Professional Documents
Culture Documents
William P. Ziemer
Department of Mathematics, Indiana University, Bloomington, In-
diana
E-mail address: ziemer@indiana.edu
Contents
Preface 7
Chapter 1. Preliminaries 1
1.1. Sets 1
1.2. Functions 3
1.3. Set Theory 7
Exercises for Chapter 1 10
Chapter 2. Real, Cardinal and Ordinal Numbers 13
2.1. The Real Numbers 13
2.2. Cardinal Numbers 25
2.3. Ordinal Numbers 32
Exercises for Chapter 2 36
Chapter 3. Elements of Topology 39
3.1. Topological Spaces 39
3.2. Bases for a Topology 45
3.3. Metric Spaces 47
3.4. Meager Sets in Topology 49
3.5. Compactness in Metric Spaces 53
3.6. Compactness of Product Spaces 56
3.7. The Space of Continuous Functions 57
3.8. Lower Semicontinuous Functions 66
Exercises for Chapter 3 69
Chapter 4. Measure Theory 77
4.1. Outer Measure 77
4.2. Caratheodory Outer Measure 86
4.3. Lebesgue Measure 88
4.4. The Cantor Set 93
4.5. Existence of Nonmeasurable Sets 95
4.6. Lebesgue-Stieltjes Measure 96
3
4 CONTENTS
4.7. Hausdor Measure 100
4.8. Hausdor Dimension of Cantor Sets 104
4.9. Measures on Abstract Spaces 107
4.10. Regular Outer Measures 110
4.11. Outer Measures Generated by Measures 116
Exercises for Chapter 4 121
Chapter 5. Measurable Functions 131
5.1. Elementary Properties of Measurable Functions 131
5.2. Limits of Measurable Functions 142
5.3. Approximation of Measurable Functions 147
Exercises for Chapter 5 150
Chapter 6. Integration 153
6.1. Denitions and Elementary Properties 153
6.2. Limit Theorems 158
6.3. Riemann and Lebesgue IntegrationA Comparison 160
6.4. L
p
Spaces 163
6.5. Signed Measures 171
6.6. The Radon-Nikodym Theorem 176
6.7. The Dual of L
p
182
6.8. Product Measures and Fubinis Theorem 188
6.9. Lebesgue Measure as a Product Measure 197
6.10. Convolution 198
6.11. Distribution Functions 200
6.12. The Marcinkiewicz Interpolation Theorem 202
Exercises for Chapter 6 208
Chapter 7. Dierentiation 219
7.1. Covering Theorems 219
7.2. Lebesgue Points 224
7.3. The Radon-Nikodym Derivative Another View 228
7.4. Functions of Bounded Variation 232
7.5. The Fundamental Theorem of Calculus 237
7.6. Variation of Continuous Functions 242
7.7. Curve Length 247
7.8. The Critical Set of a Function 254
7.9. Approximate Continuity 259
Exercises for Chapter 7 263
CONTENTS 5
Chapter 8. Elements of Functional Analysis 271
8.1. Normed Linear Spaces 271
8.2. Hahn-Banach Theorem 279
8.3. Continuous Linear Mappings 283
8.4. Dual Spaces 287
8.5. Weak and Strong Convergence in L
p
295
8.6. Hilbert Spaces 300
Exercises for Chapter 8 310
Chapter 9. Measures and Linear Functionals 317
9.1. The Daniell Integral 317
9.2. The Riesz Representation Theorem 325
Exercises for Chapter 9 332
Chapter 10. Distributions 335
10.1. The Space D 335
10.2. Basic Properties of Distributions 339
10.3. Dierentiation of Distributions 342
10.4. Essential Variation 347
Exercises for Chapter 10 350
Chapter 11. Functions of Several Variables 353
11.1. Dierentiability 353
11.2. Change of Variable 357
11.3. Sobolev Functions 367
11.4. Approximating Sobolev Functions 373
11.5. Sobolev Imbedding Theorem 377
11.6. Applications 379
11.7. Regularity of Weakly Harmonic Functions 381
Exercises for Chapter 11 384
Index 387
Chapter 12. Solutions to Selected Problems 393
Preface
This text is an essentially self-contained treatment of material that is normally
found in a rst year graduate course in real analysis. Although the presentation is
based on a modern treatment of measure and integration, it has not lost sight of
the fact that the theory of functions of one real variable is the core of the subject.
It is assumed that the student has had a solid course in Advanced Calculus and has
been exposed to rigorous , arguments. Although the books primary purpose is
to serve as a graduate text, we hope that it will also serve as useful reference for
the more experienced mathematician.
The book begins with a chapter on preliminaries and then proceeds with a
chapter on the development of the real number system. This also includes an
informal presentation of cardinal and ordinal numbers. The next chapter provides
the basics of general topological and metric spaces. By the time this chapter has
been concluded, the background of students in a typical course will have been
equalized and they will be prepared to pursue the main thrust of the book.
The text then proceeds to develop measure and integration theory in the next
three chapters. Measure theory is introduced by rst considering outer measures
on an abstract space. The treatment here is abstract, yet short, simple, and basic.
By focusing rst on outer measures, the development underscores in a natural way
the fundamental importance and signicance of -algebras. Lebesgue measure,
Lebesgue-Stieltjes measure, and Hausdor measure are immediately developed as
important, concrete examples of outer measures. Integration theory is presented by
using countably simple functions, that is, functions that assume only a countable
number of values. Conceptually they are no more dicult than simple functions,
but their use leads to a more direct development. Important results such as the
Radon-Nikodym theorem and Fubinis theorem have received treatments that avoid
some of the usual technical diculties.
A chapter on elementary functional analysis is followed by one on the Daniell
integral and the Riesz Representation theorem. This introduces the student to a
completely dierent approach to measure and integration theory. In order for the
student to become more comfortable with this new framework, the linear functional
7
8 PREFACE
approach is further developed by including a short chapter on Schwartz Distribu-
tions. Along with introducing new ideas, this reinforces the students previous
encounter with measures as linear functionals. It also maintains connection with
previous material by casting some old ideas in a new light. For example, BV
functions and absolutely continuous functions are characterized as functions whose
distributional derivatives are measures and functions, respectively.
The introduction of Schwartz distributions invites a treatment of functions of
several variables. Since absolutely continuous functions are so important in real
analysis, it is natural to ask whether they have a counterpart among functions of
several variables. In the last chapter, it is shown that this is the case by developing
the class of functions whose partial derivatives (in the sense of distributions) are
functions, thus providing a natural analog of absolutely continuous functions of a
single variable. The analogy is strengthened by proving that these functions are
absolutely continuous in each variable separately. These functions, called Sobolev
functions, are of fundamental importance to many areas of research today. The
chapter is concluded with a glimpse of both the power and the beauty of Dis-
tribution theory by providing a treatment of the Dirichlet Problem for Laplaces
equation. This presentation is not dicult, but it does call upon many of the top-
ics the student has learned throughout the text, thus providing a tting end to the
book.
We will use the following notation throughout. The symbol denotes the end
of a proof and a := b means a is dened to be b. All theorems, lemmas, corollaries,
denitions, and remarks are numbered as a.b where a denotes the chapter number.
Equation numbers are numbered in a similar way and appear as (a.b). Sections
marked with
are not essential to the main development of the material and may
be omitted.
CHAPTER 1
Preliminaries
1.1. Sets
This is the rst of three sections devoted to basic denitions, notation,
and terminology used throughout this book. We begin with an elemen-
tary and intuitive discussion of sets and deliberately avoid a rigorous
treatment of set theory that would take us too far from our main
purpose.
We shall assume that the notion of set is already known to the reader, at least in
the intuitive sense. Roughly speaking, a set is any identiable collection of objects
called the elements or members of the set. Sets will usually be denoted by capital
Roman letters such as A, B, C, U, V, . . . , and if an object x is an element of A,
we will write x A. When x is not an element of A we write x / A. There are
many ways in which the objects of a set may be identied. One way is to display all
objects in the set. For example, x
1
, x
2
, . . . , x
k
is the set consisting of the elements
x
1
, x
2
, . . . , x
k
. In particular, a, b is the set consisting of the elements a and b.
Note that a, b and b, a are the same set. A set consisting of a single element x
is denoted by x and is called a singleton. Often it is possible to identify a set
by describing properties that are possessed by its elements. That is, if P(x) is a
property possessed by an element x, then we write x : P(x) to describe the set
that consists of all objects x for which the property P(x) is true. Obviously, we
have A = x : x A and x : x ,= x = , the empty set or null set.
The union of sets A and B is the set x : x A or x B and this is written
as A B. Similarly, if / is an arbitrary family of sets, the union of all sets in this
family is
(1.1) x : x A for some A /
and is denoted by
(1.2)
AA
A or as
A : A /.
1
2 1. PRELIMINARIES
Sometimes a family of sets will be dened in terms of an indexing set I and then
we write
(1.3) x : x A
for some I =
I
A
.
If the index set I is the set of positive integers, then we write (1.3) as
(2.1)
i=1
A
i
.
The intersection of sets A and B is dened by x : x A and x B and is
written as A B. Similar to (1.1) and (1.2) we have
x : x A for all A / =
AA
A =
A : A /.
A family / of sets is said to be disjoint if A
1
A
2
= for every pair A
1
and A
2
of distinct members of /.
If every element of the set A is also an element of B, then A is called a subset
of B and this is written as A B or B A. With this terminology, the possibility
that A = B is allowed. The set A is called a proper subset of B if A B and
A ,= B.
The dierence of two sets is
A B = x : x A and x / B
while the symmetric dierence is
AB = (A B) (B A).
In most discussions, a set A will be a subset of some underlying space X and
in this context, we will speak of the complement of A (relative to X) as the set
x : x X and x / A. This set is denoted by
A and this notation will be used
if there is no doubt that complementation is taken with respect to X. In case of
possible ambiguity, we write X A instead of
A. The following identities, known
as de Morgans laws, are very useful and easily veried:
(2.2)
_
I
A
=
I
_
I
A
=
I
.
We shall denote the set of all subsets of X, called the power set of X, by
T(X). Thus,
(2.3) T(X) = A : A X.
1.2. FUNCTIONS 3
The notions of limit superior (lim sup) and lim inferior (lim inf ) are
dened for sets as well as for sequences:
(2.4)
limsup
i
E
i
=
k=1
i=k
E
i
liminf
i
E
i
=
k=1
i=k
E
i
It is easily seen that
(3.1)
limsup
i
E
i
= x : x E
i
for innitely many i ,
liminf
i
E
i
= x : x E
i
for all but nitely many i .
We use the following notation throughout:
= the empty set,
N = the set of positive integers, (not including zero),
Z = the set of integers,
Q = the set of rational numbers,
R = the set of real numbers.
We assume the reader has knowledge of the sets N, Z, and Q, while R will be
carefully constructed in Section 2.1.
1.2. Functions
In this section an informal discussion of relations and functions is given,
a subject that is encountered in several forms in elementary analysis.
In this development, we adopt the notion that a relation or function is
indistinguishable from its graph.
If X and Y are sets, the Cartesian product of X and Y is
(3.2) X Y = all ordered pairs (x, y) : x X, y Y .
The ordered pair (x, y) is thus to be distinguished from (y, x). We will discuss
the Cartesian product of an arbitrary family of sets later in this section.
4 1. PRELIMINARIES
A relation from X to Y is a subset of X Y . If f is a relation, then the
domain and range of f are
domf = X x : (x, y) f for some y Y
rngf = Y y : (x, y) f for some x X .
Frequently symbols such as or are used to designate a relation. In these
cases the notation x y or x y will mean that the element (x, y) is a member of
the relation or , respectively.
A relation f is said to be single-valued if y = z whenever (x, y) and (x, z)
f. A single-valued relation is called a function. The terms mapping, map,
transformation are frequently used interchangeably with function, although the
term function is usually reserved for the case when the range of f is a subset of R.
If f is a mapping and (x, y) f, we let f(x) denote y. We call f(x) the image of x
under f. We will also use the notation x f(x), which indicates that x is mapped
to f(x) by f. If A X, then the image of A under f is
(4.1) f(A) = y : y = f(x), for some x domf A.
Also, the inverse image of B under f is
(4.2) f
1
(B) = x : x domf, f(x) B.
In case the set B consists of a single point y, or in other words B = y, we will
simply write f
1
y instead of the full notation f
1
(y). If A X and f a
mapping with domf X, then the restriction of f to A, denoted by f A, is
dened by f A(x) = f(x) for all x A domf.
If f is a mapping from X to Y and g a mapping from Y to Z, then the
composition of g with f is a mapping from X to Z dened by
(4.3) g f = (x, z) : (x, y) f and (y, z) g for some y Y .
If f is a mapping such that domf = X and rngf Y , then we write f : X Y .
The mapping f is called an injection or is said to be univalent if f(x) ,= f(x
)
whenever x, x
domf with x ,= x
k=1
or
simply x
k
. A subsequence is obtained by discarding some elements of the original
sequence and ordering the elements that remain in the usual way. Formally, we say
that x
k
1
, x
k
2
, x
k
3
, . . . , is a subsequence of x
1
, x
2
, x
3
, . . . , if there is a mapping
g : N N such that for each i N, x
k
i
= x
g(i)
and if g(i) < g(j) whenever i < j.
Our nal topic in this section is the Cartesian product of a family of sets. Let
A be a family of sets X
I
X
for each I. Each mapping x is called a choice mapping for the family A.
Also, we call x() the
th
coordinate of x. This terminology is perhaps easier
to understand if we consider the case where I = 1, 2, . . . , n. As in the preceding
paragraph, it is useful to denote the choice mapping x as a list x(1), x(2), . . . , x(n),
and even more useful if we write x(i) = x
i
. The mapping x is thus identied with
the ordered n-tuple (x
1
, x
2
, . . . , x
n
). Here, the word ordered is crucial because an
n-tuple consisting of the same elements but in a dierent order produces a dierent
mapping x. Consequently, the Cartesian product becomes the set of all ordered
n-tuples:
(5.2)
n
i=1
X
i
= (x
1
, x
2
, . . . , x
n
) : x
i
X
i
, i = 1, 2, . . . , n.
6 1. PRELIMINARIES
In the special case where X
i
= R, i = 1, 2, . . . , n, an element of the Cartesian
product is a mapping that can be identied with an ordered n-tuple of real numbers.
We denote the set of all ordered n-tuples (also referred to as vectors) by
R
n
= (x
1
, x
2
, . . . , x
n
) : x
i
R, i = 1, 2, . . . , n
1.3. SET THEORY 7
R
n
is called Euclidean n-space. The norm of a vector x is dened as
(7.1) [x[ =
_
x
2
1
+x
2
2
+ +x
2
n
;
the distance between two vectors x and y is [x y[. As we mentioned earlier in
this section, the Cartesian product of two sets X
1
and X
2
is denoted by X
1
X
2
.
7.1. Remark. A fundamental issue that we have not addressed is whether the
Cartesian product of an arbitrary family of sets is nonempty. This involves concepts
from set theory and is the subject of the next section.
1.3. Set Theory
The material discussed in the previous two sections is based on tools
found in elementary set theory. However, in more advanced areas of
mathematics this material is not sucient to discuss to even formulate
some of the concepts that are needed. An example of this occurred in
the previous section during the discussion of the Cartesian product of
an arbitrary family of sets. Indeed, the Cartesian product of families
of sets requires the notion of a choice mapping whose existence is not
obvious. Here, we give a brief review of the Axiom of Choice and some
of its logical equivalences.
A fundamental question that arises in the denition of the Cartesian product of an
arbitrary family of sets is the existence of choice mappings. This is an example of
a question that cannot be answered within the context of elementary set theory.
In the beginning of the 20
th
century, Ernst Zermelo formulated an axiom of set
theory called the Axiom of Choice, which asserts that the Cartesian product of an
arbitrary family of nonempty sets exists and is nonempty. The formal statement is
as follows.
7.2. The Axiom of Choice. If X
I
X
is nonempty.
7.3. Proposition. The following statement is equivalent to the Axiom of
Choice: If X
A
is a disjoint family of nonempty sets, then there is a set
S
A
X
such that S X
such
that f() X
AA
f(A).
and
f
1
_
AA
A
_
=
AA
f
1
(A) and f
1
_
AA
A
_
=
AA
f
1
(A).
Give an example that shows the above inclusion cannot be replaced by equality.
1.7 Consider a nonempty set X and its power set T(X). For each x X, let
B
x
= 0, 1 and consider the Cartesian product
xX
B
x
. Exhibit a natural
one-to-one correspondence between T(X) and
xX
B
x
.
1.8 Let X
f
Y be an arbitrary mapping and suppose there is a mapping Y
g
X
such that f g(y) = y for all y Y and that g f(x) = x for all x X. Prove
that f is one-to-one from X onto Y and that g = f
1
.
1.9 Show that A (B C) = (A B) (A C). Also, show that in general
A (B C) ,= (A B) (A C).
Section 1.3
1.10 Use a one-to-one correspondence between Z and N to exhibit a linear ordering
of N that is not a well-ordering.
1.11 Use the natural partial ordering of T(1, 2, 3) to exhibit a partial
1.12 For (a, b), (c, d) NN, dene (a, b) (c, d) if either a < c or a = c and b d.
With this relation, prove that N N is a well-ordered set.
EXERCISES FOR CHAPTER 1 11
1.13 Let P denote the space of all polynomials dened on R. For p
1
, p
2
P, dene
p
1
p
2
if there exists x
0
such that p
1
(x) p
2
(x) for all x x
0
. Is a linear
ordering? Is P well ordered?
1.14 Let C denote the space of all continuous functions on [0, 1]. For f
1
, f
2
C,
dene f
1
f
2
if f
1
(x) f
2
(x) for all x [0, 1]. Is a linear ordering? Is C
well ordered?
1.15 Prove that the following assertion is equivalent to the Axiom of Choice: If A
and B are nonempty sets and f : A B is a surjection (that is, f(A) = B),
then there exists a function g : B A such that g(y) f
1
(y) for each y B.
1.16 Use the following outline to prove that for any two sets A and B, either cardA
cardB or cardB cardA: Let T denote the family of all injections from
subsets of A into B. Since T can be considered as a family of subsets of AB,
it can be partially ordered by inclusion. Thus, we can apply Zorns lemma
to conclude that T has a maximal element, say f. If a A domainf and
b B f(A), then extend f to A a by dening f(a) = b. Then f remains
an injection and thus contradicts maximality. Hence, either domainf = A in
which case cardA cardB or B = range f in which case f
1
is an injection
from B into A, which would imply cardB cardA.
1.17 Complete the details of the following proposition: If cardA cardB and
cardB cardA, then cardA = cardB.
Let f : A B and g : B A be injections. If a A range g, we have
g
1
(a) B. If g
1
(a) rangef, we have f
1
(g
1
(a)) A. Continue this
process as far as possible. There are three possibilities: either the process
continues indenitely, or it terminates with an element of A range g (possibly
with a itself) or it terminates with an element of Brange f. These three cases
determine disjoint sets A
, A
A
and A
B
whose union is A. In a similar manner,
B can be decomposed into B
, B
B
and B
A
. Now f maps A
onto B
and
A
A
onto B
A
and g maps B
B
onto A
B
. If we dene h: A B by h(a) = f(a)
if a A
A
A
and h(a) = g
1
(a) if a A
B
, we nd that h is injective.
CHAPTER 2
Real, Cardinal and Ordinal Numbers
2.1. The Real Numbers
A brief development of the construction of the Real Numbers is given in
terms of equivalence classes of Cauchy sequences of rational numbers.
This construction is based on the assumption that properties of the
rational numbers, including the integers, are known.
In our development of the real number system, we shall assume that properties of
the natural numbers, integers, and rational numbers are known. In order to agree
on what the properties are, we summarize some of the more basic ones. Recall that
the natural numbers are designated as
N: = 1, 2, . . . , k, . . ..
They form a well-ordered set when endowed with the usual ordering. The order-
ing on N satises the following properties:
(i) x x for every x S.
(ii) if x y and y x, then x = y.
(iii) if x y and y z, then x z.
(iv) for all x, y S, either x y or y x.
The four conditions above dene a linear ordering on S, a topic that was in-
troduced in Section 1.3 and will be discussed in greater detail in Section 2.3. The
linear order of N is compatible with the addition and multiplication operations
in N. Furthermore, the following three conditions are satised:
(i) Every nonempty subset of N has a rst element; i.e., if , = S N, there is an
element x S such that x y for any element y S. In particular, the set N
itself has a rst element that is unique, in view of (ii) above, and is denoted
by the symbol 1,
(ii) Every element of N, except the rst, has an immediate predecessor. That is,
if x N and x ,= 1, then there exists y N with the property that y x and
z y whenever z x.
(iii) N has no greatest element; i.e., for every x N, there exists y N such that
x ,= y and x y.
13
14 2. REAL, CARDINAL AND ORDINAL NUMBERS
The reader can easily show that (i) and (iii) imply that each element of N has
an immediate successor, i.e., that for each x N, there exists y N such that
x < y and that if x < z for some z N where y ,= z, then y < z. The immediate
successor of x, y, will be denoted by x
S. Then S = N.
Proof. Suppose S is a proper subset of N that satises the hypotheses of
the theorem. Then N S is a nonempty set and therefore by (i) above, has a
rst element x. Note that x ,= 1 since 1 S. From (ii) we see that x has an
immediate predecessor, y. As y S, we have y
S. Since x = y
, we have x S,
contradicting the choice of x as the rst element of N S.
Also, we have x S since x = y
i=1
0 and that every has
a representative r
i
i=1
with r
i
,= 0 for every i.
In order to dene the sum and product of real numbers, we invoke the corre-
sponding operations on Cauchy sequences of rational numbers. This will require
the next two elementary propositions whose proofs are left to the reader.
18.2. Proposition. If r
i
and s
i
are Cauchy sequences of rational numbers,
then r
i
s
i
and r
i
s
i
are Cauchy sequences. The sequence r
i
/s
i
is also
Cauchy provided s
i
,= 0 for every i and s
i
i=1
,
0.
18.3. Proposition. If r
i
r
i
and s
i
s
i
, then r
i
s
i
r
i
s
and r
i
s
i
r
i
s
i
. Similarly, r
i
/s
i
r
i
/s
i
provided s
i
,
0, and s
i
,= 0
and s
i
,= 0 for every i.
2.1. THE REAL NUMBERS 19
19.1. Definition. If and are represented by r
i
and s
i
respectively, then
is dened by the equivalence class containing r
i
s
i
and by r
i
s
i
.
/ is dened to be the equivalence class containing r
i
/s
i
where s
i
s
i
and
s
i
,= 0 for all i, provided s
i
,
0.
Reference to Propositions 18.2 and 18.3 shows that these operations are well-
dened. That is, if
and
are represented by r
i
and s
i
, where r
i
r
and s
i
s
i
, then + =
maxN
1
, N
2
be an integer with the property that
[r
N
s
N
[ > 2k.
20 2. REAL, CARDINAL AND ORDINAL NUMBERS
Either r
N
> s
N
or s
N
> r
N
. We will show that the rst possibility leads to
conclusion (ii) of the theorem. The proof that the second possibility leads to (iii)
is similar and will be omitted. Assuming now that r
N
> s
N
, we have
r
N
> s
N
+ 2k.
It follows from (14.1) and (17.1) that
[r
N
r
i
[ < k/2 and [s
N
s
i
[ < k/2 for all i N
.
From this and (17.3) we have that
r
i
> r
N
k/2 > s
N
+ 2k k/2 = s
N
+ 3k/2 for i N
.
But s
N
> s
i
k/2 for i N
and consequently,
r
i
> s
i
+k for i N
20.1. Definition. If = r
i
and = s
i
, then we say that < if there
exist rational numbers q
1
and q
2
with q
1
< q
2
and a positive integer N such that
such that r
i
< q
1
< q
2
< s
i
for all i with i N. Note that q
1
and q
2
can be chosen
to be independent of the representative Cauchy sequences of rational numbers that
determine and .
In view of this denition, Theorem 19.2 implies that the real numbers are
comparable, which we state in the following corollary.
20.2. Theorem. Corollary If and are real numbers, then one (and only
one) of the following must hold:
(1) = ,
(2) < ,
(3) > .
Moreover, R is an Archimedean ordered eld.
The compatibility of with the eld structure of R follows from Theorem
19.2. That R is Archimedean follows from Theorem 19.2 and the fact that Q is
Archimedean. Note that the absolute value of a real numbr can thus be dened
analogously to that of a rational number.
20.3. Definition. If
i
i=1
is a sequence in R and R we dene
lim
i
i
=
2.1. THE REAL NUMBERS 21
to mean that given any real number > 0 there is a positive integer N such that
[
i
[ < whenever i N.
22 2. REAL, CARDINAL AND ORDINAL NUMBERS
22.1. Remark. Having shown that R is an Archimedean ordered eld, we now
know that Q has a natural injection into R by way of the constant sequences. That
is, if r Q, then the constant sequence r gives its corresponding equivalence class
in R. Consequently, we shall consider Q to be a subset of R, that is, we do not
distinguish between r and its corresponding equivalence class. Moreover, if
1
and
2
are in R with
1
<
2
, then there is a rational number r such that
1
< r <
2
.
The next proposition provides a connection between Cauchy sequences in Q
with convergent sequences in R.
22.2. Theorem. If = r
i
, then
lim
i
r
i
= .
Proof. Given > 0, we must show the existence of a positive integer N such
that [r
i
[ < whenever i N. Let be represented by the rational sequence
i
. Since > 0, we know from Theorem (19.2), (ii), that there exist a positive
rational number k and an integer N
1
such that
i
> k for all i N
1
. Because
the sequence r
i
is Cauchy, we know there exists a positive integer N
2
such that
[r
i
r
j
[ < k/2 whenever i, j N
2
. Fix an integer i N
2
and let r
i
be determined
by the constant sequence r
i
, r
i
, .... Then the real number r
i
is determined by
the Cauchy sequence r
j
r
i
, that is
r
i
= r
j
r
i
.
If j N
2
, then [r
j
r
i
[ < k/2. Note that the real number [ r
i
[ is determined
by the sequence [r
j
r
i
[. Now, the sequence [r
j
r
i
[ has the property that
[r
j
r
i
[ < k/2 < k <
j
for j max(N
1
, N
2
). Hence, by Denition (20.1),
[ r
i
[ < . The proof is concluded by taking N = max(N
1
, N
2
).
22.3. Theorem. The set of real numbers is complete; that is, every Cauchy
sequence of real numbers converges to a real number.
Proof. Let
i
be a Cauchy sequence of real numbers and let each
i
be
determined by the Cauchy sequence of rational numbers, r
i,k
k=1
. By the previous
proposition,
lim
k
r
i,k
=
i
.
Thus, for each positive integer i, there exists k
i
such that
(22.1) [r
i,k
i
i
[ <
1
i
.
Let s
i
= r
i,k
i
. The sequence s
i
is Cauchy because
2.1. THE REAL NUMBERS 23
[s
i
s
j
[ [s
i
i
[ +[
i
j
[ +[
j
s
j
[
1/i +[
i
j
[ + 1/j.
Indeed, for > 0, there exists a positive integer N > 4/ such that i, j N implies
[
i
j
[ < /2. This, along with (22.1), shows that [s
i
s
j
[ < for i, j N.
Moreover, if is the real number determined by s
i
, then
[
i
[ [ s
i
[ +[s
i
i
[
[ s
i
[ + 1/i.
For > 0, we invoke Theorem 22.2 for the existence of N > 2/ such that the rst
term is less than /2 for i N. For all such i, the second term is also less than
/2.
The completeness of the real numbers leads to another property that is of basic
importance.
23.1. Definition. A number M is called an upper bound for for a set A R
if a M for all a A. An upper bound b for A is called a least upper bound
for A if b is less than all other upper bounds for A. The term supremum of A
is used interchangeably with least upper bound and is written supA. The terms
lower bound, greatest lower bound, and inmum are dened analogously.
23.2. Theorem. Let A R be a nonempty set that is bounded above (below).
Then supA (infA) exists.
Proof. Let b R be any upper bound for A and let a A be an arbitrary
element. Further, using the Archimedean property of R, let M and m be positive
integers such that M > b and m > a, so that we have m < a b < M. For
each positive integer p let
I
p
=
_
k : k an integer and
k
2
p
is an upper bound for A
_
.
Since A is bounded above, it follows that I
p
is not empty. Furthermore, if a A
is an arbitrary element, there is an integer j that is less than a. If k is an integer
such that k 2
p
j, then k is not an element of I
p
, thus showing that I
p
is bounded
24 2. REAL, CARDINAL AND ORDINAL NUMBERS
below. Therefore, since I
p
consists only of integers, it follows that I
p
has a rst
element, call it k
p
. Because
2k
p
2
p+1
=
k
p
2
p
,
the denition of k
p+1
implies that k
p+1
2k
p
. But
2k
p
2
2
p+1
=
k
p
1
2
p
is not an upper bound for A, which implies that k
p+1
,= 2k
p
2. In fact, it follows
that k
p+1
> 2k
p
2. Therefore, either
k
p+1
= 2k
p
or k
p+1
= 2k
p
1.
Dening a
p
=
k
p
2
p
, we have either
a
p+1
=
2k
p
2
p+1
= a
p
or a
p+1
=
2k
p
1
2
p+1
= a
p
1
2
p+1
,
and hence,
a
p+1
a
p
with a
p
a
p+1
1
2
p+1
for each positive integer p. If q > p 1, then
0 a
p
a
q
= (a
p
a
p+1
) + (a
p+1
a
p+2
) + + (a
q1
a
q
)
1
2
p+1
+
1
2
p+2
+ +
1
2
q
=
1
2
p+1
_
1 +
1
2
+ +
1
2
qp1
_
<
1
2
p+1
(2) =
1
2
p
.
Thus, whenever q > p 1, we have [a
p
a
q
[ <
1
2
p
, which implies that a
p
is a
Cauchy sequence. By the completeness of the real numbers, Theorem 22.3, there
is a real number c to which the sequence converges.
We will show that c is the supremum of A. First, observe that c is an upper
bound for A since it is the limit of a decreasing sequence of upper bounds. Secondly,
it must be the least upper bound, for otherwise, there would be an upper bound c
with c
. Then
a
p
1
2
p
c
1
2
p
> c +c
c = c
,
which shows that a
p
1
2
p
is an upper bound for A. But the denition of a
p
implies
that
a
p
1
2
p
=
k
p
1
2
p
,
2.2. CARDINAL NUMBERS 25
a contradiction, since
k
p
1
2
p
is not an upper bound for A.
The existence of inf A in case A is bounded below follows by an analogous
argument.
A linearly ordered eld is said to have the least upper bound property if
each nonempty subset that has an upper bound has a least upper bound (in the
eld). Hence, R has the least upper bound property. It can be shown that every lin-
early ordered eld with the least upper bound property is a complete Archimedean
ordered eld. We will not prove this assertion.
2.2. Cardinal Numbers
There are many ways to determine the size of a set, the most basic
being the enumeration of its elements when the set is nite. When the
set is innite, another means must be employed; the one that we use is
not far from the enumeration concept.
25.1. Definition. Two sets A and B are said to be equivalent if there exists
a bijection f : A B, and then we write A B. In other words, there is a
one-to-one correspondence between A and B. It is not dicult to show that this
notion of equivalence denes an equivalence relation as described in Denition 4.1
and therefore sets are partitioned into equivalence classes. Two sets in the same
equivalence class are said to have the same cardinal number or to be of the same
cardinality. The cardinal number of a set A is denoted by cardA; that is, cardA is
the symbol we attach to the equivalence class containing A. There are some sets so
frequently encountered that we use special symbols for their cardinal numbers. For
example, the cardinal number of the set 1, 2, . . . , n is denoted by n, cardN =
0
,
and cardR = c.
25.2. Definition. A set A is nite if cardA = n for some nonnegative integer
n. A set that is not nite is called innite. Any set equivalent to the positive
integers is said to be denumerable. A set that is either nite or denumerable is
called countable; otherwise it is called uncountable.
One of the rst observations concerning cardinality is that it is possible for two
sets to have the same cardinality even though one is a proper subset of the other.
For example, the formula y = 2x, x [0, 1] denes a bijection between the closed
intervals [0, 1] and [0, 2]. This also can be seen with the help of the gure below.
t
t t t
t t
@
@
@
@
@
@
@@
t t
0
p
0
1
2
26 2. REAL, CARDINAL AND ORDINAL NUMBERS
Another example, utilizing a two-step process, establishes the equivalence be-
tween points x of (1, 1) and y of R. The semicircle with endpoints omitted serves
as an intermediary.
s c c
c c
1 -1
s
x y
.
....................
.....................
.....................
.....................
....................
...................
...................
....................
.....................
.....................
.....................
.................... ....................
.....................
.....................
.....................
....................
...................
...................
....................
.....................
.....................
.....................
....................
.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A bijection could also be explicitly given by y =
2x1
1(2x1)
2
, x (0, 1).
Pursuing other examples, it should be true that (0, 1) [0, 1] although in this
case, exhibiting the bijection is not immediately obvious (but not very dicult, see
Exercise 2.17). Aside from actually exhibiting the bijection, the facts that (0, 1) is
equivalent to a subset of [0, 1] and that [0, 1] is equivalent to a subset of (0, 1) oer
compelling evidence that (0, 1) [0, 1]. The next two results make this rigorous.
26.1. Theorem. If A A
1
A
2
and A A
2
, then A A
1
.
Proof. Let f : A A
2
denote the bijection that determines the equivalence
between A and A
2
. The restriction of f to A
1
, f A
1
, determines a set A
3
(actually,
A
3
= f(A
1
)) such that A
1
A
3
where A
3
A
2
. Now we have sets A
1
A
2
A
3
such that A
1
A
3
. Repeating the argument, there exists a set A
4
, A
4
A
3
such
that A
2
A
4
. Continue this way to obtain a sequence of sets such that
A A
2
A
4
A
2i
and
A
1
A
3
A
5
A
2i+1
.
For notational convenience, we take A
0
= A. Then we have
A
0
= (A
0
A
1
) (A
1
A
2
) (A
2
A
3
) (26.1)
(A
0
A
1
A
2
)
2.2. CARDINAL NUMBERS 27
and
A
1
= (A
1
A
2
) (A
2
A
3
) (A
3
A
4
) (26.2)
(A
1
A
2
A
3
)
By the properties of the sets constructed, we see that
(27.1) (A
0
A
1
) (A
2
A
3
), (A
2
A
3
) (A
4
A
5
), .
In fact, the bijection between (A
0
A
1
) and (A
2
A
3
) is given by f restricted to
A
0
A
1
. Likewise, f restricted to A
2
A
3
provides a bijection onto A
4
A
5
, and
similarly for the remaining sets in the sequence. Moreover, since A
0
A
1
A
2
, we have
(A
0
A
1
A
2
) = (A
1
A
2
A
3
).
The sets A
0
and A
1
are represented by a disjoint union of sets in (26.1) and (26.2).
With the help of (27.1), note that the union of the rst two sets that appear in the
expressions for A and in A
1
are equivalent; that is,
(A
0
A
1
) (A
1
A
2
) (A
1
A
2
) (A
2
A
3
).
Likewise,
(A
2
A
3
) (A
4
A
5
) (A
3
A
4
) (A
5
A
6
),
and similarly for the remaining sets. Thus, it is easy to see that A A
1
.
27.1. Theorem (Schroder-Bernstein). If A A
1
, B B
1
, A B
1
and
B A
1
, then A B.
Proof. Denoting by f the bijection that determines the similarity between A
and B
1
, let B
2
= f(A
1
) to obtain A
1
B
2
with B
2
B
1
. However, by hypothesis,
we have A
1
B and therefore B B
2
. Now invoke Lemma (26.1) to conclude
that B B
1
. But A B
1
by hypothesis and consequently, A B.
It is instructive to recast all of the information in this section in terms of
cardinality. First, we introduce the concept of comparability of cardinal numbers.
27.2. Definition. If and are the cardinal numbers of the sets A and B,
respectively, we say if and only if there exists a set B
1
B such that A B
1
.
In addition, we say that < if there exists no set A
1
A such that A
1
B.
With this terminology, the Schroder-Bernstein Theorem states that
and implies = .
The next denition introduces arithmetic operations on the cardinal numbers.
28 2. REAL, CARDINAL AND ORDINAL NUMBERS
27.3. Definition. Using the notation of Denition (27.2) we dene
+ = card(A B) where A B =
= card(AB)
= cardF
where F is the family of all functions f : B A.
Let us examine the last denition in the special case where = 2. If we take
the corresponding set A as A = 0, 1, it is easy to see that F is equivalent to the
class of all subsets of B. Indeed, the bijection can be dened by
f f
1
1
where f F. This bijection is nothing more than correspondence between subsets
of B and their associated characteristic functions. Thus, 2
is the cardinality of
all subsets of B, which agrees with what we already know in case is nite. Also,
from previous discussions in this section, we have
0
+
0
=
0
,
0
0
=
0
and c +c = c.
In addition, we see that the customary basic arithmetic properties are pre-
served.
28.1. Theorem. If , and are cardinal numbers, then
(i) + ( +) = ( +) +
(ii) () = ()
(iii) + = +
(iv)
(+)
=
(v)
= ()
(vi) (
The proofs of these properties are quite easy. We give an example by proving
(vi):
Proof of (vi). Assume that sets A, B and C respectively represent the car-
dinal numbers , and . Recall that (
0
= c.
Proof. First, to prove the inequality 2
0
c, observe that each real number
r is uniquely associated with the subset Q
r
:= q : q Q, q < r of Q. Thus
mapping r Q
r
is an injection from R into T(Q). Hence,
c = cardR card[T(Q)] = card[T(N)] = 2
0
because Q N.
To prove the opposite inequality, consider the set o of all sequences of the form
x
k
where x
k
is either 0 or 1. Referring to the denition of a sequence (Denition
1.2), it is immediate that the cardinality of o is 2
0
. We will see below (Corollary
31.2) that each number x [0, 1] has a decimal representation of the form
x = .x
1
x
2
. . . , x
i
0, 1.
Of course, such representations do not uniquely represent x. For example,
1
2
= .10000 . . . = .01111 . . . .
Accordingly, the mapping from o into R dened by
f(x
k
) =
_
k=1
x
k
2
k
if x
k
,= 0 for all but nitely many k
k=1
x
k
2
k
+ 1 if x
k
= 0 for innitely many k.
is clearly an injection, thus proving that 2
0
c. Now apply the Schroder-Bernstein
Theorem to obtain our result.
30 2. REAL, CARDINAL AND ORDINAL NUMBERS
The previous result implies, in particular, that 2
0
>
0
; the next result is a
generalization of this.
30.0. Theorem. For any cardinal number , 2
> .
Proof. If A has cardinal number , it follows that 2
and
j
a
= minj : a = a
k
a
,j
.
(Be aware that a could be present more than once in A. If we visualize A as an
innite matrix, then (k
a
, j
a
) represents the position of a that is furthest to the
2.2. CARDINAL NUMBERS 31
northwest in the matrix.) Consequently, A is equivalent to a subset of N N.
Further, observe that there is an injection of N N into N given by
(i, j) 2
i
3
j
.
Indeed, if this were not an injection, we would have
2
ii
3
jj
= 1
for some distinct positive integers i, i
, j, and j
i=1
I
i
= x
0
i=1
I
i
for if not, there would be some positive integer i
0
for which x
0
/ I
i
0
. Therefore,
since I
i
0
is closed, there would be an > 0 such that [x
0
y[ > for each y I
i
0
.
Since the intervals are nested, it would follow that x
0
/ I
i
for all i i
0
and thus
[x
0
x
i
[ > for all i i
0
. This would contradict (31.1) thus establishing (31.2).
We leave it to the reader to verify that x
0
is the only point with this property.
31.2. Corollary. Every real number has a decimal representation relative to
any base.
31.3. Theorem. The real numbers are uncountable.
32 2. REAL, CARDINAL AND ORDINAL NUMBERS
Proof. The proof proceeds by contradiction. Thus, we assume that the real
numbers can be enumerated as a
1
, a
2
, . . . , a
i
, . . .. Let I
1
be a closed interval of
positive length less than 1 such that a
1
/ I
1
. Let I
2
I
1
be a closed interval of
positive length less than 1/2 such that a
2
/ I
2
. Continue in this way to produce a
nested sequence of intervals I
i
of positive length than 1/i with a
i
/ I
i
. Lemma
31.1, we have the existence of a point
x
0
i=1
I
i
.
Observe that x
0
,= a
i
for any i, contradicting the assumption that all real numbers
are among the a
i
s.
2.3. Ordinal Numbers
Here we construct the ordinal numbers and extend the familiar ordering
of the natural numbers. The construction is based on the notion of a
well-ordered set.
32.1. Definition. Suppose W is a well-ordered set with respect to the ordering
. We will use the notation < in its familiar sense; we write x < y to indicate that
both x y and x ,= y. Also, in this case, we will agree to say that x is less than
y and that y is greater than x.
For x W we dene
W(x) = y W : y < x
and refer to W(x) as the initial segment of W determined by x.
The following is the Principle of Transnite Induction.
32.2. Theorem. Let W be a well-ordered set and let S W be dened as
S := x : W(x) S implies x S.
Then S = W.
Proof. If S ,= W then W S is a nonempty subset of W and thus has a
least element x
0
. Then W(x
0
) S, which by hypothesis implies that x
0
S
contradicting the fact that x
0
W S.
When applied to the well-ordered set Z of natural numbers, the hypothesis of
Theorem 32.2 appears to dier in two ways from that of the Principle of Finite
Induction, Theorem 14.1, p.14. First, it is not assumed that 1 S and second,
in order to conclude that x S we need to know that every predecessor of x
is in S and not just its immediate predecessor. The rst dierence is illusory for
suppose a is the least element of W. Then W(a) = S and thus a S. The
2.3. ORDINAL NUMBERS 33
second dierence is more signicant because, unlike the case of N, an element of an
arbitrary well-ordered set may not have an immediate predecessor.
33.0. Definition. A mapping from a well-ordered set V into a well-ordered
set W is order-preserving if (v
1
) (v
2
) whenever v
1
, v
2
V and v
1
v
2
. If, in
addition, is a bijection we will refer to it as an (order-preserving) isomorphism.
Note that, in this case, v
1
< v
2
implies (v
1
) < (v
2
); in other words, an order-
preserving isomorphism is strictly order-preserving.
Note: We have slightly abused the notation by using the same symbol to
indicate the ordering in both V and W above. But this should cause no confusion.
33.1. Lemma. If is an order-preserving injection of a well-ordered set W into
itself, then
w (w)
for each w W.
Proof. Set
S = w W : (w) < w.
If S is not empty, then it has a least element, say a. Thus (a) < a and consequently
((a)) < (a) since is an order-preserving injection; moreover, (a) , S since
a is the least element of S. By the denition of S, this implies (a) ((a)),
which is a contradiction.
33.2. Corollary. If V and W are two well-ordered sets, then there is at most
one isomorphism of V onto W.
Proof. Suppose f and g are isomorphisms of V onto W. Then g
1
f is an
isomorphism of V onto itself and hence v g
1
f(v) for each v V . This implies
that g(v) f(v) for each v V . Since the same argument is valid with the roles
of f and g interchanged, we see that f = g.
33.3. Corollary. If W is a well-ordered set, then W is not isomorphic to an
initial segment of itself
Proof. Suppose a W and W
f
W(a) is an isomorphism. Since w
f(w) for each w W, in particular we have a f(a). Hence f(a) , W(a), a
contradiction.
33.4. Corollary. No two distinct initial segments of a well ordered set W are
isomorphic.
34 2. REAL, CARDINAL AND ORDINAL NUMBERS
Proof. Since one of the initial segments must be an initial segment of the
other, the conclusion follows from the previous result.
34.1. Definition. We dene an ordinal number as an equivalence class of
well-ordered sets with respect to order-preserving isomorphisms. If W is a well-
ordered set, we denote the corresponding ordinal number by ord(W). We dene a
linear ordering on the class of ordinal numbers as follows: if v = ord(V ) and w=
ord(W), then v < w if and only if V is isomorphic to an initial segment of W. The
fact that this denes a linear ordering follows from the next result.
34.2. Theorem. If v and w are ordinal numbers, then precisely one of the
following holds:
(i) v = w
(ii) v < w
(iii) v > w
Proof. Let V and W be well-ordered sets representing v, w respectively and
let T denote the family of all order isomorphisms from an initial segment of V (or
V itself) onto either an initial segment of W (or W itself). Recall that a mapping
from a subset of V into W is a subset of V W. We may assume that V ,= , = W.
If v and w are the least elements of V and W respectively, then (v, w) T and so
T is not empty. Ordering T by inclusion, we see that any linearly ordered subset S
of T has an upper bound; indeed the union of the subsets of V W corresponding
to the elements of S is easily seen to be an order isomorphism and thus an upper
bound for S. Therefore we may employ Zorns lemma to conclude that T has a
maximal element, say h. Since h T, it is an order isomorphism and h V W.
If domainh and range h were initial segments say V
x
and W
y
of V and W, then
h
such that =
2.3. ORDINAL NUMBERS 35
ord(W
W such that W
0
. The element
0
T is the least element of T and thus the least element of
S.
35.1. Corollary. The cardinal numbers are comparable.
Proof. Suppose a is a cardinal number. Then, the set of all ordinals whose
cardinal number is a forms a well-ordered set that has a least element, call it (a).
The ordinal (a) is called the initial ordinal of a. Suppose b is another cardinal
number and let W(a) and W(b) be the well-ordered sets whose ordinal numbers
are (a) and (b), respectively. Either one of W(a) or W(b) is isomorphic to an
initial segment of the other if a and b are not of the same cardinality. Thus, one of
the sets W(a) and W(b) is equivalent to a subset of the other.
35.2. Corollary. Suppose is an ordinal number. Then
= ord( : is an ordinal number and < ).
Proof. Let W be a well-ordered set such that = ord(W). Let < and
let W() be the initial segment of W whose ordinal number is . It is easy to
verify that this establishes an isomorphism between the elements of W and the set
of ordinals less than .
We may view the positive integers N as ordinal numbers in the following way.
Set
1 = ord(1),
2 = ord(1, 2),
3 = ord(1, 2, 3),
.
.
.
= ord(N).
We see that
(35.1) n < for each n N.
If = ord(W) < , then W must be isomorphic to an initial segment of N, i.e.,
= n for some n N. Thus is the rst ordinal number such that (35.1) holds
and is thus the rst innite ordinal.
36 2. REAL, CARDINAL AND ORDINAL NUMBERS
Consider the set of all ordinal numbers that have either nite or denumerable
cardinal numbers and observe that this forms a well-ordered set. We denote the
ordinal number of this set by . It can be shown that is the rst nondenumerable
ordinal number, cf. Exercise 2.20. The cardinal number of is designated by
1
. We have shown that 2
0
>
0
and that 2
0
= c. A fundamental question
that remains open is whether 2
0
=
1
. The assertion that this equality holds is
known as the continuum hypothesis. The work of Godel [?] and Cohen [?], [?]
shows that the continuum hypothesis and its negation are both consistent with the
standard axioms of set theory.
At this point we acknowledge the inadequacy of the intuitive approach that we
have taken to set theory. In the statement of Theorem 34.3 we were careful to refer
to the class of ordinal numbers. This is because the ordinal numbers must not be
a set! Suppose, for a moment, that the ordinal numbers form a set, say O. Then
according to Theorem (34.3), O is a well-ordered set. Let = ord(O). Since O
we must conclude that O is isomorphic to an initial segment of itself, contradicting
Corollary 33.3. For an enlightening discussion of this situation see the book by P.
R. Halmos [H].
Exercises for Chapter 2
Section 2.1
2.1 Use the fact that
N = n : n = 2k for some k N n : n = 2k + 1 for some k N
to prove c c = c. Consequently, card (R
n
) = c for each n N.
2.2 Suppose , and are cardinal numbers. Prove that
+
=
.
2.3 Prove that the set of numbers whose dyadic expansions are not unique is count-
able.
2.4 Prove that the equation x
2
2 = 0 has no solutions in the eld Q.
2.5 Prove: If x
n
n=1
is a bounded, increasing sequence in an Archimedean ordered
eld, then the sequence is Cauchy.
2.6 Prove that each Archimedean ordered eld contains a copy of Q. Moreover,
for each pair r
1
and r
2
of the eld with r
1
< r
2
, there exists a rational number
r such that r
1
< r < r
2
.
2.7 Consider the set r + q
k=0
a
k
x
k
and Q(x) =
m
j=0
b
j
x
j
where the a
k
and b
j
are in Q with a
n
,= 0 and b
m
,= 0. We order
F by saying that
P(x)
Q(x)
is positive if and only if a
n
b
m
is a positive rational
number. Prove that F is an ordered eld which is not Archimedean.
2.9 Consider the set 0, 1 with + and given by the following tables:
+ 0 1
0 0 1
1 1 0
0 1
0 0 0
1 0 1
Prove that 0, 1 is a eld and that there can be no ordering on 0, 1 that
results in a linearly ordered eld.
2.10 Prove: For real numbers a and b,
(a) [a +b[ [a[ +[b[,
(b) [[a[ [b[[ [a b[
(c) [ab[ = [a[ [b[.
Section 2.2
2.11 Show that an arbitrary function R
f
R has at most a countable number of
removable discontinuities: that is, prove that
A := a R : lim
xa
f(x) exists and lim
xa
f(x) ,= f(a)
is at most countable.
2.12 Show that an arbitrary function R
f
R has at most a countable number of
jump discontinuities: that is, let
f
+
(a) := lim
xa
+
f(x)
and
f
(a) := lim
xa
f(x).
Show that the set a R : f
+
(a) ,= f
U : U o T ,
(iii) If o is any nite subcollection of T , then
U : U o T .
The collection T is called a topology for the space X and the elements of T are
called the open sets of X. An open set containing a point x X is called a
neighborhood of x. The interior of an arbitrary set A X is the union of all
open sets contained in A and is denoted by A
o
. Note that A
o
is an open set and
that it is possible for some sets to have an empty interior. A set A X is called
closed if X A: =
A is open. The closure of a set A X, denoted by A, is
A = X x : U A ,= for each open set U containing x
39
40 3. ELEMENTS OF TOPOLOGY
and the boundary of A is A = A A
o
. Note that A A.
These denitions are fundamental and will be used extensively throughout this
text.
40.1. Definition. A point x
0
is called a limit point of a set A X provided
AU contains a point of A dierent from x
0
whenever U is an open set containing
x
0
. The denition does not require x
0
to be an element of A. We will use the
notation A
.
(vii) A B A B whenever A, B X.
(viii) A set A X is closed if and only if A = A.
(ix) A = A A
Proof. Parts (i) and (ii) constitute a restatement of the denition of a topo-
logical space. Parts (iii) and (iv) follow from (i) and (ii) and de Morgans laws,
2.2.
(v) Since A A B, we have A A B. Similarly, B A B, thus proving
A B A B. By contradiction, suppose the converse if not true. Then there
exists x A B with x / A B and therefore there exist open sets U and V
containing x such that U A = = V B. However, since U V is an open set
containing x, it follows that
, = (U V ) (A B) (U A) (V B) = ,
a contradiction.
(vi) This follows from the same reasoning used to establish the rst part of (v).
(vii) This is immediate from denitions.
(viii) If A = A, then
A is open (and thus A is closed) because x , A implies
that there exists an open set U containing x with U A = ; that is, U
A.
Conversely, if A is closed and x
A, then x belongs to some open set U with
U
A. Thus, U A = and therefore x / A. This proves
A (A)
or A A.
But always A A and hence, A = A.
(ix) is left as exercise 3.2.
42 3. ELEMENTS OF TOPOLOGY
41.2. Definition. Let (X, T ) be a topological space and x
i
i=1
a sequence
in X. The sequence is said to converge to x
0
X if for each neighborhood U of
x
0
there is a positive integer N such that x
i
U whenever i N.
It is important to observe that the structure of a topological space is so general
that a sequence could possibly have more than one limit. For example, every
sequence in the space with the indiscrete topology (Example 40.2 (ii)) converges
to every point in X. This cannot happen if an additional restriction is placed
on the topological structure, as in the following denition. (Also note that the
only sequences that converge in the discrete topology are those that are eventually
constant.)
42.1. Definition. A topological space X is said to be a Hausdor space if
for each pair of distinct points x
1
, x
2
X there exist disjoint open sets U
1
and U
2
containing x
1
and x
2
respectively. That is, two distinct points can be separated
by disjoint open sets.
42.2. Definition. Suppose (X, T ) and (Y, o) are topological spaces. A func-
tion f : X Y is said to be continuous at x
0
X if for each neighborhood V
containing f(x
0
) there is a neighborhood U of x
0
such that f(U) V . The function
f is said to be continuous on X if it is continuous at each point x X.
The proof of the next result is given as Exercise 3.4.
42.3. Theorem. Let (X, T ) and (Y, o) be topological spaces. Then for a func-
tion f : X Y , the following statements are equivalent:
(i) f is continuous.
(ii) f
1
(V ) is open in X for each open set V in Y .
(iii) f
1
(K) is closed in X for each closed set K in Y .
42.4. Definition. A collection of open sets, T, in a topological space X is
said to be an open cover of a set A X if
A
UF
U.
The family T is said to admit a subcover, (, of A if ( T and ( is a cover of
A. A subset K X is called compact if each open cover of K possesses a nite
subcover of K. A space X is said to be locally compact if each point of X is
contained in some open set whose closure is compact.
It is easy to give illustrations of sets that are not compact. For example, it is
readily seen that the set A = (0, 1] in R is not compact since the collection of open
3.1. TOPOLOGICAL SPACES 43
intervals of the form (1/i, 2), i = 1, 2, . . ., provides an open cover of A admits no
nite subcover. On the other hand, it is true that [0, 1] is compact, but the proof
is not obvious. The reason for this is that the denition of compactness is usually
not easy to employ directly. Later, in the context of metric spaces (Section 3.3),
we will nd other ways of dealing with compactness.
The following two propositions reveal some basic connections between closed
and compact subsets.
43.1. Proposition. Let (X, T ) be a topological space. If A and K are respec-
tively closed and compact subsets of X with A K, then A is compact.
Proof. If T is an open cover of A, then the elements of T along with X A
form an open cover of K. This open cover has a nite subcover, (, of K since K is
compact. The set XA may possibly be an element of (. If XA is not a member
of (, then ( is a nite subcover of A; if X A is a member of (, then ( with X A
omitted is a nite subcover of A.
43.2. Proposition. A compact subset of a Hausdor space (X, T ) is closed.
Proof. We will show that X K is open where K X is compact. Choose a
xed x
0
X K and for each y K, let V
y
and U
y
denote disjoint neighborhoods
of y and x
0
respectively. The family
T = V
y
: y K
forms an open cover of K. Hence, T possesses a nite subcover, say V
y
i
: i =
1, 2, . . . , N. Since V
y
i
U
y
i
= , i = 1, 2, . . . , N, it follows that
N
i=1
U
y
i
N
i=1
V
y
i
= .
Since K
N
i=1
V
y
i
it follows that
N
i=1
V
y
i
is an open set containing x
0
that does not
intersect K. Thus, X K is an open set, as desired.
The characteristic property of a Hausdor space is that two distinct points can
be separated by disjoint open sets. The next result shows that a stronger property
holds, namely, that a compact set and a point not in this compact set can be
separated by disjoint open sets.
43.3. Proposition. Suppose K is a compact subset of a Hausdor space X
space and assume x
0
, K. Then there exist disjoint open sets U and V containing
x
0
and K respectively.
Proof. This follows immediately from the preceding proof by taking
U =
N
i=1
U
y
i
and V =
N
i=1
V
y
i
.
44 3. ELEMENTS OF TOPOLOGY
43.4. Definition. A family E
F
E
,= .
44.1. Lemma. A topological space X is compact if and only if every family of
closed subsets of X having the nite intersection property has a nonempty intersec-
tion.
Proof. First assume that X is compact and let C
:= XC
is a family, T, of open
sets. If
:= XU
.
If U
U G V
x
: x
U.
and observe that the intersection of all sets in T is empty, for otherwise, we would
be faced with impossibility of some x
0
U G that also belongs to V
x
0
. Lemma
3.2. BASES FOR A TOPOLOGY 45
44.1 (or Remark 44.2) implies there is some nite subfamily of T that has an empty
intersection. That is, there exist points x
1
, x
2
, . . . , x
k
U such that
U G V
x
1
V
x
k
= .
The set
V = G V
x
1
V
x
k
satises the conclusion of our theorem since
K V V G V
x
1
V
x
k
U.
3.2. Bases for a Topology
Often a topology is described in terms of a primitive family of sets,
called a basis. We will give a brief description of this concept.
45.1. Definition. A collection B of open sets in a topological space (X, T )
is called a basis for the topology T if and only if B is a subfamily of T with the
property that for each U T and each x U, there exists B B such that
x B U. A collection B of open sets containing a point x is said to be a basis at
x if for each open set U containing x there is a B B such that x B U. Observe
that a collection B forms a basis for a topology if and only if it contains a basis at
each point x X. For example, the collection of all sets B(x, r), r > 0, x R
n
,
provides a basis for the topology on R
n
as described in (iii) of Example 40.2.
The following is a useful tool for generating a topology on a space X.
45.2. Proposition. Let X be an arbitrary space. A collection B of subsets of
X is a basis for some topology on X if and only if each x X is contained in some
B B and if x B
1
B
2
, then there exists B
3
B such that x B
3
B
1
B
2
.
Proof. It is easy to verify that the conditions specied in the Proposition are
necessary. To show that they are sucient, let T be the collection of sets U with
the property that for each x U, there exists B B such that x B U. It is
easy to verify that T is closed under arbitrary unions. To show that it is closed
under nite intersections, it is sucient to consider the case of two sets. Thus,
suppose x U
1
U
2
, where U
1
and U
2
are elements of T . There exist B
1
, B
2
B
such that x B
1
U
1
and x B
2
U
2
. We are given that there is B
3
B such
that x B
3
B
1
B
2
, thus showing that U
1
U
2
T .
45.3. Definition. A topological space (X, T ) is said to satisfy the rst axiom
of countability if each point x X has a countable basis B
x
at x. It is said to
satisfy the second axiom of countability if there is a countable basis B.
46 3. ELEMENTS OF TOPOLOGY
The second axiom of countability obviously implies the rst axiom of count-
ability. The usual topology on R
n
, for example, satises the second axiom of
countability.
46.0. Definition. A family S of subsets of a topological space (X, T ) is called
a subbase for the topology T if the family consisting of all nite intersections of
members of S forms a base for the topology T .
In view of Proposition 45.2, every nonempty family of subsets of X is the sub-
base for some topology on X. This leads to the concept of the product topology.
46.1. Definition. Given an index set A, consider the Cartesian product
A
X
where each (X
, T
A
X
dened by P
(x) = x
where x
is the
th
coordinate of x, (See (5.1) and its
following remarks.) Consider the collection S of subsets of
A
X
given by
P
1
(V
)
where V
are
continuous.
It is easily seen that a function f from a topological space (Y, T ) into a product
space
A
X
i=1
in a product space
A
X
converges to a point x
0
of the product space if and only if the sequence P
(x
i
)
i=1
converges to P
(x
0
)
for each A. See Exercises 3.8 and 3.9.
3.3. METRIC SPACES 47
3.3. Metric Spaces
Metric spaces are used extensively throughout analysis. The main pur-
pose of this section is to introduce basic denitions.
We already have mentioned two structures placed on sets that deserve the
designation space, namely, vector space and topological space. We now come to
our third structure, that of a metric space.
47.1. Definition. A metric space is an arbitrary set X endowed with a metric
: X X [0, ) that satises the following properties for all x, y and z in X:
(i) (x, y) = 0 if and only if x = y,
(ii) (x, y) = (y, x),
(iii) (x, y) (x, z) +(z, y).
We will write (X, ) to denote the metric space X endowed with a metric . Often
the metric is called the distance function and a reasonable name for property (iii)
is the triangle inequality. If Y X, then the metric space (Y, (Y Y )) is
called the subspace induced by (X, ).
The following are easily seen to be metric spaces.
47.2. Example.
(i) Let X = R
n
and with x = (x
1
, . . . , x
n
), y = (y
1
, . . . , y
n
) R
n
, dene
(x, y) =
_
n
i=1
[x
i
y
i
[
2
_
1/2
.
(ii) Let X = R
n
and with x = (x
1
, . . . , x
n
), y = (y
1
, . . . , y
n
) R
n
, dene
(x, y) = max[x
i
y
i
[ : i = 1, 2, . . . , n.
(iii) The discrete metric on and arbitrary set X is dened as follows: for x, y
X,
(x, y) =
_
_
_
1 if x ,= y
0 if x = y .
(iv) Let X denote the space of all continuous functions dened on [0, 1] and for
f, g C(X), let
(f, g) =
_
1
0
[f(t) g(t)[ dt.
48 3. ELEMENTS OF TOPOLOGY
(v) Let X denote the space of all continuous functions dened on [0, 1] and for
f, g C(X), let
(f, g) = max[f(x) g(x)[ : x [0, 1]
48.1. Definition. If X is a metric space with metric , the open ball centered
at x X with radius r > 0 is dened as
B(x, r) = X y : (x, y) < r.
The closed ball is dened as
B(x, r) := X y : (x, y) r.
In view of the triangle inequality, the family S = B(x, r) : x X, r > 0 forms a
basis for a topology T on X called the topology induced by . The two metrics
in R
n
dened in Examples 47.2, (i) and (ii) induce the same topology on R
n
. Two
metrics on a set X are said to be topologically equivalent if they induce the
same topology on X.
48.2. Definition. Using the notion of convergence given in Denition 41.2,
p.42, the reader can easily verify that the convergence of a sequence x
i
i=1
in a
metric space (X, ) becomes the following:
lim
i
x
i
= x
0
if and only if for each positive number there is a positive integer N such that
(x
i
, x
0
) < whenever i N.
We often write x
i
x
0
for lim
i
x
i
= x
0
.
The notion of a fundamental or a Cauchy sequence is not a topological one
and requires a separate denition:
48.3. Definition. A sequence x
i
i=1
is called Cauchy if for every > 0,
there exists a positive integer N such that (x
i
, x
j
) < whenever i, j N. The
notation for this is
lim
i,j
(x
i
, x
j
) = 0.
Recall the denition of continuity given in Denition 42.2, p.42. In a metric
space, it is convenient to have the following characterization whose proof is left as
an exercise.
48.4. Theorem. If (X, ) and (Y, ) are metric spaces, then a mapping
f : X Y is continuous at x X if for each > 0, there exists > 0 such that
[f(x), f(y)] < whenever (x, y) < .
3.4. MEAGER SETS IN TOPOLOGY 49
48.5. Definition. If X and Y are topological spaces and if f : X Y is a
bijection with the property that both f and f
1
are continuous, then f is called
a homeomorphism and the spaces X and Y are said to be homeomorphic.
A substantial part of topology is devoted to the investigation of those properties
that remain unchanged under the action of a homeomorphism. For example, in
view of Exercise 3.21, it follows that if U X is open, then so is f(U) whenever
f : X Y is a homeomorphism; that is, the property of being open is a topological
invariant. Consequently, so is closedness. But of course, not all properties are
topological invariants. For example, the distance between two points might be
changed under a homeomorphism. A mapping that preserves distances, that is,
one for which
[f(x), f(y)] = (x, y)
for all x, y X is called an isometry,. In particular, it is a homeomorphism, The
spaces X and Y are called isometric if there exists a surjection f : X Y that
is an isometry. In the context of metric space topology, isometric spaces can be
regarded as being identical.
It is easy to verify that a convergent sequence in a metric space is Cauchy, but
the converse need not be true. For example, the metric space, Q, consisting of the
rational numbers endowed with the usual metric on R, possesses Cauchy sequences
that do not converge to elements in Q. If a metric space has the property that
every Cauchy sequence converges (to an element of the space), the space is said
to be complete. Thus, the metric space of rational numbers is not complete,
whereas the real numbers are complete. However, we can apply the technique that
was employed in the construction of the real numbers (see Section 2.1, p.13) to
complete an arbitrary metric space. A precise statement of this is incorporated in
the following theorem, whose proof is left as Exercise 3.30.
49.1. Theorem. If (X, ) is a metric space, there exists a complete metric
space (X
i=1
E
i
.
Let B(x
1
, r
1
) be an open ball with radius r
1
< 1. Since E
1
is not dense in any open
set, it follows that B(x
1
, r
1
) E
1
,= . This is a nonempty open set, and therefore
there exists a ball B(x
2
, r
2
) B(x
1
, r
1
)E
1
with r
2
<
1
2
r
1
. In fact, by also choosing
r
2
smaller than r
1
(x
1
, x
2
), we may assume that B(x
2
, r
2
) B(x
1
, r
1
) E
1
.
Similarly, since E
2
is not dense in any open set, we have that B(x
2
, r
2
) E
2
is a
nonempty open set. As before, we can nd a closed ball with center x
3
and radius
r
3
<
1
2
r
2
<
1
2
2
r
1
B(x
3
, r
3
) B(x
2
, r
2
) E
2
_
B(x
1
, r
1
) E
1
_
E
2
= B(x
1
, r
1
)
2
j=1
E
j
.
Proceeding inductively, we obtain a nested sequence B(x
1
, r
1
) B(x
1
, r
1
)
B(x
2
, r
2
) B(x
2
, r
2
) . . . with r
i
<
1
2
i
r
1
0 such that
(51.1) B(x
i+1
, r
i+1
) B(x
1
, r
1
)
i
j=1
E
j
for each i. Now, for i, j > N, we have x
i
, x
j
B(x
N
, r
N
) and therefore (x
i
, x
j
)
2r
N
. Thus, the sequence x
i
is Cauchy in X. Since X is assumed to be complete,
it follows that x
i
x for some x X. For each positive integer N, x
i
B(x
N
, r
N
)
for i N. Hence, x B(x
N
, r
N
) for each positive integer N. For each positive
integer i it follows from (51.1) that
x B(x
i+1
, r
i+1
) B(x
1
, r
1
)
i
j=1
E
j
.
In particular, for each i N
x ,
i
j=1
E
j
and therefore
x ,
j=1
E
j
= X
a contradiction.
52 3. ELEMENTS OF TOPOLOGY
51.1. Definition. A function f : X Y where (X, ) and (Y, ) are metric
spaces is said to bounded if there exists 0 < M < such that (f(x), f(y)) M
for all x, y X. A family T of functions f : X Y is called uniformly bounded
if (f(x), f(y)) M for all x, y X and for all f T.
An immediate consequence of the Baire Category Theorem is the following re-
sult, which is known as the uniform boundedness principle. We will encounter
this result again in the framework of functional analysis, Theorem 283.1. It states
that if the upper envelope of a family of continuous functions on a complete met-
ric space is nite everywhere, then the upper envelope is bounded above by some
constant on some nonempty open subset. In other words, the family is uniformly
bounded on some open set. Of course, there is no estimate of how large the open
set is, but in some applications just the knowledge that such an open set exists, no
matter how small, is of great importance.
52.1. Theorem. Let T be a family of real-valued continuous functions dened
on a complete metric space X and suppose
(52.1) f
(x): = sup
fF
[f(x)[ <
for each x X. That is, for each x X, there is a constant M
x
such that
f(x) M
x
for all f T.
Then there exist a nonempty open set U X and a constant M such that [f(x)[
M for all x U and all f T.
52.2. Remark. Condition (52.1) states that the family T is bounded at each
point x X; that is, the family is pointwise bounded by M
x
. In applications, a
diculty arises from the possibility that sup
xX
M
x
= . The main thrust of the
theorem is that there exist M > 0 and an open set U such that sup
xU
M
x
M.
Proof. For each positive integer i, let
E
i,f
= x : [f(x)[ i, E
i
=
fF
E
i,f
.
Note that E
i,f
is closed and therefore so is E
i
since f is continuous. From the
hypothesis, it follows that
X =
i=1
E
i
.
Since X is a complete metric space, the Baire Category Theorem implies that there
is some set, say E
M
, that is not nowhere dense. Because E
M
is closed, it must
contain an open set U. Now for each x U, we have [f(x)[ M for all f T,
which is the desired conclusion.
3.5. COMPACTNESS IN METRIC SPACES 53
52.3. Example. Here is a simple example which illustrates this result. Dene
a sequence of functions f
k
: [0, 1] R by
f
k
(x) =
_
_
k
2
x, 0 x 1/k
k
2
x + 2k, 1/k x 2/k
0, 2/k x 1
Thus, f
k
(x) k on [0, 1] and f
j=1
B(x
j
, ),
thus producing a sequence x
i
with (x
i
, x
j
) whenever i ,= j. Clearly, x
i
i=1
A
B
i
is innite for each k = 1, 2, . . . and therefore contains a point x
k
distinct from
x
1
, x
2
, . . . , x
k1
. Observe that 0 < (x
k
, x
l
) < 2/k whenever l k, thus implying
that x
k
is a Cauchy sequence which, by assumption, converges to some x
0
A.
It is easy to verify that x
0
is a limit point of A.
(iv) implies (i): Let U
. To establish our claim, suppose for each positive integer i, there is a ball,
B
i
, of radius 1/i such that
B
i
A ,= ,
B
i
is not contained in any U
. (55.1)
For each positive integer i, select x
i
B
i
A. Since A satises the Bolzano-
Weierstrass property, the sequence x
i
possesses a limit point and therefore it has
a subsequence x
i
j
that converges to some x A. Now x U
. If i
j
is chosen so large that
(x
i
j
, x) <
2
and
1
i
j
<
4
, then for y B
i
j
we have
(y, x) (y, x
i
j
) +(x
i
j
, x) < 2
4
+
2
=
which shows that B
i
j
B(x, ) U
, say U
a
i
and therefore we have
A
k
i=1
B
i
k
i=1
U
i
,
which proves that a nite number of the U
covers A.
56.1. Corollary. A set A R
n
is compact if and only if A is closed and
bounded.
Proof. Clearly, A is bounded if it is compact. Proposition 43.2 shows that it
is also closed.
Conversely, if A is closed, it is complete, (see Exercise 3.15); it thus suces to
show that any bounded subset of R
n
is totally bounded. (Recall that bounded sets
in an arbitrary metric space are not generally totally bounded; see Exercise 3.32.)
Since any bounded set is contained in some cube
Q = [a, a]
n
= x R
n
: max([x
1
[ , . . . , [x
n
[ a),
it is sucient to show that Q is totally bounded. For this purpose, choose > 0
and let k be an integer such that k >
na/. Then Q can be expressed as the
union of k
n
congruent subcubes by dividing the interval [a, a] into k equal pieces.
The side length of each of these subcubes is 2a/k and hence the diameter of each
cube is 2
.
Let P
: X X
and
A is a subbase for the product topology on X.
The proof of Tychonos theorem will utilize the nite intersection property
introduced in Denition 43.4 and Lemma 44.1.
In the following proof, we utilize the Hausdor Maximal Principle, see p. 8.
56.2. Lemma. Let / be a family of subsets of a set Y having the nite inter-
section property and suppose / is maximal with respect to the nite intersection
property, i.e., no family of subsets of Y that properly contains / has the nite
intersection property. Then
3.7. THE SPACE OF CONTINUOUS FUNCTIONS 57
(i) / contains all nite intersections of members of /.
(ii) If S Y and S A ,= for each A /, then S /.
Proof. To prove (i) let B denote the family of all nite intersections of mem-
bers of /. Then / B and B has the nite intersection property. Thus by the
maximality of /, it is clear that / = B.
To prove (ii), suppose S A ,= for each A /. Set ( = /S. Then, since
( has the nite intersection property, the maximality of / implies that ( = /.
We can now prove Tychonos theorem.
57.1. Theorem (Tychonos Product Theorem). If X
: A is a family
of compact topological spaces and X =
A
X
(B) : B B of subsets of X
such that
x
BB
P
(B).
For any A, let U
be an open subset of X
containing x
. Then
B
P
1
(U
) ,=
for each B B. In view of Lemma 56.2 (ii) we see that P
1
(U
) B. Thus by
Lemma 56.2 (i), any nite intersection of sets of this form is a member of B. It
follows that any open subset of X containing x has a nonempty intersection with
each member of B. Since ( B and each member of ( is closed, it follows that
x C for each C (.
3.7. The Space of Continuous Functions
In this section we investigate an important metric space, C(X), the
space of continuous functions on a metric space X. It is shown that this
space is complete. More importantly, necessary and sucient conditions
for the compactness of subsets of C(X) are given.
Recall the discussion of continuity given in Theorems 42.3 and 48.4. Our dis-
cussion will be carried out in the context of functions f : X Y where (X, ) and
58 3. ELEMENTS OF TOPOLOGY
(Y, ) are metric spaces. Continuity of f at x
0
requires that points near x
0
are
mapped into points near f(x
0
). We introduce the concept of oscillation to assist
in making this idea precise.
58.0. Definition. If f : X Y is an arbitrary mapping, then the oscillation
of f on a ball B(x
0
is dened by
osc [f, B(x
0
, r)] = sup[f(x), f(y)] : x, y B(x
0
, r).
Thus, the oscillation of f on a ball B(x
0
, r) is nothing more than the diam-
eter of the set f(B(x
0
, r)) in Y . The diameter of an arbitrary set E is dened
as sup(x, y) : x, y E. It may possibly assume the value +. Note that
osc [f, B(x
0
, r)] is a nondecreasing function of r for each point x
0
.
We leave it to the reader to supply the proof of the following assertion.
58.1. Proposition. A function f : X Y is continuous at x
0
X if and
only if
lim
r0
osc [f, B(x
0
, r)] = 0.
The concept of oscillation is useful in providing information concerning the set
on which an arbitrary function is continuous. For this we need the notions of G
and F
set.
Proof. For each integer i, let
G
i
= X x : inf
r>0
osc [f, B(x, r)] < 1/i.
From the Proposition above, we know that f is continuous at x if and only if
lim
r0
osc [f, B(x, r)] = 0. Therefore, the set of points at which f is continuous is
given by
A =
i=1
G
i
.
To complete the proof we need only show that each G
i
is open. For this, observe
that if x G
i
, then there exists r > 0 such that osc [f, B(x, r)] < 1/i. Now for each
y B(x, r), there exists t > 0 such that B(y, t) B(x, r) and consequently,
osc [f, B(y, t)] osc [f, B(x, r)] < 1/i.
This implies that each point y of B(x, r) is an element of G
i
. That is, B(x, r) G
i
and since x is an arbitrary point of G
i
, it follows that G
i
is open.
3.7. THE SPACE OF CONTINUOUS FUNCTIONS 59
58.3. Theorem. Let f be an arbitrary function dened on [0, 1] and let E :=
x [0, 1] : fis continuous at x. Then E cannot be the set of rational numbers in
[0, 1].
Proof. It suces to show that the rationals in [0, 1] do not constitute a G
set.
Since continuity is such a fundamental notion, it is useful to know those proper-
ties that remain invariant under a continuous transformation. The following result
shows that compactness is a continuous invariant.
59.1. Theorem. Suppose X and Y are topological spaces and f : X Y is a
continuous mapping. If K X is a compact set, then f(K) is a compact subset of
Y .
Proof. Let T be an open cover of f(K); that is, the elements of T are open
sets whose union contains f(K). The continuity of f implies that each f
1
(U) is
an open subset of X for each U T. Moreover, the collection f
1
(U) : U T
provides an open cover of K. Indeed, if x K, then f(x) f(K), and therefore
that f(x) U for some U T. This implies that x f
1
(U). Since K is compact,
T possesses a nite subcover for K, say, f
1
(U
1
), . . . , f
1
(U
k
). From this it
easily follows that the corresponding collection U
1
, . . . , U
k
is an open cover of
f(K), thus proving that f(K) is compact.
59.2. Corollary. Assume that X is a compact topological space and suppose
f : X R is continuous. Then, f attains its maximum and minimum on X; that
is, there are points x
1
, x
2
X such that f(x
1
) f(x) f(x
2
) for all x X.
Proof. From the preceding result and Corollary 56.1, it follows that f(X) is
a closed and bounded subset of R. Consequently, by Theorem 23.2, f(X) has a
least upper bound, say y
0
, that belongs to f(X) since f(X) is closed. Thus there is
a point, x
2
X, such that f(x
2
) = y
0
. Then f(x) f(x
2
) for all x X. Similarly,
there is a point x
1
at which f attains a minimum.
We proceed to examine yet another implication of continuous mappings dened
on compact spaces. The next denition sets the stage.
60 3. ELEMENTS OF TOPOLOGY
59.3. Definition. Suppose X and Y are metric spaces. A mapping f : X Y
is said to be uniformly continuous on X if for each > 0 there exists > 0
such that [f(x), f(y)] < whenever x and y are points in X with (x, y) < .
The important distinction between continuity and uniform continuity is that in
the latter concept, the number depends only on and not on and x as in
continuity. An equivalent formulation of uniform continuity can be stated in terms
of oscillation, which was dened in Denition 57.2, p. 58: for each number r > 0,
let
f
(r): = sup
xX
osc [f, B(x, r)].
The function
f
is called the modulus of continuity of f. It is not dicult to
show that f is uniformly continuous on X provided
lim
r0
f
(r) = 0.
60.1. Theorem. Let f : X Y be a continuous mapping. If X is compact,
then f is uniformly continuous on X.
Proof. Choose > 0. Then the collection
T = f
1
(B(y, )) : y Y
is an open cover of X. Let denote the Lebesgue number of this open cover, (see
Exercise 3.35). Thus, for any x X, we have B(x, /2) is contained in f
1
(B(y, ))
for some y Y . This implies
f
(/2) .
60.2. Definition. For (X, ) a metric space, let
(60.1) d(f, g): = sup([f(x) g(x)[ : x X),
denote the distance between any two bounded, real valued functions dened on
X. This metric is related to the notion of uniform convergence. Indeed, a
sequence of bounded functions f
i
dened on X is said to converge uniformly
to a bounded function f on X provided that d(f
i
, f) 0 as i . We denote by
C(X)
the space of bounded, real-valued, continuous functions on X.
60.3. Theorem. The space C(X) is complete.
Proof. Let f
i
be a Cauchy sequence in C(X). Since
[f
i
(x) f
j
(x)[ d(f
i
, f
j
)
for all x X, it follows for each x X that f
i
(x) is a Cauchy sequence of real
numbers. Therefore, f
i
(x) converges to a number, which depends on x and is
3.7. THE SPACE OF CONTINUOUS FUNCTIONS 61
denoted by f(x). In this way, we dene a function f on X. In order to complete
the proof, we need to show that f is an element of C(X) and that the sequence f
i
f
i
(r) = 0.
That is, for each > 0 and for each i, there exists
i
> 0 such that
(62.1)
f
i
(r) < for r <
i
.
However, since f
i
converges uniformly to f, we claim that there exists > 0 inde-
pendent of f
i
such that (62.1) holds with
i
replaced by . To see this, observe that
since f is uniformly continuous, there exists
. Consequently, if we let
= min
1
, . . . ,
N1
,
it follows from (62.1) and (62.2) that for each positive integer i,
[f
i
(x) f
i
(y)[ <
whenever (x, y) < , thus establishing our claim.
This argument shows that the functions, f
i
, are not only uniformly continuous,
but that the modulus of continuity of each function tends to 0 with r, uniformly
with respect to i. We use this to formulate the next denition,
62.1. Definition. A family, T, of functions dened on X is called equicontin-
uous if for each > 0 there exists > 0 such that for each f T, [f(x) f(y)[ <
whenever (x, y) < . Alternatively, T is equicontinuous if for each f T,
f
(r) < whenever 0 < r < . Sometimes equicontinuous families are dened
pointwise; see Exercise 3.51.
3.7. THE SPACE OF CONTINUOUS FUNCTIONS 63
We are now in a position to give a characterization of compact subsets of C(X)
when X is a compact metric space.
62.2. Theorem (Arzela-Ascoli). Suppose (X, ) is a compact metric space.
Then a set T C(X) is compact if and only if T is closed, bounded, and equicon-
tinuous.
Proof. Suciency: It suces to show that T is sequentially compact. Thus,
it suces to show that an arbitrary sequence f
i
in T has a convergent subse-
quence. Since X is compact it is totally bounded, and therefore separable. Let
D = x
1
, x
2
, . . . denote a countable, dense subset. The boundedness of T implies
that there is a number M
for some M
+M
for all
i and for all x.
Our rst objective is to construct a sequence of functions, g
i
, that is a subse-
quece of f
i
and that converges at each point of D. As a rst step toward this end,
observe that f
i
(x
1
) is a sequence of real numbers that is contained in the compact
interval [M, M], where M := M
+ M
i=1
B
i
(/2). Then selecting
any y
i
B
i
(/2) D it follows that
X =
k
i=1
B(y
i
, ).
Let D
:= y
1
, y
2
, . . . , y
k
and note D
such that [x y
m
[ < . Thus, by equiconti-
nuity, it follows that
[g
i
(x) g
i
(y
m
)[ <
for all positive integers i. Therefore, we have
[g
i
(x) g
j
(x)[ [g
i
(x) g
i
(y
m
)[ +[g
i
(y
m
) g
j
(y
m
)[
+[g
j
(y
m
) g
j
(x)[
< + + = 3,
provided i, j N. This shows that
d(g
i
, g
j
) < 3 for i, j N.
That is, g
i
is a Cauchy sequence in T. Since C(X) is complete (Theorem 60.3)
and T is closed, it follows that g
i
converges to an element g T. Since g
i
is
a subsequence of the original sequence f
i
, we have shown that T is sequentially
compact, thus establishing the suciency argument.
Necessity: Note that T is closed since T is assumed to be compact. Fur-
thermore, the compactness of T implies that T is totally bounded and therefore
bounded. For the proof that T is equicontinuous, note that T being totally bounded
3.7. THE SPACE OF CONTINUOUS FUNCTIONS 65
implies that for each > 0, there exist a nite number of elements in T, say
f
1
, . . . , f
k
, such that any f T is within /3 of f
i
, for some i 1, . . . , k. Conse-
quently, by Exercise 3.42, we have
(65.0)
f
(r)
f
i
(r) + 2d(f, f
i
) <
f
i
(r) + 2/3.
Since X is compact, each f
i
is uniformly continuous on X. Thus, for each i, i =
1, . . . , k, there exists
i
> 0 such that
f
i
(r) < /3 for r <
i
. Now let =
min
1
, . . . ,
k
. By (64.1) it follows that
f
(r) < whenever r < , which proves
that T is equicontinuous.
In many applications, it is not of great interest to know whether T itself is
compact, but whether a given sequence in T has a subsequence that converges
uniformly to an element of C(X), and not necessarily to an element of T. In other
words, the compactness of the closure of T is the critical question. It is easy to see
that if T is equicontinuous, then so is
T. This leads to the following corollary.
65.1. Corollary. Suppose (X, ) is a compact metric space and suppose that
T C(X) is bounded and equicontinuous. Then
T is compact.
Proof. This follows immediately from the previous theorem since
T is both
bounded and equicontinuous.
In particular, this corollary yields the following special result.
65.2. Corollary. Let f
i
be an equicontinuous, uniformly bounded sequence
of functions dened on [0, 1]. Then there is a subsequence that converges uniformly
to a continuous function on [0, 1].
We close this section with a result that will be used frequently throughout the
sequel.
65.3. Theorem. Suppose f is a bounded function on [a, b] that is either non-
decreasing or nonincreasing. Then f has at most a countable number of disconti-
nuities.
Proof. We will give the proof only in case f is nondecreasing, the proof for f
nonincreasing being essentially the same.
Since f is nondecreasing, it follows that the left and right-hand limits exist at
each point and the discontinuities of f occur precisely where these limits are not
equal. Thus, setting
f(x
+
) = lim
yx
+
f(y) and f(x
) = lim
yx
f(y),
66 3. ELEMENTS OF TOPOLOGY
the set D of discontinuities of f in (a, b) is given by
D = (a, b)
_
k=1
x : f(x
+
) f(x
) >
1
k
_
.
For each k the set
x : f(x
+
) f(x
) >
1
k
is a closed set.
3.4 Prove Theorem 42.3.
Section 3.2
3.5 Prove that the product topology on R
n
agrees with the Euclidean topology on
R
n
.
70 3. ELEMENTS OF TOPOLOGY
3.6 Suppose that X
i
, i = 1, 2 satisfy the second axiom of countability. Prove that
the product space X
1
X
2
also satises the second axiom of countability.
3.7 Let (X, T ) be a topological space and let f : X R and g : X R be contin-
uous functions. Dene F : X R R by
F(x) = (f(x), g(x)), x X.
Prove that F is continuous.
3.8 Show that a function f from a topological space (X, T ) into a product space
A
X
i=1
in a product space
A
X
converges to a
point x
0
of the product space if and only if the sequence P
(x
i
)
i=1
converges
to P
(x
0
) for each A.
Section 3.3
3.10 In a metric space, prove that B(x, ) is an open set and that B(x, ) is closed.
Is B(x, ) = B(x, )?
3.11 Suppose X is a complete metric space. Show that if F
1
F
2
. . . are nonempty
closed subsets of X with diameter F
i
0, then there exists x X such that
i=1
F
i
= x
3.12 Suppose (X, ) and (Y, ) are metric spaces with X compact and Y complete.
Let C(X, Y ) denote the space of all continuous mappings f : X Y . Dene a
metric on C(X, Y ) by
d(f, g) = sup(f(x), g(x)) : x X.
Prove that C(X, Y ) is a complete metric space.
3.13 Let (X
1
,
1
) and (X
2
,
2
) be metric spaces and dene metrics on X
1
X
2
as
follows: For x = (x
1
, x
2
), y := (y
1
, y
2
) X
1
X
2
, let
(a) d
1
(x, y) :=
1
(x
1
, y
1
) +
2
(x
2
, y
2
)
(b) d
2
(x, y) :=
_
(
1
(x
1
, y
1
))
2
+ (
2
(x
2
, y
2
))
2
(i) Prove that d
1
and d
2
dene identical topologies.
(ii) Prove that (X
1
X
2
, d
1
) is complete if and only if X
1
and X
2
are complete.
(iii) Prove that (X
1
X
2
, d
1
) is compact if and only if X
1
and X
2
are compact.
3.14 Suppose A is a subset of a metric space X. Prove that a point x
0
, A is a limit
point of A if and only if there is a sequence x
i
in A such that x
i
x
0
.
3.15 Prove that a closed subset of a complete metric space is a complete metric
space.
EXERCISES FOR CHAPTER 3 71
3.16 A mapping f : X X with the property that there exists a number 0 < K < 1
such that (f(x), f(y)) < K(x, y) for all x ,= y is called a contraction. Prove
that a contraction on a complete metric space has a unique xed point.
3.17 Suppose (X, ) is metric space and consider a mapping from X into itself,
f : X X. A point x
0
X is called a xed point for f if f(x
0
) = x
0
. Prove
that if X is compact and f has the property that (f(x), f(y)) < (x, y) for all
x ,= y, then f has a unique xed point.
3.18 As on p.49, a mapping f : X X with the property that (f(x), f(y)) = (x, y)
for all x, y X is called an isometry. If X is compact, prove that an isometry
is a surjection. Is compactness necessary?
3.19 Show that a metric space X is compact if and only if every continuous real-
valued function on X attains a maximum value.
3.20 If X and Y are topological spaces, prove f : X Y is continuous if and only
if f
1
(U) is open whenever U Y is open.
3.21 Suppose f : X Y is surjective and a homeomorphism. Prove that if U X
is open, then so is f(U).
3.22 If X and Y are topological spaces, show that if f : X Y is continuous, then
f(x
i
) f(x
0
) whenever x
i
is a sequence that converges to x
0
. Show that
the converse is true if X and Y are metric spaces.
3.23 Prove that a subset C of a metric space X is closed if and only if every conver-
gent sequence x
i
in C converges to a point in C.
3.24 Prove that C[0, 1] is not a complete space when endowed with the metric given
in (iv) of Example 47.2, p.47.
3.25 Prove that in a topological space (X, T ), if A is dense in B and B is dense in
C, then A is dense in C.
3.26 A metric space is said to be separable if it has a countable dense subset.
(i) Show that R
n
with its usual topology is separable.
(ii) Prove that a metric space is separable if and only if it satises the second
axiom of countability.
(iii) Prove that a subspace of a separable metric space is separable.
(iv) Prove that if a metric space X is separable, then cardX c.
3.27 Let (X, ) be a metric space, Y X, and let (Y, (Y Y )) be the induced
subspace. Prove that if E Y , then the closure of E in the subspace Y is the
same as the closure of E in the space X intersected with Y .
3.28 Prove that the discrete metric on X induces the discrete topology on X.
Section 3.4
72 3. ELEMENTS OF TOPOLOGY
3.29 Prove that a set E in a metric space is nowhere dense if and only if for each
open set U, there is an non empty open set V U such that V E = .
3.30 If (X, ) is a metric space, prove that there exists a complete metric space
(X
i=1
in [0, 1] can be written as [0, 1]
N
. The Tychono
Theorem asserts that the product topology on [0, 1]
N
is compact. Prove that
the function dened by
(x
i
, y
i
) =
i=1
1
2
i
[x
i
y
i
[ for x
i
, y
i
[0, 1]
N
is a metric on [0, 1]
N
and that this metric induces the product topology on [0, 1]
N
.
Prove that every sequence of sequences in [0, 1] has a convergent subsequence in
the metric space ([0, 1]
N
, ). This space is sometimes called the Hilbert Cube.
Section 3.7
3.38 Prove that the set of rational numbers in the real line is not a G
set.
3.39 Prove that the two denitions of uniform continuity given in Denition 59.3
are equivalent.
3.40 Assume that (X, ) is a metric space with the property that each function
f : X R is uniformly continuous.
EXERCISES FOR CHAPTER 3 73
(a) Show that X is a complete metric space
(b) Give an example of a space X with the above property that is not compact.
(c) Prove that if X has only a nite number of isolated points, then X is
compact. See p.54 for the denition of isolated point.
3.41 Prove that a family of functions F is equicontinuous provided there exists a
nondecreasing real valued function such that
lim
r0
(r) = 0
and
f
(r) (r) for all f F.
3.42 Suppose f, g are any two functions dened on a metric space. Prove that
f
(r)
g
(r) + 2d(f, g).
3.43 Prove that a Lipschitzian function is uniformly continuous.
3.44 Prove: If F is a family of Lipschitzian functions from a bounded metric space
X into a metric space Y such that M is a Lipschitz constant for each member
of F and f(x
0
) : f F is a bounded set in Y for some x
0
X, then F is a
uniformly bounded, equicontinuous family.
3.45 Let (X, ) and (Y, ) be metric spaces and let f : X Y be uniformly contin-
uous. Prove that if X is totally bounded, then f(X) is totally bounded.
3.46 Let (X, ) and (Y, ) be metric spaces and let f : X Y be an arbitrary
function. The graph of f is a subset of X Y dened by
G
f
:= (x, y) : y = f(x).
Let d be the metric d
1
on X Y as dened in Exercise 3.13. If Y is compact,
show that f is continuous if and only if G
f
is a closed subset of the metric
space (X Y, d). Can the compactness assumption on Y be dropped?
3.47 Let Y be a dense subset of a metric space (X, ). Let f : Y Z be a uniformly
continuous function where Z is a complete metric space. Show that there is a
uniformly continuous function g : X Z with the property that f = g Y .
Can the assumption of uniform continuity be relaxed to mere continuity?
3.48 Exhibit a bounded function that is continuous on (0, 1) but not uniformly
continuous.
3.49 Let f
i
be a sequence of real-valued, uniformly continuous functions on a
metric space (X, ) with the property that for some M > 0, [f
i
(x) f
j
(x)[ M
for all positive integers i, j and all x X. Suppose also that d(f
i
, f
j
) 0 as
i, j . Prove that there is a uniformly continuous function f on X such
that d(f
i
, f) 0 as i .
74 3. ELEMENTS OF TOPOLOGY
3.50 Let X
f
Y where (X, ) and (Y, ) are metric spaces and where f is contin-
uous. Suppose f has the following property: For each > 0 there is a compact
set K
n=0
[10
n
x]
10
n
where [y] denotes the distance from the greatst integer in y to y.
(b)
f(x) :=
k=0
a
k
cos
b
x
where 1 < ab < b. Weierstrass was the rst to prove the existence of
continuous nowhere dierentiable functions by conceiving of this function
and then proving that it is nowhere dierentiable for certain values of a
and b, [?]. Later, Hardy proved the same result for all a and b, [?].
Section 3.8
3.62 Let f
i
be a decreasing sequence of upper semicontinuous functions dened
on a compact metric space X such that f
i
(x) f(x) where f is lower semi-
continuous. Prove that f
i
f uniformly.
3.63 Show that Theorem 67.2, (iv), remains true for lower semicontinuous functions
that are bounded below. Show also that this assumption is necessary.
CHAPTER 4
Measure Theory
4.1. Outer Measure
An outer measure on an abstract set X is a monotone, countably sub-
additive function dened on all subsets of X. In this section, the notion
of measurable set is introduced, and it is shown that the class of mea-
surable sets forms a -algebra, i.e., measurable sets are closed under the
operations of complementation and countable unions. It is also shown
that an outer measure is countably additive on disjoint measurable sets.
In this section we introduce the concept of outer measure that will underlie and
motivate some of the most important concepts of abstract measure theory. The
length of set in R, the area of a set in R
2
or the volume of a set in R
3
are
notions that can be developed from basic and strongly intuitive geometric principles
provided the sets are well-behaved. If one wished to develop a concept of volume
in R
3
, for example, that would allow the assignment of volume to any set, then
one could hope for a function 1 that assigns to each subset E R
3
a number
1(E) [0, ] having the following properties:
(i) If E
i
k
i=1
is any nite sequence of mutually disjoint sets, then
1
_
k
i=1
E
i
_
=
k
i=i
1(E
i
)
(ii) If two sets E and F are congruent, then 1(E) = 1(F)
(iii) 1(Q) = 1 where Q is the cube of side length 1.
However, these three conditions are inconsistent. In 1924, Banach and Tarski [?]
proved that it is possible to decompose a ball in R
3
into six pieces which can be
reassembled by rigid motions to form two balls, each the same size as the original.
The sets in this decomposition are pathological and require the axiom of choice for
their existence. If condition (i) is changed to require countable additivity rather
than mere nite additivity; that is, require that if E
i
i=1
is any innite sequence
of mutually disjoint sets, then
i=i
1(E
i
) = 1
_
i=1
E
i
_
.
77
78 4. MEASURE THEORY
this too suers from the same inconsistency, and thus we are led to the conclusion
that there is no function 1 satisfying all three conditions above. Later, we will also
see that if we restrict 1 to a large class of subsets of R
3
that omits only the truly
pathological sets, then it is possible to incorporate 1 in a satisfactory theory of
volume.
We will proceed to nd this large class of sets by considering a very general
context and replace countable additivity by countable subadditivity.
78.1. Definition. A function dened for every subset A of an arbitrary set
X is called an outer measure on X provided the following conditions are satised:
(i) () = 0,
(ii) 0 (A) whenever A X,
(iii) (A
1
) (A
2
) whenever A
1
A
2
,
(iv)
_
i=1
A
i
_
i=1
(A
i
) for any countable collection of sets A
i
in X.
Condition (iii) states that is monotone while (iv) states that is countably
subadditive. As we mentioned earlier, suitable additivity properties are necessary
in measure theory; subadditivity, in general, will not suce to produce a useful
theory. We will now introduce the concept of a measurable set and show later
that measurable sets enjoy a wide spectrum of additivity properties.
The term outer measure is derived from the way outer measures are con-
structed in practice. Often one uses a set function that is dened on some family
of primitive sets (such as the family of intervals in R) to approximate an arbitrary
set from the outside to dene its measure. Examples of this procedure will be
given in Sections 4.3 and 4.4. First, consider some elementary examples of outer
measures.
78.2. Examples. (i) In an arbitrary set X, dene (A) = 1 if A is nonempty
and () = 0.
(ii) Let (A) be the number (possibly innite) of points in A.
(iii) Let (A)=
_
_
_
0 if cardA
0
1 if cardA >
0
(iv) If X is a metric space, x > 0. Let (A) be the smallest number of balls of
radius that cover A.
(v) Select a xed x
0
in an arbitrary set X, and let
(A) =
_
_
_
0 if x
0
, A
1 if x
0
A
is called the Dirac measure concentrated at x
0
.
4.1. OUTER MEASURE 79
Notice that the domain of an outer measure is T(X), the collection of all
subsets of X. In general it may happen that the equality (AB) = (A) +(B)
fails when AB = . This property and more generally, property (iv) of Denition
78.1, will require a more restrictive class of subsets of X, called measurable sets,
which we now dene.
79.1. Definition. Let be an outer measure on a set X. A set E X is
called -measurable if
(A) = (A E) +(AE)
for every set A X. In view of property (iv) above, observe that -measurability
only requires
(79.1) (A) (A E) +(AE)
This denition, while not very intuitive, says that a set is -measurable if it
decomposes an arbitrary set, A, into two parts for which is additive. We use
this denition in deference to Caratheodory, who established this property as an
alternative characterization of measurability in the special case of Lebesgue measure
(see Denition 88.3 below). The following characterization of -measurability is
perhaps more intuitively appealing.
79.2. Lemma. A set E X is -measurable if and only if
(P Q) = (P) +(Q)
for any sets P and Q such that P E and Q
E.
Proof. Suciency: Let A X. Then with P : = A E E and Q: =
AE
E we have A = P Q and therefore
(A) = (P Q) = (P) +(Q) = (A E) +(AE).
Necessity: Let P and Q be arbitrary sets such that P E and Q
E. Then,
by the denition of -measurability,
(P Q) = [(P Q) E] +[(P Q)
E]
= (P E) +
_
Q
E
_
= (P) +(Q).
79.3. Remark. Recalling Examples 78.2, one veries that only the empty set
and X are measurable for (i) while all sets are measurable for (ii). Only sets of
measure zero are measurable in (iii).
80 4. MEASURE THEORY
Now that we have an alternate denition of -measurability, we investigate the
properties of -measurable sets. We start with the following theorem which is basic
to the theory. A set function that satises property (iv) below on any sequence of
disjoint sets is said to be countably additive.
80.1. Theorem. Suppose is an outer measure on an arbitrary set X. Then
the following four statements hold.
(i) E is -measurable whenever (E) = 0,
(ii) and X are -measurable.
(iii) E
1
E
2
is -measurable whenever E
1
and E
2
are -measurable,
(iv) If E
i
is a countable collection of disjoint -measurable sets, then
i=1
E
i
is
-measurable and
(
i=1
E
i
) =
i=1
(E
i
).
More generally, if A X is an arbitrary set, then
(A) =
i=1
(A E
i
) +
_
A
S
_
where S =
i=1
E
i
.
Proof. (i) If A X , then (AE) = 0. Thus, (A) (AE) +
_
A
E
_
=
_
A
E
_
(A).
(ii) This follows immediately from Lemma 79.2.
(iii) We will use Lemma 79.2 to establish the -measurability of E
1
E
2
. Thus,
let P E
1
E
2
and Q
_
E
1
E
2
_
=
E
1
E
2
and note that Q =
(Q E
2
) (QE
2
). The -measurability of E
2
implies
(80.1)
(P) +(Q) = (P) +[(Q E
2
) (QE
2
)]
= (P) +(Q E
2
) +(QE
2
).
But P E
1
, QE
2
E
1
, and the -measurability of E
1
imply
(80.2)
(P) +(Q E
2
) +(QE
2
)
= (Q E
2
) +[P (QE
2
)].
Also, Q E
2
E
2
, P (QE
2
)
E
2
, and the -measurability of E
2
imply
(80.3)
(Q E
2
) +[P (QE
2
)]
= [(Q E
2
) (P (QE
2
))]
= (Q P) = (P Q).
4.1. OUTER MEASURE 81
Hence, by (80.1), (80.2) and (80.3) we have
(P) +(Q) = (P Q).
(iv) Let S
k
=
k
i=1
E
i
and let A be an arbitrary subset of X. We proceed by nite
induction and rst note that the result is obviously true for k = 1. For k > 1
assume S
k
is -measurable and that
(81.1) (A)
k
i=1
(A E
i
) +
_
A
S
k
_
,
for any set A. Then
(A) = (A E
k+1
) +(A
E
k+1
) because E
k+1
is -measurable
= (A E
k+1
) +(A
E
k+1
S
k
)
+(A
E
k+1
S
k
) because S
k
is -measurable
= (A E
k+1
) +(A S
k
)
+(A
S
k+1
) because S
k
E
k+1
k+1
i=1
(A E
i
) +(A
S
k+1
) use (81.1) with A replaced by A S
k
.
By the countable subadditivity of , this shows that
(A) (A S
k+1
) +(A
S
k+1
);
this, in turn, implies that S
k+1
is -measurable. Since we now know for any
set A X and for all positive integers k that
(A)
k
i=1
(A E
i
) +(A
S
k
)
and that
S
k
S, we have
(A)
i=1
(A E
i
) +(A
S)
(A S) +(A
S).
(81.2)
Again, the countable subadditivity of was used to establish the last inequal-
ity. This implies that S is -measurable, which establishes the rst part of
(iv). For the second part of (iv), note that the countable subadditivity of
82 4. MEASURE THEORY
yields
(A) (A S) +(A
S)
i=1
(A E
i
) +(A
S).
This, along with (81.2), establishes the last part of (iv).
The preceding result shows that -measurable sets are closed under the set-
theoretic operations of taking complements and countable disjoint unions. Of
course, it would be preferable if they were closed under countable unions, and
not merely countable disjoint unions. The proposition below addresses this issue.
But rst, we will prove a lemma that will be frequently used throughout. It states
that the union of any countable family of sets can be written as the union of a
countable family of disjoint sets.
82.1. Lemma. Let E
i
be a sequence of arbitrary sets. Then there exists a
sequence of disjoint sets A
i
such that each A
i
E
i
and
i=1
E
i
=
i=1
A
i
.
In case each E
i
is -measurable, so is A
i
.
Proof. For each positive integer j, dene S
j
=
j
i=1
E
i
. Note that
i=1
E
i
= S
1
_
k=1
(S
k+1
S
k
)
_
.
Now take A
1
= S
1
and A
i+1
= S
i+1
S
i
for all integers i 1.
In case each E
i
is -measurable, the same is true for each S
j
. Indeed, referring
to Theorem 80.1 (iii), we see that S
2
is -measurable because S
2
= E
2
(E
1
E
2
) is
the disjoint union of -measurable sets. Inductively, we see that S
j
= E
j
(S
j1
E
j
) is the disjoint union of -measurable sets and therefore the sets A
i
are also
-measurable.
82.2. Theorem. If E
i
is a sequence of -measurable sets in X, then
i=1
E
i
and
i=1
E
i
are -measurable.
Proof. From the previous lemma, we have
i=1
E
i
=
i=1
A
i
where each A
i
is a -measurable subset of E
i
and where the sequence A
i
is
disjoint. Thus, it follows immediately from Theorem 80.1 (iv), that
i=1
E
i
is -
measurable.
4.1. OUTER MEASURE 83
To establish the second claim, note that
X (
i=1
E
i
) =
i=1
E
i
.
The right side is -measurable in view of Theorem 80.1 (ii), (iii) and the rst claim.
By appealing again to Theorem 80.1 (iv), this concludes the proof.
Classes of sets that are closed under complementation and countable unions
play an important role in measure theory and are therefore given a special name.
83.1. Definition. A nonempty collection of sets E satisfying the following
two conditions is called a -algebra:
(i) if E , then
E ,
(ii)
i=1
E
i
provided each E
i
.
Note that it easily follows from the denition that a -algebra is closed under
countable intersections and nite dierences. Note also that the entire space and
the empty set are elements of the -algebra since = E
E . The following is
an immediate consequence of Theorem 80.1 and Theorem 82.2.
83.2. Corollary. If is an outer measure on an arbitrary set X, then the
class of -measurable sets forms a -algebra.
Next we state a result that exhibits the basic additivity and continuity prop-
erties of outer measure when restricted to its measurable sets. These properties
follow almost immediately from Theorem 80.1.
83.3. Corollary. Suppose is an outer measure on X and E
i
a countable
collection of -measurable sets.
(i) If E
1
E
2
with (E
1
) < , then
(E
2
E
1
) = (E
2
) (E
1
).
(See Exercise 4.3.)
(ii) (Countable additivity) If E
i
is a disjoint sequence of sets, then
(
i=1
E
i
) =
i=1
(E
i
).
(iii) (Continuity from the left) If E
i
is an increasing sequence of sets, that is, if
E
i
E
i+1
for each i, then
i=1
E
i
_
= ( lim
i
E
i
) = lim
i
(E
i
).
84 4. MEASURE THEORY
(iv) (Continuity from the right) If E
i
is a decreasing sequence of sets, that is, if
E
i
E
i+1
for each i, and if (E
i
0
) < for some i
0
, then
i=1
E
i
_
= ( lim
i
E
i
) = lim
i
(E
i
).
(v) If E
i
is any sequence of -measurable sets, then
(liminf
i
E
i
) liminf
i
(E
i
).
(vi) If
(
i=i
0
E
i
) <
for some positive integer i
0
, then
(limsup
i
E
i
) limsup
i
(E
i
).
Proof. We rst observe that in view of Corollary 83.2, each of the sets that
appears on the left side of (ii) through (vi) is -measurable. Consequently, all sets
encountered in the proof will be -measurable.
(i): Observe that (E
2
) = (E
2
E
1
) + (E
1
) since E
2
E
1
is -measurable
(Theorem 80.1 (iii)).
(ii): This is a restatement of Theorem 80.1 (iv).
(iii): We may assume that (E
i
) < for each i, for otherwise the result
follows from the monotonicity of . Since the sets E
1
, E
2
E
1
, . . . , E
i+1
E
i
, . . .
are -measurable and disjoint, it follows that
lim
i
E
i
=
i=1
E
i
= E
1
_
i=1
(E
i+1
E
i
)
i=1
(E
i+1
E
i
).
Since the sets E
i
and E
i+1
E
i
are disjoint and -measurable, we have (E
i+1
) =
(E
i+1
E
i
) +(E
i
). Therefore, because (E
i
) < for each i, we have from (4.1)
( lim
i
E
i
) = (E
1
) +
i=1
[(E
i+1
) (E
i
)]
= lim
i
(E
i+1
),
which proves (iii).
(iv): By replacing E
i
with E
i
E
i
0
if necessary, we may assume that (E
1
) < .
Since E
i
is decreasing, the sequence E
1
E
i
is increasing and therefore (iii)
4.1. OUTER MEASURE 85
implies
(84.1)
i=1
(E
1
E
i
)
_
= lim
i
(E
1
E
i
)
= (E
1
) lim
i
(E
i
).
It is easy to verify that
i=1
(E
1
E
i
) = E
1
i=1
E
i
,
and therefore, from (iii) of Theorem ??, we have
i=1
(E
1
E
i
)
_
= (E
1
)
_
i=1
E
i
_
which, along with (84.1), yields
(E
1
)
_
i=1
E
i
_
= (E
1
) lim
i
(E
i
).
The fact that (E
1
) < allows us to conclude (iv).
(v): Let A
j
=
i=j
E
i
for j = 1, 2, . . . . Then, A
j
is an increasing sequence of -
measurable sets with the property that lim
j
A
j
= liminf
i
E
i
and therefore,
by (iii),
(liminf
i
E
i
) = lim
j
(A
j
).
But since A
j
E
j
, it follows that
lim
j
(A
j
) liminf
j
(E
j
),
thus establishing (v).
The proof of (vi) is similar to that of (v) and is left as Exercise 4.1.
85.1. Remark. We mentioned earlier that one of our major concerns is to
determine whether there is a rich supply of measurable sets for a given outer mea-
sure . Although we have learned that the class of measurable sets constitutes a
-algebra, this is not sucient to guarantee that the measurable sets exist in great
numbers. For example, suppose that X is an arbitrary set and is dened on X as
(E) = 1 whenever E X is nonempty while () = 0. Then it is easy to verify
that X and are the only -measurable sets. In order to overcome this diculty,
it is necessary to impose an additivity condition on . This will be developed in
the following section.
86 4. MEASURE THEORY
4.2. Caratheodory Outer Measure
In the previous section, we considered an outer measure on an arbi-
trary set X. We now restrict our attention to a metric space X and
impose a further condition (an additivity condition) on the outer mea-
sure. This will allow us to conclude that all closed sets are measurable.
85.2. Definition. An outer measure dened on a metric space (X, ) is
called a Caratheodory outer measure if
(86.1) (A B) = (A) +(B)
whenever A, B are arbitrary subsets of X with d(A, B) > 0. The notation d(A, B)
denotes the distance between the sets A and B and is dened by
d(A, B): = inf(a, b) : a A, b B.
86.1. Theorem. If is a Caratheodory outer measure on a metric space X,
then all closed sets are -measurable.
Proof. We will verify the condition in Denition 79.1 whenever C is a closed
set. Because is subadditive, it suces to show
(86.2) (A) (A C) +(A C)
whenever A X. In order to prove (86.2), consider A X with (A) < and
for each positive integer i, let C
i
= x : d(x, C) 1/i. Note that
d(A C
i
, A C)
1
i
> 0.
Since A (A C
i
) (A C), (85.2) implies
(86.3) (A)
_
(A C
i
) (A C)
_
(A C
i
) +(A C).
Because of this inequality, the proof of (86.2) will be concluded if we can show that
(86.4) lim
i
(A C
i
) = (A C).
For each positive integer i, let
T
i
= A
_
x :
1
i + 1
< d(x, C)
1
i
_
and note that since C is closed, x , C if and only if d(x, C) > 0 and therefore that
(86.5) A C = (A C
j
)
_
i=j
T
i
_
for each positive integer j. This, in turn, implies
(86.6) (A C) (A C
j
) +
i=j
(T
i
).
4.2. CARATH
i=1
(T
i
) < ,
because then
lim
j
_
_
i=j
(T
i
)
_
_
= 0,
which in turn, implies (86.4). To establish this, rst observe that d(T
i
, T
j
) > 0 if
[i j[ 2. Thus, we obtain from (86.1) that for each positive integer m,
m
i=1
(T
2i
) =
_
m
i=1
T
2i
_
(A) < ,
m
i=1
(T
2i1
) =
_
m
i=1
T
2i1
_
(A) < .
This establishes (86.7) and thus concludes the proof.
87.1. Definition. In a topological space, the elements of the smallest -algebra
that contains all open sets are called Borel Sets. The term smallest is taken in
the sense of inclusion and it is left as an exercise (Exercise 4.4) to show that such
a smallest -algebra exists.
The following proposition provides a useful description of the Borel sets.
87.2. Theorem. Suppose B is a family of subsets of a topological space X that
contains all open and all closed subsets of X. Suppose also that B is closed under
countable unions and countable intersections. Then B contains all Borel sets.
Proof. Let
H = B A :
A B.
Observe that H contains all closed sets. Moreover, it is easily seen that H is closed
under complementation and countable unions. Thus, H is a -algebra that contains
the closed sets and therefore contains all Borel sets.
As a direct result of Corollary 83.2 and Theorem 86.1 we have the main result
of this section.
87.3. Theorem. If is a Caratheodory outer measure on a metric space X,
then the Borel sets of X are -measurable.
88 4. MEASURE THEORY
In case X = R
n
, it follows that the cardinality of the Borel sets is at least
as great as that of the closed sets. Since the Borel sets contain all singletons of
R
n
, their cardinality is at least c. We thus have shown that not only do the -
measurable sets have nice additivity properties (they form a -algebra), but in
addition, there is a plentiful supply of them in case is an Caratheodory outer
measure on R
n
. Thus, the diculty that arises from the example in Remark 85.1
is avoided. In the next section we discuss a concrete illustration of such a measure.
4.3. Lebesgue Measure
Lebesgue measure on R
n
is perhaps the most important example of a
Caratheodory outer measure. We will investigate the properties of this
measure and show, among other things, that it agrees with the primitive
notion of volume on sets such as n-dimensional intervals.
For the purpose of dening Lebesgue outer measure on R
n
, we consider closed
n-dimensional intervals
(88.1) I = x : a
i
x
i
b
i
, i = 1, 2, . . . , n
and their volumes
(88.2) v(I) =
n
i=1
(b
i
a
i
).
With I
1
= [a
1
, b
1
], I
2
= [a
2
, b
2
], . . . , I
n
= [a
n
, b
n
], we have
I = I
1
I
2
I
n
.
Notice that n-dimensional intervals have their edges parallel to the coordinate axes
of R
n
. When no confusion arises, we shall simply say interval rather than n-
dimensional interval.
In preparation for the development of Lebesgue measure, we state two elemen-
tary propositions concerning intervals whose proofs will be omitted.
88.1. Theorem. Suppose each edge I
k
= [a
k
, b
k
] of an n-dimensio-nal interval
I is partitioned into
k
subintervals. The products of these intervals produce a
partition of I into : =
1
2
n
subintervals I
i
and
v(I) =
i=1
v(I
i
).
88.2. Theorem. For each interval I and each > 0, there exists an interval J
whose interior contains I and
v(J) < v(I) +.
4.3. LEBESGUE MEASURE 89
88.3. Definition. The Lebesgue outer measure of an arbitrary set E R
n
,
denoted by
(E), is dened by
(E) = inf
_
IS
v(I)
_
where the inmum is taken over all countable collections o of closed intervals I
such that
E
IS
I.
It may be necessary at times to emphasize the dimension of the Euclidean space
in which Lebesgue outer measure is dened. When clarication is needed, we will
write
n
(E) in place of
(E) .
Our next result shows that Lebesgue outer measure is an extension of volume.
89.1. Theorem. For a closed interval I R
n
,
(I) = v(I).
Proof. The inequality
k=1
be a sequence
of closed intervals such that
(89.1) I
k=1
I
k
and
k=1
v(I
k
) <
(I) +
For each k, refer to Proposition 88.2 to obtain an interval J
k
whose interior contains
I
k
and
v(J
k
) v(I
k
) +
2
k
.
We therefore have
k=1
v(J
k
)
k=1
v(I
k
) +.
Let T = interior (J
k
) : k N and observe that T is an open cover of the compact
set I. Let be the Lebesgue number for T (see Exercise 3.35). By Proposition
88.1, there is a partition of I into nitely many subintervals, K
1
, K
2
, . . . , K
m
, each
with diameter less than and having the property
I =
m
i=1
K
i
and v(I) =
m
i=1
v(K
i
).
Each K
i
is contained in the interior of some J
k
, say J
k
i
, although more than one
K
i
may belong to the same J
k
i
. Thus, if N
m
denotes the smallest number of the
90 4. MEASURE THEORY
J
k
i
s that contain the K
i
s, we have N
m
m and
v(I) =
m
i=1
v(K
i
)
N
m
i=1
v(J
k
i
)
k=1
v(J
k
)
k=1
v(I
k
) +.
From this and (89.1) it follows that
v(I)
(I) + 2,
which yields the desired result since is arbitrary.
We will now show that Lebesgue outer measure is a Caratheodory outer measure
as dened in Denition 85.2. Once we have established this result, we then will
be able to apply the important results established in Section 4.2, such as Theorem
87.3, to Lebesgue outer measure.
90.1. Theorem. Lebesgue outer measure,
, dened on R
n
is a Caratheodory
outer measure.
Proof. We rst verify that
i=1
A
i
. We
may as well assume that
(A
i
) < for i = 1, 2, . . . , for otherwise the conclusion
is obvious. Choose > 0. For each i, the denition of Lebesgue outer measure
implies that there exists a countable family of closed intervals, I
(i)
j
j=1
, such that
A
i
j=1
I
(i)
j
(90.1)
and
j=1
v(I
(i)
j
) <
(A
i
) +
2
i
. (90.2)
Now A
i,j=1
I
(i)
j
and therefore
(A)
i,j=1
v(I
(i)
j
) =
i=1
j=1
v(I
(i)
j
)
i=1
(
(A
i
) +
2
i
) =
i=1
(A
i
) +.
Since > 0 is arbitrary, the countable subadditivity of
is established.
Finally, we verify (85.2) of Denition 85.2. Let A and B be arbitrary sets with
d(A, B) > 0. From what has just been proved, we know that
(AB)
(A) +
(B). To prove the opposite inequality, choose > 0 and, from the denition of
4.3. LEBESGUE MEASURE 91
Lebesgue outer measure, select closed intervals I
k
whose union contains A B
such that
k=1
v(I
k
)
(A B) +
By subdividing each interval I
k
into smaller intervals if necessary, we may assume
that the diameter of each I
k
is less than d(A, B). Thus, the family I
k
consists
of two subfamilies, I
k
and I
k
, where the elements of the rst have nonempty
intersections with A while the elements of the second have nonempty intersections
with B. Consequently,
(A) +
(B)
k=1
v(I
k
) +
k=1
v(I
k
) =
k=1
v(I
k
)
(A B) +.
Since > 0 is arbitrary, this shows that
(A) +
(B)
(A B)
which completes the proof.
91.1. Remark. We will henceforth refer to
to the
family of Lebesgue measurable sets. Thus, whenever E is a Lebesgue measurable
set, we have by denition (E) =
, on R
n
.
(i) E R
n
is
-measurable.
92 4. MEASURE THEORY
(ii) For each > 0, there is an open set U E such that
(U E) < .
(iii) There is a G
(U E) = 0.
(iv) For each > 0, there is a closed set F E such that
(E F) < .
(v) There is a F
(E F) = 0.
Proof. (i) (ii). We rst assume that (E) < . For arbitrary > 0, the
denition of Lebesgue outer measure implies the existence of closed
n-dimensional intervals I
k
whose union contains E and such that
k=1
v(I
k
) <
(E) +
2
.
Now, for each k, let I
k
be an open interval containing I
k
such that v(I
k
) < v(I
k
) +
/2
k+1
. Then, dening U =
k=1
I
k
, we have that U is open and from (4.3), that
(U)
k=1
v(I
k
) <
k=1
v(I
k
) +/2 <
(E) +.
Thus, (U) <
(U E) =
(U)
(E) < .
In case (E) = , for each positive integer i, let E
i
denote E B(i) where B(i) is
the open ball of radius i centered at the origin. Then E
i
is a Lebesgue measurable
set of nite measure and thus we may apply the previous step to nd an open
set U
i
E
i
such that (U
i
E
i
) < /2
i
. Let U =
i=1
U
i
and observe that
U E
i=1
(U
i
E
i
). Now use the subadditivity of to conclude that
(U E)
i=1
(U
i
E
i
) <
i=1
2
i
= ,
which establishes the implication (i) (ii).
(ii) (iii). For each positive integer i, let U
i
denote an open set with the
property that U
i
E and
(U
i
E) < 1/i. If we dene U =
i=1
U
i
, then
(U E) =
i=1
(U
i
E)] lim
i
1
i
= 0.
(iii) (i). This is obvious since both U and (U E) are Lebesgue measurable
sets with E = U (U E).
(i) (iv). Assume that E is a measurable set and thus that
E is measurable.
We know that (ii) is equivalent to (i) and thus, given > 0, there is an open set
U
E such that (U
E) < . Note that
E
U = E U = U
E.
4.4. THE CANTOR SET 93
Since
U is closed,
U E and (E
U) < , we see that (iv) holds with F =
U.
The proofs of (iv) (v) and (v) (i) are analogous to those of (ii) (iii)
and (iii) (i), respectively.
93.1. Remark. The above proof is direct and uses only the denition of
Lebesgue measure to establish the various regularity properties. However, another
proof, which is not so long but is perhaps less transparent, proceeds as follows.
Using only the denition of Lebesgue measure, it can be shown that for any set
A R
n
, there is a G
j=1
2
j1
k=1
I
j,k
.
With C denoting the Cantor set, we dene its complement by
I C =
j=1
2
j1
k=1
I
j,k
.
94 4. MEASURE THEORY
Note that C is a closed set and that its Lebesgue measure is 0 since
(I C) =
1
3
+ 2
_
1
3
2
_
+ 2
2
_
1
3
3
_
+
=
k=0
1
3
_
2
3
_
k
= 1.
Note that C is nowhere dense since C does not contain any open set. If it did, its
measure would be positive.
Thus, the Cantor set is small both in the sense of measure and topology. We
now will determine its cardinality.
Every number x [0, 1] has a ternary expansion of the form
x =
i=1
x
i
3
i
where each x
i
is 0, 1, or 2 and we write x = .x
1
x
2
. . . . This expansion is unique
except when
x =
a
3
n
where a and n are positive integers with 0 < a < 3
n
and where 3 does not divide
a. In this case x has the form
x =
n
i=1
x
i
3
i
where x
i
is either 1 or 2. If x
n
= 2, we will use this expression to represent x.
However, if x
n
= 1, we will use the following representation for x:
x =
x
1
3
+
x
2
3
2
+ +
x
n1
3
n1
+
0
3
n
+
i=n+1
2
3
i
.
Thus, with this convention, each number x [0, 1] has a unique ternary expansion.
Let x I and consider its ternary expansion x = .x
1
x
2
. . . , bearing in mind
the convention we have adopted above. Observe that x , I
1,1
if and only if x
1
,= 1.
Also, if x
1
,= 1, then x , I
2,1
I
2,2
if and only if x
2
,= 1. Continuing in this way,
we see that x C if and only if x
i
,= 1 for each positive integer i. Thus, there is a
one-one correspondence between elements of C and all sequences x
i
where each
x
i
is either 0 or 2. The cardinality of the latter is 2
0
which, in view of Theorem
29.1, is c.
The Cantor construction is very general and its variations lead to many in-
teresting constructions. For example, if 0 < < 1, it is possible to produce a
Cantor-type set C
is nowhere
dense and has cardinality c.
4.5. Existence of Nonmeasurable Sets
The existence of a subset of R that is not Lebesgue measurable is inter-
twined with the fundamentals of set theory. Vitali showed that if the
Axiom of Choice is accepted, then it is possible to establish the existence
of nonmeasurable sets. However, in 1970, Solovay proved that using the
usual axioms of set theory, but excluding the Axiom of Choice, it is
impossible to prove the existence of a nonmeasurable set.
95.1. Theorem. There exists a set E R that is not Lebesgue measurable.
Proof. We dene a relation on elements of the real line by saying that x and
y are equivalent (written x y) if x y is a rational number. It is easily veried
that is an equivalence relation as dened in Denition 4.1. Therefore, the real
numbers are decomposed into disjoint equivalence classes. Denote the equivalence
class that contains x by E
x
. Note that if x is rational, then E
x
contains all rational
numbers. Note also that each equivalence class is countable and therefore, since R is
uncountable, there must be an uncountable number of equivalence classes. We now
appeal to the Axiom of Choice, Proposition 7.3 (p.7), to assert the existence of a set
S such that for each equivalence class E, S E consists precisely of one point. If x
and y are arbitrary elements of S, then xy is an irrational number, for otherwise
they would belong to the same equivalence class, contrary to the denition of S.
Thus, the set of dierences, dened by
D: = x y : x, y S,
is a subset of the irrational numbers and therefore cannot contain any interval.
Since the Lebesgue outer measure of any set is invariant under translation and R is
the union of the translates of S by every rational number, it follows that
(S) ,= 0.
Thus, if S were a measurable set, we would have (S) > 0, which contradicts the
following lemma.
95.2. Lemma. If S R is a Lebesgue measurable set of positive and nite
measure, then the set of dierences D := x y : x, y S contains an interval.
Proof. For each > 0, there is an open set U S with (U) < (1 +)(S).
Now U is the union of a countable number of disjoint, open intervals,
U =
k=1
I
k
.
96 4. MEASURE THEORY
Therefore,
S =
k=1
S I
k
and (S) =
k=1
(S I
k
).
Since (U) < (1 + )(S), it follows that (I
k
0
) < (1 + )(S I
k
0
) for some k
0
.
With the choice of =
1
3
, we have
(95.1) (S I
k
0
) >
3
4
(I
k
0
).
Now select any number t with 0 < [t[ <
1
2
(I
k
0
) and consider the translate of the
set SI
k
0
by t, denoted by (SI
k
0
)+t. Then (SI
k
0
)((SI
k
0
)+t) is contained
within an interval of length less than
3
2
(I
k
0
). Using the fact that the Lebesgue
measure of a set remains unchanged under a translation, we conclude that the sets
S I
k
0
and (S I
k
0
) +t must intersect, for otherwise we would contradict (95.1).
This means that for each t with [t[ <
1
2
(I
k
0
), there are points x, y S I
k
0
such
that x y = t. That is, the set
D x y : x, y S I
k
0
f
(E) = inf
_
h
k
F
f
(h
k
)
_
,
where the inmum is taken over all countable collections T of half-open intervals
h
k
of the form (a
k
, b
k
] such that
E
h
k
F
h
k
.
Later in this section, we will show that there is an identication between Lebesgue-
Stieltjes measures and nondecreasing, right-continuous functions. This explains
why we use half-open intervals of the form (a, b]. We could have chosen intervals of
the form [a, b) and then we would show that the corresponding Lebesgue-Stieltjes
measure could be identied with a left-continuous function.
97.1. Remark. Also, observe that the length of each interval (a
k
, b
k
] that
appears in (96.2) can be assumed to be arbitrarily small because
f
((a, b]) = f(b) f(a) =
N
k=1
[f(a
k
) f(a
k1
)] =
N
k=1
f
((a
k1
, a
k
])
whenever a = a
0
< a
1
< < a
N
= b.
97.2. Theorem. If f : R R is a nondecreasing function, then
f
is a Cara-
theodory outer measure on R.
Proof. Referring to Denitions 78.1 and 85.2, we need only show that
f
is
monotone, countably subadditive, and satises property (85.2). Verication of the
remaining properties is elementary.
For the proof of monotonicity, let A
1
A
2
be arbitrary sets in R and assume,
without loss of generality, that
f
(A
2
) < . Choose > 0 and consider a countable
family of half-open intervals h
k
= (a
k
, b
k
] such that
A
2
k=1
h
k
and
k=1
f
(h
k
)
f
(A
2
) +.
Then, since A
1
k=1
h
k
,
f
(A
1
)
k=1
f
(h
k
)
f
(A
2
) +
which establishes the desired inequality since is arbitrary.
The proof of countable subadditivity is virtually identical to the proof of the
corresponding result for Lebesgue measure given in Theorem 90.1 and thus will not
be repeated here.
98 4. MEASURE THEORY
Similarly, the proof of property (86.1) of Denition 85.2, p. 86, runs parallel
to the one given in the proof of Theorem 90.1 for Lebesgue measure. Indeed,
by Remark 97.1, we can may assume that the length of each (a
k
, b
k
] is less than
d(A, B).
Now that we know that
f
is a Caratheodory outer measure, it follows that the
family of
f
-measurable sets contains the Borel sets. As in the case of Lebesgue
measure, we denote by
f
the measure obtained by restricting
f
to its family of
measurable sets.
In the case of Lebesgue measure, we proved that (I) = vol (I) for all intervals
I R
n
. A natural question is whether the analogous property holds for
f
.
98.1. Theorem. If f : R R is nondecreasing and right-continuous, then
f
((a, b]) = f(b) f(a).
Proof. The proof is similar to that of Theorem 89.1 and, as in that situation,
it suces to show
f
((a, b]) f(b) f(a).
Let > 0 and select a cover of (a, b] by a countable family of half-open intervals,
(a
i
, b
i
] such that
(98.1)
i=1
f(b
i
) f(a
i
) <
f
((a, b]) +.
Since f is right-continuous, it follows for each i that
lim
t0
+
f
((a
i
, b
i
+t]) =
f
((a
i
, b
i
]).
Consequently, we may replace each (a
i
, b
i
] with (a
i
, b
i
] where b
i
> b
i
and f(b
i
)
f(a
i
) < f(b
i
) f(a
i
) + /2
i
thus causing no essential change to (98.1), and thus
allowing
(a, b]
i=1
(a
i
, b
i
).
Let a
, b]
i=1
(a
i
, b
i
).
Let be the Lebesgue number of this open cover of the compact set [a
, b] (see
Exercise 3.35). Partition [a
, b] =
m
k=1
[t
k1
, t
k
],
4.6. LEBESGUE-STIELTJES MEASURE 99
where t
0
= a
and t
m
= b and each [t
k1
, t
k
] is contained in some element of the
open cover in (98.2), say (a
i
k
, b
i
k
]. Furthermore, we can relabel the elements of our
partition so that each [t
k1
, t
k
] is contained in precisely one (a
i
k
, b
i
k
]. Then
f(b) f(a
) =
m
k=1
f(t
k
) f(t
k1
)
k=1
f(b
i
k
) f(a
i
k
)
k=1
f(b
i
) f(a
i
)
f
((a, b]) + 2.
Since is arbitrary, we have
f(b) f(a
)
f
((a, b]).
Furthermore, the right continuity of f implies
lim
a
a
+
f(a
) = f(a)
and hence
f(b) f(a)
f
((a, b]),
as desired.
We have just seen that a nondecreasing function f gives rise to a Borel measure
on R. The converse is readily seen to hold, for if is a nite Borel outer measure
on R (that is, (R) < and all Borel sets are -measurable), let
f(x) = ((, x]).
Then, f is nondecreasing, right-continuous (see Exercise 4.35) and
((a, b]) = f(b) f(a) whenever a < b.
(Incidentally, this now shows why half-open intervals are used in the development.)
With f dened this way, note from our previous result, Theorem 98.1, that the
corresponding Lebesgue- Stieltjes measure,
f
, satises
f
((a, b]) = f(b) f(a),
thus proving that and
f
agree on all half-open intervals. Since every open set in
R is a countable union of disjoint half-open intervals, it follows that and
f
agree
on all open sets. Consequently, it seems plausible that these measures should agree
on all Borel sets. In fact, this is true because both and
f
are outer measures
100 4. MEASURE THEORY
with the approximation property described in Theorem 112.1 below. Consequently,
we have the following result.
100.0. Theorem. Suppose is a nite Borel outer measure on R and let
f(x) = ((, x]).
Then the Lebesgue- Stieltjes measure,
f
, agrees with on all Borel sets.
4.7. Hausdor Measure
As a nal illustration of a Caratheodory measure, we introduce s-
dimensional Hausdor (outer) measure in R
n
where s is any nonneg-
ative real number. It will be shown that the only signicant values of
s are those for which 0 s n and that for s in this range, Hausdor
measure provides meaningful measurements of small sets. For example,
sets of Lebesgue measure zero may have positive Hausdor measure.
100.1. Definitions. Hausdor measure is dened in terms of an auxiliary set
function that we introduce rst. Let s, be nonnegative numbers and let A R
n
.
Dene
(100.1) H
s
(A) = inf
_
i=1
(s)2
s
( diam E
i
)
s
: A
i=1
E
i
, diam E
i
<
_
,
where (s) is a normalization constant dened by
(s) =
s
2
(
s
2
+ 1)
,
with
(t) =
_
0
e
x
x
t1
dx, 0 < t < .
It follows from denition that if
1
<
2
, then H
s
1
(E) H
s
2
(E). This allows the
following, which is the denition of s-dimensional Hausdor measure:
H
s
(E) = lim
0
H
s
(E) = sup
>0
H
s
(E).
When s is a positive integer, it turns out that (s) is the Lebesgue measure of the
unit ball in R
s
. This makes it possible to prove that H
s
assigns to elementary sets
the value one would expect. For example, it can be shown that the H
s
measure of
the unit ball in R
s
is (s).
Before deriving the basic properties of H
s
, a few observations are in order.
100.2. Remark.
(i) Hausdor measure could be dened in any metric space since the essential
part of the denition depends on only the notion of diameter of a set.
4.7. HAUSDORFF MEASURE 101
(ii) The sets E
i
in the denition of H
s
j=1
E
i,j
, diam E
i,j
,
j=1
(s)2
s
(diam E
i,j
)
s
< H
s
(A
i
) +
2
i
.
Then, as i and j range through the positive integers, the sets E
i,j
produce a
countable covering of A and therefore,
H
s
i=1
A
i
_
i=1
j=1
(s)2
s
( diam E
i,j
)
s
=
i=1
_
H
s
(A
i
) +
2
i
_
.
Now H
s
(A
i
) H
s
(A
i
) for each i so that
H
s
i=1
A
i
_
i=1
H
s
(A
i
) +.
102 4. MEASURE THEORY
Now taking limits as 0 we obtain
H
s
_
i=1
A
i
_
i=1
H
s
(A
i
),
which establishes countable subadditivity.
Now we will show that condition (85.2) is satised. Choose A, B R
n
with
d(A, B) > 0 and let be any positive number less than d(A, B). Let E
i
be a
covering of AB with diamE
i
. Thus no set E
i
intersects both A and B. Let /
be the collection of those E
i
that intersect A, and B those that intersect B. Then
j=1
(s)2
s
( diam E
i
)
s
EA
(s)2
s
( diam E)
s
+
EB
(s)2
s
( diam E)
s
H
s
(A) +H
s
(B).
Taking the inmum over all such coverings E
i
, we obtain
H
s
(A B) H
s
(A) +H
s
(B)
where is any number less than d(A, B). Finally, taking the limit as 0, we
have
H
s
(A B) H
s
(A) +H
s
(B).
Since we already established (countable) subadditivity of H
s
, property (85.2) is
thus established and the proof is concluded.
Since H
s
is a Caratheodory outer measure, it follows from Theorem 87.3 that
all Borel sets are H
s
-measurable. We next show that H
s
is, in fact, a Borel regular
outer measure in the sense of Denition 110.1 below.
102.1. Theorem. For each A R
n
, there exists a Borel set B A such that
H
s
(B) = H
s
(A).
Proof. From the previous comment, we already know that H
s
is a Borel
outer measure. To show that it is a Borel regular outer measure, recall from (ii)
in Remark 100.2 above that the sets E
i
in the denition of Hausdor measure
can be taken as closed sets. Suppose A R
n
with H
s
(A) < , thus implying
that H
s
i=1
E
i,j
, and
i=1
(s)2
s
( diam E
i,j
)
s
H
s
j
(A) +
j
.
Set
A
j
=
i=1
E
i,j
and B =
j=1
A
j
.
Then B is a Borel set and since A A
j
for each j, we have A B. Furthermore,
since
B
i=1
E
i,j
for each j, we have
H
s
j
(B)
i=1
(s)2
s
( diam E
i,j
)
s
H
s
j
(A) +
j
.
Since
j
0 as j , we obtain H
s
(B) H
s
(A). But A B so that we have
H
s
(A) = H
s
(B).
103.1. Remark. The preceding result can be improved. In fact, there is a G
i=1
(s)2
s
( diam E
i
)
s
H
s
(A) + 1 H
s
(A) + 1.
Then
H
t
(A)
i=1
(t)2
t
( diam E
i
)
t
=
(t)
(s)
2
st
i=1
(s)2
s
( diam E
i
)
s
( diam E
i
)
ts
(t)
(s)
2
st
ts
[H
s
(A) + 1].
Now let 0 to obtain H
t
(A) = 0.
104 4. MEASURE THEORY
103.3. Definition. The Hausdor dimension of an arbitrary set A R
n
is
that number n
A
0 such that
A
:= sups : H
s
(A) > 0 = sups : H
s
(A) =
= inft : H
t
(A) < = inft : H
t
(A) = 0
In other words, the Hausdor dimension
A
is that unique number such that
s <
A
implies H
s
(A) =
t >
A
implies H
t
(A) = 0.
The existence and uniqueness of
A
follows directly from Theorem 103.2.
104.1. Remark. If s =
A
and then any one of of the following three possibil-
ities may occur: H
s
(A) = 0, H
s
(A) = , 0 < H
s
(A) < . On the other hand, if
0 < H
s
(A) < , then
A
= s.
The notion of Hausdor dimension is not very intuitive. Indeed, the Hausdor
dimension of a set need not be an integer. Moreover, if the dimension of a set is an
integer k, the set need not resemble a k-dimensional surface in any usual sense.
See Falconer [FA] or Federer [F] for examples of pathological Cantor-like sets with
integer Hausdor dimension. However, we can at least be reassured by the fact that
the Hausdor dimension of an open set U R
n
is n. To verify this, it is sucient
to assume that U is bounded and to prove that
(104.1) 0 < H
n
(U) < .
Exercise 4.38 deals with the proof of this. Also, it is clear that any countable set
has Hausdor dimension zero; however, there are uncountable sets with dimension
zero, Exercise 4.43.
4.8. Hausdor Dimension of Cantor Sets
In this section the Hausdor dimension of Cantor sets will be determined. Note
that for H
1
dened in R that the constant (s)2
s
in (100.1) equals 1.
104.2. Definition (General Cantor Set). Let 0 < < 1/2 and denote I
0,1
=
[0, 1]. Let I
1,1
and I
1,2
denote the intervals [0, ] and [1 , 1] respectively. They
result by deleting the open middle interval of length 1 2. At the next stage,
delete the open middle 1 2 of each of the intervals I
1,1
and I
1,2
. There remains
2
2
closed intervals each of length
2
. Continuing this process, at the k
th
stage there
are 2
k
closed intervals each of length
k
. Denote these intervals by I
k,1
, . . . , I
k,2
k.
4.8. HAUSDORFF DIMENSION OF CANTOR SETS 105
We dene the generalized Cantor set as
C() =
k=0
2
k
j=1
I
k,j
.
Note that C(1/3) is the Cantor set discussed in Section 4.4.
I
0,1
I
1,1
I
1,2
I
2,1
I
2,2
I
2,3
I
2,4
Since C()
2
k
j=1
I
k,j
for each k, it follows that
H
s
k(C())
2
k
j=1
l(I
k,j
)
s
= 2
k
ks
= (2
s
)
k
,
where
l(I
k,j
) denotes the length of I
k,j
.
If s is chosen so that 2
s
= 1, (if s = log 2/ log(1/)) we have
(105.1) H
s
(C()) = lim
k
H
s
k(C()) 1.
It is important to observe that our choice of s implies that the sum of the s-power
of the lengths of the intervals at any stage is one; that is,
(105.2)
2
k
j=1
l(I
k,j
)
s
= 1.
Next, we show that H
s
(C()) 1/4 which, along with (105.1), implies that the
Hausdor dimension of C() = log(2)/ log(1/). We will establish this by showing
106 4. MEASURE THEORY
that if
C()
i=1
J
i
is an open covering C() by intervals J
i
then
(106.1)
i=1
l(J
i
)
s
1
4
.
Since this is an open cover of the compact set C(), we can employ the Lebesgue
number of this covering to conclude that each interval I
k,j
of the k
th
stage is
contained in some J
i
provided k is suciently large. We will show for any open
interval I and any xed , that
(106.2)
I
,i
I
l(I
,i
)
s
4l(I)
s
.
This will establish (106.1) since
4
i
l(J
i
)
s
I
k,j
J
i
l(I
k,j
)
s
by (106.2)
j=1
l(I
k,j
)
s
= 1 because
2
k
i=1
I
k,i
i=1
J
i
and by (105.2).
To verify (106.2), assume that I contains some interval I
,i
from the
th
stage. and
let k denote the smallest integer for which I contains some interval I
k,j
from the
k
th
stage. Then k . By considering the construction of our set C(), it follows
that no more than 4 intervals from the k
th
stage can intersect I for otherwise, I
would contain some I
k1,i
. Call the intervals I
k,k
m
, m = 1, 2, 3, 4. Thus,
4l(I)
s
m=1
l(I
k,k
m
)
s
=
4
m=1
I
,i
I
k,k
m
l(I
,i
)
s
I
,i
I
l(I
,i
)
s
.
which establishes (106.2). This proves that the dimension of C() = log 2/ log(1/).
106.1. Remark. It can be shown that (106.1) can be improved to read
(106.3)
i
l(J
i
)
s
1
which implies the precise result H
s
(C()) = 1 if
s =
log 2
log(1/)
.
4.9. MEASURES ON ABSTRACT SPACES 107
106.2. Remark. The Cantor sets C() are prototypical examples of sets that
possess self-similar properties. A set is self-similar if it can be decomposed into
parts which are geometrically similar to the whole set. For example, the sets C()
[0, ] and C() [1 , 1] when magnied by the factor 1/ yield a translate of
C(). Self-similaritry is the characteristic property of fractals..
4.9. Measures on Abstract Spaces
Given an arbitrary set X and a -algebra, M, of subsets of X, a non-
negative, countably additive set function dened on M is called a mea-
sure. In this section we extract the properties of outer measures when
restricted to their measurable sets.
Before proceeding, recall the development of the rst three sections of this
chapter. We began with the concept of an outer measure on an arbitrary set X
and proved that the family of measurable sets forms a -algebra. Furthermore, we
showed that the outer measure is countably additive on measurable sets. In order
to ensure that there are situations in which the family of measurable sets is large,
we investigated Caratheodory outer measures on a metric space and established
that their measurable sets always contain the Borel sets. We then introduced
Lebesgue measure as the primary example of a Caratheodory outer measure. In
this development, we begin to see that countable additivity plays a central and
indispensable role and thus, we now call upon a common practice in mathematics
of placing a crucial concept in an abstract setting in order to isolate it from the
clutter and distractions of extraneous ideas. We begin with a Denition
107.1. Definition. Let X be a set and / a -algebra of subsets of X. A
measure on / is a function : / [0, ] satisfying the properties
(i) () = 0,
(ii) if E
i
is a sequence of disjoint sets in /, then
i=1
E
i
_
=
i=1
(E
i
).
Thus, a measure is a countably additive set function dened on /. Sometimes
the notion of nite additivity is useful. It states that
(ii)
If E
1
, E
2
, . . . , E
k
is any nite family of disjoint sets in /, then
_
k
i=1
E
i
_
=
k
i=1
(E
i
).
If satises (i) and (ii
i=1
E
i
where (E
i
) < for each i. We emphasize that the notation (E) implies that
E is an element of /, since is dened only on /. Thus, when we write (E
i
)
as in the denition above, it should be understood that the sets E
i
are necessarily
elements of /.
108.1. Examples. Here are some examples of measures.
(i) (R
n
, /, ) where is Lebesgue measure and / is the family of Lebesgue
measurable sets.
(ii) (X, /, ) where is an outer measure on an abstract set X and / is the
family of -measurable sets.
(iii) (X, /,
x
0
) where X is an arbitrary set and
x
0
is an outer measure dened
by
x
0
(E) =
_
_
_
1 if x
0
E
0 if x
0
, E.
The point x
0
X is selected arbitrarily. It can easily be shown that all
subsets of X are
x
0
-measurable and therefore, / is taken as the family of
all subsets of X.
(iv) (R, /, ) where / is the family of all Lebesgue measurable sets, x
0
R and
is dened by
(E) = (E x
0
) +
x
0
(E)
whenever E /.
(v) (X, /, ) where / is the family of all subsets of an arbitrary space X and
where (E) is dened as the number (possibly innite) of points in E /.
The proof of Corollary 83.3 (p.83) utilized only those properties of an outer
measure that an abstract measure possesses and therefore, most of the following do
not require a proof.
108.2. Theorem. Let (X, /, ) be a measure space and suppose E
i
is a
sequence of sets in /.
(i) (Monotonicity) If E
1
E
2
, then (E
1
) (E
2
).
(ii) (Subtractivity) If E
1
E
2
and (E
1
) < , then (E
2
E
1
) = (E
2
)(E
1
).
4.9. MEASURES ON ABSTRACT SPACES 109
(iii) (Countable Subadditivity)
i=1
E
i
_
i=1
(E
i
).
(iv) (Continuity from the left) If E
i
is an increasing sequence of sets, that is, if
E
i
E
i+1
for each i, then
i=1
E
i
_
= ( lim
i
E
i
) = lim
i
(E
i
).
(v) (Continuity from the right) If E
i
is a decreasing sequence of sets, that is, if
E
i
E
i+1
for each i, and if (E
i
0
) < for some i
0
, then
i=1
E
i
_
= ( lim
i
E
i
) = lim
i
(E
i
).
(vi)
(liminf
i
E
i
) liminf
i
(E
i
).
(vii) If
(
i=i
0
E
i
) <
for some positive integer i
0
, then
(limsup
i
E
i
) limsup
i
(E
i
).
Proof. Only (i) and (iii) have not been established in Corollary 83.3, p. 83.
For (i), observe that if E
1
E
2
, then (E
2
) = (E
1
) +(E
2
E
1
) (E
1
).
(iii) Refer to Lemma 82.1, p. 82, to obtain a sequence of disjoint measurable
sets A
i
such that A
i
E
i
and
i=1
E
i
=
i=1
A
i
.
Then,
i=1
E
i
_
=
_
i=1
A
i
_
=
i=1
(A
i
)
i=1
(E
i
).
One property that is characteristic to an outer measure but is not enjoyed by
abstract measures in general is the following: if (E) = 0, then E is -measurable
and consequently, so is every subset of E. A measure with the property that
all subsets of sets of -measure zero are measurable, is said to be complete and
(X, /, ) is called a complete measure space. Not all measures are complete,
110 4. MEASURE THEORY
but this is not a crucial defect since every measure can easily be completed by
enlarging its domain of denition to include all subsets of sets of measure zero.
110.0. Theorem. Suppose (X, /, ) is a measure space. Dene / = AN :
A /, N B for some B / such that (B) = 0 and dene on / by
(A N) = (A). Then, / is a -algebra, is a complete measure on /, and
(X, /, ) is a complete measure space. Moreover, is the only complete measure
on / that is an extension of .
Proof. It is easy to verify that / is closed under countable unions since this
is true for sets of measure zero. To show that / is closed under complementation,
note that with sets A, N, and B as in the denition of /, it may be assumed that
AN = because AN = A(N A) and N A is a subset of a measurable set
of measure zero, namely, B A. It can be readily veried that
A N = (A B) ((
B N) (A B))
and therefore
(A N)
= (A B)
((
B N) (A B))
= (A B)
((B
N) (A B)
)
= (A B)
((B N) A B).
Since (A B)
i=1
A
i
_
= lim
i
(A
i
),
(ii) If AB is -measurable, (A) < , (B) < and (AB) = (A)+(B),
then both A and B are -measurable.
Proof. (i): Choose -measurable sets C
i
A
i
with (C
i
) = (A
i
). The
-measurable sets
B
i
:=
j=i
C
j
form an ascending sequence that satisfy the conditions A
i
B
i
C
i
as well as
i=1
A
i
_
_
i=1
B
i
_
= lim
i
(B
i
) lim
i
(C
i
) = lim
i
(A
i
).
Hence, it follows that
i=1
A
i
_
lim
i
(A
i
).
The opposite inequality is immediate since
i=1
A
i
_
(A
k
) for each k N.
(ii): Choose a -measurable set C
A such that (C
i=1
V
i
where each V
i
is an open set with (V
i
) < . Then, for each > 0, there is an
open set W B such that
(W B) < .
Proof. For the proof of the rst part, select a Borel set B with (B) <
and dene a set function by
(112.1) (A) = (A B)
whenever A X. It is easy to verify that is an outer measure on X whose
measurable sets include all -measurable sets (see Exercise 4.6) and thus all open
sets. The outer measure is introduced merely to allow us to work with an outer
measure for which (X) < .
Let T be the family of all -measurable sets A X with the following property:
For each > 0, there is a closed set F A such that (A F) < . The rst
part of the Theorem will be established by proving that T contains all Borel sets.
Obviously, T contains all closed sets. It also contains all open sets. Indeed, if U is
an open set, then the closed sets
F
i
= x : d(x,
U) 1/i
4.10. REGULAR OUTER MEASURES 113
have the property that F
1
F
2
. . . and
U =
i=1
F
i
and therefore that
i=1
(U F
i
) = .
Therefore, since (X) < , Corollary 83.3 (iv) yields
lim
i
(U F
i
) = 0,
which shows that T contains all open sets U.
Since T contains all open and closed sets, according to Proposition 87.2, p.87,
we need only show that T is closed under countable unions and countable inter-
sections to conclude that it also contains all Borel sets. For this purpose, suppose
A
i
is a sequence of sets in T and for given > 0, choose closed sets C
i
A
i
with
(A
i
C
i
) < /2
i
. Since
i=1
A
i
i=1
C
i
i=1
(A
i
C
i
)
and
i=1
A
i
i=1
C
i
i=1
(A
i
C
i
)
it follows that
i=1
A
i
i=1
C
i
i=1
(A
i
C
i
)
<
i=1
2
i
= (113.1)
and
lim
k
i=1
A
i
k
i=1
C
i
=
_
i=1
A
i
i=1
C
i
i=1
(A
i
C
i
)
< (113.2)
Consequently, there exists a positive integer k such that
(113.3)
_
i=1
A
i
k
i=1
C
i
< .
We have used the fact that
i=1
A
i
and
i=1
A
i
are -measurable, and in (113.2), we
again have used (iv) of Corollary 83.3. Since the sets
i=1
C
i
and
k
i=1
C
i
are closed
subsets of
i=1
A
i
and
i=1
A
i
respectively, it follows from (113.1) and (113.3) that
T is closed under the operations of countable unions and intersections.
To prove the second part of the Theorem, consider the Borel sets V
i
B and
use the rst part to nd closed sets C
i
(V
i
B) such that
[(V
i
C
i
) B] = [(V
i
B) C
i
] <
2
i
.
114 4. MEASURE THEORY
For the desired set W in the statement of the Theorem, let W =
i=1
(V
i
C
i
) and
observe that
(W B)
i=1
[(V
i
C
i
) B] <
i=1
2
i
= .
Moreover, since B V
i
V
i
C
i
, we have
B =
i=1
(B V
i
)
i=1
(V
i
C
i
) = W.
114.1. Corollary. If two nite Borel measures agree on all open (or closed)
sets, then they agree on all Borel sets. In particular, in R, if they agree on all
half-open inervals, then they agree on all Borel sets.
114.2. Remark. The preceding theorem applies directly to any Caratheodory
outer measure since its measurable sets contain the Borel sets. In particular, the
result applies to both Lebesgue- Stieltjes measure
f
and Lebesgue measure and
thereby furnishes an alternate proof to Theorem 91.2, p. 91.
114.3. Remark. In order to underscore the importance of Theorem 112.1, let
us return to Theorem 99.1 (p. 100). There we are given a Borel outer measure
with (R) < . Then dene a function f by
f(x) := ((, x])
and observe that f is nondecreasing and right-continuous. Consequently, f pro-
duces a Lebesgue- Stieltjes measure
f
with the property that
f
((a, b]) = f(b) f(a)
for each half-open interval. However, it is clear from the denition of f that also
enjoys the same property:
((a, b]) = f(b) f(a).
Thus, and
f
agree on all half-open intervals and therefore they agree on all open
sets, since any open set is the disjoint union of half-open intervals. Hence, from
Corollary 114.1, they agree on all Borel sets. This allows us to conclude that there
is unique correspondance between nondecreasing, right-continuous functions and
nite Borel measures on R.
It is natural to ask whether the previous theorem remains true if B is only
assumed to be -measurable rather than being a Borel set. In general the answer
is no, but it is true if is assumed to a Borel regular outer measure. To see this,
4.10. REGULAR OUTER MEASURES 115
observe that if is a Borel regular outer measure and A is a -measurable set with
(A) < , then there exist Borel sets B
1
and B
2
such that
(114.1) B
2
A B
1
and (B
1
B
2
) = 0.
Proof. For this, rst choose a Borel set B
1
A with (B
1
) = (A). Then
choose a Borel set D B
1
A such that (D) = (B
1
A). Note that since A
and B
1
are -measurable, we have (B
1
A) = (B
1
) (A) = 0. Now take
B
2
= B
1
D. Thus, we have the following corollary.
115.1. Corollary. In the previous theorem, if is assumed to be a Borel
regular outer measure, then the conclusions remain valid if the phrase for each
Borel set B is replaced by for each -measurable set B.
Although not all Caratheodory outer measures are Borel regular, the following
theorems show that they do agree with Borel regular outer measures on the Borel
sets.
115.2. Theorem. Let be a Caratheodory outer measure. For each set A X,
dene
(115.1) (A) = inf(B) : B A, B a Borel set.
Then is a Borel regular outer measure on X, which agrees with on all Borel
sets.
Proof. We leave it as an easy exercise (Exercise 4.7) to show that is an outer
measure on X. To show that all Borel sets are -measurable, suppose D X is a
Borel set. Then, by Denition 79.1, (p.79), we must show that
(115.2) (A) (A D) +(A D)
whenever A X. For this we may as well assume (A) < . For > 0, choose
a Borel set B A such that (B) < (A) + . Then, since is a Borel outer
measure (Theorem 87.3), we have
+(A) (B) (B D) +(B D)
(A D) +(A D),
which establishes (115.2) since is arbitrary. Also, if B is a Borel set, we claim that
(B) = (B). Half the claim is obvious because (B) (B) by denition. As
for the opposite inequality, choose a sequence of Borel sets D
i
X with D
i
B
116 4. MEASURE THEORY
and lim
i
(D
i
) = (B). Then, with D = liminf
i
D
i
, we have by Corollary
83.3 (v)
(B) (D) liminf
i
(D
i
) = (B),
which establishes the claim. Finally, since and agree on Borel sets, we have for
arbitrary A X,
(A) = inf(B) : B A, B a Borel set
= inf(B) : B A, B a Borel set.
For each positive integer i, let B
i
A be a Borel set with (B
i
) < (A) + 1/i.
Then
B =
i=1
B
i
A
is a Borel set with (B) = (A), which shows that is Borel regular.
4.11. Outer Measures Generated by Measures
Thus far we have seen that with every outer measure there is an associ-
ated measure. This measure is dened by restricting the outer measure
to its measurable sets. In this section, we consider the situation in re-
verse. It is shown that a measure dened on an abstract space generates
an outer measure and that if this measure is -nite, the extension is
unique. An important consequence of this development is that any nite
Borel measure is necessarily regular.
We begin by describing a process by which a measure generates an outer mea-
sure. This method is reminiscent of the one used to dene Lebesgue- Stieltjes
measure. Actually, this method does not require the measure to be dened on
a -algebra, but rather, only on an algebra of sets. We make this precise in the
following Denition.
116.1. Definitions. An algebra in a space X is dened as a nonempty col-
lection of subsets of X that is closed under the operations of nite unions and
complements. Thus, the only dierence between an algebra and a -algebra is that
the latter is closed under countable unions. By a measure on an algebra, /, we
mean a function : / [0, ] satisfying the properties
(i) () = 0,
(ii) if A
i
is a disjoint sequence of sets in / whose union is also in /, then
i=1
A
i
_
=
i=1
(A
i
).
4.11. OUTER MEASURES GENERATED BY MEASURES 117
Consequently, a measure on an algebra / is a measure (in the sense of Denition
107.1) if and only if / is a -algebra. A measure on / is called -nite if X can
be written
X =
i=1
A
i
with A
i
/ and (A
i
) < .
A measure on an algebra / generates a set function
(E) := inf
_
i=1
(A
i
)
_
where the inmum is taken over countable collections A
i
such that
E
i=1
A
i
, A
i
/.
Note that this denition is in the same spirit as that used to dene Lebesgue
measure or more generally, Lebesgue- Stieltjes measure.
117.1. Theorem. Let be a measure on an algebra / and let
be the corre-
sponding set function generated by . Then
(i)
is an outer measure,
(ii)
-measurable.
Proof. The proof of (i) is similar to showing that
i=1
A
i
.
Set
B
i
= A A
i
(A
i1
A
i2
A
1
).
These sets are disjoint. Furthermore, B
i
/, B
i
A
i
, and A =
i=1
B
i
. Hence,
by the countable additivity of ,
(A) =
i=1
(B
i
)
i=1
(A
i
).
Since, by denition, the inmum of the right-side of this expression tends to
(A),
this shows that (A)
(A).
(iii) For A /, we must show that
(E)
(E A) +
(E A)
118 4. MEASURE THEORY
whenever E X. For this we may assume that
i=1
A
i
and
i=1
(A
i
) <
(E) +.
Since is additive on /, we have
(A
i
) = (A
i
A) +(A
i
A).
In view of the inclusions
E A
i=1
(A
i
A) and E A
i=1
(A
i
A),
we have
(E) + >
i=1
(A
i
A) +
i=1
(A
i
A)
>
(E A) +
(E A).
Since is arbitrary, the desired result follows.
118.1. Example. Let us see how the previous result can be used to produce
Lebesgue- Stieltjes measure. Let / be the algebra formed by including , R, all
intervals of the form (, a], (b, +) along with all possible nite disjoint unions
of these and intervals of the form (a, b]. Suppose that f is a nondecreasing, right-
continuous function and dene on intervals (a, b] in / by
((a, b]) = f(b) f(a),
and then extend to all elements of / by additivity. Then we see that the outer
measure
(A) =
(A) for all A /, which agrees with Theorem 98.1, p. 98.
118.2. Remark. In the previous example, the right-continuity of f is needed
to ensure that is in fact a measure on /. For example, if
f(x) :=
_
_
_
0 x 0
1 x > 0
4.11. OUTER MEASURES GENERATED BY MEASURES 119
then ((0, 1]) = 1. But (0, 1] =
k=1
(
1
k+1
,
1
k
] and
k=1
(
1
k + 1
,
1
k
]
_
=
k=1
((
1
k + 1
,
1
k
])
= 0,
which shows that is not a measure.
Next is the main result of this section, which in addition to restating the results
of Theorem 117.1, ensures that the outer measure generated by is unique.
119.1. Theorem (Caratheodory-Hahn Extension Theorem). Let be a mea-
sure on an algebra /, let
be the
-algebra of
-measurable sets.
(i) Then /
/ and
= on /,
(ii) Let / be a -algebra with / / /
(E) since if A
i
is any countable
collection in / whose union contains E, then
(E)
_
i=1
A
i
_
i=1
(A
i
) =
i=1
(A
i
).
To prove equality let A / with (A) < . Then we have
(119.1) (E) +(A E) = (A) =
(A) =
(E) +
(A E).
Note that A E /, and therefore (A E)
(A E)
whenever A / with (A) < . Since is -nite, thee exist A
i
/ such that
X =
i=1
A
i
with (A
i
) < for each i. We may assume that the A
i
are disjoint (Lemma 82.1)
and therefore
(E) =
i=1
(E A
i
) =
i=1
(E A
i
) =
(E).
120 4. MEASURE THEORY
Let us consider a special case of this result, namely, the situation in which /
is the family of Borel sets in a metric space X. If is a nite measure dened on
the Borel sets, the previous result states that the outer measure,
, generated by
agrees with on the Borel sets. Theorem 112.1 (p.112) asserts that
enjoys
certain regularity properties. Since and
on an arbitrary set E X by
(120.1)
(E) = inf(B) : B E, B /.
We have the following result.
120.2. Theorem. Consider a measure space (X, /, ). The set function
and
be the outer measures generated by as described in (117.1) (p. 117) and (120.1),
respectively. Then, for each E X with (E) < there exists B / such that
B E,
(B) =
(B) =
(E) =
(E).
Proof. We will show that for any E X there exists B / such that
B E and (B) =
(B) =
(E) =
(E).
Note that for each > 0, and any set E, there exists a sequence A
i
/
such that
E
i=1
A
i
and
i=1
(A
i
)
(E) +.
Setting A = A
i
, we have
(A) <
(E) +.
EXERCISES FOR CHAPTER 4 121
For each positive integer k, use this observation with = 1/k to obtain a set
A
k
/ such that A
k
E and (A
k
) <
k=1
A
k
.
Then B / and since E B A
k
, we have
(E)
(B) (B) (A
k
) <
(E) + 1/k.
Since k is arbitrary, it follows that (B) =
(B) =
(E).
Exercises for Chapter 4
Section 4.1
4.1 Prove (vi) of Corollary 83.3.
4.2 (a) What are the -measurable sets in each of Examples 78.2, (p. 78)?
(b) In example (iv), let
(A).
What is (A) and what are the corresponding -measurable sets?
4.3 In (i) of Corollary 83.3, it was shown that (E
2
E
1
) = (E
2
)(E
1
) provided
E
1
E
2
are -measurable with (E
1
) < . Prove this result still remains
true if E
2
is not assumed to be -measurable.
Section 4.2
4.4 Prove that there is a smallest -algebra that contains all of the closed sets in
R
n
. That is, prove there is a -algebra that contains all closed sets and has
the property that if
1
is another -algebra containing all closed sets, then
1
.
4.5 In a topological space X the family of Borel sets, B, is by denition, the -
algebra generated by the closed sets. The method below is another way of
describing the Borel sets using transnite induction. You are to ll in the
necessary steps:
(a) For an arbitrary family T of sets, let
T
k=1
E
k
: where either E
i
T or
E
i
T for all i N
Let denote the smallest uncountable ordinal. We will use transnite
induction to dene a family c
has been
dened for each such that 0 < . Dene
c
:=
_
0<
c
and dene
/ :=
0<
c
.
(c) Show that each c
B.
(d) Show that / B
(e) Now show that / is a -algebra to conclude that / = B.
(i) Show that , X /.
(ii) Let A / = A c
A c
k=1
A
k
/. (Hint:
Each A
k
/
k
for some
k
< . We know there is < such that
>
k
for each k N ) .
4.6 Prove that the set function dened in (112.1) is an outer measure whose
measurable sets include all open sets.
4.7 Prove that the set function dened in (115.1) is an outer measure on X.
4.8 An outer measure on a space X is called -nite if there exists a countable
number of sets A
i
with (A
i
) < such that X
i=1
A
i
. Assuming is a
-nite Borel regular outer measure on a metric space X, prove that E X is
-measurable if and only if there exists an F
t
(E) = ([t[ E),
prove that
t
(N) = 0 whenever (N) = 0.
4.18 Let I, I
1
, I
2
, . . . , I
k
be intervals in R such that I
k
i=1
I
i
. Prove that
v(I)
k
i=1
v(I
i
)
where v(I) denotes the length of the interval I.
4.19 Complete the proofs of (iv) (v) and (v) (i) in Theorem 91.2.
4.20 Let P denote an arbitrary n 1-dimensional hyperplane in R
n
. Prove that
(P) = 0.
4.21 In this problem, we want to show that any Lebesgue measurable subset of R
must be densely populated in some interval. Thus, let E R be a Lebesgue
measurable set. For each > 0, show that there exists an interval I such that
(E I)
(I)
> 1 .
4.22 Suppose E R
n
is an arbitrary set with the property that there exists an
F
(A).
124 4. MEASURE THEORY
4.25 Let E R and for each real number t, let E
t
= x + t : x E. Prove that
(E) =
(E
t
). From this show that if E is Lebesgue measurable, then so is
E
t
.
4.26 Prove that Lebesgue measure on R
n
is independent of the choice of coordinate
system. That is, prove that Lebesgue outer measure is invariant under rigid
motions in R
n
.
4.27 Let E
k
be a sequence of Lebesgue measurable sets contained in a compact
set K R
n
. Assume for some > 0, that (E
k
) > for all k. Prove that there
is some point that belongs to innitely many E
k
s.
4.28 Let / denote the class of -measurable sets of an outer measure dened on
a set X. If (X) < , prove that family
T := A / : (A) > 0
is at most countable.
Section 4.4
4.29 Prove that the Cantor-type set C
f
(A) = inf
_
h
k
F
[f(b
k
) f(a
k
)]
_
,
where the inmum is taken over all countable collections T of closed intervals
of the form h
k
:= [a
k
, b
k
] such that
E
h
k
F
h
k
.
In other words, the denition of
f
(A) is the same as
f
(A) except that closed
intervals [a
k
, b
k
] are used instead of half-open intervals (a
k
, b
k
].
As in the case of Lebesgue- Stieltjes measure it can be easily seen that
f
is a Carathedory measure. (You need not prove this).
(a) Prove that
f
(A)
f
(A) for all sets A R
n
.
(b) Prove that
f
(B) =
f
(B) for all Borel sets B if f is left-ontinuous.
Section 4.7
4.38 For A R
n
, prove that
(n)2
n
n
n/2
(A) H
n
(A) (n)2
n
n
n/2
(A).
4.39 For each A R
n
, prove that there exists a G
) where A
(A) := inf
_
i=1
h(diam E
i
) : A
i=1
E
i
R
n
, h(diam E
i
) <
_
,
and
H
h
(A) := lim
0
H
h
(A).
With the Cantor set C(S
h
) that was constructed above, it follows that
(126.3)
1
4
H
h
(C(S
h
)) 1.
The proof of this proceeds precisely the same way as in Section 4.8.
EXERCISES FOR CHAPTER 4 127
(a) With s = 0, our function h in (126.1) becomes h(r) = 1/ log(1/r) and
we obtain a corresponding Cantor set C(S
h
). With the help of (126.3)
prove that the Hausdor dimension of C(S
h
) is zero, thus showing that the
converse of Problem 1 is not true.
(b) Now take s = 1 and then our function h in (126.1) becomes h(r) =
r log(1/r) and again we obtain a corresponding Cantor set C(S
h
). Prove
that the Hausdor dimension of C(S
h
) is 1 which shows that there are sets
other than intervals in R that have dimension 1.
4.44 For each arbitrary set A R
n
, prove that there exists a G
set B A such
that
(127.1) H
s
(B) = H
s
(A).
127.1. Remark. This result shows that all three of our primary measures,
namely, Lebesgue measure, Lebesgue- Stieltjes measure and Hausdor measure,
share the same important regularity property (127.1).
4.45 If A R is an arbitrary set, show that H
1
(A) =
(A).
127.2. Remark. This result is also true in R
n
but more dicult to prove.
That is, if A R
n
, then H
n
(A) =
(A) where
denotes n-dimensional
Lebesgue measure.
Also, in this problem, you will see the importance of the constant that
appears in the denition of Hausdor measure. The constant (s) that appears
in the denition of H
s
-measure is equal to 2 when s = 1. That is, (1) = 2,
and therefore, the denition of H
1
(A) can be written as
H
1
(A) = lim
0
H
1
(A)
where
H
1
(A) = inf
_
i=1
diam E
i
: A
i=1
E
i
R
n
, diam E
i
<
_
,
4.46 If A R
n
is an arbitrary set and 0 t n, prove that if H
t
on all subsets of
X in the following way: For an arbitrary set A X let
(A) := inf
_
i=1
(E
i
)
_
where the inmum is taken over all countable collections E
i
such that
A
i=1
E
i
, E
i
/.
Prove that / = /
where /
-measurable sets
and that =
on /.
4.53 In an abstract measure space (X, /, ), if A
i
is a countable disjoint family
of sets in /, we know that
(
i=1
A
i
) =
i=1
(A
i
).
Prove that converse is essentially true. That is, under the assumption that
(X) < , prove that if A
i
is a countable family of sets in / with the
property that
(
i=1
A
i
) =
i=1
(A
i
),
then (A
i
A
j
) = 0 whenever i ,= j.
4.54 Recall that an algebra in a space X is a nonempty collection of subsets of X
that is closed under the operations of nite unions and complements. Also recall
that a measure on an algebra, /, is a function : / [0, ] satisfying the
properties
(i) () = 0,
(ii) if A
i
is a disjoint sequence of sets in / whose union is also in /, then
i=1
A
i
_
=
i=1
(A
i
).
EXERCISES FOR CHAPTER 4 129
Finally, recall that a measure on an algebra / generates a set function
(E) := inf
_
i=1
(A
i
)
_
where the inmum is taken over countable collections A
i
such that
E
i=1
A
i
, A
i
/.
Assuming that (X) < , prove that
. Prove:
(a)
_
, x > 0
0, x = 0,
, x < 0
for each x R and let
() () = + and () () = .
131
132 5. MEASURABLE FUNCTIONS
The operations
(), ( ) and ( )
are undened.
We endow R with a topology called the order topology in the following manner.
For each a R let
L
a
= R x : x < a and R
a
= R x : x > a.
The collection o = L
a
: a R R
a
: a R is taken as a subbase for this
topology. A base for the topology is given by
o R
a
L
b
: a, b R, a < b.
Observe that the topology on R induced by the order topology on R is precisely
the usual topology on R.
Suppose X and Y are topological spaces. Recall that a mapping f : X Y
is continuous if and only if f
1
(U) is open whenever U Y is open. We dene a
measurable mapping analogously.
132.1. Definitions. Suppose (X, /) and (Y, ^) are measure spaces. A map-
ping f : X Y is called measurable with respect to / and ^ if
(132.1) f
1
(E) / whenever E ^.
If there is no danger of confusion, reference to / and ^ will be omitted, and we
will simply use the term measurable mapping.
In case Y is a topological space, a restriction is placed on ^. In this case it
is always assumed that ^ is the -algebra of Borel sets. Thus, in this situation, a
mapping (X, /)
f
(Y, ^) is measurable if
(132.2) f
1
(E) / whenever E Y is a Borel set.
The reason for imposing this condition is to ensure that continuous mappings will be
measurable. That is, if both X and Y are topological spaces, X
f
Y is continuous
and / contains the Borel sets of X, then f is measurable, since f
1
(E) /
whenever E is a Borel set, see Exercise 5.1. An important example of this is when
X = R
n
, / is the class of Lebesgue measurable sets and Y = R. Here, it is
required that f
1
(E) is Lebesgue measurable whenever E R is Borel, in which
case f is called a Lebesgue measurable function.
The following observation will be useful in the development. Dene
(132.3) = E : E R and f
1
(E) /.
5.1. ELEMENTARY PROPERTIES OF MEASURABLE FUNCTIONS 133
Note that is closed under countable unions. It is also closed under complemen-
tation since
(132.4) f
1
[(E R)
] = [f
1
(E)]
f
1
f
1
/
for E / and thus, is a -algebra.
One of the most important situations is when Y is taken as R (endowed with
the order topology) and (X, /) is a topological space with / the collection of
Borel sets. Then f is called a Borel measurable function. In view of (132.3)
and (132.4), note that a continuous mapping is a Borel measurable function. The
denitions imply that E is a measurable set if and only if
E
is a measurable
function.
In case f : X R, it will be convenient to characterize measurability in terms
of the sets X x : f(x) > a for a R. To simplify notation, we simply write
f > a to denote these sets. The sets f > a are called the superlevel sets of
f. The behavior of a function f is to a large extent reected in the properties of
its superlevel sets. For example, if f is a continuous function on a metric space X,
then f > a is an open set for each real number a. If the function is nicer, then
we should expect better behavior of the superlevel sets. Indeed, if f is an innitely
dierentiable function f dened on R
n
with nonvanishing gradient, then not only
is each f > a an open set, but an application of the Implicit Function Theorem
shows that its boundary is a smooth manifold of dimension n 1 as well.
We begin by showing that the denition of an R-valued measurable function
could just as well be stated in terms of its level sets.
133.1. Theorem. Let f : X R where (X, /) is a measure space. The
following conditions are equivalent:
(i) f is measurable.
(ii) f > a / for each a R.
(iii) f a / for each a R.
(iv) f < a / for each a R.
(v) f a / for each a R.
Proof. (i) implies (ii) by denition since f > a = f
1
(a, ). In view of
f a =
k=1
f > a1/k, (ii) implies (iii). The set f < a is the complement
of f a, thus establishing the next implication. Similar to the proof of the
rst implication, we have f a =
k=1
f < a + 1/k, which shows that (iv)
implies (v). For the proof that (v) implies (i), in view of (132.3) and (132.4)
it is sucient to show that f
1
(U) / whenever U R is open. Since f
1
preserves unions and intersections, we need only consider f
1
(J) where J assumes
134 5. MEASURABLE FUNCTIONS
the form J
1
= [, a), J
2
= (a, b), and J
3
= (b, ] for a, b R. By assumption,
f b / and therefore f
1
(J
3
) /. Also,
J
1
=
k=1
[, a
k
]
where a
k
< a and a
k
a as k . Hence, f
1
(J
1
) /. Finally, f
1
(J
2
) /
since J
2
= J
1
J
3
.
134.1. Theorem. A function f : X R is measurable if and only if
(i) f
1
/ and f
1
/ and
(ii) f
1
(a, b) / for all open intervals (a, b) R.
Proof. If f is measurable, then (i) and (ii) are satised since , and
(a, b) are Borel subsets of R.
A set E R is Borel if and only if ER is a Borel subset of R. Thus, in order
to prove f is measurable, it suces to show that f
1
(E) / whenever E is a
Borel subset of R. Since f
1
preserves unions of sets and since any open set in R
is the disjoint union of open intervals, we see that f
1
(U) / whenever U R
is an open set. With dened as in (132.3), we see that is a -algebra that
contains the open sets of R and therefore it contains all Borel sets.
We now proceed to show that measurability is preserved under elementary
arithmetic operations on measurable functions. For this, the following will be useful.
134.2. Lemma. If f and g are measurable functions, then the following sets are
measurable:
(i) X x : f(x) > g(x),
(ii) X x : f(x) g(x),
(iii) X x : f(x) = g(x)
Proof. If f(x) > g(x), then there is a rational number r such that f(x) >
r > g(x). Therefore, it follows that
f > g =
rQ
(f > r g < r) ,
and (i) easily follows. The set (ii) is the complement of the set (i) with f and g
interchanged and it is therefore measurable. The set (iii) is the intersection of two
measurable sets of type (ii), and so it too is measurable.
Since all functions under discussion are extended real-valued, we must take
some care in dening the sum and product of such functions. If f and g are
5.1. ELEMENTARY PROPERTIES OF MEASURABLE FUNCTIONS 135
measurable functions, then f + g is undened at points where it would be of the
form . This diculty is overcome if we dene
(134.1) (f +g)(x): =
_
_
_
f(x) +g(x) x X B
x B
where R is chosen arbitrarily and where
(135.1) B: = (f
1
g
1
)
(f
1
g
1
).
Similarly, we dene
(135.2) (fg)(x): =
_
_
_
f(x)g(x) x X B
x B
where R. with denition we have the following.
135.1. Theorem. If f, g : X R are measurable functions, then f +g and fg
are measurable.
Proof. We will treat the case when f and g have values in R. The proof is
similar in the general case and is left as an exercise, (Exercise 5.2).
To prove that the sum is measurable, dene F : X R R by
F(x) = (f(x), g(x))
and G: R R R by
G(x, y) = x +y.
Then G F(x) = f(x) + g(x), so it suces to show that G F is measurable.
Referring to Theorem 134.1, we need only show that (G F)
1
(J) / whenever
J R is an open interval. Now U : = G
1
(J) is an open set in R
2
since G is
continuous. Furthermore, U is the union of a countable family, T, of 2-dimensional
intervals I of the form I = I
1
I
2
where I
1
and I
2
are open intervals in R. Since
F
1
(I) = f
1
(I
1
) g
1
(I
2
),
we have
F
1
(U) = F
1
_
IF
I
_
=
IF
F
1
(I),
which is a measurable set. Thus G F is measurable since
(G F)
1
(J) = F
1
(U).
The product is measurable by essentially the same proof.
136 5. MEASURABLE FUNCTIONS
135.2. Remark. In the situation of abstract measure spaces, if
(X, /)
f
(Y, ^)
g
(Z, T)
are measurable functions, the denitions immediately imply that the composition
g f is measurable. Because of this, one might be tempted to conclude that the
composition of Lebesgue measurable functions is again Lebesgue measurable. Lets
look at this closely. Suppose f and g are Lebesgue measurable functions:
R
f
R
g
R
Thus, here we have X = Y = Z = R. Since Z = R, our convention requires
that we take T to be the Borel sets. Moreover, since g is assumed to be Lebesgue
measurable, the denition requires / to contain the Lebesgue measurable sets. If
f g were to be Lebesgue measurable, it would be necessary that f
1
(E) /
whenever E ^; that is, it would be necessary for f
1
to preserve Lebesgue
measurable sets. However, the following example shows this is not generally true.
136.1. Example. (The Cantor-Lebesgue Function) Our example is based on
the construction of the Cantor ternary set. Recall (p. 93) that the Cantor set C
can be expressed as
C =
j=1
C
j
where C
j
is the union of 2
j
closed intervals that remain after the j
th
step of the
construction. Each of these intervals has length 3
j
. Thus, the set
D
j
= [0, 1] C
j
consists of those 2
j
1 open intervals that are deleted at the j
th
step. Let these
intervals be denoted by I
j,k
, k = 1, 2, . . . , 2
j
1, and order them in the obvious
way from left to right. Now dene a continuous function f
j
on [0,1] by
f
j
(0) = 0,
f
j
(1) = 1,
f
j
(x) =
k
2
j
for x I
j,k
,
and dene f
j
linearly on each interval of C
j
. The function f
j
is continuous, nonde-
creasing and satises
[f
j
(x) f
j+1
(x)[ <
1
2
j
for x [0, 1].
Since
[f
j
f
j+m
[ <
j+m1
i=j
1
2
i
<
1
2
j1
5.1. ELEMENTARY PROPERTIES OF MEASURABLE FUNCTIONS 137
it follows that the sequence f
j
is Cauchy in the space of continuous functions and
thus converges uniformly to a continuous function f, called the Cantor-Lebesgue
function.
Note that f is nondecreasing and is constant on each interval in the complement
of the Cantor set. Furthermore, f : [0, 1] [0, 1] is onto. In fact, it is easy to see
that f(C) = [0, 1] because f(C) is compact and f([0, 1] C) is countable.
We use the Cantor-Lebesgue function to show that the composition of Lebesgue
measurable functions need not be Lebesgue measurable. Let h(x) = f(x) + x
and observe that h is strictly increasing since f is nondecreasing. Thus, h is a
homeomorphism from [0, 1] onto [0, 2]. Furthermore, it is clear that h carries the
complement of the Cantor set onto an open set of measure 1. Therefore, h maps the
Cantor set onto a set P of measure 1. Now let N be a non-Lebesgue measurable
subset of P, see Exercise 4.34. Then, with A = h
1
(N), we have A C and
therefore A is Lebesgue measurable since (A) = 0. Thus we have that h carries a
measurable set onto a nonmeasurable set.
Note that h
1
is measurable since it is continuous. Let g := h
1
. Observe
that A is not a Borel set, for if it were, then g
1
(A) would be a Borel set. But
g
1
(A) = h(A) = N and N is not a Borel set. Now
A
is a Lebesgue measurable
function since A is a Lebesgue measurable set. But
A
h
1
=
N
,
which shows that this composition of Lebesgue measurable functions is not Lebesgue
measurable. To summarize the properties of the Cantor-Lebesgue function, we have:
137.1. Corollary. The Cantor-Lebesgue function, f and its associate h(x) :=
f(x) +x, described above has the following properties:
(i) f(C) = [0, 1]; that is, f maps a set of measure 0 onto a set of positive measure.
(ii) The inverse of the function h(x) := f(x)+x maps a Borel set onto a Lebesgue
measurable set that is not a Borel set; that is, there exist homeomorphisms
that do not preserve Borel sets. In particular, continuous mappings need not
preserve Borel sets.
(iii) h maps a Lebesgue measurable set onto a non-measurable set.
(iv) The composition of Lebesgue measurable functions need not be Lebesgue mea-
surable.
(v) The continuous image of a Borel set need not be Borel.
Although the example above shows that Lebesgue measurable functions are not
generally closed under composition, a positive result can be obtained if the outer
138 5. MEASURABLE FUNCTIONS
function in the composition is assumed to be Borel measurable. The proof of the
following theorem is a direct consequence of the denitions.
138.0. Theorem. Theorem Suppose f : X R is Lebesgue measurable and
: R R is Borel measurable. Then f is Lebesgue measurable
The function is required to have R as its domain of denition because f is
extended real-valued; however, any Borel function dened on R can be extended
to R by assigning arbitrary values to and .
As a consequence of this result, we have the following corollary which comple-
ments Theorem 135.1.
138.1. Theorem. Corollary Let f : X R be a measurable function.
(i) Let g(x) = [f(x)[
p
, 0 < p < , where f(x) is nite, and let g assume arbitrary
extended values on the sets f
1
(), and f
1
(). Then g is measurable.
(ii) Let g(x) =
1
f(x)
where f(x) is nite and let g assume arbitrary extended values
on the sets f
1
(0), f
1
() and f
1
(). Then g is measurable.
Proof. For (i), dene (t) = [t[
p
for t R and assign arbitrary values to
() and (). Now apply the previous theorem.
For (ii), proceed in a similar way by dening (t) =
1
t
when t ,= 0, , and
assign arbitrary values to (0), (), and ().
For much of much of the development thus far, the measure in (X, /, )
has played no role. We have only used the fact that / is a -algebra. Later it
will be necessary to deal with functions that are not necessarily dened on all of
X, but only on the complement of some set of -measure 0. That is, we will deal
with functions that are dened only -almost everywhere. A measurable set N
is called a -null set if (N) = 0. A property that holds for all x X except
for those x in some -null set is said to hold -almost everywhere. The term
-almost everywhere is often written in abbreviated form, -a.e. . If it is clear
from context that the measure is under consideration, we will simply use the
terms null set and almost everywhere.
The next result shows that a measurable function on a complete measure space
remains measurable if it is altered on an arbitrary set of measure 0.
138.2. Theorem. Let (X, /, ) be a complete measure space and let f, g be
extended real-valued functions dened on X. If f is measurable and f = g almost
everywhere, then g is measurable.
5.1. ELEMENTARY PROPERTIES OF MEASURABLE FUNCTIONS 139
Proof. Let N = x : f(x) = g(x). Then (
N) = 0 and thus,
N as well as
all subsets of
N are measurable. For a R, we have
g > a =
_
g > a N
_
_
g > a
N
_
=
_
f > a N
_
_
g > a
N
_
/.
139.1. Remark. In case is a complete measure, this result allows us to
attach the meaning of measurability to a function f that is dened merely almost
everywhere. Indeed, if N is the null set on which f is not dened, we modify
the denition of measurability by saying that f is measurable if f > a
N is
measurable for each a R. This is tantamount to saying that f is measurable,
where f is an extension of f obtained by assigning arbitrary values to f on N.
This is easily seen because
f > a =
_
f > a N
_
_
f > a
N
_
;
the rst set on the right is of measure zero because is complete, and therefore
measurable. Furthermore, for functions f, g that are nite-valued at -almost every
point, we may dene f +g as (f +g)(x) = f(x) +g(x) for all those x X at which
both f and g are dened and do not assume innite values of opposite sign. Then,
if both f and g are measurable, f +g is measurable. A similar discussion holds for
the product fg.
It therefore becomes apparent that functions that coincide almost everywhere
may be considered equivalent. In fact, if we dene f g to mean that f = g almost
everywhere, then denes an equivalence relation as discussed in Denition 17.6
and thus, a function could be regarded as an equivalence class of functions.
It should be kept in mind that this entire discussion pertains only to the situa-
tion in which the measure space (X, /, ) is complete. In particular, it applies in
the context of Lebesgue measure on R
n
, the most important example of a measure
space.
We conclude this section by returning to the context of an outer measure
dened on an arbitrary space X as in Denition 78.1. If f : X R, then according
to Theorem 133.1, f is -measurable if f a is a -measurable set for each
a R. That is, with E
a
= f a, the -measurability of f is equivalent to
(139.1) (A) = (A E
a
) +(AE
a
)
for any arbitrary set A X and each a R. The next result is often useful
in applications and gives a characterization of -measurability that appears to be
weaker than (139.1).
140 5. MEASURABLE FUNCTIONS
139.2. Theorem. Suppose is an outer measure on a space X. Then an
extended real-valued function f on X is -measurable if and only if
(139.2) (A) (A f a) +(A f b)
whenever A X and a < b are real numbers.
Proof. If f is -measurable, then (139.2) holds since it is implied by (139.1).
To prove the converse, it suces to show for any real number r, that (139.2)
implies
E = x : f(x) r
is -measurable. Let A X be an arbitrary set with (A) < and dene
B
i
= A
_
x : r +
1
i + 1
f(x) r +
1
i
_
for each positive integer i. First
1
, we will show
(140.1) > (A)
_
k=1
B
2k
_
=
k=1
(B
2k
).
The proof is by induction, so assume (140.1) is valid as k runs from 1 to j 1. That
is, assume
(140.2)
_
j1
k=1
B
2k
_
=
j1
k=1
(B
2k
).
Let
A
j
=
j1
k=1
B
2k
.
Then, using (139.2), the induction hypothesis, and the fact that
(140.3) (B
2j
A
j
)
_
f r +
1
2j
_
= B
2j
and
(140.4) (B
2j
A
j
)
_
f r +
1
2j 1
_
= A
j
,
1
Note the similarity between the technique used in the following argument and the proof of
Theorem 86.1, from (86.3) to the end of that proof
5.1. ELEMENTARY PROPERTIES OF MEASURABLE FUNCTIONS 141
we obtain
_
j
k=1
B
2k
_
= (B
2j
A
j
)
_
(B
2j
A
j
)
_
f r +
1
2j
__
+
_
(B
2j
A
j
)
_
f r +
1
2j 1
__
by (139.2)
= (B
2j
) +(A
j
) by (140.3) and (140.4)
= (B
2j
) +
j1
k=1
(B
2k
) by induction hypothesis (140.2)
=
j
k=1
(B
2k
).
Thus, (140.1) is valid as k runs from 1 to j for any positive integer j. In other
words, we obtain
> (A)
_
k=1
B
2k
_
_
j
k=1
B
2k
_
=
j
k=1
(B
2k
),
for any positive integer j. This implies
> (A)
k=1
(B
2k
).
Virtually the same argument can be used to obtain
> (A)
k=1
(B
2k1
),
thus implying
> 2(A)
k=1
(B
k
).
Now the tail end of this convergent series can be made arbitrarily small; that is,
for each > 0 there exists a positive integer m such that
>
i=m
(B
i
) =
_
i=m
B
i
_
_
A
_
r < f < r +
1
m
__
142 5. MEASURABLE FUNCTIONS
For ease of notation, dene an outer measure (S) = (S A) whenever S X.
With this notation, we have shown
>
__
r < f < r +
1
m
__
=
_
r < f
_
f < r +
1
m
__
(f > r)
__
f r +
1
m
__
.
The last inequality is implied by the subadditivity of . Therefore,
(A E) +(AE) = (E) +(
E)
= (E) +(f > r)
(E) +
__
f r +
1
m
__
+
= (A E) +
_
A
_
f r +
1
m
__
+
(A) +. by (139.2)
Since is arbitrary, this proves that E is -measurable.
5.2. Limits of Measurable Functions
In order to be useful in applications, it is necessary for measurability to
be preserved by virtually all types of limit operations on sequences of
measurable functions. In this section, it is shown that measurability is
preserved under the operations of upper and lower limits of sequences
of functions as well as upper and lower envelopes. It is also shown that
on a nite measure space, pointwise a.e. convergence of a sequence of
measurable functions implies uniform convergence on the complements
of sets of arbitrarily small measure (Egoros Theorem). Finally, the
relationship between convergence in measure and pointwise a.e. conver-
gence is investigated.
Throughout this section, it will be assumed that all functions are R-valued,
unless otherwise stated.
142.1. Definition. Let (X, /, ) be a measure space, and let f
i
be a se-
quence of measurable functions dened on X. The upper and lower envelopes
of f
i
are dened respectively as
sup
i
f
i
(x) = supf
i
(x) : i = 1, 2, . . .
and
inf
i
f
i
(x) = inff
i
(x) : i = 1, 2, . . ..
5.2. LIMITS OF MEASURABLE FUNCTIONS 143
Also, the upper and lower limits of f
i
are dened as
limsup
i
f
i
(x) = inf
j1
_
sup
ij
f
i
(x)
_
and
liminf
i
f
i
(x) = sup
j1
_
inf
ij
f
i
(x)
_
.
143.1. Theorem. Let f
i
be a sequence of measurable functions dened on
the measure space (X, /, ). Then sup
i
f
i
, inf
i
f
i
, limsup
i
f
i
, and liminf
i
f
i
are all
measurable functions.
Proof. For each a R the identity
X x : sup
i
f
i
(x) > a =
i=1
_
X f
i
(x) > a
_
implies that sup
i
f
i
is measurable. The measurability of the lower envelope follows
from
inf
i
f
i
(x) = sup
i
_
f
i
(x)
_
.
Now that it has been shown that the upper and lower envelopes are measurable, it
is immediate that the upper and lower limits of f
i
are also measurable.
We begin by investigating what information can be deduced from the pointwise
almost everywhere convergence of a sequence of measurable functions on a nite
measure space.
143.2. Definition. A sequence of measurable functions, f
i
, with the prop-
erty that
lim
i
f
i
(x) = f(x)
for -almost every x X is said to converge pointwise almost everywhere (or
more briey, converge pointwise a.e.) to f.
The following is one of the main results of this section.
143.3. Theorem (Egoro). Let (X, /, ) be a nite measure space and sup-
pose f
i
and f are measurable functions that are nite almost everywhere on X.
Also, suppose that f
i
converges pointwise a.e. to f. Then for each > 0 there
exists a set A / such that (
A) < and f
i
f uniformly on A.
First, we will prove the following lemma.
144 5. MEASURABLE FUNCTIONS
143.4. Theorem (Egoro). Assume the hypotheses of the previous theorem.
Then for each pair of numbers , > 0, there exist a set A / and an integer i
0
such that (
A) < and
[f
i
(x) f(x)[ <
whenever x A and i i
0
.
Proof. Choose , > 0. Let E denote the set on which the functions f
i
, i =
1, 2, . . . , and f are dened and nite. Also, let F be the set on which f
i
converges
pointwise to f. With A
0
: = E F, we have by hypothesis, (
A
0
) = 0. For each
positive integer i, let
A
i
= A
0
x : [f
j
(x) f(x)[ < for all j i.
Then, A
1
A
2
. . . and
i=1
A
i
= A
0
and consequently,
A
1
A
2
. . . with
i=1
A
i
=
A
0
. Since (
A
1
) (X) < , it follows from Theorem 108.2 (v) that
lim
i
(
A
i
) = (
A
0
) = 0.
The result follows by choosing i
0
such that (
A
i
0
) < and A = A
i
0
.
Proof of Egoroffs Theorem. Choose > 0. By the previous lemma, for
each positive integer i, there exist a positive integer j
i
and a measurable set A
i
such that
(
A
i
) <
2
i
and [f
j
(x) f(x)[ <
1
i
for all x A
i
and all j j
i
. With A dened as A =
i=1
A
i
, we have
A =
i=1
A
i
and
(
A)
i=1
(
A
i
) <
i=1
2
i
= .
Furthermore, if j j
i
, then
sup
xA
[f
j
(x) f(x)[ sup
xA
i
[f
j
(x) f(x)[
1
i
for every positive integer i. This implies that f
i
f uniformly on A.
144.1. Corollary. In the previous theorem, assume in addition that X is a
metric space and that is a Borel measure with (X) < . Then A can be taken
as a closed set.
5.2. LIMITS OF MEASURABLE FUNCTIONS 145
Proof. The previous theorem provides a set A such that A /, f
i
con-
verges uniformly to f and (
A) < /2. Since is a nite Borel measure, we see from
Theorem 120.1 (p.120) that there exists a closed set F A with (A F) < /2.
Hence, (
F) < and f
i
f uniformly on F.
145.0. Definition. Because of its importance, we attach a name to the type
of convergence exhibited in the conclusion of Egoros Theorem. Suppose that f
i
and f are measurable functions that are nite almost everywhere. We say that f
i
converges to f almost uniformly if for every > 0, there exists a set A / such
that (
A) < and f
i
converges to f uniformly on A. Thus, Egoros Theorem
states that pointwise a.e. convergence on a nite measure space implies almost
uniform convergence. The converse is also true and is left as Exercise 5.12.
145.1. Remark. The hypothesis that (X) < is essential in Egoros Theo-
rem. Consider the case of Lebesgue measure on R and dene a sequence of functions
by
f
i
=
[i,)
,
for each positive integer i. Then, lim
i
f
i
(x) = 0 for each x R, but f
i
does
not converge uniformly to 0 on any set A whose complement has nite Lebesgue
measure. Indeed, for any such set, it would follow that
A does not contain any
[i, ); that is, for each i, there would exist x [i, ) A with f
i
(x) = 1, thus
showing that f
i
does not converge uniformly to 0 on A.
145.2. Definition. A sequence of measurable functions f
i
dened relative
to the measure space (X, /, ) is said to converge in measure to a measurable
function f if for every > 0, we have
lim
i
_
X x : [f
i
(x) f(x)[
_
= 0.
We already encountered a result (Lemma 143.4) that essentially shows that
pointwise a.e. convergence on a nite measure space implies convergence in mea-
sure. Formally, it is as follows.
145.3. Theorem. Let (X, /, ) be a nite measure space, and suppose f
i
A) < and
[f
i
(x) f(x)[ <
146 5. MEASURABLE FUNCTIONS
whenever x A and i i
0
. Thus,
X x : [f
i
(x) f(x)[
A
if i i
0
. Since (
if both I and I
. Otherwise, dene I I
.
Now put the elements of T into a one-to-one order preserving correspondence with
the positive integers. With the elements of T labeled as I
k
, k = 1, 2, . . ., dene a
sequence of functions f
k
by f
k
=
I
k
. Then it is easy to see that f
k
0 in
measure but that f
k
(x) does not converge to 0 for any x [0, 1].
Although the sequence f
k
converges nowhere to 0, it does have a subsequence
that converges to 0 a.e., namely the subsequence
f
1
, f
2
, f
4
, . . . , f
2
k1, . . . .
In fact, this sequence converges to 0 at all points except x = 0. This illustrates the
following general result.
146.2. Theorem. Let (X, /, ) be a measure space and let f
i
and f be
measurable functions such that f
i
f in measure. Then there exists a subsequence
f
i
j
such that
lim
j
f
i
j
(x) f(x)
for -a.e. x X
Proof. Let i
1
be a positive integer such that
_
X x : [f
i
1
(x) f(x)[ 1
_
<
1
2
.
Assuming that i
1
, i
2
, . . . , i
k
have been chosen, let i
k+1
> i
k
be such that
_
X
_
x :
f
i
k+1
(x) f(x)
1
k + 1
__
1
2
k+1
.
Let
A
j
=
k=j
_
x : [f
i
k
(x) f(x)[
1
k
_
and observe that the sequence A
j
is descending. Since
(A
1
) <
k=1
1
2
k
< ,
5.3. APPROXIMATION OF MEASURABLE FUNCTIONS 147
with B =
j=1
A
j
, it follows that
(B) = lim
j
(A
j
) lim
j
k=j
1
2
k
= lim
j
1
2
j1
= 0.
Now select x
B. Then there exists an integer j = j
x
such that
x
A
j
x
=
k=j
x
_
X
_
y : [f
i
k
(y) f(y)[ <
1
k
__
.
If > 0, choose k
0
such that k
0
j
x
and
1
k
0
. Then for k k
0
, we have
[f
i
k
(x) f(x)[ <
1
k
,
which implies that f
i
k
(x) f(x) for all x
B.
There is another mode of convergence, fundamental in measure, which is
discussed in Exercise 5.13.
5.3. Approximation of Measurable Functions
In Section 3.2 certain fundamental approximation properties of Cara-
theodory outer measures were established. In particular, it was shown
that each Borel set B of nite measure contains a closed set whose mea-
sure is arbitrarily close to that of B. The structure of a Borel set can be
very complicated, but yet this result states that a complicated set can be
approximated by one with an elementary topological property. In this
section we pursue an analogous situation by showing that each measur-
able function on a metric space of nite measure is almost continuous
(Lusins Theorem). That is, every measurable function is continuous
on sets whose complements have arbitrarily small measure. This result
is in the same spirit as Egoros Theorem, which states that pointwise
a.e. convergence implies almost uniform convergence on a nite measure
space.
The characteristic function of a measurable set is the most elementary example
of a measurable function. The next level of complexity involves linear combinations
of such functions. A simple function on X is one that assumes only a nite
number of values: Thus, the range of a simple function, f, is a nite subset of R.
If rng f = a
1
, a
2
, . . . , a
k
, and A
i
= f
1
a
i
, then f can be written as
f =
k
i=1
a
i
A
i
.
We begin by proving that any measurable function is the pointwise a.e. limit of
measurable simple functions.
147.1. Theorem. Let f : X R be an arbitrary (possibly nonmeasurable)
function. Then the following hold.
148 5. MEASURABLE FUNCTIONS
(i) There exists a sequence of simple functions, f
i
, such that
f
i
(x) f(x) for each x X,
(ii) If f is nonnegative, the sequence can be chosen so that f
i
f,
(iii) If f is bounded, the sequence can be chosen so that f
i
f uniformly on X,
(iv) If f is measurable, the f
i
can be chosen to be measurable.
Proof. Assume rst that f 0. For each positive integer i, partition [0, i)
into i 2
i
half-open intervals of the form
_
k 1
2
i
,
k
2
i
_
, k = 1, 2, . . . , i 2
i
. Label
these intervals as H
i,k
and let
A
i,k
= f
1
(H
i,k
) and A
i
= f
1
_
[i, ]
_
.
These sets are pairwise disjoint and form a partition of X. The approximating
simple function f
i
on X is dened as
f
i
(x) =
_
_
k1
2
i
x A
i,k
i x A
i
If f is measurable, then the sets A
i,k
and A
i
are measurable and thus, so are the
functions f
i
. Moreover, it is easy to see that
f
1
f
2
. . . f and [f
i
(x) f(x)[ <
1
2
i
for x
A
i
.
Since
A
1
A
2
. . . and
i=1
A
i
= f
1
() ,
it follows that
lim
i
f
i
(x) = f(x) for x X.
Suppose f is bounded by some number, say M; that is, suppose f(x) M for
all x X. Then A
i
= for all i > M and therefore [f
i
(x) f(x)[ < 1/2
i
for all
x X, thus showing that f
i
f uniformly if f is bounded. This establishes the
Theorem in case f 0.
In general, let f
+
(x) = max(f(x), 0) and f
_
A
i,j
F
i,j
_
< /2
i+j
, j = 1, 2, . . . . Let
E
k
= X
k
j=1
F
i,j
for k = 1, 2, . . . , . (Keep in mind that i is xed, so it is not necessary to indicate
that E
k
depends on i). Then E
1
E
2
. . . ,
k=1
E
k
= E
, and
(E
) =
_
X
j=1
F
i,j
_
=
j=1
_
A
i,j
F
i,j
_
<
2
i
.
Since (X) < , it follows that
lim
k
(E
k
) = (E
) <
2
i
.
Hence, there exists a positive integer J = J(i) such that
(E
J
) =
_
X
J
j=1
F
i,j
_
<
2
i
.
For each H
i,j
, select an arbitrary point y
i,j
H
i,j
and let B
i
=
J
j=1
F
i,j
. Then
dene a continuous function g
i
on the closed set B
i
by
g
i
(x) = y
i,j
whenever x F
i,j
, j = 1, 2, . . . , J.
The functions g
i
are continuous (relative to B
i
) because the closed sets F
i,j
are
disjoint. Note that [f(x) g
i
(x)[ < 1/i for x B
i
. Therefore, on the closed set
F =
i=1
B
i
with (X F)
i=1
(X B
i
) < ,
it follows that the continuous functions g
i
converge uniformly to f, thus proving
that f is continuous on F.
150 5. MEASURABLE FUNCTIONS
The assumption (X) < can be replaced by the following: Let X be a metric
space that can be expressed as a countable union of open sets U
i
with U
i
U
i+1
and (U
i
) < , i = 1, 2, . . . . See Exercise 5.14.
We close this chapter with a table that reects the interaction of the various
types of convergences that we have encountered so far. A convergence-type listed
in the rst column implies one in the rst row if the corresponding entry of the
matrix is indicated by (along with the appropriate hypothesis).
Fundamental
in measure
Convergence
in measure
Almost
uniform
convergence
Pointwise
a.e. conver-
gence
Fundamental
in measure
For a sub-
sequence if
(X) <
For a subse-
quence
Convergence
in measure
For a sub-
sequence if
(X) <
For a subse-
quence
Almost
uniform
convergence
Pointwise
a.e. conver-
gence
if (X) <
if (X) <
if (X) <
k
(E) for each measurable set E. With dened as (E) = lim
k
k
(E),
prove that is a measure.
5.5 Let (X, /, ) be a measure space and suppose T : X Y is a mapping onto a
set Y . Let B be the family of all sets B Y with the property that T
1
(B)
/. Dene (B) = [T
1
(B)]. Prove that (Y, B, ) is a measure space.
5.6 Prove that a nondecreasing function dened on [0, 1] is Lebesgue measurable.
5.7 Let f be a measurable function on a measure space (X, /). Dene
f
(s) =
([f[ > s). The nonincreasing rearrangement of f on (0, ) is dened
as
f
(t): = infs :
f
(s) t.
Prove the following:
(i) f
i=1
and f are mea-
surable functions. Prove that f
i
f in measure if and only if each subsequence
of f
i
has a subsequence that converges to f -a.e.
5.11 Show that the supremum of an uncountable family of measurable R-valued
functions can fail to be measurable.
5.12 Suppose (X, /, ) is a nite measure space. Prove that almost uniform con-
vergence implies convergence almost everywhere.
5.13 A sequence f
i
of a.e. nite-valued measurable functions on a measure space
(X, /, ) is fundamental in measure if, for every > 0,
(x : [f
i
(x) f
j
(x)[ ) 0
as i and j . Prove that if f
i
is fundamental in measure, then there
is a measurable function f to which the sequence f
i
converges in measure.
Hint: Choose integers i
j+1
> i
j
such that
f
i
j
f
i
j+1
> 2
j
< 2
j
. The
152 5. MEASURABLE FUNCTIONS
sequence f
i
j
converges a.e. to a function f. Then it follows that
[f
i
f[
f
i
f
i
j
/2
f
i
j
f
/2.
By hypothesis, the measure of the rst term on the right is arbitrarily small if
i and i
j
are large, and the measure of the second term tends to 0 since almost
uniform convergence implies convergence in measure.
Section 5.3
5.14 Prove that Lusins Theorem can be extended to the following situation. Let
X be a metric space that can be expressed as a countable union of open sets
U
i
with U
i
U
i+1
and (U
i
) < , i = 1, 2, . . . . Assuming that (X, /, )
otherwise satises the same hypotheses as Theorem 149.1, prove that the same
conclusion holds.
5.15 Suppose f : [0, 1] R is Lebesgue measurable. For any > 0, show that there
is a continuous function g on [0, 1] such that
([0, 1] x : f(x) ,= g(x)) < .
5.16 Let (X, /, ) be a -nite measure space and suppose that f, f
k
, k = 1, 2, . . . ,
are measurable functions with
lim
k
f
k
(x) = f(x)
for almost all x X. Prove that there are measurable sets
E
0
, E
1
, E
2
, . . . , such that (E
0
) = 0,
X =
i=0
E
i
,
and f
k
f uniformly on each E
i
, i > 0.
CHAPTER 6
Integration
6.1. Denitions and Elementary Properties
Based on the ideas of H. Lebesgue, a far-reaching generalization of Rie-
mann integration has been developed. In this section we dene and
deduce the elementary properties of integration with respect to an ab-
stract measure.
We rst extend the notion of simple function to allow better approximation of
unbounded functions. Throughout this section and the next, we will assume the
context of a general measure space (X, /, ).
153.1. Definition. A function f : X R is called countably-simple if
it assumes only a countable number of values, including possibly . Given
a measure space (X, /, ), the integral of a nonnegative measurable countably-
simple function f : X R is dened to be
_
X
f d =
i=1
a
i
(f
1
a
i
)
where the range of f = a
1
, a
2
, . . . and, by convention, 0 = 0 = 0. Note
that the integral may equal .
153.2. Definitions. For an arbitrary function f : X R, we dene
f
+
(x) := f(x) if f(x) 0
f
and [f[ = f
+
+f
.
If f is a measurable countably-simple function and at least one of
_
X
f
+
d or
_
X
f
d is nite, we dene
_
X
f d: =
_
X
f
+
d
_
X
f
d.
If f : X R (not necessarily measurable), we dene the upper integral of f by
_
X
f d := inf
__
X
g d : g is measurable, countably-simple and g f -a.e.
_
153
154 6. INTEGRATION
and the lower integral of f by
_
X
f d := sup
__
X
g d : g is measurable, countably-simple and g f -a.e.
_
.
The integral (with respect to the measure ) of a measurable function f : X R
is said to exist if
_
X
f d =
_
X
f d,
in which case we write
_
X
fd
for the common value. If this value is nite, f is said to be integrable.
154.1. Remark. Observe that our denition requires f to be measurable if it
is to be integrable. See Exercise 6.8 which shows that measurability is necessary
for a function to be integrable provided the measure is complete.
154.2. Remark. If f is a countably-simple function such that
_
X
f
d is nite
then the denitions immediately imply that integral of f exists and that
(154.1)
_
X
f d =
i=1
a
i
(f
1
a
i
)
where the range of f = a
1
, a
2
, . . .. Clearly, the integral should not depend on
the order in which the terms of (154.1) appear. Consequently, the series converges
unconditionally, possibly to + (see Exercise 6.1). An analogous statement holds
if
_
X
f
+
d is nite.
154.3. Remark. It is clear from the denitions of upper and lower integrals
that if f = g -a.e. , then
_
X
f d =
_
X
g d, etc. From this observation it follows
that if both f and g are measurable, f = g -a.e. , and f is integrable then g is
integrable and
_
X
f d =
_
X
g d.
154.4. Definition. If A X (possibly nonmeasurable), we write
_
A
f d: =
_
X
f
A
d
and use analogous notation for the other integrals.
154.5. Theorem.
(i) If f is an integrable function, then f is nite -a.e.
6.1. DEFINITIONS AND ELEMENTARY PROPERTIES 155
(ii) If f and g are integrable functions and a, b are constants, then af + bg is
integrable and
_
X
(af +bg) d = a
_
X
f d +b
_
X
g d,
(iii) If f and g are integrable functions and f g -a.e., then
_
X
f d
_
X
g d,
(iv) If f is a integrable function and E /, then f
E
is integrable.
(v) A measurable function f is integrable if and only if [f[ is integrable.
(vi) If f is a integrable function, then
_
X
f d
_
X
[f[ d.
Proof. Each of the assertions above is easily seen to hold in case the functions
are countably-simple. We leave these proofs as exercises.
(i) If f is integrable, then there are integrable countably-simple functions g and
h such that g f h -a.e. Thus f is nite -a.e.
(ii) Suppose f is integrable and c is a constant. If c > 0, then for any integrable
countably-simple function g
cg cf if and only if g f.
Since
_
X
cg d = c
_
X
g d it follows that
_
X
cf d = c
_
X
f d
and
_
X
cf d = c
_
X
f d.
Clearly f is integrable and
_
X
f d =
_
X
f d. Thus if c < 0, then
cf = [c[ (f) is integrable and
_
X
cf d = [c[
_
(f) d = [c[
_
f d = c
_
X
f d.
Now suppose f, g are integrable and f
1
, g
1
are integrable countably-simple functions
such that f
1
f, g
1
g -a.e. Then f
1
+g
1
f +g -a.e. and
_
(f +g) d
_
X
(f
1
+g
1
) d =
_
X
f
1
d +
_
X
g
1
d.
Thus
_
X
f d +
_
X
g d
_
X
(f +g) d.
An analogous argument shows
_
X
(f +g) d
_
X
f d +
_
X
g d
156 6. INTEGRATION
and assertion (ii) follows.
(iii) If f, g are integrable and f g -a.e., then, by (ii), g f is integrable and
g f 0 -a.e. . Clearly
_
X
(g f) d =
_
E
d
for E /. Thus
_
X
f
E
d
_
X
f
E
d <
and since
<
_
X
g
E
d
_
X
f
E
d
_
X
f
E
d
_
X
h
E
d <
it follows that f
E
is integrable.
(v) If f is integrable, then by (iv) f
+
: = f
{x:f(x)>0}
and
f
: = f
{x:f(x)<0}
are integrable and by (ii) [f[ = f
+
+ f
is integrable. If
[f[ is integrable, then, by (iv), f
+
= [f[
{x:f(x)>0}
and f
= [f[
{x:f(x)<0}
are
integrable and hence f = f
+
f
is integrable.
(vi) If f is integrable, then by (v) f
_
X
f d
_
X
f
+
d
_
X
f
_
X
f
+
d +
_
X
f
d =
_
X
[f[ d.
The next result, whose proof is very simple, is remarkably strong in view of
the weak hypothesis. In particular, it implies that any bounded, nonnegative,
measurable function is -integrable. This exhibits a striking dierence between the
Lebesgue and Riemann integral (see Theorem 162.1 below).
156.1. Theorem. If f is -measurable and f 0 -a.e. , then the integral of
f exists: that is,
_
X
f d =
_
X
f d.
Proof. If the lower integral is innite, then the upper and lower integrals
are both innite. Thus we may assume that the lower integral is nite and, in
particular, (x : f(x) = ) = 0. For t > 1 and k = 0, 1, 2, . . . set
E
k
= x : t
k
f(x) < t
k+1
k=
t
k
E
k
.
Since each set E
k
is measurable, it follows that g
t
is a measurable countably-simple
function and g
t
f tg
t
-a.e. Thus
_
X
f d
_
X
tg
t
d = t
_
X
g
t
d t
_
X
f d.
for each t > 1 and therefore on letting t 1
+
,
_
X
f d
_
X
f d,
which implies our conclusion since
_
X
f d
_
X
f d,
is always true.
157.1. Theorem. If f is a nonnegative measurable function and g is a inte-
grable function, then
_
X
(f +g) d =
_
X
(f +g) d =
_
X
f d +
_
X
g d.
Proof. If f is integrable, the assertion follows from Theorem 154.5, so assume
that
_
X
fd = . Let h be a countably-simple function such that 0 h f, and
let k be an integrable countably-simple function such that k [g[. Then, by
Exercise 6.1
_
X
(f +g) d
_
X
(f [g[) d
_
X
(h k) d =
_
X
hd
_
X
k d
from which it follows that
_
X
(f +g) d = and the assertion is proved.
One of the main applications of this result is the following.
157.2. Corollary. If f is measurable, and if either f
+
or f
is integrable,
then the integral exists:
_
X
f d.
Proof. For example, if f
+
is integrable, take g := f
+
and f := f
in the
previous theorem to conclude that the integrals exist:
_
X
(f
f
+
) d =
_
X
f d
and therefore that
_
X
f d
exists.
158 6. INTEGRATION
157.3. Theorem. If f is -measurable, g is integrable, and [f[ [g[ -a.e. then
f is integrable.
Proof. This follows immediately from Theorem 154.5 (v) and Lemma 156.1.
f and lim
i
f
i
related? There are three funda-
mental results that address this question: Fatous lemma, the Monotone
Convergence Theorem, and Lebesgues Dominated Convergence Theo-
rem. These will be discussed along with associated results.
Our rst result concerning the behavior of sequences of integrals is Fatous Lemma.
Note the similarity between this result and its measure-theoretic counterpart, The-
orem 108.2, (vi).
We continue to assume the context of a general measure space (X, /, ).
158.1. Lemma (Fatous Lemma). If f
k
k=1
is a sequence of nonnegative -
measurable functions, then
_
X
liminf
k
f
k
d liminf
k
_
X
f
k
d.
Proof. Let g be any measurable countably-simple function such that 0 g
liminf
k
f
k
-a.e. . For each x X, set
g
k
(x) = inff
m
(x) : m k
and observe that g
k
g
k+1
and
lim
k
g
k
= liminf
k
f
k
g -a.e.
Write g =
j=1
a
j
A
j
where A
j
:= g
1
(a
j
). Therefore A
j
A
i
= if i ,= j and
X =
k=1
A
j
. For 0 < t < 1 set
B
j,k
= A
j
x : g
k
(x) > ta
j
.
Then each B
j,k
/ , B
j,k
B
j,k+1
and since lim
k
g
k
g -a.e., we have
k=1
B
j,k
= A
j
and lim
k
(B
j,k
) = (A
j
)
for j = 1, 2, . . . . Noting that
j=1
ta
j
B
j,k
g
k
f
m
6.2. LIMIT THEOREMS 159
for each m k, we nd
t
_
X
g d =
j=1
ta
j
(A
j
) = lim
k
j=1
ta
j
(B
j,k
) liminf
k
_
X
f
k
d.
Thus, on letting t 1
, we obtain
_
X
g d liminf
k
_
X
f
k
d.
By taking the supremum of the left-hand side over all countable simple functions
g with g liminf
k
f
k
, we have
_
X
liminf
k
f
k
d liminf
k
_
X
f
k
d.
Since liminf
k
f
k
is a nonnegative measurable function, we can apply Theorem
156.1 to obtain our desired conclusion.
159.1. Theorem (Monotone Convergence Theorem). If f
k
k=1
is a sequence
of nonnegative -measurable functions such that f
k
f
k+1
for k = 1, 2, . . . , then
lim
k
_
X
f
k
d =
_
X
lim
k
f
k
d.
Proof. Set f = lim
k
f
k
. Then f is -measurable,
_
X
f
k
d
_
X
f d for k = 1, 2, . . .
and
lim
k
_
X
f
k
d
_
X
f d.
The opposite inequality follows from Fatous Lemma.
159.2. Theorem. If f
k
k=1
is a sequence of nonnegative -measurable func-
tions, then
_
X
k=1
f
k
d =
k=1
_
X
f
k
d.
Proof. With g
m
:=
m
k=1
f
k
we have g
m
k=1
f
k
and the conclusion follows
easily from the Monotone Convergence Theorem and Theorem 154.5 (ii).
159.3. Theorem. If f is integrable and E
k
k=1
is a sequence of disjoint mea-
surable sets such that X =
k=1
E
k
, then
_
X
f d =
k=1
_
E
k
f d.
Proof. Assume rst that f 0, set f
k
= f
E
k
, and apply the previous
corollary. For arbitrary integrable f use the fact that f = f
+
f
.
160 6. INTEGRATION
159.4. Corollary. If f 0 is integrable and if is a set function dened by
(E) :=
_
E
f d
for every measurable set E, then is a measure.
160.1. Theorem (Lebesgues Dominated Convergence Theorem). Suppose g is
integrable, f is measurable, f
k
k=1
is a sequence of -measurable functions such
that [f
k
[ g -a.e. for k = 1, 2, . . . and
lim
k
f
k
(x) = f(x)
for -a.e. x X. Then
lim
k
_
X
[f
k
f[ d = 0.
Proof. Clearly [f[ g -a.e. and hence f and each f
k
are integrable. Set
h
k
= 2g [f
k
f[. Then h
k
0 -a.e. and by Fatous Lemma
2
_
X
g d =
_
X
liminf
k
h
k
d liminf
k
_
X
h
k
d
= 2
_
X
g d limsup
k
_
X
[f
k
f[ d.
Thus
limsup
k
_
X
[f
k
f[ d = 0.
6.3. Riemann and Lebesgue IntegrationA Comparison
The Riemann and Lebesgue integrals are compared, and it is shown that
a bounded function is Riemann integrable if and only if it is continuous
almost everywhere.
We rst recall the denition and some elementary facts concerning Riemann
integration. Suppose [a, b] is a closed interval in R. By a partition T of [a, b] we
mean a nite set of points x
i
m
i=0
such that a = x
0
< x
1
< < x
m
= b. Let
|T| : = maxx
i
x
i1
: 1 i m.
For each i 1, 2, . . . , m let x
i
be an arbitrary point of the interval [x
i1
, x
i
]. A
bounded function f : [a, b] R is Riemann integrable if
lim
P0
m
i=1
f(x
i
)(x
i
x
i1
)
6.3. RIEMANN AND LEBESGUE INTEGRATIONA COMPARISON 161
exists, in which case the value is the Riemann integral of f over [a, b], which we will
denote by
(R)
_
b
a
f(x) dx.
Given a partition T = x
i
m
i=0
of [a, b] set
U(T) =
m
i=1
_
sup
x[x
i1
,x
i
]
f(x)
_
(x
i
x
i1
)
L(T) =
m
i=1
_
inf
x[x
i1
,x
i
]
f(x)
_
(x
i
x
i1
).
Then
L(T)
m
i=1
f(x
i
)(x
i
x
i1
) U(T)
for any choice of the x
i
. Since the supremum (inmum) of
m
i=1
f(x
i
)(x
i
x
i1
)
over all choices of the x
i
is equal to U(T) (L(T)) we see that a bounded function
f is Riemann integrable if and only if
(161.1) lim
P0
(U(T) L(T)) = 0.
We next examine the eects of using a ner partition. Suppose then, that
T = x
i
m
i=1
is a partition of [a, b], z [a, b] T, and Q = T z. Thus T Q
and Q is called a renement T. Then z (x
i1
, x
i
) for some 1 i m and
sup
x[x
i1
,z]
f(x) sup
x[x
i1
,x
i
]
f(x),
sup
x[z,x
i
]
f(x) sup
x[x
i1
,x
i
]
f(x).
Thus
U(Q) U(T).
An analogous argument shows that L(T) L(Q). It follows by induction on the
number of points in Q that
L(T) L(Q) U(Q) U(T)
whenever T Q. Thus, U does not increase and L does not decrease when a
renement of the partition is used.
We will say that a Lebesgue measurable function f on [a, b] is Lebesgue in-
tegrable if f is integrable with respect to Lebesgue measure on [a, b].
161.1. Theorem. If f : [a, b] R is a bounded Riemann integrable function,
then f is Lebesgue integrable and
(R)
_
b
a
f(x) dx =
_
[a,b]
f d.
162 6. INTEGRATION
Proof. Let T
k
k=1
be a sequence of partitions of [a, b] such that T
k
T
k+1
and |T
k
| 0 as k . Write T
k
= x
k
j
m
k
j=0
. For each k dene functions l
k
, u
k
by setting
l
k
(x) = inf
t[x
k
i1
,x
k
i
)
f(t)
u
k
(x) = sup
t[x
k
i1
,x
k
i
)
f(t)
whenever x [x
k
i1
, x
k
i
), 1 i m
k
. Then for each k the functions l
k
, u
k
are
Lebesgue integrable and
_
[a,b]
l
k
d = L(T
k
) U(T
k
) =
_
[a,b]
u
k
d.
The sequence l
k
is monotonically increasing and bounded. Thus
l(x) := lim
k
l
k
(x)
exists for each x [a, b] and l is a Lebesgue measurable function. Similarly, the
function
u := lim
k
u
k
is Lebesgue measurable and l f u on [a, b]. Since f is Riemann integrable, it
follows from Lebesgues Dominated Convergence Theorem 160.1 and (161.1) that
(162.1)
_
[a,b]
(u l) d = lim
k
_
[a,b]
(u
k
l
k
) d = lim
k
(U(T
k
) L(T
k
) = 0.
Thus l = f = u -a.e. on [a, b] (see Exercise 6.4), and invoking Theorem 160.1 once
more we have
_
[a,b]
f d = lim
k
_
[a,b]
u
k
d = lim
k
U(T
k
) = (R)
_
b
a
f(x) dx.
162.1. Theorem. A bounded function f : [a, b] R is Riemann integrable if
and only if f is continuous -a.e. on [a, b].
Proof. Suppose f is Riemann integrable and let T
k
be a sequence of par-
titions of [a, b] such that T
k
T
k+1
and lim
k
|T
k
| = 0. Set N =
k=1
T
k
. Let
l
k
, u
k
be as in the proof of Theorem 161.1. If x [a, b] N, l(x) = u(x) and > 0,
then there is an integer k such that
u
k
(x) l
k
(x) < .
Let T
k
= x
k
j
m
k
j=0
. Then x (x
k
j1
, x
k
j
) for some j 1, 2, . . . , m
k
. For any
y (x
k
j1
, x
k
j
)
[f(y) f(x)[ u
k
(x) l
k
(x) < .
6.4. L
p
SPACES 163
Thus f is continuous at x. Since l(x) = u(x) for -a.e. x [a, b] and (N) = 0 we
see that f is continuous at -a.e. point of [a, b].
Now suppose f is bounded, N [a, b] with (N) = 0, and f is continuous at
each point of [a, b] N. Let T
k
be any sequence of partitions of [a, b] such that
lim
k
|T
k
| = 0. For each k dene the Lebesgue integrable functions l
k
and u
k
as in the proof of Theorem 161.1. Then
L(T
k
) =
_
[a,b]
l
k
d
_
[a,b]
u
k
d = U(T
k
).
If x [a, b] N and > 0, then there is a > 0 such that
[f(x) f(y)[ < /2
whenever [y x[ < . There is a k
0
such that |T
k
| <
2
whenever k > k
0
. Thus
u
k
(x) l
k
(x) 2 sup[f(x) f(y)[ : [y x[ < <
whenever k > k
0
. Thus
lim
k
(u
k
(x) l
k
(x)) = 0
for each x [a, b] N. By the Dominated Convergence Theorem 160.1 it follows
that
lim
k
(U(T
k
) L(T
k
)) = lim
k
_
[a,b]
(u
k
l
k
) d = 0,
thus showing that f is Riemann integrable.
6.4. L
p
Spaces
The L
p
spaces appear in many applications of analysis. They are also
the prototypical examples of innite dimensional Banach spaces which
will be studied in Chapter VIII. It will be seen that there is a signicant
dierence in these spaces when p = 1 and p > 1.
163.1. Definition. For 1 p and E /, let L
p
(E, /, ) denote the
class of all measurable functions f on E such that |f|
p,E;
< where
|f|
p,E;
:=
_
_
__
E
[f[
p
d
_
1/p
if 1 p <
infM : [f[ M -a.e. on E if p = .
The quantity |f|
p,E;
will be called the L
p
norm of f on E and, for conve-
nience, written |f|
p
when E = X and the measure is clear from the context. The
fact that it is a norm will proved later in this section. We note immediately the
following:
(i) |f|
p
0 for any measurable f.
(ii) |f|
p
= 0 if and only if f = 0 -a.e.
164 6. INTEGRATION
(iii) |cf|
p
= [c[ |f|
p
for any c R.
For convenience, we will write L
p
(X) for the class L
p
(X, /, ). In case X
is a topological space, we let L
p
loc
(X) denote the class of functions f such that
f L
p
(K) for each compact set K X.
The next lemma shows that the classes L
p
(X) are vector spaces or, as is more
commonly said in this context, linear spaces.
164.1. Theorem. Suppose 1 p .
(i) If f, g L
p
(X), then f +g L
p
(X).
(ii) If f L
p
(X) and c R, then cf L
p
(X).
Proof. Assertion (ii) follows from property (iii) of the L
p
norm noted above.
In case p is nite, assertion (i) follows from the inequality
(164.1) [a +b[
p
2
p1
([a[
p
+[b[
p
)
which holds for any a, b R, 1 p < . For p > 1, inequality (164.1) follows from
the fact that t t
p
is a convex function on (0, ) (see Exercise 6.11) and therefore
_
a +b
2
_
p
1
2
(a
p
+b
p
).
In case p = , assertion (i) follows from the triangle inequality [a +b[ [a[+[b[
since if [f(x)[ M -a.e. and [g(x)[ N -a.e. , then [f(x) +g(x)[ M + N -
a.e.
To deduce further properties of the L
p
norms we will use the following arith-
metic inequality.
164.2. Lemma. For a, b 0, 1 < p < and p
= 1
we have
ab
a
p
p
+
b
p
.
Equality holds if and only if a
p
= b
p
.
Proof. Recall that ln(x) is an increasing, strictly concave function on (0, ),
i.e.,
ln(x + (1 )y) > ln(x) + (1 ) ln(y)
for x, y, (0, ), x ,= y and [0, 1].
Set x = a
p
, y = b
p
, and =
1
p
(thus (1 ) =
1
p
) to obtain
ln(
1
p
a
p
+
1
p
b
p
) >
1
p
ln(a
p
) +
1
p
ln(b
p
) = ln(ab).
6.4. L
p
SPACES 165
Clearly equality holds in this inequality if and only if a
p
= b
p
.
For p [1, ] the number p
dened by
1
p
+
1
p
= when p = 1 and p
= 1
when p = .
165.1. Theorem (Holders inequality). If 1 p and f, g are measurable
functions, then
_
X
[fg[ d =
_
X
[f[ [g[ d |f|
p
|g|
p
.
Equality holds if and only if
|f|
p
p
[g[
p
= |g|
p
[f[
p
-a.e..
(Recall the convention that 0 = 0 = 0.)
Proof. In case p = 1
_
[f[ [g[ d |g|
_
[f[ d = |g|
|f|
1
and an analogous inequality holds in case p = .
In case 1 < p < the assertion is clear unless 0 < |f|
p
, |g|
p
< . In this
case, set
f =
f
|f|
p
and g =
g
|g|
p
so that |
f|
p
= 1 and | g|
p
= 1 and apply Lemma 164.2 to obtain
1
|f|
p
|g|
p
_
[f[ [g[ d =
_
[ g[ d
1
p
_
_
_
f
_
_
_
p
p
+
1
p
| g|
p
.
The statement concerning equality follows immediately from the preceding lemma
when 1 < p < . The statement also holds if p = 1, p
= and
0 =
_
X
[[f[ M] [f[ [g[ d =
_
X
[f[ [M [g[] d
where M := |g|
= 1, then
(165.1) |f|
p
= sup
__
fg d : |g|
p
1
_
.
166 6. INTEGRATION
Proof. Suppose f is measurable. If g L
p
1 and
_
fg d =
_
[f[ d = |f|
1
.
Now consider the case 1 < p < . If |f|
p
= 0, then f = 0 a.e. and the desired
inequality is clear. If 0 < |f|
p
< , set
g =
[f[
p/p
sign(f)
|f|
p/p
p
.
Then |g|
p
= 1 and
_
fg d =
1
|f|
p/p
p
_
[f[
p/p
+1
d =
|f|
p
p
|f|
p/p
p
= |f|
p
.
If |f|
p
= , let X
k
k=1
be an increasing sequence of measurable sets such
that (X
k
) < for each k and X =
k=1
X
k
. For each k set
h
k
(x) =
X
k
min([f(x)[ , k)
for x X. Then h
k
L
p
(X), h
k
h
k+1
, and lim
k
h
k
= [f[. By the Monotone
Convergence Theorem lim
k
|h
k
|
p
= . Since we may assume without loss
of generality that |h
k
|
p
> 0 for each k, there exist, by the result just proved,
g
k
L
p
E
M
E
sign(f).
6.4. L
p
SPACES 167
Then |g
M
|
1
= 1 and
_
fg
M
d =
1
(E
M
E)
_
E
M
E
[f[ d M.
Thus
sup
_
fg d : |g|
1
1 |f|
.
167.1. Theorem (Minkowskis inequality). Suppose 1 p and f,g
L
p
(X). Then
|f +g|
p
|f|
p
+|g|
p
.
Proof. The assertion is clear in case p = 1 or p = , so suppose 1 < p < .
Then, applying rst the triangle inequality and then Holders inequality, we obtain
|f +g|
p
p
=
_
[f +g[
p
d =
_
[f +g[
p1
[f +g[
_
[f +g[
p1
[f[ d +
_
[f +g[
p1
[g[ d
__
([f +g[
p1
)
p
_
1/p
__
[f[
p
d
_
1/p
+
__
([f +g[
p1
)
p
_
1/p
__
[g[
p
d
_
1/p
__
[f +g[
p
_
(p1)/p
__
[f[
p
d
_
1/p
+
__
[f +g[
p
_
(p1)/p
__
[g[
p
d
_
1/p
= |f +g|
p1
p
|f|
p
+|f +g|
p1
p
|g|
p
= |f +g|
p1
p
(|f|
p
+|g|
p
).
The assertion is clear if |f +g|
p
= 0. Otherwise we divide by |f +g|
p1
p
to
obtain
|f +g|
p
|f|
p
+|g|
p
.
As a consequence of Theorem 167.1 and the remarks following Denition 163.1
we can say that for 1 p the spaces L
p
(X) are, in the terminology of Chapter
8, normed linear spaces provided we agree to identify functions that are equal
-a.e. The norm ||
p
induces a metric on L
p
(X) if we dene
(f, g) := |f g|
p
for f, g L
p
(X) and agree to interpret the statement f = g as f = g -a.e.
168 6. INTEGRATION
167.2. Definitions. A sequence f
k
k=1
is a Cauchy sequence in L
p
(X) if
given any > 0 there is a positive integer N such that
|f
k
f
m
|
p
<
whenever k, m > N. The sequence f
k
k=1
converges in L
p
(X) to f L
p
(X) if
lim
k
|f
k
f|
p
= 0.
168.1. Theorem. If 1 p , then L
p
(X) is a complete metric space under
the metric , i.e., if f
k
k=1
is a Cauchy sequence in L
p
(X), then there is an
f L
p
(X) such that
lim
k
|f
k
f|
p
= 0.
Proof. Suppose f
k
k=1
is a Cauchy sequence in L
p
(X). There an integer N
such that |f
k
f
m
|
p
< 1 whenever k, m N. By Minkowskis inequality
|f
k
|
p
|f
N
|
p
+|f
k
f
N
|
p
|f
N
|
p
+ 1
whenever k N. Thus the sequence |f
k
|
p
k=1
is bounded.
Consider the case 1 p < . For any > 0, let A
k,m
:= x : [f
k
(x) f
m
(x)[
. Then,
_
A
k,m
[f
k
f
m
[
p
d
p
(A
k,m
);
that is,
p
(x : [f
k
(x) f
m
(x)[ ) |f
k
f
m
|
p
p
.
Thus f
k
is fundamental in measure, and consequently by Exercise 5.13 and Theo-
rem 146.2, there exists a subsequence f
k
j
j=1
that converges -a.e. to a measurable
function f. By Fatous lemma,
|f|
p
p
=
_
[f[
p
d liminf
j
_
f
k
j
p
d < .
Thus f L
p
(X).
Let > 0 and let M be such that |f
k
f
m
|
p
< whenever k, m > M. Using
Fatous lemma again we see
|f
k
f|
p
p
=
_
[f
k
f[
p
d liminf
j
_
f
k
f
k
j
p
d <
p
whenever k > M. Thus f
k
converges to f in L
p
(X).
The case p = is left as Exercise 6.24.
As a consequence of Theorem 168.1 we see for 1 p , that L
p
(X) is a
Banach space; i.e., a normed linear space that is complete with respect to the
metric induced by the norm.
6.4. L
p
SPACES 169
Here we include a useful result relating norm convergence in L
p
and pointwise
convergence.
169.0. Theorem (Vitalis Convergence Theorem). Suppose f
k
, f L
p
(X),
1 p < . Then |f
k
f|
p
0 if the following three conditions hold:
(i) f
k
f -a.e.
(ii) For each > 0, there exists a measurable set E such that (E) < and
_
E
[f
k
[
p
d < , for all k N
(iii) For each > 0, there exists > 0 such that (E) < implies
_
E
[f
k
[
p
d < for all k N.
Conversely, if |f
k
f|
p
0, then (ii) and (iii) hold. Furthermore, (i) holds for a
subsequence.
Proof. Assume the three conditions hold. Choose > 0 and let > 0 be the
corresponding number given by (iii). Condition (ii) provides a measurable set E
with (E) < such that
_
E
[f
k
[
p
d <
for all positive integers k. Since (E) < , we can apply Egoros Theorem to
obtain a measurable set B E with (EB) < such that f
k
converges uniformly
to f on B. Now write
_
X
[f
k
f[
p
d =
_
B
[f
k
f[
p
d
+
_
EB
[f
k
f[
p
d +
_
E
[f
k
f[
p
d.
The rst integral on the right can be made arbitrarily small for large k, because of
the uniform convergence of f
k
to f on B. The second and third integrals will be
estimated with the help of the inequality
[f
k
f[
p
2
p1
([f
k
[
p
+[f[
p
),
see (164.1), p. 164. From (iii) we have
_
EB
[f
k
[
p
< for all k N and then Fatous
Lemma shows that
_
EB
[f[
p
< as well. The third integral can be handled in a
similar way using (ii). Thus, it follows that |f
k
f|
p
0.
Now suppose |f
k
f|
p
0. Then for each > 0 there exists a positive integer
k
0
such that |f
k
f|
p
< /2 for k > k
0
. With the help of Exercise 6.25, there
exist measurable sets A and B of nite measure such that
_
A
[f[
p
d < (/2)
p
and
_
B
[f
k
[
p
d < ()
p
for k = 1, 2, . . . , k
0
.
170 6. INTEGRATION
Minkowskis inequality implies that
|f
k
|
p,
A
|f
k
f|
p,
A
+|f|
p,
A
< for k > k
0
.
Then set E = A B to obtain the necessity of (ii).
Similar reasoning establishes the necessity of (iii).
According to Exercise 6.26 convergence in L
p
implies convergence in measure.
Hence, (i) holds for a subsequence.
Finally, we conclude this section by considering how L
p
(X) compares with
L
q
(X) for 1 p < q . For example, let X = [0, 1] and let := . In
this case it is easy to see that L
q
L
p
, for if f L
q
then [f(x)[
q
[f(x)[
p
if
x A := x : [f(x)[ 1. Therefore,
_
A
[f[
p
d
_
A
[f[
q
d <
while
_
[0,1]\A
[f[
p
d 1 ([0, 1]) < .
This observation extends to a more general situation via Holders inequality.
170.1. Theorem. If (X) < and 1 p q , then L
q
() L
p
() and
|f|
q;
(X)
1
p
1
q
|f|
p;
.
Proof. If q = , then result is immediate:
|f|
p
p
=
_
X
[f[
p
d |f|
p
_
X
1 d = |f|
p
(X) < .
If q < Holders inequality with conjugate exponents q/p and q/(q p) implies
that
|f|
p
p
=
_
X
[f[
p
1 d |[f[
p
|
q/p
|1|
q/(qp)
= |f|
p
q
(X)
(qp)/q
< .
170.2. Theorem. If 0 < p < q < r , then L
p
() L
r
() L
q
() and
|f|
q;
|f|
p;
|f|
1
r;
where is dened by he equation
1
q
=
p
+
1
r
.
6.5. SIGNED MEASURES 171
Proof. If r < , use Holders inequality with conjugate indices p/q and
r/(1 )q169g to obtain
_
X
[f[
q
=
_
X
[f[
q
[f[
(1)q
_
_
_[f[
q
_
_
_
p/q
_
_
_[f[
(1)q
_
_
_
r/(1)q
=
__
X
[f[
p
_
q/p
__
X
[f[
r
_
(1)q/r
= |f|
q
p
|f|
(1)q
r
.
We obtain the desired result by taking q
th
roots of both sides.
When r = , we have
_
X
[f[
q
|f|
qp
_
X
[f[
p
and so
|f|
q
|f|
p/q
p
|f|
1(p/q)
= |f|
p
|f|
1
.
6.5. Signed Measures
We develop the basic properties of countably additive set functions of ar-
bitrary sign, or signed measures. In particular, we establish the decom-
position theorems of Hahn and Jordan, which show that signed measures
and (positive) measures are closely related.
Let (X, /, ) be a measure space. Suppose f is measurable, at least one of f
+
,
f
k=1
is a disjoint sequence of measurable sets then
(
k=1
E
k
) =
k=1
(E
k
)
where the series on the right either converges absolutely or diverges to
(see Exercise 6.36),
(iv) (E) = 0 whenever (E) = 0.
In Section 6.6 we will show that the properties (i)(iv) characterize set functions of
the type (171.1).
172 6. INTEGRATION
171.1. Remark. An extended real-valued function dened on / is a signed
measure if it satises properties (i)(iii) above. If in addition it satises (iv) the
signed measure is said to be absolutely continuous with respect to , written
. In some contexts, we will underscore that a measure is not a signed
measure by saying that it is a positive measure. In other words, a positive
measure is merely a measure in the sense dened in Denition 107.1, p. 107.
172.1. Definition. Let be a signed measure on /. A set A / is a
positive set for if (E) 0 for each measurable subset E of A. A set B / is
a negative set for if (E) 0 for each measurable subset E of B. A set C /
is a null set for if (E) = 0 for each measurable subset E of C.
Note that any measurable subset of a positive set for is also a positive set
and that analogous statements hold for negative sets and null sets. It follows that
any countable union of positive sets is a positive set. To see this suppose P
k
k=1
is a sequence of positive sets. Then there exist disjoint measurable sets P
k
P
k
such that P :=
k=1
P
k
=
k=1
P
k
(Lemma 82.1). If E is a measurable subset of P,
then
(E) =
k=1
(E P
k
) 0
since each P
k
is positive for .
It is important to observe the distinction between measurable sets E such that
(E) = 0 and null sets for . If E is a null set for , then (E) = 0 but the converse
is not generally true.
172.2. Theorem. If is a signed measure on /, E / and 0 < (E) <
then E contains a positive set A with (A) > 0.
Proof. If E is positive then the conclusion holds for A = E. Assume E is not
positive, and inductively construct a sequence of sets E
k
as follows. Set
c
1
:= inf(B) : B /, B E < 0.
There exists a measurable set E
1
E such that
(E
1
) <
1
2
max(c
1
, 1) < 0.
For k 1, if E
k
j=1
E
j
is not positive, then
c
k+1
:= inf(B) : B /, B E
k
j=1
E
j
< 0
6.5. SIGNED MEASURES 173
and there is a measurable set E
k+1
E
k
j=1
E
j
such that
(E
k+1
) <
1
2
max(c
k+1
, 1) < 0.
Note that if c
k
= , then (E
k
) < 1/2.
If at any stage E
k
j=1
E
j
is a positive set, let A = E
k
j=1
E
j
and observe
(A) = (E)
k
j=1
(E
j
) > (E) > 0.
Otherwise set A = E
k=1
E
k
and observe
(E) = (A) +
k=1
(E
k
).
Since (E) > 0 we have (A) > 0. Since (E) is nite the series converges abso-
lutely, (E
k
) 0 and therefore c
k
0 as k . If B is a measurable subset of
A, then B
E
k
= for k = 1, 2, . . . and hence
(B) c
k
for k = 1, 2, . . . . Thus (B) 0. This shows that A is a positive set and the lemma
is proved.
173.1. Theorem (Hahn Decomposition). If is a signed measure on /, then
there exist disjoint sets P and N such that P is a positive set, N is a negative set,
and X = P N.
Proof. By considering in place of if necessary, we may assume (E) <
for each E /. Set
:= sup(A) : A is a positive set for .
Since is a positive set, 0. Let A
k
k=1
be a sequence of positive sets for
which
lim
k
(A
k
) = .
Set P =
k=1
A
k
. Then P is positive and hence (P) . On the other hand each
P A
k
is positive and hence
(P) = (A
k
) +(P A
k
) (A
k
).
Thus (P) = < .
174 6. INTEGRATION
Set N = X P. We have only to show that N is negative. Suppose B is a
measurable subset of N. If (B) > 0, then by Lemma 172.2 B must contain a
positive set B
such that (B
P is positive and
(B
P) = (B
) +(P) > ,
contradicting the choice of .
Note that the Hahn decomposition above is not unique if has a nonempty
null set.
The following denition describes a relation between measures that is the an-
tithesis of absolute continuity.
174.1. Definition. Two measures
1
and
2
dened on a measure space
(X, /) are said to be mutually singular (written
1
2
) if there exists a
measurable set E such that
1
(E) = 0 =
2
(X E).
174.2. Theorem (Jordan Decomposition). If is a signed measure on / then
there exists a unique pair of mutually singular measures
+
and
, at least one of
which is nite, such that
(E) =
+
(E)
(E)
for each E /.
Proof. Let P N be a Hahn decomposition of X with P N = , P positive
and N negative for . Set
+
(E) = (E P)
(E) = (E N)
for E /. Clearly
+
and
. The
measures
+
and
(X N). That
at least one of the measures
+
,
+
(X) = (P) and
2
and A / such that
1
(X A) = 0 =
2
(A), then
1
(X P) =
1
((X A) P)
= ((X A) P) +
2
((X A) P)
=
(X A) 0.
6.5. SIGNED MEASURES 175
Thus
1
(X P) = 0. Similarly
2
(P) = 0. For any E / we have
+
(E) = (E P) =
1
(E P)
2
(E P) =
1
(E).
Analogously
=
2
.
Note that if is the signed measure dened by (171.1) then the sets P = x :
f(x) > 0 and N = x : f(x) 0 form a Hahn decomposition of X for and
+
(E) =
_
EP
f d =
_
E
f
+
d
and
(E) =
_
EN
f d =
_
E
f
d
for each E /.
175.1. Notation. We will write [[ =
+
+
.
We conclude this section by examining alternate characterizations of absolutely
continuous measures. We leave it as an exercise to prove that the following three
conditions are equivalent:
(175.1)
(i)
(ii) [[
(iii)
+
and
.
175.2. Theorem. Let be a nite signed measure and a positive measure
on (X, /). Then if and only if for every > 0 there exists > 0 such that
[(E)[ < whenever (E) < .
Proof. Because of condition (ii) in (175.1) and the fact that [(E)[ [[ (E),
we may assume that is a nite positive measure. Since the , condition is
easily seen to imply that , we will only prove the converse. Proceeding by
contradiction, suppose then there exists > 0 and a sequence of measurable sets
E
k
such that (E
k
) < 2
k
and (E
k
) > for all k. Set
F
m
=
k=m
E
k
and
F =
m=1
F
m
.
Then, (F
m
) < 2
1m
, so (F) = 0. But (F
m
) for each m and, since is
nite, we have
(F) = lim
m
(F
m
) ,
thus reaching a contradiction.
176 6. INTEGRATION
6.6. The Radon-Nikodym Theorem
If f is an integrable function on the measure space (X, M, ), then the
signed measure
(E) =
E
f d
dened for all E M is absolutely continuous with respect to . The
Radon-Nikodym theorem states that essentially every signed measure ,
absolutely continuous with respect to , is of this form. The proof of
Theorem 176.1 below is due to A. Schep, [?].
176.1. Theorem. Suppose (X, /, ) is a nite measure space and is a mea-
sure on (X, /) with the property
(E) (E)
for each E /. Then there is a measurable function f : X [0, 1] such that
(176.1) (E) =
_
E
f d, for each E /.
More generally, if g is a nonnegative measurable function on X, then
_
X
g d =
_
X
gf d.
Proof. Let
H :=
_
f : f measurable, 0 f 1,
_
E
f d (E) for all E /
_
.
and let
M := sup
__
X
f d : f H
_
.
Then, there exist functions f
k
H such that
_
X
f
k
d > M k
1
.
Observe that we may assume 0 f
1
f
2
. . . because if f, g H, then so is
maxf, g in view of the following:
_
E
maxf, g d =
_
EA
maxf, g d +
_
E(X\A)
maxf, g d
=
_
EA
f d +
_
X\A
g d
(E A) +(E (X A)) = (E),
where A := x : f(x) g(x). Therefore, since f
k
is an increasing sequence, the
limit below exists:
f
(x) := lim
k
f
k
(x).
6.6. THE RADON-NIKODYM THEOREM 177
Note that f
d = lim
k
_
X
f
k
= M
for each E / and
_
E
f
k
d (E).
for each k and E /. So f
H.
The proof of the theorem will be concluded by showing that (176.1) is satised
by taking f as f
d < (E)
for some E /. Let
E
0
= x E : f
(x) < 1
E
1
= x E : f
(x) = 1.
Then
(E) = (E
0
) +(E
1
)
>
_
E
f
d
=
_
E
0
f
d +(E
1
)
_
E
0
f
d +(E
1
),
which implies
(E
0
) >
_
E
0
f
d.
Let
E
0
d < (E
0
)
and let F
k
:= x E
0
: f
< 1 on E
0
, we have
k=1
F
k
= E
0
and therefore that (F
k
) (E
0
).
Furthermore, it follows from (177.2) that
_
E
0
f
E
0
d < (E
0
)
178 6. INTEGRATION
for some > 0. Therefore, since [f
F
k
]
F
k
[f
E
0
]
E
0
, there exists k
such that
_
F
k
f
F
k
d
=
_
X
[f
F
k
]
F
k
d
_
E
0
f
E
0
d by Monotone Convergence Theorem
< (E
0
)
< (F
k
) for all k k
.
For all such k k
and := min(
F
k
H
The validity of this claim would imply that
_
f
F
k
d = M + (F
k
) >
M, contradicting the denition of M which would mean that our contradiction
hypothesis, (177.1), is false, thus establishing our theorem.
So, to nish the proof, it suces to prove (178.1) for any k k
, which will
remain xed throughout the remainder of the proof. For this, rst note that 0
f
F
k
1. To show that
_
E
f
F
k
d (E) for all E /,
we proceed by contradiction; if not, there would exist a measurable set G X such
that
(G F
k
) +
_
GF
k
f
F
k
d
_
G\F
k
f
d +
_
GF
k
f
F
k
d
=
_
G\F
k
f
F
k
d +
_
GF
k
f
F
k
d
=
_
G
f
F
k
d > (G).
This implies
(178.2)
_
GF
k
(f
F
k
) d > (G) (G F
k
) = (G F
k
).
Hence, we may assume G F
k
.
Let T
1
be the collection of all measurable sets G F
k
such that (178.2) holds.
Dene
1
:= sup(G) : G T
1
and let G
1
T
1
be such that (G
1
) >
1
1.
Similarly, let T
2
be the collection of all measurable sets G F
k
G
1
such that
6.6. THE RADON-NIKODYM THEOREM 179
*) holds for G. Dene
2
:= sup(G) : G T
2
and let G
2
T
2
be such that
(G
2
) >
2
1
2
2
. Proceeding inductively, we obtain a decreasing sequence
k
and
disjoint measurable sets G
j
where (G
j
) >
j
1
j
2
. Observe that
j
0, for if
j
a > 0, then (G
j
) a. Since
j=1
1
j
2
< this would imply
(
G
j
) =
j=1
(G
j
) >
j=1
_
j
1
j
2
_
= ,
contrary to the niteness of . This implies
(F
k
j=1
G
j
) = 0
for if not, there would be two possibilities:
(i) there would exist a set T F
k
j=1
G
j
of positive measure for which
(178.2) would hold with T replacing G. Since (T) > 0, there would exist
k
such
that
k
< (T)
and since
T F
k
j=1
G
j
F
k
j1
i=1
G
j
this would contradict the denition of
j
. Hence,
(F
k
) >
_
F
k
f
F
k
d
=
j
_
G
j
f
F
k
d
>
j
(G
j
) = (F
k
),
which is impossible and therefore (i) cannot occur.
(ii) If there were no set T as in (i), then F
k
j
G
j
could not satisfy (178.1) and
thus
(179.1)
_
F
k
\
j
G
j=1
f
F
k
d (F
k
j=1
G
j
).
With S := F
k
j=1
G
j
we have
_
S
f
F
k
d (S).
Since
_
G
j
f
F
k
> (G
j
)
180 6. INTEGRATION
for each j N, it follows that F
k
S must also satisfy (178.1), which contradicts
(179.1). Hence, both (i) and (ii) do not occur and thus we conclude that
(F
k
j=1
G
j
) = 0,
as desired.
180.1. Notation. The function f in (176.1) (and also in (180.2) below) is
called the Radon-Nikodym derivative of with respect to and is denoted by
f :=
d
d
.
The previous theorem yields the notationally convenient result
(180.1)
_
X
g d =
_
X
g
d
d
d.
180.2. Theorem (Radon-Nikodym). If (X, /, ) is a -nite measure space
and is a -nite signed measure on / that is absolutely continuous with respect to
, then there exists a measurable function f such that either f
+
or f
is integrable
and
(180.2) (E) =
_
E
f d
for each E /.
Proof. We rst assume, temporarily, that and are nite measures. Re-
ferring to Theorem 176.1, there exist Radon-Nikodym derivatives
f
:=
d
d( +)
and f
:=
d
d( +)
.
Dene A = X f
(x) = 0. Then
(B) =
_
B
f
d( +) = 0
and therefore, (B) = 0 since . Now dene
f(x) =
_
_
_
f
(x)
f
(x)
if x A
0 if x B.
If E is a measurable subset of A, then
(E) =
_
E
f
d( +) =
_
E
f f
d( +) =
_
E
f d
by (180.1). Since both and are 0 on B, we have
(E) =
_
E
f d
for all measurable E.
6.6. THE RADON-NIKODYM THEOREM 181
Next, consider the case where and are -nite measures. There is a sequence
of disjoint measurable sets X
k
k=1
such that X =
k=1
X
k
and both (X
k
) and
(X
k
) are nite for each k. Set
k
:= X
k
,
k
:= X
k
for k = 1, 2, . . . . Clearly
k
and
k
are nite measures on / and
k
k
. Thus there exist nonnegative
measurable functions f
k
such that for any E /
(E X
k
) =
k
(E) =
_
E
f
k
d
k
=
_
EX
k
f
k
d.
It is clear that we may assume f
k
= 0 on XX
k
. Set f :=
k=1
f
k
. Then for any
E /
(E) =
k=1
(E X
k
) =
k=1
_
EX
k
f d =
_
E
f d.
Finally suppose that is a signed measure and let =
+
be the Jordan
decomposition of . Since the measures are mutually singular there is a measurable
set P such that
+
(X P) = 0 =
+
(E) = (E P) = 0
(E) = (E P) = 0
since (EP) +(EP) = (E) = 0. Thus
+
and
such that
(E) =
_
E
f
d
for each E /. Since at least one of the measures
is -integrable. Set f = f
+
f
. In view of Theorem
157.1, p. 157,
(E) =
_
E
f
+
d
_
E
f
d =
_
E
f d
for each E /.
An immediate consequence of this result is the following.
181.1. Theorem (Lebesgue Decomposition). Let and be -nite measures
dened on the measure space (X, /). Then there is a decomposition of such that
=
0
+
1
where
0
and
1
. The measures
0
and
1
are unique.
Proof. We employ the same device as in the proof of the preceding theorem
by considering the Radon-Nikodym derivatives
f
:=
d
d( +)
and f
:=
d
d( +)
.
182 6. INTEGRATION
Dene A = X f
0
(E) := (E B) and
1
(E) := (E A) =
_
EA
f
d
provide our desired decomposition. First, note that =
0
+
1
. Next, we have
0
(A) = 0 and so
0
. Finally, to show that
1
, consider E with (E) = 0.
Then
0 = (E) =
_
E
f
d =
_
AE
f
d.
Thus, f
F
_
f
|f|
p
_
|F|
i.e.,
[F(f)[ |F| |f|
p
whenever f L
p
(X). In particular for any f, g L
p
(X)
[F(f g)[ |F| |f g|
p
and hence F is uniformly continuous on L
p
(X).
6.7. THE DUAL OF L
p
183
On the other hand if F is continuous at 0, then there exists a > 0 such that
[F(f)[ 1
whenever |f|
p
. Thus if f L
p
(X) with |f|
p
> 0, then
[F(f)[ =
|f|
p
F
_
|f|
p
f
_
|f|
p
whence |F|
1
.
183.1. Theorem. If 1 p ,
1
p
+
1
p
= 1 and g L
p
(X), then
F(f) =
_
fg d
denes a bounded linear functional on L
p
(X) with
|F| = |g|
p
.
Proof. That F is a bounded linear functional on L
p
(X) follows immediately
from Holders inequality and the elementary properties of the integral. The rest of
the assertion follows from Theorem 165.2 since
|F| = sup
__
fg d : |f|
p
1
_
= |g|
p
.
Note that while the hypotheses of Theorem 165.2, p. 165, include a -niteness con-
dition, that assumption is not needed to establish (25) if the function is integrable.
Thats the situation we have here since it is assumed that g L
p
.
The next theorem shows that all bounded linear functionals on L
p
(X) (1
p < ) are of this form.
183.2. Theorem. If 1 < p < and F is a bounded linear functional on
L
p
(X), then there is a g L
p
(X), (
1
p
+
1
p
= 1) such that
(183.1) F(f) =
_
fg d
for all f L
p
(X). Moreover |g|
p
= |F| and the function g is unique in the
sense that if (183.1) holds with g L
p
E
)
for E /. Suppose E
k
k=1
is a sequence of disjoint measurable sets and let
E :=
k=1
E
k
. Then for any positive integer N
184 6. INTEGRATION
(E)
N
k=1
(E
k
)
F(
k=1
E
k
)
F(
k=N+1
E
k
)
|F| ((
k=N+1
E
k
))
1
p
and (
k=N+1
E
k
) =
k=N+1
(E
k
) 0 as N since
(E) =
k=1
(E
k
) < .
Thus,
(E) =
k=1
(E
k
)
and since the same result holds for any rearrangement of the sequence E
k
k=1
,
the series converges absolutely. It follows that is a signed measure and since
[(E)[ |F| ((E))
1
p
we see that . By the Radon-Nikodym Theorem there
is a g L
1
(X) such that
F(
E
) = (E) =
_
E
g d
for each E /. From the linearity of both F and the integral, it is clear that
(184.1) F(f) =
_
fg d
whenever f is a simple function. Suppose M > 0 and set E
M
:= x : g(x) > M.
Then
M(E
M
)
_
E
M
g d = F(
E
M
) |F| (E
M
).
Thus (E
M
) = 0 if M > |F|, from which it follows that |g
+
|
|F|. Similarly
|g
|f|
1
|g|
k=1
be an increasing sequence of nonnegative simple functions such
that lim
k
h
k
= [g[. Set g
k
= h
p
1
k
sign(g). Then
(185.1)
|h
k
|
p
=
_
[h
k
[
p
d
_
g
k
g d = F(g
k
)
|F| |g
k
|
p
= |F|
__
h
p
k
d
_1
p
= |F| |h
k
|
p
p
p
.
We wish to conclude that g L
p
(X), 1 p < .
Now let f L
p
(X) and let f
k
be a sequence of simple functions such that
|f f
k
|
p
0 as k (see Theorem 147.1, p.147). Then
F(f)
_
fg d
[F(f f
k
)[ +
_
(f
k
f)g d
|F| |f f
k
|
p
+|f f
k
|
p
|g|
p
2 |F| |f f
k
|
p
for all k; whence,
F(f) =
_
fg d
for all f L
p
(X). By Theorem 165.2, p. 165, we have |g|
p
= |F|. Thus, using
step 1 also, we conclude that the proof is complete under the assumptions that
(X) < , 1 p < .
Step 3. Assume (X) and 1 p < .
Suppose Y / is -nite. Let Y
k
be an increasing sequence of measurable
sets such that (Y
k
) < for each k and such that Y =
k=1
Y
k
. Then, from Steps
1 and 2 above (see also Exercise 6.9), for each k there is a measurable function g
k
such that
_
_
g
k
Y
k
_
_
p
|F| and
F(f
Y
k
) =
_
f
Y
k
g
k
d
186 6. INTEGRATION
for each f L
p
(X). We may assume g
k
= 0 on Y Y
k
. If k < m then
F(f
Y
k
) =
_
fg
m
Y
k
d
for each f L
p
(X). Thus
_
f(g
k
g
m
Y
k
) d = 0
for each f L
p
(X). By Theorem 165.2, p. 165, this implies that g
k
= g
m
-a.e. on
Y
k
. Thus g
k
converges -a.e. to a measurable function g and by Fatous Lemma
(186.1) |g|
p
liminf
k
|g
k
|
p
|F| ,
which shows that g L
p
.
Fix f L
p
(X) and set f
k
:= f
Y
k
. Then f
k
converges to f
Y
and [f f
k
[
2 [f[. By the Dominated Convergence Theorem |(f
Y
f
k
)|
p
0 as k .
Thus, since g
k
= g -a.e. on Y
k
,
F(f
Y
)
_
fg d
[F(f
Y
f
k
)[ +
_
[f f
k
[ [g[ d
2 |F| |f
Y
f
k
|
p
and therefore
(186.2) F(f
Y
) =
_
fg d
for each f L
p
(X).
If is -nite, we may set Y = X and deduce that
F(f) =
_
fg d
whenever f L
p
(X). Thus, in view of Theorem 183.1,
|F| = sup[F(f)[ : |f|
q
= 1 = |g|
p
When 1 < p < and is not assumed to be -nite, we will conclude the
proof by making a judicious choice for Y in (186.2) so that
(186.3) F(f) = F(f
Y
) for each f L
p
(X) and |g|
p
= |F|.
For each positive integer k there is h
k
L
p
(X) such that |h
k
|
p
1 and
|F|
1
k
< [F(h
k
)[ .
Set
Y =
k=1
x : h
k
(x) ,= 0.
6.7. THE DUAL OF L
p
187
Then Y is a measurable, -nite subset of X and thus, by Step 2, there is a
g L
p
Y
) =
_
fg d
for each f L
p
(X). Since for each k
|F|
1
k
< F(h
k
) =
_
h
k
g d |g|
p
,
we see that |g|
p
Y
) = sup
f
p
1
_
fg d = |g|
p
,
which establishes the second part of (186.3),
To establish the rst part of (186.3), with the help of (186.4), it suces to
show that F(f) = F(f
Y
) for each f L
p
(X). By contradiction, suppose there is
a function f
0
L
p
(X) such that F(f
0
) ,= F(f
0
Y
). Set Y
0
= x : f
0
(x) ,= 0 Y .
Then, since Y
0
is -nite, there is a g
0
L
p
Y
0
) =
_
fg
0
d
for each f L
p
(X). Since g and g
0
are non-zero on disjoint sets, note that
|g +g
0
|
p
= |g|
p
+|g
0
|
p
and that |g
0
|
p
> 0 since
_
f
0
g
0
d = F(f
0
[1
Y
]) = F(f
0
) F(f
0
Y
) ,= 0.
Moreover, since
F(f
Y Y
0
) =
_
f(g +g
0
) d
for each f L
p
(X), we see that |g +g
0
|
p
|F|. Thus
|F|
p
= |g|
p
< |g|
p
+|g
0
|
p
= |g +g
0
|
p
|F|
p
.
This contradiction implies that
F(f) =
_
fg d
for each f L
p
(X), which establishes the rst part of (186.3), as desired.
188 6. INTEGRATION
If g L
p
(X) is such that
_
f(g g) d = 0
for all f L
p
(X), then by Corollary (165.2) |g g|
p
= 0 and thus g = g -
a.e. thus establishing the uniqueness of g.
6.8. Product Measures and Fubinis Theorem
In this section we introduce product measures and prove Fubinis theo-
rem, which generalizes the notion of iterated integration of Riemannian
calculus.
Let (X, /
X
, ) and (Y, /
Y
, ) be two complete measure spaces. In order to
dene the product of and we rst dene an outer measure on X Y in terms
of and .
188.1. Definition. For each S X Y set
(188.1) (S) = inf
_
_
_
j=1
(A
j
)(B
j
)
_
_
_
where the inmum is taken over all sequences A
j
B
j
j=1
such that A
j
/
X
,
B
j
/
Y
for each j and S
j=1
(A
j
B
j
).
188.2. Theorem. The set function is an outer measure on X Y .
Proof. It is immediate from the denition that 0 and () = 0. To see
that is countably subadditive suppose S
k=1
S
k
and assume that (S
k
) <
for each k. Fix > 0. Then for each k there is a sequence A
k
j
B
k
j
j=1
with
A
k
j
/
X
and B
k
j
/
Y
for each j such that
S
k
j=1
(A
k
j
B
k
j
)
and
j=1
(A
k
j
)(B
k
j
) < (S
k
) +
2
k
.
Thus
(S)
k=1
j=1
(A
k
j
)(B
k
j
)
k=1
((S
k
) +
2
k
)
k=1
(S
k
) +
6.8. PRODUCT MEASURES AND FUBINIS THEOREM 189
for any > 0.
Since is an outer measure, we know its measurable sets form a -algebra (See
Corollary 83.2, p. 83) which we denote by /
XY
. Also, we denote by the
restriction of to /
XY
. The main objective of this section is to show that
may appropriately be called the product measure corresponding to and , and
that the integral of a function over X Y with respect to can be computed
by iterated integration. This is the thrust of the next result.
189.1. Theorem (Fubinis Theorem). Suppose (X, /
X
, ) and (Y, /
Y
, ) are
complete measure spaces.
(i) If A /
X
and B /
Y
, then AB /
XY
and
( )(AB) = (A)(B).
(ii) If S /
XY
and S is -nite with respect to then
S
y
= x : (x, y) S /
X
, for a.e. y Y,
S
x
= y : (x, y) S /
Y
, for a.e. x X,
y (S
y
) is /
Y
-measurable,
x (S
x
) is /
X
-measurable,
( )(S) =
_
X
(S
x
) d(x) =
_
X
__
Y
S
(x, y)d(y)
_
d(x)
=
_
Y
(S
y
)d(y) =
_
Y
__
X
S
(x, y) d(x)
_
d(y).
190 6. INTEGRATION
(iii) If f L
1
(X Y, /
XY
, ), then
y f(x, y) is -integrable for -a.e. x X,
x f(x, y) is -integrable for -a.e. y Y ,
x
_
Y
f(x, y)d(y) is -integrable,
y
_
X
f(x, y) d(x) is -integrable,
_
XY
fd( ) =
_
X
__
Y
f(x, y)d(y)
_
d(x)
=
_
Y
__
X
f(x, y) d(x)
_
d(y).
Proof. Let T denote the collection of all subsets S of X Y such that
S
x
:= y : (x, y) S /
Y
for -a.e. x X
and that the function
x (S
x
) is /
X
-measurable.
For S T set
(S) =
_
X
(S
x
) d(x) =
_
X
__
Y
S
(x, y)d(y)
_
d(x).
Another words, T is precisely the family of sets that makes is possible to dene .
Proof of (i). First, note that is monotone on T. Next observe that if
j=1
S
j
is
a countable union of disjoint S
j
T, then clearly
j=1
S
j
T and the Monotone
Convergence Theorem implies
(190.1)
j=1
(S
j
) = (
j=1
S
j
); hence
i=1
S
i
T.
Finally, if S
1
S
2
are members of T then
(190.2)
j=1
S
j
T
and if (S
1
) < , then Lebesgues Dominated Convergence Theorem yields
(190.3) lim
j
(S
j
) =
_
j=1
S
j
_
; hence
j=1
S
j
T.
6.8. PRODUCT MEASURES AND FUBINIS THEOREM 191
Set
P
0
= AB : A /
X
and B /
Y
P
1
=
j=1
S
j
: S
j
P
0
for j = 1, 2, . . .
P
2
=
j=1
S
j
: S
j
P
1
for j = 1, 2, . . .
Note that if A /
X
and B /
Y
, then AB T and
(191.1) (AB) = (A)(B)
and thus P
0
T. If A
1
B
1
, A
2
B
2
X Y , then
(191.2) (A
1
B
1
) (A
2
B
2
) = (A
1
A
2
) (B
1
B
2
)
and
(191.3) (A
1
B
1
) (A
2
B
2
) = ((A
1
A
2
) B
1
) ((A
1
A
2
) (B
1
B
2
)).
It follows from Lemma 82.1, (p.82), (191.2) and (191.3) that each member of P
1
can be written as a countable disjoint union of members of P
0
and since T is closed
under countable disjoint unions, we have P
1
T. It also follows from (191.2) that
any nite intersection of members of P
1
is also a member of P
1
. Therefore, from
(190.2), P
2
T. In summary, we have
(191.4) P
0
, P
1
, P
2
T.
Suppose S X Y , A
j
/
X
, B
j
/
Y
and S R =
j=1
(A
j
B
j
).
Using (191.1) and that R P
1
T, we obtain
(R)
j=1
(A
j
B
j
) =
j=1
(A
j
)(B
j
).
Thus, by the denition of , (188.1),
(191.5) inf(R) : S R P
1
(S).
To establish the opposite inequality, note that if S R =
j=1
(A
j
B
j
) where the
sets A
j
B
j
are disjoint, then referring to (190.1)
(S)
j=1
(A
j
)(B
j
) = (R)
and consequently, with (191.5), we have
(191.6) (S) = inf(R) : S R P
1
192 6. INTEGRATION
for each S X Y . If A /
X
and B /
Y
, then A B P
0
T and hence,
for any R P
1
with AB R.
(AB) (A)(B) by (188.1)
= (AB) by (191.1)
(R) because is monotone.
Therefore, by (191.6) and (191.1)
(192.1) (AB) = (AB) = (A)(B).
Moreover if T R P
1
T, then using the additivity of (see(190.1)) it follows
that
(T (AB)) +(T (AB))
(R (AB)) +(R (AB)) by (191.6)
= (R) since is additive.
In view of (191.6) we see that (T (AB)) +(T (AB)) for any T X Y
and thus, (see Denition 79.1, p. 79) that AB is -measurable; that is, AB
/
XY
. Thus assertion (i) is proved.
Proof of (ii). Suppose S X Y and (S) < . Then there is a sequence
R
j
P
1
such that S R
j
for each j and
(192.2) (S) = lim
j
(R
j
).
Set
R =
j=1
R
j
P
2
.
Since P
2
T and (S) < the Dominated Convergence Theorem implies
(192.3) (R) = lim
m
_
m
j=1
R
j
_
.
Thus, since S R
m
j=1
R
j
P
2
for each nite m, by (192.3) we have
(S) lim
m
_
m
j=1
R
j
_
= (R) lim
m
(R
m
) = (S),
which implies that
(192.4) for each S X Y there is R P
2
such that S R and (S) = (R).
We now are in a position to nish the proof of assertion (ii). First suppose
S X Y , S R P
2
and (R) = 0. Then (R
x
) = 0 for -a.e. x X and
6.8. PRODUCT MEASURES AND FUBINIS THEOREM 193
S
x
R
x
for each x X. Since is complete, S
x
/
Y
for -a.e. x X and
S T with (S) = 0. In particular we see that if S X Y with (S) = 0, then
S T and (S) = 0.
Now suppose S /
XY
and ( )(S) < . Then, from (192.4), there is
an R P
2
such that S R and
( )(S) = (S) = (R).
From assertion (i) we see that R /
XY
, and since ( )(S) <
( )(R S) = 0.
This in turn implies that R S T and (R S) = 0. Since is complete and
R
x
S
x
/
Y
for -a.e. x X we see that S
x
/
Y
for -a.e. x X and thus that S T with
( )(S) = (S) =
_
X
(S
x
) d(x).
If S /
XY
is -nite with respect the measure , then there exists a
sequence S
j
of disjoint sets S
j
/
XY
with ( )(S
j
) < for each j such
that
S =
j=1
S
j
.
Since the sets are disjoint and each S
j
T we have S T and
( )(S) =
j=1
( )(S
j
) =
j=1
(S
j
) = (S).
Of course the above argument remains valid if the roles of and are interchanged,
and thus we have proved assertion (ii).
Proof of (iii). Assume rst that f L
1
(X Y, /
XY
, ) and f 0.
Fix t > 1 and set
E
k
= (x, y) : t
k
< f(x, y) t
k+1
.
for each k = 0, 1, 2, . . . . Then each E
k
/
XY
with ( )(E
k
) < . In
view of (ii) the function
f
t
=
k=
t
k
E
k
194 6. INTEGRATION
satises the rst four assertions of (ii) and
_
f
t
d( ) =
k=
t
k
( )(E
k
)
=
k=
t
k
(E
k
) by (192.1)
=
k=
t
k
_
X
__
Y
E
k
(x, y)d(y)
_
d(x)
=
_
X
_
k=
t
k
_
Y
E
k
(x, y)d(y)
_
d(x)
=
_
X
__
Y
f
t
(x, y)d(y)
_
d(x)
by the Monotone Convergence Theorem. Similarly
_
f
t
d( ) =
_
Y
__
X
f
t
(x, y) d(x)
_
d(y).
Since
1
t
f f
t
f
we see that f
t
(x, y) f(x, y) as t 1
+
for each (x, y) XY . Thus the function
y f(x, y)
is /
Y
-measurable for -a.e. x X. It follows that
1
t
_
Y
f(x, y)d(y)
_
Y
f
t
(x, y)d(y)
_
Y
f(x, y)d(y)
for -a.e. x X, the function
x
_
Y
f(x, y)d(y)
is /
X
-measurable and
1
t
_
X
__
Y
f(x, y)d(y)
_
d(x)
_
X
__
Y
f
t
(x, y)d(y)
_
d(x)
_
X
__
Y
f(x, y)d(y)
_
d(x).
Thus we see
_
fd( ) = lim
t1
+
_
f
t
(x, y)d( )
= lim
t1
+
_
X
__
Y
f
t
(x, y)d(y)
_
d(x)
=
_
X
__
Y
f(x, y)d(y)
_
d(x).
6.8. PRODUCT MEASURES AND FUBINIS THEOREM 195
Since the rst integral above is nite we have established the rst and third parts
of assertion (iii) as well as the rst half of the fth part. The remainder of (iii)
follows by an analogous argument.
To extend the proof to general f L
1
(X Y, /
XY
, ) we need only
recall that
f = f
+
f
where f
+
and f
k
_
Q
k
[f[ =
whereas
_
1
0
f(x, y) d
1
(x) =
_
1
0
f(x, y) d
1
(y) = 0.
However, this integrability hypothesis is not necessary if f 0 and if f is
measurable in each variable separately. The proof of this follows readily from the
proof of Theorem 189.1.
196 6. INTEGRATION
195.2. Corollary (Tonelli). If f is a nonnegative /
XY
-measurable function
and (x, y) : f(x, y) ,= 0 is -nite with respect to the measure , then the
function
y f(x, y)
is /
Y
-measurable for -a.e. x X,
x
_
Y
f(x, y) d(y)
is /
X
-measurable, and
_
XY
f d( ) =
_
X
__
Y
f(x, y) d
_
d
in the sense that either both expressions are innite or both are nite and equal.
Proof. Let f
k
be a sequence of nonnegative real-valued measurable func-
tions with nite range such that f
k
f
k+1
and lim
k
f
k
= f. By assertion (ii)
of Theorem 189.1 the conclusion of the corollary holds for each f
k
. For each k let
N
k
be a /
X
-measurable subset of X such that (N
k
) = 0 and
y f
k
(x, y)
is /
Y
-measurable for each x X N
k
. Set N =
k=1
N
k
. Then (N) = 0 and
for each x X N
y f(x, y) = lim
k
f
k
(x, y)
is /
Y
-measurable and by the Monotone Convergence Theorem
_
Y
f(x, y) d(y) = lim
k
_
Y
f
k
(x, y) d(y)
for x X N.
Finally, again by the Monotone Convergence Theorem,
_
XY
f d( ) = lim
k
_
XY
f
k
(x, y) d( )
= lim
k
_
X
__
Y
f
k
(x, y) d(y)
_
d(x)
=
_
X
__
Y
f(x, y) d(y)
_
d(x).
6.9. LEBESGUE MEASURE AS A PRODUCT MEASURE 197
6.9. Lebesgue Measure as a Product Measure
We will now show that n-dimensional Lebesgue measure on R
n
is a
product of lower dimensional Lebesgue measures.
For each positive integer k let
k
denote Lebesgue measure on R
k
and let /
k
denote the -algebra of Lebesgue measurable subsets of R
k
.
197.0. Theorem. For each pair of positive integers n and m
n+m
=
n
m
.
Proof. Let denote the outer measure on R
n
R
m
dened as in Denition
188.1 with =
n
and =
m
. We will show that =
n+m
.
If A /
n
and B /
m
are bounded sets and > 0, then there are open sets
U A, V B such that
n
(U A) <
m
(V B) <
and hence
(197.1)
n
(U)
m
(V )
n
(A)
m
(B) +(
n
(A) +
m
(B)) +
2
.
Suppose E is a bounded subset of R
n+m
and A
k
B
k
k=1
a sequence of subsets
of R
n
R
m
such that A
k
/
n
, B
k
/
m
, and E
k=1
A
k
B
k
. Assume that
the sequences
n
(A
k
)
k=1
and
m
(B
k
)
k=1
are bounded. Fix > 0. In view of
(197.1) there exist open sets U
k
, V
k
such that A
k
U
k
, B
k
V
k
, and
k=1
n
(A
k
)
m
(B
k
)
k=1
n
(U
k
)
m
(V
k
) .
It is not dicult to show that each of the open sets U
k
V
k
can be written as a
countable union of nonoverlapping closed intervals
U
k
V
k
=
l=1
I
k
l
J
k
l
where I
k
l
, J
k
l
are closed intervals in R
n
, R
m
respectively. Thus for each k
n
(U
k
)
m
(V
k
) =
l=1
n
(I
k
l
)
m
(J
k
l
) =
l=1
n+m
(I
k
l
J
k
l
).
It follows that
k=1
n
(A
k
)
m
(B
k
)
k=1
n
(U
k
)
m
(V
k
)
n+m
(E)
198 6. INTEGRATION
and hence
(E)
n+m
(E)
whenever E is a bounded subset of R
n+m
.
In case E is an unbounded subset of R
n+m
we have
(E)
n+m
(E B(0, j))
for each positive integer j. Since
n+m
is a Borel regular outer measure, (see
Exercise 4.(24)), there is a Borel set A
j
EB(0, j) such that (A
j
) =
m+n
(E
B(0, j)). With A :=
j=1
A
j
, we have
n+m
(E)
n+m
(A) = lim
j
n+m
(A
j
) = lim
j
n+m
(E B(0, j))
n+m
(E)
and therefore
lim
j
m+n
(E B(0, j)) =
m+n
(E).
This yields
(E)
m+n
(E)
for each E R
n+m
.
On the other hand it is immediate from the denitions of the two outer measures
that
(E)
m+n
(E)
for each E R
n+m
.
6.10. Convolution
As an application of Fubinis Theorem, we determine conditions on func-
tions f and g that ensure the existence of the convolution fg and deduce
the basic properties of convolution.
198.1. Definition. Given two Lebesgue measurable functions f and g on R
n
we dene the convolution f g of f and g to be the function dened for each
x R
n
by
(f g)(x) =
_
R
n
f(y)g(x y) dy.
Here and in the remainder of this section we will indicate integration with respect
to Lebesgue measure by dx, dy, etc.
We rst observe that if g is a nonnegative Lebesgue measurable function on
R
n
, then
_
R
n
g(x y) dy =
_
R
n
g(y) dy
for any x R
n
. This follows readily from the denition of the integral and the fact
that
n
is invariant under translation.
6.10. CONVOLUTION 199
To study the integrability properties of the convolution of two functions we will
need the following lemma.
199.0. Lemma. If f is a Lebesgue measurable function on R
n
, then the function
F dened on R
n
R
n
= R
2n
by
F(x, y) = f(x y)
is
2n
-measurable.
Proof. First, dene F
1
: R
2n
R by F
1
(x, y) := f(x) and observe that F
1
is
2n
-measurable because for any Borel set B R, we have F
1
(B) = f
1
(B) R
n
.
Then dene T : R
2n
R
2n
by T(x, y) = (xy, x+y). Note that T is a non-singular
linear transformation and therefore both T and T
1
are Lipschitz transformations.
Hence, it follows that F
1
T = F is
2n
-measurable. Indeed, if B R is a Borel
set, then E := F
1
1
(B) is
2n
-measurable and thus can be expressed as E = B
1
N
where B
1
R
2n
is a Borel set and
2n
(N) = 0, see Exercise 4.47. Consequently,
T
1
(E) = T
1
(B
1
) T
1
(N), which is the union of a Borel set and a set of
2n
-
measure zero.
We now prove a basic result concerning convolutions. We will write L
p
(R
n
) for
L
p
(R
n
, /
n
,
n
) and |f|
p
for |f|
L
p
(R
n
)
.
199.1. Theorem. If f L
p
(R
n
), 1 p and g L
1
(R
n
), then f g
L
p
(R
n
) and
|f g|
p
|f|
p
|g|
1
.
Proof. Observe that [f g[ [f[[g[ and thus it suces to prove the assertion
for f, g 0. Then by Lemma 198.2 the function
(x, y) f(y)g(x y)
is nonnegative and /
2n
-measurable and by Corollary 195.2
_
(f g)(x) dx =
_ _
f(y)g(x y) dy dx
=
_
f(y)
__
g(x y) dx
_
dy
=
_
f(y) dy
_
g(x) dx.
Thus the assertion holds if p = 1.
In case p = we see
(f g)(x) |f|
_
g(x y) dy = |f|
|g|
1
200 6. INTEGRATION
whence
|f g|
|f|
|g|
1
.
Finally suppose 1 < p < . Then
(f g)(x) =
_
f(y)(g(x y))
1
p
(g(x y))
1
1
p
dy
__
f
p
(y)g(x y) dy
_1
p
__
g(x y) dy
_
1
1
p
= (f
p
g)
1
p
(x) |g|
1
1
p
1
.
Thus
_
(f g)
p
(x) dx
_
(f
p
g)(x) dx|g|
p1
1
= |f
p
|
1
|g|
1
|g|
p1
1
= |f|
p
p
|g|
p
1
and the assertion is proved.
If we x g L
1
(R
n
) and set
T(f) = f g,
then we may interpret the theorem as saying that for any 1 p ,
T : L
p
(R
n
) L
p
(R
n
)
is a bounded linear mapping. Such mappings induced by convolution will be further
studied in Chapter 9.
6.11. Distribution Functions
Here we will study an interesting and useful connection between abstract
integration and Lebesgue integration.
Let (X, /, ) be a complete -nite measure space. Let f be an measurable
function on X and for t R set
E
t
= x : f(x) > t /
and
A
f
(t) = (E
t
).
The nonincreasing function A
f
() is called the distribution function of f. An
interesting relation between f and its distribution function can be deduced from
Fubinis Theorem.
6.11. DISTRIBUTION FUNCTIONS 201
200.1. Theorem. If f is nonnegative and measurable, then
(200.1)
_
X
f d =
_
[0,)
A
f
d.
Proof. Let
/ denote the -algebra of measurable subsets of X R corre-
sponding to . Set
W = (x, t) : 0 < t < f(x) X R.
Since f is measurable there is a sequence f
k
of measurable countably-simple
functions such that f
k
f
k+1
and lim
k
f
k
= f pointwise on X. If f
k
=
j=1
a
k
j
E
k
j
where for each k the sets E
k
j
are disjoint and measurable, then
W
k
= (x, t) : 0 < t < f
k
(x) =
j=1
E
k
j
(0, a
k
j
)
/.
Since
W
= lim
k
W
k
we see that W
/. Thus by Corollary 195.2
_
[0,)
A
f
d =
_
R
_
X
W
(x, t) d(x) d(t)
=
_
X
_
R
W
(x, t) d(t) d(x)
=
_
X
(t : 0 < t < f(x)) d(x)
=
_
X
f d.
Thus a nonnegative measurable function f is integrable over X with respect to
if and only if its distribution function A
f
is integrable over [0, ) with respect
to one-dimensional Lebesgue measure .
If (X) < , then A
f
is a bounded monotone function and thus continuous -
a.e. on [0, ). In view of Theorem 162.1 this implies that A
f
is Riemann integrable
on any compact interval in [0, ) and thus that the right-hand side of (200.1) can
be interpreted as an improper Riemann integral.
The simple idea behind the proof of Theorem 200.1 can readily be extended as
in the following theorem.
201.1. Theorem. If f is measurable and 1 p < ,then
_
X
[f[
p
d = p
_
[0,)
t
p1
(x : [f(x)[ > t) d(t).
Proof. Set
W = (x, t) : 0 < t < [f(x)[
and note that the function
(x, t) pt
p1
W
(x, t)
202 6. INTEGRATION
is
/-measurable. Thus by Corollary 195.2
_
X
[f[
p
d =
_
X
_
(0,|f(x)|)
pt
p1
d(t) d(x)
=
_
X
_
R
pt
p1
W
(x, t) d(t) d(x)
=
_
R
_
X
pt
p1
W
(x, t) d(x) d(t)
= p
_
[0,)
t
p1
(x : [f(x)[ > t) d(t).
202.1. Remark. A useful mnemonic relating to the previous result is that if f
is measurable and 1 p < , then
_
X
[f[
p
d =
_
0
([f[ > t) dt
p
.
6.12. The Marcinkiewicz Interpolation Theorem
In the previous section, we employed Fubinis theorem extensively to
investigate the properties of the distribution function. We close this
chapter by pursuing this topic further to establish the Marcinkiewicz
Interpolation Theorem, which has important applications in diverse ar-
eas of analysis, such as Fourier analysis and nonlinear potential theory.
Later, in Chapter 7, we will see a beautiful interaction between this
result and the Hardy-Littlewood Maximal function, Deniton 224.1.
In preparation for the main theorem of this section, we will need two preliminary
results. The rst, due to Hardy, gives two inequalities that are related to Jensens
inequality Exercise 6.32. If f is a non-negative measurable function dened on the
positive real numbers, let
F(x) =
1
x
_
x
0
f(t)dt, x > 0.
G(x) =
1
x
_
x
f(t)dt, x > 0.
Jensens inequality states that, for p 1, [F(x)]
p
|f|
p;(0,x)
for each x > 0
and thus provides an estimate of F(x)
p
; Hardys inequality (202.1) below, gives an
estimate of a weighted integral of F
p
.
202.2. Lemma (Hardys Inequalities). If 1 p < , r > 0 and f is a
non-negative measurable function on (0, ), then with F and G dened as above,
_
0
[F(x)]
p
x
pr1
dx
_
p
r
_
p
_
0
[f(t)]
p
t
pr1
dt. (202.1)
_
0
[G(x)]
p
x
p+r1
dx
_
p
r
_
p
_
0
[f(t)]
p
t
p+r1
dt. (202.2)
6.12. THE MARCINKIEWICZ INTERPOLATION THEOREM 203
Proof. To prove (202.1), we apply Jensens inequality (Exercise 6.32) with
the measure t
(r/p)1
dt, we obtain
__
x
0
f(t)dt
_
p
=
__
x
0
f(t)t
1(r/p)
t
(r/p)1
dt
_
p
(202.3)
_
p
r
_
p1
x
r(11/p)
_
x
0
[f(t)]
p
t
pr1+r/p
dt. (202.4)
Then by Fubinis theorem,
_
0
__
x
0
f(t)dt
_
p
x
pr1
dx
_
p
r
_
p1
_
0
x
1(r/p)
__
x
0
[f(t)]
p
t
pr1+(r/p)
dt
_
dx
=
_
p
r
_
p1
_
0
[f(t)]
p
t
pr1+(r/p)
__
t
x
1(r/p)
dx
_
dt
=
_
p
r
_
p
_
0
[f(t)t]
p
t
r1
dt.
The proof of (202.2) proceeds in a similar way.
203.1. Lemma. If f 0 is a non-increasing function on (0, ), 0 < p
and p
1
p
2
, then
__
0
[x
1/p
f(x)]
p
2
d(x)
x
_
1/p
2
C
__
0
[x
1/p
f(x)]
p
1
d(x)
x
_
1/p
1
where C = C(p, p
1
, p
2
).
Proof. Since f is non-increasing, we have for any x > 0
x
1/p
f(x) C
_
_
x
x/2
[(x/2)
1/p
f(x)]
p
1
d(y)
y
_
1/p
1
C
_
_
x
x/2
[y
1/p
f(x)]
p
1
d(y)
y
_
1/p
1
C
_
_
x
x/2
[y
1/p
f(y)]
p
1
d(y)
y
_
1/p
1
C
__
0
[y
1/p
f(y)]
p
1
d(y)
y
_
1/p
1
,
which implies the desired result when p
2
= . The general result follows by writing
_
0
[x
1/p
f(x)]
p
2
d(x)
x
sup
x>0
[x
1/p
f(x)]
p
2
p
1
_
0
[x
1/p
f(x)]
p
1
d(x)
x
.
203.2. Definition. A Borel measure on a topological space X that is nite
on compact sets is called a Radon measure. Thus, Lebesgue measure on R
n
is a
Radon measure, but s-dimensional Hausdor measure, 0 s < n, is not.
204 6. INTEGRATION
203.3. Definition. Let be a nonnegative Radon measure dened on R
n
and
suppose f is a -measurable function dened on R
n
. Its distribution function,
A
f
(), is dened by
A
f
(t) :=
_
x : [f(x)[ > t
_
.
The non-increasing rearrangement of f, denoted by f
, is dened as
(204.1) f
(t) = inf : A
f
() t.
For example, if is taken as Lebesgue measure, then f
and A
f
are non-increasing and right continuous. Since A
f
is right continuous, it follows that the inmum in (204.1) is attained. Therefore if
(204.2) f
(t) = , then A
f
(
) > t where
< .
Furthermore,
f
|
p
=
__
0
[f
(t)[
p
dt
_
1/p
=
__
R
n
[f(x)[
p
d
_
1/p
= |f|
p;
.
204.1. Remark. Observe that the L
p
norm of f relative to the measure is
thus expressed as the norm of f
(t)
_
t for all t > 0.
204.2. Lemma. For any t > 0, > 0, suppose an arbitrary function f L
p
(R
n
)
is decomposed as follows: f = f
t
+f
t
where
f
t
(x) =
_
_
_
f(x) if [f(x)[ > f
(t
)
0 if [f(x)[ f
(t
)
and f
t
:= f f
t
. Then
(204.5)
(f
t
)
(y) f
(y) if 0 y t
(f
t
)
(y) = 0 if y > t
(f
t
)
(y) f
(y) if y > t
(f
t
)
(y) f
(t
) if 0 y t
(y) =
f
t
>
[f[ >
,
which would imply
y < A
f
t (
) A
f
(
) y,
a contradiction.
In the second inequality, assume y > t
. Now f
t
= f
{|f|>f
(t
)}
and f
(t
) =
where A
f
() t
as in (204.2). Thus, A
f
[f
(t
)] = A
f
() t
and therefore
(
f
t
>
) = ([f[ > f
(t
)) t
< y
for all
(y) = 0.
205.1. Definition. Suppose (X/, ) is a measure space and let (p, q) be a
pair of numbers such that 1 p, q < . Also, let be a Radon measure dened
on X and suppose T is an sub-additive operator dened on L
p
(X) whose values
are -measurable functions. Thus, T(f) is a -measurable function on X and we
will write Tf := T(f). The operator T is said to be of weak-type (p, q) if there is
a constant C such that for any f L
p
(X, ) and > 0,
(x : [(Tf)(x)[ > ) (
1
C|f|
p;
)
q
.
T is said to be of strong type (p, q) if there is a constant C such that |Tf|
q;
C |f|
p;
for all f L
p
(X, ).
205.2. Theorem (Marcinkiewicz Interpolation Theorem). Let (p
0
, q
0
) and
(p
1
, q
1
) be pairs of numbers such that 1 p
i
q
i
< , i = 0, 1, and q
0
,= q
1
. Let
be a Radon measure dened on R
n
and suppose T is an sub-additive operator
dened on L
p
0
(R
n
) + L
p
1
(R
n
) whose values are -measurable functions. Suppose
T is simultaneously of weak-types (p
0
, q
0
) and (p
1
, q
1
). If 0 < < 1, and
(205.1)
1/p =
1
p
0
+
p
1
1/q =
1
q
0
+
q
1
,
then T is of strong type (p, q); that is,
|Tf|
q;
C |f|
p;
, f L
p
(R
n
),
where C = C(p
0
, q
0
, p
1
, q
1
, ).
206 6. INTEGRATION
Proof. The easiest case arises when p
0
= p
1
and is left as an exercise.
Henceforth, assume p
0
< p
1
. Let (Tf)
<
, A
Tf
(
C
0
_
A
Tf
(
)
_
1/q
0
|f|
p
0
;
< C
0
t
1/q
0
|f|
p
0
;
whenever f L
p
0
(R
n
). Since
< C
0
t
1/q
0
|f|
p
0
;
for all
< = (Tf)
(t), it
follows that
(206.1) (Tf)
(t) C
0
t
1/q
0
|f|
p
0
;
.
A similarly argument shows that if f L
p
1
(R
n
), then
(206.2) (Tf)
(t) C
1
t
1/q
1
|f|
p
1
;
.
We now appeal to Lemma 204.2 where is taken as
(206.3) :=
1/q
0
1/q
1/p
0
1/p
=
1/q 1/q
1
1/p 1/p
1
.
Recall the decomposition f = f
t
+f
t
; since p
0
< p < p
1
, observe from (204.5) that
f
t
L
p
0
(R
n
) and f
t
L
p
1
(R
n
). Also, we leave the following as an exercise 6218.1:
(206.4) (Tf)
(t) (Tf
t
)
(t/2) + (Tf
t
)
(t/2)
Since p
i
q
i
, i = 0, 1, by (205.1) we have p q. Thus, we obtain
|(Tf)
|
q
=
__
0
_
t
1/q
(Tf)
(t)
_
q dt
t
_
1/q
C
__
0
_
t
1/q
(Tf)
(t)
_
p dt
t
_
1/p
by Lemma 203.1
C
__
0
_
t
1/q
(Tf
t
)
(t)
_
p dt
t
_
1/p
+C
__
0
_
t
1/q
(Tf
t
)
(t)
_
p dt
t
_
1/p
by (206.4) (206.5)
C
__
1
0
_
t
1/q1/q
0
_
_
f
t
_
_
p
0
_
p dt
t
_1/p
by (206.1) (206.6)
+C
__
1
0
_
t
1/q1/q
1
|f
t
|
p
1
_
p dt
t
_1/p
by (206.2) . (206.7)
6.12. THE MARCINKIEWICZ INTERPOLATION THEOREM 207
With dened by (206.3) we estimate the last two integrals with an appeal to
Lemma 203.1 and write
_
_
f
t
_
_
p
0
=
__
0
_
y
1/p
0
(f
t
)
(y)
_
p
0
d(y)
y
_
1/p
0
C
_
t
0
y
1/p
0
(f
t
)
(y)
d(y)
y
C
_
t
0
y
1/p
0
f
(y)
d(y)
y
Inserting this estimate for |f
t
|
p
into (206.6) we obtain
__
1
0
_
t
1/q1/q
0
_
_
f
t
_
_
p
0
_
p
dt
t
_1/p
C
_
_
1
0
_
(t
1/q1/q
0
_
t
0
y
1/p
0
f
(y)
d(y)
y
_
p
dt
t
_
1/p
Thus, to estimate (206.6) we analyze
_
0
_
t
1/q1/q
0
_
t
0
y
1/p
0
f
(y)
d(y)
y
_
p
dt
t
which, under the change of variables t
s, becomes
1
_
0
_
s
1/p1/p
0
_
s
0
y
1/p
0
1
f
(y) d(y)
_
p
ds
s
.
which is equal to
1
_
0
_
s
1/p1/p
0
1/p
_
s
0
y
1/p
0
1
f
(y) d(y)
_
p
ds.
Now apply Hardys inequality (202.1) with r1 = p/p
0
and f(y) = y
1/p
0
1
f
(y)
to obtain
__
1
0
_
t
1/q1/q
0
_
_
f
t
_
_
p
0
;
_
p dt
t
_1/p
C(p, r)
__
0
f(y)
p
y
pr1
_
1/p
= C(p, r) |f
|
p
.
The estimate of
_
0
_
t
1/q1/q
1
_
t
0
y
1/p
1
f
(y)
dy
y
_
p
dt
t
proceeds in a similar way, and thus our result is established.
208 6. INTEGRATION
Exercises for Chapter 6
Section 6.1
6.1 A series
i=1
c
i
is said to converge unconditionally if it converges, and for any
one-to-one mapping of N onto N the series
i=1
c
(i)
converges to the same
limit. Verify the assertion in Remark (154.2). That is, suppose N
1
and N
2
are
both innite subsets of N such that N
1
N
2
= and N
1
N
2
= N. Suppose
a
i
: i N are real numbers such that a
i
: i N
1
are all nonpositive and
that a
i
: i N
2
are all positive numbers. If
iN
1
a
i
< and
iN
2
a
i
=
prove that
(i)N
a
(i)
=
for any bijection : N N. Also, show that
(i)N
a
(i)
< and
i=1
[a
i
[ <
if
iN
2
a
i
< . Use the assertion to show that if f is a nonnegative countably-
simple function and g is an integrable countably-simple function, then
_
X
(f +g) d =
_
X
f d +
_
X
g d.
6.2 Verify the assertions of Theorem 154.5 for countably-simple functions.
6.3 Suppose f is a nonnegative measurable function. Show that
_
X
f d = sup
N
k=1
( inf
xE
k
f(x))(E
k
)
where the supremum is taken over all nite measurable partitions of X, i.e.,
over all nite collections E
k
N
k=1
of disjoint measurable subsets of X such that
X =
N
k=1
E
k
.
6.4 Suppose f is a nonnegative, integrable function with the property that
_
X
f d = 0
Show that f = 0 -a.e.
6.5 Suppose f is an integrable function with the property that
_
E
f d = 0
whenever E is an -measurable set. Show that f = 0 -a.e.
EXERCISES FOR CHAPTER 6 209
6.6 Show that if f is measurable, g is -integrable, and f g, then f
is -
integrable and
_
X
f d =
_
X
f d =
_
X
f
+
d
_
X
f
d.
6.7 Let (X, /, ) be an arbitrary measure space. For an arbitrary X
f
R prove
that there is a measurable function with g f -a.e such that
_
X
g d =
_
X
f d
6.8 Suppose (X, /, ) is a measure space, f : X R, and
_
X
f d =
_
X
f d <
Show that there exists an integrable (and measurable) function g such that
f = g -a.e. Thus, if (X, /, ) is complete, f is measurable.
6.9 Suppose (X, /, ) is a measure space and Y /. Set
Y
(E) = (E Y )
for each E /. Show that
Y
is a measure of (X, /) and that
_
g d
Y
=
_
g
Y
d
for each nonnegative measurable function g on X.
6.10 Suppose f is a Lebesgue integrable function on R
n
. Prove that for each > 0
there is a continuous function g on R
n
such that
_
R
n
[f(y) g(y)[ d(y) < .
6.11 A function f : (a, b) R is convex if
f[(1 t)x +ty] (1 t)f(x) +tf(x)
for all x, y (a, b) and t [0, 1]. Prove that this is equivalent to
f(y) f(x)
y x
f(z) f(y)
z y
whenever a < x < y < z < b.
Section 6.2
6.12 Suppose f
k
is a sequence of measurable functions, g is a -integrable function,
and f
k
g -a.e. for each k. Show that
_
liminf
k
f
k
d liminf
k
_
f
k
d.
210 6. INTEGRATION
6.13 Let (X, /, ) be an arbitrary measure space. For arbitrary nonnegative func-
tions
f
i
: X R, prove that
_
X
liminf
i
f
i
d liminf
i
_
X
f
i
d
Hint: See Exercise 6.7.
6.14 If f
k
is an increasing sequence of measurable functions, g is -integrable, and
f
k
g -a.e. for each k, show that
lim
k
_
f
k
d =
_
lim
k
f
k
d.
6.15 Show that there exists a sequence of bounded Lebesgue measurable functions
mapping R into R such that
liminf
i
_
R
f
i
d <
_
R
liminf
i
f
i
d.
6.16 Let f be a bounded function on the unit square Q in R
2
. Suppose for each
xed y, that f is a measurable function of x. For each (x, y) Q let the partial
derivative
f
y
exist. Under the assumption that
f
y
is bounded in Q, prove
that
d
dy
_
1
0
f(x, y) d(x) =
_
1
0
f
y
d(x).
Section 6.3
6.17 Give an example of a nondecreasing sequence of functions mapping [0, 1] into
[0, 1] such that each term in the sequence is Riemann integrable and such that
the limit of the resulting sequence of Riemann integrals exists, but that the
limit of the sequence of functions is not Riemann integrable.
6.18 From here to Exercise 22 we outline a development of the Riemann-Stieltjes
integral that is similar to that of the Riemann integral. Let f and g be
two real-valued functions dened on a nite interval [a, b]. Given a partition
T = x
i
m
i=0
of [a, b], for each i 1, 2, . . . , m let x
i
be an arbitrary point
of the interval [x
i1
, x
i
]. We say that the Riemann-Stieltjes integral of f with
respect to g exists provided
lim
P0
m
i=1
f(x
i
)(g(x
i
) g(x
i1
))
exists, in which case the value is denoted by
_
b
a
f(x) dg(x).
EXERCISES FOR CHAPTER 6 211
Prove that if f is continuous and g is continuously dierentiable on [a, b], then
_
b
a
f dg =
_
b
a
fg
dx
6.19 Suppose f is a bounded function on [a,b] and g nondecreasing. Set
U
RS
(T) =
m
i=1
_
sup
x[x
i1
,x
i
]
f(x)
_
(g(x
i
) g(x
i1
))
L
RS
(T) =
m
i=1
_
inf
x[x
i1
,x
i
]
f(x)
_
(g(x
i
) g(x
i1
)).
Prove that if T
is a renement of T, then L
RS
(T
) L
RS
(T) and U
RS
(T
)
U
RS
(T). Also, if T
1
and T
2
are any two partitions, then L
RS
(T
1
) U
RS
(T
2
).
6.20 If f is continuous and g nondecreasing, prove that
_
b
a
f dg
exists. Thus establish the same conclusion if g is assumed to be of bounded
variation.
6.21 Prove the following integration by parts formula. If
_
b
a
f dg exists, then so does
_
b
a
g df and
f(b)g(b) f(a)g(a) =
_
b
a
f dg +
_
b
a
g df.
6.22 Using the proof of Theorem 161.1 as a guide, show that the Riemann-Stieltjes
and Lebesgue-Stieltjes integrals are in agreement. That is, if f is bounded, g
is nondecreasing and right-continuous, and if the Riemann-Stieltjes integral of
f with respect to g exists, then
_
b
a
f dg =
_
[a,b]
f d
g
where
g
is the Lebesgue-Stieltjes measure induced by g as in Section 4.6.
Section 6.4
6.23 Show that if f L
p
(X) (1 p < ), then there is a sequence f
k
of
measurable countably-simple functions such that [f
k
[ [f[ for each k and
lim
k
|f f
k
|
L
p
(X)
= 0.
6.24 Prove Theorem 168.1 in case p = .
6.25 Suppose (X, /, ) is an arbitrary measure space, |f|
p
< , 1 p < , and
> 0. Prove that there is a measurable set E with (E) < such that
_
E
[f[
p
d < .
212 6. INTEGRATION
6.26 Prove that convergence in L
p
implies convergence in measure.
6.27 Let (X, /, ) be a -nite measure space. Prove that there is a function
f L
1
() such that 0 < f < 1 everywhere on X.
6.28 Suppose and are measures on (X, /) with the property that (E) (E)
for each E /. For p 1 and f L
p
(X, ), show that f L
p
(X, ) and
that
_
X
[f[
p
d
_
X
[f[
p
d.
6.29 Suppose f L
p
(X, /, ), 0 < p < . Then for any t > 0,
([f[ > t) t
p
|f|
p
p;
.
This is known as Chebyshevs Inequality.
6.30 Prove that a dierentiable function f on (a, b) is convex if and only if f
is
monotonically increasing.
6.31 Prove that a convex function is continuous.
6.32 (a) Prove Jensens inequality: Let f L
1
(X, /, ) where (X) < and
suppose f(X) [a, b]. If is a convex function on [a, b], then
_
1
(X)
_
X
f d
_
1
(X)
_
X
( f) d.
Thus, (average(f)) average( f). Hint: let
t
0
= [(X)
1
]
_
X
f d. Then t
0
(a, b).
Furthermore, with
:= sup
t(a,t
0
)
(t
0
) (t)
t
0
t
,
we have (t) (t
0
) (t t
0
) for all t (a, b). In particular, (f(x))
(t
0
) (f(x) t
0
) for all x X. Now integrate.
(b) Observe that if (t) = t
p
, 1 p < , then Jensens inequality follows from
Holders inequality:
[(X)]
1
_
X
f 1 d |f|
p
[(X)]
1
p
1
= |f|
p
[(X)]
1/p
=
_
1
(X)
_
X
f d
_
p
1
(X)
_
X
([f[
p
) d.
(c) However, Jensens inequality is stronger than Holders inequality in the
following sense: If f is dened on [0, 1]then
e
X
f d
_
X
e
f(x)
d.
EXERCISES FOR CHAPTER 6 213
(d) Suppose : R R is such that
__
1
0
f d
_
_
1
0
(f) d
for every real bounded measurable f. Prove that is convex.
(e) Thus, we have
__
1
0
f d
_
_
1
0
(f) d
for each bounded, measurable f if and only if is convex.
6.33 In the context of a measure space (X, /, ), suppose f is a bounded measurable
function with a f(x) b for -a.e. x X. Prove that for each integrable
function g, there exists a number c [a, b] such that
_
X
f [g[ d = c
_
X
[g[ d.
6.34 If f L
p
(R
n
), 1 p < , then prove
lim
|h|0
|f(x +h) f(x)|
p
= 0.
Also, show that this result fails when p = .
6.35 Let p
1
, p
2
, . . . , p
m
be positive real numbers such that
m
i=1
p
i
= 1.
For f
1
, f
2
, . . . , f
m
L
1
(X, ), prove that
f
p
1
1
f
p
2
2
f
p
m
m
L
1
(X, )
and
_
X
(f
p
1
1
f
p
2
2
f
p
m
m
) d |f
1
|
p
1
1
|f
2
|
p
2
1
|f
m
|
p
m
1
.
Section 6.5
6.36 Prove property (iii) that follows (171.1).
6.37 Prove that the three conditions in (175.1) are equivalent.
6.38 Let (X, /, ) be a -nite measure space, and let f L
1
(X, ). In particular,
f is /-measurable. Suppose /
0
/ be a -algebra. Of course, f may
not be /
0
-measurable. However, prove that there is a unique /
0
-measurable
function f
0
such that
_
fg d/u =
_
f
0
g d/u
for each /
0
-measurable g for which the integrals are nite. Hint: Use the
Radon-Nikodym Theorem.
214 6. INTEGRATION
6.39 Suppose that and are -nite measures on (X, /) such that << and
<< . Prove that
d
d
,= 0
almost everywhere and
d
d
= 1/
d
d
.
6.40 Let f : R R be a nondecreasing, continuously dierentiable function and let
f
be the corresponding Lebesgue-Stieltjes measure, see Denition 96.2, p. 97.
Prove:
(a)
f
<< .
(b)
d
f
d
= f
.
6.41 Let (X, /, ) be a nite measure space with (X) < . Let
k
be a sequence
of nite measures on / (that is,
k
(X) < for all k) with the property that
they are uniformly absolutely continuous with respect to ; that is, for
each > 0, there exists > 0 and a positive integer K such that
k
(E) < for
all k K and all E / for which (E) < . Assume that the limit
(E) := lim
k
k
(E), E /
exists. Prove that is a -nite measure on /.
6.42 Let (X, /, ) a nite measure space and dene a metric space
/ as follows:
for A, B /, dene
d(A, B) := (AB) where AB denotes symmetric dierence.
The space
/ is dened as all sets in / where sets A and B are identied if
(AB) = 0.
(a) Prove that (
k
(E),
k
<< for each k.
(a) Prove that each
k
is well dened and continuous on the space (
/, d).
EXERCISES FOR CHAPTER 6 215
(b) For > 0, let
/
i,j
:= E / : [
i
(E)
j
(E)[
3
, i, j = 1, 2, . . .
and
/
p
:=
i,jp
/
i,j
, p = 1, 2, . . . .
Prove that /
p
is a closed set in (
/, d).
(c) Prove that there is some q such that /
q
contains an open set, call it U.
(d) Prove that the
k
are uniformly absolutely continuous with respect to
, as in Problem 6.41.
Hint: Let A be an interior point in U and for B /, write
k
(B) =
q
(B) + [
k
(B)
q
(B)]
and use the identity
i
(B) =
i
(BA)+
i
(BA), i = 1, 2 . . . to estimate
i
(B).
(e) Prove that is a nite measure.
Section 6.6
6.45 Prove the uniqueness assertion in Theorem 181.1.
Section 6.7
6.46 Suppose (X, /, ) is a -nite measure space and that g is a given function
with g L
p
, 1 p < . Assume
_
X
fg d <
for each f L
p
(X). Prove that |f|
p
< . Hint: See the proof of Theorem
52.1, p. 52.
6.47 Suppose f is a nonnegative measurable function. Set
E
t
= x : f(x) > t
and
g(t) = (E
t
)
for t R. Show that
_
f d =
_
0
t d
g
(t)
where
g
is the Lebesgue-Stieltjes measure induced by g as in Section 4.6.
6.48 Let X be a well-ordered set (with ordering denoted by <) that is a represen-
tative of the ordinal number and let / be the -algebra consisting of those
sets E with the property that either E or its complement is at most countable.
Let be the measure dened on E / as (E) = 0 if E is at most countable
and (E) = 1 otherwise. Let Y := X, := and
216 6. INTEGRATION
(a) Show that if A := (x, y) XY : x < y, then A
x
and A
y
are measurable
for all x and y.
(b) Show that both
_
Y
__
X
A
(x, y) d(x)
_
d(y)
and
_
X
__
Y
A
(x, y) d(y)
_
d(x)
exist.
(c) Show that
_
Y
__
X
A
(x, y) d(x)
_
d(y) ,=
_
X
__
Y
A
(x, y) d(y)
_
d(x)
(d) Why doesnt Fubinis Theorem apply in this example?
6.49 Let f and g be integrable functions on a measure space (X, /, ) with the
property that
[f > tg > t] = 0
for -a.e. t. Prove that f = g -a.e.
6.50 Let f be a Lebesgue measurable function on [0, 1] and let Q := [0, 1] [0, 1].
(a) Show that F(x, y) := f(x) f(y) is measurable with respect to Lebesgue
measure in R
2
.
(b) If F L
1
(Q), show that f L
1
([0, 1]).
Section 6.10
6.51 (a) For p > 1 and p
(R
n
),
then
f g(x) |f|
p
|g|
p
.
for any x R
n
.
(b) Suppose that f L
p
(R
n
) and g L
p
(R
n
). Prove that f g vanishes at
innity. That is, prove that for each > 0, there exists R > 0 such that
f g(x) < for all [x[ > R.
6.52 Let be a Radon measure on R
n
, x R
n
and 0 < < n. Then
_
R
n
d(y)
[x y[
n
= (n )
_
0
r
n1
(B(x, r)) dr,
provided that
_
R
n
d(y)
[x y[
n
< .
EXERCISES FOR CHAPTER 6 217
6.53 Let be a non-negative, real-valued function in C
0
(R
n
) with the property
that
_
R
n
(x)dx = 1, spt B(0, 1).
An example of such a function is given by
(x) =
_
_
_
C exp[1/(1 [x[
2
)] if [x[ < 1
0 if [x[ 1
where C is chosen so that
_
R
n
= 1. For > 0, the function
(x) :=
n
(x/) belongs to C
0
(R
n
) and spt
B(0, ).
is called a regularizer
(or mollier) and the convolution
u
(x) :=
u(x) :=
_
R
n
(x y)u(y)dy
dened for functions u L
1
loc
(R
n
) is called the regularization (mollication) of
u. As a consequence of Fubinis theorem, we have
|u v|
p
|u|
p
|v|
1
whenever 1 p , u L
p
(R
n
) and v L
1
(R
n
).
Prove the following:
(a) If u L
1
loc
(R
n
), then for every > 0, u
(R
n
).
(b) If u is continuous, then u
s and b
P is also a polynomial.
In other words, it isnt possible to make a polynomial anymore smooth than
itself.
6.55 If u L
p
(R
n
), 1 p < , then u
L
p
(R
n
), |u
|
p
|u|
p
, and
lim
0
|u
u|
p
= 0.
6.56 Consider (X, /, ) where is -nite and complete and suppose f L
1
(X)
is nonnegative. Let
G
f
:= (x, y) X [0, ] : 0 y f(x).
Prove the following:
(a) The set G
f
is
1
-measurable.
(b)
1
(G
f
) =
_
X
f d.
218 6. INTEGRATION
This shows that the area under the graph is the integral of the function.
Section 6.11
6.57 Suppose f L
1
(R
n
) and let A
t
:= x : [f(x)[ > t. Prove that
lim
t
_
A
t
[f[ d = 0
Section 6.12
6.58 Prove the Marcinkiewicz Interpolation Theorem in the case when p
0
= p
1
.
6.59 If f = f
1
+f
2
prove that
(218.1) (Tf)
(t) (Tf
1
)
(t/2) + (Tf
2
)
(t/2)
CHAPTER 7
Dierentiation
7.1. Covering Theorems
Certain covering theorems, such as the Vitali Covering Theorem, will
be developed in this section. These covering theorems are of essential
importance in the theory of dierentiation of measures.
We depart from the theory of abstract measure spaces encountered in previous
chapters and focus on certain aspects of functions dened in R. A major result
in elementary analysis is the fundamental theorem of calculus, which states that
a C
1
function can be expressed as the integral of its derivative. One of the main
objectives of this chapter is to show that this result still holds for a more general
class of functions. In fact, we will determine precisely those functions for which
the fundamental theorem holds. We will take a broader view of dierentiation by
developing a framework for dierentiation of measures. This will include the usual
notion of dierentiability of a function. The following result, whose proof is left as
an exercise, will serve to motivate our point of view.
219.1. Remark. Suppose is a Borel measure on R and let
F(x) = ((, x]) for x R.
Then the following two statements are equivalent:
(i) F is dierentiable at x
0
and F
(x
0
) = c
(ii) For every > 0 there exists > 0 such that
(I)
(I)
c
<
whenever I is an open interval containing x
0
of length less than .
Condition (ii) may be interpreted as the derivative of with respect to Lebesgue
measure, . This concept will be developed more fully throughout this chapter. As
an example in this framework, let f L
1
(R) be nonnegative and dene a measure
by
(219.1) (E) =
_
E
f(y) d(y)
219
220 7. DIFFERENTIATION
for every Lebesgue measurable set E. Then the function F introduced above can
be expressed as
F(x) =
_
x
f(y) d(y).
Of course, the derivative of F at x
0
is the limit
(220.1) lim
h0
1
h
_
x
0
+h
x
0
f(y) d(y).
This, in turn, is equivalent to statement (ii) above. The analog of this in R
n
is
lim
r0
1
[B(x
0
, r)]
_
B(x
0
,r)
f(y) d(y)
where now f L
1
R
n
. With as in (219.1), this is the same as
(220.2) lim
r0
[B(x
0
, r)]
[B(x
0
, r)]
.
At this stage we know nothing about the existence of the limit.
220.1. Remark. Strictly speaking, (220.2) is not the precise analog of (220.1)
since we have considered only balls B(x
0
, r) centered at x
0
. In R this would preclude
the use of intervals whose left endpoint is x
0
as required by (220.1). Nevertheless,
in our development we choose to use the family of open concentric balls for several
reasons. First, they are slightly easier to employ than nonconcentric balls; second,
we will see that it is immaterial to the main results of the theory whether or not
concentric balls are used. Finally, in the development of the derivative of relative
to an arbitrary measure , it is important that concentric balls are used. Thus, we
formally introduce the notation
(220.3) D
(x
0
) = lim
r0
[B(x
0
, r)]
[B(x
0
, r)]
which is the derivative of with respect to .
One of the major objectives of the next section is to prove that the limit in
(220.3) exists almost everywhere and to see how it relates to the results sur-
rounding the Radon-Nikodym Theorem. In particular, in view of Theorems 98.1
and (219.1), it will follow from our development that a nondecreasing function is
dierentiable almost everywhere. In the following we will sue the following nota-
tion: Given B = B(x, r) we will call B(x, 5r) the enlargement of B and denote it
by
B. The next lemma states that any collection of balls (that may be either open
or closed) whose radii are bounded has a countable disjoint subcollection with the
property that the union of their enlargements contains the union of the original
collection. We emphasize here that the point of the lemma is that the subcollec-
tion consists of disjoint elements, a very important consideration since countable
additivity plays a central role in measure theory.
7.1. COVERING THEOREMS 221
221.0. Theorem. Let ( be a family of closed or open balls in R
n
with
R := supdiamB : B ( < .
Then there is a countable subfamily T ( of pairwise disjoint elements such that
B : B (
B : B T
In fact, for each B ( there exists B
T such that B B
,= 0 and B
B
.
Proof. Throughout this proof, we will adopt the following notation: If A is
set and T a family of sets, then we use the notation A T , = to mean that
A B ,= for some B T.
Let a be a number such that 1 < a < a
1+2R
< 2. For j = 1, 2, . . . let
(
j
=
_
B := B(x, r) ( : a
|x|j
<
r
R
a
|x|j+1
_
,
and observe that ( =
j=1
(
j
. Since r/R < 1 and a > 1, observe that for any
B = B(x, r) (
j
we have
[x[ < j and r > a
j
R.
Hence the elements of (
j
are centered at points x B(0, j) and their radii are
bounded away from zero; this implies that there is a number M
j
> 0 depending
only on a, j and R such that any disjoint subfamily of (
j
has at most M
j
elements.
The family T will be of the form
T =
j=0
T
j
where the T
j
are nite, disjoint families dened inductively as follows. We set
T
0
= . Let T
1
be the largest (in the sense of inclusion) disjoint subfamily of (
1
.
Note that T
1
can have no more than M
1
elements. Proceeding by induction, we
assume that T
j1
has been determined, and then dene H
j
as the largest disjoint
subfamily of (
j
with the property that B T
j1
= for each B H
j
. Note that
the number of elements in H
j
could be 0 but no more than M
j
. Dene
T
j
:= T
j1
H
j
We claim that the family T :=
j=1
T
j
has the required properties: that is, we will
show that
(221.1) B
B : B T for each B (.
222 7. DIFFERENTIATION
To verify this, rst note that T is a disjoint family. Next, select B := B(x, r)
( which implies B (
j
for some j. If BH
j
,= , then there exists B
:= B(x
, r
)
H
j
such that B B
,= , in which case
(222.1)
r
R
a
[x
[j
.
On the other hand, if BH
j
= , then BT
j1
,= , for otherwise the maximality
of H
j
would be violated. Thus, there exists B
= B(x
, r
) T
j1
such that
B B
R
> a
[x
[j+1
> a
[x
[j
.
Since
r
R
a
|x|j+1
,
it follows from (222.1) and (222.2) that
r a
|x|j+1
R a
(|x|[x
[+1)
r
.
Since a was chosen so that 1 < a < a
1+2R
< 2, we have
r a
1+|x||x
|
r
a
1+|xx
|
r
a
1+2R
r
2r
.
This implies that B
B
, then
[z x
[ [z x[ +[x y[ +[y x
[
r +r +r
5r
.
If we assume a bit more about ( we can show that the union of elements in
T contains almost all of the union
B : B (. This requires the following
denition.
222.1. Definition. A collection ( of balls is said to cover a set E R
n
in the
sense of Vitali if for each x E and each > 0, there exists B ( containing x
whose radius is positive and less than . We also say that ( is a Vitali covering
of E. Note that if ( is a Vitali covering of a set E R
n
and R > 0 is arbitrary,
then (
B : B T
B : B T T
T.
7.1. COVERING THEOREMS 223
Proof. Since ( is a Vitali covering of E, there is no loss of generality if
we assume that the radius of each ball in ( is less than some xed number R.
Let T be as in Theorem 220.2 and let T
there exists B (
such that x B and B [B : B T
since B B
1
,= . Therefore
x
B
1
B : B T T
.
223.1. Remark. The preceding result and the next one are not needed in the
sequel, although they are needed in some of the Exercises, such as 7.38. We include
them because they are frequently used in the analysis literature and because they
follow so easily from the main result, Theorem 220.2.
Theorem 222.2 states that any nite family T
B : B T) = 0.
Proof. First, assume that E is a bounded set. Then we may as well assume
that each ball in ( is contained in some bounded open set H E. Let T be the
subfamily of disjoint balls provided by Theorem 220.2 and Corollary 222.2. Since
all elements of T are disjoint and contained in the bounded set H, we have
(223.1)
BF
(B) (H) < .
Now, by Corollary 222.2, for any nite subfamily T
T, we obtain
(E B : B T)
(E B : B T
B : B T T
BF\F
B)
5
n
BF\F
(B).
Referring to (223.1), we see that the last term can be made arbitrarily small by an
appropriate choice of T
k=1
of disjoint open cubes C
k
such that
_
R
n
k=1
C
k
_
= 0.
The details are left to the reader.
7.2. Lebesgue Points
In integration theory, functions that dier only on a set of measure zero
can be identied as one function. Consequently, with this identica-
tion a measurable function determines an equivalence class of functions.
This raises the question of whether it is possible to dene a measurable
function at almost all points in a way that is independent of any repre-
sentative in the equivalence class. Our investigation of Lebesgue points
provides a positive answer to this question.
224.1. Definition. With each f L
1
(R
n
), we associate its maximal func-
tion, Mf, which is dened as
Mf(x) := sup
r>0
_
B(x,r)
[f[ d
where
_
E
[f[ d :=
1
[E]
_
E
[f[ d
denotes the integral average of [f[ over an arbitrary measurable set E. In other
words, Mf(x) is the upper envelope of integral averages of [f[ over balls centered
at x.
Clearly, Mf : R
n
R is a nonnegative function. Furthermore, it is Lebesgue
measurable. To see this, note that for each xed r > 0,
x
_
B(x,r)
[f[ d
is a continuous function of x. (See Exercise 4.9.) Therefore, we see that Mf > t
is an open set for each real number t, thus showing that Mf is lower semicontinuous
and therefore measurable.
The next question is whether Mf is integrable over R
n
. In order for this to be
true, it follows from Theorem 200.1 that it would be necessary that
_
0
(Mf > t) d(t) < . (224.1)
It turns out that Mf is never integrable unless f is identically zero (see Exercise
4.8). However, the next result provides an estimate of how the measure of the set
Mf > t becomes small as t increases. It also shows that inequality (224.1) fails
to be true by only a small margin.
7.2. LEBESGUE POINTS 225
224.2. Theorem (Hardy-Littlewood). If f L
1
(R
n
), then
[Mf > t]
5
n
t
_
R
n
[f[ d
for every t > 0.
Proof. For xed t > 0, the denition implies that for each x Mf > t
there exists a ball B
x
centered at x such that
_
B
x
[f[ d > t
or what is the same
(225.1)
1
t
_
B
x
[f[ d > (B
x
).
Since f is integrable and t is xed, the radii of all balls satisfying (225.1) is bounded.
Thus, with ( denoting the family of these balls, we may appeal to Lemma 220.2 to
obtain a countable subfamily T ( of disjoint balls such that
Mf > t
B : B T.
Therefore,
(Mf > t)
_
BF
B
_
BF
(
B)
= 5
n
BF
(B)
<
5
n
t
BF
_
B
[f[ d
5
n
t
_
R
n
[f[ d
which establishes the desired result.
We now appeal to the results of Section 6.12 concerning the Marcinkiewicz
Interpolation Theorem. Clearly, the operator M is sub-additive and our previous
result shows that it is of weak type (1, 1), (see Denition 205.1, p. 205). Also, it is
clear that
|Mf|
|f|
for all f L
_
B(x,r)
f(y) d(y) f(x)
= limsup
r0
_
B(x,r)
[f(y) g(y)] d(y)
+
_
_
B(x,r)
g(y) d(y) g(x)
_
+ [g(x) f(x)]
_
B(x,r)
f(y) d(y) f(x)
> t,
F
t
= x : [f(x) g(x)[ > t,
and
H
t
= x : M(f g)(x) > t.
Then, by (226.3), E
t
F
t/2
H
t/2
. Furthermore,
t(F
t
)
_
F
t
[f(y) g(y)[ d(y) <
and Theorem 224.2 implies
(H
t
)
5
n
t
.
Hence
(E
t
) 2
t
+ 2
5
n
t
.
Since is arbitrary, we conclude that (E
t
) = 0 for all t > 0, thus establishing the
conclusion.
The theorem states that
(227.1) lim
r0
_
B(x,r)
f(y) d(y)
exists for a.e. x and that the limit denes a function that is equal to f almost
everywhere. The limit in (227.1) provides a way to dene the value of f at x that
is independent of the choice of representative in the equivalence class of f. Observe
that (226.2) can be written as
lim
r0
_
B(x,r)
[f(y) f(x)] d(y) = 0.
It is rather surprising that Theorem 226.1 implies the following apparently stronger
result.
227.1. Theorem. If f L
1
loc
(R
n
), then
(227.2) lim
r0
_
B(x,r)
[f(y) f(x)[ d(y) = 0
for a.e. x R
n
.
Proof. For each rational number apply Theorem 226.1 to conclude that
there is a set E
. Thus, with
E :=
Q
E
,
we have (E) = 0. Moreover, for x , E and Q, then, since [f(y) f(x)[ <
[f(y) [ +[f(x) [, (227.3) implies
limsup
r0
_
B(x,r)
[f(y) f(x)[ d(y) 2 [f(x) [ .
Since
inf[f(x) [ : Q = 0,
the proof is complete.
A point x for which (227.2) holds is called a Lebesgue point of f. Thus,
almost all points are Lebesgue points for any f L
1
loc
(R
n
).
An important special case of Theorem 226.1 is when f is taken as the charac-
teristic function of a set. For E R
n
a Lebesgue measurable set, let
(228.1) D(E, x) = limsup
r0
(E B(x, r))
(B(x, r))
(228.2) D(E, x) = liminf
r0
(E B(x, r))
(B(x, r))
.
228.1. Theorem (Lebesgue Density Theorem). If E R
n
is a Lebesgue mea-
surable set, then
D(E, x) = 1 for -almost all x E,
and
D(E, x) = 0 for -almost all x
E.
Proof. For the rst part, let B(r) denote the open ball centered at the origin
of radius r and let f =
EB(r)
. Since E B(r) is bounded, it follows that f is
integrable and then Theorem 226.1 implies that D(E B(r), x) = 1 for -almost
all x E B(r). Since r is arbitrary, the result follows.
For the second part, take f =
EB(r)
and conclude, as above, that D(
E
B(r), x) = 1 for -almost all x
E B(r). Observe that D(E, x) = 0 for all such
x, and thus the result follows since r is arbitrary.
7.3. The Radon-Nikodym Derivative Another View
We return to the concept of the Radon-Nikodym derivative in the setting
of Lebesgue measure on R
n
. In this section it is shown that the Radon-
Nikodym derivative can be interpreted as a classical limiting process,
very similar to that of the derivative of a function.
7.3. THE RADON-NIKODYM DERIVATIVE ANOTHER VIEW 229
We now turn to the question of relating the derivative in the sense of (220.3)
to the Radon-Nikodym derivative. Consider a -nite measure on R
n
that is abso-
lutely continuous with respect to Lebesgue measure. The Radon-Nikodym Theorem
asserts the existence of f L
loc
1(R
n
) (the Radon-Nikodym derivative) such that
can be represented as
(E) =
_
E
f(y) d(y)
for every Lebesgue measurable set E R
n
. Theorem 226.1 implies that
(229.1) D
(x) = lim
r0
[B(x, r)]
[B(x, r)]
= f(x)
for -a.e. x R
n
. Thus, the Radon-Nikodym derivative of with respect to and
D
(x) = 0
for -almost all x R
n
.
Proof. Since we know that is concentrated on a Borel set A with
(
k=1
E
k
and (A) = 0. Referring to Theorems
120.1 and 112.1, it follows that for every > 0 there exists an open set U
E
k
such that (U
and [B(x, r)] < k[B(x, r)]. The collection of all such
balls B(x, r) provides a covering of E
k
. Now employ Theorem 220.2 with R = 1 to
obtain a disjoint collection of balls, T, such that
E
k
BF
B.
230 7. DIFFERENTIATION
Then
(E
k
)
_
BF
B
_
5
n
BF
(B)
< 5
n
k
BF
(B)
5
n
k(U
)
5
n
k.
Since is arbitrary, this shows that (E
k
) = 0.
This result together with (229.1) establishes the following theorem.
230.1. Theorem. Suppose is a Radon measure on R
n
. Let = + be its
Lebesgue decomposition with and . Finally, let f denote the Radon-
Nikodym derivative of with respect to . Then
lim
r0
[B(x, r)]
[B(x, r)]
= f(x)
for -a.e. x R
n
.
230.2. Definition. Now we address the issue raised in Remark 220.1 concern-
ing the use of concentric balls in the denition of (220.3). It can easily be shown
that nonconcentric balls or even a more general class of sets could be used. For
x R
n
, a sequence of Borel sets E
k
(x) is called a regular dierentiation basis
at x provided there is a number
x
> 0 with the following property: There is a
sequence of balls B(x, r
k
) with r
k
0 such that E
k
(x) B(x, r
k
) and
(E
k
(x))
x
[B(x, r
k
)].
The sets E
k
(x) are in no way related to x except for the condition E
k
B(x, r
k
).
In particular, the sets are not required to contain x.
The next result shows that Theorem 230.1 can be generalized to include regular
dierentiation bases.
230.3. Theorem. Suppose the hypotheses and notation of Theorem 230.1 are
in force. Then for almost every x R
n
, we have
lim
k
[E
k
(x)]
[E
k
(x)]
= 0
and
lim
k
[E
k
(x)]
[E
k
(x)]
= f(x)
7.3. THE RADON-NIKODYM DERIVATIVE ANOTHER VIEW 231
whenever E
k
(x) is a regular dierentiation basis at x.
Proof. In view of the inequalities
(231.1)
x
[E
k
(x)]
[E
k
(x)]
[E
k
(x)]
[B(x, r
k
)]
[B(x, r
k
)]
[B(x, r
k
)]
the rst conclusion of the Theorem follows from Lemma 229.2.
Concerning the second conclusion, Theorem 227.1 implies
lim
r
k
0
_
B(x,r
k
)
[f(y) f(x)[ d(y) = 0
for almost all x and consequently, by the same reasoning as in (231.1),
lim
k
_
E
k
(x)
[f(y) f(x)[ d(y) = 0
for almost all x. Hence, for almost all x it follows that
lim
k
[E
k
(x)]
[E
k
(x)]
= lim
k
_
E
k
(x)
f(y) d(y) = f(x)
This leads immediately to the following theorem which is fundamental to the
theory of functions of a single variable. For the companion result for functions
of several variables, see Theorem 354.1. Also, see Exercise 4.11 for a completely
dierent proof.
231.1. Theorem. Let f : R R be a nondecreasing function. Then f
(x)
exists at -a.e. x R.
Proof. By Exercise 4.10 there is a right-continuous, nondecreasing function
g that agrees with f everywhere except possibly for a countable set. Now refer to
Theorem 98.1 to obtain a Borel measure such that
((a, b]) = g(b) g(a)
whenever a b. For each x R take as a regular dierentiation basis an arbitrary
sequence of half-open intervals I
k
such that either the left or right endpoint of I
k
is x. Thus the parameter
x
that appears in Denition 230.2 is equal to 1/2. The
previous result states that
lim
k
(I
k
)
(I
k
)
exists for almost all x and that the limit is independent of the regular dierentiation
basis chosen. Furthermore, as in (219.1), it is immediate that this limit is equal to
g
(x). Finally, if g
(x) and g
(x) = f
< h < h
such that h
/h
and x +h
. Then
g(x +h
) g(x)
h
h
f(x +h) g(x)
h
g(x +h
) g(x)
h
h
.
Since h
/h < 1/(1 + ), h
(x)
(1 +)
liminf
h0
f(x +h) g(x)
h
limsup
h0
f(x +h) g(x)
h
g
(x)(1 +).
Since is arbitrary, we have g
(x) = f
(x).
Another consequence of the above results is the following theorem concerning
the derivative of the indenite integral.
232.1. Theorem. Suppose f is a Lebesgue integrable function dened on [a, b].
For each x [a, b] let
F(x) =
_
x
a
f(t) d(t).
Then F
(x) is given by
F
(x) = lim
h0
1
h
_
x+h
x
f(t) d(t).
Let be the measure dened by
(E) =
_
E
f d
for every measurable set E. Using intervals of the form I
h
(x) = [x + h, x] as a
regular dierentiation basis, it follows from Theorem 230.3 that
lim
h0
1
h
_
x+h
x
f(t) d(t) = lim
h0
[I
h
(x)]
[I
h
(x)]
= f(x)
for almost all x [a, b].
7.4. Functions of Bounded Variation
The main objective of this and the next section is to completely de-
termine the conditions under which the following equation holds on an
interval [a, b]:
f(x) f(a) =
x
a
f
(t) dt for a x b.
This formula is well known in the context of Riemann integration and
our purpose is to investigate its validity via the Lebesgue integral. It will
be shown that the formula is valid precisely for the class of absolutely
continuous functions. In this section we begin by introducing functions
of bounded variation.
7.4. FUNCTIONS OF BOUNDED VARIATION 233
In the elementary version of the Fundamental Theorem of Calculus, it is as-
sumed that f
L
1
. But this is obviously
not true in view of the Cantor-Lebesgue function, f; see Remark 135.2. We have
seen that it is continuous, nondecreasing on [0, 1], and constant on each interval
in the complement of the Cantor set. Consequently, f
(t) dt = 0.
The quantity f(1)f(0) indicates how much the function varies on [0, 1]. Intuitively,
one might have guessed that the quantity
_
1
0
[f
[ d
provides a measurement of the variation of f. Although this is false in general, for
what class of functions is it true? We will begin to investigate the ideas surrounding
these questions by introducing functions of bounded variation.
233.1. Definitions. Suppose a function f is dened on I = [a, b]. The total
variation of f from a to x, x b, is dened by
V
f
(a; x) = sup
k
i=1
[f(t
i
) f(t
i1
)[
where the supremum is taken over all nite sequences a = t
0
< t
1
< < t
k
= x. f
is said to be of bounded variation (abbreviated, BV) on [a, b] if V
f
(a; b) < . If
there is no danger of confusion, we will sometimes write V
f
(x) in place of V
f
(a; x).
Note that if f is of bounded variation on [a, b] and x [a, b], then
[f(x) f(a)[ V
f
(a; x) V
f
(a; b)
from which we see that f is bounded.
It is easy to see that a bounded function that is either nonincreasing or non-
decreasing is of bounded variation. Also, the sum (or dierence) of two functions
of bounded variation is again of bounded variation. The converse, which is not so
immediate, is also true.
234 7. DIFFERENTIATION
233.2. Theorem. Suppose f is of bounded variation on [a, b]. Then f can be
written as
f = f
1
f
2
where both f
1
and f
2
are nondecreasing.
Proof. Let x
1
< x
2
b and let a = t
0
< t
1
< < t
k
= x
1
. Then
(234.1) V
f
(x
2
) [f(x
2
) f(x
1
)[ +
k
i=1
[f(t
i
) f(t
i1
)[ .
Now,
V
f
(x
1
) = sup
k
i=1
[f(t
i
) f(t
i1
)[
over all sequences a = t
0
< t
1
< < t
k
= x
1
. Hence,
(234.2) V
f
(x
2
) [f(x
2
) f(x
1
)[ +V
f
(x
1
).
In particular,
V
f
(x
2
) f(x
2
) V
f
(x
1
) f(x
1
) and V
f
(x
2
) +f(x
2
) V
f
(x
1
) +f(x
1
).
This shows that V
f
f and V
f
+ f are nondecreasing functions. The assertions
thus follow by taking
f
1
=
1
2
(V
f
+f) and f
2
=
1
2
(V
f
f).
234.1. Theorem. Suppose f is of bounded variation on [a, b]. Then f is Borel
measurable and has at most a countable number of discontinuities. Furthermore, f
is Lebesgue measurable,
(234.3) [f
(x)[ = V
(x)
for a.e. x [a, b], and
(234.4)
_
b
a
[f
(x)[ d(t) V
f
(b).
In particular, if f is nondecreasing on [a, b], then
(234.5)
_
b
a
f
is addressed in
(234.6).
Similarly, since V
f
is a nondecreasing function, we have that V
f
exists almost
everywhere. Furthermore, with f = f
1
f
2
as in the previous theorem and recalling
236 7. DIFFERENTIATION
that f
1
, f
2
0 almost everywhere, it follows that
[f
[ = [f
1
f
2
[ [f
1
[ +[f
2
[ = f
1
+f
2
= V
f
almost everywhere on [a, b].
To prove (234.3) we will show that
E := [a, b]
_
t : V
f
(t) > [f
(t)[
_
has measure zero. For each positive integer m let E
m
be the set of all t E such
that
1
t
2
with 0 <
2
1
<
1
m
implies
(236.1)
V
f
(
2
) V
f
(
1
)
2
1
>
[f(
2
) f(
1
)[
2
1
+
1
m
.
Since each t E belongs to E
m
for suciently large m we see that
E =
m=1
E
m
and thus it suces to show that (E
m
) = 0 for each m. Fix > 0 and let
a = t
0
< t
1
< < t
k
= b be a partition of [a, b] such that [t
i
t
i1
[ <
1
m
for each
i and
(236.2)
k
i=1
[f(t
i
) f(t
i1
)[ > V
f
(b)
m
.
For each interval in the partition, (234.2) states that
(236.3) V
f
(t
i
) V
f
(t
i1
) [f(t
i
) f(t
i1
)[
while (236.1) implies
(236.4) V
f
(t
i
) V
f
(t
i1
) [f(t
i
) f(t
i1
)[ +
t
i
t
i1
m
.
if the interval contains a point of E
m
. Let T
1
denote those intervals of the partition
that do not contain any points of E
m
and let T
2
denote those intervals that do
contain points of E
m
. Then, since (E
m
)
IF
2
b
I
a
I
,
V
f
(b) =
k
i=1
V
f
(t
i
) V
f
(t
i1
)
=
IF
1
V
f
(b
I
) V
f
(a
I
) +
IF
2
V
f
(b
I
) V
f
(a
I
)
=
IF
1
f(b
I
) f(a
I
) +
IF
2
f(b
I
) f(a
I
) +
b
I
a
I
m
by (236.3) and (236.4)
i=1
[f(t
i
) f(t
i1
)[ +
(E
m
)
m
V
f
(b)
m
+
(E
m
)
m
, by (236.2)
7.5. THE FUNDAMENTAL THEOREM OF CALCULUS 237
and therefore (E
m
) , from which we conclude that (E
m
) = 0 since is
arbitrary. Thus (234.3) is established.
Finally we apply (234.3) and (234.5) to obtain
_
b
a
[f
[ d =
_
b
a
V
d V
f
(b) V
f
(a) = V
f
(b),
and the proof is complete.
7.5. The Fundamental Theorem of Calculus
We introduce absolutely continuous functions and show that they are
precisely those functions for which the Fundamental Theorem of Calcu-
lus is valid.
237.1. Definition. A function f dened on an interval I = [a, b] is said to
be absolutely continuous on I (briey, AC on I) if for every > 0 there exists
> 0 such that
k
i=1
[f(b
i
) f(a
i
)[ <
for any nite collection of nonoverlapping intervals [a
1
, b
1
], [a
2
, b
2
], . . . , [a
k
, b
k
] in I
with
k
i=1
[b
i
a
i
[ < .
Observe if f is AC, then it is easy to show that
i=1
[f(b
i
) f(a
i
)[
for any countable collection of nonoverlapping intervals with
(237.1)
i=1
[b
i
a
i
[ < .
Indeed, if f is AC, then (237.1) holds for any partial sum, and therefore for the
limit of the partial sums. Thus, it holds for the whole series.
From the denition it follows that an absolutely continuous function is uniformly
continuous. The converse is not true as we shall see illustrated later by the Cantor-
Lebesgue function. (Of course it can be shown directly that the Cantor-Lebesgue
function is not absolutely continuous: see Exercise 7.14.) The reader can easily
verify that any Lipschitz function is absolutely continuous. Another example of an
AC function is given by the indenite integral; let f be an integrable function on
[a,b] and set
(237.2) F(x) =
_
x
a
f(t) d(t).
238 7. DIFFERENTIATION
For any nonoverlapping collection of intervals in [a,b] we have
k
i=1
[F(b
i
) F(a
i
)[
_
[a
i
,b
i
]
[f[ d.
With the help of Exercise 6.36, we know that the set function dened by
(E) =
_
E
[f[ d
is a measure, and clearly it is absolutely continuous with respect to Lebesgue mea-
sure. Referring to Theorem 175.2, we see that F is absolutely continuous.
238.1. Notation. The following notation will be used frequently throughout.
If I R is an interval, we will denote its endpoints by a
I
, b
I
; thus, if I is closed,
I = [a
I
, b
I
].
238.2. Theorem. An absolutely continuous function on [a, b] is of bounded
variation.
Proof. Let f be absolutely continuous. Choose = 1 and let > 0 be the
corresponding number provided by the denition of absolute continuity. Subdivide
[a,b] into a nite collection T of nonoverlapping subintervals I = [a
I
, b
I
] each of
whose length is less than . Then,
[f(b
I
) f(a
I
)[ < 1
for each I T. Consequently, if T consists of M elements, we have
IF
[f(b
I
) f(a
I
)[ < M.
To show that f is of bounded variation on [a,b], consider an arbitrary partition
a = t
0
< t
1
< < t
k
= b. Since the sum
k
i=1
[f(t
i
) f(t
i1
)[
is not decreased by adding more points to this partition, we may assume each
interval of this partition is a subset of some I T. But then, for each I T, it
follows that
k
i=1
I
(t
i
)
I
(t
i1
) [f(t
i
) f(t
i1
)[ < 1
(this is simply saying that the sum is taken over only those intervals [t
i1
, t
i
] that
are contained in I) and consequently,
k
i=1
[f(t
i
) f(t
i1
)[ < M,
thus proving that the total variation of f on [a,b] is no more than M.
7.5. THE FUNDAMENTAL THEOREM OF CALCULUS 239
Next, we introduce a property that is of great importance concerning absolutely
continuous functions. Later we will see that this property is one among three that
characterize absolutely continuous functions (see Theorem 345.1). This concept is
due to Lusin, now called condition N.
239.1. Definition. A function f dened on [a, b] is said to satisfy condition
N if f preserves sets of Lebesgue measure zero; that is, [f(E)] = 0 whenever
E [a, b] with (E) = 0.
239.2. Theorem. If f is an absolutely continuous function on [a, b], then f
satises condition N.
Proof. Choose > 0 and let > 0 be the corresponding number provided by
the denition of absolute continuity. Let E be a set of measure zero. Then there is
an open set U E with (U) < . Since U is the union of a countable collection
T of disjoint open intervals, we have
IF
(I) < .
The closure of each interval I contains an interval I
= [a
I
, b
I
] at whose
endpoints f assumes its maximum and minimum on the closure of I. Then
[f(I)] = [f(b
I
) f(a
I
)[
and the absolute continuity of f along with (237.1) imply
[f(E)]
IF
[f(I)] =
IF
[f(b
I
) f(a
I
)[ < .
Since is arbitrary, this shows that f(E) has measure zero.
239.3. Remark. This result shows that the Cantor-Lebesgue function is not
absolutely continuous, since it maps the Cantor set (of Lebesgue measure zero)
onto [0, 1]. Thus, there are continuous functions of bounded variation that are not
absolutely continuous.
239.4. Theorem. Suppose f is an arbitrary function dened on [a, b]. Let
E
f
:= (a, b) x : f
(x) = 0.
Then [f(E
f
)] = 0.
Proof. Step 1:
Initially, we will assume f is bounded; let M := supf(x) : x [a, b] and m :=
240 7. DIFFERENTIATION
inff(x) : x [a, b]. Choose > 0. For each x E
f
there exists = (x) > 0
such that
[f(x +h) f(x)[ < h
and
[f(x h) f(x)[ < h
whenever 0 < h < (x). Thus, for each x E
f
we have a collection of intervals of
the form [x h, x + h], 0 < h < (x), with the property that for arbitrary points
a
, b
I = [x h, x +h], then
(240.1)
[f(b
) f(a
)[ [f(b
) f(x)[ +[f(a
) f(x)[
< [b
x[ + [a
x[
(I).
We will adopt the following notation: For each interval I let I
be an interval
such that I
U and I
.
Since f is bounded, each interval I
) and f(a
I
) < m
I
+ (I
) where > 0 is
chosen as above, Thus,
f(I
) [m
I
, M
I
] [f(a
I
) +(I
), f(b
I
) (I
)]
and thus, by (240.1)
[f(I
)] [M
I
m
I
[ [f(b
I
) f(a
I
)[ + 2(I
) < 3(I
).
Then, using the fact that T is a disjoint family,
_
f(E
f
)
_
_
f
_
IF
I
__
IF
(f(I
)) <
IF
(I)
= 5
IF
(I)
5(U) < 5((E
f
) +).
Since > 0 is arbitrary, we conclude that (f(E
f
)) = 0 as desired.
Step 2:
Now assume f is an arbitrary function on [a, b]. Let H: R [0, 1] be a smooth,
7.5. THE FUNDAMENTAL THEOREM OF CALCULUS 241
strictly increasing function with H
(x) = H
(f(x))f
(x)
holds whenever either g or f is dierentiable at x. Then f
(x) = 0 i g
(x) = 0
and therefore E
f
= E
g
. From Step 1, we know [g(E
g
)] = 0 = [g(E
f
)]. But
f(A) = H
1
[g(A)] for any A R, we have f(E
f
) = H
1
[g(E
f
)] and therefore
[f(E
f
)] = 0 since H
1
preserves sets of measure zero.
241.1. Theorem. If f is absolutely continuous on [a, b] with the property that
f
(x) = 0
so that [a, b] = E N where N is of measure zero. Then [f(E)] = [f(N)] = 0
by the previous result and Theorem 239.2. Thus, since f([a,b]) is an interval and
also of measure zero, f must be constant.
We now have reached the main objective.
241.2. Theorem (The Fundamental Theorem of Calculus). f : [a, b] R is
absolutely continuous if and only if f
(t) d(t).
Then F is absolutely continuous (as in (237.2)). By Theorem 232.1 we have that
F
= f
[ d
for each x [a, b].
242 7. DIFFERENTIATION
Proof. We know that V
f
is a bounded, nondecreasing function and thus of
bounded variation. From Theorem 234.1 we know that
_
x
a
[f
[ d =
_
x
a
V
f
d V
f
(x)
for each x [a, b].
Fix x [a, b] and let > 0. Choose a partition a = t
0
< t
1
< < t
k
= x so
that
V
f
(x)
k
i=1
[f(t
i
) f(t
i1
)[ +.
In view of the Theorem 241.2
[f(t
i
) f(t
i1
)[ =
_
t
i
t
i1
f
_
t
i
t
i1
[f
[ d
for each i. Thus
V
f
(x)
k
i=1
_
t
i
t
i1
[f
[ d +
_
x
a
[f
[ d +.
Since is arbitrary we conclude that
V
f
(x) =
_
x
a
[f
[ d.
for each x [a, b].
7.6. Variation of Continuous Functions
One possibility of determining the variation of a function is the following. Con-
sider the graph of f in the (x, y)-plane and for each y, let N(y) denote the number
of times the horizontal line passing through (0, y) intersects the graph of f. It seems
plausible that
_
R
N(y) d(y)
should equal the variation of f on [a,b]. In case f is a continuous nondecreasing
function, this is easily seen to be true. The next theorem provides the general
result. First, we introduce some notation: If f : R R and E R, then
(242.1) N(f, E, y)
denotes the (possibly innite) number of points in the set E f
1
y. Thus,
N(f, E, y) is the number of points in E that are mapped onto y.
242.1. Theorem. Let f be a continuous function dened on [a, b]. Then
N(f, [a, b], y) is a Borel measurable function (of y) and
V
f
(b) =
_
R
1
N(f, [a, b], y) d(y).
7.6. VARIATION OF CONTINUOUS FUNCTIONS 243
Proof. For brevity throughout the proof, we will simply write N(y) for
N(f, [a, b], y).
Let m N(y) be a nonnegative integer and let x
1
, x
2
, , x
m
be points that
are mapped into y. Thus, x
1
, x
2
, . . . , x
m
f
1
y. For each positive integer
i, consider a partition, T
i
= a = t
0
< t
1
< . . . < t
k
= b of [a,b] such that the
length of each interval I is less than 1/i. Choose i so large that each interval of T
i
contains at most one x
j
, j = 1, 2, . . . , m. Then
m
IP
i
f(I)
(y).
Consequently,
(243.1) m liminf
i
_
IP
i
f(I)
(y)
_
.
Since m is an arbitrary positive integer with m N(y), we obtain
(243.2) N(y) liminf
i
_
IP
i
f(I)
(y)
_
.
On the other hand, for any partition T
i
we obviously have
(243.3) N(y)
IP
i
f(I)
(y)
provided that each point of f
1
(y) is contained in the interior of some interval
I T
i
. Thus (243.3) holds for all but nitely many y and therefore
N(y) limsup
i
_
IP
i
f(I)
(y)
_
.
holds for all but countably many y. Hence, with (243.2), we have
(243.4) N(y) = lim
i
_
IP
i
f(I)
(y)
_
for all but countably many y.
Since
IP
i
f(I)
is a Borel measurable function, it follows that N is also Borel
measurable. For any interval I P
i
,
[f(I)] =
_
R
1
f(I)
(y) d(y)
244 7. DIFFERENTIATION
and therefore by (243.3),
IP
i
[f(I)] =
IP
i
_
R
1
f(I)
(y) d(y)
=
_
R
1
IP
i
f(I)
(y) d(y)
_
R
1
N(y) d(y).
which implies
(244.1) limsup
i
_
IP
i
[f(I)]
_
_
R
1
N(y) d(y).
For the opposite inequality, observe that Fatous lemma and (243.4) yield
liminf
i
_
IP
i
[f(I)]
_
= liminf
i
_
IP
i
_
R
1
f(I)
(y) d(y)
_
= liminf
i
_
_
R
1
IP
i
f(I)
(y) d(y)
_
_
R
1
_
liminf
i
IP
i
f(I)
(y)
_
d(y)
=
_
R
1
N(y) d(y).
Thus, we have
(244.2) lim
i
_
IP
i
[f(I)]
_
=
_
R
1
N(y) d(y).
We will conclude the proof by showing that the limit on the left-side is equal to
V
f
(b). First, recall the notation introduced in Notation 238.1: If I is an interval
belonging to a partition T
i
, we will denote the endpoints of this interval by a
I
, b
I
.
Thus, I = [a
I
, b
I
]. We now proceed with the proof by selecting a sequence of
partitions T
i
with the property that each subinterval I in T
i
has length less than
1
i
and
lim
i
IP
i
[f(b
i
) f(a
i
)[ = V
f
(b).
Then,
V
f
(b) = lim
i
_
IP
i
[f(b
I
) f(a
I
[)
_
liminf
i
_
IP
i
[f(I)]
_
.
7.6. VARIATION OF CONTINUOUS FUNCTIONS 245
We now show that
(244.3) limsup
i
_
IP
i
[f(I)]
_
V
f
(b),
which will conclude the proof. For this, let I
= [a
I
, b
I
] be an interval contained in
I = [a
i
, b
i
] such that f assumes its maximum and minimum on I at the endpoints
of I
. Let Q
i
denote the partition formed by the endpoints of I T
i
along with
the endpoints of the intervals I
. Then
IP
i
[f(I)] =
IP
i
[f(b
I
) f(a
I
)[
IQ
i
[f(b
i
) f(a
i
)[
V
f
(b),
thereby establishing (244.3).
245.1. Corollary. Suppose f is a continuous function of bounded variation
on [a, b]. Then the total variation function V
f
() is continuous on [a, b]. In addition,
if f also satises condition N, then so does V
f
().
Proof. Fix x
0
[a, b]. By the previous result
V
f
(x
0
) =
_
R
N(f, [a, x
0
], y) d(y).
For a x
0
< x b,
N(f, [a, x], y) N(f, [a, x
0
], y) = N(f, (x
0
, x], y)
for each y such that N(f, [a, b], y) < , i.e., for a.e. y R because N(f, [a, b], ) is
integrable. Thus
(245.1)
0 V
f
(x) V
f
(x
0
) =
_
R
N(f, [a, x], y) d(y)
_
R
N(f, [a, x
0
], y) d(y)
=
_
R
N(f, (x
0
, x], y) d(y)
_
f((x
0
,x])
N(f, [a, b], y) d(y).
Since f is continuous at x
0
, [f((x
0
, x])] 0 as x x
0
and thus
lim
xx
+
0
V
f
(x) = V
f
(x
0
).
246 7. DIFFERENTIATION
A similar argument shows that
lim
xx
0
V
f
(x) = V
f
(x
0
).
and thus that V
f
is continuous at x
0
.
Now assume that f also satises condition N and let A be a set with (A) = 0
so that (f(A)) = 0. Observe that for any measurable set E R, the set function
(E) :=
_
E
N(f, [a, b], y) dy
is a measure that is absolutely continuous with respect to Lebesgue measure. Con-
sequently, for each > 0 there exists > 0 such that (E) < whenever (E) < .
Hence if U f(A) is an open set with (f(A)) < , we have
(246.1)
_
U
N(f, [a, b], y) dy < .
The open set f
1
(U) can be expressed as the countable disjoint union of intervals
i=1
I
i
A. Thus, with the notation I
i
:= [a
i
, b
i
] we have
(V
f
(A))
_
V
f
(
i
I
i
)
_
i=1
(V
f
(I
i
))
=
i=1
V
f
(b
i
) V
f
(a
i
)
=
i=1
_
R
N(f, (a
i
, b
i
], y) dy by (245.1)
=
_
i=1
f((a
i
,b
i
])
N(f, [a, b], y) dy since the I
i
s are disjoint
=
_
U
N(f, [a, b], y),
< by (246.1)
which implies that (V
f
(A)) = 0.
246.1. Theorem. If f is a nondecreasing function dened on [a, b], then it is
absolutely continuous if and only if the following two conditions are satised:
(i) f is continuous,
(ii) f satises condition N.
Proof. Clearly condition (i) is necessary for absolute continuity and Theorem
239.2 shows that condition (ii) is also necessary.
7.7. CURVE LENGTH 247
To prove that the two conditions are sucient, let g(x) := x+f(x), so that g is
a strictly increasing, continuous function. Note that g satises condition N since f
does and (g(I)) = (I) +(f(I)) for any interval I. Also, if E is a measurable set,
then so is g(E) because E = F N where F is an F
= sup
k
i=1
[(t
i
) (t
i1
)[,
where the supremum is taken over all nite sequences a = t
0
< t
1
< < t
k
= b.
Note that (x) is a vector in R
n
for each x [a, b]; writing (x) in terms of its
component functions we have
(x) = (
1
(x),
2
(x), . . . ,
n
(x)).
Thus, in (248.1) (t
i
) (t
i1
) is a vector in R
n
and [(t
i
) (t
i1
)[ is its length.
For x [a, b], we will use the notation L
(b) < .
We will show that there is a strong parallel between the notions of length and
bounded variation. In case is a curve in R, i.e, in case : [a, b] R, the two
notions coincide. More generally, we have the following.
248.1. Theorem. A continuous curve : [a, b] R
n
is rectiable if and only
if each component function,
i
, is of bounded variation on [a, b].
Proof. Suppose each component function is of bounded variation. Then there
are numbers M
1
, M
2
, , M
n
such that for any nite partition T of [a, b] into
nonoverlapping intervals I = [a
I
, b
I
],
IP
[
j
(b
I
)
j
(a
I
)[ M
j
, j = 1, 2, . . . , n.
Thus, with M = M
1
+M
2
+ +M
n
we have
IP
[(b
I
) (a
I
)[
=
IP
_
(
1
(b
I
)
1
(a
I
))
2
+ (
2
(b
I
)
2
(a
I
))
2
+ + (
n
(b
I
)
n
(a
I
))
2
1/2
IP
[
1
(b
I
)
1
(a
I
)[ +
IP
[
2
(b
I
)
2
(a
I
)[
+ +
IP
[
n
(b
I
)
n
(a
I
)[
M.
Since the partition T is arbitrary, we conclude that the length of is less than or
equal to M.
7.7. CURVE LENGTH 249
Now assume that is rectiable. Then for any partition T of [a, b] and any
integer j [1, n],
IP
[(b
I
) (a
I
)[
IP
[
j
(b
I
)
j
(a
I
)[ ,
thus showing that the total variation of
j
is no more than the length of .
In elementary calculus, we know that the formula for the length of a curve
= (
1
,
2
) dened on [a, b] is given by
L
=
_
b
a
_
(
1
(t))
2
+ (
2
(t))
2
dt.
We will proceed to investigate the conditions under which this formula holds using
our denition of length. We will consider a curve in R
n
; thus we have : [a, b]
R
n
with (t) = (
1
(t),
2
(t), . . . ,
n
(t)) and we recall the notation L
(t) introduced
earlier that denotes the length of from a to t.
249.1. Theorem. If is rectiable, then
L
(t) =
_
(
1
(t))
2
+ (
2
(t))
2
+ + (
n
(t))
2
for almost all t [a, b].
Proof. The number [L
(t +h) L
(t +h) L
,
L
i=1
[(t
i
) (t
i1
)[ .
Since
L
(x) = sup
_
k
i=1
[(t
i
) (t
i1
)[
_
over all partitions a = t
0
< t
1
< < t
k
= x, it follows
L
(x).
A similar inequality holds for h < 0 and therefore we obtain
(249.2) [L
(t +h) L
(t +h) L
1/2
whenever t +h, t [a, b]. Consequently,
[L
(t +h) L
(t)[
h
_
_
1
(t +h)
1
(t)
h
_
2
+ +
_
n
(t +h)
n
(t)
h
_
2
_
1/2
.
Taking the limit as h 0, we obtain
(250.1) L
(t)
_
(
1
(t))
2
+ (
2
(t))
2
+ + (
n
(t))
2
whenever all derivatives exist, which is almost everywhere in view of Theorem
248.1. The remainder of the proof is completely analogous to the proof of (234.3)
in Theorem 234.1 and is left as an exercise.
The proof of the following result is completely analogous to the proof of Theo-
rem 241.3 and is also left as an exercise.
250.1. Theorem. If each component function
i
of the curve : [a, b] R
n
is
absolutely continuous, the function L
(x) =
_
x
a
L
d =
_
x
a
_
(
1
)
2
+ (
2
)
2
+ + (
n
)
2
d
for each x [a, b].
250.2. Example. Intuitively, one might expect that the trace of a continuous
curve would resemble a piece of string, perhaps badly crumpled, but still like a piece
of string. However, Peano discovered that the situation could be far worse. He was
the rst to demonstrate the existence of a continuous mapping : [0, 1] R
2
such
that [0, 1] occupies the unit square, Q. In other words, he showed the existence
of an area-lling curve. In the gure below, we show the rst three stages of the
construction of such a curve. This construction is due to Hilbert.
Each stage represents the graph of a continuous (piecewise linear) mapping
k
: [0, 1] R
2
. From the way the construction is made, we nd that
sup
t[0,1]
[
k
(t)
l
(t)[
2
2
k
7.7. CURVE LENGTH 251
where k l. Hence, since the space of continuous functions with the topology of
uniform convergence is complete, there exists a continuous mapping that is the
uniform limit of
k
. To see that [0, 1] is the unit square Q, rst observe that
each point x
0
in the unit square belongs to some square of each of the partitions of
Q. Denoting by T
k
the k
th
partition of Q into 4
k
subsquares of side-length (1/2)
k
,
we see that each point x
0
Q belongs to some square, Q
k
, of T
k
for k = 1, 2, . . ..
For each k the curve
k
passes through Q
k
, and so it is clear that there exist points
t
k
[0, 1] such that
(251.0) x
0
= lim
k
k
(t
k
).
For > 0, choose K
1
such that [x
0
k
(t
k
)[ < /2 for k K
1
. Since
k
uniformly, we see by Theorem 61.2, p.62, that there exists = () > 0 such that
[
k
(t)
k
(s)[ < for all k whenever [s t[ < . Choose K
2
such that [t t
k
[ <
for k K
2
. Since [0, 1] is compact, there is a point t
0
[0, 1] such that (for a
subsequence) t
k
t
0
as k . Then x
0
= (t
0
) because
[x
0
(t
0
)[ [x
0
k
(t
k
)[ +[
k
(t
k
)
k
(t
0
)[ <
for k max K
1
, K
2
. One must be careful to distinguish a curve in R
n
from the
point set described by its trace. For example, compare the curve in R
2
given by
(x) = (cos x, sinx), x [0, 2]
to the curve
(x) = (cos 2x, sin2x), x [0, 2].
Their traces occupy the same point set, namely, the unit circle. However, the length
of is 2 whereas the length of is 4. This simple example serves as a model
for the relationship between the length of a curve and the 1-dimensional Hausdor
measure of its trace. Roughly speaking, we will show that they are the same if
one takes into account the number of times each point in the trace is covered. In
particular, if is injective, then they are the same.
251.1. Theorem. Let : [a, b] R
n
be a continuous curve. Then
(251.1) L
(b) =
_
N(, [a, b], y) dH
1
(y)
where N(, [a, b], y) denotes the (possibly innite) number of points in the set
1
y [a, b]. Equality in (251.1) is understood in the sense that either both sides
are nite and equal or both sides are innite.
252 7. DIFFERENTIATION
Proof. Let M denote the length of on [a, b]. We will rst prove
(251.2) M
_
N(, [a, b], y) dH
1
(y),
and therefore, we may as well assume M < . The function, L
(), is nondecreas-
ing, and its range is the interval [0, M]. As in (249.3), we have
(251.3) [L
(x) L
s is an interval
(possibly degenerate) for all s [0, M]. In fact, there are only countably many s for
which L
1
s is a nondegenerate interval
and if x
1
, x L
1
s, then (x
1
) = (x), thus ensuring that g(s) is well-dened.
Also, for x L
1
s, y L
1
(x) L
(y)[ = [s t[ ,
so that g is a Lipschitz function with Lipschitz constant 1. From the denition of
g, we clearly have = g L
s
is a nondegenerate interval. Then it follows that
N(, [a, b], y) = N(g, [0, M], y)
for all y , g(S). Since S is countable and therefore also g(S), we have
(252.3)
_
N(, [a, b], y) dH
1
(y) =
_
N(g, [0, M], y) dH
1
(y).
We will appeal to the proof of Theorem 242.1 to show that
M
_
N(g, [0, M], y) dH
1
(y).
Let T
i
be a sequence of partitions of [0, M] each having the property that its
intervals have length less than 1/i. Then, using the argument that established
(243.4), we have
N(g, [0, M], y) = lim
i
_
IP
i
g(I)
(y)
_
.
Since
H
1
(g(I)) =
_
g(I)
(y) dH
1
(y),
for each interval I, we adapt the proof of (244.2) to obtain
(252.4) lim
i
_
IP
i
H
1
[g(I)]
_
=
_
N(g, [0, M], y) dH
1
(y).
7.7. CURVE LENGTH 253
Now use (252.2) and Exercise 7.27 to conclude that H
1
[g(I)] (I) so that (252.4)
yields
(252.5) liminf
i
_
IP
i
(I)
_
_
0
N(g, [0, M], y) dH
1
(y).
It is necessary to use the liminf here because we dont know that the limit exists.
However, since
IP
i
(I) = ([0, M]) = M,
we see that the limit does exist and that the left-side of (252.5) equals M. Thus,
we obtain
M
_
0
N(g, [0, M], y) dH
1
(y),
which, along with (252.3), establishes (251.2).
We now will prove
(253.1) M
_
0
N(, [a, b], y) dH
1
(y)
to conclude the proof of the theorem. Again, we adapt the reasoning leading to
(252.4) to obtain
(253.2) lim
i
_
IP
i
H
1
[(I)]
_
=
_
N(, [a, b], y) dH
1
(y).
In this context, T
i
is a sequence of partitions of [a, b], each of whose intervals has
maximum length 1/i. Each term on the left-side involves H
1
[(I)]. Now (I) is the
trace of a curve dened on the interval I = [a
I
, b
I
]. By Exercise 7.27, we know that
H
1
[(I)] is greater than or equal to the H
1
measure of the orthogonal projection
of (I) onto any straight line, l. That is, if p: R
n
l is an orthogonal projection,
then
H
1
[p((I))] H
1
[(I)].
In particular, consider the straight line l that passes through the points (a
I
) and
(b
I
). Since I is connected and is continuous, (I) is connected and therefore,
its projection, p[(I)], onto l is also connected. Thus, p[(I)] must contain the
interval with endpoints (a
I
) and (b
I
). Hence [(b
I
) (a
I
)[ H
1
[(I)]. Thus,
from (253.2), we have
lim
i
_
IP
i
[(b
I
) (a
I
)[
_
_
N(, [a, b], y) dH
1
(y).
By Exercise 7.28 the expression on the left is the length of on [a, b], which is M,
thus proving (253.1).
254 7. DIFFERENTIATION
253.1. Remark. The function g dened in (252.1) is called a para-metrization
of with respect to arc-length. The purpose of g is to give an alternate and
equivalent description of the curve . It is equivalent in the sense that the trace and
length of g are the same as those of . In case is not constant on any interval,
then L
(x) = 0.
Then [f(E)] = 0.
Now we will investigate the behavior of a function on the complement of its
critical set and show that good things happen there. We will prove that the set on
which a continuous function has a non-zero derivative can be decomposed into a
countable collection of disjoint sets on each of which the function is bi-Lipschitzian.
That is, on each of these sets the function is Lipschitz and injective; furthermore,
its inverse is Lipschitz on the image of each such set.
7.8. THE CRITICAL SET OF A FUNCTION 255
254.2. Theorem. Suppose f is dened on [a, b] and let A be dened by
A := [a, b] x : f
(x) ,= 0.
Then for each > 1, there is a countable collection E
k
of disjoint Borel sets such
that
(i) A =
k=1
E
k
(ii) For each positive integer k there is a positive rational number r
k
such that
r
k
[f
(x)[ r
k
for x E
k
,
[f(y) f(x)[ r
k
[x y[ for x, y E
k
, (255.1)
[f(y) f(x)[
r
k
(y x) for x, y E
k
. (255.2)
Proof. Since > 1 there exists > 0 so that
1
+ < 1 < .
For each positive integer k and each positive rational number r let A(k, r) be the
set of all points x A such that
_
1
+
_
r [f
(x)[ ( )r
With the help of the triangle inequality, observe that if x, y [a, b] with x A(k, r)
and [y x[ < 1/k, then
[f(y) f(x)[ r [(y x)[ +[f
(x)(y x)[ r [y x[
and
[f(y) f(x)[ r [y x[ +[f
(x)(y x)[
r
[y x[ .
Clearly,
A =
A(k, r)
where the union is taken as k and r range through the positive integers and positive
rationals, respectively. To ensure that (255.1) and (255.2) hold for all a, b E
k
(dened below), we express A(k, r) as the countable union of sets each having
diameter 1/k by writing
A(k, r) =
sQ
[A(k, r) I(s, 1/k)]
256 7. DIFFERENTIATION
where I(s, 1/k) denotes the open interval of length 1/k centered at the rational
number s. The sets [A(k, r) I(s, 1/k)] constitute a countable collection as k, r,
and s range through their respective sets. Relabel the sets [A(k, r) I(s, 1/k)] as
E
k
, k = 1, 2, . . . . We may assume the E
k
are disjoint by appealing to Lemma 82.1,
p.82, thus obtaining the desired result.
The next result diers from the preceding one only in that the hypothesis now
allows the set A to include critical points of f. There is no essential dierence in
the proof.
256.1. Theorem. Suppose f is dened on [a, b] and let A be dened by
A = [a, b] x : f
(x) exists.
Then for each > 1, there is a countable collection E
k
of disjoint sets such that
(i) A =
k=1
E
k
,
(ii) For each positive integer k there is a positive rational number r such that
[f
(x)[ r for x E
k
,
and
[f(y) f(x)[ r [y x[ for x, y E
k
,
Before giving the proof of the next main result, we take a slight diversion that
is concerned with the extension of Lipschitz functions. We will give a proof of
a special case of Kirzbrauns Theorem whose general formulation we do not
require and is more dicult to prove.
256.2. Theorem. Let A R be an arbitrary set and suppose f : A R is a
Lipschitz function with Lipschitz constant C. Then there exists a Lipschitz function
f(x) = inff(a) +C [x a[ : a A.
Clearly,
f = f on A because if b A, then
f(b) f(a) C [b a[
for any a A. This shows that f(b)
f(b). On the other hand,
f(b) f(b)
follows immediately from the denition of
f. Finally, to show that
f has Lipschitz
7.8. THE CRITICAL SET OF A FUNCTION 257
constant C, let x, y R. Then
(x) ,= 0.
Then, for every Lebesgue measurable set E A,
(257.1)
_
E
[f
[ d =
_
N(f, E, y) d(y).
Equality is understood in the sense that either both sides are nite and equal or both
sides are innite.
Proof. We apply Theorem 254.2 with A replaced by E. Thus, for each k N
there is a positive rational number r such that
r
(E
k
)
_
E
k
[f
[ d r(E
k
)
and f restricted to E
k
satises a Lipschitz condition with constant r. Therefore by
Lemma 256.2, f has a Lipschitz extension to R with the same Lipschitz constant.
From Exercise 7.17, we obtain
(257.2) [f(E
k
)] r(E
k
).
Theorem 254.2 states that f restricted to E
k
is univalent and that its inverse
function is Lipschitz with constant /r. Thus, with the same reasoning as before,
(E
k
)
r
[f(E
k
)].
Hence, we obtain
(257.3)
1
2
(f[E
k
])
_
E
k
[f
[ d
2
[f(E
k
)].
Each E
k
can be expressed as the countable union of compact sets and a set of
Lebesgue measure zero. Since f restricted to E
k
satises a Lipschitz condition, f
maps the set of measure zero into a set of measure zero and each compact set is
258 7. DIFFERENTIATION
mapped into a compact set. Consequently, f(E
k
) is the countable union of compact
sets and a set of measure zero and therefore, Lebesgue measurable. Let
g(y) =
k=1
f(E
k
)
(y),
so that g(y) is the number of sets f(E
k
) that contain y. Observe that g is
Lebesgue measurable. Since the sets E
k
are disjoint, their union is E, and f
restricted to each E
k
is univalent, we have
g(y) = N(f, E, y).
Finally,
1
k=1
(f[E
k
])
_
E
[f
[ d
2
k=1
[f(E
k
)]
and, with the aid of Corollary 159.3
_
N(f, E, y) d(y) =
_
g(y) d(y)
=
_
k=1
f(E
k
)
(y) d(y)
=
k=1
_
f(E
k
)
(y) d(y)
=
k=1
[f(E
k
)].
The result now follows from (257.3) since > 1 is arbitrary.
258.1. Corollary. If f satises (257.1), then f satises condition N on the
set A.
We conclude this section with a another result concerning absolute continuity.
Theorem 256.1 states that a function possesses some regularity properties on the
set where it is dierentiable. Therefore, it seems reasonable to expect that if f is
dierentiable everywhere in its domain of denition, then it will have to be a nice
function. This is the thrust of the next result.
258.2. Theorem. Suppose f : (a, b) R has the property that f
exists every-
where and f
[ d =
_
N(f, E
k
, y) d(y)
for k = 1, 2, . . . . Hence,
_
b
a
[f
[ d =
_
> L(x
0
), we conclude that t is not a
lower bound for T and therefore that L(x
0
) is the greatest lower bound; that is,
(260.2) limsup
xx
0
f(x) = L(x
0
) = inf T.
With the above serving as motivation, we proceed with the measure-theoretic
counterpart of (260.2): If f is a Lebesgue measurable function dened on R
n
, the
upper (lower) approximate limit of f at a point x
0
is dened by
aplimsup
xx
0
f(x) = inft : D(A
t
, x
0
) = 0
(apliminf
xx
0
f(x) = supt : D(B
t
, x
0
) = 0
We speak of the approximate limit of f at x
0
when
aplimsup
xx
0
f(x) = apliminf
xx
0
f(x)
and f is said to be approximately continuous at x
0
if
ap lim
xx
0
f(x) = f(x
0
).
Note that if g = f a.e., then the sets A
t
and B
t
corresponding to g dier from
those for f by at most a set of Lebesgue measure zero, and thus the upper and
lower approximate limits of g coincide with those of f everywhere.
In topology, a point x is interior to a set E if there is a ball B(x, r) E. In
other words, x is interior to E if it is completely surrounded by other points in E.
In measure theory, it would be natural to say that x is interior to E (in the measure
theoretic sense) if D(E, x) = 1. See Exercise 7.36.
7.9. APPROXIMATE CONTINUITY 261
The following is a direct consequence of Theorem 226.1, which implies that
almost every point of a measurable set is interior to it (in the measure-theoretic
sense).
261.0. Theorem. If E R
n
is a Lebesgue measurable set, then
D(E, x) = 1 for -almost all x E,
D(E, x) = 0 for -almost all x
E.
Recall that a function f is continuous at x if for every open interval I containing
f(x), x is interior to f
1
(I). This remains true in the measure-theoretic context.
261.1. Theorem. Suppose f : R
n
R is Lebesgue measurable function. Then
f is approximately continuous at x if and only if for every open interval I containing
f(x), D[f
1
(I), x] = 1.
Proof. Assume f is approximately continuous at x and let I be an arbitrary
open interval containing f(x). We will show that D[f
1
(I), x] = 1. Let J = (t
1
, t
2
)
be an interval containing f(x) whose closure is contained in I. From the denition
of approximate continuity, we have
D(A
t
2
, x) = D(B
t
1
, x) = 0
and therefore
D(A
t
2
B
t
1
, x) = 0.
Since
R
n
f
1
(I) A
t
2
B
t
1
,
it follows that D(R
n
f
1
(I)) = 0 and therefore that D[f
1
(I), x] = 1 as desired.
For the proof in the opposite direction, assume D(f
1
(I), x) = 1 whenever I
is an open interval containing f(x). Let t
1
and t
2
be any numbers t
1
< f(x) < t
2
.
With I = (t
1
, t
2
) we have D(f
1
(I), x) = 1. Hence D[R
n
f
1
(I), x] = 0. This
implies D(A
t
2
, x) = D(B
t
1
, x) = 0, which implies that the approximate limit of f
at x is f(x).
The next result shows the great similarity between continuity and approximate
continuity.
261.2. Theorem. f is approximately continuous at x if and only if there exists
a Lebesgue measurable set E containing x such that D(E, x) = 1 and the restriction
of f to E is continuous at x.
262 7. DIFFERENTIATION
Proof. We will prove only the dicult direction. The other direction is left to
the reader. Thus, assume that f is approximately continuous at x. The denition of
approximate continuity implies that there are positive numbers r
1
> r
2
> r
3
> . . .
tending to zero such that
_
B(x, r)
_
y : [f(y) f(x)[ >
1
k
__
<
[B(x, r)]
2
k
, for r r
k
.
Dene
E = R
n
k=1
_
B(x, r
k
) B(x, r
k+1
)
_
y : [f(y) f(x)[ >
1
k
__
.
From the denition of E, it follows that the restriction of f to E is continuous at
x. In order to complete the assertion, we will show that D(
E, x) = 0. For this
purpose, choose > 0 and let J be such that
k=J
1
2
k
< . Furthermore, choose
r such that 0 < r < r
J
and let K J be that integer such that r
K+1
r < r
K
.
Then,
[(R
n
E) B(x, r)]
_
B(x, r) y : [f(y) f(x)[ >
1
K
_
+
k=K+1
_
B(x, r
k
) B(x, r
k+1
)
_
y : [f(y) f(x)[ >
1
k
__
[B(x, r)]
2
K
+
k=K+1
[B(x, r
k
)]
2
k
[B(x, r)]
2
K
+
k=K+1
[B(x, r)]
2
k
[B(x, r)]
k=K
1
2
k
[B(x, r)] ,
which yields the desired result since is arbitrary.
262.1. Theorem. Assume f : R
n
R is Lebesgue measurable. Then f is
approximately continuous - almost everywhere.
Proof. First, we will prove that there exist disjoint compact sets K
i
R
n
such that
_
R
n
i=1
K
i
_
= 0
EXERCISES FOR CHAPTER 7 263
and f restricted to each K
i
is continuous. To this end, set B
i
= B(0, i) for each
positive integer i . By Lusins Theorem, there exists a compact set K
1
B
1
with (B
1
K
1
) 1 such that f restricted to K
1
is continuous. Assuming that
K
1
, K
2
, . . . , K
j
have been constructed, appeal to Lusins Theorem again to obtain
a compact set K
j+1
such that
K
j+1
B
j+1
j
i=1
K
i
,
_
B
j+1
j+1
i=1
K
i
_
1
j + 1
,
and f restricted to K
j+1
is continuous. Let
E
i
= K
i
x : D(K
i
, x) = 1
and recall from Theorem 228.1 that (K
i
E
i
) = 0. Thus, E
i
has the property that
D(E
i
, x) = 1 for each x E
i
. Furthermore, f restricted to E
i
is continuous since
E
i
K
i
. Hence, by Theorem 261.2 we have that f is approximately continuous at
each point of E
i
and therefore at each point of
i=1
E
i
.
Since
_
R
n
i=1
E
i
_
=
_
R
n
i=1
K
i
_
= 0,
we obtain the conclusion of the theorem.
Exercises for Chapter 7
Section 7.1
7.1 (a) Prove for each open set U R
n
there exists a collection T of disjoint closed
balls contained in U such that
(U
B : B T) = 0.
(b) Thus, U =
BF
B N where (N) = 0. Prove that N ,= .
7.2 Let be a nite Borel measure on R
n
with the property that [B(x, 2r)]
c[B(x, r)] for all x R
n
and all 0 < r < , where c is a constant independent
of x and r. Prove that the Vitali covering theorem, Theorem 223.2, is valid
with replaced by .
7.3 Supply the proof of Theorem 219.1.
7.4 Prove the following alternate version of Theorem 223.2. Let E R
n
be an
arbitrary set, possibly nonmeasurable. Suppose ( is a family of closed cubes
with the property that for each x E and each > 0 there exists a cube C (
264 7. DIFFERENTIATION
containing x whose diameter is less than . Prove that there exists a countable
disjoint subfamily T ( such that
(E
C : C T) = 0.
7.5 With the help of the preceding exercise, prove that for each open set U R
n
,
there exists a countable family T of closed, disjoint cubes, each contained in U
such that
(U
C : C T) = 0.
Section 7.2
7.6 Let f be a measurable function dened on R
n
with the property that, for some
constant C, f(x) C [x[
n
for [x[ 1. Prove that f is not integrable on R
n
.
Hint: One way to proceed is to use Theorem 200.1.
7.7 Let f L
1
(R
n
) be a function that does not vanish identically on B(0, 1).
Show that Mf , L
1
(R
n
) by establishing the following inequality for all x with
[x[ > 1:
Mf(x)
_
B(x,|x|+1)
[f[ d
1
C([x[ + 1)
n
_
B(0,1)
[f[ d
>
1
2
n
C [x[
n
_
B(0,1)
[f[ d,
where C = [B(0, 1)].
7.8 Prove that the maximal function Mf is not integrable on R
n
unless f is iden-
tically 0. (cf. Exercise 7.7)
7.9 Let f L
1
loc
(R
n
). Prove for each xed r > 0, that
_
B(x,r)
[f[ d
is a continuous function of x.
Section 7.3
7.10 Let f : R R be a nondecreasing function. Prove that there is a right-
continuous, nondecreasing function g that agrees with f everywhere except
possibly for a countable set.
7.11 Here is an outline of an alternative proof of Theorem 231.1, which states that
a nondecreasing function f dened on (a, b) is dierentiable (Lebesgue) almost
everywhere. For this, we introduce the Dini Derivatives:
EXERCISES FOR CHAPTER 7 265
D
+
f(x
0
) = limsup
h0
+
f(x
0
+h) f(x
0
)
h
D
+
f(x
0
) = liminf
h0
+
f(x
0
+h) f(x
0
)
h
D
f(x
0
) = limsup
h0
f(x
0
+h) f(x
0
)
h
D
f(x
0
) = liminf
h0
f(x
0
+h) f(x
0
)
h
If all four Dini derivatives are nite and equal, then f
(x
0
) exists and is equal
to the common value. Clearly, D
+
f(x
0
) D
+
f(x
0
) and D
f(x
0
) D
f(x
0
).
To prove that f
f(x)
has measure zero. A similar argument would apply to any two Dini derivatives.
For any two rationals r, s > 0, let
E
r,s
= x : D
+
f(x) > r > s > D
f(x).
The proof reduces to showing that E
r,s
has measure zero. If x E
r,s
, then
there exist arbitrarily small positive h such that
f(x h) f(x)
h
< s.
Fix > 0 and use the Vitali Covering Theorem to nd a countable family of
closed, disjoint intervals [x
k
h
k
, x
k
], k = 1, 2, . . . , such that
f(x
k
) f(x
k
h
k
) < sh
k
,
(E
r,s
k=1
[x
k
h
k
, x
k
]
_
= (E
r,s
),
m
k=1
h
k
< (1 +)(E
r,s
).
From this it follows that
k=1
[f(x
k
) f(x
k
h
k
)] < s(1 +)(E
r,s
).
For each point
y A := E
r,s
k=1
(x
k
h
k
, x
k
)
_
,
266 7. DIFFERENTIATION
there exist arbitrarily small h > 0 such that
f(y +h) f(y)
h
> r.
Employ the Vitali Covering Theorem again to obtain a countable family of
disjoint, closed intervals [y
j
, y
j
+h
j
] such that
each [y
j
, y
j
+h
j
] lies in some [x
k
h
k
, x
k
],
f(y
j
+h
j
) f(y
j
) > rh
j
j = 1, 2, . . . ,
j=1
h
j
(A).
Since f is nondecreasing, it follows that
j=1
[f(y
j
+h
j
) f(y
j
)]
k=1
[f(x
k
) f(x
k
h
k
)],
and from this a contradiction is readily reached.
Section 7.4
7.12 Prove that a function of bounded variation is a Borel measurable function.
7.13 If f is of bounded variation on [a, b], Theorem 233.2 states that f = f
1
f
2
where both f
1
and f
2
are nondecreasing. Prove that
f
1
(x) = sup
_
k
i=1
(f(t
i
) f(t
i1
))
+
_
where the supremum is taken over all partitions a = t
0
< t
1
< < t
k
= x.
Section 7.5
7.14 Prove directly from the denition that the Cantor-Lebesgue function is not
absolutely continuous.
7.15 Prove that a Lipschitz function on [a, b] is absolutely continuous.
7.16 Prove directly from denitions that a Lipschitz function satises condition N.
7.17 Suppose f is a Lipschitz function dened on R having Lipschitz constant C.
Prove that [f(A)] C(A) whenever A R is Lebesgue measurable.
7.18 Let f
k
be a uniformly bounded sequence of absolutely continuous functions
on [0, 1]. Suppose that f
k
f in L
1
[0, 1] and that f
k
is Cauchy in L
1
[0, 1].
Prove that f is absolutely continuous on [0, 1].
7.19 Let f be a strictly increasing continuous function dened on (0, 1). Prove that
f
g dx = f(b)g(b) f(a)g(a)
_
b
a
fg
dx.
7.24 Give an example of a function f : (0, 1) R that is dierentiable everywhere
but is not absolutely continuous. Compare this with Theorem 258.2.
7.25 Given a Lebesgue measurable set E, prove that x : D(E, x) = 1 is a Borel
set.
Section 7.7
7.26 In the proof of Theorem 257.1 we established (257.2) by means of Lemma 256.2.
Another way to obtain (257.2) is the following. Prove that if f is Lipschitz on
a set E with Lipschitz constant C, then [f(A)] C(A) whenever A E
is a Lebesgue measurable. This is the same as Exercise 7.17 except that f is
dened only on E, not necessarily on R. First prove that Lebesgue measure
can be dened as follows:
(A) = inf
_
i=1
diam E
i
: A
i=1
E
i
_
where the E
i
are arbitrary sets.
7.27 Suppose f : R
n
R
m
is a Lipschitz mapping with Lipschitz constant C. Prove
that
H
k
[f(E)] C
k
H
k
(E)
for E R
n
.
7.28 Prove that if : [a, b] R
n
is a continuous curve and T
i
is a sequence of
partitions of [a, b] such that
lim
i
max
IP
i
[b
I
a
I
[ = 0,
then
lim
i
IP
i
[(b
I
) (a
I
)[ = L
(b).
7.29 Prove that the example of an area lling curve in Section 7.7 actually has the
unit square as its trace.
268 7. DIFFERENTIATION
7.30 It follows from Theorem 251.1 that the area lling curve is not rectiable. Prove
this directly from the construction of the curve.
7.31 Give an example of a continuous curve that lls the unit cube in R
3
.
7.32 Give a proof of Lemma 250.1.
7.33 Let : [0, 1] R
2
be dened by (x) = (x, f(x)) where f is the Cantor-
Lebesgue function described in Example 136.1 on p.136. Thus, describes the
graph of the Cantor function. Find the length of .
Section 7.8
7.34 Show that the conclusion of Theorem 258.2 still holds if the following assump-
tions are satised: f
+
_
r [f
(x)[ ( )r
Show that A
1
(r) is a Borel set.
(b) Let f be a continuous function on [a, b]. Let
F
1
(x, y) :=
f(y) f(x)
x y
for all x, y [a, b] with a ,= b.
Prove that F
1
is a Borel function on [a, b] [a, b] (x, y) : x = y.
(c) Let F
2
(x, y) := f
is density open. In
particular, it must be shown that E is measurable.
EXERCISES FOR CHAPTER 7 269
7.37 Prove that a function f : R
n
R is approximately continuous at every point
if and only if f is continuous in the density open topology of Exercise 36.
7.38 Let f be an arbitrary function with the property that for each x R
n
there
exists a measurable set E such that D(E, x) = 1 and f restricted to E is
continuous at x. Prove that f is a measurable function
7.39 Suppose f is a bounded measurable function on R
n
that is approximately con-
tinuous at x
0
. Prove that x
0
is a Lebesgue point for f. Hint: Use Denition
259.1.
7.40 Show that if f has a Lebesgue point at x
0
, then f is approximately continuous
at x
0
.
CHAPTER 8
Elements of Functional Analysis
8.1. Normed Linear Spaces
We have already encountered examples of normed linear spaces, namely
the L
p
spaces. Here we introduce the notion of abstract normed linear
spaces and begin the investigation of the structure of such spaces.
271.1. Definition. A a linear space (or vector space) is a set X is endowed
with with two operations, addition and scalar multiplication, that satisfy the
following conditions: for any x, y, z X and , R
(i) x +y = y +x X.
(ii) x + (y +z) = (x +y) +z.
(iii) There is an element 0 X such that x + 0 = x for each x X.
(iv) For each x X there is an element w X such that x +w = 0.
(v) x X.
(vi) (x) = ()x.
(vii) (x +y) = x +y.
(viii) ( +)x = x +y.
(ix) 1x = x.
We note here some immediate consequences of the denition of a linear space.
If x, y, z X and
x +y = x +z,
then, by conditions (i) and (iv), there is a w X such that w +x = 0 and hence
y = 0 +y = (w +x) +y = w + (x +y) = w + (x +z) = (w +x) +z = 0 +z = z.
Thus, in particular, for each x X there is exactly one element w X such that
x+w = 0. We will denote that element by x and we will write y x for y +(x).
If R and x X, then x = (x + 0) = x + 0, from which we can
conclude 0 = 0. Similarly x = ( + 0)x = x + 0x, from which we conclude
0x = 0.
If ,= 0 and x = 0, then
x =
_
_
x =
1
_
x
_
=
1
0 = 0.
271
272 8. ELEMENTS OF FUNCTIONAL ANALYSIS
272.0. Definition. A subset Y of a linear space X is a subspace of X if for
any x, y Y and , R we have x +y Y .
Thus if Y is a subspace of X, then Y is itself a linear space with respect to the
addition and scalar multiplication it inherits from X. The notion of subspace that
we have dened above might more properly be called linear subspace to distinguish
it from the notion of topological subspace in case X is also a topological space. In
this chapter we will use the term subspace in the sense of the denition above.
If we have occasion to refer to a topological subspace we will mention it explicitly.
If S is any nonempty subset of a linear space X, then the set Y of all elements
of X of the form
1
x
1
+
2
x
2
+ +
m
x
m
where m is any positive integer and
j
R, x
j
S for 1 j m is easily seen to
be a subspace of X. The subspace Y will be called the subspace spanned by S. It
is the smallest subspace of X that contains S.
272.1. Definition. If S = x
1
, x
2
, . . . , x
m
is a nite subset of a linear space
X, then S is linearly independent if
1
x
1
+
2
x
2
+ +
m
x
m
= 0
implies
1
=
2
= =
m
= 0. In general, a subset S of X is linearly independent
if every nite subset of S is linearly independent.
Suppose S is a subset of a linear space X. If S is linearly independent and X
is spanned by S, then for any x X there is a nite subset x
i
m
i=1
of S and a
nite sequence of real numbers
i
m
i=1
such that
x =
m
i=1
i
x
i
.
where
i
,= 0 for each i.
Suppose that for some other choice y
j
k
j=1
of elements in S and real numbers
k
j=1
x =
k
j=1
j
y
j
,
where
j
,= 0 for each j. Then
m
i=1
i
x
i
k
j=1
j
y
j
= 0.
If for some i 1, 2, . . . , m there were no j 1.2. . . . , k such that x
i
= y
j
,
then since S is linearly independent we would have
i
= 0. Similarly, for each j
8.1. NORMED LINEAR SPACES 273
1.2. . . . , k there is an i 1, 2, . . . , m such that x
i
= y
j
. Thus the two sequences
x
i
m
i=1
and y
j
k
j=1
must contain exactly the same elements. Renumbering the
j
,
if necessary, we see that
m
i=1
(
i
i
)x
i
= 0.
Since S is linearly independent we have
i
=
i
for 1 i k. Thus each x X
has a unique representation as a nite linear combination of elements of S.
273.1. Definition. A subset of a linear space X that is linearly independent
and spans X is a basis for X. A linear space is nite dimensional if it has a
nite basis.
The proof of the following result is a consequence of the Hausdor Maximal
Principle; see Exercise 8.1.
273.2. Theorem. Every linear space has a basis.
273.3. Examples. (i) The set R
n
is a linear space with respect to the addi-
tion and scalar multiplication dened by
(x
1
, x
2
, . . . , x
n
) + (y
1
, y
2
, . . . , y
n
) = (x
1
+y
1
, . . . , x
n
+y
n
),
(x
1
, x
2
, . . . , x
n
) = (x
1
, x
2
, . . . , x
n
)
(ii) If A is any set, the set of all real-valued functions on A is a linear space with
respect to the addition and scalar multiplication
(f +g)(x) = f(x) +g(x)
(f)(x) = f(x)
(iii) If S is a topological space, then the set C(S) of all real-valued continuous
functions on S is a linear space with respect to the addition and scalar mul-
tiplication dened in (ii).
(iv) If (X, ) is a measure space, then by, Lemma 164.1, L
p
(X, ) is a linear space
for 1 p with respect to the addition and scalar multiplication dened
in (ii).
(v) For 1 p < let l
p
denote the set of all sequences a
k
k=1
of real numbers
such that
k=1
[a
k
[
p
converges. Each such sequence may be viewed as a
real-valued function on the set of positive integers. If addition and scalar are
dened as in example (ii), then each l
p
is a linear space.
274 8. ELEMENTS OF FUNCTIONAL ANALYSIS
(vi) Let l
denote the set of all bounded sequences of real numbers. Then with
respect the addition and scalar multiplication of (v), l
is a linear space.
274.1. Definition. Let X be a linear space. A function || : X R is a
norm on X if
(i) |x +y| |x| +|y| for all x, y X,
(ii) |x| = [[ |x| for all x X and R,
(iii) |x| 0 for each x X,
(iv) |x| = 0 only if x = 0. A real-valued function on X satisfying conditions (i),
(ii), and (iii) is a semi-norm on X. A linear space X equipped with a norm
|| is a normed linear space.
Suppose X is a normed linear space with norm ||. For x, y X set
(x, y) = |x y| .
Then is a nonnegative real-valued function on X X, and from the properties of
|| we see that
(i) (x, y) = 0 if and only if x = y,
(ii) (x, y) = (y, x) for any x, y X,
(iii) (x, z) (x, y) +(y, z) for any x, y, z X.
Thus is a metric on X, and we see that a normed linear space is also a metric
space; in particular, it is a topological space. Functional Analysis (in normed linear
spaces) is essentially the study of the interaction between the algebraic (linear)
structure and the topological (metric) structure of such spaces.
In a normed linear space X we will denote by B(x, r) the open ball with center
at x and radius r, i.e.,
B(x, r) = y X : |y x| < r.
274.2. Definition. A normed linear space is a Banach space if it is a com-
plete metric space with respect to the metric induced by its norm.
274.3. Examples. (i) R
n
is a Banach space with respect to the norm
|(x
1
, x
2
, . . . , x
n
)| = (x
2
1
+x
2
2
+ +x
2
n
)
1
2
.
(ii) For 1 p the linear spaces L
p
(X, ) are Banach spaces with respect to
the norms
8.1. NORMED LINEAR SPACES 275
|f|
L
p
(X,)
= (
_
X
[f[
p
d)
1
p
if 1 p <
|f|
L
p
(X,)
= infM : (x X : [f(x)[ > M) = 0 if p = .
This is a rephrasing of Theorem 168.1.
(iii) For 1 p the linear spaces l
p
of Examples 273.3(v), (vi) are Banach
spaces with respect to the norms
|a
k
|
l
p
= (
k=1
[a
k
[
p
)
1
p
if 1 p < ,
|a
k
|
l
p
= sup
k1
[a
k
[ if p = .
This is a consequence of Theorem 168.1 with appropriate choices of X and .
(iv) If X is a compact metric space, then the linear space C(X) of all continuous
real-valued functions on X is a Banach space with respect to the norm
|f| = sup
xX
[f(x)[ .
If x
k
is a sequence in a Banach space X, the series
k=1
x
k
converges to
x X if the sequence of partial sums s
m
=
m
k=1
x
k
converges to x, i.e.,
|x s
m
| =
_
_
_
_
_
x
m
k=1
x
k
_
_
_
_
_
0
as m .
275.1. Proposition. Suppose X is a Banach space. Then any absolutely con-
vergence series is convergent, i.e., if the series
k=1
|x
k
| converges in R, then the
series
k=1
x
k
converges in X.
Proof. If m > l, then
_
_
_
_
_
m
k=1
x
k
l
k=1
x
k
_
_
_
_
_
=
_
_
_
_
_
m
k=l+1
x
k
_
_
_
_
_
k=l+1
|x
k
| .
Thus if
k=1
|x
k
| converges in R the sequence of partial sums
m
k=1
x
k
m=1
is
Cauchy in X and therefore converges to some element.
275.2. Definitions. Suppose X and Y are linear spaces. A mapping T : X
Y is linear if for any x, y X and , R
T(x +y) = T(x) +T(y).
276 8. ELEMENTS OF FUNCTIONAL ANALYSIS
If X and Y are normed linear spaces and T : X Y is a linear mapping, then
T is bounded if there exists a constant M such that
|T(x)| M |x|
for each x X.
276.1. Theorem. Suppose X and Y are normed linear spaces and T : X Y
is linear. Then
(i) The linear mapping T is bounded if and only if
sup|T(x)| : x X, |x| = 1 < .
(ii) The linear mapping T is continuous if and only if it is bounded.
Proof. (i) If M is a constant such that
|T(x)| M |x|
for each x X, then for any x X with |x| = 1 we have |T(x)| M.
On the other hand, if
K = sup|T(x)| : x X, |x| = 1 < ,
then for any 0 ,= x X we have
|T(x)| = |x|
_
_
_
_
T
_
x
|x|
__
_
_
_
K |x| .
(ii) If T is continuous at 0, then there exists a > 0 such that |T(x)| 1
whenever x B(0, 2). If |x| = 1, then
|T(x)| =
1
|T(x)|
1
.
Thus
sup|T(x)| : x X, |x| = 1
1
.
On the other hand, if there is a constant M such that
|T(x)| M |x|
for each x X, then for any > 0
|T(x)| <
whenever x B(0,
M
). Let x
0
X. If |y x
0
| <
M
, then
|T(y) T(x
0
)| = |T(y x
0
)| < .
Thus T is continuous on X.
8.1. NORMED LINEAR SPACES 277
276.2. Definition. If T : X Y is a linear mapping from a normed linear
space X into a normed linear space Y we set
|T| = sup|T(x)| : x X, |x| = 1.
This choice of notation will be justied by Theorem 278.1, where we will show that
|T| is a norm on an appropriate linear space.
277.1. Proposition. Suppose X and Y are normed linear spaces, T
k
is a
sequence of bounded linear mappings of X into Y , and T : X Y is a mapping
such that
lim
k
|T
k
(x) T(x)| 0
for each x X. Then T is a linear mapping and
|T| liminf
k
|T
k
| .
Proof. For x, y X and , R
|T(x +y) T(x) T(y)| |T(x +y) T
k
(x +y)|
+|(T
k
(x) T(x)) +(T
k
(y) T(y))|
|T(x +y) T
k
(x +y)|
+[[ |T
k
(x) T(x)| +[[ |T
k
(y) T(y)|
for any k 1. Thus T is a linear mapping.
If x X with |x| = 1, then
|T(x)| |T
k
(x)| +|T(x) T
k
(x)| |T
k
| +|T(x) T
k
(x)|
for any k 1. Thus
|T(x)| liminf
k
|T
k
|
for any x X with |x| = 1 and hence
|T| = sup|T(x)| : x X, |x| = 1 liminf
k
|T
k
| .
Suppose X and Y are linear spaces, and let L(X, Y ) denote the set of all linear
mappings of X into Y . For T, S L(X, Y ) and , R dene
(T +S)(x) = T(x) +S(x)
(T)(x) = T(x)
278 8. ELEMENTS OF FUNCTIONAL ANALYSIS
for x X. Note that these are the usual denitions of the sum and scalar
multiple of functions. It is left as an exercise to show that with these operations
L(X, Y ) is a linear space.
278.0. Notation. If X and Y are normed linear spaces, denote by B(X, Y )
the set of all bounded linear mappings of X into Y . Clearly B(X, Y ) is a subspace
of L(X, Y ). In case Y = R we will refer to the elements of L(X, R) as linear
functionals on X.
278.1. Theorem. Suppose X and Y are normed linear spaces and for T
B(X, Y ) set
(278.1) |T| = sup|T(x)| : x X, |x| = 1.
Then (278.1) denes a norm on B(X, Y ). If Y is a Banach space, then B(X, Y ) is
a Banach space with respect to this norm.
Proof. Clearly |T| 0. If |T| = 0, then T(x) = 0 for any x X with
|x| = 1. Thus if 0 ,= x X, then
T(x) = |x| T
_
x
|x|
_
= 0
i.e., T = 0.
If T, S B(X, Y ) and R, then
|T| = sup[[ |T(x)| : x X, |x| = 1 = [[ |T|
and
|T +S| = sup|T(x) +S(x)| : x X, |x| = 1
sup|T(x)| +|S(x)| : x X, |x| = 1
|T| +|S| .
Thus (278.1) denes a norm on B(X, Y ).
Suppose Y is a Banach space and T
k
is a Cauchy sequence in B(X, Y ). Then
|T
k
| is bounded, i.e., there is a constant M such that |T
k
| M for all k.
For any x X and k, m 1
|T
k
(x) T
m
(x)| |T
k
T
m
| |x| .
Thus T
k
(x) is a Cauchy sequence in Y . Since Y is a Banach space there is an
element T(x) Y such that
|T
k
(x) T(x)| 0
8.2. HAHN-BANACH THEOREM 279
as k . In view of Proposition 277.1 we know that T is a linear mapping of X
into Y . Moreover, again by Proposition 277.1
|T| = liminf
k
|T
k
| M
whence T B(X, Y ).
8.2. Hahn-Banach Theorem
In this section we prove the existence of extensions of linear functionals
from a subspace Y of X to all of X satisfying various conditions.
279.1. Theorem (Hahn-Banach Theorem: semi-norm version). Suppose X is
a linear space and p is a semi-norm on X. Let Y be a subspace of X and f : Y R
a linear functional such that f(x) p(x) for all x Y . Then there exists a linear
functional g : X R such that g(x) = f(x) for all x Y and g(x) p(x) for all
x X.
Proof. Let T denote the family of all pairs (W, h) where W is a subspace
with Y W X and h is a linear functional on W such that h = f on Y and
h p on W. For each (W, h) T let
G(W, h) = (x, r) : x W, r = h(x) X R.
Observe that if (W
1
, h
1
), (W
2
, h
2
) T, then G(W
1
, h
1
) G(W
2
, h
2
) if and
only if
W
1
W
2
(279.1)
h
2
= h
1
on W
1
. (279.2)
Set c = G(W, h) : (W, h) T. If T is a subfamily of c that is linearly ordered
by inclusion, set
W
is a subspace with Y W
on W
by setting
h
, h
, h
= x +x
0
: x W
0
, R.
Then W
is a subspace of X containing W
0
. If x, x
W
0
, ,
R and
x +x
0
= x
x
0
,
then
x x
= (
)x
0
.
If
and
= . Thus each
element of W
on W
(x +x
0
) = h
0
(x) +c
for each x W
0
, R. Clearly h
= h
0
on W
0
. We now choose c so that h
p
on W
.
Observe that for x, x
W
0
h
0
(x) +h
0
(x
) = h
0
(x +x
0
+x
x
0
) p(x +x
0
) +p(x
x
0
)
and thus
h
0
(x
) p(x
x
0
) p(x +x
0
) h
0
(x)
for any x, x
W
0
.
In view of the last inequality there is a c R such that
suph
0
(x) p(x x
0
) : x W
0
c infp(x +x
0
) h
0
(x) : x W
0
.
With this choice of c we see that for any x W
0
and ,= 0
h
(x +x
0
) = h
(
x
+x
0
) = (h
0
(
x
) +c);
if > 0, then c p(x +x
0
) h
0
(x) and therefore
h
(x +x
0
) (h
0
(
x
) +p(
x
+x
0
) h
0
(
x
))
= p(
x
+x
0
) = p(x +x
0
).
8.2. HAHN-BANACH THEOREM 281
If < 0, then
h
(x +x
0
) = (h
0
(
x
) +c)
= [[ (h
0
(
x
[[
) c)
[[ (h
0
(
x
[[
) +p(
x
[[
x
0
) h
0
(
x
[[
))
= [[ p(
x
[[
x
0
) = p(x +x
0
).
Thus (W
, h
) T and G(W
0
, h
0
) is a proper subset of G(W
, h
), contradicting
the maximality of G(W
0
, h
0
). This implies our assumption that W
0
,= X must be
false and so it follows that g := h
0
is a linear functional on X such that g = f on
Y and g p on X.
281.1. Remark. The proof of Theorem 279.1 actually gives more than is as-
serted in the statement of the Theorem. A careful reading of the proof shows that
the function p need not be a semi-norm; it suces that p be subadditive, i.e.,
p(x +y) p(x) +p(y)
and positively homogeneous, i.e.,
p(x) = p(x)
whenever 0. In particular p need not be nonnegative.
As an immediate consequence of Theorem 279.1 we obtain
281.2. Theorem (Hahn-Banach Theorem: norm version). Suppose X is a
norm-ed linear space and Y is a subspace of X. If f is a linear functional on Y
and M is a positive constant such that
[f(x)[ M |x|
for each x Y , then there is a linear functional g on X such that g = f on Y and
[g(x)[ M |x|
for each x X.
Proof. Observe that p(x) = M |x| is a semi-norm on X. Thus by Theorem
279.1 there is a linear functional g on X that extends f to X and such that
g(x) M |x|
for each x X. Since g(x) = g(x) and |x| = |x| it follows immediately that
[g(x)[ M |x|
282 8. ELEMENTS OF FUNCTIONAL ANALYSIS
for each x X.
The following is a useful consequence of Theorem 281.2,
282.1. Theorem. Suppose X is a normed linear space, Y is a subspace of X,
and x
0
X such that
= inf
yY
|x
0
y| > 0,
i.e., the distance from x
0
to Y is positive. Then there is a bounded linear functional
f on X such that
f(y) = 0 for all y Y, f(x
0
) = 1 and |f| =
1
.
Proof. We will use the following observation throughout the proof, namely,
that since Y is a vector space, then
= inf
yY
|x
0
y| = inf
yY
|x
0
(y)| = inf
yY
|x
0
+y| .
Set
W = y +x
0
: y Y, R.
Then, as noted in the proof of Theorem 279.1, W is a subspace of X containing
Y and each element of W has a unique representation in the form y + x
0
with
y Y and R. Thus we can dene a linear functional g on W by
g(y +x
0
) = .
If ,= 0 then
|y +x
0
| = [[
_
_
_
y
+x
0
_
_
_ [[
whence
[g(y +x
0
)[
1
|y +x
0
| .
Thus g is a bounded linear functional on W such that g(y) = 0 for y Y , [g(w)[
1
. There is a sequence
y
k
in Y such that
lim
k
|y
k
+x
0
| = .
Let x
k
=
y
k
+x
0
|y
k
+x
0
|
. Then |x
k
| = 1 and
|f(x
k
)| = |g(x
k
)| =
1
|y
k
+x
0
|
for all k, whence |f| =
1
.
8.3. CONTINUOUS LINEAR MAPPINGS 283
8.3. Continuous Linear Mappings
In this section we deduce from the Baire Category Theorem three im-
portant results concerning continuous linear mappings between Banach
spaces.
We rst prove a linear version of the Uniform Boundedness Principle; Theo-
rem 52.1.
283.1. Theorem (Uniform Boundedness Principle). Let T be a family of con-
tinuous linear mappings from a Banach space X into a normed linear space Y such
that
sup
TF
|T(x)| <
for each x X. Then
sup
TF
|T| < .
Proof. Observe that for each T T the real-valued function x |T(x)| is
continuous. In view of Theorem 52.1, p.52, there is a nonempty open subset U of
X and an M > 0 such that
|T(x)| M
for each x U and T T.
Fix x
0
U and let r > 0 be such that B(x
0
, r) U. If z = x
0
+ y B(x
0
, r)
and T T, then
|T(z)| = |T(x
0
) +T(y)| |T(y)| |T(x
0
)|
i.e.,
|T(y)| |T(z)| +|T(x
0
)| 2M
for each y B(0, r).
If x X with |x| = 1, then x B(0, r) for all 0 < < r, in particular for
= r/2. Therefore, for each T T
|T(x)| =
2
r
|T(rx/2|
4M
r
.
Thus
sup
TF
|T|
4M
r
.
k=1
T(B(0, k)).
=
k=1
T(B(0, k)).
Since Y is a complete metric space, the Baire Category Theorem asserts that one
of these closed sets has a non-empty interior; that is, there exist k
0
1, y Y and
> 0 such that
T(B(0, k
0
) B(y, ).
First, we will show that the origin is in the interior of T(B(0, 2)). For this
purpose, note that if z B(
y
k
0
,
k
0
) then
_
_
_
_
z
y
k
0
_
_
_
_
< O
k
0
i.e.,
|k
0
z y| < .
Thus k
0
z B(y, ) T(B(0, k
0
)) which implies that z T(B(0, )). Setting
y
0
=
y
k
0
and
0
=
k
0
we have
B(y
0
,
0
) T(B(0, )).
If w B(0,
0
), then z = y
0
+ w B(y
0
,
0
) and there exist sequences x
k
and
x
k
in B(0, ) such that
|T(x
k
) y
0
| 0
|T(x
k
) z| 0
8.3. CONTINUOUS LINEAR MAPPINGS 285
and hence
|T(x
k
x
k
) w| 0
as k . Thus, since |x
k
x
k
| < 2,
(284.1) B(0,
0
) T(B(0, 2)).
Now we will show that the origin is interior to T(B(0, 2)). So x 0 <
0
<
and let
k
be a decreasing sequence of positive numbers such that
k=1
k
<
0
.
In view of (284.1), for each k 0 there is a
k
> 0 such that
B(0,
k
) T(B(0,
k
)).
We may assume
k
0 as k . Fix y B(0,
0
). Then y is arbirarily close to
elements of T(B(0,
0
)) and therefore there is an x
0
B(0,
0
) such that
|y T(x
0
)| <
1
i.e.,
y T(x
0
) B(0,
1
) T(B(0,
1
)).
Thus there is an x
1
B(0,
1
) such that
|y T(x
0
) T(x
1
)| <
2
whence y T(x
0
) T(x
1
) T(B(0,
2
)). By induction there is a sequence x
k
such that x
k
B(0,
k
) and
_
_
_
_
_
y T(
m
k=0
x
k
)
_
_
_
_
_
<
m+1
.
Thus T(
m
k=0
x
k
) converges to y as m . For all m > 0, we have
m
k=0
|x
k
| <
k=0
k
<
0
,
which implies the series
k=0
x
k
converges absolutely; since X is a Banach space
and since x
k
B(0,
0
) for all k, the series converges to an element x in the closure
of B(0,
0
) which is contained in B(0, 2
0
), see Proposition 275.1. The continuity
of T implies T(x) = y. Since y is an arbitrary point in B(0,
0
), we conclude that
(285.1) B(0,
0
) T(B(0, 2
0
)) T(B(0, 2)),
which shows that the origin is interior to T(B(0, 2).
Finally suppose U is an open subset of X and y = T(x) for some x U. Let
> 0 be such that B(x, ) U. (285.1) states that there exists a > 0 such that
B(0, ) T(B(0, )).
286 8. ELEMENTS OF FUNCTIONAL ANALYSIS
From the linearity of T
T(B(x, )) = T(x) +T(w); w B(0, )
= y +T(w) : w B(0, )
y +z : z B(0, ) = B(y, ).
As an immediate consequence of the Open Mapping Theorem we have the
following corollary.
286.1. Corollary. If T is a one-to-one bounded linear mapping of a Banach
space X onto a Banach space Y , then T
1
: Y X is a bounded linear mapping.
Proof. The existence and linearity of T
1
are evident. If U is an open subset
of X, then the inverse image of U under T
1
is simply T(U), which is open by
Theorem 284.1. Thus, T
1
is bounded, by Theorem 276.1.
286.2. Definition. The graph of a mapping T : X Y is the set
(x, T(x)) : x X X Y.
It is left as an exercise to show that the graph of a continuous linear mapping
T of a normed linear space X into a normed linear space Y is a closed subset of
X Y . The following theorem shows that, when X and Y are Banach spaces, the
converse is true, (cf. Exercise 8.11).
286.3. Theorem (Closed Graph Theorem.). If T is a linear mapping of a
Banach space X into a Banach space Y and the graph of T is closed in X Y ,
then T is continuous.
Proof. For each x X set
|x|
1
:= |x| +|T(x)| .
It is readily veried that | |
1
is a norm on X. Let us show that it is complete.
Suppose x
k
is a Cauchy sequence in X with respect to | |
1
. Then x
k
is a
Cauchy sequence in X and T(x
k
) is a Cauchy sequence in Y . Since X and Y
are Banach spaces there exist x X and y Y such that |x
k
x| 0 and
|T(x
k
) y| 0 as k . Since the graph of T is closed we must have
(x, y) = (x, T(x)).
This implies |T(x
k
) T(x)| 0 as k , and hence that |x x
k
|
1
0 as
k . Thus X is a Banach space with respect to the norm | |
1
.
8.4. DUAL SPACES 287
Consider two copies of X, the rst one with X equipped with norm ||
1
and
the second with X equipped with ||. Then the identity mapping I : (X, ||
1
)
(X, ||) is continuous since |x| |x|
1
for any x X. According to Corollary
286.1, I is an open map which means that the inverse mapping of I is also contin-
uous. Hence, there is a constant C such that
|x| +|T(x)| = |x|
1
C |x|
for all x X. Evidently C 1. Thus
|T(x)| (C 1) |x| .
from which we conclude that T is continuous.
8.4. Dual Spaces
Here we introduce the important concept of the dual space of a normed
linear space and the associated notion of weak topology.
287.1. Definition. Let X be a normed linear space. The dual space X
of X
is the linear space of all bounded linear functionals on X equipped with the norm
|f| = sup[f(x)[ : x X, |x| = 1.
In view of Theorem 278.1 we know that X
is separable.
Proof. Let f
k
k=1
be a countable dense subset of X
such that
f(w) = 0 for all w W and f(x
0
) = 1.
288 8. ELEMENTS OF FUNCTIONAL ANALYSIS
Since f
k
is dense in X
there is a subsequence f
k
j
for which
lim
j
_
_
f
k
j
f
_
_
= 0.
However, since
_
_
x
k
j
_
_
= 1,
_
_
f
k
j
f
_
_
_
_
f
k
j
(x
k
j
) f(x
k
j
)
_
_
=
_
_
f
k
j
(x
k
j
)
_
_
1
2
_
_
f
k
j
_
_
for each j. Thus
_
_
f
k
j
_
_
0 as j , which implies that f = 0, contradicting the
fact that f(x
0
) = 1. Thus W = X.
288.1. Definition. If X and Y are linear spaces and T : X Y is a one-to-
one linear mapping of X onto Y we will call T a linear isomorphism and say that
X and Y are linearly isomorphic. If, in addition, X and Y are normed linear
spaces and |T(x)| = |x| for each x X, then T is an isometric isomorphism
and X and Y are isometrically isomorphic.
Denote by X
dened by
(288.1) (x)(f) = f(x)
for each f X
. Since
[(x)(f)[ |f| |x|
the linear functional (x) is bounded, in fact |(x)| |x|. Thus (x) X
. It
is readily veried that is a bounded linear mapping of X into X
with || 1.
The following result is the key to understanding the relation between X and
X
.
288.2. Proposition. Suppose X is a normed linear space. Then for each
x X
|x| = sup[f(x)[ : f X
, |f| = 1.
Proof. Fix x X. If f X
, |f| = 1 = |x|
for each x X.
8.4. DUAL SPACES 289
288.3. Theorem. The mapping is an isometric isomorphism of X onto
(X).
Proof. In view of Proposition 288.2 we have
|(x)| = sup[f(x)[ : f X
, |f| = 1 = |x|
for each x X.
The mapping is called the natural imbedding of X in X
.
289.1. Definition. A normed linear space X is said to be reexive if
(X) = X
.
Since X
(X, ) (L
p
(X, ))
dened by
(g)(f) =
_
X
gf d
for g L
p
(X, ) and f L
p
(X, ) is an isometric isomorphism of L
p
(X, )
onto (L
p
(X, ))
, then (L
p
(X, ))
. Thus there is an f L
p
(X, )
such that
(g) =
_
X
fg d = (g)(f)
for all g L
p
(X, ). Since (L
p
(X, )) = (L
p
(X, ))
we see that
() = (f) = (f)()
for all (L
p
(X, ))
is isometrically
isomorphic to L
(, ) is not separable, we
see that L
1
(, ) is not reexive.
290 8. ELEMENTS OF FUNCTIONAL ANALYSIS
In addition to the topology induced by the norm on a normed linear space it
is useful to consider a smaller, i.e., weaker topology. The weak topology on a
normed linear space X is the smallest topology on X with respect to which each
f X
for 1 i m
forms a base for T . From this observation it is evident that a sequence x
k
in X
converges weakly (i.e., with respect to the topology .w to x X if and only if
lim
k
f(x
k
) = f(x)
for each f X
.
In order to distinguish the weak topology from the topology induced by the
norm, we will refer to the latter as the strong topology.
290.1. Theorem. Suppose X is a normed linear space and the sequence x
k
. Since f(x
k
) is a convergent sequence in R
sup[f(x
k
)[ : k = 1, 2, . . . < ,
which may be written as
sup
1k<
[(x
k
)(f)[ < .
Since this is true for each f X
and since X
let f
Y
denote the restriction of f to Y . Then evidently
f
Y
Y
and |f
Y
| |f|. For Y
let
Y
: X
R be given by
Y
(f) = (f
Y
)
for each f X
. Then
Y
is a linear functional on X
and
Y
X
since
[
Y
(f)[ = [(f
Y
)[ || |f
Y
| || |f|
for each f X
Y
(f) = (x
0
)(f) = f(x
0
)
for each f X
. If x
0
, Y , then by Theorem 282.1 there exists an f X
such
that f(x
0
) = 1 and f(y) = 0 for each y Y . This implies that f
Y
= 0 and hence
we arrive at the contradiction
1 = f(x
0
) =
Y
(f) = (f
Y
) = 0.
Thus x
0
Y .
For any g Y
such that f
Y
= g. Thus
g(x
0
) = f(x
0
) =
Y
(f) = (f
Y
) = (g).
Thus the image of x
0
under the natural imbedding of Y into Y
is . Since Y
is arbitrary, Y is reexive.
292 8. ELEMENTS OF FUNCTIONAL ANALYSIS
291.2. Theorem. If X is a reexive Banach space, then the closed unit ball
B := x X : |x| 1 is sequentially compact in the weak topology.
Proof. Assume rst that X is separable. Then since X is reexive, X
is
separable and, by Theorem 287.2, X
is separable. Let f
m
m=1
be dense in X
and x
k
B. Since f
1
(x
k
) is bounded in R there be a subsequence x
1
k
of x
k
such that f
1
(x
1
k
) converges in R. Since the sequence f
2
(x
1
k
) is bounded in R
there is a subsequence x
2
k
of x
1
k
such that f
2
(x
2
k
) converges in R. Continuing
in this way we obtain a sequence of subsequences of x
k
such that x
m
k
is a
subsequence of x
m1
k
for m > 1 and f
m
(x
m
k
) converges as k for each
m 1. Set y
k
= x
k
k
. Then y
k
is a subsequence of x
k
such that f
m
(y
k
)
converges as k for each m 1. For arbitrary f X
[f(y
k
) f(y
l
)[ [f(y
k
) f
m
(y
k
)[ +[f
m
(y
k
) f
m
(y
l
)[ +[f
m
(y
l
) f(y
l
)[
|f
m
f| (|y
k
| +|y
l
|) +[f
m
(y
k
) f
m
(y
l
)[
for any k, l, m. Given > 0 there exists an m such that
|f
m
f| <
4M
where M = sup
k1
|x
k
| < . Since f
m
(y
k
) is a Cauchy sequence, there is a positive
integer K such that
[f
m
(y
k
) f
m
(y
l
)[ <
2
whenever k, l > K. Thus, from the previous three inequalities,
[f(y
k
) f(y
l
)[
4M
2M +
2
<
whenever k, l > K and we see that f(y
k
) is a Cauchy sequence in R. Set
(f) = lim
k
f(y
k
)
for each f X
and, since
(f) |f| M
for each f X
we have X
. Thus y
k
converges to x in the weak topology.
8.4. DUAL SPACES 293
Now suppose that X is a reexive Banach space, not necessarily separable, and
suppose x
k
B. It suces to show that there exists x B and a subsequence
such that x
k
x weakly. Let Y denote the closure in the strong topology of the
subspace of X spanned by x
k
. Then Y is obviously separable and, by Theorem
291.1, Y is a reexive Banach space. Thus there is a subsequence x
k
j
of x
k
. For any f X
let f
Y
denote the restriction of f to Y . As in the
proof of Theorem 291.1 we see that f
Y
Y
. Thus
f(x) = f
Y
(x) = lim
j
f
Y
(x
k
j
) = lim
j
f(x
k
j
).
Thus x
k
j
converges weakly to x in X. Furthermore, since |x
k
| 1, the same is
true for x by Theorem 290.1 and so x B.
293.1. Definition. If X is a normed linear space we may also consider the
weak topology on X
. The weak
topology on X
is
continuous. Here is the natural imbedding of X into X
topology on X
: [f(x
i
) f
0
(x
i
)[ <
i
for 1 i m
where m is any positive integer, f
0
X
, and x
i
X,
i
> 0 for 1 i m. Thus
a sequence f
k
in X
topology to an element f X
if
and only if
lim
k
f
k
(x) = f(x)
for each x X. Of course, if X is a reexive Banach space, the weak and weak
topologies on X
coincide.
293.2. Theorem. If X is a separable normed linear space then any bounded
subsequence in X
is weak
sequentially compact.
Proof. Let f
k
be a bounded sequence in X
, and let x
m
be a dense subset
of X. Since f
k
(x
1
) is bounded in R there is a subsequence f
1
k
of f
k
such
that f
1
k
(x
1
) converges in R. Then the sequence f
1
k
(x
2
) is bounded and thus
there is a subsequence f
2
k
of f
1
k
such that f
2
k
(x
2
) converges in R. In this way
294 8. ELEMENTS OF FUNCTIONAL ANALYSIS
we obtain by induction a family of subsequences f
m
k
of f
k
such that f
m
k
is
a subsequence of f
m1
k
for each m > 1 and f
m
k
(x
m
) converges in R for each
m 1. Set g
k
= f
k
k
. Then g
k
is a subsequence of f
k
and g
k
(x
m
) converges
as k for each m 1. Now let x be any element of X. Then
[g
k
(x) g
l
(x)[ [g
k
(x) g
k
(x
m
)[ +[g
k
(x
m
) g
l
(x
m
)[ +[g
l
(x
m
) g
l
(x)[
for any k, l, m. Given > 0 there is an m such that |x x
m
| <
4M
where
M = sup
k1
|f
k
| < .
For xed m the sequence g
k
(x
m
) is convergent and hence is a fundamental se-
quence. Thus there is a positive integer K such that
[g
k
(x
m
) g
l
(x
m
)[ <
2
whenever k, l > K and hence
[g
k
(x) g
l
(x)[ (|g
k
| +|g
l
|) |x x
m
| +
2
< .
whenever k, l > K. Thus the sequence g
k
(x) is a fundamental sequence in R for
any x X.
Set
g(x) = lim
k
g
k
(x)
for each x X. The function g is easily seen to be linear, and since [g
k
(x)[ M |x|
for each k we see that |g| M, i.e., g X
.
We remark that the base for the weak
topology on X
is similar in form
to the base for the weak topology on X except that the roles of X and X
are
interchanged.
The importance of the weak
: |f| 1 of X
topology.
Proof. If f B, then f(x) [|x| , |x|] for each x X. Set I
x
=
[|x| , |x|] for x X. Then, according to Tychonos Theorem 57.1, the product
P =
xX
I
x
,
with the product topology, is compact. Recall that P is by denition all functions
f dened on X with the pr0perty that f(x) I
x
for each x X. Thus the set B
can be viewed as a subset B
by
the product topology is easily seen to coincide with the relative topology induced
8.5. WEAK AND STRONG CONVERGENCE IN L
p
295
on B by the weak
is
a closed subset of P in the product topology.
Let f be an element in the closure of B
such that
[f(x) g(x)[ < , [f(y) g(y)[ < , [f(z) g(z)[ < .
Thus since g is linear
[f(z) f(x) f(y)[
[f(z) g(z)[ +[[ [f(x) g(x)[ +[[ [f(y) g(y)[
(1 +[[ +[[),
from which it follows that
f(x +y) = f(x) +f(y),
i.e., f is linear. In view of (295.1) f B
. Thus B
topology.
The proof of the following corollary of Theorem 294.1 is left as an exercise.
Also, see Exercise 8.12.
295.1. Corollary. The unit ball in a reexive Banach space is both compact
and sequentially compact.
8.5. Weak and Strong Convergence in L
p
Although it is easily seen that strong convergence implies weak con-
vergence in L
p
, it is shown below that under certain conditions, weak
convergence implies strong convergence.
296 8. ELEMENTS OF FUNCTIONAL ANALYSIS
We now apply some of the results of this chapter in the setting of L
p
spaces.
To begin we note that if 1 p < , then, in view of Example 289.2 (ii), a sequence
f
k
k=1
in L
p
(X, /, ) converges weakly to f L
p
(X, /, ) if and only if
lim
k
_
X
f
k
g d =
_
X
fg d
for each g L
p
(X, /, ).
296.1. Theorem. Let (X, /, ) be a measure space and suppose f and f
k
k=1
are functions in L
p
(X, /, ). If 1 p < , and |f
k
f|
p
0, then f
k
f
weakly in L
p
.
Proof. This is an immediate consequence of Holders inequality.
If f
k
k=1
is a sequence of functions with |f
k
|
p
M for some M and all k,
then since L
p
(X) is reexive for 1 < p < , Theorem 291.2 asserts that there is
a subsequence that converges weakly to some f L
p
(X). The next result shows
that if it is also known that f
k
f -a.e., then the full sequence converges weakly
to f.
296.2. Theorem. Let 1 < p < . If f
k
f -a.e., then f
k
f weakly in L
p
if and only if |f
k
|
p
is a bounded sequence.
Proof. Necessity follows immediately from Theorem 290.1.
To prove suciency, let M 0 be such that |f
k
|
p
M for all positive integers
k. Then Fatous Lemma implies
(296.1)
|f|
p
p
=
_
X
[f[
p
d =
_
X
lim
k
[f
k
[
p
d
liminf
i
_
X
[f
k
[
p
d M
p
.
Let > 0 and g L
p
d
_
1/p
<
6M
whenever E / and (E) < . We claim there exists a set F / such that
(F) < and
(296.3)
__
F
[g[
p
d
_
1/p
<
6M
.
To verify the claim, set A
t
: = x : [g(x)[
p
A
t
[g[
p
d =
_
X
[g[
p
d
8.5. WEAK AND STRONG CONVERGENCE IN L
p
297
and thus (296.3) holds with F = A
t
for suciently small positive t.
We can now apply Egoros Theorem (Theorem 143.3) on F to obtain A /
such that A F, (F A) < , and f
k
f uniformly on A. Let k
0
be such that
k k
0
implies
(297.1)
__
A
[f f
k
[
p
d
_
1/p
|g|
p
<
3
.
Setting E = F A in (296.2), we obtain from (296.1), (296.3), (297.1), and Holders
inequality,
_
X
fg d
_
X
f
k
g d
_
X
[f f
k
[ [g[ d
=
_
A
[f f
k
[ [g[ d +
_
FA
[f f
k
[ [g[ d
+
_
F
[f f
k
[ [g[ d
|f f
k
|
p,A
|g|
p
+|f f
k
|
p
(|g|
p
,FA
+|g|
p
F
)
3
+ 2M(
6M
+
6M
)
=
for all k k
0
.
If the hypotheses of the last result are changed to include that |f
k
|
p
|f|
p
,
then we can prove that |f
k
f|
p
0. This is an immediate consequence of the
following theorem.
297.1. Theorem. Let 1 p < . Suppose f and f
k
k=1
are functions in
L
p
(X, /, ) such that f
k
f -a.e. and the sequence |f
k
|
p
k=1
is bounded.
Then
lim
k
(|f
k
|
p
p
|f
k
f|
p
p
) = |f|
p
p
.
Proof. Set
M: = sup
k1
|f
k
|
p
<
and note that by Fatous Lemma, |f|
p
M.
Fix > 0 and observe that the function
h
(t): = [[t + 1[
p
[t[
p
[ [t[
p
298 8. ELEMENTS OF FUNCTIONAL ANALYSIS
is continuous on R and
lim
|t|
h
(t) = .
Thus there is a constant C
(t) < C
[b[
p
for any real numbers a and b.
Set
G
k
: =
_
[f
k
[
p
[f
k
f[
p
[f[
p
[f
k
f[
p
+
and note that G
k
0 -a.e. as k .
Setting a = f
k
f and b = f in (297.2) we see that
[f
k
[
p
[f
k
f[
p
[f[
p
[f
k
[
p
[f
k
f[
p
+[f[
p
[f
k
f[
p
+ (C
+ 1) [f[
p
from which it follows that
G
k
(C
+ 1) [f[
p
and, by the Dominated Convergence Theorem,
lim
k
_
X
G
k
d = 0.
Since
[[f
k
[
p
[f
k
f[
p
[f[
p
[ G
k
+ [f
k
f[
p
,
we see that
limsup
k
_
X
[[f
k
[
p
[f
k
f[
p
[f[
p
[ d (2M)
p
.
Since is arbitrary the proof is complete.
The following provides a summary of results.
298.1. Theorem. The following statements hold in a general measure space
(X, /, ).
(i) If f
i
f -a.e., then
|f|
p
liminf
i
|f
i
|
p
, 1 p < .
(ii) If f
i
f -a.e. and |f
i
|
p
is a bounded sequence, then f
i
f weakly in
L
p
, 1 < p < .
(iii) If f
i
f -a.e. and there exists a function g L
p
such that for each i,
[f
i
[ g -a.e. on X, then |f
i
f|
p
0, 1 p < .
(iv) If f
i
f -a.e. and |f
i
|
p
|f|
p
, then |f
i
f|
p
0, 1 p < .
8.5. WEAK AND STRONG CONVERGENCE IN L
p
299
Proof. Fatous Lemma implies (i), (ii) follows from Theorem 296.1, (iii) fol-
lows from Lebesgues Dominated Convergence Theorem, and (iv) is a restatement
of the previous theorem.
We conclude this chapter with another proof of the Radon-Nikodym Theorem,
which is based on the Riesz Representation Theorem.
299.1. Theorem. Let and be -nite measures on (X, /) with .
Then there exists a function h L
1
loc
(X, /, ) such that
v(E) =
_
X
h d
for all E /.
Proof. First, we will assume that both measures are nite and prove that
there is a measurable function g on X such that 0 g < 1 and
_
X
f(1 g) d =
_
X
fg d
for all f L
2
(X, /, +). For this purpose, dene
(299.1) T(f) =
_
X
f d
for f L
2
(X, /, +). By Holders inequality, it follows that
f L
1
(X, /, +)
and therefore f L
1
(X, /, ) also. Hence, we see that T is nite for f
L
2
(X, /, + ) and thus is a well-dened linear functional. Furthermore, it is
a bounded linear functional because
[T(f)[
__
X
[f[
2
d
_
1/2
((X))
1/2
= |f|
2;
((X))
1/2
|f|
2;+
((X))
1/2
.
Referring to Theorem 304.1 (or Theorem 183.2), there is a function L
2
( +)
such that
(299.2) T(f) =
_
X
f d( +)
for all f L
2
(+). Observe that 0 (+)-a.e, for otherwise we would obtain
T(
A
) < 0
300 8. ELEMENTS OF FUNCTIONAL ANALYSIS
where A: = < 0, which is impossible. Now, (299.1) and (299.2) imply
(299.3)
_
X
f(1 ) d =
_
X
f d
for f L
2
( +). If f is taken as
E
where E: = 1, we obtain
0 (E) =
_
X
E
d
_
X
E
d =
_
X
E
(1 ) d 0.
Hence, we have (E) = 0 and consequently, (E) = 0. Setting g =
E
, we have
0 g < 1 and g = almost everywhere with respect to both and . Reference
to (299.3) yields
_
X
f(1 g) d =
_
X
fg d.
Since g is bounded, we can replace f by (1 +g +g
2
+ +g
k
)
E
in this equation
for any positive integer k and E measurable. Then we obtain
_
E
(1 g
k+1
) d =
_
E
g(1 +g +g
2
+ +g
k
)d.
Since 0 g < 1 almost everywhere with respect to both and , the left side
tends to (E) while the integrands on the right side increase monotonically to
some function measurable h. Thus, by the Monotone Convergence Theorem, we
obtain
(E) =
_
E
h d
for every measurable set E. This gives us the desired result in case both and
are nite measures.
The proof of the general case proceeds as in Theorem 180.2.
8.6. Hilbert Spaces
We consider in this section Hilbert spaces, i.e., Banach spaces in which
the norm is induced by an inner product. This additional structure
allows us to study the representation of elements of the space in terms
of orthonormal systems.
300.1. Definition. An inner product on a linear space X is a real-valued
function (x, y) x, y) on X X such that for x, y, z X and , R
x, y) = y, x)
x +y, z) = x, z) +y, z)
x, x) 0
x, x) = 0 if, and only if, x = 0
8.6. HILBERT SPACES 301
300.2. Theorem. Suppose that X is a linear space on which an inner product
, ) is dened . Then
|x| =
_
x, x)
denes a norm on X.
Proof. It follows immediately from the denition of an inner product that
|x| 0 for all x X
|x| = 0 if, and only if x = 0
|x| = [[ |x| for all R, x X.
Only the triangle inequality remains to be proved. To do this, we rst prove the
Schwarz Inequality
[x, y)[ |x| |y| .
Suppose x, y X and R. From the properties of an inner product,
0 |x y|
2
= x y, x y)
= |x|
2
2x, y) +
2
|y|
2
,
and thus
2x, y)
1
|x|
2
+|y|
2
for any > 0. Assuming y ,= 0 and setting =
x
y
, we see that
(301.1) x, y) |x| |y| .
Note that (301.1) also holds in case y = 0. Since
x, y) = x, y) |x| |y| ,
we see that
(301.2) [x, y)[ |x| |y|
for all x, y X.
For the triangle inequality, observe that
|x +y|
2
= |x|
2
+ 2x, y) +|y|
2
|x|
2
+ 2 |x| |y| +|y|
2
= (|x| +|y|)
2
.
302 8. ELEMENTS OF FUNCTIONAL ANALYSIS
Thus
|x +y| |x| +|y|
for any x, y X, from which we see that || is a norm on X.
Thus we see that a linear space equipped with an inner product is a normed
linear space. The inequality (301.2) is called the Schwarz Inequality.
302.1. Definition. A Hilbert space is a linear space with an inner product
that is a Banach space with respect to the norm induced by the inner product (as
in Theorem 300.2).
302.2. Definition. We will say that two elements x, y in a Hilbert space H
are orthogonal if x, y) = 0. If M is a subspace of H, we set
M = x H : x, y) = 0 for all y M.
It is easily seen that M is a subspace of H.
We next investigate the geometry of a Hilbert space.
302.3. Theorem. Suppose M is a closed subspace of a Hilbert space H. Then
for each x
0
H there exists a unique y
0
M such that
|x
0
y
0
| = inf
yM
|x
0
y| .
Moreover y
0
is the unique element of M such that x
0
y
0
M
.
Proof. If x
0
M the assertion is obvious, so assume x
0
H M. Since M
is closed and x
0
, M,
d = inf
yM
|x
0
y| > 0.
There is a sequence y
k
in M such that
lim
k
|x
0
y
k
| = d.
For any k, l
|(x
0
y
k
) (x
0
y
l
)|
2
+|(x
0
y
k
) + (x
0
y
l
)|
2
= 2 |x
0
y
k
|
2
+ 2 |x
0
y
l
|
2
|y
k
y
l
|
2
= 2(|x
0
y
k
|
2
+|x
0
y
l
|
2
) 4
_
_
_
_
x
0
1
2
(y
k
+y
l
)
_
_
_
_
2
.
2(|x
0
y
k
|
2
+|x
0
y
l
|
2
) 4d
2
.
8.6. HILBERT SPACES 303
Since the right hand side of the last inequality above tends to 0 as k, l we see
that y
k
is a Cauchy sequence in H and consequently converges to an element y
0
.
Since M is closed , y
0
M.
Now let y M, R and compute
d
2
|x
0
(y
0
+y)|
2
= |x
0
y
0
|
2
2x
0
y
0
, y) +
2
|y|
2
= d
2
2x
0
y
0
, y) +
2
|y|
2
whence
x
0
y
0
, y)
2
|y|
2
for any > 0. Since is otherwise arbitrary we conclude that
x
0
y
0
, y) 0
for each y M. But then
x
0
y
0
, y) = x
0
y
0
, y) 0
for each y M and thus x
0
y
0
, y) = 0 for each y M, i.e., x
0
y
0
M
.
If y
1
M is such that
|x
0
y
1
| = inf
yM
|x
0
y| ,
then the above argument shows that x
0
y
1
M
. Hence, y
1
y
0
= (x
0
y
0
)
(x
0
y
1
) M
such that
x = y +z.
Proof. We may assume that M ,= H. Let x H. According to Theorem
302.3 there is a y M such that z = x y M
. Thus
|y
1
y
2
|
2
= y
1
y
2
, y
1
y
2
) = 0
whence y
1
= y
2
, This in turn implies that z
1
= z
2
.
If y H is xed and we dene the function
f(x) = y, x)
for x H, then f is a linear functional on H. Furthermore from Schwarzs inequal-
ity
[f(x)[ = [y, x)[ |y| |x| ,
which implies that |f| |y|. If y = 0, then |f| = 0. If y ,= 0, then
f(
y
|y|
) = |y| .
Thus |f| = |y|. Using Theorem 302.3, we will show that every continuous linear
functional on H is of this form.
304.1. Theorem (Riesz Representation Theorem). Suppose H is a Hilbert
space. Then for each f H
are isometrically
isomorphic.
Proof. Suppose f H
with x
0
,= 0. Since x
0
, M, f(x
0
) ,= 0. Since for each x H
f
_
x
f(x)
f(x
0
)
x
0
_
= 0,
we see that
x
f(x)
f(x
0
)
x
0
M
for each x H. Thus
x
f(x)
f(x
0
)
x
0
, x
0
) = 0
for each x H. This last equation may be rewritten as
f(x) =
f(x
0
)
|x
0
|
2
x
0
, x).
8.6. HILBERT SPACES 305
Thus if we set y =
f(x
0
)
|x
0
|
2
x
0
we see that (304.1) holds. We have already observed
that the norm of a linear functional f satisfying (304.1) is |y|.
If y
1
, y
2
H are such that y
1
, x) = y
2
, x) for all x H, then
y
1
y
2
, x) = 0
for all x X. Thus, in particular,
|y
1
y
2
|
2
= y
1
y
2
, y
1
y
2
) = 0
whence y
1
= y
2
. This shows that the y that represents f in (304.1) is unique.
We may rephrase the above results as follows. Let : H H
be dened for
each x H by
(x)(y) = x, y)
for all y H. Then is a one-to-one linear mapping of H onto H
. Furthermore
|(x)| = |x| for each x H. Thus is an isometric isomorphism of H onto H
305.1. Theorem. Any Hilbert space H is a reexive Banach space and conse-
quently the set x H : |x| 1 is compact in the weak topology.
Proof. We rst show that H
follows
immediately from the properties of . Furthermore
f, f) =
1
(f),
1
(f)) = |f|
2
for each f H
such that
(f) = g, f)
for each f H
B(y,
1
2
) =
whenever x, y are distinct elements of T. If c is a countable dense subset of H,
then for each x T the set B(x,
1
2
) must contain an element of c.
We next study the properties of orthonormal families beginning with countable
orthonormal families.
306.3. Theorem. Suppose x
k
k=1
is an orthonormal sequence in a Hilbert
space H. Then the following assertions hold:
(i) For each x H
k=1
x, x
k
)
2
|x|
2
.
(ii) If
k
is a sequence of real numbers, then
_
_
_
_
_
x
m
k=1
x, x
k
)x
k
_
_
_
_
_
_
_
_
_
_
x
m
k=1
k
x
k
_
_
_
_
_
for each m 1.
8.6. HILBERT SPACES 307
(iii) If
k
is a sequence of real numbers then
k=1
k
x
k
converges in H if, and
only if,
k=1
2
k
converges in R, in which case the sum is independent of the
order in which the terms are arranged, i.e., the series converges uncondi-
tionally, and
_
_
_
_
_
k=1
k
x
k
_
_
_
_
_
2
=
k=1
2
k
.
Proof. (i) For any positive integer m
0
_
_
_
_
_
x
m
k=1
x, x
k
)x
k
_
_
_
_
_
2
= |x|
2
2x,
m
k=1
x, x
k
)x
k
) +
_
_
_
_
_
m
k=1
x, x
k
)x
k
_
_
_
_
_
2
= |x|
2
2
m
k=1
x, x
k
)
2
+
m
k=1
x, x
k
)
2
= |x|
2
k=1
x, x
k
)
2
.
Thus
m
k=1
x, x
k
)
2
|x|
2
for any m 1 from which assertion (i) follows.
(ii) Fix a positive integer m and let M denote the subspace of H spanned by
x
1
, x
2
, . . . , x
m
. Then M is nite dimensional and hence closed; see Exercise 8 16.
Since for any 1 k m
x
k
, x
m
k=1
x, x
k
)x
k
) = 0
we see that
x
m
k=1
x, x
k
)x
k
M
.
In view of Theorem 302.3 this implies assertion (ii) since
m
k=1
k
x
k
M.
(iii) For any positive integers m and l with m > l
_
_
_
_
_
m
k=1
k
x
k
l
k=1
k
x
k
_
_
_
_
_
2
=
_
_
_
_
_
m
k=l+1
k
x
k
_
_
_
_
_
2
=
m
k=l+1
2
k
.
308 8. ELEMENTS OF FUNCTIONAL ANALYSIS
Thus the sequence
m
k=1
k
x
k
m=1
is a Cauchy sequence in H if and only if
k=1
2
k
converges in R.
Suppose
k=1
2
k
< . Since for any m
_
_
_
_
_
m
k=1
k
x
k
_
_
_
_
_
2
=
m
k=1
2
k
,
we see that
_
_
_
_
_
k=1
k
x
k
_
_
_
_
_
2
=
k=1
2
k
.
Let
k
j
be any rearrangement of the sequence
k
. Then for any m
(308.1)
_
_
_
_
_
_
m
k=1
k
x
k
m
j=1
k
j
x
k
j
_
_
_
_
_
_
2
=
m
k=1
2
k
2
k
j
m
2
k
j
+
m
j=1
2
k
j
.
Since the sum of the series
k=1
2
k
is independent of the order of the terms, the
last member of (308.1) converges to 0 as m .
308.1. Theorem. Suppose T is an orthonormal system in a Hilbert space H
and x H. Then
(i) The set y T : x, y) , = 0 is at most countable,
(ii) The series
yF
x, y)y converges unconditionally in H.
Proof. (i) Let > 0 and set
T
cannot exceed
|x|
2
2
and
thus T
k=1
y T : [x, y)[ >
1
k
,
we see that y T : x, y) ,= 0 is at most countable. (ii) Set y
k
k=1
= y T :
x, y) , = 0. Then, according to Theorem 306.3(i), (iii), the series
k=1
x, y
k
)y
k
converges unconditionally in H.
308.2. Theorem. Suppose T is a complete orthonormal system in a Hilbert
space H. Then for each x H
x =
yF
x, y)y
where all but countably many terms in the series are equal to 0 and the series
converges unconditionally in H.
8.6. HILBERT SPACES 309
Proof. Let Y denote the subspace of H spanned by T and let M denote the
closure of Y . If z M
= 0.
Fix x H. By Theorem 308.1 (ii), the series
yF
x, y)y converges uncondi-
tionally to an element x H.
Set y
k
k=1
= y T : x, y) , = 0. If y T and x, y) = 0, then
x
m
k=1
x, y
k
)y
k
, y) = 0
for each m 1. Then
[x x, y)[ =
x
m
k=1
x, y
k
)y
k
, y) x x, y)
x
m
k=1
x, y
k
)y
k
, y)
_
_
_
_
_
x
m
k=1
x, y
k
)y
k
_
_
_
_
_
|y|
and we see that x x, y) = 0.
If 1 l m, then
x
m
k=1
x, y
k
)y
k
, y
l
) = 0
and hence x x, y
l
) = 0 for every l 1.
We have thus shown that
(309.1) x x, y) = 0
for each y T from which it follows immediately that (309.1) holds for each y Y .
If w M, then there is a sequence w
k
in Y such that |w w
k
| 0 as
k . Thus
x x, w) = lim
k
x x, w
k
) = 0
which means that x x M
= 0.
309.1. Examples. (i) The Banach space R
n
is a Hilbert space with inner
product
(x
1
, x
2
, . . . , x
n
), (y
1
, y
2
, . . . , y
n
)) = x
1
y
1
+x
2
y
2
+ +x
n
y
n
.
310 8. ELEMENTS OF FUNCTIONAL ANALYSIS
(ii) The Banach space L
2
(X, ) is a Hilbert space with inner product
f, g) =
_
X
fg d.
(iii) The Banach space l
2
is a Hilbert space with inner product
a
k
, b
k
) =
k=1
a
k
b
k
.
Suppose H is a separable Hilbert space containing a countable orthonormal
system x
k
k=1
. For any x H the sequence x, x
k
) l
2
and
k=1
x, x
k
)
2
= |x|
2
.
On the other hand, if a
k
l
2
, then according to Theorem 306.3 the series
k=1
a
k
x
k
converges in H. Set x =
k=1
a
k
x
k
. Then
x, x
l
) = lim
m
k=1
a
k
x
k
, x
l
) = a
l
for each l and hence
k=1
a
2
k
= |x|
2
.
Thus the linear mapping T : H l
2
given by
T(x) = x, x
k
)
is an isometric isomorphism of H onto l
2
.
If is a open subset of R
n
and denotes Lebesgue measure on then L
2
(, )
is separable and hence isometrically isomorphic to l
2
.
Exercises for Chapter 8
Section 8.1
8.1 Use the Hausdor Maximal Principle to show that every linear space has a
basis. Hint: observe that a linearly independent subset of a linear space X
spans X if, and only if, it is maximal with respect to set inclusion i.e., it is not
contained in any other linearly independent subset of X.
8.2 For i = 1, 2, . . . , m, let X
i
be a Banach space with norm ||
i
. The Cartesian
product
X =
m
i=1
X
i
EXERCISES FOR CHAPTER 8 311
consisting of points x = (x
1
, x
2
, . . . , x
m
) with x
i
X
i
is a vector space under
the denitions
x +y = (x
1
+y
1
, . . . , x
m
+y
m
), cx = (cx
1
, . . . , cx
m
).
Prove that X is a Banach space with respect to any of the equivalent norms
|x| =
_
m
i=1
|x
i
|
p
i
_
1/p
, 1 p < .
Section 8.2
8.3 Suppose Y is a closed subspace of a normed linear space X, Y ,= X, and > 0.
Show that there is an element x X such that |x| = 1 and
inf
yY
|x y| > 1 .
8.4 Let f : X R
1
be a linear functional on the normed linear space X. Show
that f is bounded if and only if f is continuous at one point.
8.5 The kernel of a linear f : X R
1
is the set x : f(x) = 0. Prove that f is
bounded if and only if the kernel of f is closed in X.
Section 8.3
8.6 Suppose T : X Y is a continuous linear mapping of a normed linear space
X into a normed linear space Y . Show that the graph of T is closed in X Y .
8.7 Suppose T : X Y is a univalent, continuous linear mapping of a Banach
space X onto a Banach space Y . Show that the inverse mapping T
1
: Y X
is also continuous.
8.8 Suppose T : X Y is a univalent, continuous linear mapping of a Banach
space X into a Banach space Y . Prove that T(X) is closed in Y if and only if
|x| C |T(x)|
for each x X.
8.9 (a) Let T
k
be a sequence of bounded linear operators T
k
: X Y where X is
a Banach space aand Y is a normed linear space. Suppose
lim
k
T
k
(x)
exists for each x X. Prove that there exists a bounded linear operator
T : X Y such that
lim
k
T
k
(x) = T(x)
for each x X.
(b) Give an example that shows part (a) is false is X is not assumed to be a
Banach space.
312 8. ELEMENTS OF FUNCTIONAL ANALYSIS
8.10 Let M be an arbitrary closed subspace of a normed linear space X. Let us say
that x, y X are equivalent, written as x y, if x y M. We will denote
by [x] all elements y X such that x y.
(a) With [x] + [y] := [x + y] and [cx] := c[x] where c R, prove that these
operations are well-dened and that these cosets form a vector space.
(b) Let us dene
|[x]| := inf
yM
|x y| .
Prove that |[]| is a norm on the space, /, of all cosets [x].
(c) The space / is called the quotient space and is denoted by X/M. Prove
that if M is a closed subspaece of a Banach space X, then X/M is also a
Banach space.
8.11 Let X denote the set of all sequences a
k
k=1
such that all but nitely many
of the a
k
= 0.
(a) Show that X is a linear space under the usual denitions of addition and
scalar multiplication.
(b) Show that |a
k
| = max
k1
[a
k
[ is a norm on X.
(c) Dene a mapping T : X X by
T(a
k
) = ka
k
.
Show that T is a linear mapping.
(d) Show that the graph of T is closed in X X.
(e) Show that T is not continuous.
Section 8.4
8.12 Show that the unit ball in a reexive Banach space is both compact and se-
quentially compact.
8.13 Suppose X is a normed linear space and f
k
is a sequence in X
that converges
in the weak
topology to f X
. Show that
sup
k1
|f
k
| < ,
and
|f| liminf
k
|f
k
| .
8.14 Show that a subspace of a normed linear space is closed in the strong topology
if and only if it is closed in the weak topology.
8.15 Prove that any subspace of a normed linear space cannot be open.
8.16 Show that a nite dimensional subspace of a normed linear space is closed in
the strong topology.
EXERCISES FOR CHAPTER 8 313
8.17 Show that if X is an innite dimensional normed linear space, then there is
a bounded sequence x
k
in X of which no subsequence is convergent in the
strong topology. (Hint: Use Exercise 8.16 and 8.3.) Thus conclude that the
unit ball in any innite dimensional normed linear space is not compact.
8.18 Show that any Banach space X is isometrically isomorphic to a closed linear
subspace of C() (cf. Example 289.2(iv)) where is a compact Hausdor space.
(Hint: Set
= f X
: |f| 1
with the weak
.)
8.19 As usual, let C[0, 1] denote the space of continuous functions on [0, 1] endowed
with the sup norm. Prove that if f
k
is a sequence of functions in C[0, 1] that
converge weakly to f, then the sequence is bounded and f
k
(t) f(t) for each
t [0, 1].
8.20 Suppose is an open subset of R
n
, and let denote Lebesgue measure on .
Set
T = x = (x
1
, x
2
, . . . , x
n
) R
n
: x
j
is rational for each 1 j n
and let Q denote the set of all open cubes in R
n
with edges parallel to the
coordinate axes and vertices in T. (i) Show that if E is a Lebesgue measurable
subset of R
n
with (E) < and > 0, then there exists a disjoint nite
sequence Q
k
m
k=1
with each Q
k
Q such that
_
_
_
_
_
k=1
Q
k
_
_
_
_
_
L
p
(,)
< .
(ii) Show that the set of all nite linear combinations of elements of
Q
: Q
Q with rational coecients is dense in L
p
(, ).
8.21 Suppose is an open subset of R
n
and let denote Lebesgue measure on .
Show that L
B(x,r
1
)
B(x,r
2
)
_
_
_
L
(,)
= 1.)
8.22 Referring to Exercise 8 .2, prove that there is a natural isomorphism between
X
and
m
i=1
X
i
. Thus conclude that X is reexive if each X
i
is reexive.
Section 8.5
8.23 Suppose f
i
f weakly in L
p
(X, /, ), 1 < p < and that f
i
f pointwise
-a.e. Prove that f
+
i
f
+
and f
i
f
weakly in L
p
.
8.24 Show that the previous Exercise 8 is false if the hypothesis of pointwise con-
vergence is dropped.
314 8. ELEMENTS OF FUNCTIONAL ANALYSIS
8.25 Suppose that (X, /, ) is a nite measure space and that f is a given measure
function. Assume that:
_
X
fgd <
for each g L
p
(X), 1 p . Prove that |f|
p
< .
8.26 Let C := C[0, 1] denote the space of continuous functions on [0, 1] endowed
with the usual sup norm and let X be a linear subspace of C that is closed
relative to the L
2
norm.
(i) Prove that X is also closed in C.
(ii) Show that |f|
2
|f|
for each f X
(iii) Prove that there exists M > 0 such that |f|
M |f|
2
for all f X.
(iv) For each t [0, 1], show that there is a function g
t
L
2
such that
f(t) =
_
g
t
(x)f(x) dx
for all f X.
(v) Show that if f
k
f weakly in L
2
where f
k
X, then f
k
(x) f(x) for
each x X.
(vi) As a consequence of the above, show that if f
k
f weakly in L
2
where
f
k
X, then f
k
f strongly in L
2
; that is |f
k
f|
2
0.
8.27 A subset K X in a linear space is called convex if for any two points
x, y K, the line segment tx + (1 t)y, t [0, 1], also belong to K.
(i) Prove that the unit ball in any normed linear space is convex.
(ii) If K is a convex set, a point x
0
K is called an extreme point of K
if x
0
is not in the interior of any line segment that lies in K; that is, if
x
0
= ty + (1 t)z where 0 < t < 1, then either y or z is not in K. Show
that any f L
p
[0, 1], 1 < p < , with |f|
p
= 1 is an extreme point of
the unit ball.
(iii) Show that the extreme points of the unit ball in L
k=1
where
x
m
k
=
_
_
_
1 ifk = m
0 otherwise.
Show that the sequence x
m
m=1
in l
p
(cf. Example 289.2(iii)) converges to 0
in the weak topology but does not converge in the strong topology.
CHAPTER 9
Measures and Linear Functionals
9.1. The Daniell Integral
Theorem 183.2 states that a function in L
p
can be regarded as a
bounded linear functional on L
p
. Here we show that a large class of
measures can be represented as bounded linear functionals on the space
of continuous functions. This is a very important result that has many
useful applications and provides a fundamental connection between mea-
sure theory and functional analysis.
Suppose (X, /, ) is a measure space. Integration denes an operation that is
linear, order preserving, and continuous relative to increasing convergence. Specif-
ically, we have
(i)
_
(kf) d = k
_
f d whenever k R and f is an integrable function
(ii)
_
(f +g) d =
_
f d +
_
g d whenever f, g are integrable
(iii)
_
f d
_
g d whenever f, g are integrable functions with f g -a.e.
(iv) If f
i
is an nondecreasing sequence of integrable functions, then
lim
i
_
f
i
d =
_
lim
i
f
i
d.
The main objective of this section is to show that if a linear functional, dened
on an appropriate space of functions, possesses the four properties above, then it can
be expressed as the operation of integration with respect to some measure. Thus,
we will have shown that these properties completely characterize the operation of
integration.
It turns out that the proof of our main result is no more dicult when cast in
a very general framework, so we proceed by introducing the concept of a lattice.
317.1. Definition. If f, g are real-valued functions dened on a space X, we
dene
(f g)(x) := min[f(x), g(x)]
(f g)(x) := max[f(x), g(x)].
A collection L of real-valued functions dened on an abstract space X is called a
lattice provided the following conditions are satised: if 0 c < and f and g
317
318 9. MEASURES AND LINEAR FUNCTIONALS
are elements of L, then so are the functions f +g, cf, f g, and f c. Furthermore,
if f g, then g f is required to belong to L. Note that if f and g belong to L
with g 0, then so does f g = f + g f g. Therefore, f
+
belongs to L. We
dene f
= f
+
f and so f
L. We let L
+
denote those functions f in L for
which f 0. Clearly, if L is a lattice, then so is L
+
.
For example, the space of continuous functions on a metric space is a lattice as
well as the space of integrable functions, but the collection of lower semicontinuous
functions is not a lattice because it is not closed under the -operation (see Exercise
9.1).
318.1. Theorem. Suppose L is a lattice of functions on X and let T : L R
be a functional satisfying the following conditions for all functions in L:
(i) T(f +g) = T(f) +T(g)
(ii) T(cf) = cT(f) whenever 0 c <
(iii) T(f) T(g) whenever f g
(iv) T(f) = lim
i
T(f
i
) whenever f : = lim
i
f
i
is a member of L
(v) and f
i
is nondecreasing.
Then, there exists an outer measure on X such that for each f L, f is -
measurable and
T(f) =
_
X
f d.
In particular, f > t is -measurable whenever f L and t R.
Proof. First, observe that since T(0) = T(0 f) = 0 T(f) = 0, (iii) above
implies T(f) 0 whenever f L
+
. Next, with the convention that the inmum of
the empty set is , for an arbitrary set A X, we dene
(A) = inf
_
lim
i
T(f
i
)
_
where the inmum is taken over all sequences of functions f
i
with the property
(318.1) f
i
L
+
, f
i
i=1
is nondecreasing, and lim
i
f
i
A
.
Such sequences of functions are called admissible for A. In accordance with our
convention, if there is no admissible sequence of functions for A, we dene (A) =
. It follows from denition that if f L
+
and f
A
, then
(318.2) T(f) (A).
On the other hand, if f L
+
and f
A
, then
(318.3) T(f) (A).
9.1. THE DANIELL INTEGRAL 319
This is true because if f
i
is admissible for A, then g
i
: = f
i
f is a nondecreasing
sequence with lim
i
g
i
= f. Hence,
T(f) = lim
i
T(g
i
) lim
i
T(f
i
)
and therefore
T(f) (A).
The rst step is to show that is an outer measure on X. For this, the only
nontrivial property to be established is countable subadditivity. For this purpose,
let
A
i=1
A
i
.
For each xed i and arbitrary > 0, let f
i,j
j=1
be an admissible sequence of
functions for A
i
with the property that
lim
j
T(f
i,j
) < (A
i
) +
2
i
.
Now dene
g
k
=
k
i=1
f
i,k
and obtain
T(g
k
) =
k
i=1
T(f
i,k
)
i=1
lim
m
T(f
i,m
)
i=1
_
(A
i
) +
2
i
_
.
After we show that g
k
is admissible for A, we can take limits of both sides as
k and conclude
(A)
i=1
(A
i
) +.
Since is arbitrary, this would prove the countable subadditivity of and thus
establish that is an outer measure on X.
To see that g
k
is admissible for A, select x A. Then x A
i
for some i and
with k: = max(i, j), we have g
k
(x) f
i,j
(x). Hence,
lim
k
g
k
(x) lim
j
f
i,j
(x) 1.
The next step is to prove that each element f L is a -measurable function.
Since f = f
+
f
being similar. For this we appeal to Theorem 139.2 which asserts that it is
sucient to prove
(319.1) (A) (A f
+
a) +(A f
+
b)
whenever A X and a < b are real numbers. Since f
+
0, we may as well take
a 0. Let g
i
be an admissible sequence of functions for A and dene
h =
[f
+
b f
+
a]
b a
, k
i
= g
i
h.
Observe that h = 1 on f
+
b. Since g
i
is admissible for A it follows that k
i
is admissible for A f
+
b. Furthermore, since h = 0 on f
+
a, we have
that g
i
k
i
is admissible for A f
+
a. Consequently,
lim
i
T(g
i
) = lim
i
[T(k
i
) +T(g
i
k
i
)] (A f
+
b) +(A f
+
a).
Since g
i
is an arbitrary admissible sequence for A, we obtain
(A) (A f
+
b) +(A f
+
a),
which proves (319.1).
The last step is to prove that
T(f) =
_
f d for f L.
We begin by considering f L
+
. With f
t
: = f t, t 0, note that if > 0 and
k is a positive integer, then
0 f
k
(x) f
(k1)
(x) for x X,
and
f
k
(x) f
(k1)
(x) =
_
_
_
for f(x) k
0 for f(x) (k 1).
Therefore,
T(f
k
f
(k1)
) (f k) by (318.2)
_
(f
(k+1)
f
k
) d since f
(k+1)
f
k
{fk}
(f (k + 1)) since f
(k+1)
f
k
=
{fk}
on f (k + 1)
T(f
(k+2)
f
(k+1)
). by (318.3)
Now taking the sum as k ranges from 1 to n, we obtain
T(f
n
)
_
(f
(n+1)
f
) d T(f
(n+2)
f
2
).
9.1. THE DANIELL INTEGRAL 321
Since f
n
is a nondecreasing sequence with lim
n
f
n
= f, we have
T(f)
_
(f f
) d T(f f
2
).
Also, f
0 as 0
+
so that
T(f) =
_
f d.
Finally, if f L, then f
+
L
+
, f
L
+
, thus yielding
T(f) = T(f
+
) T(f
) =
_
f
+
d
_
f
d =
_
f d.
Now that we have established the existence of an outer measure corresponding
to the functional T, we address the question of its uniqueness. For this purpose, we
need the following lemma, which asserts that possesses a type of outer regularity.
321.1. Lemma. Under the assumptions of the preceding Theorem, let / denote
the -algebra generated by all sets of the form f > t where f L and t R.
Then, for any A X there is a set W / such that
A W and (A) = (W).
Proof. If (A) = , take W = X. If (A) < , we proceed as follows.
For each positive integer i let f
i,j
j=1
be an admissible sequence for A with the
property
lim
j
T(f
i,j
) < (A) +
1
i
,
and dene, for each positive integer j,
g
i,j
: = inff
1,j
, f
2,j
, . . . , f
i,j
,
B
i,j
: = x : g
i,j
(x) > 1
1
i
.
Then,
g
i+1,j
g
i,j
g
i,j+1
and B
i+1,j
B
i,j
B
i,j+1
.
Furthermore, with the help of (318.2),
(1 1/i)(B
i,j
) T(g
i,j
) T(f
i,j
).
Now let
V
i
=
j=1
B
i,j
.
Observe that V
i
V
i+1
A and V
i
/for all i. Indeed, to verify that V
i
A, it is
sucient to show g
i,j
(x
0
) > 11/i for x
0
A whenever j is suciently large. This
is accomplished by observing that there exist positive integers j
1
, j
2
, . . . , j
i
such that
f
1,j
(x
0
) > 1 1/i for j j
1
, f
2,j
(x
0
) > 1 1/i for j j
2
, . . . , f
i,j
(x
0
) > 1 1/i
322 9. MEASURES AND LINEAR FUNCTIONALS
for j j
i
. Thus, for j larger than maxj
1
, j
2
, . . . , j
i
, we have g
i,j
(x
0
) > 1 1/i.
For each positive integer i,
(321.1)
(A) (V
i
) = lim
j
(B
i,j
)
(1 1/i)
1
lim
j
T(f
i,j
)
(1 1/i)
1
[(A) + 1/i],
and therefore,
(322.1) (A) = lim
i
(V
i
).
Note that (321.1) implies (V
i
) < . Therefore, if we take
W =
i=1
V
i
,
we obtain A W, W /, and
(W) =
_
i=1
V
i
_
= lim
i
(V
i
)
= (A).
The preceding proof reveals the manner in which T uniquely determines .
Indeed, for f L
+
and for t, h > 0, dene
(322.2) f
h
=
f (t +h) f t
h
.
Let h
i
be a nonincreasing sequence of positive numbers such that h
i
0. Observe
that f
h
i
is admissible for the set f > t while f
h
i
{f>t}
for all small h
i
. From
Theorem 318.1 we know that
T(f
h
i
) =
_
f
h
i
d.
Thus, the Monotone Convergence Theorem implies that
lim
i
T(f
h
i
) = (f > t),
and this shows that the value of on the set f > t is uniquely determined by
T. From this it follows that (B
i,j
) is uniquely determined by T, and therefore by
(321.1), the same is true for (V
i
). Finally, referring to (322.1), where it is assumed
that (A) < , we have that (A) is uniquely determined by T. The requirement
that (A) < will be be ensured if
A
f for some f L, see (318.2). Thus, we
have the following corollary.
9.1. THE DANIELL INTEGRAL 323
322.1. Corollary. If T is as in Theorem 318.1, the corresponding outer mea-
sure is uniquely determined by T on all sets A with the property that
A
f for
some f L.
Functionals that satisfy the conditions of Theorem 318.1 are called monotone
(or alternatively, positive). In the spirit of the Jordan Decomposition Theorem, we
now investigate the question of determining those functionals that can be written
as the dierence of monotone functionals.
323.1. Theorem. Suppose L is a lattice of functions on X and let T : L R
be a functional satisfying the following four conditions for all functions in L:
T(f +g) = T(f) +T(g)
T(cf) = cT(f) whenever 0 c <
supT(k) : 0 k f < for all f L
+
T(f) = lim
i
T(f
i
) whenever f = lim
i
f
i
and f
i
is nondecreasing.
Then, there exist positive functionals T
+
and T
(f)
for all f L
+
.
Proof. We dene T
+
and T
on L
+
as follows:
T
+
(f) = supT(k) : 0 k f,
T
(f) = infT(k) : 0 k f
To prove (323.1), let f, g L
+
with f g. Then f f g and therefore
T
(f) T(f).
Similarly, since f g f, T(f g) T
+
(f) so that T(f) = T(g) + T(f g)
T(g) +T
+
(f). Taking the inmum over all 0 g f implies
T(f) T
(f) +T
+
(f).
Hence we obtain
T(f) = T
+
(f) T
(f).
324 9. MEASURES AND LINEAR FUNCTIONALS
Now we will prove that T
+
satises the conditions of Theorem 318.1 on L
+
.
For this, let f, g, h L
+
with f + g h and set k = inff, h. Then f k and
g h k; consequently,
T
+
(f) +T
+
(g) T(k) +T(h k) = T(h).
Since this holds for all h f +g with h L
+
, we have
T
+
(f) +T
+
(g) T
+
(f +g).
It is easy to verify the opposite inequality and also that T
+
is both positively
homogeneous and monotone.
To show that T
+
satises the last condition of Theorem 318.1, let f
i
be a
nondecreasing sequence in L
+
such that f
i
f. If k L
+
and k f, then the
sequence g
i
= inff
i
, k is nondecreasing and converges to k as i . Hence
T(k) = lim
i
T(g
i
) lim
i
T
+
(f
i
),
and therefore
T
+
(f) lim
i
T
+
(f
i
).
The opposite inequality is obvious and thus equality holds.
That T
(f) = T(f) T
+
(f) < for each f L
+
. Thus what we have
just proved for T
+
applies to R
+
as well. In particular, R
+
satises the conditions
of Theorem 318.1.
324.1. Corollary. Let T be as in the previous theorem. Then there exist
outer measures
+
and
measurable and
T(f) =
_
f d
+
_
f d
.
Proof. Theorem 318.1 supplies outer measures
+
,
such that
T
+
(f) =
_
f d
+
,
T
(f) =
_
f d
for all f L
+
. Since each f L can be written as f = f
+
f
K
f
U
.
Proof. Let r
1
, r
2
, . . . be an enumeration of the rationals in (0, 1) where we
take r
1
= 0 and r
2
= 1. By Theorem 44.3, there exist open sets V
0
and V
1
with
compact closures such that
K V
1
V
1
V
0
V
0
V.
Proceeding by induction, we will associate an open set V
r
k
with each rational r
k
in the following way. Assume that V
r
1
, . . . , V
r
k
have been chosen in such a manner
that V
r
j
V
r
i
whenever r
i
< r
j
. Among the numbers r
1
, r
2
, . . . , r
k
let r
m
denote
326 9. MEASURES AND LINEAR FUNCTIONALS
the smallest that is larger than r
k+1
and let r
n
denote the largest that is smaller
than r
k+1
. Referring again to Theorem 44.3, there exists V
r
k+1
such that
V
r
n
V
r
k+1
V
r
k+1
V
r
m
.
Continuing this process, for rational number r [0, 1] there is a corresponding
open set V
r
. The countable collection V
r
satises the following properties: K
V
1
, V
0
U, V
r
is compact, and
(326.1) V
s
V
r
for r < s.
For rational numbers r and s, dene
f
r
: = r
V
r
and g
s
: =
V
s
+s
XV
s
.
Referring to Denition 69.2, it is easy to see that each f
r
is lower semicontinuous
and each g
s
is upper semicontinuous. Conesequently, with
f(x): = supf
r
(x) : r Q (0, 1 ) and g : = infg
s
(x) : r Q (0, 1 )
it follows that f is lower semicontinuous and g is upper semicontinuous. Note that
0 f 1, f = 1 on K, and that spt f V
0
.
The proof will be completed when we show that f is continuous. This is
established by showing f = g. To this end, note that f g, for otherwise there
exists x X such that f
r
(x) > g
s
(x) for some rationals r and s. But this is possible
only if r > s, x V
r
, and x , V
s
. However, r > s implies V
r
V
s
.
Finally, we observe that if f(x) < g(x) for some x, then there would exist
rationals r and s such that
f(x) < r < s < g(x).
This would imply x , V
r
and x V
s
, contradicting (326.1). Hence, f = g.
326.1. Theorem. Suppose K is compact and V
1
, . . . , V
n
are open sets in a
locally compact Hausdor space X such that
K V
1
V
n
.
Then there are continuous functions g
i
with 0 g
i
1 and spt g
i
V
i
, i =
1, 2, . . . , n such that
g
1
(x) +g
2
(x) + +g
n
(x) = 1 for all x K.
9.2. THE RIESZ REPRESENTATION THEOREM 327
Proof. By Theorem 44.3, each x K is contained in an open set U
x
with
compact closure such that U
x
V
i
for some i. Since K is compact, there exist
nitely many points x
1
, x
2
, . . . , x
k
in K such that
K
k
i=1
U
x
i
.
For 1 j n, let G
j
denote the union of those U
x
i
that lie in V
j
. Urysohns
Lemma, Lemma 325.1, provides continuous functions f
j
with 0 f
j
1 such that
G
j
f
j
V
j
.
Now
n
j=1
f
j
1 on K, so applying Urysohns Lemma again, there exists f
C
c
(X) with f = 1 on K and spt f
n
j=1
f
j
> 0. Let f
n+1
= 1 f so that
n+1
j=1
f
j
> 0 everywhere. Now dene
g
i
=
f
i
n+1
j=1
f
j
to obtain the desired conclusion.
The theorem on the existence of the Daniell integral provides the following as
an immediate application.
327.1. Theorem (Riesz Representaton Theorem). Let X be a locally compact
Hausdor space and let L denote the lattice C
c
(X) of continuous functions with
compact support. If T : L R is a linear functional such that
(327.1) supT(g) : 0 g f <
whenever f L
+
, then there exist Baire outer measures
+
and
_
X
f d
.
The outer measures
+
and
i=1
K
i
=
and therefore the
K
i
form an open covering of X, and in particular, of spt f. Since
spt f is compact, there is an index i
0
such that
spt f
i
0
i=1
K
i
=
K
i
0
.
Since f
i
is a nondecreasing sequence and g 1 on spt f, this implies that
f(x) < f
i
(x) +g(x)
for all i i
0
and all x X. Note (327.1) implies that
M: = sup[T(k)[ : k L, 0 k g < .
Thus, we obtain
0 f f
i
g, [T(f f
i
)[ M < .
Since is arbitrary, we conclude T(f
i
) T(f) as required.
Concerning the assertion of uniqueness, select a compact set K and use Theo-
rem 44.3 to nd an open set U K with compact closure. Consider the lattice
L
U
: = C
c
(X) f : spt f U
and let T
U
(f): = T(f) for f L
U
. Then we obtain outer measures
+
U
and
U
such that
T
U
(f) =
_
f d
+
U
_
f d
U
for all f L
U
. With the help of Urysohns Lemma and Corollary 322.1, we nd that
+
U
and
U
are uniquely determined and so
+
U
(K) =
+
(K) and
U
(K) =
(K).
may be un-
dened. That is, both
+
(E) and
, then
|T|
_
_
T
+
_
_
+
_
_
T
_
_
= T
+
(1) +T
(1).
The opposite inequality is also valid, for if f C(X) is any function with 0 f 1,
then [2f 1[ 1 and
|T| T(2f 1) = 2T(f) T(1).
Taking the supremum over all such f yields
|T| 2T
+
(1) T(1)
= T
+
(1) +T
(1).
Hence,
(329.1) |T| = T
+
(1) +T
(1).
329.1. Corollary. Let X be a compact Hausdor space. Then for every
bounded linear functional T : C(X) R, there exists a unique, signed, Baire outer
measure on X such that
T(f) =
_
X
f d
for each f C(X). Moreover, |T| = [[ (X). Thus, the dual of C(X) is isometri-
cally isomorphic to the space of signed, Baire outer measures on X.
This will be used in the proof below.
Proof. A bounded linear functional T on C(X) is easily seen to imply prop-
erty (327.1), and therefore Theorem 327.1 implies there exist Baire outer measures
+
and
,
T(f) =
_
X
f d
for all f C(X). Thus,
[T(f)[
_
[f[ d [[
|f| [[ (X),
and therefore, |T| [[ (X). On the other hand, with
T
+
(f): =
_
X
f d
+
and T
(f): =
_
X
f d
,
330 9. MEASURES AND LINEAR FUNCTIONALS
we have with the help of (329.1)
[[ (X)
+
(X) +
(X)
= T
+
(1) +T
(1) = |T| .
Hence, |T| = [[ (X).
Baire sets arise naturally in our development because they comprise the smallest
-algebra that contains sets of the form f > t where f C
c
(X) and t R. These
are precisely the sets that occur in Theorem 318.1 when L is taken as C
c
(X). In view
of Exercise 9.7, note that the outer measure obtained in the preceding Corollary is
Borel. In general, it is possible to have access to Borel outer measures rather than
merely Baire measures, as seen in the following result.
330.1. Theorem. Assume the hypotheses and notation of the Riesz Represen-
tation Theorem. Then there is a signed Borel outer measure on X with the
property that
(330.1)
_
f d = T(f) =
_
f d
for all f C
c
(X).
Proof. Assuming rst that T satises the conditions of Theorem 318.1, we
will prove there is a Borel outer measure such that (330.1) holds for every f L
+
.
Then referring to the proof of the Riesz Representation Theorem, it follows that
there is a signed outer measure that satises (330.1).
We dene in the following way. For every open set U let
(330.2) (U): = supT(f) : f L
+
, f 1 and spt f U
and for every A X, dene
(A): = inf(U) : A U, U open.
Observe that and agree on all open sets.
First, we verify that is countably subadditive. Let U
i
be a sequence of
open sets and let
V =
i=1
U
i
.
If f L
+
, f 1, and spt f V , the compactness of spt f implies that
spt f
N
i=1
U
i
9.2. THE RIESZ REPRESENTATION THEOREM 331
for some positive integer N. Now appeal to Lemma 373.1 to obtain functions
g
1
, g
2
, . . . , g
N
L
+
with g
i
1, spt g
i
U
i
, and
N
i=1
g
i
(x) = 1 whenever x spt f.
Thus,
f(x) =
N
i=1
g
i
(x)f(x),
and therefore
T(f) =
N
i=1
T(g
i
f)
N
i=1
(U
i
) <
i=1
(U
i
).
This implies
i=1
U
i
_
i=1
(U
i
).
From this and the denition of , it follows easily that is countably subadditive
on all sets.
Next, to show that is a Borel outer measure, it suces to show that each
open set U is is measurable. For this, let > 0 and let A X be an arbitrary
set with (A) < . Choose an open set V A such that (V ) < (A) + and
a function f L
+
with f 1, spt f V U, and T(f) > (V U) . Also,
choose g L
+
with g 1 and spt g V spt f so that T(g) > (V spt f) .
Then, since f +g 1, we obtain
(A) + (V ) T(f +g) = T(f) +T(g)
(V U) +(V spt f)
(A U) + (AU) 2
This shows that U is -measurable since is arbitrary.
Finally, to establish (330.1), by Theorem 201.1 it suces to show that
(f > s) = (f > s)
for all f L
+
and all s R. It follows from denitions that
(U) = (U) (U)
whenever U is an open set. In particular, we have (U
s
) (U
s
) where U
s
: =
f > s. To prove the opposite inequality, note that if f L
+
and 0 < s < t, then
(U
s
)
T[f (t +h) f t]
h
for h > 0
332 9. MEASURES AND LINEAR FUNCTIONALS
because each
f
h
: =
f (t +h) f t
h
=
_
_
1 if f t +h
ft
h
if t < f < t +h
0 if f t
is a competitor in (330.2). Furthermore, since f
h
1
f
h
2
when h
1
h
2
and
lim
h0
+ f
h
=
{f>t}
, we may apply the Monotone Convergence Theorem to obtain
(f > t) = lim
h0
+
_
X
f
h
d
= lim
h0
+
T[f (t +h) f t]
h
(U
s
) = (U
s
).
Since
(U
s
) = lim
ts
+
(f > t),
we have (U
s
) (U
s
).
From Corollary 322.1 we see that and agree on compact sets and therefore,
is nite on compact sets. Furthermore, it follows from Lemma 321.1 that for
an arbitrary set A X, there exists a Borel set B A such that (B) = (A).
An outer measure with these properties is called a Radon outer measure. Note
that this denition is compatible with that of Radon measure given in Denition
203.2.
Exercises for Chapter 9
Section 9.1
9.1 Show that the space of lower semicontinuous functions on a metric space is not
a lattice because it is not closed under the -operation.
9.2 Let L denote the family of all functions of the form u p where u: R R is
continuous and p: R
2
R is the orthogonal projection dened by p(x
1
, x
2
) =
x
1
. Dene a functional T on L by
T(u p) =
_
1
0
u d.
Prove that T meets the conditions of Theorem 318.1 and that the measure
representing T satises
(A) = [p(A) [0, 1]]
for all A R
2
. Also, show that not all Borel sets are -measurable.
EXERCISES FOR CHAPTER 9 333
9.3 Prove that the sequence f
h
i
is admissible for the set f > t where f
h
is dened
by (322.2).
9.4 Let L = C
c
(R).
(i) Prove Dinis Theorem: If f
i
is a nonincreasing sequence in L that
converges pointwise to 0, then f
i
0 uniformly.
(ii) Dene T : L R by
T(f) =
_
f
where the integral denotes the Riemann integral. Prove that T is a Daniell
integral.
9.5 Roughly speaking, it can be shown that the measures
+
and
that occur
in Corollary 324.1 are carried by disjoint sets. This can be made precise in
the following way. Prove, for an arbitrary f L
+
, there exists a -measurable
function g on X such that
g(x) =
_
_
_
f(x) for
+
-a.e. x
0 for
-a.e. x
Section 9.2
9.6 Provide an alternate (and simpler) proof of Urysohns Lemma (Lemma 325.1)
in the case when X is a locally compact metric space.
9.7 Suppose X is a locally compact Hausdor space.
(i) If f C
c
(X) is nonnegative, then a f is a compact G
set for
all a > 0.
(ii) If K X is a compact G
sets.
9.8 Prove that the Baire sets and Borel sets are the same in a locally compact
Hausdor space satisfying the second axiom of countability.
9.9 Consider (X, /, ) where X is a compact Hausdor space and / is the family
of Borel sets. If
i
is a sequence of Radon measures with
i
(X) M for
some positive M, then Alaoglus Theorem (Theorem 294.1) implies there is
a subsequence
i
j
that converges weak
i
(F) (F) whenever F is a closed set.
(ii) liminf
i
i
(U) (U) whenever U is an open set.
334 9. MEASURES AND LINEAR FUNCTIONALS
(iii) lim
i
i
(E) = (E) whenever E is a measurable set with
(E) = 0.
9.11 Assume that X is a locally compact Hausdor space that can be written as the
countable union of compact sets. Then, under the hypotheses and notation of
the Riesz Representation Theorem, prove that there is a (nonnegative) Baire
outer measure and a -measurable function g such that
(i) [g(x)[ = 1 for -a.e. x, and
(ii) T(f) =
_
X
fg d for all f C
c
(X).
CHAPTER 10
Distributions
10.1. The Space D
In the previous chapter, we saw how a bounded linear functional on the
space C
c
(R
n
) can be identied with a measure. In this chapter, we will
pursue this idea further by considering linear functionals on a smaller
space, thus producing objects called distributions that are more general
than measures. Distributions are of fundamental importance in many
areas such as partial dierential equations, the calculus of variations,
and harmonic analysis. Their importance was formally acknowledged
by the mathematics community when Laurent Schwartz, who initiated
and developed the theory of distributions, was awarded the Fields Medal
at the 1950 International Congress. In the next chapter some applica-
tions of distributions will be given. In particular, it will be shown how
distributions are used to obtain a solution to a fundamental problem in
partial dierential equations, namely, the Dirichlet Problem. We begin
by introducing the space of functions on which distributions are dened.
Let R
n
be a bounded open set. We begin by investigating a space that
is much smaller than C
c
(), namely, the space T() of all innitely dierentiable
functions whose supports are contained in . We let C
k
(), 1 k , denote
the class of functions dened on whose partial derivatives of all orders up to and
including k are continuous. Also, we denote by C
k
c
() those functions in C
k
()
whose supports are contained in .
It is not immediately obvious that such functions exist, so we begin by analyzing
the following function dened on R:
f(x) =
_
_
_
e
1/x
x > 0
0 x 0.
Observe that f is C
(x) =
_
_
_
1
x
2
e
1/x
x > 0
0 x < 0
and therefore
lim
x0
f
(x) = 0.
335
336 10. DISTRIBUTIONS
Moreover, since
lim
h0
f(h) f(0)
h
lim
h0
e
1/h
h
= 0,
it follows that f
r(h)e
1/h
h
= 0,
thus showing that f
(k)
exists and is continuous at x = 0. To obtain a smooth
function with compact support in R
n
let
F(x) = f(1 [x[
2
).
With x = (x
1
, x
2
, . . . , x
n
), observe that 1[x[
2
is the polynomial 1(x
2
1
+x
2
2
+ +
x
2
n
) and therefore, that F is an innitely dierentiable function of x. Moreover, F
is nonnegative and is zero for [x[ 1.
It is traditional in many parts of analysis to denote C
i=1
i
.
Using this notation, higher order derivatives are denoted by
D
=
||
1
1
, . . . , x
n
n
For example,
xy
= D
(1,1)
and
3
2
xy
= D
(2,1)
.
10.1. THE SPACE D 337
Also, we let (x) denote the gradient of at x, that is,
(336.2)
(x) =
_
x
1
(x),
x
2
(x), . . . ,
x
n
(x)
_
= (D
1
(x), D
2
(x), . . . , D
n
(x)).
We have shown that there exist C
(x) =
n
_
x
_
.
Then
(x) d(x) = 1.
Thus, we have shown that T() is nonempty; we will often call functions in T()
test functions.
We now show how the functions
func-
tions with compact support. Given a function f L
1
loc
(R
n
), recall the denition
(Denition 198.1) of the convolution f
given by
(337.3) f
(x) =
_
R
n
f(x y)
(y) d(y)
for all x R
n
. We will use the notation f
: = f
. The function f
is called the
mollier of f.
337.1. Theorem.
(i) If f L
1
loc
(R
n
), then for every > 0, f
(R
n
).
(ii)
lim
0
f
(x) = f(x)
whenever x is a Lebesgue point for f. In case f is continuous, then f
con-
verges to f uniformly on compact subsets of R
n
.
(iii) If f L
p
(R
n
), 1 p < , then f
L
p
(R
n
), |f
|
p
|f|
p
, and lim
0
|f
f|
p
= 0.
Proof. Recall that D
i
f denotes the partial derivative of f with respect to the
i
th
variable. The proof that f
is C
f) = (D
i
) f
338 10. DISTRIBUTIONS
for each i = 1, 2, . . . , n. To see this, assume (337.4) for the moment. The right side
of the equation is a continuous function (Exercise 2), thus showing that f
is C
1
.
Then, if denotes the n-tuple with [[ = 2 and with 1 in the i
th
and j
th
positions,
it follows that
D
f) = D
i
[D
j
(
f)]
= D
i
[(D
j
) f]
= (D
) f,
which proves that f
C
2
since again the right side of the equation is continuous.
Proceeding this way by induction, it follows that f C
.
We now turn to the proof of (337.4). Let e
1
, . . . , e
n
be the standard basis of
R
n
and consider the partial derivative with respect to the i
th
variable. For every
real number h, we have
f
(x +he
i
) f
(x) =
_
R
n
[
(x z +he
i
)
(x z)]f(z) d(z).
Let (t) denote the integrand:
(t) = [
(x z +te
i
)
(x z)]f(z).
Then is a C
(t) =
(x z +te
i
) e
i
f(z)
= D
i
(x z +te
i
)f(z).
Since
(h) (0) =
_
h
0
(t) dt,
we have
f
(x +he
i
) f
(x)
=
_
R
n
[
(x z +he
i
)
(x z)]f(z) d(z)
=
_
R
n
(h) (0) d(z)
=
_
R
n
_
h
0
D
i
(x z +te
i
)f(z) dt d(z)
=
_
h
0
_
R
n
D
i
(x z +te
i
)f(z) d(z) dt.
It follows from Lebesgues Dominated Convergence Theorem that the inner integral
is a continuous function of t. Now divide both sides of the equation by h and take
10.2. BASIC PROPERTIES OF DISTRIBUTIONS 339
the limit as h 0 to obtain
D
i
f
(x) =
_
R
n
D
i
(x z)f(z) d(z) = (D
i
) f(x),
which establishes (337.4) and therefore (i).
In case (ii), observe that
(339.1)
[f
(x) f(x)[
_
R
n
(x) f(x)[ M,
where M is the product of max
R
n and the Lebesgue measure of the unit ball.
Since is arbitrary, this shows that f
converges uniformly to f on K.
The rst part of (iii) follows from Theorem 199.1 since |
|
1
= 1.
Finally, addressing the second part of (iii), for each > 0, select a continuous
function g with compact support on R
n
such that
(339.2) |f g|
p
<
(see Exercise 10). Because g has compact support, it follows from (ii) that |g g
|
p
<
for all suciently small. Now apply Theorem 337.1 (iii) and (339.2) to the dif-
ference g f to obtain
|f f
|
p
|f g|
p
+|g g
|
p
+|g
|
p
3.
This shows that |f f
|
p
0 as 0.
10.2. Basic Properties of Distributions
339.1. Definition. Let R
n
be an open set. A linear functional T on T()
is a distribution if and only if for every compact set K , there exist constants
C and N such that
[T()[ C sup
xK
||N
D
(x)
340 10. DISTRIBUTIONS
for all test functions T() with support in K. If the integer N can be chosen
independent of the compact set K, and N is the smallest possible choice, the
distribution is said to be of order N. We use the notation
||
K;N
: = sup
xK
||N
D
(x).
Thus, in particular, ||
K;0
denotes the sup norm of on K.
Here are some examples of distributions. First, suppose is a signed Radon
measure on . Dene the corresponding distribution by
T() =
_
(x)d(x)
for every test function . Note that the integral is nite since is nite on compact
sets, by denition. If K is compact, and C = [[ (K) (recall the notation in
175.1), then
[T()[ C ||
L
= C ||
K;0
for all test functions with support in K. Thus, the distribution T corresponding
to the measure is of order 0. In the context of distribution theory, a Radon
measure will be identied as a distribution in this way. In particular, consider the
Dirac measure whose total mass is concentrated at the origin:
(E) =
_
_
_
1 0 E
0 0 , E.
The distribution identied with this measure is dened by
(340.1) T() = (0),
for every test function . test As another example, let f L
1
loc
(). The distribution
corresponding to f is dened as
T() =
_
(x)f(x) d(x).
Thus, a locally integrable function can be considered as an absolutely continuous
measure and is therefore identied with a distribution of order 0.
We have just seen that a Radon measure is a distribution of order 0. The
following result shows that we can actually identify measures and distributions of
order 0.
340.1. Theorem. A distribution T is a Radon measure if and only if T is of
order 0.
10.2. BASIC PROPERTIES OF DISTRIBUTIONS 341
Proof. Assume that T is a distribution of order 0. We will show that T can
be extended in a unique way to a linear functional T
on the space C
c
() so that
(340.2) [T
()[ C ||
;0
(see also Exercise 340.2). Then, by appealing to the Riesz Representation Theorem,
it will follow that there is a (signed) Radon measure such that
T
() =
_
d
for each C
c
(). In particular, we will have
T() = T
() =
_
d
whenever T(), thus establishing that is the measure identied with T.
In order to prove (340.2), select a continuous function with compact support
in . By using molliers (introduced in (337.3)) and Theorem 337.1, it follows that
there is a sequence of test functions
i
T() whose supports are contained in
a xed compact neighborhood of spt such that
|
i
|
;0
0 as i .
Now dene
T
() = lim
i
T(
i
).
The limit exists because when i, j , we have
[T(
i
) T(
j
)[ = [T(
i
j
)[ C |
i
j
|
;0
0.
Similar reasoning shows that the limit is independent of the sequence chosen. Fur-
thermore, since T is of order 0, it follows that
T
() C ||
;0
,
which establishes (340.2).
We conclude this section with a simple but very useful condition that ensures
that a distribution is a measure.
341.1. Definition. A distribution on is positive if T() 0 for all test
functions on satisfying 0.
341.2. Theorem. A distribution T on is positive if and only if T is a positive
measure.
342 10. DISTRIBUTIONS
Proof. From previous discussions, we know that a Radon measure is a dis-
tribution is of order 0, so we need only consider the case when T is a positive
distribution. Let K be a compact set. Then, by Exercise 1, there exists a
function C
c
() that equals 1 on a neighborhood of K. Now select a test
function whose support is contained in K. Then we have
||
K;0
(x) (x) ||
K;0
(x) for all x,
and therefore
||
K;0
T() T() ||
K;0
T().
Thus, [T()[ [T()[ ||
K;0
, which shows that T of order 0 and therefore a
measure.
10.3. Dierentiation of Distributions
One of the primary reasons why distributions were created was to pro-
vide a notion of dierentiability for functions that are not dierentiable
in the classical sense. In this section, we dene the derivative of a dis-
tribution and investigate some of its properties.
In order to motivate the denition of a distribution, consider the following sim-
ple case. Let denote the open interval (0, 1) in R, and let f denote an absolutely
continuous function on (0, 1). If is a test function in , we can integrate by parts
(Exercise 20) to obtain
(342.1)
_
f
(x)(x) d(x) =
_
f(x)
(x) d(x),
since by denition, (1) = (0) = 0. If we consider f as a distribution T, we have
T() =
_
f(x)(x) d(x)
for every test function in (0, 1). Now dene a distribution S by
S() =
_
f(x)
(x) d(x)
and observe that it is a distribution of order 1. From (342.1) we see that S can be
identied with f
should be dened
as
T
() = T(
)
for every test function . More generally, we have the following denition.
342.1. Definition. Let T be a distribution of order N dened on an open set
R
n
. The partial derivative of T with respect to the i
th
coordinate direction is
dened by
T
x
i
() = T(
x
i
).
10.3. DIFFERENTIATION OF DISTRIBUTIONS 343
Observe that since the derivative of a test function is again a test function, the
dierentiated distribution is a linear functional on T(). It is in fact a distribution
since
T
_
x
i
_
C ||
N+1
is valid whenever is a test function supported by a compact set K on which
[T()[ C ||
N
.
Let us consider some examples. The rst one has already been discussed above,
but is repeated for emphasis.
1. Let = [a, b], and suppose f is an absolutely continuous function dened on
[a, b]. If T is the distribution corresponding to f, we have
T() =
_
b
a
f d
for each test function . Since f is absolutely continuous and has compact
support in [a, b] (so that (b) = (a) = 0), we may employ integration by parts
to conclude
T
() = T(
) =
_
b
a
f d =
_
b
a
f
d.
Thus, the distribution T
.
The next example shows how it is possible for a function to have a derivative
in the sense of distributions, but not be dierentiable in the classical sense.
2. Let = R and dene
f(x) =
_
_
_
1 x > 0
0 x 0
With T dened as the distribution corresponding to f, we obtain
T
() = T(
) =
_
R
f d =
_
0
d = (0).
Thus, the derivative T
x(x) d(x),
344 10. DISTRIBUTIONS
so that, after integrating by parts, we obtain
T
() =
_
0
(x) d(x)
_
0
(x) d(x)
=
_
R
(x)g(x) d(x)
where
g(x): =
_
_
_
1 x > 0
0 x 0
This shows that the derivative of f is g in the sense of distributions.
One would hope that the basic results of calculus carry over within the frame-
work of distributions. The following is the rst of many that do.
344.1. Theorem. If T is a distribution in R with T
= 0, then T is a constant.
That is, T is the distribution that corresponds to a constant function.
Proof. Let be a test function and observe that =
where
(x): =
_
x
(t) dt.
Since has compact support, it follows that there is a number x
, x
] and therefore
_
(t) dt =
_
x
(t) dt.
Consequently (y) = 0 for all [y[ x
,
we have
_
y
(t) dt =
_
y
x
(t) dt = 0.
We conclude that has compact support precisely when
(344.1)
_
(t) dt = 0.
Thus, is the derivative of another test function if and only if (344.1) holds. To
prove the theorem, choose an arbitrary test function with the property
_
(x) d(x) = 1.
Any test function can be written as
(x) = [(x) a(x)] +a(x)
10.3. DIFFERENTIATION OF DISTRIBUTIONS 345
with a R, in particular with
a :=
_
(t) d(t).
The function in brackets is a test function, call it , and therefore (x) = (x) +
a(x). Observe that
_
= 0 and therefore is the derivative of another test
function: =
= 0, we know that
0 = T
() = T(
) = T() = T( a),
and therefore it follows that
T() = aT() =
__
(t) d(t)
_
T()
=
_
T()(t) d(t)
which shows that T corresponds to the constant T().
This result along with Example 1 gives an interesting characterization of abso-
lutely continuous functions.
345.1. Theorem. Suppose f L
1
loc
(a, b). Then f is equal almost everywhere
to an absolutely continuous function if and only if the derivative of the distribution
corresponding to f is a function.
Proof. In Example 1 we have already seen that if f is absolutely continuous,
then its derivative in the sense of distributions is again a function. Indeed, the
function associated with the derivative of the distribution is f
.
Now suppose that T is the distribution associated with f and that T
= g.
In order to show that f is equal almost everywhere to an absolutely continuous
function, let
h(x) =
_
x
a
g(t) d(t).
Observe that h is absolutely continuous and that h
() =
_
h
=
_
h
=
_
g.
346 10. DISTRIBUTIONS
Thus, S
= g. Since T
() = T(
) =
_
f
d .
Since f is nondecreasing, it generates a Lebesgue-Stieltjes measure
f
. By the
integration by parts formula for such measures (see Exercises 21 and 22), we have
_
f
d =
_
d
f
,
which shows that T
f
= T
f
1
T
f
2
=
f
1
f
2
.
Consequently, the signed measure
f
1
f
2
corresponds to the distributional de-
rivative of f.
10.4. ESSENTIAL VARIATION 347
Conversely, suppose the derivative of f is a signed measure . By the Jor-
dan Decomposition Theorem, we can write as the dierence of two nonnegative
measures, =
1
2
. Let
f
i
(x) =
i
((, x]) i = 1, 2.
From Theorem 99.1, we have that
i
agrees with the Lebesgue-Stieltjes measure
f
i
on all Borel sets. Furthermore, utilizing the formula for integration by parts,
we obtain
T
f
i
() = T
f
i
(
) =
_
f
i
d =
_
d
f
i
=
_
d
i
.
Thus, with g : = f
1
f
2
, we have that g is of bounded variation and that its
distributional derivative is
1
2
= . Since f
being a measure. We
will see that the needed ingredient is a concept of variation that remains unchanged
when the function is altered on a set of measure zero.
347.1. Definition. The essential variation of a function f dened on (a, b)
is
ess V
b
a
f = sup
_
k
i=1
[f(t
i+1
) f(t
i
)[
_
where the supremum is taken over all nite partitions a < t
1
< < t
k+1
< b such
that each t
i
is a point of approximate continuity of f.
Recall that the variation of a function is dened similarly, but without the
restriction that f be approximately continuous at each t
i
. Clearly, if f = g a.e. on
(a, b), then
ess V
b
a
f = ess V
b
a
g.
348 10. DISTRIBUTIONS
A signed Radon measure on (a, b) denes a linear functional on the space of
continuous functions with compact support in (a, b) by
T
() =
_
b
a
d.
The total variation of , ||, is dened as the norm of T
; that is,
(348.1) || = sup
__
d : C
c
(a, b), [[ 1
_
.
Notice that the supremum could just as well be taken over C
functions with
compact support since they are dense in C
c
(a, b) in the topology of uniform con-
vergence.
348.1. Theorem. Suppose f L
1
(a, b). Then f
is given by |f
| = ess V
b
a
f.
Proof. First, under the assumption that ess V
b
a
f < , we will prove that f
| ess V
b
a
f.
For this purpose, choose > 0 and let f
i=1
[f
(t
i+1
) f
(t
i
)[ =
m
i=1
(s)(f(t
i+1
s) f(t
i
s)) d(s)
(s)
m
i=1
[f(t
i+1
s) f(t
i
s)[ d(s)
(s) ess V
b
a
f d(s)
ess V
b
a
f.
In obtaining the last inequality, we have used the fact that
_
(s) d(s) = 1.
Now take the supremum of the left side over all partitions and obtain
_
b
a+
[(f
[ d ess V
b
a
f.
10.4. ESSENTIAL VARIATION 349
Let C
c
(a, b) with [[ 1. Choosing > 0 such that spt (a +, b ), we
obtain
_
b
a
f
d =
_
b
a
(f
d
_
b
a+
[(f
[ d ess V
b
a
f.
Since f
converges to f in L
1
(a, b), it follows that
(349.1)
_
b
a
f
d ess V
b
a
f.
Thus, the distribution S dened for all test functions by
S() =
_
b
a
f
d
is a distribution of order 0 and therefore a measure since by (349.1),
S() =
_
b
a
f
d =
_
b
a
f
||
(a,b);0
d ess V
b
a
f ||
(a,b);0
,
where we have taken
=
||
(a,b);0
.
Since S = f
, we have that f
is a measure with
|f
| = sup
_
_
b
a
f
d : C
c
(a, b), [[ 1
_
ess V
b
a
f,
thus establishing (348.2) as desired.
Now for the opposite inequality. We assume that f
c
(I
h
) with [[ 1. Then, for all
suciently small > 0, we have the mollier
c
(a, b). Thus, with
the help of (337.4) and Fubinis Theorem,
_
I
h
f
d =
_
b
a
f
d =
_
b
a
(f
d
=
_
b
a
f(
d =
_
b
a
f
d |f
| ,
since [
|
L
1
;I
h
|f
| ,
because
_
I
h
f
d =
_
I
h
f
d.
For arbitrary y, z I
h
, we have
f
(z) = f
(y) +
_
z
y
f
d
350 10. DISTRIBUTIONS
and therefore, taking integral averages with respect to y,
_
I
h
[f
(z)[ d
_
I
h
[f
[ d +
_
b
a
[f
[ d.
Thus, from Theorem 337.1 (iii) and (349.2)
[f
[ (z)
3
b a
|f|
L
1
(a,b)
+|f
|
for z I
h
. Since f
(a, b).
Now that we know that f is bounded on (a, b), we see that each point of ap-
proximate continuity of f is also a Lebesgue point (see Exercise 39). Consequently,
for each partition a < t
1
< < t
m+1
< b where each t
i
is a point of approximate
continuity of f, reference to Theorem 337.1 (ii) yields
m
i=1
[f(t
i+1
) f(t
i
)[ = lim
0
m
i=1
[f
(t
i+1
) f
(t
i
)[
limsup
0
_
b
a
[f
[ d
|f| (a, b). by(349.2)
2 Now take the supremum over all such partitions to conclude that
ess V
b
a
f |f
| .
Exercises for Chapter 10
Section 10.1
10.1 Let K be a compact subset of an open set . Prove that there exists f
C
c
() such that f 1 on K.
10.2 Suppose f L
1
loc
(R
n
) and a continuous function with compact support.
Prove that f is continuous.
10.3 If f L
1
loc
(R
n
) and
_
R
n
f dx = 0
for every C
c
(R
n
), show that f = 0 almost everywhere. Hint: use
Theorem 337.1.
10.4 Prove that a linear map L: R
n
R
m
is Lipschitz.
10.5 Show that a linear map L: R
n
R
n
satises condition N and therefore leaves
Lebesgue measurable sets invariant.
Section 10.2
10.6 Use the Hahn-Banach theorem to provide an alternative proof of (340.2).
Section 10.3
EXERCISES FOR CHAPTER 10 351
10.7 If a distribution T dened on R has the property that T
= 0, what can be
said about T?
CHAPTER 11
Functions of Several Variables
11.1. Dierentiability
Because of the central role played by absolutely continuous functions
and functions of bounded variation in the development of the Funda-
mental Theorem of Calculus in R, it is natural to ask whether they
have analogues among functions of more than one variable. One of the
main objectives of this chapter is to show that this is true. We have
found that the BV functions in R comprise a large class of functions
that are dierentiable almost everywhere. Although there are functions
on R
n
that are analogous to BV functions, they are not dierentiable
almost everywhere. Among the functions that are often encountered in
application, Lipschitz functions on R
n
form the largest class that are
dierentiable almost everywhere. This result is due to Rademacher and
is one of the fundamental results in the analysis of functions of several
variables.
The derivative of a function f at a point x
0
satises
lim
h0
f(x
0
+h) f(x
0
) f
(x
0
)h
h
= 0.
The linear function L dened by
L(h) = f
(x
0
)h
serves as an approximation to the dierence quotient
f(x
0
+h) f(x
0
)
h
.
We use this observation to motivate the denition of dierentiability for functions of
several variables. Thus, if f : R
n
R, we say that f is dierentiable at x
0
R
n
provided there is a linear function L: R
n
R with the property that
(353.1) lim
h0
[f(x
0
+h) f(x
0
) L(h)[
[h[
= 0.
The linear function L is called the derivative of f at x
0
and is denoted by df(x
0
). It
is commonly accepted to use the term dierential interchangeably with derivative.
Recall that the existence of partial derivatives at a point is not sucient to en-
sure dierentiability. For example, consider the following function of two variables:
f(x, y) =
_
_
_
xy
2
+x
2
y
x
2
+y
2
if (x, y) ,= (0, 0)
0 if (x, y) = (0, 0).
353
354 11. FUNCTIONS OF SEVERAL VARIABLES
Both partial derivatives are 0 at (0, 0), but yet the function is not dierentiable
there. We leave it to the reader to verify this.
We now consider Lipschitz functions f on R
n
, see (67.1); thus there is a constant
C
f
such that
[f(x) f(y)[ C
f
[x y[
for all x, y R
n
. The next result is fundamental in the study of functions of several
variables.
354.1. Theorem. If f : R
n
R is Lipschitz, then f is dierentiable almost
everywhere.
Proof. For v R
n
with [v[ = 1, and x R
n
, let (t) = f(x +tv). Since f is
Lipschitz, so is and therefore it is dierentiable for almost all t R.
Let df(x, v) denote the directional derivative of f at x. Thus, df(x, v) =
(0)
whenever
. As a result of (354.1), we
see that f(x) exists for -almost all x. To establish (354.2), we begin with an
11.1. DIFFERENTIABILITY 355
observation that follows directly from Theorem 358.1, which will be established
later:
_
R
n
_
f(x +tv) f(x)
t
_
(x) d(x)
=
_
R
n
f(x)
_
(x) (x tv)
t
_
dx
whenever C
c
(R
n
). Letting t assume the values t = 1/k for all nonnegative
integers k, we have
f(x +
1
k
v) f(x)
1
k
C
f
[v[ = C
f
.
Therefore, with Lebesgues Dominated Convergence Theorem and (354.1), we ob-
tain
_
R
n
df(x, v)(x)dx =
_
R
n
f(x)d(x, v)dx
=
_
R
n
f(x)(x) v dx
=
n
j=1
_
R
n
f(x)D
j
(x)v
j
dx
=
n
j=1
_
R
n
D
j
f(x)(x)v
j
dx
=
_
R
n
(x)f(x) v dx.
Note that we have used the absolute continuity of f on lines. Because this is
valid for all C
c
(R
n
), we have that
(355.1) df(x, v) = f(x) v, a.e. x R
n
(see Exercise 3). Now let v
1
, v
2
, . . . be a countable dense subset of S
n1
where
S
n1
= R
n
v : [v[ = 1.
Observe that there is a set E with (E) = 0 such that
(355.2) df(x, v
k
) = f(x) v
k
for all x R
n
E, k = 1, 2, . . ..
We will now show that the conclusion of our theorem holds at all points of
R
n
E. For this purpose, let x R
n
E, [v[ = 1, t > 0 and consider the dierence
quotient
Q(x, v, t): =
f(x +tv) f(x)
t
f(x) v.
356 11. FUNCTIONS OF SEVERAL VARIABLES
For v, v
S
n1
, and t > 0 note that
(355.3)
[Q(x, v, t) Q(x, v
, t)[
[f(x +tv) f(x +tv
)[
t
+[(v v
) f(x)[
C
f
[v v
[ +[v v
[ [f(x)[
C
f
(n + 1)[v v
[
Since the sequence v
i
is dense in S
n1
, for each > 0 there exists an integer K
such that
(356.1) [v v
k
[ <
2(n + 1)C
f
for some k 1, 2, . . . , K
whenever v S
n1
. For x
0
R
n
E, we have from (355.2) the existence of > 0
such that
(356.2) [Q(x
0
, v
k
, t)[ <
2
for 0 < t < , k 1, 2, . . . , K.
Since
[Q(x
0
, v, t)[ [Q(x
0
, v
k
, t)[ +[Q(x
0
, v, t) Q(x
0
, v
k
, t)[
for k 1, 2, . . . , K, it follows from (356.2), (355.3), and (356.1) that
[Q(x
0
, v, t)[ <
2
+
2
=
whenever [v[ = 1 and 0 < t < .
The concept of dierentiability for a transformation T : R
n
R
m
is virtually
the same as in (353.1). Thus, we say that T is dierentiable at x
0
R
n
if there
is a linear mapping L: R
n
R
m
such that
(356.3) lim
h0
[T(x
0
+h) T(x
0
) L(h)[
[h[
= 0.
As in the case when T is real-valued, we call L the derivative of T at x
0
. We will
denote the linear function L by dT(x
0
). Thus, dT(x
0
) is a linear transformation,
and when it is applied to a vector v R
n
, we will write dT(x
0
)(v). Writing T in
terms of its coordinate functions, T = (T
1
, T
2
, . . . , T
m
), it is easy to see that T is
dierentiable at a point x
0
if and only if each T
i
is. Consequently, the following
corollary is immediate.
356.1. Corollary. If T : R
n
R
m
is a Lipschitz transformation, then T is
dierentiable -almost everywhere.
11.2. CHANGE OF VARIABLE 357
11.2. Change of Variable
We give a treatment of the behavior of the integral when the inte-
grand is subjected to a change of variables by a Lipschitz transformation
T : R
n
R
n
.
Consider T : R
n
R
n
with T = (T
1
, T
2
, . . . , T
n
). If T is dierentiable at x
0
,
then it follows immediately from denitions that each of the partial derivatives
T
i
x
j
i, j = 1, 2, . . . , n.
exists at x
0
. The linear mapping L = dT(x
0
) in (356.3) can be represented by the
n n matrix
dT(x
0
) =
_
_
T
1
x
1
T
1
x
n
.
.
.
.
.
.
T
n
x
1
T
n
x
n
_
_
where it is understood that each partial derivative is evaluated at x
0
. The deter-
minant of [dT(x
0
)] is called the Jacobian of T at x
0
and is denoted by JT(x
0
).
Recall from elementary linear algebra that a linear map L: R
n
R
n
can be
identied with the n n matrix (L
ij
) where
L
ij
= L(e
i
) e
j
, i, j = 1, 2, . . . , n
and where e
i
is the standard basis for R
n
. The determinant of L
ij
is denoted by
det L. If M: R
n
R
n
is also a linear map,
(357.1) det(M L) = (det M) (det L.).
Every nonsingular n n matrix (L
ij
) can be row-reduced to the identity matrix.
That is, L can be written as the composition of nitely many linear transformations
of the following three types:
(i) L
1
(x
1
, . . . , x
i
, . . . , x
n
) = (x
1
, . . . , cx
i
, . . . , x
n
), 1 i n, c ,= 0
(ii) L
2
(x
1
, . . . , x
i
, . . . , x
n
) = (x
1
, . . . , x
i
+cx
k
, . . . , x
n
) 1 i n, k ,= i, c ,= 0.
(iii) L
3
(x
1
, . . . , x
i
, . . . , x
j
, . . . , x
n
) = (x
1
, . . . , x
j
, . . . , x
i
, . . . , x
n
), 1 i < j n.
This leads to the following geometric interpretation of the determinant.
357.1. Theorem. If L: R
n
R
n
is a nonsingular linear map, then L(E) is
Lebesgue measurable whenever E is Lebesgue measurable and
(357.2) [L(E)] = [det L[ (E).
Proof. The fact that L(E) is Lebesgue measurable is left as Exercise 5. In
view of (357.1), it suces to prove (357.2) when L is of the three types mentioned
above. In the case of L
3
, we use Fubinis Theorem and interchange the order of
358 11. FUNCTIONS OF SEVERAL VARIABLES
integration. In the case of L
1
or L
2
, we integrate rst respect to x
i
to arrive at the
formulas of the form
[c[
1
(A) =
1
(cA)
1
(A) =
1
(c +A).
358.1. Theorem. If f L
1
(R
n
) and L: R
n
R
n
is a nonsingular linear
mapping, then
_
R
n
f L [det L[ d =
_
R
n
f d.
Proof. Assume rst that f 0. Note that
f > t = L(f L > t)
and therefore
(f > t) = (L(f L > t)) = [det L[ (f L > t)
for t 0. Now apply Theorem 201.1 to obtain our desired result. The general case
follows by writing f = f
+
f
.
Our next result deals with an approximation property of Lipschitz transforma-
tions with nonvanishing Jacobians. First, we recall that a linear transformation
L: R
n
R
n
can be identied with its n n matrix. Let denote the family of
n n matrices whose entries are rational numbers. Clearly, is countable. Fur-
thermore, given an arbitrary nonsingular linear transformation L and > 0, there
exists R such that
(358.1)
[L(x) R(x)[ <
L
1
(x) R
1
(x)
<
whenever [x[ 1. Using linearity, this implies
[L(x) R(x)[ < [x[
L
1
(x) R
1
(x)
< [x[
for all x R
n
. Also, (358.1) implies
[L R
1
(x y)[ (1 +) [x y[
and
R L
1
(x y)
(1 +) [x y[
11.2. CHANGE OF VARIABLE 359
for all x, y R
n
. That is, the Lipschitz constants (see (67.1)) of LR
1
and RL
1
satisfy
(358.2) C
LR
1 < 1 + and C
RL
1 < 1 +.
359.1. Theorem. Let T : R
n
R
n
be a continuous transformation and set
B = x : dT(x) exists, JT(x) ,= 0.
Given t > 1 there exists a countable collection of Borel sets B
k
k=1
such that
(i) B =
k=1
B
k
,
(ii) The restriction of T to B
k
(denoted by T
k
) is univalent,
(iii) For each positive integer k, there exists a nonsingular linear transformation
L
k
: R
n
R
n
such that the Lipschitz constants of T
k
L
1
k
and L
k
T
1
k
satisfy
C
T
k
L
1
k
t, C
L
k
T
1
k
t
and
t
n
[det L
k
[ [JT(x)[ t
n
[det L
k
[
for all x B
k
.
Proof. For xed t > 1 choose > 0 such that
1
t
+ < 1 < t .
Let C be a countable dense subset of B and let (as introduced above) be the
family of linear transformations whose matrices have rational entries. Now, for
each c C, R and each positive integer i, dene E(c, R, i) to be the set of all
b B B(c, 1/i) that satisfy
(359.1)
_
1
t
+
_
[R(v)[ [dT(b)(v)[ (t ) [R(v)[
for all v R
n
and
(359.2) [T(a) T(b) dT(b)(a b)[ [R(a b)[
for all a B(b, 2/i). Since T is continuous, each partial derivative of each coordinate
function of T is a Borel function. Thus, it is an easy exercise to prove that each
E(c, R, i) is a Borel set. Observe that (359.1) and (359.2) imply
(359.3)
1
t
[R(a b)[ [T(a) T(b)[ t [R(a b)[
for all b E(c, R, i) and a B(b, 2/i).
360 11. FUNCTIONS OF SEVERAL VARIABLES
Next, we will show that
(359.4)
_
1
t
+
_
n
[det R[ [JT(b)[ (t )
n
[det R[ .
for b E(c, R, i). For the proof of this, let L = dT(b). Then, from (359.1) we have
(360.1)
_
1
t
+
_
[R(v)[ [L(v)[ (t ) [R(v)[
whenever v R
n
and therefore,
(360.2)
_
1
t
+
_
[v[
L R
1
(v)
(t ) [v[
for v R
n
. This implies that
L R
1
[B(0, 1)] B(0, t )
and reference to Theorem 357.1 yields
det (L R
1
)
k=1
.
To show that property (i) holds, choose b B and let L = dT(b) as above.
Now refer to (358.2) to nd R such that
C
LR
1 <
_
1
t
+
_
1
and C
RL
1 < t .
Using the denition of the dierentiability of T at b, select a nonnegative integer i
such that
[T(a) T(b) dT(b) (a b)[
C
R
1
[a b[ [R(a b)[
for all a B(b, 2/i). Now choose c C such that [b c[ < 1/i and conclude that
b E(c, R, i). Since this holds for all b B, property (i) holds.
To prove (ii), choose any set B
k
. It is one of the sets of the form E(c, R, i) for
some c C, R , and some nonnegative integer i. According to (359.3),
1
t
[R(a b)[ [T(a) T(b)[ t [R(a b)[
11.2. CHANGE OF VARIABLE 361
for all b B
k
, a B(b, 2/i). Since B
k
B(c, 1/i) B(b, 2/i), we thus have
(360.3)
1
t
[R(a b)[ [T(a) T(b)[ t [R(a b)[
for all a, b B
k
. Hence, T restricted to B
k
is univalent.
With B
k
of the form E(c, R, i) as in the preceding paragraph, we dene L
k
= R.
The proof of (iii) follows from (360.3) and (359.4). They imply
C
T
k
L
1
k
t, C
L
k
T
1
k
t,
and
t
n
[det L
k
[ [JT
k
[ t
n
[det L
k
[ ,
since is arbitrary.
We now proceed to develop the analog of Banachs Theorem (Theorem 242.1)
for Lipschitz mappings T : R
n
R
n
. As in (242.1), for E R
n
we dene
N(T, E, y)
as the (possibly innite) number of points in E T
1
(y).
361.1. Lemma. Let T : R
n
R
n
be a Lipschitz transformation. If
E := x R
n
: T is dierentiable at x and JT(x) = 0,
then
(T(E)) = 0
Proof. By Rademachers theorem, we know that T is dierentiable almost
everywhere. Since each entry of dT is a measurable function, it follows that the
set E is measurable and therefore, so is T(E) since T preserves sets of measure
zero. It is sucient to prove that T(E
R
) has measure zero for each R > 0 where
E
R
:= E B(0, R). Let (0, 1) and x x E
R
. Since T is dierentiable at x,,
there exists 0 <
x
< 1 such that
(361.1) [T(y) T(x) L(y x)[ [y x[ for all y B(x,
x
)
where L := dT(x). We know that JT(x) = 0, and so L is represented by a singular
matrix. Therefore, there is a linear subspace H of R
n
of dimension n1 such that
L maps R
n
into H. From (361.1) and the triangle inequality, we have
[T(y) T(x)[ [L(y x)[ + [y x[
362 11. FUNCTIONS OF SEVERAL VARIABLES
Let
L
denote the Lipschitz constant of L: [L(y)[
L
[y[ for all y R
n
. Also, let
T
denote the Lipschitz constant of T. Hence, for x E
R
, 0 < r <
x
< 1 and
y B(x, r) it follows that
[L(y) L(x)[
L
[y x[
L
[[y[ +R]
=
L
[[y x +x[ +R]
L
[r + 2R].
This implies
[L(y)[ [L(x)[ +
L
[r + 2R]
L
R +
L
[r + 2R] =
L
[r + 3R].
With v := T(x) L(x) we see that [v[ [T(x)[ + [L(x)[
L
[x[ +
L
[x[
R(
T
+
L
) and that (361.1) becomes
[T(y) L(y) v[ [y x[ ;
that is, T(y) is within distance r of L(y) translated by the vector v. This means
that the set T(B(x, r)) is contained within an r-neighborhood of a bounded set in
an ane space of dimension n1. The bound on this set depends only on r, R,
L
and
T
. Thus there is a constant M = M(R,
L
,
T
) such that
(T(B(x, r))) rMr
n1
.
The collection of balls B(x, r) with x E
R
and 0 < r <
x
determines a Vitali
covering of E
R
and accordingly there is a countable disjoint subcollection B
i
:=
B(x
i
, r
i
) whose union contains almost all of E
R
:
(F) = 0 where F := E
R
i=1
B
i
.
Since T is Lipschitz we know that (T(F)) = 0. Therefore, because
E
R
F
i=1
B
i
,
we have
T(E
R
) T(F)
i=1
T(B
i
),
and thus,
(T(E
R
))
i=1
T(B
i
)
i=1
r
i
Mr
n1
i
M
i=1
(B(x
i
, r
i
)).
11.2. CHANGE OF VARIABLE 363
Since the balls B(x
i
, r
i
) are disjoint and all contained in B(0, R+1) and since > 0
was chosen arbitrarily, we conclude that (T(E
R
)) = 0, as desired.
363.1. Theorem. Let T : R
n
R
n
be Lipschitz. Then for each Lebesgue
measurable set E R
n
,
_
E
[JT(x)[ d(x) =
_
R
n
N(T, E, y) d(y).
Proof. Since T is Lipschitz, we know that T carries sets of measure zero into
sets of measure zero. Thus, by Rademachers Theorem, we might as well assume
that dT(x) exists for all x E. Furthermore, it is easy to see that we may assume
(E) < .
In view of Lemma 361.1 it suces to treat the case when [JT[ , = 0 on E. Fix
t > 1 and let B
k
denote the sets provided by Theorem 359.1. By Lemma 82.1,
we may assume that the sets B
k
are disjoint. For the moment, select some set
B
k
and set B = B
k
. For each positive integer j, let Q
j
denote a decomposition of
R
n
into disjoint, half-open cubes with side length 1/j and of the form [a
1
, b
1
)
[a
n
, b
n
). Set
F
i,j
= B E Q
i
, Q
i
Q
j
.
Then, for xed j the sets F
i,j
are disjoint and
B E =
i=1
F
i,j
.
Furthermore, the sets T(F
i,j
) are measurable since T is Lipschitz. Thus, the func-
tions g
j
dened by
g
j
=
i=1
T(F
i,j
)
are measurable. Now g
j
(y) is the number of sets F
i,j
such that F
i,j
T
1
(y) ,= 0
and
lim
j
g
j
(y) = N(T, B E, y).
An application of the Monotone Convergence Theorem yields
(363.1) lim
j
i=1
[T(F
i,j
)] =
_
R
n
N(T, B E, y) d(y).
Let L
k
and T
k
be as in Theorem 359.1. Then, recalling that B = B
k
, we obtain
(363.2)
[T(F
i,j
)] = [T
k
(F
i,j
)]
= [(T
k
L
1
k
L
k
)(F
i,j
)] t
n
[L
k
(F
i,j
)].
(363.3)
[L
k
(F
i,j
)] = [(L
k
T
1
k
T
k
)(F
i,j
)]
t
n
[T
k
(F
i,j
)] = t
n
[T(F
i,j
)].
364 11. FUNCTIONS OF SEVERAL VARIABLES
and thus
t
2n
[T(F
i,j
)] t
n
[L
k
(F
i,j
)] by (363.2)
= t
n
[det L
k
[ (F
i,j
) by Theorem 357.1
_
F
i,j
[JT[ d by Theorem 359.1
t
n
[det L
k
[ (F
i,j
) by Theorem 359.1
= t
n
[L
k
(F
i,j
)] by Theorem357.1
t
2n
[T(F
i,j
)]. by (363.3)
With j xed, sum on i and use the fact that the sets F
i,j
are disjoint:
(364.1) t
2n
i=1
[T(F
i,j
)]
_
BE
[JT[ d t
2n
i=1
[T(F
i,j
)].
Now let j and recall (363.1):
t
2n
_
R
n
N(T, B E, y) d(y)
_
BE
[JT[ d
t
2n
_
R
n
N(T, B E, y) d(y).
Since t was initially chosen as any number larger than 1, we conclude that
(364.2)
_
BE
[JT[ d =
_
R
n
N(T, B E, y) d(y).
Now B was dened as an arbitrary set of the sequence B
k
that was provided by
Theorem 359.1. Thus (364.2) holds with B replaced by an arbitrary B
k
. Since the
sets B
k
are disjoint and both sides of (364.2) are additive relative to B
k
, the
Monotone Convergence Theorem implies
_
E
[JT[ d =
_
R
n
N(T, E, y) d(y).
We have reached this conclusion under the assumption that [JT[ ,= 0 on E.
The proof will be concluded if we can show that
[T(E
0
)] = 0
where E
0
= E x : JT(x) = 0. Note that E
0
is a Borel set. Note also that
nothing is lost if we assume E
0
is bounded. Choose > 0 and let U E
0
be a
bounded, open set with (U E
0
) < . For each x E
0
, there exists
x
> 0 such
that B(x,
x
) U and
[T(x +h) T(x)[ < [h[
for all h R
n
with [h[ <
x
. In other words,
(364.3) T[B(x, r)] B(T(x), r)
11.2. CHANGE OF VARIABLE 365
whenever r <
x
. The collection of all balls B(x, r) where x E
0
and r <
x
provides a Vitali covering of E
0
in the sense of Denition 222.1. Thus, by Theorem
223.2, there exists a disjoint, countable subcollection B
i
such that
_
E
0
i=1
B
i
_
= 0.
Note that B
i
U. Let us suppose that each B
i
is of the form B
i
= B(x
i
, r
i
).
Then, using the fact that T carries sets of measure zero into sets of measure zero,
[T(E
0
)] =
i=1
[T(B
i
)]
i=1
[B(T(x
i
), r
i
)] by (364.3)
=
i=1
(n)(r
i
)
n
=
n
i=1
(n)r
n
i
=
n
i=1
(B
i
)
n
(U). since the B
i
are disjoint
Since (U) < (because U is bounded) and is arbitrary, we have [T(E
0
)] =
0.
365.1. Corollary. If T : R
n
R
n
is Lipschitz and univalent, then
_
E
[JT(x)[ d(x) = [T(E)]
whenever E R
n
is measurable.
365.2. Theorem. Let T : R
n
R
n
be a Lipschitz transformation and suppose
f : R
n
R is Lebesgue measurable. Then
_
R
n
f T [JT[ d =
_
R
n
f(y)N(T, R
n
, y) d(y).
Proof. First, suppose f is the characteristic function of a measurable set
A R
n
. Then f T =
T
1
(A)
and
_
R
n
f T [JT[ d =
_
T
1
(A)
[JT[ d =
_
R
n
N(T, T
1
(A), y) d(y)
=
_
R
n
A
(y)N(T, R
n
, y) d(y).
=
_
R
n
f(y)N(T, R
n
, y) d(y).
366 11. FUNCTIONS OF SEVERAL VARIABLES
Clearly, this also holds whenever f is a simple function. Reference to Theorem
147.1 and the Monotone Convergence Theorem shows that it then holds whenever
f is nonnegative. Finally, writing f = f
+
f
= (x
2
, . . . , x
n
) where
= [x
[ = r sin
1
. The coordinates of x are
x
1
= r cos
1
: = T
1
(r,
1
, . . . ,
n
)
x
2
= r sin
1
cos
2
: = T
2
(r,
1
, . . . ,
n
)
.
.
.
x
n1
= r sin
1
sin
2
. . . sin
n2
cos
n1
: = T
n1
(r,
1
, . . . ,
n
)
x
n
= r sin
1
sin
2
. . . sin
n2
sin
n1
: = T
n
(r,
1
, . . . ,
n
).
The mapping T = (T
1
, T
2
, . . . , T
n
) is a bijection of
(0, ) (0, )
n2
(0, 2)
onto
R
n
(R
n2
[0, ) 0).
A straightforward calculation shows that the Jacobian is
JT(r,
1
, . . . ,
n1
) = r
n1
sin
n2
1
sin
n3
2
. . . sin
n2
.
Hence, with = (
1
, . . . ,
n1
), we obtain
(367.1)
_
A
f T(r, )r
n1
sin
n2
1
sin
n3
2
. . . sin
n2
dr d
1
. . .
n1
=
_
T(A)
f d.
11.3. Sobolev Functions
It is not obvious that there is a class of functions dened on R
n
that is
analogous to the absolutely continuous functions on R. However, it is
tempting to employ the multi-dimensional analogue of Theorem 247.1
which states roughly that a function is absolutely continuous if and only
if its derivative in the sense of distributions is a function. We will follow
this direction and by considering functions whose partial derivatives are
functions and show that this denition leads to a fruitful development.
367.1. Definition. Let R
n
be an open set and let f L
1
loc
(). We use
Denition 342.1 to dene the partial derivatives of f in the sense of distributions.
Thus, for 1 i n, we say that a function g
i
L
1
loc
() is the i
th
partial
derivative of f in the sense of distributions if
(367.2)
_
f
x
i
d =
_
g
i
d
for every test function C
c
(). We will write
f
x
i
= g
i
;
368 11. FUNCTIONS OF SEVERAL VARIABLES
thus,
f
x
i
is merely dened to be the function g
i
.
At this time we cannot assume that the partial derivative
f
x
i
exists in the
classical sense for the Sobolev function f. This existence will be discussed later
after Theorem 370.3 has been proved. Consistent with the notation introduced in
(336.1), we will sometimes write D
i
f to denote
f
x
i
.
The denition in (367.2) is a restatement of Denition 342.1 with the require-
ment that the derivative of f (in the sense of distributions) is again a function
and not merely a distribution. This requirement imposes a condition on f and the
purpose of this section is to see what properties f must possess in order to satisfy
this condition.
In general, we can dene what it means for a higher order derivative of f to be
a function. If is a multi-index (see Section 9.1), then g
L
1
loc
() is called the
th
distributional derivative of f if
_
fD
d = (1)
||
_
d
for every test function C
c
(). We will write D
f : = g
. For 1 p and
k a nonnegative integer, we say that f belongs to the Sobolev Space
W
k,p
()
if D
f L
p
() for each multi-index with [[ k. In particular, this implies
that f L
p
(). Similarly, the space
W
k,p
loc
()
consists of all f with D
f L
p
loc
() for [[ k.
In order to motivate the denition of the distributional partial derivative, we
recall the classical Gauss-Green Theorem.
368.1. Theorem. Suppose that U is an open set with smooth boundary and
let (x) denote the unit exterior normal to U at x U. If V : R
n
R
n
is a
transformation of class C
1
, then
(368.1)
_
U
divV d =
_
U
V (x) (x) dH
n1
(x)
where divV , the divergence of V = (V
1
, . . . , V
n
), is dened by
divV =
n
i=1
V
i
x
i
.
11.3. SOBOLEV FUNCTIONS 369
Now suppose that f : R is of class C
1
and C
c
(). Dene V : R
n
R
n
to be the transformation whose coordinate functions are all 0 except for the i
th
one,
which is f. Then
divV =
(f)
x
i
= f
x
i
+
f
x
i
.
Since the support of is a compact set contained in , it is possible to nd an
open set U with smooth boundary containing the support of such that U .
Then,
_
divV d =
_
U
divV d =
_
U
V dH
n1
= 0.
Thus,
(369.1)
_
f
x
i
d =
_
f
x
i
d,
which is precisely (367.2) in case f C
1
(). Note that for a Sobolev function f,
the formula (369.1) is valid for all test functions C
c
() by the denition of
the distributional derivative. The Sobolev norm of f W
1,p
() is dened by
(369.2) |f|
1,p;
: = |f|
p;
+
n
i=1
|D
i
f|
p;
,
for 1 p < and
|f|
1,;
: = ess sup
([f[ +
n
i=1
[D
i
f[).
One can readily verify that W
1,p
() becomes a Banach space when it is endowed
with the above norm (Exercise 5).
369.1. Remark. When 1 p < , it can be shown (Exercise6) that the norm
(369.3) |f| : = |f|
p;
+
_
n
i=1
|D
i
f|
p
p;
_
1/p
is equivalent to (369.2). Also, it is sometimes convenient to regard W
1,p
() as a
subspace of the Cartesian product of spaces L
p
(). The identication is made by
dening P : W
1,p
()
n+1
i=1
L
p
() as
P(f) = (f, D
1
f, . . . , D
n
f) for f W
1,p
().
In view of Exercise 8.2, it follows that P is an isometric isomorphism of W
1,p
()
onto a subspace W of this Cartesian product. Since W
1,p
() is a Banach space, W
is a closed subspace. By Exercise 8.22, W is reexive and therefore so is W
1,p
().
370 11. FUNCTIONS OF SEVERAL VARIABLES
369.2. Remark. Observe f W
1,p
() is determined only up to a set of
Lebesgue measure zero. We agree to call the Sobolev function f continuous,
bounded, etc. if there is a function f with these respective properties such that
f = f almost everywhere.
We will show that elements in W
1,p
() have representatives that permit us
to regard them as generalizations of absolutely continuous functions on R
1
. First,
we prove an important result concerning the convergence of regularizers of Sobolev
functions.
370.1. Notation. If
to
signify that the closure of
is a subset of .
Also, we will frequently use dx instead of d(x) to denote integration with respect
to Lebesgue measure.
370.2. Lemma. Suppose f W
1,p
(), p 1. Then the molliers, f
, of f (see
(337.3)) satisfy
lim
0
|f
f|
1,p;
= 0
whenever
.
Proof. Since
, ). For <
0
, we may dierentiate under the integral sign (see the proof
of (337.4)) to obtain for x
and 1 i n,
f
x
i
(x) =
n
_
x
i
_
x y
_
f(y) d(y)
=
n
_
y
i
_
x y
_
f(y) d(y)
=
n
_
_
x y
_
f
y
i
d(y) by(367.2)
=
_
f
x
i
_
(x).
Our result now follows from Theorem 337.1.
Since the denition of Sobolev functions requires that their distributional deriva-
tives belong to L
p
, it is natural to inquire if they have partial derivatives in the
classical sense. To this end, we begin by showing that their partial derivatives exist
almost everywhere. That is, in keeping with Remark 369.2, we will show that there
is a function f
such that f
exist
almost everywhere.
In the next theorem, the set R is an interval of the form
(370.1) R: = (a
1
, b
1
) (a
n
, b
n
).
11.3. SOBOLEV FUNCTIONS 371
370.3. Theorem. Suppose f W
1,p
(), p 1. Let R . Then f has
a representative f
f|
1,p;R
0 as 0. Choose a sequence
k
0 and let f
k
: = f
k
. Also
write x R as x = (x
, t) where
x
R
i
: = (a
1
, b
1
) (a
i
, b
i
)
omitted
(a
n
, b
n
)
and t (a
i
, b
i
), 1 i n. Then it follows that
lim
k
_
R
i
_
b
i
a
i
[f
k
(x
, t) f(x
, t)[
p
+[f
k
(x
, t) f(x
, t)[
p
dt d(x
) = 0.
As in the classical setting, we use f to denote (D
1
f, . . . , D
n
f) which is dened in
terms of the distributional derivatives of f. The inner integral is a function of x
.
Since mean convergence implies a.e convergence for some subsequence (which will
still be denoted as the full sequence), we have
(371.1) lim
k
_
b
i
a
i
[f
k
(x
, t) f(x
, t)[
p
+[f
k
(x
, t) f(x
, t)[
p
dt = 0
for
n1
-almost all x
R
i
. From Holders inequality we have
_
b
i
a
i
[f
k
(x
, ) f(x
, )[ dt
(b
i
a
i
)
1/p
_
_
b
i
a
i
[f
k
(x
, t) f(x
, t)[
p
dt
_
1/p
.
If [a, b] [a
i
, b
i
], then
(371.2)
[f
k
(x
, b) f
k
(x
, a)[
_
b
a
[f(x
, t)[ dt
+
_
b
i
a
i
[f
k
(x
, t) f(x
, t)[ dt.
Consequently, it follows from (371.1) and (371.2) that there is a constant M
x
such that
[f
k
(x
, b) f
k
(x
, a)[
_
b
a
[f(x
, t)[ dt + 1
whenever k > M
x
and a, b [a
i
, b
i
]. This implies that for each x
under consider-
ation, the functions f
k
(x
IF
[f
k
(x
, b
I
) f
k
(x
, a
I
)[ <
for all positive integers k. Here, as in Denition 237.1, the endpoints of the in-
terval I are denoted by a
I
, b
I
. In particular, the sequence is equicontinuous on
[a
i
, b
i
]. Passing to another subsequence, it follows from (371.1) that the sequence
f
k
(x
, t) converges for
1
-almost every t [a
i
, b
i
]. Hence, we can conclude from
(371.2) that the sequence is uniformly bounded on [a
i
, b
i
]. Now the Arzel`a-Ascoli
Theorem applies, and we nd that there is a subsequence that converges uniformly
to a function on [a
i
, b
i
], call it f
i
(x
i
(x
, ) is absolutely continuous.
To summarize what has been done so far, recall that for each interval R
of the form (370.1) and each 1 i n, there is a representative of f, f
i
, that
is absolutely continuous in t for
n1
-almost all x
R
i
. This representative was
obtained as the pointwise a.e. limit of a subsequence of molliers of f. Observe
that there is a single subsequence and a single representative that can be found for
all i simultaneously. This can be accomplished by rst obtaining a subsequence and
a representative for i = 1. Then a subsequence can be extracted from this one that
can be used to dene a representative for i = 1 and 2. Continuing in this way, the
desired subsequence and representative are obtained. Thus, there is a sequence of
molliers, denoted by f
k
, and a function f
, )
converges uniformly to f
(x
, ) for
n1
-almost all x
. Furthermore, f
(x
, t) is
absolutely continuous in t for
n1
-almost all x
.
The proof that classical partial derivatives of f
c
(R). Since f
D
i
dx =
_
R
i
_
b
i
a
i
f
D
i
dx
i
dx
=
_
R
i
_
b
i
a
i
(D
i
f
) dx
i
dx
=
_
R
(D
i
f
) dx
11.4. APPROXIMATING SOBOLEV FUNCTIONS 373
Since f = f
D
i
dx
and therefore,
_
R
D
i
f dx =
_
R
D
i
f
dx
for every C
c
(R). This implies that the partial derivatives of f
agree with
the distributional derivatives almost everywhere (see Exercise 10.3). To prove the
converse, suppose that f has a representative f
c
(R), f
)
x
i
(x
, t) dt = 0
for
n1
-almost all x
R
i
and therefore,
_
b
i
a
i
f
(x
, t)
x
i
(x
, t) dt =
_
b
i
a
i
f
x
i
(x
, t) (x
, t) dt.
Fubinis Theorem implies
_
R
f
x
i
d =
_
R
f
x
i
d
for all C
c
(R). Since f
x
i
11.4. Approximating Sobolev Functions
We will show that the Sobolev space W
1,p
() can be characterized as
the closure of C
c
(R
n
) such that 0 f 1 and
(i) For each f T, there exists U ( such that (spt) f U,
(ii) If K E is compact, then spt f K ,= 0 for only nitely many f T,
(iii)
fF
f(x) = 1 for each x E. The family T is called a smooth partition of
unity of E subordinate to the open covering (.
374 11. FUNCTIONS OF SEVERAL VARIABLES
Proof. In case E is compact, our desired result follows from the proof of
Theorem 373.1 and Exercise 1.
Now assume E is open and for each positive integer i dene
E
i
= E B(0, i) x : dist (x, E)
1
i
.
Thus, E
i
is compact, E
i
int E
i+1
, and E =
i=1
E
i
. Let (
i
denote the collection
of all open sets of the form
U int E
i+1
E
i2
,
where U ( and where we take E
0
= E
1
= . The family (
i
forms an open
covering of the compact set E
i
int E
i1
. Therefore, our rst case applies and we
obtain a smooth partition of unity, T
i
, subordinate to (
i
, which consists of nitely
many elements. For each x R
n
, dene
s(x) =
i=1
gF
i
g(x).
The sum is well dened since only a nite number of terms are nonzero for each x.
Note that s(x) > 0 for x E. Corresponding to each positive integer i and to each
function g (
i
, dene a function f by
f(x) =
_
_
_
g(x)
s(x)
if x E
0 if x , E
The partition of unity T that we want is comprised of all such functions f.
If E is arbitrary, then any partition of unity for the open set U : U ( is
also one for E.
Clearly, the set
S = C
() f : |f|
1,p;
<
is contained in W
1,p
(). Moreover, the same is true of the closure of S in the
Sobolev norm since W
1,p
() is complete. The next result shows that S = W
1,p
().
374.1. Theorem. If 1 p < , then the space
C
() f : |f|
1,p;
<
is dense in W
1,p
().
Proof. Let
i
be subdomains of such that
i
i+1
and
i=1
i
= . Let
T be a partition of unity of subordinate to the covering
i+1
i1
, where
11.4. APPROXIMATING SOBOLEV FUNCTIONS 375
1
is dened as the null set. Let f
i
denote the sum of the nitely many f T
with spt f
i+1
i1
. Then f
i
C
c
(
i+1
i1
) and
(374.1)
i=1
f
i
1 on .
Choose > 0. For f W
1,p
(), refer to Lemma 370.2 to obtain
i
> 0 such
that
spt ((f
i
f)
i
)
i+1
i1
, (375.1)
|(f
i
f)
i
f
i
f|
1,p;
< 2
i
.
With g
i
: = (f
i
f)
i
, (375.1) implies that only nitely many of the g
i
can fail to
vanish on any given
. Therefore g : =
i=1
g
i
is dened and is an element
of C
(). For x
i
, we have
f(x) =
i
j=1
f
j
(x)f(x),
and by (375.1)
g(x) =
i
j=1
(f
j
f)
j
(x).
Consequently,
|f g|
1,p;
i
j=1
_
_
(f
j
f)
j
f
j
f
_
_
1,p;
< .
Now an application of the Monotone Convergence Theorem establishes our desired
result.
The previous result holds in particular when = R
n
, in which case we get the
following apparently stronger result.
375.1. Corollary. If 1 p < , then the space C
c
(R
n
) is dense in W
1,p
(R
n
).
Proof. This follows from the previous result and the fact that C
c
(R
n
) is
dense in
C
(R
n
) f : |f|
1,p
<
relative to the Sobolev norm (see Exercise 9).
Recall that if f L
p
(R
n
), then |f(x +h) f(x)|
p
0 as h 0. A similar
result provides a very useful characterization of W
1,p
.
376 11. FUNCTIONS OF SEVERAL VARIABLES
375.2. Theorem. Let 1 < p < and suppose R
n
is an open set. If
f W
1,p
() and
, then
h
1
h
1
).
Proof. Assume f W
1,p
() and let
() functions f
k
such that |f
k
f|
1,p;
0 as k .
For each g C
(), we have
g(x +h) g(x)
[h[
=
1
[h[
_
|h|
0
g
_
x +t
h
[h[
_
h
[h[
dt,
so by Jensens inequality (Exercise 32),
1
[h[
_
|h|
0
g
_
x +t
h
[h[
_
p
dt
whenever x
, ). Therefore,
|g(x +h) g(x)|
p
p;
[h[
p
1
[h[
_
|h|
0
_
g
_
x +t
h
[h[
_
p
dx dt
[h[
p1
_
|h|
0
_
[g(x)[
p
dx dt,
or
|g(x +h) g(x)|
p;
[h[ |g|
p;
for all h < . As this inequality holds for each f
k
, it also holds for f.
For the proof of the converse, let e
i
denote the i
th
unit basis vector. By as-
sumption, the sequence
_
f(x +e
i
/k) f(x)
1/k
_
is bounded in L
p
(
)
such that
f(x +e
i
/k) f(x)
1/k
f
i
weakly in L
p
(
). Thus, for C
0
(
),
_
f
i
dx = lim
k
_
_
f(x +e
i
/k) f(x)
1/k
_
(x)dx
= lim
k
_
f(x)
_
(x e
i
/k) (x)
1/k
_
dx
=
_
fD
i
dx.
11.5. SOBOLEV IMBEDDING THEOREM 377
This shows that D
i
f = f
i
in the sense of distributions. Hence, f W
1,p
(
).
11.5. Sobolev Imbedding Theorem
One of the most useful estimates in the theory of Sobolev functions is
the Sobolev inequality. It implies that a Sobolev function is in a higher
Lebesgue class than the one in which it was originally dened. In fact,
for 1 p < n, W
1,p
() L
p
() where p
= np/(n p).
We use the result of the previous section to set the stage for the next denition.
First, consider a domain R
n
whose boundary has Lebesgue measure zero.
Recall that Sobolev functions are only dened almost everywhere. Consequently,
it is not possible in our present state of development to dene what it means for a
Sobolev function to be zero (pointwise) on the boundary of a domain . Instead,
we dene what it means for a Sobolev function to be zero on in a global sense.
377.1. Definition. Let R
n
be an arbitrary open set. The space W
1,p
0
()
is dened as the closure of C
c
() relative to the Sobolev norm. Thus, f W
1,p
0
()
if and only if there is a sequence of functions f
k
C
c
() such that
lim
k
|f
k
f|
1,p;
= 0.
377.2. Theorem. Let 1 p < n and let R
n
be an open set. There is a
constant C = C(n, p) such that for f W
1,p
0
(),
|f|
p
;
C |f|
p;
.
Proof. First, we will assume that p = 1 and f C
0
(). Appealing to the
Fundamental Theorem of Calculus and using the fact that f has compact support,
it follows for each integer i, 1 i n, that
f(x
1
, . . . , x
i
, . . . , x
n
) =
_
x
i
f
x
i
(x
1
, . . . , t
i
, . . . , x
n
) dt
i
,
and therefore,
[f(x)[ =
_
f
x
i
(x
1
, . . . , t
i
, . . . , x
n
) dt
i
.
Consequently,
[f(x)[
n
n1
i=1
__
[f(x
1
, . . . , t
i
, . . . , x
n
)[ dt
i
_ 1
n1
.
This can be rewritten as
[f(x)[
n
n1
__
[f(x)[ dt
1
_ 1
n1
i=2
__
[f(x
1
, . . . , t
i
, . . . , x
n
)[ dt
i
_ 1
n1
.
378 11. FUNCTIONS OF SEVERAL VARIABLES
Only the rst factor on the right is independent of x
1
. Thus, when the inequality
is integrated with respect to x
1
we obtain, with the help of generalized Holders
inequality (see Exercise 35),
_
[f[
1
n1
dx
1
__
[f[ dt
1
_ 1
n1
_
i=2
__
[f[ dt
i
_ 1
n1
dx
1
__
[f[ dt
1
_ 1
n1
_
n
i=2
_
[f[ dx
1
dt
i
_ 1
n1
.
Similarly, integration with respect to x
2
yields
_
[f[
n
n1
dx
1
dx
2
__
[f[ dx
1
dt
2
_ 1
n1
__
[f[ dt
1
dx
2
_ 1
n1
i=3
__
[f[ dx
1
dx
2
dt
i
_ 1
n1
.
Continuing this way for the remaining n 2 steps, we nally arrive at
_
R
n
[f[
n
n1
dx
n
i=1
__
R
n
[f[ dx
_ 1
n1
or
(378.1) |f| n
n1
_
R
n
[f[ dx,
which is the desired result in the case p = 1. The case when 1 p < n is treated by
replacing f by positive powers of [f[. Thus, for q to be determined later, apply our
previous step to g : = [f[
q
.
1
Technically, the previous step requires g C
c
().
However, a close examination of the proof reveals that we only need g to be an
absolutely continuous function in each variable separately. Then,
|f
q
|
n/(n1)
_
R
n
[[f[
q
[ dx
q
_
R
n
_
R
n
[f[
q1
[f[ dx
q
_
_
f
q1
_
_
p
|f|
p
where we have used Holders inequality in the last inequality. Now let q = (n
1)p/(n p) to obtain our result if 1 p < n and f C
c
(). Finally, if f
W
1,p
0
(), let f
i
be a sequence of functions in C
0
() converging to f in the
1
*
11.6. APPLICATIONS 379
Sobolev norm. Then, with p
;
C |f
i
f
j
|
1,p;
.
We leave it as an exercise for the reader to conclude that f
i
f in L
p
(). In
view of the fact that [f
i
[ [f[ in L
p
(), our result now follows.
11.6. Applications
A basic problem in the calculus of variations is to nd a harmonic func-
tion in a domain that assumes values that have been prescribed on the
boundary. Using results of the previous sections, we will discuss a so-
lution to this problem. Our rst result shows that this problem has a
weak solution, that is, a solution in the sense of distributions.
379.1. Definition. A function f C
2
() is called harmonic in if
2
f
x
2
1
(x) +
2
f
x
2
2
(x) + +
2
f
x
2
n
(x) = 0
for each x .
A straightforward calculation shows that this is equivalent to div(f)(x) = 0.
Taking V = f in (368.1) we have
_
f dx = 0
whenever C
c
().
In general, if we let denote the operator
=
2
x
2
1
+
2
x
2
2
+ +
2
x
2
n
,
then
(379.1)
_
f dx =
_
divV f dx =
_
f dx
whenever f C
2
() and C
0
(). Returning to the context of distributions,
recall that a distribution can be dierentiated any number of times. In particular, it
is possible to dene T whenever T is a distribution. If T is taken as an integrable
function f, we have
f() =
_
f dx
for all C
c
(). We will say that f is harmonic in the sense of distributions
or weakly harmonic if f() = 0 for every test function . If f W
1,2
() is
weakly harmonic, then (379.1) implies (with f and interchanged) that
(379.2)
_
f dx = 0
380 11. FUNCTIONS OF SEVERAL VARIABLES
for every test function C
c
() (see Exercise 10). Since f W
1,2
(), note that
(379.2) remains valid with W
1,2
0
().
380.0. Theorem. Suppose R
n
is a bounded open set and let W
1,2
().
Then there exists a weakly harmonic function f W
1,2
() such that f
W
1,2
0
().
The theorem states that for a given function W
1,2
(), there exists a weakly
harmonic function f that assumes the same values (in a weak sense) on as does
. That is, f and have the same boundary values and
_
f dx = 0
for every test function C
c
(). Later we will show that f is, in fact, harmonic
in the sense of Denition 379.1. In particular, it will be shown that f C
c
().
Proof. Let
(380.1) m: = inf
__
[f[
2
dx : f W
1,2
0
()
_
.
This denition requires f W
1,2
0
(). Since W
1,2
(), note that f must be
an element of W
1,2
() and therefore that [f[ L
2
().
Our rst objective is to prove that the inmum is attained. For this purpose,
let f
i
be a sequence such that f
i
W
1,2
0
() and
(380.2)
_
[f
i
[
2
dx m as i .
Now apply both Holders and Sobolevs inequality (Theorem 377.2) to obtain
|f
i
|
2;
()
(
1
2
1
2
)
|f
i
|
2
;
C |(f
i
)|
2;
This along with (380.2) shows that |f
i
|
1,2;
i=1
is a bounded sequence. Referring
to Remark 6, we know that W
1,2
() is a reexive space and therefore Corollary
295.1 implies that there is a subsequence (denoted by the full sequence) and f
W
1,2
() such that f
i
f weakly in W
1,2
(). Furthermore, it follows from the
lower semicontinuity of the norm (Theorem 290.1) that
_
[f[
2
dx liminf
i
_
[f
i
[
2
dx.
To show that f is a valid competitor in (380.1) we need to establish that
f W
1,2
0
(), for then we will have
_
[f[
2
dx = m,
11.7. REGULARITY OF WEAKLY HARMONIC FUNCTIONS 381
thus establishing that the inmum in (380.1) is attained. To show that f assumes
the correct boundary values, note that (for a subsequence) f
i
g weakly for
some g W
1,2
0
() since |f
i
|
1,2;
is a bounded sequence. But f
i
f
weakly in W
1,2
() and therefore f = g W
1,2
0
().
The next step is to show that f is weakly harmonic. Choose C
c
() and
for each real number t, let
(t) =
_
[(f +t)[
2
dx
=
_
[f[
2
+ 2tf +t
2
[[
2
dx.
Note that has a local minimum at t = 0. Furthermore, referring to Exercise
16, we see that it is permissible to dierentiate under the integral sign to compute
(0) = 2
_
f dx,
which shows that f is weakly harmonic.
11.7. Regularity of Weakly Harmonic Functions
We will now show that the weak solution found in the previous section
is actually a classical C
solution.
381.1. Theorem. If f W
1,2
loc
() is weakly harmonic, then f is continuous in
and
f(x
0
) =
_
B(x
0
,r)
f(y) dy
whenever B(x
0
, r) .
Proof. For convenience take x
0
= 0 and suppose B(0, T) and let 0 <
t < T. Let be an innitely dierentiable function supported in the open interval
(t, T) with the property that
_
T
t
(r) dr = 0.
For each real number r, dene
y(r) =
_
r
t
(s) ds
and dene by
(r) =
_
r
t
y(s)
s
n1
ds.
382 11. FUNCTIONS OF SEVERAL VARIABLES
Finally, let (r) = (r) (T). Note that is innitely dierentiable and assumes
constant values (r) and 0 on [0, t] and [T, ) respectively. Now dene a test
function by
(x) = ([x[),
so that C
c
(). Since f is weakly harmonic, we have
(382.1)
_
f dx =
_
f dx = 0.
A direct calculation shows
(382.2) (x) =
([x[) +
(n 1)
[x[
([x[).
Keeping in mind that we have taken x
0
= 0, we now introduce spherical coordinates
(r, ) so that (r, ) is a point of B(0, r) (see (367.1)). Then, with r = [x[, (382.2)
leads to
(x)r
n1
= (r
n1
(r))
= y
(r)
= (r).
Let
F(r) =
_
||=1
f(r, ) dH
n1
().
Then, bearing in mind that is independent of , (382.1) becomes
0 =
_
T
t
F(r)[r
n1
(r)]
dr =
_
T
t
F(r)(r) dr.
That is, the last integral is 0 whenever C
0
(t, T) with integral 0. By Exercise
3, we conclude that F is constant, say , on (t, T). Since t and T are arbitrary, we
conclude that F is constant on (0, ) where is the distance from x
0
to . Then
n
=
_
0
F(r)r
n1
dr
=
_
0
_
||=1
f(r, )r
n1
d dr
=
_
B(x
0
,)
f(x) dx.
This shows that for each x
0
, the integral average of f over balls of radius
centered at x
0
is constant. Since almost every x
0
is a Lebesgue point for f
(see Theorem 226.1), by letting 0 we have
(382.3) f(x
0
) =
_
B(x
0
,r)
f(y) dy
11.7. REGULARITY OF WEAKLY HARMONIC FUNCTIONS 383
for any ball with B(x
0
, r) . For xed r, the right side of (382.3) is a continuous
function of x
0
. Therefore, it is clear that f can be redened on a set of measure
zero in such a way as to ensure its continuity in .
383.1. Definition. A function : B(x
0
, r) R is called radial relative to x
0
if (x x
0
) = (y x
0
) for all x, y B(x
0
, r) with [x x
0
[ = [y x
0
[.
383.2. Corollary. Suppose f W
1,2
() is weakly harmonic and B(x
0
, r)
. If C
c
(B(x
0
, r)) is nonnegative and radial relative to x
0
with
_
B(x
0
,r)
(x) dx = 1,
then
f(x
0
) =
_
B(x
0
,r)
f(x)(x) dx.
Proof. For convenience, assume x
0
= 0. Since is radial, note that each
superlevel set > t is a ball centered at 0; let r(t) denote its radius. Then, with
M: = sup
B(0,r)
and with the help of Theorem 200.1, we obtain
_
B(0,r)
f(x)(x) dx
=
_
B(0,r)
(f
+
(x) f
(x))(x) dx
=
_
M
0
_
{>t}
(f
+
(x) f
(x)) dx dt
=
_
M
0
_
B(0,r(t))
f(x) dx dt
=
_
M
0
f(0)[B(0, r(t))] dt
= f(0)
_
M
0
[ > t] dt
= f(0).
383.3. Theorem. A weakly harmonic function f W
1,2
() is of class C
().
Proof. For each domain
(x) for
x
. Since f
(x): =
n
(x/) and is as in (337.1) although we also require to be radial
with respect to 0. Finally, we take small enough to ensure that f
is dened
on
. Since
(x) =
_
f(x y)
(y) dy = f(x).
384 11. FUNCTIONS OF SEVERAL VARIABLES
Theorem 379.2 states that if is a bounded open set and W
1,2
() a given
function, then there exists a weakly harmonic function f W
1,2
() such that
f W
1,2
0
(). That is, f assumes the same values as on the boundary of
in the sense of Sobolev theory. The previous result shows that f is a classically
harmonic function in . If is known to be continuous on the boundary of , a
natural question is whether f assumes the values continuously on the boundary.
The answer to this is well understood and is the subject of other areas in analysis.
Exercises for Chapter 11
Section 11.1
384.1 Use Fubinis Theorem to prove
2
f
xy
=
2
f
yx
if f C
2
(R
2
). Use this result to conclude that all second order mixed partials
are equal if f C
2
(R
n
).
384.2 Let C
1
[0, 1] denote the space of functions on [0, 1] that have continuous
derivatives on [0, 1], including one-sided derivatives at the endpoints. De-
ne a norm on C
1
[0, 1] by
|f| : = sup
x[0,1]
[f(x)[ + sup
x[0,1]
f
(x).
Prove that C
1
[0, 1] with this norm is a Banach space.
Section 11.2
384.3 Prove that the sets E(c, R, i) dened by (359.1) and (359.2) are Borel sets.
384.4 Give another proof of Theorem 358.1.
Section 11.3
384.5 Prove that W
1,p
() endowed with the norm (369.2) is a Banach space.
384.6 Prove that the norm
|f| : = |f|
p;
+
_
n
i=1
|D
i
f|
p
p;
_
1/p
is equivalent to (369.2).
384.7 Show that W
1,p
() can be regarded as a closed subspace of the Cartesian
product of L
p
() spaces. With the help of Exercise 22, prove that W
1,p
()
is reexive if 1 < p < .
384.8 Suppose f W
1,p
(R
n
). Prove that f
+
and f
are in W
1,p
(R
n
). Hint: use
Theorem 370.3.
Section 11.4
EXERCISES FOR CHAPTER 11 385
384.9 Relative to the Sobolev norm, prove that C
c
(R
n
) is dense in
C
(R
n
) f : |f|
1,p
< .
385.10 Let f W
1,2
() have the property
_
f dx = 0
for every C
c
(). Prove
_
f dx = 0
for every test function . Hint: Use Theorem 374.1 along with (368.1).
385.11 Let R
n
be an open connected set and suppose f W
1,p
() has the
property that f = 0 almost everywhere in . Prove that f is constant on
.
Section 11.5
385.12 Suppose |f
i
f
j
|
q;
0 and |f
i
f|
p;
0 where R
n
and q > p.
Prove that |f
i
f|
q;
0.
Section 11.6
385.13 Suppose f W
1,2
() is weakly harmonic and let x
. Choose
r dist (
set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
L
p
norm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
G
set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
A
absolute value. . . . . . . . . . . . . . . . . . . . . . . . . . . 15
absolutely continuous . . . . . . . . . . . . . . . . . . 168
admissible function . . . . . . . . . . . . . . . . . . . . 312
Alaoglus Theorem. . . . . . . . . . . . . . . . . . . . . 288
algebra of sets. . . . . . . . . . . . . . . . . . . . . 114, 126
almost everywhere . . . . . . . . . . . . . . . . . . . . . 134
almost uniform convergence . . . . . . . . . . . . 141
approximate limit . . . . . . . . . . . . . . . . . . . . . . 254
approximately continuous . . . . . . . . . . . . . . 254
B
Baire outer measure . . . . . . . . . . . . . . . . . . . 319
Banach space . . . . . . . . . . . . . . . . . . . . . 164, 268
basis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43, 267
bijection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Bolzano-Weierstrass property . . . . . . . . . . . 51
Borel measurable function. . . . . . . . . . . . . . 129
Borel measure . . . . . . . . . . . . . . . . . . . . . . . . . 105
Borel outer measure . . . . . . . . . . . . . . . . . . . 109
Borel regular outer measure . . . . . . . . . . . . 109
Borel Sets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
boundary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
bounded. . . . . . . . . . . . . . . . . . . . . . . . 50, 51, 272
bounded linear functional . . . . . . . . . . . . . . 178
C
Cantor set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Cantor-Lebesgue function . . . . . . . . . . . . . . 133
Caratheodory outer measure . . . . . . . . . . . . 84
Caratheodory-Hahn Extension Thm. . . . 116
cardinal number . . . . . . . . . . . . . . . . . . . . . . . . 23
Cartesian product . . . . . . . . . . . . . . . . . . . . . 3, 5
Cauchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
cauchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Cauchy sequence . . . . . . . . . . . . . . . . . . . . . . . . 46
Cauchy sequence in L
p
. . . . . . . . . . . . . . . . . 163
characteristic function . . . . . . . . . . . . . . . . . 129
Chebyshevs Inequality. . . . . . . . . . . . . . . . . 206
choice mapping . . . . . . . . . . . . . . . . . . . . . . . . . . 5
class of ordinal numbers. . . . . . . . . . . . . . . . . 34
closed. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
closed ball . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Closed Graph theorem. . . . . . . . . . . . . . . . . 280
closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
compact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
comparable, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
complement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
complete . . . . . . . . . . . . . . . . . . . . . . . . . . . 47, 300
measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
complete measure space . . . . . . . . . . . . . . . . 107
composition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
continuous at . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
continuous on . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
continuum hypothesis . . . . . . . . . . . . . . . . . . . 34
contraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
converge
in measure. . . . . . . . . . . . . . . . . . . . . . . . . . . 141
387
388 INDEX
pointwise almost everywhere . . . . . . . . . 139
converge in measure. . . . . . . . . . . . . . . . . . . . 141
converge pointwise almost everywhere . . 139
converge to . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
converge uniformly. . . . . . . . . . . . . . . . . . . . . . 58
converges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
converges in L
p
(X) . . . . . . . . . . . . . . . . . . . . 164
converges weakly. . . . . . . . . . . . . . . . . . . . . . . 284
convex. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
convolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
coordinate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
countable . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
countably additive . . . . . . . . . . . . . . . . . . . . . . 78
countably subadditive . . . . . . . . . . . . . . . . . . . 76
countably-simple function. . . . . . . . . . . . . . 149
D
Daniell intgral . . . . . . . . . . . . . . . . . . . . . . . . . 319
de Morgans laws. . . . . . . . . . . . . . . . . . . . . . . . . 2
dense . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
dense in an open set . . . . . . . . . . . . . . . . . . . . 48
denumerable. . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
derivative of with respect to . . . . . . . 214
diameter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
dierence of sets . . . . . . . . . . . . . . . . . . . . . . . . . 2
dierentiable. . . . . . . . . . . . . . . . . . . . . . . . . . . 350
dierential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Dini derivatives . . . . . . . . . . . . . . . . . . . . . . . . 258
Dirac measure . . . . . . . . . . . . . . . . . . . . . . . . . . 76
discrete metric . . . . . . . . . . . . . . . . . . . . . . . . . . 45
discrete topology. . . . . . . . . . . . . . . . . . . . . . . . 38
disjoint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
distance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
distribution function . . . . . . . . . . . . . . 196, 198
distributional derivative. . . . . . . . . . . . . . . . 362
domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
dual space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
E
Egoros Theorem . . . . . . . . . . . . . . . . . . . . . 140
empty set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
enlargement of a ball . . . . . . . . . . . . . . . . . . . 214
equicontinuous . . . . . . . . . . . . . . . . . . . . . . . . . . 60
equicontinuous at . . . . . . . . . . . . . . . . . . . . . . . 71
equivalence class . . . . . . . . . . . . . . . . . . . . . . . . . 5
equivalence class of Cauchy sequences . . . 17
equivalence relation . . . . . . . . . . . . . . . . . . . . . . 4
equivalence relation. . . . . . . . . . . . . . . . . . . . . 17
equivalent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
essential variation. . . . . . . . . . . . . . . . . . . . . . 341
Euclidean n-space . . . . . . . . . . . . . . . . . . . . . . . . 6
extended real numbers . . . . . . . . . . . . . . . . . 127
F
eld. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .15
nite. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
nite additivity . . . . . . . . . . . . . . . . . . . . . . . . 105
nite dimensional . . . . . . . . . . . . . . . . . . . . . . 267
nite intersection property . . . . . . . . . . . . . . 42
nite measure. . . . . . . . . . . . . . . . . . . . . . . . . . 105
nite set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
nitely additive measure . . . . . . . . . . . . . . . 105
rst axiom of countability. . . . . . . . . . . . . . . 43
rst category. . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
xed point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
fractals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
function. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Borel measurable . . . . . . . . . . . . . . . . . . . . 129
characteristic . . . . . . . . . . . . . . . . . . . . . . . 129
countably-simple . . . . . . . . . . . . . . . . . . . . 149
integrable . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
integral exists . . . . . . . . . . . . . . . . . . . . . . . 150
simple . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
fundamental in measure. . . . . . . . . . . . . . . . 147
fundamental sequence . . . . . . . . . . . . . . . . . . . 46
G
general Cantor set . . . . . . . . . . . . . . . . . . . . . 102
generalized Cantor set. . . . . . . . . . . . . . . . . . .93
graph. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
graph of a function. . . . . . . . . . . . . . . . . . . . . . 71
H
Holders inequality. . . . . . . . . . . . . . . . . . . . . 161
harmonic function . . . . . . . . . . . . . . . . . . . . . 373
Hausdor dimension . . . . . . . . . . . . . . . . . . . 101
Hausdor measure . . . . . . . . . . . . . . . . . . . . . . 98
Hausdor space . . . . . . . . . . . . . . . . . . . . . . . . . 40
Hilbert Cube. . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Hilbert space . . . . . . . . . . . . . . . . . . . . . . . . . . 296
homeomorphic . . . . . . . . . . . . . . . . . . . . . . . . . . 47
INDEX 389
homeomorphism. . . . . . . . . . . . . . . . . . . . . . . . . 47
I
image . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
immediate successor, . . . . . . . . . . . . . . . . . . . . 14
indiscrete topology. . . . . . . . . . . . . . . . . . . . . . 38
induced topology. . . . . . . . . . . . . . . . . . . . . . . . 38
inmum. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
innite. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
initial ordinal . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
initial segment . . . . . . . . . . . . . . . . . . . . . . . . . . 30
injection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
inner product . . . . . . . . . . . . . . . . . . . . . . . . . . 294
integrable function. . . . . . . . . . . . . . . . . . . . . 150
integral average . . . . . . . . . . . . . . . . . . . . . . . . 218
integral exists . . . . . . . . . . . . . . . . . . . . . . . . . . 150
integral of a simple function. . . . . . . . . . . . 149
interior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
intersection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
inverse image . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
is harmonic in the sense of distributions373
isolated point . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
isometric. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
isometric isomorphism . . . . . . . . . . . . . . . . . 282
isometrically isomorphic . . . . . . . . . . . . . . . 282
isometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
isometry, . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
isomorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
J
Jacobian. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
Jensens inequality . . . . . . . . . . . . . . . . . . . . . 207
L
lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
least upper bound. . . . . . . . . . . . . . . . . . . . . . . 23
least upper bound for . . . . . . . . . . . . . . . . . . . 21
Lebesgue conjugate . . . . . . . . . . . . . . . . . . . . 161
Lebesgue integrable . . . . . . . . . . . . . . . . . . . . 157
Lebesgue measurable function. . . . . . . . . . 128
Lebesgue measure . . . . . . . . . . . . . . . . . . . . . . . 89
Lebesgue number . . . . . . . . . . . . . . . . . . . . . . . 70
Lebesgue outer measure . . . . . . . . . . . . . . . . . 87
Lebesgue-Stieltjes . . . . . . . . . . . . . . . . . . . . . . . 94
less than . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
lim inf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
lim sup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
limit point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
linear . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
linear functional . . . . . . . . . . . . . . . . . . . . . . . 178
linear functionals . . . . . . . . . . . . . . . . . . . . . . 272
linear isomorphism. . . . . . . . . . . . . . . . . . . . . 282
linearly isomorphic. . . . . . . . . . . . . . . . . . . . . 282
Lipschitz. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Lipschitz constant. . . . . . . . . . . . . . . . . . . . . . . 65
locally compact . . . . . . . . . . . . . . . . . . . . . . . . . 40
lower bound, greatest lower bound, . . . . . . 21
lower envelope . . . . . . . . . . . . . . . . . . . . . . . . . 138
lower integral . . . . . . . . . . . . . . . . . . . . . . . . . . 149
lower limit of a function. . . . . . . . . . . . . . . . . 64
lower semicontinuous . . . . . . . . . . . . . . . . 64, 67
lower semicontinuous function. . . . . . . . . . . 64
Lusins Theorem. . . . . . . . . . . . . . . . . . . . . . . 145
M
Marcinkiewicz Interpolation Theorem . . 200
maximal function . . . . . . . . . . . . . . . . . . . . . . 218
measurable mapping . . . . . . . . . . . . . . . . . . . 128
measurable sets . . . . . . . . . . . . . . . . . . . . . . . . 105
measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Borel outer . . . . . . . . . . . . . . . . . . . . . . . . . . 109
nite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
mutually singular . . . . . . . . . . . . . . . . . . . . 170
outer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
regular outer . . . . . . . . . . . . . . . . . . . . . . . . 109
measure on an algebra, A. . . . . . . . . 114, 126
measure space . . . . . . . . . . . . . . . . . . . . . . . . . 105
metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
modulus of continuity of a function. . . . . . 58
mollication . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
mollier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
monotone . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
monotone Daniell integral . . . . . . . . . . . . . . 319
monotone functional . . . . . . . . . . . . . . . . . . . 317
mutually singular measures . . . . . . . . . . . . 170
N
N. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
natural imbedding . . . . . . . . . . . . . . . . . . . . . 283
negative set . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
neighborhood . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
390 INDEX
non-increasing rearrangement . . . . . . . . . . 199
norm. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
normed linear spaces . . . . . . . . . . . . . . . . . . . 163
nowhere dense . . . . . . . . . . . . . . . . . . . . . . . . . . 48
null set . . . . . . . . . . . . . . . . . . . . . . . . . 1, 134, 168
O
onto Y . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
open ball . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
open cover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Open mapping theorem. . . . . . . . . . . . . . . . 278
open sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
ordinal number . . . . . . . . . . . . . . . . . . . . . . . . . 32
orthogonal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
orthonormal family . . . . . . . . . . . . . . . . . . . . 300
outer measure. . . . . . . . . . . . . . . . . . . . . . . . . . . 76
P
partial ordering . . . . . . . . . . . . . . . . . . . . . . . . . . 7
positive functional . . . . . . . . . . . . . . . . . . . . . 317
positive measure . . . . . . . . . . . . . . . . . . . . . . . 168
positive set . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
power set of X . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
principle of mathematical induction, . . . . 14
Principle of Transnite Induction. . . . . . . . 30
product topology. . . . . . . . . . . . . . . . . . . . . . . . 44
proper subset . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
R
radial function . . . . . . . . . . . . . . . . . . . . . . . . . 377
radon measure . . . . . . . . . . . . . . . . . . . . . . . . . 198
Radon-Nikodym derivative . . . . . . . . . . . . . 175
Radon-Nikodym Theorem. . . . . . . . . . . . . . 172
range . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
real number . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
reexive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
regularization . . . . . . . . . . . . . . . . . . . . . . . . . . 212
regularizer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
relative topology . . . . . . . . . . . . . . . . . . . . . . . . 38
restriction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Riemann integrable . . . . . . . . . . . . . . . . . . . . 156
Riemann partition . . . . . . . . . . . . . . . . . . . . . 156
Riemann-Stieltjes integral . . . . . . . . . . . . . . 205
S
Schwarz Inequality. . . . . . . . . . . . . . . . . . . . . 296
second axiom of countability . . . . . . . . . . . . 43
second category . . . . . . . . . . . . . . . . . . . . . . . . . 48
self-similar set . . . . . . . . . . . . . . . . . . . . . . . . . 104
separable. . . . . . . . . . . . . . . . . . . . . . . . . . . 69, 281
separated . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
sequentially compact . . . . . . . . . . . . . . . . . . . . 51
signed measure . . . . . . . . . . . . . . . . . . . . . . . . 168
simple function . . . . . . . . . . . . . . . . . . . . . . . . 143
single-valued . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
singleton . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
smooth partition of unity . . . . . . . . . . . . . . 367
Sobolev inequality . . . . . . . . . . . . . . . . . . . . . 371
Sobolev norm. . . . . . . . . . . . . . . . . . . . . . . . . . 363
Sobolev Space . . . . . . . . . . . . . . . . . . . . . . . . . 362
strong topology . . . . . . . . . . . . . . . . . . . . . . . . 284
strong-type (p, q) . . . . . . . . . . . . . . . . . . . . . . 200
subbase . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
subcover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
subsequence. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
subset. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
subspace induced . . . . . . . . . . . . . . . . . . . . . . . 45
superlevel sets of f . . . . . . . . . . . . . . . . . . . . . 129
supremum of . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
surjection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
symmetric dierence . . . . . . . . . . . . . . . . . . . . . 2
T
topological invariant. . . . . . . . . . . . . . . . . . . . . 47
topological space . . . . . . . . . . . . . . . . . . . . . . . . 37
topologically equivalent . . . . . . . . . . . . . . . . . 46
topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
topology induced. . . . . . . . . . . . . . . . . . . . . . . . 46
total . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
totally bounded . . . . . . . . . . . . . . . . . . . . . . . . . 51
U
unconditionally . . . . . . . . . . . . . . . . . . . . . . . . 301
uncountable. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
uniform boundedness principle . . . . . . . . . . 50
uniform convergence . . . . . . . . . . . . . . . . . . . . 58
uniformly absolutely continuous . . . . . . . . 209
uniformly bounded. . . . . . . . . . . . . . . . . . . . . . 50
INDEX 391
uniformly continuous on X . . . . . . . . . . . . . . 57
union . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
univalent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
upper bound for . . . . . . . . . . . . . . . . . . . . . . . . 21
upper envelope. . . . . . . . . . . . . . . . . . . . . . . . . 138
upper integral . . . . . . . . . . . . . . . . . . . . . . . . . . 149
upper limit of a function . . . . . . . . . . . . . . . . 64
upper semicontinuous . . . . . . . . . . . . . . . . . . . 67
upper semicontinuous function . . . . . . . . . . 64
V
Vitali covering . . . . . . . . . . . . . . . . . . . . . . . . . 216
W
weak topology . . . . . . . . . . . . . . . . . . . . . . . . . 284
weak
topology . . . . . . . . . . . . . . . . . . . . . . . . 287
weak-type (p, q) . . . . . . . . . . . . . . . . . . . . . . . . 200
weakly harmonic . . . . . . . . . . . . . . . . . . . . . . . 373
well-ordered. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
well-ordered set . . . . . . . . . . . . . . . . . . . . . . . . . 13
CHAPTER 12
Solutions to Selected Problems
Chapter 6
Solution to problem 6.32 Prove Jensens inequality: Let f L
1
(X, /, )
where (X) < and suppose f(X) [a, b]. If is a convex function on [a, b],
then
_
1
(X)
_
X
f d
_
1
(X)
_
X
( f) d.
Thus, (average(f)) average( f). Hint: let
t
0
= [(X)
1
]
_
X
f d. Then t
0
(a, b).
Furthermore, with
:= sup
t(a,t
0
)
(t
0
) (t)
t
0
t
,
we have (t) (t
0
) (t t
0
) for all t (a, b). In particular, (f(x)) (t
0
)
(f(x) t
0
) for all x X. Now integrate.
Solution:
_
X
(f(x)) (t
0
) d(x) =
_
X
( f) d
_
1
(X)
_
X
f d
_
(X)
_
X
(f(x) t
0
) d(x)
=
__
X
f(x) d(x) t
0
(X)
_
= [t
0
(X) t
0
(X)] = 0
= Our desired inequality now follows.
393.1. Remark. If (t) = t
p
, 1 p < , then Jensens inequqality follows
from Holders inequality:
[(X)]
1
_
X
f 1 d |f|
p
[(X)]
1
p
1
= |f|
p
[(X)]
1/p
=
_
1
(X)
_
X
f d
_
p
1
(X)
_
X
([f[
p
) d.
393
394 12. SOLUTIONS TO SELECTED PROBLEMS
However, Jensens inequality is stronger than Holders inequality in the follow-
ing sense:
394.0. Corollary (of Jensens inequality). If f is dened on [0, 1]then
e
X
f d
_
X
e
f(x)
d.
Soltion to 8.25 Suppose that (X, /, ) is a nite measure space and that f is
a given measure function. Assume that:
_
X
fgd <
for each g L
p
(X), 1 p . Prove that |f|
p
< .
Proof:
(1) First dene F(f) :=
_
X
fgd. Obviously it is a real-valued linear function
on L
p
(X). We may assume that f here is nonnegative because [[f[g[ = [fg[ and
fg is integrable (by assumption). First we need to nd an appropriate domain or
metric space to accommodate our following reasoning.
Claim: L
p
+
:= g L
p
.
Proof of the Claim: Let g
i
be a sequence of nonnegative L
p
-functions and
|g
i
g| 0 when i for some g L
p
(recall that L
p
is complete). Now we
need to show that g L
p
+
. By appealing to Vatalis Convergence Theorem, we get
that there is a subsequence g
ij
g a.e. when j . It implies that g L
p
+
.
Now we re-prove Theorem 3.35 on L
p
+
.
Claim There is a nonempty open set U L
p
+
and a constant M such that
[F(g)[ M for all g U.
Proof of the claim: For each positive i dene E
i
:= g L
p
+
[F(g) i. First
we show that this set is closed. Let g
i
s be a sequence in E
i
and |g
i
g| 0
when i for some g L
p
+
. By applying Vitali Theorem, we get that there is a
subsequence g
ij
j
g when j . That is to say liminf
j
g
ij
= g. According
to Fatous Theorem,
_
X
fgd liminf
j
_
X
fg
ij
d i. It follows from the
assumption that L
p
+
=
E
i
. Since L
p
+
is a closed subset of a complete metric space,
it is also complete w.r.t. the induced metric space. By Baire Category Theorem,
there is E
M
which is not nowhere dense. It must contains an open set. Therefore we
have shown the claim. Here we may take U to be a ball B =: g L
p
+
: |gg
0
|
for some g
0
L
p
+
.
Claim: For any g L
p
, if |g g
0
| , then [F(g)[ M.
Proof of the claim: Assume that g L
p
and |g g
0
| . Then it follows that
|[g[ g
0
| because [[g[ g
0
[ [g g
0
[. Since [g[ L
p
+
, F([g[) M. Besides,
[F(g)[ F([g[) M.
12. SOLUTIONS TO SELECTED PROBLEMS 395
That is to say, F is uniformly bounded by M on the open set B
=: g
L
p
[|g g
0
|
p
for some g
0
L
p
+
. Note that B
B.
Claim: |F| <
Proof of the claim: Now consider the closed ball B(0, ) where 0 is the zero
function in *L
p
L
p
(X), if |g
|
p
1, then [F(g
)[ = [ F(1/ g
)[ (M +[F(g
0
)[). Finally
we get that F is a bounded linear functional. According to Theorem 6.46, there is
a f
L
p
such that
F(g) =
_
X
f
gd for each g L
p
.
Let f and f
gd for each
g L
p
, then g = g
a.e.
Proof of the claim: Take f
0
:= sign(f f
). Obviously f
0
is measurable. Since
[f
0
[ 1 and the measure space is nite, f
0
L
p
. Note that f
0
(f f
) = [f f
[.
By the assumption, we have that
_
X
(f f
. This implies
that
_
X
[f f
[d = 0. It follows that f = f
-a.e.
By appealing to Theorem 6.45, we get that |f|
p
= |f
|
p
= |F| < .
Alternate proof that |F| = |f|
p
:
From the previous step we know that the linear functional F : L
p
(X) R
1
dened by
F(g) =
_
X
fg d
is bounded. Now we can appeal to Theorem 6.24 directly to conclude that |F| =
|f|
p
.