.9 is Proposition 2.12 below, which is similar
in essence to the familiar theorem from group theory that the range of a group
homomorphism ' W G ! H is isomorphic to the quotient group G= ker '.
Definition 2.11. A topological space Y is a quotient of a topological space X if
there exists an equivalence relation on X such that Y and the quotient space X=
are homeomorphic.
Proposition 2.12. If X and Y are compact Hausdorff spaces, and if f W X ! Y is a
continuous surjection, then Y is a quotient of X.
Proof. Define a relation
on X by x1
x2 , if f .x1 / D f .x2 /. It is plain to see that
is an equivalence relation. Define a map g W .X=

/ ! Y by g.Px/ D f .x/. Note that
g is well defined and that f D g q, where q W X ! X=
is the canonical quotient
map. Hence, by Proposition 1.97, the continuity of f and q imply the continuity of g.
The continuous map g is plainly a bijection. Therefore, in light of the fact that X=
is compact (Exercise 1.129) and Y is Hausdorff, Proposition 2.9 implies that g is a

homeomorphism.
t
u
The following proposition, which asserts that a continuous real-valued function
on a compact space achieves both its maximum and minimum values, is one of the
single-most important results concerning continuous functions on compact sets.
2.1 Compact Spaces
43
Proposition 2.13. If f W X ! R is a continuous map of a compact space X, then

there are x0 ; x1 2 X such that f .x0 / f .x/ f .x1 / for all x 2 X.
Proof. Let " > 0 be given, and consider the open cover fB" .f .x//gx2X of f .X/. By
Proposition 2.9, f .X/ is a compact subset of R; thus, there are x1 ; : : : ; xn 2 X such
that fB" .f .xj //gnjD1 covers f .X/. Hence, f .X/ is contained within the union of a finite
number of finite open intervals, which implies that f .X/ is bounded both below and
above. Therefore, by the completeness property of the real numbers, c D inf f .X/
and d D sup f .X/ exist.
Suppose that c 62 f .X/; then .c; c C "/ \ f .X/ ; for every " > 0. Therefore,
f.y; 1/gy2f .X/ is an open cover of f .X/, and this cover admits a finite subcover
f.yj ; 1/gnjD1 for some y1 ; : : : ; yn 2 f .X/. Let y D minfy1 ; : : : ; yn g. However, y 2
f .X/ .y; 1/ and y 62 .y; 1/ is a contradiction. Hence, it must be that c 2 f .X/,
and so c D f .x0 /, for some x0 2 X. A similar argument yields d D f .x1 /, for some
x1 2 X.
t
u
The closed interval 0; 1 is compact, and one might imagine that the same would
be true of the closed unit square, 0; 1 0; 1, and, in higher dimensions, of the
closed unit n-cube 0; 1n . Such is indeed the case as a consequence of the following
theorem of Tychonoff, which is a powerfully general result.
Theorem 2.14Q
(Tychonoff). If fX g2 is a family of compact spaces, then the
product space 2 X is compact in the product topology.
Q
Proof. Let X D X , and suppose that G D fG g2 is a family of subsets of X
with the finite intersection\
property. By Exercise 2.82, to prove that X is compact it
is sufficient to prove that
G ;.
2
In what follows, E shall denote an arbitrary family of subsets of X. Consider

S D fE P.X/ j E G and E has the finite intersection propertyg :
Impose a partial order on G by inclusion: E E 0 if E E 0 .
[
Suppose that L S is an arbitrary totally ordered subset, and define U D
E.
E 2L
Select n 2 N, and any E1 ; : : : ; En 2 U . Because L is totally ordered, there exists

E 2 L such that Ej 2 E for j D 1; : : : ; n. The finite intersection property of the
n
\
family E yields
Ej ;, which proves that U has the finite intersection property.
jD1
Moreover, U plainly contains G . Hence, U 2 S and E U for every E 2 L; that

is, U is an upper bound in S for the totally ordered subset L. Therefore, by Zorns
Lemma (Theorem 1.8), S contains a maximal element, which we denote by M .
Two observations concerning M are:
(i) if Y X satisfies Y \ M ; for every M 2 M , then Y 2 M ; and
n
\
(ii) if M1 ; : : : ; Mn 2 M , then
Mj 2 M .
jD1
44
The first observation above is verified by noting that the hypothesis implies that
fYg [ M is in S, and that M fYg [ M , which can occur only if M D fYg [ M .
Therefore, Y 2 M . The second observation is a consequence of the first: let Y be the
n
\
Mj , and make use of the fact that M has the finite intersection property.
set Y D
jD1
For each 2 , consider the family E D fp .M/gM2M of subsets of X , where

each p W X ! X is the projection map. Fix , and select finitely many sets in
n
\
Mj 2 M by observation (ii), it is also true
E , say p .M1 /; : : : ; p .Mn /. Because
that
n
\
jD1
p .Mj / ;. Thus, the family E has the finite intersection property in the
jD1
compact space X , and so, by Exercise 2.82,
p .M/ ;.
M2M
Suppose now that x D .x /2 , where each x 2
p .M/. If 2 and V is
M2M
a neighbourhood of x in X , then U D p1

.V / is a basic open set in X containing
x. Moreover, because x 2 p .M/ for every M 2 M , we deduce that V \ p .M/
; for every M 2 M . That is, for each M 2 M there is a yM 2 X such that yM 2
p1
.V / \ M. Hence, U \ M ; for every M 2 M . By observation (i), we deduce
that U 2 M .
Lastly, suppose that B X is a basic open set containing x. Thus, there are
n
n
\
\
p1
.V
/
D
Uj , where Vj is a neighbourhood
1 ; : : : ; n 2 such that B D
j
1
jD1
jD1
of xj in Xj . Because each Uj 2 M , observation (i) yields B 2 M . Hence, x 2 B

\
M. Now because M G ,
and B \ M ; for every M 2 M implying that x 2
it is also true that
M2M
G ;, thereby proving that X is compact.
t
u
G2G
Another use of compactness in analysis arises from the notion of convergence.

Definition 2.15. A sequence fxn gn2N of elements xn in a topological space X
is convergent if there exists an element x 2 X with the property that for every
neighbourhood U of x there is a positive integer kU 2 N such that xn 2 U for every
n kU . Such an element x is called a limit of the sequence.
In Hausdorff spaces, limits of convergent sequences are unique.
Proposition 2.16. If x and x0 are limits of a convergent sequence fxn gn2N in a
Hausdorff space X, then x0 D x.
t
u
In metric spaces, compactness has numerous advantageous features, such as the

uniform continuity of continuous functions (Proposition 2.18 below). A key lemma
for the analysis of compact sets in metric spaces is the following result, the proof of
which is outlined in Exercise 2.85.
2.1 Compact Spaces
45
Lemma 2.17. If fU g2 is an open cover of a compact metric space .X; d/,
then there exists a > 0 such that for each nonempty subset S X for which
supfd.s1 ; s2 / j s1 ; s2 2 Sg < there is a 0 2 such that S U0 .
t
u
Proposition 2.18 (Uniform Continuity). Suppose that f W X ! Y is a continuous

function, where .X; dX / and .Y; dY / are metric spaces. If X is compact, then for every
" > 0 there exists a > 0 such that dY .f .x1 /; f .x2 // < " for all x1 ; x2 2 X that satisfy
dX .x1 ; x2 / < .
Proof. Let " > 0 and consider the open covering ff 1 .B" .y//gy2Y of X. By
Lemma 2.17, there is a 0 > 0 for which any nonempty subset S X that
satisfies supfd.s1 ; s2 / j s1 ; s2 2 Sg < 0 also satisfies S f 1 .B" .y// for at least one
y 2 Y. Therefore, if D 0 =2 and if x1 ; x2 2 X are such that dX .x1 ; x2 / < , then
taking S D B .x1 / leads to B .x1 / f 1 .B" .y// for some y 2 Y, implying that
dY .f .x1 /; f .x2 // < ".
t
u
Another major feature of compactness in metric spaces is the following characterisation of compactness in terms of convergent subsequences.
Theorem 2.19. The following statements are equivalent for a metric space
.X; d/:
1. X is compact;
2. if fxn gn2N is a sequence in X, then there is a subsequence fxnj gj2N of fxn gn2N that
is convergent to some x 2 X.
Proof. Suppose first that X is compact. If fxn gn2N is a finite set, then some elements
in the sequence are repeated infinitely often. The constant subsequence extracted
from an element x 2 fxn gn2N that is repeated infinitely often is trivially convergent
to x. Therefore, suppose that Y D fxn gn2N is an infinite set of elements of X.
Consider the set L.Y/ of limit points of Y, and recall that x 2 L.Y/ if and only if
.U n fxg/ \ Y ; for every neighbourhood U of x. Hence, if x 2 L.Y/, then for each
j 2 N there is an element xnj 2 .B1=j .x/ n fxg/ \ Y, and so the subsequence fxnj gj2N of
fxn gn2N converges to x.
The only issue left to resolve is whether it is indeed true that L.Y/ ;. This is
settled by using the compactness and Hausdorff properties of X. If L.Y/ contained
an element x, then .U n fxg/ \ Y ; for every neighbourhood U of x. Furthermore,
more is true: namely, U \ Y is infinite for every neighbourhood U of x. To prove
this assertion, suppose that U is a neighbourhood of x and that U \ Y is finite. Thus,
.U n fxg/ \ Y is also a finite set, and therefore, by Proposition 2.7, .U n fxg/ \ Y is
a closed set. The complement W of .U n fxg/ \ Y is open and, hence, so is V D
U \ W. Because V is a neighbourhood of x 2 L.Y/, .V n fxg/ \ Y is nonempty. But
.V n fxg/ \ Y ; is in contradiction to .V n fxg/ \ Y D W c \ W D ;. Therefore, it
must be that U \ Y is infinite.
46
To complete the verification that L.Y/ ;, suppose, on the contrary, that the set
L.Y/ D ;. Thus, the equality of Y and Y [ L.Y/ implies that Y is a closed set in
a compact space X; therefore, Y is also compact. The fact that xn 62 L.Y/, for each
n 2 N, implies that there is a neighbourhood Un of xn such that Un \ Y is a finite
set (by the previous paragraph). From the open cover fUn gn2N of the compact space
m
[
.
Thus,
Y

Unj . However, Y \ Unj if finite
Y, extract a finite subcover fUnj gm
jD1
jD1
for each j, and so Y itself must be a finite set, which is a contradiction. Therefore, it
must be that L.Y/ ;.
Conversely, suppose that every sequence in X admits a convergent subsequence.
Let fU g2 be an arbitrary open cover of X. We claim that the conclusion of
Lemma 2.17 holds: namely, that there exists a > 0 such that, for each nonempty
subset S X for which supfd.s1 ; s2 / j s1 ; s2 2 Sg < , there is a S 2 such that
S US . If this were not true, then for each n 2 N there would exist a subset
Sn X such that d.x; y/ < 1n for all x; y 2 Sn and Sn 6 U for every 2 . Selecting
an element xn from each set Sn yields a sequence fxn gn2N and, by hypothesis, a
convergent subsequence fxnj gj2N . If x is the limit of the convergent subsequence
fxnj gj2N , then x 2 U0 for some 0 2 and there exists a " > 0 such that B" .x/ U0 .
Furthermore, there exists a N1 2 N such that xnj 2 B"=2 .x/ for all nj N1 . By the
assumption on the sets Sn , there also exists N2 N1 such that Snj B"=2 .xnj / if
nj N2 . This would then imply that Snj U0 for any nj that satisfies nj N2 ,
which is in contradiction to the assumption.
As shown in the previous paragraph, the covering fU g2 of X yields a > 0
such that for each nonempty subset S X for which supfd.s1 ; s2 / j s1 ; s2 2 Sg <
there is a S 2 such that S US . Let " D =3 and consider the open cover of X
given by fB" .x/gx2X . If fB" .x/gx2X were not to admit a finite subcover of X, then for
any x1 2 X there would exist an element x2 2 X such that x2 62 B" .x1 /; likewise, there
would exist x3 2 X such that x3 62 B" .x2 / [ B" .x1 /. Indeed, continuing by induction,
if x1 ; : : : ; xn 2 X are chosen so that d.xi ; xj / " whenever j 6D i, then there would also
exist xnC1 2 X such that xnC1 62 [njD1 B" .xj /. However, this would yield a sequence
fxn gn2N for which d.xi ; xj / " for all i; j 2 N with j 6D i, thereby making it impossible
for fxn gn2N to admit a convergent subsequence. Hence, it must be that fB" .x/gx2X
admits a finite subcover fB" .zj /gnjD1 of X for some elements z1 ; : : : ; zn 2 X. Because,
for fixed j, d.x; zj / < 23 < for all x 2 B" .zj /, there exists j 2 with B" .zj /
Uj . Hence, the covering fU g2 of X yields a finite subcovering fUj gnjD1 , which
implies that X is compact.
t
u
Compactness is a global feature of a topological space. A variant of compactness,
defined below, is a local feature.
Definition 2.20. A topological space X is locally compact if for each x 2 X there is
a neighbourhood U of x and a compact subset K X such that x 2 U K.
By definition, every compact space is locally compact. The following example is
a very familiar non-compact space that exhibits the local compactness property.
2.1 Compact Spaces
47
Example 2.21. Rn is locally compact.

Proof. Choose any x D .x1 ; : : : ; xn / 2 Rn . For each j let .aj ; bj / R be an open
n
n
Y
Y
interval that contains xj , and let U D
.aj ; bj / and K D
aj ; bj . By definition
jD1
jD1
of product topology, U is open; and by Tychonoffs Theorem, K is compact. As

x 2 U K, this shows that Rn is locally compact.
u
t
If one imagines the real line R as being anchored on the left by 1 and on
the right by C1, then joining these two ends yields a circle, thereby embedding
the noncompact space R into the compact space S1 . This conceptual idea is made
rigorous in the following manner.
Theorem 2.22 (One-Point Compactification). If X is a non-compact locally
compact Hausdorff space, then there is a compact Hausdorff space XQ such that
1. X XQ and XQ n X is a singleton set, and
Q
2. X is open and dense in X.
Proof. Let 1 denote an element that is not in X and let XQ D X [ f1g. Define the
Q
following collection B of subsets of X:
B D fU X j U is open in Xg
fXQ n K j K is compact in Xg :
To prove that B is a basis, note that, if x 2 X, then x 2 U for some open subset U
of X. Also, 1 2 XQ n fxg. Hence, for every x 2 XQ there is a B 2 B such that x 2 B.
Suppose now that B1 ; B2 2 B and that x 2 B1 \ B2 . If B1 and B2 are open sets in
X, then let B D B1 \ B2 2 B to obtain x 2 B B1 \ B2 . Next, if B1 is open in X
and if B2 D XQ n K for some compact K X, then necessarily x 6D 1 and we may
let B D B1 \ .X n K/ to obtain a B 2 B with x 2 B B1 \ B2 . Lastly, assume that
Bj D XQ n Kj for some compact Kj X, j D 1; 2. Let K D K1 [ K2 , which is compact
in X, so that XQ n K D B1 \ B2 , yielding an element B 2 B with x 2 B B1 \ B2 .
Let T be the topology on XQ with basis B. Choose any neighbourhood V of 1,
and let B 2 B be a basic open set with 1 2 B V. Thus, B D XQ n K for some
compact set K. Because K 6D X, there is an x 2 X such that x 62 K. Hence, V \ X
Q T /). Hence, X is open,
B \ X 6D ;, which implies that 1 2 X (the closure of X in .X;
Q
because X 2 B, and dense in X.
The proof that XQ is Hausdorff is left as an exercise (Exercise 2.93). To show
Q Thus, there is a 2 for which
compactness, let fV g2 be an open cover of X.
1 2 V . Choose any B 2 B for which 1 2 B V ; thus, B D XQ n K for some
compact subset K X. As fX \ V g2 is an open cover of K, there are 1 ; : : : :n
n
[
S
Q
such that K .X \ Vj /. Hence fV g fVj gnjD1 is a finite subcover of X.
t
u
jD1
Definition 2.23. The compact Hausdorff space XQ in Proposition 2.22 is called the
one-point compactification of X.
48
2.2 Topological Properties Described by Nets

Formally, a sequence in a set X is a function f W N ! X. One can of course replace
N by some other set . However, N is not simply a set, it also comes with a linear
order . Therefore, to extend the notion of sequence to more general contexts, it is
useful to consider certain partially ordered sets .; /.
Definition 2.24. If is a partial order on a partially ordered set , then is a
directed set if for each pair of ; 2 there is a 2 such that and .
Definition 2.25. A net in a set X is a function ' W ! X for some partially ordered
directed set .; /.
As with sequences, it is notationally economical to express a net ' W ! X by
its values x D './, 2 . Thus, a net in X is a family fx g2 of elements x 2 X
for some directed set .; /.
Definition 2.26. If X is a topological space, then a net fx g2 in X is convergent
to x 2 X if for every open neighbourhood U X of x there is a 0 2 such that
x 2 U for all 2 satisfying 0 .
The notation x D lim x will be used to signify that a net fx g2 in X is
convergent to x 2 X.
The following proposition shows the relationship between topological closure
and the convergence of nets.
Proposition 2.27. If Y is a nonempty subset of a topological space X, then the
following statements are equivalent for an element x 2 A:
1. x 2 Y;
2. there exists a net fy g2 in Y such that x D lim y .
Proof. The proof of (2) implies (1) is immediate from the definition of convergent
net (Exercise 2.95).
Therefore, suppose that x 2 Y. Let be the set of all open subsets U X for
which x 2 U, and let U V denote V U, for U; V 2 . Note that is a partial
order on . Furthermore, if U; V 2 and if W D U \ V, then W 2 and U W
and V W. Hence, is a directed set.
For each U 2 , choose yU 2 U \ Y (such an element yU 2 Y exists because
x 2 Y), and define the map ' W ! Y by '.U/ D yU . The net fyU gU2 has the
property that for every open neighbourhood V of x there exists an element V0 2
(namely, V0 D V) such that yW 2 U for all W 2 for which V0 W. That is, the net
t
u
fyU gU2 in Y converges to x.
The use of nets allows for the following rephrasing of the property of continuity
for functions on topological spaces.
Proposition 2.28. If X and Y are topological spaces, then the following statements
are equivalent for a function f W X ! Y:
2.2 Topological Properties Described by Nets
49
1. f is continuous;
2. for every convergent net fx g2 in X with limit x 2 X, ff .x /g2 is a convergent
net in Y with limit f .x/.
t
u
In analysis, the language of nets gives a very convenient method for certain
proofs. Possibly the most important of these methods is provided by Proposition 2.31 below, which characterises compactness in terms of convergent subnets.
Definition 2.29. A cofinal subset of a directed set .; / is a subset with
the property that for each 2 there exists 2 such that .
Note that if is a cofinal subset of a directed set .; /, and if 1 ; 2 2 , then
in considering 1 and 2 as elements of there necessarily exists 2 with 1
and 2 . Because is cofinal, there in turn exists 2 with . Hence, 1
and 2 , which shows that .; / is itself a directed set.
Definition 2.30. Assume that ' W ! X, for some directed set .; /. A subnet of
the net ' W ! X is a function # W ! X of the form # D ' , where W !
Q such that
is a function on a directed set .; /
1.
2.
./ is a cofinal subset of , and

Q 2.
.!1 / .!2 /, for all !1 ; !2 2 with !1 !
For the purposes of simplified notation, if fx g2 denotes a net, then fx! g!2
shall denote a subset of fx g2 .
space X:
1. X is compact;
2. every net in X admits a convergent subnet.
Proof. We shall make T
use of the criterion of Proposition 2.5: namely, that X is
compact if and only if 2 F 6D ; for every family fF g2 of closed sets with
the finite intersection property.
To begin, suppose that X is compact and let fx g2 be an arbitrary net in X.
Formally, there is a function ' W ! X on a directed set .; / such that './ D x
for every 2 . For each 2 , let S D fx 2 X j g. Because is a directed
set, S1 \ S2 6D ; for all 1 ; 2 2 . Hence, the collection fS j 2 g has the finite
intersection property and so, by the compactness of X, there exists x 2 X such that
x 2 S for every 2 .
Let O D fU X j U is an open set, and x 2 Ug and, for each U 2 O, define U D
f 2 j x 2 Ug. Each of the sets U is cofinal in . To verify this assertion, choose
U 2 O and 2 . Because x 2 S , the open set U has nonempty intersection with
S . Thus, there is some x 2 S \ U. Because the element satisfies and is,
by definition, an element of U , the subset U is therefore cofinal.
50
Consider the subset of O that consists of all .; U/ for which 2 U , and

Q on by
define a partial order
Q
.; U/.;
V/; if and V U:
If .; U/; .; V/ 2 , then let 2 be an element for which and .
Let W D U \ V. Because W is confinal in , there is a 2 W such that .
Q W/ and
Hence, the ordered pair .; W/ is an element of and satisfies .; U/.;
Q W/, which proves that .; /
Q is a directed set.
.; V/.;
The function W ! defined by .; U/ D plainly gives rise to a subnet
fx.;U/ g.;U/2 of fx g2 . To show that this subnet converges to x, select any open
set U X that contains x. Pick 0 2 . Because U is cofinal, there is a 2 U
Q
such that 0 . Therefore, .; U/ 2 . If .; V/ 2 satisfies .; U/.;
V/, then
x.;V/ 2 V U. Hence, the subnet fx.;U/ g.;U/2 of fx g2 is convergent.
Conversely, suppose that every net in X admits a convergent subnet. Suppose
that fF g2 is a family of closed sets that has the finite intersection property. Let
F D fF j F is finiteg and define on F by F G if G F. Because fF g2
has
T the finite intersection property, .F ; / is a directed set. For each F 2 F , the set
2F F is nonempty; thus, select xF in this intersection. Now define ' W F ! X
by '.F/ D xF ; that is, consider the net fxF gF2F . By hypothesis, fxF gF2F admits a
convergent subset net fx! g!2 , with limit x 2 X, where without loss of notational
generality we may assume that is a cofinal subset of F .
T
Fix 2 . We shall prove that x 2 F , which will imply that 2 F is
nonempty. To this end, let U be any open neighbourhood U of x. Thus, there is
a G0 2 such that xG 2 U for all G 2 for which G0 G. Now because fg 2 F ,
there is an F 2 F with fg F and G0 F. As is cofinal in F , we may in
fact assume that F 2 . Hence, F fg implies that xF 2 F . Furthermore, G0 F
implies that xF 2 U. Hence, U \ F is nonempty. Because U is an arbitrary open
neighbourhood of x, we deduce that x 2 F D F .
t
u
Observe that the statement of Proposition 2.31 is almost superficially trivial,
while all of the underlying topology embodied by the statement is buried within
the proof. This is what makes the use of nets so compelling in analysis.
The Hausdorff property may also be characterised by a property of nets; however,
the proof of Proposition 2.32 below is not nearly as subtle as the proof of
Proposition 2.31.
space X:
1. X is Hausdorff;
2. every convergent net fx g2 in X has a unique limit point.
t
u
2.3 Normal Spaces
51
2.3 Normal Spaces

Definition 2.33. A topological space X is normal if
1. fxg is a closed set for every x 2 X, and
2. for every pairs of disjoint closed sets C; F X there exist disjoint open sets
U; V X with C U and F V.
The notion of normality is a separation property, not unlike the separation
property that defines a Hausdorff space.
Proposition 2.34. Every compact Hausdorff space is normal.
Proof. Suppose that X is a compact Hausdorff space. By the Hausdorff property,
fxg is closed for every x 2 X (Proposition 2.7). Therefore, assume that C; F X are
closed sets such that C \ F D ;. Because C and F are closed and X is compact, both
C and F are compact.
Select x 2 F. For each y 2 C there are neighbourhoods Vy of x and Uy of y such
that Uy \ Vy D ;. The family fUy gy2C is an open cover of C and therefore admits a
n
n
\
[
finite subcover fUyj gnjD1 for some y1 ; : : : ; yn 2 C. Let Vx D
Vyj and Ux D
Uyj ;
jD1
jD1
thus, Vx is a neighbourhood of x, Ux C, and Ux \ Vx D ;.

Carrying out the procedure above for every x 2 F leads to an open cover
fVx gx2F of F. By the compactness of F, there is a finite subcover fVxi gm
iD1 for some
m
m
[
\
Vxi and U D
Uxi , which are open sets for which C U,
x1 ; : : : ; ym 2 F. Let V D
iD1
iD1
F V, and U \ V D ;. Hence, X is normal.
t
u
Another class of normal spaces is that of metric spaces.

Proposition 2.35. Every metric space is normal.
Proof. Suppose that X is a metric space. By the Hausdorff property of metric spaces,
fxg is closed for every x 2 X.
Next, suppose that C; F X are closed sets such that C \ F D ;. Select x 2 C.
Then x is contained in the open set F c , and so there is a "x > 0 such that B"x .x/\
F D ;. Similarly, for each y 2 F there is a "y > 0 such that B"y .y/ \ C D ;. Let
UD
[
x2C
B"x =2 .x/
and
V D
B"y =2 .y/:
y2F
Thus, U and V are open sets such that C U and F V. If, for some x 2 C and
y 2 F, B"x =2 .x/ \ B"y =2 .y/ were nonempty, then via z 2 B"x =2 .x/ \ B"y =2 .y/ we would
obtain
d.x; y/ d.x; z/ C d.z; y/ < "x =2 C "y =2 maxf"x ; "y g;
52
implying that y 2 B"x .x/, or x 2 B"y .y/, both of which are in contradiction of the fact
that B"y .y/ \ C D B"x .x/ \ F D ;. Hence, B"x =2 .x/ \ B"y =2 .y/ D ; for all x 2 C and
y 2 F, and therefore U \ V D ;.
t
u
An alternate characterisation of normality is given by the following proposition.
space X:
1. X is normal;
2. X has the properties that
a. fxg is a closed set, for all x 2 X, and
b. for every closed set F and open set U for which F U, there is an open set V
with F V V U.
t
u

The following theorem captures the most important feature of normal spaces.
Theorem 2.37 (Tietze Extension Theorem). If A is a closed subset of a normal

topological space X, and if f W A ! 0; 1 is a continuous function, then there exists
a continuous function F W X ! 0; 1 such that F.x/ D f .x/ for every x 2 A.
The function F above is called an extension of f .
Proof. Select p 2 Q and set Ap D f 1 ..1; p/ and Bp D f 1 .p; 1//, which are
closed subsets of A and hence of X (since A is closed); let Up D Bcp , which is an open
subset of X. Observe that, if p; q 2 Q satisfy p q, then Ap Aq and Up Uq ; and,
if p < q, then Ap Uq .
Let be the partial order on Q Q in which .p; q/ .p0 ; q0 /, if p p0 and q q0 ,
and let P D f.p; q/ 2 Q Q j 0 p < q 1g.
The set P is countable, and so there is an enumeration f.pn ; qn /gn2N of its
elements. With the first of these elements, .p1 ; q1 /, the inclusion Ap1 Uq1 and the
normality of X imply that there is an open set V1 X such that Ap1 V1 V 1 Uq1
(Proposition 2.36). Consider now the next element of P, namely .p2 ; q2 /. The
partial order on Q Q is not a total order; therefore, we must consider the
following subcases.
(i) .p1 ; q1 / .p2 ; q2 /: The inclusions Ap1 Ap2 Uq2 and V 1 Uq1 Uq2 imply,
by Proposition 2.36, the existence of an open set V2 X such that

Ap2 [ V 1 V2 V 2 Uq2 :
In particular, V 1 V2 .
(ii) .p2 ; q2 / .p1 ; q1 /: We have that Ap2 Ap1 Uq1 , because p2 p1 < q1 , and we
know already that Ap1 V1 . Thus, Ap2 Uq1 \ V1 . Again, by Proposition 2.36,
there is an open set V2 X for which
2.3 Normal Spaces
53

Ap2 V2 V 2 Uq1 \ V1 :
In particular, V 2 V1 .
(iii) Neither .p1 ; q1 / .p2 ; q2 / nor .p2 ; q2 / .p1 ; q1 /: In this case, p2 < q2 , and
Proposition 2.36 implies that there is an open set V2 X such that Ap2 V2
V 2 Uq2 .
Therefore, from what has been argued thus far, we have that: (a) Apk Vk V k
Uqk for k D 1; 2; (b) V 1 V2 , if .p1 ; q1 / .p2 ; q2 /; (c) V 2 V1 , if .p2 ; q2 / .p1 ; q1 /.
Based on the arguments above, proceed by induction. Suppose that open sets
Vk X have been constructed for k D 1; : : : ; n 1 with the properties
(a) Apk Vk V k Uqk for k D 1; : : : ; n 1, and
(b) V j Vk , if .pj ; qj / .pk ; qk /.
Define

Jn D j j 1 j n 1 and .pj ; qj / .pn ; qn / and
Kn D fk j 1 k n 1 and .pn ; qn / .pk ; qk /g :
By Proposition 2.36, there exists an open set Vn X such that
0
Apn [ @
V j A Vn V n 4Uqn \ @
j2Jn
13
Uqk A5 :
k2Kn
Relabel the sets Vn as Vpq , where p D pn and q D qn . Thus, by the principle of

mathematical induction, there exists a family fVpq g.p;q/2P of open sets Vpq X for
0 0
which Ap V\
pq V pq Uq , for all .p; q/ 2 P, and V pq Vp0 q0 , if .p; q/ .p ; q /.
V pq for each p 2 Q, and observe that Xp D ; for every p < 0, and
Let Xp D
q>p
that Xp D X for every p 1. For each .p; q/ 2 P, choose t 2 Q such that p < t < q;
in so doing, we have the inclusions:
Xp V pt Vtq
V qs D Xq :
s>q
Note also that Ap
V pq D Xp , for every p 2 Q \ 0; 1/. Thus,
q>q
Ap D Xp \ A D
\
q>p
!
V pq \ A
!
Uq \ A
D Ap :
q>p
Therefore, fXp gp2Q is a family of closed sets such that Xp int Xq , for all q > p, and
Xp \ A D Ap , for all p.
54
The extension F W X ! 0; 1 of f W A ! 0; 1 is defined as follows: for each x 2 X,

let
F.x/ D inf p 2 Q j x 2 Xp :
Because
Xp D ; for every p < 0, and Xp D X for every p 1, for each x 2 X the set
p 2 Q j x 2 Xp is nonempty and is bounded below by 0 and above by 1; hence, F

is a well-defined function.
To show that F is indeed an extension of f , recall that Xp \ A D Ap . Thus, if x 2 A,
then
F.x/ D inf p 2 Q j x 2 Xp D inf p 2 Q j x 2 Ap D f .x/;

by the continuity of f .
All that remains, therefore, is to verify that F is continuous. Observe that F has
the following two properties:
1. if x 2 Xp , then F.x/ p, and
2. if x 62 int Xq for some q 2 Q, then q F.x/
The first of these properties above follows from the definition of F. For the second
property, suppose that x 62 int Xq and F.x/ < q. Because F is defined as an infimum,
there must be a rational p < q for which x 2 Xp . But p < q implies that Xp int Xq ,
in contradiction to the hypothesis. Therefore, it must be true that q F.x/.
Fix x 2 X, and let W be a neighbourhood in R of F.x/. Select p; q 2 Q such that
p < q and F.x/ 2 .p; q/ W. Define U D int Xq \ Xpc . Because p < F.x/, x 62 Xp ;
and because F.x/ < q, x 2 int Xq . Thus, x 2 U, and so U is a neighbourhood of x.
Lastly, if z 2 U, then F.z/ < q, because z 2 int Xq , and p < F.z/, because z 62 Xp .
Hence, F.U/ W, which proves that F is continuous at x 2 X. Since the choice of
x is arbitrary, we conclude that F is continuous at every x 2 X.
t
u
Corollary 2.38. If A is a closed subset of a normal topological space X, and if
f W A ! C is a bounded continuous function, then there exists a continuous function
F W X ! C such that F.x/ D f .x/ for every x 2 A and sup jF.x/j D sup jf .a/j.
x2X
a2A
Proof. Let D sup jf .a/j so that the range of f1 D 1 f is in the closed unit disc of
a2A
C. The range of the real and imaginary parts <f1 and =f1 of f1 , which themselves
are continuous functions on A, lie in the closed interval 1; 1. Therefore, the
ranges of g D 12 .<f1 C 1/ and h D 12 .=f1 C 1/ are contained in the closed interval
0; 1. By the Tietze Extension Theorem, each of the continuous functions g and h
admits continuous extensions G W X ! 0; 1 and H W X ! 0; 1, respectively, and so
F D ..2G 1/ C i.2H 1// is a continuous extension of f with the property that
sup jF.x/j D sup jf .a/j.
t
u
x2X
a2A
2.3 Normal Spaces
55
Corollary 2.39 (Urysohns Lemma). If A and B are disjoint closed subsets of a

normal space X, then there exists a continuous function f W X ! 0; 1 such that
f .A/ D f0g and f .B/ D f1g.
Proof. Let h W A [ B ! 0; 1 be defined by h.x/ D 0, if x 2 A, and h.x/ D 1, if x 2 B.
Because A \ B D ;, there is no ambiguity in the definition of h. Moreover, as h1 .F/
is closed in A [ B for every closed F 0; 1, the map h is continuous. By the Tietze
Extension Theorem, h admits a continuous extension f W X ! 0; 1.
The next result, Proposition 2.41 below, establishes is a widely used technical
tool in analysis.
Definition 2.40. If X is a topological space, then the support of a continuous
function f W X ! C is the set supp f X defined by
supp f D fx 2 X j f .x/ 6D 0g:
Note that the range f .X/ of f satisfies
f .supp f / f .X/ f .supp f / [ f0g:
Therefore, if f has compact support, then f .supp f / is compact, and so the range
f .X/ of f is also compact.
Proposition 2.41 (Partitions of Unity). If fUj gnjD1 is an open cover of a normal
space X, then there exist continuous functions h1 ; : : : ; hn W X ! R such that
1. 0 hj .x/ 1, for every x 2 X, and each j D 1; : : : ; n,
2. supp hj Uj , for j D 1; : : : ; n, and
n
X
3.
hj .x/ D 1, for every x 2 X.
jD1
1
0
n
[
Proof. Let F1 D @ Uj A, which is a closed subset of U1 . Because X is normal,
jD2
there is an open set V1 such that F1 V1 V 1 U1 . Note that F1 V1 implies

that fV1 g [ fUj gnjD2 is an open cover of X. By induction, if open sets V1 ; : : : Vk1
n
have been constructed such that fV` gk1
`D1 [ fUj gjDk is an open cover of0X such that
1c
!c
k1
n
[
[
V ` U` for all ` D 1; : : : ; k 1, then the closed set Fk D
Vj \ @
Uj A
`D1
jDkC1
will lie in some open set Vk such that V k Uk (by the normality of X), and therefore,
fV` gk`D1 [ fUj gnjDkC1 is an open cover of X. Hence, after n steps, there exists an open
cover fVj gnjD1 of X with the property that V j Uj , for each j.
Applying the argument of the previous paragraph to the open cover fVj gnjD1 of X
produces an open cover fWj gnjD1 of X such that W j Vj , for each j. Use Urysohns
56
Lemma to select, for each j, a continuous function gj W X ! 0; 1 with gj .W j / D f1g

and gj .Vjc / D f0g. Hence, the support of gj is contained in V j Uj . Because fWj gnjD1
n
X
covers X,
gj .x/ 1 for every x 2 X. Thus, for each j, the continuous function
jD1
hj W X ! R defined by
hj .x/ D
n
X
!1
g` .x/;
gj .x/
`D1
satisfies supp hj Uj and 0 hj .x/ 1, for all x 2 X. Moreover, if x 2 X, then

n
X
hj .x/ D 1.
t
u
jD1
Definition 2.42. The family fh1 ; : : : ; hn g in Proposition 2.41 is called a partition of

unity of X subordinate to the open cover fUj gnjD1 .
Locally compact Hausdorff spaces need not be compact; however, it is still
possible to establish a Tietze Extension Theorem in this setting, which has important
consequences in analysis (such as Theorem 5.43).
Theorem 2.43 (Tietze Extension Theorem, II). If X is a locally compact Hausdorff space, if K X is compact, and if K U for some open set U X, then every
continuous function f W K ! C extends to a continuous bounded function F W X ! C
such that
1. supp F is a compact subset of U,
2. F.x/ D 0 for all x 2 U c , and
3. max jF.x/j D max jf .y/j.
x2X
y2K
Proof. By Exercise 2.91, there exists an open subset W X such that K W

U and W is compact. Let XQ be the one-point compactification of X. Because XQ
is compact and Hausdorff, XQ is normal. Note that X n K a basic open set in the
Q and so K is a closed subset of X.
Q Hence, K is compact subset of X.
Q
topology of X,
Q
Q
By Corollary 2.38, f W K ! C admits a continuous extension f W X ! C such that
max jfQ .x/j D max jf .y/j. In addition, because XQ nW is closed in XQ and is disjoint from
x2XQ
y2K
Q
Q XQ n
K, Urysohns Lemma yields a function hQ W XQ ! 0; 1 such that h.K/
D f1g and h.
W/ D f0g. Hence, the continuous and bounded function F W X ! C given by F D
.hQ fQ /jX satisfies F.x/ D f .x/ for all x 2 K, F.x/ D 0 for all x 62 W, and sup jF.x/j D
x2X
max jf .y/j. Furthermore, because supp F W and W is compact, we deduce that F

y2K
has compact support and, therefore, the supremum sup jF.x/j is achieved at some
point x0 2 X.
x2X
t
u
The following consequence of Theorem 2.43 will be used extensively in subsequent chapters.
2.4 Properties of Metric Spaces
57
Corollary 2.44 (Urysohns Lemma, II). If K and U are nonempty subsets of a

locally compact Hausdorff space X, and if K is compact, if U is open, and if K U,
then there exists a continuous function f W X ! 0; 1 such that f .K/ D f1g and supp f
is a compact subset of U.
Proof. Let f0 W K ! R be given by f0 .x/ D 1, for all x 2 K, and apply Theorem 2.43
to produce the desired extension f W X ! R with all the stated properties. That the
range of f lies in the interval 0; 1 is a consequence of the proof of Theorem 2.43
and the Tietze Extension Theorem.
t
u

Topological spaces that are very large topologically can behave poorly. A case
in point is a product space with the box topology. One measure of topological
smallness is compactness. Another smallness quality involves countability features
of a space.
Definition 2.45. A topological space .X; T / is:
1. separable, if there is a countable subset Y X such that Y is dense in X (that is,
Y D X);
2. second countable, if there is a countable basis B for the topology T of X.
Proposition 2.46. Every second countable space is separable.
Proof. Suppose that B D fBn gn2N is a countable basis for the topology T of a
topological space .X; T /. For each n select xn 2 Bn and let Y D fxn gn , which is
a countable subset of X. Suppose now that x 2 X and U is a neighbourhood of x.
Because B is a basis, there is a basic open set Bn for which x 2 Bn U. Now since
xn 2 Bn , we have that xn 2 U \ Y, implying that U \ Y 6D ;. Hence, Y D X.
t
u
The converse to Proposition 2.46 is not true in general, but it is true in an
important special case.
Proposition 2.47. Every separable metric space is second countable.
t
u
There are separable spaces that fail to be second countable (see Proposition 2.79,
for example). The following important theorem determines precisely when a
compact Hausdorff space is metrisable.
Theorem 2.48 (Compact Metrisable Spaces). The following statements are
equivalent for a compact Hausdorff space X:
1. X is metrisable;
2. X is second countable.
58
Proof. Suppose that X is metrisable. Let d be a metric on X that induces the topology
of X. For each n 2 N, the family fB1=n .x/gx2X is an open
[ covering ofn X; thus, there
k
k
is a finite subcovering fB1=n .xn;j /gnjD1
of X. Let B D
fB1=n .xn;j /gjD1
, which is a
n2N
countable collection. We claim that B is a basis for the topology on X. To this end,
let U 2 X be open and consider x 2 U. Because X is a metric space, there is an
n 2 N such that B1=n .x/ U. Likewise, there is an x0 2 fx3n;1 ; : : : ; x3n;.3n/k g for which
x 2 B1=3n .x0 /. Now since B1=3n .x0 / B1=n .x/, there is a set B 2 B with x 2 B U.
Hence, B is a basis for X, which implies that X is second countable.
Conversely, suppose that X is second countable, and let B D fBn gn2N be a
countable basis for the topology of X. Let
I D f.m; n/ 2 N N j Bm Bn g:
By Proposition 2.34, every compact Hausdorff space is normal. Hence, by
Urysohns Lemma, for each .m; n/ 2 I , there is a continuous map f.m;n/ W X ! 0; 1
for which f.m;n/ .Bm / D f1g and f.m;n/ .Bcn / D f0g.
Select x 2 X and a neighbourhood U of x. Thus, there is a basic open set Bn 2 B
with x 2 Bn U. Because fxg is a closed set and X is normal, Proposition 2.36 asserts
that there exists an open set V with x 2 V V Bn . Since V is a neighbourhood of x,
there is some Bm 2 B with x 2 Bm V. Hence, Bm V Bn ; therefore, .m; n/ 2 I ,
f.m;n/ .x/ D 1, and f.m;n/ .U c / D f0g.
The previous paragraph shows that there is a countable family fgn gn of continuous functions gn W X ! 0; 1 with the property that, for each x 2 X and
neighbourhood U of x, there is some n 2 N such that gn .x/ D 1 and gn .U c / D f0g.
The product topology on 0; 1N coincides with the subspace topology that
is induced by .RN ; T / (Proposition 1.51). The space .RN ; T / is metrisable
(Proposition 1.57), and therefore so is the subspace 0; 1N . Define f W X ! 0; 1N by
f .x/ D .gn .x//n :
Each coordinate function gn is continuous, and so, because 0; 1N has the product
topology, the function f is continuous (Proposition 1.92). The function f is also
injective, for if x; y 2 X are distinct, then they are separated by disjoint open sets U
and V, and so there is a function gn which maps x to 1 and y to 0, which implies that
f .x/ 6D f .y/.
The map f is a continuous bijection from X to f .X/ 0; 1N , and the subspace
f .X/ is a subspace of a metric space, and is therefore Hausdorff. Thus, Proposition 2.9 asserts that f is a homeomorphism, which proves that the topology on X is
metrisable.
t
u
The Cantor ternary function is a continuous map of the Cantor ternary set C
onto the closed unit interval 0; 1. There is a very interesting property that is shared
by all compact metric spaces.
59
Proposition 2.49. If X is a compact metric space, then there exists a continuous

function f W C ! X, where C is the Cantor ternary set, such that f is surjective.
Proof. By Exercise 1.133, the compact Hausdorff spaces C and C N (in the product
topology) are homeomorphic. Because the map Q W C N ! 0; 1N defined by
Q ..xn /n / D ..xn //n is continuous and surjective, the compact metric space 0; 1N
is a continuous image h.C / of C via some continuous function h. The proof of
Theorem 2.48 shows that X is homeomorphic to a closed subset of 0; 1N . Therefore,
without loss of generality, assume that X is a closed subset of 0; 1N and let
L D h1 .X/, which is a closed subset of C . If it can be shown that L D g.C / for
some continuous g, then f D h g would be a continuous surjection of C onto X.
In the product topology, f0; 1gN and f0; 5gN are obviously homeomorphic. By
Exercise 1.133, the former is homeomorphic to the Cantor ternary set C , while the
latter is homeomorphic to the set C2=3 defined by
(
C2=3 D
1
X
k
kD1
6k
)
j k 2 f0; 5g 0; 1:
Hence, C and C2=3 are homeomorphic, which implies that L is homeomorphic to

some closed subset K C2=3 .
The set C2=3 may be viewed as the Cantor two-thirds set, which has the
property that it does not contain the midpoint between any two of its elements. Let
dK W C2=3 ! R be the continuous function dK .x/ D sup jx yj. Because K is compact,
y2K
the supremum is achieved at some point yx 2 K. By the no midpoint geometry of

C2=3 , for each x 2 C2=3 the point yx 2 K is uniquely determined by x, and so the map
gQ W C2=3 ! K defined by gQ .x/ D yx is a continuous surjection. Hence, L and X are
continuous images of C .
t
u
Inspired by the metric completeness of the real numbers, one can consider the
metric completeness of more general metric spaces.
Definition 2.50. Assume that .X; d/ is a metric space.
1. A sequence fxk gk2N of elements xk 2 X is convergent to x 2 X if for every " > 0
there exists N" 2 N such that d.x; xn / < " for all n N" .
2. A sequence fxk gk2N of elements xk 2 X is a Cauchy sequence if for every " > 0
there exists N" 2 N such that d.xn ; xm / < " for all m; n N" .
The convergence of a sequence fxk gk2N to x is denoted by x D lim xk .
k!1
Definition 2.51. A metric space .X; d/ is complete if for every Cauchy sequence
fxk gk2N of elements xk 2 X there exists x 2 X such that x D lim xk .
k!1
Because metric spaces are Hausdorff, if a Cauchy sequence in .X; d/ is convergent, then the limit of this sequence is unique.
Example 2.52. Rn is a complete metric space in the Euclidean metric d2 .
60
Proof. Denote the canonical basis vectors of Rn by e1 ; : : : ; en , and suppose that fwk gk
is a Cauchy sequence of elements in Rn , where
wk D
n
X
.k/
j ej ;
8k 2 N:
jD1
.k/
.m/
.k/
Fix j. The inequality jj j j d2 .wk ; wm / implies that the sequence fj gk is

a Cauchy sequence in R. Theorem 1.14 asserts that R is a complete metric space in
.k/
the absolute-value metric; hence, there is an j 2 R such that j D lim j . This is
true for each j, and so let w D
n
X
k!1
j ej . Since
jD1
v
uX
u n
.k/
d2 .w; wk / D t .j j /2 ;
jD1
the sequence fwk gk converges to w. Hence, .Rn ; d2 / is a complete metric space.
t
u
Although not every metric space need be complete, the following theorem shows
that every metric space is a dense subset of some complete metric space.
Q d/
Q and a
Theorem 2.53. For every metric space .X; d/, there is a metric space .X;
Q
continuous injective function f W X ! X such that:
Q d/
Q is a complete metric space;
1. .X;
Q .x/; f .y// D d.x; y/, for all x; y 2 X; and
2. d.f
Q
3. f .X/ is a dense subset of X.
Proof. Let Z D
1
Y
X, the Cartesian product of countably many copies of X, which
nD1
we think of as the space of all sequences x D .xn /n2N in X. Let C Z be the set of all
Cauchy sequences and define a relation
on C by .xn /n
.yn / if limn d.xn ; yn / D 0.
It is straightforward to verify that
is an equivalence relation, and so consider the
space XQ D C=
of equivalence classes, whose elements we denote by sP for each
s D .sn /n 2 C. Let q W C ! XQ denote the quotient map q.s/ D sP .
If s; t 2 C, then the sequence fd.sn ; tn /gn is a Cauchy sequence in R (Exercise 2.101). The limit limn d.sn ; tn / exists, because R is a complete metric space,
Further, suppose that s; s0 ; t; t0 2 C satisfy sP D sP0 and Pt D tP0 . Then, for a fixed n 2 N,
d.sn ; tn / d.sn ; s0n / C d.s0n ; tn0 / C d.tn0 ; tn /;
and
d.s0n ; tn0 / d.s0n ; sn / C d.sn ; tn / C d.tn ; tn0 /:
61
Hence, limn d.sn ; tn / D limn d.s0n ; tn0 /, and so the formula

Q s; Pt/ D lim d.sn ; tn /
d.P
n!1
yields a well-defined function dQ W XQ XQ ! R. It is simple to verify that dQ is in fact a

Q
metric on X.
Consider the function W X ! C that sends each x 2 X to the constant (and
Cauchy) sequence .x/ D .x; x; x; : : : /, and define f W X ! XQ by f D q . Observe
Q .x/; f .y// D d.x; y/; thus, f is a continuous and injective map
that, for all x; y 2 X, d.f
(Exercise 2.102). If s 2 C belongs to the equivalence class of .x/, for an element
x 2 X, then x D limn sn in X.
Choose any sP 2 XQ and " > 0. If s D .sn /s 2 C is a representative of the class sP ,
then, by virtue of the fact that s is a Cauchy sequence, there is an N 2 N such that
Q .x/; sP / D
d.sn ; sm / < " for all n; m N. Let x D sN and consider f .x/. Because d.f
Q
limmN d.sN ; sm / < ", we deduce that B" .Ps/ \ f .X/ 6D ;. Hence, f .X/ D X.
Q d/
Q is a complete metric space, let fPsk gk be a Cauchy sequence
To show that .X;
Q sk ; sP ` / < "=3 for
Q and suppose that " > 0. Thus, there is an N0 2 N for which d.P
in X,
Q and so, for each k 2 N, there exists xk 2 X
all k; ` N0 . The set f .X/ is dense in X,
Q .xk /; sP k / < 1 . Thus,
such that d.f
k
Q .xk /; sP k / C d.P
Q sk ; sP ` / C d.P
Q s` ; f .x` //
Q .xk /; f .x` // d.f
d.f
1=k C "=3 C 1=`:
Q .xk /; f .x` // < ",
If N N0 is such that 1k < 3" and 1` < 3" for all k; ` N, then d.f
Q Furthermore,
for all k; ` N, which proves that ff .xk /gk is a Cauchy sequence in X.
Q
the equation d.f .xk /; f .x` // D d.xk ; x` / shows the sequence x D fxk gk is a Cauchy
Q The inequalities
sequence in X also. Thus, x 2 C and we may consider xP 2 X.
Q x; sP k / d.f
Q .xk /; sP k / < 1 ;
d.P
k
demonstrate that the Cauchy sequence fPsk gk converges in XQ to xP.
t
u
Definition 2.54. Let X be a topological space.

1. A subset G X is a G -set if there is a countable family fUk gk2N of open sets
Uk X such that
GD
Uk :
k2N
2. A subset F X is an F -set if there is a countable family fFk gk2N of closed sets

Fk X such that
62
FD
Fk :
k2N
The Baire Category Theorem below is a remarkable result illustrating the nature
of complete metric spaces, as well as being widely relevant in applications of
topology to analysis.
Theorem 2.55 (Baire Category Theorem). If fUk gk2N is a sequence of open sets
in a complete metric space .X; d/, and if each Uk is dense in X, then the G -set
\
Uk
k2N
is also dense in X.
Proof.\Choose x0 2 X and let " > 0. We aim to prove that B" .X0 / \ G 6D ;, where
GD
Uk .
k2N
By hypothesis, U1 is open and dense in V. Thus, there is a 1 2 .0; 1/ and an

element x1 2 B" .x0 / \ U1 such that B1 .x1 / B" .x0 / \ U1 . Likewise, U2 is open and
dense in X, and so there is a 2 2 .0; 12 / and an element x2 2 B1 .x1 / \ U2 such that
B2 .x2 / B1 .x1 / \ U2 .
It is clear that this process may be continued by induction to obtain sequences
fxn gn2N in X and fn gn2N in R such that

1
n 2 0; n1
2
and
Bn .xn / Bn1 .xn1 / \ Un :
By construction,
Bk .xk / Bn .xn / Bn .xn / B" .x0 / ;
8k > n:
Fix n 2 N and let k; m > n. By the inclusions above,

xk ; xm 2 Bn .xn / Bn .xn / :
1
, which shows that fxn gn2N is a Cauchy sequence. Since
Therefore, d.vk ; vm / < 2n2
X is a complete metric space, there is a limit x 2 X to this sequence. Choose any
n 2 N. If k > n, then
xk 2 Bn .xn / .Bn1 .xn1 / \ Un / Un :

Hence, x 2 Bn .xn / Un and x 2 Bn .xn / B" .x0 /, and so x 2 G \ B" .x0 /.
t
u
Definition 2.56. A subset F X is nowhere dense in a topological space X if the

interior of the closure F of F is the empty set.
2.5 Connected Spaces
63
Corollary 2.57. In a complete metric space,

1. the intersection of a countable family of dense G -sets is a dense G -set, and
2. the union of a countable family of nowhere dense F sets is a nowhere dense F
set.
t
u

Definition 2.58. A separation of a topological space X is a pair .U; V/ of open
nonempty subsets U; V X such that U \ V D ; and U [ V D X. If X admits a
separation, then X is said to be a disconnected space; and if no separation of X
exists, then X is called a connected space.
The standard geometric model for the continuum R is that of an unbroken
continuous line. The following example provides the topological justification for
this model.
Example 2.59. Every closed interval a; b of real numbers is connected.
Proof. Assume a; b is not connected. Thus, there is a separation .U; V/ of a; b.
Without loss of generality, assume that a 2 U and define L D fx 2 R j a; x Ug.
Some observations concerning L are: (i) a 2 L, (ii) b 62 L (because U c D V is a
nonempty subset of a; b), (iii) x b for all x 2 L, and (iv) if x 2 L and y 2 a; x,
then y 2 L.
As L is a nonempty set bounded above by b, the supremum of L exists, which
we denote by c. Thus, for every " > 0 there is a z 2 L such that c " < z, which
shows that W \ L 6D ; for every neighbourhood W of c. That is, c 2 L. From L U
we obtain L U. Further, U and U c D V are open, and so U is also closed, whence
U D U. Therefore, c 2 U.
We now show that a < c < b. First of all, because U is open, there is some " > 0
for which a; a C "/ U, which implies that a C " c. Likewise, if it were true that
c D b, then U would contain a; b " for all 0 < " < .b a/, and so U would contain
a; b/, thereby forcing the nonempty open set V to be V D fbg (which is not open in
a; b). Thus, c < b.
Now because U is open and c 2 U, there exists " > 0 such that .c "; c C "/ U.
As c D sup L, there is an element x 2 L such that x 2 .c "; c U. Select z 2
.c; c C "/ U. Thus, a; x U and .x; z .c "; c C "/ U imply that a; z D
a; x [ .x; z U. But c < z sup L D c is a contradiction. Hence, it must be that
a; b is connected.
t
u
To construct additional examples of connected spaces, the following result is
useful.
64
Proposition 2.60. \
If fX g2 is a family
[ of connected subspaces of a topological
space X, such that
X ;, then
X is connected.
Proof. Without loss of generality assume that X D
X , and suppose that U and
V are open subsets of X such that U \ V D ; and U [ V D X. Let U D U \ X

and V D V \ X so that \
U and V are open in X and satisfy U \ V D ; and
U [ V D X . Select x 2
X . Thus, x is an element of exactly one of U or V, say
U. Because x 2 X for every 2 , we deduce that x 2 U for all . Hence U ;

implies V D ; because X is connected. Thus, U D X and V D ;, which implies
that X does not admit a separation.
t
u
Example 2.61. R is connected.
[
\
Proof. R D
k; k and
k; k D 1; 1 6D ;. Therefore, Example 2.59 and
k2N
k2N
Proposition 2.60 imply that R is connected.
t
u
Concerning mappings of connected spaces, we have:

Theorem 2.62 (Intermediate Value Theorem). If X is a connected topological
space and if f W X ! R is a continuous function, then for any x1 ; x2 2 X and real
number r between f .x1 / and f .x2 / there is an x 2 X with f .x/ D r.
Proof. Without loss of generality we may assume f .x1 / f .x2 /. Choose r 2
f .x1 /; f .x2 /. If, contrary to what we aim to prove, r 62 f .X/, then f .X/ .1; r/ [
.r; 1/. The open sets U D f 1 .1; r/ and V D f 1 .r; 1/ satisfy U \ V D ;,
U [ V D X, x1 2 U, and x2 2 V. Thus, .U; V/ is a separation of X, in contradiction
to the connectivity of X.
t
u
In a similar vein:
Proposition 2.63. If f W X ! Y is a continuous function and if X is connected, then
f .X/ is a connected subspace of Y.
t
u
The connectivity of the closed interval 0; 1 gives a useful way to determine

whether a space is connected.
Proposition 2.64 (A Connectivity Criterion). If X a topological space X has the
property that for every pair x0 ; x1 2 X there is a continuous function f W 0; 1 ! X
such that f .0/ D x0 and f .1/ D x1 , then X is a connected space.
Proof. Assume, contrary to what we aim to prove, that X is disconnected. Thus,
there is a separation .U; V/ of X. Select x0 2 U and x1 2 V, and let f W 0; 1 ! X be
a continuous function such that f .0/ D x0 and f .1/ D x1 . Hence, f 1 .U/; f 1 .V/
is a separation of 0; 1, in contradiction to the connectivity of 0; 1. Hence, it must
be that X is connected.
t
u
65
Proposition 2.64 does not characterise connected spaces in that there exist
connected topological spaces X for which the hypothesis of Proposition 2.64 is not
satisfied. Therefore, spaces that satisfy this connectivity criterion are called path
connected.
Definition 2.65. A topological space X is path connected if X satisfies the hypothesis of Proposition 2.64.
Example 2.66. If n 2, then Rn n f0g is a path-connected subspace of Rn .
Proof. Let x0 ; x1 2 Rn . If, considered as vectors, x0 and x1 are linearly independent,
then no nontrivial linear combination of x0 and x1 yields the zero vector. Hence,
the map f W 0; 1 ! Rn defined by f .t/ D .1 t/x0 C tx1 is continuous, satisfies
f .0/ D x0 and f .1/ D x1 , and has range contained in Rn n f0g. On the other hand,
if x0 and x1 are linearly dependentthat is, x1 D x0 for some 2 Rthen there
is a nonzero z 2 Rn such that z is linearly independent of x0 and, thus, of x1 . So,
by what has already been proved, there are continuous g; h W 0; 1 ! Rn n f0g such
that g.0/ D x0 , g.1/ D z, h.0/ D z, and h.1/ D x1 . Now let f W 0; 1 ! Rn n f0g be
given by f .t/ D g.2t/, if t 2 0; 1=2, and by f .t/ D h.2t 1/, if t 2 1=2; 1. Then f
is continuous (Exercise 1.127) and satisfies f .0/ D x0 and f .1/ D x1 .
t
u
At the other end of the connectivity spectrum lies the notion of a totally
disconnected space.
Definition 2.67. A topological space X is totally disconnected if for every pair of
distinct x; y 2 X there exist subsets U; V X such that
1. x 2 U and y 2 V,
2. U and V are both open and closed in X, and
3. U \ V D ;.
To better understand the definition above, first note that a space A is disconnected
if A can be written as A D U [V, for some nonempty open disjoint subsets U; V A;
in this case, U and V are necessarily both closed and open. Therefore, if X is totally
disconnected, and if x and y are distinct elements of X, then there is no connected
subset A of X that contains both x and y.
Example 2.68. The Cantor set is totally disconnected.
Proof. The Cantor ternary function is monotone increasing and is a continuous
surjection of the Cantor set C onto 0; 1 (Proposition 1.86). Suppose that x; y 2 C
are such that x < y. Recall that the open set 0; 1 n C is a countable union of pairwise
disjoint open intervals. Of these intervals, select one, say .a; b/, for which x a and
b y. Now select r; s 2 .a; b/ such that r < s, and consider the open subsets U and
V of C given by U D 1 .0; r// and V D 1 ..s; 1/. Because is monotone
increasing, we deduce that x 2 U, y 2 V, and U \ V D ;; and because is constant
on every open interval in 0; 1 n C (Corollary 1.87), we observe that 1 .0; r// D
1 .0; r/ and 1 ..s; 1/ D 1 .s; 1/. Hence, U and V are also closed sets, by
the continuity of .
t
u
66
Compactification
2.6 Stone-Cech
Earlier, in Proposition 2.22, we noted that every non-compact, locally compact
Hausdorff space X can be embedded into a compact Hausdorff space XQ by adding
a single point (the so-called point at infinity). The goal of the present section is
to show that another embedding of X into a compact Hausdorff space is possible,
and this embedding has the advantage that every bounded continuous real-valued
function on X extends to a continuous function on the larger space.
The first issue to address is setting the precise meaning of compactification of
a topological space.
Definition 2.69. A compactification of a topological space X is a pair .K; X;K /
consisting of
1. a compact space K, and
2. a continuous function X;K W X ! K such that
a. X;K is a homeomorphism of X and X;K .X/, and
b. X;K .X/ is dense in K.
By identifying X with its image X;K .X/ in K, one can view K as being a compact
space that contains X as a dense subspace.
The definition of compactification does not ask, for example, that X (or X;K .X/)
be open in K. However, if X is locally compact and K is Hausdorff, then such will
be the case.
Proposition 2.70. If a locally compact space X is dense in a Hausdorff space Y,
then X is an open subset of Y.
Proof. As in Proposition 1.64, the closure in X of a subset A X will be denoted
by AX , and the closure in Y of a subset B Y will be denoted by BY .
Select x0 2 X. Because X is locally compact, there are an open set U in X and a
compact subset K of X such that x0 2 U K. Therefore, U X is a compact subset of
K and, hence, is compact in X. If one takes an open cover fV g2 in Y of the set
U X , then fX \ V g2 is an open cover of U X in X. Hence, by compactness, finitely
many X \ V cover U X , and so finitely many V cover U X , which proves that U X is
a compact subset of Y. Because Y is Hausdorff, compact subsets of Y are closed in
Y (Proposition 2.8); therefore, U X is a closed subset of Y. The set U X contains U
and is a closed subset of Y; thus, U X U Y . On the other hand, U X U Y , by virtue
of X Y. Hence, U X D U Y .
The set U has the form U D X \ V, for some open set V in Y. Therefore, U Y D
X \ V Y D V Y (because X is dense in Y). Moreover, U Y D U X X implies that V Y
X. Therefore, V X and, hence, U D X \ V D V. This proves that, for every x0 2 X,
there is an open set V in Y such that x0 2 V X. In other words, X is an open subset
of Y.
t
u

2.6 Stone-Cech
Compactification
67
The following theorem asserts the existence of a compactification K of a locally

compact Hausdorff space X in which every bounded continuous function X ! C
extends to a continuous function K ! C.
Theorem 2.71 (Stone-Cech

Compactification: Existence). If X is a locally
compact Hausdorff space, then there exists a compactification .X; X / of X such
that
1. X is Hausdorff, and
2. for every bounded continuous function f W X ! C there exists a unique continuous
fQ W X ! C such that fQ X .x/ D f .x/, for all x 2 X.
Proof. The first part of the proof is devoted to the construction of the compact
Hausdorff space .X/. To this end, let Cb .X/ denote the set of all continuous
functions f W X ! C such that sup jf .x/j < 1. For each such f 2 Cb .X/, let
x2X
kf k D sup jf .x/j and consider the compact Hausdorff subspace Kf of C given by

x2X
Kf D f 2 C j jj kf kg. Endow the product space

Y
ZD
Kf
f 2Cb .X/
with the product topology. By Tychonoffs Theorem, Z is a compact Hausdorff

space.
Define X;Z W X ! Z by
X;Z .x/ D
f .x/;
f 2Cb .X/
for x 2 X. Each component map f W X ! Kf is continuous, and so the map X;Z is also
continuous (Proposition 1.92). Furthermore, the local compactness of X implies (by
Exercise 2.92) that there exists an element f 2 Cb .X/ such that f .x/ 6D f .y/, for any
two distinct x; y 2 X. Hence, the function X;Z is necessarily injective.
To prove that X;Z is a homeomorphism between X and its range X;Z .X/ in
Z, it is sufficient to prove that X;Z maps open sets in X to open sets in X;Z .X/.
Therefore, suppose that U X is an open set, select z0 2 X;Z .X/, and let x0 2 U
be the unique element for which z0 D X;Z .x0 /. Consider the function g0 W fx0 g ! C
given by g0 .x0 / D 1. The point set fx0 g is compact and is contained in the open set
U. Hence, by the Tietze Extension Theorem (Theorem 2.43), there is a bounded
continuous function g W X ! C such that g.x0 / D 1 and g.U c / D f0g. The canonical
projection pg W Z ! Kg C is continuous (Proposition 1.92); therefore, the set
V D p1
g .C n f0g/ is open in Z. Let W D V \ X;Z .X/, which is open in X;Z .X/
and contains z0 (because pg .z0 / 6D 0). If z 2 W, then there is a unique x 2 X with
z D X;Z .x/. Thus, as 0 6D pg .x/ D g.x/, we deduce that x 62 U c ; hence, x 2 U.
Therefore, z D X;Z .x/ 2 X;Z .U/. In other words, for each z0 2 X;Z .U/ there is an
open set W in X;Z .X/ such that z0 2 W X;Z .U/. Thus, X;Z .U/ is open in X;Z .X/,
which completes the proof that X and X;Z .X/ are homeomorphic.
68
Denote the closure X;Z .X/ of X;Z .X/ in Z by X, and let X denote X;Z . Observe
that X is a compact Hausdorff space and contains X .X/ as a dense subspace. Thus,
.X; X / is a compactification of X.
Suppose now that f W X ! C is a bounded continuous function, and define fQ W
.X/ ! C by fQ D pf j.X/ . Note that, if x 2 X, then fQ X .x/ D f .x/. To show the
uniqueness of the extension fQ , suppose that fQ and fQQ are two extensions of f . If z 2
Z, then there is a net fX .x /g convergent to z. Thus, fQ X .z/ D lim f .x / and
fQQ X .z/ D lim f .x /. Because convergent nets have unique limits in the Hausdorff
space C, we deduce that fQ D fQQ .
t
u
For the moment, the following definition of the Stone-Cech

compactification of
a locally compact space will suffice.
Definition 2.72. The Stone-Cech

compactification of a locally compact Hausdorff
space X is the compactification .X; X / constructed in the proof of Theorem 2.71.
It is natural to wonder about the conclusion of Theorem 2.71 in the case where
the locally compact Hausdorff space X is compact.
Proposition 2.73. If X is a compact Hausdorff space, then X ' X.
Proof. By construction, X is the closure of X .X/ in the compact product space Z.
Furthermore, X .X/ and X are homeomorphic with respect to the subspace topology
of X .X/. Because X .X/ is a compact subspace of a Hausdorff space, the set X .X/
is closed in Z (Proposition 2.9). Hence, X D X .X/ D X .X/ ' X.
t
u
The following proposition captures a very elegant feature of the Stone-Cech

compactification.
Proposition 2.74. If X and Y are locally compact Hausdorff spaces, and if h W X !
Y is a continuous function, then there is a continuous function H W X ! Y such
that the following commutative diagram holds:
X
iX
bX
Y
iY
bY .
That is, H X D Y h.
Proof. For a fixed g 2 Cb .Y/, the function g h W X ! C is bounded and continuous.
Therefore, by Theorem 2.71, there is a continuous function gO W X ! C such that
gO .X .x// D g .h.x//, for all x 2 X. Let ZY be the compact product space constructed in
the proof of Theorem 2.71 and which contains Y as a subspace. Define a function
H W X ! ZY by
H.!/ D .Og.!//g2Cb .Y/ ; for all ! 2 X:

2.6 Stone-Cech
Compactification
69
By definition of H, it is clear that H X D Y h. To show that H is continuous,

observe that pg H D gO , where pg denotes the canonical projection ZY ! C of ZY
onto its g-th coordinate. Because gO is continuous, Proposition 1.92 implies that H is
continuous.
To this point, we know only that H maps X into ZY . However,

H.X/ D H X .X/ H .X .X// D Y h.X/ Y h.Y/ D Y
shows that the range of H is indeed within Y. (Note that the first of the inclusions
above is a consequence of the continuity of H.)
t
u
A variant of Proposition 2.74 is the following result, with essentially the same
proof.
Proposition 2.75. Assume that .K; X;K / is a Hausdorff compactification of a
locally compact Hausdorff space X with the property that for every bounded
continuous function f W X ! C there exists a unique continuous fQ W K ! C such
that fQ X;K D f . If L is an arbitrary compact Hausdorff space and if h W X ! L is
a continuous function, then there exists a continuous function H W K ! L such that
H X;K D h.
Y
Kg , each g 2
Proof. Because L is a subspace of the product space ZL D
g2Cb .L/
Cb .L/ yields a function g h W X ! C that is bounded and continuous and which, by

hypothesis, extends to a function gO W K ! C. Following the proof of Proposition 2.74
verbatim leads to a continuous map H0 W K ! L ZL . Proposition 2.73 shows that
L D L .L/, where L W L ! ZL is a homeomorphism of L and L .L/. Set H D 1
L H0
to get the map H W K ! L with the desired properties.
t
u
One final note, before continuing further, is that the method of proof employed
in Theorem 2.71 and the tangential results above (Propositions 2.74 and 2.75) also
yield the following noteworthy fact.
Proposition 2.76. If X is a locally compact Hausdorff space, then there exists a set
such that X is homeomorphic to a subspace of the compact Hausdorff hypercube
.0; 1 ; T /.
Proof. Exercise 2.112
t
u
It is preferable to define the Stone-Cech

compactification by a property rather
than by a construction. To do so, it is necessary to understand what other compactifications of X possess properties (1) and (2) of Theorem 2.71 and to understand how
these relate to X.
Theorem 2.77 (Stone-Cech

Compactification: Uniqueness). Assume that X is a
locally compact Hausdorff space and that .K; X;K / is a compactification of X in
which K is Hausdorff and for every bounded continuous function f W X ! C there
exists a unique continuous fQ W K ! C such that fQ X;K .x/ D f .x/, for all x 2 X. Then
there exists a homeomorphism W X ! K such that X D X;K . That is,
70
idX
i X,K
iX
bX
K.
Proof. If Q is a set, then idQ shall denote the (identity) function Q ! Q in which
idQ .q/ D q, for every q 2 Q.
We apply Proposition 2.75 twice, as both compactifications K and X satisfy
the hypothesis of Proposition 2.75. In the first instance, Proposition 2.75 yields
a continuous map IX W K ! X such that IX X;K D X . In the second instance,
Proposition 2.75 yields a continuous map IX;K W X ! K such that IX;K X D X;K .
Hence,
IX;K IX X;K D IX;K X D X;K and IX IX;K X D IX X;K D X :
Therefore, because X;K .X/ is dense in K and X .X/ is dense in X, the functional
equations above lead to
IX;K IX D idK and IX IX;K D idX :
Hence, IX;K is a bijection with continuous inverse IX . Therefore, with D IX;K we
have a homeomorphism W X ! K such that X D X;K .
t
u
In light of Theorem 2.77 and Proposition 2.70, the Stone-Cech

compactification
of a locally compact Hausdorff space X is any compact Hausdorff space K that
contains X as a dense open subspace and has the property that every bounded
continuous function f W X ! C extends to a unique continuous fQ W K ! C.
The final general property of the Stone-Cech

compactification to be remarked
upon here is given by the following proposition, which asserts that X is the largest
of all possible compactifications of a given locally compact Hausdorff space X.
Proposition 2.78. If X is a locally compact Hausdorff space and if K is a compact
Hausdorff space that contains X as a dense subspace, then K is a quotient of X.
t
u
The complete determination of X using existing knowledge about X is very

difficult, if not impossible. Indeed, even if X is well understood and has good
topological properties, X can be tremendously different, as the following result
indicates.
Proposition 2.79. If X is an infinite discrete space, then X is nonmetrisable and
totally disconnected.
Proof. Because X is an infinite set with the discrete topology, X is a locally compact
Hausdorff space; moreover, without loss of generality, we may assume that X is a
dense open subspace of X.
Assume, contrary to what we aim to prove, that X is metrisable. Let fxn gn2N
be any sequence in X consisting of infinitely many distinct elements. As X is
Problems
71
metrisable, Theorem 2.19 asserts that the sequence fxn gn2N admits a convergent
subsequence fxnj gj2N consisting of distinct points and with limit z 2 X. Because
X is discrete, the sets A D fxn2k j k 2 Ng and B D fxn2kC1 j k 2 Ng are closed, and
so the function f0 W A [ B ! 0; 1 in which f0 .x/ D 0, if x 2 A, and f0 .x/ D 1,
if x 2 B, has a continuous extension f W X ! 0; 1 by Theorem 2.43. Further, f
has a continuous extension f W X ! R. However, f .z/ D limk f .xn2k / D 0 and
f .z/ D limk f .xn2kC1 / D 1, which because of z D limj xnj is a contradiction of the
continuity of f . Therefore, it cannot be true that fxn gn2N admits a convergent
subsequence, which by Theorem 2.19 implies that X is not metrisable.
Select any nonempty proper subset Y of X. Because X has the discrete topology,
both Y and X n Y are open in X. Therefore, the function f W X ! 0; 1 defined by
f .x/ D 1 if x 2 Y and f .x/ D 0 if x 62 Y is continuous. If Y \ X n Y is nonempty,
where A denotes the closure in X of a subset A X, then there exist z 2 X

and nets fy g2 and fx g2 in Y and X n Y, respectively, such that z D lim y D
lim x . Therefore, the continuous extension f of f to X has the property that
f .z/ D lim f .y / D 1 and f .z/ D lim f .x / D 0, which is a contradiction. Hence:
disjoint subsets of X have disjoint closures in X.
Select any two distinct z1 ; z2 2 X. Because X is Hausdorff, there are open sets
U1 ; U2 X with each xj 2 Uj and U1 \ U2 D ;. Because X is open in X, both
X \ U1 and X \ U2 are open sets, and thus their closures X \ U1 and X \ U2 are
open in X, by Exercise 2.116. Therefore, all that remains is to show that each
zj 2 X \ Uj . To this end, assume that z 2 X and U is any open set in X that
contains z. Because X is dense in X, for every open set W containing z there is an
element xW 2 X such that xW 2 U \ W. Hence, fxW gW is a net in X \ U converging
to z, and so z 2 X \ U .
t
u
Corollary 2.80. N is separable, but not second countable.
Proof. The countable set N is dense in N, and so N is separable. If N were
second countable, then the compact Hausdorff space N would be metrisable
(Theorem 2.48), in contradiction to Proposition 2.79.
t
u
Observe that N also furnishes us with an example that shows Theorem 2.19
need not hold beyond metric spaces. Specifically, using the proof of Proposition 2.79, no sequence in N admits a convergent subsequence in N.
Problems
2.81. Prove that the following statements are equivalent for a topological
space X:
1. X\is compact;
2.
F ; for every family fF g2 of closed sets F X with the finite
2
intersection property.
72
space X:
1. X\is compact;
2.
E ; for every family fE g2 of sets E X with the finite intersection
2
property.
2.83. Prove that if F is a finite set in a Hausdorff space X, then F is a closed set.
2.84. Prove that if Y X in a Hausdorff space X, and that if x 2 Y and U X is a
neighbourhood of x, then U \ Y is an infinite set.
2.85. Suppose that .X; d/ is a compact metric space and that U1 ; : : : ; Un are open
n
[
Uj D X.
subsets of X such that
jD1
1X
1. Prove that the function f W X ! R defined by f .x/ D
dist; .x; Ujc / is
n jD1
n
continuous and satisfies f .x/ > 0 for every x 2 X.

2. Prove that if D infff .x/ j x 2 Xg, then > 0.
3. Prove that if S X is a nonempty set for which supfd.s1 ; s2 / j s1 ; s2 2 Sg < , then
S Uj for some j 2 f1; : : : ; ng.
2.86. In the Euclidean metric space .Rn ; d2 /, prove that a subset K Rn is compact
if and only if K is closed and sup d2 .x; 0/ < 1.
x2K
2.87. Prove that the n-spheres Sn are compact.

2.88. Prove that if x and x0 are limits of a convergent sequence fxn gn2N in a
Hausdorff space X, then x0 D x.
2.89. Recall from Proposition 1.57 that the topological space .RN ; T / is metrisable with respect to the metric

jxn yn j 1
:
;
..xn /n ; .yn /n / D sup min
n
n
n
Show that S D fx 2 RN j .x; 0/ D 1g is closed by not compact.
2.90. Determine whether .RN ; T / is a locally compact space.
2.91. Prove that in a locally compact Hausdorff space X that if K U, where K is
compact and U is open, then there exists an open subset W X such that K W U
and W is compact.
2.92. Assume that x and y are distinct points in a locally compact Hausdorff space
X. Prove that there exists a bounded continuous function f W X ! C such that
f .x/ 6D f .y/.
Problems
73
2.93. Prove that the topology T on XQ in Proposition 2.22 is Hausdorff.

Q and the circle S1 are homeomorphic, where R
Q is the one-point
2.94. Prove that R
compactification of R.
2.95. Assume that Y is a nonempty subset of a topological space X. Suppose that
fy g2 is a net in Y with limit x 2 X. Prove that x 2 Y.
2.96. If X and Y are topological spaces, then prove that the following statements
are equivalent for a function f W X ! Y:
1. f is continuous;
2. for every convergent net fx g2 in X with limit x 2 X, ff .x /g2 is a convergent
net in Y with limit f .x/.
space X:
1. X is Hausdorff;
2. every convergent net fx g2 in X has a unique limit point.
2.98. Prove that a topological space X is normal if and only if (i) fxg is a closed set,
for all x 2 X, and (ii) for every closed set F and open set U for which F U there is
an open set V with F V V U.
2.99. Prove that every separable metric space is second countable.
2.100. Prove that the metric spaces .Rn ; d1 / and Rn ; d1 / are complete.
2.101. Prove that if .X; d/ is a metric space and if fxk gk and fyk gk are two Cauchy
sequences in X, then fd.xk ; yk /gk is a Cauchy sequence in R.
2.102. Suppose that .X1 ; d1 / and .X2 ; d2 / are metric spaces and that f W X1 ! X2 is
an isometry, which is to say that d2 .f .x/; f .y// D d1 .x; y/ for all x; y 2 X1 . Prove that
f is continuous and injective.
2.103. Prove that in a separable metric space, every open covering of an open set
admits a countable subcovering.
2.104. Prove that in a complete metric space the intersection of a countable family
of dense G -sets is a dense G -set and that the union of a countable family of
nowhere dense F sets is a nowhere dense F set.
2.105. Prove or find a counterexample to the following assertion: if X is a compact
metric space, then X is a complete metric space.
2.106. Prove that if Y is a connected subspace of a topological space X, then Y is
connected.
2.107. Prove that if f W X ! Y is a continuous function and if X is connected, then
f .X/ is a connected subspace of Y.
2.108. Prove that the n-spheres Sn are path connected.
74
2.109. Prove that every connected open set U in Rn is path connected.

2.110. Prove that every countable metric space is totally disconnected.
2.111. Show that the bounded continuous function f .x/ D sin.1=x/ on the locally
compact space X D .0; 1/ does not extend to a continuous function fQ on the compact
set X D 0; 1.
2.112. Prove that if X is a locally compact Hausdorff space, then there exists a set
such that X is homeomorphic to a subspace of the compact Hausdorff hypercube
.0; 1 ; T /.
2.113. Prove that if X is a locally compact Hausdorff space and if K is a compact
Hausdorff space that contains X as a dense subspace, then K is a quotient of X.
2.114. Show that the one-point compactification of N is not homeomorphic to the
Stone-Cech
compactification N of N.
2.115. Assume that X is a non-compact, locally compact space X. Prove that the
Stone-Cech
compactification X of X is not metrisable.
2.116. A topological space is extremely disconnected if the closure of every open
set is open. Show that if X is an infinite discrete space, then X is extremely
disconnected.
Part II
Measure Theory and Integration
Chapter 3
Measure Theory
If topology derives its inspiration from the qualitative features of geometry, then the
subject of the present chapter, measure theory, may be thought to have its origins
in the quantitative concepts of length, area, and volume. However, a careful theory
of area, for example, turns out to be much more delicate than one might expect
initially, as any given set may possess an irregular feature, such as having a jagged
boundary or being dispersed across many subsets. Even in the setting of the real
line, if one has a set E of real numbers, then in what sense can the length of the set
E be defined and computed? Furthermore, to what extent can we expect the length
(or area, volume) of a union A [ B of disjoint sets A and B to be the sum of the
individual lengths (or areas, volumes) of A and B?
This present chapter is devoted to measure theory, which, among other things,
entails a rigorous treatment of length, area, and volume. However, as with the subject
of topology, the context and results of measure theory reach well beyond these basic
geometric quantities.
3.1 Measurable Spaces and Functions

Definition 3.1. If X is a set, then a -algebra on X is a collection of subsets of
X with the following properties:
1. X 2 ;
2. Ec 2 for every E 2 ; and
3. for every countable family fEk gk2N of sets Ek 2 ,
[
Ek 2 :
k2N

DOI 10.1007/978-3-319-45633-1_3
77
78
3 Measure Theory
The pair .X; / is called a measurable space, and the elements E of are called
measurable sets.
The smallest and largest -algebras on a set X are, respectively, D f;; Xg and
D P.X/, the power set P.X/ of X. The following definition, while abstract in
essence, allows for the determination of more interesting, intermediate examples of
-algebras.
Definition 3.2. If S is any collection of subsets of X, then the intersection of all
-algebras on X that contain S is called the -algebra generated by S .
It is elementary to verify that the -algebra generated by a collection of S of
subsets of X is a -algebra in the sense of Definition 3.1.
Definition 3.3. If .X; T / is a topological space, then the -algebra generated by T
is called the -algebra of Borel sets of X.
Let us now consider functions of interest for measure theory.
Definition 3.4. If .X; / is a measurable space, then a function f W X ! R is
measurable if f 1 .U/ 2 , for every open set U R.
Proposition 3.5. If .X; / is a measurable space, then the following statements are
equivalent for a function f W X ! R:
1.
2.
3.
4.
f 1 . .; 1/ / 2 for all 2 R;
f 1 . ; 1/ / 2 for all 2 R;
f 1 . .1; / / 2 for all 2 R;
f 1 . .1; / 2 for all 2 R.
Proof. To begin, observe that (2) follows from (1), because

!

\
\
1
1
1
1
f .; 1// D f
. ; 1/ D
f 1 . ; 1/ 2 :
k
k
k2N
k2N
Statement (3) follows easily from (2), since
f 1 ..1; // D f 1 .; 1//c 2 :
Next, we see that (3) implies (4), because
f
1
..1; / D f
1
!

\
1
1
1
.1; C / 2 :
.1; C / D
f
k
k
k2N
k2N
\
Statement (4) implies (1), because

f 1 ..; 1// D f 1 ..1; /c 2 ;
which completes the proof.
t
u
An additional equivalent condition for the measurability of a function is set aside,

for future reference, as the following result.
79
Proposition 3.6 (Criterion for Measurability). If .X; / is a measurable space,

then a function f W X ! R is measurable if and only if f 1 . .; 1/ / 2 , for all
2 R.
Proof. By definition of measurable function, f 1 . .; 1/ / 2 for all 2 R
because each .; 1/ is open in R.
Conversely, assume that f 1 . .; 1/ / 2 , for all 2 R. Let U R be an open
set. By Cantors Lemma (Proposition [
1.30), there is a family of[
pairwise disjoint
1
open intervals fJk gk2N such that U D
Jk . Because f .U/ D
f 1 .Jk / and
k
is closed under countable unions, it is enough to prove that f 1 .J/ 2 for every
open interval J. For open intervals of the form .; 1/ and .1; /, this is handled
by Proposition 3.5. If one has an open interval of the form J D .; /, then f 1 .J/
is given by f 1 .J/ D f 1 ..1; /c \ f 1 .; 1//c , which by Proposition 3.5 is the
intersection of two sets in .
t
u
Proposition 3.7. If .X; / is a measurable space, if f ; g W X ! R are measurable
functions, and if 2 R, then f C g, f , jf j, and fg are measurable functions. If, in
addition, g.x/ 6D 0 for every x 2 X, then f =g is measurable function.
Proof. The equivalent criteria for measurability of Proposition 3.5 will be used in
each case. We begin with a proof that f C g is measurable.
Fix 2 R and consider the set S D fx 2 X j f .x/ C g.x/ > g. Because f and g
are measurable, for each q 2 Q we have
fx 2 X j f .x/ > qg 2
and
fx 2 X j q > g.x/g 2 :
Hence, as is closed under intersections and countable unions,

[
.fx 2 X j f .x/ > qg \ fx 2 X j q > g.x/g/ 2 :
(3.1)
q2Q
Let G denote the set in (3.1); we shall prove that S D G. If y 2 S , then f .y/ >
g.y/. In fact, by the density of Q in R, there is a rational number qy 2 Q such
that f .y/ > qy > g.y/. Thus,
y 2 fx 2 X j f .x/ > qy g \ fx 2 X j qy > g.x/g ;
which shows that S G. Conversely, if y 2 G, then there is a rational qy 2 Q such
that y 2 fx 2 X j f .x/ > qy g \ fx 2 X j qy > g.x/g. Thus, f .y/ > qy > g.y/
implies that f .y/ C g.y/ > , whence y 2 S and, consequently, G S . This proves
that f C g is measurable.
The proof that f is measurable is clear, and we move to the proof that jf j is
measurable. Note that if 2 R, then jf j1 ..; 1// D X if < 0, and
jf j1 ..; 1// D f 1 ..; 1// [ f 1 ..1; // ; if 0 :
In either case, jf j1 ..; 1// 2 , which proves that jf j is measurable.
80
3 Measure Theory
To prove that the product fg is measurable, first assume that h W X ! R is a

2
measurable function and consider h2 . If p2 R, then
fx 2 X j h.x/ > g D X if < 0,
2
1
otherwise fx 2 X j h.x/ > g D jhj . ; 1/ . In either case, the sets belong to
. This proves that the square of a measurable function is measurable. To conclude
that fg is measurable, express fg as
fg D

1
.f C g/2 .f g/2 :
4
(3.2)
As the sums, squares, and scalar multiples of measurable functions are measurable,
equation (3.2) demonstrates that fg is measurable.
If g.x/ 6D 0 for every x 2 X, then 1=g is measurable (Exercise 3.79), which implies
that the function f =g D f .1=g/ is measurable.
t
u
Using the algebraic features exhibited in Proposition 3.7, one deduces that the
following functions are measurable as well.
Corollary 3.8. Suppose that f ; g W X ! R are measurable functions.
1. If max.f ; g/ is the function whose value at each x 2 X is the maximum of f .x/ and
g.x/, and if min.f ; g/ is the function whose value at each x 2 X is the minimum of
f .x/ and g.x/, then max.f ; g/ and min.f ; g/ are measurable.
2. f C is the function max.f ; 0/ and f is the function min.f ; 0/, then f C and f
are measurable.
Proof. By Proposition 3.7, the sum, difference, and absolute value of measurable
functions are measurable. Therefore, the formulae
max.f ; g/ D 1=2.f C g C jf gj/;
min.f ; g/ D 1=2.f C g jf gj/;
f C D 1=2.jf j C f /; and
f D 1=2.jf j f /
imply the asserted conclusions.
t
u
The purpose of the following result is to use sequences of measurable functions

to determine new measurable functions.
Proposition 3.9. Suppose that fk W X ! R is a measurable function for each k 2 N. If
S
LS
I
LI
L
D fx 2 X j supk fk .x/ existsg;

D fx 2 X j lim supk fk .x/ existsg;
D fx 2 X j infk fk .x/ existsg;
D fx 2 X j lim infk fk .x/ existsg; and
D fx 2 X j limk fk .x/ existsg;
81
then each of the sets S, LS, I, LI, and L is measurable. Moreover,

1.
2.
3.
4.
5.
supk fk is a measurable function on S,

lim supk fk is a measurable function on LS,
infk fk is a measurable function on I,
lim infk fk is a measurable function on LI, and
lim fk is a measurable function on L.
Proof. The set fk1 ..1; q// is measurable for every k 2 N and q 2 Q; therefore,
so is
[\
fk1 ..1; q// D S :
q2Q k2N
Consider now the function supk fk defined on the (measurable) set S with values in
R. For every 2 R,
[
fx 2 S j sup fk .x/ > g D
fx 2 S j fk .x/ > g 2 .S/ :
k
k2N
Hence, supk fk is measurable as a function S ! R.

The proofs that I is a measurable set and that infk fk is a measurable function
I ! R are handled in a similar fashion. For example, in this case, I is given by
ID
[\
fk1 ..q; 1// :
q2Q k2N
For each k 2 N consider the measurable function gk W S ! N defined by

gk .x/ D sup fn .x/ ;
x 2 S:
nk
For every x 2 LS, lim supk fk is precisely infk gk . Moreover, by the discussion of the
previous paragraph, inf gk is a measurable function on the (measurable) set
[\
g1
k ..q; 1// D LS :
q2Q k2N
Hence, as a function LS ! R, lim supk fk is measurable.

The proofs that LI is a measurable set and that lim infk fk is a measurable function
LI ! R are similarly handled.
Consider the measurable set E D LS \ LI and let h W E ! R be the function
h.x/ D lim sup fk .x/ lim inf fk .x/ :
k
Note that h is measurable and that

L D fx 2 X j lim sup fk .x/ D lim inf fk .x/g D h1 .f0g/ ;
k
82
3 Measure Theory
which is a measurable set because

h1 .f0g/ D En h1 .1; 0/ [ h1 .0; 1/ :
Finally, for every 2 R,
fx 2 L j lim fk .x/ > g D L \ fx 2 L j lim sup fk .x/ > g 2 .L/ :
k
Therefore, limk fk is measurable as a function from L to R.
t
u
Definition 3.10. If X is a set, then the characteristic function of a subset E X is

the function E W X ! R defined by
E .x/ D 1; if x 2 E; and E .x/ D 0; if x 62 E:
From the definition above, the following proposition is immediate:
Proposition 3.11. If .X; / is a measurable space and if E X, then the characteristic function E W X ! R is a measurable function if and only if E 2 .
Characteristic functions can be used to restrict or extend the domain of functions
(Exercise 3.82).
Definition 3.12. If .X; / is a measurable space, then a simple function is a
measurable function ' W X ! R such that ' assumes at most a finite number of
values in R.
Suppose that ' is a simple function on a measurable space .X; /. If '.X/ D
f1 ; : : : ; n g R, then let Ek D ' 1 .fk g/ (which is a measurable set, as ' is a
measurable function) so that
'D
n
X
k Ek
kD1
represents ' as a linear combination of the characteristic functions Ek .

Definition 3.13. A sequence ffk gk2N of real-valued functions fk on a set X is a
monotone increasing sequence if fk .x/ fkC1 .x/ for every k 2 N and every x 2 X.
The analysis of measurable functions depends, to a very large extent, on the
following approximation theorem.
Theorem 3.14 (Approximation of Nonnegative Measurable Functions). For
every nonnegative measurable function f on a measurable space .X; /, there is
a monotone increasing sequence f'k gk2N of nonnegative simple functions 'k on X
such that
lim 'k .x/ D f .x/;
k!1
for every x 2 X.
3.2 Measure Spaces
83
Proof. Let n W 0; n ! Z be the function whose value at t is the unique j 2 Z for

which t 2 2jn ; jC1
/, and define !n W R ! Q by !n .t/ D n .t/=2n , if t 2 0; n, and by
2n
!n .t/ D 0 if t 2 .n; 1/. The functions !n satisfy !n .t/ t for every t 2 0; 1/ and
!n .t/ !nC1 .t/ for all n 2 N and t 2 0; 1/. Now let 'n D !n f . Thus, f'n gn is a
monotone increasing sequence of nonnegative functions, each with finite range. For
each x 2 X there is some n 2 N for which f .x/ 2 0; n/.Thus, for every k > n,
f .x/ 'k .x/ <
1
;
2k
which proves that lim 'n .x/ D f .x/. All that remains is to verify that 'n is
n!1
measurable. To this end, select n 2 N and let

j1 j
1
1
En D f .n; 1// and Enj D f
; for 1 j 2n n :
;
2n 2n
These are measurable sets and
2 n
X
j1
n
'n D
jD1
2n
Enj C nEn :
Hence, 'n is a simple function.
t
u
By decomposing a real-valued function f into a difference its positive and

negative parts, namely f D f C f , where
fC D
jf j C f
2
and
f D
jf j f
;
2
(3.3)
we obtain the following approximation result for arbitrary measurable functions.

Corollary 3.15. If .X; / is a measurable space and if f W X ! R is a measurable
function, then there is a sequence f k gk2N of simple functions k W X ! R such that
lim
k!1
k .x/
D f .x/ ;
8x 2 X :
3.2 Measure Spaces

Before continuing further, the values 1 and C1 will be added to the arithmetic
system of R. Formally, the extended real number system are the elements of the set
denoted by 1; C1 and defined by f1g [ R [ fC1g. (Here, 1 and C1
are meant to denote the ends of the real axis.) The arithmetic of 1; C1 is
prescribed by the following laws:
1. r s and r C s are the usual product and sum in R, for all r; s 2 R;
2. 0 .1/ D 0 .C1/ D 0;
3. r .1/ D 1 and r .C1/ D C1, for all r 2 R with r > 0 and for r D C1;
84
3 Measure Theory
4. r .1/ D C1 and r .C1/ D 1, for all r 2 R with r < 0 and for r D 1;
5. r C .1/ D 1 and r C .C1/ D C1, for all r 2 R.
The sum of 1 and C1 is not defined in the extended real number system,
which is a small fact that will be of note in our study of signed measures in
Section 3.7.
Henceforth, 0; 1 denotes the subset of the extended real numbers given by
0; 1 D 0; 1/ [ fC1g:
The terminology below concerning families of sets will be used extensively,
beginning with the definition of measure in Definition 3.17.
Definition 3.16. A family fX g2 of subsets of a given set X is a family of pairwise
disjoint sets if X \ X D ; for all ; 2 such that 6D .
Definition 3.17. A measure on a measurable space .X; / is a function W !
0; C1 such that .;/ D 0 and
!
Ek D
k2N
.Ek / ;
(3.4)
k2N
for every sequence fEk gk2N of pairwise disjoint sets Ek 2 . Furthermore,

1. if .X/ < 1, then is said to be a finite measure, and
2. if .X/ D 1, then is said to be a probability measure.
The .X; ; / is called a measure space.
Measures are not easy to construct or determine in general, but there are some
very simple examples nevertheless.
Example 3.18. Consider the measurable space .X; / in which X is an uncountable infinite set and is the -algebra of all subsets E X that have the property
that E or Ec is countable (see Exercise 3.71). If W ! 0; C1 is defined by
.E/ D 0 if E is countable, and .E/ D 1 if Ec is countable;
then is a measure on .X; /.
Example 3.19 (Dirac Measures). If is a -algebra of subsets of X in which
fxg 2 for every x 2 X, then for each x 2 X the function x W ! 0; 1 given by
x .E/ D 1 if x 2 E; and x .E/ D 0 if x 62 E;
is a probability measure on .X; /. The measures x are called Dirac measures or
point mass measures.
3.2 Measure Spaces
85
Example 3.20 (Counting Measure). Consider the measurable space .N; P.N//,
where P.N/ is the power set of N. If W ! 0; C1 is the function defined by
.E/ D the cardinality of E ;
then is a measure on .N; P.N// and is called counting measure.
We turn now to some general properties of measures and measure spaces.
Proposition 3.21 (Monotonicity of Measure). Let .X; ; / denote a measure
space. Suppose that E; F 2 are such that E F. Then .E/ .F/. Furthermore,
if .F/ < 1, then .FnE/ D .F/ .E/.
Proof. Because E F, we may express F as F D E [ .Ec \ F/, which is a union
of disjoint sets E and Ec \ F, each of which belongs to . Hence, .F/ D .E/ C
.Ec \ F/ .E/.
t
u
Proposition 3.22 (Continuity of Measure). Let .X; ; / denote a measure space.
Suppose that fAk gk2N and fEk gk2N are sequences of sets Ek 2 .
1. If Ak AkC1 , for all k 2 N, then
!
Ak D lim .Ak / :
k!1
k2N
(3.5)
2. If Ek EkC1 , for all k 2 N, and if .E1 / < 1, then

!
Ek D lim .Ek / :
k!1
k2N
(3.6)
Proof. (1) Equation (3.5) plainly holds if .Ak / D 1 for at least one k; hence,
assume that .Ak / < 1 for all k 2 N. The sequence fAk gk2N is nested and
ascending, and so it is simple to produce from it a sequence of pairwise disjoint
sets Gk 2 by taking set differences: that is, define G1 to be A1 and let
Gk D Ak nAk1 ;
8k 2:
Observe that .Ak / < 1 implies that .Gk / D .Ak / .Ak1 /, by Proposik
[
Gn ,
tion 3.21. Furthermore, the sets Gk are pairwise disjoint. Because Ak D
nD1
we have
[
k2N
Ak D
[
k2N
Gk :
86
3 Measure Theory
Thus, by the countable additivity of on disjoint unions,

!
!
[
[

Ak D
Gk
k2N
k2N
X
.Gk /
D
k2N
n
X
.Ak / .Ak1 /
D .A1 / C lim
n!1
kD2
h
i
D .A1 / C lim .An / .A1 / ;
n!1
which establishes formula (3.5).

(2) The sequence fEk gk2N is nested and descending, and so it is simple to produce
from it a sequence of pairwise disjoint sets Fk 2 by taking set differences: that
is, let
Fk D Ek nEkC1 ;
8k 2 N:
Observe that Ek D EkC1 [ Fk and that EkC1 \ Fk D ;. Thus, by the countable

additivity of on disjoint unions,
.Ek / D .EkC1 / C .Fk / ;
Because
\
!
Ek
k2N
!
Fk D ; ;
k2N
and
E1 D
8k 2 N:
\
k2N
!
Ek
!
Fk ;
k2N
the countable additivity of on disjoint unions yields

!
!
\
[
.E1 / D
Ek C
Fk
k2N
k2N
!
\
X
Ek C
.Fk /
D
k2N
! k2N n
\
X
.Ek / .EkC1 /
D
Ek C lim
n!1
kD1
k2N
!
\
D
Ek C .E1 / lim .EnC1 / ;
k2N
which establishes formula (3.6).
n!1
t
u
3.2 Measure Spaces
87
As an application of the continuity of measure, the following result shows that

if a measurable function f on a finite measure space is unbounded, then the set on
which the values of f are very large has arbitrarily small measure.
Proposition 3.23. If .X; ; / is a finite measure space and if f W X ! R is
measurable, then for each " > 0 there is an n 2 N such that
.fx 2 X j jf .x/j > ng/ < ":
Proof. Let En D fx 2 X j jf .x/j > ng, for each n 2 N. Note that .E
\ 1 / .X/ < 1
and EnC1 En for every n. Hence, by Proposition 3.22, if E D
En , then .E/ D
n2N
lim .En /. Now because, in this particular case, E D ; and thus .E/ D 0, we
n!1
deduce that for each " > 0 there is an n 2 N such that .En / < ".
t
u
One might not have a sequence of pairwise disjoint sets at hand. Nevertheless, it
is possible to obtain an estimate on the measure of their union.
Proposition 3.24 (Countable Subadditivity of Measure). Let .X; ; / denote a
measure space. Suppose that fEk gk2N is any sequence of sets Ek 2 . Then,
!
[
X

Ek
.Ek / :
(3.7)
k2N
k2N
Proof. For each k 2 N, let
k1
[
Fk D Ek n @
Ej A :
jD1
Note that the sequence fFk gk2N consists of pairwise disjoint elements of and that
each Fk Ek . Thus, .Fk / .Ek /, by Proposition 3.21. Also,
[
[
Ek D
Fk :
k2N
Thus,
!
Ek D
k2N
which proves inequality (3.7).
[
k2N
k2N
!
Fk D
X
k2N
.Fk /
.Ek / ;
k2N
t
u
There is a rather significant difference between those measure spaces .X; ; /

in which .X/ is finite and those for which .X/ D 1. A hybrid between these two
alternatives occurs with the notion of a -finite space.
88
3 Measure Theory
Definition 3.25. A measure space .X; ; / is -finite if there is a sequence

[
fXn gn2N of measurable sets Xn 2 such that .Xn / < 1 for every n and X D
Xn .
n2N
3.3 Outer Measures

Having examined to this point some properties of measures, we turn now to the issue
of constructing measures. This will be done by first defining an outer measure.
Definition 3.26. If X is a set, then a function W P.X/ ! 0; 1 on the power set
P.X/ of X is an outer measure on X if
1. .;/ D 0,
2. .S1 / ! .S2 /, if S1 S2 , and
[
X
3.
Sk
.Sk / for every sequence fSk gk2N of subsets Sk X.
k2N
k2N
An outer measure is generally not a measure. And note that the domain of an
outer measure is the power set P.X/, rather than some particular -algebra of
subsets of X.
Definition 3.27. A sequential cover of X is a collection O of subsets of X with
the properties that ; 2 O and for every S X there is a countable subcollection
fIk gk2N O such that
[
Ik :
S
k2N
Sequential covers lead to outer measures as follows.

Proposition 3.28. Assume that O is a sequential cover of a set X. If W O ! 0; 1/
is any function for which .;/ D 0, then the function W P.X/ ! 0; 1 defined by
( 1
)
[
X

.S/ D inf
.Ik / j fIk gk2N O and S
Ik
(3.8)
kD1
k2N
is an outer measure on X.
Proof. Clearly .;/ D 0. If S T, then any fIk gk2N O that covers the set T also
covers the set S, and so .S/ .T/. Thus, all that remains is to verify that is
countable subadditive.
To this end, suppose that fSk gk2N is a sequence of subsets Sk X. Since we aim
to show that
!
X
[

Sk
.Sk / ;
k2N
k2N
3.3 Outer Measures
89
P
only the case where the sum k .Sk / converges need be considered. For this case,
suppose
S that " > 0. For each k 2 N there is a countable family fIkj gj2N O such that
Sk j Ikj and
X
.Ikj / .Sk / C
"
:
2k
Thus, fIkj gk;j2N forms a countable subcollection of sets from O that cover [k Sk and
satisfies
!
X
[
XX
"
X

.Sk / C k
Sk
.Ikj /
.Sk / C " :
2
k
j
k
k2N
k2N
As " > 0 is chosen arbitrarily, is indeed countably subadditive.
t
u
The value of an outer measure is two-fold: (i) it is frequently easier to define

an outer measure on the power set of X than it is to define a measure on some algebra of subsets of X (indeed, determining nontrivial -algebras on X is in itself a
nontrivial task), and (ii) if one has an outer measure at hand, then there is a -algebra
of subsets of X for which the restriction of to is a measure on .X; /. This
latter fact is the content of the following theorem.
Theorem 3.29 (Carathodory). If is an outer measure on a set X, then
1. the collection M .X/ of all subsets E X for which
.S/ D .E \ S/ C .Ec \ S/ ;
8S X ;
is a -algebra, and
2. the function W M .X/ ! 0; 1 defined by .E/ D .E/, E 2 M .X/, is a
measure on the measurable space .X; M .X//.
The criterion (1) in Theorem 3.29 for membership in M .X/ is called the
Carathodory criterion. The proof of Theorem 3.29 requires the following lemma.
Lemma 3.30. If E1 ; : : : ; En 2 M .X/, then
n
[
Ek 2 M .X/ :
kD1
Moreover, if E1 ; : : : ; En 2 M .X/ are pairwise disjoint, then

#!
" n
n
\ [
X

S
Ek
.S \ Ek /
8S X :
D
kD1
(3.9)
kD1
Proof. It is sufficient to consider the case n D 2, as the remaining cases follow by

induction on n.
90
3 Measure Theory
We shall prove that E1 [ E2 2 M .X/, for all E1 ; E2 2 M .X/ Let S X be
arbitrary and note that S \ .E1 [ E2 / can be written as
S \ .E1 [ E2 / D .S \ E1 / [ .S \ E2 / D .S \ E1 / [ .S \ E1c \ E2 / :
Likewise,
S \ .E1 [ E2 /c D S \ .E1c \ E2c / D .S \ E1c / \ E2c :
Thus,
.S/ .S \ .E1 [ E2 // C .S \ .E1 [ E2 /c /

D .S \ E1 / [ .S \ E1c \ E2 / C S \ E1c \ E2c

.S \ E1 / C S \ E1c \ E2 / C S \ E1c \ E2c
D .S \ E1 / C .S \ E1c /
D .S/ ;
where the final two equalities are because of E2 2 M .X/ and E1 2 M .X/,
respectively. Hence,
.S/ D .S \ .E1 [ E2 // C .S \ .E1 [ E2 /c / ;
8S X :
This proves that E1 [ E2 2 M .X/.

Next, let E1 ; E2 X be disjoint elements of M .X/. If S X, then
S \ .E1 [ E2 / \ E2 D S \ E2 and
S \ .E1 [ E2 / \ E2c D S \ E1 :
(3.10)
Thus, by using (3.10) together with the fact that E2 2 M .X/, we obtain
.S \ E1 / C .S \ E2 / D .S \ .E1 [ E2 // ;
t
u
We are now equipped to prove Theorem 3.29.

Proof. To prove (1), namely that M .X/ is a -algebra, recall that a subset E X
is an element of M .X/ if and only if Ec 2 M .X/. Hence, M .X/ is closed
under complements. Further, the empty set ; clearly belongs to M .X/. Thus, all
that remains is to prove that M .X/ is closed under countable unions.
Lemma 3.30 states that M .X/ is closed under finite unions. To get the same
result for finite intersections, note that
3.3 Outer Measures
91
E1 ; E2 2 M .X/ H) E1c ; E2c 2 M .X/

H) E1c [ E2c 2 M .X/
H) .E1c [ E2c /c 2 M .X/
H) .E1c /c \ .E2c /c D E1 \ E2 2 M .X/ :
That is, E1 \ E2 2 M .X/. By induction, E1 \ \ En 2 M .X/, for all
E1 ; : : : ; En 2 M .X/.
Now let fAk gk2N be a sequence for which Ak 2 M .X/ for all k 2 N. Let E0 D ;
and
Ek D Ak n
k1
[
Aj ;
8k 2 N:
jD1
As M .X/ is closed under finite unions and intersections, Ek 2 M .X/ for all
k 2 N. Furthermore, by Exercise 3.76, fEk gk2N is a sequence of pairwise disjoint
sets for which
[
[
Ek D
Ak :
k2N
k2N
Let
ED
Ek
and
Fn D
n
[
Ek ;
8n 2 N:
kD1
k2N
Because Fn E, we have that Ec Fnc . The sets Fn are elements of M .X/; thus,
for any subset S X,
.S/ D .S \ Fn / C .S \ Fnc /
.S \ Fn / C .S \ Ec / :
Equation (3.9) of Lemma 3.30 yields
.S \ Fn / D
n
X
.S \ Ek / :
kD1
Thus, this equation and the inequality .S/ .S \ Fn / C .S \ Ec / imply that

.S/
n
X
.S \ Ek / C .S \ Ec / ;
8n 2 N:
kD1
Therefore, by making use of the fact that is countably subadditive,
92
3 Measure Theory
1
X
.S/
.S \ Ek / C .S \ Ec /
kD1
.S \ E/ C .S \ Ec /
.S/ :
Hence, .S/ D .S \ E/ C .S \ Ec /, which proves that E 2 M .X/.
To prove (2), namely that restricted to M .X/ is a measure, note first that
.;/ D 0 and that the range of is obviously all of 0; 1.

Suppose
S now that fEk gk2N is a sequence inPM .X/ of pairwise disjoint sets and
let E D k Ek . We aim to prove that .E/ D k .Ek /. Outer measure is countably
subadditive; thus,
1
X
.E/ D .E/
.Ek / D
kD1
1
X
.Ek / :
kD1
Let S X be arbitrary. By Lemma 3.30,

" n
\ [
#!
D
Ek
kD1
n
X
.S \ Ek /
for every n 2 N :
kD1
In particular, for S D X, this yields, for every n 2 N,

n
[
!
Ek D
kD1
n
[
!
Ek D
kD1
n
X
.Ek / D
kD1
n
X
.Ek / :
kD1
Thus,
1
X
.Ek /
kD1
for every n 2 N, and so
1
[
!
Ek
kD1
1
[
kD1
!
Ek D
n
[
kD1
1
X
!
Ek D
n
X
.Ek / ;
kD1
.Ek /.
t
u
kD1
One useful consequence is the following simple result.

Proposition 3.31. Suppose that E; F 2 M.X/. If E F and if .F/ < 1, then
.FnE/ D .F/ .E/.
Proof. Write F as F D .FnE/ [ E, which is a disjoint union of elements of M.X/.
Both .F/ and .FnE/ are finite. Thus, .F/ D .FnE/ C .E/, by (3.9)
[with S D X].
t
u
3.3 Outer Measures
93
The next definition and proposition indicate that sets that have zero outer measure
are measurable.
Definition 3.32. If is an outer measure on X, then a subset S R is -null if
.S/ D 0.
Proposition 3.33. If is an outer measure on X and if E X is -null, then
E 2 M .X/.
Proof. Let E X be a -null set. If S X, then E \ S E and so 0 .E \ S/
.E/ D 0. Hence, by the subadditivity of outer measure,
.S/ .E \ S/ C .Ec \ S/ D 0 C .Ec \ S/ .S/ :
That is, .S/ D .E \ S/ C .Ec \ S/ for every S X.
t
u
What other subsets E X will belong to the -algebra M .X/? The answer to
this question depends, of course, on the character of the outer measure . A useful
answer in the setting of metric spaces is given by Proposition 3.35 below, for which
following definition will be required.
Definition 3.34. If .X; d/ is a metric space and if A and B are nonempty subsets
of X, then the distance between A and B is the quantity denoted by dist .A; B/ and
defined by
dist .A; B/ D inffd.a; b/ j a 2 A; b 2 Bg:
If, in a metric space .X; d/, the distance between subsets A and B is positive, then
A and B are disjoint and .A [ B/ .A/ C .B/. If equality is achieved in all
such cases, then the induced -algebra M .X/ will contain the Borel sets of X.
Proposition 3.35. If an outer measure on a metric space .X; d/ has the
properties that .X/ < 1 and that
.A [ B/ D .A/ C .B/;
for all subsets A; B X for which dist .A; B/ > 0, then every Borel set of X belongs
to the -algebra M .X/ induced by .
Proof. By the Carathodory criterion of Theorem 3.29, our objective is to show that,
for every open subset U X, the equation
.S/ D .S \ U/ C .S \ U c /
holds for all S X.
To this end, select a nonempty subset S of X. If S \ U D ;, then the equation
.S/ D .S \ U/ C .S \ U c / holds trivially. Thus, assume that S \ U 6D ;, and
for each n 2 N let

1
c
:
Sn D x 2 U \ S j dist .fxg; U /
n
94
3 Measure Theory
Observe that Sn SnC1 for all n 2 N and that U \ S D
Sn . By the hypothesis on
n2N
c
, the distance inequalities dist .Sn ; S \ U c / dist .Sn ; U /
1
n
> 0 imply that
.S/ ..S \ U c / [ Sn / D .S \ U c / C .Sn /:

Because .X/ < 1 and because the sets Sn form an ascending sequence, the
limit lim .Sn / exists and is bounded above by .S \ U/. If it were known that
n!1
lim .Sn / D .S \ U/, then the inequality above would lead to
n!1
.S/ .S \ U c / C .S \ U/;
which, when coupled with the inequality .S/ .S \ U c / C .S \ U/ arising
from the subadditivity of , would imply .S/ D .S \ U/ C .S \ U c /.
Therefore, all that remains is to prove that lim .Sn / D .S \ U/.
n!1
For every n 2 N, let An D SnC1 n Sn . If m; n 2 N satisfy jm nj 2, then the

distance between Am and An is positive, and so .Am [ An / D .Am / C .An /.
Therefore, by induction,
n
X
.A2k / D
n
[
kD1
so the series
A2k .S2nC1 / .S \ U/ < 1:
kD1
Hence, the series

1
X
1
X
1
X
.A2k / converges. Likewise,
kD1
.A2kC1 / converges, and
kD1
.Ak / converges. Therefore, by the countable subadditivity of ,
kD1
1
X
.Sn / .S \ U/ .Sn / C
.Ak /;
kDnC1
and so
j .S \ U/ .Sn /j
1
X
.Ak /:
kDnC1
The convergence of
1
X
kD1
.Ak / yields lim j .S \ U/ .Sn /j D 0.

n!1
t
u
3.4 Lebesgue Measure
95

The original motivation for the development of measure theory was to put the notion
of length, area, volume, and so forth on rigorous mathematical footing, with the
understanding that the sets to be measured may not be intervals, rectangles, or boxes.
The measures that captures length, area, and volume are called Lebesgue measures.
Proposition 3.36. The collection
(
On D
n
Y
)
.ai ; bi / j ai ; bi 2 R; ai < bi
iD1
is a sequential cover of Rn .
Proof. Let S Rn . For each[
x 2 S there is a neighbourhood Ux of x of the form
Qn
Ux D iD1 .ai ; bi /. Let V D
Ux , which is an open set. By Proposition 1.26, the
x2S
set B of all finite open intervals with rational end points is a basis for the topology
of R. Thus,
( n
)
Y
Bn D
.pi ; qi / j pi ; qi 2 Q; pi < qi
iD1
is a basis for the topology of Rn . By Proposition 1.24, every open set is a union of
basic open sets. Thus, since
[ Bn is countable,
[there is a countable family fIk gk2N
Bn On such that V D
Ik , whence S
Ik .
t
u
k2N
k2N
Definition 3.37. Lebesgue outer measure on Rn is the function m on P.Rn /

defined by
(
m .S/ D inf
1
X
.Ik / j fIk gk2N On and S
kD1
)
Ik ;
k2N
where On is the sequential cover of Rn given by Proposition 3.36 and the function
is defined by

n
Y
iD1
!
.ai ; bi /
n
Y
.bi ai / :
iD1
Observe that if E Rn is an open box in Rn (that is, E 2 On ), then .E/ is the

volume of E and m .E/ D .E/.
The first proposition shows that, in the case n D 1, m is a length function for all
finite intervals, open or otherwise.
96
3 Measure Theory
Proposition 3.38. If a; b 2 R are such that a < b, then

m . a; b / D m . .a; b / D m . a; b/ / D m . .a; b/ / D b a :
Proof. Because m is an outer measure, m .S1 / m .S2 / if S1 S2 . Therefore,
m . .a; b/ / m . .a; b / m . a; b / and m . .a; b/ / m . a; b/ / m . a; b /.
Since by definition, m . .a; b/ / D b a, it is enough to prove that m .a; b/ D b a.
To this end, observe that, for every " > 0, a; b .a "; b C "/. Because this open
interval covers a; b, we have that m .a; b/ . .a "; b C "/ / D b a C 2". As
this is true for every " > 0, one concludes that m .a; b/ b a D m . .a; b/ /. u
t
Similarly, one has:
Proposition 3.39. If E 2 On , then m .E/ D m .E/.
The notion of -null set, for an outer measure on a set X, was introduced
earlier. To simplify the terminology here, we shall say a subset S R is a null set if
its Lebesgue outer measure m .S/ is 0. Thus, from Proposition 3.33, every null set
S Rn is necessarily Lebesgue measurable.
Example 3.40 (Some Null Sets). The following subsets of Rn are null sets:
1.
2.
3.
4.
every finite or countably infinite set;

every countable union of null sets;
every subset of a null set;
the Cantor ternary set in R.
Proof. The details of these examples are left as an exercise (Exercise 3.89), but the
case of the Cantor set is described here.
T
The Cantor ternary set C is given by C D n2N Cn , where each Cn is a union of
2n pairwise disjoint closed intervals Fn;j of length .1=3/n . Thus,
0 n
1
n
2
2n
[
X
2
m .C / m .Cn / D m @ Fn;j A
m .Fn;j / D
:
3
jD1
jD1
As the inequality above holds for all n 2 N, m .C / D 0.
t
u
Proposition 3.41. If c is the cardinality of the continuum, then the cardinality of

M.R/ is 2c (the cardinality of the power set of R).
Proof. The Cantor ternary set C has the cardinality if the continuum (Proposition 1.83) and every subset of C is Lebesgue measurable. Hence, the cardinality
of M.R/ is the cardinality of the power set of R.
t
u
In addition to null sets, every open set is a Lebesgue measurable set.
Proposition 3.42. Every open set in Rn is Lebesgue measurable.
97
Proof. If W D .a; b/ and U D .p; q/ are open intervals, and if a < p < b < q, then
W \ U is the interval I D .p; b/ with length .I/ D .b p/ and W \ U c is an interval
J D .a; p/ with length .J/ D .p a/. Thus,
b a D m .W/ D m .I/ C m .J/ D m .W \ U/ C m .W \ U c /:
The equation above holds in cases where the inequalities a < p < b < q are not
satisfied, because either one of W \ U or W \ U c is empty, or W \ U and W \ U c
are nonempty disjoint open intervals whose lengths sum to b a.
n
n
Y
Y
A similar feature holds in Rn . If W D .aj ; bj / and U D .pj ; qj / are elements
jD1
jD1
of On , then either one of W \ U or W \ U c is empty, or W \ U and W \ U c are

n
Y
.bj aj /. Hence,
nonempty disjoint elements of On whose volumes sum to
jD1
m .W/ D m .W \ U/ C m .W \ U c /
for all W; U 2 On .
To prove that every open set in Rn is Lebesgue measurable, assume that V
n
R is an open set. Because Rn has a countable basis for its topology, every open
set is a countable union of open sets. Therefore, we may assume without loss of
n
Y
generality that V is a basic open set: V D .aj ; bj /, for some aj ; bj 2 Q. Let S Rn
jD1
be
Xarbitrary and assume that " > 0. Select a covering fUk gk On of S such that
.Uk / m .S/ C ". Because
k
S\V
[
[
.Uk \ V/ and S \ V c
.Uk \ V c /;
k
we have that
m .S \ V/ C m .S \ V c /
D
P
P

c
k .m .Uk \ V/ C m .Uk \ V //

k m .Uk /
m .S/ C ":
As " > 0 is arbitrary, we deduce that m .S/ D m .S \ V/ C m .S \ V c /, which
proves that the open set V is Lebesgue measurable.
t
u
Corollary 3.43. Every Borel subset of Rn is Lebesgue measurable.
If E and F are Lebesgue-measurable sets, then it is necessarily true that
98
3 Measure Theory
m.E [ F/ C m.E \ F/ D m.E/ C m.F/:

Proposition 3.44 below extends this property to outer measure of arbitrary sets, but
at the expense of weakening the equality above to an inequality.
Proposition 3.44. For any subsets A; B Rn ,
m .A [ B/ C m .A \ B/ m .A/ C m .B/:
Proof. Let " > 0 be given, and let fIk gk and fJi gi be coverings of A and B,
respectively, by open boxes Ik and Ji such that
X
`.Ik / m .A/ C "
and
Let U D
`.Ji / m .B/ C ":
Ik and V D
Ji . Thus, A U and B V, and A [ B U [ V and
A \ B U \ V. Because U and V are open sets, they are Lebesgue measurable and,
hence,
m .A [ B/ C m .A \ B/ m .U [ V/ C m .U \ V/ D m.U [ V/ C m.U \ V/
D m.U/ C m.V/
X
k
m.Ik / C
m.Ji / m .A/ C m .B/ C 2":
Because " > 0 is arbitrary, we have m .A [ B/ C m .A \ B/ m .A/ C m .B/.
t
u
The notion of -finite measure space was introduced in Definition 3.25 as a

hybrid of finite measure space and infinite measure space. Lebesgue measure on
Rn is a concrete example of a -finite space.
Proposition 3.45. The measure space .Rn ; M.Rn /; m/ is -finite.
Proof. If Kj D
n
Y
j; j for each j 2 N, then Kj is measurable of finite measure
m.Kj / D .2j/n , and Rn D
Kj .
t
u
j2N
Every Borel subset of Rn is Lebesgue measurable, and Borel sets are determined
by open subsets. Therefore, it seems natural to expect that the measures of arbitrary
Lebesgue-measurable sets can be approximated by the measures of open and/or
closed setsthis is the notion of regularity. The idea of translation invariance of
measure is related to the fact, for example, that if one moved an n-cube C in Rn to
some other position in space, the volume of C would not change.
A tool in analysing the regularity and translation invariance of Lebesgue measure
is the following proposition.
Proposition 3.46. The following statements are equivalent for a subset E Rn .
99
1. E is a Lebesgue-measurable set.
2. For every " > 0 there is an open set U Rn such that E U and m .UnE/ < ".
3. For every " > 0 there is a closed set F Rn such that F E and m .EnF/ < ".
Proof. The logic of proof is slightly unusual in that following implications will be
established: (1))(2) and (1))(3), then followed by (3))(1) and (2))(1).
To prove that (1) implies (2), suppose that E Rn is Lebesgue measurable and let
" > 0 be given. The cases where m .E/ is finite or infinite will be treated separately.
In the first case, assume that m .E/ < 1. By definition, there is countable
covering fIk gk2N On of E such that
1
X
.Ik / < m .E/ C " :
kD1
Let U D
Ik , which is an open (and, hence, Lebesgue measurable) set containing
E. Note that
m .E/ m .U/
1
X
.Ik / < m .E/ C " :
kD1
Because m .U/ < 1 and E U is a containment of Lebesgue-measurable sets,

Proposition 3.31 states that
m .UnE/ D m .U/ m .E/
1
X
.Ik / m .E/ < " ;
kD1
which proves (2) in the case where m .E/ < 1.T

Assume now that m .E/ D 1. Define Ek D E .k; kn /, for each k 2 N. Hence,
[
m .Ek / .2k/n and E D
Ek :
k2N
Because m .Ek / < 1, the first case implies[

there are open sets Uk Rn such that
"

Ek Uk and m .Uk nEk / < 2kC1 . Let U D
Uk , which is open and contains E.
k
Thus,
UnE
Uk nEk
k2N
and
m .UnE/ m
[
k2N
!
Uk nEk
1
X
kD1
m .Uk nEk /
1X "
< ";
2 kD1 2k
which proves (2) in the case where m .E/ D 1.

For the proof of (1) implies (3), suppose that E Rn is Lebesgue measurable
and let " > 0 be given. As E is Lebesgue measurable, so is Ec . Apply (1))(2) to
100
3 Measure Theory
Ec to conclude that there is an open set U such that Ec U and m .UnEc / < ". Let
F D U c , which is a closed set contained in E. Thus,
m .EnF/ D m .E \ F c / D m .E \ U/ D m .UnEc / < " ;
thereby proving that (1) implies (3).
To prove that (3) implies (1), assume hypothesis (3) and let " > 0 be given. By
hypothesis, there is a closed set F such that F E and m .EnF/ < . Let S Rn be
any set. Note that .S \ E/ \ F D S \ F and .S \ E/ \ F c E \ F c ; hence,
m .S \ E/ D m . .S \ E/ \ F/ C m . .S \ E/ \ F c /
m .S \ F/ C m .E \ F c /
(3.11)
D m .S \ F/ C m .EnF/
m .S \ F/ C " :
The inclusion F E implies that
m .S \ Ec / m .S \ F c / :
(3.12)
Therefore, (3.11) and (3.12) combine to produce

m .S/ m .S \ E/ C m .S \ Ec /
" C m .S \ F/ C m .S \ F c /
(3.13)
D " C m .S/ :
(The final equality arises from the fact that Fbeing closedis Lebesgue measurable.) As " is arbitrary, the inequalities (3.13) imply that m .S/ D m .S \ E/ C
m .S \ Ec /. That is, E is Lebesgue measurable.
Lastly, the proof of (2) implies (1) is similar to the proof of the (3))(1) and is,
therefore, omitted.
t
u
Proposition 3.47 (Regularity of Lebesgue Measure). Lebesgue measure m on Rn
has the following properties:
1. m.K/ < 1 for every compact subset K Rn ;
2. .E/ D inff.U/ j U Rn is open and E Ug for every measurable set E;
3. .E/ D supf.K/ j K is compact and K Eg, for every measurable set E.
Proof. Assume that K Rn is compact. For each x 2 K there is an open box Wx 2 On
of volume 1 such that x 2 Wx . From the open cover fWx gx2K of the compact set K
101
extract a finite subcover fWxj g`jD1 and deduce that

m.K/
`
X
m.Wxj / D ` < 1:
jD1
Next, assume that E Rn is a Lebesgue-measurable set and that " > 0. By

Proposition 3.46, there is an open set U R such that E U and m.UnE/ < ".
Thus,
m.U/ D m.E/ C m.UnE/ m.E/ C " :
Hence, .E/ D inff.U/ j U Rn is open and E Ug.
Now assume that E is a Lebesgue-measurable set such that the closure E of E
is compact. Let " > 0 be given. By the previous paragraph there is an open set U
containing E n E such that m.U/ < m.E n E/ C ". Let K D E \ U c , which is a closed
subset of the compact set E; hence, K is compact. Furthermore, if x 2 K, then x 2 E
and x 62 E \ Ec , which is to say that x 2 E. Thus, K E. Because

m.E/ m.K/ D m.E/ m.E/ m.U/ D m.U/ < m.E/ m.E/ C ";
we deduce that m.E/ < m.K/ C " and .E/ D supf.K/ j K is compact and K Eg.
n
Y
For each k 2 N, the set Bk D k; k is compact. If Ek D E \ Bk , then fEk gk2N is
1
an ascending sequence of sets such that E D
Ek . Thus, by continuity of measure,
k2N
m.E/ D lim m.Ek /. Choose any positive r 2 R such that r < m.E/. Thus, there is a
k!1
k 2 N such that r < m.Ek / < m.E/. Because Ek is compact, the previous paragraph
shows that there is a compact subset K of Ek such that r < m.K/. Now since Ek E,
K is also a subset of E. As the choice of r < m.E/ is arbitrary, this shows that
.E/ D supf.K/ j K is compact and K Eg.
t
u
If x 2 Rn and S Rn , then x C S denotes the subset of Rn defined by
x C S D fx C y j y 2 Sg:
Proposition 3.48 (Translation Invariance of Lebesgue Measure). If E Rn is
Lebesgue measurable and if x 2 R, then x C E is Lebesgue measurable and
m.x C E/ D m.E/ :
Proof. If I 2 On is the open box I D
(3.14)
n
Y
.aj ; bj /, then x C I and I have the same
jD1
volume. Thus, for any subset S 2 Rn and x 2 Rn , the outer measures of S and x C S
102
3 Measure Theory
coincide. Therefore, we aim to prove that x C E is a Lebesgue-measurable set if E is

a Lebesgue-measurable set. To this end, we shall employ Proposition 3.46.
Let " > 0. Because E is measurable, Proposition 3.46 states that there is an open
set U Rn such that E U and m .UnE/ < ". Thus, there is a countable covering
of UnE by open boxes Ik such that
X
.Ik / < " :
k2N
For each k, x C Ik is an open box of volume .x C Ik / D .Ik /. Furthermore, because

U is a countable union of basic opens (all of which are open boxes), the set x C U is
open, the inclusion x C E r C U is clear, and
[
.x C U/ n .x C E/ D fx C y j y 2 UnEg D x C .UnE/
.x C Ik / :
k2N
Thus,
m ..x C U/ n .x C E//
.x C Ik / D
k2N
.Ik / < " :
k2N
Hence, x C E satisfies the hypothesis of Proposition 3.46, thereby completing the

proof that x C E is Lebesgue measurable.
u
t
It is natural to wonder whether every subset of R is Lebesgue measurable. That
is not the case, as the following theorem shows. Because the proof of the theorem
below requires the Axiom of Choice, the result is existential rather than constructive.
Theorem 3.49 (Vitali). There is a subset V of R such that V is not Lebesgue
measurable.
Proof. Consider the relation
on R defined by x
y if and only if y x 2 Q. It
is not difficult to verify that
is an equivalence relation, and so the equivalence
classes xP of x 2 R form a partition of R. Note that xP D x C Q, for each x 2 R.
For each x 2 .1; 1/, let Ax D xP \.1; 1/. Of course, if x1 ; x2 2 .1; 1/, then either
Ax1 D Ax2 or Ax1 \ Ax2 D ;. By the Axiom of Choice, there is a set V such that, for
every x 2 .1; 1/, V \ Ax is a singleton set.
The set Q \ .2; 2/ is countable; hence, we may write
Q \ .2; 2/ D fqk j k 2 Ng :
For each k 2 N, consider qk C V . Suppose that x 2 .qk C V / \ .qm C V /, for some
k; m 2 N. Then there are ck ; cm 2 V such that qk C ck D qm C cm ; that is, ck cm D
qm qk 2 Q, which implies that ck 2 Acm . As V \ Acm is a singleton set, it must be
that ck D cm and qk D qm . Hence, fqk C V gk2N is a countable family of pairwise
disjoint sets, each of which is obviously contained in the open interval .3; 3/.
Let x 2 .1; 1/ and consider Ax . By construction of V , there is precisely one
element y 2 .1; 1/ that is common to both Ax and V . Thus, x and y are equivalent,
which is to say that x y 2 Q. Because x; y 2 .1; 1/, x y 2 .2; 2/; hence, x y D
qk , for some k 2 N. Therefore, x 2 qk C V .
103
The arguments above establish that

[
.1; 1/
.qk C V / .3; 3/ :
(3.15)
k2N
If V were Lebesgue measurable, then each qk C V would be Lebesgue measurable, by Proposition 3.48, and m.qk C V / would equal m.V /. Therefore, if V were
Lebesgue measurable, then
!
[
X
X
m
.qk C V / D
m.qk C V / D
m.V /
k2N
k2N
k2N
would hold. Furthermore, computation of Lebesgue measure in (3.15) would yield

2<
1
X
m.V / < 6 :
(3.16)
kD1
But there is no real number m.V / for which (3.16) can hold. Therefore, it cannot be
that V is a Lebesgue-measurable set.
t
u
Corollary 3.50. Outer measure m on R is not countably additive. That is, there is
a sequence fSk gk2N of pairwise disjoint subsets Sk R such that
!
[
X

m
Sk <
m .Sk / :
k2N
k2N

Proof. Let Sk D qk C V , as in the proof
[ of Theorem 3.49. Because m is countably
subadditive and because .1; 1/
.qk C V /,
k2N
2m
!
Sk :
k2N
Therefore, because m .qk C V / D m .V /, we have that. m .Sk / D m .S1 /, for all

k 2 N, and the inequality above shows that m .S1 / 6D 0. Thus,
X
m .Sk / D 1 :
k2N
On the other hand,

[
k2N
Hence,
Sk .3; 3/
H)
[
k2N
!
Sk
6:
104
3 Measure Theory
[
k2N
!
Sk
<
m .Sk / ;
k2N
t
u
as claimed.
Vitalis Theorem produces a nonmeasurable subset of .1; 1/; the argument can
be modified to produce a nonmeasurable subset of any measurable set of positive
measure.
Theorem 3.51. If E 2 M.R/ and if m.E/ > 0, then there is a subset V E such
that V 62 M.R/.
Proof. Let Ak D E \ k; k for each k 2 N. The sequence fAk gk2N is an ascending
sequence in M.R/ with union E. Hence, by continuity of measure,
0 < m.E/ D lim m.Ak / ;
k!1
and so m.Ak0 / > 0 for some k0 2 N. Now apply the argument of Theorem 3.49
using E \ k0 ; k0 in place of .1; 1/ to determine a nonmeasurable subset V of
E \ k0 ; k0 .
t
u
The Borel sets and null sets determine the structure of Lebesgue-measurable sets.
Proposition 3.52. The following statements are equivalent for a subset E R:
1. E is a Lebesgue-measurable set;
2. there exist B; E0 R such that:
a.
b.
c.
d.
B is a Borel set,
E0 is a null set,
E0 \ B D ;, and
E D B [ E0 .
t
u
Proposition 3.52 shows how Borel sets can be used to characterise Lebesguemeasurable sets. Much less obvious is the following theorem, which indicates that
these two -algebras are in fact distinct.
Theorem 3.53 (Suslin). There exist Lebesgue-measurable subsets of R that are
not Borel sets. In fact, there are Lebesgue-measurable subsets of the Cantor ternary
set that are not Borel sets.
Proof. Let Q denote an extension of the Cantor ternary function (see Proposition 1.86) W 0; 1 ! 0; 1 to a function R ! 0; 1 by setting Q D 0 on .1; 0/,
Q D on 0; 1, and Q D 1 on .1; 1/. Let f W R ! R be given by
Q
f .x/ D .x/
C x;
8x 2 R:
Observe that f is continuous and monotone increasing.

Define a collection of subsets of R as follows:
105
D fS R j f .S/ 2 B.R/g :
We now show that is a -algebra of subsets of R. Because R D f .R/, R 2 .
Moreover, if A 2 , then f .Ac / \ f .A/ D ; and R D f .R/ D f .A/ [ f .Ac / imply that
f .Ac / D f .A/c , whence Ac 2 . That is, is closed under complementation. Now
suppose that fAk gk2N ; then
f
[
k2N
Ak D
f .Ak / 2 B.R/ :
k2N
Hence, is a -algebra.
If p; q 2 Q and p < q, then the continuity of f and the fact that f is monotone
increasing leads to f . .p; q/ / D .f .p/; f .q//. Therefore, .p; q/ 2 for all p; q 2 Q.
Because is a -algebra and contains the base for the topology on R,
necessarily contains the Borel sets of R. Hence,
f .B/ 2 B.R/ ;
8 B 2 B.R/ :
(3.17)
In particular, if C is the Cantor ternary set, then f .C / is a Borel set. We now show
that f .C / has positive measure.
To this end note that 0; 1nC is a union of countably many pairwise disjoint
intervals .ak ; bk /, where ak ; bk 2 C for all k 2 N. Proposition 1.86 shows that is
constant on each such open interval. Therefore,
2 D m .0; 2/ D m .f .0; 1//
D m .f .C [ .0; 1nC ///
D m .f .C // C m .f .0; 1nC //
D m .f .C // C
m ..ak C .ak /; bk C .bk //
k2N
D m .f .C // C
.bk ak /
k2N
D m .f .C // C m.0; 1nC /
D m .f .C // C 1 :
Thus, m .f .C // D 1 > 0 and so, by Theorem 3.51, f .C / contains a subset V that is
not Lebesgue measurable. Let Q D f 1 .V /. Because f is an injective function, the
preimage Q of V under f must be contained in C . Thus, Q is a null set and, hence,
106
3 Measure Theory
is a Lebesgue-measurable set. However, Q is not a Borel set. (If Q were a Borel set,
then inclusion 3.17 would imply that f .Q/ D V would be a Borel setbut it is not.)
Hence, Q 2 M.R/ and Q 62 B.R/.
t
u
Similar results hold in higher dimensions.
Theorem 3.54. Not every subset of Rn is Lebesgue measurable, and there exist
Lebesgue-measurable subsets of Rn that are not Borel measurable.
t
u
3.5 Atomic and Non-Atomic Measures

There are a variety of ways to distinguish between qualitative properties of
measures, and in this section we consider atomic measures and their polar opposites,
non-atomic measures.
Definition 3.55. Assume that .X; ; / is a measure space.
1. A measurable subset E X is an atom for if .E/ > 0 and one of .E \ F/ or
.E \ F c / is 0, for every F 2 .
2. The measure on .X; / is atomic if every measurable set of positive measure
contains an atom for .
3. The measure is non-atomic if has no atoms.
Thus, counting measure on N is atomic, whereas Lebesgue measure on R is nonatomic (Exercises 3.93 and 3.94). Every measure can be decomposed uniquely as
a sum of two such measures, as shown by Proposition 3.57 below. The proof will
make use of the following concept of singularity.
Definition 3.56. If and Q are measures on .X; /, then is singular with respect
to Q for each E 2 there exists a set F 2 with the properties that
1. F E,
2. .E/ D .F/, and
3. .F/
Q
D 0.
The notation S Q indicates that is singular with respect to ,
Q and if both S Q
and S
Q occur, then and Q are said to be mutually singular.
Proposition 3.57. Every measure on a measurable space .X; / has the form
D a C na , for some mutually singular atomic measure a and non-atomic
measure na on .X; /. Moreover, if Q a and Q na are atomic and non-atomic
measures on .X; / such that D Q a C Q na , then Q a D a and Q na D na .
Proof. Let D be the family of all countable unions of sets that are atoms for . For
each E 2 , define
a .E/ D supf.E \ D/ j D 2 Dg
3.6 Measures on Locally Compact Hausdorff Spaces
107
na .E/ D supf.E \ N/ j a .N/ D 0g:

Observe that a and na are measures on .X; / and satisfy D a C na . If
a .N/ D 0, then .N \ D/ D 0 for all D 2 D; hence, na .D/ D 0 for all D 2 D,
and so a S na . By definition of na , na .D/ D na .D \ N/ and a .N/ D 0 imply
that a .D \ N/ D 0; hence, na S a .
To show that a is atomic, suppose E 2 such that a .E/ 0. Because
a .E/ D 0 when .E/ D 0, we deduce that .E/ 0. Furthermore, by the definition
of a , a .E/ 0 implies there is some
[ D 2 D such that .E \ D/ 0. By the
definition of D, we can write D D
Dn , where each Dn is an atom for , and
n2N
so .E \ Dn / 0 for some n 2 N. Since E \ Dn is an atom for such that

a .E \ Dn / 0, and because a .E/ D 0 when .E/ D 0, we see that E \ Dn is
an atom for a . Thus, a is an atomic measure.
To show that na is non-atomic, suppose that na .E/ 0. Therefore, .E \ N/
0 for some N 2 with a .N/ D 0. The set E \ N is not an atom for , because
a .E \ N/ 0 if E \ N were an atom. Since .E \ N/ 0 and because E \ N is not
an atom for , there exists F 2 such that .E \N \F/ 0 and ..E \N/nF/ 0.
Hence, na .E \ F/ 0 and na .E n F/ 0, implying that na is non-atomic.
The proof of the uniqueness of the decomposition is left as Exercise 3.95.
t
u

If one considers the Borel sets of a topological space X, then it is natural to expect
that certain topological features of X play a role in the measure theory of X. But for
this to occur, the particular measure under consideration needs to be aware of the
topology. One class of measures that is sensitive to topology is the class of regular
measures.
Definition 3.58. Let .X; T / be a topological space and consider a measurable
space .X; / in which contains the -algebra B.X/ of Borel sets of X. A measure
on .X; / is said to be a regular measure if
1. .K/ < 1 for every compact subset K X,
2. .E/ D inff.U/ j U is open and E Ug, for every E 2 , and
3. .U/ D supf.K/ j K is compact and K Ug, for every open set U.
Observe that Proposition 3.47 asserts that Lebesgue measure is regular.
The third property above for the measure of an open set extends to arbitrary
measurable sets of finite measure.
Proposition 3.59. Assume that is a regular measure on .X; /, where X is a
topological space and where contains the Borel sets of X. If E 2 satisfies
.E/ < 1, then
108
3 Measure Theory
.E/ D supf.K/ j K is compact and K Eg:

Proof. Assume that E 2 has finite measure and let " > 0. Because .E/ D
inff.U/ j U is open and E Ug, there is an open set U X such that E U and
.U/ < .E/ C "=2. Hence, .U/ < 1 and .U n E/ < "=2. Because U is open and
has finite measure, the same type of argument shows that there is a compact set A
with A U and .U/ < .A/ C "=2. Lastly, since .U n E/ < "=2, there is an open
set W with U n E W and .W/ < "=2. The set W c is closed and is contained in
U c [E. Thus, K D A\W c is a closed subset of a compact set and is, hence, compact.
Further,
K D A \ W c A \ .U c [ E/ D .A \ U c / [ .A \ E/ D A \ E E
and
.E/ .U/
< .A/ C "=2
D .A \ W/ C .A \ W c / C "=2
< "=2 C .K/ C "=2:
Hence, K E and .E/ < .K/ C " implies that .E/ is the least upperbound of all
real numbers .K/ in which K is a compact subset of E.
t
u
Proposition 3.59 admits a formulation for -finite spaces, which will be of use in
our analysis of Lp -spaces.
Proposition 3.60. If .X; ; / is a -finite measure space in which X is a topological space, contains the Borel sets of X, and is regular, then
.E/ D supf.K/ j K is compact and K Eg
for every E 2 .
t
u
Continuous functions are, from the point of view of analysis, fairly well understood. In comparison, measurable functions appear to be harder to grasp because
of the existential nature of measurability. Therefore, in this light, the following
two theorems are striking, for they show that, under the appropriate conditions,
measurable functions within " of being continuous.
Theorem 3.61. Assume that is a regular finite measure on .X; /, where X is a
compact Hausdorff space and where contains the Borel sets of X. If f W X ! R is a
109
bounded measurable function, then for every " > 0 there exist a continuous function
g W X ! R and a compact set K such that gjK D fjK and .K c / < ".
Proof. To begin with, assume f is a simple function with range f1 ; : : : ; n g,
where 1 ; : : : ; n 2 0; 1 are distinct real numbers. Let Ej D f 1 .fj g/, which is a
measurable set; note that Ej \Ei D ; if i 6D j. Let " > 0 be given. By Proposition 3.59,
for each j there is a compact subset Kj Ej with .Kj / C n" > .Ej /. Thus,
n
[
"
.Ej n Kj / < n for every j. Because Ej \ Ei D ; for i 6D j, if K D
Kj , then
jD1
1
0
n
n
[
X
c
A
@
Ej n Kj D
.Ej n Kj / < ":
.K / D
jD1
jD1
The restriction fjK of f to the compact set K is plainly continuous. Since X is normal,
the Tietze Extension Theorem asserts that fjK has a continuous extension g W X !
0; 1.
Assume now that f is an arbitrary measurable function with 0 f .x/ 1 for
every x 2 X. By Proposition 3.14 there is a monotone increasing sequence f'n gn2N
of nonnegative simple functions 'n on X such that lim 'n .x/ D f .x/ for every x 2
n!1
X. In fact, because f .X/ 0; 1, the convergence of f'n gn to f is uniform on X

(Exercise 3.83). By the previous paragraph, for each n 2 N there is a compact set Kn
N
X
c
n
such that .Kn / < "=2 and 'njKn is continuous. Observe that '1 .x/ C
.'n .x/
'n1 .x// D 'N .x/, and so '1 C
continuous on K D
1
\
1
X
nD2
N
X
.'n 'n1 / D f . Because '1 C
.'n 'n1 / is
nD2
nD2
1
X
Kn , and because '1 C
.'n 'n1 / converges uniformly to
nD1
nD2
f , the measurable function f is continuous on K. By the Tietze Extension Theorem,

fjK has a continuous extension g W X ! 0; 1. Because
.K c /
1
X
.Knc / < ";
nD1
this completes the proof of the theorem in the case where f .X/ 0; 1.
For the case of general f , select 2 R such that f .X/ 1; 1, and decompose
f as .f /C .f / , where .f /C and .f / are measurable functions with ranges
contained in 0; 1. Thus, the case of general f follows readily from the case of f
with f .X/ 0; 1.
t
u
Theorem 3.62 (Lusin). Assume that is a regular measure on .X; /, where X
is a locally compact Hausdorff space and where contains the Borel sets of X. If
f W X ! R is a measurable function with the property that fjEc D 0 for some E 2
with finite measure, then for every " > 0 there exists a continuous and bounded
110
3 Measure Theory
function g W X ! R such that

.fx 2 X j f .x/ g.x/g/ < ":
Proof. By hypothesis, fjEc D 0; thus, En D fx 2 X j jf .x/j > ng is a subset of E for
every n 2 N. Because .E/ < 1, Proposition 3.23 implies that .En / < "=6 for
some n 2 N. Hence, if F D E \ Enc , which is a set of finite measure, then fjF is
bounded.
By Proposition 3.59, there is a compact subset Y F such that .F n Y/ < "=6.
Consider the bounded measurable function fjY . By Theorem 3.61, there is a compact
subset K Y and a continuous function g0 W Y ! R such that g0jK D fjK and .Y n
K/ < "=6. Thus,
E n K D En [ .F n Y/ [ .Y n K/
yields .E n K/ < "=2.
The regularity of again implies the existence of an open set U X for which
E U and .U/ < .E/ C "=2. Hence, .U/ is finite and .U n E/ < "=2. Because
K is compact and K U, Theorem 2.43 asserts that g0 admits a continuous and
bounded extension g W X ! R such that g.x/ D 0 for all x 62 U. Therefore, 0 D
gjUc D fjUc and, hence,
.fx 2 X j f .x/ g.x/g/ D .fx 2 E j f .x/ g.x/g/ C .fx 2 Ec j f .x/ g.x/g/
< "=2 C "=2 D ";
t
u
3.7 Signed and Complex Measures

Extending the notions of length, area, volume, and other arbitrary measures to
real- and complex-valued quantities results in the concepts of signed measure and
complex measure.
Definition 3.63. A function ! W ! 1; C1 on a measurable space .X; / is
called a signed measure if !.;/ D 0 and
!
1
[
kD1
!
Ek D
1
X
!.Ek /;
kD1
for every sequence fEk gk2N of pairwise disjoint sets Ek 2 .
111
The definition of signed measure entails some subtleties. First of all, arithmetic
in the extended real number system 1; C1 does not admit sums of the form
.1/ C .C1/ or .C1/ C .1/, which implies that for each E 2 at most one
of !.E/ or !.Ec / can have an infinite value (because !.X/ D !.E/ C !.Ec /). In
particular, this means that if there exists a measurable set E with !.E/ D C1, then
necessarily !.X/ D C1; and if there exists a measurable set E with !.E/ D 1,
then !.X/ D 1 necessarily. Therefore, !.X/ can achieve at most one of the values
1 or C1. If does not achieve either of these infinite values, then ! is said to
be a finite signed measure. The triple .X; ; !/ is called a signed measure space.
Definition 3.64. If .X; ; !/ is a signed measure space, and if P; N 2 , then
1. P is said to be positive with respect to ! if !.E \ P/ 0 for every E 2 , and
2. N is said to be negative with respect to ! if !.E \ N/ 0 for every E 2 .
Interestingly, a signed measure partitions a signed measure space into a positive
part and a negative part, as shown by the Hahn Decomposition Theorem below.
Theorem 3.65 (Hahn Decomposition of Signed Measures).
signed measure space, then there exist P; N 2 such that
If .X; ; !/ is a
1. P is positive with respect to ! and N is negative with respect to !,

2. P \ N D ;, and
3. X D P [ N.
Proof. We may assume without loss of generality that 1 is not one of the values
assumed by !. Let D inff!.E/ j E 2 is a negative setg. (Because ; is a negative
set, the infimum is defined.) Let fEk gk2N be
of measurable sets for which
1
0 a sequence
k1
[
D limk !.Ek /. For each k let Nk D Ek n @ Ek A so that fNk gk2N is a sequence of
jD1
pairwise disjoint negative sets such that D infk !.Nk /. Thus, with N D
have for every j 2 N that !.N/ D
1
X
kD1
1
[
Nk , we
kD1
!.Nk / !.Nj /. Hence, !.N/ D and N is a
negative set. Because 1 is not in the range of !, it must be that !.N/ 2 R. Hence,
is the minimum measure of all negative subsets of X.
Let P D N c . Assume, contrary to what we aim to prove, that P is not a positive
set. Thus, there exists a measurable subset E P such that !.E/ < 0. The set E is
not negative because, if it were, then N [ E would also be a negative set of measure
!.N [E/ D C!.E/ < , which contradicts the fact that is the minimum measure
of all negative subsets of X. Hence, E must possess a measurable subset F of positive
measure. Let n1 2 N denote the smallest positive integer for which there exists a
measurable subset F1 E of measure !.F1 / 1=n1 . Since E n F1 and F1 are disjoint
and have union E, !.E/ D !.E n F1 / C !.F1 /. That is, !.E n F1 / D !.E/ !.F1 /
!.E/n1 1 < !.E/. For the very same reasons given earlier, the set E nF1 cannot be
negative; thus, E n F1 contains a measurable subset of positive measure. Let n2 2 N
112
3 Measure Theory
denote the smallest positive integer for which there exists a measurable subset F2
.E n F1 / of measure !.F2 / 1=n2 . Repeating this argument inductively produces a
subset fnk gk2N N and a sequence fFk gk2N of pairwise disjoint measurable subsets
1
[
Fk satisfies
Fk E such that the set F D
kD1
1
X
1
X
1
!.Fk /
> 0:
!.F/ D
n
kD1
kD1 k
Therefore, the subset G D E n F of E satisfies !.G/ !.G/ C !.F/ D !.E/ < 0.

Since 1 is not in the range of !, !.G/ is a negative real number, and so
0<
1
1
X
X
1

!.Fk / D !.F/ D !.E/ !.G/ < j!.G/j < 1
n
kD1 k
kD1
implies that limk nk 1 D limk !.Fk / D 0. Therefore, if Q is a measurable subset of

G, then
!
1
1
\
\
c
c
Q G D E\F D E\
Fk D
.E n Fk /
kD1
kD1
implies that Q E n Fk for every k 2 N. If it were true that !.Q/ > 0, then for some
1
j 2 N we would have !.Q/ > nj 1
, which is to say that Q is a subset of E n Fj of
measure !.Q/ >
1
nj 1
>
1
,
nj
in contradiction to the property of nj being the smallest
positive integer for which E nFj has a subset A of measure !.A/ > n1j . Hence, !.Q/
0 and the fact that Q is an arbitrary measurable subset of G implies that G is a
negative set. But G P implies that G \ N D ; and so the negative subset G [ N
satisfies !.G [ N/ < , which is in contradiction to the fact that is the minimum
measure of all negative subsets of X. Therefore, it must be that P is a positive set.
t
u
The sets P and N that arise in Theorem 3.65 are said to be a Hahn decomposition of the signed measure space .X; ; !/. While this decomposition need
not be unique, Exercise 3.99 shows that if .P1 ; N1 / and .P2 ; N2 / are two Hahn
decompositions of a signed measure space .X; ; !/, then
!.E \ P1 / D !.E \ P2 /
and
!.E \ N1 / D !.E \ N2 /
for all E 2 . Therefore, the functions !C ; ! W ! 0; C1 defined by

!C .E/ D !.E \ P/
and
! .E/ D !.E \ N/
(3.18)
are measures on .X; / and are independent of the choice of Hahn decomposition
.P; N/ of .X; ; !/. Note, also, that at least one of !C and ! is a finite measure.
These observations give rise to the next theorem.
113
Theorem 3.66 (Jordan Decomposition Theorem). For every signed measure !

on a measurable space .X; /, there exist measures !C and ! on .X; / such
that
1. at least one of !C and ! is a finite measure, and
2. !.E/ D !C .E/ ! .E/, for every E 2 .
Furthermore, if ; are measures on .X; /, where at least one of which is finite,
and if !.E/ D .E/ .E/ for every E 2 , then !C .E/ .E/ and ! .E/ .E/,
for all E 2 .
t
u
Turning now to complex measures, the definition below departs from the
definitions of measure and signed measure in that it is assumed from the outset
that the measure is finite.
Definition 3.67. A function W ! C on a measurable space .X; / is called a
complex measure if .;/ D 0 and

1
[
!
Ek D
kD1
1
X
.Ek /;
kD1
for every sequence fEk gk2N of pairwise disjoint sets Ek 2 .

By decomposing a complex measure into its real and imaginary parts < and
=, two finite signed measures are obtained, each of which is a difference of finite
measures. Hence, there are finite measures j on .X; /, for j D 1; : : : ; 4, such that
D .1 2 / C i.3 4 /:
By considering the function E 7! j.E/j, something close to a measure is
obtainedbut the triangle inequality makes this function countably subadditive
rather than additive on sequences of pairwise disjoint sets. Therefore, to obtain a
measure from a complex measure requires slightly more effort.
Definition 3.68. In a measurable space .X; /, a measurable partition of a measurable set E X
Sis a family PE of countably many subsets A 2 such that A E
for all A 2 PE , A2PE A D E, and A \ B D ; whenever A; B 2 PE are distinct.
Proposition 3.69. If is a complex measure on a measurable space .X; / and if
jj W ! 0; 1 is defined by
jj.A/ D sup
8
<X
:
E2PA
9
=
j.E/j j PA is a measurable partition of A ;
;
then jj is a finite measure on .X; /.
114
3 Measure Theory
Proof. Because P; D fA; Bg, where A D B D ;, is measurable partition of the

empty set ;, we have that .;/ D .;/ C .;/ in C and so .;/ D 0.
To prove that jj is countably additive,
S let fEk gk2N be a sequence of pairwise
disjoint measurable sets and let E D k2N
XEk . For each k, consider an arbitrary
measurable partition fFkj gj2N of Ek ; thus,
j.Fkj /j jj.Ek /. Because fFkj gk;j2N
j
is an arbitrary measurable partition of E,

1 X
1
X
j.Fkj /j jj.E/:
kD1 jD1
For each k, jj.Ek / is the supremum of

fFkj gj2N
j j.Fkj /j
over all measurable partitions

1
X
of Ek , and therefore the inequality above yields
jj.Ek / jj.E/.
kD1
Conversely, select an arbitrary measurable partition fA` g`2N of E. Because the

sets fEk gk2N are pairwise disjoint, fA` \ Ek gk2N is a partition of A` for every ` 2 N,
and fA` \ Ek g`2N is a partition of Ek for every k 2 N. Thus,
1
X
j.A` /j
`D1
1 X
1
X
j.A` \ Ek /j D
`D1 kD1
and so jj.E/
1
X
1 X
1
X
j.A` \ Ek /j
kD1 `D1
1
X
j.Ek /j;
kD1
j.Ek /j. Hence, jj is countably additive.
kD1
As indicated previously, there are finite measures j on .X; /, for j D 1; : : : ; 4,

such that D .1 2 / C i.3 4 /. Thus, for any measurable set E 2 , j.E/j
4
X
j .E/. Therefore, if PX is a partition of X, then
jD1
j.E/j
E2PX jD1
E2PX
Hence, jj.X/
4
X X
4
X
j .E/ D
4 X
X
jD1 E2PX
j .E/ D
4
X
j .X/ < 1:
jD1
j .X/, which proves that jj is a finite measure.
t
u
jD1
Definition 3.70. In Proposition 3.69 above, the measure jj on .X; / induced by
the complex measure is called the total variation of .
Problems
115
Problems
3.71. Show that the collection of all subsets E of an infinite set X for which E or
the complement Ec of E is countable is a -algebra.
\
is a 3.72. Prove that if is a family of -algebras on a subset X, then
2
algebra.
3.73. Assume that is a -algebra of subsets of X. Show that, for each E 2 , the
collection .E/ of subsets of X defined by
.E/ D fE \ A j A 2 g
is a -algebra on E.
3.74. Let be a -algebra of subsets of a nonempty set X, and let Ek 2 for
k 2 N. Define

T S
lim sup Ek D k1
E ;
S Tnk n
lim inf Ek D k1
nk En :
Prove the following statements.
1. lim sup Ek and lim inf Ek belong to . [
2. If E1 E2 E3 : : : , then lim sup Ek D
Ek D lim inf Ek
k
3.75. Let Ek denote the closed interval Ek D 0; 1 C .1/

. Determine the sets
k
lim sup Ek and lim inf Ek . (Suggestion: consider the cases k even and k odd separately.)
k
3.76. Let X be a nonempty set X and let fAk gk2N be a sequence of subsets of X.
Define E0 D ; and, for n; m 2 N,
Em D
m
[
Ak ;
Fn D An nEn1 :
kD1
Prove the following statements.

1. fEn gn is a monotone increasing sequence of sets (that is, En EnC1 for all n).
sequence[of pairwise disjoint sets.
2. [
fFn gn is a[
3.
En D
Fn D
An .
n
3.77. Prove that if a -algebra on an infinite set X has infinitely many elements,
then is uncountable.
116
3 Measure Theory
3.78. Prove that if .X; T / is a topological space, and if T is the -algebra

generated by T , then, with respect to the measurable space .X; T /, every
continuous function f W X ! R is a measurable function.
3.79. Suppose that .X; / is a measurable space and that h W X ! R is a measurable
function for which h.x/ 6D 0, for all x 2 X. Prove that the function 1=h is measurable.
3.80. Prove that if .X; / is a measurable space and if E X, then the characteristic
function E W X ! R is a measurable function if and only if E 2 .
3.81. Let U be a nonempty subset of N (see Section 2.6), and consider the
characteristic function U . Prove that U is continuous if and only if both U and
U are open in N.
3.82. Assume that .X; / is a measurable space and that E 2 .
1. If f W E ! R is a measurable function relative to the measurable space .E; .E//,
then prove that the extension fQ W X ! R of f defined by fQ D f E is a measurable
function with respect to the measurable space .X; /.
2. Conversely, if fQ W X ! R is a measurable function with respect to the measurable
space .X; /, and if f D fQjE (the restriction of fQ to E), then prove that f W E ! R
is a measurable function with respect to the measurable space .E; .E//.
3.83. If f W X ! 0; 1 is a measurable function, then prove that there is a monotoneincreasing sequence of nonnegative simple functions 'n W X ! 0; 1 such that
lim 'n .x/ D f .x/ uniformlythat is, for every " > 0 there is an N" 2 N such that
n!1
jf .x/ 'n .x/j < " for all n N" and all x 2 X.
3.84. Let X be an infinite set and let be the -algebra in Exercise 3.71. Define a
function W ! 0; 1 by .E/ D 0 if E 2 is countable and .E/ D 1 if E 2
is uncountable. Show that is a measure on .X; /.
3.85. Consider the measurable space .N; P.N//, where P.N/ is the power set of
N. Prove that the function W ! 0; 1 defined by
.E/ D the cardinality of E
defines a measure on .N; P.N//.
3.86. A function W ! 0; 1/ on a measurable space .X; / is finitely additive
if, for all finite sub-collections fEk gnkD1 of pairwise disjoint measurable sets Ek ,
n
[
kD1
!
Ek D
n
X
kD1
.Ek /:
Problems
117
Prove that if a finitely additive function also satisfies lim .Ak / D 0, for every
k
descending sequence A1 A2 A3 of sets Aj 2 in which

is in fact a measure on .X; /.
1
\
Ak D ;, then
kD1
3.87. Assume that is a measure on a measurable space .X; /. Prove that

.E [ F/ C .E \ F/ D .E/ C .F/ ;
for all E; F 2 .
3.88. Prove that if E 2 On , then m .E/ D m .E/, where
n
Y
On D f .ai ; bi / j ai ; bi 2 R; ai bi g:
iD1
3.89. Prove that each of the following subsets of Rn is a null set.

1. Every finite or countably infinite set.
2. Every countable union of null sets.
3. Every subset of a null set.
3.90. Prove that if E Rn is Lebesgue measurable such that m.E/ > 0, then E
contains a nonmeasurable subset.
3.91. Prove that the following statements are equivalent for a subset E R:
1. E is a Lebesgue-measurable set;
2. there exist B; E0 R such that:
a.
b.
c.
d.
B is a Borel set,
E0 is a null set,
E0 \ B D ;, and
E D B [ E0 .
3.92. Prove that there exist subsets S of Rn that are not Lebesgue measurable, and
that there exist Lebesgue-measurable subsets E of Rn that are not Borel measurable.
3.93. Determine the atoms for counting measure on N.
3.94. Prove that Lebesgue measure on Rn is non-atomic.
3.95. Suppose that D a C na D Q a C Q na are two decompositions of a measure
on .X; / as the sum of an atomic measure and a non-atomic measure, where
a S na , na S a , Q a S Q na , and Q na S Q a .
1. Show that na S Q a Q a S na
2. Show that Q na .E/ na .E/ 0 and a .E/ Q a .E/ 0 dor every E 2
3. Show that Q a D a and Q na D na .
118
3 Measure Theory
3.96. Prove that if .X; ; / is a -finite measure space in which X is a topological

space, contains the Borel sets of X, and is regular, then
.E/ D supf.K/ j K is compact and K Eg
for every E 2 .
3.97. Let denote the Borel sets of X D 0; 1 and define a function W ! 0; 1
by .E/ D m.E/, if 0 62 E, and .E/ D 1, if 0 2 E.
1. Prove that is a measure on .X; /.
2. Prove that .X; ; / is not a -finite measure space.
3.98. Show that, in a signed measure space .X; ; !/, the union and intersection of
finitely many positive sets are positive sets, and that the union and intersection of
finitely many negative sets are negative sets.
3.99. Suppose that .P1 ; N1 / and .P2 ; N2 / are Hahn decompositions of a signed
measure space .X; ; !/. Prove that, for every E 2 ,
!.E \ P1 / D !.E \ P1 \ P2 / D !.E \ P2 /:
3.100. Assume that .X; ; !/ is a signed measure space with Hahn decomposition
.P; N/. Show that the functions !C and ! defined by
!C .E/ D !.E \ P/
and
! .E/ D !.E \ N/;
for E 2 are measures on .X; / with the following properties:

1. at least one of !C and ! is a finite measure;
2. !.E/ D !C .E/ ! .E/, for every E 2 ;
3. if ; are measures on .X; /, where at least one of which is finite, and if !.E/ D
.E/ .E/ for every E 2 , then !C .E/ .E/ and ! .E/ .E/, for all
E 2 .
Chapter 4
Integration
This chapter is devoted to the main results concerning the Lebesgue integral. There
are many reasons for considering a more robust theory of integration than that
afforded by the classical theories of Cauchy and Riemann, and one of the most
compelling of these reasons arises from the inclination to view the integral as a
continuous linear transformation from a vector space of functions into the real or
complex numbers. Continuity of integration, in this regard, may be considered to be
the property that the integral of a convergent sequence of functions fn is the limit of
the integrals of fn . Limiting properties such as these are possible with the Lebesgue
integral, but not with the Riemann integral in general, and this is one reason why
the Lebesgue integral plays a central role in functional analysis.
The classical approach to the integration of continuous real-valued functions
on a closed interval a; b R is based on a partitioning of the domain, a; b,
into n-subintervals, and to then use Riemann sums to approximate the integral.
Lebesgues approach to the integration of continuous real-valued functions is to
partition the range of a function rather than its domain. This simple idea extends
beyond continuous functions to measurable functions, and leads to a remarkably
powerful theory of integration.
4.1 Integration of Nonnegative Functions

Definition 4.1. Suppose that .X; ; / is a measure space, and that ' W X ! R is a
simple function with range f1 ; : : : ; n g 0; 1/. Let Ek D ' 1 .fk g/, for each k.
1. The canonical form of ' is the representation of ' given by
'D
n
X
k Ek :
(4.1)
kD1

DOI 10.1007/978-3-319-45633-1_4
119
120
4 Integration
2. The Lebesgue integral, or simply the integral, of ' isZ the quantity in the extended
' d and defined by
nonnegative real number system 0; 1 denoted by

X
Z
' d D
n
X
k .Ek / :
(4.2)
kD1
Although 1 is
Z one value that the Lebesgue integral of ' could take , if .X/ is
' d, for every nonnegative simple function '.
finite, then so is
X
The first proposition about integration states that the definition of the integral of
' does not depend on whether or not ' is expressed in canonical form.
Proposition 4.2. If .X; ; / is a measure space and if ' W X ! R is a nonnegative
simple function of the form
'D
m
X
k Fk ;
kD1
where F1 ; : : : ; Fm are pairwise disjoint measurable sets for which F1 [ [ Fm D X,

then
Z
m
X
' d D
k .Fk / :
X
kD1
Proof. If (4.1) denotes the canonical form of ',

n
X
kD1
k Ek D
m
X
k Fk :
kD1
For all pairs .k; j/ 2 f1; : : : ; ng f1; : : : ; mg, consider Akj D Ek \ Fj . The sequence
fAkj gk;j consists of pairwise disjoint measurable sets whose union is X. If x 2 Akj , then
'.x/ D k D j . Thus, the sets f1 ; : : : ; n g and f1 ; : : : ; m g coincide. Moreover,
Z
' d D
X
n
X
k .Ek / D
kD1
D
D
D
n
X
m
X
.Ek \ Fj /
kD1
jD1
n X
m
X
k .Akj /
kD1 jD1
m X
n
X
j .Akj /
jD1 kD1
m
X
j .Fj / ;
jD1
t
u
121
Corollary 4.3. Suppose that .X; ; / is a measure space, and that ';
are nonnegative simple functions.
WX!R
.x/ '.x/, for all x 2 X, then
1. If
Z
d
' d :
2. If ; 2 R are nonnegative, then ' C
is a nonnegative simple function and
.' C / d D
X
' d C
d :
t
u
Theorem 3.14 states that nonnegative measurable functions can be realised as

the limit (pointwise) of a monotone increasing sequence of simple functions. Thus,
simple functions lead the way to the definition of integral for nonnegative functions.
Definition 4.4. If .X; ; / is a measure space and f W X ! R is a nonnegative
measurable function,
Z then the Lebesgue integral, or integral, of f is the quantity in
0; 1 denoted by
f d and defined by
X
Z

' d j ' is simple, and 0 '.x/ f .x/; for all x 2 X :
f d D sup
X
At times one prefers to integrate a function over a measurable subset of X rather

than over the whole space.
Definition 4.5. Let .X; ; / be a measure space and f W X ! R be a nonnegative
measurable function. If
Z E 2 , then the Lebesgue
Z integral
Z of f over E is the quantity
in 0; 1 denoted by
f d D
f d and defined by
E
characteristic function of E.
E f d, where E is the

X
By defining the integral of a nonnegative measurable function as in Definition 4.4, one has the following continuity property.
Theorem 4.6 (Monotone Convergence Theorem). If ffk gk2N is a monotone
increasing sequence of nonnegative measurable functions on a measure space
.X; ; / such that limk fk .x/ exists for all x 2 X, then
Z
Z
lim
k!1 X
fk d D
. lim fk / d :
X k!1
Proof. Let f W X ! R denote the limiting function: f .x/ D limk fk .x/, x 2 X. By

Theorem 3.9, f is measurable. Because fk .x/ fkC1 .x/ f .x/, for all x 2 X,
122
4 Integration
Exercise 4.63 shows that

Z
Z
Z
fk d
fkC1 d
f d ; 8 k 2 N :
X
Thus,
Z
Z
lim
k!1 X
fk d
. lim fk / d :
X k!1
Conversely, let ' W X ! R be a simple function such that 0 '.x/ f .x/, for
every x 2 X. Assume that the canonical form of ' is
'D
n
X
k Ek :
kD1
Fix any 2 .0; 1/ and consider the set

Fk D fx 2 X j fk .x/ '.x/g :
Because Fk D .fk '/1 .0; 1//, Fk is a measurable set. By definition of Fk ,
'.x/ fk .x/ f .x/
8 x 2 Fk :
Hence, again by Exercise 4.63,

Z
Z
Fk 'k d
X
Z
Fk fk d
fk d :
(4.3)
Moreover, because the sequence of functions fk is monotone increasing, the sets Fk

are monotone increasing. In fact,
[
Fk D X :
(4.4)
k2N
To prove (4.4), select x 2 X. Because < 1, we have that '.x/ < '.x/ f .x/.
Furthermore, limk fk .x/ D f .x/ and ffk gk2N is a monotone-increasing sequence of
functions; thus, there
[ is a k0 2 N such that '.x/ < fk0 .x/ f .x/, which implies that
Fk D X.
x 2 Fk0 . Hence,
k2N
The continuity of (Proposition 3.22) and equation (4.4) yield .X/ D

lim .Fk /, and so .Ej / D lim .Ej \ Fk /, for each j D 1; : : : ; n. Furthermore,
k!1
k!1
Z

123
n
n
X
X
j .Ej / D
j lim .Ej \ Fk / D lim
j .Ej \ Fk /
k!1
k!1
jD1 Z
jD1
jD1
Z
Z
D lim
Fk ' d lim
Fk fk d lim
fk d :
' d D
X
n
X
k!1 X
k!1 X
k!1 X
(The final two inequalities above are on account of (4.3).) Therefore,
Z
' d D lim
!1
' d lim
k!1 X
fk d :
Hence,

Z
Z
' d j ' is simple, and 0 '.x/ f .x/; for all x 2 X lim
fk d :
sup
k!1 X
That is,
Z
Z
lim fk d lim
X k!1
k!1 X
fk d ;
thereby completing the proof of the Monotone Convergence Theorem.
t
u
The Monotone Convergence Theorem does not apply to monotone decreasing

sequences, as demonstrated by the example below.
Example 4.7. There is a monotone-decreasing sequence fgk gk2N of Lebesguemeasurable nonnegative functions gk W R ! R such that limk gk .x/ exists for all
x 2 R, but
Z
Z
lim
gk dm 6D
. lim gk / dm:
k!1 R
R k!1
Proof. Let gk D k;1/ . For each x 2 R, g1 .x/ g2 .x/ 0 and limk gk .x/ D 0.
Thus,
Z
Z
. lim gk / dm D 0 dm D 0 :
R k!1
Z
On the other hand,
R
gk dm D 1 for every k 2 N.
t
u
Example 4.7 indicates that the Monotone Convergence Theorem does not extend
to convergent sequences that are not monotone increasing. Nevertheless, there is a
useful partial result, known as Fatous Lemma.
Theorem 4.8 (Fatous Lemma). Suppose that .X; ; / is a measure space and
that fn W X ! R is a measurable nonnegative function for each n 2 N. If lim infn fn .x/
124
4 Integration
exists, for all x 2 X, then

Z
Z
.lim inf fn / d lim inf
n!1
n!1 X
fn d:
(4.5)
Proof. For each k 2 N, let gk .x/ D infffm .x/ j m kg, for every x 2 X. By Theorem 3.9, gk and lim infn!1 fn are measurable. Moreover, fgk gk2N is a monotoneincreasing sequence of nonnegative functions such that
lim gk .x/ D lim inf fn .x/ ;
8x 2 X :
n!1
k!1
Therefore, by the Monotone Convergence Theorem,

Z
Z
lim
k!1 X
gk d D
.lim inf fn / d :
n!1
On the other hand, gk .x/ fk .x/, for all x 2 X, and fgk gk2N monotone increasing
implies that
Z
Z
gk d
X
fm d
8m k;
and so
Z
lim
k!1 X
Z
gk d lim inf
n!1 X
fn d :
t
u
This completes the proof of inequality (4.5).
Corollary 4.9. If ffk gk2N is a sequence of nonnegative measurable functions on a

measure space .X; ; / such that limk fk .x/ exists, for all x 2 X, then
Z
Z
. lim fk / d lim inf
X k!1
k!1 X
fk d :
Basic algebraic properties of the integral can now be investigated,

Theorem 4.10. Suppose that .X; ; / is a measure space and that f and g are
nonnegative measurable functions X ! R. If ; 2 0; 1/, then
Z
.f C g/ d D
X
f d C
X
g d :
X
Proof. By Proposition 3.14, there are monotone increasing sequences f'k gk2N and
f k gk2N of nonnegative simple functions such that, for every x 2 X,
lim 'k .x/ D f .x/
k!1
and
lim
k!1
k .x/
D g.x/ :
125
By the Monotone Convergence Theorem,

Z
Z
f d D lim
X
k!1 X
'k d
Z
g d D lim
and
k!1 X
k d :
(4.6)
Let #k D 'k C k , for every k 2 N. Then f#k gk2N is a monotone-increasing

sequence of simple functions such that
Z
#k d D
X
'k d C
k d
and
lim #k .x/ D .f C g/.x/ ;
8x 2 X :
k!1
The Monotone Convergence Theorem yields

Z
Z
.f C g/ d D lim
k!1 X
#k d ;
which is simplified to
Z
.f C g/ d D
X
f d C
X
g d ;
X
t
u
by the equations in (4.6).
By induction, Theorem 4.10 extends to finite linear combinations of nonnegative

functions via nonnegative coefficients.
Corollary 4.11. Suppose that .X; ; / is a measure space and that f1 ; : : : ; fn are
nonnegative measurable functions X ! R. If 1 ; : : : ; n 2 0; 1/, then
Z
0
1
n
n Z
X
X
@
j fj A d D
j fj d :
jD1
jD1
Another very useful fact is the next result.

Proposition 4.12. Suppose that .X; ; / is a measure space and that f is a
nonnegative measurable function X ! R. If E1 ; : : : ; En 2 are pairwise disjoint,
then
Z
f d D
E1 [ [En
n Z
X
jD1
f d :
Ej
126
4 Integration
Proof. Let E D E1 [ [En and note that E f D

Z
E f d D
X
n Z
X
jD1
Ej f . Hence, by Corollary 4.11,
jD1
Z
Ej f d. That is,
n
X
f d D
n Z
X
E1 [ [En
f d.
t
u
Ej
jD1
Corollary 4.13. Suppose that .X; ; / is a measure space and that f is a

nonnegative measurable function X ! R. If E 2 is such that .E/ D 0, then
Z
Z
f d D 0
Z
f d D
and
f d :
XnE
Proof. Because .E/ D 0, .E \ F/ D 0 for every F 2 . Hence, if ' is a simple

n
X
k Ek , then
function in canonical form ' D
kD1
Z
' d D
E
n
X
k .E \ Ek / D 0 :
kD1
Thus, by the definition of integral of f as the supremum of a set of integrals of simple

functions, one concludes immediately that
Z
f d D 0 :
E
This fact, together with Theorem 4.12, shows that

Z
f d D
X
f d C
XnE
Z
f d D
f d ;
XnE
t
u
4.2 Integrable Functions

Definition 4.14. If X is a set, and if f W X ! R is a function, then the Jordan
decomposition of f is the equation
f D f C f ;
(4.7)
where
fC D
jf j C f
2
and
f D
jf j f
:
2
(4.8)
127
Observe that f C and f are nonnegative functions such that

f C f D 0 and f C C f D jf j:
Moreover, if .X; / is a measurable space, and if f W X ! R is a measurable function,
then f C and f are measurable functions.
Using the Jordan decomposition, the integral of any measurable function can now
be defined.
Definition 4.15. If .X; ; / is a measure space, and if f W X ! R is a measurable
function, then the Lebesgue integral, or simply the integral,
of f is the quantity in
Z
the extended real number system 1; 1 denoted by
f d and defined by
X
Z
f d D
X
f d;
f d
X
where f D f C f is the Jordan decomposition of f .
f d is a
f d
Our interest is with the class of functions for which

X
(finite) real number.
Definition 4.16. If .X; ; / is a measure space, and if f W X ! R is a measurable

function, then f is an integrable function if
Z
f d < 1:
f d < 1 and
X
Proposition 4.17. Assume that .X; ; / is a measure space and f ; g W X ! R are

integrable functions. For all ; 2 R, the function f C g is integrable, and
Z
.f C g/ d D
X
f d C
X
g d :
X
Proof. Let h1 ; h2 W X ! R be nonnegative, integrable functions, and let h D h1 h2 .

It may be that hC 6D h1 and h 6D h2 ; nevertheless, it is true that
Z
h d D
X
h1 d
X
h2 d :
(4.9)
To prove this, express h in its Jordan decomposition: h D hC h . Because hC

h D h1 h2 , we have that hC C h2 D h C h1 . By additivity of the integral for
nonnegative functions (Theorem 4.10),
Z
hC d C
X
Z
h2 d D
X
h d C
Z
h1 d :
X
128
4 Integration
As all four integrals above are finite, by the hypothesis that h, h1 , and h2 are
integrable,
Z
hC d
X
h d D
Z
h1 d
h2 d ;
X
which proves equation (4.9).

The argument above applies to f C g via
f C g D .f C C gC / .f C g / :
(This need not be the Jordan decomposition .f C g/C .f C g/ .) Lastly, now that
we know that
Z
Z
Z
.f C g/ d D f d C g d ;
X
the case of f Cg is handled by expressing each scalar in its Jordan decomposition:

t
u
D C and D C .
By Proposition 4.17, the set of all integrable functions has the structure of a
vector space.
Corollary 4.18. If .X; ; / is a measure space, then the set LR1 .X; ; / of all
integrable
functions f W X ! R is a vector space over the field R, and the map
Z
f 7!
X
f d is a linear transformation of LR1 .X; ; / onto R.
The next result gives a characterisation of integrability in terms of the absolutevalue function.
Proposition 4.19. Assume that .X; ; / is a measure space and f W X ! R is a
measurable function. Then f is integrable if and only if jf j is integrable. Moreover,
if f is integrable, then the following triangle inequality holds:
Z
f d
jf j d :
Proof. Assume that f is integrable. Thus,

jf j D f C C f , Theorem 4.17 shows that
Z
jf j d D
X
That is, jf j is integrable.
Xf
.f C f / d D
X
d and
Xf
d are finite. Because
f d < 1 :
f d C
X
129
Conversely, assume that jf j is integrable. The equation jf j D f C C f yields

f .x/ jf j.x/ and f .x/ jf j.x/, for all x 2 X. Thus,
C
f C d
Z
jf j d < 1
and
f d
Z
jf j d < 1 :
X
Hence, f is integrable.
To prove the triangle inequality, assume that f is integrable. The triangle
inequality in real numbers and is j C j jj C jj. Applying this to integrals
leads to
Z
Z
Z
Z
Z
Z
f d D f C d f d f C d C f d jf j d;
t
u
Proposition 4.19 above demonstrates the integrability condition is similar to that

of absolute convergence in the theory of series. Indeed, if one considers the measure
space .N; P.N/; /, where is the counting measure, then a function f W N ! R
has Lebesgue integral
Z
1
X
f d D
N
f .k/;
kD1
and so f is integrable if and only if

1
X
jf .k/j < 1 :
kD1
Theorem 4.20 (Dominated Convergence Theorem). Suppose that .X; ; / is a

measure space, and that ffk gk2N is a sequence of measurable functions fk W X ! R
such that lim fk .x/ exists, for all x 2 X. If there is a nonnegative integrable function
k!1
g W X ! R such that jfk .x/j g.x/, for every k 2 N and x 2 X, and if f W X ! R

is the measurable function defined by f .x/ D lim fk .x/, for every x 2 X, then f is
k!1
integrable, and
Z
Z
lim
k!1 X
fk d D
f d and lim
k!1 X
jfk f j d D 0:
Proof. Because f .x/ D limk fk .x/, for all x 2 X, f is measurable and jf .x/j g.x/ for
every x 2 X. Thus,
Z
Z
jf j d
X
g d < 1 ;
X
130
4 Integration
which implies that f is integrable. For each k 2 N, let hk D 2g jfk f j. Note that the
triangle inequality in real numbers gives jfk f j.x/ 2g.x/, and so the functions hk
are nonnegative. Moreover, limk jfk f j.x/ D 0; thus, limk hk D 2g. Fatous Lemma
yields
Z
Z
2g d D
.lim inf hk / d
k!1
Z
lim inf
k!1 X
hk d
Z
D
2g d C lim inf
k!1
Z

jfk f j d
X
Z
2g d lim sup
jfk f j d :
k!1 X
Therefore,
Z
0 lim sup
jfk f j d 0 :
(4.10)
k!1 X
If a sequence fk gk of positive numbers does not converge to 0, then it must

necessarily be that lim supk k > 0. Thus, inequality (4.10) implies that
Z
0 lim
k!1 X
jfk f j d 0 :
For each k 2 N,
Z
Z
Z
fk d f d
jfk f j d :
Hence, as k ! 1 we obtain
Z
Z
lim
k!1 X
fk d D
f d ;
X
thereby completing the proof.
t
u
Some of the convergence results in integration can be reformulated for convergence on complements of sets of measure zero.
Definition 4.21. If f ; g W X ! R are measurable functions on a measure space
.X; ; /, then f D g almost everywhere if
.fx 2 X j f .x/ 6D g.x/g/ D 0:
131
That is, f and g are equal almost everywhere if the set on which they are not
equal is a set of measure zero. This is of significance in integration because of
Corollary 4.13, which states that if .X; ; / is a measure space and if f is a
nonnegative measurable function X ! R, then, for any E 2 such that .E/ D 0,
Z
Z
Z
f d D 0 and
f d D
f d :
E
XnE
By passing to differences of positive functions, the line above extends to any realvalued integrable function.
The point
is this:
Z
Z sets of measure zero have no role in the value of the integral.
f d D
That is,
X
g d, if f D g almost everywhere.

X
Definition 4.22. If .X; ; / is a measure space and if fk W X ! R is measurable

function, for all k 2 N, then the sequence ffk gk2N converges almost everywhere if

fx 2 X j lim fk .x/ does not exist g D 0 :
k!1
The following result, which is a partial converse to the Monotone Convergence

Theorem, demonstrates one way in which almost everywhere convergence can arise.
Proposition 4.23. Suppose that .X; ; / is a measure space, and that ffk gk2N is
a monotone
increasing sequence of nonnegative integrable functions fk W X ! R.
Z
If lim
fk d exists, then there is a

k!1 X
lim fk .x/ exists for all x 2 XnD, and
k!1
measurable set D X such that .D/ D 0,
lim
k!1 X
fk d D
Z
. lim fk / d :
XnD k!1
Proof. Because ffk gk2N is a monotone-increasing sequence,

Z
Z
fk d
fkC1 d;
X
for every
k 2 N. The hypothesis implies the existence of a real number W > 0 such
Z
fk d W for every k.
that
X
Let D D fx 2 X j limk fk .x/ does not existg. Then D D XnL, where L is the set of
points x 2 X for which limk fk .x/ exists. By Theorem 3.9, L 2 and, hence, D 2 .
Let " > 0 be arbitrary and define, for each k 2 N, Dk D fx 2 X j fk .x/ > W="g. Because
the sequence ffk gk is monotone increasing, so is the sequence fDk gk . If x 2 D, then
there is a k0 2 N such that fk0 .x/ > W=", and so x 2 Dk0 . Hence,
D
[
k2N
Dk :
(4.11)
132
4 Integration
Therefore, the continuity of the measure (Proposition 3.22) applied to the

inclusion (4.11) yields
!
[
.D/
Dk D lim .Dk / :
k!1
k2N
But, for each k 2 N,
.Dk / D
Dk d
X
"
W
Z
Dk fk d
X
"
W
Z
fk d ":
X
Thus, .D/ D 0.
Let f W X ! R be given by f .x/ D 0 for x 2 D and f .x/ D limk!1 fk .x/ for x 62 D.
The sequence fXnD fk gk2N is monotone increasing and f D limk XnD fk . Thus, f is
measurable and, by the Monotone Convergence Theorem,
Z
Z
lim
XnD fk d D f d :
k!1 X
g d D 0, because .D/ D 0. Thus,
For every nonnegative measurable function g,

Z
Z
lim
k!1 X
fk d D lim
fk d C
k!1
XnD

Z
fk d D lim
k!1 XnD
Z
D lim
k!1 X
XnD fk d D
fk d
Z
f d D
. lim fk / d;
XnD k!1
t
u
It is somewhat clumsy to make explicit reference to the set D of measure zero in

the integral
Z
. lim fk / d ;
XnD k!1
especially as null sets do not contribute to the value of the integral. Therefore, a
notational convention is adopted: if ffk gk2N converges almost everywhere, then
Z
. lim fk / d
X k!1
is understood to represent the quantity

Z
. lim fk / d ;
XnD k!1
where D X is the set (of measure zero) of points for which ffk .x/gk2N has no limit.
4.3 Complex-Valued Functions and Measures
133
With this notational convention, the Monotone Convergence Theorem and the
Dominated Convergence Theorem have versions for almost-everywhere convergence. For example, the Dominated Convergence Theorem is formulated as follows.
Theorem 4.24. Suppose that .X; ; / is a measure space, and that ffk gk2N is a
sequence of integrable functions such that lim fk .x/ exists almost everywhere. If
k!1
there is a nonnegative integrable function g W X ! R such that, for every k 2 N and

x 2 X,
jfk .x/j g.x/;
then
Z
Z
lim
k!1 X
fk d D
. lim fk / d :
X k!1
t
u

The goal of this section is to outline how integration of real-valued functions extends
to integration of complex-valued functions by using the real and imaginary parts of a
complex-valued function. The main item of note is that the Dominated Convergence
Theorem extends to complex-valued functions.
The role of signed and complex measures in integration theory is less prominent
than the role played ordinary measures. However, our study of Banach space duality
will require integration to be extended so as to encompass both complex-valued
functions and complex measures. Therefore, this section concludes with a brief
description of how such an extension is formulated and achieved.
Definition 4.25. If .X; / is a measurable space, then a function f W X ! C is said
to be complex measurable if f 1 .U/ 2 , for all open subsets U C.
If f W X ! C, then let f and jf j denote the functions defined by f .x/ D f .x/ and
jf j.x/ D jf .x/j, for all x 2 X. As with complex numbers, a function f can be expressed
in its real and imaginary parts:
f D <f C i=f ;
where
1
1
<f D .f C f / and =f D .f f / :
2
2i
Note that <f and =f are real-valued functions on X.

Proposition 4.26. If .X; / is a measurable space, then a function f W X ! C is
complex measurable if and only if <f and =f are measurable functions X ! R.
134
4 Integration
t
u
Definition 4.27. If .X; ; / is a measure space, then f W X ! C is integrable if <f

and =f are integrable functions X ! R, and the integral of f is defined by
Z
Z
f d D
X
Z
<f d C i
=f d :
X
Note that <f and =f are integrable only if <f and =f are measurable. In such
cases, f is automatically complex measurable (by Proposition 4.26), and this is why
complex measurability is not mentioned in the definition of integrability.
Proposition 4.28. If .X; ; / is a measure space, then f W X ! C is integrable if
and only if jf j is integrable, in which case
Z
f d
jf j d :
Proof. Write f in terms of its real and imaginary parts: f D <f C i=f . If f is
integrable, then each of <f and =f is integrable (by definition). Hence, j<f j and
j=f j are integrable. Because jf j j<f j C j=f j, one concludes that jf j is integrable.
Conversely, assume that jf j is integrable. Since j<f j jf j and j=f j jf j, both j<f j
and j=f j are integrable. Thus, f is integrable.
To show the triangle inequality, assume that f is integrable. For any complex
number , a complex number ! satisfies j!j jj if and only if ei <! jj for all
2 0; 2
. Note that ei < f D <.ei f / jei f j. Thus,
Z
ei <
Z
Z
Z
f d D <.ei f / d
jei f j d D jf j d ;
for all 2 0; 2
. Hence, f d
jf j d.
X
t
u
Finally, the Dominated Convergence Theorem extends to complex-valued functions.

Theorem 4.29 (Dominated Convergence Theorem: Complex Case). Suppose
that .X; ; / is a measure space, and that ffk gk2N is a sequence of measurable
functions fk W X ! C such that lim fk .x/ exists, for all x 2 X. If there is a nonnegative
k!1
integrable function g W X ! R such that jfk .x/j g.x/, for every k 2 N and x 2 X,
and if f W X ! C is the measurable function defined by f .x/ D lim fk .x/, for every
k!1
x 2 X, then f is integrable, and

Z
Z
lim
k!1 X
fk d D
f d and lim
X
k!1 X
jfk f j d D 0:
135
If ! is a signed measure on a measurable space .X; /, then expressing ! in

its Jordan Decomposition (see Theorem 3.66) as a difference of two measures, !C
and ! , allows one to define the integral of a measurable function f W X ! C with
respect to the signed measure ! by
Z
f d! D
X
as each of
Xf
d!C and
Xf
f d!C
f d! ;
d! is well defined.
Definition 4.30. If ! is a signed measure on .X; / with Jordan decomposition

! D !C ! , then a measurable function f W X ! C is integrable if
Z
Z
jf j d!C and
X
jf j d!
X
are finite.
In defining j!j by j!j D !C C ! , we obtain a measure j!j on .X; / and for
each integrable function f , we have a triangle inequality:
Z
Z
Z
Z
Z
Z

f d! D f d!C f d! jf j d!C C jf j d! D jf j dj!j:
One can approach integration with respect to a complex measure in a similar

way, by first considering the signed measures < and = induced by the real and
imaginary parts of . In so doing, two finite signed measures are obtained, each
of which is a difference of finite measures. Hence, there are finite measures j on
.X; /, for j D 1; : : : ; 4, such that
D .1 2 / C i.3 4 /:
The integral of a measurable function f W X ! C with respect to the complex measure
is defined by
Z
Z
f d D
X
Z
f d.1 2 / C i
f d.3 4 /:
X
Definition 4.31. If is a complex measure on .X; /, expressed as D .1

2 / C i.3 4 /, where 1 2 and 3 4 are Jordan decompositions of the
real and imaginary parts of , respectively, then a measurable function f W X ! C is
integrable if
Z
jf j dj < 1
X
for each j D 1; : : : ; 4.
136
4 Integration
A version of the triangle inequality is noted in Exercise 4.72.

We will not have any significant need of such integralsZin this book, except for
f d with respect to a
in the discussion of duality, where the use of the integral

X
complex measure is convenient.
4.4 Continuity
A starting point for our discussion of continuity is the following important observation:
Proposition 4.32. If .X; ; / is a measurable space and g W X ! R is a nonnegative measurable function, then the function g W ! R defined by
Z
g .E/ D g d; E 2 ;
E
is a measure on .X; /.
t
u
Proposition 4.32 shows that there are an abundance of measures one can
introduce on a measurable space, and that finite measures g are achieved from
using nonnegative integrable functions g, even if itself is not finite. In this case, as
shown by the next result, the pair of measures and g exhibit a continuity feature
(see Proposition 4.37 as well).
Proposition 4.33. If f W X ! C is an integrable function on
Z a measure space
.X; ; /, then for every " > 0 there exists a > 0 such that jf j d < " for all
E
measurable sets E X with .E/ < .
Proof. For each n 2 N, let En D fx 2 X j jf .x/j < ng and define gn W X ! R by

gn D jf jEn C nEnC :
Thus, fgn gn2N is a monotone increasing sequence of nonnegative measurable
functions with limn gn .x/R D jf .x/j for every x 2 X. Therefore, by the Monotone
Convergence
Theorem, f X gn dgn2N is a monotone increasing sequence in R with
R
limit X jf j d. Hence, if " > 0, then there is a K 2 N such that
Z
Z
"
jf j d < gK d C :
3
X
X
Thus,
Z
Z
jf j d <
Z
gK d C
EK
"
gK d C ;
C
3
EK
4.4 Continuity
137
which implies that

Z
jf j d <
E
Z
gK d C
E\EK
C
E\EK
gK d C
"
"
K.E/ C K.E/ C :
3
3
Z
jf j d < ".
Therefore, if D "=.3K/ and .E/ < , then
t
u
For integrable function f on a measure space .X; ; /, the induced measure jf j
on .X; / is finite and has the continuity property exhibited in Proposition 4.33
above. This property of a pair of such measures can be characterised in the abstract
in the following way.
Proposition 4.34. The following statements are equivalent for measures and ,
where is finite, on a measurable space .X; /:
1. for every " > 0 there exists a > 0 such that .E/ < " for each E 2 that satisfies
.E/ < ;
2. .E/ D 0 for each E 2 that satisfies .E/ D 0.
Proof. The proof of (1) ) (2) is left as an exercise. To prove the converse, we
shall prove the contrapositive: if there exists a " > 0 such that for each > 0 there
exists a set E 2 with .E / < and .E / ", then there is also a set E 2
for which .E/ D 0 and .E/ > 0. To this end, assume such a " > 0 exists and for
k
each
[ k 2 N let Ek 2 satisfy .Ek / < 2 and .Ek / ". For every n 2 N, let An D
Ek so that fAn gn2N is a descending sequence of measurable sets. Observe that
kn
.An /
1
X
.Ek /
kDn
1
X
2k D 2n1 < 2n and .An / .En / ". In particular,
kDn
.A1 / < 1 and, by hypothesis, .A1 / < 1. Therefore, if E D
An , then
n2N
.E/ D lim .An / lim 2n D 0 and

n!1
n!1
by continuity of measure (Proposition 3.22).
.E/ D lim .An / ";

n!1
t
u
The second of the equivalent conditions in Proposition 4.34 is called absolute

continuity, the relevance of which will be explained by the Radon-Nikodm
Theorem.
Definition 4.35. If .X; / is a measurable space and if and are measures on
.X; /, then is absolutely continuous with respect to if .E/ D 0 for every
E 2 for which .E/ D 0.
Theorem 4.36 (Radon-Nikodm). Suppose that and are finite measures on a
measurable space .X; /. If is absolutely continuous with respect to , then there
is a nonnegative measurable function g W X ! R such that
138
4 Integration
Z
.E/ D
g d ;
8 E 2 :
Proof.
Let G be the set of all measurable functions f W X ! 0; 1/ for which
Z
f d .E/ for every E 2 . The set G is nonempty (because 0 2 G / and
E
max.fR 1 ; f2 / 2 G for all f1 ; f2 2 G . Because .X/ < 1, the supremum of the

set f X f d j f 2 G g exists. In particular, for each n 2 N there exists fn 2 G such
R
that X fn d > 1n . Define a sequence fgn gn2N Zby gn D max.f1 ; : : : ; fn / and
note that fgn gn2N is monotone increasing and lim
n!1 X
gn d D . Therefore, by
Proposition 4.23, there exists a measurableZ function g W X ! 0; 1 such that

g.x/ D limn gn .x/ for almost every x 2 X and g d D .
X
Z
Z
Because g d D sup gn d .E/ for every E 2 , the function Q W ! R
defined by
Z
.E/
Q
D .E/
g d
E
is a finite measure. We aim to show that Q is identically zero. Therefore, assume that
Q is not the zero measure. Thus, .X/
Q
> 0. Fix " > 0 such that .X/
Q
> ".X/ and
let ! D Q ". Because ! is a difference of finite measures, ! is a signed measure.
Furthermore, !.X/ > 0.
The Hahn Decomposition (Theorem 3.65) asserts that there exist P; N 2 such
that P is positive with respect to !, N is negative with respect to !, P \ N D ;,
and X D P [ N. Thus, for E 2 , 0 !.E \ P/ implies that ".E \ P/ .E
Q \ P/.
Hence,
Z
Z
.E/ D .E/
Q
C g d ".E
Q \ P/ C g d
E
".E
Q \ P/ C
g d D
E
.g C "P / d:
E
The inequality above shows that .g C "P / 2 G . However,

Z
Z
.g C "P / d >
X
g d
X
Z
g d D . Therefore, it must be that .E/
Q
D 0 for every
contradicts the fact that

E 2 .
t
u
4.5 Fundamental Theorem of Calculus
139

As in calculus, Lebesgue integrals display certain continuity and differential
properties that facilitate calculations. The first of these properties is the following
familiar continuity feature.
Proposition 4.37. If f W a; b ! C is integrable with respect to Lebesgue measure,
then the function F W a; b ! C defined by
Z
F.x/ D
f dm;
a;x
for x 2 a; b, is continuous on a; b.

Proof. Choose
x0 2 a; b and let " > 0. By Proposition 4.33, there exists a > 0
Z
such that jf j dm < " if E a; b is a measurable set with m.E/ < . If x 2 a; b
E
satisfies jx x0 j < and if x0 x, then

Z
Z
Z
f dm
f dm D
jF.x/ F.x0 /j D
a;x
a;x0
x0 ;x
f dm <
x0 ;x
jf j dm < ":
The same inequality holds if x x0 by replacing x0 ; x with x; x0 . Hence, for each

" > 0, there exists > 0 such that jF.x/ F.x0 /j < " whenever x 2 a; b satisfies
jx x0 j < .
t
u
Continuity of the integrand at a point leads to differentiability of the integral at
that same point.
Proposition 4.38. If f W a; b ! R is integrable with respect to Lebesgue measure,
and if f is continuous at x0 2 .a; b/, then the function F W a; b ! R defined by
Z
f dm;
F.x/ D
a;x
for x 2 a; b, is differentiable at x0 and

dF
.x0 / D f .x0 / :
dx
Proof. Let " > 0. The continuity of f at x0 implies the existence of a > 0 with the
properties that .x0 ; x0 C / .a; b/, and that jf .x/ f .x0 /j < ", for all jx x0 j < .
If > h > 0 (the case where h is negative is similar), then
140
4 Integration
jF.x0 C h/ F.x0 / hf .x0 /j D
x0 ;x0 Ch
f dm hf .x0 /
Z
f dm
x0 ;x0 Ch
x0 ;x0 Ch
f .x0 / dm
Z

x0 ;x0 Ch
jf f .x0 /j dm
< "h :
(For the final inequality: jx x0 j h < implies that jf .x/ f .x0 /j < ".) Dividing
the inequalities above by h yields shows that each " > 0 there is a > 0 such that
jhj <
H)
F.x0 C h/ F.x0 /
f .x0 / < " :
dF
.x0 / D f .x0 /.
t
u
dx
The results above allow for the calculation of integrals in the style of Cauchy, via
antiderivatives.
That is, F is differentiable at x0 and
Theorem 4.39 (Fundamental Theorem of Calculus). If F W a; b ! R is differentiable on .a; b/, and if dF=dx is integrable and continuous on .a; b/, then
Z
dF
dm D F./ F.a/;
a; dx
(4.12)
for every 2 a; b.

Proof. Let G W a; b ! R be the function
Z
G./ D
dF
dm;
a; dx
for 2 a; b. Note that G.a/ D 0 and that dG

D dF
at every point of .a; b/ (by the
dx
dx
dF
hypothesis that dx is continuous on .a; b/ and by Proposition 4.38). Let H D G F
and fix 2 .a; b. By the Mean Value Theorem in differential calculus, there is a
point x0 2 .a; b/ such that
H./ H.a/
dH
D
.x0 / :
a
dx
However,
dH
dx
dG
dx
dF
implies that
dx
dH
dx
is identically zero on .a; b/. Thus,
141
0 D H./ H.a/ D G./ F./ .G.a/ F.a//

D G./ F./ C F.a/
Z
D F./ C F.a/ C
dF
dm :
a; dx
t
u
That is, equation (4.12) holds.
Corollary 4.40. If F W a; b ! R is differentiable on .a; b/ and if dF=dx is

continuous on .a; b/, then
Z
dF
dm D F.b/ F.a/ :
a;b dx
(4.13)
Moreover, if f W a; b ! R is continuous, then

Z
f dm D
a;b
f .x/ dx ;
(4.14)
where the integral on the right is the CauchyRiemann integral.

Proof. Equation (4.13) follows from Theorem 4.39.
Assume now that f is continuous. Define F by
Z
F.x/ D
a;x
f dm ;
x 2 a; b :
Then F is differentiable on .a; b/ and dF=dx D f (by Proposition 4.38). Hence,

Z
F.x/ D
dF
dm D
a;x dx
Z
a;b
f dx D F.b/ F.a/ :
Equation (4.13) for the Lebesgue integral is precisely what the Fundamental
Theorem of Calculus yields for the CauchyRiemann integral, namely that
Z
a
This completes the proof.
dF
dx D F.b/ F.a/ :
dx
t
u
We note below that the Fundamental Theorem of Calculus does not admit an
almost everywhere version.
142
4 Integration
Example 4.41. The Cantor ternary function is differentiable at almost every

point of 0; 1, and d
D 0 almost everywhere, yet
dx
Z
d
dm 6D .1/ .0/:
0;1 dx
Proof. The function is constant on any open interval .; / 0; 1nC . Hence,
the derivative of is zero on .; /. As the Cantor set C is a set of measure zero,
is differentiable at almost every point of 0; 1. Therefore, the integral of d
is zero,
dx
whereas .1/ .0/ D 1 0 D 1, and so the Fundamental Theorem of Calculus
does not hold.
t
u
4.6 Series
Proposition 4.42. On a measure space .X; ; /, assume that fun gn2N is a
1
X
un .x/ converges for
sequence of nonnegative integrable functions such that
all x 2 X. If
1
X
nD1
un is integrable, then
nD1
1
X
Proof. Let g D
1
X
un d D
1 Z
X
un d :
(4.15)
nD1 X
nD1
un . If fk D
nD1
k
X
, for each k 2 N, then fk .x/ g.x/ for all x 2 X.
nD1
Because g is integrable, the Dominated Convergence Theorem states that

Z
Z
. lim fk / d D lim
X k!1
k!1 X
fk d ;
t
u
which is precisely equation (4.15).

A partial converse to Proposition 4.42 is:
Proposition 4.43. If .X; ; / is a measure space and if fun gn2N is a sequence

1
1
X
X
of nonnegative integrable functions such that
un is integrable, then
un .x/
nD1
converges for almost all x 2 X and

Z
1
X
X
nD1
!
un d D
1 Z
X
nD1 X
nD1
un d :
4.6 Series
143
Proof. Apply the Monotone Convergence Theorem and Proposition 4.23 to the
partial sums of the series.
u
t
The following example demonstrates how Propositions 4.42 and 4.43 can be
invoked for calculations.
Example 4.44. If log x denotes the natural logarithm function, and if f W 0; 1 ! R
is defined by
f .0/ D 0 I
f .1/ D 1 I
x log x
8 x 2 .0; 1/ ;
x1
f .x/ D
then f is continuous, and

Z
1
0
f .x/ dx D 1
1
X
1
:
2 .n 1/
n
nD2
Proof. Note that f is continuous on .0; 1/ and that

lim f .x/ D 0
lim f .x/ D 1 :
and
x!0C
x!1
Thus, f is continuous on 0; 1 and therefore

Z
0
Z
f .x/ dx D
0;1
f dm :
Let fk D 1 ;1 1 f , for all k 2 N. The sequence ffk gk2N of nonnegative measurable
k
k
functions is monotone increasing and limk fk D f . By the Monotone Convergence
Theorem,
Z
f dm D lim
0;1
k!1 0;1
fk dm D lim
k!1
1 1k
1
k
x log x
dx :
x1
Change variables: let ! D 1 x to obtain

Z
1 1k
1
k
x log x
dx D
x1
1 1k
1
k

1!
log.1 !/ d! :
!
The function log.1 !/ has a power series expansion that converges (uniformly,
although we need only pointwise) on 1k ; 1 1k . Thus,
144
4 Integration

1
1 ! X !n
1!
log.1 !/ D
!
! nD1 n
D 1
1
X

! n1
nD2
D 1
1
1

n1 n
1
X
! n1
:
.n 1/n
nD2
By Proposition 4.42, this final series can be integrated term by term, which yields
Z
1 1k
1
k
1!
log.1 !/ d! D
!
D
X

Z 1 1
1
k
2
1

1
! n d!
1
k
.n

1/n
k
nD2
X

1
.1 1k /n . 1k /n
2
:

1
k
n2 .n 1/
nD2
Thus,
Z
f dm D lim
0;1
k!1
D 1
1
2 X .1 1k /n . 1k /n
1
k nD2 n2 .n 1/
1
X
nD2
n2 .n 1/
;
t
u
as claimed.
4.7 Integral Inequalities

If x and y are elements in a real or complex vector space V, then the line segment
Lx;y in V that joins x with y has parametric form
Lx;y D f x C .1 /y j 2 0; 1g:
This is, of course, only one of infinitely many ways to parameterise the line segment
Lx;y , but it is perhaps the simplest one. In particular, if J is an interval of real
numbers, then Lx;y J for all x; y 2 J.
Definition 4.45. If J R is an interval, then a function # W J ! R is:
145
1. a convex function if, for all x; y 2 J and 2 .0; 1/,

# . x C .1 /y/ #.x/ C .1 /#.y/ I
2. a concave function if, for all x; y 2 J and 2 .0; 1/,
#.x/ C .1 /#.y/ # . x C .1 /y/ :
Note that a function # is convex if and only if # is concave. The general shape
of the graph of a convex function on an interval J R is concave up, whereas
the shape of concave functions is concave down. This is made precise by the
following proposition.
Proposition 4.46. If J R is an open interval, and # W J ! R has a continuous
nonnegative second derivative at every point of J, then # is a convex function.
Proof. Assume that d2 #=dt2 is nonnegative on J. Then, d#=dt is monotone
increasing on J. To prove that # is convex, choose any x; y 2 J and 2 .0; 1/. Let
D x C .1 /y. By the Fundamental Theorem of Calculus, and by the fact that
d#=dt is monotone increasing,
#./ #.x/ D
Z
d#
dt

d#
. x/ :
dt

d#
.y / :
dt
dt
Likewise,
#.y/ #./ D
Z y

d#
dt

dt
Because x D .1 /.y x/ and y D .y x/, the second terms in each of the

two inequalities above can be expressed in terms of .y x/, leading to

d#
.1 /.y x/
dt

d#
.y x/ :
dt
#./ #.x/ C
#./ #.y/
Hence,

#./ #.x/ C

d#
.1 /.y x/
dt
.1 /#./ .1 /#.y/

d#
.1 /.y x/ :
dt
146
4 Integration
Adding these two inequalities leads to

#. x C .1 /y/ #.x/ C .1 /#.y/ :
This proves that # is a convex function.
t
u
Proposition 4.46 allows one to readily produce examples of convex functions.

Among the most important convex functions are the ones below.
Corollary 4.47. The following functions # are convex:
1. #.t/ D et on R, for any 2 R;
2. #.t/ D tp on .0; 1/, where p 2 R is such that p 1;
3. #.t/ D log t on .0; 1/.
Proposition 4.48. If # W J ! R is a convex function, and if x; y; z 2 J satisfy x < y <
z, then
#.y/ #.x/
#.z/ #.y/

:
yx
zy
(4.16)
Proof. There is a unique 2 .0; 1/ such that y D x C .1 /z. Thus,

zD
y x
1
and
zyD

.y x/ :
1
Because # is a convex function, #.y/ #.x/ C .1 /#.z/, and so

.#.z/ #.x// #.z/ #.y/ :
That is,
#.z/ #.x/
1
.#.z/ #.y// :

Hence,
#.y/ #.x/
#.x/ C .1 /#.z/ #.x/

yx
yx
D

D
D
.1 / .#.z/ #.x//
yx
.1 /

.#.z/ #.x//
yx
#.z/ #.y/

.y
1
x/
#.z/ #.y/
:
zy
t
u
147
If one views each side of inequality (4.16) as a difference quotient, then

inequality (4.16) says that the derivative of # (if it exists) is an increasing function.
The defining condition for a convex function is an inequality. Therefore, it is not
surprising that convex functions lead to a variety of inequalities, and one of the most
fundamental and generic of such inequalities is the following inequality of Jensen.
Theorem 4.49 (Jensens Inequality). Suppose that .X; ; / is a measure space
such that .X/ D 1. If f W X ! a0 ; b0 .a; b/ is a measurable function, then, for
every convex function # W .a; b/ ! R,
Z
#
Z
f d
X
Proof. Note that
Xf
# f d :
X
d 2 .a; b/ because
a < a0 f .x/ b0 < b ;
8x 2 X ;
implies that
Z
a d <
aD
X
f d <
X
b d D b :
X
Z
Let D
f d and let
X
D sup
z2.a;/
#./ #.z/
:
z
Hence,
#.z/ #./ C .z / ;
8 z 2 .a; / :
Because # is a convex function, Proposition 4.48 implies that

#.y/ #./
;
y
8 y 2 .; b/ :
Thus,
#.y/ .y / C #./ ;
8 y 2 .; b/ :
#.t/ #./ C .t / ;
8 t 2 .a; b/ :
Conclusion:
In particular,
# .f .x// #./ f .x/ C 0 ;
8x 2 X :
(4.17)
148
4 Integration
Z
On passing to integrals, and noting that .X/ D 1 and D
f d, inequality (4.17)

X
yields
Z

f d C 0 :
# f d #
X
That is,
Z
Z
f d
X
# f d ;
X
t
u
Some very basic yet far reaching inequalities may be derived from Jensens
inequality. Among the most elementary of these are the arithmetic-geometric mean
inequality and Youngs inequality.
Proposition 4.50. Suppose that 1 ; : : : ; n are positive real numbers.
1. Arithmetic-Geometric Mean Inequality:
.1 n /1=n
1
.1 C C n / :
n
2. Youngs Inequality: If p1 ; : : : ; pn are positive and

1 n
1
p1
C C p1n D 1, then
1 p1
1
C C npn :
p1
pn
Proof. For the proof of the Arithmetic-Geometric Mean Inequality, let X D

f1; 2; : : : ; ng, D P.X/, and be
.E/ D
1
jEj ;
n
8E X ;
where jEj denotes the cardinality of E. Thus, .X/ D 1.

Let 1 ; : : : ; n 2 R be such that ek D k , for each k. If f W X ! R is defined by
f .k/ D k , for k 2 X, then
Z
f d D
X
n
X
1X
k :
n kD1
n
f .k/.fkg/ D
kD1
The function #.t/ D et is convex on R and

Z
X
1 X f .k/ 1 X k
e D
e :
n kD1
n kD1
n
# f d D
149
Jensens inequality states that

Z
Z
f d
# f d :
#
X
Hence,
1 X k
1X
e D
k :
n kD1
n kD1
n
.1 2 n /1=n D e1 =nC Cn =n

The proof of Youngs inequality is left to the reader as Exercise 4.87.
t
u
Proposition 4.50 is a rather straightforward application of Jensens inequality.

A somewhat more sophisticated application of Jensens inequality leads to the
fundamental inequalities of Hlder and Minkowksi (Theorems 4.53 and 4.54).
Definition 4.51. If .X; ; / is a measure space and if p 1, then a measurable
function f W X ! C is said to be p-integrable if f p is integrable.
Definition 4.52. Two positive real numbers p; q 2 R are said to be conjugate if
1 1
C D 1:
p q
If p and q are conjugate real numbers, then q to p is uniquely determined by p, and
p > 1 and q > 1. The notion of conjugate positive numbers is nothing more than
the association of a pair of convex coefficients, namely 1p and 1q , with the positive
real numbers p and q.
Theorem 4.53 (Hlders Inequality). Suppose that p and q are conjugate real
numbers. If .X; ; / is a measure space and if f ; g W X ! C are such that f is pintegrable and g is q-integrable, then fg W X ! C is integrable and
Z
Z
jfgj d
jf j d
1=p Z
jgj d
1=q
(4.18)
Proof. Because f p and gq are integrable, so are jf jp and jgjq . Let F; G W X ! R be

the functions defined by
jf .x/j
F.x/ D R
X
1=p
jf jp d
and
G.x/ D R
Thus,
Z
Z
F d D
Gq d D 1:
jg.x/j
q
X jgj d
1=q :
150
4 Integration
By Youngs Inequality (Proposition 4.50), for each x 2 X,

1
1
F.x/G.x/ F.x/p C G.x/q :
p
q
Hence,
Z
FG d
X
1
p
Z
F p d C
X
1
q
Z
Gq d D
X
1 1
C D 1:
p q
That is,
Z
jfgj d
R
X jf j
p d
X
1=p R
X jgj
1;

q d 1=q
t
u
Theorem 4.54 (Minkowskis Inequality). Suppose that p 1. If .X; ; / is a

measure space and if f ; g W X ! C are p-integrable, then f C g is p-integrable and
Z
jf C gjp d
1=p
Z
jf jp d
1=p
Z
C
jgjp d
1=p
(4.19)
Proof. The theorem is true for p D 1 because the sum of integrable functions is
integrable and because inequality (4.19) is simply a consequence of the triangle
inequality in real and complex numbers.
Assume, therefore, that p > 1 and consider the function # W RC ! RC defined
by #.t/ D tp . Because # is convex,

#
1
1
jf .x/j C jg.x/j
2
2

1
1
#.jf .x/j/ C #.jg.x/j/ ;
2
2
8x 2 X :
Hence,
p
1
1
1
.jf j C jgj/p jf jp C jgjp :
2
2
2
Because the sum of integrable functions is integrable, the inequality above shows
that .jf j C jgj/p is integrable. But, by the triangle inequality, jf C gjp .jf j C jgj/p ;
hence, f C g is p-integrable.
Let q 2 RC be conjugate to p. Thus, p D .p 1/q and
q

.jf j C jgj/p1 D .jf j C jgj/p :
151
Hence, .jf j C jgj/p1 is q-integrable. Therefore, one can apply Hlders Inequality
to obtain
Z
Z
jf j.jf j C jgj/
p1
d
jf j d
1=p Z
.p1/q
.jf j C jgj/
1=q
d
and
Z
Z
jgj.jf j C jgj/p1 d
jgjp d
1=p Z
.jf j C jgj/.p1/q d
1=q
Because jf jp C jgjp D jf j.jf j C jgj/p1 C jgj.jf j C jgj/p1 , summing the two inequalities above yields a new inequality whose left-hand side is
Z
.jf j C jgj/p d
X
and whose right-hand side is

Z
.p1/q
.jf j C jgj/
1=q "Z
jf j d
1=p
d
Z
C
Divide the new inequality through by

.p 1/q to obtain
Z
.jf j C jgj/p d
11=q
Because 1p D 1
inequality (4.19).
R
1=q
.p1/q
d
X .jf j C jgj/
Z
jf jp d
1=p
X
1
q
jgj d
1=p #
and use that p D
Z
C
jgjp d
1=p
and jf C gjp .jf j C jgj/p , the inequality above implies

t
u
Definition 4.55. Assume that p 2 R satisfies p 1. A sequence fk gk2N in C is

p-summable if
X
jk jp < 1 :
k2N
The Hlder and Minkowski inequalities also have formulations for sequences of
complex numbers.
Theorem 4.56. Suppose that fk gk2N and fk gk2N are sequences in C. Assume
that p 1.
1. (Hlder) If p and q are conjugate real numbers, and if fk gk2N is p-summable
and fk gk2N is q-summable, then fk k gk2N is summable and
152
4 Integration
jk k j
k2N
!1=p
jk j
k2N
!1=q
jk j
k2N
2. (Minkowski) If fk gk2N and fk gk2N are p-summable, then fk C k gk2N is psummable and
X
!1=p
jk C k jp
k2N
!1=p
jk jp
k2N
!1=p
jk jp
k2N
Proof. Apply Theorems 4.53 and 4.54 to the case where the measure space
.X; ; / is given by X D N, D P.N/, and is counting measure.
t
u
Note that the Hlder and Minkowski inequalities are nontrivial even in the cases
where the sequences fk gk2N and fk gk2N have only finitely many nonzero elements.
Appendix: The Issue of How to Integrate

Some concepts in mathematics develop rapidly, while other ideas take many decades
to mature and settle. The theory of integration is of the latter type.
In elementary calculus, integration is usually carried out in the style of Cauchy,
mostly for the purpose of making explicit calculations. However, the more theoretical aspects of analysis require a theory of integration in which certain properties of
integralsespecially properties of limitsare known to hold. For such purposes,
the integration theories of Cauchy and, as well, Riemann, do not do quite all that we
require of them. These difficulties, while recognised by many outstanding mathematicians of the late nineteenth-century, were not so easily overcome. Nevertheless,
after many new and innovative results had been obtained by several mathematicians,
a satisfactory theory of integration eventually emerged in the early twentieth-century
in a series of works by Henri Lebesgue.
Riemanns Approach
Assume that a; b 2 R satisfy a < b and that f W a; b ! R is a bounded function. That
is, there is a positive real number M for which jf .x/j M, for all x 2 a; b. If n 2 N,
then an n-partition of a; b is a collection of n finite intervals of the form .k1 ; k ,
where k D 1; : : : ; n and
a D 0 < 1 < : : : < n D b :
(4.20)
153
Let D .0 ; 1 ; : : : ; n / and denote the partition (4.20) of a; b by P . As f is

bounded, the real numbers mk and Mk , defined below, exist for all 1 k n:
mk D infff .x/ j x 2 .k1 ; k g
Mk D supff .x/ j x 2 .k1 ; k g :
and
Thus, every n-partition P of a; b induces a lower Riemann sum s , namely

s D
n
X
mk .k k1 / ;
kD1
and an upper Riemann sum S , which in this case is

S D
n
X
Mk .k k1 / :
kD1
The lower and upper Riemann integrals of f are defined, respectively, as follows:
Z

f .x/ dx D sup s j P is an n-partition of a; b; n 2 N

f .x/ dx D inf S j P is an n-partition ofa; b; n 2 N
Z
a
Definition 4.57. A bounded function f W a; b ! R is Riemann integrable if

Z
Z
f .x/ dx D
f .x/ dx :
(4.21)
In this case, the Riemann integral of f is the value of (4.21) and is denoted by
Z
f .x/ dx :
It is not difficult to see that any function f W a; b ! R that assumes only one
valuethat is, any constant function f is Riemann integrable. However, as soon
as we consider functions that assume two values, things go wrong.
Proposition 4.58. There is a bounded function f on 0; 1 that assumes only two
values yet fails to be Riemann integrable.
Proof. Consider the function f W 0; 1 ! R given by f D Q\0;1 . That is, f .q/ D 1
for every rational q 2 0; 1 and f .r/ D 0 for every irrational r 2 0; 1. As f assumes
only two values, f is bounded. Take any n-partition P of 0; 1. In each interval
154
4 Integration
.k1 ; k there are both rationals and irrationals, and so mk D 0 and Mk D 1 on each
.k1 ; k . Hence,
Z
Z
f .x/ dx D 0
and
f .x/ dx D 1 :
t
u
Therefore, f is not Riemann integrable.
What is the problem with the function f in Proposition 4.58 ? Well, for one thing,
no matter has small the interval .k1 ; k may be, the oscillation of f on .k1 ; k
(namely, between 0 and 1) is large relative to the length of .k1 ; k . Bounded
functions with uncontrollable oscillations generally fail to be Riemann integrable.
On the other hand, the oscillations of continuous functions diminish as the size of
an interval .k1 ; k decreases, and so continuous functions are Riemann integrable.
(This last explanation is not, of course, a proof.)
The next result points toward the more serious issue of limits.
Proposition 4.59. There exists a sequence of Riemann-integrable functions fk W
0; 1 ! R such that
1. 0 fk .x/ fkC1 .x/ 1, for all x 2 0; 1 and all k 2 N,
2. the Riemann integrals of fk converge to 0, that is,
Z 1
fk .x/ dx D 0 ; and
lim
k!1 0
3. the function f .x/ D limk fk .x/, x 2 0; 1 is not Riemann integrable.

Proof. Enumerate Q \ 0; 1 as fqn gn2N . For each k 2 Q, let fk D q1 ;:::;qk . Because fk
is continuous and equal to zero except at the k points q1 ; : : : ; qk , it is simple to verify
that fk is Riemann integrable and
Z 1
fk .x/ dx D 0 ; 8 k 2 N :
0
It is also clear that fk .x/ fkC1 .x/, for all x and all k, and that the limiting function f
is the characteristic function Q\0;1 , which is not Riemann integrable (by the proof
of Proposition 4.58).
t
u
The point of the Proposition 4.59 is that the hoped for formula
Z b
Z b
lim
fk .x/ dx D
lim fk .x/ dx
k!1 a
does not hold.
a k!1
(4.22)
155
Lebesgues Approach
To overcome the troubles suggested by Propositions 4.58 and 4.59, Lebesgue
thought to partition the y-axis rather than the x-axis. That is, suppose that
f W a; b ! R is a bounded function with range in the closed interval c; d. Let
P be an n-partition of c; d and consider the sets
Ek D fx 2 a; b j f .x/ 2 .k1 ; k g ;
1 k n:
Here, E1 ; : : : ; En are pairwise disjoint and their union is a; b. Assume, further,
that there is a length function m defined on subsets of a; b such that when m
is evaluated at any interval .x; y/, m gives the actual length of an interval .x; y/:
m. .x; y/ / D y x. With these assumptions, let
`k D infff .x/ j x 2 Ek g
Lk D supff .x/ j x 2 Ek g ;
and
and
t D
n
X
`k m.Ek /
T D
and
kD1
n
X
Lk m.Ek / :
kD1
Define lower and upper Lebesgue integrals, respectively, by

Z

f dm D sup t j P is an n-partition of c; d; n 2 N

f dm D inf T j P is an n-partition of c; d; n 2 N
Z
a
and say that f is Lebesgue integrable if

Z
a
Z
f dm D
f dm :
The value of partitioning c; d rather than a; b is that now oscillations of f over an

interval of a; b have no impact. However, one now requires the length function m,
and this function m has to be defined in such a way that m.E/ can be evaluated even
if E is a quite complicated subset. Thus, measure theory arose.
156
4 Integration
Lebesgues Criterion for Riemann Integration

Although the following theorem of Lebesgue will not be proved here, it is certainly
striking in that it shows that a bounded function f is Riemann integrable if and only
if f is continuous almost-everywhere.
Theorem 4.60. A bounded function f W a; b ! R is Riemann integrable if and only
if the set D a; b of points of discontinuity of f is a null set.
The Improper Riemann Integral

Another difference between Riemann and Lebesgue integration may be found in the
notion of an improper Riemann integral.
Definition 4.61. If f W .0; 1/ ! R is a Riemann-integrable function on a; b, for
every 0 < a R< b, then the improper Riemann integral of f over .0; 1/ is the quantity
1
denoted by 0 f .x/dx and defined by
Z 1
Z n
f .x/ dx D lim
f .x/ dx :
n!1 1=n
The theory of improper Riemann integrals is a continuous analogue of the theory

of conditionally convergent series. For example, in calculus it is shown that
Z 1
sin x
dx D
(as an improper Riemann integral) :
x
2
0
However, the measurable function f .x/ D sinx x on .0; 1/ is not (Lebesgue) integrable
on .0; 1/. To prove this, one may argue by contradiction. Assume that f is integrable
on .0; 1/. Thus, by Theorem 4.19, so is jf j:
jf j.x/ D
j sin xj
:
x
Let hk D .
;k
jf j, for each k 2 N. Then, fhk gk2N is a monotone-increasing sequence
of measurable nonnegative functions. In fact, each hk is integrable and jf j.x/ D
limk hk .x/ for every x 2 .
; 1/. Thus, by the Monotone Convergence Theorem,
Z
Z
hk dm D
jf j dm < 1 :
lim
k!1 .
;1/
.
;1/
We shall derive a contradiction by demonstrating that

Z
lim
k!1 .
;1/
hk dm D C1 :
Problems
157
To do this, decompose the interval .

; k
as union of pairwise disjoint subintervals
.j
; .j C 1/
, for j D 1; : : : ; k 1. Note that hk .x/ > j sin xj=..j C 1/
/ for all x 2
.j
; .j C 1/
. Thus,
Z
.
;1/
hk dm D
k1 Z
X
jD1
fk dm
k1 Z
X
jD1
.j
;.jC1/
j sin xj
dx
.j
C 1/
.j
;.jC1/
k1
2X 1
;
jD1 .j C 1/
which diverges as k ! 1.
As this discussion above points out, Lebesgues approach to integration is analogous to the theory of absolutely convergent series, whereas Riemanns approach is
more like the theory of conditionally convergent series.
Problems
4.62. Assume that .X; ; / is a measure space and that ';
nonnegative simple functions.
1. If
W X ! R are
.x/ '.x/, for all x 2 X, then prove that

Z
Z
d
' d :
2. If ; 2 R are nonnegative, then prove that ' C is a nonnegative simple

function and that
Z
Z
Z
.' C / d D ' d C
d :
X
4.63. If f ; g W X ! R are nonnegative and measurable, and if f .x/ g.x/, for all
x 2 X, then prove that
Z
Z
f d
X
g d :
X
158
4 Integration
4.64. Suppose that f W X ! R is measurable, nonnegative, and

Z
f d D 0 :
X
Prove that f D 0 almost everywhere.

4.65. Assume that .X; ; / is a measure space and that fn W X ! R is a nonnegative
measurable function for each n 2 N. Assume that
X
f .x/ D
fn .x/ ; 8 x 2 X :
nD1
If the series converges for every x 2 X, then prove that

Z
1 Z
X
f d D
fn d :
X
nD1 X
4.66. Assume that .X; ; / is a measure space and that f W X ! R is a nonnegative

measurable function. Define a function W ! 0; 1 by
Z
f d ;
.E/ D
8E 2 :
Prove that is a measure on .X; / and that .E/ D 0 for every E 2 for which
.E/ D 0.
4.67. Assume that .X; ; / is a measure space and that fk W X ! R is a nonnegative
measurable function for each k 2 N. Suppose that, for every x 2 X, limk fk .x/ exists
and let f D limk fk . If f .x/ fk .x/, for every x 2 X and every k 2 N, then use Fatous
Lemma to prove that
Z
Z
f d D lim
fk d :
k!1 X
4.68. Assume that .X; ; / is a measure space and that f W X ! R is an integrable

function. If .X/ < 1, then prove that for every " > 0 there is a simple function
' W X ! R such that
Z
jf 'j d < " :
X
Is the assumption that .X/ be finite a necessary assumption ? Explain.

4.69. Assume that .X; ; / is a measure space and that f ; fk W X ! R are
measurable functions such that

lim sup jfk .x/ f .x/j D 0 :
k!1
x2X
Problems
159
(That is, fk ! f uniformly on X.) If .X/ < 1 and each fk is integrable, then prove
that f is integrable and that
Z
Z
f d D lim
k!1 X
fk d :
4.70. Assume that .X; ; / is a measure space and that fk W X ! R is an integrable

function, for each k 2 N. Furthermore, assume that
lim fk .x/ exists almost everywhere :
k!1
If there is a nonnegative integrable function g W X ! R such that, for every k 2 N,

jfk .x/j g.x/
8x 2 X ;
then prove that

Z
Z
lim
k!1 X
fk d D
. lim fk / d :
X k!1
4.71. If .X; / is a measurable space, then prove that a function f W X ! C is

complex measurable if and only if <f and =f are measurable functions X ! R.
4.72. Prove that if is a complex measure on .X; /, and if f W X ! C is an
integrable function, then
Z
Z
f d jf j djj;
where jj is the total variation of (see Definition 3.70).

4.73. Prove that if .X; ; / is a measurable space and g W X ! R is a nonnegative
measurable function, then the function g W ! R defined by
Z
g .E/ D
g d; E 2 ;
E
is a measure on .X; /.
4.74. Compute the integral of Cantors ternary function over 0; 1.
Z
4.75. Compute f dm for each of the following sets E and functions f :
E
1. E D 0; 1/ and f .x/ D .1 x2 /1=2 ;

2. E D .0; 1 and f .x/ D x log x;
3. E D .0; 1/ and f .x/ D ex ;
160
4 Integration
4. E D .0; 1/ and f .x/ D .x/1 , where x denotes the largest integer n for which
n x;
5. E D R and f .x/ D .1 C x2 /1 .
4.76. Prove that
Z n
lim
n!1 0
1C
x
n 2x
e dx
n
exists
x
n x=2
e dx
n
exists
and evaluate the limit.

4.77. Prove that
Z n
lim
n!1 0
1
and evaluate the limit.

4.78. Let un be the characteristic function of .0; 1n , for each n 2 N. Show that
1
1
X
X
un .x/ converges for all x 2 R, but
un fails to be integrable.
nD1
nD1
4.79. Assume that .X; ; / is a measure space. Prove or disprove the following
1
X
un converges
assertion: If un W X ! R is integrable, for each n 2 N, and if
absolutely on X to a function u W X ! R, then u is integrable and
Z
1
X
X
!
un d D
1 Z
X
nD1
un d :
nD1 X
nD1
4.80. Consider the function f W 0; 1 ! R given by the series

f .x/ D
1
X
x
:
.1
C
x/n
nD0
1. Show that f .x/ D x C 1, if x 2 .0; 1, and that f .0/ D 0.

2. Does the series converge uniformly on 0; 1 ?
3. Is it true that
!
Z 1 X
1
1 Z 1
X
x
x
dx
dx D
n
.1 C x/
.1 C x/n
0
nD0
nD0 0
4.81. Prove that the Radon-Nikodm Theorem (Theorem 4.36) holds for -finite
measures on a measurable space .X; /.
Problems
161
4.82. Prove that if f#k gk2N is a sequence of convex functions on an open interval
J R such that supk #k .x/ exists for all x 2 J, then supk #k is a convex function.
4.83. Prove that every convex function is continuous.
4.84. Assume that J R is an open interval and that # W J ! R has a continuous
second derivative d2 #=dt2 at every point of J. Prove that is # is a convex function,
then d2 #=dt2 is nonnegative on J.
4.85. Assume that J is an open interval and that # W J ! R is a convex function.
Prove that if t1 ; : : : ; tn 2 0; 1 satisfy t1 C C tn D 1, and if x1 ; : : : ; xn 2 J, then
#
n
X
!
tk xk
kD1
n
X
tk #.xk / :
kD1
4.86. A function # W J ! R is strictly convex if

#. x C .1 /y/ < #.x/ C .1 /#.y/ ;
8 2 .0; 1/ and 8 x 6D y :
1. Prove that #.t/ D et is strictly convex on R for every nonzero 2 R.

2. Prove that #.t/ D tp on .0; 1/ for every p > 1.
3. Prove that, for positive real numbers 1 ; : : : ; n ,
.1 n /1=n D
1
.1 C C n /
n
if and only if 1 D D n .
4.87. Let 1 ; : : : ; n be positive real numbers and let p1 ; : : : pn be positive real
numbers that satisfy p11 C C p1n D 1.
1. Prove Youngs inequality:
1 : : : n
n
p
X
k
k
kD1
pk
2. Characterise the cases of equality in Youngs inequality.

4.88. Suppose that .X; ; / is a finite measure space and that f W X ! C is
integrable. Furthermore, suppose thatZU C is an open set for which E D f 1 .E/
has positive measure. Prove that
1
.X/
f d 2 U.
E
Part III
Banach Spaces
Chapter 5
Banach Spaces
Various collections of functions carry the structure of a vector space, and functional
analysis is devoted to the study of such vector spaces, mainly from an analytic rather
than linear algebraic perspective. When working with vector spaces, one is required
to specify the underlying base field. In analysis, the natural choices are the fields
R and C, which are preferable to the field Q or some finite field F, because of the
completeness properties enjoyed by the real and complex fields. However, because
the field C is algebraically closed, whereas R is not, there is a richer and more
widely used theory in the case of complex vector spaces. Thus, the base field for all
vector spaces under consideration is assumed, with very few exceptions, to be the
field C of complex numbers.
By equipping a vector space with additional structure, such as a topology, linear
transformations of the space and the vector space itself are poised to be studied from
the point of view of analysis.
5.1 Normed Vector Spaces

The most basic vehicle for introducing a topological structure to vector spaces is
through the use of a norm.
Definition 5.1. A norm on a (complex) vector space V is a function k k W V ! R
such that, for all v; w 2 V and 2 C,
1. kvk 0, and kvk D 0 only if v D 0,
2. k vk D jj kvk,
3. kv C wk kvk C kwk.

DOI 10.1007/978-3-319-45633-1_5
165
166
5 Banach Spaces
The inequality kv C wk kvk C kwk is called the triangle inequality. Observe

that a norm k k on a normed vector space V induces a metric topology, which is
also called a norm topology, on V via the metric d W V V ! R defined by
d.v1 ; v2 / D kv1 v2 k ;
v1 ; v2 2 V :
Proposition 5.2. If k k is a norm on V, then a basis B for the norm topology T

on V is given by
B D fBr .v/ j v 2 V; r 2 R; r > 0g ;
where Br .v/ D fw 2 V j kw vk < rg.
Proposition 5.2 above is a reformulation, for normed vector spaces rather than
arbitrary metric spaces, of Proposition 1.39.
In equipping a vector space with a normand, hence, a metric topologyone
can ask whether the vector space operations of scalar multiplication and vector
addition are continuous.
Proposition 5.3. If V is a normed vector space, then the maps a W V V ! V and
m W C V ! V defined by
a.v1 ; v2 / D v1 C v2
and
m.; v/ D v; 8 2 C; v; v1 ; v2 2 V;
are continuous.
v
uX
u n
Example 5.4. The equation kk D t
jj j2 , for
t
u
jD1
3
1 ;
6 7
D 4 ::: 5 2 Cn ;
2
n
defines a normcalled the Euclidean normon Cn .
Proof. The only nontrivial verification is that of the triangle inequality. If ; 2 Cn ,
then
1
0
n
X
j j A C kk2 :
k C k2 D kk2 C 2< @
jD1
167
The complex version of the classical Cauchy-Schwarz inequality (1.2) is
j j
jD1
v0
10
1
u
n
n
u X
X
u@
t
jj j2 A @
jj j2 A :
jD1
jD1
Because < jj

Pfor every
complex number , the Cauchy-Schwarz inequality
n
2
2
above yields 2<
jD1 j j 2kk kk. Hence, k C k .kk C kk/ , which
proves the triangle inequality.
t
u
Example 5.5. Assume that is a finite Borel measure on a compact Hausdorff
space X and let C.X/ be the vector space of all continuous maps f W X ! C. Then
each of the equations
Z
kf k1 D
jf j d
kf k1 D max jf .x/j
x2X
defines norms k k1 and k k1 on C.X/.

Proof. Because X is compact and jf j 2 C.X/ for each f 2 C.X/, the maximum
modulus of each f 2 C.X/ is achieved at some point x 2 X (Proposition 2.13).
Furthermore, bounded continuous functions are integrable on a finite measure space.
Hence, the definitions of k k1 and k k1 make sense for the vector space C.X/.
As before, it is only the triangle inequality that is in need of verification, as the
other requirements for a norm are clearly met by each of k k1 and k k1 . The
triangle inequality kf Cgk
k1ZCkgk1 is a consequence of the triangle inequality
Z 1 kf
for integralsnamely h d jhj dand the triangle inequality for complex

numbers, while the triangle inequality kf C gk1 kf k1 C kgk1 is a consequence
of the triangle inequality for complex numbers and for the max function.
t
u
There are many norms of interest on Cn . For example, by taking X D f1; : : : ; ng
and to be counting measure on X, Example 5.5 implies that
kk1 D max jj j
1jn
kk1 D
n
X
jj j
jD1
are norms on Cn . (This can be easily verified directly as well.) To distinguish the
Euclidean norm from the two norms above, we write kk2 for the Euclidean norm
k k of Example 5.4
168
5 Banach Spaces
Example 5.5 demonstrates how natural it is to consider different norms on a given

vector space. In moving from one norm to another, the shape of the closed unit ball
will change, and it may happen that V is complete in one norm but not in another
norm. With equivalent norms, this pathology does not arise.
Definition 5.6. If k k and k k0 are norms on a vector space V, then k k and k k0
are equivalent norms if there are positive constants c and C such that
ckvk kvk0 Ckvk;
for every vector v 2 V.
If one uses the notation k k
k k0 to mean that there are positive constants c
and C such that ckvk kvk0 Ckvk for every v 2 V, then it is easily proved that
is an equivalence relation on the set of all norms on a given vector space V.

Proposition 5.7. If T and T 0 are the norm topologies on a vector space V
induced, respectively, by equivalent norms k k and k k0 on V, then T D T 0 .
t
u

In finite-dimensional vector spaces, all norms are equivalent.
Proposition 5.8. All norms on a finite-dimensional vector space are equivalent.
Proof. Fix a linear basis fv1 ; : : : ; vn g of V and let k k2 be defined by

v

X
uX

n
u n

t
v
D
jj j2 :
j j

jD1
jD1
2
By Example 5.4, k k2 is a norm on V. Because equivalence of norms is an

equivalence relation, it is sufficient to show that if k k is a norm on V, then
k k2
k k.
v
uX
u n
To this end, let C D t
kvj k2 . Consider S D f 2 Cn j kk2 D 1g and the
jD1

X

n

function f W S ! R defined by f ./ D

v
j j . We claim that f is continuous.

jD1

Fix 2 S and consider a neighbourhood W R of f ./. Thus, there is a " > 0
such that .f ./ "; f ./ C "/ W. Let V Cn be the set of all 2 Cn for which
k k2 < "=C. Thus, U D S \ V is a neighbourhood of in S and if 2 U, then
169

X
X

n
n

jf ./ f ./j D

v

v
j j
j j

jD1
jD1

n

X

.

/v
j
j j

jD1

n
X
[by Exercise 5.104]
jj j j kvj k
jD1
v
v
uX
uX
u n
u n
2
t
jj j j t
kvj k2

jD1
[by Cauchy-Schwarz]
jD1
D k k2 C
< ";
which implies that f ./ 2 W. Hence, f is continuous at every 2 S, which implies f
is a continuous function.
Because jj2 D j<j2 C j=j2 for every 2 C, S may be identified with the
Euclidean sphere S2n1 in R2n . Thus, S is a compact set and, hence, the continuous
map f achievesP
its minimum value c at some element of S.
Now if v D j j vj 2 V for some 1 ; : : : ; n 2 C not all zero and if the coordinates
P
of 2 S are j D j =. k jk j2 /1=2 , then c f ./. Thus,
0
c@
n
X
jD1
11=2
jj j2 A

11=2
0
X
X
n
n
X
n

j vj
jj j kvj k C @
jj j2 A :

jD1

jD1
jD1
That is, ckvk2 kvk Ckvk2 for every v 2 V, which proves that k k2
k k.
t
u
A slightly more general concept than that of a norm is the notion of a seminorm.
Definition 5.9. A seminorm on a vector space V is a function W V ! R such that,
for all v; w 2 V and 2 E,
1. .v/ 0,
2. .v/ D jj.v/,
3. .v C w/ .v/ C .w/.
Seminorms differ from norms in the following way: with a seminorm , the
equation .v/ D 0 can hold for nonzero v; however, with a norm k k, the equation
kvk D 0 holds only for v D 0. Nevertheless, one can obtain a norm from a seminorm
as follows.
170
5 Banach Spaces
Proposition 5.10. If is a seminorm on a vector space, and if

is the relation on
V defined by
v
w
.v w/ D 0 ;
if
then
is an equivalence relation. Moreover, if the equivalence classes of elements
of V are denoted by
vP D fw 2 V j w
vg ;
then:
1. the set V=
of equivalence classes is a vector space under the operations
P w/ ; v; w 2 V ;
vP C wP D .v C
P
vP D .v/ ; 2 C ; v 2 V I
2. the function k k W V=
! R defined by
kvk
P D .v/ ;
v 2V;
is a norm on V=
.
t
u
Through the use of a norm, one can introduce the notions of convergent
sequences and series, as well as the idea of completeness.
Definition 5.11. A normed vector space .V; k k/ is a Banach space if .V; d/ is a
complete metric space, where d is the metric d.v; w/ D kv wk.
Normed vector spaces of finite dimension provide the simplest examples of
Banach spaces.
Proposition 5.12. Every finite-dimensional normed vector space is a Banach
space.
Proof. By Proposition 5.8 and its proof, if fv1 ; : : : ; vn g is a basis for a normed vector
space V, then there are positive constants c and C such that

0
0
11=2
11=2
n

n
n
X
X
X

@
c@
jj j2 A
j vj
jj j2 A ;

C

jD1
jD1
jD1
(5.1)
for all 1 ; : : : ; n 2 C. Assume now that fwk gk is a Cauchy sequence of elements

in V. Write
wk D
n
X
jD1
.k/
j vj ;
8k 2 N:
171
.k/
Then inequality (5.1) implies that for each j D 1; : : : ; n, the sequence fj gk conP
verges in C to some j (because C is complete). Let w D njD1 j vj . Inequality (5.1)
implies, again, that the sequence fwk gk converges in V to w. Hence, V is a Banach
space.
t
u
Definition 5.13. If X and Y are nonempty subsets of a vector spaces V, then X C Y
denotes the set of all vectors of the form x C y, where x 2 X and y 2 Y.
The completeness axiom for Banach spaces has numerous consequences, such as
the following topological property exhibited by compact sets.
Proposition 5.14. Suppose that K and C are subsets of a Banach space V such that
K is compact and C is closed. If K \ C D ;, then there exists " > 0 such that
.K C B" .0// \ C D ;:
Proof. If the conclusion is not true, then there are vn 2 K and wn 2 V of norm
kwn k < 1n such that vn Cwn 2 C, for every n 2 N. Since K is compact, fvn gn2N admits
a convergent subsequence fvnk gk2N with limit v 2 K. Note that kv .vnk C wnk /k
kv vnk k C n1k ; thus, v 2 C, as C is closed. But this contradicts K \ C D ;; therefore,
the conclusion must hold.
t
u
The notion of basis for a vector space, which is a common feature of linear algebra, has a limited role in Banach space theory because the defining conditions are
purely algebraic and do not take into account the topological or analytic properties
of the space. Nevertheless, it is interesting to examine the facts concerning linear
bases, as in Proposition 5.16 below, for instance.
Recall that if S is a nonempty subset of a vector space V, then:
1. the elements of S are linearly independent if, for every finite subset fv1 ; : : : ; vn g
n
X
j vj D 0 holds, for 1 ; : : : ; n 2 C, only if each j D 0; and
S, the equation
jD1
2. the span of S is the set of all linear combinations

v1 ; : : : ; vn 2 S.
n
X
j vj , for all finitely many
jD1
The important point to keep in mind is that linear combinations are finite sums, even
though in analysis one considers infinite sums.
Definition 5.15. A linear basis of a vector space V is a subset B V such that the
elements of B are linearly independent and span V.
Recall from linear algebra that if a vector space V admits a finite linear basis
B, then every linear basis of V has cardinality equal to that of B and this integer
is called the dimension of V. In particular, V is finite dimensional if V has a finite
linear basis, and V is infinite dimensional if V has no finite linear basis. In the case
of infinite-dimensional Banach spaces, the question of existence of linear bases
172
5 Banach Spaces
is addressed by the following proposition. The proposition also shows that the
completeness of Banach spaces (as metric spaces) forces linear bases of Banach
spaces to be of sufficiently large cardinality.
Proposition 5.16. If V is an infinite-dimensional Banach space, then V has a linear
basis and every linear basis of V is an uncountable set.
Proof. By hypothesis, there is an infinite set B0 of linearly independent vectors in
V. Let S be the set of all subsets Y V of linearly independent vectors for which
Y B0 , and define a partial order on S by set inclusion: that is, X Y if and
only if X; Y 2 S satisfy X Y.
[
Let E S be a linearly ordered subset and consider the set Y D
E. To
E2E
show that the elements of Y are linearly independent, select y1 ; : : : ; yn 2 Y. Thus,

there exist sets E1 ; : : : ; En 2 E such that yj 2 Ej for all j. Because E is linearly
ordered, either E1 E2 or E2 E1 . Thus, there is an i2 2 f1; 2g such that y1 ; y2 2
Ei2 . Likewise, Ei2 E3 or E3 Ei2 ; hence, there is an i3 2 f1; 2; 3g such that
y1 ; y2 ; y3 2 Ei3 . Continuing by induction we obtain an integer in 2 f1; : : : ; ng for
which y1 ; : : : ; yn 2 Ein . As Ein consists of linearly independent vectors, we deduce
that y1 ; y2 ; : : : ; yn 2 Y are linearly independent. Furthermore, Y E B0 for every
E 2 E, and so Y is an element of S and is an upperbound in S of the linearly ordered
set E. Hence, by Zorns Lemma (Theorem 1.8), S has a maximal element B.
To show that B is a linear basis of V, consider the linear submanifold M D
Span B. If M 6D V, then there is a unit vector v 2 V with v 62 M, which implies that
v is linearly independent of every vector in M and, in particular, of every vector in
Q D B [ fvg is a linearly independent set that properly contains B. But
B. Thus, B
Q and B 6D B
Q contradict the maximality of B in S. Hence, it must be that
BB
M D V, which implies that B is a linear basis of V.
Now suppose that B is an arbitrary linear basis of V and assume, contrary to
what we aim to prove, that B is a countable set B D fvn j n 2 Ng. Thus, if Vn D
Span fv1 ; : : : ; vn g, then
VD
Vn :
n2N
Each Vn has finite dimension and is, therefore, closed in the metric topology of V.
We claim that Vn is nowhere densethat is, that the interior Un of Vn is empty.
To this end, select v 2 Un and " > 0 such that B" .v/ Un . On the line segment
f.1t/v CtvnC1 j t 2 0; 1g connecting v and vnC1 , select t 2 .0; 1/ sufficiently small
such that k.1 t/v C tvnC1 vk D tkvnC1 vk < ". Hence, there is a wn 2 B" .v/
such that vnC1 2 Spanfv; wn g Vn , which is a contradiction. Therefore, it must be
that Un D ; for every n 2 N, which implies that each Vn is nowhere dense. By
Corollary 2.57 of the Baire Category Theorem (Theorem 2.55), a countable union
of nowhere dense sets in a complete metric space is nowhere dense. Hence, V is
nowhere dense, which is impossible because V is both open and closed in V. Thus,
no linear basis of V can be countable.
t
u
5.2 Subspaces, Quotients, and Bases
173
5.2 Subspaces, Quotients, and Bases

Because Banach spaces are vector spaces, their subspaces are of interest. However,
the presence of an underlying topology means that one needs to distinguish
subspaces that are closed in the topology from those that are not. Thus, the following
terminology is adopted.
Definition 5.17. Suppose that V is a Banach space and that L V.
1. L is a linear submanifold, or linear manifold, of V if
1 w1 C 2 w2 2 L ;
8 1 ; 2 2 C; w1 ; w2 2 L :
2. L is a subspace of V if L is a linear submanifold of V and L is a closed set in the

topology of V.
If L is a subspace of a Banach space V, then the elements of the quotient vector
space V=L are denoted by v,
P for v 2 V, and are given by
vP D fw 2 V j v w 2 Lg:
Proposition 5.18. If L is a subspace of a Banach space V, then the function k k on
the quotient space V=L, defined by
kvk
P D inffkv yk j y 2 Lg ;
(5.2)
is a norm on V=L such that, with respect to this norm, V=L is a Banach space.
Proof. Observe that, by definition, kvk
P kvk for every v 2 V. To show kvk
P D0
P suppose that kvk
only if vP D 0,
P D 0. Thus, there is a sequence of vectors yn 2 L such
that kv yn k ! 0. Because L is closed and v is the limit of the sequence fyn gn2N , v
P It is readily apparent that k vk
must belong to L, which yields vP D 0.
P D jj kvk
P and
kvP C wk
P kvk
P C kwk,
P and so the function k k defined by equation (5.2) is indeed a
norm on the vector space V=L.
Now suppose that fvP k gk2N be a Cauchy sequence in V=L, and select a subsequence fvP kj gj2N with the property that kvP kj vP kj1 k < 2j for every j 2 N. Therefore,
for each j there exists yj 2 .vP kj vP kj1 / with kyj k kvP kj vP kj1 k < 2j . Hence,
1
1
X
X
kyj k is convergent in R, which implies that
yj is convergent in V to some
jD2
vector y. Consider yP C vP k1 2 V=L. For any j 2 N,
jD2
j

X

vP k .Py C vP k / D .vP k vP k / yP D .vP k vP k / yP

j
1
j
1
i
i1

iD2
j
j

X
X

D
yP i yP
yi y :

iD2
iD2
174
5 Banach Spaces
P
Because 1
P kj gj converges to yP C vP k1 in V=L.
iD2 yi converges to y 2 V, the sequence fv
However, as fvP kj gj is a convergent subsequence of the Cauchy sequence fvP k gk2N , the
sequence fvP k gk must also converge in V=L to yP C vP k1 .
t
u
The next two lemmas give important analytical information about infinitedimensional Banach spaces, and we shall make use of them frequently in this study.
Lemma 5.19 (Increasing Subspace Chain Lemma). If V is an infinitedimensional Banach space and if fMn gn2N is a sequence of subspaces such that
Mn MnC1 (proper containment), then for each 2 .0; 1/ there is a sequence of
vectors vn 2 V such that
1.
2.
3.
4.
vn 2 MnC1 and vn 62 Mn ,
kvn k D 1,
kvn uk for all vectors u 2 Mn , and
kvk vj k for all j; k 2 N such that k 6D j.
Proof. Fix n 2 N. Because the containment Mn MnC1 is proper, the quotient space
MnC1 =Mn is nonzero; hence, there is a vector of norm . Since < 1, there are
vectors fn 2 Mn and gn 2 MnC1 such that kPgn k D and kgn fn k D 1. Let vn D gn fn .
Because vP n D gP n and D kvP n k D inffkvn f k j f 2 Mn g, we deduce that kvn uk
for all vectors u 2 Mn .
Having produced such a unit vector vn 2 MnC1 for each n 2 N, now select j; k 2 N
with k 6D j. Without loss of generality, assume that j < k. Thus, vj 2 MjC1 Mk and,
therefore, kvk vj k .
t
u
The second lemma is proved by precisely the same type of argument.
Lemma 5.20 (Decreasing Subspace Chain Lemma). If V is an infinitedimensional Banach space and if fMn gn2N is a sequence of subspaces such that
Mn MnC1 (proper containment), then for each 2 .0; 1/ there is a sequence of
vectors vn 2 V such that
1.
2.
3.
4.
vn 2 Mn and vn 62 MnC1 ,
kvn k D 1,
kvn uk for all vectors u 2 MnC1 , and
kvk vj k for all j; k 2 N such that k 6D j.
A notable consequence of Lemma 5.19 is:
Proposition 5.21. If V is an infinite-dimensional Banach space, then there exists

a sequence fvn gn2N of (distinct) unit vectors vn 2 V such that fvn gn2N has no
convergent subsequences.
Proof. Select 2 .0; 1/. By Lemma 5.19 there is a sequence fvn gn2N of (distinct)
unit vectors vn 2 V such that kvk vj k for every j 6D k. Thus, fvn gn2N has no
Cauchy subsequences and, therefore, fvn gn2N has no convergent subsequences. u
t
5.3 Banach Spaces of Continuous Functions
175
Proposition 5.21 is noted here for future reference. The first use of the result is
in the following characterisation of the compactness of the closed unit ball in terms
of dimension.
Proposition 5.22. The closed unit ball in a Banach space V is compact if and only
if V has finite dimension.
Proof. Suppose that V has dimension n 2 N. By Proposition 5.8, the closed unit balls
of V and `1 .n/ are homeomorphic. Because the unit ball of `1 .n/ is a product of
n copies of the closed unit disc D D fz 2 C j jzj 1g, and because D is compact, the
closed unit balls of `1 .n/ and V are compact.
If V has infinite dimension, then Proposition 5.21 shows that there exists
a sequence fvn gn2N of (distinct) unit vectors vn 2 V such that fvn gn2N has no
convergent subsequences. Because sequences in compact metric spaces admit
convergence subsequences (Proposition 2.19), the unit sphere is not a compact set,
nor is any closed set that contains the closed unit sphere. Hence, the closed unit ball
of V is compact only if V has finite dimension.
t
u

Recall from Definition 2.20 that a topological space X is locally compact if, for every
x 2 X, there is an open set U X containing x such that U is compact. Assuming
X is locally compact, let Cb .X/ denote the set of all functions f W X ! C that are
continuous and bounded. Thus, for every f 2 Cb .X/ means that there is an R > 0
such that jf .x/j < R for all x 2 X.
Theorem 5.23. If X is a locally compact space, then Cb .X/ is a Banach space,
where the vector space operations are given by the usual pointwise operations, and
where the norm of f 2 Cb .X/ is defined by
kf k D sup jf .x/j:
(5.3)
x2X
Proof. It is elementary that Cb .X/ is a vector space and that (5.3) defines a norm
on Cb .X/. Thus, it remains only to show that every Cauchy sequence in Cb .X/ is
convergent in Cb .X/.
Let ffk gk2N Cb .X/ denote a Cauchy sequence. For each x 2 X,
jfn .x/ fm .x/j sup jfn .y/ fm .y/j D kfn fm k :
y2X
Since ffk gk2N is a Cauchy sequence in Cb .X/, ffk .x/gk2N is a Cauchy sequence in C
for each x 2 X. Because C is complete, limk fk .x/ exists for every x 2 X. Therefore,
define f W X ! C by f .x/ D limk fk .x/, for each x 2 X. We aim to show (i) that f is
continuous and bounded, and (ii) that ffk gk2N converges to f in Cb .X/.
176
5 Banach Spaces
Let " > 0. Because ffk gk2N is a Cauchy sequence, there exists N" 2 N such that
kfn fm k < " for all n; m N" . Assume that n N" . Choose any x 2 X; thus,
jf .x/ fn .x/j jf .x/ fm .x/j C jfm .x/ fn .x/j
jf .x/ fm .x/j C kfm fn k :
As the inequalities above are true for all m 2 N,
jf .x/ fn .x/j inf .jf .x/ fm .x/j C kfm fn k/
m2N
0C":
This right-hand side of the inequality above is independent of the choice of x 2 X.
Hence, if n N" is fixed, then f fn is a bounded function X ! C and
sup jf .x/ fn .x/j " :
x2X
Since f is uniformly within " of a continuous function, f is continuous at each x 2 X.

Furthermore, since the sum of bounded functions is bounded, fn C .f fn / D f is
bounded. This proves that f 2 Cb .X/. Finally, since f 2 Cb .X/ satisfies kf fn k "
for all n N" , the Cauchy sequence ffk gk2N converges in Cb .X/ to f 2 Cb .X/.
t
u
Notational Convention For a compact space X, we denote Cb .X/ by C.X/.
If X is compact and f 2 C.X/, then f .X/ is a compact subset of C; hence, f is
bounded and attains its supremum at some point of X. Therefore, the norm kf k is
given by
kf k D max jf .x/j:
x2X
Because C has multiplication as well as addition, we may multiply f ; g 2 Cb .X/

to produce a function fg W X ! C whose value .fg/x at each x 2 X is defined by
.fg/x D f .x/g.x/ :
It is not difficult to see that fg 2 Cb .X/ and that kfgk kf k kgk.
Definition 5.24. An associative algebraor, more simply, an algebrais a
complex vector space A endowed by with a product (or multiplication) operation
such that, for all a; b; c 2 A and all 2 C,
.a C b/c D .ac C bc/;
a.b C c/ D ab C ac;
and
. a/.b/ D a. b/ D .ab/:
a.bc/ D .ab/c;
177
Furthermore, if ab D ba, for all a; b 2 A, then A is called an abelian algebra, and if

there is an element 1 2 A such that a1 D 1a D a, for every a 2 A, then A is said to
be a unital algebra and 1 is the multiplicative identity of A.
It is not difficult to show that, if 1 and 10 are multiplicative identities for an
algebra A, then 10 D 1.
Definition 5.25. A Banach algebra is a complex associative algebra A together
with a norm k k on A such that
1. kxyk kxkkyk, for all x; y 2 A, and
2. A is a Banach space under the norm k k.
Furthermore, if A is a unital algebra, then A is a unital Banach algebra if k1k D 1.
Thus, if X is a compact space, then C.X/ is a Banach algebra. The (constant)
function that sends each x 2 X to 1 2 C is denoted by 1 and it serves as the
multiplicative identity for C.X/ in the sense that f 1 D f for every f 2 C.X/.
Definition 5.26. A uniform algebra on a compact space X is a subset A C.X/
such that:
1. A is a Banach subalgebra of C.X/;
2. 1 2 A;
3. A separates the points of Xthat is, if x1 ; x2 2 X are distinct, then there exists a
function f 2 A such that f .x1 / 6D f .x2 /.
Discussion of uniform algebras makes sense only if X is Hausdorff:
Proposition 5.27. C.X/ is a uniform algebra on a compact space X if and only if
X is Hausdorff.
Proof. Suppose that A is any uniform algebra on X and that x1 ; x2 2 X are distinct. By
hypothesis, there is a function f 2 A such that f .x1 / 6D f .x2 /. In C there are disjoint
open sets V1 and V2 that contain f .x1 / and f .x2 /, respectively. Thus, by continuity
of f , U1 D f 1 .V1 / and U2 D f 1 .V2 / disjoint open sets in X that contain x1 and x2 ,
respectively, which proves that X is a Hausdorff space.
Conversely, suppose that X is Hausdorff. Because X is a normal space (Proposition 2.34), if x0 ; x1 2 X are distinct, then Urysohns Lemma applied to the point sets
fx0 g and fx1 g (which are closed because X is Hausdorff) yields a function f 2 C.X/
such that f .x0 / D 0 6D 1 D f .x1 /. Hence, C.X/ is a uniform algebra.
t
u
The elements of C.X/ are complex-valued functions. Therefore, for each f 2
C.X/ one can consider the continuous function f W X ! C defined by
f .x/ D f .x/ ;
8x 2 X :
Definition 5.28. A nonempty subset S C.X/ is selfadjoint if f 2 S for every f 2 S.
178
5 Banach Spaces
The Stone-Weierstrass Theorem, Theorem 5.30 below, asserts that if A is a

selfadjoint uniform algebra of continous functions f W X ! C, where X is a compact
Hausdorff space, then A D C.X/.
Lemma 5.29. Suppose X is compact and that A C.X/ is a unital, closed
subalgebra of C.X/. If A is selfadjoint, and if f ; f1 ; : : : ; fn 2 A are real-valued
functions, then A also contains the following real-valued functions:
(i)
(ii)
(iii)
(iii)
jf j;
min.f1 ; : : : ; fn /;
max.f1 ; : : : ; fn /; and
f C and f , where f C D 12 .jf j C f / and f D 12 .jf j f /.
Proof. If f D 0, then jf j D 0 and so jf j 2 A trivially. Assume, therefore, that f 6D 0; by

normalising, we may assume without loss of generality that kf k D 1, which implies
that f .x/ 2 1; 1 for all x 2 X. By Newtons Binomial Theorem,
1
p
1 3 .2n 3/ n
t X
t ;
1t D 1 C
.1/n
2 nD2
2n n
which converges absolutely on 1; 1 and uniformly on compact subintervals of

.1; 1/.pFor notational convenience, let ' denote the function on 1; 1 given by
'.t/ D 1 t and write the power series expansion above of ' as
'.t/ D
1
X
n tn :
nD0
For each 2 .0; 1/ let g 2 C.X/ be given by g .x/ D C .1 /f .x/2 ; that is, g D
C .1 /f 2 , where 2 A is the constant function x 7! . Because A is an algebra,
f 2 2 A and g 2 A. Furthermore, f .x/2 2 0; 1 for all x 2 X, and so 0 g 1 and
0 1 g D .1 /.1 f 2 / 1 . That is,
1 g .x/ 2 0; 1 ;
8x 2 X :
Fix k 2 N and define f;k by

f;k D
k
X
n .1 g /n ;
nD0
where 0 ; : : : ; k 2 R are the coefficients in the power series expansion of '. Thus,
f;k 2 A and
k
n
1=2
n .1 g .x// '.1 g .x//
kf;k .g / k D max
x2X
nD0
n
max
n t '.t/ :
t20;1
nD0
179
By Newtons Binomial Theorem, this final limit tends to zero as k ! 1. Hence,

.g /1=2 2 A (as A is norm closed). Note that kf 2 g k D k1 C f 2 kp! 0 as ! 0;
that is, g ! f 2 uniformly on X as ! 0. The function .t/ D t is uniformly
continuous on the compact set 0; 1, and so g ! f 2 uniformly on X as
! 0. Because g D .g /1=2 2 A and f 2 D jf j, the limit k .g /1=2 jf j k ! 0
implies that jf j 2 A. This completes the proof that jf j 2 A for every real-valued f 2 A.
As a consequence of the arguments above, if f1 ; f2 2 A are real-valued, then the
continuous functions 12 .f1 C f2 C jf1 f2 j/ and 12 .f1 C f2 jf1 f2 j/ are elements of A.
That is, max.f1 ; f2 / 2 A and min.f1 ; f2 / 2 A. By induction,
max.f1 ; : : : ; fm / D max .max.f1 ; : : : ; fn1 /; fn /
and
min.f1 ; : : : ; fm / D min .min.f1 ; : : : ; fn1 /; fn /
are elements of A. Likewise, f C ; f 2 A.
t
u
Theorem 5.30 (Stone-Weierstrass Theorem). If X is a compact Hausdorff space

and if A is a selfadjoint uniform algebra on X, then A D C.X/.
Proof. First of all, because A is selfadjoint, A is spanned by real-valued functions.
Indeed, if f 2 A, then <f D 12 .f Cf / and =f D 2i1 .f f / are real-valued elements of A
and f D <f C i=f . Therefore it is sufficient to prove that f 2 A for every real-valued
function f 2 C.X/.
To this end, assume that f 2 C.X/ and let " > 0 be arbitrary. Fix x0 2 X and select
any x1 2 X for which x1 6D x0 . Because A separates points, there is a function h 2 A
such that h.x1 / 6D h.x0 /. At least one of <h.x1 / 6D <h.x0 / or =h.x1 / 6D =h.x0 / holds,
and so we may assume, because A is self-adjoint, that h is a real-valued function.
Consider now the real-valued function gx1 2 C.X/ defined by

gx1 .y/ D f .x0 / C .f .x1 / f .x0 //

h.y/ h.x0 /
;
h.x1 / h.x0 /
8y 2 X :
In particular, for y D x0 and y D x1 we obtain gx1 .x0 / D f .x0 / and gx1 .x1 / D f .x1 /.
Now this construction holds as long as x1 6D x0 ; we define gx0 to be f . What has been
proved, then, is the following assertion: given a fixed x0 2 X, there exists, for every
x 2 X, a real-valued gx 2 A such that gx .x0 / D f .x0 / and gx .x/ D f .x/.
Continuing with the assumption that x0 2 X is fixed, note that gx f 2 C.X/ for
every x 2 X. Hence, if Wx C is the open set Wx D fz 2 C j <z < "g, then
Ux D .gx f /1 .Wx / D fy 2 X j gx .y/ f .y/ < "g
is open in X. Furthermore, gx .x/ D f .x/ implies that x 2 Ux . Hence, fUx gx2X is
an open cover X. The compactness of X implies that this covering admits a finite
subcover, say, Ux1 , . . . , Uxn . The functions gxj that define these n open sets determine
180
5 Banach Spaces
another element of A: namely, minfgx1 ; : : : ; gxn g, by Lemma 5.29. Because all of this
has depended on the fixed element x0 2 X, this minimum function shall be denoted
by hx0 . That is, hx0 D minfgx1 ; : : : ; gxn g and hx0 has the property
hx0 .y/ < f .y/ C " ;
8y 2 X :
(5.4)
Now allow x0 to vary throughout X, producing for each point x0 the continuous
function hx0 2 A described above. For each x0 2 X, consider the open subset Vx0 of
X defined by
Vx0 D fy 2 X j hx0 .y/ f .y/ > "g :
Since hx0 .x0 / f .x0 / D 0, x0 2 Vx0 . Moreover,
hx0 .y/ > f .y/ " ;
8 y 2 Vx0 :
Therefore, fVx0 gx0 2X is an open cover of X and, by the compactness of X, there is a

finite subcover: Vx1 , . . . , Vxm . Let D maxfhx1 ; : : : ; hxm g, which is an element of A
by Lemma 5.29. Thus, for each j D 1; : : : ; m,
hj .y/ .y/ > f .y/ " ;
8y 2 X :
(5.5)
Combining inequalities (5.4) and (5.5) leads to

f .y/ " < .y/ < f .y/ C " ;
8y 2 X :
That is, kf k < ".
t
u
Corollary 5.31 (Classical Weierstrass Approximation Theorem). If f W a; b !

C is a continuous function, then for every " > 0 there is a polynomial p with complex
coefficients such that jf .t/ p.t/j < ", for all t 2 a; b.
Proof. Let A be the closure in C.a; b/ of the ring Ct of polynomials in one
indeterminate t. Thus, A is a norm-closed subalgebra of C.a; b/ and 1 2 A.
Moreover A separates the points of a; b, for if x1 ; x2 2 a; b are distinct, then
q.x1 / D 0 and q.x2 / 6D 0 for the element q 2 A given by q.t/ D t x1 . Therefore, the
Stone-Weierstrass Theorem yields A D C.a; b/. In particular, by the construction
of A, if f 2 C.a; b and if " > 0, then there is a polynomial p such that kf pk < ".
t
u
In the case of non-compact, locally compact spaces, there are functions that, in
certain respects, mimic continuous functions on compact spaces.
Recall from Definition 2.40 that the support of a continuous function f W X ! C
on a topological space X is the set supp f X defined by
supp f D fx 2 X j f .x/ 6D 0g:
5.4 Banach Spaces of p-Integrable Functions
181
Definition 5.32. If X is a locally compact Hausdorff space, then let

Cc .X/ D ff W X ! C j f is continuous and supp f is compactg;
the set of all continuous complex-valued functions on X having compact support.
Proposition 5.33. If X is a locally compact space, then Cc .X/ is a subspace
of Cb .X/.
t
u
Another useful Banach space of continuous functions is the space consisting of

continuing functions that vanish at infinity.
Definition 5.34. Suppose that X is locally compact space. A continuous function
f W X ! C vanishes at infinity if the set
fx 2 X j jf .x/j "g
is compact in X for every " > 0.
The final result of this section makes note of the fact that C0 .X/ has some
additional algebraic structure, and that it is a closed set in the topology of Cb .X/.
Proposition 5.35. If X is a locally compact space, then
1. C0 .X/ is a subspace of Cb .X/, and
2. fg 2 C0 .X/, for all f 2 C0 .X/ and g 2 Cb .X/.
t
u

Proposition 5.36. Suppose that .X; ; / is a measure space, and that p 1. If
L p .X; ; / D ff W X ! C j f is p-integrable g ;
then L p .X; ; / is a complex vector space. Furthermore, if W L p .X; ; / ! R
is given by
Z
.f / D
jf j d
p
1=p
(5.6)
for all f 2 L p .X; ; /, then is a seminorm on L p .X; ; /.

Proof. It is clear that f 2 L p .X; ; /, for every 2 C and f 2 L p .X; ; /. If
f ; g 2 L p .X; ; /, then f C g 2 L p .X; ; /, by Minkowskis inequality. Hence,
L p .X; ; / is a vector space.
182
5 Banach Spaces
To verify that is a seminorm, the only nontrivial fact to confirm is the triangle
inequality holds. To this end, Minkowskis inequality yields:
Z
1=p
p
.f C g/ D
jf C gj d
X
Z
jf j d
1=p
Z
C
jgj d
1=p
D .f / C .g/ :
Hence, is a seminorm.
t
u
The seminorm of Proposition 5.36 need not be a norm. For example, if f is the
characteristic function on the set Q of rational numbers, and if m denotes Lebesgue
measure on R, then f 6D 0, yet
Z
1=p
p
.f / D
jf j dm
D m.Q/1=p D 0 :
R
On the other hand, Proposition 5.10 demonstrates that a bona fide normed vector
space can be obtained by passing to equivalence classes.
Definition 5.37. If p 1, and if .X; ; / is a measure space, then Lp .X; ; /
denotes the normed vector space
Lp .X; ; / D L p .X; ; /=
;
where
is the equivalence relation f
g if .f g/ D 0 and where is the seminorm
in (5.6). The vector space Lp .X; ; / is called an Lp -space.
Conceptual and Notational Convention The vector space Lp .X; ; / is a vector
space of equivalence classes of p-integrable functions f W X ! C. Thus, one properly
denotes the elements of Lp .X; ; / by fP , where f 2 L p .X; ; /. However, it
is a standard practice to denote the elements of Lp .X; ; / as simply f rather
than fP . Nevertheless, we have adopted the notation L p .X; ; / for the purpose
of designating functions, and so, in the interests of clarity, we retain the use of the
notation fP for the equivalence class of f 2 L p .X; ; / in the normed vector space
Lp .X; ; /.
The following result gives rise to another class of Banach spaces.
Theorem 5.38 (Riesz). Lp .X; ; / is a Banach space, for every p 1.
Proof. Suppose that ffPk gk2N is a Cauchy sequence in Lp .X; ; /, for some sequence
of p-integrable functions fk 2 L p .X; ; /. Because this sequence is Cauchy, one
can extract from it a subsequence ffPkj gj2N such that
kfPkjC1 fPkj k <
j
1
;
2
8j 2 N:
183
For every i 2 N, let gi 2 L p .X; ; / be given by

gi D
i
X
jfkjC1 fkj j :
jD1
Observe that fgi gi2N is a monotone-increasing sequence and that

kPgi k
i
X
i
X
k jfPkjC1 fPkj j k
jD1
2j
jD1
1
X
2j D 1 :
jD1
Thus,
Z
p
gi d 1 ;
8i 2 N:
The converse to the Monotone Convergence Theorem (Theorem 4.23) implies that
limi gi .x/p exists for almost all x 2 X; thus, limi gi .x/ exists for almost all x 2 X and
the limit functioncall it gis p-integrable. Let L X denote the set of points x in
which limi gi .x/ exists; thus,
lim gi .x/ D lim
i!1
i!1
i
X
jfkjC1 .x/ fkj .x/j :
(5.7)
jD1
If f W L ! C denotes the function defined by

f .x/ D fk1 .x/ C
1
X

fkjC1 .x/ fkj .x/ ;
jD1
then series above converges absolutely, by (5.7), for every x 2 L. Extend f to all of
X by setting f .x/ D 0 if x 2 XnL. The .i 1/-th partial sum of f is precisely fki , and
i1
X
jfki j D fk1 C
.fkjC1 fkj / jfk1 j C gi1 2g :
jD1
Therefore, jfki jp 2p gp for all i 2 N. As gp is integrable and limi fki .x/p D f .x/p for
all x 2 L, the Dominated Convergence Theorem asserts that f p is integrable. This
proves that f 2 L p .X; ; /.
What remains is to show limk kfP fPk k D 0. To this end, note that
jf fki jp .jf j C jfki j/p 2p gp
8i 2 N :
184
5 Banach Spaces
Therefore, by the Dominated Convergence Theorem, jf fki jp is integrable for every

i and the limit functionin this case 0satisfies
Z
lim j0 jf fki jp j d D 0 I
i!1 X
that is,
Z
lim
i!1 X
jf fki jp d D 0 :
Hence, kfP fPki k ! 0 as i ! 1. To show that the entire Cauchy sequence ffPk gk2N
converges in Lp .X; ; / to fP , let " > 0. Since ffPk gk2N is a Cauchy sequence, there
exists N" 2 N such that lim inf kfPki fPN" k < ". Because f D limi fki almost everywhere,
i2N
Fatous Lemma yields the sought-for conclusion: namely, for any m N" ,
kfP fPm kp D
Z
jf fm jp d
X
Z
lim inf
i2N X
jfki fm jp d
(Fatous Lemma)
D lim inf kfPki fPm kp

i2N
lim inf kfPki fPN" kp

i2N
< "p :
This completes the proof that Lp .X; ; / is a Banach space.
t
u
By specialising to counting measure on N, one obtains the Banach spaces of

`p .N/ of p-summable sequences of complex numbers:
(
` .N/ D a D fk gk2N j k 2 C; for all k 2 N; and
p
1
X
)
jk j < 1 :
p
kD1
In this case, the seminorm of Proposition 5.36 is in fact a norm on this space.
Corollary 5.39. If p 1, then `p .N/ is a Banach space with respect to the norm
kak D
X
k2N
!1=p
jk j
185
Notational Convention If n 2 N, then `p .n/ denotes the normed vector space of

sequences
a D fk gnkD1
of complex numbers, considered as a finite-dimensional subspace of `p .N/.
Which measurable functions belong to L p .X; ; /? In the case of simple
functions, the criterion is quite basic.
Lemma 5.40. If ' is a simple
function on
a measure space .X; ; /, then ' is
p-integrable if and only if ' 1 .C n f0g/ < 1.
t
u

The following proposition is a type of approximation result.
Proposition 5.41. The linear submanifold f'P j ' is simple and p-integrableg is
dense in Lp .X; ; /, for every p 1.
Proof. If f 2 L p .X; ; /, then f is also p-integrable, which implies that both
the real and imaginary parts of f are p-integrable. Moreover, every real-valued
p-integrable function is a difference of nonnegative p-integrable functions, by
equation (3.3). Therefore, it is sufficient to prove that if f is a nonnegative pintegrable
function and if " > 0, then there is a p-integrable simple function ' such
Z
jf 'jp d < "p .
that
X
Assuming f is nonnegative, there exists, by Theorem 3.14, a sequence of simple

functions 'n W X ! R such that 0 'n .x/ f .x/ and limn 'n .x/ D f .x/ for every x 2
X. Therefore, each 'n is p-integrable. Moreover, because 0 f .x/ 'n .x/ f .x/ for
p
all x 2 X and n 2 N, each f 'n is p-integrable and
Z limn .f .x/ 'n .x// D 0. Hence,
by the Dominated Convergence Theorem, lim .f 'n /p d D 0. In particular,
Zn X
given " > 0, there exists an n 2 N such that jf 'n jp d < "p , which shows that
kfP 'Pn k < ".
t
u
The latter part of the proof of Proposition 5.41 does not use the property that the
functions 'n are simple; indeed, the argument shows that the following useful and
rather strong approximation property holds.
Proposition 5.42. If p 1, if f 2 L p .X; ; / is nonnegative, and if ffn gn2N is a
sequence of measurable functions for which 0 fn .x/ f .x/ and lim fn .x/ D f .x/
for all x 2 X, then each fn 2 L p .X; ; / and lim kfP fPn k D 0.
n!1
n!1
We conclude this section with another approximation result, which addresses the
cases of primary interest in analysis.
If f 2 Cc .X/, then the support of f has finite measure, because regular measures
take on finite values on compact sets. Therefore, if M D maxx2X jf .x/j, then
186
5 Banach Spaces
jf jp d M p .supp f / < 1:
jf jp d D
X
supp f
That is, f 2 L p .X; ; / and, thus, determines an element fP 2 Lp .X; ; /. It is in

this sense, then, that in Theorem 5.43 below the vector space Cc .X/ is viewed as
linear submanifold of Lp .X; ; /.
Theorem 5.43. If is a regular measure on a locally compact Hausdorff space X,
and if is a -algebra that contains the Borel sets of X, then Cc .X/ is dense in
Lp .X; ; /.
Proof. Suppose that " > 0 and choose a measurable set E 2 of finite measure. By
the regularity of and Proposition 3.59, there are K; U X such that K is closed,
"p
"p
U is open, K E U, .EnK/ < 2pC1
, and .UnE/ < 2pC1
. Thus,
.UnK/ D .UnE/ C .EnK/ < "p =2p :
The function EjK W K ! 0; 1 is continuous and has, by the Tietze Extension
Theorem (Theoerm 2.43), an extension to a continuous function h W X ! C of
compact support such that supp h U and maxx2X jh.x/j D 1. Thus,
P pD
kPE hk
Z
jE hjp d
X
Z
D
jE hj d C
jE hj d C
UnK
Z
D
jE hjp d
Uc
Z
jE hj d
.1 C 1/p d D 2p .U n K/ < "p :
UnK
UnK
P < ", which proves that the characteristic elements PE are in the
That is, kPE hk
closure of Cc .X/ in Lp .X; ; /.
n
X
Suppose next that ' is an arbitrary p-integrable simple function: ' D
k Ek ,
kD1
where each k 6D 0. By Lemma 5.40, each Ek has finite measure. Let " > 0 and
M D maxk jk j. For each k there exists gk 2 Cc .X/ such that kP Ek gP k k < "=.nM/.
Therefore, with g D 1 g1 C C n gn 2 Cc .X/ we have that
k'P gP k
n
X
jk j kPEk gPk k < " :
kD1
This proves that every simple function is in the closure of Cc .X/ in Lp .X; ; /. By
Proposition 5.41, we deduce that the closure of Cc .X/ is Lp .X; ; /.
t
u
5.5 Essentially-Bounded Measurable Functions
187
5.5 Essentially-Bounded Measurable Functions

Having spent some effort in studying the Banach spaces Lp .X; ; /, for p 2 R with
p 1, this section is devoted to what might be viewed as the case p D 1.
Definition 5.44. If .X; ; / is a measure space, and if f W X ! C is a measurable
function, then the essential range of f is the set ess-ran f of all 2 C for which

f 1 .U/ > 0;
for every neighbourhood U C of .
An alternate description of the essential range is as follows.
Proposition 5.45. If f W X ! C is a measurable function, then
ess-ran f D
f .E/ :
E2; .Ec /D0
Proof. If 2 ess-ran f and if E 2 is such that .Ec / D 0, then necessarily

2 f .E/. If not, then there is an open set U containing that is disjoint from
the closed set f .E/, and so f 1 .U/ Ec ; however, Ec has measure zero while
f 1 .U/ has positive measure (since 2 ess-ran f ), which is a contradiction. Hence,
ess-ran f f .E/ for every measurable set E with .Ec / D 0.
Conversely, if 62 ess-ran f , then .f 1 .U// D 0 for some neighbourhood U of
. Let E D f 1 .U/c . We claim that 62 f .E/. If, on the contrary, were contained
in f .E/, then f .E/ \ U would be non-empty, in contradiction to E \ Ec D ;. Hence,
f .E/ ess-ran f for every E 2 such that .Ec / D 0.
t
u
Definition 5.46. If .X; ; / is a measure space, then the essential supremum of a
measurable function f W X ! C is the quantity
ess-sup f D sup fj j j 2 ess-ran f g:
If the essential supremum of f is finite, then f is said to be essentially bounded.
If ess-sup f D M < 1, then the set fx 2 X j jf .x/j > Mg has measure zero; thus,
it is natural to use the term essentially bounded in describing the function f , even
though f may very well be an unbounded function.
The following proposition is essentially (ha!) self-evident.
Proposition 5.47. If .X; ; / is a measure space, then the set L 1 .X; ; / of
all essentially bounded measurable functions X ! C is a complex vector space.
Furthermore, the function W L 1 .X; ; / ! R defined by
.f / D ess-sup jf j;
is a seminorm on L 1 .X; ; /.
188
5 Banach Spaces
Definition 5.48. If
denotes the equivalence relation on L 1 .X; ; / defined by
f
g if and only if ess-sup jf gj D 0, and if
L1 .X; ; / D L 1 .X; ; /=
;
then L1 .X; ; / is called an L1 -space.
By definition, if f 2 L 1 .X; ; /, then the norm of fP 2 L1 .X; ; / is given by
kfP k D ess-sup jf j:
Theorem 5.49. L1 .X; ; / is a Banach space.
Proof. Assume that ffPn gn2N is a Cauchy sequence in L1 .X; ; /, where each fn 2
L 1 .X; ; /. For all k; n; m 2 N, let
Ek D fx 2 X j jfk .x/j > ess-ran f g and
Fn;m D fx 2 X j jfn .x/ fm .x/j > ess-ran .fn fm /g:
The measurable sets Ek and Fn;m have measure zero, and therefore so does G, where
GD
!
Ek
[ [
!
Fn;m :
n;m
If x 2 Gc , then jfn .x/ fm .x/j kfPn fPm k implies that ffk .x/gk2N is convergent in
C to some complex number denoted by x . Thus, if f W X ! C is defined so that
f .x/ D x for x 2 Gc and f .x/ D 0 for x 2 G, then f is bounded and measurable, and
t
u
kfP fPk k ! 0.
As with Lp -spaces, one can consider sequence spaces:

` .N/ D a D fk gk2N j k 2 C; for all k 2 N; and sup jk j < 1 :
1
k2N
Corollary 5.50. The sequence space `1 .N/ is a Banach space with respect to the
norm
kak D sup jk j:
k2N
For finite-length sequences, we use the notation `1 .n/, as we have already done
with finite-dimensional `p -spaces.
Every function on the measurable space .N; P.N// is continuous; therefore,
using counting measure , we may consider the sequential analogue of C0 .N/,
which is a Banach space denoted by c0 .N/.
5.6 Banach Spaces of Complex Measures
189
Definition 5.51. The set c0 .N/ is defined to be

c0 .N/ D a D fk gk2N j k 2 C; for all k 2 N; and lim jk j D 0 :
k!1
The next result is the L1 -version of Proposition 5.41.

Proposition 5.52. The linear manifold f'P j ' is simple and essentially boundedg is
dense in L1 .X; ; /.
Proof. As in the proof of Proposition 5.41, it is sufficient to prove that if f is a
nonnegative essentially bounded function and if " > 0, then there is a simple function
' such that ess-supjf 'j < ".
The proof of Theorem 3.14 shows that if f is a nonnegative essentially bounded
function, then there exists a monotone-increasing sequence of simple functions 'n
such that 'n .x/ n if f .x/ n and 'n .x/ D j1
if j1
f .x/ < 2jn , for j D 1; : : : ; 2n n.
2n
2n
Now let C D ess-sup f and choose n0 2 N such that n0 > C and n0 > log2 ". Thus,
the set E D fx 2 X j f .x/ n0 g has measure zero and
0 f .x/ 'n0 .x/ <
1
< "; 8 x 62 E:
2n0
Hence, ess-supjf 'n0 j < ".
t
u
5.6 Banach Spaces of Complex Measures

The examples of Banach spaces to this point have involved vectors with finitely
many or countably infinitely many coordinates (entries), or have concerned functions or equivalence classes of functions on a topological or measurable space. The
purpose of this section is to give an example of a Banach space that arises from
measures theory; in this case, the functions involved are defined on a -algebra.
Definition 5.53. If .X; / is a measurable space, then the set
M.X; / D f j is a complex measure on .X; /g
is called the space of complex measures on .X; /.
As noted earlier, the set M.X; / carries the structure of a complex vector space
in that, if j 2 C and j 2 M.X; / for j D 1; 2, then
.1 1 C 2 2 / .E/ D 1 1 .E/ C 2 2 .E/;
for every E 2 . Recall also from Proposition 3.69 that, if 2 M.X; /, then the
function jj W ! 0; 1 defined by
190
5 Banach Spaces
jj.A/ D sup
8
<X
:
E2PA
9
=
j.E/j j PA is a measurable partition of A
;
is a measure on .X; / with jj.X/ < 1.

Theorem 5.54. The function k k on M.X; / defined by kk D jj.X/, for 2
M.X; /, is a norm under which M.X; / is a Banach space.
Proof. The verification that kk D jj.X/ defines a norm on M.X; / is left as an
exercise (Exercise 5.117).
Assume that fk gk2N is a Cauchy sequence in M.X; /. Because j.E/j
jj.E/ jj.X/ D kk for every E 2 and every complex measure , for each
E 2 the sequence fk .E/gk2N is a Cauchy sequence in C. Because C is complete,
for every E 2 there exists a unique .E/ 2 C such that .E/ D limk k .E/. Hence,
we aim to show that the function E 7! .E/ is a complex measure on .X; / and
that limk k k k D 0.
Obviously .;/ D 0. If fEi gìD1 is a finite sequence of pairwise disjoint measurable sets, then

`
[
`
[
Ei D lim k
k!1
iD1
iD1
!
Ei D lim
k!1
`
X
k .Ei / D
iD1
`
X
iD1
lim k .Ei / D
k!1
`
X
.Ei /:
iD1
That is, is finitely additive.

Suppose now that fEi gi2N is a countably infinite sequence of pairwise disjoint
1
[
Ei to obtain a sequence fFj gj2N of
measurable sets. For every j 2 N, define Fj D
iDj
measurable sets with the properties that Fj FjC1 , for every j, and
were true that limj .Fj / D 0, then it would follow from

1
[
!
Ei
iD1
that
1
[
iD1
!
Ei
1
X
iD1
1
\
Fj D ;. If it
jD1
" j1 #
! j1
[
[
X
Ei
.Ei / C .Fj /
Fj D
D
iD1
iD1
.Ei /. To prove that limj .Fj / D 0, let " > 0 be selected
arbitrarily and let N 2 N be such that km n k < " for all m; n N. Thus, in letting
m ! 1 and fixing n N, we obtain j.E/ n .E/j " for all E 2 . Because the
sequence fFj gj2N is descending and has empty intersection, and because the measure
n is finite, Proposition 3.22 on the continuity of measure (routinely modified to
apply to complex measures) yields limj n .Fj / D 0. Thus, there exists a j0 2 N such
that n .Ej / < " for every j j0 . Consequently, if j j0 , then
5.7 Separable Banach Spaces
191
j.Fj /j D j.Fj / n .Fj / C n .Fj /j j.Fj / n .Fj /j C jn .Fj /j 2":

Therefore, limj .Fj / D 0, which completes the proof that is countably additive.
Hence, 2 M.X; /.
To show that limk k k k D 0, suppose that " > 0 and that N 2 N is such that
km n k < " for all m; n N. If fEi griD1 is an arbitrary finite measurable partition
of X and m; n N, then
r
X
jm .Ei / n .Ei /j km n k < ":
iD1
Thus,
r
X
iD1
j.Ei / n .Ei /j D lim
m!1
r
X
jm .Ei / n .Ei /j ":
iD1
Because the inequality above holds regardless of the size r of the partition of X, the
same inequality holds for any countable measurable partition PX of X, and hence
it also holds for the supremum over all countable measurable partitions X. That is,
k n k " for every n N. Therefore, the Cauchy sequence fk gk2N is convergent
in M.X; / to .
t
u

As with topological spaces, the notion of separability is an important feature that a
Banach space might possess.
Definition 5.55. A Banach space V is separable if V has a countable dense subset.
Not surprisingly, finite-dimensional normed vector spaces are separable.
Proposition 5.56. Every finite-dimensional Banach space is separable.
Proof. By Proposition 5.8, for any fixed basis if fv1 ; : : : ; vn g of V there are positive
constants c and C such that

0
0
11=2
11=2
X

n
n
n
X
X

@
c@
jj j2 A
j vj
jj j2 A ;

C

jD1
jD1
jD1
for all 1 ; : : : ; n 2P
C. Because the countable set .Q C iQ/n is dense in the Euclidean
n
space C , if v D njD1 j vj 2 V and if " > 0, then there are 1 ; : : : ; n 2 .Q C iQ/
such that
192
5 Banach Spaces

0
11=2
n

n
n
X
X
X

@
j vj
j vj
jj j j2 A < " :

C
jD1

jD1
jD1
That is, the countable set of all vectors of the form
n
X
j vj , where j 2 .Q C iQ/
jD1
for each j, is dense in V.
t
u
There is a close relation between the separability of C.X/ and the topology of X,
if X is compact and Hausdorff.
Theorem 5.57. The following statements are equivalent for a compact Hausdorff
space X:
1. C.X/ is separable;
2. X is second countable;
3. X is metrisable.
Proof. Assume that C.X/ is separable and
that S D ffn gn2N is a countable dense
subset of C.X/. For each n let Vn D fn1 B1=3 .1/ , which is an open set in X, and
let B D fVn gn2N . Suppose now that x 2 X and U is a neighbourhood of x. Because
X is normal, there is a neighbourhood V of x such that fxg V V U. Hence,
by Urysohns Lemma, there is a continuous function g W X ! 0; 1 with g.x/ D 1
and g.U c / D f0g. By hypothesis, there exists fn 2 S with kfn gk < 13 . In particular,
1
> jfn .x/ g.x/j D jfn .x/ 1j. Thus, if z 2 Vn , then fn .z/ 0, which implies that
3
z 2 .U c /c . Thus, by letting B D Vn , this shows that there exists B 2 B such that
x 2 B U. Hence, B is a basis for the topology of X and, therefore, X is second
countable.
Because Theorem 2.48 asserts that a compact Hausdorff space second countable
if and only if it is metrisable, we assume now that d is a metric on X that induces
the topology of X and we aim to prove that C.X/ is separable. For each n 2 N
consider the open cover fB1=n .x/gn2N of X.[
Because X is compact, we may extract a
k
k
finite subcover: fB1=n .xn;j /gnjD1
. Let U D
fB1=n .xn;j /gnjD1
and consider the subset
D U U in which
n2N

B1=n .xn;j /; B1=m .xm;` / 2 D if and only if B1=n .xn;j / \ B1=m .xm;` / D ;:
Because U is countable, so is D. By Urysohns Lemma, that there exist functions
fn;j;m;` W X ! 0; 1 with the property that
fn;j;m;` .B1=n .xn;j // D f1g and fn;j;m;` .B1=m .xm;` // D f0g;

provided B1=n .xn;j /; B1=m .xm;` / 2 D. Let F be the (countable) set of all such
functions fn;j;m;` and consider the sequence F k defined as follows:
F 0 D f1g; where 1 denotes the constant function x 7! 1; 8 x 2 X;
193
and, for k 2 N,
k
Y
F k D f fj j f1 ; : : : ; fk 2 F g:
jD1
Thus the set S D SpanQCiQ
1
[
!
F
is countable and is closed under sums and
kD0
products. Now let A D S. It is clear that A is a vector space over C, but it is also an
algebra. To verify this last assertion, suppose f ; g 2 A and let " > 0 be given. Then
there are f0 ; g0 2 S such that kf f0 k < " and kg g0 k < ". Further,
kfg f0 g0 k D kfg f0 g C f0 g f0 g0 k kf f0 k kgk C kf0 k kg g0 k
kf f0 k kgk C .kf k C "/kg g0 k
< .kgk C kf k/" C "2 :
Thus, because f0 g0 2 S, we deduce that fg 2 S D A, and hence A is an algebra. We
now show that A separates the points of X.
Let x; y 2 X be distinct and choose n 2 N such that 1n < 12 d.x; y/. Because
k
fB1=n .xn;j /gnjD1
is a cover of X, there are j; i such that x 2 B1=n .xn;j / and y 2 B1=n .xn;i /.
The condition 1n < 12 d.x; y/ implies that B1=n .xn;j / \ B1=n .xn;i / D ;. By the same
reasoning, if m1 < 12 . 1n d.y; xn;i //, then y 2 B1=m .xm;` / B1=n .xn;i / for some xm;` .
Thus, fn;j;m;` .x/ D 1 and fn;j;m;` .y/ D 0, which implies that A separates the points
of X.
Because every element of F is a real-valued function, A is closed under complex
conjugation f 7! f . Moreover, A contains all the constant functions, since F0 A.
Finally, given that A separates the points of X, the Stone-Weierstrass Theorem yields
A D C.X/, and so S is a countable dense subset of C.X/. Hence, C.X/ is separable.
t
u
A similar theme prevails for certain Lp -spaces.
Proposition 5.58. If X a compact metrisable space and if is a regular Borel
measure on a -algebra that contains the Borel sets of X, then Lp .X; ; / is
separable, for all p 2 R such that p 1.
Proof. Let f 2 L p .X; ; / and suppose that " > 0. Theorem 5.43 asserts that kfP
gP k < "=2 for some g 2 C.X/. By Theorem 5.57, there exists a countable subset
C C.X/ such that C is dense in C.X/. In particular, there exists h 2 C such that
maxx2X jg.x/ h.x/j < "=2.X/1=p . Thus,
Z
"p
P p D jg hjp d <
.X/ D ."=2/p :
kPg hk
p
2 .X/
X
P < ", and so the countable set fhP j h 2 C g is dense in Lp .X; ; /.
Thus, kfP hk
t
u
194
5 Banach Spaces
The results established to this point can be used to prove that Lp .Rn ; M.Rn /; mn /
is separable, where mn denotes Lebesgue measure on Rn .
Theorem 5.59. Lp .Rn ; M.Rn /; mn / is separable for all n 2 N, and all p 2 R with
p 1.
t
u

1
In contrast to the results above, the Banach space L .X; ; / is generally

non-separable. Indeed, the sequence space `1 .N/ is not a separable space (Exercise 5.119).
To conclude the discussion on separable spaces, we consider a Banach space of
continuous complex-valued functions which at first glance is not of the form C.X/.
Specifically, let C.R=2
Z/ denote the set of all continuous 2
-periodic functions
f W R ! C. It is straightforward to verify that under the norm
kf k D max jf .t/j;
t2R
C.R=2
Z/ is a Banach algebra with identity 1, the constant function. An important
subset of C.R=2
Z/ is the set T of trigonometric polynomials.
Definition 5.60. A trigonometric polynomial is a 2
-periodic function p W R ! C
of the form
p.t/ D
m
X
k ei kt ;
t 2 R;
(5.8)
kDn
where n m in Z and n ; : : : ; m 2 C.
The set T of all trigonometric polynomials is a complex vector space, closed
under multiplication and complex conjugation, and contains the constant functions.
Therefore, one anticipates that the Stone-Weierstrass Theorem may have a formulation in this context, and indeed it does, yielding the classical theorem of Weierstrass
on the uniform approximation of periodic continuous functions by trigonometric
polynomials.
Proposition 5.61 (Trigonometric Weierstrass Approximation Theorem). If f W
R ! C is a continuous 2
-periodic function, then for every " > 0 there is a
trigonometric polynomial p 2 T such that
jf .t/ p.t/j < "; for every t 2 R :
Proof. Recall that S1 is the unit circle in R2 , which in the present context we view
as the unit circle T D fz 2 C j jzj D 1g in the complex plane. Every f 2 C.R=2
Z/
determines a unique function F 2 C.T/ whereby
F.eit / D f .t/; for every t 2 R :
5.8 Hilbert Space
195
Conversely, each F 2 C.T/ determines a unique f 2 C.R=2

Z/ such that f .t/ D
F.eit / for every t 2 R. In particular, let T C.T/ be the set of all the functions
P of the form P.eit / D p.t/, for trigonometric polynomials p 2 T. Observe T
C.T/ contains the constant functions and is closed under complex conjugation.
0
Furthermore, because eit D eit if and only if t0 t is an integer multiple of 2
,
the function P.eit / D eit separates the points of T. Therefore, the closure of T in
C.T/ is selfadjoint uniform algebra on T, which by the Stone-Weierstrass Theorem
can only be C.T/. Hence, if f 2 C.R=2
Z/ and " > 0, then there is a P 2 T such
that
jf .t/ p.t/j D jF.eit / P.eit /j < "; for every t 2 R;
where F 2 C.T/ is the unique function determined by f and where p 2 T is the
unique trigonometric polynomial determined by P.
u
t
Corollary 5.62. C.R=2
Z/ is a separable Banach space.
t
u
5.8 Hilbert Space

If p and q are conjugate real numbers, then Hlders inequality asserts that the
product of a p-integrable function f with a q-integrable function g is integrable.
In one special case, namely the case in which p D q D 2, the functions f and g come
2
from the same function space, L 2 .X;
Z ; /. For any function f 2 L .X; ; /, we
may write jf j2 as f f to obtain kfP k2 D f f d. More generally, if f ; g 2 L 2 .X; ; /,
X
then f g is integrable, by Hlders inequality, and so we may consider the function

2
on the Cartesian product L 2 .X;
Z ; / L .X; ; / that sends an ordered pair
.f ; g/ to the complex number
f g d. This function is linear in f and conjugateX
linear in g, and has the property that .f ; f / is mapped to kfP k2 . Such a function on
L 2 .X; ; / L 2 .X; ; / is called a sesquilinear form, and these properties of
L2 -spaces motivate the definition of an abstract Hilbert space.
Definition 5.63. An inner product on a complex vector space H is a complexvalued function h; i on the Cartesian product H H satisfying the following
properties for all vectors ; 1 ; 2 ; ; 1 ; 2 2 H and scalars 2 C:
1.
2.
3.
4.
h; i 0, with h; i D 0 if and only if D 0;

h; i D h; i;
h1 C 2 ; i D h1 ; i C h2 ; i and h; 1 C 2 i D h; 1 i C h; 2 i;
h ; i D h; i and h; i D N h; i.
The vector space H, when considered with the inner product h; i, is called an inner
product space.
The simplest example of an inner product space is Cn .
196
5 Banach Spaces
Example 5.64. The vector space Cd is an inner product space, where h; i is
defined by
h; i D
d
X
k k
kD1
3
2 3
1
1
6 :: 7
6 :: 7
for D 4 : 5 ; D 4 : 5 2 Cd .
2
d
d
The Hlder Inequality in the case of L2 .X; ; / is what is called the CauchySchwarz Inequality in abstract Hilbert space.
Proposition 5.65 (Cauchy-Schwarz Inequality). If h; i is an inner product on a
vector space H, then
jh; ij h; i1=2 h; i1=2 ;
(5.9)
for all ; 2 H. Furthermore, if 6D 0, then jh; ij D h; i1=2 h; i1=2 if and only if
D for some 2 C.
Proof. The result is trivially true if h; i D 0. Assume, therefore, that h; i 6D 0.
For any 2 C,
0 h ; i D h; i h; i h; i C j j2 h; i
D h; i 2< . h; i/ C j j2 h; i :
For
D
h; i
;
h; i
the inequality above becomes

0 h; i C
h; i2 h; i

;
jh; ij2
which yields inequality (5.9). Further, note that if jh; ij D h; i1=2 h; i1=2 , then
h;i
h ; i D 0 for D h;i
, which yields D .
t
u
With our experience with L2 -spaces, the following proposition is a natural
consequence of the Cauchy-Schwarz inequality.
Proposition 5.66. If h; i is an inner product on a vector space H, then
kk D h; i1=2
defines a norm k k on H.
(5.10)
5.8 Hilbert Space
197
Proof. If ; 2 H, then h; i C h; i D 2 < .h; i/, and so

k C k2 D h C ; C i D jk2 C 2<. h; i / C kk2
jk2 C 2 jh; ij C kk2 jk2 C 2 kk kk C kk2
D .kk C kk/2 :
Hence, equation (5.10) defines a norm on H.
t
u
Observe that the Cauchy-Schwarz Inequality (5.9) now takes the form
jh; ij kk kk
and leads to the following useful proposition.
Proposition 5.67 (Continuity of the Inner Product). If 0 2 H and " > 0, then for
each 2 H there is a > 0 such that
jh; i h0 ; ij < "
for all 2 H with k 0 k < .
Proof. The assertion is clear if D 0. If 6D 0, then let D "kk1 . Thus, by the
Cauchy-Schwarz inequality,
jh; i h0 ; ij D jh 0 ; ij k 0 k kk < ";
for all 2 H with k 0 k < .
t
u
Definition 5.68. A Hilbert space is an innerpproduct space H such that H is a

Banach space with respect to the norm kk D h; i.
The following example is the most generic example of a Hilbert space.
Example 5.69. The Banach space L2 .X; ; / is a Hilbert space with respect to
the inner product
Z
hfP ; gP i D f g d;
X
where f ; g 2 L 2 .X; ; /.

Proof. We need only note that hfP ; fP i D
jf j2 d, which is the square of the norm
of fP in the Banach space L2 .X; ; /. Hence, the norm on the Banach space
L2 .X; ; / is induced by the inner product given above.
t
u
The concepts of Euclidean geometry greatly influence Hilbert space theory,
starting with the idea of perpendicular vectors.
Definition 5.70. In an inner product space H, two vectors ; 2 H are orthogonal
if h; i D 0.
198
5 Banach Spaces
Proposition 5.71 (Pythagorean Theorem). k C k2 D kk2 C kk2 , for all pairs
of orthogonal vectors ; 2 H.
Proof. Use k C k2 D h C ; C i D kk2 C 2< .h; i/ C kk2 and the fact that
h; i D 0 to obtain k C k2 D kk2 C kk2 .
t
u
A distinguished geometric property of Hilbert space is the parallelogram law
below.
Proposition 5.72 (Jordan-von Neumann Theorem).
space H and if ; 2 H, then
If H is an inner product
k C k2 C k k2 D 2kk2 C 2kk2 :
(5.11)
Furthermore, if V is a normed vector space in which the parallelogram law (5.11)

holds for all ; 2 V, then there is an inner product that induces the norm on V,
Proof. The parallelogram law is verified by expanding the appropriate inner
products. Therefore, suppose now that V is a normed vector space and that
equation (5.11) holds for all ; 2 V. Define h; i W V V ! C by
h; i D k C k2 k k2 C ik C ik2 ik ik2 ;
for ; 2 V. Thus, h; i is an inner product and kk D h; i1=2 for all 2 V.
t
u
The notion of convexity appeared earlier in the context of Jensens inequality.

This fundamental geometric idea is especially important in Hilbert space theory, as
demonstrated by Theorem 5.75 below, which has many important consequences for
Hilbert spaces that are not necessarily true for arbitrary Banach spaces.
Definition 5.73. A subset K of a vector space V is a convex set if
v C .1 /w 2 K
for all 2 0; 1 and for all v; w 2 K.
Definition 5.74. If K is a nonempty subset of a normed vector space V and if v0 2
V, then the distance from v0 to K is denoted by dist .v0 ; K/ and is defined by
dist .v0 ; K/ D inffkv0 vk j v 2 Kg :
(5.12)
The main convexity theorem in Hilbert space is the following result.

Theorem 5.75. If K is a nonempty closed convex subset of a Hilbert space H, and
if 0 2 H, then there is a unique 2 K such that
dist .0 ; K/ D k0 k :
Proof. The convexity of K will be used repeatedly in the following guise: if 1 ; 2 2
K, then 12 .1 C 2 / 2 K.
5.8 Hilbert Space
199
By definition of distance, for each k 2 N there is a vector k 2 K such that

1
k0 k k2 < .dist .0 ; K//2 C :
k
Thus,
k20 .n C m /k2 C km n k2 D k.0 n / C .0 m /k2
Ck.0 n / .0 m /k2
D 2k0 n k2 C 2k0 m k2
< 4 .dist .0 ; K//2 C 1n C m1 ;
where the second equality above follows from the parallelogram law. On the other
hand,
1
4 .dist .0 ; K//2 4k0 .n C m /k2 D k20 .n C m /k2 :
2
Hence, km n k2 < 1n C m1 , which proves that fk gk2N is a Cauchy sequence. Let
2 H denote the limit of this sequence. Because K is closed, lies in K and
dist .0 ; K/ k0 k D k0 k C k k k0 k k C kk k
<
q
.dist .0 ; K//2 C 1k C kk k :
In letting k ! 1, the inequalities above sandwich to the equation

dist .0 ; K/ D k0 k :
To prove the uniqueness of the best approximant , let 0 2 K satisfy dist .0 ; K/ D
k0 0 k. Thus, k0 k D k0 0 k and, by the parallelogram law,

k.0 0 / C .0 /k2 C k.0 0 / .0 /k2 D 2 k0 0 k2 C k0 k2 :
Therefore,
k 0 k2 D 4k0 k2 4k0 12 . C 0 /k2
D 4 .dist .0 ; K//2 4k0 12 . C 0 /k2
4 .dist .0 ; K//2 4 .dist .0 ; K//2
D 0;
and so 0 D .
t
u
200
5 Banach Spaces
Definition 5.76. If S is a nonempty subset of a Hilbert space H, then the orthogonal

complement of S is the subset of H denoted by S? and defined by
S? D f 2 H j h; i D 0; 8 2 Sg :
By linearity of the inner product in the first variable, it is not difficult to see
that S? is a vector subspace of H. By the Cauchy-Schwarz inequality (that is, by
continuity of the inner product in the first variable), S? is closed. Thus, S? is a
subspace of H for every nonempty subset S H.
Proposition 5.77. Let M H be a subspace of a Hilbert space H and let 2 H and
2 M. The following statements are equivalent:
1. dist .; M/ D k k;
2. 2 M ? .
Proof. Assume that dist .; M/ D k k. Let 0 2 M. We aim to prove that h
; 0 i D 0. To this end, consider any 2 C. Because C 0 2 M, k . C 0 /k
dist .; M/ D k k. Thus,

k k2 k. / 0 k2 D k k2 2< h0 ; i C jj2 k0 k2 ;
which yields

2< h0 ; i jj2 k0 k2 :
(5.13)
The complex number h0 ; i is either zero or it is not. If it is zero, then is

orthogonal to 0 , as desired. Thus, assume that h0 ; i 6D 0 and let D t h0 ; i,
for some t > 0. Inequality (5.13) becomes
2t jh0 ; ij2 t2 jh0 ; ij2 k0 k2 :
Because h0 ; i 6D 0 and t > 0, the inequality above implies that
2 t kk2 ;
which is clearly impossible if t ! 0C . Hence, it must be that h0 ; i D 0, which
proves that is orthogonal to every vector 0 2 M. That is, 2 M ? .
Conversely, assume that 2 M ? . If 0 2 M, then is orthogonal to 0
because 0 2 M. Invoking the Pythagorean theorem yields
k 0 k2 D k. / C . 0 /k2 D k k2 C k 0 k2 k k2 :
Thus,
k k inffk 0 k j 0 2 Mg D dist .; M/ ;
which proves that k k D dist .; M/.
t
u
5.8 Hilbert Space
201
Recall from linear algebra that if M and N are linear submanifolds of H, then
M C N denotes the set
M C N D f! C j ! 2 M and 2 Ng
and M C N is itself a linear submanifold of H.
Proposition 5.78. If M and N are subspaces of a Hilbert space H such that M
N ? , then M \ N D f0g and the linear submanifold M C N is a subspace.
Proof. If 2 M \ N, then the hypothesis M N ? implies that 2 N \ N ? , whence
0 D h; i and so D 0.
To show that the linear submanifold M C N is closed in the topology of H, select
a vector in the closure of M C N and suppose that fk gk2N is a sequence in M C N
converging to . Thus, fk gk is necessarily a Cauchy sequence.
For each k 2 N, there are !k 2 M and k 2 N such that k D !k C k . Because the
vectors of M are orthogonal to the vectors of N, the Pythagorean Theorem applies:
km n k2 D k.!m !n / C .m n /k2 D k!m !n k2 C km n k2 :
Thus, the sequences f!k gk2N M and fk gk2N N are Cauchy sequences. Let ! 2
M and 2 N be the limits, respectively, of these sequences. Then D ! C and so
M C N is closed.
t
u
This situation described in Proposition 5.78 is formalised by the following
definition.
Definition 5.79. If M; N H are subspaces of a Hilbert space H such that M N ? ,
then the orthogonal direct sum of M and N is the subspace M C N and is denoted by
M N.
Proposition 5.80. If M H is a subspace, then H D M M ? .
Proof. Obviously M M ? H. Conversely, suppose that 2 H. Because M is a
subspace, M is closed and convex. Thus, by Theorem 5.75, there is a unique 2 M
for which k k D dist .; M/. Therefore, 2 M ? , by Theorem 5.77. Let ! D
and D to get ! 2 M, 2 M ? , and D ! C 2 M M ? . This proves that
H M M?.
t
u
The direct sum decomposition in Proposition 5.80 above is internal in the sense
that one decomposes an existing space H as an orthogonal direct sum of two
subspaces. One could repeat this process finitely or countable infinitely many times
inductively. Alternatively, one frequently has a finite or countable family of Hilbert
spaces and aims to construct their (external) direct sum to create a new Hilbert space.
Proposition 5.81. If is a finite or countable set, and if fHk gk2 is a family of
Hilbert spaces in which h; ; ik denotes the inner product of Hk , then the vector
space
X
M
Hk D f.k /k2 j
kk k2 < 1g
k2
k2
202
5 Banach Spaces
is a Hilbert space with respect to the inner product

h.k /k ; .k /k i D
X
hk ; k ik :
k2
t
u
If Hk D H for every k and for some fixed Hilbert space H, and if D N, for
example, then `2H .N/ denotes the Hilbert space in Proposition 5.81.
5.9 Orthonormal Bases of Hilbert Spaces

Definition 5.82. The elements of a collection O of vectors in an inner-product
space H are said to be orthonormal if each 2 O has norm k k D 1 and h ; i D 0
for all ; 0 2 O in which 0 6D .
If 1 ; : : : ; k are orthonormal and if 2 Span f 1 ; : : : ; k g, then there are complex
numbers 1 ,. . . ,k such that
D
k
X
j j :
jD1
Because the inner product is linear in its first variable, for each ` 2 f1; : : : ; kg we
have that
* k
+
d
X
X
h; ` i D
j j ; ` D
j h j ; ` i D ` :
jD1
jD1
Hence, every 2 Span f 1 ; : : : ; k g is expressed by uniquely by its Fourier series:

D
k
X
h; j i j :
(5.14)
jD1
The Fourier series (5.14) also demonstrates that orthonormal vectors are necessarily linear independent.
Proposition 5.83 (Gram-Schmidt Process). If v1 ; : : : ; vk 2 H are linearly independent vectors in an inner-product space H, then there are orthonormal vectors
1 ; : : : ; k 2 H such that
Spanf 1 ; : : : ; k g D Spanfv1 ; : : : ; vk g:
203
Proof. Let 1 D kv1 k1 v1 and, inductively, for each j let j D kvj wj k1 .vj
Pj1
wj /, where wj D iD1 hvj ; i i i . The vectors 1 ; : : : ; k are orthonormal and are
contained in the k-dimensional subspace Spanfv1 ; : : : ; vk g. Thus, by the linear independence of orthonormal vectors, Spanf 1 ; : : : ; k g and Spanfv1 ; : : : ; vk g coincide.
t
u
As noted earlier in Theorem 5.16, the concept of a linear basis is not really suited
to analysis.
Definition 5.84. A collection B of orthonormal vectors in a Hilbert space H is
called an orthonormal basis if B 0 D B for every set B 0 of orthonormal vectors for
which B 0 B.
In other words, an orthonormal basis is a maximal set of orthonormal vectors.
Proposition 5.85. If B is a set of orthonormal vectors in Hilbert space H, then B
is an orthonormal basis of H if and only if Span B is dense in H.
t
u
The question of existence of orthonormal bases is not unlike that of linear bases
in that it requires Zorns Lemma to prove this fact.
Theorem 5.86. Every nonzero Hilbert space has an orthonormal basis.
Proof. Mimic the proof of Theorem 5.16.
t
u
The cardinality of the basis is related to the topology of the Hilbert space by way
of the following proposition.
Proposition 5.87. A Hilbert space is separable if and only if it has a countable
orthonormal basis. Moreover, all orthonormal bases of a separable Hilbert space
are in bijective correspondence.
Proof. Let H be a Hilbert space. If H has a countable orthonormal basis f k gk2N ,
then the countable set
W D SpanQCiQ f k j k 2 Ng
is dense in H by Proposition 5.85 and by the fact that Q C iQ is dense in C.
Conversely, assume that H is separable. If S is any countable, dense subset of H,
then the linear submanifold
W D SpanC f! j ! 2 Sg
is dense in H. The spanning set S must contain an algebraic basis for the vector
space W. Thus, W has a countable basis and to this basis one can apply the GramSchmidt process to obtain a countable set f k gk2N of orthonormal vectors whose
linear span O is dense in the closure W of W. But W D H, which indicates that
O ? D f0g. In other words, there are no nonzero vectors orthogonal to f k gk2N , which
proves that f k gk2N is a maximal set of orthonormal vectors. That is, f k gk2N is an
orthonormal basis of H.
204
5 Banach Spaces
If H has finite dimension n, then n is an algebraic invariant of the space H: all

linear bases of H must have the same cardinality, whence any two orthonormal bases
of H are in bijective correspondence. If H has infinite dimension and is separable,
then all orthonormal bases of H are countably infinite and are, therefore, in bijective
correspondence with one another.
t
u
It is also true that if H is a nonseparable Hilbert space, then all orthonormal bases
of H are in bijective correspondence. The proof of this fact requires some cardinal
arithmetic; however, we will not need to use such a theorem in what follows, we
will not pursue this result here.
Hilbert spaces with a countable orthonormal basis are especially easy to analyse.
The following theorem, which is an abstraction of classical Fourier series, illustrates
this fact.
Theorem 5.88. If f k gk2N is an orthonormal basis of a Hilbert space H, then, for
every 2 H,

n

X

(5.15)
lim
h; k i k D 0 :
n!1

kD1
Proof. For each n 2 N let

n D
n
X
h; k i k :
kD1
Observe that
0 k n k2 D h n ; n i D kk2
n
X
jh; k ij2 :
kD1
Hence,
lim
n
X
n!1
jh; k ij2 kk2 ;
kD1
which implies that the sequence fn gn2N is a Cauchy sequence. Because H is a
Hilbert space, this sequence has a limit 0 2 H. We shall prove that 0 D .
Choose k 2 N. Direct computation shows that h n ; k i D 0 for every n k.
Hence, if n k, then
j h. 0 /; k i j D j h. n C n 0 /; k i j
D j h. n /; k i C h.n 0 /; k i j
D j h.n 0 /; k i j
kn 0 k :
205
(The final inequality is the Cauchy-Schwarz inequality.) Thus,
0 h. 0 /; k i lim kn 0 k D 0 :

n!1
Therefore, 0 is orthogonal to every vector k in the orthonormal basis. Thus,

0 D 0.
t
u
Equation (5.15) asserts that if 2 H and f k gk2N is an orthonormal basis for H,
then

n

X

lim
h; k i k D 0 :
n!1

kD1
This will be expressed as

D
X
h; k i k :
(5.16)
k2N
Definition 5.89. If f k gk2N is an orthonormal basis of a separable Hilbert space H,

then the series
X
D
h; k i k
k2N
is called a Fourier series decomposition of 2 H, and the complex numbers h; k i

are called the Fourier coefficients of .
Proposition 5.90. Assume that f k gk2N is an orthonormal basis for a separable
Hilbert space H.
1. (Parsevals Equation) For every ; 2 H,
h; i D
h; k ih; k i :
k2N
2. For every 2 H,
kk2 D
jh; k ij2 :
k2N
Proof. Express in its Fourier series decomposition

X
D
h; k i k ;
k2N
and consider
n D
n
X
kD1
h; k i k :
(5.17)
206
5 Banach Spaces
Because limn!1 kn k D 0, Proposition 5.67 shows that

h; i D lim hn ; i
n!1
D lim
n!1
n
X
h; k i h k ; i
kD1
X
h; k ih; k i :
k2N
This proves Parsevals Equation. The equation

kk2 D
jh; k ij2
k2N
follows from Parsevals Equation because kk2 D h; i.
t
u
We conclude with some examples of orthonormal bases.

q
Example 5.91 (Legendre Polynomials). If 0 .t/ D 12 and
r
k .t/ D

2k C 1 1 dk 2
.t 1/k ;
k 2k k dtk
k 2 N;
2
then f P k g1
kD0 is an orthonormal basis of L .1; 1; M; m/.
Proof. The functions k are obtained from the linearly independent functions
1; t; t2 ; : : : by the Gram-Schmidt process. Thus,
Span L D Span f1; t; t2 ; t3 ; : : : g ;
where L denotes the set of Legendre polynomials. By the Weierstrass Approximation Theorem, if # 2 C.1; 1/ and " > 0, then there is a element f 2
Span L such that j#.t/ f .t/j < " for all t 2 1; 1. Furthermore, Theorem 5.43
states that C.1; 1/ is dense in L2 .1; 1; M; m/. Hence, Span L is dense in
L2 .1; 1; M; m/, and so L is an orthonormal basis of L2 .1; 1; M; m/.
t
u
The next example is drawn from classical Fourier series.
Example 5.92 (Classical Fourier Series). If the function n W
;
! C
given by
ei nt
n .t/ D p ;
2
for every n 2 Z, then f P n gn2Z is an orthonormal basis of L2 .

;
; M; m/.
(5.18)
207
Proof. Because
Z
h n ; m i D
n .t/ m .t/ dt D
1
2
ei .nm/t dt;
we have h n ; m i D 1 if m D n and h n ; m i D 0 otherwise. Thus, f P n gn2Z is a set

of orthonormal vectors in L2 .
;
; M; m/.
Theorem 5.61 shows that Span f n gn2Z is uniformly dense in C.R=2
Z/, the
space of all continuous 2
-periodic functions R ! C. Furthermore, C.
;
/ is
a dense linear submanifold of L2 .
;
; M; m/ (by Theorem 5.43). Therefore, it
is sufficient to show that every f 2 C.
;
/ can be approximated to within "
in the norm of L2 .
;
; M; m/ by a 2
-periodic continuous function h. To this
end, choose f 2 C.
;
/ and let " > 0. Let M D max fjf .t/j j t 2
;
g and
"2
choose > 0 such that < 8M
2 . Let h 2 C.
;
/ be the function that agrees
with f on
C ;
, is a straight line from the point .
; 0/ to the point
.
C ; f .
C //, and is a straight line from the point .
; f .
// to the
point .
; 0/. Thus, jf .t/ h.t/j D 0 for t 2
C ;
and jf .t/ h.t/j 2M for
all t 62
C ;
. Hence,
kf hk2 D
jf hj2 dm C

;
C
;
jf hj2 dm 8M 2 :
That is, kf hk < ".
t
u
The classical Fourier coefficients of f 2 L 2 .

;
; M; m/ are the complex
numbers fO .k/ defined by
fO .k/ D hfP ; P k i :
In particular, if f W
;
! C is a continuous 2
-periodic function, then the
classical Fourier series
X
fO .k/ ei kt
k2Z
of f converges to f in the sense that

Z
lim
n!1

;
jf .t/
n
X
!
Of .k/ ei kt j2 dt D 0 :
kDn
Very old books on analysis refer to this as convergence in the mean to f . But
from our perspective, convergence in the mean is more plainly understood to be
convergence in Hilbert space, as explained by Theorem 5.88.
208
5 Banach Spaces
5.10 Sums of Vectors and Hilbert Spaces

The definition of Banach space involves Cauchy sequences, but it is sometimes
useful to extend this idea to nets.
Definition 5.93. If .; / is a directed set, then a net fv g2 in a normed vector
space V is said to be a Cauchy net if for every " > 0 there exists 0 2 such that
kv v k < ", for all ; 2 such that 0 and 0 .
Proposition 5.94. In a normed vector space V, every convergent net is a Cauchy
net. Conversely, if V is a Banach space, then every Cauchy net is a convergent net.
t
u
If is a set, not necessarily ordered, then F ./ D fF j F is a finite setg is a

directed set under inclusion.
Definition 5.95. Assume that V is a normed vector space and that fv g2 is a
collection of vectors in V.
X
1. A partial sum of the collection fv g2 is a vector vF of the form vF D
v ,
2F
for some finite subset F .

2. The set fvF gF2F ./ of partial sums is convergent if the net
XfvF gF2F ./ is
convergent in V; the limit of this convergent net is denoted by
v .
2
Proposition
5.96. If fv g2 is aX
family of vectors in a Banach space V, and if
X
kv k is convergent in R, then
v is convergent in V.
2
Proof. Let " > 0. Because
2
kv k is convergent, the net of partial sums is Cauchy.
2
Thus, there is an F0 2 F ./ such that the partial sums
kv k and
2F1
X
2F2
kv k
F1 and F0 F2 . Indeed, the same is true of the partial

differ byX
less than " of F0X
sums
kv k and
kv k. Therefore, by the triangle inequality,
2.F1 [F2 /
2F0

X
X
X

v
v
v

2F1
2F2
2.F1 [F2 /nF0
X
X

kv k
kv k
2.F1 [F2 /
2F0
< ":
5.10 Sums of Vectors and Hilbert Spaces
209
Hence, the net fvF gF2F ./ of partial sums of the series
v is Cauchy, and so the
2
net is convergent by Proposition 5.94.

Proposition 5.81 indicates how the direct sum
t
u
Hn of a countable family
n2N
fHn gn2N of Hilbert spaces Hn is itself a Hilbert space. It useful to be able to carry
out a direct sum constructionY
for arbitrary families
M fH g2 of Hilbert spaces H .
In the Cartesian product
H , denote by
H the set of all D . /2 for
which
M
2
2
k k2 is convergent. By the usual Hilbert space inequalities, we see that
2
H is a complex vector space and that, if D . / ; D . / 2
2
H , then
2
h; i D
X
h ; i
defines an inner product on
H .
2
Definition 5.97. The inner product space

family fH g2 of Hilbert spaces H .
H is called the direct sum of the
2
Proposition 5.98. The direct sum of a family of Hilbert spaces is a Hilbert space.
n
n
n
Proof.
M Suppose that f gn2N is a Cauchyn sequence of vectors D . / in
H . Thus, for each , the sequence f gn2N is a Cauchy sequence in H ; let
2
2 H denote the limit of this sequence and consider D . / 2
H .
Let " > 0. Because f n gn2N is a Cauchy sequence, there exists n0 2 N such that
k n k < " for all m; n n0 . Thus, for any F 2 F ./ and m n0 ,
m
km n0 k2 k m n0 k2 < "2 :
2F
Thus, in letting m ! 1,
X
2F
k n0 k2 "2 ;
210
5 Banach Spaces
for any finite subset F . Therefore,

that n0 2
H and, thus, 2
2
k n0 k2 converges, which implies
M2
H . The inequality k n0 k " shows
2
that is the limit of the Cauchy sequence f n gn2N . Hence,
H is a Hilbert
2
space.
t
u
Problems
5.99. Prove that if V is a normed vector space, then the maps a W V V ! V and
m W C V ! V defined by
a.v1 ; v2 / D v1 C v2
and
m.; v/ D v; 8 2 C; v; v1 ; v2 2 V;
are continuous.
5.100. Let V be a normed vector space.
1. Prove that the open ball Br .v0 / is a convex set, for every v0 2 V and r > 0.
2. Prove that the closure C of a convex subset C V is convex.
5.101. Suppose that V and W are Banach spaces. On the Cartesian product V W
define
k.v; w/k D max fkvk; kwkg :
1. Show that k k is a norm on the vector space V W under which V W is a

Banach space.
2. Prove that the norm topology on V W coincides with the product topology on
V W.
5.102. If T and T 0 are the norm topologies on a vector space V induced,
respectively, by equivalent norms k k and k k0 on V, then prove that T D T 0 .
5.103. Suppose that X is a compact topological space and that is a finite measure
on the Borel sets of X.
1. Prove that there exists C > 0 such that kf k1 Ckf k1 for every f 2 C.X/.
2. Give an example of a compact space X and a finite measure on the Borel sets
of X in which k k1 and k k1 are not equivalent norms on C.X/.
5.104. In a normed vector space V, prove that
j kv1 k kv2 k j kv1 v2 k ;
8 v1 ; v2 2 V :
Problems
211
5.105. Let be a seminorm on a vector space V.

1. Prove that the function d W V V ! 0; 1/ defined by
d.v; w/ D .v w/ ;
v; w 2 V ;
is a pseudo-metric on V.
2. With respect to the topology on V induced by the seminorm , prove that
the vector space operations (scalar multiplication and vector addition) are
continuous. That is, prove that V is a topological vector space.
5.106. If is a seminorm on a topological vector space, and if
is the relation on
V defined by
v
w
if
.v w/ D 0 ;
then prove that

is an equivalence relation. Moreover, if the equivalence classes of
elements of V are denoted by
vP D fw 2 V j w
vg ;
then prove the following assertions:
1. the set V=
of equivalence classes is a vector space under the operations
P w/ ; v; w 2 V ;
vP C wP D .v C
P
vP D .v/ ; 2 C ; v 2 V I
2. the function
kvk
P D .v/ ;
v 2V;
is a norm on V=
.
5.107. Let .X; ; / be a measure space and p; q 2 .1; 1/ be conjugate. Assume
that f ; g W X ! R are nonnegative measurable functions. Prove that if f is p-integrable
and g is q-integrable, then
1=p Z
Z
Z
fg d D
X
1=q
f d
X
g d
X
if and only if there is a complex number such that f p D gq or gq D f p almost

everywhere.
p
5.108. Consider the function f .t/ D t on a closed interval 0; b, b > 0. Prove that
there is a sequence of polynomials pn such that
1. pn .0/D 0, for every n 2N, and
p
2. lim max j t pn .t/j D 0.
n!1
t20;b
212
5 Banach Spaces
5.109. Prove that if X R is compact and if f W X ! C is a continuous function,

then for every " > 0 there is a polynomial p with complex coefficients such that
jf .t/ p.t/j < " for all t 2 X.
5.110. Prove that if X is a locally compact space, then Cc .X/ is a subspace of Cb .X/.
5.111. If X is a locally compact space, then prove that
1. C0 .X/ is a subspace of Cb .X/, and
2. fg 2 C0 .X/, for all f 2 C0 .X/ and g 2 Cb .X/.
5.112. Prove that if ' is a simple
function on a measure space .X; ; /, then ' is

p-integrable if and only if ' 1 .C n f0g/ < 1.
5.113. Suppose that f is an essentially bounded function on a measure space
.X; ; /. Prove that there exists E 2 such that .E/ D 0 and sup jf .t/j < 1
for all t 2 X n E.
5.114. Prove that if f is an essentially bounded function on a measure space
.X; ; /, then

ess-sup f D inf 2 R j jf j1 .; 1/ has measure zero :
5.115. Assume that f 2 L 1 .X; ; / is nonnegative and that ffn gn2N is a
monotone-increasing sequence of measurable functions for which 0 fn .x/ f .x/
and lim fn .x/ D f .x/ for all x 2 X. Prove that each fn 2 L 1 .X; ; / and that
n!1
lim kfP fPn k D 0.
n!1
5.116. Prove that L1 .X; ; / is an abelian Banach algebra.

5.117. Prove that the function k k on M.X; / defined by kk D jj.X/ is a norm.
d
5.118. Prove that Lp .Rd ; M.R
S /; md / is adseparable Banach space, for every p 1.
d
(Suggestion: Note that R D n2N n; n .)
5.119. Prove that the Banach space `1 .N/ is not separable.

5.120. Prove that the Banach space C.R=2
Z/ of continuous 2
-periodic
complex-valued functions is separable.
5.121. Prove that in a Hilbert space H, S? is a subspace, for every S H.
5.122. In a Hilbert space H, let M1 and M2 be closed subspaces.
1. Prove that .M1 C M2 /? D M1? \ M2? .
2. Prove or find a counterexample to .M1 \ M2 /? D M1? C M2? .
5.123. If is a finite or countable set, and if fHk gk2 is a family of Hilbert spaces
in which h; ; ik denotes the inner product of Hk , then prove that the vector space
X
M
Hk D f.k /k2 j
kk k2 < 1g
k2
k2
Problems
213
is a Hilbert space with respect to the inner product

h.k /k ; .k /k i D
X
hk ; k ik :
k2
5.124. Prove that if B H is a set of orthonormal vectors in a Hilbert space H,

then B H is an orthonormal basis of H if and only if Span B is dense in H.
5.125. Prove that every nonzero Hilbert space has an orthonormal basis. In
particular, if O H is a set of orthonormal vectors, then prove that H has an
orthonormal basis B that contains O.
5.126. Suppose that in a Hilbert space H, 1 ; : : : ; n are orthonormal vectors. Prove
that if 1 ; : : : ; n 2 H satisfy kj j k < n1=2 for each j, then 1 ; : : : ; n are linearly
independent.
5.127. Consider the Fourier series expansions of f1 .t/ D t and f2 .t/ D et in
L 2 .
;
/. Calculate the Fourier series of each fj and use Parsevals identity to
establish each of the following formulae:
1
X
1
2
I
D
n2
6
nD1
1
X
nD1
n2 C 2
coth.
/ :
5.128. Assume that V is a normed vector space.

1. Prove that every convergent net in V is a Cauchy net.
2. If V is a Banach space, then prove that every Cauchy net in V is a convergent net.
Chapter 6
Dual Spaces
In considering the elements of the vector space Rd as column vectors, the vector
space .Rd /t of row vectors is obviously related to Rd , but is not necessarily identical
to it. What, therefore, is one to make of row vectors and of the transpose 7! t
operation applied to column vectors ? The usual product of matrices and vectors
indicates that each t is a linear transformation Rd ! R via t ./ D t , for
(column) vectors 2 Rd . This is perhaps the simplest instance of duality, which
is an association of a closely related vector space V of linear maps V ! R to a
given real vector space V. This idea is at the heart of functional analysis (indeed,
this is where the functional part of functional analysis enters the picture), and
this notion is developed and explored in the present chapter.
6.1 Operators
Recall that a linear transformation from a vector space V to a vector space W (both
over some field F) is a function T W V ! W in which T is additive and homogenous;
that is, T.1 v1 C2 v2 / D 1 T.v1 /C2 T.v2 / for all v1 ; v2 2 V and 1 ; 2 2 F. Unless
the context leads to ambiguity, the notation Tv shall be used in place of T.v/ for
linear transformations T W V ! W and v 2 V.
Our interest is with vector spaces that are Banach spaces; thus, we shall require
that linear transformations between Banach space be continuous. In this regard, the
essential concept is that of boundedness of a linear transformation.
Definition 6.1. If V and W are normed vector spaces, then a linear transformation
T W V ! W is bounded if there is a constant M > 0 such that kTvk Mkvk, for
every v 2 V; otherwise, T is unbounded.

DOI 10.1007/978-3-319-45633-1_6
215
216
6 Dual Spaces
A bounded linear transformation is called a bounded operator, or simply

an operator, while an unbounded linear transformation is called an unbounded
operator.
In principle, a given normed vector space can admit both bounded and unbounded
operators, as the following example shows.
Example 6.2. Let Ct be the normed vector space of complex polynomials with
norm kf k D maxt20;1 jf .t/j, and consider the linear transformations D W Ct ! Ct
R1
and J W Ct ! C defined by Df D dfdt and Jf D 0 f .t/dt, respectively. Then J is
bounded and D is unbounded.
Proof. If fn 2 Ct is fn .t/ D tn , then Dfn D nfn1 for n 2. Thus, kfn k D 1 for every
n, but kDfn k D n 1 D .n 1/kfn k; that is, no constant M > 0 exists for which
kDf k Mkf k for every f 2 V. Hence, D is an unbounded operator.
On the other hand, for every f 2 V,
Z 1
Z 1
Z 1
kJf k D f .t/ dt
jf .t/j dt
kf k dt D kf k;
0
which shows that J is a bounded operator.
t
u
As is often the case, finite-dimensional spaces do not exhibit any exotic features:
all linear maps are necessarily bounded.
Proposition 6.3. Assume that V and W are normed vector spaces.
1. If V has finite dimension, then every linear transformation T W V ! W is bounded.
2. If V has infinite dimension, and if W 6D f0g, then there is a linear transformation
T W V ! W such that T is unbounded.
Proof. Suppose that V has finite dimension, and fix a basis B D fv1 ; : : : ; vn g
of V. Set
0
11=2
n
X
D@
kTvj k2 A :
jD1
By the triangle inequality and the Cauchy-Schwarz inequality,

0

1
11=2
0
n

n
n
n
X
X
X
X

T @
j vj A
j Tvj
jj j kTvj k @
jj j2 A :
D

D
jD1

jD1
jD1
jD1
Proposition 5.8 and its proof show that there is a constant C > 0 such that

11=2
0
X

n
n
X

2A

@
jj j
C
j vj
:
jD1

jD1
Hence, with M D C , we obtain kTvk Mkvk, for every v 2 V.
6.1 Operators
217
Next, suppose that V has infinite dimension and let B be a linear basis of V. From
B select a countably infinite subset fyk j k 2 Ng B. Without loss of generality, each
vector of B may be assumed to be of norm 1. Because W is a nonzero vector space
there is at least one vector w 2 W of norm kwk D 1: Define a linear transformation
T W V ! W by the following action on the vectors of the linear basis B:
T.yk / D k w ; for each k 2 N I
T.y/ D 0 ; for each y 2 Bnfyk j k 2 Ng :
Note that kT.yk /k D k kwk D k. Because kyk k D 1, for every k 2 N, the linear
transformation T is unbounded.
t
u
The importance of boundedness for linear maps is that boundedness is a synonym
for continuity.
Proposition 6.4. The following statements are equivalent for a linear transformation T W V ! W between normed vector spaces V and W:
1. T is bounded;
2. T is continuous.
Proof. Assume that M > 0 is such that kT.v/k Mkvk, for all v 2 V. Fix v0 2 V.
By the linearity of T, kT.v/ T.v0 /k D kT.v v0 /k Mkv v0 k. Thus, if " > 0
and if D "=M, then kv v0 k < implies that kT.v/ T.v0 /k < ". That is, T is
continuous at v0 . Hence, as the choice of v0 2 V is arbitrary, T is continuous on V.
Conversely, assume that T is continuous. In particular, T is continuous at 0. Thus,
for " D 1 there is a > 0 such that kwk < implies kT.w/k < 1. Let M D 2=. If
v 2 V is nonzero, then let w D 2kvk

v. Thus, kwk < and so kT.w/k < 1. That is,
kTvk Mkvk.
t
u
The next proposition is the first step toward the study of operators in the context
of analysis.
Definition 6.5. If V and W are normed vector spaces, then the set of all operators
T W V ! W is denoted by B.V; W/. In the case where W D V, the notation B.V/ is
used in place of B.V; V/.
Proposition 6.6. Assume that V and W are normed vector spaces.
1. The set B.V; W/ is a vector space under the operations .T1 C T2 /.v/ D T1 .v/ C
T2 .v/ and . T/.v/ D T.v/ for T; T1 ; T2 2 B.V; W/ and v 2 V, 2 C.
2. The function k k W B.V; W/ ! R defined by
kT.v/k
06Dv2V kvk
kTk D sup
is a norm on B.V; W/.

3. If W is a Banach space, then so is B.V; W/.
(6.1)
218
6 Dual Spaces
Proof. It is clear that B.V; W/ has the indicated structure of a vector space, and so
we turn to the proof that (6.1) defines a norm on B.V; W/. To this end, note that
(using the fact that every linear transformation sends zero to zero) equation (6.1) is
equivalent to
kT.v/k kTk kvk ;
8 v 2 V:
(6.2)
Thus, kTk D 0 only if kT.v/k D 0 for every v 2 V; hence, T.v/ D 0 for all v 2 V.
This proves that T is the zero transformation: T D 0. It is also clear that kTk D
jj kTk, and so we now establish the triangle inequality. Let T1 ; T2 2 B.V; W/. For
every v 2 V,
k.T1 C T2 /vk D kT1 .v/ C T2 .v/k
kT1 .v/k C kT2 .v/k
.kT1 k C kT2 k/ k.v/k;
and so
kT1 C T2 .v/k
kT1 k C kT2 k:
kvk
06Dv2V
kT1 C T2 k D sup
This completes the proof that equation (6.1) defines a norm on B.V; W/.
Suppose now that W is a Banach space and select any Cauchy sequence fTk gk2N
in B.V; W/. Inequality (6.2) implies that fTk .v/gk2N is a Cauchy sequence in W for
every v 2 V. Because W is a Banach space, each sequence fTk .v/gk2N is convergent
in W; denote the limit by T.v/. Note that the map v 7! T.v/ is indeed a linear
transformation T W V ! W. It remains to show that T is bounded and that limk kTk
Tk D 0.
Let " > 0. Because fTk gk2N is a Cauchy sequence, there exists N" 2 N such that
kTn Tm k < " for all n; m N" . If v 2 V and if n N" , then
kT.v/ Tn .v/k kT.v/ Tm .v/k C kTm .v/ Tn .v/k
kT.v/ Tm .v/k C kTm Tn k kvk :
As the inequalities above are true for all m 2 N,
kT.v/ Tn .v/k inf .kT.v/ Tm .v/k C kTm Tn k kvk/
m2N
0 C " kvk:
6.1 Operators
219
Hence, if n N" is fixed, then T Tn is bounded and kT Tn k ". Therefore,

Tn C .T Tn / D T is bounded and kT Tn k < " for all n N" . This proves that the
Cauchy sequence fTk gk2N converges in B.V; W/ to T 2 B.V; W/.
t
u
Two linear structures affiliated with an operator T 2 B.V; W/ are recalled below
from linear algebra.
Definition 6.7. If T W V ! W is a linear transformation between vector spaces V
and W, then
1. the kernel of T is the set ker T D fv 2 V j Tv D 0g, and
2. the range of T is the set ran T D fTv 2 W j v 2 Vg.
It is clear that both ker T and ran T are linear submanifolds of V and W,
respectively. If T is continuous (that is, bounded), then ker T is normed closed and,
hence, is a subspace; however, the range of a bounded operator need not be closed
in general. There is one important instance in which ran T is always a subspace, and
that occurs with operators T that preserve the norms of vectors.
Definition 6.8. An operator T W V ! W acting on normed vector spaces V and W
is an isometry if kTvk D kvk for every v 2 V.
Of course, every isometry T is of norm kTk D 1. Moreover, kTvk D kvk implies
that Tv D 0 if and only if v D 0, and so every isometry is an injection. On the other
hand, by scaling an operator T by D kTk1 , every nonzero operator is a scalar
multiple of an operator of norm 1, and so not every operator of norm 1 need be an
isometry.
Proposition 6.9. If V is a Banach space and if T 2 B.V; W/ is an isometry, then
ran T is a subspace of W.
Proof. Let w 2 W be in the closure of the range of T. Thus, kw T.vn /k ! 0,
for some sequence of vectors vn 2 V, and so fT.vn /gn2N is a Cauchy sequence in W.
Because kTvn Tvm k D kT.vn vm /k D kvn vm k, the sequence fvn gn2N is Cauchy
in V. As V is a Banach space, there is a limit v 2 V to this sequence. Moreover,
kTv wk kT.v vn /k C kTvn wk D kv vn k C kTvn wk;
8 n 2 N:
Hence, w D Tv. That is, the range of T contains all of its limit points, implying that
the range of T is closed.
t
u
Definition 6.10. Two normed vector spaces V and W are said to be isometrically
isomorphic if there is a linear isometry T W V ! W such that T is a surjection.
Thus, if V and W are Banach spaces, and if T W V ! W is an isometry, then V is
isometrically isomorphic to the range of T (which by, Proposition 6.9, is a subspace
of W). Hence, T embeds V into W and in so doing preserves the Banach space
structure of V. In such cases, we say that W contains a copy of V because inside
W that copy of V appears (as a Banach space) no different from V itself.
220
6 Dual Spaces
Example 6.11. If ' W R ! C is a continuous 2

-periodic function such that
j'.t/j D 1 for all t 2 R, and if 1 p < 1, then the linear map M' W Lp .T/ ! Lp .T/
P / is an isometric isomorphism.
given by T fP D .'f
Further study of operators on Banach and Hilbert spaces will be taken up in later
chapters; the remainder of the present chapter is devoted to the quite special case
of operators that map complex normed vector spaces V to the 1-dimensional vector
space C.
6.2 Linear Functionals

Definition 6.12. If V is a normed vector space, then a linear functional on V is an
operator ' W V ! C.
The familiar operation of integration is a basic example of a linear functional.
Example 6.13. If .X; ; / is a measure space, then the map ' W L1 .X; ; / ! C
defined by
Z
P
' .f / D f d;
X
1
for f 2 L .X; ; /, is a linear functional on L1 .X; ; /.

Proof. The map ' is obviously linear. To show it is bounded, use the triangle
inequality:
Z
Z
j' .fP /j D f d jf j d D kfP k:

X
Hence, ' is bounded of norm k' k 1.
t
u
Another familiar example of a linear functional is drawn from linear algebra.

Example 6.14. If 2 Cn , then the map ' W Cn ! C defined by
' ./ D
n
X
j j ;
jD1
for 2 Cn , is a linear functional on Cn . Conversely, if ' is a linear functional on

Cn , then there exists a unique 2 Cn such that ' D ' .
Proof. It is clear that ' is a linear functional. Conversely, any linear map of Cn
onto C will be given by a 1 n matrix whose action on Cn is precisely that of ' ,
for some 2 Cn .
t
u
The norm of k' k depends upon the choice of norm for Cn .
6.2 Linear Functionals
221
Definition 6.15. The set B.V; C/ of all linear functionals on V is called the dual
space of V and is denoted by V .
By Proposition 6.6, the dual space V is a Banach space for every normed
vector space V. The determination of V from V (or vice versa) is an important
and nontrivial problem. The next result is an example of an instance in which V is
determined precisely from V.
Proposition 6.16 below is the first of many results that have the moniker Riesz
Representation Theorem. Other Riesz Representation Theorems in the present
book appear as Theorems 6.40, 6.51, and 10.1.
Proposition 6.16 (Riesz). Suppose that p and q are conjugate real numbers.
1. If g D .gk /k2N 2 `q .N/, then the function ' W `p .N/ ! C defined by
'.f / D
1
X
fk gk ;
kD1
for f D .fk /k2N 2 `p .N/, is a linear functional on `p .N/ of norm k'k D kgk.
2. For each ' 2 .`p .N// there is a unique g 2 `q .N/ of norm kgk D k'k such that
'.f / D
1
X
fk gk ;
kD1
for every f D .fk /k2N 2 `p .N/.

Proof. The proof of the first assertion is left to the reader as Exercise 6.54.
To prove the second assertion, select ' 2 .`p .N// . If ek 2 `p .N/ is the vector
with 1 in position k and 0 elsewhere, then let gk D '.ek / for each k 2 N. Select and
fix n 2 N, and consider the element f D .fk /k2N 2 `p .N/ in which fk D 0 for every
k > n and fk D jgk jq1 sgn .gk / for each k D 1; : : : ; n, where, for any 2 C, sgn is
given by 0 if D 0 and by jj
otherwise. Thus,
fk gk D jgk jq1 .gk sgn .gk // D jgk jq and jfk jp D jgk jpCqp D jgk jq :
Thus,
'.f / D '
n
X
!
fk ek D
kD1
n
X
fk '.ek / D
kD1
n
X
jgk jq ;
kD1
and so
n
X
kD1
jgk jq D j'.f /j k'k kf k D k'k
n
X
kD1
!1=p
jfk jp
D k'k
n
X
kD1
!1=p
jgk jq
222
6 Dual Spaces
Thus,
k'k
n
X
!1 1p
jgk jq
n
X
kD1
!1=q
jgk jq
kD1
Because the inequality above is true for arbitrary n 2 N, we deduce that g 2 `q .N/
and that kgk k'k.
Next consider hk D jgk jq1 sgn .gk / for every k 2 N. Because
1
X
!1=p
jhk j
kD1
1
X
!1=p
jgk j
D kgkq=p ;
kD1
we have that h D .hk /k2N 2 `p .N/. Thus, by the continuity of ', the triangle
inequality, and Hlders Inequality, we have
! 1
1
1
X
X
X

j'.h/j D '
hk ek D
hk gk
jhk j jgk j khk kgk;

kD1
kD1
kD1
which implies that k'k kgk. This proves that k'k D kgk.
Lastly, because each f 2 `p .N/ is expressed uniquely in terms of the vectors
fek gk2N `p .N/, the choice of g 2 `q .N/ is also uniquely determined from the
equations gk D '.ek /, k 2 N.
t
u
Corollary 6.17. If p and q are conjugate real numbers, then .`p .N// and `q .N/
are isometrically isomorphic.
Proof. Let W `q .N/ ! .`p .N// be given by g D 'g , where 'g 2 .`p .N//
satisfies
'g .f / D
1
X
fk gk ;
kD1
for every f D .fk /k2N 2 `p .N/. The map is plainly linear and is also, by Proposition 6.16, isometric and surjective. Hence, is a linear isometric isomorphism of
.`p .N// and `q .N/.
t
u
Although Proposition 6.16 and Corollary 6.17 address the case of `p spaces over
N, precisely the same results hold true for the finite-dimensional Banach spaces
`p .n/, for every n 2 N.
Some other accessible examples of interesting dual spaces (such as the dual of
`1 .N/) are addressed in Exercises 6.57 and 6.56. There is an integral version of
Proposition 6.16, but establishing it requires somewhat more effort, and so further
discussion of it is deferred to Section 6.6.
6.3 Hahn-Banach Extension Theorem
223

Notwithstanding the explicit determination of the duals of `p -spaces in the previous
section, the analysis of the dual space V of abstract infinite-dimensional normed
vector spaces V is somewhat delicate. For example, to begin with, it is not at all
obvious that V has any nonzero elements whatsoever. Proving that V has elements
in abundance is the principal objective of the present section, and the main result
developed here, the Hahn-Banach Extension Theorem, is surely the most important
foundational theorem in functional analysis. At its heart is the following linearalgebraic theorem.
Theorem 6.18 (Hahn-Banach Extension Theorem). Assume that V is a real
vector space and that p W V ! R is a function such that, for all v; v1 ; v2 2 V and
all 0,
1. p.v1 C v2 / p.v1 / C p.v2 /, and
2. p.v/ D p.v/.
If L is a linear submanifold of V and if ' W L ! R is a linear transformation for which
j'.v/j p.v/ for every v 2 L, then there is a linear transformation W V ! R such
that jL D ' and p.v/ .v/ p.v/ for every v 2 V.
Proof. The linearity of ' implies that p.v/ '.v/ p.v/ for every v 2 L. If
' D 0, then we may take to be D 0; therefore, assume that ' 6D 0.
Define a set S consisting of all ordered pairs .M; #/ such that M is a linear
submanifold of V containing L and # W M ! R is a linear transformation satisfying
p.v/ #.v/ p.v/ for every v 2 M. The set S is nonempty since .L; '/ 2 S.
Define a partial order on S by
.M1 ; #1 / .M2 ; #2 /
if and only if M1 M2 and #2jM1 D #1 :
With respect to this partial order, let F be any linearly ordered subset of S. Hence,
there is a linearly ordered set such that
F D f.M ; # / j 2 g :
Define M V by
MD
M ;
2
and note that M is a linear submanifold of V containing L. Furthermore, the function

# W M ! R given by #.v/ D # .v/, if v 2 M , is well defined, linear, and satisfies
p.v/ #.v/ p.v/ for all v 2 M. Thus, .M; #/ 2 S and .M; #/ is an upper
bound for F. Hence, by Zorns Lemma, S has a maximal element.
224
6 Dual Spaces
Let .M; #/ denote one such maximal element of S. Suppose M 6D V; thus, there
exists (nonzero) v0 2 V n M. If x; y 2 M, then
#.x/ C #.y/ D #.x C y/ p.x C y/
D p ..x v0 / C .v0 C y//
p.x v0 / C p.x0 C y/:
Thus,
#.x/ p.x v0 / p.v0 C y/ #.y/:
Let 0 2 R satisfy
sup .#.x/ p.x v0 // 0 inf .p.v0 C y/ #.y// :
y2M
x2M
Thus,
#.x/ 0 p.x v0 /
and
#.y/ C 0 p.v0 C y/
for all x; y 2 M.
Consider now the linear submanifold M1 D fx C v0 j x 2 M; 2 Rg of V, and
note that M1 contains M and M 6D M1 . Define a function #1 W M1 ! R by
#1 .x C v0 / D #.x/ C 0 ;
and observe that #1 is linear. If > 0, then

#1 .x C v0 / D #. 1 x/ C 0 p. 1 x C v0 / D p.x C v0 /;
while if < 0, then

#1 .x C v0 / D jj #. 1 x/ 0 jjp. 1 x v0 / D p.x C v0 /:
Therefore, by the linearity of #1 , p.v/ #1 .v/ p.v/ for every v 2 V,
which shows that .M1 ; #1 / 2 S. But the relation .M; #/ .M1 ; #1 / with M1 6D M
contradicts the maximality of .M; #/ in S. Therefore, it must be that M D V and so
D # is one desired extension of '.
t
u
The function p in the statement of the Hahn-Banach Extension Theorem is called
a sublinear functional.
Definition 6.19. A function p W V ! R for which p.v1 C v2 / p.v1 / C p.v2 / and
p.v/ D p.v/, for all v; v1 ; v2 2 V and all 0, is called a sublinear functional.
225
A more specialised version of the Hahn-Banach Extension Theorem, in which the

sublinear functional p is given by p.v/ D kvk for some norm k k, is very frequently
used. Before stating this form of the result, we first indicate how to pass from a
complex normed vector space to a real normed vector space and then back again.
Thus, suppose, as usual, that V is a complex vector space and let ' W V ! C be
an arbitrary linear transformation. The functions 1 ; 2 W V ! R defined by
1 .v/
1
'.v/ C '.v/
2
and
2 .v/
1
'.v/ '.v/
2i
satisfy j .v1 C v2 / D j .v1 / C j .v2 / and j .rv/ D r j .v/ for v; v1 ; v2 2 V, r 2 R,

and j D 1; 2. Thus, considering V as a real vector space, 1 and 2 are R-linear
transformations such that '.v/ D 1 .v/ C i 2 .v/ for every v 2 V. Therefore, we
denote 1 and 2 by <' and =', respectively.
Now if V is a normed vector space and ' 2 V , then for every unit vector v 2 V
we have
j<' .v/j
1
j'.v/j C j'.v/j D j'.v/j;
2
which shows that <' is a bounded linear transformation and that k<'k k'k.
On the other hand, given a unit vector v 2 V, there is a real number such that
j'.v/j D ei '.v/. Thus, with w D ei v 2 V, we have

j<' .w/j D < '.ei v/ D < ei '.v/ D j'.v/j;
which shows that
k'k D
sup
v2V; kvkD1
j'.v/j
sup
j'.w/j D k<'k:
w2V; kwkD1
Hence, k<'k D k'k. A similar argument shows that k='k D k'k.

The discussion above is summarised by the following lemma.
Lemma 6.20. If V is a normed vector space and if ' 2 V , then <' and =' are
R-linear maps V ! R such that k<'k D k='k D k'k.
Conversely, suppose now that W V ! R is a (real) linear transformation, where
V is considered as a vector space over R. Motivated by the fact that D <./
i<.i/ for every 2 C, define 1 W V ! R by 1 .v/ D .iv/, for v 2 V, and
' W V ! C by ' D C i 1 . Observe that 1 is a R-linear transformation, ' is Clinear transformation, and <' D and =' D 1 . If, in addition, V is a normed
vector space and is bounded, then the discussion preceding Lemma 6.20 shows
that ' is bounded and k'k D k k. Thus:
Lemma 6.21. If V is a normed vector space and if W V ! R is R-linear and
bounded, then the function ' W V ! C defined by '.v/ D .v/ i .iv/, for v 2 V,
is C-linear, bounded, and satisfies k'k D k k.
226
6 Dual Spaces
The specialised form of the Hahn-Banach Extension Theorem is now ready to be

formulated and proved.
Theorem 6.22 (Hahn-Banach Extension Theorem for Normed Spaces). If L is
a linear submanifold of a normed vector space V, and if ' 2 L , then there exists
2 V such that jL D ' and kk D k'k.
Proof. Write ' D <' C i='. Because <' and =' are R-linear functionals on L
of norm k<'k D k='k D k'k (by Lemma 6.20), it is sufficient to find R-linear
extensions 1 and 2 of <' and =', respectively, such that kj k D k'k for j D 1; 2,
and to then define by D 1 C i2 , using Lemma 6.21.
Therefore, we assume without loss of generality that
W L ! R is R-linear,
nonzero, and bounded of norm k k. Thus, Q D k k1 is R-linear of norm
k Q k D 1.
Let p W V ! R be given by p.v/ D kvk, for v 2 V. Thus, for every v; v1 ; v2 2 V
and 0, we have p.v1 C v2 / p.v1 / C p.v2 /, p.v/ D p.v/, and j Q .v/j p.v/,
for every v 2 L. Hence, by Theorem 6.18, there is an R-linear map Q W V ! R
extending Q and such that p.v/ Q .v/ p.v/ for every v 2 V; that is, such
that kvk Q .v/ kvk for all v 2 V. Hence, kQ k 1 D k Q k kQ k implies that
k k D k k, where D k kQ is the desired R-linear extension of from L to V.
t
u
The Hahn-Banach Extension Theorem leads to the following result which shows
that the dual space V of an infinite-dimensional normed vector space V contains
(many) elements other than 0.
Corollary 6.23. If v is a nonzero element of a normed vector space V, then there is
a ' 2 V such that k'k D 1 and '.v/ D kvk.
Proof. On the 1-dimensional subspace L D fv j 2 Cg, let '0 W L ! C be given by
'0 .v/ D kvk. Then '0 is a linear transformation and '0 .v/ D kvk. The norm of
'0 is 1, since j'0 .v/j D jj kvk D kvk. By the Hahn-Banach Theorem, there is
an extension of '0 to a linear functional ' 2 V such that k'k D k'0 k D 1.
t
u
6.4 The Second Dual

By Corollary 6.23, for every v in a normed vector space V, there is a ' 2 V such
that k'k D 1 and '.v/ D kvk. Therefore, a normed vector space admits a rich family
of linear functionals. Treating the dual V as a normed vector space itself, then its
dual .V / is likewise large. This second dual of V is important in many regards,
not the least of which is because the second dual of V contains V in a natural way,
which leads to a fruitful conceptual perspective in which the elements of V act on
the elements of V (rather than vice versa, according to the definition of V ).
Definition 6.24. The second dual of a normed vector space V is the dual space
.V / of V , and is denoted by V .
6.5 Weak Topologies
227
The following proposition shows that V contains an isometric copy of V.

Proposition 6.25. If V is normed vector space, then there exists a linear isometry
W V ! V such that v.'/ D '.v/ for every v 2 V and ' 2 V .
Proof. For each v 2 V, let !v W V ! C be the linear transformation defined by
!v .'/ D '.v/, for ' 2 V . Because j'.v/j k'k kvk, for every ' 2 V , the linear
map !v is bounded. Hence, !v 2 V , for every v 2 V.
It is straightforward to verify that the map W V ! V that sends each v 2 V to
the function !v is linear (i.e., !v1 C !v2 D !v1 Cv2 and !v D ! v, for all v1 ; v2 ; v 2
V and 2 C). Thus, is a linear transformation. Moreover, for each v 2 V,
kvk D
sup
'2V ; k'kD1
j!v .'/j D
sup
'2V ; k'kD1
j'.v/j D kvk :
Thus, is a linear isometry.
t
u
Definition 6.26. A Banach space V is said to be reflexive if the operator in

Proposition 6.25 is a surjection.
The most immediate example of a reflexive Banach space is afforded by finitedimensional spaces.
Proposition 6.27. Every finite-dimensional Banach space is reflexive.
Proof. To prove this, suppose that B D fv1 ; : : : ; vn g is a basis of V,P
then each v 2 V
has unique representation as linear combination of v1 , . . . , vn : v D j j vj . For each
P
D k . Clearly 'k 2 V and 'k .vj / D 0, if
v
k, let 'k W V ! C be defined by 'k
j
j
j
j 6D k and 'k .vk / D 1. To show that '1 ; : : : ; 'n are linearly
independent,
suppose that
P
P
P
v
'
D
0.
Then,
for
any
k
2
f1;
2;
:
:
:
;
ng,
0
D
'
D
k
j j j
j j j
j j 'j .vk / D k .
Hence, '1 ; : : : ; 'n are linearly independent.
Using the basis f'1 ; : : : ; 'n g of V , repeat the argument above to produce a basis
f1 ; : : : ; n g of V that has the property that k .'k / D 1 and k .'j / D 0 if j 6D k.
Define a function W V ! V on the basis of V by .vk / D k and extend this by
linearity to all of V. The operator plainly satisfies v.'/ D '.v/ for every v 2 V
and ' 2 V .
t
u
Proposition 6.16 provides another set of examples of reflexive spaces:
Example 6.28. If p > 1, then `p .N/ is a reflexive Banach space.
6.5 Weak Topologies

Suppose that X and Y are topological spaces and that F is a family of functions
f W X ! Y. Recall that Proposition 1.88 shows that the collection

B D f11 .U1 / \ \ fn1 .Un / j n 2 N; Uj Y is an open set; fj 2 F
228
6 Dual Spaces
is a basis for a topology (called the weak topology induced by F ) on X with respect
to which each function f 2 F is continuous. In this section we shall consider two
particular choices of F when X and Y are certain normed vector spaces.
Definition 6.29. If V is a normed vector space, then the weak topology on V is the
the weak topology on V induced by the family F D V , and the elements of this
topology are called weakly open sets.
Note the use of the term weak topology in the setting of Banach space differs
slightly from the use of the term in topology in that reference to the choice of family
F is dropped. That is, when saying that V has the weak topology it is understood
implicitly that the family of functions inducing the topology is the family V of all
bounded linear functionals on V.
Suppose now that a normed vector space has the weak topology. If v0 2 V and
U V is a weakly open set, then there is a basic weakly open set B such that
v0 2 B U. That is, there are '1 ; : : : ; 'n 2 V and open sets W1 ; : : : ; Wn C such
that
v0 2
n
\
'j1 .Wj / U:
jD1
As '.v0 / 2 Wj C for each j, there are positive real numbers "1 ; : : : ; "n such that,
for each j,
f 2 C j j 'j .v0 /j < "j g Wj :
Hence,

v0 2 v 2 V j j'j .v/ 'j .v0 /j < "j ; 8 j D 1; : : : ; n U:
Proposition 6.30. If V is a finite-dimension normed vector space, then the weak
topology and the norm topology on V coincide.
t
u
In contrast to Proposition 6.30, the weak topology and the norm topology are
strikingly different in the case of infinite-dimensional spaces. For example, in the
norm topology of an infinite-dimensional Banach space V there are bounded open
sets U V that contain 0 2 V (the open unit ball, for example); however, this is not
at all true for the weak topology.
Proposition 6.31. If V is an infinite-dimensional Banach space, and if U V is a
weakly open set such that 0 2 U, then U is unbounded. In fact, there is an infinitedimensional subspace L V such that L U.
6.5 Weak Topologies
229
Proof. Choose a basic weakly open set B such that 0 2 B U. Thus, there are
n
\
'j1 .Wj /. Let L D
'1 ; : : : ; 'n 2 V and open sets W1 ; : : : ; Wn C such that 0 2
n
\
jD1
ker 'j , which is a subspace of V contained in U. We need only verify that L has
jD1
infinite dimension. To this end, let W V ! Cn be the linear transformation

3
'1 .v/
7
6
.v/ D 4 ::: 5 ;
2
v 2 V:
'n .v/
Note that ker D L. The First Isomorphism Theorem in linear algebra asserts that
the quotient space V=L is isomorphic to the range of , which is a subspace of Cn .
Because the quotient of an infinite-dimensional vector space by a finite-dimensional
subspace cannot have finite dimension, it must be that L has infinite dimension. u
t
When it comes to the dual space V of V, it is of less interest to endow V with
the weak topology induced by the family V than it is to endow V with the weak
topology induced by the subfamily .V/ V indicated in Proposition 6.25
namely, those functions f W V ! C for which there exist a v 2 V such that f .'/ D
'.v/ for every ' 2 V .
Definition 6.32. If V is a normed vector space, then the weak topology on V
is the weak topology on V induced by the subfamily .V/ V indicated in
Proposition 6.25.
To be clear, if '0 2 V , then a basic weak open subset B V that contains '0
has the form
B D f' 2 V j j'.vj / '0 .vj /j < "j ; for all j D 1; : : : ; ng;
for some n 2 N, v1 ; : : : ; vn 2 V, and positive real numbers "1 ; : : : "n .
The most important fundamental property of the weak topology is established
by the following theorem.
Theorem 6.33 (Banach-Alaoglu). If V is a normed vector space and if X V is
the closed unit ball of V , then X is compact in the weak topology of V .
Y
Proof. For each v 2 V, let Kv D f 2 C j j j kvkg. Consider the space K D
Kv ,
v2V
endowed with the product topology. By Tychonoffs Theorem (Theorem 2.14), K is

a compact set. Furthermore, K is Hausdorff because each Kv is Hausdorff.
Define f W X ! K by f .'/ D .'.v//v2V , and note that f is an injective function.
Select ' 2 X and consider an open set W K that contains f .'/. Thus, there are
open subsets Wv Kv such that '.v/ 2 Wv , for every v 2 V, and Wv D Kv for all
230
6 Dual Spaces
but at most a finite number of vectors in V, say v1 ,. . . , vn , and W D
Wv . Because
v2V
each .vj / is continuous on V , the set .vj /1 .Wvj / is open in V , which implies
n
\
that Uj D f 2 X j .vj / 2 Wvj g is open in X for each j. Thus, if U D
Uj , then
jD1
U is an open set containing ' for which f . / 2 W for every 2 U. Hence, f is

continuous at ', and so f is a continuous function on X.
On the other hand, if U X is an arbitrary open set and if ' 2 U, then there is
a basic open set B such that ' 2 B U. By definition, there are v1 ; : : : ; vn 2 V and
"1 ; : : : ; "n > 0 such that, for 2 X, we have 2 B if and only if j .vj / '.vj /j < "j
for each j D 1; : : : ; n. Hence, if Wvj D f 2 Kvj j j '.vj /j < "j g and if Wv D Kv for
Y
Wv is open in K and f .'/ 2 W' f .U/.
every v 2 V n fv1 ; : : : ; vn g, then W' D
Thus, f .U/ D
v2V
W' , which shows that f .U/ is open. Hence, f 1 is continuous, and
'2U
therefore f is a homeomorphism between X and f .X/.

We now show that f .X/ is a closed subset of K. Let D . v /v2V 2 K be in the
closure of f .X/. Suppose that v1 ; v2 2 V and 1 ; 2 2 C, and let " > 0. Define
UDf
2 X j j .v/ v j < "; v 2 fv1 ; v2 ; 1 v1 C 2 v2 gg ;
which is an open subset of X. Thus, W D f .U/ is open in K (because f is a

homeomorphism). The open set W contains , and is in the closure of f .X/. Hence,
there exists ' 2 X such that f .'/ 2 W, and so
j'.v1 / v1 j < " ;
j'.v2 / v2 j < ";
and
j'.1 v1 C 2 v2 / 1 v1 C2 v2 j < ":
Therefore,
j 1 v1 C2 v2 1 v1 2 v2 j < .1 C j1 j C j2 j/ ":
As " > 0 is arbitrary, we deduce that 1 v1 C2 v2 D 1 v1 C 2 v2 . Therefore, the
map v 7! v is linear and satisfies j v j kvk for all v 2 V, implying that this map
is an element of X and that 2 f .X/. This proves that f .X/ is closed in K.
Because K is compact, Hausdorff, and f .X/ is closed in K, we deduce that f .X/
is compact and Hausdorff; hence, X is compact and Hausdorff.
t
u
The following striking theorem shows that all Banach spaces arise as subspaces
of C.X/ for various choices of compact Hausdorff spaces X.
Proposition 6.34. For every Banach space V there is a compact Hausdorff space
X such that V is isometrically isomorphic to a subspace of C.X/.
6.5 Weak Topologies
231
Proof. Let X be the closed unit ball of V , which is compact and Hausdorff when
endowed with the weak topology (Theorem 6.33). By Proposition 6.25, there is
an isometric embedding W V ! V whereby v.'/ D '.v/, for all ' 2 X and
v 2 V. Thus, we need only show that v 2 C.X/ for every v 2 V. To this end, select
v 2 V and consider the function v W X ! C. To show that v is continuous at
' 2 X, assume that W is an open set containing v.'/ D '.v/. Thus, there exists
" > 0 such that W0 D f 2 C j j '.v/j < "g W. If U D v 1 .W0 / D f 2
X j j .v/ '.v/j < "g, which is a basic open set in X, then v.U/ W, which
proves that v is continuous at '. Hence, v is continuous on X.
t
u
If V is a separable Banach space, then one would hope that the topological
space X that arises in Proposition 6.34 above is a compact metric space, for then
the enveloping Banach space C.X/ that contains V as a subspace would also be
separable (Theorem 5.57). This is indeed the case by the following result.
Proposition 6.35. If V is a separable Banach space and X is the closed unit ball of
V , then X is metrisable.
Proof. By hypothesis, there is a countable set that is dense in V; hence, there is a
countable set fvn gn2N X that is dense in the closed unit ball of V.
The compact set Y
D D fz 2 C j jzj 1g is a subset of the metric space C, and so the
product space D D
D of countably many copies of D is compact and metrisable
n2N
in the product topology (by Tychonoffs Theorem and Proposition 1.57).

Let f W X ! D be given by f .'/ D .'.vn //n2N . To show that f is continuous at
each point of
QX, select '0 2 X and let W D be an open set that contains f .'0 /.
Thus, W D n2N Wn for some open sets Wn D for which Wn D D for all n with
the exception of at most finitely many n1 ; : : : ; nm 2 N. Thus, there are positive real
numbers "1 ; : : : ; "m such that f 2 D j j '0 .vnj /j < "j g, for j D 1; : : : ; m. The set
U D f' 2 X j j'.vnj / '0 .vnj /j < "j ; j D 1; : : : ; ng is an open subset of X that contains
'0 and satisfies f .U/ W. Hence, f is continuous at '0 , which proves that f is a
continuous map X ! D.
The continuous image of a compact set is compact (Proposition 2.9), which
implies that f .X/ is compact. Furthermore, because D is metrisable, so is f .X/.
Therefore, in particular, f .X/ is Hausdorff. Because fvn gn2N is dense in the closed
unit ball of V and because each ' 2 X is continuous, the map f is injective. Hence,
f is a bijective continuous map from a compact space X onto a Hausdorf space
f .X/. By Proposition 2.9, f is necessarily a homeomorphism, which implies that X
is metrisable.
t
u
Corollary 6.36. For every separable Banach space V there exists a compact metric
space X and a subspace L C.X/ such that V and L are isometrically isomorphic.
Proof. Let X be the topological space given by the closed unit ball of V in the
weak -topology. By Proposition 6.35, X is metrisable; and, by Proposition 6.34 V
is isometrically isomorphic to a subspace L of C.X/.
t
u
232
6 Dual Spaces
6.6 Linear Functionals on Lp and L1

For every p 1, the elements of Lp .X; ; / are determined by p-integrable
functions f W X ! C. We shall write f 0 for f 2 L p .X; ; / if f .x/ 0 for every
x 2 X, and write fP 0 if there is a function g 2 L p .X; ; / with g 0 and gP D fP .
The notation gP fP is used to denote fP gP 0.
Definition 6.37. A linear functional ' W Lp .X; ; / ! C is said to be positive if
'.fP / 0 for every fP 2 Lp .X; ; / with fP 0.
Lemma 6.38. Every linear functional on Lp .X; ; /, where p 2 R satisfies p 1,
is a linear combination of four positive linear functionals.
Proof. By Lemmas 6.20 and 6.21, every ' 2 Lp .X; ; / is a linear combination
' D <' C i=' of two continuous R-linear maps <'; =' W Lp .X; ; / ! R.
Thus, consider Lp .X; ; / as a Banach space over R and let D <', which is
continuous and R-linear. Define a real-valued function C on the positive elements
of Lp .X; ; / by
P D supf .Pg/ j g 2 L p .X; ; /; 0 g f g:
C .f /
To confirm that the supremum above exists, note that 0 g f and p 1 imply
that jgjp jf jp , and so kPgk kfP k; thus, j .Pg/j k k kPgk k k kfP k and, therefore,
P
P
P
C .f / exists and is such that C .f / k k kf k.
p
Suppose now that g1 ; g2 2 L .X; ; / satisfy 0 gj f . By the linearity of
, we have .Pg1 / C .Pg2 / D .Pg1 C gP 2 / C .fP1 C fP2 /; hence, C .fP1 / C C .fP2 /
p
P
P
C .f1 C f2 /. On the other hand, if h 2 L .X; ; / satisfies 0 h f1 C f2 , then
let g1 .x/ D maxfh.x/ f2 .x/; 0g and g2 .x/ D minfh.x/; f2 .x/g to obtain p-integrable
P D
functions g1 and g2 with 0 gj fj , for j D 1; 2, and g1 C g2 D h. Thus, .h/
P
P
P
P
P
P
.Pg1 / C .Pg2 / C .f1 / C C .f2 / yields C .f1 C f2 / C .f1 / C C .f2 /. Hence,
p
C is an additive function on the positive elements of L .X; ; /. The function
P
P
C is also plainly positive-homogeneous in the sense that C . f / D C .f / for
p
P
every 0 in R and f 0 in L .X; ; /.
Let R D ffP j f 2 L p .X; ; / such that f .x/ 2 R for all x 2 Xg, which is a real
subspace of Lp .X; ; /. Express each real-valued function f 2 L p .X; ; / as the
difference f D f C f of positive p-integrable functions f C and f as prescribed in
equation (3.3), and define C on R by
P D
C .f /
PC/
C .f
P /:
C .f
In expressing each 2 R as a difference D C of positives, we see that

p
P
P
C . f / D C .f / for every real-valued f 2 L .X; ; /. As C is plainly additive
on R, we deduce that C is a continuous R-linear map on R. Likewise, by defining
on R, the map is R-linear, continuous, and satisfies .fP / 0
D C
P
for every f 0.
233
Because every f 2 L p .X; ; / has the form f D <f C i=f , the maps C and
p
P
P
P
extend from R to all of L .X; ; / via
C .f / D C .<f / C i C .=f / and
P
P
P
.f / D .<f / C i .=f /, in each case yielding a continuous, positive, C-linear
map Lp .X; ; / ! C with the property that
D C . Hence, <' is the
difference of two positive linear functionals. A similar argument shows that ='
is a difference of two positive linear functionals, which implies that ' is a linear
combination of four positive linear functionals.
t
u
The main result that we aim to prove in this section is Theorem 6.40, which is
based in part on the density of simple functions in Lp . A proof technique related to
the density of simple functions is encompassed by the following lemma.
Lemma 6.39. Assume that p 1 and that ' is a positive linear functional on
Lp .X; ; /. Suppose that g W X ! R is a nonnegative measurable function and that
f ; fk 2 L p .X; ; /, for k 2 N, have the following properties:
1. 0 fk .x/ fkC1 .x/ f .x/, for all x 2 X and k 2 N;
2. lim fk .x/ D f .x/, for all x 2 X;
k!1
Z
P
3. '.fk / D fk g d for every k 2 N.
X
Then '.fP / D
Z
fg d.
X
Proof. By Proposition 5.42, the first two of the three conditions above imply that
lim kfP fPk k D 0. Because 0 fk .x/g.x/ f .x/g.x/ and limk fk .x/g.x/ D f .x/g.x/
k!1
for every x 2 X, the Dominated Convergence Theorem yields

Z
lim
k!1 X
jfg fk gj d D 0:
Hence,
Z
Z
'.fP / fg d D '.fP / '.fPk / C '.fPk / fg d
kfP fPk k k'k C .fk f /g d

X
kfP fPk k k'k C
Z
jfk g fgj d:
X
Thus, by letting k ! 1, we obtain '.fP / D
Z
fg d.
t
u
The following result is our second instance of a Riesz Representation Theorem.
234
6 Dual Spaces
Theorem 6.40 (Riesz). Suppose that p and q are conjugate real numbers. If
.X; ; / is a -finite measure space, and if W Lq .X; ; / ! Lp .X; ; / is
defined by
.Pg/fP D
Z
fg d;
(6.3)
for all f 2 L p .X; ; / and g 2 L q .X; ; /, then is a linear isometric

isomorphism.
Proof. Let g 2 L q .X; ; / be fixed. By Hlders inequality (Proposition 4.53),
Z
Z
jfgj d
jf j d
1=p Z
jgj d
1=q
8 f 2 L p .X; ; / :
R
Hence, the function 'g W Lp .X; ; / ! C defined by 'g .fP / D X fg d is a linear
functional on Lp .X; ; / for which k'g .fP /k kfP k kPgk for all fP 2 Lp .X; ; /.
Therefore, the function indeed takes values in the dual of Lp .X; ; / and is
given unambiguously by .Pg/ D 'g . The map is plainly linear, and the inequality
k.Pg/k D k'g k kPgk implies that is continuous.
To show that is isometric, we need to show that kPgk k'g k. Let g W X ! C
be the function g D sgn gnamely, the measurable function whose value at x 2 X
g.x/
is 0, if g.x/ D 0, and is jg.x/j
, if g.x/ 6D 0. Thus, jg.x/j D g .x/g.x/ for all x 2 X.
Let f W X ! C be defined by f D jgjq=p g . Observe that f is p-integrable and that
q
fg D jgjq=p jgj D jgj1C p D jgjq . Thus,
Z
kPgk D
fg d D j'g .fP /j k'g k kfP k
jgj d D
q
Z
D k'g k
jf jp d
1=p
Z
D k'g k
jgjq d
1=p
D k'g k kPgkq=p :
Hence, because q qp D 1, we deduce that kPgk k'g k, which proves that is a

linear isometry.
What remains, therefore, is to prove is that is surjective. By Lemma 6.38, it is
sufficient to show that every positive linear functional on Lp .X; ; / is in the range
of . To this end, suppose that ' is a nonzero positive linear functional.
Assume, in the first instance, that .X/ < 1. Define a function W ! R by
.E/ D '.P E /, and note that .E/ 0 (because P E 0). Suppose that fEk gk2N is
n
[
[
Ek . If Gn D
Ek ,
a sequence of pairwise disjoint measurable sets and let E D
then P Gn D
n
X
kD1
P Ek and, thus, .Gn / D
n
X
kD1
and Fn are disjoint with union E, and so
k2N
kD1
.Ek /. Therefore, if Fn D E n Gn , then Gn
.Fn / D '.P E /
235
n
X
!
'.P Ek / D .E/
n
X
kD1
!
.Ek / :
kD1
The sequence fGn gn2N is monotone increasing to E , and therefore the Monotone
Convergence Theorem yields lim kP Gn P E k D 0. Hence, by the continuity of ',
n!1
.E/ D lim
n!1
n
X
.Ek / D
kD1
1
X
.Ek /:
kD1
This proves that is countably additive. That is, is a measure on .X; /.

Now suppose that E 2 satisfies .E / D 0. Thus, kP E k D 0 and so .E/ D
'.P E / D 0. Therefore, is absolutely continuous with respect to . By the RadonNikodm Theorem (Theorem 4.36), there exists g 2 L 1 .X; ; / such that g.x/ 0,
for every x 2 X, and
Z
.E/ D g d ; 8 E 2 :
E
Hence, if h is a simple function in canonical form (see equation (4.1)), then

Z
P D hg d:
'.h/
X
Suppose that f 2 L .X; ; / satisfies f .x/ 0 for every x 2 X. Proposition 5.41

shows that there is a monotone-increasing sequence fhk gk2N of simple functions
hk such that 0 hk .x/ f .x/ forRall x 2 X and limk hk .x/ D f .x/ for each x 2 X.
Therefore, the equations '.hP kZ/ D X hk g d for every k 2 N and Lemma 6.39 yield
p
the desired formula '.fP / D
fg d. Because every element of Lp .X; ; / is a

Z
P
linear combination of positive elements, the formula '.f / D fg d holds for every
X
f 2 L p .X; ; /.
To show that g is q-integrable, let Ek D fx 2 X j g.x/ kg for each k 2 N and define
fk D Ek g gq1 . Because g .x/ 2 f0; 1g for all x 2 X, the function fk is bounded.
P
RFurthermore, .Ek / .X/ < 1q implies that fk is p-integrable, and so '.fk / D
X fk g d. Therefore, using fk g D g Ek , we have
Z
Z
gq d D fk g d D '.fPk / k'k kfPk k
Ek
Z
D k'k
jfk jp d
1=p
Z
D k'k
1=p
gq d
This proves that

Z
q
g d
Ek
.1 1p /
Z
D
1=q
q
g d
Ek
k'k:
236
6 Dual Spaces
Z
q
Because
X
gk d D
gq d, where gk D Ek g, and because g.x/q D limk gk .x/q

Ek
for every x 2 X, Fatous Lemma (Corollary 4.9) yields the first of the inequalities
below:
Z
Z
q
q
g d lim inf gk d k'kq :
k
Hence, g 2 L q .X; ; /, which thereby completes the proof (under the assumption
that .X/ < 1) that is surjective.
Assume now that .X/ D 1 and that ' 2 Lp .X; ; / is a positive linear
functional. Because .X; ; / is -finite, there is an increasing
sequence fEk gk2N
[
of measurable sets Ek of finite measure such that X D
Ek . If, for each k 2 N,
k2N
k D fEk \ A j A 2 g and k D jk , then the linear map f 7! fQ whereby a function

f W Ek ! C is sent to a function fQ W X ! C in which fQ .x/ D f .x/, for x 2 Ek , and fQ .x/ D
0, for x 2 Ekc , induces a natural linear isometry Tk;p W Lp .Ek ; k ; k / ! Lp .X; ; /.
Therefore, we view Lp .Ek ; k ; k / as a subspace of Lp .X; ; / consisting of
equivalence classes of p-integrable functions on X that vanish on Ekc . This is also
true for the conjugate q, via Tk;q W Lq .Ek ; k ; k / ! Lq .X; ; /.
Setting 'k D 'jLp .Ek ;k ;k / yields an element of the dual space of Lp .Ek ; k ; k /.
Note that if m > n, then 'mjLp .En ;n ;n / D 'n . Furthermore, because .Ek / < 1,
the linear isometry k W Lq .Ek ; k ; k / ! Lp .Ek ; k ; k / given by equation (6.3)
yields an element k1 .'k / of Lq .Ek ; k ; k /. For each k 2 N select a representative
gk 2 L q .Ek ; k ; k / such that gP k D k1 .'k /. Hence,
'k .fPk / D
fQk gQ k d;
fk gk dk D
Ek
for every fk 2 L p .Ek ; k ; k /. Because any two representatives in L q .Ek ; k ; k /

for a single equivalence class in Lq .Ek ; k ; k / will differ only on a set of measure
zero, we see that for m > n the set fx 2 En j gm .x/ 6D gn .x/g has measure zero. Let
FD
1 [
1
[
fx 2 En j gm .x/ 6D gn .x/g;
nD1 mDn
which is a measurable set with .F/ D 0. Define g W X ! C by g.x/ D 0 if x 2 F,

and for x 2 F c by g.x/ D gk .x/ for any k 2 N that satisfies x 2 Ek . Therefore g is a
measurable function and g.x/ D limk gQ k .x/ for all x 2 X n F. Because
Z
q
Ek
gk dk D k'k kq k'kq
237
for every k, Fatous Lemma again yields

Z
Z
Z
q
q
gq d lim inf gQ k d D
gk dk k'kq ;
k
Ek
which proves that g is q-integrable.

To complete the proof, select any nonnegative f 2 L p .X; ; / and, for each
k 2 N, let fk D f Ek . Thus, fPk is an element of the subspace Lp .Ek ; k ; k / and
Z
Z
Z
P
P
'.fk / D 'k .fk / D
fk gk d D
fk g d D fk g d:
Ek \F c
Ek
Furthermore, ffk gk2N is a monotone increasing sequence of nonnegative p-integrable

functions in which fk .x/ f .x/
Z and limk fk .x/ D f .x/ for all x 2 X. Hence, by
Lemma 6.39, we have '.fP / D fg d. Hence, because the positive elements span
XZ
Lp .X; ; /, the formula '.fP / D fg d holds for every f 2 L p .X; ; /.
t
u
X
Corollary 6.41. If positive p and q are conjugate real numbers, and if .X; ; /
is a -finite measure space, then for every ' 2 LZp .X; ; / there exists a unique
gP 2 Lq .X; ; / such that kPgk D k'k and '.fP / D fg d, for every fP 2 Lp .X; ; /.
X
The hypothesis that .X; ; / be a -finite measure space, in Theorem 6.40

above, can be removed; doing so, however, is rather subtle. The monograph of Bartle
[6] details how such an extension of Theorem 6.40 is achieved.
The case of L1 .X; ; / is similar to that of Lp , but there are some key
differences, which we make note of below.
Definition 6.42. An element fP 2 L1 .X; ; / is said to be positive if gP D fP for some
nonnegative function g 2 L 1 .X; ; /.
Note that if f 2 L 1 .X; ; / is such that ess-ran f 0; 1/, then fP is a positive
element of L1 .X; ; /. As with Lp -spaces, we shall write write fP 0 for positive
elements, and use the notation gP fP to denote fP gP 0.
Definition 6.43. A linear functional ' W L1 .X; ; / ! C is said to be positive if
'.fP / 0 for every fP 2 L1 .X; ; / with fP 0.
Lemma 6.44. Every linear functional on L1 .X; ; / is a linear combination of
four positive linear functionals.
t
u
Theorem 6.45 (Riesz). If .X; ; / is a -finite measure space, then the function
W L1 .X; ; / ! L1 .X; ; / defined by
.Pg/fP D
Z
fg d;
X
(6.4)
238
6 Dual Spaces
for all f 2 L 1 .X; ; / and g 2 L 1 .X; ; /, is a linear isometric isomorphism.

Proof. Let g 2 L 1 .X; ; / be fixed. Because
Z
Z
jfgj d
X

.ess-sup jgj/jf j d D kgk kf k;
R
the function 'g W L1 .X; ; / ! C defined by 'g .fP / D X fg d is a linear functional
on L1 .X; ; / and k'g .fP /k kfP k kPgk for all fP 2 L1 .X; ; /. Therefore, the
function takes values in the dual of L1 .X; ; / and is given unambiguously by
.Pg/ D 'g . The map is linear, and the inequality k.Pg/k D k'g k kPgk implies
that is bounded.
Assume, to begin with, that .X/ < 1. To show that is isometric, let " > 0
and A" D fx
En D fx 2 X j jg.x/ ng.
[2 X j k'k C " < jg.x/jg. For each n 2 N, let
g
Thus, A D
En \ A for every A 2 . Set fn D En \A" jgj and note that
n2N
Z
jfn j d D
X
En \A"

g
d D .En \ A" / .X/ < 1:
jgj
Therefore, fn 2 L 1 .X; ; / and

Z
Z
P
k'k.En \ A" / D k'k kfn k jfn gj d D
X
En \A"
jgj d .k'k C "/.En \ A" /:
By continuity of measure,
k'k.A" / .k'k C "/k.A" /:
Hence, .A" / D 0, which implies that ess-sup jgj k'k C ". However, as the choice
of " > 0 is arbitrary, ess-sup jgj k'k, whence kPgk k'g k. That is, is an
isometry.
To prove is that is surjective, suppose that ' 2 L1 .X; ; / is a nonzero
positive linear functional. Define a function W ! R by .E/ D '.P E /. As shown
in the proof of Theorem 6.40, is a measure on .X; ; /, absolutely continuous
with respect to . Therefore, by the Radon-Nikodm Theorem (Theorem 4.36),
there
exists measurable function g such
R
R that g.x/ 0, for every x 2 X, and .E/ D
P D hg d for every simple function h.
g
d,
for
all
E
2
.
Hence,
'.
h/
E
X
Suppose that f 2 L 1 .X; ; / satisfies f .x/ 0 for every x 2 X. Using Proposition 5.41 we find a monotone-increasing sequence fhk gk2N of simple functions
hk such that 0 hk .x/ f .x/ forRall x 2 X and limk hk .x/ D f .x/ for each x 2 X.
Therefore, the equations '.hP kZ/ D X hk g d for every k 2 N and Lemma 6.39 yield
the desired formula '.fP / D
fg d. Because every element of L1 .X; ; / is a

Z
linear combination of positive elements, the formula '.fP / D fg d holds for every
X
f 2 L 1 .X; ; /. The proof above, where it is shown that is isometric, also shows
6.7 Linear Functionals on C.X/
239
that ess-sup jgj k'k, and therefore g 2 L 1 .X; ; /. Hence, is a surjective,
thereby completing the proof of the theorem in the case where .X/ is finite.
The proof of the remainder of the theorem is similar to the corresponding part of
the proof of Theorem 6.40 and is, therefore, left as an exercise (Exercise 6.69). u
t
1

Corollary 6.46. If .X; ; / is -finite, then
Z for each ' 2 L .X; ; / there is a
unique gP 2 L1 .X; ; / such that '.fP / D fg d, for every fP 2 L1 .X; ; /, and
kPgk D k'k.

As with Lp and L1 spaces, the notion of positive elements has a key role in
determining features of linear functionals on C.X/, where X is a compact Hausdorff
space.
Definition 6.47. If X is a compact Hausdorff space, an element f 2 C.X/ is said to
be positive if f .x/ 0 for every x 2 X.
As in the previous section, if f ; g 2 C.X/ are real-valued functions, then the
notation f g is used to denote that g f 0.
Definition 6.48. A linear functional ' W C.X/ ! C is said to be positive if '.f / 0
for every f 2 C.X/ with f 0.
The next result illustrates a somewhat surprising fact: if a linear transformation
on C.X/ preserves positivity, then the linear transformation is necessarily continuous.
Proposition 6.49. If X is a compact space and if ' W C.X/ ! C is a linear
transformation for which '.f / 0 for every positive f 2 C.X/, then ' is continuous.
Proof. Let 1 2 C.X/ denote the constant function x 7! 1 2 C. For any f 2 C.X/
satisfying f .x/ 0 for all x 2 X, we have that 0 f kf k1, which implies that
0 '.f / kf k'.1/ in R.
Suppose now that g 2 C.X/ is real valued and write g D gC g , where gC ; g 2
C.X/ are given by gC D 12 .jgj C g/ and g D 12 .jgj g/ and satisfy 0 gC and
0 g . Thus, kgC k kgk and kg k kgk, and

j'.g/j j'.gC /j C j'.g /j kgC k C kg k '.1/ 2kgk'.1/:
Now let h 2 C.X/ be arbitrary and write h D < h C i= h. Because the real-valued
functions < h and = h are given by < h D 12 .h C h/ and = h D 2i1 .h h/, we have that
k< hk khk and k= hk khk. Thus,
j'.h/j j'.<h/j C j'.=h/j 2k< hk'.1/ C 2k= hk'.1/ 4khk'.1/:
Hence, ' is bounded.
t
u
240
6 Dual Spaces
A straightforward adaption of the proof of Lemma 6.38 yields:

Lemma 6.50. Every linear functional on C.X/ is a linear combination of four
positive linear functionals.
Examples of Zpositive linear functionals are point evaluations f 7! f .x0 /, and
integration f 7! f d. However, the two examples are related, for if X is compact
X
Z
and Hausdorff and if x0 2 X, then f d D f .x0 /, if is the point-mass measure
X
D fx0 g .
The following theorem, Theorem 6.51, is another Riesz Representation Theorem,
and is without doubt one of the major achievements of analysis. The version of
the theorem that is proved here requires the compact Hausdorff space X to be
second countable; under this assumption, the topology on X is metrisable, which
allows one to invoke Proposition 3.35 to show that the -algebra constructed from
a particular outer measure includes the Borel sets of X. Even so, Theorem 6.51 is
true for arbitrary compact Hausdorff spaces, but showing that the measurable space
that is constructed in the proof actually contains the Borel sets of X is a much more
delicate task without the assumption that X be second countable (see [10, 50]). When
appealing to Theorem 6.51 at later points of the present book, it will always be the
case that the topological space under consideration is a second countable compact
Hausdorff space, and in many ways the specific version of theorem proved here is
in fact the most important of all cases.
Theorem 6.51 (Riesz). If X is a second countable compact Hausdorff space, and
if ' is a positive linear functional on C.X/, then there exists a unique regular Borel
measure on the Borel sets of X such that
Z
'.f / D f d;
X
for every f 2 C.X/.

Proof. An important topological feature of (locally) compact Hausdorff spaces was
noted in the second version of Urysohns Lemma (Corollary 2.44) in Chapter 2: if
K and U are nonempty subsets of X such that K is compact, U is open, and K U,
then there exists a continuous function f W X ! 0; 1 such that f .K/ D f1g and supp f
is a compact subset of U. This result is key to linking the topology of X to elements
of C.X/ and, ultimately, to the functional '.
We begin with the construction of from the positive linear functional '. Let
I D ff 2 C.X/ j 0 f .x/ 1; 8 x 2 Xg and let S W T ! R (where T is the topology
of X) be defined by
s.U/ D supf'.f / j f 2 I and supp f Ug:
Note that the compactness of X implies that the support of any f 2 C.X/ is compact.
Now define a function W P.X/ ! R on the power set P.X/ of X by
241
.S/ D inffs.U/ j U 2 T and S Ug:

Our claim is that is an outer measure.
Because '.f / 0 for every f 2 I , the function s has nonnegative values.
Furthermore, if U1 and U2 are open sets such that U1 U2 , then clearly s.U1 /
s.U2 /; thus, .U/ D s.U/, for every
[U 2 T . Suppose the fUk gk2N is a countable
collection of open sets and let U D
Uk . If f 2 I satisfies supp f U, then the
k2N
compactness of supp f implies that there are finite many Uk that cover supp f , say
Uk1 ,. . . , Ukn . Let fh1 ; : : : ; hn g be a partition of unity of supp f subordinate to fUkj gnjD1
(Proposition 2.41). Thus, fhj 2 I and supp fhj Ukj , and so
0
'.f / D ' @f
n
X
1
hj A D
jD1
n
X
'.fhj /
jD1
n
X
.Ukj /
jD1
1
X
.Uk /:
kD1
Therefore,
Uk D sup '.f / j f 2 I and supp f
k2N
)
Uk
1
X
k2N
.Uk /;
kD1
which shows that is countably subadditive on open sets. To handle the case
of arbitrary subsets of X, assume that fSk gk2N is a countable collection of subsets
Sk X, and let " > 0. By definition, for each k 2 N there is an open set Uk containing
Sk and such that .Uk / < .Ek / C 2"k . Therefore,
[
k2N
!
Ek
[
k2N
!
Uk " C
1
X
.Ek /:
kD1
As " > 0 is arbitrary, the inequality above implies that is countably subadditive,
which completes the proof that is an outer measure on X.
One other feature of to mention before proceeding further is: if U1 and U2 are
disjoint open sets, then .U1 [ U2 / D .U1 / C .U2 /. To see this, let " > 0 and
select f1 ; f2 2 I such that supp fj Uj and .Uj / < '.fj / C "=2. Because U1 \ U2 D
;, the function f1 C f2 is an element of I and supp .f1 C f2 / .U1 [ U2 /. Thus,
.U1 / C .U2 / '.f1 / C '.f2 / C " D '.f1 C f2 / C " .U1 [ U2 / C ":
Hence, .U1 / C .U2 / .U1 [ U2 / .U1 / C .U2 / implies that is
additive on the union of two disjoint open sets.
To show that every Borel set of X belongs to the -algebra M .X/ induced by
, recall that, because the compact Hausdorff space X is second countable, the
topology on X is induced by a metric d on X (Theorem 2.48). Suppose that A1 and
A2 are subsets of X such that dist .A1 ; A2 / > 0. By continuity of the metric d, it will
242
6 Dual Spaces
also be true that dist .A1 ; A2 / > 0. Because X is normal, there are open disjoint sets
V1 and V2 such that Aj Vj . Let " > 0 be given. If W is an open set containing
A1 [ A2 , then let Uj D W \ Vj so that U1 \ U2 D ; and Aj Uj . Thus,
.A1 / C .A2 / .U1 / C .U2 / D .U1 [ U2 / .W/:
The infimum of the right-hand side over all open sets W that contain A1 [ A2 yields
.A1 [ A2 /, which shows that .A1 [ A2 / D .A1 / C .A2 /. Therefore, by
Proposition 3.35, every Borel set of X belongs to the -algebra M .X/ induced
by .
Thus, let denote the -subalgebra of M .X/ generated by the topology of X,
and let W ! R be the measure defined by .E/ D .E/, for every E 2 . We
shall now show that is a regular measure. By definition of , we already have
that
.E/ D inff.U/ j U is open and U Eg;
(6.5)
for every E 2 . To complete the proof of the regularity of , we need to show

(by Definition 3.58) that .U/ D supf.K/ j K is compact and K Ug for every
open set U X. Now if U D X, then there is nothing to show because U is compact.
Likewise, if .U/ D 0, then .K/.U/ D 0 for every compact K U. Thus, assume
that U 6D X and that .U/ > 0. Clearly, supf.K/ j K is compact and K Ug
.U/. Conversely, suppose that " > 0 satisfies 0 < .U/ "; by definition, there
exists a f 2 I with support supp f U and '.f / > .U/ ". Let K D supp f ;
we shall compute .K/ using equation (6.5). To this end, let W be an open set
containing K. Because K D supp f W, the definition of yields .W/ '.f /
.U/ ". Hence,
.U/ .K/ D inff.W/ j W is open and W Kg .U/ ";
which implies that .U/ D supf.K/ j K is compact and K Ug. Hence, is a
regular Borel measure.
We now show that if K X is compact, then
.K/ D inff'.f / j f 2 I and f .K/ D f1gg:
(6.6)
To this end, select f 2 I such that f .K/ D f1g, and let U D f 1 ..; 1//, for each
2 .0; 1/. Note that K U , for each . Suppose that g 2 I has support supp g
U . If x 2 supp g, then 0 g.x/ f .x/; and if x 62 supp g, then 0 D g.x/ D
f .x/. Thus, g f in C.X/, and so
'.g/ D '.g/ '.f /:
Thus,
.U / D supf'.g/ j g 2 I and supp g U g
'.f /
:
243
As the inequality above holds for every 2 .0; 1/, we have

1
.K/ inf .U / inf '.f / D '.f /:

Hence,
.K/ inff'.f / j f 2 I and f .K/ D f1gg:
To show that the inequality above can be reversed, select " > 0. By definition, there
is an open set W containing K and such that .W/ < .K/ C ". By Corollary 2.44,
there exists f 2 I with supp f W and f .K/ D f1g. Therefore, '.f / .W/, by
definition of . Hence,
.K/ inff'.f / j f 2 I and f .K/ D f1gg .W/ < .K/ C ";
which implies equation (6.6).
The integral representation of ' may now
Select any f 2 I . Fix
be established.

n 2 N and for k D 1; : : : ; n define Uk D f 1 . k1
;
1/
,
and
let
U
0 D X and UnC1 D ;.
n
These sets form a descending sequence
X D U 0 D U0 U 1 U1 U 2 U2 U n Un U nC1 D UnC1 D ;:
/Uk nUkC1 . Because the sets Uj are open, each fk is
Define fk D 1n UkC1 C .f k1
n
continuous. If x 2 Uk n UkC1 , then k1
< f .x/ nk , and so
n

k k1
1
k1

D ;
0 < f .x/
n
n
n
n
for all x 2 Uk n UkC1 . Thus,
Z
X
fk d D 1n .UkC1 / C
Z
Uk nUkC1
.f k1
/ d
n
1n .UkC1 / C 1n .Uk / .UkC1 /

D 1n .Uk /:
The fact that fk .x/ D 1=n for every x 2 UkC1 yields
1
.UkC1 /
n
by summing over all k we obtain the inequalities

1X
.UkC1 /
n kD1
n
1X
.Uk /:
n kD1
Z
fk d. Hence,
X
f d
X
(6.7)
244
6 Dual Spaces
The function fk is zero on Ukc ; thus, fx 2 X j fk .x/ 6D 0g Uk , which implies that

the support of fk satisfies supp fk U k Uk1 . Therefore, by definition of , and
noting that nfk and fk have the same support, we have that '.nfk / .Uk1 /. On the
other hand, by continuity, fk .x/ D 1n for all x 2 U kC1 . Therefore, by the compactness

of U kC1 and because nfk 2 I satisfies nfk U kC1 D f1g, equation (6.6) shows that
.UkC1 / '.nfk /. Hence, by summing over all k and dividing by n we obtain the
inequalities
1X
1X
.UkC1 / '.f /
.Uk1 /:
n kD1
n kD1
n
(6.8)
Therefore, using the fact that j j .b C c/ 2a if 2 a; b and 2 a; c,

inequalities (6.7) and (6.8) yield
Z
f d '.f / 1 ..U0 / C .U1 / .Un // 2.X/ :
n
n
X
Z
Because the choice of n 2 N is arbitrary, we obtain '.f / D
f d. The integral

X
formula for arbitrary f 2 C.X/ follows from the fact that the positive functions span
C.X/ and the fact both ' and the integral are linear maps.
To prove the uniqueness
of , suppose that Q is another regular Borel measure
Z
f d,
Q for every f 2 C.X/. Choose any open set U and let f 2 I
for which '.f / D

X
have support contained in U. Thus,

Z
Z
Z
'.f / D f dQ D f dQ
dQ D .U/:
Q
X
By definition of , the inequality above yields .U/ .U/.

Q
To prove the reverse
inequality, let " > 0 and select a compact set K such that K U and .U/
Q
< .K/
Q
C
". Select f 2 I with supp f U and f .K/ D f1g. Therefore, equation (6.6) and the
definition of yield
.U/
Q
< .K/
Q
C " '.f / C " .U/ C ":
Hence, .U/
Q
D .U/ for all open sets U. By regularity of the measures, we deduce
that .E/
Q
D .E/ for all Borel sets E.
t
u
Corollary 6.52. If X is a second countable compact Hausdorff space and if ' is a
linear functional on C.X/, then there exist a regular Borel measures 1 ; : : : ; 4 on
the Borel sets of X, such that, for every f 2 C.X/,
Z
'.f / D f d;
X
where is the complex measure D .1 2 / C i.3 4 /.
Problems
245
Proof. Lemma 6.50 and its proof show that ' has the form ' D .'1 '2 / C
i.'3 '4 /, for some positive linear functionals 'j on C.X/. Theorem 6.51 gives
a representing regular Borel measure j for each 'j .
t
u
R
For each complex measure on the Borel sets of X, the linear map f 7! X f d
is bounded. Thus, the dual space of C.X/ may be identified with the Banach
space M.X; / of complex measures. One would like such an identification to be
isometric, and such is the content of the next theorem which characterises the dual
space of C.X/.
Theorem 6.53. If X is a second countable compact Hausdorff space, then the map
W M.X; / ! C.X/ defined by
Z
./f D
f d;
X
for every f 2 C.X/, is a linear isometric surjection.

t
u
Problems
6.54. Suppose that p and q are conjugate real numbers. Prove that if g D .gk /k2N 2
`q .N/, then the function ' W `p .N/ ! C defined by
'.f / D
1
X
fk gk ;
kD1
for f D .fk /k2N 2 `p .N/, is a linear functional on `p .N/ of norm k'k D kgk.
6.55. Consider the Banach space `1 .N/.
1. Prove that if g D .gk /k2N 2 `1 .N/, then the function 'g W `1 .N/ ! C defined by
'g .f / D
1
X
fk gk ;
kD1
for f D .fk /k2N 2 `1 .N/, is a linear functional

`1 .N/ of norm k'k D kgk.
1
on

1
2. Prove that the function W ` .N/ ! ` .N/ defined by .g/ D 'g (as above)

is a linear isometric isomorphism of `1 .N/ and `1 .N/ .
6.56. Recall that c0 .N/ is the subspace of `1 .N/ given by
c0 .N/ D f.fk /k2N j lim fk D 0g:
k!1
246
6 Dual Spaces
1. Prove that if g D .gk /k2N 2 `1 .N/, then the function 'g W c0 .N/ ! C defined by
'g .f / D
1
X
fk gk ;
kD1
for f D .fk /k2N 2 c0 .N/, is a linear functional on c0 .N/ of norm k'k D kgk.
2. Prove that the function W `1 .N/ ! .c0 .N// defined by .g/ D 'g (as above)
is a linear isometric isomorphism of `1 .N/ and .c0 .N// .
6.57. Prove that if g D .gk /k2N 2 `1 .N/, then the function 'g W `1 .N/ ! C
defined by
'g .f / D
1
X
fk gk ;
kD1
1
for f D .fk /k2N 2 `1 .N/, is a linear functional on
1` .N/
of norm k'k D kgk.
1
Furthermore, prove that the function W ` .N/ ! ` .N/ defined by .g/ D 'g

(as above) is a linear isometric isomorphism of `1 .N/ and `1 .N/ .
6.58. Prove that if V and W are Banach spaces, then the Banach space B.V; W/ is
nonzero.
6.59. Suppose that M is a proper subspace of a Banach space V and that v 2 V is
nonzero and v 62 M. Prove that there exists ' 2 V such that '.v/ D 1 and '.w/ D 0
for every w 2 M. (Suggestion: consider the linear submanifold L D fw C v j w 2
M; 2 Cg and the linear map '0 W L ! C defined by '0 .w C v/ D .)
6.60. Assume that V is a real vector space and that p W V ! R is a sublinear
functional. Modify the proof of Theorem 6.18 to show that if L is a linear
submanifold of V and if ' W L ! R is a linear transformation for which '.v/ p.v/
for every v 2 L, then there is a linear transformation W V ! R such that jL D '
and p.v/ .v/ p.v/ for every v 2 V.
6.61. Suppose that V is a Banach space.
1. (a) Prove that if V is a separable, then V is separable.
2. (b) Show by example that there are separable Banach spaces V for which V is
nonseparable.
6.62. Prove that `p .N/ is a reflexive Banach space, for all p 2 R such that p > 1.
6.63. Prove that neither c0 .N/ nor `1 .N/ is a reflexive Banach space.
6.64. Let V be a normed vector space and suppose that '1 ; : : : ; 'n 2 V . Let
LD
n
\
ker 'j :
jD1
Assume that ' 2 V satisfies './ D 0, for every 2 L.
Problems
247
3
'1 .v/
7
6
1. Let W V ! Cn be given by .v/ D 4 ::: 5, v 2 V. Show that there is a linear
2
'n .v/
W Cn ! C such that '.v/ D ..v//, for every v 2 V.
n
X
j 'j .
2. Show that there are 1 ; : : : ; n 2 C such that ' D
functional
jD1
6.65. Prove that the weak topology and the norm topology on a finite-dimensional
normed vector space coincide.
6.66. Suppose that H is an infinite-dimensional separable Hilbert space.
1. Prove that the closed unit ball of H is weakly compact.
2. Prove that the zero vector is in the weak closure of the unit sphere of H.
6.67. Prove that the vector space operations on V are continuous in the weak
topology, for every normed vector space V.
6.68. Prove that every linear functional on L1 .X; ; / is a linear combination of
four positive linear functionals.
6.69. Prove that if Theorem 6.45 is true for finite measure spaces, then it is also
true for -finite measure spaces.
6.70. Assume that X is a second countable compact Hausdorff space, and define a
linear map W M.X; / ! C.X/ by ./ D ' , where
Z
' .f / D
f d;
X
for every f 2 C.X/. Prove that is an isometric surjection.

6.71. Suppose that is a finite regular Borel measure on R, and Zconsider the
functions gn W R ! C, for n 2 Z, defined by gn .t/ D eint . Prove that if gn d D 0
for every n 2 Z, then D 0.
Chapter 7
Convexity
The linear character of functional analysis underscores the entire subject. In addition
to geometric structures such as subspaces, it can be important to consider subsets C
of vector spaces V that are locally linear in the sense that C contains the line segment
in V between every pair of points of C. That is, if u; v 2 C, then so is tu C .1 t/v
for every t 2 0; 1. Such sets are said to be convex and they are crucial structures in
functional analysis, useful for both the geometrical and topological information that
they reveal about a space V.
7.1 Convex Sets

Recall that a subset C of a vector space V is convex if tuC.1t/v 2 C for all u; v 2 C
and every t 2 0; 1. More generally, a convex combination of elements v1 ; : : : ; vn in
a vector space V is a sum of the form
n
X
tj vj ;
(7.1)
jD1
where each tj 2 0; 1 and
n
X
tj D 1.
jD1
Proposition 7.1. A subset C is convex if and only if C contains every convex

combination of its elements.
t
u
The scalars that arise in (7.1) are real, and so convexity can be studied completely
within the realm of real vector spaces. However, functional analysis is mostly carried

DOI 10.1007/978-3-319-45633-1_7
249
250
7 Convexity
out over the complex field. Thus, in this chapter, we shall use real vector spaces
when it is helpful to do so and then make use of the fact that every complex vector
space is also a real vector space.
Of course any subspace of a vector space is a convex set. Some basic convex sets
of real and complex numbers are:
1.
2.
3.
4.
any interval J of real numbers;

the open unit disc D of complex numbers;
the closed unit disc D of complex numbers;
the right halfplane fz 2 C j <z 0g of C.
For every subset S V there is a smallest convex set C that contains S.
Definition 7.2. If S V, then the set of all convex combinations of elements of S

is called the convex hull of S and is denoted by Conv S.
The following theorem is very useful in the convexity theory of finitedimensional space.
Theorem 7.3 (Carathedory). If V is an n-dimensional vector space over R and
if S V is nonempty, then for each v 2 Conv S there are v1 ; : : : ; vm 2 S such that v
is a convex combination of v1 ; : : : ; vm and m n C 1.
Proof. Let v 2 Conv S. Thus, v is a convex combination of v1 ; : : : ; vm 2 S, say
vD
m
X
tj vj ;
jD1
where each tj 6D 0. If m .n C 1/, then the desired conclusion is reached. Therefore,

Q for 1 j m.
suppose that m > .n C 1/. Let VQ D R V and let vQ j D .1; vj / 2 V,
Q the vectors vQ 1 ; : : : ; vQ m 2 W
Q are linearly dependent. Thus,
Since m > n C 1 D dimV,
P
P
therePare 1 ; : : : ; m 2 R, not all zero, such that j j vQ j D 0; hence, j j D 0 in R
and j j vj D 0 in V.
Let i be such that

j i
; 8 j 6D i ;
t t
j
i
and set sj D tj
j ti
for every j. Then
i
sj 2 0; 1; si D 0;
m
X
sj D
jD1
But the number of nonzero summands in
m
X
tj D 1 ; and v D
jD1
m
X
jD1
m
X
sj vj :
jD1
sj vj is less than m since si D 0. Hence,
if v can be expressed as a convex combination of m > n C 1 elements of S, then v
7.2 Separation Theorems
251
can be expressed as a convex combination of m 1 of these same elements. This

shows, by iteration of the argument, that the number of summands can be reduced
to n C 1.
t
u
The upper bound of n C 1 in Carathedorys Theorem is sharp, as one sees by
considering a triangle C in R2 with vertex set S.
Corollary 7.4. If V is an n-dimensional vector space over C and if S V is
nonempty, then for each v 2 Conv S there are u1 ; : : : ; um 2 S such that v is a convex
combination of u1 ; : : : ; um and m 2n C 1.

This section establishes a cornerstone result called the Hahn-Banach Separation
Theorem that provides analytic information from the knowledge that two convex
sets are disjoint. This theorem is applied to subsets of Banach spaces in a variety
of topologies (norm, weak, weak , etc.), and so the development below will not
immediately make use of Banach spaces and linear functionals, but rather topological vector spaces V and linear transformations ' W V ! C that are continuous in the
topology of V.
Recall that a vector space V is a topological vector space if V is a topological
space for which addition and scalar multiplication are continuous functions V V !
V and C V ! V, respectively.
Proposition 7.5. If V is a topological vector space, then
1. W is an open subset of V if and only if there exists w0 2 W and open
neighbourhood U of 0 such that W D fw0 g C U, and
2. for each open neighbourhood U of 0 2 V and v 2 V there exists a " > 0 such that
v 2 U for all 2 C with j j < ".
Proof. For (1), select w0 2 W and let U D fw0 w j w 2 Wg. Because scalar
multiplication and vector addition are continuous, U is an open set and W D
fw0 g C U. For (2), because scalar multiplication m W C V ! V is continuous, the
map mv W C ! V defined by mv ./ D m.; v/ D v is continuous for each fixed
v 2 V. Thus if U V is an open neighbourhood of 0 2 V, then m1
v .U/ is an open
neighbourhood of 0 2 C. Hence, there exists a " > 0 such that f 2 C j j j < "g
m1
t
u
v .U/.
The next analytical tool is a sublinear functional (Definition 6.19) known as the
Minkowski functional.
Proposition 7.6. If C is an open convex subset of a topological vector space V such
that 0 2 C, then the map p W V ! R defined by
p.v/ D inf ft > 0 j t1 v 2 Cg;
for v 2 V, is a sublinear functional and C D fv 2 V j p.v/ < 1g.
(7.2)
252
7 Convexity
Proof. The sublinearity of p is evident. Suppose that v 2 V satisfies p.v/ < 1. Then
there exists t 2 .0; 1/ and u 2 C such that 1t v D u. That is, v D tu D tu C .1 t/0 2
C. Conversely, if v 2 C, then there is an open set W such that v 2 W C. By
Proposition 7.5, W may be taken to be W D fvg C U for some open neighbourhood
U of 0 for which 0 2 U C. Proposition 7.5 shows that there is a " > 0 such that
"v 2 U. Thus, .1 C "/v D v C "v 2 W C implies that p.v/ .1 C "/1 < 1.
t
u
Theorem 7.7 (Hahn-Banach Separation Theorem). Assume that C1 and C2 are
nonempty, disjoint convex subsets of a topological vector space V such that C1 is
open. Then there are a linear transformation ' W V ! C and a 2 R such that ' is
continuous and
<'.v1 / < <'.v2 / ;
Proof. Let C D
8 v1 2 C1 ; v2 2 C2 :
(7.3)
fv1 v2 j v2 2 C2 g, and note that C is open and convex. Further,
v1 2C1
0 62 C, as C1 \ C2 D ;. Select v0 2 C and let C0 D fv0 v j v 2 Cg. Thus, C0 is open,

convex, and 0 2 C0 . Let p be the Minkowski functional (7.2) associated with C0 .
Because 0 62 C implies v0 62 C0 , Proposition 7.6 yields p.v0 / 1.
Now let L D SpanR fv0 g and define 0 W L ! R by 0 . v0 / D p.v0 /, for all
2 R. The function is linear over R and satisfies 0 . v0 / D p. v0 / if 0. If
< 0, then 0 . v0 / D p.v0 / < 0 p. 0 /. Thus, 0 .w/ p.w/ for every w 2 L. By
the version of the Hahn-Banach Extension Theorem in Exercise 6.60, 0 extends to
a linear transformation W V ! R such that p.v/ .v/ p.v/ for every v 2 V.
Because C0 is open and contains 0, for each v 2 V there exists > 0 such that
v 2 C0 ; thus,
1 > p.v/ D .p.v// .v/ D .v/ p.v/ < 1:
Therefore, j.u/j < 1 for every u 2 U D C0 \ .C0 /, an open neighbourhood of
0. Hence, for every " > 0 such that " < 1, W" D ."; "/ is a neighbourhood of
.0/ D 0 in R and the set U" D "U U is a neighbourhood of 0 2 V such that
.U" / W" . Thus, is continuous at 0 2 V. By Proposition 7.5, is therefore
continuous at every v 2 V, which implies that W V ! R is a continuous linear
transformation. Because scalar multiplication and vector addition are continuous,
the linear transformation ' W V ! C defined by '.v/ D .v/ i.iv/, for v 2 V, is
continuous and D <'.
Let N D ker D 1 .f0g/, which is closed by the continuity of . If v 2 C, then
v0 v 2 C0 and so
1 < p.v0 v/ .v0 v/ D .v0 / .v/ D p.v0 / .v/I
that is, .v/ > p.v0 / 1 0, which implies .v/ 6D 0. Thus, ker \ C D ;.
Therefore, because C is convex and is continuous, .C/ is a connected subset
of R. But ker \ C D ; implies that .C/ .1; 0/ or .C/ .0; 1/. Without
253
loss of generality, assume that .C/ .1; 0/. Hence, .v1 / < .v2 /, for all
v1 2 C1 , v2 2 C2 . Therefore, there is a 2 R such that
sup .v1 / inf .v2 / :
v2 2C2
v1 2C1
Because C1 is an open convex set and because is a continuous linear transformation, the interval .C1 / is open in R. Hence, 62 .C1 / and so the inequalities (7.3)
hold.
t
u
In the case of Banach spaces, the following Hahn-Banach Separation Theorem
is of particular importance.
Corollary 7.8. If C1 and C2 are convex sets in a Banach space V such that C1 \
C2 D ;, C1 is compact, and C2 is closed, then there exist ' 2 V and 1 ; 2 2 R such
that
<'.v1 / < 1 < 2 < <'.v2 / ;
8 v1 2 C1 ; v2 2 C2 :
(7.4)
Proof. Assume that C1 ; C2 V are disjoint convex sets and that C1 is compact and
C2 is closed. By Proposition 5.14, there exists " > 0 such that .C1 CB" .0//\C2 D ;.
Note that C1 C B" .0/ is convex and open. Thus, by Theorem 7.7, there exist ' 2 V
and 2 R such that
<'.v1 / < <'.v2 / ;
8 v1 2 C1 C B" .0/; v2 2 C2 :
However, as C1 is compact, <'.C1 / is compact and <' .C1 C B" .0// has compact
closure, disjoint from C2 . Therefore, there are 1 ; 2 2 R such that
<'.v1 / < 1 < 2 < <'.v2 /;
for all v1 2 C1 and v2 2 C2 .
t
u
A somewhat surprising consequence of Hahn-Banach Separation Theorem is the

following result about topology, which demonstrates the essential role convexity
theory plays in understanding Banach spaces.
Proposition 7.9. Let C be a nonempty convex subset of a Banach space V. If C and
wk
wk
C denote the closures of C in the norm and weak topologies of V, then C D C .
Proof. Because C is convex and closed, if v0 62 C, then by the Hahn-Banach
Separation Theorem (Corollary 7.8) there exist ' 2 V and 2 R such that, for
all v 2 C, .v/ < < .v0 /, where D <'. Therefore, there exists " > 0 such that
j.v/ .v0 /j " for all v 2 C. In particular, for v 2 C,
j'.v/ '.v0 /j2 D ..v/ .v0 //2 C ..iv0 / .iv//2 ..v/ .v0 //2 "2 :
wk
Hence, v0 62 C . This proves that C
wk
C.
254
7 Convexity
wk
Conversely, for any set C, it is always the case that C C .
t
u
To make additional concrete applications of the Hahn-Banach Separation Theorem, one needs to have a sufficient amount of information concerning continuous
linear maps V ! C on topological vector spaces V. The following proposition is
one such example.
Proposition 7.10. If V is the dual of a normed vector space V, then
1. V is a Hausdorff topological vector space in the weak -topology, and
2. a linear transformation W V ! C is continuous with respect to the weak topology if and only if there exists v 2 V such that .'/ D '.v/ for every ' 2 V .
Proof. The proof of the first statement is left as an exercise (Exercise 7.30).
Linear maps of the form ' 7! '.v/, for fixed v 2 V, are continuous by the
definition of weak -topology. Suppose, conversely, that W V ! C is an arbitrary
linear transformation and is continuous with respect to the weak -topology. Thus, if
D C is the open unit disc in C, then there is a basic weak -open set U V such
that 0 2 U 1 .D/. By definition, such a set U has the form
UD
n
\
f
2 V j j .vj /j < "g;
jD1
for some v1 ; : : : ; vn 2 V and " > 0. Therefore, j. /j < 1 for all

if ' 2 V is arbitrary and is nonzero on at least one vj , then
2 U. In particular,
"
' 2 U;
2 max` j'.v` /j
and so j.'/j < "1 max` j'.v` /j. Hence
j.'/j
2 max` j'.v` /j
;
"
(7.5)
for every ' 2 V .

Consider the linear transformation T W V ! Cn defined by
3
'.v1 /
7
6
T.'/ D 4 ::: 5 :
2
'.vn /
On the range of T, define a function 0 W ran T ! C by 0 .T'/ D .'/. This
function 0 is well defined because, if '1 ; '2 2 V satisfy T'1 D T'2 , then .'1
'2 /.vj / D 0 for each j D 1; : : : ; n, and so inequality (7.5) gives .'1 '2 / D 0.
Because 0 is also linear, 0 extends to a linear functional on Cn . Hence, is
given by an n 1 matrix D 1 2 n and, for every ' 2 V ,
7.3 Extreme Points
255
.'/ D hT'; i D
where v D
n
X
n
X
0
1
n
X
j '.vj / D ' @
j vj A D '.v/;
jD1
j1
j vj .
t
u
jD1
With Proposition 7.10 in hand, the following Hahn-Banach Separation Theorem

holds for dual spaces.
Proposition 7.11. Assume that V is the dual of a normed vector space V. If
K V is weak -compact and convex, and if '0 62 K, then there exist 2 R and
v0 2 V such that
<'0 .v0 / < <'.v0 /
for every ' 2 K.
Proof. The dual space V is a Hausdorff topological vector space in the weak topology. By the Hausdorff property, for each ' 2 K there are weak -open sets U'
and W' such that '0 2 U' , ' 2 W' , and U' \ W' D ;. Because K is compact, there
are finitely members W'1 : : : ; W'n of the open cover fW' g'2K that cover K. Hence, if
UD
n
\
U'j and W D
jD1
n
[
W'j ;
jD1
then U and V are disjoint open sets with '0 2 U and K V. The proof of
Theorem 7.7 shows that there is a convex open set U0 that contains the origin
and Proposition 7.5 shows that there is a " > 0 such that C1 D f'0 g C "U0 U.
Hence, C1 is an open convex set disjoint from K. By Theorem 7.7 there exist a
weak -continuous linear transformation W V ! C and 2 R such that <.'0 / <
<.'/ for all ' 2 K. However, by Proposition 7.10, the map is given by
evaluation at some point v0 2 V. Hence, <'0 .v0 / < <'.v0 / for every ' 2 K.
t
u
7.3 Extreme Points

Triangles and squares are determined by their vertices, a disc by its boundary circle,
and a Euclidean ball in R3 by its boundary surface (sphere). The general concept in
convexity theory that captures these phenomena is that of an extreme point.
Definition 7.12. An element v in a convex subset C V is an extreme point of C
if the equation
vD
n
X
jD1
tj vj ; where each vj 2 C;
n
X
tj D 1; and each tj 2 .0; 1/;
(7.6)
jD1
holds only for v1 D D vn D v. The set of all extreme points of C is denoted by

ext C.
256
7 Convexity
Thus, if C is convex, then v 2 C is an extreme point if v is not a proper convex

combination of elements other than itself. Geometrically this means:
Proposition 7.13. Let C be a convex set and v 2 C. The following statements are
equivalent:
1. v is an extreme point of C;
2. v is not interior to any line segment contained in C (i.e., if there are v1 ; v2 2 C
and t 2 .0; 1/ such that v D tv1 C .1 t/v2 , then v1 D v2 D v).
t
u
Proposition 7.13 implies that the extreme points of a convex set C must lie on
the topological boundary of C. If not all points of the boundary are extreme points,
then the boundary contains line segments. On the other hand, if every point on the
boundary is an extreme point, then we may think of the convex set as having a
curved or rounded boundary.
Example 7.14. The extreme points of the closed unit disc in the complex plane are
precisely the points on the unit circle.
Proof. To verify this assertion, note that if 2 ext D, then necessarily jj has
modulus 1 by Proposition 7.13. Conversely, suppose that 2 D is such that jj D 1.
Thus, there exists 2 R such that D cos C i sin . Assume that D t C .1 t/,
for some t 2 .0; 1/ and ; 2 D. The triangle inequality yields j j D jj D 1 and so
D cos C i sin and D cos C i sin for some ; 2 R. Hence,
1 D cos2 C sin2 D 1 C 2t2 2t C 2t.1 t/ .cos. // ;
which implies that cos. / D 1; that, is D C .2k C 1/
for some k 2 Z.
Thus, D D .
t
u
Example 7.15. The set of extreme points of the closed unit ball of a Hilbert space
H is the set of unit vectors of H.
Proof. Let H1 denote the closed unit ball of H. Choose any unit vector 2 H and
assume that D t1 C .1 t/2 for some 1 ; 2 2 H1 and t 2 .0; 1/. Thus,
1 D h; i D th1 ; i C .1 t/h2 ; i
expresses the number 1 as a proper convex combination of the complex numbers
h1 ; i and h2 ; i in the closed unit disc D C (as jhj ; ij kj k kk 1).
Example 7.14 shows that 1 is an extreme point of D, and therefore 1 D hj ; i for
each j D 1; 2, which give cases of equality in the Cauchy-Schwarz Inequality. Hence,
j D j for some j 2 C; but 1 D hj ; i D j h; i D j yields 1 D 2 D , and so
2 ext H1 .
t
u
Convex sets that are not topologically closed may fail to have extreme points.
For example, the open unit ball in a Hilbert space does not have any extreme points.
But even closedness of a convex set is an insufficient condition for extreme points
to exist.
7.3 Extreme Points
257
Example 7.16. The closed unit ball of L1 .0; 1; M; m/ has no extreme points.
Proof. Extreme points of the closed unit ball in any Banach space, if they exist,
1
necessarily lie on the unit sphere. Thus, suppose that f 2 L
Z .0; 1; M; m/ is such
that kfP k D 1 and select a Borel subset E 0; 1 such that jf j dm D 1=2. Define
E
P gP 2 L1 .0; 1; M; m/ are

h; g 2 L 1 .0; 1; M; m/ by h D 2f E and g D 2f Ec . Then h;
1 P
P
P
P
unit vectors and f D 2 .h C gP /. Because gP 6D f , f is not an extreme point of the closed
unit ball of L1 .0; 1; M; m/.
t
u
The main value in knowing the extreme points in a convex set (if they exist) is
that they can be used to recover the convex set itself by way of the convex hull. For
example, in the closed unit ball of a Hilbert space, every point is either an extreme
point or an average of two extreme points. Indeed, the zero vector is evidently an
average of and , for any unit vector 2 H. Suppose that 2 H is of norm
kk < 1. Let 1 D kk1 and 2 D kk1 , so that 1 and 2 are extreme points
of H1 . With t D 12 .1 kk/ we obtain D t1 C .1 t/2 , and so the closed unit ball
of H is the convex hull of the unit spherethat is, H1 D Conv .ext H1 /.
In contrast to Example 7.16, a compact convex set will possess extreme points
and, moreover, knowledge of the extreme points of a compact convex set K is
sufficient to recover the entire set K.
Theorem 7.17 (Kren-Milman). If K is a nonempty compact convex subset of a
Banach space space V, then
1. the set ext K of extreme points of K is nonempty, and
2. K is the closure of the convex hull of the set ext K,
Proof. Consider a convex subset F K with the property that if v 2 F and tv1 C
.1 t/v2 2 F, for some t 2 .0; 1/ and v1 ; v2 2 K, then v1 ; v2 2 F. (Such sets F are
called faces.) Assume further that F is compact and define
SF D fG F j G is a nonempty compact face of Fg:
Use reverse inclusion to partially order SF : G1 G2 if and only if G2 G1 . Let L
be a linearly ordered subset of SF . Hence, if G1 ; : : : ; Gn 2 L, then there is a j0 such
that Gi Gj0 for all 1 i n. Hence, ; 6D Gj0 G1 \ \ Gn . Therefore, the family
L of compact sets has the finite intersection property, and so F0 6D ;, where
\
F0 D
G:
G2L
As F0 is a nonempty compact face of F, F0 is an upperbound in SF for L. Zorns

Lemma implies that SF has a maximal element, say E.
Let ' 2 V , D maxf<'.v/ j v 2 Eg, and E' D fv 2 E j '.v/ D g. If t 2 .0; 1/
and v1 ; v2 2 E are such that tv1 C .1 t/v2 2 E' , then
D t<'.v1 / C .1 t/<'.v2 / t C .1 t/ D ;
258
7 Convexity
which implies that v1 ; v2 2 E' . Hence, E' is a face of E and, as well, a face of F.
Furthermore, E' is closed, and so it is compact (since it is a closed subset of a
compact set). Thus, E' 2 SF and E E' . By the maximality of E in SF , E D E' .
We conclude, therefore, that for every ' 2 V and all v1 ; v2 2 E, <'.v1 / D <'.v2 /.
The formula
'.v/ D <'.v/ i<'.iv/
yields '.v1 v2 / D 0 for every ' 2 V , which proves that v1 D v2 . Hence, E is a
singleton set fv0 g. But fv0 g is a face of K if and only if v0 is an extreme point of K,
and so the set ext K of extreme points of K is nonempty.
Next let C D Conv .ext K/ and consider the compact convex subset C of K. If C 6D
K, then let w0 2 K n C. By the Hahn-Banach Separation Theorem (Theorem 7.7),
there are ' 2 V and 1 ; 2 2 R such that
<'.v/ < 1 < 2 < <'.v0 / ;
8v 2 C:
If D maxf<'.w/ j w 2 Kg, then 2 < <'.w0 / . The set

K' D fw 2 K j <'.w/ D g
is a compact face of K. Therefore, by the proof of the first statement, there is an
extreme point v0 of K in K' . Thus, <'.v0 / D and <'.v0 / < (since v0 2 extK
C), which is a contradiction. Hence, it must be that C D K.
By changing the topology on a Banach space, one may produce different versions
of the Kren-Milman Theorem. One of the most useful versions occurs with the dual
space in its weak -topology.
Theorem 7.18 (Kren-Milman Theorem: Weak -Topology Version). If K is a
nonempty weak -compact convex subset of the dual space V of a normed vector
space V, then
1. the set ext K of extreme points of K is nonempty, and
2. K is the weak -closure of the convex hull of the set ext K,
Proof. The proof proceeds as in the proof of Theorem 7.17 to produce, using Zorns
Lemma, a minimal face E of K.
Fix v 2 V and define D maxf<'.v/ j ' 2 Eg and let Cv D f' 2 E j <'.v/ D g.
The set Cv is a weak -closed convex subset of K; thus, Cv is weak -compact. As in
the proof of Theorem 7.17, Cv is a face of K such that Cv E. Therefore, Cv D E
and, hence, E is a singleton set, whence K has an extreme point.
As in the proof of Theorem 7.17, the Hahn-Banach Separation Theorem is
required for the second statement. In the case of the weak -topology, it is the version
of the Hahn-Banach Separation Theorem given in Proposition 7.11 that yields the
result.
7.4 Extremal Regular Probability Measures
259
One consequence of Theorem 7.18 is that it tells us something about the

geometry of the closed unit ball of dual spaces. Thus, Banach spaces that fail to
have this property are not the duals of other Banach spaces.
Example 7.19. L1 .0; 1; M; m/ is not isometrically isomorphic to the dual space
V of any normed vector space V.
Proof. Theorem 7.18 says that the closed unit ball of V has an extreme point.
However, Example 7.16 shows that the closed unit ball of L1 .0; 1; M; m/ has no
extreme points.
t
u
7.4 Extremal Regular Probability Measures

The purpose of this section is to give an interesting example in which the extreme
points of a compact convex set may be determined completely.
Assume that .X; / is a measurable space in which X is a compact Hausdorff
space and is the -algebra of Borel sets of X. Consider the set
P.X; / D f 2 M.X; / j .E/ 0; 8 E 2 ; .X/ D 1g;
which is a subset of the closed unit sphere of the Banach space M.X; / of
regular complex measures on .X; /. Because M.X; / is the dual space of
C.X/, M.X; / carries a weak -topology. Furthermore, because P.X; / is weak compact (Exercise 7.36), the Kren-Milman Theorem implies that P.X; / is the
weak -closure of the closed convex hull of the extreme points of P.X; /.
Recall that if K X is a closed subset, then the -algebra K of Borel sets of K
is given by K D fK \ E j E 2 g.
Definition 7.20. The support of a regular measure on a measurable space .X; /,
where X is a compact Hausdorff space and is the -algebra of Borel sets of X, is
the smallest closed subset K X for which .X n K/ D 0.
Lemma 7.21. Assume K X is the support of 2 P.X; /. Then is an extreme
point of P.X; / if and only if the restriction jK of to K is an extreme point
of P.K ; K /.
Proof. Assume that is an extreme point of P.X; /. Let 0 ; 1 ; 2 2 P.K ; K /
1
and such that jK D 0 D .1 C 2 /. Define Q j W ! R by Q j .E/ D j .E \ K /
2
for all E 2 , to obtain Q j 2 P.X; /. Because K is the support of , .E/ D .E \
K / for all E 2 ; thus, D 12 .Q 1 C Q 2 /, and so D Q 1 D Q 2 , and so 0 D 1 D 2 .
Conversely, assume that 0 D jK is an extreme point of P.K ; K /. Let D
1
.1 C 2 / for 1 ; 2 2 P.X; /. If E 2 satisfies .E/ D 0, then 0 D .E/
2
1
0 implies that j D 0. Thus, j . If we show that the support of each j
2 j
is contained in the support of , then we conclude that 1 D 2 D .
260
7 Convexity
Thus, it remains to prove that if !; 2 P.X; / is such that ! , then

K! K . To this end, let U D .X n K / \ .X n K! /, which is open, and let K D
.X n U/ \ K! , which is closed. Thus,
!.X n K/ D ! .U [ .X n K! // !.U/ C !.X n K! / D !.U/:
Now since U X n K , we have .U/ .X n K / D 0. Thus, ! implies that
!.U/ D 0 and so !.X n K/ D 0. Hence, K K! and !.X n K/ D 0 which implies
that K D K! by definition of support and by the above arguments. Hence, K! D K D
.X n U/ \ K! D .K [ K! / \ K! implies that K! K .
t
u
The following theorem is a cornerstone of probability theory.
Theorem 7.22. The following statements are equivalent for 2 P.X; /:
1. is an extreme point of P.X; /;
2. D fx0 g for some x0 2 X.
Proof. Assume that is an extreme point of P.X; /. By Lemma 7.21, we may
replace X with the support of , and so we assume without loss of generality
that X D K . Suppose, contrary to what we aim to prove, that the support X of
contains at least two points, x and y. Because X is a normal topological space
(Proposition 2.34), there are open subsets Ux and Vy containing x and y, respectively,
and such that Ux \ Uy D ;. Because X is the support of , both Ux and Vy have
positive measure. Therefore, if f D Ux and g D Uy , then fP and gP are linearly
independent elements of L1 .X; ; /. Moreover, fP is clearly linearly independent
P where 1 D X . Hence, by Exercise 6.59, there is a linear functional ' on
of 1,
P D 0.
L1 .X; ; / such that '.fP / D 1 and '.1/
Theorem 6.45 asserts that linear functionals on L1 .X; ; / are determined by
1
P
elements of L1 .X; ;
R /. Hence, there exists 2 L .X; ; / Rsuch that k k D 1
1
in L .X; ; / and X d D 0. Define Q W ! R by .E/
Q
D E d. Let 1 D
C Q and 2 D .
Q Note that
Z
1 .E/ D
d. C /
Q D
E
d C
.1 C / d:
d D
Since 1 C is nonnegative for almost all x 2 X, this final integral above is

nonnegative. Likewise 2 .E/ is nonnegative. Further,
Z
1 .X/ D
Z
d. C /
Q D .X/ C
d D .X/ C 0 D 1:
X
Hence, 1 ; 2 2 P.X; / and D 12 1 C 12 2 . Because 1 (as P 0), the

measure is not an extreme point of P.X; /. This contradiction implies that X
must be a singleton set, which is to say that is a point-mass measure.
The proof that point-mass measures fx0 g are extremal is left as an exercise
(Exercise 7.36).
t
u
7.5 Integral Representations of Compact Convex Sets
261

The Kren-Milman Theorem asserts that every element x0 of a compact convex set
K in a Banach space V is a limit point of the set of all convex combinations of the
extreme points of K. However, suppose that x0 is already a convex combination
m
X
j xj , where x1 ; : : : ; xm 2 ext K and where
of extreme points of K, say x0 D
1 ; : : : ; m 2 0; 1 satisfy
jD1
j j D 1. Then, for every linear functional ' W V ! C,
'.x0 / D
m
X
j '.xj /:
jD1
That is,
Z
'.x0 / D
' d;
(7.7)
for every ' 2 V , where is the probability measure D
m
X
j fxj g on the Borel
jD1
sets of K. Observe that not only is a measure on the Borel sets of K, but that it is
in fact supported on the Borel set E D fx1 ; : : : ; xm g (in the sense that .K n E/ D 0.
If it were possible, for each x0 2 K, that the integral equation (7.7) held for some
regular Borel probability measure supported on the extreme points of K, then
this would represent a sharpening of the Kren-Milman Theorem. The goal of this
section is to prove such a result (Theorem 7.25), due to Choquet, in the case where
the topology of K is metrisable.
Definition 7.23. If C is a convex set, then a function h W C ! C is an affine
function if
h. x C .1 /y/ D h.x/ C .1 /h.y/;
for every x; y 2 C and 2 0; 1.
In cases where the convex set C is a subset of a vector space V, one of the most
immediate ways to produce an affine function h on C is to take any linear map
' W V ! C on V and then consider h D 'jC .
Recall that f W C ! R is convex if f . x C .1 /y/ f .x/ C .1 /f .y/ for all
x; y 2 C and 2 0; 1, and that f W C ! R is concave if the function f is convex.
Thus, every affine function h on a convex set C is both convex and concave.
Lemma 7.24. If f W K ! R is a function on a compact, convex topological space K
such that f is bounded above by q 2 R, then the function f defined by
f .x/ D inffh.x/ j h 2 Aff K; f hg; x 2 X;
is concave, upper-semicontinuous, and bounded above by q. Furthermore, if f itself
is concave and upper-semicontinuous, then f D f .
262
7 Convexity
Proof. By hypothesis, there is a real number q 2 R with f .x/ q for every x 2 K.

Therefore, if F D fh 2 Aff K j f hg, then the constant function r 2 F and, hence,
f .x/ r for all x 2 K, thereby establishing that f is bounded above by r.
To prove that f is concave, select x; y 2 K and 2 0; 1. Thus,
f . x C .1 /y/ D inf fh . x C .1 /y/ j h 2 F g
D inf f h.x/ C .1 /h.y/ j h 2 F g
inf fh.x/ j h 2 F g C .1 / inf fh.y/ j h 2 F g
D f .x/ C .1 /f .y/:
Thus, f is a concave function.
1
1
Now select 2 R and consider f ..1; //. If x 2 f ..1; //, then f .x/ <
implies that there exists h 2 F such that f .x/ h.x/ < , which implies that
x 2 h1 ..1; //. If, on the other hand, x 2 h1 ..1; // for some h 2 F , then
1
x 2 f ..1; //. Hence,
f
1
..1; // D
h1 ..1; // ;
h2F
which is an open set since each h 2 F is continuous.

Now assume that f W K ! R is upper-semicontinuous and concave. Because f is
upper-semicontinuous, the graph G.f / D f.x; f .x// 2 V R j x 2 Kg of f is a closed
set. Further, the concavity of f implies that G.f / is convex.
Suppose, contrary to what we aim to prove, that f 6D f . Thus, f .x1 / < f .x1 / for
at least one x1 2 K, which implies that .x1 ; f .x1 // is separated from the closed
convex set G.f /. Therefore, by the Separation Theorem (Theorem 7.7), there is a
(real)
linear functional ' W V R ! R and a 2 R such that ' ..x; f .x/// < <
' .x1 ; f .x1 // for every x 2 K. In particular, using that V R is a vector space,

0 < ' .x1 ; f .x1 // .x1 ; f .x1 // D ' .0; f .x1 / f .x1 //

D f .x1 / f .x1 / '..0; 1//;
and therefore '..0; 1// > 0.
For each x 2 K, let `x W R ! R denote the function given by `x .s/ D '..x; s//. If
s < s0 , then `x .s0 / `x .s/ D '..0; s0 s// D .s0 s/'..0; 1//, and so `x is a strictly
increasing function with lims!1 `x .s/ D 1. Thus, because `x .f .x// < , there is
a unique r 2 R for which `x .r/ D . Therefore, let h W K ! R denote the function
h.x/ D r, where r 2 R is the unique real number such that '..x; r// D . Note that
the continuity of ' implies that h is continuous. Further, if x; y 2 K and 2 0; 1,
and if r D h.x/ and s D h.y/, then
263
D ' ..x; r// C .1 /' ..y; s//

D ' .. x; r/ C ..1 /y; .1 /s//
D ' .. x C .1 /y; r C .1 /s// :
Hence, h. x C .1 /y/ D h.x/ C .1 /h.y/, which shows that h is an affine
function on K, and so h 2 Aff K. Because, for every x 2 K, `x is strictly increasing
and `x .f .x// D '..x; f .x// < , we have that f .x/ < h.x/ for all x 2 K. Hence,
f .x/ h.x/ for each x 2 K. However, '..x1 ; h.x1 // D and '..x1 ; f .x1 // > that
is, `x1 .h1 .x1 // < `x1 .f .x1 //implies that h1 .x1 / < f .x1 /, by the strict monotonicity
of `x1 . But this contradicts f h. Therefore, this contradiction demonstrates that it
must be that f D f .
t
u
Theorem 7.25 (Choquet). Assume that K is a convex subset of a real topological
vector space V. If K is compact, Hausdorff, and second countable, then the set ext K
of extreme points of K is a G -set and for every x0 2 K there exists a regular Borel
probability measure supported on ext K such that
Z
'.x0 / D
' d
ext K
for every ' 2 V .

Proof. The topological conditions on K imply, by Theorem 2.48, that there is a
metric d on K that induces the topology of K. For each n 2 N, consider the subset

1
1
:
Kn D x 2 K j x D .y C z/; y; z 2 K; d.y; z/
2
n
Each Kn is a closed subset of K and x 62 ext K if and only if there is some n 2 N for
which x 2 Kn . Hence,
ext K D
Knc :
n2N
That is, ext K is a G set; in particular, ext K is Borel measurable.

The topological conditions on K also imply that C.K/ is a separable Banach
space (Theorem 5.57); thus, Aff K is also separable and, therefore, there is a
countable subset fhn gn2N Aff K such that fhn gn2N is dense in the unit ball of
1
X
Aff K. Because khn k 1 for all n, the series
2n .hn /2 converges uniformly to
nD1
some f 2 C.K/. As the function .t/ D t2 has strictly positive second derivative
on R, the function is strictly convex (Proposition 4.46); therefore, because hn is
affine, the function hn D .hn /2 is strictly convex. Thus, f is a convex function.
264
7 Convexity
In fact, f itself is strictly convex. Indeed, if x; y 2 K are distinct, then there is an

n 2 N such that hn .x/ hn .y/ (because Aff K separates the points of K). Hence, for
all 2 .0; 1/,
hn . x C .1 /y/ < hn .x/2 C .1 /hn .y/2 ;
which implies that f . x C .1 /y/ < f .x/ C .1 /f .y/.
Fix x0 2 K and define p W CR .K/ ! R by p.g/ D g.x0 /, where g is the upper
envelope of g. Observe that p.g1 C g2 / p.g1 / C p.g2 / and p.rg/ D rp.g/, for all
g1 ; g2 ; g 2 CR .K/ and all r 2 0; 1/. Hence p is a sublinear functional on CR .K/.
Assume that h 2 Aff K. Because h is affine, continuous, and bounded,
Lemma 7.24 asserts that h D h. And if r 0, then rf D rf and, therefore,
h C rf D h C rf . On the other hand, if r < 0, then h C rf is concave and so
h C rf D h C rf , again by Lemma 7.24. But since f f and r < 0, we have
that h C rf D h C rf h C rf . Hence, if L D fh C rf j h 2 Aff K; r 2 Rg, then
L is a linear submanifold in CR .K/ and the function ! W L ! R defined by
!.h C rf / D h.x0 / C rf .x0 / is linear and satisfies !.h C rf / p.h C rf / for every
hCrf 2 L. Therefore, by the HahnBanach Extension Theorem, there is a linear map
W CR .K/ ! R such that jL D ! and .g/ p.g/ D g.x0 / for every g 2 CR .K/.
Suppose that g 2 CR .K/ is such that g 0 and let q D g. Then q 0 implies
that q 0 and so .q/ q.x0 / 0. That is, .g/ 0, and so is continuous,
by Proposition 6.49. The Riesz RepresentationR Theorem (Theorem 6.51) yields
a regular Borel measure such that .g/ D K g d for every g 2 CR .K/. The
constant function x 7! 1 is affine and so .1/ D 1 D .K/, which shows that is a
probability measure.
Integration preserves order; thus, if h 2 Aff K satisfies f h, then f h and
Z
Z
Z
f .x0 / D !.f / D f d f d h d D !.h/ D h.x0 / D h.x0 /:
K
Hence
Z
f d inffh.x0 / j h 2 Aff K; f hg D f .x0 /;
K
R
R
which proves that K f d D K f d.
R
R
Let E D fx 2 K j f .x/ D f .x/g. Because f f and K f d D K f d, we have that
R
c
c
Ec .f f / d D 0. Hence, as f f is strictly positive on E , .E / D 0. Therefore,
is supported on E.
If x 2 K n ext K, then there are distinct y; z 2 K such that x D 12 .y C z/. Because f
is strictly convex,

1
1
1
f .x/ < .f .x/ C f .y// f .x/ C f .y/ f
.y C z/ D f .x/;
2
2
2
where the final inequality above holds because f is concave. Hence, E ext K, which
proves that is supported on the Borel set ext K.
7.6 The Range of Non-Atomic Measures
265
Lastly, if ' 2 V , then 'jK is a continuous affine function K ! R, and so

Z
'.x0 / D 'jK .x0 / D !.'jK / D
ext K
' d;
t
u
7.6 The Range of Non-Atomic Measures

If one considers the counting measure on a finite set X, then the range R D
f.E/ j E 2 g of is also a finite set. On the other hand, if m is Lebesgue measure
on a; b, then the range of m is a continuum, namely the closed interval 0; b a.
From the point of view of convexity, Rm is convex, while R is not convex (assuming
X has at least two elements). The explanation for the convexity of Rm goes beyond
the particulars of Lebesgue measurerather, it is the property of Lebesgue measure
being non-atomic that is at play here.
Recall from Definition 3.55 that a measure on a measurable space .X; / is
non-atomic if there are no atoms for ; that is, there are no sets E 2 with the
property that .E/ > 0 and one of .E \ F/ or .E \ F c / is 0 for every F 2 .
Theorem 7.26 (Lyapunov). If is a finite non-atomic measure on a measurable
space .X; /, then the range of is the closed interval 0; .X/.
Proof. The Banach space L1 .X; ; / is the dual of L1 .X; ; / (Theorem 6.45),
and so L1 .X; ; / carries a weak -topology. If
I Df
2 L 1 .X; ; / j ess-ran
0; 1g and I D f P 2 L1 .X; ; / j
2 I g;
then I is a convex subset of the closed unit ball of L1 .X; ; /, which is
compact with respect to the weak -topology (Theorem 6.33). Therefore, the weak compactness of I will follow from showing that I is weak -closed. To this end,
suppose that 2 L 1 .X; ; / is such that P 62 I. Thus, there is a 2 ess-ran
such that 62 0; 1. Select an open set V C that contains but does not
intersect 0; 1; by definition of essential range, the measurable set E D 1 .V/ has
1
positive measure. The function g D .X/
E 2 L 1 .X; ; / induces a weak -open
neighbourhood U L1 .X; ; / of P given by

Z
U D 'P 2 L1 .X; ; / j 'g d D
X
1
.X/

' d 2 V :
Note
that ess-ran ' 6 0; 1 for every 'P 2 U; on the other hand, if # 2 I , then
R
#
d
2 0; .X/ and so #P 62 U. Hence, U \ I D ;, which proves that I c is a
E

weak -open set.
266
7 Convexity
R
Next, consider the map E W I ! R defined by E . P / D X d. Observe that E
is an affine function and that it is continuous with respect to the weak -topology
Ron I. Hence, the range K of E is a compact convex subset of R. Indeed, because
d 2 0; .X/ for all 2 I , and because 0 and .X/ are plainly elements of
X
the set K, we deduce from the convexity of K that K D E .I/ D 0; .X/.
Select a 2 K and consider the set Ia D f P 2 I j E . P / D ag. Because Ia D I \
E 1 .fag/ is the intersection of two weak -closed sets, the set Ia is a weak -closed
subset of the closed unit ball of L1 .X; ; /; hence, Ia is weak -compact. The set
Ia is also convex. Therefore, by the Kren-Milman Theorem, Ia has an extreme point
'.
P Assume, for theRmoment, that
R there is a measurable set E X for which 'P D P E ;
thus, a D E .'/
P D X E d D E d D .E/. The choice of a 2 K D 0; .X/ being
arbitrary would yield 0; .X/ f.E/ j E 2 g 0; .X/, thereby completing
the proof.
Therefore, it remains to show that there is a measurable set E X for which
'P D P E . Assume, on the contrary, that 'P 6D P E for every measurable set E. Thus, the
essential range of ' contains at least one point different from 0 and 1. Therefore,
there exists " > 0 such that ."; 1 "/ is an open neighbourhood in R of . If V is any
open set in C for which V \ R D ."; 1 "/, then, by definition of essential range,
the measurable set E D ' 1 .V/ has positive measure. Because is a non-atomic
measure, there is a measurable proper subset F of E such that both F and E n F
have positive measure. Likewise, there are measurable proper subsets G1 F and
G2 .E n F/ such that 0 < .G1 / < .F/ and 0 < .G2 / < .E n F/. In the 1dimensional real vector space R any two real numbers are linearly dependent. Thus,
there are ; 2 R not both zero such that
..G1 / .F// C ..E n F/ .G2 // D 0:
By multiplying the equation above by an appropriate constant, we may assume that
and have been scaled so as to satisfy jj < " and jj < ". Consider now the
measurable function

# D .G1 F / C EnF G2 ;
R
which has theR properties Rthat X # d D 0 and ' # 2 I . If '1 D ' C # and '2 D
' #, then X '1 d D X '2 d D a, which is to say that 'P1 ; 'P2 2 Ia . However,
'P 6D 'Pj for each j and 'P D 12 .'P1 C 'P2 / contradict the fact that 'P is an extreme point
of Ia . Therefore, it must be that 'P D P E for some E 2 .
t
u
Problems
7.27. Prove that a subset C of a vector space V is convex if and only if C contains
every convex combination of its elements.
7.28. Determine the extreme points of a closed interval a; b in R.
Problems
267
7.29. Let V be a topological vector space and assume that C V is an open convex
set. If W V ! R is a continuous linear transformation for which .v/ 6D 0 for every
v 2 C, then prove that .C/ is an open interval of R.
7.30. Prove that the dual space V of a normed vector space V is a Hausdorff
topological vector space in the weak -topology.
7.31. A cone in a finite-dimensional normed vector space V is a convex subset
C V such that v 2 C, for every 2 RC and v 2 C. Let C D f' 2 V j '.v/
0 ; 8 v 2 Cg.
1. Prove that C is a cone in the dual space V .
2. Prove that C D C.
7.32. Let C be a convex set and v 2 C. Prove that the following statements are
equivalent:
1. v is an extreme point of C;
2. if there are v1 ; v2 2 C and 2 .0; 1/ such that v D v1 C .1 /v2 , then v1 D
v2 D v.
7.33. Let C D 0; 1 0; 1 0; 1 R3 .
1. Determine all the faces of C.
2. Of the faces found, identify those that correspond to extreme points of C.
7.34. Let C be a convex set in a vector space V. Show that if F1 C is a face of C
and F2 F1 is a face of F1 , then F2 is a face of C.
7.35. Let fen gn2N denote the canonical coordinate vectors of `1 .N/. Prove that the
extreme points of the closed unit ball of `1 .N/ are precisely the vectors of the form
ei en , for some 2 R and n 2 N.
7.36. Assume that X is a compact Hausdorff space and is the -algebra of Borel
sets of X, and consider the set
P.X; / D f 2 M.X; / j .E/ 0; 8 E 2 ; .X/ D 1g;
which is a subset of the closed unit sphere of the Banach space M.X; / of regular
complex measures on .X; /.
1. Prove that P.X; / is weak -compact.
2. Prove that if x0 2 X, then the point-mass measure fx0 g is an extreme point of
P.X; /.
7.37. Prove that the vector space operations on V are continuous in the weak
topology, for every normed vector space V.
7.38. Suppose that V is a locally convex topological vector space and that K V
is a convex, compact Hausdorff space. Prove that if x0 2 K is an Rextreme point of
K and is a regular Borel probability measure such that '.x0 / D K ' d for every
' 2 V , then D fx0 g , a point-mass measure concentrated on fx0 g.
268
7 Convexity
7.39. If C is a convex set in a Banach space V, then an element v0 2 C is exposed

if there is a ' 2 V such that < '.v/ < <'.v0 /, for all v 2 C n fv0 g. Furthermore,
an exposed point v0 of C is strongly exposed if, for any sequence fvk gk2N C, the
sequence f<'.vk /gk2N converges to <'.v0 / only if fvk gk2N converges to v0 .
1. Prove that every exposed point of C is an extreme point of C.
2. Determine the strongly exposed points of `1 .N/.
3. Determine the strongly exposed points of `p .N/, for p > 1.
Part IV
Operator Theory
Chapter 8
Banach Space Operators
If Banach spaces are viewed as the metric analogue of the notion of vector space,
then the concept of an operator is correspondingly viewed as the continuous
analogue of the notion of a linear transformation of a vector space.
Recall that one can compose linear transformation T W V ! W and S W W ! Z to
produce linear transformation ST W V ! W defined by ST.v/ D S.Tv/, for v 2 V.
Similarly, one has the obvious notions of sum T1 C T2 and scalar multiplication T
(where 2 C) for linear transformations T; T1 ; T2 W V ! W. Of particular importance
are cases in which W D V, for in these cases the set of all linear transformations
V ! V has the structure of an associative algebra. In considering only those linear
transformations T W V ! V of a Banach space V that are continuous, it turns out that
the resulting set is also an associative algebra, known as a Banach algebra.
8.1 Examples of Operators

To this point we have encountered operators in the form of linear functionals and as
surjective isometries between certain Banach spaces. This section is a brief sampling
of operators of a more general type.
8.1.1 Matrices
The most accessible and most familiar examples of operators are to be found with
matrices. An m n matrix T D tij 1im; 1jn , where each tij 2 C, is a linear
transformation Cn ! Cm . The entries of the i-th column of T represent the entries in

DOI 10.1007/978-3-319-45633-1_8
271
272
8 Banach Space Operators
the vector Tei 2 Cm , where e1 ; : : : ; en 2 Cn are the canonical basis vectors (canonical
in the sense that ei is a column vector with 1 in the i-th entry and 0 in every
other entry). If Cn and Cm are endowed with norms, then T is an operator by
Proposition 6.3.
Another convenient way to analyse matrices is to consider Cn as a space of
functions. Specifically, let X D f1; 2; : : : ; ng and let be the power set of X.
Consider L1 .X; ; /, where is counting measure. Thus, each f 2 L1 .X; ; /
is a function f W X ! C in which
kf k D maxff .k/ j 1 k ng:
Now assume Y D f1; : : : ; mg, is the power set of Y, and is counting measure
on Y. As vector spaces, there are the obvious isomorphisms
Cn L1 .X; ; /
and
Cm L1 .Y; ; /:
Hence, an m n complex matrix T is an operator

T W L1 .X; ; / ! L1 .Y; ; /:
There is nothing special about the choice of the L1 -norm, and one could just as
easily consider a matrix T as an operator
0
T W Lp .X; ; / ! Lp .Y; ; /
for any choice of p; p0 2 1; 1/.
8.1.2 Operators on Finite-Dimensional Banach Spaces

If V and W are finite-dimensional vector spaces, then by choosing bases for V and
W each operator T will have a matrix representation with respect to these bases.
Specifically, if BV D fv1 ; : : : ; vn g and BW D fw1 ; : : : ; wm g are bases for V and W,
respectively, then, for each i D 1; : : : ; n,
Tvi D
n
X
tij wj
jD1
for some unique choice of ti1 ; : : : ; tim 2 C. The m n matrix TQ D tij 1im; 1jn is a
representation of T with respect to the bases BV and BW .
273
8.1.3 Finite-Rank Operators

An operator T 2 B.V/ is said to have finite rank if the range of T has finite
dimension. In such cases, the rank of T is defined to be the dimension of the range
of T. If 2 V and w 2 V are nonzero, then the operator F ;w on V defined by
F ;w .v/ D .w/v is of rank 1. Conversely, if F is a rank-1 operator, then there exist
nonzero 2 V and w 2 V such that F D F ;w . Furthermore, if T has finite rank
n
X
Fj . The
n 2 N, then there are rank-1 operators F1 ; : : : ; Fn 2 B.V/ such that T D
jD1
proofs of these facts are left to the reader (Exercise 8.60).
8.1.4 Integral Operators

Let p 2 1; 1/ and consider Banach spaces Lp .X; ; / and Lp .Y; ; / for some
finite measure spaces .X; ; / and .Y; ; /. Assume that W X Y ! C is a
bounded measurable function and that M D supfj.s; t/j j .s; t/ 2 X Yg. For each
s 2 Y, let s W X ! C be given by s .t/ D .s; t/. Thus, for each f 2 L p .X; ; / we
obtain a function .K f / W Y ! C defined by
Z
s f d:
.K f /.s/ D
X
Because js .t/j M for every .s; t/ 2 X Y, we deduce that

Z Z
jK f jp d D
X
js jp jf jp d d M p .Y/p
Y
jf jp d:
X
Hence, the function .K f / is p-integrable. Moreover, the map f 7! K f is plainly

linear, and so we obtain an operator K W Lp .X; ; / ! Lp .Y; ; / defined by
K.fP / D gP , where g D K f 2 L p .Y; ; /, of norm kKk M.
8.1.5 Multiplication Operators on C.X/

Let X be a compact Hausdorff space and fix a function
2 C.X/. The map
M W C.X/ ! C.X/ defined by M f D f is a bounded operator and it is clear
that kM k k k D maxfj .x/j j x 2 Xg. By taking f .x/ D 1 for all x 2 X, we see
that kM f k D k k kf k. Hence, kM k D k k.
274
8.1.6 Integral Operators on C.0; 1/

If W 0; 1 0; 1 ! C is continuous, then the linear transformation K W C.0; 1/ !
C.0; 1/ defined by
Z 1
Kf .s/ D
.s; t/f .t/ dt;
0
for f 2 C.0; 1/ is a linear transformation in which

Z 1
Z 1
kKf k D max .s; t/f .t/ dt max j.s; t/j

jf .t/j dt kk kf k;
s20;1
s;t20;1
where kk is the norm of as an element of C.0; 1 0; 1/. Operators of the form
K are called integral operators.
Another type of integral operator occurs with functions W 0; 1 0; 1 ! C that
are continuous on 0 t s 1. In particular, if .s; t/ D 1 for 0 t s 1 and
.s; t/ D 0 otherwise, then the corresponding integral operator K is given by
Z s
f .t/ dt;
Kf .s/ D
0
for f 2 C.0; 1/. This particular operator is known as the Volterra integral operator
and is frequently denoted by V rather than K.
8.1.7 Multiplication Operators on Lp .X; ; /, 1 p < 1

Assume that .X; ; / is a -finite measure space. If W X ! C is an essentially
bounded Borel measurable function and if fP 2 Lp .X; ; /, where p 1, then
Z
Z
j f j d
X
Z
.ess-sup / jf j d D .ess-sup /
jf jp d < 1
p
X
implies that
induces a linear transformation M W Lp .X; ; / ! Lp .X; ; /
via M .fP / D . P f / for all fP 2 Lp .X; ; /. The inequality above shows that M
is bounded and that
kM k k P k1 D ess-sup :
To show, conversely, that k P k1 kM k, assume first that is a simple function
such that P 0. Thus, if 2 C satisfies jj D ess-sup and if E D 1 .fg/, then
.E/ > 0. Because X is -finite, there S
is a sequence of measurable sets Fn 2 ,
each with finite measure, such that X D n Fn . Hence, there is at least one n 2 N for
which F D E \ Fn is a nonempty set of finite measure. Let f D .F/1=p F and note
that kfP k D 1 and kM fP k D jj D k P k1 . Hence, kM k D k P k1 if is a simple
function with P 0, and kM k D k P k1 is trivially true for P D 0.
275
Assume now that

is an arbitrary essential bounded function. By Proposition 5.52, there is a sequence f'gn2N of simple functions such that limn k P
'Pm k1 D 0. Because kM fP M'n fP k D kM 'n fP k k P 'Pn k1 kfP k for every fP 2
Lp .X; ; /, we deduce that kM M'n k k P 'Pn k1 . Hence, limn kM M'n k D
0 implies that kM k D limn kM'n k D limn k'Pn k1 D k P k1 .
8.1.8 Weighted Unilateral Shift Operators

Consider an element D ./k2N 2 `1 .N/ and the linear transformation S W
`p .N/ ! `p .N/, for p 2 1; 1, defined by
2
3
02 31
0
v1
6 1 v1 7
B6 v2 7C
6
7
B6 7C
S v D S B6 v 7C D 6 v 7 ; v 2 `p .N/:
4 2 25
@4 3 5A
::
::
:
:
The linear transformation S is clearly bounded of norm kS k D supk jk j, and S is
called a weighted unilateral shift operator.
If k D 1 for every k 2 N, then the resulting weighted unilateral shift operator S
is denoted simply by S and is called the unilateral shift operator on `p .N/. Note that
S is an isometry, but is not surjective.
8.1.9 Adjoint Operators

Theorem 8.1. If V and W are Banach spaces, and if T 2 B.V; W/, then there is a
unique operator T W W ! V with the property that
T .v/ D
.Tv/ ;
2 W ; v 2 V :
(8.1)
Furthermore, if T1 ; T2 ; T 2 B.V; W/ and 1 ; 2 2 C, then

1. kT k D kTk,
2. .1 T1 C 2 T2 / D 1 T1 C 2 T2 , and
3. if W D V, then .T1 T2 / D T2 T1 .
Proof. To prove the first assertion, observe that equation (8.1) above defines a linear
transformation of W into V such that
kT k D
sup
v2V; kvkD1
j .Tv/j k k kTk:
Thus, T is a bounded linear transformation of norm kT k kTk.
276
Conversely, for every " > 0 there is a unit vector v 2 V such that kTk < kTvk C ".
Further, Corollary 6.23 of the Hahn-Banach Extension Theorem shows that there is
a 2 W of norm k k D 1 such that .Tv/ D kTvk. Thus,
kTk < kTvk C " D j .Tv/j C " D kT .v/k C " kT k k k kvk C " D kT k C ":
Hence, kTk kT k.
To prove the uniqueness of T, suppose that T 0 2 B.W ; V / satisfies equation (8.1) for all 2 W and v 2 V. Thus, for a fixed v 2 V, 0 D .Tv/ .T 0 v/ D
.Tv T 0 v/ for all 2 W . By Corollary 6.23, this means that Tv T 0 v D 0.
Because the choice of v 2 V is arbitrary, we deduce that T 0 D T.
The proofs of the remaining statements are left as an exercise (Exercise 8.67).
t
u
Definition 8.2. The operator T is called the adjoint of T.
The explicit determination of the adjoint of a given operator T 2 B.V; W/ depends
to a certain extent on how well one understands the dual spaces of V and W. For
example, if p; q > 1 satisfy p1 C q1 D 1, then by identifying the dual of `p as `q
one sees that the adjoint of the weighted unilateral shift operator S W `p .N/ ! `p .N/
is the operator S W `q .N/ ! `q .N/ defined by
02
31
3
2
'1
1 '2
B6 '2 7C
B6 7C
7
6
.S / ' D S B6 ' 7C D 4 2 '3 5 ;
@4 3 5A
::
::
:
:
' 2 `q .N/:
(8.2)
Indeed, if v 2 `p .N/ and ' 2 `q .N/ D `p .N/ , then

'.S v/ D
1
X
kD1
k vk 'kC1 D
1
X

k 'kC1 vk D S ' .v/;
kD1
which shows that S has the form given by (8.2).
8.2 Mapping Properties of Operators

Because operators are continuous, the kernel ker T D fv 2 V j Tv D 0g of an operator
T W V ! W is necessarily closed, as ker T is the pre-image of the closed set f0g
in W. It is natural, therefore, to ask about the range ran T D fTv j v 2 Vg of an
operator. A simple sufficient condition for an operator to have closed range is that it
be bounded below in the sense of the following definition.
8.2 Mapping Properties of Operators
277
Definition 8.3. An operator T 2 B.V; W/ is bounded below if there exists > 0

such that
kvk kTvk;
8v 2 V :
The positive real number is called a lower bound for T.

Proposition 8.4. If V and W are Banach spaces, then every operator T W V ! W
that is bounded below will have closed range.
Proof. The proof is a variant of the proof of Proposition 6.9 and is, therefore, left as
an exercise (Exercise 8.61).
t
u
If one has a collection of operators T W V ! W in which V is a Banach space,
it may or may not be true that is a bounded set in the sense that there is a K > 0
such that kTk K for all T 2 . The following theorem indicates that if fails to
be bounded, then there is a dense G set G V for which the norms of Tv, for v 2 G
and T 2 , are arbitrarily large.
Theorem 8.5 (Principle of Uniform Boundedness). If B.V; W/ is a
nonempty set of operators, where V is a Banach space and W is a normed vector
space, then exactly one of the following two statements holds:
1. there is a K > 0 such that kTk K for every T 2 ; or
2. there is a dense G -set G V such that
sup kTvk D 1
8v 2 G:
T2
Proof. For each T 2 let fT W V ! R be the (nonlinear) continuous function fT .v/ D

kTvk, for v 2 V. By continuity, fT1 .0; n/ D fv 2 V j kTvk ng is a closed set and,
therefore,
Kn D
fT1 .0; n/
T2
is a closed subset of V for every n 2 N. Let Un D V n Kn , which is open. Either every

Un is dense in V or there is at least one n 2 N for which Un is not dense.
Case #1: Un is not dense for some n. In this case, fix such an n and choose v0 2
Kn D VnUn so that v0 lies outside the closure of Un . Thus, there is a > 0 with
B .v0 / \ Un D ;. Hence, if " D =2, then v0 C v 2 Kn for all v 2 V that satisfy
kvk ". That is, if kvk " and T 2 , then
kTvk D k.Tv0 C Tv / Tv0 k kT.v0 C v/k C kTvk 2n :
Hence,
kTk
2n
;
"
which proves that is a bounded set.
8T 2 ;
278
Case #2: Un is dense in V for every n 2 N.

defined by
GD
In this case, let G V be the G -set
Un :
n2N
By the Baire Category Theorem, G is a dense in V. By definition, if v 2 Un , then

there is a T 2 with kTvk > n. That is, if v 2 G, then supT2 kTvk D 1.
t
u
The next mapping property of operators to be considered originates in topology.
Definition 8.6. If X and Y are topological spaces and if f W X ! Y is a function,
then f is called an open map if f .U/ is open in Y for every open set U in X.
Using the terminology of open maps, one can say that if a continuous bijection
f W X ! Y is an open map, then f is a homeomorphism. Our aim is to achieve a
similar statement for linear operators, and key result in achieving this aim is the
following fundamental theorem.
Theorem 8.7 (Open Mapping Theorem).
Banach spaces is an open map.
Every surjective operator between
Proof. Assume that V and W are Banach spaces and that T W V ! W is a surjective
operator. Our first objective is to prove that there exists a > 0 such that
fw 2 W j kwk < g fTv j v 2 V and kvk < 1g :
(8.3)
Assuming that inclusion (8.3) holds, there is an open ball Q0 of radius in W such
that 0 2 Q0 T.U/, where U is the open unit ball in V. Because in a normed vector
space every open ball is obtained by translation and scaling of an open ball about
the origin, we deduce that for every w 2 T.U/ there is an open ball Qw of radius
w about w such that Qw T.U/, thereby establishing that T.U/ is open. Similarly,
an open ball P in V may be translated and scaled to the open unit ball U, which
shows that T.P/ is open. Thus, the proof of the theorem hinges on establishing the
inclusion (8.3).
SetSUk D fv 2 V j kvk < kg for each k 2 N and note that, because T is surjective,
W D k T.Uk / Thus, if Kk is the closure of each T.Uk /, then W D [k Kk . By the
Baire Category Theorem, there is at least one n 2 N for which the interior of Kn is
nonempty. Select w0 2 int Kn ; thus, there is a > 0 such that w0 C wQ 2 int Kn Kn D
T.Un / for every wQ 2 W with kwk
Q < . Let D =.4n/.
Choose w from the open ball B .0/ in W and set wQ D 12 kwk1 w. Because w0
and w0 C wQ are in the closure of T.Un /, there are sequences fuj gj and fzj gj in Un
with w0 D limj Tuj and w0 C wQ D limj Tzj . Thus, with vj D zj uj we have that fvj gj
is a sequence in U2n with wQ D limj Tvj . Therefore, there is some v0 2 U2n with
Q < 12 kwk1 ". Hence, if v D .2kwk 1 /v0 , then kvk < kwk= < 1 and
kTv0 wk
kTv wk < ".
8.3 Inversion of Operators
279
The previous paragraph shows that for each nonzero w 2 B .0/ there exists v 2 V
of norm kvk < kwk= < 1 such that kTv wk < ". We aim to replace kTv wk < "
with an equality. The key point is that is independent of ". Thus, assume that " > 0
also satisfies " < 1. Fix w 2 B .0/ and set "1 D 12 ". By the previous paragraph, there
exists v1 2 V with kv1 k < 1 kwk and kTv1 wk < "1 . Because w Tv1 2 B .0/,
we apply the previous paragraph again with "2 D 12 "1 to obtain a vector v2 2 V
with kv2 k < 1 kw Tv1 k and kTv2 .w Tv1 /k < "2 . Repetition of the argument
yields, inductively, a sequence fvj gj2N in V such that, for every j 2 N,
kvj k <
"
2j1
and
kw
k
X
Tvj k <
jD1
"
:
2k
k
1
X
X
vj gk is a Cauchy sequence in V with limit v D
vj of norm
Hence, f
jD1
jD1
kvk
1
X
kvj k kv1 k C "
jD1
1
X
1
< 1 C ":
j1
2
jD2
By continuity of T, Tv D w.
The arguments of the paragraph above demonstrate that for every w 2 B .0/ and
0 < " < 1 there is a vector v 2 V of norm kvk < 1 C " such that Tv D w. Suppose
that w 2 B .0/ is of norm kwk < 1C"

and let v 2 V be a vector of norm kvk < 1 C "
that satisfies Tv D .1 C "/w. Therefore, vQ D .1 C "/1 v is in the open unit ball of V
and T vQ D w. Because " is an arbitrary positive number in the interval .0; 1/,
fw 2 W j kwk < g D
[
w 2 W j kwk <
0<"<1
1C"

fTv j v 2 V; kvk < 1g;
t
u
8.3 Inversion of Operators

Our first goal in this section is to prove that if T W V ! W is a bijective operator, then
the inverse linear transformation T 1 W W ! V is also an operator.
Proposition 8.8. If V and W are Banach spaces, then the following statements are
equivalent for T 2 B.V; W/.
1. T is a bijection;
2. T is bounded below and has dense range.
280
Proof. (1) ) (2). Assume that T is a bijection. Because T is surjective, it is trivial

that T has dense range; further, the Open Mapping Theorem asserts that there is a
> 0 for which
fw 2 W j kwk < g fTv j v 2 V and kvk < 1g:
In other words, using that T is also injective, for every w 2 W such that kwk < there
is a unique v 2 V with kvk < 1 and Tv D w. The contrapositive of this statement is:
kTvk for every v 2 V such that kvk 1. Hence, if v 2 V is nonzero, then kvk1 v
is a unit vector and so kT.kvk1 v/k ; that is, kTvk kvk, which proves (2).
(2) ) (1). Assume that T is bounded below by > 0 and that T has dense range.
Because T is bounded below, T has closed range (Proposition 8.4) and is obviously
injective. But to have both dense range and closed range is to say that T is surjective.
Hence, T is a bijection.
t
u
Proposition 8.8 has an important consequence: if an operator is bijective, then its
inverse is also an operator.
Corollary 8.9. If V and W are Banach spaces and if T 2 B.V; W/ is a bijection,
then the linear transformation T 1 W W ! V is bounded.
Proof. Ordinary linear algebra demonstrates that the inverse function T 1 W W ! V
of T is a linear transformation. By Proposition 8.8, T is bounded below by some
> 0 (and ran T is dense). Thus, T 1 is bounded and kT 1 k 1 by the following
computation:
kT 1 wk D kT 1 .Tv/k D kvk 1 kwk;
where w D Tv.
t
u
Two criteria for the singularity of an operator are:

Corollary 8.10. If V and W are Banach spaces, then an operator T 2 B.V; W/ fails
to be invertible if
1. there is a sequence of unit vectors vk 2 V with infk kTvk k D 0, or
2. if the range of T is not dense in W.
Another consequence of the inversion theorem for operators is a useful result
called the Closed Graph Theorem. Recall from Exercise 5.101 that the Cartesian
product V W of Banach spaces V and W is a Banach space in its product topology.
Definition 8.11. If X and Y are topological spaces and if f W X ! Y is a function,
then the graph of f is the set G.f / D f.x; f .x// 2 X Y j x 2 Xg.
Theorem 8.12 (Closed Graph Theorem). If the graph of a linear transformation
T W V ! W of Banach spaces V and W is closed, then T is continuous.
Proof. Because T is linear, the graph G.T/ of T is a linear submanifold of V W.
The hypothesis that G.T/ is closed implies that G.T/ is itself a Banach space.
8.4 Idempotents and Complemented Subspaces
281
Let p1 and p2 denote the projection maps of V W onto V and W (the first and
second coordinates), respectively. Recall (or observe) that p1 and p2 are continuous
functions. In considering p1 restricted to the graph G.T/ of T we obtain a continuous
linear bijection s D p1jG.T/ W G.T/ ! V. By Corollary 8.9, the linear inverse s1 W
V ! G.T/ of s is continuous. Hence, so is the linear map T D p2 s1 .
t
u

Definition 8.13. A linear transformation E W V ! V of a vector space V is
idempotent if E2 D E.
If E is an idempotent, then so is 1 E and their product satisfies E.1 E/ D
.1 E/E D 0. If an idempotent E is continuous, then both ran E and ker E are
subspaces. The latter is clear; and to show the former, note that ran E is the kernel
of the continuous idempotent 1 E.
The most important of the algebraic properties of idempotent operators are
described in the next proposition.
Proposition 8.14. The following properties hold for idempotents E and F acting
on a Banach space V:
1. E C F is an idempotent if and only if EF D FE D 0;
2. E F is an idempotent if and only if EF D FE D F;
3. if EF D FE, then EF is idempotent with range ran E \ ran F and kernel ker E C
ker F.
Proof. The proof of (1) is left as an exercise (Exercise 8.64).
To prove (2), assume that EF D FE D F. Thus, .E F/2 D E2 EF FE C
2
F D E F, which shows that E F is an idempotent. Conversely, if E F is an
idempotent, then EF C FE D 2F, and so 2F EF FE D .1 E/F C F.1 E/ D 0.
Hence, ..1 E/ C F/2 D .1 E/2 C .1 E/F C F.1 E/ C F 2 D .1 E/ C F. Thus,
.1 E/ C F is idempotent. Therefore, assertion (1) implies that 1 E and F are each
idempotent and satisfy .1 E/F D F.1 E/ D 0, Hence, EF D FE D F.
To prove (3), note that EF D FE implies that .EF/2 D EF 2 E D E2 F D FE,
which implies that EF is idempotent. Because ran EF D ran FE, ran EF ran E, and
ran FE ran F, we deduce that ran EF ran E \ ran F. Conversely, if v 2 ran E \
ran F, then Ev D Fv D v and so v 2 ran EF. This proves that ran EF D ran E \ ran F.
If v 2 ker E and w 2 ker F, then EF.v C w/ D FEv C EFw D 0, which implies that
ker E C ker F ker EF. Conversely, if v 2 ker EF, then v D Fv C .1 F/v. Because
EFv D 0 only if Fv 2 ker E and F.1 E/v D 0 only if .1 F/v 2 ker F, we see that
EFv D 0 implies that v 2 ker E C ker F. Hence, ker EF ker E C ker F.
t
u
The range of an idempotent is a subspace (Exercise 8.65), and so it is not
surprising that idempotents are used to reflect algebraically certain geometric
aspects of Banach spaces.
282
Definition 8.15. Two subspaces M and N of a Banach space V are complementary

subspaces if
1. M \ N D f0g and
2. M C N D V.
The most immediate example of a complementary pair of subspaces is that
furnished by two mutually orthogonal subspaces M and N D M ? of a Hilbert
space H.
There is an intimate connection between complementary pairs of subspaces and
continuous linear idempotents.
Proposition 8.16. The following statements are equivalent for a pair of subspaces
M and N of a Banach space V:
1. M and N are a complementary pair;
2. there exists an idempotent E 2 B.V/ such that ran E D M and ker E D N.
Proof. Suppose that M and N are a complementary pair of subspaces of V. Thus, for
each v 2 V there are unique u 2 M and w 2 N such that v D u C v. Hence, the linear
transformation E W V ! V defined by E.u C w/ D u, for all u 2 M and w 2 N, is an
idempotent with ran E D M and ker E D N. Therefore, all that is required to show is
that E is continuous. We shall use the Closed Graph Theorem (Theorem 8.12) to do
so.
Let .v; z/ 2 V V be an element in the closure of the graph G.E/ of E. Thus, there
is a sequence f.vk ; Evk /gk2N in G.E/ such that limk kv vk k D limk kz Evk k D
0. Each vk is expressed uniquely as vk D uk C wk for some uk 2 M and wk 2 N.
Therefore, Evk D uk and so limk kz uk k D limk kz Evk k D 0; because M is closed,
this implies that z 2 M. Likewise, v z D limk ..uk C wk / uk / D limk wk 2 N, as N
is closed. Because N D ker E we have 0 D E.v z/ D Ev Ez D Ev z, which
implies that .v; z/ 2 G.E/. Therefore, the graph G.E/ of E is closed, and therefore
E is continuous.
Conversely, if E 2 B.V/ is an idempotent operator such that ran E D M and
ker E D N, then 1 D E C .1 E/ implies that M C N D V. If v 2 M \ N, then
v D Ev D .1 E/v and so v D Ev D E.1 E/v D 0v D 0.
t
u
Proposition 8.17. If M and N are a complementary pair of subspaces of a Banach
V, then the Banach spaces V=M and N are isomorphic.
Proof. By Proposition 8.16, there exists an idempotent E 2 B.V/ such that ran E D N
and ker E D M. Thus, the function T W V=M ! N given by T vP D Ev is a well-defined
linear bijection. The inverse S W N ! V=M of T is given by Sw D w,
P for w 2 N.
Because kSwk D kwk
P kwk, S is continuous; hence, so is T (by Corollary 8.9). u
t
If one has a single subspace M of a Banach space V at hand, it is natural to ask
whether M is part of a complementary pair or not. Such a subspace M is said to be
complemented.
283
Definition 8.18. A subspace M of a Banach space V is said to be complemented in

V if there is a subspace N of V, called the complement of M, such that M and N
form a complementary pair of subspaces of V.
Besides subspaces of Hilbert spaces, other examples of complemented subspaces
are finite-dimensional subspaces.
Proposition 8.19. Every finite-dimensional subspace of a Banach space is complemented.
Proof. Let fv1 ; : : : ; vn g be a basis of a finite-dimensional space M of a Banach
space V. As in the proof of Proposition 6.27, there are linear functionals 'Qj W M ! C
such that 'Qj .vk / D 0, for j 6D k, and 'Qj .vj / D 1, for each 1 j n. By the HahnBanach Extension Theorem, each of these linear functionals has an extension to a
n
\
ker 'j , which is a subspace of V
continuous linear function 'j W V ! C. Let N D
jD1
P
such that M \ N D f0g. Choose any v 2 V and let w D jD1 'j .v/vj 2 M. Consider
z D v w. Because 'j .z/ D 'j .v/ 'j .w/ D 'j .v/ 'j .v/ D 0, we deduce that z 2 N.
Thus, v D w C z 2 M C N.
t
u
However, as the next result shows, not every subspace of a Banach space need be
complemented.
Proposition 8.20. The subspace c0 .N/ has no complement in `1 .N/.
Proof. Write c0 for c0 .N/ and `1 for `1 .N/. Let X D .0; 1/ \ Q and let W N ! X
be a bijection. By Exercise 1.101, there is an uncountable family fX g 2 of infinite
subsets X X such that X \X 0 is a finite set for all ; 0 2 for which 0 6D . For
each 2 , let f D .f .n//n2N 2 `1 satisfy f .n/ D 0, if .n/ 62 X , and f .n/ D 1,
if .n/ 2 X . Because f f 0 62 c0 whenever 0 6D , the family ffP g 2 `1 =c0 is
uncountable.
Choose ' 2 .`1 =c0 / and, for every n 2 N, let Z' .n/ D ffP j j'.fP /j 1n g. Suppose
that fP 1 ; : : : ; fP m 2 Z' .n/. For each j, let

'.fP j /
j D sgn '.fP j / D
:
j'.fP j /j
Let v D
m
X
j f j 2 `1 . Because X j \ X k is a finite set for k 6D j,
jD1
kvk
P D inf fkv hk j h 2 c0 g D max jj j D 1:
1jm
Furthermore,
'.v/
P D
m
X
jD1
j '.fP j / D
m
X
jD1
j'.fP j /j
m
X
1
jD1
m
:
n
284
Thus, k'k D kvk

P k'k j'.v/j
P D '.v/
P m=n implies that m is bounded above
[ and,
hence, that Z' .n/ is a finite set. Therefore, the set Z' D ffP j '.fP / 6D 0g D
Z' .n/
is countable. Consequently,
n2N
Z'k is countable for every countable set f'k gk2N of
k2N
linear functionals on `1 =c0 .

Assume, contrary to what we aim to prove, that there exists a complement
N to c0 in `1 . By Proposition 8.17, there is a continuous linear isomorphism
T W `1 =c0 ! N. For each k 2 N let k 2 N be the linear functional for which
k .g/ D gk , for all g D .gn /n2N 2 N. Observe that if k .g/ D 0 for every k 2 N,
then necessarily g D 0. Now let 'k D k T, which is a linear functional on `1 =c0
for every k 2 N.
On the one hand, because T is an isomorphism, if w 2 `1 =c0 is such that
'k .w/ D 0 for every k 2 N, then necessarily w D 0. On the other hand, by the
argument of the previous paragraph, the set f 2 j 9 k 2 N such that 'k .fP / 6D 0g
is countable. Hence, because is uncountable, there is a 2 such that the
nonzero element fP of `1 =c0 satisfies 'k .fP / D 0 for every k 2 N. Therefore, this
contradiction implies that N and `1 =c0 cannot be isomorphic; that is, c0 is not
complemented in `1 .
t
u
Corollary 8.21. c0 .N/ is not the range of any idempotent E 2 B .`1 .N// :
Part of the definition for a pair of subspaces M and N to be a complementary pair
is that M C N D Vin other words, M C N is also closed. This leads naturally to
the question of whether M C N is closed for every two subspaces M and N such that
M \ N D f0g. The answer is yes in one of the most important cases of all:
Proposition 8.22. Suppose that M is a proper subspace of a Banach space V and
that v 2 V is nonzero and v 62 M. Then M C Span fvg is closed.
Proof. By Exercise 6.59, there is a ' 2 V such that '.v/ D 1 and '.w/ D 0, for
every w 2 M. Suppose that y 2 V is in the closure of M C Span fvg; thus, there
are sequences fwn gn2N in M and fn gn2N in C such that k.wn C n v/ yk ! 0.
Hence, n D '.wn C n v/ approaches '.y/ as n ! 1, which implies that fwn gn2N
converges to y '.y/v. Because M is closed, y '.y/v must therefore belong to M,
and so y D .y '.y/v/ C '.y/v belongs to M C Span fvg.
t
u
Corollary 8.23. If M and N are subspaces of V such that M \ N D f0g and N has
finite dimension, then M C N is a subspace.
8.5 Compact Operators

Definition 8.24. An operator K on a Banach space V is said to be compact if for
every sequence fvn gn2N of unit vectors vn 2 V there is a subsequence fKvnj gj2N of
fKvn gn2N such that fKvnj gj2N is convergent.
285
An alternate definition for compactness is as follows (Exercise 8.74): an operator

K on a Banach space V is compact if and only if the image of the closed unit sphere
of V under the operator K has compact closure.
If V has finite dimension, then Proposition 5.22 shows that the closed unit
ball of V is compact. Therefore, by Theorem 2.19, any sequence fvn gn2N of unit
vectors vn 2 V admits a convergent subsequence fvnj gj2N . Thus, using the continuity
of operators, for each T 2 B.V/ the sequence fTvn gn2N admits a convergent
subsequencenamely, fTvnj gj2N . That is, if v D limj vnj , then

lim Tvnj D T
j!1
lim vnj D Tv:
j!1
Hence, every operator T 2 B.V/ is compact if V has finite dimension.

A variant of the argument above is the following observation, which will be used
frequently.
Proposition 8.25. If V is a Banach space, 2 C is nonzero, and if T D 1, then T
is a compact operator only if V has finite dimension.
Proof. Assume that V has infinite dimension and choose 2 .0; 1/. By Proposition 5.21, there exists a sequence fvn gn2N of unit vectors vn 2 V such that fvn gn2N
does not admit any Cauchy subsequences. Thus, the sequence fTvn gn2N does
not admit any Cauchy subsequences and, hence, does not admit any convergent
subsequences. Therefore, T is compact only if V has finite dimension.
t
u
Compact operators exhibit the following algebraic and analytic features.
Proposition 8.26. The following operators are compact:
1.
2.
3.
4.
5.
K, for every 2 C and all compact operators K;

K1 C K2 , for all compact operators K1 and K2 ;
KT and TK, for all compact operators K and all T 2 B.V/;
F, for all F 2 B.V/ with finite-dimensional range;
K, for all operators K for which there exists a sequence fKn gn2N of compact
operators Kn such that limn kK Kn k D 0.
Proof. The proof of the first four assertions is left as an exercise (Exercise 8.70). To
prove the final statement, assume that there exists a sequence fKn gn2N of compact
operators Kn such that kK Kn k < 1n for very n 2 N. As K1 is compact, there is a
subsequence fv1;j gj of fvi gi such that fK1 v1;j gj is convergent. As K2 is compact, there
is a subsequence fv2;j gj of fv1;j gj such that fK2 v2;j gj and fK1 v2;j gj are convergent.
Inductively, for each n 2 N there is a sequence fvn;j gj such that
1. fvn;j gj is a subsequence of fvn1;j gj , and
2. fK` vn;j gj is convergent for all 1 ` n.
Now let " > 0. Choose p 2 N such that kK Kp k < ". Claim: fKvn;n gn is a Cauchy
sequence. Note that if n p, then fKp vn;n gn is a subsequence of the convergent
sequence fKp vp;n gn . Hence, fKp vn;n gn is convergent and, therefore, Cauchy. Thus,
286
kKvn;n Kvm;m k k.K Kp /vn;n k C kKp vn;n Kp vm;m k C k.K Kp /vm;m k

< 2" C kKp vn;n Kp vm;m k
Thus, fKvn;n gn is a Cauchy sequence and, hence, convergent. This proves that K is
compact.
t
u
Corollary 8.27. The set of compact operators acting on a Banach space V is a
subspace of B.V/.
Proof. The conclusion is a direct consequence of assertions (1), (2), and (5) of
Proposition 8.26.
u
t
If M is a subspace of V, then the range of the identity operator restricted to M
(as a map M ! V) obviously has closed range. The next proposition asserts that
the same is true for any perturbation 1 K of the identity operator 1 by a compact
operator K.
Proposition 8.28. If K is a compact operator acting on a Banach space V, and if
M V is a subspace, then
fw Kw j w 2 Mg
is a subspace of V.
Proof. Let M V be a subspace and assume that M \ ker.1 K/ D f0g. (The case
where M \ ker.1 K/ 6D f0g will be handled at the end of the proof.) Let
L D fw Kw j w 2 Mg
and suppose that y 2 L. We aim to prove that y 2 fw Kw j w 2 Mg. Thus, there is a
sequence of vectors yn 2 fw Kw j w 2 Mg with limit y; that is, there is a sequence of
vectors wn 2 M such that ky .wn Kwn /k ! 0. Assume that the sequence fwn gn2N
admits a subsequence of vectors wnk in which kwnk k ! 1. If this is the case, then
let vk 2 M denote the unit vector kwnk k1 wnk . The compactness of K implies that
the sequence fKvk gk2N admits a convergent subsequence fKvkj gj2N with limit, say,
z 2 V. Note that
kvkj Kvkj k D
1
kwnkj Kwnkj k:
kwnkj k
As j ! 1 we have kwnkj Kwnkj k ! kyk and kwnkj k ! 1. Therefore, fvkj gj2N

converges to z, which implies that z 2 M. However, we now have that z 2 M and z
Kz D limj .vkj Kvkj / D 0; that is, z 2 M \ ker.1 K/ D f0g. This is a contradiction
because 0 D z D limj vkj , where each kvkj k D 1. Therefore, it cannot happen that the
sequence fwn gn2N admits a subsequence of vectors wnk in which kwnk k ! 1.
287
Because there is a > 0 such that kwn k for all n 2 N, and because K
is a compact operator, the sequence fKwn gn2N admits a convergent subsequence
fKwnk gk2N with limit, say, x 2 V. Now since
Kwnk ! x
and
.wnk Kwnk / ! y ;
we conclude that wnk ! x C y, whence x C y 2 M. Thus, if w D x C y, then

.1 K/w D .1 K/.x C y/ D lim .wnk Kwnk / D y ;
k!1
which proves that the linear submanifold fw Kw j w 2 Mg is closed.

Suppose next that M \ ker.1 K/ 6D f0g. Thus, if F D M \ ker.1 K/, then F is
a finite-dimensional subspace. (If not, then F is an infinite-dimensional space upon
which the compact operator K acts as the identity, in contradiction to the fact that the
identity operator 1 is not compact.) Proposition 8.19 asserts that F is complemented
in M; hence, there is a subspace N M such that N \ F D f0g and M D N C F.
Consequently, each w 2 M has the form w1 C w0 , where w1 2 N and .1 K/w0 D 0.
Hence,
fw Kw j w 2 Mg D fw1 Kw1 j w1 2 Ng :
(8.4)
Since N \ ker.1 K/ D f0g, the linear submanifold in (8.4) is closed by the

arguments developed initially.
t
u
Corollary 8.29. If K is a compact operator on a Banach space V, then the range
of 1 K is closed.
For operators acting on finite-dimensional spaces, injectivity implies surjectivity.
Although this fails to be true in infinite-dimensional spaces (even for compact
operators), this property holds for compact perturbations of the identity operator.
Proposition 8.30. If K is a compact operator acting on a Banach space V such that
1 K is injective, then 1 K is surjective.
Proof. Assume, contrary to what we aim to prove, that 1 K is not a surjection. Set
M0 D V and let
Mn D f.1 K/n v j v 2 Vg D f.1 K/w j w 2 Mn1 g ;
8n 2 N:
Thus, M0 M1 M2 Mn MnC1 : : : is a descending sequence of

subspaces (by Proposition 8.28). This sequence is in fact proper, by the following
argument. Suppose for some n 2 N we have Mn D MnC1 . Because 1 K is not
surjective, there is a v0 2 Vnran .1K/. On the other hand, .1K/n v0 2 Mn D MnC1 ,
and so there is a vector w0 2 V such that .1 K/nC1 w0 D .1 K/v0 . That is,
.1 K/n .1 K/w0 v0 D 0 :
288
Because 1 K and, hence, .1 K/n are injective, we conclude that v0 D .1 K/w0 .

But this would place v0 in the range of 1 K, in contradiction to v0 2 Vnran .1 K/.
Hence, it must indeed be true that the sequence fMn gn2N[f0g is properly descending.
Moreover, if v 2 Mn , then v D .1 K/n w for some w 2 V; hence, Kv D .1 K/n Kw
(as K and 1 K commute). In other words, each Mn is invariant under K.
Let 2 .0; 1/ be fixed. For each n 2 N there is a vector in the quotient space
Mn1 =Mn of norm . Since < 1, this means that there is a unit vector vn 2 Mn1
such that kvn f k for all f 2 Mn . If j < k, then .1 K/vj 2 .1 K/.Mj1 / D Mj
and Kvk 2 Mk Mj . Thus, .1 K/vj C Kvk 2 Mj and so
kvj .1 K/vj C Kvk k D kKvj Kvk k:
Therefore, the sequence fKvn gn2N does not admit a Cauchy subsequence and,
hence, does not admit a convergent subsequence. This contradicts the fact that K
is compact. Thus, the original assumption that 1 K is not surjective cannot hold.
That is, 1 K is necessarily surjective.
t
u
An important feature of compact operators is that the adjoint K 2 B.V / is
compact if K 2 B.V/ is compact.
Proposition 8.31. If K 2 B.V/ is a compact operator, then K 2 B.V / is a compact
operator.
Proof. Let BV and BV denote the closed unit balls of V and V. Suppose that
f'n gn2N is a sequence in BV . Because BV is weak compact, by the BanachAlaoglu Theorem (Theorem 6.33), there is a ' 2 BV such that for every weak
open neighbourhood U of ' in BV and j 2 N there is an n j such that 'n 2 U.
(That is, ' is a weak limit point of f'n gn2N .)
Let " > 0 and fix j 2 N. Because K.BV / is compact, there are w1 ; : : : ; wm 2 K.BV /
such that
K.BV /
m
[
fw 2 V j kw wk k < "g :
kD1
Consider the weak open neighbourhood U BV of ' that is given by

U D f
2 BV j j .wk / '.wk /j < "; 1 k mg :
Because ' is a weak limit point of f'n gn2N , there is a nj j such that 'nj 2 U. If
v 2 Bv , then there is a 1 k m such that kKv wk k < ". Hence,
jK '.v/ K 'nj .v/j D j'.Kv/ 'nj .Kv/j
j'.Kv wk / 'nj .Kv wk /j
C j'.wk / 'nj .wk /j
< 3" :
8.6 Operator Algebra
289
Hence,
kK ' K'nj k D sup jK '.v/ K 'nj .v/j < 3" :
kvk1
As the choice of " > 0 and j 2 N are arbitrary, this proves that there is a subsequence
f'nj gj of f'n gn such that fK 'nj gj is convergent (to K '). Hence, K is a compact
operator.
t
u
Proposition 8.32. If K is a compact operator acting on a Banach space V such that
1 K is surjective, then 1 K is injective.
Proof. Because K is compact, so are K and K (Proposition 8.31). Because
1 K is surjective, the defect spectrum d .1 K/ is empty. But since d .1 K/ D
p .1K /, we conclude that 1K is injective. As K is compact, Proposition 8.30
implies 1 K is surjective. Therefore, the defect spectrum is 1 K , and so the
point spectrum of .1 K / is empty. In other words, 1 K is injective. The
restriction of 1K to the subspace V of V is precisely 1K; as 1K remains
injective on any smaller domain, 1 K is, therefore, an injective operator.
t
u

The notion of a Banach algebra has already been encountered in our discussion
of the Stone-Weierstrass Theorem. The algebraic basis for the concept of Banach
algebra is that of an associative algebra. Recall from Definition 5.24 that an algebra
a complex vector space A that has a product (or multiplication) such that, for all
a; b; c 2 A and all 2 C,
.a C b/c D .ac C bc/;
a.b C c/ D ab C ac;
a.bc/ D .ab/c;
and
. a/.b/ D a. b/ D .ab/:
Recall that, if ab D ba, for all a; b 2 A, then A is called an abelian algebra, and if
there is an element 1 2 A such that a1 D 1a D a, for every a 2 A, then A is said to
be a unital algebra and 1 is the multiplicative identity of A.
Definition 5.25 asserts that a Banach algebra is an algebra A together with a
norm k k on A such that
1. kxyk kxkkyk, for all x; y 2 A, and
2. A is a Banach space under the norm k k.
Recall that if A is a unital algebra, then A is a unital Banach algebra if k1k D 1.
290
A norm k k that satisfies the condition kxyk kxkkyk, for all x; y 2 A, is called a
submultiplicative norm.
Definition 8.33. A subset J of a Banach algebra A is an ideal of A if
1. J is a subspace of A, and
2. ax 2 J and xa 2 J, for all a 2 A and x 2 J.
If an ideal J of A satisfies J 6D f0g and J 6D A, then J is called a proper ideal of A.
Observe that, by definition, ideals are closed in the norm topology. We shall also
have need of the purely algebraic notion of ideal, which in the context of Banach
algebras are called algebraic ideals.
Definition 8.34. A subset I of a Banach algebra A is an algebraic ideal of A if
1. I is a linear submanifold of A, and
2. ax 2 I and xa 2 I, for all a 2 A and x 2 I.
If an algebraic ideal I of A satisfies I 6D f0g and I 6D A, then I is called a proper
algebraic ideal of A.
The set B.V/, where V is a normed vector space, is a unital associative algebra
whereby the product ST of operators S; T 2 B.V/ is simply compositionnamely,
ST.v/ D S.Tv/, for all v 2 Vand the multiplicative identity 1 of B.V/ is the
identity operator 1v D v, for all v 2 V.
Theorem 8.35. Assume that V is a Banach space.
1. The set B.V/ is a unital Banach algebra.
2. If V has infinite dimension, then the set K.V/ of all compact operators on V is a
proper ideal of B.V/.
3. If V has infinite dimension, then the set F.V/ of all finite-rank operators on V is
a proper algebraic ideal of K.V/.
4. The norm-closure F.V/ of F.V/ is an ideal of B.V/ and F.V/ J for every
nonzero ideal J of B.V/.
Proof. The norm on B.V/ is plainly submultiplicative and the norm of the identity
operator 1 is 1. By Proposition 6.6, B.V/ is a Banach space. Hence, B.V/ is a unital
Banach algebra.
Proposition 8.26 shows that K.V/ is an ideal of B.V/, while Proposition 8.25
indicates that 1 62 K.V/ if V has infinite dimension. Proposition 8.26 also shows
that the algebraic ideal F.V/ of B.V/ is a subset of K.V/. Therefore, because F.V/
is plainly an algebraic ideal of K.V/, we need only verify that there is a compact
operator on V of infinite rank.
Let M1 D V and select a unit vector v1 2 M1 . By Proposition 8.19, the
1-dimensional subspace Span fv1 g has a complement M2 . Select a unit vector
v2 2 M2 . In the Banach space M2 , the 1-dimensional subspace Span fv2 g has a
complement M3 . Note that M3 is also a complement to Span fv1 ; v2 g in V. Because
V has infinite dimension, proceeding by induction yields a properly descending
291
sequence of subspaces M1 M2 : : : and unit vectors vk 2 Mk such that vk 62 MkC1 .

Indeed, for each k 2 N the subspace
Nk D .Span fv1 ; : : : ; vk1 g/ C MkC1
is closed (by Proposition 8.23 or by noting that Nk is a complement to Span fvk g in
V) and vk 62 Nk . Therefore, by Exercise 6.59, there exists 'k 2 V of norm k'k k D 1
such that 'k .vk / 6D 0 and '.w/ D 0 for all w 2 Nk in particular, 'k .v` / D 0 for all
` 6D k.
Define, for each n 2 N, the finite-rank operator Kn on V given by
Kn v D
n
X
2k 'k .v/vk ;
kD1
for v 2 V. If m > n and v 2 V, then

k.Km Kn /vk
m
X
2k j'k .v/j kvk k D kvk
kDnC1
where fsj gj2N is the sequence of partial sums sj D

1
X
n
X
2k D kvk.sm sn /;
kDnC1
j
X
2k of the convergent series
kD1
2k . Hence, fKn gn2N is a Cauchy sequence of compact operators, and so this
kD1
sequence converges to some compact operator K. If k 2 N is fixed and n k, then

there is a nonzero k 2 R such that Kn vk D k vk and so Kvk D limn Kn vk D k vk .
This proves that the range of the compact operator K contains the countable set
fvk gk2N of linearly independent vectors, which implies that K has infinite rank.
It is straightforward to prove that the norm closure F.V/ of F.V/ is also an ideal
of K.V/and, hence, of B.V/. Suppose now that J is a nonzero ideal of B.V/.
Consider an arbitrary operator F 2 F.V/ of rank-1. By Exercise 8.60, there exist
(nonzero) w 2 V and ' 2 V such that F D F';w , where F';w v D '.v/w for all v 2 V.
Because J 6D f0g, there are nonzero T 2 J and u 2 V such that Tu 6D 0; and, by
Corollary 6.23, there exists 2 V such that .Tu/ D 1. Let S D F ;w TF';u ; because
J is an ideal, S is an element of J. However, Sv D '.v/w for all v 2 V implies
that S D F, and thus F 2 J. This proves that J contains every rank-1 operator.
Exercise 8.60 asserts that every finite rank operator is a (finite) sum of rank-1
operators, which implies that F.V/ J. Thus, because J is closed, F.V/ J.
t
u
Theorem 8.35 above shows that F.V/ is a minimal ideal of B.V/. Interestingly,
for certain separable Banach spaces V the ideal F.V/ is a proper ideal of K.V/,
while for other separable Banach spaces (such as separable Hilbert spaces) F.V/
coincides with K.V/. An example of a space V in which F.V/ is a proper ideal of
K.V/ is given by P. Enflo in [24].
292
8.7 The Spectrum of an Operator

Another instance whereby the theory of operators on Banach spaces takes a great
deal of inspiration from the theory of linear transformations is in the notion of
the spectrum of an operator, which is the algebraic analogue of the concept of
eigenvalue in linear algebra. However, owing to the possible infinite dimensionality
of the domain of an operator, the study of the spectrum in operator theory relies a
great deal more upon analysis than it does upon algebra. At the heart of the matter is
the development of a mathematical result for operator theory that essentially fulfills
the role that the fundamental theorem of algebra plays in linear algebra.
Earlier we determined that the linear inverse T 1 of a bijective linear operator
T W V ! W between Banach spaces is also an operator. In the case where W D V, the
inverse operator T 1 2 B.V/ satisfies T 1 T D TT 1 D 1, where 1 2 B.V/ denotes the
identity operator 1v D v, for v 2 V. This gives rise to the usual algebraic formulation
of inverse:
Definition 8.36. If V is a Banach space, then an operator T 2 B.V/ is invertible
in B.V/ if there exists an operator S 2 B.V/ such that ST D TS D 1 (the identity
operator 1v D v, for all v 2 V).
Of course, if such an operator S 2 B.V/ in which ST D TS D 1 exists, then S is
necessarily unique and is denoted by T 1 .
Recall, from linear algebra, the following theorem:
Theorem 8.37. If T is an operator acting on a finite-dimensional Banach space V,
then the following statements are equivalent for a complex number :
1. is an eigenvalue of T;
2. T 1 is not invertible in B.V/.
The eigenvalue problem in linear algebra is settled by the characteristic polynomial in the sense that the eigenvalues of a matrix T coincide precisely with
the roots of the characteristic polynomial cT .z/ D det.z1 T/. Therefore, it is
a consequence of the Fundamental Theorem of Algebra is that every complex
matrix has an eigenvalue. Thus, the eigenvalue problem for matrices is as much
a problem in algebra as it is in operator theory. The corresponding theorem in
operator theory states that every operator on a Banach space has nonempty spectrum
(Theorem 8.42). One proof of the Fundamental Theorem of Algebra is via complex
analysis (specifically, Liouvilles Theorem on bounded entire functions) and, not
surprisingly, it is by a very similar route that Theorem 8.42 is proved.
Lemma 8.38. If T 2 B.V/ and 2 C are such that j j > kTk, then .T 1/1 exists
and
lim k.T 1/1 k D 0:
j j!1
293
Proof. Observe that T 1 D .1 1 T/. By hypothesis, k 1 Tk < 1; thus,

1
X
kTk k
kD0
and so the series
1
X
j j
1
j j
D
;
1
j j

kTk
1 k Tk
j j
k T k converges to some S 2 B.V/ with kSk j jkTk
. Because
kD0

1
1
1
1 k
1 T
1 C T C C k T D 1 kC1 T kC1 ;

we conclude that S D .1 1 T/1 , which implies that T 1 is invertible. Moreover,
1
1
1
k.1 T/1 k
k.T 1/ k D
j j

j j
1
1 k 1 Tk
1
!
D
1
:
j j kTk
Hence, lim k.T 1/1 k D 0.
t
u
j j!1
Lemma 8.39. If T 2 B.V/ satisfies kTk < 1, then

k1 .1 T/1 k
kTk
:
1 kTk
Proof. Lemma 8.38 shows that .1 T/1 D
1
X
T k and k.1 T/1 k
1
.
1kTk
kD0
Further,

1
1

X
X

k1 .1 T/1 k D 1
Tk D
Tk

kD0
kD1

!
1

X

k
D T
T D k T.1 T/1 k

kD0
kTk k.1 T/1 k :

Hence, k1 .1 T/1 k
kTk
.
1kTk
t
u
Lemma 8.40. If T 2 B.V/ and 0 62 .T/, then .T 1/1 exists for all 2 C for
which j 0 j < k.T 0 1/1 k1 .
294
Proof.
For all 2 C suchthat j 0 j < k.T 0 1/1 k1 , Lemma 8.39 asserts that

1 C . 0 /.T 0 1/1 is invertible. However,

T 1 D T 0 1 . 0 /1 D .T 0 1/ 1 C . 0 /.T 0 1/1 ;
Thus, .T 1/1 exists for all 2 C for which j 0 j < k.T 0 1/1 k1 .
t
u
Lemma 8.41. Suppose that T 2 B.V/, ' 2 B.V/ , and 0 62 .T/. There is an " >
0 such that if D f 2 C j j 0 j < "g, then \ .T/ D ; and the function
f W ! C defined by f . / D ' .T 1/1 is differentiable at 0 .
Proof. Let " D k.T 0 1/1 k1 and D f 2 C j j 0j < "g. By Lemma 8.39,
\ .T/ D ;. Define f W ! C by f . / D ' .T 1/1 . For every 2 ,

.T 1/1 .T 0 /1 D . 0 / .T 1/1 .T 0 1/1 :
This yields a difference quotient:

1
.T 1/1 .T 0 /1 D .T 1/1 .T 0 1/1 :
0
This limit, as ! 0 , appears to be .T 0 1/2 . This is indeed true, since
k.T 0 1 C . 0 /1/1 .T 0 1/1 C . 0 /.T 0 1/2 k
2k.T 0 1/1 k3 j 0 j2 :
Hence,
f . / f . 0 /
D '..T 0 1/2 / ;
! 0
0
lim
which implies that f is differentiable at 0 2 C.
t
u
We are now prepared to prove every operator T on a Banach space V has a

nonempty compact spectrum.
Theorem 8.42. If V is a Banach space and T 2 B.V/, then .T/ is a nonempty
compact subset of f 2 C j jj kTkg.
Proof. Lemma 8.38 shows that .T/ f 2 C j jj kTkg; in addition, C n .T/ is
an open set, by Lemma 8.40. Thus, .T/ is bounded and closed, and hence .T/ is
compact. It remains to show that .T/ is nonempty.
Assume, contrary to what we aim to prove, that .T/ D ;. Choose any ' 2 B.V/
and let f W C ! C be defined by f . / D '..T 1/1 /. By Lemma 8.41, f is
differentiable at each 0 2 C. Hence, f is holomorphic on C. Lemma 8.38 shows
that lim k.T 1/1 k D 0; therefore, as ' is bounded, lim f . / D 0 as well.
j j!1
j j!1
295
On the compact set f 2 C j jj kTkg, the continuous map f is bounded, and on
the complement of this set f tends to 0 as j j ! 1. Therefore, f is bounded on its
entire domain C. But the only bounded entire functions are the constant functions.
Thus, there is a 2 C such that f . / D , for all 2 C. Because lim f . / D 0, the
!1
constant must in fact be zero. This, therefore, proves that '..T 1/1 / D 0 for all
' 2 B.V/ and 2 C. By the Hahn-Banach Theorem, this implies that .T 1/1 D
0, for each , which is impossible since 0 is not invertible. Therefore, it must be that
.T/ 6D ;.
t
u
The methods used to prove Theorem 8.42 also lead to the following continuity
assertion: namely, that the set-valued map T 7! .T/ is upper semicontinuous.
Proposition 8.43. If T 2 B.V/, then for every open set U C that contains .T/
there exists > 0 such that .S/ U for every S 2 B.V/ for which kS Tk < .
Proof. Assume that U is an open subset of C that contains the spectrum of T. By
Lemma 8.38, the function 7! k.T 1/1 k tends to 0 as j j ! 1. Hence, there
exists M > 0 such that j.T 1/1 k < M for all 62 U. Let D M 1 . If S 2 B.V/
satisfies kS Tk < , and if 62 U, then
k.S 1/ .T 1/k D kS Tk < D
1
1
<
:
M k.T 1/1 k
The invertibility of T 1 and this inequality above show, by Lemma 8.40, that
S 1 is invertible. Therefore, if 2 .S/, then cannot be exterior to U.
t
u
Although .T/ lies in the closed disc of radius kTk and centre 0 2 C, there could
be a smaller disc with the same centre that contains .T/. The radius of the smallest
of such discs is called the spectral radius.
Definition 8.44. The spectral radius of T 2 B.V/ is the quantity spr T defined by
spr T D max j j :
2.T/
Theorem 8.45. If T 2 B.V/, then lim kT n k1=n exists and

n
spr T D lim kT n k1=n :

n
Proof. If 2 .T/, then

.T n n 1/ D .T 1/
n
X
jD1
1
0
n
X
j1 T nj D @
j1 T nj A .T 1/ :
jD1
(8.5)
296
If T n n 1 were invertible, then by the equations above .T 1/ would have a left

and a right inverse and, thus, be invertible (Exercise 8.69). Therefore, T n n 1 is
not invertible. Thus, n 2 .T n / and j jn kT n k. Hence,
spr T lim inf kT n k1=n :
n
Let ; C be the open sets

D f 2 C j jj .spr T/ < 1g
and
D f 2 C j j j kTk < 1g :
Note that because spr T kTk.

Now, choose any ' 2 B.V/ and define f W ! C by

f . / D ' .1 T/1 :
If 2 , then .1 T/1 is a geometric series (Lemma 8.38); hence,
f . / D
1
X
n '.T n / ;
8 2 :
nD0
On the other hand, if 2 is nonzero, then

f ./ D

1
1
' . 1 T/1 :

By Lemma 8.41, f is differentiable at each nonzero point of . Thus, since 0 2

, f is holomorphic on the disc . By the uniqueness of the power series expansion
about the origin, we obtain
f ./ D
1
X
n '.T n / ;
8 2 :
nD0
Hence, lim j'. n T n /j D 0 for every 2 . Thus, for each 2 there is an M;' > 0
n
such that j'. n T n /j M;' for all n 2 N.

Now fix 2 and consider the family f n T n j n 2 Ng. Because B.V/ embeds
into B.V/ isometrically as a Banach space of operators on B.V/ , we consider the
family f n T n j n 2 Ng as acting on B.V/ in this waynamely,
n T n .'/ D '. n T n / ;
' 2 B.V/ :
By the Uniform Boundedness Principle (Theorem 8.5), either there is an R > 0 such
that k n T n k R for all n 2 N or supn k n T n 'k D 1 for a dense set of ' 2 B.V/ .
However, the latter situation cannot occur, since j'. n T n /j M;' for all n 2 N.
Hence, there is an R > 0 such that k n T n k R for all n 2 N. Thus,
1=n
jj kT n k1=n R :
297
Now choose a nonzero 2 . Therefore, spr T < 1=jj and

1=n
lim sup kT n k1=n
lim supn R
jj
1
:
jj
Hence,
lim sup kT n k1=n
n
inf
2nf0g
1
D spr T :
jj
This proves that

lim sup kT n k1=n spr T lim inf kT n k1=n :
n
That is, lim kT n k1=n exists and equals spr T.

n
t
u
The next result concerns the relationship between the spectra of ST and TS.
Proposition 8.46. .ST/ [ f0g D .TS/ [ f0g, for all S; T 2 B.V/.
Proof. If 1 ST is invertible, then .1 TS/.1 C TZS/ D .1 C TZS/.1 TS/ D 1,
where Z D .1 ST/1 . Interchanging the roles of S and T leads to: 1 ST is
invertible if and only if 1TS is invertible. Hence, if 6D 0, then ST 1 is invertible
if and only if TS 1 is invertible.
t
u
If X C, and if f is a polynomial, then f .X/ denotes the set ff ./ j 2 Xg.
Theorem 8.47 (Polynomial Spectral Mapping Theorem). If T 2 B.V/, then
f . .T// D .f .T//
for every complex polynomial f .
Proof. Let f be a complex polynomial and suppose that 2 .T/. Let g.t/ D f .t/
f . /. As g. / D 0, there is a polynomial h such that g.t/ D .t /h.t/. Hence,
f .T/f . /1 D g.T/ D .T 1/h.T/ D h.T/.T 1/. If f .T/f . /1 were invertible,
then these equations imply that T 1 has a left and right inverse, from which we
would conclude that T 1 is invertible (Exercise 8.69). Therefore, as T 1 is not
invertible, it must be that f .T/ f . /1 is not invertible. That is, f . / 2 .f .T//. This
proves that f . .T// .f .T//.
Conversely, suppose that ! 62 f . .T//. Thus, ! 6D f . /, for all 2 .T/. Let
h.t/ D f .t/ ! and factor h: h.t/ D .t !1 /n1 .t !m /nm . Since h. / 6D 0 for all
2 .T/, 6D !j for every j and 2 .T/. That is, !j 62 .T/ for each j. The
factorisation h leads to the following expression for h.T/:
f .T/ !1 D .T !1 1/n1 .T !m 1/nm :
Because !j 62 .N/ for each j, f .T/ !1 is a product of invertible operators and is,
hence, invertible. Thus, ! 62 .f .T//, and so .f .T// f . .T//.
t
u
298
8.8 Eigenvalues and Approximate Eigenvalues

By Proposition 8.8, an operator T on a Banach space V is invertible if and only if T
is bounded below and has dense range. This fact leads to the following definitions
for subsets of the spectrum.
Definition 8.48. Assume that T is an operator on a Banach space V.
1. The point spectrum of T is the set p .T/ of all 2 C for which the operator
T 1 on V is not injective.
2. The approximate point spectrum of T is the set ap .T/ of all 2 C for which the
operator T 1 on V is not bounded below.
3. The defect spectrum of T is the set d .T/ of all 2 C for which the operator
T 1 on V does not have dense range.
The elements of p .T/ are precisely the eigenvalues of Tthat is, the set of
2 C for which there is nonzero vector v (called an eigenvector) such that Tv D v.
The elements of ap .T/ are called approximate eigenvalues of T. If 2 ap .T/ and
if fvk gk2N is a sequence of unit vectors for which limk kTvk vk k D 0, then the
vectors vk are called approximate eigenvectors.
Proposition 8.49. If V is a Banach space and T 2 B.V/, then d .T/ D p .T /.
Proof. Suppose that 2 d .T/; thus, ran .T 1/ is a proper subspace of V. Let
W D V=ran .T 1/ and let q W V ! W be the canonical quotient map q.v/ D v,
P
for v 2 V. Since W is nonzero, there exists 2 W of norm k k D 1. Let ' 2 V
be given by ' D q and note that ran .T 1/ ker '. Thus, T ' D ', which
shows that 2 p .T /.
Conversely, assume that 2 p .T /. Thus, there exists ' 2 V of norm 1 such
that T ' D '; that is, '.Tv v/ D 0 for all v 2 V. If the range of T 1 were
dense, then we would conclude that '.w/ D 0, for all w 2 V, and this would imply
that ' D 0, by Corollary 6.23. But since ' 6D 0, the range of T 1 cannot be dense.
Hence, 2 d .T/.
t
u
Proposition 8.50. If V is a Banach space and T 2 B.V/, then ap .T/ is compact
and @ .T/ ap .T/.
Proof. To show that ap .T/ is closed, let 2 Cnap .T/. Thus, there is a > 0 such
that k.T 1/vk kvk, for every v 2 V. Let " D =2. If 2 C satisfies j j < ",
then, for every v 2 V,
kvk k.T 1/vk D k.T 1/v C . /vk k.T 1/vk C j j kvk ;
which implies that T 1 is bounded below by =2. Hence, the complement of
ap .T/ is open, which proves that ap .T/ is closed. As .T/ is bounded, ap .T/ is
compact.
Suppose that 2 @ .T/. Therefore, there is a sequence n 2 Cn .T/ such that
j n j ! 0. If there were a > 0 for which k.T n 1/1 k < for every n 2 N,
299
then it would be true that T 1 is invertible (by Lemma 8.40). Specifically, if n0 is

such that j n0 j < 1= , then the distance between the invertible operator T n0 1
and the operator T 1 satisfies
k.T 1/ .T n0 1/k D j n0 j < 1= < k.T n0 1/1 k1 :
This would imply that T 1 is invertible, contrary to the fact that 2 .T/.
Hence, it must be that fk.T n 1/1 kgn2N is an unbounded sequence; that is,
k.T n 1/1 k1 ! 0.
By definition of norm, for each n 2 N there is a unit vector wn 2 V such that
k.T n 1/1 k < k.T n 1/1 wn k C
1
:
n
Let
vn D
1
.T n 1/1 wn ;
k.T n 1/1 wn k
8n 2 N:
Thus, kvn k D 1 and

.T 1/vn D
1
wn C . n /vn :
k.T n 1/1 wn k
Hence,
k.T 1/vn k
1
C j n j ;
k.T n 1/1 k1 .1=n/
which converges to 0 as n ! 1. That is, 2 ap .T/.
t
u
Corollary 8.51. Every Banach space operator has an approximate eigenvalue.

We turn now to some sample calculations of the eigenvalue and approximate
eigenvalue problem.

Z t Z 1
Example 8.52. The eigenvalue problem
f .u/ dm.u/ dm.s/ D f .t/, for
0
almost all t 2 0; 1, in the Hilbert space L2 .0; 1; M; m/ admits countably many
solutions f k gk2N given by
k D
4
;
.2k 1/2
2
where each eigenvalue k has a corresponding eigenvector

p
fk .t/ D 2i sin.t= k /:
300
Proof. Consider the unit square 0; 1 0; 1 R2 and denote by L 2 and L2 ,

respectively, the spaces L 2 .0; 1; M; m/ and L2 .0; 1; M; m/. Define a linear
transformation K W L 2 ! L 2 by
K f .t/ D
Z t Z
0

f .s/ dm.s/ dm.t/ ;
f 2 L2:
The linear transformation K induces an operator K W L2 ! L2 in which kK fP k kfP k,

for all f 2 L 2 . The eigenvalue problem K fP D fP in L2 corresponds to an eigenvaluetype problem in L 2 : namely,
Z t Z
0

f .u/ dm.u/ dm.s/ D f .t/; for almost all t 2 0; 1 :
As the left-hand side of the equation above is twice differentiable almost everywhere, differentiation once leads to
Z 1
df
f .u/ dm.u/ D :
dt
t
Differentiation a second time yields
f D
d2 f
:
dt2
Evaluation of the two differential equations above at the boundary values for t gives
f .0/ D f 0 .1/ D 0.
Notice that hK gP ; gP i 0, for every gP 2 L2 , and so 0 hK fP ; fP i D kp
fP k2 implies that
p
00
it=
0. Therefore, the general solution of f C f D 0 is f .t/ D 1 e p C 2 eit= .
But f .0/ D f 0 .1/ D 0 implies that 2 D 1 . Hence, f .t/ D 2i sin.t= /. To satisfy
both f 6D 0 and f 0 .1/ D 0, it is necessary and sufficient that p1 D
2 .2k 1/ for some

k 2 N. Hence, the eigenvalues of K are
k D
4
;
.2k 1/2
2
k 2 N;
and the corresponding

eigenvectors fPk 2 L2 are determined by the functions fk .t/ D
p
2i sin.t= k /.
t
u
Recall that the essential range of an essentially bounded function
a measure space .X; ; / is the set ess-ran of all 2 C for which

f 1 .U/ > 0;
W X ! C on
for every neighbourhood U C of , and that Proposition 5.45 shows that

ess-ran
\
E2; .Ec /D0
.E/ :
301
Example 8.53. If .X; ; / is a measure space, then the approximate point spectrum of the multiplication operator M on Lp .X; ; /, where 1 p < 1, is given by
ap .M / D ess-ran :
Proof. Recall from Example 8.1.7 that if 2 L 1 .X; ; /, then M is defined
by M fP D Pf , for fP 2 Lp .X; ; / and has norm kM k ess-sup . If, in addition,
.X; ; / is a -finite measure space, then kM k D ess-sup ; however, this equality
is not required for the spectral calculation below, and so we need not assume
anything special of the measure space .X; ; /.
Suppose that 2 ess-ran . Choose any " > 0 and for each E 2 for which
.Ec / D 0 let FE E be the measurable subset of X defined by
FE D
1
.B" . // \ E:
Note that each FE is nonempty because 2
.E/. Now let
FD
FE
E
X; .XnE/D0
and consider the function f D F . Thus, .F/ > 0 and

k.M 1/fP kp D
j .t/ jp jf .t/jp d.t/ "p kfP kp :

F
Hence, M 1 is not bounded below, which proves that 2 ap .M /. Thus,

ess-ran ap .M /.
Conversely, suppose that 62 ess-ran . Thus, there exists at least one E 2 such
that .Ec / D 0 and 62 .E/, and soRthere is a > 0 such that j .t/ j > 0 for
all t 2 E. Since 0 D .Ec /, kfP kp D E jf jp d, for every f 2 L p .X; ; /. Further,
for any f 2 L p .X; ; /,
k.M 1/fP kp D
Z
j j jf j d
p
jf jp d D p kfP kp :
Thus, M 1 is bounded below, and so 62 ap .M /. Hence, ap .M / ess-ran ,

which completes the proof that ap .M / D ess-ran .
t
u
It can happen that M has no eigenvalues. For example, if X D 0; 1, D M,
D m, and .t/ D t, then M fP D fP implies that 0 D tf .t/ f .t/ for almost all t 2
0; 1, which can only happen if f .t/ D 0 almost everywhere. But, in this case, fP D 0
in Lp .X; ; /, thereby violating the requirement that an eigenvector be nonzero.
302
8.9 The Spectra and Invariant Subspaces of Compact

Operators
The nonzero points of the spectrum of a compact operator behave rather closely to
the eigenvalues of a matrix, as shown by the Fredholm Alternative below.
Theorem 8.54 (The Fredholm Aternative). Assume that K is a compact operator
on a Banach space V and that 2 C is nonzero. Then exactly one of the following
statements holds:
1. Kv D v for some nonzero v 2 V.
2. K 1 is invertible.
Proof. Propositions 8.30 and 8.32 show that 1 K is injective if and only if 1 K is
surjective. Therefore, if 2 C is nonzero, then replacing K by the compact operator
1
K and using the fact that .K/ D ap .K/[d .K/, we obtain 2 p .K/ or 62 .K/.

t
u
The Fredholm Alternative has implications for what properties the spectrum of a
compact operator may exhibit.
Theorem 8.55. If K is a compact operator acting on an infinite-dimensional
Banach space V, then
1.
2.
3.
4.
0 2 .K/,
each nonzero 2 .K/ is an eigenvalue of finite geometric multiplicity,
.K/ is a finite or countably infinite set, and
if .K/ is infinite, then 0 is the only limit point of .K/.
Proof. The fact that V has infinite dimension implies that K.V/ is a proper ideal
of B.V/ (Theorem 8.35), and so no compact operator can possess an inverse. Thus,
0 2 .K/.
Assume next that 2 .K/ is nonzero. Theorem 8.54 (Fredholm Alternative)
asserts that is an eigenvalue of K. By Proposition 8.25, the dimension of ker.K
1/ must be finite.
To show that .K/ is a finite or countably infinite set, it is enough to assume
that .K/ is infinite and to prove that 0 is the only limit point of .K/. Therefore,
assume that .K/ is infinite. Assume that 0 is not a limit point of .K/. There
exist, therefore, 1 > > 0 and a sequence f n gn .K/ (distinct elements) with the
property that j n j for every n 2 N. Each n is an eigenvalue of K; let vn 2 V
be corresponding eigenvectors of length 1. Fix m 2 N and let f1 ; : : : ; fm 2 C t be
polynomials such that fi . j / D 0, for j 6D i, and fi . i / D 1. (Such polynomials exist;
for example, one could use the Lagrange interpolation
1 them.) Therefore,
0to construct
m
m
m
X
X
X
if 1 ; : : : ; m 2 C satisfy
j vj D 0, then 0 D f .K/ @
j vj A D
j f . j /vj , for
jD1
jD1
jD1
every f 2 C t. In particular, if f D fi , then 0 D i vi , and so i D 0. This proves that

the sequence fvn gn consists of linearly independent vectors.
8.9 The Spectra and Invariant Subspaces of Compact Operators
303
For each m 2 N let Mm D Span fv1 ; : : : ; vm g, which yields an ascending sequence

M1 M2 of finite-dimensional subspaces of V. By Lemma 5.19, there is a
sequence of unit vectors wn 2 V such that wn 2 MnC1 , kwn uk for all u 2 Mn ,
and kwn wm k if m 6D n. Each subspace Mn is spanned by eigenvectors of K;
thus, K maps Mn back into itself. Therefore, K n 1 maps Mn into Mn1 , by the
following calculation:
1
0
n
n
n1
X
X
X
.K n 1/ @
j vj A D
j . j n /vj D
j . j n /vj :
jD1
jD1
jD1
Therefore, if ` < n, then u D Kw` .K n 1/wn 2 Mn1 . Hence,

kKwn Kw` k D k n wn C .K n 1/wn Kw` k D k n wn uk D j n j kwn 1n uk
j n j 2 > 0:
This means that fKwn gn does not admit a convergent subsequence, in contradiction
to the compactness of K. Therefore, it must be that 0 is a limit point of .K/. The
same argument shows that no nonzero 2 .K/ could possibly be a limit point of
.K/.
t
u
Definition 8.56. If S B.V/ is a nonempty subset, then a subspace M V is
invariant for S if Sv 2 M for all S 2 S and v 2 M. If an invariant subspace M for
S is neither f0g nor V, then M is a nontrivial invariant subspace for S . The set of
all subspaces M V that are invariant under S is denoted by Lat S and is called
the invariant-subspace lattice of S .
If S is a singelton set S D fTg, for some operator T on V, then the notation
Lat T is used in place of Lat S .
One of the most basic examples of invariant subspaces comes from eigenvectors
(if they exist): if for some nonzero v 2 V and 2 C one has Tv D v, then the
1-dimensional subspace Span fvg is T-invariant. Indeed, if is an eigenvalue of T,
then entire eigenspace ker.T 1/ is T-invariant. More generally, ker.T 1/k 2
Lat T for every k 2 N.
Definition 8.57. A subspace M V is hyperinvariant for an operator T 2 B.V/ if
M is an invariant subspace for fTg0 , where
fTg0 D fS 2 B.V/ j ST D TSg:
Because T 2 fTg0 , any hyperinvariant subspace for T is an invariant subspace
for T.
Proposition 8.58. If is an eigenvalue of T, then ker.T 1/ is hyperinvariant
for T.
t
u
304
The existence of nontrivial invariant subspaces depends on both the particular

Banach space at hand and on the size of the set S of operators. But even if S is a
singleton set S D fTg, the operator T may fail to admit any nontrivial invariant
subspaces [46]. It is still an open problem whether every operator acting on a
separable Hilbert space has nontrivial subspaces. A positive result, however, is the
following theorem of Lomonosov.
Theorem 8.59 (Lomonosov). If K 2 B.V/ is a nonzero compact operator on
an infinite-dimensional Banach space V, then K has a nontrivial hyperinvariant
subspace.
Proof. Because K 6D 0, we may assume without loss of generality that kKk D 1.
If K has an eigenvalue , then ker.K 1/ is hyperinvariant for K. If 6D 0, then
the compactness of K implies the finite dimensionality of ker.K 1/, which means
that this hyperinvariant subspace is indeed nontrivial. If D 0, then K 6D 0 implies
that ker K 6D V, and so again ker K is a nontrivial hyperinvariant subspace.
Assume, therefore, that K has no eigenvalues. Let v0 2 V be any vector for which
kKv0 k > 1. Observe that, because kKk D 1, the norm of v0 necessarily satisfies
kv0 k > 1. Let U be the open set U D fv 2 V j kv v0 k < 1g and for each S 2 fKg0
let WS be the open set WS D S1 .U/ D fv 2 V j kSv v0 k < 1g. If v 2 U, then
kv v0 k 1 and so kKv Kv0 k 1. Therefore, the condition kKv0 k > 1 implies
that the zero vector does not belong to K.U/ or its closure C D K.U/, which by
the compactness of the operator K is a compact subset of V. Hence, C is a compact
subset of V n f0g.
Because kv0 k > 1, the zero vector
[ neither belongs to U nor to
[any of the open
0
sets WS , for S 2 fKg . Thus, 0 62
WS . Suppose, though, that
WS D V n f0g.
S2fKg0
S2fKg0
Thus, fWS gS2fKg0 is an open cover of the compact set C and so there are S1 ; : : : ; Sn 2
n
[
WSj . The vector Kv0 is an element of K.U/ and, hence, of
fKg0 such that C
jD1
C; thus, Kv0 2 WSi1 for some i1 2 f1; : : : ; ng. Thus, Si1 .Kv0 / 2 U and KSi1 Kv0 2
C, which therefore implies that KSi1 Kv0 2 WSi2 , for some i2 2 f1; : : : ; ng, and that
Si2 KSi1 Kv0 2 U. Continuing inductively and using the fact that Sj K D KSj for all j
we deduce that for every m 2 N there is an im 2 f1; : : : ; ng such that
Sim Sim1 Si1 K m v0 D Sim KSim1 Si1 Kv0 2 U:
Now if D maxj kSj k and if SQ ik D 1 Sik , then
SQ im SQ im1 SQ i1 .K/m v0 2 U
and is of norm no greater than k.K/m k kv0 k. The fact that K has no eigenvalues
implies, by the Fredholm Alternative, that the spectrum of K is the singleton
set .K/ D f0g. Therefore, .K/ D f0g which implies, by the Spectral Radius
Formula, that 0 D limm k.K/m k1=m . Hence, limm k.K/m k D 0 which in turn
Problems
305
implies that 0 2 U because SQ im SQ im1 SQ i1 .K/m v0 2 U for all m. However,

[ this
WS D
conclusion is in contradiction to 0 62 U. Therefore, the assumption that
S2fKg0
V n f0g cannot be true, and so there is at least one nonzero vector v 2 V for which
v 62 WS for every S 2 fKg0 .
With this vector v define a subspace M of V by
M D fSv j S 2 fKg0 g:
Note that M is hyperinvariant for K and that 0 6D v 2 M; thus, the only question
is whether M 6D V. If it so happened that M D V, then fSv j S 2 fKg0 g would be
dense in V, and in particular there would exist some operator S 2 fKg0 for which
kSv v0 k < 1. However, this can never happen because v 62 WS for every S 2 fKg0 .
Hence, M 6D V.
t
u
Problems
8.60. Recall that an operator T 2 B.V/ has finite rank if the range of T has finite
dimension. In such cases, the rank of T is defined to be the dimension of the range
of T.
1. Prove that if 2 V and w 2 V are nonzero, then the operator F ;w on V defined
by F ;w .v/ D .w/v is of rank 1.
2. Prove that if F is a rank-1 operator, then there exist nonzero 2 V and w 2 V
such that F D F ;w .
3. Prove that if T has finite rank n 2 N, then there are rank-1 operators F1 ; : : : ; Fn 2
n
X
Fj .
B.V/ such that T D
jD1
8.61. Prove that if V and W are Banach spaces, then every operator T W V ! W that
is bounded below will have closed range.
8.62. If ' 2 L 1 .X; ; / has the property that j'.x/j D 1 for almost all x 2 X,
then show the multiplication M' W Lp .X; ; / ! Lp .X; ; /, for a fixed p 2 1; 1/,
defined by
M' .fP / D 'P fP
is an isometric operator.
8.63. Let H be a Hilbert space and suppose that ; k 2 H, k 2 N, satisfy
lim hk ; i D h; i ;
k!1
8 2 H :
306
1. Prove that the set S D fk j k 2 Ng is uniformly bounded.

2. Prove that kk lim infk kk k.
8.64. Prove that if E and F are idempotents acting on a Banach space V, then E C F
is an idempotent if and only if EF D FE D 0.
8.65. Prove that the range of an idempotent operator E on a Banach space V is
closed.
8.66. Suppose that V is a Banach space and that T 2 B.V/ is invertible. Prove that
kT 1 k kTk1 .
8.67. Prove the following properties of the adjoint T of an operator T 2 B.V; W/:
1.
2.
3.
4.
kT k D kTk,
.T1 C T2 / D T1 C T2 ,
if T is invertible, then T is invertible and .T /1 D .T 1 / , and
if W D V, then .T1 T2 / D T2 T1 .
8.68. Let V be a Banach space and consider V as a subspace of V . Define T to

be .T / . Prove that the restriction of T to V is T. That is, T jV D T.
8.69. Assume that R; S; T 2 B.V/ satisfy ST D TR D 1. Prove that T is invertible.
8.70. Prove that the following operators are compact:
1.
2.
3.
4.
K, for every 2 C and all compact operators K;

K1 C K2 , for all compact operators K1 and K2 ;
KT and TK, for all compact operators K and all T 2 B.V/;
F, for all F 2 B.V/ with finite-dimensional range;
8.71. Let X be a compact Hausdorff space, select 2 C.X/, and let T W C.X/ !
C.X/ be the operator of multiplication by : T f D f , for all f 2 C.X/. For each
2 C, let K D ft 2 X j .t/ 6D g (the closure in X).
1. Prove that ft 2 X j .t/ 6D g is an open set, for every 2 C.
2. Prove that if 2 p .T / if and only if K 6D X.
8.72. Assume that p; q 2 .1; 1/ satisfy 1=p C 1=q D 1. Define a linear transformation S W `p .N/ ! `p .N/ by
2
31
3
0
v1
6 1 v1 7
B6 v2 7C
6
B6 7C
7
6
B6 v3 7C
7
Sv D S B6 7C D 6 2 v2 7 ;
6 v 7
B6 v 7C
4 3 35
@4 4 5A
::
::
:
:
02
8v 2 `p .N/ ;
where fk gk2N C is a sequence for which supk jk j < 1.

1. Prove that S is an operator on `p .N/.
2. Determine an explicit form for the adjoint operator S on `q .N/.
Problems
307
3. If 2j D 0 and 2j1 D 1, for all j 2 N, then compute kSk.

8.73. Prove that k1 . 1 T/1 .S T/k 1, if S; T 2 B.V/ and 2 .S/ \ .T/c .
8.74. Prove that an operator K on a Banach space V is compact if and only if the
set fKv j v 2 V; kvk D 1g has compact closure.
8.75. Consider the unit square X D 0; 1 0; 1 R2 , and let 2 C.X/. Fix 1
p < 1 and let Lp denote Lp .0; 1; m/ (Lebesgue measure). Consider the integral
operator T 2 B.Lp /.
1. Prove that if is a polynomial, which means has the form
.t; s/ D
m X
n
X
ij ti sj ;
for some ij 2 C ;
iD0 jD0
then T is an operator of finite rank.

2. If 2 C.X/ is arbitrary, prove that for every " > 0 there is a polynomial q 2 Ct; s
such that j.t; s/ q.t; s/j < ", for all t; s 2 0; 1.
3. If 2 C.X/ is arbitrary, prove that for every " > 0 there is a finite rank operator
F 2 B.Lp / such that kT Fk < " .
Prove that if S; T 2 B.V/ satisfy ST D TS, then
8.76. Prove that if T 2 B.V/ satisfies TK D KT for every compact operator K, then
T D 1 for some 2 C.
8.77. Prove that if is an eigenvalue of T, then ker.T 1/ is hyperinvariant for T.
8.78. Suppose that T 2 B.V/. Prove that if V has finite dimension or is nonseparable, then T has a nontrivial hyperinvariant subspace.
8.79. Determine the hyperinvariant subspaces of the operator S on `2 .n/ given by
2
0 1 0
6
6 0 0 1 :::
6
6
S D 6 0 0 ::: :::
6
6:
:: ::
4 ::
: :
0 ::: ::: 0
3
0
:: 7
:7
7
7
:
07
7
7
15
0
Chapter 9
Spectral Theory in Banach Algebras
A Banach algebra is an associative algebra A with a norm k k such that A is a

Banach space in this norm, and the norm satisfies kxyk kxk kyk on all products
xy of elements x; y 2 A. Two examples of Banach algebras encountered to this point
are: (i) C.X/, the algebra of continuous functions on a compact Hausdorff space
X, and (ii) B.V/, the algebra of all bounded linear operators acting on V. A third
class of Banach algebras has also been encountered, but the algebraic structure of
these algebras has not yet been studied; these are the Banach algebras L1 .X; ; /,
whose algebraic structure is noted in Exercises 9.43 and 9.44.
In the case of the algebra B.V/, the notion of spectrum for Banach space
operators was analysed in detail in the Chapter 8, motivated by the notion of
eigenvalue from linear algebra. However, the initial properties of the spectrum of an
operator T carried out in Chapter 8 made little or no specific reference to the action
of T on the space V, but rather made reference to the behaviour of T as an element
of the algebra B.V/. Following this mode of thinking, the present chapter develops
spectral theory within the general context of Banach algebras, and concludes with a
few applications.
9.1 Banach Algebras

Recall from Definition 5.25 that a Banach algebra is a complex associative algebra
A together with a norm k k on A such that kxyk kxkkyk, for all x; y 2 A, and A is
a Banach space under the norm k k. Furthermore, if A is a unital algebra, then A is
a unital Banach algebra if k1k D 1. A Banach algebra A is abelian if xy D yx for
every x 2 A.
The submultiplicativity of the norm, which is to say that kxyk kxkkyk, for all
x; y 2 A, ensures that multiplication is a continuous map A A ! A (Exercise 9.45).
The natural maps between Banach algebras are homomorphisms.
DOI 10.1007/978-3-319-45633-1_9
309
310
9 Spectral Theory in Banach Algebras
Definition 9.1. If A and B are Banach algebras, then a function W A ! B is

1. a homomorphism, if is a bounded linear operator and .xy/ D .x/ .y/ for all
x; y 2 A, and is
2. an isomorphism, if is a bijective homomorphism.
If A and B are unital Banach algebras, then a homomorphism W A ! B for which
.1A / D 1B is called a unital homomorphism.
Definition 9.2. Suppose that A is a unital Banach algebra.
1. An element x 2 A is invertible if there is an element y 2 A such that xy D yx D 1.
2. The spectrum of an element x 2 A is the set .x/ defined by
.x/ D f 2 C j x 1 is not invertible in Ag:
3. The spectral radius of x is the quantity spr x defined by
spr x D sup j j :
2.x/
As with operators in the Banach algebra B.V/, if an element x 2 A is invertible,

then the element y 2 A for which xy D yx D 1 is necessarily unique and is denoted by
x1 and is called the inverse of x. Furthermore, the results about invertible operators
and their spectra established in Section 8.7 made use only of the fact that B.V/ is a
unital Banach algebrathat is, at no point was the action of an operator T on the
Banach space V considered. Therefore, the results of Section 8.7 carry over to unital
Banach algebras verbatim, including the following theorem.
Theorem 9.3. If A is a unital Banach algebra and if x; y 2 A, then
1.
2.
3.
4.
x is invertible if k1 xk < 1,
.x/ is a nonempty compact subset of f 2 C j j j kxkg,
.xy/ [ f0g D .yx/ [ f0g, and
spr x D limn kxn k1=n
Recall that a division ring is a unital ring in which every nonzero element has an
inverse. Fields are of course division rings, but there exist division rings, such as the
ring of quaternions, that are nonabelian.
Corollary 9.4 (Gelfand-Mazur). If a unital Banach algebra A is a division ring,
then A is isometrically isomorphic to the Banach algebra of complex numbers.
Proof. If x 2 A, then x has at least one spectral element 2 .x/. Hence, x 1
is not invertible. Because A is a division ring, it must be that x 1 D 0, and so
x D 1. Note that j j D kxk and that the map x 7! is an isometric isomorphism of
A with C.
t
u
Proposition 9.6 below adds to the list of Banach algebra properties mentioned
above in Theorem 9.3.
9.1 Banach Algebras
311
Definition 9.5. If A is a unital Banach algebra, then the general linear group of A
is the set GL.A/ consisting of all invertible elements of A.
Proposition 9.6. If A is a unital Banach algebra, then
1. GL.A/ is a multiplicative group,
2. GL.A/ is an open subset of A, and
3. the inverse map x 7! x1 is continuous on GL.A/.
Proof. The set GL.A/ is evidently a multiplicative group in the product of A, and
so we now prove that GL.A/ is an open set. Select x 2 GL.A/. If y 2 A satisfies
ky xk < kx1 k1 , then
k1 x1 yk D kx1 .x y/k kx1 k kx yk < 1;
and Proposition 9.3 shows that x1 y is invertible. Thus, y D xg for some g 2 GL.A/,
implying that y is invertible. Hence, GL.A/ is an open set.
Suppose now that x 2 GL.A/ and that " > 0 is given. Select > 0 such that
< kx1 k1 .1 C "1 kx1 k/1 and suppose that y 2 A satisfies ky xk < . Because
< kx1 k1 , the previous paragraph shows that y 2 GL.A/. Furthermore, since k1
x1 yk < 1,
y1 x D .x1 y/1 D
1
X
.1 x1 y/n ;
nD0
implying that
1
1
X
!
1
.1 x y/
1
Dx
1
1
X
C
.1 x1 y/n x1 :
nD0
nD1
Therefore,
ky1 x1 k kx1 k
1
X
kx1 kn kx ykn
nD1
D kx1 k2 kx yk
1
X
kx1 kn kx ykn
nD0
kx1 k2 kx yk
:
1 kx1 k kx yk
Thus, kx yk < < kx1 k1 .1 C "1 kx1 k/1 yields

kx1 k C "1 kx1 k2 < kx yk1
312
and
kx1 k2
1 kx1 k kx yk
< kx yk1 kx1 k D
:
"
kx yk
Hence, if ky xk < , then ky1 x1 k < ", thereby proving that inversion is
continuous.
u
t
Corollary 9.7. The group of invertible operators acting on a Banach space is an
open set.
Definition 9.8. If A is a Banach algebra, then a subset B A is a Banach subalgebra
of A if B is a Banach algebra in the norm and the algebra operations of A. If A is a
unital Banach algebra and a Banach subalgebra B of A contains the multiplicative
identity of A, then B is a unital Banach subalgebra of A.
If B is a unital Banach subalgebra of a unital Banach algebra A, and if x 2 B, then
there are two possibilities for the spectrum of x:
1. a spectrum denoted by A .x/ consisting all 2 C for which x 1 has no inverse
in A, and
2. a spectrum denoted by B .x/ consisting all 2 C for which x 1 has no inverse
in B.
Evidently, if x 1 is invertible in B, then it is invertible in A. Hence, C n B .x/
C n A .x/, implying that A .x/ B .x/. On the other hand, if x 2 B is invertible in A,
there it can happen that x1 62 B. A concrete case of this is provided by the following
example.
Example 9.9. Suppose that D is the open unit disc in C, and let A.D/ denote the
disc algebra
A.D/ D ff j f 2 C.D/ and fjD is an analytic functiong;
which is a unital Banach subalgebra of C.@D/. If f 2 A.D/ is given by f .z/ D z, for
z 2 D, then f is invertible in C.@D/, but not in A.D/.
Proof. The function inverse of the function f in the algebra C.@D/ is precisely the
function f 1 .z/ D z. However, the map z 7! z is not analytic on D, which implies
that f 1 62 A.D/.
t
u
The most general assertion relating the spectra A .x/ and B .x/ is given below.
Proposition 9.10 (Spectral Permanence). If B is a unital Banach subalgebra of
a unital Banach algebra A, then
A .x/ B .x/ and @B .x/ @A .x/
for every x 2 B.
9.1 Banach Algebras
313
Proof. The inclusion A .x/ B .x/ has already been noted. By definition of
boundary, if 2 @B .x/, then there is a sequence of distinct points k 2 .C n B .x//
such that j k j ! 0. Therefore, each x k 1 is invertible in A and
lim k.x 1/ .x k 1/k D 0:
k!1
Suppose that it is not true that 2 A .x/. Then the element x 1 is invertible in
A and is a limit of invertible elements x k 1. Therefore, using the continuity of
inversion in GL.A/ (Proposition 9.6), we have that
lim k.x 1/1 .x k 1/1 k D 0:
k!1
However, each .x k 1/1 is an element of B, and therefore .x 1/1 2 B also,

in contradiction of the fact that x 1 has no inverse in B. Therefore, it must be
that 2 A .x/. Furthermore, because D limk k and each k lies outside A .x/, we
deduce that 2 @A .x/.
t
u
It sometimes happens that equality in the spectral inclusion A .x/ B .x/ is
achieved.
Proposition 9.11. If B is a unital abelian Banach subalgebra of a unital Banach
algebra A, and if B has the property that
fy 2 A j yz D zy; for all z 2 Bg B;
then A .x/ D B .x/, for every x 2 B.
Proof. Let x 2 B; the inclusion A .x/ B .x/ is known from Proposition 9.10.
Therefore, suppose that 62 A .x/. Thus, for some y 2 A, .x 1/y D y.x 1/ D 1.
Thus, for any z in the abelian algebra B, we have z.x 1/ D .x 1/z, and so
yz D yz.x 1/y D y.x 1/zy D zy:
Therefore, by the hypothesis on B, y must be an element of B, which shows that
t
u
62 B .x/. Hence, B .x/ A .x/.
The algebra B that appears in Proposition 9.11 is called a maximal abelian
subalgebra of A.
Definition 9.12. A unital Banach subalgebra B of a unital Banach algebra A is
called a maximal abelian subalgebra of A if
1. B is abelian, and
2. fy 2 A j yz D zy; for all z 2 Bg B.
314
9.2 Ideals and Quotients

As was the case with B.V/, it is interest to consider two-sided ideals of Banach
algebras.
Definition 9.13. An ideal of a Banach algebra A is a subset J A such that
1. J is a subspace of A, and
2. ax 2 J and xa 2 J for all a 2 J and x 2 A.
Furthermore, if an ideal J of A is satisfies J 6D A, then J is said to be a proper ideal
of A.
As we had done in our study of the Banach algebra B.V/, if a linear submanifold
J of A satisfies ax 2 J and xa 2 J for all a 2 J and x 2 A, then J is called an algebraic
ideal of A.
Example 9.14. Suppose that X is a compact Hausdorff space, and let A be any
Banach subalgebra of C.X/. If Y X is a closed subset of X such that Y 6D X, then
the set
JY D ff 2 A j f .t/ D 0; for every t 2 Yg
is an ideal of A. Furthermore, if A is unital, then JY is a proper ideal of A.
Proof. If f 2 JY and g 2 A, and if t 2 Y, then fg.t/ D f .t/g.t/ D 0g.t/ D 0, which
implies that fg 2 JY . Likewise, 1 f1 C 2 f2 2 JY , for all f1 ; f2 2 JY and 1 ; 2 2 C.
Thus, JY is an algebraic ideal of A. To show that JY is closed, observe that, if f 2 JY ,
and if ffn gn2N is a sequence in JY converging in A to f , then for every t 2 Y and every
n 2 N, the inequality
jf .t/j D jf .t/ fn .t/j kf fn k
implies that f .t/ D 0. Hence, JY is an ideal of A.
If A is unital, then 1 62 JY because 1.t/ D 1 6D 0, for every t 2 Y. Hence, JY 6D A,
which implies that JY is proper.
t
u
Proposition 9.15. If A is a Banach algebra, and if J is an algebraic ideal of A, then
J is an ideal of A. Moreover, if A is unital and if the algebraic ideal J satisfies J A,
then k1 xk 1, for every x 2 J, and J is a proper ideal of A.
Proof. By continuity of multiplication, scalar multiplication, and sum, the set J is
also an algebraic ideal of A.
Suppose that A is unital and J A. If there were an element x 2 J such that
k1 xk < 1, then x would be invertible, and so 1 D x1 x would be an element of
J, implying that J D A, which is in contradiction to the hypothesis on A. Hence, it
must be that k1 xk 1, for every x 2 J. Therefore, it is also true that k1 yk 1
for every y 2 J. Hence, 1 62 J, which implies that J is a proper ideal.
t
u
315
If J is an ideal of a Banach algebra A, then A=J is a Banach space in the quotient

norm. The following result shows that the quotient space is also a Banach algebra.
Proposition 9.16. If J is an ideal of a Banach algebra A, then A=J is a Banach
P
algebra with respect to the quotient norm and multiplication defined by xP yP D .xy/.
Furthermore, if A is a unital Banach algebra and J 6D A, then A=J is a unital Banach
algebra.
P is a well-defined associative product on A=J is a
Proof. The fact that xP yP D .xy/
simple standard fact from ring theory. It is also clear that A=J is not just a ring, but
an associate algebra as well. By Proposition 5.18, A=J is a Banach space. Therefore,
the only issue left to address is the submultiplicativity of the quotient norm.
If x; y 2 A, then
P D inf kxy ak inf k.x a/.y b/k
k.xy/k
a2J

a;b2J

inf kx ak inf kxy bk
a2J
b2J
D kPxk kPyk:
Now if A is a unital Banach algebra with J 6D A, then 1P is a multiplicative identity
for the ring A=J. Furthermore, Theorem 9.3 states that a is invertible for every
element a 2 J that satisfies k1 ak < 1. However, as J 6D A, the ideal J contains
no invertible elements. Hence, k1 ak 1 for every a 2 J, and so
P D inf k1 ak 1;
1 k1k
a2J
implying that A is a unital Banach algebra.
t
u
Definition 9.17. An ideal N of a Banach algebra A is a maximal ideal if

1. N is a proper ideal of A, and
2. M D N, for every proper ideal M of A for which N M.
Example 9.18. If X is a compact Hausdorff space, and if x0 2 X, then the set
Jfx0 g D ff 2 C.X/ j f .x0 / D 0g
is a maximal ideal of C.X/.
Proof. The point set fx0 g is closed in X; thus, Example 9.14 shows that Jfx0 g is a
proper ideal of C.X/. Suppose that M is an ideal of C.X/ such that Jfx0 g M and
M Jfx0 g . Select h 2 M such that h 62 Jfx0 g ; therefore, h.x0 / 0.
On the Banach algebra C.X/=Jfx0 g , define a function W C.X/=Jfx0 g ! C by
.Pg/ D g.x0 /. If g1 ; g2 2 C.X/ satisfy gP 1 D gP 2 , then g1 .x0 / D g2 .x0 /, and so is a
316
well-defined function. Indeed, is a unital, surjective homomorphism. Let D

P Hence,
as is surjective, there is a kP 2 C.X/=Jfx0 g such that D '.k/.
1
;
.hP /
P
P D .k/ .
P h/
P D 1 D .1/;
.kP h/
which implies that 1P kP hP 2 ker ; that is, 1 kh 2 Jfx0 g . Since h 2 M and Jfx0 g M,
we deduce that 1 2 M , and so M D A. Therefore, the ideal Jfx0 g is maximal.
t
u
As with rings (such as the ring of even integers), some Banach algebras may fail
to possess a maximal ideal. However, in the case of unital Banach algebras, maximal
ideals always exist.
Proposition 9.19. If J is a proper ideal of a unital Banach algebra A, then there
exists a maximal ideal N of A such that J N.
Proof. Let S be the set of all proper ideals I of A such that J I and I 6D A. Define a
partial order on S by I K if I; K 2 S satisfy I K. Suppose that E is a linearly
[
[
ordered subset S. Define K D
I. Because E is linearly ordered, the set
I is
I2E
I2E
an algebraic ideal of A; hence, K is an ideal of A. If it were true that K 2 S, then K

would be an upper bound for E. Therefore, S would satisfy the hypotheses of Zorns
Lemma, implying that S has a maximal element N. Clearly a maximal element N
of S is a maximal ideal of A.
Therefore, all that is left to prove is that K 2 S. To this end, note that J K,
[and
so it remains to show that K 6D A. Assume, on the contrary, that K D A. Since
I is
I2E
dense in K, there exists I 2 E and y 2 I such that k1 yk < 1. On the other, because
I A, we have that k1 yk 1 (Proposition 9.15), in contradiction to k1 yk < 1.
Hence, it must be that K A.
t
u
9.3 Abelian Banach Algebras

The study of unital abelian Banach algebras is very closely related to the space of
maximal ideals in such an algebra, and maximal ideals arise from a special type of
homomorphism which is called a character.
Definition 9.20. If A is a unital abelian Banach algebra, then a unital homomorphism W A ! C is called a character of A.
Proposition 9.21. Suppose that A is a unital abelian Banach algebra. If is a
character of A, then ker is a maximal ideal of A. Conversely, if N is a maximal
ideal of A, then there exists a unique character W A ! C such that N D ker .
Proof. The kernel of every homomorphism of A is an ideal of A and the map
P W A= ker ! C in which .P
P x/ D .x/ is a well-defined isomorphism of Banach
317
algebras. Hence, A= ker is a field. If M A is an ideal of A that properly contains

ker , then there is an x 2 M such that xP is nonzero in A= ker . Because A= ker is
a field, there is a y 2 A such that yP D xP 1 . Thus, xy 1 2 ker M yields 1 2 M
(because xy 2 M), and so M D A. Hence, ker is a maximal ideal of A.
Conversely, suppose that N is a maximal ideal of A. Select a nonzero xP 2 A=N;
thus, x 62 N. Define
M0 D fax C y j a 2 A; y 2 Ng:
Evidently, M0 is an algebraic ideal of A, and so M D M0 is an ideal of A that contains
N. As x 2 M and x 62 N, the maximality of N implies that M be must A. Hence,
there are a 2 A and y 2 N such that k1 .ax C y/k < 1. Passing to the quotient and
P we deduce that k1P aP xP k < 1, and so aP xP is invertible. Hence, there
noting that yP D 0,
P The commutativity of A yields xP .PazP/ D .PazP/Px D 1,
P
exists z 2 A such that .PaxP /Pz D 1.
implying that xP is invertible. Therefore, A=N is a division ring. By the GelfandMazur Theorem, there is an isometric isomorphism # W A=N ! C. Thus, if
W A !
A=N is the canonical quotient homomorphism given by
.x/ D xP , for all x 2 A, then
D #
is a character on A with kernel ker D N.
To prove the uniqueness of , suppose that Q W A ! C is a character such that
ker Q D ker . For every x 2 A, the element x .x/1
Q
belongs to ker .
Q Therefore,
x .x/1
Q
also belongs to ker , which implies that
0 D .x .x/1/
Q
D .x/ .x/;
Q
and so Q D .
t
u
Proposition 9.21 identifies a bijective correspondence between maximal ideals

and characters, which leads to a bijective correspondence between the sets MA and
RA , defined below.
Definition 9.22. If A is a unital abelian Banach algebra, then
1. the maximal ideal space of A is the set
MA D fN A j N is a maximal ideal of Ag; and
2. the character space of A is the set
RA D f 2 A j is a unital homomorphism A ! Cg:
An important fact about the character space is that it is a compact Hausdorff
space.
Proposition 9.23. The character space RA of a unital abelian Banach algebra is a
weak -compact subset of the unit sphere of the dual space A of A.
318
Proof. If 2 RA , then .1/ D 1 and, thus, kk 1. Suppose it is true that kk > 1;
thus, there exists x 2 A with kxk D 1 and .x/ > 1. Let z D .x/1 x. Because
1D
the series
1
X
.x/
1
>
D kzk;
.x/ .x/
zn converges in A to an element y that satisfies y D z C zy. Thus, .y/ D
nD1
.z/ C .z/.y/ D 1 C .y/, implying that 0 D 1. Therefore, it must be that kk D 1,

which proves that RA is a subset of the unit sphere of A .
To prove that RA is weak -closed, suppose that # 2 A is in the weak -closure
of A . Thus, if x; y 2 A and " > 0, then the weak -open neighbourhood
U D f 2 A j j .z/ #.z/j < "; z 2 f1; x; y; xygg
intersects RA , and so there exists 2 RA such that 2 U. Thus,
j#.xy/ #.x/#.y/j j#.xy/ .xy/j C j.y/ #.y/j j.x/j
Cj.x/ #.x/j j#.y/j
< ".1 C kxk C k#k kyk/:
Hence, as " > 0 is arbitrary, #.xy/ D #.x/#.y/, which proves that # is a character
on A.
The unit ball of A is a weak -compact and Hausdorff; hence, the weak -closed
subset RA has these same two topological properties.
t
u
Let I W RA ! MA be defined by I./ D ker . By Proposition 9.21, I is a bijection.
Therefore, using Proposition 9.21, one can endow the maximal ideal space MA with
a topology T via
T D fU MA j I 1 .U/ is open in RA g;
where RA has the weak -topology.
Proposition 9.24. The maximal ideal space MA of a unital abelian Banach algebra
A is a compact Hausdorff space.
Proof. The map I W RA ! MA is a homeomorphism.
t
u
The importance of characters in spectral theory is demonstrated by the following

result.
Proposition 9.25. If A is a unital abelian Banach algebra, and if x 2 A, then
.x/ D f.x/ j 2 RA g:
319
Proof. If 2 RA and D .x/, then .x 1/ 2 ker . Because ker is a proper ideal

of A, it contains no invertible elements. Hence, 2 .x/.
Conversely, suppose that 2 .x/. Because x 1 is not invertible, there is a
maximal ideal N of A such that x 1 2 N (Exercise 9.51). By Proposition 9.21,
there exists a character W A ! C such that ker D N. Hence, .x/ D .
t
u
The following result is a basic application of Proposition 9.25 that is useful for
studying equations in Banach algebras. (See Exercise 9.48, for example.)
Proposition 9.26. If A is a unital Banach algebra, not necessarily abelian, and if
x; y 2 A satisfy xy D yx, then
.x y/ f j 2 .x/; 2 .y/g:
Proof. Suppose that there exists a maximal abelian subalgebra B of A that contains
both x and y. In this case, Proposition 9.11 shows that .z/ D B .z/, for every z 2 B.
In particular, with z D x y, this fact and Proposition 9.25 yield
.x y/ D B .x y/ D f.x y/ j 2 RB g D f.x/ .y/ j 2 RB g
f j 2 B .x/; 2 B .y/g
D f j 2 .x/; 2 .y/g:
Therefore, we need only establish the existence of the maximal abelian subalgebra B.
If A itself is abelian, then take B D A. Thus, assume that A is nonabelian and let B0
m X
n
X
be the closure of the unital abelian algebra of all elements of the form
kj xj yk ,
jD0 kD0
where each jk 2 C. Thus, B0 is a unital abelian Banach subalgebra of A. Let S

denote the set of all unital abelian Banach subalgebras C of A for which B0 C.
The set S is nonempty, as B0 2 S. Let be defined by C1 C2 , if C1 ; C2 2 S
satisfy C1 C2 . Suppose
that L is a linearly ordered subset of S, and define CQ to
[
be the closure of
C. By continuity of the algebraic operations, CQ 2 S and is an
C2L
upper bound for L. Hence, by Zorns Lemma, S has a maximal element B.

To show that B has the property fa 2 A j az D za; for all z 2 Bg B, select an
element a 2 A for which az D za, for every z 2 B. Thus, the closure BQ of the set
ff .a/z1 C z2 j z1 ; z2 2 B; f 2 Ctg is a unital abelian Banach subalgebra that contains
Q and because B is maximal in S, we deduce that BQ B and, in
B0 . Since, B B,
particular, that a 2 B.
t
u
We are now prepared for the major fundamental result of this section.
Theorem 9.27 (Gelfand). If A is a unital abelian Banach algebra, then there exists
a contractive unital homomorphism W A ! C.RA / such that, for every x 2 A,
320
1. .x/ D .x/, for all 2 RA ,

2. the range of the function .x/ is .x/, and
3. k .x/k D spr x.
\
N.
Furthermore, ker D
N2MA
Proof. The map W A ! C.RA / is plainly unital, linear, and multiplicative.

For x 2 A, the function A ! C given by ' 7! '.x/ is weak -continuous. Hence,
the map 7! .x/ determines a continuous function .x/ on the compact Hausdorff
space RA . Proposition 9.25 shows that the range of .x/ is .x/, and so the maximal
modulus of .x/, as varies through the character space RA , is the spectral
radius of x. Therefore, k .x/k D spr x kxk, which implies that k k 1. Note
that .x/ D 0 if and only if .x/ D 0 for all 2 RA ; that is, .x/ D 0 if and only
if x 2 ker for every 2 RA . That is, by Proposition 9.21, .x/ D 0 if and only if
x 2 N for every maximal ideal N of A.
t
u
Definition 9.28. If A is a unital abelian Banach algebra, then the Gelfand transform of A is the contractive unital homomorphism W A ! C.RA / described in
Theorem 9.27.
We turn now to some examples of character spaces, beginning with an example
that involves C.X/ itself.
Example 9.29. If X is a compact Hausdorff space, then the character space of C.X/
is homeomorphic to X.
Proof. Example 9.18 and Exercise 9.52 show that the maximal ideals of X are of
the form
Jfx0 g D ff 2 C.X/ j f .x0 / D 0g;
for x0 2 X. Therefore, by Proposition 9.21, the characters of C.X/ are all functions
x W C.X/ ! C of the form x .f / D f .x/, for f 2 C.X/, and x 2 X. Thus, we aim
to prove that the map W X ! RC.X/ , defined by .x/ D x , is a homeomorphism.
Since RC.X/ is a compact Hausdorff space, it is sufficient, by Proposition 2.9, to
prove that is continuous and bijective.
The surjectivity of
is apparent. If x1 ; x2 2 X are distinct, then by Urysohns
Lemma there is a function f 2 C.X/ with f .x1 / D 1 and f .x2 / D 0; hence, x1 and
x2 take different values at f , which shows that is injective.
To prove the continuity of , fix x0 2 X and consider a basic weak open
neighbourhood U of .x0 /:
U D f 2 RC.X/ j j.fj / x0 .fj /j < "j ; for all j D 1; : : : ; ng;
for some n 2 N, f1 ; : : : ; fn 2 C.X/, and positive real numbers "1 ; : : : ; "n . The equations
j.fj / x0 .fj /j D j.fj / fj .x0 /j imply that

1
321
.U/ D fx 2 X j jfj .x/ fj .x0 /j < "j ; for all j D 1; : : : ; ng

D
n
\

fj1 B"j .fj .x0 // ;
jD1
which is open in X by the continuity of the functions fj . Thus,

which completes the proof that is a homeomorphism.
is continuous,
t
u
The next example makes use of the Stone-Cech

compactification. Recall that
Cb .X/ denotes the unital abelian Banach algebra of bounded, continuous complexvalued functions on a locally compact Hausdorff space X.
Example 9.30. If X is a locally compact Hausdorff space, then the character space
of Cb .X/ is homeomorphic to X.
Proof. By Theorem 2.71, for every f 2 Cb .X/ there exists a unique fQ 2 C.X/ such
that fQ X .x/ D f .x/, for every x 2 X, where X W X ! X is a topological embedding
of X into X as a dense open subset, homeomorphic with X. Therefore, the function
W Cb .X/ ! C.X/ defined by

.f / D fQ is a unital isometric isomorphism of abelian
Banach algebras. Therefore, the character space of Cb .X/ is homeomorphic to the
character space of C.X/. Because X is compact, Example 9.29 shows that X
is homeomorphic to the character space of C.X/, and so the same is true of the
character space of Cb .X/ (Exercise 9.53).
t
u
Example 9.31. The maximal ideal space of `1 .N/ is homeomorphic to N.
Proof. By endowing N with the discrete topology, every function N ! C is
continuous. Thus, the bounded functions are precisely those given by the elements
of `. N/. Therefore, because `1 .N/ D Cb .N/, Example 9.30 shows that the character
space (and, hence, the maximal ideal space) of `1 .N/ is homeomorphic to N. u
t
Another algebraic structure of interest is the kernel of the Gelfand transform
Definition 9.32. The radical of a unital abelian Banach algebra is the ideal Rad A
of A defined by
Rad A D
N:
N2MA
If Rad A D f0g, then A is called a semisimple Banach algebra.

With this terminology:
Proposition 9.33. The Gelfand transform of a unital abelian Banach algebra A is
an injection if and only if A is semisimple.
If X is a locally compact Hausdorff space, then Cb .X/ is a semisimple Banach
algebra (Exercise 9.56). In contrast, the next example illustrates a situation where
the radical is so large as to be of co-dimension 1.
322
Example 9.34. Fix n 2 N such that n 2, and for each n-tuple .z0 ; z1 ; : : : ; zn1 / of
complex numbers, let T.z0 ; z1 ; : : : ; zn1 / denote the n n upper triangular Toeplitz
matrix
3
z0 z1 z2 zn1
6
: 7
6 0 z0 z1 : : : :: 7
7
6
7
6
T.z0 ; z1 ; : : : ; zn1 / D 6 0 0 : : : : : : z2 7 :
7
6
7
6 :
:: ::
4 ::
: : z1 5
2
0 ::: ::: 0
z0
If Tn D fT.z0 ; z1 ; : : : ; zn1 / j z0 ; z1 ; : : : ; zn1 2 Cg, then Tn is a unital abelian Banach

algebra of operators acting on the Hilbert space `2 .n/ such that:
1. the character space of Tn is given by RTn D fg, where
.T.z0 ; z1 ; : : : ; zn1 // D z0 ;
for all T.z0 ; z1 ; : : : ; zn1 / 2 Tn ;
2. Rad Tn D fT.0; z1 ; : : : ; zn1 / j z1 ; : : : ; zn1 2 Cg; and
3. Tn =Rad Tn is isometrically isomorphic to C.
Proof. If T.z0 ; z1 ; : : : ; zn1 / and T.w0 ; w1 ; : : : ; wn1 / are elements of Tn , then
T.z0 ; z1 ; : : : ; zn1 / C T.w0 ; w1 ; : : : ; wn1 / D T.z0 C w0 ; 1 ; : : : ; n1 /;
for some j 2 C, and
T.z0 ; z1 ; : : : ; zn1 /T.w0 ; w1 ; : : : ; wn1 / D T.w0 ; w1 ; : : : ; wn1 /T.z0 ; z1 ; : : : ; zn1 /
D T.z0 w0 ; !1 ; : : : ; !n1 /;
for some !j 2 C. Thus, the map W Tn ! C, in which .T.z0 ; z1 ; : : : ; zn1 // D z0 , is
linear, multiplicative, unital, and bounded. That is, is a character on Tn .
Gelfands Theorem asserts that the spectrum of each matrix in Tn is obtained
by evaluating all characters on Tn at that matrix. Therefore, since the spectrum of
T.z0 ; z1 ; : : : ; zn1 / is the singleton set fz0 g, any other character on Tn necessarily has
the same value as at each matrix T.z0 ; z1 ; : : : ; zn1 /. Hence, RTn D fg.
The radical Rad Tn of Tn is precisely the kernel of the Gelfand transform .
Gelfands Theorem shows that the norm of evaluated at T.z0 ; z1 ; : : : ; zn1 / is the
spectral radius of T.z0 ; z1 ; : : : ; zn1 /, namely jz0 j. Thus, T.z0 ; z1 ; : : : ; zn1 / 2 ker
if and only if z0 D 0. Hence, the equivalence class in Tn =Rad Tn of each matrix
T.z0 ; z1 ; : : : ; zn1 / is determined by z0 , and so Tn =Rad Tn is a division ring, implying
that Tn =Rad Tn is isometrically isomorphic to C, by the Gelfand-Mazur Theorem.
t
u
9.4 Absolutely Convergent Fourier Series
323
9.4 Absolutely Convergent Fourier Series

Any norm-closed subalgebra of C.X/ that contains the constant functions will be an
example of a unital abelian Banach algebra. A rather different type of example is
considered below, leading to a nontrivial application of the Gelfand theory of unital
abelian Banach algebras to absolutely convergent Fourier series.
If f W R ! C is a 2
-periodic continuous function, then the Fourier coefficients
fO .n/ of f are computed according to the usual fashion in the Hilbert space L2 .T/:
1
fO .n/ D
2
f .t/eint dt;
for n 2 Z.
Definition 9.35. The Fourier coefficients fO .n/ of a 2
-periodic continuous function
1
X
jfO .n/j < 1.
f W R ! C are summable if
1
The summability of the Fourier coefficients of continuous functions in a property

that is not enjoyed by all continuous 2
-periodic functions.
Proposition 9.36. If a 2
-periodic continuous function f W R ! C has summable
1
X
Fourier coefficients, then the series
fO .n/eint converges uniformly on R to f .
1
t
u
Let AC.T/ denote the set of all 2

-periodic continuous functions f W R ! C
for which the Fourier coefficients of f are summable. By Proposition 9.36, such
functions have absolutely convergent Fourier series. By the linearity of the Fouriercoefficient map f 7! fO .n/, the set AC.T/ is a vector space.
Proposition 9.37. AC.T/ is a unital abelian Banach algebra with respect to the
norm
kf k D
1
X
jfO .n/j;
1
for f 2 AC.T/.
Proof. By the triangle inequality in C, the formula for kf k is indeed a norm on
1
AC.T/. Thus, AC.T/
is a
normed vector space and the function W W AC.T/ ! ` .Z/
is a linear isometry. In fact this isometry is onto, for
defined by Wf D fO .n/
n2Z
if D .n /n2Z 2 `1 .Z/, then the series
1
X
1
n eint converges uniformly on R to a
continuous 2
-periodic function f such that fO .n/ D n for every n 2 Z. Hence, W
324
is an isometric isomorphism of AC.T/ and `1 .Z/, which proves that AC.T/ is a

Banach space.
To show that the norm is submultiplicative, select f ; g 2 AC.T/. Since the series
XX
fO .k/Og.m/ei.kCm/t
k2Z m2Z
converges uniformly to fg, we have that f .t/g.t/ D
XX
fO .k/Og.n k/eint and
n2Z k2Z
X
n2Z
O
j.fg/.n/j
XX
jfO .k/Og.n k/j D
n2Z k2Z
!
jfO .k/j
k2Z
!
jOg.n/j :
n2Z
Hence, kfgk kf k kgk, implying that AC.T/ is a Banach algebra. Lastly, the
O
constant function 1 2 AC.T/ has Fourier coefficient 1.n/
D 0 for every nonzero
O
n 2 Z and 1.0/ D 1. Thus, k1k D 1, which implies that AC.T/ is a unital Banach
algebra. The fact that AC.T/ is abelian is obvious.
t
u
With the Gelfand theory one can compute the spectra of absolutely convergent
Fourier series as follows.
Proposition 9.38. If f 2 AC.T/, then .f / D ff .t/ j t 2 Rg.
Proof. The main issue is to identify the character space RAC.T/ of AC.T/. To begin,
note that if t0 2 R, then the function t0 W AC.T/ ! C defined by t0 .f / D f .t0 / is
a character on AC.T/. Conversely, select any 2 RAC.T/ and let D .eit /. Thus,
j j kk keit k D 1. Because is a character, 1 D .eit / and so j 1 j 1 also.
Hence, j j D 1 and so D eit0 for some t0 2 R. If f 2 AC.T/, then define for each
m 2 N the trigonometric polynomial
fm .t/ D
m
X
n eint :
nDm
By linearity of , .fm / D fm .t0 / for every m. Because lim kf fm k D 0 implies (i)

m!1
that .fm / ! .f / and (ii) that fm ! f uniformly on R, we deduce that .f / D f .t0 /.
Hence, .f / D ff .t/ j t 2 Rg.
t
u
Absolutely convergent Fourier series are of classical interest and have been
studied by the methods of hard analysis. It is interesting, therefore, to note that
by viewing these series as elements of the Banach algebra AC.T/ one can obtain
some classically difficult results. A striking example is the following theorem of
N. Wiener.
Corollary 9.39 (Wiener). If f W R ! C is a continuous 2
-periodic function such
that the Fourier coefficients of f are summable, and if f .t/ 6D 0 for every t 2 R, then
the Fourier coefficients of 1=f are summable.
9.5 The Exponential Function
325
Proof. If f .t/ 6D 0 for every t 2 R, then 0 62 .f / and so f is invertible in AC.T/. If

g D f 1 , then g.t/f .t/ D 1 for every t 2 R. Therefore, g D 1=f , implying that the
continuous 2
-periodic function 1=f has summable Fourier coefficients.
t
u
9.5 The Exponential Function

The simplest of all Banach algebras is the unital abelian Banach algebra C. Among
the many holomorphic functions defined on all of C, the exponential function is one
of the most important. To begin this section, we shall see below how the exponential
function and its usual property of sending sums to products extends to the level of
Banach and abelian Banach algebras.
Proposition 9.40. If A is a unital Banach algebra, and if for each n 2 N the
n
X
1 k
function sn W A ! A is defined by sn .x/ D
x , where x0 D 1, then there exists a
k
kD0
continuous function exp W A ! A such that
1. fsn gn2N converges uniformly on fx 2 A j kxk rg to exp, for every r > 0, and
2. kexp.x/k ekxk , for all x 2 A.
Proof. Fix r > 0. The sequence ftn .r/gn2N of partial sums tn .r/ D
n
X
rk
kD0
1 k
X
r
of the
is a Cauchy sequence. If x 2 A satisfies kxk r, then by the

k
triangle inequality and the submultiplicativity of the norm,
series er D
kD0
ks` .x/ sk .x/k
`
`
X
X
kxkm
rm

D jt` .r/ tk .r/j :
m
m
mDkC1
mDkC1
Hence, the sequence fsn .x/gn2N is Cauchy and, therefore, converges in A to the
element that we denote by exp.x/. Observe that the inequalities above imply that
the converge of the functions sn W A ! A to exp is uniform on the closed ball of
radius r that is centred at 0 2 A, and that kexp.x/k ekxk .
t
u
Henceforth, we adopt the commonly used notation ex to denote exp.x/.
Proposition 9.41. If A is a unital Banach algebra, then
1.
2.
3.
4.
e0 D 1,
ex is invertible and .ex /1 D ex , for all x 2 A,
1
egxg D gex g1 , for all x 2 A and invertible g 2 A, and
exCy D ex ey , if xy D yx.
326
Proof. Suppose that xy D yx. Apply the binomial formula to obtain

k
k
k
k
k
X
X
X
1
1 X
m
1 `X 1 j
.x C y/m D
x` ym` D
x
y:
m
m
`.m `/
` jD0 j
mD0
mD0
`D0
`D0
Therefore, upon passing to limits and using the continuity of multiplication in the
algebra A, exCy D ex ey .
It is clear that e0 D 1. Suppose that x 2 A is arbitrary. The elements x and x
commute to produce 1 D e0 D ex ex , and so ex is invertible with inverse .ex /1 D
ex .
Lastly, note that sn .gxg1 / D gsn .x/g1 for all invertible g and all n 2 N, and so
gxg1
D gex g1 , for all x 2 A and invertible g 2 A.
t
u
e
The formula exCy D ex ey may fail if xy 6D yx.
The following proposition is of interest for one-parameter continuous groups
in A.
Proposition 9.42. If A is a unital Banach algebra and if x 2 A is fixed, then the
function W R ! A defined by .t/ D etx is a homomorphism of the additive group
.R; C/ into the multiplicative group GL.A/ (that is, .s C t/ D .s/ .t/, for all
s; t 2 R). Furthermore,
1
lim . .t/ .0// D x:
t!0 t
Proof. Proposition 9.41 shows that .s C t/ D .s/ .t/, as sx and tx commute. By
the power series expansion of etx we have
!
!
1
1 k2
X
1 X1 k k
1
t
. .t/ .0// D
t x 1 D xCt
xk ;
t
t kD0 k
k
kD2
1
and so lim . .t/ .0// D x.
t!0 t
t
u
Problems
9.43. Suppose that .X; ; / is a measure space.
1. Prove that L 1 .X; ; / is a complex, associative abelian algebra with respect
to the product 1 2 .t/ D 1 .t/ 2 .t/, for all 1 ; 2 2 L1 .X; ; / and t 2 X.
2. Prove that L1 .X; ; / is a Banach algebra, where the product on L1 .X; ; /
is induced by the product on L 1 .X; ; /.
3. Prove that the Banach algebra L1 .X; ; / is unital and abelian.
9.44. Determine the multiplicative identities of the unital abelian Banach algebras
`1 .N/ and `1 .n/.
Problems
327
9.45. Prove that if fxn gn2N and fyn gn2N are convergent sequences in a Banach
algebra A, and if x and y are the respective limits of these sequences, then the
sequence fxn yn gn2N is convergent to xy.
9.46. Assume that A is a unital Banach algebra, and that a1 ; : : : ; an 2 A. Prove the
following statements. (Suggestion: use induction.)
1. If ai aj D 0, for all i; j with i < j, then
0
1
n
n
X
[
@
aj A
.aj /:
jD1
jD1
2. If ai aj D 0, for all i; j with i 6D j, then

0
0
1
1
n
n
X
[
@
aj A n f0g D @ .aj /A n f0g:
jD1
jD1
9.47. Suppose that A is a unital Banach algebra, and that a 2 A. Define functions
à W A ! A and ra W A ! A by à .x/ D ax and ra .x/ D xa, for x 2 A.
1. Prove that à and ra are bounded linear operators on A.
2. In considering à and ra as elements in the unital Banach algebra B.A/, prove
that .à / .a/ and that .ra / .a/.
3. Prove that à and rb are commuting operators, for every b 2 A.
9.48 (Sylvester). Suppose that A is a unital Banach algebra, and that a; b 2 A. Prove
that if .a/ \ .b/ D ;, then for every y 2 A there exists a unique x 2 A such that
ax xb D y. (Suggestion: Consider the operator D on A defined by D D à rb .)
9.49. Suppose that A is a unital Banach algebra, and that a; b 2 A are such that
.b/ f 2 C j jj < g and .a/ f 2 C j jj > g, for some > 0.
1. Prove that, for each y 2 A, the series
2. Prove that if y 2 A and if x D
1
X
1
X
an1 ybn converges in A.
nD0
an1 ybn , then x is the unique solution to the
nD0
Sylvestre equation ax xb D y.
9.50. Suppose that A is a uniform algebra on a compact Hausdorff space X, and

suppose that f ; g 2 A are such that k1 C f C gk < 1. Prove the following assertions
for h D f C g:
1. k1 C <hk < 1;
2. there exists " > 0 such that 1 C "h.t/ 2 D, for every t 2 X;
3. h is invertible in A.
328
9.51. Prove that if A is a unital abelian Banach algebra, and if a 2 A is a nonzero

noninvertible element, then there exists a maximal ideal N of A such that a 2 N.
9.52. Prove that if N is a maximal ideal of C.X/, where X is a compact Hausdorff
space, then there exists a point x0 2 X such that N D ff 2 C.X/ j f .x0 / D 0g.
9.53. Suppose that A and B are unital abelian Banach algebras.
1. Prove that if W A ! B is a unital isometric isomorphism, then the map
W RB ! RA , defined by ./a D . .a//, for a 2 A, is a homeomorphism
of maximal ideal spaces.
2. Prove that if a continuous bijection W RB ! RA is a homeomorphism, then the
map W C.RA / ! C.RB /, defined by .f / D f , for f 2 C.RA /, is a unital
isometric isomorphism.
3. Prove that A and B are unitally isometrically isomorphic if and only if the
character spaces of A and B are homeomorphic.
9.54. An ideal K of a Banach algebra A is said to be an essential ideal if K 6D A and
K \ J 6D f0g for every nonzero ideal J of A. Prove that if X is a compact Hausdorff
space, the K is an essential ideal of C.X/ if and only if there is a dense open subset
U of X such that K D ff 2 C.X/ j f .x/ D 0 for every x 2 U c g.
9.55. If A is a unital abelian Banach algebra such that kx2 k D kxk2 for each x 2 A,
then prove that the Gelfand transform is an isometry.
9.56. Prove that if X is a locally compact Hausdorff space, then Cb .X/ is a
semisimple Banach algebra.
9.57. If A is a unital abelian Banach algebra, then prove that
Rad A D fx 2 A j .xy/ D f0g; for every y 2 Ag:
P for every unital abelian Banach algebra A.
9.58. Prove that Rad .A=Rad A/ D f0g,
9.59. Determine the character space of the disc algebra A.D/.
9.60. Prove that if a 2
-periodic continuous function f W R ! C has summable
1
X
Fourier coefficients, then the series
fO .n/eint converges uniformly on R to f .
1
9.61. Determine the matrix eS , where S is the matrix

3
2
0 1 0 0
6
:7
6 0 0 1 : : : :: 7
7
6
7
6
S D 6 0 0 ::: ::: 0 7:
7
6
6:
:: :: 7
4 ::
: : 15
0 ::: ::: 0 0
Chapter 10
Hilbert Space Operators
The theory of bounded linear operators acting on Hilbert spaces has a special place
in functional analysis. In many regards, it is a very specialised part of the subject;
yet, it is impressively rich in both theory and application. While the results already
established for operators acting on Banach spaces apply automatically to Hilbert
space operators, there is at least one aspect in which there is a slight but important
departure from Banach space operator theory, and it is the first issue addressed in
the present chapter.
10.1 Hilbert Space Duality and Adjoint Operators

Proposition 6.16 concerning the dual spaces of `p describes, in the case p D 2, an
isomorphism of Hilbert spaces. In fact, in the setting of abstract Hilbert space, all
linear functionals are determined by vectors, which is the content of the final Riesz
Representation Theorem in this book.
Theorem 10.1 (Riesz). Suppose that H is a Hilbert space. For every ' 2 H , there
corresponds a unique 2 H such that k'k D kk and './ D h; i, for all 2 H.
Conversely, for each 2 H, the formula './ D h; i, for 2 H, determines a
unique ' 2 H of norm k'k D kk.
Proof. First of all, suppose that ' 2 H . If ' D 0, then take D 0 and we obtain,
trivially, that './ D h; i for all 2 H. Therefore, assume that ' 6D 0. Because
H D ker ' .ker '/? , the subspaces ker ' and .ker '/? form a complementary pair,
and so the quotient space H= ker ' and the Hilbert space .ker '/? are isomorphic
P D './
(Proposition 8.17). Because the linear map 'P W H= ker ' ! C defined by '.
P /
is a well-defined linear isomorphism, the Banach space H= ker ' is 1-dimensional;
hence, so is .ker '/? .

DOI 10.1007/978-3-319-45633-1_10
329
330
10 Hilbert Space Operators
Thus, ker ' is spanned by some nonzero vector w 2 H. Let D '.w/kwk2 w.

Because each 2 H has the form D C w for some 2 ker ' and 2 C, we
obtain './ D '.w/ and
*
+
'.w/
hw; wi
h; i D C w;
w D '.w/
D './ :
jjwjj2
jjwjj2
Furthermore, j'./j D jh; ij kk kk implies that k'k kk. On the other hand,
if D kk1 , then kk D 1 and './ D kk. Hence, k'k D kk.
To show the uniqueness of , assume 0 2 H is another vector for which './ D
h; 0 i for every 2 H. Then, h; 0 i D h; i implies h; 0 i D 0 for every 2 H.
Such is the case for D 0 in particular; hence, k 0 k2 D 0.
The converse is clear.
t
u
As a consequence of the Riesz Representation Theorem, Hilbert spaces a selfdual. However, the natural notion of adjoint for Hilbert space operators is slightly
different from the adjoint of Banach space operators because, in the Hilbert space
setting, one needs to account for the fact that the inner product is not bilinear
rather, it is conjugate linear in the second variable. In this regard, the notion of
adjoint in Theorem 8.1 is not the same as the adjoint that is shown to exist in the
proposition below.
Proposition 10.2. If T is an operator on a Hilbert space H, then there is a unique
operator T on H such that
hT; i D h; T i ;
8 ; 2 H:
(10.1)
Proof. Fix 2 H and define ' W H ! C by ' ./ D hT; i for all 2 H. Because
' is plainly linear, the Riesz Representation Theorem (Theorem 10.1) states that
there is a unique vector, which we will denote by , such that
hT; i D ' ./ D h; i ;
8 2 H :
(10.2)
Thus, represents ' .

Now consider the function T W H ! H that sends each 2 H to 2 H. It
is straightforward to verify that T is a linear transformation. Therefore equation (10.2) becomes
hT; i D h; T i ;
8 ; 2 H :
(10.3)
The Riesz Representation Theorem states that k k D k' k. Hence, for 2 H,
331
kT k D k k D k' k D sup jhT; ij

kkD1
!
sup kTk kk D kTk kk:

kkD1
That is, T is bounded and kT k kTk.

All that remains now is to show that the transformation T that satisfies
equation (10.3) is unique. Suppose that S is an operator on H such that hT; i D
h; Si for all ; 2 H. Then for any 2 H, hT S; i D 0 for all 2 H. In
particular, T S is orthogonal to itself and so T S D 0. Therefore, S D T .
t
u
Definition 10.3. The operator T defined by equation (10.1) is called the adjoint of
the operator T and the map T 7! T on B.H/ is called the involution on B.H/.
A result related to Proposition 10.2 concerns bounded sesquilinear forms.
Definition 10.4. A function
if
W H H ! C is a bounded sesquilinear form on H
1. .1 1 C 2 2 ; / D 1 .1 ; / C 2 .2 ; /,

2. .; 1 1 C 2 2 / D 1 .; 1 / C 2 .; 2 /, and
3. there exists C > 0 such that j .; /j Ckk kk for all ; 2 H.
for all ; 1 ; 2 ; ; ; 2 2 H and 1 ; 2 ; 1 ; 2 2 C.
is a bounded sesquilinear form on H, then there exists a
unique operator T such that .; / D hT; i, for all ; 2 H.
Proof. As in the proof of Proposition 10.2, fixing 2 H and defining ' W H ! C
by ' ./ D .; /, for all 2 H, results in an element ' 2 H which, by
Theorem 10.1, has the form ' ./ D h; i for some unique 2 H. The function
T W H ! H that sends each 2 H to 2 H is a bounded linear operator, and yields
.; / D h; T i, for all ; 2 H. Hence, by Proposition 10.2, .; / D hT; i,
for all ; 2 H.
t
u
Proposition 10.6. The involution on B.H/ has the following properties for all
S; T 2 B.H/ and 2 C:
1.
2.
3.
4.
T D T;
. T/ D T ;
.S C T/ D S C T ;
.ST/ D T S .
Proof. To prove (1), the adjoint T of T is, by (10.3), the unique operator on
H for which hT #; i D h#; T iequivalently, hT #; i D h#; T ifor all
#; 2 H. In setting D and # D , it follows that h; T i D hT ; i, for all
; 2 H. Because h; T i D hT; i, if is fixed, then for every 2 H we have that
332
hT T ; i D 0. Thus, the vector T T is orthogonal to itself, which means

that T T D 0. As this is true for every , we deduce that T D T. The proofs
of the remaining algebraic statements are straightforward.
t
u
Through the use of the Hilbert space inner product, the norm of an operator can
be achieved as in Proposition 10.7 below. Another very important feature of the
norm of T 2 B.H/, which is also shown below, is that kTk2 D kT Tk.
Proposition 10.7. If T 2 B.H/, then
kTk D
sup
kkDkkD1
jhT; ij:
Furthermore, kT k D kTk and kT Tk D kTk2 .

Proof. If ! 2 H, then k!k D supfj'.!/j j ' 2 H ; k'k D 1g (Corollary 6.23). Thus,
k!k D supfjh!; ij j 2 H; kk D 1g, by the Riesz Representation Theorem. Hence,
kTk D sup kTk D
sup
kkD1
kkDkkD1
jhT; ij :
Since jhT; ij D jh; T ij, for all ; 2 H, we obtain kT k D kTk immediately.
The norm on B.H/ satisfies kSTk kSk kTk for every S; T 2 B.H/. Therefore,
kT Tk kT k kTk D kTk kTk D kTk2 :
Conversely, if ; 2 H are unit vectors, then the Cauchy-Schwarz inequality yields
jhT; ij2 kTk2 kk2 D hT; Ti D hT T; i kT Tk :
Thus, kTk2 kT Tk.
t
u
Proposition 10.8. If T 2 B.H/, then

1. ker T D .ran T /? and
2. ran T D .ker T /? .
Proof. For the proof of the first assertion, assume that 2 ker T. Any vector in ran T
has the form T , for some 2 H. Since h; T i D hT; i D 0, we conclude that
2 .ran T /? .
Conversely, suppose that 2 .ran T /? . Thus, for every 2 H, 0 D h; T i. In
particular, if D T, then 0 D h; T i D h; T Ti D hT; Ti D kTk2 , and so
2 ker T.
The proof of the second assertion is left as an exercise (Exercise 10.105).
t
u
Another aspect of the Hilbert space adjoint to be aware ofespecially in light of
what has come before in the study of operators on Banach spacesis that the defect
spectrum is characterised as follows:
333
Proposition 10.9. If T 2 B.H/, then 2 d .T/ if and only if 2 p .T /.

t
u
The following little fact is very useful in the study of Hilbert space operators, as
seen in Corollary 10.11 below.
Proposition 10.10 (Polarisation Identity). For every pair of vectors and in
an inner product space,
h; i D
1
.h C ; C i h ; i C h C i; C ii h i; ii/ :
4
Proof. A straightforward computation confirms the result.
t
u
Corollary 10.11. If S; T 2 B.H/ satisfy hS; i D hT; i for every 2 H, then

S D T.
Proof. By the Polarisation Identity, the hypothesis yields hS; i D hT; i for
every ; 2 H, and so h.S T/; i D 0 for every ; 2 H yields S T D 0 by
Proposition 10.7.
t
u
Isometries of Hilbert spaces are characterised by a succinct algebraic condition:
Proposition 10.12. An operator V 2 B.H/ is an isometry if and only if V V D 1.
Proof. If V is an isometry and 2 H, then kk2 D kVk2 D hV V; i implies that
h.1 V V/; i D 0 for every 2 H. By the Polarisation Identity, this implies that
V V D 1.
The converse is clear.
t
u
Definition 10.13. A surjective isometry U 2 B.H/ is called a unitary operator.
The following facts about unitary operators are readily confirmed.
Proposition 10.14. The following statements are equivalent for U 2 B.H/:
1.
2.
3.
4.
U is a unitary operator;
U U D UU D 1;
for some orthonormal basis f i gi of H, fU i gi is also an orthonormal basis;
for every orthonormal basis f i gi of H, fU i gi is also an orthonormal basis.
t
u
Isometries and unitaries are among the most fundamental of all Hilbert space
operators. Of equal importance are the projections.
Definition 10.15. An operator P 2 B.H/ is a projection if P D P D P2 .
In contrast to Banach spaces, where there can be subspaces without complements, every subspace M of a Hilbert space H has a complement, the most important
of which is M ? . Hence, Proposition 8.16 has the following form in Hilbert space:
334
Proposition 10.16. For every subspace M of H there exists a projection P 2 B.H/

such that M D ran P and M ? D ker P. Conversely, if P 2 B.H/ is a projection and if
M D ran P, then ran .1 P/ D ker P D M ? .
Proof. Because the subspaces M and M ? are a complementary pair, Proposition 8.16 shows that there exists an idempotent E 2 B.H/ such that ran E D M
and ker E D M ? . Such an idempotent E necessarily satisfies E D E because if
; 2 M M ? , then
hE. /; i D h; i D h ; E. /i:
That is, E D E and so the idempotent E is a projection.
Conversely, if P 2 B.H/ is a projection, then it is also an idempotent operator,
and so its range M is a subspace of H. If 2 M and 2 H, then P D and P D P
yield
h; .1 P/i D h; i h; Pi D h; i hP; i D h; i h; i D 0:
That is, .1 P/ 2 M ? and so ran .1 P/ M ? . And, if 2 M ? , then
k .1 P/ k2 D hP; P i D hP P; i D hP; i D 0;
which shows that D .1 P/ and, hence, that M ? ran .1 P/.
t
u
Proposition 8.14 on the algebraic features of idempotents translates into the

following proposition about projections.
Proposition 10.17. The following properties hold for projections P; Q 2 B.H/:
1. P C Q is a projection if and only if PQ D QP D 0;
2. P Q is a projection if and only if PQ D QP D Q;
3. if PQ D QP, then QP is a projection with range ran P \ ran Q and kernel ker P C
ker Q.
Another convenient relationship between T and T occurs with invariant subspaces.
Proposition 10.18. A subspace M H is invariant under an operator T 2 B.H/ if
and only if M ? is invariant under T .
Proof. If M is invariant under T and if 2 M ? , then for every 2 M we have that
0 D hT; i D h; T i, which yields T 2 M ? . That is, M ? is T -invariant. The
converse is proved in the obvious similar manner.
t
u
With regard to invariant subspaces, we have the following useful algebraic
characterisations.
Proposition 10.19. Suppose that T; P 2 B.H/, where P is a projection.
1. The range of P is invariant under T if and only if PTP D TP.
2. The range of P is invariant under T and T if and only if TP D PT.
10.2 Examples
335
Proof. For (1), if ran P is invariant under T, then for every 2 H, TP 2 ran P and
so P.ran P/ D TP. Conversely, if PTP D TP, for each 2 ran P we have T D
TP D PTP 2 ran P, and so ran P is T-invariant.
To prove (2), if ran P is invariant under T and T , then by (1) we have that PTP D
TP and PT P D T P. Taking the adjoint of the second equation give PTP D PT;
hence, TP D PTP D PT. Conversely, TP D PT implies PTP D P2 T D PT and, by
taking adjoints, that PT D T P, whence PT P D T P2 D T P.
t
u
10.2 Examples
10.2.1 Matrix Representations
Suppose that T 2 B.H/ and that B D f k gk2N is an orthonormal basis of a separable
Hilbert space H. If T D ij is the (infinite) matrix representation of T with respect
to the orthonormal basis B, then the .i; j/-entry of T is determined via
ij D hT j ; i i :
In using this for T in place of T we conclude that the .p; q/-entry of the matrix
representation of T with respect to B is hT q ; p i. Furthermore, hT q ; p i is
given by
hT q ; p i D h p ; T q i D hT p ; q i D qp :
Thus, the matrix representation T of T is determined by transposing T , the
matrix representation of T, and then conjugating each entry. In other words, T is
the conjugate transpose of T .
10.2.2 Multiplication Operators

Recall from Section 8.1 that if .X; ; / is a -finite measure space and if 2
L 1 .X; ; /, then
induces an operator M W L2 .X; ; / ! L2 .X; ; / via
M .fP / D . P f / for all fP 2 L2 .X; ; /. Because .X; ; / is -finite, Example 8.1.7
shows that kM k D ess-sup
To calculate the adjoint of M , note that if f ; g 2 L 2 .X; ; /, then
hfP ; M gP i D
Z
f g d D
f gd D hM fP ; gP i:
Hence, by the uniqueness of the Hilbert space adjoint, M D M .
336
10.2.3 The Bilateral Shift Operator

Let L2 .T/ denote the Hilbert space L2 .
;
; M; m/. Recall that an orthonormal
basis f P k gk2Z of L2 .T/ arises from the continuous functions k W
;
! C, k 2 Z,
whereby
1
k .t/ D p ei kt ;
2
t 2
;
:
If f 2 L 2 .T/, then fP 2 L2 .T/ has a Fourier series decomposition

fP D
fO .k/ P k ;
k2Z
which is convergent in L2 .T/ and where

fO .k/ D hfP ; P k i D
f .t/ei kt dm.t/ :
Let .t/ D eit . The adjoint of the multiplication operator M is M . Therefore,

P D .M / fP , then
because .t/ D eit , if gP D M fP and hh
gO .k/ D fO .k 1/
and
O
h.k/
D fO .k C 1/ ;
8k 2 Z :
Thus, if B D M , then B shifts the Fourier coefficients of fP 2 L2 .T/ forward by one

position, and its adjoint shifts the Fourier coefficients backwards one position. For
this reason, B is called the bilateral shift operator .
Put in terms of the action of B on the orthonormal basis of L2 .T/, we have
B P k D P kC1
and
B P k D P k1 ;
8k 2 Z:
One last observation: because BB D B B D 1, the bilateral shift operator B is a

unitary operator.
10.2.4 Toeplitz Operators

Continuing with the notation L2 .T/ for the Hilbert space L2 .
;
; M; m/, the
linear submanifold
n
o
H 2 .T/ D fP 2 L2 .T/ j fO .k/ D 0; 8 k < 0
10.2 Examples
337

is a subspace of L2 .T/ called the Hardy space. Let PH 2 2 B L2 .T/ denote the
projection whose range is the Hardy space H 2 .T/. If 2 L 1 .T/, then the Toeplitz
operator T on H 2 .T/ is the operator T D PH 2 M . The function is called the
symbol of the operator T . Because .M / D M , if fP ; gP 2 H 2 .T/, then
hT fP ; gP i D hPH 2 M fP ; gP i D hM fP ; PH 2 gP i D hM fP ; gP i
D hfP ; M gP i D hPH 2 fP ; M gP i D hfP ; PH 2 M gP i
D hfP ; T gP i:
Thus, .T / D T .
In particular, if .t/ D eit and S D T , then S shifts each element of the
orthonormal basis f P k gk0 forward one position, from P k to P kC1 . In this case the
Toeplitz operator S is called the unilateral shift operator.
With respect to the orthonormal basis f P k gk0 of H 2 .T/, the matrix representation
T of a ToeplitzPoperator T has a rather special form. Express P as a Fourier series
in L2 .T/: P D n2Z n P n . Observe that if k; j 0, then k-th Fourier coefficient of
the product .t/ j .t/ is the same as the k-th Fourier coefficient of ei jt .t/, which in
L2 .T/ is the Fourier coefficient arising from
h. P j /; P k i D hBj P ; P k i;
where B is the bilateral shift. Thus, the .k; j/-entry of the matrix for T is given by
hT P j ; P k i D hPH 2 M P j ; P k i D hM P j ; P k i D h. P j /; P k i
D hBj P ; P k ih P ; .Bj / P k i D h P ; P kj i D kj :
Note that the .k C 1; j C 1/-entry of T is also kj . Hence, the matrix representation
for T is given by
3
0 1 2 3
7
6
6 1 0 1 2 : : : 7
7
6
6
:: 7
7:
6
:
T D 6 2 1 0 1
7
7
6
6 3 2 1 0 : : : 7
5
4
:: :: :: ::
:
: :
::: :
2
338
In particular, the unilateral shift S has matrix representation

3
0 0 0 0
7
6
6 1 0 0 0 ::: 7
7
6
6
:: 7
6
S D6 0 1 0 0 :7
7:
7
6
:
6 0 0 1 0 :: 7
5
4
: : : :
::: :: :: :: ::
2
10.2.5 Weighted Unilateral Shift Operators

The following class of operators were mentioned in Section 8.1.
If D ./k2N 2 `1 .N/ and S W `2 .N/ ! `2 .N/ is defined by
2
31
3
0
v1
6 1 v1 7
B6 v2 7C
6
7
B6 7C
S v D S B6 v 7C D 6 v 7 ;
4 2 25
@4 3 5A
::
::
:
:
02
v 2 `p .N/;
then S is is called a weighted unilateral shift operator. The (Hilbert space) adjoint
.S / of S is given by
2
31
3
v1
1 v2
6 2 v3 7
B6 v2 7C
6
B6 7C
7
S v D S B6 v 7C D 6 v 7 ;
3
3
4
4
@4 5A
5
::
::
:
:
02
v 2 `p .N/;
Note that the Hilbert space adjoint of S is slightly different from the Banach space
adjoint of S that is given in equation (8.2) of Section 8.1.
10.2.6 Rank-1 Operators

An operator T 2 B.H/ is of rank 1 if T has 1-dimensional range. Let 2 H be a
vector that spans the range of T. Then, for each 2 H, there is a unique 2 C for
which T D . The map H ! C given by 7! is easily seen to be linear and
bounded, and so by the Riesz Representation Theorem there is a vector 2 H for
which T D h; i , for every 2 H.
10.3 Hermitian Operators
339
This form of a rank-1 operator T is expressed by T D . That is,

T D . / D h; i ;
for every 2 H.
10.2.7 Direct Sums of Operators

If Tj 2 B.Hj /, for j D 1; : : : ; n, and if H is the Hilbert space direct sum H D
n
M
Hj
jD1
(see Proposition
1 then T D T1 Tn denotes the unique operator on H
0 5.81),
n
n
M
M
j A D
Tj . Observe that T has norm kTk D maxj kTj k and
defined by T @
jD1
jD1
that the adjoint of T is given by T D T1 Tn . More generally, if
H is
2
the direct sum of a family fH g2 of Hilbert spaces (as in Proposition 5.98), and
if fT g2 is a familyM
of operators
MT 2 B.H / such that sup kT k < 1, then the
linear transformation
T on
H defined by
2
2
!
T . / D .T / ;
2
for . / 2
H , is an operator of norm sup kT k and adjoint
2
T .
2

Much of the theory of Hilbert space operators is devoted to the ways in which T and
T interact. The first case of interest occurs when T and T are in fact the same.
Such operators are probably the most important in all of operator theory and its
applications.
Definition 10.20. An operator T 2 B.H/ is hermitian if T D T.
Example 10.21. A multiplication operator M on L2 .X; ; / is hermitian if and
only if ess-ran R.
Proof. Recall that if .X; ; / is a measure space, then M D M . However, M D
M if and only if .t/ D .t/ for almost all t 2 X.
t
u
340
For any T 2 B.H/, T D 12 .T C T / C i 2i1 .T T /. Since T C T and i.T T /

are hermitian, the hermitian operators span B.H/. This is one reason why hermitian
operators are of importance.
If T 2 B.H/ is hermitian, then hT; i D h; T i D h; Ti, for every ; 2 H. In
particular, if D , this implies that hT; i D hT; i, for all 2 H; that is, the form
7! hT; i is necessarily real valued if T is hermitian. This necessary condition is
also sufficient.
Proposition 10.22. An operator T 2 B.H/ is hermitian if and only if hT; i 2 R
for all 2 H.
Proof. If T is hermitian, then for every vector ,
hT; i D h; T i D h; Ti D hT; i;
and therefore hT; i 2 R.
Conversely, suppose that hT#; #i 2 R for all #; then hT#; #i D h#; T#i. However,
T also satisfies hT#; #i D h#; T #i, for every # 2 H. Therefore, h#; T#i D
h#; T #i, for every # 2 H, and so T D T, by the Polarisation Identity (Corollary 10.11).
t
u
Corollary 10.23. If T 2 B.H/ is hermitian and if M H is a T-invariant subspace,
then the restriction TjM of T to M is hermitian.
Proof. If 2 M, then hTjM ; i D hT; i 2 R. Hence, by Proposition 10.22, TjM is
hermitian.
t
u
The next set of propositions reveals some striking features of the spectra of
hermitian operators.
Proposition 10.24. If T 2 B.H/ is a hermitian operator, then .T/ D ap .T/ R.
Proof. Recall that .T/ D ap .T/ [ d .T/. By Proposition 10.9, 2 d .T/ if and
only if 2 p .T /. As T D T, if we show that every eigenvalue of T is real, then
we will obtain d .T/ R. To this end, let T D for some 2 C and unit vector
2 H. Because hT; i 2 R (Proposition 10.22), we obtain
D h; i D h ; i D hT; i 2 R :
Hence, d .T/ D p .T/ R and
.T/ D ap .T/ [ d .T/ D ap .T/ [ p .T/ ap .T/ .T/ :
We now show that ap .T/ R. Suppose that 2 C n R. For any nonzero 2 H,
we have
341
0 < j j kk2 D jh.T 1/; i h.T 1/; ij

D jh.T 1/; i h; .T 1/ij
2k.T 1/k kk :
Hence, k.T 1/k 12 j j kk for all 2 H. This proves that .T 1/
is bounded below and, hence, 62 ap .T/. This concludes the proof of
.T/ D ap .T/ R.
u
t
Proposition 10.25. If T 2 B.H/ is hermitian, then
kTk D max fj j j 2 .T/g:
That is, kTk D spr T.
Proof. Let D kTk. For any unit vector 2 H,
k.T 2 2 1/k2 D h.T 2 2 1/; .T 2 2 1/i
D kT 2 k2 2 2 kTk2 C 4 kk2
2 kTk2 2 2 kTk2 C 4
D 4 2 kTk2 :
(10.4)
By definition of the norm of an operator, there are unit vectors n 2 H such that
kTn k ! kTk. Hence, by inequality (10.4), limn k.T 2 2 1/n k2 exists and is equal
to 0. Therefore, 2 2 .T 2 /.
Because T 2 2 1 D .T C 1/.T 1/, at least one of the two operators on the
right-hand side of this expression must fail to be invertible. Thus, 2 .T/ or 2
.T/. In either case, there is a 2 .T/ such that j j D kTk. On the other hand,
j j kTk, for all 2 .T/ (Theorem 8.42), which completes the proof.
t
u
Proposition 10.26. Assume that T 2 B.H/ is hermitian and let
m` D inf hT; i
kkD1
and
mu D sup hT; i :
kkD1
Then m` ; mu 2 .T/ and .T/ m` ; mu .

Proof. If T 0 D T m` 1, then T 0 is hermitian and m` 2 .T/ if and only if 0 2 .T 0 /.
Therefore, we assume without loss of generality that m` D 0. Under this assumption,
the sesquilinear form ; W H H ! C defined by ; D hT; i satisfies the
342
Cauchy-Schwarz inequality
j; j ; 1=2 ; 1=2 ;
8 ; 2 H :
Therefore, with D T,

kTk4 D jhT; Tij2 hT; i hT 2 ; Ti hT; i kTk3 kk :
Hence, inf kTk D 0, which proves that T is not bounded below. That is, 0 2 .T/.
kkD1
On the other hand, if < 0, then

k.T 1/k2 D kTk2 2 hT; i C 2 kk2 2 kk2
implies that T 1 is bounded below. Hence, 62 ap .T/ D .T/. This proves that
.T/ 0; 1/.
The proof that mu 2 .T/ and that .T/ .1; mu is left to the reader
(Exercise 10.116).
t
u
Corollary 10.27. If T 2 B.H/ is hermitian, then Conv .T/ D fhT; i j kk D 1g.
Proof. The unit sphere S is a Hilbert space H is a path-connected set, and so the
continuous map S ! R given by 7! hT; i has a path-connected range. Because
the infimum and supremum of the range of this map are the minimum and maximum
elements of the spectrum of T, we deduce the equality of the sets Conv .T/ and
fhT; i j kk D 1g.
t
u
By way of the classical Weierstrass Approximation Theorem, the usual polynon
X
mial functional calculus T 7! f .T/ D
j T j , where T is a Banach space operator
and f .t/ D
n
X
jD0
j tj is a polynomial with complex coefficients, extends to continuous
jD0
functions in cases where T is a hermitian Hilbert space operator. The main result in
this direction is stated below, and it is a major tool in the analysis of Hilbert space
operators.
Theorem 10.28 (Continuous Functional Calculus for Hermitian Operators). If
T 2 B.H/ is a hermitian operator, then for every continuous function f W .T/ ! R
there is a unique hermitian operator f .T/ with the property that
kf .T/ qn .T/k ! 0
for every sequence fqn gn2N of polynomials qn 2 R t for which

lim
n!1

max jf .t/ qn .t/j D 0:
t2.T/
(10.5)
343
Furthermore, for all continuous functions f ; g W .T/ ! R and 2 R,

1.
2.
3.
4.
kf .T/k D max 2.T/ jf . /j,

f .T/ D .f .T//,
.f C g/.T/ D f .T/ C g.T/, and
fg.T/ D f .T/g.T/.
Proof. Select a real-valued f 2 C . .T//. By the Weierstrass Approximation Theorem (Exercise 5.109), there exists a sequence fqn gn2N of polynomials with complex
coefficients such that limn maxt2.T/ jf .t/ qn .t/j D 0. By considering 12 .qn C qn /,
we may assume without loss of generality that each qn has real coefficients.
The convergent sequence fqn gn2N is necessarily Cauchy in C . .T//. Furthermore,
qm .T/ qn .T/ is hermitian, for all m; n 2 N, and so the Spectral Mapping Theorem
shows that
.qm .T/ qn .T// D fqm . / qn . / j 2 .T/g:
Thus, kqm .T/ qn .T/k D max 2.T/ jqm . / qn . /j (Proposition 10.36) and therefore fqn .T/gn2N is a Cauchy sequence of hermitian operators in B.H/. Denote the
limit of this sequence by f .T/, and observe that f .T/ is independent of the choice of
approximating sequence fqn gn2N .
Now if 2 R and f and g are polynomials with real coefficients, then
kf .T/k D max 2.T/ jf . /j, f .T/ D .f .T//, .f C g/.T/ D f .T/ C g.T/, and
fg.T/ D f .T/g.T/. Hence, by the approximation in equation (10.5), this properties
also hold for continuous functions f ; g W .T/ ! R.
t
u
The map f 7! f .T/ in Theorem 10.28 is called continuous functional calculus for T. Not all Hilbert space operators admit continuous functional calculus
(Exercise 10.131), and so Theorem 10.28 is quite particular to hermitian (and
hermitian-like) operators.
A useful application of the continuous functional calculus concerns isolated
points in the spectrum of a hermitian operator. Recall from Definition 1.68 that a
limit point of a subset Y in a topological space X is a point x 2 X such that for every
open set U containing x there is an element y 2 Y such that y 2 U and y 6D x. By
Proposition 1.69, the closure of Y is given by Y D Y [ L.Y/, where L.Y/ denotes the
set of limit points of Y.
Definition 10.29. If Y is a subset of a topological space X, then an element y 2 Y is
an isolated point of Y if y 2 Y n L.Y/.
Proposition 10.30. An isolated point of the spectrum of a hermitian operator T is
necessarily an eigenvalue of T.
Proof. The set f g is a closed subset of .T/. Because is an isolated point of
.T/, the characteristic function f D f g , as a map .T/ ! C, is a nonzero realvalued continuous function such that f 2 D f . Therefore, by the continuous functional
calculus, the operator P D f .T/ is hermitian, nonzero, and satisfies P2 D P; in other
words, P is a nonzero projection. Now consider the function h W .T/ ! C given by
344
h.t/ D tf .t/, for t 2 .T/. Because h.t/ D if t D and h.t/ D 0 otherwise, we see
that h.t/ D f .t/. Hence, the continuous functional calculus yields TP D Tf .T/ D
f .T/ D P, and so every 2 ran P satisfies T D .
t
u
The spectrum of a compact operator K is finite or countable, and the nonzero
points of .K/ correspond to eigenvalues of finite geometric multiplicity (Theorem 8.55). If a compact operator K acts on a Hilbert space H and if K D K, then
the eigenvectors and corresponding to distinct eigenvalues and of K are
orthogonal by the following computation:
. /h; i D h ; i h; i D hK; i h; Ki D hK; i hK; i D 0;
which implies that h; i D 0 because 6D 0. Hence, ker.K 1/? ker.K 1/
for every pair of distinct eigenvalues and of K.
Theorem 10.31 (Spectral Theorem #1). If K 2 B.H/ is hermitian and compact,
then H has an orthonormal basis consisting of eigenvectors of K.
Proof. Express H as H D ker K .ker K/? D ker K .ran K / D ker K .ran K/. If
one finds orthonormal bases for each of ker K and .ran K/, then the union of these
bases will be an orthonormal basis for H. Because ker K is the eigenspace of K
corresponding to the eigenvalue D 0, it is enough to prove that .ran K/ has an
orthonormal basis of eigenvectors of K. Furthermore, because .ran K/ is a Hilbert
space invariant under K, we may assume without loss of generality that ker K D f0g
and that the range of K is dense in H.
Because ker.K 1/? ker.K 1/ for every pair of distinct eigenvalues and
of K, we may form a direct sum (as in Proposition 5.98) of all the eigenspaces of
K. To this end, let
M
MD
ker.K 1/:
2.K/; 6D0
An orthonormal basis BM for M is obtained by taking an orthonormal basis B

(necessarily finite) for each ker.K 1/ and then setting
BM D
B :
2.K/; 6D0
The elements of BM are obviously eigenvectors of K. Therefore, it remains to prove

that M ? D f0g.
The subspace M is K-invariant, and so M ? is K -invariant (Proposition 10.18);
that is, M ? is K-invariant. Because the restriction KjM? of K to M ? is a compact
operator, the spectrum of KjM? consists of 0 and nonzero eigenvalues. However, any
eigenvalue of KjM? would also be an eigenvalue of K, and since the eigenvectors of
K corresponding to its nonzero eigenvalues all lie in M, it cannot be that KjM?
has nonzero eigenvalues. Therefore, because KjM? is hermitian (Corollary 10.23)
and because the norm and spectral radius of a hermitian operator coincide (Corol-
345
lary 10.27), we deduce that KjM? D 0. But if M ? were nonzero, then nonzero vectors
in M ? would be elements of ker K, which would contradict ker K D f0g. Hence,
M ? D f0g and, therefore, M D H.
t
u
A second formulation of the spectral theorem may be viewed as a spectral
decomposition of compact hermitian operators.
Theorem 10.32 (Spectral Theorem #2). The following statements are equivalent
for a compact operator K acting on a separable Hilbert space H:
1. K is hermitian;
2. there are a bounded sequence f j grjD1 of nonzero real numbers, where r is finite
or infinite, and a sequence f j grjD1 of pairwise-orthogonal unit vectors in H such
that
a. K j D j j , for each j,
b. limj j D 0, if r is infinite, and
r
X
c. K D
j h; j i j , for every 2 H.
jD1
Proof. Assume (1). Theorem 10.31 asserts that H has an orthonormal basis
consisting of eigenvectors of K. The proof of Theorem 10.31 indicate that this basis
consists of two parts: (i) an orthonormal basis f k gskD1 of ker K and an orthonormal
basis f j grjD1 of
MD
ker.K 1/;
2.K/; 6D0
where M D ran K. Let j 2 R denote the eigenvalue corresponding to the eigenvector

j , which thereby yields property (2a). If r is infinite, then there must be infinitely
distinct points in the sequence f j grjD1 because, by Theorem 8.55, (i) the eigenspace
associated with any individual nonzero eigenvector is finite dimensional (meaning
an individual point j in the sequence is repeated in the sequence at most finitely
many times) and (ii) the point 0 2 R is the only cluster point of .K/. Hence,
property (2b) holds.
To prove property (2c), express H as H D ker K M so that 2 H has the form
D C , where 2 ker K and 2 M. Write in its Fourier decomposition:
D
r
X
h; j i j :
jD1
Note that K D K C K D K and that h; j i D h; j i for every j.Thus,

K D K D
r
r
r
X
X
X
h; j i K j D
j h; j i j D
j h; j i j ;
jD1
which proves property (2c).
jD1
jD1
346
Conversely, if an operator K 2 B.H/ satisfies (2), then, for every 2 H,

*
hK; i D
r
X
+
r
r
X
X
j h; j i j ;
h; i i i D
j jh; j ij2 2 R
jD1
iD1
jD1
because each j is real. Hence, K D K.
t
u
10.4 Normal Operators

The relationship between an operator T and its adjoint T is an important one,
and indeed T shares many of the properties that T possesses (such as the norm).
However, there may be very little relation between the two operators with regard to
how they act on the Hilbert space. This section considers one case in which there
is a rather strong algebraic link between T and T , and this specific relationship is
called normality. The class of normal operators includes every hermitian operator
and every unitary operator.
Definition 10.33. An operator N 2 B.H/ is a normal operator if N N D NN .
Example 10.34. The bilateral shift operator B on `2 .Z/ is normal, but the unilateral shift operator S on `2 .N/ is nonnormal.
Proof. Because the bilateral shift operator B on `2 .Z/ is unitary, B is plainly normal.
However, the unilateral shift operator S on `2 .N/ is nonnormal, as S Se1 D e1 but
SS e1 D 0 (where e1 is the first canonical basis vector of `2 .N/).
t
u
The unilateral shift operator S is a good example of the challenges in quantifying
nonnormality. On the one hand, S is rather close to being a normal operator because
S S SS D e1 e1 is a rank-1 operator. On the other hand, S is rather far from
normal in the sense of norm, given that kS S SS k D 1.
Proposition 10.35. The following statements are equivalent for T 2 B.H/:
1. T is normal;
2. kT k D kTk, for all 2 H;
3. the real and imaginary parts of T commute.
t
u
The third equivalent statement of Proposition 10.35 above points to an intimate

relationship between normality and the property of being hermitian. The next set of
results illustrate how some of the spectral features of hermitian operators are shared
by normal operators.
347
Proposition 10.36. If N 2 B.H/ is normal, then spr N D kNk.

Proof. By the spectral radius formula (Theorem 8.45),
spr N D lim kN n k1=n :
n
In particular,
1
spr N D lim kN 2k k 2k :
k
(10.6)
Because N is normal, .N 2 / .N 2 / D .N N/2 . Thus,

kN 2 k2 D k.N 2 / .N 2 /k D k.N N/.N N/k D kN Nk2 D kNk4 ; ;
which implies that kN 2 k D kNk2 . By induction, kN 2k k D kNk2k , for all k 2 N.
Hence, by (10.6), spr N D kNk.
u
t
Proposition 10.37. If N is a normal Hilbert space operator, then .N/ D ap .N/.
Proof. Because .N/ D ap .N/ [ d .N/ D ap .N/ [ p .N / , it is sufficient to show
that if 2 p .N /, then 2 p .N/. Therefore, assume that 2 p .N / and let 2 H
be nonzero with N D . Because N is normal, it is also true that N 1 is a normal
operator. Consequently, Proposition 10.35 implies that
0 D k.N 1/k D k.N 1/ k D k.N 1/k :
Hence, 2 p .N/ ap .N/.
t
u
Using the properties of normal operators above, we can deduce precise information concerning multiplication operators.
Example 10.38. If .X; ; / is a -finite measure space, then every multiplication
operator M on L2 .X; ; /, where 2 L 1 .X; ; /, is a normal operator with
spectrum
.M / D ess-ran
and norm
kM k D spr M D ess-sup :
Proof. The multiplication operator M on L2 .X; ; / has adjoint M D M , and it
is clear that M commutes with every multiplication operator M% , not just with M .
In any case, M is a normal operator.
348
Example 8.1.7 shows that kM k D ess-sup , whereas Example 8.53 proves that
ap .M / D ess-ran . The normality of M yields ap .M / D .M / (Proposition 10.37), and so kM k D spr M .
t
u
Example 10.38 is indicative of the general case, for if N is a normal operator acting on a separable Hilbert space H, then there is a -finite measure space .X; ; /
and a surjective isometry U W H ! L2 .X; ; / such that UNU 1 is the multiplication
operator M on L2 .X; ; /, for some 2 L 1 .X; ; / (Corollary 11.34).
The proof of Proposition 10.37 reveals that N D if and only if N D .
Therefore, if N D and N D for distinct eigenvalues and of N and
nonzero vectors and , then
. /h; i D h ; i h; i D hN; i h; N i
D hN; i hN; i D 0;
which implies that h; i D 0. Hence, ker.N 1/ ? ker.N 1/ for distinct
eigenvalues and of N. With this observation, the proofs of the spectral theorems
for compact operators carry over verbatim, with only change being the requirement
(in the hermitian case of the theorem) that the eigenvalues be real.
Theorem 10.39 (Spectral Theorem for Compact Normal Operators). The
following statements are equivalent for a compact operator K acting on a separable
Hilbert space H:
1. K is normal;
2. there are a bounded sequence f j grjD1 of nonzero complex numbers, where r is
finite or infinite, and a sequence f j grjD1 of pairwise-orthogonal unit vectors in H
such that
a. K j D j j , for each j,
b. limj j D 0, if r is infinite, and
r
X
c. K D
jD1
As with hermitian operators, normal operators admit a continuous functional

calculus. Below, C s; t denotes the commutative ring of complex polynomials in
two variables.
Theorem 10.40 (Continuous Functional Calculus for Normal Operators). If N
is a normal operator, then for every continuous function f W .N/ ! C there is a
unique hermitian operator f .N/ with the property that
kf .N/ qn .N; N /k ! 0
(10.7)
349
for every sequence fqn gn2N of polynomials qn 2 C s; t for which

lim
n!1

max jf .z/ qn .z; z/j D 0:
z2.N/
Furthermore, for all continuous functions f ; g W .N/ ! C and 2 C,

1.
2.
3.
4.
kf .N/k D max 2.N/ jf . /j,

f .N/ D .f .N//,
.f C g/.N/ D f .N/ C g.N/, and
fg.N/ D f .N/g.N/.
Proof. The proof is identical to the proof of Theorem 10.28 except for the
approximation indicated in (10.7). The use of polynomials q in two variables s and
t is necessary to invoke the Stone-Weirerstrass Theorem. That is, if S is the set of
all continuous functions on .N/ of the form z 7! q.z; z/, where q 2 C s; t, then S
contains the constants, is self-adjoint, and separates the points of .N/. Hence, S
is dense in C . .N//.
t
u
Theorem 10.40 has an even stronger form, in which the use of continuous
functions is extended to bounded Borel functions on the spectrum of N. This
extension of continuous functional calculus to Borel functional calculus will not
be needed for the topics studied in this text, and so we shall not develop it here.
With Theorem 10.40 at hand, one can prove assertions such as the following
result.
Proposition 10.41. If is an isolated point in the spectrum of a normal operator
N, then is an eigenvalue of N.
t
u
If a normal operator N leaves the subspace Span fg invariant, for some nonzero
vector 2 H, then Span fg is invariant under the action of N also. But this feature
does not apply to all invariant subspaces of a normal operator.
Example 10.42. The subspace `2 .N/ of `2 .Z/ is invariant for the bilateral shift
operator B, but not for B .
Proof. Because B is a forward shift, which is to say that Bek D ekC1 for every k 2 Z,
it is clear that B 2 `2 .N/ for every 2 `2 .N/. However, with the vector e0 2 `2 .N/,
we have that B e0 D e1 62 `2 .N/, and so `2 .N/ is not B -invariant.
t
u
Motivated by the dual-invariance feature exhibited by normal operators with
respect to the eigenvectors, one is led to the following class of operators.
Definition 10.43. An operator T 2 B.H/ is a reductive operator if M ? is Tinvariant for every T-invariant subspace M H.
350
Evidently every hermitian operator is reductive. However, as Example 10.42

demonstrates, not every normal operator is reductive. If we consider only compact
normal operators, then the following positive result holds.
Proposition 10.44. Every compact normal operator is reductive.
t
u
Do there exist reductive operators that are nonnormal? This is an open question,
equivalent to the long-standing open question of whether every operator on a
separable Hilbert space has a nontrivial invariant subspace. Therefore, we might
ask whether there exist compact reductive operators that are nonnormal.
Theorem 10.45 (Rosenthal). If a compact operator K is reductive, then K is
normal.
Proof. The first step of the proof is to show that there is a unit vector 2 H such
that is an eigenvector of both K and K . This is achieved by a Zorns Lemma
argument.
Because K is compact, K has a nontrivial invariant subspace (Theorem 8.59).
Consider the family F of all nonzero K-invariant subspaces M H such that
kKjM k D kKk. The family F is nonempty because H 2 F . Furthermore, the relation
- on F defined by
L-M
if and only if L M
is a partial order on F .
Let fMg2 be linearly ordered chain in F and consider the subspace
\
ND
M :
(10.8)
2
Note that the subspace N is reducing for K. If it can be shown that N 2 F , then N
will be an upper bound in F for the linearly ordered chain fMg2 .
Because H is a separable metric space, every open covering of an open set admits
a countable subcovering (by Exercise 2.103). Taking set-theoretic complements in
equation (10.8) leads to
H nN D
HnM ;
2
which is a covering of the open set H n N by the family of open sets H n M . Hence,
by Exercise 2.103, fH n M g admits a countable covering fH n Mn gn2N of H n N.
Therefore,
\
Mn :
(10.9)
ND
n2N
351
Because the original descending chain was linearly ordered, the sequence of
subspaces Mn can be assumed to be ordered so as to satisfy
Mn MnC1 ;
for every n 2 N :
(10.10)
Every compact operator achieves its norm on some unit vector. Therefore, for
each n 2 N there is a unit vector n 2 Mn such that
kKn k D kKjMn k D kKk :
Because K is compact, there is a subsequence fnk gk of fn gn such that fKnk gk
converges to a vector 2 H. The norm of is necessarily kKk, since kKnk k D
kKk for every k 2 N. Equations (10.9) and (10.10) still hold if one replaces the
sequence of subspaces Mn by the subsequence fMnk gk2N . Therefore, without loss
of generality, it may be assumed that the original sequence fKn gn2N converges to
2 H.
The closed unit ball of H is weakly compact by the Banach-Alaoglu Theorem.
Thus, the sequence fn gn2N admits a subsequence fnk gk2N that is weakly convergent
to some vector 2 H of norm kk 1. Fix k. For every j > k, the vector nj belongs to
Mnk . Thus, the weak limit is also the weak limit of the sequence fnj gjk in Mnk .
Therefore the vector belongs to Mnk . As this is true for any k, equation (10.10)
yields 2 N.
For each k 2 N,
k Knk k2 D kk2 C kKnk k2 2<hKnk ; i D 2kKk2 2<hKnk ; i :
As k ! 1, the equation above yields <hK; i D kKk2 . Thus,
kKk2 jhK; ij kKk kk D kKk kKk kKk2 ;
and so kKk D kKk. This shows that 2 N is a nonzero vector (in fact it is a unit
vector). Hence, the orthogonally reducing subspace N is at least one-dimensional
and K achieves its norm on N. That is, N 2 F , and so N is an upper bound in F for
the linearly ordered chain fMg2 . By Zorns Lemma, F has a maximal element,
say M.
Since M 2 F , M is nonzero. We shall show that M is one-dimensional. Suppose
that dim M > 1. Because KjM is compact, there is a nontrivial subspace L M
that is invariant under KjM . Since the subspace L is K-invariant and because K is a
reductive operator, L? is K-invariant. Hence, L and L? \ M are invariant under KjM .
This implies that KjN achieves its norm on L or on L? \ M. Either case contradicts
the maximality of M in F . Therefore, M must be one-dimensional. Thus, if 2 M
is a unit vector, then is an eigenvector of K and K .
The second step in the proof shows that K is normal. Let E be a maximal family
of orthonormal vectors 2 H for which is an eigenvector of K and K . If E is an
orthonormal basis of H, then K has an orthonormal basis consisting of eigenvectors
352
and is, therefore, normal. Suppose that E is not an orthonormal basis of H. Since the
closure of the span of E is K-invariant, the subspace E ? is also K-invariant, since
K is a reductive operator. By the arguments in the first step, there is a unit vector
2 E ? which is an eigenvector of KjE ? and .KjE ? / . Since K is reductive, this
vector is an eigenvector of K and K . But 2 E ? , which is in contradiction to
the maximality of the family E . Hence, it must be that E is an orthonormal basis
of H.
t
u
10.5 Positive Operators

To this point we have seen that hermitian operators have properties connected to
the real numbers, while normal operators have the flavour of arbitrary complex
numbers. In this section, the operator-theoretic analogue of a nonnegative real
number is introduced, which is one of the most important features of Hilbert space
operator theory.
Definition 10.46. An operator A 2 B.H/ is positive if A is hermitian and .A/
0; 1/.
Corollary 10.27 provides an alternate criterion for the positivity of an operator:
Proposition 10.47. A 2 B.H/ is positive if and only if hA; i 0 for every 2 H.
Corollary 10.48. T T is a postive operator, for every operator T.
Positivity also captures information about the norm of an operator.
Proposition 10.49. Ifp is a positive real number and T 2 B.H/, then 1 T T is
positive if and only if kTk.
Proof. For each 2 H,

h 1 T T ; i D kk2 kTk2 :
Thus, h.1 T T/ ; i 0 for every 2 H if and
p only if kTk
2 H. That is, 1 T T is positive if and only if kTk.
kk for all

t
u
One of the most useful features of positive operators is that they possess (unique)
positive square roots.
Theorem 10.50. If A 2 B.H/ is positive, then
1. there is a positive operator R 2 B.H/ such that R2 D A, and
2. if R1 2 B.H/ is a positive operator such that R21 D A, then R1 D R.
p
Proof. Since .A/ 0; 1/ and the function f .t/ D t is continuous on the
spectrum of A. Consider the hermitian operator R D f .A/, which satisfies, by
Theorem 10.28, R2 D A. We now show that R is positive.
353
By scaling A we may assume without loss of generality that kAk D 1. Thus,

.A/ p0; 1. For each n 2 N, let qn be the n-th Bernstein polynomial approximating
f .t/ D t: namely,
qn .t/ D
n
X
kD1
r
k n k
t .1 t/nk :
n k
Therefore, qn .t/ 0 for all t 2 0; 1, which implies

that .qn .A// 0;
1/,
by the Spectral Mapping Theorem; furthermore, limn maxt20;1 jqn .t/ f .t/j D 0
([14, 10.3]). Thus, the sequence fqn .A/gn of positive operators converges to R.
Proposition 10.26 implies that the smallest element m` in .R/ has the form
m` D inf hR; i :
kkD1
If m` < 0, then there must be a unit vector and an n 2 N such that hqn .A/; i < 0.
On the other hand, as qn .A/ is positive, hqn .A/; i 0 by Proposition 10.26.
This contradiction implies that m` 0 and so R is positive. Thus, .R/ 0; 1/.
Furthermore, since 0 qn .t/ 1 for all t 2 0; 1, each kqn .A/k 1 and so kRk 1.
2
Assume that R1 is positive and R21 D A. Since
p 0; 1 .A/ D f j 2 .R1 /g,
we see that
p .R1 / 0; 1. Note that qn .t/ ! t uniformly on 0; 1 implies that
qn .t2 / ! t2 D t uniformly. Thus, qn .R21 / ! R1 . That is,
R1 D lim qn .R21 / D lim qn .A/ D R ;
n
which proves that A has a unique positive square root.
t
u
Notational Convention If A 2 B.H/ is positive, then A1=2 will denote the unique
positive square root of A.
The proof of Theorem 10.50 establishes the following result, which we record
here formally for future use.
Proposition 10.51. If A is a positive operator and if f W .A/ ! R is a nonnegative
continuous function, then f .A/ is a positive operator.
Definition 10.52. The Loewner ordering on the set B.H/sa of hermitian operators
is the partial ordering in which S T, for hermitian S and T, if and only if T S
is positive.
Note that S T, for hermitian operators S and T, if and only if hS; i hT; i,
for every 2 H. An elementary but useful fact about the Loewner ordering is:
Proposition 10.53. If S and T are hermitian operators for which S T, then
X SX X TX for every operator X 2 B.H/.
t
u
354
The continuous functional calculus of Theorem 10.28 demonstrates that algebraic features of continuous maps f and g carry over to operators f .T/ and g.T/
(that is, the continuous functional calculus preserves sums, products, and scalar
multiplication). The situation is rather different when considering the preservation
of the Loewner order (see Exercise 10.132). There are, however, some positive
results, and the first of these (below) is amongst the most important.
Proposition 10.54. If A and B are positive operators and if A B, then A1=2 B1=2 .
Proof. By the uniqueness of the positive square root of a positive operator, it is
enough to prove that if A; B 2 B.H/ are positive operators such that A2 B2 , then
A B. Under these assumptions, note that if 2 H is a unit vector and if 2 R, then
hB2 ; i hB; i is a real number and hB; Ai is a real or complex number such
that
<hB; Ai jhB; Aij kBk kAk D hB2 ; i1=2 hA2 ; i1=2 hB2 ; i:
Now, to show that B A is positive, it is sufficient, by Proposition 10.26, to prove
that 0 for each 2 .B A/. To this end, select 2 .B A/. Because B A is
hermitian, is necessarily real and an approximate eigenvalue (Proposition 10.24).
Thus, there is a sequence of unit vectors n such that limn k.B A/n n k D 0.
For every n we have that
jhBn ; .B 1/n An ij kBk k.B A/n n k;
and therefore limn hBn ; .B 1/n An i D 0. Because every n is a unit vector,
each of the sequences fhB2 n ; n ign2N , fhBn ; n ign2N , and fhBn ; An ign2N lies in a
compact subset of C. Hence, there is a subsequence fnj gj2N of fn gn2N such that
fhB2 nj ; nj igj2N , fhBnj ; nj igj2N , and fhBnj ; Anj igj2N are convergent. Thus,

lim hB2 nj ; nj i hBnj ; nj i D lim hBnj ; Anj i
j!1
j!1
D lim <hBnj ; Anj i

j!1
lim hB2 nj ; nj i;

j!1
which implies that

lim hBnj ; nj i 0:
j!1
If limj hBnj ; nj i 6D 0, then necessarily 0. If, however, limj hBnj ; nj i D 0, then
0 D lim hBnj ; nj i D lim hB1=2 nj ; B1=2 nj i D lim kB1=2 nj k2
j!1
j!1
j!1
355
yields
0 D lim kB1=2 .B1=2 nj /k2 D lim hB2 nj ; nj i:
j!1
j!1
Because, for every j 2 N,

hB2 nj ; nj i hA2 nj ; nj i D kAnj k2 0;
we deduce from the inequality and limit equation above that limj kAnj k2 exists and
is equal to 0. Therefore, limj k.B A/nj nj k D 0 holds only if D 0.
t
u
If A is a positive operator, then the compactness of the spectrum implies that two
scenarios are possible: either (i) 0 2 .A/ or (ii) there exists > 0 such that
for all 2 .A/. In the latter case, Proposition 10.26 yields 0 < hA; i for every
unit vector 2 Hin other words, 1 A. This leads to a simple criterion for the
invertibility of positive operators:
Proposition 10.55. A positive operator A 2 B.H/ is invertible if and only if there
exists a real number > 0 such that 1 A.
The next proposition asserts that the function t 7! t1 on .0; 1/ is operator
monotone.
Proposition 10.56. If A and B are invertible positive operators such that A B,
then A1 and B1 are positive operators and B1 A1 .
Proof. If T is an invertible positive operator, then for every 2 H we have, using
D T 1 ,
hT 1 ; i D hT 1 .T/; Tii D h; Ti 0:
Hence, T 1 is a positive operator, by Proposition 10.47. (Alternatively, one could
argue via Proposition 10.51.)
p
By hypothesis, 1 A B for some real number > 0. Hence, 1 A1=2
B1=2 , by Proposition 10.54, which implies that B1=2 is invertible. Let T D A1=2 B1=2
and choose any 2 H. Thus, there is a unique 2 H for which D B1=2 , and so
kTk2 D kA1=2 k2 D hA; i hB; i D kB1=2 k2 D kk2 :
Therefore, kTk 1 and, hence, kT k 1. Thus, kB1=2 A1=2 #k2 k#k2 for every
# 2 H.
Select any 2 H and let # 2 H denote the unique vector for which D A1=2 #.
Thus,
hB1 ; i D kB1=2 k2 D kB1=2 A1=2 #k2 k#k2 D kA1=2 k2 D hA1 ; i;
which proves that B1 A1 .
t
u
356
Propositions 10.54 and 10.56 belong to a wider set of results

p on operator
monotone functions. These propositions say that the functions t 7! t and t 7! t1
are operator monotone on 0; 1 and .0; 1/, respectively. In contrast t 7! t2 is not
operator monotone in the sense that there exist positive operators A and B such that
A B but A2 6 B2 (Exercise 10.132).
A common technique in measure theory is to write an arbitrary real-valued
function as a difference of two nonnegative functions whose product is zero. That
idea carries over, via functional calculus, to hermitian operators.
Proposition 10.57. If A is a hermitian operator, then there are positive operators
AC and A 2 AC such that A D AC A and AC A D A AC D 0.
Proof. Let X D kAk; kAk, which is a compact set that contains .A/ and 0.
Consider the functions f ; g 2 C.X/ defined by f .t/ D .t C jtj/=2 and g.t/ D f .t/.
The functions f and g are nonnegative and vanish at 0; thus, by Proposition 10.51,
the operators f .A/ and g.A/ are positive. Let AC D f .A/ and A D g.A/. Because
t D f .t/g.t/ and f .t/g.t/ D 0 for all t 2 X, the continuous functional calculus yields
A D AC A and AC A D A AC D 0.
t
u
For compact positive operators, the min-max variational principle exhibited by
equation (10.11) below is very useful in the analysis and estimation of eigenvalues.
Theorem 10.58 (Courant-Fischer Theorem). If A 2 B.H/ is a positive compact
operator, then there are a bounded sequence f j grjD1 of real numbers, where r is
such that
1. A j D j j , for each j,
2. j jC1 > 0, for all j,
3. limj j D 0, if r is infinite, and
r
X
4. A D
jD1
Furthermore, for each j such that 1 j r,

j D
min
LH; dim LDj1

max
hA ; i :
2L? ; k kD1
(10.11)
Proof. Theorem 10.32 provides the spectral decomposition of A. By relabelling the

indices, we may assume that the elements of the sequence f j grjD1 are ordered so
that j jC1 , for every j. Therefore, all that remains is to prove equation (10.11).
By Proposition 10.26 the spectral radius 1 of A is given by
1 D sup hA; i :
kkD1
357
But since A is compact, the unit eigenvector 1 corresponding to the eigenvalue 1

also satisfies hA 1 ; 1 i D 1 and the supremum above is in fact a maximum:
1 D max hA; i :
kkD1
Let M1 D Span f 1 g, which is A-invariant. Thus, M1? is A-invariant and the

restriction AjM? of A to M1? is positive and compact. Therefore, the spectral radius
1
of AjM? is given by
1
spr AjM? D
1
max
2M1? ; kkD1
hAjM? ; i D
1
max
2M1? ; kkD1
hA; i:
Note that j 2 M1? for all j 2, and so spr AjM? j for all j 2. But since spr AjM?
1
1
is also an eigenvalue of A, it must be that spr AjM? D 2 . By induction, if Mj1 D
1
Span f 1 ; : : : ; j1 g, then
j D
max
?
2Mj1
; kkD1
hA; i:
Suppose now that L H is a subspace of dimension j 1 and that f 1 ; : : : ; j1 g

is an orthonormal basis of L. The matrix Z D h r ; s i1rj;1sj1 is a linear
map of Cj into Cj1 , and so ker Z 6D f0g. Select a unit vector 2 ker Z and let
j
X
D
` ` , which is a unit vector in Mj . The condition Z D 0 implies that
`D1
2 L? . Furthermore,
hA; i D
j
j
X
X
kD1 `D1
k ` hA k ; ` i D
j
X
j` j2 ` j
`D1
j
X
j` j2 D j :
`D1
Hence, j maxfhA; i j 2 L? ; kk D 1g for every subspace L of dimension j 1.

This completes the proof of equation (10.11).
t
u
Definition 10.59. The unit operator interval is the subset I.H/ of B.H/ defined by
I.H/ D fA 2 B.H/ j 0 A 1g:
Thus, I.H/ is the set of positive operators A of norm kAk 1or, equivalently,
the set of positive operators A for which 1 A is positive (Proposition 10.49).
The unit operator interval is plainly a convex set that contains every projection.
Proposition 10.60. An operator is an extreme point of the unit operator interval
I.H/ if and only if it is a projection.
358
Proof. Suppose that P is a projection and express P as a proper convex combination

P D A1 C .1 /A2 of A1 ; A2 2 I.H/, for some 2 .0; 1/. If 2 ker P, then 0 D
1=2
hP; i D hA1 P; iC.1 /hA2 ; i 0 gives kAj j k2 D hAj ; i D 0 for j D 1; 2,
and so A1 D A2 D 0. Because it is also true that the projection 1P is given by 1
P D .1 A1 / C .1 /.1 A2 /, the same argument shows that ker.1 P/ ker.1
A1 / \ ker.1 A2 /. Thus, the action of Aj on each of ker P and .ker P/? D ker.1 P/
coincides with the action of P on these subspaces. Because H D ker P ker.1 P/,
we deduce that A1 D A2 D P. Hence, every projection P is an extreme point of I.H/.
Conversely, suppose that A 2 I.H/ is not a projection. This means, by Proposition 10.30 that A has a point of spectrum in the open interval .0; 1/, say . Select
f 2 C. .A// such that 0 t f .t/ 1 for every t 2 .A/ and f . / 6D 0. Thus,
f .A/ 6D 0 and A1 D A C f .A/ and A2 D A f .A/ are elements of I.H/ such that
1
A C 12 A2 D A but neither A1 nor A2 equal A. Thus, A is not an extreme point of
2 1
I.H/. Hence, every extreme point of I.H/ must be a projection.
t
u
If H is finite-dimensional, then I.H/ is a compact convex set and so, by the KrenMilman Theorem, I.H/ is the closed convex hull of the set of projections. However,
a much sharper statement can be made.
Proposition 10.61. If H has finite dimension, then the unit operator interval I.H/
is the convex hull of the set of projections on H.
P
Proof. Select A 2 I.H/ and write A in its spectral decomposition: A D m
jD1 j Pj ,
where 1 ; : : : ; m are the distinct eigenvalues of A and each Pj is a projection with
range ker.A j 1/. If j is a unit eigenvector of A corresponding to the eigenvalue j ,
then j D h j j ; j i D hAj ; j i 2 0; 1. Thus, we may assume the eigenvalues of A are
ordered so that 1 1 > 2 > > m 0. Set i D i iC1 for 1 i < m,P
m D m ,
and mC1 D 1 1 ; thus, 1 ; : : : ; mC1 are convex coefficients such that m
iDj i D
P
j . For i D 1; : : : ; m, let Qi D ijD1 Pj and let QmC1 D 0. Thus, Q1 ; : : : ; QmC1 are
PmC1
projections and iD1
i Qi D A.
t
u
10.6 Polar Decomposition

In working with complex numbers z, it is sometimes advantageous to express z in
its polar form z D ei jzj, where is the argument of z. One can do the same with
operators on Hilbert space, and the result is a major structure theorem for arbitrary
operators called Polar Decomposition.
Definition 10.62. For any T 2 B.H/, the modulus jTj of an operator T 2 B.H/ is
the positive operator jTj D .T T/1=2 .
One could of course elect to have defined jTj by .TT /1=2 , which results in a
different operator than .T T/1=2 (if, for example, T is the unilateral shift operator).
The adoption of .T T/1=2 for jTj is made so that the polar form T D UjTj appears,
359
at least on the surface, in exactly the same form as the traditional way of expressing
the polar form of a complex number znamely, as z D ei jzj.
Definition 10.63. An operator V 2 B.H/ is a partial isometry if there exists a
subspace M of H such that VjM is an isometry and VjM? D 0.
The subspace M is called the initial space and of V, and the range of V is called
the final space of V.
Proposition 10.64. If V 2 B.H/ is a partial isometry with initial space M and final
space V.M/, then
1. V V is a projection with range M, and
2. VV is a projection with range V.M/.
Proof. Let P 2 B.H/ be the projection with ran P D M. To show that V V D P, first
note that V is a contraction because, for every 2 H, .1 P/ 2 M ? D ker V and
P 2 M and so for every 2 M,
kVk D kV.P C .1 P/k D kVPk D kPk kk:
Therefore, if 2 M is a unit vector, then
1 D kVk2 D hV V; i kV Vk kk kV Vk 1
gives a case of equality in the Cauchy-Schwarz inequality; hence, V V D , which
shows that V V and P agree on ran P. If 2 M ? D .ran P/? D ker P D ran .1 P/,
then V V D 0 since M ? D ker V. Thus, V V and P agree on .ran P/? . Hence, V V
and P agree on H.
Let Q D VV . Because V D V.V V C .1 V V// D VV V D QV, the range of V
is contained in the range of Q. But Q D VV implies that the range of Q is contained
in the range of V. Thus, ran Q D ran V D V.M/. Lastly, Q is plainly hermitian and
Q2 D .VV /.VV / D .VV V/V D VV D Q.
t
u
It is also true that if V is an operator such that V V is a projection, then V is a
partial isometry (Exercise 10.138).
Theorem 10.65 (Polar Decomposition). For every T 2 B.H/ there exists a partial
isometry V 2 B.H/ such that
1. the initial space of V is ran jTj,
2. the final space of V is ran T, and
3. T D VjTj.
Furthermore, if T D V1 R1 for some positive operator R1 and partial isometry V1
with initial space ran R1 , then R1 D jTj and V1 D V.
Proof. For every 2 H,
k jTjk2 D hjTj; jTji D hjTj2 ; i D hT T; i D kTk2 :
Therefore, ker jTj D ker T.
(10.12)
360
Let V0 W ranjTj ! H be the function that maps each jTj 2 ranjTj to T 2 ran T.
Since jTj1 D jTj2 only if 1 2 2 ker jTj D ker T, V0 is a well-defined linear
surjection ranjTj ! ran T. Because kV0 k D k k for all 2 ran jTj, V0 extends (by
continuity) to an isometry ran jTj ! H, denoted again by V0 . Therefore, the range
of the isometry V0 is closed and coincides with ran T. Now extend V0 to a partial
?
isometry V 2 B.H/ by defining V D 0 for all 2 ran jTj D ker jTj D ker T. Hence,
V is a partial isometry with initial space ran jTj, final space ran T, and satisfies
VjTj D T.
Suppose next that T D V1 R1 for some positive operator R1 and partial isometry
V1 with initial space ran R1 . Because, for every 2 H,
hT T; i D kTk2 D kV1 R1 k2 D kR1 k2 D hR21 ; i ;
T T D R21 by the Polarisation Identity. Thus, jTj D .T T/1=2 D .R21 /1=2 D R1 by the
uniqueness of the positive square root. Hence, VjTj D V1 jTj. That is, V and V1 agree
?
on ran jTj. But since the initial space of R1 is ran jTj, R1 is zero on ran jTj . Hence,
V and V1 agree on all of H.
t
u
Definition 10.66. The polar decomposition of an operator T 2 B.H/ is the unique
decomposition of T as T D VjTj, where V is a partial isometry with initial space
ran jTj and final space ran T.
Two properties of the polar decomposition are noted below as corollaries for
future reference.
Corollary 10.67. If T D VjTj is the polar decomposition of T, then V T D jTj.
Proof. The operator V V is a projection with range ran T, and so V V D for
every 2 ran T.
t
u
Corollary 10.68. If T 2 B.H/ is invertible, then the partial isometry V in the polar
decomposition T D VjTj of T is a unitary operator.
Proof. Equation (10.12) shows that there is a sequence of unit vectors fn gn2N with
limn kTn k D 0 if and only if limn k jTjn k D 0that is, 0 2 ap .T/ if and only if
0 2 ap .jTj/. Thus, if 0 62 ap .T/, then 0 62 ap .jTj/ D .jTj/, which implies that jTj
is invertible. Therefore, V D TjTj1 is invertible and the initial space of V is H,
which means that V V D 1. Hence, V D V 1 , which implies that V is unitary. u
t
The polar decomposition informs the theory of Hilbert space operators in a
variety of manners. For example, the polar decomposition yields the following
information about the geometry of the closed unit ball of B.H/.
Proposition 10.69. If an invertible contraction T is not unitary, then T is the
average of two unitaries.
Proof. Corollary 10.68 shows that the polar decomposition of T is of the form T D
UjTj, for a unitary operator U. Because T T 1, 1 jTj2 D 1 T T is a positive
361
operator. Note that T T 6D 1, because T is invertible but nonunitary. Hence, 1 jTj2

is a nonzero positive operator. Define an operator W by W D jTj C i.1 jTj2 /1=2 and
observe that

W W D jTj2 C 1 jTj2 C i jTj.1 jTj2 /1=2 .1 jTj2 /1=2 jTj ; and

WW D jTj2 C 1 jTj2 C i jTj.1 jTj2 /1=2 C .1 jTj2 /1=2 jTj :

The operators jTj and 1 jTj2 obviously commute, and so jTj and f 1 jTj2
commute for every polynomial f . Hence, by the continuous functional calculus,
jTj and .1 jTj/1=2 commute, which implies that W W D WW D 1. Thus, the
operators U1 D UW and U2 D UW are unitary and distinct, and

1
1

.U1 C U2 / D U
.W C W / D U<.W/ D UjTj D T
2
2
expresses T as an average of unitaries U1 and U2 .
t
u
Returning to properties of the polar decomposition, a useful fact about the

modulus of a complex number z is that the real part <z of z satisfies <z jzj. This is
not true verbatim in the case of operators, but a very closely related property holds.
Proposition 10.70. If Z 2 B.H/, then <.Z/ VjZjV for some isometry V.
Proof. Let Z D UjZj be the polar decomposition of Z, where U is a partial isometry.
Decompose the hermitian operator <.Z/ as a difference <.Z/ D YC Y of positive
operators YC and Y such that YC Y D Y YC D 0. Let Q 2 B.H/ denote the
projection with range ran YC . Note that QY D Y Q D 0. Now let R D Q.Z C jZj/
and let R D WjRj be the polar decomposition of R in which W is a partial isometry
with final space ran R. Note that ran R ran Q, and so ran R ran Q. Therefore,
because the projection WW has range ran R, we deduce that ran .WW / ran Q.

?
The orthogonal complement of ran .WW / in H is ran R D ker R . If 2 ker R \
ran Q, then
0 D h; R i D hR; i D hQ.Z C jZj/; i D h.Z C jZj/; Qi D h.Z C jZj/; i
D < .h.Z C jZj/; i/ D h.<Z C jZj/; i h<.Z/; i D hYC ; i;
where the final inequality is on account of 2 ran Q D ran YC and QY D 0. Thus,
hYC ; i D 0 and YC positive yield YC D 0. However, ker YC ? ran YC implies that
YC D 0 only if D 0. Hence,

?
ran WW \ ran Q D ker R \ ran Q D f0g:
Therefore, WW Q and ran WW D ran Q together yield WW D Q.
362
Let P D W W, the projection with range .ker W/? D ran jRj. Thus,
ker W W \ ran Q D . ran jRj/? \ ran Q D ker R \ ran Q D f0g:
Thus, W Wjran Q is injective. Therefore, if 2 ker.1 WW / \ ran .1 W W/, then
D WW 2 ran Q and W W D 0 imply that D 0. Hence, the projection 1 Q is
injective on the range of 1 P.
Consider .1 Q/.1 P/. Because ker.1 Q/.1 P/ D ker.1 P/, in the polar
decomposition .1 Q/.1 P/ D W0 j.1 Q/.1 P/j the range of the projection
W0 W0 is ran .1 P/ and the range of W0 W0 is contained in ran .1 Q/. Because
ran W D ran R ran Q, we have .ran Q/? .ran W/? . Thus,
ran W0 ran .1 Q/ D ker Q D .ran Q/? .ran W/? ;
which implies that hW0 W; i D 0 for all ; 2 H. Hence, if V D W C W0 , then
V V D W W C W W0 C W0 W C W0 W0 D P C 0 C 0 C .1 P/ D 1:
That is, V is an isometry.
Recall that Z D UjZj, <.Z/ D YC Y , Q<.Z/Q D YC , R D Q.Z C jZj/, and
R D WjRj. Thus,
4YC D 2Q.Z C Z /Q D 2Q.UjZj C jZjU /
D Q .1 C U/jZj.1 C U/ .1 U/jZj.1 U/
Q.1 C U/jZj.1 C U/ Q:
Note that Q.1 C U/jZj.1 C U/ Q2 D Q.1CU/jZj.1CU/ Q.1CU/jZj.1CU/ Q.
Therefore, by the uniqueness of the positive square root, we obtain

1=2
4YC Q.1 C U/jZj.1 C U/ Q.1 C U/jZj.1 C U/ Q
:
Let X D .1 C U /Q and note that kXk kQk C kU k kQk 2. Thus, X X
kX Xk1 D kXk2 1 4 1, and so
Q.1 C U/jZj.1 C U/ Q.1 C U/jZj.1 C U/ Q D Q.1 C U/jZjX XjZj.1 C U/ Q

4 Q.1 C U/jZj2 .1 C U/ Q :
363
As the square root function is operator monotone, we now have that

4YC .Q.1 C U/jZj.1 C U/ Q.1 C U/jZj.1 C U/ Q/1=2

1=2
2 Q.1 C U/jZj2 .1 C U/ Q
D 2 .RR /1=2 :
Because ran W0 .ran W/? D .ran R/? D ker R , the operator R W0 D 0; thus,
W0 R D 0 and RV D R.W C W0 / D RW . By passing to adjoints, VR D WR .
Therefore,
RR D WVjRj2 W D W.R R/W D V.R R/V D VjZj.1 C U/ Q.1 C U/jZjV ;
and so RR 4VjZj2 V . Hence, using that the square root is operator monotone,
4YC 2 .RR /1=2 4.VjZj2 V /1=2
D 4.VjZjV VjZjV /1=2
D 4VjZjV :
Hence, <Z YC VjZjV .
t
u
An important consequence of the proposition above is the following triangle

inequality for Hilbert space operators.
Theorem 10.71 (Triangle Inequality). If S; T 2 B.H/, then there are isometries
V; W 2 B.H/ such that
jS C Tj VjSjV C WjTjW :
Proof. Let S C T D UjS C Tj be the polar decomposition of S C T, where U 2 B.H/
is a partial isometry. Therefore, U .S C T/ D jS C Tj and so
jS C Tj D < .jS C Tj/ D <.U S/ C <.U T/:
Because kU k D 1, we have that UU 1 and therefore X UU X X X for every
X 2 B.H/. Hence, by Proposition 10.54, jU Xj D .X UU X/1=2 .X X/1=2 D jXj
for each X 2 B.H/. Further, Proposition 10.70 asserts that there exist isometries V
and W such that <.U S/ VjU SjV and <.U T/ WjU TjW . Hence,
jS C Tj D <.U S/ C <.U T/ VjU SjV C WjU TjW VjSjV C WjTjW ;
t
u
364
10.7 Strong and Weak Operator Topologies

Recall from Proposition 1.88 that if X is a set, f.Y ; T /g2 is a family of
topological spaces, and if g W X ! Y is a function, for each 2 , then there
is a coarsest topology on X in which each function g W X ! Y is continuous. If
we let the set X be B.H/, D H, .Y ; T / D H (in the usual topology of H), and if
each g W B.H/ ! H is the function g .T/ D T, then we obtain a topology on B.H/
called the strong operator topology.
Definition 10.72. The strong operator topology on B.H/ is the coarsest topology
on B.H/ in which the functions g W B.H/ ! H defined by g .T/ D T, for T 2 B.H/,
are continuous for every 2 H.
Observe that in the strong operator topology (SOT) a basic open set containing a
given operator T0 2 B.H/ is a set of the form
U1 ;:::;m I"1 ;:::"m D fT 2 B.H/ j kTk T0 k k < "k for all k D 1; : : : ; mg;
for some m 2 N, 1 ; : : : ; m 2 H, and "1 ; : : : "m 2 .0; 1/. In particular, if fTk gk2N is a
sequence of operators such that, for some operator T 2 B.H/, limk kTk Tk D 0
for every 2 H, then T is the limit of the sequence fTk gk in the strong operator
topology. (Here, T is the limit of fTk gk means that for every SOT-open set U there
is an n0 2 N such that Tk 2 U for every k n0 .)
Another application of Proposition 1.88 leads to the weak operator topology.
Definition 10.73. The weak operator topology on B.H/ is the coarsest topology
on B.H/ in which the functions f; W B.H/ ! C defined by f; .T/ D hT; i, for
T 2 B.H/, are continuous for every .; / 2 H H.
In the weak operator topology, a basic open set containing T0 2 B.H/ is a set of
the form
W1 ;:::;m I1 ;:::;m I"1 ;:::"m D fT 2 B.H/ j jhTk T0 k ; k ij < "k for all k D 1; : : : ; mg;
for some m 2 N, 1 ; : : : ; m ; 1 ; : : : ; m 2 H, and "1 ; : : : "m 2 .0; 1/. Thus, if T 2 B.H/
and if fTk gk2N is a sequence of operators such that limk jhTk T; ij D 0 for every
; 2 H, then T is the limit of the sequence fTk gk in the weak operator topology.
One of the most useful features of the weak operator topology is the compactness
of the closed unit ball, which is an Alaoglu-type theorem in both its statement and
its method of proof.
Theorem 10.74. The set fT 2 B.H/ j kTk 1g is compact and Hausdorff in the
weak operator topology.
Proof. For each ordered pairY
.; / 2 H H, let K.;/ D f 2 C j j j kk kkg.
Consider the space K D
K.;/ , endowed with the product topology. By
.;/2HH
365
Tychonoffs Theorem (Theorem 2.14), K is a compact set. Furthermore, K is

Hausdorff because each K.;/ is Hausdorff.
Let X D fT 2 B.H/ j kTk 1g and consider X as a topological space in which the
topology on X is induced by the weak operator topology of B.H/. Define f W X ! K
by f .T/ D .hT; i/.;/ , and note that f is an injective function. Select T 2 X and
consider an open set W K that contains f .T/. Thus, there are open subsets W.;/
K.;/ such that hT; i 2 W.;/ , for every .; / 2 H H, and W.;/ D K.;/ for all
but at most
Y a finite number of elements in H Hsay .1 ; 1 /,. . . , .n ; n /and
WD
W.;/ . Hence there are positive real numbers "1 ; : : : ; "n such that
.;/2HH
W.j ;j / D fz 2 C j jz hT; j ij < "j g

for every j D 1; : : : ; n. Therefore,
f 1 .W/ D
n
\
fS 2 X j jh.S T/j ; j ij < "j g;
jD1
which is a basic WOT-open neighbourhood of T 2 X. Hence, f is continuous at every

T 2 X, which implies that f is a continuous function on X.
On the other hand, if U X is an arbitrary open set and if T 2 U, then
there is a basic WOT-open set B such that T 2 B U. By definition, there are
.1 ; 1 /; : : : ; .n ; n / 2 H H and positive real numbers "1 ; : : : ; "n such that, for
S 2 X, we have S 2 B if and only if jh.S T/j ; j ij < "j for each j D 1; : : : ; n.
Hence, if W.j ;j / D f 2 K.j ;j / j j hTj ; j ij < "j g and if W.;/ D K.;/ for every
Y
W.;/ is open in K and
.; / 2 H H n f.1 ; 1 /; : : : ; .n ; n /g, then WT D
f .T/ 2 WT f .U/. Thus, f .U/ D
[
T2U
.;/2HH
WT , which shows that f .U/ is open. Hence,
f 1 W f .X/ ! X is continuous, and therefore f is a homeomorphism between X and

f .X/.
We now show that f .X/ is a closed subset of K. Let D . .;/ /.;/ 2 K be in the
closure of f .X/, and define a function W H H ! C by .; / D .;/ . Claim: is
a bounded sesquilinear form. To prove this claim, select vectors 0 ; 1 ; 2 ; 0 ; 2 ; 2 2
H and scalars 1 ; 2 ; 1 ; 2 2 C. Let " > 0 be arbitrary. Define subsets W.;/ K.;/
as follows:
366

W.1 1 C2 2 ;0 / D z 2 K.1 1 C2 2 ;0 / j jz .1 1 C2 2 ;0 / j < " I

W.0 ;1 1 C2 2 / D z 2 K.0 ;1 1 C2 2 / j jz .0 ;1 1 C2 2 / j < " I
n
W.0 ;j / D z 2 K.0 ;j / j jz .0 ;j / j <
"
jj j
n
W.j ;0 / D z 2 K.j ;0 / j jz .j ;0 / j <
"
jj j
o
o
; for j D 1; 2I
; for j D 1; 2I
W.;/ D K.;/ in all other cases:

Thus, W D
W.;/ is open in K and contains . Because is in the closure
.;/2HH
of f .X/, there is an operator S 2 X with f .S/ 2 W. Therefore,

j1 .1 ; 0 / 1 hS1 ; 0 ij < ";
j2 .2 ; 0 / 2 hS2 ; 0 ij < ";
and
j .1 1 C 2 2 ; 0 / hS.1 1 C 2 2 /; 0 ij < ":
Hence,
j .1 1 C 2 2 ; 0 / 1 .1 ; 0 / 2 .2 ; 0 /j < 3":
A similar argument shows that
j .0 ; 1 1 C 2 2 / 1 .0 ; 1 / 2 .0 ; 2 /j < 3":
The choice of " > 0 being arbitrary yields
.1 1 C 2 2 ; 0 / D 1 .1 ; 0 / C 2 .2 ; 0 /; and
.0 ; 1 1 C 2 2 / D 1 .0 ; 1 / C 2 .0 ; 2 /:
Hence, is a sesquilinear form.
The boundedness of
is immediate from j .; /j kk kk. By Proposition 10.5, there is a unique T 2 B.H/ such that .; / D hT; i for all ; 2 H.
Because kTk is the supremum of all jhT; ij as and range through unit vectors,
kTk 1. This proves that T 2 X and, hence, that D f .T/ 2 f .X/.
Because f .X/ is closed in K and since K is compact and Hausdorff, we deduce
that f .X/ is compact and Hausdorff; hence, X is compact and Hausdorff.
t
u
The following example helps distinguish the two topologies on B.H/.
367
Example 10.75. The involution T 7! T is continuous with respect to the weak

operator topology, but not with respect to the strong operator topology.
Proof. The proof of the first assertion is left as an exercise (Exercise 10.142).
Let S denote the unilateral shift operator on the Hilbert space `2 .N/, and let
Tn D .S /n , for every n 2 N. Note that if D .k /k2N 2 H and n 2 N, then
kTn k D
1
X
!1=2
jk j
kDnC1
Thus, lim kTn k D 0, for every 2 `2 .N/; that is, the sequence fTn gn converges
n!1
to 0 with respect to the strong operator topology. However, kTn k D kSn k D kk,
because S is an isometry, and so 0 is not the SOT-limit of the sequence fTn gn ,
implying that the involution fails to be continuous with respect to the strong operator
topology.
t
u
In contrast to Example 10.75, B.H/ admits the same set of continuous linear
maps into C regardless of whether B.H/ has the strong operator topology or the
weak operator topology.
Proposition 10.76. The following statements are equivalent for a linear transformation ' W B.H/ ! C:
1. ' is continuous with respect to the weak operator topology on B.H/;
2. ' is continuous with respect to the strong operator topology on B.H/;
3. there exist n 2 N and nonzero vectors 1 ; : : : ; n ; 1 ; : : : ; n 2 H such that
'.T/ D
n
X
hTj ; j i;
jD1
for every T 2 B.H/.

Proof. (1) ) (2). Assume that ' is continuous with respect to the weak operator
topology on B.H/. Suppose that V C is a nonempty open set, and select T0 2 B.H/
and " > 0 so that B" .z0 / V, where z0 D '.T0 /. Because ' is weakly continuous,
there is a basic WOT-open set W about T0 , say
W D fT 2 B.H/ j jhTj T0 j ; j ij < "j for all j D 1; : : : ; ng;
for some nonzero vectors 1 ; : : : ; n ; 1 ; : : : ; n and positive real numbers "1 ; : : : "n ,
such that '.W/ B" .z0 /. For each j let "Qj D kj k1 ", and consider the SOT-open set
UT0 D fT 2 B.H/ j kTj T0 j k < "Qj for all j D 1; : : : ; ng;
368
By the Cauchy-Schwarz inequality, if T 2 W for every T 2 UT0 . Hence, ' 1 .V/ is a

union of SOT-open sets UT0 , and therefore ' is continuous with respect to the strong
operator topology.
(2) ) (3). Assume that ' is continuous with respect to the strong operator
topology on B.H/. Therefore, using the open unit disc D in C, the set ' 1 .D/ is
SOT-open, and hence there exists a basic SOT-open set U about 0 2 B.H/ of the
form U D U1 ;:::;n I"1 ;:::"n for some nonzero j 2 H and "j > 0. Let " D minj "j ; thus, if
T 2 B.H/ satisfies kTj k < " for each j, then j'.T/j < 1.
Let C D 2" and suppose that R 2 B.H/ satisfies Rj 6D 0 for at least one j. Let
n
X
DC
!1=2
kRk k
kD1
Thus, for any j,

1
Rj P "kRj k " < ";

n
2 1=2
2
2
kD1 kRk k
and therefore j'.R/j < . Hence, for every R 2 B.H/,

!1=2
n
X
2
kRk k
:
j'.R/j C
kD1
By replacing each k with Ck in the inequality above, we may assume without
further change of notation that
!1=2
n
X
2
j'.R/j
kRk k
;
kD1
for every R 2 B.H/.
n
M
In the Hilbert space H .n/ D
H (the direct sum of n copies of H), consider the
1
L
linear submanifold L0 D f j Tj j T 2 B.H/g. If S; T 2 B.H/ are such that Sj D Tj
for j D 1; : : : ; n, then
!1=2
n
X
2
j'.S/ '.T/j D j'.S T/j
k.S T/k k
D 0:
kD1
Therefore, the map

j Tj 7! '.T/ is well defined, linear, and contractive. Thus,
by the Hahn-Banach Theorem, there is a contractive linear functional on H such
0
@
M
j
1
Tj A D '.T/;
10.8 Matrices of Operators
369
for all T 2 B.H/. The Riesz Representation Theorem yields a vector D

H .n/ that implements . In particular, for each T 2 B.H/,
*
'.T/ D
n
M
Tj ;
jD1
n
M
+
j D
jD1
n
X
j j
in
hTj ; j i:
jD1
The proof of (3) ) (1) is obvious.
t
u
Equipped with Proposition 10.76 and the Hahn-Banach Separation Theorem, the
following fundamental fact about B.H/ is deduced.
Proposition 10.77. If K B.H/ is a convex set, then C WOT D C SOT .
Proof. It is clear that C SOT C WOT . To prove the inclusion C WOT C SOT , select
T 2 C WOT . If, contrary to what we aim to prove, T 62 C SOT , then the Hahn-Banach
Separation Theorem implies that there are a SOT-continuous ' W B.H/ ! C and a
2 R such that
< .'.R// < '.T/; 8 R 2 C:
But Proposition 10.76 implies that ' is also WOT-continuous, and so the inequality
above implies that T 62 C WOT , which is a contradiction.
t
u

Through the use of matrices of operators, a number of properties concerning
individual operators are revealed.
Proposition 10.78. If T 2 B.H/, then the operator

1 T
AD
T 1
is a positive operator on H H if and only if kTk 1.

Proof. By Exercise 10.128, if Q is hermitian and X is invertible, then
Q is positive
1 T
as
if and only if X QX is positive. Factor the hermitian operator A D
T 1

AD
1 0
T 1
1
0
0 1 T T

1T
D X QX:
01
370
1T
01
Because X D
is invertible, we have that A is positive if and only if the

1
0
matrix Q D
is positive. But Q is positive if and only if 1 T T is
0 1 T T
positive, which is equivalent to saying that Q is positive if and only if kTk 1, by
Proposition 10.49.
t
u
A 3 3 version of Proposition 10.78 is:
Proposition 10.79. If T1 ; T2 2 B.H/, then the operator
2
3
1 T1 0
A D 4 T1 1 T2 5
0 T2 1
is a positive operator on H H H if and only if 1 T1 T1 T2 T2 is positive.
Proof. Factor A as
2
32
32
3
1 00
1 T1 0
1
0
0
A D 4 T1 1 0 5 4 0 .1 T1 T1 / T2 5 4 0 1 0 5 :
0
T2
1
0 01
0 0 1
Thus, A is positive ifand only if the middle factor is positive, which in turn is
.1 T1 T1 / T2
is positive. This matrix is equal to
positive if and only if
T2
1

01
10
1 0
T2 1
1
0
0 .1 T1 T1 T2 T2 /
1 T2
0 1

01
;
10
which is positive if and only if 1 T1 T1 T2 T2 is positive.
t
u
The next theorem is one of the first ever dilation, or matrix completion,
theorems established for Hilbert space operators. Below,

given a contraction T,
T
the unspecified entries in the 2 2 operator matrix
are determined so that

the completed matrix is a unitary operator.
Proposition 10.80 (Halmos). If T 2 B.H/ satisfies kTk 1, then the matrix

.1 TT /1=2
T
UD
1=2
.1 T T/
T
is a unitary operator on H H.
371
Proof. By Proposition 10.49, the condition kTk 1 is equivalent to the positivity of

1 T T. Because kT k D kTk 1, we also have that 1 TT is positive. Therefore,
in the definition of the matrix U, the (1,2) and (2,1) entries are well defined.
Computation of U U and UU leads to

TT C .1 TT /
T .1 T T/1=2 .1 TT /1=2 T
:
.1 T T/1=2 T T .1 TT /1=2
.1 T T/ C T T
T T C .1 T T/
T .1 TT /1=2 .1 T T/1=2 T
U UD
1=2
1=2
.1 TT / T T .1 TT /
.1 TT / C TT
and
UU D
Therefore, it is enough to prove that T .1 T T/1=2 D .1 TT /1=2 T. Let A and B

denote the positive contractions A D .1 T T/1=2 and B D .1 TT /1=2 . Observe
that
TA2 D T TT T D B2 T:
Thus, Tf .A2 / D f .B2 /T for every polynomial f 2 C t. Hence, if ffn gn2N is
a sequence of polynomials pconverging uniformly on the interval 0; 1 to the
square-root function h.t/ D t, then Th.A2 / D h.B2 /T; that is, T .1 T T/1=2 D
.1 TT /1=2 T, which proves that U U D UU D 1 2 B.H H/.
t
u
Proposition 10.80 has numerous interesting applications, one of which concerns
the weak operator topology.
Proposition 10.81. If H is an infinite-dimensional Hilbert space, then the closure
of the set fU 2 B.H/ j Uis unitaryg in the weak operator topology of B.H/ is the set
of all T 2 B.H/ for which kTk 1.
Proof. Select T 2 B.H/ such that kTk 1. Consider a basic WOT-open set W
containing T, which by the definition of the weak operator topology is a set of the
form
WD
m
\
fS 2 B.H/ j jh.S T/j ; j ij < "j g;
jD1
for some m 2 N, 1 ; : : : ; m ; 1 ; : : : ; m 2 H, and "1 ; : : : "m 2 .0; 1/. We aim to prove

that W contains some unitary operator U.
Let H0 D Spanf1 ; : : : ; m ; 1 ; : : : ; m g. Because H has infinite dimension, we may
consider the finite-dimensional Hilbert space H0 H0 as a subspace of H; hence,
H decomposes as H D .H0 H0 / H1 , where H1 D .H0 H0 /? . Let P 2 B.H/
denote the projection with range H0 and consider the contraction PTP acting on
H0 . By Proposition 10.80, there is a unitary operator U0 2 B.H0 H0 / such that
372

.PTP/ X
U0 D
. Extend U0 to a unitary operator U acting on H D .H0 H0 /H1 ,
Y Z
where

UD
U0 0
0 1jH1
3
.PTP/ X 0
D 4 Y Z 0 5:
0 0 1jH1
Note that PUP D PTP and so, for each j D 1; : : : ; m,

hUj ; j i D hUPj ; Pj i D hPUPj ; j i D hPTPj ; j i D hTPj ; Pj i D hTj ; j i:
WOT
Thus, U 2 W. Hence, fT 2 B.H/ j kTk 1g fU 2 B.H/ j Uis unitaryg

.
Conversely, if T 2 B.H/ satisfies kTk > 1, then there are unit vectors ; 2 H
such that jhT; ij > 1. On the other hand, jhU; ij kUk kk D kk kk D 1 for
WOT
every unitary operator U. Hence, T 62 fU 2 B.H/ j U is unitaryg
.
t
u
The following two results of this section are in the spirit of Propositions 10.80
and 10.81.
Proposition 10.82. If A 2 B.H/ is positive and kAk 1, then

.A.1 A//1=2
A
PD
1=2
.A.1 A//
1A
is a projection operator on H H.
t
u
Proposition 10.83. If H is an infinite-dimensional Hilbert space, then the closure

of the set fP 2 B.H/ j Pis a projectiong in the weak operator topology of B.H/ is the
unit operator interval I.H/.
t
u
One of the most striking applications of matrices of operators involves an infinite

matrix. Let H be a Hilbert space and suppose that T 2 B.H/ is a contraction.
Consider the Hilbert space `2H .Z/ of sequences D .n /n2Z of vectors n 2 H for
which
X
n2Z
kn k2 D lim
k!1
k
X
kn k2 < 1:
nDk
P
(The inner product is h.xn /n ; ./n i D n2Z hn ; n i.) With respect to this sequence
2
space, consider the operator U W `H .Z/ ! `2H .Z/ defined by the following lowertriangular infinite matrix, with entries indexed by Z Z, of operators acting on H:
10.9 Singular Values and Trace-Class Operators
373
::
7
6 :
7
6:
7
6 :: 0
7
6
7
6
1 0
7
6
7;
UD6
X T
7
6
7
6

7
6
T Y 0
7
6
7
6
1
0
5
4
:: ::
: :
where X D .1TT /1=2 and Y D .1T T/1=2 , and where the operator T is the (0,0)entry of the matrix U, all remaining diagonal entries are 0, all subdiagonal entries
are 1 except for X and Y, and all other matrix entries are 0. By using arguments
like those employed in the proof of Proposition 10.80, one sees that the operator U
satisfies U U D UU D 1.
Let P0 be the projection on `2H .Z/ with range given by the 0-th copy of H in
2
`H .Z/. Thus, P0 Ujran P0 D T. Moreover, because U is in lower-triangular form,
P0 .U k /jran P0 D T k for every positive integer k. This leads to the following important
theorem.
Theorem 10.84 (Sz.-Nagy Dilation Theorem). If T 2 B.H/ satisfies kTk 1,
Q that contains H as a subspace and a unitary operator
then there is a Hilbert space H
Q
U on H such that
P.U k /jH D T k
Q is the projection with range H.
for every positive integer k, where P 2 B.H/
Corollary 10.85 (von Neumanns Inequality). If T 2 B.H/ satisfies kTk 1, then
kf .T/k 1 for every polynomial f 2 C t for which max jf .z/j 1.
jzj1
t
u

The spectral theory of compact hermitian operators leads to a general structure
theorem for arbitrary compact Hilbert space operators known as the singular value
decomposition.
Theorem 10.86 (Singular Value Decomposition). If K 2 B.H/ is a compact
operator of rank r 2 N [ f1g, then there exist a sequence fsj grjD1 of real numbers
and orthonormal sets f j grjD1 and f j grjD1 of vectors such that
1. sj sjC1 > 0, for all j,
374
2. 0 D limj sj , if r D 1, and
r
X
3. K D
sj h; j i j , for every 2 H.
jD1
Proof. Apply Theorem 10.32 to the compact positive operator K K to obtain

0
a bounded sequence f j grjD1 of (nonzero) positive real numbers and a sequence
0
f j grjD1 of pairwise-orthogonal unit vectors in H such that K K j D j j , for each
0
j, limj j D 0, if r is infinite, and K K D
r
X
j h; j i j , for every 2 H. Note
jD1
r q
X
that r D rank .K K/. Furthermore, if R 2 B.H/ is given by R D
j h; j i j ,
0
jD1
for every 2 H, then R is compact, positive, and R2 D K K. Therefore,

by the
p
uniqueness of the positive square root, R D .K K/1=2 ; thus, if sj D j for each j,
then
r q
X
jKj D
j h; j i j ;
0
jD1
for every 2 H. Hence, if K D VjKj is the polar decomposition of K, and if

for each j, then
D V j
K D
r
X
sj h; j i j ;
jD1
for all 2 H. Because the range of V is isometric on ran jKj and has range ran K,
0
f k grkD1 is a set of orthonormal vectors and r0 D rank K.
u
t
The singular decomposition of a compact operator K refers to the representation
in (3) of Theorem 10.86 of the action of K on the Hilbert space H.
Definition 10.87. The singular values of a compact operator K of rank r 2 N[f1g
acting on a separable Hilbert space of dimension d 2 N [ f1g are the nonnegative
real numbers sj .K/ defined by
sj .K/ D sj ;
if 1 j r and where fsj grjD1 are the positive numbers arising in the singular value
decomposition (3) of K, and by
sj .K/ D 0;
if j 2 f1; : : : ; dg is such that j > r.
375
Using the notation to denote rank-1 operators (of the form 7! h; i ),
the singular value decomposition of a compact operator K can be expressed as
KD
d
X
sj .K/ j
j:
(10.13)
jD1
In the case where H has infinite dimension, the series
d
X
sj .K/ j
converges in
jD1
the norm of B.H/ to K. Indeed, if 2 H and N 2 N, then

N
X

K
sj .K/. j

jD1
2

X

sj .K/2 jh;
j / D

j>N
j ij
sN .K/2 kk2 :
Thus,

N
X

K
sj .K/. j

jD1

/
j sN .K/;

which converges to zero because lim sN .K/ D lim

N!1
N!1
N .K K/ D 0. Hence, the
following proposition has been proved.

Proposition 10.88. If H is a separable Hilbert space, then the algebraic ideal F.H/
of finite-rank operators on H is dense in the ideal K.H/ of compact operators on H.
The following elegant application of Proposition 10.88 returns us to the notion
of complementation (see Definition 8.15) in Banach space theory.
Proposition 10.89 (Conway). If H is an infinite-dimensional separable Hilbert
space, then the subspace K.H/ is not complemented in B.H/.
Proof. Fix an orthonormal basis f n gn2N of H. Let `1 and c0 denote `1 .N/ and
c0 .N/. For each 2 `1 , define M on H by
M D
1
X
.n/h; n i;
nD1
for 2 H. The linear map

W `1 ! B.H/ in which
. / D M is plainly linear and
isometric. Now if f 2 `1 has the property that f .n/ 6D 0 for at most a finite number
of n 2 N, then Mf has finite rank. Such functions are dense in c0 ; therefore, if f 2 c0
and if ffk gk2N is sequence in `1 converging to f and such that, for each k, fk .n/ 6D 0
for at most a finite number of n 2 N, then limk kMf Mjk k D limk kf fk k D 0 shows
that Mf is a compact operator.
376
Let W B.H/ ! `1 be the contractive linear map defined by .T/ D .hT n ; n i/n .
Suppose that K is a rank-1 operator, say K D , for some nonzero vectors
and . Then, for a fixed n, hK n ; n i D h n ; ih; n i. Because h ; n i and h; n i
of the n-th Fourier coefficients of
and , respectively, these complex numbers
converge to 0 as n ! 1. Hence, .K/ 2 c0 for every rank-1 operator K, and so by
linearity .K/ 2 c0 for all finite-rank operators K. Because F.H/ is dense in K.H/,
and because is continuous and c0 is closed in `1 , we deduce that .K/ 2 c0 for
every K 2 K.H/.
Assume, contrary to what we aim to prove, that K.H/ is complemented in B.H/.
Hence, there exists an idempotent operator E W B.H/ ! B.H/ with range K.H/
(Proposition 8.16). Define now an operator E W `1 ! `1 by E D E
and
observe that E is an idempotent with range c0 . Therefore, by Proposition 8.16, c0
is a complemented subspace of `1 , which is in contradiction to Proposition 8.20.
Therefore, it cannot be that K.H/ is complemented in B.H/.
t
u
Returning to the study of singular values, we begin with two basic properties.
Proposition 10.90. If K; K1 ; K2 ; S; T 2 B.H/ and if K; K1 ; K2 are compact, then
1. sj .jKj/ D sj .K/ D sj .K / and
2. sj .SKT/ kSk kTksj .K/
for every j.
Proof. Note that the singular values of K are simply the square roots of the
eigenvalues of K K, labelled in non-ascending order. Thus,
sj .jKj/ D j .jKj/ D sj .K/
for every j.
Suppose that K D VjKj is the polar decomposition of K. Thus, KK D
V.K K/V . Because
ran K K D ran jKj2 ran jKj;
V V.K K/ D K K. Hence, KK D V.K K/V yields f .K K/ D Vf .K K/V for
all polynomials f 2 R t, and so jK j D VjKjV by continuous functional calculus.
Suppose now that jKj D for some 2 .0; 1/ and nonzero 2 H. Thus,
2 ran jKj. Hence, if D V, then k k D kk and V V D . Thus,
jK j
D VjKjV
D VjKjV V D VjKj D V D :
That is, is a nonzero eigenvalue of jKj if and only if is a nonzero eigenvalue of

jK j.
The argument above shows that if f1 ; : : : ; m g is an orthonormal basis for
ker.jKj 1/, then fV1 ; : : : ; Vm g is an orthonormal subset of ker.jK j 1/, and so
the multiplicity of as an eigenvalue of jKj is bounded above by the multiplicity of
377
as an eigenvalue of jK j. By letting L D K and invoking the argument again

shows that ker.jKj 1/ and ker.jK j 1/ have the same dimension. Hence,
sj .K/ D sj .K / for every j.
To prove the second assertion, because
jSKj2 D .SK/ .SK/ D K .S S/K kSk2 K K D kSk2 jKj2 ;
Theorem 10.54 implies that jSKj kSk jKj. Therefore, the min-max variation
principle (equation (10.11)) in the Courant-Fischer Theorem (Theorem 10.58)
yields
sj .SK/ D j .jSKj/ kSk j .jKj/ D sj .K/
for every j. Thus,
sj .SKT/ kSksj .KT/ D kSksj .T K / kSk kT ksj .K / D kSk kTksj .K/
t
u
for every j.
Definition 10.91. A compact operator K acting on a separable Hilbert space of

dimension d 2 N [ f1g is a trace-class operator if
d
X
sj .K/ < 1:
jD1
Let T.H/ denote the set of trace-class operators acting on a separable Hilbert
space H. Thus, we have
T.H/ K.H/ B.H/:
If H has infinite dimension, then the inclusions above are sharp. The proper
inclusion of K.H/ into B.H/ has already been noted (as the identity operator is not
compact), and so consider the inclusion T.H/ K.H/. Select an orthonormal basis
f j gj2N and consider the compact positive operator K for which K j D j1 j for all
1
X
1
j 2 N. Because sj .K/ D j for each j, the sum
sj .K/ diverges, and so K 62 T.H/.
jD1
On the other hand, observe that

K 2 T.H/ jKj 2 T.H/ K 2 T.H/
and that RKS 2 T.H/ if K 2 T.H/ and R; S 2 B.H/.
The use of the adjective trace in Definition 10.91 above will be explained
shortly, but note that every operator of finite rank is a trace-class operator; in
378
particular, if H has finite dimension, then every operator on H is a trace-class

operator.
In linear algebra, the trace of an n n matrix T D tij ni;jD1 is defined by Tr T D
n
X
tii , the sum of the diagonal elements of T. If T is the matrix representation of an
iD1
operator on an n-dimensional Hilbert space with respect to some orthonormal basis

f i gniD1 of H, then
Tr T D
n
X
hT i ; i i:
iD1
This motivates the definition of trace for operators on infinite-dimensional separable

Hilbert spaces, starting first with the cone B.H/C of positive operators.
Definition 10.92. Let B D f i g1
iD1 denote an orthonormal basis of an infinitedimensional separable Hilbert space H. The function B W B.H/C ! 0; 1
defined by
B .A/ D
1
X
hA i ; i i:
iD1
is called a canonical tracial weight on B.H/C .

Besides its linearity, a distinguishing property of the trace of matrices is that
Tr .ST/ D Tr .TS/ for all matrices S and T.
Proposition 10.93. If B is a given orthonormal basis of an infinite-dimensional
separable Hilbert space H, then
1.
2.
3.
4.
B .1 A1 C 2 A2 / D 1 B .A1 / C 2 B .A2 /, for all Aj 2 B.H/C and j 2 RC ,

B .TT / D B .T T/, for every T 2 B.H/,
B .U AU/ D B .A/, for every positive operator A and unitary operator U, and
B D B0 for every orthonormal basis B 0 of H.
Proof. It is clear that the property B .1 A1 C 2 A2 / D 1 B .A1 / C 2 B .A2 / holds,

for all A1 ; A2 2 B.H/C and 1 ; 2 2 RC .

Let B D f i g1
iD1 . By way of the Fourier series decompositions of T i and T j ,
B .TT / D
D
i hTT
i ; i i D
P DP
i
j hT
E PP
2
i ; j iT j ; i D i j hT j ; i i
2 P
P
PP
P
j
i hT j ; i i D
j T j ;
i hT j ; i i i D
i hT T i ; i i
D B .T T/:
Hence, B .TT / D B .T T/.
379
Suppose that A is positive and U is unitary; let T D A1=2 U. Thus,

B .U AU/ D B .U A1=2 A1=2 U/ D B .T T/
D B .TT / D B .A1=2 UU A1=2 /
D B .A/:
Lastly, if B 0 D f i0 gdiD1 is an orthonormal basis of H, then let U 2 B.H/ be the
unitary operator for which U i D i0 for every i 2 f1; : : : ; dg. Hence, for every A 2
B.H/C ,
B .A/ D B .U AU/ D B0 .A/;
t
u
In light of Proposition 10.93, we may drop the reference to the orthonormal basis
B when discussing a canonical tracial weight B since there is exactly one such
functionand denote the canonical tracial weight on B.H/ by .
The domain of definition of can be extended to the R-vector space B.H/sa of
hermitian operators as follows. For any difference C D A1 A2 of positive operators
A1 and A2 , define .C/ by
.C/ D .A1 / .A2 /:
This is a well-defined function on B.H/sa because, if A1 A2 D A01 A02 , then A01 C
A2 D A1 C A02 2 B.H/C and therefore
.A01 / C .A2 / D .A01 C A2 / D .A1 C A02 / D .A1 / C .A02 /I
hence, .A1 / .A2 / D .A01 / .A02 / in the extended real number system.
Definition 10.94. If H is a separable Hilbert space, then the canonical trace on
B.H/ is the function Tr on B.H/ defined by
Tr ..A1 A2 / C i.B1 B2 // D . .A1 / .A2 // C i ..B1 / .B2 // ;
for all positive operators A1 ; A2 ; B1 ; B2 .
Alternatively, the canonical trace on B.H/ is the map T 7!
(and every) orthonormal basis f i g1
iD1 of H.
1
X
hT i ; i i for some
iD1
380
Proposition 10.95. If H is an infinite-dimensional separable Hilbert space and if

K 2 B.H/ is a trace-class operator, then the sum
X
hK!; !i
(10.14)
!2B
is absolutely convergent in C for every orthonormal basis B of H.

Proof. By the singular value decomposition of K, there exist orthonormal sets
1
f j g1
jD1 and f j gjD1 in H such that
KD
1
X
sj .K/ j
j:
jD1
If the set B0 D f j g1
jD1 , which is an orthonormal basis for ran jKj, is not already an
orthonormal basis of H, then it may be extended to one, say B D B0 [ B1 , where
B1 is an orthonormal basis of .ran jKj/? D ker jKj D ker K. Thus, K! D 0 for every
! 2 B1 , and so

r
X
X
hK!; !i D
hK!; !i D hK i ; i i
iD1
!2B
!2B0
+
r * r
X X
D
sj .K/h i ; j i j ; i
iD1 jD1
1
X
si .K/ jh i ; i ij
iD1
d
X
si .K/ < 1:
iD1
t
u
Hence, the sum (10.14) is absolutely convergent.
Corollary 10.96. If K is a trace-class operator acting on a separable Hilbert space

H of dimension d 2 N [ f1g, then
jTr Kj Tr jKj D
d
X
sj .K/:
jD1
Proof. The proof of Proposition 10.95 shows that

jTr Kj
r
X
iD1
si .K/:
381
However, because i .jKj/ D si .jKj/ D si .K/ for all i such that 1 i r, this latter
sum is precisely Tr jKj by the spectral theorem (Theorem 10.32). Since sj .jKj/ D 0
d
X
for r < j d,
sj .jKj/ D Tr jKj.
t
u
jD1
Every rank-1 operator is a trace-class operator, and therefore T. / is

also a trace-class operator, for every T 2 B.H/. The traces of such operators (of rank
0 or 1) are easily determined as follows.
Example 10.97. If ;
2 H are nonzero and if T 2 B.H/ is arbitrary, then

Tr .T. // D hT ; i:
Proof. We may assume that k k D 1, for if not we could replace with Q D k k

and with Q D k k1 to obtain D Q Q . Select an orthonormal basis
f j gj of H in which 1 D . Thus,
Tr .T. // D
X
X

T h j ; i ; j D
h j ; ihT ; j i D hT ; i;
j
t
u
which completes the calculation.

In particular, the computation above yields k k1 D jh ; ij.
The first major result about trace-class operators is the following theorem.
Theorem 10.98. The set T.H/ of all trace-class operators acting on a separable
Hilbert space H of dimension d 2 N [ f1g is an algebraic ideal of B.H/ and the
function k k1 W T.H/ ! R defined by
kKk1 D
d
X
sj .K/
(10.15)
jD1
is a norm on T.H/. Furthermore, with respect to the norm k k1 , T.H/ is a separable

Banach space and the algebraic ideal F.H/ of finite-rank operators is dense in
T.H/.
Proof. We shall assume for the proof that H has infinite dimension.
Proposition 10.90 shows that T.H/ is closed under scalar multiplication, the
involution , and under products of the form RKS, where K 2 T.H/ and R; S 2 B.H/.
All that remains, therefore, to show that T.H/ is an algebraic ideal is to show
that K1 C K2 2 T.H/ for every K1 ; K2 2 T.H/. To this end, let K1 ; K2 2 T.H/ and
consider jK1 CK2 j. By the Triangle Inequality (Theorem 10.71), there are isometries
V; W 2 B.H/ such that
jK1 C K2 j VjK1 jV C WjK2 jW D X X C Y Y;
382
where X D jK1 j1=2 V and Y D jK2 j1=2 W . Proposition 10.93 asserts that .G G/ D
.GG / for every G 2 B.H/, and therefore
.jK1 C K2 j/ .X X C Y Y/ D .X X/ C .Y Y/ D .XX / C .YY /:
Because XX D jK1 j and YY D jK2 j, the inequality above yields
Tr jK1 C K2 j Tr jK1 j C Tr jK2 j:
By Corollary 10.96,
d
X
sj .K1 C K2 / D Tr jK1 C K2 j Tr jK1 j C Tr jK2 j D
jD1
d
X
sj .K1 / C
jD1
d
X
sj .K2 /:
jD1
Thus, K1 C K2 2 T.H/ and kK1 C K2 k1 kK1 k1 C kK2 k1 . Hence, T.H/ is an

algebraic ideal of B.H/ and k k1 satisfies the triangle inequality on T.H/.
If K 2 T.H/ satisfies kKk1 D 0, then the only eigenvalue of jKj is 0, and so
the spectral radius of jKj is 0. Hence, jKj D 0 and therefore K D 0. The property
kKk1 D jj kKk1 is trivial. Since the triangle inequality was established in the
previous paragraph, k k1 is a norm on T.H/.
Suppose that fKn gn2N is a Cauchy sequence in T.H/. Because
kKn Km k1 s1 .Kn Km / D kjKn Km jk D kKn Km k;
fKn gn2N is a Cauchy sequence in B.H/ and is, hence, convergent to some K 2 B.H/.
Because each Kn is compact and K.H/ is norm-closed, the limit operator K must
also be compact. Furthermore,
kKn Kn K Kk D kKn .Kn K/ C .Kn K/ Kk kKn kkKn Kk C kKn K kkKk
implies that Kn Kn converges to K K. Thus, by the continuous functional calculus,
kjKn j jKjk ! 0.
Let f j gj be an orthonormal set of eigenvectors of jKj corresponding to the
nonzero eigenvalues j .jKj/ of K. If N 2 N, then
N
X
jD1
sj .K/ D
N
X
j .jKj/ D
jD1
D lim
N
X
hjKj j ; j i
jD1
n!1
N
X
hjKn j j ; j i lim inf Tr jKn j
jD1
D lim inf kKn k1 :

n
383
Now since limn kKn k1 exists, lim inf kKn k1 D lim kKn k1 , and so the sum
n
N
X
sj .K/ is
jD1
bounded above for all N. Hence, K is a trace-class operator.

If " > 0, then there exists N 2 N such that kKn K` k1 < " for all `; n N. Thus,
if ` N, then
kK K` k1 D lim kKn K` k1 < ":
n!1
Hence, fKn gn2N converges in T.H/ to K, thus proving that T.H/ is a Banach space.
To show that F.H/ is dense in T.H/, let K 2 T.H/ have infinite rank; thus, the
positive trace-class operator jKj also has positive rank. Express jKj in its spectral
decomposition:
jKj D
1
X
j .jKj/ j j ;
jD1
for some orthonormal set
f j g1
jD1 .
Consider jKjN D
N
X
sj .K/ j .jKj/ j j , which
jD1
is a positive operator of finite rank and satisfies jKjN jKj in the Loewner ordering.
Thus,
X
j .jKj/:
kjKj jKjN k1 D Tr .j jKj jKjN j/ D Tr .jKj jKjN / D
j>N
1

Because j .jKj/ jD1 2 `1 .N/, the partial sums of the eigenvalues converge to the
X
j .jKj/ D 0 and lim kjKj jKjN k1 D 0.
sum of the eigenvalues; hence, lim
N!1
N!1
j>N
Express K in its polar decomposition: K D VjKj and let GN D VjKjN , which is

an operator of finite rank. Thus,
kK GN k1 D kV.jKj jKjN /k1 kVkkjKj jKjN k1 D kjKj jKjN k1 ;
and so lim kK GN k1 D 0.
N!1
The proof of the separability of T.H/ is left as an exercise (Exercise 10.145).
t
u
The norm k k1 on T.H/ is called the trace norm. Note that the proof of
Theorem 10.98 shows that if
KD
1
X
jD1
sj .K/ j
384
is the singular value decomposition of K 2 T.H/, then the sum converges in the trace
norm.
Because T.H/ is a Banach space, it is of interest to understand its dual. The
following theorem shows that the dual space of T.H/ is (isometrically isomorphic
to) B.H/.
Theorem 10.99. If H is a separable Hilbert space, then for each T 2 B.H/ the
function 'T W T.H/ ! C defined by 'T .K/ D Tr .TK/ is a bounded linear functional.
Furthermore, the map W B.H/ ! T.H/ defined by .T/ D 'T is a linear
isometric isomorphism.
Proof. Only the case in which H has infinite dimension will be treated.
If T 2 B.H/ and K 2 T.H/, then
jTr .TK/j Tr jTKj D
1
X
sj .TK/
jD1
1
X
kTksj .K/ D kTk kKk1 :
jD1
Furthermore, K 7! Tr TK is plainly linear in K. Hence, the function 'T is a bounded

linear functional on T.H/ of norm k'T k kTk. If " > 0, then by Proposition 10.7
there exist unit vectors ; 2 H such that kTk " jhT ; ij. Thus,
kTk " jhT ; ij D j'T . /j k'T kk k D k'T k k kk k D k'T k;
which proves that k'T k D kTk. Because the map W B.H/ ! T.H/ defined by
.T/ D 'T is linear in T, we deduce that is a linear isometry.
To show that is surjective, let ' 2 T.H/ and define ' W H H ! C by
' .; / D '. /. Because j ' .; /j k'kk k D k'kkkkk and
' is
plainly a sesquilinear form, Proposition 10.5 implies that there exists an operator
T' 2 B.H/ such that '. / D hT' ; i, for all ; 2 H.
1
X
sj .K/ j j is the singular value decomposition of a trace-class
If K D
jD1
operator K, then the sum converges in T.H/ and, by the continuity of ' and the
trace,
'.K/ D
1
X
sj .K/'. j
j/
jD1
1
X
1
X
jD1

sj .K/Tr T' . j
jD1
D Tr .T' K/:
j/
sj .K/hT' j ;
ji
385
In the notation established above, this means that .T' / D '. Hence, the linear
isometry W B.H/ ! T.H/ is surjective.
t
u
As a dual space, B.H/ has a weak -topology, which is also commonly referred
to as the ultraweak topology or the -weak topology.
Definition 10.100. The ultraweak topology on B.H/, where H is a separable
Hilbert space, is the weak -topology on B.H/ induced by the isometric isomorphism
B.H/ T.H/ .
We noted in Proposition 10.76 that a linear transformation ' W B.H/ ! C is
weakly continuous if and only if it is strongly continuous, and that such linear maps
have the form
'.T/ D
n
X
hTj ; j i;
jD1
for some finite sets f1 ; : : : ; n g and f1 ; : : : ; n g of nonzero vectors. The situation is
slightly different with the ultraweak topology.
Proposition 10.101. If H is an infinite-dimensional separable Hilbert space, then
a linear transformation ' W B.H/ ! C is continuous with respect to the ultraweak
topology of B.H/ if and only if there are sequences fj gj2N and fj gj2N of vectors
such that
1
X
kj k2 and
jD1
1
X
kj k2
jD1
converge and
'.T/ D
1
X
hTj ; j i;
jD1
for every T 2 B.H/.

t
u

It so happens that T.H/ is itself a dual space.
Theorem 10.102. If H is a separable Hilbert space, then for each S 2 T.H/ the
function 'S W K.H/ ! C defined by 'S .K/ D Tr .SK/ is a bounded linear functional.
Furthermore, the map ! W T.H/ ! K.H/ defined by !.S/ D 'S is a linear isometric
isomorphism.
Proof. If S 2 T.H/ and K 2 K.H/, then
jTr .SK/j Tr .jSKj/ D
1
X
jD1
sj .SK/ kKk
1
X
jD1
sj .S/ D kKk kSk1 :
386
Therefore, the linear transformation 'S on K.H/ is bounded of norm k'S k kSk1 .
Hence, ! is a contractive linear map of T.H/ into the dual space K.H/ .
If ' 2 K.H/ , then define a function W H H ! C by .; / D '. /.
Because ./ D . / for all 2 C, the function is a sesquilinear form
and satisfies j .; /j k'k kk kk. Hence, by Proposition 10.5, there exists an
operator S 2 B.H/ such that .; / D hS; i, for all ; 2 H.
Recall from Example 10.97 that if ; 2 H, then Tr .S. // D hS ; i. Thus,
n
X
j j , then
if F is a finite-rank operator expressed as F D
jD1
Tr .SF/ D
n
X
Tr .S. j
jD1
j //
n
X
hS j ;
ji
jD1
n
X

' j

j
D '.F/:
jD1
Let S D VjSj denote the polar decomposition of S, and write jSj as jSj D V S. If
k
X
j j , which is the projection
f j gj2N is an orthonormal basis of H, and if Pk D
jD1
with range Span f 1 ; : : : ; k g, then the operator Pk W is a finite-rank contraction.

Because . j j /V D j .V j /, we have that
X
X
k
k

k'k j'.Pk W /j D hS j ; V j i D hV S j ; j i D Tr .jSjPk /:

jD1
jD1
Therefore, lim Tr .jSjPk / D Tr .jSj/ exists, which implies that jSj and S are tracek!1
class operators of norm kSk1 k'k. Hence, ! is surjective and isometric.
t
u
To conclude, the final result describes the extreme points of the closed unit ball
of trace-class operators.
Theorem 10.103. The following statements are equivalent for S 2 T .H/:
1. S is an extreme point of the closed unit ball of T .H/;
2. rank S D 1 and Tr.S S/ D 1.

Proof. Denote the closed unit balls of T .H/ and `1 .N/ by T .H/1 and `1 .N/ 1 ,
respectively.
Assume that (1) holds. Express S in its singular value decomposition:
T D
r
X
sj h; j i
j;
for every 2 H ;
(10.16)
jD1
where r D rank S, s1 sr > 0, and f 1 ; : : : ; r g and f 1 ; : : : ; r g are orthonormal

systems in H. Let fen gn2N denote the canonical coordinate vectors of `1 .N/. The
extreme points of the closed unit ball of `1 .N/ are precisely the vectors of the form
ei en , for some 2 R and n 2 N (Exercise 7.35).
387
Now let s 2 `1 .N/ be the vector of the singular

values
of S. We aim to show that

s is an extreme point of the closed unit ball `1 .N/ 1 of `1 .N/. To this end, assume

that s D C .1 /, for some ; 2 `1 .N// 1 and 2 .0; 1/. Let A and B be
operators defined by the equations
A D
d
X
j h; j i
jD1
and
B D
d
X
j h; j i
j;
jD1
Thus, S D A C .1 /B. Furthermore, A; B 2 T .H/1 because kAk1 D kk1 and

kBk1 D kk1 . Hence, A D B D S, as S is an extreme point of T .H/1 . But this
occurs only if D D s, which implies that s is an extreme point of the unit ball of
`1 .N/. Thus, r D 1 and s1 D 1, which yields rank S D 1 and Tr.S S/ D 1.
Conversely, assume that (2) holds: namely, S has rank 1 and Tr.S S/ D 1. The
polar decomposition S D UjSj, for some partial isometry W, shows that S and jSj
have the same rank. By the Spectral Theorem, jSj and S S D jSj2 have the same
rank. Now since S S has exactly one nonzero eigenvalue and Tr.S S/ D 1, this sole
nonzero eigenvalue is 1, which implies that S S is a rank-1 projection.
Let P D S S and let 2 H be a unit vector that spans the range of P. Define a
linear functional W T .H/ ! C by

.T/ D Tr PW T ; T 2 T .H/ :
Thus, for any T 2 T .H/1 ,
j .T/j D jTr.PW T/j D jhW T ; ij kW k kTk kTk1 1 :
In particular, <.T/ 1 for every T 2 T .H/ with kTk1 1. The value of at S is
.S/ D Tr.PW S/ D Tr.S S/ D 1. Suppose now that S D A C .1 /B for some
A; B 2 T .H/1 and 2 .0; 1/. Thus,
1 D < .S/ D < .A/ C .1 /< .B/ C .1 / D 1 :
Hence, < .A/ D < .B/ D 1; however, because j .A/j and j .B/j are at most
1, we deduce that in fact .A/ D .B/ D 1. In particular for A, this means that
hW A ; i D 1, which is a case of equality in the Cauchy-Schwarz inequality and
so W A D , whence A D W D WP D WjSj D S . Furthermore, the equation
A D W implies that kA k D k k D 1 and so the operator norm of A is kAk D
1. Thus, 1 kAk kAk1 1 yields 1 D kAk D kAk1 D 1. Because kAk is the
spectral radius of jAj and k jAj k1 is the trace of jAj, the equation spr jAj D Tr.jAj/ D 1
implies that jAj has exactly one nonzero eigenvalue (namely, 1). Thus, jAj is a rank1 projection and A D VjAj (polar decomposition) has rank-1. Hence, if ! D S , then
A D h; i! D S for every 2 H. By a similar argument, B D S as well. Hence, S
is an extreme point of T .H/1 .
t
u
388
Problems
10.104. Prove that T; S 2 B.H/ are equal if and only if hT; i D hS; i, for all
2 H.
10.105. Prove that ran T D .ker T /? , for every T 2 B.H/.
10.106. If S is the unilateral shift operator on `2 .N/, and if R is an operator on
`2 .N/ for which RS D SR and RS D S R, then prove that R has the form R D 1
for some 2 C.
10.107. For T 2 B.H/, prove that 2 d .T/ if and only if 2 p .T /.
10.108. Prove that the following statements are equivalent for an operator U 2
B.H/:
1.
2.
3.
4.
U is unitary;
U U D UU D 1;
for some orthonormal basis f i gi of H, fU i gi is also an orthonormal basis;
fU i gi is an orthonormal basis for H for every orthonormal basis f i gi of H.
10.109. Suppose that V 2 B.H/ is an isometry.

1. Prove that ap .T/ T.
2. If V is unitary, prove that .T/ T.
10.110. Prove that if P 2 B.H/ is a projection different from 0 and 1, then .P/ D
f0; 1g.
10.111. Suppose that P 2 B.H/ is a nonzero projection acting on a separable Hilbert
space H and that f k gdkD1 is an orthonormal basis for the range of P, where d is either
finite or infinite. Prove that
P D
d
X
h; k i k ;
kD1
for every 2 H.
10.112. Assume that T 2 B.H/ is an operator of rank m 2 N.
1. Prove that if m D 1, then there are unit vectors ; 2 H such that T D h; i ,
for all 2 H.
2. Prove that the rank of T is m.
10.113. Prove that if f k g1
kD0 is the canonical orthonormal basis of the Hardy space
H 2 .T/ and if S 2 B.H 2 .T// is the unilateral shift operator, then S satisfies S k D
k1 , for all k 2 N and S 0 D 0.
10.114. With respect to the canonical orthonormal basis f k g1
kD0 of the Hardy
space H 2 .T/, find the matrix representation S of the unilateral shift operator
Problems
389
S 2 B.H 2 .T//. Viewing S as acting on `2 .N [ f0g/, determine the action of the

matrix S on a vector 2 `2 .N [ f0g/.
10.115. Prove that if T 2 B.H/ is hermitian, then
kTk D sup jhT; ij :
kkD1
10.116. Assume that T 2 B.H/ is hermitian and let

m` D inf hT; i
kkD1
and
mu D sup hT; i :
kkD1
Complete the proof of Proposition 10.26 by proving the following statements.

1. mu 2 .T/.
2. .T/ .1; mu
3. .T/ m` ; mu .
10.117. Suppose that T is a hermitian operator and that .T/ D 1 [ 2 , where 1
and 2 are compact subsets of R with 1 \ 2 D ;. Prove that there are subspaces
M1 and M2 of H such that
1. H D M1 M2 ,
2. Mj is invariant under T, for j D 1; 2,
3. the operator TjMj is hermitian and has spectrum j , for j D 1; 2.
10.118. Assume that H is a separable Hilbert space and that for an operator K 2
B.H/ there are a bounded sequence f j grjD1 of nonzero real numbers, where r is
such that
1. K j D j j , for each j,
2. limj j D 0, if r is infinite, and
r
X
3. K D
jD1
Prove that K is a compact operator.

10.119. Prove that the following statements are equivalent for T 2 B.H/:
1. T is normal;
2. kT k D kTk, for all 2 H;
3. T T D TT .
10.120. Prove that if is an isolated point in the spectrum of a normal operator N,
then is an eigenvalue of N.
10.121. Let B denote the bilateral shift operator on L2 .T/. Prove that the Hardy
space H 2 .T/ is invariant under B and that, with respect to the decomposition
390
L2 .T/ D H 2 .T/ H 2 .T/? , B is represented by an operator matrix of the form

S B12
;
0 B22
where S is the unilateral shift operator on H 2 .T/ and B12 6D 0.
10.122. Prove that the following statements are equivalent for an operator T and
projection P on a Hilbert space H:
1. ran P is reducing for T;
2. .1 P/T P D .1 P/TP D 0;
3. TP D PT.
10.123. A subspace L H is said to be nontrivial if L is neither f0g nor H. Suppose
that T 2 B.H/.
1. Prove that if H has finite dimension, then T has a nontrivial invariant subspace.
(Hint: think about eigenvectors.)
2. Prove that if H is nonseparable, then T has a nontrivial invariant subspace. (Hint:
if 2 H is nonzero, consider the subspace generated by T k for k 2 N.)
10.124. Prove that every compact normal operator is reductive.
10.125. Let S be the unilateral shift operator on `2 .N/.
1. Prove that if T 2 B.`2 .N// satisfies TS D ST and TS D S T, then T D 1 for
some 2 C.
2. Prove that the only subspaces L `2 .N/ that are invariant under both S and S
are L D f0g and L D `2 .N/.
10.126. Assume that N 2 B.H/ is normal. If 1 ; 2 2 ap .N/ are distinct, and if
n ; n 2 H are unit vectors for which
lim k.N 1 1/n k D lim k.N 2 1/n k D 0 ;
n
then prove that

lim hn ; n i D 0 :
n
10.127. Prove that if S and T are hermitian operators for which S T, then X SX
X TX for every operator X 2 B.H/.
10.128. Prove that if A is hermitian and X is invertible, then A is positive if and only
if X AX is positive.
10.129. Prove that, for every operator T 2 B.H/, the operator 1 C T T is invertible
and positive, and
.1 C T T/1=2 T.1 C T T/1=2 D T:
Problems
391
10.130. Suppose that A and B are positive operators such that AB D BA D 0. Prove
that if A0 and B0 are positive operators such that A0 A and B0 B, then A0 B0 D
B0 A0 D 0.

01
10.131. Let T D
2 B.C2 /.
00
1. Prove that .T/ D f0g.
2. Prove that there are no operators R 2 B.C2 / such that R2 D T.
10.132. Let H D C2 and consider hermitian operators A; B 2 B.H/.

1. If A D
2 B.H/, where ; ; 2 C, then prove that A is positive if and only

if and are nonnegative real numbers such that k j2 .
2. Given an example of positive A; B 2 B.H/ for which A B but A2 6 B2 .
10.133. Prove that if A 2 B.H/ is positive and kAk 1, then

.A.1 A//1=2
A
PD
.A.1 A//1=2
1A
is a projection operator on H H.
10.134. The commutant of a nonempty subset S B.H/ is the set
S 0 D fT 2 B.H/ j ST D TS 8 s 2 S g:
1. Prove that S 0 is closed in the weak operator topology.
2. Prove that S 0 is an associative subalgebra of B.H/.
3. Prove that if S 2 S for every S 2 S , then T 2 S 0 for every T 2 S 0 .
10.135. Prove that if H is an infinite-dimensional Hilbert space, then the closure of
the set fP 2 B.H/ j Pis a projectiong in the weak operator topology of B.H/ is the set
of all positive operators A 2 B.H/ for which kAk 1.
10.136 (von Neummans Inequality). Use the Sz.-Nagy Dilation Theorem to
prove that if T 2 B.H/ satisfies kTk 1, then kf .T/k 1 for every polynomial
f 2 C t for which max jf .z/j 1.
jzj1
10.137. Prove that if T 2 B.H/ is invertible, then T D UjTj for some unitary
operator U 2 B.H/.
10.138. Prove that if V 2 B.H/ is a partial isometry, then V V is a projection.
10.139. Suppose that T1 ; T2 2 B.H/ are hermitian operators such that T1 T2 D
T2 T1 D 0. Prove that h; i D 0 for all 2 ran T1 and 2 ran T2 .
10.140. Assume that T 2 B.H/ and 2 C. Prove that
ker.T 1/ 2 Lat T
and
ran.T 1/ 2 Lat T :
392
10.141. Prove that, for every T 2 B.H/,

M 2 Lat T
if and only if
M ? 2 Lat T :
10.142. Prove that the involution T 7! T is a continuous function on B.H/ with

respect to the weak operator topology.
10.143. Let W 0; 1 ! 0; 1 be given by .t/ D t and consider the (hermitian)
multiplication operator M on L2 .0; 1; M; m/.
1. Prove that M has no eigenvalues.
2. Prove that M has no finite-dimensional invariant subspaces.
3. Find one nontrivial subspace of L2 .0; 1; M; m/ that is invariant under M .
10.144. Prove that the space K.H/ compact operators on a separable Hilbert space
H is separable.
10.145. Prove that the space T.H/ trace-class operators on a separable Hilbert
space H is separable with respect to the trace norm k k1 .
10.146. Assume that Tij , for i; j D 1; 2, are trace-class operators acting on a
separable Hilbert space H.

T T
1. Prove that T D 11 12 is a trace-class operator on H H.
T21 T22
2. Prove that kT11 k1 C kT22 k1 kTk1 .
10.147. Assume that H is an infinite-dimensional separable Hilbert space and that
' W B.H/ ! C is a linear transformation.
1. Prove that if B.H/ has the weak operator topology, then ' is continuous if and
only if there are finite sets f1 ; : : : ; n g and f1 ; : : : ; n g of vectors such that
'.T/ D
n
X
hTj ; j i;
jD1
for every T 2 B.H/.

2. Prove that if B.H/ has the ultraweak operator topology, then ' is continuous if
and only if there are sequences fj gj2N and fj gj2N of vectors such that
1
X
kj k2 and
jD1
1
X
kj k2
jD1
converge and
1
X
hTj ; j i;
'.T/ D
jD1
for every T 2 B.H/.
Chapter 11
Algebras of Hilbert Space Operators
Collections of Hilbert space operators lead to a rich palette of algebraic structures.

Of principal interest in this chapter are certain associative algebras of operators,
called -algebras, that are closed under the involution T 7! T . However, one
could also quite readily consider other algebraic structures, such as semigroups of
operators, Lie algebras of operators, or vector spaces of operators. Our focus on
-algebras of Hilbert space operators (and their abstractions known as C -algebras)
stems from the fact that such algebras are widely employed and studied, and
exhibit special features that are not present in more generic algebraic structures. The
monograph of Paulsen [42] has a good treatment of the theory of general operator
algebras and discusses a wide variety of applications to operator theory.
Operator algebras, in the sense formulated in this chapter, are sometimes
considered as the basis for noncommutative topology and noncommutative measure
theory, thereby completing the arc of this text by bringing us back to the books
topological and measure-theoretic beginnings.
11.1 Examples
Definition 11.1. A -algebra of operators is a subset A B.H/ that is closed under
addition, product, scalar multiplication, and the adjoint operation. Furthermore, a
-algebra A is called:
1. a C -algebra of operators, if A is closed with respect to the norm topology of
B.H/, and
2. a von Neumann algebra if A is closed with respect to the strong operator topology
of B.H/.
By the term operator algebra we shall henceforth mean a -algebra of Hilbert
space operators.
DOI 10.1007/978-3-319-45633-1_11
393
394
11 Algebras of Hilbert Space Operators
One might wonder why there is not a third category of operator algebra, defined
by the requirement that it be closed with respect to the weak operator topology. The
reason for there being no third category of -algebra is that nothing new is gained:
if A is a -algebra of operators, then the closure of A in the weak operator topology
coincides with the closure of A in the strong operator topology, by the convexity of
A and Proposition 10.77. Thus, every von Neumann algebra is closed with respect
to the weak operator topology.
Notational Convention In keeping with notation that is standard in the operator
algebra literature, lowercase letters are used in this chapter to denote individual
operators, whereas uppercase letters denote algebras of operators.
Definition 11.2. A -algebra A of operators is abelian if xy D yx for all x; y 2 A.
Before considering operator algebra theory, a few basic examples are considered.
Example 11.3. Group Algebras.
Proof. Suppose that G is a countable discrete group, with identity element e. Define
a Hilbert space `2 .G/ by
`2 .G/ D ff W G ! C j
jf .h/j2 < 1g:
h2G
An orthonormal basis for this Hilbert space is given by the set fg gg2G , where
g .h/ D 1 if h D g and 0 if h 6D g. Thus,
`2 .G/ D
8
<X
:
g g j g 2 C;
g2G
X
g2G
9
=
jg j2 < 1 ;
;
and the inner product on `2 .G/ is given by

*
X
g
g g ;
X
g2G
+
g g D
g g :
g2G
For each h 2 G, let h W `2 .G/ ! `2 .G/ be the operator that sends f 2 `2 .G/ to
the function whose value at k 2 G is f .h1 k/. Note that h is an isometry, and that

h1 D 1
h D h ; thus, h is a unitary operator, for each h 2 G.
2
The action of h on the basiselements
of ` .G/ is given by h g D hg , and so
2
, considered as a map G ! B ` .G/ in which h 7! h , is called the left regular
representation of G.
The group algebra CG, which is the set all products of finite linear combinations
of elements from the set f h j h 2 Gg, is a -algebra of operators. The normclosure in B `2 .G/ of CG is denoted by C .G/, and is called the reduced group
C -algebra of G. The SOT-closure of CG is denoted by V .G/, and is called the
11.1 Examples
395
group von Neumann algebra of G. Evidently, C .G/ is a C -algebra of operators,

and V .G/ is a von Neumann algebra of operators.
There is also the possibility of allowing G to act on the right, leading to operators
h on `2 .G/ for which h .g / D gh , for every g 2 G. Hence, once again each h is a
unitary operator, and the map W G ! B.H/ is the right regular representation of G.
As with the left regular representation, one obtains a C -algebra C .G/ and a von
Neumann algebra V .G/ D C .G/
SOT
t
u
Observe that C .G/ (equivalently, V .G/) is an abelian operator algebra if and

only if G is an abelian group. We shall mainly concern ourselves with C .G/ and
V .G/, using V .G/ only for the purpose of studying V .G/.
Example 11.4. The C -algebra K.H/ of compact operators.
Proof. Theorem 8.35 shows that K.H/ is a Banach algebra of operators, while
Proposition 8.31 implies that K.H/ is closed under the adjoint. Thus, K.H/ is a
C -algebra of operators.
Note, however, that K.H/ is not a von Neumann algebra if H has infinite
dimension, as 1 62 K.H/ but 1 is the SOT-limit of a net of compact operators
(Exercise 11.107).
t
u
Example 11.5. Matrix operator algebras.
Proof. Suppose that A is a -algebra of operators acting on H, and consider the set
Mn .A/ of n n matrices with entries from A. Given a matrix X D aij ni;jD1 2 Mn .A/,
define X by X D aji nij;D1 . The map X 7! X is an involution on Mn .A/ and, thus
under the usual algebra operations on matrices, Mn .A/ is a -algebra of operator
acting on the direct sum H .n/ of n copies of H.
If A is norm closed, then so is Mn .A/. To prove this, let X D aij i;j be in the
.k/
norm-closure of Mn .A/. Thus, there is a sequence of elements Xk D aij i;j such that
.k/
kX Xk k ! 0 as k ! 1. Hence, for a given pair of i and j, kaij aij k ! 0 as

k ! 1. Thus, aij 2 A, and therefore X 2 Mn .A/.
By analogy, the same type of argument in which a SOT-convergent net of
matrices is reduced to n2 SOT-convergent nets of operators shows that Mn .A/ is
SOT-closed if A is SOT-closed.
t
u
Definition 11.6. A -algebra A of operators acting on a Hilbert space H is unital if
the identity operator 1 2 B.H/ is an element of A.
In the examples above, the group algebra CG is unital, where the identity on the
Hilbert space `2 .G/ is given by the identity element of the group G; consequently,
C .G/ and V .G/ are unital operator algebras. If H has infinite dimension, then
K.H/ is not unital, whereas if A is a unital -algebra, then so is the matrix algebra
Mn .A/, where the identity of Mn .A/ is the diagonal matrix in which each entry is the
identity of A.
Definition 11.7. If S B.H/, then the -algebra generated by S is the smallest
-subalgebra of B.H/ that contains S , and is denoted by -Alg S .
396
The definitions of C -algebra and von Neumann algebra generated by a set of

operators are straightforward; however, note that the definition of von Neumann
algebra requires that it be unital.
Definition 11.8. Assume that S is a set of operators acting on a Hilbert space H.
1. The C -algebra generated by S is denoted by C .S / and is defined to be the
norm-closure of -Alg S .
2. The unital C -algebra generated by S is denoted by C .S ; 1/ and is defined to
be the norm-closure of -Alg .S [ f1g/.
3. The von Neumann algebra generated by S is denoted by W .S / and is defined
to be the SOT-closure of -Alg .S [ f1g/.
If S consists of a single operator, x, then
-Alg.x/ D
8
n X
m
<X
:
ij xi .x /j C
iD0 jD1
s X
t
X
kD0 `D1
9
=
k` .x /k x` j n; m; s; t 2 N; ij ; k` 2 C
;
and
-Alg.x; 1/ D
8
n X
m
<X
:
iD0 jD0
ij xi .x /j C
s X
t
X
kD0 `D0
9
=
k` .x /k x` j n; m; s; t 2 N; ij ; k` 2 C :
;
In particular, any one of the algebras -Alg.x/, C .x/, or W .x/ is abelian if and
only if x x D xx .
Example 11.9. Algebras of multiplication operators.
Proof. Assume that .X; ; / is a -finite measure space and consider the Hilbert
space L2 .X; ; /. For every essentially bounded Borel function W X ! C, the
multiplication operator M on H has norm kM k D k P k (see Section 8.1) and
adjoint M D M (see Section 10.2). It is clear that M 1 2 D M 1 M 2 D M 2 M 1
and that M1 1 C2 2 D 1 M 1 C2 M 2 , and so the set
fM j P 2 L1 .X; ; /g
is an abelian -algebra of operators. Because the map L1 .X; ; / !
B.L2 .X; ; // given by P 7! M is multiplicative and a linear isometry, this
set is a norm closed algebra, isometrically isomorphic to L1 .X; ; /. (This
isomorphism is also compatible with the conjugation on L1 .X; ; / and involution
on B.L2 .X; ; // in the sense that M D M .) Hence, the set of all such
multiplication operators on L2 .X; ; / form an abelian C -algebra.
t
u
11.2 von Neumann Algebras
397

One of the most immediate ways by which examples of von Neumann algebras are
obtained is through the use of commutants. Recall from Exercise 10.134 that the
commutant of a nonempty subset S B.H/ is the set
S 0 D fy 2 B.H/ j sy D ys; 8 s 2 S g:
Furthermore, Exercise 10.134 shows that if s 2 S for every s 2 S , then S 0 is a
unital C -subalgebra of B.H/, closed in the weak operator topology.
Example 11.10. If G is a countable discrete group, then V .G/ V .G/0 .
Proof. Select any basis element g of `2 .G/. If h; k 2 G, then
h k g D h gk D hgk D k hg D k h g :
Hence, h k D k h . As the choice of h; k 2 G is arbitrary, and because commutants
are closed in the weak operator topology, xy D yx for x 2 V .G/ and y 2 V .G/. u
t
Definition 11.11. The double commutant of a nonempty subset S B.H/ is the
set denoted by S 00 and defined by S 00 D .S 0 /0 .
Observe that it always the case that S S 00 .
The following example shows how the use of commutants can play a role in
showing that certain C -algebras are von Neumann algebras.
Example 11.12. If .X; ; / is a finite measure space, then L1 .X; ; / is a von
Neumann algebra, when considered as a C -algebra of multiplication operators
acting on the Hilbert space L2 .X; ; /.
Proof. Let H D L2 .X; ; / and let M D fM j 2 L 1 .X; ; /g. Example 11.9
shows that M is a unital C -algebra of operators acting on H. As M is abelian,
M M0.
Conversely, suppose that z 2 M 0 . Let fP 2 H by given by fP D z.P X /, where X 2
1
L .X; ; / is the characteristic function of X. Select any 2 L 1 .X; ; /;
because M z D zM , we have that
. Pf / D M fP D M z.P X / D zM .P X / D z. P /:
Thus, k. Pf /k2 D kz. P /k2 kzk2 k P k2 . Now if E 2 , then
Z
jf j2 d D kz.P E /k2 kzk2 .E/:

E
For each > 0, let E D jf j1 ..; 1//. Thus, the inequality above yields
2
.E /
E
jf j2 d kzk2 .E /:
398
Therefore, if .E / > 0, then kzk. Hence, f 2 L 1 .X; ; / and the multiplication operator Mf and the operator z agree on L1 .X; ; /. Because L1 .X; ; /
is dense in L2 .X; ; /, the operators Mf and z necessarily agree on all of H, which
shows that z 2 M. Hence, M 0 M, which proves that M 0 D M.
Because M D M 0 , M is necessarily closed in the weak operator topology, by
Exercise 10.134. Hence, M is also closed in the strong operator topology, and so M
is a von Neumann algebra.
t
u
It is also true that L1 .X; ; / is a von Neumann algebra if .X; ; / is a
-finite measure space (Exercise 11.108). However, there exist examples of measure
spaces .X; ; / that are not -finite and for which the C -algebra L1 .X; ; / of
multiplication operators is not a von Neumann algebra.
The role of the commutant in Example 11.12 is not accident, as the following
fundamental theorem of von Neumann demonstrates.
Theorem 11.13 (Double Commutant Theorem). The following statements are
equivalent for a unital C -subalgebra M B.H/:
1. M D M 00 ;
2. M is a von Neumann algebra.
Proof. By Exercise 10.134, if M D M 00 , then M is necessarily closed in the weak
operator topology; hence, M is also closed in the strong operator topology, thereby
implying that M is a von Neumann algebra.
Conversely, suppose that M is closed with respect to the strong operator topology.
SOT
. Let y 2 M 00 and
Because M M 00 , we need only prove that M 00 M D M
assume that U B.H/ is a SOT-open set containing y. Thus, there exist " > 0 and
1 ; : : : ; n 2 H such that
B";1 ;:::;n .y/ D fx 2 B.H/ j kyk xk k < "; k D 1; : : : ; ng U:
Consider the Hilbert space H .n/ obtained as the n-fold direct sum of H, and let
Q B.H .n/ / be the set of all n n diagonal matrices
D 1 n 2 H .n/ . Let M
Q
D of operators in which each diagonal entry djj is a fixed element of M: that is, M
consists of all operators of the form Dx D x x, where x 2 M. The commutant
Q consists of all n n matrices of operators such that each entry of the matrix
of M
Q 00 consists of all n n diagonal operators D whose
is an element of M 0 , whereas M
00
Q DM
Q 00 .
entries come from M and which satisfy dii D djj , for all 1 i; j n. Hence, M
.n/
Let p 2 B.H / be the projection onto the closure L1 of the linear submanifold
L0 D fDx j x 2 Mg. If z 2 M, then Dz .Dx / 2 L0 for all x 2 M, and so L0 and
Q Because M
Q is self-adjoint, the
its closure, L1 , are invariant under the algebra M.
Q implies that p 2 M
Q 0 ; hence, pDy D Dy p. In addition, 2 L0
invariance of L1 under M
because 1 2 M, and so Dy D Dy .p/ D p.Dy / 2 L1 , which implies that Dy is
within " of some vector in L0 . That is, kDy Dx k2 < "2 for some x 2 M, which
implies that the basic SOT-open neighbourhood B"; .y/ of y intersects M. Hence,
SOT
D M, thereby proving that M 00 M.
t
u
S2M
399
The Double Commutant Theorem has an interesting consequence for the polar
decomposition of operators.
Proposition 11.14. If M is a von Neumann algebra and if the polar decomposition
of x 2 M is denoted by x D vjxj, then both jxj and v are elements of M.
Proof. Because jxj D .x x/1=2 D limn fn .x x/ for some sequence of polynomials fn
for which fn .0/ D 0 for every n 2 N, we see that jxj 2 C .x/ M. Therefore, it
remains to show that v 2 M.
Suppose that y 2 C .x; 1/0 . If 2 ker x, then xy D yx D 0 implies that yx 2
ker x, and so ker x is invariant under y. Equation (10.12) in the proof of the polar
decomposition (Theorem 10.65) shows that k jxj k D kxk for all 2 H, and so
ker jxj D ker x. Therefore, the restriction of y to the y-invariant subspace ker jxj
commutes with the restriction of x to x-invariant subspace ker jxj.
Suppose now that 2 ran jxj, say D jxj for some 2 H. Thus,
vy D vyjxj D vjxj.y / D xy D yx D yv.jxj / D yv:
Therefore, v and y commute on the linear submanifold ran jxj, and so they commute
on the subspace ran jxj. Because H D ran jxj ker jxj, we deduce that y and v
commute on H.
Hence, v lies in the commutant of C .x; 1/0 , and so by the Double Commutant
Theorem,
v 2 C .x; 1/00 D C .x; 1/
SOT
M;
which proves the result.
t
u
Note that the proof of Proposition 11.14 in fact shows that v and jxj belong to the
von Neumann algebra W .x/ generated by x.
The final fundamental result in von Neumann algebra theory considered in this
section concerns approximation. If A is a untial C -algebra of operators, and if one
SOT
generated by A, then it far from
considers the von Neumann algebra M D A00 D A
apparent, for example, that an operator in the closed unit ball of M is the SOT-limit
of a net of operators from the closed unit ball of A. That such a fact is true is part of
what the density theorem of Kaplansky (Theorem 11.16) asserts below.
The proof of Theorem 11.16 requires the following technical fact.
Lemma 11.15. Let B.H/sa D fx 2 B.H/ j x D xg and define f W B.H/sa ! B.H/sa
by f .x/ D .1 C x2 /1=2 .2x/.1 C x2 /1=2 , for x 2 B.H/sa . Then:
1. f .x/ D 2x.1 C x2 /1 and kf .x/k 1, for every x 2 B.H/sa , and
2. f is continuous with respect to the strong operator topology.
Proof. If x D x, then .1 C x2 / D f1 C 2 j 2 .x/g, by the Spectral Mapping
Theorem. Thus, as .x/ R, 0 62 .1 C x2 /. Because x commutes with 1 C x2 , x
also commutes with any continuous function in x, including .1 C x2 /1=2 . Hence,
f .x/ D .1 C x2 /1=2 .2x/.1 C x2 /1=2 D 2x.1 C x2 /1 .
400
The spectral radius of 1 C x2 is at least 1, and so the spectral radius of .1 C x2 /1

is at most 1; therefore, k.1 C x2 /1 k 1. Furthermore, elementary calculus shows
2t
that the real-valued function t 7! 1Ct
2 is strictly increasing on the interval 1; 1
and has an absolute maximum value of 1 over all t 2 R. Therefore, the equality of
the norm and spectral radius for hermitian operators yields
sup kf .x/k 1:
x2B.H/sa
To prove (2), note that if y; z 2 B.H/sa , then

f .y/ f .z/ D 2.1 C y2 /1 y.1 C z2 / .1 C y2 /z .1 C z2 /1
D 2.1 C y2 /1 .y z C y.z y/z/ .1 C z2 /1 :
Hence, for every 2 H,

kf .y/ f .z/k 2.1 C y2 /1 .y z/.1 C z2 /1

C 2.1 C y2 /1 .y.z y/z/.1 C z2 /1 :
Now fix z 2 B.H/sa and suppose that .y /2 is a net of hermitian operators that
converges in the strong operator topology to z. Because k.1 C y2 /1 k 1, for each
2 H we obtain

2.1 C y2 /1 .y z/.1 C z2 /1 2k.y z/..1 C z2 /1 /k;

and so lim 2.1 C y2 /1 .y z/.1 C z2 /1 D 0. Similarly,

k2.1 C y2 /1 .y .z y /z/.1 C z2 /1 k kf .y /k .z y /z.1 C z2 /1

.z y /z.1 C z2 /1 :

Thus, lim 2.1 C y2 /1 .y z/.1 C z2 /1 D 0.
Combining the two limiting argument above yields lim kf .y / f .z/k D 0,
which shows that f is continuous with respect to the strong operator topology.
u
t
Theorem 11.16 (Density Theorem). Assume that A is a unital C -subalgebra of
SOT
B.H/ and that M D A . If
A1 D fa 2 A j kak 1g;
A1;sa D fa 2 A j a D a; kak 1g;
11.3 Irreducible Operator Algebras
401
and if
M1;sa D fx 2 M j x D x; kxk 1g;
M1 D fx 2 M j kxk 1g;
then M1;sa D A1;sa
SOT
and M1 D A1
SOT
Proof. Let Msa D fx 2 M j x D xg and Asa D fa 2 A j a D ag. Example 10.75

asserts that the involution is continuous with respect to the weak operator topology,
WOT
SOT
WOT
. The set Asa is convex, and so Asa
D Asa
, by
and so Msa D Msa
SOT
WOT
DA
M imply that
Proposition 10.77. Therefore, the inclusions M D A
Msa D Asa
WOT
D Asa
SOT
SOT
:
WOT
Likewise, the set A1;sa is convex, and so A1;sa

D A1;sa
, by Proposition 10.77.
Continuity of the involution with respect to the weak operator topology yields
WOT
and so
M1;sa D M1;sa
A1;sa
SOT
M1;sa
SOT
D M1;sa
WOT
D M1;sa :
Conversely, select x 2 M1;sa . Let g W 1; 1 ! 1; 1 denote the inverse of the
2t
strictly increasing function f .t/ D 1Ct
2 on the interval 1; 1. Thus, g.x/ 2 M1;sa .
SOT
Because Msa D Asa

, there exists a net .a /2 of hermitian operators a 2 Asa
that converges in the strong operator topology to g.x/. Hence, by Lemma 11.15, the
net .f .a //2 of hermitian contractions f .a / 2 A converges in the strong operator
SOT
.
topology to f .g.x// D x, which proves that M1;sa A1;sa
SOT
is outlined in Exercise 11.113.
t
u
The proof that M1 D A1

If A is a -algebra of operators acting on a Hilbert space H, and if L H is a
subspace invariant under A, which is to say that x 2 L for every 2 L and x 2 A,
then for every x 2 A the orthogonal complement L? is invariant under x . Because
A is closed under the involution, we see that if L is invariant under A, then so is L? .
x x
Writing H D L L? and each operator x 2 B.H/ as a matrix 11 12 of operators,
x21 x22
each element a 2 A therefore is a diagonal operator matrix:

a1 0
aD
;
0 a2
402
where a1 D ajL and a2 D ajL? . Because A1 D fajL j a 2 Ag and A2 D fajL? j a 2 Ag

are -subalgebras of B.L/ and B.L? /, respectively, the existence of a nontrivial
invariant subspace for A yields two -algebras A1 and A2 that constitute A and for
which A A1 A2 B.L/ B.L? /.
If A cannot be reduced in this way, then A is said to be irreducible.
Definition 11.17. A -subalgebra A B.H/ is an irreducible operator algebra if
H has dimension at least 2 and the only subspaces L of H that are invariant under A
are L D f0g and L D H.
The following condition characterise irreducible operator algebras.
Proposition 11.18. If H has dimension at least 2, then the following statements are
equivalent for a -subalgebra A B.H/:
1. A is an irreducible operator algebra;
2. if p 2 A0 is a projection, then p D 0 or p D 1; and
3. A0 D f 1 j 2 Cg.
Proof. The equivalence of (1) and (2) is a simple exercise (Exercise 11.110).
Evidently, if (3) holds, then the projections in A0 D f 1 j 2 Cg occur when D 0
or D 1, and so (2) holds. Therefore, assume that (2) holds, and we shall prove that
(3) holds.
If A0 6D f 1 j 2 Cg, then there exists a hermitian operator y 2 A0 with at least two
points of spectrum; hence, by translation by a scalar if necessary, we may assume
that both .y/ \ .1; 0/ and .y/ \ .0; 1/ are nonempty. Therefore, in writing
y D yC y , for positive yC ; y 2 C .y; 1/ A0 with yC y D y yC D 0, it must be
that neither yC nor y is the zero operator. Thus, yC y D 0 implies that the range of
y is contained in the kernel of yC , and so the range of yC is not dense. Therefore,
the partial isometry v in A0 with final space ran yC yields a projection vv 2 A0 such
that vv is neither 0 nor 1, in contradiction to assumption (2).
t
u
Example 11.19. The C -algebra generated by the unitaleral shift operator is an
irreducible operator algebra.
Proof. If s denotes the unilateral shift operator on `2 .N/, then Exercise 10.106
shows that the only operators y that commute with both s and s are operators of
the form y D 1, for 2 C. Therefore, C .s/0 D f 1 j 2 Cg, and so C .s/ is an
irreducible operator algebra.
t
u
The following concept, that of transitivity, is an import from the subject of pure
algebra, where analysis usually has no role.
Definition 11.20. A -subalgebra A B.H/ is transitive if, for each pair of vectors
; 2 H in which 6D 0, there is an operator x 2 A such that x D .
Quite remarkably, irreducible C -algebras are transitive.
Theorem 11.21. If A B.H/ is an irreducible C -algebra, then A is a transitive
C -algebra.
403
Proof. By Proposition 11.18, the commutant of A is f 1 j 2 Cg, and so the double

SOT
D B.H/ by the Double Commutant
commutant of A is B.H/. Therefore, A
Theorem. Hence, if " > 0 is given and if y 2 B.H/ and 2 H, then by the Density
Theorem (Theorem 11.16) there is an element a 2 A of norm kak kyk such that
ka yk < ". We shall make use of this last fact repeatedly in the argument below.
Select any ; 2 H with 6D 0; because A is closed under scalar multiplication,
we may assume without loss of generality that kk D 1. If D 0, then a D 0 2 A
has the property that a D . Therefore, assume that 6D 0. Consider the operator
y0 2 B.H/ given by y0 D ; thus, y0 D h; i for all 2 H, and so ky0 k D kk
and y0 D . As explained in the previous paragraph, there is an operator a0 2 A of
norm ka0 k kk such that ka0 k < 12 .
Now consider the operator y1 2 B.H/ given by y1 D . a0 / , which has the
properties ky1 k 12 and y1 D a0 . Therefore, there is an operator a1 2 A of
norm ka1 k ky1 k < 12 such that ka1 . a0 /k < 14 . Continue this construction
inductively so that at the completion of the n-th step one has a0 ; : : : ; an 2 A, each of
norm kaj k 2j , and such that

0
1

X

n
1
@
A
aj
< 2nC1 :

jD0
Now let a D
1
X
aj , which converges in A because
1
X
kaj k is bounded above by a

0
1

n

X
@
A ! 0 as n ! 1, the operator
a
convergent geometric series. Because
j

jD0
a 2 A satisfies a D , which proves that A is a transitive C -algebra.
t
u
jD0
jD0
Kadisons theorem on transitivity leads to the following useful result.

Proposition 11.22. If an irreducible C -subalgebra A B.H/ contains a nonzero
compact operator, then A contains every compact operator.
Proof. Let x 2 A be a nonzero compact operator. Thus, at least one of <.x/ and =.x/
is nonzero, which implies that A contains a nonzero hermitian operator h. Therefore,
h has a nonzero eigenvalue !. Express h in its spectral decomposition as h D !qCg,
where q is the projection onto ker.h !1/ and where g D h !q has eigenvalues
different from !. The characteristic function ! of the set f!g is continuous on .h/,
because f!g is an open subset of .h/. Hence, the continuous functional calculus
yields the operator q D ! .h/ 2 A. Therefore, A contains an operator of finite rank.
Let y 2 A be an operator of least positive finite rank. If y is not a rank-1 operator,
then there are ; 2 H such that y and y are linearly independent. Because A is an
irreducible C -algebra, Theorem 11.21 shows that A is transitive. Hence, there exists
an operator a 2 A such that ay D . The vectors yay and y are, therefore, linearly
independent. The operator ya plainly leaves the finite-dimensional subspace ran y
404
invariant. Therefore, the linear mapping ya W ran y ! ran y has an eigenvalue , and
so the rank of the restriction of .ya 1/ to ran y is less than the dimension of ran y,
and so the rank of the operator yay y 2 A is less than the rank of y. But y has the
smallest positive rank in A, and so it must be that yay y D 0; hence, yay D y,
which is in contradiction of the fact that yay and y are linearly independent. It
must be, therefore, that y is indeed a rank-1 operator.
As a rank-1 operator, y has the form y D , for some ; 2 H. Without loss of
generality, we may assume that h; i D 1. Let L D f# 2 H j # 2 Ag and suppose
that there exists a nonzero 2 L? . Because A is transitive, there is an operator b 2 A
such that b D . Because the rank-1 operator yb 2 A has the form yb D .b/,
the vector b is orthogonal to L; in particular, 0 D hb; i D h; i 6D 0, which is
a contradiction. Therefore, it must be that L D H. Hence, A contains all rank-1
operators of the form #, for all # 2 H.
Select an arbitrary rank-1 operator, say 0 #0 , on H. By the transitivity of A,
there exists an operator c 2 A with 0 D c. Because both c and #0 belong to A,
so does c. #0 / D 0 #0 . Hence, A contains every rank-1 operator, and so A also
contains the norm-closure of the linear span of the set of rank-1 operators; that is, A
contains K.H/.
t
u
Corollary 11.23. If A is an irreducible C -algebra acting on a finite-dimensional
Hilbert space H, then A D B.H/.
Example 11.24. The C -algebra generated by the unilateral shift operator is unital
and contains every compact operator.
Proof. If s denotes the unilateral shift operator on `2 .N/, then s is an isometry,
which implies that s s D 1. Given that s s 2 -Alg.s/, we deduce that 1 2 C .s/.
Example 11.19 shows that C .s/ is an irreducible C -algebra; hence, C .s/ is also
a transtive C -algebra.
The operator ss fixes every component of a vector in `2 .N/, except for the first
component, which is sent to zero. Thus, the element p D 1 ss 2 C .s/ is the
projection of rank 1 onto the subspace spanned by the canonical unit basis vector
e1 . Therefore, C .s/ is a transitive C -algebra that contains a nonzero compact
operator, and so Proposition 11.22 shows that C .s/ contains every compact
operator on `2 .N/.
t
u
If A1 and A2 are operator algebras acting on Hilbert spaces H1 and H2 ,
respectively, then A D A1 A2 D fa1 a2 jaj 2 Aj g acts on H D H1 H2 . However,
if it were to happen that A2 D f0g, then the presence of H2 is somewhat artificial,
making H overly large relative to the action of A upon it. For this reason, one is most
commonly interested in nondegenerate algebras.
Definition 11.25. A -subalgebra A B.H/ is nondegenerate if the only 2 H for
which x D 0 for every x 2 A is D 0.
In contrast to degeneracy is the following concept.
405
Definition 11.26. A vector 2 H is a cyclic vector for an operator algebra A

B.H/ if the linear submanifold fx j x 2 Ag is dense in H.
Example 11.27. If G is a countable discrete group, then the unit vector e of `2 .G/
is a cyclic vector for the group C -algebra C .G/.
Proof. For each h 2 G, the action of the unitary operator h on e produces the
unit vector g . Therefore, the linear submanifold fxe j x 2 C .G/g contains the
orthonormal basis vectors for `2 .G/. Hence, e is a cyclic vector for C .G/.
t
u
The relevance of cyclic vectors to operator algebra theory is indicated by the next
result.
Proposition 11.28. If H is a separable Hilbert space and if A B.H/ is a
nondegenerate C -subalgebra, then there exist a finite or countable family fHn gn
of pairwise-orthogonal subspaces Hn H and unit vectors n 2 Hn such that
1. each Hn is invariant under A,
2. fxn jM
x 2 Ag is dense in Hn , and
3. H D
Hn .
n
Proof. Select a unit vector 2 H and denote the A-invariant subspace fxn j x 2 Ag
by H . Suppose that p 2 B.H/ is the projection with range H . The invariance of
H under x; x 2 A implies that xp D px (Proposition 10.19). Thus, for every x 2 A,
.1 p/x D x.1 p/, and so
\
.1 p/ 2
ker x D f0g;
x2A
where the equality of the intersection of kernels ker x with f0g is because of the
nondegeneracy of A. Hence, D p 2 H , and this is the case for every unit vector
2 H.
Let S consist of all sets O of unit vectors from H such that H ?H for any pair
of distinct ; 2 O. Order S by set inclusion, and apply a Zorns Lemma argument
to produce a maximal element O in S. Because H is separable,
the set O is finite
M
or countably infinite, and so we denote O by fn gn . Let H0 D
Hn . If H0 6D H,
n
then there is a subspace H1 H such that H D H0 H1 and H1 is A-invariant.

Therefore, applying our arguments to AjH1 would yield a unit vector 2 H1 which
is cyclic on some A-invariant subspace H11 of H1 , which would therefore imply that
O [ fg 2 S, in contradiction to the maximality of O.
t
u
Corollary 11.29. If A is an irreducible operator algebra, then every nonzero vector
is a cyclic vector for A.
Proof. Because A is irreducible, the direct sum decomposition in Proposition 11.28
is trivial in the sense that n must be 1. However, the algebra A1 was constructed in
the proof of Proposition 11.28 by selecting any unit vector 2 H and considering
the subspace H1 D fx j x 2 Ag. As H1 D H, the vector is cyclic for A.
t
u
406
11.4 Abelian Operator Algebras

The Gelfand Representation Theorem (Theorem 9.27) for abelian Banach algebras
certainly applies to any abelian C -algebra. However, the presence of the Hilbert
space adjoint yields an even stronger form of Gelfands theorem.
Definition 11.30. If A and B are C -algebras of operators, then a -homomorphism
from A to B is a map W A ! B such that, for all x; y 2 A and ; 2 C,
1. .x C y/ D .x/ C .y/,
2. .xy/ D .x/ .y/, and
3. .x / D .x/ .
Moreover,
4. If A and B are unital C -algebras of operators, and if maps the identity operator
in A to the identity operator in B, then is said to be a unital map.
5. If a -homomorphism is bijective, then is called an -isomorphism.
In what follows, we shall be considering the complex number system C as a
1-dimensional C -algebra of operators (with adjoint D and norm kk D jj).
Moreover, if X is a compact Hausdorff space, then we shall regard C.X/ as a
unital abelian C -algebra acting on the Hilbert space L2 .X; ; /, where is a
fixed regular Borel probability measure on the Borel sets of X, and where
f 2 C.X/ is identified with the multiplication operator Mf on L2 .X; ; /. Under
this identification, f is identified with Mf D Mf and the norms of f and Mf coincide.
Theorem 11.31 (Gelfand). If A B.H/ is a unital abelian C -algebra and if RA
is the maximal ideal space of A, then the Gelfand transform W A ! C .RA / is an
isometric -isomorphism of the C -algebras A and C .RA /.
Proof. If x 2 A, then
2

kx2 k2 D k.x2 / .x2 /k D k.x x/ .x x/k D kx xk2 D kxk2 :
Thus, kx2 k D kxk2 for each x 2 A and, therefore, is an isometry by Exercise 9.55.
The maximal ideal space RA consists of all unital continuous linear maps W A !
C for which .xy/ D .x/.y/ for all x; y 2 A. Because A is abelian, for all z1 ; z2 2 A,
1
X
1 n
the exponential map z 7! ez D
z satisfies ez1 Cz2 D ez1 ez2 (Proposition 9.41).
n
nD0
In particular, if h 2 A is a hermitian operator and 2 R, then ei h ei h D 1 implies

1 D .1/ D ei h ei h D ei.h/ ei.h/ D jei.h/ j2 :
As the equation above is true for every 2 R, .h/ must be a real number. Thus,
if x 2 A is arbitrary, then expressing x D h C ig, where h; g 2 A are the real and
11.4 Abelian Operator Algebras
407
imaginary parts of x, gives

.x / D .h/ i .g/ D . .h/ C i .g// D . .x// ;
implying that the homomorphism is a -homomorphism.
Because is a unital -preserving isometric homomorphism, the range of is a
unital self-adjoint Banach subalgebra of C .RA /. Because .A/ clearly separates the
points of RA , the Stone-Weierstrass Theorem asserts that .A/ D C .RA /, thereby
proving that is surjective.
t
u
Theorem 11.31 demonstrates that the Gelfand transform of a unital abelian C algebra of operators is an isometric -isomorphism with C.X/, for an appropriate
compact Hausdorff space X. However, if the abelian C -algebra of operators is a von
Neumann algebra M, say M D L1 .X; ; /, then expressing M as a C -algebra of
multiplication operators by continuous functions on some compact Hausdorff space
Y is slightly unnatural since M is already an algebra of multiplication operators.
Therefore, the Gelfand theory for an abelian von Neumann algebra should take a
measure-theoretic form, which is accomplished here in Theorem 11.33 below.
For simplicities of cardinal arithmetic, we will assume that the Hilbert spaces
upon which these abelian von Neumann algebras act are separable.
Theorem 11.32. Assume that M is an abelian von Neumann algebra acting on a
separable Hilbert space H. If M has a cyclic vector, then there exist a compact
metrisable space X, a regular probability measure on the -algebra of Borel
sets of X, and a surjective isometry u W H ! L2 .X; ; / such that the linear map
W L1 .X; ; / ! B.H/ defined by
.M / D u1 M u;
is an isometric -isomorphism of the von Neumann algebra L1 .X; ; / of multiplication operators on L2 .X; ; / and the von Neumann algebra M. Furthermore,
is continuous with respect to the strong operator topology on each of L1 .X; ; /
and M.
Proof. Because H is separable, Exercise 11.114 shows that the strong operator
topology on the closed unit ball of B.H/ is separable and metrisable; hence, the same
is true of the closed unit ball M1 of M with respect to the strong operator topology.
Let fan gn2N be a countable SOT-dense subset of M1 , and let A D C .fan gn ; 1/,
SOT
D M. By the Density
which is a separable unital abelian C -algebra for which A
SOT
D M1 ; and via the Gelfand transform , A is isometrically Theorem, A1
isomorphic to C.X/, where because of the separability of A the space X is necessarily
compact and metrisable (Theorem 5.57).
Let 2 H be a unit cyclic vector for M and define a linear functional ' on C.X/ by
'.f / D h 1 .f /; i. By the Riesz Representation Theorem (Theorem 6.51), there
exists a regular Borel measure on the -algebra of Borel sets of X such that
Z
'.f / D f d;
X
408

for every f 2 C.X/. Define
W C.X/ ! B L2 .X; ; / by
.f / D Mf , the operator
of multiplication by f . The map
is plainly a unital -homomorphism, but it is
also isometric by Example 10.38. Thus, A is isometrically -isomorphic with the
operator algebra fMf j f 2 C.X/g acting on L2 .X; ; /.
If 2 H and " > 0 is given, then there exists x 2 M with k xk < "=2, and there
exists a 2 A with kx ak < "=2. Therefore, the linear submanifold H0 D fa j a 2
Ag is dense in H. If a1 ; a2 2 A satisfy a1 D a2 , then for every a 2 A we have that
.a1 a2 /a D a.a1 a2 / D 0, which implies that a1 D a2 for every 2 H. Thus,
the function u0 W H0 ! C.X/ defined by u0 .a/ D .a/ is well defined, linear, and
surjective. In viewing C.X/ as a linear submanifold of L2 .X; ; /, we have that the
norm of u0 .a/ in L2 .X; ; / satisfies
ku0 .a/k2 D
j .a/j2 d D
X

.a a/ d D ' .a a/ D ha a; i D kak2 :
Therefore, u0 is a linear isometry of H0 onto C.X/, and so passing to closures in

each of H and L2 .X; ; / yields a surjective isometry
u W H ! C.X/
k k2
D L2 .X; ; /:
The map u also has the following property: given a 2 A, then for every b 2 A,
ua.b/ D u.ab/ D .ab/ D .a/ .b/ D M .a/ . .b// D M .a/ u.b/:
Thus, ua D M .a/ u, for every a 2 A. If x 2 M and if .a /2 is a net of operators
1
a 2 A converging
strongly
to uxu1 .
to x, then .ua u / converges
strongly
2

Therefore, the net M .a / is strongly convergent in B L .X; ; / to an operator
SOT
1
in C.X/ D L .X; ; / (Exercise 11.109). The same argument shows1that if
M .a / is strongly convergent to M , then .a / converges strongly to u M u.
Hence, the isometric -isomorphism 0 W C.X/ ! A given by 0 .Mf / D 1 f
has the property that u0 .Mf / D Mf u and extends to an isometric -isomorphism
SOT
SOT
W C.X/
!A
and satisfies u.M / D M u for all essentially bounded
measurable functions .
t
u
The case of noncyclic abelian von Neumann algebras may now be examined.
Theorem 11.33. If M is an abelian von Neumann algebra acting on a separable
Hilbert space H, then there exists a -finite measure space .X; ; / and a
surjective isometry u W H ! L2 .X; ; / such that the linear map W L1 .X; ; / !
B.H/ defined by
.M / D u1 M u;
is an isometric -isomorphism of the von Neumann algebra L1 .X; ; / of
multiplication operators on L2 .X; ; / and the von Neumann algebra M.
11.5 C -Algebras
409
Proof. By Proposition 11.28, there exist a finite or countable family fHn gn of

pairwise-orthogonal subspaces Hn H and unit vectors n 2M
Hn such that each Hn
is invariant under M, fxn j x 2 Mg is dense in Hn , and H D
Hn . Moreover, for
n
each n the restriction of M to Hn is an abelian von Neumann algebra acting on

Hn with cyclic vector n . Therefore, by Theorem 11.32, MjHn L1 .Xn ; n ; n /,
for some Borel probability measure n on the Borel sets n of some compact
[
P
Xn , the disjoint union of the family fXn gn2N ,
metrisable space Xn . Let X D
n
X
[
P
n . Define W ! 0; 1 by .E/ D
.En /. Because
and let D
n
N, the measure spaceM

.X; ; / is -finite.
.Xn / D n .Xn / D 1 for every n 2 M
2
2
Furthermore, L .X; ; / is given by
L .Xn ; n ; n /
Hn D H, where the
n
Hilbert space isomorphism between each Hn and L2 .Xn ; n ; n / is implemented

by
M
2
a surjective linear isometry un W Hn ! L .Xn ; n ; n / and where u D
un is a
M n
surjective isometry H ! L2 .X; ; /. Hence, .M / D u1 M u D
u1
n M jXn un
is an isometric -homomorphism of L1 .X; ; / onto M.
t
u
Theorem 11.33 has the following important consequence: every normal operator
is unitarily equivalent to a multiplication operator.
Corollary 11.34 (Spectral Theorem for Normal Operators). If N is a normal
operator acting on a separable Hilbert space H, then there is a -finite measure
space .X; ; / and a surjective isometry U W H ! L2 .X; ; / such that UNU 1 is
the multiplication operator M on L2 .X; ; /, for some P 2 L1 .X; ; /.
Proof. Let M D W .N/, the von Neumann algebra generated by N. Because N is
normal, W .N/ is abelian. Thus, apply Theorem 11.33 to obtain the result.
u
t
A more specific form of Theorem 11.34 is possible when W .N/ admits a cyclic
vector; see Exercise 11.115.
11.5 C -Algebras
While the study of operator algebras has to this point been quite satisfactory, there is
some limit to what one can achieve using a purely operator-theoretic approach. For
example, if H is an infinite-dimensional Hilbert space, then the compact operators
form a proper ideal K.H/ of B.H/, and so one can consider the quotient Banach
algebra B.H/=K.H/. One would be correct in thinking that the involution on
B.H/ (and on the ideal K.H/) would lead to an involution on the quotient space
B.H/=K.H/; further, as we shall see, the quotient norm behaves just like the operator
410
norm in the sense that kPx xP k D kPxk2 , for every xP 2 B.H/=K.H/. Therefore, it is both
natural and necessary to consider Banach algebras that share the involutive and norm
properties of B.H/.
Definition 11.35. A complex associate algebra A is said to be involutive if there
exists a function x 7! x (called an involution) such that, for all x; y 2 A and 2 C,
1.
2.
3.
4.
.x / D x,
.x C y/ D x C y ,
.x/ D x , and
.xy/ D y x .
Definition 11.36. An involutive Banach algebra A is called a C -algebra if, for

every x 2 A, kx xk D kxk2 . If, in addition, A admits a multiplicative identity 1, then
A is said to be a unital C -algebra.
Evidently, every C -algebra of Hilbert space operators is a C -algebra as defined
above. Algebras of continuous functions offer another example, and they can be
considered in their original form rather than in the guise of a C -algebra of
multiplication operators.
Example 11.37. If X is a locally compact Hausdorff space, then the algebra C0 .X/
of continuous functions f W X ! C that vanish at infinity is a C -algebra under the
norm kf k D maxx2X jf .x/j and involution f .t/ D f .t/, for t 2 X.
Let us now explore a few more consequences of the axioms. If A is a C -algebra
and if x 2 A, then
kxk2 D kx xk kx k kxk and kx k2 D kx x k kx k kx k D kxk kx k;
implying that kxk kx k and kx k kxk. That is, kx k D kxk and so the involution
on a C -algebra is isometric.
If A is a C -algebra with multiplicative identity 1 2 A, then k1k D 1 by an
argument that is similar to the one above. Thus, C -algebras with identity are unital
Banach algebras in the sense of Definition 5.25. Furthermore, if x 2 A, then
1 x D .1 x/ D .x 1/ D x D x:
Likewise, x1 D x. By the uniqueness of the multiplicative identity in a unital ring,
1 D 1.
Definition 11.38. If A is a C -algebra, then a subset B A is a C -subalgebra of A
if B is a C -algebra with respect to the sum, product, involution, and norm of A. If
A is unital and if the multiplicative identity of A belongs to B, then B is said to be a
unital C -subalgebra of A.
If F A is a subset of a C -algebra A, then the C -algebra generated by F
is the smallest C -subalgebra of A that contains F and is denoted by C .F /. Of
special interest is the case in which F D fxg for some x 2 A. In this regard, each of
the elements of the following type will generate abelian C -algebras.
11.5 C -Algebras
411
Definition 11.39. Assume that A is a C -algebra and x 2 A.

1. If x D x, then x is said to be hermitian.
2. If x x D xx , then x is normal.
3. If A is unital and if x x D xx D 1, then x is unitary.
As noted earlier, if A is a unital C -algebra, then the multiplicative identity 1 2 A
is hermitian.
Definition 11.40. The set of hermitian elements in a C -algebra A is denoted
by Asa .
As with Hilbert space operators, real and imaginary parts of every x 2 A are
defined
1
1
< x D .x C x / and = x D .x x /:
2
2i
Hence, Asa is a real vector space and SpanC Asa D A.
In the category of C -algebras, the natural maps between C -algebras are called
-homomorphisms.
Definition 11.41. If A and B are C -algebras, then a -homomorphism from A to B
is a map W A ! B such that, for all x; y 2 A and ; 2 C,
1. .x C y/ D .x/ C .y/,
2. .xy/ D .x/ .y/, and
3. .x / D .x/ .
Moreover,
4. If A and B are unital C -algebras, and if maps the identity of A to the of B, then
is said to be a unital map.
5. If a -homomorphism is bijective, then is called an -isomorphism.
The Gelfand Theorem for unital abelian C -algebras has exactly the same form
as the version for abelian C -algebras of operators (and has exactly the same proof).
Theorem 11.42 (Gelfand). If A is a unital abelian C -algebra, then the Gelfand
transform W A ! C .RA / is an isometric -isomorphism of A and C .RA /.
Gelfands Theorem has many consequences, including a determination of the
C -algebra generated by a normal operator (see Theorem 11.47).
Example 11.43. If A is a unital C -algebra, and if x 2 A is normal operator, then
the character space of C .x; 1/ is the spectrum of x, and the unital abelian C algebra C .x; 1/ is isometrically -isomorphic to C . .x//.
Proof. As x x D xx , the algebra C .x; 1/ is abelian. Let W C .x; 1/ ! C.RC .x;1/ /
denote the Gelfand transform. Therefore,
.x/ D f.x/ j 2 RC .x;1/ g;
412
and the function W RC .x;1/ ! .x/ defined by ./ D .x/ is surjective. To prove
that is injective, suppose that 1 .x/ D 2 .x/. Then it is also true that 1 .y/ D 2 .y/
for every y 2 C .x; 1/, because fx; x ; 1g generates C .x; 1/ and 1 and 2 are unital
-homomorphisms. Thus, is injective
If f g2 is a net in RC .x;1/ converging to 2 RC .x;1/ , then, by definition of
weak -topology, .y/ D lim .y/, for every y 2 C .x; 1/. Hence, is a continuous
function. By Proposition 2.9, any continuous bijection from a compact space onto a
Hausdorff space is a homeomorphism.
t
u
The following example is of interest from the point of view of topology.
Example 11.44. If .X; ; / is a -finite measure space, then there exists a
compact Hausdorff space Y such that L1 .X; ; / and C.Y/ are isometrically
-isomorphic.
Proof. Let Y be the character space of the abelian von Neumann algebra
L1 .X; ; /, and apply Gelfands Theorem.
t
u
In reference to the example above, we know that L1 .X; ; / has a multitude
of projections, whereas a projection in C.Y/ corresponds to the characteristic
function of some subset E of Y. Therefore, because continuous functions preserve
connectivity, the space Y must possess a high degree of disconnectivity. The next
example shows this fact explicitly, since N is a nonmetrisable, totally disconnected
compact Hausdorff space (Proposition 2.79).
Example 11.45. The
-isomorphic.
C -algebras
`1 .N/
and
C.N/
are
isometrically
Proof. By Example 9.31, the maximal ideal space of `1 .N/ is homeomorphic

to N. Therefore, by Theorem 11.42, the C -algebras `1 .N/ and C.N/ are
isometrically -isomorphic.
t
u
The next result shows that the spectrum of a C -algebra element x does not
depend on the particular C -algebra that contains x.
Proposition 11.46 (Spectral Permanence). If B is a unital C -subalgebra of a
unital C -algebra A, then A .x/ D B .x/ for every x 2 B.
Proof. Let x 2 B. The inclusion A .x/ B .x/ has already been noted in Proposition 9.10. To prove the containment B .x/ A .x/ it is sufficient to show that
0 2 B .x/ implies 0 2 A .x/. This is most simply done by proving the contrapositive:
if x 2 B has an inverse x1 in A, then x1 2 B.
Therefore, assume that x 2 B is invertible in A. Hence, x is invertible as well,
since 1 D xz D zx implies that 1 D z x D x z . Consequently, x x 2 B is invertible
in A.
Let C D C .x x; 1/, the unital abelian C -subalgebra of B (and of A) generated
by x x. Applying the Gelfand transform on C to the hermitian element x x yields
C .x x/ R. Proposition 9.10 on spectral permanence in abelian Banach algebras
now implies the following inclusions:
11.5 C -Algebras
413
C .x x/ D @C .x x/ A .x x/ C .x x/:

Therefore, C .x x/ D A .x x/. Consequently, the invertibility of x x in A implies
the invertibility of x x in C B. A left inverse for x in B is .x x/1 x x; because
x is in fact invertible in A, this left inverse is necessarily the inverse x1 2 A of x.
Hence, x1 D .x x/1 x 2 B.
t
u
The Gelfand Theorem (Theorem 11.42) carries the continuous functional calculus for normal operators on Hilbert spaces to the more abstract setting of normal
elements in C -algebras. Note that Proposition 11.46 allows us to adopt the notation
.x/, for the spectrum of x, unambiguously.
Theorem 11.47 (Continuous Functional Calculus). If x is a normal element of a
unital C -algebra A, then the unital C -subalgebra C .x; 1/ of A generated by x is
abelian and
1. the character space RC .x;1/ of C .x; 1/ is homeomorphic to the spectrum of x,
2. there is an isometric isomorphism W C . .x// ! C .x; 1/ such that ./ D x,
where 2 C . .x// is the function .t/ D t, and
3. (Spectral Mapping Theorem) for each f 2 C . .x//, the spectrum of .f / 2
C .x; 1/ is
..f // D ff . / j 2 .x/g:
Proof. Example 11.43 shows that the character space of C .x; 1/ is homeomorphic
to .x/, via the homeomorphism W RC .x;1/ ! .x/ defined by ./ D .x/.
Let W C. .x// ! C.RC .x;1/ / be defined, for f 2 C. .x//, by
.f / D f . .// D f ..x// ;
for each 2 RC .x;1/ . The map is evidently a unital -isomorphism and
k.f /k D
max j.f /j D
2RC .x;1/
max jf ..x//j D max jf . /j D kf k
2RC .x;1/
2.x/
for every f 2 C. .x//.

The map W C. .x// ! C .x; 1/ defined by D 1 is a unital isometric
-isomorphism. If 2 C . .x// denotes the function .t/ D t, then 2 C . .x// is
the function .t/ D t. The Gelfand transform evaluated at x 2 C .x; 1/ yields the
function .x/ 2 C.RC .x;1/ / defined by .x/ D .x/ for every character . Hence,
1 maps x to , implying that x D ./ and x D . /. More generally,
0
1
m X
m X
n
n
X
X
@
kj tk tj A D
kj xk .x /j :
kD0 jD0
kD0 jD0
(11.1)
414
The spectrum of the element given in (11.1) is the range of the continuous function
0
1
m X
n
X
@
kj xk .x /j A 2 C.RC .x;1/ /;
kD0 jD0
namely,
8
m X
n
<X
:
kD0 jD0
9 8
9
m X
n
=
<
=
X
j
j
kj .x/k .x/ j ! 2 RC .x;1/ A D
kj k j 2 .x/ :
; :
;
kD0 jD0
Since the ring of all polynomials in commuting variables t and t is uniformly dense
in C. .x// (by the Stone-Weierstrass Theorem), equation (11.1) shows that
..f // D ff . / j 2 .x/g;
t
u
Notational Convention In applications of Theorem 11.47 it is customary to denote

.f / by f .x/, for each f 2 C . .x//.
Proposition 11.48. Suppose that A is a unital C -algebra and h 2 A is hermitian.
Let X R be a compact set such that X .h/ [ f0g. If f 2 C.X/ satisfies f .0/ D 0,
then f .h/ 2 C .h/.
Proof. By Exercise 11.123, the condition f .0/ D 0 implies that there is a sequence
of polynomials fn for which fn .0/ D 0 and jf .t/ fn .t/j ! 0 uniformly on X (and,
thus, on .h/ as well). Since fn .h/ 2 C .h/,

lim kf .h/ fn .h/k D lim
n!1
n!1

max jf .t/ fn .t/j D 0 ;
t2.h/
and so f .h/ 2 C .h/.
t
u
In Proposition 11.48 above, the algebra C .h/ does not necessarily contain the
identity of A; thus, the conclusion f .h/ 2 C .h/ is sharper than the conclusion of
Theorem 11.47, which is that f .h/ 2 C .h; 1/.
The use of nonunital C -algebras is necessarily in many settings; however, each
such algebra may be realised as a C -subalgebra, of co-dimension 1, of a unital
C -algebra.
Proposition 11.49. If A is a nonunital C -algebra, then on the set
A1 D A C D f.a; / j a 2 A; 2 Cg
11.5 C -Algebras
415
define an involution and vector space operations through the involution and the
vector space operations in each coordinate, and define multiplication by
.a; / .b; / D .ab C b C a; / :
Furthermore, let k k0 W A1 ! R be defined by
kzk0 D supfkzbk j b 2 A; kbk 1g;
(11.2)
for z 2 A1 . Then:
1.
2.
3.
4.
k k0 is a norm;
A1 is a C -algebra with respect to k k0 ;
the ordered pair .0; 1/ is a multiplicative identity for A1 ; and
kak0 D kak, for every a 2 A.
Proof. It is clear that A1 is a Banach space and that .0; 1/ is a multiplicative identity
for A1 . Identify A with the subalgebra f.a; 0/ j; a 2 Ag of A1 . As vector spaces,
A1 =A C, and so A has codimension 1 in A1 . Moreover, if z 2 A1 and a 2 A, then
za 2 A.
Suppose that z 2 A1 satisfies kzk0 D 0. If z 2 A, then kzk0 D 0 implies that kzbk D 0
for every b 2 A. In particular, kzz k D 0, whence kz k D 0. Since the involution on
A is an isometry, kzk D 0. This proves that z D 0 if z 2 A and kzk0 D 0.
Next, consider the possibility that z 6D 0 yet kzk0 D 0. The paragraph above shows
that z 62 A (for otherwise z would be 0). Thus, z D .a; / for some a 2 A and nonzero
2 C. The hypothesis kzk0 D 0 again implies that z b D 0 for all b 2 Athat is,
ab C b D 0 for every b 2 A. Hence, 1 a is a left multiplicative identity for A.
By passing to adjoints, . 1 a/ is a right multiplicative identity of A. Thus,

1 a D 1 a 1 a D 1 a :
In other words, 1 a is a multiplicative identity for A, which is in contradiction to
the hypothesis that A is a nonunital algebra. Therefore, it must be that kzk0 D 0 only
if z D 0. The remaining properties required for k k0 to be a submultiplicative norm
are straightforward to verify and, thus, are omitted.
To verify the property kz zk0 D .kzk0 /2 for all z 2 A1 , let z 2 A1 and b 2 A. Because
A is an algebraic ideal of A1 , zb 2 A; thus,
kzbk2 D k.zb/ .zb/k D kb .z z/bk D kb .z z/bk0 kbk2 kz zk0 :
(11.3)
To show that kz zk0 .kzk0 /2 , note that for each " > 0 there is a b 2 A with kbk 1
such that kzbk > .1 "/kzk0 . Thus, kz zk0 > .1 "/2 .kzk0 /2 by (11.3). As " > 0
is arbitrary, the inequality kz zk0 .kzk0 /2 must hold. Conversely, because k k0 is
submultiplicative and is an isometry on A1 with respect to k k0 , the inequality
kz zk0 kz k0 kzk0 D .kzk0 /2 leads to the conclusion that kz zk0 D .kzk0 /2 .
416
To show k k0 extends the norm k k on A, let a 2 A. For every b 2 A with kbk 1,

kabk kak kbk kak; thus, kak0 kak. On the other hand, if a is normalised so as
to have norm kak D 1, then kak0 kaa k D ka k2 D kak2 D 1 D kak. This proves
that kak0 D kak, for every a 2 A.
t
u
Definition 11.50. If A is a nonunital C -algebra, then the C -algebra A1 is called
the unitisation of A and the norm k k0 on A1 is simply denoted by k k.
Proposition 11.46 indicates that the notation .x/ for the spectrum of x is
unambiguous. Therefore, one can define the spectrum for elements of nonunital
C -algebras.
Definition 11.51. If A is a nonunital C -algebra, and if x 2 A, then the spectrum of
x is the set
.x/ D f 2 C j x 1 is not invertible in A1 g :
Another consequence of spectral permanence is that the norm on a C -algebra is
necessarily unique.
Proposition 11.52. If A is a C -algebra with norm k k and if k k0 on a norm such
that A is a C -algebra with respect to k k0 , then kxk0 D kxk for every x 2 A.
Proof. By (4) of Proposition 11.49, we may suppose without loss of generality that
A is a unital C -algebra.
Let x 2 A and consider x x. The Gelfand transform corresponding to the
unital abelian C -algebra C of A1 generated by x x is an isometry. Hence, kx xk
is the norm of .x x/ in C.RC /that is, kx xk is the maximum modulus of the
elements in the range of .x x/ and is, therefore, given by the spectral radius of
x x (since C .x x/ D .x x/ by spectral permanence). The invertibility of elements
in A1 is based upon a purely algebraic criterion; therefore, the spectrum of x x is
independent of the norm on A. Consequently, kx xk0 D spr .x x/ as well. Thus, for
each x 2 A,
kxk02 D kx xk0 D spr .x x/ D kx xk D kxk2 ;
which completes the argument.
t
u
11.6 Positive Elements and Functionals

The definition of a positive operator T acting on a Hilbert space H is not purely
algebraic in that the definition takes into the action of T on H, as well as the way the
inner product is defined. Therefore, in the abstract setting of C -algebras a different
approach is required.
Definition 11.53. An element h 2 A is positive if h D h and .h/ 0; 1/.
417
In addition to positive operators acting on a Hilbert space, one has the following
example, which is a consequence of the definitions.
Example 11.54. If X is a compact Hausdorff space, then f 2 C.X/ is a positive
element of the unital C -algebra C.X/ if and only if f .t/ 0 for every t 2 X.
Let AC be the set
AC D fh 2 A j h is a positive element of Ag :
Note that if A is a C -subalgebra of a C -algebra B, then AC BC (since the
spectrum of an element x is independent of the C -algebra that contains x).
The first main objectives of this section are to show (i) that every positive element
has a unique positive square root, (ii) that AC is a pointed convex cone, and (iii) that
x x 2 AC for every x 2 A. This latter fact is not as obvious as one might expect.
The final goal of this section is to examine those linear functionals on A that take on
nonnegative real values on positive elements of A.
Proposition 11.55. The following statements are equivalent for a hermitian element h in a unital C -algebra A:
1.
2.
3.
4.
h 2 AC ;
h D b2 for some b 2 AC such that b 2 C .h/;
k1 hk , for every khk;
k0 1 hk 0 , for some 0 khk.
Proof. (1) ) (2). Because

p h is positive, .h/ RC . Let X D 0; khk and let f 2
C.X/ be given by f .t/ D t. By Proposition 11.48,pthe hermitian element b D f .h/
is an element of C .h/. Furthermore, .b/ D ff . j 2 .h/g 0; 1/, which
implies that b 2 C .h/C , and so b 2 AC .
(2) ) (3). Assume that h D b2 for some positive b 2 C .h/. Choose any khk.
By spectral mapping, .b2 / D f 2 j 2 .b/g; thus, .b2 / RC . Since the norm of
a positive element is its spectral radius, 0 khk implies that D
j j for every 2 .h/. Hence, spr .1 h/ D k1 hk.
(3) ) (4). This is trivial.
(4) ) (1). Assume that 0 khk satisfies 0 k0 1 hk. Thus, if 2 .h/, then
j j 0 and j0 j 0 ; that is, 0.
t
u
If h 2 A is positive, then the positive element b 2 C .h/ that satisfies b2 D h in
assertion (2) of Proposition 11.55 is unique, as shown by the following proposition.
Proposition 11.56. If b1 ; b2 2 AC are such that b21 D b22 , then b1 D b2 .
Proof. Let > 0 be large enough so that .h/ [ .b1 / [ .b2 / 0; , where
h D b21 D b22 . Therefore, for any g 2 C.0; /,
kg.bj /k D max jg. /j max jg.t/j:
2.bj /
0t
418
By Exercise 11.123, there

p is a sequence of polynomials fn such that fn .0/ D 0, for
all n 2 N, and jfn .t/ tj ! 0 uniformly on 0; as n ! 1. Thus, each fn .h/ 2 A
and kfn .h/ f .h/k ! 0. Note that fn .h/ D fn .b2j / 2 A, and so
kf .h/ bj k D lim kfn .h/ bj k D lim kfn .b2j / bj k D lim kgn .bj /k;
n
where gn .t/ D fn .t2 / t, for each n. Since gn ! 0 uniformly on 0; as n ! 1, we

have that kgn .bj /k ! 0, whence f .h/ D b1 D b2 .
t
u
Definition 11.57. If h 2 AC , then the unique b 2 AC for which b2 D h is called the
positive square root of h and is denoted by h1=2 .
The decomposition of a hermitian operator as a difference of positive operators
with product 0 applies to the C -algebra setting as well. The result is proved again
below because of the need to account for nonunital algebras.
Proposition 11.58. If h is a hermitian element of a C -algebra A, then there are
positive hC ; h 2 AC such that h D hC h and hC h D h hC D 0.
Proof. First assume that A is unital. The C -algebra C .h; 1/ is a unital, abelian C subalgebra of A; moreover, C .h; 1/ and C. .h// are isometrically -isomorphic.
Let X D khk; khk, a compact set that contains .h/ and 0. Consider the functions
f ; g 2 C.X/ defined by f .t/ D .t C jtj/=2 and g.t/ D f .t/. The functions f and
g are nonnegative and vanish at 0; thus, by the Spectral Mapping Theorem and
Proposition 11.48 the elements f .h/ and g.h/ are positive and belong to C .h/. Let
hC D f .h/ and f D g.h/. Because t D f .t/ g.t/ and f .t/g.t/ D 0 for all t 2 X, the
Continuous Functional Calculus yields h D hC h and hC h D h hC D 0.
If A is nonunital, then consider A as a C -subalgebra of its minimal unitisation A1 .
The argument above yields hC ; h 2 C .h/C AC .A1 /C such that h D hC h
and hC h D h hC D 0, thereby completing the proof.
t
u
Our second objective for this section is achieved by the next result.
Proposition 11.59. If A is a C -algebra, then AC is a pointed convex cone. That is,
if ; 2 0; 1/ and if h; k 2 AC , then
1. h C k 2 AC , and
2. h 2 AC only if h D 0.
t
u
We now arrive at our third objective for this section.

Theorem 11.60. AC D fx x j x 2 Ag, for every C -algebra A.
Proof. If h 2 AC , then assertion (2) of Proposition 11.55 yields a positive element
b 2 C .h/ such that b2 D h. Hence, h D b b 2 fx x j x 2 Ag.
Conversely, let x 2 A. By Proposition 11.58, the hermitian element x x 2 A may
be expressed as x x D bC b , where bC ; b 2 AC and bC b D b bC D 0. To
show that x x 2 AC it is sufficient to prove that b D 0.
419
p
Let c D .b /1=2 2 AC and a D xc. By Exercise 11.123, f .t/ D t can be
approximated uniformly on .b / by polynomials p such that p.0/ D 0. Since
p.b /bC D bC p.b / D 0 for any polynomial for which p.0/ D 0, we conclude that
cbC D bC c D 0. Hence,
a a D cx xc D c.bC b /c D cb c D b2 ;
which implies that .aa / RC . Thus,
.a a/ .RC / :
(11.4)
Let u; v 2 A be the real and imaginary parts of a; thus, a D u C iv. By the

Spectral Mapping Theorem, u2 and v 2 are positive and so, by Proposition 11.59,
u2 C v 2 2 AC . Therefore, a a C aa 2 AC as well, since a a C aa D 2.u2 C v 2 /.
By Proposition 11.59 once again, we have that a a C aa C b2 2 AC . But
a a C aa C b2 D a a C aa a a D aa I
this shows that aa 2 AC and so
.aa / R :
(11.5)
Theorem 9.3 asserts that .aa / [ f0g D .a a/ [ f0g. Therefore, (11.4) and (11.5)
combine to give
.a a/ .RC / \ RC D f0g :
Therefore, the spectral radius of a a is 0. Since the spectral radius and norm
coincide for hermitian elements, a a D 0. That is, 0 D ka ak D kak2 , which proves
that a D 0. Since b2 D a a D 0 and b is positive, we obtain b D b2 1=2 D
t
u
01=2 D 0.
Definition 11.61. If h; k 2 Asa , then h k if k h 2 AC .
The relation on Asa has the following properties (Exercise 11.126). If a; b; c 2
Asa , then:
1. a a;
2. if a b and b a, then b D a; and
3. if a b and b c, then a c.
That is, is a partial order on the R-vector space Asa .
Proposition 11.62. If h; k 2 Asa satisfy h k, then x hx x kx for every x 2 A.
Proof. If x 2 A, then
x kx x hx D x .k h/x D x .k h/1=2 .k h/1=2 x D z z 2 AC ;
where z D .k h/1=2 x.
As with Hilbert space operators, one has the following definition.
t
u
420
Definition 11.63. If A is a C -algebra and if x 2 A, then the modulus of x is the

element jxj 2 A defined by .x x/1=2 .
If X is a compact Hausdorff space and if is a probability measure on the Borel
sets of X, then the linear functional ' W C.X/ ! C defined by
Z
'.f / D f d;
X
satisfies '.f / 0 whenever f .t/ 0 for all t 2 X. This example motivates the
following definition.
Definition 11.64. A linear functional ' W A ! C on a C -algebra A is a positive
linear functional if '.h/ 0 for every h 2 AC . If, in addition, k'k D 1, then the
positive linear functional ' is called a state on A. The state space of A is the set
S.A/ of all states on A.
Example 11.65. If fn gn2N is a sequence of vectors for which
then the function ' W B.H/ ! C defined by
'.x/ D
1
X
kn k2 converges,
nD1
1
X
hxn ; n i
nD1
is a positive linear functional on B.H/.

Positive linear functionals on A necessarily map Asa onto R, which can be seen
via writing h 2 Asa as h D hC h , where hC ; h 2 AC . Therefore, by expressing
any x 2 A in terms of its real and imaginary parts, we obtain
'.x / D '.x/ ;
8 x 2 A ; 8 ' 2 S.A/ :
Proposition 11.66 (Schwarz Inequality). If ' 2 S.A/ and x; y 2 A, then

j'.y x/j2 '.x x/ '.y y/:
(11.6)
Proof. The equation x; y D '.y x/ defines a sesquilinear form on A A. Therefore,

the proof of the inequality can be achieved by arguing as in the proof of the CauchySchwarz inequality in Hilbert space.
Choose x; y 2 A. If x; y D 0, then the inequality holds trivially. Thus, assume that
x; y 6D 0. Note that x x; y y 2 AC imply that x; x; y; y 2 RC . For any 2 C,
0 x y; x y D x; x 2< . y; x/ C j j2 y; y :
For
D
x; x
;
y; x
421
the inequality above becomes

0 x; x C
x; x2 y; y
;
jx; yj2
which yields inequality (11.6).
t
u
If A is a unital C -algebra, then there is a relatively simple criterion for a linear

functional to be a state.
Proposition 11.67. The following statements are equivalent for a linear functional
' of norm k'k D 1 on a unital C -algebra A:
1. ' is a state on A;
2. '.1/ D 1.
Proof. Assume that ' is a state on A. Because 1 D 1 1 2 AC and k1k D 1, we
have that 0 '.1 1/ D '.1/ k'k k1k D 1. To show that 1 '.1/, choose any
x 2 A with kxk 1. Thus, kx xk 1. Since kx xk D r.x x/ and .x x/ RC , the
hermitian element 1 x x is positive in A. Thus, 0 '.1 x x/ D '.1/ '.x x/,
which implies that '.x x/ '.1/. Therefore, by an application of the Schwarz
inequality,
j'.x/j D j'.1 x/j '.x x/'.1 1/ '.1/2 1 ;
since '.1/ 1. Hence j'.x/j 1, for all x 2 A with kxk 1, implies that k'k '.1/.
By hypothesis, k'k D 1; therefore, '.1/ D 1.
Conversely, suppose that '.1/ D 1; thus, k'k D '.1/ D 1. It must happen that
'.Asa / D R, for if not then there is a hermitian element h 2 Asa such that '.h/ D
C i, where ; 2 R and 6D 0. Therefore, with k D 1 .h 1/ 2 Asa , we would
have that '.k/ D i and, for each 2 R,
. C 1/2 D ji C ij2 D j'.k C i1/j2 k'k2 kk C i1k2
D k.k C i1/ .k C i1/k D kk2 C 2 1k D kk2 k C 2 :
Thus, .2 C 1/ kk2 k for all 2 R. But this is impossible; therefore, it must be
that '.h/ is real for every h 2 Asa . If h 2 AC , then k'k D 1 and '.h/ 2 R imply that
'.h/ 2 khk; khk. Thus, khk khk '.h/ 0, which implies that '.h/ 0. u
t
For every ' 2 S.A/ and x 2 A we have the basic inequality j'.x/j kxk. If x
is positive, then equality is achieved for some state ', as shown by the following
result.
Proposition 11.68. For every h 2 AC there is a state ' on A with '.h/ D khk.
Proof. If A is nonunital, then consider the unitisation A1 of A; otherwise, let A1
denote A in the case where A is unital.
422
If h 2 AC , then h 2 .A1 /C as well. Consider the unital abelian C -algebra C .h; 1/

generated by h. Via the Gelfand transform, there is a character W C .h; 1/ ! C
such that .h/ D khk. Of course, .1/ D kk D 1. By the Hahn-Banach Theorem,
extends to a linear function W A1 ! C. Since kk D .1/ D 1, is a state on A1 by
Proposition 11.67. Thus, if A is unital, we may take ' D . If A is nonunital, then let
' D jA . Note that '.k/ 0 for all k 2 AC and that k'k 1. With k D khk1 h 2 AC ,
we have kkk D 1 and '.k/ D 1. Hence k'k D 1, and so ' is a state on A.
t
u

Ideals of C -algebras inherit many properties of the ambient C -algebra. First and
foremost of these is that every ideal of a C -algebra is itself a C -algebra, which is
proved as Theorem 11.70 below.
Lemma 11.69. If J is an ideal of a C -algebra A and if x 2 J, then there is a
sequence fen gn2N JC such that .en / 0; 1, for all n 2 N, and kxen xk ! 0.
Proof. First suppose that x 2 A. If A is unital and if e 2 AC satisfies .e/ 0; 1, then
k1 ek 1 (Exercise 11.132). Thus, kx xek2 D k.1 e/x x.1 e/k kx x.1
e/k D kx x x xek. If A is nonunital, then one can embed A into A1 to produce the
same inequality. Therefore, regardless of whether A is unital or not,
kx xek2 kx x x xek ;
8 e 2 AC with .e/ 0; 1 :
(11.7)
Suppose now that x 2 J. Because J is an ideal, x x 2 J. Let h D x x. For each

n 2 N, let fn .t/ D nt=.1Cnt/; thus, fn 2 C. .h//, 0 fn .t/ 1, for all t, and fn .0/ D 0.
Let en D fn .h/. Theorem 11.47 and Proposition 11.48 show that en 2 JC and .en /
0; 1. We aim to verify that kh hen k ! 0. To this end, note that if t 2 .h/, then
t tfn .t/ D
t
D
1 C nt
nt
1 C nt

1
1
< ;
n
n
8 t 2 .h/ :
Therefore, by the fact that continuous functional calculus is an isometric

-homomorphism, kh hen k < 1=n. Hence, by inequality (11.7),
kx xen k2 kx x x en k <
That is, lim kxen xk D 0.
n!1
1
:
n
t
u
Theorem 11.70. If J is an ideal of a C -algebra A, then J is a C -subalgebra of A.

Proof. All that needs to be verified is that x 2 J for every x 2 J. By Lemma 11.69,
there is a sequence fen gn2N JC such that .en / 0; 1, for all n 2 N, and kxen
xk ! 0. Note that en x 2 J for every n 2 N. The C -norm is isometric, and so
423
lim ken x x k D lim kxen xk D 0 :
n!1
n!1
Because J is closed and because each en x 2 J, we conclude that x 2 J.
t
u
Define a function on A=J by

P D .xP / :
.x/
It is clear that this definition above yields an involution on the associative algebra
A=J.
Recall that A=J is a Banach algebra under the quotient norm
kPxk D inf fkx bk j b 2 Jg :
The new fact that is proved below is that the quotient norm satisfies the C -norm
axiom kPxk2 D kPx xP k.
Theorem 11.71. If J is an ideal of a C -algebra A, then the quotient Banach
algebra A=J is a C -algebra with respect to quotient norm and the involution
xP D .xP /.
Proof. Because J is closed under the involution, the function xP 7! .xP / is a welldefined involution on the quotient A=J. Because A=J is a Banach algebra in the
quotient norm, the only issue remaining to be verified is that the quotient norm
satisfies kPxk2 D kPx xP k. To this end, fix x 2 A and define
E D fe 2 JC j .e/ 0; 1g :
If A is unital and if e 2 E, then k1 ek 1 and, for any b 2 J, kx C bk k.x C
b/.1 e/k D k.x xe/ C .b be/k. If A is nonunital, then one can embed A into A1
to produce the same inequality. Hence,
kx C bk k.x xe/ C .b be/k ;
8 e 2 E; b 2 J ;
regardless of whether A is unital or not. By definition of the quotient norm,

kPxk inf fkx xek j e 2 Eg :
(11.8)
To show that equality holds in (11.8), let b 2 J. By Lemma 11.69, there is a

sequence fen gn2N E such that kben bk ! 0. Thus, for every n 2 N,
kx C bk k.x xen / C .b ben /k ;
and so
kx C bk lim inf kx xen k inf kx xek kPxk :
n
e2E
424
Therefore,
kPxk D inf kx C bk kPxk
b2J
implies that kPxk D inf fkx ek j e 2 Eg, for every x 2 A. Consequently, by invoking
inequality (11.7) we obtain
kPxk2 D inf kx xek2 inf kx x x xek D kPx xP k;
e2E
e2E
implying that
kPxk2 kPx xP k kPx k kPxk :
(11.9)
Conversely, kPx k D inffkx b k j b 2 Jg D inffkx bk j b 2 Jg D kPxk, since J is

-closed. Therefore, inequality (11.9) is an equality.
u
t
Quotient algebras occur by way of the kernels of -homomorphisms. The main
features of -homomorphisms are described by the following result.
Proposition 11.72. If A and B are C -algebras, and if W A ! B is a
-homomorphism, then
1.
2.
3.
4.
is continuous and kk 1,

is an isometry if and only if ker D f0g,
the kernel of is an ideal of A, and
the range of is a C -subalgebra of B.
Proof. By Exercise 11.136, spr .x x/ spr .x x/, for all x 2 A. Thus,
k.x/k2 D k.x/ .x/k D spr .x x/ spr .x x/ D kx xk D kxk2 :
That is, is bounded and kk 1, which proves (1).
For (2), it is trivial that isometries are injective, and so only the converse is proved
here. Thus, assume that ker D f0g. Assume, contrary to what we aim to prove,
that there is an element x 2 A with k.x/k < kxk. Then, k.h/k < khk, where h D
x x 2 AC . Let f W 0; khk ! R be any continuous function such that f .t/ D 0 for t 2
0; k.h/k and f .khk/ D 1. By the Spectral Mapping Theorem, kf ..h//k D 0 and
kf .h/k 1. Because f ..h// D .f .h// (by the continuity of and the Weierstrass
Approximation Theorem), it must be that k.f .h//k D 0. Since is injective, this
means that f .h/ D 0in contradiction of kf .h/k 1. Therefore, it must be that is
isometric if ker D f0g, which proves (2).
Since is continuous, ker is closed. As the kernel of any -homomorphism is
an algebraic ideal, we conclude that ker is an ideal, thereby proving (3).
For the proof of (4), consider the quotient C -algebra A= ker and let W
A= ker ! B be defined by .Px/ D .x/, for every x 2 A. Then is a well-defined
-homomorphism with trivial kernel and range equal to the range of . Thus, by (2),
is an isometry, and so the range of is closed. Hence, the range of is closed. u
t
11.8 Representations and Primitive Ideals
425
Corollary 11.73. If two C -algebras are -isomorphic, then they are isometrically
-isomorphic.
One of the most important quotient C -algebras occurs in operator theory.
Definition 11.74. If H is an infinite-dimensional Hilbert space, then the Calkin
algebra is the quotient C -algbera Q.H/ D B.H/=K.H/.
Example 11.65 shows that B.H/ admits positive linear functionals ' of the form
'.x/ D
1
X
hxn ; n i;
nD1
for sequences fn gn in H in which
1
X
kn k2 converges. With nonzero positive linear
nD1
functionals of this form, one can always find a compact operator k 2 B.H/ such that
'.k/ 6D 0 (Exercise 11.139). In light of the following example, not all states on B.H/
are given by such a formula.
Example 11.75. If H is an infinite-dimensional Hilbert space, then there exists a
state ' on B.H/ such that '.k/ D 0 for every compact operator k 2 B.H/.
Proof. Let q W B.H/ ! Q.H/ be the quotient map q.x/ D xP , mapping B.H/ onto
the Calkin algebra Q.H/. Because H has infinite dimension, Q.H/ 6D f0g. Select a
nonzero positive h 2 Q.H/. By Proposition 11.68, there is a state on Q.H/ with
.h/ D khk. Let ' D q, which is a positive linear functional such that '.1/ D 1.
P in Q.H/, the state ' annihilates every compact operator
Because q .K.H// D f0g
in B.H/.
t
u

Some C -algebras, such as the Calkin algebra, occur abstractly rather than as a
C -algebra of Hilbert space operators. The goal of this section is to show, for any
C -algebra A, the existence of -homomorphisms
W A ! B.H/ (for an appropriate
choice of Hilbert space H) where by
is isometric; in so doing, A and the C algebra
.A/ of operators acting on H are isometrically -isomorphic.
Definition 11.76. A representation of a C -algebra A on a Hilbert space H is a
-homomorphism
W A ! B.H/. Further,
is:
1.
2.
3.
4.
unital, if A is a unital C -algebra and

.1/ D 1;
nondegenerate, if the only 2 H that satisfies
.a/ D 0 for all a 2 A is D 0;
cyclic, if there is a vector 2 H such that f
.a/ j a 2 Ag is dense in H;
irreducible, if the commutant of
.A/ in B.H/ is f 1 j 2 Cg.
The following theorem reveals a close relationship between states and representations of C -algebras.
426
Theorem 11.77 (Gelfand-Naimark-Segal). Assume that ' is a state on a unital

C -algebra A.
1. (Existence) There exists a unital representation
of A on a Hilbert space H
and
a unit vector 2 H
such that
a. is a cyclic vector for
.A/, and
b. '.x/ D h
.x/; i for every x 2 A.
2. (Uniqueness) Given a triple .H
;
; / as in (1), if W A ! B.H / is a unital
representation of A, if 2 H is a unit cyclic vector for .A/, and if '.x/ D
h.x/; i, for every x 2 A, then there is a surjective isometry u W H
! H such
that u D and u
.x/ D .x/u, for every x 2 A.
Proof. Let L D fx 2 A j '.y x/ D 0 for every y 2 Ag. By the Schwarz inequality for
states, j'.b a/j '.a a/'.b b/ for every a; b 2 A, This inequality implies that L is
a closed set; hence, because L is also a vector space, L is a subspace of A.
For each x 2 L and a; y 2 A,

' y .ax/ D ' .y a/x D ' .a y/ x D 0;
implying that ax 2 L.
Define a sesquilinear form h; i on the quotient vector space A=L by
P D '.b a/
hPa; bi
A straightforward computation shows that this form is well defined. Moreover, if
hPa; aP i D 0, then '.a a/ D 0 and, by the Schwarz inequality, '.y a/ D 0 for every
P which proves that h; i is an inner product
y 2 A. Thus, hPa; aP i D 0 only if aP D 0,
on A=L. In the metric on A=L induced by the norm kPxk D hPx; xP i1=2 , let H
be the
completion of A=L. Thus, H
is a Hilbert space that contains A=L as a dense linear
submanifold.
For each x 2 A, let
0 .x/ W A=L ! A=L be the (well-defined) linear transformation
P Because
0 .x/Pa D .xa/.
k
0 .x/Pak2 D '.a .x x/a/ kx xk'.a a/ D kxk2 kPak2 ;
the linear transformation
0 .x/ extends to an operator
.x/ on H
of norm at most
kxk. Furthermore,

P D ' b .xa/ D ' .x b/ a D hPa;
0 .x /bi
P
h
0 .x/Pa; bi
implies that
.x/ D
.x/. Because
0 is a contractive -homomorphism of A into
B.H
/,
is also a contractive -homomorphism.
Let D 1P 2 A=L H
and note that h
.x/; i D '.1 x/ D '.x/ for every x 2 H
and that
f
.x/' j a 2 Ag D A=L:
Hence, ' is a cyclic vector for
.A/, which establishes the existence of the trip
.H
;
; / with stated properties in (1).
427
The proof of the uniqueness assertion (2) is left as Exercise 11.138.
t
u
Definition 11.78. If ' is a state on a unital C -algebra, then a triple .H

;
; /
consisting of a Hilbert space H
, a representation
of A on H
, and a unit vector
2 H
is called a GNS-triple for ' if (i) is a cyclic vector for
.A/ and (ii)
'.x/ D h
.x/; i, for every x 2 A.
The main fundamental fact about representations C -algebras is the following
theorem.
Theorem 11.79 (Gelfand-Naimark). For every unital C -algebra A there exists a
Hilbert space H and a unital representation
W A ! B.H/ such that
is injective.
Moreover, if A is separable, then H can be taken to be a separable Hilbert space.
Proof. Fix z 2 A. By Proposition 11.68, there is a state '0 on A such that '0 .z z/ D
kz zk. Let Cz D f' 2 S.A/ j '.z z/ D kz zkg, which is a convex and weak -closed
subset of the unit sphere in the dual space of A. Therefore, the Kren-Milman
Theorem (Theorem 7.18) asserts that Cz has an extreme point, say '. This state
' is also an extreme point of S.A/, for if ' D 12 .'1 C '2 /, then the fact that states
are contractive implies that '1 .z z/ D '2 .z z/ D kz zk; hence, '1 ; '2 2 Cz , implying
that '1 D '2 D '.
Suppose now that .H;
; / is a GNS-triple for '. If
is not an irreducible
representation of A, then there exists a projection p 2 B.H/ such that p 6D 0,
p 6D 1, and p
.x/ D
.x/p for all x 2 A. If p were 0, then it would be true that
p.
.x// D
.x/p D for every x 2 A; however, such vectors form a dense subspace
of H and this would imply that p D 0, contrary to the assumption on p. Likewise,
1 p 6D 0. Let t D kpk2 so that 1 t D k.1 p/k2 and t 2 .0; 1/. Define states '1
and '2 by
1
1
h
.x/.1 p/; .1 p/i:
'1 .x/ D h
.x/p; pi and '2 .x/ D
t
1t
Because ' D t'1 C .1 t/'2 , we have that '1 D '2 D '. In particular, the equation
h
.x/; i D t1 h
.x/p; pi D t1 h
.x/; pi
for every x 2 A implies that
.x/; p t1 D 0
for every x 2 A. Because the f
.x/ j x 2 Ag is dense, the equation above yields
p D t1 , which implies t1 is an eigenvalue of p. However, t1 62 f0; 1g D .p/
and, therefore, it must be that the only projections that commute with
.A/ are 0
and 1. Hence,
is a unital irreducible representation and it has the property that
kzk2 k
.z/k2 D k
.z z/k h
.z z/; i D '.z z/ D kz zk D kzk2 :
Hence, k
.z/k D kzk.
428
Denote the GNS-triple associated with the irreducible representation

M
in the
previous paragraph by .Hz ;
z ; z /. Consider the Hilbert space H D
Hz and the
representation
D
z2A
z and note that

W A ! B.H/ is a unital representation of
z2A
A for which
.x/ D 0 only if x D 0.
and if fan gn2N is a countable dense subset of A, then with H D
M
MIf A is separable
Han and
D
an we have a representation
of A on the separable Hilbert
n2N
n2N
space H for which

.x/ D 0 only if x D 0.
t
u
Corollary 11.80. Every C -algebra is isometrically -isomorphic to a C -algebra

of Hilbert space operators.
Proof. If A is a unital C -algebra, then Theorem 11.79 applies immediately to
achieve the assertion. If A is nonunital, then apply Theorem 11.79 to the unitisation
A1 to achieve an isometric -isomorphism of A with a C -algebra of Hilbert space
operators.
t
u
A noteworthy fact that is a consequence of the proof of Theorem 11.79 is:
Proposition 11.81. If .H
;
; / is a GNS-triple for a state ' on a unital C -algebra
A, and if ' is an extreme point of the state space of A, then
is an irreducible
representation.
The converse of Proposition 11.81 is also true.
We begin our consideration of primitive ideals by noting a basic relationship
between irreducible representations of A and its ideals J.
Proposition 11.82. Assume that J is an ideal of a C -algebra A.
1. If
W A ! B.H/ is an irreducible representation of A and if J 6 ker
, then
jJ
is an irreducible representation of J.
2. If W J ! B.H/ is an irreducible representation of J, then extends to an
irreducible representation
W A ! B.H/ of A.
Proof. Assume that
W A ! B.H/ is irreducible. Let 2 H be any unit vector and
consider HJ D f
.x/ j x 2 Jg. Since J is an ideal of A, HJ is
.A/-invariant. But
.A/ is an irreducible operator algebra, and so HJ D f0g or HJ D H. We show that

only the latter condition holds. If it were true that HJ D f0g, then
.x/
.a/ D 0
for every x 2 J and a 2 A; but vectors of the form
.a/, a 2 A, are dense in H, and
so
.x/ would be zero for every x 2 J, in contradiction to J 6 ker
. Thus, HJ D H.
The choice of 2 H being arbitrary shows that
jJ is an irreducible representation
of J.
For the second statement, assume W J ! B.H/ is an irreducible representation
of J. Choose any unit vector x 2 H. For a 2 A define
.a/ on the dense linear
submanifold f.x/ j x 2 Jg by
.a/ .x/ D .ax/. Then
.a/ extends to an
operator on H and the map a 7!
.a/ determines an irreducible representation of
A on H.
t
u
429
Definition 11.83. Assume that J is an ideal of a C -algebra A.

1. J is a primitive ideal if J D ker
for some irreducible representation
of A.
2. J is a prime ideal if, for any ideals I and K of A, the inclusion I \ K J holds
only if I J or K J.
The set of all primitive ideals of A is denoted by PrimA.
Proposition 11.84. J is prime, for every J 2 PrimA.
Proof. Suppose that J 2 PrimA and assume that I and K are ideals of A such
that I \ K J. Suppose that K 6 J. Then K 6 ker
, where
W A ! B.H/ is
an irreducible representation of A with ker
D J. Hence,
jK W K ! B.H/ is an
irreducible representation of K. Therefore, if 2 H is a fixed unit vector, then
H D
.K/ . In particular if x 2 I and y 2 K, then
.x/ .
.y// D
.xy/ D 0,
as xy 2 I \ K J D ker
. But vectors of the form
.y/, y 2 K, are dense in H;
thus,
.x/ D 0 for every x 2 I, which proves that I J.
t
u
The converse to Proposition 11.84 is false in general, although it is true if A is
separable.
Proposition 11.85. If I is an ideal of A, then
\
ID
J:
IJ; J2PrimA
Proof. Clearly I is a subset of the ideal on the right-hand side of the equation above.
To show the other inclusion, assume that x 62 I. Thus, 0 6D xP 2 A=I and so there is
an irreducible representation of A=I such that k.Px/k D kPxk. If q W A ! A=I is the
canonical quotient homomorphism, then
D q is an irreducible representation
of A with I ker
and x 62 ker
. Therefore, x is not an element of the right-hand
side of the equation above.
t
u
Definition 11.86. Assume that F is a nonempty subset of PrimA. The closure of
F , which is denoted by F , is the set
\
F D fJ 2 PrimA j
I Jg :
I2F
A subset F PrimA is closed if F D F .

The closure operation satisfies the following properties:
(i)
(ii)
(iii)
(iv)
; D ;;
F F;
F D F;
F1 [ F2 D F1 [ F2 .
Proposition 11.87. There exists a unique topology T on PrimA in which the closed
sets F of PrimA are precisely those in which F D F , where F is given by
Definition 11.86.
430
t
u
The topology T in Proposition 11.87 is called the Jacobson topology, or the

hull-kernel topology, on PrimA.
Definition 11.88. Two representations
1 and
2 of a C -algebra A on Hilbert
spaces H1 and H2 , respectively, are equivalent, which is denoted by
1

2 , if
there is a surjective isometry u W H1 ! H2 such that
2 .a/u D u
1 .a/, for all a 2 A.
The next proposition is basically self evident.
Proposition 11.89. If Irr A denotes the set of irreducible representations of A, then
is an equivalence relation on Irr A. Furthermore, if

1

2 , then ker
1 D ker
2 .
Because equivalent irreducible representations have the same kernels, it is
convenient to identify such representations by passing to the space Irr A=
.
O is the set Irr A=
of equivaDefinition 11.90. The spectrum of A, denoted by A,
lence classes of irreducible representations of A.
Elements of AO will be denoted by
,
P where
2 Irr A.
Definition 11.91. Assume that J is an ideal of a C -algebra A.
1. J is a primitive ideal J if J D ker
for some irreducible representation
of A.
2. J is a prime ideal if, for all ideals I1 ; I2 A, the inclusion I1 \ I2 J holds only
if I1 J or I2 J.
The set of primitive ideals of A will be denoted by PrimA. Note that the map
AO ! PrimA ;
P 7! ker
;
is a surjection, and via this surjection one endows the spectrum of A with a topology
as follows.
Definition 11.92. A subset U AO is an open set if fker
j
P 2 Ug is open in the
Jacobson topology of PrimA.
Thus, the surjection AO ! PrimA is an open, continuous map.
For the remainder of this section, the spectra and primitive ideal spaces defined
above are used to describe -homomorphisms between abelian C -algebras and to
analyse the structure of ideals in such algebras.
Proposition 11.93. If A D C0 .X/, where X is locally compact and Hausdorff, then
X ' AO ' PrimA.
Proof. For each t0 2 X let t0 W A ! C be given by t0 .f / D f .t0 /, for all f 2 A.
Although AO D ft j t 2 Xg as sets, it is not yet obvious that AO and RA have the same
topologies. We first show that X is homeomorphic to PrimA by identifying the closed
sets of PrimA with closed sets of X.
Note that J 2 PrimA if and only if there is a t0 2 X such that J D ff 2 A j f .t0 / D
0g. Suppose that F PrimA is arbitrary. Thus, there is a subset F X such that
431
F D fker t j t 2 Fg. Claim:

F D F. To prove the claim, let t0 2 F. By definition,
T
ker t 2 F if ker t I2F I. But this is true, since (by continuity) we have
\
I D ff 2 C0 .T/ j f .t/ D 0 ; 8 t 2 Fg ff 2 C0 .T/ j f .t0 / D 0g :
I2F
Conversely, assume that J D ker t0 2 PrimA is such that J 62 fker t j t 2 Fg. Thus,
t0 62 F. By Theorem
2.43, there is an f 2 CT
0 .X/ such that f .t0 / D 1 and f .F/ D f0g.
T
That is, f 2 I2F I but f 62 J; hence, J 6 I2F I, which is to say that J 62 F . This
proves that T and PrimA have the same closed sets, and so PrimA is locally compact
and Hausdorff.
Because the open, continuous, surjective map AO ! PrimA is injective (for if t1 6D
t2 , then there is an f 2 C0 .X/ such that f .t1 / 6D f .t2 /, and so g D f f .t1 / belongs
to ker t1 but g 62 ker t2 ), it is also true that AO and PrimA are homeomorphic.
t
u
Proposition 11.94 (Poincar Duality). Assume that X and Y are locally compact
Hausdorff spaces. Then a map
W C0 .X/ ! C0 .Y/ is a -homomorphism if and only
if there is a continuous map W Y ! X such that
1. 1 .K/ is compact in Y for all every compact K X, and
2.
.f / D f , for all f 2 C0 .X/.
Moreover, is injective if and only if
is surjective, and
if
is injective.
is surjective if and only
Proof. The sufficiency of the two conditions is clear, as the second defines a
-homomorphism C0 .X/ ! C.Y/, and the first condition shows that f vanishes
at infinity, so that indeed f 2 C0 .Y/.
Conversely, given a -homomorphism
W C0 .X/ ! C0 .Y/, when
is composed
with a point evaluation y W C0 .Y/ ! C the result is a nonzero -homomorphism
C0 .X/ ! C. Hence, y
D x for some (uniquely determined) x 2 X. Let W Y ! X
be the function that sends y to x. To show that is continuous, we take advantage
of the homeomorphisms X ' PrimC0 .X/ and Y ' PrimC0 .Y/.
Let C X be a closed set; thus, fker x j x 2 Cg is closed in PrimC0 .X/. Choose
y0 2 1 .C/ and let x0 2 X be such that x0 D y0
. Because Y ' PrimC0 .Y/,
ker y0 is in the closure of fker y j y 2 1 .C/g in PrimC0 .Y/. That is,
\
y2
ker y ker y0 :
1 .C/
Hence, it is also true that

\
y2
1 .C/
ker y
ker y0
:
432
That is, ker x0 is in the closure of fker x j x 2 Cg in PrimC0 .X/. But this is equivalent
to x0 2 C. As C D C, we obtain y0 2 1 .C/, which proves that 1 .C/ is closed.
Hence, W Y ! X is a continuous function.
Let K X be compact. As X is Hausdorff, K is closed; thus, 1 .K/ is closed in
Y, by the continuity of . By Theorem 2.43, there is a function f 2 C0 .X/ such that
f .K/ D f1g. Thus,
1
fy 2 Y j j
.f /.y/j g
2
1
.K/ :
Because
.f / vanishes at infinity, the set on the left-hand side above is compact,
which implies that 1 .K/ is compact.
Because
.f / D f , for all f 2 C0 .X/, we obtain the desired formula for
.
The remaining assertions of the proposition are straightforward to deduce.
t
u
\
f 1 f0g. Then
Proposition 11.95. Assume that I C0 .X/ is an ideal. Let Z D
f 2I
I D fg 2 C0 .X/ j gjZ D 0g.
Proof. Clearly I fg 2 C0 .T/ j gjZ D 0g. To prove the converse, assume that h 62 I.
Every ideal is the intersection of primitive ideals that contain it, and so
I D
ker t :
ker t I
Therefore, h 62 I implies that h 62 ker t0 for some t0 2 X with I ker t0 . Thus, t0 2 Z
and h.t0 / 6D 0 imply that h 62 fg 2 C0 .X/ j gjZ D 0g.
t
u
Proposition
11.96. If I C0 .X/ is an ideal, then I C0 .U/, where U D X n Z and
\
ZD
f 1 f0g.
f 2I
Proof. The C -algebra C0 .X/=I is abelian. Let Y denote its maximal ideal space.
We obtain a surjective homomorphism
W C0 .X/ ! C0 .Y/ via
C0 .X/ ! C0 .X/=I C0 .Y/ :
By Poincar Duality, there is a continuous injective function
W Y ! X such
that
.f / D f for every f 2 C0 .X/. Note that I D ker
, which means (by
Theorem 2.43) that Z D .Y/. Consider the isometric -homomorphism W I !
C.U/ whereby .f / D fjU for every f 2 I. If " > 0, then
ft 2 U j jf .t/j "g D ft 2 U [ Z j jf .t/j "g I
the set on the right-hand side is compact, since f 2 C0 .X/. Therefore, .f / 2 C0 .U/.
To see that is surjective, let f 2 C0 .U/C and define F W X ! RC by F.t/ D f .t/,
if t 2 U, and by F.t/ D 0, if t 2 Z. Let r 2 RC . The set F 1 .1; r is Z [ f 1 .0; r,
which is closed. Hence, F is lower semicontinuous. The set F 1 r; 1/ D f 1 r; 1/,
11.9 Traces and Factors
433
which is also closed; therefore, F is upper semicontinuous. Thus F is continuous.

Moreover, F vanishes at infinity and f D .F/. Since C0 .U/ is spanned by functions
in C0 .U/C , this proves that is surjective.
t
u

The set Mn .C/ of n n complex matrices is a unital C -algebra, since Mn .C/ D
B.Cn /, the C -algebra of bounded linear operators on the Hilbert space Cn . Of the
many important functions defined on matrices, the trace functional is especially
relevant to functional analysis, as it is a positive linear functional that behaves as
though Mn .C/ were commutative. To be more precise, the function Tr W Mn .C/ ! C
defined by
n

X
Tr ij ni;jD1 D
kk ;
kD1
is linear, has the property that Tr.x x/ 0 for all x 2 Mn .C/, and satisfies Tr.xy/ D
Tr.yx/, for all x; y 2 Mn .C/. There is one additional feature: the trace is faithful,
which is to say that Tr.x x/ D 0 only if x D 0.
Concerning Mn .C/ as an operator algebra, it is of course a von Neumann algebra
and its centre, Z .Mn .C// is trivial in the sense that Z .Mn .C// D f 1 j 2 Cg, where,
for any ring R, the centre of R is the abelian subring Z.R/ of R defined by
Z.R/ D fx 2 R j xy D yx; 8 y 2 Rg:
One additional feature of Mn .C/ is that its only ideals are f0g and Mn .C/, making it
a simple algebra.
The purpose of this section is to introduce and examine operator algebras that
exhibit these same algebraic and functional-analytic properties that are present in
the matrix algebra Mn .C/. The first fundamental concept is that of a trace.
Definition 11.97. A state on a C -algebra A is a trace if .xy/ D .yx/, for all
x; y 2 A. If, in addition, .x x/ D 0 only if x D 0, then is said to be a faithful trace.
Example 11.98. If G is a countable discrete group, then the group von Neumann
algebra V .G/ has a faithful trace.
Proof. Let 2 `2 .G/ be the unit vector D e , where e is the identity of G, and
define W V .G/ ! C by .x/ D hx; i. Thus, is a state on V .G/, and it remains
to show that is a faithful trace.
If g; h 2 G, then g h e D gh , and so g h e ; e D hgh ; e i, which

is nonzero
if
and
only
if
gh
D
e.
But
gh
D
e
if
and
only
if
hg
D
e;
therefore,

e ; e D
h
g
g h e ; e . Hence, if x; y 2 V .G/ are elements of the group algebra CG, then

.xy/ D .yx/. Because multiplication in a Banach algebra is norm continuous
(Exercise 9.45), we deduce that .xy/ D .yx/ for all x; y 2 C .G/.
434
To show that .xy/ D .yx/ for x; y 2 V .G/, it is sufficient to assume that x

and y are hermitian and that kxk D kyk D 1 (because the hermitian operators of
V .G/ span V .G/). By the Density Theorem, there are nets fx g and fy g of
hermitian operators in the unit ball of C .G/ that are SOT-convergent to x and y,
respectively. In particular, with the unit vector and making use of the fact that
hy ; x i D .x y / D .y x / D hx ; y i for all and , we see that

j .xy/ .yx/j 2 kx x k C ky y k ;
and so j .xy/ .yx/j D 0. Thus, is a trace on V .G/.
Suppose now that x 2 V .G/ satisfies .x x/ D 0. Thus, 0 D hx xe ; e i D kxe k2 .
Select anyother orthonormal
basis vector, say g . Thus, g D eg D .g/e , where

W G ! B `2 .G/ is the right regular representation. Because g 2 V .G/0 , we have
that xg D xg e D g xe D 0. Hence, x maps every orthonormal basis vector g to
0, which proves that x D 0.
t
u
The proof employed in Example 11.98 shows how a certain type of trace
SOT
functional on a C -algebra A extends to a trace on A . This fact is recorded below
for later reference.
Proposition 11.99. If is a unit cyclic vector for a unital C -algebra A acting on
a Hilbert space H, and if the state W B.H/ ! C defined by .x/ D hx; i is a trace
SOT
on A, then is also a trace on M D A .
Proposition 11.99 is particularly useful in the following form.
Corollary 11.100. If .H
;
; / is a GNS-triple for a trace on a unital C -algebra
SOT
A, then there is a trace M on the von Neumann algebra M D
.A/
such that
D M
.
Again motivated by the situation with the matrix algebra Mn .C/, von Neumann
algebras with trivial centres are of particular interest.
Definition 11.101. A von Neumann algebra M is a factor if Z.M/ D f 1 j 2 Cg.
The von Neumann algebra B.H/ is a factor (Exercise 11.141), and it is known,
for infinite discrete groups G, that V .G/ is a factor if and only if the conjugacy
class fh1 gh j h 2 Gg of g 2 G is infinite for every g 6D e. (The free group Fn on
n 2 generators is one easy example of a group that has infinite conjugacy classes.)
Another manner in which factors are obtained is through the tracial state space of a
C -algebra (Proposition 11.103 below).
Definition 11.102. Assume that A is a unital C -algebra.
1. The set T.A/ of all traces on A is called the tracial state space of A.
SOT
2. If .H
;
; / is a GNS-triple for a trace 2 T.A/, and if
.A/
is a factor, then
is said to be a factorial trace.
435
It may happen that a C -algebra does not admit a trace (for example, B.H/ does
not, if H has infinite dimension), or it may admit a continuum of traces.
Proposition 11.103. If A is a unital C -algebra, then the tracial state space T.A/
is weak*-compact and convex. Moreover, every extreme point of T.A/ is a factorial
trace.
Proof. The assertion that T.A/ is weak*-compact and convex is left as Exercise 11.142.
Suppose that is an extreme point of T.A/, and suppose that .H
;
; / is a GNSSOT
triple for , and let M D
.A/ . Assume, contrary to what we aim to prove, that
is not factorial. Thus, the centre Z.M/ of M is nontrivial. Because Z.M/ is an
abelian von Neumann algebra and is nontrivial, there is a projection p 2 Z.M/ such
that p 6D 0 and p 6D 1. If 1 D p, then because f
.x/ j x 2 Ag is dense in H
, and
because
.x/1 D
.x/p D p
.x/ for all x 2 A, it cannot be that 1 D 0. Similarly,
if 2 D .1 p/1 , then 2 6D 0. Thus, k1 k2 D 1 k2 k2 < 1.
For each j 2 f1; 2g, let j D kj k1 j and j .x/ D h
.x/j ; j i, for x 2 A. Thus, j
SOT
WOT
is a state on A. Since p 2 Z.M/ M D

.A/
D
.A/
, there is a net fx g of
hermitian elements in A such that f
.x /g is convergent to p in the weak operator
topology of B.H
/. Therefore, for fixed x; y 2 A, we have that
h
.xy/1 ; 1 i D hp
.xy/; i D limh
.x /
.xy/; i D limh
.x xy/; i
D lim .x xy/ D lim .yx x/ D limh

.x /
.x/;
.y/ i
D hp
.x/;
.y/ i D h
.y/p
.x/; i D hp
.yx/; i
D h
.yx/1 ; 1 i:
Hence, 1 is a trace on A. Similarly, 2 2 T.A/.
Let sj D kj k2 2 .0; 1/ so that s1 C s2 D 1 and D s1 1 C s2 2 . Because
each .x / D h
.x /; i, the net f .x /g is convergent to hp; i D k1 k2 < 1,
whereas 1 .x / D k1 k2 h
.x /1 ; 1 i implies that net f1 .x /g is convergent to
k1 k2 hp1 ; 1 i D k1 k2 k1 k2 D 1. Therefore, 1 6D , in contradiction to the
hypothesis that is an extreme point of T.A/. Hence, it must be that Z.M/ is trivial,
which is to say that is factorial.
t
u
With the theory and examples developed to this point, we conclude this chapter
with the construction of a unital simple C -algebra with a unique trace.
n

1 X
jj . Thus, n is a trace on
Let n W M2n .C/ ! C be given by ij ni;jD1 D n
2 jD1
M2n .C/ and basic linear algebra shows that it is the only element of the tracial state
space T .M2n .Cn //. Denote the zero and identity matrices in M2n .C/ by 0n and 1n ,
respectively.
436
For n 2 N, let #n W M2n .C/ ! M2nC1 .C/ be defined by

#n .x/ D x x;
and note that #n is an injective -homomorphism and is trace preserving in the sense
that n D nC1 #n . Consider now
B D f.xn /n2N j xn 2 M2n .C/ and sup kxn k < 1g
n
and
J D f.xn /n2N 2 B j lim kxn k D 0g:
n
The set B is clearly a unital C -algebra with respect to the algebraic operations
and involution induced by each coordinate, and the norm k.xn /n k D sup kxn k, while
n
J is evidently an ideal of B. Let Q D B=J and denote the identity of Q by 1

and the canonical quotient -homomorphism B ! Q by q. For each n 2 N, let
n W M2n .Cn / ! B and
n W M2n .Cn / ! Q be defined by
n .x/ D .01 ; : : : ; 0n1 ; x; #n .x/; #nC1 #n .x/; #nC2 #nC1 #n .x/; : : : /
and
n D q n :
Both n and
n are injective -homomorphisms, and
n .1n / D 1 (see Exercise 11.143). Let An D
n .M2n .Cn //, which is a finite-dimensional unital C subalgebra of Q, and observe that An AnC1 for every n (Exercise 11.143). Let
A0 D
An
and
A D A0 :
n2N
The unital C -subalgebra A of Q above is called the Fermion C -algebra.

Proposition 11.104. The Fermion C -algebra is a simple unital C -algebra with
trivial centre and a unique trace . Furthermore, the trace is faithful and factorial.
Proof. Let K A be an ideal of A such that K 6D f0g. If K \ An D f0g for every
[
.An \ K/ is dense in K, we would have K D f0g; therefore, it
n 2 N, then because
n2N
must be that K \ An 6D f0g, for some n. Because An M2n .C/ is simple, we deduce
that K \ An D An . However, as 1 2 An , we also have that 1 2 K, and so K D A, which
proves that A is a simple C -algebra. The fact that Z.A/ D f 1 j 2 Cg is left as
Exercise 11.143.
437
If x 2 An for some n, then x D

n .a/ for some matrix a 2 M2n .C/, and we may
define the trace 0 .x/ of x by n .a/. If k > n, then there is a unique b 2 M2k .C/
with x D
k .b/, namely b D #k1 #n .a/ 2 M2k .C/. (The uniqueness of b is on
account of the injectivity of
k .) Moreover, the trace k .b/ of b is the same as the
trace n .a/ of a. Hence, the function 0 W A0 ! C defined by 0 .x/ D n .a/, if x D
n .a/ for some n and some a 2 M2n .C/, is well defined, linear, positive, bounded,
and satisfies 0 .xy/ D 0 .yx/ for all x; y 2 A0 . Hence, 0 extends to a trace on A.
The uniqueness of follows from the fact that A0 is dense in A and from the fact
that each matrix algebra M2n .C/ has a unique tracial state. Therefore, T.A/ D fg,
which implies that is an extreme point of T.A/. Therefore, by Proposition 11.103,
is a factorial trace.
To show that is faithful, consider the set D D fx 2 AC j .x/ D 0g AC . If
x 2 D and y 2 AC , then j .yx/j D j .xy/j D .x1=2 yx1=2 / kyk.x/ D 0. Therefore,
if x 2 Span D and if y 2 AC , then it is also true that j.yx/j D j.xy/j D 0. Hence,
because AC spans A, and using the continuity of , we obtain .xy/ D 0 for every
x 2 K D Span D and every y 2 A. In other words, K is an ideal of A. Evidently 1 62 K;
thus, by the fact that A is simple, it must be that K D f0g. Hence, D D f0g, which
proves that is a faithful trace.
t
u
Corollary 11.105. If .H
;
; / is a GNS-triple for the unique faithful trace on
SOT
the Fermion algebra A, then R D
.A/
is a factor and has a faithful trace R
such that R .
.a// D .a/, for every a 2 A.
Proof. Proposition 11.100 shows the existence of the trace R with the stated
property R
D , defined by R .x/ D hx; i for x 2 R, while Proposition 11.104
asserts that is factorial. All that remains is to show that the trace R is faithful.
Suppose that x 2 R satisfies R .x x/ D 0. Thus, 0 D hx x; i D kxk2 , and so
x D 0. For any a 2 A,
kx
.a/k2 D hx
.a/; x
.a/i D R .
.a/ x x
.a//
D R .
.a/
.a/ x x/ D h
.a/
.a/ x x; i
D 0;
as x D 0. Thus, x is zero on the dense linear submanifold f
.a/ j a 2 Ag, and
therefore x D 0.
t
u
Definition 11.106. The factor R described by Corollary 11.105 is called the
hyperfinite II1 -factor.
Of course, the definition of the hyperfinite II1 -factor R depends on the choice of
GNS triple for the trace on the Fermion algebra A; however, this is not really an
issue because of the uniqueness assertion in the GNS theorem. The distinguishing
features of R are that it is a factor, it possesses a faithful trace, and it has a unital
separable strongly dense C -subalgebra attained from an increasing sequence of
finite-dimensional factors (that is, simple matrix algebras).
438
Problems
11.107. Assume that H is a Hilbert space, and consider the C -algebra K.H/ of
compact operators.
1. Show that, with respect to the strong operator topology of B.H/, the identity
operator 1 2 B.H/ is the limit of a net of compact operators.
2. Show that K.H/ is a von Neumann algebra if and only if H has finite dimension.
11.108. Prove that if .X; ; / is a -finite measure space, then L1 .X; ; / is a
von Neumann algebra when considered as a C -algebra of multiplication operators
acting on the Hilbert space L2 .X; ; /.
11.109. Assume that X is a compact Hausdorff space and that is a finite regular
Borel measure on the -algebra of Borel sets of X. Consider the following unital
operator algebras C and M acting on L2 .X; ; / as algebras of multiplication
operators:
C D fMf j f 2 C.X/g
and
M D fM j
2 L 1 .X; ; /g:
1. Prove that C0 D M.
SOT
2. Prove that C
D M.
11.110. Prove that a -subalgebra A B.H/ is irreducible if and only if the only
projections p 2 B.H/ that belong to the commutant A0 of A are p D 0 and p D 1.
11.111. Show that the von Neumann algebra generated by the unilateral shift
operator on `2 .N/ is B `2 .N/ .
11.112. Show that the von Neumann algebra L1 .0; 1; M; m/ is not irreducible.
11.113. Let A be a unital C -subalgebra of B.H/ and let M D A
SOT
1. Prove that the matrix operator algebras M2 .A/ and M2 .M/ acting on H H satisfy
SOT
M2 .N/ D M2 .A/ .
2.
Prove
that if x 2 M has norm kxk 1, then the hermitian operator matrix X D
0 x
has norm kXk 1.
x 0
3. Prove that there exists a net of hermitian, contractive operator matrices .B /2
in M2 .A/ converging to X in the strong operator topology of M2 .B.H//.
4. Prove that there exists a net .b /2 of contractive operators b 2 A converging
to x in the strong operator topology of B.H/.
11.114. Assume that H is an infinite-dimensional separable Hilbert space. Let
B.H/1 denote the closed unit ball of B.H/, and let TSOT denote the strong operator
topology on B.H/1 .
Problems
439
1. Use the Kaplansky Density Theorem to show that .B.H/1 ; TSOT / is a separable
topological space.
2. Show that .B.H/1 ; TSOT / is a metrisable topological space. (Suggestion: select a
countable dense subset fn gn of the closed unit ball of H and show that d, defined
1
X
1
by d.x; y/ D
kxn yn k, is a metric inducing the topology TSOT .)
2n
nD1
11.115. Suppose that a normal operator N on a separable Hilbert space has the
property that the von Neumann algebra N generates has a cyclic vector. Prove that
there exists a regular Borel probability measure on the Borel sets of .N/ and a
unital isometric -isomorphism W W .N/ ! L1 . .N/; ; / such that .N/ D P ,
where W .N/ ! C is given by .t/ D t.
11.116. If M is a von Neumann algebra acting on a Hilbert space H, and if a; b 2
MC , then prove that the following statements are equivalent for x 2 M:

a x
1.
is a positive operator on H H;
x b
2. x D a1=2 yb1=2 for some y 2 M with kyk 1.
11.117. Prove that a unital abelian C -algebra A is a semisimple (Definition 9.32).
11.118. Prove that if A is a nonunital C -algebra, then 0 2 .x/, for all x 2 A.
11.119. Suppose that A is a C -algebra with norm k k. Prove that if k k0 is a norm
on A that satisfies all of the axioms of a C -norm, then kxk0 D kxk for all x 2 A.
11.120. If A is a unital C -algebra and if u 2 A is unitary, then prove that .u/ @D,
where D is the open unit disc of the complex plane.
11.121. If A is a unital C -algebra, then prove that the set of unitary elements of A
if the form eih for some hermitian h 2 A is a path-connected set.
11.122. If A is a unital C -algebra and if x 2 A (not necessarily normal), then prove
or find a counterexample to each of the following statements.
1. .x / D f j 2 .x/g.
2. x x is invertible if x is invertible.
3. x is invertible if x x is invertible.
11.123. Suppose that X R is a compact set such that 0 2 X. Prove that if f 2 C.X/
satisfies f .0/ D 0 and if " > 0, then there is a polynomial p such that p.0/ D 0 and
jf .t/ p.t/j < " for all t 2 X.
11.124. In a unital Banach algebra A, an element x 2 A is quasinilpotent if
.x/ D f0g, and x is properly quasinilpotent if .xy/ D f0g for all y 2 A. Prove
that if A is a unital C -algebra, then the only properly quasinilpotent element x 2 A
is x D 0.
440
11.125. If A is a C -algebra, and if 2 RC and h; k 2 AC , then prove the following

assertions:
1. h 2 AC ;
2. h C k 2 AC ; and
3. h 2 AC only if h D 0.
11.126. Suppose that a; b; c 2 Asa . Prove the following assertions:
1. a a;
2. if a b and b a, then b D a; and
3. if a b and b c, then a c.
11.127. Suppose that A is a C -algebra with x 2 A, h 2 AC , and xh D hx. Prove that
xh1=2 D h1=2 x.
11.128. If A is a C -algebra and if a; b 2 AC satisfy a b and ab D ba, then prove
that a2 b2 . Show by example that a b does not always imply a2 b2 if ab 6D ba.
11.129. If A is a unital C -algebra, prove that 1 C x x is invertible for every x 2 A.
11.130. Prove that .ab/ RC for all positive elements a and b in a C -algebra A.
11.131. Prove that if A is a unital C -algebra and if x 2 A is invertible, then there is
a unitary u 2 A such that x D ujxj.
11.132. Prove that if h 2 Asa , where A is a unital C -algebra, and if khk, then
1. h 1, and
2. k1 hk .
11.133. Suppose that J is an ideal of a nonunital C -algebra A. Consider the
inclusion of A in its unitisation A1 . Show that J is an ideal of A1 .
11.134. Suppose that J is a proper ideal of a unital C -algebra A. Define J C C1 by
J C C1 D fx C 1 j x 2 J; 2 Cg :
1. Prove that J C C1 is a unital C -subalgebra of A.
2. Prove or find a counterexample to the following statement: the C -algebras J 1
and J C C1 are isometrically isomorphic.
11.135. Prove that if A and B are unital C -algebras and if W A ! B is a
-homomorphism, then ..x// .x/, for all x 2 A.
11.136. Prove that if A and B are C -algebras and if W A ! B is a
-homomorphism, then spr .x x/ spr .x x/, for all x 2 A. (A and B are not
assumed to be unital.)
11.137. Let q W B.H/ ! B.H/=K.H/ be the unital -homomorphism in which
q.x/ D xP , for every x 2 B.H/. Compute the spectrum of q.s/, where s 2 B.H/ is
the unilateral shift operator on H D `2 .N/.
Problems
441
11.138. Assume that ' is a state on a unital C -algebra A and that .H

;
; / and
.H ; ; / are GNS-triples for '. Prove that there is a surjective isometry u W H
! H
such that u D and u
.x/ D .x/u, for every x 2 A.
11.139. If fn gn is a sequence in a Hilbert space H such that
1
X
kxn k2 converges,
nD1
and if n 6D 0 for at least one n, then prove that the positive linear functional ' on
B.H/ defined by
'.x/ D
1
X
hx; i;
nD1
has the property that '.k/ 6D 0 for at least one compact operator k 2 B.H/.
11.140. Prove that there exists a unique topology T on PrimA in which a subset F
of PrimA is closed if and only if
F D fJ 2 PrimA j
I Jg :
I2F
11.141. Prove that B.H/ is a factor.

11.142. Prove that the tracial state space T.A/ of a unital C -algebra A is weak*compact and convex.
11.143. Let J D f.xn /n2N 2 B j limn kxn k D 0g, where
B D f.xn /n2N j xn 2 M2n .C/ and sup kxn k < 1g;
n
and let Q D B=J. Denote the quotient map B ! Q by q and the identity elements of
M2n .C/ and Q by 1n and 1, respectively. Let #n W M2n .C/ ! M2nC1 .C/ be defined by
#n .x/ D x x.
1. Prove that
n W M2n .C/ ! Q is an isometric -homomorphism such that
n .1n / D
1, where
n D q n and n W M2n .C/ ! B is defined by
n .x/ D .01 ; : : : ; 0n1 ; x; #n .x/; #nC1 #n .x/; : : : /:
2. If An D
n .M2n .C// for each n 2 N, then prove that An AnC1 .
[
An , then prove that the centre Z.A/ of A is Z.A/ D f 1 j 2 Cg.
3. If A D
n2N
11.144. Prove that the trace R on the hyperfinite II1 -factor R is unique.
References
1. Akemann, C., Anderson, J., & Pedersen, G. K. (1982). A triangle inequality in operator
algebras. Linear Multilinear Algebra, 11, 1679178.
2. Ando, T. (1963). Note on invariant subspaces of a compact normal operator. Archiv der
Mathematik, 14, 337340.
3. Arveson, W. B. (1976). An invitation to C -algebra. New York: Springer-Verlag.
4. Arveson, W. B. (2002). A short course in spectral theory. New York: Springer-Verlag.
5. Alexander, H., & Wermer, J. (1998). Several complex variables and Banach algebras.
New York: Springer-Verlag.
6. Bartle, R. G. (1966). The elements of integration. New York: John Wiley & Sons, Inc.
7. Bourdon, P. S. (2016). Spectra of composition operators with symbols in S .2/. Journal of
Operator Theory, 75, 321348.
8. Brown, J. W., & Churchill, R. V. (2009). Complex variables and applications. New York:
McGraw-Hill, Inc.
9. Chaktoura, M. (2013). Personal communication.
10. Cohn, D. L. (2013). Measure theory (2nd ed.). New York: Birkhuser.
11. Conway, J. B. (1972). The compact operators are not complemented in B.H/. Proceedings of
the American Mathematical Society, 32, 549550.
12. Conway, J. B. (1985). A course in functional analysis. New York: Springer-Verlag.
13. Conway, J. B. (2000). A course in operator theory. Providence: American Mathematical
Society.
14. Davidson, K. R., & Donsig, A. P. (2002). Real analysis with real applications. Upper Saddle
River, NJ: Prentice Hall, Inc.
15. Davies, E. B. (1976). Quantum theory of open systems. New York: Academic Press.
16. Davies, E. B. (2007). Linear operators and their spectra. Cambridge: Cambridge University
Press.
17. Davis, C. (1963). Notions generalizing convexity for functions defined on spaces of matrices.
Proceedings of Symposia in Pure Mathematics, 7, 1987201.
18. DiBenedetto, E. (2002). Real analysis. Boston: Birkhuser.
19. Douglas, R. G. (1964). On extremal measures and subspace density. Michigan Mathematical
Journal, 11, 243246.
20. Douglas, R. G. (1998). Banach algebra techniques in operator theory (2nd ed.). New York:
Springer-Verlag.
21. Dugundji, J. (1966). Topology. Boston: Allyn and Bacon, Inc.
22. Dunford, N., & Schwartz, J. T. (1957). Linear operators, Part I. New York: John Wiley &
Sons, Inc.

DOI 10.1007/978-3-319-45633-1
443
444
References
23. Duren, P. (2012). Invitation to classical analysis. Providence: American Mathematical Society.
24. Enflo, P. (1973). A counterexample to the approximation problem in Banach spaces. Acta
Mathematica, 130, 309317.
25. Fabian, M., Habala, P., Hjek, P., Montesinos, V., & Zizler, V. (2011). Banach space theory.
New York: Springer-Verlag.
26. Farenick, D. R. (2000). Algebras of linear transformations. New York: Springer-Verlag.
27. Farenick, D., Kavruk, A., & Paulsen, V. I. (2013). C -algebras with the weak expectation
property and a multivariable analogue of Andos theorem on the numerical radius. Journal of
Operator Theory, 70, 573590.
28. Gillman, L., & Jerison, M. (1960). Rings of continuous functions. New York: Van Nostrand.
29. Gohberg, I. C., & Kren, M. G. (1969). Introduction to the theory of linear nonselfadjoint
operators. Providence: American Mathematical Society.
30. Hadwin, D., & Ma, X. (2008). A note on free products. Operators and Matrices, 2, 5365.
31. Halmos, P. R. (1957). Introduction to Hilbert space (2nd ed.). New York: Chelsea.
32. Halmos, P. R. (1982). A Hilbert space problem book (2nd ed.). New York: Springer-Verlag.
33. Herstein, I. M. (1964). Topics in algebra. New York: Blaisdell.
34. Enflo, P. (1970). Atomic and nonatomic measures. Proceedings of the American Mathematical
Society, 25, 650655.
35. Kadison, R. V., & Ringrose, J. R. (1983). Fundamentals of the theory of operator algebras
(Vol. I). New York: Academic Press.
36. Katznelson, Y. (2004). An introduction to harmonic analysis (3rd ed.). Cambridge: Cambridge
University Press.
37. Lang, S. (1993). Real and functional analysis (3rd ed.). New York: Springer-Verlag.
38. Lindenstrauss, J. (1966). A short proof of Liapounoffs convexity theorem. Journal of
Mathematics and Mechanics, 15, 971972.
39. McCann, S. J. (2008). Equivalent forms of the Connes embedding problem. M.Sc. Thesis,
University of Regina.
40. Michaels, A. J. (1977). Hildens simple proof of Lomonosovs invariant subspace theorem.
Advances in Mathematics, 25, 5658.
41. Munkres, J. R. (2000). Topology (2nd ed.). Upper Saddle River, NJ: Prentice Hall, Inc.
42. Paulsen, V. I. (2001). Completely bounded maps on operator algebras. Cambridge: Cambridge
University Press.
43. Phelps, R. R. (2001). Lectures on Choquet theory (2nd ed.). Lecture Notes in Mathematics,
Vol. 1757. New York: Springer-Verlag.
44. Radjavi, H., & Rosenthal, P. (1973). Invariant subspaces. New York: Springer-Verlag.
45. Raeburn, I., & Williams, D. P. (1998). Morita equivalence and continuous-trace C -algebras.
Mathematical Surveys and Monographs, Vol. 60. Providence: American Mathematical Society.
46. Read, C. J. (1986). A short proof concerning the invariant subspace problem. Journal of the
London Mathematical Society, 34, 335348.
47. Rosenholtz, I. (1976). Another proof that any compact metric space is the continuous image of
the Cantor set. American Mathematical Monthly, 83, 646647.
48. Rosenthal, P. (1968). Completely reducible operators. Proceedings of the American Mathematical Society, 19, 826830.
49. Rudin, W. (1964). Principles of mathematical analysis. New York: McGraw-Hill, Inc.
50. Rudin, W. (1987). Real and complex analysis (3rd ed.). New York: McGraw-Hill, Inc.
51. Rudin, W. (1991). Functional analysis (2nd ed.). New York: McGraw-Hill, Inc.
52. Stein, E. M., & Shakarchi, R. (2005). Real analysis. Princeton: Princeton University Press.
53. Stein, E. M., & Shakarchi, R. (2011). Functional analysis. Princeton: Princeton University
Press.
54. Takesaki, M. (1979). Theory of operator algebras, I. New York: Springer-Verlag.
55. Thompson, R. C. (1976). Convex and concave functions of singular values of matrix sums.
Pacific Journal of Mathematics, 66, 285290.
Index
Symbols
.X; ; /, 84
.X; T /, 8
-homomorphism, 406, 411
-isomorphism, 406, 411
A.D/, 312
AC.T/, 323
B.V/, 217
B.V; W/, 217
C.X/, 176
C.R=2
Z/, 195
Cb .X/, 175
Cc .X/, 181
F.H/, 375
F.V/, 290
F -set, 61
G -set, 61
K.V/, 290
L1 -space, 188
Lp .X; ; /, 182
1
LR
.X; ; /, 128
M.X; /, 189
T.H/, 377
`1 -space, 188
`1 .N/, 188
`1 .n/, 188
`p -space, 184
`p .N/, 184
`p .n/, 185
L p .X; ; /, 181
L 1 .X; ; /, 187
T,8
ess-ran f , 187
GL.A/, 311
ess-sup f , 187
-algebra, 77
-finite measure space, 88
c0 .N/, 188
A
abelian, 394
absolute continuity of measures, 137
absolutely continuous measure, 137
absolutely convergent Fourier series, 323
adjoint operator, 276, 331
affine function, 261
algebra, 176, 289
abelian, 177, 289
associative, 176, 289
Banach, 177, 289, 309
simple, 433
uniform, 177
unital, 177, 289
algebraic ideal, 290, 314
almost everywhere convergence, 131
almost everywhere equal functions,
130
approximate eigenvalue, 298
approximate eigenvector, 298
arithmetic-geometric mean inequality,
148
atom, 106
atomic measure, 106
B
Banach algebra, 177, 289, 309
abelian, 309
maximal abelian, 313
semisimple, 321

DOI 10.1007/978-3-319-45633-1
445
446
Banach space, 170
reflexive, 227
separable, 191
Banach subalgebra, 312
basic open set, 10
basis
for a topology, 10, 57
linear, 171
of subsets, 10
orthonormal, 203
bijective correspondence, 5
bilateral shift operator, 336
Borel set, 78
boundary, 24
bounded linear map, 215
box topology, 19
C
C -algebra, 393, 410
unital, 410
C -subalgebra, 410
unital, 410
Calkin algebra, 425
Cantor set, 27, 96
Cantor ternary function, 29, 104
Carathodorys criterion, 89
Carathedorys Convexity Theorem, 250
Cauchy net, 208
Cauchy sequence, 8
Cauchy-Schwarz inequality, 15, 196
character, 316
character space, 317
characteristic function, 82, 116
Closed Graph Theorem, 280, 282
closed set, 22
closure, 22
co-countable topology, 9
co-finite topology, 9
cofinal, 49
commutant, 391, 397
double, 397
compact operator, 284, 395
compact set, 39
compactification, 66
Stone-Cech,
67, 69
complemented subspace, 282, 283, 375
complete metric space, 59
complex measure, 113, 189
concave function, 145, 261
cone, 267
conjugate real numbers, 149
connected space, 63
Index
continuous function, 25
continuous function vanishing at infinity, 181
convergent net, 48
convergent sequence, 44
convex combination, 249
convex function, 145, 261
convex hull, 250
convex set, 198, 249
face, 257
countable set, 6
Courant-Fischer Theorem, 356
cyclic vector, 405
D
dense set, 23, 57
dilation, 370, 373
Dirac measure, 84
direct sum, 201, 209
orthogonal, 201
directed set, 48
disc algebra, 312, 328
discrete topology, 9
distance between sets, 93
distance from a point to a set, 198
division ring, 310
Dominated Convergence Theorem, 129, 134,
159
double commutant, 397
Double Commutant Theorem, 398
dual space, 221
second, 226
E
eigenvalue, 298
approximate, 298
eigenvector, 298
approximate, 298
equivalent norms, 168
equivalent representations, 430
essential ideal, 328
essential range of a measurable function, 187,
300
essential supremum, 187
essentially bounded function, 187
Euclidean metric, 14
Euclidean norm, 166
exponential function, 325
exposed point, 268
extended real number system, 83
extreme point, 255
extremely disconnected, 74
Index
F
face of a convex set, 257
factor, 434, 441
factorial trace, 434
faithful trace, 433
Fatous Lemma, 123
Fermion algebra, 436
Fermion C -algebra, 436
final space, 359
finite intersection property, 40
finite measure, 84
finite signed measure, 111
finite-rank operator, 273
Fourier coeffcients
summable, 323
Fourier series, 202, 205
Fredholm alternative, 302
function
p-integrable, 149
affine, 261
Cantor ternary, 29, 104
characteristic, 82, 116
concave, 145, 261
continuous, 25
convex, 145, 261
essentially bounded, 187
integrable, 127, 134, 135
Jordan decomposition of, 126
logarithm, 143
measurable, 78, 133
monotone increasing, 29, 104
Riemann integrable, 153
simple, 82
uniformly continuous, 45
vanishing at infinity, 181
Fundamental Theorem of Algebra, 292
Fundamental Theorem of Calculus, 140
G
Gelfand Representation Theorem, 319
Gelfand transform, 320
Gelfand-Mazur Theorem, 310, 317
general linear group, 311
GNS triple, 427, 441
Gram-Schmidt Process, 202
group algebras, 394, 397, 405, 433
group von Neumann algebra, 395
H
Hlders inequality, 149
Hahn decomposition of a signed measure
space, 112
Hahn Decomposition Theorem, 111
447
Hahn-Banach Extension Theorem, 223, 226,
246
Hahn-Banach Separation Theorem, 252
Hardy space, 337
Hausdorff space, 41
hermitian
element of a C -algebra, 411
hermitian operator, 339
Hilbert space, 197
homeomorphism, 33
homomorphism, 310
unital, 310
hyperinvariant subspace, 303
I
ideal, 290, 314
algebraic, 290, 314
essential, 328
maximal, 315
maximal ideal space, 317
proper, 290, 314
proper algebraic, 290
radical, 321
idempotent operator, 281
improper Riemann integral, 156
inequality
arithmetic-geometric mean, 148
Cauchy-Schwarz, 196
Hlder, 149
Jensen, 147
Minkowski, 150, 182
Young, 148
initial space, 359
inner product, 195
inner product space, 195
integrable function, 127, 135
p-integrable, 149
integral, 120, 121, 127, 134
integral operator, 273, 274
interior, 23
Intermediate Value Theorem, 64
invariant subspace, 303
inverse, 310
invertible operator, 292
involution, 410
on B.H/, 331
involutive algebra, 410
irreducible operator algebra, 402
isolated point, 343
isometrically isomorphism, 219
isometry, 73, 219, 333
isomorphism, 310
isometric, 219
448
J
Jacobson topology, 441
Jensens inequality, 147
Jordan decomposition of a function, 126
Jordan Decomposition Theorem, 135
Jordan Decomposition Theorem of a signed
measure, 113
K
Kaplanskys Density Theorem, 400
kernel, 219
kernel of an operator, 219
Kren-Milman Theorem, 257, 258
L
Lp -space, 182
Lebesgue integral, 120, 121, 127, 134
complex-valued functions, 134
nonnegative functions, 121
nonnegative simple functions, 120
real-valued functions, 127
Lebesgue outer measure, 95
left regular representation, 394
Legendre polynomials, 206
limit point, 24, 343
linear basis, 171
linear functional, 220
positive, 232, 237, 239, 420
linear manifold, 173
linear order, 6
linear submanifold, 173
locally compact space, 46, 175
Loewner ordering, 353
logarithm, 143
Lomonosovs Theorem, 304
lower-limit topology, 35
Lusins Theorem, 108, 109
Lyapunovs Theorem, 265
M
matrices of operators, 395
matrix completion, 370
maximal abelian subalgebra, 313
maximal ideal, 315
maximal ideal space, 317
measurable function, 78, 133
measurable partition, 113
measurable set, 78
measurable space, 78
measure, 84
absolute continuity, 137
Index
absolutely continuous, 137
atomic, 106
complex, 113, 189
counting, 85
Dirac, 84
finite, 84
non-atomic, 106, 265
point mass, 84
probability, 84
regular, 107
signed, 110
signed, finite, 111
singular, 106
support of, 259
total variation, 114, 159
measure space, 84
-finite, 88
signed, 111
metric, 13
Euclidean, 14
pseudo-, 211
uniform, 19
metric space, 13, 16
complete, 59
metric topology, 16, 166
metrisable space, 17, 21
Min-Max Variational Principle, 356
Minkowski functional, 251
Minkowskis inequality, 150, 182
modulus, 420
modulus operator, 358
Monotone Convergence Theorem, 121
monotone increasing function, 29, 104
monotone increasing sequence, 82
multiplication operator, 273, 274
N
neighbourhood, 24
net, 48, 208
Cauchy, 208
convergent, 48
non-atomic measure, 106, 265
nondegenerate operator algebra, 404
nonmeasurable set, 102
norm, 165
equivalent, 168
Euclidean, 166
submultiplicative, 290
trace, 383
norm topology, 166
normal
normal operator, 346
Index
normal space, 51
nowhere dense set, 62
null set, 93
O
one-point compactification, 47
open cover, 39
open map, 278
Open Mapping Theorem, 278
open set, 9
basic, 10
weakly, 228
operator, 216
adjoint, 276, 331
bilateral shift, 336
bounded, 215
bounded below, 277
compact, 284
finite-rank, 273
hermitian, 339
idempotent, 281
integral, 274
invertible, 292
isometric, 219, 333
kernel, 219
lower bound, 277
modulus, 358
normal, 346
partial isometry, 359
positive, 352
projection, 333
range, 219
reductive, 349
trace-class, 377
unbounded, 216
unilateral shift, 275, 337
unitary, 333
Volterra, 274
weighted unilateral shift, 275, 338
orthogonal complement, 200
orthogonal vectors, 197
orthonormal basis, 203
outer measure, 88
Lebesgue, 95
P
parallelogram law, 198
Parsevals Equation, 205
partial isometry, 359
partial order, 6
partial sum of vectors, 208
partition of unity, 56, 241
449
path connected, 65
Peano axioms, 5
point mass measures, 84
polar decomposition, 359, 360
polarisation identity, 333
polynomial
Legendre, 206
trigonometric, 194
positive
positive linear functional, 232, 237, 239, 420
positive operator, 352
positive square root, 352, 354
Principle of Uniform Boundedness, 277
product topology, 19
projection map, 31
projection operator, 333
proper algebraic ideal, 290
proper ideal, 290, 314
pseudo-metric, 211
Pythagorean Theorem, 198
Q
quasinilpotent, 439
quotient
of a topological space, 42, 70
quotient norm, 173
quotient space, 42, 70, 173, 315, 423
quotient topology, 32
R
radical, 321
Radon-Nikodm Theorem, 137, 160
range, 219
range of an operator, 219
reduced group C -algebra, 394
reductive operator, 349
reflexive Banach space, 227
regular measure, 107
representation, 425
Riemann integrable function, 153
Riemann integral, 153
improper, 156
Riesz Representation Theorem, 221, 233, 240,
264, 329
right regular representation, 395
S
Schroeder-Bernstein Theorem, 5, 29
second countable, 57
second dual, 226
450
selfadjoint subset, 177
seminorm, 169
semisimple Banach algebra, 321, 439
separable Banach space, 191
separable topological space, 57
sequence, 7
p-summable, 151
Cauchy, 8, 59
convergent, 44, 59
sequential cover, 88
sesquilinear form, 195, 331
set
-null, 93
cofinal, 49
countable, 6
directed, 48
linearly ordered, 6
nonmeasurable, 102
null, 96
partially ordered, 6
sigma-algebra, 77
signed measure, 110
signed measure space, 111
simple function, 82
canonical form, 119
singular measure, 106
singular value decomposition, 373
singular values, 374
Spectral Mapping Theorem
Polynomial, 297
spectral permanence, 312
spectral radius, 295, 310
Spectral Theorem, 344, 345, 348
spectrum, 416
of a C -algebra, 430
sphere
n-sphere Sn , 22
standard topology of R, 11
state, 420
state space, 420
Stone-Cech
compactification, 67, 69
Stone-Weierstrass Theorem, 179, 194
strictly monotone increasing function, 29
strong operator topology, 364
strongly exposed point, 268
subalgebra
Banach, 312
subbasis, 13
subcover, 39
sublinear functional, 224
subnet, 49
subspace, 173
complementary pairs, 282
complemented, 283
Index
subspace topology, 17
sum of sets, 171
summable Fourier coefficients, 323
support, 55, 180, 259
Suslins Theorem, 104
Sylvestres equation, 327
Sz.-Nagy Dilation Theorem, 373
T
Tietze Extension Theorem, 52
topological boundary, 24
topological space, 9
T2 , 40
connected, 63
extremely disconnected, 74
Hausdorff, 41
locally compact, 46, 175
normal, 51
second countable, 57
separable, 57
totally disconnected, 65
topological spaces
homeomorphic, 33
topology, 8
box, 19
co-countable, 9
co-finite, 9
coarser than, 9
discrete, 9
finer than, 9
indiscrete, 9
Jacobson, 441
lower-limit, 35
metric, 16, 166
metrisable, 17, 21
norm, 166
product, 19
quotient, 32
standard, of R, 11
subspace, 17
ultraweak, 385
uniform metric, 19
weak, 31
weak , 229
total variation of a complex measure, 114,
159
totally disconnected, 65
trace, 378, 433
factorial, 434
faithful, 433
trace norm, 383
trace-class operator, 377
tracial state space, 434
Index
451
tracial weight, 378

triangle inequality, 13, 135, 159,
166
trigonometric polynomial, 194
V
Volterra integral operator, 274
von Neumann algebra, 393
von Neumanns inequality, 373
U
ultraweak topology, 385
uniform, 327
uniform algebra, 177
uniform continuity, 45
uniform metric, 19
uniform metric topology, 19
unilateral shift operator, 337
unit operator interval, 357
unitary
unitary operator, 333
unitisation, 416
upper semicontinuity of spectra,
295
Urysohns Lemma, 55
W
weak operator topology, 364
weak topology, 31, 228
weak topology, 229
Weierstrass Approximation Theorem, 179,
180, 212
weighted unilateral shift operator, 275
Wieners Theorem, 324
Y
Youngs inequality, 148
Z
Zorns Lemma, 6

Fundamentals of Functional Analysis

Uploaded by

Document Informationclick to expand document informationFundamentals of Functional Analysis

Document Informationclick to expand document information

Copyright:

Available Formats

Fundamentals of Functional Analysis

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fundamentals of Functional Analysis

Uploaded by

Copyright:

Available Formats

Universitext

Universitext is a series of textbooks that presents material from a wide variety of

More information about this series at http://www.springer.com/series/223

Functional analysis tightly interweaves concepts from analysis, topology, geometry,

countability, and the Stone-Cech

I have discussed many aspects of functional analysis with numerous colleagues,

Topological Spaces with Special Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Part II Measure Theory and Integration

Part III Banach Spaces

Banach Space Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Spectral Theory in Banach Algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Hilbert Space Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

10.3 Hermitian Operators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Algebras of Hilbert Space Operators. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1.1 Sets and Partial Orderings

Springer International Publishing Switzerland 2016

With respect to Cartesian products, union and intersection satisfy

1.1 Sets and Partial Orderings

With respect to complements, DeMogans Laws hold:

1.2 Completeness of the Real Number System

1.2 Completeness of the Real Number System

The following theorem establishes the completeness of R as a metric space.

1.3 Topological Spaces

for every family fU g2 of sets U 2 T ;

1.3 Topological Spaces

for every finite collections fUk gnkD1 of sets Uk 2 T .

F is a subset of each of the sets F and is, therefore, countable.

This shows that Tcocntbl is closed under arbitrary unions. Finally, if U1 ; : : : ; Un 2

Definition 1.19. A collection B of subsets of a set X is called a basis of subsets of

any x 2 U; thus, x 2 U for some 2 . Because U 2 T , there is a B 2 B such that

In light of Propositions 1.20 and 1.22, the following definition is made.

1.3 Topological Spaces

Proposition 1.24. If T is a topology on X, and if B is a basis for the topology T ,

Proof. Exercise 1.104.

Proof. Exercise 1.105.

1.4 Metric Topologies

The concept of subbasis is important in topology because of the next result.

1.4 Metric Topologies

The pair .X; d/ is called a metric space.

The Euclidean metric on R2 is derived from the Pythagorean theorem, and it

and so the discriminant .2

1.4 Metric Topologies

D .d2 .x; y/ C d2 .y; z//2 :

Corollary 1.36. The formula d.x; y/ D jx  yj, for x; y 2 R, defines a metric on R.

for real numbers u1 ; : : : ; un ; v1 ; : : : ; vn is called the Cauchy-Schwarz inequality.

d1 .x; y/ D max jxj  yj j:

Proof. Exercise 1.111.

The reason for labelling these metrics as d1 , d2 , and d1 will be apparent in

Proposition 1.39. If d is a metric on a set X and if

where Br .x/ D fy 2 X j d.x; y/ < rg, then B is a basis.

which implies that d2 .x; y/ d1 .x; y/ for all x; y 2 Rn .

1.5 Subspaces and Product Spaces

On the other hand, the Cauchy-Schwarz inequality (1.2) yields

is a metric X. Furthermore, the metric topologies Td and Tdb on X coincide.

Proof. Exercise 1.113.

The standard topology on R was first introduced here by specifying a certain

1.5 Subspaces and Product Spaces

It is straightforward to verify that TY satisfies the axioms for a topology. Hence,

Corollary 1.36. The formula d.x; y/ D jx yj, for x; y 2 R, defines a metric on R.

d1 .x; y/ D max jxj yj j:

< r D min fjxj pxj j; jxj qxj jg

Proposition 1.61. If Y is a subspace of a topological space X, then F Y is closed

5. if Y Z, then Z n Y Z n Y, and equality need not be attained.

(3) ) (1). If V Y is open, then f 1 .V c / is closed, by hypothesis. Because

respectively, where gj 2 f0; 2g for all 1 j .n 1/. Thus,

implies j.y0 / .x/j < ", proving the continuity of at y0 .

1.115. Prove that if B is a basis for a topology T on X and if Y X, then the

1.117. If Y X is a subset of a topological space X, then prove that Y D int .Y c /

2. If fU g2 is an open cover of Y and if , then the family fU! g!2 is a

x 2 .a; b j there is a finite subset F such that a; x