Secchi S A Circleline Study of Mathematical Analysis

UNITEXT 141
Simone Secchi
A Circle-Line Study
of Mathematical
Analysis
UNITEXT
La Matematica per il 3+2
Volume 141
Editor-in-Chief
Alfio Quarteroni, Politecnico di Milano, Milan, Italy; École Polytechnique Fédérale
de Lausanne (EPFL), Lausanne, Switzerland
Series Editors
Luigi Ambrosio, Scuola Normale Superiore, Pisa, Italy
Paolo Biscari, Politecnico di Milano, Milan, Italy
Ciro Ciliberto, Università di Roma “Tor Vergata”, Rome, Italy
Camillo De Lellis, Institute for Advanced Study, Princeton, New Jersey, USA
Massimiliano Gubinelli, Hausdorff Center for Mathematics, Rheinische
Friedrich-Wilhelms-Universität, Bonn, Germany
Victor Panaretos, Institute of Mathematics, École Polytechnique Fédérale de
Lausanne (EPFL), Lausanne, Switzerland
Lorenzo Rosasco, DIBRIS, Università degli Studi di Genova, Genova, Italy
Center for Brains Mind and Machines, Massachusetts Institute of Technology,
Cambridge, Massachusetts, USA; Istituto Italiano di Tecnologia, Genova, Italy
The UNITEXT - La Matematica per il 3+2 series is designed for undergraduate
and graduate academic courses, and also includes advanced textbooks at a research
level.
Originally released in Italian, the series now publishes textbooks in English
addressed to students in mathematics worldwide.
Some of the most successful books in the series have evolved through several
editions, adapting to the evolution of teaching curricula.
Submissions must include at least 3 sample chapters, a table of contents, and
a preface outlining the aims and scope of the book, how the book fits in with the
current literature, and which courses the book is suitable for.
For any further information, please contact the Editor at Springer:
francesca.bonadei@springer.com
THE SERIES IS INDEXED IN SCOPUS
***
UNITEXT is glad to announce a new series of free webinars and interviews
handled by the Board members, who rotate in order to interview top experts in their
field.
Click here to subscribe to the events!
https://cassyni.com/events/TPQ2UgkCbJvvz5QbkcWXo3
Simone Secchi
A Circle-Line Study of
Mathematical Analysis
Simone Secchi
Department of Mathematics and
Applications
University of Milano Bicocca
Milan, Italy
ISSN 2038-5714 ISSN 2532-3318 (electronic)

UNITEXT
ISSN 2038-5722 ISSN 2038-5757 (electronic)
La Matematica per il 3+2
ISBN 978-3-031-19737-6 ISBN 978-3-031-19738-3 (eBook)
https://doi.org/10.1007/978-3-031-19738-3
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland
AG 2022
This work is subject to copyright. All rights are solely and exclusively licensed 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, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
Cover illustration: The circular structure of the book. Painted by Francesca Vettori
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
This book is dedicated to Francesca, for her
constant and doubtless support
Preface
My first encounter with Mathematical Analysis was on the celebrated Principles

of Mathematical Analysis by Walter Rudin. I immediately fell in love with this
discipline, so elegant and polished in Rudin’s pages. Seventy years ago, when that
book was written, the common approach to Mathematical Analysis was deeply
influenced by the topological revolution introduced by the Bourbaki group. Metric
spaces were the basement of the building, and a general effort towards abstraction
and generality was evident.
Nowadays many things have changed. I have been teaching Calculus and Mathe-
matical Analysis for more than 15 years, and I have read many books on this subject.
A current trend is to isolate Calculus from Analysis, and a rather disappointing
corollary of this approach is that the topological revolution has been forgotten. The
field of real numbers is so rich that many students do not understand that limits and
continuity are not necessarily related to algebraic structures. Sequences are strange
objects that appear and disappear without any good reasons, since they are not used
to describe the topology of the real line.
In this book I have tried to place the puzzle pieces where they have to stay,
starting from an exposition of aware Basic Analysis in .R and reaching the rigorous
development of more advanced branches of Analysis. In my own experience,
students learn limits, continuity and the main properties of the real line, but they
hardly realize that a fair amount of analysis ideas are just general topological
facts applied to the topological space .R. So to speak, our students would need a
return ticket to modern mathematical analysis, but they just get a one-way ticket
to Cauchy’s calculus. Figure 1 gives an idea of our view of modern analysis (and
mathematics as well!).
Our journey starts with some naïve set theory and an axiomatic introduction of
the real numbers. Natural numbers are introduced as a subset of .R, and mathematical
induction is fully justified. I believe that there is no concrete alternative to this
approach, since sets and numbers are those mathematical objects which can be
used before learning their coherent definitions. Let me be clear: when you are an
experienced working mathematician, you may very well believe that real numbers
should be constructed from the rational numbers, and natural numbers should be
vii
viii Preface
Fig. 1 The tree of this book
defined in terms of sets. But this viewpoint would prevent any student
√ from using
numbers on a daily basis. Students are not supposed to think of . 2 as a Dedekind
cut or as an equivalence class of sequences of rational numbers, when they first meet
limits, derivatives and integrals. And I would bet that no experienced analyst ever
learned analysis (or fell in love with analysis) this way. It is a matter of fact that
many students will never feel the need of studying Dedekind cuts, and they will be
happy to know that .R can be constructed in one way or another.
A first taste of topology on the real line then comes into play, and this is enough
to have a rigorous definition of all the main tools of analysis in one dimension:
limits, sequences and series, derivatives and Riemann integrals. Our first definition
of the derivative is formally the same definition that one would give in normed vector
spaces. We prove that the derivative is the usual limit of the incremental ratio, but we
stress that a Calculus definition must be changed as early as we want to differentiate
functions of two variables. Linearization is the keyword of differential calculus, and
we use this in full strength.
Metric spaces are not overestimated in this book. It is a common opinion
that metric spaces are the natural environment for doing analysis. This is surely
supported by some recent theories, but we prefer to present them for what they
Preface ix
are: a special case of topological spaces. Most textbooks avoid the generality of
topological spaces, and students often feel that mathematical analysis can exist only
if distances can be measured. This is largely false, as we will see. Furthermore, I
totally dislike the habit of keeping analysis and topology disjoint: some branches of
Functional Analysis do need abstract topological structures.
The chapter on differentiation may look short. Since this is not a Calculus
textbook, we do not spend time to compute dozens of elementary derivatives. The
purpose of the chapter is to present the theory of differentiation in one real variable,
so I omit any discussion about the physical meaning of derivatives. I assume that
the average reader of my book knows that “velocity is the derivative of motion, and
acceleration is the derivative of velocity.” Although Mathematics is continuously
stimulated by other sciences, it is indeed a science itself. My students are not
supposed to define derivatives as slopes of a line and integrals as energies.
The chapter on Riemann integration is probably longer than usual: I present a few
results that most textbooks only propose in the special case of continuous functions.
Indeed, I think that the class of Riemann integrable functions should be kept distinct
from the subclass of continuous functions as long as possible. The most important
feature of modern mathematics is the capability to arrange objects according to their
properties: should we deal with differentiable functions only, since they form an
interesting subclass of continuous functions? I do not think so.
A chapter on the so-called elementary functions is necessary in any rigorous
exposition of mathematical analysis. On one hand, it is true that differentiating the
cosine function before providing a definition of the cosine function is nonsense.
On the other hand, however, it would be unrealistic to remove any example which
involves elementary functions from a textbook. In my humble opinion, this is a
problem without solution: we must teach differential and integral calculus before
teaching the construction of the elementary functions. But a formal approach to
such functions is not enough for a good analyst, who needs to learn why elementary
functions actually exist.
Our journey can now proceed backwards, so that we look again at the basic ideas
from a more advanced viewpoint. In particular, I propose a quick-and-dirty review
of axiomatic set theory. Set theory is, somehow, what every mathematician pretends
to know. Students are not exposed to the difficulties of the axiomatization of this
discipline, and they usually ignore what entities are really primitive. A function is
more often than not proposed as a black box that converts a number into another
number. Once sets have been introduced or assumed to exist, there is no need to
suppose that functions are a primitive notion. This may be convenient, but it is
definitely unnecessary.
Axiomatic Set Theory looks close to philosophy, since different and non-
equivalent approaches have been proposed over the decades. I present a review of
John Kelley’s set theory, which I find particularly suited to the analyst’s mindshape.
For the sake of completeness, I also list the axioms of Zermelo and Fraenkel, which
are probably dominant among experts.
I then introduce some general topology: open and closed sets, neighborhoods,
and of course limits. I am proud to define nets, a generalization of sequences
x Preface
that completely characterize the topology of any space. Convergence is usually

the most appreciated idea of mathematical analysis, and we see that the whole
Calculus can be described in terms of converging nets. Compact sets and connected
sets are defined, so that the reader can appreciate several classical theorems about
continuous functions of a single variable as particular cases of abstract properties.
Compactness is ubiquitous in analysis, and I hope that our survey may help readers
understand its use. The reader will notice that the chapter on topology is not written
in a systematical way, and ideas appear when they are needed. The realm of analysis
is not so polished as the realm of geometry, because analysis is more concerned with
borderline cases.
I have wondered for a long time whether a chapter on abstract differentiation
in normed vector spaces was a good idea. It needs some Functional Analysis to
be understood, and this is not a book on Functional Analysis. Anyway I came up
with the conclusion that linearization is seldom taught in a unitary way: students
differentiate functions of one variable with one definition, then functions of two
variables with another definition, the functions on manifolds with again a different
definition. In this respect I wanted to show that differentiation is the same in .R,
.R , .R , and in any vector space that allows distances to be measured. A Global
2 3
Inversion Theorem is proved in Hausdorff spaces, and this is a nice result that
is seldom proposed in Analysis books. My gratitude goes to Antonio Ambrosetti
and Giovanni Prodi, who popularized abstract differential calculus in the Italian
mathematical community.
The two chapters on integration and measure theory are a journey inside the
journey. I present a flavor of Integration Theory based on the Daniell approach:
the integral becomes a suitable extension of a linear functional with some weak
continuity condition. This extension leans on an elementary integral which we may
imagine as the Cauchy integral of continuous functions with compact support. I
believe that this functional-analytic construction may be of great utility to young
mathematicians who do not need all the pathologies of Geometric Measure Theory.
The subsequent chapter returns to the integral via a different path, based on
suitable families of sets called .σ -algebras. I would like the reader to understand that
there is a path which connects measurable sets and integrable functions, and that we
can decide in what direction we prefer to go along this path. But even in this second
chapter, I completely avoided the Carathéodory machinery of outer measures, and
the concrete Lebesgue measure is constructed via a Riesz Representation Theorem
à la Rudin.
This book contains a few figures, often realized in a sketchy way. The use of
personal computers would surely allow us to produce perfect figures, but I wanted
to draw the same pictures I would draw on the blackboard. Most drawings were
made by Dr. Francesca Vettori, to whom I am indebted.
Who will read this book? I hope that it can be useful to students of Mathematics
and Physics who wish to go further than standard Calculus. If compared to other
similar textbooks, the main difference here is that we offer a second glance—and
indeed more than a glance—on every traditional topic. Of course, I also hope that
some colleagues may find the book useful for preparing their lectures.
Preface xi
Instructors will surely need to complement this text with some additional
examples. In my opinion, the same example can be enlightening for a student but
obscure for another student. A good instructor knows his/her audience, and can
suitably illustrate abstract ideas with concrete examples. Exercises appear after
some important results, so that the reader is invited to solve them before proceeding
further. Several chapters end with a short collection of problems, which may be
solved by collecting the ideas of the whole chapter to which they refer.
I would like to express my gratitude Dr. Francesca Bonadei and Dr. Francesca
Ferrari of Springer Nature for their support during the preparation of the manuscript.
I am obviously responsible for any misprint appearing in this book. If you are
reading it and you have just found a (serious) error in the text, feel free to contact
me at simone.secchi@unimib.it.
Milano, Italy Simone Secchi

July 2022
Acknowledgements
Every scientific book is the result of several months of hard work, but also of several
years of study. First of all I need to acknowledge the support of my family during
the last year. I also want to express my gratitude to my university for providing me
with all the necessary tools that were used during the preparation of the manuscript.
xiii
Contents
Part I First Half of the Journey

1 An Appetizer of Propositional Logic . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
1.1 The Propositional Calculus . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
1.2 Quantifiers .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 5
2 Sets, Relations, Functions in a Naïve Way . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7
2.1 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 15
Reference .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 16
3 Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 17
3.1 The Axioms of R. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 17
3.2 Order Properties of R . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 20
3.3 Natural Numbers .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 22
3.4 Isomorphic Copies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 27
3.5 Complex Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 27
3.6 Polar Representation of Complex Numbers . . .. . . . . . . . . . . . . . . . . . . . 30
3.7 A Construction of the Real Numbers . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 32
3.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 35
3.9 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 35
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 36
4 Elementary Cardinality .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 37
4.1 Countable and Uncountable Sets . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 37
4.2 The Schröder-Bernstein Theorem .. . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 41
4.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 43
4.4 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 44
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 45
5 Distance, Topology and Sequences on the Set of Real Numbers. . . . . . . 47
5.1 Sequences and Limits .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 50
5.2 A Few Fundamental Limits . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 58
5.3 Lower and Upper Limits. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 59
xv
xvi Contents
5.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 62
5.5 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 65
Reference .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 65
6 Series . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 67
6.1 Convergence Tests for Positive Series . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 71
6.2 Euler’s Number as the Sum of a Series . . . . . . . .. . . . . . . . . . . . . . . . . . . . 74
6.3 Alternating Series .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 76
6.3.1 Product of Series . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 77
6.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 79
6.5 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 80
Reference .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 80
7 Limits: From Sequences to Functions of a Real Variable . . . . . . . . . . . . . 81
7.1 Properties of Limits . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 86
7.2 Local Equivalence of Functions .. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 87
7.3 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 89
8 Continuous Functions of a Real Variable . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 91
8.1 Continuity and Compactness .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 95
8.2 Intermediate Value Property .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 98
8.3 Continuous Invertible Functions . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 99
8.4 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 101
9 Derivatives and Differentiability .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 103
9.1 Rules of Differentiation, or the Algebra of Calculus .. . . . . . . . . . . . . 105
9.2 Mean Value Theorems .. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 109
9.3 The Intermediate Property for Derivatives.. . . .. . . . . . . . . . . . . . . . . . . . 113
9.4 Derivatives at End-Points .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 114
9.5 Derivatives of Derivatives . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 114
9.6 Convexity .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 117
9.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 119
9.8 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 122
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 122
10 Riemann’s Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 123
10.1 Partitions and the Riemann Integral .. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 123
10.2 Integrable Functions as Elements of a Vector Space . . . . . . . . . . . . . . 128
10.3 Classes of Integrable Functions . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 130
10.4 Antiderivatives and the Fundamental Theorem.. . . . . . . . . . . . . . . . . . . 134
10.5 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 137
10.6 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 138
11 Elementary Functions .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 139
11.1 Sequences and Series of Functions.. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 139
11.2 Uniform Convergence . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 141
11.3 The Exponential Function . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 145
11.4 Sine and Cosine .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 148
Contents xvii
11.5 Polynomial Approximation.. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 151

11.6 A Continuous Non-differentiable Function .. . .. . . . . . . . . . . . . . . . . . . . 153
11.7 Asymptotic Estimates for the Factorial Function.. . . . . . . . . . . . . . . . . 155
11.8 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 158
Part II Second Half of the Journey

12 Return to Set Theory .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 163
12.1 Kelley’s System of Axioms.. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 164
12.2 From Sets to N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 178
12.3 A Summary of Kelley’s Axioms . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 180
12.4 Set Theory According to J.D. Monk . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 181
12.5 ZF Axioms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 186
12.6 From N to Z . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 188
12.7 From Z to Q .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 189
12.8 From Q to R. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 191
12.9 About the Uniqueness of R . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 195
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 199
13 Neighbors Again: Topological Spaces . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 201
13.1 Topological Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 201
13.2 The Special Case of RN . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 206
13.3 Bases and Subbases .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 208
13.4 Subspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 209
13.5 Connected Spaces .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 210
13.6 Nets and Convergence . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 214
13.7 Continuous Maps and Homeomorphisms . . . . .. . . . . . . . . . . . . . . . . . . . 223
13.8 Product Spaces, Quotient Spaces, and Inadequacy
of Sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 226
13.9 Initial and Final Topologies . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 232
13.10 Compact Spaces. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 234
13.10.1 The Fundamental Theorem of Algebra . . . . . . . . . . . . . . . . . 245
13.10.2 Local Compactness . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 247
13.11 Compactification of a Space .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 248
13.12 Filters and Convergence . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 250
13.13 Epilogue: The Limit of a Function .. . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 254
13.14 Separation and Existence of Continuous Extensions .. . . . . . . . . . . . . 256
13.15 Partitions of Unity and Paracompact Spaces . .. . . . . . . . . . . . . . . . . . . . 262
13.16 Function Spaces .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 272
13.17 Cubes and Metrizability . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 280
13.18 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 290
13.19 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 293
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 294
xviii Contents
14 Differentiating Again: Linearization in Normed Spaces . . . . . . . . . . . . . . . 295

14.1 Normed Vector Spaces.. . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 295
14.2 Bounded Linear Operators . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 298
14.3 The Hahn-Banach Theorem . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 299
14.4 Baire’s Theorem and Uniform Boundedness . .. . . . . . . . . . . . . . . . . . . . 306
14.5 The Open Mapping Theorem.. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 309
14.6 Weak and Weak* Topologies .. . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 311
14.7 Isomorphisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 315
14.8 Continuous Multilinear Applications . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 316
14.9 Inner Product Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 320
14.10 Linearization in Normed Vector Spaces . . . . . . .. . . . . . . . . . . . . . . . . . . . 327
14.11 Derivatives of Higher Order .. . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 336
14.12 Partial Derivatives.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 340
14.13 The Taylor Formula .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 341
14.14 The Inverse and the Implicit Function Theorems . . . . . . . . . . . . . . . . . 342
14.14.1 Local Inversion.. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 343
14.15 A Global Inverse Function Theorem .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 346
14.16 Critical and Almost Critical Points . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 353
14.17 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 356
14.18 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 358
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 359
15 A Functional Approach to Lebesgue Integration Theory .. . . . . . . . . . . . . 361
15.1 The Riemann Integral in Higher Dimension .. .. . . . . . . . . . . . . . . . . . . . 361
15.2 Elementary Integrals .. . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 363
15.3 Null and Full Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 363
15.4 The Class L+ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 365
15.5 The Class L of Integrable Functions .. . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 369
15.6 Taking Limits Under the Integral Sign .. . . . . . . .. . . . . . . . . . . . . . . . . . . . 371
15.7 Measurable Functions and Measurable Sets . . .. . . . . . . . . . . . . . . . . . . . 374
15.8 Integration Over Measurable Sets . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 378
15.9 The Concrete Lebesgue Integral . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 382
15.10 Integration on Product Spaces . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 386
15.11 Spaces of Integrable Functions .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 390
15.12 The Space L∞ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 397
15.13 Changing Variables in Multiple Integrals .. . . . .. . . . . . . . . . . . . . . . . . . . 398
15.14 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 400
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 400
16 Measures Before Integrals .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 401
16.1 General Measure Theory . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 401
16.2 Convergence Theorems .. . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 409
16.3 Complete Measures .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 412
16.4 Different Types of Convergence.. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 415
16.5 Measure Theory on Product Spaces.. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 419
Contents xix
16.6 Measure, Topology, and the Concrete Lebesgue Measure . . . . . . . . 423

16.6.1 The Concrete Lebesgue Measure . . .. . . . . . . . . . . . . . . . . . . . 434
16.7 Mollifiers and Regularization . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 439
16.8 Compactness in Lebesgue Spaces . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 445
16.9 The Radon-Nykodim Theorem .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 446
16.10 A Strong Form of the Fundamental Theorem of Calculus .. . . . . . . 451
16.11 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 468
16.12 Comments .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 469
References .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 469
Part I
First Half of the Journey
Chapter 1
An Appetizer of Propositional Logic
Abstract Mathematics is based on the language of proposition logic: every state-

ment is a combination of logical propositions, and theorems are simply true
statements that can be deduced according to the rules of logic. We follow the first
chapter of Mendelson (Introduction to mathematical logic. CRC Press, Boca Raton,
2015).
1.1 The Propositional Calculus
Sentences are just statements to which it is possible to attach a binary value: true
(T) or false (F). For example, “Roses are flowers” is a sentence, “dogs have five
legs” is another sentence. But “Any cat is” is not. Sometimes sentences depend on
free variables, as in “The integer n is a prime”, or “The real number x is irrational”.
The variable n and x are free in the sense that they may take any (admissible) value:
compare with “For every integer n, .n + 1 > n”. In this sentence, the variable n
is quantified by “For each”, and is not a free variable. Another example is “There
exists a positive real number r such that .r 2 = 2”.
Definition 1.1 (Negation) If A is a sentence, its negation .¬A is the sentence
governed by the following table:
A ¬A
. T F
F T
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 3

S. Secchi, A Circle-Line Study of Mathematical Analysis,
La Matematica per il 3+2 141, https://doi.org/10.1007/978-3-031-19738-3_1
4 1 An Appetizer of Propositional Logic
Definition 1.2 (Conjunction) If A and B are sentences, then their conjunction .A ∧

B is the sentence governed by the following table:
A B A∧B
T T T
.T F F
F T F
F F F
Definition 1.3 (Disjunction) If A and B are sentences, then their disjunction .A∨B
is the sentence governed by the following table:
A B A∨B
T T T
.T F T
F T T
F F F
Remark 1.1 Although the mathematical conjunction agrees with the use of “and” in
everyday language, the mathematical disjunction reflects a use of “or” which may
differ from the use in common language. To be explicit, we may formulate a golden
rule: .A ∨ B corresponds to “either A, or B, or both”. In common language we tend
to understand “either A or B, but not both.”
Definition 1.4 (Implication) If A and B are sentences, the sentence .A ⇒ B is
defined by the following table:
A B A ⇒ B
T T T
.F T T
T F F
F F T
Remark 1.2 Logical implication may be written in different ways: .A ⊃ B was

common in Logic textbooks a few years ago, but also .A → B is often found. The
symbol .A ⇒ B is pronounced “If A, then B”, and we also call it a conditional.
1.2 Quantifiers 5
Definition 1.5 (Logical Equivalence) If A and B are sentences, the sentence

A ⇐⇒ B is the sentence governed by the following table:
.
A B A ⇐⇒ B
T T T
.F T F
T F F
F F T
Logical equivalence is also denoted by .A ≡ B.
1.2 Quantifiers
As we said before, sentences may contain one or more free variables.

Example 1.1 The sentence .A(x, y) defined by “the real number x is strictly smaller
than the real number y” is sentence with two free variables x and y. From a logical
viewpoint, .A(x, y) is indistinguishable from .A(α, β) or .A(♣, ♠). Of course we
cannot replace free variables with symbols that are already taken: .A(x, y) is not
the same as .A(1, 4) or .A(cos, log). However .A(π, e) may be acceptable, provided
that we do not understand .π as the number .3.14159... and e as the Napier number
.2.718281... As a stronger example, think of .A(i): is .i
2 = −1 as in Complex
Analysis, or is i a free variable?

The truth value of a sentence depending on free variables may depend on the
choice of these variables. If .A(x, y) is defined by “the real number x is strictly
smaller than the real number y”, then .A(1, 2) is certainly true, while .A(4, 0) is
false.
Definition 1.6 (Universal Quantifier) The universal quantifier .∀ means “for all”,
or “for each”.
Definition 1.7 (Existential Quantifier) The existential quantifier .∃ means “there
exists”.
Important: .∃ vs. .∃!

In mathematics, “there exists” always means “there exists at least one”. The sentence
“there exists a solution .x ∈ R to the equation .x 2 = 1” is true, although we know
that there exist exactly two real solutions to the equation .x 2 = 1. Since existence
and uniqueness is often important, the symbol .∃! is reserved for the sentence “there
exists a unique”.
6 1 An Appetizer of Propositional Logic
The syntax of sentences with quantifiers is not completely universal. The

sentence “For each x the sentence .A(x) holds” can be written in different ways:
. ∀x A(x)
(∀x)A(x)
∀x, A(x)
(x) A(x).
The last one is clearly the most economic, and the first one is affordable. Logicians
tend to avoid brackets as far as they may, and also commas are seen as inessential
objects. It is a matter of facts that most mathematicians use brackets freely on their
blackboards, and commas are also ubiquitous.
Remark 1.3 The existential quantifier is not a primitive symbol, since the sentence
.∃x A(x) is logically equivalent to (and actually defined as) .¬(∀x ¬A(x)). As a
consequence, the negation of .∃x A(x) is precisely .∀x ¬A(x), and the negation of
.∀x A(x) is precisely .∃x ¬A(x).
If we are given a sentence .A(x1, . . . , xn ) defined by n free variables and one or

more of them is quantified by either .∀ or .∃, the quantified variables become bound
variables. Bound variables essentially disappear from the arguments of the sentence.
Example 1.2 Suppose that .A(x1 , x2 ) means “.x1 −x2 = 0”. The sentence A contains
two free variables, but the sentence .∃x2 A(x1, x2 )—which means “there exists .x2
such that .x1 − x2 = 0”—contains one free variable. The sentence .∀x1 ∃x2 A(x1 , x2 )
does not contain any free variable, and means “For every .x1 there exists (at least
one) .x2 such that .x1 − x2 = 0”.
We follow the first chapter of Mendelson (Introduction to mathematical logic.
CRC Press, Boca Raton, 2015).
Chapter 2
Sets, Relations, Functions in a Naïve Way
Abstract We start our journey with naïve set theory. In the second half of the book
we will provide a rigorous foundation of these ideas.
We begin this book in the worst possible manner: we introduce a meaningless

definition.
Definition 2.1 (Sets) A set is a collection of elements.
Important: Sets Remain Undefined

It should be clear in the reader’s mind that the previous sentence is far from being a
mathematical definition. A set is defined through the word “collection”, but we do
not provide any primitive definition of collections. In other words, we are assuming
that the concept of set is already present in our minds. More formally, we can say
that our set theory is based on two primitive objects: sets and elements.
We write .x ∈ X to mean that x is an element of the set X, and we say that x is

an element of X, or that x belongs to X. We will avoid the reversed symbol .X x,
since . is sometimes used in mathematics with a different meaning.
The typical way of constructing a set is as follows:
.X = {x | some proposition about x}.
The variable x is a dummy variable, in the sense that it can be replaced by any other
symbol without affecting the validity of the definition of the set X.
Example 2.1 To clarify the use of dummy variables, consider
. {x | x is a cat} = {C | C is a cat} .

8 2 Sets, Relations, Functions in a Naïve Way
On both sides we are introducing the set of all cats, no matter how we name the
generic cat.
By definition, .X = {x | x ∈ X}. Two sets X and Y are equal when they share the
same elements: .x ∈ X if and only if .x ∈ Y .
Definition 2.2 (Empty Set) The empty set is
∅ = {x | x = x},
.
Exercise 2.1 Prove that and .∅ contains no element at all. Hint: for every x, the
statement .x = x is false.
It should be remarked that the definition of the empty set is meaningful, in the
sense that it does not rely on some intuitive knowledge. The empty set could be
equally defined by means of any statement which is false, for instance

.∅ = x ∈ R x 2 = −1
= {n ∈ N | n is neither odd nor even}

= {f | f is a function which is both bounded and unbounded}
Example 2.2 Why don’t we define the opposite of the empty set, namely
U = {x | x = x}?
.
This object would contain anything, since anything is equal to itself by definition
of equality. It would be desirable to have such a “set”. wouldn’t it? Unfortunately
.U cannot be a set, as Russel showed in his celebrated paradox. Let us consider
.R = {x | x ∈ / x}, the set of all sets which do not belong to themselves. What can we
say about the relation .R ∈ R?
Well, if .R ∈ R, then R is a set which does not belong to itself, so that .R ∈ / R.
Viceversa, if .R ∈
/ R, then R is not a set which does not belong to itself, hence .R ∈ R.
Formally, .R ∈ R if and only if .R ∈ / R. The consequence of this logical equivalence
is that sets cannot be described unrestrictedly, and the universe .U cannot be a set
in the naïve sense. We will see in the second part of this book that Axiomatic Set
Theory can be used to speak of sets without facing Russel’s paradox. But most
mathematicians think of sets naïvely, and so will we do for the moment. The only
recommendation is to avoid any use of the universe.
Definition 2.3 (Subsets) If A and B are sets, then A is a subset of B if and only if
each element of A is an element of B: in symbols,
∀x(x ∈ A ⇒ x ∈ B).
.
In this situation we write .A ⊂ B or .B ⊃ A. A set A is a proper subset of B if

A ⊂ B and .A = B. We remark that .A = B if and only if .(A ⊂ B) ∧ (B ⊂ A).
.
2 Sets, Relations, Functions in a Naïve Way 9
Important: Proper Inclusion

It must be observed that .A ⊂ B is compatible with .A = B. Since many
mathematicians do not like this occurrence, the notation .A ⊆ B is often found
in the literature, so that .A ⊂ B means .A ⊆ B and .A = B. In this book we will
never understand .⊂ in this restrictive sense.
Definition 2.4 (Union and Intersection) The union of two sets A and B is the set
A ∪ B of all points that are element of either A or B (or both):
.
A ∪ B = {x | (x ∈ A) ∨ (x ∈ B)}.
.
The intersection of two sets A and B is the set .A ∩ B of all points that are elements
of both A and B:
A ∩ B = {x | (x ∈ A) ∧ (x ∈ B)}.
.
Two sets A and B are disjoint if .A ∩ B = ∅.

Definition 2.5 (Complement) The absolute complement of a set A is the set .A =
{x | x ∈/ A}. We remark that .A = A. The relative complement of a set A with
respect to a set X is .X \ A = X ∩ A.
Figures 2.1, 2.2, and 2.3 describe visually the basic operations on sets.
Definition 2.6 (Singleton) The set that contains only the element x is denoted by
{x} and called singleton x.
.
Fig. 2.1 Intersection of two

sets
Fig. 2.2 Union of two sets

Fig. 2.3 Difference of two \

sets
Let us suppose that for each element .α of a set A, which is called the index set,
we are given a set .Xα . We can extend our definition of union and intersection as
follows:

. {Xα | α ∈ A} = Xα = {x | ∃α(α ∈ A ∧ x ∈ Xα )}. (2.1)
α∈A

{Xα | α ∈ A} = Xα = {x | ∀α(α ∈ A ∧ x ∈ Xα )}. (2.2)
α∈A
A particular case arises when the index set is a collection .A of sets, and in this case
we can write

. {A | A ∈ A} = {x | x ∈ A for some A ∈ A}
and similarly

. {A | A ∈ A} = {x | x ∈ A for each A ∈ A}.
Exercise 2.2 For each positive real numbers .α and .β, let .Qα,β be the rectangle
[0, α] × [0, β] in the plane. Describe the sets
.

. {Qα,β | α > 0, β > 0}, {Qα,β | α > 0, β > 0}.
Theorem 2.1 Let A be an index set, and for each .α ∈ A let .Xα be a subset of a
fixed set Y . Then
(a) If B is a subset of A, then

. {Xβ | β ∈ B} ⊂ {Xα | α ∈ A},
and

. {Xβ | β ∈ B} ⊃ {Xα | α ∈ A}.

\ {Xα | α ∈ A} = {Y \ Xα | α ∈ A}, and .Y \ {Xα | α ∈ A} =
(b) .Y
{Y \ Xα | α ∈ A}.
Proof

(a) If .x ∈ {Xβ | β ∈ B}then there exists .β ∈ B such that .x ∈ Xβ . By assumption
.β ∈ A, and thus .x ∈ {Xα | α ∈ A}. If .x ∈ {Xα | α ∈ A} then .x ∈ Xα for
each .α ∈ A, so that in particular .x ∈ Xβ for each .β ∈ B. Thus .x ∈ {Xβ | β ∈

B}.
(b) If .x ∈ Y \ {Xα | α ∈ A} then .x ∈ Y and x is not an element of any .Xα ,
.α ∈ A. Hence x belongs to Y and for each .α ∈ A there holds .x ∈ / Xα . This

means that .x ∈ {Y \ Xα | α ∈ A}. Reversing this argument we prove the first
identity. Now, if .x ∈ Y \ {X α | α ∈ A} then .x ∈ Y and there exists .α ∈ A
such that .x ∈
/ Xα . Hence .x ∈ {Y \ Xα | α ∈ A}. Reversing this argument we
prove the second identity.

An ordered pair is a new object .(x, y) characterized by the following property:
two ordered pairs .(x, y) and .(u, v) are equal if and only if .x = u and .y = v.
Actually an ordered pair may be defined in terms of sets as follows.
Definition 2.7 (Ordered Pair)
(x, y) = {{x}, {x, y}}.

.
Exercise 2.3 Prove that indeed .(x, y) = (u, v) if and only if .x = u and .y = v.
Hint: by assumption .{{x}, {x, y}} = {{u}, {u, v}}. Consider first the case .x = y,
then deal with the general case.
Definition 2.8 (Relations) A relation is a set of ordered pairs: a relation is
therefore a set whose elements are ordered pairs.
If R is a relation, we usually write .xRy instead of the more formal .(x, y) ∈ R,

and we say that x is related to y via R.
Definition 2.9 The domain of a relation R is the set .{x | ∃y((x, y) ∈ R)}. The
range of a relation R is the set .{x | ∃x((x, y) ∈ R)}. The field of a relation R is the
union of the domain and of the range of R.
One of the simplest relations is the set of ordered pairs .(x, y) such that x is a
member of a fixed set A, and y is a member of a fixed set B. This relation reduces
Fig. 2.4 A cartesian product
therefore to
A × B = {(x, y) | (x ∈ A) ∧ (y ∈ B)} ,
.
and is called the cartesian product of A and B: see Fig. 2.4. It is clear that any
relation is a subset of the cartesian product of its domain and its range.
Remark 2.1 The identification of sets and relations usually sounds strange to
students. In this book we will never think of relations or functions like black boxes
which transform elements of some set into elements of some other set.
The inverse of a relation R, denoted by .R −1 , is the relation obtained by swapping
each of the ordered pairs belonging to R. Formally,
R −1 = {(y, x) | (x, y) ∈ R} ,
.
or equivalently .yR −1 x if and only if xRy.

The composition of two relations R and S is
R ◦ S = {(x, z) | ∃y((x, y) ∈ S ∧ (y, z) ∈ R)} .

.
We remark that, roughly speaking, first comes S, then comes R, and not viceversa.
The domain of .R ◦ S is the domain of S, while the range of .R ◦ S is the range of R.
This will be of crucial importance when we introduce functions.
Definition 2.10 Suppose that R is a relation and X is the set of all points that are
elements of either the domain or the range of R. We say that R is
• reflexive, if each element of X is in relation R with itself;
• symmetric, if xRy whenever .yRx;
• antisymmetric, if xRy and yRx imply .x = y;
• transitive, if xRy and yRz imply xRz.
Definition 2.11 An equivalence relation is a reflexive, symmetric and transitive

relation. An order relation is a reflexive, antisymmetric and transitive relation.
It is customary to use the symbol .∼ for equivalence relations, and .≤ for order
relations.
A function is a relation such that no two distinct members have the same first
coordinate. More explicitly, a relation f is a function if for each element x of its
domain there exists a unique element y of its range such that .(x, y) ∈ f , see Fig. 2.5.
Uniqueness means that if .(x, y) ∈ f and .(x, z) ∈ f , then .y = z. For a function it
is customary to abandon the general notation .(x, y) ∈ f (or xfy) in favor of .y =
f (x). Then .f (x) is the image of the element x of the domain of f . In mathematical
analysis a function .f ⊂ X × Y is denoted by the (more complicated) symbol
f : X → Y,
. x → f (x).
A function .f : X → Y is injective if distinct points of X have distinct images in Y .

Equivalently, .f (x1 ) = f (x2 ) implies .x1 = x2 . A function .f : X → Y is surjective
if the range of f coincides with Y . Equivalently, for each .y ∈ Y there exists .x ∈ X
such that .f (x) = y. Finally, a function .f : X → Y is bijective if it is both injective
and surjective.
Exercise 2.4 Let X and Y be sets. Prove that the map .f : X × Y → Y × X defined
by .f (x, y) = (y, x) for each .(x, y) ∈ X × Y is a bijection. In this sense, .X × Y and
.Y × X are essentially the same object.
If A is a set and f is a function, the set
f (A) = {y | ∃x(x ∈ A ∧ f (x) = y)} = {f (x) | x ∈ A}

.
Fig. 2.5 Intuition of a function

is called the image of the set A under f . Similarly, if B is a set and f is a function,
f −1 (B) = {x | ∃y(x ∈ B ∧ f (x) = y)}

.
is called the pre-image of B under f . We notice that .f −1 (B) is just the image of
the set B under the inverse relation .f −1 . Clearly .f (A) is a subset of the range of f ,
while .f −1 (B) is a subset of the domain of f .
Theorem 2.2 If f is a function and A and B are sets, then
(a) .f −1 (A \ B) = f −1 (A) \ f −1 (B);
(b) .f −1 (A ∪ B) = f −1 (A) ∪ f −1 (B);
(c) .f −1 (A ∩ B) = f −1 (A) ∩ f −1 (B).
More generally, if we are given a set .Xα for each member .α of a non-empty index
set C, then

(d) .f −1 ({Xα | α ∈ C}) = {f −1 (Xα ) | α ∈ C};
(e) .f −1 ( {Xα | α ∈ C}) = {f −1 (Xα ) | α ∈ C}.
Proof We prove part (e),leaving the rest of the proof as a simple exercise. A point
x is an element of .f −1 ( {Xα | α ∈ C}) if and only if .f (x) is an element of this
intersection, in which case .f (x) ∈ Xα for each .α ∈C. But the latter condition is
equivalent to .x ∈ f −1 (Xα ) for each .α ∈ C, i.e. .x ∈ {f −1 (Xα ) | α ∈ C}.

Remark 2.2 Any function f is invertible as a relation. However the inverse relation
f −1 need not be again a function: this happens if and only if for each y there exists
.
a unique x such that .yf −1 x, i.e. .f (x) = y. We have proved that the relation .f −1
is a function if and only if f is a bijective function. It is customary to say that a
function .f : X → Y is invertible if it is bijective.
Remark 2.3 Any injective function .f : X → Y can be somehow inverted, in the
sense that we can define a function .g : f (X) → X such that .g(y) = x if and only
if .f (x) = y. In general the domain of g is a proper subset of Y , but the rule which
defines g is exactly the same rule which defines .f −1 . Many mathematicians do not
require surjectivity in order to define invertible functions. This is fairly reasonable,
since .f (X) is the largest subset of Y on which we can define the inverse function of
the injective function f .
Exercise 2.5 Let .f : X → U and .g : Y → V denote two functions. Prove that
(x, y) → (f (x), g(y)) defines a function .f × g : X × Y → U × V , which we call
.
the Cartesian product of f and g. Prove the following statements:

(i) if f and g are injective, then so is .f × g;
(ii) if f and g are surjective, then so is .f × g.
2.1 Comments 15
Important: Sets or Subsets?

Most surveys of naïve set theory for mathematical analysis only deal with subsets
of a given universe. We followed another route, and this may have been surprising.
The use of a given universe is motivated by Russel’s paradox, but for the moment
this remains irrelevant to us. As we will see, denying the set of all sets is not the
only escape from Russel’s paradox.
2.1 Comments
We have presented a quick survey of Set Theory from a non-axiomatic viewpoint.

Most textbook in Mathematical Analysis contain similar information, with only
minor differences in the language. As an example, functions are typically defined
as rules of assignment instead of special relations between two sets. A standard
reference is [1], a book which goes however much beyond the level suggested by
the title.
Before proceeding further, we should stop and think about notation. It is a matter
of facts that most instructors discourage the abstract use of
. {x | P (x)} (2.3)
for the definition of a set. In this book we may seem to be lazy, since such a notation
is allowed and even typical. Let us try to elaborate on this issue.
From a very abstract viewpoint, (2.3) contains the troublesome formula
. {x | x = x} ,
which leads to the paradox of the universe. On the contrary, the more precise
formula
. {x | x ∈ U ∧ P (x)} ,
often written as .{x ∈ U | P (x)}, is admissible, since it defines a subset of a (given)

set .U. Nowadays, most introductory discussions about (naïve) Set Theory are
based on axiomatic theories which discard arbitrarily large sets, like ZF (Zermelo-
Fraenkel), and this accounts for the recommendation against the use of .{x | P (x)}.
On the contrary, we will discuss a different Axiomatic Theory of Sets which
allows large objects (called classes). In some sense, we should say that .{x | x = x}
exists as a class, but not as a set. Since the algebra of classes is quite similar to the
algebra of sets, at a first stage we forget the distinction and we allow a more relaxed
notation.
Reference
1. P.R. Halmos, Naive Set Theory (Dover Publications, 2017)

Chapter 3
Numbers
Abstract Classical mathematical analysis is actually analysis over the field of real
numbers. In a later chapter we will construct the set .R of real numbers from the
axiom of set theory, as a completion of the set .Q of rational numbers. These are
in turn constructed from the set .Z of (signed) integers, which are constructed from
the set .N of positive integers, or natural numbers. However, this approach is time-
consuming, and we prefer to quickly introduce real numbers axiomatically.
Important: Warning
The existence of the natural numbers 1, 2, 3, 4, . . . will be taken as granted. This is
a reasonable compromise in a first approach to Analysis. In the second half of the
book we will show that numbers can be defined in terms of sets.
3.1 The Axioms of R
The system of real numbers is a set R, together with a distinguished subset P

and two functions from R × R into R, called operations. The first operation
is the sum, and is denoted by (x, y) → x + y. The second operation is the
product, and is denoted by (x, y) → xy = x · y.
Remark 3.1 The subset P should be thought of as the subset of positive numbers.
The choice of isolating a subset P instead of introducing an order relation is clearly
idiosyncratic.

18 3 Numbers
Axiom 1 For each x ∈ R and y ∈ R there results x +y = y +x and xy = yx.
Axiom 2 For each x, y and z in R, there results x + (y + z) = (x + y) + z

and x(yz) = (xy)z.
Axiom 3 For each x, y and z in R, there results x(y + z) = xy + xz.
Axiom 4 There exist two distinct elements 0 and 1 in R such that 0 + x = x

and 1x = x for each x ∈ R.
Axiom 5 For each x ∈ R there exists a unique −x ∈ R such that x + (−x) =

0. If x ∈ R and x = 0, there exists a unique x −1 ∈ R such that xx −1 = 1. We
will write x − y instead of x + (−y), and x/y instead of xy −1 for y = 0.
One summarizes the first five axioms by saying that R is a field. Abstract fields
are algebraic structures, and we will not discuss them in this book. It is noteworthy
that these axioms allow us to recover the basic algebraic rules of manipulation of
numbers.
Theorem 3.1 If x, y, z and w are real numbers, and w = 0, then x + z = y + z
implies x = y. Furthermore, xw = yw implies x = y.
Proof Indeed
x = 1x = x1 = x(ww−1 ) = (xw)w−1
.
= (yw)w−1 = y(ww−1 ) = y1 = 1y = y.
Similarly,
x = x + (z + (−z)) = (x + z) + (−z)
.
= (y + z) + (−z) = y + (z + (−z)) = y.

3.1 The Axioms of R 19
Theorem 3.2 If x, y, z and w are real numbers such that z = 0 and w = 0, then
1. x0 = 0;
2. −(−x)) = x;
3. (w−1 )−1 = w;
4. (−1)x = −x;
5. x(−y) = −xy = (−x)y;
6. (−x) + (−y) = −(x + y);
7. (−x)(−y) = xy;
8. (x/z)(y/w) = (xy)/(zw);
9. (x/z) + (y/w) = (wx + zy)/(zw).
The proof is left as an exercise. For instance, x0 +x0 = x(0 +0) = x0 = 0 +x0,
thus x0 = 0. Or x + (−x) = 0 = (−x) + (−(−x)) = −(−x) + (−x), thus
x = −(−x).
Axiom 6 The three sets {0}, P and −P = {x ∈ R | − x ∈ P } are pairwise

disjoint, and their union is R:
.R = P ∪ {0} ∪ (−P ).
Remark 3.2 Equivalently, we could require that R = P ∪ (−P ) and P ∩ (−P ) =

{0}.
Axiom 7 If x and y are element of P , then x + y ∈ P and xy ∈ P .
We call P the subset of positive real numbers. The set −P is the subset of
negative real numbers. It is customary to write x < y instead of y − x ∈ P ,
or equivalently x − y ∈ −P . Moreover x ≤ y will mean that either x < y or
x = y. Finally, y ≥ x is the same as x ≤ y, and y > x is the same as x < y.
Theorem 3.3 For each x, y and z in R, we have

(i) x < y and y < z imply x < z;
(ii) exactly one of the relations x < y, x = y, x > y holds;
(iii) x < y implies x + z < y + z;
(iv) x < y, z > 0 imply xz < yz;
(v) x < y, z < 0 imply xz > yz;
20 3 Numbers
(vi) 1 > 0 and −1 < 0;

(vii) z > 0 implies 1/z > 0;
(viii) 0 < x < y implies 0 < 1/y < 1/x.
Proof (i) Since y − x ∈ P and z − y ∈ P , their sum z − x ∈ P : x < z. (ii)
according to Axiom 6, the number y − x must lie exactly in one of the three sets
P , {0}, −P . The proofs of (iii), (iv) and (v) are similar. (vi) We want to show that
1 ∈ P . If not, since 1 = 0, 1 ∈ −P and −1 ∈ P . Thus 1 = (−1)(−1) ∈ P . (vii)
Suppose z > 0 and 1/z < 0. Then 1 = z · (1/z) < 0z = 0, a contradiction. Finally,
(1/x) − (1/y) = (y − x)/(xy) ∈ P as a product of two positive numbers, and (viii)
follows.

It is now clear that ≤ is an order relation on R. This order is called total since it
satisfies (ii) of Theorem 3.3. In view of the first seven axioms, we may say that R is
a totally ordered field. Unfortunately this is not enough to distinguish R from other
different sets, like Q. We need the last axiom, that we call Dedekind completeness.
Axiom 8 Let A and B be subsets of R such that A = ∅, B = ∅, A ∩ B = ∅,

A ∪ B = R and for each a ∈ A, b ∈ B there results a < b. Then there exists
a unique element x ∈ R such that
(i) a ∈ R and a < x imply a ∈ A;
(ii) b ∈ R and x < b imply b ∈ B.
The element x is often called the separator of A and B. Since either x ∈ A or

x ∈ B, but of course not both, it follows that either A = {t ∈ R | t ≤ x} and
B = R \ A, or B = {t ∈ R | x ≤ t} and A = R \ B.
3.2 Order Properties of R
Definition 3.1 Let E be a non-empty subset of R. A number b ∈ R is an upper

bound for E [resp. a lower bound], if x ≤ b [resp. b ≤ x] for each x ∈ E. The set
E is bounded from above [resp. from below] if an upper [resp. lower] bound for E
exists. The set E is bounded if it is bounded both from above and from below. A
number s ∈ R is called the supremum (or least upper bound) of E, and we write
s = sup E, if s is the smallest upper bounds of E. Similarly a number s ∈ R is
called the infimum (or greatest lower bound) of E, and we write s = inf E, if s is
the largest lower bound of E.1
1 The symbols glb E for inf E and lub E for sup E are old-fashioned.
3.2 Order Properties of R 21
Concretely, s = sup E if and only if s is an upper bound for E, and for each
upper bound b of E there results s ≤ b. A similar statement holds for the infimum
of E, and is left as an exercise.
Example 3.1 Consider E = {x | (∃n ∈ N) (x = 1/(n + 1))}. With the convention
that N = {0, 1, . . .}, we claim that inf E = 0 and sup E = 1. Indeed, 1 ∈ E with the
choice n = 0, and clearly 1/(n + 1) < 1 for each n ≥ 1. On the other hand, 0 is a
lower bound for E, since x ≥ 0 for each x ∈ E. We fix any ε > 0 and prove that ε
is not a lower bound for E. Indeed, there exists n0 ∈ N so large that 1/(n0 + 1) < ε:
any positive integer larger that 1/ε − 1 will suffice.2 Hence 0 = inf E.
Definition 3.2 Let E be a non-empty subset of R. If b ∈ E is an upper bound of E,
then we call it the maximum of E, denoted by b = max E. Similarly, if b ∈ E is a
lower bound of E, then b is called the minimum of E, denoted by b = min E.
Definition 3.3 For each x ∈ R we define the absolute value of x to be
x if x ≥ 0
.|x| =
−x if x < 0.
It follows that |x| = max{x, −x}.

Example 3.2 The absolute value has a few algebraic properties that follow directly
from the definition. Clearly |x| ≥ 0 for each x. Furthermore, |xy| = |x||y| for each
x and y. The proof is trivial if both x and y are positive. If x < 0 and y < 0, then
xy > 0, and so |x||y| = (−x)(−y) = xy = |xy|. If x < 0 and y > 0, then xy < 0
so that |x||y| = (−x)y = −xy = |xy|. Every other case can be reconducted to
these ones.
Theorem 3.4 (Triangle Inequality) For each real numbers x and y, there results
|x + y| ≤ |x| + |y|.
.
Proof We consider two cases. The first one is x + y ≤ 0, so that |x + y| = x + y ≤

|x|+|y|. The second case is x +y < 0, so that |x +y| = −(x +y) = (−x)+(−y) ≤
|x| + |y|.

Exercise 3.1 Prove the following reversed triangle inequality: for each real num-
bers x and y, there results ||x| − |y|| ≤ |x − y|. Hint: write x = (x − y) + y and
deduce that |x| ≤ |x − y| + |y|. Now swap x and y.
Theorem 3.5 (Supremum Principle) Every non-empty set E ⊂ R that is bounded
from above has a supremum in R.
2 This proof actually leans on Theorem 3.11.

22 3 Numbers
Proof Let B the set of all upper bounds of E, and set A = R \ B. Then B = ∅, and
if x ∈ E, then x − 1 ∈ A, so that A = ∅. Let a ∈ A, b ∈ B; then a is not an upper
bound of E, so there exists x ∈ E such that a < x ≤ b and thus a < b. It follows
now from Axiom 8 that there exists a unique real number s such that a ≤ s for each
a ∈ A and s ≤ b for each b ∈ B. If s ∈ A, then there would exist x ∈ E with s < x.
Setting a = (s + x)/2 we would derive a ∈ A because a < x ∈ E, and s < a:
contradiction. Therefore s ∈ B, and is an upper bound of E. Since s is smaller than
any upper bound b of E, it follows that s = sup E.

Corollary 3.1 Every non-empty set F ⊂ R that is bounded from below has an
infimum in R.
Proof Let E be the set of all lower bounds of F . It follows that E = ∅ and that
each element of F is an upper bound of E. Hence there exists s = sup E. Let us
now prove that s = inf F . Indeed, it is clear that s is larger than any each lower
bound of F . To conclude we need to check that s is a lower bound of F . Suppose
not. Then there exists y ∈ F such that y < s. Then y is not an upper bound of E,
so there exists x ∈ E with y < x. Since y ∈ F , this is impossible, and s is a lower
bound of F .

The following is a useful characterization of the supremum. The reader is invited
to prove this result, and to provide a similar statement for the infimum.
Theorem 3.6 Let E be a non-empty subset of R. The real number s is the supremum
of E if and only if
1. for each x ∈ E, x ≤ s;
2. for each ε > 0 there exists x ∈ E with s − ε < x.
Exercise 3.2 Consider subsets A and B of R. If A ⊂ B, prove that inf B ≤ inf A ≤
sup A ≤ sup B. Hint: Trivially, inf A ≤ sup A. Pick a ∈ A and observe that a ∈ B.
Hence a ≤ sup B. Since a ∈ A is arbitrary, deduce that sup A ≤ sup B. Now
complete the proof.
3.3 Natural Numbers
Although any mature mathematician should be aware that natural numbers are set-
theoretic objects, many analysts prefer to consider them as real numbers. This is
how we introduce them now, postponing a set-theoretic definition to a later chapter.
Definition 3.4 (Inductive Sets) We say that a set .I ⊂ R is inductive if and only if
(i) .1 ∈ I and
(ii) .x ∈ I implies .x + 1 ∈ I .
3.3 Natural Numbers 23
Definition 3.5 The set .N is the smallest (in the sense of set inclusion) inductive
subset of .R, i.e.

.N = {I | I is an inductive subset of R} .
Remark 3.3 Plainly 0 is not a natural number according to our definition. This is
just a matter of choice, since one could replace 1 by 0 in the definition of inductive
sets. To be honest, names are just names.
Important: 0 or 1?
Whether .N must contain 0 is an exhausting discussion. Many—if not most—
mathematicians tend nowadays to include 0, but there is a (funny) issue. We have
defined the natural numbers by picking 1 as the first element, and then by adding it
recursively. If we choose 0 as the first element, we cannot add it recursively, since
0 is the neutral element of the addition in .R. To sum up, starting with 0 breaks the
construction of .N as those numbers obtained by adding up the first element as many
times as we wish.
Example 3.3 The set .R is inductive (trivially). The set
R+ = {x ∈ R | x ≥ 0}
.
is also inductive.
Exercise 3.3 Prove that each half-line .[a, +∞), where .a ∈ R, is an inductive set.
Proposition 3.1 .N is an inductive set.
Proof Indeed .1 ∈ N, since 1 is an element of each inductive subset of .R. Now let
x ∈ N, so that x is an element of each inductive subset I of .R. Then .x + 1 ∈ I , and
.
since I is arbitrary we get .x + 1 ∈ N.

The following is a mere restatement of the inductive property of .N. Anyway, it
deserves a special name for historical reasons.
Theorem 3.7 (Induction Principle) Suppose .S ⊂ N is such that .1 ∈ S and .x ∈ S
implies .x + 1 ∈ S. Then .S = N.
Proof Indeed S is an inductive set of .R, then .N ⊂ S. But .S ⊂ N by assumption,
and equality follows.

Example 3.4 We prove that .1 + 2 + · · · + n = 12 n(n + 1) for each natural number
n. We formally put

1
A = n ∈ N 1 + 2 + · · · + n = n(n + 1) ,
.
2
24 3 Numbers
and show that .A = N. Since .A ⊂ N, it is enough to prove that A is inductive. For

.n = 1 we have .1 = 12 · 1 · 2, and .1 ∈ A. We suppose that .n ∈ A, and we check that
.n + 1 ∈ A. Indeed,
1
1 + 2 + ···+ n+ n+ 1 =
. n(n + 1) + (n + 1)
2
1
= (n + 1) k+1
2
1
= (n + 1)(n + 2),
2
and we conclude that .n + 1 ∈ A. Hence A is an inductive set, thus .A ⊃ N. Since
A ⊂ N, we necessarily have .A = N.
.
Example 3.5 We prove that .2n > n2 for each integer .n ≥ 5. Indeed, our statement
is equivalent to .2m+4 > (m + 4)2 for each integer .m ≥ 1. Let

A = m ∈ N 2m+4 > (m + 4)2 .
.
By direct computation, .1 ∈ A. Suppose that .k ∈ A; then
((k + 1) + 4)2 = ((k + 4) + 1)2

.
= (k + 4)2 + 2(k + 4) + 1
< (k + 4)2 + 2(k + 4) + (k + 4)
= (k + 4)2 + 3(k + 4)
< (k + 4)2 + (k + 4)(k + 4)
= 2(k + 4)2 < 2 · 2k+4 = 2(k+1)+4.
Thus .k + 1 ∈ A, and .A = N by induction.

Theorem 3.8 (Binomial Theorem) For every real numbers a and b and for every
n ∈ N there results
.

n
n k n−k
. (a + b)n = a b ,
k
k=0

where . nk = n!
k!(n−k)! .
Exercise 3.4
(i) Prove that for every .n ∈ N and every integer .k ∈ {1, . . . , n} there results
n n n+1
. + = .
k−1 k k
3.3 Natural Numbers 25
(ii) Using mathematical induction, prove Theorem 3.8.

Theorem 3.9 For each .n ∈ N there results
(i) .1 ≤ n;
(ii) .n > 1 implies .n − 1 ∈ N;
(iii) .x ∈ R, .x > 0, .x + n ∈ N imply .x ∈ N;
(iv) .m ∈ N, .m > n imply .m − n ∈ N;
(v) .a ∈ R, .n − 1 < a < n imply .a ∈

/ N.
Proof (i) The set .{x ∈ R | x ≥ 1} is an inductive set, hence it contains .N. (ii) Let
.S = {1} ∪ {n ∈ N | n − 1 ∈ N}. Plainly .1 ∈ S, and if .n ∈ S then .(n + 1) − 1 = n ∈
S ⊂ N. Hence .n + 1 ∈ S, and S is inductive. It follows that .S = N. Let now T be
the set of points .n ∈ N such that (iii) holds. It follows from (ii) with .x + 1 instead
of n that .1 ∈ T . Let .n ∈ T and .x > 0 be such that .x + (n + 1) ∈ N. Now .x + 1 > 0
and .(x + 1) + n ∈ N. Hence .x + 1 ∈ N. Again (ii) implies .x ∈ N. This proves that
.n + 1 ∈ T , and thus .T = N. (iv) follows from (iii) upon setting .x = m − n. Finally,
if (v) were false, then we would have .a ∈ N, .n < a + 1 and .(a + 1) − n < 1. From
(iv) it would follow that .a + 1 − n ∈ N, and this contradicts (i).

The following is a fundamental property of the natural numbers. It looks almost
trivial, but we encourage the reader to keep in mind the well-order property, since it
will come back in a more complicated way.
Theorem 3.10 (.N Is Well-Ordered) Any non-empty subset of .N has a minimum.
Proof Let A be a non-empty subset of .N, and suppose it has no smallest element.
Set
S = {n ∈ N | ∀a(a ∈ A ⇒ n < a)}.

.
It is clear that .1 ∈ S, otherwise 1 would be the smallest element of A. Assume that

n ∈ S. If .n + 1 ∈
. / S, there exists .a ∈ A such that .a ≤ n + 1. Since .n ∈ S, .n < a,
and thus a is a natural number lying between n and .n + 1. The previous theorem
yields .a = n + 1, and therefore a is the smallest element of A. This contradiction
shows that .n + 1 ∈ S. By induction, .S = N. Let a be any element of A: then .a < a,
a contradiction. We conclude that A has a smallest element.

The next result shows a deep interplay between .R and .N.
Theorem 3.11 (Archimedean Property of .R) If a and b are real numbers and
a > 0, then there exists .n ∈ N such that .na > b. As a particular case, .N is not
.
bounded from above.

Proof If the conclusion is false, then the set .E = {na | n ∈ N} is bounded from
above by b. Let .s = sup E. It follows from the definition of supremum that there
exists .n ∈ N with .na > s −a. Then .(n+1)a ∈ E and .(n+1)a > s, in contradiction
with the choice of s. Choosing .a = 1 yields the last statement.

We may now add a sign to the natural numbers.
26 3 Numbers
A real number x is an integer number if either .x = 0, or .x ∈ N, or .−x ∈ N.

The set of integer numbers is denoted by .Z.
Theorem 3.12 Let .x ∈ R. Then there exists a unique integer n such that .n ≤ x <
n + 1, .x − 1 < n ≤ x.
Proof Suppose the conclusion holds true for different integers .n < m. Then .n <
m ≤ x and .x < n + 1, thus .n < m < n + 1. Hence there would exist an integer m
between the consecutive integers n and .n + 1, a contradiction to (v) of Theorem 3.9.
Let a be the smallest element of .N that is greater than .|x|. If .x ≥ 0, we take
.n = a − 1. If .x < 0, we take .n = −a + 1 or .−a according as x is an integer or
not.

Definition 3.6 The unique integer n of the previous theorem is called the integral
part of the real number x, and is denoted by .[x].
A rational number is any real number of the form .a/b, for some integers a
and b, .b = 0. The set of rational numbers is denoted by the symbol .Q. The
elements of .R \ Q are called irrational numbers.
The reader will easily check that .Q satisfies Axioms 1–5, so that it is a field itself.
A classical result shows that irrational numbers exist.
√
Theorem 3.13 . 2 ∈ / Q.
√
Proof Assume that . 2 = m/n for some integers m, n, .n = 0. We may assume that
m and n are coprime, in the sense that the fraction .m/n cannot be further reduced.
Then .2 = m2 /n2 , i.e. .m2 = 2n2 . Hence .m2 is an even number, and so is m. Hence
.m = 2k for some integer k, and .m = 4k . Thus .4k = 2n , or .2k = n . This
2 2 2 2 2 2
2
yields that .n is even, and so is n. But m and n are coprime, a contradiction.

Thus rational numbers do not exhaust .R. However, there are no “holes” between
rational numbers.
Theorem 3.14 (Density of .Q in .R) If x, y are real numbers with .x < y, there
exists .z ∈ Q such that .x < z < y.
Proof The Archimedean property yields .b ∈ N such that .b > (y − x)−1 . Then
.b
−1 < y − x. Let .a = [bx] + 1 ∈ Z. Hence .a − 1 ≤ bx < a, i.e. .a/b < x + b −1 and
.x < a/b. We conclude that .x < a/b < x + b

−1 < x + (y − x) = y. The conclusion
follows with .z = a/b.

Exercise 3.5 Prove that .Q intersects any open interval of .R. This property will be
called topological density of .Q in .R.
3.5 Complex Numbers 27
3.4 Isomorphic Copies
Theorem 3.15 Every ordered field contains sets isomorphic to the natural num-
bers, the integers, and the rational numbers.
Proof Let us consider any ordered field (X, +, ·, ≤). By induction we may define a
function ϕ : N → X recursively as follows:
ϕ(1) = 1
.
ϕ(n + 1) = ϕ(n) + 1.
Let p and q be different natural numbers, say p < q. There exists n ∈ N such that
q = p + n: we claim that ϕ(p) < ϕ(q).
Indeed, for n = 1 we just have q = p + 1 and ϕ(q) = ϕ(p) + 1 > ϕ(p). For
a general n ∈ N we have ϕ(p + n + 1) = ϕ(p + n) + 1 > ϕ(p + n), and so
ϕ(p + n) > ϕ(p) implies ϕ(p + n + 1) > ϕ(p). By induction ϕ(p + n) > ϕ(p)
and we see that ϕ is injective.
Again by induction we can show that ϕ(p + q) = ϕ(p) + ϕ(q) and ϕ(pq) =
ϕ(p)ϕ(q). Thus ϕ is a bijective function from N onto a subset of X which preserves
sums, products and the order relation. By taking differences of natural numbers we
obtain Z as a subset of X, and by taking quotients of integers we obtain Q as a
subset of X. The proof is complete.

3.5 Complex Numbers
As we have seen, up to isomorphisms we may always assume that .N ⊂ Z ⊂ Q ⊂ R.

We close this chapter on number systems with a quick introduction to the filed of
complex numbers. For the first time we define a set larger than .R.
Definition 3.7 The field of complex numbers is the set .C of ordered pairs of real
numbers, together with two operations. The sum of two complex numbers .z = (a, b)
and .w = (c, d) is defined to be
z + w = (a + c, b + d).
.
The product of z and w is defined to be
.zw = (ac − bd, ad + bc).
The set .C is a field under these operations. Indeed the complex number .(0, 0)
satisfies .(a, b) + (0, 0) = (a, b) for any .(a, b) ∈ C. The complex number .(1, 0)
satisfies .(a, b)(1, 0) = (a, b) for any .(a, b) ∈ C. If .z = (a, b), then .−z = (−a, −b)
28 3 Numbers
is such that .z + (−z) = (0, 0). If .z = (a, b) = (0, 0), then the number
a b
z−1 =
. ,− 2
a2 +b 2 a + b2
is such that .zz−1 = (1, 0), namely the multiplicative inverse of z.

If .z = (a, b) is a complex number, we call a the real part of z, and b the
imaginary part of z. The symbols are .a = z, .b = z. The modulus of z is

|z| =
. a 2 + b2 ,
and the conjugate of z is
z = (a, −b).
.
Definition 3.8 The imaginary unit is the complex number .i = (0, 1).
The main reason for introducing .i is that .(a, b) = (a, 0) + (b, 0)(0, 1). If we
identify .(a, 0) with a, .(b, 0) with b, we can formally write .(a, b) = a + bi.
Exercise 3.6 Let .z = (a, b) and w be complex numbers. Prove that
1. .z = a − bi
2. .z+w =z+w
3. .zw = zw
4. .z + z = 2z and .z − z = 2i z
.zz = |z|
5. 2
6. .z
−1 = z/|z|2 provided that .z = 0.
Proposition 3.2 Let z, w be complex numbers. Then
(i) .|z| > 0 unless .z = 0, and .|0| = 0;
(ii) .|zw| = |z| |w|;
(iii) .|z| = |z|;
(iv) .|z| ≤ |z|, .| z| ≤ |z|;
(v) .|z + w| ≤ |z| + |w|.
Proof (i) is obvious from the properties of real numbers. Let .z = a +bi, .w = c+di.
Then .|zw|2 = (ac − bd)2 + (ad + bc)2 = (a 2 + b2 )(c2 + d 2 ) = |z|2 |w|2 . Since the
modulus cannot be negative, (ii) follows.
√ (iii) is trivial. To prove (iv) we just remark
that .a 2 ≤ a 2 + b2 , so that .|a| ≤ a 2 + b 2 . The same holds the imaginary part b.
Finally, .zw + zw = 2(zw). Hence
.|z + w|2 ≤ (z + w)(z + w)

= zz + zw + zw + ww
= |z|2 + 2(zw) + |w|2
3.5 Complex Numbers 29
≤ |z|2 + 2|zw| + |w|2

= |z|2 + 2|z||w| + |w2 |
= (|z| + |w|)2 .
We conclude again by the positivity of the modulus.

Theorem 3.16 (Cauchy-Schwartz Inequality) Let .a1 ,. . . , .an and .b1 ,. . . ,.bn be
complex numbers. Then
n 2

n
n

. ak b k ≤ |ak |2 |bk |2

k=1 k=1 k=1
n n n
Proof Define .A = k=1 |ak |
2 , .B = k=1 |bk |
2 , .C = k=1 ak bk . Then

n
n
0≤
. |Bak − Cbk | = (Bak − Cbk )(Ba k − CBk )
2
k=1 k=1

n
n
= B2 |ak |2 − BC ak b k
k=1 k=1

n
n
= BC a k bk + |C|2 |bk |2
k=1 k=1
= B 2 A − B|C|2
= B(AB − |C|2 ).
We conclude that .AB ≥ |C|2 , since .B ≥ 0.

Important: Order Properties of the Complex Numbers

We might suspect that the order properties of .R could “pass over” to the set .C, since
real numbers can be identified with complex numbers whose imaginary part is zero.
However this is impossible: indeed we know that the product of two positive real
numbers is always positive, and in particular .x 2 ≥ 0 for each .x ∈ R. Hence .i2
should be a positive number, but .i2 = −1. Mathematical analysis in the field of
complex numbers is deeply influenced by this fact.
30 3 Numbers
3.6 Polar Representation of Complex Numbers
Every complex number .z = (x, y) = x + iy = (0, 0) can be written in the form
. z = r (cos ϑ + i sin ϑ) ,

for some numbers .r ≥ 0 and .ϑ ∈ [0, 2π). Indeed, .r = |z| = x 2 + y 2 (since
.cos ϑ + sin ϑ = 1), and .ϑ is defined as follows:
2 2
(a) if .x = 0, then
y
ϑ = arctan
. + kπ,
x
where
⎧
⎪
⎨0 if x > 0 and y ≥ 0
⎪
k=
. 1 if x < 0 and y ∈ R
⎪
⎪
⎩2 if x > 0 and y < 0.
(b) If .x = 0 and .y = 0, then
π
if y > 0
ϑ=
.
2
3
2π if y < 0.
It is impossible to associate an angle .ϑ to the complex number .z = 0 in a coherent

way, as Fig. 3.1 shows.
The representation
z = |z| (cos ϑ + i sin ϑ)

.
is called the trigonometric form of the complex number .z = 0. The number .ϑ ∈

[0, 2π) is called the principal argument of z, and it is often denoted by .Arg z.
Exercise 3.7 Prove that two complex numbers
z1 = |z1 | (cos ϑ1 + i sin ϑ1 )

.
z2 = |z2 | (cos ϑ2 + i sin ϑ2 )
such that .z1 = 0 and .z2 = 0 are equal if and only if .|z1 | = |z2 | and .ϑ1 − ϑ2 = 2kπ
for some integer k.
3.6 Polar Representation of Complex Numbers 31
Fig. 3.1 Trigonometric intuition of a complex number
Proposition 3.3 Suppose that
z1 = |z1 | (cos ϑ1 + i sin ϑ1 )

.
z2 = |z2 | (cos ϑ2 + i sin ϑ2 ) .
Then
.z1 z2 = |z1 ||z2 | (cos(ϑ1 + ϑ2 ) + i sin(ϑ1 + ϑ2 )) .
Proof Indeed,
z1 z2 = |z1 ||z2 | (cos ϑ1 + i sin ϑ1 ) (cos ϑ2 + i sin ϑ2 )

.
= |z1 ||z2 | (cos ϑ1 cos ϑ2 − sin ϑ1 sin ϑ2 + i(sin ϑ1 cos ϑ2 + sin ϑ2 cos ϑ1 ))
= |z1 ||z2 | (cos(ϑ1 + ϑ2 ) + i sin(ϑ1 + ϑ2 )) .

32 3 Numbers
Theorem 3.17 (De Moivre’s Formula) If .z = |z|(cos ϑ + i sin ϑ) and .n ∈ N, then
zn = |z|n (cos(nϑ) + i sin(nϑ)) .

.
Proof It is enough to recall that .zn = z · · · z (n times) and to apply Proposition 3.3.

3.7 A Construction of the Real Numbers
When I began to write this book, I did not consider any discussion about the
construction of .R in the first part. As I said before, it seems that a purely axiomatic
definition of the real numbers is more than enough as a first approach to Mathemati-
cal Analysis. But then some colleagues convinced me that an authoritative approach
(“Believe me, real numbers do exist!”) may not be the best choice for an instructor:
the pace of teaching is not the pace of logic.
In this section we (the reader and I) will meet the basic ideas of a popular
construction of .R from .Q. Let me try to introduce the topic.
Imagine you know and use rational numbers (as √ fractions), but you have no idea
about irrational numbers. What is, for example, . 2? Well, Dedekind proposed to
identify real numbers
with subsets of .Q, in an appropriate way. Roughly speaking,
√ √
. 2 = r ∈ Q r < 2 .
This definition should be prosecuted by the law, of course: nothing should be
defined recursively! But look at the following proposal:
√

. 2 = r ∈ Q r ≤ 0 or r 2 < 2 .
This looks much better, doesn’t it? A (real) number is a set of rational numbers; but
what sets?
Definition 3.9 (Dedekind Cuts) A subset L of .Q is a Dedekind cut, if
(a) .L = ∅ and .L = Q;
(b) for every .x ∈ L there exists .y ∈ L such that .x < y;
(c) if .x ∈ L and .y < x, then .y ∈ L.
In our minds, a cut is a proper, non-empty subset of .Q which has no largest
element (condition (b)) and looks like a half-line starting from .−∞ (condition (c)).
The term cut can be explained as follows: if L is a Dedekind cut, then .Q is cut by L
in two parts, L and .Q \ L, such that any element of L is smaller than any element of
.Q \ L.
Example 3.6 For every .r ∈ Q, the set
Lr = {x ∈ Q | x < r}
.
3.7 A Construction of the Real Numbers 33
is a Dedekind cut. However, our initial discussion suggests that this type of subsets
does not exhaust the class of cuts.
We are tempted to define .R as the collection of all Dedekind cuts. The main issue
is that .R must be an ordered field.
We can now abandon the letter L (as in Left) to denote Dedekind cuts, and
use Greek letters instead. Hence .α, .β, .γ , . . . will be real numbers, i.e. cuts.
Definition 3.10 (Order) Let .α ∈ R and .β ∈ R. We say that .α ≤ β if and only if

α ⊂ β.
.
It is not difficult to convince oneself that .≤ is indeed a reflexive, antisymmetric

and transitive relation on .R. In other words, it is an order relation. As a first step, we
can be satisfied.
Now algebra comes into play, since .R must be endowed with two operations.
Definition 3.11 (Sum) Let .α ∈ R and .β ∈ R. The sum of .α and .β is defined as
α + β = {p + q | p ∈ α, q ∈ β} .
.
Of course we should check that the set .{p + q | p ∈ α, q ∈ β} is a Dedekind

cut, but this is a straightforward exercise.
Exercise 3.8 Let .α ∈ R. Show that
. − α = {r ∈ Q | ∃s(s > r ∧ −s ∈
/ α)}
is a Dedekind cut, and that .α + (−α) = 0, where .0 = {r ∈ Q | r < 0}.

The product should be defined carefully, since we the product of negative
numbers is expected to be a positive number.
Definition 3.12 (Product) Let .α ∈ R and .β ∈ R. The product of .α and .β is firstly
defined in the case .α ≥ 0, .β ≥ 0 as
αβ = {pq | p ∈ α, q ∈ β} .
.
Then
⎧
⎪
⎨− ((−α)β)
⎪ if α < 0 and β ≥ 0
.αβ = − (α(−β)) if α ≥ 0 and β < 0
⎪
⎪
⎩(−α)(−β) if α < 0 and β < 0.
34 3 Numbers
Clearly enough, we have adjusted the signs so that the elementary properties of
the algebraic product with respect to the order relation are satisfied.
Exercise 3.9 The real number 1 is defined as
1 = {r ∈ Q | r < 1} ,
.
where the number 1 in the right-hand side is the rational number .1 = 1/1. Show
that for any .α ∈ R, .α = 0, there exists .α −1 ∈ R such that .αα −1 = 1.
We will not prove in detail that .R with these two algebraic operations is a field:
the proofs are straightforward but boring, and can be a good exercise for the reader.
Instead, a proof of (some version of) completeness is more interesting. Let us state
it as follows.
Theorem 3.18 The ordered field .R (obtained from Dedekind cuts) has the upper
bound property: any non-empty subset, bounded from above, has a least upper
bound in .R.
Proof Let .A ⊂ R be a non-empty set, and let .β ∈ R an upper bound for A. Let us
set

.γ = {α | α ∈ A} .
This definition is meaningful, since the elements of A are sets. We are going to show
that .γ = sup A.
Pick any .α0 ∈ A, so that .α0 = ∅. Since .α0 ⊂ γ , .γ = ∅. Then .γ ⊂ β by
construction, hence .γ = Q. Let now .p ∈ γ , hence there exists .α1 ∈ A such that
.p ∈ α1 . If .q < p, then .q ∈ α1 and thus .q ∈ γ . If we finally choose .f ∈ α1 such that
.r > p, we see that .r ∈ γ : we have proved that .γ is a Dedekind cut, or equivalently
.γ ∈ R.
The very definition of .γ implies that .α ⊂ γ for every .α ∈ A. To prove that .γ is

the least upper bound of A, we fix .δ < γ . Hence there exists .s ∈ γ such that .s ∈ / δ.
Moreover, there exists .α ∈ A such that .s ∈ α. Hence .δ < α, and .δ is not an upper
bound for A. The proof is now complete.

Dedekind cuts are probably the most elementary model of the real numbers as
a complete ordered field. It is important to keep in mind that any construction of
.R must start from a consistent definition of the rational numbers. From a logical
viewpoint, the axiomatization of .R is equivalent to the axiomatization of .N: we

have defined .N as a subset of .R, but we have also proved that .R can be constructed
from .Q, and hence from .N.
3.9 Comments 35
3.8 Problems
3.1 If r is a rational number different than zero, and if x is an irrational number,

prove that both r + x and rx are irrational numbers.
3.2 Fix a real number b > 1.
1. If m, n, p and q are integers, n > 0, q > 0, and r = m/n = p/q, prove that
1/n 1/q
. bm = bp .
In particular we may define br = (bm )1/n .

t prove that b = b b .
2. If r and s are rational numbers, r+s r s

3. If x ∈ R, define B(x) = b t ∈ Q, t ≤ x . Prove that
br = sup B(r)
.
whenever r is rational. In particular we may define bx = sup B(x) for each

x ∈ R.
4. Prove that bx+y = bx by for each real numbers x and y.
This problem provides a rigorous definition of the power of a real number.
3.3 Let z ∈ C be a complex number with the property that |z| = 1. Compute
|1 + z|2 + |1 − z|2 .
.
3.4 Compute the infimum and the supremum of the set

1 1 1
E= + + {k, m, n} ⊂ N .
5n
.
2k 3m
3.9 Comments
The chapter on real numbers is always the most important one in textbooks about
(Real) Mathematical Analysis, and often the less self-contained one. The reason is
that an rigorous definition of .R requires a strong background in Set Theory and in
Abstract Algebra. We will see that natural numbers stem directly from Set Theory,
while rational numbers can be defined in terms of natural numbers with an algebraic
construction. These steps are usually omitted, since most students are satisfied with
intuitive definitions like
36 3 Numbers
Natural numbers are 0, 1, 2, 3, . . .
or
Rational numbers are just quotient of integers.
Needless to say, the first definition is based on the hope that every student
can decide whether an object is a natural number, and the second definition is
meaningless until quotients are defined. In other words, rational numbers are
quotient only when rational numbers already exist, or when real numbers already
exist.
However, it turns out that such an intuition about numbers does suffice to develop
Calculus of one or more real variables. As a result, I believe that only few graduate
students can construct .R from .Q, and only a small minority of them can define .N
in terms of sets. I was one of those students, and this is why this book contains a
chapter on Axiomatic Set Theory.
The books [1] and [2] are good sources about numbers.
References
1. W. Rudin, Principles of Mathematical Analysis. International Series in Pure and Applied

Mathematics, 3rd edn. (McGraw-Hill Book Co., New York, 1976)
2. K.R. Stromberg, An Introduction to Classical Real Analysis (AMS Chelsea Publishing, 2015)
Chapter 4
Elementary Cardinality
Abstract What does it mean that two sets have the same number of elements? This
may appear clear if we can write down all the members in a finite list. The answer
becomes complicated if the sets contain infinitely many elements. In this chapter we
propose a definition of cardinality in an elementary fashion.
4.1 Countable and Uncountable Sets
Definition 4.1 (Sequences) A sequence is any function whose domain is of the

form N \ F , for some finite subset F of N. If X is a set, a sequence in X is any
function which takes values in X and whose domain is of the form N \ F , for some
finite subset F of N.
If s is a sequence, it is customary to abridge the notation s(n) to sn . Hence we
will also write {sn }n for a sequence, but we remark that n is a dummy variable:
{sn }n = {sj }j = {sk }k = . . .
Important: Notation for Sequences

Since a sequence is a function, one might wonder why we make so many efforts to
avoid the natural use of functional notation. This sounds as a reasonable question,
because historical habit remains the only answer. Sequence are often denoted by
(sn )n or "sn #n , to distinguish the sequence from the set of its values.
We try to illustrate our definition of sequences.

Theorem 4.1 Let N be a subset of N. The following statements are equivalent:
(a) N = N \ F for some finite subset F of N;
(b) N contains an interval of the form N ∩ [n0 , +∞) for some n0 ∈ N.
Proof If (a) holds, we call n0 − 1 the largest positive integer which does not belong
to N. Then (b) holds. Conversely, we suppose that (b) holds and we consider the

38 4 Elementary Cardinality
finite set {1, 2, . . . , n0 − 1}. Thus at most finitely many positive integers do not
belong to N, and (a) holds.

In other words, our sequences may be considered as functions from an
unbounded interval N ∩ [n0 , +∞) for some n0 ∈ N. In the Comments at the
end of the chapter we will discuss again our definition.
Definition 4.2 (Subsequences) Let s be a sequence, and let k : N → N a sequence
of positive integers with the property that kn < kn+1 for each n ∈ N. Then the
composition s ◦ k is called a subsequence of s. Explicitly, s ◦ k = {skn }n .
Remark 4.1 In a subsequent chapter we will see that a weaker condition on
the sequence k could be assumed in order to define subsequences. The strong
monotonicity kn < kn+1 is however more popular in the literature.
Definition 4.3 (Equal Cardinality) Two sets A and B are equinumerous (or have
the same cardinality), if there exists a bijective function F : A → B. In this case we
will write A ∼ B, or even #A = #B.
It is an easy exercise in set theory to check that ∼ is actually an equivalence
relation between sets. We will use this fact in the rest of the chapter.
Definition 4.4 We say that a set A has cardinality n, if A ∼ {1, 2, . . . , n}. By
extension, the cardinality of the empty set is zero. A set A is finite, if there exists a
positive integer n such that A has cardinality n. Otherwise it is called infinite. A set
A is countably infinite if A ∼ N, and it is countable if it is either finite or countably
infinite. If A is not countable, we say that A is uncountable.
Important: Finite or Countable?

The use of the adjective “countable” is not completely universal. Several mathemati-
cians actually think of countable sets as countably infinite sets. Hence they would
not say that {5, 7, 11, 23} is a countable set. In my opinion, such an agreement is
popular among analysts, who seldom work with finite structures. For this reason, it
may happen that in this book the word countable can be used instead of countably
infinite. The reader should not have any trouble in recognizing such an abuse of
language.
Exercise 4.1 Prove that the Cartesian product of two finite sets is a finite set. Hint:
this is essentially a “matrix” proof. If X has n members and Y has m members, you
can write down X × Y as a matrix of n rows and m columns. Then just... count the
entries of this matrix.
A countably infinite set S can always be described as S = {s1 , s2 , . . .}, where s
is the bijective function that describes the fact that A ∼ N. In this sense, a countably
infinite set can be seen as a labeled list of points.
4.1 Countable and Uncountable Sets 39
Theorem 4.2 Every subset of a countable set is countable.

Proof Let S be a countable set, and let A ⊂ S. If A is finite, there is nothing to
prove. We may therefore assume that A is infinite, and S is infinite as well. We
select a sequence s = {sn }n of distinct points such that S = {s1 , s2 , . . .}. We define
a function as follows: let k1 be the smallest positive integer such that sk1 ∈ A. If
k2 , k3 , . . . , kn−1 have been selected, we choose kn as the smallest positive integer
> kn−1 such that skn ∈ A. It is evident that kn < kn+1 for each n. The composition
s ◦ k is defined on N and its range is A. Since skn = skm implies kn = km (because
the points s1 , s2 , . . . are distinct) and this implies n = m, we see that s ◦k is injective.
The proof is complete.

Theorem 4.3 The cartesian product N × N is countably infinite.
Proof For each (m, n) ∈ N × N we set f (m, n) = 2m 3n . This is an injective
function whose range is contained in N. Since this range is countable by the previous
theorem and N × N is clearly infinite, the proof is complete.

What about the cardinality of Q? To answer this question we need some
preliminary result about unions of countable sets.
We say that a family F of sets is a collection of disjoint sets, if any two elements
of F are disjoint.
Theorem 4.4 If F is a countable collection
disjoint sets, say F = {A1 , A2 , . . .},
of
such that each An is countable, then F = ∞ n=1 An is also countable.
∞
Proof For each n, let An = {a1,n , a2,n , a3,n , . . .}. Call S = n=1 An . Every
element x of S must lie in some An , thus x = am,n for some pair of integers (m, n).
This pair is uniquely determined, since F is a collection of disjoint sets. This defines
a function f : S → N × N via f (x) = am,n . We have just seen that f is injective,
so its range is countable. We conclude that S is also countable.

We want to remove the assumption that F should be a collection of disjoint sets.
This is possible, but it requires some attention.
Theorem 4.5 If F is a countable collection of countable sets, then the union of all
the members of F is also countable.
Proof We need to reduce to the case of a collection of disjoint sets. A standard way
to achieve this result is as follows: put B1 = A1 , and, for n > 1,

n−1
Bn = An \
. Ak .
k=1

Clearly
∞ G = {B1 , B2 , B3 , . . .} is a disjoint collection. Setting A = ∞ n=1 An , B =
B
n=1 n , we show that A = B. If x ∈ A, then x ∈ A k for some k. Let n be the
smallest k with this property, so that x ∈ / Ak for k < n. This implies x ∈ Bn , and in
turn x ∈ B. Viceversa, if x ∈ B, then x ∈ Bn for some n, and in particular x ∈ An

for the same n. The proof is complete.

Corollary 4.1 The set Q of rational numbers is countably infinite.
Proof We call An the set of all positive
rational numbers whose denominator is
n. The set Q is therefore equal to ∞n=1 n , a union of countable sets. The result
A
follows from the previous theorem and the trivial remark that Q is an infinite set.

We already know that R = Q as sets. We can now show that R has actually more
elements than Q.
Theorem 4.6 The set R is uncountable.
Proof Since the interval (0, 1) = {x ∈ R | 0 < x < 1} is a subset of R, it suffices
to show that (0, 1) is uncountable. Suppose not, so that there exists a sequence
s = {sn }n whose range is (0, 1). We show that this is impossible by constructing
a real number in (0, 1) which is not a term of the sequence s. As a starting point,
we assume that each real number can be uniquely written as an infinite decimal,
and in particular sn = 0.un,1 un,2 un,3 . . . Each un,i is a digit, i.e. an element of
{0, 1, 2, 3, 4, 5, 6, 7, 8, 9}. Consider the number y = 0.v1 v2 v3 . . . where
1 if un,n = 1
vn =
.
2 if un,n = 1.
We claim that no term of the sequence {sn }n can equal y. Indeed y differs from s1
in the first digit, differs from s2 in the second digit, and in general differs from sn in
the n-th digit. But 0 < y < 1 by construction, and this contradicts the assumption
that (0, 1) is countable.

Example 4.1 Every open subset (a, b) of R has the same cardinality as R. Indeed,
we choose a number c ∈ (a, b) and we define f : (a, b) → R as
x−c
if c ≤ x < b
f (x) =
.
b−x
x−c
x−a if a < x ≤ c.
It is easy to check that f is a bijective map.

Exercise 4.2 Let P be the set of all positive real numbers. Prove that (0, 1) and P
have the same cardinality by using the function f : (0, 1) → P defined by
x if 0 < x ≤ 1/2
f (x) =
.
1
4(1−x) if 1/2 < x < 1.
Exercise 4.3 Prove that any infinite set contains a countably infinite subset. Hint:
let X be an infinite set. Pick any x1 ∈ X. Since X is infinite, there exists x2 ∈
4.2 The Schröder-Bernstein Theorem 41
X \ {x1 }. For the same reason, there exists x3 ∈ X \ {x1 , x2 }, and so on. In this way
we construct a subset {xj | j ∈ N} of X which is clearly countably infinite.
Let us call c the cardinality of R and ℵ0 for the cardinality of N. From our
discussion it is clear that
ℵ0 < c,
.
in the sense that there exists an injective function from N into R, but there cannot
exist a bijection between these two sets.
Important: Question
Is there any set whose cardinality is strictly larger than ℵ0 and strictly smaller than
c?
The answer is more than difficult: it is actually impossible! To be more precise,

let us state the following
Continuum Hypothesis There exists no set whose cardinality κ satisfies ℵ0 <
κ < c.
Although David Hilbert proposed a proof that the continuum hypothesis was
actually true, it soon turned out that his proof was incorrect. Some years later, Gödel
showed that the continuum hypothesis cannot be disproved in the framework of
any consistent theory of sets. The debate was closed in 1963 by Paul Cohen, who
showed that the continuum hypothesis cannot be proved in the framework of any
consistent theory of sets, either. Roughly speaking, and since we always assume
to have a consistent Set Theory at our disposal, the continuum hypothesis remains
independent: it is a matter of taste whether we want to include it among our axioms.
Luckily enough, it is rather hard to single out a milestone of Mathematical Analysis
which depends on the continuum hypothesis. For this reason, we will not pursue
further this topic in the book.
4.2 The Schröder-Bernstein Theorem
We have decided that two sets have the same cardinality if a bijective map exists
which takes one set onto the other. A celebrated result by Schröder and Bernstein
simplifies our task.
Theorem 4.7 (Schröder-Bernstein) If there is a one-to-one function on a set A to
a subset of a set B and there is also a one-to-one function on B to a subset of A,
then A and B have the same cardinality.
Proof Suppose that .f : A → B and .g : B → A are two injective maps. We may

assume without loss of generality that .A ∩ B = ∅. We say that a point x of either A
or B is an ancestor of a point y if and only if y can be obtained from x by successive
application of f and g, or of g and f . Now we split A into three subsets: .AE
consisting of all points of A which have an even number of ancestors, .AO consisting
of all points of A which have an odd number of ancestors, and .AI consisting of all
points of A which have infinitely many ancestors. The set B can be split in the same
way. We finally define .F : A → B as follows:
f on AE ∪ AI
F =
.
g −1 on AO
is a bijective map.

Remark 4.2 How do we interpret the previous proof? We have actually constructed
the map F by an inductive process:
E0 = A \ g(B)
.
E1 = g(f (E0 ))
E2 = g(f (E1 ))
...
En+1 = g(f (En )),

and so on. Then we set .E = n En . The function F is constructed in such a way
that .F = f on A, and .F = g −1 on .A \ E.
We present a second proof of this important result in Set Theory. We need a
preliminary tool.
Lemma 4.1 Let .X be an ordered set such that every non-empty subset has a
greatest lower bound. If .f : X → X is such that
1. there exists .x ∈ X such that .f(x) ≤ x;
2. for every .x ∈ X, .y ∈ X, .x ≤ y implies .f(x) ≤ f(y),
then .f has a fixed point, i.e. there exists .a ∈ X such that .f(a) = a.
Proof The set
A = {x ∈ X | f(x) ≤ x}
.
is non-empty, hence there exists a greatest lower bound .a ∈ X for A. If .x ∈ A,

then .a ≤ x, hence assumption 2 implies .f(a) ≤ f(x) ≤ x. Thus .f(a) ≤ a, since
.a = inf A. Using again 2, we see that .f(f(a)) ≤ f(a), hence .f(a) ∈ A and so
.a ≤ f(a). The proof is complete.

4.3 Problems 43
Proof (of Theorem 4.7) Let .f : X → Y and .g : Y → X be injective functions. We

claim that there exists a subset A of X such that .g(Y \ f (A)) = X \ A. Once this
claim is proved, the construction of a bijective application of X onto Y is easy.
Let us define .F : 2X → 2X such that
A → X \ g(Y \ f (A)).
.
Lemma 4.1 can be applied with .X = 2X , ordered by inclusion .⊂, and .f = F , since
F satisfies condition 2. Condition 1 is also satisfied, since .2X contains a largest
element. Thus .F (A) = A for some .A ⊂ X, and the proof follows.

A remarkable fact is that given a set A, one can always construct another set
whose cardinality is different than the cardinality of A. We call .P(A) the set of all
subsets of A.
Theorem 4.8 (Cantor) If .A = ∅, then there exists no surjective map .f : A →
P(A). In particular, A and .P(A) do not have the same cardinality.
Proof Let .f : A → P(A); we will prove that the set
.S = {x ∈ A | x ∈
/ f (x)}
does not belong to the image of f . Suppose that .S ∈ f (A), so that .S = f (s) for
some member .s ∈ A. If .s ∈ S, then .s ∈
/ f (s) = S; if .s ∈
/ S, then .s ∈ f (s) = S. In
any case we reach a contradiction.

Exercise 4.4 Suppose that .A = {x}. What is the cardinality of .P(A)? Think
carefully!
4.3 Problems
4.1 A complex number z is an algebraic number if there exist integers a0 , . . . , an ,

not all zero, such that
a0 zn + a1 zn−1 + · · · + an−1 z + an = 0.
.
Prove that the set of all algebraic numbers is countable. Hint: given N ∈ N, there
exist only finitely many equations with n + |a0 | + · · · + |an | = N.
4.2 Is the set R \ Q countable?
4.3 Prove that a set E is infinite if and only if E has the same cardinality of a proper
subset of E. Hint: one direction is Exercise 4.3. Conversely, if f : E → E is an
injective function and a ∈ E \ f (E), define recursively a1 = f (a), an+1 = f (an ).
4.4 Comments
The rigorous definition of sequences is more problematic than we might suspect.

Most textbooks propose to call sequence in a set X any function from √ .N to X. But a
problem immediately arises: with this definition the function .n → n2 − 9 should
not be termed sequence. Our definition clearly absorbs the previous one.
A more refined definition appears in [1]: a sequence in a set X is any function
defined on an infinite subset of .N, taking values in X. It is easy to check that infinite
subsets of .N are characterized as follows.
Theorem 4.9 Let N be a subset of .N. The following statements are equivalent:
(a) N is an infinite set;
(b) for every .n ∈ N there exists .p ∈ N such that .p ≥ n.
We will see later that (b) is actually the characterizing property of nets, a
generalization of sequences.
Comparing sets by counting their elements obviously leads to a rather rough
classification. However, this is the first appearance of the concept of infinity, which
students consider from a philosophical viewpoint. We have proposed a standard
approach to elementary cardinality of sets, and in particular we have avoided
any explicit reference to the complicated issue of choosing elements from non-
empty sets. This immediately leads to the Axiom of Choice and to the exhausting
discussions about the necessity of using it.
Luckily, I have never found a student who needed an axiom to label the elements
of a countable collection of countable sets, although such an operation requires some
flavor of the Axiom of Choice. To clarify this point, we should always compare the
sentences
1. A is a countable set;
2. let .{s1 , s2 , s3 , . . .} be the elements of the countable set A.
The first statement is intrinsic, and we understand that an enumeration of the
elements of A exists. The second statement already contains the choice of an
enumeration of A, since the same countable set can be enumerated in infinitely
many different ways. To summarize, the Axiom of Choice is not needed to define
countable sets, but it comes into play as soon as we want to write down an
enumeration of a countable set.
The Schröder-Bernstein Theorem is a useful result which can be proved in several
ways. The first proof appears in [2] (but the author attributes it to G. Birkhoff and S.
Mac Lane), while the second in based on the fixed point Lemma 4.1. I believe that
both proofs are elegant and readable at an early stage.
References 45
References
1. S. Dolecki, F. Mynard, Convergence Foundations of Topology (World Scientific, 2016)

2. J.L. Kelley, General Topology. Graduate Texts in Mathematics, No. 27 (Springer, New York,
1975). Reprint of the 1955 edition [Van Nostrand, Toronto, Ont.]
Chapter 5
Distance, Topology and Sequences on the
Set of Real Numbers
Abstract The set .R has a rich algebraic structure. What is even more important for
Analysis is that its structure of ordered field with the Completeness Axiom may be
used to generate a topological environment.
We start with a fairly general definition that describes the possibility of measuring a
distance.
Definition 5.1 Let X be a set. A distance on X is a function .d : X × X → R with
the following properties:
1. .d(x, y) ≥ 0 for each .(x, y) ∈ X × X; .d(x, y) = 0 if and only if .x = y.
2. .d(x, y) = d(y, x) for each .(x, y) ∈ X × X.
3. (triangle inequality) .d(x, y) ≤ d(x, z) + d(z, y) for each x, y and z in X.
If a distance d is given on X, we say that .(X, d) is metric space.
Definition 5.2 The standard (or Euclidean) metric on .R is defined as .d(x, y) =
|x − y| for each .(x, y) ∈ R × R.
Whenever a distance is available, we can introduce the idea of neighborhood.
We will come back to this in greater generality; for the time being we stick to the
particular case of the standard metric in .R.
Definition 5.3 An open interval is any set of the form .(a, b) = {x ∈ R | a < x <
b} for some real numbers .a < b. If S is a subset of .R and .x0 is a point, we say
that .x0 is an interior point of S whenever there exists an open interval I such that
.x0 ∈ I ⊂ S. A set S is called open, if each point of S is an interior point of S. A set
S is called closed, if the complement .R \ S is open.

We will also need closed intervals, i.e. set of the form .[a, b] = {x ∈ R | a ≤
x ≤ b} for some real numbers .a ≤ b. Half-lines can be considered to be improper

48 5 Distance, Topology and Sequences on the Set of Real Numbers
intervals, for example
(a, +∞) = {x ∈ R | a < x}

.
(−∞, b) = {x ∈ R | x < b}
[a, +∞) = {x ∈ R | a ≤ x}
and so on.
The sentence “The set I is an interval” means that I is an interval of any kind:
open, closed, half-open, a half-line or even the whole real line.
Definition 5.4 A neighborhood of a point .x0 in .R is any set U such that an open
interval .(a, b) exists with .x0 ∈ (a, b) ⊂ U .
A neighborhood of .x0 is therefore any set that contains an open set that contains
x0 .
.
Example 5.1 A neighborhood need not be open: .U = [0, 1] is a neighborhood of

x0 = 1/2, but U is not open.
.
Exercise 5.1 Prove that if .x0 is a positive real number, there exists a neighborhood
of .x0 whose points are positive numbers. This harmless result will be used to prove
some results about limits of sequences and functions.
Definition 5.5 Let S be a subset of .R. A point .x0 ∈ R is an accumulation (or limit,
or cluster) point of S, if each neighborhood of .x0 contains a point of S different than
.x0 itself (see Fig. 5.1).
Proposition 5.1 A point .x0 is an accumulation point of S if and only if each

neighborhood of .x0 contains infinitely many points of S.
Proof We need to show that if a point .x0 is an accumulation point of S, then each
neighborhood of .x0 contains infinitely many points of S. The other implication is
clearly trivial. Suppose not: there exists a neighborhood U of .x0 that contains only
finitely many elements of S, say .s1 , .s2 , . . . , .sk . Pick a number .δ such that
.0 < δ < min{d(x0, sj ) | j = 1, 2, . . . , k}.
The set .(x0 − δ, x0 + δ) is an open neighborhood of .x0 that contains no points of

S different than .x0 , and this contradicts the assumption that .x0 is an accumulation
point of S.

5 Distance, Topology and Sequences on the Set of Real Numbers 49
Fig. 5.1 Intuition of accumulation points
Definition 5.6 The set of accumulation points of .S ⊂ R is denoted by .S % .1 The

closure .S of S is defined to be .S ∪ S % .
Example 5.2 Let .S = (0, 1). Each point of .[0, 1] is an accumulation point of S,
for if .x0 ∈ [0, 1], then each open interval centered at .x0 contains infinitely many
points of S. Furthermore, no other point of .R is an accumulation point of S. Indeed,
if .x0 ∈ (−∞, 0) ∪ (1, +∞), it is easy to construct an open neighborhood of .x0
which does not intersect S. Suppose for instance that .1 < x0 < +∞. Then the open
interval .(x0 − 12 (x0 − 1), x0 + 12 (x0 − 1)) is disjoint from S. A similar construction
applies to points of .(−∞, 0).
Theorem 5.1 Suppose S is a subset of .R. Then
(i) .S is closed;
(ii) .S = S if and only if S is closed
(iii) .S ⊂ F for each closed subset F of .R such that .S ⊂ F .
Proof
(i) Let .x ∈ R, .x ∈/ S. Then x is neither a point of S nor an accumulation point
of S. Therefore there exists a neighborhood of x that does not intersect S. This
proves that .S is open, i.e. .S is closed.
(ii) If .S = S, then S is closed by (i). On the other hand, if S is closed and x is
an accumulation point of S, then .x ∈ S. Indeed, supposing that .x ∈ / S would
lead to a neighborhood U of x such that .x ∈ U ⊂ S since .S is open. This
neighborhood would not contain any point of S at all, and x would not be an
1 The notation .DS is sometimes used instead of .S % .

accumulation point of S. We have proved that .S % ⊂ S if S is closed, and thus

.S = S.
(iii) If F is closed and .F ⊃ S, then .F ⊃ F % and thus .F ⊃ S % . Hence .F ⊃ S.

Theorem 5.2 Let S be a subset of .R that is bounded above. If .y = sup S, then
y ∈ S.
.
Proof If .y ∈ S, then .y ∈ S by the very definition. Assume .y ∈ / S, and we prove

that y is an accumulation point of S. For each .ε > 0 there exists a point .x ∈ S such
that .y − ε < x < y. Hence the arbitrary open neighborhood .(y − ε, y + ε) contains
a point x of S different than y, This proves that y is an accumulation point of S.
A similar result holds true for the infimum of a subset bounded below.
5.1 Sequences and Limits
Recall that a sequence in a set X is any function whose domain is .N \ F for some
finite subset F of .N and whose values lie in X (see Fig. 5.2).
Definition 5.7 Let .s = {sn }n be a sequence in .R. We say that s converges to a limit
L, and we write
L = lim sn ,
.
n→+∞
if for every neighborhood U of L there exists a positive integer .n0 such that .n > n0
implies .sn ∈ U .
Fig. 5.2 The graph of a sequence of real numbers

5.1 Sequences and Limits 51
Fig. 5.3 Intuition about uniqueness of limits
Equivalently, s converges to L if and only if for every .ε > 0 there exists a

positive integer .n0 such that .n > n0 implies .d(sn , L) < ε, i.e. .|sn − L| < ε. This
formulation is the most popular in undergraduate courses, but it does not work in
general topological spaces. Notice that we will also write .sn → L as .n → +∞
instead of .L = limn→+∞ sn .2
Exercise 5.2 Let .{sn }n be a sequence of real numbers. Prove that .sn → L as .n →
+∞ if and only if .|sn − L| → 0 as .n → +∞.
Exercise 5.3 Suppose that .{sn }n and .{tn }n are two sequences such that .sn = tn for
every .n ∈ N except for a finite number of values. Prove that .sn → L if and only if
.tn → L.
Theorem 5.3 (Uniqueness) If .sn → L1 and .sn → L2 as .n → +∞, then .L1 = L2 .

Proof We will use the following fact: if .L1 and .L2 are distinct points on .R there
exist a neighborhood .U1 of .L1 and a neighborhood .U2 of .L2 such that .U1 ∩ U2 = ∅.
Indeed, let .r = d(L1 , L2 ) = |L1 − L2 | > 0. The sets .U1 = (L1 − r/2, L1 + r/2)
and .U2 = (L2 − r/2, L2 + r/2) are clearly disjoint neighborhoods of .L1 and .L2 ,
respectively.
We therefore choose .U1 and .U2 in such a way, and by assumption there exist two
positive integers .n1 and .n2 such that .n > n1 implies .sn ∈ U1 , and .n > n2 implies
.sn ∈ U2 . Therefore .n > max{n1 , n2 } implies .sn ∈ U1 ∩ U2 = ∅, a contradiction.
We conclude that .L1 = L2 . Figure 5.3 should clarify our proof.

2Since we use s to name a sequence, it would be natural to write .s → L. This, albeit correct,
might be easily confused with the identical symbol that describes the fact that a variable s tends to
L independently, like in .lims→L f (s).
Theorem 5.4 (Squeezing Principle) Let .s = {sn }n be a sequence in .R. The

sequence s converges to the limit L if and only if .d(sn , L) = |sn − L| → 0 as
.n → +∞. In particular, if .{εn }n is a sequence of positive numbers that converges
to zero, and if .d(sn , L) < εn for each n, then .sn → L as .n → +∞.

Proof If s converges to L, for every .ε > 0 there exists a positive integer .n0 such
that .n > n0 implies .d(sn , L) < ε. Hence .d(sn , L) → 0 as .n → +∞. Reversing this
argument we prove the first statement. Suppose now that .εn → 0 as .n → +∞, and
that .d(sn , L) < εn for each n. For every .ε > 0 there exists a positive integer .n0 such
that .n > n0 implies .0 < εn < ε. Hence for any .n > n0 we get that .d(sn , L) < ε, or
.sn → L as .n → +∞.

Theorem 5.5 Any sequence that converges to a limit is bounded.
Proof Let .s = {sn }n be a real sequence such that .limn→+∞ sn = L. Consider the
neighborhood .(L − 1, L + 1) of L, and select a positive integer .n0 such that .n > n0
implies .sn ∈ (L − 1, L + 1). Now define .δ = min{d(L, sj ) | j = 1, 2, . . . , n0 }. The
union .U = (L − 1, L + 1) ∪ (L − δ, L + δ) is an open neighborhood of L such that
.sn ∈ U for each n. We have proved that the range of the sequence s is contained in
a bounded set, and the conclusion follows.

If .{sn }n and .{tn }n are two sequences of real numbers, we can form their sum .{sn +
tn }n and their product .{sn tn }n . These algebraic operations are stable with respect to
limits.
Theorem 5.6 Suppose that .sn → L and .tn → M as .n → +∞. Then, as .n →
+∞,
1. .sn + tn → L + M
2. .sn tn → LM
3. .sn /tn → L/M whenever .tn = 0 and .M = 0.
Proof Let .ε > 0. There exist positive integers .n1 and .n2 such that .n > n1 implies
|sn − L| < ε, and .n > n2 implies .|tn − M| < ε. If .n > max{n1 , n2 }, then
.
|(sn + tn ) − (L + M)| = |(sn − L) + (tn − M)| ≤ |sn − L| + |tn − M| < 2ε.

.
Since .ε > 0 is arbitrary, this proves 1. Similarly,
|sn tn − LM| = |sn tn − sn M + sn M − LM| ≤ |sn (tn − M)| + |M(sn − L)|.

.
The sequence .{sn }n is bounded by Theorem 5.5, thus there exists .C > 0 such that
|sn | < C for each n. Therefore
.
|sn tn − LM| ≤ |sn (tn − M)| + |M(sn − L)| ≤ C|tn − M| + |M||sn − L|

.
≤ Cε + |M|ε = (C + |M|) ε.
Since .C + |M| is independent of .ε and of n, the proof of 2. is complete. To prove

3. we proceed as follows. First of all, since .M = 0, there exists a positive integer .n3
such that .|tn | > |M|/2 > 0 for any .n > n3 . This follows from the definition of limit
with the neighborhood .(M − |M|/2, M + |M|/2). If .n > max{n1 , n2 , n3 },

sn
− L = sn M − Ltn = sn M − ML + ML − Ltn
.
t M tn M tn M
n

(sn − L)M + L(M − tn )
≤

tn M
|sn − L| |L| |tn − M|
≤ +
|tn | |M| |tn |
2 |L| 2
< ε+ ε
|M| |M| |M|
and 3. follows again from the arbitrariness of .ε > 0.

Theorem 5.7 (Monotone Sequences Have a Limit) Let .{sn }n be a bounded
increasing sequence of real numbers. Then .sn → supk sk as .n → +∞.
Proof Since .{sn }n is bounded, the quantity .L = supk sk is well defined in .R. If .ε >
0, then there exists a term .sn0 such that .L − ε < sn0 ≤ L. Since .{sn }n is increasing,
.n > n0 implies .sn > sn0 , so that .L − ε < sn ≤ L. This proves that .sn → L as
.n → +∞.

Exercise 5.4 Let .{sn }n be a bounded decreasing sequence of real numbers. Then
sn → infk sk as .n → +∞.
.
Monotone sequences in .R provide the following important result.

Theorem 5.8 (Nested Intervals) Suppose that for each positive integer n, .In =
[an , b
n ] is a closed interval, and suppose furthermore that .In ⊃ In+1 for each n.
Then . ∞ n=1 In = ∅.
Proof The sequence .{an }n is increasing and bounded above
(by .b1 , for instance).
Hence it converges to .L = supk ak . We claim that .L ∈ ∞ n=1 In . Indeed .L ≥ an
for each n, by definition of supremum. On the other hand, if n and p are positive
integers, then .an ≤ an+p ≤ bn+p ≤ bn . Letting .p → +∞ we get .an ≤ L ≤ bn , so
that .L ∈ [an , bn ] for each n. The proof is complete.

Exercise
∞ 5.5 Show that there exists a sequence .{In }n of nested intervals such that
. I
n=1 n contains infinitely many points.
Sometimes a more quantitative generalization is needed, as we are going to see.
Theorem 5.9 (Nested Intervals of Infinitesimal Length) Let .{In }n be a sequence
of closed intervals of .R such that (i) .In+1 ⊂ In for each n, and (ii) for every .ε > 0
there
∞ exists a positive integer .nε such that the length of .Inε is smaller than .ε. Then
. n=1 In is a singleton.

Proof We have already proved that . n∈N In = ∅. Let us suppose that two distinct
numbers .z < w belong to this intersection.
We may choose .ε = w − z > 0 in (ii)
and obtain a contradiction. Hence . n∈N In contains one and only one element.
Example 5.3 We prove that .R is uncountable as a consequence of Theorem 5.9. As
already noticed before, we need to prove that a closed interval .[a, b] is uncountable.
We suppose on the contrary that .[a, b] is countable. Of course .[a, b] contains
infinitely many elements, so we may suppose that .{xn | n = 1, 2, 3, . . .} is an
enumeration of .[a, b]. We are going to construct a number .z ∈ [a, b] which is
different than every term .xn of this enumeration. We divide .[a, b] into three intervals
of equal length, and we choose one of them, called .I1 , such that .x1 ∈ / I1 . If .In
has been chosen, we split it into three intervals of equal length, and we call .In+1
that interval which does not contain .xn+1 . The length of .In converges to zero as
.n → +∞, and by construction .In+1 ⊂ In for each n. Hence there exists a unique
real number z that belongs to every .In . In particular .z = xn for each n, since .z ∈ In
but .xn ∈
/ In . This contradiction proves the statement.
Theorem 5.10 Let S be a subset of .R, and let x be an accumulation point of S.
Then there exists a sequence .{sn }n of points of S such that .sn → x as .n → +∞.
Proof Consider the open neighborhood .U1 = (x − 1, x + 1) of x, and select a
point .s1 ∈ S such that .s1 ∈ U1 . This is possible because x is an accumulation
point of S. After point .s2 , .s3 , . . . , .sn−1 are selected in S, choose .sn in S such that
.sn ∈ Un = (x − 1/n, x + 1/n). Then .|sn − x| < 1/n for each n, and therefore
.sn → x as .n → +∞.

We record the following fact, which should be an easy exercise: a sequence .{sn }n
converges to a limit L if and only if every subsequence of .{sn }n converges to L.
Definition 5.8 A set .K ⊂ R is called sequentially compact, if every sequence in K
has a subsequence that converges to a point of K.
Theorem 5.11 (Bolzano-Weierstrass) Every bounded sequence in .R contains a
converging subsequence.
Proof Let .s = {sn }n be a bounded sequence of real numbers. For some .M > 0, this
means that .sn ∈ [−M, M] for each n. If the range of the sequence s is a finite set (in
the sense that s takes on only a finite number of different values), then there exists
infinitely many positive integers .n1 < n2 < n3 < . . . such that .sn1 = sn2 = sn3 =
. . .. Hence s contains a constant subsequence, which is clearly convergent to a point
of K. We may now assume that the range of s is an infinite set.
We set .I0 = [−M, M]. We divide .I0 into two parts: at least one of them must
contain infinitely many terms of the sequence, otherwise the range of s would be
finite. Let us call .I1 this part, and we choose the smallest positive integer .n1 such that
.sn−1 ∈ I1 . Now we split .I1 into two parts as before, we call .I2 the part that contains
infinitely many terms of the sequence s, and we choose the smallest positive integer
.n2 > n1 such that .sn2 ∈ I2 . Proceeding this way, we construct a subsequence .{snk }k
of S.
By construction, .I1 ⊃ I2 ⊃ I3 ⊃. . ., and by the principle of nested intervals we

know that there exists a point .x ∈ ∞ k=1 Ik . We claim that .snk → x as .k → +∞.
We fix a number .ε > 0. By construction the length of the interval .Ik is .m/2k−1 . If
the positive integer k is so large that .M/2k−1 < ε, then the length of .Ik is smaller
than .ε. Since .snk and x lie in .Ik , then .d(snk , x) < ε, and the proof is complete.

− 0
1
2
The Bolzano-Weierstrass Theorem can be proved along a different and delightful

strategy. We now show that any real-valued sequence has a monotone subsequence.
And since we already know that (bounded) monotone sequences converge. . .
Theorem 5.12 (Any Sequence Has a Monotone Subsequence) Every sequence
{sn }n of real numbers has a monotone subsequence.
.
Proof We provisionally agree that the n-th term of our sequence is dominant if
sm < sn for each .m > n. There are now two cases. It might happen that there exist
.
infinitely many dominant terms, and let .{snk }k be a subsequence consisting solely
of dominant terms. Then .snk+1 < snk for each k, and thus .{snk }k is monotonically
increasing.
The second case is that only finitely many terms are dominant. In particular there
exists a positive integer .n1 which is greater than any dominant term. There must
exist .n2 > n1 such that .sn2 ≥ sn1 . Suppose that .n3 , . . . , nk−1 have been selected so
that .n3 < n4 < . . . < nk−1 and .sn1 ≤ sn2 ≤ sn3 ≤ . . . ≤ snk−1 . As before, since
only finitely many terms are dominant, there exists .nk > nk−1 such that .snk ≥ snk−1 .
By induction we have constructed an increasing subsequence of .{sn }n .

Remark 5.1 Theorem 5.12 is an ingenious trick that essentially reduces the theory
of sequence in .R to the theory of monotone sequences. Unfortunately this trick relies
on the order property of the real numbers, and cannot be generalized to metric or
topological spaces.
We now prove the basic characterization of compact subsets of the real line.
Theorem 5.13 A set .K ⊂ R is sequentially compact if and only if it is closed and
bounded.
Proof Suppose K is sequentially compact. If K is not bounded, then there exists
x1 ∈ K such that .|x1 | > 1. Likewise, there exists .x2 ∈ K such that .|x2 | > 2.
.
In general, for each positive integer n there exists .xn ∈ K such that .|xn | > n.
From this we deduce that no converging subsequence of .{xn }n may exist. Indeed,
any converging subsequence .{xnk }k must be bounded by Theorem 5.5. But this
contradicts the fact that .|xnk | > k for each k. Hence K is bounded.
To prove that K is closed, we show that K contains its accumulation points. So,
let .{xn }n be a sequence in K that converges to a point x of .R. We have to show
that .x ∈ K. Since K is sequentially compact, there exists a subsequence .{xnk }k that
converges to a point .y ∈ K. As we have remarked above, there results .y = x since
.{xnk }k is a subsequence of .{xn }n . In particular .x ∈ K, and K is a closed set.
Conversely, suppose that K is closed and bounded, and let .{xn }n be a sequence
in K. Since K is bounded, by Bolzano-Weierstrass there exists a subsequence that
converges to a point x of .R. Since K is closed, .x ∈ K, and thus K is sequentially
compact. The proof is complete.

The definition of limit (Definition 5.7) has a big weakness: we can check if a
number is the limit of a sequence, but we need to have a good candidate at our
disposal. In the framework of real numbers with the Euclidean metric, the existence
of a limit can be ensured by a condition that only involves the terms of the sequence.
Definition 5.9 We say that a sequence .{sn }n of real numbers is a Cauchy sequence
if and only if for every .ε > 0 there exists a positive integer .n0 such that .n > n0 ,
.m > n0 imply .d(sn , sm ) < ε.
Proposition 5.2 Any converging sequence is a Cauchy sequence.

Proof If .{sn }n is a converging sequence and L is its limit, then for every .ε > 0 there
exists a positive integer .n0 such that .n > n0 implies .d(sn , L) < ε. Hence .n > n0
and .m > n0 imply by the triangle inequality .d(sn , sm ) ≤ d(sn , L) + d(L, sm ) < 2ε.

The surprising fact is that we can also reverse this statement: any Cauchy
sequence of real numbers must have a limit. The proof is however more delicate.
Proposition 5.3 Every Cauchy sequence is bounded.
Proof Recalling the definition with .ε = 1, we find a positive integer .n0 such that
.n > n0 and .m > n0 imply .d(sn , sm ) < 1. In particular .|sn | < |sn0 +1 | + 1 for each
.n > n0 . It follows that
M = max{|s1 |, |s2 |, . . . , |sn0 |, |sn0 +1 | + 1}

.
satisfies .sn ∈ [−M, M] for each positive integer n.

Theorem 5.14 (.R Is a Complete Metric Space) A real-valued sequence converges
if and only if it is a Cauchy sequence.
Proof We have already proved one half of the statement. Suppose now that .{sn }n is
a Cauchy sequence in .R. We know that it is bounded, and by Bolzano-Weierstrass
it has a converging subsequence .{snk }k : let L be its limit. We claim that the whole
sequence .{sn }n converges to L. Indeed, let .ε > 0, and choose a positive integer .n0
such that .n > n0 and .m > n0 imply .|sn − sm | < ε. Since .snk → L as .k → +∞, we
may select an index .k0 such that

. snk0 − L < ε.
For each .n > n0 we have .|sn − L| ≤ |sn − snk0 | + |snk0 − L| < 2ε, and the proof is
complete.

Exercise 5.6 Suppose that .{sn }n is a sequence of real numbers such that .|sn+1 −
sn | < 2−n for each n. Prove that .{sn }n is a Cauchy sequence. Is this result true if we
suppose that .|sn+1 − sn | < 1/n for each n?
We complete this discussion with a few words about divergent sequences.
Definition 5.10 A sequence .{sn }n is divergent if it is not convergent.
This definition is more appreciated if one thinks that sequences may be more
general than sequences of real numbers. In Calculus courses, the term divergent
is usually associated to sequences that “have an infinite limit”. When we think of
sequences of terms in a general set, the meaning of infinity becomes hard to define,
if not impossible at all.
Definition 5.11 Let .s = {sn }n be a sequence of real numbers. We say that s diverges
to .+∞ if for every .ε > 0 there exists a positive integer .n0 such that .n > n0 implies
.sn > 1/ε. In this case we write .limn→+∞ sn = +∞, or simply .sn → +∞ as
.n → +∞.
Similarly, we say that s diverges to .−∞ if for every .ε > 0 there exists a positive
integer .n0 such that .n > n0 implies .sn < −1/ε. In this case we write .limn→+∞ sn =
−∞, or simply .sn → −∞ as .n → +∞.
We remark that the symbols .+∞ and .−∞ do not represent elements of a numeric
set. We will see that we could extend the set .R by adding them in such a way the
previous definition becomes a particular case of a general definition of limit for
sequences.
Theorem 5.15 For a sequence .{sn }n of positive real numbers, we have
. lim sn = +∞
n→+∞
if and only if
1
. lim = 0.
n→+∞ sn
Proof Let us suppose that our sequence diverges to .+∞. Let .ε > 0 and .M = 1/ε.
By assumption there exists a positive integer N such that .n > N implies .sn > M.
Therefore .n > N implies .ε > 1/sn > 0. This proves that .1/sn converges to zero.
Conversely, let .M > 0 and .ε = 1/M. Since .1/sn → 0, there exists a positive
integer N such that .n > N implies .1/sn < ε, or .sn > M. This concludes the proof.

Exercise 5.7 Suppose that there exists a positive integer .N0 such that .sn ≤ tn for
each .n > N0 .
(a) Prove that if .limn→+∞ sn = +∞, then .limn→+∞ tn = +∞.
(b) Prove that if .limn→+∞ tn = −∞, then .limn→+∞ sn = −∞.
(c) Prove that if .limn→+∞ sn and .limn→+∞ tn exist, then .limn→+∞ sn ≤
limn→+∞ tn .

Exercise 5.8 Suppose that .sn = 0 for each n, and that .L = limn→+∞ sn+1
sn
exists.
(a) Show that if .L < 1 then .limn→+∞ sn = 0. Hint: fix a number a such that
.L < a < 1, and obtain a positive integer N such that .|sn+1 | < a|sn | for each
.n > N. Deduce that .|sn | < a

n−N |s | for each .n > N.
N
(b) Show that if .L > 1, then .limn→+∞ |sn | = +∞. Hint: apply (a) to the sequence
.tn = 1/|sn |.
5.2 A Few Fundamental Limits
We collect a few statements that follow from elementary estimates based on the
Binomial Theorem. The reader is invited to appreciate the proofs.
Proposition 5.4 If .p > 0 is a real number, then
1
. lim = 0.
n→+∞ np
Proof Given .ε > 0, just take .n > (1/ε)1/p . This is possible by the Archimedean
property of .R.

Proposition 5.5 If .p > 0 is a real number, then
√
. lim n
p = 1.
n→+∞
√
Proof If .p > 1, put .xn = n p−1. Clearly .xn > 0 and Theorem 3.8 yields .1+nxn ≤
(1+xn )n = p. Hence .0 < xn ≤ (p−1)/n, and the conclusion follows by squeezing.
If .p = 1 the conclusion is trivial. If .0 1 and the claim
is reconducted to the previous case.

5.3 Lower and Upper Limits 59
Proposition 5.6
√
. lim n
n = 1.
n→+∞
√
Proof Let .xn = n
n − 1, so that .xn ≥ 0. By Theorem 3.8
n(n − 1) 2
n = (1 + xn )n ≥
. xn .
2
Hence

2
0 ≤ xn ≤
.
n−1
for .n ≥ 2. We conclude by squeezing.

Proposition 5.7 If .p > 0 and .α are real numbers, then
nα
. lim = 0.
n→+∞ (1 + p)n
Proof Fix a positive integer .k > α. For .n > 2k
n k n(n − 1) · · · (n − k + 1) k nk pk
(1 + p)n >
. p = p > k .
k k! 2 k!
Hence
nα 2k k!
0<
. < k nα−k
(1 + p) n p
for .n > 2k. Since .α − k < 0, we conclude by squeezing and Proposition 5.4.

Proposition 5.8 If x is a real number and .−1 < x < 1, then
. lim x n = 0.
n→+∞
Proof Just take .α = 0 in Proposition 5.7.

5.3 Lower and Upper Limits
We now try to describe the loss of convergence for real sequences. The question is:
why can a sequence be divergent?
Definition 5.12 Let .{an }n be a sequence of real numbers. If it is not bounded above,
we declare that
. lim sup an = +∞.

n→+∞
Similarly, if it is not bounded below, we declare that
. lim inf an = −∞.

n→+∞
We say that a number M is an eventual upper bound [resp. lower bound] for the
sequence if there exists a positive integer .ν such that .an ≤ M [resp. .an ≥ M]
for each .n ≥ ν. The limsup of the sequence .{an }n is the infimum of the set .M of
eventual upper bounds:
. lim sup an = inf M.

n→+∞
In a similar fashion, the liminf of .{an }n is the supremum of the set .N of eventual
lower bounds:
. lim inf an = sup N.

n→+∞
Plainly .lim infn→+∞ an ≤ lim supn→+∞ an in any case. The inequality can be
strict: if .an = (−1)n , then .lim infn→+∞ an = −1 and .lim supn→+∞ an = 1.
Theorem 5.16 The sequence .{an }n converges to a limit L if and only if
. lim inf an = L = lim sup an .

n→+∞ n→+∞
Proof The cases .L ∈ {−∞, +∞} are clear by the initial definition. We now
focus on the case .L ∈ R. Assume that .an → L: given .ε > 0, there exists
a positive integer .ν such that .L − ε < an < L + ε for each .n > ν. Hence
.L + ε ∈ M, and .lim supn→+∞ an ≤ L + ε by definition of infimum. Similarly
.L − ε ≤ lim infn→+∞ an . Since .ε > 0 is arbitrary, this yields .lim infn→+∞ an =
lim supn→+∞ an = L.
Viceversa, assume that .lim infn→+∞ an = L = lim supn→+∞ an . Fix any .ε > 0.
We know that .L + ε ∈ M and .L − ε ∈ N. Hence there exists a positive integer .ν
such that for each .n > ν we must have .L − ε ≤ an ≤ L + ε. Therefore .an → L,
and the proof is complete.

Exercise 5.9 Prove that the sequence .{(−1)n }n does not converge.
We provide a useful characterization of liminf and limsup. Sometimes this is
taken as a definition.
5.3 Lower and Upper Limits 61
Theorem 5.17 For a bounded sequence .{an }n ,
. lim sup an = lim sup ak = inf sup ak

n→+∞ n→+∞ k≥n n∈N k≥n
lim inf an = lim inf ak = inf inf ak .

n→+∞ n→+∞ k≥n n∈N k≥n
Proof The two statements are similar, and we prove the second one. We
set .λn = infk≥n ak , and remark that .λn+1 = inf{an+1 , an+2 , . . .} ≥
inf{an , an+1 , an+2 , . . .} = λn . The sequence .n → λn is thus increasing.
Furthermore, .ak ≥ λn for each .k ≥ n, so that .λn is an eventual lower bound.
By definition, .lim infn→+∞ an ≥ λn and finally .lim infn→+∞ an ≥ supn λn . We
need to prove the opposite inequality.
Fix any eventual lower bound .: there exists a positive integer .ν such that .ak ≥
for each .k ≥ ν. Hence . ≤ infk≥ν ak = λν ≤ supn λn . The element . ∈ N is
arbitrary, and so .lim infn→+∞ an ≤ supn λn .

Theorem 5.18 (Monotonicity of liminf and limsup) If .an ≤ bn for each n, then
lim infn→+∞ an ≤ lim infn→+∞ bn and .lim supn→+∞ an ≤ lim infn→+∞ bn .
.
Proof If .{an }n is unbounded below, or if .{bn }n is unbounded above, the conclusion

is trivial. Let M be an eventual upper bound for .{bn }n . There exists a positive integer
.ν such that .n ≥ ν implies .bn ≤ M. Then .an ≤ M for the same indices n, and thus
.lim supn→+∞ an ≤ M. But M is arbitrary, so that .lim supn→+∞ an ≤ inf M =
lim supn→+∞ bn . The other statement is similar.

The following result is often used in the theory of numerical series.
Theorem 5.19 Suppose .{an }n is a sequence of real numbers. There results

an+1 an+1
. lim inf ≤ lim inf |an | 1/n
≤ lim sup |an | 1/n
≤ lim sup .
n→+∞ an n→+∞ n→+∞ n→+∞ an

In particular, if .limn→+∞ an+1
an exists and equals L, then also .limn→+∞ |an |
1/n
exists and equals L.

Proof We prove the last inequality, and leave the first as an exercise. The
middle

inequality is clear. Put .α = lim supn→+∞ |an |1/n and .L = lim supn→+∞ an+1
an . If
.L = +∞, the proof is complete, so we may assume that .L ∈ R. Furthermore, it
suffices to show that .α ≤ L + ε for each .ε > 0. By definition,

an+1
L= lim sup n > N < L + ε,
an
.
N→+∞
there exists a positive integer N such that

an+1
sup n ≥ N < L + ε,
a
.
n
and thus

an+1

.
a <L+ε
n
for each .n ≥ N. For .n > N we can write

an an−1 aN+1
.|an | =
a · a ··· a .
n−1 n−2 N
Hence
|an | < (L + ε)n−N |aN |

.
for each .n > N. It follows that for .n > N we have

n−N
|an |1/n < (L + ε)
. n |aN |1/n ,
and letting .n → +∞ we see that .α ≤ L + ε.

5.4 Problems
5.1 Let {sn }n be a set of real numbers. The arithmetic mean σn are defined by
s0 + s1 + · · · + sn
σn =
. .
n+1
1. If limn→+∞ sn = s ∈ R, prove that limn→+∞ σn = s.

2. Construct a sequence {sn }n which does not converge, but such that there results
limn→+∞ σn = 0.
3. Define an = sn − sn−1 for each n ≥ 1. Prove that
1
n
.sn − σn = kak .
n+1
k=1
Suppose that limn→+∞ nan = 0 and that {σn }n converges. Prove that {sn }n
converges.
5.4 Problems 63
4. Prove the same statement as in 3. under the weaker assumption that {nan }n
is bounded. As a hint, you may use the following approach. Suppose that
limn→+∞ σn = σ and that |nan | ≤ M for each n. If m < n, then
m+1 1
n

sn − σn =
. (σn − σm ) + sj − σj .
n−m n−m
j =m+1
For each j ∈ {m + 1, . . . , n},
(n − j )M (n − m − 1)M
. |sj − σj | ≤ ≤ .
j +1 m+2
For each ε > 0 and each positive integer n, let m be the positive integer such that
n−ε
m≤
. < m + 1.
1+ε
Then (m + 1)/(n − m) ≤ 1/ε and |sn − sj | < Mε. Hence
. lim sup |sn − σ | ≤ Mε.

n→+∞
√
5.2 Let b > 0 be a given real number. Choose any real number x1 > b, and
define recursively
1 b
xn+1 =
. xn + .
2 xn
√
√ {xn }n is decreasing, and that limn→+∞ xn =
1. Prove that the sequence b.
2. Define εn = xn − b, and prove that
εn2 ε2
. εn+1 = < √n .
2xn 2 b
√
Setting β = 2 b, deduce that
2n
ε1
εn+1 < β
.
β
for n = 1, 2, 3, . . .
This problem describes a numerical algorithm for computing the square root of a
given number.
√
5.3 Let b > 1 be a given real number. Choose any real number x1 > b, and
define recursively
b + xn b − xn2
xn+1 =
. = xn + .
1 + xn 1 + xn
1. Prove that x1 > x3 > x5 > . . .

2. Prove that x2 < x4 < x6 <√...
3. Prove that limn→+∞ xn = b.
5.4 Compute the upper and the lower limit of the sequence {sn }n defined recursively
by
.s1 = 0
s2n−1
s2n =
2
1
s2n+1 = + s2n .
2
5.5 Evaluate

n
1
. lim √ .
n→+∞
k=1 n2 + k
√
Hint: prove that the sum is smaller than 1 and larger than n/ n2 + n.
5.6 We define the sequence of Fibonacci numbers recursively by u0 = 0, u1 = 1
and
un+2 = un + un+1 .
.
Set xn = un+1 /un for each n ∈ N.

(a) Prove that x1 < x3 < x5 < . . . < x6 < x4 < x2 .
(b) Prove that limn→+∞ (x2n − x2n−1 ) = 0.
(c) Compute limn→+∞ xn .
(d) If α and β are the roots of the polynomial x 2 = x + 1 and if wn = aα n + bβ n ,
prove that the sequence {wn }n satisfies wn+2 = wn+1 + wn for each n ∈ N.
(e) Deduce an explicit expression of un in terms of n.
(f) Use the result of (e) to solve (c).
Reference 65
5.5 Comments
Elementary Calculus can be introduced in a way that hides the topological (metric)
structure of the set .R. This approach, in my opinion, is too radical: the passage from
analysis in .R to analysis in .RN , .N ≥ 2, requires arguments that I prefer to introduce
from the first time. The topology of the real line coincides with the order topology
induced by the usual ordering .≤, but this is false in higher dimension. The fact that
intervals are basic examples of: open sets, convex sets, connected sets, relatively
compact sets should be seen both as a positive and a negative feature of .R.
The book [1] is a wonderful example of modern treatment of Calculus and
Analysis based on topological tools. In my opinion, this book remains a masterwork
in its field, although students may need to work hard before appreciating it.
Reference

Chapter 6
Series
Abstract Series are just a special type of sequences. The main feature of numerical
series is that they lead us to finding convergence theorems which do not involve the
value of the limit.
If .a = {an }n is a sequence of real numbers, we use the symbol

q
. an
n=p
to denote the finite sum .ap + ap+1 + · · · + aq−1 + aq . We use the sequence a to
construct a new sequence .s = {sn }n by means of the formula

n
sn =
. ak .
k=1
The sequence s is called the sequence of partial sums of a. It is customary to

introduce a different notation for the sequence s:
∞

s=
. an .
n=1
In a really formal world, a series should be defined asan ordered couple .(a, s)
such that a is a sequence, s is a sequence, and .sn = nk=1 ak for each .n ∈ N.
Remark 6.1 The language about series is very unprecise. In a completely rigorous
world, we should probably remove the word series and continue to use the word

68 6 Series
sequence, as in

n k
consider the sequence . k=1 k 2 +1 n
.
N
Furthermore, several mathematicians interpret . ∞ n=1 an as .limN→+∞ n=1 an ,
which is either a real number of a symbol of infinity. Despite these difficulties,
tradition rules, and in this chapter we will freely abuse of language and define a
series with the symbol . n an .

Definition 6.1 We say that the series . ∞ n=1 an converges to s if .limn→+∞ sn = s.
In this case, we will often say that s is the sum of the series.
Remark 6.2 It should be clear that sequences and series are the same object.
Indeed, series are sequences by definition. Conversely, the sequence .{an }n can be
recovered from the sequence .{sn }n by writing .an = sn − sn−1 . Of course this logical
equivalence is not a good reason to forget about numerical series at all.

We will often write . n an or even . an to denote a series. We agree that the
first index of the sum may also be different than 1, as in . ∞ n=7 an . Clearly, the
convergence of a series does not depend on the first terms that we add or discard:
remember that the character of a sequence is not altered by the modification of
finitely many terms.
Example 6.1 Let us consider the series
∞
1
. .
n(n − 1)
n=2
Since
1 1 1
. = − ,
n(n − 1) n−1 n
we see that

n
1 1
sn =
. =1− →1
k(k − 1) n
k=2
as .n → +∞. Hence the series converges to the sum 1.

Example 6.2 The previous example can be easily generalized. Suppose that we are
given the series
∞

. (bn+1 − bn ) ,
n=1
6 Series 69
where .{bn }n is a sequence such that .limn→+∞ bn = b. Then

∞

. (bn − bn+1 ) = b1 − b.
n=1
These are called telescoping series.

Since a closed formula for the partial sums of a sequence is usually unavailable,
the whole theory of convergence must be based on some indirect approach. A very
general one is the Cauchy characterization of convergence.

Theorem 6.1 (Cauchy for Series) A series . an converges if and only if for every
.ε > 0 there exists a positive integer N such that
m

. ak < ε

k=n
for any .m ≥ n > N.

Proof Since
m

. ak = |sm − sn−1 | ,

k=n
the conclusion follows from Theorem 5.14.

Corollary 6.1 (Necessary Condition for Convergence) If . an converges, then
limn→+∞ an = 0.
.
Proof We take .m = n in the previous theorem.

Remark 6.3 We will see that this corollary cannot be reversed. For instance the
harmonic series . ∞
n=1 n diverges, although .1/n → 0 as .n → +∞.
1
The necessary condition for convergence confirms an intuitive fact: you

cannot sum infinitely many numbers and obtain a finite result, unless the
numbers you add get smaller and smaller. As usual, intuitive results in
mathematics are weak results.

Theorem 6.2 Suppose that .an ≥ 0 for each n. The series . an converges if and
only if its partial sums form a bounded sequence.
Proof For a series of non-negative terms, we clearly have
sn+1 = sn + an+1 ≥ sn
.
70 6 Series
for every n. In other words, the sequence of partial sums is increasing. The
conclusion follows from Theorem 5.7.

The most important test of convergence is based on comparison. We will see that
actually all convergence tests are based on some comparison argument.
Theorem 6.3 (Comparison Test)

(a) If .|an | ≤ cn for
.n ≥ N0 , where .N0 is some fixed positive integer, and if . cn
converges, then . an converges aswell.
(b) If .an ≥ dn ≥ 0 for .n ≥ N0 , and if . dn diverges, then . an diverges as well.
Proof
(a) Given .ε > 0, there exists a positive .nm0 ≥ N0 such
m integer mthat .m ≥ n > n0
implies m
c ≤ ε. Hence a ≤ |a | ≤
. k=n k . k=n k k=n k k=n ck < ε.
(b) If . an converges, by (a) . dn converges. Contradiction.

An important corollary is described in the next result.

Theorem 6.4 (Asymptotic Comparison Test) Let . an and . bn be series of
positive terms, and suppose that
an
. lim = 1.
bn
n→+∞

The series . an converges if and only if . bn converges.
Proof Indeed, there exists a positive integer .N0 such that .1/2 < an /bn < 3/2 for
every .n > N0 . Hence . b2n < an < 32 bn for .n > N0 . The conclusion follows from the
Comparison test.

Example 6.3 The series
∞
1
.
n2
n=1
converges. Indeed,
1 1
. ≤
n2 n(n − 1)
for .n = 2, 3, . . . We conclude by comparison with Example 6.1.

Remark 6.4 If .an = 1/n2 and .bn = n(n−1)1
, we have .limn→+∞ an /bn = 1. This

shows a typical application of the asymptotic comparison test to the series . ∞ 1
n=1 n2 ,
which often requires less care in checking the validity of the comparison.
6.1 Convergence Tests for Positive Series 71
The triangle inequality always ensures that

m
m

. ak ≤ |ak |, (6.1)

k=n k=n
leading us to the following definition via the Cauchy condition for convergence.

Definition 6.2 (Absolute Convergence) We say that the series . an converges
absolutely, if the series . |an | is convergent.
An easy but not trivial consequence of (6.1) is the next result.
Theorem 6.5 Every absolutely convergent series is convergent.

Proof Let . an be an absolutely convergent series. By (6.1), the series . an
satisfies the Cauchy condition, and is therefore convergent.

The converse is false, as Exercise 6.4 shows.
6.1 Convergence Tests for Positive Series
Theorem 6.2 says that series of positive terms are somehow easier to deal with, since
no oscillation phenomenon can arise. In this section we develop several convergence
tests for positive series, i.e. series of positive terms.
Important: Negative Series

Of course the very same tests can be applied to series of negative terms, just by
changing signs to each term. For the sake of definiteness, we will always deal with
positive series.
Let us start with a milestone of the theory.

Theorem 6.6 (Geometric Series) If .0 ≤ x < 1, then
∞
1
. xn = .
1−x
n=0

If .x ≥ 1, the series . ∞ n
n=0 x diverges.
n
Proof If .x = 1, then . k=0 1k = n + 1, and the series diverges. Suppose .x = 1,
and compute

n
1 − x n+1
. xk = .
1−x
k=0
72 6 Series
Indeed
(1 − x)(1 + x + x 2 + · · · + x n ) = 1 + x − x + x 2 − x 2 + · · · + x n − x n − x n+1
.
= 1 − x n+1 .
The conclusion follows by letting .n → +∞.

Exercise 6.1 Prove the identity
(1 − x)(1 + x + x 2 + · · · + x n ) = 1 − x n+1
.
by induction.
The following test is usually a difficult one for students. It states a rather
surprising fact: under a monotonicity assumption, only those terms of a very
particular subsequence decide whether a series converges.
Theorem 6.7 (Condensation Test) Suppose that .a1 ≥ a2 ≥ a3 ≥ . . . ≥ 0. The
series . ∞ ∞ k
n=1 an is convergent if and only if the series . k=0 2 a2k is convergent.
Proof It suffices to prove that the partial sums of the two series are simultaneously
bounded from above. Set
sn = a1 + · · · + an
.
tk = 20 a20 + 21 a21 + · · · + 2k a2k .
We consider two cases. If .n < 2k , then
sn ≤ a1 + (a2 + a3 ) + · · · + (a2k + · · · + a2k+1 −1 )

.
≤ a1 + 2a2 + · · · + 2k a2k
= tk
by the monotonicity of .{an }n . Notice that we have grouped terms in blocks that
begin with a power of 2 and end one step before the subsequent power of 2. We
deduce that .sn ≤ tk .
On the other hand, if .2k < n, we group terms in a different way:
. sn ≥ a1 + a2 + (a3 + a4 ) + · · · + (a2k−1 +1 + · · · + a2k )

1
≥ a1 + a2 + 2a4 + · · · + 2k−1 a2k
2
1
= tk .
2
6.1 Convergence Tests for Positive Series 73
In this case, .tk ≤ 2sn . In any case the sequences .{sn }n and .{tk }k are both bounded or
unbounded above, and the proof is complete.

Example 6.4 As a fundamental application, we consider the generalized harmonic
series
∞
1
. ,
np
n=1
where p is a fixed real number. Clearly .p ≤ 0 implies divergence of the series, since
the general term does not converge to zero. For .p > 0 we use the condensation test,
and look at the series
∞
∞
1
. 2k k p
= 2(1−p)k .
(2 )
k=0 k=0
This is a geometric series, and we know that the latter series converges if and only
if .21−p < 1, i.e. .p > 1.
We propose the following tests for historical reasons. They are based on a
comparison with a geometric series, and we will comment on the weakness of these
tests after the proof.

Theorem 6.8 (Root and Ratio Tests) The series . an
√
(a) converges, if .lim supn→+∞√n |an | < 1;
(b) diverges, if .lim supn→+∞ n|an | > 1;

(c) converges, if .lim supn→+∞ an+1 an < 1;

(d) diverges, if . an+1
an ≥ 1 for each .n ≥ n0 , where .n0 is some fixed positive integer.
√
Proof Put .α = lim supn→+∞ n |an |. If√.α < 1, we can choose .β such that .α < β <
1, and a positive integer N such that . n |an | < β for each .n ≥ N. Hence .n ≥ N
implies .|an | < β n√. Since .β < 1, the comparison test leads to (a).
If .α > 1, then . n |an | > 1 for infinitely many indices n (otherwise 1 would be an
eventualupper bound). This prevents .an from converging to 0 as .n → +∞, and the
series . an is divergent. This proves
(b).

Suppose that .lim supn→+∞ an+1 an < 1: we can find .β < 1 and a positive integer

N such that . an+1
an < β for each .n ≥ N. In particular
|aN+1 | < β|aN |

.
|aN+2 | < β|aN+1 | < β 2 |aN |

..
.
|aN+p | < β p |aN |
74 6 Series
for each positive integer p. Writing .n = N + p we discover that
|an | < |aN |β −N · β n

.
for each .n ≥ N. Again (c) follows from the comparison theorem. Finally, if .|an+1 | ≥
|an | for .n ≥ n0 , then the condition .an → 0 fails, and the series . an is divergent.

1
The root and the ration tests are popular but weak. We know that the series . n

diverges while . n12 converges. The ratio and the root tests are both inconclusive,
since the limsup equals 1.
Remark 6.5 It follows from Theorem 5.19 that the root test is stronger than the ratio
test. In particular, if the root test is inconclusive, the ratio test must be inconclusive
as well.

Example 6.5 Consider the series . n n2n+3 . If we put .an = n2n+3 , there results
an+1 n + 1 n2 + 3
. = .
an n n2 + 2n + 4
√
We deduce that .limn→+∞ |an+1 /an | = 1. Similarly, .limn→+∞ n |an | = 1. The
root test and the ratio test are inconclusive, although the series is divergent by
comparison:
n 1
an ≥
. = .
n2 + 3n2 4n
Once more, we remark that a clever direct comparison is often preferable to a

standard test.

Exercise 6.2 Prove that if a series . n an of nonnegative numbers converges, then
p
the series . n an converges for every real number .p > 1. Hint: the inequality .an < 1
must hold eventually.

Exercise 6.3 Prove that . n√an and . n bn are convergent series
√ of nonnegative
numbers, then the series . n an bn converges. Hint: prove that . an bn ≤ an + bn .
6.2 Euler’s Number as the Sum of a Series
The typical Calculus approach to the definition of the number e is via the
“fundamental limit”
n
1
. lim 1+ .
n→+∞ n
6.2 Euler’s Number as the Sum of a Series 75
Unfortunately the existence of this limit is not straightforward. In the next theorem
we propose a different approach.

Theorem 6.9 (The Euler Number) The series . ∞ 1
n=0 n! converges to a
limit that is
n
denoted by e and called the Euler number. Furthermore, .e = limn→+∞ 1 + n1 .
Proof Recall that .0! = 1 and, for any positive integer n, the factorial of n is defined
as .n! = 1 · 2 · · · (n − 1)n. Since
1 1 1 1
sn = 1 +
. + + + ··· +
1 1·2 1·2·3 1 · 2···n
1 1 1
< 1 + 1 + + 2 + · · · + n−1 < 3,
2 2 2

the series . ∞ 1
n=0 n! converges to a limit .e < 3. To prove the second part, we
introduce the sequences

n
1 1 n
sn =
. , tn = 1 + .
k! n
k=0
The binomial formula

n
n n−k k
n
n!
(a + b)n =
. a b = a n−k bk
k k!(n − k)!
k=0 k=0
yields
1 1 1 1 2
tn = 1 + 1 +
. 1− + 1− 1− + ···
2! n 3! n n
1 1 2 n−1
+ 1− 1− ··· 1 − .
n! n n n
Then .tn ≤ sn and .lim supn→+∞ tn ≤ e. If .n ≥ m,
1 1 1 1 m−1
tn ≥ 1 + 1 +
. 1− + ··· + 1− ··· 1 − ,
2! n m! n n
so that .sm ≤ lim infn→+∞ tn for any m. Letting .m → +∞, .e ≤ lim infn→+∞ tn ,

∞ 1
The definition .e = n=0 n! is rather flexible, and allows us to derive a theoretical
property of the Euler number.
Theorem 6.10 The number e is irrational.
76 6 Series

Proof We begin with an estimate of the convergence of the series . 1/n! to e.
Letting .sn denote the n-th partial sum of this series, we have
1 1 1
e − sn =
. + + + ···
(n + 1)! (n + 2)! (n + 3)!
1 1 1
< 1+ + + ···
(n + 1)! n + 1 (n + 1)2
1
= .
n!n
1
Therefore .0 < e−sn < n!n for each positive integer n. Now suppose that .e = p/q is
a rational number, where p and q are positive integers. Then .0 < q!(e − sq ) < 1/q.
The number .q!e must be an integer, since e is rational. Also
1 1
q!sq = q! 1 + 1 +
. + ···+
2! q
is an integer. Hence .q!(e − sq ) is an integer between 0 and 1: contradiction. The

number e is therefore irrational.

6.3 Alternating Series
The reader should suspect that a complete analysis of series whose terms do not have
constant sign is out of reach. In this section we focus our attention on a particular
class of series of variable sign. We begin with a general result which reminds us of
the popular formula of integration by parts.
by Parts) Two sequences .{an }n and .{bn }n are given.

Proposition 6.1 (Summation
Put .A−1 = 0 and .An = nk=0 ak for .n ≥ 0. For each positive integers .p ≤ q we
have

p
q−1
. an bn = An (bn − bn+1 ) + Aq bq − Ap−1 bp .
n=p n=p
Proof Since .an = An − An−1 , we write

p
q
q
q−1
. an bn = (An − An−1 ) bn = An bn − An bn+1 .
n=p n=p n=p n=p−1
6.3 Alternating Series 77

The last difference is equal to . q−1
n=p An (bn − bn+1 ) + Aq bq − Ap−1 bp , and the
proof is complete.

Theorem 6.11 (Dirichlet’s Test) Suppose

(a) the partial sums .An of . an form a bounded sequence;
(b) .b0 ≥ b1 ≥ b2 ≥ . . .;
(c) .limn→+∞ bn = 0.

Then the series . an bn is convergent.
Proof There exists .M > 0 such that .|An | ≤ M for each n. Let .ε > 0, and
pick a positive integer .ν such that .bν ≤ ε/(2M). For .ν ≤ p ≤ q we have by
Proposition 6.1
q q−1

. an bn ≤ An (bn − bn+1 ) + Aq bq − Ap−1 bp
n=p n=p

q−1

≤ M (bn − bn+1 ) + bq + bp
n=p
= 2Mbp ≤ 2Mbν ≤ ε.

The series . an bn converges by the Cauchy theorem.

Choosing .an = (−1)n+1 and .bn = |cn | in the previous theorem yields a popular
test for alternating series.
Theorem 6.12 (Leibnitz Theorem for Alternating Series) Suppose that
(a) .|c1 | ≥ |c2 | ≥ |c3 | ≥ . . .
(b) .c2m−1 ≥ 0, .c2m ≤ 0 for .m = 1, 2, 3, . . .
(c) .limn→+∞ cn = 0

Then the series . cn is convergent.

Exercise 6.4 Prove that the series . ∞ (−1)n
n=1 n converges, but it does not converge
absolutely. This fact often seems to be surprising, but we must remember that the
factor .(−1)n contributes to a huge balancing of the terms in the series.
6.3.1 Product of Series
Numerical series can be multiplied together. The definition is reminiscent of the

product of two polynomials .p(x) and .q(x), in which terms are grouped according
to the power of the unknown x.
78 6 Series
Definition
6.3 (Cauchy Product
of Two Series) The Cauchy product of the series
. an and . bn is the series . cn defined by

n
cn =
. ak bn−k .
k=0
Remark 6.6 Properly speaking, the Cauchy product of two series is a discrete
convolution product. Since we do not assume the reader to be familiar with integral
convolutions, we will not use this language in the book.
The convergence of a product of two series is a delicate issue. Consider for
example the series
∞
(−1)n 1 1 1
. √ = 1 − √ + √ − √ + ···
n=0
n+1 2 3 4
Convergence follows from Theorem 6.12. Let us now multiply this series by itself,
obtaining
∞
1 1 1 1 1
. cn = 1 − √ +√ + √ +√ √ +√ + ···
n=0
2 2 3 2 2 3
∞

n
1
= (−1)n √ .
n=0 k=0
(n − k + 1)(k + 1)
But
n 2 n 2 n 2
(n − k + 1)(k + 1) =
. +1 − −k ≤ +1 ,
2 2 2
and

n
2 2(n + 1)
|cn | ≥
. = .
n+2 n+2
k=0

Since the necessary condition .cn → 0 is violated, the series . cn must diverge.
Here comes the basic convergence result about the product of convergent series.
Theorem 6.13 (Mertens) Suppose that

(a) .∞ n=0 an converges absolutely
(b) .∞ n=0 an = A
(c) . ∞ n=0bn = B
(d) .cn = nk=0 ak bn−k .
6.4 Problems 79

Then . ∞
n=0 cn converges.
Proof We follow [1], and set

n
An =
. ak
k=0

n
Bn = bk
k=0

n
Cn = ck
k=0
βn = Bn − B.
We compute
Cn = a0 b0 + (a0 b1 + a1 b0 ) + · · · + (a0 bn + a1 bn−1 + · · · + an b0 )

.
= a0 Bn + a1 Bn−1 + · · · + an B0
= a0 (B + βn ) + · · · + an (B + β0 )
= An B + a0 βn + a1 βn−1 + · · · + an β0 .
To conclude the proof, we must show that .limn→+∞ γn = 0, where .γn = a0 βn +

a1 βn−1 + · · · + an β0 . Let .α = ∞
n=0 |an |. Notice that this is the first time we invoke
assumption (a). Given any .ε > 0, we can choose a positive integer .ν such that
.|βn | ≤ ε for each .n ≥ ν. Thus
|γn | ≤ |β0 an + · · · + βν an−ν | + |βν+1 an−ν−1 + · · · + βn a0 |

.
≤ |β0 an + · · · + βn an−ν | + εα.
Since .lim supn→+∞ (β0 an + · · · + βn an−ν ) = 0, we find .lim supn→+∞ |γn | ≤ εα,
and the conclusion follows.

6.4 Problems
6.1 Decide whether the series

∞

. sin(α) sin(2α) · · · sin(nα)
n=1
is convergent, for any fixed value of α ∈ R.

80 6 Series
6.2 Let {an }n be a sequence with the property that there exists a real number h < 1
such that |an+1 − an | ≤ h|an − an−1 | for each n. Prove that the sequence converges.
6.3 Using the previous problem, show that the sequence defined by choosing any
two real numbers a1 and a2 , and defining
an−1 + an
an+1 =
.
2
converges. Compute its limit.
∞
6.4 Let {an }n be a sequence of positive
realannumbers. Prove that the series n=1 an
converges if and only if the series ∞ n=1 1+an converges.
6.5 Starting from
∞
1
. = xn
1−x
n=1
and using Cauchy products, prove that
∞
1
. = nx n−1
(1 − x) 2
n=1
for each real number x with |x| < 1.
6.5 Comments
Once upon a time, the treatment of numerical series used to fill up long chapters in
Calculus textbooks. As I have tried to show, the theory of series is indeed a long
collection of sufficient conditions for the convergence of particular sequences of
numbers. In recent years this awareness has become prevalent, and we no longer
annoy our students with awful convergence tests. Last but not least, many of these
tests are based on the algebraic properties of real numbers, and they do not extend
to series of complex numbers, for instance.
Reference

Chapter 7
Limits: From Sequences to Functions
of a Real Variable
Abstract From a really abstract point of view, the whole theory of limits for
functions of a real variable is an immediate consequence of the theory of limits
for sequences. Furthermore, the definition of limit for a real-valued function can
be easily reduced to a continuity request. In this chapter we introduce limits via
sequential limits, and prove a standard characterization in terms of neighborhoods.
Definition 7.1 Let E be a nonempty subset of .R, and let p be an accumulation

point of E. We say that a number .L ∈ R is the limit of .f (x) as .x → p, if for every
sequence .{xn }n of points of E such that
(xn → p) ∧ (∀n)(xn = p),

.
one has that .f (xn ) → L as .n → +∞. In this case we write
f (x) → L as x → p,
.
or
. lim f (x) = L.
x→p
Remark 7.1 Since .p ∈ E % , at least one sequence of points .xn from E that converge
to p while being always different than p exists. We point out that L has nothing to
do with the value .f (p), and indeed p need not even belong to the domain E of f .1
Figure 7.1 shows a typical example of a function with a finite limit. We now
introduce two more possible definitions of limits, and we finally prove that these
definitions are indeed equivalent.
1 We follow here the traditional definition of limit. In the recent French tradition, the condition
.xn = p is omitted. We will come back to this discussion later.

82 7 Limits: From Sequences to Functions of a Real Variable
Fig. 7.1 A function with a limit
− +
Fig. 7.2 .ε-.δ intuition of the limit
Definition 7.2 (.ε-.δ) Let E be a nonempty subset of .R, and let p be an accumulation
point of E. We say that a number .L ∈ R is the limit of .f (x) as .x → p, if for every
.ε > 0 there exists .δ > 0 such that .x ∈ E and .0 < |x −p| < δ imply .|f (x)−L| < ε.
See Fig. 7.2.
Definition 7.3 (Topological Limit) Let E be a nonempty subset of .R, and let p be
an accumulation point of E. We say that a number .L ∈ R is the limit of .f (x) as
7 Limits: From Sequences to Functions of a Real Variable 83
x → p, if for every neighborhood V of L there exists a neighborhood U of p such

.
that .f (U \ {p}) ⊂ V .
Theorem 7.1 Definitions 7.1, 7.2, and 7.3 are logically equivalent.
Proof Since any neighborhood of a point x contains an interval of the form
.(x − r, x + r) for some .r > 0, it is clear that Definitions 7.2 and 7.3 are logically
equivalent statements. We prove that Definitions 7.1 and 7.3 are equivalent. Suppose
that .f (x) → L as .x → p in the sense of Definition 7.3, and let .{xn }n be a sequence
of points from E such that .xn = p for each n, and .xn → p. If V is any neighborhood
of L, we choose a neighborhood U of p as in Definition 7.3. There exists a positive
integer .n0 such that .n > n0 implies .xn ∈ U \ {p}. Then .f (xn ) ∈ V for .n > n0 , and
this proves that .f (xn ) → L as .n → +∞.
Conversely, suppose that Definition 7.3 fails to hold. Then there exists some
neighborhood V of L with the property that no neighborhood U of p satisfies the
condition of Definition 7.3. Fixing such V , it follows that for each positive integer
n there is some .xn ∈ E such that .|xn − p| < 1/n, .xn = p, and .f (xn ) ∈ / V . Thus
.{xn }n is a sequence in .E \ {p} that converges to p, but the sequence .{f (xn )}n does
not converge to L. Hence Definition 7.1 fails as well.

Remark 7.2 Definition 7.2 is for sure the most popular in Calculus courses. It has
many computational advantages, but it cannot be extended to topological structures
without a distance. We will come back to the theory of convergence in a later
chapter.
Time has come to introduce another fundamental property of functions, which
will produce important and useful consequences.
Definition 7.4 Let E be a nonempty subset of .R, and let .f : E → R be a function.
For a point .p ∈ E, we say that f is continuous at p if and only if for each
neighborhood V of .f (p) there exists a neighborhood U of p such that .f (U ) ⊂ V .
We say briefly that f is continuous on E if it is continuous at each point of E.
As a matter of facts, continuity and limits are essentially the same thing. Let us
try to be more precise.
Theorem 7.2 Let E be a nonempty subset of .R, let p be an accumulation point of
E, and let .f : E → R be a function. The following statements are equivalent:
(a) The limits .limx→p f (x) exists;
(b) there exists a number L such that the function defined on .E ∪ {p} by
f (x) if x = p
f˜(x) =
.
L if x = p
is continuous at p.
Proof Suppose .L = limx→p f (x). Then the function .f˜ is continuous at p,

since .limx→p f˜(x) = limx→p f (x) = L = f˜(x). Conversely, suppose that
L is chosen so that (b) holds. Then .f˜ coincides with f on .E \ {p}, and thus
.limx→p f (x) = limx→p f˜(x). But .f˜ is continuous at p, and we conclude that
.limx→p f (x) = f˜(p) = L. In particular .limx→p f (x) exists.

Remark 7.3 According to this result, we could consider the whole theory of
limits as an application of the theory of continuous functions. We won’t, both for
pedagogical and theoretical reasons.
Sometimes2 limits are just evaluation, as the next result shows.
Theorem 7.3 Let E be a nonempty subset of .R, and let .f : E → R be a function.
The function f is continuous at a point .p ∈ E if and only if either p is not an
accumulation point of E, or p is an accumulation point of E and .limx→p f (x) =
f (p).
Proof The case in which p is not an accumulation point of E is clear. Indeed, in this
case there is a neighborhood U of p such that no point of E other than p belongs
to U . If V is any neighborhood of .f (p), then .f (U ) = {f (p)} ⊂ V , and f is
continuous at p. We may therefore assume that p is an accumulation point of E. It
suffices to compare Definitions 7.3 and 7.4.

We extend the definition of limit in the following way.
Definition 7.5
• Suppose that E contains a half-line .(a, +∞) for some a. We say that the function
.f : E → R converges to .L ∈ R as .x → +∞, if for any sequence .{xn }n of
points in .(a, +∞) such that .xn → +∞ as .n → +∞, there results .f (xn ) →
L as .n → +∞. In this case we write .f (x) → L as .x → +∞, or briefly
.limx→+∞ f (x) = L.
• Suppose that E contains a half-line .(−∞, b) for some b. We say that the function
.f : E → R converges to .L ∈ R as .x → −∞, if for any sequence .{xn }n of
points in .(−∞, b) such that .xn → −∞ as .n → +∞, there results .f (xn ) →

L as .n → +∞. In this case we write .f (x) → L as .x → −∞, or briefly
.limx→−∞ f (x) = L.
• Suppose that p is an accumulation point of E. We say that the function .f : E →

R diverges to .+∞ as .x → p, if for every sequence .{xn }n of points in E such that
.∀n(xn = p) and .xn → p, there results .f (xn ) → +∞ as .n → +∞.
• Suppose that p is an accumulation point of E. We say that the function .f : E →

R diverges to .−∞ as .x → p, if for every sequence .{xn }n of points in E such that
.∀n(xn = p) and .xn → p, there results .f (xn ) → −∞ as .n → +∞.
2 Too many students tend to believe that this is indeed the general case. No, it isn’t.
7 Limits: From Sequences to Functions of a Real Variable 85
Similar definitions may be provided to describe the fact that .limx→+∞ f (x) =
+∞ or .limx→+∞ f (x) = −∞ and so on. The details are left to the reader.
Dealing with several different definitions of limits may seem troublesome. An
interesting way out consists in extending the set .R so that it contains .+∞ and .∞.
Definition 7.6 We call the set .R∗ = R ∪ {−∞, +∞} the extended real line. A
neighborhood of .−∞ is any set of the form .(−∞, b) for some .b ∈ R. Analogously,
a neighborhood of .+∞ is any set of the form .(a, +∞) for some .a ∈ R. If .E ⊂ R∗
and .p ∈ R∗ , we say that p is an accumulation point of E if any neighborhood of p
contains a point of E, different than p itself.
The extended real line allows us to summarize a all the possible definitions of
limit into a single topological definition.
Definition 7.7 (Limits in .R∗ ) Let .E ⊂ R∗ , let .p ∈ R∗ an accumulation point of E,
and let .L ∈ R∗ . We say that .f (x) tends to L as .x → p, if for every neighborhood
V of L in .R∗ there exists a neighborhood U of p in .R∗ such that .f (U \ {p}) ⊂ V .
Arithmetic operations in the extended real line present some difficulties. Indeed,
2
it is clear that .limx→+∞ xx = 1, while .limx→+∞ xx = +∞ and .limx→+∞ xx2 = 0.
There is no hope to define sums and products in .R∗ without any exceptional case.
The algebra of limits is completely satisfactory only for finite limits.
Theorem 7.4 (Algebra of Finite Limits) Retain the assumptions of Definition 7.7.
If
. lim f (x) = L ∈ R
x→p
lim g(x) = M ∈ R,
x→p
then
. lim (f (x) + g(x)) = L + M

x→p
lim f (x)g(x) = LM,

x→p
and
f (x) L
. lim =
x→p g(x) M
provided that .M = 0.
Proof The conclusion follows from the corresponding statements for sequences,
see Theorem 5.6.

The circumstance that the algebra of limits cannot be completely extended to

R∗ should not be seen as a weakness. Limits are a topological object, and such
.
is the set .R∗ . On the contrary the algebra of limits describes an interplay of
topology with the algebraic structure of real numbers. As a matter of facts, we
cannot define algebraic operations with .±∞ without losing some properties.
To rephrase this, there is no value of .∞/∞ or of .0 · ∞ that is compatible with
our definition of limits.
7.1 Properties of Limits
Clearly enough, many properties of limits can be deduced from similar properties of
limits for sequences. We just state a few important results that the reader may have
studied in Calculus courses.
Theorem 7.5 (Limits and Order) Let .E ⊂ R, .p ∈ R∗ be an accumulation point
of E, and .L ∈ R.
(a) If .L > 0, there exists a neighborhood U of p such that .f (x) > 0 for each
.x ∈ U \ {p}.
(b) If .f (x) > 0 for each .x = p that belongs to some neighborhood of p, then
.L ≥ 0.
Proof We first deal with the case .L ∈ R, and we use Definition 7.2. To prove (a), we
select .ε = L/2 so that a neighborhood U of p exists such that .L − L/2 < f (x) <
L + L/2 for each .x ∈ U \ {p}. This shows that .f (x) > L/2 > 0 for such values of
x. If (b) were false, then .L < 0. Applying (a) to .−f would yield a neighborhood of
p in which .−f would be positive, i.e. f would be negative. This is a contradiction.
The case .L = +∞ is easier. Indeed, For each .M > 0 there exists a neighborhood
U of p such that .f (x) > M for each .x ∈ U \ {p}. Conclusion (b) follows again by
(a) as above.

Theorem 7.6 (Squeezing Property) Let .p ∈ R∗ be an accumulation point of a set

E. Suppose three functions f , g and .σ are defined in some neighborhood of p, and
that .|f (x) − g(x)| < σ (x) in that neighborhood. If .σ (x) → 0 and .g(x) → L ∈ R
as .x → p, then also .f (x) → L as .x → p.
Proof Fix any .ε > 0. By assumption there exists a neighborhood U of p such
that .σ (x) < ε and .|g(x) − L| < ε provided that .x ∈ U \ {p}. For these x’s,
.|f (x) − L| ≤ |f (x) − g(x)| + |g(x) − L| < σ (x) + ε < 2ε.

Remark 7.4 The condition .|f (x) − g(x)| < σ (x) is equivalent to .g(x) − σ (x) <
f (x) < g(x) + σ (x), see Fig. 7.3. Under the assumptions of Theorem 7.6, .g(x) −
σ (x) and .g(x) + σ (x) both converge to the finite limit L as .x → p, and this implies
that .f (x) → L as .x → p.
7.2 Local Equivalence of Functions 87
Fig. 7.3 Squeezing of

functions
7.2 Local Equivalence of Functions
Since limits and continuity are related to the local behavior of functions, it seems
rather natural to classify functions according to their asymptotic properties around
a point.
Definition 7.8 Let I be an interval, let .c ∈ I or .c ∈ {−∞, +∞} (in case I is
unbounded), and let f , g be two functions defined on I . We say that f and g are
equivalent at c if and only if there exists a function .u : I → R such that .f (x) =
u(x)g(x) for each .x ∈ I and .limx→c u(x) = 1. In this case we write .f ∼ g as
.x → c.
Remark 7.5 If necessary, we will also write .f ∼ g as .x → c− of .x → c+ , with an

obvious meaning of symbols. We point out that the excluding the point c from the
test of .f = ug is not particularly meaningful. Indeed, since .u(x) → 1 as .x → c,
u can always be (re)defined at c so that .u(c) = 1. This forces .f (c) = g(c), and
therefore we may assume that f and g are defined at c without loss of generality.
Important: Simplified Definition

Several textbooks propose the following definition of local equivalence: .f ∼ g as
x → c if and only if
.
f (x)
. lim = 1.
x→c g(x)
Unfortunately this becomes troublesome when .g = 0 infinitely often around c,

while there is no need to exclude this possibility.
Theorem 7.7 Let .F(c) be the set of all functions defined on I . Then .∼ is an
equivalence relation on .F(c).
Proof For each .f ∈ F(c), .f ∼ f in a trivial way. Next, if .f ∼ g, there exists u
such that .f = ug on I , with .u(x) → 1 as .x → c. In particular .u = 0 near c, and
thus .g = (1/u)f near c. Hence .g ∼ f . To conclude, if .f ∼ g and .g ∼ h, we can
find functions u and v such that .f = ug and .g = vh with .u(x) → 1, .v(x) → 1 as
.x → c. Then .f = uvh and .u(x)v(x) → 1 as .x → c. Hence .f ∼ h.

Exercise 7.1 Suppose that .limx→c f (x) = limx→c g(x) = y0 = 0. Prove that
f ∼ g as .x → c. Show with a counterexample that the condition .y0 = 0 is essential.
.
Local equivalence is just the first ingredient of local comparison.

Definition 7.9 Let I be an interval, let .c ∈ I or .c ∈ {−∞, +∞} (in case I is
unbounded), and let f , g be two functions defined on I .
(o) We say that .f = o(g) as .x → c if and only if there exists a function .σ such
that .f (x) = σ (x)g(x) on I and .limx→c σ (x) = 0. In this case we also say that
f is negligible compared to g as .x → c.
(O) We say that .f = O(g) as .x → c if and only if there exists a bounded function
.β such that .f (x) = β(x)g(x) on I .
Exercise 7.2 Suppose that g does not vanish in a neighborhood of c. Prove that
f = o(g) if and only if
.
f (x)
. lim = 0.
x→c g(x)
Prove also that .f = O(g) if and only if .x → f (x)/g(x) remains bounded in a

neighborhood of c.
We summarize the basic algebraic properties of local comparison in the following
exercise. The proofs are simple consequences of the definition.
Definition 7.10 We say that a function g is bounded away from zero around a point
x if and only if there exists a number .ε > 0 such that .|g(x)| ≥ ε for all x in a
neighborhood of c.
Exercise 7.3 Prove that the following statements are true as .x → c.
1. If .f = o(h) and .g = o(h), then .f + g = o(h) and .fg = o(h).
2. If .f = o(h) and .α ∈ R, then .αf = o(h).
3. If .f = o(h) and g is bounded away from zero, then .f/g = o(h).
4. If .f = o(g) and .g = O(h), then .f = o(h).
5. If .f = o(1), then .1/(1 + f ) = 1 − f + o(f ).
6. If .f = O(h) and .g = O(h), then .f + g = O(h) and .fg = O(h).
7. If .f = O(h) and .α ∈ R, then .αf = O(h).
8. If .f = O(h) and g is bounded away from zero, then .f/g = O(h).
9. If .f = O(g) and .g = O(h), then .f = O(h).
7.3 Comments 89
10. If .f = o(1) and .g = O(h), then .fg = o(h).

11. If .f = O(1) and .g = O(h), then .fg = O(h).
Needless to say, the language of local comparison is a perfect dialect for speaking
about limits.
Example 7.1
1. Since3
ex − 1
. lim = 1,
x→0 x
we can write .ex = 1 + x + o(x) as .x → 0.

2. For a similar reason, .sin x = x +o(x) as .x → 0, and .cos x = 1−(1/2)x 2 +o(x 2 )
as .x → 0.
3. If .f1 = o(f ) and .g1 = o(g) as .x → c, then
f (x) + f1 (x) f (x)

. lim = lim .
x→c g(x) + g1 (x) x→c g(x)
Indeed,
f1 (x)
f (x) + f1 (x) f (x) 1 + f (x)
. lim = lim .
x→c g(x) + g1 (x) x→c g(x) 1 + g1 (x)
g(x)
This is often called the principle of negligible terms. Its use in computing limits
is ubiquitous.
7.3 Comments
We will discuss again the definition of limit in the chapter about topology. For the
moment, I point out that our definition remains the most common in contemporary
literature, although some alternatives actually exist. In particular, a few authors
propose the following variant:
.limx→p f (x) = q if and only if for every .ε > 0 there exists .δ > 0 such that .|x − p| < δ
implies .|f (x) − q| < ε.
3We assume that the reader is familiar with a few limits that involve the elementary functions.
Formal proofs will be given later on, when we discuss these functions from an advanced viewpoint.
The difference is that the case .x = p is not excluded. It is therefore clear that the
condition .f (p) = q is a necessary condition for the existence of the limit with this
definition. If you like this approach, you should always remember that limits are no
longer independent of the value of the functions at the point p.
Chapter 8
Continuous Functions of a Real Variable
Abstract Calculus students tend to believe that continuous functions are those
functions which “vary a little when the independent variable varies a little.” In this
chapter we define continuity in a rigorous way, and we invite the readers to convince
themselves that the previous sentence is actually false.
First of all, we reformulate Definition 7.4 in a quantitative way.

Definition 8.1 Let E be a subset of .R, and let .f : E → R be a function. We say
that f is continuous at the point .p ∈ E if for every .ε > 0 there exists .δ > 0 such
that .x ∈ E and .|x − p| < δ imply .|f (x) − f (p)| < ε. If f is continuous at any
point of E, we say that f is continuous on E.
Two cases are possible. If the point p is also an accumulation point of E (hence
p ∈ E ∩ E % ), then the previous definition merely says that .limx→p f (x) = f (p).
.
The second case is .p ∈ E but not an accumulation point of E. This means that p is
isolated, in the sense that there exists a neighborhood U of p such that .U ∩E = {p}.
In this situation, the continuity of f at p is always granted. Indeed, if .ε > 0, we
may choose .δ > 0 so small that the condition .|x − p| < δ is satisfied only by .x = p,
and therefore .|f (x) − f (p)| = |f (p) − f (p)| = 0 < ε.
Remark 8.1 Most Calculus books propose .limx→p f (x) = f (p) as the definition
of continuity, but this is equivalent to ours only under the assumption that p is an
accumulation point of the domain of f .
Recalling Theorem 7.1 and Definition 7.4, we may state
Theorem 8.1 Let .f : E → R. The function f is continuous at the point .p ∈ E if
and only if one of the following conditions is met:
(i) for every .ε > 0 there exists .δ > 0 such that .x ∈ E and .|x − p| < δ imply
.|f (x) − f (p)| < ε;
(ii) For every neighborhood V of .f (p) there exists a neighborhood U of p such

that .f (U ∩ E) ⊂ V ;

92 8 Continuous Functions of a Real Variable
(iii) for every sequence .{xn } of point from E that converges to p, there results
.f (xn ) → f (p) as .n → +∞.
Corollary 8.1 If there exists a sequence .{xn }n ⊂ E that converges to p, but

{f (xn )}n does not converge to .f (p), then f is discontinuous at p.
.
The word discontinuity is used as the negation of continuity. In this book we will
not enter into the troublesome classification of discontinuity points, since we believe
that this is of little interest.
Exercise 8.1 Let .f : E → R be a function defined on .E ⊂ R. Prove that f is
continuous at .x0 ∈ E if and only if for every monotonic sequence .{xn }n converging
to .x0 , we have .f (xn ) → f (x0 ). Hint: use Theorem 5.12.
Example 8.1 Let .E = R, .p = 0 and
x
|x| if x = 0
f (x) =
.
0 if x = 0.
Consider the two sequences defined by .xn = 1/n and .yn = −1/n for .n ∈ N. Then
we form the sequence .{zn }n by the rule
x1 , y1 , x2 , y2 , . . . , xn , yn , . . .
.
Clearly .zn → 0 as .n → +∞, but the sequence .{f (zn )}n is
1, −1, 1, −1, . . .
.
This sequence cannot converge to .f (0) = 0, since .lim infn→+∞ f (zn ) = −1 =

1 = lim supn→+∞ f (zn ).
Example 8.2 Let us define f on .[0, 1) ∪ (1, 2] by the following rule:
0 if 0 ≤ x < 1
f (x) =
.
1 if 1 < x ≤ 2.
For each .x ∈ [0, 1) ∪ (1, 2], the function f is continuous at x. However, if x is very
close to 1 but smaller than 1, the value of .f (x) is zero. If x is very close to 1 but
larger than 1, the value of .f (x) is 1. We are allowed to say that f changes a lot when
x is varied a little! In other words, there is much more in the definition of continuity
than the basic idea of “a little on the x-axis becomes a little on the y-axis.”
8 Continuous Functions of a Real Variable 93
The algebraic rules for computing (finite) limits immediately implies the follow-
ing result.
Theorem 8.2 Suppose that .f : E → R and .g : E → R are continuous at .p ∈ E.
Then
(i) .x → kf (x) is continuous at p for every .k ∈ R;
(ii) .x → f (x) + g(x) is continuous at p;
(iii) .x → f (x)g(x) is continuous at p;
(iv) .x → f (x)/g(x) is continuous at p, provided that the quotient is defined.
As a consequence, any polynomial function .x → a0 + a1 x + a2 x 2 + · · · + an x n
is continuous on .R, as the sum of continuous functions.
Example 8.3 We consider the function .h(x) = [x], where .[x] denotes the largest
integer .n ∈ Z such that .n ≤ x. Given a point .p ∈ Z, we consider the sequence
.xn = p − 1/n. Clearly .xn → p, but .h(xn ) → p − 1, which is different than
.h(p) = p. Hence h is discontinuous at all integer points. On the contrary, h is
continuous at all non-integers. Indeed, let us fix .p ∈ / Z. Given .ε > 0, we want

to find a neighborhood U of p such that .h(U ) ⊂ (h(p) − ε, h(p) + ε). Since p
is not an integer, there exists an integer n such that .n < p < n + 1. Let us take
.δ = min{p − n, (n + 1) − p}: it follows from the definition of h that .h(x) = h(p)
for every .x ∈ (p − δ, p + δ). Thus we certainly have .h(x) ∈ (h(p) − ε, h(p) + ε)

for every .x ∈ (p − δ, p + δ).
Since elementary functions (polynomials, sine, cosine, exponentials and loga-
rithms) are continuous functions in their natural domains of definition, no arithmetic
operation on them can produce discontinuous functions. We now observe that even
their composition must be continuous.
Theorem 8.3 (Continuity of Composite Functions) Given .f : E → R and
.g : F → R, assume that the range .f (E) is contained in F , so that the composition
.g ◦ f is defined on E. If f is continuous at a point .p ∈ E and if g is continuous at
the point .f (p) ∈ F , then .g ◦ f is continuous at p.

Proof We use Theorem 8.1. Let .{xn }n be a sequence in E that converges to p. By
assumption .f (xn ) → f (p), and so .g(f (xn )) → g(f (p)) since g is continuous at
.f (p) and .f (xn ) → f (p). This shows that .g ◦ f (xn ) → g ◦ f (p), namely that .g ◦ f
is continuous at p.

The previous theorem is indeed a particular case of the following formula about
change of variables in limits.
Theorem 8.4 (Changing Variables in Limits) Let D and E be subsets of .R∗ =
R ∪ {−∞} ∪ {+∞}, let c and p be accumulation points of D and E respectively.
Let .f : D → R∗ be a function, and let .ϕ : E → D be a bijective function such that
. lim ϕ(t) = c, lim ϕ −1 (x) = p.

t →p x→c
Under these assumptions, .limx→c f (x) exists in .R∗ if and only if .limt →p f ◦ ϕ(t)
exists in .R∗ , and in such a case these limits coincide.
Proof We refer to Fig. 8.1. Let us assume that .L = limx→c f (x) exists in .R∗ . Let
V be a neighborhood of L in .R∗ , and let U be a neighborhood of c in .R∗ such that
.f (U ∩ D \ {c}) ⊂ V . Furthermore, given the neighborhood U of c there exists a
neighborhood W of p in .R∗ such that .ϕ(W ∩ E \ {p}) ⊂ U . The crucial remark

is now that .ϕ is bijective: there exists one and only one element .a ∈ E such that
.ϕ(a) = c. In case .a = p, we may take a smaller neighborhood W so that .a ∈ / W . In
this way we are sure that .ϕ(W ∩ E \ {p}) ⊂ U \ {c}. As a consequence .f (ϕ(t)) ∈ V
whenever .t ∈ W ∩ E \ {p}. This shows that .L = limt →p f (ϕ(t)).
To prove the converse, we simply apply the result to the function .f ◦ ϕ and .ϕ −1 ,
and we conclude because .(f ◦ ϕ) ◦ ϕ −1 = f .

The rough interpretation of Theorem 8.4 is that .x = ϕ(t); hence .x → c if and
only if .t → p, and there is no difference between the variable x and the variable t
when limits are considered.
Example 8.4 The following limits are equivalent:
ex − 1
. lim =1
x→0 x
Fig. 8.1 Change of variable in the limit

8.1 Continuity and Compactness 95
and
log(x + 1)
. lim = 1.
x→0 x
Indeed, setting .et − 1 = x, we get .t = log(x + 1) and .x → 0 if and only if .t → 0.

The following example introduces a quantitative approach to continuity of real-
valued functions.
Example 8.5 Let .f : R → R be a function. For each .x ∈ R, we define
ωf (x) = inf sup {|f (y) − f (z)| | {y, z} ⊂ (x0 − δ, x0 + δ)} .

.
δ>0
We claim that the set .{x ∈ R | ωf (x) < ε} is open for each .ε > 0. Indeed, we
suppose that .ωf (x0 ) < ε, and we must prove that the same inequality holds in an
open interval containing .x0 . By definition, there exists .δ0 such that
. sup {|f (y) − f (z)| | {y, z} ⊂ (x0 − δ, x0 + δ)} < ε.
For every .x ∈ (x0 − δ/2, x0 + δ/2) we have .(x0 − δ/2, x + δ/2) ⊂ (x0 − δ, x0 + δ),
and therefore
ωf (x) ≤ sup {|f (y) − f (z)| | y, z ∈ x − δ/2, x + δ/2)}

.
≤ sup {|f (y) − f (z)| | y, z ∈ (x0 − δ, x0 + δ}

< ε.
Exercise 8.2 Prove that a function f is continuous at a point .x0 if and only if
ωf (x0 ) = 0.
.
8.1 Continuity and Compactness
The most interesting properties of continuous functions are related to the compact-
ness of their domains.
Theorem 8.5 (Preservation of Compactness) Let .f : E → R be a continuous
function, and suppose that K is a (sequentially) compact subset of A. Then .f (K) is
a (sequentially) compact subset of .R.
Proof Let .{yk }k be a sequence in the range .f (K). Hence for each positive integer k,
there exists an element .xk ∈ K such that .f (xk ) = yk . The sequence .{xk }k possesses
a converging subsequence .{xnk }k . Call x its limit. By continuity of f at x, there
results .f (x) = limk→+∞ f (xnk ). Hence .ynk → f (x) as .k → +∞.

Preservation of compactness under continuous functions ensures the existence of

maxima and minima. The following is a fundamental result of Real Analysis.
Theorem 8.6 (Weierstrass) If .f : K → R is a continuous function on a (sequen-
tially) compact subset K of .R, then f attains a maximum and a minimum value. In
particular, there exist points .x0 and .x1 in K such that .f (x0 ) ≤ f (x) ≤ f (x1 ) for
every .x ∈ K.
Proof Since the range .f (K) is (sequentially) compact, we can set .α = sup f (K)
and deduce that .α ∈ f (K). Indeed .f (K) is a closed set, and .α is an accumulation
point of .f (K) (by Theorem 5.2). This shows that f attains a maximum value. In a
similar fashion we can show that f attains a minimum value.

Continuity is essentially a local property, since a function is continuous on a set
if and only if it is continuous at each point of the set. We introduce now a genuinely
global definition which extends continuity in a proper way.
Definition 8.2 Let E be a subset of .R, .f : E → R a function. We say that f is
uniformly continuous on E if for every .ε > 0 there exists .δ > 0 such that .x ∈ E,
.y ∈ E and .|x − y| < δ imply .|f (x) − f (y)| < ε.
It is worth noticing that uniform continuity implies at once continuity on E: just

fix .y ∈ E and use the definition. On the other hand, the following example shows
that the two definitions remain distinct.
Example 8.6 The function .g : x → x 2 is not uniformly continuous on .R. Indeed,
we can consider the two sequences .xn = n and .yn = n + 1/n. Clearly
.limn→+∞ xn = limn→+∞ yn = +∞, while .limn→+∞ |xn − yn | = 0. Now,

2
2 1
.|f (xn ) − f (yn )| = n − n+
n

1
= 2 + 2 .
n
Hence .limn→+∞ |f (xn ) − f (yn )| = 2, against the definition of uniform continuity.

The previous argument can be turned into a general test for uniform continuity.
Theorem 8.7 A function .f : E → R fails to be uniformly continuous on E if and
only if there exist a positive number .ε0 > 0 and two sequences .{xn }n , .{yn }n in E
that satisfy .|xn − yn | → 0, .|f (xn ) − f (yn )| ≥ ε0 .
Proof If .ε0 and the two sequences exist, it is immediate to check that f cannot be
uniformly continuous. Actually, no matter how close .xn and .yn are, their images
.f (xn ) and .f (yn ) remain far away from each other.
Conversely, we suppose that f fails to be uniformly continuous on E. Negating

the definition, we discover that there exists .ε0 such that for every .δ > 0 there must
exist points .xδ and .yδ in E satisfying .|xδ − yδ | < δ, but .|f (xδ ) − f (yδ )| ≥ ε0 .
8.1 Continuity and Compactness 97
Now .δ > 0 is a free variable, and we can select .δ = δn = 1, 2, 3, . . .. Hence two

sequences .{xδn }n and .{yδn }n are born such that .|xδn − yδn | < δn , but .|f (xδn ) −
f (yδn )| ≥ ε0 .

Again compactness does the magic: uniform continuity boils down to continuity
on (sequentially) compact sets.
Theorem 8.8 A function that is continuous on a (sequentially) compact subset K
of .R is uniformly continuous.
Proof Let f be a continuous function on K. We argue by contradiction, assuming
that f fails to be uniformly continuous on K. Then there exist .ε0 > 0 and two
sequences .{xn }n , .{yn }n in E that satisfy .|xn − yn | → 0, .|f (xn ) − f (yn )| ≥ ε0 .
Compactness comes into play: a subsequence .{xnk }k of .{xn }n converges to a point
.x ∈ K. Then also .ynk converges to the same limit x, since .|xnk − ynk | → 0 (write
.|ynk − x| = |ynk − xnk + xnk − x| and use the triangle inequality). Since f is
continuous at x, we know that .f (x) = limk→+∞ f (xnk ) = limk→+∞ f (ynk ). This

contradicts the condition .|f (xnk ) − f (ynk )| ≥ ε0 .

Exercise 8.3 Prove that uniformly continuous functions map Cauchy sequences
into Cauchy sequences. More precisely, if f is uniformly continuous on a set S
and if .{xn }n is a Cauchy sequence in S, then .{f (xn )}n is a Cauchy sequence. Hint:
just apply the definition of uniform continuity.
Theorem 8.9 A real-valued function f on .(a, b) is uniformly continuous on .(a, b)
if and only if it can be extended to a continuous function .f˜ on .[a, b].
Proof If f can be extended, then .f˜ is automatically uniformly continuous on the
compact set .[a, b], and thus f is uniformly continuous on .(a, b) as a restriction.
Conversely, we need to define .f˜(a) and .f˜(b) in such a way that the extended
function is continuous. We construct the value .f˜(a), since the other case is similar.
Let .{xn }n any sequence in .(a, b) such that .xn → a. The sequence .{xn }n is a Cauchy
sequence, and therefore .{f (xn )}n is a Cauchy sequence by the previous exercise.
Thus .limn→+∞ f (xn ) exists: we set
f˜(a) = lim f (xn ).

.
n→+∞
Now, the value of .f˜(a) depends on the choice of the sequence .{xn }n . But this is not
the case. Indeed, let .{yn }n be another sequence in .(a, b) such that .yn → a. Then the
sequence .{un }n defined by
{x1 , y1 , x2 , y2 , . . . , xn , yn , . . .}
.
is a sequence in .(a, b) which converges to a. We deduce that .limn→+∞ f (un )

exists, and since .{f (xn )}n and .{f (yn )}n are subsequences of .{f (un )}n , the limits
.limn→+∞ f (xn ) and .limn→+∞ f (yn ) must be equal. We have thus proved that the
definition of .f˜(a) is unambiguous, and plainly .f˜ is continuous at a.

8.2 Intermediate Value Property
The next results are of topological nature, and in a perfect mathematical world they
should be obtained from the properties of connected sets. We present them in this
chapter, and we invite the interested reader to think back of them after studying the
chapter on General Topology.
Theorem 8.10 Let .f : [a, b] → R be a continuous function. If L is a real number
with either .f (a) < L < f (b) or .f (a) > L > f (b), then there exists a point
.c ∈ (a, b) such that .f (c) = L.
Proof By considering .f˜ : x → f (x) − L instead of f , we can always assume that

L = 0. We prove the Theorem under the assumption .f (a) < 0 < f (b), the other
.
case being similar. Therefore we look for a point .c ∈ (a, b) such that .f (c) = 0. Let
us introduce the set
K = {x ∈ [a, b] | f (x) ≤ 0} .
.
The number b is an upper bound for K, and .K = ∅ because .a ∈ K. Then .c = sup K

exists in .R. Let us prove that .f (c) = 0. If not, either .f (c) > 0 of .f (c) < 0. Suppose
.f (c) > 0: by continuity, .f > 0 in a neighborhood .(c − δ, c + δ) of c. In particular
.K ⊂ [a, c−δ], against the definition of c as the supremum of K. By the same token,
one excludes the case .f (c) < 0, and the conclusion follows.

Exercise 8.4 In this exercise we propose a second proof of Theorem 8.10. Suppose
.L = 0 and .f (a) < 0 < f (b). Define .I0 = [a, b] and consider the mid-point
.z = (a + b)/2. If .f (z) ≥ 0, set .a1 = a and .b1 = z. If .f (z) < 0, set .a1 = z and
.b1 = b. This gives rise to an interval .I1 ⊂ I0 . Use this scheme and the principle of
infinitesimal nested intervals to prove Theorem 8.10.

The previous result suggests the following question: if a function attains every
value between its infimum and its supremum, is it a continuous function?
Definition 8.3 A function f has the intermediate value property on an interval
[a, b] if for all .x < y in .[a, b], and all L between .f (x) and .f (y), it is possible
.
to find a point .c ∈ (x, y) such that .f (c) = L.

We have seen in Theorem 8.10 that any continuous function has the intermediate
value property. it is a common mistake to assume that the intermediate value
property is a characterization of continuous functions.
Example 8.7 Define .g : (−∞, +∞) → R by
sin(1/x) if x = 0
.g(x) =
0 otherwise.
8.3 Continuous Invertible Functions 99
For each .L ∈ [−1, 1], the equation .sin z = L possesses infinitely many solutions
z = 0, and therefore g has the intermediate value property. However g is not
.
continuous at 0 (why?).
Theorem 8.11 If f is increasing on .[a, b] and satisfies the intermediate value
property, then f is continuous on .[a, b]. The same results holds for a decreasing
function.
Proof We consider a point .c ∈ (a, b): the continuity of f at a and b is left as an
exercise. By monotonicity,
f (c−) = lim f (x) = sup {f (x) | a ≤ x < c}

.
x→c−
f (c+) = lim f (x) = inf {f (x) | c < x ≤ b} ,
x→c+
and both limits are finite. Again by monotonicity, .f (c−) ≤ f (c+). We need to
prove that actually .f (c−) = f (c+). We suppose on the contrary that .f (c−) <
f (c+), and pick a number L such that .f (c−) < L < f (c+). We have now two
cases: (i) if .f (c) = L, we reach a contradiction with the intermediate value property.
If (ii) .f (c) = L, we select .L% = L and fall into case (i). The proof is complete.
8.3 Continuous Invertible Functions
A word of warning: the results of this section are typical of functions of a single
variable. We will see in the chapter on General Topology that the continuity of an
inverse function is not for free. Here the algebraic properties of the real line add a
remarkable richness.
Example 8.8 Let A be a subset of .R, and let f be an injective continuous function
from A to .R. We show that .f −1 need not be continuous. Consider .A = [0, 1]∪[2, 3]
and
x if 1 ≤ x ≤ 2
f (x) =
.
x − 1 if 2 ≤ x ≤ 3.
Clearly f is continuous on A, but its inverse .f −1 is defined on .[0, 2] by
y if 0 ≤ y ≤ 1
f −1 (y) =
.
y + 1 if 1 ≤ y ≤ 2.
The point .y = 1 is a discontinuity point.

Theorem 8.12 If f is a continuous, invertible function on an interval I , then f is

strictly monotonic.
Proof Let .x0 ∈ I be a fixed point. Suppose that .f (x̄) > f (x0 ) for some .x̄ > x0 . If
a point .x % > x0 exists such that .f (x % ) ≤ f (x0 ), then Theorem 8.10 yields a point .ξ
between .x0 and .x̄ with .f (ξ ) = f (x0 ). This is impossible, since f is invertible. We
have proved that .x → f (x) − f (x0 ) keeps the same sign in .I ∩ [x0 , +∞).
Let us consider any two points a and b of I , with .a f (a), we see from the previous argument that
.f (x) > f (a) for every .x > a. In particular .f (b1 ) > f (a). By the same token,
.f (x) < f (b1 ) for every .x < b1 , hence .f (x1 ) < f (b1 ). Finally, .f (x1 ) < f (x) for
every .x > x1 , and in particular .f (x1 ) < f (x2 ). The claim is proved.
If .f (a) > f (b), in a similar way we prove that f is monotonically decreasing
on I . The proof is complete.

Theorem 8.13 Let g be a function defined on an interval J , and monotonic on J .
The function g is continuous on J if and only if the range .g(J ) is an interval.
Proof Without loss of generality we assume that g is increasing on J . If g is
continuous, then .g(J ) is an interval by Theorem 8.10.
We suppose now that .g(J ) is an interval. If g is discontinuous at some point
.x0 ∈ J , then
.l = lim g(x) < lim g(x) = L.

x→x0− x→x0+
For any .x < x0 we have .g(x) < l, while for any .x > x0 we have .g(x) > L. Among
all points of .(l, L), at most one can belong to .g(J ), a set that contains both points
smaller than l and points larger than L. This contradicts the assumption that .g(J ) is
an interval. The proof is complete.

Theorem 8.14 (Continuity of the Inverse Function) If a function f is continuous
on an interval I and invertible, then .f −1 is continuous on .f (I ).
Proof By Theorem 8.12 f is strictly monotonic on I . By Theorem 8.10 .J = f (I )
is an interval. The function .f −1 is defined on J and is monotonic on J . Its range is
I . By Theorem 8.13 .f −1 is therefore continuous.

Example 8.9 The previous results justify several elementary statements. For
instance, the continuity of the exponential function .x → ex is equivalent to the
continuity of the logarithm function .x → log x, since they both map an interval into
an interval.
Important: Warning
We will not spend too much time in proving the continuity of elementary functions.
Most traditional proofs are actually circular, the reason being the lack of a rigorous
8.4 Problems 101
definition of the functions themselves. Consider the following sketch of a proof that
x → ex is continuous at .x = 0. For each .ε > 0, the inequalities
.
1 − ε < ex < 1 + ε
.
is equivalent to
. log (1 − ε) < x < log (1 + ε) ,
at least when .0 < ε < 1. Hence (?) the continuity is proved.

Unfortunately we have tacitly used several properties of the exponential without
any authorization. Monotonicity is the most evident, but monotonicity is indeed the
main reason why the exponential function is continuous! In a future chapter we
will define several elementary functions in such a way that their properties can be
proved.
8.4 Problems
8.1 Let y be a point of R and let f be a function defined on an interval containing

y. We define
. lim sup f (x) = inf {sup {f (x) | 0 < |x − y| < δ} | δ > 0}

x→y
lim inf f (x) = sup {inf {f (x) | 0 < |x − y| < δ} | δ > 0.}
x→y
Prove the following statements, and conjecture similar statements for lim inf.
(a) lim supx→y f (x) ≤ A if and only if for every ε > 0 there exists δ > 0 such that
0 < |x − y| < δ implies f (x) ≤ A + ε.
(b) lim supx→y f (x) ≥ A if and only if for every ε > 0 and for every δ > 0 there
exists a point x such that 0 < |x − y| < δ and f (x) ≥ A − ε.
(c) lim infx→y f (x) ≤ lim supx→y f (x) with equality if and only if limx→y f (x)
exists.
(d) If lim supx→y f (x)= A and if {xn }n converges to y, then lim supn→+∞ f (xn ) =
A.
(e) If lim supx→y f (x) = A, then there exists a sequence {xn }n such that xn → y
and A = limn→+∞ f (xn ) = A.
8.2 A function f : R → R is homogeneous of degree one if f (λx) = λf (x) for
each x ∈ R and each λ ∈ R. Prove that f is continuous.
8.3 Let {a

n }n and {bn }n be two sequences of real numbers. Suppose that each an ≥
0 and that n an converges to a finite sum a. A function fn is defined for each n by
0 if x < bn
fn (x) =
.
an if x ≥ bn .

Let f (x) = n fn (x) for all x. prove that
1. f is non-decreasing.
2. f is discontinuous on the set A = {bn | n ∈ N}.
3. f is continuous on R \ A.
8.4 Let f and g be real-valued functions, uniformly continuous on a set A.
1. Prove that f + g is uniformly continuous on A.
2. Prove that f ◦ g is uniformly continuous on g(A) ∩ A.
3. Prove that fg is uniformly continuous if A = [a, b]. Give a counterexample if A
is not a compact interval.
8.5 Let A be a non-empty subset of R. A function fA : R → R is defined by
fA (x) = inf {|x − a| | a ∈ A}. Prove that fA is uniformly continuous on A.
8.6 Let K be a compact subset of R, and let f : K → R be a continuous function.
Prove that for each ε > 0 there exists M ∈ R such that |f (x)−f (y)| ≤ M|x −y|+ε
for each x, y ∈ K.
8.7 A function f : [0, 1] → R is upper semicontinuous if given x ∈ [0, 1] and
ε > 0 there exists δ > 0 such that |y − x| < δ implies f (y) < f (x) + ε. Prove
that an upper semicontinuous function on [0, 1] is bounded above and attains its
maximum value at some point of [0, 1].
Chapter 9
Derivatives and Differentiability
Abstract Derivatives are usually introduced by fully exploiting the possibility of

dividing real numbers. We propose an approach that can be extended almost literally
to function defined on general normed spaces.
Definition 9.1 Let .f : (a, b) → R be a function, and let .x0 ∈ (a, b) be a

distinguished point. We say that f is differentiable at .x0 if a real number A exists
with the property that
.f (x) = f (x0 ) + A(x − x0 ) + o(|x − x0 |) as x → x0 . (9.1)
The number A is called the derivative of f at .x0 , and is denoted by any of the
symbols
f % (x0 ),
. Df (x0 ), df (x0 ), f˙(x0 ).
The reader should recall that (9.1) is an equivalent way of requiring
f (x) − f (x0 ) − A(x − x0 )

. lim = 0.
x→x0 |x − x0 |
Remark 9.1 The assumption that f be defined on an open interval .(a, b) is

essentially for definiteness. Equation (9.1) shows that .x0 must be an accumulation
point of the domain of f , but it should also belong to it. It would be possible to
differentiate functions defined on a closed interval .[a, b], for instance, but at the
end-points the derivative would lose several properties. For this reason we define
the derivative of a function only at interior points of its domain.
Exercise 9.1 Suppose that f is differentiable at .x0 . We want to prove that the
number A in (9.1) is uniquely determined. For the sake of contradiction, assume
104 9 Derivatives and Differentiability
that
f (x) = f (x0 ) + A(x − x0 ) + o(|x − x0 |)

.
f (x) = f (x0 ) + B(x − x0 ) + o(|x − x0 |)
as .x → x0 . Deduce that .A − B = o(1) as .x → x0 , and conclude that .A = B.

If (9.1) holds, then
f (x) − f (x0 )
. lim = A. (9.2)
x→x0 x − x0
On the other hand, if
f (x) − f (x0 )
. lim = A,
x→x0 x − x0
then
f (x) − f (x0 )
. = A + o(1) as x → x0 ,
x − x0
or .f (x) − f (x0 ) = A(x − x0 ) + o(|x − x0 |) as .x → x0 . We have proved

Theorem 9.1 For a function .f : (a, b) → R the following conditions are equiva-
lent:
(i) f is differentiable at .x0 ∈ (a, b) and .f % (x0 ) = A;
(ii) the limit .limx→x0 f (x)−f
x−x0
(x0 )
exists as a real number and is equal to A.
Remark 9.2 Calculus books usually propose the derivative as the limit of the
incremental ratio, namely (9.2). Our Definition 9.1 can be formally generalized to
the case in which the function f is defined on a normed vector space, like .Rn for
.n ≥ 2. Equivalent definitions may be used as they are needed: we will see that (9.1)
is the most convenient characterization of the derivative for proving the chain rule.
We record a third definition of the derivative in terms of continuous functions.
Theorem 9.2 A function .f : (a, b) → R is differentiable at .x0 ∈ (a, b) if and only
if there exists a continuous function .ω : (a, b) → R such that
.f (x) = f (x0 ) + ω(x)(x − x0 ) for every x ∈ (a, b). (9.3)
In this case, .f % (x0 ) = ω(x0 ).

Proof Condition (9.3) simply means that the function
f (x) − f (x0 )
x →
. , x = x0
x − x0
9.1 Rules of Differentiation, or the Algebra of Calculus 105
can be extended at .x = x0 continuously. Of course this is true if and only if the limit
f (x) − f (x0 )
. lim
x→x0 x − x0
exists as a real number. This means that f is differentiable at .x0 , and by (9.3) we
must have .ω(x0 ) = limx→x0 f (x)−f
x−x0
(x0 )
.

Corollary 9.1 If a function is differentiable at a point, then it is continuous at that
point.
Proof This is immediate from (9.3).

Exercise 9.2 Prove the previous Corollary by using each of the equivalent defini-
tions of the derivative.
9.1 Rules of Differentiation, or the Algebra of Calculus
If two functions f and g are defined on a neighborhood of a point .x0 , we can define
pointwise the functions .f +g and .f ·g: indeed .x → f (x)+g(x) and .x → f (x)g(x)
are well defined in a neighborhood of .x0 . If .g = 0 in a neighborhood of .x0 , then the
quotient .x → f (x)/g(x) is also defined.
Theorem 9.3 (Differentiation Rules) Suppose that f and g are defined on a
neighborhood .(a, b) of the point .x0 . Then
(i) the function .f + g is differentiable at .x0 , and .(f + g)% (x0 ) = f % (x0 ) + g % (x0 );
(ii) the function .f · g is differentiable at .x0 , and
.(f · g)% (x0 ) = f % (x0 )g(x0 ) + f (x0 )g % (x0 );
(iii) if .g(x0 ) = 0, then the function .f/g is differentiable at .x0 , and
f % (x0 )g(x0 ) − f (x0 )g % (x0 )

(f/g)% (x0 ) =
. .
g(x0 )2
Proof The proof of (i) is left as an easy exercise. A standard proof of (ii) is as
follows:
f (x)g(x) − f (x0 )g(x0 ) f (x)g(x) − f (x0 )g(x) + f (x0 )g(x) − f (x0 )g(x0 )
. =
x − x0 x − x0
f (x) − f (x0 ) g(x) − g(x0 )
= g(x) + f (x0 )
x − x0 x − x0
→ f % (x0 )g(x0 ) + f (x0 )g % (x0 ).
Another proof, based instead on Definition 9.1, starts from the assumptions .f (x) =
f (x0 ) + f % (x0 )(x − x0 ) + o(1), .g(x) = g(x0 ) + g % (x0 )(x − x0 ) + o(1) and then

f (x)g(x) = f (x0 ) + f % (x0 )(x − x0 ) + o(1) g(x0 ) + g % (x0 )(x − x0 ) + o(1)
.
= f (x0 )g(x0 ) + f % (x0 )g(x0 )(x − x0 )

+ f (x0 )g % (x0 )(x − x0 ) + [. . .] o(1),
where .[. . .] contains all the terms that are multiplied by some .o(1) in the algebraic
expansion. It follows that the linearization of fg at .x0 (exists and) is .f % (x0 )g(x0 ) +
f (x0 )g % (x0 ).
The proof of (iii) is more traditional. First of all, since differentiability implies
continuity, the condition .g(x0 ) = 0 implies that .g = 0 in a neighborhood of .x0 .
Then we construct
f (x) f (x0 )
g(x) − g(x0 ) 1 f (x)g(x0 ) − f (x0 )g(x)
. =
x − x0 x − x0 g(x)g(x0 )
1 f (x)g(x0 ) − f (x0 )g(x0 ) + f (x0 )g(x0 ) − f (x0 )g(x)
=
x − x0 g(x)g(x0 )
f % (x0 )g(x0 ) − f (x0 )g % (x0 )
→ .
g(x0 )2

Remark 9.3 Formula (iii) is not easily proved by means of Definition 9.1. The
trouble is that it is not trivial to extract a linearization formula from the quotient
f (x0 ) + f % (x0 )(x − x0 ) + o(1)

. .
g(x0 ) + g % (x0 )(x − x0 ) + o(1)
Exercise 9.3 Try to deduce formula (iii) from
f (x0 ) + f % (x0 )(x − x0 ) + o(1)

. .
g(x0 ) + g % (x0 )(x − x0 ) + o(1)
As a hint, you may begin with the identity
1
. = 1 + z + z2 + z3 + · · · = 1 + z + O(z2 ).
1−z
Theorem 9.4 (Chain Rule) Let .f : (a, b) → R be differentiable at a point .x0 ∈

(a, b), and let g be a function defined on a neighborhood of the range .f ((a, b)). If
g is differentiable at the point .f (x0 ), then .g ◦ f is differentiable at .x0 , and
. (g ◦ f )% (x0 ) = g % (f (x0 ))f % (x0 ). (9.4)

9.1 Rules of Differentiation, or the Algebra of Calculus 107
Proof From Definition 9.1, we know that there exists function .σ and .τ such that
σ (x) = o(1) as .x → x0 , .τ (y) = o(1) as .y → f (x0 ), and
.
f (x) = f (x0 ) + f % (x0 )(x − x0 ) + σ (x)|(−x0)

.
g(y) = g(f (x0 )) + g % (f (x0 ))(y − f (x0 ) + τ (y)(y − f (x0 )).
Then
g ◦ f (x) = g(f (x0 )) + g % (f (x0 ))(f (x) − f (x0 )) + τ (f (x)(f (x) − f (x0 ))
.
= g(f (x0 )) + g % (f (x0 ))(f % (x0 ) + σ (x)|x − x0 |)

+τ (f (x)(f (x) − f (x0 ))
= g(f (x0 )) + g % (f (x0 ))(f % (x0 ) + σ (x)|x − x0 |)
+τ (f (x))(f % (x0 )(x − x0 ) + σ (x)(x − x0 ))
= g(f (x0 )) + g % (f (x0 )f % (x0 )(x − x0 ) +
+ ([. . .] σ (x) + [. . .] τ (f (x))) (x − x0 ).
As .x → x0 , it is immediate to check that .[. . .] σ (x) + [. . .] τ (f (x)) → 0, and the

conclusion follows.

Important: Warning
The Calculus “proof” of the Chain Rule goes as follows:
g(f (x)) − g(f (x0 )) g(f (x)) − g(f (x0 )) f (x) − f (x0 )
. =
x − x0 f (x) − f (x0 ) x − x0
→ g % (f (x0 ))f % (x0 ).
There is a subtle flaw in this computation, since division by .f (x) − f (x0 ) is

legitimate only if .f (x) = f (x0 ) in a neighborhood of .x0 . Unfortunately the
assumptions of the Theorem do not ensure that this additional condition is satisfied
by f . There is a way out, but we do not emphasize this approach, since it cannot be
generalized to higher dimension.
Exercise 9.4 Provide a proof of the Chain Rule according to Theorem 9.2.
Theorem 9.5 (Differentiation of the Inverse Function) Suppose that f is an
invertible function on an interval .(a, b). If f is differentiable at a point .x0 ∈ (a, b)
and .f % (x0 ) = 0, then .f −1 is differentiable at .y0 = f (x0 ). Moreover,
1
Df −1 (y0 ) =
. .
f % (x 0)
Fig. 9.1 Differentiating the

inverse function
Proof Since f is continuous and invertible on an interval, its inverse function .f −1

is continuous on the range of f . Then
f −1 (y) − f −1 (y0 ) 1
. lim = lim
y→y0 y − y0 x→x0 f (x)−f (x0 )
x−x0
1
= .
f % (x0 )

The necessity of all the assumptions should be clear from Fig. 9.1.
Example 9.1
1. The function f defined by
x sin x1 if x = 0
f (x) =
.
0 otherwise
is differentiable at any .x = 0: indeed
1 1 1
f % (x) = sin
. − cos .
x x x
However
f (x) − f (0) 1
. = sin
x−0 x
does not converge as .x → 0. Hence f is not differentiable at .x = 0.

2. The function f defined by
x 2 sin x1 if x = 0
f (x) =
.
0 otherwise
9.2 Mean Value Theorems 109
is differentiable at .x = 0, since
f (x) − f (0) 1
. = x sin ,
x−0 x
and .0 ≤ |x sin(1/x)| ≤ x for every x. Therefore .f % (0) = 0.

Exercise 9.5 Suppose that .f : (a, b) → R is differentiable at a point a. Show that
f (a + h) − f (a − h)
f % (a) = lim
. .
h→0 2h
Provide an example of a function such that the limit on the right-hand side exists,
but the function is not differentiable at a.
9.2 Mean Value Theorems
The derivative is a local object that can provide global properties of functions. This
is essentially the basis of mean value theorems.
Theorem 9.6 (Fermat) Let f be a function defined on the interval .[a, b]. If f has
a local maximum or a local minimum at a point .x0 ∈ (a, b), and if .f % (x0 ) exists,
then .f % (x0 ) = 0.
Proof Suppose that .x0 is a local maximum of f , so that there exists .δ > 0 such that
a < x0 − δ < x0 < x0 + δ < b. If .x0 − δ < x, x0 , then
.
f (x) − f (x0 )
. ≥ 0,
x − x0
since f attains a local maximum at .x0 . Letting .x → x0 in the last inequality, we get
f % (x0 ) ≥ 0. Similarly, if .x0 < x < x0 + δ, then
.
f (x) − f (x0 )
. ≤ 0,
x − x0
and letting .x → x0 we get .f % (x0 ) ≤ 0. Necessarily .f % (x0 ) = 0. The proof for a

local minimum reduces to the previous one by considering .−f instead of f .

Theorem 9.7 (Cauchy) Let f and g be two functions defined on .[a.b], which are
differentiable on .(a, b) and continuous on .[a, b]. Then there exists a point .c ∈ (a, b)
such that
[f (b) − f (a)]g % (c) = [g(b) − g(a)]f % (c).

.
Proof Let us introduce the function h defined on .[a, b] by
h(x) = [f (b) − f (a)]g % (x) − [g(b) − g(a)]f % (x).

.
obviously h is continuous and differentiable on .(a, b), and .h(a) = h(b). It remains
to prove that the derivative of h vanishes somewhere inside .(a, b). If h turns out to
be constant on .[a, b], then the proof is complete.
If .h(x) > h(a) for some .x ∈ (a, b), then h must attain a global maximum inside
%
.(a, b), and at this point .h vanishes by Theorem 9.6.
If .h(x) < h(a) for some .x ∈ (a, b), then h must attain a global minimum inside
%
.(a, b), and at this point .h vanishes again by Theorem 9.6.

The simple choice .g = id is surprisingly important: see Fig. 9.2.
Theorem 9.8 (Lagrange) Let f be a function defined on .[a.b], which is differen-
tiable on .(a, b) and continuous on .[a, b]. Then there exists a point .c ∈ (a, b) such
that
.[f (b) − f (a)]g % (c) = (b − a)f % (c).
Corollary 9.2 (Monotonicity) Suppose f is differentiable on .(a, b).

(a) If .f % ≥ 0 on .(a, b), then f is monotonically increasing on .(a, b).
(b) If .f % ≤ 0 on .(a, b), then f is monotonically decreasing on .(a, b).
(c) if .f % = 0 identically on .(a, b), then f is constant on .(a, b).
0 1
Fig. 9.2 Lagrange’s theorem

9.2 Mean Value Theorems 111
Proof For any points .x1 and .x2 in .(a, b), Theorem 9.8 yields .f (x2 ) − f (x1 ) =
(x2 − x1 )f % (c) for some c between .x1 and .x2 . It is now immediate to conclude
according to the sign of .f % .

Exercise 9.6 Suppose that f is a differentiable function such that .f % (x) = f (x)
for each .x ∈ R. If .f (0) = 1, prove that .f (x) = ex for each .x ∈ R.
Exercise 9.7 Let f be a differentiable function on .R with

L = sup |f % (x)| x ∈ R < 1.
.
(a) Fix any .s0 ∈ R, and define .sn = f (sn−1 ) for each .n = 1, 2, . . . Prove that the
sequence .{sn }n is convergent. Hint: show that .{sn }n is a Cauchy sequence.
(b) Prove the Banach-Caccioppoli Fixed Point Theorem: there exists a point .x ∈ R
such that .f (x) = x.
Mean value theorems are typically used in Calculus to derive criteria for the
existence of limits. The well-known result which goes under the name of De
l’Hospital is the most celebrated one.1 We follow [2] for the proof.
Theorem 9.9 (De l’Hospital) Suppose f and g are differentiable on .(a, b), and
g % (x) = 0 for all .x ∈ (a, b), where .−∞ ≤ a < b ≤ +∞. Suppose
.
f % (x)
. lim = A, (9.5)
x→a g % (x)
where .A ∈ R̃. If either
. lim f (x) = lim g(x) = 0 (9.6)

x→a x→a
or
. lim g(x) = +∞, (9.7)

x→a
then
f (x)
. lim = A. (9.8)
x→a g(x)
An analogous statement holds as .x → b.

We remark that A may be infinite.
1We write De l’Hospital since this is the ancient and original name. Nowadays it is customary to
write De l’Hôpital.
Proof Let us start with the case .−∞ ≤ A < +∞. Let .q > A be any real number,
and choose r such that .A < r < q. By (9.5) there exists a point .c ∈ (a, b) such that
.a < x < c implies
f % (x)
. < r. (9.9)
g % (x)
If .a < x < y < b, Theorem 9.7 yields a point .t ∈ (x, y) such that
f (x) − f (y) f % (t)

. = % < r. (9.10)
g(x) − g(y) g (t)
Suppose that (9.6) holds. When .x → a in (9.10) we see that .a < y < c implies
f (y)
. ≤r<q (9.11)
g(y)
Suppose now that (9.7) holds. We fix y in (9.10) and select .c1 ∈ (a, y) such that
g(x) > g(y) and .g(x) > 0 for every .x ∈ (a, c1 ). Then it follows from (9.10) that
.
f (x) g(y) f (y)

. <r −r + (9.12)
g(x) g(x) g(x)
for every .x ∈ (a, c1 ). Letting .x → a in (9.12) we see that the right-hand side of
(9.12) converges to r, and therefore there exists a point .c2 ∈ (a, c1 ) such that
f (x)
. ≤r<q (9.13)
g(x)
for every .x ∈ (a, c2 ). Equations (9.11) and (9.13) imply that .f (x)/g(x) < q for
every .x ∈ (a, c2 ).
If .−∞ < A ≤ +∞, a completely similar argument shows that, given any .p < A,
there exists a point .c3 such that .p < f (x)/g(x) for every .x ∈ (a, c3 ). Since p and
q are arbitrary, we have proved that .f (x)/g(x) → A as .x → a.

Exercise 9.8 For every .x ∈ R, consider the functions
f (x) = x + sin x cos x

.
g(x) = esin x f (x).
(a) Show that .limx→+∞ f (x) = limx→+∞ g(x) = +∞.

(b) Show that .f % (x) = 2 cos2 x and .g % (x) = esin x cos x (2 cos x + f (x)).
− sin x
(c) Show that .f % (x)/g % (x) = 22ecos x+fcos(x)x if .cos x = 0 and .x > 3.
9.3 The Intermediate Property for Derivatives 113
2e− sin x cos x

(c) Show that .limx→+∞ 2 cos x+f (x) = 0 and yet .limx→+∞ f (x)/g(x) does not
exist.
This exercise shows that the assumption “.g % (x) = 0” is necessary in Theorem 9.9.
Remark 9.4 Calculus books often write that De l’Hospital’s theorem is a tool for the
analysis of indeterminate forms .[0/0] and .[∞/∞]. As we have seen, no condition
on f is needed when .g(x) → +∞ as .x → a.
Exercise 9.9 Suppose that .f (x) + f % (x) → L as .x → +∞. Prove that .f (x) → L
as .x → a and .f % (x) → 0 as .x → +∞. Hint: write .f (x) = ex f (x)/ex , and remark
that .ex → +∞ as .x → +∞. De l’Hospital’s theorem applies even if we have no
information about .ex f (x) as .x → +∞.
9.3 The Intermediate Property for Derivatives
Although a differentiable function may have a discontinuous derivative, it is

interesting that derivatives always have the intermediate value property.
Theorem 9.10 (Darboux) Suppose that f is a real differentiable function on
[a, b], and suppose that .f % (a) < λ < f % (b). Then there exists a point .ξ ∈ (a, b)
.
such that .f % (ξ ) = λ.
Proof We define .g(x) = f (x) − λx. By assumption .g % (a+) < 0 and .g % (b−) >
0. It follows that .g(t1 ) < g(a) and .g(t2 ) < g(b) for some .t1 , .t2 in .(a, b). As a
consequence, the function g must attain its minimum at some point .ξ ∈ (a, b). We
already know that .g % (ξ ) = 0, and thus .f % (ξ ) = λ.

Example 9.2 Define the polynomial .P (x) = (x 2 − 1)2 . Let .f : [0, 1] → R be the
function such that .f (0) = 0 and
1 1 1
f (x) =
. P (2n(n + 1)x − 2n − 1) if ≤x≤ .
n3/2 n+1 n
Clearly f is a differentiable function, but .f % is not continuous, and event not

bounded on .[0, 1]. Indeed .f+% (0) = 0, but .f (bn ) → +∞ at the points
4n + 1
bn =
. → 0.
4n(n + 1)
Nevertheless, Theorem 9.10 applies, and .f % attains every positive value .γ on every
interval .[0, bn ].
9.4 Derivatives at End-Points
The idea of linearization is fruitful at inner points of the domain of definition: the
function can be identified, at an infinitesimal scale, with a linear function. It is
nonetheless convenient, from time to time, to extend the definition of derivative
at end-points.
Definition 9.2 Let .f : [a, b] → R be a function. We say that f is differentiable at
a if the limits
f (x) − f (x0 )
f % (a+) = lim
.
x→x0+ x − x0
exists as a real number. In this case, we call .f % (a+) the right-derivative of f at a.

Similarly we define the left derivative of f at b.
Unfortunately, several fundamental results of differential calculus do not extend
to end-point derivatives. As an example, we propose the function .f : [a, b] → R
such that .f (x) = mx + q, where .m = 0 and q are real numbers. It is easy to check
that a and b are global extremum points of f (their nature depends on the sign of
m), but .f % (a+) = m = f % (b−) are different than zero. In other words, Fermat’s
Theorem does not hold at end-points.
9.5 Derivatives of Derivatives
A nice feature of derivatives in one variable is that we can easily differentiate

derivatives. We will see that this requires much more attention in higher dimension,
since the derivative is no longer a real number.
Definition 9.3 Suppose that a function f is defined on an interval .(a, b), and
that the derivative .f % of f exists at every point of .(a, b). Hence the function
% % %
.f : (a, b) → R is defined in such a way that .f : x → f (x) for every .x ∈ (a, b).
%
We say that f is twice differentiable at .x ∈ (a, b) if .f is differentiable at x. In this
case we denote by .f %% (x) or .D 2 f (x) or .d2 f (x) the derivative of .f % at x, and call it
the second derivative of f at x.
More generally, if f is differentiable n times at every point of .(a, b), we say that
f is differentiable .(n + 1)-times at .x ∈ (a, b) if the function .f (n) is differentiable
at x. In this case we denote by .f (n+1) (x), or .D n+1 f (x) or .dn+1 f (x) the derivative
of .f (n) at x, and call it the derivative of f of order .n + 1 at x.
Definition 9.4 (Regularity Classes) Let .f : (a, b) → R be a function. We write
f ∈ C 0 (a, b) if f is continuous on .(a, b). For .n ∈ N, we write .f ∈ C n (a, b)
.
if the derivatives .f % , .f %% , . . . , .f (n−1) exist on .(a, b), and if .f n exists and is a

continuous function on .(a, b). We formally write .f ∈ C ∞ (a, b) to mean that
9.5 Derivatives of Derivatives 115

f ∈ ∞
.
n
n=0 C (a, b). Hence, a function is of class .C
∞ if and only if it can be
differentiated as many times as we please.

We now attach a very specific polynomial to every function with a high degree
of differentiability.
Definition 9.5 (Taylor Polynomial) Let .f : (a, b) → R be n-times differentiable
at a point .x0 ∈ (a, b). The Taylor polynomial of degree n at .x0 is defined to be

n
f (k) (x0 )
P (n, x0 ; x) =
. (x − x0 )k
k!
k=0
f %% (x0 )
= f (x0 ) + f % (x0 )(x − x0 ) + (x − x0 )2 + · · ·
2!
f (n) (x0 )
··· + (x − x0 )n . (9.14)
n!
Taylor polynomials express the local behavior of functions, and generalize the
concept of linear approximation which was introduced in the definition of the first
derivative.
Theorem 9.11 (Local Polynomial Approximation) Let .f : (a, b) → R be n-
times differentiable at a point .x0 ∈ (a, b), and let .P (n, x0 ; ·) be its Taylor
polynomial of degree n. Then
f (x) = P (n, x0 ; x) + (x − x0 )n o(1) as x → x0 .

. (9.15)
If, in addition, .f (n+1) (x0 ) exists, then
ζ(x) f (n+1) (x0 )

. lim = .
x→x0 x − x0 (n + 1)!
Proof We consider the function
f (x) − P (n, x0 ; x)
. ζ(x) = , (9.16)
(x − x0 )n
defined for .x = x0 . It is easy to check that all the derivatives of order .1 ≤ j ≤ n − 1

of the numerator of .ζ vanish at .x0 . We apply Theorem 9.9 .n − 1 times to (9.16), to
get
f (n−1) (x) − f (n−1) (x0 ) − (x − x0 )f (n) (x0 )

. lim ζ (x) = lim , (9.17)
x→x0 x→x0 n!(x − x0 )
provided the last limit exists. By definition,
f (n−1) (x) − f (n−1) (x0 )

. lim = f (n) (x0 ),
x→x0 x − x0
thus from (9.17) we deduce .limx→x0 ζ(x) = 0. We can define .ζ(x0 ) = 0, so that .ζ
becomes a continuous function on .(a, b). By applying again Theorem 9.9 n times,
we finally see that
ζ (x) 1 f (n) (x) − f (n) (x0 ) f (n+1) (x0 )

. lim = lim = .
x→x0 x − x0 (n + 1)! x→x0 (x − x0 ) (n + 1)!

Theorem 9.12 (Lagrange Remainder) Let .f : (a, b) → R be .n + 1 times
differentiable on .(a, b), and let .x0 ∈ (a, b). For each .x ∈ (a, b), .x = x0 , there
exists a point .ξ between .x0 and x such that
f (n+1) (ξ )
.f (x) = P (n, x0 ; x) + (x − x0 )n+1 .
(n + 1)!
Proof Suppose without loss of generality that .x0 < x. We define .F : [x0 , x] → R,
f %% (t) f (n) (t)

F (t) = f (x) − f (y) − f % (t)(x − t) −
. (x − t)2 − · · · − (x − t)n .
2! n!
Then
f (n+1) (t)
F % (t) = −
. (x − t)n .
n!
Next we introduce the function .G : [x0 , x] → R,
(x − t)n+1
.G(t) = .
(n + 1)!
We have .F (x) = G(x) = 0, and .F % (t)/G% (t) = f (n+1) (t). We now apply
Theorem 9.7 to F and G on .[x0 , x], and find a point .ξ between .x0 and x such
that
F (x0 ) F (x) − F (x0 ) F % (ξ )

. = = % = f (n+1) (ξ ).
G(x0 ) G(x) − G(x0 ) G (ξ )

9.6 Convexity 117
[ 1 + (1 − ) 2]
( 1) + (1 − ) ( 2)
( )
1 1 + (1 − ) 2 2
Fig. 9.3 A convex function
9.6 Convexity
Convexity is a fundamental property in mathematical analysis. Most Calculus books

propose the definition of convexity as a property of the second derivative. We are
going to see that this is just the top of the iceberg.
Definition 9.6 Let I be an interval2 (open, closed, bounded or unbounded), and let
.f : I → R be a function. We say that f is a convex function on I , if for each .x1 ∈ I ,
.x2 ∈ I , and for all real numbers .λ ≥ 0, .μ ≥ 0 such that .λ + μ = 1, there results
f (λx1 + μx2 ) ≤ λf (x1 ) + μf (x2 ).

. (9.18)
The function f is concave on I if and only if the function .−f is convex on I .

Exercise 9.10 Prove that a function f is convex on an interval I if and only if for
every .x1 ∈ I , .x2 ∈ I and .λ ∈ [0, 1] there results
f (λx1 + (1 − λ)x2 ) ≤ λf (x1 ) + (1 − λ)f (x2 ).

.
See Fig. 9.3.
Remark 9.5 The crucial fact is that .λx1 + μx2 must be an element of I , as soon as
x1 and .x2 belong to I and .λ + μ = 1. It is easy to check that this is actually correct,
.
since I is an interval.
2 Hence I is characterized by the following property: if .x ∈ I , .y ∈ I and .x < z < y, then .z ∈ I .

Let us manipulate the convexity inequality (9.18). Let x be any point between .x1
and .x2 . We set .x = λx1 + μx2 . Since .λ + μ = 1, we see that
x2 − x x − x1
λ=
. , μ= .
x2 − x1 x2 − x1
Hence (9.18) is equivalent to
x2 − x x − x1
f (x) ≤
. f (x1 ) + f (x2 ). (9.19)
x2 − x1 x2 − x1
By symmetry, we can always suppose that .x2 > x1 . Then we get
(x2 − x1 )f (x) ≤ (x2 − x)f (x1 ) + (x − x1 )f (x2 ).

. (9.20)
Writing .x2 − x = (x2 − x1 ) − (x − x1 ) in (9.20), we find
f (x) − f (x1 ) f (x2 ) − f (x1 )

. ≤ . (9.21)
x − x1 x2 − x1
Writing .x2 − x1 = (x2 − x) + (x − x1 ) in (9.20) we find
f (x1 ) − f (x) f (x2 ) − f (x)

. ≤ . (9.22)
x1 − x x2 − x
Comparing (9.18), (9.21) and (9.22), we have proved

Theorem 9.13 The function f is convex on the interval I if and only if, for every
x0 ∈ I , the map
.
f (x) − f (x0 )
x →
.
x − x0
is (defined for .x = x0 and) monotonically increasing.

We thus see that convexity is just another way of stating that the incremental ratio
is an increasing function. Since monotone functions always have one-sided limits,
we deduce
Corollary 9.3 A convex function f defined on an interval .(a, b) is left- and right-
differentiable at every .x0 ∈ (a, b). Moreover .f % (x0 −) ≤ f % (x0 +).
If we remember (9.22) and let .x → x1 and then .x → x2 , we see that
f (x2 ) − f (x1 )
f % (x1 +) ≤
. ≤ f % (x2 −). (9.23)
x2 − x1
We can finally relate convexity to derivatives.

9.7 Problems 119
Theorem 9.14 Let f be a differentiable function in the interval .[a, b]. A necessary
and sufficient condition for f to be convex is that .f % be monotonically increasing.
Proof If f is convex, then (9.23) implies .f % (x1 ) ≤ f % (x2 ), so that .f % is increasing.
In the other direction, we remark that convexity is equivalent to (9.22) for all points
%
.x1 , x and .x2 such that .x1 < x < x2 . If .f is increasing, then there exists points .ξ1
and .ξ2 such that .x1 < ξ1 < x < ξ2 < x2 and
f (x1 ) − f (x) f (x2 ) − f (x)

. = f % (ξ1 ), = f % (ξ2 ).
x1 − x x2 − x
Since .f % (x1 ) ≤ f % (x2 ), the conclusion follows.

In particular, it is indeed true that convex functions are those functions whose
second derivative is positive, but there is no need to restrict our definitions to twice
differentiable functions.
Corollary 9.4 Let f be twice differentiable in .[a, b]. The function f is convex if
and only if .f %% (x) ≥ 0 for every x.
Proof This follows immediately from the characterization of increasing functions
in terms of the first derivative, see Corollary 9.2.

Example 9.3 Prove that the function .x → |x| is convex on .I = R. Of course this
conclusion would be meaningless if we had defined convex functions through the
sign of the second derivative.
9.7 Problems
9.1 Suppose that f is differentiable at the point x0 . Prove that

α β
. lim n f x0 + −f x0 − = (α + β)f % (x0 ).
n→+∞ n n
Give an example to show that the existence of the previous limit does not imply the
differentiability of f at x0 .
9.2 Suppose that f is differentiable at the point x0 . let {hn }n and {kn }n be two
nonincreasing sequences which converge to x0 . Prove that
f (x0 + hn ) − f (x0 − kn )
. lim = f % (x0 ).
n→+∞ hn + kn
Give an example to show that the existence of the previous limit does not imply the
differentiability of f at x0 .
9.3 Suppose that f % is continuous on an interval [a, b]. Prove that for every ε > 0
there exists δ > 0 such that

f (t) − f (x)
− f (x) < ε
%
.
t −x
whenever 0 < |t − x| < δ and x ∈ [a, b], t ∈ [a, b].

9.4 Let f be differentiable on [a, b]. Suppose that 0 < m ≤ f % (x) ≤ M for each
x ∈ [a, b], and that f (a) < 0 < f (b). Given x1 ∈ [a, b], define a sequence {xn }n
by
f (xn )
xn+1 = xn −
.
M
for n = 1, 2, 3, . . . Prove that {xn }n converges to a limit x0 such that f (x0 ) = 0.
Furthermore, prove that
f (x1 ) m n
. |xn+1 − xn | ≤ 1− .
m M
9.5 Suppose f is a real-valued function defined on the half-line (a, +∞). Suppose
that f is twice differentiable on (a, +∞), and define
M0 = sup |f (x)|
.
x>a
M1 = sup |f % (x)|
x>a
M2 = sup |f %% (x)|.
x>a
Prove that M12 ≤ M0 M2 as follows: for each h > 0 deduce from Taylor’s expansion
that there exists ξ ∈ (x, x + 2h) such that
f (x + 2h) − f (x)
f % (x) =
. − hf % (ξ ).
2h
Therefore
M0
|f % (x)| ≤ hM1 +
. .
h
Now optimize the right-hand side with respect to h > 0.
9.6 If f is twice differentiable on (0, +∞), f %% is bounded and limx→+∞ f (x) =
0, prove that limx→+∞ f % (x) = 0. Hint: consider the limit a → +∞ in the previous
problem.
9.7 Problems 121
9.7 (a) Prove that for each x > 0, x = 1, we have
x−1
. < log x < x − 1.
x
(b) For each j ∈ N, j > 1, prove that
j +1 1 j
. log < < log .
j j j −1
(c) For each n ∈ N, k ∈ N, n > 1, prove that
1
kn
1 k
. log k + < < log k +
n j n−1
j =n
and

kn
i
. lim = log k.
n→+∞ j
j =n
(d) Deduce from (c) and from the identity

2n
(−1)j +1
2n
1
n
1
. = −2
j j 2k
j =1 j =1 k=1
that
∞
(−1)j +1
. = log 2.
j
j =1
9.8 If x > 1, x = e, prove that there exists one and only one number f (x) > 0
such that f (x) = x and
x f (x) = (f (x))x .
.
log x log y
Hint: x y = y x if and only if x = y .
9.8 Comments
The standard definition of derivative as the limit of the incremental ratio should
not be considered as the one used by mathematicians from the beginning of
Calculus. They would rather use a principle of disappearing quantities which
roughly correspond to an expansion of functions at first order as in
f (t + Δt) = f (t) + f % (t)Δt + f %% (t)(Δt)2 ≈ f (t) + f % (t)Δt.

.
In this sense, the one-dimensional derivative has progressively lost its definition as
a linearization procedure in favor of an iconic limit:
f (t + Δt) − f (t)
f % (t) = lim
. .
Δt →0 Δt
There are several good reasons to define the derivative as a linearization, and the
most important one is that the derivative of a function of several variables is not a
number.
The theory of convex functions is a long but elementary exercise, in the case
of functions of a single real variable. The topic becomes much more exciting in
higher dimensions, where intervals must be replaced by convex sets and a new fact
comes into play: it is possible to draw conclusion about a function on a convex set
by assuming a property of that function on every straight line. The interested reader
may start from [1].
References
1. R.T. Rockafellar, Convex Analysis (Princeton University Press, 1970)

Chapter 10
Riemann’s Integral
Abstract The basic theory of definite integration can be named after Bernhard
Riemann. In this chapter we will propose a rigorous introduction to it. Although
we have in mind Rudin’s lucid chapter in his Principles of Mathematical Analysis,
we prefer to avoid the additional complication of the Stieltjes generalization. By the
way, a later chapter will show that the much more flexible integral of Lebesgue can
be presented without too much effort.
10.1 Partitions and the Riemann Integral
We will systematically consider bounded functions defined on bounded

intervals.
This is by far the worst weakness of the Riemann integral. The rough idea is to
construct finite sums of values of a function at suitable points, and then pass to the
limit is a suitable sense. The next definition gives a name to the selection of the
suitable points.
Definition 10.1 A partition P of a closed and bounded interval .[a, b] is a finite set
of points .x0 , .x1 , . . . , .xn such that .a = x0 < x1 < x2 < . . . < xn−1 < xn = b. We
write .xi = xi − xi−1 for .i = 1, . . . , n.
Let .f : [a, b] → R be a bounded function. To any partition P of .[a, b] we attach
the quantities
Mi = sup {f (x) | xi−1 ≤ x ≤ xi }

.
mi = inf {f (x) | xi−1 ≤ x ≤ xi }
124 10 Riemann’s Integral
Fig. 10.1 Riemann sums
25
20
15
10
1 2 3 4 5

n
U (f, P ) =
. Mi xi
i=1

n
L(f, P ) = mi xi ;
i=1
see Fig. 10.1.

Definition 10.2 The upper and the lower Riemann integral of f on .[a, b] are
defined respectively by
b b
. f dx = inf U (f, P ), f dx = sup L(f, P )
a a
where inf and sup are taken over all partitions P of .[a, b].
Remark 10.1 Since .mi ≤ Mi , we see that .L(f, P ) ≤ U (f, P ) for every partition
P . Since f is bounded, say .m ≤ f ≤ M on .[a, b], then .m(b − a) ≤ L(f, P ) ≤
U (f, P ) ≤ M(b − a). As a consequence, the upper and the lower integrals of f are
finite. Of course this is the reason why boundedness cannot be dispensed with.
10.1 Partitions and the Riemann Integral 125
Definition 10.3 A bounded function f is R-integrable on .[a, b] if and only if
b b
. f dx = f dx.
a a
In this case the common value of the upper and the lower integral is denoted by
!b
.
a f dx, and it is called the Riemann integral of f on .[a, b].
Remark 10.2 There are good reasons to criticize the symbol associated to the
!b
Riemann integral. Indeed, the letter x in . a f dx is completely useless. The symbol
!b
.
a f would be a natural choice, but we prefer to use the traditional notation.
Unlike the derivative, the Riemann integral is a global object, in the sense that
it involves the values of the function f on the whole interval .[a, b]. It is a matter
of fact that Definition 10.3 is not particularly concrete. We need to deploy some
general condition for integrability.
Definition 10.4 A partition .P ∗ is a refinement of a partition P of .[a, b], if .P ⊂ P ∗ .
If .P1 and .P2 are two partitions of .[a, b], their union .P ∗ = P1 ∪ P2 is called the
common refinement of .P1 and .P2 .
Exercise 10.1 Prove that the common refinement of two partitions .P1 and .P2 is the
smallest partition which refines both .P1 and .P2 .
Theorem 10.1 If .P ∗ is a refinement of P , then
.L(f, P ) ≤ L(f, P ∗ ), U (f, P ∗ ) ≤ U (f, P ).
Proof We suppose first that .P ∗ contains just one point more that P . If this point is
∗ ∗
.x , for some index i we must have .xi−1 < x < xi . Let
.w1 = inf{f (x) | xi−1 ≤ x ≤ x ∗ }. (10.1)

∗
w2 = inf{f (x) | x ≤ x ≤ xi }. (10.2)
By set inclusion, .mi ≤ w1 and .mi ≤ w2 , hence
L(f, P ∗ ) − L(f, P ) = w1 [x ∗ − xi−1 ] + w2 [xi − x ∗ ] − mi [xi − xi−1 ]

.
= (w1 − mi )[x ∗ − xi−1 ] + (w2 − mi )[xi − x ∗ ] ≥ 0.
In the general case, a finite number of points are added to those of P . We just repeat
the same construction for each of them, and we conclude. The proof of the inequality
∗
.U (f, P ) ≤ U (f, P ) is similar, and we omit the details.

!b !b
Corollary 10.1 . f dx ≤ a f dx.
a
Proof Let .P1 and .P2 be arbitrary partitions of .[a, b], and let .P ∗ their common
refinement. As we have just seen,
L(f, P1 ) ≤ L(f, P ∗ ) ≤ U (f, P ∗ ) ≤ U (f, P2 ).

.
!b
Taking the supremum over .P1 , we see that . f dx ≤ U (f, P2 ). Taking now the
a
infimum over .P2 we conclude the proof.

Theorem 10.2 (Integrability Condition) A bounded function f is R-integrable on
[a, b] if and only if for every .ε > 0 there exists a partition .Pε of .[a, b] such that
.
U (f, Pε ) − L(f, Pε ) < ε.

. (10.3)
Proof If P is any partition, then
b b
L(f, P ) ≤
. f dx ≤ f dx ≤ U (f, P ).
a a
!b !b !b !b
Then (10.3) implies .0 ≤ a f dx − f dx < ε, and . a f dx = f dx because
a a
.ε > 0 is arbitrary.
Conversely, let f be integrable, and let .ε > 0 be fixed. There exist two partitions
.P1 and .P2 such that
b ε b ε
.U (f, P2 ) < f dx + , L(f, P1 ) > f dx − .
a 2 a 2
Choosing .Pε as the common refinement of .P1 and .P2 , then
b ε
U (f, Pε ) ≤ U (f, P2 ) <
. f dx + < L(f, P1 ) + ε ≤ L(f, Pε ) + ε,
a 2
and (10.3) is proved.

Exercise 10.2
(a) Prove that a bounded function f is integrable on .[a, b] if and only if there exists
a sequence .{Pn }n of partitions of .[a, b] such that
. lim (U (f, Pn ) − L(f, Pn )) = 0,

n→+∞
!b
and in this case . a f dx = limn→+∞ U (f, Pn ) = limn→+∞ L(f, Pn ). Hint:
apply (10.3) with .ε = εn → 0 as .n → +∞.
10.1 Partitions and the Riemann Integral 127
(b) For each n, let .Pn be the partition of .[0, 1] into n equally spaced points. Find a
closed formula for .U (f, Pn ) and .L(f, Pn ) in case .f (x) = x. Hint: first prove
by induction that .1 + 2 + . . . + n = n(n + 1)/2.
!1
(c) Deduce that .f (x) = x is integrable on .[0, 1], and compute . 0 x dx.
We can now relate our definition of the Riemann integral to the usual approach
of Calculus via Riemann sums. In undergraduate introductions to the integral, it is
customary to select arbitrary points between two consecutive nodes of a partition,
so that sums like

n
. f (ti )xi
i=1
are considered for points .ti ∈ [xi−1 , xi ].

Theorem 10.3
(i) If (10.3) holds for a partition P and a value of .ε > 0, then (10.3) holds for
any refinement of P , with the same value of .ε.
(ii) If (10.3) holds for a partition .P = {x0 , x1 , . . . , xn } and if .si and .ti are points
belonging to .[xi−1 , xi ], then

n
. |f (si ) − f (ti )| xi < ε.
i=1
(iii) If f is R-integrable on .[a, b] and the assumptions of (ii) hold, then

n
b

. f (ti )xi − f dx < ε.
a
i=1
Proof Part (i) follows immediately from Theorem 10.1. Let us prove part (ii). Of
course .f (si ) and .f (ti ) lie in .[mi , Mi ], so that .|f (si )−f (ti )| ≤ Mi −mi . This yields

n
. |f (si ) − f (ti )|xi ≤ U (f, P ) − L(f, P ) < ε.
i=1
Part (iii) follows from the relations

n
L(f, P ) ≤
. f (ti )xi ≤ U (f, P )
i=1
b
L(f, P ) ≤ f dx ≤ U (f, P ).
a
( )
= 0 −1 +1 =
Fig. 10.2 Intuition of Theorem 10.3

Remark 10.3 Part (iii) actually says that the Riemann integral can be approximated
by a Riemann sum as soon as it exists, see Fig. 10.2.
10.2 Integrable Functions as Elements of a Vector Space
Functions can be added pointwise, and pointwise multiplied by constants. This fact
induces a vector space structure to the class of R-integrable functions, as the next
result states.
10.2 Integrable Functions as Elements of a Vector Space 129
Theorem 10.4
(a) If .f1 and .f2 are R-integrable on .[a, b] and if c is a real number, then .f1 + f2
and .cf1 are R-integrable; moreover
b b b
. (f1 + f2 ) dx = f1 dx + f2 dx,
a a a
b b
cf1 dx = c f1 dx.
a a
!b !b
(b) If .f1 ≤ f2 on .[a, b], then . a f1 dx ≤ a f2 dx.
(c) If f is R-integrable on .[a, b] and if .a < c < b, then f is R-integrable on .[a, c]
!b !c !b
and on .[c, b], and . a f dx = a f dx + c f dx.
Proof For .f = f1 + f2 and any partition P of .[a, b], we notice that
L(f1 , P ) + L(f2 , P ) ≤ L(f, P ) ≤ U (f, P ) ≤ U (f1 , P ) + U (f2 , P ).

.
Let .ε > 0, and choose partitions .P1 , .P2 so that
.U (f1 , P1 ) − L(f2 , P2 ) < ε,
and
. U (f2 , P2 ) − L(f2 , P2 ) < ε.
Call P the common refinement of .P1 and .P2 . Then .U (f, P ) − L(f, P ) < 2ε.
Furthermore
b
U (fj , P ) <
. fj dx + ε, j = 1, 2
a
Hence
b b b
. f dx ≤ U (f, P ) < f1 dx + f2 dx + 2ε,
a a a
and
b b b
. f dx ≤ f1 dx + f2 dx
a a a
follows from the arbitrariness of .ε. Replacing .f1 and .f2 with .−f1 and .−f2 , we see
!b !b !b
that . a f dx ≥ a f1 dx + a f2 dx. The claim about cf is easier, and we leave it
as an exercise. We have thus proved (a).
To prove (b) we notice that, for every partition P of .[a, b],
L(f1 , P ) ≤ L(f2 , P ),
. U (f1 , P ) ≤ U (f2 , P ).
!b !b !b
Hence . a f1 dx ≤ U (f2 , P ), and thus . a f1 dx ≤ a f2 dx.
To prove (c), write .f = g + h in such a way that .g = 0 on .[c, b], .h = 0 on
!b !c
.[a, c]. It is clear that g and h are R-integrable on .[a, b], and . g dx = a f dx,
!b !b a
while . a h dx = c f dx. The conclusion follows from (a).

Exercise 10.3 Suppose that .f (x) > 0 for each .x ∈ [a, b]. If f is integrable on
!b !b
[a, b], prove that . a f dx > 0. Hint: recall that .L(f, P ) ≤ a f dx ≤ U (f, P ) for
.
each partition P of .[a, b]. What is the sign of .L(f, P )?
10.3 Classes of Integrable Functions
Example 10.1 Dirichlet’s function is defined on [0, 1] by
0 if x ∈ [0, 1] ∩ Q
h(x) =
.
1 otherwise.
Let P be any partition of [0, 1]. Since Q is dense in R, between two consecutive
nodes xi−1 and xi of the partition there are infinitely many rational points, and
infinitely many irrational points. If follows that U (f, P ) = 1 and L(f, P ) = 0.
This clearly shows that h is not R-integrable on [0, 1].
The previous example suggests the following question: are there general proper-
ties that imply the R-integrability of functions?
Theorem 10.5 (Monotonic Functions Are R-integrable) Let f be a bounded
monotonic function on [a, b]. Then f is R-integrable on [a, b].
Proof For the sake of definiteness, we assume that f is increasing. Pick any ε > 0,
and split [a, b] into a number n of equal parts so that (b − a)/n < ε/[f (b) − f (a)].
This gives rise to a partition P of [a, b]. Monotonicity implies that mi = f (xi−1 ),
Mi = f (xi ), so that

n
b−a
U (f, P ) − L(f, P ) =
. [f (xi ) − f (xi−1 )]xi = [f (b) − f (a)] < ε.
n
i=1
The conclusion follows from Theorem 10.2.

Theorem 10.6 (Continuous Functions Are R-integrable) If f is continuous on
[a, b], then f is R-integrable on [a, b].
10.3 Classes of Integrable Functions 131
Proof Pick any ε > 0. By Theorem 8.8 f is uniformly continuous. There exists
δ > 0 such that |f (x) − f (t)| < ε/(b − 1) whenever |x − t| < δ. Let P a partition
of [a, b] such that xi < δ for every i. Clearly Mi − mi < ε/(b − a) for every i,
and thus

n
ε
U (f, P ) − L(f, P ) =
. [Mi − mi ]xi ≤ (b − a) = ε.
b−a
i=1
The conclusion follows from Theorem 10.2.

Remark 10.4 The passage from pointwise continuity to uniform continuity should
not be a surprise: integration is not a local property, and we cannot just sum up
infinitely many local inequalities to get a global inequality.
Too many discontinuity points can prevent integrability, as we have seen in
Example 10.1. However, discontinuous functions may well be integrable.
Theorem 10.7 Suppose f is a bounded function on [a, b] that possesses finitely
many discontinuity points. Then f is R-integrable.
Proof Pick any ε > 0, and write M = supx∈[a,b] |f (x)|. By assumption, the set E
of points at which f is not continuous
is a finite set. Let us cover E by finitely many
intervals [uj , vj ] such that j |vj − uj | < ε. Without loss of generality, we can
choose [uj , vi ] such that each points of E is an interior point of some [uj , vj ].
By removing the open sets (uj , vj ) from [a, b] we get a compact set K, on which
f is uniformly continuous: there exists δ > 0 such that |f (s) − f (t)| < ε whenever
s ∈ K, t ∈ K, |s − t| < δ. Now we construct a partition P = {x0 , . . . , xn } as
follows: each uj is in P . Each vj is in P . No point of any segment (uj , vj ) is in P .
If xi−1 is not one of the uj , then xi < δ.
Since Mi − mi ≤ 2M for every j , and Mi − mi ≤ ε unless xi−1 is one of the uj ,
we see that
U (f, P ) − L(f, P ) ≤ (b − a)ε + 2Mε.

.
The conclusion follows again from Theorem 10.2.

Despite the technicalities of the proof, we strongly urge the reader to understand
the general idea: we remove small neighborhoods of each point of discontinuity,
obtaining a compact subset K. We know that f is integrable on K by continuity.
A finite number of small intervals remains, but we have chosen the size of these
intervals so small that their contribution to the Riemann sum is arbitrarily small.
Theorem 10.8 (Integrability of a Composite Function) Suppose that f is R-
integrable on [a, b], m ≤ f ≤ M on [a, b], φ is continuous on [m, M], and
h(x) = φ(f (x)) on [a, b]. Then h is R-integrable on [a, b].
Proof Let ε > 0. Since [m, M] is a compact set, φ is uniformly continuous. Thus
we get δ > 0 such that δ < ε and |φ(s) − φ(t)| < ε whenever |s − t| ≤ δ in [m, M].
Select a partition P = {x0, . . . , xn } of [a, b] such that
U (f, P ) − L(f, P ) < δ 2 .

. (10.4)
We use the symbols mi and Mi as before, and we denote
m∗i = inf{h(x) | xi−1 ≤ x ≤ xi },

.
Mi∗ = sup{h(x) | xi−1 ≤ x ≤ xi }.
The idea is to split the partition in two parts: a part on which Mi − mi < δ, and
the remaining part. Precisely, for i = 0, . . . , n we say that i ∈ A if Mi − mi < δ,
and i ∈ B if Mi − mi ≥ δ. For i ∈ A we have Mi∗ − m∗i ≤ ε. For i ∈ B, we have
Mi∗ − m∗i ≤ 2K, where K = supx∈[a,b] |φ(x)|. By (10.4) we have

.δ xi ≤ (Mi − mi )xi ≤ δ 2 .
i∈B i∈B
Hence

.U (h, P ) − L(h, P ) = (Mi∗ − m∗i )xi + (Mi∗ − m∗i )xi
i∈A i∈B
≤ (b − a)ε + 2Kδ < ε(b − a + 2K).
Since ε > 0 is arbitrary, the proof is complete.

Corollary 10.2 If f and g are two R-integrable functions on [a, b], then
(a) the product fg is R-integrable;
! !
b b
(b) |f | is R-integrable, and a f dx ≤ a |f | dx.
Proof We take φ(t) = t 2 in Theorem 10.8. This yields that f 2 is R-integrable, and
the identity
4fg = (f + g)2 − (f − g)2

.
shows (a).
Taking φ(t) = |t| in Theorem 10.8 shows that |f | is R-integrable. Let c ∈
!b
{−1, +1} be chosen so that c a f dx ≥ 0. Then

b b b b
. f dx = c f dx = cf dx ≤ |f | dx.

a a a a
This completes the proof of (b).

10.3 Classes of Integrable Functions 133
Exercise 10.4 Let f : E → R be a bounded function on a set E. We define
f + = max{f, 0},
. f − = max{−f, 0},
.M = sup{f (x) | x ∈ E}
m = inf{f (x) | x ∈ E}
M % = sup{|f (x)| | x ∈ E}
m% = inf{|f (x)| | x ∈ E}.
(i) Prove that f = f + − f − , |f | = f + + f − on E. Deduce that 2f − = |f | − f .

(ii) Show that 2 sup{f − (x) | x ∈ E} = M % − m and that 2 inf{f − (x) | x ∈ E} =
m% − M. In particular M − m ≥ M % − m% .
(iii) Use (ii) to prove (b) of Corollary 10.2.
Exercise 10.5 Suppose f and g are continuous functions on [a, b] such that
!b !b
a f dx = a g dx. Prove that there exists a point x ∈ [a, b] such that f (x) = g(x).
Theorem 10.9 (Change of Variable) Let ϕ be a strictly increasing functions from
an interval [A, B] onto an interval [a, b]. Let f be R-integrable on [a, b], and
suppose that ϕ % is R-integrable on [A, B]. Defining g on [A, B] by g(y) = f (ϕ(y)),
!B !b
the function gϕ % is R-integrable on [A, B], and A gϕ % dy = a f dx.
Proof Let ε > 0, and pick a partition Q = {y0 , . . . , yn } of [A, B] such that
U (ϕ % , Q) − L(ϕ % , Q) < ε. The mean value theorem furnishes points ti ∈ [yi−1 , yi ]
such that ϕ(yi ) − ϕ(yi−1 ) = ϕ % (ti )yi for every i = 1, . . . , n.
We observe that any partition Q = {y0 , . . . , yn } of [A, B] corresponds to a
partition P = {x0 , . . . , xn } of [a, b] such that xi = ϕ(yi ). Furthermore the values
taken by f on [xi−1 , xi ] are the same as the values taken by g on [yi−1 , yi ]. In
particular

n
U (f, Q) =
. (sup{g(y) | yi−1 ≤ y ≤ yi }) [ϕ(yi ) − ϕ(yi−1 )]. (10.5)
i=1

n
L(f, Q) = (inf{g(y) | yi−1 ≤ y ≤ yi }) [ϕ(yi ) − ϕ(yi−1 )]. (10.6)
i=1
Let si ∈ [yi−1 , yi ], and observe that

n
. |ϕ % (si ) − ϕ % (ti )|yi < ε
i=1
by Theorem 10.3. If M = sup |f |, it follows from

n
n
. g(si )[ϕ(yi ) − ϕ(yi−1 )] = g(si )ϕ % (ti )yi
i=1 i=1
that
n

n

. g(si )[ϕ(yi ) − ϕ(yi−1 )] − g(si )ϕ % (si )yi ≤ Mε.

i=1 i=1
In particular we get

n
. g(si )[ϕ(yi ) − ϕ(yi−1 )] ≤ U (gϕ % , P ) + Mε
i=1
for all choices of si ∈ [yi−1 , yi ]. In a similar way we also see that

n
%
U (gϕ , P ) ≤
. g(si )[ϕ(yi ) − ϕ(yi−1 )] + Mε
i=1
for all choices of si ∈ [yi−1 , yi ]. Comparing with (10.5) we conclude that

. U (f, Q) − U (gϕ % , P ) ≤ Mε.

Analogously we see that L(f, Q) − L(gϕ % , P ) ≤ Mε, and the conclusion follows.

10.4 Antiderivatives and the Fundamental Theorem
How do we actually compute Riemann integrals? Despite all the definitions and all
the results we have proved, there is just one universal approach: we need to find an
antiderivative.
Definition 10.5 A function F is an antiderivative of a function f on the interval
.[a, b], if F is differentiable on .[a, b], and .F % = f at every point.
! !
It is customary to collect under the symbol . f (or . f (x) dx) the set of all
antiderivatives of f (on an interval that is not specified in the notation). The symbol
.D
−1 f would probably be a better choice, but it is not customary in the literature.
10.4 Antiderivatives and the Fundamental Theorem 135
Remark 10.5 Although legitimate, we do not define antiderivatives of a function on

more general sets like the union of disjoint intervals. The main reason is that the
description of the set of all antiderivatives becomes less explicit, and could induce
mistakes in applications.
Theorem 10.10 Two antiderivatives of the same function differ by a constant on
any interval.
Proof Let .F1 and .F2 be two antiderivatives of a function f on the interval .[a, b].
Since .(F1 − F2 )% = F1% − F2% = f − f = 0 on .[a, b], the function .F1 − F2 is constant
on .[a, b]. The conclusion follows.

Example 10.2 For any choice of the real numbers .c1 and .c2 , the function
c1 if 0 ≤ x ≤ 1
f (x) =
.
c2 if 2 ≤ x ≤ 3
is an antiderivative of the zero function on .[0, 1] ∪ [2, 3]. This shows that
Theorem 10.10 does not hold on domains different than intervals.
Theorem 10.11 (Existence of Antiderivatives) Let f be R-integrable on .[a, b].
For .a ≤ x ≤ b we define
x
. F (x) = f (t) dt.
a
The function F is uniformly continuous on .[a, b]. If f is continuous at a point .x0 ∈

[a, b], then F is also differentiable at .x0 , and there results .F % (x0 ) = f (x0 ).
Proof As an integrable function, f must be bounded. So we can put .M = sup |f |.
If .a ≤ x < y ≤ b, then

y
. |F (x) − F (y)| ≤ f (t) dt ≤ M(y − x).

x
Hence for every .ε > 0 we see that .|y − x| < ε/M implies .|F (x) − F (y)| < ε.
Now suppose that f is continuous at .x0 . Given .ε > 0 there exists .δ > 0 such that
.|x − x0 | < δ implies .|f (x) − f (x0 )| < ε. As a consequence, if .x0 − δ < s ≤ x0 ≤
t < x0 + δ and .a ≤ s < t ≤ b, then

F (t) − F (s) 1 t
− f (x
0 =
) [f (u) − f (x0 )] du < ε.
.
t −s t −s s
It follows that .F % (x0 ) = f (x0 ).

Theorem 10.12 (The Fundamental Theorem of Calculus) If f is R-integrable
on .[a, b], and if there exists a differentiable function F on .[a, b] such that .F % = f ,
then
b
. f dx = F (b) − F (a).
a
Proof Fix any .ε > 0. A partition .P = {x0 , . . . , xn } of .[a, b] exists such that
U (f, P ) − L(f, P ) < ε. The mean value theorem
. yields points .ti ∈ [xi−1 , xi ] such
that .F (xi ) − F (xi−1 ) = f (ti )xi . It follows that . ni=1 f (ti )xi = F (b) − F (a).
Theorem 10.3 now ensures that

b

. F (b) − F (a) − f dx < ε.

a
!b
The arbitrariness of .ε > 0 implies that .F (b) − F (a) = a f dx.

!1
Example 10.3 Following Archimedes, we can easily claim that . −1 x
2 dx = 2/3,
since .F (x) = 13 x 3 is an antiderivative of .f (x) = x 2 on .[−1, 1].
Theorem 10.13 (Integration by Parts) Suppose that F and G are differentiable
functions on .[a, b], and .F % = f , .G% = g are R-integrable. Then
b b
. F (x)g(x) dx = F (b)G(b) − F (a)G(a) − f (x)G(x) dx.
a a
!b
Proof Let .H = F G, so that Theorem 10.12 yields . a H % (x) dx = H (b) − H (a).
But .H % = f G + F g, and we conclude.

Remark 10.6 It is customary to present integration by parts in a more symmetric
manner:
b b
. uv % dx = u(b)v(b) − u(a)v(a) − u% v dx.
a a
Exercise 10.6
(a) Choose .u(x) = arctan x and .v(x) = x 2 /2 to compute the integral
!1
.
0 x arctan x dx.
(b) Choose now .u(x) = arctan x and .v(x) = (x 2 + 1)/2 and repeat the
computations.
Remark 10.7 Theorem 10.12 provides an easier proof of Theorem 10.9. Indeed one
observes that, if .F % = f , then .(F ◦ ϕ)% = (f ◦ ϕ)ϕ % . Hence
B b
. f (ϕ(y))ϕ % (y) dy = F (ϕ(B)) − F (ϕ(A)) = F (b) − F (a) = f dx.
A a
10.5 Problems 137
However this approach requires that an antiderivative F of f exists, while Theo-

rem 10.9 does not need this assumption.
10.5 Problems
10.1 Assume that f is a bounded function defined on [a, b]. If x → f (x)2 is R-

integrable on [a, b], can we deduce that f is R-integrable?
10.2 Let f : (a, b] → R be a function such that f is R-integrable on each interval
[c, b] with a < c < b. Define
b b
. f dx = lim f dx,
a c→a+ c
provided that this limit exists as a real number.

1. If f is R-integrable on [a, b], prove that this definition coincides with the
definition of the Riemann integral of f .
2. Construct a function f such that the previous limit exists, but the same limit for
|f | does not exist.
This exercise suggests a reasonable definition for the improper Riemann integral of
an unbounded function f on a bounded interval.
10.3 Let f be R-integrable on every interval [a, b] with b > a. Define
+∞ b
. f dx = lim f dx,
a b→+∞ a
provided that this limit exists as a real number. Suppose that f is monotonically
decreasing on [1, +∞) and that f (x)! ≥ 0 for each x ≥ 1. Prove that the series
+∞ +∞
n=1 f (n) converges if and only if 1 f dx exists as a finite number. This is
sometimes called the test of the improper integral for numerical series.
10.4 Let p and q be two positive real numbers such that 1
p + 1
q = 1. Prove the
following statements.
1. If u ≥ 0 and v ≥ 0, then
up vq
uv ≤
. + .
p q
Furthermore the equality sign holds if and only if up = v q .

2. If f and g are R-integrable, f ≥ 0 and g ≥ 0, then

1 1
b b p b q
. fg dx ≤ p
f dx q
g dx .
a a a
3. Deduce that if f and g are R-integrable functions, then
1 1
b b p b q
. fg dx ≤ |f | dx |g| dx
p q
.
a a a
This is called the Hölder inequality with conjugate exponents p and q. The case
p = q = 2 is usually called the Cauchy-Schwarz inequality.
10.6 Comments
My personal position is that a young (and also an old) mathematician should have
a very clear view of Riemann’s integral for functions of one variable, while a rough
idea of its extension to functions of several variables is more than enough. The
construction I have proposed in this chapter is standard, and can be found on many
textbooks.
Riemann integration theory is weak and easy in one dimension, but it becomes
weak and tricky in two or more dimensions. For this reason I always recommend
colleagues and students to replace Riemann’s integral in .Rn with the (concrete)
Lebesgue integral as soon as possible.
Chapter 11
Elementary Functions
Abstract Most of us make use of elementary functions in a formal way. We graph

exponentials, logarithms, sines, cosines, we differentiate and integrate them. But
Calculus does not teach us an acceptable definition of these functions. We accept
their existence, and we keep using their properties. In this chapter we offer a more
advanced description of the most important functions, and show that their definition
is indeed far from being elementary.
11.1 Sequences and Series of Functions
A particular type of sequence is of fundamental importance for defining the ele-

mentary functions in a rigorous way: sequences (and series) of functions. Although
these are nothing else than sequences in a set of functions, we follow here a classical
approach which does not lean on General Topology.
Suppose that E is a set and that for every positive integer n we have a function
.fn defined on E. We can say that .{fn }n is a sequence of functions on E. When we
1
speak of sequences, sooner or later we speak of limits.

Definition 11.1 Let .{fn }n be a sequence of functions on a set2 E into .R. We say
that this sequence converges pointwise to a function f , if the numerical sequence
.{fn (x)}n converges for every .x ∈ E. The function .f : E → R defined by
f (x) = lim fn (x)

.
n→+∞
is the (pointwise) limit of the sequence .{fn }n .
1 It should be remarked that E does not depend on the index n. From a theoretical view point we
could consider sequences of functions .fn : En → R in which every term is defined on a set .En .
For our purposes such a generality can be troublesome.
2 The nature of E is not particularly relevant. In many concrete cases, E is a subset of .R of .C.
140 11 Elementary Functions
In a similar way—and indeed in an equivalent way—we can define series of

functions. Indeed we can consider the numerical series . n fn (x) for every.x ∈ E: if
this series converges, the (pointwise) limit of the series . n fn is .f (x) = n fn (x).
Example 11.1 For .m = 1, 2, 3, . . . we define
fm (x) = lim (cos(m!πx))2n .

.
n→+∞
When .m!x ∈ Z, we have .fm (x) = 1. Otherwise we have .fm (x) = 0. Let .f (x) =
limm→+∞ fm (x).
If x is irrational, then .fm (x) = 0 for every m, so .f (x) = 0. For rational values
of .x = p/q, we see that .m!x is an integer for every .m ≥ q, and therefore .f (x) = 1.
To summarize,
0 if x is irrational
f (x) =
.
1 if x is rational.
This shows that the pointwise limit of very smooth functions may well be a very
irregular function.
Exercise 11.1 Consider
x2
.fn (x) =
(1 + x 2 )n
∞
for .x ∈ R and .n = 0, 1, 2, 3, . . . Let .f (x) = n=0 fn (x). Show that f is defined
for every real x, and that
0 if x = 0
f (x) =
.
1 + x2 if x = 0.
Deduce that the pointwise limit of a sequence of continuous functions is in general

a discontinuous function.
Example 11.2 Let .fn : [0, 1] → R be defined by .fn (x) = n2 x(1 − x 2 )n . Since
.fn (0) = 0, we have that
f (0) = lim fn (0) = 0.

.
n→+∞
For .0 < x ≤ 1, we trivially have .limn→+∞ n2 x(1 − x 2 )n = 0. By the Fundamental

Theorem of Calculus,
1 1
. x(1 − x 2 )n dx = ,
0 2n + 2
11.2 Uniform Convergence 141
so that
1 n2
. lim fn (x) dx = lim = +∞.
n→+∞ 0 n→∞ 2n + 2
!1 !1
In particular .0 = 0 f (x) dx = limn→+∞ 0 fn (x) dx.
Exercise 11.2 Consider instead .fn (x) = nx(1 − x 2 )n for .x ∈ [0, 1]. Show that
!1
.limn→+∞
0 fn (x) dx = 1/2, while .f (x) = limn→+∞ fn (x) = 0 for every .x ∈
[0, 1].
To summarize: the pointwise limit of a sequence of function does not preserve
continuity, differentiability, integrability. The natural question is whether we can
replace our pointwise convergence with another convergence which preserves these
properties.
11.2 Uniform Convergence
Definition 11.2 A sequence {fn }n of functions defined on a set E converges

uniformly to a limit f on E, if and only if for every ε > 0 there exists a positive
integer N such that x ∈ E and n ≥ N imply
. |fn (x) − f (x)| < ε.

When dealing with series of functions, we say that n fn converges
uniformly on
E if the sequence {sn }n of partial sums defined by sn (x) = nj=1 fj (x) converges
uniformly (to some limit).
Remark 11.1 It is easy to check that {fn }n converges uniformly to f on E if and
only if
. lim sup {|fn (x) − f (x)| | x ∈ E} = 0.

n→+∞
The quantity sup {|fn (x) − f (x)| | x ∈ E} is often denoted by *fn − f *∞,E .
Theorem 11.1 (Cauchy Criterion for Uniform Convergence) A sequence {fn }n
of functions defined on a set E converges uniformly if and only if the Cauchy
condition holds: for every ε > 0 there exists a positive integer N such that m ≥ N,
n ≥ N, x ∈ E imply
. |fn (x) − fm (x)| < ε.

Proof Suppose that {fn }n converges uniformly on E to a limit f , and let ε > 0
be fixed. By Definition 11.2 there exists a positive integer N such that n ≥ N and
x ∈ E imply |fn (x) − f (x)| < ε/2. Thus
|fn (x) − fm (x)| ≤ |fn (x) − f (x)| + |fm (x) − f (x)| < ε
.
for every n ≥ N, m ≥ N, x ∈ E.
On the contrary, suppose that the Cauchy condition holds. For every x ∈ E, the
numerical sequence {fn (x)}n is then a Cauchy sequence in R, so it converges to
some limit that we call f (x). We need to prove that this convergence is uniform on
E. Let ε > 0 be given, and choose a positive integer N such that |fn (x) − fm (x)| <
ε for every m ≥ N, n ≥ N, x ∈ E. Letting m → +∞, since fm (x) → f (x), we
deduce that
.|fn (x) − f (x)| ≤ ε
for every n ≥ N and x ∈ E. Hence {fn }n converges uniformly on E to the function

f.

The Cauchy condition immediately implies a useful test for the uniform convergence
of series of functions.
Theorem 11.2 (Weierstrass’ M-Test for Series) Suppose that {fn }n is a sequence
of functions that |fn (x)| ≤ Mn for every x ∈ E and every
on a set E, and suppose
n ∈ N. If n Mn converges, then n fn converges uniformly on E.
For any ε > 0, let N be an integer such that n ≥ N, m ≥ N and m ≥ n
Proof
imply m j =n Mj < ε. For every x ∈ E we see that

m
m
f (x) ≤ Mj ≤ ε,
.
j
j =n j =n
and the conclusion follows from Theorem 11.1.

We will now present a few statements which relate uniform convergence with limits,
derivatives and integrals.
Theorem 11.3 (Uniform Convergence and Continuity) Suppose that fn → f
uniformly on a set E in a metric space. Let x be an accumulation point of E, and
suppose that
. lim fn (t) = An
t →x
11.2 Uniform Convergence 143
for n = 1, 2, 3, . . . Then {An }n converges, and
. lim fn (t) = lim An .

t →x n→+∞
Proof Fix ε > 0. By uniform convergence there exists N such that n ≥ N, m ≥ N,

t ∈ E imply
|fn (t) − fm (t)| < ε.

.
As t → x, |An −Am | ≤ ε. Hence {An }n is a Cauchy sequence in R, and it converges

to a limit A. Now
|f (t) − A| ≤ |f (t) − fn (t)| + |fn (t) − An | + |An − A|.

.
We choose n0 ≥ N so that |f (t)−fn0 (t)| < ε/3 for all t ∈ E, and |An0 −A| < ε/3.
With this n0 we choose a neighborhood V of x such that t ∈ V ∩ E, t = x imply
|fn0 (t) − An | < ε/3. It follows that |f (t) − A| < ε for every t ∈ V ∩ E, t = x. The
proof is complete.

Corollary 11.1 If {fn }n is a sequence of continuous functions on E that converges
uniformly to a limit f , then f is a continuous function.
Theorem 11.4 (Uniform Convergence and Differentiation) Suppose that {fn }n
is a sequence of functions, differentiable on [a, b] and such that there exists a point
x0 ∈ [a, b] such that {fn (x0 )}n converges. If {fn% }n converges uniformly on [a, b],
then {fn }n converges uniformly on [a, b] to a limit f , and
f % (x) = lim fn% (x)

.
n→+∞
for every x ∈ [a, b].

In other words, uniform convergence of the derivatives a pointwise convergence of
the functions at some point x0 imply uniform convergence of the functions.
Proof We fix ε > 0. By assumption there exists a positive integer N such that
ε
|fn (x0 ) − fm (x0 )| <
.
2
and
ε
|fn% (t) − fm% (t)| <
. (11.1)
2(b − a)
for every n ≥ N, m ≥ N. Let us apply the mean value theorem to fn −fm : Eq. (11.1)
yields
|x − t|ε ε
. |fn (x) − fm (x) − (fn (t) − fm (t))| ≤ ≤ (11.2)
2(b − a) 2
for every x ∈ [a, b], t ∈ [a, b], n ≥ N, m ≥ N. Splitting
|fn (x) − fm (x)| ≤ |fn (x) − fm (x) − fn (x0 ) + fm (x0 )| + |fn (x0 ) − fm (x0 )|
.
we deduce that x ∈ [a, b], n ≥ N, m ≥ N imply
|fn (x) − fm (x)| ≤ ε.

.
This proves that {fn }n converges uniformly to a limit that we call f . We fix a point
x ∈ [a, b] and introduce the incremental ratios
fn (t) − fn (x) f (t) − f (x)

.φn (t) = , φ(t) =
t −x t −x
for any t ∈ [a, b] \ {x0 }. Clearly limt →x φn (t) = fn% (x). From (11.2) we see that
n ≥ N, m ≥ N imply
ε
|φn (t) − φm (t)| ≤
. .
2(b − a)
Hence {φn }n converges uniformly on [a, b] \ {x}. But {fn }n converges to f , hence
limn→+∞ φn (t) = φ(t) uniformly on [a, b] \ {x}. We conclude from Theorem 11.3
that limt →x φ(t) = limn→+∞ fn% (x). The proof is complete.

Theorem 11.5 (Uniform Convergence and Integration) Suppose that each fn is
R-integrable on [a, b], and suppose that fn → f uniformly on [a, b]. Then f is
!b !b
R-integrable on [a, b] and a f dx = limn→+∞ a fn dx.
Proof Let
εn = sup {|fn (x) − f (x)| | x ∈ [a, b]} .

.
Hence fn − εn ≤ f ≤ fn + εn , and this implies
b b b b
. (fn − εn ) dx ≤ f dx ≤ f dx ≤ (fn + εn ) dx.
a a a a
11.3 The Exponential Function 145
Hence
b b
0≤
. f dx − f dx ≤ 2εn (b − a).
a a
!b !b
Since εn → 0 as n → +∞, we conclude that f dx = af dx, and f is R-
a
integrable. As before,

b b b
. f dx − fn dx = (f − fn ) dx ≤ εn (b − a),

a a a
!b !b
and it follows that a fn dx → a f dx as n → +∞.

Remark 11.2 The rigidity of the Riemann integral under passage to the limit is one
of the reasons why it has been superseded by more flexible integrals. We will meet
Lebesgue’s generalization in Chap. 15.
11.3 The Exponential Function
It has been said that the exponential function is the most important function in
Mathematical Analysis. We propose a definition which entails a lot of useful
properties.
Definition 11.3 For each .z ∈ C, we define
∞ n
z
. exp z = .
n!
n=0
By the Ratio Test, the series converges absolutely. The function
. exp : C → C
is the exponential function.3

Proposition 11.1
1. For every .z ∈ C, .w ∈ C, there results .exp(z + w) = exp z · exp w.
2. .exp 0 = 1 and .exp 1 = e.
3. .exp z =
0 for every .z ∈ C.
3We refrain from writing .ez instead of .exp z. This is only a pedagogical choice, since we want to
prevent the reader from believing that all properties of this function are trivial since we are deling
with an ordinary power.
4. .exp(−z) = 1/ exp z for every .z ∈ C.

5. .exp is a continuous function.
Proof We form the Cauchy product according to Definition 6.3:
∞
n
zk wn−k
. exp z · exp w =
k! (n − k)!
n=0 k=0
∞
1
n
n!
= zk wn−k
n! k!(n − k)!
n=0 k=0
∞
1
= (z + w)n = exp(z + w).
n!
n=0
This proves 1. Part 2 is clear from the definitions. Using 1 and 2 we see that
exp(−z) exp z = 1 for every .z ∈ C, and both 3 and 4 follow. To prove 5, we fix
.
a point .z ∈ C and .ε > 0. Let
ε
.δ = min 1, .
2| exp z|
Hence .h ∈ C and .|h| < δ imply
. |exp(z + h) − exp z| = | exp z| · | exp h − 1|

∞
ε hn
∞
ε |h|n
≤ ≤
2δ n! 2δ n!
n=1 n=1
∞
ε 1 ε
< = (e − 1) < ε.
2 n! 2
n=1
The proof of 5 is complete.

Theorem 11.6 There results
1. .exp x > 0 for every .x ∈ R.
2. .exp is strictly increasing on .R.
3. .limx→+∞ exp x = +∞.
4. .limx→−∞ exp x = 0.
5. .exp(R) = (0, +∞).
6. .limx→+∞ x
−n exp x = +∞ for every integer .n ≥ 0 and every .x ∈ R.
Proof We first notice that .exp(R) ⊂ R. Since .x > 0 implies .exp x > 1 + x, 1
follows for .x ≥ 0. If .x < 0, .exp(−x) = 1/ exp x, and 1 follows also in this case. To
prove 2, we fix .x ∈ R and .h > 0. Then .exp(x + h) = exp x · exp h > exp x since
.exp h > 1. Similarly, .limx→+∞ exp x ≥ limx→+∞ (1 + x) = +∞, and 3 follows.
11.3 The Exponential Function 147
Recalling again that
1
. lim exp x = lim exp(−x) = lim ,
x→−∞ x→+∞ x→+∞ exp x
we see that 4 follows from 3. The proof of 6 goes as follows: .x > 0 implies
x n+1 x
.x −n exp x > x −n =
(n + 1)! (n + 1)!
To conclude, 5 follows from 1, 3, 4 and the fact that .exp is injective on .R.

Theorem 11.7 The restriction of .exp to .R is differentiable at every point, and there
results .exp% = exp.
Proof Just differentiate the series of function that defines the exponential, and use
Theorem 11.4.

One an exponential function has been introduced, the logarithm comes into play as
its inverse.
Definition 11.4 The (real) logarithm is defined as
−1
. log = exp|R .
As a function, .log : (0, +∞) → R.

Theorem 11.8 There results
1. .log((0, +∞)) = R.
2. .log is strictly increasing on .(0, +∞).
3. .log is continuous on .(0, +∞).
4. .limx→+∞ log x = +∞.
5. .limx→0+ log x = −∞.
6. .log 1 = 0 and .log e = 1.
7. .log(ab) = log a + log b for every .a > 0, .b > 0.
.log(a ) = n log a for every .a > 0 and every .n ∈ Z.

8. n
9. .log(a
1/n ) = (1/n) log a for every .a > 0 and .n ∈ N.
log x
10. .limx→+∞ √ n x = 0 for every .n ∈ N.
Proof Properties from 1 to 6 follow from the analogous properties of the exponen-
tial. For .a > 0 and .b > 0 we have
. exp (log a + log b) = exp(log a) · exp log b = ab,
and this proves 7. Properties 8 and 9 are left as a simple exercise about induction.
To prove 10 we fix .ε > 0 and .n ∈ N. By Theorem 11.6 we can choose .α > 1 such
that .y > α implies .y −n exp y > ε−n . By property 4 there exists .β > 1 such that
x > β implies .log x > α. Therefore .x > β implies

.
. (log x)−n x > ε−n ,
or
log x
0< √
.
n
< ε.
x

Theorem 11.9 The function .log is differentiable at every point of .(0, +∞), and
there results .(log)% (y) = 1/y for every .y > 0.
Proof Exercise on the derivative of the inverse function!

Remark 11.3 Another approach is possible, which avoids the use of series. The first
idea is to define a function .log : (0, +∞) → R by
x dt
. log x = .
1 t
!b !a
Here we are using the convention . a = − b . The main properties of the real
logarithm follows at once, in particular the fact that .log has a continuous inverse.
We call the inverse the real exponential function.
11.4 Sine and Cosine
Definition 11.5 The functions sin and cos are defined on C by

∞
z2n+1
. sin z = (−1)n
(2n + 1)!
n=0
∞
z2n
cos z = (−1)n
(2n)!
n=0
for every z ∈ C. The convention 00 = 1 is used in these formulas.

Theorem 11.10 (Euler) For every z ∈ C we have
1. exp(iz) = cos z + i sin z.
2. exp(−iz) = cos z − i sin z.
3. sin z = exp(iz)−exp(−iz)
2i .
exp(iz)+exp(−iz)
4. cos z = 2 .
11.4 Sine and Cosine 149
Proof Observe that (−1)n z2n = (iz)2n and i(−1)n z2n+1 = (iz)2n+1 . Then 1
follows from the definition of sin and cos. Property 2 follows by replacing z by
−z in 1. If we solve the system
exp(iz) = cos z + i sin z

.
exp(−iz) = cos z − i sin z
with respect to sin z and cos z, we find 3 and 4.

The next result may also be taken as a definition of π.
Theorem 11.11 There exists one and only one real number π such that
(i) π > 0;
(ii) cos(π/2) = 0;
(iii) 0 < x < π/2 implies cos x > 0.
Furthermore, π < 4.
Proof If π and π % are different numbers which satisfy (i)–(iii) and π < π % , then
cos(π/2) > 0. This contradiction shows the uniqueness of π.
To prove the existence, we reason as follows. For every n > 1 we have
22n 22n+2
. > ,
(2n)! (2n + 2)!
hence
22 24 26
. cos 2 = 1 − + − + ···
2! 4! 6!
16 2
= −1 + − · · · < −1 + < 0.
24 3
Since x → cos x is a continuous real-valued function on [0, 2] and cos 0 = 1 > 0,
it follows that the set
A = {x ∈ [0, 2] | cos x = 0}
.
is non-empty. We define π/2 = inf A. As an accumulation point of the closed set A,

we have π/2 ∈ A, and π/2 > 0 since cos 0 = 1. As a consequence π satisfies (i)
and (ii). Suppose (iii) was false. Then another application of the Intermediate Value
Theorem would produce an element of A which would be smaller than π/2. Finally,
since cos 2 < −1 + 2/3, we find π/2 < 2, and the proof is complete.

From Euler’s Theorem we can recover all the properties of the trigonometric
functions. The interested reader can expand the details and prove the main results.
It is however remarkable that it would be impossible to postpone the use of these
functions until they can be rigorously defined. Calculus is built around elementary
functions, but it does not provide sufficient tools to define them without any
reference to geometric or intuitive facts.
Remark 11.4 The approach to trigonometric functions without power series is
slightly involved. A possible approach is to begin with arctan: R → (−π/2, π/2)
in terms of a definite R-integral:
x dt
. arctan x = .
0 1 + t2
This function is invertible, so that tan is defined on (−π/2, π/2) as arctan−1 .

Then sin and cos are recovered as suggested by the identity sin2 + cos2 = 1, i.e.
tan2 +1 = 1/ cos2 .
We end this section with a discussion about periodic functions.
Definition 11.6 A function f : R → R is periodic if and only if there exists a real
number T = 0 such that f (x + T ) = f (x) for every x ∈ R.
Exercise 11.3 Let P = {T ∈ R \ {0} | T is a period of f } be the set of all periods
of a given function f . Prove that the sum and the difference of two elements of P
are elements of P . We can summarize this by saying that P is an additive subgroup
of R.
We now investigate additive subgroups of R.
Theorem 11.12 If H is an additive subgroup of R, then either H = R or there
exists T ∗ = 0 such that H = {mT ∗ | m ∈ Z}.4
Proof Since H is an additive subgroup, H ∩ [0, +∞) = ∅. We define
.η = inf(H ∩ [0, +∞)).
Two cases are possible. If η > 0, then we pick h ∈ H and m ∈ Z such that
mη ≤ |h| < (m + 1)η.

.
Since |h| − mη ∈ H and 0 ≤ |h| − mη < (m + 1)η − mη = η, the definition of η

implies |h| − mη = 0, or h = ±mη. Hence H consists of all integer multiples of η.
If η = 0, we must prove that H is dense in R. Let r ∈ R and let ε > 0.
Since η = 0, there exists hε ∈ H ∩ [0, ε]. We may suppose r ≥ 0, the case
r < 0 being similar. By the Archimedean property, there exists k ∈ N such that
kh ≤ r < (k + 1)h. Since hk ∈ H and 0 ≤ r − kh < (k + 1)h − kh = h ≤ ε, we
see that |r − kh| ≤ ε, which shows that H is dense in R. The proof is complete.

4 Algebraists say that H is a cyclic group.

11.5 Polynomial Approximation 151
Theorem 11.13 Let f : R → R be a periodic continuous function. If f is non-

constant, then there exists a smallest positive period of f .
Proof Let P be the set of all periods of f . We are going to rule out the possibility
that P is dense in R. Indeed, if this were true, then f would be constant on the dense
subset P , and thus f would be globally constant by continuity. Since this contradicts
the assumption, we conclude that all periods are integer multiples of some period
T ∗ = 0. Since −T ∗ is also a period, we may assume that T ∗ > 0, and by Theorem
11.12 T ∗ is the smallest positive period of f . The proof is complete.

Important: Periodic Functions Without a Smallest Period

Many Calculus students believe that any periodic function possesses the period, i.e.
a unique number like π for the sine or cosine. Although some textbooks restrict
the definition of periodicity to functions which do have a smallest period, in this
book we will always think of constant functions as periodic functions. We invite the
reader to elaborate on the function
0 if x is rational
f (x) =
.
1 if x is irrational.
It is clear that no smallest positive period exists, although f is surely periodic in the
sense of Definition 11.6.
11.5 Polynomial Approximation
Polynomials are the most elementary functions of mathematical analysis. They are
built on arithmetic operations, and they turn out to be a flexible class of infinitely
differentiable functions. Of course not all functions are polynomial.
Exercise 11.4 Prove rigorously that not all functions are polynomials. Hint: if P is
a non-constant polynomial, either .limx→±∞ |P (x)| = +∞.
Nevertheless, polynomials do approximate continuous functions in a strong way.
Theorem 11.14 (Weierstrass Approximation Theorem) If .f : [a, b] → R
is a continuous function and .ε > 0, there exists a polynomial P such that
.supx∈[a,b] |f (x) − P (x)| < ε.
Proof We will prove an equivalent statement: there exists a sequence .{Pn }n of

polynomials such that .Pn → f uniformly on .[a, b]. Considering an affine change
of variable, we may assume that .[a, b] = [0, 1]. Furthermore, we may also assume
that .f (0) = f (1) = 0. Indeed, the function f may be replaced by the function
x → f (x) − f (0) − x(f (1) − f (0)).

.
This function differs from f by a polynomial (of degree .≤ 1), so that a uniform
approximation of this function by means of polynomials implies that f is approxi-
mated by polynomials.
Finally, for technical reasons, we define .f (x) = 0 for each .x ∈ R \ [0, 1]. Hence
f is defined on the whole real line. For each .n = 1, 2, 3, . . . define
1
cn = ! 1
.
−1 (1 − x 2 )n dx
and
. Qn (x) = cn (1 − x 2 )n .
!1
Trivially, . −1 Qn (x) dx = 1 for each n. Furthermore,
1 1
. (1 − x 2 )n dx = 2 (1 − x 2 )n dx
−1 0
√
1/ n
≥2 (1 − x 2 )n dx
0
√
1/ n
≥2 (1 − nx 2 ) dx
0
4 1
= √ > √ .
3 n n
√
As a consequence, .cn < n for each n. Now, fix any .δ > 0, and observe that
.δ ≤ |x| ≤ 1 implies
√
Qn (x) ≤
. n(1 − δ 2 )n .
The right-hand side converges to zero, hence .Qn converges to zero uniformly in the
region .δ ≤ |x| ≤ 1, i.e. away from zero.
We introduce the sequence of functions
1
Pn (x) =
. f (x + t)Qn (t) dt,
−1
11.6 A Continuous Non-differentiable Function 153
defined for .0 ≤ x ≤ 1. A change of variable shows that
1
Pn (x) =
. f (t)Qn (t − x) dt,
0
which is a polynomial function. We claim that .{Pn }n is the approximating sequence

of polynomials we are looking for. Indeed, given .ε > 0 we choose .δ > 0 so that
.|y − x| < δ implies .|f (y) − f (x)| < ε/2. Here we are exploiting the uniform
continuity of f on .[0, 1]. Call .M = sup{|f (x)| | 0 ≤ x ≤ 1}. Recalling that

.Qn ≥ 0 we see that for each .0 ≤ x ≤ 1 we have

1
. |Pn (x) − f (x)| = (f (x + t) − f (x)) Qn (t) dt
−1
1
≤ |f (x + t) − f (x)| Qn (t) dt
−1
−δ ε δ 1
≤ 2M Qn (t) dt + Qn (t) dt + 2M Qn (t) dt
−1 2 −δ δ
√ ε
≤ 4M n(1 − δ 2 )n +
2
<ε
provided that n is big enough.

Example 11.3 Consider the function .f : [−1, 1] → R defined by .f (x) = |x|.
By the previous result, a sequence .{P̃n }n of polynomials exists which converges
uniformly to f on .[−1, 1]. Setting .Pn (x) = P̃n (x) − P̃n (0), we see that .{Pn }n is a
sequence of polynomials that converges to f on .[−1, 1] and such that .Pn (0) = 0
for every n. Notice that f is not differentiable at .x = 0: in some sense, the sequence
.{Pn } is a smooth uniform approximation of f .
11.6 A Continuous Non-differentiable Function
Every student learns that a continuous function may fail to be differentiable at

all points, and the simplest example is usually the absolute value .x → |x|. Karl
Weierstrass proved a much stronger result about a continuous function exists which
is nowhere differentiable. Intuitively, such a function cannot be represented by a
simple formula. In this section we propose a reasonable construction.
Theorem 11.15 There exists a continuous function on the real line .R which is
nowhere differentiable .
Proof Let us start with .ϕ(x) = |x|, for each .x ∈ [−1, 1]. Then we extend it by
periodicity to .R, i.e. .ϕ(x + 2) = ϕ(x) for each .x ∈ R. It is clear that
|ϕ(s) − ϕ(t)| ≤ |s − t|
. (11.3)
for each s, t in .R. The function .ϕ : R → R is uniformly continuous.

Now we define
∞
3 n
f (x) =
. ϕ 4n x .
4
n=0
The M-test 11.2 shows that f is defined by a series that converges uniformly on .R,
so that f is a continuous function. We claim that f is differentiable at no point of
.R. To prove this claim, we pick any .x ∈ R and any positive integer m. Let
1
δm = ± · 4−m ,
.
2
where the sign is chosen so that the interval .[4m x, 4m (x + δm )] contains no integer.
Since .4m |δm | = 1/2, this is indeed possible. Consider now the incremental ratio
ϕ(4n (x + δm )) − ϕ(4n x)
γn =
. .
δm
If .n > m, the number .4n δm is an even integer, so that .γn = 0. If .0 ≤ n ≤ m, it

follows from (11.3) that .|γn | ≤ 4n . Recalling that .|γm | = 4m , we see that
∞
f (x + δm ) − f (x) 3 n

. = γn
δm 4
n=0

m−1
≥ 3m − 3n
n=0
1 m
= 3 +1 .
2
Letting .m → +∞ we get .δm → 0. Hence f is not differentiable at x, and the proof
is complete.

Remark 11.5 Such a function is a typical example of a fractal curve, whose graph
is essentially impossible to sketch. Our function f is based on the function .ϕ, which
is already irregular at countably many points. However, Weierstrass constructed a
11.7 Asymptotic Estimates for the Factorial Function 155
more complicated example of the form

∞

f (x) =
. a n cos b n πx ,
n=0
where .0 < a < 1 and b is a positive odd integer such that
3
.ab > 1 + π.
2
The function f is then a trigonometric series, and each term of the infinite sum is a
smooth function. Once more we see that a limit of regular functions may be a very
irregular function.
11.7 Asymptotic Estimates for the Factorial Function
Asymptotic estimates are a major tool in several fields of mathematics. We want to

present a couple of results which describe the behavior of the factorial .n! as n gets
larger and larger. Since the factorial has been introduced as a discrete function, our
road is not really straight.
Definition 11.7 (Double Factorial) We define inductively on .n ∈ N,
(−1)!! = 1,
. 0!! = 1, (n + 1)!! = (n + 1) · (n − 1)!!
Exercise 11.5 Prove that .n!! is the product of all odd numbers .m ≤ n when n is
odd, and it is the product of all even numbers .m ≤ n when n is even.
Theorem 11.16 (Wallis Integrals) If .n ∈ N, then
π/2 (2n)!!
. (sin x)2n+1 dx = . (11.4)
0 (2n + 1)!!
π/2 π (2n − 1)!!
(sin x)2n dx = . (11.5)
0 2 (2n)!!
π/2 π/2 π/2
(sin x)2n+1 dx ≤ (sin x)2n dx ≤ (sin x)2n−1 dx. (11.6)
0 0 0
Proof For every .x ∈ [0, π/2] we have .0 ≤ sin x ≤ 1. Hence, for any such x,
the sequence .n → (sin x)n is decreasing. In particular (11.6) follows at once from
the monotonicity properties of the Riemann integral. We prove (11.4) and (11.5) by
induction on n. They clearly hold true when .n = 0, and we integrate by parts as
follows:
π/2 π/2
. (sin x)m dx = (sin x)m−2 (1 − cos2 x) dx
0 0
π/2 π/2
= (sin x)m−2 dx − cos x · (sin x)m−2 x cos x dx
0 0
π/2 1 π/2
= (sin x)m−2 dx − (sin x)m dx,
0 m−1 0
deducing the identity
π/2 m−1 π/2

. (sin x)m dx = (sin x)m−2 dx
0 m 0
for every .m ≥ 2. The induction step is now easy.

Theorem 11.17 (Wallis Formulas) As .n → +∞,
(2n)!! √
. ∼ nπ . (11.7)
(2n − 1)!!
2n 22n
∼ √ (11.8)
n nπ
2 · 2 · 4 · 4 · 6 · 6 · 8 · 8 · · · 2n · 2n π
. = + o(1). (11.9)
1 · 3 · 3 · 5 · 5 · 7 · 7 · · · (2n − 1) · (2n − 1) 2
Proof We set
√
nπ
q=
.
(2n)!!
(2n−1)!!
and we see from (11.4) and (11.5) that
2n π/2 2n π/2
q= √
. (sin x)2n dx ≥ √ (sin x)2n+1 dx
nπ 0 nπ 0
2n (2n)!! 1 1 2n
= √ =
nπ (2n − 1)!! 2n + 1 q 2n + 1
11.7 Asymptotic Estimates for the Factorial Function 157
and
2n π/2 2n π/2
.q= √ (sin x)2n dx ≤ √ (sin x)2n−1 dx
nπ 0 nπ 0
2n (2n)!! 1 1
= √ = .
nπ (2n − 1)!! 2n q
We have thus proved that

" √ #2
2n nπ
. ≤ ≤1
2n + 1 (2n)!!
(2n−1)!!
for every .n ≥ 1. Now (11.7) follows easily. Since
2n (2n)!! 22n (2n − 1)!! 22n

. = = ∼ √ ,
n (n!)2 (2n)!! nπ
also (11.8) follows. To conclude, we observe that
2 · 2 · · · (2n) · (2n) (2n)!!(2n)!! nπ π

. ∼ ∼ ∼ .
1 · 3 · · · (2n − 1) · (2n − 1) (2n − 1)!!(2n − 1)!!(2n + 1) 2n + 1 2

Theorem 11.18 (Stirling) As .n → +∞,
√
. n! ∼ nn e−n 2nπ.
Proof We define the sequence
xn = n!en n−n−1/2 .
.
A direct calculation shows that

xn+1 − n+ 12 log n+1
n +1
. =e .
xn
By taking logarithms we see that
xn+1 1 n+1 1
. log xn+1 − log xn = log = − n+ log +1∼−
xn 2 n 12n2

as .n → +∞. As a consequence, the series . n log xn+1 − log xn converges, and
k = lim log xn
.
n→+∞
√
exists in .R. Thus .n! ∼ nn e−n nek . Inserting this into (11.8) we see that
√ √ √
22n 2n (2n)! (2n)2n e−2n 2nek 22n 2 22n 2π
.√ ∼ = ∼ = √ k =√ ,
nπ n (n!) 2 2n
n e ne−2n 2k ne nπ ek
√
and necessarily .ek = 2π. The proof is complete.

11.8 Problems
11.1 Consider the series of functions

∞
1
f (x) =
. .
1 + n2 x
n=1
For what values of x does the series converge absolutely? In what intervals does the
series converge uniformly? If the series converges, is f a continuous function?
11.2 For each n = 1, 2, 3, . . . let
⎧
⎪
⎪ 1
⎨0 if x < n+1
.fn (x) = ≤x≤
1 1
sin2 π
if
⎪
⎪ x n+1 n
⎩0 if x > n1 .
Prove that the sequence {fn }n converges pointwise to a continuous function, but the
converges is not uniform.
11.3 For each real number x, let {x} = x − [x], where [x] denotes the integer part
of x. Let
∞
{nx}
f (x) =
. .
n2
n=1
Find all discontinuity points of f , and prove that these points form a dense,
countable subset of R. Prove also that f is R-integrable on each bounded interval.
11.8 Problems 159
11.4 Let f be a continuous function on [0, 1], and suppose that
1
. f (x)x n dx = 0 for n = 0, 1, 2, 3, . . .
0
Prove that f (x) = 0 for each x ∈ [0, 1].

11.5 Suppose that a sequence {fn }n converges pointwise to f on a compact set K,
and suppose moreover that fn (x) ≤ fn+1 (x) for each x ∈ K and each n ∈ N.
1. By setting g = f − fn , reduce to the case of a sequence of function which
converges pointwise to zero in a decreasing way.
2. Assume that f and each fn are continuous functions on K. Fix any ε > 0 and
define for each n the set Kn = {x ∈ K | gn (x) ≥ ε}. Prove that
K1 ⊃ K2 ⊃ K3 ⊃ . . . ,
.
and conclude that fn → f uniformly on K.

Part II
Second Half of the Journey
Chapter 12
Return to Set Theory
Abstract This chapter introduces the second part of the book. Why should we
bother about set theory, again? And what is axiomatic set theory? In some sense,
learning mathematics has a privileged direction: we take something for granted, and
then we proceed. Mathematical analysts are usually satisfied with a good knowledge
of naïve set theory, since this is all they need to do their job. Going backwards
is another story. We are always worried by primitive knowledge, in the sense of
something that we agree to know before we start out journey. So, what is a set?
Why do textbooks begin with the definition of a function as a black box that turns
an element (of a set) into another element (of another set)? And, after all: is this the
only way to begin?
The roots of mathematics are close to philosophy, in the sense that a beginning must
exist. If nothing exists, how can we exist? We only (!) have to choose where to start
from.
It is a general agreement that set theory is indeed the first mathematical chapter
in the big book of all Mathematics. The rigorous construction of sets, functions and
all that has been a long and recent process. If we agree that mathematics should set
axioms and deduce theorems, we should isolate the axioms of set theory.
Nowadays the most popular axiomatization of set theory is ZF, or Zermelo-
Fraenkel. Another axiomatization is due to Bernays, Gödel and Von Neumann. It
is not our purpose to discuss the pros and cons of each theory, since this is a deep
aspect that soon involves mathematical logic. Since this is a book in mathematical
analysis, we prefer to present an overview of two less known theory of sets. The first
one is due to John Kelley, see [3]. The second one is actually a variation on Kelley’s
theme, due to J. D. Monk. These theories share the simple approach of considering
only classes, an undefined category that contains sets. The idea of elements is not a
different object: everything is a class. For the reader’s convenience, a short account
of ZF is also provided.
164 12 Return to Set Theory
Morse-Kelley theory is suitable for mind shape of a mathematical analyst,

because it provides a straightforward escape from one of the worst paradoxes of
mathematics:
No matter what you suppose to know about sets, the set of all sets cannot be a set.
We will come back to this paradox later in the chapter.

Remark 12.1 From a more advanced viewpoint, Morse-Kelley axiomatization has
been superseded by the NBG axiomatization of Von Neumann, Bernays and Gödel.
The classification axiom scheme that we will introduce below is the most flexible
feature of Morse-Kelley, although it is so general that experts of mathematical logic
prefer to replace it with something weaker but easier from their perspective. This is
however an issue that does not bother us.
12.1 Kelley’s System of Axioms
We will see that the word “class” does not appear in any axiom of Kelley’s theory:
instead of assuming the existence of sets, Kelley assumes the existence of classes,
and sets are just special classes.
The natural question now is: why don’t we simply agree that “set = class”? Of
course this is a matter of language, but we always keep in mind that the set of all
sets is not a set.
Important: Variables
We stick here to Kelley’s original habit of using variables in a broad sense, so that
any object should be considered as a class unless otherwise stated. Another popular
approach is to reserve lower case letters to sets, and upper case letters to classes.
Axiom of Extent For each x and each y, it is true that .x = y if and only if
for each z, .z ∈ x if and only if .z ∈ y.
Two classes are equal if and only if each element of each is a member of the
other. The reader will notice that this is the usual definition of equal sets. And here
is the rigorous definition of a set, at last! Sets are just elements of some class.
Definition 12.1 A class x is a set if and only if there exists a class y such that .x ∈ y.
Remark 12.2 Maybe some reader will remember a common approach to naïve set
theory: whenever we name a set, we must agree that it is a subset of some larger set.
This is a naïve response to the paradox of the set of all sets: since we agree to work
12.1 Kelley’s System of Axioms 165
with subsets of a fixed large set, we will never have to deal with the set of all sets.
This is a useful agreement, but it does not face the deep meaning of the paradox.
Now that we have sets, we want to describe the use of the classifier .{. . . | . . .}.
The are two blanks: the first blank for a variable, and a second blank for a formula.
Axiom of the Classifier For each u and y, .u ∈ {x | x ∈ y} if and only if u is

a set and .u ∈ y.
This axiom is of fundamental importance: it states that whenever we write .u ∈

. . ., we immediately understand that u is a set. More generally, each statement of
the form
u ∈ {x | . . . x . . .}
.
is considered to be an axiom with the meaning that u is a set and .. . . u. Here
. ...x ...
is a formula and
. ...u...
is the formula which is obtained from it by replacing each occurrence of x with u.

For example, .u ∈ {x | x ∈ y and z ∈ x} if and only if u is a set, .u ∈ y and .z ∈ u.
The axiom scheme that we have just introduced obviously corresponds to the
familiar way of constructing sets by specifying properties that characterize its
elements. However, we added the requirement “u is a set”. This is surely unnatural,
but again: .{x | x is a set} would coincide with the set of all sets, and this is not a
set. With our axiom scheme, .u ∈ {x | x is a set} just means that u is a set, and no
paradox arises.
Formulae We agree that:
(a) The result of replacing .α and .β by variables is, for each of the following, a
formula: .α = β and .α ∈ β.
(b) The result of replacing .α and .β by variables and A and B by formulae is, for
each of the following, a formula:
1. if A then B,
2. A if and only if B,
3. it is false that A,
4. A and B,
5. A or B,
6. for every .α, A

7. for some .α, A
8. .β ∈ {α | A}
9. .{α | A} ∈ β
10. .{α | A} ∈ {β | B}
(c) Formulae are constructed recursively, beginning with the primitive formulae of
(a) and then proceeding via the constructions allowed by (b).
Classification Axiom Scheme An axiom results if in the following .α and .β

are replaced by variables, A by a formula .A and B by the formula obtained
from .A by replacing each occurrence of the variable that replaced .α by the
variable that replaced .β:
For each .β, .β ∈ {α | A} if and only if .β is a set and B.
Important: Braces
In old books a peculiar use of braces was common. For instance, one would write
.M = {m} to mean the “typical” element of the class/set M is m. In modern
mathematics and according to our use of the classifier .{. . . | . . .}, .M = {m} means
.{x | x = m}, so that the class/set M contains exactly one element, i.e. m. Please
avoid old-fashioned notation if you do not want to produce tragic results.
Luckily enough, our axioms already allow us to introduce new definitions and
prove some theorems. Let us see some of them.
Definition 12.2 (Union of Classes) .x ∪ y = {z | z ∈ x or z ∈ y}.
Definition 12.3 (Intersection of Classes) .x ∩ y = {z | z ∈ x and z ∈ y}.
Theorem 12.1 For each z, .z ∈ x ∪ y if and only if .z ∈ x or .z ∈ y. For each z,
z ∈ x ∩ y if and only if .z ∈ x and .z ∈ y.
.
Proof It follows from the classification axiom that .z ∈ x ∪ y if and only if z is a set,
and .z ∈ x or .z ∈ y. Recalling Definition 12.1, .z ∈ x or .z ∈ y and z is a set if and
only if .z ∈ x or .z ∈ y. By the same token one proves the second statement about
the intersection.

Remark 12.3 Some reader may think that we are playing a strange game. This is
more or less the case. We should always remember that mathematics is concerned
with proving theorems from axioms. In this particular case, it would be much
stranger if our theorems were unexpected ones! We are trying to define rigorously
what we handle every day, this is the point.
Theorem 12.2 For each x, .x ∪ x = x, and .x ∩ x = x.
Proof Every element of x is a member of .x ∪ x, by definition of the union.
Conversely, if .z ∈ x ∪ x, then z is a set and .z ∈ x or .z ∈ x. Hence .z ∈ x. The
second statement is left as an exercise.

Exercise 12.1 Prove the theorem .x ∩ x = x by mimicking the previous proof.
We collect now three statements that follow directly from the properties of logical
quantifiers “and”, “or”.
Theorem 12.3 For each x, y and z,
1. .x∪y =y∪x
2. .x∩y =y∩x
3. .(x ∪ y) ∪ z = x ∪ (y ∪ z)
4. .(x ∩ y) ∩ z = x ∩ (y ∩ z)
5. .x ∩ (y ∪ z) = (x ∩ y) ∪ (x ∩ z)
6. .x ∪ (y ∩ z) = (x ∪ y) ∩ (x ∪ z).
Exercise 12.2 Prove the previous Theorem, by showing that any two classes
separated by the symbol .= have the same members.
Definition 12.4 For each x and y, .x ∈
/ y if and only if it is false that .x ∈ y.
Definition 12.5 For each x, .x = {y | y ∈/ x}. The class .x is the complement of
the class x.

Theorem 12.4 For each x, . x = x.

Proof By Definition 12.5, . x is the class
{y | y ∈
. / x} = {y | it is false that y ∈ x} = {y | y ∈ x}.

Theorem 12.5 (De Morgan Laws) For each x and y, .(x ∪ y) = x ∩ y, and
(x ∩ y) = x ∪ y.
.
Proof The second statement is left as an exercise. Let us see why the previous
statement is true. For each z, .z ∈ (x ∪ y) if and only if z is a set and it is false
that .z ∈ x ∪ y. Recalling Theorem 12.1, .z ∈ x ∪ y if and only if .z ∈ x or .z ∈ y.
Consequently, .z ∈ (x ∪ y) if and only if z is a set and .z ∈
/ x and .z ∈
/ y. This means
exactly that .z ∈ x and .z ∈ y. By Theorem 12.1, .z ∈ x and .z ∈ y if and only if
.z ∈ x ∩ y. We conclude by the axiom of extent.

Exercise 12.3 Prove that .(x ∩ y) = x ∪ y by mimicking the previous proof.
Definition 12.6 For each x and y, .x \ y = x ∩ y.
Remark 12.4 We use the modern symbol .\ throughout the book. Other symbols are
found in the literature, like .x − y or .x ∼ y. Kelley systematically writes .∼ x instead
of .x.
Remark 12.5 In naïve set theory, the class .x is usually undefined as a primitive
object. The reason is that all sets must be subsets of some universe U . Hence .x
must reduce to .U \ x. In other words, only the relative complement of a set must be
defined in naïve set theory.
Definition 12.7 (Empty Class) .∅ = {x | x = x}.
Remark 12.6 The empty class is denoted by 0 in [3]. We prefer to use the dedicated
symbol .∅, since 0 is used with a lot of different meanings in mathematics. Another
popular notation is .{}.
Theorem 12.6 For each x, .x ∈
/ ∅.
Proof By definition equality is reflexive, in the sense that for each x, .x = x. Hence
it is false that .x = x, and therefore .∅ cannot contain any element.

Exercise 12.4 Prove that for each x, .∅ ∪ x = x and .∅ ∩ x = ∅.
Definition 12.8 (Universal Class) .U = {x | x = x}. The class .U is called the
universe.
Important: Warning
Beware! .U is not a set! Look ahead in this chapter.
Theorem 12.7 For each x, .x ∈ U if and only if x is a set.

Proof .x ∈ U if and only if x is a set and .x = x, hence if and only if x is a set.

We may now say that .U is the class that contains every set, although .U is not a
set itself. This is essentially the reason why you are reading this chapter.
Exercise 12.5 Prove that for each x, .x ∪ U = U and .x ∩ U = x. Roughly
speaking, .U is so large that we cannot add anything that is not already an element,
and intersecting with .U is equivalent to doing nothing.
Theorem 12.8 .∅ = U and .U = ∅.
Proof Indeed .z ∈ ∅ if and only if z is a set and it is false that .z ∈ ∅. By
Theorem 12.6, we conclude that .z ∈ ∅ if and only if z is a set, namely if and
only if .z ∈ U. The proof of the second equality is similar.

Definition 12.9 The intersection the members of x is

. x = {z | (∀y)(y ∈ x ⇒ z ∈ y)}.
Definition 12.10 The union of the members of x is

. x = {z | (∃y)(z ∈ y and y ∈ x)}.
Remark 12.7 We are using here an intrinsic notation. Most readers are probably
familiar with the bound variable notation for a “set of sets”, as we have seen in (2.1)
and (2.2).
Important: Sets or Elements?

In my experience, the hardest step of learning axiomatic Set Theory consists in
getting rid of the naïve idea that an element cannot be a set. Consider the following
identity:

. {a, b, c} = a ∪ b ∪ c.

This is actually correct, since .x ∈ {a, b, c} if and only if .x ∈ a or .x ∈ b or
.x ∈ c. However, at the beginning, we tend to believe that the equality is non-sense,
since the right-hand side looks like a union of elements, which is naïvely undefined
(what is .0 ∪ 1, if 0 and 1 represent the usual natural number we first met at school?).
Well, this is precisely how andwhen abstraction is needed. To be honest, in naïve
set theory nobody ever writes . {a, b, c}, and the issue disappears. But a, b and c
may very well be sets (or classes)!
The popular belief that lower-case letters are elements and upper-case letters are
sets/classes is, in this perspective, tragic. It should be encouraged only if a basic
knowledge of Set Theory is sufficient.

Theorem 12.9 . ∅ = U and . ∅ = ∅.

Proof For each x, .x ∈ ∅ if and only if x is a set and x belongs to each member
of .∅. We already know that .∅ contains no member at all, hence .x ∈ ∅ if and only
if x is a set, i.e. .x ∈ U.
Similarly, for each x, .x ∈ ∅ if and only if x is a set and x belongs
to some
member of .∅. Since .∅ has no member, such an x cannot exist. Hence . ∅ contains
no member.

Definition 12.11 For each x and y, .x ⊂ y if and only if for each z, if .z ∈ x then
z ∈ y. In this case we say that x is a subclass of y, or that the class x is contained in
.
the class y, or that the class y contains the class x.

It should be noticed that .x ⊂ y does not exclude that .x = y. We will not use
x ⊂ y in the sense of “every element of x is an element of y and .x = y”.
.
Remark 12.8 It is tempting to confuse .∈ and .⊂. The language does not help, since
x ∈ y is often read “y contains x”. However it would be definitely wrong to use .⊂
.
as a replacement of .∈.
Exercise 12.6 Prove that .∅ ⊂ ∅, but .∅ ∈
/ ∅.
We collect some basic properties of subclasses, whose proofs are a straightfor-
ward application of the definition.
Theorem 12.10 For each x, y and z,
1. .∅ ⊂ x and .x ⊂ U.
2. .x = y if and only if .x ⊂ y and .y ⊂ x.
3. If .x ⊂ y and .y ⊂ z, then .x ⊂ z.
4. .x ⊂ y if and only if .x ∪ y = y.
5. .x ⊂ y if and only if .x ∩ y = x.

6. If .x ⊂ y, then . x
⊂ y and . y ⊂ x.
7. If .x ∈ y, then .x ⊂ y and . y ⊂ x.
Exercise 12.7 Prove the previous Theorem. When union and intersection of classes
are involved,
it could be
helpful to temporarily switch to a bound variable notation,
e.g. . α∈A xα instead of . x.
Question Do sets exist?
Axiom of Subsets If x is a set, there is a set y such that for each z, if .z ⊂ x,

then .z ∈ y.
In words, given a set x there exists a set y such that any subclass of x is a member
of y. Let us see a useful consequence of this axiom.
Theorem 12.11 If x is a set and .z ⊂ x, then z is a set.
Proof If x is a set, there is a set y such that .z ⊂ x implies .z ∈ y. Hence z is a
set.

Exercise 12.8 Read carefully the previous proof, and notice that the axiom of
subsets was not used in its full strength. Hint: by definition, if y is a class and .z ∈ y,
then z is a set. Did we use every property of the classes in the axiom of subsets?

Theorem 12.12 .∅ = U and .U = U.

Proof If .x ∈ U, then x is a set and since .∅ ⊂ x, it follows that .∅ is also
a set.
Then .∅ ∈ U and each member of . U belongs to .∅. It now follows that . U has
no member.
To prove the second statement, Theorem 12.10 implies that . U ⊂ U. On the
other hand, if .x ∈ U (i.e. if x is a set) by the axiom of subsets there exists a set y
that, if .z ⊂ x, then .z ∈ y.

such In particular .x ∈ y, and since .y ∈ U it follows that
x ∈ U. Consequently .U ⊂ U, and the proof is complete.
.

Theorem 12.13 (A Set Exists) If .x = ∅, then . x is a set.
Proof
If .x = ∅, then there exists y such that .y∈ x. Then y is a set, and since
. x ⊂ y, by the axiom of subsets it follows that . x is also a set.

Definition 12.12 (Power Set) For each x,
2x = {y | y ⊂ x} .
.
Theorem 12.14 .U = 2U .
Proof Every element of .2U is a set and therefore belongs to .U. Conversely, every
member of .U is a set and is contained in .U, so that it belongs to .2U .

Theorem 12.15 If x is a set, then .2x is a set, and for each y, .y ⊂ x if and only if
.y ∈ 2 .
x
Exercise 12.9 Prove the previous Theorem, by using the fact that .U = 2U .
We are now ready for the Russel paradox, which has been haunting us so far.
Example 12.1 Let .R = {x | x ∈ / x}. By the classification axiom, .R ∈ R if and only
if .R ∈
/ R and R is a set. Therefore R is not a set.
Theorem 12.16 .U is not a set.
Proof Otherwise .R ⊂ U, and R would be a set by Theorem 12.11.

Definition 12.13 (Singleton) For each x, .{x} = {z | if x ∈ U, then z = x}.
Theorem 12.17 If x is a set, for each y, .y ∈ {x} if and only if .y = x.
Exercise 12.10 Prove the previous Theorem.
Theorem 12.18 1. If x is a set, then .{x} is a set.
2. For each x, .{x} = U if and only if x is not a set.
Proof If x is a set, then .{x} ⊂ 2x and .2x is a set. Hence .{x} is a set. This proves 1.
Let us prove 2. If x is a set, then .{x} is a set and therefore is not equal to .U (since
.U is not a set). If x is not a set, then .x ∈
/ U and .x ∈ {x} by definition.

Axiom of Finite Union If x is a set and y is a set, then .x ∪ y is a set.
Definition 12.14 (Unordered Pair) For each x and y,
{x, y} = {x} ∪ {y}.

.
Remark 12.9 The symbol .{xy} has some advantages over .{x, y}. However we
prefer the standard notation with a comma which cannot be confused with “the
singleton of the product xy”, as soon as a product of x and y is defined and denoted
by xy.
If x and y are sets, by the axiom of union also .{x, y} is a set. In general, however,
{x, y} is a class.
.
Definition 12.15 (Ordered Pair) For each x and y,
(x, y) = {{x}, {x, y}}

.
is the ordered pair of x and y.

Theorem 12.19 (Equality of Ordered Pairs) If x is a set, y is a set and .(x, y) =
(u, v), then .x = u and .y = v.
Proof Since x and y are sets, so is .(x, y) as the unordered pair of two sets. If
.x = y, then .(x, x) = {{x}}, and also .u = v. If .x = y, then .(x, y) = (u, v) implies
.{x} = {u} and hence .x = u. This in turn implies .{x, y} = {x, v}, and thus .y = v.

Theorem 12.19 contains the essential information about ordered couples. For the
sake of completeness we record now a deeper survey of both unordered and ordered
couples. We omit their proofs, since we will not refer to them in the sequel.
Theorem 12.20
1. If x is a set and y is a set, then .{x, y} is a set and .z ∈ {x, y} if and only if .z = x
or .z = y. Furthermore, .{x, y} = U if and only if x is not a set or y is not a set.
2. For each x and y, .(x, y) is a set if and only if x is a set and y is a set. If .(x, y) is
not a set, then .(x, y) = U.
Theorem 12.21 For all x and y, there results

y) = x;−1
1. . (x,
2. . {(x, y)} = y.
Proof Indeed statement 1. follows from

. (x, y) = {{x}, {x, y}} = {x} = x.

Statement 2. follows from 1. and the fact that . {(x, y)}−1 = (y, x).

We are therefore led to the following definition, which is mainly interesting in
the case of an ordered pair.
Definition 12.16 For each x,

π1 x =
. x

π2 x = {x}−1 .
We call .π1 x the first coordinate of x, and .π2 x the second coordinate of x.
We now want to define functions. As we saw in Chap. 1, a function is a particular
type of relation, and relations were defined as sets of ordered pairs. It is by now
clear that we can formulate the following definition.
Definition 12.17 A class r is a relation if and only if for each member z of r there
exist x and y such that .z = (x, y).
Remark 12.10 We state clearly that a relation is not a set of ordered pairs. It is a
class of ordered pairs.
Definition 12.18 The composition .r ◦ s of the relations r and s is

for some x, some y and some z, u =
r ◦ s = u
.
(x, z), (x, y) ∈ s and (y, z) ∈ r
Remark 12.11 To save space, we will abbreviate .{u | for some x, some z, .u =
(x, z) and . . . .} with .{(x, z) | . . .}. In particular, .r ◦ s = {(x, z) | for some y,
.(x, y) ∈ s and .(y, z) ∈ r}.
Definition 12.19 For each relations r, .r −1 = {(x, y) | (y, x) ∈ r}. The relation
r −1 is the inverse relation of r.
.
Exercise 12.11 Prove that, for each relation r and each relation s, .(r −1 )−1 = r,
.(r ◦ s)
−1 = s −1 ◦ r −1 .
Definition 12.20 A relation f is a function if and only if for each x, each y and
each z, if .(x, y) ∈ f and .(x, z) ∈ f , then .y = z.
Two functions can always be composed as relations. But they can also be
composed as functions.
Theorem 12.22 If f is a function and g is a function, then .g ◦ f is a function.
Proof Assume that .(x, y) ∈ g ◦f , .(x, z) ∈ g ◦f . There exists .y1 such that .(x, y1 ) ∈
f and .(y1 , y) ∈ g. Similarly there exists .y2 such that .(x, y2 ) ∈ f and .(y2 , z) ∈ g.
Since f is a function, .y1 = y2 . For the same reason, .y = z, and the proof is
complete.

Definition 12.21 For each f ,
. dom f = {x | (∃y)(x, y) ∈ f }
and
. ran f = {y | (∃x)(x, y) ∈ f }.
The class .dom f is the domain of f , the class .ran f is the range of f .
Exercise 12.12 Prove that .dom U = U = ran U. Hint: If .x ∈ U, then x is a set
and both .(x, ∅), .(∅, x) belong to .U.
Definition 12.22 For each f and each x, we define

f (x) =
. {y | (x, y) ∈ f }.
Hence .z ∈ f (x) if and only if z belongs to the second coordinate of each member
of f whose first coordinate is x. The class .f (x) is the value of f at x.
Remark 12.12

. {y | (x, y) ∈ f } = {z | for each w, if w ∈ {y | (x, y) ∈ f } then z ∈ w},
and this is—in general—different than

{y | ∃z(z ∈ x) ∧ (y = f (z))} =
. {π2 z | π1 z ∈ x}.
Remark 12.13 The previous Definition is usually a source of bad nightmares. It

looks totally different than naïve definition of .f (x) as the unique y such that .f (x) =
y. Nightmares disappear as soon as we realize that MK theory does not distinguish
between sets and points. By the way, a point is just another name for an element
of some class, and therefore a point is just another name for a set. But there is also
another reason behind Definition 12.22: the image of x is always defined, no matter
if x is an element of .dom f , and no matter if f is actually a function. Consider for
example
f = {(1, 1), (1, 2), (3, 4), (9, 9)} .

.
Clearly f is not a function, and yet .f (3) = 4, .f (9) = 9 are defined. But what
is the image of 1? In the naïve sense, 1 this question is meaningless, since f is
not a function. In MK theory .f (1) must belong to both .{1} and .{2}, i.e. .f (1) =
∅. Similarly, .f (0) = ∅, since 0 is not an element of the domain of f . Needless
to say, nobody really needs to define the image of a point outside the domain, in
everyday mathematics. But Axiomatic Set Theory exists because we need a coherent
deductive theory that does not lean on intuition to prove or disprove statements about
sets.
Theorem 12.23 If .x ∈
/ dom f , then .f (x) = U. If .x ∈ dom f , then .f (x) ∈ U.
Proof In the first case, .{y | (x, y) ∈ f } = ∅, and the intersection of the empty class
is .U. In the second case, .{y | (x, y) ∈ f } = ∅, and Theorem 12.13 implies that

. {y | (x, y) ∈ f } is a set.

Theorem 12.24 If f is a function, then .f = {(x, y) | y = f (x)}.
Exercise 12.13 Prove the previous Theorem.
Theorem 12.25 (Equality of Functions) For each function f and each function
g, .f = g if and only if .f (x) = g(x) for each x.
Proof Indeed .f = g if and only if .{(x, y) | y = f (x)} = {(x, y) | y = g(x)} and
this happens if and only if .f (x) = g(x) for each x.

We are in a position to enlarge our collection of axioms.
Axiom of Substitution If f is a function and .dom f is a set, then .ran f is a

set.

Axiom of Amalgamation If x is a set then . x is a set.
Remark 12.14 We should compare Theorem 12.13 with the axiom of amalgama-
tion.
Definition 12.23 For each x and each y,
.x × y = {(u, v) | u ∈ x, v ∈ y}.
The class .x × y is the cartesian product of the classes x and y.

Theorem 12.26
(a) If u and y are sets, then .{u} × y is a set.
(b) If x and y are sets, then .x × y is a set.
Proof Consider the function .{(w, z) | w ∈ y and z = (u, w)}. Its domain is y and
its range is .{u} × y. By the axiom of substitution, .{u} × y is a set. This proves (a).
To prove (b), let f be the function .f = {(u, z) | u ∈ x and z = {u} × y}. By the
axiom of substitution, .ran f is a set. But .ran f = {z | ∃u(u ∈ x) ∧ (z = {u} × y)}.

Since .x × y = ran f , the axiom of amalgamation implies the result.

Theorem 12.27 If f is a function and .dom f is a set, then f is a set.
Proof Indeed .f ⊂ dom f × ran f . Since the right-hand class is a set, also f is a
set.

Definition 12.24 For each x and each y,
y x = {f | f is a function, dom f = x and ran f ⊂ y}.

.
Theorem 12.28 If x and y are sets, then .y x is a set.

Proof For each .f ∈ y x we have .f ⊂ x × y, and .x × y is a set. Hence .f ∈ 2x×y ,
and .2x×y is a set. We conclude that .y x ⊂ 2x×y , and the axiom of subsets implies
that .y x is a set.

A remarkable fact is that we a function is not a triple .(X, Y, f ) in which X is the
domain, Y is the codomain, and f is the law. However, it is convenient to introduce
some terminology that may be recall the Calculus approach.
Definition 12.25
(i) f is on x if and only if f is a function and .dom f = x.
(ii) f is to y if and only if f is a function and .ran f ⊂ y.
(iii) f is onto y if and only if f is a function and .ran f = y.
The last axiom of Kelley’s set theory is a returning object of contemporary
mathematics.
Definition 12.26 (Choice Function) A class c is a choice function if and only if c
is a function and .c(x) ∈ x for each .x ∈ dom c.
Remark 12.15 The existence of a choice function ensures the possibility of simul-
taneously select a member from each set that belongs to the domain of the choice
function. This looks like an almost trivial fact, but it cannot be deduced from the
previous axioms.
Axiom of Choice There exists a choice function c whose domain is .U \ ∅.
Remark 12.16 Equivalently, the axiom of choice gives us a function c such that
c(x) ∈ x for each non-empty set x.
.
The axiom of choice is not necessary to develop all modern mathematics. It

is however essential to prove a few results that are of fundamental importance in
mathematical analysis. We present now a list of axioms which are equivalent to the
axiom of choice, although we will not prove these equivalences in this book. We
will follow [2], but a richer statement is also presented in [3].
Well-Ordering Principle Every set can be well-ordered. More precisely, any non-
empty subset has a smallest element.
Definition 12.27 A subset C of a partially ordered set .(P , <) is chain in P if and
only if C is totally ordered by .<. An element .u ∈ P is an upper bound of C if and
only if .c ≤ u for every .c ∈ C. Finally, an element .a ∈ P is a maximal element if
there is no .x ∈ P such that .a < x.
Zorn’s Lemma 1 Let .(P , <) be a non-empty partially ordered set. If every chain
in P has an upper bound, then P has a maximal element.
Definition 12.28 A collection .F of sets has finite character if and only if the
following condition holds: for every X, .X ∈ F if and only if every finite subset
of X belongs to .F.
Tukey’s Lemma 2 If a non-empty collection .F of sets has finite character, then .F
has a maximal element with respect to the inclusion .⊂.
Theorem 12.29 The following statements are equivalent:
(i) the Axiom of Choice;
(ii) the Well-ordering Principle;
(iii) Zorn’s Lemma;
(iv) Tukey’s Lemma.
Proof Suppose (i), and let S be any set. Let F be a choice function on the familyof
all non-empty subsets of S. Now we define .a0 = F (S), .aξ = F (S \ aη η < ξ ).
The construction stops as soon as we exhaust all elements of S. Hence (ii) holds.
Now assume that (ii) holds, and let .(P , <) be a non-empty partially ordered
set. Assume that every chain of P has an upper bound. To construct a maximal
element, we start from the assumption that P can be well-ordered, so that there is
an enumeration

P = p0 , p1 , . . . , pξ , . . .
. ξ <α
for some ordinal number .α. We set .c0 = p0 and .cξ = pγ , where .γ is the smallest

ordinal such that.pγ is an upper bound of the chain .C = cη η < ξ and .pγ ∈ / C.

We remark that . cη η < ξ is always a chain, and that .pγ exists unless .cξ −1 is
a maximal element of P . In the end, the construction must stop, and we obtain a
maximal element of P . This proves (iii).
Suppose now that (iii) holds. We consider a non-empty family .F of sets and we
assume that .F has finite character. Clearly .F is partially ordered by inclusion. If .C is
a chain in .F and if .A = {X | X ∈ C}, then every finite subset of A belongs to .F and
therefore .A ∈ F. It follows at once that A is an upper bound of .C. We may apply
Zorn’s Lemma and obtain a maximal element of the collection .F. This proves the
validity of (iv).
Finally, assume that (iv) holds, and let .F be a collection of non-empty sets. We
need to construct a choice function on .F. To this aim we consider the collection
G = {f | f is a choice function on some E ⊂ F} .

.
Since a subset of a choice function is a choice function, it follows that .G has finite
character. By assumption .G possesses a maximal element F . By maximality, the
domain of F is .F, and the proof is complete.

12.2 From Sets to N
Our next task is to construct the (positive) integers from set theory. As we will
see, this is somehow a painful task, while the subsequent steps—from integers to
rationals, and from rationals to real numbers—are rather standard.
Axiom of Regularity For each x, if .x = ∅, there is .y ∈ x such that .x ∩y = ∅.
Remark 12.17 Using Kelley’s words, the axiom of regularity avoids the possibility
that there exist a class z whose member exist by “taking in each other’s laundry, in
the sense that every member of z consists of members of z.”
Axiom of Infinity For some y, y is a set, .∅ ∈ y, and .x ∪ {x} ∈ y whenever

x ∈ y.
.
Remark 12.18 An immediate consequence of this axiom is that .∅ is actually a set,

since it is an element of the set y. Notice that .∅ was merely a class, before adding
the axiom of infinity.
Theorem 12.30 For each x, .x ∈
/ x.
Proof If not, .x ∈ x, x is a non-empty set and is the only member of .{x}. By the
axiom of regularity there exists y in .{x} such that .y ∩ {x} = ∅, and necessarily
.y = x. But then .y ∈ y ∩ {x}, against .y ∩ {x} = ∅.

Theorem 12.31 The statement
∀x∀y(x ∈ y ∧ y ∈ x)
.
is false.
Proof If .x ∈ y and .y ∈ x, then both x and y are sets and are the only members of
{z | z = x or z = y}. To this class we apply the axiom of regularity, and we reach a
.
contradiction exactly as in the proof of Theorem 12.30.

12.2 From Sets to N 179
Good. Let us now sketch the basic idea to construct (positive) integers from set
theory. We want to start from .0 = ∅, then define .1 = 0∪{0}, .2 = 1∪{1}, .3 = 2∪{2},
and so on. Of course this is not a full definition, since the previous “and so on”
requires explanation. We must now provide a rigorous definition of the ordinals.
Definition 12.29 (The .∈-Relation) .E = {(x, y) | x ∈ y}.
Theorem 12.32 E is not a set.
Proof Assume that .E ∈ U, then .{E} ∈ U and .(E, {E}) ∈ E. We recall that
.(x, y) = {{x}, {x, y}} and, if .(x, y) is a set, .z ∈ (x, y) if and only if .z = {x} or
.z = {x, y}. We deduce that .E ∈ {E} ∈ {{E}, {E, {E}}} ∈ E. Let us summarize:
we know that .a ∈ b ∈ c ∈ a. A contradiction now arises if we apply the axiom of

regularity to the class .{z | .z = a or .z = b or .z = c}. Hence E cannot be a set.

Definition 12.30 (Complete Classes) For each x, x is complete if and only if for
each .y ∈ x, .y ⊂ x.
Remark 12.19 We prefer to use the word “complete” instead of Kelley’s “full”. If it
is true that “complete” refers to several mathematical properties, we believe that it
is impossible to systematically select different words.
Definition 12.31 (Ordinals) A class x is an ordinal if and only if x is complete and
the following condition holds:
(C) when u and v belong to x, either .(u, v) ∈ E, or .(v, u) ∈ E, or .u = v.
Definition 12.32 .R = {x | x is an ordinal}. A class x is an ordinal number if and
only if .x ∈ R.
In particular, an ordinal number is a set, being a member of R.
Definition 12.33 (Successor) For each x, .x + 1 = x ∪ {x}.
To move from ordinals to integers, we need to explore relation E more deeply.
Definition 12.34 We say that .E −1 well-orders x if and only if condition (C) is
satisfied, and for each .y ⊂ x, .y = ∅, there exists .z ∈ y such that for each .w ∈ y, it
is false that .(w, z) ∈ E −1 .
Definition 12.35 (Positive Integers) A class x is an integer if and only if x is an
ordinal and .E −1 well-orders x.
Definition 12.36 .N = {x | x is an integer}.
Remark 12.20 In axiomatic set theory, the letter .ω is used instead of .N. We believe
that .N is a better choice from the viewpoint of a mathematical analyst.
The next result contains the celebrated Peano axioms of positive integers. A
mathematical analyst need not know its proof, which therefore we omit.
Theorem 12.33 (Peano Axioms)

1. If .x ∈ N, then .x + 1 ∈ N.
2. ∅ ∈ N, and if .x ∈ N, then .∅ = x + 1.
.
3. If x and y are members of .N and .x + 1 = y + 1, then .x = y.

4. If .x ⊂ N, .0 ∈ x and .u + 1 ∈ x whenever .u ∈ x, then .x = N.
Theorem 12.34 .N ∈ R. In particular, .N is a set.
Proof By the axiom of infinity, there exists a set y such that .∅ ∈ y and .x + 1 ∈ y
whenever .x ∈ y. By Theorem 12.33, statement 4, .N ∩ y = N and hence .N is a set
because .N ⊂ y. Since .N consists of ordinal numbers, Definition 12.32 is satisfied
by .N, and therefore .N ∈ R.

Definition 12.37 (Ordering) For each x and y, .x < y if and only if .x ∈ y, i.e.
(x, y) ∈ E. Similarly, .x ≤ y if and only if .x ∈ y or .x = y.
.
With this definition in mind, we see that positive integers are ordered in a way
that is compatible with the usual naïve definition of inequality between numbers.
12.3 A Summary of Kelley’s Axioms
For the reader’s convenience, we summarize here Kelley’s axioms.
Axiom of Extent For each x and each y, it is true that .x = y if and only if
for each z, .z ∈ x if and only if .z ∈ y.
Axiom of the Classifier For each u and y, .u ∈ {x | x ∈ y} if and only if u is

a set and .u ∈ y.
Classification Axiom Scheme An axiom results if in the following .α and .β

are replaced by variables, A by a formula .A and B by the formula obtained
from .A by replacing each occurrence of the variable that replaced .α by the
variable that replaced .β:
For each .β, .β ∈ {α | A} if and only if .β is a set and B.

12.4 Set Theory According to J.D. Monk 181
Axiom of Subsets If x is a set, there is a set y such that for each z, if .z ⊂ x,

then .z ∈ y.
Axiom of Union If x is a set and y is a set, then .x ∪ y is a set.
Axiom of Substitution If f is a function and .dom f is a set, then .ran f is a

set.

Axiom of Amalgamation If x is a set then . x is a set.
Axiom of Choice There exists a choice function c whose domain is .U \ ∅.
12.4 Set Theory According to J.D. Monk
As we said at the beginning of this chapter, we want to introduce a variation of

Kelley’s construction. Although essentially equivalent, Monk’s revisited Set Theory
is based on a stronger Axiom of Choice, and on a more restrictive definition of
relations and functions.
Definition 12.38 The primitive notions are those of classes and membership. We
will use capital letters like A, B,. . . , X, Y , Z for classes, and the usual symbol .∈ for
membership. The negation of .∈ will be denoted by .∈. /
Axiom 1 (Extensionality Axiom) .∀A .∀B .(∀C(C ∈ A ⇐⇒ C ∈ B) ⇒

A = B).
By this Axiom, two classes A and B are the same class if and only if they share
the same members.
Definition 12.39 (Sets) A class A is a set if and only if there exists a class B such
that .A ∈ B. A is a proper class if and only if A is not a set.
As a rule, sets will be denoted by lower case letters: a, b,. . . , x, y, z.
Exercise 12.14 Prove that .∀a .∃B .(a ∈ B).
Exercise 12.15 Prove that .∀x(x ∈ A ⇐⇒ x ∈ B) ⇒ A = B.
Definition 12.40 The expressions .A = A, .A = B, .A = C, . . . , .B = A, .B = B,
.B = C, . . . .C = A, .C = B, .C = C,. . . are all set-theoretical formulas, as are .A ∈ A,
.A ∈ B, .A ∈ C, . . . , .B ∈ A, .B ∈ B, .B ∈ C, . . . .C ∈ A, .C ∈ B, .C ∈ C,. . .
If .ϕ and .ψ are set-theoretical formulas, so are .¬ϕ, .ϕ ∨ ψ, .ϕ ∧ ψ, .ϕ ⇒ ψ,

.ϕ ⇐⇒ ψ, .∃Aϕ, .∃Bϕ,. . . , .∀Aϕ, .∀Bϕ, . . . Set-theoretical formulas can only be
obtained by finitely many applications of the processes just mentioned.
Axiom 2 (Class-Building Axiom) If .ϕ(X) is a set-theoretical formula not

involving the letter A, then the following is an axiom:
∃A∀X(X ∈ A ⇐⇒ X is a set and ϕ(X)).
.
Similarly, if .ϕ(X) is a set-theoretical formula not involving the letter B, then

the following is an axiom:
∃B ∀X(X ∈ B ⇐⇒ X is a set and ϕ(X)),
.
and so on for other letters. Letters different than X may also be used.
This axiom allows us to define classes by specifying the properties which each
member must satisfy.
Example 12.2 If .ϕ(X) is .X ∈ X, the class-building axiom allows us to construct
the class of all sets.
Definition 12.41 For any set-theoretical formula .ϕ(X) not involving A, let
. {X | ϕ(X)}
be the unique class A such that
∀X(X ∈ A ⇐⇒ X is a set and ϕ(X)).

.
As before, letters other than A and X may be used.

Definition 12.42 .A ⊂ B if and only if .∀C .(C ∈ A ⇒ C ∈ B). We say that A is
a subclass of B. If A is a set, then A is a subset of B.
Theorem 12.35 .A ⊂ B if and only if .∀x(x ∈ A ⇐⇒ x ∈ B).
Proof If .A ⊂ B, certainly the conclusion is true. Conversely, assume that .∀x(x ∈

A ⇐⇒ x ∈ B). If C is an arbitrary class such that .C ∈ A, then C is a set and
.C ∈ B. Since C is arbitrary, .∀X(X ∈ A ⇐⇒ X ∈ B), and thus .A ⊂ B.

Axiom 3 (Power-Set Axiom) .∀a∃b∀C(C ⊂ a ⇒ C ∈ b).
Axiom 4 (Pairing Axiom) .∀a∀b∃c(a ∈ c ∧ b ∈ c)
Beware that the set c need not contain only a and b.
Axiom 5 (Union Axiom) .∀a∃b∀C(C ∈ a ⇒ C ⊂ b).
This axiom becomes clearer if we imagine that a is a family of sets, so that there
exists another set b which contains every member of a,
Definition 12.43 (Empty Class) .∅ = {x | x = x}.
Theorem 12.36 For all X, .X ∈
/ ∅.
Proof If .X ∈ ∅, then X is a set and .X = X, a contradiction.

Definition 12.44 The intersection of two classes A and B is defined as
.A ∩ B = {x | x ∈ A ∧ x ∈ B} .
The following axiom is usually considered as an useless one. However classes

are so large that counterintuitive possibilities may arise, such as the existence of a
class A with .A ∈ A.
Axiom 6 (Regularity Axiom) .∀A(A = ∅ ⇒ ∃X(X ∈ A ∧ X ∩ A = ∅)).
Definition 12.45 The successor of a class A is the class
S(A) = {x | x ∈ A ∨ x = A} .
.
Remark 12.21 We cannot write .S(A) = A ∪ {A}, since singletons have not been
introduced yet.
Theorem 12.37 If A is a proper class, then .S(A) = A.
Proof Suppose that X is any class. If .X ∈ A, then X is a set, and .X ∈ A or .X = A.
Hence .X ∈ S(A). If .X ∈ S(A), then .X ∈ A or .X = A, and X is a set. But A is not
a set, hence .X = A. By the axiom of extensionality, .A = S(A).

Axiom 7 (Axiom of Infinity) .∃a(∅ ∈ a ∧ ∀X(X ∈ a ⇒ S(X) ∈ a)).
So a set a exists such that .∅ ∈ A, .S(∅) ∈ a, .S(S(∅)) ∈ a, and so on.

Definition 12.46 (Unordered Pairs) The unordered pair1 of two classes A and B
is
. {A, B} = {x | x = A ∨ x = B} .
Theorem 12.38 If a and b are sets, so is .{a, b}.

Proof By the pairing axiom, a set c exists such that .a ∈ c and .b ∈ c. Thus .{a, b} ⊂
c. By the power-set axiom, a set d exists such that .∀X(X ⊂ c ⇒ X ∈ d). In
particular we have .{a, b} ∈ d, hence .{a, b} is a set.

Definition 12.47 The singleton A is .{A} = {A, A}.
The following result describes the equality of unordered pairs.
Theorem 12.39 If .{a, b} = {c, d}, then .a = c and .b = d, or .a = d and .b = c.
Proof Clearly .a ∈ {a, b}. By assumption .a ∈ {c, d}, so that .a = c or .a = d. These
cases are symmetric, hence we may assume that .a = c. Similarly .b ∈ {a, b}, and
.b = c or .b = d. If .b = d, the proof is complete. If .b = c, then .a = b = c.
Again, .d ∈ {a, b}, and by the same token .d = a or .d = b. Thus .a = b = c = d,

and the desired conclusion has been reached.

We now propose a familiar definition of ordered pair.
Definition 12.48 The ordered pairs of two classes A and B is
(A, B) = {{A} , {A, B}} .

.
By imitating Theorem 12.38, one can prove that the ordered pair of two sets is a
set.
1 The old-fashioned term doubleton has been used as well.

Theorem 12.40 If .(a, b) = (c, d) then .a = c and .b = d.

Proof Since .(a, b) = {{a}, {a, b}} and .(c, d) = {{c}, {c, d}} and .{a}, .{a, b}, .{c}
and .{c, d} are all sets, by Theorem 12.39 two cases may happen. In the first case,
.{a} = {c} and .{a, b} = {c, d}. Now .a ∈ {a}, thus .a ∈ {c} and .a = c. It remains to
show that .b = d. We apply Theorem 12.39 again and we conclude that either .a = c
and .b = d, or .a = d and .b = c. In both cases .a = c and .b = d.
In the second case, .{a} = {c, d} and .{a, b} = {c}. Now .c ∈ {c, d}, so that .c ∈ {a}
and consequently .c = a. Similarly, .d = a and .b = c, so that .a = c and .b = d. The
proof is complete.

Let us turn to relations.
Definition 12.49 A class R is a relation if and only if
∀A (A ∈ R ⇒ ∃c∃d(A = (c, d))) .

.
The domain of a relation R is the class
. Dom R = {x | ∃y((x, y) ∈ R)} .
The range of R is the class
. Ran R = {y | ∃x((x, y) ∈ R)} .
Definition 12.50 A class F is a function if and only if F is a relation and
∀x∀y∀z((x, y) ∈ F ∧ (x, z) ∈ F ⇒ y = z).

.
As usual, a relation is just a any class of ordered pairs. A function is just a “rule”
which assigns to each set of its domain a unique set of its codomain.
Axiom 8 (Axiom of Substitution) If F is a function and .Dom F is a set,

then .Ran F is a set.
Monk’s final axiom is a very strong form of the Axiom of Choice. It is indeed
stronger than the usual one.
Axiom 9 (Relational Axiom of Choice) If R is a relation, there exists a

function F such that .F ⊂ R and .Dom F = Dom R.
With these axioms one can construct the familiar boolean algebra of sets and of
relations, with minor differences with respect to our previous discussion. Monk’s
Theory of Sets is actually equivalent to Kelley’s one. The strength of the Relational
Axiom of Choice may be used to prove stronger results directly in Monk’s theory,
but most mathematicians do not usually see any concrete difference.
12.5 ZF Axioms
As we have stated at the beginning, the most popular axiomatization of Set Theory
goes under the name of Zermelo and Fraenkel. Although we have preferred another
approach, we believe it may be useful for the reader to have at least an account of
ZF. It should be noticed that ZF requires both sets and elements.
Axiom 10 (Extensionality Axiom) Two sets are equal if and only if they
have the same elements. In symbols:
∀A∀B(A = B ⇐⇒ ∀x(x ∈ A ⇐⇒ x ∈ B))

.
Axiom 11 (Empty Set) There exists a set with no elements. In symbols:
∃A∀x(x ∈
. / A)
Axiom 12 (Subset Axiom) Let .ϕ(x) be a formula. For every set A there
exists a set S that consists of all the elements .x ∈ A such that .ϕ(x) holds.
In symbols,
∀A∃S∀x(x ∈ S ⇐⇒ (x ∈ A ∧ ϕ(x))
.
Axiom 13 (Pairing Axiom) For every u and v there exists a set that consists
of just u and v. In symbols:
∀u∀v∃A∀x(x ∈ A ⇐⇒ (x = u ∨ x = v))
.
12.5 ZF Axioms 187
Axiom 14 (Union Axiom) For every set .F there exists a set U that consists
of all the elements that belong to at least one set in .F. In symbols,
∀F∃U ∀x(x ∈ U ⇐⇒ ∃C(x ∈ C ∧ C ∈ F))

.
Axiom 15 (Power Set Axiom) For every set A there exists a set P that
consists of all the sets that are subsets of A. In symbols:
∀A∃P ∀x(x ∈ P ⇐⇒ ∀y(y ∈ x ⇒ y ∈ A))

.
Axiom 16 (Infinity Axiom) There exists a set I that contains the empty set
as an element and whenever .x ∈ I , then .x ∪ {x} ∈ I . In symbols:
∃I (∅ ∈ I ∧ ∀x(x ∈ I ⇒ x ∪ {x} ∈ I ))
.
Axiom 17 (Replacement Axiom) Let .ψ(x, y) be a formula. For every set A,

if for each .x ∈ A there exists a unique y such that .ψ(x, y), then there exists a
set S that consists of all the elements y such that .ψ(x, y) for some .x ∈ A. In
symbols:
∀A((∀x ∈ A)∃!y ψ(x, y) ⇒ ∃S∀y(y ∈ S ⇐⇒ (∃x ∈ A)ψ(x, y)))

.
Axiom 18 (Regularity Axiom) Every non-empty set A has an element that

is disjoint from A. In symbols:
∀A(A = ∅ ⇒ ∃x(x ∈ A ∧ x ∩ A = ∅))

.
12.6 From N to Z
Once we have a rigorous foundation of positive integers, the construction of relative

integers is standard. Let us set .X = N × N, the cartesian product of the positive
integers with themselves.
Definition 12.51 For each .(m, n) and .(m% , n% ) in X, we say that .(m, n) ∼ (m% , n% )
if and only if .m + n% = m% + n.
Remark 12.22 The intuition behind the previous definition is that we want to
identify the “numbers” .m − n and .m% − n% . Actually, .−3 is both equal to .2 − 5 and
.7 − 10. However this intuition cannot be a rigorous definition o relative integers,
since .m − n is not—in general—a positive integer. In other words, we know what

we want, but we cannot just take it for granted that what we want does exist.
Exercise 12.16
1. Prove that .∼ is an equivalence relation.
2. Prove that .(m, n) ∼ (m% , n% ) if and only if .(m + p, n + p) ∼ (m% + p, n% + p)
for each .p ∈ N.
3. Prove that .m = n if .(m, p) ∼ (n, p) or if .(p, m) ∼ (p, n).
Definition 12.52 (Addition) For each .(m, n) and .(m% , n% ) in X, we define
(m, n) + (m% , n% ) = (m + m% , n + n% ).
.
Definition 12.53 (The Relative Integers) The symbol .Z denotes the set .X/ ∼, i.e.
the set of equivalence classes determined by .∼. The equivalence class of the element
.(m, n) is denoted by .[(m, n)].
Definition 12.54 .0 = [(1, 1)].

Definition 12.55 For each .a ∈ Z and .b ∈ Z, we define .a + b as follows. We write
a = [(m, n)] and .b = [(p, q)] for some .(m, n) and .(p, q) in X. Then
.
a + b = [(m + p, n + q)].
.
Exercise 12.17 Prove that the sum .a + b is independent of the particular ordered
couples .(m, n) and .(p, q) that describe the classes a and b. More precisely, show
that .(m + p, n + q) ∼ (m% + p% , n% + p% ) whenever .(m, n) ∼ (m% , n% ) and .(p, q) ∼
(p% , q % ).
Definition 12.56 (Opposite) For each .a ∈ Z, we define .−a as the unique element
x ∈ Z such that .a + x = 0.
.
Exercise 12.18 Prove that the previous definition is consistent, in the sense that .−a
does exist for each .a ∈ Z.
Definition 12.57 (Difference) For each a and b in .Z, we define .a − b = a + (−b).
12.7 From Z to Q 189
We can now embed .N into .Z as follows.

Definition 12.58 .f+ : N → Z is defined by .f+ (n) = [(n + 1, 1)]. We define .Z+ =
ran f+ .
Exercise 12.19 Prove that .[(n + 1, 1)] = [(n% + 1, 1)] implies .n = n% . Deduce that
.f+ is injective, and therefore its restriction .f+ : N → Z+ is a bijection.
By the previous exercise, for each .a ∈ Z+ there exists a unique .n ∈ N such that
f+ (n) = a. We write .n = f+−1 (a).
.
Definition 12.59 Let .s : N → N the standard successor function defined by .s(n) =

n + 1. We define .s % : Z+ → Z+ by .s % (a) = f+ (s(f+−1 (a))).
To conclude, we define the product of relative integers.
Definition 12.60 An operation .∗ is defined on X by
(m, n) ∗ (p, q) = (mp + nq, mq + np).

.
Exercise 12.20 Prove that .(m, n) ∼ (m% , n% ) and .(p, q) ∼ (p% , q % ) imply .(m, n) ∗
(p, q) ∼ (m% , n% ) ∗ (p% , q % ).
By the previous exercise, the product on X can be extended to a product on .Z.
Definition 12.61 For each .a = [(m, n)] and .b = [(p, q)] in .Z, we define
ab = a ∗ b = [(mp + nq, mq + np)].

.
Remark 12.23 In general this product does not possess a inverse element, in the
sense that, given a and b in .Z, .a = [(0, 0)], the equation .a ∗ x = b is not solvable
in .Z.
12.7 From Z to Q
The next extension of our number sets will be the set of rational numbers. While
everybody thinks of a rational number as a fraction like .p/q with .p ∈ Z and .q ∈
N \ {0}, this approach is purely naïve. Indeed we did not—and could not—define a
division operation in .Z. Again, the road goes through equivalence classes.
Definition 12.62 .V = {(r, s) ∈ Z × Z | s = 0}. We define an equivalence relation
∼ on V by setting .(r, s) ∼ (t, u) if and only if .ru = st. Notice that we are writing
.
rs instead of .r ∗ s, and so on. The equivalence class of the element .(r, s) will be
denoted by .[(r, s)].
Definition 12.63 (Rational Numbers) A rational number is any element of .V / ∼.
If .[(r, s)] ∈ V / ∼, we simply write .r/s. The set .Q is the set whose members are all
the rational numbers.
Definition 12.64 .ι : Z → Q is defined by .ι(r) = [(r, 1)] for each r.

Theorem 12.41 .ι is an injective function.
Proof Assume that .ι(x) = ι(y) for some x, y in .Z. Then .[(x, 1)] = [(y, 1)], i.e.
(x, 1) ∼ (y, 1), and this means that .x ∗ 1 = y ∗ 1, hence .x = y.
.

Definition 12.65 (Sum of Rational Numbers) For each .[(r, s)] and .[(t, u)] in .Q,
we define
[(r, s)] + [(t, u)] = [(ru + st, su)].

.
Exercise 12.21 Prove that the previous definition is consistent: if .(r % , s % ) ∼ (r, s)
and if .(t % , u% ) ∼ (t, u), then .(ru + st, su) ∼ (r % u% + s % t % , s % u% ).
Definition 12.66 (Multiplication of Rational Numbers) For each .[(r, s)] and
[(t, u)] in .Q, we define
.
[(r, s)] ∗ [(t, u)] = [(rt, su)].

.
Exercise 12.22 Prove that the previous definition is consistent.

Remark 12.24 It must be observed that our operations of sum and multiplication is
compatible with the corresponding operation of .Z. More precisely,
ι(x) + ι(y) = [(x, 1)] + [(y, 1)]

.
= [(x ∗ 1 + 1 ∗ y, 1 ∗ 1)] = [(x + y, 1)]

= ι(x + y)
and
ι(x) ∗ ι(y) = [(x, 1)] ∗ [(y, 1)]

.
= [(xy, 1)]
= ι(x ∗ y)
Exercise 12.23
1. Prove that .[(1, 1)] is neutral for multiplication: for each .[(r, s)] there results
.[(r, s)] ∗ [(1, 1)] = [(r, s)].
2. Prove that for each .[(r, s)] = [(0, 1)], there exists a unique .[(x, y)] such that
.[(r, s)] ∗ [(x, y)] = [(1, 1)]. Hint: a good candidate is .x/y = s/r.
It is now easy to prove the following result.

Theorem 12.42 The set .Q is a field under the operations of sum and multiplication.
12.8 From Q to R 191
Remark 12.25 Every rational number can be uniquely written as .r/s with .s ≥ 1.
Indeed, if .s ≤ −1, we recall that .[(r, s)] = [(−r, −s)]. In the rest of this section,
we will systematically assume that rational numbers are presented in this form.
Definition 12.67 (Ordering) A relation .< is defined on .Q by .[(r, s)] < [(t, u)] if
and only if .ru < st.
Remark 12.26 If .r/s = r % /s % and .t/u = t % /u% , then .rs % = r % s and .tu% = t % u.
Therefore .[(r, s)] < [(t, u)] if and only if .ru < st if and only if .rus % u% < sts % u%
if and only if .f % suu% < ss % t % u if and only if .r % u% < s % t % if and only if .[(r % , s % )] <
[(t % , u% )]. We deduce that the ordering definition is consistent in .Q.
12.8 From Q to R
Our last step in set and number theory is a concrete construction of real numbers.
We have introduced .R as a totally ordered field which satisfies Dedekind’s axiom.
But this definition remains somehow vague until we prove that such a numerical
structure exists. To this aim several approaches have been proposed. A classical one
is via Dedekind cuts, see [4].
We present a different approach that lies on a more analytical construction, due
to Cantor.
Definition 12.68 A sequence .x = {xn }n of rational numbers is a Cauchy sequence
of rational numbers if for each rational number .a > 0, there exists .N ∈ N such that
.|xm − xn | < a for each .m ≥ N and each .n ≥ N. We denote by .C the set of all
Cauchy sequences of rational numbers.

Definition 12.69 A sequence .x = {xn }n of rational numbers is a zero sequence of
rational numbers if and only if for each rational number .a > 0 there exists .N ∈ N
such that .|xn | < a for each .n ≥ N. We denote by .Z the set of all zero sequences of
rational numbers.
Exercise 12.24 Prove that .Z ⊂ C.
Definition 12.70 Each rational number q induces a constant sequence .q̄ : N → Q
by setting .q̄n = q for each n. We denote by .Q the set of all constant sequences of
rational numbers.
Definition 12.71 If .{xn }n and .{yn }n are sequences in .Q, we define their sum as
n → xn + yn and their product as .n → xn yn .
.
Theorem 12.43 Let .{xn }n and .{yn }n be sequences in .Q.

1. If .{xn }n ∈ C and .{xn }n ∈ C then their product is a member of .C.
2. If .{xn }n ∈ Z and .{xn }n ∈ Z then their product is a member of .Z.
3. If .{xn }n ∈ C and .{xn }n ∈ Z then their product is a member of .Z.
Proof To prove 1, we first remark that .{xn }n and .{yn }n are bounded sequences as in
Proposition 5.3. Then we merely use the triangle inequality:
|xn yn − xm ym | ≤ |(xn − xm )yn | + |xm (yn − ym )| = |xn − xm ||yn | + |xm ||yn − ym |.

.
Since the two sequences and in .C and bounded, the right-hand side of the last
inequality can be made smaller than any given positive rational number by taking m
and n sufficiently large. In a similar way we prove 2 and 3.

Definition 12.72 An equivalence relation on .C is defined as follows: for each .x ∈ C
and each .y ∈ C, .x ∼ y if and only if their difference .x − y is an element of .Z. We
denote by .[x] the equivalence class of the sequence .x ∈ C.
Definition 12.73 A real number is any element of .C/ ∼, i.e. any equivalence class
of a sequence in .C. We denote by .R the set of all real numbers.
Let us stop for a moment: what are we doing? Well, the intuition is that a real
number is the “limit” of a sequence of rational numbers. But this would not be a
good definition, since—of course—there are sequences of rational numbers that do
not converge to a rational number. We are forced to take limits in .Q, but we need to
go outside .Q! The only way out is to remove the requirement that our sequences of
rational numbers have a limit, and this is done by using the Cauchy condition. Then
we identify rational sequences that “converge” to the same number, in the sense that
their difference converges to zero. This is the equivalence relation .∼.
In this way we have avoided any use of the term limit, which would be somehow
misleading at this stage. But we have preserved the original intuition of “adding to
.Q the limits of rational sequences that satisfy the Cauchy property.”
Theorem 12.44 Let x, .x % , y and .y % be elements of .C. If .x ∼ x % and .y ∼ y % , then

% %
.x + y ∼ x + y and .xy ∼ x y .
% %
Proof Let .x % = x + p and .y % = y + q for some p and q in .Z. Then .(x % + y % ) −

(x + y) = p + q ∈ Z by Theorem 12.43. Also .x % y % − xy = py + xq + pq ∈ Z by
Theorem 12.43. The conclusion follows.

Definition 12.74 A sequence .x : N → Q is eventually positive if there exists .N ∈
N such that .xn > 0 for each .n ≥ N.
Definition 12.75 A real number .[x] is positive if each .y ∈ [x] is an eventually
positive sequence of rational numbers. The subset of positive real numbers is
denoted by .R+ .
Remark 12.27 Notice that .R+ is what we usually denote by .(0, +∞). The real
number 0 does not belong to .R+ . Indeed .y ∈ [0] if and only if .y ∈ Z. The constant
rational sequence .0̄ belongs to .Z, but it is not eventually positive.
The following theorem shows that .R is indeed a totally ordered field in the sense
of Chap. 2. Its proof is not difficult, but it is somehow boring. We leave the details
as an exercise for the interested reader.
12.8 From Q to R 193
Theorem 12.45 The set .R with the operations defined by
[x] + [y] = [x + y],

. [x] ∗ [y] = [x ∗ y]
for each .[x] and each .[y] in .R, is a field. It is also totally ordered by the
relation .[x] < [y] if and only if .y − x is a positive real number in the sense of
Definition 12.75.
Luckily enough, .Q may be identified with a subset of .R. Indeed, let .ϕ : Q → R
be the function defined by .ϕ(p) = [p̄], where .p̄ is the constant sequence .n → p.
It can be easily proved that .ϕ is injective, so that .ϕ : Q → ϕ(Q) is a bijection.
Furthermore .ϕ(p + q) = ϕ(p) + ϕ(q) and .ϕ(pq) = ϕ(p)ϕ(q).
The last—and the most intriguing—property of .R is clearly the fact that .R has
the least upper bound property: if a subset E of .R has an upper bound, then there
exists the least upper bound of E. The proof of this fact requires some work.
Theorem 12.46 If .r = [x] is a real number and if there exists .N ∈ N such that
xN ≥ 0, then .r ≥ 0.
.
Proof Indeed, either .r = 0, or r is positive, or .−r is positive. The last possibility

is ruled out by assumption, since every element of the class r would be eventually
positive. But we know that .−x ∈ r and .−x is not eventually positive. We deduce
that either .r = 0 or r is a positive real number.

Theorem 12.47 If r is a positive real number, then there is .p ∈ Q such that .0 0 and some .N ∈ N, either .xn > a or .xn < −a
for each .n ≥ N. Since x is eventually positive, the second alternative cannot hold.
Hence .xn > a whenever .n ≥ N. By Theorem 12.46 .0 < a ≤ r. Let .p = a/2, then
.0 < a/2 < a ≤ r, and the proof is complete.

Theorem 12.48 (Density or Rational Numbers) For each real numbers r, s with
r < s, there exists .q ∈ Q such that .r < q < s.
.
Proof By assumption, .0 < s−r. By the previous results, we can choose .p ∈ Q such
that .0 xm + p for each .m ≥ M. Since x is a
Cauchy sequence, there exists .N ≥ M, .N ∈ N, such that .|xN − nn | ≤ p/4 for each
.n ≥ N.
Let .q1 = xN + p/4 and .q2 = xN + 3p/4. For each .n ≥ N, .q1 − xn = p/4 +
xN − xn ≥ p/4 + |xN − xn | ≥ 0 implies .r ≤ q1 , and similarly .q2 ≤ s. The rational
number .q = (q1 + q2 )/2 satisfies .r < q < s.

Theorem 12.49 Let A be a nonempty subset of .R. If A has an upper bound, then a
number .p ∈ Q exists such that p is not an upper bound of A but .p + 1 is an upper
bound of A.
Proof Pick .a ∈ A. Surely .r = a − 1 is not an upper bound of A, since .r < a. Let

.s < r, .s ∈ Q. A fortiori s is not an upper bound of A. The sequence .n → sn = s + n
is unbounded from above. Hence there exists .n ∈ N such that .s+n is an upper bound
of A. By induction, we know that there exists a smallest positive integer m such that
.q = s + m is an upper bound of A. Then .p = q − 1 = s + m − 1 is not an upper
bound of A.

Theorem 12.50 Let A be a nonempty subset of .R. If A has an upper bound, then
there exist two sequences of rational numbers .{pn }n and .{qn }n such that, for each
.n ∈ N, .pn is not an upper bound of A and .qn is an upper bound of A. Furthermore,
.qn − pn = 2
1−n .
Proof Let p and q be as in Theorem 12.49. We construct the two sequences

inductively. Let .p1 = p and .q1 = q. Suppose that .pn and .qn have been chosen, and
let .sn = (pn + qn )/2. If .sn is not an upper bound of A, then we set .pn+1 = sn ,
.qn+1 = qn . Otherwise we set .pn+1 = pn and .qn+1 = sn . It follows that
.qn+1 − pn+1 = 2
−n , and the proof is complete.

Theorem 12.51 The sequences .{pn }n and .{qn }n are equivalent Cauchy sequences.
Proof By construction, .pn ≤ pn+1 ≤ qn+1 ≤ qn for each n. It follows that .pn ≤
pn+k ≤ qn+k ≤ qn for each k. Thus, if .m ≥ n, .|pn − pm | ≤ |pn − qn | ≤ 21−n .
Since .{21−n }n is a zero sequence, we see that .{pn }n is a Cauchy sequence. By the
same token, .{qn }n is a Cauchy sequence. Since .qn − pn = 21−n , .{pn }n ∼ {qn }n .

Theorem 12.52 (Existence of the Least Upper Bound) Let A be a non-empty
subset of .R. If A has an upper bound, then there exists in .R a least upper bound of
A.
Proof We begin with the sequences .{pn }n and .{qn }n constructed above. Since they
are equivalent Cauchy sequences, they represent the same real number r. We claim
that .r = sup A.
Suppose r is not an upper bound of A. Then there exists .a ∈ A such that .r < a.
By Theorem 12.48 we can select a rational number q with .r < q < a. The sequence
.n → q − qn represents the number .q − a ∈ R. Hence this sequence must be
eventually positive., and there exists .N ∈ N such that .q > qn for each .n ≥ N. This
implies .qn < q < a for each .n ≥ N, so that .qn is not an upper bound of A, against
the property of .qn . This contradiction proves that r is an upper bound of A.
To see that .r = sup A, we suppose that A possesses an upper bound .s ∈ R with
.s < r. Again, we find .p ∈ Q such that .s p > s for large .n ∈ N. This implies that .pn is an upper
bound of A, against the main property of .pn . This contradiction shows that no real
number less than r can be an upper bound of A, i.e. .r = sup A.

12.9 About the Uniqueness of R 195
12.9 About the Uniqueness of R
One of the most folkloristic statements of Mathematical Analysis says that the
system of real numbers is unique, up to isomorphisms. However, the vast majority
of the textbooks that I have used in my life omit a rigorous proof of this (true) fact.
Following the survey [1], we will learn that all models (i.e. constructions) of .R are
indeed equivalent up to renaming objects.
Definition 12.76 Let .K = (K, +, ·, ≤) be a totally ordered field, in the sense
described in Chap. 3. For any non-empty subsets X and Y of .K, we write .X ≤ Y to
mean that
∀x∀y(x ∈ X ∧ y ∈ Y ⇒ x ≤ y).
.
Definition 12.77 (Dedekind Completeness) The field .K is Dedekind complete if

and only if for every non-empty subsets X and Y of .K such that .X ≤ Y , there exists
.z ∈ K with the property that
∀x∀y(x ∈ X ∧ y ∈ Y ⇒ x ≤ z ≤ y).
.
The element z separates X and Y , in the sense of the order relation .≤.
As a first step, we prove that Dedekind completeness is equivalent to the property
of the least upper bound.
Theorem 12.53 Let .K be a totally ordered field. The following properties are
equivalent:
(i) .K is Dedekind complete;
(ii) every non-empty subset X of .K which is bounded from above possesses a least
upper bound in .K;
(iii) every non-empty subset Y of .K which is bounded from below possesses a least
upper bound in .K;
Proof Suppose (i) holds true, and let X be a non-empty subset of .K which is
bounded from above. The set .U of all upper bounds of X is non-empty, and trivially
.X ≤ U. By (i), there exists .z ∈ K such that .X ≤ {z} ≤ U. Since .z ∈ U, we
conclude that .z = min U, which implies that z is the least upper bound of X. Hence
(ii) is proved.
Suppose now that (ii) holds, and let X, Y be two non-empty subsets of .K such
that .X ≤ Y . Since every element of Y is an upper bound of X, the set .U of all upper
bound of X is non-empty. By (ii), there exists .z ∈ K such that .z = sup X = min U.
But .Y ⊂ U, hence .z = min Y and .{z} ≤ Y . On the other hand, .X ≤ {z} because
.z ∈ U, and thus (i) is proved.
The proof that (i) is equivalent to (iii) is similar, and it is left as an exercise.

Here comes the most technical part: we will be somehow sketchy, and refer to
[1] for the details. First of all, any ordered field contains an isomorphic copy of the
rational numbers.
Theorem 12.54 Let .K be an ordered field. Then there exists a subfield .Q(K) of .K
isomorphic to .Q.
Proof We temporarily denote by .1K the unit element (with respect to the multipli-
cation) of .K. First we embed .N into .K as follows:
.ϕ : n ∈ N → 1K + · · · + 1K ,
where the right-hand side contains n terms. It is easy to check that .ϕ : N → K is an

injective homomorphism which preserves the order.
Then we define .ψ : Z → K as follows:
⎧
⎪
⎪ if n > 0
⎨ϕ(n)
.ψ(n) = 0 if n = 0
⎪
⎪
⎩(−1 )ϕ(−n) if n < 0.
K
Finally, the required copy of the rational numbers is the range of the function
.e : Q → K defined as follows: for every .r = m/n ∈ Q such that .m ∈ Z, .∈ N,
.n > 0 and either .m = 0 or m and n are coprime, we let
e(r) = ψ(m) (ϕ(n))−1 .

.
It can be verified that e is an injective homomorphism which preserves the order

relation.

It follows from the previous result that we may systematically write .Q to denote
the copy of the standard rational numbers embedded into a given ordered field.2
With a minor modification of the proof of Theorem 3.14, we can see that the
Archimedean property of .K implies the density of .Q in .K.
Theorem 12.55 Let .K be a totally ordered field with the Archimedean property. Let
ϕ : K → K be an increasing function, in the sense that
.
∀x∀y(x ∈ K ∧ y ∈ K ∧ x < y ⇒ ϕ(x) < ϕ(y)).

.
If .ϕ(x) = x for every .x ∈ Q, then .ϕ(x) = x for every .x ∈ K.
2 More precisely, we identify .Q(K) and .e(Q).

12.9 About the Uniqueness of R 197
Proof Pick .x ∈ K and .ε > 0. Since .Q is dense in .K, there exist numbers .p ∈ Q
and .q ∈ Q such that
x − ε < p < x < q < x + ε.

.
The assumptions imply that
ϕ(p) = p < ϕ(x) < q = ϕ(q),

.
which in turn imply .x − ε < ϕ(x) < x + ε. Since .ε does not depend on x, we
conclude that .ϕ(x) = x, and the proof is complete.

We are ready to learn that Dedekind complete totally ordered fields are unique up
to monotonic isomorphisms. Although the meaning of the terms could be already
clear, we formalize some definitions.
Definition 12.78 Let .K1 = (K1 , +1 , ·1 , ≤1 ) and .K2 = (K2 , +2 , ·2 , ≤2 ) be ordered
fields. A function .ϕ : K1 → K2 is an increasing isomorphism if and only if .ϕ is
bijective and
∀x∀y(x ∈ K1 ∧ y ∈ K2 ∧ x ≤1 y ⇒ ϕ(x) ≤2 ϕ(y)).

.
Theorem 12.56 (Uniqueness up to Isomorphisms) For every Dedekind complete

totally ordered fields .K1 and .K2 there exists an increasing isomorphism .ϕ : K1 →
K2 .
Proof Let .K1 = (K1 , +1 , ·1 , ≤1 ) and .K2 = (K2 , +2 , ·2 , ≤2 ) be Dedekind
complete totally ordered fields. To ease notation, we will omit the subscripts 1 and
2 in the rest of the proof.
For every .x ∈ K1 we consider
Lx = {q ∈ Q | q ≤ x} .
.
It follows that .Lx is bounded from above in .Q. Understanding the identification
of the rational numbers in .K1 and .K2 , we deduce that .Lx is bounded from above
also in .K2 . The Dedekind completeness of .K2 allows us to define .ϕ : K1 → K2 by
declaring that .ϕ(x) is the least upper bound in .K2 of .Lx .
If .x < y in .K1 , then .Lx ⊂ Ly by density of .Q in .K1 , so that .ϕ is increasing.
Furthermore, if .x ∈ Q, then x is trivially the maximum of .Lx in .K2 , and then
.ϕ(x) = x.
By the same token, we can construct an increasing function .ψ : K2 → K1 such

that .ψ is the identity on .Q. If we consider the functions .ψ ◦ ϕ and .ϕ ◦ ψ, we may
apply Theorem 12.55 and conclude that these two compositions coincide with the
identities of .K1 and .K2 , respectively. In conclusion, .ϕ turns out to be an order-
preserving bijection.
To complete the proof, we must ensure that .ϕ also preserves the algebraic
operations., namely
(a) .ϕ(x + y) = ϕ(x) + ϕ(y);
(b) .ϕ(xy) = ϕ(x)ϕ(y)
for every .x ∈ K1 , .y ∈ K1 . We will exploit the density of the rational numbers: more
precisely, we claim that, given .x ∈ K1 , .y ∈ K2 and .ε > 0, there results
. ϕ(x) + ϕ(y) − ε < ϕ(x + y) < ϕ(x) + ϕ(y) + ε, . (12.1)

ϕ(x)ϕ(y) − ε < ϕ(xy) < ϕ(x)ϕ(y) + ε. (12.2)
Fix rational numbers .p1 , .q1 , .p2 and .q2 such that
ε ε ε ε
x−
. < p1 < x < p2 < x + , y− < q1 < y < q2 < x + .
2 2 2 2
It follows from the properties of .ϕ that
ε
ϕ(x) < ϕ p1 +
. ,
2
or
ε
ϕ(x) −
. < p1 .
2
Similarly it follows that
ε
ϕ(x) −
. < p1
2
ε
p2 < ϕ(x) +
2
ε
ϕ(y) − < q1
2
ε
q2 < ϕ(y) + .
2
Putting these inequalities together, we deduce that
ϕ(x) + ϕ(y) − ε < p1 + q1 ,

. p2 + q2 < ϕ(x) + ϕ(y) + ε.
But .p1 + q1 < x + y < p2 + q2 , hence .p1 + q1 < ϕ(x + y) < p2 + q2 , and the
proof of (12.1) follows.
We prove (12.2) under the additional assumptions that .0 ≤ x and .0 ≤ y in .K1 .
The general case follows by replacing x and y with .−x and .−y.
References 199
Let .r ∈ Q be such that .r > 0 and

ε ε
r < x,
. r < y, r< , r2 < .
2(x + y) 2
Fix rational numbers .p1 , .q1 , .p2 and .q2 such that
x − r < p1 < x < p2 < x + r,

. y − r < q1 < y < q2 < x + r.
It follows that .p1 q1 < xy < p2 q2 , .p2 − r < x < p1 + r, .q2 − r < y < q1 + r. The
properties of .ϕ imply that .p1 q1 < ϕ(xy) < p2 q2 , and .p2 − r < ϕ(x) < p1 + r,
.q2 − r < ϕ(y) < q1 + r. Then
(p2 − r)(q2 − r) < ϕ(x)ϕ(y) < (p1 + r)(q1 + r).

.
This implies that
ϕ(xy) − ϕ(x)ϕ(y) < p2 q2 − (p2 − r)(q1 + r) = r(p2 + q2 − r)

.
< r(x + y +r)

ϕ(x)ϕ(y) − ϕ(x)ϕ(y) > p1 q1 − (p1 + r)(q1 + r) = −r(p1 + q1 + r)
> −r(x + y + r).
We conclude that
. |ϕ(xy) − ϕ(x)ϕ(y)| < r(x + y) + r 2 < ε,
and the proof of (12.2) is complete.

At the end of this journey, we must observe that sentences like
.R is the unique totally ordered field with the least upper bound property
remain logically ambiguous. Theorem 12.56 ensures that we can speak of real
numbers only up to relabeling elements in an increasing manner. Since Analysts
need to make computations, what really matters to their eyes are models of .R. We
have seen Dedekind cuts and equivalence classes of rational Cauchy sequences, but
other modes are possible. In everyday life, most mathematicians simply use the
formal properties of the real numbers to carry on.
References
1. G. Devillanova, G. Molica Bisci, The fabuloud destiny of Richard Dedekind. Atti Accad.
Peloritana Pericolanti 99(S1), A18 (2021)
2. T.J. Jech, The Axiom of Choice (North-Holland Publishing Company, 1973)
3. J.L. Kelley, General Topology. Graduate Texts in Mathematics, No. 27 (Springer, New York,
1975). Reprint of the 1955 edition [Van Nostrand, Toronto, Ont.]
Chapter 13
Neighbors Again: Topological Spaces
Abstract The theory of Calculus can be successfully developed in the restricted

framework of the real line. We have seen that .R enjoys several properties that are
related to the concept of distance between two numbers. In this chapter we propose
a robust introduction to General Topology, which is the branch of mathematics that
deals with the idea of proximity.
I am happy to say that the disease of axiomatic topology has been almost totally cured. Right
now I don’t care a bit whether every .β-capsule of type . is also a T-spot of the second kind.
(Edwin Hewitt)
13.1 Topological Spaces
Definition 13.1 A topology is a family τ of sets which satisfies two conditions:

(τ1 ) the intersection of any two members of τ is a member of τ ;
(τ2 ) the union of the members of any subfamily of τ is a member of τ .
A subset U of X is open if and only if U ∈ τ . The set

X=
. τ= {U | U ∈ τ }
is usually called the space of the topology τ , and τ is a topology for X. The pair
(X, τ ) is a topological space.

Remark 13.1 Our conditions imply that X = τ is necessarily a member of τ ,
since τ is a subfamily of itself and every member of τ is a subset of X. Similarly,
the empty set ∅ is a member of τ , since it is the union of all the elements of the
empty subfamily of τ .
202 13 Neighbors Again: Topological Spaces
Important: About Language

We admit that our language here may sound peculiar. Most textbooks of General
Topology begin with a set X and define a topology τ as a family of subsets of X
which contains X, ∅, arbitrary unions and finite intersections of its members. The
two approaches are indeed equivalent, since X is always the union of all its open
subsets.
Example 13.1 For a given set X, the topology τ = {X, ∅} is called the indiscrete
topology. This topology possesses only two open sets.
Example 13.2 For a given set X, the topology τ = {U | U ⊂ X} is called the
discrete topology. Every set is open in this topology.
Exercise 13.1
1. Let X = {1, 2, 3} and τ = {∅, {1, 2, 3}, {1, 2}, {3}}. Show that τ is a topology on
X.
2. Let X = {1, 2, 3} and τ = {∅, {1, 2, 3}, {1, 2}, {1, 3}}. Show that τ is not a
topology on X.
3. Let X = {1, 2}. Write down all sets of subsets of X, and decide which are
topologies on X.
4. Let τ consist of ∅, R, and all sets of the form [a, +∞), a ∈ R. Show that τ is not
a topology on R.
Exercise 13.2 Show that τ = {S ⊂ R | R \ S is a finite set} ∪ {∅} is a topology on
R. This is the topology of finite complements.
Definition 13.2 Let τ1 and τ2 be two topologies for a set X. If τ1 ⊂ τ2 , we say that
τ1 is smaller than τ2 .
Remark 13.2 If τ1 is smaller than τ2 , many mathematicians say that τ1 is stronger
than τ2 , or equivalently that τ2 is weaker than τ1 . This is somehow counterintuitive,
since a smallness is seldom associated to strongness. For this reason we will try to
use the set-theoretic language of smaller and larger in the rest of this chapter.
Exercise 13.3 Let X be a set, and let τ be a topology for X. Show that any open
set for τ is also an open set for the discrete topology, and that any open set for
the indiscrete topology is also an open set for τ . Deduce that the indiscrete and the
discrete topology for a set X are respectively the largest and the smallest topology
for X.
Example 13.3 Let us go back to the real line. The usual topology of R is the family
of all subsets U of R such that for any x ∈ U there exists an open interval (a, b)
such that (a, b) ⊂ U .
13.1 Topological Spaces 203
Exercise 13.4 Prove that the usual topology of R is indeed a topology in the sense
of our definition. In particular, notice that any open interval is an open set for the
usual topology.
Exercise 13.5 We define a family σ of subsets of R as follows: W ∈ σ if either
W = ∅, or R \ W is a finite set. Prove that σ is a topology on R, and that this
topology is smaller that the usual topology of R.
Definition 13.3 A set U of a topological space (X, τ ) is a neighborhood of a point
x if and only if U contains an open set to which x belongs.
Remark 13.3 A neighborhood need not be an open set. The reader should be aware
that several mathematicians say that U is a neighborhood of x if and only if x ∈ U
and U is open. The two definitions are not equivalent, of course.
Example 13.4 In R with the usual topology, (0, 1) is clearly a neighborhood of
1/2. But also [0, 1) is a neighborhood of 1/2. On the contrary, (0, 1) is not a
neighborhood of 0, and [0, 1) is not a neighborhood of 0.
Exercise 13.6 If x ∈
/ U , can U be a neighborhood of x? If x ∈ U , is U necessarily
a neighborhood of x? Discuss.
The whole topology of a space can be described in terms of its neighborhoods.
The next result characterizes open sets.
Theorem 13.1 A set is open if and only if it contains a neighborhood of each of its
points.
Proof Let A a subset of a topological space. If A is open, then A ⊂ A, so that it
contains a neighborhood of each of its points. Conversely, suppose that A contains a
neighborhood of each of its points. The union U of all open subsets of A is an open
set, according to the definition of a topology. Then each x ∈ A belongs to some
open subset of A, and hence x ∈ U . Hence A = U , and A is open.

Definition 13.4 The neighborhood system of a point is the family of all neighbor-
hoods of that point.
Definition 13.5 (Closed Sets) In a topological space (X, τ ), a subset A is closed if
and only if X \ A is an open set.
Exercise 13.7 Recall that X \ (X \ A) = A. Deduce that the complement of a
closed set is an open set. In conclusion, a subset is open if and only if its relative
complement is closed.
Exercise 13.8 Luckily enough, any closed interval [a, b] is a closed set of R with
its usual topology. Prove this.
Exercise 13.9 What are the closed sets for the discrete and the indiscrete topologies
on a give set X?
Exercise 13.10 Using De Morgan’s laws from Set Theory, prove that the union
of any two closed subsets is a closed subset. Furthermore, the intersection of any
subfamily of closed sets is a closed set.
Example 13.5 The topology of upper semicontinuity on R is the topology whose
open sets are all the subsets of the form (−∞, a), where a ∈ R.
Example 13.6 The cofinite topology on a set X is the topology whose closed sets
are those subsets C such that either X = C or C is finite.
Definition 13.6 (Accumulation Points) A point x is an accumulation point of a
subset A of a topological space (X, τ ) if and only if every neighborhood of x
contains points of A different than x. The set A% of all the accumulation points
of A is called the derived set of A.1
Theorem 13.2 A subset is closed if and only if it contains its derived set.
Proof Indeed, if A is a subset, each neighborhood of a point x intersects A if and
only if either x ∈ A or x is an accumulation point of A.

Theorem 13.3 If A is a subset, A ∪ A% is a closed set.
Proof If x is neither a point of A nor an accumulation point of A, then there exists
an open neighborhood U of x which does not intersect A. But U is a neighborhood
of each of its points, and none of these are accumulation points of A. This proves
that the union A ∪ A% is the complement of an open set.

Definition 13.7 (Closure) The closure A of a subset A of a topological space
(X, τ ) is the intersection of all closed sets that contain A. Hence A is the smallest2
closed set containing A.
Theorem 13.4 The closure of any set if the union of the set and of its derived set.
Proof Every accumulation point of a set A in an accumulation point of each set
containing A, and is therefore a member of each closed set containing A. Hence A
contains both A and A% . Conversely, A∪A% is closed by Theorem 13.3, and therefore
it contains A.

Exercise 13.11 Let A = (−1, 3) ∪ [5, 6) be a subset of R with the usual topology.
What is A?
Definition 13.8 (Dense Subsets) A set A is dense in a topological space X if and
only if A = X.
Example 13.7 Q is dense in R, since for any real numbers x < y there exists a
rational number r such that x < r < y.
1 Sometimes denoted by DA or DA.

2 In the sense of inclusion of sets.
13.1 Topological Spaces 205
Definition 13.9 (Interior Points) A point x of a subset A of a topological space

is an interior point of A if and only if A is a neighborhood of x. The set A◦ of all
interior points of A is called the interior of A.
We collect several properties of the interior in a single statement.
Theorem 13.5 Let A be a subset of a topological space X. Then the interior of A
is open, and it is the largest open subset of A. A set A is open if and only if A = A◦ .
The set of all points that are not accumulation points of X \ A is A◦ . The closure of
X \ A is X \ A◦ .
Proof If a point x belongs to A◦ , then x belongs to some open subset U of A.
Every element of U is also an element of A◦ , and consequently A◦ contains a
neighborhood of each of its points, and it is therefore open. Now, if V in an open
subset of A and y ∈ V , then A is a neighborhood of y, so y ∈ A◦ . This shows that
A◦ contains each open subset of A and it is the largest open subset of A. If A is
open, it is clearly equal to the largest open subset of A. Hence A is open if and only
if A + A◦ .
If x is a point of A which is not an accumulation point of X \ A, then there exists
a neighborhood U of x which does not intersect X \ A and is therefore a subset of
A.Then A◦ is a neighborhood of x and x ∈ A◦ . Conversely, A◦ is a neighborhood
of each of its points and A◦ does not intersect X \ A. Hence there is no point A◦
which is an accumulation point of X \ A.
To conclude, since A◦ consists of the points of A which are not accumulation
points of X \A, the complement X \A◦ consists of the points which are either points
of X \ A or accumulation points of X \ A; In other words, A◦ is the complement of
X \ A.

Definition 13.10 (Boundary of a Subset) The boundary of a subset A of a
topological space X is the set of all points x such that every neighborhood of x
intersects both A and X \ A. The boundary of A is denoted by ∂A.
Exercise 13.12 Prove that ∂A is precisely the set of all points that are interior to
neither A nor X \ A.
Exercise 13.13 Prove that ∂A = ∂(X \ A).
Example 13.8 If X has the indiscrete topology and if A = X, A = ∅, then ∂A = X.
If X has the discrete topology, the boundary of every subset is empty.
Exercise 13.14 Let [a, b] a proper interval in R with the usual topology. Check that
∂[a, b] = {a, b}.
Exercise 13.15 Prove that the boundary of both Q and R \ Q is R. Hint: remember
that between any two distinct real numbers there is a rational number.
13.2 The Special Case of RN
In the first part of the book we have extensively described the real line .R and its
properties. We have seen that a distance function exists on it: .d(x, y) = |x − y|.
In terms of this distance we have defined open sets, accumulation points, and so
forth. Our aim now is to introduce finitely many additional dimensions to the space
.R = R , a process that is needed is almost every application of modern mathematics
1
to the real world.

Definition 13.11 Let .N ≥ 1 be an integer. The (vector) space .RN is the set of all
N-tuples .x = (x1 , . . . , xN ) of real numbers, together with the following operations:
a sum defined by
x + y = (x1 + y1 , . . . , xN + yN )
.
for each .x = (x1 , . . . , xN ) and .y = (y1 , . . . , yN ), and a multiplication by a scalar

defined by
λx = (λx1 , . . . , λxN )
.
for each .x = (x1 , . . . , xN ) and .λ ∈ R.

Remark 13.4 It is sometimes desirable to distinguish .RN as a set and .RN as a vector
space. For instance, the set .C of complex numbers is built on the set .R2 , but no
multiplication by a real scalar is necessary at the beginning. More correctly, our
vector space .RN should be denoted by .(RN , +, ·), but we will not be so pedantic in
the sequel.
Here is how we introduce a topology on .RN .
Definition 13.12 The usual distance on .RN is defined by
$
%
%N
%
.d(x, y) = & |xj − yj |2
j =1
for each .x = (x1 , . . . , xN ) and .y = (y1 , . . . , yN ). This distance is often called the
Euclidean distance. The open ball centered at some .x ∈ RN with radius .r > 0 is

B(x, r) = Br (x) = y ∈ RN d(x, y) < r ;
.
see Fig. 13.1. A subset A of .RN is open if for each .x ∈ A there exists .r > 0 such
that .Br (x) ⊂ A; see Fig. 13.2.
13.2 The Special Case of RN 207
Fig. 13.1 The ball .B(x, r)
Fig. 13.2 An open subset
Remark 13.5 From the topological viewpoint, .RN is a metric space. But the vector
space structure is often so important that mathematicians prefer to see .RN as a
normed vector space. For this reason the usual distance .d(x, y) is often denoted by
.*x − y*. This is indeed the standard notation for any distance induced by a norm
.* · *. We will come back to normed vector spaces in the chapter about differentiation
in abstract spaces.
Once we can recognize open sets in .RN , we can speak of closed sets, accumu-
lation points, closure, interior, and so on. If not explicitly stated, we will always
endow .RN with the usual distance and the usual topology induced by this distance.
13.3 Bases and Subbases
As a rule, in order to define a topology it is not strictly necessary to specify the

family of all open subsets. Exactly as a maximal set of linearly independent vectors
spans the whole vector space (via finite linear combinations), a maximal family of
subsets generates all open sets.
Definition 13.13 (Base) A family .B is a base for a topology .τ if and only if .B ⊂ τ
and for each point x of the space, and each neighborhood U of x, there exists a
member V of .B such that .x ∈ V ⊂ U .
Example 13.9 The collection of all open intervals is a base for the usual topology
of .R. Notice that this is essentially a restatement of the definition of open sets in .R.
Example 13.10 We consider the family of all intervals .[a, b) in .R. As a and b range
over .R, this family if the base for a topological space called the Sorgenfrey line.
Exercise 13.16 Prove that the topology of the Sorgenfrey
line is larger than the
usual topology of .R. Hint: just observe that .(a, b) = c>a [c, b).
Thus a base for a topology must be a family of open subsets. Can we characterize
those families of open subsets that are a base for the topology? The answer is
contained in the next theorem.
Theorem
13.6 A family .B of sets is a base for some topology for the set .X =
{B | B ∈ B} if and only if for every two elements U and V of .B and each point
.x ∈ U ∩ V , there exists .W ∈ B such that .x ∈ W ⊂ U ∩ V .
Proof Suppose first that .B is a base for some topology, U and V are members of
.B, and .x ∈ U ∩ V . Since .U ∩ V is open, some .W ∈ B contains x and is contained
in .U ∩ V . Conversely, let .τ be the union of all members of .B. Clearly any union of
members of .τ is a union of members of .B, hence it is a member of .τ . We only need
to show that the intersection of any two members U , V of .τ is a member of .τ . But if
.x ∈ U ∩ V , we can choose .U1 and .V1 in .B and select .W ∈ B such that .x ∈ U1 ⊂ U
and .x ∈ V1 ⊂ V . Hence there exists .W ∈ B such that .x ∈ W ⊂ U1 ∩ V1 ⊂ U ∩ V .

This shows that .U ∩ V is the union of members of .B, and .τ is a topology.

Definition 13.14 (Subbase) A family .S of sets is a subbase for a topology .τ if and
only if the family of all finite intersections of members of .S is a base for .τ .
Example 13.11 Since .(a, b) = (−∞, b) ∩ (a, +∞), the family of all half-lines is a
subbase for the usual topology of .R.
Subbases are a very general and flexible tool for topologizing sets.
Theorem 13.7 Let X be a non-empty set, and let . = {Aα | α ∈ A} be any
collection of subsets of X. There exists a unique topology .τ () on X which contains
.. This topology is the smallest topology with this property, and it can be described
as follows: it consists of .∅, X, all finite intersections of elements of ., and all unions
of these finite intersections.
13.4 Subspaces 209
We say that .τ () is the topology on X generated by the collection ..

Proof It is obvious that .2X (the collection of all subsets of X) is a topology on X
containing .. Therefore the intersection .τ () of all topologies on X which contain
. is non-empty. It is a simple exercise to check that .τ () is a topology on X
containing ., and it is both unique and the smallest one with such a property. To
prove the stated description of open sets, we notice that .τ () contains ., and thus
it must contain all the sets listed in the statement of the theorem. On the other hand,
since unions distribute over intersections, the sets listed actually form a topology
on X containing ., and therefore such a topology must contain .τ (). The proof is
complete.

13.4 Subspaces
Imagine we have a topological space .(X, τ ), and that Y is a subset of X. A natural

question at this point is whether Y can be endowed with some topology related to
.τ . We answer this question in the following definition.
Definition 13.15 (Relative Topology) Let .(X, τ ) be a topological space, and let
Y be a subset of X. The induced (or relative) topology on Y is the family of all
intersections of members of .τ with Y , and is denoted by .τ |Y . Explicitly,
τ |Y = {V ∩ Y | V ∈ τ } .
.
We say that .(Y, τ |Y ) is a subspace of .(X, τ ).

Example 13.12 Let .Y = [0, 1]. Each set of the form .(a, b) ∩ [0, 1] is an open set in
the induced topology on Y . Observe in particular that .[0, 1/2) is open in the relative
topology of Y , but not in the usual topology of .R.
This examples shows a general fact: if .A ⊂ Y ⊂ X, the set A may be open in
the relative topology of Y , but this does not imply that A is open as a subset of X.
Trivially, Y is always open — and closed — in the relative topology of Y .
Exercise 13.17 Describe the topology induced by the indiscrete and the discrete
topologies on a generic subset.
Theorem 13.8 Let .(X, τ ) be a topological space, and let Y be a subset of X. If A
is a subset of Y , then:
(i) the set A is closed in .τ |Y if and only if it is the intersection of Y and a set
closed in .τ .
(ii) A point .y ∈ Y is an accumulation point of A in .τ |Y if and only if it is an
accumulation point of A in .τ .
(iii) The closure of A in .τ |Y is the intersection of Y and the closure of A in .τ .
Proof The set A is closed in Y if and only if .Y \ A has the form .V ∩ X for some
V ∈ τ . This is true if and only if .A = (X \ V ) ∩ Y , and (i) is proved. Statement (ii)
.
follows directly from the definitions of relative topology and accumulation point.
Finally, the closure of A in .τ |Y is the union of A and the set of accumulation points
of A in .τ |Y . Then (iii) follows from (ii).

Exercise 13.18 Prove that if Y is open in X, each open set of .(Y, τ |Y ) is also open
in X. Is this true for closed sets, provided that Y is closed in X?
As a matter of fact, the information that a subset A is open in the relative topology
of Y tells very little about the properties of A as a subset of X.
Example 13.13 Suppose that .X = Y ∪ Z, and .A ⊂ X is such that .A ∩ Y is open
in Y , .A ∩ Z is open in Z. Although we might expect that A be open in X, this is
false. Just take .A = Y , .Z = X \ Y : clearly .Y ∩ Y and .Y ∩ Zare open in Y and Z
respectively, but Y may well be any subset of X.
13.5 Connected Spaces
Connected topological spaces are, from an analyst’s viewpoint, the straightforward

generalization of intervals in the real line. Recall that a (nonempty) set A in .R is an
interval if and only if, for each .x < y in A, each number z with .x < z < y is a
member of A. Roughly speaking, intervals cannot have holes.
Definition 13.16 (Connected Spaces) A topological space X is connected if and
only if the only subsets of X which are both open and closed are X and .∅.
Disconnected is the logical negation of connected. A set .Y ⊂ X is connected if
and only if Y is connected in the induced topology.
Exercise 13.19 Prove that a topological space X is connected if and only if X is
not the union of two disjoint open subsets. Equivalently, X is connected if and only
if X is not the union of two disjoint closed subsets. See Fig. 13.3.
This is by far the most elegant definition of connectedness. However an equivalent
definition can be provided in terms of separation.
Definition 13.17 Two subsets A and B are separated in a topological space X if
and only if .A ∩ B = ∅ and .A ∩ B = ∅.
Theorem 13.9 Let X be a topological space. A subspace Y of X is connected if
and only if Y is not the union of two non-empty separated sets.
13.5 Connected Spaces 211
Fig. 13.3 Connected vs. disconnected
Proof Suppose that Y is disconnected, so that .Y = A ∪ B, and A, B are non-empty

open subsets of Y such that .A ∩ B = ∅. Clearly A and B are separated in X, since
A ∩ Cl(B, X) = (A ∩ Y ) ∩ Cl(B, X)
.
= H ∩ (Y ∩ Cl(B, X))
= A ∩ Cl(B, Y ) = ∅,
and similarly for .Cl(A, X) ∩ B.3 Conversely, if A and B are separated in X and
.Y = A ∪ B, then
. Cl(A, Y ) = Y ∩ Cl(A, X) = (A ∪ B) ∩ Cl(A, X)

= (A ∩ Cl(A, X)) ∪ (B ∩ Cl(A, X))
= A,
so that A is closed in Y . A similar computation proves that B is closed in Y .

The topology of the real line is intimately related to the order properties, and this
characterizes connected subsets.
Theorem 13.10 A subset A of .R (with the usual topology) is connected if and only
if it has the following property: if .x ∈ A, .y ∈ A and .x < z < y, then .z ∈ A.
3 We have denoted .Cl(A, X) the closure of A in X, since we need to distinguish the closure in X
and the closure in Y .

Proof Suppose there exist .x ∈ A, .y ∈ A and .z ∈ (x, y) with .z ∈

/ A. Then .A =
Uz ∪ Vz with
Uz = A ∩ (−∞, z),
. Vz = A ∩ (z, +∞).
Clearly .Uz and .Vz are nonempty. Since .Uz ⊂ (−∞, z) and .Vz ⊂ (z, +∞), they
are separated. Hence A is the union of two nonempty separated sets, and is not
connected.
Conversely, suppose that A is not connected, and has the form .A = U ∪ V for
some nonempty separated sets U and V . Let .x ∈ U , .y ∈ V , assuming .x < y for the
sake of definiteness. We define .z = sup (U ∩ [x, y]). As a supremum, we know that
z is in the closure of U , so that .z ∈
/ V because U and V are separated. In particular,
.x ≤ z < y. If .z ∈ / U , it follows that .x < z < y and .z ∈
/ A. If .z ∈ U , then .z ∈
/ V,
so that there exists .z1 such that .z < z1 < y and .z1 ∈ / V . Again .x < z1 < y and
.z1 ∈/ A.

Important: Beware!
The last result shows that the only connected subsets of the real line (with the usual
topology) are the intervals, of any kind. This is no longer true if we replace .R with
.Q with the induced topology. Indeed, for any irrational number a, the two sets .{x ∈
Q | x < a} and .{x ∈ Q | x > a} are separated in this topology.
Remark 13.6 In a geometric language, we have proved the connected subsets of .R

coincide with the convex subsets of .R. Indeed .A ⊂ R is convex if and only if
λx + (1 − λ)y ∈ A
.
for every .x ∈ A, .y ∈ A, .λ ∈ [0, 1].

Exercise 13.20 Prove the last statement of the warning above.
Connectedness is stable under closure and unions.
Theorem 13.11
(a) The closure of a connected set is connected.
(b) If .A is a family of connected
sets of a topological space, and if no two members
of .A are separated, then . A is connected.
Proof Suppose that Y is a connected subset of a topological space X, and that .Y =

A ∪ B for some sets A and B that are both open and closed in .Y . Then each of .A ∩ Y
and .B ∩ Y is open and closed in Y , and since Y is connected, one of these two sets
must be empty. Suppose .B ∩ Y is empty. Then Y is a subset of A and .Y is a subset
of A because A is closed in .Y . Hence .B = ∅, and it follows that .Y is connected.
This proves (a).
13.5 Connected Spaces 213
To prove (b), we denote by C the union of the members of .A. Suppose that D
is both open and closed in C. For each .A ∈ A, we have that .A ∩ D is open and
closed in A, and since A is connected we conclude that either .A ⊂ D or .A ⊂ C \ D.
We claim that either each member of .A is contained in D and .C \ D = ∅, or it
is contained in .C \ D and .D = ∅. Indeed, if A and B are members of .A, it is
impossible that .A ⊂ D and .B ⊂ C \ D, for in this case A and B would be separated
as subsets of the separated sets D and .C \ D. This contradicts the assumption, and
proves the claim and the proof.

Definition 13.18 (Connected Components) A (connected) component of a topo-
logical space is a maximal connected subset. More precisely, it is a connected subset
which is properly contained in no other connected subset.
Example 13.14 If a topological space is connected, then it is the only connected
component. Indeed, if A is a connected subset, then A is contained in the connected
topological space, and A cannot be maximal.
Example 13.15 In a discrete topological space, each connected component is a
singleton.
Theorem 13.12 Each connected subset of a topological space is contained in a
connected component. Each component is closed. If A and B are distinct connected
components of a space, then A and B are separated.
Proof Let A be a nonempty connected subsets of a space X, and let C be the union
of all the connected sets containing A. As a consequence of Theorem 13.11, C is
connected. If D contains C and D is connected, then .D ⊂ C, so that .D = C.
Hence C is a maximal connected subset, i.e. a connected component of the space.
Each component C is connected by definition, so that its closure is connected
by Theorem 13.11. Therefore .C = C, and C is closed. Finally, if A and B are
disjoint components but they are not separated, then .A ∪ B is a connected subset, a
contradiction to the maximality of A and B.

Definition 13.19 (Totally Disconnected Spaces) A topological space X is totally
disconnected if and only if the connected component containing any point .x ∈ X
coincides with .{x}.
The intuition behind connectedness is often expressed by saying that any two
points of the space can be connected by a path. Although it can be proved the this is
not an equivalent definition, it is nonetheless an interesting definition.
Definition 13.20 (Arc-Wise Connected Spaces) A topological space X is arc-
wise connected if and only if for each points .x0 and .x1 in X there exists a continuous
map .α : [0, 1] → X such that .α(0) = x0 and .α(1) = x1 . See Fig. 13.4.
Fig. 13.4 A path joining two points
Theorem 13.13 An arc-wise connected space is always connected.

Proof Let X be a connected space, and suppose that .X = A ∪ B for some open sets
A and B. We will prove that .A∩B = ∅. Pick any two points .x0 ∈ A and .x1 ∈ B, and
consider a path .α that joins them as in the previous definition. The subsets .α −1 (A)
and .α −1 (B) are open in .[0, 1], non-empty, and their union is .[0, 1]. Since .[0, 1] is
connected, we must have .α −1 (A) ∩ α −1 (B) = ∅. But then .A ∩ B = ∅.

Example 13.16 The space .RN is arc-wise connected: if .x0 and .x1 are two points,
the path .α(t) = tx0 + (1 − t)x1 is a continuous map that joins them. Using polar
coordinates, it is easy to check that the unit sphere .SN−1 = {x ∈ RN | *x* = 1} is
arc-wise connected (by a geodesic arc).
13.6 Nets and Convergence
So far we have described topologies in terms of their open sets. This is by far the
most common approach to General Topology. But there are other viewpoints which
can be even more useful to mathematical analysts.
If we think of our previous chapters, we immediately see that elementary real
analysis in one variable leans on a single idea: that of limits. In any reasonable
generalization, limits should therefore play a crucial role. And indeed limits can be
defined as soon as a topology exists on a set. However, in this section we want to
define a broader definition that can include sequences, functions, Riemann sums,
and much more.
Recall that a sequence is just a function defined on the complement of a finite
subset N of .N. With an abuse of notation, and since our interest is towards the
theory of convergence, we will assume that sequences are functions defined on the
whole set .N of natural numbers. Indeed, any two sequences which differ for finitely
many terms have the same character.
The value of a sequence S at .n ∈ N is denoted either by the functional symbol
.S(n) of by the traditional symbol .Sn . We say that a sequence S is in a set A if and
only if .Sn ∈ A for each n.

13.6 Nets and Convergence 215
Definition 13.21 A sequence S is eventually in a set A if and only if there exists an

integer m such that .Sn ∈ A for each .n ≥ m.
Example 13.17 A sequence S of real numbers converges to a limit .L ∈ R if and
only if .Sn is eventually in each neighborhood of L.
The definition of subsequence is usually hard to grasp at first sight. Since we
are going to introduce a more general definition, it is advisable to look back to
subsequences. Suppose a sequence S is given, and that we want to select an integer
.Ni for each integer i, such that .SNi = S(N(i)) converges. A good candidate could
be .Ni = 1 for each i, since .S(N(i)) = S(1) for each i, and convergence would be
ensured. But this is not an interesting candidate, and some additional condition on
.i → N(i) should be imposed.
The usual condition, which we actually used in a previous chapter, is that .i < j
imply .N(i) < N(j ), i.e. a strict monotonicity condition. If this condition is satisfied,
we then say that .{SNi }i is a subsequence of S. But is this really the good condition?
Well, not really. What we actually need is that .Ni becomes large as i becomes large.
Definition 13.22 (Generalized Subsequences) T is a subsequence of S if and only
if there exists a map .N : N → N such that .T = S ◦ N, and for each integer m there
exists an integer n such that .i ≥ n implies .Ni ≥ m.
The previous discussion introduces the following problem: we want to construct
generalized sequences on any topological space in such a way that the convergence
of such generalized sequences characterize the topology of the space.
Definition 13.23 A binary relation .≥ directs a set D if and only if .D = ∅ and
(a) if m, n and p are elements of D such that .m ≥ n and .n ≥ p, then .m ≥ p;
(b) if .m ∈ D then .m ≥ m;
(c) if m and n are elements of D, then there exists .p ∈ D such that .p ≥ m and
.p ≥ n.
Example 13.18 The real numbers and the natural numbers are directed by their
usual order .≥.
Example 13.19 The family of all neighborhoods of a point in a topological space
is directed by reverse inclusion: .U ≥ V if and only if .U ⊂ V . Condition
(c) is a consequence of the fact that the intersection of two neighborhoods is a
neighborhood. This is a particularly important example, since it will join the usual
definition of convergence is a topological space with the definition of convergence
of a net.
Example 13.20 The family of all finite partitions of the closed interval .[a, b] is a
directed set, when ordered by the relation .P2 ≥ P1 if and only if .P2 is a refinement
of .P1 .
Exercise 13.21 Prove that the family of all finite subsets of a set is directed by
direct inclusion: .A ≥ B if and only if .A ⊃ B.
Definition 13.24 (Nets) A directed set is a pair .(D, ≥) such that .≥ directs D. A
net is a pair .(S, ≥) such that S is a function and .≥ directs the domain of S. If S is a
function whose domain contains D and D is directed by .≥, then .{Sn , n ∈ D, ≥} is
the net .(S|D , ≥). A net .{Sn , n ∈ D, ≥} is in a set A if and only if .Sn ∈ A for each
.n ∈ D; it is eventually in A if and only if there exists .m ∈ D such that .Sn ∈ A for
each .n ≥ m. The net is frequently in A if and only if for each .m ∈ D there exists
.n ∈ D such that .n ≥ m and .Sn ∈ A.
Important: Notation
The full notation .{Sn , n ∈ D, ≥} is too cumbersome to be currently used. We will
write .{Sn , n ∈ D} when the domain of S plays an explicit role and no confusion can
arise about the direction .≥ in D.
We have formally recovered most set-theoretic definitions of sequences in a

general topological setting. We are ready for the most important one.
Definition 13.25 (Convergent Nets) A net .(S, ≥) in a topological space .(X, τ )
converges to a point .s ∈ X if and only if for each neighborhood U of s, S is
eventually in U .
Important: Nets Are not Sequences!

When working with sequences, convergence is often stated in the following way: a
sequence .{Sn | n ∈ N} converges to x if and only if for each neighborhood V of x,
the sequence .{Sn | n ∈ N} is in V for all but finitely many indices n. This equivalent
definition is based on the fact that the complement of each tail .{n ∈ N | n ≥ n0 } is
a finite set. This is in general false, if .N is replaced by a directed set D.
Example 13.21 Let .{an }n be a sequence in .R. The sequence converges to a (finite)
limit L if and only if the associated net .(a, ≥) converges to L in the sense of
Definition 13.25. Indeed, both statement mean: for each neighborhood U of L there
exists an integer N such that .n ≥ N implies .an ∈ U .
Example 13.22 As another interesting construction based on nets, we consider
summability. Let A be a set, and .f : A → R be a real-valued function. We direct
finite subsets of A by .⊃, and for each finite set .F ⊂ A we define

S(F, f ) =
. {f (a) | a ∈ F } .
In this way we construct a net .(S(F, f ), F, ⊃). We say that f is summable on A if

this net converges to a real number I . This appears to be an unordered sum, since
the elements of finite sets need not be arranged in any increasing order. But let us
think of the familiar case .A = N, so that f is actually a sequence of real numbers.

Finite subsets of .N need not be finite intervals, i.e. sets of the form .{1, 2, 3, . . . , N}
for some .N ∈ N. When we introduced the idea of numerical series, we considered
the convergence of partial sums, which are constructed by summing the values .f (n)
as n ranges from 1 to some N. Here we are summing the values .f (n) as n ranges
over finite subsets of the natural numbers. Is there any difference?
The two approaches lead to the same idea of convergence. Indeed, we know that
if two numerical sequences are identical from a certain index onwards, then they
have the same character. The same holds for numerical series, which are nothing but
special numerical sequences. Using this simple remark, saying that for each .ε > 0
there exists a finite subset .Fε of .N with the property that .|S(f, f ) − I | < ε holds for
each finite subset F of .N such that .F ⊃ Fε is equivalent to saying that for each .ε > 0
there exists a finite interval .Fε = {1, 2, . . . , Nε } such that .|S(F, f ) − I | < ε holds
for each finite interval F of .N such that .F = {1, 2, . . . , Nε , . . . , n} with .n > Nε .
This in turn means that

n
. lim f (k) = I.
n→+∞
k=1
Example 13.23 We already know that the set .P of all finite partitions of an interval
.[a, b] is a directed set. If .f : [a, b] → R is a given function, we can define a net
.{L(f, P ), P ∈ P} by letting .L(f, P ) denote the lower Riemann sum of f with
respect to the partition P . The same can be done for the upper Riemann sum, of
course. Convergence of both nets to a common limit .c ∈ R is then equivalent to the
!b
integrability of f on .[a, b], and . a f dx = c.
Exercise 13.22 Let .f : [0, 1] → R be a bounded function. We define the set D
of all ordered pairs .(P , ξ ) such that .P = {a0 , . . . , an } is a partition of .[0, 1] and
.ξ = {ξ1 , . . . , ξn } is a finite set of points such that .ak−1 ≤ ξk ≤ ak for .k = 1, . . . , n.
If .(P , ξ ) and .(Q, η) are elements of D, we define .(P , ξ ) ≥ (Q, η) if and only if P
is a refinement of Q. It is easy to prove that .≥ directs D. We define the Riemann net
.S : D → R such that

n
.S(P , ξ ) = f (ξk )(ak − ak−1 ).
k=1
Finally, let

n
U (P ) =
. sup {f (x) | ak−1 ≤ x ≤ ak }
k=1

n
L(P ) = inf {f (x) | ak−1 ≤ x ≤ ak }
k=1
μ(P ) = sup {ak − ak−1 | k = 1, . . . , n}

δ(P ) = inf {ak − ak−1 | k = 1, . . . , n} .
1. For every .(P , ξ ) ∈ D, prove that .L(P ) ≤ S(P , ξ ) ≤ U (P ). Prove also that, for
every .ε > 0 and for every partition P , there exist .ξ and .η such that .(P , ξ ) ∈ D,
.(P , η) ∈ D, .S(P ξ ) > U (P ) − ε and .S(P , ξ ) < L(P ) + ε.
2. If P , Q are partitions of .[0, 1] and if P is a refinement of Q, prove that .μ(P ) ≤

μ(Q). Deduce that if the Riemann integral of f exists and equals I , then the
Riemann net S converges to I .
3. Prove that the Riemann integral of f exists if and only if for every .ε > 0 there
exists .δ > 0 such that whenever P is a partition of .[0, 1] with .μ(P ) < δ, then
.U (P ) − L(P ) < ε.
4. If P , Q are partitions of .[0, 1] such that .μ(P ) < (1/2)δ(Q), prove that .U (P ) −
L(P ) < 2(U (Q) − L(Q)). Hint: observe that every interval of P is contained in
at most two intervals of Q.
5. Prove that if the Riemann net S converges to I , then the integral of f exists and
equals I . Hint: it follows from 3 and 4 that it is sufficient to construct a partition
Q such that .U (Q) − L(Q) is small.
Exercise 13.23 Let X be a discrete space; prove that a net S converges to x if and
only if it is eventually equal to x. If X is an indiscrete space, prove that any net
converges to any point of X.
Theorem 13.14 Let X be a topological space.
(a) A point x is an accumulation point of a subset A of X if and only if there exists
a net in .A \ {x} which converges to s.
(b) A point x belongs to the closure of a subset A of X if and only if there is a net
in A converging to x.
Proof Suppose x is an accumulation point of A. For every neighborhood U of x
there exists a point .SU of A which belongs to .U \ {x}. Recall that the family .U of
all neighborhoods of x is directed by .⊂. Now, if U and V are neighborhoods of x
such that .V ⊂ U , then .SV ∈ V ⊂ U . The net .{SU , U ∈ U, ⊂} therefore converges
to x. Conversely, if a net in .A \ {x} converges to x, then this net has values in every
neighborhood of x, and .A \ {x} intersects each neighborhood of x.
To prove (b), we remember that the closure of a subset consists of A together
with the set of all accumulation points of A. If x is an accumulation point of A, by
the preceding discussion there exists a net in A converging to x. Furthermore, if x is
a point of A, the net which is constantly equal to x converges to x. In any case, each
point of the closure of A has a net in A which converges to it. On the other hand,
if a net in A converges to x, then every neighborhood of x intersects A, and x thus
belongs to .A.

Fig. 13.5 Two points .x1 and

.x2separated by disjoint
neighborhoods
Remark 13.7 The previous proof is a typical example of the use of the Axiom of
Choice. Although this is usually taken as granted, when we define the net .U → SU
we actually choose a point .SU from the neighborhood U , and this choice can only
be motivated by the Axiom of Choice.
We have seen that in an indiscrete space, any net converges to any point. This is
in striking contrast to the familiar case of .RN , in which a sequence can have at most
one limit. Honestly, it is hard to imagine a sequence converging to any point, since
our mind immediately draws picture in the physical space .R3 .
Definition 13.26 A topological space is a Hausdorff space, or a .T2 -space, if and
only if whenever x and y are distinct points of X, there exist disjoint neighborhoods
of x and y. See Fig. 13.5.
In our intuition, this separation property is usually taken for granted. As we said
before, we always think in .R3 , which is a metric space. This makes the difference.
Example 13.24 For any integer N, the space .RN (with the usual topology) is a
Hausdorff space. Indeed, let x and y be two distinct point of .RN . We call .δ = *x −
y* > 0 their distance. The sets .B(x, δ/3) and .B(y, δ/3) are disjoint neighborhoods
of x and y.
Theorem 13.15 A topological space is a Hausdorff space if and only if each net in
the space converges to at most one point.
Proof One half of the proof is standard. Indeed, assume that X is a Hausdorff space
and x, y are distinct points of X. By definition, we can pick disjoint neighborhoods
U and V of x and y. Since a net cannot be eventually in each of the disjoint sets U
and V , it is clear that a converging net in X cannot have distinct limits. Conversely,
we proceed by contradiction, assuming that X is not a Hausdorff space and x and y
are distinct points such that every neighborhood of x intersects every neighborhood
of y. We now construct a net in X that converges to both x and y. Let .Ux be the
family of all neighborhoods of x, and .Uy be the corresponding family of y. Both
families are directed by inclusion. The cartesian product .Ux × Uy is ordered by
agreeing that .(T , U ) ≥ (V , W ) if and only if .T ⊂ V and .U ⊂ W . The cartesian

product is directed by .≥, as is easy to check. By assumption, for each .(T , U ) of
.Ux × Uy , the intersection .T ∩ U is non-empty; hence we select a point .S(T ,U )
in this intersection. If .(V , W ) ≥ (T , U ), then .S(V ,W ) ∈ V ∩ W ⊂ T ∩ U , and

consequently the net .{ST ,U ) , (T , U ) ∈ Ux × Uy , ≥} converges to both x and y.

Definition 13.27 Let .(X, τ ) be a Hausdorff space. If a net .{Sn , n ∈ D, ≥}
converges to a point x, we write
τ − lim{Sn , n ∈ D, ≥} = x.
.
When no confusion may arise, we simply write .limn Sn = x.

Remark 13.8 Following [5], we reserve the symbol of limit to net in a Hausdorff
space. It is tempting to denote by .τ − lim{Sn , n ∈ D, ≥} the set of all limits of the
net. This is possible, but we must be coherent and systematically write .x ∈ τ −
lim{Sn , n ∈ D, ≥}. This piece of notation would be too different than the familiar
.x = limn sn that we learn in Calculus.
Exercise 13.24 Prove that the trick introduced in the previous proof is a general
one. Precisely, if .(D, ≥D ) and .(E, ≥E ) are directed sets, prove that .D×E is directed
by .(d, e) ≥D×E (f, g) if and only if .d ≥E f and .e ≥E g. This is the product
directed set.
Example 13.25 Cartesian products ' can be directed even in a much more general
situation. The cartesian product . {Da | a ∈ A} of a family of sets is the set of all
functions d on A such that .d(a) ' ∈ Da for each .a ∈ A. Suppose now each .Da is
directed by .≥a . The product . {Da | a ∈ A} is directed by .d ≥ e if and only if
.d(a) ≥ e(a) for each .a ∈ A. This is a very natural order relation on the product set.
We now verify that we have constructed a directed set. Let d and e be two members
of the product set. For each .a ∈ A there exists a member .f (a) of .Da such that
.f (a) ≥a d(a) and .f (a) ≥a e(a). Consequently the function f whose value at a is
.f (a) follows both d and e in the product order.

'
Exercise 13.25 Consider the product . {D | a ∈ A} of all functions defined on
A whose values are in D. Prove that the product order coincides with the familiar
order on functions: for f , .g : A → D, .f ≥ g if and only if .f (a) ≥ g(a) for every
.a ∈ A.
Let us go back to the problem of subsequences.

Definition 13.28 A net .{Tm , m ∈ E} is a subnet of a net .{Sn , n ∈ D} if and only if
there exists a function .N : E → D such that
1. .T = S ◦ N, or equivalently .Ti = SNi , for each .i ∈ E;
2. for each .m ∈ D there exists .n ∈ E with the property that, if .p ≥ n, then
.N(p) ≥ m.
Remark 13.9 The word “subnet”should not be taken too seriously. While a subse-
quence is always indexed by a selection of natural numbers, a subnet may be indexed
by a set which has nothing to do with the set of indices of the original net. This is
an important fact to remember, if we want to avoid silly mistakes.
We have already seen that the usual approach to subsequences is to require that
the function N be strictly increasing. This of course guarantees that the second
condition is satisfied.
We now try to relate cluster points and convergent subnets in general topological
spaces.
Lemma 13.1 Let S be a net and .A be a family of sets such that S is frequently in
each member of .A, and such that the intersection of two members of .A contains a
member of .A. Then there exists a subnet of S that is eventually in each member of
.A.
Proof By assumption, .A is directed by .⊂. Let .{Sn , n ∈ D} be a net which is

frequently in each member of .A, and let E be the set of all pairs .(m, A) such that
.m ∈ D, .A ∈ A, and .Sm ∈ A. Using the assumption that the intersection of any
two members of .A contains a member of A, it is easy to check that E is directed

by the product order of .D × A. For every .(m, A) of E, let .N(m, A) = m. The map
.N : E → D is increasing; since the net S is frequently in each member of .A, .S ◦ N
is a subnet of S. To complete the proof, suppose .A ∈ A and .m ∈ D is such that

.Sm ∈ A. If .(n, B) ≥ (m, A), then .S ◦ N(n, B) = Sn ∈ B ⊂ A, and the net .S ◦ N is
eventually in A.

Definition 13.29 A point x of the space is a cluster point of a net S if and only if S
is frequently in every neighborhood of x.
Theorem 13.16 A point x is a topological space is a cluster point of a net S if and
only if some subnet of S converges to x.
Proof Let x be a cluster point of a net S, and let .U be the family of all
neighborhoods of x. The intersection of two members of .U is again a member of .U.
The preceding Lemma applies, and there is a subset of S which is eventually in each
neighborhood of x. This means that this subnet converges to x. On the contrary, if
x is not a cluster point of S, then there exists a neighborhood U of x such that S
is not frequently in U , and therefore S is eventually in the complement of U . Then
each subnet of S is eventually in the complement of U and hence cannot converge
to x.

We conclude this section with some remarks about the limit of a function in
general spaces. We have seen that nets are a genuine generalization of sequences, in
the sense that any sequence is a net since .N is directed by the usual order. But what
about limits of functions?
Example 13.26 Suppose that Y is a topological space, .(X, d) is a metric space, A
is a subset of X, .x0 an accumulation point of A, and .f : A → Y a function. The set
.A \ {x0 } is directed as follows: .x ≥ y if and only if .d(x, x0 ) ≤ d(y, x0 ). Since .x0
is an accumulation point of A, it is easy to check (exercise!) that .≥ directs .A \ {x0 }.

How do we interpret the convergence of the net .{f (x), x ∈ A \ {x0 }, ≥}? This net
converges to a point .y0 ∈ Y if and only if for each neighborhood V of .y0 there
exists .xV ∈ A \ {x0 } such that .x ∈ A \ {x0 } and .x ≥ xV implies .f (x) ∈ V . This
in turn means that .x ∈ A, .0 < d(x, x0 ) ≤ d(xV , x0 ) imply .f (x) ∈ V . Setting
.δ = d(xV , x0 ) > 0 we recover the usual definition of .limx→x0 f (x) = y0 .
The last example shows that we can use nets for defining limits of functions if
the domain of the function is a metric space. In particular we recover the theory
of limits for functions of a real variable. However, the situation is less clear if the
domain of the function is merely a topological space. The following is the ad hoc
definition that we can find in many textbooks.
Definition 13.30 Suppose that Y is a topological space, X is a topological space, A
is a subset of X, .x0 an accumulation point of A, and .f : A → Y a function. We say
that .y0 ∈ Y is a limit of f as x tends to .x0 if and only if for every neighborhood V
of .y0 there exists a neighborhood U of .x0 such that .f (A ∩ U \ {x0 }) ⊂ V .
If Y is a Hausdorff space, it is easy to prove that the limit, if it exists, must be unique.
Hence we can write .y0 = limx→x0 f (x), and we are happy.
The question now is: can we interpret this definition of limit in the setting of
nets? Well, in order that .{f, A \ {x0 }} be a net, the domain .A \ {x0 } must be directed
in such a way that .x ≥ y means “x is closer to .x0 than y.” In general there is no
canonical way to ensure that such an order relation exists: topologies do not always
allow to measure the distance between points. Should we give up? No, although the
answer is probably not so elegant as we may hope.
Definition 13.31 (Limits Through Nets) Suppose that Y is a topological space,
X is a topological space, A is a subset of X, .x0 an accumulation point of A, and
.f : A → Y a function. We say that .y0 ∈ Y is a limit of f as x tends to .x0 if and only
if for every net .{xα , α ∈ D} in .A \ {x0 } converging to .x0 , the net .{f (xα ), α ∈ D}
converges to .y0 in Y .
Notice that this is the straightforward generalization of Definition 7.1.
Exercise 13.26 Prove that Definitions 13.30 and 13.31 are indeed equivalent. Hint:
in one direction you will need to construct a suitable net in .A \ {x0 } by using the
family of neighborhoods of .x0 directed by reverse inclusion. Compare with Theorem
7.1.
Remark 13.10 Let A, B be disjoint subsets of a topological space X, and let .Y =
A ∩ B. The set A (resp. B) is closed in Y if and only if for each net .{Sn , n ∈ D} in A
(resp. in B) converging to a limit x in the relative topology of Y there results .x ∈ A
(resp. .x ∈ B). Now, .Sn → x if and only if .Sn eventually lies in every open (with
respect to the topology of Y ) neighborhood of x, i.e. if and only if for every open
(with respect to the topology of X) neighborhood U of x there results .Sn ∈ U ∩ Y
13.7 Continuous Maps and Homeomorphisms 223
eventually. Since .U ∩ Y = (U ∩ A) ∪ (U ∩ B) and .A ∩ B = ∅, this means that x is an

accumulation point of either A or B in the topology of X. We conclude that .x ∈ A
(resp. .x ∈ B) if and only if .A ∩ B = ∅ (resp. .A ∩ B = ∅). This is al alternative
proof of Theorem 13.9 with the language of nets.
13.7 Continuous Maps and Homeomorphisms
Topology is often described as the branch of mathematics which studies invariant

properties under the action of continuous functions. The first task, therefore, is to
define continuity.
Definition 13.32 A function f of a topological space .(X, τ ) to a topological space
(Y, σ ) is continuous if and only if .f −1 (V ) ∈ τ for every .V ∈ σ . In words, the
.
preimage of every open set of Y must be an open set of X.

Topologists often prefer the global approach to the local one. Analysts often do
the opposite: in Calculus we all learn that continuity is a definition that applies to a
single point.
Definition 13.33 A function f of a topological space X to a topological space Y is
continuous at the point .x ∈ X if and only if the preimage of any neighborhood of
.f (x) is a neighborhood of x.
Exercise 13.27 Prove that a function is continuous if and only if it is continuous at

every point of the domain.
Theorem 13.17
(a) If f is continuous from X to Y , and if g is continuous from Y to Z, then .g ◦ f
is continuous from X to Z.
(b) If f is continuous from X to Y , and if A is a subset of X, then .f|A is continuous
from A (with the relative topology) to Y .
Proof Let V be an open set in Z; since .(g ◦ f )−1 (V ) = f −1 (g −1 (V )), using first
the continuity of g and then the continuity of f we see that .(g ◦ f )−1 (V ) is open in
X. This proves (a).
−1 −1
If V is open in Y , then .f|A (V ) = A ∩ f −1 (V ), and therefore .f|A (V ) is an open
set in the relative topology of A.

We summarize eight equivalent definitions of continuity. Which one to choose is
clearly a matter of taste
Theorem 13.18 If X, Y are topological spaces and .f : X → Y is a function, the
following statements are equivalent:
1. f is continuous;
2. the preimage of each closed set is closed;
3. the preimage of each member of a subbase for the topology of Y is open;
4. for each .x ∈ X the preimage of every neighborhood of .f (x) is a neighborhood

of x;
5. for each .x ∈ X and each neighborhood V of .f (x), there exists a neighborhood
U of x such that .f (U ) ⊂ V ;
6. for each net .{Sn , n ∈ D} in X which converges to a point x, the composition
.{f ◦ Sn , n ∈ D} converges to .f (x);
7. for each subset A of X, there results .f (A) ⊂ f (A);

8. for each subset B of Y , there results .f −1 (B) ⊂ f −1 (B).
Proof 1 is equivalent to 2: a trivial consequence of the identity .f −1 (Y \ B) =
X \ f −1 (B) for every .B ⊂ Y .
1 is equivalent to 3: if f is continuous, condition 3 is satisfied because any
member of a subbase is an open set. Conversely, since every open set V in Y is
the union of finite intersections of subbase members, .f −1 (Y ) is the union of finite
intersections of the preimages of subbase members; if these are open, then .f −1 (V )
is open.
1 implies 4: for each .x ∈ X and each neighborhood V of .f (x), V contains an
open neighborhood W of .f (x) and .f −1 (W ) is an open neighborhood of x which is
a subset of .f −1 (V ). Therefore .f −1 (V ) is a neighborhood of x.
4 implies 5: if V is a neighborhood of .f (x), then .f −1 (V ) is a neighborhood of
x such that .f (f −1 (V )) ⊂ V .
5 implies 6: consider a net .{Sn , n ∈ D} in X that converges to x. If V is a
neighborhood of .f (x), there exists a neighborhood U of x such that .f (U ) ⊂ V ,
and since S is eventually in U , .f ◦ S is eventually in V .
6 implies 7: Let A be a subset of X and x a point of .A. There exists a net S in A
which converges to x, and .f ◦ S converges to .f (x). Hence .f (x) ∈ f (A).
7 implies 8: if .A = f −1 (B), then .f (A) ⊂ f (A) ⊂ B. Hence .A ⊂ f −1 (B).
Equivalently, .f −1 (B) ⊂ f −1 (B).
8 implies 2: If B is any closed set in Y , then .f −1 (B) ⊂ f −1 (B) = f −1 (B), and
.f
−1 (B) is thus a closed set.

Definition 13.34 A homeomorphism is a continuous bijective function f from a
topological space X onto a topological space Y such that .f −1 is continuous. If this
is the case, X and Y are called homeomorphic, or topologically equivalent.
Example 13.27 Two discrete spaces X and Y are homeomorphic if and only if
there exists a bijective function from X to Y . Therefore two discrete spaces are
homeomorphic if and only if they have the same cardinality. The same statement
holds for two indiscrete spaces.
Example 13.28 The space .R with the usual topology is homeomorphic to the
interval .(0, 1) with the relative topology. Indeed the map .x → (2x − 1)/(x(x − 1))
from .(0, 1) to .R is clearly a homeomorphism.
Exercise 13.28 Prove that any two open intervals of .R are homeomorphic. Hint: by
modifying the previous example, prove first that any open interval is homeomorphic
to .R.
13.7 Continuous Maps and Homeomorphisms 225
How can we prove that two spaces are, or are not, homeomorphic? It doesn’t
come as a surprise that this is a difficult task. As we said before, topology is exactly
the study of properties that are left invariant under homeomorphisms.
Definition 13.35 A topological invariant is a property which when possessed by a
space is also possessed by each homeomorphic space.
This definition offers the following consequence: if we can exhibit a topological
invariant which is possessed by one space but not by another space, then the two
spaces are hot homeomorphic.
Theorem 13.19 If X is a connected space and .f : X → Y is a continuous function,
then .f (X) is also connected.
Proof First of all, we may assume that f is surjective, since the map .f : X → f (X)
is continuous and surjective. Assume that B is both open and closed in Y . Hence
.f
−1 (B) is both open and closed in X by Theorem 13.18. Since X is connected,
either .f −1 (B) = ∅, or .f −1 (B) = X. Recalling that f is surjective, we conclude

that either .B = ∅ or .B = Y .

As an immediate corollary, if X and Y are homeomorphic and X is connected,
then Y is also connected. We can use this fact to disprove that two spaces are
homeomorphic.
Example 13.29 The interval .[0, 1] is not homeomorphic to .[0, 1/3] ∪ [1/2, 1].
Indeed .[0, 1] is connected, while .[0, 1/3] ∪ [1/2, 1] has two connected components.
A characterization of connected spaces in terms of continuous function is also
possible, as the next result shows.
Theorem 13.20 A topological space X is connected if and only if every continuous
function .f : X → R assumes for each pair of values .f (a) < f (b) also every .η
satisfying .f (a) ≤ η ≤ f (b).
Proof First we suppose the existence of a continuous function f from X to .R
which leaves out some value .η lying between .f (a) and .f (b). Hence .f (X) is not
an interval, thus .f (X) is not connected as a subset of .R. It follows that X is not
connected.
Conversely, we suppose that X is not connected. Hence there exists two non-
empty open sets .U1 and .U2 of X such that .X = U1 ∪ U2 and .U1 ∩ U2 = ∅. We
define .f (x) = 1 for every .x ∈ U1 and .f (x) = 2 for every .x ∈ U2 . If .x ∈ X, then x
belongs to either .U1 or .U2 , and f is identically equal to either 1 or 2, respectively. In
any case there exists a neighborhood U of x such that .f = f (x) on U . This shows
that f is continuous at x. It is evident that f cannot assume those values .η ∈ (1, 2),

13.8 Product Spaces, Quotient Spaces, and Inadequacy

of Sequences
Once we have a good topological definition of continuity, we can use it to construct

new spaces from old ones. There are essentially two abstract constructions: the first
one is based on cartesian products, in such a way that all the projections onto the
factors of the product should be continuous. The second one is based on a single
surjective function that generates the largest topology on its codomain for which the
function is continuous.
Definition 13.36 Let .X0 , .X1 , . . . , .Xn be topological spaces. Let .B the collection of
all sets of the form .U0 ×U' 1 ×· · ·×Un such that each .Ui is open in .Xi , .i = 0, . . . , n.
n
The
'n product topology on . i=0 Xi is the topology for which .B is a base: a subset of
. i=0 X i is open if and only if it is the union of members of .B.
'
Exercise 13.29 Prove that a set W is open in . ni=0 Xi if and only if for each
element .(x0 , x1 , . . . , xn ) ∈ U there exist open subsets .Ui ⊂ Xi , .i = 0, . . . , n,
such that .(x0 , x1 , . . . , xn ) ∈ U0 × U1 × · · · × Un ⊂ W . See Fig. 13.6.
'
Definition 13.37 The canonical projections on . ni=0 Xi are the functions
(
n
Pi :
. Xi → Xi
i=0
defined by .Pi (x0 , . . . , xn ) = xi for .i = 0, . . . , n.
Fig. 13.6 An open subset of .X1 × X2 × X3

13.8 Product Spaces, Quotient Spaces, and Inadequacy of Sequences 227
'
Theorem 13.21 If . ni=0 Xi is endowed with the product topology, then each .Pi is
a continuous function.
Proof Indeed, if .Ui is an open subset of .Xi , then
Pi−1 (Ui ) = X0 × · · · × Xi−1 × Ui × Xi+1 × · · · × Xn ,

.
which is open in the product topology.

The product topology is actually the smallest topology on the cartesian product
such that each projection is continuous. We prove this in the particular case of two
factors, but there is no difficulty in generalizing the proof to any finite number of
factors.
Theorem 13.22 Let X and Y be two topological spaces, The product topology of
.X × Y is the smallest topology such that the two projections .PX : X × Y → X and
.PY : X × Y → Y are continuous.
Proof Let .τ be any topology on .X × Y such that .PX and .PY are both continuous.If
U is open in X and V is open in Y , then .U × V is open in .τ , since .U × V =
PX1 (U ) ∩ PY−1 (V ), and this intersection is open as the intersection of two open
subsets. Hence the product topology is smaller than .τ , and we conclude by the
arbitrariness of .τ .

Example 13.30 The space .RN is the product space of N copies of .R' 1 = R, each
factor being endowed with the usual topology. The N-cells of the form . N i=0 (ai , bi ),
where .ai and .bi are real numbers, are a base of the product topology of .RN .
Exercise 13.30 Using some geometric intuition, prove that the product topology of
RN is actually equivalent to the usual topology generated by the distance
.
$
%N
%
.d(x, y) =
& |xi − yi |2 .
i=0
As a hint, you should prove that each open ball contains a N-cell and is contained
in another N-cell. This is fairly obvious, but you should make an effort to prove it.
In some applications to Functional Analysis, finite products of spaces are not
enough. A typical situation is that of function spaces and weak topologies. As we
have seen in Set Theory, the cartesian product of any collection of sets is already a
set of functions.
Definition 13.38 (Product Set) Suppose

' we are given ' a set .Xa for each a of some
index set A. The cartesian product . a∈A Xa = {Xa | a ∈ A} is the set of all
functions x on A such that .x(a) ∈ Xa for each .a ∈ A, i.e.
(

. Xa = x : A → {Xa | a ∈ A} ∀a ∈ A x(a) ∈ Xa
a∈A
' of .x(a). The projections .Pa are defined by .Pa (x) =

We will often write .xa instead
xa for each .a ∈ A and .x ∈ a∈A Xa .
Definition 13.39 Suppose that each set .Xa is endowed with a topology .τa , .a ∈ A.
The family of sets of the form .Pa−1 (U ), where U is'an open subset of .Xa , is a
subbase of a topology, called the product topology of . a∈A Xa .
A consequence of this definition is that a base for the product topology is the
family of all finite intersections of members of the subbase. Hence a typical member
of this base is a set of the form

.U = Pa−1 (Ua ) a ∈ F = {x | xa ∈ Ua for each a ∈ F } ,
where F is a finite subset of A, and .Ua is open in'.Xa for each .a ∈ F . A strange fact
arises from this discussion: a subset of the form . a∈A Ua , where each .Ua is open in
.Xa , need not be open in the product topology! This however is true if only finitely
many sets .Ua are different than .Xa .

'
Exercise 13.31 Prove that the projections .Pa : a∈A Xa → Xa are continuous.
Exercise 13.32 Compare the previous construction with Theorem 13.7.
'
Theorem 13.23 A function f from a topological space to a product space . a∈A Xa
is continuous if and only if the composition .Pa ◦ f is continuous for each projection
.Pa .
Proof If f is continuous, the .Pa ◦f is the composition of two continuous functions.

If .Pa ◦ f is continuous for each .a ∈ A, for each open subset U of .Xa the set
.(Pa ◦ f )
−1 (U ) = f −1 (P −1 (U )) is open. Since finite intersections of .P −1 (U )
a a a
form a base of the product topology, it follows that the inverse image under f of
any member of a base is an open subset. Hence f is continuous.

We now consider nets in a product space.
Theorem 13.24 A net S in a product space converges to a point x if and only if its
projections onto each factor converges to the projection of x.
'
Proof If .{Sn , n ∈ D} is a net in the product space . a∈A xa which converges to a
point x, since each projection is continuous it follows that the net .{Pa ◦ SN , n ∈ D}
converges to .Pa (x) = xa . Conversely, suppose that .{Sn , n ∈ D} is a net in the
product space such that .Pa (Sn ) converges to .xa for each .a ∈ A. Given any open
neighborhood .Ua of .xa , .{Pa (Sn ), n ∈ D} is eventually in .Ua , and therefore .{Sn , n ∈
D} is eventually in .Pa−1 (Ua ). We deduce that .{Sn , n ∈ D} must be eventually in any

finite intersection of sets of the form .Pa−1 (Ua ). Since these sets form a base of the
product topology, the net .{Sn , n ∈ D} converges to x.

Important: The Box Topology Is not Good

'
The previous proof explains why the so-called box topology on a product . a∈A Xa
is not good enough for most purposes. It is defined by stating that a set is open if
and only if it has the form
(
. Ua ,
a∈A
for some open sets .Ua ∈ Xa , .a ∈ A. This looks rather natural, since it is the
straightforward generalization of the product topology in .X × Y (or in any product
space with finitely many factors). Nevertheless, there is no reason why Theorem
13.24 should hold for such a topology. Indeed, looking at the proof, it is true that
−1 −1
.{Sn , n ∈ D} is eventually in .Pa (Ua ), say .Sn ∈ Pa (Ua ) as soon as .n ≥ na .
If however A contains infinitely many elements, it may very well happen that .na
“escapes to infinity”, and it is therefore impossible to select a single element .n ∈ D
such that .Sn ∈ Pa−1 (Ua ) when .n ≥ n for every .a ∈ A. It was possible for a finite
number of indices a, of course.
Remark 13.11 The previous result somehow motivates the terminology of point-
wise convergence for convergence in the product topology.
' This is even clearer when
we are dealing with a product of identical factors, . a∈A X = XA , which is simply
the collection of all functions from A to X. A net .{Fn , n ∈ D} converges to f in .XA
if and only if .{Fn (a), n ∈ D} converges to .f (a) for each .a ∈ A.
Theorem 13.25 The product of Hausdorff spaces is a Hausdorff space.
'
Proof Let x and y be distinct members of . a∈A Xa . Then .xa = ya for some a.
By assumption, there exist disjoint open neighborhoods U and V of .xa and .ya
respectively. Then .Pa−1 (U ) and .Pa−1 (V ) are disjoint open neighborhoods of x and
y respectively, so that the product space is Hausdorff.

We now take a break, and go back to the comparison of sequences and nets. We
have insisted that nets completely describe a topology, while sequences in general
do not. But why? The following is an example that leans on the product topology.
Example 13.31 We consider the space .X = RR with the product topology. Let
E = {f ∈ X | f (x) ∈ {0, 1} and f (x) = 0 only finitely often}

.
and .g ∈ X be the function which is 0 everywhere. A generic neighborhood of g is

therefore of the form
U = {h ∈ X | |h(y) − g(y)| < ε if y ∈ F }

.
for some finite set .F ⊂ R and some .ε > 0. Such a neighborhood intersects E in
the function h which is 0 on elements of F and 1 elsewhere. Hence .g ∈ E. We
now claim that no sequence in E can converge to g. Indeed, if .{fn }n is a sequence
in E with .fn being equal to zero on a set .An , then
∞any function which is a limit of
.{fn }n can be zero at most on the countable set . n=1 An . Clearly g does not meet
this condition, and the claim is proved.
Sequences do not suffice to assign a topology on a set. A natural question is
whether we can add any additional requirement in order to restore the full power of
converging sequences.
Definition 13.40 A topological space satisfies the first countability axiom, or
briefly is first countable, if the neighborhood system of each point of the space has
a countable base.
Theorem 13.26 Let X be a first countable space.
(a) A point x is an accumulation point of a subset A if and only if there exists a
sequence in .A \ {x} which converges to x.
(b) A subset A is open if and only if each sequence which converges to a point of A
is eventually in A.
(c) If x is an accumulation point of a sequence S, there is a subsequence of S
converging to x.
Proof Suppose first that x is an accumulation point of a subset A, and that .U0 ,
.
1n,. . . , .Un ,. . . is a countable base of the neighborhood system at x. Let .Vn =
U
i=0 Ui . Then the sequence .V0 , V1 , . . . , Vn , . . . is also a base of the neighborhood
system at x with the additional property that .Vn+1 ⊂ Vn for each n. For each n we
select a point .Sn in .Vn ∩ (An \ {x}), thus obtaining a sequence .{Sn }n which clearly
convergens to x. The converse of (a) is trivial.
If a subset A is not open in X, then there is a sequence in .X \ A which converges
to a point of A. Such a sequence fails to be eventually in A, and (b) is proved.
Finally, suppose that x is an accumulation point of a sequence S and that
.V0 , V1 , . . . , is a countable base for the neighborhood system at x such that .Vn+1 ⊂
Vn for each n. For any integer i we choose .Ni such that .Ni ≥ i and .SNi ∈ Vi . Thus
.{SNi }i is a subsequence of S which converges to x.

Remark 13.12 Any metric space .(X, d) satisfies the first axiom of countability.
Indeed the sequence

1
Un = x d(x, x0 ) <
.
n
is a countable base of the neighborhood system at a given point .x0 ∈ X. This is the
reason why mathematical analysis in .RN can be completely explained in terms of
sequences.
A nice consequence of the density of the rational numbers in .R is contained in
the following result.
Theorem 13.27 (Lindelöf) Let .C be a collection of open sets of .R. Then there
exists a countable sub-collection .{On }n of .C such that
∞

. {O | O ∈ C} = On .
n=1

Proof We call .U = {O | O ∈ C} and we pick any .x ∈ U . As such, there exists
.O ∈ C such that .x ∈ O. But O is an open set, hence there exists an open interval
.Ix such that .x ∈ Ix ⊂ O. By the density of .Q in .R, we can construct an open
interval .Jx whose end-points are rational numbers and which satisfies .x ∈ Jx ⊂ Ix .
The collection of all open intervals with rational
end-points is countable, we see
that .{Jx | x ∈ U } is countable and .U = {Jx | x ∈ U }. Now, for each interval
in .{Jx | x ∈ U } we select a set O from .C which contains it. In this way we have
constructed a countable sub-collection .{On }n of .C such that .U = ∞ n=1 On , and the
proof is complete.

This suggests a general definition connecting open covers and countability proper-
ties of the local neighborhoods.
Definition 13.41 A topological space X is a Lindelöf space if and only if every
open cover of X has a countable sub-cover.
After this discussion on the importance of nets, we return to the second
construction of a new space from an old one. Suppose that we are given a function
f from a topological space X onto a set Y . Can we topologize Y so that f is
continuous?
Definition 13.42 Let X be a topological space, and f be a function defined on X
with range Y . The family .U of all sets .U ⊂ Y such that .f −1 (U ) is open in X is the
quotient topology of Y induced by f .
Exercise 13.33 By using the elementary properties of preimages, prove that the
quotient topology is indeed a topology, i.e. that .U satisfies the axioms of a topology.
Suppose that Y has a topology .τ such that f is a continuous function. For each
U ∈ τ , we then have that .f −1 (U ) is open in X, which proves that U is also open for
.
the quotient topology. We have thus proved that the quotient topology is the largest
topology on Y such that f is a continuous function.
Exercise 13.34 Prove that a set B is closed in the quotient topology if and only if
f −1 (B) is a closed subset of X. Hint: .f −1 (Y \ B) = X \ f −1 (B).
.
We now establish a counterpart of Theorem 13.23 for the quotient topology.

Theorem 13.28 Let f be a continuous map of a space X onto a space Y and
let Y have the quotient topology. A function g from Y to a topological space Z
is continuous if and only if the composition .g ◦ f is continuous.
Proof The composition of two continuous functions is always a continuous func-
tion. To prove the converse, let U be open in Z and .g ◦ f be continuous. Then
.(g ◦ f )
−1 (U ) = f −1 (g −1 (U )) is open in X, and therefore .g −1 (U ) is open in the
quotient topology of Y by definition.

13.9 Initial and Final Topologies
Cartesian product and quotient spaces are actually special cases of more general
constructions of topologies which preserve the continuity of given families of
functions.
Problem Suppose that .Xα is a topological space for each index .α ∈ A, and suppose
that Y is a set. Functions .fα : Xα → Y are given. We wish to find a topology on Y
such that each .fα is a continuous function.
There is a trivial answer, of course, so we add a requirement: we wish to find the
largest topology on Y such that each .fα is continuous.
Definition 13.43 The largest topology on Y such that each .fα is continuous is
called the final topology on Y with respect to .{fα | α ∈ A}. This topology can
be explicitly described: a subset U of Y is open if and only if .fα−1 (U ) is open in .Xα
for each .α ∈ A.
The following result characterizes the final topology.
Theorem 13.29 (Universal Property of the Final Topology) For each .α ∈ A,
suppose that .Xα is a topological space. The topology of a space Y is the final
topology with respect to the functions .fα : Xα → Y if and only if the following
condition is satisfied: for any topological space Z, a function .g : Y → Z is
continuous if and only if .g ◦ fα : Xα → Z is continuous for every .α ∈ A.
Proof Suppose first that Y is endowed with the final topology with respect to
{fα | α ∈ A}. If g is continuous, then .g ◦ fα is continuous for every .α as the
.
composition of two continuous functions. Conversely, if .g ◦ fα is continuous for

any .α, then

fα−1 g −1 (U ) = (g ◦ fα )−1 (U )
.
is open in .Xα for any .α and for any open .U ⊂ Z. Hence .g −1 (U ) is open in Y and g
is continuous.
13.9 Initial and Final Topologies 233
Suppose now that the topology of Y satisfies the condition in the Theorem. Then
each .fα : Xα → Y is continuous, since the identity map on Y is continuous. Calling
.Ȳ the set Y endowed with the final topology with respect to .{fα | α ∈ A}, the
functions .fα : Xα → Ȳ are continuous. If .ι : Y → Ȳ is the identity function, then

the composition .ι ◦ fα : Xα → Ȳ is continuous for every .α. By our assumption
.ι is continuous. Also, each composition .ι
−1 ◦ f : X → Y is continuous. Since
α α
the topology of .Ȳ is the final topology of .Ȳ with respect to .{fα | α ∈ A}, .ι−1 is
continuous. Thus .ι : Y → Ȳ is a homeomorphism, and .Y = Ȳ .

If any .Xα is a subspace of Y , and if Y already has a topology, we expect some
coherence of the initial topology with respect to the inclusions .ια : Xα → Y . This
suggests the following definition.
Definition 13.44 Let Y be a topological space, and .{Xα | α ∈ A} be a collection of
subspaces of Y . The topology of Y is coherent with .{Xα | α ∈ A} if and only if it
coincides with the initial topology with respect to the inclusion maps .ια : Xα → Y .
The proof of the next characterization is left as a simple exercise.
Theorem 13.30 A necessary and sufficient condition that a space Y has a topology
coherent with a collection of its subspaces .{Xα | α ∈ A} is that .U ⊂ Y is open
(resp. closed) if and only if .U ∩ Xα is open (resp. closed) in .Xα for every index
.α ∈ A.
Let us now consider a dual situation.

Definition 13.45 Suppose that X is a set and .Yα is a topological space for every
index .α ∈ A. For each .α ∈ A, a function .fα : X → Yα is assigned. The initial
topology induced by .{fα | α ∈ A} is the smallest topology on X such that each .fα
is a continuous function. Equivalently, the family

. fα−1 (Uα ) α ∈ A, Uα is open in Yα
is a subbasis of this topology.

This definition should also be compared to Theorem 13.7.
'
Example 13.32 If .X = α∈A Yα , the initial topology induced by the canonical
projections .Pα : X → Yα is just the product topology.
Theorem 13.31 (Universal Property of the Initial Topology) Assume that X is
endowed with the initial topology induced by the functions .fα : X → Yα , .α ∈ A. A
function .g : Z → X, Z a topological space, is continuous if and only if .fα ◦g : Z →
Yα is continuous for each .α ∈ A. Moreover, this property characterizes the initial
topology of X.
Proof Suppose first that g is continuous. Each .fα : g is then continuous. On the
other hand, suppose that each .fα : g is continuous. Then, for any open .Uα ⊂ Yα ,

g −1 fα−1 (Uα ) = (fα ◦ g)−1 (Uα )
.
is open in Z. By the definition of the initial topology, g is then continuous.

To establish the last statement, suppose that X has the initial topology with
respect to .{fα | α ∈ A}, and let .X̄ be the set X endowed with a topology such that
a function .g : Z → X is continuous if and only if .fα ◦ g : Z → Yα is continuous
for each .α ∈ A. The functions .fα : X̄ → Yα are then continuous, since the identity
map on .X̄ is continuous. If .ι : X → X̄ is the identity map, then the continuity of
the compositions .fα ◦ ι = fα : X → Yα implies that .ι is continuous. Similarly,
the continuity of the compositions .fα ◦ ι−1 : X̄ → Yα implies that .ι−1 : X̄ → X is
continuous. Hence X and .X̄ are homeomorphic, and the proof is complete.

A word of warning: the initial and the final topologies are often known under
different names. The term weak topology is sometimes used for both of them, while
we will reserve this name for a very special topology on Banach spaces. The names
projective topology, induced topology are also used in the literature.
13.10 Compact Spaces
Mathematical analysts tend to believe that compactness is the most important

topological property at all. In some sense this is correct: a good compactness
property is of fundamental importance in many fields, from Functional Analysis
to Calculus of Variations. We have already encountered sequential compactness in
.R, although the characterization in terms of closedness and boundedness somehow
hides the role of compactness. In a topological space it is generically impossible to

provide a necessary and sufficient condition for a set to be compact.
Definition 13.46 (Open Cover) Let X be a topological space, and let E be a
subset. An open cover of A is a family4 .{Uα | α ∈ A} of open subsets such that

E⊂
. Uα .
α∈A
Definition 13.47 (Compact Space) A topological space X is compact if and only

if every open cover of X has a finite subcover. A subset E is compact if and only if
it is compact as a space with the relative topology induced by X.
4 The indexed notation is preferable. Otherwise

we should say that a collection .U of open subsets
is an open cover of E if and only if .E ⊂ U.
13.10 Compact Spaces 235
More explicitly, a space X is compact if and only if for every open cover
{Uα | α ∈ A} of X it is possible to find a finite set .F = {α1 , . . . , αn } ⊂ A such
.
that
X ⊂ Uα1 ∪ · · · ∪ Uαn .
.
Remark 13.13 The difficulty of the definition is that there is no restriction on the
cardinality of the open cover. Compactness is a highly demanding property.
As a first step, we characterize compactness in terms of an intersection property.
Definition 13.48 A family .A of sets has the finite intersection property if and only
if the intersection of the members of each finite subfamily of .A is non-empty.
Theorem 13.32 A topological space is compact if and only if each family of closed
subsets which has the finite intersection property has a non-empty intersection.
Proof If .A is a family of subsets of a topological space X, then

X\
. A= (X \ A) .
A∈A A∈A
As a consequence, the family .A is a cover of X if and only if the intersection of

the complements of the members of .A is empty. The space X is compact if and
only if each family of open sets such that no finite subfamily covers X fails to be a
cover. This is true if and only if each family of closed sets which possesses the finite
intersection property has a non-empty intersection.

In the framework of .R, we have defined compactness via the existence of conver-
gent subsequences. The next result shows, in greater generality, that compactness is
indeed a property of converging subnets.
Theorem 13.33 A topological space X is compact if and only if each net in X has
a subnet which converges to some point of X.
Proof It suffices to show that each net in X has an accumulation point. Indeed,
a point is an accumulation point of a net if and only if some subnet converges to
it. Let .{Sn , n ∈ D} be a net in the compact space X, and for each .n ∈ D let .An =
{Sm | m ≥ n}.5 The family .{An }n has the finite intersection property, since .≥ directs
D. Consequently the family .{An }n has the finite intersection property. Recalling that
X is compact, there exists a point x which belongs to every .An , and such a point is
an accumulation point of .{Sn , n ∈ D}.
Conversely, we suppose that X is a topological space in which every net has
an accumulation point. Let .A be a family of closed subsets of X that has the finite
intersection property. We define .B as the family of all finite intersections of members
5 The set .An can be called the n-tail of the net .{Sn , n ∈ D}.
of .A. Clearly .B has the finite intersection property, and .A ⊂ B. It is therefore

sufficient to prove that .{B | B ∈ B} = ∅. Observe that the intersection of two
members of .B is again a member of .B, so that .B is directed by .⊂. We choose6 an
element .SB from each .B ∈ B, obtaining a net .{SB , B ∈ B, ⊂} in X. By assumption
this net has an accumulation point x. Let B and C be any two members of .B such
that .C ⊂ B; then .SC ∈ B, and thus the net .{Sb , B ∈ B} is eventually in the closed
subset B. This implies that the accumulation point x lies in B. Since B was arbitrary,
the point x is a member of each member of .B, and the intersection of all members
of .B is thus non-empty.

The following a simple but important corollary.
Theorem 13.34 A closed subset of a compact space is compact.
Proof Let X be a compact space, and A be a closed subset of X. We suppose that a
net .{Sn , n ∈ D} satisfies .Sn ∈ A for each .n ∈ D. Since X is compact, there exists a
subnet .S ◦ N which converges to a point .x ∈ X. But A is closed, thus .x ∈ A, and
we conclude.

Exercise 13.35 Prove the last theorem by means of open covers. Hint: let A be
closed in the compact space X. If .{Uα | α ∈ A} is an open cover of A, we can add
.X \ A to it and get an open cover of X. Now use the compactness of X.
The converse statement is less trivial, and generally false. It becomes true under
a separation assumption.
Theorem 13.35 If A is a compact subset of a Hausdorff space X and x is a point of
X \A, then there are disjoint neighborhoods of x and A. In particular, each compact
.
subset of a Hausdorff space is closed.

Proof Since X is a Hausdorff space, there is a neighborhood U of each point y of
A such that x does not belong to the closure .U . Because A is compact, there exists
a finite family .U0 , .U1 , . . . , .Un of
open sets covering A and such that .x ∈
/ U i for
n
.i ∈ {0, 1, . . . , n}. Letting .V = U
i=0 i , then .A ⊂ V and .x ∈
/ V . Consequently
.X \ V and V are disjoint neighborhoods of x and A.

Theorem 13.36 Let X be a compact space, Y be a topological space, and .f : X →
Y a continuous function. The Y is compact. Furthermore, if Y is Hausdorff and f is
bijective, then f is a homeomorphism.

Proof If .A is an open cover of Y , then . f −1 (A) A ∈ A is an open cover of
X which must have a finite sub-cover. The collection of the images (under f ) of
the members of this sub-cover is a finite sub-collection of .A which covers Y , and
consequently Y is compact. Suppose that Y is Hausdorff and f is bijective. Consider
any closed .A ⊂ X; then A is compact and .f (A) is compact as well. Hence .f (A)
6 This is an evident application of the Axiom of Choice, since we need a function .B → SB .

is closed in Y . This proves the .(f −1 )−1 (A) is closed, and the continuity of .f −1
follows.

Exercise 13.36 Provide an alternative proof of the previous result by using nets and
Theorem 13.33.
Theorem 13.37 If A and B are disjoint compact subsets of a Hausdorff space X,
then there exist disjoint neighborhoods of A and B.
Proof By Theorem 13.35, to each .x ∈ A there corresponds a neighborhood of
x and a neighborhood of B which are disjoint. As a consequence there exists a
neighborhood U of x such that .U ∩ B = ∅, and since B is compact there exists a
family .Ui , .i = 0, . . . , n such that .Ui ∩ B = ∅ for .i = 0, . . . , n and .A ⊂ V =
finite
{Ui | i = 0, . . . , n}. Then V is a neighborhood of A and .X \ V is a neighborhood
of B which is disjoint from V .

The fundamental example of compact sets for a mathematical analyst is clearly
the N-cell in .RN . We develop here a proof which does not make use of Tychonoff’s
theorem on the product of compact sets.
Definition 13.49 An N-cell in .RN is a cartesian product of N closed and bounded
intervals, i.e. a set of the form

N
. x ∈ R for each i ∈ {1, . . . , N}, ai ≤ xi ≤ bi ,
where .ai and .bi are real numbers. Hence an N-cell is the cartesian product
[a1 , b1 ] × [a2 , b2 ] × · · · × [aN , bN ].

.
Theorem 13.38 Let N be a positive integer. If .{In }n is a sequence of N-cells such

that .In+1 ⊂ In for each n, then . ∞
n=1 In = ∅.
Proof We first prove the statement for .N = 1. Suppose that .In = [an , bn ], and
let E be the set of all .an . Clearly .an ≤ b1 , so that E is bounded from above and
non-empty. We can therefore set .α = sup E. If now m and n are positive integers,
then .an ≤ am+n ≤ bm+n ≤ bm , and we conclude that .α ≤ bm for each m. Therefore
.am ≤ α ≤ bm for each m, and therefore .α ∈ Im for each m.
We now consider the general case .N > 1. Suppose that .In consists of all points
.x = (x1 , . . . , xN ) such that .an,j ≤ xj ≤ bn,j for each n and .j ∈ {1, . . . , N}.
We define .In,j = [an,j , bn,j ]. For fixed j , the sequence .In,j of 1-cells has non-
empty intersection. Hence there exist real numbers .xj∗ , .j = 1, . . . , N, such that
∗ ∗ ∗ ∗
.an,j ≤ xn ≤ bn,j for each n and .j ∈ {1, . . . , N. The point .x = (x , . . . , x ) lies
1 N
in each .In , and the proof is complete.

Unlike most topological spaces, the Euclidean space .RN possesses a complete
characterization of all compact subsets. This is an important result for analysis, and
we provide a statement that anticipates a more general result about sequences in
compact spaces.
Theorem 13.39 For a subset K of .RN the following statements are equivalent to
each other:
(a) K is closed and bounded;
(b) K is compact;
(c) every sequence in K has an accumulation point in K.
Proof Since any bounded subset is contained in a suitable N-cell, we use Theorem
13.34 to show that (a) implies (b). Suppose now that K is compact and E is an
infinite subset of K. If no point of K is an accumulation point of E, then each point
of K has an open neighborhood which contains at most one point of E. The union
of all these neighborhoods is an open cover of K which has no finite subcover,
since E is an infinite set. Thus (b) implies (c). We prove that (c) implies (a). If K
is unbounded, we can construct a sequence of points .xn ∈ K such that .|xn | > n
for each n. The set of these points is infinite and has no accumulation point in
K, in contradiction with (c). Hence K is bounded. Suppose that K is not closed,
i.e. there exists a point .x0 of .RN which is an accumulation point of K but does
not belong to K. For each positive integer n, there exist points .xn ∈ K such that
.|xn − x0 | < 1/n. Let .S = {xn | n ∈ N}. The set S is infinite, otherwise the positive
number .|xn − x0 | would be constant for infinitely many values of n. Furthermore

.x0 is an accumulation point of S, and we claim that no other point of .R
N is an
accumulation point of K. Indeed, if .y ∈ R , .y = x0 , then

N
|xn − y| ≥ |x0 − y| − |xn − x0 |

.
1 1
≥ |x0 − y| − ≥ |x0 − y|.
n 2
for all but finitely many values of n. This shows that y cannot be an accumulation
point of S. We have reached a contradiction with (c), so that K must be closed.
What about sequences? Recall that we called sequentially compact any subset K
of .R with the following property: any sequence in K has a converging subsequence
in K. It is evident that the last property can be generalized to any setting.
Definition 13.50 A topological space is sequentially compact if and only if every
sequence in the space has a subsequence which converges to some point of the
space. A subset is sequentially compact if and only if it is sequentially compact in
the relative topology.
We already know that sequences are not sufficient to describe the topology in
the general case. As a matter of facts, sequential compactness is not equivalent to
compactness in a generic topological space.
Definition 13.51 A topological space X satisfies the second axiom of countability
if and only if the topology of X has a countable base.
Theorem 13.40 If X satisfies the second axiom of countability, then the following
statements are equivalent:
(a) Every sequence in X has an accumulation point;
(b) for each sequence in X there exists a subsequence converging to a point of X;
(c) X is compact.
Proof If the topology of X has a countable base, then every open cover of X has a
countable subcover. It is sufficient to prove that (a) implies (c), since (a) and (b) are
equivalent by previous results. So, we must show that every open cover of X has a
finite subcover. By assumption we may assume that the open cover is a sequence of
open sets
A0 , A1 , A2 , . . . , An , . . .
.
By induction we set .B0 = A0 , and for each .p ∈ N we define .Bp as the first member
of the cover which is not covered by
B0 ∪ B1 ∪ · · · ∪ Bp−1 .
.
If such a choice is impossible at any stage, we conclude that .{B0 , . . . , Bp−1 } is a

finite subcover. Otherwise it is possible to select a point .bp ∈ Bp for each .p ∈ N
such that .bp ∈
/ Bi for .i < p. Let x be an accumulation point of this sequence, so
that there exists p with .x ∈ Bp . Since x is an accumulation point, we must have
.bq ∈ Bp for some .q > p, a contradiction.

Using Theorem 13.32 we can generalize Weierstrass’ theorem on the existence
of minima and maxima.
Definition 13.52 Let X be a topological space, and let .f : X → R be a function.
We say that f is lower semicontinuous if and only if the set .f ≤α = {x ∈ X | f (x) ≤
α} is closed for each .α ∈ R. We say that f is upper semicontinuous if and only if
the set .f ≥α = {x ∈ X | f (x) ≥ α} is closed for each .α ∈ R.
Theorem 13.41 (Generalized Weierstrass) Suppose that X is a compact topolog-
ical space, .f : X → R is a function, and .m = infX f . Then .m ∈ R and .m ∈ f (X).
Proof In principle .m ∈ [−∞, +∞). For each .α > m the set .f ≤α is closed. The
family .{f ≤α | α > m} has the finite intersection property, since its members are
clearly ordered by inclusion. The compactness of X implies the existence of a point
x which belongs to each .f ≤α , .α > m. Hence .m ≤ f (x) ≤ α for each .α > m,
which means .f (x) = m. In particular .m > −∞.

We conclude this section with one of the most important results of General
Topology. Roughly speaking we want to prove that the cartesian product of compact
spaces is compact. The proof is reasonably elementary in the case of two compact
spaces. The proof in the case of a generic product, finite or infinite, countable or
uncountable, requires much more care. We begin with a more refined criterion for
compactness via nets.
Definition 13.53 A net .{Sn , n ∈ D} in a topological space X is a universal net if
and only if for each .E ⊂ X, it is either eventually in E or eventually in .X \ E.
Remark 13.14 An universal net in a topological space converges to any of its
accumulation points. Indeed, if the net is frequently in a set, then it is eventually
in this set.
Theorem 13.42 (Kelley) Let X be a non-empty set. Every net in X has a universal
subnet.
Proof Let .S = {Sn , n ∈ D} be a net in X and let
= {F ⊂ X | S is eventually in F } .
.
It is clear that .A ∈ and .B ∈ imply .A ∩ B ∈ . As a consequence the following

properties hold for .:
(1) S is frequently in every element of .;
(2) . has the finite intersection property.
Let .S be the set of all families of sets in X which contain . and have the properties
(1) and (2). We order .S by inclusion .⊂, and remark that every totally ordered
collection in .S has an upper bound given by the union of all of its elements. Hence
Zorn’s Lemma applies and provides us with a maximal collection . which contains
. and has properties (1) and (2).
Let .A ⊂ X and suppose that A does not belong to .. We claim that .X \ A ∈ .
Indeed, either S is eventually in .X \ A or there exists .B ∈ such that .A ∩ B = ∅. If
S is eventually in .X \ A then .X \ A ∈ ⊂ . Therefore we assume that there exists
.B ∈ such that .A ∩ B = ∅. Then S must be frequently in .X \ A or else it could
not be frequently in B and .B ⊂ X \ A. As a consequence we can add .X \ A to get

a larger collection .% which also has the properties (1) and (2). But . is maximal
with respect to these properties, hence .% = and .X \ A ∈ .
Since S is frequently in every element of . and . has the finite intersection
property, there exists a subset of .S ◦ N which is eventually in every element of ..
Let .A ⊂ X. If .A ∈ , then .S ◦N is eventually in A. If, on the contrary, A is not in .,
then we have seen before that .X \ A ∈ and therefore .S ◦ N is eventually in .X \ A.
This means that .S ◦ N is a universal subnet of S, and the proof is complete.

Remark 13.15 Kelley’s result can be proved in a more straightforward way.7 We
describe here the main steps. As a rule, if D is a directed set, we write
D≥n = {p ∈ D | p ≥ n} .
.
7 This proof appears in [1].

1. Let .S be the collection of all subnets of the net S. Since .S ∈ S, we see that
.S = ∅.
2. If .T1 ∈ S and .T2 ∈ S, we define .T1 ≥ T2 if and only if .T1 is a subnet of .T2 . It
can be proved that .≥ is a partial order on .S.
3. Let .{Ti | i ∈ I } be8 a totally ordered subset of .S, where each .Ti is defined on a
directed set .E i . We set

.E = E≥m mi ∈ E , i ∈ I ,
i i
i
j
and we order E by defining .E≥mj ≥ E≥m
i
i
if and only if .Tj ≥ Ti and
j
Tj (E≥mj ) ⊂ Ti (E≥m
.
i
i
).
and .E≥mj are given, and if .Tj ≥ Ti , we can find .m%j ∈ E j such that
i j
4. If .E≥mi
m%j ≥ mj and .Tj (E≥m% ) ⊂ Ti (E≥m

j j j j
.
i
i
). Hence .E≥m% ≥ E≥m
i
i
and .E≥m% ≥ E≥mj ,
j j j
and therefore E is a directed set.
5. We define .T ∗ : E → X by
T ∗ (E≥m
.
i
i
) = Ti (mi ).
we have .T ∗ (E≥mj ). Thus .T ∗ ≥ Ti for every .i ∈ I , or .T ∗ is an

j j
If .E≥mj ≥ E≥m
i
i
upper bound for the collection .{Ti | i ∈ I }, since .T ∗ is evidently a subnet of S.
6. Zorn’s Lemma applies, and there exists a maximal element .T ∈ S. It is easy to
check that T is a universal net and a subnet of S.
Theorem 13.43 A topological space X is compact if and only if each universal net
in X converges.
Proof We already know that X is compact if and only if each net has an
accumulation point. By the previous result, this happens if and only if each universal
net converges in X.

Theorem 13.44 (Tychonoff) A nonempty product space is compact if and only if
each factor space is compact.
'
Proof Let . α∈A Xα be a nonempty product space. If it is compact, and since all
the projection maps are continuous, then each .Xα , .α ∈'A, is a compact space.
Conversely, let .{Sn , n ∈ D} be a universal net in . α∈A Xα . For each .α ∈ A,
the net .{Pα ◦ Sn , n ∈ D} is a universal net in the compact space .Xα , and therefore
it converges to some point. We have proved that each component of the original
8Here we are using a bound-variable notation of a collection of nets. The variable i is not the
dummy variable which runs over a directed set, but a dummy variable which labels the elements
of the collection. Unfortunately it would be quite difficult to switch to an intrinsic notation.
universal net converges, and therefore the universal net itself must converge by
Theorem 13.24.

We present a slightly different proof of Tychonoff’s Theorem, due to Paul R.
Chernoff [2].
Proof Let .{Xα | α ∈ A} be an indexed family of non-empty topological spaces,
' of which is compact. A basic neighborhood N of an element .f ∈ X =
each
α∈A Xα is determined by a finite subset .F ⊂ A, together with neighborhoods
.Uα of .f (α) in .Xα for every .α ∈ F . Hence N consists of all .h ∈ X such that for
all .α ∈ F , .h(α) ∈ Uα . We will say that N is supported on F , and we will write

.N = N{Uα | α ∈ F }. A partially defined element g of X is a function g with
domain .J ⊂ A such that, for every .α ∈ J , .g(α) ∈ Xα .

To prove the theorem, let .{Sn , n ∈ D}, be a net in X. Suppose that g, with domain
.J ⊂ A, is a partially defined element of X. We say that g is a partial accumulation
point of our net if and only if for every .n ∈ D, for every

' finite subset .F ⊂ J and for
every basic neighborhood .N{Uα | α ∈ F } of g in . α∈J Xα , there exists .m ∈ N,
.m ≥ n, such that .Sm (α) ∈ Uα for every .α ∈ J . Of course, if the domain of g
coincides with A, then g is an accumulation point in X of the net .{Sn , n ∈ D}. We

claim that such a g exists.
Let .P be the set of all partial accumulation points of the given net .{Sn , n ∈ D}.
Since .∅ ∈ P, we see that .P = ∅. We introduce a partial order on .P as follows:
.g1 ≤ g2 if and only if the domain of .g1 is contained in the domain of .g2 , and
.g1 = g2 on their common domain. Let .L = {gα | λ ∈ } be a totally ordered

9
subset of .P, and define

.g0 = {gλ | λ ∈ } .
Since any two elements of .L must agree on their common domain, .g0 is a partially
defined element of X. Furthermore, .g0 ∈ P, since every basic neighborhood of .g0
has finite support F , and thus F is contained in the domain of .gλ for some .λ ∈ .
To summarize, .g0 ∈ P and .g0 is an upper bound of .L.
We can now use Zorn’s Lemma, which yields a maximal element g in .P. We
want to show that the domain J of g coincides 'with A. Otherwise, we may choose
.k ∈ A \ J . Now g is an accumulation point in .
Xα ofthe net .{(Sn )|J | n ∈ D},
α∈J
and thus g is the limit of some subnet . (Sϕ(β) )|J β∈B .

Now, every .Xk is a compact space, the net . Sϕ(β) β ∈ B has an accumulation
point .p ∈ Xk . We define a function h with domain .J ∪{k} by setting .h = g on J and
.h(k) = p. It is clear that h is a partial accumulation point of the net .{Sn , n ∈ D},
9 Such a definition will return in the proof of the Hahn-Banach Theorem.

hence .h ∈ P and h is strictly larger than g. This contradicts the maximality of g in

P, hence the domain of g is A, g is an accumulation point of .{Sn , n ∈ D}, and the
.
proof is complete.

Yet another proof of Tychonoff’s Theorem is based on the following result, of
independent interest.
Theorem 13.45 (Alexander Sub-Base Theorem) Let X be a topological space
with a sub-base .B. Then the following are equivalent:
(i) Every open cover has a finite subcover (i.e. X is compact);
(ii) Every sub-basic open cover has a finite subcover.
With a clear choice of words, a sub-basic open cover is merely a cover which
consists of elements taken from the sub-base .B.
Proof We propose a proof by T. Tao. Call an open cover bad if it has no finite
subcover, and good otherwise. It suffices to show that if every sub-basic open cover
is good, then every basic open cover is also good, where basic refers to the basis
B∗ = {B1 ∩ · · · ∩ Bk | B1 , . . . , Bk ∈ B, k ∈ N}
.
is the standard basis associated to the sub-basis .B. Suppose for contradiction that
every sub-basic open cover was good, but at least one basic open cover was bad. If
we order the bad basic open covers by set inclusion, observe that every chain of bad
basic open covers has an upper bound that is also a bad basic open cover, namely the
union of all the covers in the chain. Thus, by Zorn’s lemma, there exists a maximal
bad basic open cover
C = {Uα | α ∈ A} .
.
Thus this cover has no finite subcover, but if one adds any new basic open set to this
cover, then there must now be a finite subcover.
Pick a basic open set .Uα from this cover .C. Then we can write
Uα = B1 ∩ · · · ∩ Bk
.
for some choice of the sub-basic open sets .B1 , . . . , Bk . We claim that at least one of
the .B1 , . . . , Bk also lies in .C. Suppose not, and observe that adding any of the .Bi to
.C enlarges the basic open cover and thus creates a finite subcover; thus .Bi together
with finitely many sets from .C cover X, or equivalently one can cover .X \ Bi with
finitely many sets from .C. Thus one can also cover

k
X \ Uα =
. (X \ Bi )
i+1
with finitely many sets from .C and thus X itself can be covered by finitely many
sets from .C, a contradiction.
From the above discussion and the axiom of choice, we see that for each basic
set .Uα in .C there exists a sub-basic set .Bα containing .Uα that also lies in .C. (Two
different basic sets .Uα , .Uβ could lead to the same sub-basic set .Bα = Bβ , but this
will not concern us.) Since the .Uα cover X, the .Bα do also. By hypothesis, a finite
number of .Bα can cover X, and so .C is good, which gives the desired a contradiction.

'
Proof of Tychonoff’s Theorem via Sub-Bases Let .X = {Xα | α ∈ A} a product
of compact spaces. In virtue of the Alexander sub-base Theorem, it suffices
to show
−1
that any open cover of X by sub-basic open sets . παβ (Uβ )) β ∈ B has a finite
sub-cover, where B is some index set, and for each .β ∈ B, .αβ ∈ A and .Uβ is open
in .Xαβ .
For each .α ∈ A, consider the sub-basic open sets .πα−1 (Uβ ) that are associated to
those .β ∈ B with .αβ = α. If the open sets .Uβ here cover .Xα , then by compactness
of .Xα , a finite number of the .Uβ already suffice to cover .Xα , and so a finite number
of the .πα−1 (Uβ ) cover X, and we are done. So we may assume that the .Uβ do not
cover .Xα , thus there exists .xα ∈ Xα that avoids all the .Uβ with .αβ = α. One then
sees that the point .(xα )α∈A in X avoids all of the .πα−1 (Uβ ), a contradiction. The
claim follows.

H. Lebesgue proved an interesting result: if .U is an open cover of a closed
interval of .R, then there exists a radius .r > 0 such that, if .|x − y| < r, then x and y
are both contained in some member of the cover .U. It is not so easy to provide an
intuitive proof without mentioning compactness: if it is evident that each open set
of .R contains an open interval of some length, in general this length depends on the
member of the open cover.
We prove a generalization of Lebesgue’s result valid in any metric space.
Theorem 13.46 (Lebesgue Covering Lemma) If .U is an open cover of a compact
subset A of a metric space .(X, d), then there exists a positive number r such that
the open sphere of radius r about each point of A is contained in some member of
.U.
Proof By compactness, we may assume that .U = {U1 , . . . , Un }. We set
fi (x) = d(x, X \ Ui ) = inf {d(x, y) | y ∈ X \ Ui } ,

.
f (x) = max{f1 (x), . . . , fn (x)}.
Each .fi is a continuous function, and consequently f is a continuous function. Each

point x of A belongs to some member .Ui of .U, hence .f (x) ≥ fi (x) > 0. The set
.f (A) is a compact subset of .(0, +∞), so there exists .r > 0 such that .f (A) ⊂
(r, +∞). Therefore, for each .x ∈ A there is an index i such that .fi (x) > r, and it
follows that the open sphere of radius r about x is contained in .Ui .

Compactness can imply cardinality properties. We say that a point x of a

topological space X is isolated, if the set .{x} is open in X.
Theorem 13.47 Suppose that X is a compact Hausdorff space. If no isolated points
exist in X, then X is uncountable.
Proof First of all, we show that to any non-empty open .U ⊂ X and any point .x ∈ X
there corresponds an open .V ⊂ U such that .x ∈ / V . Indeed, let .y ∈ U be a point
with .y = x. This choice is possible if .x ∈ U because x is not an isolated point, and
it is also possible if .x ∈
/ U because U is non-empty. By assumption we can choose
open neighborhoods .Wx and .Wy of x and y respectively, such that .Wx ∩ Wy = ∅.
Then .V = U ∩ Wy is the desired open set.
To complete the proof, we pick any function .f : N → X and we prove that f
cannot be surjective. This clearly implies that X is uncountable. For every .n ∈ N,
write .xn = f (n). We apply the previous claim to .U = X and choose a non-empty
open .V1 ⊂ X such that .V1 does not contain .x1 . By induction, if .Vn−1 has been
selected, we choose a non-empty open .Vn ⊂ Vn−1 such that .Vn does not contain
.xn . This construction produces a nested sequence of closed sets .V1 ⊃ V2 ⊃ · · · .

Recalling that X is a compact space, there exists a point .x ∈ ∞ n=1 Vn . This point x
is different from every .xn , since .xn ∈
/ Vn . The proof is complete.

Corollary 13.1 Every closed interval of .R is uncountable. In particular, .R is
uncountable.
Proof Immediate, since singletons are not open in the standard topology of .R.

13.10.1 The Fundamental Theorem of Algebra
Every Calculus student knows that a polynomial equation like
P (x) = 0
.
may not have a solution .x ∈ R. For instance, the equation .x 2 + 1 + 0 does not
have any real solution, by the obvious fact that .x 2 + 1 ≥ 0 + 1 = 1 > 0.
The Fundamental Theorem of Algebra states that every polynomial with complex
coefficients possesses at least a complex solution. The proof of this important result
is often postponed to a course in Complex Analysis.
In this section we present an elementary proof due to Charles Fefferman, see [3].
Theorem 13.48 Let .n ∈ N, .a0 , . . . , an ∈ C and
P (z) = a0 + a1 z + · · · + an zn
.
be a polynomial in the complex indeterminate z. Then P has a zero.

Proof We first prove that the function .z ∈ C → |P (z)| attains a minimum. To prove
this claim, we notice that

n an−1 a0
.|P (z)| = |z|
a n + z + · · · + z n
for every .z ∈ C \ {0}. Hence there exists a number .M > 0 such that
|z| > M ⇒ |P (z)| ≥ |a0 |.

. (13.1)
Since the set .B[0, M] = {z ∈ C | |z| ≤ M} is compact in .C, the continuous function
z → |P (z)| attains a global minimum at some .z0 ∈ B[0, M]. Hence
.
|P (z)| ≥ |P (z0 )|
. for every z ∈ B[0, M]. (13.2)
Since .0 ∈ B[0, M], we see that .|P (z0 )| ≤ |P (0)| = |a0 |, and (13.1) implies that
|P (z0 )| ≤ |P (z)| as soon as .|z| > M. A comparison with (13.2) shows that
.
|P (z0 )| ≤ |P (z)|
. for every z ∈ C. (13.3)
The claim is then proved. As a second and last step, we will show that .P (z0 ) = 0.
Indeed, we exploit the identity .P (z) = P (z0 + (z − z0 )) to write .P (z) as a sum
of powers of .z − z0 . More formally, there exists a polynomial Q such that
P (z) = Q(z − z0 ).
.
Hence (13.3) becomes
|Q(0)| ≤ |Q(z)|
. for every z ∈ C.
We need to prove that .Q(0) = 0. Let j the smallest positive integer such that .zj has
a non-zero coefficient in the expansion of the polynomial Q. Then we can write
Q(z) = c0 + cj zj + · · · + cn zn ,
.
where .cj = 0. Factoring .zj +1 in the last terms, we may write
Q(z) = c0 + cj zj + zj +1 R(z),
.
for some complex polynomial R. Writing

c0
. − = reiθ ,
cj
the number .z1 = r 1/j eiθ/j satisfies

j
cj z1 = −c0 .
. (13.4)
Let .ε > 0, so that

j +1
Q(εz1 ) = c0 + cj εj z1 + εj +1 z1
j
. R(εz1 ).
Pick .N > 0 so large that .|R(εz1 )| ≤ N for every .ε ∈ (0, 1). Recalling (13.4) we
see that

j j j +1
.|Q(εz1 ) ≤ c0 + cj ε z + ε
1 |z1 |j +1 |R(εz1 )|

j
≤ c0 + εj (cj z1 ) + εj +1 |z1 |j +1 N

= c0 + εj (−c0 ) + εj +1 |z1 |j +1 N
= (1 − εj )|c0 | + εj +1 |z1 |j +1 N
= |c0 | − εj |c0 | + εj +1 |z1 |j +1 N.
Suppose now that .c0 = 0. Since .ε is arbitrary, we can pick it so small that
.|Q(εz1 )| ≤ |c0 | − εj |c0 | + εj +1 |z1 |j +1 N < |c0 | = |Q(0)|.
This contradicts the fact that Q attains a global minimum at 0. Hence .c0 = 0, and
the proof is complete.

13.10.2 Local Compactness
Compactness is a very strong property of a topological space. If we think back of

the real line .R with its usual topology, we may observe that any open neighborhood
contains a compact neighborhood. Indeed, if U is a neighborhood of a point x, then
there exists an open ball .B(x, r) contained in U . Then the closed ball .B(x, r/2) is a
compact subset contained in U , and clearly it contains the open ball .B(x, r/4). This
is a very specific example of the following definition.
Definition 13.54 A topological space is locally compact if and only if each point
has at least a compact neighborhood.
Exercise 13.37 Prove that a compact space is locally compact.
Exercise 13.38 Prove that every discrete space is locally compact.
Exercise 13.39 Prove that each closed subspace of a locally compact space is
locally compact. Hint: the intersection of a closed set and a compact set is a closed
subset of the latter, and therefore a compact subset.
Example 13.33 It is false that the continuous image of a locally compact space
must be locally compact. This follows from the interesting fact that any topological
space is the continuous one-to-one image of a discrete space. Indeed, let .(X, τ ) be
a topological space, and let Y be the topological space consisting of the set X with
the discrete topology. The identity map from Y to X is continuous and bijective. But
every discrete space is locally compact, while X is an arbitrary topological space.
Definition 13.55 (Nowhere Dense) A set in a topological space is nowhere dense
if and only if its closure has an empty interior.
'
Theorem 13.49 Let .X = {Xα | α ∈ A} be a topological product space. If an
infinite number of the coordinate spaces .Xα are non-compact, then each compact
subset of X is nowhere dense.
Proof Suppose B is a compact subset of X with an interior
−1 point x. ThenB contains
a neighborhood U of x which is of the form .U = Pα (Vα ) α ∈ F , for some
finite subset F of A and some open sets .Vα in .Xα . If .β ∈ A \ F , then .Pβ (B) = Xβ
and .Xβ is compact as the continuous image of a compact space. As a consequence,
all but finitely many of the coordinate spaces .Xα are compact. The proof is complete.

Theorem 13.50 (Local Compactness of Product Spaces) If a product space is
locally compact, then each coordinate space is locally compact and all but a finite
number of coordinate spaces are compact.
Proof Suppose that a product space is locally compact. Since the projection into a
coordinate space is an open map, each coordinate space is locally compact. Indeed,
if a function is both continuous and open, the image of a compact neighborhood of
a point is a compact neighborhood of the image point.
If infinitely many coordinate spaces are non-compact, then each compact subset
of the product space is nowhere dense by Theorem 13.49. Hence no point can have
a compact neighborhood. The proof is complete.

13.11 Compactification of a Space
As a matter of facts, non-compact spaces exist. So the question is: can we somehow
embed a non-compact space into a compact one? To be more precise: can we
construct a compact space which contains the non-compact space as a subspace?
This is a typical problem in mathematical analysis. For instance, it is common
to attach two “points” .−∞ and .+∞ to the real line, so that the resulting set is a
compact space. In Complex Analysis, the complex unit sphere is constructed by
adjoining a single point .∞ to the bi-dimensional space .C and specifying that the
neighborhoods of .∞ are the complements of bounded subsets of .C. In this section
we present an abstract construction along the same lines.
13.11 Compactification of a Space 249
Fig. 13.7 The compactification of .R is homeomorphic to the unit circle
Definition 13.56 Let X be a topological space. The one-point compactification of

X is the set .X∗ = X ∪ {∞} with the topology whose members are the open subsets
of X and all subsets U of .X∗ such that .X∗ \ U is a closed compact subset of X.
Remark 13.16 We have used the undefined symbol .∞ to denote any element which
is not a member of X. For instance .∞ = X can be a good choice, although very few
analysts think of .∞ in this way.
Figure 13.7 provides an intuition of the one-point compactification of the real line.
Theorem 13.51 (Alexandroff) The one-point compactification .X∗ of a topological
space X is compact and X is a subspace. The space .X∗ is Hausdorff if and only if
X is locally compact and Hausdorff.
Proof We follow [5]. A set U is open in .X∗ if and only if (a) .U ∩ X is open
in X and (b) whenever .∞ ∈ U , then .X \ U is compact. As a consequence, finite
intersections and arbitrary unions of sets open in .X∗ intersect X in open sets. If .∞ is
a member of the intersection of two open subsets of .X∗ , then the complement of the
intersection is the union of two closed compact subsets of X and is therefore closed
and compact. If .∞ belongs to the union of the members of a family of open subsets
of .X∗ , then .∞ belongs to some member U of the family, and the complement of
the union is a closed subset of the compact set .X \ U and is therefore closed and
compact. Consequently .X∗ is a topological space and X is a subspace.
Let .U be an open cover of .X∗ . Then .∞ is a member of some U in .U and .X \U is
compact, and hence there is a finite subcover of .U. This proves that .X∗ is compact.
If .X∗ is a Hausdorff space, then its open subspace X is a locally compact
Hausdorff space. To conclude, we must show that .X∗ is a Hausdorff space if X
is a locally compact Hausdorff space. It is sufficient to show that, if .x ∈ X, then
there exist disjoint neighborhoods of x and .∞. Since X is locally compact and
Hausdorff, there is a closed compact neighborhood U of x in X, and .X∗ \ U is the
required neighborhood of .∞.

Exercise 13.40 Prove that X is a compact space if and only if .∞ is an isolated

point of the one-point compactification .X∗ of X. In this sense, the one-point
compactification is “useless” (since X is already compact) if and only if we are
adjoining an isolated point to X.
13.12 Filters and Convergence
As we have seen, nets are a powerful generalization of sequences, and a topology

can be recovered in terms of convergent nets. Furthermore, nets depict topologies in
a rather dynamic way.
Another tool of convergence exists besides nets: it was introduced by H. Cartan
under the name of filters. We provide now a quick introduction to this topic,
following the book [8].
Definition 13.57 (Filters) Let X be any set. We say that a family .F of subsets of E
is a filter if and only if
1. .∅ ∈/ F;
2. if .A ∈ F and .B ∈ F, then .A ∩ B ∈ F;
3. if .A ∈ F and .A ⊂ B, then .B ∈ F;
Example 13.34 The following are example of filters.
(a) The family of all subsets of X which contain a fixed subset .X0 of X.
(b) The family of all neighborhoods of a point in a topological space X: this is the
neighborhood filter of that point.
(c) The family of all the complements of finite subsets of an infinite set X. In the
special case .X = N this is the Fréchet filter.
Definition 13.58 A family .B of subsets of X is a filter basis if and only if
1. if .A ∈ B and .B ∈ B, there exists .C ∈ B such that .C ⊂ A ∩ B.
2. .∅ ∈/ B.
Definition 13.59 A family .B of subsets of X is a filter basis of the filter .F if and
only if for each .V ∈ F, there exists .A ∈ B such that .A ⊂ V . We will say that the
filter .F is generated by .B, or that .B generates .F.
Concisely, the filter generated by a filter basis .B consists of all supersets of the
elements of .B.
Exactly as topologies, filters can be compared.
Definition 13.60 A filter .F1 is finer than a filter .F2 if .F2 ⊂ F1 . In other words,
every element of .F1 is also an element of .F2 . Similarly, .F1 is strictly finer than .F2
if .F2 ⊂ F1 and .F2 = F1 .
Remark 13.17 Again, it would be preferable to say that .F1 is larger than .F2 . We
adhere to the French tradition which uses .F1 est plus fin que .F2 .
Definition 13.61 A collection .U of subsets of X is a ultrafilter if and only if .U is a
filter and there is no filter strictly finer than .U.
13.12 Filters and Convergence 251
Example 13.35 Let X be a set, and a be a point of X. The family .U = {A | (A ⊂

X) ∧ (a ∈ A)} is a ultrafilter. This is a rather useless ultrafilter, and its name is
actually the trivial ultrafilter defined by the point a.
Exercise 13.41 Prove that the trivial ultrafilter defined by a is indeed a ultrafilter.
Hint: suppose .U% is a strictly finer filter. Then .U% must contain a set .A ⊂ X which
does not contain a. Since .{a} ∈ U, the ultrafilter .U% must contain .∅, which is
impossible.
The existence of non-trivial ultrafilters leans on the axiom of choice, or better on
Zorn’s Lemma.
Theorem 13.52 For each filter, there exists a ultrafilter which contains it.
Proof Let .F be a filter, and let .S be the set of all filters strictly finer than .F. We
claim that .S is inductive, in the sense that any totally ordered subset of .S has an
upper bound. Indeed, a family .{Fi | i ∈ I } of filters has an upper bound if for each
sub-family .{Fi | i ∈ J }, .J ⊂ I finite, and for each system .Ai ∈ Fi , the intersection

. {Ai | i ∈ J } = ∅. Now, if the family .{Fi | i ∈ I } is totally ordered, then the
previous condition is clearly true, and therefore .S is inductive.
By Zorn’s Lemma, .S has a maximal element which is a ultrafilter.

Example 13.36 Let X be a Hausdorff topological space, and let a be a point of X
which is not an isolated point. This amounts to saying that .{a} is not an open set.
The family
. {V \ {a} | V is a neighborhood of a}
is a filter base which defines a filter .F. As we have seen, there exists a ultrafilter .U
finer that .F. The elements of .F are mutually disjoint, and the same must be true for
the elements of .U. Actually, if .b = a, there exists a neighborhood V of a which
does not contain b, so that .V \ {a} belongs to .F and does not contain b. Clearly,
.V \ {a} does not contain a.
Theorem 13.53 Let X be a set. A filter .U is a ultrafilter if and only if for each
A ⊂ X, either .A ∈ U or .X \ A ∈ U.
.
Proof Let .U be a filter which contains either A or .X \A for each .A ⊂ X. If .U is not

a ultrafilter, there exists a filter .F strictly finer than .U. Hence there exists a subset A
of X such that .A ∈ F and .A ∈ / U. Then .X \ A ∈ U and therefore .X \ A ∈ F since
.F is finer than .U. As a consequence, .∅ = A ∩ (X \ A) ∈ F, a contradiction. Hence
.U is a ultrafilter.
Conversely, let .U be a ultrafilter, and suppose that .A ∈ / U. We set .B = X \ A.

For each .V ∈ U, V is not a subset of A, so that .V ∩ B = ∅. The set of all .V ∩ B as
V ranges over .U is a filter base which generates a filter .W.Now, we deduce from
.V ∩ B ⊂ V that .W is finer than .U, and therefore .W = U since .U is a ultrafilter.
Finally, .B = X ∩ B is an element of .W since it is an element of .U.

We now see how filters behave under the action of functions.
Definition 13.62 Let .f : X → Y and let .B be a filter base on Y . If .f −1 (B) = ∅

for each .B ∈ B, then .f −1 (B) = {f −1 (B) | B ∈ B} is a filter base on X which we
call the counter-image of .B under f . The filter generated by .f −1 (B) is called the
counter-image of the filter generated by .B.
Example 13.37 When .X ⊂ Y , we can consider f as the inclusion map. Every filter
.F on Y has a counter-image under f if and only if for each .B ∈ F, there results
.X ∩ B = ∅. This is the filter on X induced by .F. Compare it with the induced
topology on a subset.
Definition 13.63 Let .B be a filter base on X, and .f : X → Y be a map. The family
f (B) = {f (A) | A ∈ B} is a filter base on Y called the direct image of .B under f .
.
The filter generated by .f (B) is called the direct image of the filter generated by .B.
It should be remarked that .f (B) need not be a filter, even in the favorable case in
which .B is a filter.
Example 13.38 Let .{xn }n be a sequence in X, i.e. a function from .N into X. The
Fréchet filter on .N has a direct image .F in X, which we call the Fréchet filter of the
sequence, or the elementary filter associated to the given sequence: this is the set of
all subsets of X which contain all but finitely many terms .xn .
Theorem 13.54 If .B is a ultrafilter base of X, its direct image under a map
f : X → Y is a ultrafilter base on Y .
.
Proof Let .B ⊂ Y ; the sets .f −1 (B) and .f −1 (Y \ B) are complementary, hence at

least one of them belongs to the filter generated by .B, since .B is a ultrafilter base.
Hence either B of .Y \B belongs to the filter generated by .f (B), which is a ultrafilter
base.

Filters were introduced for the same reason as nets: to describe convergence in a
general setting.
Definition 13.64 Let X be a topological space. We say that a filter .F converges to
a point .x ∈ X if and only if .F is finer than the neighborhood filter of x. Concretely,
this means that every neighborhood of x belongs to .F. In this case we write .x ∈
lim F or .F → x.10
Remark 13.18 The convergence of a filter extends the definition of convergent
sequence. Indeed, the filter associated to a sequence .{xn }n converges to a point x
if and only if each neighborhood of x contains all but finitely many terms of the
sequence .{xn }n . In other words, this filter converges to x if and only if .xn → x in
the topology of the space.
Theorem 13.55 In a Hausdorff space, a filter can converge to at most one point.
10 Here we are pedantic: without any further assumption on the topology, a filter can converge to
different points, and this is why we write .x ∈ lim F. Nevertheless, the notation .x = lim F is often
used in the literature.
13.12 Filters and Convergence 253
Proof Indeed, if a filter .F converges to both x and y, there exists disjoint

neighborhoods of x and y whose intersection belongs to .F. But the intersection
is empty, a contradiction.

Filters characterize accumulation points.
Theorem 13.56 Let X be a topological space, and A be a subset of X. A point
a ∈ X is an accumulation point of A if and only if there exists a filter .F which has
.
a base consisting of subsets of A, or such that .A ∈ F, and which converges to a.

Proof Let us suppose that such a filter .F exists. Now .A ∈ F, and since .F is finer that
the neighborhood filter of the point a, we have that .A∩V ∈ F for each neighborhood
V of a. Hence .a ∈ A.
Conversely, if .a ∈ A, we have .A ∩ V = ∅ for each neighborhood V of a. The
neighborhood filter of a induces a filter base .B on A, consisting of subsets of A. On
the other hand, if V is a neighborhood of a, V belongs to the filter .F generated by
.B, because V contains .A ∩ V . In particular .F → a.

We conclude with the characterization of continuity in terms of converging filters.
Theorem 13.57 Let .f : X → Y be a map between two topological spaces. The
map f is continuous if and only if for each filter .F converging to x in X, the filter
.f (F) converges to .f (x) in Y .
Proof Let us suppose that the last condition is satisfied. In particular we can select
the neighborhood filter .F of x in X. Its image .f (F) converges to .f (x), and thus f
is continuous at x.
Conversely, let us suppose that f is continuous at x. For each neighborhood W
of .f (x) in Y we can pick a neighborhood V of x in X such that .f (V ) ⊂ W . This
shows that W belongs to the image of the neighborhood filter of x, and therefore this
image filter is finer than the neighborhood filter of .f (x) in Y . Now, if a filter .F is
finer than the neighborhood filter of x, its image .f (F) is finer than the neighborhood
filter of .f (x), and thus it converges to .f (x).

An interesting remark we make is that filters and nets are somehow homomorphic
structures.
Theorem 13.58 Consider a set X.
(a) If .F is a filter on X then the set .IF of pairs .(A, p) such that .A ∈ F and .p ∈ A
is directed by .(A, p) ≤ (B, q) if and only if .A ⊃ B. Furthermore the function
(A, p) ∈ IF → p ∈ X
.
is a net associated to the filter .F.

(b) If .{Sn , n ∈ D} is a net in X, then11
F = {A ⊂ X | ∃N ∈ D ∀n ∈ D(n ≥ N ⇒ Sn ∈ A)}
.
is a filter on X, associated to .{Sn , n ∈ D}. The subsets

An = Sj j ≥ n
.
are a basis for the filter .F.
Exercise 13.42 Prove Theorem 13.58. Establish a comparison between the conver-
gence of a net and the convergence of the associated filter.
We will not pursue further the study of filters, since it is by now apparent that they
produce the same results as nets. Choosing either nets or filters for a description of
the topology is essentially a matter of taste.
13.13 Epilogue: The Limit of a Function
A serious objection to any reasonable definition of the limit of a function between

topological spaces is that it not elegant. Although this is a matter of taste, let us
compare several possible definitions. In each of them, X and Y are topological
spaces, .f : X → Y is a function, and p is an accumulation point of X. The value q
of the limit is a point of Y .
Definition 13.65 (Traditional Definition) We say that .limx→p f (x) = q if and
only if for each neighborhood V of q there exists a neighborhood U of p such that
.f (V ∩ X \ {p}) ⊂ V .
Definition 13.66 (Limit à la Cartan) Let .F be a filter in X. We say that .limF f =

q if and only if the filter .f (F) converges to q. In particular, if .F is the filter
of neighborhoods of the point p, .limx→p f (x) = q if and only if for each
neighborhood V of q there exists a neighborhood U of p such that .f (U ) ⊂ V .
Let us compare the two definitions. If p does not belong to X (but is an accumulation
point of X), these definitions agree. On the other hand, if .p ∈ X, these definitions are
not equivalent. In particular, the limit à la Cartan exists if and only if f is continuous
at p, and therefore .q = f (p). The question is: can we tolerate this fact?
11 .A ∈ F if and only if the net .{Sn , n ∈ D} is eventually in A.

13.13 Epilogue: The Limit of a Function 255
While most textbooks in the USA propose the traditional definition, in France the
school of Bourbaki imposed the second definition from high schools to universities.
The reason is fairly philosophical: Bourbaki believes that (Real) Analysis is a
daughter of General Topology, and in General Topology the true concept is
continuity. What many people call removable discontinuities is a strange animal
that we can discard from the theory of limits.
If we are in love with nets, we may propose the following definition.
Definition 13.67 (Limits with Nets) We say that .limx→p f (x) = q if and only if
for each net .{xn , n ∈ D} such that .xn = p for each .n ∈ D and which converges to
p, the net .{f (xn ), n ∈ D} converges to q.
This turns out to be equivalent to the traditional definition, and this does not come
as a surprise: it is a mere generalization of the characterization of limits in metric
spaces in terms of sequences. But the price to pay is that the condition
xn = p for each n ∈ D
.
does not belong to the general theory of convergent nets. If the domain X is a
subspace of a metric space, there is a nice way out: we can say that .x > y if
and only if .|x − p| < |y − p|. In this way we exclude p in an elegant fashion,
and convergence with respect to the direction .> reduces to the traditional definition
of limit. What about filters? Well, the first idea is to use the “filter” of punctured
neighborhoods of p, but this is not a filter: the point p may get back through the
window if we pass to a superset of a punctured neighborhood.
So, the only way out is to “screw up” the whole topological space X to which p
belongs by removing p once and for all: now punctured neighborhoods of p form a
filter. But this amounts to considering the new function .f|X\{p} which agrees with f
away from the point p. But then .q = limx→p f (x) is equivalent to the requirement
that
f (x) if x = p
.g(x) =
q if x = p
be continuous at p.
Let us summarize:
In a general theory, the traditional definition of limit is artificial both in the

language of filters and in the language of nets. Excluding the limit point p
must be an additional requirement, and we must choose if we can accept it.
13.14 Separation and Existence of Continuous Extensions
A common problem in Analysis is related to the existence of continuous functions

that extend a given (continuous) function on a subset. Furthermore, the extension
should also preserve important properties of the extended function. In this Section
we will prove some general results which relate the solvability of such a problem to
the separation properties of the underlying topological space.
Definition 13.68 Let X be a topological space.
(a) We say that X is regular if and only if for each point x and for each
neighborhood U of x there exists a closed neighborhood V of x such that
.x ∈ V ⊂ U . Equivalently, X is regular if and only if for each point x and
for each closed .A ⊂ X, there exist disjoint open sets U and V such that .x ∈ U ,
.A ⊂ V .
(b) We say that X is normal if the following condition is satisfied: if A and B are
closed subsets of X, there exist open sets U , V such that .A ⊂ U , .B ⊂ V , and
.U ∩ V = ∅. See Fig. 13.8.
For reference we summarize the most useful separation properties in General

Topology.
Definition 13.69 A topological space X is
(i) a .T0 -space if and only if for every couple x, y of distinct points in X there
exists an open set which contains only one of them;
(ii) a .T1 -space if and only if for every couple x, y of distinct points in X there exist
open sets .U1 and .U2 such that .x ∈ U1 , .y ∈ U2 , but .y ∈
/ U1 and .x ∈
/ U2 ;
(iii) a .T2 -space, or a Hausdorff space, if and only if for every couple x, y of distinct
points in X there exists open sets .U1 , .U2 such that .x ∈ U1 , .y ∈ U2 , and
.U1 ∩ U2 = ∅;
(iv) a .T3 -space if and only if X is .T1 and regular;

(v) a .T4 -space if and only if X is .T1 and normal.
Fig. 13.8 Separation in normal spaces

13.14 Separation and Existence of Continuous Extensions 257
It should be noted that intermediate separation properties have been introduced, but
they are of little importance in our setting.
Exercise 13.43 Prove that a space X is .T1 if and only if every singleton .{x} is a
closed set.
In the basic topological space .R (endowed with its natural distance), and even in any
metric space, a point y is an accumulation point for a subset A if and only if each
neighborhood of y contains infinitely many elements of A. The proof is standard:
just consider open balls .B(y, 1/n) which shrink to y as .n → +∞. However
this argument depends on the metrizability of the topology. The following result
shows that a rather weak separation assumption ensures the validity of the previous
characterization of accumulation points.
Theorem 13.59 Suppose that y is an accumulation point of a subset A of a .T1 -
space X. Then every neighborhood of y contains infinitely many points of A.
Proof Let U be a neighborhood of y, and let .F = A ∩ (U \ {y}). We claim that
F is an infinite set. Otherwise, .X \ F would be an open set (since singletons are
closed in a .T1 -space), and .y ∈ X \ F . But the neighborhood .X \ F of y cannot
contain points of A other than y itself, and this contradicts the assumption that y is
an accumulation point of A. The proof is complete.

The following technical Lemmas will be useful in the sequel.
Lemma 13.2 Suppose that for each element t of a dense subset D of .(0, +∞), .Ft
is a subset of a set X such that
implies .Ft ⊂ Fs ,
(a) .t < s
(b) .X = {Ft | t ∈ D}.
For .x ∈ X let .f (x) = inf {t ∈ D | x ∈ Ft }. Then

. {x ∈ X | f (x) < s} = {Ft | t ∈ D, t < s}

{x ∈ X | f (x) ≤ s} = {Ft | t ∈ D, t > s}
for each real number s.

Proof By definition,
. {x ∈ X | f (x) < s} = {x ∈ X | inf {t ∈ D | x ∈ Ft } < s} ,
and the properties of the infimum imply that .{x ∈ X | f (x) < s} is the set of all
.x ∈ X such that for some .t ∈ D, .t < s and .x ∈ Ft . This proves the first identity of
the conclusion.
To prove the second one, we remark that .inf {t ∈ D | x ∈ Ft } ≤ s if and only
if for each .u > s there exists .t s it is true that .x ∈ Ft , then .inf {t ∈ D | x ∈ Ft } ≤ s
because D is dense in .(0, +∞). We conclude that the set of all x such that
f (x) = inf {t ∈ D | x ∈ Ft } ≤ s coincides with
.
. {x ∈ X | if t ∈ D, t > s, then x ∈ Ft } ,

Lemma 13.3 Suppose that for each element t of a dense subset D of .(0, +∞), .Ft
is an open subset of a topological space X such that
implies .Ft ⊂ Fs ,
(a) .t < s
(b) .X = {Ft | t ∈ D}.
Then the function f such that .f (x) = inf {t ∈ D | x ∈ Ft } is continuous.
Proof It is sufficient to prove that the set .{x ∈ X | f (x) < s} is open and the
set .{x ∈ D | f (x) ≤ s} is closed for each .s ∈ R. By Lemma 13.2, the set
.{x ∈ X | f (x) < s} is the union of open sets. Moreover,

. {x ∈ D | f (x) ≤ s} = {Ft | t ∈ D, t > s} ,
and the proof will be complete once we show that this set is identical with

. Ft t ∈ D, t > s .
For each .t ∈ D, .Ft ⊂ Ft , so that

. {Ft | t ∈ D, t > s} ⊂ Ft t ∈ D, t > s .
On the other hand, for each .t ∈ D with .t > s there exists .r ∈ D such that .s < r < t,
and thus such that .Fr ⊂ Fs . he reverse inclusion follows, and the proof is complete.

Theorem 13.60 (Urysohn’s Lemma) Let Y be a Hausdorff space. The following
1. Y is normal;
2. for each pair A, B of disjoint closed sets in Y , there exists a continuous function
.f : Y → R, which we call a Urysohn function, such that (a) .0 ≤ f ≤ 1 on Y ,
(b) .f = 0 on A, (c) .f = 1 on B.
Proof We show that 2. implies 1. Fix two closed sets A and B such that .A ∩ B = ∅.
Denoting by f a Urysohn function for this pair, we set .U = {y ∈ Y | f (y) < 1/2}
and .V = {y ∈ Y | f (y) > 1/2}. Then U and V are disjoint open sets such that
.A ⊂ U , .B ⊂ V .
The converse implication is proved as follows. Let D be the set of positive dyadic
rational numbers, i.e. the set of all numbers of the form .p2−q and p and q range over
all positive integers. For every .t ∈ D with .t > 1, we let .F (t) = X, .F (1) = X \ B,
and .F (0) be an open set which contains A and such that .F (0) ∩ B = ∅. For every
.t ∈ D with .0 < t < 1, we write .t = (2m+1)2
−n and choose, inductively on n, .F (t)
to be an open set which contains .F (2m2 ) and such that .F (t) ⊂ F ((2m + 2)2−n ).
−n
Of course this construction if possible because X is a normal space. Finally, let
.f (x) = inf {t ∈ D | x ∈ F (t)}. Lemma 13.3 shows that f is a continuous function.
The function f is zero on A because .A ⊂ F (t) for every .t ∈ D, and f is one on

B because .F (t) ⊂ X \ B for every .t ≤ 1 and .F (t) = X for .t > 1. Hence f is a
Urysohn function, and the proof is complete.

Although normality is a complete characterization of the existence of Urysohn’s
functions, sometimes a sufficient condition is needed.
Theorem 13.61 If X is a locally compact Hausdorff space, then the family of closed
compact neighborhoods of each point is a base for its neighborhood system.
Proof Fix any point x of X, and let C be a compact neighborhood of x, U be an
arbitrary neighborhood of x. Since X is Hausdorff and W is the interior of .U ∩ C,
then .W is a compact Hausdorff space, and W contains a closed compact set V which
is a neighborhood of x in .W . But V is also a neighborhood of x in W and is therefore
a neighborhood of x in X.

Theorem 13.62 If X is a regular topological space, A is a compact subset, and U
is an open set containing A, then there exists a closed neighborhood V of A such
that .V ⊂ U . In particular, a compact regular space is normal.
Proof Since X is regular, to each point .x ∈ A there corresponds an open
neighborhood W of x such that .W ⊂ U , and by compactness we may assume that
there exist finitely many
such
neighborhoods .W0 , .W1 ,. . . , .Wn such that .Wi ⊂ U for
each i. Then .V = Wi i = 0, . . . , n is the required closed neighborhood of
A.

Theorem 13.63 If U is a neighborhood of a closed compact subset A of a
regular locally compact topological space X, then there exists a closed compact
neighborhood V of A such that .A ⊂ V ⊂ U .
Moreover, there exists a continuous function .f : X → [0, 1] such that .f = 0 on
A and .f = 1 on .X \ V .
Proof To each .x ∈ A there corresponds a neighborhood W which is a closed
compact subset of U . Since A is compact, a finite union of such neighborhoods
covers A, and their union is a closed compact neighborhood V of A. Then V with the
relative topology is a regular compact spaces, which is normal by Theorem 13.62.
Hence there exists a continuous function .g : V → [0, 1] such that .g = 0 on A and
◦ ◦
.g = 1 on .V \ V , where .V denotes the interior of V . Let f be equal to g on V and
equal to 1 on .X \ V . Since .V ◦ and .X \ V are separated sets, it follows easily that f

is continuous on X.

We conclude this Section with another strong characterization of normal spaces.
We state a useful lemma.
Lemma 13.4 Let A be a closed subset of a Hausdorff normal space X, let .g : A →

R be a continuous function such that .|g(x)| ≤ c for every .x ∈ A. Then there exists
a continuous function .h : X → R such that
(a) .|h(x)| ≤ c/3 for every .x ∈ X;
(b) .|g(x) − h(x)| ≤ (2/3)c for every .x ∈ A.
Proof Let
c c

. A+ = x ∈ a g(x) ≥ , A− = x ∈ A g(a) ≤ − .
3 3
These two sets are disjoint and closed in the closed .A ⊂ X, so that both .A− and .A+
are closed in X. Since X is normal, a Urysohn function .h : X → R exists having
value .c/3 on .A+ and .−c/3 on .A− . Furthermore .−c/3 ≤ h(x) ≤ c/3 for every
.x ∈ X.

Theorem 13.64 (Tietze Extension Theorem) Let X be a Hausdorff topological
space. The following statements are equivalent:
1. X is normal;
2. for every closed set .A ⊂ X and every continuous function .f : A → R, there
exists a function .F : X → R such that F coincides with f on A. Furthermore, if
.|f (x)| < c for every .x ∈ A, then .|F (x)| < c for every .x ∈ X.
The function F is an extension of f , since coincides with f on the common domain.

Proof Suppose that 2. holds, and let A, B be disjoint closed subsets of X. The map
f : A ∪ B → R which sends A to a value .y0 and B to a value .y1 = y0 has an
.
extension to a continuous .F : X → R. If U and V are open neighborhoods of .y0

and .y1 respectively, then .F −1 (U ), .F −1 (V ) are disjoint open neighborhoods of A
and B, respectively.
Let us prove that 1. implies 2.
Step 1. .|f (x)| ≤ c for every .x ∈ A. We apply Lemma 13.4 with f in place of g,
and call .h0 : X → R the corresponding function. On A we thus have .|f − h0 | ≤
(2/3)c. We apply the same Lemma once more to the function .f − h0 on A, to
get .h1 : X → R such that
1 2
|h1 (x)| ≤
. · c x∈X
3 3
2 2
|f (x) − h0 (x) − h1 (x)| ≤ · c x ∈ A.
3 3
Now we proceed by induction, and assume that .h0 , .h1 , . . . , hn have been defined.
Lemma 13.4 applied to .g = f − h0 − · · · − hn on A yields .hn+1 : X → R such
that
n
1 2
|hn+1 (x)| ≤
. · c x∈X
3 3
n
2 2
|f (x) − h0 (x) − · · · − hn+1 (x)| ≤ · c x ∈ A.
3 3
have a function .hn : X → R for each .n ∈ N. The function F such that

We thus
F (x) = ∞
. n=0 hn (x) is continuous on X, .F (x) = f (x) for every .x ∈ A, and
∞
1 2 n
. |F (x)| ≤ c = c.
3 3
n=0
Step 2. .|f (x)| < c for every .x ∈ A. Indeed, the extension F constructed in Step
1 satisfies .|F (x)| ≤ c for every .x ∈ X. We set .A0 = {x ∈ X | |F (x)| = c}. This
set is closed in X and disjoint from A. Therefore there exists a Urysohn function
.ϕ : X → R having value 1 on A and value 0 on .A0 , with .0 ≤ ϕ ≤ 1 everywhere.
We define .G(x) = ϕ(x)F (x), a continuous function such that .G(x) = F (x) =
f (x) for .x ∈ A. Thus G extends f ; furthermore .|G(x)| < c for every .x ∈ X.
Indeed .G(x) = 0 if .x ∈ A0 , while .|ϕ(x)| ≤ 1 if .X ∈ X \ A0 , and .|F (x)| < c.
Step 3. f is not necessarily bounded. In this case we introduce the function
.h : R → (−1, 1) such that
x
h(x) =
. .
1 + |x|
By Step 2, the map .h◦f : A → (−1, 1) possesses an extension .F : X → (−1, 1)

and then .h−1 ◦ F is an extension of f , since .h−1 ◦ F : x ∈ A → h−1 ◦ h ◦ f (x) =
f (x). The proof is now complete.

The previous results require some refined separation property like normality or
regularity. For the purposes of Measure Theory, it will be useful to prove Urysohn’s
Lemma in a very particular environment. We propose here a proof based on the
on-point compactification. Later on we will propose a more concrete proof.
Let X be a locally compact Hausdorff (LCH) space. We denote by .X∗ its one-
point compactification, see Theorem 13.51.
Proposition 13.1 If K is a compact set and if U is an open set such that .K ⊂ U ⊂
X, then there exists an open set V with compact closure such that .K ⊂ V ⊂ V ⊂ U .
Proof As a compact subset of .X∗ , K is closed in .X∗ , while U is open in .X∗ . Since
∗
.X is a compact Hausdorff space, it is normal, and there exists an open subset V of
X∗ such that .K ⊂ V ⊂ V ⊂ U . The closure of V in .X∗ coincides with the closure

.
of V in X: indeed the former is a subset of X and the latter is equal to the former
intersected with X. Since .V is closed in the compact space .X∗ , it is also compact,
and since .V ⊂ X is open in .X∗ , it is also open in X. Thus V is the desired open
set.

Theorem 13.65 (Urysohn’s Lemma for LCH Spaces) If X is a locally compact
Hausdorff space, .K ⊂ U ⊂ X, K is compact and U is open, then there exists a
continuous function f on X such that .0 ≤ f ≤ 1, .f ≡ 1 on K, and .supp f ⊂ V .
Proof By the previous Proposition, we may choose an open set V with compact
closure such that .K ⊂ V ⊂ V ⊂ U , since K and .X∗ \ V are disjoint closed
subspaces of the normal space X. By Theorem 13.60, there is a continuous function
g on X such that .0 ≤ g ≤ 1, .g ≡ 1 on K, and .g ≡ 0 on .X∗ \ V . We define f as the
restriction of g to X. Clearly f is continuous on X, .0 ≤ f ≤ 1, and .f ≡ 1 on K.
Since g vanishes outside .V , so does f , and this implies
. supp f = {x ∈ X | f (x) = 0} ⊂ V ⊂ U,
since .V is closed. The proof is complete.

13.15 Partitions of Unity and Paracompact Spaces
One of the most important tools of Mathematical Analysis is the possibility of gluing
together functions with compact support. As a rough idea, several constructions
like Analysis on Manifolds or Partial Differential Equations proceed from local to
global: around every point we are able to construct something, and the we would
like to extend such a construction to the whole space.
If we work in a compact setting, it is intuitive that a compact space behaves much
like a finite space, in the sense that every open cover may be assumed to be finite
from the beginning. However this is only the most favorable case, and quite often
compactness may not be assumed. In this section we introduce the definition of
paracompactness, originally due to J. Dieudonné. We will show that it generalizes
the definition of compactness and that it allows us to define partitions of unity: a
collection of continuous, compactly supported functions whose sum equals one at
any point.
Definition 13.70 A collection .A of subsets of a topological space X is locally finite
in X if and only if every point of X possesses a neighborhood which intersects only
finitely many elements of .A.
Exercise 13.44 In .R with the Euclidean topology, prove that .A = {(n, n + 2) | n
∈ Z} is locally finite.
13.15 Partitions of Unity and Paracompact Spaces 263
Exercise 13.45 Prove that .A = {(0, 1/n) | n = 1, 2, 3, . . .} is locally finite in

(0, 1) (with the induced topology), but not in .R with the Euclidean topology.
.
Definition 13.71 Let .A be a collection of subsets of a topological space X. A

collection .B of subsets of X is a refinement of .A if and only if for every element B
of .B there exists an element A of .A such that .B ⊂ A. If all the elements of .B are
open sets, we say that .B is an open refinement of .A.
Definition 13.72 A topological space X is paracompact if and only if every open
cover .A of X possesses a locally finite open refinement .B which covers X.
Theorem 13.66 The space .Rn with the Euclidean topology is paracompact.
Proof Consider an open cover .A of .Rn . We begin with .B0 = ∅ and for every
positive integer m we call .Bm the open ball of radius m centered at the origin. Since
.Bm is compact, we choose finitely many elements of .A which cover .Bm and we
∞each of them with .R \ Bm−1 . We call such a finite collection .Cm . Clearly
intersect n
.C = m=1 Cm is an open refinement of .A. It is evidently locally finite, since the

open ball .Bm intersects only finitely many elements of .C, i.e. those elements which
belong to .C1 ∪ · · · ∪ Cm . Finally, for every .x ∈ Rn we select the smallest positive
integer m such that .x ∈ Bm . It follows that x belongs to an element of .Cm , and
therefore .C covers .Rn . The proof is complete.

Theorem 13.67 Every closed subspace of a paracompact space is paracompact.
Proof Let Y be a closed subspace of a paracompact space X, and let .A be an open
(relative to the induced topology) cover of Y . To each .A ∈ A we associate an open
set .A1 of X such that .A1 ∩ Y = A. We now cover X by the open sets .A1 and by
.X \ Y . By assumption there exists a locally finite open refinement .B which covers
X. Hence the collection .C = {B ∩ Y | B ∈ B} is the required locally finite open

refinement of .A.

Theorem 13.68 Every paracompact Hausdorff space X is normal.
Proof Consider a point a of X and a closed set B which does not contain a. For
every .b ∈ B we choose an open neighborhood .Ub of b such that .Ub does not contain
a. Putting together these sets .Ub and .X\B we cover X, so that we may take a locally
finite refinement .C which covers X. We consider the collection .D of those elements
of .C which intersect B. Then .D covers X, and the closure .D of every .D ∈ D is
disjoint from a. Since D intersects B, D must lie in some set .Ub whose closure is
disjoint from a. Let

V =
. {D | D ∈ D} ;
then V is open and contains B. Since .D is locally finite,

V =
. DD∈D ,
and .V is disjoint from a. We have proved that X is a regular space. To prove

normality, we repeat the previous argument, replacing a by a closed set A and using
regularity instead of the Hausdorff property.

Although we have proved that .Rn is paracompact by hands, it is a particular case of
a more general result. The proof is rather difficult.
Theorem 13.69 (A. H. Stone) Every metric space is paracompact.
Proof Let X be a metric space. We say that a collection .A of subsets of X is
countably locally finite if and only if .A is a countable union of collections .An ,
.n ∈ N, each of which is locally finite.
We claim that if .A is any open cover of X, there exists an open cover .E of X

which is a refinement of .A and which is countably locally finite.
Indeed, by the Well-ordering principle we may choose a well ordering .< for the
collection .A. Consider a positive integer n. If U belongs to .A, we define
Sn (U ) = {x ∈ X | B(x, 1/n) ⊂ U } .
.
In order to exploit the well-ordering property of .<, we introduce

Tn (U ) = Sn (U ) \
. {V | V < U } .
If V and W are distinct elements of .A, then .d(x, y) ≥ 1/n for every .x ∈ Tn (V )
and every .y ∈ Tn (W ). Indeed, since .< is a well-ordering for .A, we may assume
that .V < W . Since .x ∈ Tn (V ), we have .x ∈ Sn (V ), so the open ball centered at x
with radius .1/n lies in V . On the other hand, since .y ∈ Tn (W ) and .V < W , we see
that .y ∈/ V . Hence y does not lie in the open ball of center x and radius .1/n.
Since there is no reason why the sets .Tn (U ) should be open, we enlarge them a
little bit, and we set

1 1

.En (U ) = x ∈ R d(x, Tn (U )) < = B x, x ∈ Tn (U ) ,
3n 3n
where .d(x, Tn (U )) = inf {d(x, z) | z ∈ Tn (U )}. It is a simple exercise to show that

En (U ) and .En (V ) are disjoint provided that U and V are distinct elements of .A.
.
Finally, we set
En = {En (U ) | U ∈ A} .
.
The collection .En is a locally finite collection of open sets which refines .A. Indeed,
En (V ) ⊂ V for every .V ∈ A. Furthermore, for every .x ∈ X, the open ball of center
.
x and radius .1/(6n) intersects at most one element of .En . To obtained the desired
countably locally finite refinement of .A, we define
∞

E=
. En .
n=1
We only need to check that .E covers X. If x is any point of X, let U be the smallest
element of .A which contains x, with respect to the well-order .<. Since U is an
open set, there exists a positive integer n such that .B(x, 1/n) ⊂ U . By definition
.x ∈ Sn (U ). It follows that .x ∈ Tn (U ), since U is the smallest element of .A which
contains x, and thus .x ∈ En (U ) ∈ En , and the claim is proved.

The conclusion follows now from Theorem 13.70.

Theorem 13.70 Suppose X is a regular topological space. The following condi-
tions are equivalent: Every open cover of X has a refinement which is
(a) an open cover of X and countably locally finite;
(b) a cover of X and locally finite;
(c) a closed cover of X and locally finite;
(d) an open cover of X and locally finite.
Proof Since (d) implies (a) in a trivial way, we will prove that (a) implies (b) implies
(c) implies (d).
(a) implies (b) Consider any open cover .A of X, and let .B be an open refinement
∞ covers X and is countably locally finite. We may suppose that .B =
which
n=1 Bn , where each .Bn is locally finite. For every index i we define .Vi =
{U | U ∈ Bi }. For every positive integer n and every .U ∈ B)n, define

Sn (U ) = U \
. Vi .
i<n
The collection .Cn = {Sn (U ) |U ∈ Bn } is a refinement of .Bn , since .U ∈ Bn

implies .Sn (U ) ⊂ U . If .C = ∞ n=1 Cn , we claim that .C is the desired locally
finite refinement of .A. To prove that .C covers X, we pick .x ∈ X and we select
the smallest positive integer N such that x belongs to some element U of .BN .
Since N is the smallest positive integer with such a property, .x ∈ SN (U ) ∈ C.
Furthermore, for every .n = 1, . . . , N we can select a neighborhood .Wn of x
which intersects only finitely many elements of .Bn . If .Wn ∩ Sn (V ) = ∅, then
.Wn must intersect any element V of .Bn . As a consequence .Wn intersects only
finitely many elements of .Cn . But U is in .BN , hence U does not intersect any
element of .Cn for .n > N. To summarize, the neighborhood
W1 ∩ W2 ∩ · · · ∩ WN ∩ U
.
of the point x intersects only finitely many elements of .Cn .

(b) implies (c) Let .A be an open cover of X, and let .B be the family of all open
sets .U ⊂ X such that .U is contained in some element of .A. Since X is a regular
space, .B is an open cover of X. By assumption (b), there exists a refinement
.C
of .B which covers X and is also locally finite. Setting .D = C C ∈ C , we see

that .D is a cover of X which refines .A and is locally finite.
(c) implies (d) This proof is slightly longer. We begin with an open cover .A of
X and we use (c) to construct a refinement .B which covers X and is locally
finite. The trick here is to “enlarge” the elements of .B in order to get open sets.
The geometric intuition is somewhat misleading, since we do not have a distance
function to play with. The proof is subtle.
So, for every .x ∈ X there exists an open neighborhood of X which intersects
only finitely many elements of .B. The union of these neighborhoods is therefore
an open cover of X. By assumption (c), we may define a closed refinement .C of
such an open cover which is still a cover of X and is also locally finite. Clearly,
every element of .C can intersect only finitely many elements of .B.
Now, for every .B ∈ B, we define
C(B) = {C | C ∈ C ∧ C ⊂ X \ B} ,
.
and

E(B) = X \
. {C | C ∈ C(B)} .
The set .E(B) is an open subset of X.12 By definition, .B ⊂ E(B). The collection
of all sets .E(B) need not be a refinement of .A, but we can always choose
some .F (B) ∈ A which contains B. Then .D = {E(B) ∩ F (B) | B ∈ B} is a
refinements of .A which covers X, since .B ⊂ E(B) ∩ F (B).
To conclude, we still have to prove that .D is locally finite. Pick a point .x ∈ X,
and choose an open neighborhood W of x which intersects only finitely many
elements .C1 , . . . , Ck of .C. Since .C is a cover of X, .W ⊂ ki=1 Ci . To prove
that W intersects only finitely many elements of .D, it is now sufficient to prove
that every .C ∈ C intersects only finitely many elements of .D. Suppose that C
intersects .E(B) ∩ F (B). Then C is not contained in .X \ B, which means that
.C ∩ B = ∅. But C intersects only finitely many elements of .B, hence it can
intersect no more than the same number of elements of .D. The proof is complete.

Proof of Theorem 13.69, Different Flavor Consider an open cover .{Ui | i ∈ I } of
the metric space X. As we saw above, we may suppose that the index set I is well-
ordered by .≤.Inparticular,
every .x ∈ X there exists a unique index .i ∈ I such
for
that .x ∈ Ui \ Uj j < i . Explicitly,

i = min j ∈ I x ∈ Uj .
.
For every .i ∈ I and every positive integer n, we define
12 Here
the fact that .C is a locally finite collection of closed sets is essential to ensure that
. {C | C ∈ C(B)} is a closed set. Recall that arbitrary unions of closed sets are not a closed set, in
general.

Vi,n =
. B(x, 2−n ) x ∈ Xi,n

/ Uj j < i ,
B(x, 3 · 2−n ) ⊂ Ui , x ∈
Xi,n = x ∈ X .
x∈/ Vj,k j ∈ I, k < n

Our claim is that . Vi,n i ∈ I, n ∈ N is a locally finite open refinement of
.{Ui | i ∈ I }. Of course each element of . Vi,n i ∈ I, n ∈ N is open. By definition,
−n ) ⊂ B(x, 3·2−n ) ⊂ U for every ball contributing to .V , hence .V
.B(x, 2
i i,n i,n ⊂ Ui :
this proves that . Vi,n i ∈ I, n ∈ N is an open refinement of .{Ui | i ∈ I }. To
complete the proof, we must show that thisis a locally
finite refinement.
Pickany .x ∈ X, fix .i = min j ∈ I x ∈ Uj , and select .n ∈ N such that

.Vi,n = B(x, 2−n ) x ∈ Xi,n . Two cases are possible:
either .x ∈ Vj,k
for some
.j ∈ I and .k < n, or .x ∈ Xi,n ⊂ Vi,n . In any case, . Vi,n i ∈ I, n ∈ N is a cover
of X.
For every .x ∈ X, define .i = min j ∈ I x ∈ Vj,n n ∈ N . We can choose
positive integers k and n such that .B(x, 2−k ) ⊂ Vi,n . We are going to prove the
following statements:
(i) if . ≥ n + k, then .B(x, 2−n−k ) does not intersect any .Vj, ;
(ii) if . < n + k, then .B(x, 2−n−k ) intersects .Vj, for at most one index .j ∈ I .
These claims imply that the open neighborhood .B(x, 2
−n−k ) can meet at most

n + k − 1 elements of . Vi,n i ∈ I, n ∈ N , and the latter is then a locally finite
.
refinement.
To prove (i), we pick .y ∈ Xj, . Since . > n, .y ∈/ Vi,n . But .B(x, 2−k ) ⊂ Vi,n ,
hence .d(x, y) ≥ 2−k . Since . ≥ k + 1 and .n + k ≥ k + 1, the condition
z ∈ B(x, 2−n−k ) ∩ B(y, 2− )

.
would imply
d(x, y) ≤ d(x, z) + d(z, y) ≤ 2−n−k + 2− ≤ 2−k−1 + 2−k−1 = 2−k ,

.
which is a contradiction. Thus .B(x, 2−n−k ) is disjoint from each ball .B(y, 2− ),
.y ∈ Xj, . Since these balls cover .Vj, , (i) is proved.
Let us now turn to (ii). For .i < j we pick .x ∈ Vi, , .y ∈ Vj, . So there exist points
% %
.x and .y in X such that
.x ∈ B(x % , 2− ) ⊂ Vi, , y ∈ B(y % , 2− ) ⊂ Vj, .
Hence .B(x % , 3 · 2− ) ⊂ Ui , but .y % ∈

/ Ui . Therefore .d(x % , y % ) ≥ 3 · 2− , and the
triangle inequality yields
3 · 2− ≤ d(x % , y % ) ≤ d(x % , x) + d(x, y) + d(y, y % ) ≤ d(x, y) + 2− + 2− ,

.
or .d(x, y) > 2− . As a trivial consequence, .Vi, ∩ Vj, = ∅ whenever .i = j ,

and every ball of radius .2−−1 intersects .Vi, for at most one index .i ∈ I . Since
. < n + k, i.e. .n + k ≥ + 1, the same conclusion holds for every ball .B(x, 2
−n−k ).

Important: About A. H. Stone’s Theorem

In its astonishing simplicity, Theorem 13.69 remains one of the most important
result of Topology and also of Abstract Analysis, and was obtained in [9]. The first
proof we have presented follows Munkres [6], and is based on the Well-ordering
Principle. The second proof is essentially due to [7], with the only difference that
M. E. Rudin began her proof by indexing the open cover .{Ui } over the ordinals.
Since the ordinals are well-ordered by definition, Rudin actually uses the Axiom of
Choice in place of the Well-ordering Principle.
The main application of paracompactness in Analysis is the construction of

general partitions of unity. In the chapter on Measure Theory we will need a
convenient version in locally compact spaces, which will be easier since local
compactness essentially reduces the proof to a finite case.
Definition 13.73 Let .{Uα | α ∈ J } be an open cover of a topological space X.13 A
collection of functions .φα : X → [0, 1], .α ∈ J , is a partition of unity dominated by
.{Uα | α ∈ J } if
(1) .supp φα ⊂ Uα for every .α ∈ J ;

collection .{supp φα | α ∈ J } is locally finite;
(2) the
(3) . α∈J φα (x) = 1 for every .x ∈ X.14
To prove the existence of partitions of unity in paracompact spaces, we begin with
a sort of “shrinking” result.
Lemma 13.5 Suppose that X is a paracompact Hausdorff space and .{Uα | α ∈ J }
is an open cover of X. Then there exists a locally finite open cover .{Vα | α ∈ J } of
X such that .Vα ⊂ Uα for every .α ∈ J .15
Proof Let us define

.A = A A is open and A is contained in .
some element of {Uα | α ∈ J }
13 We prefer here the use of indices to label the elements of the cover, since an intrinsic notation
would hide the correspondence between the index of the cover and the index of the function in the
partition of unity.
14 The sum is indeed a finite sum by condition (2), in the sense that only finitely many terms are
different than zero.

15 Sometimes .{V | α ∈ J } is called a precise refinement of .{U | α ∈ J }.
α α
Since X is regular, see Theorem 13.68, .A is a cover of X. By paracompactness

we can finite refinement .B of open sets which cover X. We write
find a locally
.B = Bβ β ∈ K . Recalling that .B is a refinement of .A, we define a function
.f : K → J by choosing, for every .β ∈ K, and element .f (β) ∈ J such that
.Bβ ⊂ Uf (β) .
16 For every .α ∈ J , let .V be the union of all the elements of
α

Bα = Bβ f (β) = α .
.
Of course we set .Vα = ∅ if the condition .f (β) = α cannot be satisfied.

By definition, for every .Bβ in .Bα we have .Bβ ⊂ Uα . Since .Bα is locally finite,
.Vα coincides with the closure of the union of the elements of .Bα , hence .Vα ⊂ Uα .
To conclude, for every point .x ∈ X we choose a neighborhood W of x which

intersects only finitely many elements .Bβ1 , . . . , Bβk . Then W intersects .Vα only if
.α belongs to the finite set .{f (β1 ), . . . , f (βk )}, and therefore .{Vα | α ∈ J } is locally
finite.

Theorem 13.71 Suppose that X is a paracompact Hausdorff space. If .{Uα | α ∈ J }
is an open cover of X, then there exists a partition of unity dominated by
.{Uα | α ∈ J }.
Proof Invoking the previous Lemma twice, we construct locally finite open covers
{Wα | α ∈ J } and .{Vα | α ∈ J } such that
.
Wα ⊂ Vα ,
. V α ⊂ Uα
for every .α ∈ J . By normality, to every .α ∈ J we may attach a continuous function

.ψα : X → [0, 1] such that .ψα (Wα ) = {1} and .ψα (X \ Vα ) = {0}. Then the support
of .ψα is contained in .Vα , and hence in .Uα . The collection .{Vα | α ∈ J } is locally
finite, the collection .{supp ψα | α ∈ J } is also locally finite. Furthermore, for any
.x ∈ X at least one of the functions .ψα must be positive at x, since .{Wα | α ∈ J }
covers X.
Hence, for every .x ∈ X we can form the sum

(x) =
. ψα (x),
α∈J
since x has a neighborhood which intersects only finitely many of the sets .supp ψα .
Being a finite sum of continuous functions, . is a continuous function at any point
of X. If we set
ψα (x)
φα (x) =
. ,
(x)
16 Here we are using the Axiom of Choice.

it is immediate to check that .{φα | α ∈ J } is the desired partition of unity dominated

by .{Uα | α ∈ J }. The proof is complete.

Partitions of unity are typically used to construct continuous17 functions which
enjoy additional requirements. Here is a nice example.
Theorem 13.72 Let X be a paracompact space, and let .g : X → R be a lower
semicontinuous function, .h : X → R be an upper semicontinuous function. If
.h(x) < g(x) for every .x ∈ X, there exists a continuous function .p : X → R
such that .h(x) < p(x) < g(x) for every .x ∈ X.

Proof Let
)

.U = Uα Uα is open and sup h < inf g .
Uα Uα
We already know that there exists a partition of unity .{φα | α ∈ J } dominated by .U.
For each .α ∈ J and each .x ∈ X such that .ϕα (x) = 0, we choose a number .kα with
.h(x) < kα < g(x). The function .p : X → R defined by

p(x) =
. kα ϕα (x)
α∈J
is continuous and clearly satisfies .h(x) < p(x) < g(x) at any .x ∈ X.

Some constructions in mathematical analysis and also in differential geometry
are based on a common idea: one can “fill up” the Euclidean space .Rn by a sequence
.{Kn }n of compact sets such that .Kn ⊂ K
◦
n+1 , i.e. each .Kn is continued in the interior
of .Kn+1 . This is a trivial fact, since one can consider .Kn = {x ∈ Rn | *x* ≤ n}. We
now show that the same idea also works in more general situations.
Definition 13.74 (Exhaustion by Compact Sets) An exhaustion by compact sets
in a locally compact Hausdorff space X consists of a sequence .{Kn }n of compact
sets of X such that
∞

◦
Kn ⊂ Kn+1
. and X = Kn .
n=1
Theorem 13.73 Suppose that X is a locally compact Hausdorff space.

1. If the topology of X has a countable base, then X has an exhaustion by compact
sets.
2. If X is paracompact and connected, then X has an exhaustion by compact sets.
3. If X has an exhaustion by compact sets, then X is paracompact.
17 In many situations, continuity in not sufficient, and smooth partitions of unity must be used.
This happens in Differential Geometry, but also in the theory of Function Spaces. Smoothness is
ensured by convolution with suitable kernels.
Proof We choose a countable base .{Bn }n of X, such that each .Bn is compact. We
define .K1 = B1 and we assume that .K2 , . . . , Kn have been chosen. Let .m > n be
the smallest integer such that .Kn ⊂ B1 ∪ · · · ∪ Bm , and let
Kn+1 = B1 ∪ · · · ∪ Bm .
.
It is easy to check that .{Kn }n is an exhaustion of X by compact sets. This proves 1.

To prove 2, we choose a locally finite open cover .{Ui | i ∈ I } of X, such that
each .Ui is compact. Then every compact subset of X intersects only finitely many
of the .Ui . Choose .i(1) ∈ I such that .Ui(1) = ∅ and define .K1 = Ui(1) . Then .K1 is
compact and therefore intersects only finitely many of the .Ui . We now enumerate
these, i.e.
.K1 ∩ Ui(j ) = ∅ for i(1), . . . , i(j2 ) ∈ I.
We define .K2 = Ui(1) ∪ · · · ∪ Ui(j2 ) . As before, .K2 is compact and intersects finitely
many of the .Ui , i.e.
.K2 ∩ Ui(j ) = ∅ for i(1), . . . , i(j3 ) ∈ I.
We then define .K3 = Ui(1) ∪ · · · ∪ Ui(j3 ) . Recursively, we obtain a sequence of

compact sets .Kn such that
◦
Kn ⊂ Ui(1) ∪ · · · ∪ Ui(n+1) = Kn+1
. .
∞
In particular, .X = n=1 Kn is open in X. Let X be an accumulation point
of . ∞ n=1 Kn . A compact neighborhood of x in X intersects only finitely many
of the .Ui(j ) , so x lies in the finite union of the closures of these .Ui(j ) . This
union is contained in one of the .Kn , and so is x. We have proved that . ∞ n=1 Kn
contains
∞ its accumulation points, hence
∞ it is closed in X. Since X is connected and
. n=1 Kn = ∅, it follows that .X = n=1 Kn .
To prove 3, we fix an exhaustion .{Kn }n of X by compact sets. Let .U =
{Ui | i ∈ I } be an open cover of X. Each .Kn \ Kn+1 ◦ is compact, hence there exist

◦
finite open covers . Vnj 1 ≤ j ≤ kn of .Kn \ Kn+1 with the properties that all .Vnj
◦
are contained in18 .Kn+1 \ Kn+2 and for every .Vnj there exists a .Uj with .Vnj ⊂ Uj .
The collection of all the .Vnj is then an open cover of X, it is locally finite, and it is a
refinement of the cover .U. Hence X is paracompact, and the proof is complete.
18 We define here .K−1 = K0 = ∅.

13.16 Function Spaces
One of the most important applications of General Topology to Mathematical

Analysis concerns spaces of maps, i.e. topological spaces whose elements are
functions.
Definition 13.75 Let X, Y be sets. The set
= Map(X, Y )
.
is the set of all continuous functions .f : X → Y .

We want to topologize ., and this can be done in several ways. First of all, let us
recall that the set .Y X of all functions .f : X → Y (continuous or not) has a natural
topology as a product space as soon as Y is a topological space.
Definition 13.76 Suppose that Y is a topological space. The topology induced on
. by the product topology of .Y X is called the topology of pointwise convergence in
..
Exercise 13.46 For every .x ∈ X and every open set .U ⊂ Y , define

E(x, U ) = f ∈ Y X f (x) ∈ U
.
M(x, U ) = {f ∈ Map(X, Y ) | f (x) ∈ U } .
1. Prove that the sets .E(x, U ), as .x ∈ X and .U ⊂ Y is open, form a subbasis of the
topology of .Y X .
2. Since .M(x, U ) = ∩ E(x, U ), deduce that .{M(x, U ) | x ∈ X, U ⊂ Y is open}
is a subbasis of the pointwise topology of ..
Proposition 13.2 For each finite set .F ⊂ X and every open set .U ⊂ Y , let

M(F, U ) = {f ∈ Map(X, Y ) | f (F ) ⊂ U } =
. {M(x, U ) | x ∈ F } .
The set .M(F, U ) is open in the topology of pointwise convergence in .. In

particular, the collection of the sets .M(F, U ) as F ranges of the finite subsets of
X and U ranges over the open sets in Y is a subbasis of the topology of pointwise
convergence in ..
Proof The proof is immediate: indeed, .M(F, U ) is open in the topology of
pointwise convergence as a finite intersection of open sets of the form .M(x, U ),
for .x ∈ F .

The topology of pointwise convergence is often referred to as the finite-open
topology, as the previous result suggests.
13.16 Function Spaces 273
The basic weakness of the finite-open topology on . is clearly the fact that
the construction is independent of any topological structure of the set X. This
topology typically contains very few open sets, and few open sets correspond to
few continuous maps.
Definition 13.77 Let X and Y be topological spaces. For any compact .K ⊂ X and
any open .U ⊂ Y , we define
M(K, U ) = {f ∈ | f (K) ⊂ U } .
.
The compact-open topology of . is the smallest topology on . which contains all

sets of the form .M(K, U ). In particular, a set is open in the compact-open topology
if and only if it is the union of a collection of finite intersections of sets .M(K, U ).
We now want to investigate some topologies .τ on . = Map(X, Y ). We begin
with the natural evaluation map
ω : (, τ ) × X → Y
.
ω(f, x) = f (x).
Definition 13.78 A topology .τ on . is admissible if and only if the map .ω is

continuous.
We cannot expect the compact-open topology to be always admissible. The next
results show, however, that this is correct under some regularity assumption on the
space.
Definition 13.79 A Hausdorff space X is regular if each .x ∈ X each closed set A
not containing x have disjoint neighborhoods. More precisely, if .A ⊂ X is closed
and if .x ∈
/ A, there exist open sets U containing x and .V ⊃ A such that .U ∩ V = ∅.
In a metric space, we can always shrink neighborhoods by playing with their radii.
The category of regular spaces is precisely the category of topological spaces in
which the same trick is possible.
Proposition 13.3 The following statements are equivalent:
1. the topological space X is regular;
2. for each .x ∈ X and for each neighborhood U of x, there exists a neighborhood
V of x such that .x ∈ V ⊂ V ⊂ U ;
3. for each .x ∈ X and for each closed A not containing x, there exists a
neighborhood V of x such that .V ∩ A = ∅.
Proof Given U , by definition x and .X \ U have disjoint neighborhoods V and W .
Thus .V ⊂ X \ W , so that .V ⊂ X \ W . From .V ∩ W ⊂ V ∩ W = ∅ we deduce
.V ⊂ U . Hence 1. implies 2.
Using now x and its neighborhood .X \ A, we find an open set V such that .x ∈
V ⊂ V ⊂ X \ A. Hence .V ∩ A = ∅, and 2. implies 3.
Let A be closed, .x ∈
/ A. Pick a neighborhood V of x such that .V ∩ A = ∅. Then
A ⊂ X \ V , and .V ∩ (X \ V ) = ∅. This proves that 3. implies 1.
.

Theorem 13.74 (Arens) If X is a locally compact regular space, then the compact-
open topology on . is admissible. More precisely, it is the smallest of all admissible
topologies on ..
Proof Pick .f ∈ , .x ∈ X, and an open set W of Y containing .f (x). Since f is
continuous, .f −1 (W ) is an open set in X which contains x. Since X is a regular
locally compact space, there exists an open neighborhood V of x such that .V is
compact in X and is contained in .f −1 (W ). Then .U = M(V , W ) is an element of
the subbasis of the compact-open topology of . which contains f . It follows that
the evaluation .ω sends .U × V into W , so that the compact-open topology of . is
admissible.
To complete the proof, we will show that the compact-open topology of . is
smaller than any other admissible topology .τ on . without using our assumptions
on the space X. Pick any set of the form .M(K, W ), where K is compact in X and W
is open in Y , and pick an element .f ∈ M(K, W ). Since the topology .τ is admissible,
the evaluation map .ω : (, τ ) × X → Y is continuous. Hence, for every .x ∈ K,
there exist an open neighborhood .Vx of x in X and an open neighborhood .Ux of f
in .(, τ ) such that .ω (Ux × Vx ) ⊂ W . Since K is compact, there are finitely many
points .x1 , . . . , xn such that
K ⊂ Vx 1 ∪ · · · ∪ Vx n .
.
Let
U = Ux 1 ∩ · · · ∩ Ux n .
.
Then U is an open neighborhood of f in .(, τ ). We claim that .U ⊂ M(K, W ).

Indeed, let .g ∈ U . We have .g(K) = ω(g × K) ⊂ W . This implies that .g ∈
M(K, W ) and hence that .U ⊂ M(K, W ). We have proved that .M(K, W ) is open
in .(, τ ). It follows that the compact-open topology is smaller than .τ , and the proof
is complete.

Corollary 13.2 The topology of pointwise convergence is not admissible if it is
different than the compact-open topology of ..
Proof This follows at once from the fact that the properties of X are not used in the
proof of the second statement of Theorem 13.74.

We have seen in a previous chapter of this book that the pointwise convergence
of a sequence of continuous functions (of a real variable) is typically insufficient
to ensure good properties of the limit function. On the contrary, continuity, integra-
bility and (to some extent) differentiability are stable under uniform convergence.
We want to describe now a topology on a space of maps which describes uniform
convergence of sequences.
Although a natural setting for this purpose would be the theory of uniform
structures, we believe that a simplified setting is more than enough to the Analyst’s
eye. So we will consider a metric space Y . The following result brings in another
useful property.
Proposition 13.4 Every metric space is homeomorphic to a bounded metric space.
Proof Let .(Y, d) be a metric space. We define .d ∗ : Y × Y → [0, +∞) such that
d(x, y)
d ∗ (x, y) =
. , (x, y) ∈ Y × Y.
1 + d(x, y)
It is easy to check that .d ∗ is a metric on Y . Indeed, .d ∗ (x, y) = 0 means .d(x, y) = 0.

Moreover,
d(x, y) d(x, z)
d ∗ (x, y) + d ∗ (y − z) ≥
. +
1 + d(x, y) 1 + d(x, y) + d(x, z)
d(x, y) + d(x, z) 1
= =
1 + d(x, y) + d(x, z) 1+ 1
d(x,y)+d(x,z)
1
≥ = d ∗ (y, z).
1+ 1
d(y,z)
Since .t → t/(1 + t) is a homeomorphism of .[0, +∞) onto itself, the spaces .(Y, d)
and .(Y, d ∗ ) are homeomorphic. Finally, the fact that .d ∗ (x, y) ≤ 1 for every .x ∈ Y ,
∗
.y ∈ Y shows that .(Y, d ) is a bounded metric space.

In virtue of the previous Proposition we may assume that .(Y, d) is a bounded
metric space.
Definition 13.80 Consider . = Map(X, Y ). The function .d ∗ : × → [0, +∞)
such that
d ∗ (f, g) = sup {d(f (x), f (y)) | x ∈ X}

.
is a metric on . which induces a topology called the topology .u of uniform

convergence.
Exercise 13.47 Prove that .d ∗ is a metric on ..
Theorem 13.75 The topology .u is admissible.
Proof We need to prove that the evaluation .ω : (, u) × X → Y is continuous. Let

f0 ∈ and .x0 ∈ X be given, and pick any .δ > 0. We denote .y0 = f (x0 ). Since .f0 is
.
continuous, there exists a neighborhood V of .x0 in X such that .d(y0 , f0 (x)) < δ/2
for every .x ∈ V . Let

δ
U = f ∈ d ∗ (f, f0 ) <
. .
2
There results
d(ω(f0 , x0 ), ω(f, x)) = d(f0 (x), f (x))

.
≤ d(f0 (x0 ), f0 (x)) + d(f0 (x), f (x))

≤ d(y0 , f0 (x)) + d ∗ (f0 , f ) < δ
for every .f ∈ U and every .x ∈ V . This proves the continuity of .ω at .(f0 , x0 ).

A natural question is whether the topology of uniform convergence coincides with
the compact-open topology.
Theorem 13.76 If X is a compact space, the compact-open topology and the
topology of uniform convergence coincide on ..
Proof By Theorem 13.75, every open set for the compact-open topology is also
open for the topology of uniform convergence. Conversely, we need to prove that,
for any .f ∈ and any .δ > 0, there exists an open set V for the compact-open
topology such that

f ∈ V ⊂ U = g ∈ d ∗ (f, g) < δ .
.
For .x ∈ X, we denote

δ
Wx = y ∈ Y d(f (x), y) <
. .
2
Since f is continuous and .f (x) ∈ Wx , there exists an open neighborhood .Gx of x

in X such that .f (Kx ) ⊂ Wx , where .Kx = Gx . Now the compactness of X comes
into play: there exist finitely many points .x1 , . . . , xn in X such that
X = Gx 1 ∪ · · · ∪ Gx n .
.
Each set .Kxi , .i = 1, . . . , n, is compact as a closed subset of a compact space. Hence
V = M(Kx1 , Wx1 ) ∩ · · · ∩ M(Kxn , Wxn )

.
is an open set of the compact-open topology of ., and .f ∈ V . It remains to prove

that .V ⊂ U . Let .g ∈ V and .x ∈ X. For some .i = 1, . . . , n there results .x ∈ Gxi ⊂
Kx1 . This implies .f (xi ) ∈ Wxi and .g(x) ∈ Wxi . Then
δ δ
d(f (xi ), f (x)) <
. , d(f (xi ), g(x)) < .
3 3
We obtain
2
d(f (x), g(x)) ≤ d(f (x), f (xi )) + d(f (xi ), g(x)) <
. δ.
3
Since .x ∈ X is arbitrary, .d ∗ (f, g) ≤ (2/3)δ < δ, and thus .g ∈ U . The proof is

complete.

Remark 13.19 We point out that the boundedness of .(Y, d) is irrelevant in Theorem
13.76. Indeed the boundedness of .d ∗ is ensured by the compactness of X.
As we have already seen, compactness is probably the most useful property
of a topological space for applications to Analysis. We want to introduce now a
celebrated compactness condition in the space of continuous functions.
We begin with a general property of compact products, and with some definitions
which are typical of metric spaces.
Theorem 13.77 (Wallace) Let X, Y be topological spaces, .A ⊂ X, .B ⊂ Y be
compact subspaces and .W ⊂ X × Y be an open set such that .A × B ⊂ W . Then
there exist open sets .U ⊂ X, .V ⊂ Y such that .A ⊂ U , .B ⊂ V , and .U × V ⊂ W .
Proof We first prove the theorem in the particular case .A = {a}. For any point
b ∈ B there exist two open sets .Ub ⊂ X, .Vb ⊂ Y , such that .(a, b) ∈ (Ub , Vb ) ⊂ W .
.
The collection of open sets .{Vb | b ∈ B} covers B, and by compactness there are
points .b1 , . . . , bn ∈ B such that
B ⊂ Vb 1 ∪ · · · ∪ Vb n .
.
Then

n
{a} × B ⊂ U × V ⊂
. Ubi × Vbi ⊂ W.
i=1
We now prove the statement in the general case. We have just proved that for every
.a ∈ A there exist open sets .Ua ⊂ X, .Va ⊂ Y , such that .{a} × B ⊂ Ua × Va ⊂ W .
Again, the collection .{Ua | a ∈ A} of open sets covers A, and by compactness
.A ⊂ Ua1 ∪ · · · Uan for some points .a1 , . . . , an ∈ A.
To conclude, the open sets .U = Ua1 ∪ · · · ∪ Uan and .V = Va1 ∩ · · · ∩ Van satisfy
A × B ⊂ U × V ⊂ W.
.

Definition 13.81 A metric space is totally bounded if it can be covered by a finite
number of open balls of radius r, for every .r > 0.
Part of the following result was proved in a more general setting. We present a
version that contains a new characterization of compactness in metric spaces.
Theorem 13.78 If K is a closed subset of a complete metric space .(X, d), the
following statements are equivalent:
(a) K is compact;
(b) every infinite subset of K has an accumulation point;
(c) K is totally bounded.
Proof Suppose (a). If E is an infinite subset of K and no point of K is an

accumulation point of E, there is an open cover .{Vα | α ∈ I } of K such that each
.Vα contains at most one point of E. Therefore .{Vα | α ∈ I } has no finite subcover,
contradicting the compactness of K.

Suppose (b). Let .ε > 0. Pick .x1 ∈ K at random. Suppose .x1 , . . . , xn have been
chosen in K with the condition that .d(xi , xj ) ≥ ε if .i = j . If possible, choose
.xn+1 ∈ K so that .d(xi , xn+1 ) ≥ ε for .1 ≤ i ≤ n. This process must stop after a
finite number of steps, because of assumption (b). Thus the open balls with radius .ε
centered at .x1 , . . . , xn cover K, and K is totally bounded.
Suppose now (c). Consider an open cover . of K which has no finite subcover.By
assumption (c), K is a union of finitely many closed sets of diameters less than or
equal to one. One of these, say .K1 , cannot be covered by finitely many members of
. . Now we repeat this scheme with .K1 instead of K, and continue. The result is a
sequence of closed set .Ki such that

(i) .K ⊃ K1 ⊃ K2 ⊃ . . .
(ii) .diam(Kn ) = sup {d(x, y) | x ∈ Kn , y ∈ Kn } ≤ 1/n for .n = 1, 2, . . .
(iii) no .Kn can be covered by finitely many members of . .
Let .xn ∈ Kn . By (i) and (ii), the sequence .{xn }n is
a Cauchy sequence in the
complete metric space X, so it converges to a limit .x ∈ ∞n=1 Kn . Hence .x ∈ V for
some .V ∈ . By (ii), .Kn ⊂ V provided that n is sufficiently large. This contradicts
(iii), and K is therefore compact.

Definition 13.82 Let X be a topological space, and .(Y, d) be a metric space. A
family .F of elements of .Map(X, Y ) is equicontinuous if for every .x0 ∈ X and every
.ε > 0 there exists a neighborhood U of .x0 in X such that .d(f (x), f (x0 )) < ε for
every .f ∈ F and every .x ∈ U . The family .F is pointwise totally bounded if the set
.{f (x) | f ∈ F} is totally bounded in Y for every .x ∈ X.
Theorem 13.79 (Ascoli-Arzelà) Let X be a compact topological space and .(Y, d)

be a complete metric space. A family .F ⊂ Map(X, Y ) is relatively compact in the
topology of uniform convergence if and only if:
(i) .F is equicontinuous, and
(ii) .F is pointwise totally bounded.
Proof We prove that (i) and (ii) are necessary conditions, and we assume that
.F is contained in a compact set .K ⊂ Map(X, Y ). Since the evaluation map
.ω(·, x) : Map(X, Y ) → Y defined by .f → f (x) is continuous for every .x ∈ X,
the set .{f (x) | f ∈ F} is contained in .ex (K), a compact set.

Let .x0 ∈ X be fixed, and consider the continuous function .α : Map(X, Y )×X →
[0, +∞) such that
α(f, x) = d(f (x0 ), f (x)).

.
For every .ε > 0 the compact set .K × {x0 } is a subset of the open set .α −1 ([0, ε)),
and Wallace’s Theorem 13.77 yields an open set .U ⊂ X such that .x0 ∈ U and
.K × U ⊂ α
−1 ([0, ε)). In particular .d(f (x ), f (x)) < ε for every .f ∈ F and any
0
.x ∈ U .
We now prove the converse implication, assuming that (i) and (ii) hold. It suffices
to prove that .F is totally bounded in .Map(X, Y ). Pick any .ε > 0; the equicontinuity
of the family .F implies that for any .x ∈ X there exists an open neighborhood .Ux of
x such that .d(f (x), f (y)) < ε for every .f ∈ F and every .y ∈ Ux . Since the space
X is compact, there exist finitely many points .x1 , . . . , xn ∈ X such that
X = Ux 1 ∪ · · · ∪ Ux n .
.
The image of the map .F → Y ×'· · · × Y such that .f → (f (x1 ), . . . , f (xn )) is

therefore contained in the product . ω(F, xi ) of totally bounded sets, so it is totally
bounded. We can find a finite set .F ⊂ F such that , for every .f ∈ F, there exists
.g ∈ F with the property that .d(f (xi ), g(xi )) < ε for every .i = 1, . . . , n.
We claim that .F is the union of open balls centered at .g ∈ F with radius .3ε.
Indeed, pick .f ∈ Fand .g ∈ F such that .d(f (xi ), g(xi )) < ε for every i. Then for
any .x ∈ X there exists an index i such that .x ∈ Uxi , whence
d(f (x), g(x)) ≤ d(f (x), f (xi )) + d(f (xi ), g(xi )) + d(g(xi ), g(x)) < 3ε,
.
and the claim follows.

Corollary 13.3 (Classical Version of Ascoli-Arzelà’s Theorem) Let X be a
compact space. A subspace .F of .Map(X, Rn ) is relatively compact in the topology
of uniform convergence if and only if .F is equicontinuous and pointwise totally
bounded (in the standard metric topology of .Rn ).
Corollary 13.4 (Arzelà) Let X be a compact space, and let .{fk }k be a sequence of
continuous functions from X to .Rn . If the sequence .{fk }k is pointwise bounded and
equicontinuous, then there exists a uniformly convergent subsequence.
13.17 Cubes and Metrizability
Metric spaces are for sure the most ubiquitous topological structure in Mathematical
Analysis. We dare say that our minds always think in terms of a distance, although
this might be conceptually wrong.
Anyway, metric topologies are particularly easy to work with, and in this section
we try to analyze them and discover when a given topology is induced by a metric.
Definition 13.83 A pseudo-metric (or ècart) for a set X is a function .d : X×X → R
satisfying the following conditions: for every .x ∈ X, .y ∈ X and .z ∈ X,
(a) .d(x, y) = d(y, x);
(b) .d(x, y) ≤ d(x, z) + d(z, y) (triangle inequality);
(c) .d(x, y) = 0 if .x = y.
The function d is a metric if, in addition to the previous properties, it also satisfies
(d) if .d(x, y) = 0, then .x = y.
A pseudo-metric space is a couple .(X, d) in which X is a set and d is a pseudo-
metric on X.
Remark 13.20 Although the non-degeneracy condition (d) is usually assumed in
any Analysis textbook, it turns out that its relevance from a topological viewpoint is
really small. Of course the results of this section hold for any metric space.
Definition 13.84 (Balls) The open ball of center x and radius .r > 0 in a pseudo-
metric space X is the set
B(x, r) = {y ∈ X | d(x, y) < r} .

.
The closed ball of center x and radius .r > 0 is the set
. B(x, r) = {y ∈ X | d(x, y) ≤ r} .
Definition 13.85 (Topology Defined by a Pseudo-Metric) A subset U of a

pseudo-metric space X is open if and only if for every point .x ∈ U there exists an
open ball .B(x, r) such that .B(x, r) ⊂ U . This property gives rise to a topology for
X, called the pseudo-metric topology induced by d.
13.17 Cubes and Metrizability 281
Definition 13.86 (Distance from a Set) Let A be a subset of a pseudo-metric space

X. The distance from a point .x ∈ X to A is
d(x, A) = inf {d(x, a) | a ∈ A} .

.
Theorem 13.80 Let .A ⊂ X. The function .x → d(x, A) is continuous with respect

to the pseudo-metric topology of X.
Proof This follows easily from the triangle inequality. Indeed, since .d(x, a) ≤
d(x, y) + d(y, a), we can minimize over .a ∈ A and get
d(x, A) ≤ d(x, y) + d(y, A).

.
Swapping x and y we see that .d(y, A) ≤ d(y, x) + d(x, A), and therefore
. |d(x, A) − d(y, A)| ≤ d(x, y).
If .y ∈ B(x, r), then .|d(x, A) − d(y, A)| ≤ r, and the proof is complete.

Theorem 13.81 In a pseudo-metric space X, the closure of A is the set of point
whose distance to A is equal to zero.
Proof We have to prove that
A = {x ∈ X | d(x, A) = 0} .
.
By continuity, the set .{x ∈ X | d(x, A) = 0} is closed in X and trivially contains

A. Hence .A ⊂ {x ∈ X | d(x, A) = 0}. Conversely, if .y ∈ / A, there exists a
neighborhood of y, which we may assume to be an open sphere .B(y, r) for
some .r > 0, such that .B(y, r) ∩ A = ∅. Thus .d(y, A) ≥ r > 0, and
.{x ∈ X | d(x, A) = 0} ⊂ A. The proof is complete.

Theorem 13.82 Every pseudo-metric space is normal.
Proof Consider two disjoint closed subsets A and B of a pseudo-metric space X.
We define
U = {x ∈ X | d(x, A) − d(x, B) < 0}

.
V = {x ∈ X | d(x, A) − d(x, B) > 0} .
By continuity, U and V are open sets, and they are clearly disjoint. Since A and B
are closed, it follows from Theorem 13.81 that .A ⊂ U and .B ⊂ V . This concludes
the proof.

Theorem 13.83 A net .{Sn , n ∈ D} is a pseudo-metric space X converges to a point
x if and only if the net .{d(x, Sn ), n ∈ D} converges to zero.
Proof Indeed, a net .{Sn , n ∈ D} converges to x if and only if it is eventually in every

open ball .B(x, r), and this happens if and only if .{d(x, Sn ), n ∈ D} is eventually in
the open interval .(−r, r) centered at .0 ∈ R.

Definition 13.87 (Diameter) The diameter of a subset A of a pseudo-metric space
X is
. diam A = sup {d(x, y) | x ∈ A, y ∈ A} .
If the supremum on the right-hand side is infinite, we way that A has infinite
diameter.
We have already proved the following result in Proposition 13.4. We propose here a
different proof.
Theorem 13.84 Let .(X, d) be a pseudo-metric space, and let
e(x, y) = min {1, d(x, y)} .

.
Then .(X, e) is a pseudo-metric space whose topology coincides with the topology
induced by d.
Proof Let .a ≥ 0, .b ≥ 0 and .c ≥ 0 be such that .a + b ≥ c. We claim that
. min{1, a} + min{1, b} ≥ min{1, c}. (13.5)
If either .min{1, a} of .min{1, b} is equal to 1, then the claim reduces to .min{1, c} ≤ 1,

which is surely true. If neither of these is equal to 1, the validity of (13.5) follows
from .a + b ≥ c ≥ min{1, c}.
If we set .a = d(x, y), .b = d(y, z) and .c = d(z, y), we see that e is a pseudo-
metric on X. Now, the collection of all open balls of radius .r ≤ 1 is a base for the
topology induced by any pseudo-metric. Since this collection is the same for d and
for e, the topologies induced by d and e must coincide. This concludes the proof.

Theorem 13.85 (Countable Products of Pseudo-Metric Spaces) Consider a
sequence .{(Xn , dn ) | n ∈ N} of pseudo-metric spaces, each of diameter at most
one, and define d by
∞
dn (xn , yn )
d(x, y) =
.
2n
n=1
'∞
Then d is a pseudo-metric for .X = n=1 Xn , and the topology induced by d is the
product topology.
Proof The fact that d is indeed a pseudo-metric on the cartesian product is an easy
exercise. We only show that the product topology coincides with the pseudo-metric
topology. To this aim, we remark that if .V = B(x, 2−p ) is an open ball centered at
.x ∈ X and if

−n−p−2
.U = y ∈ X dn (xn , yn ) < 2 for every n ≤ p + 2 ,
there results .U ⊂ V . Indeed, if .y ∈ U , then

p+2 ∞

1 1 1
d(x, y) <
. + n
< p.
2n+p+2 2 2
n=1 n=p+3
It follows that each set which is open in the pseudo-metric topology is also open in
the product topology of X. Conversely, we consider any element U of the sub-base
which defines the product topology of X. Hence .U = {x ∈ X | xn ∈ W } for some
open set .W ⊂ Xn . Any .x ∈ U has an open ball of radius r, centered at .xn and
contained in W . Since
dn (xn , yn )
d(x, y) ≥
. ,
2n
the open ball centered at x with radius .r/2n is a subset of U . This proves that any
element of the sub-base of the product topology is open relative to the pseudo-metric
topology, and the proof is complete.

Definition 13.88 (Isometries) A function .f : X → Y between the pseudo-metric
spaces .(X, d) and .(Y, e) is an isometry if and only if
e(f (x), f (y)) = d(x, y)

.
holds for every .x ∈ X, .y ∈ X.

Exercise 13.48 Prove that any isometry is continuous. Prove that the composition
of two isometries is again an isometry, and that the inverse of an isometry (if it
exists) is an isometry.
So far we have seen that all the basic properties of metric spaces are properties of
pseudo-metric spaces. There is however a point at which condition (d) of Definition
13.83 is needed to exclude undesired limitations.
Definition 13.89 (Distance Between Sets) The distance of two non-empty subsets
A and B of a pseudo-metric space X is
d(A, B) = inf {d(a, b) | a ∈ A, b ∈ B} .

.
Unfortunately, we immediately realize that .d(X, A) = 0 for every .A ⊂ X, and

the triangle inequality fails. To overcome this issue, we introduce an equivalence
relation on X.
Definition 13.90 If .(X, d) is a pseudo-metric space, we define the relation
R = {(x, y) | d(x, y) = 0}
.
on .X×X. The quotient space .X/R is endowed with the quotient topology associated
to R.
Theorem 13.86 Let .(X, d) be a pseudo-metric space, and let

D = {x} x ∈ X .
.
For every .A ∈ D and .b ∈ D we define
.d(A, B) = inf {d(a, b) | a ∈ A, b ∈ B} .
Then .(D, d) is a metric space whose topology is the quotient topology of .X/R.
Furthermore, the projection of X onto .D is an isometry.
Proof We begin with a remark: .u ∈ {x} if and only if .d(u, x) = 0, which is true if
and only if .x ∈ {u}. If .u ∈ {x} and .v ∈ {y}, then
d(u, v) ≤ d(u, x) + d(x, y) + d(y, v) = d(x, y).

.
But .x ∈ {u} and .y ∈ {v}, hence .d(u, v) = d(x, y). As a consequence, for every
.A ∈ D, .B ∈ D, the value of .d(A, B) coincides with the value of .d(x, y) for every
.x ∈ A and .y ∈ B. This proves that .D, d) is a metric space, and the projection of X
onto .D is an isometry.
Let U be an open set in X and .x ∈ U . There exists .r > 0 such that .x ∈ B(x, r) ⊂
U , hence .{x} ⊂ U . The projection of X onto .D is thus an open map with respect
to the quotient topology, but the projection is also an open map with respect to the
metric topology induced by the distance d between sets. Hence these topologies
must coincide.

A basic question arises at this point: when is a given topology on a set X the
topology associated to a (pseudo)metric?
Definition 13.91 (Metrizable Spaces) A topological space X is metrizable if and
only if its topology coincides with the topology induced by a metric on X. Similarly,
X is pseudo-metrizable if and only if f its topology coincides with the topology
induced by a pseudo-metric on X.
Exercise 13.49 Prove that a pseudo-metric is a metric if and only if the associated
topology is .T1 . Deduce that a topological space is metrizable if and only if it is
pseudo-metrizable and its topology is .T1 .
A typical approach to metrizability results is via good embeddings of a given
topological space.
Definition 13.92 (Cubes) Any cartesian product of the unit interval .[0, 1],
endowed with the product topology, is called a cube.
Explicitly, a cube is a topological space of the form .[0, 1]A , where A is a set, and
its topology is the topology of pointwise convergence. As a product space, each
element of a cube is a function whose domain is a specified set. In view of this
generality, cubes may fail to have good properties, and this is the reason why we
need to add suitable assumptions to the functions of our cubes.
Definition 13.93 Let X be a topological space, and let F be a collection of
functions .fj : X
'→ Yj , .i ∈ J, such that .Yj is a topological space. The evaluation
map .e : X → Yj j ∈ J is defined as follows: for every .x ∈ X, .e(x) is the
function .j ∈ J → fj (x).
Thus, roughly speaking the j -th coordinate of .e(x) is .fj (x), or .e(x)j = fj (x).

Definition 13.94 The collection .F = fj : X → Yj j ∈ J distinguishes points
if and only if for every .x ∈ X, .y ∈ X such that .x = y there exists .j ∈ J such that
.fj (x) = fj (y).

Definition 13.95 The collection .F = fj : X → Yj j ∈ J distinguishes points
and closed sets if and only if for every .x ∈ X and every closed set A such that
.x ∈ X \ A there exists .j ∈ J such that .fj (x) ∈
/ f (A).
We summarize in the next result the main topological features of the evaluation map.

Theorem 13.87 Let .F = fj : X → Yj j ∈ J be a collection of continuous
functions. Then
'
(a) the evaluation map .e : X → Yj j ∈ J is continuous;
(b) The evaluation e is an open map of X onto .e(X) if and only if F distinguishes
points and closed sets.
(c) The evaluation map e is injective if and only if F distinguishes points.
Proof Since .Pj ◦ e(x) = fj (x) for every .j ∈ J , by definition of the product
topology it follows that e is continuous, and (a) is proved. To prove (b), we show
that the image under e of an open neighborhood U of a point x contains the
intersection of .e(X) and a neighborhood of .e(x) in the product topology. Let .j ∈ J
such that .fj (x) does not belong to the closure of .fj (X \ U ). Now, the set of all
'
.y ∈ Yj j ∈ J such that .Pj (y) ∈ / fj (X \ U ) is open, and evidently the
intersection with .e(X) is a subset of .e(U ). This proves that e is a open map of
X onto .e(X).
Finally, .e(x) = e(y) if and only if .fj (x) = fj (y) for every .j ∈ J . Hence e is
injective if and only if F distinguishes points.

In virtue of the last result, a topological space can be embedded into a cube (i.e.
it is homeomorphic to a subset of a cube) provided that it is possible to construct a
sufficiently rich collection of continuous functions defined on the space. By the very
definition, the existence of such collections seems to be related to some separation
properties of the topology.
Definition 13.96 (Completely Regular Spaces) A topological space X is com-

pletely regular if and only if for every .x ∈ X and every neighborhood U of x,
there exists a continuous function .f : X → [0, 1] such that .f (x) = 0 and .f = 1 on
.X \ U . A completely regular .T1 space is a Tychonoff space.
Exercise 13.50 Prove that the collection of all continuous functions .f : X → [0, 1]
defined on a completely regular space X distinguishes points and closed sets.
Exercise 13.51 Suppose X is a Tychonoff space, and let F be the collection of
all continuous functions from X to .[0, 1]. Prove that the evaluation map e is a
homeomorphism of X with a subspace of .[0, 1]F . Hint: use Theorem 13.87.
Our task now is to prove the converse of the last exercise. We need a preliminary
result.
Theorem 13.88 Any product of Tychonoff spaces is a Tychonoff space.
Proof We say that a continuous function .f : X → [0, 1] defined on a topological
space X is for a pair .(x, U ) if and only if x is a point of X, U is a neighborhood of
x, .f (x) = 0 and .f = 1 on .X \ U . Now, if .f1 , . . . , fn are for .(x, U1 ), . . . , (x, Un ),
then we can set
g(x) = sup {fi (x) | i = 1, . . . , n}

.

and conclude that g is for . x, ni=1 Ui . This shows that X is completely regular if
for every x and every neighborhood U of x belonging to a sub-base of the topology
of X, there exists a function for the pair .(x, U ). '
Consider now the case in which X is a product . {Xα | α ∈ A} of Tychonoff
spaces. Let .x ∈ X and .Ua be a neighborhood of .xa = Pa (x) in .Xa . If f is a
function for .(xa , Ua ), then .f ◦ Pa is a function for .(x, Pa−1 (Ua )). The collection
of all sets .Pa−1 (Ua ) forms a sub-base of the product topology of X, hence X is
completely regular. The fact that any product of .T1 spaces is a .T1 space completes
the proof.

Theorem 13.89 (Embedding into Cubes) For a topological space X, the follow-
ing are equivalent:
(a) X is a Tychonoff space;
(b) X is homeomorphic to a subspace of a cube.
Proof Clearly (b) implies (a). Conversely, we remark that the space .[0, 1] is a
Tychonoff space, thus any cube is a Tychonoff space by Theorem 13.88. Therefore
any subspace of a cube is a Tychonoff space. Exercise 13.51 and Theorem 13.87
show that the evaluation map e is a homeomorphism of X into a cube, and the proof
is complete.

We are ready to prove a sufficient condition for a topology to be metrizable.
Theorem 13.90 (Urysohn) A regular .T1 topological space whose topology has a
countable base is homeomorphic to a subspace of the cube .[0, 1]R . In particular, it
is metrizable.
Proof Let us explain the strategy of the proof. A product of countably many pseudo-
metrizable spaces is pseudo-metrizable, see Theorem 13.85. By Theorem 13.87, if
F is a collection of continuous function on a .T1 space X, where an'element
f of F
maps X to .Yf , then the evaluation map is continuous from X to . Yf f ∈ F ,
and it is a homeomorphism as soon as F distinguishes points and closed sets. To
show that X is metrizable, it suffices to construct a countable family (i.e. a sequence)
F of continuous functions, each from X into .[0, 1], such that F distinguishes points
and closed sets.
Let .B be a countable base for the topology of X, and let

A = (U, V ) U ∈ B, V ∈ B, U ⊂ V .
.
It is clear that .A is a countable set. To each .(U, V ) ∈ A we associate a continuous

function .f : X → [0, 1] such that .f = 0 on U and .f = 1 on .X \ V . We call
F the collection of all such functions. Since F is indexed over a countable set, F
is countable. We claim that F distinguishes points and closed sets. So, let B be
a closed set and .x ∈ X \ B. We choose a neighborhood .V ∈ B of x such that
.x ∈ V ⊂ X \ B. Furthermore, we choose .U ∈ B such that .U ⊂ V (this is possible
since X is regular). But then .(U, V ) ∈ A, and if f is the element of F associated to

.(U, V ), then .f (x) = 0 ∈ / {1} = f (B). The proof is complete.

A more general metrizability result can be proved, but we need to introduce a
new definition and a couple of preliminary results.
Definition 13.97 A collection .G of subsets of a topological space is .σ -locally finite
if and only if .G is the union of a countable collection of locally finite collections .Gk ,
.k ∈ N.
We recall that a collection .G of subsets of a topological space X is locally finite

if and only if each point .x ∈ X has a neighborhood which intersects only finitely
many elements of .G.
Theorem 13.91 (Stone) Every open cover .U of a metric space .(X, d) has a .σ -
locally finite refinement .V = {Vk | k ∈ N}.
Proof For every .k ∈ N and every .U ∈ U we define

1
Uk = x ∈ X d(x, X \ U ) ≥ k
. ⊂ U.
2
An immediate consequence of the triangle inequality is that
1 1 1
d(Uk , X \ Uk+1 ) ≥
. − k+1 = k+1 .
2k 2 2
Here comes the hard part of the proof: we select a well-order ., on .U. For every
k ∈ N and every .U ∈ U we introduce
.

Uk% = Uk \
. {Vk+1 | V ∈ U, V , U } ⊂ Uk .
In the well-order .,, any pair .(U, V ) of distinct sets in .U satisfies either .U , V
or .V , U , but not both. Hence either .Uk% ⊂ X \ Vk+1 or .Vk% ⊂ X \ Uk+1 . In other
words, either
1
d(Uk% , Vk% ) ≥ d(X \ Vk+1 , Vk ) ≥
.
2k+1
or
1
d(Uk% , Vk% ) ≥ d(Uk , X \ Uk+1 ) ≥
. .
2k+1
Then

1
%%
.Uk = x ∈ X d(x, Uk% ) ≤ k+3
2
is contained in U . Moreover, for .U = V ,
1 1 1
d(Uk%% , Vk%% ) ≥
. −2· = .
2k+1 2k+3 2k+2

Thus each collection .Vk = Uk%% U ∈ U is locally finite. Indeed, for every point
.x ∈ X, the open ball .B(x, 2
−k−3 ) intersects at most one element of .V .
k
The collection .V = {Vk | k ∈ N} is a cover of X. Every .x ∈ X belongs to some
.U ∈ U, hence it belongs to .Uk for some .k ≥ 1. As a consequence, .x ∈ U ⊂ U ,
% %%
k k
%%
when U is the smallest element of .U which contains x. Since .Uk is open and is a
subset of U , the collection .V is a .σ -locally finite open refinement of .U. The proof
is complete.

Theorem 13.92Every regular topological space X whose topology has a .σ -locally
finite base .B = {Bk | k ∈ N} is a normal space.
Proof Fix two disjoint closed subsets A and B of X. By regularity, a cover .U of
A exists made by open sets that have closure disjoint from B, and a cover .V of B
exists made by open sets that have closure disjoint from A. By assumption we may
express

U=
. {Uk | k ∈ N} , V= {Vk | k ∈ N} ,
where .Uk ⊂ Bk and .Vk ⊂ Bk for every k. For every .k ∈ N, we set

Uk =
. {U | U ∈ Uk } , Vk = {V | V ∈ Vk } .
Since .Bk is locally finite,

Uk =
. U U ∈ Uk , Vk = V V ∈ Vk .
We thus see that A has the countable cover .{Uk | k ∈ N} by open sets with closure
.U k ⊂ X \ B, and B has the countable cover .{Vk | k ∈ N} by open sets with closure
.V k ⊂ X \ A. By replacing .Uk and .Vk with the union of their predecessors in the
indexing, we obtain two covers of A and B by nested open sets:
U1 ⊂ U2 ⊂ . . . ⊂ Uk ⊂ . . .
.
V1 ⊂ V2 ⊂ . . . ⊂ Vk ⊂ . . .
These sets still have the properties that .U k ⊂ X \ B and .V k ⊂ X \ A. Then the sets

U=
. Uk \ V k k ∈ N , V = Vk \ U k k ∈ N
are open neighborhoods of A and B, respectively. But they are also disjoint, since

Uk ∩ X \ U n ⊃ Uk \ V )k ∩ Vn \ U n ⊂ X \ V k ∩ Vn ,
.

where .Uk ∩ X \ U n = ∅ when .k ≤ n, and . X \ V k ∩ Vn = ∅ when .n ≤ k.

Theorem 13.93 (Nagata-Smirnov) For a topological space X the following prop-
erties are equivalent:
(a) X is metrizable;
(b) X is regular and has a .σ -locally finite base .B = {Bk | k ≥ 1}.
Proof (a) implies (b). Indeed, for every integer .n ≥ 1 the opencover .U(b) by open
finite refinement .B(n) = {Bk (n) | k ≥ 1} by
balls of radius .1/n has a .σ -locally
Stone’s Theorem. Then .B = {B(n) | n ≥ 1} is a .σ -locally finite base for the
metric topology of X.
(b) implies (a). Indeed, we begin with a .σ -locally finite base .B = {Bk | k ≥ 1}
for the topology of X. For every couple .(m, n) ∈ N × N and every .U ∈ Bm , we set

. G= V ∈ Bn V ⊂ U .
Since .Bn is locally finite,

.G= V ∈ Bn V ⊂ U ,
so that .G ⊂ U . By Theorem 13.92, X is normal. We select a Urysohn function

uU : X → [0, 1] such that .uU (G) = {1} and .uU (X \ U ) = {0}. We consider the
.
function .dm,n : X × X → R such that

.dm,n (x, y) = {|uU (x) − uU (y)| | U ∈ Bm } .
Since .Bm is locally finite, each element .(x, y) ∈ X×X has a neighborhood on which
the summation in the definition of .dm,n is a finite sum. As a consequence, .dm,n is
a continuous function. It is easy to check that .dm,n is a pseudo-metric on X. We
call .Ym,n the set X endowed with the topology induced by the pseudo-metric .dm,n ,
and let .gm,n : X → Ym,n be the inclusion map. It is clear that .gm,n is a continuous
function.
We claim that the countable collection .F = gm,n (m, n) ∈ N × N separates
points and closed sets. Indeed, pick any .x ∈ X and any closed set A which does
not contain x. There exists a member .U ∈ Bm ⊂ Bof the base such that .x ∈ U ⊂
X \ A. By regularity, there exists a second member .V ∈ Bm of the base such that
.x ∈ V ⊂ V ⊂ U ⊂ X \ A. But then .dm,n (x, A) ≥ 1, since .uU (V ) = {1} and .uU
vanishes on A. As a consequence, .gm,n (x) = x does not belong to the closure (with
respect to .dm,n ) of .gm,n (A) = A. The claim is proved.
The evaluation map
(
e: X →
. Ym,n (m, n) ∈ N × N
is an embedding of X onto a subspace of the countable ' product of pseudo-

metrizable spaces. By Theorem 13.85 the product . Ym,n (m, n) ∈ N × N is
pseudo-metrizable by a metric d. But this pseudo-metric is actually a metric on the
image .e(X), since .e(X) consists of constant sequences. Indeed, there exists a pair
.(m, n) ∈ N × N such that .dm,n (x, y) ≥ 1 whenever .x = y, by taking .A = {y} in
the separation property of the collection .F. The proof is complete.

13.18 Problems
13.1 Let S be a subset of a topological space X. Show that a sequence {sn }n in S

converges to s ∈ S in the relative topology if and only if, considered as a sequence
in X, the sequence {sn }n converges to s.
13.2 Prove that the space
R2 \ {(x, y) | x ∈ N ∧ y ∈ N}
.
13.18 Problems 291
is homeomorphic to the space

1
.R2 \ (x, y) (∃n)(∃m)(n ∈ N ∧ m ∈ N) ⇒ (x − n)2 + (y − m)2 < .
10
13.3 Let (X, d) be a compact metric space, and let f : X → X be an isometry,

i.e. d(x, y) = d(f (x), f (y)) for every x ∈ X, y ∈ X. Prove that f is surjective.
Hint: set A = f (X), and suppose there exists x0 ∈ X \ A. Define inductively
xn = f (xn−1 ). Now elaborate on the sequence {xn }n .
13.4 Let (X, d) be a metric space. In analogy with the field of real numbers, we say
that a sequence {xn }n in X is a Cauchy sequence if and only if for each ε > 0 there
exists N ∈ N such that N ≥ N, m ≥ N imply d(xn , xn ) < ε. Furthermore, we
say that (X, d) is a complete metric space if and only if each Cauchy sequence in X
converges to a point of X. Prove that X is complete if it is a compact metric space.
13.5 Let (X, d) be a metric space. The diameter of any non-empty subset E of X
is defined to be the supremum of d(x, y) as x and y range over E. Prove that a
sequence {xn }n in X is a Cauchy sequence if and only if
. lim diam {xn , xn+1 , xn+2 , . . .} = 0.

n→+∞
13.6 Let (X, d) be a metric space, and let E be a subset of X. Prove that diam E =
E.
13.7 Let (X, d) be a metric space.
1. Call two Cauchy sequences {xn }n , {yn }n in X equivalent if and only if
. lim d(xn , yn ) = 0.
n→+∞
Prove that this is indeed an equivalence relation.

2. Let X∗ be the set of all equivalence classes so obtained. If P ∈ X∗ and Q ∈ X∗ ,
define
(P , Q) = lim d(xn , yn ),

.
n→+∞
where P = [{xn }n ] and Q = [{yn }n ]. Prove that this limit exists, and that
(P , Q) depends only on P and Q, but not on the representatives {xn }n and
{yn }n of P and Q.
3. Prove that (X∗ , ) is a complete metric space.
4. For each point x ∈ X, there is a Cauhcy sequence all of whose terms are equal
to x: let Px be the element of X∗ which contains this sequence. Prove that
(Px , Py ) = d(x, y)
.
for all x, y ∈ X. In other words, the mapping ϕ defined by ϕ(x) = Px is an

isometry of X into X∗ .
5. Prove that ϕ(X) is dense in X∗ , and that ϕ(X) = X∗ if X is complete. The metric
space (X∗ , ) is called the completion of (X, d).
13.8 A proper interval in R is a half-open, open, or closed interval which contains
more than one point. If A is an arbitrary family of proper intervals, prove that there
exists a countable subfamily B of A such that

. {B | B ∈ B} = {A | A ∈ A} .
Hint: prove that a disjoint family of proper

intervals is countable, and show that all
but a countable number of points of {A | A ∈ A} are interior points of members
of A.
13.9 Let X be a topological space. If A is dense and U is open, prove that U ⊂
A ∪ U.
13.10 For each a ∈ Z, b ∈ Z, let
Na.b = {a + kb | k ∈ Z} .
.

1. Prove that the collection B = Na,b a ∈ Z, b ∈ Z, b > 0 is a base for a
topology τ on Z.
2. Prove that each Na,b is both open and closed in (Z, τ ).
3. Let P = {2, 3, . . .} the set of prime numbers. Prove that

Z \ {−1, 1} =
. N0,p p ∈ P .
Deduce that if P were a finite set, then {−1, 1} would be an open set in (Z, τ ).
Hence P is an infinite set.
13.11 Let Y be a dense subset of a Hausdorff topological space X. If Y is locally
compact, prove that Y is open in X.
13.12 Let (X, d) be a metric space. If A ⊂ X and ε > 0, we set
U (A, ε = {x ∈ X | d(x, a) < ε for every a ∈ A} .

.
Let H be the collection of all non-empty closed, bounded subsets of X. For A,

B ∈ H, we set
D(A, B) = inf {ε > 0 | A ⊂ U (B, ε) and B ⊂ U (A, ε)} .

.
13.19 Comments 293
(a) Prove that D is a metric on H, which we call the Hausdorff metric.

(b) Prove that if (X, d) is complete, so is (H, D). Hint: Let {An }n be a
Cauchy sequence in H; by passing to a subsequence, we may assume that
D(An , An+1 ) < 2−n . Define A to be the set of all points x which are the limits
of sequences {xn }n such that xn ∈ An for every n and d(xn , xn+1 ) < 2−n . Prove
that An → A with respect to D.
(c) Prove that if (X, d) is totally bounded, so is (H, D). Hint: Given ε > 0,
pick δ < ε and let S be a finite subset of X such that X is covered by
{B(x, δ) | x ∈ S}. Let A be the collection of all non-empty subsets of S.
Prove that {B(A, ε) | A ∈ A} covers H.
(d) If X is compact in the metric d, prove that H is compact in the Hausdorff metric
D.
13.13 Let X be a topological space, and (Y, d) be a metric space. If the subset F
of Map(X, Y ) is totally bounded with respect to the metric of uniform convergence,
prove that F is equicontinuous.
13.14 A real number p > 0 is called an almost-period relative to ε > 0 for the
function f : R → R if and only if |f (x + p) − f (x)| < ε for every x ∈ R. The
function f is almost-periodic if and only if for every ε > 0 there exists a length
> 0 such that in every interval of length it is possible to find an almost-period p
relative to ε. Prove that an almost-periodic function must be bounded.
13.19 Comments
I believe that the chapter on General Topology has a fundamental role in a book like
this one. This is why this chapter is particularly long and full of ideas and results.
The reader may have noticed that I did not insist on a strictly economic exposition:
a few definitions appear twice, and the order of appearance of the main characters is
not always coherent with the tradition. As an example, separation axioms just come
into play when they are needed from the viewpoint of a Mathematical Analyst. As
long as it looks possible, the Hausdorff separation axiom is used alone, because
this is exactly the basic condition ensuring the most important fact of Analysis:
a function converges to at most one point. When the construction of continuous
functions which separate sets becomes necessary, we introduce normality and
regularity. The locally compact Hausdorff case is dealt with separately, although
one might optimize several proofs by reducing to the normal case. I have decided to
do so because locally compact Hausdorff space are the natural setting of Measure
Theory, a topic that will be discussed in the next chapters.
While writing this chapter, I had in mind the classical reference [5]. Kelley’s
book remains a great source for the young Analyst who wishes to learn “what every
Analyst should know about topology”, but its main feature is that the whole book
should be read like a romance, from cover to cover. This is nowadays uncommon,
since textbooks are written for the time-lacking reader.
Another standard reference is [6], whose style is opposed to Kelley’s. There are
plenty of pictures, diagrams, sketches, although the book is somehow redundant for
our purposes.
The definitive bible of General Topology is surely [4], whose bibliography is an
encyclopedia of references.
References
1. J.F. Aarnes, P.R. Andenaes, On nets and filters. Math. Scand. 31, 285–292 (1972)
2. P.R. Chernoff, A simple proof of Tychonoff’s theorem via nets. Am. Math. Mon. 99(10), 932–
934 (1992)
3. C. Fefferman, An easy proof of the fundamental theorem of algebra. Am. Math. Mon. 74, 854–
855 (1967)
4. R. Engelking, General Topology. Sigma Series in Pure Mathematics Series Profile, vol. 6
(Heldermann Verlag, Berlin, 1989)
5. J.L. Kelley, General Topology. Graduate Texts in Mathematics, vol. 27 (Springer-Verlag, New
York-Berlin, 1975). Reprint of the 1955 edition [Van Nostrand, Toronto, Ont.]
6. J.R. Munkres, Topology (Prentice Hall, Upper Saddle River, 2000), xvi, 537 p.
7. M.E. Rudin, A new proof that metric spaces are paracompact. Proc. Am. Math. Soc. 20, 603
(1969)
8. L. Schwartz, Analyse, in Topologie Générale et Analyse Fonctionnelle (Hermann, Paris, 1993)
9. A.H. Stone, Paracompactness and product spaces. Bull. Am. Math. Soc. 54, 977–982 (1948)
Chapter 14
Differentiating Again: Linearization
in Normed Spaces
Abstract This chapter is devoted to an overview of basic linear analysis in normed

spaces.
14.1 Normed Vector Spaces
Let us start with some old friends.

Definition 14.1 A vector space V over .R is a commutative group under a binary
operation .+, together with an operation of scalar multiplication .R × V → V ,
denoted by juxtaposition, such that
1. .a(x + y) = ax + ay
2. .a + b)x = ax + bx
3. .a(bx) = (ab)x
4. .1x = x
for every .x ∈ V , .y ∈ V , .a ∈ R, .b ∈ R. The symbol 0 will be used both for the real
number zero and for the additive identity in V .
Remark 14.1 The systematic use of bold-face fonts to denote vectors, like in .x or .v,
is no longer popular among mathematicians.
Definition 14.2 Let V be a vector space. Given .E ⊂ R, .a ∈ R, .A ⊂ V , .B ⊂ V ,
x0 ∈ V , we set
.
A + B = {x + y | x ∈ A, y ∈ B}
.
x0 + B = = {x0 + y | y ∈ B}
EA = {ax | a ∈ E, x ∈ A}
aA = {ax | x ∈ A} .
296 14 Differentiating Again: Linearization in Normed Spaces
Definition 14.3 Let V be a vector space, and let .W ⊂ V . We say that W is a vector
subspace of V if .W +W ⊂ W and .RW ⊂ W . Furthermore, we say that W is convex
if .ax + (1 − a)y ∈ W whenever .a ∈ R, .x ∈ W , .y ∈ W and .0 ≤ a ≤ 1. We say
that W is symmetric if .−1W = −W = W , and that W is balanced if .aW ⊂ W for
every .R such that .|a| ≤ 1.
Exercise 14.1 Prove that W is a vector subspace of V if and only if for every .a ∈ R,
b ∈ R, .u ∈ W and .v ∈ W there results .au + bv ∈ W .
.
Definition 14.4 Let V be a vector space, and suppose that .p : V → R. We say that
p is a seminorm on V if
1. .p(x + y) ≤ p(x) + p(y)
2. .p(ax) = |a|p(x)
for every .x ∈ V , .y ∈ V , .a ∈ R. Of course the first property is the triangle inequality
for p, while the second property is a homogeneity property.
Exercise 14.2 Let p be a seminorm on the vector space V . Prove that .p(0) = 0,
p(x) ≥ 0 for every .x ∈ V , and .p(x − y) ≥ |p(x) − p(y)| for every .x ∈ V , .y ∈ V .
.
Definition 14.5 Let V be a vector space, and let p be a seminorm on V . If
∀x(x ∈ V ∧ p(x) = 0) ⇒ (x = 0),

.
i.e. .x = 0 is the unique zero of the seminorm p, we say that p is a norm on V .

If p is a norm on V , we will usually employ a less generic notation. It is
customary to write .* · * instead of .p(·). Normed vector spaces will be our primary
object of interest in the rest of this chapter.
Remark 14.2 If .(V , * · *) is a normed vector space, it is a trivial exercise to check
that V inherits a topology as a metric space with the metric .d : V × V → R such
that
d(x, y) = *x − y*.
.
This topology will be called the norm topology.

Definition 14.6 A normed vector space is a Banach space if it is complete (as a
metric space) in the norm topology.
Let us present a few examples that play an important role in Analysis.
Example 14.1 Let X be a Hausdorff topological space. By .C(X), .C0 (X) and .Cc (X)
we denote, respectively, the vector spaces of all continuous real-valued functions
on X that are bounded, vanish at infinity, or have compact support. The algebraic
operations are, of course, defined pointwise on X.
The set .C0 (X) requires some further explanation, since X does not carry a metric
structure. We say that a function .f : X → R vanishes at infinity if, for every .ε > 0,
14.1 Normed Vector Spaces 297
there exists a compact subset .Kε of X such that .|f (x)| < ε for every .x ∈ X \ Kε .
If X is a normed vector space (with norm .* · *), this means that for every .ε > 0
there exists .M ∈ R such that .*x* > M implies .|f (x)| < ε, or equivalently that
.lim*x*→+∞ f (x) = 0.
If f is an element of any of these spaces, we set
*f * = *f *∞ = sup {f (t) | t ∈ X} .
.
Exercise 14.3 Prove that .C(X) and .C0 (X) are Banach spaces. If X is a compact
Hausdorff space, prove that the normed spaces .C(X), .C0 (X) and .Cc (X) coincide.
Example 14.2 Let .a < b and let .n = 0, 1, 2, . . . We denote by .C n ([a, b]) the vector
space of n-times continuously differentiable functions from .[a, b] to .R. We define
n *
*
* (k) *
*f *n =
. *f *
∞
k=0
for .f ∈ C n ([a, b]), where .f (k) stands for the k-th derivative of f , .0 ≤ k ≤ n.
Finally, we define
∞

C ∞ ([a, b]) =
. C n ([a, b]).
n=0
Exercise 14.4 Prove that .C n ([a, b]) is a Banach space.

Remark 14.3 The space .C ∞ ([a, b]) has a more complicated topological structure
which we are not going to describe in this book.
Example 14.3 Let .0 < p < ∞ be a fixed number. We define .
p = p (N) as the
vector space of all sequences .{xn }n of real numbers such that . ∞ n=1 |xn | ∈ R. In
p
this case, for .x = {xn }n we define the norm

" ∞
#1/p

. *x*p = |xn | p
.
n=1
If .p = ∞, the space .∞ = ∞ (N) is the vector space of all bounded sequences of
real numbers. The norm of .x = {xn }n ∈ ∞ is
*x*∞ = sup |xn |.

.
n∈N
Exercise 14.5 Prove that .p and .∞ are Banach spaces.
Definition 14.7 Let V be a (real) vector space. An inner (or a scalar) product on
V is a function which assigns to every .(x, y) ∈ V × V a real number ."x | y# such
that
1. (linearity) ."ax + by | z# = a"x | z# + b"y | z#
2. (symmetry) ."x, y# = "y | x#
3. (positivity) ."x | x# ≥ 0
4. (non-degeneracy) ."x | x# = 0 implies .x = 0
for every .x ∈ V , .y ∈ V , .z ∈ V , .a ∈ R, .b ∈ R. The norm induced by the inner
product is defined by

*x* =
. "x | x# x ∈ V.
A Hilbert space is a vector space which is a complete metric space with respect to
the norm induced by an inner product.
Remark 14.4 If V is a complex vector space, in the sense that vectors are multiplied
by complex numbers, condition 2 in the definition of an inner product must be
changed to
2. ."x, y# = "y | x#;
in this case the inner product is usually called hermitian instead of symmetric.
14.2 Bounded Linear Operators
Important: Norms on Different Spaces

With an abuse of notation, more often than not we use the same symbol for norms
of different spaces. The context usually permits to avoid any confusion.
Definition 14.8 Let X, Y be Banach spaces. A function T : X → Y is a bounded

linear operator if
1. T is linear, i.e. T (ax + by) = aT x + bT y for every x ∈ X, y ∈ X, a ∈ R,
b ∈ R;
2. there results
*T x*
. sup ∈ R. (14.1)
x∈X\{0} *x*
14.3 The Hahn-Banach Theorem 299
If this is the case, the real number in (14.1) is called the operator norm of T , and is
denoted by *T * (as usual!). The set of all bounded linear operators from X to Y is
denoted by L(X, Y ).
Exercise 14.6 Prove that L(X, Y ) is a Banach space under the operator norm. It
should be remarked that only the completeness of Y is necessary here.
Definition 14.9 Let X be a Banach space. The (topological) dual space of X is
X∗ = L(X, R), i.e. the Banach space of all bounded linear operators from X to R.
Any such operator will be called a bounded linear functional on X. The operator
norm of f ∈ X∗ will be also denoted by *f *X∗ .
Remark 14.5 Sometimes X% is used instead of X∗ , but we will always stick to our
symbol.
Theorem 14.1 Let X, Y be Banach spaces, and let T : X → Y be a linear operator.
The following are equivalent:
(a) T is continuous at some point x0 ∈ X;
(b) T is continuous at 0 ∈ X;
(c) T is a bounded operator;
(d) T is uniformly continuous.
Proof Obviously (d) implies (c) implies (b) implies (a). Suppose now that (a) holds.
If V is an open neighborhood of the origin of Y , then V1 = V + T x0 is an open
neighborhood of T x0 . By assumption there exists an open neighborhood U1 of x0
such that T (U1 ) ⊂ V1 . But then U = U1 − x0 is an open neighborhood of the origin
of X, and if y and z are elements of X such that y − z ∈ U , then
T y − T z = T (y − z) ∈ V1 − T x0 = V .
.
This shows that T is uniformly continuous, and thus (a) implies (d). The proof is
complete.

14.3 The Hahn-Banach Theorem
Definition 14.10 Let V be a vector space. A gauge on X is a function p : V → R

such that
1. p(x + y) ≤ p(x) + p(y)
2. p(λx) = λp(x)
for every x ∈ V , y ∈ V , λ > 0.
Theorem 14.2 (Hahn-Banach, Analytic Form) Suppose p is a gauge on a vector

space V , and suppose that W is a vector subspace of V . Let g : W → R be such
that
g(x) ≤ p(x)
. for all x ∈ W.
Then there exists a linear map1 f : V → R such that g(x) = f (x) for every x ∈ W ,
and
. f (x) ≤ p(x) for all x ∈ V .
Proof We consider the set

)
h : D(h) ⊂ V → R where D(h) is a vector

.P = h subspace of V , h is linear, G ⊂ D(h), h = g .

on W and h(x) ≤ p(x) for every x ∈ D(h)
An order relation is defined on P as follows: h1 ≤ h2 if and only if D(h1 ) ⊂ D(h2 )

and h1 = h2 on D(h1 ). Then P = ∅, since g ∈ P . Let Q ⊂ P be a totally ordered
subset. We set Q = {hi | i ∈ I } and

D(h) =
. D(hi )
i∈I
h(x) = hi (x) if x ∈ D(hi ).
It is easy to check that h ∈ P , and that h is an upper bound of Q. We can apply

Zorn’s Lemma to produce a maximal element f . We claim that D(f ) = V , so that
the proof will be complete.
We assume on the contrary that D(f ) = V , and let x0 ∈ V \ D(f ). We define
D(h) = D(f ) + Rx0 and h(x + tx0 ) = f (x) + tα for every x ∈ D(f ), ∈ R. Here
α is a real number that will be fixed conveniently in a moment. Hence h ∈ P , and
we need to prove that
.f (x) + tα ≤ p(x + tx0 )
for every x ∈ D(f ) and t ∈ R. Since p is a gauge, it is sufficient to prove that
f (x) + α ≤ p(x + x0 )
.
f (x) − α ≤ p(x − x0 ).
1 Here we use linearity in a pure algebraic sense. No reference to any norm is understood.
This is certainly true provided that
. sup {f (y) − p(y − x0 ) | y ∈ D(f )} ≤ α ≤ inf {p(x + x0 ) − f (x) | x ∈ D(f )} .
Such a choice of α is possible because
f (y) − p(y − x0 ) ≤ p(x + x0 ) − f (x)

.
for every x ∈ D(f ), y ∈ D(f ), which follows from
f (x) + f (y) ≤ p(x + y) ≤ p(x + x0 ) + p(y − x0 ).

.
We have just proved that f ≤ h, against the maximality of f . The proof is now
complete.

Corollary 14.1 Let G be a vector subspace of X and let g : G → R be a continuous
linear map whose norm is

g(x)
.*g*G = sup
∗ x ∈ G, x = 0 .
*x*
Then there exists f ∈ X∗ such that f = g on G and *f *X∗ = *g*G∗ .

Proof We just apply Theorem 14.2 with the gauge p(x) = *g*G∗ *x*.

Corollary 14.2 For every x0 ∈ X there exists f0 ∈ X∗ such that
*f0 | = *x0 *,
. f0 (x0 ) = *x0 *2 .
Proof Let G = Rx0 and g(tx0 ) = t*x0 *2 . We have |g*G∗ = *x0 *2 , so that
Corollary 14.1 applies.

Corollary 14.3 For every x ∈ X there results

|f (x)| |f (x)|
f ∈ X∗ .
.*x* = sup
f ∈ X∗ = max
*f *X ∗ |f *X∗
Proof The conclusion is trivial if x = 0, since f (0) = 0 whenever f ∈ X∗ . Assume

x = 0. Clearly

|f (x)|
. sup f ∈ X∗ ≤ *x|.
*f *X∗
On the other hand, there exists f0 ∈ X∗ such that *f0 * = *x* and f0 (x) = *x*2 .
We define f1 = *x*−1 f0 in such a way that *f1 | = 1 and f1 (x) = *x*.

Theorem 14.2 can be also presented in a more geometric way.

Definition 14.11 A hyperplane is a set of the form
H = {x ∈ X | f (x) = α} ,
.
where f : X → R is a linear map2 and α ∈ R. We will always assume that f is not

identically zero.
For the sake of brevity, we will also write
[f = α] = {x ∈ X | f (x) = α} .
.
Theorem 14.3 The hyperplane [f = α] is closed if and only if f ∈ X∗ .

Proof If f ∈ X∗ , then [f = α] is closed as the preimage of the closed subset {α}
in R. Conversely, suppose that H = [f = α] is closed in X. Fix x0 ∈ X \ H , and
suppose for definiteness that f (x0 ) < α. Pick r > 0 such that B(x0 , r) ⊂ X \ H ,
where
B(x0 , r) = {x ∈ X | *x − x0 * < r} .
.
We claim that f (x) < α for all x ∈ B(x0 , r). Indeed, suppose that f (x1 ) > α for
some x1 ∈ B(x0 , r). For every t ∈ [0, 1], the point xt = (1 − t)x0 + tx1 lies in
B(x0 , r), so that f (xt ) = α for every t ∈ [0, 1]. But f (xt ) = α for
f (x1 ) − α
t=
. ,
f (x1 ) − f (x0 )
which is a contradiction. The claim is proved, and it follows that f (x0 + rz) < α
for every z ∈ B(0, 1). As a consequence f ∈ X∗ with
α − f (x0 )
|f *X∗ <
. .
r

Definition 14.12 Let A ⊂ X, B ⊂ X. The hyperplane H = [f = α] separates A
and B if and only if
x ∈ A ⇒ f (x) ≤ α
.
x ∈ B ⇒ f (x) ≥ α.
2 Continuity is not required.

Fig. 14.1 Separation of closed convex subsets
We say that H strictly separates A and B if there exists ε > 0 such that
x ∈ A ⇒ f (x) ≤ α − ε
.
x ∈ B ⇒ f (x) ≥ α + ε.
Figure 14.1 describes the definition of separating hyperplane.

Lemma 14.1 Let C ⊂ X be an open convex subset such that 0 ∈ C. We define the
gauge of C as the function pC : X → R such that

1
pC (x) = inf α > 0 x ∈ C
.
α
for every x ∈ X. Then pC is a gauge. Furthermore there exists M ∈ R such that

0 ≤ pC (x) ≤ M*x* for every x ∈ X, and
. C = {x ∈ X | pC (x) < 1} .
Proof Since C is open, there exists r > 0 such that B(0, r) ⊂ C. By definition,
1
pC (x) ≤
. |x| for every x ∈ X.
r
Hence M = 1/r works. The first property of a gauge is trivial.
Now, assume that x ∈ C. since C is open, we have (1 + ε)x ∈ C for ε > 0
sufficiently small. Hence pC (x) ≤ 1/(1 + ε) < 1. Conversely, if pC (x) < 1 then
there exists 0 < α < 1 such that ∗1/α)x ∈ C and thus
1
x=α·
. x + (1 − α) · 0 ∈ C.
α
Finally, let x ∈ X, y ∈ X, ε > 0. It follows from the previous considerations that

x y
. ∈ C, ∈ C.
pC (x) + ε pC (y) + ε
Therefore
tx (1 − t)y
. + ∈C
pC (x) + ε pC (y) + ε
for every t ∈ [0, 1], since C is convex. Choosing
pC (x) + ε
t=
.
pC (x) + pC (y) + 2ε
we obtain
x+y
. ∈ C.
pC (x) + pC (y) + 2ε
Hence
x+y
pC
. < 1,
pC (x) + pC (y) + 2ε
or pC (x + y) < pC (x) + pC (y) + 2ε, for every ε > 0. We deduce that pC (x + y) ≤

pC (x) + pC (y), and the proof is complete.

Lemma 14.2 Let C ⊂ X be a non-empty, open and convex set, and let x0 ∈ X \ C.
There exists f ∈ X∗ such that f (x) < f (x0 ) for every x ∈ C. In particular, the
hyperplane [f = f (x0 )] separates {x0 } and C.
Proof Without loss of generality, we assume 0 ∈ C (otherwise we replace C by
C − x0 ). We set G = Rx0 and g : G → R such that
.g(tx0 ) = tpC (x0 ),
where pC is the gauge of C. If t > 0, then g(tx0 ) = tpC (x0 ) = pC (tx0 ). Since
0 = pC (0) = pC (x0 − x0 ) ≤ pC (x0 ) + pC (−x0 ) implies pC (−x0 ) ≤ pC (x0 ), it
follows that g(tx0 ) ≤ pC (tx0 ) for every t ≤ 0. In any case, g(x) ≤ pC (x) for every
x ∈ G.
By the Hahn-Banach Theorem, there exists f ∈ X∗ such that f = g on G
and f (x) ≤ pC (x) for every x ∈ X. In particular f (x0 ) = 1. Lemma 14.1 yields
f (x) < 1 for every x ∈ C.

Theorem 14.4 (Hahn-Banach, Geometric Form) Let A ⊂ X and B ⊂ X be

non-empty, convex and disjoint sets.
(a) If A is open, then there exists a closed hyperplane which separates A and B.
(b) If A is closed and B is compact, then there exists a closed hyperplane which
strictly separates A and B.
Proof
(a) We set C = A − B, so that C is convex. Since A ∩ B = ∅ and

C=
. {(A − y) | y ∈ B} ,
/ C and C is open. Lemma 13.2 yields f ∈ X∗ such that f (z) < 0

we see that 0 ∈
for every z ∈ C, i.e. f (x) < f (y) for every x ∈ A, y ∈ B. Pick a number
. sup {f (x) | x ∈ A} ≤ α ≤ sup {f (y) | y ∈ B} ,
and conclude that [f = α] separates A and B.

(b) Let ε > 0 be given. We define
Aε = A + B(0, ε)
.
Bε = B + B(0, ε),
so that Aε and Bε are convex, open and non-empty sets. We claim that if ε > 0 is
sufficiently small, then Aε ∩Bε = ∅. This is the only place where the topological
assumptions on A and B are used. Indeed, if the claim were false, there would
exist sequences εn → 0, xn ∈ A and yn ∈ B such that *xn − yn * ≤ 2ε0 . By
compactness, there would exist a subsequence ynk → x ∈ B, and therefore
xnk → x as well. But A is closed, hence x ∈ A ∩ B = ∅, a contradiction. We
can now use part (a) to produce f ∈ X∗ such that [f = α] separates Aε and Bε .
As a consequence,
f (x + εz) ≤ α ≤ f (y + εz)
.
for every x ∈ A, y ∈ B, z ∈ B(0, 1). It follows that
f (x) + ε*f * ≤ α ≤ f (y) − ε*f *

.
for every x ∈ A, y ∈ B. Since *f * = 0, we see that A and B are strictly

separated by the hyperplane [f = α], and the proof is complete.

A very useful consequence of the Hahn-Banach separation theorem is a sufficient
condition for the (topological) density of a vector subspace.
Corollary 14.4 Let Y be a vector subspace such that Y = X. Then there exists
f ∈ X∗ , f = 0, such that
. f (y) = 0 for every y ∈ Y.
Proof Indeed, pick x0 ∈ X \ Y . We apply the Hahn-Banach Theorem in geometric

form with A = Y and B = {x0 }. There exists f ∈ X∗ , f = 0, such that the
hyperplane [f = α] strictly separates A and B. Hence
f (y) < α < f (x0 )

. for all y ∈ Y,
and this implies f (y) = 0 for every y ∈ Y , since λf (y) < α for every λ ∈ R.

14.4 Baire’s Theorem and Uniform Boundedness
We begin with a general result in metric topology.

Theorem 14.5 (Baire’s Theorem) Let X be a complete metric space, and let
.{Un |n ∈ N} be a sequence of open subsets of X. If each .Un is dense in X, then
.G = n∈N Un is dense in X.
Proof Let .ω be a non-empty open set in X. We will show that .ω ∩ G = ∅. Pick any
point .x0 ∈ ω and a number .r0 > 0 such that
.B(x0 , r0 ) ⊂ ω.
Since .U1 is dense, it must intersect the open set .B(x0 , r0 ), so we may choose a point
x1 ∈ B(x0 , r0 ) ∩ U1 and a number .r1 > 0 such that .0 < r1 < r0 /2 and
.
B(x1 , r1 ) ⊂ B(x0 , r0 ) ∩ U1 .
.
Iterating, we construct sequences .{rn }n and .{xn }n such that

rn
.0 < rn+1 < , B(xn+1 , rn+1 ) ⊂ B(xn , rn ) ∩ Un+1 .
2
It follows immediately that .{xn }n is a Cauchy sequence, and therefore .xn → as
n → +∞. Since .xn+p ∈ B(xn , rn ) for every .n ∈ R and every .p ≥ 0, we obtain in
.
the limit as .p → +∞
∈ B(xn , rn )
. for all n ∈ N.
In particular . ∈ ω ∩ G, and the proof is complete.

14.4 Baire’s Theorem and Uniform Boundedness 307
Corollary 14.5 Let X be a non-empty complete metric space. If .{Xn | n ∈ N} is a

sequence of closed subsets such that
∞

X=
. Xn ,
n=1
then there exists .n0 ∈ N such that .Xn0 has non-empty interior.

Proof Consider .Un = X \ Xn . By assumption . n∈N Un is not dense, so there exists
.n0 such that .Un0 is not dense, and this means that the interior of .Xn0 is non-empty.

We introduce the principle of uniform boundedness in a non-linear setting.
Although this is not the common statement in Functional Analysis, we believe that
it fits into our topological approach to mathematical analysis.
Definition 14.13 Let X be a topological space. A function .f : X → R is lower
semicontinuous if the set .{x ∈ X | f (x) ≤ t} is closed in X for evert .t ∈ R.
Theorem 14.6 (Osgood) Let X be a topological space, and suppose that .{fα | α ∈
A} is any family of real-valued lower semicontinuous functions defined on X. If for
every .x ∈ X there exists .Mx > 0 such that
. sup {fα (x) | α ∈ A} ≤ Mx ,
then there exist a non-empty open set .U ⊂ X and a number .M > 0 such that
. sup {fα (x) | α ∈ A, x ∈ U } ≤ M.
Proof For every .n ∈ N, let
Xn = {x ∈ X | fα (x) ≤ n for every α ∈ A} .

.

Each .Xn is closed because .fα is lower semicontinuous. Moreover, .X = n∈N Xn .
Thus there exists .n0 ∈ N such that the interior of .Xn0 is non-empty. Denoting by U
the interior of .Xn0 , there results
. sup {fα (x) | α ∈ A, x ∈ U } ≤ n0 = M.

Remark 14.6 The meaning of Osgood’s Theorem is that a uniform bound on a
suitable open set U can be deduced from a pointwise bound (recall that .Mx depends
on x).
We now apply Osgood’s Theorem in a linear setting. The role of the open set U
becomes irrelevant as a consequence of linearity.
Theorem 14.7 (Uniform Boundedness Theorem) Suppose that X and Y are

Banach space, and suppose that .{Tα | α ∈ A} is any family of bounded linear
operators. If for every .x ∈ X there exists .Mx > 0 such that
. sup {*Tα x* | α ∈ A} ≤ Mx ,
then there exists .M > 0 such that

. sup *Tα *L(X,Y ) α ∈ A ≤ M.
Proof For every .α ∈ A, the function .fα (x) = *T x* is lower semicontinuous

(and indeed continuous) on X. For every .x ∈ X, .sup {fα (x) | α ∈ A} ≤ Mx . By
Osgood’s Theorem there exist a number .M % > 0 and a non-empty open set .U ⊂ X
such that
. sup {*Tα x* | α ∈ A, x ∈ U } ≤ M % .
In particular, there exist some .x0 ∈ U and some .δ > 0 such that .B(x0 , δ) ⊂ U and
. sup {*T x* | α ∈ A, x ∈ B(x0 , δ)} ≤ M % .
If .y ∈ X, .*y* ≤ 1, then .y + x0 ∈ B(x0 , δ), and so for every .α ∈ A,
.*Tα y* ≤ *Tα (y + x0 )* + *Tα x0 * ≤ 2M % .
Finally, if .z ∈ X, .z = 0, we set
δ
y=
. z.
2*z*
Clearly .*y* ≤ δ, and the previous equation shows that
2*z* 4M %
*T z* =
. *Tα y* ≤ *z*
δ δ
for every .α ∈ A. This implies .*Tα *L(X,Y ) ≤ M = 4M % /δ for every .α ∈ A, and the
proof is complete.

Corollary 14.6 Let X, Y be two Banach spaces, and let .{Tn }n be a sequence of
bounded linear operators from X to Y such that for every .x ∈ X, .Tn x converges as
.n → +∞ to a limit T x. Then there results
(a) .sup {*Tn * | n ∈ N} ∈ R,

(b) .T ∈ L(X, Y ),
(c) .*T * ≤ lim infn→+∞ *Tn *L(X,Y ) .
14.5 The Open Mapping Theorem 309
Proof Statement (a) follows directly from the Uniform Boundedness Theorem, so
there exists .M > 0 such that
(n ∈ N) ∧ (x ∈ X) ⇒ *Tn x* ≤ *x*.
.
Taking the limit as .n → +∞ we find .*T x* ≤ M*x* for every .x ∈ X. Since the
linearity of T follows from the basic algebra of limits, the proof of (b) is complete.
Finally we have
*Tn x* ≤ *Tn *L(X,Y ) · *x*

.
for every .x ∈ X, and (c) follows at once. The proof is complete.

14.5 The Open Mapping Theorem
Theorem 14.8 Let X and Y be Banach spaces. If T : X → Y is a bounded linear

operator and if T is surjective, then there exists c > 0 such that
B(0, c) ⊂ T (B(0, 1)).

.
Proof We claim that there exists c > 0 such that
B(0, 2c) ⊂ T (B(0, 1)).

.
positive integer n we set Xn = nT (B(0, 1)). The surjectivity of T

Indeed, for every
implies that Y = ∞ n=1 Xn , and Baire’s Lemma provides us with an integer n0 such
that Xn0 has non-empty interior. Trivially, T (B(0, 1)) has non-empty interior, and
therefore there exist c > 0 and y0 ∈ Y such that
B(y0 , 4c) ⊂ T (B(0, 1)).

.
Since y0 ∈ T (B(0, 1)) we have by symmetry that −y0 ∈ T (B(0, 1)). It follows that
B(0, 4c) ⊂ T (B(0, 1)) + T (B(0, 1)) = 2T (B(0, 1)),

.
since T (B(0, 1)) is a convex set. The claim is proved.

We now complete the proof via an iterative scheme. Let y ∈ Y such that *y* < c.
We are looking for x ∈ X such that T x = y. The previous claim ensures that for
every ε > 0 there exists z ∈ X such that *z* < 1/2 and *y − T z* < ε. We choose
ε = c/2 and we get z1 ∈ X such that |z1 * < 1/2 and *y − T z1 * < c/2.
Applying the same reasoning to y − T z1 instead of y and with ε = c/2, we get

z2 ∈ X such that *z2 * < 1/4 and *y − T z1 − T z2 * < c/4. In this way we construct
a sequence {zn }n of points of X such that *zn * < 2−n and
c
. *y − T (z1 + · · · + zn )* <
2n
for every n. Hence the sequence xn = z1 +· · ·+zn is a Cauchy sequence which must
converge to some x ∈ X. In particular *x* < 1 and y = T x since T is continuous.

Exercise 14.7 Justify the name of the Open Mapping Theorem: let U an open
subset of X, and pick y0 ∈ T (U ). Prove that there exists a ball B(y0 , δ) ⊂ T (U ).
Hint: let x0 ∈ X such that T x0 = y0 , and fix r > 0 such that B(x0 , r) ⊂ U . Then
y0 + B(0, r) ⊂ T (U ), and the Open Mapping Theorem yields a number c > 0 such
that B(0, rc) ⊂ T (B(0, r)). Hence B(y0 , rc) ⊂ T (U ).
Theorem 14.9 (Isomorphism Theorem) Let X and Y be Banach spaces. If
T : X → Y is a bijective bounded linear operator, then T −1 is a bounded linear
operator.
Proof The statement of the Open Mapping Theorem can be formulated as follows:
for every x ∈ X such that *T x* < c, there results *x* < 1. Hence
1
. *x* < *T x* for every x ∈ X.
c
This clearly means that T −1 is a bounded linear operator. The proof is complete.

Example 14.4 Let X be a Banach space, and let * · *1 , * · *2 be two norms on X.
Suppose that X is complete under both norms, and suppose also that there exists
C ≥ 0 such that
.*x*2 ≤ C*x*1 for every x ∈ X.
Then there exists c > 0 such that
.*x*1 ≤ C*x*2 for every x ∈ X.
In other words, the two norms are equivalent. Indeed, the identity operator between
(X, * · *1 ) and (X, * · *2 ) is bounded and bijective. By the Isomorphism Theorem,
its inverse is also bounded, and the existence of the constant c follows.
Theorem 14.10 (Closed Graph Theorem) Let X, Y be Banach spaces, and
T : X → Y be a linear operator. If
G(T ) = {(x, T x) | x ∈ X}
.
is a closed subset of X × Y , then T is continuous.

14.6 Weak and Weak* Topologies 311
Proof We endow X with two norms
*x*1 = *x* + *T x*,

. *x*2 = *x*.
Since G(T ) is closed (X, * · *1 ) is complete, and trivially * · *2 ≤ * · *1 . Hence

there exists c > 0 such that *x*1 ≤ c*x*2 for every x ∈ X. This implies that
*T x* ≤ c*x* for every x ∈ X, and the proof is complete.

14.6 Weak and Weak* Topologies
A fundamental application of the constructions proposed in Sect. 13.9 yields a new

topology on a Banach space X and a new topology on the dual space .X∗ .
Definition 14.14 Let X be a Banach space, and let .x ∗ ∈ X∗ be an element of the
dual space of X. We consider the function .ϕx ∗ : X → R such that
ϕx ∗ (x) = f (x)
. for every x ∈ X.
The weak topology .σ (X, X∗ ) is the smallest topology on X such that every function
∗ ∗
.ϕx ∗ , .x ∈ X , is continuous.
To be more explicit, we are considering here the setting of Definition 13.45 with
Yα = R for every .α, and .A = X∗ . Hence the weak topology on X is the initial
.
topology induced by the dual space .X∗ .

Theorem 14.11 Let .x0 ∈ X. A neighborhood base at .x0 for the weak topology
σ (X, X∗ ) consists of the sets of the form
.

V = x ∈ X |xi∗ (x − x0 )| < ε for every i ∈ I ,
.
where I is a finite set, .xi∗ ∈ X∗ for every .i ∈ I , and .ε > 0.

Proof Clearly enough, a set of the form

V =
. ϕx−1
∗ ((a i − ε, a i + ε)) i∈I
i
with .ai = xi∗ (x0 ) is an open set for .σ (X, X∗ ). Conversely, let U be a neighborhood
of .x0 for .σ (X, X∗ ). By definition of the initial topology, there exists a neighborhood
W of .x0 such that .W ⊂ U and

.W = ϕx−1
∗ (ωi ) i ∈ I ,
i
where I is a finite set and .ωi is a neighborhood of the number .ai = xi∗ (x0 ) in .R.
Therefore there exists .ε > 0 such that .(ai − ε, ai + ε) ⊂ ωi for every .i ∈ I , and this
implies that .x0 ∈ V ⊂ W ⊂ U . The proof is complete.

Corollary 14.7 A net .{Sn , n ∈ D} in X converges weakly to a point x if and only if
x ∗ (Sn ) → x ∗ (x) for every .x ∗ ∈ X∗ .
.
Theorem 14.12 The weak topology .σ (X, X∗ ) is Hausdorff.

Proof Let .x1 ∈ X, .x2 ∈ X be two distinct points. According to the Hahn-Banach
Theorem, there exists a closed hyperplane which separates .{x1 } and .{x2 } strictly.
Hence there exist .x ∗ ∈ X∗ and .α ∈ R such that
x ∗ (x1 ) < α < x ∗ (x2 ).

.
If we define
∗
U1 = x ∈ X
. x (x) < α
∗
U2 = x ∈ X x (x) > α ,
then .U1 and .U2 are open sets for .σ (X, X∗ ), and clearly .x1 ∈ U1 , .x2 ∈ U2 , .U1 ∩U2 =
∅.

In a similar way we introduce a new topology on the dual space .X∗ . Let us recall
once more that .X∗ is always topologized by the operator norm

*x ∗ * = sup |x ∗ (x)| *x* ≤ 1
. for every x ∗ ∈ X∗ .
We can then consider the bidual space .X∗∗ of X, i.e. the dual of .X∗ , and this space
is endowed with the norm

*x* = sup |x ∗ (x)| x ∗ ∈ X∗ , *x ∗ * ≤ 1
Theorem 14.13 There exists a canonical injection .J : X → X∗∗ such that
*J (x)* = *x*
Proof Indeed, for every .x ∈ X we define the linear bounded map
x ∗ ∈ X∗ → x ∗ (x).
.
In this way we associate to every .x ∈ X a unique element .J (x) ∈ X∗∗ . Concisely,
J (x)(x ∗ ) = x ∗ (x)
. for every x ∈ X and x ∗ ∈ X∗ .
14.6 Weak and Weak* Topologies 313
By direct computation,

*J (x)* = sup |J (x)(x ∗)| *x ∗ * ≤ 1 = sup |x ∗ (x)| *x ∗ * ≤ 1 = *x*
.
as a consequence of Corollary 14.3. The proof is complete.

Important: Warning
In general, .J : X → X∗∗ is not surjective. If it is, the space X is called reflexive. Be
careful that a space is reflexive if and only if J is a bijective map; different bijective
maps may exist between X and .X∗∗ , but they do not enter into the definition of
reflexivity.
Definition 14.15 For every .x ∈ X we consider the application .ϕx : X∗ → R such

that .ϕx (x ∗ ) = x ∗ (x) for every .x ∗ ∈ X∗ . The weak* topology of .X∗ , also denoted
by .σ (X∗ , X), is the smallest topology such that all the applications .ϕx , .x ∈ X, are
continuous.
Once again, the weak* topology is an initial topology with .A = X and .Yα = R
for every .α ∈ A. As before, we may provide a useful characterization of open sets
in .σ (X∗ , X).
Theorem 14.14 Let .x0∗ ∈ X∗ . A neighborhood base at .x0∗ for .σ (X∗ , X) consists of
all sets of the form

V = x ∗ ∈ X∗ |(x ∗ − x0∗ )(xi )| < ε for every i ∈ I
.
where I is a finite set, .xi ∈ X for every .i ∈ I , and .ε > 0.

Theorem 14.15 The topology .σ (X∗ , X) is Hausdorff.
Proof Let .x1∗ ∈ X∗ and .x2∗ ∈ X∗ two distinct points of the dual space .X∗ . By
definition of equality of two functions, there exists at least one point .x ∈ X such
that .x1∗ (x) = x2∗ (x). Without loss of generality, we may assume that .x1∗ (x) < x2∗ (x).
Fix a number .α such that
x1∗ (x) < α < x2∗ (x).

.
If
∗
U1∗ = x ∗ ∈ X∗
. x (x) < α
∗
U2∗ = x ∗ ∈ X∗ x (x) > α ,
then .U1∗ and .U2∗ are disjoint open sets such that .x1∗ ∈ U1∗ and .x2∗ ∈ U2∗ . The proof is
complete.

As an application of Tychonoff’s Theorem, we prove a fundamental result in the

theory of Banach spaces.
Theorem 14.16 (Alaoglu) If X is a Banach space, then the unit ball .B ∗ of .X∗ is
compact with respect to .σ (X∗ , X).
Proof In this proof we denote with B and .B ∗ the unit balls of X and .X∗ ,
respectively. By Tychonoff’s Theorem, the topological space .[−1, 1]B consisting
of functions from B to .[−1, 1] with the product topology is a compact space. We
define the restriction application .R : B ∗ → [−1, 1]B by .R(ψ) = ψ|B for every
∗
.ψ ∈ B . Suppose that
(i) .R(B ∗ ) is a closed subset of .[−1, 1]B , and

(ii) R is a topological homeomorphism from .B ∗ with the weak* topology onto
∗
.R(B ) with the product topology.
Once (i) and (ii) have been established, we deduce that .R(B ∗ ) is compact. Hence
∗ ∗
.B must be compact as a homeomorphic copy of .R(B ). It remains to prove that (i)
and (ii) hold.

First of all, R is injective: if .R(ψ) = R(η) for some .ψ ∈ B ∗ , .η ∈ B ∗ such that
.ψ = η, then there exists a point .x ∈ B such that .ψ(x) = η(x). Hence .R(ψ) =
R(η). The fact that R is a homeomorphism of .B ∗ onto .R(B ∗ ) follows from a direct
comparison of the basic neighborhoods in the weak* topology and in the product
topology. To conclude, we need to prove that .R(B ∗ ) is a closed subset of .[−1, 1]B
in the product topology.
Let .f : B → [−1, 1] be a point of the closure of .R(B ∗ ) in the product topology.
By definition of the application R, we only need to check that for every .u ∈ B,
.v ∈ B, .λ ∈ R such that .u + v ∈ B and .λu ∈ B, there results
f (u + v) = f (u) + f (v),
. f (λu) = λf (u).
Now, for every .ε > 0, the weak* neighborhood consisting of those .g ∈ [−1, 1]B
such that
. |g(u) − f (u)| < ε

|g(v) − f (v)| < ε
|g(u + v) − f (u + v)| < ε
contains some element .R(ψε ), and since .ψε is linear, we must have
. |f (u + v) − f (u) − f (v)| < 3ε.
The proof that .|f (λu) − λf (u)| < 2ε is similar, and thus .f ∈ R(B ∗ ). The proof is
complete

14.7 Isomorphisms 315
14.7 Isomorphisms
Theorem 14.17 Let X be a Banach space, and let T ∈ L(X, X) be such that
*T * < 1. Then the operator I − T is invertible, and its inverse operator belongs to
L(X, X).
Proof Consider the series
∞

. Tn = I + T + T2 + ···+ Tn + ··· ,
n=0
where T n = T ◦ T n−1 for n = 2, 3, . . . Since *T

* ≤ *Tn*3 , this series converges
n n
in X by comparison with the geometric series ∞ n=0 *T * . We call S the sum of

this series,
and we remark that S ∈ L(X, X). Then S ◦ T = T ◦ S is the sum of the
series ∞ n=1 T n , and
S ◦ (I − T ) = (I − T ) ◦ S = I,
.
and this implies that S is the inverse operator of I − T . The proof is complete.

Definition 14.16 If X and Y are Banach spaces, we denote by Iso(X, Y ) the subset
of L(X, Y ) consisting of all T such that T is invertible and the inverse of T belongs
to L(Y, X).
Remark 14.7 Members of Iso(X, Y ) are usually called isomorphisms between X
and Y . In abstract algebra the word isomorphism refers to invertible functions which
preserves some prescribed algebraic structure in X and in Y . In our definition we
add a topological condition, i.e. the continuity of T and of its inverse, and we should
look for a less generic word. It must be said that continuity is somehow the smallest
property that we want to preserve besides linearity in Functional Analysis, and for
this reason we force the word isomorphism to include the continuity preservation.
Theorem 14.18 Let X and Y be Banach spaces.
(a) Iso(X, Y ) is an open subset of L(X, Y ).
(b) The function T → T −1 of Iso(X, Y ) to L(Y, X) is continuous.
Proof Since the empty set is always open, we will assume that Iso(X, Y ) = ∅. Pick
T0 ∈ Iso(X, Y ). For T : X → Y to be an isomorphism it is necessary and sufficient
that T0−1 ◦ T : X → Y be an isomorphism. We set T0−1 ◦ T = I − . If we can
* * * *
3 More precisely, * nk=m T k * ≤ nk=m *T k * ≤ nk=m *T *k , and the conclusion follows from
the completeness of L(X, Y ).
ensure that ** < 1, then Theorem 14.17 implies that is an isomorphism. Since
= I − T0−1 ◦ T = T0−1 ◦ (T0 − T ),

.
we have ** ≤ *T0−1 **T − T0 *. Therefore, if
1
*T − T0 * <
. ,
*T0−1 *
then ** < 1 and also T is an isomorphism. This proves (a).

To prove (b), we remark that
T −1 = (T0 ◦ (I − T ))−1 = (I − )−1 ◦ T0−1 ,

.
hence
T −1 − T0−1 = [(I − )−1 − I ] ◦ T0−1 .

.
∞ ∞
But (I − )−1 = n
n=0 , hence (I − )−1 − I = n
n=1 , and
* * ∞
||
* *
. *(I − )−1 − I * ≤ **n = .
1 − **
n=1
This implies that

* * ||
* −1 *
. *T − T0−1 * ≤ *T0−1 * .
1 − **
As T → T0 , ** → 0, and thus T −1 → T0−1 . This proves that T −1 is a continuous

function of T if T remains in Iso(X, Y ). The proof is complete.

14.8 Continuous Multilinear Applications
Let .X1 , . . . , Xn and Y be vector spaces. A function .f : X1 × · · · × Xn → Y is

multilinear if and only if for every .k ∈ [1, n] ∩ N and every system .ai ∈ Xi , .i = k,
of vectors, the application
x → f (a1 , . . . , ak−1 , x, ak+1 , . . . , an )

.
is linear from .Xk to Y . Roughly speaking, multilinear means linear in each variable
separately.
14.8 Continuous Multilinear Applications 317
Exercise 14.8 If f is multilinear, prove that
f (λ1 x1 , . . . , λn xn ) = λ1 λ2 · · · λn f (x1 , . . . , xn )
.
for every .λi ∈ R, .i = 1, . . . , n.

Suppose now that all vector spaces are Banach spaces. How can we describe the
continuity of a multilinear application? The answer is similar to the case of a linear
operator.
Theorem 14.19 Let .X1 , . . . , Xn and Y be Banach spaces, and let .f : X1 × · · · ×
Xn → Y be multilinear. The following statements are equivalent:
(a) f is continuous at any point;
(b) f is continuous at .(0, . . . , 0);
(c) the set
. {*f (x1 , . . . , xn )* | *xi * ≤ 1 for i = 1, . . . , n}
is bounded.
Proof Clearly (a) implies (b). Suppose that (b) holds. Since f is continuous at
the origin, the pre-image of the unit ball of Y is an open neighborhood of the
origin in .X1 × · · · × Xn . Hence there exists .r > 0 such that .|xi * ≤ r for every
i implies .*f (x1 , . . . , xn )* ≤ 1. By homogeneity, .*xi * ≤ 1 for every i implies
.*f (x1 , . . . , xn )* ≤ r
−n . Thus (c) holds.
Assume now that (c) holds, and let .M > 0 be such that .*f (x1 , . . . , xn )* ≤ M
whenever .*xi * ≤ 1 for every i. By homogeneity, for every .x1 , . . . , xn ,
*f (x1 , . . . , xn )* ≤ M*x1 * · · · *xn *.

.
Let .(a1 , . . . , an ) be a given point. We write
f (x1 , . . . , xn ) − f (a1 , . . . , an )
.
= f (x1 − a1 , x2 , . . . , xn ) + f (a1 , x2 − a2 , x3 , . . . , xn )
+ · · · + f (a1 , . . . , an−1 , xn − an ).
Hence
*f (x1 , . . . , xn ) − f (a1 , . . . , an )*
.
≤ M*x1 − a1 **x2 * · · · *xn * + M*x2 − a2 **a2 **x3 * · · · *xn *

+ · · · + M*xn − an **a1 * · · · *an−1 *.
Suppose that .*xi − ai | ≤ ε for every i. It follows that .*xi * ≤ *ai * + ε, and there
exists .A > 0 such that .*xi − ai | ≤ ε for every i implies .*xi * ≤ A for every i. We
deduce that
" n #

.*f (x1 , . . . , xn ) − f (a1 , . . . , an )* ≤ MA *xi − ai * ≤ nMAn−1 ε.
n−1
i=1
whenever .*xi − ai * ≤ ε for every i. Since we may choose .A > 0 which does
not depend on .ε provided that .ε > 0 is sufficiently small, the continuity of f at
.(a1 , . . . , an ) follows.

Definition 14.17 If .X1 , . . . , Xn and Y are Banach spaces, we denote by
.L(X1 , . . . , Xn ; Y )
the set of all continuous multilinear applications from .X1 × · · · × Xn to Y . For any
f ∈ L(X1 , . . . , Xn ; Y ) we define the norm
.

*f * = sup |f (x1 , . . . , xn ) *xi *Xi ≤ 1 for i = 1, . . . , n .
.
Exercise 14.9 Prove that .L(X1 , . . . , Xn ; Y ) is a Banach space with respect to the
norm just defined.
Example 14.5 Consider three Banach spaces X, Y and Z. We define .ϕ : L(Y, Z) ×
L(X, Y ) → L(X, Z) such that .ϕ(f, g) = g ◦ f . It is (almost) obvious that .ϕ is
bilinear. For every .x ∈ X we notice that
*g ◦ f (x)* = *g(f (x))* ≤ *g* · *f (x)*

.
≤ *g* · *f * · *x*.
Hence .ϕ is continuous, and .*ϕ* ≤ 1.

We now introduce a canonical isomorphism which we will exploit in order to
interpret in different ways the definition of derivatives of order higher than one.
Let X, Y and Z be three Banach spaces. We define
ϕ : L(X, Y ; Z) → L(X, L(Y, Z))

.
in the following way: any .f ∈ L(X, Y ; Z) is a bilinear application .f (x, y) of the

variables .x ∈ X and .y ∈ Y . If we keep x fixed, the application .fx : y → f (x, y) is
linear from Y to Z. Furthermore
*fx * = *f (x, y)* ≤ *f * · *x* · *y*,

.
14.8 Continuous Multilinear Applications 319
so that .*fx * ≤ *f * · *x*. This inequality proves the continuity of .fx , and allows
us to define the application .g : x ∈ X → fx ∈ L(Y, Z). The previous inequality
becomes now
*g(x)* ≤ *f * · *x*.
.
Therefore .*g* ≤ *f *.
To summarize, to each .f ∈ L(X, Y ; Z) we have associated an application
.g : X → L(Y, Z), which we may denote by .ϕ(f ). It is immediate to check that
.ϕ is linear and that .*ϕ* ≤ 1. Our next step consists in constructing an inverse of .ϕ,
namely a continuous application
.ψ : L(X, L(Y, Z)) → L(X, Y ; Z)
which inverts .ϕ. We start with .g : X → L(Y, Z), and we notice that g associates to
each .x ∈ X a bounded linear application .g(x) from Y to Z. Hence
f : (x, y) ∈ X × Y → g(x)(y)
.
is bilinear from .X × Y to Z. Moreover
. *g(x)* ≤ *g* · *x*,
which implies
.*f (x, y)* = *g(x)(y)* ≤ *g(x)* · *y* ≤ *g* · *x* · *y*.
This shows that .f ∈ L(X, Y ; Z) and .*f * ≤ *g*. We summarize as follows: the
application .ψ associates to each .g ∈ L(X, L(Y, Z)) the application .f ∈ L(X, Y ; Z)
in such a way that .*ψ* ≤ 1.
Theorem 14.20 There exists an isometry4 between .L(X, Y ; Z) and .L(X, L(Y, Z)).
Proof It is clear that .ϕ ◦ ψ and .ψ ◦ ϕ are the identities in the corresponding spaces.
In particular the operator norm of .ψ ◦ ϕ must be equal to one. Hence .1 = *ψ ◦ ϕ* ≤
*ψ* · *ϕ*, and the fact that .*ϕ* ≤ 1 and .*ψ* ≤ 1 implies .*ψ* = 1, .*ϕ* = 1. This
proves that .ϕ is a bounded linear operator which preserves norms, and the proof is
complete.

4 A bounded linear application with bounded inverse, which preserves norms.

14.9 Inner Product Spaces
Theorem 14.21 (Cauchy-Schwarz Inequality) Let H be an inner product space.

For every u ∈ H and v ∈ H , the following inequality holds:

. |"u, v#| ≤ "u, u# "v, v#.
Proof Let us define p : R → R such that p(λ) = "λu + v, λu + v#. From the
definition of the inner product it follows that p(λ) ≥ 0 for every λ. But
p(λ) = "u, u#λ2 + 2"u, v#λ + "v, v#.

.
We thus see that p(λ) is a non-negative polynomial of degree two in λ, hence it

cannot have real zeroes. As a consequence

4 "u, v#2 − "u, u#"v, v# < 0,
.
which immediately gives the conclusion.

√√
Exercise 14.10 Deduce from the previous proof that |"u, v#| = "u, u# "v, v#
holds if and only if the vectors u and v are linearly dependent. Hint: the discriminant
of the polynomial p must be zero in this case.
Theorem 14.22 (Triangle Inequality) Let H be an inner product space. For every
u ∈ H and v ∈ H , the following inequality holds:

."u + v, u + v# ≤ "u, u# + "v, v#.
Proof Indeed,
0 ≤ "u + v, u + v# = "u, u# + 2"u, v# + "v, v#

.

≤ "u, u# + 2 "u, u# "v, v# + "v, v#
2
= "u, u# + "v, v# .

Definition 14.18 (Norms in Inner Product Spaces) The norm induced by the
inner product of a space H is defined by

. *u* = "u, u# (14.2)
for every u ∈ H .
14.9 Inner Product Spaces 321
The fact that this is actually a norm on H follows at once from the triangle
inequality and the algebraic properties of the inner product. Unless otherwise stated,
the norm of an inner product space will refer to (14.2).
Theorem 14.23 (Parallelogram Identity) If H is an inner product space, and if
u ∈ H , v ∈ H , then
* * * *
* u + v *2 * u − v *2
* = *u* + *v* .
2 2
* * + *
.
* 2 * * 2 * 2
Proof The identity follows from an expansion of the left-hand side according to the
bilinearity properties of the inner product. The details are left as an exercise.

Definition 14.19 (Hilbert Spaces) An inner product H is a Hilbert space if and
only if it is a complete metric space with respect to the distance d(u, v) = *u − v*
associated to the norm (14.2).
Theorem 14.24 (Projection on Closed Convex Subsets) Let H be a Hilbert
space, and let K be a closed convex subset of H . For every f ∈ H there exists
one and only one element u ∈ K such that
*f − u* = min {*f − v* | v ∈ K} .
.
Furthermore, u ∈ K is characterized by the properties
u ∈ K,
.
"f − u, v − u# ≤ 0 for every v ∈ K.
Proof Let us set d = min {*f − v* | v ∈ K}, and let {vn }n be a sequence in K such
that dn = |f − vn * → d as n → +∞. The parallelogram identity shows that
* *2 * *
* vn + vm * * vn − vm *2
* = dn + dm
2 2
*
. f − * *
+*
* 2 * 2 * 2
* *
for every n, m. Since K is convex, vn +vm
2 ∈ K, and therefore *f − vn +vm *2
2 ≥ d 2.
Hence
* *
* vn − vm *2 dn2 + dm
2
* * ≤ − d 2.
.
* 2 * 2
This shows that *vn − vm *2 can be made as small as we please by choosing n and
m sufficiently large. In other words, {vn }n is a Cauchy sequence in H . Since H is a
complete metric space, vn → u as n → +∞, and u ∈ K because K is closed. But
then d = *f − u*, and we have proved the existence of the desired element of K.
Suppose now that u ∈ K satisfies *f − u* = min {*f − v* | v ∈ K}, and fix

w ∈ K. For every t ∈ (0, 1] we have v = (1 − t)u + tw ∈ K, hence
*f − u* ≤ *f − ((1 − t)u + tw)| = *(f − u) − t (w − u)*.

.
It follows that
|f − u*2 ≤ *f − u*2 − 2t"f − u, w − u# + t 2 |w − u*2 .

.
Simplifying we see that 2"f − u, w − u# ≤ t*w − u*2 , and we conclude by letting

t → 0.
Conversely, suppose that u ∈ K and "f − u, v − u# ≤ 0 for every v ∈ K. Then
*u − f *2 − *v − f *2 = 2"f − u, v − u# − *u − v*2 ≤ 0
.
for every v ∈ K, and this shows that *f − u* = min {*f − v* | v ∈ K}. To

complete the proof, we must now show that the element u ∈ K is unique. Assume
that u1 ∈ K and u2 ∈ K satisfy "f − u1 , v − u1 # ≤ 0 and "f − u2 , v − u2 # ≤ 0 for
every v ∈ K. In particular
"f − u1 , u2 − u1 # ≤ 0
.
"f − u2 , u1 − u2 # ≤ 0
and thus *u1 − u2 *2 ≤ 0, Hence u1 = u2 .

Definition 14.20 (Projection Operator) With the same notation as in Theo-
rem 14.24, we call u ∈ K the projection of f on K, and we denote it with the
symbol PK f .
The condition
"f − u, v − u# ≤ 0
. for every v ∈ K
is a condition about the angle between the vectors f − u and v − u. If K is a vector

subspace of H , it is not surprising that this condition becomes stronger.
Theorem 14.25 Let H be a Hilbert space.
(i) Suppose that K is a closed convex subset of H . If f1 ∈ H and f2 ∈ H , then
. *PK f1 − Pk f2 * ≤ *f1 − f2 * .
(ii) If K is a closed subspace of H and f ∈ H , then u = PK f is characterized by
u∈H
.
"f − u, v# = 0 for every v ∈ K.
In this case, PK is a linear operator.

Proof Writing u1 = PK f1 , u2 = PK f2 , we have by definition
"f1 − u1 , v − u1 # ≤ 0 for every v ∈ K

.
"f2 − u2 , v − u2 # ≤ 0 for every v ∈ K.
Choosing v = u2 in the first inequality and v = u1 in the second inequality, we see

that
*u1 − u2 *2 ≤ "f1 − f2 , u1 − u2 #.
.
The Cauchy-Schwarz inequality yields *u1 − u2 * ≤ *f1 − f2 *. This proves (i).

To prove (ii), we start with
"f − u, v − u# ≤ 0 for every v ∈ K.

.
Since K is a vector space, λv ∈ K for every λ ∈ R, so that
"f − u, λv − u# ≤ 0 for every v ∈ K and every λ ∈ R.

.
As a consequence, "f − u, v# = 0 for every v ∈ K. Suppose conversely that u ∈ K

and "f − u, v# = 0 for every v ∈ K. Then "f − u, v − u# = 0 for every v ∈ K, and

Theorem 14.26 (Riesz-Fréchet) Let H be a Hilbert space. For every ϕ ∈ H ∗ there
exists a unique f ∈ H such that "f, v# = ϕ(v) for every v ∈ H . In addition,
*f * = *ϕ*H ∗ .
Proof Let us set K = ker ϕ = ϕ −1 ({0}). If K = H , then ϕ = 0 identically, and
clearly f = 0 suffices. We may now suppose that H \ K = ∅, so that there exists
g0 ∈ H such that g0 ∈/ K. We define g1 = PK g0 and
g0 − g1
g=
. .
*g0 − g1 *
It follows immediately that *g* = 1, g ∈

/ K, and "g, w# = 0 for every w ∈ K.
Now, every element v ∈ H can be decomposed as v = λg + w for some λ ∈ R
and some w ∈ K: indeed,
ϕ(v)
λ=
. , w = v − λg.
ϕ(g)
From the properties of g it follows that
0 = "g, w# = "g, v − λg#,

.
or
ϕ(v)
"g, v# = λ =
. .
ϕ(g)
We conclude by defining f = ϕ(g)g.

The Riesz-Fischer Theorem is a representation theorem, since it provides a
complete description, up to isometries, of the dual space of the Hilbert space H .
We conclude this section with an important result, due to the Italian mathemati-
cian Guido Stampacchia, which is of great importance in the analysis of Partial
Differential Equations. Before stating and proving it, we introduce a celebrated
result about fixed points in complete metric spaces.
Theorem 14.27 (Banach-Caccioppoli) Let (X, d) be a complete metric space. If
T : X → X is a contractive map, i.e. there exists a number 0 ≤ L < 1 such that
d(T (x), T (y)) ≤ Ld(x, y) for every x ∈ X, y ∈ X, then T has one and only one
fixed point z ∈ X such that T (z) = z.
Proof The case L = 0 is trivial. Suppose now 0 < L < 1, and pick any point
x0 ∈ X, and define recursively xn+1 = T (xn ) for n = 1, 2, . . . An easy induction
argument shows that
d(xn+1 , xn ) ≤ Ln d(x1 , x0 )
.
for every n ∈ N, n ≥ 1. Let m > n be two integers. Then
.d(xm , xn ) ≤ d(xm , xm−1 ) + d(xm−1 , xm−2 ) + · · · + d(xn+1 , xn )

≤ Lm−1 d(x1 , x0 ) + Lm−2 d(x1, x0 ) + · · · + Ln d(x1 , x0 )

m−n−1
= L d(x1 , x0 )
n
Lk
k=0
∞

≤ Ln d(x1 , x0 ) Lk
k=0
Ln
= d(x1, x0 ) .
1−L
If d(x1 , x0 ) = 0, then x1 = T (x0 ) = x0 and x0 is a fixed point of T . Otherwise, for

every ε > 0 there exists N ∈ N such that
ε(1 − L)
LN <
. .
d(x1 , x0 )
Hence, if m > N and n > N, we see that
Ln ε(1 − L) d(x1 , x0 )
d(xm , xn ) ≤ d(x1 , x0 )
. < = ε.
1−L d(x1 , x0 ) 1 − L
Therefore {xn }n is a Cauchy sequence, and by completeness it must converge to

some point z ∈ X. But
z = lim xn = lim T (xn−1 ) = T

. lim xn−1 = T (z),
n→+∞ n→+∞ n→+∞
since T is continuous. This proves the existence of a fixed point for T . Uniqueness
is easy, since T (z1 ) = z1 and T (z2 ) = z2 imply d(z1 , z2 ) = d(T (z1 ), T (z2 )) <
Ld(z1 , z2 ), so that d(z1 , z2 ) = 0. The proof is complete.

Theorem 14.28 (Stampacchia) Let H be a Hilbert space, and let a : H × H → R
a continuous bilinear form. We assume that there exists α > 0 such that
a(u, u) ≥ α*u*2
. for every u ∈ H.
Let K be a non-empty closed convex subset of H . For every ϕ ∈ H ∗ there exists a

unique element u ∈ K such that
a(u, v − u) ≥ ϕ(v − u)
. for every v ∈ K.
If a(v, w) = a(w, v) for every v and w, then u is characterized by the variational

property
⎧
⎨u ∈ K

.
⎩ 12 a(u, u) − ϕ(u) = min
1
2 a(v, v) − ϕ(v) v ∈ K .
Proof We may represent ϕ ∈ H ∗ by a unique element f ∈ H in the sense that

ϕ(v) = "f, v# for every v ∈ H . Similarly, for every u ∈ H , the function v →
a(u, v) is linear and continuous on H , hence there exists a unique element Au ∈ H
such that a(u, v) = "Au, v# for every v ∈ H . It is easy to check that A : H → H is
linear, continuous, and
"Au, u# ≥ α*u|u2
. for every u ∈ H.
To complete the proof, it suffices to show that there exists a unique u ∈ K such that
."Au, v − u# ≥ "f, v − u# for every v ∈ K.
Let ! > 0 be a number that will be chosen hereafter. We can rewrite our inequality
as
"!f − !Au + u − u, v − u# ≤ 0 for every v ∈ K.

.
But this just means that
u = PK (!f − !Au + u).

.
This is a fixed point problem. More precisely, we are looking for a fixed point u ∈ K
of the function S : v ∈ K → PK (!f − !Av + v). We may now play with the free
parameter ! > 0 to ensure that S is a strict contraction on the complete metric space
K.
Indeed, we already know that
. *Sv1 − Sv2 * ≤ *(v1 − v2 ) − !(Av1 − Av2 )* ,
so that
. *Sv1 − Sv2 *2 ≤ *v1 − v2 *2 − 2!"Av1 − Av2 , v1 − v2 # + !2 *Av1 − Av2 *2

≤ *v1 − v2 *2 1 − 2!α + *A*2L(H,H ) !2 .
We now choose
2α
0<!<
. ,
*A*L(H,H )
so that 1 − 2!α + *A*2L(H,H ) !2 < 1. We may now apply the Banach-Caccioppoli

Theorem 14.27 to conclude that S has a unique fixed point u ∈ K.
We suppose now that the bilinear form a is symmetric. In particular, a is an
inner product on H which induces a norm equivalent to the original one. Hence H
is a Hilbert space with respect to the inner product a. The Riesz-Fréchet Theorem
yields g ∈ H such that ϕ(v) = a(g, v) for every v ∈ H . The relation a(u, v − u) ≥
ϕ(v − u) for every v ∈ K reduces to
a(g − u, v − u) ≤ 0
. for every v ∈ K,
14.10 Linearization in Normed Vector Spaces 327
or u = PK g, where the projection is understood in the sense of the inner product a.

As we know, the element u is a solution to the minimization problem

. min a(g − v, g − v) v ∈ K ,
which is equivalent to the minimization on K of a(g − v, g − v), which is in turn

equivalent to the minimization of
1
. a(v, v) − ϕ(v)
2
with respect to v ∈ K. The proof is complete.

Theorem 14.29 (Lax-Milgram) Let H be a Hilbert space, a : H × H → R be a
continuous bilinear form. Suppose that there exists α > 0 such that
a(u, u) ≥ α*u*2
.
for every u ∈ H . For every ϕ ∈ H ∗ there exists a unique element u ∈ H such that
a(u, v) = ϕ(v)
. for every v ∈ H.
Furthermore, if a is symmetric, i.e. a(v, w) = a(w, v) for every v and w, then the
element u is characterized by
⎧
⎨u ∈ H

⎩ 12 a(u, u) − ϕ(u) = min 12 a(v, v) − ϕ(v) v ∈ H .
.
Proof The conclusion follows from Stampacchia’a Theorem and Theorem 14.25.

14.10 Linearization in Normed Vector Spaces
Let X and Y be Banach spaces. In the very particular case .X = Y = Rn , the notion
of derivative has been studied in the first part of the book. As we said there, it is
customary to think of the derivative of a function f at a point a as a real number,
defined as
f (x) − f (a)
f % (a) = lim
. ,
x→a x−a
provided that this limit exists in .R. In the general case we need to avoid the division
by .x − a, which is now a vector.
Definition 14.21 Let U be an open subset of X. Two functions .f1 : U → Y and

f2 : U → Y are tangent at the point .a ∈ U if and only if the function
.
m(r) = sup {*f1 (x) − f2 (x)* | *x − a* ≤ r} ,

.
which is defined for any .r > 0 sufficiently small, satisfies
m(r)
. lim = 0.
r→0+ r
Exercise 14.11 We say that .f1 ∼ f2 if and only if .f1 and .f2 are tangent at a. Prove
that this is an equivalence relation.
Exercise 14.12 Prove that if .f1 and .f2 are tangent at a, then .f1 −f2 is continuous at
a. If, in particular, .f1 is continuous at a, then .f2 is continuous as well, and .f1 (a) =
f2 (a).
Example 14.6 Let g be a linear function .X → Y , and let .f (x) = g(x − a). It is
clear that f is tangent to zero at a if and only if .*g* = 0, i.e. if and only if g is
identically zero. Indeed,
m(r) = *g*r.
.
The previous example proves an important result.

Theorem 14.30 If .f : U → Y and .a ∈ U , then there exists at most one linear
application .g : X → Y such that
x → f (x) − f (a)
.
and
x → g(x − a)
.
are tangent at a. If such a g exists, it is continuous at zero if and only if f is

continuous at a.
Definition 14.22 Let U be open in X. A function .f : U → Y is differentiable in
the sense of Fréchet (or F-differentiable) at the point .a ∈ U if and only if
(i) f is continuous at a, and
(ii) there exists a linear application .g : X → Y such that .x → f (x) − f (a) and
.x → g(x − a) are tangent at a.
In this case, the application g is the derivative of f at a, and it is denoted by one of

the symbols .f % (a), .Df (a), .df (a).
Finally, the function f is differentiable on U if and only if it is differentiable at
any point of U .
If f is differentiable at the point a, the derivative .f % (a) is a continuous linear

operator. The standard notation for the value of .f % (a) at .h ∈ X should be
either .f % (a)(h) or .f % (a)h. The first one is the ordinary symbol for functions,
while the second one is the common symbol for linear operators. Some
authors also use .f % (a)[h].
Exercise 14.13 Sometimes the definition of differentiability is stated as follows: f

is differentiable at a if and only if there exists .g ∈ L(X, Y ) such that
*f (x) − f (a) − g(x − a)* = o(*x − a*) as x → a.

.
Prove that these definitions are actually equivalent.

If f is differentiable on U , a new map is defined by
f % : U → L(X, Y ),
. x → f % (x).
This is the derivative of f . Recalling that .L(X, Y ) is a Banach space with respect
to the operator norm, the following definition makes sense.
Definition 14.23 The function .f : U → Y is of class .C 1 (U ) if and only if
%
.f : U → L(X, Y ) is a continuous function.
5
A very natural question is whether the differentiability property depends on the

norms used to topologize X and Y .
Definition 14.24 Two norms .* · * and .* · *1 on a vector space X are equivalent if
and only if there exists a constant .M > 0 such that
1
. *x* ≤ *x*1 ≤ M*x*
M
for every .x ∈ X.
Exercise 14.14 Prove that equivalent norms induce the same topology on X. Hint:
any ball for the first norm contains and is contained in a ball for the second norm.
Hence both norms produce the same open neighborhoods.
Theorem 14.31 Suppose that .* · *1 is an equivalent norm on X and that .* · *2 is
an equivalent norm on Y . A function .f : U → Y is differentiable at .a ∈ U with
respect to the norms .* · *1 and .* · *2 if and only if it is differentiable with respect to
the original norms.
5 It is then understood that f is differentiable on U .

Proof Since there is a perfect symmetry between the old and the new norms, it is
enough to prove that the differentiability with respect to the original norms implies
the differentiability with respect to the equivalent norms. Assume therefore that
there exists .g ∈ L(X, Y ) such that
*f (x) − f (a) − g(x − a)*

. lim = 0.
x→a *x − a*
By assumption
1 1
. ≤M .
*x − a*1 *x − a*
Since .* · *2 is an equivalent norm on Y , there exists .M % > 0 such that
*f (x) − f (a) − g(x − a)*2 ≤ M % *f (x) − f (a) − g(x − a)*.

.
Hence
*f (x) − f (a) − g(x − a)*2 *f (x) − f (a) − g(x − a)*
. ≤ M · M% .
*x − a*1 *x − a*
The conclusion is now immediate.

Exercise 14.15 Suppose that f and g are differentiable at a. Prove that .f + g is
differentiable at a, and fg is differentiable at a for any .k ∈ R. How do you express
the derivatives of .f + g and fg in terms of .f % (a) and .g % (a)?
The last exercise shows that differentiability passes through sums and products.
Quotients are troublesome, since it is forbidden to divide a vector by a vector.
Theorem 14.32 (Chain Rule) Suppose X, Y and Z are Banach spaces, that U is
open in X, V is open in Y , and .a ∈ U . Let .f : U → Y and .g : V → Z be continuous
functions, and assume that .b = f (a) ∈ V . Hence the composition .g ◦ f is defined
on an open neighborhood .U % of a. If f is differentiable at a and if g is differentiable
at b, then .h = g ◦ f is differentiable at a, and there results
h% (a) = g % (a) ◦ f % (a).

.
Proof By assumption
f (x) = f (a) + f % (a)(x − a) + ϕ(x − a)

. (14.3)
where .*ϕ(x − a)* = o(*x − a*) as .x → a. Similarly,
g(y) = g(b) + g % (b)(y − b) + ψ(y − b)

.
where .*ψ(y − b)* = o(*y − b*) as .y → b. Putting these identities together we

have
h(x) − h(a) = g % (f (a))(f (x) − f (a)) + ψ(f (x) − f (a)).

.
Replacing .f (x) − f (a) by its value from (14.3) we get
h(x) − h(a) = (g % (f (a)) ◦ f % (a))(x − a)

.
+ g % (f (a))(ϕ(x − a))
+ ψ(f (x) − f (a)).
We need to prove that the second and the third line satisfy
*g % (f (a))(ϕ(x − a))* = o(*x − a*)

.
*ψ(f (x) − f (a))* = o(*x − a*)
as .x → a. The first estimate follows at once from
*g % (f (a))(ϕ(x − a))* ≤ *g % (f (a))* · *ϕ(x − a)*.

.
The second estimates follows from the fact that .*ψ(f (x) − f (a))* = o(*f (x) −
f (a)*) and .*f (x) − f (a)* ≤ 2*f % (a)* · * − a* holds as long as x is sufficiently
close to a. The proof is complete.

Example 14.7 If U is open in X and if .f : U → Y is the restriction of a continuous
linear application, then f is differentiable and .f % (x) = f for every .x ∈ U . Indeed,
.f (x) − f (a) = f (x − a) = f (f − a) + 0 by linearity, and the conclusion follows.
Theorem 14.33 Let .ϕ : Iso(X, Y ) → Iso(Y, X) be such that .ϕ(u) = u−1 for every
u. Then .ϕ ∈ C 1 (Iso(X, Y )), and
ϕ % (u) : h ∈ L(X, Y ) → −u−1 ◦ h ◦ u−1 .

.
Proof We already know that .Iso(X, Y ) is open in .L(X, Y ). We can also consider .ϕ
as a map into .L(Y, X). Let us fix .u ∈ Iso(X, Y ) and .h ∈ L(X, Y ). We have
ϕ(u + h) − ϕ(u) = (u + h)−1 − u−1

.
= (u + h)−1 ◦ (u − (u + h)) ◦ u−1

= −(u + h)−1 ◦ h ◦ u−1 .
It suffices to prove that the difference between .(u + h)−1 ◦ h ◦ u−1 and .u−1 ◦ h ◦ u−1
is .o(*h*). Now,
(u + h)−1 ◦ h ◦ u−1 − u−1 ◦ h ◦ u−1 = ((u + h)−1 − u−1 ) ◦ h ◦ u−1 ,

.
whence
*(u + h)−1 ◦ h ◦ u−1 − u−1 ◦ h ◦ u−1 * ≤ *(u + h)−1 − u−1 **h**u−1 *

.
We claim that .*(u + h)−1 − u−1 * → 0 as .*h* → 0. Indeed, this is a straightforward

consequence of Theorem 14.18.
To prove that .ϕ is of class .C 1 , we must prove that the function
ϕ % : Iso(X, Y ) → L(L(X, Y ), L(Y, Y ))

.
is continuous. We will use a convenient notation: for every .v ∈ L(Y, X) and .w ∈

L(Y, X), we let
ψ(v, w) : h ∈ L(X, Y ) → −v ◦ h ◦ w ∈ L(Y, X).

.
We have proved above that .ϕ % (u) = ψ(u−1 , u−1 ). The function .(v, w) → ψ(v, w)
is bilinear from .L(Y, X) × L(X, Y ) to .L(L(X, Y ), L(Y, X)). Furthermore it is
continuous, since
*ψ(v, w)h* = *v ◦ h ◦ w* ≤ *v**h**w*.

.
Hence
*ψ(v, w)h*
*ψ(v, w)* =
. sup ≤ *v**w*.
h∈L(X,Y )\{0} *h*
These considerations show that .u → ϕ % (u) = ψ(u−1 , u−1 ) is the composition of

the continuous function .u → (u−1 , u−1 ) of .Iso(X, Y ) to .L(Y, X) × L(Y, X) and of
the continuous function .(v, w) → ψ(v, w). Hence it is continuous, and the proof is
complete.

Remark 14.8 If .X = Y = R, every linear application from .R to .R is identified
with a real number. Hence .u ∈ Iso(R, R) if and only if .u = 0, and .u−1 = 1/u via
this identification. The previous theorem shows that .u → 1/u is differentiable at
any .u ∈ R \ {0}, and its derivative is .−1/u2 . A very complicated way to derive an
elementary fact.
Example 14.8 (Derivative of Bilinear Applications) Let X, Y and Z be Banach
spaces, and let .f : X × Y → Z be bilinear and continuous. We first introduce
a norm on .X × Y as follows: for .x ∈ X, .y ∈ Y , let
*(x, y)* = *x* + *y*.

.
It is easy to check that this is indeed a norm on .X × Y , and that .X × Y becomes a

Banach space with respect to this norm.
We now claim that f is differentiable at any point .(a, b) ∈ X × Y , and that
f % (a, b) : (h, k) ∈ X × Y → f (h, b) + f (a, k).

.
Indeed, we write
.f (a + h, b + k) − f (a, b) = f (h, b) + f (a, k) + f (h, k).
It suffices to prove that .*f (h, k)* = o(*(h, k)*) as .*(h, k)* → 0. By definition,
*(h, k)* = *h* + *k*, so that
.
*f (h, k)* ≤ *f **h**k* ≤ *f *(*h* + *k*)2 .

.
Since it is evident that .(*h* + *k*)2 = o(*h* + *k*) as .*(h, k)* → 0, the proof is
complete.
Exercise 14.16 Generalize the previous example to a bounded multilinear applica-
tion .f : X1 × · · · × Xn → Z, where the norm in .X1 × · · · × Xn is defined by
*(x1 , . . . , xn )* = *x1 * + · · · + *xn *.

.
Our definition of Fréchet differentiability is the straightforward generalization of

the basic idea of linear approximation. There are circumstances in which one might
be satisfied with a weaker kind of approximation: this is typical for functions of
several variables, and leads to the definition of the directional derivative.
Definition 14.25 Let U be an open subset of a Banach space, and let Y be a Banach
space. A function .F : U → Y is differentiable in the sense of Gâteaux (or G-
differentiable) at a point .a ∈ U if and only if there exists .g ∈ L(X, Y ) such that
F (a + εh) − F (a)
. lim = g(h) (14.4)
ε→0 ε
for every .h ∈ X. If it exists, the application g is unique, and it is denoted by .DG F (a)
of by .dG F (a).6
6The symbol .FG% (a) is also used, but it may be confused with the Fréchet derivative of a function
called .FG . We prefer to avoid such a symbol.
Remark 14.9 The limit which defines the Gâteaux derivative is a limit of a function
of one real variable .ε (which takes values in Y ).
Important: Different Definitions Are Possible

A slightly more general definition of the Gâteaux derivative is often proposed in the
literature by removing the linearity of g in (14.4). In other words, the limit in (14.4)
must exist for every h, but its value need not depend linearly on h.
Exercise 14.17 Prove that F -differentiability implies G-differentiability (with the

same derivative). Consider the function .F : R2 → R2 such that
⎧
⎨ x 2 y 2 if y = 0
.F (x, y) = x 4 +y 2
⎩0 otherwise,
and conclude that a G-differentiable function need not be continuous.

Theorem 9.8 is a basic result of Calculus courses. Although such a strong result
does not extend to our infinite-dimensional setting, the following is a fundamental
replacement. Given two points u and v of a vector space, we denote by
[u, v] = {tu + (1 − t)v | 0 ≤ t ≤ 1}

.
the interval defined by u and v.

Theorem 14.34 (Mean-Value Inequality) Let .F : U → X be G-differentiable at
any point of the open subset U of X. For every u and v in U , there results
*F (u) − F (v)* ≤ sup {*DG F (w)* | w ∈ [u, v]} · *u − v*.

.
Proof We assume that .F (u) = F (v), otherwise the proof is trivial. By Corol-
lary 14.3 there exists .ψ ∈ Y ∗ such that .*ψ* = 1 and
.ψ (F (u) − F (v)) = *F (u) − F (v)*.
We define .γ (t) = tu + (1 − t)v for any .t ∈ [0, 1] and
h(t) = ψ (F (γ (t)) .
.
Since .γ (t + τ ) = γ (t) + τ (u − v) it follows that
h(t + τ ) − h(t) F (γ (t + τ )) − F (γ (t))

. =ψ
τ τ
F (γ (t) + τ (u − v)) − F (γ (t))
=ψ
τ
By assumption we may let .τ → 0 and derive
h% (t) = ψ (DG F (tu + (1 − t)v)(u − v)) .

.
Since .h : [0, 1] → R, Theorem 9.8 applies and there exists .ϑ ∈ (0, 1) such that
h(1) − h(0) = h% (ϑ).

.
To conclude we notice that
*F (u) − F (v)* = h(1) − h(0) = h% (ϑ)

.
= ψ (DG F (tu + (1 − t)v(u − v))

≤ *ψ**DG F (ϑu + (1 − ϑ)v)**u − v*.
Since .*ψ* = 1 and .ϑu + (1 − ϑ)v = w ∈ [u, v], the proof is complete.

Remark 14.10 The previous Mean Value Inequality is actually a one-dimensional
result, as the proof clearly shows. We have reduced the infinite-dimensional function
F to the function h of one real variable. However, the most interesting technique of
the proof consists in composing F with .ψ ∈ Y ∗ . This is a standard trick to project
the range of the function to .R, so that the basic Lagrange Theorem may be applied.
The simple example
ϑ ∈ [0, 1] → e2πiϑ ∈ C
.
shows that the Lagrange Theorem does not hold for vector-valued functions.
A basic use of Theorem 14.34 is explained in the next regularity result.
Theorem 14.35 Suppose that .F : U → Y is G-differentiable in U , and let
DG F : U → L(X, Y ),
. u → DG F (u)
be continuous (with the standard topologies of each space) at the point .u∗ . Then F
is F-differentiable at .u∗ , and .F % (u∗ ) = DG F (u∗ ).
Proof We only need to show that F is F-differentiable at .u∗ , since we already know
that the F-derivative must then coincide with the G-derivative. For every .h ∈ X, we
define
R(h) = F (u∗ + h) − F (u∗ ) − DG F (u∗ )h.

.
We need to prove that .R(h) = o(*h*) as .*h* → 0. For .ε > 0 sufficiently small, the
function R is G-differentiable in .B(0, ε), and the Chain Rule yields
DG R(h) : k ∈ X → DG F (u∗ + h)k − DG F (u∗ )k.

.
We apply Theorem 14.34 with .[u, v] = [0, h] to get
.*R(h)* = *R(h) − R(0)* ≤ sup *DG R(th)**h*.

0≤t ≤1
But
*DG R(th)* = *DG F (u∗ + th) − DG F (u∗ )*,

.
hence
*R(h)* ≤ sup *DG F (u∗ + th) − DG F (u∗ )**h*.

.
0≤t ≤1
The continuity of .DG F at .u∗ comes now into play for the first time, and yields
. lim sup *DG F (u∗ + th) − DG F (u∗ )* = 0.

*h*→0 0≤t ≤1
In other words, we have proved that .*R(h)* ≤ o(1)*h* as .*h* → 0, and the proof
is complete.

Remark 14.11 The previous result offers a convenient tool for checking the Fréchet-
differentiability of a function. Since the Gâteaux derivative is just a limit in one
real variable, it is usually easier to compute. Then one hopes that the G-derivative
depends continuously on the point at which it is evaluated. Of course this is only
a sufficient condition for Fréchet-differentiability, and in many situations the only
possible approach to F-differentiability is via the basic definition.
14.11 Derivatives of Higher Order
For functions of a real variable there is no need to distinguish the nature of the
first derivative and of the second derivative, since at each step we go back to some
suitable function of a real variable. The situation changes drastically for real-valued
functions of two or more variables: if the first derivative is usually defined as a
14.11 Derivatives of Higher Order 337
suitable vector, the second derivative is a matrix. Something is lurking behind the
very rich structure of .Rn , and in this section we want to investigate this situation in
a general setting.
Suppose that .F ∈ C(U, Y ) is differentiable (in the sense of Fréchet) at all points
of the open set .U ⊂ X. We consider .F % : U → L(X, Y ).
Definition 14.26 Let .u∗ ∈ U . We say that F is twice F-differentiable at .u∗ if and
only if .F % is differentiable at .u∗ . The second F-derivative of F at .u∗ is then
D 2 F (u∗ ) = DF % (u∗ ) = D(DF (u∗ )).

.
The symbol .F %% (u∗ ) is also used. If F is twice differentiable at any point of U , we

say that F is twice differentiable in U . If .F %% is continuous from U to .L(X, Y ), we
write .F ∈ C 2 (X, Y ).
This definition mimicks the basic definition of the second derivative of a function
of one real variable. But if we think twice (!) about it, we easily realize that we
are considering the differentiability of the function .F % , which acts between U and
.L(X, Y ), a space which is usually much different than Y . So, the second derivative
of F at .u∗ is a bounded linear operator between X and .L(X, Y ), i.e.
F %% (u∗ ) ∈ L(X, L(X, Y )).

.
Recalling Theorem 14.20, it is convenient to think of .F %% (u∗ ) as a continuous

bilinear application from .X × X to Y . We will often write
F %% (u∗ )(h, k)
.
instead of7
.F %% (u∗ )(h)(k).
The isomorphism between .L(X, L(X, Y )) and .L(X, X; Y ) is usually understood in

notation, and .F %% (u∗ ) is employed to represent both objects.
Exercise 14.18 Let .X = C([0, 1]) be endowed with the usual norm of uniform
convergence. For .n ∈ N, define .F : X → X such that .u → un (here .un means “u to
the power n”). Prove that .F ∈ C 2 (X) and that
F %% (u) : (h, k) ∈ X × X → n(n − 1)un−2 hk.

.
7 This notation is formally correct but awful. It means that .F %% (u∗ )(h) is a bounded linear operator
which associates to .k ∈ X the element .F %% (u∗ )(h)(k) ∈ Y . Indeed, .F %% (u∗ ) associates to each
%% ∗
.h ∈ X an element of .F (u )(h) ∈ L(X, Y ).
When .n = 2, .F %% (u) is independent of u. Compare this with the basic fact that “the
second derivative of .x → x 2 is constant.” Can you extend the results of this exercise
to the case .n ∈ R, .n ≥ 2?
The actual computation of the second derivative can be performed by “fixing”
the first increment h, and differentiating once more with respect to u. The precise
statement is as follows.
Theorem 14.36 Suppose that .F : U → Y is twice differentiable at .u∗ ∈ U . For all
%
.h ∈ X, the function .Fh : X → Y such that .Fh (u) = F (u)h is differentiable at .u
∗
and
.Fh% (u∗ ) : k ∈ X → F %% (u∗ )(h, k).
Proof We decompose .Fh as .u → F % (u) → F % (u)h by means of .F % : U → L(X, Y )

and the evaluation map which associates to any .A ∈ L(X, Y ) the value .Ah ∈ Y .
This evaluation map is linear, hence the result follows at once from the Chain Rule.

No discussion of second derivatives is complete without the investigation of the
symmetry of the bilinear application .F %% (u). The next result extends a celebrated
result about the symmetry of the Hessian matrix for functions of two real variables.
Theorem 14.37 If .F : U → Y is twice differentiable at .u ∈ U , then .F %% (u) ∈
L(X, X; Y ) is symmetric:
F %% (u)(h, k) = F %% (u)(k, h)
. for every h ∈ X, k ∈ X.
Proof Let .ε > 0 be given. For any h and k in .B(0, ε), we define
ψ(h, k) = F (u + h + k) − F (u + h) − F (u + k) + F (u)
.
γh (ξ ) = F (u + h + ξ ) − F (u + ξ ).
For every .h ∈ X, the function .gh : B(0, ε) → Y such that
gh (k) = ψ(h, k) − F %% (u)(h, k) = γh (k) − γh (0) − F %% (u)(h, k)

.
Applying Theorem 14.34 to .gh (k) − gh (0) we find
*ψ(h, k) − F %% (u)(h, k)*

. ≤ sup *γh% (tk) − F %% (u)(h)* · *k*
0≤t ≤1
= sup *F % (u + h + tk) − F % (u + tk
0≤t ≤1
−F %% (u)(h)* · *k*.
14.11 Derivatives of Higher Order 339
Now,
F % (u + h + tk) = F % (u) + F %% (u)(h + tk) + ω(h + tk),

.
F % (u + tk) = F % (u) + F %% (u)(tk) + ω(tk)
with .ω(v) = o(*v*) as .*v* → 0. Therefore
F % (u + h + tk) − F % (u + tk) = F %% (u)h + ω(h + tk) − ω(tk),

.
and
*ψ(h, k) − F %% (u)(h, k)* ≤ sup *ω(h + tk) − ω(tk)* · *k*

.
0≤t ≤1
≤ ε(*h* + *k*)*k*
provided that .*h* and .*k* are sufficiently small.

By the same token, swapping h and k, we find
. *ψ(k, h) − F %% (u)(k, h)* ≤ sup *ω(k + th) − ω(th)* · *h*

0≤t ≤1
≤ ε(*k* + *h*)*h*.
The fact that .ψ(h, k) = ψ(k, h) yields
*F %% (u)(h, k) − F %% (u)(k, h)* ≤ ε(2*k*2 + 2|h*2 + 2|h**k*)

.
≤ 3ε(*k*2 + *h*2 ).
By homogeneity, the last inequality remains true for every h and k in X. Since .ε > 0
was arbitrary, .F %% (u)(h, k) = F %% (u)(k, h), and the proof is complete.

The definition of higher order derivatives is now clear. If .n ∈ N, the function
F : U → Y is n-times differentiable at .u∗ ∈ U if and only if .D n−1 F is differentiable
.
at .u∗ . In particular, .D n F (u∗ ) can be seen as a continuous multilinear application

from .X × · · · × X (n factors) to Y via a repeated application of Theorem 14.20.
In particular
Definition 14.27 Let X and Y be Banach spaces, and let U be an open subset of
X. A function .F : U → Y is of class .C m (U, Y ) (.m ≥ 1 being an integer) if and
only if DF , .D 2 F , . . . , .D m F exist and are continuous functions on U . As in the
real-variable case, we set
∞

C ∞ (U, Y ) =
. C m (U, Y ).
m=0
14.12 Partial Derivatives
A very peculiar aspect of differential calculus in Banach spaces is that every function
is actually considered as a function of a single vector variable. When we study
Calculus in Several Variables, the fact that any point of .Rn is an n-tuple of real
numbers plays a fundamental rôle. This is one of the many places where the
algebraic structure of .Rn conflicts with its geometric structure, and things become
unexpectedly obscure. Indeed, our definition of the Fréchet derivative applies very
well to the case .X = Rn , and no need to distinguish variables appears.
There are however situations in which the domain of a function is given as
a cartesian product of two Banach spaces, and then partial derivatives become a
natural idea. This is the object of this section.
Suppose that X, Y and Z are Banach spaces, and let .(u∗ , v ∗ ) ∈ X × Y be a point.
Two applications can be naturally defined as follows:
σv ∗ (u) = (u, v ∗ )
.
τu∗ (v) = (u∗ , v).
Hence .σv ∗ : X → X × Y and .τu∗ : Y → X × Y . Both functions are linear, hence
Dσv ∗ (u) : h → (h, 0)

.
Dτu∗ (v) : k → (0, k).
Since these derivatives are independent of .u∗ , .v ∗ , u and v, we will denote them by
.σ and .τ , respectively.
We consider an open subset Q of .X × Y and a function .F : Q → Z.

Definition 14.28 The function F is partially differentiable with respect to u at
(u∗ , v ∗ ) if and only if .F ◦ σv ∗ is differentiable at .u∗ . In this case the bounded linear
.
operator .D(F ◦ σv ∗ )(u∗ ) ∈ L(Y, Z) is the partial derivative of F at .(u∗ , v ∗ ) with

respect to u, and is denoted by
∂1 F (u∗ , v ∗ ).
.
Similarly, F is partially differentiable with respect to v at .(u∗ , v ∗ ) if and only if .F ◦

τu∗ is differentiable at .v ∗ . In this case the bounded linear operator .D(F ◦ τu∗ )(v ∗ ) ∈
L(X, Z) is the partial derivative of F at .(u∗ , v ∗ ) with respect to v, and is denoted by
∂2 F (u∗ , v ∗ ).
.
14.13 The Taylor Formula 341
Remark 14.12 Of course u and v are dummy variables, and we prefer to avoid the
popular notation .∂u F , .∂v F to denote partial derivatives. What really matters here is
just the position of the variable with respect to which we are differentiating, and our
notation reflects this fact.
Exercise 14.19 Retain the setting of the previous definition. Prove that the defini-
tion of partial derivatives is equivalent to requiring that there exist .Au ∈ L(X, Z)
and .Av ∈ L(Y, Z) such that
F (u∗ + h, v ∗ ) − F (u∗ , v ∗ ) = Au h + o(*h*)

.
F (u∗ , v ∗ + k) − F (u∗ , v ∗ ) = Av k + o(*k*)
as .*h* → 0, .*k* → 0. In other words, partial derivatives are indeed what we

expect.
Exercise 14.20 Suppose that F is differentiable at .(u∗ , v ∗ ). Prove that F has partial
derivatives at .(u∗ , v ∗ ), and
∂1 F (u∗ , v ∗ ) : h → F % (u∗ , v ∗ )(h, 0)

.
∂2 F (u∗ , v ∗ ) : k → F % (u∗ , v ∗ )(0, k).
Hence partial derivatives are just evaluations of the Fré derivative, provided that F
is F-differentiable.
14.13 The Taylor Formula
Theorem 14.38 Let f : U → Y be a function of class C m (U, Y ), where U is an

open subset of a Banach space X and Y is a Banach space. If p ∈ U , u ∈ U are
such that [p, u] = {(1 − t)p + tu | 0 ≤ t ≤ 1} ⊂ U , then
1 1
F (u) = F (p) +
. Df (p)(u − p) + D 2 F (p)(u − p, u − p) + · · ·
1! 2!
1
+ D m−1 F (p)(u − p, . . . , u − p)+
(m − 1)!
1 1
+ (1 − t)m−1 D m F (p + t (u − p))(u − p, . . . , u − p) dt.
(m − 1)! 0
If Y = R, there exists a point ξ ∈ [p, u] such that ξ = p, ξ = u and
1 1
. (1 − t)m−1 D m F (p + t (u − p))(u − p, . . . , u − p) dt
(m − 1)! 0
1 m
= D F (p)(u − p, . . . , u − p).
m!
Proof The idea is to reduce to a function of a single variable. Let σ (t) = p + t (u −
p) for every t ∈ [0, 1]. If g = F ◦ σ , then g ∈ C m ([0, 1], Y ) and
g () (t) = D F (σ (t))(u − p, . . . , u − p)

.
for every positive integer ≤ m. Hence the usual Taylor formula with integral
remainder yields

m−1
g () (0) 1 1
g(1) =
. 1 + (1 − τ )m−1 g (m) (τ ) dτ.
! (m − 1)! 0
=0

14.14 The Inverse and the Implicit Function Theorems
Many problems can be translated into an equation of the form
F (u) = v,
.
in which the unknown u must be found for a given datum v. If F happens to be

invertible, then .u = F −1 (v) and the problem is solved. Unfortunately it is often
difficult to ensure that the function F be invertible by direct inspection.
Furthermore, the equation .F (u) = v is a particular case of the more general
equation .G(u, v) = 0, in which the unknown u is (hopefully) defined implicitly in
terms of v. Again, it might be very difficult to solve .G(u, v) = 0 with respect to u
in terms of elementary functions only.
These are the basic examples which introduce two8 powerful tools of Nonlinear
Analysis. We first introduce a classical version of these results, and then we
investigate the possibility of inverting a nonlinear function globally.
8It is a matter of fact that these two results are indeed equivalent, so that they could be seen as
different flavors of the same result.
14.14 The Inverse and the Implicit Function Theorems 343
14.14.1 Local Inversion
We begin with local inversion of a function around suitable points. We will always
deal with continuous functions .F : X → Y between Banach spaces. We recall that
.Iso(X, Y ) denotes the Banach space of bounded linear operators from X to Y which
possess a bounded linear inverse.

Definition 14.29 The function .F ∈ C(X, Y ) is locally invertible at .u∗ ∈ X if and
only if there exist open neighborhoods U of .u∗ and V of .v ∗ = F (u∗ ) such that
.F ∈ Iso(U, V ).
Theorem 14.39 Suppose that .F ∈ C 1 (X, Y ) and .F % (u∗ ) is a bounded linear

invertible operator. Then F is locally invertible at .u∗ with a local inverse of class
∗ ∗
.C . More precisely, there exists open neighborhoods U of .u and V of .v = F (u )
1 ∗
such that
(a) .F ∈ Iso(U, V )
(b) .F −1 ∈ C 1 (V , X) and for every .v ∈ V there results .(F −1 )% (v) = (F % (u))−1 ,
where .u = F −1 (v).
(c) If .F ∈ C k (X, Y ) for some .k > 1, then .F −1 ∈ C k (V , X)
Proof A few reductions may be convenient. Firstly, we assume that .u∗ = 0 and
∗
.v = 0: the general case follows by composition with two translations. If we define
the linear operator .A = (F % (0))−1 and we replace F with .A ◦ F , we see that we

may always consider the particular case of a function of the form
.F = I + ,
where .I = IX is the identity on X and . ∈ C 1 (X, X) satisfies . % (0) = 0. In the

rest of the proof we will retain these assumptions.
Pick .r > 0 so small that
1
.∀p(p ∈ X ∧ *p* < r) ⇒ * % (p)* < .
2
Using the Mean Value Inequality, for every p and q in the ball .B(0, r) there results

*(p) − (q)* ≤ sup * % (w)* w ∈ [p, q] *p − q*
.
1
≤ *p − q*.
2
This shows that . is a contraction and that .*(p)* ≤ (1/2)*p* whenever .*p* < r.
Now, for every .v ∈ X we introduce the auxiliary function
v (u) = v − (u).
.
The function .v is also a contraction, and for every .u ∈ B(0, r) and .v ∈ B(0, r/2)
there results
*v (u)| ≤ *v* + *(u)* ≤ r.

.
As a consequence, whenever .*v* ≤ r/2, .v maps .B(0, r) into itself and is a
contraction. It follows that .v possesses a unique fixed point .u ∈ B(0, r) which
satisfies .u = v (u), i.e.
u = v − (u).
.
This is equivalent to .F (u) = v, and therefore there exists an inverse

F −1 : B(0, r/2) → B(0, r). To prove that .F −1 is continuous, we call .u = F −1 (v)
.
and .w = F −1 (z), i.e.
u + (u) = v,
. w + (w) = z.
These equalities yield
1
|u − w| ≤ |v − z* + |(u) − (w)| ≤ *v − z* + *u − w*.
.
2
Hence .*F −1 (v) − F −1 (z)* ≤ 2*v − z*, and .F −1 is actually Lipschitz continuous.
Letting
r
V = B 0,
. , U = B(0, r) ∩ F −1 (V ),
2
we see that .F|U ∈ Iso(U, V ). This proves (a).
To prove (b), we set again .u = F −1 (v) and from .v = u + (u) we derive
.F
−1 (v) = v − (F −1 (v)). But .(u) = o(*u*) as .|u* → 0 and F is Lipschitz
continuous, hence .(F −1 (v)) = o(*v*) and .F −1 is differentiable at .v = 0 with
.(F
−1 )% (0) = I . To treat the general case, we pick .v ∈ B(0, r/2) and .u = F −1 (v);
we then translate both u and v to the origin of X, and we find that .(F −1 )% (v) =
(F % (u))−1 . The application
(F % )−1 : v → (F % (F −1 (v)))−1
.
can be factored as
v → u = F −1 (v) → F % (u) → (F −1 )−1 ,

.
which shows that .F −1 ∈ C 1 as a composition of functions of class .C 1 . Notice that

we are using Theorem 14.33. The proof of (b) is complete. The proof of (c) follows
from (b) by an easy induction argument.

14.14 The Inverse and the Implicit Function Theorems 345
As a remarkable corollary of Theorem 14.39 we will prove a local Implicit

function Theorem. Following [1] we employ an asymmetric notation: we consider
two Banach spaces X and Y , an open subset . of some Banach space T , and a
function
F : × X → Y.
.
This suggests that . should be considered as a space of parameters, which is the

typical situation in many applications to differential equations. Needless to say, this
is just a suggestion, since names are just names.
Theorem 14.40 (Implicit Function Theorem) Suppose that .F ∈ C k ( × U, Y )
for some .k ≥ 1, where U is an open subset of X. Suppose that
1. .F (λ∗ , u∗ ) = 0,
2. the partial derivative .∂2 F (λ∗ , u∗ ) is a bounded linear invertible operator.
Then there exist open neighborhoods ." of .λ∗ in T and .U ∗ of .u∗ in X and a function
.g ∈ C (", X) such that
k
(a) .F (λ, g(λ)) = 0 for every .λ ∈ ";

(b) .F (λ, u) = 0 and .(λ, u) ∈ " × U ∗ imply .u = g(λ);
(c) .g(λ) = −(∂2 F (λ, g(λ)))−1 ◦ ∂1 F (λ, g(λ)) for every .λ ∈ ".
The strategy of the proof is to reduce the conclusion to the setting of Theo-
rem 14.39. This is the technical content of the following Lemma.
Lemma 14.3 Let .(λ∗ , u∗ ) ∈ × U . Suppose that
1. F is continuous, .∂2 F exists in .×U , and .∂2 F : ×U → L(X, Y ) is continuous;
2. .∂2 F (λ∗ , u∗ ) is a bounded linear invertible operator from X to Y .
Then the application . : × U → T × Y such that
.(λ, u) = (λ, F (λ, u))
is locally invertible at .(λ∗ , u∗ ) with continuous inverse .. Furthermore, if .F ∈

C 1 ( × U, Y ), then . is of class .C 1 .
Proof An argument similar to that in the proof of Theorem 14.39 shows that . is
locally invertible with a continuous inverse .. Now, assume that .F ∈ C 1 (×U, Y )
and define
A = ∂1 F (λ∗ , u∗ ),
. B = ∂2 F (λ∗ , u∗ ).
It is clear that . ∈ C 1 ( × U, T × Y ) and that
% (λ∗ , u∗ ) : (ξ, v) → (ξ, Aξ + Bv).

.
Hence the equation . % (λ∗ , u∗ )(ξ, v) = (η, υ) yields .ξ = η and .Aη + Bv = υ.

Recalling that B is invertible, we derive .v = B −1 (υ − Aη). As a consequence
% ∗ ∗
. (λ , u ) is a bounded linear invertible operator from .T × X to .T × Y The
conclusion follows from Theorem 14.39.

Proof (of Theorem 14.40) We consider again the function .(λ, u) = (λ, F (λ, u))
introduced in Lemma 14.3. It follows that . is locally invertible at .(λ∗ , u∗ ) and
(λ∗ , u∗ ) = (λ∗ , F ∗ λ∗ , u∗ )) = (λ∗ , 0).

.
The local inverse . os . is of the form .(λ, v) = (λ, ϕ(v)) for some function
ϕ : " × V → X defined in an open neighborhood ." × V of .(λ∗ , F (λ∗ , u∗ )) and
.
satisfying .F (λ, ϕ(λ, v)) = v for every .λ ∈ ". This follows at once from the fact
that the first component of . is the identity in .. An easy induction argument shows
also that .ϕ ∈ C k provided that .F ∈ C k .
We define .g(λ) = ϕ(λ, 0) for every .λ ∈ ". It follows that
F (λ, g(λ)) = F (λ, ϕ(λ, 0)) = 0

.
for every .λ ∈ ". Hence (a) follows, and (b) follows from the fact that . is injective.
Differentiating the identity .F (λ, ϕ(λ, v)) = v yields
. ∂1 F + ∂2 F ◦ ∂1 ϕ = 0
∂2 F ◦ ∂2 ϕ = I.
Hence .∂1 ϕ = −(∂2 F )−1 ◦ ∂1 F , and (c) follows at once.

14.15 A Global Inverse Function Theorem
The nature of Theorem 14.39 in strongly local, since the assumption on the
derivative at a single point does not allow us to prove that the given function is
globally injective. It is then natural to investigate whether a global Inverse Function
Theorem may exist at all. In this section we first present a very classical result in
this direction, whose proof is inspiring in its own. As a further improvement, we
prove a strong result which requires some definitions of Algebraic Topology.
Theorem 14.41 (Hadamard-Caccioppoli) Let X and Y be Banach spaces, and let
.F ∈ C 1 (X, Y ) such that .F % (x)−1 ∈ L(Y, X) for all .x ∈ X. If there exist constants
.A > 0 and .B > 0 such that
* *
* % −1 *
. *F (x) * ≤ A*x* + B for all x ∈ X,
14.15 A Global Inverse Function Theorem 347
then F is a diffeomorphism between X and Y .9

Proof We need to prove that f is surjective and injective. The regularity of .F −1
follows as in the proof of Theorem 14.39.
Surjectivity We want to prove that for every .y ∈ Y there exists .x ∈ X such that
.f (x) = y. We consider the last equation as part of a family of equations as follows:
given .x0 ∈ X, we define .F̃ : [0, 1] × X → Y such that
F̃ (t, x) = F (x) − [(1 − t)F (x0 ) + ty].

.
We consider the set
S = {t ∈ [0, 1] | F (t, ·) = 0 is solvable} .

.
It is clear that .0 ∈ S, and Theorem 14.39 shows that S is an open set, since
(∂2 F̃ )−1 (t, x) = F % (x)−1 ∈ L(Y, X). We claim that S is closed, and this implies
.
that .S = [0, 1] by connectedness. In any connected component .(a, b) of S there

exists a branch .t → xt of solutions satisfying
F̃ (t, xt ) = 0
. for every t ∈ (a, b).
This follows from the Implicit Function Theorem. Differentiating both sides we get
dxt
F % (xt )
. = y − F (x0 ).
dt
Therefore
* *
* dxt * * *
.* * * % −1 *
* dt * ≤ *F (xt ) * *y − F (x0 )* ≤ (A *xt * + B) *y − F (x0 )*. (14.5)
Let .c = (a + b)/2, so that

t
. *xt * ≤ *xc * + *y − F (x0 )* (A *xs * + B) ds
c
for every .t > c. It follows easily10 that there exists a constant .C > 0 such that
. *xt * ≤ C for every t ∈ (a, b).
9 Recall that this means that both F and .F −1 are differentiable with continuous inverse.
10 This is often called Gronwall’s inequality.
Now (14.5) yields another constant .C1 > 0 such that

* *
* dxt *
* *
.
* dt * ≤ C for every t ∈ (a, b),
and it follows that S is a closed subset of .[0, 1]. Thus F is surjective.

Injectivity We proceed by contradiction, assuming the existence of .y ∈ Y and
.x0 ∈ X, .x1 ∈ X such that .F (x0 ) = y = F (x1 ). Let .γ : [0, 1] → X be the segment
γ (s) = (1 − s)x0 + sx1 ,

. 0 ≤ s ≤ 1.
In particular .f ◦ γ is a closed loop passing through y. We want to construct a survey

x : [0, 1] → X such that
.
x(0) = x0
.
x(1) = x1
f (x(s)) = y for every s ∈ [0, 1].
This will immediately contradict the local invertibility of F . We set .I = [0, 1] and
T : I × C0 (I, X) → C0 (I, Y ) such that
.
T : (t, u(·)) → F (γ (·) + u(·)) − ty − (1 − t)F (γ (·)),

.
where
C0 (I, X) = {u ∈ C(I, X) | u(0) = u(1) = 0} .

.
We need to solve the equation .T (t, u) = 0. It is evident that .T (0, 0) = 0; if,

moreover, we have .u ∈ C0 (I, X) satisfying .T (1, u) = 0, then .x = u + γ is what
we need. Observe now that
1. the partial derivative
∂2 T (t, u) = F % (γ (·) + u(·)) ∈ L(C0 (I, X), C0 (I, Y )),

.
which has a bounded inverse. Therefore the set
S = {t ∈ I | T (t, u) = 0 is solvable}
.
is open, as a consequence of the Implicit Function Theorem.

2. Let .s → ut (s) be a solution at .t ∈ S. Then
dut
F % (γ (s) + ut (s))
. = y − F (γ (s)).
dt
As before we obtain
* *
* dut *
* * ≤ A *ut *C0 (I,X) + B1 *y − F ◦ γ *C0 (I,Y ) ,
.
* dt *
C0 (I,X)
and again there exists a constant .C > 0 such that

* *
* dut *
* * ≤C for every t ∈ S.
.
* dt *
C0 (I,X)
Hence .1 ∈ S, and this is a contradiction. Since we have proved that F is injective,


A much stronger refinement of Theorem 14.41 is possible, but we need to
introduce some terminology from Algebraic Topology.
Definition 14.30 A topological space X is path-connected if and only if for every
x ∈ X, .y ∈ X there exists a continuous function .γ : [0, 1] → X such that .γ (0) = x
.
and .γ (1) = y.
Definition 14.31 A topological space X is simply connected if and only if it is
path-connected and for every continuous .p : [0, 1] → X, .q : [0, 1] → X such that
.p(0) = q(0), .p(1) = q(1), there exists a continuous .F : [0, 1] × [0, 1] → X such
that .F (x, 0) = p(x) and .F (1, x) = q(x) for every .x ∈ X.

Definition 14.32 Let X and Y be topological spaces. A function .f : X → Y is a
local homeomorphism if and only if for every .x ∈ X there exist open neighborhoods
U of x, V of .f (x) such that .f : U → V is a homeomorphism.
Theorem 14.42 (Global Inverse Function Theorem) Let .f : X → Y be a local
homeomorphism between two path-connected Hausdorff spaces X and Y , and let Y
be simply connected. The following statements are equivalent:
(a) f is a homeomorphism;
(b) f is a proper function, i.e. for every compact .K ⊂ Y , the pre-image .f −1 (K) is
a compact subset of X.
The proof is rather long, and requires some preliminaries. In the rest of this
section, X, Y and Z will denote Hausdorff spaces.
Definition 14.33 Let .f : X → Y be a local homeomorphism, and let .p : Z → Y
be a continuous function. A continuous function .p̃ : Z → X is a lifting of p by f if
and only if .f ◦ p̃ = p.
Proposition 14.1 Let .f : X → Y be a local homeomorphism, and let .p : Z → Y
be a continuous function. If Z is connected and if .p̃1 , .p̃2 : Z → X are both liftings
of p by f , then either .p̃1 = p̃2 , or .p̃1 (z) = p̃2 (z) for every .z ∈ Z.
Proof Let .C = {z ∈ Z | p̃1 (z) = p̃2 (z)}. We first claim that C is open in Z. If
C = ∅, there is nothing to prove; otherwise we consider .z0 ∈ C and let .x0 =
.
p̃1 (z0 ) = p̃2 (z0 ). Moreover, let U and V open neighborhoods of .x0 and .f (x0 ) such
that .f : U → V has a continuous inverse .g : V → U . The set .W = p̃1−1 (U ) ∩
p̃2−1 (U ) is an open neighborhood of .z0 and there results .p̃1 |W = p̃2 |W = g ◦ p|W .
Thus .W ⊂ C and C is an open set.
By a very easy argument the complement .Z \ C is also open, hence .C = Z by
connectedness. The proof is complete.

Definition 14.34 A local homeomorphism .f : X → Y lifts the paths if and only
if, for every continuous .α : [0, 1] → Y such that .α(0) ∈ f (X) and for every .x0 ∈
f −1 ({0}), there exists a lifting .α̃ : [0, 1] → X of .α such that .α̃(0) = x0 .
Exercise 14.21 Prove that there exists at most one lifting .α̃ as described in the
previous Definition. Hint: use Proposition 14.1.
Definition 14.35 Any continuous function .H : Z × [0, 1] → Y is a homotopy
with base .H0 : Z → Y such that .z → H (z, 0). A function .f : X → Y lifts the
homotopies if, for every homotopy H and any continuous function .H̃0 : Z → X
such that .H̃0 = H0 , there exists a continuous lifting .H̃ with base .H̃0 , i.e. .f ◦ H̃ = H
and .H̃ (z, 0) = H̃0 (z) for all .z ∈ Z.
Proposition 14.2 If a local homeomorphism .f : X → Y between Hausdorff spaces
lifts the paths, then it lifts the homotopies.
Proof Let .t → H̃ (z, t) be the unique lifting of the path .t → H (z, t) with origin
.H̃0 (z), for any .z ∈ Z. It is clear that .f ◦ H̃ = H , and that .H̃ (z, 0) = H̃0 (z). Since
.z ∈ Z is arbitrary, we have thus defined .H̃ : Z × [0, 1] → X. We claim that .H̃ is
continuous.
Let .z0 ∈ Z and let

.D = t ∈ [0, 1] H̃ is not continuous at (z0 , t) .
If .D = ∅, then we may introduce .α = inf D ≥ 0. By continuity of .t → H̃ (z0 , t), for

every open neighborhood U of .H̃ (z0 , α) there exists an open interval .J1 containing
.α such that .H̃ (z0 , t) ∈ U for every .t ∈ J1 . By shrinking U we may also assume
that f induces a homeomorphism between U and some open neighborhood V of

.H (z0 , α). Since H is continuous, there exist an open neighborhood .W1 of .z0 in Z
and an open interval .J2 containing .α such that .H (W1 × J2 ) ⊂ V . Let .J = J1 ∩ J2 ,

and let .b ∈ J be such that .b < α if .α > 0; if .α = 0, we choose .b = 0. In any
case .z → H̃ (z, b) is continuous at .z0 , and since .H̃ (z0 , b) ∈ U , there exists an open
neighborhood .W2 of .z0 in Z such that .H̃ (W2 × {b}) ⊂ U . Setting .W = W1 ∩ W2 ,
we claim that
H̃ |(W × J ) = (f |U )−1 ◦ H |(W × J ).

.
These two functions agree on .W × {b}, hence for every .z ∈ W the functions
t → H̃ (z, t)
.
t → (f |U )−1 ◦ H (z, t)
are liftings of .t → H (z, t) which coincide at .b ∈ J , and hence coincide on J . In

conclusion .H̃ is continuous at .(z0 , t) for every .t ∈ J with .t ≥ α, contradicting the
minimality of .α. The proof is complete.

The next result is the turning point of our investigation.
Proposition 14.3 Suppose that .f : X → Y is a local homeomorphism between
Hausdorff spaces, and that f lifts the paths. If X and Y are path-connected and if
Y is simply connected, then f is a homeomorphism.
Proof We prove that f is surjective. Fix .y0 ∈ f (X) and .x0 ∈ f −1 ({y0 }), and let
.α : [0, 1] → Y be any continuous function such that .α(0) = y0 and .α(1) = y. We
know that there exists a unique lifting .α̃ of .α such that .α̃(0) = x0 . Since .f ◦ α̃ = α
yields .f (α̃(1)) = y. Since .y ∈ Y is arbitrary, f is surjective.
To prove that f is injective, let .x0 and .x1 be two distinct points of X such that
.f (x0 ) = f (x1 ) = y0 . Since X is path-connected, we can consider a continuous
.σ : [0, 1] → X such that .σ (0) = x0 and .σ (1) = x1 . The formula .α = f ◦ σ
defines a continuous function with .α(0) = α(1) = y0 . Now we recall that Y is

simply connected, so that there exists a continuous .h : [0, 1] × [0, 1] → Y such that
.h(t, 0) = α(t), .h(t, 1) = y0 for every .t ∈ [0, 1], and .h(0, s) = y0 = h(1, s) for
every .s ∈ [0, 1].

By assumption f lifts the paths, hence it lifts the homotopies by Proposition 14.2.
Call .h̃ : [0, 1] × [0, 1] → X be a homotopy which lifts h and which satisfies
.h̃(t, 0) = σ (t) for every .t ∈ [0, 1].
It follows from the definitions that a constant path11 is lifted to a constant path.
In particular .h̃(0, s) = σ (0) = x0 , .h̃(1, s) = σ (1) = x1 for every .s ∈ [0, 1].
Exploiting the fact that .t → h̃(t, 1) is also constant, we have
x0 = h̃(0, 1) = h̃(1, 1) = x1 .
.
This contradiction proves that f is injective, and therefore it is a homeomorphism.

Definition 14.36 Suppose that .f : X → Y is a local homeomorphism. For a
continuous path .α : [0, 1] → Y such that .α(0) ∈ f (X) and .x0 ∈ f −1 (α(0)), we
define the maximal lifting .φ : J → X of .α with .φ(0) = x0 in the following way.
There exists a continuous function .φI : I → X, with .I = [0, b) ⊂ [0, 1], such that
11 This refers to any continuous function .t → α(t) such that .α(t) does not depend on .t ∈ [0, 1].
φI (0) = x0 and .f ◦ φI = α|I . By Proposition 14.1 the formula .φ|J = φI defines

.
the mapping .φ : J → X on the union J of all the intervals I .

Definition 14.37 Let .φ : [0, b) → X, .0 < b ≤ +∞, be a continuous function. The
ω-limit set of .φ is
.

.ωφ = φ([t, b)) 0 ≤ t < b .
Exercise 14.22 Prove that a point .x ∈ ωφ if and only if x is an accumulation point

of some sequence .{φ(tn )}n , where .tn → b and .0 ≤ tn < b for every n. If X is a
metric space, .x ∈ ωφ if and only if there exists a sequence .{tn }n such that .0 ≤ tn < b
for every n, .tn → b and .φ(tn ) → x.
Remark 14.13 The terminology is reminiscent of the terminology used in Dynami-
cal Systems. We want to stress the fact that here no differential equation is involved,
and the definition of the .ω-limit set is purely topological.
Proposition 14.4 Let .f : X → Y be a local homeomorphism between Hausdorff
spaces, and let .φ : J → X be the maximal lifting of .α : [0, 1] → Y with .φ(0) =
x0 ∈ f −1 (α(0)). If .J = [0, 1], then J has the form .[0, b) for some .b ∈ (0, 1), and
.ωφ = ∅.
Proof We assume that .J = [0, a] for some .0 < a < 1. At the point .f (φ(a)) the
function f has a local inverse, and we can extend .φ to a lifting defined on some
larger interval. This contradicts the maximality of .φ, and therefore the maximal
lifting must be defined on an interval .[0, b) with .0 < b < 1.
To prove the second statement, we assume that .x0 ∈ ωφ . Then .f (x0 ) = α(b)
since f is continuous, and thus .f (φ([t, b))) ⊂ f (φ([t, b)) and

. f (φ([t, b)) 0 ≤ t < b = {α([t, b]) | 0 ≤ t < b} = {α(b)} .
Pick open neighborhoods U and V of .x0 and .f (x0 ) respectively, such that
.f |U : U → V is a homeomorphism, and let g be the inverse of .f |U . We select
.a ∈ [0, b) such that .α([a, b]) ⊂ V and such that .φ(a) ∈ U . We then define
.ψ : [0, b] → X as a lifting of .α|[0, b] by .ψ|[0, a] = φ|[0, a] and by .ψ|(a, b] =
g ◦ α|(a, b]. Again we reach a contradiction with the maximality of .φ, and the proof
is complete.

Proof (of Theorem 14.42) Since (a) trivially implies (b), we only prove that (b)
implies (a). Let f be a proper function. If we can prove that f lifts the paths, an
application of Proposition 14.3 gives the desired result.
We argue by contradiction, assuming the existence of a continuous path
.α : [0, 1] → Y
and of a point .x0 ∈ f −1 (α(0)) such that the maximal lifting .φ of .α with .φ(0) = x0
is defined on a proper subset .[0, b) of .[0, 1]. Then we know that .ωφ = ∅. But
14.16 Critical and Almost Critical Points 353
φ([0, b)) ⊂ f −1 (α([0, 1])) and the latter set is compact since .α([0, 1]) is compact
.
in Y and f is proper.
Every finite collection of closed sets .{φ([ti , b))}i has non-empty intersection,
then

.ωφ = φ([t, b)) 0 ≤ t < b = ∅,
by the finite-intersection property of compact spaces. This contradiction proves the

statement.

14.16 Critical and Almost Critical Points
Every (or almost every) student should remember the definition of critical point for a
function .f : R → R: we say that .x0 is a critical point of f if and only if .f % (x0 ) = 0
(which implies that f must be differentiable at .x0 ). What can we do for functions
between normed vector spaces?
Definition 14.38 Let X, Y be Banach spaces, and let U be an open subset of X. A
point .u ∈ U is a critical point of a function .F : U → Y if and only if .DF (u) ∈
L(X, Y ) is not surjective.
Example 14.9 If .Y = R, a critical point is just a point .u ∈ U such that .DF (u) =
0 ∈ X∗ , i.e.
DF (u)(v) = 0
. for every v ∈ X.
The quest of critical points is a formidable task, and it is an independent branch

of mathematics called Critical Point Theory. In this chapter we want to present an
important tool that can be used to construct critical points. We follow [6].
Theorem 14.43 (Ekeland Variational Principle) Let .(M, d) be a complete metric
space and let . : M → (−∞, +∞] be a lower semicontinuous function, bounded
from below and not identical to .+∞. Let .ε > 0 and .u ∈ M such that
(u) ≤ inf + ε.
.
M
Then there exists .v ∈ M such that

(a) .(v) ≤ (u);
(b) .d(u, v) ≤ 1;
(c) for every .w = v in M, there results .(w) > (v) − εd(w, v).
Proof The relation
w ≤ v ⇐⇒ (w) + εd(w, v) ≤ (v)

.
is a partial order on M. We call .u0 = u and we suppose that .un is know. Let
Sn = {w ∈ M | w ≤ un } .
.
There exists .un+1 ∈ Sn such that
1
(un+1 ) ≤ inf +
. .
Sn n+1
Since .un+1 ≤ un we see that .Sn+1 ⊂ Sn . Moreover .Sn is a closed subset since . is
lower semicontinuous. Now, for every .w ∈ Sn+1 we have .w ≤ un+1 ≤ un , so that
1 1
εd(w, un+1 ) ≤ (un+1 ) − (w) ≤ inf +
. − inf = .
Sn n+1 Sn n+1
As a consequence,
2
. diam Sn+1 = sup {d(w1 , w2 ) | w1 ∈ Sn+1 , w2 ∈ Sn+1 } ≤ .
ε(n + 1)
But M is complete, hence

∞

. Sn = {v}
n=0
for some .v ∈ M. In particular .v ∈ S0 , i.e. .v ≤ u0 = u. This implies
(v) ≤ (u) − εd(u, v) ≤ (u)

.
and
(u) − (v) infM + ε − infM
.d(u, v) ≤ ≤ = 1.
ε ε
To conclude, it suffices to show that .w ≤ v implies .w = v.12 By definition, .w ≤ un

for every n, hence .w ∈ Sn for every n. This implies .w = v, as claimed. The proof
is complete.

Remark 14.14 The same proof can be repeated with the equivalent metric .λd, .λ >
0, instead of d. Of course .λ and .ε may be related, as in the proof of the following
corollary.
12 In this sense, v is a minimal element.

14.16 Critical and Almost Critical Points 355
Theorem 14.44 (Almost Critical Points) Let X be a Banach space, .ϕ : X → R

be a function bounded from below and differentiable on X. For every .ε > 0 and for
every .u ∈ X such that
ϕ(u) ≤ inf ϕ + ε,
.
X
there exists .v ∈ X such that
ϕ(v) ≤ ϕ(u)
.
√
*u − v* ≤ ε
√
*Dϕ(v)* ≤ ε.
Proof We apply the Ekeland Variational Principle in .M = X with . = ϕ and

λ = ε−1/2 as in the previous √
. Remark. This produces an element v, and we only
need to prove that .*Dϕ(v)* ≤ ε.
Let .w = v + th for .t > 0, .h ∈ X and .*h* = 1. The properties of v give
√
ϕ(v + th) − ϕ(v) > − εt.
.
Dividing by t and letting .t → 0 we see that

√
. − ε ≤ Dϕ(v)(h).
√ √
Replacing h with .−h, we conclude that .− ε ≤ Dϕ(v)(h) ≤ ε for every .h ∈ X,
.*h* = 1. The conclusion follows.

Corollary 14.8 Let X be a Banach space, .ϕ : X → R be a function bounded from
below and differentiable on X. For every minimizing sequence .{un }n of .ϕ there exists
a minimizing sequence .{vn }n such that
ϕ(vn ) ≤ ϕ(un )
.
*un − vn * → 0
*Dϕ(vn )* → 0.
Proof We define
ϕ(un ) − infX ϕ if ϕ(un ) − infX ϕ > 0

.εn =
1/n if ϕ(un ) − infX ϕ = 0
and we choose .un according to Theorem 14.44.

Exercise 14.23 Let .(M, d) be a complete metric space, .ϕ : M → R a lower

semicontinuous non-negative function, and .T : M → M a function such that
.d(u, T (u)) ≤ ϕ(u) − ϕ(T (u)) for every .u ∈ M. Prove that T has a fixed point.
Hint: Ekeland’s Principle with .ε = 1/2 gives .v ∈ M such that
1
. d(v, T (v)) ≥ ϕ(v) − ϕ(T (v)).
2
Hence .d(v, T (v)) ≤ 0.
Exercise 14.24 Let .(M, d) be a complete metric space, .T : M → M such that
there exists .L ∈ [0, 1) such that .d(T (u), T (v)) ≤ Ld(u, v) for every .u ∈ M and
.v ∈ M. Use the previous exercise to show that T has a fixed point. Hint: try to
choose a constant .c > 0 such that .ϕ(u) = cd(u, T (u)) satisfies the conditions of
the previous exercise.
At this point, the natural question is: does the sequence .{vn }n of Corollary 14.8
converge?
The answer is negative, in general. This is the reason why the following definition
was introduced.
Definition 14.39 Let X be a Banach space, .F : X → R be a differentiable function.
The function F satisfies the Palais-Smale condition ((PS) for short) if and only if
every sequence .{un }n in X such that .{F (un )}n is bounded and .*DF (un )* → 0 as
.n → +∞ has a convergent subsequence.
The (PS) condition—which can be considerably weakened—is the basic com-

pactness condition of Critical Point Theory. We refer to [6] for a readable exposition
of this theory.
14.17 Problems
14.1 Let (X, * · *) be a normed space, and let f : X → R be the function such that
f (u) = *u* for every u ∈ X. Prove that the Gâteaux derivative of f at u = 0 does
not exist.
14.2 Let σ : Rn \ {0} → Rn be the function such that σ (x) = x/*x* for every
x ∈ Rn \ {0}. Prove that
1 h
.Dσ (x) : h ∈ Rn → − "x | h#x + ,
*x*3 *x*
where "· | ·# denotes the usual inner product of Rn and * · * the associated Euclidean
norm.
14.17 Problems 357
14.3 Let X and Y be normed spaces, U an open subset of X, and F : U → Y a

Fréchet-differentiable function. Prove that F is a Lipschitz function if and only if
there exists a number M > 0 such that *DF (u)*L(X,Y ) ≤ M for every u ∈ U .
14.4 A subset C of a vector space X is conic if and only if tC = {tx | x ∈ C} ⊂ C
for every t > 0. A function f : C → R defined on a conic subset C is homogeneous
of degree p ∈ Z if and only if f (tx) = t p f (x) for every x ∈ C and for every t > 0.
Prove the following result, due to Eulero: suppose that C is open and conic in Rn .
A differentiable function f : C → R is homogeneous of degree p if and only if

n
pf (x) =
. xk ∂k f (x) for every x ∈ C.
k=1
14.5 A function F : X → Y between two real Banach spaces is Hadamard-

differentiable at a point u ∈ X if and only if there exists A ∈ L(X, Y ) such that
F (u + tn vn ) − F (u)
. lim = Av
n→+∞ tn
for all v ∈ X, for all sequences {tn }n in R\{0} such that tn → 0 and for all sequences
{vn }n in X such that vn → v. Similarly, F is weakly Hadamard-differentiable at
u ∈ X if and only if there exists A ∈ L(X, Y ) such that
F (u + tn vn ) − F (u)
. lim = (Av)
n→+∞ tn
for all v ∈ X and all ∈ Y ∗ , for all sequences {tn }n in R \ {0} such that tn → 0 and
for all sequences {vn }n in X such that vn $ v. Prove the following statements:
1. F is Fréchet-differentiable at u ∈ X if and only if there exists A ∈ L(X, Y ) such
that
F (u + tv) − F (u)
. lim = Av
t →0 t
uniformly with respect to v in bounded subsets of X.

2. F is Hadamard-differentiable at u ∈ X if and only if there exists A ∈ L(X, Y )
such that
F (u + tv) − F (u)
. lim = Av
t →0 t
uniformly with respect to v in compact subsets of X.

3. F is weakly Hadamard-differentiable at u ∈ X if and only if there exists A ∈

L(X, Y ) such that for all ∈ Y ∗ ,
F (u + tv) − F (u)
. lim = (Av)
t →0 t
uniformly with respect to v in weaklycompact subsets of X.

4. Let X = R and Y = L2 (R). Pick z ∈ Y \ {0} and define F : R → L2 (R) by
tz(· + t −1 ) for t = 0
F (t) =
.
0 for t = 0.
Prove that F is weakly Hadamard-differentiable at t = 0 with F % (0) = 0. Prove

that
* *
* F (t) *2
* *
* t * = *z*Y
2
. for all t = 0,
Y
and deduce that F is not Gâteaux-differentiable at t = 0.
14.18 Comments
Linear Functional Analysis is precisely concerned with topological vector spaces.

i.e. vector spaces endowed with a topology compatible with the two algebraic
operations. This means that the sum and the product with scalar numbers must be
continuous functions. Normed vector spaces are important examples of topological
vector spaces, and many mathematical analysts do use them in everyday life. We
refer to [5, 7] for a thorough study of topological vector spaces.
Our definition of tangent functions and of the Fréchet derivative follows [4]. It
seems that the French school of Bourbaki popularized the theory of differential
calculus in normed spaces. We have followed the elegant survey of [1]. The
interested reader should refer to [2] for a treatise on differential calculus in infinite-
dimensional spaces.
The remarkable global inversion theorem appears in [3], and generalizes similar
results of [1].
References 359
References
1. A. Ambrosetti, G. Prodi, A Primer of Nonlinear Analysis (Cambridge University Press, 1995)

2. H. Cartan, Cours de calcul différentiel. Collection Méthodes (Hermann, Paris, 1997)
3. G. De Marco, G. Gorni, G. Zampieri, Global inversion of functions: an introduction. NoDEA
1, 229–248 (1994)
4. J. Dieudonné, Foundations of Modern Analysis. Pure and Applied Mathematics, vol. 10
(Academic Press, New York, 1960)
5. J.L. Kelley, I. Namioka, Linear Topological Spaces. Graduate Texts in Mathematics 36
(Springer, New York, 1976)
6. J. Mawhin, M. Willem, Critical Point Theory and Hamiltonian Systems. Applied Mathematical
Sciences, vol. 74 (Springer, 1989)
7. W. Rudin, Functional Analysis. International Series in Pure and Applied Mathematics
(McGraw-Hill, New York, 1991)
Chapter 15
A Functional Approach to Lebesgue
Integration Theory
Abstract The standard approach to the Lebesgue integral is via measure theory:
we must define a set function—called a measure—on a set of suitable sets—called
measurable sets, then we can define measurable functions, and finally integrable
functions. The main advantage of this approach is that at the end we have the
highest generality. On the other hand, such a construction requires a good amount
of mathematical education before it can be understood.
15.1 The Riemann Integral in Higher Dimension
Let us present the basic construction of the Riemann integral in .Rn . We will see that
much the same ideas that we developed in .R can be adapted. We start with some
notation.
Definition 15.1 (Cell) An n-cell is a cartesian product
.B = [a1 , b1 ] × · · · × [an , bn ]
of closed and bounded intervals of .R. The volume of B is defined as in elementary

geometry:
(
n
. Vol(B) = (bk − ak ).
k=1
We often fix a basic cell, which we denote by .B.

Let .f : B → R be a bounded function. A partition P of .B is a splitting of
.B into sub-cells (i.e. each interval .[ak , bk ] is partitioned exactly as we did in our
construction of the Riemann integral in one dimension) .B1 , . . . Bp . To each partition
362 15 A Functional Approach to Lebesgue Integration Theory
P we associated the numbers
Mk = sup {f (x) | x ∈ Bk }
.
mk = inf {f (x) | x ∈ Bk } ,
for .k = 1, . . . , p. Again we define

p
U (P , f ) =
. Mk Vol(Bk )
k=1

p
L(P , f ) = mk Vol(Bk ),
k=1
and again it is easy to check that if .P % is finer than P , then
L(P , f ) ≤ L(P % , f ),
. U (P % , f ) ≤ U (P , f ).
Furthermore, if .P1 and .P2 are partitions of .B, then
.L(P1 , f ) ≤ U (P2 , f ).
The number
. f = sup {L(P , f ) | P is a partition of B}

B
is the lower Riemann integral of f over .B, and
. f = inf {L(P , f ) | P is a partition of B}

B
is the upper Riemann integral of f over .B.

Definition 15.2 The function f is R-integrable on .B if
. f = f.
B B
If this is the case, the common value is denoted by
. f,
B
and is called the Riemann integral of f on .B.

15.3 Null and Full Sets 363
The theory of the Riemann integral can now be developed as it was developed in
R, but this is not our aim. We are going to construct a different integral, called the
.
Lebesgue integral.
15.2 Elementary Integrals
Definition 15.3 Suppose that H is a set of bounded real-valued functions defined

on a set X, and suppose that
(a) H is a vector space with the usual addition and multiplication by real numbers;
(b) if h ∈ H , then |h| ∈ H .
The elements of H will be called elementary functions.
Exercise 15.1 If h ∈ H , prove that h+ = max{h, 0} and h− = max{0, −h} belong
to H . Hint: recall that 2x + = |x| + x and 2x − = |h| − x for each x ∈ R.
Definition 15.4 An elementary integral on H is a function I : H → R such that
(1) (linearity) if h, k ∈ H , α, β ∈ R, then
I (αh + βk) = αI h + βI k.
.
(2) (positivity) If h ∈ H and h ≥ 0, then I h ≥ 0.

(3) (continuity) If {hp }p is a non-increasing sequence of elementary functions
which converges pointwise to zero on X, then limp→+∞ I hn = 0.
Exercise 15.2 Let h, k be elementary functions. Prove that h ≤ k implies I h ≤ I k.

Hence positivity of I is actually monotonicity of I .
15.3 Null and Full Sets
Definition 15.5 (Measure Zero) A subset Z of X is a set of measure zero, or a null

set, if and only if for every ε > 0 there exists a non-decreasing sequence {hp }p of
elementary functions such that
I hp < ε,
.
sup hp (x) ≥ 1 for each x ∈ Z.

p
On the opposite side, a subset E of X is a set of full measure, if X \ E is a set of

measure zero.
Definition 15.6 Let P (x) be a logical statement depending on the free variable x.
If
. {x ∈ X | P (x)}
is a set of full measure, we say that the property P holds for almost every x ∈ X, or
that P holds almost everywhere in X.
Example 15.1 Let {hp }p be a sequence of elementary functions. We say that hp →
0 almost everywhere on X if hp (x) → 0 (as p → +∞) for every x is a subset of
full measure of X.
Here is a first application of our definitions. The reader is invited to compare the
next statement with the third property of the elementary integral.
Proposition 15.1 Suppose that a sequence {hp }p of non-negative elementary
functions is non-increasing and converges to zero almost everywhere. Then
. lim I hp = 0.
p→+∞
Proof Recalling that elementary functions are bounded, we set
M1 = sup {h1 (x) | x ∈ X} .

.
Let Z be the subset of X on which the sequence {hp }p does not converge to zero: by
assumption, Z is a null set. If ε > 0, there exists a non-decreasing sequence {kp }p
of elementary functions such that
ε
I kp <
.
M1
and supp kp (x) ≥ 1 for every x ∈ Z. The two limits
ε
. lim I hp ≥ 0, lim I kp ≤
p→+∞ p→+∞ M1
exist by monotonicity and boundedness. Furthermore the sequence {hp − M1 kp }p is

non-increasing and has a non-positive limit everywhere. By the continuity property
of I ,
+
.I hp − M1 kp ≤ I hp − M1 kp → 0,
and therefore
. lim I hp − M1 lim I kp ≤ 0.
p→+∞ p→+∞
15.4 The Class L+ 365
But then
ε
0 ≤ lim I hp ≤ M1 lim I kp ≤ M1 ·
. = ε.
p→+∞ p→+∞ M1
Since ε > 0 was arbitrary, we deduce that limp→+∞ I hp = 0.

Example 15.2 A countable union of subsets of measure zero is a subset of measure
zero. Indeed, if Z1 , . . . , Zn , . . . is a countable family of null subsets, for every ε > 0
and every n there exists a non-decreasing sequence {hn,p }p of elementary functions
such that I hn,p < 2−n ε and supp hn,p ≥ 1 on Zn . Then the sequence defined by

hp = max h1,p , . . . , hn,p
.
∞
is non-decreasing, supp hp ≥ 1 on n=1 Zn , and

p
I hp ≤
. I hn,p ≤ ε.
n=1
Exercise 15.3 Prove that a countable intersection of subsets of full measure is a

subset of full measure.
We will write hp - f to mean that h1 ≤ h2 ≤ h3 ≤ . . . and hp (x) converges

to f (x) for almost every x ∈ X. The symbol hp . f has an analogous
definition.
Remark 15.1 Quite often the arrows - and . mean monotone convergence at
every point. Be careful, since we will always use them with almost everywhere
convergence.
15.4 The Class L+
Definition 15.7 A function f : X → (−∞, +∞] belongs to the class L+ (or

L+ (X) when confusion may arise) if there exists a sequence {hp }p in H such that
hp - f and
C = sup I hp ∈ R.
.
p
Proposition 15.2 If f ∈ L+ , then f (x) < +∞ for almost every x ∈ X.

Proof Let Z be the set of all x ∈ X such that f (x) = +∞. Replacing hp with
hp − h1 , we may assume that each hp is non-negative. Discarding a set of measure
zero, we may then assume that {hp }p is non-decreasing and converges to +∞ on
the whole set Z.
Pick ε > 0 and x ∈ Z; then the inequality
C
hp (x) ≥
.
ε
must holds from some value of p onwards. Hence Z is covered by the countable
family of sets

C
. x hp (x) ≥ , p ∈ N.
ε
Hence
εhp (x)
. sup ≥1
p C
and
εhp ε
I
. = I hp ≤ ε.
C C
Therefore Z is a set of measure zero.

Let f ∈ L+ ; we know that f is almost everywhere the increasing limit of a
sequence {hp }p of elementary functions. Suppose that f is also almost everywhere
the increasing limit of another sequence {kp }p of elementary functions. We claim
that
. lim I hp = lim I kp .
p→+∞ p→+∞
Since the two sequences can be swapped, it suffices to prove that
. lim I hp ≤ lim I kp . (15.1)

p→+∞ p→+∞
Fix an index m, and consider the non-increasing sequence
n → hm − kn .
.
15.4 The Class L+ 367
As n → +∞, this sequence converges to hm − f ≤ f − f ≤ 0. Therefore (hm −

kn )+ . 0, and by the continuity property of the integral
I (hm − kn )+ . 0.
.
But I (hm − kn ) ≤ I (hn − km )+ , it follows that
n → I (hm − kn ) = I hm − I kn
.
is non-increasing and has a non-positive limit, which in turn implies that I hm ≤

limn→+∞ I kn . Since m was arbitrary, we deduce that (15.1) holds. It is now
legitimate to propose the following definition.
Definition 15.8 The integral of a function f ∈ L+ is defined as
If = lim I hp ,
.
p→+∞
where {hp }p is any sequence as in Definition 15.7.

Exercise 15.4 More generally, suppose that {hp }p and {kp }p are two sequences of
elementary functions such that hp - f , kp - g, f ≤ g almost everywhere, and
. sup I hp ∈ R
p
sup I kp ∈ R.
p
Prove that If ≤ Ig.

The elements of L+ have some useful features that we now describe. The standard
proofs are left to the reader.
1. If f ∈ L+ and g ∈ L+ , then f + g ∈ L+ and there results
I (f + g) = If + Ig.
.
2. If f ∈ L+ and α ∈ [0, +∞), then αf ∈ L+ and there results
I (αf ) = α If.
.
3. If f ∈ L+ and g ∈ L+ , then max{f, g} ∈ L+ and min{f, g} ∈ L+ .

Important: Instability Under Multiplication

On the contrary, the class L+ is not stable under multiplication by negative numbers,
since such a multiplication turns increasing sequences into decreasing sequences.
As a deeper property, we prove that the class L+ is closed under limits of

increasing sequences with bounded integrals.
Theorem 15.1 Suppose fn ∈ L+ for n = 1, 2, . . ., fn - f and supn Ifn ∈ R.
Then f ∈ L+ and
If = lim Ifn .
.
n→+∞
Proof Given n ∈ N, we construct sequences of elementary functions such that
h11 ≤ . . . ≤ h1n ≤ . . . ,
. h1n - f1
h21 ≤ . . . ≤ h2n ≤ . . . , h2n - f2
···
hk1 ≤ . . . ≤ hkn ≤ . . . , hkn - fk
···
according to the definition of L+ . We then set
.hn = max {h1n , . . . , hnn } .
It is clear that hn is an elementary function, and that hn ≤ hn+1 for every n. Since
hn ≤ max {f1 , . . . , fn } = fn , we see that
. sup I hn ≤ sup Ifn ∈ R.

n n
Setting f ∗ = limn→+∞ hn , by definition of L+ there results f ∗ ∈ L+ , and
. If ∗ = lim I hn .
n→+∞
We claim that f ∗ = f almost everywhere in X. Indeed, for any k ∈ N and any

n ≥ k, we have hkn ≤ hn ≤ fn , so that fk ≤ f ∗ ≤ f . But fn - f by assumption,
and the claim is proved. Finally, I hkn ≤ I hn ≤ Ifn ≤ If and I hn - If ∗ = If
imply If = limn→+∞ Ifn .

15.5 The Class L of Integrable Functions 369
A typical and nice application of the previous convergence result is to series of

functions.
Theorem 15.2 Suppose that gk ∈ L+ and gk ≥ 0 in X for every k ∈ N. If there
exists a constant C ∈ R such that
" n #

.I gk ≤ C n ∈ N,
k=1
∞
then f = k=1 gk ∈ L+ and there results

n ∞

If = lim
. Igk = Igk .
n→+∞
k=1 k=1
Proof Since each gk ≥ 0, the sequence

n
n →
. gk
k=1
is non-decreasing, and Theorem 15.1 applies.

15.5 The Class L of Integrable Functions
We can now complete the last step of our construction. We have seen that the
class .L+ is not closed under multiplication by negative numbers, so that we cannot
subtract elements of .L+ . This is a gap we need to fill by enlarging the class .L+ .
Definition 15.9 The class .L = L(X) of integrable functions on X is the set of all
functions .ϕ on X which can be represented almost everywhere as .ϕ = f − g, for
some .f ∈ L+ and .g ∈ L+ .
It is evident from the previous discussion that L enjoys the following proper-
ties:
1. if .ϕ1 = f1 − g1 and .ϕ2 = f2 − g2 are elements of L, then .ϕ1 + ϕ2 = (f1 + f2 ) −
(g1 + g2 ), and therefore .ϕ1 + ϕ2 ∈ L.
2. If .ϕ = f − g and .α ∈ R, then .αϕ ∈ L, Indeed, if .α ≥ 0, then .αϕ = αf − αg
and .αf , .αg belong to .L+ . If .α < 0, then .−α > 0 and .αϕ = (−α)g − (−αf ),
and it follows again that .αϕ ∈ L.
3. If .ϕ ∈ L, then .ϕ + , .ϕ − and .|ϕ| belong to L. Indeed, from .ϕ = f − g it follows
that
|ϕ| = max{f, g} − min{f, g}

.
belongs to L. Then
|ϕ| + ϕ
ϕ+ =
.
2
|ϕ| −ϕ
ϕ− =
2
also belong to L by linearity.
We propose a formal definition, which will immediately be justified.
Definition 15.10 The integral .I ϕ of a function .ϕ ∈ L is defined as
I ϕ = If − Ig,
.
where .ϕ = f − g, .f ∈ L+ and .g ∈ L+ .
Proposition 15.3 The integral of .ϕ ∈ L is independent on the representation .ϕ =
f − g.
Proof Let .ϕ = f − g = f1 − g1 , for suitable functions f , .f1 , g and .g1 in .L+ . Since
.f + g1 = g + f1 , we have .I (f + g1 ) = I (g + f1 ), or .If + Ig1 = Ig + If1 . This
is equivalent to
.I (f − g) = I (f1 − g1 ),
which completes the proof.

Remark 15.2 It should be noted that mathematical analysts do not usually appreci-
ate definitions based on arbitrary choices of something. These definitions are typical
of abstract algebra, whilst analysis tends to be more constructive. The previous
definition of the (abstract!) integral is an exception to the rule.
Exercise 15.5 Prove that I is a linear operator: .I (f + g) = If + Ig and .I (αf ) =
αIf for every f and g in L, and every .α ∈ R.
Proposition 15.4 If .ϕ ∈ L and .ϕ ≥ 0 in X, then .I ϕ ≥ 0.
Proof If .ϕ = f − g, we must have .f ≥ g, and therefore .If ≥ Ig by monotonicity.

Corollary 15.1 For every .ϕ ∈ L, there results .|I ϕ| ≤ I (|ϕ|).
Proof Since .ϕ ≤ |ϕ| and .−ϕ ≤ |ϕ|, the conclusion follows from the previous
Corollary.

15.6 Taking Limits Under the Integral Sign 371
Important: Notation
!
We have carefully
! avoided the use of the integral symbol . for our abstract integral.
The use of . X ϕ is clearly possible—and often useful—, but our choice of the letter
I highlights the functional nature of our integral. In other words, in our approach
the integral is a linear operator which acts on the vector space .L(X), and we prefer
to encourage this viewpoint.
15.6 Taking Limits Under the Integral Sign
When we were dealing with sequences and series of functions, we realized quite
easily that a pointwise convergence of the integrands does not imply the convergence
of the integrals. Uniform convergence was a successful replacement, but the
weakness of the Riemann integral with respect to limits remains a matter of facts.
We want to convince the reader that the Lebesgue (abstract) integral I is much more
flexible. We begin with a technical lemma.
Lemma 15.1 Any .ϕ ∈ L admits a representation with the following property: for
every .ε > 0, there exist .f ∈ L+ and .g ∈ L+ such that .ϕ = f − g, .g ≥ 0 and
.Ig < ε.
Proof By definition, .ϕ = f − g for some f and g in .L+ . Let .hn - g be a sequence

of elementary functions such that .Ig = limn→+∞ I hn . We can write
ϕ = f − g = (f − hn ) − (g − hn ) = fn − gn .
.
Now .fn ∈ L+ since .fn = f − hn is the sum of two elements of .L+ , and similarly
+
.gn = g − hn ∈ L . Clearly .gn ≥ 0 for n sufficiently large, and .Ign < ε because
.limn→+∞ Ign = 0.

Theorem 15.3 (Beppo Levi) Assume that .ϕk ∈ L, .ϕk ≥ 0 for every .k ∈ N, and
" #

n
I
. ϕk ≤ C, n∈N
k=1
∞
for some suitable constant C. Then .ϕ = k=1 ϕk belongs to L, and
∞

.Iϕ = I ϕk .
k=1
Proof We use Lemma 15.1 to decompose .ϕk = fk −gk , where .fk , .gk ∈ L+ , .gk ≥ 0
and .Igk < 2−k , for every .k = 1, 2, . . .. Since .ϕk ≥ 0,
it turns out that .fk ≥+ 0. It is
easy to check that Theorem 15.2 applies, hence .g = ∞ k=1 gk belongs to .L and
∞

Ig =
. Igk .
k=1
∞
∞ same conclusion holds also for .fk , and then .f =
The k=1 fk ∈ L+ , .If =
k=1 Ifk . Putting everything together we see that
∞
∞
∞

ϕ=
. ϕk = fk − gk = f − g ∈ L
k=1 k=1 k=1
and
∞
∞

I ϕ = If − Ig =
. I (fk − gk ) = I ϕk .
k=1 k=1

Recalling that sequences and series are the same mathematical object, we deduce
the following statement.
Theorem 15.4 If .ψn ∈ L for .n = 1, 2, . . ., .ψn - ψ and .I ψn ≤ C for every n,
then .ψ ∈ L and .I ψ = limn→+∞ I ψn .
Proof We set .ϕ1 = ψ1 , .ϕn = ψn −ψn−1 , and the conclusion follows from Theorem
15.3.

A very useful consequence of Beppo Levi’s Convergence Theorem is a sort of
characterization of nonnegative functions whose integral vanishes.
Exercise 15.6 Prove that the integral of a function .ϕ ∈ L such that .ϕ = 0 almost
everywhere is zero.
The converse implication is contained in the next result.
Theorem 15.5 Let .ϕ0 ∈ L be a non-negative function such that .I ϕ0 = 0. Then
ϕ0 = 0 almost everywhere in X.
.
Proof Define the sequence .n → ϕn = nϕ0 . Clearly .ϕn ≥ 0 and .I ϕn = nI ϕ0 = 0

for each n, and .ϕn - ϕ, where .ϕ = 0 on the set where .ϕ0 = 0, and .ϕ = +∞ on
the set where .ϕ0 > 0. By Beppo Levi’s Theorem, .ϕ ∈ L, so that .ϕ = +∞ only on
a set of measure zero, and in particular .ϕ0 > 0 only on a set of measure zero.

Exercise 15.7 Show that the previous result is generally false if we drop the non-
negativity of .ϕ0 .
15.6 Taking Limits Under the Integral Sign 373
Corollary 15.2 If .Z ⊂ X, suppose that for every .ε > 0 there exists a sequence of
integrable functions
0 ≤ ϕε,1 ≤ ϕε,2 ≤ . . . ≤ ϕε,n ≤ . . .

.
such that .I ϕε,n < ε for each n and .supn ϕε,n ≥ 1 on Z. The Z is a set of measure
zero.
Proof Of course there is nothing to prove if each .ϕn is an elementary function. In
the general case, let .ϕε = limn→+∞ ϕε,n . By Beppo Levi’s Theorem, .ϕ is integrable
and
I ϕε = lim I ϕε,n ≤ ε.
.
n→+∞
We discretize .ε = 1/m, for .m ∈ N, so that we have the functions
ψ1 = ϕ1
.
ψ2 = min {ϕ1 , ϕ2 }
...
ψn = min {ϕ1 , ϕ2 , . . . , ϕn }
...
Each function .ψm is non-negative and .ψm ≥ 1 on the set Z. Furthermore .ψ1 ≥
ψ2 ≥ ψ3 ≥ . . . and
1
I ψm ≤ I ϕ1/m ≤
. .
m
Setting .ψ = limm→+∞ ψm , we see that .ψ ∈ L and .I ψ = limm→+∞ I ψm = 0. The
limit .ψ is non-negative, and .ψ ≥ 1 on Z. Hence the set .Z % = {x ∈ X | ψ(x) > 0}
has measure zero. But then Z has measure zero.

The most powerful Convergence Theorem for our integral is due to Lebesgue.
Theorem 15.6 (Dominated Convergence Theorem) Suppose that .ϕn ∈ L for
every n, .ϕn → ϕ almost everywhere, and there exists a function .ϕ0 ∈ L such
that
|ϕn (x)| ≤ ϕ0 (x),

. x ∈ X, n = 1, 2, . . . (15.2)
Under these assumptions, .ϕ ∈ L and .I ϕ = limn→+∞ I ϕn .

Proof The assumptions imply that .−ϕ0 ≤ ϕn ≤ ϕ0 , and in turn .−I ϕ0 ≤ I ϕn ≤

I ϕ0 . On a set of full measure we have .−ϕ0 ≤ ϕ ≤ ϕ0 , and hence .ϕ ∈ L. Indeed, we
can set
ψn (x) = sup {ϕn (x), ϕn+1 (x), . . .}

.
χn (x) = inf {ϕn (x), ϕn+1 (x), . . .} .
These functions are integrable and satisfy (almost everywhere) .−ϕ0 ≤ ψn ≤ ϕ0 ,

−ϕ0 ≤ χn ≤ ϕ0 . Restricting x to the subset of X on which .ϕn (x) converges to .ϕ,
.
we have
ψn (x) ≥ lim ϕn+p (x) = ϕ(x)

.
p→+∞
χn (x) ≤ lim ϕn+p (x) = ϕ(x).

p→+∞
Since .ψn+1 ≤ ψn , .χn+1 ≥ χn , the assumption .ϕn - ϕ implies .ψn . ϕ, .χn - ϕ,

and in conclusion a set of full measure we have .−ϕ0 ≤ ϕ ≤ ϕ0 .
Observing that .I χn - I ϕ, .I ψn . I ϕ, .I χn ≤ I ϕn ≤ I ψn , we conclude that
.I ϕn → I ϕ. The proof is complete.

A natural question is what happens if we weaken (15.2). Since (15.2) implies that
I (|ϕn |) ≤ I ϕ0 = C for every n, a good replacement is
.
. I (|ϕn |) ≤ C.
Theorem 15.7 (Fatou’s Lemma) Suppose that .ϕn ∈ L, .ϕn ≥ 0, .ϕn → ϕ almost
everywhere, and for some constant .C ≥ 0 we have .I (|ϕn |) ≤ C for every n. Then
.ϕ ∈ L and
0 ≤ I ϕ ≤ C.
.
Proof We define .χn = inf {ϕn , ϕn+1 , . . .}, observing that .χn ≤ χn+1 and .χn → ϕ
almost everywhere. Furthermore .χn ≤ ϕn , .I χn ≤ I ϕn ≤ C, and Beppo Levi’s
Theorem for sequences implies that .ϕ ∈ L. Since .I χn - I ϕ, it follows in particular
that .0 ≤ I ϕ = limn→+∞ I ϕn ≤ C.

15.7 Measurable Functions and Measurable Sets
The leitmotiv of Measure Theory is the adjective measurable: measurable sets,

measurable functions. Abstract Measure Theory resembles General Topology, with
the difference that open sets are replaced by the class of measurable sets. Once these
are defined, measurable functions come out naturally.
15.7 Measurable Functions and Measurable Sets 375
On the contrary, we have developed the Theory of the Integral by embedding

“good” functions into the class .L(X). The concept of measurability becomes less
evident, and we need to define it from scratch.
Definition 15.11 A function .ϕ : X → (−∞, +∞] is a measurable function if .ϕ <
+∞ almost everywhere, and .ϕ is almost everywhere the limit of a sequence of
elementary functions.
Theorem 15.8 Suppose that .ϕ is a measurable function such that .|ϕ(x)| ≤ ϕ0 (x)
for every .x ∈ X, where .ϕ0 ∈ L, then .ϕ ∈ L.
Proof Since .ϕ is the almost everywhere limit of a sequence .{hn }n of elementary
functions, we just observe that elementary functions are integrable, and the conclu-
sion follows from the Dominated Convergence Theorem.

The class of measurable function is an open door on a fascinating new world,
which we try to introduce in the next pages. First of all we list a few immediate
properties of this class.
1. if f and g are measurable, then .αf +βg is measurable for every choice of .α ∈ R,
.β ∈ R.
2. If f is measurable, so are .|f |, .f + and .f − .

3. Every .f ∈ L+ is measurable, since it is the limit of a sequence of elementary
functions. In particular every integrable function is measurable.
Coming to convergence, we have the following result
Theorem 15.9 If .{fn }n is a non-decreasing sequence of integrable functions
converging almost everywhere to a finite limit f , then f is measurable.
Proof We consider first the case .fn ∈ L+ , and pick .hnk ∈ H such that .hnk - fn .
If
hn = max {h1n , h2n , . . . , hnn } ,

.
then .{hn }n is a non-decreasing sequence of elementary functions which converges

to a limit .f ∗ . Since .n > k implies .hnk ≤ hn ≤ fn , letting .n → +∞ yields
.fk ≤ f
∗ ≤ f . This shows that .f ∗ is almost everywhere finite. Hence .f ∗ is a
measurable function, but then .f ∗ = f almost everywhere, since .fn - f , and f is

measurable.

Much the same conclusion holds for sequences of measurable functions.
Theorem 15.10 If a sequence .{fn }n of measurable functions converges almost
everywhere to a function f , then f is a measurable function.
Proof Splitting .fn = fn+ − fn− and .f = f + − f − , we may assume without loss
of generality that .fn ≥ 0 and .f ≥ 0. Each .fn is the limit of a sequence .{hnp }p
of elementary functions which may be assumed to be non-negative with positive
integrals. Consider then the function

∞
∞
hnp
ϕo =
. cnp ,
I hnp
n=1 p=1
∞
where the real coefficients .cnp are chosen so that the series . ∞ n=1 p=1 cnp
converges. By Beppo Levi’s Theorem, .ϕ0 is integrable, and .ϕ0 (x) > 0 whenever
.fn (x) > 0. It follows that f is also the limit of the sequence
gn (x) = min {f (x), nϕ0 (x)} .

.
According to Theorem 15.9, we need to show that the measurability of .fn implies
the measurability of .gn . But
gn (x) = min {f (x), nϕ0 (x)}

.
= lim min {fm (x), nϕ0 (x)}

m→+∞
and each function .min {fm (x), nϕ0 (x)} is measurable and bounded by the integrable
function .nϕ0 . The Dominated Convergence Theorem then yields the integrability of
.gn , and the proof is complete.

Corollary 15.3 If .{fn }n is a sequence of measurable functions, then .infn fn and
.supn fn are measurable functions. If each .fn is finite almost everywhere, then
.lim infn→+∞ fn and .lim supn→+∞ fn are measurable.
Once measurable functions have been defined, measurable sets do not come as a
surprise.
Definition 15.12 A subset .E ⊂ X is measurable if and only if its characteristic
function .χE is a measurable function.
We recall that
1 if x ∈ E
χE (x) =
.
0 if x ∈ X \ E.
Definition 15.13 A subset E of X has finite measure if and only if .χE ∈ L(X). In
this case, the measure of E is the number
µ(E) = I χE .
.
If .χE is not integrable, we set .µ(E) = +∞. In a conventionally way, we set .µ(∅) =
0.
15.7 Measurable Functions and Measurable Sets 377
Exercise 15.8
1. Prove that a measurable subset of a set of finite measure, is a set of finite measure.
2. Prove that any subset of a set of measure zero is a set of measure zero.
Important: For Experts of Measure Theory

If Measure Theory is introduced before the integral, as is customary in Geometric
Measure Theory, the fact that a subset of a set of measure zero has measure zero
is not assumed. Then the issue of completeness of a measurable space comes into
play. In this sense, the Measure Theory we are constructing is slightly less general,
and always produces complete measurable spaces. We will come back to this topic
in the next chapter.
Theorem 15.11 The union, intersection and difference of two measurable sets are
measurable sets.
Proof Let E, F be measurable sets. The conclusion follows from the identities
χE∪F = max {χE , χF }

.
χE∩F = min {χE , χF }

χE\F = χE − χF .

The previous result can be generalized to countable unions as follows.
∞
Theorem 15.12 If each set .En , .n ∈ N, is measurable, then .E = n=1 En is
measurable. Moreover, if .Ei ∩ Ej = ∅ whenever .i = j , then
∞

µ(E) =
. µ(En ). (15.3)
n=1
The case .+∞ = +∞ is not excluded in (15.3)

Proof It follows from Corollary 15.3 that

χE = sup χE1 , χE2 , . . . = lim sup χE1 , . . . , χEn
.
n→+∞
is a measurable function, hence E is a measurable set. To prove (15.3), we remark

that if .µ(En ) = +∞ for some n, then .µ(E) = +∞ because .E ⊃ En . On the
other hand, if .µ(En ) ∈ R for each n, then .χE = ∞
n=1 χEn . Beppo Levi’s Theorem
implies that .χE is integrable, and

∞

I χE =
. I χEn ,
n=1
∞
provided that the series . ∞n=1 I χEn = n=1 µ(En ) converges. Conversely,
∞ if .χE is
n
integrable,
∞ then . k=1 I χ Ek ≤ I χ E for each n, and hence the series . n=1 I χEn =
n=1 µ(En ) converges. We have proved that (15.3) holds in any case.

Corollary 15.4 Suppose that .En , .n ∈ N, is a measurable set.
∞
(a) If .E1 ⊂ E2 ⊂ . . ., then .E = n=1 En is measurable, and .µ(E) =
limn→+∞ µ(E n ).
(b) The set .F = ∞ n=1 En is measurable. If .E1 ⊃ E2 ⊃ . . . and .µ(E1 ) ∈ R, then
.µ(F ) = limn→+∞ µ(En ).
Proof If some .En has infinite measure, then .µ(E) = +∞, and there is nothing to
prove. Otherwise, we write
E = E1 ∪ (E2 \ E1 ) ∪ (E3 \ E2 ) ∪ . . .
.
so that E is a countable union of disjoint subsets. By Theorem 15.12 we see that

n
µ(E) = lim
. µ(Ek \ Ek−1 ) = lim µ(En ),
n→+∞ n→+∞
k=1
having set .E0 = ∅ for convenience. This proves (a).

To prove (b) we take complements and recall that the complement of an
intersection is the union of the complements. The measurability of F follows from
(a) in this way. As before, we write
E1 = F ∪ (E1 \ E2 ) ∪ (E2 \ E3 ) ∪ . . .
.
and apply (a) again. The term .µ(E1 ) cancels since it is finite, and the proof is
complete.

15.8 Integration Over Measurable Sets
Up to now, all our integrals have been computed “on the whole space X.” In
many applications it would be convenient to integrate functions over subsets of X.
Although we already have the best possible candidate, i.e. measurable subsets, a
problem arises: consider indeed .µ(X), the measure of the whole space. There is no
need for X to be a measurable set; but even if this were the case, .µ(X) should be
15.8 Integration Over Measurable Sets 379
defined as .I χX , or equivalently as .I (1), and in this case .µ(X) would be a finite

value.
As we will soon see, this fact would prevent us from constructing the (concrete)
Lebesgue measure on .Rn , since .Rn is a set of infinite measure to any reasonable
mind. To overcome this obstacle, we need to introduce a new axiom.
Stone’s Axiom The collection H of elementary functions satisfies
(c) If .h ∈ H , then .min {h, 1} ∈ H .
(d) There exists a sequence .{hn }n of non-negative elementary functions such that
.I hn > 0 and .supn hn (x) > 0 for every .x ∈ X.
Example 15.3 Axiom (c) extends to measurable functions: if .ϕ = limn→+∞ hn is

a measurable function, then .min{ϕ, 1} = limn→+∞ min{hn , 1} is measurable.
Example 15.4 Axiom (d) yields the existence of an integrable function .ϕ0 such that
ϕ0 (x) > 0 at every .x ∈ X. Indeed,
.
∞
1 hn (x)
ϕ0 (x) =
.
n2 I hn
n=1
does the job, where .{hn }n is the sequence considered in Axiom (d).
Theorem 15.13 The function .χX is measurable, so that X is a measurable set. In
particular, the set .X \ E is measurable for every measurable set .E ⊂ X.
Proof Actually, .1 = limn→+∞ min {1, nϕ0 }, and Axiom (c) ensures that .{1, nϕ0 } is
measurable.

Corollary 15.5 If .ϕ is a measurable function and a, b, c are real numbers, then
. min {ϕ, c} , max {ϕ, c} , max {min {ϕ, b} , a}
are measurable functions.

The next result contains the characterization of measurable functions which is
introduced in Geometric Measure Theory.
Theorem 15.14 Let .ϕ be a function, almost everywhere finite. The set
. [ϕ > c] = {x ∈ X | ϕ(x) > c}
is measurable for every .c ∈ R if and only if the function .ϕ is measurable.

Proof If .ϕ is measurable, then so is the function
min{ϕ, c + 1/n} − min{ϕ, c}

ϕn,c =
.
1/n
for every c and n. Keeping c fixed, we see that
0 if ϕ(x) ≤ c
. lim ϕn,c (x) =
n→+∞ 1 if ϕ(x) > c.
Hence .χ[ϕ>c] is the limit of measurable functions, and this implies the measurability
of .[ϕ > c].
Conversely, suppose that .[ϕ > c] is measurable for every c. Then
. {x ∈ X | c < ϕ(x) ≤ d} = [ϕ > c] \ [ϕ > d]
is also measurable for every .c < d. Given .n ∈ N, we consider the function .ϕn equal
to .k/n on the measurable sets

k k+1
Ek,n
. = x ∈ X < ϕ(x) ≤ ,
n n
k ∈ Z. The function .ϕn is defined almost everywhere, and differs from .ϕ by at most
.
1/n. Moreover
.
∞
k
ϕn =
. χE ,
n k,n
k=−∞
and hence .ϕn is measurable. Since .ϕn → ϕ uniformly on X, .ϕ is a measurable

function, and the proof is complete.

Proposition 15.5 The product of two measurable functions is a measurable func-
tion.
Proof We suppose that f and g are measurable functions. Without loss of gener-
ality, we may assume that .f ≥ 0 and .g ≥ (the general case follows by writing
.f = f
+ − f − and .g = g + − g − ). Given .c ∈ R, we observe that
c

{x ∈ X | f (x)g(x) > c} =
. x ∈ X f (x) > r, g(x) > r ∈ Q, r > 0 ,
r
and the conclusion follows from the fact that .Q is a countable set.

Corollary 15.6 If f is a measurable function and E is a measurable set, then .f χE
is a measurable function. In particular, if f is integrable and E is measurable, then
.f χE is integrable.
The integral of a function on a subset is now defined naturally.

15.8 Integration Over Measurable Sets 381
Definition 15.14 A function .ϕ defined on X is integrable (or measurable) on a

measurable .E ⊂ X if .ϕχE is integrable (or measurable). In the first case, the integral
of .ϕ on E is defined to be
. ϕ dµ = I (ϕχE ) .
E
!
At this point, we will freely write . X ϕ dµ for .I ϕ, as a particular case. Of
course this is just a customary piece of notation that adds no content to the
formal .I (ϕχE ).
We collect several easy properties of the integral over measurable sets.

Proposition 15.6 If .ϕ is integrable on E and .|ϕ| ≤ M on E, then
. |ϕ| dµ ≤ Mµ(E).
E
Proof Since .χE |ϕ| ≤ MχE on X, we see that
. |ϕ| dµ = I (χE |ϕ|) ≤ MI χE = Mµ(E).

E

Proposition 15.7 If .{En | n ∈ N} is a countable family of mutually disjoint
measurable sets, and if .ϕ is integrable (resp. measurable) on .E = n∈N En , then
.ϕ is integrable (resp. measurable) on each .En . If .ϕ is integrable on E, then
∞

. ϕ dµ = ϕ dµ.
E n=1 En
Proof If .χE ϕ is measurable (resp. integrable) on X, then so is the product .χEn χE ϕ.

Moreover .χE1 + χE2 + · · · = χE , so that .χE1 ϕ + χE2 ϕ + · · · = χE ϕ. If .ϕ is
integrable, the partial sums of the series in the left-hand side are bounded by the
integrable function .χE ϕ, and we can integrate term by term. This completes the
proof.

Proposition 15.8 If .{En | n ∈ N} is a countable family of
measurable sets, and if
ϕ is measurable on each .En , then .ϕ is measurable on .E = ∞
. n=1 En .
Proof Indeed,
χE ϕ = ϕχE1 + ϕχE2 \E1 ∩E2 + ϕχE3 \E1 ∩E3 \E2 ∩E3 + · · ·

.
and each function on the right-hand side is measurable. By Theorem 15.10 the
function .χE ϕ is measurable.

Theorem 15.15 (Absolute Continuity of the Integral) Suppose that .ϕ is an
integrable function on X. For every .ε > 0 there exists .δ > 0 such that if E is a
measurable set and .µ(E) < δ, then

.
ϕ dµ < ε.
E
Proof Pick an elementary function h such that

ε
I (||ϕ| − h|) <
. .
2
Since the function h is bounded, say .0 ≤ h ≤ M, we set .δ = Mε/2 and compute

. ϕ dµ ≤ |ϕ| dµ

E E
≤ ||ϕ| − h| dµ + h dµ
E E
ε
≤ + Mδ < ε
2
whenever E is a measurable set with .µ(E) < δ.

15.9 The Concrete Lebesgue Integral
So far we have proposed a completely abstract construction of the integral. However,

there is no reason why we should call I an integral, since it appears as a map with
some selected properties. In this section we want to show that a true integration
theory on .Rn can be constructed by applying our abstract scheme to the Riemann
integral in n dimensions.
To this aim we need a good class H of elementary functions, and of course a
good map I on H which satisfies the necessary axioms. The sketch we proposed in
Sect. 15.1 is not a complete solution yet, so how do we select the vector space H ?
At least two possible constructions are possible: the first one is based on step
functions, the second one on continuous functions. The resulting integrals are
equivalent, but this must be proved. We begin with step functions.
15.9 The Concrete Lebesgue Integral 383
Definition 15.15 Let .B = B1 ∪B2 ∪· · ·∪Bn be a partition of the basic n-cell .B, and
we suppose that the different sub-cells .Bj do not have interior points in common.1
Any function h such that
⎧
⎪
⎪h1 if x ∈ B1
⎪
⎪
⎪
⎨h2 if x ∈ B2
.h(x) = ..
⎪
⎪
⎪
⎪ .
⎪
⎩h
n if x ∈ Bn
for suitable real numbers .h1 , . . . , hn is a step function. The collection of all step
functions on .B is denoted by .H (B).
Remark 15.3 The previous definition is somehow troublesome, since h might be
defined in different ways on the layer .Bi ∩ Bj , for .i = j . This is irrelevant to us, and
we might even leave step functions undefined on the interface layers of the partition.
As we will see in a moment, their values on such layers plays no role at all in our
construction.
Definition 15.16 Let .h ∈ H (B) be a step function, and consider the sets

Bj = x ∈ B h(x) = hj ,
. j = 1, . . . , n.
The integral of h is

n
Ih =
. hj Vol(Bj ).
j =1
It is now easy to check that .I : H (B) → R is an elementary integral in the sense of

Definition 15.4. Therefore our abstract extension produces a class .L(B) of integrable
functions and an integral associated to each integrable function on .B.
What happens if we start with the collection .H̃ of continuous functions defined
on .B? Clearly enough, this is a vector spaces that satisfies Definition 15.3, and each
continuous function has a Riemann integral as described in Sect. 15.1. The only
non-trivial fact to check is the continuity axiom, which follows from the next result.
Theorem 15.16 (Dini’s Lemma) A non-increasing sequence .{fn }n of continuous
functions converging pointwise to zero on a compact set K, converges uniformly to
zero on K.
Proof In Problem 11.5 a more general statement was proposed. For the reader’s
convenience we write here the proof. Fix any .ε > 0; to any point .x0 ∈ K there
1Geometrically, we are assuming here that the cells .Bj can touch each other only at their
boundaries.
corresponds an integer .m = m(x0 ) such that .fm (x0 ) < ε. By continuity, there
exists a neighborhood .U (x0 ) of .x0 such that .fm (x) < ε for every .x ∈ U (x0 ).
By monotonicity, if .p > m then .fp (x) ≤ fm (x) < ε for every .x ∈ U (x0 ).
As .x0 ranges over K, the neighborhoods .U (x0 ) form an open cover of K: let
.{U (x1 ), U (x2 ), . . . , U (xν )} be a finite subcover of the compact set K. If q denotes
the smallest index of the functions which participate in this subcover, we see that
.fr (x) < ε for every .x ∈ K and every .r > q, and the proof is complete.

As before, we can now turn on the engine of our abstract extension machine, and
obtain a space .L̃(B) of integrable functions and a corresponding integral .I˜. We
now show that .L̃ = L and .I˜ = I , so that the concrete Lebesgue integral can be
equivalently defined in two ways.
Theorem 15.17 There results .L̃(B) = L(B) and .I˜ = I .
Proof Since the proof is long, we split it into several steps.
Step 1. Every continuous function f belongs to L. Indeed, let .ε > 0 be given,
and we select a partition .P = {B1 , . . . , Bm } of .B such that

m

. f− f (ξj ) Vol(Bj ) < ε,

B j =1
whenever .ξj ∈ Bj . This is equivalent to

. f − I hP < ε,

B
where .hP is the step function whose value on .Bj is .f (ξj ). Since .hP converges
uniformly to f as the partition P is indefinitely refined, it follows from the
Dominated Convergence Theorem that .f ∈ L(B) and
If = I˜f.
.
Step 2. Every step function h belongs to .L̃. This is actually a density statement
about the approximation of step functions by continuous function. We consider a
function h which is equal to 1 on a cell B and to 0 outside B. If the dimension n of
the space is 1, a “trapezoidal” graph shows that there exists a continuous (actually
piecewise affine) function that approximates h with any prescribed precision. If
.n > 1, we just pick such an approximating function .fi = fi (xi ) in each variable
.xi , and define .(x1 , . . . , xn ) → f1 (x1 )f2 (x2 ) · · · fn (xn ). As a consequence, each
step function h can be expressed as the limit of a sequence .{fm }m of continuous

functions such that, as we have shown above, .I˜fm = Ifm for each m. Using
again the Dominated Convergence Theorem, .h ∈ L̃ and .I˜h = limm→+∞ I˜fm =
limm→+∞ Ifm = I h.
15.9 The Concrete Lebesgue Integral 385
Step 3. Both constructions lead to the same sets of measure zero.2 Indeed, let
.Z̃ be a set of measure zero with respect to the integral .I˜. Given .ε > 0, there
exists a non-decreasing sequence of non-negative continuous functions .fm such

that .I˜fm < ε and .supm fm (x) ≥ 1 for every .x ∈ Z. By Step 1, .I˜fm = Ifm for
each m. By Beppo Levi’s Theorem, .Z̃ is a set of measure zero with respect to the
integral I . Conversely, let Z be a set of measure zero with respect to the integral
I . For this reason, there exists a non-decreasing sequence of step functions .hm
such that .I hm < ε and .supm hm (x) ≥ 1 for every .x ∈ Z. By Step 2, .hm ∈ L̃
and .I˜hm = I hm , so that we conclude again that Z is a set of measure zero with
respect to .I˜ as well.
Step 4. Monotone passages to the limit and formation of differences. Pick any
+
.f ∈ L , so that f is the limit (almost everywhere) of a non-decreasing sequence
of step functions .hm with bounded integrals .I hm . Then the integrals .I˜hm = I hm
remain bounded, too, and hence by Beppo Levi’s Theorem .f ∈ L̃, .I˜f = If .
On the other hand, if .f ∈ L̃+ , then f is the limit (almost everywhere) of a non-
decreasing sequence of continuous functions .fm with bounded integrals .I˜fm . As
before the integrals .Ifm = I˜fm remain bounded, .f ∈ L and .If = I˜f . By taking
differences, .L̃ contains every function of L, and vice-versa, with equal integrals.
The proof is now complete.

For the concrete Lebesgue integral, a geometric characterization of sets of
measure zero is possible. We remark that the next result is usually taken as the
definition of measure zero in Euclidean spaces.
Theorem 15.18 Let .Z ⊂ B be a subset of the basic block .B. The following
(a) for every .ε > 0 there exists a countable (i.e. finite or countably
∞ infinite)
collection of n-cells .B1 , B2 , . . . , such that .Z ⊂ ∞
k=1 Bk and . k=1 Vol(Bk ) <
ε;
(b) for every .ε > 0 there exists a non-decreasing sequence of non-negative step
functions
(ε)
h1 ≤ · · · ≤ h(ε)
. m ≤ ···
(ε)
such that .I hm < ε for every m and
. m (x) ≥ 1 for every x ∈ Z.

sup h(ε)
m
Proof Assume that (a) holds, and call .h(ε) m the step function which is equal to 1
on the cells .B1 , . . . , Bm and equal to 0 outside these cells. Now, every point .x0 ∈
2 Hence the sentence “almost everywhere” can be interpreted with respect to both integrals.
(ε)
Z belongs to some cell .Bm , hence .hm (x0 ) = 1. The remaining properties of the
(ε)
sequence .{hm }m are trivial, and therefore (b) holds.
Conversely, suppose that (b) holds. Let .B1 , . . . , Br1 be the collection of cells
(ε) (ε)
on which the function .h1 ≥ 1/2. Then .h2 is larger than .1/2 on the same cells,
and also on some other cells .Br1 +1 , . . . , Br2 . Iterating this construction, we obtain
an infinite collection of cells .B1 , . . . , Br1 , . . . , Br2 , . . . with no interior points in
common. Since
. m (x) ≥ 1
sup h(ε) for every x ∈ Z,
m
the set Z is covered by all these cells. To complete the proof, we need to compute
the sum of the volumes of such cells. We consider only the cells .B1 , . . . , Brm on
which .h(ε) (ε)
m ≥ 1/2, it follows from .I hm < ε that

rm
. Vol(Bk ) ≤ 2ε.
k=1

Letting .m → +∞ we derive . ∞ k=1 Vol(Bk ) ≤ 2ε. We must now remark that
the cells .Bk just considered may not cover Z, since the points of Z need not lie
in the interior of such cells. But this is not an obstruction, since wemay replace
% % ∞ %
.Bk by a concentric cell .B such that .Vol B
∞ k k = 2 Vol(Bk ), .Z ⊂ k=1 Bk and
%
. k=1 Vol(Bk ) ≤ 4ε. Hence (a) holds, and the proof is complete.

15.10 Integration on Product Spaces
Every student knows that .Rn = R × R × · · · × R (n times). So a natural question

is whether the integral on .Rn can be reduced to the integral on .R by some “product
rule.” The answer is essentially affirmative, and we present it in our style: first an
abstract statement, then a concrete construction.
Theorem 15.19 (Abstract Fubini Theorem) Let X and Y be two sets, and let
W = X × Y be their cartesian product. Assume that .L(X), .L(Y ), .L(W ) are spaces
.
of integrable functions on each set, and that .IX , .IY , .IW = I are the corresponding
abstract integrals. Assume moreover that the family .H (W ) of elementary functions
which generate .L(W ) has the following properties:
(a) for every function .h ∈ H (W ), the function .x → h(x, y) belongs to .L(X) for
almost every .y ∈ Y ;
(b) for every .x ∈ X, the function .y → IX h(x, y) belongs to .L(Y );
(c) .I h = IY (IX h).
15.10 Integration on Product Spaces 387
The family .L(W ) has the same properties: every function .ϕ ∈ L(W ) is integrable in
the first variable for almost every value of the second variable, the integral .IX ϕ(x, ·)
is integrable on Y , and .I ϕ = IY (IX ϕ).
Proof The proof is long, so we split it into several steps. We define the class . of
all functions .ϕ ∈ L(W ) for which the conclusion is true. We will eventually show
that . = L(W ). Since . contains all elementary functions by assumption, we only
need to prove the inclusion .L(W ) ⊂ .
Step 1. . is closed under the formation linear combinations. This is trivial, and
we omit the details.
Step 2. . is closed under monotonic limits. Let .{ϕn }n be a sequence in . which is
monotonic, and suppose that the sequence of the integrals .I ϕn remains bounded.
Then the pointwise limit .ϕ = limn→+∞ ϕn belongs to .. Indeed, suppose
for the sake of definiteness that .{ϕn }n is non-decreasing, and put .gn (y) =
IN ϕn (·, y).The the sequence .{gn }n is also non-decreasing and the sequence of
integrals .IY gn remains bounded. Furthermore
IY gn = IY (IX ϕn ) = I ϕn - I ϕ.
.
An argument based on Beppo Levi’s Theorem shows that .gn converges to an

integrable function g which is almost everywhere finite, and moreover .IY g =
limn→+∞ IY gn = I ϕ. Let E be the subset of Y of full measure on which the
function g is finite, and fix .y ∈ E. The sequence .{ϕ(·, y)}n is non-decreasing
and the sequence of integrals .IX ϕn (·, y) remains bounded: .IX ϕn (·, y) = gn - g.
Again the limit .ϕ(·, y) is integrable on X and
. lim IX ϕn (·, y) = g(y) = IX ϕ(·, y).

n→+∞
But then .I ϕ = IY g = IY (IX ϕ), and thus .ϕ ∈ .

Step 3. . contains every function z which is almost everywhere equal to zero.
Let .Z ⊂ W be the set of measure zero on which z is not zero. As a first step,
we assume that z takes on values between 0 and 1. For every .m ∈ N we select a
non-decreasing sequence .n → hm,n of non-negative elementary functions such
that
1
I hm,n <
. , lim hm,n ≥ 1 on Z.
m n→+∞
Replacing .hm+1,n by .min{hm+1,n , hm,n } we may also assume that .hm+1,n ≤

hm,n . The limit .hm = limn→+∞ hm,n is a monotonic limit of elementary
functions of ., and then belongs to . by Step 2. For the very same reason,
h = limm→+∞ hm belongs to ., and

.
1
I hm = lim I hm,n ≤
.
n→+∞ m
I h = lim I hm = 0.
m→+∞
Now .hm ≥ z on Z implies .h ≥ z on Z, and setting .y → g(y) = IX h(·, y) we

see that
. IY g = IY (IX h) = I h = 0
by Step 2. Therefore .g = 0 almost everywhere. For almost every .y ∈ Y we

deduce that .h(·, y) is almost everywhere equal to zero, and the same must hold
for z. It follows that .IX z(·, y), and hence
I z = 0 = IY (IX z).
.
We need to remove the condition on the range of z. Let .z ≥ 0 be an arbitrary

function that vanishes outside Z, and let
1 on Z
=
.
0 on W \ Z.
There results
1
z = lim n min , z ,
.
n→+∞ n
so that .z ∈ by the previous case. Finally, writing .z = z+ − z− recovers the

case of variable sign.
Step 4. Every element of .L+ (W ) belongs to .. Indeed, by definition every .f ∈
L+ (W ) is the almost everywhere non-decreasing limit of a sequence .{hn }n of
elementary functions:
hn - f,
. I hn - If.
We call .fˆ the limit of the sequence defined by
ĥ1 = h
.
ĥ2 = max {h1 , h2 }

···
ĥn = max {h1 , h2 , . . . , hn }
···
15.10 Integration on Product Spaces 389
This sequence is everywhere non-decreasing, and the functions .ĥn , .hn coincide
almost everywhere, so that .I ĥn = I h. As a result, the function f and the function
.fˆ coincide almost everywhere, and we can write .fˆ = f + z for some function z
which is almost everywhere equal to zero. By Steps 2 and 3, both .fˆ and z belong
to ., so that .f ∈ by Step 1.
Step 5. . contains every function .ϕ ∈ L(W ). Indeed, .ϕ can be written as the
difference of two elements of .L+ (W ).

Theorem 15.20 (Tonelli) Suppose that .ϕ is a measurable function on W such that
ϕ ≥ 0. If the iterated integral .IY (IX ϕ) exists, then .ϕ ∈ L(W ) and
.
I ϕ = IY (IX ϕ).
.
Proof Call .A = IY (IX ϕ), where .ϕ = limn→+∞ hn for some elementary

functions .hn . Then the functions
.ϕn = min {ϕ, max {h1 , . . . , hn }}
are measurable; furthermore the function .ϕn is dominated by the integrable function
max {h1 , . . . , hn }. Hence .ϕn is integrable on W , and
.
I ϕn = IY (IX ϕn ) ≤ A
.
by Fubini’s Theorem. Since .ϕn - ϕ and .I ϕn ≤ A, it follows from Beppo Levi’s

Theorem that .ϕ ∈ L(W ). A second application of Fubini’s Theorem completes the
proof.

Exercise 15.9 Consider the integrals
. e−xy sin x sin y dx dy. (15.4)

(0,+∞)×(0,+∞)
x2 − y2
dx dy, (15.5)
(0,1)×(0,1) (x + y )
2 2 2
with integrals computed in the concrete Lebesgue sense. Prove that the correspond-
ing iterated integrals exist for either order of integration, and that they coincide for
(15.4) but differ for (15.5). Finally, prove that both integrands are not integrable.
Taking account of the last exercise, we may say that Theorem 15.20 is a partial
converse to Theorem 15.19, and that the non-negativity assumption in Theorem
15.20 is optimal.
Let us stop for a moment. In the two main theorems of this Section, the existence
of a class of integrable functions was assumed. The main issue is now to construct
such a class and the corresponding integral.
Definition 15.17 Let X, Y be sets equipped with abstract integrals .IX and .IY . If
W = X × Y , the class .H (W ) of elementary functions on the product W consists of
.
all functions of the form

m
.h : (x, y) ∈ W → αj χEj (x)χFj (y),
j =1
where .m ∈ N, .αj ∈ R, and the sets .Ej ⊂ X, .Fj ⊂ Y are integrable sets (i.e.
χEj ∈ L(X), .χFj ∈ L(Y )) for every .j = 1, . . . , m.
.
Remark 15.4 It is clear that we can always assume .Ei ∩ Ej = ∅, .Fi ∩ Fj = ∅

whenever .i = j .
Exercise 15.10
Prove that .H (W ) satisfies the axioms of Definition 15.3. Hint: if
.h(x, y) = m j =1 αj χEj (x)χFj (y) is an element of .H (W ) such that .Ei ∩ Ej = ∅,
m
.Fi ∩ Fj = ∅ whenever .i = j , observe that .|h(x, y)| = j =1 |αj |χEj (x)χFj (y).
Definition 15.18 The elementary integral I on .L(W ) is defined by

m
Ih =
. αj µX (Ej )µY (Fj )
j =1
m
for every .h ∈ H (W ) of the form .h(x, y) = j =1 αj χEj (x)χFj (y). Of course
.µX (Ej ) = IX χEj and .µY (Fj ) = IY χFj .
Concerning Definition 15.4, we prove that .I hn → 0 if .hn . 0 in .H (W ), since the

linearity and the positivity of I are trivial. We claim that it suffices to prove that
.IX hn (·, y) . 0 for every y. Indeed if this is the case, then .IY (IX hn ) → 0 by Beppo
Levi’s Theorem. But the assumption .hn . 0 (everywhere) implies .IX hn (·, y) .
for every .y ∈ Y by Levi’s Theorem, and the claim is proved.
Once the elementary integral on .W = X × Y has been defined, our abstract
machinery yields a class .L(W ) of integrable functions and an associated integral I
which satisfies all the assumptions of Fubini’s Theorem.
15.11 Spaces of Integrable Functions
Let us start with an easy remark: if a function f is measurable, then .|f |p is

measurable for every fixed .p > 0. Indeed, if c denotes any (positive) real number,
then

. x ∈ X |f (x)| > c = x ∈ X |f (x)| > c
p 1/p
.
15.11 Spaces of Integrable Functions 391
Definition 15.19 For every .p > 0, the set .Lp = Lp (X) is defined to be the set of
all measurable functions f for which

I |f |p =
. |f |p dµ ∈ R.
X
Theorem 15.21 For every .p > 0, .Lp (X) is a vector space.

Proof Suppose that .f ∈ Lp , .g ∈ Lp . We already know that .f + g is measurable;
since

. |f + g| ≤ |f |p + |g|p ≤ (2 sup {|f |, |g|})p

= 2p sup |f |p , |g|p

≤ 2p |f |p + |g|p ,
we see that .f + g ∈ Lp . The fact that any real multiple of f belongs to .Lp is trivial,

Definition 15.20 The standard norm of .Lp (X) is defined by
1/p 1/p
. *f *p = I |f |p = |f |p dµ .
X
The fact that .* · *p is actually a norm follows from two fundamental inequalities.
Theorem 15.22
(a) (Hölder’s inequality) If .f ∈ Lp and .g ∈ Lq for some numbers .p > 1, .q > 1
such that
1 1
. + = 1,
p q
then
. |fg| dµ ≤ *f *p *g*q .
X
(b) (Minkowski’s inequality) If .f ∈ Lp , .g ∈ Lp , then
*f + g*p ≤ *f *p + *g*p .
.
Proof Let us write
A = *f *p ,
. B = *g*q .
If .A = 0 then .f = 0 almost everywhere, and .fg = 0 almost everywhere. Therefore

we need to consider only .A > 0 and .B > 0. We introduce
|f | |g|
F =
. , G= ,
A B
! !
so that . X F p dµ = 1 = X Gq dµ. If .x ∈ X is such that .0 < F (x) < +∞ and
.0 < G(x) < +∞, then there exist real numbers s and t such that
F (x) = es/p ,
. G(x) = et /q .
The convexity of the exponential function yields
s
+t es et
e
. p q ≤ + .
p q
It follows that
F (x)s G(x)q
F (x)G(x) ≤
. + .
p q
We integrate this inequality and get (a). The proof of (b) follows easily from (a).
Indeed, let us write
. |f + g|p = |f + g|p−1 · |f + g| ≤ |f | |f + g|p−1 + |g| |f + g|p−1 .
By Hölder’s inequality,
1/p 1/q
. |f ||f + g|p−1 dµ ≤ |f |p dµ |f + g|(p−1)q dµ .
X X X
Since .(p − 1)q = p, we conclude that
1/p 1/p 1/p

. |f + g|p dµ ≤ |f |p dµ + |g|p dµ .
X X X

Theorem 15.23 (Riesz-Fischer) The space .Lp (X) is complete.
Proof Let .{ϕn }n be a Cauchy sequence in .Lp . It is sufficient to prove that some
subsequence converges in .Lp , since the whole sequence will then converge to the
same limit. We first find indices .n1 < n2 < n3 < . . . of positive integers such that
* * 1
. * ϕn − ϕnk * < k , k = 1, 2, . . . .
k+1
2

Hence the series . ∞
k=1 |ϕnk+1 − ϕnk | converges almost everywhere. Indeed,
*N *
* *
N
* *
N
* ϕ n − ϕ n * * ϕn − ϕn * < 1
.* k * < <1
* k+1
* k+1 k p
2k
k=1 p k=1 k=1
for every .N > 1, and the claim follows

from Beppo Levi’s Theorem.
As a consequence, the series . ∞ k=1 ϕnk+1 − ϕnk converges almost everywhere
with partial sums

N

. ϕnk+1 − ϕnk = ϕnN+1 − ϕn1 .
k=1

This means that the sequence . ϕnk k has a limit .ϕ almost everywhere. For any fixed
k, the function .ϕnj −ϕnk approaches .ϕ −ϕnk almost everywhere as .j → +∞. Since
* *
. *ϕn − ϕn * < 1 , k = 1, 2, . . .
j k p
2k
it follows that .ϕ − ϕnk ∈ Lp , and then .ϕ ∈ Lp . Taking the limit as .j → +∞ we

also derive
* *
. *ϕ − ϕn * < 1 .
k
2k
Letting .k → +∞ we see that .ϕnk → ϕ in .Lp , and the proof is complete.

The previous proof hides a statement of fundamental importance in several aspects
of Functional Analysis. We record it as follows.
Theorem 15.24 (Partial Converse of the Dominated Convergence Theorem)
Let .p ≥ 1 and let .ϕn → ϕ in .Lp (X). There exists a subsequence .ψk = ϕnk and a
function .g ∈ Lp (X) such that .|ψk | ≤ g and .ψk → ϕ almost everywhere.
Proof Since .{ϕn }n converges in .Lp , it is a Cauchy sequence. The subsequence
.{ψk }k = {ϕnk }k of the previous proof converges almost everywhere to .ϕ, and for
every index k,
∞

|ψk | ≤ |ψ1 | +
. |ψk+1 − ψk | ∈ Lp (X).
k=1

Theorem 15.25 (Density of Elementary Functions) The collection of elementary
functions H is dense in .Lp (X).
Proof Let .h ∈ H . We first prove that .h ∈ Lp . Since elementary functions are

bounded, we assume that .|h(x)| ≤ M for every .x ∈ X, and we consider the set
.E = {x ∈ X | |h(x)| > 1}. Then E is measurable with a finite measure, and
Mp if x ∈ E
. |h(x)|p ≤
|h(x)| if x ∈
/ E.
In other words,
. |h(x)|p ≤ M p χE (x) + |h(x)| χX\E (x) ≤ M p χE (x) + |h(x)| .
It follows at once that .|h|p is integrable.

Let now .f ∈ Lp (X) be given. Since both .f + and .f − belong to .Lp , without loss
of generality we suppose that .f ≥ 0. Define the increasing sequence of measurable
sets

1
.En = x ∈ X
n < f (x) < n , n = 1, 2, . . .
and set
f (x) if x ∈ En
fn (x) =
.
0 otherwise.
Obviously .fn - f and .(f − fn )p . 0, and therefore Beppo Levi’s Theorem yields
fn → f in .Lp . Given .ε > 0 we can choose a positive integer n such that
.
ε
. *f − fn *p < .
2
p
Since .χEn ≤ χEn ≤ np f p , the function .fn is integrable by Hölder’s inequality:
* *
* q * *
. fn dµ = χEn f dµ ≤ *χEn * *f p p .
X X q
Assume that H is dense in .L1 . Then a sequence of elementary functions .hk exists
such that .hk → fn in .L1 as .k → +∞. Replacing .hk by .h+
k we may also assume
that .hk ≥ 0. Replacing .hk by
1
. min {hn , n} = n min hk , 1 ,
n
we may also assume that .|hk | ≤ n. Here we are using Stone’s axiom. We claim that
hk → fn in .Lp . Indeed,
.
p
. *fn − hk *p = |fn − hk |p dµ = |fn − hk |p−1 |fn − hk | dµ
X X
< np−1 |fn − hk | → 0

X
as .k → +∞, and we can choose k so large that

ε
. *fn − hk *p < .
2
We conclude that
. *f − hk *p < *fn − hk *p + *f − fn *p < ε.
To complete the proof, it remains to show that H is dense in .L1 (X) = L(X). But
every function in L is the difference of two functions in .L+ , and we need to prove
the claim for functions .f ∈ L+ which are limits (in the norm of .L1 ) of a sequence
of functions .hn ∈ H . The natural choice for this sequence is the sequence which
defines f . Then .hn - f , .I hn - If , and
. *f − hn *1 = I (f − hn ) = If − I hn → 0
as .n → +∞.

Theorem 15.26 (Generalized Hölder’s Inequality) Suppose .1 < pj < +∞,
uj ∈ Lpj (X) for .1 ≤ j ≤ k. If
.
1 1 1
. + + ··· + = 1,
p1 p2 pk
then .u1 u2 · · · uk ∈ L1 and
. u1 u2 · · · uk dµ ≤ *u1 *p1 · · · *uk *pk .

X
Proof Exercise.

Theorem 15.27 (Interpolation Inequality) Suppose that .1 ≤ p < q < r < +∞,
1 1−λ λ
. = +
q p r
and .u ∈ Lp (X) ∩ Lr (X). Then .u ∈ Lq (X) and
. *u*p ≤ *u*1−λ
p *u*λr .
Proof Exercise.

Theorem 15.28 Suppose that .p ≥ 1 and .{un }n is a sequence in .Lp (X) such that
(a) .*un *p → *u*p as .n → +∞,
(b) .un → u almost everywhere as .n → +∞.
Then .*un − u*p → 0 as .n → +∞.
Proof We have almost everywhere

0 ≤ 2p |un |p + |u|p − |un − u|p .
.
Fatou’s Lemma implies

2p+1
. |u|p dµ ≤ lim inf 2p |un |p + |u|p − |un − u|p dµ
X n→+∞ X
= 2p+1 |u|p dµ − lim sup |un − u|p dµ.

X n→+∞ X
!
We conclude that .lim supn→+∞ X |un − u|p dµ ≤ 0, and the proof is complete.

Theorem 15.29 (Brexis-Lieb) Suppose that .p ≥ 1 and .{un }n is a sequence in
Lp (X) such that
.
(a) .c = supn *un *p ∈ R,

(b) .un → u almost everywhere as .n → +∞.
Then .u ∈ Lp (X) and
p p p
. lim *un *p − *un − u*p = *u*p .
n→+∞
Proof The assumptions imply .|u*p ≤ c. Fix any .ε > 0. By homogeneity, there
exists a number .C(ε) > 0 such that for every a, b in .R,

. |a + b|p − |a|p − |b|p ≤ ε|a|p + C(ε)|b|p .
Again by Fatou’s Lemma,
. C(ε)|u|p dµ ≤ lim inf ε|un − u|p + C(ε)|u|p

X n→+∞ X

− |un |p − |un − u|p − |u|p dµ
15.12 The Space L∞ 397
≤ (2c)p ε + C(ε)|u|p dµ
X

− lim sup |un |p − |un − u|p − |u|p dµ,
n→+∞ X
which means

. lim sup |un |p − |un − u|p − |u|p dµ ≤ (2c)p ε.
n→+∞ X
Since .ε > 0 is arbitrary, the proof is complete.

15.12 The Space L∞
The formal case .p = ∞ of .Lp gives rise to a very different space of functions.
Definition 15.21 Suppose that .f : X → [0, +∞] be a measurable function. We
consider the set

−1
.S = α ∈ R µ(g ((α, +∞])) = 0 .
If .S = ∅, we put .β = +∞. If .S = ∅, we put .β = sup S. The number .β is called the

essential supremum of f .
∞ −1
Exercise
15.11+Prove that .β ∈ S. Hint: .g −1 ((β, +∞]) = n=1 g
β + n , +∞ .
1
Definition 15.22 A measurable function .f : X → R belongs to .L∞ (X) if and only

if the essential supremum of .|f | is a finite real number. In this case we write .*f *∞
to denote this essential supremum. In symbols,
L∞ (X) = {f | *f *∞ < +∞}

.
)
f is measurable and there exists

= f C ∈ R such that |f | ≤ C almost .

everywhere
The inequalities of Hölder and Minkowski can be extended to the .L∞ -case without
much pain. For example,
Theorem 15.30 If .f ∈ L1 (X) and .g ∈ L∞ (X), then .fg ∈ L1 (X) and .*fg*1 ≤
*f *1 *g*∞ .
Proof For almost every .x ∈ X we have .|f (x)g(x)| ≤ *f (x)**g*∞ . The
conclusion follows by integration of this inequality.

The completeness of .L∞ is a deeper result.

Theorem 15.31 The space .L∞ (X) is a Banach space.
Proof Suppose that .{fn }n is a Cauchy sequence in .L∞ (X), and set
.Ak = {x ∈ X | |fk (x)| > *fk *∞ }

Bm,n = {x ∈ X | |fm (x) − fn (x)| > *fm − fn *∞ }

E= Ak ∪ Bn,m k, m, n ∈ N .
As a countable union of sets of measure zero, we have .µ(E) = 0. On .X \ E we have

that .fn → f uniformly to a bounded limit f . If we set .f (x) = 0 for every .x ∈ E,
the function f belongs to .L∞ (X), and .*fn − f *∞ → 0 as .n → +∞. The proof is
complete.

15.13 Changing Variables in Multiple Integrals
Following [2], we quickly discuss an important technique of Advanced Calculus.

Definition 15.23 Let . and .ω be open subsets of .Rn . A diffeomorphism .f : →
ω is a continuously differentiable bijective function such that for every .x ∈ ,
Jf (x) = det Df (x) = 0.

.
Here .Df (x) denotes the Jacobian matrix of f at .x ∈ .

Theorem 15.32 (Change of Variables) Let .f : → ω be a diffeomorphism and
u ∈ Cc (ω). The .u ◦ f ∈ Cc (), and
.
. u(f (x))|Jf (x)| dx = u(y) dy.

ω
!
Notice that we are denoting by a symbol like . . . . dx the Lebesgue integral in .Rn .
Proof We proceed by induction on the dimension .n ≥ 1.
1. .n = 1. In this case we may assume that . = (a, b). By the Fundamental
Theorem of Calculus,
b f (b)
. u(f (x))f % (x) dx = u(y) dy.
a f (a)
If .f % > 0, then .ω = (f (a), f (b)). If .f % < 0, then .ω(f (b), f (a)). In both cases
the claim is proved.
15.13 Changing Variables in Multiple Integrals 399
2. The induction step. We assume the result has been proved in dimension .n−1. Let
.a ∈ be a fixed point. Since f is a diffeomorphism, .(∂1 f1 (a), . . . , ∂n fn (a)) =
0. Without loss of generality we may assume that .∂n fn (a) = 0. By the Implicit
Function Theorem, there exist .r > 0, an open set .U ⊂ such that .a ∈ U , an
open set .V ⊂ Rn−1 , and a function .β ∈ C 1 (V × (fn (a) − r, fn (a) + r)) such
that for .|t − fn (a)| < r there results

. {fn = t} ∩ U = (x % , β(x % , t)) x % ∈ V . (15.6)
We now split3
f = (f % , fn )
.
h(x % , xn ) = (x % , fn (x % , xn ))
t (x % ) = f % (x % , β(x % , t))
g(x % , t) = (t (x % ), t).
We add the assumption that .supp u ⊂ U . Fubini’s Theorem and the induction
hypothesis ensure that4
. u(g(x))|Jg (x)| dx = dt u(t (x % ), t)|Jt (x % )| dx %

V
= dt u(y % , t) dy %
= u(y) dy.
Now let .v = (u ◦ g)|Jg |. Fubini’s Theorem and the proof in dimension .n = 1

imply that
. v(h(x))|Jh (x)| dx = dx % v(x % , fn (x % , xn ))|∂n fn (x % , xn ))| dxn
= dx % v(x % , t) dt
= u(g(x))|Jg (x)| dx.
3 The prime .% does not mean differentiation. It is used to group the first .n − 1 components of a
vector in .Rn .
4 All the integrals in the next equations may be computed on .Rn , since the integrand functions have
compact support.
Recalling that .f = g ◦ h on U , we get .Df = Dg(h)Dh and .Jf = Jg (h)Jh .

Therefore
. u(f (x))|Jf (x)| dx = u(g(h(x)))|Jg (x)| |Jh (x)| dx
= v(h(x))|Jh (x)| dx
= u(y) dy.
To conclude the proof, we need to remove the condition that the support of u
be contained in U . By assumption the support of u is a compact subset of .Rn ,
and we can cover it by a finite collection of open sets .Uj which satisfy (15.6). If
.{ψj }j is a finite partition of unity subordinated to the covering .{Uj }j , we have

.u = j ψj u, and the general case follows from the linearity of the integral.

15.14 Comments
The construction of the abstract Lebesgue integral dates back to P.J. Daniell in 1918.
It has the typical elegance of formal axiomatizations, which isolate the essential
properties of the object we want to define. A short but complete reference is [1].
The clean approach of [2] is another interesting source.
To be fairly honest, many analysts are satisfied with the basic Lebesgue integral.
If this is the main purpose, the approach we present in the next chapter might
be preferable. It is nonetheless interesting that the theory of integration can be
completely developed in terms of a functional-analytic completion process: we
construct the Lebesgue integral from the Riemann integral in the same way as we
construct the real numbers from the rational ones.
References
1. G.E. Shilov, B.L. Gurevich, Integral, Measure and Derivative: A Unified Approach (Dover
Publications, New York, 1977)
2. M. Willem, Functional Analysis: Fundamentals and Applications (Birkhäuser/Springer, New
York, 2013)
Chapter 16
Measures Before Integrals
Abstract This chapter presents a straightforward approach to abstract measure the-

ory. We will see that an abstract Lebesgue integral can be defined after introducing
the idea of measuring sets. In the end we will connect these two approaches to
integration theory.
The popular approach to the Lebesgue integral is via abstract measure theory:
we first define a set function—called a measure—on a set of suitable sets—
called measurable sets, then we define measurable functions, and finally integrable
functions. The main pro of this approach is that at the end we have the highest
generality. A troublesome con is that such a construction requires a good amount of
mathematical education before it can be understood.
Our approach follows closely [3]. Historically, the Bourbaki group rejected
abstract measure theory for a long time, since they decided to focus only on Radon
measures defined as continuous linear functionals. There is some reason for doing
this, but the stiffness of Radon measures is an obstacle to several investigations
in Calculus of Variations, Geometric Measure Theory, Differential Geometry, and
so on.
16.1 General Measure Theory
Definition 16.1 Let X be a set. A family of subsets of X is a σ -algebra if and

only if
1. X ∈ ,
2. if A ∈ ,
then X \ A ∈ ,
3. if A = ∞ n=1 An and An ∈ for every n, then A ∈ .
The elements of a σ -algebra are called measurable sets, and X is a measure space.
402 16 Measures Before Integrals
Definition 16.2 Let X be a measure space and Y be a topological space. A function

f : X → Y is measurable if and only if f −1 (V ) is measurable whenever V is open
in Y .
Exercise 16.1 Let be a σ -algebra on a set X. Prove the following statements:
1. ∅ ∈ .
2. is closed under the formation of finite and countable intersections.
3. If A ∈ and B ∈ , then A \ B ∈ .
Theorem 16.1 Let Y and Z be topological spaces, and let g : Y → Z be a
continuous function. If X is a measure space, if f : X → Y is measurable, and
if h = g ◦ f , then h : X → Z is measurable.
Proof If V is open in Z, then g −1 (V ) is open in Y , and h−1 (V ) = f −1 (g −1 (V )).
Since f is measurable, it follows that h−1 (V ) is measurable, and the proof is
complete.

Theorem 16.2 Let u and v be measurable functions on a measure space X, and let
be a continuous function from R2 to a topological space Y . Define
h(x) = (u(x), v(x)),

. x ∈ X.
Then h : X → Y is measurable.
Proof Let us write f (x) = (u(x), v(x)) for x ∈ X. Since h = ◦ f , Theorem 16.1
shows that we only need to prove that f is measurable. To this aim, we suppose that
R = (a, b) × (c, d). It follows that
f −1 (R) = u−1 ((a, b)) ∩ v −1 ((c, d)),

.
which is a measurable set by assumption. Since any open set V ⊂ R2 is a

countable union of rectangles of the form (a, b) × (c, d), we deduce that f −1 (V ) is
measurable, and the proof is complete.

Corollary 16.1 Let X be a measure space. If f , g are real-valued measurable
functions, then f + g, fg, |f | are measurable.
Proof We just consider (s, t) = s + t, (s, t) = st in Theorem 16.2, or g(x) =
|x| in Theorem 16.1.

Theorem 16.3 Let X be a measure space. A function f : X → R is measurable if
and only if any of the following statements holds true:
1. the set {x ∈X | f (x) > a} is measurable for every a ∈ R;
2. the set {x ∈X | f (x) ≥ a} is measurable for every a ∈ R;
3. the set {x ∈X | f (x) < a} is measurable for every a ∈ R;
4. the set {x ∈X | f (x) ≤ a} is measurable for every a ∈ R.
16.1 General Measure Theory 403
Proof This follows immediately from the remark that the collection of all open
half-lines of the form (a, +∞) or (−∞, b) is a sub-basis of the standard topology
of R.

Theorem 16.4 If {fn }n is a sequence of real-valued measurable functions on X,
then infn fn , supn fn , lim infn→+∞ fn , lim supn→+∞ fn are measurable.
Proof For example,
∞

. x ∈ X sup fn (x) > a = {x ∈ X | fn (x) > a} ,
n
n=1
so that supn fn is measurable. Since − supn (−fn ) = infn fn , lim supn→+∞ fn =

infm supn≥m fn and lim infn→+∞ fn = supm infn≥m fn , the conclusion follows.
Exercise 16.2 If f and g are real-valued measurable functions, prove that
max{f, g} and min{f, g} are measurable.
Theorem 16.5 (Existence of σ -Algebras) If F is any collection of subsets of X,
then there exists a smallest σ -algebra ∗ in X such that F ⊂ ∗ . This is called the
σ -algebra generated by F.
Proof Since 2X is trivially a σ -algebra, the set of all σ -algebras containing F is
not empty. Define

∗ =
. { | ∈ } .
It is obvious that F ⊂ ∗ and that ∗ is contained in every σ -algebra which contains

F. The proof will be completed once we show that ∗ is a σ -algebra.
Suppose An ∈ ∗ for n ∈ N. For every ∈ , A−n ∈ , so that ∞ n=1 An ∈ .
But is an arbitrary σ -algebra containing F, and we conclude that ∞ ∗
n=1 An ∈ .
A similar argument shows that X ∈ ∗ and that A ∈ ∗ implies X \A ∈ ∗ . Hence
∗ is a σ -algebra, and the proof is complete.

If the set X is endowed with a topology, the previous result provides us with a σ -
algebra compatible with open subsets.
Definition 16.3 Suppose that X is a topological space. The Borel σ -algebra B in
X is the σ -algebra generated by the topology of X, i.e. by the collection of open
subsets of X. Every member of B is a Borel set.
Definition 16.4 Suppose that X is a topological space. If a subset of X can be
expressed as a countable union of closed sets, it is called a Fσ set. If a subset of X
can be expressed as a countable intersection of open sets, it is called a Gδ set.
Exercise 16.3 Prove that any Fσ and any Gδ set are Borel sets.
Definition 16.5 Let (X, ) be a measure space. A function µ : → [0, +∞] is a
measure if and only if
1. µ(∅) = 0,
2. if {An | n ∈ N} is a sequence of measurable sets such that Ai ∩Aj = ∅ for i = j ,
then
"∞ # ∞

.µ An = µ(An ).
n=1 n=1
The space X is of finite measure if and only if µ(X) < +∞. The space X is σ -
finite
if and only if it is the countable union of measurable subsets of finite measure:
X= ∞ n=1 Xn with µ(Xn ) < +∞ for every n.
Example 16.1 (Counting Measure) For every subset A of a set X, we define
µ(A) = +∞ if A contains infinitely many elements, and µ(A) = #A if A contains
finitely many elements.
Example 16.2 (Dirac Measure) Let X be a set, and let x0 ∈ X be a fixed point. If
A ⊂ X contains x0 , we set µ(A) = 1; otherwise we set µ(A) = 0. This is the Dirac
measure concentrated at x0 .
Definition 16.6 A subset E of a measure space (X, , µ) has measure zero if and
only if E ∈ and µ(E) = 0. A property holds almost everywhere (a.e. for short)
if and only if it holds in X except for a subset of measure zero.
Theorem 16.6 Let (X, , µ) be a measure space.
(a) If {An }n is a sequence of measurable subsets such that An ⊂ An+1 for every n,
then
"∞ #

.µ An = lim µ(An ).
n→+∞
n=1
(b) If {An }n is a sequence of measurable subsets such that An ⊃ An+1 for every n,
and if µ(A1 ) < +∞, then
"∞ #

µ
. An = lim µ(An ).
n→+∞
n=1
Proof We may assume that µ(An ) < +∞ for every n, otherwise the conclusion is
trivial. From the identity
∞

. An = A1 ∪ (A2 \ A1 ) ∪ (A3 \ A2 ) ∪ · · ·
n=1
we derive
"∞ #

.µ An = µ(A1 ) + µ(A2 \ A1 ) + µ(A3 \ A2 ) + · · ·
n=1
= µ(A1 ) + µ(A2 ) − µ(A1 ) + µ(A3 ) − µ(A2 ) + · · ·
= lim µ(An ).
n→+∞
This proves (a). To prove (b) we set Bn = A1 \ An , so that Bn ⊂ Bn+1 for every n.
Furthermore µ(Bn ) = µ(A1 ) − µ(An ), and
∞
∞

. Bn = A1 \ An .
n=1 n=1
Hence (a) implies

" ∞
# " ∞
#

µ(A1 ) − µ
. An = µ A1 \ An
n=1 n=1
"∞ #

=µ Bn = lim µ(Bn )
n→+∞
n=1
= µ(A1) − lim µ(An ).
n→+∞
Since µ(A1 ) ∈ R, the conclusion follows.

Definition 16.7 A function s : X → R defined on a set X is a simple function if
and only if the image s(X) is a finite set.
Exercise 16.4 Prove that any simple function s is a linear combination of charac-
teristic functions. Hint: assume that s(X) =
{c1 , c2 , . . . , cn }. For 1 ≤ i ≤ n, set
Ei = {x ∈ X | s(x) = ci }. Conclude that s = ni=1 ci χEi .
Theorem 16.7 Let X be a measure space, and let s be a simple function whose
range is {c1 , . . . , cn }. The function s is measurable if and only if the sets Ei = {x ∈
X | s(x) = ci } are measurable for 1 ≤ i ≤ n.
Proof This follows immediately from Theorem 16.3.

Simple function are the bricks with which we build integration theory. Let us start
with an approximation result.
Theorem 16.8 If f : X → R is a function, then there exists a sequence {sn }n of
simple functions such that sn (x) → f (x) for every x ∈ X. If f is measurable, then
each sn can be taken measurable. If f ≥ 0, then we may suppose that sn ≤ sn+1 for
every n.
Proof We begin with the case f ≥ 0. For any n ∈ N and i ∈ {1, 2, . . . , n · 2n } we
define

i−1 i
.Eni = x ∈ X n ≤ f (x) ≤ n
2 2
Fn = {x ∈ X | f (x) ≥ n}
n

n·2
i−1
sn (x) = χEni (x) + nχFn (x).
2n
i=1
It is easy to check that sn → f pointwise in X, and that sn ≤ sn+1 for every n. In

the general case, we split f = f + − f − and repeat the same construction for f +
and f − .

Exercise 16.5 If f is a bounded function, prove that the sequence {sn }n constructed
above converges uniformly to f .
We can now introduce the (abstract) integral in a measure space (X, , µ).
Definition 16.8 (Integral of Simple Functions) Let E ⊂ X be a measurable set.
The integral on E of a simple measurable function s such that

n
s=
. ci χEi ,
i=1
ci ≥ 0 for every i, and ci = cj for i = j , is the number

n
IE (s) =
. ci µ(E ∩ Ei ).
i=1
We agree that 0 · ∞ = 0. At this stage this is actually harmless: we have in

mind the situation in which a set E ∩ Ei has infinite measure, while ci = 0.
Definition 16.9 (Integral of Measurable Functions) Let E ⊂ X be a measurable

set. The integral on E of a measurable function f : X → [0, +∞) is defined as
. f dµ = sup {IE (s) | s is a simple function s.t. 0 ≤ s ≤ f in E} .

E
Definition 16.10 Let f : X → R be a measurable function. We say that f is

integrable on E if and only if
. f + dµ < +∞, f − dµ < +∞.

E E
In this case we write f ∈ L1 (E) = L1 (E, µ), and
. f dµ = f + dµ − f − dµ.
E E E
We collect several properties of the abstract integral. The easy proofs can be
provided by the reader as an exercise.
Proposition 16.1 Let X be a measurable space, and let E be a measurable subset
of X.
1. If f is measurable and bounded on E, and if µ(E) <!+∞, then f ∈ L1 (X).
2. If µ(E) < +∞ and a ≤ f ≤ b on E, then aµ(E) ≤ E f dµ ≤ bµ(E).
! !
3. If f ∈ L1 (E) and c ∈ R, then cf ∈ L1 (E), and E (cf ) dµ = c E f dµ.
! !
4. If f ∈ L1 (E), g ∈ L1 (E) and f ≤ g on E, then E f dµ ≤ E g dµ.
5. If f ∈ L1 (E), A is measurable and A ⊂! E, then f ∈ L1 (A).
6. If µ(E) = 0 and f is measurable, then E f dµ = 0.
Any measurable function induces a new measure by means of the integral. Here
is the precise statement.
Theorem 16.9 Suppose that f is a measurable function on a measurable space
(X, , µ) such that f ≥ 0 on X. Then the function
ν : A ∈ →
. f dµ
A
is a measure on X.
Proof Consider any sequence {An }n of measurable sets such that Ai ∩ Aj = ∅ for
i = j , and A = ∞n=1 An . We need to prove that
∞

ν(A) =
. ν(An ).
n=1
We suppose that f = χE , where E is a measurable set. In this case the conclusion

follows from the identity
. χE dµ = µ(B ∩ E)
B
for every measurable B and from countable additivity of µ. It then follows that the
same conclusion holds for every simple function f .
In the general case, let s be a simple measurable function such that 0 ≤ s ≤ f
on X. We have
∞
∞

. s dµ = s dµ ≤ ν(An ).
A n=1 An n=1
Hence
∞

. f dµ = ν(A) ≤ ν(An ).
A n=1
To prove the reversed inequality, we may clearly restrict to the case ν(An ) ∈ R for
every n. Pick ε > 0 and two measurable simple functions s1 , s2 such that
ε
. s1 dµ > f dµ −
A1 A1 2
ε
s2 dµ > f dµ − .
A2 A2 2
Then

ν(A1 ∪ A2 ) ≥
. s1 χA1 + s2 χA2 dµ
A1 ∪A2
= s1 dµ + s2 dµ ≥ ν(A1 ) + ν(A2 ) − ε.
A1 A2
Since this holds for any ε > 0, we see that ν(A1 ∪ A2 ) ≥ ν(A1 ) + ν(A2 ). By
induction

n
ν(A1 ∪ · · · ∪ An ) ≥
. ν(Ak ) n ≥ 1,
k=1
n
and finally ν(A) ≥ ∞ n=1 ν(An ), since A ⊃ k=1 Ak . Since the other properties of
a measure are trivially satisfied, the proof is complete.

16.2 Convergence Theorems 409
Exercise 16.6 Suppose that !A and B are! measurable sets such that µ(A \ B) = 0.
If f is integrable, prove that A f dµ = B f dµ.
Theorem 16.10 Let f be a measurable function on X.
! !
(a) If f ∈ L1 (X), then |f | ∈ L1 (X), and X f dµ ≤ X |f | dµ.
(b) If g ∈ L1 (X) and |f | ≤ g on X, then f ∈ L1 (X).
Proof Let A = {x ∈ X | f (x) ≥ 0} and B = {x ∈ X | f (x) < 0}. Clearly X is the

disjoint union of A and B, and therefore
. |f | dµ = |f | dµ + |f | dµ
X A B
= f + dµ + f − dµ,
A B
and the last two integrals are finite. Hence |f | ∈ L1 (X). Recalling that f ≤ |f | and
−f ≤ |f |, we see that
. f dµ ≤ |f | dµ, − f dµ ≤ |f | dµ.
X X X X
The proof of (a) is thus complete. To prove (b) we remark that f + ≤ g, f − ≤ g, so

that f = f + − f − is the difference of two integrable functions.

16.2 Convergence Theorems
We consider a fixed measurable space .(X, , µ).

Theorem 16.11 (Beppo Levi) Suppose .{fn }n is a sequence of measurable func-
tions such that .fn ≥ 0 and .fn ≤ fn+1 for every n. Let .E ∈ and .f (x) =
limn→+∞ fn (x) for every .x ∈ E. Then
. f dµ = lim fn dµ.
E n→+∞ E
!
Proof The sequence! .n → E fn dµ is increasing, so there exists .α ∈ [0, +∞] such
that .α = limn→+∞ E fn dµ. Furthermore
α≤
. f dµ.
E
Let .0 < c < 1 and s be a measurable simple function such that .0 ≤ s ≤ f on E.

We define
En = {x ∈ E | fn (x) ≥ cs(x)} ,
. n ≥ 1.
.En is a measurable set, and .E1 ⊂ E2 ⊂ E3 ⊂ · · · Since .fn → f

Of course each
on E, .E = ∞ n=1 En . On the other hand, for each .n ≥ 1 we have
. fm dµ ≥ fn dµ ≥ c s dµ.
E En En
!
By Theorem 16.9 we may let .n → +∞ and conclude! that .α ≥ c E s dµ. Since this
holds for every .0 < c < 1, we also have .α ≥ E s dµ. Now the conclusion follows
from the arbitrariness of the simple function s.

Remark 16.1 The integrals appearing in Beppo Levi’s Theorem may well be
infinite. It should be remembered as a statement about limit of measurable functions,
without any reference to integrability properties.
Theorem 16.12 (Fatou’s Lemma) Suppose .fn ≥ 0 is a measurable function on a
set .E ∈ for every n, and let .f (x) = lim infn→+∞ fn (x) for every .x ∈ E. Then
. f dµ ≤ lim inf fn dµ.

E n→+∞ E
Proof For every .x ∈ E we set .gn (x) = infm≥n supn∈N fn (x). Clearly
. 0 ≤ g1 ≤ g2 ≤ g3 ≤ . . .
gn ≤ fn
lim gn (x) = f (x)
n→+∞
for every .x ∈ E. Beppo Levi’s Theorem yields
. f dµ = lim gn dµ ≤ lim inf f dµ.

E n→+∞ E n→+∞ E

Beppo Levi’s Theorem is a fundamental result for proving basic statements about
integrable functions. We just provide an example.
Proposition 16.2 Suppose that .f1 and .f2 are integrable functions on a measurable
set E. If .f = f1 + f2 , then .f ∈ L1 (E) and
. f dµ = f1 dµ + f2 dµ.
E E E
16.2 Convergence Theorems 411
Proof Suppose initially that .f1 ≥ 0 and .f2 ≥ 0. Let .{sn1 }n and .{sn2 }n be sequences
of measurable simple functions such that .sn1 → f1 , .sn2 → f2 on E. If we define
.sn = sn + sn , then
1 2
. sn dµ = sn1 dµ + sn2 dµ.

E E E
! ! !
Beppo Levi’s Theorem yields now that . E f dµ = E f1 dµ + E f2 dµ. Suppose
now that .f1 ≥ and .f2 ≤ 0. We construct the sets
A = {x ∈ E | f (x) ≥ 0} ,
. B = {x ∈ E | f (x) < 0} .
Since f , .f1 and .−f2 are non-negative on A, we have
. f1 dµ = f dµ + (−f2 ) dµ = f dµ − f2 dµ.
A A A A A
Since .−f , .f1 and .−f2 are non-negative on B, we see that
. (−f2 ) dµ = f1 dµ + (−f ) dµ.

B B B
But this means that
. f1 dµ = f dµ − f2 dµ,
B B B
and the conclusion follows in this case. The other cases are treated similarly.

Theorem 16.13 (Dominated Convergence Theorem) Let E be a measurable set,
and suppose that .{fn }n is a sequence of measurable functions such that
f (x) = lim fn (x) for every x ∈ E.

.
n→+∞
If there exists a function .g ∈ L1 (E) such that
.|fn (x)| ≤ g(x) for every x ∈ E and every n ∈ N,
then .f ∈ L1 (E) and
. f dµ = lim fn dµ.
E n→+∞ E
Proof By Theorem 16.10 we have .fn ∈ L1 (E) and .f ∈ L1 (E). Furthermore
fn + g ≥ 0 on E,
.
so that Fatou’s Lemma yields
. (f + g) dµ ≤ lim inf (fn + gn ) dµ,

E n→+∞ E
or
. f dµ ≤ lim inf fn dµ.

E n→+∞ E
Similarly, .g − fn ≥ 0, and therefore
. f dµ ≤ lim inf − fn dµ .
E n→+∞ E
This is equivalent to
. f dµ ≥ lim sup fn dµ,

E n→+∞ E

16.3 Complete Measures
We have already remarked that sets of measure zero should be completely invisible
in integration theory. Formally, we may agree that two measurable functions f and
g are equivalent if and only if the set
. {x ∈ X | f (x) = g(x)}
has measure zero in X. Then integrable equivalent functions have the same integral.
Furthermore, it would be quite natural to relax the assumptions of Beppo Levi’s
Theorem, or of the Dominated Convergence Theorem, to allow pointwise limits
almost everywhere. Unfortunately this may be troublesome, since a subset of a set
of measure zero need not be measurable. We thus introduce a reasonable definition.
Definition 16.11 A measurable space is complete if and only if the following
condition is satisfied: if E is a measurable set of measure zero and if .F ⊂ E,
then F is a measurable set (and of course the measure of F is then zero).
16.3 Complete Measures 413
Theorem 16.14 Suppose that .(X, , µ) is a measurable space. We define . ∗ as

the set of all .E ⊂ X such that there exist measurable sets A and B such that
∗
.A ⊂ E ⊂ B and .µ(B \ A) = 0. For every .E ∈ we define .µ(E) = µ(A). Then
∗ ∗
. is a .σ -algebra and .µ is a measure on . .
Proof First of all we check that .µ is well-defined on . ∗ . Assume that .A ⊂ E ⊂ B

and .A1 ⊂ E ⊂ B1 , and .µ(B1 \ A1 ) = µ(B \ A) = 0. From the inclusions .A \ A1 ⊂
E \ A1 ⊂ B1 \ A1 we see that .µ(A \ A1 ) = 0, hence .µ(A) = µ(A ∩ A1 ). For the
very same reason, .µ(A1 ) = µ(A ∩ A1 ), and we conclude that .µ(A) = µ(A1 ).
Now we have to prove that . ∗ is a .σ -algebra. Clearly enough, .X ∈ ∗ since
∗
.X ∈ and . ⊂ . Furthermore, if .A ⊂ E ⊂ B, then .X \ B ⊂ X \ E ⊂ X \ A.
Hence .E ∈ implies .X\E ∈ ∗ , because .(X\A)\(X\B) = (X\A)∩B= B \A.

∗
∞
∞ if the sets.A∞n ⊂ En ⊂ Bn for every .n ∈ N, and if .E = n=1 En ,
Finally,
.A = n=1 An , .B = n=1 , then .A ⊂ E ⊂ B and
∞
∞

B \A=
. (Bn \ A) ⊂ (Bn \ An ).
n=1 n=1
Since countable unions of sets of measure zero are sets of measure zero, it follows
that .E ∈ ∗ is each .En ∈ ∗ .
If, in particular, the sets .En are disjoint, then the sets .An are disjoint, so that
∞
∞

µ(E) = µ(A) =
. µ(An ) = µ(En ),
n=1 n=1
and .µ is a measure on . ∗ . This completes the proof.

Remark 16.2 The previous theorem shows a remarkable property of measurable
spaces. If a given measure may fail to be complete, it is however true that this
measure can be suitably completed.
With Theorem 16.14 in mind, we can propose a relaxed definition of measurable
functions which takes into account sets of measure zero.
Definition 16.12 A function f defined on a set .E ∈ is measurable if and only if
µ(X \ E) = 0 and .f −1 (V ) ∩ E is measurable for every open set V .
.
We may then define .f = 0 on .X \ E, which is a measurable function according to

our previous definition. And if the measure is complete, we may even define f on
.X \ E in an arbitrary manner, and still we get a measurable function.
In this way our statements can take into accounts function defined almost
everywhere, without altering the conclusions. Let us prove a useful example.
Theorem 16.15 Suppose that .{fn }n is a sequence of measurable functions defined

almost everywhere on X. If
∞

. |fn | dµ < +∞,
n=1 X
∞
then the series .f (x) = n=1 fn (x) converges for almost every .x ∈ X, .f ∈ L1 (X),
and
∞

. f dµ = fn dµ
X n=1 X
Proof Each function .µ(X \ Sn ) = 0.

.fn is defined on a subset
∞.Sn of X such that
For every .x ∈ S = ∞ n=1 Sn we set .ϕ(x) = n=1 |fn (x)|. Then .µ(X \ S) = 0. By
1
Beppo Levi’s Theorem, .ϕ ∈ L1 (S). If .E = {x ∈ S | ϕ(x) < +∞}, then .µ(X\E) =

0.
The series . ∞ n=1 fn (x) converges absolutely for every .x ∈ E, and .|f (x)| ≤ ϕ(x)
on E. Hence .f ∈ L1 (E). Introducing the partial sum .gn = f1 + · · · + fn , we see
that .|gn | ≤ ϕ on E, that .gn → f on E, and the Dominated Convergence Theorem
yields
∞

. f dµ = fn dµ.
E n=1 E
But .µ(X \ E) = 0, and the conclusion follows.

!
Theorem 16.16 If .f : X → [0, +∞] is measurable, .E ∈ and . E f dµ = 0,
then .f = 0 almost everywhere on E.
Proof For every positive integer n, We define .An = {x ∈ E | f (x) > 1/n}. The set
An is measurable, and
.
1
. µ(An ) ≤ f dµ ≤ f dµ = 0,
n An E

which implies .µ(An ) = 0 for every n. Since .{x ∈ E | f (x) > 0} = ∞
n=1 An , it
follows that .f = 0 almost everywhere.

!
Theorem 16.17 If .f ∈ L (X) and . E f dµ = 0 for every measurable .E ⊂ X,
1
then .f = 0 almost everywhere on X.
1With our original definition of measurable functions, this implication would require the completes
of the measurable space X.
16.4 Different Types of Convergence 415
!
Proof We apply the assumption to .E = {x ∈ X | f (x) ≥ 0}, so that . E f + dµ = 0.
Hence .f + = 0 almost everywhere on X. Similarly .f − = 0 almost everywhere on
X, and the proof is complete.

16.4 Different Types of Convergence
For a sequence of functions, pointwise convergence is often too less, while uniform
convergence is usually too much in applications. Measure Theory provides some
intermediate types of convergence which are of great convenience in mathematical
analysis. In this section we want to review some of them. As a rule, all functions
will be defined on a measurable space .(X, , µ).
Definition 16.13 A sequence of measurable functions .{fn }n converges almost
uniformly to a measurable limit f if and only if for every .ε > 0 there exists a
measurable set E such that .µ(X \ E) < ε and .fn → f uniformly on E.
Exercise 16.7 Prove that almost uniform convergence implies almost everywhere
convergence.
For a converse of the previous exercise, an additional assumption is needed.
Theorem 16.18 (Severini-Egorov) Suppose that .µ(X) < +∞. If a sequence of
measurable functions .{fn }n converges almost everywhere to f in X, then .fn → f
almost uniformly.
Proof Removing a set of measure zero, we may assume that .fn → f pointwise in
the whole X. For every .k ∈ N and .m ∈ N we define the set

1
E(m, k) =
. x ∈ X |fn (x) − f (x)| < .
n>m
k

We have .E(1, k) ⊂ E(2, k) ⊂ E(3, k) ⊂ · · · , and . m∈N E(m, k) = X. As a
consequence, to each .k ∈ N there corresponds an integer .mk ∈ N such that
" #
1 ε
µ
. x ∈ X |fn (x) − f (x)| < < k.
n>mk
k 2
If we set

1
E=
. x ∈ X |fn (x) − f (x)| < ,
k
k∈N n>mk
∞
it follows that .µ(X \ E) ≤ k=1 ε · 2−k = ε and .fn → f uniformly on E. The
proof is complete.

Exercise 16.8 When did we use the assumption .µ(X) < +∞ in the previous
proof?
Definition 16.14 A sequence .{fn }n of measurable functions converges in measure
to the limit f if and only if for every .ε > 0 there results
. lim µ ({x ∈ X | |fn (x) − f (x)| > ε}) = 0.

n→+∞
Example 16.3 Let .X = [0, +∞) with the concrete Lebesgue measure, and define
fn on X such that
.
0 if 0 ≤ x < n
fn (x) =
.
1 if n ≤ x < +∞.
Then .fn → 0 pointwise in X, but it does not converge to zero in measure. Indeed,
for .0 < ε < 1, .fn (x) > ε if and only if .x > n.
Theorem 16.19 If .fn → f in measure, then there exists a subsequence .{fnk }k
which converges to f almost everywhere. If, in addition, .µ(X) < +∞, then
convergence almost everywhere implies convergence in measure.
Proof Pick a sequence .n1 < n2 < n3 < . . . of positive integers such that the
measure of the set

1
.Ek = x ∈ X fnk (x) − f (x) >
k
is smaller than .1/k 2 . Thus

" ∞
# ∞
1
µ
. Ek ≤
k2
k=m k=n
and
∞
∞
E=
. Ek
m=1 k=m
has measure zero. Since

∞
∞
1
X\E =
. x ∈ X fnk (x) − f (x) ≤ ,
k
m=1 k=m
we see that .f (x) = limk→+∞ fnk (x) for every .x ∈ X \ E, and the first assertion is
proved.
16.4 Different Types of Convergence 417
To complete the proof, we assume that X has finite measure and we use Theorem
16.18. For each .η > 0 there exists a measurable set E such that .µ(X \ E) < η and
.fn → f uniformly on E. Then
. {x ∈ X | |fn (x) − f (x)| > ε}

= {x ∈ E | |fn (x) − f (x)| > ε} ∪ {x ∈ X \ E | |fn (x) − f (x)| > ε}
is a measurable set whose measure if smaller than .η provided that n is sufficiently

large. Since .η > 0 was arbitrary, the proof is complete.

In many applications a direct proof of .L1 -convergence remains out-of-reach. We
now introduce a fundamental result due to Vitali which may be useful in such
situations.
Definition 16.15 A sequence .{fn }n in .L1 (X) is called equi-integrable if and only
if the following condition is satisfied: for every .ε > 0 there exist a number .δ > 0
and a measurable set A such that .µ(A) < δ and
!
1. for every .n ∈ N there results . X\A |fn | dµ < ε;
!
2. for every measurable set E such that .µ(E) < δ there results . E |fn | dµ < ε for
every n.
The first condition is clearly empty when the whole space has finite measure.
We first establish a general result about integrable functions.
Proposition 16.3 (Absolute Continuity of the Integral) Let .f ∈ L1 (X). For
every .ε !> 0 there exists .δ > 0 such that for any measurable set E, .µ(E) < δ
implies . E |f | dµ < ε. Furthermore, for every .ε > ! 0 there exists a measurable
subset .X0 of X such that .X0 has finite measure and . X\X0 |f | dµ < ε.
Proof We may assume .f ≥ 0 without loss of generality, so that .|f!| = f . In the
general case we split .f = f + − f − . Fix any .ε > 0. By definition of . X f! dµ, there
!exists a measurable simple function s such that .0 ≤ s ≤ f and .0 ≤ X f dµ −
X s dµ < ε/2. Since s is a bounded function, there exists .M > 0 such that .0 ≤
s ≤ M on X. Therefore, for every measurable .E ⊂ X of finite measure, we have
ε ε
. f dµ = s dµ + (f − s) dµ ≤ s dµ + ≤ Mµ(E) + .
E E E E 2 2
The conclusion follows with .δ = ε/(2M).

To prove the second statement, we remark that s is integrable, so that .X0 = {x ∈
X | s(x) > 0} has finite measure. Moreover
. f dµ = (f − s) dµ ≤ (f − s) dµ < ε.
X\X0 X\X0 X

Theorem 16.20 (Vitali) Let .{fn }n be a sequence of equi-integrable functions. If

fn → f almost everywhere and if .f ∈ L1 (X), then .fn → f in .L1 (X).
.
Proof Trivially, .|fn − f | ≤ |fn | + |f |. If .X0 and .X1 are measurable sets with
.X1 ⊂ X0 , then

.
(fn − f ) dµ ≤ |fn −f | dµ+ (|fn |+|f |) dµ+ (|fn |+|f |) dµ.
X X1 X0 \X1 X\X0
Fix .ε > 0. By Proposition 16.3 and the first condition of equi-integrability, there
exists a measurable set .X0 ⊂ X of finite measure such that
ε
. (|fn | + |f |) dµ = |fn | dµ + |f | dµ < .
X\X1 X\X0 X\X0 3
By Proposition 16.3 again, we may find a number .δ > 0 such that for every
measurable .E ⊂ X, .µ(E) < δ implies
ε
. (|fn | + |f |) dµ = |fn | dµ + |f | dµ < .
E E E 3
By assumption .f ∈ L1 (X). In particular .f < +∞ almost everywhere, and

.µ(X0 ) < +∞. It follows from Theorem 16.18 that there exists a measurable subset
.X1 of .X0 such that .µ(X0 \ X1 ) < δ and .fn → f uniformly on .X1 . Hence
ε
. (|fn | + |f |) dµ <
X0 \X1 3
and there exists .N ∈ N such that .n ≥ N implies
ε
. |fn − f | dµ ≤ sup {|fn (x) − f (x)| | x ∈ X1 } µ(X1 ) < .
X1 3
Collecting these estimates we conclude that .n ≥ N implies

ε ε ε
(fn − f ) dµ < + + = ε.
.
3 3 3
X

Exercise 16.9 Suppose that for every!.n ∈ N, .hn ∈ .hn ≥ 0 and .hn → 0
L1 (X),
almost everywhere on X. Prove that . X hn dµ → 0 if and only if .{hn }n is equi-
integrable. What can you deduce in the particular case .hn = |fn − f |?
16.5 Measure Theory on Product Spaces 419
16.5 Measure Theory on Product Spaces
Let be a σ -algebra on a set X and T be a σ -algebra on a set Y .a A

measurable rectangle is a set A × B such that A ∈ and B ∈ T . The
class E of elementary sets consists of all sets Q = R1 ∪ · · · ∪ Rn where each
Rj is a measurable rectangle and Ri ∩ Rj = ∅ when i = j .
a
Here T should be seen as the capital τ of the Greek alphabet.
Definition 16.16 The product σ -algebra × T is the smallest σ -algebra which

contains all measurable rectangles.
A monotone class M is a collection of sets with the following properties: if

Ai ∈ M, Bi ∈ M, Ai ⊂ Ai+1 , Bi ⊃ Bi+1 for every i = 1, 2, 3, . . . and if
∞
∞

A=
. Ai , B= Bi ,
i=1 i=1
then A ∈ M and B ∈ M. To summarize, monotone classes are closed with

respect to countable increasing unions and countable decreasing intersections.
Theorem 16.21 × T is the smallest monotone class which contains all elemen-
tary sets.
Proof Let M be the intersection of all monotone classes which contain E. It is clear
that M = ∅, since X × Y is a monotone class which contains E. But × T is a
monotone classes, so that M ⊂ × T . For A1 ∈ , A2 ∈ , B1 ∈ T , B2 ∈ T we
have
(A1 × B1 ) ∩ (A2 × B2 ) = (A1 ∩ A2 ) × (B1 ∩ B2 )

.
(A1 × B1 ) \ (A2 × B2 ) = ((A1 \ A2 ) × B1 ) ∪ ((A1 ∩ A2 ) × (B1 \ B2 )) .
It follows that P ∈ E, Q ∈ E imply P ∩ Q ∈ E and P \ Q ∈ E. Furthermore

P ∪ Q = (P \ Q) ∪ Q and (P \ Q) ∩ Q = ∅, hence P ∪ Q ∈ E.
For every P ⊂ X × Y , call (P ) the family of all Q ⊂ X × Y such that
P \ Q ∈ M, Q \ P ∈ M, and P ∪ Q ∈ M. It is evident that Q ∈ (P ) if and only
if P ∈ (Q), and that each (P ) is a monotone class.
Now select P ∈ E. As we have just seen, Q ∈ E implies Q ∈ (P ), so that
E ⊂ (P ) and finally M ⊂ (P ).
Select Q ∈ M. Since P ∈ E implies Q ∈ (P ), we also have P ∈ (Q).

Hence E ⊂ (Q) and we conclude again that M ⊂ (Q).
We have proved that P \ Q ∈ M and P ∪ Q ∈ M provided that P ∈ M and
Q ∈ M. It is easy to deduce that M is a σ -algebra on X × Y . Indeed X × Y ∈ M
because X × Y ∈ E.
If Q ∈ M, then (X × Y ) \ Q ∈ M since the difference of any two elements of
M is an element of M.
Finally, if Pi ∈ M for i = 1, 2, 3, . . . and P = ∞ i=1 Pi , we set Qn = P1 ∪
· · · ∪ Pn and remark that M is closedunder the formation of finite unions. Hence
Qn ∈ M. But Qn ⊂ Qn+1 and P = ∞ n=1 Qn , and the monotonicity of M shows
that P ∈ M.
We have proved that M is a σ -algebra such that E ⊂ M ⊂ × T . Hence
M = × T , and the proof is complete.

The introduction of a measure on X × Y is associated to the sections of a
measurable subset.
Definition 16.17 If E ⊂ X × Y , x ∈ X and y ∈ Y , we set
Ex = {y ∈ Y | (x, y) ∈ E}
.
E y = {x ∈ X | (x, y) ∈ E} .
Proposition 16.4 If E ∈ × T , then Ex ∈ T and E y ∈ for every x ∈ X, y ∈ Y .

Proof We temporarily say that E ∈ if and only if Ex ∈ T for every x ∈ X. If
E = A × B, then
B if x ∈ A
Ex =
.
∅ if x ∈
/ A.
Hence every measurable rectangle belongs to . Observe now that X × Y ∈ , that

((X × Y ) \ E)x = (X × Y ) \ Ex , and that
"∞ # ∞

. Ei = (Ei )x .
i=1 x i=1
Hence is a σ -algebra which must coincide with × T . A completely analogous

proof can be repeated for E y .

We now deal with functions of two variables in a similar manner.
Definition 16.18 Let f be a function defined on X × Y . For every x ∈ X and every
y ∈ Y we denote
fx = f (x, ·),
. f y = f (·, y).
16.5 Measure Theory on Product Spaces 421
Proposition 16.5 Let f be measurable in × T . Then fx is measurable in T and

f y is measurable in for every x ∈ X, y ∈ Y .
Proof Consider an open set V (contained in the unnamed codomain of f ), and
set Q = {(x, y) | f (x, y) ∈ V } = f −1 (V ). By assumption Q ∈ × T , and
Qx = {y ∈ Y | fx (y) ∈ V }. We have already proved that Qx ∈ T , and this shows
that fx is measurable in T . The same reasoning can be applied to f y , and the proof
is complete.

We can define the product measure on X × Y in terms of integrals.
Definition 16.19 Let (X, , µ) and (Y, T , λ) be σ -finite measurable spaces. For
every Q ∈ × T we define
.(µ × λ)(Q) = λ(Qx ) dµ(x) = (Qy ) dλ(y).

X Y
This is the product measure of µ and λ.

Of
! course the previous
! definition must be confirmed by proving that indeed
X λ(Qx ) dµ(x) = Y (Qy ) dλ(y). This is the content of the next result.
Theorem 16.22 Let (X, , µ) and (Y, T , λ) be σ -finite measurable spaces. Sup-
pose that Q ∈ × T . If
ϕ(x) = λ(Qx ),
. ψ(y) = µ(Qy ) (16.1)
for every x ∈ X, y ∈ Y , then ϕ is measurable in T , ψ is measurable in , and
. λ(Qx ) dµ(x) = (Qy ) dλ(y).

X Y
Proof We call the collection of all subsets Q of X × Y for which the conclusion
of the theorem holds.
(a) Every measurable rectangle belongs to Q. Indeed, if Q = A×B is a measurable
rectangle, then λ(Qx ) = λ(B)χA (x), µ(Qy ) = µ(A)χB (y), and both integrals
are equal to µ(A)λ(B).
(b) If Q1 ⊂ Q2 ⊂ Q3 ⊂ . . ., if Qi ∈ for every i and if Q = ∞ i=1 Qi , then
Q ∈ . Indeed, let ϕi and ψi be defined as in (16.1) with Qi instead of Q. It
follows from the countable additivity of each measure that ϕi - ϕ, ψi - ψ
pointwise as i → +∞. Beppo Levi’s Theorem yields the conclusion.
(c) If
∞ {Qi }i is a disjoint countable collection of elements of , and if Q =
i=1 Qi , then Q ∈ . Indeed, this is trivial for a finite collection, since
the characteristic function of a union of disjoint sets is just the sum of their
characteristic functions. In the general case we may use (b).
(d) If
∞ µ(A) < +∞, λ(B) < +∞, if A × B ⊃ Q1 ⊃ Q2 ⊃ Q3 ⊃ . . ., if Q =
i=1 Qi and Qi ∈ for each i, then Q ∈ . Indeed, we can repeat the proof
of (b) after replacing Beppo Levi’s Theorem with the Dominated Convergence
Theorem.
∞
Now, recall that X = ∞ n=1 Xn and Y = m=1 Ym for suitable disjoint sets Xn , Ym
of finite measure. For every m ∈ N, n ∈ N, we define Qmn = Q ∩ (Xn × Ym ). Let
us agree that Q ∈ M if and only if Qmn ∈ for every m and n. We have proved in
(b) and (d) that M is a monotone class. Furthermore E is contained in M by (a) and
(c). Since M ⊂ × T , it follows from Theorem 16.21 that M = × T .
We have proved that for every Q ∈ × T we have Qmn ∈ for every choice of
m and n. But Q = ∞ m,n=1 Qmn , and the sets Qmn are disjoint. Hence (c) implies
that Q ∈ . The proof is complete.

Exercise 16.10 Prove that µ × λ is indeed a measure on (X × Y, × T ), and that
it is a σ -finite measure.
Remark 16.3 We notice that we have defined the product measure µ × λ by means
of two (equal) iterated integrals. This might look strange, since the integral should
come after the measure. It would be possible to define µ × λ without any reference
to integration, and we would all agree that (µ × λ)(A × B) = µ(A)λ(B) for any
measurable rectangle; but a valid formula for any measurable subset of ×T would
be much harder to guess.
The equality of the two integrals in Theorem 16.22 suggests a further step. This
is the content of a celebrated result that we discussed from an abstract viewpoint in
Daniell’s approach to integration theory.
Theorem 16.23 (Fubini-Tonelli) Let (X, , µ) and (Y, T , λ) be σ -finite measur-
able spaces, and let f be a measurable function in × T .
(a) If 0 ≤ f ≤ +∞, and if
ϕ(x) =
. fx dλ, ψ(y) = f y dµ
Y X
for every x ∈ X, y ∈ Y , then ϕ is measurable in , ψ is measurable in Y , and
. ϕ dµ = f d(µ × λ) = ψ dλ.
X X×Y Y
(b) If f is real-valued, if
ϕ ∗ (x) =
. |f |x dλ
Y
!
and if X ϕ ∗ dµ < +∞, then f ∈ L1 (X × Y ).
16.6 Measure, Topology, and the Concrete Lebesgue Measure 423
(c) If f ∈ L1 (X × Y ), then fx ∈ L1 (Y ) for almost every x ∈ X, f y ∈ L1 (X) for

almost every y ∈ Y ; the functions ϕ and ψ are defined almost everywhere and
belong to L1 (X) and L1 (Y ) respectively. Furthermore,
ϕ(x) =
. fx dλ, ψ(y) = f y dµ.
Y X
Proof We consider (a). The conclusion holds if f is a simple function. In the general
case, there exists a sequence {sn }n of non-negative simple functions such that sn -
f pointwise on X × Y . We associate to each sn a function ϕn in the same way as we
associated ϕ to f . Then
. ϕn dµ = sn d(µ × λ)
X X×Y
for n = 1, 2, . . . We can apply Beppo Levi’s Theorem in Y to conclude that ! ϕn (x) -

ϕ(x)
! for every x ∈ X. A second application of Lev’s Theorem ! yields X ϕ dµ =
f d(µ × λ). If we swap the roles of x and y, we also get X×Y f d(µ × λ) =
!X×Y
Y ψ dλ.
The proof of (b) follows from (a) applied to |f |. To prove (c), we first observe that
(a) applies to f + and f − . Let us denote by ϕ1 and ϕ2 the functions associated to f +
and f − in the same way as ϕ was associated to f . The assumption f ∈ L1 (X × Y )
and the trivial inequality f + ≤ |f | imply that ϕ1 ∈ L1 (X). Similarly, ϕ2 ∈ L1 (Y ).
Notice now that fx = (f + )x − (f − )x , so that fx ∈ L1 (Y ) for every x ∈ X such
that ϕ1 (x) < +∞ and ϕ2 (x) < +∞. This must hold for almost every x ∈ X, since
ϕ1 , ϕ2 are integrable. For any such value of x, we have ϕ(x) = ϕ1 (x) − ϕ2 (x). In
particular ϕ ∈ L1 (X). The conclusion of (c) thus holds both with ϕ1 and f + , and
with ϕ2 and f − in place of ϕ and f . Subtracting these two equalities we conclude
the proof for fx . By the same token, the conclusion holds for f y .

16.6 Measure, Topology, and the Concrete Lebesgue

Measure
It is a matter of fact that no description of Measure Theory can be complete without

a construction of the concrete Lebesgue measure in Euclidean spaces. However, this
is a rather complicated construction, since the topology of .Rn plays a fundamental
role. The first step is a version of Urysohn’s Theorem 13.60 for locally compact
Hausdorff spaces. The proof we present here is more direct than the proof of
Theorem 13.65.
Definition 16.20 We write .K f to mean that K is a compact subset of X, that
f ∈ Cc (X), .0 ≤ f ≤ 1 on X, and that .f = 1 on K.
.
Similarly, we write .f V to mean that V is an open subset of X, that .f ∈

Cc (X), .0 ≤ f ≤ 1 on X, and that .supp f ⊂ V .
Theorem 16.24 (Urysohn’s Lemma for Locally Compact Hausdorff Spaces)
Suppose that X is a locally compact Hausdorff space, V is open in X, .K ⊂ V ,
and K is compact. Then there exists .f ∈ Cc (X) such that .K f V .
Proof Let .r1 = 0, .r2 = 1, and .{r3 , r4 , r5 , . . .} be an enumeration of .Q ∩ (0, 1). By
Proposition 13.1, there exist open sets .V0 and .V1 such that .V0 is compact and
K ⊂ V1 ⊂ V1 ⊂ V0 ⊂ V0 ⊂ V .
.
Suppose .n ≥ 2 and that .Vr1 , .Vr2 , . . . , .Vrn have been chosen so that .ri < rj implies
Vri ⊂ Vrj . Among the rational numbers .r1 , . . . , rn , choose the largest one, say .ri ,
.
such that .ri < rn+1 , and the smallest one, say .rj , such that .rj > rn+1 . Arguing
again as before, we select an open set .Vrn+1 such that
Vrj ⊂ Vrn+1 ⊂ Vrn+1 ⊂ Vri .

.
By induction we construct a countable family .{Vr | r ∈ Q∩[0, 1]} of open sets, such
that .K ⊂ V1 , .V0 ⊂ V , each .Vr has compact closure, and .s > r implies .Vs ⊂ Vr .
Define
r if x ∈ Vr
fr (x) =
.
0 otherwise,
1 if x ∈ Vs
. gs (x) =
s otherwise,
and .f = sup {fr | r ∈ Q ∩ [0, 1]}, .g = inf {gs | s ∈ Q ∩ [0, 1]}. It is easy to check
that f is lower semicontinuous and that g is upper semicontinuous. Clearly .0 ≤
f ≤ 1, .f (x) = 1 if .x ∈ K, and the support of f is contained in V . To conclude, we
prove that .f = g, getting the continuity of f .
The inequality .fr (x) > gs (x) is possible only if .r > s, .x ∈ Vr , and .x ∈ / Vs .
But .r > s implies .Vr ⊂ Vs . Hence .fr ≤ fs for all r and s, which yields .f ≤ g.
Suppose that .f (x) < g(x) for some x. There exist rational numbers r and s such
that .f (x) < r < s < g(x). The inequality .f (x) < r implies .x ∈
/ Vr . The inequality
.g(x) > s implies .x ∈ Vs . This is clearly a contradiction to the property that .s > r
implies .Vs ⊂ Vr .
We conclude that .f = g, and the proof is complete.

Theorem 16.25 (Finite Partition of Unity) Suppose .V1 , . . . , Vn are open subsets
of a locally compact Hausdorff space X. If K is compact and .K ⊂ V1 ∪ · · · ∪ Vn ,
then there exist functions .hi Vi , .i = 1, . . . , n, such that . ni=1 hi (x) = 1 for every
.x ∈ K.
Proof For every .x ∈ K there exists an open neighborhood .Wx with compact closure
.Wx ⊂ Vi for some index i which depends on x. By compactness, we select points
.x1 , . . . , xn such that .K ⊂ Wx1 ∪ · · · ∪ Wxn . For .1 ≤ i ≤ n, let

Hi =
. Wxj Wxj ⊂ Vi .
By Theorem 16.24, there exist functions .gi such that .Hi gi Vi . We now define
h1 = g1
.
h2 = (1 − g1 )g2
..
.
hn = (1 − g1 )(1 − g2 ) · · · (1 − gn−1 )gn .
It is now easy to check that .hi Vi and

n
. hi = 1 − (1 − g1 ) · · · (1 − gn ).
i=1
Since .K ⊂ H1 ∪ · · · Hn , at least on .gi (x) = 1 at each point .x ∈ K. The proof is

complete.

We are ready to state and proof the cornerstone of Topological Measure Theory.
Theorem 16.26 (Riesz Representation Theorem for Positive Measures) Let X
be a locally compact Hausdorff space, and let . : Cc (X) → R be positive and
linear.2 Then there exists a .σ -algebra . in X which contains all Borel sets, and
there exists a unique positive measure .µ on . such that
!
(a) .f = X f dµ for all .f ∈ Cc (X).
Furthermore
(b) .µ(K) is finite for every compact K.
(c) For every .E ∈ ,
µ(E) = inf {µ(V ) | E ⊂ V , V open} .

.
(d) The equality
µ(E) = sup {µ(K) | K ⊂ E, K compact}

.
2No continuity is assumed here. The term positive means that .f ∈ Cc (X) and .f (X) ⊂ [0, +∞)
imply .f ≥ 0.
holds for every open set E and for every measurable set E of finite measure.
(e) The measure .µ is complete.
Proof The proof is rather long, so we state some general principle: K will always
denote a compact set of X, and V an open set of X.
The uniqueness of the measure .µ is rather easy. Indeed, properties (c) and (d)
imply that .µ is determined by its values on compact sets. Let .µ1 and .µ2 two
measures for which the theorem holds, and assume that .µ1 (K) = µ2 (K) for
every K. Fix any such K, and let .ε > 0. By (b) and (c) there is V such that
.µ2 (V ) < µ2 (K) + ε. Theorem 16.24 provides a function f such that .K f V ,
hence
µ1 (K) =
. χK dµ1 ≤ f dµ1 = f = f dµ2
X X X
≤ χV dµ2 = µ2 (V ) < µ2 (K) + ε.

X
Letting .ε → 0 we find .µ1 (K) ≤ µ2 (K). Exchanging .µ1 and .µ2 yields .µ1 (K) =
µ2 (K). Hence uniqueness of .µ is proved.
We now construct . and .µ. For every open V , define
µ(V ) = sup {f | f V } .

.
Clearly .V1 ⊂ V2 implies .µ(V1 ) ≤ µ(V2 ); thus we define
µ(E) = inf {µ(V ) | E ⊂ V , V open}

.
for every .E ⊂ X. This definition agrees with the previous one if E is open. But
beware! Our set function .µ is not countably additive on the whole .2X , and for this
reason we need to introduce a good .σ -algebra. Let .F be the collection of all .E ⊂ X
such that .µ(E) is finite and
µ(E) = sup {µ(K) | K ⊂ E, K compact} .

. (16.2)
Our .σ -algebra is
= {E ⊂ X | E ∩ K ∈ F for every compact K} .

.
The rest of the proof consists in showing that . and .µ have the desired properties.
Evidently .µ is monotone, i.e. .µ(A) ≤ µ(B) if .A ⊂ B, and .µ(E) = 0 implies
.E ∈ F and then .E ∈ . Hence .µ is a complete measure, and (c) holds true by
definition.
Claim 1. If .Ei are subsets of X, .i ∈ N, then

"∞ # ∞

µ
. Ei ≤ µ(Ei ).
i=1 i=1
We first prove that .µ(V1 ∪ V2 ) ≤ µ(V1 ) + µ(V2 ) if .V1 and .V2 are open sets. Pick
any .g V1 ∪ V2 . There exists a partition of unity consisting of two functions .h1 ,
.h2 such that .hi Vi and .h1 (x) + h2 (x) = 1 for all .x ∈ supp g. Hence .hi g Vi ,
.g = gh1 + gh2 , and therefore
g = (gh1 ) + (gh2 ) ≤ µ(V1 ) + µ(V2 ).

.
Since g was arbitrary, it follows that .µ(V1 ∪ V2 ) ≤ µ(V1 ) + µ(V2 ). Let us now
consider the general case, and we assume without loss of generality that .µ(Ei )
is finite for every index i (otherwise the inequality is trivial). Let .ε > 0, and
consider open sets .Vi ⊃ Ei such that
ε
µ(Vi ) ≤ µ(Ei ) +
. .
2i

We set .V = ∞ i=1 Vi and we choose any .f V . This says that f has compact
support, so that .f V1 ∪ · · · ∪ Vn for some n. Iterating the previous case we get
f ≤ µ(V1 ∪ · · · ∪ Vn ) ≤ µ(V1 ) + · · · + µ(Vn )

.
∞

≤ µ(Ei ) + ε.
i=1

Since f was arbitrary and . ∞
i=1 Ei ⊂ V , we may conclude that
"∞ # ∞

µ
. Ei ≤ µ(V ) ≤ µ(Ei ) + ε.
i=1 i=1
We prove the Claim by letting .ε → 0.

Claim 2. If K is compact then .K ∈ F and
µ(K) = inf {f | K f } .

.
Of course this claim proves part (b) of the theorem. So, let .K f and .0 < α <
1. Define .Vα = {x ∈ X | f (x) > α}. It is clear that .K ⊂ Vα and that .αg ≤ f as
soon as .g Vα . As a consequence
1
µ(K) ≤ µ(Vα ) − sup {g | g Vα } ≤
. f.
α
Letting .α → 1− we conclude that .µ(K) ≤ f . In particular the measure of

K is finite. Now let .ε > 0. There exists an open set V such that .K ⊂ V and
.µ(V ) < µ(K) + ε. By Urysohn’s Lemma, for a suitable f we have .K f V .
Therefore .f ≤ µ(V ) < µ(K) + ε, and the proof of the claim is complete.
Claim 3. Every open set V satisfies (16.2). Hence .F containers every open set of
finite measure. Indeed, we fix a real number .α such that .α < µ(V ), and a function
.f V such that .α < f . If W is any open set such that .K = supp f ⊂ W ,
then .f W and .f ≤ µ(W ). Hence .f ≤ µ(K), and this argument provides
a compact .K ⊂ V such that .α < µ(K), so that (16.2) holds for V .
Claim 4. Suppose that .Ei ∈ F for .i ∈ N, and that .Ei ∩ Ej = ∅ when .i = j .
If .E = ∞ i=1 E i , then .µ(E) = ∞
i=1 µ(Ei ). If, in addition, .µ(E) is finite, then
also .E ∈ F .
We first prove that .µ(K1 ∪ K2 ) = µ(K1 ) + µ(K2 ) if .K1 and .K2 are disjoint
compact sets. Let .ε > 0. Urysohn’s Lemma provides .f ∈ Cc (X) such that
.f ≡ 1 on .K1 , .f ≡ 0 on .K2 , and .0 ≤ f ≤ 1. By Claim 2 there exists a function
g such that .K1 ∪ K2 g and .g < µ(K1 ∪ K2 ) + ε. We remark that .K1 fg
and .K2 (1 − f )g. The linearity of . implies
.µ(K1 ) + µ(K2 ) ≤ (fg) + λ(g − fg) = g < µ(K1 ∪ K2 ) + ε.
Letting .ε → 0 we deduce .µ(K1 ∪ K2 ) = µ(K1 ) + µ(K2 ). Now, if .µ(E) = +∞,

the claim follows from Claim 1. We assume that .µ(E) is finite, and we choose
.ε > 0. For every index i, .Ei ∈ F implies the existence of a compact set
.Hi ⊂ Ei such that
ε
µ(Hi ) > µ(Ei ) −
. .
2i
The set .Kn = H1 ∪ · · · ∪ Hn is compact, and we deduce by induction that

n
n
µ(E) ≥ µ(Kn ) =
. µ(Hi ) > µ(Ei ) − ε.
i=1 i=1

Letting first .n → +∞ and then .ε → 0, we see that .µ(E) ≥ ∞i=1 µ(Ei ), and the
conclusion follows from Claim 1. To prove that .E ∈ F we recall the definition
of convergent sequence, and deduce that

N
.µ(E) ≤ µ(Ei ) + ε
i=1
for some positive integer N. Hence .µ(E) ≤ µ(KN ) + 2ε, so that E satisfies
(16.2). The claim is now proved.
Claim 5. If .E ∈ F and .ε > 0, there exist a compact set K and an open set V
such that .K ⊂ E ⊂ V and .µ(V \ K) < ε.
Indeed, our definitions show that there are .K ⊂ E and .V ⊃ E such that
ε ε
µ(V ) −
. < µ(E) < µ(K) + .
2 2
Now, .V \ K is open and .V \ K ∈ F by Claim 3. Hence Claim 4 yields
µ(K) + µ(V \ K) = µ(V ) < µ(K) + ε.

.
Claim 6. If .A ∈ F and .B ∈ F , then .A\B ∈ F , .A∪B ∈ F , and .A∩B ∈ F .

Indeed, given .ε > 0, Claim 5 shows that there exist sets .Ki and .Vi such that
.K1 ⊂ A ⊂ V1 , .K2 ⊂ B ⊂ V2 , and .µ(Vi \ Ki ) < ε for .i = 1, 2. Since
A \ B ⊂ V1 \ K1 ⊂ (V1 \ K1 ) ∪ (K1 \ V2 ) ∪ (V2 \ K2 ),

.
it follows from Claim 1 that .µ(A \ B) ≤ ε + µ(K1 \ V2 ) + ε. Observing that

.K1 \ V2 is a compact subset of .A \ B, we see that .A \ B ∈ F . Furthermore,
.A ∪ B = (A \ B) ∪ B and .A ∩ B = A \ (A \ B), and an application of Claim 4
leads to the conclusion.

Claim 7. . is a .σ -algebra which contains all Borel sets.
Indeed, let K be any compact set in X. For every .A ∈ , .(X \ A) ∩ K =
K \ (A ∩ K), hence .(X \ A) ∩ K is the difference of twoelements of .F . Hence
n
.A ∈ implies .X \ A ∈ . Next we suppose that .A = i=1 Ai , where .Ai ∈ .
Let .B1 = A1 ∩ K and
Bn = (An ∩ K) \ (B1 ∪ · · · ∪ Bn−1 )

.
for .n ≥ 2. The collection .{Bi }i is a disjoint collection of elements of .F by

Claim 6, and .A ∩ K = ∞ i=1 Bi . Claim 4 implies .A ∩ K ∈ F , hence .A ∈ .
Finally, let C be a closed subset of X. Then .C∩K is compact, hence .C∩K ∈ F ,
so .C ∈ . As a particular case, .X ∈ . Since . is a .σ -algebra which contains
all closed sets, it must contain all the Borel sets.
Claim 8. .E ∈ F if and only if .E ∈ and .µ(E) is finite.
Indeed, let .E ∈ F . Claims 2 and 6 imply .E ∩ K ∈ F for every compact K,
hence .E ∈ . Conversely, let .E ∈ , .ε > 0. If .µ(E) is finite, then there exists
and open set V such that .E ⊂ V and .µ(V ) < +∞. By Claims 3 and 5, there
exists a compact .K ⊂ V such that .µ(V \ K) < ε. Since .E ∩ K ∈ F , there
exists a compact .H ⊂ E ∩ K such that .µ(E ∩ K) < µ(H ) + ε. Now
E ⊂ (E ∩ K) ∪ (V \ K)
.
implies
µ(E) ≤ µ(E ∩ K) + µ(V \ K) < µ(H ) + 2ε.

.
which implies .E ∈ F . The claim is proved, and part (d) of the theorem as well.
Claim 9. .µ is a measure on .. Indeed, this follows! at once from Claim 4 and 8.
Claim 10. .µ represents . in the sense that .f = X f dµ for every .f ∈ Cc (X).
This is part (a) of the theorem. !
Indeed, we observe that we may prove the inequality .f ≤ X f dµ, since by
linearity
. − f = (−f ) ≤ (−f ) dµ = − f dµ,

X X
and the equality follows. So, we call K the support of .f ∈ Cc (X), and let .[a, b]
be an intervals which contains the range of f . For every .ε > 0 we choose points
.y0 , .y1 , . . . , .yn such that .yi − yi−1 < ε and .y0 < a < y1 < . . . < yn = b. We
define the sets
Ei = {x ∈ X | yi−1 < f (x) ≤ yi } ∩ K.

.
As a continuous function, f is Borel measurable,

so that every .Ei is a Borel set.
Furthermore .Ei ∩ Ej = ∅ if .i = j , and .K = ni=1 Ei . Thus there exist open sets
.Vi ⊃ Ei such that
ε
µ(Vi ) < µ(Ei ) +
.
n
and such that .f (x) < yi + ε for
every .x ∈ Vi . We introduce a partition
n of unity
{hi }i such that .hi Vi and . ni=1 hi = 1 on K. Hence .f =
. i=1 f hi and
Claim 2 yields
" n #

n
µ(K) ≤
. hi = hi .
i=1 i=1
Observing that .hi f < (yi + ε)hi and that .yi − ε < f (x) for .x ∈ Ei , we have

n
n
f =
. (hi f ) ≤ (yi + ε)hi
i=1 i=1

n
n
= (|a| + yi + ε) hi − |a| hi
i=1 i=1

n ε
≤ (|a| + yi + ε) µ(Ei ) + − |a|µ(K)
n
i=1

n
ε
n
= (yi − ε)µ(Ei ) + 2εµ(K) + (|a| + yi + ε)
n
i=1 i=1
≤ f dµ + ε (2µ(K) + |a| + b + ε) .
X
!
The arbitrariness of .ε > 0 proves that .f ≤ X f dµ. The theorem is completely
proved.

The Riesz Representation Theorem will be our access point to the concrete
Lebesgue measure in .Rn . Although our approach might be considered rather
abstract, it has some advantages over the usual approach via an outer measure and a
Carathéodory completion.
Definition 16.21 A measure .µ defined on the .σ -algebra of Borel sets in a locally
compact Hausdorff space X is called a Borel measure on X.
Definition 16.22 A Borel set E is outer regular if
µ(E) = inf {µ(V ) | E ⊂ V , V open} .

.
The Borel set E is inner regular if
.µ(E) = sup {µ(K) | K ⊂ E, K compact} .
Finally, the measure .µ is regular if every Borel set is both inner and outer regular.
Remark 16.4 An inspection of Theorem 16.26 shows that the measure induced by
the positive linear functional . is not regular, in general. Indeed, the inner regularity
holds for open sets and for Borel sets of finite measure.
Definition 16.23 A set E in a topological space is .σ -compact if E is a countable
union of compact sets.
We can now show that .σ -compactness fills the gap of inner regularity.
Theorem 16.27 Let X be a locally compact, .σ -compact Hausdorff space, and let
, .µ be defined according to Theorem 16.26.
.
(i) If .E ∈ and .ε > 0, there exist a closed set F and an open set V such that
.F ⊂ E ⊂ V and .µ(V \ F ) < ε.
(ii) .µ is a regular measure on X.

(iii) If .E ∈ , there exist sets A and B such that A is .Fσ , B is .Gδ , .A ⊂ E ⊂ B,
and .µ(B \ A) = 0.

Proof By assumption, .X = ∞ i=1 Ki , where each .Ki is compact. Pick .E ∈ and
.ε > 0. There results .µ(Kn ∩ E) < +∞ and there exist open sets .Vn ⊃ Kn ∩ E such
that
ε
.µ (Vn \ (Kn ∩ E)) <
2n+1
∞
for .n ∈ N. We define .V = ∞i=1 Vi so that .V \ E ⊂ i=1 (Vn \ (Kn ∩ E)) and
.µ(V \E) < ε/2. The very same construction applies to .X\E in place of E, yielding
an open set .W ⊃ X \ E such that .µ(W \ (X \ E)) < ε/2. With .F = X \ W we get
F ⊂ E and .E \ F = W \ (X \ E). Conclusion (i)follows at once.
.
Every closed set F is .σ -compact, since .F = ∞ i=1 (F ∩ Ki ). Hence (i) implies

that every set .E ∈ is inner regular, and (ii) follows.
The proof of (iii) is now easy, choosing .ε = 1/j , .j = 1, 2, . . . This provides us
with closed sets .Vj such that .Fj ⊂ E ⊂ Vj and .µ(Vj \ Fj ) < 1/j .
.Fj and open sets
∞
We call .A = ∞ j =1 Fj , .B = j =1 Vj , to get .A \ E \ B and .µ(B \ A) = 0. By
definition A is .Fσ and B is .Gδ . The proof is complete.

The following regularity result will be used in the construction of the Lebesgue
measure.
Theorem 16.28 Let X be a locally compact Hausdorff space in which every open
set is .σ -compact. If .λ is a Borel measure on X such that .λ(K) is finite for every
compact set K, then .λ is regular.
Proof
! We introduce the positive linear functional . defined on .Cc (X) by .f =
X f dλ. The assumption on .λ implies that . is well-defined, and Theorem 16.26
provides us with a regular measure .µ such that
. f dλ = f dµ
X X
for every .f ∈ Cc (X). We will prove that .λ = µ, so that .λ is a regular measure as a

corollary.
Let V be an open set, so that .V = ∞ i=1 Ki , where every .Ki is compact. By
Urysohn’s Lemma we can choose .fi ∈ Cc (X) such that .Ki fi V . Let .gn =
max{f1 , . . . , fn }. Clearly .gn - χV pointwise, and Beppo Levi’s Theorem ensures
that
λ(V ) = lim
. gn dλ = lim gn dµ = µ(V ).
n→+∞ X n→+∞ X
For a generic Borel set E, we fix any .ε > 0. By Theorem 16.27, there exist a closed
set F and an open set V such that .F ⊂ E ⊂ V such that .µ(V \ F ) < ε. In particular
.µ(V ) ≤ µ(F ) + ε ≤ µ(E) + ε. We apply the previous considerations to the open
set .V \ F , and we get .λ(V \ V ) < ε and thus .λ(V ) < λ(E) + ε. We conclude that
λ(E) ≤ λ(V ) = µ(V ) ≤ µ(E) + ε

.
µ(E) ≤ µ(V ) = λ(V ) ≤ λ(E) + ε.
Hence .|λ(E) − µ(E)| < ε for every .ε > 0, and therefore .µ(E) = λ(E). The proof
is complete.

Since continuous functions with compact support appear as the basic ingredient
of the Riesz Representation Theorem, we investigate their role in Measure Theory.
The following is a basic result in the approximation of measurable functions by

means of more regular functions.
Theorem 16.29 (Lusin) Let X be a locally compact Hausdorff space, and let .µ be
a measure on X which has the properties described in Theorem 16.26. Suppose f is
a real-valued measurable function on X, that A is a measurable set such that .µ(A)
is finite and .f (x) = 0 whenever .x ∈
/ A. For every .ε > 0 there exists .g ∈ Cc (X)
such that
µ ({x ∈ X | f (x) = g(x)}) < ε.

.
Furthermore, we may select g so that
. sup {|g(x)| | x ∈ X} ≤ sup {|f (x)| | x ∈ X} .
Proof Suppose initially that .0 ≤ f < 1 and that A is compact. We consider a

sequence .{sn }n of simple functions as in Theorem 16.8. We define recursively .t1 =
s1 and .tn = sn − sn−1 . The function .2n tn is the characteristic function of a set
.Tn ⊂ A, and
∞

f (x) =
. tn (x), x ∈ X.
n=1
Fix an open set V such that .A ⊂ V and .V is compact. There exist compact sets .Kn
and open sets .Vn such that .Kn ⊂ Tn ⊂ Vn and .µ(Vn \ Kn ) < ε/2n . By Urysohn’s
Lemma there are functions .hn such that .Kn hn Vn . We define
∞
hn (x)
g(x) =
.
2n
n=1
for every .x ∈ X. This series converges uniformly on X, so that g is a continuous

−n
∞ in .V . Since .2 hn = tn except in .Vn \ Kn ,
function. The support of g is contained
we must have .g = f except in . n=1 (Vn \ Kn ). The measure of this latter set is
smaller than .ε. The first statement is proved if .0 ≤ f < 1 and if A is compact.
A simple scaling argument shows that the conclusion holds for every bounded
measurable f if A is compact. To remove the compactness condition, we remark
that the set A has finite measure, hence it must contain a compact set K such that
.µ(A \ K) is as small as we wish.
To remove
the boundedness condition on f , we set .Bn = {x ∈ X | |f (x)| > n},
so that . ∞
n=1 n = ∅. Hence .µ(Bn ) → 0 as .n → +∞. But f coincides with the
B
bounded function .(1 − χBn )f except on .Bn , and the proof of the first statement of
the theorem is complete in the general case.
Finally, let .R = sup {|f (x)| | x ∈ X}, and define
t if − R ≤ t ≤ R
ϕ(t) =
.
Rt
|t | if |t| > R.
The function .ϕ is continuous, and if g satisfies the first statement of the theorem,
then .g1 = ϕ ◦ g satisfies the second statement as well. The proof is complete.

16.6.1 The Concrete Lebesgue Measure
From now on, we want to construct the concrete Lebesgue measure using Theorem
16.26 as a starting point. This allows us to deduce some additional (regularity)
properties of the Lebesgue measure almost for free.
Due to some possible conflict with the index of several sequences, we will denote
by .k ≥ 1 the dimension of our basic Euclidean space .Rk . Let us recall that a k-cell
is any set of the form

W = x = (x1 , . . . , xk ) ∈ Rk αi < xi < βi , 1 ≤ i ≤ k
.
= (α1 , β1 ) × · · · × (αk , βk ), (16.3)
where .αi and .βi are given real numbers. Sometimes we can replace some or all
inequality signs .< by .≤: we already suspect that the difference will be a set of
measure zero. Finally, recall that the volume of the k-cell W is defined to be
(
k
. Vol(W ) = (βi − αi ) .
i=1
For our purposes, it will be better to replace the basic spherical neighborhood
B(a, δ) by a box. More precisely, for .a = (a1 , . . . , ak ) ∈ Rk and .δ > 0, we define
.
the .δ-box with corner at a as
Q(a, δ) = {x | ai ≤ xi < ai + δ, 1 ≤ i ≤ k} .
.
Definition 16.24 For .n ∈ N we define .Pn as the set of all points .x ∈ Rk whose
coordinates are integral multiples of .2−n , and we define .n as the collection of all
.2
−n -boxes with corners at points of .P .
n
∞ non-empty open set of .R is a countable union of disjoint

Theorem 16.30 Every k
boxes belonging to . i=1 i .

Proof Let V be an open set. Every point .x ∈ V lies in an open ball which lies, in
turn, in V . Hence .x ∈ Q ⊂ V for some suitable Q which belongs to some .n .
Equivalently, V is the union of all boxes which lie in V and which belong to some
.n .
From this collection of boxes we first select those boxes belonging to .1 , and
we remove those in .2 , 3 , . . . which lie in any of the selected boxes. From the
remaining collection, we select those boxes of .2 which lie in V , and we remove
those in .3 , 4 , . . . which lie in any of the selected boxes. This procedure provides
a countable collection of disjoint boxes in .1 ∪ 2 ∪ 3 ∪ · · · whose union is V .

The next theorem defines the concrete Lebesgue measure in .Rk .
Theorem 16.31 There exists a positive complete measure m defined on a .σ -algebra
M of .Rk , with the following properties:
.
(a) .m(W ) = Vol(W ) for every k-cell W .

(b) .M contains all Borel sets; more precisely, .E ∈ M if and only if there exist sets
A and B such that .A ⊂ E ⊂ B, A is .Fσ , B is .Gδ , and .m(B \A) = 0. Moreover,
m is a regular measure.
(c) m is invariant under translations.
(d) If .µ is any Borel measure on .Rk , invariant under translations and such that
.µ(K) is finite for every compact K, then there exists a constant c such that
.µ(E) = cm(E) for every .E ∈ M.
Proof Our proof, as we said, constructs .M and m via the Riesz Representation
Theorem. Clearly enough, we need a positive linear functional to begin with. This
functional is precisely the Riemann (or Cauchy) integral in .Rk , which was already
studied in Sect. 15.1. For the reader’s sake we recall here the basic ideas.
Let .f ∈ Cc (Rk ), and define for .n ∈ N the functional
1
n f =
. {f (x) | x ∈ Pn } ,
2nk
where .Pn was introduced in Definition 16.24. Let W be an open k-cell containing
the support of f . By uniform continuity, there exist an integer N and functions g,
h with support in W , such that (i) g and h are constant on each box of .N , (ii)
.g ≤ f ≤ h, (iii) .h − g < ε.
It is easy to see that
N g = n g ≤ n f ≤ n h = N h
.
for every .n > N. As a consequence,
. lim sup n f − lim inf n f < ε Vol(W ),

n→+∞ n→+∞
which implies that the limit .f = limn→+∞ n f exists (as a finite real number).
As a simple exercise, the reader can prove that . is a positive linear operator on
.Cc (R ). The .σ -algebra .M and the measure m are now defined according to Theorem
k
16.26. Since .Rk is .σ -compact, (b) is automatically true.

To prove (a) we use the k-cell W of (16.3). For any .r ∈ N, we call .Er the union
of those boxes of .r whose closure belongs is contained in W . We select .fr such
that .Er fr W and we put .gr = max{f1 , . . . , fr }. It follows from the definition
of . that
. Vol(Er ) ≤ fr ≤ gr ≤ Vol(W ).
When we let .r → +∞, .Vol(Er ) → Vol(W ) and
gr =
. gr dm → m(W )
Rk
by Beppo Levi’s Theorem, observing that .gr - χW . Thus .m(W ) = Vol(W ) for
every open k-cell W . Since any k-cell is the intersection of a decreasing sequence
of open k-cells, the proof of (a) is complete.
Let us now remark what follows: if .λ is a Borel measure on .Rk and .λ(E) = m(E)
for every box E, then the same equality holds for every open set E. This is indeed
a consequence of Theorem 16.30. Once this is established, it follows from Theorem
16.28 that the equality holds for every Borel set, since .λ and m are regular.
Consider statement (c). Let .x ∈ Rk and define .λ(E) = m(E + x). Clearly .λ is a
measure, and (a) implies that .λ coincides with m on all boxes, and thus on all Borel
sets: this means that .m(E) = m(E + x). The same equality holds for every .E ∈ M
because of (b). Hence (c) is proved.
Finally, suppose that .µ satisfies the hypotheses of (d). Let .Q0 be a 1-box, and set
.c = µ(Q0 ). Since .Q0 is the union of .2
nk disjoint .2−n -boxes which are translates of
each other, we have
2nk µ(Q) = µ(Q0 ) = cm(Q0 ) = c · 2nk m(Q)

.
for every .2−n -box Q. Theorem 16.30 implies that .µ(E) = cm(E) for all open sets
E of .Rk , and the proof of (d) is complete.

Remark 16.5 The symbolism for the concrete Lebesgue measure m is very rich.
Firstly, it may be useful to write .mk in order to denote the dimension of the
Euclidean space. But sometimes .λk is preferred3 to .mk . Some books use .Lk , and
even .|E| to denote the Lebesgue measure of E. When integrals come into play, the
3 .λ is reminiscent of the “L” of Lebesgue.

pedantic notation
. f dmk
Rk
is often replaced by
. f (x) dx,
Rk
where x is a dummy variable. However we need to point out that the bad habit of
calling .dx the Lebesgue measure, which occurs in several Calculus texts.
It is interesting to show that non-measurable sets exist for the concrete Lebesgue
measure.
Theorem 16.32 Every set of positive Lebesgue measure contains a non-
measurable subset.
This is indeed a corollary of a more general result.
Theorem 16.33 If .A ⊂ R and if every subset of A is Lebesgue measurable, then
m(A) = 0.
.
Proof The basic tool in the proof is the structure of .R as a group relative to addition.
Consider .Q as a subgroup of .R, and introduce an equivalence relation as follows:
.x ∼ y if and only if .x − y ∈ Q. For .x ∈ R, we write .[x]∼ = {y ∈ R | y ∼ x},
the equivalence class of x. Using the Axiom of Choice, we construct a set E which
contains exactly one point from each equivalence class of .Q in .R.
Claim 1. If .r ∈ Q, .s ∈ Q and .r = s, then .(E + r) ∩ (E + s) = ∅. Indeed, suppose
.x ∈ (E + r) ∩ (E + s). Then .x = y + r = z + s for some .y ∈ E, .z ∈ E, .y = z.
But .y − z = s − r ∈ Q, so that y and z lie in the same equivalence class of .Q, a

contradiction.
Claim 2. Every point .x ∈ R belongs to .E+r, for some .r ∈ Q. Indeed, there exists
one and only one point y of E such that .y ∼ x. We merely define .r = x − y, and
the claim is proved.
Let now .t ∈ Q and put .At = A ∩ (E + t). By assumption, .At is Lebesgue
measurable. Let K be a compact subset of .At , and let H be the union of all
translates .K + r, as r ranges over .Q ∩ [0, 1]. The set H is bounded, hence .m(H )
is finite. Since .K ⊂ E +t, Claim 1 shows that the sets .K +r are pairwise disjoint.
Thus

.m(H ) = m(K + r).
r
But .M(K + r) = m(K), hence .m(K) = 0. Since this holds forevery compact
K ⊂ At , we deduce that .m(At ) = 0. Claim 2 shows that .A = {At | t ∈ Q}.
.
Since .Q is a countable set, we conclude that .m(A) = 0, and the proof is complete.

Lusin’s theorem is a powerful result that allows us to approximate different classes
of functions by means of continuous functions.
Theorem 16.34 Let X be a locally compact Hausdorff space, endowed with a

measure as in Theorem 16.26. If .1 ≤ p < +∞, then .Cc (X) is dense in .Lp (X).
Proof Let S be the set of all the measurable simple functions s on X such that
.µ({x ∈ X | s(x) = 0}) < +∞. We claim that S is dense in .Lp (X). Firstly, .S ⊂ Lp .
Suppose that .f ≥ 0 is a function in .Lp , and let .{sn }n be a sequence of measurable
simple functions which monotonically converges to f . Since .0 ≤ sn ≤ f , we have
.sn ∈ L for every n, hence .sn ∈ S. But .|f − sn |
p p ≤ f p , hence the Dominated
Convergence Theorem yields .*f − sn *p → 0 as .n → +∞, and f belongs to the

closure of S in .Lp . The case of a function f with variable sign follows from the
decomposition .f = f + − f − .
Now let .ε > 0 and .s ∈ S. By Lusin’s Theorem, there exists .g ∈ Cc (X) such that
.g = s except on a set of measure .< ε, and .|g| ≤ supX |s|. Hence
*g − s*p ≤ 2ε1/p *s*p .

.
Since S is dense in .Lp , the proof is complete.

We conclude this section with the answer to a basic question: what functions are
Riemann-integrable?
Theorem 16.35 (Lebesgue) A bounded function f on an interval .[a, b] is
Riemann-integrable if and only if the set of points at which f is discontinuous
has measure zero.
Proof We introduce the lower and the upper envelope of f as follows:
.g(y) = sup {inf {f (x) | |x − y| < δ} | δ > 0}

h(y) = inf {sup {f (x) | |x − y| < δ} | δ > 0} .
The following statements are easy exercises:

(a) for every .x ∈ [a, b], .g(x) ≤ f (x) ≤ h(x). More precisely, .f (x) = g(x) if
and only if f is lower semicontinuous at x, and .f (x) = h(x) if and only if
f is upper semicontinuous at x. In particular, .g(x) = h(x) if and only if f is
continuous at x.
(b) The function g is lower semicontinuous, and the function h is upper semicon-
tinuous.
(c) If .ϕ is any lower semicontinuous function such that .ϕ ≤ f on .[a, b], then .ϕ ≤ g
on .[a, b]. Similarly, if .ψ is any upper semicontinuous function such that .f ≤ ψ
on .[a, b], then .ψ ≤ h on .[a, b].
16.7 Mollifiers and Regularization 439
Consider now any step function .ϕ ≥ f ; in particular .ϕ ≥ h except at a finite number

of points. Hence
b b
. h≤ f dx.
a a
On the other hand, by definition of the upper envelope of f , there exists a sequence
{ψn }n of step functions such that .ϕn . f . Since f is bounded on .[a, b], The
.
Dominated Convergence Theorem implies
b b b
. h dx = lim ϕn dx ≥ f dx.
a n→+∞ a a
We have thus proved that
b b
. f dx = h dx.
a a
By the same token,
b b
. f dx = g dx.
a a
!b
It follows that f is R-integrable on .[a, b] if and only if . a (h − g) dx = 0. Since
.h ≥ g, we deduce that this happens if and only if .h = g a.e. on .[a, b]. Recalling (a)
above, this is equivalent to the fact that the set of points at which f is discontinuous
has measure zero. The proof is complete.

16.7 Mollifiers and Regularization
In the whole section, we will be working on the measurable space .X = RN with the
standard Lebesgue measure.4
Our aim is to introduce a general technique for regularizing integrable functions.
As a by-product, we will provide a compactness result in .Lp . Let us start with some
notation.
Definition 16.25 Let . be an open subset of .RN , .N ≥ 1. We write

D() = u ∈ C ∞ () the support of u is a compact subset of .
.
4The choice of N as the fixed dimension of the Euclidean space allows us to use freely the index
n.
A multi-index is an element .α = (α1 , . . . , αN ) of .NN . Its length is

N
|α| =
. αj ,
j =1
and we will write

∂ α
∂j =
. , D α = ∂1α1 ∂2α2 · · · ∂NN .
∂xj
Exercise 16.11 Prove that the function .f : R → R such that
e1/x if x < 0
f (x) =
.
0 if x ≥ 0
belongs to .C ∞ (R). Hint: show by induction that for every .n ∈ R and every .x < 0
there results .D n f (0) = 0 and .D n f (x) = Pn (1/x)e1/x , where .Pn is a polynomial.
Definition 16.26 (Standard Mollifiers) Let .! : RN → R such that
⎧
⎪
1
⎪
⎪ e |x|2 −1
⎨ if |x| < 1
!(x) =
.
! 1
|x|2 −1 dx
⎪
⎪ B(0,1) e
⎪
⎩0 if |x| ≥ 1.
The standard mollifiers are the functions .!n : RN → R such that .!n (x) = nN !(nx)
for every .x ∈ RN and every .n ∈ N.
! Prove that for every .n ∈ N, .!n ∈ D(R ), .supp !n ⊂ B(0, 1),

Exercise 16.12 N
.!n ≥ 0 and .
RN !n (x) dx = 1.
Since we will often need to restrict functions to relatively compact subsets of ., we
introduce a new space.
Definition 16.27 (Local Lebesgue Spaces) Let . be an open subset of .RN , and
let .ω be such that .ω is open and .ω is a compact subset of .. For brevity, we will
write .ω ⊂⊂ .
For every .1 ≤ p < ∞ we define

Lloc () = u ∈ R u|ω ∈ Lp (ω) for every ω ⊂⊂ .
p
.
Although local Lebesgue spaces are not normed spaces,5 yet we introduce a
definition of convergent sequences.
p p
Definition 16.28 (Convergence in .Lloc ) A sequence .{un }n in .Lloc () converges
to u if and only if for every .ω ⊂⊂
. lim |un − n|p dx = 0.

n→+∞ ω
Here comes the most useful tool of Harmonic Analysis: the convolution. We
consider a particular case, which however is sufficient for our purposes.
Definition 16.29 (Convolution in Local Lebesgue Spaces) Suppose .u ∈ L1loc ()
and .v ∈ Cc (RN ) are such that
1
. supp v ⊂ B 0, .
n
For every .n ∈ N, the convolution .v ∗ u is defined on the set

1
. n = x ∈ d(x, ∂) >
n
by
v ∗ u(x) =
. v(x − y)u(y) dy = v(y)u(x − y) dy.
B(0,1/n)
Definition 16.30 (Translation Operator) If .|y| < 1/n, the translation of .u ∈

L1loc () by y is defined on .n by
τy u(x) = u(x − y).

.
The use of convolution products in regularization is explained by the next result.

Theorem 16.36 Suppose .u ∈ L1loc () and .v ∈ D(RN ) are such that
1
. supp v ⊂ B 0, .
n
5 The intuitive reason is that there is no norm that can take into account all possible sets .ω ⊂⊂
at the same time.
Then .v ∗ u ∈ C ∞ (n ), and for every multi-index .α there results
D α (v ∗ u) = (D α v) ∗ u.
.
Proof We prove the result under the additional assumption .|α| = 1. The general
case follows by induction on .|α|. Fix .x ∈ n : there exists .r > 0 such that6
.B(0, r) ⊂ n . Hence
1
ω = B x, r +
. ⊂⊂ ,
n
and if .0 < |ε| < r we have
v ∗ u(x + εα) v(x + εα − y) − v(x − y)

. = u(y) dy.
ε ω ε
Since .v(x + εα − y) − v(x − y) = εD α v(x − y) + o(ε) and

v(x + εα − y) − v(x − y)
. sup y ∈ ω, 0 < |ε| < r ,

ε
the Dominated Convergence Theorem yields
D α (v ∗ u)(x) =
. D α v(x − y)u(y) dy = (D α v) ∗ u(x).
ω

Theorem 16.37 (Continuity of Translations) Let .ω ⊂⊂ .
(a) If .u ∈ C(), then

. lim sup τy u(x) − u(x) = 0.
y→0 x∈ω
p
(b) If .u ∈ Lloc () for some .1 ≤ p < ∞, then
* *
. lim *τy u − u*p = 0.
y→0
Proof
(a) Choose an open set U such that .ω ⊂⊂ U ⊂⊂ . Since u is uniformly
continuous on U , the conclusion follows immediately from (16.4).
6 .RN is locally compact!

(b) Pick .ε > 0, and choose now an open set U such that .ω ⊂⊂ U ⊂⊂ . By
Theorem 16.34 a function .v ∈ Cc (U ) exists such that .*u − v*Lp (U ) ≤ ε. By
(a) there exists .0 < δ < d(ω, ∂U ) such that, if .|y| < 1/n, then

. sup τy u(x) − u(x) ≤ ε.
x∈ω
Hence, if .|y| < 1/n,

* * * * * *
. *τy u − u* p ≤ *τy u − τy v * p + *τy v − v * p + *v − u*Lp (ω)
L (ω) L (ω) L (ω)

≤ 2 *u − v*Lp (U ) + mN (ω) sup τy u(x) − v(x)
1/p
x∈ω

≤ 2 + mN (ω)1/p ε,
where .mN denotes the Lebesgue measure in .RN , as usual. Since .ε > 0 is
arbitrary, the proof is complete.

Theorem 16.38 (Regularization Theorem)
(a) If .u ∈ C(), then .{!n ∗ u}n converges uniformly to u on every compact subset
of ..
p
(b) If .u ∈ Lloc (), .1 ≤ p < ∞, then .{!n ∗ u}n converges to u in .Lp ().
Proof
(a) We claim that, for .n ∈ N sufficiently large,

. sup |!n ∗ u(x) − u(x)| ≤ sup sup τy u(x) − u(x) . (16.4)
x∈ω |y|< n1 x∈ω
Indeed, for n sufficiently large, .ω ⊂⊂ n . From the properties of the mollifiers,

for every .x ∈ ω we have

. |!n ∗ u(x) − u(x)| = !n (y) (u(x − y) − u(x)) dy
B(0,1/n)
≤ sup sup |u(x − y) − u(x)| ,

|y|<1/n x∈ω
and (16.4) is proved. The conclusion follows from the continuity of translations.
(b) We claim that, for every .n ∈ N sufficiently large,
* *
. *!n ∗ u − u*Lp (ω) ≤ sup *τy u − u*Lp (ω) . (16.5)
|y|<1/n
Indeed, by Hölder’s inequality, for every .x ∈ ω we have

. |!n ∗ u(x) − u(x)| = !n (y) (u(x − y) − u(x)) dy
B(0,1/n)
1/p
≤ !n (y) |u(x − u) − u(x)|p dy .
B(0,1/n)
Interchanging the order of integration,
. |!n (y) ∗ u(x) − u(x)|p dx ≤ dx !n (y) |u(x − y) − u(x)|p dy

ω ω B(0,1/n)
= dy !n (y) |u(x − y) − u(x)|p dx

B(0,1/n) ω
≤ sup |u(x − y) − u(x)|p dx,

|y|<1/n ω
and (16.5) follows. The conclusion follows from the continuity of translations.

As we promised, the convolution with mollifiers allows us to approximate integrable
functions with smooth functions.
Theorem 16.39 (Smooth Functions Are Dense in Lebesgue Spaces) If .1 ≤ p <
∞, then .D() is dense in .Lp ().
Proof Theorem 16.34 ensures the density of .Cc () in .Lp (). Fix .u ∈ Cc () and
an open set .ω such that .supp u ⊂⊂ ω ⊂⊂ . Taking n sufficiently large, the support
of .un = !n ∗ u is contained in .ω, and .un ∈ C ∞ (RN ). It follows that .un ∈ D(),
and the conclusion follows from Theorem 16.38.

It is not too difficult to convince ourselves that we cannot approximate globally
a continuous function with a smooth function, or equivalently that the uniform
convergence of .!n ∗ u to u in part (a) of Theorem 16.38 is optimal.
On the other hand, part (b) extends to the case . = RN .
Theorem 16.40 Let .1 ≤ p < ∞. If .u ∈ Lp (RN ), then .*!n ∗ u*p ≤ *u*p , and
.!n ∗ u → u in .L (R ).
p N
Proof By Hölder’s inequality,

1/p

. |!n ∗ u(x)| = u(y)!n (x − y) dy ≤ |u(y)| !n (x − y) dy .
p

RN RN
16.8 Compactness in Lebesgue Spaces 445
Using Fubini’s Theorem we see that
. |!n ∗ u(x)|p dx ≤ dx |u(y)|p !n (x − y) dy

RN RN RN
= dy |u(y)|p !n (x − y) dx
RN RN
= |u(y)|p dy.
RN
This proves the first part of the theorem. Let now .u ∈ Lp (RN ) and .ε > 0. By
Theorem 16.34 there exists .v ∈ Cc (RN ) such that .*v − u*p ≤ ε. By Theorem
16.38, .!n ∗ v → v in .Lp (RN ). Fix .m ∈ N such that, for every .n ≥ m, there results
.*!n ∗ v − v*p ≤ ε. For these values of n,
. *!n ∗ u − u*p ≤ *!n ∗ (u − v)*p + *!n ∗ v − v*p + *v − u*p ≤ 3ε,

16.8 Compactness in Lebesgue Spaces
Theorem 16.41 (M. Riesz) Let be an open subset of RN , 1 ≤ p < ∞ and

suppose that S ⊂ Lp () satisfies

(a) c = sup *u*Lp () u ∈ S ∈ R;
(b) for every ε > 0 there exists ω ⊂⊂ such that sup *u*Lp (\ω) u ∈ S ≤ ε;
* *

(c) for every ω ⊂⊂ , limy→0 sup *τy u − u* p u ∈ S = 0.
L (ω)
Then S is relatively compact in Lp ().

Proof Fix ε > 0, and let ω be as in condition (b). Using (c) we see that there exists
δ ∈ (0, d(ω, ∂) such that if |y| < δ, then
* *

. sup *τy u − u* p u ∈ S ≤ ε.
L (ω)
Choose now an integer n > 1/δ. It follows from (16.5) that

)
* * * *

. sup *τy u − u* p u ∈ S ≤ sup sup *τy u − u*Lp (ω) u ∈ S ≤ ε.
L (ω)
|y|<1/n
(16.6)
Define

.U = x ∈ R d(x, ω) < 1
N
⊂⊂ .
n
We claim that the collection F = {!n ∗ u|ω | u ∈ S} satisfies the assumptions of

Corollary 13.3.
Indeed, by condition (a), for every u ∈ S and every x ∈ ω, we have
. |!n ∗ u(x)| ≤ !n (x − z) |u(z)| dz ≤ *!n *∞ *u*L1 (U ) ≤ c1 .

U
Similarly, for every x ∈ ω and y ∈ ω we have
. |!n ∗ u(x) − !n ∗ u(y)| ≤ |!n (x − z) − !n (y − z)| |u(z)| dz

U

≤ sup |!n (x − z) − !n (y − z)| z ∈ RN
*u*L1 (U ) ≤ c2 |x − y|.
Hence F is relatively compact in the space of bounded continuous functions on ω.

But
. *v*Lp (ω) ≤ mN (ω)1/p *v*L∞ (ω) ,
F is also relatively compact in Lp (ω).7 Now (16.6) implies the existence of a finite
cover of F|ω in Lp (ω) by balls of radius 2ε. We finally use assumption (b) to ensure
the existence of a finite cover of F in Lp () by balls of radius 3ε. We have thus
proved that F is totally bounded in the complete metric space Lp (), hence its
closure is relatively compact by Theorem 13.78.

16.9 The Radon-Nykodim Theorem
Let us consider a .σ -finite measurable space .(X, ) together with a measure .µ, and
let .f ∈ L1 (X) be a non-negative function. It is easy to check that
ν : A ∈ →
. f dµ
A
7 We have proved that the space of bounded continuous functions on ω is continuously embedded
into Lp (ω).
16.9 The Radon-Nykodim Theorem 447
is a measure on X which satisfies the property
µ(A) = 0 ⇒ ν(A) = 0.
.
This motivates the following definition.

Definition 16.31 A measure .ψ on .(X, ) is absolutely continuous with respect to
µ if and only if for every .A ∈ such that .µ(A) = 0 there results .ψ(A) = 0.
.
The main result of this section is the following description of absolutely continuous
measures.
Theorem 16.42 Let .(X, ) be a .σ -finite measurable space with measure .µ. If .ν
! respect to .µ, then there
is a finite measure on .(X, ), absolutely continuous with
exists a measurable function f on X such that .f ≥ 0, . X f dµ < ∞ and
ν(E) =
. f dµ for every E ∈ .
E
The proof of this theorem requires some technical results. Let .ϕ : → R a

countably additive function, i.e. .ϕ(∅) = 0 and
" ∞
# ∞

ϕ
. En = ϕ (En )
n=1 n=1
for every sequence .{En }n of pairwise disjoint sets in .. We define
ϕ + (X) = sup {ϕ(A) | A ∈ } .

.
Lemma 16.1 There exists a set .E ∈ such that .ϕ + (X) = ϕ(E). Moreover
ϕ(A) ≥ 0 if A ⊂ E
.
ϕ(A) ≤ 0 if A ⊂ X \ E.
Proof We consider a minimizing sequence .{En }n for .ϕ + (X) such that
1
ϕ + (X) ≥ ϕ(En ) ≥ ϕ + (X) −
. for n ≥ 1.
2n
If .A ⊂ En , writing .ϕ(En ) = ϕ(A) + ϕ(En \ A), we see that
1
ϕ + (X) −
. ≤ ϕ(En ) ≤ ϕ(A) + ϕ + (X).
2n
Hence .ϕ(A) ≥ −2−n . Now, for every .n ≥ 1,

" ∞
#

ϕ
. Em = ϕ (En ∪ (En+1 \ En ) ∪ (En+2 \ En+1 ) ∪ · · · )
m=n
1 1
≥ ϕ(En ) − − n+2 − · · ·
2n+1 2
2
≥ ϕ + (X) − n .
2
It follows that
"∞ ∞ # " ∞
#

ϕ
. Em = lim ϕ Em ≥ ϕ + (X).
n→+∞
n=1 m=n m=n
Hence the set

∞
∞
E=
. Em
n=1 m=n
satisfies .ϕ(E) = ϕ + (X). To prove the second part, we notice that if .A ⊂ E, then
ϕ + (X) = ϕ(E) = ϕ(A) + ϕ(E \ A) ≤ ϕ(A) + ϕ + (X),

.
or .ϕ(A) ≥ 0. If .A ⊂ X \ E, then
ϕ + (X) ≥ ϕ(A ∪ E) = ϕ(A) + ϕ(E) = ϕ(A) + ϕ + (X),

.
or .ϕ(A) ≤ 0. The proof is complete.

Proof of Theorem 16.42 We first consider the case .µ(X) < ∞. We propose a
variational proof, based on a minimization problem. Consider indeed the set

I = g ≥ 0
. g dµ ≤ ν(E) for every E ∈ .
E
We claim that there exists .f ∈ I such that

. f dµ = sup g dµ g ∈ I = S.
X X
16.9 The Radon-Nykodim Theorem 449
For every integer .n ≥ 1, let .gn ∈ I be such that
1
. gn dµ ≥ S − .
X n
We define
fn = max {g1 , . . . , gn } ,
. f = lim fn .
n→+∞
It is not difficult to check that .fn ∈ I for every .n ≥ 1. For instance, in the case
n = 2, for every .E ∈ we have
.
. f2 dµ = g1 dµ + g2 dµ
E E∩{g1 ≥g2 } E∩{g2 >g1 }
≤ ν (E ∩ {g1 ≥ g2 }) + ν (E ∩ {g2 > g1 })

= ν(E).
The !general case is similar. Since .fn - f , Beppo Levi’s Theorem yields .f ∈ I
and . X f dµ = S. The claim is thus proved. !
Let us now show that for every .E ∈ there results
! .ν(E) =
E f dµ. By
definition of .I, we only need to show that .ν(E) ≤ E f dµ. Suppose not, so that
there exist .ε > 0 and .E0 ∈ such that
. (f + ε) dµ < ν(E0 ).
E0
Since .ν is absolutely continuous with respect to .µ, we see that .µ(E0 ) > 0. If we
define
ϕ(E) = ν(E) −
. (f + ε) dµ,
E
Lemma 16.1 yields two subset F , G of .E0 such that .E0 = F ∪ G,
ν(A) −
. (f + ε) dµ ≥ 0 if A ⊂ F
A
ν(A) − (f + ε) dµ ≤ 0 if A ⊂ G.
A
The function g defined as
f (x) if x ∈
/F
g(x) =
.
f (x) + ε if x ∈ F
belongs to .I and
. g dµ > f dµ = S.
X X
This is impossible, and the proof is complete in the particular case .µ(X) < ∞.
To deal with the general case, let .{Xi }i be a sequence of pairwise disjoint
measurable sets such that .µ(Xi ) is finite for every i and
∞

X=
. Xi .
i=1
Applying the previous step in each .Xi , we see that there exist functions .fi ∈
L1 (Xi , µ) such that .fi ≥ 0 and
ν(E) =
. fi dµ for every E ⊂ Xi .
E
If we set .f = fi on .Xi , we define a function f on X. If E is contained in a finite

union of the sets .Xi , we have
ν(E) =
. f dµ.
E
Furthermore, for every .n ≥ 1,
.
∞ f dµ ≤ ν(X),
i=1 Xi
hence .f ∈ L1 (X, µ). If now E is any measurable set, then
. f dµ = lim f dµ
E n→+∞ E∩n X
i=1 i
" #

n
= lim ν E ∩ Xi = ν(E),
n→+∞
i=1

Exercise 16.13 By inspection of the previous proof, deduce that if .ν is just .σ -finite
instead of finite, the conclusion of the Theorem holds, but f need not be integrable
with respect to .µ.
16.10 A Strong Form of the Fundamental Theorem of Calculus 451
16.10 A Strong Form of the Fundamental Theorem

of Calculus
The basic result of Riemann integration theory is for sure the formula
b
. f % (x) dx = f (b) − f (a),
a
which expresses a function as the integral of its derivative. We refer back to Theorem
10.12 for a precise statement.
Important: Question
Does there exist a Fundamental Theorem of Calculus for Lebesgue integrals?
Such a natural question requires several new ideas to be answered. In the rest of
the section, .λ = m1 will always denote the Lebesgue measure in .R, and “almost
everywhere” will always refer to this measure.
Definition 16.32 Let .a ∈ R and .δ > 0. If .ϕ : (a, a + δ) → R, we define
. lim inf ϕ(h) = sup {inf {ϕ(h) | a < h < t} | a < t ≤ a + δ}

h→a
lim sup ϕ(h) = inf {sup {ϕ(h) | a < h < t} | a < t ≤ a + δ} .
h→a
These quantities are called respectively the lower right limit and the upper right limit
of the function .ϕ at the point a. Similarly, if .ϕ : (a − δ, a) → R, we define the lower
left limit and the upper left limit of .ϕ at a as
. lim inf ϕ(h) = sup {inf {ϕ(h) | t < h < a} | a − δ ≤ t < a}

h→a
lim sup ϕ(h) = inf {sup {ϕ(h) | t < h < a} | a − δ ≤ t < a} .
h→a
Definition 16.33 (Dini’s Derivatives) Let .a ∈ R and .δ > 0. If .f : [a, a + δ) → R,

we define
f (a + h) − f (a)
D+ f (a) = lim inf
.
h→0 h
f (a + h) − f (a)
D + f (a) = lim sup .
h→0 h
If .f : (a − δ, a] → R, we define
f (a + h) − f (a)
D− f (a) = lim inf
.
h→0 h
f (a + h) − f (a)
D − f (a) = lim sup .
h→0 h
Exercise 16.14 Prove that .D + f (a) is the largest limit of a sequence
f (a + hn ) − f (a)
.
hn n
where .hn > 0 and .limn→+∞ hn = 0. Conjecture and prove similar statements for
the remaining Dini’s derivatives.
The four Dini’s derivatives describe the lack of differentiability of f at a, since it is
clear that f is differentiable at a if and only if the four Dini’s derivatives are finite
and coincide.
Theorem 16.43 Let .(a, b) be an open interval, and let f be a real-valued function
defined on .(a, b). There exist at most countably many points .x ∈ (a, b) such that
D+ f (x) = D + f (x)
.
and
D− f (x) = D − f (x)
.
both exist in .[−∞, +∞] but are different.

Proof In case .D+ f (x) = D + f (x), we call .f+% (x) the common value. and similarly
for .f−% (x). These quantities may well be infinite. Let

A = x ∈ (a, b) f+% (x) < f−% (x)
.

B = x ∈ (a, b) f+% (x) > f−% (x) .
For each point .x ∈ A we select a rational number .r(x) such that .f+% (x) < r(x) <
f−% (x). Next we select rational numbers .s(x) and .t (x) such that .a < s(x) < x <
t (x) < b,
f (y) − f (x)
. > r(x) if s(x) < y < x
y−x
and
f (y) − f (x)
. < r(x) if x < y < t (x).
y−x
Then .y = x and .s(x) < y < t (x) imply
. f (y) − f (x) < r(x)(y − x).
Hence we have defined a function .ϕ : A → Q × Q × Q such that .ϕ(x) =

(r(x), s(x), t (x)). We claim that .ϕ is injective.
For the sake of contradiction, assume that there exist points .y = x in A such
that .ϕ(x) = ϕ(y). This implies .(s(y), t (y)) = (s(x), t (x)), and x, y both lie in this
interval. It follows that
f (y) − f (x) < r(x)(y − x)

.
f (x) − f (y) < r(y)(x − y).
But .r(x) = r(y), hence .0 < 0. This contradiction proves that .ϕ is injective, and
thus A is a countable set. The proof that B is also countable is similar.

Definition 16.34 Let E be a subset of .R. A collection .V of closed intervals, each
having positive measure, is a Vitali cover of E if and only if for every .x ∈ E and
for every .ε > 0 there exists an interval .I ∈ V such that .x ∈ I and .λ(I ) < ε.
Roughly speaking, Vitali covers consist of closed interval of arbitrarily small
lengths.
Theorem 16.44 (Vitali’s Covering Theorem) Let .V be a non-empty Vitali cover
of a set .E ⊂ R. Then there exists a pairwise disjoint countable collection .{In }n ⊂ V
such that
" " ∞
##

.λ E ∩ R\ In = 0.
n=1
.λ(E) ∈ R and if .ε > 0, there exists a finite pairwise disjoint collection

If
.I1 , . . . , Ip ⊂ V such that
" " ##

p
λ E∩ R\
. In = 0.
n=1
Proof
First case: .λ(E) ∈ R. We fix an open set V which contains E and such that
.λ(V ) ∈ R. Let
V0 = {I ∈ V | I ⊂ V } .
.
The fact that .V0 is a Vitali cover of E is clear. If .I1 ∈ V0 and .E ⊂ I1 , the proof
is complete. Otherwise we proceed by induction. Assume that .I1 , .I2 , . . . , .In have
been chosen and are pairwise disjoint. If .E ⊂ nk=1 Ik , the proof is complete.
Otherwise we write

n
An =
. Ik
k=1
Un = V ∩ (R \ An ) .
The set .An is closed as a finite union of closed sets, .Un is open, and .Un ∩ E = ∅.
Define
δn = sup {λ(I ) | I ∈ V0 , I ⊂ Un } .
.
Next we select .In+1 ∈ V0 such that .In+1 ⊂ Un and .λ(In+1 ) > δn /2.
If this procedure continuous indefinitely, weget an infinite sequence .{In }n of
pairwise disjoint elements of .V0 . Let .A = ∞ n=1 In , and we claim that .λ(E ∩
(R \ A)) = 0. Indeed, for every positive integer n there exists a unique closed
interval .Jn having the same mid-point as .In and such that .λ(Jn ) = 5λ(In ). Since
" ∞
# ∞ ∞

λ
. Jn ≤ λ(Jn ) = 5 λ(In ) = 5λ(A) ≤ 5λ(V ),
n=1 n=1 n=1
we see that
" ∞
#

. lim λ Jn = 0.
p→+∞
n=p

As a consequence, it suffices to show that .E ∩ (R \ A) ⊂ ∞ n=p Jn for every
.p ∈ N.
Fix .p ∈ N and .x ∈ E ∩ (R \ A). Then .x ∈ E ∩ (R \ Ap ) ⊂ Up , hence there

exists .I ∈ V0 such that .x ∈ I ⊂ Up . Clearly .δn < 2λ(In+1 ), and .λ(In ) → 0 as
.n → +∞. Hence there exists .n ∈ N such that I is not contained in .Un . Call q
the smallest such integer. It is obvious that .p < q, so that
I ∩ Aq = ∅,
. I ∩ Aq−1 = ∅.
This yields .I ∩ Iq = ∅ and, since I is a subset of .Uq−1 , we find .λ(I ) ≤ δq−1 <
2λ(Iq ). Recalling that .λ(Jq ) = 5λ(Iq ), we see that
∞

I ⊂ Jq ⊂
. Jn ,
n=p
∞
hence .x ∈ ∞ n=p Jn . In conclusion .E ∩ (R \ A) ⊂ n=p Jn , and this shows that
.λ(E ∩ (R \ A) = 0. To conclude the first part of the proof, let .ε > 0 be given and
select a positive integer p such that

∞

. λ(In ) < ε.
n=p+1
But
∞

E ∩ R \ Ap ⊂ (E ∩ (R \ A)) ∪
. In ,
n=p+1
hence
⎛ ⎞
∞

.λ E ∩ R \ Ap ≤ 0 + λ⎝ In ⎠ < ε.
n=p+1
Second case: λ(E) = +∞. For every .n ∈ N we introduce .En = E ∩ (n, n + 1)

.
and
Vn = {I ∈ V | I ⊂ (n, n + 1)} .
.
It is a simple exercise to check that .Vn is a Vitali cover of .En . Since .λ(En )
is finite, the first part of the proof applies and yields a finite pairwise disjoint
collection .In ⊂ Vn such that

λ En ∩ R \
. In = 0 for every n ∈ Z.

To conclude, let .I = n∈Z In . Then .I is a countable pairwise disjoint sub-
collection of .V and
" #

.E ∩ R \ I ⊂Z∪ En ∩ (R \ In ) .
n∈Z
Since
∞

λ E∩ R\
. I ≤ λ(Z) + 0 = 0,
n=−∞

A fundamental result of Lebesgue measure theory is the differentiability almost
everywhere of monotone functions.
Theorem 16.45 (Lebesgue) If .f : [a, b] → R is a monotone function on a closed
interval .[a, b], then f has a finite derivative almost everywhere in .[a, b].
Proof Replacing f with .−f , we may assume that f is monotone increasing on
[a, b]. Let us define
.

E = x ∈ [a, b) D+ f (x) < D + f (x) .
.

We claim that .λ(E) = 0. Indeed, we can write .E = {E(u, v) | u ∈ Q, v ∈ Q,
. 0 < u < v}, where

E(u, v) = x ∈ [a, b) D+ f (x) 0. Fix .ε > 0 so small that
α(v − u)
0<ε<
. .
u + 2v
By the regularity properties of the Lebesgue measure, there exists an open set U
such that .E(u, v) ⊂ U and .λ(U ) < α + ε. By definition, to each .x ∈ E(u, v) there
correspond arbitrarily small numbers h such that .[x, x + h] ⊂ U ∩ [a, b] and
f (x + h) − f (x) < uh.

.
The collection of all such intervals .[x, x + h] is a Vitali cover .V of

.E(u, v), and therefore there exists a finite pairwise disjoint sub-collection
.{[xi , xi + hi ] | i = 1, . . . , m} of .V such that
" " ##

m
λ E(u, v) ∩ R \
. [xi , xi + hi ] < ε.
i=1
m
If .V = i=1 (xi , xi + hi ), then .λ(E ∩ (R \ V )) < ε. Since .V ⊂ U , we have

m
. hi = λ(V ) ≤ λ(U ) < α + ε,
i=1
and thus

m
m
. (f (xi + hi ) − f (xi )) vk.

.

As before, there exists a finite pairwise disjoint collection .[yj , yj + kj ]
. | j = 1, . . . , n} such that
⎛ ⎛ ⎞⎞

n
λ ⎝E(u, v) ∩ V ∩ ⎝R \
. [yj , yj + kj ]⎠⎠ < ε.
j =1
Summing up, we see that

n
.α = λ(E(u, v)) ≤ λ (E(u, v) ∩ (R \ V )) + λ (E(u, v) ∩ V ) < ε + ε + kj .
j =1
Next, the previous inequalities yield

n
n
. v(α − 2ε) < v kj < f (yj + kj ) − f (yj ).
j =1 j =1

Recalling that . nj=1 [yj , yj +kj ] ⊂ m
i=1 [xi , xi +hi ], the monotonicity of f implies

m
m
. f (yj + kj ) − f (yj ) ≤ f (xi + hi ) − f (xi ).
j =1 i=1
In conclusion .v(α − 2ε) < u(α + ε), which is a contradiction. We have thus proved
that .λ(E) = 0, so that the right derivative of f exists as a real number at almost
every point of .[a, b]. By the same token, the left derivative of f exists and is finite
at almost every point of .[a, b]. Theorem 16.43 implies that the derivative .f % (x)
exists for almost every .x ∈ [a, b]. To complete the proof, we must show that the set
F of points .x ∈ (a, b) where .f % (x) = +∞ is a set of measure zero.8
Let .β > 0 be given. For every .x ∈ F there exist arbitrarily small numbers h such
that .[x, x + h] ⊂ (a, b) and .f (x + h) − f (x) > βh. We can therefore construct a
countable pairwise disjoint collection .{[xn , xn + hn ] | n = 1, 2, . . .} such that
" " ∞
##

λ F ∩ R\
. [xn , xn + hn ] = 0.
n=1
We derive that
∞
∞

βλ(F ) ≤ β
. hn < f (xn + hn ) − f (xn ) ≤ f (b) − f (a).
n=1 n=1
Since .β > 0 can be arbitrarily large, we see that .λ(F ) = 0. The proof of the theorem
is complete.

Definition 16.35 Let .f : [a, b] → R be a function. We define the total variation of
f over .[a, b] as
)

n

Vab f = sup
. |f (xk ) − f (xk−1 )| a = x0 < x1 < . . . < xn = b .

k=1
The function f is a function of bounded variation over .[a, b] if and only if .Vab f ∈ R.
Remark 16.6 The previous definition can be extended to functions of several vari-
ables, but the language of Measure Theory becomes necessary, and the development
is much more involved. We will not enter into the details in this book.
Exercise 16.15 Prove the identity .Vab f + Vbc f = Vac f for every .a < b < c.
Exercise 16.16 Prove that the function .x → Vax f is non-decreasing on .[a, b].
These exercises inspire the next result.
Theorem 16.46 (Jordan Decomposition Theorem) A function of bounded varia-
tion is the difference of two non-decreasing functions.
Proof If f is of bounded variation over .[a, b], we write

. f (x) = Vax f − Vax f − f (x) .
8 Recall that f is a monotone increasing function, hence the derivative .f % (x) cannot equal .−∞.
Here we set .Vaa f = 0. The function .x → Vax f is non-decreasing. If .x1 < x2 , then

Vax2 f − f (x2 ) − Vax1 f − f (x1 ) = Vxx12 f − (f (x2 ) − f (x1 )) ≥ 0
.
by definition of .Vxx12 f . Hence .x → Vax f − f (x) is also non-decreasing, and the

proof is complete.

We are ready to state the fundamental result of differentiability for functions
of bounded variation. The proof is an immediate consequence of the previous
theorems.
Theorem 16.47 (Lebesgue) A function of bounded variation has finite derivative
almost everywhere.
Let us take a break. We have so far introduced a class of real-valued functions
whose derivative exists at almost every point. But our goal was much more
ambitious: we wanted to characterize functions which satisfy the Fundamental
Theorem of Calculus.
It turns out the road is still long, and we present an interesting tool first.
Theorem 16.48 (Fubini) Suppose .{fn }n is a sequence of monotone functions on

a common interval .[a, b]. If . ∞n=1 n (x) = s(x) exists and is finite for every .x ∈
f
[a, b], then
∞

s % (x) =
. fn% (x) for a.e. x ∈ [a, b].
n=1
Proof We will suppose without loss of generality that all functions .fn are non-
decreasing. Replacing .fn with .fn − fn (a), we also assume that .fn ≥ 0. Thus .s =
∞ %
n=1 fn is non-negative and non-decreasing. Therefore .s exists as a finite number
almost everywhere in .(a, b). Let
sn = f1 + · · · + fn ,
. rn = s − sn .
Each function .fj has finite derivative almost everywhere: there exists a set .A ⊂
(a, b) such that .λ((R \ A) ∩ (a, b)) = 0,
sn% (x) = f1% (x) + · · · + fn% (x) < +∞

.
for every .x ∈ A and every .n ∈ N, and .s % (x) exists in .R for every .x ∈ A.

If .x ∈ (a, b) and .h > 0 is such that .x + h ∈ (a, b), then
s(x + h) − s(x) sn (x + h) − sn (x) rn (x + h) − rn (x)

. = + ,
h h h
hence
sn (x + h) − sn (x) s(x + h) − s(x)
. ≤ .
h h
As .h → 0, this yields .sn% (x) ≤ s % (x) for every .x ∈ A. Since .sn% (x) ≤ sn+1
% (x) is
trivial, we deduce that
.sn% (x) ≤ sn+1

%
(x) ≤ s % (x) for every x ∈ A and n ∈ N.
∞
As a consequence .limn→+∞ sn% (x) = %
j =1 fj (x) exists almost everywhere. It
% %
.limn→+∞ sn (x) = s (x) almost everywhere.
remains to show that
The sequence . sn% (x) n is non-decreasing for every .x ∈ A, so that it suffices to

prove that . sn% n has a subsequence converging to .s % almost everywhere. We select
.n1 < n2 < n3 < · · · in .N such that
∞

. s(b) − snk (b) < +∞.
k=1
.x ∈ (a, b) we have
We remark that for every k and for every .0 ≤ s(x) − snk (x) ≤
s(b) − snk (b). Therefore the series . ∞ k=1 s(x) − snk (x) converges. Since the
terms of this series are monoton functions
with finite
% derivative almost
everywhere,
the same reasoning as above shows that . ∞ k=1 s (x) − s % (x) converges almost
nk
everywhere. It is now clear that .limk→+∞ sn% k (x) = s % (x) almost everywhere in
.(a, b). The proof is complete.

Let us now face the main problem of identifying those functions F of the form
x
F (x) =
. f (t) dt
a
for some .f ∈ L1 (a, b).

Theorem 16.49 If .f ∈ L1 (a, b), we define its integral function F on .[a, b] as
x
F : x →
. f (t) dt.
a
! x .Va F =
The function F is uniformly continuous and of bounded variation b
inta |f (t)| dt. The same conclusion holds if .f ∈ L (R) and .F (x) = −∞ f (t) dt.
b 1
Proof If .x1 < x2 are points of .[a, b], we compute

x2
. |F (x2 ) − F (x1 )| = f (t) dt .
x1
It follows from Proposition 16.3 that F is uniformly continuous. If
a = x0 < x1 < . . . < xn = b,

.
then
n

n
xk n xk b
|F (xk ) − F (xk−1 )| = f (t) dt ≤ |f (t)| dt = |f (t)| dt.
.

k=1 k=1 xk−1 k=1 xk−1 a
This shows that .Vab F ≤ *f *1 , and in particular F is a function of bounded variation.

To prove the reversed inequality, we use the density of simple functions

n
s=
. αk χ[xk−1 ,xk )
k=1
in .L1 (a, b), where .a = x0 < x1 < . . . < xn = b. We consider the function .sign f
defined by
⎧
⎪
⎪
⎨1 if f (x) > 0
. sign f (x) = 0 if f (x) = 0
⎪
⎪
⎩−1 if f (x) < 0.
Let .{sm }m be a sequence of simple functions such that
1
. *sm − sign f *1 ≤ .
m
This inequality is preserved if we replace every .αk such that .|αk | > 1 with
.αk |αk |−1 . In other words, we may always assume that .|sm | ≤ 1 for every m. Since
.sm converges to .sign f in measure, there exists a subsequence .{smj }j such that
.smj → sign f pointwise almost everywhere in .[a, b]. The Dominated Convergence
Theorem implies now that
b b b
. |f (t)| dt = f (t) sign f (t) dt = lim f (t)smj (t) dt.
a a j →+∞ a
But

b n xk

. f (t)smj (t) dt = αk f (t) dt

a k=1 xk−1
n

= αk (F (xk ) − F (xk−1 ))

k=1

n
≤ |αk ||F (xk ) − F (xk−1 )|
k=1

n
≤ |F (xk ) − F (xk−1 )|
k=1
≤ Vab F.
!b
This shows that . a |f (t)| dt ≤ Vab F , and the proof is complete.

Theorem 16.50 If A is a subset of .R, then
λ(A ∩ (x, x + k)) λ(A ∩ (x − h, x))

. lim = lim
k→0+ k h→0+ h
λ(A ∩ (x − h, x + k))
= lim =1
(h,k)→(0,0) h+k
h>0
k>0
for almost every .x ∈ A. If A is a Lebesgue-measurable set, then all the previous

limits are equal to zero for almost every .x ∈ R \ A.
Proof Since we always intersect A with a bounded interval, we may assume without
loss of generality that A is itself a bounded subset. Consider a sequence of bounded
open sets .Un such that
.U1 ⊃ U2 ⊃ · · · ⊃ Un ⊃ · · · ⊃ A
and .λ(Un ) − 2−n < λ(A). We call .a = inf U1 and consider the functions
ϕn (x) = λ(Un ∩ (a, x)),

. ϕ(x) = λ(A ∩ (a, x)).
For every .x ∈ Un and .h > 0 sufficiently small, it is easy to check that
ϕn (x + h) − ϕn (x) ϕn (x) − ϕn (x − h)
. = = 1.
h h
Hence .ϕn% (x) exists at all .x ∈ Un and .ϕn% (x) = 1. We consider the series
(ϕ1 − ϕ) + (ϕ2 − ϕ) + · · · + (ϕn − ϕ) + · · ·

.
We claim that .ϕn − ϕ is monotone for every n. Indeed, if .h > 0, then
ϕn (x + h) − ϕ(x + h) − (ϕn (x) − ϕ(x))

.
= λ(Un ∩ [x, x + h)) − λ(A ∩ (a, x + h)) + λ(A ∩ (a, x))

≥ λ(Un ∩ [x, x + h)) − λ(A ∩ [x, x + h)) ≥ 0,
as a consequence of the inequality
λ(A ∩ (a, x + h)) ≤ λ(A ∩ (a, x)) + λ(A ∩ [x, x + h))

.
and of the inclusion .A ∩ [x, x + h) ⊂ Un ∩ [x, x + h). The claim is proved.

Let .b = sup U1 , so that
1
ϕn (b) − ϕ(b) = λ(Un ) − λ(A) <
. .
2n
For every .x ∈ [a, b] we then have
∞
∞
∞
1
. (ϕn (x) − ϕ(x)) ≤ (ϕn (b) − ϕ(b)) ≤ < +∞.
2n
n=1 n=1 n=1
∞
We may set .s(x) = n=1 (ϕn (x) − ϕ(x)). By Theorem 16.48 we have that
∞
%
s % (x) =
. ϕn (x) − ϕ % (x) ∈ R
n=1
for almost every .x ∈ (a, b), and so also .limn→+∞ ϕn% (x) %
∞= ϕ (x) for almost every
%
.x ∈ (a, b). To summarize, we see that .ϕ = 1 on . n=1 Un except on a set of
measure zero, and the first statement of the theorem is proved.
If A is also Lebesgue-measurable, then
λ(A ∩ (x − h, x + k)) λ((R \ A) ∩ (x − h, x + k))

1=
. +
h+k h+k
= ψA (x) + ψR\A (x).
As .h → 0 and .k → 0 along positive values, the first part of the theorem applied
to .R \ A yields that .ψR\A (x) → 1 for a.e. .x ∈ R \ A. Hence .ψA (x) → 0 for a.e.
.x ∈ R \ A, and the proof is complete.

!
Theorem 16.51 If .f ∈ L1 (a, b) and .F (x) = a f (t) dt, then .F % (x) = f (x) for
x
almost every .x ∈ (a, b).

Proof If .f = χA for some measurable

subset A of .(a, b), then conclusion follows
from Theorem 16.50. Let .s = nk=1 αk χAk be a non-negative simple measurable
function, so that
x
n x
.S(x) = s(t) dt = αk χAk (t) dt.
a k=1 a
Again Theorem 16.50 shows that .S % (x) = s(x) for a.e. .x ∈ (a, b). If .f ∈ L1 (a, b),
there exists a sequence .{sn }n of simple measurable ! x functions such that .sn ≤ sn+1
and .sn (x) → f (x) for all .x ∈ [a, b]. If .Sn (x) = a sn (t) dt, Beppo Levi’s Theorem
yields
x x
F (x) =
. f (t) dt = lim sn (t) dt = lim Sn (x)
a n→+∞ a n→+∞
∞

= S1 (x) + (Sn+1 (x) − Sn (x))
n=1
for every .x ∈ [a, b]. We apply Theorem 16.48 and obtain

∞

%
%
F (x) =
. S1% (x) + Sn+1 (x) − Sn% (x)
n=1
= lim S % (x)
n→+∞ n+1
for a.e. .x ∈ (a, b). Therefore .limn→+∞ Sn% (x) = limn→+∞ sn (x) for a.e. .x ∈ (a, b).
Since .sn → f , we conclude that .F % (x) = f (x) for a.e. .x ∈ (a, b).

The hard question is whether the last theorem can be reversed: if .ϕ is a continuous
function whose derivative exists almost everywhere, is it true that
x
ϕ(x) − ϕ(a) =
. ϕ % (t) dt ?
a
Very technical examples show that this is typically false. The validity of the
Fundamental Theorem of Calculus is however true for a restricted class of functions.
Definition 16.36 Let J be an interval (of any kind, even unbounded), and let
f : J → R be a function. We say that f is absolutely continuous on J if and only
.
if for every .ε > 0 there exists .δ > 0 such that

n
. |f (dk ) − f (ck )| < ε
k=1
npairwise disjoint family .{(ck , dk )}k=1 of open intervals contained in

for every finite n
J such that . k=1 |dk − ck | < δ.

Proposition 16.6 Any absolutely continuous function f defined on .[a, b] has finite
variation on .[a, b].
Proof This is almost trivial: consider .ε = 1 and .δ > 0 according to the definition
of absolute continuity. Fix a positive integer .n > (b − a)/δ and split .[a, b] by points
.a = x0 < . . . < xn = b such that .xk − xk−1 = (b − a)/n < δ for every k. It follows
xk
that .Vxk−1 f ≤ 1 for every k, hence

n
Vab f ≤
. k f ≤ n.
Vxxk−1
k=1

Proposition 16.7 Every absolutely continuous function f on .[a, b] is continuous,
and can be written as the difference .f1 − f2 of two non-decreasing absolutely
continuous functions .f1 and .f2 .
Proof It is obvious that absolutely continuous functions are continuous. We set
f1 (x) = Vax f and .f2 = f1 − f . We need to prove that .f1 is absolutely continuous.
.
Pick .ε > 0 and .δ > 0 so small that

n
ε
. |f (dk ) − f (ck )| <
2
k=1

whenever the pairwise disjoint intervals .(ck , dk ) satisfy . nk=1 (dk − ck ) < δ. Since
f is of bounded variation, each .(ck , dk ) admits a subdivision
ck = a0(k) < a1(k) < . . . < a(k)

.
k
= dk
such that
k −1

(k) (k) ε
d
.Vc k f
k
< f (aj +1 ) − f (aj ) + .
2n
j =0
Hence

n
n k −1
n ε
(k) (k)
. |f1 (dk ) − f1 (ck )| = Vcdkk f < f (aj +1 ) − f (aj ) +
2
k=1 k=1 k=1 j =0
ε ε
< + = ε,
2 2

Theorem 16.52 if f is a non-decreasing function defined on .[a, b], then .f % is

Lebesgue-measurable and
b
. f % (x) dx ≤ f (b) − f (a).
a
If g is a real-valued function of bounded variation on .[a, b], then .g % ∈ L1 ([a, b]).

Proof We extend f by setting .f (x) = f (b) when .x > b. For .n ∈ N and .a ≤ x ≤ b,
we define

f x + n1 − f (x)
.fn (x) = .
1
n
Then .{fn }n is a sequence of non-negative measurable functions such that .fn (x) →
f % (x) for a.e. .x ∈ (a, b). In particular .f % is measurable. We apply Fatou’s Lemma:
b b b
. f % (x) dx = lim fn (x) dx ≤ lim inf fn (x) dx
a a n→+∞ n→+∞ a
b 1
= lim inf n f x+ − f (x) dx
n→+∞ a n
" #
b+ n1 a+ n1
= lim inf n f (x) dx − n f (x) dx
n→+∞ b a
" #
b+ n1 a+ n1
≤ lim inf n f (b) dx − n f (a) dx
n→+∞ b a
= f (b) − f (a).
To prove the second statement, we decompose g as a linear combination of two

non-decreasing functions, and we apply to each one the first statement. The proof is
complete.

We can now prove a generalization of the basic principle for differentiable functions.
Theorem 16.53 Suppose that .f : [a, b] → R is an absolutely continuous function.
If .f % = 0 a.e. on .[a, b], then f is a constant function.
Proof Let .c ∈ (a, b] be an arbitrary point; we claim that .f (c) = f (a). To prove
this claim, we pick any .ε > 0. By absolute
continuity,
there exists .δ > 0 as in
Definition 16.36. If .E = x ∈ (a, c) f % (x) = 0 , by assumption .λ(E) = c − a.
Hence to every .x ∈ E there correspond arbitrarily small values .h > 0 such that
.[x, x + h] ⊂ (a, c) and
h
. |f (x + h) − f (x)| < ε.
c−a
Since the collection of all such intervals .[x, x + h] is a Vitali cover of E, there exists
a finite pairwise disjoint collection .{[xk , xk + hk ] | k = 1, . . . , n} such that
" " ##

n
. λ E∩ R\ [xk , xk + hk ] < δ.
k=1

Hence .λ((a, c)) = λ(E) < δ + nk=1 hk . Assuming that .x1 < x2 < · · · < xn , it
follows that the sum of the lengths of the intervals
(a, x1 ), (x1 + h1 , x2 ), . . . , (xn + hn , c)

.
is smaller than .δ; our choice of .δ yields now

n−1
|f (a) − f (x1 )| +
. |f (xk + hk ) − f (xk )| + |f (xn + hn ) − f (c)| < ε.
k=1
We now conclude by the triangle inequality:

n−1
|f (a) − f (c)| ≤ |f (a) − f (x1 )| +
. |f (xk + hk ) − f (xk )|
k=1

n
+ |f (xn + hn ) − f (c)| + |f (xk + hk ) − f (xk )|
k=1

n
hk
<ε+ ε ≤ 2ε.
c−a
k=1
Since .ε > 0 is arbitrary, we conclude that .f (a) = f (c), and f is constant on .[a, b].

Theorem 16.54 (Fundamental Theorem of Calculus) If .f : [a, b] → R is an
absolutely continuous function, then .f ∈ L1 ([a, b]) and
x
f (x) = f (a) +
. f (t) dt
a
for every .x ∈ [a, b].

Proof We know that f is a function of bounded variation, hence .f % exists a.e. in
!b % %
.[a, b] and .
a f (x) dx ≤ f (b) − f (a). Hence .f ∈ L ([a, b]). Setting .g(x) =
1
!x % %
a f (t) dt, we see that g is absolutely continuous and .g = f almost every by
Theorem 16.51. It follows that .h = f − g is absolutely continuous and .h% = 0

almost everywhere, hence h is a constant function. Thus
x x
f (x) = h(x) + g(x) = h(a) +
. f % (t) dt = f (a) + f % (t) dt
a a
for every .x ∈ [a, b]. The proof is complete.

Since it is an easy exercise to prove that the indefinite integral of a function in
L1 ([a, b]) is absolutely continuous, we can summarize our results in the following
.
statement.
Theorem 16.55 A function .f : [a, b] → R has the form
x
f (x) = f (a) +
. ϕ(t) dt
a
for some .ϕ ∈ L1 ([a, b]) if and only if f is absolutely continuous on .[a, b]. In this
case there results .ϕ = f % almost everywhere in .[a, b].
!x
Exercise 16.17 Prove that a function .f : R → R has the form .f (x) = −∞ ϕ(t) dt
for some .ϕ ∈ L1 (R) if and only if f is absolutely continuous on .[−A, A] for every
+∞
.A > 0, .V−∞ f ∈ R and .limx→−∞ f (x) = 0. To this aim, proceed as follows.
!x
(1) If f has the form .f (x) = −∞ ϕ(t) dt, prove that f is absolutely continuous
+∞ !
on every interval .[−A, A] and that .V−∞ f = R ϕ(t) dt. Use the Dominated
Convergence Theorem to show that .f (x) → 0 as .x → −∞.
(2) Conversely,
!x % apply the previous theorem to deduce that .f (x) = f (−A) +
f (t) dt for every .A > 0 and every .x > −A.
−A !x
(3) Let .A → +∞ and deduce that .f (x) = limA→+∞ −A f % (t) dt.
! +∞ % !n
(4) Prove that . −∞ |f (t)| dt = limn→+∞ −n |f % (t)| dt = limn→+∞ V−n n f ≤
+∞
V−∞ f ∈ R. !x
(5) Deduce that .f % ∈ L1 (R), so that (3) gives .f (x) = −∞ f % (t) dt.
16.11 Problems
∞ Let {En }n be a sequence of measurable sets in a measurable space X such that

16.1
n=1 µ(En ) converges. Prove that almost every x ∈ X lie in at most finitely many
of the sets En .
Hint:
show that the set A of all x which lie in infinitely many En
coincides with ∞ n=1
∞
k=n Ek .
16.2 Suppose fn : X → [0, +∞] is measurable for every n, that f1 ≥! f2 ≥ f3 ≥
. . . and that fn → f for every x ∈ X. If f1 ∈ L1 (X), prove that X fn dµ →
!
X f dµ. Is the assumption that f1 ∈ L (X) really necessary?
1
References 469
16.3 In a metric setting, the proof of Urysohn’s Lemma is much easier. Let (X, d)
be a metric space. For every non-empty subset E of X we define for all x ∈ X
dE (x) = inf {d(x, y) | y ∈ E} .

.
Prove that dE is uniformly continuous on X. Now let A and B be disjoint non-empty

closed subsets, and derive Urysohn’s Lemma from the function
dA (x)
f (x) =
. .
dA (x) + dB (x)
16.4 Prove that Theorem 16.18 extends to the situation in which the sequence {fn }n
is replaced by a family {ft | t ∈ R} such that (i) ft (x) → f (x) as t → +∞, for
every x ∈ X; (b) t → ft (x) is continuous, for every x ∈ X.
!
16.5 Let u ∈ L1loc () be such that u(x)v(x) dx = 0 for every v ∈ D(). Prove
that u = 0 almost everywhere on . Hint: observe that !n ∗ u = 0 for every n.
16.12 Comments
As we said, Measure Theory is usually introduced as the study of certain functions

defined on suitable collections of sets, called .σ -algebras. In this book we did not
consider signed measures or complex measures, for which we refer to [3] or [2].
Algebras and .σ -algebras of sets are the natural families of measurable sets, but
sometimes it is natural to introduce “measures” which are defined on every subset
of a gives set. For instance one might define the “measure” of an arbitrary subset E
of .R by covering E with sequences of pair-wise disjoint intervals .Ik , summing the
lengths of all the .Ik , and taking the infimum over all such coverings of E. This leads
to the concept of outer measures, and measurable sets a those sets which satisfy a
particular condition introduced by C. Carathéodory. The interested reader is referred
to [1].
References
1. L.C. Evans, R.F. Gariepy, Measure Theory and Fine Properties of Functions. Textbooks in
Mathematics Series Profile, 2nd edn. (CRC Press, Boca Raton, 2015)
2. G.B. Folland, Real Analysis: Modern Techniques and Their Applications. Pure and Applied
Mathematics. A Wiley-Interscience Series of Texts, Monographs, and Tracts, 2nd edn. (Wiley,
New York, 1999)
3. W. Rudin, Real and Complex Analysis. McGraw-Hill Series in Higher Mathematics, xi, 412 p.
(McGraw-Hill Book Company, New York, 1966)

Secchi S A Circleline Study of Mathematical Analysis

Uploaded by

Copyright:

Available Formats

Secchi S A Circleline Study of Mathematical Analysis

Uploaded by

Document Information

Original Title

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

Secchi S A Circleline Study of Mathematical Analysis

Uploaded by

Copyright:

Available Formats

UNITEXT 141

La Matematica per il 3+2

ISSN 2038-5714 ISSN 2532-3318 (electronic)

My first encounter with Mathematical Analysis was on the celebrated Principles

Fig. 1 The tree of this book

that completely characterize the topology of any space. Convergence is usually

Milano, Italy Simone Secchi

Part I First Half of the Journey

11.5 Polynomial Approximation.. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 151

Part II Second Half of the Journey

14 Differentiating Again: Linearization in Normed Spaces . . . . . . . . . . . . . . . 295

16.6 Measure, Topology, and the Concrete Lebesgue Measure . . . . . . . . 423

Abstract Mathematics is based on the language of proposition logic: every state-

1.1 The Propositional Calculus

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 3

Definition 1.2 (Conjunction) If A and B are sentences, then their conjunction .A ∧

Remark 1.2 Logical implication may be written in different ways: .A ⊃ B was

Definition 1.5 (Logical Equivalence) If A and B are sentences, the sentence

Logical equivalence is also denoted by .A ≡ B.

As we said before, sentences may contain one or more free variables.

Analysis, or is i a free variable?

Important: .∃ vs. .∃!

The syntax of sentences with quantifiers is not completely universal. The

If we are given a sentence .A(x1, . . . , xn ) defined by n free variables and one or

We begin this book in the worst possible manner: we introduce a meaningless

Important: Sets Remain Undefined

We write .x ∈ X to mean that x is an element of the set X, and we say that x is

.X = {x | some proposition about x}.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 7

= {n ∈ N | n is neither odd nor even}

In this situation we write .A ⊂ B or .B ⊃ A. A set A is a proper subset of B if

Important: Proper Inclusion

Two sets A and B are disjoint if .A ∩ B = ∅.

Fig. 2.1 Intersection of two

Fig. 2.2 Union of two sets

Fig. 2.3 Difference of two \

each .α ∈ A, so that in particular .x ∈ Xβ for each .β ∈ B. Thus .x ∈ {Xβ | β ∈

(x, y) = {{x}, {x, y}}.

If R is a relation, we usually write .xRy instead of the more formal .(x, y) ∈ R,

Fig. 2.4 A cartesian product

or equivalently .yR −1 x if and only if xRy.

R ◦ S = {(x, z) | ∃y((x, y) ∈ S ∧ (y, z) ∈ R)} .

Definition 2.11 An equivalence relation is a reflexive, symmetric and transitive

A function .f : X → Y is injective if distinct points of X have distinct images in Y .

If A is a set and f is a function, the set

f (A) = {y | ∃x(x ∈ A ∧ f (x) = y)} = {f (x) | x ∈ A}

Fig. 2.5 Intuition of a function

f −1 (B) = {x | ∃y(x ∈ B ∧ f (x) = y)}

the Cartesian product of f and g. Prove the following statements:

Important: Sets or Subsets?

We have presented a quick survey of Set Theory from a non-axiomatic viewpoint.

often written as .{x ∈ U | P (x)}, is admissible, since it defines a subset of a (given)

1. P.R. Halmos, Naive Set Theory (Dover Publications, 2017)

3.1 The Axioms of R

The system of real numbers is a set R, together with a distinguished subset P

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2022 17

Axiom 1 For each x ∈ R and y ∈ R there results x +y = y +x and xy = yx.

Axiom 2 For each x, y and z in R, there results x + (y + z) = (x + y) + z

Axiom 3 For each x, y and z in R, there results x(y + z) = xy + xz.

Axiom 8 Let A and B be subsets of R such that A = ∅, B = ∅, A ∩ B = ∅,

since I is arbitrary we get .x + 1 ∈ N.

follows with .z = a/b.

(iv) .|z| ≤ |z|, .| z| ≤ |z|;

We conclude again by the positivity of the modulus.

We conclude that .AB ≥ |C|2 , since .B ≥ 0.

Every complex number .z = (x, y) = x + iy = (0, 0) can be written in the form

(b) If .x = 0 and .y = 0, then

is called the trigonometric form of the complex number .z = 0. The number .ϑ ∈