Compact Textbooks in Mathematics

Martin Brokate
Götz Kersting

Measure
and
Integral
Compact Textbooks in Mathematics
Compact Textbooks in Mathematics

This textbook series presents concise introductions to current topics in math-

ematics and mainly addresses advanced undergraduates and master students.
The concept is to offer small books covering subject matter equivalent to 2- or
3-hour lectures or seminars which are also suitable for self-study. The books pro-
vide students and teachers with new perspectives and novel approaches. They
feature examples and exercises to illustrate key concepts and applications of the
theoretical contents. The series also includes textbooks specifically speaking to
the needs of students from other disciplines such as physics, computer science,
engineering, life sciences, finance.

More information about this series at

http://www.springer.com/series/11225
Martin Brokate
Götz Kersting

Measure and Integral

Martin Brokate Götz Kersting
TU München Institut für Mathematik
Garching, Germany Goethe-Universität Frankfurt
Frankfurt am Main, Germany

Revised and updated translation from the German language edition: Maß und Integral by
Martin Brokate and Götz Kersting, © Springer Basel AG 2011.

ISSN 2296-4568 ISSN 2296-455X (electronic)

Compact Textbooks in Mathematics
ISBN 978-3-319-15364-3 ISBN 978-3-319-15365-0 (eBook)
DOI 10.1007/978-3-319-15365-0

Library of Congress Control Number: 2015941155

2010 Mathematics Subject Classification: 28-XX, 28-01

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made.

Cover design: deblik, Berlin

Printed on acid-free paper

Springer International Publishing AG Switzerland is part of Springer Science+Business Media

(www.springer.com)
Preface

This text is a translation of the German edition. It closely follows the original; some
errors and misprints were corrected.

München and Frankfurt/Main Martin Brokate

January 2015 Götz Kersting

Preface to the German Edition

Modern measure and integration theory is a prominent descendant of Cantor’s set

theory, and it played an important role for the formation of the latter. The roots
of measure and integration theory thus are found in areas usually attributed to
pure mathematics. Nevertheless, it has gained importance particularly for areas
of mathematics strongly linked to applications—for functional analysis, partial
differential equations, applied analysis and control theory, numerical mathematics,
potential theory, ergodic theory, probability theory, and statistics. Measure and
integration theory thus cannot be subsumed so easily under the paradigm pure versus
applied mathematics (a paradigm which nowadays tends to become less and less
persuasive anyway).
It is under this view that we have written our textbook. Indeed we have in mind
readers who want to utilize the theory elsewhere and are interested in a concise
exposition of the most important results. At the same time, we aim at presenting
measure and integration theory as a coherent and transparent system of assertions
on areas, volumes, and integrals. We think that this can be done in a compact manner
so that it can be integrated into a standard bachelor’s curriculum in mathematics.
From the standpoint of mathematics, the core of measure and integration theory
has largely reached its final form. Nevertheless, we think that concerning its
presentation, there is still room for accentuation. Our arrangement of the content
does not follow the format chosen by other authors. Here are some explanations.
We do not start with the existence and uniqueness theorems for measures. We
believe that such an approach better fits the needs of the students: Initially, the
convergence results for integrals are important; the construction of measures—
however nicely it works out following Carathéodory—may be postponed for the

v
vi Preface

start. For this reason we treat these constructions only at the end of our textbook
(which does not prevent a lecturer from reorganizing the material, of course). There
we have opted for a presentation which directly leads to the goal, avoiding the
usual discussions of set systems like algebras of sets, semi-rings, etc. At some other
places, too, there are new features.
We do not intend to present the theory in all its ramifications. We concentrate
on the core (as we understand it) and, beyond that, display results which provide
links to other areas of mathematics. Regarding analysis, this pertains, e.g., to
the smoothing of functions by convolution as well as Jacobi’s transformation
formula. Concerning geometric measure theory, we discuss the Hausdorff measure
and dimension. For probability theory, among other things, we treat kernels and
measures on infinite products following Kolmogorov. With the final two chapters,
we try to exhibit some connections to functional analysis which we find useful for
understanding measure and integration theory. To guide the reader, we have marked
some sections with an asterisk (); they may be skipped at first reading.
As a prerequisite, we assume knowledge of the contents of the first-year bachelor
courses in mathematics (as they are typically given in our home country). From
topology, without comment we only use elementary concepts (open, closed, com-
pact, neighborhood, continuity) in the setting of metric spaces. Anything exceeding
that, we discuss by some means or other. Historical notes are found in footnotes.
A concise text as the one we aimed at cannot substitute any comprehensive expo-
sition. We therefore do not intend to replace established textbooks like Elstrodt’s [2],
much less classical texts like those of Halmos [4] or Bauer [1]. In the appendix we
mention these and other introductions to the theory. From all of them, we have
benefitted a lot; we take the liberty not to document this in detail, as should be
permitted in a textbook. We gladly have incorporated suggestions for the text as
well as corrections due to Christian Böinghoff and Henning Sulzbach. We thank
Birkhäuser for the pleasant and smooth collaboration.

München and Frankfurt/Main Martin Brokate

March 2010 Götz Kersting
Contents

1 Introduction .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 1
2 Measurability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 7
3 Measures .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 19
4 The Integral of Nonnegative Functions . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 29
5 Integrable Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 41
6 Convergence .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 53
7 Uniqueness and Regularity of Measures . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 63
8 Multiple Integrals and Product Measures . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 73
9 Absolute Continuity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 87
10 The Jacobi Transformation Formula .. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 109
11 Construction of Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 117
12 Hilbert Spaces .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 137
13 Banach Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 153

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 169

Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 171

vii
Introduction
1

To determine specific contents of area and volume as well as integrals is a very

old theme in mathematics. Unsurpassed are the achievements of Archimedes, in
particular his computation of the volume of the unit ball as 4 =3 and of the area of
the
R 1 unit sphere as 4 . Starting from Euler, problems like determining the value of
0
sin x
x dx (which is  =2) have kept the analysts busy.
To the end of the nineteenth century, this subject became less and less important,
as there was not much left to be discovered. At that moment, measure and integration
theory entered the stage. It, too, deals with contents or (as we will call it in the
following) measures of sets, as well as with integrals of functions, but the question
has changed. It no longer reads “what is the measure of this or that set?” but rather
“which sets can be measured, which functions can be integrated?”. To which sets
one thus can assign a measure, to which functions an integral? Their specific value
becomes secondary, general rules of integration come to the fore. The relation to
differential calculus, which for a long period since Newton and Leibniz was in the
foreground, loses its dominant role.
Such a change of perspective is not uncommon in mathematics. In our case, it
arose in the context that one no longer considered integrals on their own, but rather
needed them as tools in other mathematical investigations. Historically one should
mention in particular the Fourier analysis of functions, the decomposition of real-
valued functions into sinusoidal oscillations. Their coefficients (amplitudes) can be
expressed by certain integrals—soon, one realized that for this purpose one needed
properties of integration which could not be provided by the notions of integrals
being available at that time.
Measure and integration theory according to Lebesgue arose by and large
between the years 1900 and 1915, based on essential preliminary work of Borel1

1
ÉMILE BOREL, 1871–1956, born in Saint-Affrique, active in Paris at the École Normale
Supérieure and the Sorbonne. He significantly contributed not only to the foundations of measure
theory, but also to complex analysis, set theory, probability theory, and to applications of

© Springer International Publishing Switzerland 2015 1

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_1
2 1 Introduction

from 1894. Right from the start, the pioneers during that time directed their attention
towards the fundamental properties of measure and integral. Borel was the first
to demand that measures should be not only additive, but also ¢-additive. This
means that not only for finitely many disjoint measurable sets B1 ; B2 ; : : :  Rd
with measures œ.B1 /; œ.B2 /; : : : the union B D B1 [ B2 [    is measurable and has
measure

œ.B/ D œ.B1 / C œ.B2 / C    ;

but that moreover this property holds for every infinite sequence B1 ; B2 ; : : : of
disjoint measurable sets. Borel realized that only under this assumption a fertile
mathematical theory arises. Particular cases like the circle in the figure

of course do not yield anything new. Lebesgue,2 the founder of modern integration
theory, in his fundamental treatise on integration from the year 1901 started from
six properties that integrals should reasonably satisfy.
Measure and integration theory is based on set theory and cannot dispense with
its ways of reasoning. Only with the aid of set theory a path was found leading to
the full system of measurable subsets of Rd and of other spaces. Yet this approach
is comparatively abstract and indirect. To realize that it is justified, for a start it is
perhaps appropriate to take a look at other more descriptive approaches, even though
they finally were not conclusive.

mathematics. He combined this work with a political career as member of the parliament, minister
of the navy, and finally member of the Résistance.
2
HENRI LEBESGUE, 1875–1941, born in Beauvais, active in Paris at the Sorbonne and the Collège
de France. His foundation of integration theory is a landmark in mathematics, he could resort to
preliminary work of Borel and Baire. With his methods he then obtained results on Fourier series.
1 Introduction 3

LetSus look at the approach due to Jordan.3 His idea is intuitively appealing: Let
V D kjD1 Ij be a union of finitely many disjoint d-dimensional intervals Ij  Rd ,
thus Ij D Œaj1 ; bj1 /      Œajd ; bjd / (it turns out to be useful, though not strictly
necessary, to work with semi-open intervals). One obtains its measure œ.V/ by
adding the products of the edge lengths of the individual intervals:

X
k
œ.V/ WD .bj1  aj1 /    .bjd  ajd / :
jD1

Following Jordan, the exterior and the interior measure of a subset B  Rd result
from covering resp. exhausting B by a union of intervals:

Expressed in formulas,

œ .B/ WD inffœ.V/ W V  Bg ; œ .B/ WD supfœ.V/ W V  Bg :

If both expressions have the same value, then the set B is called a Jordan set, and
œ.B/ WD œ .B/ D œ .B/ is called the Jordan measure of B. This definition is
analogous to that of the Riemann integral of a function.
Without a doubt, this approach assigns to a Jordan set its “correct” measure.
The deficiency of this approach lies elsewhere, on the structural level. Indeed, finite
unions, finite intersections, and complements of Jordan sets are again Jordan sets.
But it turns out that, in general, a countable union of Jordan sets is not necessarily
a Jordan set. One easily sees, for example, that every set which consists of just a
single element is a Jordan set of measure 0, while the set of rational numbers in
Œ0; 1 is not a Jordan set (its inner and outer measures are 0 resp. 1). The ¢-additivity
is lacking.
This deficiency is fatal. All attempts to modify Jordan’s definition in order to
remove this deficiency have failed.
But perhaps it is not really necessary to define measurability of sets through an
explicit construction. Is it maybe possible to assign a measure to each subset of

3
CAMILLE JORDAN , 1838–1922, born in Lyon, active in Paris at the École Polytechnique and
the Collège de France. Better known than his contributions to measure theory is his work on
group theory. The Jordan normal form of matrices as well as Jordan curves demonstrate his wide
mathematical interests.
4 1 Introduction

Rd in a reasonable manner, no matter whether in a direct or an indirect fashion?

Already Lebesgue posed that question. The answer is negative, as was discovered
by Vitali4 and Hausdorff.5 Later, Hausdorff’s result was extended by Banach6 and
Tarski.7 It is somewhat perplexing and thus nowadays known as the Banach-Tarski
paradox. These two mathematicians proved in 1924: Any two bounded subsets B
and B0 of Rd , d  3, with nonempty interior, for example two balls of different radii,
can be decomposed into an equal number of disjoint subsets B D C1 [    [ Ck
and B0 D C01 [    [ C0k such that all the parts C1 ; : : : ; Ck ; C01 ; : : : ; C0k are pairwise
congruent, that is, they can be transformed into each other by translations, rotations
and reflexions. One then is inclined to conclude that all parts have the same measure
due to congruency, and therefore B and B0 have the same measure by virtue of
additivity. This would be paradoxical. How can one realise such decompositions?
Intuitively this is inconceivable.
The answer is the following: The theorem of Banach-Tarski is a result of set
theory, and set theory (in particular, when the axiom of choice is employed) admits
the formation of rather exotic subsets of Rd which are no longer accessible through
imagination. This is the meaning of the theorem: The system of all subsets of Rd
is so extensive that it is impossible to assign measures to every subset such that
they are invariant under congruency as well as additive. Therefore, the conclusion
mentioned above cannot be drawn. Thus the paradox dissolves. These results due to
Vitali, Hausdorff, Banach and Tarski are significant in the history of measure theory;
nowadays they rather are a special theme.
Let us record: Attempting to view measurable subsets as single items does not lead
to a sound mathematical theory. Therefore, we no longer look at individual subsets,
but focus instead on systems B of measurable subsets. Their properties are simple.
Following Borel, two properties are indispensable:
[
B 2 B ) Bc 2 B and B1 ; B2 ; : : : 2 B ) Bn 2 B
n1

4
GIUSEPPE VITALI , 1875–1932, born in Ravenna, active in Modena, Padova and Bologna. He
provided distinguished contributions to measure theory, but also to complex analysis.
5
FELIX HAUSDORFF, 1868–1942, born in Breslau, active in Leipzig, Greifswald, and Bonn.
Hausdorff made fundamental contributions to set theory, topology, and measure theory. His
monograph on set theory had enormous influence. Under the alias Paul Mongré he published
essayistic and literary works. Due to his Jewish origin, Hausdorff was forced to retire in 1935.
To escape deportation he took his own life in 1942.
6
STEFAN BANACH , 1892–1945, born in Krakow, active in Lemberg. He established modern
functional analysis. The Lemberg school of mathematicians formed around him and Hugo
Steinhaus.
7
ALFRED TARSKI , 1902–1983, born in Warsaw, active in Warsaw and Berkeley. He is regarded
as one of the most famous logicians due to, for instance, his papers on model theory. He also
contributed to set theory, measure theory, algebra, and topology. Because of his Jewish origin,
after the German invasion of Poland he remained in the United States.
1 Introduction 5

must hold for the complement Bc of B and for finite as well as infinite sequences
B1 ; B2 ; : : : Such systems of sets are of fundamental importance in measure theory;
following Hausdorff, they are called ¢-algebras. Now the task arises to exhibit a
¢-algebra which is large enough and is such that ¢-additivity holds when assigning
a measure to its elements.
This task can be tackled in different ways. One possibility is to start from a
system E of sets to which a measure can be assigned in an obvious manner. For this,
the system of all (semi-open) intervals of Rd qualifies. One then enlarges E to the
system E 0 of all countable unions of sets from E together with the complements of
those unions. Using ¢-additivity, a measure can be assigned to all elements of E 0 . If
E 0 is not yet a ¢-algebra, one repeats this step until a ¢-algebra B d has emerged. This
path can be (and initially has been) entered, however it turns out that uncountably
many steps are required to attain the goal. This not only stresses our intuition, but
moreover one has to utilize advanced methods of set theory, namely, the theory
of well-ordered sets and transfinite induction. No view emerges of how a typical
measurable set looks like.
Fortunately, an elementary and much simpler approach was found soon: one
directly focuses on B d by characterizing it as the smallest ¢-algebra which contains
E. It is called the Borel ¢-algebra, and its elements B  Rd are called Borel sets.
We will see how one assigns a measure to every Borel set so that ¢-additivity holds,
and how an integration theory is established whose rules are transparent and easy to
apply.
One has to pay a price: in order to smoothly manipulate measurable sets and
integrable functions one also has to deal with sets and functions, which in no
way conform to classical perceptions. Back then, leading mathematicians faced this
development in a reserved or even hostile manner, Hermite,8 for example, spoke
about the “deplorable plague” of functions not possessing derivatives. Nevertheless,
the ideas of Borel and Lebesgue prevailed. Their theory is one of the most important
accomplishments of set theory.
As individual elements, measurable sets can hardly be controlled, one gets hold
of them only through their affiliation to systems of sets. This also means that nobody
can say how a “typical” Borel set looks like. In contrast, one may imagine of a
typical Jordan set as the above figure suggests. Nevertheless, in the following we
will no longer bring up Jordan sets, while Borel sets will remain in the focus of
our considerations. In measure and integration theory one has to get used to operate
with systems of sets and of functions, not with individual sets and functions.
Since its emergence, during the age of Newton and Leibniz, the integral has
evolved into a fundamental tool to be employed in many areas within and outside
of mathematics. Among them are the description of processes taking place in
the continuum—e.g. the space-time continuum—in the corresponding areas of

8
CHARLES HERMITE, 1822–1901, born in Dieuze, active in Paris at the École Polytechnique and at
the Sorbonne. He significantly contributed to algebra and number theory, orthogonal polynomials,
and elliptic functions.
6 1 Introduction

(mathematical) analysis, the description of random phenomena in probability

theory, as well as the description of algorithms for computer approximation and
simulation of such processes in numerical mathematics and scientific computing.
In all those contexts the Lebesgue integral has turned out to be the most
adequate notion of an integral. Concerning analysis and numerical mathematics,
the main reason is that the functions whose p-th power possesses a Lebesgue
integral form a complete space (that is, every Cauchy sequence converges) with
respect to the integral norm. In the case p D 2 the integral moreover yields a
scalar product, and we obtain a Hilbert space. These spaces, called Lp spaces,
and their descendants—for example, the Sobolev spaces—provide the predominant
mathematical framework for problems in the continuum.
While Lebesgue integration theory does not concern itself with the computation
of specific integrals, some of its results nevertheless assist this purpose. The results
pertaining to the interchange of integrals and limits (on monotone and dominated
convergence) have manifold applications, for example they clarify under which
circumstances derivatives and integrals can be interchanged. Analogously, this is
true for the theorems of Fubini9 and Tonelli10 concerning interchanging the order
of integration for multiple integrals. Some specific important integrals will be dealt
with in the text.

9
GUIDO FUBINI , 1879–1943, born in Venice, active in Catania, Turin, and Princeton. He worked
on real analysis, differential geometry, and complex analysis. 1939 he emigrated to the USA after
he had lost his chair in Turin in the course of the antisemitic politics under Mussolini.
10
LEONIDA TONELLI , 1885–1946, born in Gallipoli near Lecce, active in Cagliari, Parma,
Bologna, and Pisa. He worked in many areas of analysis and is known mainly for his contributions
to the calculus of variations.
Measurability
2

In this chapter we introduce measurable sets and measurable functions. As explained

in the introduction, the objects we operate with are mainly systems of sets, and not
individual sets. In doing so, there will arise finite as well as infinite sequences of
sets. In both cases and, regardless
S of their length, we denote such sequences as
A1 ; A2 ; : : :, their union as n1 An , and so on.

Definition
A system A of subsets of a nonempty set S with the properties

(i) S 2 A ,
(ii) A 2 A ) Ac WD S nSA 2 A ,
(iii) A1 ; A2 ; : : : 2 A ) n1 An 2 A ,

is called a ¢-algebra or a ¢-field in S. The pair .S; A/ is called a measurable

space. The elements of A are termed measurable subsets of S.

As a consequence,

(iv) ¿ D Sc 2 A , T S
(v) A1 ; A2 ; : : : 2 A ) n1 An D . n1 An / 2 A ,
c c

(vi) A1 ; A2 2 A ) A1 n A2 WD A1 \ A2 2 A ,
c

(vii) A1 ; A2 2 A ) A1 A2 WD .A1 [ A2 / n .A1 \ A2 / 2 A .

Definition
Let .S; A/, .S0 ; A0 / be measurable spaces. A mapping ® W S ! S0 is called
measurable, more precisely A-A0 -measurable, if preimages of measurable sets
are themselves measurable, that is, if

®1 .A0 / 2 A for all A0 2 A0 :

© Springer International Publishing Switzerland 2015 7

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_2
8 2 Measurability

When dealing with a measurable space based on a set S, the choice of the ¢-algebra
A will usually be obvious, and thus it will be clear which subsets of S are
measurable. Therefore, in the following we will not always specify A explicitly.

Example (Trace ¢-Algebra)

Given a measurable subset S1 in a measurable space .S; A/, the system A1 WD fA  S1 W A 2 Ag
becomes a ¢-algebra on S1 . It is called the trace ¢-algebra of A on S1 , or alternatively the induced
¢-algebra on S1 . A mapping ® W S ! S0 is A-A0 -measurable if and only if the restrictions of ® on
S1 and on S2 WD Sc1 are measurable with respect to the trace ¢-algebras A1 and A2 , thanks to the
formula

®1 .A0 / D .®1 .A0 / \ S1 / [ .®1 .A0 / \ S2 / :

Proposition 2.1 (Composition of measurable mappings). Let .S; A), .S0 ; A0 /

and .S00 ; A00 / be measurable spaces, ® W S ! S0 be a A-A0 -measurable mapping,
and § W S0 ! S00 be a A0 -A00 -measurable mapping. Then § ı ® W S ! S00 is a
A-A00 -measurable mapping.

Proof. For any measurable subset A00 of S00 , the set A0 WD §1 .A00 / is measurable in
S0 by assumption, and therefore .§ ı ®/1 .A00 / D ®1 .A0 / is measurable in S. u t

Generators of ¢-Algebras, Borel ¢-Algebras

When the set S is countable, the usual choice of the ¢-algebra is the power set, the
set of all subsets of S. For uncountable sets S, however, this approach has turned out
to be unsuitable. Instead, in that case one specifies a ¢-algebra by a generator.

Definition
A system E of subsets of S is called a generator of the ¢-algebra A in S, if A is
the smallest ¢-algebra in S which contains E (that is, if for every ¢-algebra AQ on
S with AQ  E we also have AQ  A). A is called the ¢-algebra generated by E
and is denoted by A D ¢.E/.

Every system of subsets generates a ¢-algebra.

Proposition 2.2 (Generated ¢-algebras). For every system E of subsets in S

there is a smallest ¢-algebra which contains E. We obtain it as the intersection of
all ¢-algebras containing E:

¢.E/ D fA  S W A 2 AQ for every ¢-algebra AQ in S with AQ  Eg :

Generators of ¢-Algebras, Borel ¢-Algebras 9

Proof. The system of all ¢-algebras containing E is nonempty, since the system of
all subsets of S belongs to it. The intersection A of all those ¢-algebras is itself a ¢-
algebra. Indeed, A 2 A means that A 2 AQ for all ¢-algebras AQ  E. It follows that
Ac 2 AQ for all AQ  E and therefore Ac 2 A. The other properties of a ¢-algebra
are derived analogously. Moreover, we have A  E as well as A  AQ for every
¢-algebra AQ  E. The proposition is proved. t
u

When working with generated ¢-algebras the following statements are used rou-
tinely.

Proposition 2.3 (Equality of ¢-algebras). Let E1 and E2 be generators of the

¢-algebras A1 and A2 in S. Then A1 D A2 holds if E1  A2 and E2  A1 .

Proof. From E1  A2 we conclude that A1  A2 , and vice versa. t

u

Proposition 2.4 (Measurability criterion). Let .S; A/, .S0 ; A0 / be measurable

spaces, and let E 0 generate A0 . Then ® W S ! S0 is an A-A0 -measurable mapping,
if

®1 .A0 / 2 A for all A0 2 E 0 :

˚ 
Proof. AQ WD A0 2 A0 W ®1 .A0 / 2 A is a ¢-algebra, as a brief computation
shows. By assumption, E 0  AQ  A0 . Since A0 is the smallest ¢-algebra containing
E 0 , we conclude that AQ D A0 , and the assertion follows. t
u

Rather frequently one considers the ¢-algebra generated by the open subsets in a
Euclidean space or, more generally, in a metric space.

Definition
Let .S; d/ be a metric space with metric d and let O be the system of its open
subsets. Its Borel-¢-algebra B WD ¢.O/ is defined as the ¢-algebra generated
by the open subsets of S. Its elements are called Borel sets. A mapping between
two metric spaces is called Borel measurable, if it is measurable w.r.t. the Borel
¢-algebras.

In a topological space, too, the ¢-algebra generated by the open sets is called the
Borel ¢-algebra. We restrict our treatment to metric spaces, where the circumstances
remain clear.
We now have at our disposal a method for constructing measurable sets which
is highly indirect. In general it does not give us any indication which subsets of S
actually belong to ¢.E/ resp. ¢.O/. In contrast to, e.g., the open sets in a metric
10 2 Measurability

space, they cannot be characterized “individually”. However, this does not create
any serious problems; one just works with systems of sets instead of individual sets.

Example

1. By virtue of Proposition 2.3, the Borel ¢-algebra is also generated by the system
of all closed subsets (the complements of the open sets).
2. Every continuous mapping between two metric spaces is Borel measurable. This
follows from Proposition 2.4 because, for continuous mappings, the preimages
of open sets are again open and hence Borel measurable.
3. We denote the Borel ¢-algebra of the Euclidean space Rd by B d . It is generated,
too, by the system of all d-dimensional open intervals of the form

.1; b/ WD .1; b1 /      .1; bd / ; b D .b1 ; : : : ; bd / 2 Rd :

Indeed, from those intervals we may obtain every finite half-open interval
Œa; b/ D Œa1 ; b1 /      Œad ; bd / according to

[
d
Œa; b/ D .1; b/ n .1; ci /
iD1

with ci WD .b1 ; : : : ; bi1 ; ai ; biC1 ; : : : ; bd /, and furthermore every open set O as a

countable union of half-open intervals according to
[˚ 
OD Œa; b/ W Œa; b/  O; a; b 2 Qd I

note that, because the rational numbers are dense in R, for every open set O and
every x 2 O there is an interval Œa; b/ with x 2 Œa; b/  O and a; b 2 Qd .
Therefore, the finite half-open intervals Œa; b/ generate the Borel ¢-algebra, too.
In the same manner B d is also generated by all finite open or by all finite closed
intervals, and moreover by all intervals .1; b, b 2 Rd .
4. As a consequence, every monotone mapping ® W R ! R is Borel measurable,
since the preimage of an interval under ® is again an interval, and hence a Borel
set.
5. Let ®1 ; ®2 ; : : : be an infinite sequence of measurable mappings from a measurable
space S with ¢-algebra A to a metric space S0 with metric d and Borel ¢-algebra
B. We assume that the sequence converges pointwise to a mapping ® W S ! S0 ,
thus d.®n .x/; ®.x// ! 0 holds for all x 2 S. Then ® is measurable. Indeed, let
B  S0 , © > 0, and let U© .B/ WD fy 2 S0 W d.y; z/ < © for a z 2 Bg be the “open
©-neighbourhood” of B. If B is closed, then for every seqence ©1  ©2     > 0
converging to 0 we have
Generators of ¢-Algebras, Borel ¢-Algebras 11

1
\ ˚ 
®1 .B/ D x 2 S W ®n .x/ 2 U©k .B/ except for finitely many n
kD1
1 [
\ 1
˚ 
D x 2 S W ®n .x/ 2 U©k .B/ for all n  m
kD1 mD1

\ 1 \
1 [ 1
 
D ®1
n U©k .B/ 2 A ;
kD1 mD1 nDm

and the assertion follows from Proposition 2.4. This convergence property is a
feature which distinguishes the class of measurable functions from other classes
of functions (like e.g., the continuous functions), compare Exercise 7.4.

We may also generate ¢-algebras from mappings.

Definition
Let .Si ; Ai /; i 2 I, be measurable spaces and §i W S0 ! Si ; i 2 I, mappings. The
smallest ¢-algebra A0 in S0 such that all §i are A0 -Ai -measurable mappings is
called the ¢-algebra generated by the .§i /. We denote it by A0 D ¢.§i ; i 2 I/.
S ˚ 1 
The ¢-algebra ¢.§i ; i 2 I/ is generated by E 0 D i2I §i .Ai / W Ai 2 Ai .

Example
The Borel ¢-algebra B in a metric space S with metric d coincides with the ¢-algebra B0
generated by all continuous functions § W S ! R. On the one hand, continuous functions
are Borel measurable, therefore B0  B. On the other hand, for all sets B  S the function
x 7! §B .x/ WD inffd.x; z/ W z 2 Bg (the “distance” between x and B) is continuous from S to R,
because j§B .x/  §B .y/j  d.x; y/. If B is closed we have in addition that x 2 B , §B .x/ D 0,
thus B D §B1 .f0g/. Therefore, B0 includes all closed sets, and by Proposition 2.3 we conclude
that B  B0 .

The following statement corresponds to the measurability criterion.

Proposition 2.5. Let .S; A/, .S0 ; A0 / and .Si ; Ai /, i 2 I, be measurable spaces,
and let A0 be generated by the mappings §i W S0 ! Si , i 2 I. A mapping ® W S ! S0
is A-A0 -measurable if and only if §i ı ® is A-Ai -measurable for all i.

Proof. The “only if”-part follows because the composition of measurable mappings
is again measurable. For the converse, let §i ı ® be measurable for all i, thus .§i ı
®/1 .Ai / 2 A for all Ai 2 Ai . This means that ®1 .A0 / 2 A for all A0 D §i1 .Ai /
with Ai 2 Ai . Those sets A0 generate the ¢-algebra A0 . The measurability of ® now
follows from the measurability criterion. t
u
12 2 Measurability

Product Spaces

We now apply our construction method for ¢-algebras to finite or countably infinite
Cartesian products
Y
S D Sn D S1  S2    
n1

Let A1 ; A2 ; : : : be ¢-algebras on S1 ; S2 ; : : : We call a subset of S of the form

A1  A2     with An 2 An

a measurable rectangle.

Definition
The ¢-algebra A˝ in S generated by all measurable rectangles is called the
product ¢-algebra of A1 ; A2 ; : : :. We call .S ; A˝ / the product space of .Sn ; An /
and write
O
A˝ D An D A1 ˝ A2 ˝    :
n1

If, in particular, S1 D S2 D    D S und A1 D A2 D    D A, we write

Sd and Ad (2.1)

instead of S and A˝ . Here d denotes the length of the sequence S1 ; S2 ; : : : The

case d D 1 is included; S1 is just the set of infinite sequences in S.
Alternatively, we may describe the product ¢-algebra by the projection mappings
 i W S ! Si , i  1, given by

 i .x1 ; x2 ; : : :/ WD xi :

Since  1
i .Ai / D S1      Si1  Ai  SiC1     ,  i is an A˝ -Ai -measurable
mapping. Moreover, A1  A2     D  1 1
1 .A1 / \  2 .A2 / \    , therefore we
may characterize the product ¢-algebra as the ¢-algebra generated by the projection
mappings:

A˝ D ¢. i ; i  1/ :

Example (Euclidean spaces)

The ¢-algebra Bd in Rd , 2  d < 1, can be regarded as either a Borel ¢-algebra (thus, generated
by the open sets) or as a product ¢-algebra, because on Rd D Rd1      Rdk , d D d1 C    C dk ,
Product Spaces 13

it holds that

B d D B d1 ˝    ˝ B dk : (2.2)

To prove this we first note that every open set O  Rd is a countable union of measurable
rectangles, e.g., as above,
[˚ 
OD Œa; b/ W Œa; b/  O ; a; b 2 Qd :

Thus O belongs to the product ¢-algebra. Since Bd is the smallest ¢-algebra including all open sets,
it follows that Bd  Bd1 ˝    ˝ Bdk . Conversely, the projection mappings  i W Rd ! Rdi are
continuous and thus Bd -Bdi -measurable, and therefore Bd1 ˝    ˝ Bdk D ¢. 1 ; : : : ;  k /  Bd .

Example (The extended real line)

When considering suprema and infima of countably many measurable real functions it is
N WD R [ f1; 1g. We equip R
convenient to extend the range to R N with the ¢-algebra
˚ 
N j B \ R is a Borel set in R ;
BN WD B  R

called the Borel ¢-algebra in R N (cf. Exercise 2.6), and R

N d with the product ¢-algebra BN d . Here, d
is either a natural number, or d D 1. The functions

sup W R N ;
Nd ! R Nd ! R
inf W R N ;

which to every finite or infinite sequence x1 ; x2 ; : : : assign its infimum and supremum, respectively,
then become BNd -BN-measurable. This follows from
 
sup1 Œ1; x D Œ1; x  Œ1; x     ;
 
inf1 Œx; 1 D Œx; 1  Œx; 1     ;

the measurability criterion, and from the fact that BN (similarly to the Borel ¢-algebra on the real
axis) is generated by the intervals Œ1; x, and just as well by the intervals Œx; 1.
As a price to be paid, however, it is no longer possible to subtract and divide arbitrary elements
N without entangling oneself in contradictions.
of R
No difficulties arise with the rules

1 C 1 WD 1 ; 0  1 WD 0 ; a  1 WD 1 for a > 0 ; .1/  1 D 1 I

we will use them in the sequel. In contrast, one has to avoid the expressions
1
11; I
1
they are (and remain) undefined.

Product ¢-algebras have the important property that mappings into a Cartesian
product are measurable if and only if the same is true for all their components.
14 2 Measurability

Proposition 2.6. Let .S; A/ be a measurable space, let ®i W S ! Si be mappings,

i  1. Then the mapping ® WD .®1 ; ®2 ; : : :/ from S to S is A-A˝ -measurable if
and only if all mappings ®i are A-Ai -measurable.

Proof. This is a special case of the preceding proposition, as ®i D  i ı ®. t

u

Real Functions

We summarize:

Proposition 2.7. If .S; A/, .Si ; Ai /, i  1, .S0 ; A0 / are measurable spaces, and
if the mappings ®i W S ! Si are A-Ai -measurable and § W S1  S2     ! S0 is
A˝ -A0 -measurable, then § ı .®1 ; ®2 ; : : :/ is A-A0 -measurable.

Using this result one may ascertain the measurability of many mappings and sets.
We demonstrate this for the particularly important case of functions with values in
R and RN D Œ1; 1 (Rd and R N are equipped with the Borel ¢-algebra B d resp. B).
N
Here, the simplest functions are the characteristic functions 1A of subsets A of S,
taking the value 1 on A and 0 on Ac . 1A is a measurable function if and only if A is
a measurable subset.
Let now f1 ; f2 W S ! R be measurable functions and let c1 ; c2 2 R. Then the
linear combination c1 f1 C c2 f2 is a measurable function. This follows from the
representation

c1 f1 C c2 f2 D ® ı .f1 ; f2 / ;

where ®.x; y/ WD c1 x C c2 y, due to continuity, is a Borel measurable mapping from

R2 to R. In the same way one obtains the measurability of

f1  f2 ; max.f1 ; f2 / ; min.f1 ; f2 /

and, for every measurable f, the measurability of

fC WD max.f; 0/ ; f WD max.f; 0/ ; jfj D fC C f :

For measurable functions f1 ; f2 W S ! R the measurability of the set

˚ 
ff1 D f2 g WD x 2 S W f1 .x/ D f2 .x/

results from the fact that ff1 D f2 g D .f1 ; f2 /1 .D/ with the “diagonal” D WD
f.x; y/ 2 R2 W x D yg, since D, being a closed subset of R2 , is Borel measurable.
Real Functions 15

Analogously one obtains the measurability of sets like

˚ 
ff1 f2 g WD x 2 S W f1 .x/ f2 .x/

or ff1 ¤ f2 g, ff1 < f2 g.

In the same manner one may construct new measurable functions from infinite
sequences f1 ; f2 ; : : : of given measurable functions from S to R, extending R in the
process to RN when necessary. We have shown that the mappings sup; inf W R N1 ! R N
are measurable; therefore, for measurable f1 ; f2 ; : : : their pointwise supremum and
infimum

sup fn D sup ı.f1 ; f2 ; : : :/ ; inf fn D inf ı.f1 ; f2 ; : : :/

n1 n1

are measurable. As a consequence, the pointwise limit superior and limit inferior

lim sup fn D inf sup fn ; lim inf fn D sup inf fn ;

n!1 m1 nm n!1 m1 nm

are measurable. Moreover, flimn fn existsg is a measurable set, since

flim fn existsg D flim sup fn D lim inf fn g \ f1 < lim sup fn < 1g :
n n n n

If the sequence f1 ; f2 ; : : : converges pointwise, we have limn fn D lim supn fn , and

thus limn fn is a measurable function. With that property of measurable mappings
we are already acquainted.
For the theory of integration, the following characterization of measurable
nonnegative functions will be important. Using it we will extend properties of the
integral from specific sets of measurable functions to all measurable functions.
Nonnegative functions are always understood as functions values in R N C D Œ0; 1.

Proposition 2.8 (Monotonicity principle). Let .S; A/ be a measurable space

and let K be a set of functions f W S ! R
N C with the properties

(i) f1 ; f2 2 K ; c1 ; c2 2 RC ) c1 f1 C c2 f2 2 K ;
(ii) f1 ; f2 ; : : : 2 K ; f1 f2    ) supn fn 2 K ;
(iii) 1A 2 K for all A 2 A .

N C,
Then K includes all nonnegative measurable functions on S (with values in R
according to our terminology).
16 2 Measurability

Proof. Let f W S ! R N C be measurable. The sets Ak;n WD fk2n < f .k C 1/2n g

belong to A, for all natural numbers k and n. The functions

X
n2n
k
fn WD 1A C n 1ffD1g
kD1
2n k;n

then belong to K by virtue of (i) and (iii).

One has that f1 f2    and supn1 fn D f, therefore using (ii) we obtain f 2 K,

as asserted. u
t

Exercises
2.1 Let S be a set. Which ¢-algebra is generated by the subsets of S consisting of a single element
only ? What are the measurable functions f W S ! R? S
2.2 Let E1 ; E2 ; : : : be a partition of S, that is, a sequence of disjoint subsets of S with n1 En D S.
Let A be the ¢-algebra generated by those sets. Describe all sets which belong to A.
2.3 Let A1 ; A2 be ¢-algebras on S. Is A1 \ A2 a ¢-algebra? What about A1 [ A2 ?
Hint: One may construct counterexamples from ¢-algebras with 4 elements.
2.4 Prove: The ¢-algebra BN on R N is generated by the intervals Œ1; b, b 2 R.
2.5 Let S be a metric space with metric d. Prove:
(i) Every closed set F  S can be obtained as an intersection of countably many open sets
(one says that F is a G• -set).
(ii) The Borel ¢-algebra in S equals the smallest system B0 ofSsets which includes T all open
sets and moreover, for every sequence B1 ; B2 ; : : :, the sets n1 Bn and n1 Bn .
Hint: Consider the system fB 2 B0 W Bc 2 B0 g.
2.6 Let m W R N ! R be strictly monotone and bounded. Prove that dN .x; y/ WD jm.x/  m.y/j
defines a metric dN on R N and that the corresponding Borel ¢-algebra equals BN .
Hint: The system O  R N of open sets depends upon whether and where m has jumps!
2.7 The graph of a measurable mapping Let ®; §; §0 W S ! S0 be A-A0 -measurable
mappings and assume that the “diagonal” D WD f.x; y/ 2 S0  S0 W x D yg belongs to
A0 ˝ A0 . Prove that f§ D §0 g 2 A and conclude that
˚ 
.x; y/ 2 S  S0 W y D ®.x/ 2 A ˝ A0 :

2.8 Let S be uncountable and set A WD fA  S W A or Ac is countableg. Prove:

(i) A is a ¢-algebra.
Real Functions 17

(ii) For every A0 2 A ˝ A, either A0 or .A0 /c is thin. Here we say that A0  S2 is “thin”, if
A0  .A  S/ [ .S  A/ for some countable set A  S.
(iii) The diagonal D WD f.x; y/ 2 S  S W x D yg does not belong to A ˝ A.
2.9 A function g W Rd ! R N is called upper semicontinuous, if

lim sup g.y/  g.x/

y!x

holds for all x 2 Rd . Prove:

(i) g is upper semicontinuous if and only if the sets fg < ag WD fx 2 Rd W g.x/ < ag are
open for all real numbers a.
(ii) Upper semicontinuous functions are Borel measurable.
(iii) For every (not necessarily measurable) function f W Rd ! R, the functions

g.x/ WD lim sup f.y/ ; h.x/ WD lim inf f.y/ ; x2R;

©#0 jyxj© ©#0 jyxj©

are upper semicontinuous, resp. lower semicontinuous (that is, h is upper semicontinu-
ous). Prove: The set C  Rd of points of continuity of f is a Borel set, and f1C is Borel
measurable.
(iv) A function f W Rd ! R having at most countably many points of discontinuity is Borel
measurable.
Measures
3

Carrying measures is an essential purpose of measurable spaces.

Definition
Let .S; A/ be a measurable space. A mapping  which to every A 2 A assigns a
number .A/  0, or possibly the value .A/ D 1, is called a measure, if:

(i) .¿/ D 0 ;  S  P
(ii) ¢-additivity:  n1 An D n1 .An / for every finite or infinite sequence
A1 ; A2 ; : : : of pairwise disjoint measurable sets.

The triple .S; A; / is called a measure space. If .S/ D 1, then  is called a

probability measure. More generally,  is called finite if .S/S< 1, and ¢-finite
if there exist measurable sets A1  A2     such that n1 An D S and
.An / < 1 holds for all n.

In the Introduction we have been guided by the idea that .A/ is the volume of
A. One also may think of  as describing a mass distribution on S, and then .A/
equals the mass of A. In probability theory one interprets the elements A of the
¢-algebra as observable events occuring with probability .A/.
¢-finite measures are of interest for two reasons. Firstly, some important
measures are ¢-finite, e.g., the Lebesgue measure, on Rd which we will address
soon. Secondly, properties of finite measures often extend to the ¢-finite case.
One achieves this by replacing a given ¢-finite measure  with finite measures
n ./ WD . \ An / and then passing to the limit n ! 1. Often this does not
present any difficulties whatsoever, so that one may omit the details.

© Springer International Publishing Switzerland 2015 19

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_3
20 3 Measures

Example

1. A Dirac measure1 is a probability measure whose total mass is concentrated in

a single point. The Dirac measure •x in the point x 2 S of a measurable space is
defined as
(
1 ; if x 2 A ;
•x .A/ WD
0 ; if x … A :

As values it takes on only 0 and 1.

2. A measure  is called discrete if its total mass is concentrated in a countable
measurable set, that is, if .Cc / D 0 holdsfor some countable set C  S. In this
case,  is specified by its weights x WD  fxg , x 2 C, according to the formula
X
.A/ D x :
x2A\C

Conversely, from any family .x /x2C of nonnegative numbers, using this formula
one obtains a discrete measure .

The following proposition summarizes several essential properties of measures. For

sets A; A1 ; A2 ; : : :  S we write
[
An " A ; if A1  A2     and A D An ;
n1
\
An # A ; if A1  A2     and A D An :
n1

Proposition 3.1. For any measure  und any measurable sets A; A1 ; A2 ; : : :

there holds:

(i) Monotonicity: .A 1S/ .A2 /, ifPA1  A2 ,

(ii) ¢-subadditivity:  n1 An n1 .An /,
(iii) ¢-continuity: If An " A, then .An / ! .A/ for n ! 1.
If An # A and moreover .A1 / < 1, then .An / ! .A/ for n ! 1 as
well.

Proof. (i) In the case A1  A2 , A2 equals the disjoint union of A1 and A2 n A1 ,

and .A1 / .A1 / C .A2 n A1 / D .A2 / follows by additivity.

1
PAUL DIRAC , 1902–1984, born in Bristol, active in Cambridge. He is famous in particular for his
contributions to the foundations of quantum mechanics. In 1933 he was awarded the Nobel prize
for physics.
3 Measures 21

(ii) To begin with, we have .A1 [ A2 / D .A1 / C .A2 n A1 / .A1 / C .A2 /

due to additivity und monotonicity. For finite unions it follows by induction that
.A1 [    [ Ak / .A1 [    [ Ak1 / C .Ak / .A1 / C    C .Ak1 / C
.Ak /. It remains to prove
P the assertion for infinite unions. For this, one passes
in .A1 [    [ Ak / n1 .An / to the limit k ! 1, using the ¢-continuity
of measures, which we will prove next.
(iii) Assuming An " A,S the sets A01 WDSA1 , A0k WD Ak n Ak1 , k  2, are disjoint
and we have An D nkD1 A0k , A D 1 0
kD1 Ak . Consequently, as n ! 1,

[
n  Xn 1
X [
1 
.An / D  A0k D .A0k / ! .A0k / D  A0k D .A/ :
kD1 kD1 kD1 kD1

This yields the first assertion. Assuming An # A we get A00n " A1 n A for
A00n WD A1 n An , n  1. Consequently,

.An / C .A00n / D .A1 / D .A/ C .A1 n A/ :

Passing to the limit n ! 1 yields the second assertion, using the first assertion
as well as the assumption .A1 / < 1.
t
u

I Remark The condition .A1 / < 1 in the last assertion cannot be omitted
without replacement. A counterexample is provided by the sequence of sets
An WD fm 2 N W m  ng. The An all have measureT 1 for the counting measure 
on N defined by .A/ WD #A. On the other hand, n1 An D ¿ has measure zero.

Measures can be mapped to other measure spaces via measurable mappings. This
issue will become important presently.

Definition
Let .S; A/, .S0 ; A0 / be measurable spaces, let ® W S ! S0 be measurable, and let
 be a measure on A. The measure 0 on S given by
 
0 .A0 / WD  ®1 .A0 / ; A0 2 A0 ;

is called the image measure of  under the mapping ®. We write 0 D ®./.

A short S
computation
 shows
 S that 1 0 indeed  isPa measure:
0
 1  0.¿/ DP.¿/ 0D 0 0
and 0 A
n1 n
0
D  n1 ® .A 0
n / D n1  ® .A n / D n1  .An /
for pairwise disjoint sets A01 ; A02 ; : : : 2 A0 . Just as quickly one convinces oneself
that

.§ ı ®/./ D §.®.// :
22 3 Measures

From ®.x/ D y it follows that ®.•x / D •y , we thus have transferred ® to measures

in a canonical manner.

Null Sets

We now broach the topic of those measurable sets which a given measure does not
distinguish from the empty set.

Definition (Null set)

Let .S; A; / be a measure space. A set A  S is called a null set, more precisely
a -null set, if A 2 A and .A/ D 0.

The system N  A of all null sets of a not identically vanishing measure  has the
following properties, as a consequence of monotonicity and ¢-subadditivity:

¿2N ; S…N ;
A 2 N ; A0 2 A; A0  A ) A0 2 N ;
[
A1 ; A2 ; : : : 2 N ) An 2 N :
n1

If a property holds for all elements of S except for elements in some null set, then
one says that the property holds almost everywhere.

Definition
Let .S; A; / be a measure space. Two measurable mappings ®; § W S ! S0 are
called equal almost everywhere, more precisely equal -almost everywhere, if
f® ¤ §g is a null set. We write

® D § a.e.

and say that ®.x/ D §.x/ for -almost all x.

In probability theory one speaks of almost sure equality. It is an equivalence relation.

In the same manner one writes

® § a.e. W, f® > §g is a null set ;

whenever the range S0 of ® and § is equipped with an order relation .

Null sets will become important for us specifically in the context of convergence.
The Lebesgue Measure on Rd 23

Definition
Let .S; A; / be a measure space, let .S0 ; d0 / be a metric space, and let
®; ®1 ; ®2 ; : : : be measurable mappings. We say that ®n converges to ® almost
everywhere, and write

®n ! ® a.e. ;

if f®n 6! ®g WD fx 2 S W ®n .x/ 6! ®.x/g is a null set.

I Remark For any measure space .S; A; /, the system

˚ 
AQ WD AQ  S W 9 A1 ; A2 2 A with A1  A
Q  A2 and .A2 n A1 / D 0

is a ¢-algebra in S which contains A. One may also describe it as the ¢-algebra

generated by A [ NQ , where NQ denotes the system of all subsets of all null sets.
Moreover, for any AQ 2 A,
Q

. Q WD .A1 / D .A2 / ;
Q A/

is well defined. 
Q is a measure which extends  to A. Q The measure space .S; A;
Q /
Q
is called the completion of .S; A; /. (Proof as exercise)

The Lebesgue Measure on Rd

Now we will see that the concept of a measure on a ¢-algebra works out well in
an especially important case. The following nontrivial result says that there exists
a unique measure on the Borel ¢-algebra B d on Rd (d finite) which associates to
every d-dimensional interval its “natural” volume. One usually considers half-open
intervals

Œa; b/ WD Œa1 ; b1 /      Œad ; bd / ;

where a D .a1 ; : : : ; ad /, b D .b1 ; : : : ; bd / 2 Rd . Half-open intervals have the

advantage that they cover the whole space Rd completely and without intersections.
In the following picture, the bold edges belong to the corresponding interval.
24 3 Measures

Proposition 3.2. On the Borel sets of Rd there exists a unique measure, denoted
by œd , with the property that for all a1 < b1 ; : : : ; ad < bd
 
œd Œa; b/ D .b1  a1 /    .bd  ad / ;

where a D .a1 ; : : : ; ad / and b D .b1 ; : : : ; bd /.

We postpone the proof, we will show uniqueness in Chap. 7 and existence in

Chap. 11. œd is called the Lebesgue measure on B d (one also speaks of the Lebesgue-
Borel measure). Its completion, too, is termed Lebesgue measure. In the case d D 1,
we more briefly write B and œ for the Borel ¢-algebra B 1 and the measure œ1 .
We treat some important properties of the Lebesgue measure.

Proposition 3.3. The Lebesgue measure œd on B d is the only measure on B d

satisfying the following two properties:

(i) Translation invariance: œd .B/ D œd .B0 /, if B; B0 2 B d map into each other

by translation.  
(ii) Normalization: œd Œ0; 1/d D 1 for the d-dimensional unit cube Œ0; 1/d .

Proof. Only (i) has to be proved. We consider, for fixed v 2 Rd , the translation
mapping x 7! ®.x/ WD x C v on Rd and the image measure  WD ®.œd /.
Translation
  maps intervals to intervals of equal measure, that is, it holds that
 Œa; b/ D .b1  a1 /    .bd  ad /. Thus  satisfies the property which characterizes
 
the Lebesgue measure. It follows that  D œd , therefore œd .B/ D œd ®1 .B/ ,
yielding assertion (i).
Conversely, let  be any measure satisfying (i) and (ii). Then for every natural
number n we have
 
 Œ0; n1 /d D nd ;

since the cube Œ0; 1/d decomposes in nd subcubes, which all arise from Œ0; n1 /d
by translation and therefore have identical measure, by the assumption. From
such cubes one can compose all those half-open d-dimensional intervals whose
boundaries a and b consist of rational components only. By additivity,
 
 Œa; b/ D .b1  a1 /    .bd  ad /

follows for rational numbers ai < bi . Because the rational numbers form a dense
subset of the real numbers, we may enclose, from outside as well as from inside,
arbitrary intervals by intervals with rational vertices. The latter formula then follows
The Lebesgue Measure on Rd 25

for arbitrary ai < bi , by the monotonicity property of measures. Therefore,  enjoys

the property characterizing the Lebesgue measure, and so  D œd . t
u

Proposition 3.4. For the Lebesgue measure it holds that:

(i) œd .H/ D 0 for every hyperplane H  Rd .

(ii) If ® W Rd ! Rd is linear, then ®.B/ is a Borel set for every Borel set B  Rd ,
and
 
œd ®.B/ D j det ®j  œd .B/ :

Phrased in a different manner: If ® maps the canonical unit vectors e1 ; : : : ; ed to

the vectors v1 ; : : : ; vd 2 Rd , then it maps the unit cube Œ0; 1/d to the (half-open)
parallelepiped spanned by the vectors v1 ; : : : ; vd :

nX
d o
PŒv1 ; : : : ; vd  WD ci vi 2 Rd W 0 ci < 1 ; i D 1; : : : ; d :
iD1

Since det ® D detŒv1 ; : : : ; vd , we get

  ˇ ˇ
œd PŒv1 ; : : : ; vd  D ˇ detŒv1 ; : : : ; vd ˇ :

This is not surprising since the determinant detŒv1 ; : : : ; vd  can be interpreted as the
oriented volume of a parallelepiped, as we know from linear algebra. Except for the
orientation (the sign of the determinant), measure theory thus yields the same result.

Proof. (i) Every hyperplane H can be covered by countably many sets which
arise by translation from a single .d  1/-dimensional rectangle Q spanned by
some orthogonal vectors b2 ; : : : ; bd . Let b1 be orthogonal to b2 ; : : : ; bd ; the sets
Q C rb1 , 0 r 1, are pairwise disjoint and have identical Lebesgue measure
due to translation invariance. That measure has to be equal to zero, because
otherwise the rectangle spanned by b1 ; : : : ; bd would have infinite measure.
Thus Q, and therefore H too, has measure 0.
26 3 Measures

(ii) The case det ® D 0 is already covered by (i). So let us assume that ® has an
inverse § which, being a linear mapping, is continuous and therefore Borel
measurable. We conclude that ®.B/ D §1 .B/ is Borel measurable whenever
B is a Borel set.
The bijectivity of ® has additional consequences: ./ WD œd .®.// defines
a measure. Since ®.B C v/ D ®.B/ C ®.v/ we have .B C v/ D .B/ for
every v2 Rd ; thus is translation invariant. Moreover, 0 < c < 1 holds for
c WD  Œ0; 1/d , as ® Œ0; 1/d contains small cubes and is contained in a large
cube. The characterization of the Lebesgue measure in the previous proposition
now yields  D cœd .
It remains to determine the value of c. We begin by considering two simple
cases:
First, let ¢ be a linear mapping having the unit vectors e1 ; : : : ; ed as
eigenvectors, with eigenvalues ©1 ; : : : ; ©d > 0. Then Œ0; 1/d transforms into the
interval Œ0; ©1 /      Œ0; ©d /, and we directly obtain c D ©1    ©d , which in turn
equals det ¢.
Secondly, let £ be an orthogonal mapping. Then £ maps the unit ball B onto
itself. Since B can be sandwiched between cubes, 0 < œd .B/ < 1. In this case
we therefore have c D 1; on the other hand, the determinant of an orthogonal
mapping is known to be ˙1.
The assertion now results from the fact that every linear mapping ® can be
represented as ® D £1 ı ¢ ı £2 , where ¢ is as above and £1 ; £2 are orthogonal
mappings (“singular value decomposition”). Indeed, the assertion follows from
the special cases considered above and known properties of determinants:
    ˇ ˇ
œd ®.B// D œd £1 .¢.£2 .B/// D ˇ det £1 det ¢ det £2 ˇœd .B/ D j det ®jœd .B/ :

(In Exercise 3.9, we will recap the singular value decomposition of matrices.)
t
u

Exercises
S S
3.1 Let A D n1 An and A0 D n1 A0n . Check whether for every measure  it is true that
X X
.A n A0 /  .An n A0n / ; .AA0 /  .An A0n / :
n1 n1

3.2 Let 1  2     be a sequence of measures on a ¢-algebra, that is, 1 .A/  2 .A/    

holds for every measurable set A. Prove that .A/ WD limn n .A/ defines a measure .
3.3 Let A be the system of all sets A  N for which the limit

1
š.A/ WD lim #.A \ f1; 2; : : : ; ng/
n!1 n
exists. Prove: (i) š is additive, but not ¢-additive, (ii) A is not a ¢-algebra.
The Lebesgue Measure on Rd 27

3.4 Existence of non-measurable sets, due to Vitali Let N  Œ0; 1 a set having the property
that for every real number a there exists a unique number b 2 N such that a  b is rational.
Prove:
0 0
(i) N C r and SN C r are disjoint for any rational numbers r ¤ r .
(ii) Œ0; 1  r2Q\Œ1;1 .N C r/  Œ1; 2.
(iii) N is not a Borel set.
Remark: N is a complete set of representatives for the equivalence relation on R given by
a b W, a  b 2 Q. One obtains N using the axiom of choice from set theory.
3.5 Egorov’s theorem Let  be a finite measure, and let f1 ; f2 ; : : : converge -a.e. to f. We want
to prove that for each © > 0 there exists a measurable set A  S such that f1 ; f2 ; : : : converges
uniformly to f on A and thatS.Ac /  ©. For this purpose,
T prove:  
(i) Let • > 0 and A0m WD nm fjfn fj > •g. Then m1 A0m  ffn 6! fg and  A0m ! 0
for m ! 1.
k
(ii) For each S© > 0 there exist natural numbers m1 < m2 <    such that .Ak /  ©2 for
Ak WD nmk fjfn  fj > 1=kg. T
(iii) f converges uniformly on A WD k1 Ack , and .Ac /  ©.
3.6 On R, we consider the Borel measurable functions f D 1Q and g D 1Œ0;1 . Which of these
functions are (i) a.e. continuous, (ii) a.e. equal to a continuous function (with respect to the
Lebesgue measure)?
3.7 Let ® W Rd ! Rd be linear and bijective. Prove that for the image of the Lebesgue measure
under ® it holds that ®.œd /./ D j det ®j1 œd ./.
3.8 Let B  Œ0; 1/ be a Borel set. Prove that for  each © > 0 there exist (half-open) intervals
Sk
I1 ; : : : ; Ik  Œ0; 1/ such that œ1 B jD1 Ij < ©. In addition, consider the d-dimensional
case.
Hint: Consider the system of all sets B having the stated property.
3.9 Singular value decomposition Let M be an invertible dd-matrix, and let M be its adjoint.
Prove:
(i) M M is selfadjoint and invertible, with strictly positive eigenvalues ©21 ; : : : ; ©2d . Thus there
exists an orthogonal matrix O such that M M D O D2 O; here D denotes the diagonal
matrix with entries ©1 ; : : : ; ©d .
(ii) The mapping DOx 7! Mx, x 2 Rd , is well-defined, linear and orthogonal, that is,
jDOxj2 D jMxj2 for all x.
(iii) There exists an orthogonal matrix V such that M D VDO (“singular value decomposi-
tion”).
The Integral of Nonnegative Functions
4

Given a measure  on a measurable space .S; A/, we now define the integral for
arbitrary measurable functions

f0:

Here we consider measurable functions on S taking values in RN C D Œ0; 1.

The integral will be defined with the aid of elementary functions. These are
measurable functions h  0 having at most finitely many different real values. Thus
X
h D z  1fhDzg ;
z

where the summation runs over the finitely many real function values z of h. In
the case S D R the elementary functions include the step functions, for which
the sets fh D zg are intervals or finite union of intervals, but additionally, due to

the diversity of the Borel sets, quite different functions which no longer can be
represented graphically.

© Springer International Publishing Switzerland 2015 29

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_4
30 4 The Integral of Nonnegative Functions

The integral of f  0 arises through exhaustion from below by means of

elementary functions, as follows:

Definition
Let f W S ! R N C be measurable. The integral of f w.r.t. the measure  (more
precisely, the Lebesgue integral) is defined as
Z nX o
f d WD sup z  .h D z/ W h  0 is elementary; h f :
z

 
Here, for  fh D zg we simply write .h D z/. The integral may possibly have
the value 1. Sometimes, in particular when the integrand f also depends on other
variables in addition to x, one has to specify clearly with respect to which variable
one integrates. In that case one writes the integral as
Z
f.x/ .dx/ :

One can interpret the integral as the “content” of the region between 0 and f w.r.t.
 (we will come back to this in Exercise 8.4). In the case of a probability measure
the integral may be interpreted as the “mean value” of f w.r.t. . If,
R in particular,  is
a probability measure on RC interpreted as a mass distribution, x .dx/ becomes
its center of mass. In probability theory one uses integrals in a similar manner, in
order to define expectation values.
From the definition we at once draw a simple but important conclusion.

Proposition 4.1 (Markov’s Inequality1 ). Let f  0 be measurable, and let z be

a nonnegative number. Then
Z
z  .f  z/ f d :

Proof. For the elementary function h WD z  1ffzg we have 0 h f. t

u

The following properties of the integral are immediate consequences of its defini-
tion.

1
ANDREI M ARKOV , 1856–1922, born in Ryazan, active in St. Petersburg. He is mainly known for
his fundamental contributions to probability theory.
4 The Integral of Nonnegative Functions 31

Proposition 4.2. For arbitrary measurable functions f; g  0 it holds that:

R R
(i) f g a:e: ) R f d R g d,
(ii) fR D g a:e: ) f d D g d,
(iii) R f d D 0 , f D 0 a:e:,
(iv) f d < 1 ) f < 1 a:e:

0
Proof. (i) If h  0 is elementary P with h f, then also hP WD h1ffgg is elementary
0 0
and
P h g. By assumption, z z  .h D z/ D z z  .h D z; f g/ D
z z  .h D z/, and the assertion follows from the definition of the integral.
(ii) Follows from (i).
(iii) The implicationR ( follows from (ii) and the definition of the integral.
Conversely, let f d D 0. For any n 2 N, Markov’s inequality yields
.f  1=n/ D 0. Since ff  1=ng " ff > 0g and due to ¢-continuity,
.f > 0/ D 0 follows, and therefore f D 0 a.e. R
(iv) From h WD z  1ffD1g R f for all z > 0 it follows that z  .f D 1/ f d
for all z > 0. From f d < 1 we therefore get .f D 1/ D 0, which yields
the assertion. t
u

The following proposition, also called Beppo Levi’s2 theorem, is a key result element
in the theory of the Lebesgue integral.

Proposition 4.3 (Monotone Convergence Theorem). Let 0 f1 f2   

hold for measurable functions f1 ; f2 ; : : :, and set f WD supn1 fn . Then
Z Z
f d D lim fn d :
n!1

R
Proof.
R By Proposition
R 4.2 (i), the sequence fn d increases monotonically, and
limn fn d f d. To prove the reverse inequality, let h  0 be elementary with
h f and let © > 0. The elementary functions

hn WD .h  ©/C  1ffn >h©g

2
BEPPO LEVI , 1875–1961, born in Turin, active in Piacenza, Cagliari, Parma, Bologna, and
Rosario. He wrote papers in areas as distinct as algebraic geometry, set theory, integration theory,
projective geometry, and number theory. Because of his Jewish origin he went into exile to
Argentina in 1939.
32 4 The Integral of Nonnegative Functions

(setting gC WD max.g; 0/) then satisfy 0 hn fn . From the definition of the

integral it follows that
X Z
.z  ©/C .h D z; fn > h  ©/ fn d :
z

By the assumption, we have ffn > h  ©g " S and therefore .h D z; fn > h  ©/ !
.h D z/ thanks to ¢-continuity. Consequently,
X Z
C
.z  ©/ .h D z/ lim fn d ;
n!1
z

P R
and letting © ! 0 we finally obtain z zR .h D z/ R limn fn d. Using the
definition of the integral we conclude that f d limn fn d, as claimed. t
u

A useful variant of the monotone convergence theorem is given by the following

result.

Proposition 4.4 (Fatou’s Lemma3 ). Let f; f1 ; f2 ; : : :  0 be measurable func-

tions satisfying the inequality f lim infn fn a.e. Then
Z Z
f d lim inf fn d :
n!1

Proof. Setting gn WD infmn fm we get 0 g1 g2    , supn1 gn D

lim infn!1 fn , and gn fn . Using the monotone convergence theorem we conclude
that
Z Z Z Z
f d lim inf fn d D lim gn d lim inf fn d :
n!1 n!1 n!1

t
u

In Fatou’s lemma, one cannot avoid the limit inferior: even if f equals the pointwise
limit of fn , in general we cannot replace the lim inf of the integrals with the lim in
the assertion, as the following example shows.

3
PIERRE FATOU , 1878–1929, born in Lorient, active as astronomer at the Paris observatory. To
him we owe applications of Lebesgue integration to Fourier series and complex analysis.
4 The Integral of Nonnegative Functions 33

Example
Let .an / be an arbitrary sequence of positive numbers. Then fn WD an n1.0;1=n defines a sequence of
Borel
R measurable mappings from R to R which converges to 0 pointwise. The Lebesgue integral
fn dœ equals an and therefore may not converge. The following picture shows that the same effect
can be achieved with continuous functions, too.

In order to guarantee the convergence of integrals one therefore needs additional

conditions, like monotonicity in the monotone convergence theorem. In the fol-
lowing chapter we will encounter a different convergence criterion, namely the
dominated convergence theorem.
We now compute the integral for functions taking finitely or countably infinitely
many values.

Proposition 4.5. For any measurable function f  0 taking only countably many
values (possibly including the value 1) it holds that
Z X
f d D y  .f D y/ :
y

The sum ranges over all values y of f and does not depend on the order of
summation.

Proof. Assume at first that f is elementary. If moreover h, too, is elementary and

0 h f, it follows that .f D y; h D z/ D .¿/ D 0 whenever z > y, therefore
X XX
z  .h D z/ D z  .h D z; f D y/
z z y
XX X
y  .f D y; h D z/ D y  .f D y/ :
y z y

For elementary f we thus obtain the assertion directly from the definition of the
integral.
In the general case, let y1 ; y2 ; : : : be an arbitrary enumeration of the real values
of f, and let 0 z1 z2 : : : be a divergent sequence of real numbers not including
34 4 The Integral of Nonnegative Functions

any value of f. We set

X
n
fn WD yk 1ffDyk g C zn  1ffD1g :
kD1

Then 0 f1 f2    are elementary functions with f D supn fn . The assertion

now carries over from fn to f with the aid of the monotone convergence theorem. ut

We now employ monotone convergence in order to prove the additivity and positive
homogeneity of the integral.

Proposition 4.6. For any measurable functions f; g  0 and any real numbers
a; b  0 we have
Z Z Z
.af C bg/ d D a f d C b g d :

Proof. For functions f; g having countably many values the assertion follows from
Proposition 4.5, due to ¢-additivity:
X   X X  
z   af C bg D z D z  f D u; g D v
z z u;v
auCbvDz
XX  
D .au C bv/   f D u; g D v
u v
X X
D a u  .f D u/ C b v  .g D v/ :
u v

In the general case, in addition to f and g we consider the functions

X1
k
fn WD  1fk=2n <f.kC1/=2n g C 1  1ffD1g
kD1
2n

and analogously gn . This implies 0 f1 f2    as well as supn1 fn D f.

Analogous properties hold for gn , therefore the assertion results from passing to the
limit in
Z Z Z
.afn C bgn / d D a fn d C b gn d

by virtue of the monotone convergence theorem. t

u
4 The Integral of Nonnegative Functions 35

Combining additivity and monotone convergence yields the following version of the
monotone convergence theorem.

Proposition 4.7. For arbitrary measurable functions fn  0 we have

Z X
1 1 Z
X
fn d D fn d :
nD1 nD1

The following proposition features an alternative formula for integrals.

Proposition 4.8. Let f  0 be measurable. Then

Z Z 1  
f d D  f > t dt :
0

The
R integral
 on the right hand side is to be understood as the Lebesgue integral
Œ0;1/  f > t œ.dt/. In the next chapter we will get on to the relation between
Lebesgue- and Riemann integral.
P1
Proof. Again we work with fn WD k
kD1 2n  1fk=2n <f.kC1/=2n g C 1  1ffD1g , now
represented in the form
1
X
fn D 2n 1ff>k=2n g :
kD1

Using Proposition 4.7 we get

Z 1
X Z 1
n
   
fn d D 2  f > k=2 D
n
 f > dt2n e=2n dt :
kD1 0

For the left-hand side it holds that 0 f1 f2    and f D supn1 fn , while for
the right-hand side we have dt2n e=2n # t and ff > dt2n e=2ng " ff > tg. Passing
to the limit n ! 1, the assertion follows due to ¢-continuity and the monotone
convergence theorem. t
u

Already now we clearly recognize the central role played by monotone convergence
within integration theory. As a method for proofs one often utilizes it in the form
of the monotonicity principle, Proposition 2.8. We will illustrate this method in the
next two subsections.
36 4 The Integral of Nonnegative Functions

The Transformation Formula

Let  be a measure on the measurable space .S; A/, let ® W S ! S0 be a A-A0 -

measurable mapping, and let

0 WD ®./

be the image measure of  under ®.

Proposition 4.9 (Transformation formula). For any measurable f W S0 ! R

NC
we have
Z Z
f d0 D f ı ® d :

Proof. We consider
n Z Z o
0
K WD f  0 W f is measurable ; f d D f ı ® d :

K satisfies the conditions (i)–(iii) of the monotonicity principle (Proposition 2.8),

by virtue of the Propositions 4.3 and 4.6 and of the definition of . Therefore K
includes all measurable functions f  0. This is the assertion. t
u

Densities

We now introduce the notation

Z Z
f d WD 1A f d
A

for any measurable A  S.

Definition
Let  and  be measures on the measurable space .S; A/. A measurable function
h  0 is called density of  w.r.t. , if
Z
.A/ D h d
A

holds for all measurable sets A  S.

Densities 37

We then write, in short,

d D h d

or (in the style of the differential calculus)

h D d=d :

Given a measure  and a measurable function h  0, one may regard

Z
.A/ WD h d ; A2A
A

as an equation defining . Indeed,  is a measure on A, the ¢-additivity being a

consequence of Proposition 4.7.

Proposition 4.10. Let d D h d, and let f  0 be measurable. Then

Z Z
f d D fh d :

Proof. We set
n Z Z o
K WD f  0 W f is measurable, f d D fh d :

By Proposition 4.6, Proposition 4.3 and the definition of a density, the assumptions
(i)–(iii) of the monotonicity principle (Proposition 2.8) are satisfied. The assertion
follows. t
u

If, in particular,  D h d and ¡ D k d, then

Z Z Z
f d¡ D fk d D fkh d

resp.

d¡ D kh d :

This rule is symbolically also written as

d¡ d¡ d
D :
d d d
38 4 The Integral of Nonnegative Functions

Note that, in general, densities are not uniquely determined, because if h is a density,
then so is h0 whenever h0 D h -a.e. holds. In the ¢-finite case, however, densities
are uniquely determined a.e.

Proposition 4.11. Let d D h d D h0 d and let  be ¢-finite. Then h D h0 -a.e.

Proof. Assume first that  is a finite measure. By Proposition 4.6,

Z Z Z
0 0 C 0
.h > h / C .h  h / d D h d C .h  h0 /C d
fh>h0 g fh>h0 g
Z
D h d D .h > h0 / :
fh>h0 g

R
Since  is finite, .h  h0 /C d D 0 follows, and so .h  h0 /C D 0 -a.e. by
Proposition 4.2 (iii). This means that h h0 -a.e. R The reverse inequality follows
analogously. In the ¢-finite case one first considers An .hh0 /C d with .An / < 1
and then passes to the limit n ! 1. t
u

We will get back to densities in Chap. 9 on absolute continuity.

Exercises
R
4.1 Let •x be the Dirac measure in x 2 S. Determine f d•x when f  0 is measurable.
4.2 Prove that for any measurable f  0 and any real number a > 0
Z Z 1
fa d D a ta1 .f > t/ dt :
0
R
N C be a Borel measurable function satisfying
4.3 Let f W R ! R f dœ < 1, and let a > 0. Prove
that
1
X
na f.nx/ < 1
nD1

holds for œ-almost

R all x 2 R.
Hint: Determine fn dœ for fn .x/ WD na f.nx/.
4.4 For measurable sets A1 ; A2 ; : : :  S we define
[ \
lim inf An WD fx 2 S W x 2 An except for finitely many ng D An :
n!1
m1 nm
Densities 39

For any measure , derive from the Fatou Lemma that

 
 lim inf An  lim inf .An / :
n!1 n!1

4.5 Borel-Cantelli Lemma For measurable sets A1 ; A2 ; : : :  S let

\ [
lim sup An WD fx 2 S W x 2 An for 1 many ng D An :
n!1 m1 nm

P
Prove that .lim R supn!1 An / D 0, assuming that n1 .An / < 1.
P
Hint: Consider f d for f.x/ WD n1 1An .x/, the number of those n with x 2 An .
R measure  on S is ¢-finite if and only if there exists a measurable function f  0 satisfying
4.6 A
f d < 1 as well as f.x/ > 0 for all x 2 S. Prove this assertion.
4.7 An abstract view onto the integral Let  be a measure on S, and let I be a mapping, which
assigns to every measurable function f  0 a number I.f/  0, possibly 1, and which fulfils
(i) f1 ; f2  0, measurable, c1 ; c2 2 RC ) I.c1 f1 C c2 f2 / D c1 I.f1 / C c2 I.f2 / ;
(ii) 0  f1  f2     measurable ) I.supn fn / D supn I.fn / ,
(iii) I.1A / DR.A/ for all measurable sets A  S .
Then I.f/ D f d for all measurable f  0.
Integrable Functions
5

Integration of measurable functions f W S ! R N is reduced to integration of

nonnegative measurable functions as follows. We decompose f into a positive and a
negative part:

f D fC  f ; where fC WD max.f; 0/ and f WD max.f; 0/ :

Definition
N be a measurable function such that
R C be a measure
Let R on S and let f W S ! R
f d and f d are not both equal to 1. We define
Z Z Z
C
f d WD f d  f d :

In the following we will focus on functions whose integral is finite. We restrict

ourselves to real-valued functions, in order to be able to add and multiply them
without any restrictions.

Definition
Let  be a measure on S. A measurable
R function f RW S ! R f is called integrable,
more precisely -integrable, if fC d < 1 and f d < 1.

Since

jfj D fC C f ;

© Springer International Publishing Switzerland 2015 41

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_5
42 5 Integrable Functions

Proposition 4.6 shows that

Z Z Z
C
jfj d D f d C f d :

This yields the following integrability criterion.

RProposition 5.1. A measurable function f W S ! R is -integrable if and only if

jfj d < 1, and it follows that
ˇZ ˇ Z
ˇ ˇ
ˇ f dˇ jfj d :

Further properties of the integral arise from the results of the preceding chapter.

Proposition 5.2 (Monotonicity). If f; g are integrable and satisfy f g a.e., we

have
Z Z
f d g d :

C   C
Proof. f Rg a.e. implies
R f C g R f C g R a.e. From Propositions 4.2 (i) and 4.6
we get that f d C g d f d C gC d. Rearranging terms we obtain
C

the assertion. This is permitted since all integrals are finite. t

u

Proposition 5.3 (Linearity). If f; g are integrable and a; b are real numbers,

then af C bg, too, is integrable, and
Z Z Z
.af C bg/ d D a f d C b g d:
5 Integrable Functions 43

Proof. Propositions 4.2 (i) and 4.6 yield the estimate

Z Z Z Z
 
jf C gj d jfj C jgj d D jfj d C jgj d < 1 ;

therefore f C g is integrable. From .f C g/C  .f C g/ D f C g D fC  f C gC  g

it follows that .f C g/C C f C g D .f C g/ C fC C gC . Integrating R this equation
according
R R to Proposition 4.6 and
R rearrangingRthe terms one obtains .f C g/ d D
f d C g d. The equation .af/ d D a f d is proved analogously. t
u

Finally, the following assertion, also called Lebesgue’s convergence theorem, holds.

Proposition 5.4 (Dominated convergence theorem). Let f1 ; f2 ; : : : be a

sequence of measurable functions converging a.e. to a Rmeasurable function f.
If, in addition, for some measurable function g  0 with g d < 1 we have

jfn j g a.e.
R
for all n, then fn and f are integrable, jfn  fj d ! 0, and
Z Z
fn d ! f d

for n ! 1.

R
R By assumption, jfj g a.e. According to Proposition 4.2 (i), jfn j d < 1
Proof.
and jfj d < 1 follow, so fn and f are integrable. Moreover, 2g  jfn  fj  0 a.e.,
therefore Fatou’s Lemma yields that
Z Z Z Z
2g d lim inf .2g  jfn  fj/ d D 2g d  lim sup jfn  fj d :
n!1 n!1

R R
Since 2g d is finite by assumption,R it follows that lim
R supn jfn  fj d 0.
Obviously
R we R 0 lim infn jfn  fj d, thus jfn  fj d ! 0. Since
R also have
j fn d  f dj jfn  fj d, we obtain the assertion. t
u

Additionally, we now present a generalization of the monotone convergence theo-

rem. This result is sometimes important (e.g., in probability theory), but it will not
be needed later.
A sequence f1 ; f2 ; : : : of R-valued
N measurable functions is called equiintegrable,
if for every © > 0 there exists an integrable function g  0 such that
Z
sup jfn j d © :
n1
fjfn j>gg
44 5 Integrable Functions

Proposition 5.5. Let the functions f1 ; f2 ; : : : be a.e. convergentRto f and equiinte-

grable. Then fn and f are integrable, and for n ! 1 we have jfn  fj d ! 0
and
Z Z
fn d ! f d :

Proof. Since ffn g is equiintegrable, so are ffC 

n g and ffn g, and they converge a.e. to
C 
f resp. f . We therefore may assume that fn ; f  0.
Let © > 0, letR g  0 beRchosen according to the equiintegrability assumption.
Then we have fn d g d C ©, therefore fn is integrable. Moreover, f is
integrable, since due to f1ff>gg lim infn fn 1ffn >gg a.e., Fatou’s Lemma implies that
Z Z
f d lim inf fn d © :
n!1
ff>gg ffn >gg

From jfn  fj .fn  min.g; fn // C j min.g; fn /  min.g; f/j C .f  min.g; f// we get
that
Z Z Z Z
jfn  fj d fn d C j min.g; fn /  min.g; f/j d C f d :
ffn >gg ff>gg

By the monotone convergence theorem, the middle integral on the right converges
to 0, therefore
Z
lim sup jfn  fj d 2© :
n!1

R
Letting © ! 0 we obtain jfn  fj d ! 0. This yields the assertion. t
u

In the next chapter, we will get backRto the role played by equiintegrability. In par-
ticular, we will see that, conversely, jfn  fj d ! 0 implies the equiintegrability
of f1 ; f2 ; : : :.
Two Inequalities 45

Example
R
Let  be a finite measure, let ˜ > 0 and jfn j1C˜ d  s for some s < 1. Then for all real
numbers c > 0 the estimate
Z Z
1 s
jfn j d  ˜ jfn j1C˜ d  ˜
c c
fjfn j>cg fjfn j>cg

holds. When  is finite, the equiintegrability of f1 ; f2 ; : : : results (for further elaboration see
Exercise 5.5).

Two Inequalities

As a consequence of monotonicity and linearity of the integral we prove two

inequalities which are based on convexity.

Proposition 5.6 (Hölder’s Inequality1 ). Let f; g be measurable real func-

R p and let p; q R> 1 be
tions, conjugate real numbers, that is, 1=p C 1=q D 1. If
jfj d < 1 and jgjq d < 1, then fg is integrable, and
ˇZ ˇ Z 1=p  Z 1=q
ˇ ˇ
ˇ fg dˇ p
jfj d jgjq d :

In the case p D q D 2 this is just the Cauchy-Schwarz Inequality2;3

Z 2 Z Z
fg d f2 d g2 d :

Proof. Since the logarithm is a concave function, for any numbers a; b  0 we have
 
1 1 1 p 1 q
p
log ab D log a C log b log q
a C b ;
p q p q

resp. ab 1p ap C 1q bq . For any ’; “ > 0 it follows that

jfj jgj 1 jfjp 1 jgjq

 C :
’ “ p ’p q “q

1
OTTO HÖLDER , 1859–1937, born in Stuttgart, active in Göttingen and Tübingen. He gave
important contributions, in particular to group theory.
2
AUGUSTIN -LOUIS CAUCHY , 1789–1857, born in Paris, active in Paris at the École Polytechnique
and the Collège de France. He is a pioneer of real and complex analysis, all the way from the
foundations to applications.
3
HERMANN AMANDUS SCHWARZ, 1843–1921, born in Hermsdorf, Silesia, active in Zürich,
Göttingen, and Berlin. His most important contributions pertain conformal mappings and the
calculus of variations.
46 5 Integrable Functions

R R
The particular choice ’ D . jfjp d/1=p and “ D . jgjq d/1=q , in case ’; “ > 0
results after integration in
Z
1 1 1
jfgj d C D 1;
’“ p q

yieldingRthe assertion. The case when ’ or “ is equal to 0 has to be treated separately.

If, e.g., jfjp d D 0, R Proposition 4.2 (iii) implies that f D 0 a.e. and therefore
fg D 0 a.e. as well as fg d D 0. t
u

The following inequality holds for normed measures only, in general.

Proposition 5.7 (Jensen’s Inequality4 ). Let  be a probability measure, let f

be integrable and let the function k W R ! R be convex. Then k ı f possesses a
well-defined integral, and
Z  Z
k f d k ı f d :

Important special cases are given by

ˇZ ˇ Z Z 2 Z
ˇ ˇ
ˇ f dˇ jfj d ; f d f2 d :

Proof. Any convex function k.x/ enjoys the property of having a supporting straight
line at every point a. This means that for every a 2 R there exists a real number b
such that

k.x/  k.a/ C b.x  a/ for all x 2 R

4
JOHAN JENSEN , 1859–1925, born in Nakskov, active in Copenhagen for the Bell Telephone
Company. He also contributed to complex analysis.
Parameter Dependent Integrals 47

and consequently

k ı f  k.a/ C b.f  a/ :
R
It follows that .k ı f/R .k.a/ C b.f  a// as well as .k 
R ı f/ d < 1, since
f is integrable. Thus, k ı f d is well-defined. In the case .k ı f/C d D 1 the
assertion now obviously holds, and so we may assume that k ı f is integrable. Due
to monotonicity, linearity, and since .S/ D 1 we see that
Z Z 
k ı f d  k.a/ C b f d  a :

R
Setting a D f d, the assertion follows. t
u

Parameter Dependent Integrals*

As an application of the dominated convergence theorem we investigate functions

of the form
Z
F.u/ WD f.u; x/ .dx/ ; u 2 U ; where U  Rd ;

concerning their continuity and differentiability.

Proposition 5.8. Let  be a measure on S, let u0 2 U and f W U  S ! R such that

(i) u 7! f.u; x/ is continuous in u0 for -almost all x 2 S,

(ii) x !
7 f.u; x/ is measurable for all u 2 U,
(iii) jf.u; x/j g.x/ for all u; x, for some -integrable function g  0.

Then F is continuous in u0 .

R
Proof.
R R f.u; x/ .dx/ is integrable for all u. The convergence of
Due to (iii),
f.un ; x/ .dx/ to f.u0 ; x/ .dx/ along every sequence un ! u0 now immediately
follows from the dominated convergence theorem. t
u

Proposition 5.9. Let  be a measure on S, let U  Rd be open, and let

f W U  S ! R be a function having the following properties for some given
i 2 f1; : : : ; dg:

(i) x 7! f.u; x/ is -integrable for all u,

(ii) f has a partial derivative w.r.t. ui , and there exists a -integrable function
48 5 Integrable Functions

g  0 such that for all u 2 U, x 2 S

ˇ @f ˇ
ˇ ˇ
ˇ .u; x/ˇ g.x/ :
@ui
@f
Then F has a partial derivative w.r.t. ui , x 7! @u i
.u; x/ is -integrable for all
u 2 U, and
Z
@F @f
.u/ D .u; x/ .dx/ :
@ui @ui

Proof. Since when forming partial derivatives the remaining variables are kept
constant, we may assume that d D 1 and that U is an open interval, without loss
of generality. Let h1 ; h2 ; : : : be a sequence converging to 0. By assumption (ii) and
the mean value theorem, for all u 2 U we have
ˇ f.u C h ; x/  f.u; x/ ˇ
ˇ n ˇ
ˇ ˇ g.x/ :
hn

The assertion therefore results from

Z
F.u C hn /  F.u/ f.u C hn ; x/  f.u; x/
D .dx/ ;
hn hn

passing to the limit n ! 1 with the aid of the dominated convergence theorem. u
t

In combination with other rules of integration, one can use the preceding result to
compute certain specific integrals. Examples can be found in the exercises.

Lebesgue and Riemann Integral*

The Lebesgue integral of an integrable function f w.r.t. the Lebesgue measure we

also denote by
Z Z Z Z
f dœd D f.x/ dx resp. f dœd D f.x1 ; : : : ; xd / dx1 : : : dxd

and, in the case d D 1, also by

Z Z b
f dœ D f.x/ dx :
Œa;b a
Lebesgue and Riemann Integral 49

Here we employ the notations commonly used for the Riemann integral5 (we will
recall the definition of the latter during the following proof.). Namely, it turns
out that the Riemann and the Lebesgue integrals of a function f coincide if both
integrals exist. The following figure visualizes the different procedures employed
when integrating according to Riemann and Lebesgue, thus making it clear.

More precisely, one has the following result.

Proposition 5.10. Let f W Œa; b ! R be a bounded (not necessarily measurable)

function, and let C; D  Œa; b denote its sets of continuity resp. discontinuity
points. Then:

(i) C and D are Borel sets and f  1C is Borel measurable.

(ii) f is Riemann integrable if and only if D is a null set for the Lebesgue measure
œ, and in this case its Riemann integral satisfies
Z b Z
f.x/ dx D f  1C dœ :
a

For Riemann integrable functions f this does not necessarily mean that f1D is Borel
measurable. However, f1D and f are measurable w.r.t. the completion of the Borel
Rb
¢-algebra
R w.r.t. the Lebesgue measure. For this reason, the equality a f.x/ dx D
Œa;b f dœ makes sense.

Proof. (i) Let a D x0 < x1 <    < xk D b be a partition P of the given interval
having mesh size w.P/ WD maxj .xj  xj1 /. We set
˚  ˚ 
ij WD inf f.x/ W xj1 x xj ; sj WD sup f.x/ W xj1 x xj

5
BERNHARD RIEMANN , 1826–1866, born in Breselenz near Hannover, active in Göttingen. His
famous publications are concerned, in particular, with complex analysis, geometry, and number
theory.
50 5 Integrable Functions

for j D 1; : : : ; k, and

X
k X
k
gP WD ij 1.xj1 ;xj  ; hP WD sj 1.xj1 ;xj  :
jD1 jD1

As is well known, the lower and upper sums of f for P are defined as

X
k Z X
k Z
UP WD ij .xj  xj1 / D gP dœ ; OP WD sj .xj  xj1 / D hP dœ :
jD1 jD1

In the following, P1 ; P2 ; : : : denotes a sequence of partitions such that w.Pn /

converges to 0 and that PnC1 is a refinement of Pn for all n. Then it holds that
gP1 gP2    f    hP2 hP1 . For the Borel measurable functions

g WD sup gPn ; h WD inf hPn

n n

we get g f h. Since w.Pn / converges to 0,

fg < hg  D  fg < hg [ Q ;

where Q denotes the set of all partition points in P1 ; P2 ; : : :. The set Q being
countable, since fg < hg is a Borel set, then so is D, and

œ.g < h/ D œ.D/ :

Moreover f  1C D g  1C holds, thus f  1C is Borel measurable, and (i) is proved.

(ii): By the dominated convergence theorem,
Z Z
g dœ D lim UPn ; h dœ D lim OPn :
n n

Since g h, we have g D h a.e. if and only if limn UPn D limn OPn . In the
latter case, f is called Riemann integrable (usually one considers equidistant
partitions, but as we see this does not matter). Therefore, D is a null set if and
only if f is Riemann integrable. Then moreover g D f  1C a.e. and

Zb Z Z
f.x/ dx D lim UPn D g dœ D f  1C dœ ;
n
a

as claimed.
t
u
Lebesgue and Riemann Integral 51

The preceding result remains valid for the d-dimensional Lebesgue measure. That
the set of continuity points is Borel measurable we already know from Exercise 2.5.
The Riemann integral, while being popular in teaching, has deficiencies which
render it useless for many purposes in analysis and probability theory. It lacks
essential properties like the monotone convergence theorem. The Lebesgue integral
mends those drawbacks.

Exercises
5.1 Let f be -integrable. Prove (for instance, using dominated convergence) that n.jfj  n/ ! 0
as n ! 1.
5.2 Let a > 1. Prove: The measurable function f W S ! R is -integrable if and only if
1
X  
ai  ai1  jfj < ai < 1 :
iD1

When  is a finite measure, the condition is equivalent to

1
X  
 jfj  n < 1 :
nD1

5.3 Prove that for n ! 1

Z n Z 1
.1  x=n/n dx ! ex dx :
0 0

Hint: Use 1  t  et . R

5.4 Let f  0 be a measurable function satisfying 0 < c WD f d < 1, and let 0 < a < 1.
Prove:
8
Z ˆ
<1 ; if a < 1;
lim n log.1 C .f=n/a / d D c ; if a D 1;
n!1 :̂
0; if a > 1:

Hint: Use Fatou’s Lemma and the dominated convergence theorem. One has log.1 C xa /  ax
for all x  0; a  1.
N
5.5 Let  be a finite measure and f1 ; f2 ; : : : be a sequence of R-valued measurable functions. Prove
the equivalence of the following assertions:
(i) f1 ; f2 ; : : : is equiintegrable.
(ii) For every © > 0 there exists a real number c > 0 such that
Z
sup jfn j d  © :
n1 fjfn j>cg
52 5 Integrable Functions

(iii) There exists a nonnegative function ® W RC ! RC satisfying ®.x/=x ! 1 for all

x ! 1 as well as
Z
sup ®.jfn j/ d < 1 :
n1

One can choose ® to be convex.

P
Hint: Use the ansatz ®.x/ D i1 .x  ci /C with 0  ci " 1.
5.6 Prove the equality
Z 1
ex  eux
dx D log u ; u>0;
0 x

using differentiation.
5.7 Show that
Z t
2
2 Z 1 2 2
et .x C1/
F.t/ WD ex dx C dx
0 0 x2 C 1
0
satisfies F.0/ D  =4 and F .t/ D 0 for all t  0. Conclude that
Z 1
2 p
ex dx D  :
1

5.8 Show that

Z 1
2 =2
F.t/ WD ex cos.tx/ dx
1
p
satisfies F.0/ D 2  and F0 .t/ C tF.t/ D 0. Conclude that
Z 1 p
2 =2 2
ex cos.tx/ dx D 2 et =2 :
1
Convergence
6

So far we had in mind two notions of convergence for measurable functions:

monotone convergence and convergence almost everywhere. Both are notions which
result from convergence of the values taken by functions at fixed (but arbitrary)
points of the domain. This is no longer the case for the two important notions
of convergence discussed in the present chapter, convergence in the mean and
convergence in measure. However, we will see that convergence almost everywhere
will nevertheless come into play.

Convergence in the Mean and the Spaces Lp ./

Definition
Let p  1. Let  be a measureRand assume that f Rand f1 ; f2 ; : : : are real-valued
measurable functions such that jfjp d < 1 and jfn jp d < 1 for all n. The
sequence f1 ; f2 ; : : : converges in p-mean (or in Lp ) to f if, for n ! 1,
Z
jfn  fjp d ! 0 :

In this case we write

p
fn ! f :

It is of fundamental significance that we may interpret convergence in the mean

as convergence with respect to some seminorm. We will now elaborate on this
aspect. Readers who want to look up the notion of a (semi-) norm are referred to
the beginning of Chap. 13.

© Springer International Publishing Switzerland 2015 53

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_6
54 6 Convergence

For 1 p < 1 let

n Z o
Lp ./ D Lp .SI / WD f W S ! R W f is measurable; jfjp d < 1 :

L1 ./ is just the set of all integrable functions. From the estimates

jf C gjp .jfj C jgj/p .2jfj/p C .2jgj/p resp.

Z Z Z
jf C gj d 2
p p
jfj d C 2
p p
jgjp d

it follows that Lp ./ is a vector space. We set

Z 1=p
Np .f/ WD jfjp d :

In addition, let
˚ 
L1 ./ WD f W S ! R W f is measurable; jfj c a.e. for some c < 1

and let
˚ 
N1 .f/ WD inf c > 0 W jfj c -a.e.

denote the essential supremum of jfj. It holds that Np .f/ ! N1 .f/ for p ! 1
(exercise).
It turns out that the quantities Np .f/ enjoy essential properties of a norm.
Obviously,

Np .af/ D jajNp .f/

for every 1 p 1 and every real number a. It is less obvious that the triangle
inequality holds.

Proposition 6.1 (Minkowski’s Inequality1 ). For any measurable functions

f; g W S ! R and any 1 p 1,

Np .f C g/ Np .f/ C Np .g/ :

1
HERRMANN M INKOWSKI , 1864–1909, born in Kaunas, active in Bonn, Königsberg, Zürich, and
Göttingen. He became famous for his contributions to number theory, convex geometry, and the
theory of relativity.
Convergence in the Mean and the Spaces Lp ./ 55

Proof. For p D 1 the assertion follows directly from the fact that jf C gj jfj C jgj.
Likewise, the case p D 1 has an easy proof.
Thus, let 1 < p < 1. Then 1=p C 1=q D 1 for q WD p=.p  1/ > 1. It follows
that
Z Z Z
jf C gj d jfjjf C gj d C jgjjf C gjp1 d
p p1

and, by Hölder’s inequality,

Z h Z 1=p  Z 1=p i Z 1=q
p
jf C gj d p
jfj d C p
jgj d jf C gj.p1/q

Since
R .p  1/q D p andR1  1=q D 1=p, the assertion
R follows. The particular cases
jf C gjp d D 0 and jfjp d D 1 resp. jgjp d D 1 have to be considered
separately, and they are trivial. t
u

Another important fact is that the convergence in the mean enjoys completeness.

Proposition 6.2 (Riesz2 -Fischer3 Theorem). Let 1 p 1, and let f1 ; f2 ; : : :

be a Cauchy sequence in Lp ./, that is,

lim Np .fm  fn / D 0 :
m;n!1

Then there exists an f 2 Lp ./ such that

lim Np .fn  f/ D 0 :
n!1

The core of the proof consists in establishing a connection to a.e. convergence by

passing to a suitable subsequence.

Lemma. Let  be a measure, let f1 ; f2 ; : : : be a sequence of real-valued

measurable functions satisfying
 
lim  jfm  fn j > © D 0
m;n!1

for every © > 0. Then this sequence possesses an a.e. convergent subsequence.

2
FRIGYES RIESZ, 1880–1956, born in Györ, active in Klausenburg, Szeged, and Budapest. He is
mainly known due to his significant contributions to functional analysis.
3
ERNST FISCHER , 1875–1954, born in Vienna, active in Brno, Erlangen, and Köln. He also played
an influential role in the development of abstract algebra. He had been forced to retire in 1938
because of his Jewish origin, but resumed teaching in Köln in 1945.
56 6 Convergence

Proof. By assumption, there exists a sequence 1 n1 < n2 <    such that, for all
m > nk ,
 
 jfm  fnk j > 2k 2k :
P
It follows that .jfnkC1  fnk j > 2k / 2k . For g WD k1 1fjfnkC1 fnk j>2k g ,
k
R
which is the number of indices k with jfnkC1
P  f nk j > 2 , it holds that g d D
k
k1 .jfnkC1  fnk j > 2 / < 1. It follows that g < 1 a.e., thus

 
 jfnkC1  fnk j > 2k for infinitely many k D 0 :
P
Consequently, the series k1 jfnkC1 fnk j converges a.e., and therefore the sequence
Pm1
fnm D fn1 C kD1 .fnkC1  fnk / converges a.e. to a measurable function f. This
completes the proof. t
u

Proof of the Riesz-Fischer Theorem. For p < 1, due to the Markov inequality we
have for every © > 0
Z
 
1
 jfm  fn j > © p jfm  fn jp d :
©

Thus there exist, by virtue of the assumption and the preceding lemma, a measurable
function f and a subsequence fn1 ; fn2 ; : : : which converges a.e. to f. By Fatou’s
Lemma it follows that, for all m  1,
Z Z Z
jfm  fj d lim inf jfm  fnk j d sup jfm  fn jp d :
p p
k!1 nm

This supremum is finite due to the assumption, thus f  fm belongs to Lp ./, and
therefore so does f. Moreover, by the assumption this supremum converges to 0 as
m ! 1, and the assertion follows. The case p D 1 is treated in a similar manner.
t
u

The spaces Lp ./ thus enjoy properties well known in the context of Euclidean
spaces. Moreover, the space Rd is subsumed in a natural manner. Indeed, set

.A/ WD #A for A  S WD f1; : : : ; dg :

Then for f W f1; : : : ; dg ! R we have

X
d 1=p
Np .f/ D jf.i/jp ;
iD1

and in the case p D 2 we arrive at the usual Euclidean norm on Rd .

Convergence in the Mean and the Spaces Lp ./ 57

One property of norms, however, does not hold: Np .f/ D 0 in general does not
imply f D 0. By Proposition 4.2 one may conclude at least that jfjp D 0 a.e. and
thus f D 0 a.e. In the same manner, the limit of a sequence converging in p-mean
is uniquely determined only in the a.e. sense. Therefore one cannot draw all the
conclusions one is accustomed to in Rd .
In order to remedy this deficiency one introduces new spaces. One makes use
of the fact that equality a.e. is an equivalence relation, and works with equivalence
classes

Œf WD fg W g D f a.e.g

instead of functions f. For 1 p 1 we set

Lp ./ WD fŒf W f 2 Lp ./g

and then for f; g 2 Lp ./, a; b 2 R define

aŒf C bŒg WD Œaf C bf ;

kŒfkp WD Np .f/ :

Obviously, all those quantities are well defined.

We may summarize our considerations as follows.

Proposition 6.3. For 1 p 1, the space Lp ./ endowed with k  kp is

a Banach space, that is, a normed vector space which is complete w.r.t. the
convergence induced by the norm.

In the case p D 2 we may, due to the Cauchy-Schwarz inequality, introduce a scalar

product
Z
.Œf; Œg/ WD fg d :

This turns L2 ./ into a Hilbert space, and the analogy with the Euclidean vector
spaces is perfect. We will further elaborate this viewpoint in Chaps. 12 and 13.
The spaces Lp ./ are readily designated as function spaces, and the equivalence
classes are written as functions. In this manner one writes kfk and .f; g/ instead of
kŒfk and .Œf; Œg/. For one thing, one often performs calculations using representa-
tives instead of equivalence classes; for another thing, equivalence classes including
a smooth function may be identified with that function.
To an equivalence class, however, one cannot in general associate a value at any
single point x 2 S having measure 0, in contrast to what one can do for functions.
Namely, for any representative one may prescribe an arbitrary value at this point.
58 6 Convergence

Convergence in Measure

Convergence in measure is particularly important in probabilistic settings (there one

speaks of convergence in probability).

Definition
Let  be a measure, let f; f1 ; f2 ; : : : be real-valued measurable functions on S. We
say that the sequence f1 ; f2 ; : : : converges in measure , or briefly in measure to
f if
 
lim  jfn  fj > © D 0
n!1

for every © > 0.

The N
 limit f isa.e. uniquely determined. Namely,
 if f is another limit, then we have
 jf  fj > © D 0, and letting © ! 0 we get  jf  Nfj > 0/ D 0.
N
One may motivate convergence in measure as a notion which compensates for a
peculiarity of a.e. convergence. Namely, in general, it is not the case that a sequence
converges a.e. if and only if every subsequence possesses an a.e. convergent
subsequence. On the other hand, the following relationship holds.

Proposition 6.4. Let  be a measure, let f; f1 ; f2 ; : : : be measurable real-valued

functions on S. Consider the assertions

(i) The sequence f1 ; f2 ; : : : converges in measure to f,

(ii) Each subsequence of f1 ; f2 ; : : : contains a subsubsequence converging a.e. to
f.

Then (i) ) (ii). For finite measures we even have (i) , (ii).

Proof. Assume that (i) holds. Then we may find (similarly as in the proof of the
preceding lemma) for each subsequence of the natural numbers a subsubsequence
1 n1 < n2 <    such that
 
 jfnk  fj > 2k 2k :
P R
For g WD k1 1fjfnk fj>2k g it follows that g d < 1 and therefore g < 1 a.e.,
thus
 
 jfnk  fj > 2k for infinitely many k D 0 :

This means that fn1 ; fn2 ; : : : converges a.e. to f. Thus (ii) is proved.
Convergence in Measure 59

Conversely, let (ii) be satisfied and let 1 n1 < n2 <    be a subsubsequence

as in (ii). For any © > 0 it then follows that 1fjfnk fj>©g ! 0 a.e. as k ! 1. In the
case of a finite measure, the dominated convergence theorem yields
Z
 
 jfnk  fj > © D 1fjfnk fj>©g d ! 0 :

 
Therefore, every subsequence of the sequence  jfn  fj > © contains a subsub-
sequence converging to 0. As a consequence, the whole sequence converges to 0.
Thus (i) holds. t
u

In particular, for finite measures every sequence converging a.e. also converges in
measure. The converse does not hold.

Example
Let f1 ; f2 ; : : : be an enumeration, in any order, of the characteristic functions 1Ik;m of the intervals
Ik;m D Πk1 m
; mk /, with k; m 2 N and 1  k  m, for example fn D 1Ik;m with n D kCm.m1/=2.
The sequence f1 ; f2 ; : : : converges nowhere in the interval Œ0; 1/, but it converges in measure to 0
w.r.t. the Lebesgue measure restricted to Œ0; 1/.

Convergence in measure is superior to convergence a.e. also regarding the following

completeness property.

Proposition 6.5. Let  be a measure and let f1 ; f2 ; : : : be measurable real-valued

functions with the property that
 
lim  jfm  fn j > © D 0
m;n!1

for every © > 0. Then there exists a measurable function f W S ! R such that the
sequence f1 ; f2 ; : : : converges in measure to f.

Proof. By the lemma above there exists a subsequence fn1 ; fn2 ; : : : which converges
a.e. to a function f. It follows that 1fjfm fj>©g lim infk!1 1fjfm fnk j>©g a.e. Using
Fatou’s lemma we obtain for every m  1 that
     
 jfm  fj > © lim inf  jfm  fnk j > © sup  jfm  fn j > © :
k!1 nm

Letting m ! 1 the assertion follows in view of the assumption. t

u
60 6 Convergence

Convergence in measure is metrizable; we will come back to this point in Exer-

cise 6.3. We may therefore express the preceding proposition as follows: any Cauchy
sequence (in such a metric) is convergent.

The Connection Between the Two Notions of Convergence*

We now relate convergence in the mean to convergence in measure. The former

notion is the stronger one. More precisely, the following proposition due to F. Riesz
holds.

Proposition 6.6. Let f and f1 ; f2 ; : : : be elements of Lp ./ for some 1 p < 1.

The following assertions are equivalent:
p
(i) fn ! f, R R
(ii) f1 ; f2 ; : : : converges in measure to f, and jfn jp d ! jfjp d as n ! 1.

Proof. (i) ) (ii): The convergence in measure follows the Markov inequality
Z
  1
 jfn  fj  © p jfn  fjp d :
©

The Minkowski inequality implies that

ˇ ˇ
ˇNp .fn /  Np .f/ˇ Np .fn  f/ ! 0
R R
and thus that jfn jp d ! jfjp d.
(ii) ) (i): By Proposition 6.4, for each subsequence of the natural numbers there
exists a subsubsequence 1 n1 < n2 <    such that fn1 ; fn2 ; : : : converges a.e.
to f. Fatou’s lemma, applied to 2p .jfjp C jfnk jp /  jfnk  fjp  0, yields
Z  Z Z Z 
2 p
2  jfj d lim inf 2
p p
jfj d C 2
p p
jfnk j d  jfnk  fjp d :
p
k!1

R
By assumption, on the right and on the left the term 2p jfjp d appears twice.
By assumption it is finite, consequently
Z
lim sup jfnk  fjp d 0 :
k!1

Altogether each subsequence contains a subsubsequence along which Np .fn  f/

converges to 0. This is equivalent to Np .fn  f/ ! 0. t
u
The Connection Between the Two Notions of Convergence 61

Condition (ii) of Proposition 6.6 may be further reshaped with the aid of a notion
which, in a somewhat more specialized form, has already appeared in the previous
chapter.

Definition
Let p  1. A sequence f1 ; f2 ; : : : in Lp ./ is called equiintegrable (or uniformly
integrable), more precisely equi-p-integrable,
R if for each © > 0 there exists a
measurable g  0 satisfying jgjp d < 1, such that
Z
sup jfn jp d < © :
n1
fjfn j>gg

Replacing g by g C jf1 j C    C jfk j, the first k integrals under the supremum become
equal to zero, for arbitrary k  1. In this way we realize that the last requirement is
equivalent to
Z
lim sup jfn jp d < © :
n!1
fjfn j>gg

We will use this condition in a moment.

Proposition 6.7 (Vitali’s Convergence Theorem). Let f; f1 ; f2 ; : : : 2 Lp ./ for

some 1 p < 1. Then the following assertions are equivalent:
p
(i) fn ! f,
(ii0 ) f1 ; f2 ; : : : are equiintegrable and converge to f in measure.

Proof. (ii) ) (ii0 ): g WD 2jfj belongs to Lp ./. As in the preceding proof, let
1 n1 < n2 <    be a subsubsequence such that fn1 ; fn2 ; : : : converges a.e. to f.
Then jfnk jp 1fjfnk jgg converges a.e. to jfjp , and due to the dominated convergence
theorem (along subsubsequences and therefore along the whole sequence),
Z Z
jfn jp d ! jfjp d :
fjfn jgg

From (ii) it follows that

Z
jfn jp d ! 0 ;
fjfn j>gg

thus the equiintegrability.

62 6 Convergence

(ii0 ) ) (ii): Given © > 0, choose g 2 Lp ./ according to the equiintegrability

condition. If we replace g by g0 WD g C 2jfj, we may conclude as before that
Z Z
jfn jp d ! jfjp d :
fjfn jg0 g

This yields
ˇZ Z ˇ Z
ˇ ˇ
lim sup ˇ jfn j d  jfjp dˇ lim sup
p
jfn jp d < © :
n!1 n!1
fjfn j>g0 g

Letting © ! 0 we obtain (ii). t

u

Exercises
6.1 Prove the Riesz-Fischer Theorem in the case p D 1.
6.2 Let f1  f2     be a sequence of measurable functions which converge in measure to a
function f. Prove that the sequence converges a.e. to f.
6.3 Given measurable functions f; g W S ! R and a measure  on S, let
˚   
d.f; g/ WD inf © > 0 W  jf  gj > ©  © :

Prove: d is a pseudometric, that is, d is symmetric and satisfies the triangle inequality. d
metrizes the convergence in measure, that is, d.fn ; f/ ! 0 if and only if fn ! f in measure .
Uniqueness and Regularity of Measures
7

Uniqueness theorems in measure and integration theory serve to determine and

identify measures. The most important of those clarifies when two measures on
a ¢-algebra A which coincide on a generator E of A are actually equal on
all of ˚A. This is not
 always the case. On f1; 2; 3; 4g, for example, the system
E WD f1; 2g; f2; 3g generates the ¢-algebra of all subsets, and the two probability
measures  and  with weights 1 D 2 D 3 D 4 D 1=4 and 1 D 3 D 1=2,
2 D 4 D 0 coincide on E.
For this reason, there comes into play the new condition that E is a \-stable
system, meaning that

E; E0 2 E ) E \ E0 2 E

holds.

Proposition 7.1 (Uniqueness theorem for measures). Let E be a \-stable

generator of a ¢-algebra A on S, and let ,  be two measures on A. If

(i) .E/ D .E/ for every E 2 E,

(ii) .S/ D .S/ < 1 or .En / D .En / < 1 for some sets E1 ; E2 ; : : : 2 E
with En " S,

then  D .

In the case .S/ D .S/ < 1 one may readily adjoin S to the generator. Thus one
notices that in (ii) the second condition is more general.

© Springer International Publishing Switzerland 2015 63

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_7
64 7 Uniqueness and Regularity of Measures

Example (Lebesgue Measure)

The system of all finite intervals Œa; b/, a; b 2 Rd is a \-stable generator of the Borel ¢-algebra
Bd . It includes the intervals Œn; n/d , n  1; they have finite Lebesgue measure and their union
exhausts Rd . Therefore the Lebesgue measure œd is uniquely determined by its values on the
intervals. This proves a part of Proposition 3.2.

Example (Borel ¢-algebras)

The Borel ¢-algebra Bd on Rd (or, more generally, on a metric space S) is generated by the open
sets. Since S itself is open and the open sets form a \-stable system, by the theorem above a finite
measure  on Bd is uniquely determined by its values .O/ R for open set O  S.
Moreover,  is uniquely determined by all integrals f d of bounded continuous functions f.
Indeed, for any open O  S the distance between x and Oc ,

g.x/ WD d.x; Oc / D inffjx  zj W z 62 Og ;

is a continuous function (more precisely, we have jg.x/  g.y/j  jx  yj). The bounded
continuous functions fn .x/ WD min.1; ng.x// convergeR pointwise and monotonically towards 1O ,
and the dominated convergence theorem implies that fn d ! .O/. Therefore,  is uniquely
determined.

In order to prove the theorem we go back to the calculus of systems of sets with
which we have already become acquainted in Chap. 2.

Definition
A system D of subsets of a nonempty set S is called a Dynkin system,1 if

(i) S 2 D,
(ii) A 2 D ) Ac 2 D, S
(iii) A1 ; A2 ; : : : 2 D ) n1 An 2 D, whenever the sets A1 ; A2 ; : : : are
pairwise disjoint.

Dynkin systems are used (in contrast to ¢-algebras) as a technical device only. One
needs them in order to identify certain systems of sets as ¢-algebras. Here it comes
in handily that Dynkin systems inherit the property of \-stability from generators.
This is the essence of the following fact.

Proposition 7.2. Let D be a Dynkin system and A a ¢-algebra with a \-stable

generator E. Then E  D  A implies that D D A.

1
EVGENII DYNKIN , born 1924 in Leningrad, active in Moscow and Cornell. He made essential
contributions to Lie algebras and probability theory.
Regularity 65

In other words, a Dynkin system generated by a \-stable system of sets is

necessarily a ¢-algebra.

Proof. We may assume that D is the smallest Dynkin system containing E. We want
to prove that whenever E is a \-stable system of sets, then so is D. In order to show
this, for any D 2 D we consider the system

DD WD fA 2 D W A \ D 2 Dg :

DD , too, is a Dynkin system: Properties (i) are (iii) are obviously satisfied. Moreover,
for any A 2 DD the disjoint union .A \ D/ [ Dc and hence its complement Ac \ D
belongs to D. Thus property (ii) holds.
Now let E 2 E. We get E  DE , since E is \-stable by assumption. The
minimality of D yields DE D D, in other words: D \ E 2 D for every D 2 D,
E 2 E. This means that E  DD holds for every D 2 D. Again, the minimality of D
yields the equality DD D D, this time for every D 2 D. By definition, this equality
means that D is \-stable, as claimed.
Now, in D we may convert every countable union into a disjoint union, according
to the scheme
[ [
An D A1 [ .An \ Ac1 \    \ Acn1 / :
n1 n2

Therefore, D is a ¢-algebra. Since A is the smallest ¢-algebra containing E, the

assertion follows. u
t

Proof of the uniqueness theorem. Let En 2 E such that .En / D .En / < 1 and
En " S. By the properties of measures,
˚ 
Dn WD A 2 A W .A \ En / D .A \ En /

is a Dynkin system. Since E is \-stable, it follows that E  Dn  A and moreover,

due to the preceding proposition, that Dn D A. Therefore .A \ En / D .A \ En /
holds for each A 2 A. Passing to the limit n ! 1 we obtain the assertion. t
u

Regularity*

We now treat situation where the values .E/ of a measure on a generator E and its
other values .A/ are related in a more explicit manner. To this purpose we form
the expression
nX [ o
 .A/ WD inf .Em / W E1 ; E2 ; : : : 2 E; A  Em ; AS;
m1 m1
66 7 Uniqueness and Regularity of Measures

which is determined solely by the restriction of  on E. One thus considers finite

or countably infinite coverings of A with elements from the generator such that the
sum of their measures is as small as possible.

From the properties of measures (monotonicity and sub-¢-additivity) it follows that

.A/  .A/

for every A 2 A. Moreover,

.E/ D  .E/

for every E 2 E, because E is covered by itself. When is it possible to conclude

.A/ D  .A/ for other sets A 2 A, too? This question leads us, following
Carathéodory,2 to the following definition (it is a bit more general than usual: we
do not restrict ourselves to Borel ¢-algebras).

Definition
Let  be a measure on a ¢-algebra A, let E be a generator of A.  is called outer
regular (with respect to E), if
nX [ o
.A/ D inf .Em / W E1 ; E2 ; : : : 2 E; A  Em
m1 m1

holds for every A 2 A.

Some generators are immediately ruled out at this point, for example the generator
of the Borel ¢-algebra in R consisting of the intervals .1; x  R, which cannot
be used to cover perfectly arbitrary Borel sets. However, even when using more
suitable generators not every measure is outer regular.

2
CONSTANTIN CARATHÉODORY , 1873–1950, born in Berlin, active at several German univer-
sities, in Athens, and finally from 1924 in Munich. He made essential contributions to measure
and integration theory, the calculus of variations, complex analysis, and the axiomatic treatment
of thermodynamics. During the period 1920–1922 he acted as founding rector of the university at
Smyrna.
Regularity 67

Example
The counting measure .B/ WD #B on the Borel ¢-algebra in R, as well as the ¢-finite measure
.B/ WD #B \ Q, do not possess outer regularity with respect to the generator formed by the open
sets.

Usually one deals with measures which are outer regular with respect to a clearly
specified generator. This is true at least for measures constructed by Carathéodory’s
method. We will discuss this in Chap. 11.
When outer regularity is present, we may supplement the uniqueness theorem for
measures with the following comparison result, which is sometimes useful.

Proposition 7.3 (Comparison Theorem). Let  and  be measures on a ¢-

algebra A with generator E. If

.E/ .E/

for every E 2 E, and if  is outer regular w.r.t. E, then  .

S
Proof. Let A 2 A and A  m1 Em with Em 2 E. By the properties of measures
and the assumption,
X X
.A/ .Em / .Em / :
m1 m1

Taking the infimum over all coverings of A we conclude from the outer regularity
that .A/ .A/, as asserted. t
u

In particular, an outer regular measure is maximal among all measures that coincide
on E.
We now pursue the question how we may read off outer regularity from the
generator.

Proposition 7.4. Let E be a \-stable generator of a ¢-algebra A on S satisfying

¿ 2 E. Let  be a measure on A, and assume that there exist sets E1 ; E2 ; : : : 2 E
such that En " S and .En / < 1 for every n  1. If

.E0 n E/ D  .E0 n E/ for every E; E0 2 E with E  E0 ;

then  is outer regular w.r.t. E.

We will present the proof in Chap. 11.

68 7 Uniqueness and Regularity of Measures

Example (Semirings, Outer Regularity of the Lebesgue Measure)

A \-stable system E of sets with ¿ 2 E is called a semiring, ifSfor every E; E0 2 E with E 
0 0
P there exist disjoint0 sets E1 ; E2 ; : : : 2 E such that E n E D m1 Em . In this case we have
E
m1 .Em / D .E n E/ and consequently

 .E0 n E/ D .E0 n E/ :

If, in addition, S can be exhausted with elements En , n  1, from E with finite measure, then the
assumptions of the preceding proposition are satisfied and  is outer regular.
In particular this shows that the d-dimensional Lebesgue measure is outer regular with respect
to the generator E of the Borel ¢-algebra Bd consisting of all d-dimensional intervals

E D Œa; b/ ; a; b 2 Rd :
S  
Obviously, E is a semiring, and moreover Rd D m1 Œm; m/
d
and œd Œm; m/d < 1.

We now discuss the important case where the generator consists of all open subsets
of a metric space. Proposition 7.4 yields the following result.

Proposition 7.5. Let  be a measure on the Borel ¢-algebra of a metric space

S, and assume that there exist open sets E1 ; E2 ; : : :  S satisfying En " S and
.En / < 1 for each n  1. Then  is outer regular. More precisely, for every
Borel set B we have
˚ 
.B/ D inf .O/ W O  B; O is open

as well as
˚ 
.B/ D sup .A/ W A  B; A is closed :

Proof. First, let  be finite. We check the assumptions of the preceding proposition:
The open sets form a \-stable system of sets which includes the empty set.
Furthermore, let O  O0 be open. Then A WD Oc is closed, so for any null
sequence ©1 > ©2 >    > 0 of real numbers we have
1
\
Oc D A©n
nD1

where A© WD fx 2 S W d.x; y/ < © for some y 2 Ag (the open ©-neighbourhood of

A in the metric d). Since  is finite, using ¢-continuity we get that .O0 n O/ D
limn!1 .A©n \ O0 /. Moreover, O0 n O is covered by the open sets A© \ O0 , for each
© > 0. In conclusion,

 .O0 n O/ D .O0 n O/ ;
Regularity 69

therefore by the preceding proposition  is outer regular. The first assertion now
follows since a union of open sets is open. Passing to complements reveals that the
second assertion is equivalent to the first.
More generally, let now E1  E2     be open sets of finite measure exhausting
S. As we have just proved, the proposition applies to the finite measures . \ Em /.
For any Borel set B  S and any © > 0 there now exist closed sets Am and open
m satisfying Am  BS
sets OS  Om and .Om \ Em / < .Am \ Em / C ©2m . Setting
A WD m1 Am and O WD m1 Om \Em we get A  B  O and .O/ < .A/C©.
 Sn  Sn
Moreover, by ¢-continuity,  mD1 Am ! .A/ as n ! 1. Since mD1 Am is
closed and O is open, the assertion follows. t
u

For measures on Borel ¢-algebras one further expands the notion of regularity.

Definition
A measure  on a Borel ¢-algebra is called outer regular, if for every Borel set B,
˚ 
.B/ D inf .O/ W O  B; O is open :

 is called inner regular, if for every Borel set B,

˚ 
.B/ D sup .K/ W K  B; K is compact :

If both properties hold,  is called regular.

Proposition 7.6. Let  be a measure on a metric space S which satisfies the

assumptions of the preceding proposition. Suppose that S is a K¢ -set, that is,
there exist compact sets Kn  S, n  1 with Kn " S. Then  is regular.

Proof. Due to ¢-continuity, the assumption implies that .A \ Kn / ! .A/ as

n ! 1. When A is closed, the sets A \ Kn are compact. The assertion thus follows
from the preceding proposition. t
u

Example (Regularity of the Lebesgue measure)

Setting Kn D Œn; nd , the measure œd obviously satisfies the assumptions of the proposition.

For the further development of measure theory on topological spaces, Radon

measures play a prominent role. Radon measures are those regular measures on
Borel ¢-algebras, which are locally finite, that is, for which every x 2 S possesses
an open neighbourhood of finite measure. We do not dwell further on this issue.
70 7 Uniqueness and Regularity of Measures

Density of the Continuous Functions*

As an application of the regularity of the Lebesgue measure obtained just above, we

prove that the continuous functions are dense in the spaces Lp .œd /. Recall that the
support of a continuous function g W Rd ! R is defined as the topological closure
of the set fx 2 Rd W g.x/ ¤ 0g.

Proposition 7.7. Let f 2 Lp .œd /, where 1 p < 1. Then for every © > 0 there
exists a continuous function g W Rd ! R with compact support such that
Z
ˇ ˇ
ˇf.x/  g.x/ˇp dx < © :

Proof. We first consider the case f D 1B , where B  Rd is any Borel set with
œd .B/ < 1. Since the Lebesgue measure is regular, for every © > 0 there exist a
compact set K and an open set O satisfying K  B  O and œd .O/ < œd .K/ C ©.
Due to compactness, there exists a • > 0 such that jx  yj  • for all x 2 K; y … O.
The function

g.x/ WD .1  •1 d.x; K//C ; where d.x; K/ WD inffjx  yj W y 2 Kg

is continuous. Its support is contained in the closed •-neighbourhood of K and

therefore compact. g takes values between 0 and 1, on K the value 1 and on Oc
the value 0. This implies j1B  gjp 1OnK D 1O  1K and
Z
j1B  gjp dœd .O/  .K/ < © ;

which proves the assertion for f D 1B .

When Rf 2 Lp .œd / is arbitrary, forP
every © > 0 there exist natural numbers m; n
such that jf  f0 jp dœd < © for f0 D nkDn mk 1fk=mf<.kC1/=mg . The summands can
be approximated as described above by continuous functions with compact support.
In this manner, the assertion follows for general f (here the Minkowski inequality
helps). t
u

Exercises
7.1 Let  be a finite measure on S1  S2 (endowed with the product ¢-algebra). Let 1 and 2 be
the two image measures of  under the projection mappings  1 und  2 . Prove by means of an
example that  is not uniquely determined by 1 and 2 (even though  1 and  2 generate the
product ¢-algebra).
7.2 Let S be a finite set and D the system of all subsets consisting of an even number of elements.
When is D a Dynkin system? Is D then a ¢-algebra?
Density of the Continuous Functions 71

7.3 Let D be a Dynkin system and A; A0 2 D where A0  A. Prove that A n A0 also belongs to D .
Hint: Consider .A n A0 /c .
7.4 Let M be the smallest set of functions f from a metric space S to R with the properties
(i) fn 2 M; fn ! f pointwise ) f 2 M
(ii) M includes all continuous functions.
Prove that M equals the set of all Borel measurable functions.
Hint: In order to show that M is a vector space, for any given g 2 M, a; b 2 R consider
the set Mg;a;b WD ff 2 M W af C bg 2 Mg, at first for continuous g, and then for arbitrary
g 2 M. Prove in addition that D WD fB 2 B W 1B 2 Mg is a Dynkin system which contains
the open sets. Rˇ ˇ
7.5 Show that for any Lebesgue integrable function f W R ! R we have ˇf.xCt/f.x/ˇ dx ! 0
as t ! 0.
Hint: First consider the case of a continuous function with compact support.
7.6 Steinhaus’ Theorem Let B  R be a Borel set with œ.B/ > 0. Prove that BB WD fxy W
x; y 2 Bg contains an interval .•; •/ for some • > 0. 
Hint: Conclude from the preceding exercise that œ B \ .B C t/ ! œ.B/ as t ! 0.
Multiple Integrals and Product Measures
8

It is not particularly surprising that one may integrate measurable functions multiply
with respect to different variables. But it did irritate mathematicians like Cauchy
that the result may depend on the sequential arrangement of the integrals. When
computing derivatives this is usually not the case.
Only with the advent of Lebesgue’s integration theory it turned out that in
integration, too, the result usually does not depend on the order in which the
integrals are taken. This is the content of Fubini’s theorem, a core result of this
chapter. It has theoretical significance, but is also relevant when computing specific
integrals. Some important examples will be found in the text, others in the exercises.
Multiple integrals have many applications. We will construct product measures
and discuss the convolution and smoothing of functions. Finally, we will address a
more general situation: the integration of kernels.

Double Integrals

Multiple integration rests on the following fact.

Lemma. Let .S0 ; A0 /, .S00 ; A00 / be measurable spaces,  a ¢-finite measure on

A00 , and f W S0  S00 ! R
N C a nonnegative A0 ˝ A00 -B-measurable
N function. Then
the following assertions hold.

(i) The mapping y 7! Rf.x; y/ is A00 -B-measurable

N for each x 2 S0 . Conse-
quently, the integralR f.x; y/ .dy/ is well-defined for each x 2 S0 .
(ii) The mapping x 7! f.x; y/ .dy/ is nonnegative and A0 -B-measurable.
N

Proof. We restrict ourselves to the case of a finite measure  (the ¢-finite case is a
consequence). We consider the system D of those sets A 2 A0 ˝ A00 for which the

© Springer International Publishing Switzerland 2015 73

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_8
74 8 Multiple Integrals and Product Measures

function f D 1A satisfies the assertions (i) and (ii). By the properties of measurable
mappings and Proposition 4.7, D includes all unions of disjoint sets A1 ; A2 ; : : : from
D. Since  is assumed to be finite, D also includes the complement Ac for each A
in D. Finally, S0  S00 belongs to D, therefore D is a Dynkin system.
Moreover, A0  A00 2 D for all A0 2 A0 , A00 2 A00 , as one concludes from the
equality 1A0 A00 .x; y/ D 1A0 .x/1A00 .y/. Because those product sets form a \-stable
generator of the product ¢-algebra, it follows from Proposition 7.2 that D coincides
with the product ¢-algebra.
Let K be the system of all nonnegative A0 ˝ A00 -B-measurable
N functions f W
0 00
S  S ! R satisfying assertions (i) and (ii). By what we just proved and due to
N
the properties of measurable functions and integrals, K satisfies the conditions of
the monotonicity principle (Proposition 2.8). Therefore, K includes all nonnegative
A0 ˝ A00 -B-measurable
N functions f W S0  S00 ! R.N The assertion is proved. t
u

For ¢-finite measures  and  and nonnegative measurable functions f the double
integral
Z Z  Z Z 
f.x; y/ .dy/ .dx/ D f.x; y/ .dy/ .dx/ (8.1)
S0 S00

is thus well-defined, as is the double integral with the order of integrations

interchanged.
It is a fundamental fact that the order in which the integrations are performed
does not matter.

Proposition 8.1 (Fubini). For ¢-finite measures  and  on ¢-algebras A0 and

A00 and nonnegative measurable functions f W S0  S00 ! R
N C we have
Z Z  Z Z 
f.x; y/ .dy/ .dx/ D f.x; y/ .dx/ .dy/ :
Double Integrals 75

Proof. Once more we restrict ourselves to the case of finite measures. We consider
the system D of those sets A 2 A0 ˝ A00 for which our assertion holds with 1A in
place of f. Using the properties of integrals we again conclude that D is a Dynkin
system. For f.x; y/ WD 1A0 A00 .x; y/ D 1A0 .x/1A00 .y/ both integrals are equal to
.A0 /.A00 /, thus A0  A00 2 D, and D again coincides with the product ¢-algebra.
We form the system K of all nonnegative A0 ˝ A00 -B-measurable
N functions f W
0 00
S  S ! R for which the asserted equation holds. By what we just proved and due
N
to the properties of integrals, K satisfies the conditions in Proposition 2.8, and our
proposition follows. t
u

Example
We have
Z Z  Z Z
1 1
2 /y2
1 1
2 /z 1 
e.1Cx y dy dx D e.1Cx dz dx
0 0 0 0 2
Z
1 1
1 1
1  
D dx D arctan x 0 D
2 0 1 C x2 2 4

and
Z 1 Z 1  Z 1 Z 1 
2 /y2 2 2
e.1Cx y dx dy D ey e.xy/ y dx dy
0 0 0 0
Z 1 Z 1  Z 1 2
y2 z2 2
D e e dz dy D ez dz :
0 0 0

By Fubini’s Theorem the two expressions are equal, and we obtain the important formula
Z 1
2 p
ez dz D  :
1

This argument is due to Laplace, the formula itself had already been obtained earlier by Euler.2
1

So far we have written the double integrals with parentheses, in order to be precise.
In the following, according to common usage we will omit them.
We now introduce double integrals for measurable real-valued functions f.x; y/
which may take negative values, too. As in the case of single integrals this is not

1
PIERRE-SIMON LAPLACE, 1749–1827, born in Beaumont-en-Auge, active in Paris at the
École Militaire and École Polytechnique. His main research areas were celestial mechanics and
probability theory.
2
LEONARD EULER , 1707–1783, born in Basel, active in St. Petersburg and Berlin. He shaped
mathematics far beyond his century.
76 8 Multiple Integrals and Product Measures

always possible; here the additional assumption is

“
ˇ ˇ
ˇf.x; y/ˇ .dx/.dy/ < 1 ;

where according to the theorem of Fubini the order of integration is immaterial.

Whenever this holds in addition to measurability as above, f is called integrable.
But even under
R this assumption a small
R cliff has to be circumvented: It is still
possible that fC .x; y/ .dy/ as well as f .x; y/ .dy/ may attain the
R value 1 for
some values of x, so that we may not be able to form the integral f.x; y/ .dy/
as we did before. However the set of those x’s is a -null set. More precisely, the
following lemma holds.

0 00
’ ˇ Let fˇ W S  S ! R be measurable, le  and  be ¢-finite measures
Lemma.
and ˇf.x; y/ˇ .dx/.dy/ < 1. Then there exists a measurable function Of W
S0  S00 ! R with the following properties:

(i) It holds Of D f a.e., that is, Of.x; / D f.x; / -a.e. for -almost all x 2 S0 ,
0
(ii) y ! R y/ is -integrable for all x 2 S ,
7 Of.x;
(iii) x ! 7 Of.x; y/ .dy/ is -integrable.

Rˇ ˇ
Proof. Let A0 be the set of those x 2 S0 for which ˇf.x; y/ˇ .dy/ < 1. We
’ˇ ˇ
set Of.x; y/ WD f.x; y/1 A0 .x/. By assumption, ˇf.x; y/ˇ .dy/.dx/ < 1. By
Proposition 4.2 (iv),  .A0 /c D 0 follows. This yields assertion (i). (ii) holds due
’ˇ ˇ ’ˇ ˇ
to the choice of A0 . From ˇf.x; y/ˇ .dy/.dx/ D ˇOf.x; y/ˇ .dy/.dx/ we get
Z ˇZ ˇ “
ˇ O ˇ ˇ ˇ
ˇ f.x; y/ .dy/ˇ.dx/ ˇOf.x; y/ˇ.dy/.dx/ < 1 ;

and thus (iii) holds. t

u
R
In particular, the lemma implies that f.x; y/ .dy/ exists for -almost all x 2 S0 .
’
For Of, due to (ii) and (iii) we can form the double integral Of.x; y/ .dy/.dx/ (with
the order of integration as indicated!). If Qf is another measurable function with the
properties stated in the lemma, then applying Proposition 4.2 (ii) twice it follows
according to property (i) that
“ “
Qf.x; y/ .dy/.dx/ D Of.x; y/ .dy/.dx/:
Double Integrals 77

’ˇ ˇ
Thus if we assume that ˇf.x; y/ˇ .dx/.dy/ < 1, the integral
“ “
f.x; y/ .dy/.dx/ WD Of.x; y/ .dy/.dx/

is well-defined. Its value is finite. In an analogous manner one obtains the double
integral in reverse order.
The properties of the double integral again R result from decomposing
R C f into
its positive and negative parts. One has Of.x; y/ .dy/ D Of .x; y/ .dy/ 
R 
Of .x; y/ .dy/. If f is integrable, those integrals viewed as functions of x are -
integrable. The linearity of the integral yields
“ “ “
Of.x; y/ .dy/.dx/ D OfC .x; y/ .dy/.dx/  Of .x; y/ .dy/.dx/ :

In addition, OfC and fO are a.e. equal to fC and f , therefore we obtain – in this case
not by the definitions, but via the detour of integrating Of – the equation
“ “ “
C
f.x; y/ .dy/.dx/ D f .x; y/ .dy/.dx/  f .x; y/ .dy/.dx/ :

In the right-hand side we may apply the standard integration rules and thus arrive
at the properties of double integrals. In particular, we obtain a second version of
Fubini’s Theorem.

Proposition 8.2 (Fubini). Let the measurable

’ real-valued function f W S0 S00!R
N
and the ¢-finite measures ,  satisfy jf.x; y/j .dy/.dx/ < 1. Then
“ “
f.x; y/ .dy/.dx/ D f.x; y/ .dx/.dy/ :

Example (Reordering of absolutely convergent series)

P P
For any doubly-indexed sequence f.m; n/ of real numbers with m1 n1 jf.m; n/j < 1
(absolute convergence) it holds that
XX XX
f.m; n/ D f.m; n/ :
m1 n1 n1 m1

We may view this as a particular case of Fubini’s Theorem, applied to the ¢-finite counting
measures .A/ D .A/ D #A, A  N. The requirement of absolute convergence cannot just
be omitted, as the example f.m; m/ D 1, f.m; m C 1/ D 1 and f.m; n/ D 0 otherwise shows.
78 8 Multiple Integrals and Product Measures

Here we have
XX XX
f.m; n/ D 1 ¤ 0 D f.m; n/ :
m1 n1 n1 m1

O y/ D f.x; y/.
As in this example, in many specific instances one may choose f.x;
The above-mentioned problem concerning the existence of integrals does not arise.
Multiple integrals may easily be reduced to double integrals. Details are left to
the reader.

Product Measures

Double integrals give rise to new measures on the product ¢-algebra.

Proposition 8.3. Let  and  be ¢-finite measures on the ¢-algebras A0 and A00 .
Then
“
 .A/ WD 1A .x; y/ .dy/.dx/ ; A 2 A0 ˝ A00 ;

defines a measure   on the product ¢-algebra. We have

Z “
f d  D f.x; y/ .dy/.dx/

for every measurable function f  0.

Proof. Obviously  .¿/ D 0 holds, and the ¢-additivity results from applying
Proposition 4.7 twice. In order to prove the second assertion we consider
n Z “ o
K WD f  0 W f d  D f.x; y/ .dy/.dx/ :

Due to the definition of  , K includes all elements A of the product ¢-algebra.

By virtue of the rules of integration, the other two conditions of the monotonicity
principle (Proposition 2.8) are satisfied, too. Therefore K includes all nonnegative
measurable functions, and the assertion follows. t
u

According to the exposition in the foregoing section it also holds that

Z “
f d  D f.x; y/ .dx/.dy/ ;
Product Measures 79

that is one can reverse the order of integration.

One calls   the product measure of  and , and writes

 D˝ or  .dx; dy/ D .dx/ ˝ .dy/ :

f.x; y/ is  ˝ -integrable if and only if

Z “
ˇˇ ˇ ˇ
ˇfˇ d. ˝ / D ˇf.x; y/ˇ .dx/.dy/ < 1 :

In this case, integrals w.r.t. the product measure can be reduced to double integrals,
the order of integration being arbitrary. This fact, too, is called the Fubini’s
Theorem.

I Remark A set A 2 A0 ˝RA00 is a  ˝ -null set if and only if the double integral
’
1A .x; y/ .dy/.dx/ D .Ax / .dx/ equals 0, employing the “cross section”
Ax WD fy 2 S00 W .x; y/ 2 Ag. In other words, A is a  ˝ -null set if and only if Ax
is a -null set for -almost all x 2 S0 . This is in complete accordance with the a.e.
notion used in item (i) of the foregoing lemma.

The following proposition illuminates why one speaks of “product measures”.

Proposition 8.4. Let  and  be ¢-finite measures. Then

. ˝ /.A0  A00 / D .A0 /  .A00 /

for all A0 2 A0 , A00 2 A00 . These product formula determines  ˝  uniquely.

Proof. The product formula results from the double integral of the function
1A0 A00 .x; y/ D 1A0 .x/1A00 .y/. The other assertion follows from the uniqueness
theorem for measures, as  and  are assumed to be ¢-finite, and the measurable
sets of the form A0  A00 form a \-stable generator of the product ¢-algebra. t
u
80 8 Multiple Integrals and Product Measures

Example (Lebesgue Measure)

We recall that the Borel ¢-algebras in Rd D Rd1  Rd2 (thus d D d1 C d2 ) satisfy Bd D Bd1 ˝ Bd2 .
The Cartesian product Œa1 ; b1 /  Œa2 ; b2 /  Rd1  Rd2 of semi-open intervals is again a semi-open
interval, and
     
œd Œa1 ; b1 /  Œa2 ; b2 / D œd1 Œa1 ; b1 /  œd2 Œa2 ; b2 / :
   
Consequently, œd Œa; b/ D œd1 ˝ œd2 Œa; b/ holds for each Œa; b/  Rd . Since these semi-open
intervals form a \-stable generator of the Borel ¢-algebra, the uniqueness theorem for measures
implies that

œd D œd1 ˝ œd2 :

In this manner Lebesgue integrals may be reduced to multiple integrals, and we obtain the formula
Z Z Z
f dœd D  f.x1 ; : : : ; xd / dx1 : : : dxd :

Example (Volume of the d-dimensional unit ball)

We want to determine the volume

vd WD œd .B1 /

of the unit ball B1 WD fx 2 Rd W jxj  1g in Rd through reduction to the -function

Z 1
.t/ WD ez zt1 dz ; t>0:
0

For this purpose we consider the image measure  D ®.œd / of the Lebesgue measure under the
mapping ® W Rd ! RC given by ®.x/ D jxj2 . Due to the transformation formula for integrals in
Chap. 4 we have
Z Z
2
ey .dy/ D ejxj œd .dx/ :

Let us calculate the two integrals. Since œd is a product measure, a multiple application of Fubini’s
Theorem yields that
Z Z 1 Z 1 Z 1 d
2 2 2 2
ejxj œd .dx/ D  ex1    exd dx1 : : : dxd D eu du :
1 1 1
   
For the other integral we use the formula  Œ0; z D œd z1=2 B1 D zd=2 vd , z > 0. From Fubini’s
R 1
Theorem and since ey D y ez dz it follows that
Z Z 1 Z 1
ey .dy/ D ez 1fyzg dz .dy/
0 0
Z Z Z d 
1 1 1  
D ez 1fyzg .dy/ dz D ez  Œ0; z dz D vd  C1 :
0 0 0 2
Convolution and Smoothing 81

The comparison of the two integrals is revealing already in the cases d D 1 and 2: As is well
R 1 we2 have v2 D   (area of the unit circle), and .2/ D 1 (partial integration). It follows that
known,
.p1 eu du/2 D  , a formula which we have derived already. From v1 D 2 it then follows that
  D 2.3=2/.
In conclusion

 d=2
vd D  :
 d=2 C 1

We may evaluate the -function inductively using the formula .t C 1/ D t.t/ p (partial
integration)pas well as the values just obtained, namely .2/ D 1 and .3=2/ D  =2 resp.
.1=2/ D  . Details are left to the reader.

Convolution and Smoothing*

We apply multiple integration to a particular situation. Let g; h W Rd ! R be

Lebesgue integrable functions, thus
Z Z
ˇ ˇ ˇ ˇ
ˇg.x/ˇ dx ; ˇh.x/ˇ dx < 1 :

2d
R WD g.x  y/h.y/ is Borel measurable on R , and since we have
RThe function f.x; y/
jg.x  y/j dx D jg.x/j dx we get
“ Z Z
ˇ ˇ ˇ ˇ ˇ ˇ
ˇg.x  y/h.y/ˇ dxdy D ˇg.x/ˇ dx ˇh.y/ˇ dy < 1 : ()

In the section above on double integrals we have seen that the convolution integral
Z
g.x  y/h.y/ dy

then exists for almost all x, resp. it defines a Lebesgue integrable function uniquely
for a.e. x 2 Rd . Moreover, the convolution integral remains unchanged if g or h are
modified on Lebesgue null sets. Therefore it is natural to understand g; h and their
convolution integral as equivalence classes of measurable functions, as elements of
L1 .œd /. We thus give the following definition:

Definition
Let g; h 2 L1 .œd /. Their convolution g  h 2 L1 .œd / is defined as
Z
g  h.x/ WD g.x  y/h.y/ dy :
82 8 Multiple Integrals and Product Measures

By () we have

kg  hk1 kgk1 khk1 :

Convolutions arise in various situations.

Example
Let a > 0, and let k W Œ0; 1/ ! R be continuous. The solution of the inhomogeneous linear
differential equation

f0 .x/ D af.x/ C k.x/ ; x0;

together with the boundary condition f.0/ D 0 is given by

Z x Z
f.x/ D k.y/ea.xy/ dy D g.x  y/h.y/ dy
0

where g.x/ WD eax , h.x/ WD k.x/ for x  0 and g.x/ D h.x/ WD 0 for x < 0. One can verify this
directly by differentiation.

Convolution is important also because of its nice algebraicR properties. When we

substitute y 7! x  y, the convolution integral changes into g.y/h.x  y/ dy, and
consequently

ghDhg:

Moreover,

.g  h/  k D g  .h  k/ ; g  .h C k/ D g  h C g  k :

We leave the proof as an exercise for the reader.

We now explain how functions can be smoothed by convolution, and thus show that
the smooth functions are dense in Lp .œd /. For every • > 0 we choose a so-called
“mollifier” ›• W Rd ! R with the following properties:

(a) ›• is nonnegative and differentiable of infinite order,

(b) ›
R • .x/ D 0 for jxj  •,
(c) ›• .x/ dx D 1.

We may take, for example, ›• .x/ WD •d ›.•1 x/ with

(  
c exp  .1  jxj2 /1 ; if jxj < 1 ;
›.x/ WD
0; if jxj  1 ;

where c > 0 is a suitably chosen normalization constant.

Convolution and Smoothing 83

R
For any measurable function f W Rd ! R we define, whenever jfjp dœd < 1
for some p  1, the functions

f• WD f  k• ;

or
Z
f• .x/ D f.y/k• .x  y/ dy :

The integral exists in the case p D 1 because ›• is bounded, and in the case
p > 1 as a consequence of Hölder’s inequality. Applying Proposition 5.9 on the
differentiation of integrals we see that f• has derivatives of any order.

Proposition 8.5 (Smoothing Theorem). Let 1 p < 1. Then for f 2 Lp .œd /

we have

kf  f  ›• kp ! 0

as • ! 0.

Proof. First we prove the result for continuous g with compact support. As is
known, g then is uniformly continuous. For any given © > 0 it holds for sufficiently
small • > 0 that jg.x/  g.x  y/j © for each jyj •. Consequently,
Z
ˇ ˇ ˇ ˇ
ˇg.x/  g  ›• .x/ˇ ˇg.x/  g.x  y/ˇk• .y/ dy © :

Therefore, g  ›• converges uniformly to g. Moreover, g.x/ and thus g  ›• .x/ are

nonzero in a bounded domain only. It follows that kg  ›•  gkp ! 0 as • ! 0, as
one checks with the aid of the dominated convergence theorem.
The transition from a continuous gR with compact support to arbitary f 2 Lp .œp /
we achieve with an estimate. Since k• dœd D 1, Jensen’s inequality shows that
(note that t 7! jtjp is convex for p  1)
Z ˇZ ˇp “
ˇ ˇ ˇ ˇ
kf  ›• kpp D ˇ f.x  y/›• .y/ dyˇ dx ˇf.x  y/ˇp ›• .y/ dydx
Z Z
ˇ ˇ 
D ˇf.x  y/ˇp dx ›• .y/ dy D kfkp :
p

According to Proposition 7.7, for any given © > 0 we choose a continuous g with
compact support such that kf  gkp < ©. It follows that

kf  f  ›• kp kf  gkp C kg  g  ›• kp C k.g  f/  ›• kp 2© C kg  g  ›• kp :
84 8 Multiple Integrals and Product Measures

Letting • ! 0 we get lim sup•!0 kf  f  ›• kp 2©, and letting © ! 0 we obtain

the assertion. t
u

Kernels*

We discuss a generalization
R R which is important in probability theory: within the
double integral . f.x; y/ .dy//.dx/ one allows the measure  to depend on x.
In order that the outer integral be defined, one needs a regularity assumption.

Definition
Let .S0 ; A0 /, .S00 ; A00 / be measurable spaces. A family
 
 D .x; dy/ x2S0

of finite measures .x; dy/ on A00 is called a kernel of .S0 ; A0 / to .S00 ; A00 /, if for
every A00 2 A00 the function

x 7! .x; A00 /

is A0 -B 1 -measurable.

Lemma. Let  be a kernel of .S0 ; A0 / to .S00 ; A00 / and let f W S0  S00 ! R

N C be a
nonnegative A0 ˝ A00 -B-measurable
N function. Then the function
Z
x 7! f.x; y/ .x; dy/

is A0 -B-measurable.
N

Proof. As before, we consider the system D of sets A 2 A0 ˝ A00 for which the
function f D 1A satisfies the assertion. By the properties of measurable functions
and Proposition 4.7, D includes the union of disjoint sequences A1 ; A2 ; : : : as well
as complements of sets in D. Finally, S0  S00 is included in D by the measurability
properties of kernels, therefore D is a Dynkin system.
0 00 0 0 00 00
R we have A  A 2 D for all A 200 A , A 2 A , as one sees from the
Moreover,
equation 1A0 A00 .x; y/ x .dy/ D 1A0 .x/.x; A /. Because these product sets form
a \-stable generator of the product ¢-algebra, we conclude from Proposition 7.2
that D coincides with the product ¢-algebra.
The assertion now follows in the same manner as in the proof of the lemma at
the beginning of this chapter. t
u
Kernels 85

Thus one again may form double integrals. For reasons of clarity we use the notation
Z Z
.dx/ .x; dy/ f.x; y/ :

Once more,
Z Z
A 7! .dx/ .x; dy/ 1A.x; y/

defines a measure on the product ¢-algebra, denoted as

˝ resp. .dx/ ˝ .x; dy/ :

It is an interesting question which measures can be obtained by this procedure,

that is, under which conditions can a given measure   on the product ¢-algebra be
expressed as   D  ˝  for some measure  and some kernel . One speaks of
disintegration of the measure  . On Borel ¢-algebras this is possible under rather
general conditions. We do not dwell further on this subject.

Exercises
8.1 Prove and comment on the following observation, due to Cauchy: the double integrals
Z Z Z Z
x2  y2 x2  y2
dxdy ; dydx
.0;1/ .0;1/ .x2 C y2 /2 .0;1/ .0;1/ .x2 C y2 /2

are well-defined and different from each other.

Hint: .x2  y2 /.x2 C y2 /2 D @2 arctan.x=y/=@x@y.
8.2 Let  be the counting measure on R, that is, .B/ WD #B for Borel sets B  R, and let D be
the diagonal in R2 , D D f.x; y/ 2 R2 W x D yg. Prove and comment:
“ “
1D .x; y/ œ.dx/.dy/ ¤ 1D .x; y/ .dy/œ.dx/ :

8.3 Let 1 .dx/ D h1 .x/ 1 .dx/, 2 .dy/ D h2 .y/ 2 .dy/. What is the density of 1 ˝ 2 w.r.t.
1 ˝ 2 ?
8.4 Integrals “measure the area below a function” Let f W S ! R N C be measurable. Prove the
formula
Z Z 1
f d D  ˝ œ.Af / D .f > t/ dt ;
0

where Af D f.x; t/ 2 S  R WR0  t < f.x/g.

Hint: It holds that f.x/ D 1f0t<f.x/g dt. Concerning the measurability of Af compare
Exercise 2.7.
86 8 Multiple Integrals and Product Measures

8.5 We want to derive the formula (improper Riemann integral)

Z 1
sin x  
dx D :
0 x 2
RaR1
First, prove that 0 0 x exp.xy/ dydx < 1 for every 0  a < 1. Conclude that
Z a Z 1 Z a
sin x
dx D sin x exy dxdy :
0 x 0 0
Ra
Compute the inner integral, most simply as the imaginary part of 0 exp..i  y/x/ dx, and pass
to the limit a ! 1.
8.6 In the same manner prove that for a > 0 one has
Z 1
ex  eax
dx D log a :
0 x

8.7 The Beta function The Beta function is defined as

Z 1
B.x; y/ WD sx1 .1  s/y1 ds ; x; y > 0 :
0
R1
We want to express it with the aid of the Gamma function .x/ WD 0 tx1 et dt, x > 0.
Prove that
Z 1 Z t 
.x C y/B.x; y/ D ux1 .t  u/y1 du et dt :
0 0

Using Fubini’s Theorem and a shift of variable conclude that

.x/.y/
B.x; y/ D :
.x C y/

8.8 Prove that .g  h/  k D g  .h  k/.

Absolute Continuity
9

In this chapter we discuss under which circumstances measures and functions have
densities.
In the first situation considered, two measures  and  are given on some ¢-
algebra, and we ask for conditions under which a measureable function h exists
such that d D h d, that is,
Z
.A/ D h d
A

holds for all measurable sets A. In the second situation, a function f W Œa; b ! R is
given and one asks for the existence of a Borel measurable function h W Œa; b ! R
such that
Z x
f.x/ D h.z/ dz
a

holds for all x 2 Œa; b.

The two questions are related. This becomes clear when one chooses for  the
Lebesgue measure, restricted to the interval
  Œa; b, and for  another measure on the
Borel sets in Œa; b. Setting f.x/ WD  Œa; x and A WD Œa; x, the first equation turns
into the second.
Therefore, it is possible to treat both problem statements simultaneously. How-
ever, we want to consider two different methods; for measures a “global” exhaustion
procedure, for functions a “local” method which is more involved, but provides a
connection to differentiation and to the fundamental theorem of calculus.
For measures, one may easily formulate a necessary condition for the existence
of a density. One obviously has to require that

.A/ D 0 ) .A/ D 0 :

© Springer International Publishing Switzerland 2015 87

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_9
88 9 Absolute Continuity

We will show that for ¢-finite measures this condition is also sufficient. As we will
see in the exercises, the following seemingly stronger condition is equivalent:

8© > 0 9• > 0 8A 2 A W .A/ • ) .A/ © :

For functions, we will consider an analogous condition.

Absolute Continuity and Singularity of Measures

This section is about the following pair of complementary notions.

Definition
Let  and  be two measures on a ¢-algebra A.

(i)  is called absolutely continuous w.r.t. , written as


;

if, for each A 2 A, .A/ D 0 implies that .A/ D 0. If  and  are both
absolutely continuous w.r.t. each other,  and  are called equivalent.
(ii)  and  are called mutually singular, written as

? ;

if there exists an A 2 A such that .A/ D 0 and .Ac / D 0.

Absolute continuity can be characterized as follows.

Proposition 9.1 (Radon1 -Nikodym2 Theorem). Let  and  be ¢-finite mea-

sures on a ¢-algebra A. Then the following assertions are equivalent:

(i) 
,
(ii) d D h d for some measurable function h W S ! RC .

The density h then is -a.e. finite and -a.e. unique.

1
JOHANN RADON , 1887–1956, born in Tetschen, active a.o. in Hamburg, Breslau, and Vienna. His
working areas were measure and integration theory, functional analysis, calculus of variations, and
differential geometry.
2
OTTON NIKODÝM, 1887–1974, born in Zablotow, active in Kraków, Warsaw and at Kenyon
College, Ohio. He worked on measure theory and functional analysis.
Absolute Continuity and Singularity of Measures 89

One may after all drop the requirement that  is ¢-finite. Concerning  this is not
the case (compare Exercise 9.2).
From the various proofs available we choose a lucid classical approach. It uses a
result which moreover is of independent interest.

Proposition 9.2 (Hahn Decomposition3 ). Let  and ¡ be finite measures on a

¢-algebra A. Then there is a measurable set A , its complement being A WD
S n A , such that

.A/ ¡.A/ for all A  A ;

.A/  ¡.A/ for all A  A :

Proof. We set •.A/ WD .A/  ¡.A/ for A 2 A. Then like a measure, • satisfies
•.¿/ D 0 and is ¢-additive, but •.A/ may be negative. For later purposes we allow
•.A/ to assume the value 1, but not the value 1.

(i) A measurable set N  S we term negative if •.A/ 0 for all A  N. We

want to construct A as the largest possible negative set. S It fits our purpose
that whenever
S N 1 ; N 2 ; : : :  S are negative, then so is k1 Nk . Namely, for
c c
A  k1 Nk the set AP k WD A \ N k \ N 1 \    \ N k1 is a subset of Nk , thus
•.Ak / 0 and •.A/ D k1 •.Ak / 0 follows.
(ii) First, in the case •.S/ < 1 we construct a negative subset N  S satisfying
•.Nc /  0. We obtain N by removing successively certain disjoint measurable
sets Bk , k  1, with •.Bk /  0, for which •.Bk / is sufficiently large. We set
B1 WD ¿. Having chosen B1 ; : : : ; Bk we let sk be the supremum of all values
•.A/, taken over those measurable sets A which are disjoint with B1 ; : : : ; Bk .
It holds sk  •.¿/ D 0. Now we choose the set BkC1 as disjoint to B1 ; : : : ; Bk ,
and such that •.BkC1 /  sk =2 in case sk < 1, in particular BkC1 D ¿ in case
sk D 0, and •.BkC1 /  1Sin case sk D 1. P
Let now N WD S n k1 Bk . Then we have •.Nc / D k1 •.Bk /, thus
•.Nc /  0. From •.N/ C •.Nc / D •.S/ < 1 it follows that •.Nc / < 1.
This implies •.Bk / ! 0 and thus sk ! 0. If A  N, then A is disjoint with
B1 ; : : : ; Bk , and therefore •.A/ sk . Passing to the limit k ! 1 we obtain
•.A/ 0. Thus N is negative.
(iii) More generally, we claim that if S0  S is measurable and •.S0 / < 1, then
there exists a negative set N0  S0 such that •.S0 n N0 /  0 and therefore
•.N0 / •.S0 /. This follows from (ii) when we consider the restriction •0 of •
to the measurable subsets of S0 .

3
HANS HAHN , 1879–1934, born in Vienna, active in Chernovitz, Bonn, and Vienna. He made
essential contributions to functional analysis, measure theory, and real analysis. He played a
leading role in the Vienna Circle, a group of positivist philosophers and scientists.
90 9 Absolute Continuity

(iv) Let now ’ WD inff•.A/ W A 2 Ag, thus ’ 0. Let Sk  S, k  1, be

measurable subsets satisfying •.Sk / < 1 and •.Sk / ! ’. According to (iii),
there exist negative sets NkS Sk such that •.Nk / •.Sk /. It follows that
•.Nk / ! ’. We set A WD k1 Nk . By (i) the set A is negative. Therefore
we have •.A / D •.A n Nk / C •.Nk / •.Nk / for all k and thus •.A / D ’.
It follows that ’ > 1. We now finish the proof as follows:
Let A  A . Since A is negative, we have •.A/ 0, resp. .A/ ¡.A/.
This is one part of the assertion. On the other hand, let A  S n A . Then we
have •.A/ D •.A [ A /  •.A /  ’  ’ D 0. This is the other part of the
assertion.
t
u

Proof of the Radon-Nikodym Theorem. The implication (ii) ) (i) is obviously true.
To prove (i) ) (ii) let us first assume that  and  are finite. We consider the set
n Z o
F WD f  0 W f d .A/ for all A 2 A
A

of measurable functions and set

Z
“ WD sup f d :
f2F

As  is finite, we have “ .S/ < 1. We want to obtain the sought-after density h

as an element of F satisfying
Z
h d D “ :

To this end we claim that max.f; f0 / 2 F whenever f; f0 2 F . Indeed, the latter gives
Z Z Z
max.f; f0 / d D f d C f0 d
A A\fff0 g A\ff<f0 g
   
 A \ ff  f0 g C  A \ ff < f0 g D .A/ :
R
If f1 ; f2 ; : : : are elements of F satisfying fn d ! “, we may assume without loss
of generality that 0 f1 f2    , otherwise replace fn with max.fR1 ; : : : ; fn /. For
h WD supn fn using monotone convergence we R obtain that h 2 F and h d D “.
For any A0 2 A we therefore have that A0 h d .A0 /. To prove the reverse
inequality we consider, for any given © > 0, the finite measure ¡ having the density
d¡ D .h C ©1A0 / d, and additionally, according to the foregoing proposition, the
Hahn decomposition A , A for  and ¡. On A ,  is dominated by ¡, and thus we
Absolute Continuity and Singularity of Measures 91

get the estimate

Z
.A0 \ A / ¡.A0 \ A / ¡.A0 / D h d C ©.A0 / :
A0

On A , ¡ remains below . Therefore, g WD h C ©1A0 \A belongs to F , since for

any measurable A we have
Z Z
g d D ¡.A \ A / C h d .A \ A / C .A \ A / D .A/ :
A A\A

R
From g d D “ C ©.A0 \ A / it follows that .A0 \ A / D 0 and, since 
,
we get .A0 \ A / D 0. All in all, it follows that
Z
0
.A / h d C ©.A0 / ;
A0

and letting © ! 0 we obtain the desired inequality.

Thus, d D h d. In particular, .h D 1/ D 1  .h D 1/. Since  is finite,
we also get h < 1-a.e. The -a.e. uniqueness was established above.
These results easily carry over to ¢-finite measures if we exhaust S by a sequence
of sets of finite measure. t
u

The Radon-Nikodym Theorem has a number of applications. In probability theory,

the following application is of particular importance.

Example (Conditional Expectation)

Let  be a finite measure on the ¢-algebra A and h  0 be a -integrable function. Then the
measure , given as d D h d, is finite, too. Let moreover A0 be a ¢-algebra contained in A.
Restricting  and  to A0 we obtain finite measures 0 and 0 . Since 
 we have 0
0 . By
the Radon-Nikodym Theorem, there exists a A0 -measurable function h0  0 such that d0 D h0 d0 .
This means that
Z Z
h d D h0 d
A0 A0

holds for all A0 2 A0 . We thus have adapted the measurability of the density to the ¢-algebra A0 .
In probability theory, h0 is called the conditional expectation of h given A0 ; it is -a.e. unique.
The case of an arbitrary -integrable function h can be treated by decomposing it into its positive
and negative part. In Chap. 12 we will make aquaintance with a different approach to conditional
expectations which is based on the completeness of the space L2 ./ instead of the Radon-Nikodym
Theorem.

Another application of the theorem is concerned with the decomposition of a

measure into an absolutely continuous and a singular part.
92 9 Absolute Continuity

Proposition 9.3 (Lebesgue Decomposition). Let  and  be ¢-finite measures

on a ¢-algebra A. Then there exist measures a and s with the properties:

(i)  D a C s ,
(ii) a
 and s ? .

a and s are uniquely determined.

Proof. Obviously  is absolutely continuous w.r.t. the measure  C . According to

the Radon-Nikodym Theorem,  has a density h  0 w.r.t.  C , that is,
Z Z
.A/ D h d C h d
A A

holds for any A 2 A. We set

   
a .A/ WD  A \ fh > 0g ; s .A/ WD  A \ fh D 0g :
R
Then (i) obviously holds. If A is a -null set, A h d
 D 0 follows. Therefore, we
have h1A D 0 -a.e. resp. 1A\fh>0g D 0 -a.e. or  A \ fh R> 0g/ D 0. This shows
that a
. In addition, s .h > 0/ D 0 and .h D 0/ D fhD0g h d. C / D 0,
therefore s ?  holds.
Let now  D 0a C 0s be another decomposition with the properties (i) and
(ii). Then there are measurable sets N; N0 satisfying s .N/ D 0s .N0 / D 0, and
whose complements are -null sets. Therefore we get a .Nc / D a .N0 /c D 0. For
measurable sets A it follows that

a .A/ D a .A \ N \ N0 / D .A \ N \ N0 / :

An analogous equality holds for 0a , and so a D 0a .

In the case .A/ < 1 we obtain from (i) that s .A/ D 0s .A/. Since  is
assumed to be ¢-finite, we also get s D 0s . u
t

A Singular Measure on the Cantor Set*

We consider measures  which are singular with respect to the Lebesgue measure
œ on R. An example is given by the Dirac measure  D •x whose whole mass
is concentrated in x 2 R. A point x for which .fxg/ > 0 is called an atom
of . Discrete measures composed from countably many atoms are obviously
singular with respect to the Lebesgue measure. It is less obvious that there also exist
measures which are singular to the Lebesgue measure but do not possess atoms.
A Singular Measure on the Cantor Set 93

In order to construct such a measure we consider a variant of the Cantor set,4

a subset C of the semi-open interval Œ0; 1/ within R. Geometrically C is easily
accessible: one decomposes the interval C0 WD Œ0; 1/ into parts Œ0; 1=3/, Œ1=3; 2=3/
and Œ2=3; 1/ of equal length and removes the middle part:

C1 WD Œ0; 1=3/ [ Œ2=3; 1/ :

With the two remaining intervals one proceeds analogously:

C2 WD Œ0; 1=9/ [ Œ2=9; 1=3/ [ Œ2=3; 7=9/ [ Œ8=9; 1/

[ [ 
D a1 =3 C a2 =9; a1 =3 C a2 =9 C 1=9 :
a1 2f0;2g a2 2f0;2g

In a picture this looks as

Having removed the middle interval n times we arrive at the set

[ [ hX
n X
n 
Cn WD  ak 3k ; ak 3k C 3n ;
a1 2f0;2g an 2f0;2g kD1 kD1

thus C1  C2     . We define our version of the Cantor set as the result after 1
many steps,
1
\
C WD Cn :
nD1

(If we construct C from closed instead of semi-open intervals, as is the common

procedure, we obtain the usual Cantor set which moreover is compact. Here such
subtleties are irrelevant; our procedure avoids having to deal with nonunique b-nary
representations of numbers.)

4
GEORG CANTOR , 1845–1918, born in St. Petersburg, active in Halle. He was the founder of set
theory.
94 9 Absolute Continuity

C is a null set, indeed we always remove one third, so that œ.CnC1 / D 23 œ.Cn /. It
follows that œ.Cn / D .2=3/n and

œ.C/ D 0 :

To describe C more precisely, we utilize the b-nary representation (to the basis b D
2; 3; : : :)
1
X
xD xk bk
kD1

of numbers x 2 Œ0; 1/. We assume that the sequence x1 ; x2 ; : : : belongs to

˚ 
Db WD .xk /k1 W xk 2 f0; 1; : : : ; b  1g ; xk ¤ b  1 1-often :

As we know, in this way we achieve uniqueness in the representation of x. Then

Œ0; 1=3/, Œ1=3; 2=3/, Œ2=3; 1/ are the regions for which x in ternary representation
(b D 3) has the coefficient x1 equal to 0, to 1, and to 2. Thus it holds that
nX o
C1 D xk 3k W .xk /k1 2 D3 ; x1 ¤ 1
k1

and iteratively
nX o
Cn D xk 3k W .xk /k1 2 D3 ; x1 ; : : : ; xn ¤ 1
k1

and finally
nX o
CD xk 3k W .xk /k1 2 D3 ; x1 ; x2 ; : : : ¤ 1 :
k1

Therefore C not only is nonempty, but has the same cardinality as the interval Œ0; 1/:
the assignment
1
X 1
X
y WD yk 2k $ 2yk 3k DW ®.y/ ; .yk /k1 2 D2 ;
kD1 kD1

gives rise to a bijection ® W Œ0; 1/ ! C. It is strictly monotone, since y < y0 holds if

and only if there exists an n such that yn < y0n and yk D y0k for any k < n, and this
implies that ®.y/ < ®.y0 /.
Differentiability 95

The singular measure  we are looking for can now be found as the image measure
of the Lebesgue measure (restricted to Œ0; 1/) under the mapping ®, so
 
.B/ WD œ ®1 .B/

for Borel sets B  Œ0; 1/. As œ does not possess atoms and ® is injective,  too has
no atoms. Their mutual singularity follows from œ.C/ D 0, .Cc / D 0.

Differentiability*

We now change over to consider functions f W Œa; b ! R. We want R x to ascertain

which functions admit an integral representation f.x/ D f.a/ C a h.z/ dz. It is
natural to obtain h by differentiating f. Therefore, initially we concern ourselves
with differentiation namely of monotone functions.

Proposition 9.4 (Lebesgue). Let a < b be real numbers and let f W Œa; b ! R be
an increasing function. Then f is differentiable a.e. (w.r.t. the Lebesgue measure),
and there exists a measurable function f0 W Œa; b ! RC such that f0 .x/, for almost
every x 2 .a; b/, is equal to the derivative of f at the point x. Moreover it holds
that
Z b
f0 .z/ dz f.b/  f.a/ :
a

The proof rests on comparing, for any a < x < b, the following four “right and left
sided, upper and lower” derivatives:

f.x C h/  f.x/ f.x C h/  f.x/

f0ro .x/ WD lim sup ; f0ru .x/ WD lim inf ;
h#0 h h#0 h
f.x/  f.x  h/ f.x/  f.x  h/
f0lo .x/ WD lim sup ; f0lu .x/ WD lim inf :
h#0 h h#0 h

Since f is monotone, all four expressions are nonnegative. Differentiability in x

means that they have the same finite value.
The question of measurability does not create problems: Since f is monotone, we
have suph2.0;r .f.x C h/  x/=h D suph2.0;r\Q .f.x C h/  x/=h, and so

f.x C h/  f.x/
f0ro .x/ D lim sup :
n!1
h2.0;n1 \Q h
96 9 Absolute Continuity

From the usual properties of measurable functions we obtain the Borel measurability
of f0ro W .a; b/ ! RN C and in the same manner that of f0ru ; f0 and f0 . The Borel
lo lu
measurability of the set Df of all points x 2 .a; b/ where f is differentiable follows
from the fact that
˚ 
Df D x 2 .a; b/ W f0lu .x/ D f0lo .x/ D f0ru .x/ D f0ro .x/ < 1 :

The proof of the remaining part of Lebesgue’s Theorem is more difficult. By

means of a simple particular case we want to make it plausible that discrepancies
between the derivatives lead to a contradiction if they are too large. Let us assume
that there are numbers r < s such that f0ru .x/ < r < s < f0lo .x/ for all x 2 .a; b/.
Thus for each x there exists an h > 0 satisfying f.x C h/  f.x/ rh. Therefore it
seems natural that we can find a partition a D x0 < x1 <    < xm1 < xm D b
with f.xj /  f.xj1 / r.xj  xj1 / for every j D 1; : : : ; m. We then would have
decomposed Œa; b into intervals Ij D .xj1 ; xj / where f has only small increments,
and could conclude that

f.b/  f.a/ r.b  a/ :

But using the other part of the assumption in the same manner we could find another
partition a D y0 < y1 <    < yn1 < yn D b with f.yj /  f.yj1 /  s.yj  yj1 / for
every j D 1; : : : ; n, a decomposition into intervals I0j of larger increments of f, and
we would obtain

f.b/  f.a/  s.b  a/ :

All in all this yields a contradiction.

These considerations may be transferred in a similar manner to subintervals and
to the other derivatives. This makes it plausible that contradictions can be avoided
only if f0lu ; f0ru ; f0lo and f0ro coincide almost everywhere. In what follows we want
to elaborate this argument, however in general the choice of suitable intervals of
smaller or bigger increments of f is a bit more complicated. We prepare this step by
the following lemma concerning Vitali coverings of Borel sets.

Lemma (Vitali’s Covering Lemma). Let B  .a; b/ be a Borel set, and let V
be a set of intervals I  .a; b/ with œ.I/ > 0 having the property: For each x 2 B
and each © > 0 there exists an I 2 V such that x 2 I and œ.I/ ©. Then for every
© > 0 there exist finitely many disjoint intervals I1 ; : : : ; In 2 V such that

 [
n 
œ Bn Ij © :
jD1
Differentiability 97

Proof. We inductively construct the intervals I1 ; I2 ; : : : 2 V. I1 is chosen arbitrarily

from V. Having chosen I1 ; : : : ; Ik , we set

n [
k o
sk WD sup œ.I/ W I 2 V; I  .a; b/ n Ij :
jD1

S
If B  kjD1 INj (NIj being the topological closure of Ij ), the construction terminates,
otherwise sk > 0 holds by the assumption of the lemma. We then choose IkC1 2 V
such that œ.IkC1 /  sk =2.
If the construction terminates after n steps, the intervals I1 ; : : : ; In obviously
satisfy our assertion. If the construction does not terminate, since the intervals are
disjoint we get
1
X 1 
[
œ.Ij / D œ Ij b  a < 1 :
jD1 jD1

It follows that œ.Ik / ! 0 and that sk ! 0P

when k ! 1. In addition, for any © > 0
there exists a natural number n such that l>n œ.Il / ©=5. We show that for this n
the assertion of the lemma holds.
To this end we prove that

[
n [
Bn NIj  Il ;
jD1 l>n

where Il denotes

S the interval which has the S same midpoint as Il , but is 5 times longer.
Let x 2 B n njD1 NIj be arbitrary. Since njD1 NIj is closed, there exists an I 2 V with
x 2 I such that I; I1 ; : : : ; In are disjoint intervals. If I would be disjoint with all
the intervals Ik , it would follow that œ.I/ sk for all k and therefore œ.I/ D 0, a
contradiction. Thus there exists an l > n such that I \ Il ¤ ¿ and I \ Ij D ¿ for all
j < l. It follows that œ.I/ sl1 2œ.Il /. From this and since I \ Il ¤ ¿ we get
that Il , stretched by a suitable factor, covers the interval I. More precisely, I  Il
holds for the interval Il of 5-fold length defined above. As x 2 I, x 2 Il follows.
This yields the assertion.
We conclude that
 [
n  X X
œ Bn Ij œ.Il / D 5 œ.Il / © :
jD1 l>n l>n

The lemma is proved. t

u
98 9 Absolute Continuity

Proof of Lebesgue’s Theorem. Let r < s be real numbers. The main part of the proof
consists in showing that
˚ 
Nrs WD x 2 .a; b/ W f0ru .x/ < r < s < f0lo .x/

is a Lebesgue null set.

Let © > 0. Due to the outer regularity of the Lebesgue measure according to
Proposition 7.5, there exists an open set O such that Nrs  O  .a; b/ and œ.O/
œ.Nrs / C ©. We consider the system V of all intervals .x; x C h/  O satisfying
x 2 Nrs , h > 0 and f.x C h/  f.x/ rh. By the definition of Nrs , the system V
fulfils the conditions of Vitali’s Covering Lemma for B D Nrs , therefore there exist
disjoint intervals I1 D .x1 ; x1 C h1 /; : : : ; Im D .xm ; xm C hm / such that

 [
m 
œ Nrs n Ij ©
jD1

and

X
m
  X
m [
m   
f.xj C hj /  f.xj / r hj D rœ Ij rœ.O/ r œ.Nrs / C © :
jD1 jD1 jD1

S
We moreover consider the system V 0 of all intervals .y  k; y/  m jD1 Ij with
y 2 Nrs , k > 0 and f.y/  f.y  k/  sk. The systemS V 0 , too, by the definition of Nrs
satisfies the conditions of the lemma for B D Nrs \ m jD1 Ij , thus there exist disjoint
intervals I01 D .y1  k1 ; y1 /,. . . ,I0n D .yn  kn ; yn / such that

 [
m
 [n 
œ Nrs \ Ij n I0l ©
jD1 lD1

and

X
n
  X
n [
n   
f.yj /  f.yj  kj /  s kj D sœ I0l  s œ.Nrs /  2© :
jD1 jD1 lD1

Since each I0l is contained in one of the Ij , and since f is monotone,

X
n
  Xm
 
f.yj /  f.yj  kj / f.xj C hj /  f.xj / :
jD1 jD1

   
Altogether we get s œ.Nrs /  2© r œ.Nrs / C © . Since r < s we obtain, letting
© ! 0, that œ.Nrs / D 0, as claimed.
Absolutely Continuous Functions 99

As the rational numbers are dense in R, we have

˚  [
x 2 .a; b/ W f0ru .x/ < f0lo .x/ D Nrs ;
r;s2Q;r<s

and by virtue of ¢-subadditivity we get œ.f0ru < f0lo / D 0, that is f0lo f0ru a.e. In the
same manner f0ro f0lu a.e. follows (interchanging intervals to the right resp. left). In
addition, f0ru f0ro and f0lu f0lo obviously hold, and we obtain

f0lo f0ru f0ro f0lu f0lo a.e.

Thus the four derivatives are equal a.e., and f is a.e. differentiable; it may happen
that the derivatives have the value 1.
In order to show that the derivatives are finite a.e., we consider
 
fn .x/ WD n f.x C 1=n/  f.x/ 1.a;b1=n/ .x/ :

We have limn!1 fn .x/ D f0ro .x/ a.e. By Fatou’s Lemma and monotonicity, it follows
that
Z b Z b
f0ro .z/ dz lim inf fn .z/ dz
a n!1 a
 Z b Z aC1=n 
D lim inf n f.z/ dz  n f.z/ dz f.b/  f.a/
n!1 b1=n a

Consequently, f0ro < 1 a.e., and f possesses an a.e. finite derivative. Setting f0 .x/ WD
f0ro .x/ when x 2 Df and f0 .x/ WD 0 otherwise, the assertion follows. t
u

Example (Cantor’s Function)

In the previous section we have constructed a strictly increasing function ® from Œ0; 1/ onto the
Cantor set C. Its inverse function § W C ! Œ0; 1/ can be extended to a monotone function f W
Œ0; 1/ ! Œ0; 1/. For this purpose we recall that Œ0; 1/ n C consists of countably many disjoint
intervals Œan ; bn /. For x 2 Œan ; bn / we set f.x/ WD f.bn /. Then f is monotone and surjective, and this
implies the continuity of f.
Obviously, f0 .x/ D 0 for all x 2 .an ; bn /. As C is a null set, it follows that f0 D 0 a.e. Hence, in
R1
this example we have that 0 f0 .z/ dz < f.1/  f.0/.
There are even strictly monotone, continuous functions f W Œ0; 1/ ! Œ0; 1/ whose derivative
vanishes a.e. Such functions are more difficult to construct.

Absolutely Continuous Functions*

Now we want to characterize those monotone functions for which in the foregoing
proposition concerning derivatives of monotone functions we even have equality
100 9 Absolute Continuity

Rx
f.x/ D f.a/ C a f0 .z/ dz. To this end we introduce the following notion (not
restricted to monotone functions) which strengthens the notions of continuity and
uniform continuity.

Definition
A function f W Œa; b ! R is called absolutely continuous, if for each © > 0 there
exists a • > 0 such that for every a x1 < y1 x2 < y2    xn < yn b it
holds that

X
n X
n
ˇ ˇ
.yi  xi / • ) ˇf.yi /  f.xi /ˇ © :
iD1 iD1

Lipschitz continuous functions, for example, are absolutely continuous. These

are functions f for which there exists an L < 1 such that jf.x/f.y/j jxyj holds
for all x; y. Functions whose derivative exists everywhere and is bounded belong to
this class.

Proposition 9.5. A monotone increasing function f W Œa; b ! R is absolutely

continuous if and only if there exists a nonnegative Lebesgue integrable function
h W Œa; b ! R such that
Z x
f.x/ D f.a/ C h.z/ dz :
a

In this case h.x/ D f0 .x/ holds for almost all x 2 .a; b/.

Proof. (i) First, let us assume that f possesses the stated integral representation.
Then for any S a x1 < y1 x2 < y2    xn < yn b and c > 0 it holds,
setting A WD niD1 Œxi ; yi , that

X Z Z
n
ˇ ˇ
ˇf.yi /  f.xi /ˇ D h.z/ dz cœ.A/ C h.z/ dz :
iD1 A fh>cg

For any given © > 0 we choose c large

Pn enough such that the rightmost integral
becomes smaller
P than
ˇ ©=2. If now
ˇ iD1 .yi  xi / D œ.A/ • with • WD ©=.2c/,
we get that niD1 ˇf.yi /  f.xi /ˇ ©. Therefore, f is absolutely continuous.
(ii) Next we show that f a.e. has the derivative h if we assume the integral
representation to hold. By the results of the previous section, f is differentiable
a.e., thus
 
fn .x/ WD n f.x C 1=n/  f.x/  1.a;b1=n/ .x/

converges a.e. to f0 .x/  0. We have to prove that f0 D h a.e.

Absolutely Continuous Functions 101

We first consider the case where h.x/ c for some c < 1 and all x. Then
0 fn .x/ c follows, and the dominated convergence theorem yields for a < x < b
that
Z x Z x
0
 
f .z/ dz D lim n f.z C 1=n/  f.z/ dz
a n!1 a

 Z xC1=n Z aC1=n 
D lim n f.z/ dz  n f.z/ dz
n!1 x a
Z x
D f.x/  f.a/ D h.z/ dz :
a

This means that the two measures on Œa; b, given by the densities f0 dœ and h dœ,
coincide on all subintervals of Œa; b. These intervals form a \-stable generator
of the Borel ¢-algebra, therefore the two measures, too, coincide according to the
uniqueness theorem. Thus the densities f0 and h are equal a.e.
The general case now can be treated using the decomposition
Z x Z x
f.x/  f.a/ D f1 .x/ C f2 .x/ WD h1 .z/ dz C h2 .z/ dz ;
a a

where h1 WD h1fhcg , h2 WD h1fh>cg, and c > 0 is given. f2 is monotone increasing

and thus has a.e. a nonnegative derivative. As h1 is bounded by c, from what we
just proved it follows that h1 .x/ D f01 .x/ f0 .x/ a.e. Since h is finite a.e., letting
c ! 1 we obtain that h f0 a.e. On the other hand, the proposition concerning the
differentiation of monotone functions implies that
Z b Z b
h.z/ dz D f.b/  f.a/  f0 .z/ dz :
a a

All in all this yields h D f0 a.e. and thus the assertion.

(iii) Finally, let f be absolutely continuous. We have to prove that f has the integral
representation as stated. To this end we will show that the function
Z x
g.x/ WD f.x/  f0 .z/ dz
a

has the constant value f.a/.

By what we have seen so far g has the following properties: by Rvirtue of the
proposition concerning the derivatives of monotone functions we have x f0 .z/ dz
y

f.y/  f.x/ for x < y, therefore g is monotonically increasing. It then follows that
jg.x/  g.y/j jf.x/  f.y/j, thus g is absolutely continuous because so is f. Finally,
by (ii) we have g0 .x/ D f0 .x/  f0 .x/ a.e., that is, the derivative of g vanishes a.e.
102 9 Absolute Continuity

Let B be the Borel set of all x 2 .a; b/ with g0 .x/ D 0, and let © > 0. We consider
the system V of all intervals Œy; z  .a; b/ satisfying y < z and g.z/  g.y/
©.z  y/. For each x 2 B and each • > 0 there exists an interval I 2 V with x 2 I
and œ.I/ •. By Vitali’s Covering Lemma, S forevery • > 0 we can find  disjoint
intervals Ij D Œyj ; zj  2 V such that œ B n njD1 Ij •. Since œ Œa; b n B D 0, this
means that

X
n1
.y1  a/ C .yiC1  zi / C .b  zn / • :
iD1

If we choose • (depending on ©) sufficiently small, the absolute continuity of g

implies that

X
n1
 
g.y1 /  g.a/ C g.yiC1 /  g.zi / C g.b/  g.zn / © :
iD1

According to the definition of the intervals Ij we moreover have that

X
n
  Xn
g.zj /  g.yj / ©.zj  yj / ©.b  a/ :
jD1 jD1

Adding the previous two inequalities yields that g.b/  g.a/ © C ©.b  a/, and
letting © ! 0 we obtain g.b/ g.a/ D f.a/. Since on the other hand g increases
monotonically, it follows that g.x/ D g.a/ D f.a/ for all x 2 Œa; b. This is the
desired integral representation. t
u

Functions of Bounded Variation*

We now want to drop the assumption of monotonicity, which played an important

role in the previous two sections, and pass to functions which can be represented as
differences of monotone functions.

Definition
A function f W Œa; b ! R is said to have bounded variation (or finite variation),
if there exists a c > 0 such that for all n 2 N and all partitions a D x0 x1
   xn1 xn D b of length n,

X
n
ˇ ˇ
ˇf.xi /  f.xi1 /ˇ c :
iD1
Functions of Bounded Variation 103

Proposition 9.6 (Jordan Decomposition). A function f W Œa; b ! R has

bounded variation if and only if it is equal to the difference of two monotonically
increasing functions f1 ; f2 W Œa; b ! R:

f D f1  f2 :

Proof. First, let f be the difference of the monotonically increasing functions f1 ; f2 .

Then

X
n
ˇ ˇ X n
  Xn
 
ˇf.xi /  f.xi1 /ˇ f1 .xi /  f1 .xi1 / C f2 .xi /  f2 .xi1 /
iD1 iD1 iD1

D f1 .b/  f1 .a/ C f2 .b/  f2 .a/ :

Thus f has bounded variation.

Conversely, assume that f has bounded variation. For a y < z b the
nonnegative quantity

X
n
ˇ ˇ
v.y; z/ WD sup ˇf.xi /  f.xi1 /ˇ
yDx0 x1 xn1 xn Dz
iD1

is termed the variation of f on the interval Œy; z. Obviously, it is finite for functions
of bounded variation. If y < u < z, we may always adjoin u to the partition x0
x1    xn1 xn , because the corresponding sums become larger and the
supremum remains unchanged. Since we may select the partition below and above
u separately, it follows that

v.y; z/ D v.y; u/ C v.u; z/ :

We set

f1 .y/ WD v.a; y/ ; f2 .y/ WD v.a; y/  f.y/ ;

thus f1  f2 D f. For y < z it holds that f1 .z/  f1 .y/ D v.y; z/  0 and

f2 .z/  f2 .y/ D v.y; z/  f.z/ C f.y/  v.y; z/  jf.z/  f.y/j  0 :

Therefore, f1 and f2 are monotonically increasing. t

u

According to the proposition on differentiation of monotone functions, every

function of bounded variation can be differentiated a.e. For absolute continuous
functions the following stronger result is valid.
104 9 Absolute Continuity

Proposition 9.7. Any absolutely continuous function f W Œa; b ! R can be

represented as the difference f D f1  f2 of two monotonically increasing,
absolutely continuous functions f1 ; f2 .

Proof. As in the previous proof, we work with the variation v.y; z/. Absolute
continuity of f means that for each © > 0 there exists a • > 0 such that v.y; z/ ©
for z  y •. Since v.y; z/ D v.y; u/ C v.u; z/, it follows that v.y; z/ n© for
z  y n• and all n 2 N. In particular, v.y; z/ < 1 holds for all a y < z b.
Absolutely continuous functions therefore have bounded variation.
We proceed as in the previous proof and obtain monotone functions f1 .y/ WD
v.a; y/, f2 .y/ WD v.a; y/  f.y/, such that f D f1  f2 . It remains to show that f1 (and
therefore f2 D f1  f) is absolutely continuous.
P Let •; © > 0 and a y1 < z1
y2 < z2    yn < zn b such that njD1 .zi  yi / •. By the definition of the
supremum, there exist partitions yi D xi;0 xi;1    xi;ni D zi such that

X
ni
ˇ ˇ
v.yi ; zi / 2 ˇf.xi;j /  f.xi;j1 /ˇ :
jD1

We get

X
n X
ni X
n
.xi;j  xi;j1 / D .zi  yi / • :
iD1 jD1 iD1

Due to the absolute continuity of f it follows that

Xn X ni
ˇ ˇ
ˇf.xi;j /  f.xi;j1 /ˇ © ;
iD1 jD1
2

if • is sufficiently small. We obtain

X
n
  Xn
f1 .zi /  f1 .yi / D v.yi ; zi / © ;
iD1 iD1

thus f1 is absolutely continuous as claimed. t

u

Generalizing from the situation of monotone functions, we present the following

characterization of absolutely continuous functions.
Signed Measures 105

Proposition 9.8. A function f W Œa; b ! R is absolutely continuous if and only

if there exists a Lebesgue integrable function h W Œa; b ! R such that
Z x
f.x/ D f.a/ C h.z/ dz :
a

In this case h.x/ D f0 .x/ for almost all x 2 .a; b/.

Proof. If f is absolutely continuous, we have f D f1  f2 for someR monotone

x
absolutely continuous function f1 ; f2 . For those functions fi .x/ D fi .a/ C a hi .z/ dz,
and we obtain the integral representation for f setting h WD h1  h2 .
If, conversely, the integral representation R xholds, it follows that Rf xD f1  f2 with
the monotone functions f1 .x/ WD f.a/ C a hC .z/ dz, f2 .x/ WD a h .z/ dz. The
functions f1 and f2 are absolutely continuous, and thus so is f.
The final claim results from the corresponding assertions for f1 and f2 . t
u

Signed Measures*

When treating measures, one may also drop the monotonicity requirement, similarly
as we did for functions in the previous section.

Definition
A mapping • W A ! R N from a ¢-algebra on a measurable space .S; A/ to R
N D
R [ f1; 1g is called a signed measure, if •.¿/ D 0 and if for every (finite or
infinite) sequence of disjoint sets A1 ; A2 ; : : : 2 A one has that
[  X
• An D •.An / :
n1 n1

It is part of the definition that the sum on the right-hand side is always well-defined.
On one hand, this means that the sum does not depend on the order of summation.
On the other hand, 1 and 1 are not allowed to appear both during summation.
This excludes the possibility that there are two sets A; A0 2 A with •.A/ D 1 and
•.A0 / D 1. (Then •.A \ A0 / would have to be finite, and the disjoint sets A n A0
and A0 n A would have the value 1 and 1.) Therefore either 1 or 1 is not
present among the values of •.
Obviously, a signed measure arises when one considers the difference • D   
of two measures, at least one of them being finite. It turns out that one obtains all
signed measures in this manner. More precisely, the following proposition holds.
106 9 Absolute Continuity

Proposition 9.9 (Jordan Decomposition of Signed Measures). Let • be a

signed measure. Then there are measures •C und • , at least one of them
being finite, such that • D •C  • and •C ? • . These measures are uniquely
determined, and it holds that

•C .A/ D sup •.A0 / ; • .A/ D  inf

0
•.A0 / :
A0 A A A

•C and • are termed positive and negative variation of •. Thus one may think of
a signed measure as a charge distribution in the space S, with positive and negative
charges (as one may think of a measure as a mass distribution in the space). The
proof of the proposition is based on the Hahn decomposition for signed measures.

Proposition 9.10 (Hahn Decomposition). Let • be a signed measure on a ¢-

algebra. Then there are measurable sets A and A D S n A such that for all
measurable sets A we have

•.A/  0 for A  A ;
•.A/ 0 for A  A :

Proof. It suffices to treat the case where •.A/ > 1 for all measurable A. In that
case we may carry over completely the proof of Proposition 9.2. t
u

Proof of the Jordan decomposition. Letting A ; A be a Hahn decomposition of •,

we set

•C .A/ WD •.A \ A / ; • .A/ WD •.A \ A / :

•C and • satisfy • D •C  • and •C ? • .

Concerning uniqueness: Let • D    and ?. For measurable sets A0  A
then

•.A0 / .A0 / .A/

holds. Moreover, there exists a measurable set B such that .B/ D .Bc / D 0. It
follows that

•.A \ B/ D .A \ B/ D .A/ :

Signed Measures 107

Taken together the two assertions yield

.A/ D sup •.A0 / :

A0 A

Analogously,

.A/ D  inf
0
•.A0 / :
A A

Therefore,  and  are uniquely determined by •, and these formulas are valid for
•C resp. • , too. t
u

Exercises
9.1 Let  and  be ¢-finite. Prove that 
 is equivalent to the condition

8© > 0 9• > 0 W .A/  • ) .A/  © :

Hint: The Radon-Nikodym theorem helps. From d D h d it follows for all c > 0 that
Z Z
.A/  c d C h d  c.A/ C .h > c/ :
A\fhcg A\fh>cg

9.2 Let S be uncountable, let A be the ¢-algebra of all A  S which are countable or whose
complement is countable, and let h W S ! R be a nonnegative function. We consider the
measures  and  on A, given by .A/ WD #A and
(P
x2A h.x/ ; if A countable;
.A/ WD
1; otherwise:

(i) When does 

 hold? (ii) When has  a density w.r.t. ? (Compare Exercise 2.1)
9.3 Let B  R be a Borel set. Prove that for almost all x 2 B
 
œ Œx  h; x C h \ B
lim D1:
h#0 2h

One says that almost all elements of B are points of density of B.

9.4 Does the continuous function f.x/ WD x sin.1=x/, f.0/ WD 0 have bounded variation on the
interval Œ0; 1? What about g.x/ WD xf.x/?
9.5 Let • D   , where  and  are measures (one of them being finite). Then •C .A/  .A/
and • .A/  .A/ for all measurable A.
9.6 For a signed measure • one defines its variation as the measure j•j WD •C C • . Prove that

nX
n [
n o
j•j.A/ D sup j•.Ak /j W A1 ; : : : ; An are disjoint ; Ak  A :
kD1 kD1
The Jacobi Transformation Formula
10

We have determined the volume of parallelotopes in Euclidean space with the

aid of determinants in Proposition 3.4. In this chapter we present a far-reaching
generalization of this issue which dates back to Jacobi.1
Let G; H be open subsets of Rd , and let

®WG!H

be a C1 -diffeomorphism, that is, a bijective mapping between G and H which is

continuously differentiable in both directions. For any fixed x 2 G we thus have
 
®.x C v/ D ®.x/ C ®0x .v/ C o jvj ; (10.1)

as v 2 Rd tends to 0. Here ®0x denotes, for any x, a linear mapping from Rd to Rd .

By the inverse function theorem, ®0x is bijective for every x, and ®0x .v/ is jointly
continuous in x and v. The inverse mapping § W H ! G has analogous properties,
and
0
§®.x/ D .®0x /1 :

Generalizing Proposition 3.4 we now prove the following result which goes back to
Jacobi.

1
CARL GUSTAV JACOBI , 1804–1851, born in Potsdam, active in Königsberg and Berlin. He
worked in number theory, elliptic functions, and mechanics.

© Springer International Publishing Switzerland 2015 109

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_10
110 10 The Jacobi Transformation Formula

Proposition 10.1. For any C1 -diffeomorphism ® W G ! H and any Borel set

B  G we have
Z
 
œ ®.B/ D j det ®0x j dx :
d
B

Since ®.B/ D §1 .B/ and § is Borel measurable, ®.B/ is a Borel set. A large
portion of the proof is concerned with a geometric property of diffeomorphisms,
namely that the images of rectangles under ® (as shown in the following figure) can
be enclosed from outside as well as from inside by parallelepipeds, in fact more and
more accurately as the rectangles become smaller and smaller.

Let Q WD Œc; c/, c D .c1 ; : : : ; cd / 2 RdC be a d-dimensional interval centered at 0.

Dilating it in all directions by the factor ¢ > 0 and translating it by x 2 Rd , we obtain
the rectangle x C ¢Q. Its image ®.x C ¢Q/ can be nested using the parallelotope
®0x .¢Q/. More precisely, the following fact holds.

Lemma. Let K  G be compact and 0 < ˜ < 1. If ¢ > 0 is sufficiently small,

then

®.x/ C .1  ˜/®0x .¢Q/  ®.x C ¢Q/  ®.x/ C .1 C ˜/®0x .¢Q/

for all x 2 K.

Proof. (i) As a preparation, we show that (10.1) holds uniformly on compacta. As

K  G is compact, there exists a › > 0 such that x C v 2 G for any x 2 K and
v 2 Rd satisfying jvj ›. We claim that for every © > 0 there exists a • 2 .0; ›
such that
ˇ ˇ
ˇ®.x C v/  ®.x/  ®0 .v/ˇ ©jvj (10.2)
x

for every x 2 K and v 2 Rd with jvj •.

10 The Jacobi Transformation Formula 111

To prove this we remark that the mapping .x; w/ 7! ®0x .w/ is continuous
and hence uniformly continuous on the compact set f.x C v; w/ W x 2 K; jvj
›; jwj D 1g. Thus, for every © > 0 there exists a • 2 .0; › such that j®0xCv .w/ 
®0x .w/j ©=2 for all x 2 K, jvj •, and jwj D 1. For any x 2 K, jvj • we
now consider the function
 
g.t/ WD ® x C tv  ®.x/  t®0x .v/ ; 0 t 1:

Due to our differentiability assumptions, dg.t/=dt D ®0xCtv .v/  ®0x .v/. From
jg.t C h/j2  jg.t/j2 D .g.t C h/  g.t//  .g.t C h/ C g.t// (the dot denotes
the scalar product) it follows that djg.t/j2=dt D 2 ˇ®0xCtv .v/ ˇ ®0x .v/  g.t/.
Using the Cauchy-Schwarz inequality we get that ˇdjg.t/j2=dtˇ ©jvjjg.t/j.
Integration yields jg.t/j2 ©jvjt sup0st jg.s/j, thus .sup0t1 jg.t/j/2
©jvj sup0t1 jg.t/j and therefore jg.1/j ©jvj. This is (10.2).
We transform (10.2) in a twofold manner. Firstly we claim that for any given
© > 0 there exists a • > 0 such that
ˇ 0  ˇ
ˇ§ ®.x C v/  ®.x/  ®0x .v/ ˇ ©jvj (10.3)
®.x/

for all x 2 K, jvj •. To show this we use that the continuous mapping
0
.x; w/ 7! j§®.x/ .w/j attains a finite maximum m1 on the compact set f.x; w/ W
0
x 2 K; jwj D 1g. Consequently, j§®.x/ .v/j m1 jvj for all x 2 K, v 2 Rd , the
assertion thus follows from (10.2) if we replace there © by ©=m1 .
Secondly we claim that for every given © > 0 there exists a • > 0 such that
ˇ ˇ
ˇ§.®.x/ C ®0 .v//  x  v ˇ ©jvj (10.4)
x

for all x 2 K, jvj •. To show this we use that the continuous mapping
.x; w/ 7! j®0x .w/j attains a finite maximum m2 on f.x; w/ W x 2 K; jwj D 1g. If
• > 0 is sufficiently small, we therefore have ®.x/C®0x .v/ 2 H whenever x 2 K
and v 2 Rd with jvj •. Now (10.2) yields for ®.K/, §, ®.x/ and ®0x .v/ in place
of K, ®, x and v, the inequality j§.®.x/ C ®0x .v//  x  vj ©j®0x .v/j ©m2 jvj.
The claim follows if we replace © with ©=m2 .
(ii) We now prove the assertion of the lemma.
Concerning the right inclusion: We investigate when for any x 2 K and any
v 2 ¢Q there exists a u 2 ˜¢Q such that ®.x C v/ D ®.x/ C ®0x .v C u/ holds.
If ¢ is sufficiently small, we have x C ¢Q  G for all x 2 K. For any x 2 K,
v 2 ¢Q we may then form
0
 
u WD §®.x/ ®.x C v/  ®.x/  ®0x .v/ :

For any given © > 0, by (10.3) one has juj ©˜jvj whenever ¢ is sufficiently
small. If © is chosen sufficiently small, in view of the shape of rectangles and
because v 2 ¢Q it follows that u 2 ˜¢Q, and therefore v C u 2 .1 C ˜/¢Q.
112 10 The Jacobi Transformation Formula

According to the definition of u,

®.x C v/ D ®.x/ C ®0x .v C u/ ;

therefore ®.x C ¢Q/  ®.x/ C .1 C ˜/®0x .¢Q/ holds, as asserted.

Concerning the left inclusion: One has that ®.x/ C ®0x .v/ 2 H for all x 2 K
and v 2 ¢Q whenever ¢ is sufficiently small. We may then form
 
uN WD § ®.x/ C .1  ˜/®0x .v/  x  .1  ˜/v

for all x 2 K, v 2 ¢Q. By (10.4), jNuj ©˜jvj if ¢ is sufficiently small. If © is

sufficiently small, it follows that uN 2 ˜¢Q and .1  ˜/v C uN 2 ¢Q. According
to the definition of uN we obtain
 
®.x/ C .1  ˜/®0x .v/ D ® x C .1  ˜/v C uN ;

and so ®.x/ C .1  ˜/®0x .¢Q/  ®.x C ¢Q/.

t
u

Proof of the proposition. Let moreover Q D Œc; c/. We first determine the
Lebesgue measure of ®.z C Q/, assuming that the topological closure K of z C Q
is contained in G. To this end we use that z C Q can be partitioned, for any natural
number n, into nd disjoint rectangles Qin D xin C n1 Q, i D 1; : : : ; nd , with xin 2 K.
The following figure illustrates the case d D 3; n D 2.

Since ® is bijective, the partition can be transferred to ®.zCQ/. Using the additivity,
monotonicity, and translation invariance of the Lebesgue measure, we conclude
from the lemma that for sufficiently large n

X
d
n
   
œd .1  ˜/®0xin .n1 Q/ œd ®.z C Q/   
iD1
10 The Jacobi Transformation Formula 113

where we have omitted the upper estimate (with ˜ instead of ˜). We know the
behaviour of the Lebesgue measure under linear mappings from Proposition 3.4, so
it follows that

X
d
n
 
.1  ˜/ d
j det ®0xin jœd .n1 Q/ œd ®.z C Q/ .1 C ˜/d   
iD1

or, written by means of an integral,

Z X
nd
 
.1  ˜/ d
j det ®0xin j1Qin dœd œd ®.z C Q/ .1 C ˜/d   
iD1

Since j det ®0x j is continuous, the integrands are uniformly bounded by a constant,
and for n ! 1 they converge to j det ®0x j1zCQ . By the dominated convergence
theorem,
Z
 
.1  ˜/d j det ®0x j1zCQ dx œd ®.z C Q/ .1 C ˜/d   

and letting ˜ ! 0 we finally obtain that

Z
 
œ ®.z C Q/ D
d
j det ®0x j dx :
zCQ

This proves the formula for half-open rectangles. Consequently, it also holds for
any finite disjoint union of such rectangles whose topological closure is contained
in G. The system of all such unions forms a \-stable generator of the ¢-algebra
of all Borel sets B  G. In addition, it satisfies the assumptions of the uniqueness
theorem, applied to the measures
Z
 
.B/ WD œd ®.B/ ; .B/ WD j det ®0x j dx
B

with B  G, because the open set G can be represented as a countable union of such
rectangles. This yields the assertion. t
u

With the aid of the monotonicity principle (Proposition 2.8) we now obtain the
following “substitution formula” for integration.
114 10 The Jacobi Transformation Formula

Corollary (Transformation Formula of Jacobi). For any C1 -diffeomorphism

® W G ! H and any nonnegative measurable function f W H ! R N C it holds
that
Z Z
f.y/ dy D f.®.x//  j det ®0x j dx :
H G

Proof. For the Borel set B0 D ®.B/, Proposition 10.1 can be rewritten as
Z Z
1B0 .y/ dy D 1B0 ı ®.x/  j det ®0x j dx :
H G

The assertion now follows from the monotonicity principle, Proposition 2.8. t
u

For integrable functions an analogous formula holds. The following example

includes a well-known application.

Example (Polar coordinates)

The mapping

x D .r; ’/ 7! y D .u; v/ WD .r cos ’; r sin ’/

defines a C1 -diffeomorphism from G WD .0; 1/  .0; 2 / to H D R2 n f0g  RC . The figure

shows that the dilation is variable and equals r. Indeed, we obtain for the Jacobi determinant
 @u=@r @u=@’   cos ’ r sin ’ 
det ®0x D det D det Dr:
@v=@r @v=@’ sin ’ r cos ’

One obtains an interesting application of the transformation formula for

f.y/ D exp.jyj2 / D exp.u2  v 2 / :

The formula yields

Z Z
exp.u2 / exp.v 2 / dudv D exp.r2 /r d’dr ;
R2 G
10 The Jacobi Transformation Formula 115

where we have extended H to R2 by the null set f0g  RC . By Fubini’s Theorem from Chap. 8
we may replace both two-dimensional Lebesgue integrals with double integrals in the respective
variables, thus
Z 1 Z 1 Z 2  Z 1
1
exp.u2 / du exp.v 2 / dv D d’ exp.r2 /r dr D 2   :
1 1 0 0 2

We obtain, for a second time, the formula

Z 1 p
exp.u2 / du D  :
1

This argument goes back to Gauss.2

Exercises
10.1 Compute
“
x2 y2 dxdy ;
B

where B WD f.x; y/ 2 R2 W x2 C y2  1g.

2
CARL FRIEDRICH GAUSS, 1777–1855, born in Braunschweig, active in Braunschweig and at the
observatory in Göttingen. His contributions shape the whole of mathematics until the present time.
For astronomy, physics and geodesy, too, he has lasting merits.
Construction of Measures
11

Let A be a ¢-algebra on S with generator E, and let

 WE !R
NC

be a mapping which associates a nonnegative number  .E/ (or possibly the value
1) to each element E of the generator. In this section we want to specify conditions
which allow us to extend   to a measure  on A. More precisely, following
Carathéodory we ask under which circumstances we may use for this purpose the
mapping

NC
WA!R

related to  , given by
nX [ o
.A/ WD inf  .Em / W E1 ; E2 ; : : : 2 E; A  Em :
m1 m1

As usual we set inf ¿ WD 1. (While we follow up our discussion concerning the

regularity of measures in Chap. 7, we do not need its results in the following.)
Thus, the idea is to obtain the measure of A by outer approximation, covering A
with finitely or infinitely many elements E1 ; E2 ; : : : from E

© Springer International Publishing Switzerland 2015 117

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_11
118 11 Construction of Measures

the sum of their measures becoming as small as possible. The question is under
which conditions this procedure works. We will also discuss some applications.
As a preparation, we first present a general method to construct a measure from
an outer measure, dating back to Carathéodory. This method has a large scope of
application and in particular yields, for example, the Hausdorff measures discussed
at the end of this chapter.

Outer Measures

Definition
A mapping

˜ W P.S/ ! R
NC

on the power set P.S/ of S is called an outer measure if the following holds:

(i) ˜.¿/ D 0, P
(ii) ¢-subadditivity:
S ˜.A/ n1 ˜.An / for all A; A1 ; A2 ; : : :  S which satisfy
A  n1 An .

A subset A  S is called ˜-measurable if for every C  S

˜.C \ A/ C ˜.C \ Ac / D ˜.C/

holds.

In particular, the ¢-subadditivity entails the property of

(iii) Monotonicity: ˜.A/ ˜.A0 /, whenever A  A0 .

˜-measurability of A means that one may separate ˜ into two parts on A and on Ac ,
from which one may get back ˜ by addition. For the ˜-measurability of A it suffices
Outer Measures 119

that ˜.C \ A/ C ˜.C \ Ac / ˜.C/ holds for all C  S, because subadditivity yields
the reverse inequality.
The following result holds.

Proposition 11.1 (Carathéodory). Let ˜ be an outer measure on S. Then the

system A˜ of all ˜-measurable sets is a ¢-algebra, and the restriction of ˜ to A˜
is a measure.

Proof. Immediately obvious are the properties

S 2 A˜ and A 2 A˜ ) Ac 2 A˜ :

Let A1 ; A2 2 A˜ . Repeated application of the defining property of ˜-measurable

subsets yields

˜.C/ D ˜.C \ A1 / C ˜.C \ Ac1 / ()

D ˜.C \ A1 / C ˜.C \ Ac1 \ A2 / C ˜.C \ Ac1 \ Ac2 /
D ˜.C \ .A1 [ A2 / \ A1 / C ˜.C \ .A1 [ A2 / \ Ac1 /
C ˜.C \ .A1 [ A2 /c /
D ˜.C \ .A1 [ A2 // C ˜.C \ .A1 [ A2 /c / :

It follows that A1 [ A2 2 A˜ as well as A1 \ A2 D .Ac1 [ Ac2 /c 2 A˜ . If A1 and A2

are disjoint, we obtain from row (), replacing C by C \ .A1 [ A2 /, the additivity
property

˜.C \ .A1 [ A2 // D ˜.C \ A1 / C ˜.C \ A2 / :

Let moreover A1 ; A2 ; : : : 2 A˜ be pairwise disjoint. Using the additivity just

proved, as well as the monotonicity of ˜, we obtain that for any natural number r

 [
r   [
r c 
˜.C/ D ˜ C \ An C ˜ C \ An
nD1 nD1

X
r     [ c 
 ˜ C \ An C ˜ C \ An :
nD1 n1
120 11 Construction of Measures

Letting r ! 1 and using ¢-subadditivity we get

X   [ c 
˜.C/  ˜.C \ An / C ˜ C \ An ()
n1 n1
 [    [ c 
˜ C\ An C ˜ C \ An
n1 n1

 ˜.C/ :
S
Thus, equality
S holds everywhere above, and n1 An 2 A˜ follows. The particular
choice C D n1 An in row () yields
[  X
˜ An D ˜.An / ;
n1 n1

that is, ˜ is ¢-additive on A˜ . Finally, arbitrary countable unions can be reduced to

disjoint unions according to
[ [
An D An \ Ac1 \    \ Acn1 ;
n1 n1

so that A˜ indeed is a ¢-algebra. t

u

Extension to a Measure

We utilize outer measures to prove the following proposition.

Proposition 11.2 (Extension). Let E generate the ¢-algebra A on S, and let

 WE !R N C be a mapping. Then
nX [ o
.A/ WD inf  .Em / W E1 ; E2 ; : : : 2 E; A  Em ; A2A;
m1 m1

defines a measure  on A coinciding with   on E, if and only if the conditions

(i) .¿/ D 0,
(ii) .E/ D  .E/ for all E 2 E,
(iii) .E0 \ E/ C .E0 \ Ec /  .E0 / for all E; E0 2 E

are satisfied.
Extension to a Measure 121

Since .E/  .E/ always holds, (ii) can be replaced with .E/   .E/. The
verification of this seemingly innocuous condition typically requires a substantial
effort. According to the definition of  it is equivalent to the condition
P S
(ii’)  .E/ m1  .Em / for E; E1 ; E2 ; : : : 2 E and E  m1 Em .

We will see how, in order to verify it, one replaces the infinite covering of E
with other, more easily tractable, finite coverings. Reasonings like that, based on
compactness arguments, go back to Borel, who established the topological concept
of compactness in mathematics.

Proof. The conditions are obviously necessary. To prove their sufficiency we extend
 to the whole power set by
nX [ o
˜.A/ WD inf  .Em / W E1 ; E2 ; : : : 2 E; A  Em for all A  S :
m1 m1

S
˜ is ¢-subadditive: Let A; A1 ; A2 ; : : :  S be such that A  n1 An holds. By
S of ˜, for every © > 0 there exist elements E1n ; E2n ; : : : of E such that
definition
An  m1 Emn and
X
 .Emn / ˜.An / C ©2n :
m1

S
It follows that A  m;n1 Emn and
X X X
˜.A/  .Emn / .˜.An / C ©2n / ˜.An / C © :
m;n1 n1 n1

Letting © ! 0 we obtain the ¢-subadditivity. By (i) we moreover have ˜.¿/ D 0,

thus ˜ is an outer measure.
We show next Sthat each E 2 E is a ˜-measurable set. Let C  S and E1 ; E2 ; : : : 2 E
satisfying C  m1 Em . Since ˜ is ¢-subadditive,
X X
˜.C/ ˜.C \ E/ C ˜.C \ Ec / ˜.Em \ E/ C ˜.Em \ Ec / ;
m1 m1

and due to (iii)

X
˜.C/ ˜.C \ E/ C ˜.C \ Ec /  .Em / :
m1
122 11 Construction of Measures

According to the definition of P˜, we may choose, for any given © > 0, the sets
E1 ; E2 ; : : : in such a way that m1  .Em / ˜.C/ C © holds. It follows that

˜.C/ ˜.C \ E/ C ˜.C \ Ec / ˜.C/ C © :

Letting © ! 0 we see that E is ˜-measurable.

We can now apply the preceding proposition. Since E generates A, it follows
at first that A  A˜ and secondly that  is a measure. Due to condition (ii), 
coincides with   on E as claimed. t
u

By its very definition, the measure obtained in the extension theorem is outer regular
w.r.t. E. The theorem has important applications.

Example (Locally finite measures on R)

We consider measures on R which are finite on bounded sets. Such measures  are, as a
consequence of the uniqueness theorem, uniquely determined by the values
 
 .a; b ; 1 < a  b < 1 :

One always can exhibit a “primitive” F W R ! R such that

 
 .a; b D F.b/  F.a/
   
holds, e. g., F.a/ WD  .0; a , resp.  .a; 0 , depending on whether a  0 or a < 0. In
addition, F (like primitives in calculus) is uniquely determined by  up to a constant. F is obviously
monotone and, by virtue of the ¢-continuity of , right-continuous.
Here we want to show that, conversely, for every monotone right-continuous function F there
exists a measure  such that the stated connection stands. For this purpose, on the generator

E WD f.a; b W 1 < a  b < 1g

of the Borel ¢-algebra in R we consider the functional   W E ! R given by

 
  .a; b WD F.b/  F.a/ :

We want to show that the conditions of the extension theorem are satisfied.
Obviously  .¿/ D 0 holds, thus (i) is satisfied. Moreover, for any E0 D .a0 ; b0  and E 2 E
there exist numbers a0  a  b  b0 such that

E0 \ E D .a; b ; E0 \ Ec D .a0 ; a [ .b; b0  :

Consequently, .E0 \ E/   ..a; b/ and .E0 \ Ec /   ..a0 ; a/ C  ..b; b0 /, whence

.E0 \ E/ C .E0 \ Ec /  F.b0 /  F.a0 / D  .E0 / :

Therefore (iii) is satisfied.

S
Finally, let .a; b  m1 .am ; bm . As mentioned before, in order to prove (ii’) we will pass
from the countable covering to suitable finite coverings: Since F is right-continuous, for any given
Outer Regularity 123

© > 0 there S exist numbers ©m > 0 such that F.bm C ©m /  F.bm / C ©2m . It follows that
Œa C ©; b  m1 .am ; bm C ©m /. Thus, we exhibited an open covering of a compact set which
Sn
accordingly contains a finite subcovering. We thus have .a C ©; b  mD1 .am ; bm C ©m  for a
sufficiently large natural number n, and consequently

X
n
  Xn
 
F.b/  F.a C ©/  F.bm C ©m /  F.am /  F.bm /  F.am / C © :
mD1 mD1

Letting first n ! 1 and then © ! 0 we obtain that

  X  
  .a; b    .am ; bm  :
m1

Therefore, (ii’) is satisfied.

By virtue of the extension theorem, there exists a measure  on the Borel ¢-algebra satisfying
 
 .a; b D F.b/  F.a/ ;

as claimed.

Example (Lebesgue measure)

In the particular case F.a/ D a, as in the foregoing example, one obtains the 1-dimensional
Lebesgue measure. The d-dimensional Lebesgue measure can be constructed analogously, using
d-dimensional instead of 1-dimensional intervals, or alternatively as a product measure from the
1-dimensional Lebesgue measure.

Outer Regularity*

Now we can prove the proposition concerning outer regularity in Chap. 7. Let us
repeat its assertion (with a change in notation).
Let E be a \-stable generator of the ¢-algebra A on S with ¿ 2 E. Let  be a
measure on A such that there exist E1 ; E2 ; : : : 2 E with Em " S and .Em / < 1 for
all m  1. Set
nX [ o
.A/ WD inf .Em / W E1 ; E2 ; : : : 2 E; A  Em ; A 2 A :
m1 m1

If

.E0 n E/ D .E0 n E/ for all E; E0 2 E with E  E0 ;

then  is outer regular w.r.t. E, that is, .A/ D .A/ for all A 2 A.
124 11 Construction of Measures

Proof. We show that the conditions of the extension theorem hold with   WD jE.
Condition (ii’) is satisfied automatically because  is a measure. Thus (ii) holds and,
as ¿ 2 E, also (i).
Concerning (iii): Let E; E0 2 E. As E is \-stable, .E0 \ E/ D .E0 \ E/ holds.
By assumption we moreover have .E0 \ Ec / D .E0 n E \ E0 / D .E0 n E \ E0 / D
.E0 \ Ec /. Since  is additive,

.E0 \ E/ C .E0 \ Ec / D .E0 / D  .E0 /

follows. Thus (iii) is verified.

By the extension theorem,  is a measure which coincides with  on E. Applying
the uniqueness theorem for measures we obtain that  D . t
u

The Riesz Representation Theorem

R
Any measure  induces a functional f 7! `.f/ WD f d on the space of -integrable
functions. This functional is linear and positive, that is,

`.’f C “g/ D ’`.f/ C “`.g/ for all ’; “ 2 R ; `.f/  0 for all f  0 :

If f g, it follows that `.f/ `.g/, since `.g/  `.f/ D `.g  f/  0.

Conversely, we may ask which positive linear functionals can be represented as
integrals. We consider the simplest, yet most important case. By C.S/ we denote the
linear space of all real-valued continuous functions on a metric space S.

Proposition 11.3 (Riesz Representation Theorem). Let S be a compact metric

space endowed with the Borel ¢-algebra B, and let ` W C.S/ ! R be a positive
linear functional. Then there exists a unique measure  on .S; B/ satisfying
Z
`.f/ D f d

for every f 2 C.S/.

We obtain such a measure with the aid of the extension Proposition 11.2. As
generator E of B we choose the system O of all open subsets of S. We define a
set function   W O ! RC by

 .O/ D sup `.f/ :

0f1O

Since 0 1O 1, we have 0  .O/ `.1/. We directly conclude that  .¿/ D 0,

 .O/  .O0 / whenever O  O0 , as well as `.f/  .O/ `.g/ whenever we have
The Riesz Representation Theorem 125

f 1O g. (We remark that the definition  .O/ D `.1O / is not possible as 1O is,
in general, not continuous.)
We define
nX [ o
.A/ D inf  .Om / W O1 ; O2 ;    2 O; A  Om ; A 2 B :
m1 m1

The extension Proposition 11.2 for measures says that  defines a measure on B if
the conditions

(i)  .¿/ D 0,
(ii) .O/ D  .O/ for all O 2 O,
(iii) .O0 \ O/ C .O0 \ Oc /  .O0 / for all O; O0 2 O,

are satisfied, where

P S
(ii0 )  .O/ m1  .Om / for all O; O1 ; O2 ; : : : 2 O with O  m1 Om ,

is equivalent to (ii).
As a prerequisite we provide a connection between monotonicity and uniform
convergence in C.S/.

Lemma (Dini’s1 Theorem). Let f1 ; f2 ; : : : be a sequence in C.S/ with fn " f and

f 2 C.S/, S being a compact metric space. Then fn converges uniformly to f.

Proof. Let © > 0. Given x 2 S, we choose nx such that jf.x/  fnx .x/j < ©. Due to
continuity there exists an open neighbourhood Ox of x such that jf  fnx j < © holds
on Ox . By virtue of compactness, S can be covered by finitely many such Ox , say
for points xj , 1 j m. It follows that kf  fn k1 < © for any n  maxj nxj . t
u

We return to the task of proving properties (i) to (iii) and consider, for any O 2 O,
the functions
 
®n;O .x/ D min 1; nd.x; Oc/ :

They are continuous and satisfy 0 ®1;O ®2;O    , as well as

1O D sup ®n;O ;  .O/ D sup `.®n;O / :

n1 n1

1
ULISSE DINI , 1845–1918, born in Pisa, active in Pisa. He did research in real analysis.
126 11 Construction of Measures

The latter is a consequence of Dini’s Theorem, since for any f 1O and fn D

min.f; ®n;O / one has that fn " f, and thus

`.f/ D sup `.fn / sup `.®n;O /  .O/ ;

n1 n1

which after passing to the supremum w.r.t. f implies the asserted equality.

Lemma. The set function   satisfies (ii0 ) and consequently (ii).

S
Proof. Let
PO; O1 ; O2 ; : : : 2 O satisfying O  m1 Om be given, let f 1O . We
set gn D nmD1 ®n;Om and fn D min.f; gn /. Then

X
n X
n
`.fn / `.gn / D `.®n;Om /  .Om /
mD1 mD1

holds. Since f 1O supn1 gn we have fn " f, thus `.f/ D supn1 `.fn /

P
m1  .Om / by Dini’s Theorem, and the assertion follows after passing to the
supremum w.r.t. f. t
u

Lemma. The set function   satisfies (iii).

In this proof, for any O 2 O, instead of ®n;O we use

 
§n;O .x/ WD min 1; .nd.x; Oc/  1/C :

§n;O has the same properties as those we just derived for ®n;O . In addition,
d.x; Oc / > 1=n whenever §n;O .x/ > 0.

Proof. Let O; O0 be open sets. We set g WD §n;O0 \O and

V WD fx 2 O0 W d.x; Oc / < 1=ng :

One has that fg > 0g \ V D ¿.

The Riesz Representation Theorem 127

Let © > 0. We choose n large enough such that  .O0 \ O/ `.g/ C ©. Since V is
open, there exists an h 1V satisfying  .V/ `.h/ C ©. We have 0 g C h 1O0 ,
as V  O0 and g.x/ D 0 for any x 2 V. Since moreover O0 \ Oc  V, it follows
that

.O0 \ O/ C .O0 \ Oc /  .O0 \ O/ C  .V/ `.g C h/ C 2©

 .O0 / C 2© :

Passing to the limit © ! 0 yields the assertion. t

u

Proof of Riesz’ theorem. The preceding lemmata show that the assumptions of the
R  is a measure on B, and that
extension Proposition 11.2 are satisfied, implying that
.S/ D `.1/ < 1. It remains to show that `.f/ D f d holds for all f 2 C.S/. Let
f  0 be continuous. For any n  1 and any k  0 we set
1 k 
fkn D min ; .f  /C :
n n
P
The functions fkn are continuous, and for every n one has f D k0 fkn , where at
most finitely many summands are nonzero, as well as

1 1
1ff>.kC1/=ng fkn 1ff>k=ng :
n n
According to the definition of  ,

1 1
 .ff > .k C 1/=ng/ `.fkn /  .ff > k=ng/ ;
n n
thus
X X1 X kC1
`.f/ D `.fkn / .fk=n < fg/ D .fk=n < f .k C 1/=ng/
k0 k0
n k0
n
Z
1
f d C .S/ :
n
R
Letting n ! 1 we get R`.f/ f d. The reverse inequality results in an analogous
manner, and so `.f/ D f d for any f  0 and, via a decomposition in positive and
negative part, for arbitrary f 2 C.S/, as it was claimed. The uniqueness of  was
already proved earlier in Chap. 7. t
u

We remark that the measure , constructed in the Riesz representation theorem, is

regular by Proposition 7.6.
128 11 Construction of Measures

Extension of Measures on Infinite Products*

The following result, which goes back to Kolmogorov,2 is of interest in probability

theory. The question is this: Let there be given finite measures d on the Borel ¢-
algebra B d of Rd , d  1. Under which conditions does there exist a measure  on
the product space .R1 ; B 1 / which extends the measures d in the sense that

.B  R1 / D d .B/ ; B 2 Bd

holds? Such a  is called the projective limit of the measures d . The measures d
obviously have to be related to each other in the following manner.

Definition
A sequence d , d  1, of finite measures on Rd is called consistent, if

dC1 .B  R/ D d .B/

for every d  1 and every Borel set B 2 B d .

Example (Product measures)

If dC1 D d ˝ dC1 holds for some probability measures 2 ; 3 ; : : :, then 1 ; 2 ; : : : are
consistent measures.

Proposition 11.4 (Kolmogorov’s Theorem). Every consistent sequence

1 ; 2 ; : : : of finite measures possesses a unique projective limit .

Proof. By E we denote the system of all sets O  R1  R1 , where O is an open

subset of some Rd with d D 1; 2; : : :. We define the set function   W E ! RC by

 .O  R1 / WD d .O/ :

Here we have to take into account that every E 2 E allows different representations,
namely E D O  R1 can also be written as E D O0  R1 with O0 D O  Re , e  1.
Nevertheless   is well defined due to the consistency condition.
E generates the product ¢-algebra B 1 on R1 . We define  as in the measure
extension theorem and therefore have to verify the conditions of the latter.
Concerning condition (iii): For any E; E0 2 E there exist a (common!) d  1 and
O; O0 2 B d such that E D O  R1 , E0 D O0  R1 . Moreover, we consider the open

2
ANDREI N. KOLMOGOROV , 1903–1987, born in Tambov, active in Moscow. He made seminal
contributions to probability theory, topology, dynamical systems, mechanics, and turbulent flows.
Extension of Measures on Infinite Products 129

sets On WD fx 2 Rd W jxyj < 1=n for some y 2 Oc g, the open 1=n-neighbourhoods

of the closed set Oc . Setting En D On  R1 we get for every n  1

.E0 \ E/ C .E0 \ Ec /  .E0 \ E/ C  .E0 \ En / D d .O0 \ O/ C d .O0 \ On / ;

and passing to the limit n ! 1 we obtain, by virtue of ¢-continuity,

.E0 \ E/ C .E0 \ Ec / d .O0 \ O/ C d .O0 \ Oc / D d .O0 / D  .E0 / :

This yields (iii). Condition (i) follows from (ii) since ¿ 2 E.

It remains to prove (ii0 ): Let E D O  R1 , with O  Rd open. As we already did
previously, we will pass from a countable covering of E to suitable finite coverings.
To this end we choose © > 0 and, according to Proposition 7.6, for every n  1 a
SO  R such that  .E/ D dCn .O  R / <  dCn .Kn / C ©.
n n
compact set Kn 
Let now E  m1 Em satisfying Em 2 E and Em D Om  R1 . We want to show
that there exists an n  1 such that

[
n
Kn  R1  Em :
mD1

Contrarily, assume
Sn that there exist x1 ; x2 ; : : : in R1 satisfying xn 2 Kn  R1
and xn … mD1 Em for all n  1. Then one may pass to a subsequence
converging componentwise, by the following scheme: As K1 is compact, there exists
a subsequence xi;1 2 R1 , i  1, whose first d C 1 components converge. As K2
is compact, we find a subsubsequence xi;2 , i  1, for which also the .d C 2/-th
component converges. One continues as follows: In the kth subsubsequence xi;k ,
i  1, the first d C k components converge. Following Cantor, we finally pass to the
diagonal sequence xi;i 2 R1 , i  1, which eventually traverses every subsequence
S converge to some limit y D .y1 ; y2 ; : : :/. It
and for which therefore all components
follows that y 2 K1  R1  E  m1 Em and thus y 2 Ej for some j  1. Since
Oj is open we conclude that xi;i 2 Ej whenever i is sufficiently large. As the diagonal
sequence is a subsequence
S of the original sequence xn , n  1, there exists an n  j
such that xn 2 nmD1 Em . This is a contradiction.
Therefore, there exists an n  1 such that the inclusion stated above holds. In
other words, there exist a k  n C d S and open sets Om 2 Rk , m n such that
1
Em D Om  R and Kn  R knd
 nmD1 Om . Due to the subadditivity of k it
follows that

X
n
 .E/  © dCn .Kn / D k .Kn  R knd
/ k .Om / ;
mD1

Pn
therefore  .E/ mD1  .Em / C ©. Passing to the limit n ! 1 and then © ! 0
we obtain (ii0 ).
130 11 Construction of Measures

In this manner the extension theorem yields a measure  satisfying .OR1 / D

d .O/ for all open O  Rd . Using the uniqueness theorem .B  R1 / D d .B/ for
all Borel sets B  Rd . Therefore,  is the projective limit of the measures d , d  1.
As E is a \-stable generator of B 1 , the projective limit is uniquely determined. u t

The compactness argument in the proof above may also be based on Tikhonov’s
Theorem which states that infinite Cartesian products of compact sets are them-
selves compact. This would shorten the proof.
Kolmogorov’s theorem can be generalized in several directions. The space R
may be replaced by spaces in which open subsets can be approximated from the
interior by compact sets, at least in measure. This works for all complete separable
metric spaces (Ulam’s Theorem). The result may also be transferred to uncountable
products without a major effort.

Hausdorff Measures*

The Lebesgue measure is not the only translation invariant measure on the Borel
sets of Rd . As a conclusion of this chapter we want to present a whole family
of translation invariant measures. Only if the unit cube has finite measure, one is
dealing (except for a normalization factor) with the Lebesgue measure.
A basic idea is to cover a subset A of Rd by balls and other sets of bounded
diameter

and to obtain a number measuring A from their count and their diameter. There are
different possibilities to do so, one may assign a positive measure also for “sparse”
sets with Lebesgue measure 0. It appears to be natural to cover A with sets of very
small diameter only – we will see that there are sound mathematical reasons for this.
Our approach makes use of outer measures ˜s which depend on a given parameter
s > 0.
We define the diameter of A  Rd as
˚ 
d.A/ WD sup jx  yj W x; y 2 A :
Hausdorff Measures 131

In a first step we prescribe (besides s) a • > 0 and set

nX [ o
˜s;• .A/ WD inf d.Am /s W A  Am ; d.Am / • ; A  Rd :
m1 m1

For covering we thus use arbitrary sets of diameter at most •.

The set function ˜s;• is an outer measure; this is proved as for the extension
theorem above. However, in general one does not know what are the corresponding
measurable sets. Therefore, in a second step one passes to

˜s .A/ WD sup ˜s;• .A/ ; A  Rd :

•>0

This means that henceforth we consider only small •, since ˜s;• .A/ increases
monotonically as • decreases. Obviously ˜s is translation invariant.
˜s , too, S
is an outer
 measure:
P since ˜s;• .¿/
P D 0 we also have ˜s.¿/ S D 0, and
from
P ˜ s;• n1 A n n1 ˜ s;• .A n / n1 ˜ s .A n / follows ˜ s n1 An
n1 ˜s .An /.
For ˜s an additional property comes into play: Let
˚ 
a.A0 ; A00 / WD inf jx  yj W x 2 A0 ; y 2 A00 ;

denote the distance between two subsets A0 , A00 of Rd (using the convention
inf ¿ D 1, the distance to the empty set equals 1). Following Carathéodory, an
outer measure ˜ on Rd is called metric if it satisfies the condition

a.A0 ; A00 / > 0 ) ˜.A0 [ A00 / D ˜.A0 / C ˜.A00 / :

S
The outer measures ˜s are metric. Indeed, let A0 [ A00  m1 Am where
d.Am / •. If • < a.A; A0 /=2, then each Am intersects at most one of the sets
A0 , A00 . Therefore we may partition
S the sequence Am into
S two subsequences A0m ,
00 0 0 00 00
Am , m  1, such that A 
P m1 Am and A  m1 Am . It follows that
0 00 0 00 0 00
m1 d.Am /  ˜s;• .A / C ˜s;• .A /, therefore ˜s;• .A [ A /  ˜s;• .A / C ˜s;• .A /
s
0 00 0 00
and, letting • ! 0, finally ˜s .A [ A /  ˜s .A / C ˜s .A /. The reverse inequality
holds because ˜s is an outer measure.
The importance of metric outer measures stems from the following characteriza-
tion.

Proposition 11.5. An outer measure ˜ on Rd is metric if and only if all Borel

sets are ˜-measurable.
132 11 Construction of Measures

Proof. First assume that all Borel sets are ˜-measurable. If a.A0 ; A00 / > 0 holds for
sets A0 ; A00 , then O WD fy 2 Rd W jy  xj < a.A0 ; A00 / for some x 2 A0 g is an open
set. Its ˜-measurability implies that
   
˜.A0 [ A00 / D ˜ .A0 [ A00 / \ O C ˜ .A0 [ A00 / \ Oc :

Moreover, we have A0  O, A00  Oc , and we obtain ˜.A0 [ A00 / D ˜.A0 / C ˜.A00 /.

Therefore ˜ is a metric outer measure.
Conversely, let ˜ be metric. We show that in this case every closed set A  Rd is
˜-measurable, so that ˜.C/  ˜.C \ A/ C ˜.C \ Ac / holds for all C  Rd . Without
loss of generality, we may assume that ˜.C/ < 1. To prove the claim we construct
sets D1  D2      C \ Ac such that a.C \ A; Dn / > 0 and ˜.Dn / ! ˜.C \ Ac /.
Since ˜ is metric and .C \ A/ [ Dn  C, it then follows that
 
˜.C \ A/ C ˜.Dn / D ˜ .C \ A/ [ Dn ˜.C/ ;

and passing to the limit n ! 1 yields the assertion.

To carry out this train of thoughts we choose a sequence of real numbers ©1 >
©2 >    > 0 converging to zero and set

Dn WD fx 2 C \ Ac W jx  yj  ©n for all y 2 Ag :

Now a.C \ A; Dn /  ©n > 0 holds, as desired.

In order to prove the other property of the sets Dn , we additionally consider the
sets En WD DnC1 nDn , n  1. For any m  1 we have a.EnCm ; En1 /  ©n ©nC1 > 0.
Since ˜ is metric, it follows that

X
n [
n 
˜.E2k / D ˜ E2k ˜.C/
kD1 kD1

Pn P
and analogously kD1 ˜.E2k1 / ˜.C/, and we obtain k1 ˜.Ek / < 1, because
we have assumed ˜.C/ < 1 to hold. S
Since A is assumed to be closed, we have C \ Ac D Dn [ mn Em and
consequently, by virtue of ¢-subadditivity,
X
˜.Dn / ˜.C \ Ac / ˜.Dn / C ˜.Em / :
mn

As n ! 1 the rightmost expression tends to 0, and so ˜.Dn / ! ˜.C \ Ac /.

Therefore, the sets D1 ; D2 ; : : : have the desired properties, and so all closed sets are
˜-measurable. This holds for all Borel sets, too, because the closed sets generate the
Borel ¢-algebra. t
u
Hausdorff Measures 133

The metric outer measures ˜s resp. the measures arising from the restriction to the
Borel ¢-algebra are called Hausdorff measures. For geometric investigations one
rather uses them as a family, the value of the parameter s is chosen for each A  Rd
separately.

Lemma. For every A  Rd there exists a number 0 hA d such that

(
1; if s < hA ;
˜s .A/ D
0; if s > hA :

Proof. According to the definition of ˜s;• we have for all © > 0

˜sC©;• .A/ •© ˜s;• .A/ :

When ˜s .A/ < 1, passing to the limit • ! 0 yields ˜sC© .A/ D 0. Thus, a number
hA 2 Œ0; 1 exists which is related to ˜s .A/ as claimed.
It remains to prove that hA d. The unit cube Œ0; 1/d can beppartitioned in an
obvious manner into nd subcubes of sidelength 1=n and diameter d=n. Therefore,
  p d
˜d;pd=n Œ0; 1/d nd d=n D dd=2 ;
 
and letting
 n ! 1 we obtain ˜d Œ0; 1/d < 1. For every © > 0 it follows that
˜dC© Œ0; 1/d D 0, and by virtue of ¢-additivity we get ˜dC© .Rd / D 0. This proves
that hA d for all A  Rd . t
u

The number hA is called the Hausdorff dimension of A. In geometric measure theory

the Hausdorff dimension and measure are studied in more detail. It turns out that in
all cases for which one can assign a dimension to A in an intuitive manner, this
dimension coincides with the Hausdorff dimension. Moreover, in the d-dimensional
case the Hausdorff measure for s D d coincides with the Lebesgue measure, except
for a positive normalization constant which however is not easy to determine. We
do not discuss this further and close this section with an example.

Example (Cantor set)

The Hausdorff dimension of the Cantor set C can be determined easily through a heuristic scaling
argument. For any set A  R and any c > 0 we set cA WD fcx W x 2 Ag. Then we have (compare
Exercise XI.2)

˜s .cA/ D cs ˜.A/ :
134 11 Construction of Measures

Obviously C D C0 [ C00 for some disjoint sets C0 and C00 which result from C by scaling with the
factor c D 1=3 and by translation. Hence,

˜s .C/ D ˜s .C0 / C ˜s .C00 / D 2  3s ˜s .C/ :

Assuming that 0 < ˜h .C/ < 1 holds for the Hausdorff dimension h D hC of C, it follows that
1 D 2  3h or
log 2
hD D 0; 631 :
log 3

We now want to show that for this number h we indeed have 1=2  ˜h .C/  1. On the one
hand, C is contained in Cn , the disjoint union of 2n intervals of length 3n . Therefore,

˜h;3n .C/  2n .3n /h D 1

and ˜h .C/  1.
For the other estimate we utilize the bijection ® W Œ0; 1/ ! C which we have introduced in
Chap. 9 in the section dealing with the Cantor set. For all y; y0 2 Œ0; 1/ we have
ˇ ˇh
2ˇ®.y/  ®.y0 /ˇ  jy  y0 j :
P
Indeed, let n be the first position in the binary representations y D yk 2k and y0 D
P 0 k
k1
k1 yk 2 where yn ¤ y0n holds. Then
X  X 
jy  y0 j  2k D 2nC1 ; j®.y/  ®.y0 /j  2 3n  3k D 3n ;
kn k>n

and the assertion follows since .3n /h D 2n . For any interval A  R this yields
   
2d.A/h  d ®1 .A/ D œ ®1 .A/ :
S
If now C  m1 Am holds for some intervals A1 ; A2 ; : : :, the ¢-continuity of the Lebesgue
S
measure and the fact that Œ0; 1/  m1 ®1 .Am / imply that
X X  
2 d.Am /h  œ ®1 .Am /  1 :
m1 m1

Because in the one-dimensional case it is obviously sufficient to consider coverings consisting

of intervals only, we obtain that ˜h .C/  1=2. By the way, a more precise analysis reveals that
˜h .C/ D 1.

Exercises
11.1 Let  be the measure used in the proof of the extension theorem which results from restricting
the outer measure ˜ to the ¢-algebra A˜ . Prove that if  is finite (or, at least, ¢-finite), then 
coincides with the completion of .
0 0
Hint: First show that for each A  S there S an A  A such that .A / D ˜.A/.
T exists
0 0
A can be chosen as having the form A D n1 m1 Emn where Emn 2 E .
Hausdorff Measures 135

11.2 Prove that the Hausdorff measure satisfies

˜s .cA/ D cs ˜s .A/ :

Conclude that in the d-dimensional case ˜s is different from the Lebesgue measure when
s ¤ d, and moreover cannot be made to coincide with it by scaling.
Hilbert Spaces
12

We return to the space L2 .SI / of square integrable functions whose basic

properties we have discussed in Chap. 6. They are related to some geometric issues
which we now want to learn about. Those are the properties of a Hilbert space,1 for
which the space L2 .SI / serves as a prototype.
A Hilbert space is a vector space where not only a length is associated to each
vector, but moreover any two vectors – by means of a scalar product – enclose an
angle between them, in particular it is possible to state whether they are orthogonal
to each other. Its additional geometric properties make it possible to find, for any
convex closed set K and any point x not belonging to K, a point in K whose distance
to x is minimal. This leads to widely used orthogonal decompositions, the most
prominent being the Fourier series.
We recall the definition of a scalar product in a vector space over the real or
complex field. For any ’ 2 C we denote by ’ the complex number conjugate to ’;
note that ’’ D j’j2 .

Definition
A scalar product is a mapping which to any two elements x; y of a vector space
X associates a number .x; y/ with the following properties:

(i) Positive definiteness: .x; x/ > 0 for any x ¤ 0,

(ii) .y; x/ D .x; y/ for all vectors x; y 2 X,
(iii) .’x C “y; z/ D ’.x; z/ C “.y; z/ for all x; y; z 2 X and all scalars ’; “.

1
DAVID HILBERT, 1862–1943, born at Königsberg, active in Königsberg and Göttingen. The
23 problems presented by him in Paris in 1900 and named after him have deeply influenced
the development of mathematics. With him and his activity, which extended to all branches of
mathematics, Göttingen became the world center of mathematics.

© Springer International Publishing Switzerland 2015 137

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_12
138 12 Hilbert Spaces

Properties (ii) and (iii) directly imply .x; 0/ D .0; x/ D 0 and

.x; ’y C “z/ D ’.x; y/ C “.x; z/

for any vectors x; y; z and any scalars ’; “. In the real case, a scalar product thus is
just a symmetric positive definite bilinear form.

Example

1. With .x1 ; x2 ; : : : / and .y1 ; y2 ; : : : / being two sequences of scalars of length d, the
expression

X
d
.x; y/ D xn yn
nD1

defines a scalar product on the space Rd , resp. Cd , if d is finite; in the case d D 1

we arrive at the space
˚ X 
`2 WD .x1 ; x2 ; : : :/ W jxn j2 < 1
n

of all square summable real, resp. complex sequences.

2. The definition
Z
.f; g/ D fg d

yields a scalar product in the space L2 .SI / of all square integrable functions
on a measure space .S; A; /. Here, the integral of a complex-valued function
h D h1 C ih2 with h1 ; h2 2 L1 .SI / is defined as
Z Z Z
h d D h1 d C i h2 d :

We set
p
kxk WD .x; x/ :

The following fact is known from analysis and linear algebra.

Proposition 12.1. In a vector space X with scalar product .; /, k  k defines a
norm, and there holds the Cauchy-Schwarz inequality

j.x; y/j kxkkyk for all x; y 2 X :

12 Hilbert Spaces 139

In particular, d.x; y/ WD kx  yk defines a metric d on X. With respect to this metric

we thus may speak of convergence xn ! x of sequences xn 2 X to a limit x 2 X, of
closed subsets of X, and so on. Since .xn ; yn /  .x; y/ D .xn  x; yn  y/ C .xn 
x; y/ C .x; yn  y/ and therefore j.xn ; yn /  .x; y/j kxn  xkkyn  yk C kxn 
xkkyk C kxkkyn  yk, it holds that

.xn ; yn / ! .x; y/ ; if xn ! x ; yn ! y :

The scalar product is thus continuous, and therefore so is the norm. In the next
chapter we will recall the notion of a norm and its implications in more detail.
In a vector space X with scalar product there holds the parallelogram identity

kx C yk2 C kx  yk2 D 2.kxk2 C kyk2 / ;

an immediate consequence of the formula kx˙yk2 D kxk2 Ckyk2 ˙Œ.x; y/C.y; x/.
Moreover, the definitions directly imply that in the real case the scalar product
satisfies the identity

1
.x; y/ D .kx C yk2  kx  yk2 / ()
4

for any x; y 2 X. If, conversely, k  k is a norm on X satisfying the parallelogram

identity for all x; y 2 X, then after some computation one sees that, in the real case,
() indeed defines a scalar product. In the complex case, a different formula is valid
(Exercise 12.1). This passage between the square of the norm and the scalar product
is called polarization.
Two vectors x; y on a Hilbert space X are called orthogonal if .x; y/ D 0. From
.x C y; x C y/ D .x; x/ C .x; y/ C .y; x/ C .y; y/ we obtain for orthogonal vectors
x; y 2 X the “Pythagoras Theorem”

kx C yk2 D kxk2 C kyk2 :

For M  X, the set

M? D fx W .x; y/ D 0 for all y 2 Mg

is called the orthogonal complement of M. We obviously have N?  M? whenever

?
N  M, as well as M D M? , because .xn ; y/ D 0 for every n and xn ! x imply
that .x; y/ D 0. We also note that M? is a closed subspace of X.

Definition
A vector space X endowed with a scalar product is called Hilbert space, if it is
complete w.r.t. the corresponding norm k  k, that is, if every Cauchy sequence
w.r.t. the metric d.x; y/ WD kx  yk is convergent.
140 12 Hilbert Spaces

The Projection Theorem

If K is a closed convex subset of the plane, for every point x of the plane we can find
a unique point y in K which minimizes the distance to x, as the figure illustrates.

This fact remains true in general Hilbert spaces.

Proposition 12.2 (Projection Theorem I). Let K be a closed, convex and

nonempty subset of a Hilbert space X. Then for every x 2 X there exists a unique
y 2 K such that

kx  yk D min kx  zk :
z2K

The point y is called the projection of x onto K, written as y D PK x. If x 2 K we

have y D x.

Proof. To prove existence, for any given x 2 X we choose a minimizing sequence

fyn g in K satisfying limn kx  yn k D infz2K kx  zk DW d. Using the parallelogram
identity we get

2.kx  yn k2 C kx  ym k2 / D k2x  .yn C ym /k2 C kyn  ym k2 :

Since .yn C ym /=2 2 K by convexity, it follows that kx  .yn C ym /=2k  d and

therefore kyn  ym k2 2.kx  yn k2 C kx  ym k2 /  4d2 ! 0 as n; m ! 1.
Thus, yn is a Cauchy sequence. Since X is complete, there exists y D limn yn . As
K is closed, y belongs to K, and the continuity of the norm implies that kx  yk D
limn kx  yn k D d.
To prove uniqueness, let yQ 2 K with kx  yQ k D d. As above, the parallelogram
identity implies that

ky  yQ k2 D 2.kx  yk2 C kx  yQ k2 /  k2x  y  yQ k2 D 4d2  4kx  .y C yQ /=2k2 0 ;

because .y C yQ /=2 2 K. This yields y D yQ . t

u
The Projection Theorem 141

As the next figure shows, the angle between the difference vectors x  y and z  y
amounts to at least 90ı , for any z 2 K.

The projection can be characterized by this property.

Proposition 12.3 (Projection Theorem II). Let K be a closed, convex and

nonempty subset of a real Hilbert space X. Then for every x 2 X there is a
unique solution y 2 K of the inequalities

.x  y; z  y/ 0 for all z 2 K ; ()

and it holds y D PK x.

The system () of inequalities is usually called a variational inequality. One may
interpret it as the variational form of an inequality for the vector x  y.

Proof. If y; yQ 2 K solve () for a given x 2 X, we have .x  y; yQ  y/ 0 as well

as .x  yQ ; y  yQ / 0. Adding these inequalities yields 0  .x  y  x C yQ ; yQ  y/ D
kQy  yk2 , which implies uniqueness. We prove that y D PK x is a solution. For any
z 2 K and any t 2 .0; 1/ we have zt WD .1  t/y C tz 2 K, thus we obtain due to
x  zt D .x  y/ C t.y  z/ that

kx  yk2 kx  zt k2 D kx  yk2 C 2.x  y; t.y  z// C t2 kz  yk2

and therefore 0 2.x  y; y  z/ C tkz  yk2 after dividing by t. Passing to the limit
t ! 0 yields the assertion. t
u

In the case of a complex Hilbert space, too, the projection is characterized by a

variational inequality, namely

Re .x  y; z  y/ 0 for all z 2 K :

The proof is analogous.

142 12 Hilbert Spaces

Let y D PK x and yQ D PK xQ be the projections of the two points x; xQ 2 X. Adding

the inequalities Re .x  y; yQ  y/ 0 and Re .Qx  yQ ; y  yQ / 0 and using the
Cauchy-Schwarz inequality we get

kQy  yk2 D .Qy  y; yQ  y/ Re .Qx  x; yQ  y/ kQx  xkkQy  yk ;

and therefore

kPK xQ  PK xk kQx  xk :

This means that the projection PK W X ! K is Lipschitz continuous. Since PK x D x

for x 2 K, the Lipschitz constant equals 1 if K consists of more than a single point.
One says that the projection PK is non-expansive.
If in particular K D U is a closed subspace of X, the variational inequality turns
into the variational equation

.x  y; v/ D 0 for all v 2 U :

We obtain it by inserting z D y ˙ v and, in the complex case, also z D y ˙ iv

into the variational inequality. The projection PU is linear in that case, because the
variational inequalities for y D PU x and yQ D PU xQ and for arbitrary scalars ’ and “
immediately imply the variational equation

.Œ’x C “Qx  Œ’y C “Qy; v/ D 0 for all v 2 U ;

and therefore PU .’x C “Qx/ D ’PU x C “PU xQ . Summarizing these considerations we

obtain the following result.

Lemma. Let U be a closed subspace of a Hilbert space X. Then the projection

PU is a linear continuous mapping from X to U.

Example
If U is a closed subspace of the real Hilbert space L2 .SI /, and if f 2 L2 .SI /, then by the
projection theorem PU f equals the uniquely determined function in U which satisfies
Z Z
fg d D PU f  g d for all g 2 U : ()

A certain special case is of relevance in probability theory. Let  be a probability measure on

.S; A/, let A0 be a ¢-algebra with A0  A. For L2 .SI / DW L2 .SI A; / we consider the subspace
U D L2 .SI A0 ; / of all real-valued functions which are square integrable on S and measurable
The Projection Theorem 143

w.r.t. A0 . Since U itself is a Hilbert space, U is closed in L2 .SI /. The characterization () of the
projection given in this example can be written equivalently (Exercise 12.2) as
Z Z
f d D PU f d for all A0 2 A0 : ()
A0 A0

Together with the A0 -measurability of PU f, () thus asserts that PU f is the conditional expectation
of f.

We return to the general situation of a closed subspace U in a Hilbert space X. The

variational equation

.x  PU x; v/ D 0 for all v 2 U

means that x  PU x is orthogonal to U, thus x  PU x 2 U? . Therefore, PU also is

called orthogonal projection. Setting u D PU x and u? D x  u, we thus obtain an
orthogonal decomposition

x D u C u? ; u 2 U; u ? 2 U? ; ()

for which by Pythagoras

kxk2 D kuk2 C ku? k2 :

Proposition 12.4 (Orthogonal Decomposition). Let U be a closed subspace of

a Hilbert space X. Each x 2 X can be decomposed uniquely in the form (), and
one has u D PU x and u? D PU? x, as well as

kPU xk kxk :

Proof. We have U \ U? D f0g since .v; v/ D 0 and therefore v D 0 holds for all
v 2 U \ U? . This implies uniqueness, because for any two such decompositions
x D u C u? D uQ C uQ ? we get u  uQ D uQ ?  u? 2 U \ U? . It remains to show
that u? D PU? x. For any w 2 U? we have .x  u? ; w/ D .u; w/ D 0, therefore u?
solves the variational equation which characterizes PU? x. t
u

The orthogonal projection enables us to characterize the continuous linear function-

als on a Hilbert space. By a continuous linear functional we mean a continuous and
linear mapping ` from X to the scalar field R or C. The set of all those functionals `
144 12 Hilbert Spaces

forms the dual space X0 of X, we will treat it in more detail in the next chapter. For
a given ` 2 X0 one defines

k`k WD sup j`.x/j :

kxk1

There holds the following characterization which goes back to F. Riesz.

Proposition 12.5 (Riesz Representation Theorem). Let y be an element of a

Hilbert space X. Then

x 7! .x; y/

defines a continuous linear functional. Conversely, any ` 2 X0 can be represented

in the form `.x/ D .x; y/ for some y 2 X. Here, y is uniquely determined by `,
and k`k D kyk.

Proof. We already know that the first part of the assertion is true. For the converse
we consider for any given ` 2 X0 its kernel U D `1 .f0g/ which is a closed subspace
of X since ` is continuous. If ` D 0 then we must have y D 0, otherwise we may
choose some w 2 U? with `.w/ D 1. For every x 2 X we have x  `.x/w 2 U,
since `.x  `.x/w/ D `.x/  `.x/`.w/ D 0. We then get

.x; w/ D .x  `.x/w; w/ C .`.x/w; w/ D `.x/kwk2 :

The vector y D kwk2 w thus has the required property. If on the other hand 0 D
.x; y/  .x; yQ / D .x; y  yQ / for all x 2 X, then in particular 0 D .y  yQ ; y  yQ / and
therefore y D yQ . Finally we get k`k D kyk, since j`.x/j kykkxk kyk whenever
kxk 1, and `.y=kyk/ D kyk if y ¤ 0. t
u

With the preceding result we also have characterized all closed hyperplanes H in
a Hilbert space, because such hyperplanes coincide with the level sets f` D cg
associated to continuous linear functionals.

Corollary. Let  be a finite measure on a measurable space .S; A/. Every

continuous linear functional ` on the real Hilbert space L2 .SI / is of the form
Z
`.f/ D fg d

for some g 2 L2 .SI /, and k`k D kgk2 .

Bases in Hilbert Spaces 145

In the following chapter we will generalize this result to the spaces Lp .SI / for
1 p < 1.

Bases in Hilbert Spaces

When X is a vector space, a (vector space) basis B is a system of linearly

independent vectors in X, such that every x 2 X can be represented uniquely as
a linear combination
X
xD ’b b ()
b2B

with finitely many nonzero scalars ’b . This notion, while being central for the
treatment of finite-dimensional spaces, is largely useless in the context of infinite-
dimensional spaces. Instead, one considers representations for which (*) becomes a
series that converges in a suitable sense. The situation is particularly neat in Hilbert
space, because one has the scalar product at one’s disposal, and thus one is able to
form orthonormal systems.

Definition
A subset E of a Hilbert space X is called orthonormal system, if kek D 1 for each
e 2 E and .e; f/ D 0 for all e; f 2 E with e ¤ f.

Example

1. In the space `2 of square-summable sequences, the set E D fek W k 2 Ng of all

j
unit vectors (ek D •kj ) is an orthonormal system.
2. We consider the space L2 . ;  / WD L2 .. ;  /I œ/ and write more precisely
LC R
2 . ;  / and L2 . ;  / in order to specify the field of scalars. The set E D
fek W k 2 Zg, where

1
ek .t/ D p eikt ;
2 

is an orthonormal system in LC
2 . ;  /, since for any k ¤ j
Z   ˇtD 
1 1 1 ˇ
.ek ; ej / D ei.kj/t dt D ei.kj/t ˇ D0
2    2  i.k  j/ tD 

and obviously .ek ; ek / D 1. The set E D fQek W k 2 Zg, where

1 1 1
eQ 0 .t/ D p ; eQ k .t/ D p cos kt ; eQ k .t/ D p sin kt ; k  1;
2     
146 12 Hilbert Spaces

is also an orthonormal system in LC R

2 . ;  / and therefore, too, in L2 . ;  /, as
one sees from the formulas eQ 0 D e0 ,

1 1
eQ k D p .ek C ek / ; eQ k D p .ek  ek / ; k  1;
2 i 2

and from the properties of the scalar product, or directly via partial integration.

If ’1 ; : : : ; ’n are scalars and e1 ; : : : ; en are distinct elements of an orthonormal

system E, then

Xn 2 Xn
 
 ’k k
e D j’k j2 :
kD1 kD1

P P Pn Pn Pn
Indeed,
Pn . nkD1 ’k ek ; nlD1 ’l el / D kD1 lD1 ’k ’l .ek ; el / D kD1 ’k ’k . If
kD1 ’k e k D 0, we must have ’1 D    D ’n D 0. Therefore, any orthonormal
system is linearly independent.
In addition,
P orthonormal systems are useful for forming convergent series. A
series k1 yk in a Hilbert space P (or, more general, in a normed space) X is called
n
convergent, if the sequence sn D P kD1 yk of the partial sums converges in X. The
limit y D limn sn is also denoted by k1 yk .

Lemma. Let fe1 ; e2 ; : : :g be a countably infinite orthonormal system

P in the
Hilbert space X, and let ’1 ; ’2 ; : : : be a scalar sequence. Then k1 ’k ek
P
converges in X if and only if k1 j’k j2 < 1. We then have
X 2 X
 
 ’k ek  D j’k j2 :
k1 k1

Pn
Proof. For sn WD kD1 ’k ek and any m < n we have

 X
n 2 X
n
 
ksn  sm k2 D  ’k ek  D j’k j2 :
kDmC1 kDmC1

P
Therefore, fsn g is a Cauchy sequence if and only if k1 j’k j2 converges. As X is
complete, this is equivalent to the first assertion. P
The second assertion results from the fact that ksn k2 D nkD1 j’k j2 , passing to
the limit n ! 1 in view of the continuity of the norm. t
u

To apply the lemma we utilize the following proposition.

Bases in Hilbert Spaces 147

Proposition 12.6 (Bessel’s2 Inequality). Let fe1 ; e2 ; : : :g be a finite or count-

ably infinite orthonormal system in X. Then for every x 2 X we have
X
j.x; ek /j2 kxk2 :
k1

Pn Pn
Proof. Setting sn WD kD1 .x; ek /ek we get ksn k2 D kD1 j.x; ek /j2 , and therefore

X
n
.x; sn / D .x; ek /.x; ek / D ksn k2 D .sn ; sn / :
kD1

It follows that .x  sn ; sn / D 0, and the Pythagoras Theorem yields

X
n
kxk2 D ksn k2 C kx  sn k2  ksn k2 D j.x; ek /j2 :
kD1

This proves the assertion for the finite case and, letting n ! 1, also for the infinite
case. t
u

The Bessel inequality

P in combination with the foregoing lemma implies that
the expression k1 .x; ek /ek has a well-defined value for any x 2 X and any
orthonormal system fe1 ; e2 ; : : :g, either as a finite sum or as a convergent series.
We may interpret those expressions as projections on subspaces.
For this purpose, let span.E/ denote the subspaceP of X spanned by the set E  X.
It consists of all linear combinations of the form e2E ’e e with scalars ’e , only
finitely many of them being nonzero. We denote its closure by span.E/. If E is
finite, then span.E/ D span.E/, because every finite-dimensional normed space is
complete. If E is infinite, a result of functional analysis says that span.E/ ¤ span.E/
if X is complete. (If that would not be true, one might work with E as a vector space
basis.)

Proposition 12.7. Let U be a closed subspace and E D fe1 ; e2 ; : : :g a finite

or countably infinite orthonormal system in the Hilbert space X, assume that
span.E/ D U. Then for all x 2 X one has that
X X
PU x D .x; ek /ek ; kPU xk2 D j.x; ek /j2 :
k1 k1

2
FRIEDRICH W ILHELM BESSEL, 1784–1846, born in Minden, active at the observatory in
Königsberg. He worked in astronomy, mathematics, and geodesics.
148 12 Hilbert Spaces

P
Proof. Let y WD k1 .x; ek /ek . It then holds that (in the finite case and also, due to
the continuity of the scalar case, in the infinite case)
X
.x  y; el / D .x; el /  .x; ek /.ek ; el / D 0 :
k1

It follows that .x  y; z/ D 0 for all z 2 span.E/ and, due to the continuity of

the scalar product, .x  y; z/ D 0 for all z 2 span.E/ D U. Therefore, y satisfies
the variational equation characterizing the projection, and the first assertion follows.
The second assertion is a consequence of the foregoing lemma. t
u

Example

1. In the sequence space `2 we consider the orthonormal system E D fek W k 2 Ng

consisting of the standard unit vectors. For any x D .x1 ; x2 ; : : : / 2 `2 we have
.x; ek / D xk , and
X
n X
n
PU x D .x; ek /ek D x k ek
kD1 kD1

is the orthogonal projection onto U D span .fe1 ; : : : ; en g/.

2. In the function space LC2 . ;
p  / we investigate the orthonormal system given by
the functions ek .t/ D .1= 2 /eikt . For any f 2 LC2 . ;  /,
Z  
1
ck D .f; ek / D p f.t/eikt dt ; k 2 Z;
2   

is called the k-th Fourier coefficient3 of f. Setting U D span .fen ; : : : ; en g/, the
orthogonal projection

X
n X
n
PU f D .f; ek /ek D ck ek
kDn kDn

P
is just the n-th partial sum of the Fourier series k2Z ck ek of f. Concerning the
convergence of the Fourier series, the results below will provide information.

We now arrive at the notion which in Hilbert space replaces the notion of a vector
space basis.

3
JOSEPH FOURIER , 1768–1830, born in Auxerre, active in Paris at the École Polytechnique. In
the context of his fundamental contribution to heat conduction he utilized, for the first time,
trigonometric series for the representation of general functions.
Bases in Hilbert Spaces 149

Definition
An orthonormal system E is called orthonormal basis4 of X, if span.E/ is dense
in X, that is, if span.E/ D X.

Thus one only requires that every x 2 X can be represented as the limit of a sequence
in span.E/. Consequently, in Hilbert space any element x can be represented as the
limit of a series whose partial sums belong to span.E/.

Proposition 12.8. For a countably infinite orthonormal system E D fe1 ; e2 : : : g

in a Hilbert space X the following assertions are equivalent:

(i) E? D f0g.
(ii) X D span.E/, that is, E is an orthonormal basis.
(iii) There holds
1
X
xD .x; ek /ek for all x 2 X :
kD1

(iv) There holds

1
X
.x; y/ D .x; ek /.ek ; y/ for all x; y 2 X :
kD1

(v) The Parseval5 identity holds:

1
X
kxk2 D j.x; ek /j2 for all x 2 X :
kD1

Proof. (i) ) (ii): Let U WD span.E/. From E  U it follows that U?  E? , thus

U? D f0g and therefore U D X.
(ii) ) (iii): From U WD span.E/ D X it follows that PU x D x for all x 2 X and
thus the assertion, by the preceding proposition.
(iii) ) (iv): The series on the right side of (iv) converges absolutely, since
1
X 1
X 1
X
j.x; ek /.ek ; y/j j.x; ek /j2  j.y; ek /j2 kxk2 kyk2
kD1 kD1 kD1

4
Instead of an orthonormal basis one also speaks of a complete orthonormal system.
5
M ARC -ANTOINE PARSEVAL, 1755–1836, born in Rosière-aux-Salines, active in Paris.
150 12 Hilbert Spaces

due to the Cauchy-Schwarz inequality in `2 and the Bessel inequality. The assertion
now follows when on both sides of (iii) we take the scalar product with y and use
the continuity of the scalar product.
(iv) ) (v): We set y D x in (iv).
(v) ) (i): For every x 2 E? we have .x; ek / D 0 for all k, and therefore kxk D 0
by (v). t
u

Example
In the sequence space `2 , the orthonormal system E D fek W k 2 Ng consisting of the standard
1 2 2
P is ank orthonormal basis, since for any x D .x ; x ; : : : / 2 ` the sequence given
unit vectors
by sn D kn x ek belongs to span .E/ and converges to x; thus condition (ii) in the preceding
proposition is satisfied.

In order
p to prove that the orthonormal system given by the functions ek .t/ D
.1= 2 /eikt actually is an orthonormal basis of LC 2 . ;  /, we utilize arguments
from analysis. We P
owe to Fejér the idea of investigating, instead of the sequence of
partial sums sn D jkjn .f; ek /ek , the sequence defined by their arithmetic means

1 X X
m n
am WD .f; ek /ek :
m C 1 nD0 kDn

Proposition 12.9 (Fejér6 ). Let f W Œ ;   ! C be a continuous function

satisfying f. / D f. /. Then am converges uniformly to f on Œ ;  .

Proof. We have

Z   Z  
1 X X 1
m n
ik£ 1
am .t/ D f.£/e ikt
d£  e D f.£/Fm .t  £/ d£ ;
m C 1 nD0 kDn 2    2   

with the Fejér kernel

1 X X ik£
m n
Fm .£/ D e :
m C 1 nD0 kDn

6
LIPÓT FEJÉR , 1880–1959, born in Pécs, active in Klausenburg and Budapest. He worked in
harmonic analysis and potential theory.
Bases in Hilbert Spaces 151

R  R 
We have   Fm .£/ d£ D 2 , since   eik£ d£ D 0 for any k ¤ 0. Since Fm is a
2 -periodic function,
Z  
1
am .t/ D f.t  £/Fm .£/ d£ ;
2   

where we have extended f periodically outside Œ ;  . As f. / D f. /, this
extension preserves continuity.
Due to a certain trigonometric identity (Exercise 12.4) we have

1 sin2 . mC12
£/
Fm .£/ D : ()
m C 1 sin2 . 12 £/

For any 0 < • <   we estimate on Œ ;  :

Z
1 ˇˇ   ˇ
ˇ
jf.t/  am .t/j D ˇ .f.t/  f.t  £//Fm .£/ d£ˇ
2   
Z • Z Z  
1 1  •
jf.t/  f.t  £/jFm .£/ d£C C jf.t/f.t  £/jFm .£/ d£ :
2  • 2    •

For any given © > 0 we choose • > 0, according to the uniform continuity of f, such
that jf.t/  f.t  £/j < © for any j£j < •, and we choose m0 by virtue of () such that
Fm .£/ < © for all £ satisfying • j£j   and all m  m0 . It follows that

kf  am k1 .1 C 2kfk1 /©

for all m  m0 , and thus the assertion is proved. t

u

p
Corollary. The functions ek .t/ D .1= 2 /eikt , k 2 Z, constitute an orthonor-
mal basis of LC
2 . ;  /. The functions

1 1 1
eQ 0 .t/ D p ; eQ k .t/ D p cos kt ; eQ k .t/ D p sin kt ; k  1;
2     

constitute an orthonormal basis of LR

2 . ;  /.

p
Proof. Let U D span fek W k 2 Zg. Then am 2 U and kf  am k2 2 kf  am k1 .
By Fejér’s result, U is dense in the subspace V of LC 2 . ;  / consisting of the
continuous functions satisfying f. / D f. /. Modifying it near a boundary point,
we may approximate any arbitrary continuous function by functions from V with
an arbitrarily small error in the L2 norm, and because the continuous functions are
dense in LC 2 . ;  / by Proposition 7.7, this remains valid for arbitrary functions
152 12 Hilbert Spaces

from L2 . Thus, condition (ii) in the proposition characterizing orthonormal bases

holds. Since the Fejér kernel Fm is real-valued, und thus am , too, is real-valued
for any real-valued f, and because every real-valued function in U is a real linear
combination of the functions eQ k , the assertion also follows for LR
2 . ;  /. t
u
P
Summarizing, for functions f 2 L2 the Fourier series k2Z .f; ek /ek converges to
f in the sense of the norm of L2 .
It is a famous result due to Carleson that for any f 2 L2 the whole Fourier series
converges to f almost everywhere (and not only a subsequence of the partial sums,
a consequence of Propositions 6.4 and 6.6).

Exercises
12.1 Prove that in a complex Hilbert space X, the scalar product satisfies the identity

1
.x; y/ D .kx C yk2  kx  yk2 C ikx C iyk2  ikx  iyk2 /
4
for all x; y 2 X.
12.2 Let  be a probability measure on .S; A/, let A0  A be another ¢-algebra, let X D
L2 .SI A; / and U D L2 .SI A0 ; /. Prove that for every f 2 X and every h 2 U,
Z Z
f d D PU f d for all A0 2 A0
A0 A0

implies that h D PU f.
12.3 Let E D fe1 ; e2 ; : : : g be a countably infinite orthonormal system in a Hilbert space X, let
x 2 X. Then
X X
.x; e .k/ /e .k/ D .x; ek /ek
k1 k1

P mapping   W N ! N. This property is termed

for any reordering of E given by a bijective
unconditional convergence of the series k1 .x; ek /ek .
12.4 Prove that for every    £   , £ ¤ 0 the trigonometric identity

X
m X
n
sin2 . mC1 £/
2
eikt D
nD0 kDn
sin2 . 12 £/

is valid. Hint: Use the formula giving the partial sums of the geometric series as well as the
trigonometric identity
h1 i2
4 sin2 ® D 4 .ei®  ei® / D 2  e2i®  e2i® :
2i
Banach Spaces
13

In the preceding chapters, several times already we have interpreted functions as

elements of function spaces. We now deepen this view, looking more closely at
continuous linear functionals on such spaces. We will characterize them in two
important cases intimately linked to integration theory, namely for the spaces of
p-integrable functions and of continuous functions. The notion of a Banach spaces
provides the appropriate functional analytic framework.
To this end we first want to acquaint the reader a little closer with Banach spaces.
We recall the definition of a norm in a vector space.

Definition
A norm is a mapping on a real or complex vector space X which to each vector
x 2 X associates a nonnegative number kxk with the properties

(i) Definiteness: kxk D 0 if and only if x D 0.

(ii) Positive homogeneity: k’xk D j’jkxk for all scalars ’.
(iii) Triangle inequality: kx C yk kxk C kyk for all x; y 2 X.

We call X, or more precisely .X; k  k/, a normed space. If (ii) and (iii) hold, but
not necessarily (i), we speak of a seminorm on X.

From the triangle inequality, due to kxk kx  yk C kyk we immediately obtain

the reverse triangle inequality

jkxk  kykj kx  yk ; x; y 2 X :

Example
As shown in Chap. 6, the spaces Lp .SI / of the p-integrable (1  p < 1) resp., in the case
p D 1, measurable and essentially bounded (equivalence classes of) functions on a measure

© Springer International Publishing Switzerland 2015 153

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0_13
154 13 Banach Spaces

space .S; A; / are real normed spaces equipped with the p-norms
Z 1=p
kfkp D jfjp d ; 1  p < 1; kfk1 D N1 .f/ :

If we admit complex-valued functions f, we likewise obtain complex normed spaces. (A complex-

valued function is called measurable, resp. integrable if its real and its imaginary part are
measurable, resp. integrable. The properties of the norm are proved in the same way as in the
real case.)

The spaces Rd and Cd with d < 1, equipped with the norms

X
d 1=p
kxkp D jxk jp ; 1 p < 1; kxk1 D sup jxk j ;
kD1 k

can be viewed as special cases of Lp spaces. For  we choose the counting measure
on S D f1; : : : ; dg; here xk denotes the k-th component of the vector x. In the case
d D 1 we obtain the scalar field, interpreted as a normed space with kxk D jxj.
For d D 1 we obtain the spaces `p of sequences which are summable to the
p-th power resp. bounded, consisting of those sequences x D .x1 ; x2 ; : : : / for which
kxkp is finite. With the choice S D N and the counting measure for , they too
become special cases of the spaces Lp ./.

Example
For an arbitrary set S, the vector space of all bounded (real- or complex-valued) functions on S is
a normed space if equipped with kfk1 D supx2S jf.x/j. When S is a compact metric space, the
same definition yields a norm on the vector space C.S/ of all continuous functions on S, too.

The preceding example illustrates the fact that every subspace U of a normed space
X becomes a normed space if we restrict the norm on X to U.

Definition
A sequence x1 ; x2 ; : : : in a normed space X is said to converge to the limit x 2 X,
written as

x D lim xn ; or xn ! x ;
n!1

if limn!1 kxn  xk D 0.

In the case X D Lp .SI / for 1 p < 1 this is just the convergence in p-mean.
Convergence in the sup-norm k  k1 in a function space is synonymous to uniform
convergence (resp. to uniform convergence almost everywhere).
13 Banach Spaces 155

It is an immediate consequence of the definition of a norm that sums and scalar

multiples of convergent sequences converge to the sum, resp. multiple of their limits.
In a normed space X, the closed ball of radius r > 0 around a point x 2 X is given
by fy W ky  xk rg, we denote it by Br .x/. Instead of Br .0/ we briefly write Br ,
instead of B1 simply B; the latter is called the (closed) unit ball in X. We obviously
have Br .x/ D x C rB.
Setting d.x; y/ D kx  yk, any norm on a vector space X generates a translation
invariant metric, that is, it holds that d.x C z; y C z/ D d.x; y/ for all x; y; z 2 X.
Via restriction of d, any subset M of X becomes a metric space.

Definition
A complete normed space is called a Banach space.

The above-mentioned spaces Lp .SI / (S measure space) and C.S/ (S compact

metric space) are Banach spaces. For Lp we have proved this in Chap. 6. To prove
the completeness of C.S/ one shows that every Cauchy sequence of continuous
functions converges uniformly
p to its pointwise limit (Exercise 13.1).
Setting kxk WD .x; x/, every Hilbert space becomes a Banach space. A
subspace U of a Banach space X obviously is itself a Banach space if and only if it
is closed in X. Every finite-dimensional normed space (and therefore every finite-
dimensional subspace of a normed space, too) is a Banach space (Exercise 13.3).
The reverse triangle inequality jkxk  kykj kx  yk says that the norm is
a Lipschitz continuous function on X with Lipschitz constant equal to 1. Such
functions are called non-expansive. The distance function, defined for M  X and
x 2 X by

d.x; M/ D inf d.x; z/ D inf kx  zk ;

z2M z2M

when viewed as a function of x, too is non-expansive (Exercise 13.2).

Definition
Two norms k  ka and k  kb on a vector space X are called equivalent if there exist
constants c1 ; c2 > 0 such that

c1 kxka kxkb c2 kxka ; for all x 2 X.

This notion indeed yields an equivalence relation on the set of all norms on X, as
one immediately checks. Whenever two norms are equivalent, they generate the
same topology, that is, in either norm the same sequences are convergent, the same
sets are open and closed, and so on.

Proposition 13.1. On Rd and Cd , d 2 N, all norms are equivalent.

156 13 Banach Spaces

Proof. It suffices to show that each norm k  k is equivalent to the supremum norm
k  k1 . Let ei be the standard unit vectors in X, X D Rd or Cd . Then

X
d X
d
kxk jxi jkei k c2 kxk1 ; c2 WD kei k :
iD1 iD1

Moreover, since

jkxn k  kxkj kxn  xk c2 kxn  xk1 ;

the real-valued function f.x/ D kxk is continuous on .X; k  k1 /, and thus on the
compact set S WD fx W kxk1 D 1g it attains its minimum c1 , which is strictly
positive since the norm is definite. Consequently, for all x ¤ 0 in X, c1 kkxk1
1 xk
and therefore c1 kxk1 kxk. t
u

In infinite-dimensional spaces the assertion of the preceding proposition does not

hold. For example, let us consider on C.Œ0; 1/, besides the supremum norm, the
R1
integral norm kfk1 D 0 jf.x/j dx; there are sequences f1 ; f2 ; : : : with kfn k1 D 1,
but kfn k1 ! 0. (See Exercise 13.2.) Different norms thus yield different conver-
gence statements.
The following pictures exhibit the unit balls of the p-norms in R2 for p D 1; 2; 1.

Continuous Linear Mappings

Between finite-dimensional spaces, all linear mappings are continuous. In infinite

dimensions this is no longer true. A counterexample is given by any linear mapping
T W `2 ! R satisfying Ten D n for the standard unit vectors en . Indeed, setting
xn D n1 en we obtain xn ! 0 because of kxn k D n1 , but Txn D 1 ¤ 0 D T.0/.
The continuity of a linear mapping T between normed spaces X and Y can be
characterized by several equivalent properties. A number C > 0 is called a bound
for a subset M of X, if kxk C for all x 2 M; if such a bound exists, M is called
bounded (in X). One immediately realizes that (finite) sums and scalar multiples
of bounded sets are again bounded. The mapping T is called bounded on M if the
image T.M/ is bounded in Y.
Continuous Linear Mappings 157

Proposition 13.2. For any linear mapping T between normed spaces X and Y
the following statements are equivalent:

(i) T is continuous on X.
(ii) T is continuous in 0.
(iii) There exists a ball Br on which T is bounded.
(iv) The image T.M/ of every bounded set M is bounded.
(v) There exists a C > 0 such that kTxkY CkxkX for all x 2 X.

Proof. It is obvious that (i) implies (ii). To prove (ii) ) (iii) we use contraposition.
Let x1 ; x2 ; : : : be a sequence satisfying 0 < kTxn k ! 1 and w.l.o.g. kxn k D r;
setting zn D kTxn k1 xn we get that zn ! 0 as well as kTzn k D 1, thus T is not
continuous in 0. To deduce (iv) from (iii), let M be bounded. Then M  tBr for a
suitable t > 0 and T.M/  tT.Br /, thus T.M/ is bounded. To deduce (v) from (iv)
we note that for any x ¤ 0 we have kTxkY D kxkX kT.kxk1 X x/kY CkxkX , if C
is a bound for T.B1 / in Y. To deduce (i) from (v), let x1 ; x2 ; : : : be a sequence with
xn ! x, then we have kTxn  TxkY D kT.xn  x/kY Ckxn  xkX ! 0. t
u

Definition
By L.XI Y/ we denote the set of all continuous linear mappings between normed
spaces X and Y. When Y is the scalar field, we call it the dual space of X,
denoted by X0 . Elements of X0 are called functionals, elements of L.XI Y/ are
called operators.

For any T 2 L.XI Y/, the set fx 2 X W Tx D yg, with y 2 Y given, is a closed affine
subspace of X. If, in particular, ` W Rd ! R is linear (d < 1) and c is a scalar,
we obtain hyperplanes H D f` D cg, decomposing Rd into two open half-spaces
f` > cg and f` < cg. This fact remains valid for functionals ` 2 X0 on arbitrary
normed spaces X and serves as a starting point for geometric considerations in
Banach spaces.
Since sums and scalar multiples of continuous linear mappings are again
continuous and linear, X0 , and more generally L.XI Y/ are vector spaces. The
characterization (v) of their continuity in the preceding proposition yields that

kTxk
kTk WD sup kTxk D sup kTxk D sup
kxk1 kxkD1 kxk¤0 kxk

is a finite nonnegative number; it is called the operator norm of T 2 L.XI Y/. We

obviously have

kTxk kTkkxk
158 13 Banach Spaces

for all x 2 X, and kTk is the smallest constant C with the property that kTxk Ckxk
for all x. This implies that the composition S ı T of two continuous linear mappings
satisfies, because k.S ı T/xk kSkkTxk kSkkTkkxk,

kS ı Tk kSkkTk :

Proposition 13.3. Equipped with the operator norm, L.XI Y/ becomes a

normed space. If Y is complete, L.XI Y/ is a Banach space. In particular, the
dual space X0 is a Banach space.

Proof. Definiteness holds, since kTk D 0 if and only if Tx D 0 for all x, which is the
same as T D 0. Positive homogeneity and the triangle inequality are consequences
of elementary properties of the supremum. Let T1 ; T2 ; : : : be a Cauchy sequence in
L.XI Y/. Since kTn x  Tm xk kTn  Tm kkxk, T1 x; T2 x; : : : is a Cauchy sequence
in Y for every fixed x. When Y is complete, there exists limn Tn x DW Tx, and one
can verity (Exercise 13.4), that the mapping T W X ! Y thus defined is linear and
continuous, and that Tn ! T in L.XI Y/. t
u

On X0 , the operator norm is called dual norm, and for ` 2 X0 one usually terms

j`.x/j
k`k D sup j`.x/j D sup j`.x/j D sup
kxk1 kxkD1 kxk¤0 kxk

simply the norm of `.

Example
R 0
The formula `.f/ D f d defines on X D L R 1 .SI /,  being a measure, a functional ` 2 X which
satisfies j`.f/j  kfk1 as well as `.1A / D 1A d D k1A k1 for measurable A with .A/ < 1,
thus k`k D 1. If moreover S is a compact metric space,  finite, and X D .C.S/; kk1 /, then again
` 2 X0 , but this time k`k D .S/, since j`.f/j  .S/kfk1 and `.1/ D .S/. In particular, the
Dirac measure •x for x 2 S defines a functional •x 2 C.S/0 with k•x k D 1, one has •x .f/ D f.x/.
(One also terms it Dirac functional.) On the other hand, on X D L1 .SI œ/, S D .a; b/, one cannot
obtain a continuous linear functional from the Dirac measure •x , compare Exercise 13.5.

Example
If U is a closed subspace of a Hilbert space X, the orthogonal projection PU considered in the
preceding chapter defines an operator in L.X/ WD L.XI X/ with kPU k D 1 whenever U ¤ f0g.
The Dual Space of Lp .SI / 159

Example
For a finite measure  we consider the spaces X D Lp .SI / and Y D Lr .SI /, 1  r < p < 1.
If f 2 Lp .SI /, using Hölder’s inequality with the decomposition 1 D r=p C .p  r/=p we see that

Z  1r Z  1p  Z  pr
pr pr
kfkr D jfjr d  jfjp d 1 d D Ckfkp ; C D .S/ pr :

Therefore Lp .SI /  Lr .SI / holds, and the embedding of Lp .SI / into Lr .SI / defined by
T.f/ D f is linear and continuous. The inclusion is proper in general, as for example in the case
S D .0; 1/ and  D œ the function defined by f.t/ D t1=p demonstrates.

Example
We consider an integral operator of the form
Z
.Tf/.x/ D k.x; y/f.y/ .dy/ : ()

For a given kernel k it maps a function f to a function Tf. We consider measure spaces .S0 ; A0 ; /
and and .S00 ; A00 ; / as in Chap. 8 and assume that k W S0  S00 ! R is measurable. Let moreover
Z
Ck WD sup jk.x; y/j .dx/ < 1 :
y2S00

For any f 2 L1 .S00 I / we have

“ Z
jk.x; y/f.y/j.dx/.dy/  Ck jf.y/j .dy/ D Ck kfk1 < 1 : ()

As explained in Chap. 8, the right side of ./ defines an element of L1 .S0 I /. Thus, ./ defines an
operator T W L1 .S00 I / ! L1 .S0 I /. T is obviously linear; moreover, it is continuous by virtue of
the inequality kTfk1  Ck kfk1 which is valid due to ./.

Depending upon the properties of the kernel function k, integral operators of the
form ./ act on various different function spaces. The classical starting point is
given by the Hilbert space case T W L2 .0; 1/
’ ! L2 .0; 1/ with  D  D œ, in this
case it suffices for T being continuous that jk.x; y/j2 dx dy is finite.

The Dual Space of Lp .SI /

For a given measure space .S; A; / we consider the spaces Lp .SI / where p 2
Œ1; 1. Let q be the exponent dual to p, that is, 1=p C 1=q D 1 (here 1 is dual to 1,
and 1 is dual to 1). When g 2 Lq .SI /, the mapping
Z
f 7! fg d
160 13 Banach Spaces

defines a continuous linear functional on Lp .SI / because

ˇZ ˇ
ˇ ˇ
ˇ fg dˇ kgkq kfkp
S

due to Hölder’s inequality. It turns out that for p < 1 every continuous linear
functional on Lp .SI / can be represented in this way. We restrict ourselves to the
case where the measure  is finite.

Proposition 13.4. Let  be a finite measure on a measurable space .S; A/, let
1 p < 1. Every continuous linear functional ` on Lp .SI / has the form
Z
`.f/ D fg d

for some g 2 Lq .SI /. The mapping g 7! ` is linear and isometric, that is,
k`k D kgkq holds for the dual norm of `.

In other words: The dual space of Lp .SI / is isometrically isomorphic to the space
Lq .SI /.
R
Proof. For any given g 2 Lq .SI / we set G.f/ WD fg d. As we already have
seen above, G is well-defined, continuous and linear, and satisfies kGk kgkq . The
mapping g 7! G is obviously linear. To prove the reverse inequality kGk  kgkq in
the case p > 1, we consider the function

f D .sign g/jgjq1 :

We have fg D jgjq D jfjp due to p.q  1/ D q, and

ˇZ ˇ Z 1=q  Z 1=p
ˇ ˇ
ˇ fg dˇ D jgjq d jgjq d D kgkq kfkp ;
S

so altogether kGk D kgkq in the case p > 1. In the case p D 1 we set

1
An D fjgj  kgk1  g ; fn D 1An sign g :
n

We have kfn k1 D .An / and

Z Z  1  1
`.fn / D fn gd D 1An jgj d  .An / kgk1  D kfn k1 kgk1  :
n n

It follows that kGk  kgk1  1=n and thus kGk  kgk1 .

The Dual Space of Lp .SI / 161

It remains to show, and this constitutes the main part of the proof, that every
functional ` 2 Lp .SI /0 can be represented in this way.

1. We want to prove that

.A/ D `.1A / ; A  S measurable,

defines a signed finite measure on A. First, we have .;/ D `.0/ D 0. Next, let
A1 ; A2 ; : : : be any sequence of disjoint measurable sets. Setting A D [n1 An we
get

 X
m p  [
m 
 
1A  1 An  D  A n An ! 0
p
nD1 nD1

in the limit m ! 1 due to continuity of measures, and thus, since ` is

continuous,

X
m  X
m X
.A/ D `.1A / D lim ` 1An D lim `.1An / D .An / :
m!1 m!1
nD1 nD1 n1

The set function  therefore is ¢-additive and thus a signed measure satisfying
j.S/j D j`.1/j < 1.
2. Let  D C   be the Jordan decomposition of  into the measures C and
 according to Proposition 9.9, which are both finite since  is finite. We have
C
, 
, since it follows from .A/ D 0 that 0 D `.1A0 / D .A0 /
for all A0  A and therefore C .A/ D  .A/ D 0. By the Radon-Nikodym
Theorem there exist densities dC D gC d,  D g d, which are integrable
since ˙ is finite. We set g D gC  g and obtain for any measurable A
Z
`.1A / D .A/ D g d
A

for a suitable integrable function g.

3. We prove that
Z
`.f/ D fg d ()

for bounded measurable functions f. Indeed, () holds for f D 1A and therefore,
due to linearity, for signed elementary functions. Since the latter are dense in
L1 .SI / (Exercise 13.6), () holds as claimed.
162 13 Banach Spaces

4. We show that g 2 Lq .SI /. In the case p > 1 we consider the sequence of
bounded measurable functions defined by

fn D 1An .sign g/jgjq1 ; An D fjgj ng :

As shown above during the proof, we have jfn jp D 1An jgjq and by virtue of 3.
Z Z Z 1=p
1An jgj d D
q
fn g d D `.fn / k`kkfn kp D k`k 1An jgj d
q
:

It follows that k1An gkq k`k and moreover, due to monotone convergence,
kgkq k`k, as jgjq D supn 1An jgjq almost everywhere. In the case p D 1 we set
A D fjgj > k`kg and obtain, letting f D 1A sign g,
Z Z
1A jgj d D fg d D `.f/ k`kkfk1 D k`k.A/ :

R
If .A/ > 0, we would have .A/k`k < 1A jgj d by definition of A, a
contradiction. Consequently, jgj k`k almost everywhere, thus kgk1 k`k
in the case p D 1.
5. Both sides of () define continuous functionals on Lp .SI / which conincide on
the dense subset L1 of Lp , and therefore on all of Lp . Thus we have proved the
representation of ` as claimed.
t
u

The Banach Space M.S/ of Signed Finite Measures

Let .S; A/ be a measurable space. The set

M.S/ D f j  W A ! R is a signed finite measureg

becomes a real vector space when equipped with the addition and scalar multiplica-
tion

.1 C 2 /.A/ D 1 .A/ C 2 .A/ ; .’/.A/ D ’.A/ :

We consider the Jordan decomposition  D C   of  into finite measures ˙

according to Proposition 9.9:

C .A/ D sup .A0 / ;  .A/ D  inf

0
.A0 / D ./C .A/
A0 A A A
The Banach Space M.S/ of Signed Finite Measures 163

for measurable A. From this we immediately obtain that

.1 C 2 /C .A/ C C
1 .A/ C 2 .A/ ; .1 C 2 / .A/  
1 .A/ C 2 .A/ ()

for any 1 ; 2 2 M.S/. The formula jj D C C defines another finite measure,
called the variation of . The triangle inequality for the positive and negative part
extends to the variation, because by (),

j1 C 2 j.A/ j1 j.A/ C j2 j.A/ :

For scalar multiples we obtain j’j.A/ D j’jjj.A/ from the Jordan decomposition
’ D .’/C  .’/ , where in the case ’ < 0 we only have to take into account
that .’/C D ’ and .’/ D ’C . It follows from the exposition above
that

kk D jj.S/

defines a norm on M.S/, since kk D 0 implies that C .S/ D  .S/ D 0 and
therefore  D 0. For any  2 M.S/ and any measurable A we thus obtain

j.A/j jj.A/ kk : ()

Proposition 13.5. The space M.S/ is a Banach space when equipped with the
norm kk D jj.S/.

Proof. Only the completeness remains to be proved. Let .n / be a Cauchy sequence
in M.S/. For any measurable A, .n .A// is a Cauchy sequence in R because of
(). We set

.A/ D lim n .A/ :

n!1

We want to prove that the set function  is a signed finite measure. We have
.;/ D 0. Since we may interchange the limit with finite sums,  is finitely additive.
Moreover, again because of (), one has that

j.A/  n .A/j D lim jm .A/  n .A/j lim sup km  n k

m!1 m!1
164 13 Banach Spaces

for all A. In order to prove that  is ¢-additive,

S we consider a sequence A1 ; A2 ; : : :
of disjoint measurable sets and set A D k1 Ak . For all natural numbers n; l it
holds that
ˇ X
l ˇ ˇ X
l ˇ
ˇ ˇ ˇ ˇ
ˇ.A/  .Ak /ˇ j.A/  n .A/j C ˇn .A/  n .Ak /ˇ
kD1 kD1

ˇ [l  [
l ˇ
ˇ ˇ
C ˇn Ak   Ak ˇ ;
kD1 kD1

where we have made use of the finite additivity of , already proved above. Passing
to the limit superior in l while keeping n fixed yields, since n is ¢-additive,

ˇ X
l ˇ
ˇ ˇ
lim sup ˇ.A/  .Ak /ˇ 2 lim sup km  n k :
l!1 m!1
kD1

Passing once more to the

P limit superior, this time in n, gives 0 on the right side, and
consequently .A/ D k1 .Ak /. t
u

The Dual Space of C.S/

Let S be a compact metric space, equipped with the Borel ¢-algebra B, and let
C.S/ be the Banach space of all real-valued continuous functions on S. By the
representation Proposition 11.3, we may represent every positive linear functional
` on C.S/ as an integral with respect to some finite measure . If we also allow
signed measures, we can find such a representation for arbitrary continuous linear
functionals on C.S/.
A signed finite measure  is called regular, if C and  are regular (or
equivalently, if jj is regular). From Proposition 7.6 it follows that every signed
finite measure on the compact metric space S is regular.

Proposition 13.6. Let S be a compact metric space. Every continuous linear

functional ` on C.S/ can be uniquely represented in the form
Z
`.f/ D f d

with a signed finite regular measure . The mapping  7! ` is linear and

isometric, that is, k`k D kkM.S/ holds for the dual norm of `.
The Dual Space of C.S/ 165

Proof. For any given  2 M.S/ the functional ` W C.S/ ! R, given by

Z
`.f/ D f d ;

is linear. Due to the Jordan decomposition  D C   we obtain the estimate

Z Z
C
j`.f/j jfj d C jfjd kfk1 kC k C kfk1 k k kfk1 kk ;

and thus ` is continuous with k`k kk, therefore ` 2 C.S/0 . In order to prove the
reverse inequality k`k  kk, let AC and A WD AcC be the sets belonging to the
Jordan (resp. Hahn) decomposition satisfying C .A / D  .AC / D 0. Since  is
regular, for arbitrary © > 0 we find compact sets KC  AC and K  A such that
˙ .A˙ / ˙ .K˙ / C ©. We now define the continuous functions

f˙ .x/ D .1  ’1 d.x; K˙ //C ; f D fC  f ;

where ’ WD dist .KC ; K / D infx˙ 2K˙ d.xC ; x /. We have f D 1 on KC , f D 1

on K , and kfk1 1. We estimate
Z Z Z Z
f d D f d C f d C f d
S KC K .KC [K /c

 jj.KC / C jj.K /  jj..KC [ K /c / D 2.jj.KC / C jj.K //  jj.S/

 2.jj.AC / C jj.A /  2©/  jj.S/ D jj.S/  4© D kk  4© :

Therefore, k`k  `.f/  kk4©, and consequently k`k  kk, letting © ! 0. The
isometry k`k D kk just proved implies the uniqueness of  in the representation
of `, since the mapping  7! ` is obvious linear.
It remains to show that such a  exists for any given ` 2 C.S/0 . In order to
achieve this, we represent ` as the difference of two positive linear functionals and
apply the Riesz representation Proposition 11.3. We define

`C .f/ D sup `.®/ ; if f  0.

0®f

For any such ® we have k®k1 kfk1 , therefore `.®/ k`kk®k1 k`kkfk1 ,
and thus 0 `C .f/ < 1 for f  0. Immediately from the definition we obtain that

`C .f/ C `C .g/ `C .f C g/ ; `C .’f/ D ’`C .f/ ;

166 13 Banach Spaces

for any f; g  0 and any ’  0. In order to prove the reverse inequality, let 0 ®
f C g. We have

® D min.®; f/ C .®  f/C ; .®  f/C g ;

and therefore
 
`.®/ D ` min.®; f/ C `..®  f/C / `C .f/ C `C .g/

by the definition of `C . Passing to the supremum with respect to ® yields the

inequality `C .f C g/ `C .f/ C `C .g/. We conclude that

`C .f/ C `C .g/ D `C .f C g/ ; if f; g  0. ()

We now define, for arbitrary f 2 C.S/,

`C .f/ D `C .fC /  `C .f / :

The linearity of `C on C.S/ is proved in the same manner as for the Lebesgue
integral, namely we apply `C in view of () to the identities

.f C g/C C f C g D .f C g/C C fC C gC ;
.f/C C fC D .f/ C f :

Besides `C , also ` WD `C  ` is a positive linear functional on C.S/. The Riesz

representation Proposition 11.3 yields finite measures C and  such that
Z Z
C 
` .f/ D f dC ; ` .f/ D f d :

Finally,  D C   yields the sought-after representation of `. t

u

Exercises
13.1 Prove that the space C.S/ of continuous functions on a compact metric space S, equipped
with the supremum norm kfk1 D supx2X jf.x/j, is a Banach space.
13.2 Let M be a subset of a normed space X. Prove that the distance function d.x; M/ D
infz2M kx  zk is nonexpansive when viewed as a function of x.
13.3 1. Let T W X ! Y be a linear mapping between normed spaces spaces X and Y. Prove that
if X is finite-dimensional, then T is continuous.
2. Prove that every finite-dimensional normed space is a Banach space.
13.4 Completeness of L.XI Y/
Let X; Y be Banach spaces, let T1 ; T2 ; : : : be a Cauchy sequence in L.XI Y/, let T W X ! Y
be defined by Tx D limn!1 Tn x. Prove:
(i) T is linear.
The Dual Space of C.S/ 167

(ii) The set fkTn kgn2N is bounded, and T is continuous.

(iii) limn!1 kTn  Tk D 0.
13.5 We interpret the set of continuous functions f W Œ0; 1 ! R as a subspace U of the Banach
space L1 .Œ0; 1I œ/, equipped with the L1 norm. Prove:
(i) U is not closed in X, and therefore not complete.
(ii) Let x 2 Œ0; 1. The functional defined by •x .f/ WD f.x/ is not continuous on U.
13.6 Let .S; A; / be a measure space. Prove that for every f 2 L1 .SI / there exists a sequence
f1 ; f2 ; : : : of signed elementary functions such that kfn  fk1 ! 0.
Bibliography

1. H. Bauer, Measure and Integration Theory (de Gruyter, Berlin/New York, 2001)
2. J. Elstrodt, Maß- und Integrationstheorie, 6. Aufl. (Springer, 2009)
3. L.C. Evans, R.F. Gariepy, Measure Theory and Fine Properties of Functions (CRC, Boca
Raton, 1992)
4. P.R. Halmos, Measure Theory (Van Nostrand, New York, 1950/Springer, New York, 1974)
5. A. Klenke, Probability Theory, 2nd edn. (Springer, London/New York, 2013)
6. A. Pietsch, History of Banach Spaces and Linear Operators (Springer, London, 2007)
7. W. Rudin, Principles of Mathematical Analysis, 3rd edn. (McGraw-Hill, New York, 1976)
8. W. Rudin, Real and Complex Analysis, 3rd edn. (McGraw-Hill, Singapore, 1986)
9. R. Schilling, Measures, Integrals and Martingales (Cambridge University Press, Cam-
bridge/New York, 2005)
10. The MacTutor history of mathematics archive, http://www-history.mcs.st-and.ac.uk/
11. D. Werner, Funktionalanalysis, 6. Aufl. (Springer, 2007)
12. D. Werner, Einführung in die Höhere Analysis, 2. Aufl. (Springer, 2009)

© Springer International Publishing Switzerland 2015 169

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0
Index

Absolutely continuous, 88, 100 Functional, 157

Almost everywhere (a.e.), 22

Generator, 8
Banach-Tarski paradox, 4
Bilinear form, 138
Borel measurable, 9 Hausdorff dimension, 133
Bound, 156
Bounded, 156
Inequality
Completion, 23 Bessel, 147
Conditional expectation, 91, 143 Cauchy-Schwarz, 45, 138
Convergence Hölder, 45
almost everywhere, 23 Jensen, 46
in p-mean, 53 Markov, 30
in measure, 58 Minkowski, 54
in the norm, 154 Integrable, 41
Convolution, 81 equiintegrable, 44, 61
Riemann, 49
Integral, 30
Decomposition double, 73
Hahn, 89, 106 Lebesgue, 30, 48
Jordan, 106 parameter dependent, 47
Lebesgue, 92 Riemann, 49
Density, 36 Integral operator, 159
Distance function, 155
Dual norm, 158
Dual space, 157 Kernel, 84
of a Hilbert space, 144
Dynkin system, 64
Measurable mapping, 7
Measure, 19
Essential supremum, 54 Dirac, 20
discrete, 20
finite, 19
Fourier series, 148 Hausdorff, 133
Function image, 21
Cantor, 99 Jordan, 3
measurable, 14 Lebesgue, 24, 64, 80, 123
of bounded variation, 102 Lebesgue-Borel, 24

© Springer International Publishing Switzerland 2015 171

M. Brokate, G. Kersting, Measure and Integral, Compact Textbooks
in Mathematics, DOI 10.1007/978-3-319-15365-0
172 Index

locally finite, 122 Jordan, 3

mutually singular, 88 measurable, 7
outer, 118 non-measurable, 27
probability, 19 null, 22
product, 79, 128 Singular, 88
Radon, 69 Space
¢-finite, 19 Banach, 57, 155
signed, 105 Hilbert, 57, 139
Monotonicity principle, 15 measurable, 7
measure, 19
normed, 153
Non-expansive, 142, 155 product, 12
Norm, 138, 153

Theorem
Operator, 157
Beppo Levi, 31
Operator norm, 157
Carathéodory, 119
Orthogonal, 139
comparsion, 67
Orthogonal complement, 139
Dini, 125
Orthogonal decomposition, 143
dominated convergence, 43
Orthogonal projection, 143
Egorov, 27
Orthonormal basis, 149
extension, 120
Orthonormal system, 145
Fatou’s Lemma, 32
Fejér, 150
Parallelogram identity, 139 Fubini, 74, 77, 79
Projective limit, 128 Kolmogorov, 128
Lebesgue, 43, 95
monotone convergence, 31
Regular, 69 projection, 140, 141
inner, 69 Pythagoras, 139
outer, 66 Radon-Nikodym, 88
Riesz, 60
Riesz-Fischer, 55
¢-additivity, 2, 19 Riesz representation, 124, 144
¢-algebra, 5, 7 smoothing, 83
Borel, 5, 9, 13 Steinhaus, 71
generated, 8, 11 uniqueness, 63
product, 12 Vitali, 61
trace, 8 Vitali’s Covering Lemma, 96
¢-continuity, 20 Transformation formula, 36
¢-subadditivity, 20 of Jacobi, 114
Scalar product, 137 Triangle inequality, 153
Seminorm, 153
Semiring, 68
Set Variation of a measure, 163
Borel, 5, 9 Variational equation, 142
Cantor, 93, 133 Variational inequality, 141