Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
students:
measure and integration
theory
Richard F. Bass
ii
c _ Copyright 2011 Richard F. Bass
All rights reserved.
ISBN13: 9781466391574
ISBN10: 146639157X
iii
To the memory of my parents
Jay and Betty Bass
iv
Contents
Preface ix
1 Preliminaries 1
1.1 Notation and terminology . . . . . . . . . . . . . . .1
1.2 Some undergraduate mathematics . . . . . . . . . .2
2 Families of sets 7
2.1 Algebras and σalgebras . . . . . . . . . . . . . . . .7
2.2 The monotone class theorem . . . . . . . . . . . . .10
2.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . .11
3 Measures 13
3.1 Deﬁnitions and examples . . . . . . . . . . . . . . . .13
3.2 Exercises . . . . . . . . . . . . . . . . . . . . . . . .16
4 Construction of measures 19
4.1 Outer measures . . . . . . . . . . . . . . . . . . . . .20
4.2 LebesgueStieltjes measures . . . . . . . . . . . . . .24
4.3 Examples and related results . . . . . . . . . . . . .27
4.4 Nonmeasurable sets . . . . . . . . . . . . . . . . . .30
4.5 The Carath´eodory extension theorem . . . . . . . . .31
4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . .34
v
vi CONTENTS
5 Measurable functions 37
5.1 Measurability . . . . . . . . . . . . . . . . . . . . . .37
5.2 Approximation of functions . . . . . . . . . . . . . .41
5.3 Lusin’s theorem . . . . . . . . . . . . . . . . . . . . .41
5.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . .44
6 The Lebesgue integral 47
6.1 Deﬁnitions . . . . . . . . . . . . . . . . . . . . . . . .47
6.2 Exercises . . . . . . . . . . . . . . . . . . . . . . . .49
7 Limit theorems 51
7.1 Monotone convergence theorem . . . . . . . . . . . .51
7.2 Linearity of the integral . . . . . . . . . . . . . . . .52
7.3 Fatou’s lemma . . . . . . . . . . . . . . . . . . . . .54
7.4 Dominated convergence theorem . . . . . . . . . . .55
7.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . .56
8 Properties of Lebesgue integrals 63
8.1 Criteria for a function to be zero a.e. . . . . . . . . .63
8.2 An approximation result . . . . . . . . . . . . . . . .65
8.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . .66
9 Riemann integrals 69
9.1 Comparison with the Lebesgue integral . . . . . . . .69
9.2 Exercises . . . . . . . . . . . . . . . . . . . . . . . .71
10 Types of convergence 75
10.1 Deﬁnitions and examples . . . . . . . . . . . . . . . .75
10.2 Exercises . . . . . . . . . . . . . . . . . . . . . . . .78
11 Product measures 81
CONTENTS vii
11.1 Product σalgebras . . . . . . . . . . . . . . . . . . .81
11.2 The Fubini theorem . . . . . . . . . . . . . . . . . .85
11.3 Examples . . . . . . . . . . . . . . . . . . . . . . . .87
11.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . .88
12 Signed measures 93
12.1 Positive and negative sets . . . . . . . . . . . . . . .93
12.2 Hahn decomposition theorem . . . . . . . . . . . . .95
12.3 Jordan decomposition theorem . . . . . . . . . . . .97
12.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . .97
13 The RadonNikodym theorem 99
13.1 Absolute continuity . . . . . . . . . . . . . . . . . . .99
13.2 The main theorem . . . . . . . . . . . . . . . . . . .100
13.3 Lebesgue decomposition theorem . . . . . . . . . . .103
13.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . .104
14 Diﬀerentiation 107
14.1 Maximal functions . . . . . . . . . . . . . . . . . . .108
14.2 Antiderivatives . . . . . . . . . . . . . . . . . . . . .112
14.3 Bounded variation . . . . . . . . . . . . . . . . . . .113
14.4 Absolutely continuous functions . . . . . . . . . . . .118
14.5 Approach 2 – diﬀerentiability . . . . . . . . . . . . .120
14.6 Approach 2 – antiderivatives . . . . . . . . . . . . .124
14.7 Approach 2 – absolute continuity . . . . . . . . . . .126
14.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . .127
15 L
p
spaces 131
15.1 Norms . . . . . . . . . . . . . . . . . . . . . . . . . .131
15.2 Completeness . . . . . . . . . . . . . . . . . . . . . .134
viii CONTENTS
15.3 Convolutions . . . . . . . . . . . . . . . . . . . . . .136
15.4 Bounded linear functionals . . . . . . . . . . . . . . .137
15.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . .141
16 Fourier transforms 147
16.1 Basic properties . . . . . . . . . . . . . . . . . . . . .147
16.2 The inversion theorem . . . . . . . . . . . . . . . . .150
16.3 The Plancherel theorem . . . . . . . . . . . . . . . .154
16.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . .155
17 Riesz representation 157
17.1 Partitions of unity . . . . . . . . . . . . . . . . . . .158
17.2 The representation theorem . . . . . . . . . . . . . .159
17.3 Regularity . . . . . . . . . . . . . . . . . . . . . . . .163
17.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . .164
18 Banach spaces 167
18.1 Deﬁnitions . . . . . . . . . . . . . . . . . . . . . . . .167
18.2 The HahnBanach theorem . . . . . . . . . . . . . .168
18.3 Baire’s theorem and consequences . . . . . . . . . .171
18.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . .174
19 Hilbert spaces 177
19.1 Inner products . . . . . . . . . . . . . . . . . . . . .177
19.2 Subspaces . . . . . . . . . . . . . . . . . . . . . . . .180
19.3 Orthonormal sets . . . . . . . . . . . . . . . . . . . .182
19.4 Fourier series . . . . . . . . . . . . . . . . . . . . . .185
19.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . .187
Preface
Nearly every Ph.D. student in mathematics needs to pass a pre
liminary or qualifying examination in real analysis. The purpose
of this book is to teach the material necessary to pass such an
examination.
I had three main goals in writing this text:
(1) present a very clear exposition;
(2) provide a large collection of useful exercises;
(3) make the text aﬀordable.
Let me discuss each of these in more detail.
(1) There are a large number of real analysis texts already in
existence. Why write another? In my opinion, none of the exist
ing texts are ideally suited to the beginning graduate student who
needs to pass a “prelim” or “qual.” They are either too hard, too
advanced, too encyclopedic, omit too many important topics, or
take a nonstandard approach to some of the basic theorems.
Students who are starting their graduate mathematics educa
tion are often still developing their mathematical sophistication and
ﬁnd that the more details that are provided, the better (within rea
son). I have tried to make the writing as clear as possible and to
provide the details. For the sake of clarity, I present the theorems
and results as they will be tested, not in the absolutely most general
abstract context. On the other hand, a look at the index will show
that no topics that might appear on a preliminary or qualifying
examination are omitted.
All the proofs are “plain vanilla.” I avoid any clever tricks,
sneaky proofs, unusual approaches, and the like. These are the
proofs and methods that most experts grew up on.
ix
x PREFACE
(2) There are over 220 exercises. I tried to make them inter
esting and useful and to avoid problems that are overly technical.
Many are routine, many are of moderate diﬃculty, but some are
quite challenging. A substantial number are taken from prelimi
nary examinations given in the past at universities with which I
have been associated.
I thought long and hard as to whether to provide hints to the
exercises. When I teach the real analysis course, I give hints to the
harder questions. But some instructors will want a more challeng
ing course than I give and some a less challenging one. I leave it
to the individual instructor to decide how many hints to give.
(3) I have on my bookshelf several books that I bought in the
early 1970’s that have the prices stamped in them: $10$12. These
very same books now sell for $100$200. The cost of living has
gone up in the last 40 years, but only by a factor of 5 or 6, not
a factor of 10. Why do publishers make textbooks so expensive?
This is particularly troublesome when one considers that nowadays
authors do their own typesetting and frequently their own page
layout.
My aim was to make the soft cover version of this text cost less
than $20 and to provide a version in .pdf format for free. To do
that, I am selfpublishing the text.
At this point I should tell you a little bit about the subject
matter of real analysis. For an interval contained in the real line or
a nice region in the plane, the length of the interval or the area of
the region give an idea of the size. We want to extend the notion of
size to as large a class of sets as possible. Doing this for subsets of
the real line gives rise to Lebesgue measure. Chapters 2–4 discuss
classes of sets, the deﬁnition of measures, and the construction of
measures, of which one example is Lebesgue measure on the line.
(Chapter 1 is a summary of the notation that is used and the
background material that is required.)
Once we have measures, we proceed to the Lebesgue integral.
We talk about measurable functions, deﬁne the Lebesgue integral,
prove the monotone and dominated convergence theorems, look at
some simple properties of the Lebesgue integral, compare it to the
Riemann integral, and discuss some of the various ways a sequence
of functions can converge. This material is the subject of Chapters
5–10.
xi
Closely tied with measures and integration are the subjects of
product measures, signed measures, the RadonNikodym theorem,
the diﬀerentiation of functions on the line, and L
p
spaces. These
are covered in Chapters 11–15.
Many courses in real analysis stop at this point. Others also in
clude some or all of the following topics: the Fourier transform, the
Riesz representation theorem, Banach spaces, and Hilbert spaces.
We present these in Chapters 16–19.
The prerequisites to this text are a solid background in un
dergraduate mathematics. An acquaintance with metric spaces is
assumed, but no other topology. A summary of what you need to
know is in Chapter 1. All the necessary background material can
be learned from many sources; one good place is the book [3].
At some universities preliminary or qualifying examinations in
real analysis are combined with those in undergraduate analysis or
topology or complex analysis. If that is the case at your university,
you will have to supplement this text with texts in those subjects.
Further reading is always useful. I have found the books [1], [2],
and [4] helpful.
I would like to thank A. Baldenko, I. BenAri, K. Bharath, K.
Burdzy, D. Ferrone, E. Gin´e, M. Gordina, E. Hsu, G. Lawler, L.
Lu, K. Marinelli, J. Pitman, M. Poehlitz, H. Ren, L. Rogers, and
A. Teplyaev for some very useful suggestions.
If you have comments, suggestions, corrections, etc., I would be
glad to hear from you: r.bass@uconn.edu. I cannot, however,
provide hints or solutions to the exercises.
Good luck with your exam!
xii PREFACE
Chapter 1
Preliminaries
In this short chapter we summarize some of the notation and ter
minology we will use and recall a few deﬁnitions and results from
undergraduate mathematics.
1.1 Notation and terminology
We use A
c
, read “A complement,” for the set of points not in A.
To avoid some of the paradoxes of set theory, we assume all our
sets are subsets of some given set X, and to be precise, deﬁne
A
c
= ¦x ∈ X : x / ∈ A¦.
We write
A−B = A∩ B
c
(it is common to also see A¸ B) and
A∆B = (A−B) ∪ (B −A).
The set A∆B is called the symmetric diﬀerence of A and B and
is the set of points that are in one of the sets but not the other.
If I is some index set, a collection of subsets ¦A
α
¦
α∈I
is disjoint if
A
α
∩ A
β
= ∅ whenever α ,= β.
We write A
i
↑ if A
1
⊂ A
2
⊂ and write A
i
↑ A if in addition
A = ∪
∞
i=1
A
i
. Similarly A
i
↓ means A
1
⊃ A
2
⊃ and A
i
↓ A
means that in addition A = ∩
∞
i=1
A
i
.
1
2 CHAPTER 1. PRELIMINARIES
We use Q to denote the set of rational numbers, R the set of
real numbers, and C the set of complex numbers. We use
x ∨ y = max(x, y) and x ∧ y = min(x, y).
We can write a real number x in terms of its positive and negative
parts: x = x
+
−x
−
, where
x
+
= x ∨ 0 and x
−
= (−x) ∨ 0.
If z is a complex number, then z is the complex conjugate of z.
If f is a function whose domain is the reals or a subset of the
reals, then lim
y→x+
f(y) and lim
y→x−
f(y) are the right and left
hand limits of f at x, resp.
We say a function f : R → R is increasing if x < y im
plies f(x) ≤ f(y) and f is strictly increasing if x < y implies
f(x) < f(y). (Some authors use “nondecreasing” for the former
and “increasing” for the latter.) We deﬁne decreasing and strictly
decreasing similarly. A function is monotone if f is either increas
ing or decreasing.
Given a sequence ¦a
n
¦ of real numbers,
limsup
n→∞
a
n
= inf
n
sup
m≥n
a
m
,
liminf
n→∞
a
n
= sup
n
inf
m≥n
a
m
.
For example, if
a
n
=
_
1, n even;
−1/n, n odd,
then limsup
n→∞
a
n
= 1 and liminf
n→∞
a
n
= 0. The sequence
¦a
n
¦ has a limit if and only if limsup
n→∞
a
n
= liminf
n→∞
a
n
and
both are ﬁnite. We use analogous deﬁnitions when we take a limit
along the real numbers. For example,
limsup
y→x
f(y) = inf
δ>0
sup
y−x<δ
f(y).
1.2 Some undergraduate mathematics
We recall some deﬁnitions and facts from undergraduate topology,
algebra, and analysis. The proofs and more details can be found
in many places. A good source is [3].
1.2. SOME UNDERGRADUATE MATHEMATICS 3
A set X is a metric space if there exists a metric d : XX →R
such that
(1) d(x, y) = d(y, x) for all x, y ∈ X;
(2) d(x, y) ≥ 0 for all x, y ∈ X and d(x, y) = 0 if and only if x = y;
(3) d(x, z) ≤ d(x, y) +d(y, z) for all x, y, z ∈ X.
Condition (3) is called the triangle inequality.
Given a metric space X, let
B(x, r) = ¦y ∈ X : d(x, y) < r¦
be the open ball of radius r centered at x. If A ⊂ X, the interior
of A, denoted A
o
, is the set of x such that there exists r
x
> 0 with
B(x, r
x
) ⊂ A. The closure of A, denoted A, is the set of x ∈ X
such that every open ball centered at x contains at least one point
of A. A set A is open if A = A
o
, closed if A = A. If f : X → R,
the support of f is the closure of the set ¦x : f(x) ,= 0¦. f is
continuous at a point x if given ε > 0, there exists δ > 0 such
that [f(x) − f(y)[ < ε whenever d(x, y) < δ. f is continuous if
it is continuous at every point of its domain. One property of
continuous functions is that f
−1
(F) is closed and f
−1
(G) is open
if f is continuous, F is closed, and G is open.
A sequence ¦x
n
¦ ⊂ X converges to a point x ∈ X if for each
ε > 0 there exists N such that d(x
n
, x) < x whenever n ≥ N. A
sequence is a Cauchy sequence if for each ε > 0 there exists N such
that d(x
m
, x
n
) < ε whenever m, n ≥ N. If every Cauchy sequence
in X converges to a point in X, we say X is complete.
An open cover of a subset K of X is a collection ¦G
α
¦
α∈I
of
open sets such that K ⊂ ∪
α∈I
G
α
. The index set I can be ﬁnite
or inﬁnite. A set K is compact of every open cover contains a
ﬁnite subcover, i.e., there exists G
1
, . . . , G
n
∈ ¦G
α
¦
α∈I
such that
K ⊂ ∪
n
i=1
G
i
.
We have the following two facts about compact sets.
Proposition 1.1 If K is compact, F ⊂ K, and F is closed, then
F is compact.
Proposition 1.2 If K is compact and f is continuous on K, then
there exist x
1
and x
2
such that f(x
1
) = inf
x∈K
f(x) and f(x
2
) =
sup
x∈K
f(x). In other words, f takes on its maximum and mini
mum values.
4 CHAPTER 1. PRELIMINARIES
Remark 1.3 If x ,= y, let r = d(x, y) and note that d(x, r/2) and
d(y, r/2) are disjoint open sets containing x and y, resp. Therefore
metric spaces are also what are called Hausdorﬀ spaces.
Let F be either R or C. X is a vector space or linear space if
there exist two operations, addition (+) and scalar multiplication,
such that
(1) x +y = y +x for all x, y ∈ X;
(2) (x +y) +z = x + (y +z) for all x, y, z ∈ X;
(3) there exists an element 0 ∈ X such that 0 + x = x for all
x ∈ X;
(4) for each x in X there exists an element −x ∈ X such that
x + (−x) = 0;
(5) c(x +y) = cx +cy for all x, y ∈ X, c ∈ F;
(6) (c +d)x = cx +dx for all x ∈ X, c, d ∈ F;
(7) c(dx) = (cd)x for all x ∈ X, c, d ∈ F;
(8) 1x = x for all x ∈ X.
We use the usual notation, e.g., x −y = x + (−y).
X is a normed linear space if there exists a map x → x such
that
(1) x ≥ 0 for all x ∈ X and x = 0 if and only if x = 0;
(2) cx = [c[ x for all c ∈ F and x ∈ X;
(3) x +y ≤ x +y for all x, y ∈ X.
Given a normed linear space X, we can make X into a metric
space by setting d(x, y) = x −y.
A set X has an equivalence relationship “∼” if
(1) x ∼ x for all x ∈ X;
(2) if x ∼ y, then y ∼ x;
(3) if x ∼ y and y ∼ z, then x ∼ z.
Given an equivalence relationship, X can be written as the
union of disjoint equivalence classes. x and y are in the same
equivalence class if and only if x ∼ y. For an example, let X = R
and say x ∼ y if x −y is a rational number.
A set X has a partial order “≤” if
(1) x ≤ x for all x ∈ X;
(2) if x ≤ y and y ≤ x, then x = y.
Note that given x, y ∈ X, it is not necessarily true that x ≤ y
or y ≤ x. For an example, let Y be a set, let X be the collection
of all subsets of Y , and say A ≤ B if A, B ∈ X and A ⊂ B.
1.2. SOME UNDERGRADUATE MATHEMATICS 5
We need the following three facts about the real line.
Proposition 1.4 Suppose K ⊂ R, K is closed, and K is contained
in a ﬁnite interval. Then K is compact.
Proposition 1.5 Suppose G ⊂ R is open. Then G can be written
as the countable union of disjoint open intervals.
Proposition 1.6 Suppose f : R → R is an increasing function.
Then both lim
y→x+
f(y) and lim
y→x−
f(y) exist for every x. More
over the set of x where f is not continuous is countable.
For an application of Hilbert space techniques to Fourier series,
which is the last section of Chapter 19, we will use the Stone
Weierstrass theorem. The particular version we will use is the
following.
Theorem 1.7 Let X be a compact metric space and let / be a
collection of continuous complexvalued functions on X with the
following properties:
(1) If f, g ∈ / and c ∈ C, then f +g, fg, and cf are in /;
(2) If f ∈ /, then f ∈ /, where f is the complex conjugate of f;
(3) If x ∈ X, there exists f ∈ / such that f(x) ,= 0;
(4) If x, y ∈ X with x ,= y, there exists f ∈ / such that f(x) ,=
f(y). Then the closure of / with respect to the supremum norm is
the collection of continuous complexvalued functions on X.
The conclusion can be rephrased as saying that given f contin
uous on X and ε > 0, there exists g ∈ / such that
sup
x∈X
[f(x) −g(x)[ < ε.
When (3) holds, / is said to vanish at no point of X. When (4)
holds, / is said to separate points. In (3) and (4), the function f
depends on x and on x and y, resp.
For a proof of Theorem 1.7 see [1] or [3].
6 CHAPTER 1. PRELIMINARIES
Chapter 2
Families of sets
2.1 Algebras and σalgebras
When we turn to constructing measures in Chapter 4, we will see
that we cannot in general deﬁne the measure of an arbitrary set.
We will have to restrict the class of sets we consider. The class of
sets that we will want to use are σalgebras (read “sigma algebras”).
Let X be a set.
Deﬁnition 2.1 An algebra is a collection / of subsets of X such
that
(1) ∅ ∈ / and X ∈ /;
(2) if A ∈ /, then A
c
∈ /;
(3) if A
1
, . . . , A
n
∈ /, then ∪
n
i=1
A
i
and ∩
n
i=1
A
i
are in /.
/ is a σalgebra if in addition
(4) whenever A
1
, A
2
, . . . are in /, then ∪
∞
i=1
A
i
and ∩
∞
i=1
A
i
are in
/.
In (4) we allow countable unions and intersections only; we do
not allow uncountable unions and intersections. Since ∩
∞
i=1
A
i
=
(∪
∞
i=1
A
c
i
)
c
, the requirement that ∩
∞
i=1
A
i
be in / is redundant.
The pair (X, /) is called a measurable space. A set A is measurable
if A ∈ /.
7
8 CHAPTER 2. FAMILIES OF SETS
Example 2.2 Let X = R, the set of real numbers, and let / be
the collection of all subsets of R. Then / is a σalgebra.
Example 2.3 Let X = R and let
/ = ¦A ⊂ R : A is countable or A
c
is countable¦.
Verifying parts (1) and (2) of the deﬁnition is easy. Suppose
A
1
, A
2
, . . . are each in /. If each of the A
i
are countable, then
∪
i
A
i
is countable, and so is in /. If A
c
i
0
is countable for some i
0
,
then
(∪A
i
)
c
= ∩
i
A
c
i
⊂ A
c
i
0
is countable, and again ∪
i
A
i
is in /. Since ∩A
i
= (∪
i
A
c
i
)
c
, then
the countable intersection of sets in / is again in /.
Example 2.4 Let X = [0, 1] and let / = ¦∅, X, [0,
1
2
], (
1
2
, 1]¦.
Then / is a σalgebra.
Example 2.5 Let X = ¦1, 2, 3¦ and let / = ¦X, ∅, ¦1¦, ¦2, 3¦¦.
Then / is a σalgebra.
Example 2.6 Let X = [0, 1], and let B
1
, . . . , B
8
be subsets of X
which are pairwise disjoint and whose union is all of X. Let / be
the collection of all ﬁnite unions of the B
i
’s as well as the empty
set. (Thus / consists of 2
8
elements.) Then / is a σalgebra.
Lemma 2.7 If /
α
is a σalgebra for each α in some index set I,
then ∩
α∈I
/
α
is a σalgebra.
Proof. This follows immediately from the deﬁnition.
If we have a collection ( of subsets of X, deﬁne
σ(() = ∩¦/
α
: /
α
is a σalgebra, ( ⊂ /¦,
the intersection of all σalgebras containing (. Since there is at
least one σalgebra containing (, namely, the one consisting of all
subsets of X, we are never taking the intersection over an empty
class of σalgebras. In view of Lemma 2.7, ( is a σalgebra. We
call σ(() the σalgebra generated by the collection (, or say that
2.1. ALGEBRAS AND σALGEBRAS 9
( generates the σalgebra σ((). It is clear that if (
1
⊂ (
2
, then
σ((
1
) ⊂ σ((
2
). Since σ(() is a σalgebra, then σ(σ(()) = σ(().
If X has some additional structure, say, it is a metric space,
then we can talk about open sets. If ( is the collection of open
subsets of X, then we call σ(() the Borel σalgebra on X, and this
is often denoted B. Elements of B are called Borel sets and are
said to be Borel measurable. We will see later that when X is the
real line, B is not equal to the collection of all subsets of X.
We end this section with the following proposition.
Proposition 2.8 If X = R, then the Borel σalgebra B is gener
ated by each of the following collection of sets:
(1) (
1
= ¦(a, b) : a, b ∈ R¦;
(2) (
2
= ¦[a, b] : a, b ∈ R¦;
(3) (
3
= ¦(a, b] : a, b ∈ R¦;
(4) (
4
= ¦(a, ∞) : a ∈ R¦.
Proof. (1) Let ( be the collection of open sets. By deﬁnition,
σ(() is the Borel σalgebra. Since every element of (
1
is open,
then (
1
⊂ (, and consequently σ((
1
) ⊂ σ(() = B.
To get the reverse inclusion, if G is open, it is the countable
union of open intervals by Proposition 1.5. Every ﬁnite open in
terval is in (
1
. Since (a, ∞) = ∪
∞
n=1
(a, a +n), then (a, ∞) ∈ σ((
1
)
if a ∈ R and similarly (−∞, a) ∈ σ((
1
) if a ∈ R. Hence if
G is open, then G ∈ σ((
1
). This says ( ⊂ σ((
1
), and then
B = σ(() ⊂ σ(σ((
1
)) = σ((
1
).
(2) If [a, b] ∈ (
2
, then [a, b] = ∩
∞
n=1
(a −
1
n
, b +
1
n
) ∈ σ(().
Therefore (
2
⊂ σ((), and hence σ((
2
) ⊂ σ(σ(()) = σ(() = B.
If (a, b) ∈ (
1
, choose n
0
≥ 2/(b −a) and note
(a, b) = ∪
∞
n=n
0
_
a +
1
n
, b −
1
n
_
∈ σ((
2
).
Therefore (
1
⊂ σ((
2
), from which it follows that B = σ((
1
) ⊂
σ(σ((
2
)) = σ((
2
).
(3) Using (a, b] = ∩
∞
n=1
(a, b +
1
n
), we see that (
3
⊂ σ((
1
), and
as above we conclude that σ((
3
) ⊂ σ((
1
) = B. Using (a, b) =
∪
∞
n=n
0
(a, b −
1
n
], provided n
0
is taken large enough, (
1
⊂ σ((
3
),
and as above we argue that B = σ((
1
) ⊂ σ((
3
).
10 CHAPTER 2. FAMILIES OF SETS
(4) Because (a, b] = (a, ∞) − (b, ∞), then (
1
⊂ σ((
4
). Since
(a, ∞) = ∪
∞
n=1
(a, a+n], then (
4
⊂ σ((
3
). As above, this is enough
to imply that σ((
4
) = B.
2.2 The monotone class theorem
This section will be used in Chapter 11.
Deﬁnition 2.9 A monotone class is a collection of subsets / of
X such that
(1) if A
i
↑ A and each A
i
∈ /, then A ∈ /;
(2) if A
i
↓ A and each A
i
∈ /, then A ∈ /.
The intersection of monotone classes is a monotone class, and
the intersection of all monotone classes containing a given collection
of sets is the smallest monotone class containing that collection.
The next theorem, the monotone class theorem, is rather tech
nical, but very useful.
Theorem 2.10 Suppose /
0
is an algebra, / is the smallest σ
algebra containing /
0
, and / is the smallest monotone class con
taining /
0
. Then / = /.
Proof. A σalgebra is clearly a monotone class, so / ⊂ /. We
must show / ⊂ /.
Let A
1
= ¦A ∈ / : A
c
∈ /¦. Note A
1
is contained in / and
contains /
0
. If A
i
↑ A and each A
i
∈ A
1
, then each A
c
i
∈ / and
A
c
i
↓ A
c
. Since / is a monotone class, A
c
∈ /, and so A ∈ A
1
.
Similarly, if A
i
↓ A and each A
i
∈ A
1
, then A ∈ A
1
. Therefore A
1
is a monotone class. Hence A
1
= /, and we conclude / is closed
under the operation of taking complements.
Let A
2
= ¦A ∈ / : A ∩ B ∈ / for all B ∈ /
0
¦. Note
the following: A
2
is contained in / and A
2
contains /
0
because
/
0
is an algebra. If A
i
↑ A, each A
i
∈ A
2
, and B ∈ /
0
, then
A∩B = ∪
∞
i=1
(A
i
∩B). Because /is a monotone class, A∩B ∈ /,
which implies A ∈ A
2
. We use a similar argument when A
i
↓ A.
2.3. EXERCISES 11
Therefore A
2
is a monotone class, and we conclude A
2
= /. In
other words, if B ∈ /
0
and A ∈ /, then A∩ B ∈ /.
Let A
3
= ¦A ∈ / : A ∩ B ∈ / for all B ∈ /¦. As in
the preceding paragraph, A
3
is a monotone class contained in /.
By the last sentence of the preceding paragraph, A
3
contains /
0
.
Hence A
3
= /.
We thus have that / is a monotone class closed under the
operations of taking complements and taking intersections. This
shows / is a σalgebra, and so / ⊂ /.
2.3 Exercises
Exercise 2.1 Find an example of a set X and a monotone class
/ consisting of subsets of X such that ∅ ∈ /, X ∈ /, but / is
not a σalgebra.
Exercise 2.2 Find an example of a set X and two σalgebras /
1
and /
2
, each consisting of subsets of X, such that /
1
∪ /
2
is not
a σalgebra.
Exercise 2.3 Suppose /
1
⊂ /
2
⊂ are σalgebras consisting
of subsets of a set X. Is ∪
∞
i=1
/
i
necessarily a σalgebra? If not,
give a counterexample.
Exercise 2.4 Suppose /
1
⊂ /
2
⊂ are monotone classes.
Let / = ∪
∞
n=1
/
n
. Suppose A
j
↑ A and each A
j
∈ /. Is A
necessarily in /? If not, give a counterexample.
Exercise 2.5 Let (Y, /) be a measurable space and let f map
X onto Y , but do not assume that f is onetoone. Deﬁne B =
¦f
−1
(A) : A ∈ /¦. Prove that B is a σalgebra of subsets of X.
Exercise 2.6 Suppose / is a σalgebra with the property that
whenever A ∈ /, there exist B, C ∈ / with B∩C = ∅, B∪C = A,
and neither B nor C is empty. Prove that / is uncountable.
Exercise 2.7 Suppose T is a collection of realvalued functions on
X such that the constant functions are in T and f +g, fg, and cf
12 CHAPTER 2. FAMILIES OF SETS
are in T whenever f, g ∈ T and c ∈ R. Suppose f ∈ T whenever
f
n
→ f and each f
n
∈ T. Deﬁne the function
χ
A
(x) =
_
1, x ∈ A;
0, x / ∈ A.
Prove that / = ¦A ⊂ X : χ
A
∈ T¦ is a σalgebra.
Exercise 2.8 Does there exist a σalgebra which has countably
many elements, but not ﬁnitely many?
Chapter 3
Measures
In this chapter we give the deﬁnition of a measure, some examples,
and some of the simplest properties of measures. Constructing
measures is often quite diﬃcult and we defer the construction of
the most important one, Lebesgue measure, until Chapter 4
3.1 Deﬁnitions and examples
Deﬁnition 3.1 Let X be a set and / a σalgebra consisting of
subsets of X. A measure on (X, /) is a function µ : / → [0, ∞]
such that
(1) µ(∅) = 0;
(2) if A
i
∈ /, i = 1, 2, . . ., are pairwise disjoint, then
µ(∪
∞
i=1
A
i
) =
∞
i=1
µ(A
i
).
Saying the A
i
are pairwise disjoint means that A
i
∩A
j
= ∅ if i ,= j.
Deﬁnition 3.1(2) is known as countable additivity. We say a set
function is ﬁnitely additive if µ(∪
n
i=1
A
i
) =
n
i=1
µ(A
i
) whenever
A
1
, . . . , A
n
are in / and are pairwise disjoint.
The triple (X, /, µ) is called a measure space.
13
14 CHAPTER 3. MEASURES
Example 3.2 Let X be any set, / the collection of all subsets of
X, and µ(A) the number of elements in A. µ is called counting
measure.
Example 3.3 Let X = R, / the collection of all subsets of R,
x
1
, x
2
, . . . ∈ R, and a
1
, a
2
, . . . ≥ 0. Set
µ(A) =
{i:x
i
∈A}
a
i
.
Counting measure is a particular case of this is if x
i
= i and all the
a
i
= 1.
Example 3.4 Let δ
x
(A) = 1 if x ∈ A and 0 otherwise. This
measure is called point mass at x.
Proposition 3.5 The following hold:
(1) If A, B ∈ / with A ⊂ B, then µ(A) ≤ µ(B).
(2) If A
i
∈ / and A = ∪
∞
i=1
A
i
, then µ(A) ≤
∞
i=1
µ(A
i
).
(3) Suppose A
i
∈ / and A
i
↑ A. Then µ(A) = lim
n→∞
µ(A
n
).
(4) Suppose A
i
∈ / and A
i
↓ A. If µ(A
1
) < ∞, then we have
µ(A) = lim
n→∞
µ(A
n
).
Proof. (1) Let A
1
= A, A
2
= B − A, and A
3
= A
4
= = ∅.
Now use part (2) of the deﬁnition of measure to write
µ(B) = µ(A) +µ(B −A) + 0 + 0 + ≥ µ(A).
(2) Let B
1
= A
1
, B
2
= A
2
− A
1
, B
3
= A
3
− (A
1
∪ A
2
), B
4
=
A
4
−(A
1
∪ A
2
∪ A
3
), and in general B
i
= A
i
−(∪
i−1
j=1
A
j
). The B
i
are pairwise disjoint, B
i
⊂ A
i
for each i, and ∪
∞
i=1
B
i
= ∪
∞
i=1
A
i
.
Hence
µ(A) = µ(∪
∞
i=1
B
i
) =
∞
i=1
µ(B
i
) ≤
∞
i=1
µ(A
i
).
(3) Deﬁne the B
i
as in (2). Since ∪
n
i=1
B
i
= ∪
n
i=1
A
i
, then
µ(A) = µ(∪
∞
i=1
A
i
) = µ(∪
∞
i=1
B
i
) =
∞
i=1
µ(B
i
)
= lim
n→∞
n
i=1
µ(B
i
) = lim
n→∞
µ(∪
n
i=1
B
i
) = lim
n→∞
µ(∪
n
i=1
A
i
).
3.1. DEFINITIONS AND EXAMPLES 15
(4) Apply (3) to the sets A
1
− A
i
, i = 1, 2, . . .. The sets A
1
− A
i
increase to A
1
−A, and so
µ(A
1
) −µ(A) = µ(A
1
−A) = lim
n→∞
µ(A
1
−A
n
)
= lim
n→∞
[µ(A
1
) −µ(A
n
)].
Now subtract µ(A
1
) from both sides and then multiply both sides
by −1.
Example 3.6 To see that µ(A
1
) < ∞ is necessary in Proposition
3.5, let X be the positive integers, µ counting measure, and A
i
=
¦i, i + 1, . . .¦. Then the A
i
decrease, µ(A
i
) = ∞ for all i, but
µ(∩
i
A
i
) = µ(∅) = 0.
Deﬁnition 3.7 A measure µ is a ﬁnite measure if µ(X) < ∞. A
measure µ is σﬁnite if there exist sets E
i
∈ / for i = 1, 2, . . . such
that µ(E
i
) < ∞for each i and X = ∪
∞
i=1
E
i
. If µ is a ﬁnite measure,
then (X, /, µ) is called a ﬁnite measure space, and similarly, if µ
is a σﬁnite measure, then (X, /, µ) is called a σﬁnite measure
space.
If we let F
n
= ∪
n
i=1
E
i
, then µ(F
n
) < ∞ for each n and F
n
↑ X.
Therefore there is no loss of generality in supposing the sets E
i
in
Deﬁnition 3.7 are increasing.
Let (X, /, µ) be a measure space. A subset A ⊂ X is a null set
if there exists a set B ∈ / with A ⊂ B and µ(B) = 0. We do not
require A to be in /. If / contains all the null sets, then (X, /, µ)
is said to be a complete measure space. The completion of / is the
smallest σalgebra / containing / such that (X, /, µ) is complete.
Sometimes one just says that / is complete or that µ is complete
when (X, /, µ) is complete.
A probability or probability measure is a measure µ such that
µ(X) = 1. In this case we usually write (Ω, T, P) instead of
(X, /, µ), and T is called a σﬁeld, which is the same thing as
a σalgebra.
16 CHAPTER 3. MEASURES
3.2 Exercises
Exercise 3.1 Suppose (X, /) is a measurable space and µ is a
nonnegative set function that is ﬁnitely additive and such that
µ(∅) = 0. Suppose that whenever A
i
is an increasing sequence of
sets in /, then µ(∪
i
A
i
) = lim
i→∞
µ(A
i
). Show that µ is a measure.
Exercise 3.2 Let X be an uncountable set and let / be the collec
tion of subsets A of X such that either A or A
c
is countable. Deﬁne
µ(A) = 0 if A is countable and µ(A) = 1 if A is uncountable. Prove
that µ is a measure.
Exercise 3.3 Suppose (X, /, µ) is a measure space and A, B ∈ /.
Prove that
µ(A) +µ(B) = µ(A∪ B) +µ(A∩ B).
Exercise 3.4 3. Prove that if µ
1
, µ
2
, . . . are measures on a mea
surable space and a
1
, a
2
, . . . ∈ [0, ∞), then
∞
n=1
a
n
µ
n
is also a
measure.
Exercise 3.5 3. Prove that if (X, /, µ) is a measure space, B ∈ /,
and we deﬁne ν(A) = µ(A∩ B) for A ∈ /, then ν is a measure.
Exercise 3.6 Suppose µ
1
, µ
2
, . . . are measures on a measurable
space (X, /) and µ
n
(A) ↑ for each A ∈ /. Deﬁne
µ(A) = lim
n→∞
µ
n
(A).
Is µ necessarily a measure? If not, give a counterexample. What
if µ
n
(A) ↓ for each A ∈ / and µ
1
(X) < ∞?
Exercise 3.7 Let (X, /, µ) be a measure space, let A be the col
lection of null sets with respect to /, and let B = σ(/∪A). Show
that (X, B, µ) is complete and is the completion of (X, /, µ).
Exercise 3.8 Suppose X is the set of real numbers, B is the Borel
σalgebra, and m and n are two measures on (X, B) such that
m((a, b)) = n((a, b)) whenever a < b. Prove that m(A) = n(A)
whenever A ∈ B.
3.2. EXERCISES 17
Exercise 3.9 Suppose (X, /) is a measurable space and ( is an
arbitrary subset of /. Suppose m and n are two σﬁnite mea
sures on (X, /) such that m(A) = n(A) for all A ∈ (. Is it true
that m(A) = n(A) for all A ∈ σ(()? What if m and n are ﬁnite
measures?
18 CHAPTER 3. MEASURES
Chapter 4
Construction of
measures
Our goal in this chapter is give a method for constructing measures.
This is a complicated procedure, and involves the concept of outer
measure, which we introduce in Section 4.1.
Our most important example will be onedimensional Lebesgue
measure, which we consider in Section 4.2. Further results and
some examples related to Lebesgue measure are given in Section
4.3.
One cannot deﬁne the Lebesgue measure of every subset of the
reals. This is shown in Section 4.4.
The methods used to construct measures via outer measures
have other applications besides the construction of Lebesgue mea
sure. The Carath´eodory extension theorem is a tool developed in
Section 4.5 that can be used in constructing measures.
Let us present some of the ideas used in the construction of
Lebesgue measure on the line. We want the measure m of an open
interval to be the length of the interval. Since every open subset
of the reals is the countable union of disjoint open intervals (see
Proposition 1.5), if G = ∪
∞
i=1
(a
i
, b
i
), where the intervals (a
i
, b
i
) are
pairwise disjoint, we must have
m(G) =
∞
i=1
(b
i
−a
i
).
19
20 CHAPTER 4. CONSTRUCTION OF MEASURES
We then set
m(E) = inf¦m(G) : G open, A ⊂ G¦
for arbitrary subsets E ⊂ R. The diﬃculty is that m is not a
measure on the σalgebra consisting of all subsets of the reals; this
is proved in Section 4.4. We resolve this by considering a strictly
smaller σalgebra. This is the essential idea behind the construction
of Lebesgue measure, but it is technically simpler to work with
intervals of the form (a, b] rather than open intervals.
4.1 Outer measures
We begin with the notion of outer measure.
Deﬁnition 4.1 Let X be a set. An outer measure is a function
µ
∗
deﬁned on the collection of all subsets of X satisfying
(1) µ
∗
(∅) = 0;
(2) if A ⊂ B, then µ
∗
(A) ≤ µ
∗
(B);
(3) µ
∗
(∪
∞
i=1
A
i
) ≤
∞
i=1
µ
∗
(A
i
) whenever A
1
, A
2
, . . . are subsets of
X.
A set N is a null set with respect to µ
∗
if µ
∗
(N) = 0.
A common way to generate outer measures is as follows.
Proposition 4.2 Suppose ( is a collection of subsets of X such
that ∅ and X are both in (. Suppose : ( → [0, ∞] with (∅) = 0.
Deﬁne
µ
∗
(E) = inf
_
∞
i=1
(A
i
) : A
i
∈ ( for each i and E ⊂ ∪
∞
i=1
A
i
_
.
(4.1)
Then µ
∗
is an outer measure.
Proof. (1) and (2) of the deﬁnition of outer measure are obvious.
To prove (3), let A
1
, A
2
, . . . be subsets of X and let ε > 0. For
each i there exist C
i1
, C
i2
, . . . ∈ ( such that A
i
⊂ ∪
∞
j=1
C
ij
and
4.1. OUTER MEASURES 21
j
(C
ij
) ≤ µ
∗
(A
i
) +ε/2
i
. Then ∪
∞
i=1
A
i
⊂ ∪
i
∪
j
C
ij
and
µ
∗
(∪
∞
i=1
A
i
) ≤
i,j
(C
ij
) =
i
_
j
(C
ij
)
_
≤
∞
i=1
µ
∗
(A
i
) +
∞
i=1
ε/2
i
=
∞
i=1
µ
∗
(A
i
) +ε.
Since ε is arbitrary, µ
∗
(∪
∞
i=1
A
i
) ≤
∞
i=1
µ
∗
(A
i
).
Example 4.3 Let X = R and let ( be the collection of intervals
of the form (a, b], that is, intervals that are open on the left and
closed on the right. Let (I) = b − a if I = (a, b]. Deﬁne µ
∗
by
(4.1). Proposition 4.2 shows that µ
∗
is an outer measure, but we
will see in Section 4.4 that µ
∗
is not a measure on the collection
of all subsets of R. We will also see, however, that if we restrict
µ
∗
to a σalgebra / which is strictly smaller than the collection of
all subsets of R, then µ
∗
will be a measure on /. That measure is
what is known as Lebesgue measure. The σalgebra / is called the
Lebesgue σalgebra.
Example 4.4 Let X = R and let ( be the collection of intervals
of the form (a, b] as in the previous example. Let α : R → R
be an increasing right continuous function on R. Thus α(x) =
lim
y→x+
α(y) for each x and α(x) ≤ α(y) if x < y. Let (I) =
α(b) − α(a) if I = (a, b]. Again deﬁne µ
∗
by (4.1). Again Propo
sition 4.2 shows that µ
∗
is an outer measure. Restricting µ
∗
to
a smaller σalgebra gives us what is known as LebesgueStieltjes
measure corresponding to α. The special case where α(x) = x for
all x is Lebesgue measure.
In general we need to restrict µ
∗
to a strictly smaller σalgebra
than the collection of all subsets of R, but not always. For example,
if α(x) = 0 for x < 0 and 1 for x ≥ 1, then the corresponding
LebesgueStieltjes measure is point mass at 0 (deﬁned in Example
3.4), and the corresponding σalgebra is the collection of all subsets
of R.
22 CHAPTER 4. CONSTRUCTION OF MEASURES
Deﬁnition 4.5 Let µ
∗
be an outer measure. A set A ⊂ X is
µ
∗
measurable if
µ
∗
(E) = µ
∗
(E ∩ A) +µ
∗
(E ∩ A
c
) (4.2)
for all E ⊂ X.
Theorem 4.6 If µ
∗
is an outer measure on X, then the collection
/ of µ
∗
measurable sets is a σalgebra. If µ is the restriction of µ
∗
to /, then µ is a measure. Moreover, / contains all the null sets.
This is sometimes known as Carath´eodory’s theorem, but do not
confuse this with the Carath´eodory extension theorem in Section
4.5.
Proof. By Deﬁnition 4.1,
µ
∗
(E) ≤ µ
∗
(E ∩ A) +µ
∗
(E ∩ A
c
)
for all E ⊂ X. Thus to check (4.2) it is enough to show
µ
∗
(E) ≥ µ
∗
(E ∩ A) +µ
∗
(E ∩ A
c
).
This will be trivial in the case µ
∗
(E) = ∞.
Step 1. First we show / is an algebra. If A ∈ /, then A
c
∈ / by
symmetry and the deﬁnition of /. Suppose A, B ∈ / and E ⊂ X.
Then
µ
∗
(E) = µ
∗
(E ∩ A) +µ
∗
(E ∩ A
c
)
= [µ
∗
(E ∩ A∩ B) +µ
∗
(E ∩ A∩ B
c
)]
+ [µ
∗
(E ∩ A
c
∩ B) +µ
∗
(E ∩ A
c
∩ B
c
)].
The second equality follows from the deﬁnition of / with E ﬁrst
replaced by E∩A and then by E∩A
c
. The ﬁrst three summands on
the right of the second equals sign have a sum greater than or equal
to µ
∗
(E∩(A∪B)) because A∪B ⊂ (A∩B) ∪(A∩B
c
) ∪(A
c
∩B).
Since A
c
∩ B
c
= (A∪ B)
c
, then
µ
∗
(E) ≥ µ
∗
(E ∩ (A∪ B)) +µ
∗
(E ∩ (A∪ B)
c
),
which shows A∪ B ∈ /. Therefore / is an algebra.
4.1. OUTER MEASURES 23
Step 2. Next we show / is a σalgebra. Let A
i
be pairwise disjoint
sets in /, let B
n
= ∪
n
i=1
A
i
, and B = ∪
∞
i=1
A
i
. If E ⊂ X,
µ
∗
(E ∩ B
n
) = µ
∗
(E ∩ B
n
∩ A
n
) +µ
∗
(E ∩ B
n
∩ A
c
n
)
= µ
∗
(E ∩ A
n
) +µ
∗
(E ∩ B
n−1
).
Similarly, µ
∗
(E ∩ B
n−1
) = µ
∗
(E ∩ A
n−1
) + µ
∗
(E ∩ B
n−2
), and
continuing, we obtain
µ
∗
(E ∩ B
n
) ≥
n
i=1
µ
∗
(E ∩ A
i
).
Since B
n
∈ /, then
µ
∗
(E) = µ
∗
(E ∩B
n
) +µ
∗
(E ∩B
c
n
) ≥
n
i=1
µ
∗
(E ∩A
i
) +µ
∗
(E ∩B
c
).
Let n → ∞. Recalling that µ
∗
is an outer measure,
µ
∗
(E) ≥
∞
i=1
µ
∗
(E ∩ A
i
) +µ
∗
(E ∩ B
c
) (4.3)
≥ µ
∗
(∪
∞
i=1
(E ∩ A
i
)) +µ
∗
(E ∩ B
c
)
= µ
∗
(E ∩ B) +µ
∗
(E ∩ B
c
)
≥ µ
∗
(E).
This shows B ∈ /.
Now if C
1
, C
2
, . . . are sets in /, let A
1
= C
1
, A
2
= C
2
− A
1
,
A
3
= C
3
− (A
1
∪ A
2
), and in general A
i
= C
i
− (∪
i−1
j=1
A
j
). Since
each C
i
∈ / and / is an algebra, then A
i
= C
i
∩ C
c
i−1
∈ /. The
A
i
are pairwise disjoint, so from the previous paragraph,
∪
∞
i=1
C
i
= ∪
∞
i=1
A
i
∈ /.
Also, ∩
∞
i=1
C
i
= (∪
∞
i=1
C
c
i
)
c
∈ /, and therefore / is a σalgebra.
Step 3. We now show µ
∗
restricted to / is a measure. The only way
(4.3) can hold is if all the inequalities there are actually equalities,
and in particular,
µ
∗
(E) =
∞
i=1
µ
∗
(E ∩ A
i
) +µ
∗
(E ∩ B
c
).
24 CHAPTER 4. CONSTRUCTION OF MEASURES
Taking E = B, we obtain
µ
∗
(B) =
∞
i=1
µ
∗
(A
i
),
or µ
∗
is countably additive on /.
Step 4. Finally, if µ
∗
(A) = 0 and E ⊂ X, then
µ
∗
(E ∩ A) +µ
∗
(E ∩ A
c
) = µ
∗
(E ∩ A
c
) ≤ µ
∗
(E),
which shows / contains all the null sets.
4.2 LebesgueStieltjes measures
Let X = R and let ( be the collection of intervals of the form (a, b],
that is, intervals that are open on the left and closed on the right.
Let α(x) be an increasing right continuous function. This means
that α(x) ≤ α(y) if x < y and lim
z→x+
α(z) = α(x) for all x. We
do not require α to be strictly increasing. Deﬁne
((a, b]) = α(b) −α(a).
Deﬁne
m
∗
(E) = inf
_
∞
i=1
(A
i
) : A
i
∈ ( for each i and E ⊂ ∪
∞
i=1
A
i
_
.
(In this book we usually use m instead of µ when we are talking
about LebesgueStieltjes measures.) We use Proposition 4.2 to tell
us that m
∗
is an outer measure. We then use Theorem 4.6 to show
that m
∗
is a measure on the collection of m
∗
measurable sets. Note
that if K and L are adjacent intervals, that is, if K = (a, b] and
L = (b, c], then K ∪ L = (a, c] and
(K) +(L) = [α(b) −α(a)] + [α(c) −α(b)] (4.4)
= α(c) −α(a) = (K ∪ L)
by the deﬁnition of .
The next step in the construction of LebesgueStieltjes measure
corresponding to α is the following.
4.2. LEBESGUESTIELTJES MEASURES 25
Proposition 4.7 Every set in the Borel σalgebra on R is m
∗

measurable.
Proof. Since the collection of m
∗
measurable sets is a σalgebra,
it suﬃces to show that every interval J of the form (c, d] is m
∗

measurable. Let E be any set with m
∗
(E) < ∞; we need to show
m
∗
(E) ≥ m
∗
(E ∩ J) +m
∗
(E ∩ J
c
). (4.5)
Choose I
1
, I
2
, . . ., each of the form (a
i
, b
i
], such that E ⊂ ∪
i
I
i
and
m
∗
(E) ≥
i
[α(b
i
) −α(a
i
)] −ε.
Since E ⊂ ∪
i
I
i
, we have
m
∗
(E ∩ J) ≤
i
m
∗
(I
i
∩ J)
and
m
∗
(E ∩ J
c
) ≤
i
m
∗
(I
i
∩ J
c
).
Adding we have
m
∗
(E ∩ J) +m
∗
(E ∩ J
c
) ≤
i
[m
∗
(I
i
∩ J) +m
∗
(I
i
∩ J
c
)].
Now I
i
∩J is an interval that is open on the left and closed on the
right, and I
i
∩ J
c
is the union of zero, one, or two such intervals,
depending on the relative locations of I
i
and J. Using (4.4) either
zero, one, or two times, we see that
m
∗
(I
i
∩ J) +m
∗
(I
i
∩ J
c
) = m
∗
(I
i
).
Thus
m
∗
(E ∩ J) +m
∗
(E ∩ J
c
) ≤
i
m
∗
(I
i
) ≤ m
∗
(E) +ε.
Since ε is arbitrary, this proves (4.5).
After all this work, it would be upsetting if the measure of a
halfopen interval (e, f] were not what it is supposed to be. For
tunately, everything is ﬁne, due to Proposition 4.9. First we need
the following lemma.
26 CHAPTER 4. CONSTRUCTION OF MEASURES
Lemma 4.8 Let J
k
= (a
k
, b
k
), k = 1, . . . , n, be a ﬁnite collection
of ﬁnite open intervals covering a ﬁnite closed interval [C, D]. Then
n
k=1
[α(b
k
) −α(a
k
)] ≥ α(D) −α(C). (4.6)
Proof. Since ¦J
k
¦ is a cover of [C, D], there exists at least one
interval, say, J
k
1
, such that C ∈ J
k
1
. If J
k
1
covers [C, D], we stop.
Otherwise, b
k
1
≤ D, and there must be at least one interval, say,
J
k
2
, such that b
k
1
∈ J
k
2
. If [C, D] ⊂ J
k
1
∪J
k
2
, we stop. If not, then
b
k
1
< b
k
2
≤ D, and there must be at least one interval, say, J
k
3
that
contains b
k
2
. At each stage we choose J
k
j
so that b
k
j−1
∈ J
k
j
. We
continue until we have covered [C, D] with intervals J
k
1
, . . . , J
k
m
.
Since ¦J
k
¦ is a ﬁnite cover, we will stop for some m ≤ n.
By our construction we have
a
k
1
≤ C < b
k
1
, a
k
m
< D < b
k
m
,
and for 2 ≤ j ≤ m,
a
k
j
< b
k
j−1
< b
k
j
.
Then
α(D) −α(C) ≤ α(b
k
m
) −α(a
k
1
)
= [α(b
k
m
) −α(b
k
m−1
)] + [α(b
k
m−1
) −α(b
k
m−2
)] +
+ [α(b
k
2
) −α(b
k
1
)] + [α(b
k
1
) −α(a
k
1
)]
≤ [α(b
k
m
) −α(a
k
m
)] + [α(b
k
m−1
) −α(a
k
m−1
)] +
+ [α(b
k
2
) −α(a
k
2
)] + [α(b
k
1
) −α(a
k
1
)].
Since ¦J
k
1
, . . . , J
k
m
¦ ⊂ ¦J
1
, . . . , J
n
¦, this proves (4.6).
Proposition 4.9 If e and f are ﬁnite and I = (e, f], then m
∗
(I) =
(I).
Proof. First we show m
∗
(I) ≤ (I). This is easy. Let A
1
= I and
A
2
= A
3
= = ∅. Then I ⊂ ∪
∞
i=1
A
i
, hence
m
∗
(I) ≤
∞
i=1
(A
i
) = (A
1
) = (I).
4.3. EXAMPLES AND RELATED RESULTS 27
For the other direction, suppose I ⊂ ∪
∞
i=1
A
i
, where A
i
= (c
i
, d
i
].
Let ε > 0 and choose C ∈ (e, f) such that α(C) − α(e) < ε/2.
This is possible by the right continuity of α. Let D = f. For
each i, choose d
i
> d
i
such that α(d
i
) − α(d
i
) < ε/2
i+1
and let
B
i
= (c
i
, d
i
).
Then [C, D] is compact and ¦B
i
¦ is an open cover for [C, D].
Use compactness to choose a ﬁnite subcover ¦J
1
, . . . , J
n
¦ of ¦B
i
¦.
We now apply Lemma 4.8. We conclude that
(I) ≤ α(D)−α(C)+ε/2 ≤
n
k=1
(α(d
k
)−α(c
k
))+ε/2 ≤
∞
i=1
(A
i
)+ε.
Taking the inﬁmum over all countable collections ¦A
i
¦ that
cover I, we obtain
(I) ≤ m
∗
(I) +ε.
Since ε is arbitrary, (I) ≤ m
∗
(I).
We now drop the asterisks from m
∗
and call m LebesgueStieltjes
measure. In the special case where α(x) = x, m is Lebesgue mea
sure. In the special case of Lebesgue measure, the collection of m
∗

measurable sets is called the Lebesgue σalgebra. A set is Lebesgue
measurable if it is in the Lebesgue σalgebra.
Given a measure µ on R such that µ(K) < ∞ whenever K is
compact, deﬁne α(x) = µ((0, x]) if x ≥ 0 and α(x) = −µ((x, 0]) if
x < 0. Then α is increasing, right continuous, and Exercise 4.1 asks
you to show that µ is LebesgueStieltjes measure corresponding to
α.
4.3 Examples and related results
Example 4.10 Let m be Lebesgue measure. If x ∈ R, then ¦x¦
is a closed set and hence is Borel measurable. Moreover
m(¦x¦) = lim
n→∞
m((x −(1/n), x]) = lim
n→∞
[x −(x −(1/n))] = 0.
We then conclude
m([a, b]) = m((a, b]) +m(¦a¦) = b −a + 0 = b −a
28 CHAPTER 4. CONSTRUCTION OF MEASURES
and
m((a, b)) = m((a, b]) −m(¦b¦) = b −a −0 = b −a.
Since σalgebras are closed under the operation of countable
unions, then countable sets are Borel measurable. Adding 0 to
itself countably many times is still 0, so the Lebesgue measure of
a countable set is 0.
However there are uncountable sets which have Lebesgue mea
sure 0. See the next example.
Example 4.11 Recall from undergraduate analysis that the Can
tor set is constructed as follows. Let F
0
be the interval [0, 1]. We
let F
1
be what remains if we remove the middle third, that is,
F
1
= F
0
−(
1
3
,
2
3
).
F
1
consists of two intervals of length
1
3
each. We remove the middle
third of each of these two intervals and let
F
2
= F
1
−[(
1
9
,
2
9
) ∪ (
7
9
,
8
9
)].
We continue removing middle thirds, and the Cantor set F is ∩
n
F
n
.
Recall that the Cantor set is closed, uncountable, and every point
is a limit point. Moreover, it contains no intervals.
The measure of F
1
is 2(
1
3
), the measure of F
2
is 4(
1
9
), and the
measure of F
n
is (
2
3
)
n
. Since the Cantor set C is the intersection
of all these sets, the Lebesgue measure of C is 0.
Suppose we deﬁne f
0
to be
1
2
on the interval (
1
3
,
2
3
), to be
1
4
on
the interval (
1
9
,
2
9
), to be
3
4
on the interval (
7
9
,
8
9
), and so on. Deﬁne
f(x) = inf¦f
0
(y) : y ≥ x¦ for x < 1. Deﬁne f(1) = 1. Notice
f = f
0
on the complement of the Cantor set. f is increasing,
so it has only jump discontinuities; see Proposition 1.6. But if it
has a jump continuity, there is a rational of the form k/2
n
with
k ≤ 2
n
that is not in the range of f. On the other hand, by the
construction, each of the values ¦k/2
n
: n ≥ 0, k ≤ 2
n
¦ is taken by
f
0
for some point in the complement of C, and so is taken by f.
The only way this can happen is if f is continuous. This function
f is called the CantorLebesgue function or sometimes simply the
Cantor function. We will use it in examples later on. For now, we
note that it a function that increases only on the Cantor set, which
is a set of Lebesgue measure 0, yet f is continuous.
4.3. EXAMPLES AND RELATED RESULTS 29
Example 4.12 Let q
1
, q
2
, . . . be an enumeration of the rationals,
let ε > 0, and let I
i
be the interval (q
i
−ε/2
i
, q
i
+ε/2
i
). Then the
measure of I
i
is ε/2
i−1
, so the measure of ∪
i
I
i
is at most 2ε. (It is
not equal to that because there is a lot of overlap.) Therefore the
measure of A = [0, 1] −∪
i
I
i
is larger than 1 −2ε. But A contains
no rational numbers.
Example 4.13 Let us follow the construction of the Cantor set,
with this diﬀerence. Instead of removing the middle third at the
ﬁrst stage, remove the middle fourth, i.e., remove (
3
8
,
5
8
). On each
of the two intervals that remain, remove the middle sixteenths. On
each of the four intervals that remain, remove the middle interval
of length
1
64
, and so on. The total that we removed is
1
4
+ 2(
1
16
) + 4(
1
64
) + =
1
2
.
The set that remains contains no intervals, is closed, every point
is a limit point, is uncountable, and has measure 1/2. Such a set
is called a generalized Cantor set. Of course, other choices than
1
4
,
1
16
, etc. are possible.
Let A ⊂ [0, 1] be a Borel measurable set. We will show that
A is “almost equal” to the countable intersection of open sets and
“almost equal” to the countable union of closed sets. (A similar
argument to what follows is possible for subsets of R that have
inﬁnite measure; see Exercise 4.2.)
Proposition 4.14 Suppose A ⊂ [0, 1] is a Borel measurable set.
Let m be Lebesgue measure.
(1) Given ε > 0, there exists an open set G so that m(G−A) < ε
and A ⊂ G.
(2) Given ε > 0, there exists a closed set F so that m(A−F) < ε
and F ⊂ A.
(3) There exists a set H which contains A that is the countable
intersection of a decreasing sequence of open sets and m(H−A) =
0.
(4) There exists a set F which is contained in A that is the
countable union of an increasing sequence of closed sets which is
contained in A and m(A−F) = 0.
30 CHAPTER 4. CONSTRUCTION OF MEASURES
Proof. (1) There exists a set of the form E = ∪
∞
j=1
(a
j
, b
j
] such
that A ⊂ E and m(E −A) < ε/2. Let G = ∪
∞
j=1
(a
j
, b
j
+ε2
−j−1
).
Then G is open and contains A and
m(G−E) <
∞
j=1
ε2
−j−1
= ε/2.
Therefore
m(G−A) ≤ m(G−E) +m(E −A) < ε.
(2) Find an open set G such that A
⊂ G and m(G − A
) < ε,
where A
= [0, 1] −A. Let F = [0, 1] −G. Then F is closed, F ⊂ A,
and m(A−F) ≤ m(G−A
) < ε.
(3) By (1), for each i, there is an open set G
i
that contains A
and such that m(G
i
− A) < 2
−i
. Then H
i
= ∩
i
j=1
G
j
will contain
A, is open, and since it is contained in G
i
, then m(H
i
−A) < 2
−i
.
Let H = ∩
∞
i=1
H
i
. H need not be open, but it is the intersection of
countably many open sets. The set H is a Borel set, contains A,
and m(H−A) ≤ m(H
i
−A) < 2
−i
for each i, hence m(H−A) = 0.
(4) If A
= [0, 1] −A, apply (3) to A
to ﬁnd a set H containing
A
that is the countable intersection of a decreasing sequence of
open sets and such that m(H − A
) = 0. Let J = [0, 1] − H. It is
left to the reader to verify that J has the desired properties.
The countable intersections of open sets are sometimes called
G
δ
sets; the G is for geoﬀnet, the German word for “open” and
the δ for Durchschnitt, the German word for “intersection.” The
countable unions of closed sets are called F
σ
sets, the F coming
from ferm´e, the French word for “closed,” and the σ coming from
Summe, the German word for “union.”
Therefore, when trying to understand Lebesgue measure, we
can look at G
δ
or F
σ
sets, which are not so bad, and at null sets,
which can be quite bad but don’t have positive measure.
4.4 Nonmeasurable sets
Theorem 4.15 Let m
∗
be deﬁned by (4.1), where ( is the collec
tion of intervals that are open on the left and closed on the right
4.5. THE CARATH
´
EODORY EXTENSION THEOREM 31
and ((a, b]) = b − a. m
∗
is not a measure on the collection of all
subsets of R.
Proof. Suppose m
∗
is a measure. Deﬁne x ∼ y if x−y is rational.
It is easy to see that this is an equivalence relationship on [0, 1].
For each equivalence class, pick an element out of that class (we
need to use the axiom of choice to do this). Call the collection of
such points A. Given a set B, deﬁne B+x = ¦y+x : y ∈ B¦. Note
that ((a +q, b +q]) = b −a = ((a, b]) for each a, b, and q, and so
by the deﬁnition of m
∗
, we have m
∗
(A + q) = m
∗
(A) for each set
A and each q. Moreover, the sets A + q are disjoint for diﬀerent
rationals q.
Now
[0, 1] ⊂ ∪
q∈[−1,1]∩Q
(A+q),
where the sum is only over rational q, so
1 ≤
q∈[−1,1],q∈Q
m
∗
(A+q),
and therefore m
∗
(A) > 0. But
∪
q∈[−1,1]∩Q
(A+q) ⊂ [−1, 2],
where again the sum is only over rational q, so if m
∗
is a measure,
then
3 ≥
q∈[0,1],q∈Q
m
∗
(A+q),
which implies m
∗
(A) = 0, a contradiction.
4.5 The Carath´eodory extension theo
rem
We prove the Carath´eodory extension theorem in this section. This
theorem abstracts some of the techniques used above to give a tool
for constructing measures in a variety of contexts.
Let /
0
be an algebra but not necessarily a σalgebra. Saying
is a measure on /
0
means the following: (1) of Deﬁnition 3.1
32 CHAPTER 4. CONSTRUCTION OF MEASURES
holds and if A
1
, A
2
, . . . are pairwise disjoint elements of /
0
and
also ∪
∞
i=1
A
i
∈ /
0
, then (∪
∞
i=1
A
i
) =
∞
i=1
(A
i
). Sometimes one
calls a measure on an algebra a premeasure. Recall σ(/
0
) is the
σalgebra generated by /
0
.
Theorem 4.16 Suppose /
0
is an algebra and : /
0
→ [0, ∞] is
a measure on /
0
. Deﬁne
µ
∗
(E) = inf
_
∞
i=1
(A
i
) : each A
i
∈ /
0
, E ⊂ ∪
∞
i=1
A
i
_
for E ⊂ X. Then
(1) µ
∗
is an outer measure;
(2) µ
∗
(A) = (A) if A ∈ /
0
;
(3) every set in /
0
is µ
∗
measurable;
(4) if is σﬁnite, then there is a unique extension to σ(/
0
).
Proof. (1) is Proposition 4.2. We turn to (2). Suppose E ∈ /
0
.
We know µ
∗
(E) ≤ (E) since we can take A
1
= E and A
2
, A
3
, . . .
empty in the deﬁnition of µ
∗
. If E ⊂ ∪
∞
i=1
A
i
with A
i
∈ /
0
, let
B
n
= E ∩ (A
n
− (∪
n−1
i=1
A
i
)). Then the B
n
are pairwise disjoint,
they are each in /
0
, and their union is E. Therefore
(E) =
∞
i=1
(B
i
) ≤
∞
i=1
(A
i
).
Taking the inﬁmum over all such sequences A
1
, A
2
, . . . shows that
(E) ≤ µ
∗
(E).
Next we look at (3). Suppose A ∈ /
0
. Let ε > 0 and let E ⊂
X. Pick B
1
, B
2
, . . . ∈ /
0
such that E ⊂ ∪
∞
i=1
B
i
and
i
(B
i
) ≤
µ
∗
(E) +ε. Then
µ
∗
(E) +ε ≥
∞
i=1
(B
i
) =
∞
i=1
(B
i
∩ A) +
∞
i=1
(B
i
∩ A
c
)
≥ µ
∗
(E ∩ A) +µ
∗
(E ∩ A
c
).
Since ε is arbitrary, µ
∗
(E) ≥ µ
∗
(E ∩ A) + µ
∗
(E ∩ A
c
). Thus A is
µ
∗
measurable.
Finally, we look at (4). Suppose we have two extensions to
σ(/
0
), the smallest σalgebra containing /
0
. One is µ
∗
and let the
4.5. THE CARATH
´
EODORY EXTENSION THEOREM 33
other extension be called ν. We will show that if E is in σ(/
0
),
then µ
∗
(E) = ν(E).
Let us ﬁrst assume that µ
∗
is a ﬁnite measure. The µ
∗
measur
able sets form a σalgebra containing /
0
. Because E ∈ /
0
, E must
be µ
∗
measurable and
µ
∗
(E) = inf
_
∞
i=1
(A
i
) : E ⊂ ∪
∞
i=1
A
i
, each A
i
∈ /
0
_
.
But = ν on /
0
, so
i
(A
i
) =
i
ν(A
i
). Therefore if E ⊂ ∪
∞
i=1
A
i
with each A
i
∈ /
0
, then
ν(E) ≤
i
ν(A
i
) =
i
(A
i
),
which implies
ν(E) ≤ µ
∗
(E). (4.7)
Since we do not know that ν is constructed via an outer measure,
we must use a diﬀerent argument to get the reverse inequality. Let
ε > 0 and choose A
i
∈ /
0
such that µ
∗
(E) + ε ≥
i
(A
i
) and
E ⊂ ∪
i
A
i
. Let A = ∪
∞
i=1
A
i
and B
k
= ∪
k
i=1
A
i
. Observe
µ
∗
(E) +ε ≥
i
(A
i
) =
i
µ
∗
(A
i
) ≥ µ
∗
(∪
i
A
i
) = µ
∗
(A),
hence µ
∗
(A−E) ≤ ε. We have
µ
∗
(A) = lim
k→∞
µ
∗
(B
k
) = lim
k→∞
ν(B
k
) = ν(A).
Then
µ
∗
(E) ≤ µ
∗
(A) = ν(A) = ν(E) +ν(A−E)
≤ ν(E) +µ
∗
(A−E) ≤ ν(E) +ε,
using (4.7) in the next to last inequality. Since ε is arbitrary, this
completes the proof when is ﬁnite.
It remains to consider the case when is σﬁnite. Write X =
∪
i
K
i
, where K
i
↑ X and (K
i
) < ∞ for each i. By the preceding
paragraph we have uniqueness for the measure
i
deﬁned by
i
(A) =
(A∩K
i
). If µ and ν are two extensions of and A ∈ σ(/
0
), then
µ(A) = lim
i→∞
µ(A∩ K
i
) = lim
i→∞
i
(A) = lim
i→∞
ν(A∩ K
i
) = ν(A),
which proves µ = ν.
34 CHAPTER 4. CONSTRUCTION OF MEASURES
4.6 Exercises
Exercise 4.1 Let µ be a measure on the Borel σalgebra of R such
that µ(K) < ∞ whenever K is compact, deﬁne α(x) = µ((0, x])
if x ≥ 0 and α(x) = −µ((x, 0]) if x < 0. Show that µ is the
LebesgueStieltjes measure corresponding to α.
Exercise 4.2 Let m be Lebesgue measure and A a Lebesgue mea
surable subset of R with m(A) < ∞. Let ε > 0. Show there exist
G open and F closed such that F ⊂ A ⊂ G and m(G−F) < ε.
Exercise 4.3 If (X, /, µ) is a measure space, deﬁne
µ
∗
(A) = inf¦µ(B) : A ⊂ B, B ∈ /¦
for all subsets A of X. Show that µ
∗
is an outer measure. Show
that each set in / is µ
∗
measurable and µ
∗
agrees with the measure
µ on /.
Exercise 4.4 Let m be LebesgueStieltjes measure corresponding
to a right continuous increasing function α. Show that for each x,
m(¦x¦) = α(x) − lim
y→x−
α(y).
Exercise 4.5 Suppose m is Lebesgue measure. Deﬁne x + A =
¦x+y : y ∈ A¦ and cA = ¦cy : y ∈ A¦. Show that for all Lebesgue
measurable sets Awe have m(x+A) = m(A) and m(cA) = [c[m(A).
Exercise 4.6 Let m be Lebesgue measure. Suppose for each n,
A
n
is a Lebesgue measurable subset of [0, 1]. Let B consist of those
points x that are in inﬁnitely many of the A
n
.
(1) Show B is Lebesgue measurable.
(2) If m(A
n
) > δ > 0 for each n, show m(B) ≥ δ.
(3) If
∞
n=1
m(A
n
) < ∞, prove that m(B) = 0.
(4) Give an example where
∞
n=1
m(A
n
) = ∞, but m(B) = 0.
Exercise 4.7 Suppose ε ∈ (0, 1) and m is Lebesgue measure. Find
a measurable set E ⊂ [0, 1] such that the closure of E is [0, 1] and
m(E) = ε.
4.6. EXERCISES 35
Exercise 4.8 If X is a metric space, B is the Borel σalgebra, and
µ is a measure on (X, B), then the support of µ is the smallest
closed set F such that µ(F
c
) = 0. Show that if F is a closed subset
of [0, 1], then there exists a ﬁnite measure on [0, 1] whose support
is F.
Exercise 4.9 Let m be Lebesgue measure. Find an example of
Lebesgue measurable subsets A
1
, A
2
, . . . of [0, 1] such that m(A
n
) >
0 for each n, m(A
n
∆A
m
) > 0 if n ,= m, and m(A
n
∩ A
m
) =
m(A
n
)m(A
m
) if n ,= m.
Exercise 4.10 Let ε ∈ (0, 1), let m be Lebesgue measure, and
suppose A is a Borel measurable subset of R. Prove that if
m(A∩ I) ≤ (1 −ε)m(I)
for every interval I, then m(A) = 0.
Exercise 4.11 Suppose m is Lebesgue measure and A is a Borel
measurable subset of R with m(A) > 0. Prove that if
B = ¦x −y : x, y ∈ A¦,
then B contains a nonempty open interval centered at the origin.
Exercise 4.12 Let m be Lebesgue measure. Construct a Borel
subset A of R such that 0 < m(A ∩ I) < m(I) for every open
interval I.
Exercise 4.13 Let N be the nonmeasurable set deﬁned in Section
4.4. Prove that if A ⊂ N and A is Lebesgue measurable, then
m(A) = 0.
Exercise 4.14 Let m be Lebesgue measure. Prove that if A is a
Lebesgue measurable subset of R and m(A) > 0, then there is a
subset of A that is nonmeasurable.
Exercise 4.15 Let X be a set and / a collection of subsets of X
that form an algebra of sets. Suppose is a measure on / such
that (X) < ∞. Deﬁne µ
∗
using as in (4.1). Prove that a set A
is µ
∗
measurable if and only if
µ
∗
(A) = (X) −µ
∗
(A
c
).
36 CHAPTER 4. CONSTRUCTION OF MEASURES
Exercise 4.16 Suppose µ
∗
is an outer measure. Show that if A
n
↑
A, then µ
∗
(A
n
) ↑ µ
∗
(A). Given an example to show that even if
µ
∗
is ﬁnite, A
n
↓ A does not necessarily imply µ
∗
(A
n
) ↓ µ
∗
(A).
Exercise 4.17 Suppose A is a Lebesgue measurable subset of R
and
B = ∪
x∈A
[x −1, x + 1].
Prove that B is Lebesgue measurable.
Chapter 5
Measurable functions
We are now ready to move from sets to functions.
5.1 Measurability
Suppose we have a measurable space (X, /).
Deﬁnition 5.1 A function f : X →R is measurable or / measur
able if ¦x : f(x) > a¦ ∈ / for all a ∈ R. A complexvalued function
is measurable if both its real and complex parts are measurable.
Example 5.2 Suppose f is realvalued and identically constant.
Then the set ¦x : f(x) > a¦ is either empty or all of X, so f is
measurable.
Example 5.3 Suppose f(x) = 1 if x ∈ A and 0 otherwise. Then
the set ¦x : f(x) > a¦ is either ∅, A, or X. Hence f is measurable
if and only if A is in /.
Example 5.4 Suppose X is the real line with the Borel σalgebra
and f(x) = x. Then ¦x : f(x) > a¦ = (a, ∞), and so f is measur
able.
Proposition 5.5 Suppose f is realvalued. The following condi
tions are equivalent.
37
38 CHAPTER 5. MEASURABLE FUNCTIONS
(1) ¦x : f(x) > a¦ ∈ / for all a ∈ R;
(2) ¦x : f(x) ≤ a¦ ∈ / for all a ∈ R;
(3) ¦x : f(x) < a¦ ∈ / for all a ∈ R;
(4) ¦x : f(x) ≥ a¦ ∈ / for all a ∈ R.
Proof. The equivalence of (1) and (2) and of (3) and (4) follow
from taking complements, e.g., ¦x : f(x) ≤ a¦ = ¦x : f(x) > a¦
c
.
If f is measurable, then
¦x : f(x) ≥ a¦ = ∩
∞
n=1
¦x : f(x) > a −1/n¦
shows that (4) holds if (1) does. If (4) holds, then (1) holds by
using the equality
¦x : f(x) > a¦ = ∪
∞
n=1
¦x : f(x) ≥ a + 1/n¦.
This completes the proof.
Proposition 5.6 If X is a metric space, / contains all the open
sets, and f : X →R is continuous, then f is measurable.
Proof. Note that ¦x : f(x) > a¦ = f
−1
((a, ∞)) is open, and hence
in /.
Proposition 5.7 Let c ∈ R. If f and g are measurable realvalued
functions, then so are f +g, −f, cf, fg, max(f, g), and min(f, g).
Proof. If f(x) +g(x) < a, then f(x) < a −g(x), and there exists
a rational r such that f(x) < r < a −g(x). Hence
¦x : f(x) +g(x) < a¦ = ∪
r∈Q
(¦x : f(x) < r¦ ∩ ¦x : g(x) < a −r¦).
This proves f +g is measurable.
Since ¦x : −f(x) > a¦ = ¦x : f(x) < −a¦, then −f is measur
able using Proposition 5.5.
If c > 0, then ¦x : cf(x) > a¦ = ¦x : f(x) > a/c¦ shows cf is
measurable. When c = 0, cf is measurable by Example 5.2. When
5.1. MEASURABILITY 39
c < 0, write cf = −([c[f), which is measurable by what we have
already proved.
f
2
is measurable since for a < 0, ¦x : f(x) > a¦ = X, while for
a ≥ 0,
¦x : f(x)
2
> a) = ¦x : f(x) >
√
a¦ ∪ ¦x : f(x) < −
√
a¦.
The measurability of fg follows since
fg =
1
2
[(f +g)
2
−f
2
−g
2
].
The equality
¦x : max(f(x), g(x)) > a¦ = ¦x : f(x) > a¦ ∪ ¦x : g(x) > a¦
shows max(f, g) is measurable, and the result for min(f, g) follows
from min(f, g) = −max(−f, −g).
Proposition 5.8 If f
i
is a measurable realvalued function for
each i, then so are sup
i
f
i
, inf
i
f
i
, limsup
i→∞
f
i
, and liminf
i→∞
f
i
.
Proof. The result will follow for limsup and liminf once we
have the result for the sup and inf by using the deﬁnitions since
limsup
i
f
i
= inf
j
sup
i≥j
f
j
and similarly for the liminf. We have
¦x : sup
i
f
i
> a¦ = ∩
∞
i=1
¦x : f
i
(x) > a¦, so sup
i
f
i
is measurable,
and the proof for inf f
i
is similar.
Deﬁnition 5.9 We say f = g almost everywhere, written f = g
a.e., if ¦x : f(x) ,= g(x)¦ has measure zero. Similarly, we say
f
i
→ f a.e. if the set of x where f
i
(x) does not converge to f(x)
has measure zero.
If X is a metric space, B is the Borel σalgebra, and f : X →R
is measurable with respect to B, we say f is Borel measurable. If
f : R → R is measurable with respect to the Lebesgue σalgebra,
we say f is Lebesgue measurable.
We saw in Proposition 5.6 that all continuous functions are
Borel measurable. The same is true for increasing functions on the
real line.
40 CHAPTER 5. MEASURABLE FUNCTIONS
Proposition 5.10 If f : R → R is monotone, then f is Borel
measurable.
Proof. Let us suppose f is increasing, for otherwise we look at
−f. Given a ∈ R, let x
0
= sup¦y : f(y) ≤ a¦. If f(x
0
) ≤ a, then
¦x : f(x) > a¦ = (x
0
, ∞). If f(x
0
) > a, then ¦x : f(x) > a¦ =
[x
0
, ∞). In either case ¦x : f(x) > a¦ is a Borel set.
Proposition 5.11 Let (X, /) be a measurable space and let f :
X →R be a / measurable function. If A is in the Borel σalgebra
on R, then f
−1
(A) ∈ /.
Proof. Let B be the Borel σalgebra on R and ( = ¦A ∈ B :
f
−1
(A) ∈ /¦. If A
1
, A
2
, . . . ∈ (, then since
f
−1
(∪
i
A
i
) = ∪
i
f
−1
(A
i
) ∈ /,
we have that ( is closed under countable unions. Similarly ( is
closed under countable intersections and complements, so ( is a
σalgebra. Since f is measurable, ( contains (a, ∞) for every real
a, hence ( contains the σalgebra generated by these intervals, that
is, ( contains B.
Example 5.12 Let us construct a set that is Lebesgue measur
able, but not Borel measurable. Recall the Lebesgue measurable
sets were constructed in Chapter 4 and include the completion of
the Borel σalgebra.
Let f be the CantorLebesgue function of Example 4.11 and
deﬁne
F(x) = inf¦y : f(y) ≥ x¦.
Although F is not continuous, observe that F is strictly increasing
(hence onetoone) and maps [0, 1] into C, the Cantor set. Since F
is increasing, F
−1
maps Borel measurable sets to Borel measurable
sets.
Let m be Lebesgue measure and let A be the nonmeasurable
set we constructed in Proposition 4.15. Let B = F(A). Since
F(A) ⊂ C and m(C) = 0, then F(A) is a null set, hence is Lebesgue
measurable. On the other hand, F(A) is not Borel measurable, be
cause if it were, then A = F
−1
(F(A)) would be Borel measurable,
a contradiction.
5.2. APPROXIMATION OF FUNCTIONS 41
5.2 Approximation of functions
Deﬁnition 5.13 Let (X, /) be a measurable space. If E ∈ /,
deﬁne the characteristic function of E by
χ
E
(x) =
_
1 x ∈ E;
0 x / ∈ E.
A simple function s is a function of the form
s(x) =
n
i=1
a
i
χ
E
i
(x)
for real numbers a
i
and measurable sets E
i
.
Proposition 5.14 Suppose f is a nonnegative and measurable
function. Then there exists a sequence of nonnegative measurable
simple functions increasing to f.
Proof. Let
E
ni
= ¦x : (i −1)/2
n
≤ f(x) < i/2
n
¦
and
F
n
= ¦x : f(x) ≥ n¦
for n = 1, 2, . . . and i = 1, 2, . . . , n2
n
. Then deﬁne
s
n
=
n2
n
i=1
i −1
2
n
χ
E
ni
+nχ
F
n
.
In words, f
n
(x) = n if f(x) ≥ n. If f(x) is between (i − 1)/2
n
and i/2
n
for i/2
n
≤ n, we let f
n
(x) = (i −1)/2
n
.
It is easy to see that s
n
has the desired properties.
5.3 Lusin’s theorem
The following theorem is known as Lusin’s theorem. It is very
pretty but usually other methods are better for solving problems.
42 CHAPTER 5. MEASURABLE FUNCTIONS
Example 5.16 will illustrate why this is a less useful theorem than
at ﬁrst glance.
We use m for Lebesgue measure. Recall that the support of a
function f is the closure of the set ¦x : f(x) ,= 0¦.
Theorem 5.15 Suppose f : [0, 1] → R is Borel measurable, m is
Lebesgue measure, and ε > 0. There exists a closed set F ⊂ [0, 1]
such that m([0, 1] − F) < ε and the restriction of f to F is a
continuous function on F.
This theorem can be loosely interpreted as saying every measurable
function is “almost continuous.”
Proof. First let us suppose that f = χ
A
, where A is a Borel
measurable subset of [0, 1]. By Proposition 4.14 we can ﬁnd E
closed and G open such that E ⊂ A ⊂ G and m(G − A) < ε/2
and m(A − E) < ε/2. Let δ = inf¦[x − y[ : x ∈ E, y ∈ G
c
¦. Since
E ⊂ A ⊂ [0, 1], E is compact and δ > 0. Letting
g(x) =
_
1 −
d(x, E)
δ
_
+
,
where y
+
= max(y, 0) and d(x, E) = inf¦[x − y[ : y ∈ E¦, we
see that g is continuous, takes values in [0, 1], is equal to 1 on
E, and equal to 0 on G
c
. Take F = (E ∪ G
c
) ∩ [0, 1]. Then
m([0, 1] −F) ≤ m(G−E) < ε, and f = g on F.
Next suppose f =
M
i=1
a
i
χ
A
i
is simple, where each A
i
is a
measurable subset of [0, 1] and each a
i
≥ 0. Choose F
i
closed such
that m([0, 1] − F
i
) < ε/M and χ
A
i
restricted to F
i
is continuous.
If we let F = ∩
M
i=1
F
i
, then F is closed, m([0, 1] − F) < ε, and f
restricted to F is continuous.
Now suppose f is nonnegative, bounded by K, and has support
in [0, 1]. Let
A
in
= ¦x : (i −1)/2
n
≤ f(x) < i/2
n
¦.
Then
f
n
(x) =
K2
n
+1
i=1
i
2
n
χ
A
in
(x)
are simple functions increasing to f. Note that
h
n
(x) = f
n+1
(x) −f
n
(x)
5.3. LUSIN’S THEOREM 43
is also a simple function and is bounded by 2
−n
. Choose F
0
closed
such that m([0, 1] −F
0
) < ε/2 and f
0
restricted to F
0
is continuous.
For n ≥ 1, choose F
n
closed such that m([0, 1] −F
n
) < ε/2
n+1
and
h
n
restricted to F
n
is continuous. Let F = ∩
∞
n=0
F
n
. Then F, being
the intersection of closed sets, will be closed, and m([0, 1] − F) ≤
∞
n=0
m([0, 1]−F
n
) < ε. On the set F, we have f
0
(x)+
∞
n=1
h
n
(x)
converges uniformly to f(x) because each h
n
is bounded by 2
−n
.
The uniform limit of continuous functions is continuous, hence f is
continuous on F.
If f ≥ 0, let B
K
= ¦x : f(x) ≤ K¦. Since f is everywhere ﬁnite,
B
K
↑ [0, 1] as K → ∞, hence m(B
K
) > 1 −ε/3 if K is suﬃciently
large. Choose D ⊂ B
K
such that D is closed and m(B
K
−D) < ε/3.
Now choose E ⊂ [0, 1] closed such that f χ
D
restricted to E is
continuous and m([0, 1] − E) < ε/3. Then F = D ∩ E is closed,
m([0, 1] −F) < ε, and f restricted to F is continuous.
Finally, for arbitrary measurable f write f = f
+
−f
−
and ﬁnd
F
+
and F
−
closed such that m([0, 1]−F
+
) < ε/2, m([0, 1]−F
−
) <
ε/2, and f
+
restricted to F
+
is continuous and similarly for f
−
.
Then F = F
+
∩ F
−
is the desired set.
Example 5.16 Suppose f = χ
B
, where B consists of the irra
tionals in [0, 1]. f is Borel measurable because [0, 1] −B is count
able, hence the union of countably many points, and thus the union
of countably many closed sets. Every point of [0, 1] is a point of
discontinuity of f because for any x ∈ [0, 1], there are both ratio
nals and irrationals in every neighborhood of x, hence f takes the
values 0 and 1 in every neighborhood of x.
Recall Example 4.12. f restricted to the set A there is identi
cally one, hence f restricted to A is a continuous function. A is
closed because it is equal to the interval [0, 1] minus the union of
open intervals.
This does not contradict Lusin’s theorem. No claim is made
that the function f is continuous at most points of [0, 1]. What is
asserted is that there is a closed set F with large measure so that
f restricted to F is continuous when viewed as a function from F
to R.
44 CHAPTER 5. MEASURABLE FUNCTIONS
5.4 Exercises
Exercise 5.1 Suppose (X, /) is a measurable space, f is a real
valued functions, and ¦x : f(x) > r¦ ∈ / for each rational number
r. Prove that f is measurable.
Exercise 5.2 Let f : (0, 1) → R be such that for every x ∈ (0, 1)
there exist r > 0 and a Borel measurable function g, both depend
ing on x, such that f and g agree on B(x, r). Prove that f is Borel
measurable.
Exercise 5.3 Suppose f
n
are measurable functions. Prove that
A = ¦x : lim
n→∞
f
n
(x) exists¦
is a measurable set.
Exercise 5.4 If f : R → R is Lebesgue measurable, prove that
there exists a Borel measurable function g such that f = g a.e.
Exercise 5.5 Give an example of a collection of measurable non
negative functions ¦f
α
¦
α∈A
such that if g is deﬁned by g(x) =
sup
α∈A
f
α
(x), then g is ﬁnite for all x but g is nonmeasurable. (A
is allowed to be uncountable.)
Exercise 5.6 Suppose f : X → R is Lebesgue measurable and
g : R →R is continuous. Prove that g ◦ f is Lebesgue measurable.
Is this true if g is Borel measurable instead of continuous? Is this
true if g is Lebesgue measurable instead of continuous?
Exercise 5.7 Suppose f : R → R is Borel measurable. Deﬁne /
to be the smallest σalgebra containing the sets ¦x : f(x) > a¦ for
every a ∈ R. Suppose g : R → R is measurable with respect to
/, which means that ¦x : g(x) > a¦ ∈ / for every a ∈ R. Prove
that there exists a Borel measurable function h : R →R such that
g = h ◦ f.
Exercise 5.8 One can show that there exist discontinuous real
valued functions f such that
f(x +y) = f(x) +f(y) (5.1)
5.4. EXERCISES 45
for all x, y ∈ R. (The construction uses Zorn’s lemma, which is
equivalent to the axiom of choice.) Prove that if f satisﬁes (5.1)
and in addition f is Lebesgue measurable, then f is continuous.
46 CHAPTER 5. MEASURABLE FUNCTIONS
Chapter 6
The Lebesgue integral
In this chapter we deﬁne the Lebesgue integral. We only give the
deﬁnition here and consider the properties of the Lebesgue integral
in later chapters.
6.1 Deﬁnitions
Deﬁnition 6.1 Let (X, /, µ) be a measure space. If
s =
n
i=1
a
i
χ
E
i
is a nonnegative measurable simple function, deﬁne the Lebesgue
integral of s to be
_
s dµ =
n
i=1
a
i
µ(E
i
). (6.1)
Here, if a
i
= 0 and µ(E
i
) = ∞, we use the convention that
a
i
µ(E
i
) = 0. If f ≥ 0 is a measurable function, deﬁne
_
f dµ = sup
_
_
s dµ : 0 ≤ s ≤ f, s simple
_
. (6.2)
Let f be measurable and let f
+
= max(f, 0) and f
−
= max(−f, 0).
Provided
_
f
+
dµ and
_
f
−
dµ are not both inﬁnite, deﬁne
_
f dµ =
_
f
+
dµ −
_
f
−
dµ. (6.3)
47
48 CHAPTER 6. THE LEBESGUE INTEGRAL
Finally, if f = u+iv is complexvalued and
_
([u[ +[v[) dµ is ﬁnite,
deﬁne
_
f dµ =
_
udµ +i
_
v dµ. (6.4)
A few remarks are in order. A function s might be written
as a simple function in more than one way. For example s =
χ
A∪B
= χ
A
+ χ
B
if A and B are disjoint. It is not hard to check
that the deﬁnition of
_
s dµ is unaﬀected by how s is written. If
s =
m
i=1
a
i
χ
A
i
=
n
j=1
b
j
χ
B
j
, then we need to show
m
i=1
a
i
µ(A
i
) =
n
j=1
b
j
µ(B
j
). (6.5)
We leave the proof of this to the reader as Exercise 6.1.
Secondly, if s is a simple function, one has to think a moment
to verify that the deﬁnition of
_
s dµ by means of (6.1) agrees with
its deﬁnition by means of (6.2).
Deﬁnition 6.2 If f is measurable and
_
[f[ dµ < ∞, we say f is
integrable.
The proof of the next proposition follows from the deﬁnitions.
Proposition 6.3 (1) If f is a realvalued measurable function with
a ≤ f(x) ≤ b for all x and µ(X) < ∞, then aµ(X) ≤
_
f dµ ≤
bµ(X);
(2) If f and g are measurable, realvalued, and integrable and
f(x) ≤ g(x) for all x, then
_
f dµ ≤
_
g dµ.
(3) If f is integrable, then
_
cf dµ = c
_
f dµ for all complex c.
(4) If µ(A) = 0 and f is integrable, then
_
fχ
A
dµ = 0.
The integral
_
fχ
A
dµ is often written
_
A
f dµ. Other nota
tion for the integral is to omit the µ and write
_
f if it is clear
which measure is being used, to write
_
f(x) µ(dx), or to write
_
f(x) dµ(x).
When we are integrating a function f with respect to Lebesgue
measure m, it is usual to write
_
f(x) dx for
_
f(x) m(dx) and to
6.2. EXERCISES 49
deﬁne
_
b
a
f(x) dx =
_
[a,b]
f(x) m(dx).
Proposition 6.4 If f is integrable,
¸
¸
¸
_
f
¸
¸
¸ ≤
_
[f[.
Proof. For the real case, this is easy. f ≤ [f[, so
_
f ≤
_
[f[. Also
−f ≤ [f[, so −
_
f ≤
_
[f[. Now combine these two facts.
For the complex case,
_
f is a complex number. If it is 0, the
inequality is trivial. If it is not, then
_
f = re
iθ
for some r and θ.
Then
¸
¸
¸
_
f
¸
¸
¸ = r = e
−iθ
_
f =
_
e
−iθ
f.
From the deﬁnition of
_
f when f is complex, we have Re (
_
f) =
_
Re (f). Since [
_
f[ is real, we have
¸
¸
¸
_
f
¸
¸
¸ = Re
_
_
e
−iθ
f
_
=
_
Re (e
−iθ
f) ≤
_
[f[.
We do not yet know that
_
(f +g) =
_
f +
_
g. We will see this
in Theorem 7.4.
6.2 Exercises
Exercise 6.1 Verify (6.5).
Exercise 6.2 Suppose f is nonnegative and measurable and µ is
σﬁnite. Show there exist simple functions s
n
increasing to f at
each point such that µ(¦x : s
n
(x) ,= 0¦) < ∞ for each n.
Exercise 6.3 Let f be a nonnegative measurable function. Prove
that
lim
n→∞
_
(f ∧ n) →
_
f.
50 CHAPTER 6. THE LEBESGUE INTEGRAL
Exercise 6.4 Let (X, /, µ) be a measure space and suppose µ is
σﬁnite. Suppose f is integrable. Prove that given ε there exists δ
such that
_
A
[f(x)[ µ(dx) < ε
whenever µ(A) < δ.
Exercise 6.5 Suppose µ(X) < ∞and f
n
is a sequence of bounded
realvalued measurable functions that converge to f uniformly.
Prove that
_
f
n
dµ →
_
f dµ.
This is sometimes called the bounded convergence theorem.
Exercise 6.6 If f
n
is a sequence of nonnegative integrable func
tions such that f
n
(x) decreases to f(x) for every x, prove that
_
f
n
dµ →
_
f dµ.
Exercise 6.7 Let (X, /, µ) be a measure space and suppose f is a
nonnegative, measurable function that is ﬁnite at each point of X,
but not necessarily integrable. Prove that there exists a continuous
increasing function g : [0, ∞) → [0, ∞) such that lim
x→∞
g(x) = ∞
and g ◦ f is integrable.
Chapter 7
Limit theorems
The main reason the Lebesgue integral is so much easier to work
with than the Riemann integral is that it behaves nicely when
taking limits. In this chapter we prove the monotone convergence
theorem, Fatou’s lemma, and the dominated convergence theorem.
We also prove that the Lebesgue integral is linear.
7.1 Monotone convergence theorem
One of the most important results concerning Lebesgue integration
is the monotone convergence theorem.
Theorem 7.1 Suppose f
n
is a sequence of nonnegative measur
able functions with f
1
(x) ≤ f
2
(x) ≤ for all x and with
lim
n→∞
f
n
(x) = f(x)
for all x. Then
_
f
n
dµ →
_
f dµ.
Proof. By Proposition 6.3(2),
_
f
n
is an increasing sequence of
real numbers. Let L be the limit. Since f
n
≤ f for all n, then
L ≤
_
f. We must show L ≥
_
f.
Let s =
m
i=1
a
i
χ
E
i
be any nonnegative simple function less
than or equal to f and let c ∈ (0, 1). Let A
n
= ¦x : f
n
(x) ≥ cs(x)¦.
51
52 CHAPTER 7. LIMIT THEOREMS
Since f
n
(x) increases to f(x) for each x and c < 1, then A
n
↑ X.
For each n,
_
f
n
≥
_
A
n
f
n
≥ c
_
A
n
s
n
= c
_
A
n
m
i=1
a
i
χ
E
i
= c
m
i=1
a
i
µ(E
i
∩ A
n
).
If we let n → ∞, by Proposition 3.5(3) the right hand side con
verges to
c
m
i=1
a
i
µ(E
i
) = c
_
s.
Therefore L ≥ c
_
s. Since c is arbitrary in the interval (0, 1), then
L ≥
_
s. Taking the supremum over all simple s ≤ f, we obtain
L ≥
_
f.
Example 7.2 Let X = [0, ∞) and f
n
(x) = −1/n for all x. Then
_
f
n
= −∞, but f
n
↑ f where f = 0 and
_
f = 0. The reason the
monotone convergence theorem does not apply here is that the f
n
are not nonnegative.
Example 7.3 Suppose f
n
= nχ
(0,1/n)
. Then f
n
≥ 0, f
n
→ 0 for
each x, but
_
f
n
= 1 does not converge to
_
0 = 0. The reason the
monotone convergence theorem does not apply here is that the f
n
do not increase to f for each x.
7.2 Linearity of the integral
Once we have the monotone convergence theorem, we can prove
that the Lebesgue integral is linear.
Theorem 7.4 If f and g are nonnegative and measurable or if f
and g are integrable, then
_
(f +g) dµ =
_
f dµ +
_
g dµ.
7.2. LINEARITY OF THE INTEGRAL 53
Proof. First suppose f and g are nonnegative and simple, say,
f =
m
i=1
a
i
χ
A
i
and g =
n
j=1
b
j
χ
B
j
. Without loss of generality
we may assume that A
1
, . . . , A
m
are pairwise disjoint and that
B
1
, . . . , B
n
are pairwise disjoint. Since a
i
= 0 and b
j
= 0 are
permissible, we may also assume ∪
m
i=1
A
i
= X = ∪
n
j=1
B
j
. Then
f +g =
m
i=1
n
j=1
(a
i
+b
j
)χ
A
i
∩B
j
,
and we have
_
(f +g) =
m
i=1
n
j=1
(a
i
+b
j
)µ(A
i
∩ B
j
)
=
m
i=1
n
j=1
a
i
µ(A
i
∩ B
j
) +
m
i=1
n
j=1
b
j
µ(A
i
∩ B
j
)
=
m
i=1
a
i
µ(A
i
) +
n
j=1
b
j
µ(B
j
)
=
_
f +
_
g.
Thus the theorem holds in this case.
Next suppose f and g are nonnegative. Take s
n
simple and
increasing to f and t
n
simple and increasing to g. Then s
n
+ t
n
are simple functions increasing to f +g, so the result follows from
the monotone convergence theorem and
_
(f +g) = lim
n→∞
_
(s
n
+t
n
) = lim
n→∞
_
s
n
+ lim
n→∞
t
n
=
_
f +
_
g.
Suppose now that f and g are realvalued and integrable but
take both positive and negative values. Since
_
[f +g[ ≤
_
([f[ +[g[) =
_
[f[ +
_
[g[ < ∞,
then f +g is integrable. Write
(f +g)
+
−(f +g)
−
= f +g = f
+
−f
−
+g
+
−g
−
,
so that
(f +g)
+
+f
−
+g
−
= f
+
+g
+
+ (f +g)
−
.
54 CHAPTER 7. LIMIT THEOREMS
Using the result for nonnegative functions,
_
(f +g)
+
+
_
f
−
+
_
g
−
=
_
f
+
+
_
g
+
+
_
(f +g)
−
.
Rearranging,
_
(f +g) =
_
(f +g)
+
−
_
(f +g)
−
=
_
f
+
−
_
f
−
+
_
g
+
−
_
g
−
=
_
f +
_
g.
If f and g are complexvalued, apply the above to the real and
imaginary parts.
Proposition 7.5 Suppose f
n
are nonnegative measurable func
tions. Then
_
∞
n=1
f
n
=
∞
n=1
_
f
n
.
Proof. Let F
N
=
N
n=1
f
n
. Since 0 ≤ F
n
(x) ↑
∞
n=1
f
n
(x), we
can write
_
∞
n=1
f
n
=
_
lim
N→∞
N
n=1
f
n
=
_
lim
N→∞
F
N
= lim
N→∞
_
F
N
(7.1)
= lim
N→∞
N
n=1
_
f
n
=
∞
n=1
_
f
n
,
using the monotone convergence theorem and the linearity of the
integral.
7.3 Fatou’s lemma
The next theorem is known as Fatou’s lemma.
7.4. DOMINATED CONVERGENCE THEOREM 55
Theorem 7.6 Suppose the f
n
are nonnegative and measurable.
Then
_
liminf
n→∞
f
n
≤ liminf
n→∞
_
f
n
.
Proof. Let g
n
= inf
i≥n
f
i
. Then the g
n
are nonnegative and
g
n
increases to liminf
n
f
n
. Clearly g
n
≤ f
i
for each i ≥ n, so
_
g
n
≤
_
f
i
. Therefore
_
g
n
≤ inf
i≥n
_
f
i
. (7.2)
If we take the limit in (7.2) as n → ∞, on the left hand side we
obtain
_
liminf
n
f
n
by the monotone convergence theorem, while
on the right hand side we obtain liminf
n
_
f
n
.
A typical use of Fatou’s lemma is the following. Suppose we
have f
n
→ f and sup
n
_
[f
n
[ ≤ K < ∞. Then [f
n
[ → [f[, and by
Fatou’s lemma,
_
[f[ ≤ K.
7.4 Dominated convergence theorem
Another very important theorem is the dominated convergence the
orem.
Theorem 7.7 Suppose that f
n
are measurable realvalued func
tions and f
n
(x) → f(x) for each x. Suppose there exists a non
negative integrable function g such that [f
n
(x)[ ≤ g(x) for all x.
Then
lim
n→∞
_
f
n
dµ →
_
f dµ.
Proof. Since f
n
+g ≥ 0, by Fatou’s lemma,
_
f +
_
g =
_
(f +g) ≤ liminf
n→∞
_
(f
n
+g) = liminf
n→∞
f
n
+
_
g.
Since g is integrable,
_
f ≤ liminf
n→∞
_
f
n
. (7.3)
56 CHAPTER 7. LIMIT THEOREMS
Similarly, g −f
n
≥ 0, so
_
g−
_
f =
_
(g−f) ≤ liminf
n→∞
_
(g−f
n
) =
_
g+liminf
n→∞
_
(−f
n
),
and hence
−
_
f ≤ liminf
n→∞
_
(−f
n
) = −limsup
n→∞
_
f
n
.
Therefore
_
f ≥ limsup
n→∞
_
f
n
,
which with (7.3) proves the theorem.
Exercise 7.1 asks you to prove a version of the dominated con
vergence theorem for complexvalued functions.
Example 7.3 is an example where the limit of the integrals is not
the integral of the limit because there is no dominating function g.
If in the monotone convergence theorem or dominated conver
gence theorem we have only f
n
(x) → f(x) almost everywhere, the
conclusion still holds. For example, if the f
n
are measurable, non
negative, and f
n
↑ f a.e., let A = ¦x : f
n
(x) → f(x)¦. Then
f
n
χ
A
(x) ↑ fχ
A
(x) for each x. Since A
c
has measure 0, we see from
Proposition 6.3(4) and the monotone convergence theorem that
lim
n
_
f
n
= lim
n
_
f
n
χ
A
=
_
fχ
A
=
_
f.
7.5 Exercises
Exercise 7.1 State and prove a version of the dominated conver
gence theorem for complexvalued functions.
Exercise 7.2 The following generalized dominated convergence
theorem is often useful. Suppose f
n
, g
n
, f, and g are integrable,
f
n
→ f a.e., g
n
→ g a.e., [f
n
[ ≤ g
n
for each n, and
_
g
n
→
_
g.
Prove that
_
f
n
→
_
f.
Exercise 7.3 Give an example of a sequence of nonnegative func
tions f
n
tending to 0 pointwise such that
_
f
n
→ 0, but there is no
integrable function g such that f
n
≤ g for all n.
7.5. EXERCISES 57
Exercise 7.4 Suppose (X, /, µ) is a measure space, each f
n
is
integrable and nonnegative, f
n
→ f a.e., and
_
f
n
→
_
f. Prove
that for each A ∈ /
_
A
f
n
dµ →
_
A
f dµ.
Exercise 7.5 Suppose f
n
and f are integrable, f
n
→ f a.e., and
_
[f
n
[ →
_
[f[. Prove that
_
[f
n
−f[ → 0.
Exercise 7.6 Suppose f : R → R is integrable, a ∈ R, and we
deﬁne
F(x) =
_
x
a
f(y) dy.
Show that F is a continuous function.
Exercise 7.7 Let f
n
be a sequence of nonnegative Lebesgue mea
surable functions on R. Is it necessarily true that
limsup
n→∞
_
f
n
dx ≤
_
limsup
n→∞
f
n
dx?
If not, give a counterexample.
Exercise 7.8 Find the limit
lim
n→∞
_
n
0
_
1 +
x
n
_
−n
log(2 + cos(x/n)) dx
and justify your reasoning.
Exercise 7.9 Find the limit
lim
n→∞
_
n
0
_
1 −
x
n
_
n
log(2 + cos(x/n)) dx
and justify your reasoning.
Exercise 7.10 Prove that the limit exists and ﬁnd its value:
lim
n→∞
_
1
0
1 +nx
2
(1 +x
2
)
n
log(2 + cos(x/n)) dx.
58 CHAPTER 7. LIMIT THEOREMS
Exercise 7.11 Prove the limit exists and determine its value:
lim
n→∞
_
∞
0
ne
−nx
sin(1/x) dx.
Exercise 7.12 Let g : R → R be integrable and let f : R → R be
bounded, measurable, and continuous at 1. Prove that
lim
n→∞
_
n
−n
f
_
1 +
x
n
2
_
g(x) dx
exists and determine its value.
Exercise 7.13 Suppose µ(X) < ∞, f
n
converges to f uniformly,
and each f
n
is integrable. Prove that f is integrable and
_
f
n
→
_
f. Is the condition µ(X) < ∞ necessary?
Exercise 7.14 Prove that
∞
k=1
1
(p +k)
2
= −
_
1
0
x
p
1 −x
log xdx
for p > 0.
Exercise 7.15 Let ¦f
n
¦ be a sequence of realvalued functions on
[0, 1] that is uniformly bounded.
(1) Show that if A is a Borel subset of [0, 1], then there exists a
subsequence n
j
such that
_
A
f
n
j
(x) dx converges.
(2) Show that if ¦A
i
¦ is a countable collection of Borel subsets of
[0, 1], then there exists a subsequence n
j
such that
_
A
i
f
n
j
(x) dx
converges for each i.
(3) Show that there exists a subsequence n
j
such that
_
A
f
n
j
(x) dx
converges for each Borel subset A of [0, 1].
Exercise 7.16 Let (X, /, µ) be a measure space. A family of
measurable functions ¦f
n
¦ is uniformly integrable if given ε there
exists M such that
_
{x:f
n
(x)>M}
[f
n
(x)[ dµ < ε
for each n. The sequence is uniformly absolutely continuous if given
ε there exists δ such that
¸
¸
¸
_
A
f
n
dµ
¸
¸
¸ < ε
7.5. EXERCISES 59
for each n if µ(A) < δ.
Suppose µ is a ﬁnite measure. Prove that ¦f
n
¦ is uniformly
integrable if and only if sup
n
_
[f
n
[ dµ < ∞ and ¦f
n
¦ is uniformly
absolutely continuous.
Exercise 7.17 The following is known as the Vitali convergence
theorem. Suppose µ is a ﬁnite measure, f
n
→ f a.e., and ¦f
n
¦ is
uniformly integrable. Prove that
_
f
n
→
_
f.
Exercise 7.18 Suppose µ is a ﬁnite measure, f
n
→ f a.e., each
f
n
is integrable, f is integrable, and
_
f
n
→
_
f. Prove that ¦f
n
¦
is uniformly integrable.
Exercise 7.19 Suppose µ is a ﬁnite measure and for some ε > 0
sup
n
_
[f
n
[
1+ε
dµ < ∞.
Prove that ¦f
n
¦ is uniformly integrable.
Exercise 7.20 Suppose f
n
is a uniformly integrable sequence of
functions deﬁned on [0, 1]. Prove that there is a subsequence n
j
such that
_
1
0
f
n
j
g dx converges whenever g is a realvalued bounded
measurable function.
Exercise 7.21 Suppose µ
n
is a sequence of measures on (X, /)
such that µ
n
(X) = 1 for all n and µ
n
(A) converges as n → ∞ for
each A ∈ /. Call the limit µ(A).
(1) Prove that µ is a measure.
(2) Prove that
_
f dµ
n
→
_
f dµ whenever f is bounded and mea
surable.
(3) Prove that
_
f dµ ≤ liminf
n→∞
_
f dµ
n
whenever f is nonnegative and measurable.
Exercise 7.22 Let (X, /, µ) be a measure space and let f be non
negative and integrable. Deﬁne ν on / by
ν(A) =
_
A
f dµ.
60 CHAPTER 7. LIMIT THEOREMS
(1) Prove that ν is a measure.
(2) Prove that if g is integrable with respect to ν, then fg is inte
grable with respect to µ and
_
g dν =
_
fg dµ.
Exercise 7.23 Suppose µ and ν are positive measures on the Borel
σalgebra on [0, 1] such that
_
f dµ =
_
f dν whenever f is real
valued and continuous on [0, 1]. Prove that µ = ν.
Exercise 7.24 Let B be the Borel σalgebra on [0, 1]. Let µ
n
be a
sequence of ﬁnite measures on ([0, 1], B) and let µ be another ﬁnite
measure on ([0, 1], B). Suppose µ
n
([0, 1]) → µ([0, 1]). Prove that
the following are equivalent:
(1)
_
f dµ
n
→
_
f dµ whenever f is a continuous realvalued func
tion on [0, 1];
(2) limsup
n→∞
µ
n
(F) ≤ µ(F) for all closed subsets F of [0, 1];
(3) liminf
n→∞
µ
n
(G) ≥ µ(G) for all open subsets G of [0, 1];
(4) lim
n→∞
µ
n
(A) = µ(A) whenever A is a Borel subset of [0, 1]
such that µ(∂A) = 0, where ∂A = A−A
o
is the boundary of A;
(5) lim
n→∞
µ
n
([0, x]) = µ([0, x]) for every x such that µ(¦x¦) = 0.
Exercise 7.25 Let B be the Borel σalgebra on [0, 1]. Suppose
µ
n
are ﬁnite measures on ([0, 1], B) such that
_
f dµ
n
→
_
1
0
f dx
whenever f is a realvalued continuous function on [0, 1]. Suppose
that g is a bounded measurable function such that the set of dis
continuities of g has measure 0. Prove that
_
g dµ
n
→
_
1
0
g dx.
Exercise 7.26 Let B be the Borel σalgebra on [0, 1]. Let µ
n
be a
sequence of ﬁnite measures on ([0, 1], B) with sup
n
µ
n
([0, 1]) < ∞.
Deﬁne α
n
(x) = µ
n
([0, x]).
(1) If r is a rational in [0, 1], prove that there exists a subsequence
¦n
j
¦ such that α
n
j
(r) converges.
(2) Prove that there exists a subsequence ¦n
j
¦ such that α
n
(r)
converges for every rational in [0, 1].
(3) Let α(r) = lim
n→∞
α
n
(r) for r rational and deﬁne
α(x) = lim
r→x,r>x,r∈Q
α(r).
7.5. EXERCISES 61
This means, since clearly α(r) ≤ α(s) if r < s, that
α(x) = inf¦α(r) : r > x, r ∈ Q¦.
Let µ be the LebesgueStieltjes measure associated with α. Prove
that
_
f dµ
n
→
_
f dµ
whenever f is a continuous realvalued function on [0, 1].
62 CHAPTER 7. LIMIT THEOREMS
Chapter 8
Properties of Lebesgue
integrals
We present some propositions which imply that a function is zero
a.e. and we give an approximation result.
8.1 Criteria for a function to be zero a.e.
The following two propositions are very useful.
Proposition 8.1 Suppose f is realvalued and measurable and for
every measurable set A we have
_
A
f dµ = 0. Then f = 0 almost
everywhere.
Proof. Let A = ¦x : f(x) > ε¦. Then
0 =
_
A
f ≥
_
A
ε = εµ(A)
since fχ
A
≥ εχ
A
. Hence µ(A) = 0. We use this argument for
ε = 1/n and n = 1, 2, . . . to conclude
µ¦x : f(x) > 0¦ = µ(∪
∞
n=1
¦x : f(x) > 1/n¦)
≤
∞
n=1
µ(¦x : f(x) > 1/n¦) = 0.
63
64 CHAPTER 8. PROPERTIES OF LEBESGUE INTEGRALS
Similarly µ¦x : f(x) < 0¦ = 0.
Proposition 8.2 Suppose f is measurable and nonnegative and
_
f dµ = 0. Then f = 0 almost everywhere.
Proof. If f is not equal to 0 almost everywhere, there exists an n
such that µ(A
n
) > 0 where A
n
= ¦x : f(x) > 1/n¦. But since f is
nonnegative,
0 =
_
f ≥
_
A
n
f ≥
1
n
µ(A
n
),
a contradiction.
As a corollary to Proposition 8.1 we have the following.
Corollary 8.3 Let m be Lebesgue measure and a ∈ R. Suppose
f : R → R is integrable and
_
x
a
f(y) dy = 0 for all x. Then f = 0
a.e.
Proof. For any interval [c, d],
_
d
c
f =
_
d
a
f −
_
c
a
f = 0.
By linearity, if G is the ﬁnite union of disjoint intervals, then
_
G
f =
0. By dominated convergence and Proposition 1.5,
_
G
f = 0 for any
open set G. Again by dominated convergence, if G
n
are open sets
decreasing to H, then
_
H
f = lim
n
_
G
n
f = 0.
If E is any Borel measurable set, Proposition 4.14 tells us that
there exists a sequence G
n
of open sets that decrease to a set H
where H diﬀers from E by a null set. Then
_
E
f =
_
fχ
E
=
_
fχ
H
=
_
H
f = 0.
This with Proposition 8.1 implies f is zero a.e.
8.2. AN APPROXIMATION RESULT 65
8.2 An approximation result
We give a result on approximating a function on R by continuous
functions.
Theorem 8.4 Suppose f is a Borel measurable realvalued inte
grable function on R. Let ε > 0. Then there exists a continuous
function g with compact support such that
_
[f −g[ < ε.
Proof. If we write f = f
+
− f
−
, it is enough to ﬁnd continuous
functions g
1
and g
2
with compact support such that
_
[f
+
−g
1
[ <
ε/2 and
_
[f
−
− g
2
[ < ε/2 and to let g = g
1
− g
2
. Hence we may
assume f ≥ 0.
By monotone convergence
_
f χ
[−n,n]
increases to
_
f, so by
taking n large enough, the diﬀerence of the integrals will be less
than ε/2. If we ﬁnd g continuous with compact support such that
_
[f χ
[−n,n]
− g[ < ε/2, then
_
[f − g[ < ε. Therefore we may in
addition assume that f is 0 outside some bounded interval.
Suppose f = χ
A
, where A is a bounded Borel measurable set.
We can choose G open and F closed such that F ⊂ A ⊂ G and
m(G−F) < ε by Proposition 4.14. Without loss of generality, we
may assume G is also a bounded set. Since F is compact, there is
a minimum distance between F and G
c
, say, δ. Let
g(x) =
_
1 −
dist (x, F)
δ
_
+
.
Then g is continuous, 0 ≤ g ≤ 1, g is 1 on F, g is 0 on G
c
, and g
has compact support. We have
[g −χ
A
[ ≤ χ
G
−χ
F
,
so
_
[g −χ
A
[ ≤
_
(χ
G
−χ
F
) = m(G−F) < ε.
Thus our result holds for characteristic functions of bounded sets.
If f =
p
i=1
a
i
χ
A
i
, where each A
i
is contained in a bounded
interval and each a
i
> 0, and we ﬁnd g
i
continuous with compact
66 CHAPTER 8. PROPERTIES OF LEBESGUE INTEGRALS
support such that
_
[χ
A
i
− g
i
[ < ε/a
i
p, then g =
p
i=1
a
i
g
i
will
be the desired function. Thus our theorem holds for nonnegative
simple functions with compact support.
If f is nonnegative and has compact support, we can ﬁnd simple
functions s
m
supported in a bounded interval increasing to f whose
integrals increase to
_
f. Let s
m
be a simple function such that
s
m
≤ f and
_
s
m
≥
_
f − ε/2. We choose continuous g with
compact support such that
_
[s
m
− g[ < ε/2 using the preceding
paragraphs, and then
_
[f −g[ < ε.
The method of proof, where one proves a result for characteristic
functions, then simple functions, then nonnegative functions, and
then ﬁnally integrable functions, is very common.
8.3 Exercises
Exercise 8.1 This exercise gives a change of variables formula in
two simple cases. Show that if f is an integrable function on the
reals and a is a nonzero real number, then
_
R
f(x +a) dx =
_
R
f(x) dx
and
_
R
f(ax) dx = a
−1
_
R
f(x) dx.
Exercise 8.2 Let (X, /, µ) be a σﬁnite measure space. Suppose
f is nonnegative and integrable. Prove that if ε > 0, there exists
A ∈ / such that µ(A) < ∞ and
ε +
_
A
f dµ >
_
f dµ.
Exercise 8.3 Suppose A is a Borel measurable subset of [0, 1], m
is Lebesgue measure, and ε ∈ (0, 1). Prove that there exists a
continuous function f : [0, 1] →R such that 0 ≤ f ≤ 1 and
m(¦x : f(x) ,= χ
A
(x)¦) < ε.
8.3. EXERCISES 67
Exercise 8.4 Suppose f is a nonnegative integrable function on
a measure space (X, /, µ). Prove that
lim
t→∞
tµ(¦x : f(x) ≥ t¦) = 0.
Exercise 8.5 Find a nonnegative function f on [0, 1] such that
lim
t→∞
tm(¦x : f(x) ≥ t¦) = 0
but f is not integrable, where m is Lebesgue measure.
Exercise 8.6 Suppose µ is a ﬁnite measure. Prove that a measur
able nonnegative function f is integrable if and only if
∞
n=1
µ(¦x : f(x) ≥ n¦) < ∞.
Exercise 8.7 Let µ be a measure, not necessarily σﬁnite, and
suppose f is realvalued and integrable with respect to µ. Prove
that A = ¦x : f(x) ,= 0¦ has σﬁnite measure, that is, there exists
F
n
↑ A such that µ(F
n
) < ∞ for each n.
Exercise 8.8 Recall that a function f : R →R is convex if
f(λx + (1 −λ)y) ≤ λf(x) + (1 −λ)f(y)
whenever x < y ∈ R and λ ∈ [0, 1].
(1) Prove that if f is convex and x ∈ R, there exists a real number
c such that f(y) ≥ f(x) +c(y −x) for all y ∈ R. Graphically, this
says that the graph of f lies above the line with slope c that passes
through the point (x, f(x)).
(2) Let (X, /, µ) be a measure space, suppose µ(X) = 1, and let
f : R →R be convex. Let g : X →R be integrable. Prove Jensen’s
inequality:
f
_
_
g dµ
_
≤
_
X
f ◦ g dµ.
Exercise 8.9 Suppose f is a realvalued function on R such that
f
_
_
1
0
g(x) dx
_
≤
_
1
0
f(g(x)) dx
whenever g is bounded and measurable. Prove that f is convex.
68 CHAPTER 8. PROPERTIES OF LEBESGUE INTEGRALS
Exercise 8.10 Suppose g : [0, 1] →R is bounded and measurable
and
_
1
0
f(x)g(x) dx = 0
whenever f is continuous and
_
1
0
f(x) dx = 0. Prove that g is equal
to a constant a.e.
Chapter 9
Riemann integrals
We compare the Lebesgue integral and the Riemann integral. We
show that the Riemann integral of a function exists if and only if
the set of discontinuities of the function have Lebesgue measure
zero, and in that case the Riemann integral and Lebesgue integral
agree.
9.1 Comparison with the Lebesgue in
tegral
We only consider bounded measurable functions from [a, b] into R.
If we are looking at the Lebesgue integral, we write
_
f, while, tem
porarily, if we are looking at the Riemann integral, we write R(f).
Recall that the Riemann integral on [a, b] is deﬁned as follows: if
P = ¦x
0
, x
1
, . . . , x
n
¦ with x
0
= a and x
n
= b is a partition of [a, b],
let
U(P, f) =
n
i=1
[ sup
x
i−1
≤x≤x
i
f(x)] (x
i
−x
i−1
)
and
L(P, f) =
n
i=1
[ inf
x
i−1
≤x≤x
i
f(x)] (x
i
−x
i−1
).
Set
R(f) = inf¦U(P, f) : P is a partition¦
69
70 CHAPTER 9. RIEMANN INTEGRALS
and
R(f) = sup¦L(P, f) : P is a partition¦.
The Riemann integral exists if R(f) = R(f), and the common value
is the Riemann integral, which we denote R(f).
Theorem 9.1 A bounded Borel measurable realvalued function f
on [a, b] is Riemann integrable if and only if the set of points at
which f is discontinuous has Lebesgue measure 0, and in that case,
the Riemann integral is equal in value to the Lebesgue integral.
Proof. Step 1. First we show that if f is Riemann integrable, then
f is continuous a.e. and R(f) =
_
f. If P is a partition, deﬁne
T
P
(x) =
n
i=1
[ sup
x
i−1
≤y≤x
i
f(y)]χ
[x
i−1
,x
i
)
(x),
and
S
P
(x) =
n
i=1
[ inf
x
i−1
≤y≤x
i
f(y)]χ
[x
i−1
,x
i
)
(x).
We observe that
_
T
P
= U(P, f) and
_
S
P
= L(P, f).
If f is Riemann integrable, there exists a sequence of parti
tions Q
i
such that U(Q
i
, f) ↓ R(f) and a sequence Q
i
such that
L(Q
i
, f) ↑ R(f). It is not hard to check that adding points to a par
tition increases L and decreases U, so if we let P
i
= ∪
j≤i
(Q
j
∪Q
j
),
then P
i
is an increasing sequence of partitions, U(P
i
, f) ↓ R(f) and
L(P
i
, f) ↑ R(f). We see also that T
P
i
(x) decreases at each point,
say, to T(x), and S
P
i
(x) increases at each point, say, to S(x). Also
T(x) ≥ f(x) ≥ S(x). By dominated convergence (recall that f is
bounded)
_
(T −S) = lim
i→∞
_
(T
P
i
−S
P
i
) = lim
i→∞
(U(P
i
, f) −L(P
i
, f)) = 0.
We conclude T = S = f a.e. If x is not in the null set where
T(x) ,= S(x) nor in ∪
i
P
i
, which is countable and hence of Lebesgue
measure 0, then T
P
i
(x) ↓ f(x) and S
P
i
(x) ↑ f(x). We claim that
f is continuous at such x. To prove the claim, given ε, choose i
large enough so that T
P
i
(x) −S
P
i
(x) < ε and then choose δ small
enough so that (x − δ, x + δ) is contained in the subinterval of P
i
that contains x. Finally, since
R(f) = lim
i→∞
U(P
i
, f) = lim
i→∞
_
T
P
i
=
_
f
9.2. EXERCISES 71
by the dominated convergence theorem, we see that the Riemann
integral and Lebesgue integral agree.
Step 2. Now suppose that f is continuous a.e. Let ε > 0. Let P
i
be the partition where we divide [a, b] into 2
i
equal parts. If x is
not in the null set where f is discontinuous, nor in ∪
∞
i=1
P
i
, then
T
P
i
(x) ↓ f(x) and S
P
i
(x) ↑ f(x). By dominated convergence,
U(P
i
, f) =
_
T
P
i
→
_
f
and
L(P
i
, f) =
_
S
P
i
→
_
f.
This does it.
Example 9.2 Let [a, b] = [0, 1] and f = χ
A
, where A is the set of
irrational numbers in [0, 1]. If x ∈ [0, 1], every neighborhood of x
contains both rational and irrational points, so f is continuous at
no point of [0, 1]. Therefore f is not Riemann integrable.
Example 9.3 Deﬁne f(x) on [0, 1] to be 0 if x is irrational and to
be 1/q if x is rational and equals p/q when in reduced form. f is
discontinuous at every rational. If x is irrational and ε > 0, there
are only ﬁnitely many rationals r for which f(r) ≥ ε, so taking
δ less than the distance from x to any of this ﬁnite collection of
rationals shows that [f(y) − f(x)[ < ε if [y − x[ < δ. Hence f is
continuous at x. Therefore the set of discontinuities is a countable
set, hence of measure 0, hence f is Riemann integrable.
9.2 Exercises
Exercise 9.1 Find a measurable function f : [0, 1] →R such that
R(f) ,=
_
1
0
f(x) dx and R(f) ,=
_
1
0
f(x) dx.
Exercise 9.2 Find a function f : (0, 1] → R that is continuous,
is not Lebesgue integrable, but where the improper Riemann inte
gral exists. Thus we want f such that
_
1
0
[f(x)[ m(dx) = ∞ but
lim
a→0+
R(fχ
[a,1]
) exists.
72 CHAPTER 9. RIEMANN INTEGRALS
Exercise 9.3 Suppose f : [0, 1] → R is integrable, f is bounded
on (a, 1] for each a > 0, and the improper Riemann integral
lim
a→0+
R(fχ
(a,1]
)
exists. Show that the limit is equal to
_
1
0
f(x) dx.
Exercise 9.4 Divide [a, b] into 2
n
equal subintervals and pick a
point x
i
out of each subinterval. Let µ
n
be the measure deﬁned by
µ
n
(dx) = 2
−n
2
n
i=1
δ
x
i
(dx),
where δ
y
is point mass at y. Note that if f is a bounded measurable
realvalued function on [a, b], then
_
b
a
f(x) µ
n
(dx) =
2
n
i=1
f(x
i
)2
−n
(9.1)
is a Riemann sum approximation to R(f).
(1) Prove that µ
n
([0, x]) → m([0, x]) for every x ∈ [0, 1]. Conclude
by Exercise 7.24 that
_
f dµ
n
→
_
1
0
f dx whenever f is continuous.
(2) Use Exercise 7.25 to see that if f is a bounded and measurable
function on [a, b] whose set of discontinuities has measure 0, then
the Riemann sum approximation of f given in (9.1) converges to
the Lebesgue integral of f. This provides an alternative proof of
Step 2 of Theorem 9.1.
Exercise 9.5 Let f be a bounded, realvalued, and measurable
function. Prove that if
f = lim
δ→0
sup
y−x<δ,a≤y≤b
f(y),
then f = T a.e., where we use the notation of Theorem 9.1. Con
clude f is Lebesgue measurable.
Exercise 9.6 Deﬁne f = lim
δ→0
inf
y−x<δ,a≤y≤b
f(y) and let f
be deﬁned as in Exercise 9.5.
(1) Suppose that the set of discontinuities of a bounded realvalued
9.2. EXERCISES 73
measurable function f has positive Lebesgue measure. Prove that
there exists ε > 0 such that if
A
ε
= ¦x ∈ [a, b] : f(x) −f(x) > ε¦,
then m(A
ε
) > 0.
(2) Prove that U(P, f) − L(P, f) > εm(A
ε
) for every partition P
on [a, b], using the notation of Theorem 9.1. Conclude that f is
not Riemann integrable. This provides another proof of Step 1 of
Theorem 9.1.
Exercise 9.7 A realvalued function on a metric space is lower
semicontinuous if ¦x : f(x) > a¦ is open whenever a ∈ R and
upper semicontinuous if ¦x : f(x) < a¦ is open whenever a ∈ R.
(1) Prove that if f
n
is a sequence of realvalued continuous functions
increasing to f, then f is lower semicontinuous.
(2) Find a bounded lower semicontinuous function f : [0, 1] → R
such that f is continuous everywhere except at x = 1/2.
(3) Find a bounded lower semicontinuous realvalued function f
deﬁned on [0, 1] such that the set of discontinuities of f is equal to
the set of rationals in [0, 1].
(4) Find a bounded lower semicontinuous function f : [0, 1] → R
such that the set of discontinuities of f has positive measure.
(5) Does there exist a bounded lower semicontinuous function f :
[0, 1] →R such that f is discontinuous a.e.?
Exercise 9.8 Find a sequence f
n
of continuous functions mapping
[0, 1] into [0, 1] such that the f
n
increase to a bounded function f
which is not Riemann integrable. Such an example shows there
is no monotone convergence theorem or dominated convergence
theorem for Riemann integrals.
Exercise 9.9 Let M > 0 and let B be the σalgebra on [−M, M]
2
generated by the collection of sets of the form [a, b] [c, d] with
−M ≤ a ≤ b ≤ M and −M ≤ c ≤ d ≤ M. Suppose µ is a measure
on ([−M, M]
2
, B) such that
µ([a, b] [c, d]) = (b −a)(d −c).
(We will construct such a measure µ in Chapter 11.) Prove that
if f is continuous with support in [−M, M]
2
, then the Lebesgue
integral of f with respect to µ is equal to the double (Riemann)
integral of f and the two multiple (Riemann) integrals of f.
74 CHAPTER 9. RIEMANN INTEGRALS
Chapter 10
Types of convergence
There are various ways in which a sequence of functions f
n
can
converge, and we compare some of them.
10.1 Deﬁnitions and examples
Deﬁnition 10.1 If µ is a measure, we say a sequence of mea
surable functions f
n
converges almost everywhere to f and write
f
n
→ f a.e. if there is a set of measure 0 such that for x not in
this set we have f
n
(x) → f(x).
We say f
n
converges in measure to f if for each ε > 0
µ(¦x : [f
n
(x) −f(x)[ > ε¦) → 0
as n → ∞.
Let 1 ≤ p < ∞. We say f
n
converges in L
p
to f if
_
[f
n
−f[
p
dµ → 0
as n → ∞.
Proposition 10.2 Suppose µ is a ﬁnite measure.
(1) If f
n
→ f, a.e., then f
n
converges to f in measure.
(2) If f
n
→ f in measure, there is a subsequence n
j
such that
f
n
j
→ f, a.e.
75
76 CHAPTER 10. TYPES OF CONVERGENCE
Proof. Let ε > 0 and suppose f
n
→ f a.e. If
A
n
= ¦x : [f
n
(x) −f(x)[ > ε¦,
then χ
A
n
→ 0 a.e., and by dominated convergence,
µ(A
n
) =
_
χ
A
n
(x) µ(dx) → 0.
This proves (1).
To prove (2), suppose f
n
→ f in measure, let n
1
= 1, and
choose n
j
> n
j−1
by induction so that
µ(¦x : [f
n
j
(x) −f(x)[ > 1/j¦) ≤ 2
−j
.
Let A
j
= ¦x : [f
n
j
(x) − f(x)[ > 1/j¦. Then µ(A
j
) ≤ 2
−j
, and if
we set
A = ∩
∞
k=1
∪
∞
j=k
A
j
,
then by Proposition 3.5
µ(A) = lim
k→∞
µ(∪
∞
j=k
A
j
) ≤ lim
k→∞
∞
j=k
µ(A
j
) ≤ lim
k→∞
2
−k+1
= 0.
Therefore A has measure 0. If x / ∈ A, then x / ∈ ∪
∞
j=k
A
j
for some
k, and so [f
n
j
(x) −f(x)[ ≤ 1/j for j ≥ k. This implies f
n
j
→ f on
A
c
.
Example 10.3 Part (1) of the above proposition is not true if
µ(X) = ∞. To see this, let X = R and let f
n
= χ
(n,n+1)
. We have
f
n
→ 0 a.e., but f
n
does not converge in measure.
The next proposition compares convergence in L
p
to conver
gence in measure. Before we prove this, we prove an easy prelimi
nary result known as Chebyshev’s inequality.
Lemma 10.4 If 1 ≤ p < ∞, then
µ(¦x : [f(x)[ ≥ a¦) ≤
_
[f[
p
a
p
.
10.1. DEFINITIONS AND EXAMPLES 77
Proof. Let A = ¦x : [f(x)[ ≥ a¦. Since χ
A
≤ [f[
p
χ
A
/a
p
, we have
µ(A) ≤
_
A
[f[
p
a
p
dµ ≤
1
a
p
_
[f[
p
dµ.
This is what we wanted.
Proposition 10.5 If f
n
converges to f in L
p
, then it converges in
measure.
Proof. If ε > 0, by Chebyshev’s inequality
µ(¦x : [f
n
(x) −f(x)[ > ε¦) = µ(¦x : [f
n
(x) −f(x)[
p
> ε
p
¦)
≤
_
[f
n
−f[
p
ε
p
→ 0
as required.
Example 10.6 Let f
n
= n
2
χ
(0,1/n)
on [0, 1] with Lebesgue mea
sure. This gives an example where f
n
converges to 0 a.e. and in
measure, but does not converge in L
p
for any p ≥ 1.
Example 10.7 We give an example where f
n
→ f in measure and
in L
p
, but not almost everywhere. Let S = ¦e
iθ
: 0 ≤ θ < 2π¦ be
the unit circle in the complex plane and deﬁne
µ(A) = m(¦θ ∈ [0, 2π) : e
iθ
∈ A¦)
to be arc length measure on S, where m is Lebesgue measure on
[0, 2π).
Let X = S and let f
n
(x) = χ
F
n
(x), where
F
n
=
_
e
iθ
:
n
j=1
1
j
≤ θ ≤
n+1
j=1
1
j
_
.
Let f(e
iθ
) = 0 for all θ.
Then µ(F
n
) ≤ 1/(n + 1) → 0, so f
n
→ f in measure. Also,
since f
n
is either 1 or 0,
_
[f
n
−f[
p
dµ =
_
χ
F
n
dµ = µ(F
n
) → 0.
78 CHAPTER 10. TYPES OF CONVERGENCE
But because
∞
j=1
1/j = ∞, each point of S is in inﬁnitely many
F
n
, and each point of S is in S − F
n
for inﬁnitely many n, so f
n
does not converge to f at any point.
The F
n
are arcs whose length tends to 0, but such that ∪
n≥m
F
n
contains S for each m.
The following is known as Egorov’s theorem.
Theorem 10.8 Suppose µ is a ﬁnite measure, ε > 0, and f
n
→ f
a.e. Then there exists a measurable set A such that µ(A) < ε and
f
n
→ f uniformly on A
c
.
This type of convergence is sometimes known as almost uniform
convergence. Egorov’s theorem is not as useful for solving problems
as one might expect, and students have a tendency to try to use it
when other methods work much better.
Proof. Let
A
nk
= ∪
∞
m=n
¦x : [f
m
(x) −f(x)[ > 1/k¦.
For ﬁxed k, A
nk
decreases as n increases. The intersection ∩
n
A
nk
has measure 0 because for almost every x, [f
m
(x) −f(x)[ ≤ 1/k if
m is suﬃciently large. Therefore µ(A
nk
) → 0 as n → ∞. We can
thus ﬁnd an integer n
k
such that µ(A
n
k
k
) < ε2
−k
. Let
A = ∪
∞
k=1
A
n
k
k
.
Hence µ(A) < ε. If x / ∈ A, then x / ∈ A
n
k
k
, and so [f
n
(x) −f(x)[ ≤
1/k if n ≥ n
k
. Thus f
n
→ f uniformly on A
c
.
10.2 Exercises
Exercise 10.1 Suppose that f
n
is a sequence that is Cauchy in
measure. This means that given ε and a > 0, there exists N such
that if m, n ≥ N, then
µ(¦x : [f
n
(x) −f
m
(x)[ > a¦) < ε.
Prove that f
n
converges in measure.
10.2. EXERCISES 79
Exercise 10.2 Suppose µ(X) < ∞. Deﬁne
d(f, g) =
_
[f −g[
1 +[f −g[
dµ.
Prove that d is a metric on the space of measurable functions,
except for the fact that d(f, g) = 0 only implies that f = g a.e.,
not necessarily everywhere. Prove that f
n
→ f in measure if and
only if d(f
n
, f) → 0.
Exercise 10.3 Prove that if f
n
→ f in measure and each f
n
is
nonnegative, then
_
f ≤ liminf
n→∞
_
f
n
.
Exercise 10.4 Prove that if A
n
is measurable and µ(A
n
) < ∞
for each n and χ
A
n
converges to f in measure, then there exists a
measurable set A such that f = χ
A
a.e.
Exercise 10.5 Suppose for each ε there exists a measurable set F
such that µ(F
c
) < ε and f
n
converges to f uniformly on F. Prove
that f
n
converges to f a.e.
Exercise 10.6 Suppose that f
n
and f are measurable functions
such that for each ε > 0 we have
∞
n=1
µ(¦x : [f
n
(x) −f(x)[ > ε¦) < ∞.
Prove that f
n
→ f a.e.
Exercise 10.7 Let f
n
be a sequence of measurable functions and
deﬁne
g
n
(x) = sup
m≥n
[f
m
(x) −f
n
(x)[.
Prove that if g
n
converges in measure to 0, then f
n
converges a.e.
Exercise 10.8 If (X, /, µ) is a measure space and f
n
is a sequence
of realvalued measurable functions such that
_
f
n
g dµ converges
to 0 for every integrable g, is it necessarily true that f
n
converges
to 0 in measure? If not, give a counterexample.
80 CHAPTER 10. TYPES OF CONVERGENCE
Chapter 11
Product measures
We have deﬁned Lebesgue measure on the line. Now we give a
method for constructing measures on the plane, in ndimensional
Euclidean spaces, and many other product spaces. The main theo
rem, the Fubini theorem, which allows one to interchange the order
of integration, is one of the most important theorems in real anal
ysis.
11.1 Product σalgebras
Suppose (X, /, µ) and (Y, B, ν) are two measure spaces and suppose
also that µ and ν are σﬁnite measures. A measurable rectangle is
a set of the form AB, where A ∈ / and B ∈ B.
Let (
0
be the collection of ﬁnite unions of disjoint measurable
rectangles. Thus every element of (
0
is of the form ∪
n
i=1
(A
i
B
i
),
where A
i
∈ /, B
i
∈ B, and if i ,= j, then (A
i
B
i
) ∩(A
j
B
j
) = ∅.
Since (AB)
c
= (AB
c
) ∪ (A
c
Y ) and the intersection of two
measurable rectangles is a measurable rectangle, it is easy to check
that (
0
is an algebra of sets. We deﬁne the product σalgebra
/B = σ((
0
).
If E ∈ /B, we deﬁne the xsection of E by
s
x
(E) = ¦y ∈ Y : (x, y) ∈ E¦
81
82 CHAPTER 11. PRODUCT MEASURES
and similarly deﬁne the ysection:
t
y
(E) = ¦x : (x, y) ∈ E¦.
Given a function f : XY →R that is /B measurable, for each
x and y we deﬁne S
x
f : Y →R and T
y
f : X →R by
S
x
f(y) = f(x, y), T
y
f(x) = f(x, y).
Lemma 11.1 (1) If E ∈ / B, then s
x
(E) ∈ B for each x and
t
y
(E) ∈ / for each y.
(2) If f is /B measurable, then S
x
f is B measurable for each x
and T
y
f is / measurable for each y.
Proof. (1) Let ( be the collection of sets in / B for which
s
x
(E) ∈ B for each x. We will show that ( is a σalgebra containing
the measurable rectangles, and hence is all of /B.
If E = A B, then s
x
(E) is equal to B if x ∈ A and equal to
∅ if x / ∈ A. Hence s
x
(E) ∈ B for each x when E is a measurable
rectangle.
If E ∈ (, then y ∈ s
x
(E
c
) if and only if (x, y) ∈ E
c
, which
happens if and only if y / ∈ s
x
(E). Therefore s
x
(E
c
) = (s
x
(E))
c
, and
( is closed under the operation of taking complements. Similarly,
it is easy to see that s
x
(∪
∞
i=1
E
i
) = ∪
∞
i=1
s
x
(E
i
), and so ( is closed
under the operation of countable unions.
Therefore ( is a σalgebra containing the measurable rectangles,
and hence is equal to /B. The argument for t
y
(E) is the same.
(2) Fix x. If f = χ
E
for E ∈ / B, note that S
x
f(y) =
χ
s
x
(E)
(y), which is B measurable. By linearity, S
x
f is B measurable
when f is a simple function. If f is nonnegative, take / B
measurable simple functions r
n
increasing to f, and since S
x
r
n
↑
S
x
f, then S
x
f is B measurable. Writing f = f
+
− f
−
and using
linearity again shows that S
x
f is B measurable. The argument for
T
y
f is the same.
Let E ∈ /B and let
h(x) = ν(s
x
(E)), k(y) = µ(t
y
(E)).
11.1. PRODUCT σALGEBRAS 83
Proposition 11.2 (1) h is / measurable and k is B measurable.
(2) We have
_
h(x) µ(dx) =
_
k(y) ν(dy). (11.1)
Since χ
s
x
(E)
(y) = S
x
χ
E
(y) for all x and y, (11.1) could be
written as
_
_
_
S
x
χ
E
(y) ν(dy)
_
µ(dx) =
_
_
_
T
y
χ
E
(x) µ(dx)
_
ν(dy).
We will usually write this as
_ _
χ
E
(x, y) ν(dy) µ(dx) =
_ _
χ
E
(x, y) µ(dx) ν(dy).
Proof. First suppose µ and ν are ﬁnite measures. Let ( be the
collection of sets in /B for which (1) and (2) hold. We will prove
that ( contains (
0
and is a monotone class. This will prove that (
is the smallest σalgebra containing (
0
and hence is equal to /B.
If E = AB, with A ∈ / and B ∈ B, then h(x) = χ
A
(x)ν(B),
which is / measurable, and
_
h(x) µ(dx) = µ(A)ν(B). Similarly,
k(y) = µ(A)χ
B
(y) is B measurable and
_
k(y) ν(dy) = µ(A)ν(B).
Therefore (1) and (2) hold for measurable rectangles.
If E = ∪
n
i=1
E
i
, where each E
i
is a measurable rectangle and the
E
i
are disjoint, then s
x
(E) = ∪
n
i=1
s
x
(E
i
), and since the s
x
(E
i
) are
disjoint, then
h(x) = ν(s
x
(E)) = ν(∪
n
i=1
s
x
(E
i
)) =
n
i=1
ν(s
x
(E
i
)).
This shows that h is / measurable, since it is the sum of / mea
surable functions. Similarly k(y) is B measurable. If we let h
i
(x) =
ν(s
x
(E
i
)) and deﬁne k
i
(y) similarly, then
_
h
i
(x) µ(dx) =
_
k
i
(y) ν(dy)
by the preceding paragraph, and then (2) holds for E by linearity.
Therefore ( contains (
0
.
Suppose E
n
↑ E and each E
n
∈ (. If we let h
n
(x) = ν(s
x
(E
n
))
and let k
n
(y) = µ(t
n
(E
n
)), then h
n
↑ h and k
n
↑ k. Therefore h is
84 CHAPTER 11. PRODUCT MEASURES
/ measurable and k is B measurable. We have (11.1) holding when
h and k are replaced by h
n
and k
n
, resp. We let n → ∞ and use
the monotone convergence theorem to see that (11.1) holds with h
and k.
If E
n
↓ E with each E
n
∈ (, almost the same argument shows
that h and k are measurable with respect to / and B, and that
(11.1) holds. The only diﬀerence is that we use the dominated con
vergence theorem in place of the monotone convergence theorem.
This is where we need µ and ν to be ﬁnite measures.
We have shown ( is a monotone class containing (
0
. By the
monotone class theorem (Theorem 2.10), ( is equal to σ((
0
), which
is /B.
Finally suppose µ and ν are σﬁnite. Then there exist F
i
↑ X
and G
i
↑ Y such that each F
i
is / measurable and has ﬁnite µ
measure and each G
i
is B measurable and has ﬁnite ν measure.
Let µ
i
(A) = µ(A ∩ F
i
) for each A ∈ / and ν
i
(A) = ν(A ∩ G
i
) for
each B ∈ B. Let h
i
(x) = ν
i
(s
x
(E)) = ν(s
x
(E) ∩ G
i
) and similarly
deﬁne k
i
(y). By what we have proved above, h
i
is / measurable,
k
i
is B measurable, and (11.1) holds if we replace h and k by h
i
and
k
i
, resp. Now h
i
↑ h and k
i
↑ k, which proves the measurability of
h and k. Applying the monotone convergence theorem proves that
(11.1) holds with h and k.
We now deﬁne µ ν by
µ ν(E) =
_
h(x) µ(dx) =
_
k(y) ν(dy). (11.2)
Clearly µ ν(∅) = 0. If E
1
, . . . , E
n
are disjoint and in /B and
E = ∪
n
i=1
E
i
, then we saw in the proof of Proposition 11.2 that
ν(s
x
(E)) =
n
i=1
ν(s
x
(E
i
)). We conclude that
µ ν(E) =
_
ν(s
x
(E)) µ(dx) =
n
i=1
_
ν(s
x
(E
i
)) µ(dx)
=
n
i=1
µ ν(E
i
),
or µν is ﬁnitely additive. If E
n
↑ E with each E
n
∈ /B and we
let h
n
(x) = ν(s
x
(E
n
)), then h
n
↑ h, and by monotone convergence,
µ ν(E
n
) ↑ µ ν(E). Therefore µ ν is a measure.
11.2. THE FUBINI THEOREM 85
Note that if E = AB is a measurable rectangle, then h(x) =
χ
A
(x)ν(B) and so
µ ν(AB) = µ(A)ν(B),
which is what it should be.
11.2 The Fubini theorem
The main result of this chapter is the Fubini theorem, which allows
one to interchange the order of integration. This is sometimes
called the FubiniTonelli theorem.
Theorem 11.3 Suppose f : XY →R is measurable with respect
to /B. If either
(a) f is nonnegative, or
(b)
_
[f(x, y)[ d(µ ν)(x, y) < ∞,
then
(1) for each x, the function y → f(x, y) is measurable with respect
to B;
(2) for each y, the function x → f(x, y) is measurable with respect
to /;
(3) the function g(x) =
_
f(x, y) ν(dy) is measurable with respect
to /;
(4) the function h(y) =
_
f(x, y) µ(dx) is measurable with respect
to B;
(5) we have
_
f(x, y) d(µ ν)(x, y) =
_
_
_
f(x, y) dµ(x)
_
dν(y) (11.3)
=
_
_
_
f(x, y) dν(y)
_
µ(dx).
The last integral in (11.3) should be interpreted as
_
_
_
S
x
f(y) ν(dy)
_
µ(dx)
and similarly for the second integral in (11.3). Since no confusion
results, most often the brackets are omitted in (11.3).
Proof. If f is the characteristic function of a set in / B, then
(1)–(5) are merely a restatement of Lemma 11.1 and Proposition
86 CHAPTER 11. PRODUCT MEASURES
11.2. By linearity, (1)–(5) hold if f is a simple function. Since the
increasing limit of measurable functions is measurable, then writing
a nonnegative function as the increasing limit of simple functions
and using the monotone convergence theorem, we see that (1)–(5)
hold when f is nonnegative. In the case where
_
[f[ d(µν) < ∞,
writing f = f
+
− f
−
and using linearity proves (1)–(5) for this
case, too.
Observe that if we know
_ _
[f(x, y)[ µ(dx) ν(dy) < ∞,
then since [f(x, y)[ is nonnegative the Fubini theorem tells us that
_
[f(x, y)[ d(µ ν) =
_ _
[f(x, y)[ µ(dx) ν(dy) < ∞
We can then apply the Fubini theorem again to conclude
_
f(x, y) d(µ ν) =
_ _
f(x, y) dµdν =
_ _
f(x, y) dν dµ.
Thus in the hypotheses of the Fubini theorem, we could as well
assume
_ _
[f(x, y)[ dµdν < ∞ or
_ _
[f(x, y)[ dν dµ < ∞.
When f is measurable with respect to /B, we sometimes say
that f is jointly measurable.
Even when (X, /, µ) and (Y, B, ν) are complete, it will not be
the case in general that (X Y, / B, µ ν) is complete. For
example, let (X, /, µ) = (Y, B, ν) be Lebesgue measure on [0, 1]
with the Lebesgue σalgebra. Let A be a nonmeasurable set in
[0, 1] and let E = A¦1/2¦. Then E is not a measurable set with
respect to /B, or else A = t
1/2
(E) would be in / by Lemma 11.1.
On the other hand, E ⊂ [0, 1]¦1/2¦, which has zero measure with
respect to µ ν, so E is a null set.
One can take the completion of (X Y, /B, µ ν) without
great diﬃculty. See [4] for details.
There is no diﬃculty extending the Fubini theorem to the prod
uct of n measures. If we have µ
1
, . . . , µ
n
all equal to m, Lebesgue
measure on R with the Lebesgue σalgebra /, then the completion
of (R
n
, / /, m m) is called ndimensional Lebesgue
measure.
For a general change of variables theorem, see [1].
11.3. EXAMPLES 87
11.3 Examples
We give two examples to show that the hypotheses of the Fubini
theorem are necessary.
Example 11.4 Let X = Y = [0, 1] and µ and ν both being
Lebesgue measure. Let g
i
be continuous functions with support
in (1/(i + 1), 1/i) such that
_
1
0
g
i
(x) dx = 1, i = 1, 2, . . .. Let
f(x, y) =
∞
i=1
[g
i
(x) −g
i+1
(x)]g
i
(y).
For each point (x, y) at most two terms in the sum are nonzero, so
the sum is actually a ﬁnite one. If we ﬁrst integrate with respect
to y, we get
_
1
0
f(x, y) dy =
∞
i=1
[g
i
(x) −g
i+1
(x)].
This is a telescoping series, and sums to g
1
(x). Therefore
_
1
0
_
1
0
f(x, y) dy dx =
_
1
0
g
1
(x) dx = 1.
On the other hand, integrating ﬁrst with respect to x gives 0, so
_
1
0
_
1
0
f(x, y) dxdy = 0.
This doesn’t contradict the Fubini theorem because
_
1
0
_
1
0
[f(x, y)[ dxdy = ∞.
Example 11.5 For this example, you have to take on faith a bit of
set theory. There exists a set X together with a partial order “≤”
such that X is uncountable but for any y ∈ X, the set ¦x ∈ X :
x ≤ y¦ is countable. An example is to let X be the set of countable
ordinals. Deﬁne µ on X by µ(A) = 0 if A is countable and 1 if
A is uncountable. The σalgebra is the collection of subsets A of
X such that either A or A
c
is countable. Deﬁne f on X X by
f(x, y) = 1 if x ≤ y and zero otherwise. Then
_ _
f(x, y) dy dx = 1
but
_ _
f(x, y) dxdy = 0. The reason there is no contradiction is
that f is not measurable with respect to the product σalgebra.
88 CHAPTER 11. PRODUCT MEASURES
11.4 Exercises
Exercise 11.1 State and prove a version of the Fubini theorem
for complexvalued functions.
Exercise 11.2 Let (X, /) and (Y, B) be two measurable spaces
and let f ≥ 0 be measurable with respect to / B. Let g(x) =
sup
y∈Y
f(x, y) and suppose g(x) < ∞ for each x. Is g necessarily
measurable with respect to /? If not, ﬁnd a counterexample.
Exercise 11.3 Prove the equality
_
∞
−∞
[f(x)[ dx =
_
∞
0
m(¦x : [f(x)[ ≥ t¦) dt,
where m is Lebesgue measure.
Exercise 11.4 Let A be a Lebesgue measurable subset of [0, 1]
2
with m
2
(A) = 1, where m
2
is twodimensional Lebesgue measure.
Show that for almost every x ∈ [0, 1] (with respect to one di
mensional Lebesgue measure) the set s
x
(A) has onedimensional
Lebesgue measure one.
Exercise 11.5 Let f : [0, 1]
2
→R be such that for every x ∈ [0, 1]
the function y → f(x, y) is Lebesgue measurable on [0, 1] and for
every y ∈ [0, 1] the function x → f(x, y) is continuous on [0, 1].
Prove that f is measurable with respect to the the completion of
the product σalgebra / / on [0, 1]
2
. Here / is the Lebesgue
σalgebra on [0, 1].
Exercise 11.6 Suppose f is realvalued and integrable with re
spect to twodimensional Lebesgue measure on [0, 1]
2
and
_
a
0
_
b
0
f(x, y) dy dx = 0
for all a ∈ [0, 1] and b ∈ [0, 1]. Prove that f = 0 a.e.
Exercise 11.7 Prove that
_
1
0
_
1
0
x
2
−y
2
(x
2
+y
2
)
2
log(4 + sin x) dy dx
=
_
1
0
_
1
0
x
2
−y
2
(x
2
+y
2
)
2
log(4 + sin x) dxdy.
11.4. EXERCISES 89
Exercise 11.8 Let X = Y = [0, 1] and let B be the Borel σ
algebra. Let m be Lebesgue measure and µ counting measure on
[0, 1].
(1) If D = ¦(x, y) : x = y¦, show that D is measurable with respect
to B B.
(2) Show that
_
X
_
Y
χ
D
(x, y) µ(dy) m(dx) ,=
_
Y
_
X
χ
D
(x, y) m(dx) µ(dy).
Why does this not contradict the Fubini theorem?
Exercise 11.9 Let X = Y = R and let B be the Borel σalgebra.
Deﬁne
f(x, y) =
_
¸
_
¸
_
1, x ≥ 0 and x ≤ y < x + 1;
−1, x ≥ 0 and x + 1 ≤ y < x + 2;
0, otherwise.
Show that
_ _
f(x, y) dy dx ,=
_ _
f(x, y) dxdy.
Why does this not contradict the Fubini theorem?
Exercise 11.10 Find a realvalued function f that is integrable
on [0, 1]
2
such that
_
a
0
_
1
0
f(x, y) dy dx = 0,
_
1
0
_
b
0
f(x, y) dy dx = 0
for every a, b ∈ [0, 1], but f is not zero almost everywhere with
respect to 2dimensional Lebesgue measure.
Exercise 11.11 Let µ be a ﬁnite measure on R and let α(x) =
µ((−∞, x]). Show
_
[α(x +c) −α(x)] dx = cµ(R).
Exercise 11.12 Use
1
x
=
_
∞
0
e
−xy
dy
90 CHAPTER 11. PRODUCT MEASURES
and the Fubini theorem to calculate
_
b
0
_
∞
0
e
−xy
sin xdy dx
two diﬀerent ways. Then prove that
lim
b→∞
_
b
0
sin x
x
dx =
π
2
.
Recall that
_
e
au
sin udu =
e
au
(a sin u −cos u)
1 +a
2
+C.
Exercise 11.13 Let X = ¦1, 2, . . .¦ and let µ be counting measure
on X. Deﬁne f : X X →R by
f(x, y) =
_
¸
_
¸
_
1, x = y;
−1, x = y + 1;
0, otherwise.
Show that
_
X
_
X
f(x, y) µ(dx) µ(dy) ,=
_
X
_
X
f(x, y) µ(dy) µ(dx).
Why is this not a contradiction to the Fubini theorem?
Exercise 11.14 Let ¦a
n
¦ and ¦r
n
¦ be two sequences of real num
bers such that
∞
n=1
[a
n
[ < ∞. Prove that
∞
n=1
a
n
_
[x −r
n
[
converges absolutely for almost every x ∈ R.
Exercise 11.15 Let (X, /, µ) and (Y, B, ν) be measure spaces.
Prove that if λ is a measure on /B such that
λ(AB) = µ(A)ν(B)
whenever A ∈ / and B ∈ B, then λ = µ ν on /B.
11.4. EXERCISES 91
Exercise 11.16 Let S be the unit circle ¦e
iθ
: 0 ≤ θ < 2π¦ and
deﬁne a measure µ on S by µ(A) = m(¦θ : e
iθ
∈ A¦), where m is
Lebesgue measure on [0, 2π). Let m
2
be twodimensional Brownian
motion. Show that if A is a Borel subset of S and R > 0, then
m
2
(¦re
iθ
: 0 < r < R, e
iθ
∈ A¦) = µ(A)R
2
/2.
Exercise 11.17 Use Exercise 11.16 to prove that if f is a continu
ous realvalued function with support in the ball B(0, R) = ¦(x, y) :
x
2
+y
2
< R
2
¦, then
_ _
B(0,R)
f(x, y) dy dx =
_
2π
0
_
R
0
f(r cos θ, r sin θ) r dr dθ.
Exercise 11.18 Prove that
_
∞
0
e
−x
2
/2
dx =
_
π/2
by ﬁlling in the missing steps and making rigorous the following.
If I =
_
∞
0
e
−x
2
/2
dx, then
I
2
=
_
∞
0
_
∞
0
e
−(x
2
+y
2
)/2
dy dx =
_
π/2
0
_
∞
0
e
−r
2
/2
r dr dθ = π/2.
Exercise 11.19 If M = (M
ij
)
n
i,j=1
is a n n matrix and x =
(x
1
, . . . , x
n
) ∈ R
n
, deﬁne Mx to be the element of R
n
whose i
th
coordinate is
n
j=1
M
ij
x
j
. (This is just the usual matrix multipli
cation of a nn matrix and a n1 matrix.) If A is a Borel subset
of R
n
, let M(A) = ¦Mx : x ∈ A¦.
(1) If c ∈ R and
M
ij
=
_
¸
_
¸
_
c, i = j = 1;
1, i = j ,= 1;
0, i ,= j;
show
m
n
(M(A)) = [c[m
n
(A) = [ det M[m
n
(A),
where we use m
n
for ndimensional Lebesgue measure. (Multipli
cation by M multiplies the ﬁrst coordinate by c.)
(2) If 1 ≤ k ≤ n and
M
ij
=
_
¸
¸
¸
_
¸
¸
¸
_
1, i = 1 and j = k;
1, j = 1 and i = k;
1, i = j and neither equals k;
0, otherwise;
92 CHAPTER 11. PRODUCT MEASURES
show
m
n
(M(A)) = m
n
(A) = [ det M[m
n
(A).
(Multiplication by M interchanges the i
th
and k
th
coordinates.)
(3) If c ∈ R and
M
ij
=
_
¸
_
¸
_
1, i = j;
c, i = 1, j = 2;
0, otherwise,
show
m
n
(M(A)) = m
n
(A) = [ det M[m
n
(A).
(Multiplication by M replaces x
1
by x
1
+cx
2
.)
(4) Since every n n matrix can be written as the product of ma
trices each of which has the form given in (1), (2), or (3), conclude
that if M is any n n matrix, then
m
n
(M(A)) = [ det M[m
n
(A).
(5) If M is an orthogonal matrix, so that M times its transpose
is the identity, show m
n
(M(A)) = m
n
(A). (Multiplication by an
orthogonal matrix is a rotation of R
n
.)
Chapter 12
Signed measures
Signed measures have the countable additivity property of mea
sures, but are allowed to take negative as well as positive val
ues. We will see shortly that an example of a signed measure is
ν(A) =
_
A
f dµ, where f is integrable and takes both positive and
negative values.
12.1 Positive and negative sets
Deﬁnition 12.1 Let / be a σalgebra. A signed measure is a
function µ : / → (−∞, ∞] such that µ(∅) = 0 and µ(∪
∞
i=1
A
i
) =
∞
i=1
µ(A
i
) whenever the A
i
are pairwise disjoint and all the A
i
are in /.
When we want to emphasize that a measure is deﬁned as in
Deﬁnition 3.1 and only takes nonnegative values, we refer to it as
a positive measure.
Deﬁnition 12.2 Let µ be a signed measure. A set A ∈ / is called
a positive set for µ if µ(B) ≥ 0 whenever B ⊂ A and B ∈ /. We
say A ∈ / is a negative set if µ(B) ≤ 0 whenever B ⊂ A and
B ∈ /. A null set A is one where µ(B) = 0 whenever B ⊂ A and
B ∈ /.
93
94 CHAPTER 12. SIGNED MEASURES
Note that if µ is a signed measure, then
µ(∪
∞
i=1
A
i
) = lim
n→∞
µ(∪
n
i=1
A
i
).
The proof is the same as in the case of positive measures.
Example 12.3 Suppose m is Lebesgue measure and
µ(A) =
_
A
f dm
for some integrable f. If we let P = ¦x : f(x) ≥ 0¦, then P
is easily seen to be a positive set, and if N = ¦x : f(x) < 0¦,
then N is a negative set. The Hahn decomposition which we give
below is a decomposition of our space (in this case R) into the
positive and negative sets P and N. This decomposition is unique,
except that C = ¦x : f(x) = 0¦ could be included in N instead
of P, or apportioned partially to P and partially to N. Note,
however, that C is a null set. The Jordan decomposition below is
a decomposition of µ into µ
+
and µ
−
, where µ
+
(A) =
_
A
f
+
dm
and µ
−
(A) =
_
A
f
−
dm.
Proposition 12.4 Let µ be a signed measure which takes values
in (−∞, ∞]. Let E be measurable with µ(E) < 0. Then there exists
a measurable subset F of E that is a negative set with µ(F) < 0.
Proof. If E is a negative set, we are done. If not, there exists a
measurable subset with positive measure. Let n
1
be the smallest
positive integer such that there exists E
1
⊂ E with µ(E
1
) ≥ 1/n
1
.
We then deﬁne pairwise disjoint measurable sets E
2
, E
3
, . . . by in
duction as follows. Let k ≥ 2 and suppose E
1
, . . . , E
k−1
are pair
wise disjoint measurable sets with µ(E
i
) > 0 for i = 1, . . . , k − 1.
If F
k
= E −(E
1
∪ ∪ E
k−1
) is a negative set, then
µ(F
k
) = µ(E) −
k−1
i=1
µ(E
i
) ≤ µ(E) < 0
and F
k
is the desired set F. If F
k
is not a negative set, let n
k
be
the smallest positive integer such that there exists E
k
⊂ F
k
with
E
k
measurable and µ(E
k
) ≥ 1/n
k
.
12.2. HAHN DECOMPOSITION THEOREM 95
We stop the construction if there exists k such that F
k
is a
negative set with µ(F
k
) < 0. If not, we continue and let F =
∩
k
F
k
= E −(∪
k
E
k
). Since 0 > µ(E) > −∞ and µ(E
k
) ≥ 0, then
µ(E) = µ(F) +
∞
k=1
µ(E
k
).
Then µ(F) ≤ µ(E) < 0, so the sum converges.
It remains to show that F is a negative set. Suppose G ⊂ F is
measurable with µ(G) > 0. Then µ(G) ≥ 1/N for some N. But
this contradicts the construction, since for some k, n
k
> N, and
we would have chosen the set G instead of the set E
k
at stage k.
Therefore F must be a negative set.
12.2 Hahn decomposition theorem
Recall that we write A∆B for (A−B) ∪(B −A). The following is
known as the Hahn decomposition.
Theorem 12.5 (1) Let µ be a signed measure taking values in
(−∞, ∞]. There exist disjoint measurable sets E and F in / whose
union is X and such that E is a negative set and F is a positive
set.
(2) If E
and F
are another such pair, then E∆E
= F∆F
is
a null set with respect to µ.
(3) If µ is not a positive measure, then µ(E) < 0. If −µ is not
a positive measure, then µ(F) > 0.
Proof. (1) Let L = inf¦µ(A) : A is a negative set¦. Choose
negative sets A
n
such that µ(A
n
) → L. Let E = ∪
∞
n=1
A
n
. Let
B
n
= A
n
−(B
1
∪ ∪B
n−1
) for each n. Since A
n
is a negative set,
so is each B
n
. Also, the B
n
are disjoint and ∪
n
B
n
= ∪
n
A
n
= E.
If C ⊂ E, then
µ(C) = lim
n→∞
µ(C ∩ (∪
n
i=1
B
i
)) = lim
n→∞
n
i=1
µ(C ∩ B
i
) ≤ 0.
Thus E is a negative set.
96 CHAPTER 12. SIGNED MEASURES
Since E is a negative set,
µ(E) = µ(A
n
) +µ(E −A
n
) ≤ µ(A
n
).
Letting n → ∞, we obtain µ(E) = L.
Let F = E
c
. If F were not a positive set, there would exist
B ⊂ F with µ(B) < 0. By Proposition 12.4 there exists a negative
set C contained in B with µ(C) < 0. But then E ∪ C would be a
negative set with µ(E ∪ C) < µ(E) = L, a contradiction.
(2) To prove uniqueness, if E
, F
are another such pair of sets
and A ⊂ E−E
⊂ E, then µ(A) ≤ 0. But A ⊂ E−E
= F
−F ⊂
F
, so µ(A) ≥ 0. Therefore µ(A) = 0. The same argument works if
A ⊂ E
−E, and any subset of E∆E
can be written as the union
of A
1
and A
2
, where A
1
⊂ E −E
and A
2
⊂ E
−E.
(3) Suppose µ is not a positive measure but µ(E) = 0. If A ∈ /,
then
µ(A) = µ(A∩ E) +µ(A∩ F) ≥ µ(E) +µ(A∩ F) ≥ 0,
which says that µ must be a positive measure, a contradiction. A
similar argument applies for −µ and F.
Let us say two measures µ and ν are mutually singular if there
exist two disjoint sets E and F in / whose union is X with µ(E) =
ν(F) = 0. This is often written µ ⊥ ν.
Example 12.6 If µ is Lebesgue measure restricted to [0, 1/2], that
is, µ(A) = m(A∩[0, 1/2]), and ν is Lebesgue measure restricted to
[1/2, 1], then µ and ν are mutually singular. We let E = [0, 1/2] and
F = (1/2, 1]. This example works because the Lebesgue measure
of ¦1/2¦ is 0.
Example 12.7 A more interesting example is the following. Let f
be the CantorLebesgue function where we deﬁne f(x) = 1 if x ≥ 1
and f(x) = 0 if x ≤ 0 and let ν be the LebesgueStieltjes measure
associated with f. Let µ be Lebesgue measure. Then µ ⊥ ν. To
see this, we let E = C, where C is the Cantor set, and F = C
c
. We
already know that m(E) = 0 and we need to show ν(F) = 0. To do
that, we need to show ν(I) = 0 for every open interval contained
in F. This will follow if we show ν(J) = 0 for every interval of the
form J = (a, b] contained in F. But f is constant on every such
interval, so f(b) = f(a), and therefore ν(J) = f(b) −f(a) = 0.
12.3. JORDAN DECOMPOSITION THEOREM 97
12.3 Jordan decomposition theorem
The following is known as the Jordan decomposition theorem.
Theorem 12.8 If µ is a signed measure on a measurable space
(X, /), there exist positive measures µ
+
and µ
−
such that µ =
µ
+
−µ
−
and µ
+
and µ
−
are mutually singular. This decomposition
is unique.
Proof. Let E and F be negative and positive sets, resp., for µ
so that X = E ∪ F and E ∩ F = ∅. Let µ
+
(A) = µ(A ∩ F),
µ
−
(A) = −µ(A∩ E). This gives the desired decomposition.
If µ = ν
+
− ν
−
is another such decomposition with ν
+
, ν
−
mutually singular, let E
be a set such that ν
+
(E
) = 0 and
ν
−
((E
)
c
) = 0. Set F
= (E
)
c
. Hence X = E
∪F
and E
∩F
= ∅.
If A ⊂ F
, then ν
−
(A) ≤ ν
−
(F
) = 0, and so
µ(A) = ν
+
(A) −ν
−
(A) = ν
+
(A) ≥ 0,
and consequently F
is a positive set for µ. Similarly, E
is a nega
tive set for µ. Thus E
, F
gives another Hahn decomposition of X.
By the uniqueness part of the Hahn decomposition theorem, F∆F
is a null set with respect to µ. Since ν
+
(E
) = 0 and ν
−
(F
) = 0,
if A ∈ /, then
ν
+
(A) = ν
+
(A∩ F
) = ν
+
(A∩ F
) −ν
−
(A∩ F
)
= µ(A∩ F
) = µ(A∩ F) = µ
+
(A),
and similarly ν
−
= µ
−
.
The measure
[µ[ = µ
+
+µ
−
(12.1)
is called the total variation measure of µ.
12.4 Exercises
Exercise 12.1 Suppose µ is a signed measure. Prove that A is a
null set with respect to µ if and only if [µ[(A) = 0.
98 CHAPTER 12. SIGNED MEASURES
Exercise 12.2 Let µ be a signed measure. Deﬁne
_
f dµ =
_
f dµ
+
−
_
f dµ
−
.
Prove that
¸
¸
¸
_
f dµ
¸
¸
¸ ≤
_
[f[ d[µ[.
Exercise 12.3 Let µ be a signed measure on (X, /). Prove that
[µ(A)[ = sup
_¸
¸
¸
_
A
f dµ
¸
¸
¸ : f ≤ 1
_
.
Exercise 12.4 Let µ be a positive measure and ν a signed mea
sure. Prove that ν ¸ µ if and only if ν
+
¸ µ and ν
−
¸ µ.
Exercise 12.5 Let (X, /) be a measurable space. Suppose λ =
µ − ν, where µ and ν are ﬁnite positive measures. Prove that
µ(A) ≥ λ
+
(A) and ν(A) ≥ λ
−
(A) for every A ∈ /.
Exercise 12.6 Let (X, /) be a measurable space. Prove that if µ
and ν are ﬁnite signed measures, then [µ+ν[(A) ≤ [µ(A)[ +[ν(A)[
for every A ∈ /.
Exercise 12.7 Suppose that µ is a signed measure on (X, /).
Prove that if A ∈ /, then
µ
+
(A) = sup¦µ(B) : B ∈ /, B ⊂ A¦
and
µ
−
(A) = −inf¦µ(B) : B ∈ /, B ⊂ A¦.
Exercise 12.8 Suppose that µ is a signed measure on (X, /).
Prove that if A ∈ /, then
[µ[(A) = sup
_
n
j=1
[µ(B
j
)[ : each B
j
∈ /,
the B
j
are disjoint, ∪
n
j=1
B
j
= A
_
.
Chapter 13
The RadonNikodym
theorem
Suppose f is nonnegative and integrable with respect to µ. If we
deﬁne ν by
ν(A) =
_
A
f dµ, (13.1)
then ν is a measure. The only part that needs thought is the
countable additivity. If A
n
are disjoint measurable sets, we have
ν(∪
n
A
n
) =
_
∪
n
A
n
f dµ =
∞
n=1
_
A
n
f dµ =
∞
n=1
ν(A
n
)
by using Proposition 7.5. Moreover, ν(A) is zero whenever µ(A)
is.
In this chapter we consider the converse. If we are given two
measures µ and ν, when does there exist f such that (13.1) holds?
The RadonNikodym theorem answers this question.
13.1 Absolute continuity
Deﬁnition 13.1 A measure ν is said to be absolutely continuous
with respect to a measure µ if ν(A) = 0 whenever µ(A) = 0. We
write ν ¸ µ.
99
100 CHAPTER 13. THE RADONNIKODYM THEOREM
Proposition 13.2 Let ν be a ﬁnite measure. Then ν is absolutely
continuous with respect to µ if and only if for all ε there exists δ
such that µ(A) < δ implies ν(A) < ε.
Proof. Suppose for each ε, there exists δ such that µ(A) < δ
implies ν(A) < ε. If µ(A) = 0, then ν(A) < ε for all ε, hence
ν(A) = 0, and thus ν ¸ µ.
Suppose now that ν ¸ µ. If there exists an ε for which no
corresponding δ exists, then there exists E
k
such that µ(E
k
) < 2
−k
but ν(E
k
) ≥ ε. Let F = ∩
∞
n=1
∪
∞
k=n
E
k
. Then
µ(F) = lim
n→∞
µ(∪
∞
k=n
E
k
) ≤ lim
n→∞
∞
k=n
2
−k
= 0,
but
ν(F) = lim
n→∞
ν(∪
∞
k=n
E
k
) ≥ ε;
This contradicts the absolute continuity.
13.2 The main theorem
Lemma 13.3 Let µ and ν be ﬁnite positive measures on a mea
surable space (X, /). Either µ ⊥ ν or else there exists ε > 0 and
G ∈ / such that µ(G) > 0 and G is a positive set for ν −εµ.
Proof. Consider the Hahn decomposition for ν −
1
n
µ. Thus there
exists a negative set E
n
and a positive set F
n
for this measure,
E
n
and F
n
are disjoint, and their union is X. Let F = ∪
n
F
n
and
E = ∩
n
E
n
. Note E
c
= ∪
n
E
c
n
= ∪
n
F
n
= F.
For each n, E ⊂ E
n
, so
ν(E) ≤ ν(E
n
) ≤
1
n
µ(E
n
) ≤
1
n
µ(X).
Since ν is a positive measure, this implies ν(E) = 0.
One possibility is that µ(E
c
) = 0, in which case µ ⊥ ν. The
other possibility is that µ(E
c
) > 0. In this case, µ(F
n
) > 0 for
some n. Let ε = 1/n and G = F
n
. Then from the deﬁnition of F
n
,
G is a positive set for ν −εµ.
13.2. THE MAIN THEOREM 101
We now are ready for the RadonNikodym theorem.
Theorem 13.4 Suppose µ is a σﬁnite positive measure on a mea
surable space (X, /) and ν is a ﬁnite positive measure on (X, /)
such that ν is absolutely continuous with respect to µ. Then there
exists a µintegrable nonnegative function f which is measurable
with respect to / such that
ν(A) =
_
A
f dµ
for all A ∈ /. Moreover, if g is another such function, then f = g
almost everywhere with respect to µ.
The function f is called the RadonNikodym derivative of ν with
respect to µ or sometimes the density of ν with respect to µ, and
is written f = dν/dµ. Sometimes one writes
dν = f dµ.
The idea of the proof is to look at the set of f such that
_
A
f dµ ≤ ν(A) for each A ∈ /, and then to choose the one such
that
_
X
f dµ is largest.
Proof. Step 1. Let us ﬁrst prove the uniqueness assertion. For
every set A we have
_
A
(f −g) dµ = ν(A) −ν(A) = 0.
By Proposition 8.1 we have f −g = 0 a.e. with respect to µ.
Step 2. Let us assume µ is a ﬁnite measure for now. In this step
we deﬁne the function f. Deﬁne
T =
_
g measurable : g ≥ 0,
_
A
g dµ ≤ ν(A) for all A ∈ /
_
.
T is not empty because 0 ∈ T. Let L = sup¦
_
g dµ : g ∈ T¦,
and let g
n
be a sequence in T such that
_
g
n
dµ → L. Let h
n
=
max(g
1
, . . . , g
n
).
We claim that if g
1
and g
2
are in T, then h
2
= max(g
1
, g
2
) is
102 CHAPTER 13. THE RADONNIKODYM THEOREM
also in T. To see this, let B = ¦x : g
1
(x) ≥ g
2
(x)¦, and write
_
A
h
2
dµ =
_
A∩B
h
2
dµ +
_
A∩B
c
h
2
dµ
=
_
A∩B
g
1
dµ +
_
A∩B
c
g
2
dµ
≤ ν(A∩ B) +ν(A∩ B
c
)
= ν(A).
Therefore h
2
∈ T.
By an induction argument, h
n
is in T.
The h
n
increase, say to f. By monotone convergence,
_
f dµ =
L and
_
A
f dµ ≤ ν(A) (13.2)
for all A.
Step 3. Next we prove that f is the desired function. Deﬁne a
measure λ by
λ(A) = ν(A) −
_
A
f dµ.
λ is a positive measure since f ∈ T.
Suppose λ is not mutually singular to µ. By Lemma 13.3, there
exists ε > 0 and G such that G is measurable, µ(G) > 0, and G is
a positive set for λ −εµ. For any A ∈ /,
ν(A) −
_
A
f dµ = λ(A) ≥ λ(A∩ G) ≥ εµ(A∩ G) =
_
A
εχ
G
dµ,
or
ν(A) ≥
_
A
(f +εχ
G
) dµ.
Hence f +εχ
G
∈ T. But
_
X
(f +εχ
G
) dµ = L +εµ(G) > L,
a contradiction to the deﬁnition of L.
Therefore λ ⊥ µ. Then there must exist H ∈ / such that
µ(H) = 0 and λ(H
c
) = 0. Since ν ¸ µ, then ν(H) = 0, and hence
λ(H) = ν(H) −
_
H
f dµ = 0.
13.3. LEBESGUE DECOMPOSITION THEOREM 103
This implies λ = 0, or ν(A) =
_
A
f dµ for all A.
Step 4. We now suppose µ is σﬁnite. There exist F
i
↑ X such
that µ(F
i
) < ∞ for each i. Let µ
i
be the restriction of µ to F
i
,
that is, µ
i
(A) = µ(A ∩ F
i
). Deﬁne ν
i
, the restriction of ν to F
i
,
similarly. If µ
i
(A) = 0, then µ(A ∩ F
i
) = 0, hence ν(A ∩ F
i
) = 0,
and thus ν
i
(A) = 0. Therefore ν
i
¸ µ
i
. If f
i
is the function such
that dν
i
= f
i
dµ
i
, the argument of Step 1 shows that f
i
= f
j
on F
i
if i ≤ j. Deﬁne f by f(x) = f
i
(x) if x ∈ F
i
. Then for each A ∈ /,
ν(A∩ F
i
) = ν
i
(A) =
_
A
f
i
dµ
i
=
_
A∩F
i
f dµ.
Letting i → ∞ shows that f is the desired function.
13.3 Lebesgue decomposition theorem
The proof of the Lebesgue decomposition theorem is almost the
same.
Theorem 13.5 Suppose µ and ν are two ﬁnite positive measures.
There exist positive measures λ, ρ such that ν = λ+ρ, ρ is absolutely
continuous with respect to µ, and λ and µ are mutually singular.
Proof. Deﬁne T and L and construct f as in the proof of the
RadonNikodym theorem. Let ρ(A) =
_
A
f dµ and let λ = ν − ρ.
Our construction shows that
_
A
f dµ ≤ ν(A),
so λ(A) ≥ 0 for all A. We have ρ +λ = ν. We need to show µ and
λ are mutually singular.
If not, by Lemma 13.3, there exists ε > 0 and F ∈ / such that
µ(F) > 0 and F is a positive set for λ − εµ. We get a contradic
tion exactly as in the proof of the RadonNikodym theorem. We
conclude that λ ⊥ µ.
104 CHAPTER 13. THE RADONNIKODYM THEOREM
13.4 Exercises
Exercise 13.1 This exercise asks you to prove the RadonNiko
dym theorem for signed measures. Let (X, /) be a measurable
space. Suppose ν is a signed measure, µ is a ﬁnite positive measure,
and ν(A) = 0 whenever µ(A) = 0 and A ∈ /. Show there exists
an integrable realvalued function f such that ν(A) =
_
A
f dµ for
all A ∈ /.
Exercise 13.2 We deﬁne a complex measure µ on a measurable
space (X, /) to be a map from / to C such that µ(∅) = 0 and
µ(∪
∞
i=1
A
i
) =
∞
i=1
µ(A
i
) whenever the A
i
are in / and are pair
wise disjoint. Formulate and prove a RadonNikodym theorem for
complex measures.
Exercise 13.3 Let (X, /) be a measurable space and let µ and ν
be two ﬁnite measures. We say µ and ν are equivalent measures if
µ ¸ ν and ν ¸ µ. Show that µ and ν are equivalent if and only
if there exists a measurable function f that is strictly positive a.e.
with respect to µ such that dν = f dµ.
Exercise 13.4 Suppose µ and ν are two ﬁnite measures such that
ν is absolutely continuous with respect to µ. Let ρ = µ + ν. Note
that µ(A) ≤ ρ(A) and ν(A) ≤ ρ(A) for each measurable A. In
particular, µ ¸ ρ and ν ¸ ρ. Prove that if f = dµ/dρ and
g = dν/dρ, then g is strictly positive for almost every x with respect
to µ, f +g = 1, and dν = (f/g) dµ.
Exercise 13.5 If µ is a signed measure on (X, /) and [µ[ is the
total variation measure, prove that there exists a realvalued func
tion f that is measurable with respect to / such that [f[ = 1 a.e.
with respect to µ and dµ = f d[µ[.
Exercise 13.6 Suppose ν ¸ µ and ρ ¸ ν. Prove that ρ ¸ µ and
dρ
dµ
=
dρ
dν
dν
dµ
.
Exercise 13.7 Suppose λ
n
is a sequence of positive measures on
a measurable space (X, /) with sup
n
λ
n
(X) < ∞ and µ is another
ﬁnite positive measure on (X, /). Suppose λ
n
= f
n
dµ + ν
n
is
13.4. EXERCISES 105
the Lebesgue decomposition of λ
n
; in particular, ν
n
⊥ µ. If λ =
∞
n=1
λ
n
, show that
λ =
_
∞
n=1
f
n
_
dµ +
∞
n=1
ν
n
is the Lebesgue decomposition of λ.
Exercise 13.8 Let (X, T, µ) be a measure space, and suppose c
is a subσalgebra of T, that is, c is itself a σalgebra and c ⊂ T.
Suppose f is a nonnegative integrable function that is measurable
with respect to T. Deﬁne ν(A) =
_
A
f dµ for A ∈ c and let µ be
the restriction of µ to c.
(1) Prove that ν ¸ µ.
(2) Since ν and µ are measures on c, then g = dν/dµ is measurable
with respect to c. Prove that
_
A
g dµ =
_
A
f dµ (13.3)
whenever A ∈ c. g is called the conditional expectation of f with
respect to c and we write g = E[f [ c]. If f is integrable and
realvalued but not necessarily nonnegative, we deﬁne
E[f [ c] = E[f
+
[ c] −E[f
−
[ c].
(3) Show that f = g if and only if f is measurable with respect to
c.
(4) Prove that if h is c measurable and
_
A
hdµ =
_
A
f dµ for all
A ∈ c, then h = g a.e. with respect to µ.
Exercise 13.9 Suppose (X, /, µ) is a measure space and f is in
tegrable and measurable with respect to /. Suppose in addition
that B
1
, B
2
, . . . , B
n
is a ﬁnite sequence of disjoint elements of /
whose union is X and that each B
j
has positive µ measure. Let
( = σ(B
1
, . . . , B
n
). Prove that
E[f [ (] =
n
j=1
_
B
j
f dµ
µ(B
j
)
χ
B
j
.
Exercise 13.10 Suppose that (X, T, µ) is a measure space, c is a
subσalgebra of T, and T is a subσalgebra of c. Suppose f is
106 CHAPTER 13. THE RADONNIKODYM THEOREM
integrable, realvalued, and measurable with respect to T. Prove
that
E[ E[f [ c] [ T] = E[f [ T]
and
E[ E[f [ T] [ c] = E[f [ T].
Exercise 13.11 Suppose that (X, T, µ) is a measure space and
c is a subσalgebra of T. Suppose that f and fg are integrable
realvalued functions, where f is measurable with respect to T and
g is measurable with respect to c. Prove that
E[fg [ c] = gE[f [ c].
Chapter 14
Diﬀerentiation
In this chapter we want to look at when a function from R to R is
diﬀerentiable and when the fundamental theorem of calculus holds.
Brieﬂy, our results are the following.
(1) The derivative of
_
x
a
f(y) dy is equal to f a.e. if f is integrable
(Theorem 14.5);
(2) Functions of bounded variation, in particular monotone func
tions, are diﬀerentiable (Theorem 14.8);
(3)
_
b
a
f
(y) dy = f(b)−f(a) if f is absolutely continuous (Theorem
14.14).
Our approach uses what are known as maximal functions and
uses the RadonNikodym theorem and the Lebesgue decomposition
theorem. However, some students and instructors prefer a more
elementary proof of the results on diﬀerentiation. In Sections 14.5,
14.6, and 14.7 we give an alternative approach that avoids the
use of the RadonNikodym theorem and Lebesgue decomposition
theorem.
The deﬁnition of derivative is the same as in elementary calcu
lus. A function f is diﬀerentiable at x if
lim
h→0
f(x +h) −f(x)
h
exists, and the limit is called the derivative of f at x and is denoted
f
(x). If f : [a, b] → R, we say f is diﬀerentiable on [a, b] if the
107
108 CHAPTER 14. DIFFERENTIATION
derivative exists for each x ∈ (a, b) and both
lim
h→0+
f(a +h) −f(a)
h
and lim
h→0−
f(b +h) −f(b)
h
exist.
14.1 Maximal functions
In this section we consider realvalued functions on R
n
. Let B(x, r)
be the open ball with center x and radius r.
The following is an example of what is known as a covering
lemma. We use m for Lebesgue measure on R
n
throughout this
section.
Proposition 14.1 Suppose E ⊂ R
n
is covered by a collection of
balls ¦B
α
¦ of bounded diameter. There exists a disjoint sequence
B
1
, B
2
, . . . of elements of ¦B
α
¦ such that
m(E) ≤ 5
n
k
m(B
k
).
Proof. Let d(B
α
) be the diameter of B
α
. Choose B
1
such that
d(B
1
) ≥
1
2
sup
α
d(B
α
).
Once B
1
, . . . , B
k
are chosen, choose B
k+1
disjoint from B
1
, . . . , B
k
such that
d(B
k+1
) ≥
1
2
sup¦d(B
α
) : B
α
is disjoint from B
1
, . . . , B
k
¦.
The procedure might terminate after a ﬁnite number of steps or it
might not.
If
k
m(B
k
) = ∞, we have our result. Suppose
k
m(B
k
) <
∞. Let B
∗
k
be a ball with the same center as B
k
but 5 times the
radius. We claim E ⊂ ∪
k
B
∗
k
. Once we have this,
m(E) ≤ m(∪
k
B
∗
k
) ≤
k
m(B
∗
k
) = 5
n
k
m(B
k
).
14.1. MAXIMAL FUNCTIONS 109
To show E ⊂ ∪
k
B
∗
k
, it suﬃces to show each B
α
⊂ ∪
k
B
∗
k
, since
¦B
α
¦ is a cover of E. Fix α. If B
α
is one of the B
k
, we are done.
If
k
m(B
k
) < ∞, then d(B
k
) → 0. Let k be the smallest
integer such that d(B
k+1
) <
1
2
d(B
α
). B
α
must intersect one of
B
1
, . . . , B
k
, or else we would have chosen it instead of B
k+1
. There
fore B
α
intersects B
j
0
for some j
0
≤ k. We know
1
2
d(B
α
) ≤ d(B
j
0
),
and some simple geometry shows that B
α
⊂ B
∗
j
0
. In fact, let x
j
0
be the center of B
j
0
and y a point in B
α
∩ B
j
0
. If x ∈ B
α
, then
[x −x
j
0
[ ≤ [x −y[ +[y −x
j
0
[ < d(B
α
) +d(B
j
0
)/2 ≤
5
2
d(B
j
0
),
or x ∈ B
∗
j
0
. Therefore B
α
⊂ B
j
0
, and the proof is complete.
We say f is locally integrable if
_
K
[f(x)[ dx is ﬁnite whenever
K is compact. If f is locally integrable, deﬁne
Mf(x) = sup
r>0
1
m(B(x, r))
_
B(x,r)
[f(y)[ dy.
Note that without the supremum, we are looking at the average of
[f[ over B(x, r). The function Mf is called the maximal function
of f.
We now prove a weak 11 inequality, due to Hardy and Little
wood. It is so named because M does not map integrable functions
into integrable functions, but comes close in a certain sense to doing
so.
Theorem 14.2 If f is integrable, then for all β > 0
m(¦x : Mf(x) > β¦) ≤
5
n
β
_
[f(x)[ dx.
Proof. Fix β and let E
β
= ¦x : Mf(x) > β¦. If x ∈ E
β
, then there
exists a ball B
x
centered at x such that
_
B
x
[f(x)[ dx > βm(B
x
)
by the deﬁnition of Mf(x). Then
m(B
x
) ≤
_
[f[
β
,
so ¦B
x
¦ is a cover of E
β
by balls of bounded diameter. Extract
a disjoint sequence B
1
, B
2
, . . . such that 5
n
k
m(B
k
) ≥ m(E
β
).
Then
m(E
β
) ≤ 5
n
k
m(B
k
) ≤
5
n
β
k
_
B
k
[f[
110 CHAPTER 14. DIFFERENTIATION
=
5
n
β
_
∪
k
B
k
[f[ ≤
5
n
β
_
[f[,
as desired.
If we look at the function f(x) = χ
B
, where B is the unit ball,
note that Mf(x) is approximately a constant times [x[
−n
for x
large, so Mf is not integrable. Hence M does not map the class of
integrable functions into the class of integrable functions.
Theorem 14.3 Let
f
r
(x) =
1
m(B(x, r))
_
B(x,r)
f(y) dy. (14.1)
If f is locally integrable, then f
r
(x) → f(x) a.e. as r → 0.
Proof. It suﬃces to prove that for each N, f
r
(x) → f(x) for
almost every x ∈ B(0, N). Let N > 0. We may suppose without
loss of generality that f is 0 outside of B(0, 2N), and thus we may
suppose f is integrable.
Fix β > 0. Let ε > 0. Using Theorem 8.4, take g continuous
with compact support such that
_
[f − g[ dm < ε. If g
r
is deﬁned
analogously to f
r
using (14.1),
[g
r
(x) −g(x)[ =
¸
¸
¸
1
m(B(x, r))
_
B(x,r)
[g(y) −g(x)] dy
¸
¸
¸ (14.2)
≤
1
m(B(x, r))
_
B(x,r)
[g(y) −g(x)[ dy → 0
as r → 0 by the continuity of g. We have
limsup
r→0
[f
r
(x) −f(x)[ ≤ limsup
r→0
[f
r
(x) −g
r
(x)[
+ limsup
r→0
[g
r
(x) −g(x)[
+[g(x) −f(x)[.
The second term on the right is 0 by (14.2). We now use Theorem
14.2 and Lemma 10.4 to write
m(¦x : limsup
r→0
[f
r
(x) −f(x)[ > β¦)
14.1. MAXIMAL FUNCTIONS 111
≤ m(¦x : limsup
r→0
[f
r
(x) −g
r
(x)[ > β/2¦)
+m(¦x : [f(x) −g(x)[ > β/2¦)
≤ m(¦x : M(f −g)(x) > β/2¦) +
_
[f −g[
β/2
≤
2(5
n
+ 1)
β
_
[f −g[
<
2(5
n
+ 1)ε
β
,
where we use the deﬁnition of the maximal function to see that
[f
r
(x) −g
r
(x)[ ≤ M(f −g)(x)
for all r. This is true for every ε, so
m(¦x : limsup
r→0
[f
r
(x) −g
r
(x)[ > β¦) = 0.
We apply this with β = 1/j for each positive integer j, and we
conclude
m(¦x : limsup
r→0
[f
r
(x) −f(x)[ > 0¦)
≤ ∪
∞
j=1
m(¦x : limsup
r→0
[f
r
(x) −f(x)[ > 1/j¦) = 0.
This is the result we seek.
We can get a stronger statement:
Theorem 14.4 For almost every x
1
m(B(x, r))
_
B(x,r)
[f(y) −f(x)[ dy → 0
as r → 0.
Proof. For each rational c there exists a set N
c
of measure 0 such
that
1
m(B(x, r))
_
B(x,r)
[f(y) −c[ dy → [f(x) −c[
112 CHAPTER 14. DIFFERENTIATION
for x / ∈ N
c
; we see this by applying Theorem 14.3 to the function
[f(x) − c[. Let N = ∪
c∈Q
N
c
and suppose x / ∈ N. Let ε > 0 and
choose c rational such that [f(x) −c[ < ε. Then
1
m(B(x, r))
_
B(x,r)
[f(y) −f(x)[ dy
≤
1
m(B(x, r))
_
B(x,r)
[f(y) −c[ dy
+
1
m(B(x, r))
_
B(x,r)
[f(x) −c[ dy
=
1
m(B(x, r))
_
B(x,r)
[f(y) −c[ dy +[f(x) −c[
and hence
limsup
r→0
1
m(B(x, r))
_
B(x,r)
[f(y) −f(x)[ dy ≤ 2[f(x) −c[ < 2ε.
Since ε is arbitrary, our result follows.
If we apply the above to the function f = χ
E
, then for almost
all x ∈ E
1
m(B(x, r)
_
B(x,r)
χ
E
→ 1,
or
m(E ∩ B(x, r))
m(B(x, r))
→ 1,
and similarly, for almost all x / ∈ E, the ratio tends to 0. The points
where the ratio tends to 1 are called points of density for E.
14.2 Antiderivatives
For the remainder of the chapter we consider realvalued functions
on the real line R. We can use the results on maximal functions
to show that the derivative of the antiderivative of an integrable
function is the function itself. A ball B(x, h) in R is merely the
interval (x−h, x+h). We use m for Lebesgue measure throughout.
Deﬁne the indeﬁnite integral or antiderivative of an integrable
function f by
F(x) =
_
x
a
f(t) dt.
14.3. BOUNDED VARIATION 113
Recall by Exercise 7.6 that F is continuous.
Theorem 14.5 Suppose f : R → R is integrable and a ∈ R. De
ﬁne
F(x) =
_
x
a
f(y) dy.
Then F is diﬀerentiable almost everywhere and F
(x) = f(x) a.e.
Proof. If h > 0, we have
F(x +h) −F(x) =
_
x+h
x
f(y) dy,
so
¸
¸
¸
F(x +h) −F(x)
h
−f(x)
¸
¸
¸ =
1
h
¸
¸
¸
_
x+h
x
(f(y) −f(x)) dy
¸
¸
¸
≤ 2
1
m(B(x, h))
_
x+h
x−h
[f(y) −f(x)[ dy.
By Theorem 14.4, the right hand side goes to 0 as h → 0 for almost
every x, and we conclude the right hand derivative of F exists and
equals f for almost every x. The left hand derivative is handled
similarly.
14.3 Bounded variation
In this section we show that functions of bounded variation are
diﬀerentiable almost everywhere. We start with right continuous
increasing functions.
Lemma 14.6 Suppose H : R →R is increasing, right continuous,
and constant for x ≥ 1 and x ≤ 0. Let λ be the LebesgueStieltjes
measure deﬁned using the function H and suppose λ and m are
mutually singular. Then
lim
r→0
λ(B(x, r))
m(B(x, r))
= 0
for almost every x.
114 CHAPTER 14. DIFFERENTIATION
Proof. This is clear if x < 0 or x > 1. Since λ ⊥ m, there exist
measurable sets E and F such that λ(F) = 0, m(E) = 0, and
F = E
c
. Let ε > 0.
Step 1. The ﬁrst step of the proof is to ﬁnd a bounded open set
G such that F ⊂ G and λ(G) < ε. By the deﬁnition of Lebesgue
Stieltjes measure, there exist a
i
< b
i
such that F ⊂ ∪
∞
i=1
(a
i
, b
i
]
and
∞
i=1
[H(b
i
) −H(a
i
)] < ε/2.
Since H is right continuous, for each i there exists b
i
> b
i
such that
H(b
i
) −H(b
i
) < ε/2
i+1
.
If G
= ∪
∞
i=1
(a
i
, b
i
), then G
is open, G
contains F, and
λ(G
) ≤
∞
i=1
λ((a
i
, b
i
)) ≤
∞
i=1
λ((a
i
, b
i
]) =
∞
i=1
[H(b
i
) −H(a
i
)] < ε.
Since H is constant on (−∞, 0] and [1, ∞), we can take G to be
the set G = G
∩ (−1, 2).
Step 2. If β > 0, let
A
β
=
_
x ∈ F ∩ [0, 1] : limsup
r→0
λ((x, r))
m(B(x, r))
> β
_
.
The second step is to show that m(A
β
) = 0. If x ∈ A
β
, then
x ∈ F ⊂ G, and there exists an open ball B
x
centered at x and
contained in G such that λ(B
x
)/m(B
x
) > β. Use Proposition 14.1
to ﬁnd a disjoint subsequence B
1
, B
2
, . . . such that
m(A
β
) ≤ 5
∞
i=1
m(B
i
).
Then
m(A
β
) ≤ 5
∞
i=1
m(B
i
) ≤
5
β
∞
i=1
λ(B
i
) ≤
5
β
λ(G) ≤
5
β
ε.
Since ε is arbitrary, and our construction of G did not depend on
β, then m(A
β
) = 0.
Since m(A
1/k
) = 0 for each k, then
m(¦x ∈ F ∩ [0, 1] : limsup
r→0
λ(B(x, r))/m(B(x, r)) > 0¦) = 0.
14.3. BOUNDED VARIATION 115
Since m(E) = 0, this completes the proof.
Proposition 14.7 Let F : R →R be an increasing and right con
tinuous function. Then F
exists a.e. Moreover, F
is locally inte
grable and for every a < b,
_
b
a
F
(x) dx ≤ F(b) −F(a).
Proof. We will show F is diﬀerentiable a.e. on [0, 1]. Once we have
that, the same argument can be used to show that F is diﬀeren
tiable a.e. on [−N, N] for each N, and that proves that F is diﬀer
entiable a.e. on R. If we redeﬁne F so that F(x) = lim
y→0+
F(y)
if x ≤ 0 and F(x) = F(1) if x > 1, then F is still right continuous
and increasing, and we have not aﬀected the diﬀerentiability of F
on [0, 1] except possibly at the points 0 and 1.
Let ν be the LebesgueStieltjes measure deﬁned in terms of F.
By the Lebesgue decomposition theorem, we can write ν = λ + ρ,
where λ ⊥ m and ρ ¸ m. Note
ρ([0, 1]) ≤ ν([0, 1]) = F(1) −F(0).
By the RadonNikodym theorem there exists a nonnegative inte
grable function f such that ρ(A) =
_
A
f dm for each measurable
A.
Let
H(x) = λ((0, x]) = ν((0, x])−ρ((0, x]) = F(x)−F(0)−
_
x
0
f(y) dy.
By Exercise 7.6, the function x →
_
x
0
f(y) dy is continuous, so
H is right continuous, increasing, and λ is the LebesgueStieltjes
measure deﬁned in terms of H. By Lemma 14.6,
limsup
h→0+
H(x +h) −H(x)
h
≤ limsup
h→0+
H(x +h) −H(x −h)
h
= limsup
h→0+
λ((x −h, x +h])
h
≤ 4 limsup
h→0+
λ(B(x, 2h))
4h
= 0
for almost every x. The same is true for the left hand derivative,
so H
exists and equals 0 for almost every x. We saw by Theo
rem 14.5 that the function x →
_
x
0
f(y) dy is diﬀerentiable almost
everywhere, and we conclude that F is diﬀerentiable a.e.
116 CHAPTER 14. DIFFERENTIATION
We have shown that F
= f a.e. If a < b,
_
b
a
F
(x) dx =
_
b
a
f(x) dx = ρ((a, b]) ≤ ν((a, b]) = F(b) −F(a).
This completes the proof.
Here is the main theorem on the diﬀerentiation of increasing
functions.
Theorem 14.8 If F : R → R is increasing, then F
exists a.e.
and
_
b
a
F
(x) dx ≤ F(b) −F(a) (14.3)
whenever a < b.
Proof. Let G(x) = lim
y→x+
F(y). Since F is increasing, there are
at most countably many values of x where F is not continuous, so
F(x) = G(x) a.e. Since G is increasing and right continuous, G is
diﬀerentiable a.e. by Proposition 14.7. We will show that if x is a
point where G is diﬀerentiable and at the same time F(x) = G(x),
then F
(x) exists and is equal to G
(x).
Let x be such a point, let L = G
(x) and let ε > 0. Because F
and G are increasing, for any h > 0 there exists a point x
h
strictly
between x +h and x + (1 +ε)h where F and G agree, and so
F(x +h) ≤ F(x
h
) = G(x
h
) ≤ G(x + (1 +ε)h).
Then
limsup
h→0+
F(x +h) −F(x)
h
≤ limsup
h→0+
G(x + (1 +ε)h) −G(x)
h
= (1 +ε) limsup
h→0+
G(x + (1 +ε)h) −G(x)
(1 +ε)h
= (1 +ε)L.
Similarly, liminf
h→0+
[F(x + h) − F(x)]/h ≥ (1 − ε)L. Since ε
is arbitrary, the right hand derivative of F exists at x and is equal
to L. That the left hand derivative equals L is proved similarly.
14.3. BOUNDED VARIATION 117
Since F
= G
a.e., then F
is locally integrable. If a < b, take
a
n
↓ a and b
n
↑ b such that F and G agree on a
n
and b
n
. Then
using Proposition 14.7,
F(b) −F(a) ≥ F(b
n
) −F(a
n
)
= G(b
n
) −G(a
n
) ≥
_
b
n
a
n
G
(x) dx
=
_
b
n
a
n
F
(x) dx.
Now let n → ∞ and use the monotone convergence theorem.
Remark 14.9 Note that if F is the CantorLebesgue function,
then F
(x) = 0 a.e., in fact at every point of C
c
, where C is the
Cantor set. Thus
1 = F(1) −F(0) > 0 =
_
1
0
F
(x) dx,
and we do not in general have equality in (14.3).
A realvalued function f is of bounded variation on [a, b] if
sup
_
k
i=1
[f(x
i
) −f(x
i−1
)[
_
is ﬁnite, where the supremum is over all partitions a = x
0
< x
1
<
< x
k
= b of [a, b].
Lemma 14.10 If f is of bounded variation on [a, b], then f can be
written as f = f
1
− f
2
, the diﬀerence of two increasing functions
on [a, b].
Proof. Deﬁne
f
1
(y) = sup
_
k
i=1
[f(x
i
) −f(x
i−1
)]
+
_
and
f
2
(y) = sup
_
k
i=1
[f(x
i
) −f(x
i−1
)]
−
_
,
118 CHAPTER 14. DIFFERENTIATION
where the supremum is over all partitions a = x
0
< x
1
< <
x
k
= y for y ∈ [a, b]. f
1
and f
2
are measurable since they are both
increasing. Since
k
i=1
[f(x
i
) −f(x
i−1
)]
+
=
k
i=1
[f(x
i
) −f(x
i−1
)]
−
+f(y) −f(a),
taking the supremum over all partitions of [a, y] yields
f
1
(y) = f
2
(y) +f(y) −f(a).
Clearly f
1
and f
2
are increasing in y, and the result follows by
solving for f(y).
Using this lemma and Theorem 14.8, we see that functions of
bounded variation are diﬀerentiable a.e. Note that the converse
is not true: the function sin(1/x) deﬁned on (0, 1] is diﬀerentiable
everywhere, but is not of bounded variation.
Remark 14.11 If we write a function f of bounded variation as
the diﬀerence of two increasing functions f
1
and f
2
, then the quan
tity (f
1
(b) + f
2
(b)) − (f
1
(a) + f
2
(a)) is called the total variation
of f on the interval [a, b]. We make the observation that if f is of
bounded variation on the interval [a, b] and on the interval [b, c],
then it is of bounded variation on the interval [a, c].
14.4 Absolutely continuous functions
A realvalued function f is absolutely continuous on [a, b] if given ε
there exists δ such that
k
i=1
[f(b
i
)−f(a
i
)[ < ε whenever ¦(a
i
, b
i
)¦
is a ﬁnite collection of disjoint intervals with
k
i=1
[b
i
−a
i
[ < δ.
It is easy to see that absolutely continuous functions are con
tinuous and that the CantorLebesgue function is not absolutely
continuous.
Lemma 14.12 If f is absolutely continuous, then it is of bounded
variation.
Proof. By the deﬁnition of absolutely continuous function with
ε = 1, there exists δ such that
k
i=1
[f(b
i
) − f(a
i
)[ < 1 whenever
14.4. ABSOLUTELY CONTINUOUS FUNCTIONS 119
k
i=1
(b
i
−a
i
) ≤ δ and the (a
i
, b
i
) are disjoint open intervals. Hence
the total variation of f on [a+jδ, a+(j +1)δ] is less than or equal
to 1. Using Remark 14.11, we see the total variation of f on [a, b]
is ﬁnite.
Lemma 14.13 Suppose f is of bounded variation and we decom
pose f as f = f
1
−f
2
, where f
1
and f
2
are increasing functions. If
f is absolutely continuous, then so are f
1
and f
2
.
Proof. Given ε there exists δ such that
k
i=1
[f(b
i
) − f(a
i
)[ <
ε whenever
k
i=1
(b
i
− a
i
) ≤ δ and the (a
i
, b
i
) are disjoint open
intervals. Partitioning each interval (a
i
, b
i
) into subintervals with
a
i
= s
i0
< s
i1
< < s
iJ
i
= b
i
, then
k
i=1
J
i
−1
j=0
(s
i,j+1
−s
ij
) =
k
i=1
(b
i
−a
i
) ≤ δ.
Hence
k
i=1
J
i
−1
j=0
[f(s
i,j+1
) −f(s
ij
)[ ≤ ε.
Taking the supremum over all such partitions,
k
i=1
[(f
1
+f
2
)(b
i
) −(f
1
+f
2
)(a
i
)[ ≤ ε,
and our conclusion follows.
Here is the main theorem on absolutely continuous functions.
Theorem 14.14 If F is absolutely continuous, then F
exists a.e.,
and
_
b
a
F
(x) dx = F(b) −F(a).
Proof. By Lemma 14.13 it suﬃces to suppose F is increasing and
absolutely continuous. Let ν be the LebesgueStieltjes measure de
ﬁned in terms of F. Since F is continuous, F(d) −F(c) = ν((c, d)).
120 CHAPTER 14. DIFFERENTIATION
Taking a limit as k → ∞, we see that given ε there exists δ such
that
∞
i=1
[F(b
i
) −F(a
i
)[ ≤ ε whenever ¦(a
i
, b
i
)¦ is a collection of
disjoint intervals with
∞
i=1
(b
i
−a
i
) < δ. Since any open set G can
be written as the union of disjoint intervals ¦(a
i
, b
i
)¦, this can be
rephrased as saying, given ε there exists δ such that
ν(G) =
∞
i=1
ν((a
i
, b
i
)) =
∞
i=1
(F(b
i
) −F(a
i
)) ≤ ε
whenever G is open and m(G) < δ. If m(A) < δ and A is Borel
measurable, then there exists an open set G containing A such that
m(G) < δ, and then ν(A) ≤ ν(G) ≤ ε. We conclude that ν ¸ m.
Hence there exists a nonnegative integrable function f such
that
ν(A) =
_
A
f dm
for all Borel measurable sets A. In particular, for each x ∈ [a, b],
F(x) −F(a) = ν((a, x)) =
_
x
a
f(y) dy.
By Theorem 14.5, F
exists and is equal to f a.e. Setting x = b we
obtain
F(b) −F(a) =
_
b
a
F
(y) dy
as desired.
14.5 Approach 2 – diﬀerentiability
In this and the following two sections we give an alternative ap
proach to Theorems 14.5, 14.8, and 14.14 that avoids the use of
the RadonNikodym theorem, and instead proceeds via a covering
lemma due to Vitali.
Let m be Lebesgue measure. Let E ⊂ R be a measurable set
and let ( be a collection of intervals. We say ( is a Vitali cover of
E if for each x ∈ E and each ε > 0 there exists an interval G ∈ (
containing x whose length is less than ε.
The following is known as the Vitali covering lemma, and is a
reﬁnement of Proposition 14.1.
14.5. APPROACH 2 – DIFFERENTIABILITY 121
Lemma 14.15 Suppose E has ﬁnite measure and let ( be a Vitali
cover of E. Given ε > 0 there exists a ﬁnite subcollection of disjoint
intervals I
1
, . . . , I
n
∈ ( such that m(E −∪
n
i=1
I
n
) < ε.
Proof. We may replace each interval in ( by a closed one, since
the set of endpoints of a ﬁnite subcollection will have measure 0.
Let G be an open set of ﬁnite measure containing E. Since ( is
a Vitali cover, we may suppose without loss of generality that each
set of ( is contained in G. Let
a
0
= sup¦m(I) : I ∈ (¦.
Let I
1
be any element of ( with m(I
1
) ≥ a
0
/2. Let
a
1
= sup¦m(I) : I ∈ (, I disjoint from I
1
¦,
and choose I
2
∈ ( disjoint from I
1
such that m(I
2
) ≥ a
1
/2. Con
tinue in this way, choosing I
n+1
disjoint from I
1
, . . . , I
n
and in (
with length at least one half as large as any other such interval in
( that is disjoint from I
1
, . . . , I
n
.
If the process stops at some ﬁnite stage, we are done. If not, we
generate a sequence of disjoint intervals I
1
, I
2
, . . . Since they are
disjoint and all contained in G, then
∞
i=1
m(I
i
) ≤ m(G) < ∞.
Therefore there exists N such that
∞
i=N+1
m(I
i
) < ε/5.
Let R = E−∪
N
i=1
I
i
. We claim m(R) < ε. Let I
∗
n
be the interval
with the same center as I
n
but ﬁve times the length. Let x ∈ R.
Since we supposed each interval in ( was to be modiﬁed so as to
include its endpoints, then ∪
n
i=1
I
i
is closed. Hence there exists
an interval I ∈ ( containing x with I disjoint from I
1
, . . . , I
N
.
Since
n
m(I
n
) < ∞, then
n
a
n
≤ 2
n
m(I
n
) < ∞, and a
n
→
0. Hence I must either be one of the I
n
for some n > N or at
least intersect it, for otherwise we would have chosen I at some
stage. Let n be the smallest integer such that I intersects I
n
; note
n > N. We have m(I) ≤ a
n−1
≤ 2m(I
n
). Since x is in I and I
intersects I
n
, the distance from x to the midpoint of I
n
is at most
m(I) +m(I
n
)/2 ≤ (5/2)m(I
n
). Therefore x ∈ I
∗
n
.
Thus we have R ⊂ ∪
∞
i=N+1
I
∗
i
, so
m(R) ≤
∞
i=N+1
m(I
∗
i
) = 5
∞
i=N+1
m(I
i
) < ε.
122 CHAPTER 14. DIFFERENTIATION
This completes the proof.
Given a realvalued function f, we deﬁne the derivates of f at
x by
D
+
f(x) = limsup
h→0+
f(x +h) −f(x)
h
,
D
−
f(x) = limsup
h→0−
f(x +h) −f(x)
h
,
D
+
f(x) = liminf
h→0+
f(x +h) −f(x)
h
,
D
−
f(x) = liminf
h→0−
f(x +h) −f(x)
h
.
If all the derivates are equal, then f is diﬀerentiable at x and f
(x)
is the common value.
Theorem 14.16 Suppose f is increasing on [a, b]. Then f is dif
ferentiable almost everywhere, f
is integrable, and
_
b
a
f
(x) dx ≤ f(b) −f(a). (14.4)
Proof. We will show that the set where any two derivates are
unequal has measure zero. Let us consider the set
E = ¦x : D
+
f(x) > D
−
f(x)¦,
the other sets being similar. Let
E
uv
= ¦x : D
+
f(x) > v > u > D
−
f(x)¦.
If we show m(E
uv
) = 0, then observing that E ⊂ ∪
u<v,u,v∈Q
E
uv
will show that m(E) = 0.
Let s = m(E
uv
), let ε > 0, and choose an open set G such
that E
uv
⊂ G and m(G) < s + ε. For each x ∈ E
uv
there exists
an arbitrarily small interval [x − h, x] contained in G such that
f(x) − f(x − h) < uh. Use Lemma 14.15 to choose I
1
, . . . , I
N
which are disjoint and whose interiors cover a subset A of E
uv
of
measure greater than s −ε. Write I
n
= [x
n
−h
n
, x
n
]. Thus
A = E
uv
∩ (∩
N
n=1
(x
n
−h
n
, x
n
)).
14.5. APPROACH 2 – DIFFERENTIABILITY 123
Taking a sum,
N
n=1
[f(x
n
) −f(x
n
−h
n
)] < u
n
n=1
h
n
< um(G) < u(s +ε).
Each point y in the subset A is the left endpoint of an arbitrarily
small interval (y, y +k) that is contained in some I
n
and for which
f(y + k) − f(y) > vk. Using Lemma 14.15 again, we pick out a
ﬁnite collection J
1
, . . . , J
M
whose union contains a subset of A of
measure larger than s −2ε. Summing over these intervals yields
M
i=1
[f(y
i
+k
i
) −f(y
i
)] > v
M
i=1
k
i
> v(s −2ε).
Each interval J
i
is contained in some interval I
n
, and if we sum
over those i for which J
i
⊂ I
n
we ﬁnd
{i:J
i
⊂I
n
}
[f(y
i
+k
i
) −f(y
i
)] ≤ f(x
n
) −f(x
n
−h
n
),
since f is increasing. Thus
N
n=1
[f(x
n
) −f(x
n
−h
n
)] ≥
M
i=1
[f(y
i
+k
i
) −f(y
i
)],
and so u(s + ε) > v(s − 2ε). This is true for each ε, so us ≥ vs.
Since v > u, this implies s = 0.
We prove similarly that D
+
f = D
+
f a.e. and D
−
f = D
−
f a.e.
and conclude that
g(x) = lim
h→0
f(x +h) −f(x)
h
is deﬁned almost everywhere and that f is diﬀerentiable wherever
g is ﬁnite.
Deﬁne f(x) = f(b) if x ≥ b. Let
g
n
(x) = n[f(x + 1/n) −f(x)].
Then g
n
(x) → g(x) for almost all x, and so g is measurable. Since
f is increasing, g
n
≥ 0. By Fatou’s lemma and the fact that f is
124 CHAPTER 14. DIFFERENTIATION
increasing,
_
b
a
g ≤ liminf
n
_
b
a
g
n
= liminf
n
n
_
b
a
[f(x + 1/n) −f(x)]
= liminf
n
_
n
_
b+1/n
b
f −n
_
a+1/n
a
f
_
= liminf
n
_
f(b) −n
_
a+1/n
a
f
_
≤ f(b) −f(a).
We used a change of variables in the second equality. This shows
that g is integrable and hence ﬁnite almost everywhere.
We refer the reader to Lemma 14.10 to see that a function of
bounded variation is diﬀerentiable almost everywhere.
14.6 Approach 2 – antiderivatives
Continuing the alternative approach, we look at when the deriva
tive of
F(x) =
_
x
a
f(t) dt (14.5)
is equal to f(x) a.e.
Theorem 14.17 If f is integrable and F is deﬁned by (14.5), then
F
(x) = f(x) for almost every x.
Proof. By writing f = f
+
− f
−
, it suﬃces to consider the case
where f is nonnegative. In this case F is increasing, and so F
exists a.e. By Exercise 7.6 we know F is continuous.
Suppose for the moment that f is bounded by K. Then
¸
¸
¸
F(x + 1/n) −F(x)
1/n
¸
¸
¸ =
¸
¸
¸n
_
x+1/n
x
f(t) dt
¸
¸
¸
14.6. APPROACH 2 – ANTIDERIVATIVES 125
is also bounded by K. By dominated convergence,
_
c
a
F
(x) dx = lim
n
n
_
c
a
[F(x + 1/n) −F(x)] dx
= lim
n
_
n
_
c+1/n
c
F(x) dx −n
_
a+c
a
F(x) dx
_
= F(c) −F(a) =
_
c
a
f(x) dx.
We used a change of variables for the second equality and the fact
that F is continuous for the third equality. Therefore
_
c
a
[F
(x) −f(x)] dx = 0
for all c, which implies F
= f a.e. by Corollary 8.3.
We continue to assume f is nonnegative but now allow f to be
unbounded. Since f −(f ∧ K) ≥ 0, then
G
K
(x) =
_
x
a
[f −(f ∧ K)] dx
is increasing, and hence has a derivative almost everywhere. More
over,
G
K
(x) = lim
n→∞
G
K
(x + 1/n) −G
K
(x)
1/n
≥ 0
at points x where G
exists since G is increasing. By the preceding
paragraph, we know the derivative of
H
K
(x) =
_
x
a
(f ∧ K) dx
is equal to f ∧ K almost everywhere. Therefore
F
(x) = G
K
(x) +H
K
(x) ≥ (f ∧ K)(x), a.e.
Since K is arbitrary, F
≥ f a.e., and so
_
b
a
F
≥
_
b
a
f = F(b) −F(a).
Combining with (14.4) we conclude that
_
b
a
[F
− f] = 0. Since
F
−f ≥ 0 a.e., this tells us that F
= f a.e.
126 CHAPTER 14. DIFFERENTIATION
14.7 Approach 2 – absolute continuity
Finally, we continue the alternative approach to look at when
_
b
a
F
(y) dy = F(b) −F(a).
We refer the reader to Lemma 14.12 for the proof that if f is
absolutely continuous, then it is of bounded variation.
Lemma 14.18 If f is absolutely continuous on [a, b] and f
(x) = 0
a.e., then f is constant.
The CantorLebesgue function is an example to show that we
need the absolute continuity.
Proof. Let c ∈ [a, b], let E = ¦x ∈ [a, c] : f
(x) = 0¦, and let
ε > 0. Choose δ such that
K
i=1
[f(b
i
) − f(a
i
)[ < ε whenever
K
i=1
[b
i
− a
i
[ ≤ δ and the (a
i
, b
i
) are disjoint intervals. For each
point x ∈ E∩[a, c) there exist arbitrarily small intervals [x, x+h] ⊂
[a, c] such that [f(x + h) − f(x)[ < εh. By Lemma 14.15 we can
ﬁnd a ﬁnite disjoint collection of such intervals that cover all of
E except for a set of measure less than δ. We label the intervals
[a
i
, b
i
] so that a
i
< b
i
≤ a
i+1
. Except for a set of measure less
than δ, E is covered by ∪
i
(a
i
, b
i
). This implies that ∪
i
(b
i
, a
i+1
)
has measure less than δ, or
i
[a
i+1
− b
i
[ ≤ δ. By our choice of δ
and the deﬁnition of absolute continuity,
i
[f(a
i+1
) −f(b
i
)[ < ε.
On the other hand, by our choice of the intervals (a
i
, b
i
),
i
[f(b
i
) −f(a
i
)[ < ε
i
(b
i
−a
i
) ≤ ε(c −a).
Adding these two inequalities together,
[f(c) −f(a)[ =
¸
¸
¸
i
[f(a
i+1
) −f(b
i
)] +
i
[f(b
i
) −f(a
i
)]
¸
¸
¸
≤ ε +ε(c −a).
Since ε is arbitrary, then f(c) = f(a), which implies that f is
constant.
14.8. EXERCISES 127
Theorem 14.19 If F is absolutely continuous, then
F(b) −F(a) =
_
b
a
F
(y) dy.
Proof. Suppose F is absolutely continuous on [a, b]. Then F is of
bounded variation, so F = F
1
−F
2
where F
1
and F
2
are increasing,
and F
exists a.e. Since [F
(x)[ ≤ F
1
(x) +F
2
(x), then
_
[F
(x)[ dx ≤ (F
1
(b) +F
2
(b)) −(F
1
(a) −F
2
(a)),
and hence F
is integrable. If
G(x) =
_
x
a
F
(t) dt,
then G is absolutely continuous by Exercise 14.2, and hence F −G
is absolutely continuous. Then (F −G)
= F
−G
= F
−F
= 0
a.e., using Theorem 14.17 for the second equality. By Lemma 14.18,
F − G is constant, and thus F(x) − G(x) = F(a) − G(a). We
conclude
F(x) =
_
x
a
F
(t) dt +F(a).
If we set x = b, we get our result.
14.8 Exercises
Exercise 14.1 (1) Show that if f and g are absolutely continuous
on an interval [a, b], then the product fg is also.
(2) Prove the integration by parts formula:
fb)g(b) −f(a)g(a) =
_
b
a
f(x)g
(x) dx +
_
b
a
f
(x)g(x) dx.
Exercise 14.2 If f is integrable and realvalued, a ∈ R, and
F(x) =
_
x
a
f(y) dy,
prove that F is of bounded variation and is absolutely continuous.
128 CHAPTER 14. DIFFERENTIATION
Exercise 14.3 Suppose that f is a realvalued continuous function
on [0, 1] and that ε > 0. Prove that there exists a continuous
function g such that g
(x) exists and equals 0 for a.e. x and
sup
x∈[0,1]
[f(x) −g(x)[ < ε.
Exercise 14.4 Suppose f is a realvalued continuous function on
[0, 1] and f is absolutely continuous on (a, 1] for every a ∈ (0, 1). Is
f necessarily absolutely continuous on [0, 1]? If f is also of bounded
variation on [0, 1], is f absolutely continuous on [0, 1]? If not, give
counterexamples.
Exercise 14.5 A realvalued function f is Lipschitz with constant
M if
[f(x) −f(y)[ ≤ M[x −y[
for all x, y ∈ R. Prove that f is Lipschitz with constant M if and
only if f is absolutely continuous and [f
[ ≤ M a.e.
Exercise 14.6 Suppose F
n
is a sequence of increasing nonnega
tive right continuous functions on [0, 1] such that sup
n
F
n
(1) < ∞.
Let F =
∞
n=1
F
n
. Prove that
F
(x) =
∞
n=1
F
n
(x)
for almost every x.
Exercise 14.7 Suppose f is absolutely continuous on [0, 1] and
for A ⊂ [0, 1] we let f(A) = ¦f(x) : x ∈ A¦. Prove that if A has
Lebesgue measure 0, then f(A) has Lebesgue measure 0.
Exercise 14.8 If f is realvalued and diﬀerentiable at each point
of [0, 1], is f necessarily absolutely continuous on [0, 1]? If not, ﬁnd
a counterexample.
Exercise 14.9 Find an increasing function f such that f
= 0 a.e.
but f is not constant on any open interval.
Exercise 14.10 If f : [a, b] → R is continuous, let M(y) be the
number of points x in [a, b] such that f(x) = y. M(y) may be
ﬁnite or inﬁnite. Prove that M is Borel measurable and
_
M(y) dy
equals the total variation of f on [a, b].
14.8. EXERCISES 129
Exercise 14.11 Let α ∈ (0, 1). Find a Borel subset E of [−1, 1]
such that
lim
r→0+
m(E ∩ [−r, r])
2r
= α.
Exercise 14.12 Suppose f is a realvalued continuous function
on [a, b] and the derivate D
+
f is nonnegative on [a, b]. Prove that
f(b) ≥ f(a). What if instead we have that D
+
f is nonnegative on
[a, b]?
Exercise 14.13 Let
f(x) =
_
∞
−∞
e
−xy
2
1 +y
2
dy.
(1) Find the derivative of f.
(2) Find an ordinary diﬀerential equation that f solves. Find the
solution to this ordinary diﬀerential equation to determine an ex
plicit value for f(x).
Exercise 14.14 Let (X, /, µ) be a measure space and let f be a
realvalued integrable function. Deﬁne
g(x) =
_
[f(y) −x[ µ(dy)
for x ∈ R.
(1) Prove that g is absolutely continuous.
(2) Prove that lim
x→∞
g(x) = ∞ and lim
x→−∞
g(x) = ∞.
(3) Find g
(x) and prove that g(x
0
) = inf
x∈R
g(x) if and only if
µ(¦y : f(y) > x
0
¦) = µ(¦y : f(y) < x
0
¦).
Exercise 14.15 Suppose A ⊂ [0, 1] has Lebesgue measure zero.
Find an increasing function f : [0, 1] → R that is absolutely con
tinuous, but
lim
h→0
f(x +h) −f(x)
h
= ∞
for each x ∈ A.
Exercise 14.16 Suppose that µ is a measure on the Borel σ
algebra on [0, 1] and for every f that is realvalued and continuously
diﬀerentiable we have
¸
¸
¸
_
f
(x) µ(dx)
¸
¸
¸ ≤
_
_
1
0
f(x)
2
dx
_
1/2
.
130 CHAPTER 14. DIFFERENTIATION
(1) Show that µ is absolutely continuous with respect to Lebesgue
measure on [0, 1].
(2) If g is the RadonNikodym derivative of µ with respect to
Lebesgue measure, prove that there exists a constant c > 0 such
that
[g(x) −g(y)[ ≤ c[x −y[
1/2
, x, y ∈ [0, 1].
Exercise 14.17 Let p > 1 and f, g ∈ L
p
(R). Deﬁne
H(t) =
_
∞
−∞
[f(x) +tg(x)[
p
dx
for t ∈ R. Prove that H is a diﬀerentiable function and ﬁnd its
derivative.
Chapter 15
L
p
spaces
We introduce some spaces of functions, called the L
p
spaces. We
deﬁne the L
p
norm of a function, prove completeness of the norm,
discuss convolutions, and consider the bounded linear functionals
on L
p
. We assume throughout this chapter that the measure µ is
σﬁnite.
15.1 Norms
Let (X, /, µ) be a σﬁnite measure space. For 1 ≤ p < ∞, deﬁne
the L
p
norm of f by
f
p
=
_
_
[f(x)[
p
dµ
_
1/p
. (15.1)
For p = ∞, deﬁne the L
∞
norm of f by
f
∞
= inf¦M : µ(¦x : [f(x)[ ≥ M¦) = 0¦. (15.2)
Thus the L
∞
norm of a function f is the smallest number M such
that [f[ ≤ M a.e.
For 1 ≤ p ≤ ∞ the space L
p
is the set ¦f : f
p
< ∞¦. One
can also write L
p
(X) or L
p
(µ) if one wants to emphasize the space
or the measure. It is clear that f
p
= 0 if and only if f = 0 a.e.
If 1 < p < ∞, we deﬁne q by
1
p
+
1
q
= 1
131
132 CHAPTER 15. L
P
SPACES
and call q the conjugate exponent of p.
Basic to the study of L
p
spaces is H¨ older’s inequality. Note that
when p = q = 2, this is the CauchySchwarz inequality.
Proposition 15.1 (H¨ older’s inequality) If 1 < p, q < ∞ and
p
−1
+q
−1
= 1, then
_
[fg[ dµ ≤ f
p
g
q
.
This also holds if p = ∞ and q = 1.
Proof. If M = f
∞
, then [f[ ≤ M a.e. and
_
fg ≤ M
_
[g[. The
case p = ∞ and q = 1 follows.
Now let us assume 1 < p, q < ∞. If f
p
= 0, then f = 0 a.e.
and
_
[fg[ = 0, so the result is clear if f
p
= 0 and similarly if
g
q
= 0. Let F(x) = [f(x)[/f
p
and G(x) = [g(x)[/g
q
. Note
F
p
= 1 and G
q
= 1, and it suﬃces to show that
_
FGdµ ≤ 1.
The second derivative of the function e
x
is again e
x
, which is
everywhere positive. Any function whose second derivative is ev
erywhere nonnegative is convex, so if 0 ≤ λ ≤ 1, we have
e
λa+(1−λ)b
≤ λe
a
+ (1 −λ)e
b
(15.3)
for every pair of reals a ≤ b. If F(x), G(x) ,= 0, let a = p log F(x),
b = q log G(x), λ = 1/p, and 1 − λ = 1/q. We then obtain from
(15.3) that
F(x)G(x) ≤
F(x)
p
p
+
G(x)
q
q
.
Clearly this inequality also holds if F(x) = 0 or G(x) = 0. Inte
grating,
_
FGdµ ≤
F
p
p
p
+
G
q
q
q
=
1
p
+
1
q
= 1.
This completes the proof.
One application of H¨ older’s inequality is to prove Minkowski’s
inequality, which is simply the triangle inequality for L
p
.
We ﬁrst need the following lemma:
15.1. NORMS 133
Lemma 15.2 If a, b ≥ 0 and 1 ≤ p < ∞, then
(a +b)
p
≤ 2
p−1
a
p
+ 2
p−1
b
p
.
Proof. The case a = 0 is obvious, so we assume a > 0. Dividing
both sides by a
p
, letting x = b/a, and setting
f(x) = 2
p−1
+ 2
p−1
x
p
−(1 +x)
p
,
the inequality we want to prove is equivalent to showing f(x) ≥ 0
for x ≥ 0. Note f(0) > 0, f(1) = 0, lim
x→∞
f(x) = ∞, and the
only solution to f
(x) = 0 on (0, ∞) is x = 1. We conclude that f
takes its minimum at x = 1 and hence f(x) ≥ 0 for x ≥ 0.
Proposition 15.3 (Minkowski’s inequality) If 1 ≤ p ≤ ∞, then
f +g
p
≤ f
p
+g
p
.
Proof. Since [(f + g)(x)[ ≤ [f(x)[ + [g(x)[, integrating gives the
case when p = 1. The case p = ∞ is also easy. Now let us suppose
1 < p < ∞. If f
p
or g
p
is inﬁnite, the result is obvious, so
we may assume both are ﬁnite. The inequality Lemma 15.2 with
a = [f(x)[ and b = [g(x)[ yields, after an integration,
_
[(f +g)(x)[
p
dµ ≤ 2
p−1
_
[f(x)[
p
dµ + 2
p−1
_
[g(x)[
p
dµ.
We therefore have f+g
p
< ∞. Clearly we may assume f+g
p
>
0.
Now write
[f +g[
p
≤ [f[ [f +g[
p−1
+[g[ [f +g[
p−1
and apply H¨older’s inequality with q = (1 −
1
p
)
−1
. We obtain
_
[f +g[
p
≤ f
p
_
_
[f +g[
(p−1)q
_
1/q
+g
p
_
_
[f +g[
(p−1)q
_
1/q
.
Since p
−1
+q
−1
= 1, then (p −1)q = p, so we have
f +g
p
p
≤
_
f
p
+g
p
_
f +g
p/q
p
.
134 CHAPTER 15. L
P
SPACES
Dividing both sides by f+g
p/q
p
and using the fact that p−(p/q) =
1 gives us our result.
Recall the deﬁnition of normed linear space from Chapter 1.
We would like to say that by virtue of Minkowski’s inequality, L
p
is a normed linear space. This is not quite right. The L
p
norm of a
function satisﬁes all the properties of a norm except that f
p
= 0
does not imply that f is the zero function, only that f = 0 a.e.
The procedure we follow to circumvent this is to say two functions
are equivalent if they diﬀer on a set of measure 0. This is an
equivalence relation for functions. We then deﬁne the space L
p
to
be the set of equivalence classes with respect to this equivalence
relation, and deﬁne f
p
to be the L
p
norm of any function in the
same equivalence class as f. We then have that  
p
is a norm on
L
p
. We henceforth keep this interpretation in the back of our minds
when we talk about a function being in L
p
; the understanding is
that we identify functions that are equal a.e.
Recall Deﬁnition 10.1: f
n
converges to f in L
p
if f
n
−f
p
→ 0
as n → ∞. In terms of L
p
norms, this is equivalent to f
n
−f
p
p
→ 0
as n → ∞.
Related to the deﬁnition of L
∞
is the following terminology.
Given a realvalued measurable function f, the essential supremum
and essential inﬁmum are deﬁned by
ess sup f = inf¦M : µ(¦x : f(x) > M¦) = 0¦
and
ess inf f = sup¦m : µ(¦x : f(x) < m¦) = 0¦.
15.2 Completeness
We show that the space L
p
viewed as a metric space is complete.
Theorem 15.4 If 1 ≤ p ≤ ∞, then L
p
is complete.
Proof. We will do only the case p < ∞ and leave the case p = ∞
as Exercise 15.1.
Step 1. Suppose f
n
is a Cauchy sequence in L
p
. Our ﬁrst step is
to ﬁnd a certain subsequence. Given ε = 2
−(j+1)
, there exists n
j
15.2. COMPLETENESS 135
such that if n, m ≥ n
j
, then f
n
− f
m

p
≤ 2
−(j+1)
. Without loss
of generality we may assume n
j
≥ n
j−1
for each j.
Step 2. Set n
0
= 0 and deﬁne f
0
to be identically 0. Our candidate
for the limit function will be
j
[f
n
j
−f
n
j−1
]. In this step we show
absolute convergence of this series. If
A
j
= ¦x : [f
n
j
(x) −f
n
j−1
(x)[ > 2
−j/2
¦,
then from Lemma 10.4, µ(A
j
) ≤ 2
−jp/2
. We have
µ(∩
∞
j=1
∪
∞
m=j
A
m
) = lim
j→∞
µ(∪
∞
m=j
A
m
) ≤ lim
j→∞
∞
m=j
µ(A
m
) = 0.
Thus, except for a set of measure 0, for each x there exists a j
(depending on x) such that x / ∈ ∪
∞
m=j
A
m
. Hence for each x (except
for those in the null set) there is a j
0
(depending on x) such that
if j ≥ j
0
, then [f
n
j
(x) −f
n
j−1
(x)[ ≤ 2
−j
.
Set
g
j
(x) =
j
m=1
[f
n
m
(x) −f
n
m−1
(x)[.
g
j
(x) increases in j for each x. The limit is ﬁnite for almost every
x because for almost every x, [f
n
m
(x) − f
n
m−1
(x)[ ≤ 2
−m
for m
large (how large depends on x).
Step 3. We deﬁne our function f. Set
f(x) =
∞
m=1
[f
n
m
(x) −f
n
m−1
(x)].
We showed in Step 2 that this series is absolutely convergent for
almost every x, so f is well deﬁned for a.e. x. Set f(x) = 0 for any
x where absolute convergence does not hold. We have
f(x) = lim
K→∞
K
m=1
[f
n
m
(x) −f
n
m−1
(x)] = lim
K→∞
f
n
K
(x)
since we have a telescoping series. By Fatou’s lemma,
f −f
n
j

p
p
=
_
[f −f
n
j
[
p
≤ liminf
K→∞
_
[f
n
K
−f
n
j
[
p
= liminf
K→∞
f
n
K
−f
n
j

p
p
≤ 2
(−j+1)p
.
136 CHAPTER 15. L
P
SPACES
Step 4. We have thus shown that f −f
n
j

p
→ 0 as j → ∞. It is
standard that a Cauchy sequence with a convergent subsequence
itself converges. Here is the proof in our case. Given ε > 0, there
exists N such that f
n
− f
m

p
< ε if m, n ≥ N. In particular,
f
n
j
−f
m

p
< ε if j is large enough. By Fatou’s lemma,
f −f
m

p
p
≤ liminf
j→∞
f
n
j
−f
m

p
p
≤ ε
p
if m ≥ N. This shows that f
m
converges to f in L
p
norm.
Next we show:
Proposition 15.5 The set of continuous functions with compact
support is dense in L
p
(R).
Proof. Suppose f ∈ L
p
. We have
_
[f −fχ
[−n,n]
[
p
→ 0 as n → ∞
by dominated convergence, the dominating function being [f[
p
.
Hence it suﬃces to approximate functions in L
p
that have compact
support. By writing f = f
+
−f
−
we may suppose f ≥ 0. Consider
simple functions s
m
increasing to f; then we have
_
[f −s
m
[
p
→ 0
by dominated convergence, so it suﬃces to approximate simple
functions with compact support. By linearity, it suﬃces to approx
imate characteristic functions with compact support. Given E, a
Borel measurable set contained in a bounded interval, and ε > 0,
we showed in Proposition 8.4 that there exists g continuous with
compact support and with values in [0, 1] such that
_
[g −χ
E
[ < ε.
Since [g − χ
E
[ ≤ 1, then
_
[g − χ
E
[
p
≤
_
[g − χ
E
[ < ε. This com
pletes the proof.
The same proof shows the following corollary.
Corollary 15.6 The set of continuous functions on [a, b] are dense
in the space L
2
([a, b]) with respect to L
2
([a, b]) norm.
15.3 Convolutions
The convolution of two measurable functions f and g is deﬁned by
f ∗ g(x) =
_
f(x −y)g(y) dy,
15.4. BOUNDED LINEAR FUNCTIONALS 137
provided the integral exists. By a change of variables, this is the
same as
_
f(y)g(x −y) dy, so f ∗ g = g ∗ f.
Proposition 15.7 If f, g ∈ L
1
, then f ∗ g is in L
1
and
f ∗ g
1
≤ f
1
g
1
. (15.4)
Proof. We have
_
[f ∗ g(x)[ dx ≤
_ _
[f(x −y)[ [g(y)[ dy dx. (15.5)
Since the integrand on the right is nonnegative, we can apply the
Fubini theorem to see that the right hand side is equal to
_ _
[f(x −y)[ dx[g(y)[ dy =
_ _
[f(x)[ dx[g(y)[ dy (15.6)
= f
1
g
1
.
The ﬁrst equality here follows by a change of variables (see Exercise
8.1). This together with (15.5) proves (15.4). From (15.4) we
conclude that f ∗ g is ﬁnite a.e.
15.4 Bounded linear functionals
A linear functional on L
p
is a map H from L
p
to R satisfying
H(f +g) = H(f) +H(g), H(af) = aH(f)
whenever f, g ∈ L
p
and a ∈ R. (One can also have complexvalued
linear functionals, but we do not consider them in this section. See,
however, Exercise 15.28.) H is a bounded linear functional if
H = sup¦[Hf[ : f
p
≤ 1¦ (15.7)
is ﬁnite. The dual space of L
p
is the collection of all bounded linear
functionals with norm given by (15.7). Our goal in this section is
to identify the dual of L
p
.
We deﬁne the signum function or sign function by
sgn (x) =
_
¸
_
¸
_
−1, x < 0;
0, x = 0;
1, x > 0.
138 CHAPTER 15. L
P
SPACES
Note xsgn (x) = [x[.
The following is very useful.
Theorem 15.8 For 1 < p < ∞ and p
−1
+q
−1
= 1,
f
p
= sup
_
_
fg dµ : g
q
≤ 1
_
. (15.8)
When p = 1, (15.8) holds if we take q = ∞, and if p = ∞, (15.8)
holds if we take q = 1.
Proof. The right hand side of (15.8) is less than the left hand
side by H¨older’s inequality. Thus we need only show that the right
hand side is greater than the left hand side.
Case 1: p = 1. Take g(x) = sgn f(x). Then g is bounded by 1 and
fg = [f[. This takes care of the case p = 1.
Case 2: p = ∞. If f
∞
= 0, the result is trivial, so suppose
f
∞
> 0. Since µ is σﬁnite, there exist sets F
n
increasing up to
X such that µ(F
n
) < ∞ for each n. If M = f
∞
, let a be any
ﬁnite real less than M. By the deﬁnition of L
∞
norm, the measure
of A
n
= ¦x ∈ F
n
: [f(x)[ > a¦ must be positive if n is suﬃciently
large. Let
g
n
(x) =
sgn f(x)χ
A
n
(x)
µ(A
n
)
.
Then the L
1
norm of g
n
is 1 and
_
fg
n
=
_
A
n
[f[/µ(A
n
) ≥ a. Since
a is arbitrary, the supremum on the right hand side of (15.8) must
be M.
Case 3: 1 < p < ∞. We may suppose f
p
> 0. Let F
n
be
measurable sets of ﬁnite measure increasing to X, q
n
a sequence of
nonnegative simple functions increasing to f
+
, r
n
a sequence of
nonnegative simple functions increasing to f
−
, and
s
n
(x) = (q
n
(x) −r
n
(x))χ
F
n
(x).
Then s
n
(x) → f(x) for each x, [s
n
(x)[ increases to [f(x)[ for each
x, each s
n
is a simple function, and s
n

p
< ∞ for each n. Then
s
n

p
→ f
p
by monotone convergence, whether or not f
p
is
ﬁnite. For n suﬃciently large, s
n

p
> 0.
Let
g
n
(x) = (sgn f(x))
[s
n
(x)[
p−1
s
n

p/q
p
.
15.4. BOUNDED LINEAR FUNCTIONALS 139
g
n
is again a simple function. Since (p −1)q = p, then
g
n

q
=
(
_
[s
n
[
(p−1)q
)
1/q
s
n

p/q
p
=
s
n

p/q
p
s
n

p/q
p
= 1.
On the other hand, since [f[ ≥ [s
n
[,
_
fg
n
=
_
[f[ [s
n
[
p−1
s
n

p/q
p
≥
_
[s
n
[
p
s
n

p/q
p
= s
n

p−(p/q)
p
.
Since p − (p/q) = 1, then
_
fg
n
≥ s
n

p
, which tends to f
p
.
This proves the right hand side of (15.8) is at least as large as the
left hand side.
The proof of Theorem 15.8 also establishes
Corollary 15.9 For 1 < p < ∞ and p
−1
+q
−1
= 1,
f
p
= sup
_
_
fg : g
q
≤ 1, g simple
_
.
Proposition 15.10 Suppose 1 < p < ∞, p
−1
+ q
−1
= 1, and
g ∈ L
q
. If we deﬁne H(f) =
_
fg for f ∈ L
p
, then H is a bounded
linear functional on L
p
and H = g
q
.
Proof. The linearity is obvious. That H ≤ g
q
follows by
H¨older’s inequality. Using Theorem 15.8 and writing
H = sup
f
p
≤1
[H(f)[ = sup
f
p
≤1
¸
¸
¸
_
fg
¸
¸
¸ ≥ sup
f
p
≤1
_
fg = g
q
completes the proof.
Theorem 15.11 Suppose 1 < p < ∞, p
−1
+ q
−1
= 1, and H is
a realvalued bounded linear functional on L
p
. Then there exists
g ∈ L
q
such that H(f) =
_
fg and g
q
= H.
This theorem together with Proposition 15.10 allows us to iden
tify the dual space of L
p
with L
q
.
140 CHAPTER 15. L
P
SPACES
Proof. Suppose we are given a bounded linear functional H on
L
p
. First suppose µ(X) < ∞. Deﬁne ν(A) = H(χ
A
). We will
show that ν is a measure, that ν ¸ µ and that g = dν/dµ is the
function we seek.
If A and B are disjoint, then
ν(A∪ B) = H(χ
A∪B
) = H(χ
A
+χ
B
)
= H(χ
A
) +H(χ
B
) = ν(A) +ν(B).
To show ν is countably additive, it suﬃces to show that if A
n
↑ A,
then ν(A
n
) → ν(A), and then use Exercise 3.1. But if A
n
↑ A,
then χ
A
n
→ χ
A
in L
p
, and so ν(A
n
) = H(χ
A
n
) → H(χ
A
) = ν(A);
we use here the fact that µ(X) < ∞. We conclude that ν is a
countably additive signed measure. Moreover, if µ(A) = 0, then
χ
A
= 0 a.e., hence ν(A) = H(χ
A
) = 0. Using Exercise 13.1, which
is the RadonNikodym theorem for signed measures, we see there
exists a realvalued integrable function g such that ν(A) =
_
A
g for
all sets A.
If s =
i
a
i
χ
A
i
is a simple function, by linearity we have
H(s) =
i
a
i
H(χ
A
i
) =
i
a
i
ν(A
i
) =
i
a
i
_
gχ
A
i
=
_
gs.
(15.9)
By Corollary 15.9 and (15.9),
g
q
= sup
_
_
gs : s
p
≤ 1, s simple
_
= sup¦H(s) : s
p
≤ 1, s simple¦ ≤ H.
If s
n
are simple functions tending to f in L
p
(see Exercise 15.2),
then H(s
n
) → H(f), while by H¨older’s inequality
¸
¸
¸
_
s
n
g −
_
fg
¸
¸
¸ =
¸
¸
¸
_
(s
n
−f)g
¸
¸
¸ ≤ s
n
−f
p
g
q
→ 0,
so
_
s
n
g →
_
fg. We thus have H(f) =
_
fg for all f ∈ L
p
, and
g
q
≤ H. By H¨older’s inequality, H ≤ g
q
.
In the case where µ is σﬁnite, but not necessarily ﬁnite, let
F
n
↑ X so that µ(F
n
) < ∞ for each n. Deﬁne functionals H
n
by
H
n
(f) = H(fχ
F
n
). Clearly each H
n
is a bounded linear functional
on L
p
. Applying the above argument, we see there exist g
n
such
that H
n
(f) =
_
fg
n
and g
n

q
= H
n
 ≤ H. It is easy to see
15.5. EXERCISES 141
that g
n
is 0 if x / ∈ F
n
. Moreover, by the uniqueness part of the
RadonNikodym theorem, if n > m, then g
n
= g
m
on F
m
. Deﬁne
g by setting g(x) = g
n
(x) if x ∈ F
n
. Then g is well deﬁned. By
Fatou’s lemma, g is in L
q
with a norm bounded by H. Note
fχ
F
n
→ f in L
p
by dominated convergence. Since H is a bounded
linear functional on L
p
, we have H
n
(f) = H(fχ
F
n
) → H(f). On
the other hand
H
n
(f) =
_
F
n
fg
n
=
_
F
n
fg →
_
fg
by dominated convergence. Thus H(f) =
_
fg. Again by H¨older’s
inequality H ≤ g
q
.
15.5 Exercises
Exercise 15.1 Show that L
∞
is complete.
Exercise 15.2 Prove that the collection of simple functions is
dense in L
p
.
Exercise 15.3 Prove the equality
_
[f(x)[
p
dx =
_
∞
0
pt
p−1
m(¦x : [f(x)[ ≥ t¦) dt
for p ≥ 1.
Exercise 15.4 Consider the measure space ([0, 1], B, m), where B
is the Borel σalgebra and m is Lebesgue measure, and suppose f
is a measurable function. Prove that f
p
→ f
∞
as p → ∞.
Exercise 15.5 When does equality hold in H¨ older’s inequality?
When does equality hold in the Minkowski inequality?
Exercise 15.6 Give an example to show that L
p
,⊂ L
q
in general
if 1 < p < q < ∞. Give an example to show that L
q
,⊂ L
p
in
general if 1 < p < q < ∞.
142 CHAPTER 15. L
P
SPACES
Exercise 15.7 Deﬁne
g
n
(x) = nχ
[0,n
−3
]
(x).
(1) Show that if f ∈ L
2
([0, 1]), then
_
1
0
f(x)g
n
(x) dx → 0
as n → ∞.
(2) Show that there exists f ∈ L
1
([0, 1]) such that
_
1
0
f(x)g
n
(x) dx
,→ 0.
Exercise 15.8 Suppose µ is a ﬁnite measure on the Borel subsets
of R such that
f(x) =
_
R
f(x +t) µ(dt), a.e.,
whenever f is realvalued, bounded, and integrable. Prove that
µ(¦0¦) = 1.
Exercise 15.9 Suppose µ is a measure with µ(X) = 1 and f ∈ L
r
for some r > 0, where we deﬁne L
r
for r < 1 exactly as in (15.1).
Prove that
lim
p→0
f
p
= exp
_
_
log [f[ dµ
_
,
where we use the convention that exp(−∞) = 0.
Exercise 15.10 Suppose 1 < p < ∞ and q is the conjugate expo
nent to p. Suppose f
n
→ f a.e. and sup
n
f
n

p
< ∞. Prove that
if g ∈ L
q
, then
lim
n→∞
_
f
n
g =
_
fg.
Does this extend to the case where p = 1 and q = ∞? If not, give
a counterexample.
Exercise 15.11 If f ∈ L
1
(R) and g ∈ L
p
(R) for some p ∈ [1, ∞),
prove that
f ∗ g
p
≤ f
1
g
p
.
Exercise 15.12 Suppose p ∈ (1, ∞) and q is its conjugate expo
nent. Prove that if f ∈ L
p
(R) and g ∈ L
q
(R), then f∗g is uniformly
continuous and f ∗ g(x) → 0 as x → ∞ and as x → −∞.
15.5. EXERCISES 143
Exercise 15.13 Show that if f and g are continuous with compact
support, then f ∗ g is continuous with compact support.
Exercise 15.14 Suppose f ∈ L
∞
(R), f
h
(x) = f(x +h), and
lim
h→0
f
h
−f
∞
= 0.
Prove that there exists a uniformly continuous function g on R such
that f = g a.e.
Exercise 15.15 Let p ∈ [1, ∞). Prove that f ∈ L
p
(µ) if and only
if
∞
n=1
(2
n
)
p
µ(¦x : [f(x)[ > 2
n
¦) < ∞.
Exercise 15.16 Suppose µ(X) = 1 and f and g are nonnegative
functions such that fg ≥ 1 a.e. Prove that
_
_
f dµ
__
_
g dµ
_
≥ 1.
Exercise 15.17 Suppose f : [1, ∞) →R, f(1) = 0, f
exists and is
continuous and bounded, and f
∈ L
2
([1, ∞)). Let g(x) = f(x)/x.
Show g ∈ L
2
([1, ∞)).
Exercise 15.18 Find an example of a measurable f : [1, ∞) →R
such that f(1) = 0, f
exists and is continuous and bounded, f
∈
L
1
([1, ∞)), but the function g(x) = f(x)/x is not in L
1
.
Exercise 15.19 Prove the generalized Minkowski inequality: If
(X, /, µ) and (Y, B, ν) are measure spaces, f is measurable with
respect to /B, and 1 < p < ∞, then
_
_
X
_
_
Y
[f(x, y)[ ν(dy)
_
p
µ(dx)
_
1/p
≤
_
Y
_
_
X
[f(x, y)[
p
µ(dx)
_
1/p
ν(dy).
Does this extend to the cases where p = 1 or p = ∞? If not, give
counterexamples.
If Y = ¦1, 2¦, ν(dy) = δ
1
(dy)+δ
2
(dy), where δ
1
and δ
2
are point
masses at 1 and 2, resp. and we let g
1
(x) = f(x, 1), g
2
(x) = f(x, 2),
we recover the usual Minkowski inequality, Proposition 15.3.
144 CHAPTER 15. L
P
SPACES
Exercise 15.20 Let α ∈ (0, 1) and K(x) = [x[
−α
for x ∈ R. Note
that K is not in L
p
for any p ≥ 1. Prove that if f is nonnegative,
realvalued, and integrable on R and
g(x) =
_
f(x −t)K(t) dt,
then g is ﬁnite a.e.
Exercise 15.21 Suppose p > 1 and q is its conjugate exponent,
f is an absolutely continuous function on [0, 1] with f
∈ L
p
, and
f(0) = 0. Prove that if g ∈ L
q
, then
_
1
0
[fg[ dx ≤
_
1
p
_
1/p
f

p
g
q
.
Exercise 15.22 Suppose f : R → R is in L
p
for some p > 1 and
also in L
1
. Prove there exist constants c > 0 and α ∈ (0, 1) such
that
_
A
[f(x)[ dx ≤ cm(A)
α
for every Borel measurable set A ⊂ R, where m is Lebesgue mea
sure.
Exercise 15.23 Suppose f : R → R is integrable and there exist
constants c > 0 and α ∈ (0, 1) such that
_
A
[f(x)[ dx ≤ cm(A)
α
for every Borel measurable set A ⊂ R, where m is Lebesgue mea
sure. Prove there exists p > 1 such that f ∈ L
p
.
Exercise 15.24 Suppose 1 < p < ∞, f : (0, ∞) →R, and f ∈ L
p
with respect to Lebesgue measure. Deﬁne
g(x) =
1
x
_
x
0
f(y) dy.
Prove that
g
p
≤
p
p −1
f
p
.
This is known as Hardy’s inequality.
15.5. EXERCISES 145
Exercise 15.25 Suppose (X, /, µ) is a measure space and suppose
K : X X → R is measurable with respect to / /. Suppose
there exists M < ∞ such that
_
X
[K(x, y)[ µ(dy) ≤ M
for each x and
_
X
[K(x, y)[ µ(dx) ≤ M
for each y. If f is measurable and realvalued, deﬁne
Tf(x) =
_
X
K(x, y)f(y) µ(dy)
if the integral exists.
(1) Show that Tf
1
≤ Mf
1
.
(2) If 1 < p < ∞, show that Tf
p
≤ Mf
p
.
Exercise 15.26 Suppose A and B are two Borel measurable sub
sets of R with ﬁnite strictly positive Lebesgue measure. Show that
χ
A
∗χ
B
is a continuous nonnegative function that is not identically
equal to 0.
Exercise 15.27 Suppose A and B are two Borel measurable sub
sets of R with strictly positive Lebesgue measure. Show that
C = ¦x +y : x ∈ A, y ∈ B¦
contains a nonempty open interval.
Exercise 15.28 Suppose 1 < p < ∞ and q is the conjugate ex
ponent of p. Prove that if H is a bounded complexvalued lin
ear functional on L
p
, then there exists a complexvalued measur
able function g ∈ L
q
such that H(f) =
_
fg for all f ∈ L
p
and
H = g
q
.
146 CHAPTER 15. L
P
SPACES
Chapter 16
Fourier transforms
Fourier transforms give a representation of a function in terms of
frequencies. There is a great deal known about them and their
applications. We give an introduction here.
16.1 Basic properties
If f is a complexvalued function and f ∈ L
1
(R
n
), deﬁne the
Fourier transform
´
f to be the function with domain R
n
and range
C given by
´
f(u) =
_
R
n
e
iu·x
f(x) dx, u ∈ R
n
. (16.1)
We are using u x for the standard inner product in R
n
. Various
books have slightly diﬀerent deﬁnitions. Some put a negative sign
and/or 2π before the iu x, some have a (2π)
−1
or a (2π)
−1/2
in
front of the integral. The basic theory is the same in any case.
Some basic properties of the Fourier transform are given by
Proposition 16.1 Suppose f and g are in L
1
. Then
(1)
´
f is bounded and continuous;
(2)
(f +g)(u) =
´
f(u) +´ g(u);
(3)
¯
(af)(u) = a
´
f(u) if a ∈ C;
(4) if a ∈ R
n
and f
a
(x) = f(x +a), then
´
f
a
(u) = e
−iu·a
´
f(u);
147
148 CHAPTER 16. FOURIER TRANSFORMS
(5) if a ∈ R
n
and g
a
(x) = e
ia·x
g(x), then ´ g
a
(u) =
´
f(u +a);
(6) if a is a nonzero real number and h
a
(x) = f(ax), then
´
h
a
(u) =
a
−n
´
f(u/a).
Proof. (1)
´
f is bounded because f ∈ L
1
and [e
iu·x
[ = 1. We have
´
f(u +h) −
´
f(u) =
_
_
e
i(u+h)·x
−e
iu·x
_
f(x) dx.
Then
[
´
f(u +h) −
´
f(u)[ ≤
_
¸
¸
e
iu·x
¸
¸
¸
¸
e
ih·x
−1
¸
¸
[f(x)[ dx.
The integrand is bounded by 2[f(x)[, which is integrable, and
e
ih·x
− 1 → 0 as h → 0. Thus the continuity follows by domi
nated convergence.
(2) and (3) are easy by a change of variables. (4) holds because
´
f
a
(u) =
_
e
iu·x
f(x +a) dx =
_
e
iu·(x−a)
f(x) dx = e
−iu·a
´
f(u)
by a change of variables. For (5),
´ g
a
(u) =
_
e
iu·x
e
ia·x
f(x) dx =
_
e
i(u+a)·x
f(x) dx =
´
f(u +a).
Finally for (6), by a change of variables,
´
h
a
(u) =
_
e
iu·x
f(ax) dx = a
−n
_
e
iu·(y/a)
f(y) dy
= a
−n
_
e
i(u/a)·y
f(y) dy = a
−n
´
f(u/a),
as required.
One reason for the usefulness of Fourier transforms is that they
relate derivatives and multiplication.
Proposition 16.2 Suppose f ∈ L
1
and x
j
f(x) ∈ L
1
, where x
j
is
the j
th
coordinate of x. Then
∂
´
f
∂u
j
(u) = i
_
e
iu·x
x
j
f(x) dx.
16.1. BASIC PROPERTIES 149
Proof. Let e
j
be the unit vector in the j
th
direction. Then
´
f(u +he
j
) −
´
f(u)
h
=
1
h
_
_
e
i(u+he
j
)·x
−e
iu·x
_
f(x) dx
=
_
e
iu·x
_
e
ihx
j
−1
h
_
f(x) dx.
Since
¸
¸
¸
1
h
_
e
ihx
j
−1
_¸
¸
¸ ≤ [x
j
[
and x
j
f(x) ∈ L
1
, the right hand side converges to
_
e
iu·x
ix
j
f(x) dx
by dominated convergence. Therefore the left hand side converges.
Of course, the limit of the left hand side is ∂
´
f/∂u
j
.
Proposition 16.3 Suppose f : R →R is integrable, f is absolutely
continuous, and f
is integrable. Then the Fourier transform of f
is −iu
´
f(u).
The higher dimensional version of this is left as Exercise 16.4.
Proof. Since f
is integrable,
[f(y) −f(x)[ ≤
_
y
x
[f
(z)[ dz → 0
as x, y → ∞ by dominated convergence. This implies that f(y
n
) is
a Cauchy sequence whenever y
n
→ ∞, and we conclude that f(y)
converges as y → ∞. Since f is integrable, the only possible value
for the limit is 0. The same is true for the limit as y → −∞.
By integration by parts (use Exercise 14.1 and a limit argu
ment),
´
f
(u) =
_
∞
−∞
e
iux
f
(x) dx = −
_
∞
−∞
iue
iux
f(x) dx
= −iu
´
f(u),
as desired.
Recall the deﬁnition of convolution given in Section 15.3. Recall
also (15.6), which says that
_ _
[f(x −y)[ [g(y)[ dxdy = f
1
g
1
. (16.2)
150 CHAPTER 16. FOURIER TRANSFORMS
Proposition 16.4 If f, g ∈ L
1
, then the Fourier transform of f ∗g
is
´
f(u)´ g(u).
Proof. We have
f ∗ g(u) =
_
e
iu·x
_
f(x −y)g(y) dy dx
=
_ _
e
iu·(x−y)
f(x −y) e
iu·y
g(y) dxdy
=
_
´
f(u)e
iu·y
g(y) dy =
´
f(u)´ g(u).
We applied the Fubini theorem in the second equality; this is valid
because as we see from (16.2), the absolute value of the integrand
is integrable. We used a change of variables to obtain the third
equality.
16.2 The inversion theorem
We want to give a formula for recovering f from
´
f. First we need
to calculate the Fourier transform of a particular function.
Proposition 16.5 (1) Suppose f
1
: R →R is deﬁned by
f
1
(x) =
1
√
2π
e
−x
2
/2
.
Then
´
f
1
(u) = e
−u
2
/2
.
(2) Suppose f
n
: R
n
→R is given by
f
n
(x) =
1
(2π)
n/2
e
−x
2
/2
.
Then
´
f
n
(u) = e
−u
2
/2
.
Proof. (1) may also be proved using contour integration, but let’s
give a (mostly) real variable proof. Let g(u) =
_
e
iux
e
−x
2
/2
dx.
Diﬀerentiate with respect to u. We may diﬀerentiate under the in
tegral sign because (e
i(u+h)x
−e
iux
)/h is bounded in absolute value
16.2. THE INVERSION THEOREM 151
by [x[ and [x[e
−x
2
/2
is integrable; therefore dominated convergence
applies. We then obtain
g
(u) = i
_
e
iux
xe
−x
2
/2
dx.
By integration by parts (see Exercise 14.1) this is equal to
−u
_
e
iux
e
−x
2
/2
dx = −ug(u).
Solving the diﬀerential equation g
(u) = −ug(u), we have
[log g(u)]
=
g
(u)
g(u)
= −u,
so log g(u) = −u
2
/2 +c
1
, and then
g(u) = c
2
e
−u
2
/2
. (16.3)
By Exercise 11.18, g(0) =
_
e
−x
2
/2
dx =
√
2π, so c
2
=
√
2π. Sub
stituting this value of c
2
in (16.3) and dividing both sides by
√
2π
proves (1).
For (2), since f
n
(x) = f
1
(x
1
) f
1
(x
n
) if x = (x
1
, . . . , x
n
), then
´
f
n
(u) =
_
_
e
i
j
u
j
x
j
f
1
(x
1
) f
1
(x
n
) dx
1
dx
n
=
´
f
1
(u
1
)
´
f
1
(u
n
) = e
−u
2
/2
.
This completes the proof.
One more preliminary is needed before proving the inversion
theorem.
Proposition 16.6 Suppose ϕ is in L
1
and
_
ϕ(x) dx = 1. Let
ϕ
δ
(x) = δ
−n
ϕ(x/δ).
(1) If g is continuous with compact support, then g∗ϕ
δ
converges
to g pointwise as δ → 0.
(2) If g is continuous with compact support, then g∗ϕ
δ
converges
to g in L
1
as δ → 0.
(3) If f ∈ L
1
, then f ∗ ϕ
δ
−f
1
→ 0 as δ → 0.
152 CHAPTER 16. FOURIER TRANSFORMS
Proof. (1) We have by a change of variables (Exercise 8.1) that
_
ϕ
δ
(y) dy = 1. Then
[g ∗ ϕ
δ
(x) −g(x)[ =
¸
¸
¸
_
(g(x −y) −g(x)) ϕ
δ
(y) dy
¸
¸
¸
=
¸
¸
¸
_
(g(x −δy) −g(x)) ϕ(y) dy
¸
¸
¸
≤
_
[g(x −δy) −g(x)[ [ϕ(y)[ dy.
Since g is continuous with compact support and hence bounded
and ϕ is integrable, the right hand side goes to zero by dominated
convergence, the dominating function being 2g
∞
ϕ.
(2) We now use the Fubini theorem to write
_
[g ∗ ϕ
δ
(x) −g(x)[ dx =
_
¸
¸
¸
_
(g(x −y) −g(x)) ϕ
δ
(y) dy
¸
¸
¸ dx
=
_
¸
¸
¸
_
(g(x −δy) −g(x)) ϕ(y) dy
¸
¸
¸ dx
≤
_ _
[g(x −δy) −g(x)[ [ϕ(y)[ dy dx
=
_ _
[g(x −δy) −g(x)[ dx[ϕ(y)[ dy.
Let
G
δ
(y) =
_
[g(x −δy) −g(x)[ dx.
By dominated convergence, for each y, G
δ
(y) tends to 0 as δ →
0, since g is continuous with compact support. Moreover G
δ
is
bounded in absolute value by 2g
1
. Using dominated convergence
again and the fact that ϕ is integrable, we see that
_
G
δ
(y) [ϕ(y)[ dy
tends to 0 as δ → 0.
(3) Let ε > 0. Let g be a continuous function with compact
support so that f −g
1
< ε. Let h = f −g. A change of variables
shows that ϕ
δ

1
= ϕ
1
. Observe
f ∗ ϕ
δ
−f
1
≤ g ∗ ϕ
δ
−g
1
+h ∗ ϕ
δ
−h
1
.
Also
h∗ϕ
δ
−h
1
≤ h
1
+h∗ϕ
δ

1
≤ h
1
+h
1
ϕ
δ

1
< ε(1+ϕ
1
).
16.2. THE INVERSION THEOREM 153
Therefore, using (2),
limsup
δ→0
f ∗ ϕ
δ
−f
1
≤ limsup
δ→0
h ∗ ϕ
δ
−h
1
≤ ε(1 +ϕ
1
).
Since ε is arbitrary, we have our conclusion.
Now we are ready to give the inversion formula. The proof seems
longer than one might expect it to be, but there is no avoiding the
introduction of the function H
a
or some similar function.
Theorem 16.7 Suppose f and
´
f are both in L
1
. Then
f(y) =
1
(2π)
n
_
e
−iu·y
´
f(u) du, a.e.
Proof. If g(x) = a
−n
k(x/a), then the Fourier transform of g is
´
k(au). Hence the Fourier transform of
1
a
n
1
(2π)
n/2
e
−x
2
/2a
2
is e
−a
2
u
2
/2
. If we let
H
a
(x) =
1
(2π)
n
e
−x
2
/2a
2
,
we have
´
H
a
(u) = (2π)
−n/2
a
n
e
−a
2
u
2
/2
.
We write
_
´
f(u)e
−iu·y
H
a
(u) du =
_ _
e
iu·x
f(x)e
−iu·y
H
a
(u) dxdu
=
_ _
e
iu·(x−y)
H
a
(u) duf(x) dx
=
_
´
H
a
(x −y)f(x) dx. (16.4)
We can interchange the order of integration because
_ _
[f(x)[ [H
a
(u)[ dxdu < ∞
154 CHAPTER 16. FOURIER TRANSFORMS
and [e
iu·x
[ = 1. The left hand side of the ﬁrst line of (16.4) con
verges to (2π)
−n
_
´
f(u)e
−iu·y
dy as a → ∞ by dominated conver
gence since H
a
(u) → (2π)
−n
and
´
f ∈ L
1
. The last line of (16.4) is
equal to
_
´
H
a
(y −x)f(x) dx = f ∗
´
H
a
(y), (16.5)
using that
´
H
a
is symmetric. But by Proposition 16.6, setting δ =
a
−1
, we see that f ∗
´
H
a
converges to f in L
1
as a → ∞.
16.3 The Plancherel theorem
The last topic that we consider is the Plancherel theorem.
Theorem 16.8 Suppose f is continuous with compact support.
Then
´
f ∈ L
2
and
f
2
= (2π)
−n/2

´
f
2
. (16.6)
Proof. First note that if we combine (16.4) and (16.5), take y = 0,
and use the symmetry of
´
H
a
, we obtain
_
´
f(u)H
a
(u) du = f ∗
´
H
a
(0). (16.7)
Let g(x) = f(−x), where a denotes the complex conjugate of a.
Since ab = ab,
´ g(u) =
_
e
iu·x
f(−x) dx =
_
e
−iu·x
f(−x) dx
=
_
e
iu·x
f(x) dx =
´
f(u).
The third equality follows by a change of variables. By (16.7) with
f replaced by f ∗ g,
_
f ∗ g(u)H
a
(u) du = f ∗ g ∗
´
H
a
(0). (16.8)
Since
f ∗ g(u) =
´
f(u)´ g(u) = [
´
f(u)[
2
, the left hand side of (16.8)
converges by monotone convergence to (2π)
−n
_
[
´
f(u)[
2
du as a →
16.4. EXERCISES 155
∞. Since f and g are continuous with compact support, then by
Exercise 15.13, f ∗ g is also, and so the right hand side of (16.8)
converges to f ∗ g(0) =
_
f(y)g(−y) dy =
_
[f(y)[
2
dy by Proposi
tion 16.6(2).
Remark 16.9 We can use Theorem 16.8 to deﬁne
´
f when f ∈ L
2
so that (16.6) will continue to hold. The set of continuous functions
with compact support is dense in L
2
by Proposition 15.5. Given a
function f in L
2
, choose a sequence ¦f
m
¦ of continuous functions
with compact support such that f
m
→ f in L
2
. Then f
m
−f
n

2
→
0 as m, n → ∞. By (16.6), ¦
´
f
m
¦ is a Cauchy sequence in L
2
, and
therefore converges to a function in L
2
, which we call
´
f.
Let us check that the limit does not depend on the choice of
the sequence. If ¦f
m
¦ is another sequence of continuous functions
with compact support converging to f in L
2
, then ¦f
m
−f
m
¦ is a
sequence of continuous functions with compact support converging
to 0 in L
2
. By (16.6),
´
f
m
−
´
f
m
converges to 0 in L
2
, and therefore
´
f
m
has the same limit as
´
f
m
. Thus
´
f is deﬁned uniquely up to
almost everywhere equivalence. By passing to the limit in L
2
on
both sides of (16.6), we see that (16.6) holds for f ∈ L
2
.
16.4 Exercises
Exercise 16.1 Find the Fourier transform of χ
[a,b]
and in partic
ular, ﬁnd the Fourier transform of χ
[−n,n]
.
Exercise 16.2 Find a realvalued function f ∈ L
1
such that
´
f / ∈
L
1
.
Exercise 16.3 Show that if f ∈ L
1
and f is everywhere strictly
positive, then [
´
f(y)[ <
´
f(0) for y ,= 0.
Exercise 16.4 If f is integrable, realvalued, and all the partial
derivatives f
j
= ∂f/∂x
j
are integrable, prove that the Fourier
transform of f
j
is given by
´
f
j
(u) = −iu
j
´
f(u).
156 CHAPTER 16. FOURIER TRANSFORMS
Exercise 16.5 Let o be the class of realvalued functions f on R
such that for every k ≥ 0 and m ≥ 0, [x[
m
[f
(k)
(x)[ → 0 as [x[ → ∞,
where f
(k)
is the k
th
derivative of f when k ≥ 1 and f
(0)
= f. The
collection o is called the Schwartz class. Prove that if f ∈ o, then
´
f ∈ o.
Exercise 16.6 The Fourier transform of a ﬁnite signed measure µ
on R
n
is deﬁned by
´ µ(u) =
_
e
iu·x
µ(dx).
Prove that if µ and ν are two ﬁnite signed measures on R
n
(with
respect to the completion of / /, where / is the Lebesgue σ
algebra on R) such that ´ µ(u) = ´ ν(u) for all u ∈ R
n
, then µ = ν.
Exercise 16.7 If f is realvalued and continuously diﬀerentiable
on R, prove that
_
_
[f[
2
dx
_
2
≤ 4
_
_
[xf(x)[
2
dx
__
_
[f
[
2
dx
_
.
Exercise 16.8 Prove Heisenberg’s inequality (which is very useful
in quantum mechanics): there exists c > 0 such that if a, b ∈ R
and f is in L
2
, then
_
_
(x−a)
2
[f(x)[
2
dx
__
_
(u−b)
2
[
´
f(u)[
2
du
_
≥ c
_
_
[f(x)[
2
dx
_
2
.
Find the best constant c.
Chapter 17
Riesz representation
In Chapter 4 we constructed measures on R. In this chapter we
will discuss how to construct measures on more general topological
spaces X.
If X is a topological space, let B be the Borel σalgebra and
suppose µ is a σﬁnite measure on (X, B). Throughout this chap
ter we will restrict our attention to realvalued functions. If f is
continuous on X, let us deﬁne
L(f) =
_
X
f dµ.
Clearly L is linear, and if f ≥ 0, then L(f) ≥ 0. The main topic
of this chapter is to prove a converse, the Riesz representation
theorem.
We need more hypotheses on X than just that it is a topological
space. For simplicity, throughout this chapter we suppose X is
a compact metric space. In fact, with almost no changes in the
proof, we could let X be a compact Hausdorﬀ space, and with only
relatively minor changes, we could even let X be a locally compact
Hausdorﬀ metric space. See Remark 17.1. But here we stick to
compact metric spaces.
We let ((X) be the collection of continuous functions from X
to R. Recall that the support of a function f is the closure of
¦x : f(x) ,= 0¦. We write supp (f) for the support of f. If G is an
157
158 CHAPTER 17. RIESZ REPRESENTATION
open subset of X, we deﬁne T
G
by
T
G
= ¦f ∈ ((X) : 0 ≤ f ≤ 1, supp (f) ⊂ G¦.
Observe that if f ∈ T
G
, then 0 ≤ f ≤ χ
G
, but the converse
does not hold. For example, if X = [−2, 2], G = (−1, 1), and
f(x) = (1 −x
2
)
+
, then 0 ≤ f ≤ χ
G
, but the support of f, which is
[−1, 1], is not contained in G.
17.1 Partitions of unity
The reason we take our set X to be a metric space is that if K ⊂
G ⊂ X, where K is compact and G is open, then there exists
f ∈ T
G
such that f is 1 on K. If we let
f(x) =
_
1 −
d(x, K)
δ/2
_
+
,
where d(x, K) = inf¦d(x, y) : y ∈ K¦ is the distance from x to K
and δ = inf¦d(x, y) : x ∈ K, y ∈ G
c
¦, then this f will do the job.
Remark 17.1 If X is a compact Hausdorﬀ space instead of a com
pact metric one, we can still ﬁnd such an f, that is, f ∈ T
G
with
f ≥ χ
K
when K ⊂ G, K is compact, and G is open. Urysohn’s
lemma is the result from topology that guarantees such an f ex
ists. (A Hausdorﬀ space X is one where if x, y ∈ X, x ,= y, there
exist disjoint open sets G
x
and G
y
with x ∈ G
x
and y ∈ G
y
. An
example of a compact Hausdorﬀ space that is not a metric space
and cannot be made into a metric space is [0, 1]
R
with the product
topology.) See [1] for details.
We will need the following proposition.
Proposition 17.2 Suppose K is compact and K ⊂ G
1
∪ ∪G
n
,
where the G
i
are open sets. There exist g
i
∈ T
G
i
for i = 1, 2, . . . , n
such that
n
i=1
g
i
(x) = 1 if x ∈ K.
The collection ¦g
i
¦ is called a partition of unity on K, subordi
nate to the cover ¦G
i
¦.
Proof. Let x ∈ K. Then x will be in at least one G
i
. Single points
are always compact, so there exists h
x
∈ T
G
i
such that h
x
(x) = 1.
17.2. THE REPRESENTATION THEOREM 159
Let N
x
= ¦y : h
x
(y) > 0¦. Since h
x
is continuous, then N
x
is open,
x ∈ N
x
, and N
x
⊂ G
i
.
The collection ¦N
x
¦ is an open cover for the compact set K, so
there exists a ﬁnite subcover ¦N
x
1
, . . . , N
x
m
¦. For each i, let
F
i
= ∪¦N
x
j
: N
x
j
⊂ G
i
¦.
Each F
i
is closed, and since X is compact, F
i
is compact. We have
F
i
⊂ G
i
. Let us choose f
i
∈ T
G
i
such that f
i
is 1 on F
i
.
Now deﬁne
g
1
= f
1
,
g
2
= (1 −f
1
)f
2
,
g
n
= (1 −f
1
)(1 −f
2
) (1 −f
n−1
)f
n
.
Clearly g
i
∈ T
G
i
. Note g
1
+ g
2
= 1 − (1 − f
1
)(1 − f
2
), and an
induction argument shows that
g
1
+ +g
n
= 1 −(1 −f
1
)(1 −f
2
) (1 −f
n
).
If x ∈ K, then x ∈ N
x
j
for some j, so x ∈ F
i
for some i. Then
f
i
(x) = 1, which implies
n
k=1
g
k
(x) = 1.
17.2 The representation theorem
Let L be a linear functional mapping ((X) to R. Thus L(f +g) =
L(f) + L(g) and L(af) = aL(f) if f, g ∈ ((X) and a ∈ R. L is a
positive linear functional if L(f) ≥ 0 whenever f ≥ 0 on X.
Here is the Riesz representation theorem. B is the Borel σ
algebra on X, that is, the smallest σalgebra that contains all the
open subsets of X.
Theorem 17.3 Let X be a compact metric space and L a positive
linear functional on ((X). Then there exists a measure µ on (X, B)
such that
L(f) =
_
f(y) µ(dy), f ∈ ((X). (17.1)
160 CHAPTER 17. RIESZ REPRESENTATION
We often write Lf for L(f). Since X is compact, taking f identi
cally equal to 1 in (17.1) shows that µ is a ﬁnite measure.
Proof. If G is open, let
(G) = sup¦Lf : f ∈ T
G
¦
and for E ⊂ X, let
µ
∗
(E) = inf¦(G) : E ⊂ G, G open¦.
Step 1 of the proof will be to show µ
∗
is an outer measure. Step
2 is to show that every open set is µ
∗
measurable. Step 3 is to
apply Theorem 4.6 to obtain a measure µ. Step 4 establishes some
regularity of µ and Step 5 shows that (17.1) holds.
Step 1. We show µ
∗
is an outer measure. The only function in T
∅
is the zero function, so (∅) = 0, and therefore µ
∗
(∅) = 0. Clearly
µ
∗
(A) ≤ µ
∗
(B) if A ⊂ B.
To show the countable subadditivity of µ
∗
, ﬁrst let G
1
, G
2
, . . .
be open sets. For any open set H we see that µ
∗
(H) = (H). Let
G = ∪
i
G
i
and let f be any element of T
G
. Let K be the support of
f. Then K is compact, ¦G
i
¦ is an open cover for K, and therefore
there exists n such that K ⊂ ∪
n
i=1
G
i
. Let ¦g
i
¦ be a partition of
unity for K subordinate to ¦G
i
¦
n
i=1
. Since K is the support of f,
we have f =
n
i=1
fg
i
. Since g
i
∈ T
G
i
and f is bounded by 1, then
fg
i
∈ T
G
i
. Therefore
Lf =
n
i=1
L(fg
i
) ≤
n
i=1
µ
∗
(G
i
) ≤
∞
i=1
µ
∗
(G
i
).
Taking the supremum over f ∈ T
G
,
µ
∗
(G) = (G) ≤
∞
i=1
µ
∗
(G
i
).
If A
1
, A
2
, . . . are subsets of X, let ε > 0, and choose G
i
open
such that (G
i
) ≤ µ
∗
(G
i
) +ε2
−i
. Then
µ
∗
(∪
∞
i=1
A
i
) ≤ µ
∗
(∪
∞
i=1
G
i
) ≤
∞
i=1
µ
∗
(G
i
) ≤
∞
i=1
µ
∗
(A
i
) +ε.
Since ε is arbitrary, countable subadditivity is proved, and we con
clude that µ
∗
is an outer measure.
17.2. THE REPRESENTATION THEOREM 161
Step 2. We show that every open set is µ
∗
measurable. Suppose G
is open and E ⊂ X. It suﬃces to show
µ
∗
(E) ≥ µ
∗
(E ∩ G) +µ
∗
(E ∩ G
c
), (17.2)
since the opposite inequality is true by the countable subadditivity
of µ
∗
.
First suppose E is open. Choose f ∈ T
E∩G
such that
L(f) > (E ∩ G) −ε/2.
Let K be the support of f. Since K
c
is open, we can choose
g ∈ T
E∩K
c such that L(g) > (E ∩ K
c
) − ε/2. Then f + g ∈ T
E
,
and
(E) ≥ L(f +g) = Lf +Lg ≥ (E ∩ G) +(E ∩ K
c
) −ε
= µ
∗
(E ∩ G) +µ
∗
(E ∩ K
c
) −ε
≥ µ
∗
(E ∩ G) +µ
∗
(E ∩ G
c
) −ε.
Since ε is arbitrary, (17.2) holds when E is open.
If E ⊂ X is not necessarily open, let ε > 0 and choose H open
such that E ⊂ H and (H) ≤ µ
∗
(E) +ε. Then
µ
∗
(E) +ε ≥ (H) = µ
∗
(H) ≥ µ
∗
(H ∩ G) +µ
∗
(H ∩ G
c
)
≥ µ
∗
(E ∩ G) +µ
∗
(E ∩ G
c
).
Since ε is arbitrary, (17.2) holds.
Step 3. Let B be the Borel σalgebra on X. By Theorem 4.6,
the restriction of µ
∗
to B, which we call µ, is a measure on B. In
particular, if G is open, µ(G) = µ
∗
(G) = (G).
Step 4. In this step we show that if K is compact, f ∈ ((X), and
f ≥ χ
K
, then L(f) ≥ µ(K). Let ε > 0 and deﬁne
G = ¦x : f(x) > 1 −ε¦,
which is open. If g ∈ T
G
, then g ≤ χ
G
≤ f/(1−ε), so (1−ε)
−1
f −
g ≥ 0. Because L is a positive linear functional, L((1−ε)
−1
f −g) ≥
0, which leads to Lg ≤ Lf/(1 − ε). This is true for all g ∈ T
G
,
hence
µ(K) ≤ µ(G) ≤
Lf
1 −ε
.
Since ε is arbitrary, µ(K) ≤ Lf.
162 CHAPTER 17. RIESZ REPRESENTATION
Step 5. We now establish (17.1). By writing f = f
+
− f
−
and
using the linearity of L, to show (17.1) for continuous functions
we may suppose f ≥ 0. Since X is compact, then f is bounded,
and multiplying by a constant and using linearity, we may suppose
0 ≤ f ≤ 1.
Let n ≥ 1 and let K
i
= ¦x : f(x) ≥ i/n¦. Since f is continuous,
each K
i
is a closed set, hence compact. K
0
is all of X. Deﬁne
f
i
(x) =
_
¸
_
¸
_
0, x ∈ K
c
i−1
;
f(x) −
i−1
n
, x ∈ K
i−1
−K
i
;
1
n
, x ∈ K
i
.
Note f =
n
i=1
f
i
and χ
K
i
≤ nf
i
≤ χ
K
i−1
. Therefore
µ(K
i
)
n
≤
_
f
i
dµ ≤
µ(K
i−1
)
n
,
and so
1
n
n
i=1
µ(K
i
) ≤
_
f dµ ≤
1
n
n−1
i=0
µ(K
i
). (17.3)
Let ε > 0 and let G be an open set containing K
i−1
such that
µ(G) < µ(K
i−1
) +ε. Then nf
i
∈ T
G
, so
L(nf
i
) ≤ µ(G) ≤ µ(K
i−1
) +ε.
Since ε is arbitrary, L(f
i
) ≤ µ(K
i−1
)/n. By Step 4, L(nf
i
) ≥
µ(K
i
), and hence
1
n
n
i=1
µ(K
i
) ≤ L(f) ≤
1
n
n−1
i=0
µ(K
i
). (17.4)
Comparing (17.3) and (17.4) we see that
¸
¸
¸L(f) −
_
f dµ
¸
¸
¸ ≤
µ(K
0
) −µ(K
n
)
n
≤
µ(X)
n
.
Since, as we saw above, µ(X) = L(1) < ∞ and n is arbitrary, then
(17.1) is established.
Example 17.4 If f is continuous on [a, b], let L(f) be the Rie
mann integral of f on the interval [a, b]. Then L is a positive linear
functional on (([a, b]). In this case, the measure whose existence is
given by the Riesz representation theorem is Lebesgue measure.
17.3. REGULARITY 163
Remark 17.5 Let X be a metric space, not necessarily compact.
A continuous function f vanishes at inﬁnity if given ε > 0 there
exists a compact set K such that [f(x)[ < ε if x / ∈ K. (
0
(X) is
the usual notation for the set of continuous functions vanishing at
inﬁnity. There is a version of the Riesz representation theorem for
(
0
(X). See [1] for details.
17.3 Regularity
We establish the following regularity property of measures on com
pact metric spaces.
Proposition 17.6 Suppose X is a compact measure space, B is
the Borel σalgebra, and µ is a ﬁnite measure on the measurable
space (X, B). If E ∈ B and ε > 0, there exists K ⊂ E ⊂ G such
that K is compact, G is open, µ(G − E) < ε, and µ(E − K) < ε.
(K and G depend on ε as well as on E.)
Proof. Let us say that a subset E ∈ B is approximable if given ε >
0 there exists K ⊂ E ⊂ G with K compact, G open, µ(G−E) < ε,
and µ(E−F) < ε. Let 1 be the collection of approximable subsets.
We will show 1 contains all the compact sets and 1 is a σalgebra,
which will prove that 1 = B, and thus establish the proposition.
If K is compact, let G
n
= ¦x : d(x, K) < 1/n¦. Then the G
n
are
open sets decreasing to K, and if n is large enough, µ(G
n
−K) < ε.
Thus every compact set is in 1.
If E is in 1and ε > 0, then choose K ⊂ E ⊂ Gwith K compact,
G open, µ(E−K) < ε, and µ(G−E) < ε. Then G
c
⊂ E
c
⊂ K
c
, G
c
is closed, hence compact, K
c
is open, µ(K
c
−E
c
) = µ(E−K) < ε,
and µ(E
c
−G
c
) = µ(G−E) < ε. Therefore 1 is closed under the
operation of taking complements.
Suppose E
1
, E
2
, . . . ∈ 1. For each i choose K
i
compact and G
i
open such that K
i
⊂ E
i
⊂ G
i
, µ(G
i
−E
i
) < ε2
−i
, and µ(E
i
−K
i
) <
ε2
−(i+1)
. Then ∪
∞
i=1
G
i
is open, contains ∪
∞
i=1
E
i
, and
µ(∪
i
G
i
−∪
i
E
i
) ≤
∞
i=1
µ(G
i
−E
i
) < ε.
164 CHAPTER 17. RIESZ REPRESENTATION
We see that ∪
∞
i=1
K
i
is contained in ∪
∞
i=1
E
i
and similarly,
µ(∪
∞
i=1
E
i
−∪
∞
i=1
K
i
) ≤
∞
i=1
µ(E
i
−K
i
) < ε/2.
Since ∪
n
i=1
K
i
increases to ∪
∞
i=1
K
i
, we can choose n large so that
µ(∪
∞
i=n+1
K
i
) < ε/2.
Then ∪
n
i=1
K
i
, being the ﬁnite union of compact sets, is compact,
is contained in ∪
∞
i=1
E
i
, and
µ(∪
∞
i=1
E
i
−∪
n
i=1
K
i
) < ε.
This proves that ∪
i
E
i
is in 1.
Since 1 is closed under the operations of taking complements
and countable unions and ∩
i
E
i
= (∪
i
E
c
i
)
c
, then 1 is also closed
under the operation of taking countable intersections. Therefore 1
is a σalgebra.
A measure is called regular if
µ(E) = inf¦µ(G) : G open, E ⊂ G¦
and
µ(E) = sup¦µ(K) : K compact, K ⊂ E¦
for all measurable E. An immediate consequence of what we just
proved is that ﬁnite measures on (X, B) are regular when X is a
compact metric space.
17.4 Exercises
Exercise 17.1 Suppose F is a closed subset of [0, 1] and we deﬁne
L(f) =
_
1
0
fχ
F
dx
for realvalued continuous functions f on [0, 1]. Prove that if µ is
the measure whose existence is given by the Riesz representation
theorem, then µ(A) = m(A∩ F), where m is Lebesgue measure.
17.4. EXERCISES 165
Exercise 17.2 Suppose X is a compact metric space and µ is a
ﬁnite regular measure on (X, B), where B is the Borel σalgebra.
Prove that if f is a realvalued measurable function and ε > 0,
there exists a closed set F such that µ(F
c
) < ε and the restriction
of f to F is a continuous function on F.
Exercise 17.3 Let C
1
([0, 1]) be the set of functions whose deriva
tive exists and is continuous on [0, 1]. Suppose L is a linear func
tional on C
1
([0, 1]) such that
[L(f)[ ≤ c
1
f
 +c
2
f
for all f ∈ C
1
([0, 1]), where c
1
and c
2
are positive constants and the
norm is the supremum norm. Show there exists a signed measure
µ on the Borel subsets of [0, 1] and a constant K such that
L(f) =
_
f
dµ +Kf(0), f ∈ C
1
([0, 1]).
Exercise 17.4 Suppose X and Y are compact metric spaces and
F : X → Y is a continuous map from X onto Y . If ν is a ﬁnite
measure on the Borel sets of Y , prove that there exists a measure
µ on the Borel sets of X such that
_
Y
f dν =
_
X
f ◦ F dµ
for all f that are continuous on Y .
Exercise 17.5 Let X be a compact metric space. Prove that
C(X) has a countable dense subset.
Exercise 17.6 Let X be a compact metric space and let B be the
Borel σalgebra on X. Let µ
n
be a sequence of ﬁnite measures
on (X, B) and let µ be another ﬁnite measure on (X, B). Suppose
µ
n
(X) → µ(X). Prove that the following are equivalent:
(1)
_
f dµ
n
→
_
f dµ whenever f is a continuous realvalued func
tion on X;
(2) limsup
n→∞
µ
n
(F) ≤ µ(F) for all closed subsets F of X;
(3) liminf
n→∞
µ
n
(G) ≥ µ(G) for all open subsets G of X;
(4) lim
n→∞
µ
n
(A) = µ(A) whenever A is a Borel subset of X such
that µ(∂A) = 0, where ∂A = A−A
o
is the boundary of A.
166 CHAPTER 17. RIESZ REPRESENTATION
Exercise 17.7 Let X be a compact metric space and let B be the
Borel σalgebra on X. Let µ
n
be a sequence of ﬁnite measures on
(X, B) and suppose sup
n
µ
n
(X) < ∞.
(1) Prove that if f ∈ C(X), there is a subsequence ¦n
j
¦ such that
_
f dµ
n
j
converges.
(2) Let A be a countable dense subset of C(X). Prove that there
is a subsequence ¦n
j
¦ such that
_
f dµ
n
j
converges for all f ∈ A.
(3) With ¦n
j
¦ as in (2), prove that
_
f dµ
n
j
converges for all f ∈
C(X).
(4) Let L(f) = lim
n
j
→∞
_
f dµ
n
j
. Prove that L(f) is a positive
linear functional on C(X). Conclude that there exists a measure
µ such that
_
f dµ
n
j
→
_
f dµ
for all f ∈ C(X).
Chapter 18
Banach spaces
Banach spaces are normed linear spaces that are complete. We
will give the deﬁnitions, discuss the existence of bounded linear
functionals, prove the Baire category theorem, and derive some
consequences such as the uniform boundedness theorem and the
open mapping theorem.
18.1 Deﬁnitions
The deﬁnition of normed linear space X over a ﬁeld of scalars F,
where F is either the real numbers or the complex numbers, was
given in Chapter 1. Recall that a normed linear space is a metric
space if we use the metric d(x, y) = x −y.
Deﬁnition 18.1 We deﬁne a Banach space to be a normed lin
ear space that is complete, that is, where every Cauchy sequence
converges.
A linear map is a map L from a normed linear space X to a
normed linear space Y satisfying L(x + y) = L(x) + L(y) for all
x, y ∈ X and L(αx) = αL(x) for all x ∈ X and α ∈ F. We will
sometimes write Lx for L(x). Since L(0) = L(0+0) = L(0) +L(0),
then L(0) = 0.
167
168 CHAPTER 18. BANACH SPACES
Deﬁnition 18.2 A linear map f from X to R is a real linear func
tional, a linear map from X to C a complex linear functional. f is
a bounded linear functional if
f = sup¦[f(x)[ : x ∈ X, x ≤ 1¦ < ∞.
Proposition 18.3 The following are equivalent.
(1) The linear functional f is bounded.
(2) The linear functional f is continuous.
(3) The linear functional f is continuous at 0.
Proof. [f(x) −f(y)[ = [f(x −y)[ ≤ f [x −y[, so (1) implies (2).
That (2) implies (3) is obvious. To show (3) implies (1), if f is not
bounded, there exists a sequence x
n
∈ X such that x
n
 = 1 for
each n, but [f(x
n
)[ → ∞. If we let y
n
= x
n
/[f(x
n
)[, then y
n
→ 0
but [f(x
n
)[ = 1 ,→ 0, contradicting (3).
18.2 The HahnBanach theorem
We want to prove that there are plenty of linear functionals, but
ﬁrst we need Zorn’s lemma, which is equivalent to the axiom of
choice.
If we have a set Y with a partial order “≤,” a linear ordered
subset X ⊂ Y is one such that if x, y ∈ X, then either x ≤ y
or y ≤ x (or both) holds. A linearly ordered subset X ⊂ Y has
an upper bound if there exists an element z of Y (but it is not
necessary that z ∈ X) such that x ≤ z for all x ∈ X. An element
z of Y is maximal if z ≤ y for y ∈ Y implies y = z.
Here is Zorn’s lemma.
Lemma 18.4 If Y is a partially ordered set and every linearly
ordered subset of Y has an upper bound, then Y has a maximal
element.
Now we give the HahnBanach theorem for real linear function
als.
18.2. THE HAHNBANACH THEOREM 169
Theorem 18.5 If M is a subspace of a normed linear space X and
f is a bounded real linear functional on M, then f can be extended
to a bounded linear functional F on X such that F = f.
Saying that F is an extension of f means that the domain of F
contains the domain of f and F(x) = f(x) if x is in the domain of
f.
Proof. If f = 0, then we take F to be identically 0, so we
may assume that f , = 0, and then by multiplying by a constant,
that f = 1. We ﬁrst show that we can extend f by at least one
dimension.
Choose x
0
∈ X − M and let M
1
be the vector space spanned
by M and x
0
. Thus M
1
consists of all vectors of the form x +λx
0
,
where x ∈ X and λ is real.
We have for all x, y ∈ M
f(x) −f(y) = f(x −y) ≤ x −y ≤ x −x
0
 +y −x
0
.
Hence
f(x) −x −x
0
 ≤ f(y) +y −x
0

for all x, y ∈ M. Choose α ∈ R such that
f(x) −x −x
0
 ≤ α ≤ f(y) +y −x
0

for all x, y ∈ M. Deﬁne f
1
(x + λx
0
) = f(x) + λα. This is clearly
an extension of f to M
1
.
We need to verify that the norm of f
1
is less than or equal to
1. Let x ∈ M and λ ∈ R. By our choice of α, f(x) −x−x
0
 ≤ α,
or f(x) − α ≤ x − x
0
, and α ≤ f(x) + x − x
0
, or f(x) − α ≥
−x −x
0
. Thus
[f(x) −α[ ≤ x −x
0
.
Replacing x by −x/λ and multiplying by [λ[, we get
[λ[ [ −f(x)/λ −α[ ≤ [λ[  −x/λ −x
0
,
or
[f
1
(x +λx
0
)[ = [f(x) +λα[ ≤ x +λx
0
,
which is what we wanted to prove.
We now establish the existence of an extension of f to all of X.
Let T be the collection of all linear extensions F of f satisfying
170 CHAPTER 18. BANACH SPACES
F ≤ 1. This collection is partially ordered by inclusion. That is,
if f
1
is an extension of f to a subspace M
1
and f
2
is an extension of
f to a subspace M
2
, we say f
1
≤ f
2
if M
1
⊂ M
2
. Since the union of
any increasing family of subspaces of X is again a subspace, then
the union of a linearly ordered subfamily of T lies in T. By Zorn’s
lemma, T has a maximal element, say, F
1
. By the construction
of the preceding two paragraphs, if the domain of F
1
is not all of
X, we can ﬁnd an extension, which would be a contradiction to F
1
being maximal. Therefore F
1
is the desired extension.
To get a version for complex valued linear functionals is quite
easy. Note that if f(x) = u(x) + iv(x), then the real part of f,
namely, u = Re f, is a real valued linear functional. Also, u(ix) =
Re f(ix) = Re if(x) = −v(x), so that v(x) = −u(ix), and hence
f(x) = u(x) −iu(ix).
Theorem 18.6 If M is a subspace of a normed linear space X
and f is a bounded complex linear functional on M, then f can be
extended to a bounded linear functional F on X such that F = f.
Proof. Assume without loss of generality that f = 1. Let
u = Re f. Note [u(x)[ ≤ [f(x)[ ≤ x. Now use the version of the
HahnBanach theorem for real linear functionals to ﬁnd a linear
functional U that is an extension of u to X such that U ≤ 1. Let
F(x) = U(x) −iU(ix).
It only remains to show that the norm of F is at most 1. Fix
x, and write F(x) = re
iθ
. Then
[F(x)[ = r = e
−iθ
F(x) = F(e
−iθ
x).
Since this quantity is real and nonnegative,
[F(x)[ = U(e
iθ
x) ≤ U e
−iθ
x ≤ x.
This holds for all x, so F ≤ 1.
As an application of the HahnBanach theorem, given a sub
space M and an element x
0
not in M, we can deﬁne f(x +λx
0
) =
λx
0
for x ∈ M, and then extend this linear functional to all of X.
Then f will be 0 on M but nonzero at x
0
.
Another application is to ﬁx x
0
,= 0, let f(λx
0
) = λx
0
, and
then extend f to all of X. Thus there exists a linear functional f
such that f(x
0
) = x
0
 and f = 1.
18.3. BAIRE’S THEOREM AND CONSEQUENCES 171
18.3 Baire’s theorem and consequences
We turn now to the Baire category theorem and some of its con
sequences. Recall that if A is a set, we use A for the closure of A
and A
o
for the interior of A. A set A is dense in X if A = X and
A is nowhere dense if (A)
o
= ∅.
The Baire category theorem is the following. Completeness of
the metric space is crucial to the proof.
Theorem 18.7 Let X be a complete metric space.
(1) If G
n
are open sets dense in X, then ∩
n
G
n
is dense in X.
(2) X cannot be written as the countable union of nowhere dense
sets.
Proof. We ﬁrst show that (1) implies (2). Suppose we can write
X as a countable union of nowhere dense sets, that is, X = ∪
n
E
n
where (E
n
)
o
= ∅. We let F
n
= E
n
, which is a closed set, and then
F
o
n
= ∅ and X = ∪
n
F
n
. Let G
n
= F
c
n
, which is open. Since F
o
n
= ∅,
then G
n
= X. Starting with X = ∪
n
F
n
and taking complements,
we see that ∅ = ∩
n
G
n
, a contradiction to (1).
We must prove (1). Suppose G
1
, G
2
, . . . are open and dense in
X. Let H be any nonempty open set in X. We need to show there
exists a point in H ∩ (∩
n
G
n
). We will construct a certain Cauchy
sequence ¦x
n
¦ and the limit point, x, will be the point we seek.
Let B(z, r) = ¦y ∈ X : d(z, y) < r¦, where d is the metric. Since
G
1
is dense in X, H∩G
1
is nonempty and open, and we can ﬁnd x
1
and r
1
such that B(x
1
, r
1
) ⊂ H ∩ G
1
and 0 < r
1
< 1. Suppose we
have chosen x
n−1
and r
n−1
for some n ≥ 2. Since G
n
is dense, then
G
n
∩ B(x
n−1
, r
n−1
) is open and nonempty, so there exists x
n
and
r
n
such that B(x
n
, r
n
) ⊂ G
n
∩ B(x
n−1
, r
n−1
) and 0 < r
n
< 2
−n
.
We continue and get a sequence x
n
in X. If m, n > N, then x
m
and x
m
both lie on B(x
N
, r
N
), and so d(x
m
, x
n
) < 2r
N
< 2
−N+1
.
Therefore x
n
is a Cauchy sequence, and since X is complete, x
n
converges to a point x ∈ X.
It remains to show that x ∈ H ∩ (∩
n
G
n
). Since x
n
lies in
B(x
N
, r
N
) if n > N, then x lies in each B(x
N
, r
N
), and hence in
each G
N
. Therefore x ∈ ∩
n
G
n
. Also,
x ∈ B(x
n
, r
n
) ⊂ B(x
n−1
, r
n−1
) ⊂ ⊂ B(x
1
, r
1
) ⊂ H.
172 CHAPTER 18. BANACH SPACES
Thus we have found a point x in H ∩ (∩
n
G
n
).
A set A ⊂ X is called meager or of the ﬁrst category if it is
the countable union of nowhere dense sets; otherwise it is of the
second category.
An important application of the Baire category theorem is the
BanachSteinhaus theorem, also called the uniform boundedness
theorem.
Theorem 18.8 Suppose X is a Banach space and Y is a normed
linear space. Let A be an index set and let ¦L
α
: α ∈ A¦ be a
collection of bounded linear maps from X into Y . Then either
there exists a M < ∞ such that L
α
 ≤ M for all α ∈ A or else
sup
α
L
α
x = ∞ for some x.
Proof. Let (x) = sup
α∈A
L
α
x. Let G
n
= ¦x : (x) > n¦.
We argue that G
n
is open. The map x → L
α
x is a continuous
function for each α since L
α
is a bounded linear functional. This
implies that for each α, the set ¦x : L
α
x > n¦ is open. Since
x ∈ G
n
if and only if for some α ∈ A we have L
α
x > n, we
conclude G
n
is the union of open sets, hence is open.
Suppose there exists N such that G
N
is not dense in X. Then
there exists x
0
and r such that B(x
0
, r) ∩ G
N
= ∅. This can
be rephrased as saying that if x − x
0
 ≤ r, then L
α
(x) ≤ N
for all α ∈ A. If y ≤ r, we have y = (x
0
+ y) − x
0
. Then
(x
0
+y) −x
0
 = y ≤ r, and hence L
α
(x
0
+y) ≤ N for all α.
Also, of course, x
0
−x
0
 = 0 ≤ r, and thus L
α
(x
0
) ≤ N for all
α. We conclude that if y ≤ r and α ∈ A,
L
α
y = L
α
((x
0
+y) −x
0
) ≤ L
α
(x
0
+y) +L
α
x
0
 ≤ 2N.
Consequently, sup
α
L
α
 ≤ M with M = 2N/r.
The other possibility, by the Baire category theorem, is that
every G
n
is dense in X, and in this case ∩
n
G
n
is dense in X. But
(x) = ∞ for every x ∈ ∩
n
G
n
.
The following theorem is called the open mapping theorem. It is
important that L be onto. A mapping L : X → Y is open if L(U)
is open in Y whenever U is open in X. For a measurable set A, we
let L(A) = ¦Lx : x ∈ A¦.
18.3. BAIRE’S THEOREM AND CONSEQUENCES 173
Theorem 18.9 Let X and Y be Banach spaces. A bounded linear
map L from X onto Y is open.
Proof. We need to show that if B(x, r) ⊂ X, then L(B(x, r)) con
tains a ball in Y . We will show L(B(0, r)) contains a ball centered
at 0 in Y . Then using the linearity of L, L(B(x, r)) will contain
a ball centered at Lx in Y . By linearity, to show that L(B(0, r))
contains a ball centered at 0, it suﬃces to show that L(B(0, 1))
contains a ball centered at 0 in Y .
Step 1. We show that there exists r such that B(0, r2
−n
) ⊂
L(B(0, 2
−n
)) for each n. Since L is onto, Y = ∪
∞
n=1
L(B(0, n)).
The Baire category theorem tells us that at least one of the sets
L(B(0, n)) cannot be nowhere dense. Since L is linear, L(B(0, 1))
cannot be nowhere dense. Thus there exist y
0
and r such that
B(y
0
, 4r) ⊂ L(B(0, 1)).
Pick y
1
∈ L(B(0, 1)) such that y
1
− y
0
 < 2r and let z
1
∈
B(0, 1) be such that y
1
= Lz
1
. Then B(y
1
, 2r) ⊂ B(y
0
, 4r) ⊂
L(B(0, 1)). Thus if y < 2r, then y +y
1
∈ B(y
1
, 2r), and so
y = −Lz
1
+ (y +y
1
) ∈ L(−z
1
+B(0, 1)).
Since z
1
∈ B(0, 1), then −z
1
+B(0, 1) ⊂ B(0, 2), hence
y ∈ L(−z
1
+B(0, 1)) ⊂ L(B(0, 2)).
By the linearity of L, if y < r, then y ∈ L(B(0, 1)). It follows
by linearity that if y < r2
−n
, then y ∈ L(B(0, 2
−n
)). This can
be rephrased as saying that if y < r2
−n
and ε > 0, then there
exists x such that x < 2
−n
and y −Lx < ε.
Step 2. Suppose y < r/2. We will construct a sequence ¦x
j
¦ by
induction such that y = L(
∞
j=1
x
j
). By Step 1 with ε = r/4, we
can ﬁnd x
1
∈ B(0, 1/2) such that y − Lx
1
 < r/4. Suppose we
have chosen x
1
, . . . , x
n−1
such that
_
_
_y −
n−1
j=1
Lx
j
_
_
_ < r2
−n
.
Let ε = r2
−(n+1)
. By Step 1, we can ﬁnd x
n
such that x
n
 < 2
−n
and
_
_
_y −
n
j=1
Lx
j
_
_
_ =
_
_
_
_
y −
n−1
j=1
Lx
j
_
−Lx
n
_
_
_ < r2
−(n+1)
.
174 CHAPTER 18. BANACH SPACES
We continue by induction to construct the sequence ¦x
j
¦. Let
w
n
=
n
j=1
x
j
. Since x
j
 < 2
−j
, then w
n
is a Cauchy sequence.
Since X is complete, w
n
converges, say, to x. But then x <
∞
j=1
2
−j
= 1, and since L is continuous, y = Lx. That is, if
y ∈ B(0, r/2), then y ∈ L(B(0, 1)).
18.4 Exercises
Exercise 18.1 Find a measure space (X, /, µ), a subspace Y of
L
1
(µ), and a bounded linear functional f on Y with norm 1 such
that f has two distinct extensions to L
1
(µ) and each of the exten
sions has norm equal to 1.
Exercise 18.2 Show that L
p
([0, 1]) is separable, that is, there is a
countable dense subset, if 1 ≤ p < ∞. Show that L
∞
([0, 1]) is not
separable.
Exercise 18.3 For k ≥ 1 and functions f : [0, 1] → R that are k
times diﬀerentiable, deﬁne
f
C
k = f
∞
+f

∞
+ +f
(k)

∞
,
where f
(k)
is the k
th
derivative of f. Let C
k
([0, 1]) be the collection
of k times diﬀerentiable functions f with f
C
k < ∞. Is C
k
(0, 1])
complete with respect to the norm  
C
k?
Exercise 18.4 Let α ∈ (0, 1). For f a realvalued continuous
function on [0, 1] deﬁne
f
C
α = sup
x∈[0,1]
[f(x)[ + sup
x,y∈[0,1],x=y
[f(x) −f(y)[
[x −y[
α
.
Let C
α
([0, 1]) be the set of functions f with f
C
α < ∞. Is
C
α
([0, 1]) complete with respect to the norm  
C
α?
Exercise 18.5 For positive integers n let
A
n
=
_
f ∈ L
1
([0, 1]) :
_
1
0
[f(x)[
2
dx ≤ n
_
.
Show that each A
n
is a closed subset of L
1
([0, 1]) with empty inte
rior.
18.4. EXERCISES 175
Exercise 18.6 Suppose L is a linear functional on a normed linear
space X. Prove that L is a bounded linear functional if and only
if the set ¦x ∈ X : L(x) = 0¦ is closed.
Exercise 18.7 A set A in a normed linear space is convex if
λx + (1 −λ)y ∈ A
whenever x, y ∈ A and λ ∈ [0, 1].
(1) Prove that if A is convex, then the closure of A is convex.
(2) Prove that the open unit ball in a normed linear space is convex.
(The open unit ball is the set of x such that x < 1.)
Exercise 18.8 The unit ball in a normed linear space X is strictly
convex if λx + (1 − λ)y < 1 whenever f = g = 1, f ,= g,
and λ ∈ (0, 1).
(1) Let (X, /, µ) be a measure space. Prove that the unit ball in
L
p
(µ) is strictly convex.
(2) Prove that the unit balls in L
1
(µ), L
∞
(µ), and C(X) are not
strictly convex provided X consists of more than one point.
Exercise 18.9 Let f
n
be a sequence of continuous functions on R
that converge at every point. Prove there exist an interval and a
number M such that sup
n
[f
n
[ is bounded by M on that interval.
Exercise 18.10 Suppose  
1
and  
2
are two norms such that
x
1
≤ x
2
for all x in a vector space X, and suppose X is
complete with respect to both norms. Prove that there exists a
positive constant c such that
x
2
≤ cx
1
for all x ∈ X.
Exercise 18.11 Suppose X and Y are Banach spaces.
(1) Let X Y be the set of ordered pairs (x, y) with
(x
1
+x
2
, y
1
+y
2
) = (x
1
, y
1
) + (x
2
, y
2
)
for each x
1
, x
2
∈ X and y
1
, y
2
∈ Y and c(x, y) = (cx, cy) if x ∈ R.
Deﬁne (x, y) = x +y. Prove that X Y is a Banach space.
(2) Let L be a linear map from X into Y such that if x
n
→ x in
176 CHAPTER 18. BANACH SPACES
X and Lx
n
→ y in Y , then y = Lx. Such a map is called a closed
map. Let G be the graph of L, deﬁned by G = ¦(x, y) : y = Lx¦.
Prove that G is a closed subset of X Y , hence is complete.
(3) Prove that the function (x, Lx) → x is continuous, oneone,
linear, and maps G onto X.
(4) Prove the closed graph theorem, which says that if L is a linear
map from one Banach space to another that is a closed map, then
L is a continuous map.
Exercise 18.12 Let X = C
1
([0, 1]) (deﬁned in Exercise 18.3) and
Y = C([0, 1]). Deﬁne D : X → Y by Df = f
. Show that D is a
closed map but not a bounded one.
Exercise 18.13 Let A be the set of realvalued continuous func
tions on [0, 1] such that
_
1/2
0
f(x) dx −
_
1
1/2
f(x) dx = 1.
Prove that A is a closed convex subset of C([0, 1]), but there does
not exist f ∈ A such that
f = inf
g∈A
g.
Exercise 18.14 Let A
n
be the subset of the realvalued continu
ous functions on [0, 1] given by
A
n
= ¦f : there exists x ∈ [0, 1] such that
[f(x) −f(y)[ ≤ n[x −y[ for all y ∈ [0, 1]¦.
(1) Prove that A
n
is nowhere dense in C([0, 1]).
(2) Prove that there exist functions f in C([0, 1]) which are nowhere
diﬀerentiable on [0, 1], that is, f
(x) does not exist at any point of
[0, 1].
Chapter 19
Hilbert spaces
Hilbert spaces are complete normed linear spaces that have an inner
product. This added structure allows one to talk about orthonor
mal sets. We will give the deﬁnitions and basic properties. As an
application we brieﬂy discuss Fourier series.
19.1 Inner products
Recall that if a is a complex number, then a represents the complex
conjugate. When a is real, a is just a itself.
Deﬁnition 19.1 Let H be a vector space where the set of scalars
F is either the real numbers or the complex numbers. H is an
inner product space if there is a map ¸, ¸ from H H to F such
that
(1) ¸y, x¸ = ¸x, y¸ for all x, y ∈ H;
(2) ¸x +y, z¸ = ¸x, z¸ +¸y, z¸ for all x, y, z ∈ H;
(3) ¸αx, y¸ = α¸x, y¸, for x, y ∈ H and α ∈ F;
(4) ¸x, x¸ ≥ 0 for all x ∈ H;
(5) ¸x, x¸ = 0 if and only if x = 0.
We deﬁne x = ¸x, x¸
1/2
, so that ¸x, x¸ = x
2
. From the
deﬁnitions it follows easily that ¸0, y¸ = 0 and ¸x, αy¸ = α¸x, y¸.
The following is the CauchySchwarz inequality. The proof is
the same as the one usually taught in undergraduate linear algebra
177
178 CHAPTER 19. HILBERT SPACES
classes, except for some complications due to the fact that we allow
the set of scalars to be the complex numbers.
Theorem 19.2 For all x, y ∈ H, we have
[¸x, y¸[ ≤ x y.
Proof. Let A = x
2
, B = [¸x, y¸[, and C = y
2
. If C = 0, then
y = 0, hence ¸x, y¸ = 0, and the inequality holds. If B = 0, the
inequality is obvious. Therefore we will suppose that C > 0 and
B ,= 0.
If ¸x, y¸ = Re
iθ
, let α = e
iθ
, and then [α[ = 1 and α¸y, x¸ =
[¸x, y¸[ = B. Since B is real, we have that α¸x, y¸ also equals
[¸x, y¸[.
We have for real r
0 ≤ x −rαy
2
= ¸x −rαy, x −rαy¸
= ¸x, x¸ −rα¸y, x¸ −rα¸x, y¸ +r
2
¸y, y¸
= x
2
−2r[¸x, y¸[ +r
2
y
2
.
Therefore
A−2Br +Cr
2
≥ 0
for all real numbers r. Since we are supposing that C > 0, we may
take r = B/C, and we obtain B
2
≤ AC. Taking square roots of
both sides gives the inequality we wanted.
From the CauchySchwarz inequality we get the triangle in
equality:
Proposition 19.3 For all x, y ∈ H we have
x +y ≤ x +y.
Proof. We write
x +y
2
= ¸x +y, x +y¸ = ¸x, x¸ +¸x, y¸ +¸y, x¸ +¸y, y¸
≤ x
2
+ 2x y +y
2
= (x +y)
2
,
19.1. INNER PRODUCTS 179
as desired.
The triangle inequality implies
x −z ≤ x −y +y −z.
Therefore   is a norm on H, and so if we deﬁne the distance
between x and y by x −y, we have a metric space.
Deﬁnition 19.4 A Hilbert space H is an inner product space that
is complete with respect to the metric d(x, y) = x −y.
Example 19.5 Let µ be a positive measure on a set X, let H =
L
2
(µ), and deﬁne
¸f, g¸ =
_
fg dµ.
This is easily seen to be a Hilbert space. To show the completeness
we use Theorem 15.4.
If we let µ be counting measure on the natural numbers, we
get what is known as the space
2
. An element of
2
is a sequence
a = (a
1
, a
2
, . . .) such that
∞
n=1
[a
n
[
2
< ∞ and if b = (b
1
, b
2
, . . .),
then
¸a, b¸ =
∞
n=1
a
n
b
n
.
We get another common Hilbert space, ndimensional Euclidean
space, by letting µ be counting measure on ¦1, 2, . . . , n¦.
Proposition 19.6 Let y ∈ H be ﬁxed. Then the functions x →
¸x, y¸ and x → x are continuous.
Proof. By the CauchySchwarz inequality,
[¸x, y¸ −¸x
, y¸[ = [¸x −x
, y¸[ ≤ x −x
 y,
which proves that the function x → ¸x, y¸ is continuous. By the
triangle inequality, x ≤ x −x
 +x
, or
x −x
 ≤ x −x
.
The same holds with x and x
reversed, so
[ x −x
 [ ≤ x −x
,
and thus the function x → x is continuous.
180 CHAPTER 19. HILBERT SPACES
19.2 Subspaces
Deﬁnition 19.7 A subset M of a vector space is a subspace if
M is itself a vector space with respect to the same operations of
addition and scalar multiplication. A closed subspace is a subspace
that is closed relative to the metric given by ¸, ¸.
For an example of a subspace that is not closed, consider
2
and let M be the collection of sequences for which all but ﬁnitely
many elements are zero. M is clearly a subspace. Let x
n
=
(1,
1
2
, . . . ,
1
n
, 0, 0, . . .) and x = (1,
1
2
,
1
3
, . . .). Then each x
n
∈ M,
x / ∈ M, and we conclude M is not closed because
x
n
−x
2
=
∞
j=n+1
1
j
2
→ 0
as n → ∞.
A set E ⊂ H is convex if λx+(1−λx) ∈ E whenever 0 ≤ λ ≤ 1
and x, y ∈ E.
Proposition 19.8 Each nonempty closed convex subset E of H
has a unique element of smallest norm.
Proof. Let δ = inf¦x : x ∈ E¦. Since x +y
2
= ¸x +y, x +y¸
and similarly for x − y
2
, x
2
, and y
2
, a simple calculation
shows that
x +y
2
+x −y
2
= 2x
2
+y
2
.
Dividing by 4, if x, y ∈ E, then
1
4
x −y
2
=
1
2
x
2
+
1
2
y
2
−
_
_
_
x +y
2
_
_
_
2
.
Since E is convex, if x, y ∈ E, then (x +y)/2 ∈ E, and we have
x −y
2
≤ 2x
2
+ 2y
2
−4δ
2
. (19.1)
Choose y
n
∈ E such that y
n
 → δ. Applying (19.1) with x
replaced by y
n
and y replaced by y
m
, we see that
y
n
−y
m

2
≤ 2y
n

2
+ 2y
m

2
−4δ
2
,
19.2. SUBSPACES 181
and the right hand side tends to 0 as m and n tend to inﬁnity.
Hence y
n
is a Cauchy sequence, and since H is complete, it con
verges to some y ∈ H. Since y
n
∈ E and E is closed, y ∈ E. Since
the norm is a continuous function, y = limy
n
 = δ.
If y
is another point with y
 = δ, then by (19.1) with x
replaced by y
we have y −y
 = 0, and hence y = y
.
We say x ⊥ y, or x is orthogonal to y, if ¸x, y¸ = 0. Let x
⊥
,
read “x perp,” be the set of all y in X that are orthogonal to x. If
M is a subspace, let M
⊥
be the set of all y that are orthogonal to
all points in M. It is clear from the linearity of the inner product
that x
⊥
is a subspace of H. The subspace x
⊥
is closed because
it is the same as the set f
−1
(¦0¦), where f(x) = ¸x, y¸, which is
continuous by Proposition 19.6. Also, it is easy to see that M
⊥
is
a subspace, and since
M
⊥
= ∩
x∈M
x
⊥
,
M
⊥
is closed. We make the observation that if z ∈ M ∩M
⊥
, then
z
2
= ¸z, z¸ = 0,
so z = 0.
The following is sometimes called the Riesz representation theo
rem, although usually that name is reserved for Theorem 17.3. To
motivate the theorem, consider the case where H is ndimensional
Euclidean space. Elements of R
n
can be identiﬁed with n 1
matrices and linear maps from R
n
to R
m
can be represented by
multiplication on the left by a mn matrix A. For bounded linear
functionals on H, m = 1, so A is 1 n, and the y of the next
theorem is the vector associated with the transpose of A.
Theorem 19.9 If L is a bounded linear functional on H, then
there exists a unique y ∈ H such that Lx = ¸x, y¸.
Proof. The uniqueness is easy. If Lx = ¸x, y¸ = ¸x, y
¸, then
¸x, y −y
¸ = 0 for all x, and in particular, when x = y −y
.
We now prove existence. If Lx = 0 for all x, we take y = 0.
Otherwise, let M = ¦x : Lx = 0¦, take z ,= 0 in M
⊥
, and let
y = αz where α = Lz/¸z, z¸. Notice y ∈ M
⊥
,
Ly =
Lz
¸z, z¸
Lz = [Lz[
2
/¸z, z¸ = ¸y, y¸,
182 CHAPTER 19. HILBERT SPACES
and y ,= 0.
If x ∈ H and
w = x −
Lx
¸y, y¸
y,
then Lw = 0, so w ∈ M, and hence ¸w, y¸ = 0. Then
¸x, y¸ = ¸x −w, y¸ = Lx
as desired.
19.3 Orthonormal sets
A subset ¦u
α
¦
α∈A
of H is orthonormal if u
α
 = 1 for all α and
¸u
α
, u
β
¸ = 0 whenever α, β ∈ A and α ,= β.
The GramSchmidt procedure from linear algebra also works in
inﬁnitely many dimensions. Suppose ¦x
n
¦
∞
n=1
is a linearly inde
pendent sequence, i.e., no ﬁnite linear combination of the x
n
is 0.
Let u
1
= x
1
/x
1
 and deﬁne inductively
v
N
= x
N
−
n−1
i=1
¸x
N
, u
i
¸u
i
,
u
N
= v
N
/v
N
.
We have ¸v
N
, u
i
¸ = 0 if i < N, so u
1
, . . . , u
N
are orthonormal.
Proposition 19.10 If ¦u
α
¦
α∈A
is an orthonormal set, then for
each x ∈ H,
α∈A
[¸x, u
α
¸[
2
≤ x
2
. (19.2)
This is called Bessel’s inequality. This inequality implies that
only ﬁnitely many of the summands on the left hand side of (19.2)
can be larger than 1/n for each n, hence only countably many of
the summands can be nonzero.
Proof. Let F be a ﬁnite subset of A. Let
y =
α∈F
¸x, u
α
¸u
α
.
19.3. ORTHONORMAL SETS 183
Then
0 ≤ x −y
2
= x
2
−¸x, y¸ −¸y, x¸ +y
2
.
Now
¸y, x¸ =
_
α∈F
¸x, u
α
¸u
α
, x
_
=
α∈F
¸x, u
α
¸¸u
α
, x¸ =
α∈F
[¸x, u
α
¸[
2
.
Since this is real, then ¸x, y¸ = ¸y, x¸. Also
y
2
= ¸y, y¸ =
_
α∈F
¸x, u
α
¸u
α
,
β∈F
¸x, u
β
¸u
β
_
=
α,β∈F
¸x, u
α
¸¸x, u
β
¸¸u
α
, u
β
¸
=
α∈F
[¸x, u
α
¸[
2
,
where we used the fact that ¦u
α
¦ is an orthonormal set and so
¸u
α
, u
β
¸ equals 0 if α ,= β and equals 1 if α = β. Therefore
0 ≤ y −x
2
= x
2
−
α∈F
[¸x, u
α
¸[
2
.
Rearranging,
α∈F
[¸x, u
α
¸[
2
≤ x
2
when F is a ﬁnite subset of A. If N is an integer larger than nx
2
,
it is not possible that [¸x, u
α
¸[
2
> 1/n for more than N of the α.
Hence [¸x, u
α
¸[
2
,= 0 for only countably many α. Label those α’s
as α
1
, α
2
, . . .. Then
α∈A
[¸x, u
α
¸[
2
=
∞
j=1
[¸x, u
α
j
¸[
2
= lim
J→∞
J
j=1
[¸x, u
α
j
¸[
2
≤ x
2
,
which is what we wanted.
Proposition 19.11 Suppose ¦u
α
¦
α∈A
is orthonormal. Then the
following are equivalent.
(1) If ¸x, u
α
¸ = 0 for each α ∈ A, then x = 0.
(2) x
2
=
α∈A
[¸x, u
α
¸[
2
for all x.
(3) For each x ∈ H, x =
α∈A
¸x, u
α
¸u
α
.
184 CHAPTER 19. HILBERT SPACES
We make a few remarks. When (1) holds, we say the orthonor
mal set is complete. (2) is called Parseval’s identity. In (3) the
convergence is with respect to the norm of H and implies that only
countably many of the terms on the right hand side are nonzero.
Proof. First we show (1) implies (3). Let x ∈ H. By Bessel’s
inequality and the remarks following the statement of Proposition
19.10 there can be at most countably many α such that [¸x, u
α
¸[
2
,=
0. Let α
1
, α
2
, . . . be an enumeration of those α. By Bessel’s in
equality, the series
i
[¸x, u
α
i
¸[
2
converges. Using that ¦u
α
¦ is an
orthonormal set,
_
_
_
n
j=m
¸x, u
α
j
¸u
α
j
_
_
_
2
=
n
j,k=m
¸x, u
α
j
¸¸x, u
α
k
¸¸u
α
j
, u
α
k
¸
=
n
j=m
[¸x, u
α
j
¸[
2
→ 0
as m, n → ∞. Thus
n
j=1
¸x, u
α
j
¸u
α
j
is a Cauchy sequence, and
hence converges. Let z =
∞
j=1
¸x, u
α
j
¸u
α
j
. Then ¸z −x, u
α
j
¸ = 0
for each α
j
. By (1), this implies z −x = 0.
We see that (3) implies (2) because
x
2
−
n
j=1
[¸x, u
α
j
¸[
2
=
_
_
_x −
n
j=1
¸x, u
α
j
¸u
α
j
_
_
_
2
→ 0.
That (2) implies (1) is clear.
Example 19.12 Take H =
2
= ¦x = (x
1
, x
2
, . . .) :
[x
i
[
2
< ∞¦
with ¸x, y¸ =
i
x
i
y
i
. Then ¦e
i
¦ is a complete orthonormal sys
tem, where e
i
= (0, 0, . . . , 0, 1, 0, . . .), i.e., the only nonzero coordi
nate of e
i
is the i
th
one.
A collection of elements ¦e
α
¦ is a basis for H if the set of ﬁnite
linear combinations of the e
α
is dense in H.
Proposition 19.13 Every Hilbert space has an orthonormal basis.
This means that (3) in Proposition 19.11 holds.
19.4. FOURIER SERIES 185
Proof. If B = ¦u
α
¦ is orthonormal, but not a basis, let V be the
closure of the linear span of B, that is, the closure with respect to
the norm in H of the set of ﬁnite linear combinations of elements
of B. Choose x ∈ V
⊥
, and if we let B
= B ∪ ¦x/x¦, then B
is
a basis that is strictly bigger than B.
It is easy to see that the union of an increasing sequence of
orthonormal sets is an orthonormal set, and so there is a maximal
one by Zorn’s lemma. By the preceding paragraph, this maximal
orthonormal set must be a basis, for otherwise we could ﬁnd a
larger basis.
19.4 Fourier series
An interesting application of Hilbert space techniques is to Fourier
series, or equivalently, to trigonometric series. For our Hilbert
space we take H = L
2
([0, 2π)) and let
u
n
=
1
√
2π
e
inx
for n an integer. (n can be negative.) Recall that
¸f, g¸ =
_
2π
0
f(x)g(x) dx
and f
2
=
_
2π
0
[f(x)[
2
dx.
It is easy to see that ¦u
n
¦ is an orthonormal set:
_
2π
0
e
inx
e
−imx
dx =
_
2π
0
e
i(n−m)x
dx = 0
if n ,= m and equals 2π if n = m.
Let T be the set of ﬁnite linear combinations of the u
n
. We
want to show that T is a dense subset of L
2
([0, 2π)). The ﬁrst step
is to show that the closure of T with respect to the supremum norm
is equal to the set of continuous functions f on [0, 2π) with f(0) =
f(2π). We will accomplish this by using the StoneWeierstrass
theorem, Theorem 1.7.
We identify the set of continuous functions on [0, ∞) that take
the same value at 0 and 2π with the continuous functions on the
186 CHAPTER 19. HILBERT SPACES
circle. To do this, let S = ¦e
iθ
: 0 ≤ θ < 2π¦ be the unit circle in C.
If f is continuous on [0, 2π) with f(0) = f(2π), deﬁne
¯
f : S → C
by
¯
f(e
iθ
) = f(θ). Note ¯ u
n
(e
iθ
) = e
inθ
.
Let
¯
T be the set of ﬁnite linear combinations of the ¯ u
n
. S is a
compact metric space. Since the complex conjugate of ¯ u
n
is ¯ u
−n
,
then
¯
T is closed under the operation of taking complex conjugates.
Since ¯ u
n
¯ u
m
= ¯ u
n+m
, it follows that T is closed under the operation
of multiplication. That it is closed under scalar multiplication and
addition is obvious. ¯ u
0
is identically equal to 1, so
¯
T vanishes at
no point. If θ
1
, θ
2
∈ S and θ
1
,= θ
2
, then θ
1
− θ
2
is not an integer
multiple of 2π, so
¯ u
1
(θ
1
)
¯ u
1
(θ
2
)
= e
i(θ
1
−θ
2
)
,= 1,
or ¯ u
1
(θ
1
) ,= ¯ u
1
(θ
2
). Therefore T separates points. By the Stone
Weierstrass theorem (Theorem 1.7), the closure of T with respect
to the supremum norm is equal to the set of continuous complex
valued functions on S.
If f ∈ L
2
([0, 2π)), then
_
[f −fχ
[1/m,2π−1/m]
[
2
→ 0
by dominated convergence as m → ∞. By Corollary 15.6 any func
tion in L
2
([1/m, 2π −1/m]) can be approximated in L
2
by contin
uous functions which have support in the interval [1/m, 2π −1/m].
By what we showed above, a continuous function with support in
[1/m, 2π − 1/m] can be approximated uniformly on [0, 2π) by el
ements of T. Finally, if g is continuous on [0, 2π) and g
m
→ g
uniformly on [0, 2π), then g
m
→ g in L
2
([0, 2π)) by dominated
convergence. Putting all this together proves that T is dense in
L
2
([0, 2π)).
It remains to show the completeness of the u
n
. If f is orthogonal
to each u
n
, then it is orthogonal to every ﬁnite linear combination,
that is, to every element of T. Since T is dense in L
2
([0, 2π)), we
can ﬁnd f
n
∈ T tending to f in L
2
. Then
f
2
= [¸f, f¸[ ≤ [¸f −f
n
, f¸[ +[¸f
n
, f¸[.
The second term on the right of the inequality sign is 0. The ﬁrst
term on the right of the inequality sign is bounded by f −f
n
 f
19.5. EXERCISES 187
by the CauchySchwarz inequality, and this tends to 0 as n →
∞. Therefore f
2
= 0, or f = 0, hence the ¦u
n
¦ are complete.
Therefore ¦u
n
¦ is a complete orthonormal system.
Given f in L
2
([0, 2π)), write
c
n
= ¸f, u
n
¸ =
_
2π
0
fu
n
dx =
1
√
2π
_
2π
0
f(x)e
−inx
dx,
the Fourier coeﬃcients of f. Parseval’s identity says that
f
2
=
n
[c
n
[
2
.
For any f in L
2
we also have
n≤N
c
n
u
n
→ f
as N → ∞ in the sense that
_
_
_f −
n≤N
c
n
u
n
_
_
_
2
→ 0
as N → ∞.
Using e
inx
= cos nx +i sin nx, we have
∞
n=−∞
c
n
e
inx
= A
0
+
∞
n=1
B
n
cos nx +
∞
n=1
C
n
sin nx,
where A
0
= c
0
, B
n
= c
n
+c
−n
, and C
n
= i(c
n
−c
−n
). Conversely,
using cos nx = (e
inx
+e
−inx
)/2 and sin nx = (e
inx
−e
−inx
)/2i,
A
0
+
∞
n=1
B
n
cos nx +
∞
n=1
C
n
sin nx =
∞
n=−∞
c
n
e
inx
if we let c
0
= A
0
, c
n
= B
n
/2 + C
n
/2i for n > 0 and c
n
= B
n
/2 −
C
n
/2i for n < 0. Thus results involving the u
n
can be transferred
to results for series of sines and cosines and vice versa.
19.5 Exercises
Exercise 19.1 For f, g ∈ L
2
([0, 1]), let ¸f, g¸ =
_
1
0
f(x)g(x) dx.
Let H = C([0, 1]) be the functions that are continuous on [0, 1]. Is
H a Hilbert space with respect to the norm deﬁned in terms of the
inner product ¸, ¸? Justify your answer.
188 CHAPTER 19. HILBERT SPACES
Exercise 19.2 Suppose H is a Hilbert space with a countable ba
sis. Suppose x
n
 → x as n → ∞and ¸x
n
, y¸ → ¸x, y¸ as n → ∞
for every y ∈ H. Prove that x
n
−x → 0 as n → ∞.
Exercise 19.3 Prove that if M is a closed subset of a Hilbert space
H, then (M
⊥
)
⊥
= M. Is this necessarily true if M is not closed?
If not, give a counterexample.
Exercise 19.4 Prove that if H is inﬁnitedimensional, that is, it
has no ﬁnite basis, then the closed unit ball in H is not compact.
Exercise 19.5 Suppose a
n
is a sequence of real numbers such that
∞
n=1
a
n
b
n
< ∞
whenever
∞
n=1
b
2
n
< ∞. Prove that
∞
n=1
a
2
n
< ∞.
Exercise 19.6 We say x
n
→ x weakly if ¸x
n
, y¸ → ¸x, y¸ for every
y in H. Prove that if x
n
is a sequence in H with sup
n
x
n
 ≤ 1,
then there is a subsequence ¦n
j
¦ and an element x of H with
x ≤ 1 such that x
n
j
converges to x weakly.
Exercise 19.7 If A is a measurable subset of [0, 2π], prove that
lim
n→∞
_
A
e
inx
dx = 0.
This is known as the RiemannLebesgue lemma.
Exercise 19.8 The purpose of Exercise 13.4 was to show that in
proving the RadonNikodym theorem, we can assume that ν(A) ≤
µ(A) for all measurable A. Assume for the current problem that
this is the case and that µ and ν are ﬁnite measures. We use this
to give an alternative proof of the RadonNikodym theorem.
For f realvalued and in L
2
with respect to µ, deﬁne L(f) =
_
f dν.
(1) Show that L is a bounded linear functional on L
2
(µ).
(2) Conclude by Theorem 19.9 that there exists a realvalued mea
surable function g in L
2
(µ) such that L(f) =
_
fg dµ for all f ∈
L
2
(µ). Prove that dν = g dµ.
19.5. EXERCISES 189
Exercise 19.9 Suppose f is a continuous realvalued function on
R such that f(x + 1) = f(x) for every x. Let γ be an irrational
number. Prove that
lim
n→∞
1
n
n
j=1
f(jγ) =
_
1
0
f(x) dx.
Exercise 19.10 If M is a closed subspace of a Hilbert space, let
x +M = ¦x +y : y ∈ M¦.
(1) Prove that x +M is a closed convex subset of H.
(2) Let Qx be the point of x+M of smallest norm and Px = x−Qx.
P is called the projection of x onto M. Prove that P and Q are
mappings of H into M and M
⊥
, respectively.
(3) Prove that P and Q are linear mappings.
(4) Prove that if x ∈ M, then Px = x and Qx = 0.
(5) Prove that if x ∈ M
⊥
, then Px = 0 and Qx = x.
(6) Prove that
x
2
= Px
2
+Qx
2
.
190 CHAPTER 19. HILBERT SPACES
Bibliography
[1]G.B. Folland. Real Analysis: Modern Techniques and their
Applications, 2nd ed. Wiley, New York, 1999.
[2]H. Royden. Real Analysis, 3rd ed. Macmillan, New York, 1988.
[3]W. Rudin. Principles of Mathematical Analysis, 3rd ed.
McGrawHill, New York, 1976.
[4]W. Rudin. Real and Complex Analysis, 3rd ed. McGrawHill,
New York, 1987.
191
Index
A
c
, 1
A−B, 1
A∆B, 1
A
i
↑, 1
A
i
↓, 1
A
o
, 3
A, 3
B(x, r), 3
C, 2
C
k
([0, 1]), 174
C
α
([0, 1]), 174
((X), 157
L
p
, 131
Q, 2
R, 2
x ∨ y, 2
x ∧ y, 2
x
+
, 2
x
−
, 2
z, 2
σ((), 8
absolutely continuous
function, 118
measures, 99
algebra, 7
almost everywhere, 39
antiderivative, 112
Baire category theorem, 171
Banach space, 167
BanachSteinhaus theorem, 172
basis, 184
Bessel’s inequality, 182
Borel σalgebra, 9
Borel measurable, 9
Borel sets, 9
bounded convergence
theorem, 50
bounded variation, 117
Cantor function, 28
Cantor set, 28
generalized, 29
Carath´eodory extension
theorem, 31
Carath´eodory’s theorem, 22
Cauchy sequence, 3
CauchySchwarz inequality, 177
characteristic function, 41
Chebyshev’s inequality, 76
closed, 3
closed graph theorem, 176
closed map, 176
closure, 3
complete
measure space, 15
metric space, 3
orthonormal set, 184
completion, 15
conditional expectation, 105
conjugate exponent, 132
convergence
almost everywhere, 75
almost uniform, 78
in L
p
, 75
192
INDEX 193
in measure, 75
convex
function, 67, 180
set, 175
strictly, 175
convolution, 136
countable additivity, 13
counting measure, 14
covering lemma, 108
decreasing, 2
strictly, 2
density, 101
derivates, 122
derivative, 107
diﬀerentiable, 107
dominated convergence
theorem, 55
Egorov’s theorem, 78
equivalence relationship, 4
equivalent measures, 104
essential inﬁmum, 134
essential supremum, 134
Fatou’s lemma, 54
ﬁnitely additive, 13
ﬁrst category, 172
Fourier series, 185
Fubini theorem, 85
FubiniTonelli theorem, 85
generates the σalgebra, 9
graph, 176
Hahn decomposition, 95
HahnBanach theorem, 168
Hardy’s inequality, 144
Heisenberg’s inequality, 156
Hilbert space, 179
increasing, 2
strictly, 2
indeﬁnite integral, 112
inner product space, 177
integrable, 48
integration by parts, 127
interior, 3
Jensen’s inequality, 67
jointly measurable, 86
Jordan decomposition theorem,
97
Lebesgue decomposition
theorem, 103
Lebesgue measurable, 27
Lebesgue measure, 21, 27
ndimensional, 86
Lebesgue σalgebra, 21, 27
LebesgueStieltjes measure, 21,
27
liminf, 2
limsup, 2
linear functional, 137
bounded, 137, 168
complex, 168
positive, 159
real, 168
linear map, 167
linear space, 4
Lipschitz, 128
locally integrable, 109
lower semicontinuous, 73
maximal function, 109
meager, 172
measurable, 7
measurable rectangle, 81
measurable space, 7
measure, 13
complex, 104
ﬁnite, 15
regular, 164
signed, 93
measure space, 13
194 INDEX
metric space, 3
Minkowski inequality
generalized, 143
monotone, 2
monotone class, 10
monotone class theorem, 10
monotone convergence theorem,
51
µ
∗
measurable, 22
mutually singular, 96
negative set, 93
normed linear space, 4
nowhere diﬀerentiable, 176
null set, 15, 20, 93
open, 3
open ball, 3
open mapping, 172
open mapping theorem, 172
orthogonal, 181
orthonormal, 182
outer measure, 20
pairwise disjoint, 13
Parseval’s identity, 184
partial order, 4
Plancherel theorem, 154
point mass, 14
points of density, 112
positive measure, 93
positive set, 93
premeasure, 32
probability, 15
product σalgebra, 81
projection, 189
RadonNikodym
theorem, 101, 188
RadonNikodym derivative, 101
RiemannLebesgue lemma, 188
Riesz representation
theorem, 159, 181
Schwartz class, 156
second category, 172
section, 81
separable, 174
separate points, 5
σalgebra, 7
generated by a collection, 8
σﬁeld, 15
σﬁnite, 15
sign function, 137
signed measure, 93
signum function, 137
simple function, 41
StoneWeierstrass theorem, 5
strictly convex, 175
subspace, 180
closed, 180
support, 3
measure, 35
symmetric diﬀerence, 1
total variation
of a function, 118
total variation measure, 97
trigonometric series, 185
uniform boundedness
theorem, 172
uniformly absolutely
continuous, 58
uniformly integrable, 58
upper semicontinuous, 73
vanish at no point, 5
vanishes at inﬁnity, 163
vector space, 4
Vitali convergence theorem, 59
Vitali cover, 120
Vitali covering lemma, 120
weak 11 inequality, 109
weak convergence, 188
Zorn’s lemma, 168
Real analysis for graduate students: measure and integration theory
Richard F. Bass
ii
c Copyright 2011 Richard F. Bass All rights reserved. ISBN13: 9781466391574 ISBN10: 146639157X
iii
To the memory of my parents Jay and Betty Bass
iv .
. .20 LebesgueStieltjes measures . . . . . . . . . . . . . . . . . . . . . . . . . .31 e Exercises . .4 4. . . . . . . . . . . . . . . . . . . . .24 Examples and related results . . . .2 4. . . . .1 1. . . . . . . . . . . . . .1 3.2 7 ix 1 2 Families of sets 2. . . . . . . . . . . .3 Algebras and σalgebras . . . . .3 4. . . . . . . . . . . .5 4. . . . . . .27 Nonmeasurable sets . . . . . . . .2 Notation and terminology . . . .2 2. . . . . . . . . . . .6 Outer measures . . .13 Exercises . . . . . . . .2 13 Deﬁnitions and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . .34 v . . . . .10 Exercises .1 4. . . . . . . . . . .30 The Carath´odory extension theorem . .1 2. . .7 The monotone class theorem . . . . . . . . . . . . . . . . . .11 3 Measures 3. . . . . . . . .Contents Preface 1 Preliminaries 1. . . . . .16 19 4 Construction of measures 4.1 Some undergraduate mathematics .
. . . . . . . .2 Deﬁnitions .51 Linearity of the integral . . . . . .75 10. . . . .1 5. . . . . . . . . . . . . . . . . . . . . .41 Exercises . .1 Deﬁnitions and examples .56 63 8 Properties of Lebesgue integrals 8. . . . . . . .2 Comparison with the Lebesgue integral .5 Monotone convergence theorem . .3 Criteria for a function to be zero a. .78 81 11 Product measures . . .1 6. . . . . . .49 51 7 Limit theorems 7. . . . . . .66 69 9 Riemann integrals 9. . . . .41 Lusin’s theorem .1 9. . . . . . . . . . . . . .4 CONTENTS 37 Measurability . . . . . . . . .2 Exercises . . . . . . . . . . .2 5. . . . . . . . . . . . . . .3 5.e. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .44 47 6 The Lebesgue integral 6. . . . . . . . . . . . . . . .55 Exercises . . . . . . . . . . . . . .69 Exercises . . .71 75 10 Types of convergence 10. . . . . . . . . . . . . .4 7. . . . . . . . . . . .vi 5 Measurable functions 5. . . . . . . . . . . . . . . . . . . . .2 8. .3 7. . . . . . . . . . . . . . . . . . . . . . . . .54 Dominated convergence theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .63 An approximation result . .52 Fatou’s lemma . . . . . . . . . . . . . .47 Exercises . . . . .65 Exercises . . .1 7. . . . . . . . . . . . .37 Approximation of functions . . . . . . .2 7. . . .1 8. .
. . . . . . . .2 The main theorem . . . . . . . . . .8 Exercises 15 Lp spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .120 14. . . . . . . . . . . . . . . . . . . . . . . .1 Norms .134 . . . . . . . .6 Approach 2 – antiderivatives . . . . . .131 15. . . . . .4 Exercises . . .118 14. . . . . . . .3 Bounded variation . . . . . . . . . . . . . . . .4 Exercises . . . . . .1 Absolute continuity . . . . . .93 12. . . . .2 Antiderivatives . .97 99 13 The RadonNikodym theorem 13. . . . . .100 13. . . . . .7 Approach 2 – absolute continuity . .4 Exercises . . . . . . . . . . . . . . . . . . . . .95 12.1 Positive and negative sets . . . . . . . . . . . . . . . . .127 131 15. . . . . . . .3 Lebesgue decomposition theorem . . . . . . . . . . . . .113 14.108 14. . . . .88 93 12 Signed measures 12. . . . . . . . . . . . . . . . . .126 14. . . . . . .81 11. . .104 107 14 Diﬀerentiation 14. .85 11. . . . . . . . . . . . . . . . . . . . . . .CONTENTS vii 11. . . .5 Approach 2 – diﬀerentiability . . . . . . . . . . . . . . . . .124 14. . . . . . . . . . . .1 Maximal functions . . .1 Product σalgebras . . . . . . . . . . . . . . . . . . . . . . . . .99 13. . . . . . .2 Completeness .87 11. . . .4 Absolutely continuous functions . .3 Jordan decomposition theorem . . . . . .97 12. . . . . . . . . . . . . . .112 14. .3 Examples . . . . . . . . . .103 13. . . . .2 Hahn decomposition theorem . . . . . . . . . . . . . .2 The Fubini theorem . .
. . .1 Basic properties . .136 15. . . . . . . .141 147 16 Fourier transforms 16. . . . . . . . . .180 19.167 18.4 Fourier series . . . . . . . . . . . . . . .147 16.158 17. . . . . . . . . . . . . . . .1 Inner products . . . . . . . . . . . . . . . . . . .3 Regularity . . . . .2 The HahnBanach theorem . . .4 Exercises . . . . . . . . . . . . . . . . . . . . . . .3 Orthonormal sets .163 17. . . . .171 . . . . . .4 Exercises 19 Hilbert spaces . . . . . . . . . .5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .187 . . . .4 Bounded linear functionals . . .3 Baire’s theorem and consequences 18. . . . . . .4 Exercises 18 Banach spaces . . . . . . . .177 19. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .150 16. . . .5 Exercises . .168 18. . . . .164 167 18. . . . . . . .2 The representation theorem . . . . . . . . . .137 15. . . . . . . . . . . . . . .2 The inversion theorem . . . . . . .182 19. . . . . . . . .2 Subspaces . . . . . .3 The Plancherel theorem .1 Deﬁnitions . . . . . . . . . . . . . . . .3 Convolutions . . . . . . . . . . . . . . . . . . . . .1 Partitions of unity . . . . . . . . . . . . . . . .185 19. . . . . . .154 16. . . . . . . .viii CONTENTS 15. . . . . . . .159 17. . . . . . . . . . . . . . . . . . . . . . . . .174 177 19. . . . . . .155 157 17 Riesz representation 17. . . . . . . . . . . . . . .
Preface Nearly every Ph. On the other hand. All the proofs are “plain vanilla. student in mathematics needs to pass a preliminary or qualifying examination in real analysis. Why write another? In my opinion. I present the theorems and results as they will be tested. a look at the index will show that no topics that might appear on a preliminary or qualifying examination are omitted. (1) There are a large number of real analysis texts already in existence. I have tried to make the writing as clear as possible and to provide the details. unusual approaches. not in the absolutely most general abstract context. (2) provide a large collection of useful exercises.” They are either too hard. none of the existing texts are ideally suited to the beginning graduate student who needs to pass a “prelim” or “qual. too encyclopedic.” I avoid any clever tricks. (3) make the text aﬀordable. omit too many important topics. sneaky proofs. Let me discuss each of these in more detail. The purpose of this book is to teach the material necessary to pass such an examination. and the like. or take a nonstandard approach to some of the basic theorems. the better (within reason).D. These are the proofs and methods that most experts grew up on. I had three main goals in writing this text: (1) present a very clear exposition. Students who are starting their graduate mathematics education are often still developing their mathematical sophistication and ﬁnd that the more details that are provided. For the sake of clarity. too advanced. ix .
I give hints to the harder questions. (Chapter 1 is a summary of the notation that is used and the background material that is required. Doing this for subsets of the real line gives rise to Lebesgue measure. . For an interval contained in the real line or a nice region in the plane. When I teach the real analysis course. Why do publishers make textbooks so expensive? This is particularly troublesome when one considers that nowadays authors do their own typesetting and frequently their own page layout. of which one example is Lebesgue measure on the line. A substantial number are taken from preliminary examinations given in the past at universities with which I have been associated. not a factor of 10. prove the monotone and dominated convergence theorems. deﬁne the Lebesgue integral.pdf format for free. the deﬁnition of measures. Chapters 2–4 discuss classes of sets. These very same books now sell for $100$200. We talk about measurable functions. the length of the interval or the area of the region give an idea of the size. and the construction of measures. look at some simple properties of the Lebesgue integral. (3) I have on my bookshelf several books that I bought in the early 1970’s that have the prices stamped in them: $10$12. We want to extend the notion of size to as large a class of sets as possible. At this point I should tell you a little bit about the subject matter of real analysis. I leave it to the individual instructor to decide how many hints to give.) Once we have measures. I tried to make them interesting and useful and to avoid problems that are overly technical. we proceed to the Lebesgue integral. But some instructors will want a more challenging course than I give and some a less challenging one. My aim was to make the soft cover version of this text cost less than $20 and to provide a version in . This material is the subject of Chapters 5–10. I thought long and hard as to whether to provide hints to the exercises. but only by a factor of 5 or 6. but some are quite challenging. The cost of living has gone up in the last 40 years. I am selfpublishing the text. Many are routine. compare it to the Riemann integral. and discuss some of the various ways a sequence of functions can converge.x PREFACE (2) There are over 220 exercises. many are of moderate diﬃculty. To do that.
and Hilbert spaces. and [4] helpful. Bharath. e Lu. M. L. suggestions. An acquaintance with metric spaces is assumed. If that is the case at your university. Lawler. A summary of what you need to know is in Chapter 1. E. M. H. I would like to thank A. K. Gin´. Ren. The prerequisites to this text are a solid background in undergraduate mathematics. signed measures. the diﬀerentiation of functions on the line. Gordina. however.xi Closely tied with measures and integration are the subjects of product measures. At some universities preliminary or qualifying examinations in real analysis are combined with those in undergraduate analysis or topology or complex analysis. but no other topology. Burdzy. the RadonNikodym theorem. We present these in Chapters 16–19. Further reading is always useful. one good place is the book [3]. K. corrections. Baldenko. I. provide hints or solutions to the exercises. E. Rogers.. J. Ferrone. Teplyaev for some very useful suggestions. I would be glad to hear from you: r. These are covered in Chapters 11–15. If you have comments. All the necessary background material can be learned from many sources. Pitman. Good luck with your exam! . BenAri. and Lp spaces. the Riesz representation theorem. and A. Poehlitz. L. Others also include some or all of the following topics: the Fourier transform. Banach spaces. K. etc. I cannot. Hsu. [2]. I have found the books [1]. Many courses in real analysis stop at this point.edu. you will have to supplement this text with texts in those subjects. Marinelli. G.bass@uconn. D.
xii PREFACE .
a collection of subsets {Aα }α∈I is disjoint if Aα ∩ Aβ = ∅ whenever α = β. we assume all our sets are subsets of some given set X. 1. If I is some index set.” for the set of points not in A. deﬁne Ac = {x ∈ X : x ∈ A}. read “A complement. We write Ai ↑ if A1 ⊂ A2 ⊂ · · · and write Ai ↑ A if in addition A = ∪∞ Ai . The set A∆B is called the symmetric diﬀerence of A and B and is the set of points that are in one of the sets but not the other. and to be precise. / We write A − B = A ∩ Bc (it is common to also see A \ B) and A∆B = (A − B) ∪ (B − A). Similarly Ai ↓ means A1 ⊃ A2 ⊃ · · · and Ai ↓ A i=1 means that in addition A = ∩∞ Ai . To avoid some of the paradoxes of set theory.Chapter 1 Preliminaries In this short chapter we summarize some of the notation and terminology we will use and recall a few deﬁnitions and results from undergraduate mathematics. i=1 1 .1 Notation and terminology We use Ac .
The proofs and more details can be found in many places. If f is a function whose domain is the reals or a subset of the reals. A good source is [3].) We deﬁne decreasing and strictly decreasing similarly. where x+ = x ∨ 0 and x− = (−x) ∨ 0. . resp. n even. We say a function f : R → R is increasing if x < y implies f (x) ≤ f (y) and f is strictly increasing if x < y implies f (x) < f (y). 1. then lim supn→∞ an = 1 and lim inf n→∞ an = 0. n m≥n For example. For example. algebra. n odd. y) and x ∧ y = min(x. y). The sequence {an } has a limit if and only if lim supn→∞ an = lim inf n→∞ an and both are ﬁnite. We use analogous deﬁnitions when we take a limit along the real numbers. and C the set of complex numbers. and analysis. then limy→x+ f (y) and limy→x− f (y) are the right and left hand limits of f at x. then z is the complex conjugate of z. We can write a real number x in terms of its positive and negative parts: x = x+ − x− . (Some authors use “nondecreasing” for the former and “increasing” for the latter. A function is monotone if f is either increasing or decreasing. If z is a complex number.2 Some undergraduate mathematics We recall some deﬁnitions and facts from undergraduate topology. lim sup an = inf sup am . lim sup f (y) = inf y→x δ>0 y−x<δ sup f (y). if an = 1. −1/n. PRELIMINARIES We use Q to denote the set of rational numbers. n→∞ n→∞ n m≥n lim inf an = sup inf am . R the set of real numbers. Given a sequence {an } of real numbers. We use x ∨ y = max(x.2 CHAPTER 1.
n ≥ N . f is continuous if it is continuous at every point of its domain. x) < x whenever n ≥ N . F ⊂ K. z ∈ X. is the set of x such that there exists rx > 0 with B(x. rx ) ⊂ A. r) = {y ∈ X : d(x. A set A is open if A = Ao . Proposition 1. The index set I can be ﬁnite or inﬁnite. is the set of x ∈ X such that every open ball centered at x contains at least one point of A. i=1 We have the following two facts about compact sets. . denoted Ao . (3) d(x. let B(x. Condition (3) is called the triangle inequality. x) for all x. Proposition 1. and F is closed. One property of continuous functions is that f −1 (F ) is closed and f −1 (G) is open if f is continuous. If A ⊂ X. there exists G1 .1. i. there exists δ > 0 such that f (x) − f (y) < ε whenever d(x. F is closed.. denoted A. Given a metric space X. . z) ≤ d(x.2. A sequence {xn } ⊂ X converges to a point x ∈ X if for each ε > 0 there exists N such that d(xn . y. f takes on its maximum and minimum values.2 If K is compact and f is continuous on K. Gn ∈ {Gα }α∈I such that K ⊂ ∪n Gi . . closed if A = A. y ∈ X. the support of f is the closure of the set {x : f (x) = 0}. A sequence is a Cauchy sequence if for each ε > 0 there exists N such that d(xm . If every Cauchy sequence in X converges to a point in X. y) < r} be the open ball of radius r centered at x. The closure of A.1 If K is compact. If f : X → R. then there exist x1 and x2 such that f (x1 ) = inf x∈K f (x) and f (x2 ) = supx∈K f (x). . A set K is compact of every open cover contains a ﬁnite subcover. f is continuous at a point x if given ε > 0. An open cover of a subset K of X is a collection {Gα }α∈I of open sets such that K ⊂ ∪α∈I Gα . and G is open. (2) d(x. y) < δ. z) for all x. the interior of A. y) = 0 if and only if x = y. . y) ≥ 0 for all x.e. In other words. we say X is complete. y) = d(y. y ∈ X and d(x. y) + d(y. then F is compact. xn ) < ε whenever m. SOME UNDERGRADUATE MATHEMATICS 3 A set X is a metric space if there exists a metric d : X × X → R such that (1) d(x.
X can be written as the union of disjoint equivalence classes. We use the usual notation. resp. (2) if x ≤ y and y ≤ x. y ∈ X. (5) c(x + y) = cx + cy for all x. y. r/2) are disjoint open sets containing x and y. Therefore metric spaces are also what are called Hausdorﬀ spaces. (2) cx = c x for all c ∈ F and x ∈ X. (3) if x ∼ y and y ∼ z. r/2) and d(y. x − y = x + (−y). c ∈ F . y ∈ X. then y ∼ x. and say A ≤ B if A. Let F be either R or C. c. y) and note that d(x. d ∈ F . (2) if x ∼ y. (3) there exists an element 0 ∈ X such that 0 + x = x for all x ∈ X. (6) (c + d)x = cx + dx for all x ∈ X. let Y be a set. X is a normed linear space if there exists a map x → x such that (1) x ≥ 0 for all x ∈ X and x = 0 if and only if x = 0. B ∈ X and A ⊂ B. x and y are in the same equivalence class if and only if x ∼ y. Given an equivalence relationship. (2) (x + y) + z = x + (y + z) for all x. c. For an example. y ∈ X. d ∈ F . it is not necessarily true that x ≤ y or y ≤ x. then x ∼ z. z ∈ X.g. such that (1) x + y = y + x for all x.4 CHAPTER 1. e. (3) x + y ≤ x + y for all x. PRELIMINARIES Remark 1.. (7) c(dx) = (cd)x for all x ∈ X. let X = R and say x ∼ y if x − y is a rational number. Given a normed linear space X. y ∈ X. . A set X has a partial order “≤” if (1) x ≤ x for all x ∈ X. let r = d(x. (8) 1x = x for all x ∈ X. Note that given x. let X be the collection of all subsets of Y . addition (+) and scalar multiplication. y) = x − y . (4) for each x in X there exists an element −x ∈ X such that x + (−x) = 0. we can make X into a metric space by setting d(x. then x = y. For an example.3 If x = y. A set X has an equivalence relationship “∼” if (1) x ∼ x for all x ∈ X. X is a vector space or linear space if there exist two operations.
7 Let X be a compact metric space and let A be a collection of continuous complexvalued functions on X with the following properties: (1) If f. and K is contained in a ﬁnite interval.2.7 see [1] or [3]. Proposition 1. y ∈ X with x = y. Then both limy→x+ f (y) and limy→x− f (y) exist for every x. (4) If x. then f + g.5 Suppose G ⊂ R is open. (2) If f ∈ A. For an application of Hilbert space techniques to Fourier series. In (3) and (4). we will use the StoneWeierstrass theorem. which is the last section of Chapter 19.1. A is said to vanish at no point of X. there exists f ∈ A such that f (x) = 0. there exists f ∈ A such that f (x) = f (y). For a proof of Theorem 1. x∈X When (3) holds. Then the closure of A with respect to the supremum norm is the collection of continuous complexvalued functions on X. . (3) If x ∈ X. f g. Then K is compact. g ∈ A and c ∈ C. Theorem 1.6 Suppose f : R → R is an increasing function. The conclusion can be rephrased as saying that given f continuous on X and ε > 0. SOME UNDERGRADUATE MATHEMATICS We need the following three facts about the real line. the function f depends on x and on x and y. When (4) holds. and cf are in A. Proposition 1. then f ∈ A. The particular version we will use is the following. A is said to separate points. 5 Proposition 1. resp. where f is the complex conjugate of f . Moreover the set of x where f is not continuous is countable. K is closed. Then G can be written as the countable union of disjoint open intervals. there exists g ∈ A such that sup f (x) − g(x) < ε.4 Suppose K ⊂ R.
PRELIMINARIES .6 CHAPTER 1.
1 An algebra is a collection A of subsets of X such that (1) ∅ ∈ A and X ∈ A.Chapter 2 Families of sets 2. Since ∩∞ Ai = i=1 (∪∞ Ac )c . then ∪n Ai and ∩n Ai are in A. 7 . i=1 i=1 A is a σalgebra if in addition (4) whenever A1 . A) is called a measurable space. Deﬁnition 2. . the requirement that ∩∞ Ai be in A is redundant. (2) if A ∈ A. then ∪∞ Ai and ∩∞ Ai are in i=1 i=1 A. Let X be a set.1 Algebras and σalgebras When we turn to constructing measures in Chapter 4. In (4) we allow countable unions and intersections only. i=1 i i=1 The pair (X. (3) if A1 . . we will see that we cannot in general deﬁne the measure of an arbitrary set. . are in A. . We will have to restrict the class of sets we consider. . A2 . then Ac ∈ A. An ∈ A. . we do not allow uncountable unions and intersections. The class of sets that we will want to use are σalgebras (read “sigma algebras”). . A set A is measurable if A ∈ A.
. and so is in A. . C is a σalgebra. . are each in A.3 Let X = R and let A = {A ⊂ R : A is countable or Ac is countable}. This follows immediately from the deﬁnition. Lemma 2. ( 1 . 1] and let A = {∅. If Ac0 is countable for some i0 .5 Let X = {1. Verifying parts (1) and (2) of the deﬁnition is easy. 3} and let A = {X.8 CHAPTER 2. then ∩α∈I Aα is a σalgebra. Then A is a σalgebra. If we have a collection C of subsets of X. then ∪i Ai is countable. and let A be the collection of all subsets of R. . and let B1 . 1]. 2 2 Then A is a σalgebra. or say that . {2. Since ∩Ai = (∪i Ac )c . (Thus A consists of 28 elements. X. C ⊂ A}.7. Example 2. Proof. [0. A2 .6 Let X = [0.2 Let X = R. 1 ]. then i the countable intersection of sets in A is again in A. {1}. 1]}. Then A is a σalgebra. Example 2. . Example 2. the set of real numbers. Example 2. Since there is at least one σalgebra containing C. . 2. ∅. B8 be subsets of X which are pairwise disjoint and whose union is all of X. and again ∪i Ai is in A. we are never taking the intersection over an empty class of σalgebras. the one consisting of all subsets of X. namely.4 Let X = [0. the intersection of all σalgebras containing C. i then (∪Ai )c = ∩i Ac ⊂ Ac0 i i is countable.) Then A is a σalgebra. We call σ(C) the σalgebra generated by the collection C. If each of the Ai are countable. Let A be the collection of all ﬁnite unions of the Bi ’s as well as the empty set. deﬁne σ(C) = ∩{Aα : Aα is a σalgebra. .7 If Aα is a σalgebra for each α in some index set I. 3}}. Suppose A1 . FAMILIES OF SETS Example 2. In view of Lemma 2.
1 (3) Using (a. then [a. and this is often denoted B.5. Elements of B are called Borel sets and are said to be Borel measurable. b] : a. then the Borel σalgebra B is generated by each of the following collection of sets: (1) C1 = {(a. b] : a. provided n0 is taken large enough. . if G is open. then C1 ⊂ G. and consequently σ(C1 ) ⊂ σ(G) = B. To get the reverse inclusion. We will see later that when X is the real line. (3) C3 = {(a. a + n). b) = 1 ∪∞ 0 (a. it is a metric space.8 If X = R. we see that C3 ⊂ σ(C1 ). Using (a. b ∈ R}. b) : a. choose n0 ≥ 2/(b − a) and note (a. ∞) = ∪∞ (a. (4) C4 = {(a. it is the countable union of open intervals by Proposition 1.1. then we can talk about open sets. b ∈ R}. and hence σ(C2 ) ⊂ σ(σ(G)) = σ(G) = B. then σ(σ(C)) = σ(C). ALGEBRAS AND σALGEBRAS 9 C generates the σalgebra σ(C). B is not equal to the collection of all subsets of X. Hence if G is open. b) ∈ C1 . (2) C2 = {[a. then G ∈ σ(C1 ). If X has some additional structure. σ(G) is the Borel σalgebra. b + n ) ∈ σ(G). b] ∈ C2 . Proposition 2. ∞) ∈ σ(C1 ) n=1 if a ∈ R and similarly (−∞. We end this section with the following proposition. Since (a.b − ∈ σ(C2 ). say. n=1 Therefore C2 ⊂ σ(G). By deﬁnition. n=n and as above we argue that B = σ(C1 ) ⊂ σ(C3 ). n n Therefore C1 ⊂ σ(C2 ). C1 ⊂ σ(C3 ). b ∈ R}. and then B = σ(G) ⊂ σ(σ(C1 )) = σ(C1 ). If G is the collection of open subsets of X. b − n ]. Proof. 1 1 (2) If [a. then σ(C1 ) ⊂ σ(C2 ). a) ∈ σ(C1 ) if a ∈ R. b] = ∩∞ (a. then (a. ∞) : a ∈ R}. Since σ(C) is a σalgebra. It is clear that if C1 ⊂ C2 .2. b + n ). Every ﬁnite open interval is in C1 . from which it follows that B = σ(C1 ) ⊂ σ(σ(C2 )) = σ(C2 ). b] = ∩∞ (a − n . This says G ⊂ σ(C1 ). If (a. Since every element of C1 is open. (1) Let G be the collection of open sets. then we call σ(G) the Borel σalgebra on X. b) = ∪∞ 0 a + n=n 1 1 . and n=1 as above we conclude that σ(C3 ) ⊂ σ(C1 ) = B.
A∩B ∈ M. We must show A ⊂ M. FAMILIES OF SETS (4) Because (a. then C4 ⊂ σ(C3 ). Note N1 is contained in M and contains A0 .2 The monotone class theorem This section will be used in Chapter 11. (2) if Ai ↓ A and each Ai ∈ M. and the intersection of all monotone classes containing a given collection of sets is the smallest monotone class containing that collection. then A∩B = ∪∞ (Ai ∩B). A is the smallest σalgebra containing A0 . a + n]. so M ⊂ A. then C1 ⊂ σ(C4 ). and M is the smallest monotone class containing A0 . then each Ac ∈ M and i Ac ↓ Ac . this is enough n=1 to imply that σ(C4 ) = B. . and so A ∈ N1 . then A ∈ N1 . Deﬁnition 2. Let N1 = {A ∈ M : Ac ∈ M}. Then M = A. i Similarly. i=1 which implies A ∈ N2 . then A ∈ M. if Ai ↓ A and each Ai ∈ N1 . and B ∈ A0 . Ac ∈ M.10 CHAPTER 2. Since M is a monotone class. As above. Proof. Let N2 = {A ∈ M : A ∩ B ∈ M for all B ∈ A0 }. then A ∈ M. If Ai ↑ A and each Ai ∈ N1 .10 Suppose A0 is an algebra. Theorem 2. Since (a. Hence N1 = M. Therefore N1 is a monotone class. each Ai ∈ N2 . but very useful. We use a similar argument when Ai ↓ A. ∞) − (b. is rather technical. A σalgebra is clearly a monotone class. ∞). Because M is a monotone class. The next theorem. Note the following: N2 is contained in M and N2 contains A0 because A0 is an algebra. The intersection of monotone classes is a monotone class. the monotone class theorem. b] = (a. and we conclude M is closed under the operation of taking complements. If Ai ↑ A.9 A monotone class is a collection of subsets M of X such that (1) if Ai ↑ A and each Ai ∈ M. ∞) = ∪∞ (a. 2.
Deﬁne B = {f −1 (A) : A ∈ A}. Exercise 2. C ∈ A with B ∩ C = ∅. Exercise 2.2 Find an example of a set X and two σalgebras A1 and A2 .5 Let (Y.3 Suppose A1 ⊂ A2 ⊂ · · · are σalgebras consisting of subsets of a set X. N3 is a monotone class contained in M. but M is not a σalgebra. and neither B nor C is empty. if B ∈ A0 and A ∈ M. Prove that A is uncountable. Let N3 = {A ∈ M : A ∩ B ∈ M for all B ∈ M}. and cf . and we conclude N2 = M. EXERCISES 11 Therefore N2 is a monotone class. Exercise 2. By the last sentence of the preceding paragraph. Exercise 2. give a counterexample. N3 contains A0 . We thus have that M is a monotone class closed under the operations of taking complements and taking intersections. 2. and so A ⊂ M. i=1 give a counterexample. such that A1 ∪ A2 is not a σalgebra.7 Suppose F is a collection of realvalued functions on X such that the constant functions are in F and f + g. then A ∩ B ∈ M. Hence N3 = M. Suppose Aj ↑ A and each Aj ∈ M. A) be a measurable space and let f map X onto Y .6 Suppose A is a σalgebra with the property that whenever A ∈ A.2. there exist B. but do not assume that f is onetoone.3. In other words. Is ∪∞ Ai necessarily a σalgebra? If not. X ∈ M. Exercise 2.1 Find an example of a set X and a monotone class M consisting of subsets of X such that ∅ ∈ M. Exercise 2. Let M = ∪∞ Mn . Is A n=1 necessarily in M? If not. Prove that B is a σalgebra of subsets of X. f g. As in the preceding paragraph.4 Suppose M1 ⊂ M2 ⊂ · · · are monotone classes.3 Exercises Exercise 2. This shows M is a σalgebra. each consisting of subsets of X. B ∪ C = A.
but not ﬁnitely many? . Deﬁne the function χA (x) = 1. / Prove that A = {A ⊂ X : χA ∈ F} is a σalgebra. Exercise 2. g ∈ F and c ∈ R. 0. x ∈ A. FAMILIES OF SETS are in F whenever f. Suppose f ∈ F whenever fn → f and each fn ∈ F.12 CHAPTER 2. x ∈ A.8 Does there exist a σalgebra which has countably many elements.
Chapter 3 Measures In this chapter we give the deﬁnition of a measure. until Chapter 4 3. 13 . Lebesgue measure. . A. . (2) if Ai ∈ A. are pairwise disjoint. . Saying the Ai are pairwise disjoint means that Ai ∩ Aj = ∅ if i = j. .1 Deﬁnitions and examples Deﬁnition 3. The triple (X. A) is a function µ : A → [0.1 Let X be a set and A a σalgebra consisting of subsets of X. Constructing measures is often quite diﬃcult and we defer the construction of the most important one. An are in A and are pairwise disjoint. 2. µ) is called a measure space.. then ∞ µ(∪∞ Ai ) i=1 = i=1 µ(Ai ). i = 1. some examples. . and some of the simplest properties of measures. Deﬁnition 3. A measure on (X. ∞] such that (1) µ(∅) = 0. . We say a set n function is ﬁnitely additive if µ(∪n Ai ) = i=1 µ(Ai ) whenever i=1 A1 .1(2) is known as countable additivity. .
3 Let X = R. B ∈ A with A ⊂ B. Set µ(A) = {i:xi ∈A} ai . a2 . B3 = A3 − (A1 ∪ A2 ). and a1 . . The Bi j=1 are pairwise disjoint. . . This measure is called point mass at x. then µ(A) ≤ i=1 µ(Ai ). ∞ (2) If Ai ∈ A and A = ∪∞ Ai . then µ(A) ≤ µ(B). (4) Suppose Ai ∈ A and Ai ↓ A. then we have µ(A) = limn→∞ µ(An ). and in general Bi = Ai − (∪i−1 Aj ). MEASURES Example 3. A the collection of all subsets of X. Example 3.2 Let X be any set. i=1 i=1 i=1 n→∞ n→∞ . Since ∪n Bi = ∪n Ai . B2 = A2 − A1 . Example 3. ≥ 0. x2 . (3) Deﬁne the Bi as in (2). µ is called counting measure. i=1 (3) Suppose Ai ∈ A and Ai ↑ A. (1) Let A1 = A. . x1 . Proof. (2) Let B1 = A1 .14 CHAPTER 3. . i=1 i=1 Hence ∞ ∞ µ(A) = µ(∪∞ Bi ) = i=1 i=1 µ(Bi ) ≤ i=1 µ(Ai ).4 Let δx (A) = 1 if x ∈ A and 0 otherwise. A the collection of all subsets of R. B4 = A4 − (A1 ∪ A2 ∪ A3 ). . and µ(A) the number of elements in A. Then µ(A) = limn→∞ µ(An ). then i=1 i=1 ∞ µ(A) = µ(∪∞ Ai ) i=1 n = µ(∪∞ Bi ) i=1 = i=1 µ(Bi ) = lim n→∞ µ(Bi ) = lim µ(∪n Bi ) = lim µ(∪n Ai ). If µ(A1 ) < ∞. and A3 = A4 = · · · = ∅. ∈ R. Counting measure is a particular case of this is if xi = i and all the ai = 1. A2 = B − A. Now use part (2) of the deﬁnition of measure to write µ(B) = µ(A) + µ(B − A) + 0 + 0 + · · · ≥ µ(A). Proposition 3. and ∪∞ Bi = ∪∞ Ai .5 The following hold: (1) If A. Bi ⊂ Ai for each i.
let X be the positive integers. A subset A ⊂ X is a null set if there exists a set B ∈ A with A ⊂ B and µ(B) = 0. Then the Ai decrease. .5. then (X. A.6 To see that µ(A1 ) < ∞ is necessary in Proposition 3. A.}.1. . µ) is complete. and F is called a σﬁeld. . If µ is a ﬁnite measure. but µ(∩i Ai ) = µ(∅) = 0. A. 2.. i + 1. n→∞ Now subtract µ(A1 ) from both sides and then multiply both sides by −1. . If we let Fn = ∪n Ei . such that µ(Ei ) < ∞ for each i and X = ∪∞ Ei . The sets A1 − Ai increase to A1 − A. µ) is said to be a complete measure space. . if µ is a σﬁnite measure. µ counting measure. P) instead of (X. DEFINITIONS AND EXAMPLES 15 (4) Apply (3) to the sets A1 − Ai . i=1 then (X. and so µ(A1 ) − µ(A) = µ(A1 − A) = lim µ(A1 − An ) n→∞ = lim [µ(A1 ) − µ(An )]. A. Sometimes one just says that A is complete or that µ is complete when (X.3. F.7 are increasing. then µ(Fn ) < ∞ for each n and Fn ↑ X. A. . The completion of A is the smallest σalgebra A containing A such that (X. Deﬁnition 3. and similarly. µ) be a measure space. If A contains all the null sets. µ) is complete. A. i=1 Therefore there is no loss of generality in supposing the sets Ei in Deﬁnition 3. and Ai = {i. i = 1. Let (X. 2.7 A measure µ is a ﬁnite measure if µ(X) < ∞. In this case we usually write (Ω. µ) is called a σﬁnite measure space. We do not require A to be in A. A. then (X. Example 3. µ). which is the same thing as a σalgebra. A probability or probability measure is a measure µ such that µ(X) = 1. . . µ(Ai ) = ∞ for all i. . µ) is called a ﬁnite measure space. . A measure µ is σﬁnite if there exist sets Ei ∈ A for i = 1.
Prove that if µ1 . . . Exercise 3. then ν is a measure. Exercise 3.1 Suppose (X. ∈ [0. Exercise 3. µ) is a measure space and A.5 3. then n=1 an µn is also a measure. b)) whenever a < b. Prove that m(A) = n(A) whenever A ∈ B. µ). .16 CHAPTER 3. . . A. B. and we deﬁne ν(A) = µ(A ∩ B) for A ∈ A. A) is a measurable space and µ is a nonnegative set function that is ﬁnitely additive and such that µ(∅) = 0. . A. A. µ) is complete and is the completion of (X. B) such that m((a. What if µn (A) ↓ for each A ∈ A and µ1 (X) < ∞? Exercise 3. A) and µn (A) ↑ for each A ∈ A.2 Let X be an uncountable set and let A be the collection of subsets A of X such that either A or Ac is countable. µ) be a measure space. . Exercise 3. Prove that µ is a measure. Prove that if (X. give a counterexample. and m and n are two measures on (X. MEASURES 3. Deﬁne µ(A) = lim µn (A). . Show that µ is a measure. are measures on a mea∞ surable space and a1 . are measures on a measurable space (X.8 Suppose X is the set of real numbers. Exercise 3. a2 . then µ(∪i Ai ) = limi→∞ µ(Ai ). µ2 . b)) = n((a. Prove that µ(A) + µ(B) = µ(A ∪ B) + µ(A ∩ B). A.3 Suppose (X.4 3. µ) is a measure space.6 Suppose µ1 . B is the Borel σalgebra. B ∈ A. ∞). n→∞ Is µ necessarily a measure? If not. Suppose that whenever Ai is an increasing sequence of sets in A.2 Exercises Exercise 3. Deﬁne µ(A) = 0 if A is countable and µ(A) = 1 if A is uncountable. let N be the collection of null sets with respect to A.7 Let (X. µ2 . B ∈ A. Exercise 3. and let B = σ(A ∪ N ). Show that (X. . .
EXERCISES 17 Exercise 3.3. A) is a measurable space and C is an arbitrary subset of A. Suppose m and n are two σﬁnite measures on (X. Is it true that m(A) = n(A) for all A ∈ σ(C)? What if m and n are ﬁnite measures? . A) such that m(A) = n(A) for all A ∈ C.9 Suppose (X.2.
MEASURES .18 CHAPTER 3.
3. bi ) are i=1 pairwise disjoint. This is a complicated procedure. and involves the concept of outer measure. Our most important example will be onedimensional Lebesgue measure. if G = ∪∞ (ai . This is shown in Section 4. Further results and some examples related to Lebesgue measure are given in Section 4. 19 . bi ).2. The Carath´odory extension theorem is a tool developed in e Section 4. The methods used to construct measures via outer measures have other applications besides the construction of Lebesgue measure.1. where the intervals (ai . which we consider in Section 4. Since every open subset of the reals is the countable union of disjoint open intervals (see Proposition 1.Chapter 4 Construction of measures Our goal in this chapter is give a method for constructing measures. We want the measure m of an open interval to be the length of the interval. which we introduce in Section 4. we must have ∞ m(G) = i=1 (bi − ai ).4. Let us present some of the ideas used in the construction of Lebesgue measure on the line.5 that can be used in constructing measures.5). One cannot deﬁne the Lebesgue measure of every subset of the reals.
A set N is a null set with respect to µ∗ if µ∗ (N ) = 0. Proof. A common way to generate outer measures is as follows. ∞] with (∅) = 0. The diﬃculty is that m is not a measure on the σalgebra consisting of all subsets of the reals. Deﬁnition 4. . A ⊂ G} for arbitrary subsets E ⊂ R. Ci2 . .1 Let X be a set. For each i there exist Ci1 . Proposition 4. are subsets of i=1 X. This is the essential idea behind the construction of Lebesgue measure.1) Then µ∗ is an outer measure. be subsets of X and let ε > 0. . . 4. (1) and (2) of the deﬁnition of outer measure are obvious. We resolve this by considering a strictly smaller σalgebra. .1 Outer measures We begin with the notion of outer measure. .20 We then set CHAPTER 4. let A1 . CONSTRUCTION OF MEASURES m(E) = inf{m(G) : G open. but it is technically simpler to work with intervals of the form (a.4. ∞ (3) µ∗ (∪∞ Ai ) ≤ i=1 µ∗ (Ai ) whenever A1 . (2) if A ⊂ B. Suppose : C → [0. i=1 (4. A2 . then µ∗ (A) ≤ µ∗ (B). . An outer measure is a function µ∗ deﬁned on the collection of all subsets of X satisfying (1) µ∗ (∅) = 0. . this is proved in Section 4. To prove (3). . ∈ C such that Ai ⊂ ∪∞ Cij and j=1 . Deﬁne ∞ µ∗ (E) = inf i=1 (Ai ) : Ai ∈ C for each i and E ⊂ ∪∞ Ai . b] rather than open intervals. A2 .2 Suppose C is a collection of subsets of X such that ∅ and X are both in C.
OUTER MEASURES j 21 (Cij ) ≤ µ∗ (Ai ) + ε/2i .4. that if we restrict µ∗ to a σalgebra L which is strictly smaller than the collection of all subsets of R. but not always. and the corresponding σalgebra is the collection of all subsets of R. Let (I) = α(b) − α(a) if I = (a. b]. then µ∗ will be a measure on L. .4).1. intervals that are open on the left and closed on the right.j ∞ (Cij ) = i ∞ j (Cij ) ε/2i i=1 ≤ i=1 ∞ µ∗ (Ai ) + µ∗ (Ai ) + ε. Again Proposition 4. In general we need to restrict µ∗ to a strictly smaller σalgebra than the collection of all subsets of R. Again deﬁne µ∗ by (4. however. Example 4. Example 4. Let (I) = b − a if I = (a. The σalgebra L is called the Lebesgue σalgebra. then the corresponding LebesgueStieltjes measure is point mass at 0 (deﬁned in Example 3. b] as in the previous example.4 that µ∗ is not a measure on the collection of all subsets of R. Then ∪∞ Ai ⊂ ∪i ∪j Cij and i=1 µ∗ (∪∞ Ai ) ≤ i=1 i.3 Let X = R and let C be the collection of intervals of the form (a. The special case where α(x) = x for all x is Lebesgue measure. but we will see in Section 4. i=1 ∞ i=1 = Since ε is arbitrary. Thus α(x) = limy→x+ α(y) for each x and α(x) ≤ α(y) if x < y. if α(x) = 0 for x < 0 and 1 for x ≥ 1.2 shows that µ∗ is an outer measure. Restricting µ∗ to a smaller σalgebra gives us what is known as LebesgueStieltjes measure corresponding to α. For example. b].1). We will also see. Deﬁne µ∗ by (4.2 shows that µ∗ is an outer measure. Proposition 4. That measure is what is known as Lebesgue measure. µ∗ (∪∞ Ai ) ≤ i=1 µ∗ (Ai ). b].4 Let X = R and let C be the collection of intervals of the form (a. Let α : R → R be an increasing right continuous function on R. that is.1).
Then µ∗ (E) = µ∗ (E ∩ A) + µ∗ (E ∩ Ac ) = [µ∗ (E ∩ A ∩ B) + µ∗ (E ∩ A ∩ B c )] + [µ∗ (E ∩ Ac ∩ B) + µ∗ (E ∩ Ac ∩ B c )].6 If µ∗ is an outer measure on X. then µ∗ (E) ≥ µ∗ (E ∩ (A ∪ B)) + µ∗ (E ∩ (A ∪ B)c ). then the collection A of µ∗ measurable sets is a σalgebra. Moreover. Step 1. A contains all the null sets. then Ac ∈ A by symmetry and the deﬁnition of A. B ∈ A and E ⊂ X. The second equality follows from the deﬁnition of A with E ﬁrst replaced by E ∩A and then by E ∩Ac . This is sometimes known as Carath´odory’s theorem.2) . then µ is a measure. Therefore A is an algebra.5. If A ∈ A. (4. Thus to check (4.2) it is enough to show µ∗ (E) ≥ µ∗ (E ∩ A) + µ∗ (E ∩ Ac ). The ﬁrst three summands on the right of the second equals sign have a sum greater than or equal to µ∗ (E ∩ (A ∪ B)) because A ∪ B ⊂ (A ∩ B) ∪ (A ∩ B c ) ∪ (Ac ∩ B). If µ is the restriction of µ∗ to A. Theorem 4. Suppose A. By Deﬁnition 4. which shows A ∪ B ∈ A. µ∗ (E) ≤ µ∗ (E ∩ A) + µ∗ (E ∩ Ac ) for all E ⊂ X. but do not e confuse this with the Carath´odory extension theorem in Section e 4. This will be trivial in the case µ∗ (E) = ∞.5 Let µ∗ be an outer measure.1. CONSTRUCTION OF MEASURES Deﬁnition 4.22 CHAPTER 4. First we show A is an algebra. Since Ac ∩ B c = (A ∪ B)c . A set A ⊂ X is µ∗ measurable if µ∗ (E) = µ∗ (E ∩ A) + µ∗ (E ∩ Ac ) for all E ⊂ X. Proof.
Next we show A is a σalgebra. Since Bn ∈ A. let Bn = ∪n Ai .1.3) can hold is if all the inequalities there are actually equalities. then n µ (E) = µ (E ∩ Bn ) + µ (E ∗ ∗ ∗ c ∩ Bn ) ≥ i=1 µ∗ (E ∩ Ai ) + µ∗ (E ∩ B c ).3) ≥ µ (∪∞ (E ∩ Ai )) + µ∗ (E ∩ B c ) i=1 = µ∗ (E ∩ B) + µ∗ (E ∩ B c ) ≥ µ∗ (E). . OUTER MEASURES 23 Step 2. . We now show µ∗ restricted to A is a measure. i=1 i=1 c Also. ∞ µ∗ (E) ≥ i=1 ∗ µ∗ (E ∩ Ai ) + µ∗ (E ∩ B c ) (4. Let n → ∞. Similarly. and B = ∪∞ Ai . we obtain n µ∗ (E ∩ Bn ) ≥ i=1 µ∗ (E ∩ Ai ). and in general Ai = Ci − (∪i−1 Aj ). are sets in A. i=1 i=1 µ∗ (E ∩ Bn ) = µ∗ (E ∩ Bn ∩ An ) + µ∗ (E ∩ Bn ∩ Ac ) n = µ∗ (E ∩ An ) + µ∗ (E ∩ Bn−1 ). A3 = C3 − (A1 ∪ A2 ). Recalling that µ∗ is an outer measure. The Ai are pairwise disjoint. µ∗ (E ∩ Bn−1 ) = µ∗ (E ∩ An−1 ) + µ∗ (E ∩ Bn−2 ). and continuing. ∩∞ Ci = (∪∞ Ci )c ∈ A. and in particular. let A1 = C1 . Let Ai be pairwise disjoint sets in A. and therefore A is a σalgebra. . then Ai = Ci ∩ Ci−1 ∈ A. C2 .4. A2 = C2 − A1 . Now if C1 . This shows B ∈ A. . i=1 i=1 Step 3. Since j=1 c each Ci ∈ A and A is an algebra. If E ⊂ X. The only way (4. ∪∞ Ci = ∪∞ Ai ∈ A. so from the previous paragraph. ∞ µ∗ (E) = i=1 µ∗ (E ∩ Ai ) + µ∗ (E ∩ B c ).
then K ∪ L = (a.4) . We then use Theorem 4. then µ∗ (E ∩ A) + µ∗ (E ∩ Ac ) = µ∗ (E ∩ Ac ) ≤ µ∗ (E). c]. Let α(x) be an increasing right continuous function. Finally. if K = (a. that is.24 CHAPTER 4. if µ∗ (A) = 0 and E ⊂ X. which shows A contains all the null sets. Note that if K and L are adjacent intervals. Deﬁne ∞ m∗ (E) = inf i=1 (Ai ) : Ai ∈ C for each i and E ⊂ ∪∞ Ai .2 to tell us that m∗ is an outer measure. CONSTRUCTION OF MEASURES Taking E = B. b] and L = (b. we obtain ∞ µ∗ (B) = i=1 µ∗ (Ai ). ∗ 4. intervals that are open on the left and closed on the right.) We use Proposition 4. Deﬁne ((a. The next step in the construction of LebesgueStieltjes measure corresponding to α is the following.6 to show that m∗ is a measure on the collection of m∗ measurable sets. (4. c] and (K) + (L) = [α(b) − α(a)] + [α(c) − α(b)] = α(c) − α(a) = (K ∪ L) by the deﬁnition of . This means that α(x) ≤ α(y) if x < y and limz→x+ α(z) = α(x) for all x. or µ is countably additive on A. Step 4. b]) = α(b) − α(a). i=1 (In this book we usually use m instead of µ when we are talking about LebesgueStieltjes measures. that is. We do not require α to be strictly increasing. b].2 LebesgueStieltjes measures Let X = R and let C be the collection of intervals of the form (a.
Thus m∗ (E ∩ J) + m∗ (E ∩ J c ) ≤ i m∗ (Ii ) ≤ m∗ (E) + ε. or two times. we need to show m∗ (E) ≥ m∗ (E ∩ J) + m∗ (E ∩ J c ).5). d] is m∗ measurable. Since the collection of m∗ measurable sets is a σalgebra.4.2. one. . LEBESGUESTIELTJES MEASURES 25 Proposition 4. each of the form (ai .4) either zero. it suﬃces to show that every interval J of the form (c. Since ε is arbitrary. Now Ii ∩ J is an interval that is open on the left and closed on the right. . one. bi ]. (4. everything is ﬁne. Let E be any set with m∗ (E) < ∞. f ] were not what it is supposed to be.5) Choose I1 ..9. Fortunately. Using (4. such that E ⊂ ∪i Ii and m∗ (E) ≥ i [α(bi ) − α(ai )] − ε. this proves (4. After all this work. and Ii ∩ J c is the union of zero. . Proof. we see that m∗ (Ii ∩ J) + m∗ (Ii ∩ J c ) = m∗ (Ii ). First we need the following lemma. . Since E ⊂ ∪i Ii . I2 . due to Proposition 4. or two such intervals.7 Every set in the Borel σalgebra on R is m∗ measurable. Adding we have m∗ (E ∩ J) + m∗ (E ∩ J c ) ≤ i [m∗ (Ii ∩ J) + m∗ (Ii ∩ J c )]. it would be upsetting if the measure of a halfopen interval (e. we have m∗ (E ∩ J) ≤ i m∗ (Ii ∩ J) and m∗ (E ∩ J c ) ≤ i m∗ (Ii ∩ J c ). depending on the relative locations of Ii and J.
6) Proof. Proposition 4. . By our construction we have ak1 ≤ C < bk1 . Since {Jk1 . D]. If [C. Then I ⊂ ∪∞ Ai . there exists at least one interval. Jkm . At each stage we choose Jkj so that bkj−1 ∈ Jkj . Jn }. D]. and there must be at least one interval. akm < D < bkm . Jk1 . we stop. . . then bk1 < bk2 ≤ D. k=1 (4. Otherwise. say. If not. Since {Jk } is a cover of [C. . . be a ﬁnite collection of ﬁnite open intervals covering a ﬁnite closed interval [C. such that bk1 ∈ Jk2 . akj < bkj−1 < bkj . . bk ). such that C ∈ Jk1 . Jk2 . and there must be at least one interval.26 CHAPTER 4. .9 If e and f are ﬁnite and I = (e. D]. say. say. . We continue until we have covered [C.8 Let Jk = (ak . we will stop for some m ≤ n.6). Let A1 = I and A2 = A3 = · · · = ∅. we stop. and for 2 ≤ j ≤ m. Then n [α(bk ) − α(ak )] ≥ α(D) − α(C). n. D] ⊂ Jk1 ∪ Jk2 . . . Jkm } ⊂ {J1 . . First we show m∗ (I) ≤ (I). If Jk1 covers [C. then m∗ (I) = (I). . k = 1. Proof. . Then α(D) − α(C) ≤ α(bkm ) − α(ak1 ) = [α(bkm ) − α(bkm−1 )] + [α(bkm−1 ) − α(bkm−2 )] + · · · + [α(bk2 ) − α(bk1 )] + [α(bk1 ) − α(ak1 )] ≤ [α(bkm ) − α(akm )] + [α(bkm−1 ) − α(akm−1 )] + · · · + [α(bk2 ) − α(ak2 )] + [α(bk1 ) − α(ak1 )]. . hence i=1 ∞ m∗ (I) ≤ i=1 (Ai ) = (A1 ) = (I). D] with intervals Jk1 . Since {Jk } is a ﬁnite cover. This is easy. . f ]. CONSTRUCTION OF MEASURES Lemma 4. Jk3 that contains bk2 . this proves (4. bk1 ≤ D. . .
suppose I ⊂ ∪∞ Ai . Given a measure µ on R such that µ(K) < ∞ whenever K is compact. . Then [C. Moreover m({x}) = lim m((x − (1/n). n→∞ n→∞ We then conclude m([a. and Exercise 4. If x ∈ R.8. This is possible by the right continuity of α. b]) = m((a. Use compactness to choose a ﬁnite subcover {J1 . right continuous. D] is compact and {Bi } is an open cover for [C. In the special case of Lebesgue measure. di ). x]) if x ≥ 0 and α(x) = −µ((x. A set is Lebesgue measurable if it is in the Lebesgue σalgebra. We now apply Lemma 4. where Ai = (ci .3 Examples and related results Example 4. we obtain (I) ≤ m∗ (I) + ε. Taking the inﬁmum over all countable collections {Ai } that cover I. In the special case where α(x) = x. D]. choose di > di such that α(di ) − α(di ) < ε/2i+1 and let Bi = (ci . deﬁne α(x) = µ((0.1 asks you to show that µ is LebesgueStieltjes measure corresponding to α. . Since ε is arbitrary. We conclude that n ∞ (I) ≤ α(D)−α(C)+ε/2 ≤ (α(dk )−α(ck ))+ε/2 ≤ k=1 i=1 (Ai )+ε. the collection of m∗ measurable sets is called the Lebesgue σalgebra.3. di ]. i=1 Let ε > 0 and choose C ∈ (e. b]) + m({a}) = b − a + 0 = b − a . EXAMPLES AND RELATED RESULTS 27 For the other direction. 0]) if x < 0. (I) ≤ m∗ (I). 4. Let D = f . For each i. Jn } of {Bi }.4. x]) = lim [x − (x − (1/n))] = 0. Then α is increasing. We now drop the asterisks from m∗ and call m LebesgueStieltjes measure.10 Let m be Lebesgue measure. m is Lebesgue measure. f ) such that α(C) − α(e) < ε/2. . . then {x} is a closed set and hence is Borel measurable.
Let F0 be the interval [0. Suppose we deﬁne f0 to be 1 on the interval ( 1 . Adding 0 to itself countably many times is still 0. 8 ). Since the Cantor set C is the intersection 3 of all these sets. it contains no intervals. We will use it in examples later on. there is a rational of the form k/2n with k ≤ 2n that is not in the range of f . the measure of F2 is 4( 1 ). But if it has a jump continuity. 1 F1 = F0 − ( 3 . so the Lebesgue measure of a countable set is 0. uncountable. This function f is called the CantorLebesgue function or sometimes simply the Cantor function. 3 F1 consists of two intervals of length 1 each. The only way this can happen is if f is continuous. However there are uncountable sets which have Lebesgue measure 0. see Proposition 1. which is a set of Lebesgue measure 0. to be 1 on 2 3 3 4 the interval ( 1 .11 Recall from undergraduate analysis that the Cantor set is constructed as follows. and every point is a limit point. Since σalgebras are closed under the operation of countable unions. so it has only jump discontinuities. each of the values {k/2n : n ≥ 0. Deﬁne 9 9 4 9 9 f (x) = inf{f0 (y) : y ≥ x} for x < 1. 8 )]. For now. b)) = m((a. by the construction. to be 3 on the interval ( 7 . The measure of F1 is 2( 1 ). f is increasing. Notice f = f0 on the complement of the Cantor set. 1]. Example 4. 2 ). 2 ) ∪ ( 7 . we note that it a function that increases only on the Cantor set. and so on. 2 ). CONSTRUCTION OF MEASURES m((a. k ≤ 2n } is taken by f0 for some point in the complement of C. We remove the middle 3 third of each of these two intervals and let F2 = F1 − [( 1 . Moreover. . 2 ). yet f is continuous.28 and CHAPTER 4. then countable sets are Borel measurable. and so is taken by f . On the other hand.6. Deﬁne f (1) = 1. and the Cantor set F is ∩n Fn . We let F1 be what remains if we remove the middle third. and the 3 9 measure of Fn is ( 2 )n . that is. Recall that the Cantor set is closed. the Lebesgue measure of C is 0. 9 9 9 9 We continue removing middle thirds. b]) − m({b}) = b − a − 0 = b − a. See the next example.
But A contains no rational numbers. Let m be Lebesgue measure. (A similar argument to what follows is possible for subsets of R that have inﬁnite measure. Instead of removing the middle third at the ﬁrst stage. Such a set is called a generalized Cantor set. 5 ). etc. (1) Given ε > 0. Let A ⊂ [0. Of course. is uncountable. with this diﬀerence. there exists a closed set F so that m(A−F ) < ε and F ⊂ A. On each of the four intervals that remain. Then the measure of Ii is ε/2i−1 . so the measure of ∪i Ii is at most 2ε. is closed. and has measure 1/2.3. The total that we removed is 1 4 1 1 + 2( 16 ) + 4( 64 ) + · · · = 1 . EXAMPLES AND RELATED RESULTS 29 Example 4. there exists an open set G so that m(G−A) < ε and A ⊂ G. remove ( 3 . .) Proposition 4. (It is not equal to that because there is a lot of overlap. i.4.14 Suppose A ⊂ [0. .2. and so on. Example 4. every point is a limit point.) Therefore the measure of A = [0. We will show that A is “almost equal” to the countable intersection of open sets and “almost equal” to the countable union of closed sets. remove the middle interval 1 of length 64 . other choices than 1 . 1] − ∪i Ii is larger than 1 − 2ε. 1] be a Borel measurable set.e.12 Let q1 . are possible. . and let Ii be the interval (qi − ε/2i . remove the middle fourth.. . 4 1 16 . q2 . qi + ε/2i ). be an enumeration of the rationals.13 Let us follow the construction of the Cantor set. (2) Given ε > 0. 1] is a Borel measurable set. remove the middle sixteenths. (3) There exists a set H which contains A that is the countable intersection of a decreasing sequence of open sets and m(H − A) = 0. let ε > 0. On each 8 8 of the two intervals that remain. (4) There exists a set F which is contained in A that is the countable union of an increasing sequence of closed sets which is contained in A and m(A − F ) = 0. see Exercise 4. 2 The set that remains contains no intervals.
Therefore m(G − A) ≤ m(G − E) + m(E − A) < ε. is open. Let H = ∩∞ Hi . F ⊂ A. which are not so bad. the G is for geoﬀnet. apply (3) to A to ﬁnd a set H containing A that is the countable intersection of a decreasing sequence of open sets and such that m(H − A ) = 0. The set H is a Borel set. (4) If A = [0. Let J = [0. and m(A − F ) ≤ m(G − A ) < ε.” and the σ coming from e Summe. the German word for “intersection.1). the F coming from ferm´. the German word for “open” and the δ for Durchschnitt. when trying to understand Lebesgue measure. H need not be open. j=1 Then G is open and contains A and ∞ m(G − E) < j=1 ε2−j−1 = ε/2.30 CHAPTER 4. and since it is contained in Gi . and at null sets. 1] − G.” The countable unions of closed sets are called Fσ sets. which can be quite bad but don’t have positive measure. but it is the intersection of i=1 countably many open sets. (2) Find an open set G such that A ⊂ G and m(G − A ) < ε. Then F is closed. bj + ε2−j−1 ). 4. 1] − A. bj ] such j=1 that A ⊂ E and m(E − A) < ε/2. we can look at Gδ or Fσ sets. Let G = ∪∞ (aj . (1) There exists a set of the form E = ∪∞ (aj . Then Hi = ∩i Gj will contain j=1 A. the German word for “union. 1] − A.15 Let m∗ be deﬁned by (4. and m(H − A) ≤ m(Hi − A) < 2−i for each i. 1] − H. contains A. hence m(H − A) = 0. where C is the collection of intervals that are open on the left and closed on the right . there is an open set Gi that contains A and such that m(Gi − A) < 2−i . CONSTRUCTION OF MEASURES Proof.4 Nonmeasurable sets Theorem 4. the French word for “closed. (3) By (1). It is left to the reader to verify that J has the desired properties. then m(Hi − A) < 2−i . for each i. where A = [0. The countable intersections of open sets are sometimes called Gδ sets.” Therefore. Let F = [0.
and q.´ 4.5. so 1≤ q∈[−1. the sets A + q are disjoint for diﬀerent rationals q. Now [0. b]) for each a. Given a set B. we have m∗ (A + q) = m∗ (A) for each set A and each q.q∈Q m∗ (A + q). b. Note that ((a + q.1 . 4. m∗ is not a measure on the collection of all subsets of R. For each equivalence class.1]. then 3≥ m∗ (A + q). This e theorem abstracts some of the techniques used above to give a tool for constructing measures in a variety of contexts. q∈[0. Let A0 be an algebra but not necessarily a σalgebra.1].q∈Q which implies m∗ (A) = 0. Proof. 2]. so if m∗ is a measure. where the sum is only over rational q. THE CARATHEODORY EXTENSION THEOREM 31 and ((a. pick an element out of that class (we need to use the axiom of choice to do this). 1]. deﬁne B + x = {y + x : y ∈ B}. It is easy to see that this is an equivalence relationship on [0. and therefore m∗ (A) > 0. b + q]) = b − a = ((a.1]∩Q (A + q). Call the collection of such points A. But ∪q∈[−1. Saying is a measure on A0 means the following: (1) of Deﬁnition 3. Moreover. and so by the deﬁnition of m∗ . where again the sum is only over rational q. a contradiction. Deﬁne x ∼ y if x − y is rational. b]) = b − a.1]∩Q (A + q) ⊂ [−1.5 The Carath´odory extension theoe rem We prove the Carath´odory extension theorem in this section. Suppose m∗ is a measure. 1] ⊂ ∪q∈[−1.
. . we look at (4). Since ε is arbitrary. Pick B1 . One is µ∗ and let the . µ∗ (E) ≥ µ∗ (E ∩ A) + µ∗ (E ∩ Ac ). (3) every set in A0 is µ∗ measurable. We know µ∗ (E) ≤ (E) since we can take A1 = E and A2 . Suppose we have two extensions to σ(A0 ). . ∈ A0 such that E ⊂ ∪∞ Bi and i (Bi ) ≤ i=1 µ∗ (E) + ε. A3 . . (1) is Proposition 4.2. If E ⊂ ∪∞ Ai with Ai ∈ A0 . A2 . . Then ∞ ∞ ∞ µ∗ (E) + ε ≥ i=1 (Bi ) = i=1 (Bi ∩ A) + i=1 (Bi ∩ Ac ) ≥ µ∗ (E ∩ A) + µ∗ (E ∩ Ac ). We turn to (2). Therefore ∞ ∞ (E) = i=1 (Bi ) ≤ i=1 (Ai ). ∞] is µ (E) = inf i=1 ∗ (Ai ) : each Ai ∈ A0 . Theorem 4. . and their union is E. shows that (E) ≤ µ∗ (E). Suppose A ∈ A0 . . Next we look at (3).32 CHAPTER 4. empty in the deﬁnition of µ∗ . A2 . Sometimes one i=1 i=1 calls a measure on an algebra a premeasure. Thus A is µ∗ measurable. Recall σ(A0 ) is the σalgebra generated by A0 . Deﬁne ∞ : A0 → [0. Then the Bn are pairwise disjoint. . then there is a unique extension to σ(A0 ). Let ε > 0 and let E ⊂ X. . then (∪∞ Ai ) = i=1 (Ai ). Proof. . . are pairwise disjoint elements of A0 and ∞ also ∪∞ Ai ∈ A0 . CONSTRUCTION OF MEASURES holds and if A1 . Then (1) µ∗ is an outer measure. the smallest σalgebra containing A0 . E ⊂ ∪∞ Ai i=1 for E ⊂ X. (4) if is σﬁnite. i=1 they are each in A0 . (2) µ∗ (A) = (A) if A ∈ A0 . Suppose E ∈ A0 . Finally. Taking the inﬁmum over all such sequences A1 .16 Suppose A0 is an algebra and a measure on A0 . B2 . . let i=1 Bn = E ∩ (An − (∪n−1 Ai )).
5. Observe i=1 i=1 µ∗ (E) + ε ≥ i ∗ (Ai ) = i µ∗ (Ai ) ≥ µ∗ (∪i Ai ) = µ∗ (A). By the preceding paragraph we have uniqueness for the measure i deﬁned by i (A) = (A ∩ Ki ). THE CARATHEODORY EXTENSION THEOREM 33 other extension be called ν. hence µ (A − E) ≤ ε. Write X = ∪i Ki . . i ν(Ai ) = which implies ν(E) ≤ µ∗ (E). then ν(E) ≤ i ν(Ai ). It remains to consider the case when is σﬁnite. (4. i→∞ which proves µ = ν. so i (Ai ) = with each Ai ∈ A0 . then µ∗ (E) = ν(E). The µ∗ measurable sets form a σalgebra containing A0 . Let us ﬁrst assume that µ∗ is a ﬁnite measure.7) in the next to last inequality. using (4. we must use a diﬀerent argument to get the reverse inequality. k→∞ k→∞ Then µ∗ (E) ≤ µ∗ (A) = ν(A) = ν(E) + ν(A − E) ≤ ν(E) + µ∗ (A − E) ≤ ν(E) + ε. then µ(A) = lim µ(A ∩ Ki ) = lim i→∞ i→∞ i (A) = lim ν(A ∩ Ki ) = ν(A). Since ε is arbitrary. i=1 i But = ν on A0 . this completes the proof when is ﬁnite. Because E ∈ A0 . Let ε > 0 and choose Ai ∈ A0 such that µ∗ (E) + ε ≥ i (Ai ) and E ⊂ ∪i Ai .7) Since we do not know that ν is constructed via an outer measure. Therefore if E ⊂ ∪∞ Ai i=1 (Ai ). If µ and ν are two extensions of and A ∈ σ(A0 ).´ 4. E must be µ∗ measurable and ∞ µ∗ (E) = inf i=1 (Ai ) : E ⊂ ∪∞ Ai . each Ai ∈ A0 . where Ki ↑ X and (Ki ) < ∞ for each i. We have µ∗ (A) = lim µ∗ (Bk ) = lim ν(Bk ) = ν(A). Let A = ∪∞ Ai and Bk = ∪k Ai . We will show that if E is in σ(A0 ).
Show there exist G open and F closed such that F ⊂ A ⊂ G and m(G − F ) < ε. An is a Lebesgue measurable subset of [0. Show that for all Lebesgue measurable sets A we have m(x+A) = m(A) and m(cA) = cm(A). Exercise 4.4 Let m be LebesgueStieltjes measure corresponding to a right continuous increasing function α.2 Let m be Lebesgue measure and A a Lebesgue measurable subset of R with m(A) < ∞. . 1]. Let ε > 0.6 Let m be Lebesgue measure.1 Let µ be a measure on the Borel σalgebra of R such that µ(K) < ∞ whenever K is compact. Show that each set in A is µ∗ measurable and µ∗ agrees with the measure µ on A. Exercise 4.5 Suppose m is Lebesgue measure. B ∈ A} for all subsets A of X. 1) and m is Lebesgue measure. Find a measurable set E ⊂ [0. ∞ (3) If n=1 m(An ) < ∞.6 Exercises Exercise 4. y→x− Exercise 4. Let B consist of those points x that are in inﬁnitely many of the An . Exercise 4. Exercise 4. Show that µ is the LebesgueStieltjes measure corresponding to α. (1) Show B is Lebesgue measurable.34 CHAPTER 4. CONSTRUCTION OF MEASURES 4. (2) If m(An ) > δ > 0 for each n. Suppose for each n.7 Suppose ε ∈ (0. µ) is a measure space. Show that for each x.3 If (X. 1] and m(E) = ε. 0]) if x < 0. but m(B) = 0. A. m({x}) = α(x) − lim α(y). Show that µ∗ is an outer measure. deﬁne µ∗ (A) = inf{µ(B) : A ⊂ B. show m(B) ≥ δ. prove that m(B) = 0. Exercise 4. x]) if x ≥ 0 and α(x) = −µ((x. ∞ (4) Give an example where n=1 m(An ) = ∞. Deﬁne x + A = {x + y : y ∈ A} and cA = {cy : y ∈ A}. deﬁne α(x) = µ((0. 1] such that the closure of E is [0.
Deﬁne µ∗ using as in (4.1). Prove that if m(A ∩ I) ≤ (1 − ε)m(I) for every interval I. and µ is a measure on (X. Show that if F is a closed subset of [0. Exercise 4. 1] whose support is F . m(An ∆Am ) > 0 if n = m.4. Find an example of Lebesgue measurable subsets A1 . y ∈ A}.12 Let m be Lebesgue measure. Exercise 4. Exercise 4. then B contains a nonempty open interval centered at the origin. B). Exercise 4.6.9 Let m be Lebesgue measure. .14 Let m be Lebesgue measure. then there exists a ﬁnite measure on [0.11 Suppose m is Lebesgue measure and A is a Borel measurable subset of R with m(A) > 0. Exercise 4. Prove that if A is a Lebesgue measurable subset of R and m(A) > 0. B is the Borel σalgebra. . then there is a subset of A that is nonmeasurable. then m(A) = 0.15 Let X be a set and A a collection of subsets of X that form an algebra of sets. Exercise 4. 1]. . Exercise 4. and suppose A is a Borel measurable subset of R. 1). Prove that if A ⊂ N and A is Lebesgue measurable. let m be Lebesgue measure.4. Construct a Borel subset A of R such that 0 < m(A ∩ I) < m(I) for every open interval I. of [0.10 Let ε ∈ (0. Suppose is a measure on A such that (X) < ∞. EXERCISES 35 Exercise 4. . Prove that a set A is µ∗ measurable if and only if µ∗ (A) = (X) − µ∗ (Ac ). and m(An ∩ Am ) = m(An )m(Am ) if n = m. Prove that if B = {x − y : x. then the support of µ is the smallest closed set F such that µ(F c ) = 0. 1] such that m(An ) > 0 for each n.13 Let N be the nonmeasurable set deﬁned in Section 4. A2 .8 If X is a metric space. then m(A) = 0.
36 CHAPTER 4.16 Suppose µ∗ is an outer measure. An ↓ A does not necessarily imply µ∗ (An ) ↓ µ∗ (A). Exercise 4.17 Suppose A is a Lebesgue measurable subset of R and B = ∪x∈A [x − 1. Prove that B is Lebesgue measurable. then µ∗ (An ) ↑ µ∗ (A). CONSTRUCTION OF MEASURES Exercise 4. Given an example to show that even if µ∗ is ﬁnite. Show that if An ↑ A. x + 1]. .
Chapter 5 Measurable functions We are now ready to move from sets to functions.5 Suppose f is realvalued.1 A function f : X → R is measurable or A measurable if {x : f (x) > a} ∈ A for all a ∈ R. A complexvalued function is measurable if both its real and complex parts are measurable. and so f is measurable.2 Suppose f is realvalued and identically constant. A). or X. Proposition 5. Deﬁnition 5. 37 .1 Measurability Suppose we have a measurable space (X. ∞). Example 5. 5. A. Then the set {x : f (x) > a} is either empty or all of X. Then the set {x : f (x) > a} is either ∅. The following conditions are equivalent. Example 5.4 Suppose X is the real line with the Borel σalgebra and f (x) = x. so f is measurable. Example 5.3 Suppose f (x) = 1 if x ∈ A and 0 otherwise. Hence f is measurable if and only if A is in A. Then {x : f (x) > a} = (a.
7 Let c ∈ R. Proposition 5. g). Note that {x : f (x) > a} = f −1 ((a.38 (1) (2) (3) (4) CHAPTER 5. Proof.. then {x : cf (x) > a} = {x : f (x) > a/c} shows cf is measurable. If f is measurable. g).2. This proves f + g is measurable. a ∈ R. and min(f. a ∈ R.g. MEASURABLE FUNCTIONS {x : f (x) > a} ∈ A {x : f (x) ≤ a} ∈ A {x : f (x) < a} ∈ A {x : f (x) ≥ a} ∈ A for for for for all all all all a ∈ R. cf is measurable by Example 5. and f : X → R is continuous. Hence {x : f (x) + g(x) < a} = ∪r∈Q ({x : f (x) < r} ∩ {x : g(x) < a − r}). If f and g are measurable realvalued functions. e. then −f is measurable using Proposition 5. max(f. then (1) holds by using the equality {x : f (x) > a} = ∪∞ {x : f (x) ≥ a + 1/n}. n=1 This completes the proof. ∞)) is open.5. and there exists a rational r such that f (x) < r < a − g(x). Proposition 5. If f (x) + g(x) < a. a ∈ R. f g. {x : f (x) ≤ a} = {x : f (x) > a}c . Proof. and hence in A. cf . When c = 0. If c > 0. then so are f + g. If (4) holds. When . Since {x : −f (x) > a} = {x : f (x) < −a}. Proof.6 If X is a metric space. A contains all the open sets. The equivalence of (1) and (2) and of (3) and (4) follow from taking complements. then f (x) < a − g(x). −f . then {x : f (x) ≥ a} = ∩∞ {x : f (x) > a − 1/n} n=1 shows that (4) holds if (1) does. then f is measurable.
we say f is Borel measurable. We have {x : supi fi > a} = ∩∞ {x : fi (x) > a}. {x : f (x) > a} = X. The same is true for increasing functions on the real line. and lim inf i→∞ fi . if the set of x where fi (x) does not converge to f (x) has measure zero.8 If fi is a measurable realvalued function for each i. which is measurable by what we have already proved. write cf = −(cf ). f 2 is measurable since for a < 0. The measurability of f g follows since f g = 1 [(f + g)2 − f 2 − g 2 ]. and the result for min(f. 2 The equality {x : max(f (x). g) is measurable. . written f = g a.5. g(x)) > a} = {x : f (x) > a} ∪ {x : g(x) > a} shows max(f.e. √ √ {x : f (x)2 > a) = {x : f (x) > a} ∪ {x : f (x) < − a}. g) follows from min(f. i=1 and the proof for inf fi is similar. g) = − max(−f. so supi fi is measurable. lim supi→∞ fi .e. If X is a metric space. inf i fi . Deﬁnition 5. if {x : f (x) = g(x)} has measure zero. and f : X → R is measurable with respect to B. We saw in Proposition 5. B is the Borel σalgebra. MEASURABILITY 39 c < 0. If f : R → R is measurable with respect to the Lebesgue σalgebra.9 We say f = g almost everywhere.6 that all continuous functions are Borel measurable.. The result will follow for lim sup and lim inf once we have the result for the sup and inf by using the deﬁnitions since lim supi fi = inf j supi≥j fj and similarly for the lim inf. then so are supi fi . −g). we say fi → f a. we say f is Lebesgue measurable. Proposition 5. while for a ≥ 0. Proof.1. Similarly.
the Cantor set. we have that C is closed under countable unions. then {x : f (x) > a} = [x0 . .12 Let us construct a set that is Lebesgue measurable. On the other hand. MEASURABLE FUNCTIONS Proposition 5. ∞). Proof. Since f is measurable. Since F (A) ⊂ C and m(C) = 0. C contains (a. then since f −1 (∪i Ai ) = ∪i f −1 (Ai ) ∈ A. Although F is not continuous. for otherwise we look at −f . If A1 . Similarly C is closed under countable intersections and complements. . A2 . then f is Borel measurable. If f (x0 ) > a.15. because if it were. let x0 = sup{y : f (y) ≤ a}. C contains B. Let B be the Borel σalgebra on R and C = {A ∈ B : f −1 (A) ∈ A}. F −1 maps Borel measurable sets to Borel measurable sets. then A = F −1 (F (A)) would be Borel measurable. ∈ C. A) be a measurable space and let f : X → R be a A measurable function. 1] into C. Let f be the CantorLebesgue function of Example 4. Proof. hence C contains the σalgebra generated by these intervals. Let m be Lebesgue measure and let A be the nonmeasurable set we constructed in Proposition 4.40 CHAPTER 5. If A is in the Borel σalgebra on R. then f −1 (A) ∈ A. . but not Borel measurable. Recall the Lebesgue measurable sets were constructed in Chapter 4 and include the completion of the Borel σalgebra.11 Let (X. observe that F is strictly increasing (hence onetoone) and maps [0. Let us suppose f is increasing.11 and deﬁne F (x) = inf{y : f (y) ≥ x}. ∞). hence is Lebesgue measurable. In either case {x : f (x) > a} is a Borel set. If f (x0 ) ≤ a. F (A) is not Borel measurable. Proposition 5. a contradiction. Since F is increasing. Example 5. that is. Given a ∈ R. so C is a σalgebra. Let B = F (A). . then {x : f (x) > a} = (x0 . then F (A) is a null set. ∞) for every real a.10 If f : R → R is monotone.
and i = 1.2. . x ∈ E. deﬁne the characteristic function of E by χE (x) = 1 0 x ∈ E. If E ∈ A. 5.3 Lusin’s theorem The following theorem is known as Lusin’s theorem.2 Approximation of functions Deﬁnition 5. APPROXIMATION OF FUNCTIONS 41 5. . If f (x) is between (i − 1)/2n and i/2n for i/2n ≤ n. Then deﬁne n2n sn = i=1 i−1 χEni + nχFn . Proof.14 Suppose f is a nonnegative and measurable function. .5. Then there exists a sequence of nonnegative measurable simple functions increasing to f . A) be a measurable space. . n2n . / A simple function s is a function of the form n s(x) = i=1 ai χEi (x) for real numbers ai and measurable sets Ei . Proposition 5.13 Let (X. fn (x) = n if f (x) ≥ n. 2n In words. It is very pretty but usually other methods are better for solving problems. . 2. It is easy to see that sn has the desired properties. . 2. we let fn (x) = (i − 1)/2n . . Let Eni = {x : (i − 1)/2n ≤ f (x) < i/2n } and Fn = {x : f (x) ≥ n} for n = 1. .
1]. E is compact and δ > 0. and ε > 0. where each Ai is a measurable subset of [0. If we let F = ∩M Fi . Let Ain = {x : (i − 1)/2n ≤ f (x) < i/2n }. 1] − F ) ≤ m(G − E) < ε. takes values in [0. 1] − F ) < ε and the restriction of f to F is a continuous function on F . Take F = (E ∪ Gc ) ∩ [0. 1]. and equal to 0 on Gc . By Proposition 4. we see that g is continuous. 1] → R is Borel measurable. bounded by K. First let us suppose that f = χA . m([0. Then m([0. Choose Fi closed such that m([0. E) = inf{x − y : y ∈ E}. Note that hn (x) = fn+1 (x) − fn (x) . is equal to 1 on E. There exists a closed set F ⊂ [0. Recall that the support of a function f is the closure of the set {x : f (x) = 0}.14 we can ﬁnd E closed and G open such that E ⊂ A ⊂ G and m(G − A) < ε/2 and m(A − E) < ε/2. We use m for Lebesgue measure. Theorem 5. y ∈ Gc }.15 Suppose f : [0. This theorem can be loosely interpreted as saying every measurable function is “almost continuous. then F is closed. 1] − F ) < ε. 1] such that m([0. where A is a Borel measurable subset of [0.16 will illustrate why this is a less useful theorem than at ﬁrst glance. MEASURABLE FUNCTIONS Example 5. where y + = max(y. and f = g on F . E) δ + . 1] − Fi ) < ε/M and χAi restricted to Fi is continuous. and has support in [0. 0) and d(x. 1].42 CHAPTER 5. m is Lebesgue measure. Then fn (x) = i=1 K2n +1 M i χA (x) 2n in are simple functions increasing to f . Since E ⊂ A ⊂ [0. Let δ = inf{x − y : x ∈ E. 1]. Letting g(x) = 1 − d(x. and f i=1 restricted to F is continuous.” Proof. Now suppose f is nonnegative. 1]. 1] and each ai ≥ 0. Next suppose f = i=1 ai χAi is simple.
1] − E) < ε/3. 1]. Recall Example 4. choose Fn closed such that m([0. No claim is made that the function f is continuous at most points of [0. hence f takes the values 0 and 1 in every neighborhood of x. Now choose E ⊂ [0. + Example 5. If f ≥ 0. hence m(BK ) > 1 − ε/3 if K is suﬃciently large. . On the set F . 1] is a point of discontinuity of f because for any x ∈ [0. 1]. 1] closed such that f · χD restricted to E is continuous and m([0. m([0. and thus the union of countably many closed sets.3. 1]−F − ) < ε/2. 1]. being n=0 the intersection of closed sets. f is Borel measurable because [0. LUSIN’S THEOREM 43 is also a simple function and is bounded by 2−n . This does not contradict Lusin’s theorem. 1] as K → ∞. hence f restricted to A is a continuous function. f restricted to the set A there is identically one. A is closed because it is equal to the interval [0.16 Suppose f = χB . Choose F0 closed such that m([0.5.12. Then F = D ∩ E is closed. 1] − Fn ) < ε/2n+1 and hn restricted to Fn is continuous. let BK = {x : f (x) ≤ K}. Finally. For n ≥ 1. hence the union of countably many points. for arbitrary measurable f write f = f + − f − and ﬁnd F and F − closed such that m([0. BK ↑ [0. and f + restricted to F + is continuous and similarly for f − . and f restricted to F is continuous. where B consists of the irrationals in [0. 1] minus the union of open intervals. we have f0 (x)+ n=1 hn (x) converges uniformly to f (x) because each hn is bounded by 2−n . 1]−F + ) < ε/2. What is asserted is that there is a closed set F with large measure so that f restricted to F is continuous when viewed as a function from F to R. and m([0. there are both rationals and irrationals in every neighborhood of x. Then F = F + ∩ F − is the desired set. Every point of [0. Since f is everywhere ﬁnite. 1] − F ) < ε. 1]−Fn ) < ε. The uniform limit of continuous functions is continuous. hence f is continuous on F . Choose D ⊂ BK such that D is closed and m(BK −D) < ε/3. 1]−F0 ) < ε/2 and f0 restricted to F0 is continuous. m([0. 1] − B is countable. Then F . 1] − F ) ≤ ∞ ∞ n=0 m([0. will be closed. Let F = ∩∞ Fn .
7 Suppose f : R → R is Borel measurable.44 CHAPTER 5. Prove that A = {x : lim fn (x) exists} n→∞ is a measurable set. (A is allowed to be uncountable. 1) there exist r > 0 and a Borel measurable function g.1) .8 One can show that there exist discontinuous realvalued functions f such that f (x + y) = f (x) + f (y) (5. Exercise 5.6 Suppose f : X → R is Lebesgue measurable and g : R → R is continuous. both depending on x.1 Suppose (X. Deﬁne A to be the smallest σalgebra containing the sets {x : f (x) > a} for every a ∈ R. Exercise 5.e. A) is a measurable space. Prove that g ◦ f is Lebesgue measurable. Exercise 5.) Exercise 5.2 Let f : (0. Exercise 5. Prove that f is Borel measurable. f is a realvalued functions. and {x : f (x) > r} ∈ A for each rational number r. Suppose g : R → R is measurable with respect to A. Prove that there exists a Borel measurable function h : R → R such that g = h ◦ f. then g is ﬁnite for all x but g is nonmeasurable. prove that there exists a Borel measurable function g such that f = g a. Exercise 5.4 Exercises Exercise 5.3 Suppose fn are measurable functions. MEASURABLE FUNCTIONS 5.4 If f : R → R is Lebesgue measurable. 1) → R be such that for every x ∈ (0.5 Give an example of a collection of measurable nonnegative functions {fα }α∈A such that if g is deﬁned by g(x) = supα∈A fα (x). which means that {x : g(x) > a} ∈ A for every a ∈ R. such that f and g agree on B(x. Prove that f is measurable. r). Is this true if g is Borel measurable instead of continuous? Is this true if g is Lebesgue measurable instead of continuous? Exercise 5.
y ∈ R.4. . (The construction uses Zorn’s lemma.) Prove that if f satisﬁes (5. then f is continuous. which is equivalent to the axiom of choice. EXERCISES 45 for all x.1) and in addition f is Lebesgue measurable.5.
46 CHAPTER 5. MEASURABLE FUNCTIONS .
3) . (6. µ) be a measure space. deﬁne the Lebesgue integral of s to be n s dµ = i=1 ai µ(Ei ). We only give the deﬁnition here and consider the properties of the Lebesgue integral in later chapters. (6. s simple . Provided f + dµ and f − dµ are not both inﬁnite. A. If f ≥ 0 is a measurable function.Chapter 6 The Lebesgue integral In this chapter we deﬁne the Lebesgue integral. deﬁne f dµ = sup s dµ : 0 ≤ s ≤ f. 0). if ai = 0 and µ(Ei ) = ∞. (6.1 Deﬁnitions n Deﬁnition 6. If s= i=1 ai χEi is a nonnegative measurable simple function.2) Let f be measurable and let f + = max(f.1) Here. we use the convention that ai µ(Ei ) = 0. 0) and f − = max(−f. deﬁne f dµ = f + dµ − 47 f − dµ. 6.1 Let (X.
then aµ(X) ≤ f dµ ≤ bµ(X). then cf dµ = c f dµ for all complex c. then The integral f χA dµ is often written A f dµ. it is usual to write f (x) dx for f (x) m(dx) and to . Secondly. Other notation for the integral is to omit the µ and write f if it is clear which measure is being used.1. THE LEBESGUE INTEGRAL Finally. to write f (x) µ(dx). (2) If f and g are measurable. A function s might be written as a simple function in more than one way. if f = u + iv is complexvalued and (u + v) dµ is ﬁnite. we say f is The proof of the next proposition follows from the deﬁnitions. one has to think a moment to verify that the deﬁnition of s dµ by means of (6.48 CHAPTER 6.2). f χA dµ = 0. realvalued. then f dµ ≤ g dµ.4) A few remarks are in order.3 (1) If f is a realvalued measurable function with a ≤ f (x) ≤ b for all x and µ(X) < ∞. (6. Proposition 6. or to write f (x) dµ(x).2 If f is measurable and integrable. f  dµ < ∞.5) We leave the proof of this to the reader as Exercise 6. (4) If µ(A) = 0 and f is integrable.1) agrees with its deﬁnition by means of (6. (6. if s is a simple function. Deﬁnition 6. For example s = χA∪B = χA + χB if A and B are disjoint. (3) If f is integrable. deﬁne f dµ = u dµ + i v dµ. and integrable and f (x) ≤ g(x) for all x. It is not hard to check that the deﬁnition of s dµ is unaﬀected by how s is written. When we are integrating a function f with respect to Lebesgue measure m. then we need to show m n ai µ(Ai ) = i=1 j=1 bj µ(Bj ). If m n s = i=1 ai χAi = j=1 bj χBj .
Exercise 6. We will see this 6.b] f (x) m(dx). Proof. f is a complex number. we have f = Re e−iθ f = Re (e−iθ f ) ≤ f . Also −f ≤ f .3 Let f be a nonnegative measurable function. f ≤ f .2 Suppose f is nonnegative and measurable and µ is σﬁnite. so f ≤ f . Proposition 6.6. then f = reiθ for some r and θ. the inequality is trivial. For the complex case. For the real case. Now combine these two facts.5). f + g. From the deﬁnition of f when f is complex. we have Re ( f ) = Re (f ). so − f ≤ f . We do not yet know that (f + g) = in Theorem 7.1 Verify (6. Prove that lim (f ∧ n) → f. Since  f  is real. Show there exist simple functions sn increasing to f at each point such that µ({x : sn (x) = 0}) < ∞ for each n. Then f = r = e−iθ f = e−iθ f. f ≤ f .4. If it is 0.2. Exercise 6. If it is not. n→∞ . EXERCISES deﬁne a 49 b f (x) dx = [a.4 If f is integrable.2 Exercises Exercise 6. this is easy.
6 If fn is a sequence of nonnegative integrable functions such that fn (x) decreases to f (x) for every x. Exercise 6. ∞) such that limx→∞ g(x) = ∞ and g ◦ f is integrable. µ) be a measure space and suppose f is a nonnegative. . A.7 Let (X. This is sometimes called the bounded convergence theorem. Exercise 6.5 Suppose µ(X) < ∞ and fn is a sequence of bounded realvalued measurable functions that converge to f uniformly. µ) be a measure space and suppose µ is σﬁnite. THE LEBESGUE INTEGRAL Exercise 6. prove that fn dµ → f dµ. but not necessarily integrable. Prove that fn dµ → f dµ. Prove that there exists a continuous increasing function g : [0. Suppose f is integrable. ∞) → [0.4 Let (X. measurable function that is ﬁnite at each point of X. Prove that given ε there exists δ such that f (x) µ(dx) < ε A whenever µ(A) < δ.50 CHAPTER 6. A. Exercise 6.
and the dominated convergence theorem. 51 m . Let L be the limit. We also prove that the Lebesgue integral is linear. In this chapter we prove the monotone convergence theorem. Then fn dµ → Proof. Fatou’s lemma. Let s = i=1 ai χEi be any nonnegative simple function less than or equal to f and let c ∈ (0.1 Suppose fn is a sequence of nonnegative measurable functions with f1 (x) ≤ f2 (x) ≤ · · · for all x and with n→∞ lim fn (x) = f (x) f dµ. for all x. Theorem 7.Chapter 7 Limit theorems The main reason the Lebesgue integral is so much easier to work with than the Riemann integral is that it behaves nicely when taking limits. Let An = {x : fn (x) ≥ cs(x)}. then L ≤ f .3(2). fn is an increasing sequence of real numbers.1 Monotone convergence theorem One of the most important results concerning Lebesgue integration is the monotone convergence theorem. We must show L ≥ f . Since fn ≤ f for all n. 7. By Proposition 6. 1).
. Therefore L ≥ c s. Example 7. by Proposition 3.5(3) the right hand side converges to m c i=1 ai µ(Ei ) = c s. Then fn ≥ 0. we can prove that the Lebesgue integral is linear. fn → 0 for each x.2 Linearity of the integral Once we have the monotone convergence theorem. Example 7.1/n) .3 Suppose fn = nχ(0. 1).2 Let X = [0. 7. The reason the monotone convergence theorem does not apply here is that the fn are not nonnegative. For each n. then (f + g) dµ = f dµ + g dµ. then An ↑ X. Taking the supremum over all simple s ≤ f . Theorem 7. but fn = 1 does not converge to 0 = 0.52 CHAPTER 7. fn ≥ An m fn ≥ c An sn =c An i=1 m a i χ Ei ai µ(Ei ∩ An ). but fn ↑ f where f = 0 and f = 0. LIMIT THEOREMS Since fn (x) increases to f (x) for each x and c < 1. we obtain L ≥ f. Since c is arbitrary in the interval (0.4 If f and g are nonnegative and measurable or if f and g are integrable. Then fn = −∞. ∞) and fn (x) = −1/n for all x. i=1 =c If we let n → ∞. then L ≥ s. The reason the monotone convergence theorem does not apply here is that the fn do not increase to f for each x.
so that (f + g)+ + f − + g − = f + + g + + (f + g)− . . Next suppose f and g are nonnegative. Write (f + g)+ − (f + g)− = f + g = f + − f − + g + − g − . Thus the theorem holds in this case. . . . Without loss of generality we may assume that A1 . LINEARITY OF THE INTEGRAL 53 Proof. we may also assume ∪m Ai = X = ∪n Bj .7. say. Then sn + tn are simple functions increasing to f + g. Then i=1 j=1 m n f +g = i=1 j=1 (ai + bj )χAi ∩Bj . m n f = i=1 ai χAi and g = j=1 bj χBj . so the result follows from the monotone convergence theorem and (f + g) = lim n→∞ (sn + tn ) = lim n→∞ sn + lim tn = n→∞ f+ g. then f + g is integrable. Bn are pairwise disjoint. and we have m n (f + g) = i=1 j=1 m n (ai + bj )µ(Ai ∩ Bj ) m n = i=1 j=1 m ai µ(Ai ∩ Bj ) + i=1 j=1 n bj µ(Ai ∩ Bj ) = i=1 ai µ(Ai ) + j=1 bj µ(Bj ) = f+ g. . Since f + g ≤ (f  + g) = f  + g < ∞. . . Suppose now that f and g are realvalued and integrable but take both positive and negative values. . Since ai = 0 and bj = 0 are permissible.2. First suppose f and g are nonnegative and simple. Am are pairwise disjoint and that B1 . . Take sn simple and increasing to f and tn simple and increasing to g.
54
CHAPTER 7. LIMIT THEOREMS
Using the result for nonnegative functions, (f + g)+ + Rearranging, (f + g) = = = (f + g)+ − f+ − f+ (f + g)− g+ − g− f− + g− = f+ + g+ + (f + g)− .
f− + g.
If f and g are complexvalued, apply the above to the real and imaginary parts.
Proposition 7.5 Suppose fn are nonnegative measurable functions. Then
∞ ∞
fn =
n=1 n=1
fn .
Proof. Let FN = can write
∞
N n=1
fn . Since 0 ≤ Fn (x) ↑
N
∞ n=1
fn (x), we
fn =
n=1
N →∞
lim
fn
n=1
=
N →∞
lim FN = lim
N
N →∞ ∞
FN fn ,
(7.1)
= lim
N →∞
fn =
n=1 n=1
using the monotone convergence theorem and the linearity of the integral.
7.3
Fatou’s lemma
The next theorem is known as Fatou’s lemma.
7.4. DOMINATED CONVERGENCE THEOREM
55
Theorem 7.6 Suppose the fn are nonnegative and measurable. Then lim inf fn ≤ lim inf fn .
n→∞ n→∞
Proof. Let gn = inf i≥n fi . Then the gn are nonnegative and gn increases to lim inf n fn . Clearly gn ≤ fi for each i ≥ n, so gn ≤ fi . Therefore gn ≤ inf fi . (7.2)
i≥n
If we take the limit in (7.2) as n → ∞, on the left hand side we obtain lim inf n fn by the monotone convergence theorem, while on the right hand side we obtain lim inf n fn . A typical use of Fatou’s lemma is the following. Suppose we have fn → f and supn fn  ≤ K < ∞. Then fn  → f , and by Fatou’s lemma, f  ≤ K.
7.4
Dominated convergence theorem
Another very important theorem is the dominated convergence theorem. Theorem 7.7 Suppose that fn are measurable realvalued functions and fn (x) → f (x) for each x. Suppose there exists a nonnegative integrable function g such that fn (x) ≤ g(x) for all x. Then lim fn dµ → f dµ.
n→∞
Proof. Since fn + g ≥ 0, by Fatou’s lemma, f+ g= (f + g) ≤ lim inf
n→∞
(fn + g) = lim inf fn +
n→∞
g.
Since g is integrable, f ≤ lim inf
n→∞
fn .
(7.3)
56 Similarly, g − fn ≥ 0, so g− f=
CHAPTER 7. LIMIT THEOREMS
(g −f ) ≤ lim inf
n→∞
(g −fn ) =
g +lim inf
n→∞
(−fn ),
and hence − Therefore f ≥ lim sup
n→∞
f ≤ lim inf
n→∞
(−fn ) = − lim sup
n→∞
fn .
fn ,
which with (7.3) proves the theorem. Exercise 7.1 asks you to prove a version of the dominated convergence theorem for complexvalued functions. Example 7.3 is an example where the limit of the integrals is not the integral of the limit because there is no dominating function g. If in the monotone convergence theorem or dominated convergence theorem we have only fn (x) → f (x) almost everywhere, the conclusion still holds. For example, if the fn are measurable, nonnegative, and fn ↑ f a.e., let A = {x : fn (x) → f (x)}. Then fn χA (x) ↑ f χA (x) for each x. Since Ac has measure 0, we see from Proposition 6.3(4) and the monotone convergence theorem that lim
n
fn = lim
n
fn χA =
f χA =
f.
7.5
Exercises
Exercise 7.1 State and prove a version of the dominated convergence theorem for complexvalued functions. Exercise 7.2 The following generalized dominated convergence theorem is often useful. Suppose fn , gn , f , and g are integrable, fn → f a.e., gn → g a.e., fn  ≤ gn for each n, and gn → g. Prove that fn → f . Exercise 7.3 Give an example of a sequence of nonnegative functions fn tending to 0 pointwise such that fn → 0, but there is no integrable function g such that fn ≤ g for all n.
7.5. EXERCISES
57
Exercise 7.4 Suppose (X, A, µ) is a measure space, each fn is integrable and nonnegative, fn → f a.e., and fn → f . Prove that for each A ∈ A fn dµ →
A A
f dµ.
Exercise 7.5 Suppose fn and f are integrable, fn → f a.e., and fn  → f . Prove that fn − f  → 0. Exercise 7.6 Suppose f : R → R is integrable, a ∈ R, and we deﬁne x F (x) =
a
f (y) dy.
Show that F is a continuous function. Exercise 7.7 Let fn be a sequence of nonnegative Lebesgue measurable functions on R. Is it necessarily true that lim sup
n→∞
fn dx ≤
lim sup fn dx?
n→∞
If not, give a counterexample. Exercise 7.8 Find the limit
n n→∞
lim
1+
0
x n
−n
log(2 + cos(x/n)) dx
and justify your reasoning. Exercise 7.9 Find the limit
n n→∞
lim
1−
0
x n
n
log(2 + cos(x/n)) dx
and justify your reasoning. Exercise 7.10 Prove that the limit exists and ﬁnd its value:
1 n→∞
lim
0
1 + nx2 log(2 + cos(x/n)) dx. (1 + x2 )n
Prove that f is integrable and fn → f . Exercise 7. µ) be a measure space. Exercise 7. 1]. 0 Exercise 7. and each fn is integrable.13 Suppose µ(X) < ∞.14 Prove that ∞ k=1 1 =− (p + k)2 1 0 xp log x dx 1−x for p > 0. 1]. fn converges to f uniformly. A. then there exists a subsequence nj such that A fnj (x) dx converges. then there exists a subsequence nj such that Ai fnj (x) dx converges for each i. 1]. (3) Show that there exists a subsequence nj such that A fnj (x) dx converges for each Borel subset A of [0. 1] that is uniformly bounded. Prove that n n→∞ lim f 1+ −n x g(x) dx n2 exists and determine its value. and continuous at 1.15 Let {fn } be a sequence of realvalued functions on [0.16 Let (X. (1) Show that if A is a Borel subset of [0. The sequence is uniformly absolutely continuous if given ε there exists δ such that fn dµ < ε A .11 Prove the limit exists and determine its value: ∞ n→∞ lim ne−nx sin(1/x) dx. (2) Show that if {Ai } is a countable collection of Borel subsets of [0. Is the condition µ(X) < ∞ necessary? Exercise 7. measurable. Exercise 7.58 CHAPTER 7. LIMIT THEOREMS Exercise 7. A family of measurable functions {fn } is uniformly integrable if given ε there exists M such that fn (x) dµ < ε {x:fn (x)>M } for each n.12 Let g : R → R be integrable and let f : R → R be bounded.
21 Suppose µn is a sequence of measures on (X. . Prove that {fn } is uniformly integrable if and only if supn fn  dµ < ∞ and {fn } is uniformly absolutely continuous..18 Suppose µ is a ﬁnite measure. Call the limit µ(A).7. each fn is integrable. f is integrable. (2) Prove that f dµn → f dµ whenever f is bounded and measurable. Exercise 7. EXERCISES for each n if µ(A) < δ.20 Suppose fn is a uniformly integrable sequence of functions deﬁned on [0. Exercise 7.22 Let (X. 59 Suppose µ is a ﬁnite measure.. Suppose µ is a ﬁnite measure. Exercise 7. Prove that {fn } is uniformly integrable. (1) Prove that µ is a measure.5. Prove that there is a subsequence nj 1 such that 0 fnj g dx converges whenever g is a realvalued bounded measurable function.e. Prove that fn → f . (3) Prove that f dµ ≤ lim inf n→∞ f dµn whenever f is nonnegative and measurable.17 The following is known as the Vitali convergence theorem. fn → f a. fn → f a. and fn → f . A) such that µn (X) = 1 for all n and µn (A) converges as n → ∞ for each A ∈ A.19 Suppose µ is a ﬁnite measure and for some ε > 0 sup n fn 1+ε dµ < ∞. 1]. Exercise 7. µ) be a measure space and let f be nonnegative and integrable. Exercise 7. Deﬁne ν on A by ν(A) = A f dµ. Prove that {fn } is uniformly integrable. and {fn } is uniformly integrable. A. Exercise 7.e.
x]) for every x such that µ({x}) = 0. 1]. B) and let µ be another ﬁnite measure on ([0. B) such that f dµn → 0 f dx whenever f is a realvalued continuous function on [0. Deﬁne αn (x) = µn ([0. (4) limn→∞ µn (A) = µ(A) whenever A is a Borel subset of [0. 1] such that f dµ = f dν whenever f is realvalued and continuous on [0. Exercise 7. (2) lim supn→∞ µn (F ) ≤ µ(F ) for all closed subsets F of [0. 1]. 1]).26 Let B be the Borel σalgebra on [0.25 Let B be the Borel σalgebra on [0. 1]. x]) = µ([0. (5) limn→∞ µn ([0. x]). 1]. 1]. 1].r∈Q lim α(r). 1]. . 1]. (2) Prove that there exists a subsequence {nj } such that αn (r) converges for every rational in [0. Let µn be a sequence of ﬁnite measures on ([0. LIMIT THEOREMS (1) Prove that ν is a measure. Exercise 7.24 Let B be the Borel σalgebra on [0. 1]. 1]) → µ([0. Prove that the following are equivalent: (1) f dµn → f dµ whenever f is a continuous realvalued function on [0. 1]. Suppose that g is a bounded measurable function such that the set of discontinuities of g has measure 0. 1] such that µ(∂A) = 0. Exercise 7. 1]. 1]) < ∞. Prove that µ = ν. (2) Prove that if g is integrable with respect to ν. then f g is integrable with respect to µ and g dν = f g dµ. Suppose µn ([0. prove that there exists a subsequence {nj } such that αnj (r) converges. Exercise 7. where ∂A = A − Ao is the boundary of A. Suppose 1 µn are ﬁnite measures on ([0.60 CHAPTER 7. (3) lim inf n→∞ µn (G) ≥ µ(G) for all open subsets G of [0. Prove that 1 g dµn → 0 g dx. (3) Let α(r) = limn→∞ αn (r) for r rational and deﬁne α(x) = r→x. Let µn be a sequence of ﬁnite measures on ([0. 1]. (1) If r is a rational in [0.r>x. B) with supn µn ([0.23 Suppose µ and ν are positive measures on the Borel σalgebra on [0. B). 1]. 1].
EXERCISES This means.7. 1]. Prove that f dµn → f dµ whenever f is a continuous realvalued function on [0. 61 Let µ be the LebesgueStieltjes measure associated with α. r ∈ Q}. since clearly α(r) ≤ α(s) if r < s.5. . that α(x) = inf{α(r) : r > x.
62 CHAPTER 7. LIMIT THEOREMS .
Let A = {x : f (x) > ε}. 63 . and we give an approximation result.Chapter 8 Properties of Lebesgue integrals We present some propositions which imply that a function is zero a.1 Suppose f is realvalued and measurable and for every measurable set A we have A f dµ = 0. . Proof.e. Proposition 8. Hence µ(A) = 0. 2. to conclude µ{x : f (x) > 0} = µ(∪∞ {x : f (x) > 1/n}) n=1 ∞ ≤ n=1 µ({x : f (x) > 1/n}) = 0. Then 0= A f≥ A ε = εµ(A) since f χA ≥ εχA . . 8. Then f = 0 almost everywhere.e. The following two propositions are very useful. We use this argument for ε = 1/n and n = 1.1 Criteria for a function to be zero a. .
Proof. if Gn are open sets decreasing to H.1 we have the following. By linearity. Corollary 8.1 implies f is zero a. Then f = 0 a. PROPERTIES OF LEBESGUE INTEGRALS Similarly µ{x : f (x) < 0} = 0. n An a contradiction.3 Let m be Lebesgue measure and a ∈ R. For any interval [c. Then f = 0 almost everywhere. Proof. G f = 0 for any open set G. Proposition 8. d d c f= c a f− a f = 0.64 CHAPTER 8. then G f = 0. If f is not equal to 0 almost everywhere. d].14 tells us that there exists a sequence Gn of open sets that decrease to a set H where H diﬀers from E by a null set. If E is any Borel measurable set. if G is the ﬁnite union of disjoint intervals. 1 0= f ≥ f ≥ µ(An ). But since f is nonnegative.2 Suppose f is measurable and nonnegative and f dµ = 0. By dominated convergence and Proposition 1. then H f = limn Gn f = 0. As a corollary to Proposition 8.e. Then f= E f χE = f χH = H f = 0. . Suppose x f : R → R is integrable and a f (y) dy = 0 for all x.5. Again by dominated convergence. there exists an n such that µ(An ) > 0 where An = {x : f (x) > 1/n}. Proposition 4.e. This with Proposition 8.
so by taking n large enough. so g − χA  ≤ (χG − χF ) = m(G − F ) < ε. g is 0 on Gc . Without loss of generality.14. Suppose f = χA .8. We can choose G open and F closed such that F ⊂ A ⊂ G and m(G − F ) < ε by Proposition 4. Then there exists a continuous function g with compact support such that f − g < ε. Since F is compact. it is enough to ﬁnd continuous functions g1 and g2 with compact support such that f + − g1  < ε/2 and f − − g2  < ε/2 and to let g = g1 − g2 . then f − g < ε. δ. where each Ai is contained in a bounded interval and each ai > 0. Proof. g is 1 on F .2. We have g − χA  ≤ χG − χF . Theorem 8.n] − g < ε/2. Thus our result holds for characteristic functions of bounded sets. we may assume G is also a bounded set. F ) δ + . 0 ≤ g ≤ 1. there is a minimum distance between F and Gc .2 An approximation result We give a result on approximating a function on R by continuous functions. where A is a bounded Borel measurable set. Hence we may assume f ≥ 0. AN APPROXIMATION RESULT 65 8. the diﬀerence of the integrals will be less than ε/2. Therefore we may in addition assume that f is 0 outside some bounded interval. say. Then g is continuous.4 Suppose f is a Borel measurable realvalued integrable function on R. If f = i=1 ai χAi . and g has compact support.n] increases to f . By monotone convergence f · χ[−n. Let ε > 0. If we ﬁnd g continuous with compact support such that f · χ[−n. If we write f = f + − f − . and we ﬁnd gi continuous with compact p . Let g(x) = 1 − dist (x.
A.66 CHAPTER 8. Exercise 8. where one proves a result for characteristic functions. m is Lebesgue measure. Let sm be a simple function such that sm ≤ f and sm ≥ f − ε/2. 1] → R such that 0 ≤ f ≤ 1 and m({x : f (x) = χA (x)}) < ε.1 This exercise gives a change of variables formula in two simple cases. If f is nonnegative and has compact support. Prove that there exists a continuous function f : [0. there exists A ∈ A such that µ(A) < ∞ and ε+ A f dµ > f dµ. The method of proof. . 1). We choose continuous g with compact support such that sm − g < ε/2 using the preceding paragraphs. Show that if f is an integrable function on the reals and a is a nonzero real number. is very common. Exercise 8.3 Exercises Exercise 8. and ε ∈ (0. then simple functions. Prove that if ε > 0.2 Let (X. Suppose f is nonnegative and integrable. µ) be a σﬁnite measure space. Thus our theorem holds for nonnegative simple functions with compact support. then f (x + a) dx = R R f (x) dx and f (ax) dx = a−1 R R f (x) dx. PROPERTIES OF LEBESGUE INTEGRALS support such that χAi − gi  < ε/ai p.3 Suppose A is a Borel measurable subset of [0. we can ﬁnd simple functions sm supported in a bounded interval increasing to f whose integrals increase to f . then nonnegative functions. and then f − g < ε. then g = i=1 ai gi will be the desired function. 1]. p 8. and then ﬁnally integrable functions.
Prove Jensen’s inequality: f g dµ ≤ X f ◦ g dµ. and let f : R → R be convex. µ). Exercise 8. there exists a real number c such that f (y) ≥ f (x) + c(y − x) for all y ∈ R.5 Find a nonnegative function f on [0. Prove that A = {x : f (x) = 0} has σﬁnite measure.4 Suppose f is a nonnegative integrable function on a measure space (X.8 Recall that a function f : R → R is convex if f (λx + (1 − λ)y) ≤ λf (x) + (1 − λ)f (y) whenever x < y ∈ R and λ ∈ [0. Prove that f is convex. Prove that t→∞ lim tµ({x : f (x) ≥ t}) = 0. n=1 Exercise 8. Let g : X → R be integrable.9 Suppose f is a realvalued function on R such that 1 1 f 0 g(x) dx ≤ 0 f (g(x)) dx whenever g is bounded and measurable. 1] such that t→∞ lim tm({x : f (x) ≥ t}) = 0 but f is not integrable. f (x)). Exercise 8. A. and suppose f is realvalued and integrable with respect to µ. Prove that a measurable nonnegative function f is integrable if and only if ∞ µ({x : f (x) ≥ n}) < ∞. there exists Fn ↑ A such that µ(Fn ) < ∞ for each n. (2) Let (X. this says that the graph of f lies above the line with slope c that passes through the point (x. (1) Prove that if f is convex and x ∈ R.7 Let µ be a measure.8. EXERCISES 67 Exercise 8. 1]. A. not necessarily σﬁnite. where m is Lebesgue measure. that is. .3. Graphically.6 Suppose µ is a ﬁnite measure. Exercise 8. suppose µ(X) = 1. µ) be a measure space. Exercise 8.
Prove that g is equal .68 CHAPTER 8. PROPERTIES OF LEBESGUE INTEGRALS Exercise 8.10 Suppose g : [0. 1] → R is bounded and measurable and 1 f (x)g(x) dx = 0 0 whenever f is continuous and to a constant a.e. 1 0 f (x) dx = 0.
b]. Set R(f ) = inf{U (P. Recall that the Riemann integral on [a. b] is deﬁned as follows: if P = {x0 . f ) : P is a partition} 69 . f ) = i=1 [ sup f (x)] (xi − xi−1 ) i=1 xi−1 ≤x≤xi n [ xi−1 ≤x≤xi inf f (x)] (xi − xi−1 ). if we are looking at the Riemann integral. We show that the Riemann integral of a function exists if and only if the set of discontinuities of the function have Lebesgue measure zero.Chapter 9 Riemann integrals We compare the Lebesgue integral and the Riemann integral. xn } with x0 = a and xn = b is a partition of [a. x1 .1 Comparison with the Lebesgue integral We only consider bounded measurable functions from [a. b] into R. If we are looking at the Lebesgue integral. let n U (P. . . . we write R(f ). . while. 9. f ) = and L(P. we write f . temporarily. and in that case the Riemann integral and Lebesgue integral agree.
To prove the claim. It is not hard to check that adding points to a partition increases L and decreases U . RIEMANN INTEGRALS R(f ) = sup{L(P. deﬁne n TP (x) = and SP (x) = i=1 n [ sup i=1 xi−1 ≤y≤xi f (y)]χ[xi−1 . b] is Riemann integrable if and only if the set of points at which f is discontinuous has Lebesgue measure 0.70 and CHAPTER 9. [ xi−1 ≤y≤xi inf f (y)]χ[xi−1 . and SPi (x) increases at each point. By dominated convergence (recall that f is bounded) (T − S) = lim i→∞ (TPi − SPi ) = lim (U (Pi . We observe that TP = U (P. f ) ↓ R(f ) and L(Pi . The Riemann integral exists if R(f ) = R(f ). f ) = lim i→∞ i→∞ TPi = f . say. x + δ) is contained in the subinterval of Pi that contains x.e. f ) : P is a partition}. U (Pi . If x is not in the null set where T (x) = S(x) nor in ∪i Pi . choose i large enough so that TPi (x) − SPi (x) < ε and then choose δ small enough so that (x − δ. f ) ↑ R(f ). then Pi is an increasing sequence of partitions. the Riemann integral is equal in value to the Lebesgue integral. which we denote R(f ). to T (x). so if we let Pi = ∪j≤i (Qj ∪ Qj ). Proof. since R(f ) = lim U (Pi . say. Finally. f ). SP = L(P. and the common value is the Riemann integral. given ε. and R(f ) = f . We claim that f is continuous at such x. First we show that if f is Riemann integrable. which is countable and hence of Lebesgue measure 0.e.xi ) (x). i→∞ We conclude T = S = f a. then TPi (x) ↓ f (x) and SPi (x) ↑ f (x). We see also that TPi (x) decreases at each point. f ) and If f is Riemann integrable. f ) ↓ R(f ) and a sequence Qi such that L(Qi . f ) ↑ R(f ). Also T (x) ≥ f (x) ≥ S(x). Step 1.xi ) (x). f )) = 0. If P is a partition. and in that case. to S(x).1 A bounded Borel measurable realvalued function f on [a. f ) − L(Pi . Theorem 9. there exists a sequence of partitions Qi such that U (Qi . then f is continuous a.
1] → R such that 1 1 R(f ) = 0 f (x) dx and R(f ) = 0 f (x) dx.3 Deﬁne f (x) on [0. every neighborhood of x contains both rational and irrational points. Let ε > 0. Therefore f is not Riemann integrable. 1] and f = χA . Example 9. 1]. f is discontinuous at every rational. there are only ﬁnitely many rationals r for which f (r) ≥ ε.1 Find a measurable function f : [0. By dominated convergence. 1] → R that is continuous. If x is not in the null set where f is discontinuous. so taking δ less than the distance from x to any of this ﬁnite collection of rationals shows that f (y) − f (x) < ε if y − x < δ. 1]. so f is continuous at no point of [0.e.1] ) exists. Step 2. f ) = This does it. EXERCISES 71 by the dominated convergence theorem. Therefore the set of discontinuities is a countable set. If x is irrational and ε > 0. If x ∈ [0. Now suppose that f is continuous a. b] = [0. U (Pi . is not Lebesgue integrable. hence f is Riemann integrable. Exercise 9. Thus we want f such that 0 f (x) m(dx) = ∞ but lima→0+ R(f χ[a. SPi → f.2 Find a function f : (0. TPi → f Example 9. nor in ∪∞ Pi .2 Let [a. where A is the set of irrational numbers in [0. but where the improper Riemann inte1 gral exists. . Hence f is continuous at x. 9. hence of measure 0. then i=1 TPi (x) ↓ f (x) and SPi (x) ↑ f (x). 1].9. b] into 2i equal parts. 1] to be 0 if x is irrational and to be 1/q if x is rational and equals p/q when in reduced form. f ) = and L(Pi . we see that the Riemann integral and Lebesgue integral agree.2 Exercises Exercise 9. Let Pi be the partition where we divide [a.2.
(1) Prove that µn ([0. 1]. RIEMANN INTEGRALS Exercise 9. 1] → R is integrable. Conclude 1 by Exercise 7.6 Deﬁne f = limδ→0 inf y−x<δ.1.1) is a Riemann sum approximation to R(f ). Exercise 9. where we use the notation of Theorem 9. Exercise 9. then the Riemann sum approximation of f given in (9. realvalued.5 Let f be a bounded.e. then b 2n f (x) µn (dx) = a i=1 f (xi )2−n (9. Exercise 9. x]) for every x ∈ [0.72 CHAPTER 9.3 Suppose f : [0. Let µn be the measure deﬁned by 2n µn (dx) = 2 −n i=1 δxi (dx).1. 1] for each a > 0. (2) Use Exercise 7.a≤y≤b sup f (y). and measurable function.1] ) 1 0 exists. where δy is point mass at y. Show that the limit is equal to f (x) dx.. and the improper Riemann integral a→0+ lim R(f χ(a. b].5. x]) → m([0. Prove that if f = lim δ→0 y−x<δ. f is bounded on (a. b] whose set of discontinuities has measure 0.24 that f dµn → 0 f dx whenever f is continuous. Conclude f is Lebesgue measurable. This provides an alternative proof of Step 2 of Theorem 9.a≤y≤b f (y) and let f be deﬁned as in Exercise 9.25 to see that if f is a bounded and measurable function on [a.1) converges to the Lebesgue integral of f . then f = T a. b] into 2n equal subintervals and pick a point xi out of each subinterval. Note that if f is a bounded measurable realvalued function on [a. (1) Suppose that the set of discontinuities of a bounded realvalued .4 Divide [a.
This provides another proof of Step 1 of Theorem 9. 1] → R such that the set of discontinuities of f has positive measure.8 Find a sequence fn of continuous functions mapping [0. 1]. (3) Find a bounded lower semicontinuous realvalued function f deﬁned on [0. (We will construct such a measure µ in Chapter 11. (4) Find a bounded lower semicontinuous function f : [0. 1] such that the fn increase to a bounded function f which is not Riemann integrable. d]) = (b − a)(d − c). (2) Prove that U (P. 1] such that the set of discontinuities of f is equal to the set of rationals in [0. Such an example shows there is no monotone convergence theorem or dominated convergence theorem for Riemann integrals. (2) Find a bounded lower semicontinuous function f : [0. M ]2 .9. M ]2 . Prove that there exists ε > 0 such that if Aε = {x ∈ [a.e. d] with −M ≤ a ≤ b ≤ M and −M ≤ c ≤ d ≤ M . 1] into [0. b]. . Exercise 9. EXERCISES 73 measurable function f has positive Lebesgue measure.? Exercise 9. then the Lebesgue integral of f with respect to µ is equal to the double (Riemann) integral of f and the two multiple (Riemann) integrals of f .1. 1] → R such that f is continuous everywhere except at x = 1/2. B) such that µ([a. f ) > εm(Aε ) for every partition P on [a. (5) Does there exist a bounded lower semicontinuous function f : [0. b] × [c. M ]2 generated by the collection of sets of the form [a. then f is lower semicontinuous. 1] → R such that f is discontinuous a. f ) − L(P. b] : f (x) − f (x) > ε}.1. Conclude that f is not Riemann integrable.) Prove that if f is continuous with support in [−M. then m(Aε ) > 0.2.9 Let M > 0 and let B be the σalgebra on [−M. (1) Prove that if fn is a sequence of realvalued continuous functions increasing to f .7 A realvalued function on a metric space is lower semicontinuous if {x : f (x) > a} is open whenever a ∈ R and upper semicontinuous if {x : f (x) < a} is open whenever a ∈ R. using the notation of Theorem 9. Exercise 9. Suppose µ is a measure on ([−M. b] × [c.
RIEMANN INTEGRALS .74 CHAPTER 9.
we say a sequence of measurable functions fn converges almost everywhere to f and write fn → f a.Chapter 10 Types of convergence There are various ways in which a sequence of functions fn can converge. (2) If fn → f in measure.1 If µ is a measure. a. Let 1 ≤ p < ∞. then fn converges to f in measure. Proposition 10. and we compare some of them.2 Suppose µ is a ﬁnite measure. a.e.e. if there is a set of measure 0 such that for x not in this set we have fn (x) → f (x). (1) If fn → f . We say fn converges in Lp to f if fn − f p dµ → 0 as n → ∞.. 10. We say fn converges in measure to f if for each ε > 0 µ({x : fn (x) − f (x) > ε}) → 0 as n → ∞.1 Deﬁnitions and examples Deﬁnition 10. 75 . there is a subsequence nj such that fnj → f .e.
5 ∞ χAn (x) µ(dx) → 0. Then µ(Aj ) ≤ 2−j . let n1 = 1. we prove an easy preliminary result known as Chebyshev’s inequality. µ(An ) = This proves (1).. and by dominated convergence. This implies fnj → f on Ac . then x ∈ ∪∞ Aj for some / / j=k k. Example 10. and if we set A = ∩∞ ∪∞ Aj . j=k k→∞ Therefore A has measure 0. let X = R and let fn = χ(n.4 If 1 ≤ p < ∞. Lemma 10. but fn does not converge in measure. To prove (2).76 CHAPTER 10. We have fn → 0 a. µ(A) = lim k→∞ µ(∪∞ Aj ) j=k ≤ lim k→∞ µ(Aj ) ≤ lim 2−k+1 = 0.n+1) . Let ε > 0 and suppose fn → f a. Let Aj = {x : fnj (x) − f (x) > 1/j}. If An = {x : fn (x) − f (x) > ε}. If x ∈ A. Before we prove this. suppose fn → f in measure.e. To see this. then µ({x : f (x) ≥ a}) ≤ f p . TYPES OF CONVERGENCE Proof.e. ap .e. then χAn → 0 a.. The next proposition compares convergence in Lp to convergence in measure. and choose nj > nj−1 by induction so that µ({x : fnj (x) − f (x) > 1/j}) ≤ 2−j . k=1 j=k then by Proposition 3.3 Part (1) of the above proposition is not true if µ(X) = ∞. and so fnj (x) − f (x) ≤ 1/j for j ≥ k.
Let X = S and let fn (x) = χFn (x). Also. but not almost everywhere.7 We give an example where fn → f in measure and in Lp . by Chebyshev’s inequality µ({x : fn (x) − f (x) > ε}) = µ({x : fn (x) − f (x)p > εp }) ≤ as required. . Example 10. we have µ(A) ≤ A f p 1 dµ ≤ p p a a f p dµ.1. Proof. 1] with Lebesgue measure. and in measure.e. then it converges in measure.5 If fn converges to f in Lp . 2π) : eiθ ∈ A}) to be arc length measure on S. This is what we wanted. Then µ(Fn ) ≤ 1/(n + 1) → 0. since fn is either 1 or 0. Let A = {x : f (x) ≥ a}. Let S = {eiθ : 0 ≤ θ < 2π} be the unit circle in the complex plane and deﬁne µ(A) = m({θ ∈ [0. Proposition 10. where n Fn = eiθ : j=1 1 ≤θ≤ j n+1 j=1 1 .1/n) on [0. where m is Lebesgue measure on [0. 2π).10. but does not converge in Lp for any p ≥ 1. j Let f (eiθ ) = 0 for all θ. Since χA ≤ f p χA /ap . If ε > 0. fn − f p →0 εp Example 10. This gives an example where fn converges to 0 a. fn − f p dµ = χFn dµ = µ(Fn ) → 0.6 Let fn = n2 χ(0. so fn → f in measure. DEFINITIONS AND EXAMPLES 77 Proof.
and each point of S is in S − Fn for inﬁnitely many n.e. m=n For ﬁxed k. 10. We can thus ﬁnd an integer nk such that µ(Ank k ) < ε2−k . fm (x) − f (x) ≤ 1/k if m is suﬃciently large. Therefore µ(Ank ) → 0 as n → ∞. Proof. and so fn (x) − f (x) ≤ / / 1/k if n ≥ nk .78 ∞ CHAPTER 10. Prove that fn converges in measure. Ank decreases as n increases. so fn does not converge to f at any point. Egorov’s theorem is not as useful for solving problems as one might expect. then x ∈ Ank k .1 Suppose that fn is a sequence that is Cauchy in measure. This means that given ε and a > 0. and fn → f a. Let A = ∪∞ Ank k .2 Exercises Exercise 10. n ≥ N . The following is known as Egorov’s theorem. Then there exists a measurable set A such that µ(A) < ε and fn → f uniformly on Ac .8 Suppose µ is a ﬁnite measure. This type of convergence is sometimes known as almost uniform convergence. but such that ∪n≥m Fn contains S for each m. The intersection ∩n Ank has measure 0 because for almost every x. then µ({x : fn (x) − fm (x) > a}) < ε. there exists N such that if m. . ε > 0. k=1 Hence µ(A) < ε. Thus fn → f uniformly on Ac . each point of S is in inﬁnitely many Fn . If x ∈ A. Let Ank = ∪∞ {x : fm (x) − f (x) > 1/k}. The Fn are arcs whose length tends to 0. Theorem 10. and students have a tendency to try to use it when other methods work much better. TYPES OF CONVERGENCE But because j=1 1/j = ∞.
n=1 Prove that fn → f a. is it necessarily true that fn converges to 0 in measure? If not. not necessarily everywhere.10. then there exists a measurable set A such that f = χA a. µ) is a measure space and fn is a sequence of realvalued measurable functions such that fn g dµ converges to 0 for every integrable g. Prove that fn converges to f a.5 Suppose for each ε there exists a measurable set F such that µ(F c ) < ε and fn converges to f uniformly on F . Exercise 10. 1 + f − g 79 Prove that d is a metric on the space of measurable functions..6 Suppose that fn and f are measurable functions such that for each ε > 0 we have ∞ µ({x : fn (x) − f (x) > ε}) < ∞. then fn converges a. EXERCISES Exercise 10. give a counterexample. . Exercise 10.2 Suppose µ(X) < ∞. A.7 Let fn be a sequence of measurable functions and deﬁne gn (x) = sup fm (x) − fn (x). Exercise 10. Exercise 10.e.4 Prove that if An is measurable and µ(An ) < ∞ for each n and χAn converges to f in measure.e.e.e. Exercise 10.2. Prove that fn → f in measure if and only if d(fn . g) = f − g dµ. g) = 0 only implies that f = g a. m≥n Prove that if gn converges in measure to 0.3 Prove that if fn → f in measure and each fn is nonnegative. except for the fact that d(f.8 If (X. Deﬁne d(f. f ) → 0. then f ≤ lim inf n→∞ fn .e. Exercise 10.
80 CHAPTER 10. TYPES OF CONVERGENCE .
Thus every element of C0 is of the form ∪n (Ai × Bi ). We deﬁne the product σalgebra A × B = σ(C0 ). and if i = j. the Fubini theorem. and many other product spaces. in ndimensional Euclidean spaces. y) ∈ E} 81 . A. we deﬁne the xsection of E by sx (E) = {y ∈ Y : (x. µ) and (Y.Chapter 11 Product measures We have deﬁned Lebesgue measure on the line. it is easy to check that C0 is an algebra of sets. where A ∈ A and B ∈ B. Bi ∈ B. Now we give a method for constructing measures on the plane. Let C0 be the collection of ﬁnite unions of disjoint measurable rectangles. which allows one to interchange the order of integration. A measurable rectangle is a set of the form A × B. B. ν) are two measure spaces and suppose also that µ and ν are σﬁnite measures.1 Product σalgebras Suppose (X. 11. then (Ai × Bi ) ∩ (Aj × Bj ) = ∅. If E ∈ A × B. Since (A × B)c = (A × B c ) ∪ (Ac × Y ) and the intersection of two measurable rectangles is a measurable rectangle. is one of the most important theorems in real analysis. i=1 where Ai ∈ A. The main theorem.
(2) If f is A × B measurable. Lemma 11. note that Sx f (y) = χsx (E) (y). y) ∈ E}. If E = A × B. Similarly. Hence sx (E) ∈ B for each x when E is a measurable / rectangle. If E ∈ C. Sx f is B measurable when f is a simple function. y). y). k(y) = µ(ty (E)). then y ∈ sx (E c ) if and only if (x. and since Sx rn ↑ Sx f . which happens if and only if y ∈ sx (E). and / C is closed under the operation of taking complements. Let E ∈ A × B and let h(x) = ν(sx (E)). Given a function f : X × Y → R that is A × B measurable. take A × B measurable simple functions rn increasing to f . and so C is closed i=1 i=1 under the operation of countable unions. then Sx f is B measurable. which is B measurable. (1) Let C be the collection of sets in A × B for which sx (E) ∈ B for each x. The argument for Ty f is the same. then Sx f is B measurable for each x and Ty f is A measurable for each y. Writing f = f + − f − and using linearity again shows that Sx f is B measurable. The argument for ty (E) is the same. If f is nonnegative.82 CHAPTER 11. We will show that C is a σalgebra containing the measurable rectangles. If f = χE for E ∈ A × B. then sx (E) ∈ B for each x and ty (E) ∈ A for each y. (2) Fix x. then sx (E) is equal to B if x ∈ A and equal to ∅ if x ∈ A. By linearity. y) ∈ E c . it is easy to see that sx (∪∞ Ei ) = ∪∞ sx (Ei ). Therefore C is a σalgebra containing the measurable rectangles. and hence is all of A × B. PRODUCT MEASURES and similarly deﬁne the ysection: ty (E) = {x : (x.1 (1) If E ∈ A × B. for each x and y we deﬁne Sx f : Y → R and Ty f : X → R by Sx f (y) = f (x. Therefore sx (E c ) = (sx (E))c . and hence is equal to A × B. . Ty f (x) = f (x. Proof.
y) ν(dy) µ(dx) = χE (x. which is A measurable. and since the sx (Ei ) are i=1 disjoint. If we let hi (x) = ν(sx (Ei )) and deﬁne ki (y) similarly. We will prove that C contains C0 and is a monotone class. Ty χE (x) µ(dx) ν(dy). k(y) = µ(A)χB (y) is B measurable and k(y) ν(dy) = µ(A)ν(B). First suppose µ and ν are ﬁnite measures.1) Since χsx (E) (y) = Sx χE (y) for all x and y. with A ∈ A and B ∈ B. Therefore C contains C0 .1) could be written as Sx χE (y) ν(dy) µ(dx) = We will usually write this as χE (x. (11. since it is the sum of A measurable functions. then n h(x) = ν(sx (E)) = ν(∪n sx (Ei )) = i=1 i=1 ν(sx (Ei )). (11. Therefore (1) and (2) hold for measurable rectangles. PRODUCT σALGEBRAS 83 Proposition 11.2 (1) h is A measurable and k is B measurable. Similarly k(y) is B measurable. y) µ(dx) ν(dy). Therefore h is . If E = A × B. This shows that h is A measurable. Similarly. then sx (E) = ∪n sx (Ei ). Suppose En ↑ E and each En ∈ C. and then (2) holds for E by linearity. where each Ei is a measurable rectangle and the i=1 Ei are disjoint. (2) We have h(x) µ(dx) = k(y) ν(dy).11.1. then h(x) = χA (x)ν(B). then hi (x) µ(dx) = ki (y) ν(dy) by the preceding paragraph. and h(x) µ(dx) = µ(A)ν(B). Proof. If we let hn (x) = ν(sx (En )) and let kn (y) = µ(tn (En )). Let C be the collection of sets in A × B for which (1) and (2) hold. then hn ↑ h and kn ↑ k. This will prove that C is the smallest σalgebra containing C0 and hence is equal to A × B. If E = ∪n Ei .
Applying the monotone convergence theorem proves that (11. We now deﬁne µ × ν by µ × ν(E) = h(x) µ(dx) = k(y) ν(dy). By the monotone class theorem (Theorem 2. .1) holds.2 that i=1 n ν(sx (E)) = i=1 ν(sx (Ei )). or µ×ν is ﬁnitely additive.1) holds with h and k. hi is A measurable.84 CHAPTER 11. Then there exist Fi ↑ X and Gi ↑ Y such that each Fi is A measurable and has ﬁnite µ measure and each Gi is B measurable and has ﬁnite ν measure. The only diﬀerence is that we use the dominated convergence theorem in place of the monotone convergence theorem. and by monotone convergence. . ki is B measurable. Let µi (A) = µ(A ∩ Fi ) for each A ∈ A and νi (A) = ν(A ∩ Gi ) for each B ∈ B. . resp. We conclude that n µ × ν(E) = n ν(sx (E)) µ(dx) = i=1 ν(sx (Ei )) µ(dx) = i=1 µ × ν(Ei ). Therefore µ × ν is a measure. which is A × B. and that (11. then we saw in the proof of Proposition 11. then hn ↑ h.1) holding when h and k are replaced by hn and kn .2) Clearly µ × ν(∅) = 0. Let hi (x) = νi (sx (E)) = ν(sx (E) ∩ Gi ) and similarly deﬁne ki (y). which proves the measurability of h and k. En are disjoint and in A × B and E = ∪n Ei . Now hi ↑ h and ki ↑ k. This is where we need µ and ν to be ﬁnite measures. µ × ν(En ) ↑ µ × ν(E). PRODUCT MEASURES A measurable and k is B measurable. We have shown C is a monotone class containing C0 .1) holds with h and k. If E1 . We let n → ∞ and use the monotone convergence theorem to see that (11.1) holds if we replace h and k by hi and ki . By what we have proved above. resp. almost the same argument shows that h and k are measurable with respect to A and B. We have (11. . If En ↑ E with each En ∈ A×B and we let hn (x) = ν(sx (En )). . and (11. (11. C is equal to σ(C0 ). If En ↓ E with each En ∈ C. Finally suppose µ and ν are σﬁnite.10).
y) ν(dy) is measurable with respect to A. If f is the characteristic function of a set in A × B.2.11. y) dµ(x) dν(y) f (x. then (1)–(5) are merely a restatement of Lemma 11.3) should be interpreted as Sx f (y) ν(dy) µ(dx) and similarly for the second integral in (11. y) dν(y) µ(dx). y) µ(dx) is measurable with respect to B. This is sometimes called the FubiniTonelli theorem.3). then h(x) = χA (x)ν(B) and so µ × ν(A × B) = µ(A)ν(B). y) < ∞. or (b) f (x. most often the brackets are omitted in (11. Since no confusion results. y) is measurable with respect to B. the function y → f (x. If either (a) f is nonnegative. THE FUBINI THEOREM 85 Note that if E = A × B is a measurable rectangle. (2) for each y. which is what it should be. y) = = f (x. 11. (3) the function g(x) = f (x. Proof. y) d(µ × ν)(x. (4) the function h(y) = f (x.2 The Fubini theorem The main result of this chapter is the Fubini theorem.1 and Proposition . (11.3) The last integral in (11.3 Suppose f : X ×Y → R is measurable with respect to A × B. y) d(µ × ν)(x. Theorem 11. y) is measurable with respect to A. then (1) for each x.3). (5) we have f (x. which allows one to interchange the order of integration. the function x → f (x.
y) µ(dx) ν(dy) < ∞ We can then apply the Fubini theorem again to conclude f (x. y) d(µ × ν) = f (x. Since the increasing limit of measurable functions is measurable. E ⊂ [0. we see that (1)–(5) hold when f is nonnegative. y) dν dµ < ∞. ν) are complete. There is no diﬃculty extending the Fubini theorem to the product of n measures. m × · · · × m) is called ndimensional Lebesgue measure. let (X. A × B. B. y) dµ dν < ∞ or f (x. (1)–(5) hold if f is a simple function. it will not be the case in general that (X × Y. too. . Even when (X. we could as well assume f (x. A. 1]×{1/2}. B. 1] and let E = A × {1/2}. PRODUCT MEASURES 11. µn all equal to m. Thus in the hypotheses of the Fubini theorem.86 CHAPTER 11.1. . then the completion of (Rn . . µ × ν) without great diﬃculty. Let A be a nonmeasurable set in [0. Lebesgue measure on R with the Lebesgue σalgebra L. y) dµ dν = f (x. L × · · · × L. µ) = (Y. we sometimes say that f is jointly measurable. which has zero measure with respect to µ × ν. When f is measurable with respect to A × B. then writing a nonnegative function as the increasing limit of simple functions and using the monotone convergence theorem. µ × ν) is complete. y) dν dµ. ν) be Lebesgue measure on [0. A. Then E is not a measurable set with respect to A×B.2. so E is a null set. 1] with the Lebesgue σalgebra. . or else A = t1/2 (E) would be in A by Lemma 11. If we have µ1 . One can take the completion of (X × Y. Observe that if we know f (x. In the case where f  d(µ × ν) < ∞. see [1]. . writing f = f + − f − and using linearity proves (1)–(5) for this case. For a general change of variables theorem. A × B. y) d(µ × ν) = f (x. y) µ(dx) ν(dy) < ∞. y) is nonnegative the Fubini theorem tells us that f (x. µ) and (Y. See [4] for details. On the other hand. then since f (x. For example. By linearity.
y) dx dy = ∞. y) dx dy = 0. Then f (x. There exists a set X together with a partial order “≤” such that X is uncountable but for any y ∈ X. EXAMPLES 87 11. y) at most two terms in the sum are nonzero. integrating ﬁrst with respect to x gives 0. 1/i) such that 0 gi (x) dx = 1. y) dy dx = 1 but f (x. The reason there is no contradiction is that f is not measurable with respect to the product σalgebra. This doesn’t contradict the Fubini theorem because 1 0 0 1 f (x. If we ﬁrst integrate with respect to y. For each point (x. . . . Example 11. Example 11. y) dx dy = 0. On the other hand. y) dy dx = 0 g1 (x) dx = 1. the set {x ∈ X : x ≤ y} is countable. Deﬁne µ on X by µ(A) = 0 if A is countable and 1 if A is uncountable. Let ∞ f (x. so 1 0 0 1 f (x. Deﬁne f on X × X by f (x. so the sum is actually a ﬁnite one. An example is to let X be the set of countable ordinals. Therefore 1 0 0 1 1 f (x. This is a telescoping series.11. The σalgebra is the collection of subsets A of X such that either A or Ac is countable. .. we get 1 ∞ f (x. y) dy = 0 i=1 [gi (x) − gi+1 (x)].3. 1] and µ and ν both being Lebesgue measure. 2. y) = 1 if x ≤ y and zero otherwise. you have to take on faith a bit of set theory. Let gi be continuous functions with support 1 in (1/(i + 1). i = 1. y) = i=1 [gi (x) − gi+1 (x)]gi (y).4 Let X = Y = [0.3 Examples We give two examples to show that the hypotheses of the Fubini theorem are necessary.5 For this example. and sums to g1 (x).
where m is Lebesgue measure. Here L is the Lebesgue σalgebra on [0.e.1 State and prove a version of the Fubini theorem for complexvalued functions. Exercise 11. 1]. where m2 is twodimensional Lebesgue measure.4 Exercises Exercise 11. Exercise 11. Let g(x) = supy∈Y f (x. ﬁnd a counterexample. 1] (with respect to one dimensional Lebesgue measure) the set sx (A) has onedimensional Lebesgue measure one. 1]. 1] and for every y ∈ [0.6 Suppose f is realvalued and integrable with respect to twodimensional Lebesgue measure on [0. A) and (Y. y) is Lebesgue measurable on [0. Exercise 11.7 Prove that 1 0 0 1 x2 − y 2 log(4 + sin x) dy dx (x2 + y 2 )2 1 1 0 = 0 x2 − y 2 log(4 + sin x) dx dy. Exercise 11. 1] and b ∈ [0.4 Let A be a Lebesgue measurable subset of [0. (x2 + y 2 )2 . Prove that f = 0 a.2 Let (X. 1]2 with m2 (A) = 1. 1]2 . 1]. 1] the function x → f (x. y) dy dx = 0 for all a ∈ [0.3 Prove the equality ∞ ∞ f (x) dx = −∞ 0 m({x : f (x) ≥ t}) dt. y) and suppose g(x) < ∞ for each x. Prove that f is measurable with respect to the the completion of the product σalgebra L × L on [0. B) be two measurable spaces and let f ≥ 0 be measurable with respect to A × B. 1] the function y → f (x.5 Let f : [0.88 CHAPTER 11. 1]2 → R be such that for every x ∈ [0. Is g necessarily measurable with respect to A? If not. y) is continuous on [0. Exercise 11. 1]2 and a 0 0 b f (x. PRODUCT MEASURES 11. Show that for almost every x ∈ [0. Exercise 11.
1] and let B be the Borel σalgebra. y) µ(dy) m(dx) = X Y Y X χD (x. Show that f (x. Exercise 11. Show [α(x + c) − α(x)] dx = cµ(R).8 Let X = Y = [0. y) = −1. 1]. otherwise. Exercise 11. x]). y) : x = y}. y) dy dx = f (x. (1) If D = {(x.12 Use 1 = x ∞ e−xy dy 0 . Deﬁne 1. EXERCISES 89 Exercise 11. b ∈ [0. Why does this not contradict the Fubini theorem? Exercise 11.4.9 Let X = Y = R and let B be the Borel σalgebra. y) dy dx = 0 for every a. x ≥ 0 and x + 1 ≤ y < x + 2. show that D is measurable with respect to B × B. 0 0 f (x. y) m(dx) µ(dy). but f is not zero almost everywhere with respect to 2dimensional Lebesgue measure. (2) Show that χD (x.11 Let µ be a ﬁnite measure on R and let α(x) = µ((−∞. x ≥ 0 and x ≤ y < x + 1. Let m be Lebesgue measure and µ counting measure on [0.10 Find a realvalued function f that is integrable on [0.11. f (x. 1]. y) dy dx = 0. 1]2 such that a 0 0 1 1 b f (x. y) dx dy. Why does this not contradict the Fubini theorem? Exercise 11. 0.
13 Let X = {1.15 Let (X.14 Let {an } and {rn } be two sequences of real num∞ bers such that n=1 an  < ∞. x 2 Recall that eau sin u du = eau (a sin u − cos u) + C. y) µ(dy) µ(dx). y) = −1.} and let µ be counting measure on X. y) µ(dx) µ(dy) = X X X X f (x. ν) be measure spaces. A. otherwise. Exercise 11. x = y + 1. 0. . 1 + a2 Exercise 11. B. µ) and (Y.90 CHAPTER 11. Prove that if λ is a measure on A × B such that λ(A × B) = µ(A)ν(B) whenever A ∈ A and B ∈ B. . then λ = µ × ν on A × B. . . Why is this not a contradiction to the Fubini theorem? Exercise 11. f (x. Prove that ∞ an x − rn  n=1 converges absolutely for almost every x ∈ R. Then prove that b b→∞ lim 0 sin x π dx = . x = y. PRODUCT MEASURES and the Fubini theorem to calculate b 0 0 ∞ e−xy sin x dy dx two diﬀerent ways. Deﬁne f : X × X → R by 1. 2. Show that f (x.
19 If M = (Mij )n i. xn ) ∈ R . (Multiplication by M multiplies the ﬁrst coordinate by c.11. i = j = 1.R) 0 0 f (r cos θ. 2π).4. Exercise 11. . 0.) If A is a Borel subset of Rn . y) dy dx = B(0.16 Let S be the unit circle {eiθ : 0 ≤ θ < 2π} and deﬁne a measure µ on S by µ(A) = m({θ : eiθ ∈ A}). Show that if A is a Borel subset of S and R > 0.18 Prove that ∞ e−x 0 2 /2 dx = π/2 by ﬁlling in the missing steps and making rigorous the following.j=1 is a n × n matrix and x = n (x1 . Mij = 1. where we use mn for ndimensional Lebesgue measure. Exercise 11.17 Use Exercise 11. eiθ ∈ A}) = µ(A)R2 /2. i = j. (This is just the usual matrix multiplication of a n × n matrix and a n × 1 matrix. i = j and neither equals k. R) = {(x. Let m2 be twodimensional Brownian motion. . (1) If c ∈ R and c. Exercise 11. 0. deﬁne M x to be the element of Rn whose ith n coordinate is j=1 Mij xj .16 to prove that if f is a continuous realvalued function with support in the ball B(0. y) : x2 + y 2 < R2 }. otherwise. r sin θ) r dr dθ. then ∞ ∞ I2 = 0 0 e−(x 2 +y 2 )/2 π/2 ∞ dy dx = 0 0 e−r 2 /2 r dr dθ = π/2. . 2 ∞ If I = 0 e−x /2 dx. j = 1 and i = k. then m2 ({reiθ : 0 < r < R. . where m is Lebesgue measure on [0. . 1. Mij = 1. show mn (M (A)) = cmn (A) =  det M mn (A). let M (A) = {M x : x ∈ A}. then 2π R f (x. i = 1 and j = k. EXERCISES 91 Exercise 11.) (2) If 1 ≤ k ≤ n and 1. i = j = 1.
(Multiplication by M replaces x1 by x1 + cx2 . i = 1. (Multiplication by an orthogonal matrix is a rotation of Rn . 0. j = 2. so that M times its transpose is the identity. conclude that if M is any n × n matrix. show mn (M (A)) = mn (A) =  det M mn (A). Mij = c. (Multiplication by M interchanges the ith and k th coordinates. or (3). then mn (M (A)) =  det M mn (A). otherwise.) . show mn (M (A)) = mn (A).92 show CHAPTER 11.) (3) If c ∈ R and 1. i = j. (5) If M is an orthogonal matrix. (2).) (4) Since every n × n matrix can be written as the product of matrices each of which has the form given in (1). PRODUCT MEASURES mn (M (A)) = mn (A) =  det M mn (A).
Deﬁnition 12. 12. A signed measure is a function µ : A → (−∞.Chapter 12 Signed measures Signed measures have the countable additivity property of measures.2 Let µ be a signed measure. When we want to emphasize that a measure is deﬁned as in Deﬁnition 3. ∞] such that µ(∅) = 0 and µ(∪∞ Ai ) = i=1 ∞ i=1 µ(Ai ) whenever the Ai are pairwise disjoint and all the Ai are in A. 93 .1 Let A be a σalgebra. We say A ∈ A is a negative set if µ(B) ≤ 0 whenever B ⊂ A and B ∈ A. A set A ∈ A is called a positive set for µ if µ(B) ≥ 0 whenever B ⊂ A and B ∈ A. where f is integrable and takes both positive and negative values. We will see shortly that an example of a signed measure is ν(A) = A f dµ.1 and only takes nonnegative values. we refer to it as a positive measure. A null set A is one where µ(B) = 0 whenever B ⊂ A and B ∈ A.1 Positive and negative sets Deﬁnition 12. but are allowed to take negative as well as positive values.
. Let E be measurable with µ(E) < 0. If we let P = {x : f (x) ≥ 0}. however.3 Suppose m is Lebesgue measure and µ(A) = A f dm for some integrable f . . If Fk = E − (E1 ∪ · · · ∪ Ek−1 ) is a negative set. . that C is a null set. ∞]. Let n1 be the smallest positive integer such that there exists E1 ⊂ E with µ(E1 ) ≥ 1/n1 . there exists a measurable subset with positive measure. Example 12. where µ+ (A) = A f + dm and µ− (A) = A f − dm.94 CHAPTER 12. Proof. We then deﬁne pairwise disjoint measurable sets E2 . Let k ≥ 2 and suppose E1 . If Fk is not a negative set. by induction as follows. then k−1 µ(Fk ) = µ(E) − i=1 µ(Ei ) ≤ µ(E) < 0 and Fk is the desired set F . then P is easily seen to be a positive set. The Hahn decomposition which we give below is a decomposition of our space (in this case R) into the positive and negative sets P and N . . . then N is a negative set. then µ(∪∞ Ai ) = lim µ(∪n Ai ). we are done. i=1 i=1 n→∞ The proof is the same as in the case of positive measures. k − 1. Proposition 12. SIGNED MEASURES Note that if µ is a signed measure. let nk be the smallest positive integer such that there exists Ek ⊂ Fk with Ek measurable and µ(Ek ) ≥ 1/nk .4 Let µ be a signed measure which takes values in (−∞. If not. . . . E3 . except that C = {x : f (x) = 0} could be included in N instead of P . The Jordan decomposition below is a decomposition of µ into µ+ and µ− . If E is a negative set. Ek−1 are pairwise disjoint measurable sets with µ(Ei ) > 0 for i = 1. and if N = {x : f (x) < 0}. or apportioned partially to P and partially to N . This decomposition is unique. . . . . Then there exists a measurable subset F of E that is a negative set with µ(F ) < 0. Note.
5 (1) Let µ be a signed measure taking values in (−∞. Then µ(G) ≥ 1/N for some N . (1) Let L = inf{µ(A) : A is a negative set}. . The following is known as the Hahn decomposition. Let n=1 Bn = An − (B1 ∪ · · · ∪ Bn−1 ) for each n. so is each Bn . Proof. Then µ(F ) ≤ µ(E) < 0. nk > N . since for some k. then µ(F ) > 0. But this contradicts the construction. Since An is a negative set. so the sum converges.12. Choose negative sets An such that µ(An ) → L. (2) If E and F are another such pair.2. Suppose G ⊂ F is measurable with µ(G) > 0. Therefore F must be a negative set. Theorem 12. 12. Let E = ∪∞ An . If −µ is not a positive measure. ∞]. and we would have chosen the set G instead of the set Ek at stage k. then E∆E = F ∆F is a null set with respect to µ. Since 0 > µ(E) > −∞ and µ(Ek ) ≥ 0. we continue and let F = ∩k Fk = E − (∪k Ek ). HAHN DECOMPOSITION THEOREM 95 We stop the construction if there exists k such that Fk is a negative set with µ(Fk ) < 0. If not. Also. It remains to show that F is a negative set.2 Hahn decomposition theorem Recall that we write A∆B for (A − B) ∪ (B − A). then µ(E) < 0. (3) If µ is not a positive measure. then n µ(C) = lim µ(C ∩ (∪n Bi )) = lim i=1 n→∞ n→∞ µ(C ∩ Bi ) ≤ 0. then ∞ µ(E) = µ(F ) + k=1 µ(Ek ). There exist disjoint measurable sets E and F in A whose union is X and such that E is a negative set and F is a positive set. i=1 Thus E is a negative set. If C ⊂ E. the Bn are disjoint and ∪n Bn = ∪n An = E.
But f is constant on every such interval. where C is the Cantor set. But then E ∪ C would be a negative set with µ(E ∪ C) < µ(E) = L. A similar argument applies for −µ and F . (3) Suppose µ is not a positive measure but µ(E) = 0. then µ and ν are mutually singular. Let f be the CantorLebesgue function where we deﬁne f (x) = 1 if x ≥ 1 and f (x) = 0 if x ≤ 0 and let ν be the LebesgueStieltjes measure associated with f . Let us say two measures µ and ν are mutually singular if there exist two disjoint sets E and F in A whose union is X with µ(E) = ν(F ) = 0. This example works because the Lebesgue measure of {1/2} is 0. we need to show ν(I) = 0 for every open interval contained in F . But A ⊂ E − E = F − F ⊂ F . if E . SIGNED MEASURES Since E is a negative set. µ(E) = µ(An ) + µ(E − An ) ≤ µ(An ). Let F = E c . Then µ ⊥ ν. a contradiction. then µ(A) ≤ 0. and F = C c . By Proposition 12. we let E = C. Therefore µ(A) = 0. b] contained in F . 1/2] and F = (1/2. and therefore ν(J) = f (b) − f (a) = 0. We already know that m(E) = 0 and we need to show ν(F ) = 0. so µ(A) ≥ 0. where A1 ⊂ E − E and A2 ⊂ E − E. Letting n → ∞. and any subset of E∆E can be written as the union of A1 and A2 . 1/2]). To do that. If F were not a positive set. The same argument works if A ⊂ E − E. If A ∈ A. (2) To prove uniqueness. Example 12.6 If µ is Lebesgue measure restricted to [0. We let E = [0. µ(A) = m(A ∩ [0. we obtain µ(E) = L. 1]. that is. 1]. there would exist B ⊂ F with µ(B) < 0. a contradiction. To see this. Example 12. 1/2]. which says that µ must be a positive measure. then µ(A) = µ(A ∩ E) + µ(A ∩ F ) ≥ µ(E) + µ(A ∩ F ) ≥ 0. This will follow if we show ν(J) = 0 for every interval of the form J = (a. Let µ be Lebesgue measure.7 A more interesting example is the following. so f (b) = f (a). and ν is Lebesgue measure restricted to [1/2. . F are another such pair of sets and A ⊂ E − E ⊂ E.4 there exists a negative set C contained in B with µ(C) < 0.96 CHAPTER 12. This is often written µ ⊥ ν.
there exist positive measures µ+ and µ− such that µ = µ+ −µ− and µ+ and µ− are mutually singular.1 Suppose µ is a signed measure. and consequently F is a positive set for µ. If µ = ν + − ν − is another such decomposition with ν + . Let µ+ (A) = µ(A ∩ F ). if A ∈ A. let E be a set such that ν + (E ) = 0 and ν − ((E )c ) = 0. If A ⊂ F . By the uniqueness part of the Hahn decomposition theorem.3. Set F = (E )c . for µ so that X = E ∪ F and E ∩ F = ∅. then ν + (A) = ν + (A ∩ F ) = ν + (A ∩ F ) − ν − (A ∩ F ) = µ(A ∩ F ) = µ(A ∩ F ) = µ+ (A). Prove that A is a null set with respect to µ if and only if µ(A) = 0. Similarly. JORDAN DECOMPOSITION THEOREM 97 12. (12. Hence X = E ∪F and E ∩F = ∅. Proof. and so µ(A) = ν + (A) − ν − (A) = ν + (A) ≥ 0. . E is a negative set for µ. This decomposition is unique. This gives the desired decomposition. µ− (A) = −µ(A ∩ E).4 Exercises Exercise 12. F ∆F is a null set with respect to µ.. A). then ν − (A) ≤ ν − (F ) = 0.3 Jordan decomposition theorem The following is known as the Jordan decomposition theorem. Let E and F be negative and positive sets.12. Thus E . resp. The measure µ = µ+ + µ− is called the total variation measure of µ.8 If µ is a signed measure on a measurable space (X. F gives another Hahn decomposition of X. Theorem 12. and similarly ν − = µ− . Since ν + (E ) = 0 and ν − (F ) = 0.1) 12. ν − mutually singular.
98 CHAPTER 12. Suppose λ = µ − ν. Prove that if A ∈ A. then µ + ν(A) ≤ µ(A) + ν(A) for every A ∈ A. ∪n Bj = A .3 Let µ be a signed measure on (X. Exercise 12. the Bj are disjoint. Prove that if A ∈ A. where µ and ν are ﬁnite positive measures. j=1 . Exercise 12. then n µ(A) = sup j=1 µ(Bj ) : each Bj ∈ A. A) be a measurable space. B ⊂ A} and µ− (A) = − inf{µ(B) : B ∈ A.6 Let (X.7 Suppose that µ is a signed measure on (X. A). Prove that µ(A) ≥ λ+ (A) and ν(A) ≥ λ− (A) for every A ∈ A.2 Let µ be a signed measure. Exercise 12.4 Let µ be a positive measure and ν a signed measure. SIGNED MEASURES Exercise 12. Exercise 12. A). A) be a measurable space. Exercise 12. Exercise 12. Deﬁne f dµ = Prove that f dµ ≤ f  dµ. then µ+ (A) = sup{µ(B) : B ∈ A. B ⊂ A}. Prove that µ(A) = sup A f dµ : f ≤ 1 . f dµ+ − f dµ− . A). Prove that ν µ if and only if ν + µ and ν − µ.8 Suppose that µ is a signed measure on (X. Prove that if µ and ν are ﬁnite signed measures.5 Let (X.
we have ∞ ∞ ν(∪n An ) = ∪n An f dµ = n=1 An f dµ = n=1 ν(An ) by using Proposition 7. (13.1 Absolute continuity Deﬁnition 13. The only part that needs thought is the countable additivity. In this chapter we consider the converse. We write ν µ.5. 99 .1 A measure ν is said to be absolutely continuous with respect to a measure µ if ν(A) = 0 whenever µ(A) = 0. when does there exist f such that (13.1) holds? The RadonNikodym theorem answers this question.Chapter 13 The RadonNikodym theorem Suppose f is nonnegative and integrable with respect to µ. ν(A) is zero whenever µ(A) is. 13. If An are disjoint measurable sets. Moreover. If we deﬁne ν by ν(A) = A f dµ. If we are given two measures µ and ν.1) then ν is a measure.
Consider the Hahn decomposition for ν − n µ.100 CHAPTER 13. Suppose for each ε. Let F = ∩∞ ∪∞ Ek . Note E c = ∪n En = ∪n Fn = F . Then ν is absolutely continuous with respect to µ if and only if for all ε there exists δ such that µ(A) < δ implies ν(A) < ε. and their union is X. Proof. hence ν(A) = 0. 13. k=n n→∞ This contradicts the absolute continuity. E ⊂ En . .2 Let ν be a ﬁnite measure. THE RADONNIKODYM THEOREM Proposition 13. If there exists an ε for which no corresponding δ exists. A). Then from the deﬁnition of Fn . Either µ ⊥ ν or else there exists ε > 0 and G ∈ A such that µ(G) > 0 and G is a positive set for ν − εµ. One possibility is that µ(E c ) = 0. then there exists Ek such that µ(Ek ) < 2−k but ν(Ek ) ≥ ε. k=n but ν(F ) = lim ν(∪∞ Ek ) ≥ ε. 1 Proof. In this case. Since ν is a positive measure. there exists δ such that µ(A) < δ implies ν(A) < ε. G is a positive set for ν − εµ.2 The main theorem Lemma 13. this implies ν(E) = 0. Then n=1 k=n ∞ µ(F ) = lim µ(∪∞ Ek ) ≤ lim k=n n→∞ n→∞ 2−k = 0. µ(Fn ) > 0 for some n. The other possibility is that µ(E c ) > 0. Thus there exists a negative set En and a positive set Fn for this measure. Let ε = 1/n and G = Fn . If µ(A) = 0. then ν(A) < ε for all ε. Let F = ∪n Fn and c E = ∩n En . so ν(E) ≤ ν(En ) ≤ 1 n µ(En ) ≤ 1 n µ(X). For each n. En and Fn are disjoint. and thus ν µ. in which case µ ⊥ ν. Suppose now that ν µ.3 Let µ and ν be ﬁnite positive measures on a measurable space (X.
gn ). The idea of the proof is to look at the set of f such that f dµ ≤ ν(A) for each A ∈ A. 101 Theorem 13.13.2. THE MAIN THEOREM We now are ready for the RadonNikodym theorem. Let L = sup{ g dµ : g ∈ F}. Let us assume µ is a ﬁnite measure for now. . A) such that ν is absolutely continuous with respect to µ. A g dµ ≤ ν(A) for all A ∈ A . We claim that if g1 and g2 are in F. with respect to µ. In this step we deﬁne the function f .4 Suppose µ is a σﬁnite positive measure on a measurable space (X. then f = g almost everywhere with respect to µ. Proof.e. and let gn be a sequence in F such that gn dµ → L. and then to choose the one such A that X f dµ is largest.1 we have f − g = 0 a. Deﬁne F = g measurable : g ≥ 0. F is not empty because 0 ∈ F. Let us ﬁrst prove the uniqueness assertion. and is written f = dν/dµ. For every set A we have (f − g) dµ = ν(A) − ν(A) = 0. The function f is called the RadonNikodym derivative of ν with respect to µ or sometimes the density of ν with respect to µ. A By Proposition 8. Step 1. . if g is another such function. Sometimes one writes dν = f dµ. Then there exists a µintegrable nonnegative function f which is measurable with respect to A such that ν(A) = A f dµ for all A ∈ A. Step 2. A) and ν is a ﬁnite positive measure on (X. . g2 ) is . Let hn = max(g1 . . then h2 = max(g1 . Moreover.
λ is a positive measure since f ∈ F.102 CHAPTER 13. Therefore h2 ∈ F. say to f .2) for all A. Deﬁne a measure λ by λ(A) = ν(A) − A f dµ. Since ν µ. Hence f + εχG ∈ F. Therefore λ ⊥ µ. Then there must exist H ∈ A such that µ(H) = 0 and λ(H c ) = 0. there exists ε > 0 and G such that G is measurable. and write h2 dµ = A A∩B h2 dµ + A∩B c h2 dµ g2 dµ A∩B c = A∩B g1 dµ + ≤ ν(A ∩ B) + ν(A ∩ B c ) = ν(A). To see this. The hn increase. ν(A) − A f dµ = λ(A) ≥ λ(A ∩ G) ≥ εµ(A ∩ G) = A εχG dµ. For any A ∈ A. let B = {x : g1 (x) ≥ g2 (x)}. and G is a positive set for λ − εµ. By Lemma 13. X a contradiction to the deﬁnition of L. By an induction argument. then ν(H) = 0. But (f + εχG ) dµ = L + εµ(G) > L. µ(G) > 0. . THE RADONNIKODYM THEOREM also in F. hn is in F.3. or ν(A) ≥ A (f + εχG ) dµ. Next we prove that f is the desired function. Suppose λ is not mutually singular to µ. By monotone convergence. Step 3. L and f dµ ≤ ν(A) A f dµ = (13. and hence λ(H) = ν(H) − H f dµ = 0.
13.3. LEBESGUE DECOMPOSITION THEOREM This implies λ = 0, or ν(A) =
A
103
f dµ for all A.
Step 4. We now suppose µ is σﬁnite. There exist Fi ↑ X such that µ(Fi ) < ∞ for each i. Let µi be the restriction of µ to Fi , that is, µi (A) = µ(A ∩ Fi ). Deﬁne νi , the restriction of ν to Fi , similarly. If µi (A) = 0, then µ(A ∩ Fi ) = 0, hence ν(A ∩ Fi ) = 0, and thus νi (A) = 0. Therefore νi µi . If fi is the function such that dνi = fi dµi , the argument of Step 1 shows that fi = fj on Fi if i ≤ j. Deﬁne f by f (x) = fi (x) if x ∈ Fi . Then for each A ∈ A, ν(A ∩ Fi ) = νi (A) =
A
fi dµi =
A∩Fi
f dµ.
Letting i → ∞ shows that f is the desired function.
13.3
Lebesgue decomposition theorem
The proof of the Lebesgue decomposition theorem is almost the same. Theorem 13.5 Suppose µ and ν are two ﬁnite positive measures. There exist positive measures λ, ρ such that ν = λ+ρ, ρ is absolutely continuous with respect to µ, and λ and µ are mutually singular.
Proof. Deﬁne F and L and construct f as in the proof of the RadonNikodym theorem. Let ρ(A) = A f dµ and let λ = ν − ρ. Our construction shows that f dµ ≤ ν(A),
A
so λ(A) ≥ 0 for all A. We have ρ + λ = ν. We need to show µ and λ are mutually singular. If not, by Lemma 13.3, there exists ε > 0 and F ∈ A such that µ(F ) > 0 and F is a positive set for λ − εµ. We get a contradiction exactly as in the proof of the RadonNikodym theorem. We conclude that λ ⊥ µ.
104
CHAPTER 13. THE RADONNIKODYM THEOREM
13.4
Exercises
Exercise 13.1 This exercise asks you to prove the RadonNikodym theorem for signed measures. Let (X, A) be a measurable space. Suppose ν is a signed measure, µ is a ﬁnite positive measure, and ν(A) = 0 whenever µ(A) = 0 and A ∈ A. Show there exists an integrable realvalued function f such that ν(A) = A f dµ for all A ∈ A. Exercise 13.2 We deﬁne a complex measure µ on a measurable space (X, A) to be a map from A to C such that µ(∅) = 0 and ∞ µ(∪∞ Ai ) = i=1 µ(Ai ) whenever the Ai are in A and are pairi=1 wise disjoint. Formulate and prove a RadonNikodym theorem for complex measures. Exercise 13.3 Let (X, A) be a measurable space and let µ and ν be two ﬁnite measures. We say µ and ν are equivalent measures if µ ν and ν µ. Show that µ and ν are equivalent if and only if there exists a measurable function f that is strictly positive a.e. with respect to µ such that dν = f dµ. Exercise 13.4 Suppose µ and ν are two ﬁnite measures such that ν is absolutely continuous with respect to µ. Let ρ = µ + ν. Note that µ(A) ≤ ρ(A) and ν(A) ≤ ρ(A) for each measurable A. In particular, µ ρ and ν ρ. Prove that if f = dµ/dρ and g = dν/dρ, then g is strictly positive for almost every x with respect to µ, f + g = 1, and dν = (f /g) dµ. Exercise 13.5 If µ is a signed measure on (X, A) and µ is the total variation measure, prove that there exists a realvalued function f that is measurable with respect to A such that f  = 1 a.e. with respect to µ and dµ = f dµ. Exercise 13.6 Suppose ν µ and ρ ν. Prove that ρ µ and
dρ dρ dν = · . dµ dν dµ Exercise 13.7 Suppose λn is a sequence of positive measures on a measurable space (X, A) with supn λn (X) < ∞ and µ is another ﬁnite positive measure on (X, A). Suppose λn = fn dµ + νn is
13.4. EXERCISES
105
the Lebesgue decomposition of λn ; in particular, νn ⊥ µ. If λ = ∞ n=1 λn , show that
∞ ∞
λ=
n=1
fn dµ +
n=1
νn
is the Lebesgue decomposition of λ. Exercise 13.8 Let (X, F, µ) be a measure space, and suppose E is a subσalgebra of F, that is, E is itself a σalgebra and E ⊂ F. Suppose f is a nonnegative integrable function that is measurable with respect to F. Deﬁne ν(A) = A f dµ for A ∈ E and let µ be the restriction of µ to E. (1) Prove that ν µ. (2) Since ν and µ are measures on E, then g = dν/dµ is measurable with respect to E. Prove that g dµ =
A A
f dµ
(13.3)
whenever A ∈ E. g is called the conditional expectation of f with respect to E and we write g = E [f  E]. If f is integrable and realvalued but not necessarily nonnegative, we deﬁne E [f  E] = E [f +  E] − E [f −  E]. (3) Show that f = g if and only if f is measurable with respect to E. (4) Prove that if h is E measurable and A h dµ = A f dµ for all A ∈ E, then h = g a.e. with respect to µ. Exercise 13.9 Suppose (X, A, µ) is a measure space and f is integrable and measurable with respect to A. Suppose in addition that B1 , B2 , . . . , Bn is a ﬁnite sequence of disjoint elements of A whose union is X and that each Bj has positive µ measure. Let C = σ(B1 , . . . , Bn ). Prove that
n
E [f  C] =
j=1
Bj
f dµ χBj .
µ(Bj )
Exercise 13.10 Suppose that (X, F, µ) is a measure space, E is a subσalgebra of F, and D is a subσalgebra of E. Suppose f is
106
CHAPTER 13. THE RADONNIKODYM THEOREM
integrable, realvalued, and measurable with respect to F. Prove that E [ E [f  E]  D] = E [f  D] and E [ E [f  D]  E] = E [f  D]. Exercise 13.11 Suppose that (X, F, µ) is a measure space and E is a subσalgebra of F. Suppose that f and f g are integrable realvalued functions, where f is measurable with respect to F and g is measurable with respect to E. Prove that E [f g  E] = gE [f  E].
A function f is diﬀerentiable at x if f (x + h) − f (x) h h→0 lim exists. our results are the following. if f is integrable (Theorem 14. If f : [a. b] → R.Chapter 14 Diﬀerentiation In this chapter we want to look at when a function from R to R is diﬀerentiable and when the fundamental theorem of calculus holds.8).14). are diﬀerentiable (Theorem 14. and 14.7 we give an alternative approach that avoids the use of the RadonNikodym theorem and Lebesgue decomposition theorem. b (3) a f (y) dy = f (b)−f (a) if f is absolutely continuous (Theorem 14.6. x (1) The derivative of a f (y) dy is equal to f a.e.5. The deﬁnition of derivative is the same as in elementary calculus. we say f is diﬀerentiable on [a. Our approach uses what are known as maximal functions and uses the RadonNikodym theorem and the Lebesgue decomposition theorem. However. and the limit is called the derivative of f at x and is denoted f (x). (2) Functions of bounded variation. Brieﬂy. in particular monotone functions. 14.5). b] if the 107 . In Sections 14. some students and instructors prefer a more elementary proof of the results on diﬀerentiation.
α Once B1 . Let Bk be a ball with the same center as Bk but 5 times the ∗ radius. . Bk such that d(Bk+1 ) ≥ 1 2 sup{d(Bα ) : Bα is disjoint from B1 . ∗ m(E) ≤ m(∪k Bk ) ≤ k ∗ m(Bk ) = 5n k m(Bk ). choose Bk+1 disjoint from B1 . . . r) be the open ball with center x and radius r. The procedure might terminate after a ﬁnite number of steps or it might not. Proposition 14. Once we have this. Choose B1 such that d(B1 ) ≥ 1 2 sup d(Bα ). There exists a disjoint sequence B1 . If k m(Bk ) = ∞. DIFFERENTIATION derivative exists for each x ∈ (a. . B2 . . Proof. . The following is an example of what is known as a covering lemma. .108 CHAPTER 14. . We use m for Lebesgue measure on Rn throughout this section. We claim E ⊂ ∪k Bk . 14. of elements of {Bα } such that m(E) ≤ 5n k m(Bk ). Let d(Bα ) be the diameter of Bα . Bk are chosen. Bk }.1 Maximal functions In this section we consider realvalued functions on Rn . Let B(x. Suppose k m(Bk ) < ∗ ∞. . . .1 Suppose E ⊂ Rn is covered by a collection of balls {Bα } of bounded diameter. b) and both lim f (a + h) − f (a) h and lim f (b + h) − f (b) h h→0+ h→0− exist. . . . we have our result. . .
14. B(x. Fix α. . then d(Bk ) → 0. let xj0 be the center of Bj0 and y a point in Bα ∩ Bj0 . it suﬃces to show each Bα ⊂ ∪k Bk . We now prove a weak 11 inequality. Bα must intersect one of 2 B1 .r) Note that without the supremum. and the proof is complete. It is so named because M does not map integrable functions into integrable functions. r). . Therefore Bα intersects Bj0 for some j0 ≤ k. Theorem 14. MAXIMAL FUNCTIONS 109 ∗ ∗ To show E ⊂ ∪k Bk . we are done. deﬁne M f (x) = sup r>0 1 m(B(x. Fix β and let Eβ = {x : M f (x) > β}. We know 1 d(Bα ) ≤ d(Bj0 ). In fact. . 2 ∗ or x ∈ Bj0 . . . such that 5n k m(Bk ) ≥ m(Eβ ). Extract a disjoint sequence B1 . then there exists a ball Bx centered at x such that Bx f (x) dx > βm(Bx ) by the deﬁnition of M f (x).1. We say f is locally integrable if K f (x) dx is ﬁnite whenever K is compact. If x ∈ Eβ . . 2 ∗ and some simple geometry shows that Bα ⊂ Bj0 . or else we would have chosen it instead of Bk+1 . Then m(Bx ) ≤ f  . β so {Bx } is a cover of Eβ by balls of bounded diameter. If k m(Bk ) < ∞. since {Bα } is a cover of E. The function M f is called the maximal function of f . Then 5n m(Eβ ) ≤ 5n m(Bk ) ≤ f  β Bk k k . Therefore Bα ⊂ Bj0 . Proof. then for all β > 0 m({x : M f (x) > β}) ≤ 5n β f (x) dx. but comes close in a certain sense to doing so. r)) f (y) dy. B2 . If x ∈ Bα . due to Hardy and Littlewood. Bk . Let k be the smallest integer such that d(Bk+1 ) < 1 d(Bα ). . If Bα is one of the Bk . then x − xj0  ≤ x − y + y − xj0  < d(Bα ) + d(Bj0 )/2 ≤ 5 d(Bj0 ). If f is locally integrable. we are looking at the average of f  over B(x.2 If f is integrable.
2). The second term on the right is 0 by (14. and thus we may suppose f is integrable. as r → 0. fr (x) → f (x) for almost every x ∈ B(0.r) (14. 5n β CHAPTER 14. r)) [g(y) − g(x)] dy B(x. We now use Theorem 14. Let N > 0. It suﬃces to prove that for each N .2) g(y) − g(x) dy → 0 B(x.3 Let fr (x) = 1 m(B(x.4 to write m({x : lim supfr (x) − f (x) > β}) r→0 . then fr (x) → f (x) a. gr (x) − g(x) = 1 m(B(x. If gr is deﬁned analogously to fr using (14. 2N ).1). DIFFERENTIATION f  ≤ ∪k Bk 5n β f . Using Theorem 8. so M f is not integrable.1) If f is locally integrable. Theorem 14.r) (14. note that M f (x) is approximately a constant times x−n for x large. r)) 1 ≤ m(B(x. N ). Hence M does not map the class of integrable functions into the class of integrable functions.4.e.r) as r → 0 by the continuity of g. Fix β > 0. Let ε > 0. If we look at the function f (x) = χB .2 and Lemma 10. We have lim sup fr (x) − f (x) ≤ lim sup fr (x) − gr (x) r→0 r→0 + lim sup gr (x) − g(x) r→0 + g(x) − f (x).110 = as desired. r)) f (y) dy. take g continuous with compact support such that f − g dm < ε. B(x. Proof. We may suppose without loss of generality that f is 0 outside of B(0. where B is the unit ball.
β ≤ where we use the deﬁnition of the maximal function to see that fr (x) − gr (x) ≤ M (f − g)(x) for all r. r→0 f − g β/2 We apply this with β = 1/j for each positive integer j. MAXIMAL FUNCTIONS ≤ m({x : lim sup fr (x) − gr (x) > β/2}) r→0 111 + m({x : f (x) − g(x) > β/2}) ≤ m({x : M (f − g)(x) > β/2}) + 2(5n + 1) f − g β 2(5n + 1)ε < . f (y) − f (x) dy → 0 B(x.4 For almost every x 1 m(B(x. j=1 r→0 This is the result we seek. For each rational c there exists a set Nc of measure 0 such that 1 f (y) − c dy → f (x) − c m(B(x.r) . This is true for every ε.r) Proof.14. and we conclude m({x : lim supfr (x) − f (x) > 0}) r→0 ≤ ∪∞ m({x : lim sup fr (x) − f (x) > 1/j}) = 0. r)) B(x. r)) as r → 0. We can get a stronger statement: Theorem 14.1. so m({x : lim sup fr (x) − gr (x) > β}) = 0.
DIFFERENTIATION for x ∈ Nc . Let ε > 0 and / choose c rational such that f (x) − c < ε. 14. m(B(x. the ratio tends to 0.r) Since ε is arbitrary. our result follows. x+h).r) or m(E ∩ B(x. We can use the results on maximal functions to show that the derivative of the antiderivative of an integrable function is the function itself. r) B(x. for almost all x ∈ E. we see this by applying Theorem 14. r)) B(x.3 to the function / f (x) − c.r) 1 + f (x) − c dy m(B(x. h) in R is merely the interval (x−h.112 CHAPTER 14. If we apply the above to the function f = χE . B(x. r)) → 1.r) 1 = f (y) − c dy + f (x) − c m(B(x. r)) f (y) − f (x) dy ≤ 2f (x) − c < 2ε. r)) f (y) − f (x) dy B(x.r) ≤ 1 f (y) − c dy m(B(x. m(B(x. Let N = ∪c∈Q Nc and suppose x ∈ N . . then for almost all x ∈ E 1 χE → 1.r) and hence lim sup r→0 1 m(B(x. The points / where the ratio tends to 1 are called points of density for E. A ball B(x. r)) and similarly. Then 1 m(B(x. We use m for Lebesgue measure throughout. r)) B(x.2 Antiderivatives For the remainder of the chapter we consider realvalued functions on the real line R. r)) B(x. Deﬁne the indeﬁnite integral or antiderivative of an integrable function f by x F (x) = a f (t) dt.
r)) =0 m(B(x.14. so F (x + h) − F (x) 1 − f (x) = h h ≤2 x+h (f (y) − f (x)) dy x x+h 1 m(B(x. The left hand derivative is handled similarly. If h > 0. r)) r→0 for almost every x. Let λ be the LebesgueStieltjes measure deﬁned using the function H and suppose λ and m are mutually singular. Deﬁne x F (x) = a f (y) dy. .4.6 that F is continuous. We start with right continuous increasing functions. and we conclude the right hand derivative of F exists and equals f for almost every x. x−h By Theorem 14. 113 Theorem 14. the right hand side goes to 0 as h → 0 for almost every x.6 Suppose H : R → R is increasing.3 Bounded variation In this section we show that functions of bounded variation are diﬀerentiable almost everywhere. Lemma 14. right continuous. Proof. and constant for x ≥ 1 and x ≤ 0. Then lim λ(B(x.e. we have x+h F (x + h) − F (x) = x f (y) dy. 14. h)) f (y) − f (x) dy. BOUNDED VARIATION Recall by Exercise 7. Then F is diﬀerentiable almost everywhere and F (x) = f (x) a.3.5 Suppose f : R → R is integrable and a ∈ R.
i=1 Since H is right continuous. bi ] i=1 and ∞ [H(bi ) − H(ai )] < ε/2. r)) >β . 1] : lim sup λ(B(x. 2). . for each i there exists bi > bi such that H(bi ) − H(bi ) < ε/2i+1 . m(E) = 0. bi ]) = i=1 [H(bi ) − H(ai )] < ε. then G is open. there exist ai < bi such that F ⊂ ∪∞ (ai . ∞).114 CHAPTER 14. bi )) ≤ i=1 λ((ai . 1] : lim sup r→0 λ((x. r→0 . such that ∞ m(Aβ ) ≤ 5 i=1 m(Bi ). then x ∈ F ⊂ G. m(B(x. This is clear if x < 0 or x > 1. Let ε > 0. Use Proposition 14. β β Since ε is arbitrary. we can take G to be the set G = G ∩ (−1. then m(Aβ ) = 0. By the deﬁnition of LebesgueStieltjes measure. If G = ∪∞ (ai . G contains F . Then ∞ m(Aβ ) ≤ 5 i=1 m(Bi ) ≤ 5 β ∞ λ(Bi ) ≤ i=1 5 5 λ(G) ≤ ε. B2 . and our construction of G did not depend on β. bi ). r)) > 0}) = 0. The ﬁrst step of the proof is to ﬁnd a bounded open set G such that F ⊂ G and λ(G) < ε. there exist measurable sets E and F such that λ(F ) = 0. If β > 0. DIFFERENTIATION Proof. and F = E c . and i=1 ∞ ∞ ∞ λ(G ) ≤ i=1 λ((ai . and there exists an open ball Bx centered at x and contained in G such that λ(Bx )/m(Bx ) > β. Step 1. Since λ ⊥ m. r))/m(B(x. let Aβ = x ∈ F ∩ [0. . r)) The second step is to show that m(Aβ ) = 0. 0] and [1. . Since m(A1/k ) = 0 for each k. Step 2. then m({x ∈ F ∩ [0.1 to ﬁnd a disjoint subsequence B1 . If x ∈ Aβ . Since H is constant on (−∞.
then F is still right continuous and increasing. 1]) ≤ ν([0. The same is true for the left hand derivative. By Exercise 7.14. on [−N. Note ρ([0.e. Once we have that. on [0.e.e. Then F exists a. By the Lebesgue decomposition theorem. x]) = ν((0. so H is right continuous. and we conclude that F is diﬀerentiable a. BOUNDED VARIATION Since m(E) = 0. Moreover. Let ν be the LebesgueStieltjes measure deﬁned in terms of F . 1]. .e. F is locally inteb grable and for every a < b. the same argument can be used to show that F is diﬀerentiable a.5 that the function x → 0 f (y) dy is diﬀerentiable almost everywhere.3. on R. a F (x) dx ≤ F (b) − F (a). 1]) = F (1) − F (0). We saw by Theox rem 14. so H exists and equals 0 for almost every x. and λ is the LebesgueStieltjes measure deﬁned in terms of H. x]) = F (x)−F (0)− 0 x f (y) dy.e. increasing. x + h]) = lim sup h h→0+ λ(B(x. we can write ν = λ + ρ. By Lemma 14. 1] except possibly at the points 0 and 1. where λ ⊥ m and ρ m. x])−ρ((0. 115 Proposition 14.6. If we redeﬁne F so that F (x) = limy→0+ F (y) if x ≤ 0 and F (x) = F (1) if x > 1. lim sup h→0+ H(x + h) − H(x) H(x + h) − H(x − h) ≤ lim sup h h h→0+ λ((x − h. Proof.6. and we have not aﬀected the diﬀerentiability of F on [0. and that proves that F is diﬀerentiable a. By the RadonNikodym theorem there exists a nonnegative integrable function f such that ρ(A) = A f dm for each measurable A. 2h)) ≤ 4 lim sup =0 4h h→0+ for almost every x.7 Let F : R → R be an increasing and right continuous function. Let x H(x) = λ((0. N ] for each N . We will show F is diﬀerentiable a. the function x → 0 f (y) dy is continuous. this completes the proof.
b]) = F (b) − F (a).8 If F : R → R is increasing. Since F is increasing. Then lim sup h→0+ F (x + h) − F (x) G(x + (1 + ε)h) − G(x) ≤ lim sup h h h→0+ G(x + (1 + ε)h) − G(x) = (1 + ε) lim sup (1 + ε)h h→0+ = (1 + ε)L. the right hand derivative of F exists at x and is equal to L. and b F (x) dx ≤ F (b) − F (a) a (14.116 CHAPTER 14.e. Since ε is arbitrary. b]) ≤ ν((a. let L = G (x) and let ε > 0. .e. for any h > 0 there exists a point xh strictly between x + h and x + (1 + ε)h where F and G agree.7. then F (x) exists and is equal to G (x). Let G(x) = limy→x+ F (y).e.3) whenever a < b. Here is the main theorem on the diﬀerentiation of increasing functions. and so F (x + h) ≤ F (xh ) = G(xh ) ≤ G(x + (1 + ε)h). That the left hand derivative equals L is proved similarly. there are at most countably many values of x where F is not continuous. Similarly. Theorem 14. lim inf h→0+ [F (x + h) − F (x)]/h ≥ (1 − ε)L. If a < b. by Proposition 14. b b F (x) dx = a a f (x) dx = ρ((a. Because F and G are increasing. Let x be such a point. DIFFERENTIATION We have shown that F = f a. G is diﬀerentiable a. Since G is increasing and right continuous. so F (x) = G(x) a.e. then F exists a. Proof. We will show that if x is a point where G is diﬀerentiable and at the same time F (x) = G(x). This completes the proof.
Lemma 14. the diﬀerence of two increasing functions on [a. b]. . then F is locally integrable. Now let n → ∞ and use the monotone convergence theorem. If a < b. and we do not in general have equality in (14.14.. take an ↓ a and bn ↑ b such that F and G agree on an and bn . Thus 1 1 = F (1) − F (0) > 0 = 0 F (x) dx.3).3.9 Note that if F is the CantorLebesgue function. b]. BOUNDED VARIATION 117 Since F = G a.10 If f is of bounded variation on [a. b]. A realvalued function f is of bounded variation on [a. where the supremum is over all partitions a = x0 < x1 < · · · < xk = b of [a. Then using Proposition 14.e. where C is the Cantor set.7. Remark 14. then F (x) = 0 a. b] if k sup i=1 f (xi ) − f (xi−1 ) is ﬁnite..e. Proof. in fact at every point of C c . then f can be written as f = f1 − f2 . Deﬁne k f1 (y) = sup i=1 [f (xi ) − f (xi−1 )]+ and k f2 (y) = sup i=1 [f (xi ) − f (xi−1 )]− . F (b) − F (a) ≥ F (bn ) − F (an ) bn = G(bn ) − G(an ) ≥ an bn G (x) dx = an F (x) dx.
14. c]. then it is of bounded variation. It is easy to see that absolutely continuous functions are continuous and that the CantorLebesgue function is not absolutely continuous.12 If f is absolutely continuous. there exists δ such that i=1 f (bi ) − f (ai ) < 1 whenever . but is not of bounded variation. We make the observation that if f is of bounded variation on the interval [a. Lemma 14. By the deﬁnition of absolutely continuous function with k ε = 1. bi )} k is a ﬁnite collection of disjoint intervals with i=1 bi − ai  < δ. y] yields f1 (y) = f2 (y) + f (y) − f (a). we see that functions of bounded variation are diﬀerentiable a. Proof. Remark 14. then the quantity (f1 (b) + f2 (b)) − (f1 (a) + f2 (a)) is called the total variation of f on the interval [a. taking the supremum over all partitions of [a.e. Since k k [f (xi ) − f (xi−1 )]+ = i=1 i=1 [f (xi ) − f (xi−1 )]− + f (y) − f (a). b] if given ε k there exists δ such that i=1 f (bi )−f (ai ) < ε whenever {(ai . DIFFERENTIATION where the supremum is over all partitions a = x0 < x1 < · · · < xk = y for y ∈ [a. c]. Using this lemma and Theorem 14. Note that the converse is not true: the function sin(1/x) deﬁned on (0.4 Absolutely continuous functions A realvalued function f is absolutely continuous on [a. b]. then it is of bounded variation on the interval [a. and the result follows by solving for f (y). 1] is diﬀerentiable everywhere. b] and on the interval [b. f1 and f2 are measurable since they are both increasing.118 CHAPTER 14. Clearly f1 and f2 are increasing in y. b].8.11 If we write a function f of bounded variation as the diﬀerence of two increasing functions f1 and f2 .
a + (j + 1)δ] is less than or equal to 1. a Proof. bi ) are disjoint open intervals. . ABSOLUTELY CONTINUOUS FUNCTIONS k i=1 (bi −ai ) 119 ≤ δ and the (ai .13 Suppose f is of bounded variation and we decompose f as f = f1 − f2 . then so are f1 and f2 . Hence k Ji −1 f (si.4.13 it suﬃces to suppose F is increasing and absolutely continuous.14. and b F (x) dx = F (b) − F (a). then k Ji −1 k k (si. If f is absolutely continuous. By Lemma 14. Lemma 14.. Hence the total variation of f on [a + jδ. Since F is continuous.e. Using Remark 14. Given ε there exists δ such that i=1 f (bi ) − f (ai ) < k ε whenever i=1 (bi − ai ) ≤ δ and the (ai . bi ) into subintervals with ai = si0 < si1 < · · · < siJi = bi . F (d) − F (c) = ν((c. Theorem 14.14 If F is absolutely continuous. Here is the main theorem on absolutely continuous functions. bi ) are disjoint open intervals. i=1 j=0 Taking the supremum over all such partitions. we see the total variation of f on [a. b] is ﬁnite. Proof. Let ν be the LebesgueStieltjes measure deﬁned in terms of F . k (f1 + f2 )(bi ) − (f1 + f2 )(ai ) ≤ ε.j+1 − sij ) = i=1 j=0 i=1 (bi − ai ) ≤ δ. d)). then F exists a. where f1 and f2 are increasing functions. i=1 and our conclusion follows.j+1 ) − f (sij ) ≤ ε. Partitioning each interval (ai .11.
b]. Since any open set G can be written as the union of disjoint intervals {(ai . Setting x = b we obtain b F (b) − F (a) = a F (y) dy as desired. 14. In particular. We conclude that ν m. and 14. for each x ∈ [a.5. and then ν(A) ≤ ν(G) ≤ ε. Hence there exists a nonnegative integrable function f such that ν(A) = f dm A for all Borel measurable sets A. and instead proceeds via a covering lemma due to Vitali. bi )}. F exists and is equal to f a. x)) = a f (y) dy. The following is known as the Vitali covering lemma.120 CHAPTER 14.14 that avoids the use of the RadonNikodym theorem. . then there exists an open set G containing A such that m(G) < δ. By Theorem 14. we see that given ε there exists δ such ∞ that i=1 F (bi ) − F (ai ) ≤ ε whenever {(ai .1. Let m be Lebesgue measure. bi )} is a collection of ∞ disjoint intervals with i=1 (bi − ai ) < δ. this can be rephrased as saying. 14.5.5 Approach 2 – diﬀerentiability In this and the following two sections we give an alternative approach to Theorems 14. given ε there exists δ such that ∞ ∞ ν(G) = i=1 ν((ai . x F (x) − F (a) = ν((a. DIFFERENTIATION Taking a limit as k → ∞. We say G is a Vitali cover of E if for each x ∈ E and each ε > 0 there exists an interval G ∈ G containing x whose length is less than ε. bi )) = i=1 (F (bi ) − F (ai )) ≤ ε whenever G is open and m(G) < δ.e. and is a reﬁnement of Proposition 14.8. Let E ⊂ R be a measurable set and let G be a collection of intervals. If m(A) < δ and A is Borel measurable.
If not. choosing In+1 disjoint from I1 . I2 . note n > N . . ∗ Let R = E −∪N Ii . then n an ≤ 2 n m(In ) < ∞. Therefore x ∈ In . In ∈ G such that m(E − ∪n In ) < ε. . then i=1 m(Ii ) ≤ m(G) < ∞. Hence there exists i=1 an interval I ∈ G containing x with I disjoint from I1 . the distance from x to the midpoint of In is at most ∗ m(I) + m(In )/2 ≤ (5/2)m(In ). we are done. and an → 0. Let a0 = sup{m(I) : I ∈ G}. . IN . . Since G is a Vitali cover. Continue in this way. .14. Let I1 be any element of G with m(I1 ) ≥ a0 /2. Let a1 = sup{m(I) : I ∈ G. then ∪n Ii is closed. Since they are ∞ disjoint and all contained in G. We claim m(R) < ε. Since n m(In ) < ∞. . ∞ Therefore there exists N such that i=N +1 m(Ii ) < ε/5. . In and in G with length at least one half as large as any other such interval in G that is disjoint from I1 . I disjoint from I1 }. ∗ Thus we have R ⊂ ∪∞ +1 Ii . i=1 Proof. . APPROACH 2 – DIFFERENTIABILITY 121 Lemma 14. In . Since x is in I and I intersects In . . so i=N ∞ ∞ ∗ m(Ii ) = 5 i=N +1 i=N +1 m(R) ≤ m(Ii ) < ε. . Let G be an open set of ﬁnite measure containing E.15 Suppose E has ﬁnite measure and let G be a Vitali cover of E.5. . since the set of endpoints of a ﬁnite subcollection will have measure 0. . Let In be the interval i=1 with the same center as In but ﬁve times the length. . . . and choose I2 ∈ G disjoint from I1 such that m(I2 ) ≥ a1 /2. Given ε > 0 there exists a ﬁnite subcollection of disjoint intervals I1 . Let n be the smallest integer such that I intersects In . . we may suppose without loss of generality that each set of G is contained in G. for otherwise we would have chosen I at some stage. Let x ∈ R. If the process stops at some ﬁnite stage. Hence I must either be one of the In for some n > N or at least intersect it. . we generate a sequence of disjoint intervals I1 . We may replace each interval in G by a closed one. . We have m(I) ≤ an−1 ≤ 2m(In ). . . Since we supposed each interval in G was to be modiﬁed so as to include its endpoints.
xn )). the other sets being similar.4) Proof. h h→0− f (x + h) − f (x) D+ f (x) = lim inf .15 to choose I1 . Then f is differentiable almost everywhere. h h→0+ f (x + h) − f (x) D− f (x) = lim sup . . h→0− h D+ f (x) = lim sup If all the derivates are equal. Let Euv = {x : D+ f (x) > v > u > D− f (x)}. Let s = m(Euv ). If we show m(Euv ) = 0. Theorem 14. Let us consider the set E = {x : D+ f (x) > D− f (x)}. We will show that the set where any two derivates are unequal has measure zero. xn ]. and choose an open set G such that Euv ⊂ G and m(G) < s + ε. let ε > 0. For each x ∈ Euv there exists an arbitrarily small interval [x − h. Use Lemma 14. x] contained in G such that f (x) − f (x − h) < uh. h→0+ h f (x + h) − f (x) D− f (x) = lim inf .16 Suppose f is increasing on [a. b]. IN which are disjoint and whose interiors cover a subset A of Euv of measure greater than s − ε. f is integrable.v∈Q Euv will show that m(E) = 0. and b f (x) dx ≤ f (b) − f (a). . Thus A = Euv ∩ (∩N (xn − hn . n=1 . . Write In = [xn − hn . . a (14. then observing that E ⊂ ∪u<v. CHAPTER 14. then f is diﬀerentiable at x and f (x) is the common value.u. we deﬁne the derivates of f at x by f (x + h) − f (x) .122 This completes the proof. DIFFERENTIATION Given a realvalued function f .
14. {i:Ji ⊂In } since f is increasing. Since v > u.15 again.5. gn ≥ 0. we pick out a ﬁnite collection J1 . Deﬁne f (x) = f (b) if x ≥ b. Using Lemma 14. and D− f = D− f a. Let gn (x) = n[f (x + 1/n) − f (x)]. APPROACH 2 – DIFFERENTIABILITY Taking a sum.e. N n 123 [f (xn ) − f (xn − hn )] < u n=1 n=1 hn < um(G) < u(s + ε). . and so u(s + ε) > v(s − 2ε). We prove similarly that D+ f = D+ f a. Thus N M [f (xn ) − f (xn − hn )] ≥ n=1 i=1 [f (yi + ki ) − f (yi )]. Since f is increasing. By Fatou’s lemma and the fact that f is . and so g is measurable.e. y + k) that is contained in some In and for which f (y + k) − f (y) > vk. Summing over these intervals yields M M [f (yi + ki ) − f (yi )] > v i=1 i=1 ki > v(s − 2ε). Then gn (x) → g(x) for almost all x. Each interval Ji is contained in some interval In . so us ≥ vs. this implies s = 0. and conclude that g(x) = lim f (x + h) − f (x) h h→0 is deﬁned almost everywhere and that f is diﬀerentiable wherever g is ﬁnite. and if we sum over those i for which Ji ⊂ In we ﬁnd [f (yi + ki ) − f (yi )] ≤ f (xn ) − f (xn − hn ). JM whose union contains a subset of A of measure larger than s − 2ε. . . This is true for each ε. Each point y in the subset A is the left endpoint of an arbitrarily small interval (y. .
By writing f = f + − f − . We used a change of variables in the second equality. Then F (x + 1/n) − F (x) = n 1/n x+1/n f (t) dt x . We refer the reader to Lemma 14.17 If f is integrable and F is deﬁned by (14.124 increasing.6 we know F is continuous. we look at when the derivative of x F (x) = a f (t) dt (14.5).e. 14. This shows that g is integrable and hence ﬁnite almost everywhere. and so F exists a. Theorem 14.5) is equal to f (x) a. then F (x) = f (x) for almost every x. Suppose for the moment that f is bounded by K. b CHAPTER 14. it suﬃces to consider the case where f is nonnegative. DIFFERENTIATION b g ≤ lim inf a n a gn b = lim inf n n a [f (x + 1/n) − f (x)] b+1/n a+1/n = lim inf n n b f −n a a+1/n f f = lim inf f (b) − n n a ≤ f (b) − f (a). Proof.e. In this case F is increasing.10 to see that a function of bounded variation is diﬀerentiable almost everywhere. By Exercise 7.6 Approach 2 – antiderivatives Continuing the alternative approach.
c c 125 F (x) dx = lim n a n a [F (x + 1/n) − F (x)] dx c+1/n a+c = lim n n c F (x) dx − n a c F (x) dx = F (c) − F (a) = a f (x) dx. . and so b b a. GK (x + 1/n) − GK (x) GK (x) = lim ≥0 n→∞ 1/n at points x where G exists since G is increasing. F ≥ a a f = F (b) − F (a). then x GK (x) = a [f − (f ∧ K)] dx is increasing..14. Since F − f ≥ 0 a. Since K is arbitrary. which implies F = f a. and hence has a derivative almost everywhere.. we know the derivative of x HK (x) = a (f ∧ K) dx is equal to f ∧ K almost everywhere.e. By the preceding paragraph.6. Therefore F (x) = GK (x) + HK (x) ≥ (f ∧ K)(x). Since f − (f ∧ K) ≥ 0.e. We used a change of variables for the second equality and the fact that F is continuous for the third equality.e.e.e. b Combining with (14. Therefore c [F (x) − f (x)] dx = 0 a for all c. Moreover. this tells us that F = f a.3.4) we conclude that a [F − f ] = 0. We continue to assume f is nonnegative but now allow f to be unbounded. F ≥ f a. APPROACH 2 – ANTIDERIVATIVES is also bounded by K. By dominated convergence. by Corollary 8.
c] : f (x) = 0}. x+h] ⊂ [a. then f is constant.12 for the proof that if f is absolutely continuous.18 If f is absolutely continuous on [a. Choose δ such that i=1 f (bi ) − f (ai ) < ε whenever K bi − ai  ≤ δ and the (ai . Proof. For each i=1 point x ∈ E ∩[a. c] such that f (x + h) − f (x) < εh. then f (c) = f (a). let E = {x ∈ [a.. c) there exist arbitrarily small intervals [x. Except for a set of measure less than δ. The CantorLebesgue function is an example to show that we need the absolute continuity.126 CHAPTER 14. bi ). By our choice of δ and the deﬁnition of absolute continuity. Since ε is arbitrary. We label the intervals [ai . bi ] so that ai < bi ≤ ai+1 . DIFFERENTIATION 14. f (c) − f (a) = i [f (ai+1 ) − f (bi )] + i [f (bi ) − f (ai )] ≤ ε + ε(c − a). a We refer the reader to Lemma 14. E is covered by ∪i (ai .7 Approach 2 – absolute continuity Finally. Adding these two inequalities together. we continue the alternative approach to look at when b F (y) dy = F (b) − F (a). . f (ai+1 ) − f (bi ) < ε.e. b] and f (x) = 0 a. which implies that f is constant. i On the other hand. by our choice of the intervals (ai . f (bi ) − f (ai ) < ε i i (bi − ai ) ≤ ε(c − a). By Lemma 14. and let K ε > 0.15 we can ﬁnd a ﬁnite disjoint collection of such intervals that cover all of E except for a set of measure less than δ. bi ). then it is of bounded variation. Let c ∈ [a. b]. or i ai+1 − bi  ≤ δ. Lemma 14. bi ) are disjoint intervals. ai+1 ) has measure less than δ. This implies that ∪i (bi .
2. then G is absolutely continuous by Exercise 14. If x G(x) = a F (t) dt. so F = F1 − F2 where F1 and F2 are increasing. prove that F is of bounded variation and is absolutely continuous. EXERCISES Theorem 14. then F (x) dx ≤ (F1 (b) + F2 (b)) − (F1 (a) − F2 (a)). 14. using Theorem 14. Then (F − G) = F − G = F − F = 0 a. we get our result.14. b]. and thus F (x) − G(x) = F (a) − G(a). a ∈ R. F − G is constant. If we set x = b. then the product f g is also.19 If F is absolutely continuous. (2) Prove the integration by parts formula: b b f b)g(b) − f (a)g(a) = a f (x)g (x) dx + a f (x)g(x) dx. Proof.8. We conclude x F (x) = a F (t) dt + F (a). By Lemma 14.1 (1) Show that if f and g are absolutely continuous on an interval [a.8 Exercises Exercise 14. b]. and F exists a.2 If f is integrable and realvalued. and hence F is integrable. Then F is of bounded variation..17 for the second equality. Suppose F is absolutely continuous on [a. then b 127 F (b) − F (a) = a F (y) dy. and x F (x) = a f (y) dy. . and hence F − G is absolutely continuous.e.18.e. Exercise 14. Since F (x) ≤ F1 (x) + F2 (x).
5 A realvalued function f is Lipschitz with constant M if f (x) − f (y) ≤ M x − y for all x.4 Suppose f is a realvalued continuous function on [0.8 If f is realvalued and diﬀerentiable at each point of [0. Prove that M is Borel measurable and M (y) dy equals the total variation of f on [a. Exercise 14. 1] and that ε > 0. 1). Exercise 14. DIFFERENTIATION Exercise 14.3 Suppose that f is a realvalued continuous function on [0. is f necessarily absolutely continuous on [0. then f (A) has Lebesgue measure 0. y ∈ R. Exercise 14.1] Exercise 14. 1]? If f is also of bounded variation on [0. b] → R is continuous. . b].7 Suppose f is absolutely continuous on [0.e.e.e.128 CHAPTER 14. Exercise 14. b] such that f (x) = y. Prove that there exists a continuous function g such that g (x) exists and equals 0 for a. M (y) may be ﬁnite or inﬁnite. ﬁnd a counterexample. x∈[0. 1]? If not. Prove that ∞ F (x) = n=1 Fn (x) for almost every x.10 If f : [a. 1] we let f (A) = {f (x) : x ∈ A}. ∞ Let F = n=1 Fn . Is f necessarily absolutely continuous on [0. 1] and f is absolutely continuous on (a.9 Find an increasing function f such that f = 0 a. 1] and for A ⊂ [0. Prove that f is Lipschitz with constant M if and only if f is absolutely continuous and f  ≤ M a. is f absolutely continuous on [0. give counterexamples. 1] for every a ∈ (0. 1]? If not. x and sup f (x) − g(x) < ε. let M (y) be the number of points x in [a. but f is not constant on any open interval. 1]. Exercise 14. 1] such that supn Fn (1) < ∞. Prove that if A has Lebesgue measure 0. 1]. Exercise 14.6 Suppose Fn is a sequence of increasing nonnegative right continuous functions on [0.
lim r→0+ 2r Exercise 14. 1] → R that is absolutely continuous. Exercise 14. r]) = α. Find a Borel subset E of [−1.14 Let (X.13 Let ∞ f (x) = −∞ e−xy dy.15 Suppose A ⊂ [0. 1 + y2 2 (1) Find the derivative of f . b] and the derivate D+ f is nonnegative on [a. . Prove that f (b) ≥ f (a).11 Let α ∈ (0. (2) Prove that limx→∞ g(x) = ∞ and limx→−∞ g(x) = ∞. Exercise 14. but f (x + h) − f (x) lim =∞ h→0 h for each x ∈ A. Exercise 14. A. µ) be a measure space and let f be a realvalued integrable function.14. 1] has Lebesgue measure zero. 1] such that m(E ∩ [−r. Find the solution to this ordinary diﬀerential equation to determine an explicit value for f (x). b]. EXERCISES 129 Exercise 14. (3) Find g (x) and prove that g(x0 ) = inf x∈R g(x) if and only if µ({y : f (y) > x0 }) = µ({y : f (y) < x0 }). 1). (2) Find an ordinary diﬀerential equation that f solves.8. Find an increasing function f : [0.16 Suppose that µ is a measure on the Borel σalgebra on [0.12 Suppose f is a realvalued continuous function on [a. 1] and for every f that is realvalued and continuously diﬀerentiable we have 1 f (x) µ(dx) ≤ 0 f (x)2 dx 1/2 . Deﬁne g(x) = f (y) − x µ(dy) for x ∈ R. What if instead we have that D+ f is nonnegative on [a. b]? Exercise 14. (1) Prove that g is absolutely continuous.
DIFFERENTIATION (1) Show that µ is absolutely continuous with respect to Lebesgue measure on [0. Deﬁne ∞ H(t) = −∞ f (x) + tg(x)p dx for t ∈ R. y ∈ [0. Prove that H is a diﬀerentiable function and ﬁnd its derivative. . x. 1].130 CHAPTER 14. prove that there exists a constant c > 0 such that g(x) − g(y) ≤ cx − y1/2 .17 Let p > 1 and f. 1]. Exercise 14. (2) If g is the RadonNikodym derivative of µ with respect to Lebesgue measure. g ∈ Lp (R).
(15. We assume throughout this chapter that the measure µ is σﬁnite.1) For p = ∞. 15. For 1 ≤ p < ∞. We deﬁne the Lp norm of a function. deﬁne the Lp norm of f by f p = f (x)p dµ 1/p . we deﬁne q by 1 1 + =1 p q 131 . If 1 < p < ∞. and consider the bounded linear functionals on Lp . prove completeness of the norm.1 Norms Let (X.Chapter 15 Lp spaces We introduce some spaces of functions. For 1 ≤ p ≤ ∞ the space Lp is the set {f : f p < ∞}. It is clear that f p = 0 if and only if f = 0 a. One can also write Lp (X) or Lp (µ) if one wants to emphasize the space or the measure.2) Thus the L norm of a function f is the smallest number M such that f  ≤ M a. deﬁne the L∞ norm of f by f ∞ ∞ = inf{M : µ({x : f (x) ≥ M }) = 0}. called the Lp spaces.e. (15. discuss convolutions. µ) be a σﬁnite measure space.e. A.
The Now let us assume 1 < p. this is the CauchySchwarz inequality. If M = f ∞ . If f p = 0.3) that G(x)q F (x)p F (x)G(x) ≤ + . Any function whose second derivative is everywhere nonnegative is convex. LP SPACES and call q the conjugate exponent of p. p q Clearly this inequality also holds if F (x) = 0 or G(x) = 0. Basic to the study of Lp spaces is H¨lder’s inequality. q < ∞ and o p−1 + q −1 = 1. b = q log G(x). We then obtain from (15. then f  ≤ M a.3) for every pair of reals a ≤ b. and f g = 0. and it suﬃces to show that F G dµ ≤ 1. If F (x). so if 0 ≤ λ ≤ 1. F p G q 1 1 p q F G dµ ≤ + = + = 1. λ = 1/p. Proof. Note F p = 1 and G q = 1. p g q. then f = 0 a. Proposition 15.132 CHAPTER 15. and case p = ∞ and q = 1 follows. G(x) = 0. fg ≤ M g. and 1 − λ = 1/q. p q p q This completes the proof. Note that o when p = q = 2. which is everywhere positive. Let F (x) = f (x)/ f p and G(x) = g(x)/ g q . Integrating. we have eλa+(1−λ)b ≤ λea + (1 − λ)eb (15. The second derivative of the function ex is again ex . then f g dµ ≤ f This also holds if p = ∞ and q = 1.e.e.1 (H¨lder’s inequality) If 1 < p. One application of H¨lder’s inequality is to prove Minkowski’s o inequality. so the result is clear if f p = 0 and similarly if g q = 0. which is simply the triangle inequality for Lp . let a = p log F (x). q < ∞. We ﬁrst need the following lemma: .
3 (Minkowski’s inequality) If 1 ≤ p ≤ ∞. The case p = ∞ is also easy. Dividing both sides by ap . so we assume a > 0. and the only solution to f (x) = 0 on (0. then (p − 1)q = p. the result is obvious. ∞) is x = 1. We obtain o p f (x)p dµ + 2p−1 g(x)p dµ. p < ∞. If f p or g p is inﬁnite. Now let us suppose 1 < p < ∞. The case a = 0 is obvious. We conclude that f takes its minimum at x = 1 and hence f (x) ≥ 0 for x ≥ 0. letting x = b/a. .1.15. Now write f + gp ≤ f  f + gp−1 + g f + gp−1 1 and apply H¨lder’s inequality with q = (1 − p )−1 . integrating gives the case when p = 1. and setting f (x) = 2p−1 + 2p−1 xp − (1 + x)p . Proof. Note f (0) > 0. then f +g p ≤ f p + g p. b ≥ 0 and 1 ≤ p < ∞. (f + g)(x)p dµ ≤ 2p−1 We therefore have f +g 0. The inequality Lemma 15. so we have f +g p p ≤ f p + g p f +g p/q p . NORMS Lemma 15. Proposition 15.2 with a = f (x) and b = g(x) yields. the inequality we want to prove is equivalent to showing f (x) ≥ 0 for x ≥ 0. after an integration. Since (f + g)(x) ≤ f (x) + g(x). Since p−1 + q −1 = 1. then (a + b)p ≤ 2p−1 ap + 2p−1 bp . limx→∞ f (x) = ∞.2 If a. so we may assume both are ﬁnite. 133 Proof. Clearly we may assume f +g > f + gp ≤ f p f + g(p−1)q 1/q + g p f + g(p−1)q 1/q . f (1) = 0.
This is an equivalence relation for functions. 15. Recall Deﬁnition 10. Our ﬁrst step is to ﬁnd a certain subsequence. only that f = 0 a.2 Completeness We show that the space Lp viewed as a metric space is complete.1. then Lp is complete. and deﬁne f p to be the Lp norm of any function in the same equivalence class as f . Given a realvalued measurable function f .1: fn converges to f in Lp if fn − f as n → ∞. Lp is a normed linear space. Theorem 15. this is equivalent to fn −f as n → ∞. In terms of Lp norms. The procedure we follow to circumvent this is to say two functions are equivalent if they diﬀer on a set of measure 0. the understanding is that we identify functions that are equal a. We henceforth keep this interpretation in the back of our minds when we talk about a function being in Lp . there exists nj . This is not quite right.e.4 If 1 ≤ p ≤ ∞. p/q p CHAPTER 15. the essential supremum and essential inﬁmum are deﬁned by ess sup f = inf{M : µ({x : f (x) > M }) = 0} and ess inf f = sup{m : µ({x : f (x) < m}) = 0}. We will do only the case p < ∞ and leave the case p = ∞ as Exercise 15. Given ε = 2−(j+1) .134 Dividing both sides by f +g 1 gives us our result. Step 1. Proof.e. We then have that · p is a norm on Lp . p p p →0 →0 Related to the deﬁnition of L∞ is the following terminology. We then deﬁne the space Lp to be the set of equivalence classes with respect to this equivalence relation. We would like to say that by virtue of Minkowski’s inequality. The Lp norm of a function satisﬁes all the properties of a norm except that f p = 0 does not imply that f is the zero function. Suppose fn is a Cauchy sequence in Lp . LP SPACES and using the fact that p−(p/q) = Recall the deﬁnition of normed linear space from Chapter 1.
If Aj = {x : fnj (x) − fnj−1 (x) > 2−j/2 }. gj (x) increases in j for each x. We have K f (x) = lim K→∞ [fnm (x) − fnm−1 (x)] = lim fnK (x) m=1 K→∞ since we have a telescoping series. COMPLETENESS 135 such that if n. By Fatou’s lemma. Set ∞ f (x) = [fnm (x) − fnm−1 (x)]. Step 2. so f is well deﬁned for a. Step 3. Set f (x) = 0 for any x where absolute convergence does not hold. In this step we show absolute convergence of this series. for each x there exists a j (depending on x) such that x ∈ ∪∞ Am . then fnj (x) − fnj−1 (x) ≤ 2−j .e. Set n0 = 0 and deﬁne f0 to be identically 0. m ≥ nj . m=1 We showed in Step 2 that this series is absolutely convergent for almost every x. Without loss of generality we may assume nj ≥ nj−1 for each j. Set gj (x) = m=1 j fnm (x) − fnm−1 (x). then fn − fm p ≤ 2−(j+1) . Hence for each x (except / m=j for those in the null set) there is a j0 (depending on x) such that if j ≥ j0 . except for a set of measure 0. Our candidate for the limit function will be j [fnj − fnj−1 ]. . The limit is ﬁnite for almost every x because for almost every x. µ(Aj ) ≤ 2−jp/2 . x. m=j Thus. We have ∞ µ(∩∞ ∪∞ Am ) = lim µ(∪∞ Am ) ≤ lim j=1 m=j m=j j→∞ j→∞ µ(Am ) = 0. We deﬁne our function f .4.15.2. then from Lemma 10. fnm (x) − fnm−1 (x) ≤ 2−m for m large (how large depends on x). f − fnj p p = f − fnj p ≤ lim inf K→∞ p p K→∞ fnK − fnj p = lim inf fnK − fnj ≤ 2(−j+1)p .
The same proof shows the following corollary. Since g − χE  ≤ 1. . Hence it suﬃces to approximate functions in Lp that have compact support. In particular. It is standard that a Cauchy sequence with a convergent subsequence itself converges. f − fm p p ≤ lim inf fnj − fm j→∞ p p ≤ εp if m ≥ N .136 CHAPTER 15. n ≥ N .3 Convolutions The convolution of two measurable functions f and g is deﬁned by f ∗ g(x) = f (x − y)g(y) dy.4 that there exists g continuous with compact support and with values in [0. it suﬃces to approximate characteristic functions with compact support. Suppose f ∈ Lp . there exists N such that fn − fm p < ε if m. We have thus shown that f − fnj p → 0 as j → ∞. Proof.n] p → 0 as n → ∞ by dominated convergence. Here is the proof in our case. b]) with respect to L2 ([a. a Borel measurable set contained in a bounded interval. and ε > 0. LP SPACES Step 4. Next we show: Proposition 15.5 The set of continuous functions with compact support is dense in Lp (R). then g − χE p ≤ g − χE  < ε. so it suﬃces to approximate simple functions with compact support. This shows that fm converges to f in Lp norm. Consider simple functions sm increasing to f . Given ε > 0. the dominating function being f p . By linearity. By writing f = f + − f − we may suppose f ≥ 0. We have f − f χ[−n. then we have f − sm p → 0 by dominated convergence. we showed in Proposition 8. Corollary 15. b] are dense in the space L2 ([a. fnj − fm p < ε if j is large enough. By Fatou’s lemma. This completes the proof. b]) norm. 1] such that g − χE  < ε.6 The set of continuous functions on [a. 15. Given E.
4) we conclude that f ∗ g is ﬁnite a. Proposition 15. we can apply the Fubini theorem to see that the right hand side is equal to f (x − y) dx g(y) dy = = f 1 f (x) dx g(y) dy g 1. but we do not consider them in this section.) H is a bounded linear functional if H = sup{Hf  : f p ≤ 1} (15. This together with (15. so f ∗ g = g ∗ f .28. From (15.1). We deﬁne the signum function or −1. 15. g ∈ Lp and a ∈ R. x > 0. g ∈ L1 .6) The ﬁrst equality here follows by a change of variables (see Exercise 8.4) Since the integrand on the right is nonnegative. however. See. (One can also have complexvalued linear functionals. then f ∗ g is in L1 and f ∗g Proof.5) proves (15.5) 1 ≤ f 1 g 1. (15. By a change of variables.7) is ﬁnite. (15.4 Bounded linear functionals A linear functional on Lp is a map H from Lp to R satisfying H(f + g) = H(f ) + H(g).4. this is the same as f (y)g(x − y) dy. sgn (x) = 0.15. .7 If f.e.4). x = 0. We have f ∗ g(x) dx ≤ f (x − y) g(y) dy dx. BOUNDED LINEAR FUNCTIONALS 137 provided the integral exists. Our goal in this section is to identify the dual of Lp . (15. sign function by x < 0. H(af ) = aH(f ) whenever f. 1. Exercise 15.7). The dual space of Lp is the collection of all bounded linear functionals with norm given by (15.
By the deﬁnition of L∞ norm. Case 1: p = 1. Take g(x) = sgn f (x).138 Note x sgn (x) = x. Let gn (x) = (sgn f (x)) sn (x)p−1 sn p/q p . the result is trivial. rn a sequence of nonnegative simple functions increasing to f − . so suppose f ∞ > 0.8) holds if we take q = ∞. sn p > 0. the measure of An = {x ∈ Fn : f (x) > a} must be positive if n is suﬃciently large. Then sn p → f p by monotone convergence. Let Fn be measurable sets of ﬁnite measure increasing to X. µ(An ) Then the L1 norm of gn is 1 and f gn = An f /µ(An ) ≥ a.8) must be M . Case 2: p = ∞. We may suppose f p > 0.8) is less than the left hand side by H¨lder’s inequality. Then sn (x) → f (x) for each x. qn a sequence of nonnegative simple functions increasing to f + . LP SPACES Theorem 15. Thus we need only show that the right o hand side is greater than the left hand side. Case 3: 1 < p < ∞. (15. each sn is a simple function. Let sgn f (x)χAn (x) gn (x) = . and sn (x) = (qn (x) − rn (x))χFn (x). The right hand side of (15.8 For 1 < p < ∞ and p−1 + q −1 = 1. If M = f ∞ . If f ∞ = 0. This takes care of the case p = 1. sn (x) increases to f (x) for each x. there exist sets Fn increasing up to X such that µ(Fn ) < ∞ for each n. For n suﬃciently large. (15. the supremum on the right hand side of (15. Since µ is σﬁnite. f p = sup f g dµ : g q ≤1 . The following is very useful. and sn p < ∞ for each n.8) When p = 1. let a be any ﬁnite real less than M .8) holds if we take q = 1. (15. Proof. whether or not f p is ﬁnite. . Then g is bounded by 1 and f g = f . CHAPTER 15. Since a is arbitrary. and if p = ∞.
11 Suppose 1 < p < ∞.4.15. That H ≤ g H¨lder’s inequality. Since (p − 1)q = p. Using Theorem 15. Theorem 15. On the other hand. and g ∈ Lq . Proposition 15. f p = sup fg : g q ≤ 1. The proof of Theorem 15. If we deﬁne H(f ) = f g for f ∈ Lp . and H is a realvalued bounded linear functional on Lp . BOUNDED LINEAR FUNCTIONALS gn is again a simple function. which tends to f p .9 For 1 < p < ∞ and p−1 + q −1 = 1. This proves the right hand side of (15.10 allows us to identify the dual space of Lp with Lq . This theorem together with Proposition 15. The linearity is obvious. g simple .8) is at least as large as the left hand side. p Since p − (p/q) = 1. p−1 + q −1 = 1. .10 Suppose 1 < p < ∞. Proof. then gn q 139 = ( sn (p−1)q )1/q sn p/q p = p/q p p/q sn p sn = 1. Then there exists g ∈ Lq such that H(f ) = f g and g q = H .8 also establishes Corollary 15. p−1 + q −1 = 1. since f  ≥ sn . then H is a bounded linear functional on Lp and H = g q . f gn = f  sn p−1 sn p/q p ≥ sn p sn p/q p = sn p−(p/q) . then f gn ≥ sn p .8 and writing o H = sup H(f ) = sup f p ≤1 q follows by f g ≥ sup f p ≤1 fg = g q f p ≤1 completes the proof.
9 and (15. Suppose we are given a bounded linear functional H on Lp . If A and B are disjoint. We conclude that ν is a countably additive signed measure. If s = H(s) = i i ai χAi is a simple function. while by H¨lder’s inequality o sn g − so g fg = (sn − f )g ≤ sn − f p g q → 0. and so ν(An ) = H(χAn ) → H(χA ) = ν(A). (15. we see there exists a realvalued integrable function g such that ν(A) = A g for all sets A. and then use Exercise 3. by linearity we have ai H(χAi ) = i ai ν(Ai ) = i ai gχAi = gs. Moreover. which is the RadonNikodym theorem for signed measures. Clearly each Hn is a bounded linear functional on Lp . that ν µ and that g = dν/dµ is the function we seek. s simple} ≤ H . then H(sn ) → H(f ). hence ν(A) = H(χA ) = 0. We thus have H(f ) = f g for all f ∈ Lp . Applying the above argument. To show ν is countably additive. First suppose µ(X) < ∞. Deﬁne functionals Hn by Hn (f ) = H(f χFn ).1. but not necessarily ﬁnite.1. It is easy to see . then ν(A ∪ B) = H(χA∪B ) = H(χA + χB ) = H(χA ) + H(χB ) = ν(A) + ν(B). if µ(A) = 0. we use here the fact that µ(X) < ∞. Deﬁne ν(A) = H(χA ).9). g q = sup gs : s p p ≤ 1.e. In the case where µ is σﬁnite. = sup{H(s) : s If sn are simple functions tending to f in Lp (see Exercise 15. then ν(An ) → ν(A).9) By Corollary 15. We will show that ν is a measure. But if An ↑ A. let Fn ↑ X so that µ(Fn ) < ∞ for each n. H ≤ g q . By H¨lder’s inequality.. Using Exercise 13. s simple ≤ 1. then χAn → χA in Lp .140 CHAPTER 15. we see there exist gn such that Hn (f ) = f gn and gn q = Hn ≤ H . it suﬃces to show that if An ↑ A. sn g → f g. and o q ≤ H . LP SPACES Proof.2). then χA = 0 a.
if n > m. we have Hn (f ) = H(f χFn ) → H(f ). Prove that f p → f ∞ as p → ∞.5 When does equality hold in H¨lder’s inequality? o When does equality hold in the Minkowski inequality? Exercise 15.2 Prove that the collection of simple functions is dense in Lp . Exercise 15. 1]. Again by H¨lder’s o 15. Give an example to show that Lq ⊂ Lp in general if 1 < p < q < ∞. Thus H(f ) = inequality H ≤ g q . Exercise 15.1 Show that L∞ is complete. Moreover.5.4 Consider the measure space ([0. Exercise 15. g is in Lq with a norm bounded by H . Deﬁne g by setting g(x) = gn (x) if x ∈ Fn . . and suppose f is a measurable function. f g.15.3 Prove the equality ∞ f (x)p dx = 0 ptp−1 m({x : f (x) ≥ t}) dt for p ≥ 1. On the other hand Hn (f ) = Fn f gn = Fn fg → fg by dominated convergence.5 Exercises Exercise 15. Then g is well deﬁned. by the uniqueness part of the / RadonNikodym theorem. Since H is a bounded linear functional on Lp . m). Note f χFn → f in Lp by dominated convergence.6 Give an example to show that Lp ⊂ Lq in general if 1 < p < q < ∞. where B is the Borel σalgebra and m is Lebesgue measure. EXERCISES 141 that gn is 0 if x ∈ Fn . B. then gn = gm on Fm . By Fatou’s lemma. Exercise 15.
(1) Show that if f ∈ L2 ([0. . 1 0 f (x)gn (x) dx Exercise 15. prove that f ∗ g p ≤ f 1 g p. Exercise 15. LP SPACES gn (x) = nχ[0.e. Prove that lim f p = exp log f  dµ . Exercise 15. then 1 f (x)gn (x) dx → 0 0 as n → ∞. Prove that if g ∈ Lq .e. ∞) and q is its conjugate exponent. whenever f is realvalued.1).11 If f ∈ L1 (R) and g ∈ Lp (R) for some p ∈ [1. Exercise 15. give a counterexample.10 Suppose 1 < p < ∞ and q is the conjugate exponent to p. 1]). p→0 where we use the convention that exp(−∞) = 0. and supn fn p < ∞. and integrable. then n→∞ lim fn g = f g. Suppose fn → f a..7 Deﬁne CHAPTER 15.n−3 ] (x).9 Suppose µ is a measure with µ(X) = 1 and f ∈ Lr for some r > 0.142 Exercise 15. (2) Show that there exists f ∈ L1 ([0. Prove that µ({0}) = 1. Does this extend to the case where p = 1 and q = ∞? If not.8 Suppose µ is a ﬁnite measure on the Borel subsets of R such that f (x) = R f (x + t) µ(dt). then f ∗g is uniformly continuous and f ∗ g(x) → 0 as x → ∞ and as x → −∞. 1]) such that → 0.12 Suppose p ∈ (1. ∞). Exercise 15. where we deﬁne Lr for r < 1 exactly as in (15. Prove that if f ∈ Lp (R) and g ∈ Lq (R). bounded. a.
Does this extend to the cases where p = 1 or p = ∞? If not. y) ν(dy) X Y µ(dx) f (x. ∞) → R. ∞)). Let g(x) = f (x)/x. ∞)). ∞). and 1 < p < ∞. Prove that f ∈ Lp (µ) if and only if ∞ (2n )p µ({x : f (x) > 2n }) < ∞. If Y = {1.14 Suppose f ∈ L∞ (R). then f ∗ g is continuous with compact support. A.3. f exists and is continuous and bounded.18 Find an example of a measurable f : [1. we recover the usual Minkowski inequality. then p 1/p f (x. where δ1 and δ2 are point masses at 1 and 2. Show g ∈ L2 ([1. give counterexamples. n=1 Exercise 15. Exercise 15. 2}. 2).e. Proposition 15. 1). Exercise 15. .13 Show that if f and g are continuous with compact support. Prove that f dµ g dµ ≥ 1. Exercise 15. f (1) = 0. µ) and (Y. ∞)).15.5.16 Suppose µ(X) = 1 and f and g are nonnegative functions such that f g ≥ 1 a. f is measurable with respect to A × B.17 Suppose f : [1. and f ∈ L2 ([1. Exercise 15. ∞) → R such that f (1) = 0. EXERCISES 143 Exercise 15. fh (x) = f (x + h). but the function g(x) = f (x)/x is not in L1 . f exists and is continuous and bounded. Prove that there exists a uniformly continuous function g on R such that f = g a. and h→0 lim fh − f ∞ = 0. resp. and we let g1 (x) = f (x. B.15 Let p ∈ [1. g2 (x) = f (x. ν) are measure spaces.e. y)p µ(dx) 1/p ≤ Y X ν(dy). f ∈ L1 ([1.19 Prove the generalized Minkowski inequality: If (X. Exercise 15. ν(dy) = δ1 (dy)+δ2 (dy).
where m is Lebesgue measure. where m is Lebesgue measure. and f ∈ Lp with respect to Lebesgue measure. Deﬁne g(x) = Prove that g p 1 x x f (y) dy. and f (0) = 0. Prove there exists p > 1 such that f ∈ Lp . 1] with f ∈ Lp .21 Suppose p > 1 and q is its conjugate exponent. Exercise 15. 1) such that f (x) dx ≤ cm(A)α A for every Borel measurable set A ⊂ R. 1) such that f (x) dx ≤ cm(A)α A for every Borel measurable set A ⊂ R.20 Let α ∈ (0. Prove that if f is nonnegative. LP SPACES Exercise 15. and integrable on R and g(x) = then g is ﬁnite a. Note that K is not in Lp for any p ≥ 1.e.23 Suppose f : R → R is integrable and there exist constants c > 0 and α ∈ (0. then 1 f (x − t)K(t) dt. ∞) → R. 1) and K(x) = x−α for x ∈ R. . 0 ≤ p f p−1 p. This is known as Hardy’s inequality. realvalued.24 Suppose 1 < p < ∞. Exercise 15. f : (0. f g dx ≤ 0 1 p 1/p f p g q. f is an absolutely continuous function on [0.144 CHAPTER 15. Exercise 15.22 Suppose f : R → R is in Lp for some p > 1 and also in L1 . Prove there exist constants c > 0 and α ∈ (0. Exercise 15. Prove that if g ∈ Lq .
y) µ(dy) ≤ M X for each x and K(x. Exercise 15.5.26 Suppose A and B are two Borel measurable subsets of R with ﬁnite strictly positive Lebesgue measure. (1) Show that T f 1 ≤ M f 1 .28 Suppose 1 < p < ∞ and q is the conjugate exponent of p. deﬁne T f (x) = X K(x. .27 Suppose A and B are two Borel measurable subsets of R with strictly positive Lebesgue measure. Show that χA ∗χB is a continuous nonnegative function that is not identically equal to 0. then there exists a complexvalued measurable function g ∈ Lq such that H(f ) = f g for all f ∈ Lp and H = g q.25 Suppose (X. EXERCISES 145 Exercise 15. Suppose there exists M < ∞ such that K(x. µ) is a measure space and suppose K : X × X → R is measurable with respect to A × A. show that T f p ≤M f p. Prove that if H is a bounded complexvalued linear functional on Lp . (2) If 1 < p < ∞. Exercise 15. If f is measurable and realvalued. Exercise 15. Show that C = {x + y : x ∈ A. y) µ(dx) ≤ M X for each y. y ∈ B} contains a nonempty open interval. A. y)f (y) µ(dy) if the integral exists.15.
146 CHAPTER 15. LP SPACES .
deﬁne the Fourier transform f to be the function with domain Rn and range C given by f (u) = Rn eiu·x f (x) dx. (4) if a ∈ Rn and fa (x) = f (x + a).1 Basic properties If f is a complexvalued function and f ∈ L1 (Rn ). There is a great deal known about them and their applications. 16.1) We are using u · x for the standard inner product in Rn . (2) (f + g)(u) = f (u) + g(u). Various books have slightly diﬀerent deﬁnitions.1 Suppose f and g are in L1 . We give an introduction here.Chapter 16 Fourier transforms Fourier transforms give a representation of a function in terms of frequencies. Then (1) f is bounded and continuous. (3) (af )(u) = af (u) if a ∈ C. some have a (2π)−1 or a (2π)−1/2 in front of the integral. Some put a negative sign and/or 2π before the iu · x. The basic theory is the same in any case. Some basic properties of the Fourier transform are given by Proposition 16. then fa (u) = e−iu·a f (u). (16. u ∈ Rn . 147 .
Finally for (6). Proof. (4) holds because fa (u) = eiu·x f (x + a) dx = eiu·(x−a) f (x) dx = e−iu·a f (u) by a change of variables. Thus the continuity follows by dominated convergence. ga (u) = eiu·x eia·x f (x) dx = ei(u+a)·x f (x) dx = f (u + a). One reason for the usefulness of Fourier transforms is that they relate derivatives and multiplication. We have f (u + h) − f (u) = Then f (u + h) − f (u) ≤ eiu·x · eih·x − 1 f (x) dx. ei(u+h)·x − eiu·x f (x) dx. (6) if a is a nonzero real number and ha (x) = f (ax). then ha (u) = a−n f (u/a).2 Suppose f ∈ L1 and xj f (x) ∈ L1 . which is integrable. then ga (u) = f (u + a). FOURIER TRANSFORMS (5) if a ∈ Rn and ga (x) = eia·x g(x). ha (u) = eiu·x f (ax) dx = a−n eiu·(y/a) f (y) dy = a−n as required. Then ∂f (u) = i ∂uj eiu·x xj f (x) dx. where xj is the j th coordinate of x. (2) and (3) are easy by a change of variables. ei(u/a)·y f (y) dy = a−n f (u/a). Proposition 16. The integrand is bounded by 2f (x). . and eih·x − 1 → 0 as h → 0. by a change of variables. For (5).148 CHAPTER 16. (1) f is bounded because f ∈ L1 and eiu·x  = 1.
the limit of the left hand side is ∂ f /∂uj . Then the Fourier transform of f is −iuf (u). Since f is integrable. This implies that f (yn ) is a Cauchy sequence whenever yn → ∞. Proof. the right hand side converges to eiu·x ixj f (x) dx by dominated convergence. Of course. y → ∞ by dominated convergence. (16. By integration by parts (use Exercise 14.3 Suppose f : R → R is integrable. as desired.4.6). BASIC PROPERTIES Proof. Proposition 16. Recall also (15. Let ej be the unit vector in the j th direction. The higher dimensional version of this is left as Exercise 16. and f is integrable. Therefore the left hand side converges. y f (y) − f (x) ≤ x f (z) dz → 0 as x.3.1 and a limit argument).16. ∞ ∞ f (u) = −∞ eiux f (x) dx = − −∞ iueiux f (x) dx = −iuf (u). The same is true for the limit as y → −∞.2) . Since f is integrable.1. which says that f (x − y) g(y) dx dy = f 1 g 1. h 149 1 ihxj e − 1 ≤ xj  h and xj f (x) ∈ L1 . and we conclude that f (y) converges as y → ∞. f is absolutely continuous. the only possible value for the limit is 0. Recall the deﬁnition of convolution given in Section 15. Then 1 f (u + hej ) − f (u) = h h = Since ei(u+hej )·x − eiu·x f (x) dx eiu·x eihxj − 1 f (x) dx.
150
CHAPTER 16. FOURIER TRANSFORMS
Proposition 16.4 If f, g ∈ L1 , then the Fourier transform of f ∗g is f (u)g(u). Proof. We have f ∗ g(u) = = = eiu·x f (x − y)g(y) dy dx
eiu·(x−y) f (x − y) eiu·y g(y) dx dy f (u)eiu·y g(y) dy = f (u)g(u).
We applied the Fubini theorem in the second equality; this is valid because as we see from (16.2), the absolute value of the integrand is integrable. We used a change of variables to obtain the third equality.
16.2
The inversion theorem
We want to give a formula for recovering f from f . First we need to calculate the Fourier transform of a particular function. Proposition 16.5 (1) Suppose f1 : R → R is deﬁned by
2 1 f1 (x) = √ e−x /2 . 2π
Then f1 (u) = e−u
2
/2
.
(2) Suppose fn : Rn → R is given by fn (x) = Then fn (u) = e−u
2 2 1 e−x /2 . (2π)n/2
/2
.
Proof. (1) may also be proved using contour integration, but let’s 2 give a (mostly) real variable proof. Let g(u) = eiux e−x /2 dx. Diﬀerentiate with respect to u. We may diﬀerentiate under the integral sign because (ei(u+h)x − eiux )/h is bounded in absolute value
16.2. THE INVERSION THEOREM
2
151
by x and xe−x /2 is integrable; therefore dominated convergence applies. We then obtain g (u) = i eiux xe−x
2
/2
dx.
By integration by parts (see Exercise 14.1) this is equal to −u eiux e−x
2
/2
dx = −ug(u).
Solving the diﬀerential equation g (u) = −ug(u), we have [log g(u)] = g (u) = −u, g(u)
so log g(u) = −u2 /2 + c1 , and then g(u) = c2 e−u
2
/2
.
(16.3)
√ √ 2 By Exercise 11.18, g(0) = e−x /2 dx = 2π, so c2 = 2π. Sub√ stituting this value of c2 in (16.3) and dividing both sides by 2π proves (1). For (2), since fn (x) = f1 (x1 ) · · · f1 (xn ) if x = (x1 , . . . , xn ), then fn (u) = ··· ei
j
uj xj
f1 (x1 ) · · · f1 (xn ) dx1 · · · dxn
2
= f1 (u1 ) · · · f1 (un ) = e−u This completes the proof.
/2
.
One more preliminary is needed before proving the inversion theorem. Proposition 16.6 Suppose ϕ is in L1 and ϕδ (x) = δ −n ϕ(x/δ). ϕ(x) dx = 1. Let
(1) If g is continuous with compact support, then g∗ϕδ converges to g pointwise as δ → 0. (2) If g is continuous with compact support, then g∗ϕδ converges to g in L1 as δ → 0. (3) If f ∈ L1 , then f ∗ ϕδ − f
1
→ 0 as δ → 0.
152
CHAPTER 16. FOURIER TRANSFORMS
Proof. (1) We have by a change of variables (Exercise 8.1) that ϕδ (y) dy = 1. Then g ∗ ϕδ (x) − g(x) = = ≤ (g(x − y) − g(x)) ϕδ (y) dy (g(x − δy) − g(x)) ϕ(y) dy g(x − δy) − g(x) ϕ(y) dy.
Since g is continuous with compact support and hence bounded and ϕ is integrable, the right hand side goes to zero by dominated convergence, the dominating function being 2 g ∞ ϕ. (2) We now use the Fubini theorem to write g ∗ ϕδ (x) − g(x) dx = = ≤ = Let Gδ (y) = g(x − δy) − g(x) dx. (g(x − y) − g(x)) ϕδ (y) dy dx (g(x − δy) − g(x)) ϕ(y) dy dx g(x − δy) − g(x) ϕ(y) dy dx g(x − δy) − g(x) dx ϕ(y) dy.
By dominated convergence, for each y, Gδ (y) tends to 0 as δ → 0, since g is continuous with compact support. Moreover Gδ is bounded in absolute value by 2 g 1 . Using dominated convergence again and the fact that ϕ is integrable, we see that Gδ (y) ϕ(y) dy tends to 0 as δ → 0. (3) Let ε > 0. Let g be a continuous function with compact support so that f − g 1 < ε. Let h = f − g. A change of variables shows that ϕδ 1 = ϕ 1 . Observe f ∗ ϕδ − f Also h ∗ ϕδ − h
1 1
≤ g ∗ ϕδ − g
1
+ h ∗ ϕδ − h 1 .
≤ h
1+
h ∗ ϕδ
1
≤ h
1+
h
1
ϕδ
1
< ε(1 + ϕ 1 ).
16.2. THE INVERSION THEOREM Therefore, using (2), lim sup f ∗ ϕδ − f
δ→0 1
153
≤ lim sup h ∗ ϕδ − h
δ→0
1
≤ ε(1 + ϕ 1 ).
Since ε is arbitrary, we have our conclusion. Now we are ready to give the inversion formula. The proof seems longer than one might expect it to be, but there is no avoiding the introduction of the function Ha or some similar function. Theorem 16.7 Suppose f and f are both in L1 . Then f (y) = 1 (2π)n e−iu·y f (u) du, a.e.
Proof. If g(x) = a−n k(x/a), then the Fourier transform of g is k(au). Hence the Fourier transform of
2 2 1 1 e−x /2a n (2π)n/2 a 2
is e−a
u2 /2
. If we let Ha (x) =
2 2 1 e−x /2a , n (2π)
we have Ha (u) = (2π)−n/2 an e−a We write f (u)e−iu·y Ha (u) du = = = eiu·x f (x)e−iu·y Ha (u) dx du eiu·(x−y) Ha (u) du f (x) dx Ha (x − y)f (x) dx. (16.4)
2
u2 /2
.
We can interchange the order of integration because f (x) Ha (u) dx du < ∞
4) converges to (2π)−n f (u)e−iu·y dy as a → ∞ by dominated convergence since Ha (u) → (2π)−n and f ∈ L1 .3 The Plancherel theorem The last topic that we consider is the Plancherel theorem. Then f ∈ L2 and f 2 = (2π)−n/2 f 2 . 16. g(u) = = eiu·x f (−x) dx = e−iu·x f (−x) dx eiu·x f (x) dx = f (u).154 CHAPTER 16. (16. we see that f ∗ Ha converges to f in L1 as a → ∞. the left hand side of (16.8) converges by monotone convergence to (2π)−n f (u)2 du as a → . The left hand side of the ﬁrst line of (16. (16.8 Suppose f is continuous with compact support. By (16. Theorem 16.6) Proof.4) is equal to Ha (y − x)f (x) dx = f ∗ Ha (y). take y = 0.7) with f replaced by f ∗ g. First note that if we combine (16. Since ab = ab.6. (16.8) Since f ∗ g(u) = f (u)g(u) = f (u)2 .7) Let g(x) = f (−x). and use the symmetry of Ha . setting δ = a−1 .5) using that Ha is symmetric. where a denotes the complex conjugate of a.5). f ∗ g(u)Ha (u) du = f ∗ g ∗ Ha (0).4) and (16. But by Proposition 16. FOURIER TRANSFORMS and eiu·x  = 1. we obtain f (u)Ha (u) du = f ∗ Ha (0). (16. The last line of (16. The third equality follows by a change of variables.
f ∗ g is also.6(2). then f (y) < f (0) for y = 0. Thus f is deﬁned uniquely up to almost everywhere equivalence.6) will continue to hold. EXERCISES 155 ∞. and all the partial derivatives fj = ∂f /∂xj are integrable. and therefore fm has the same limit as fm . 16.2 Find a realvalued function f ∈ L1 such that f ∈ / L1 . By (16. Given a function f in L2 . The set of continuous functions with compact support is dense in L2 by Proposition 15.b] and in particular.6) holds for f ∈ L2 .8 to deﬁne f when f ∈ L2 so that (16. .1 Find the Fourier transform of χ[a. Exercise 16. then by Exercise 15.4. fm − fm converges to 0 in L2 .4 If f is integrable. By (16. and so the right hand side of (16.6).6). we see that (16. Then fm −fn 2 → 0 as m. prove that the Fourier transform of fj is given by fj (u) = −iuj f (u).16. which we call f . If {fm } is another sequence of continuous functions with compact support converging to f in L2 .13. By passing to the limit in L2 on both sides of (16.n] . choose a sequence {fm } of continuous functions with compact support such that fm → f in L2 . {fm } is a Cauchy sequence in L2 . Exercise 16.5. Let us check that the limit does not depend on the choice of the sequence. Remark 16.6). realvalued. ﬁnd the Fourier transform of χ[−n.9 We can use Theorem 16. Since f and g are continuous with compact support.4 Exercises Exercise 16. n → ∞. then {fm − fm } is a sequence of continuous functions with compact support converging to 0 in L2 . and therefore converges to a function in L2 .8) converges to f ∗ g(0) = f (y)g(−y) dy = f (y)2 dy by Proposition 16.3 Show that if f ∈ L1 and f is everywhere strictly positive. Exercise 16.
FOURIER TRANSFORMS Exercise 16.8 Prove Heisenberg’s inequality (which is very useful in quantum mechanics): there exists c > 0 such that if a. then µ = ν. The collection S is called the Schwartz class. (u−b)2 f (u)2 du ≥ c f (x)2 dx 2 .156 CHAPTER 16.7 If f is realvalued and continuously diﬀerentiable on R. Exercise 16. then f ∈ S. prove that f 2 dx 2 ≤4 xf (x)2 dx f 2 dx .6 The Fourier transform of a ﬁnite signed measure µ on Rn is deﬁned by µ(u) = eiu·x µ(dx). . where f (k) is the k th derivative of f when k ≥ 1 and f (0) = f . then (x−a)2 f (x)2 dx Find the best constant c. Exercise 16. Prove that if µ and ν are two ﬁnite signed measures on Rn (with respect to the completion of L × L. where L is the Lebesgue σalgebra on R) such that µ(u) = ν(u) for all u ∈ Rn . Prove that if f ∈ S. Exercise 16. xm f (k) (x) → 0 as x → ∞.5 Let S be the class of realvalued functions f on R such that for every k ≥ 0 and m ≥ 0. b ∈ R and f is in L2 .
and if f ≥ 0. See Remark 17. We let C(X) be the collection of continuous functions from X to R. with almost no changes in the proof. The main topic of this chapter is to prove a converse. throughout this chapter we suppose X is a compact metric space. Recall that the support of a function f is the closure of {x : f (x) = 0}. we could even let X be a locally compact Hausdorﬀ metric space. Throughout this chapter we will restrict our attention to realvalued functions. We need more hypotheses on X than just that it is a topological space. If G is an 157 . and with only relatively minor changes. then L(f ) ≥ 0. We write supp (f ) for the support of f . If X is a topological space. let B be the Borel σalgebra and suppose µ is a σﬁnite measure on (X. we could let X be a compact Hausdorﬀ space.Chapter 17 Riesz representation In Chapter 4 we constructed measures on R. Clearly L is linear. B).1. In this chapter we will discuss how to construct measures on more general topological spaces X. If f is continuous on X. let us deﬁne L(f ) = X f dµ. In fact. the Riesz representation theorem. But here we stick to compact metric spaces. For simplicity.
K) = inf{d(x. if X = [−2. 2. The collection {gi } is called a partition of unity on K. y ∈ X. there exist disjoint open sets Gx and Gy with x ∈ Gx and y ∈ Gy . If we let f (x) = 1 − d(x. We will need the following proposition. where the Gi are open sets. x = y. Urysohn’s lemma is the result from topology that guarantees such an f exists. then 0 ≤ f ≤ χG . then there exists f ∈ FG such that f is 1 on K.1 If X is a compact Hausdorﬀ space instead of a compact metric one. Proposition 17. 1]R with the product topology. .1 Partitions of unity The reason we take our set X to be a metric space is that if K ⊂ G ⊂ X. we deﬁne FG by FG = {f ∈ C(X) : 0 ≤ f ≤ 1.) See [1] for details. .2 Suppose K is compact and K ⊂ G1 ∪ · · · ∪ Gn . y) : y ∈ K} is the distance from x to K and δ = inf{d(x. but the converse does not hold. Let x ∈ K. n n such that i=1 gi (x) = 1 if x ∈ K. Then x will be in at least one Gi . so there exists hx ∈ FGi such that hx (x) = 1. There exist gi ∈ FGi for i = 1. we can still ﬁnd such an f . is not contained in G. Remark 17. and G is open. y ∈ Gc }. 17. RIESZ REPRESENTATION open subset of X. and f (x) = (1 − x2 )+ . subordinate to the cover {Gi }. where d(x. which is [−1.158 CHAPTER 17. (A Hausdorﬀ space X is one where if x. y) : x ∈ K. 2]. 1). that is. An example of a compact Hausdorﬀ space that is not a metric space and cannot be made into a metric space is [0. but the support of f . Proof. K) δ/2 + . For example. Single points are always compact. . then this f will do the job. G = (−1. 1]. Observe that if f ∈ FG . where K is compact and G is open. . supp (f ) ⊂ G}. f ∈ FG with f ≥ χK when K ⊂ G. . K is compact. then 0 ≤ f ≤ χG .
The collection {Nx } is an open cover for the compact set K. then x ∈ Nxj for some j.3 Let X be a compact metric space and L a positive linear functional on C(X). Then n fi (x) = 1. and Nx ⊂ Gi . B) such that L(f ) = f (y) µ(dy). Note g1 + g2 = 1 − (1 − f1 )(1 − f2 ). . We have Fi ⊂ Gi . .2. the smallest σalgebra that contains all the open subsets of X. f ∈ C(X). ··· gn = (1 − f1 )(1 − f2 ) · · · (1 − fn−1 )fn . THE REPRESENTATION THEOREM 159 Let Nx = {y : hx (y) > 0}. Since hx is continuous. (17. . 17. and an induction argument shows that g1 + · · · + gn = 1 − (1 − f1 )(1 − f2 ) · · · (1 − fn ). Each Fi is closed. x ∈ Nx . g ∈ C(X) and a ∈ R. which implies k=1 gk (x) = 1. Let us choose fi ∈ FGi such that fi is 1 on Fi .2 The representation theorem Let L be a linear functional mapping C(X) to R.17. so x ∈ Fi for some i. Now deﬁne g1 = f1 . then Nx is open. . and since X is compact.1) . Theorem 17. For each i. g2 = (1 − f1 )f2 . Then there exists a measure µ on (X. Here is the Riesz representation theorem. Nxm }. B is the Borel σalgebra on X. so there exists a ﬁnite subcover {Nx1 . Fi is compact. L is a positive linear functional if L(f ) ≥ 0 whenever f ≥ 0 on X. that is. Clearly gi ∈ FGi . If x ∈ K. let Fi = ∪{Nxj : Nxj ⊂ Gi }. Thus L(f + g) = L(f ) + L(g) and L(af ) = aL(f ) if f.
If G is open. Then K is compact. Step 2 is to show that every open set is µ∗ measurable. Let {gi } be a partition of i=1 unity for K subordinate to {Gi }n . and we conclude that µ∗ is an outer measure. ﬁrst let G1 . To show the countable subadditivity of µ∗ . and therefore there exists n such that K ⊂ ∪n Gi . so (∅) = 0. let µ∗ (E) = inf{ (G) : E ⊂ G. G2 . RIESZ REPRESENTATION We often write Lf for L(f ). taking f identically equal to 1 in (17. A2 . ∞ µ∗ (G) = (G) ≤ i=1 µ∗ (Gi ). Therefore n n ∞ Lf = i=1 L(f gi ) ≤ i=1 µ∗ (Gi ) ≤ i=1 µ∗ (Gi ). Since X is compact. If A1 . . Step 1 of the proof will be to show µ∗ is an outer measure. Since K is the support of f . i=1 n we have f = i=1 f gi . . The only function in F∅ is the zero function.1) shows that µ is a ﬁnite measure. Step 3 is to apply Theorem 4. . Taking the supremum over f ∈ FG . . Let K be the support of f . Since gi ∈ FGi and f is bounded by 1. Since ε is arbitrary. Clearly µ∗ (A) ≤ µ∗ (B) if A ⊂ B. and therefore µ∗ (∅) = 0.160 CHAPTER 17. Step 1. For any open set H we see that µ∗ (H) = (H). We show µ∗ is an outer measure. . Proof. Let G = ∪i Gi and let f be any element of FG . Step 4 establishes some regularity of µ and Step 5 shows that (17. are subsets of X. be open sets. G open}. let (G) = sup{Lf : f ∈ FG } and for E ⊂ X. and choose Gi open such that (Gi ) ≤ µ∗ (Gi ) + ε2−i . {Gi } is an open cover for K.1) holds. Then ∞ ∞ µ∗ (∪∞ Ai ) ≤ µ∗ (∪∞ Gi ) ≤ i=1 i=1 i=1 µ∗ (Gi ) ≤ i=1 µ∗ (Ai ) + ε. then f gi ∈ FGi . . .6 to obtain a measure µ. countable subadditivity is proved. let ε > 0.
We show that every open set is µ∗ measurable. and f ≥ χK . Since ε is arbitrary. This is true for all g ∈ FG . Let ε > 0 and deﬁne G = {x : f (x) > 1 − ε}. let ε > 0 and choose H open such that E ⊂ H and (H) ≤ µ∗ (E) + ε. Because L is a positive linear functional. Suppose G is open and E ⊂ X. It suﬃces to show µ∗ (E) ≥ µ∗ (E ∩ G) + µ∗ (E ∩ Gc ). which we call µ. which is open. we can choose g ∈ FE∩K c such that L(g) > (E ∩ K c ) − ε/2. In particular. f ∈ C(X). Since ε is arbitrary. Since K c is open. then L(f ) ≥ µ(K). Choose f ∈ FE∩G such that L(f ) > (E ∩ G) − ε/2. if G is open. µ(G) = µ∗ (G) = (G). the restriction of µ∗ to B. which leads to Lg ≤ Lf /(1 − ε). THE REPRESENTATION THEOREM 161 Step 2.2) holds when E is open. 1−ε Since ε is arbitrary. Let K be the support of f . Then f + g ∈ FE . Let B be the Borel σalgebra on X.2. and (E) ≥ L(f + g) = Lf + Lg ≥ (E ∩ G) + (E ∩ K c ) − ε = µ∗ (E ∩ G) + µ∗ (E ∩ K c ) − ε ≥ µ∗ (E ∩ G) + µ∗ (E ∩ Gc ) − ε. is a measure on B. (17.2) since the opposite inequality is true by the countable subadditivity of µ∗ . (17. By Theorem 4. hence Lf µ(K) ≤ µ(G) ≤ . First suppose E is open. If g ∈ FG . In this step we show that if K is compact. then g ≤ χG ≤ f /(1 − ε). µ(K) ≤ Lf .2) holds.17. L((1−ε)−1 f −g) ≥ 0. Step 4. so (1 − ε)−1 f − g ≥ 0. If E ⊂ X is not necessarily open. . Then µ∗ (E) + ε ≥ (H) = µ∗ (H) ≥ µ∗ (H ∩ G) + µ∗ (H ∩ Gc ) ≥ µ∗ (E ∩ G) + µ∗ (E ∩ Gc ).6. (17. Step 3.
Let n ≥ 1 and let Ki = {x : f (x) ≥ i/n}. i=0 (17. Then nfi ∈ FG . Since X is compact. K0 is all of X. i=0 (17. b].4) we see that L(f ) − f dµ ≤ µ(K0 ) − µ(Kn ) µ(X) ≤ .4 If f is continuous on [a. hence compact. then (17. RIESZ REPRESENTATION Step 5. and multiplying by a constant and using linearity.162 CHAPTER 17. b]. Note f = n i=1 fi and χKi ≤ nfi ≤ χKi−1 . By writing f = f + − f − and using the linearity of L. then f is bounded. and hence 1 n n µ(Ki ) ≤ L(f ) ≤ i=1 1 n n−1 µ(Ki ). x ∈ Ki−1 − Ki .1). to show (17. Since f is continuous. the measure whose existence is given by the Riesz representation theorem is Lebesgue measure. we may suppose 0 ≤ f ≤ 1. By Step 4. Since ε is arbitrary.4) Comparing (17.1) for continuous functions we may suppose f ≥ 0. L(nfi ) ≥ µ(Ki ). In this case.1) is established. b]). Example 17. x ∈ Ki−1 . Deﬁne c 0. i−1 fi (x) = f (x) − n .3) and (17. n n Since.3) Let ε > 0 and let G be an open set containing Ki−1 such that µ(G) < µ(Ki−1 ) + ε. We now establish (17. n 1 n n−1 and so 1 n n µ(Ki ) ≤ i=1 f dµ ≤ µ(Ki ). Then L is a positive linear functional on C([a. 1 x ∈ Ki . let L(f ) be the Riemann integral of f on the interval [a. L(fi ) ≤ µ(Ki−1 )/n. as we saw above. Therefore µ(Ki ) ≤ n fi dµ ≤ µ(Ki−1 ) . n. so L(nfi ) ≤ µ(G) ≤ µ(Ki−1 ) + ε. . each Ki is a closed set. µ(X) = L(1) < ∞ and n is arbitrary.
Then Gc ⊂ E c ⊂ K c . REGULARITY 163 Remark 17. contains ∪∞ Ei . Proposition 17. E2 . ∈ H. µ(G − E) < ε. See [1] for details. . . and µ(G−E) < ε. and i=1 i=1 ∞ µ(∪i Gi − ∪i Ei ) ≤ i=1 µ(Gi − Ei ) < ε.6 Suppose X is a compact measure space.17. G is open. and µ(E c − Gc ) = µ(G − E) < ε. Then ∪∞ Gi is open. K c is open. and µ is a ﬁnite measure on the measurable space (X. B is the Borel σalgebra.) Proof. If K is compact. K) < 1/n}. G open. . C0 (X) is / the usual notation for the set of continuous functions vanishing at inﬁnity. and if n is large enough. . Then the Gn are open sets decreasing to K. hence compact. Let H be the collection of approximable subsets. and µ(E −F ) < ε. let Gn = {x : d(x. and µ(Ei −Ki ) < ε2−(i+1) . G open. 17. For each i choose Ki compact and Gi open such that Ki ⊂ Ei ⊂ Gi . µ(G − E) < ε. Therefore H is closed under the operation of taking complements. If E ∈ B and ε > 0. then choose K ⊂ E ⊂ G with K compact. We will show H contains all the compact sets and H is a σalgebra. Let us say that a subset E ∈ B is approximable if given ε > 0 there exists K ⊂ E ⊂ G with K compact. µ(K c − E c ) = µ(E − K) < ε. and thus establish the proposition. B). µ(Gn −K) < ε.3 Regularity We establish the following regularity property of measures on compact metric spaces. A continuous function f vanishes at inﬁnity if given ε > 0 there exists a compact set K such that f (x) < ε if x ∈ K. If E is in H and ε > 0. There is a version of the Riesz representation theorem for C0 (X). and µ(E − K) < ε.3. (K and G depend on ε as well as on E. Thus every compact set is in H.5 Let X be a metric space. Suppose E1 . Gc is closed. µ(E −K) < ε. there exists K ⊂ E ⊂ G such that K is compact. not necessarily compact. which will prove that H = B. µ(Gi −Ei ) < ε2−i .
Prove that if µ is the measure whose existence is given by the Riesz representation theorem. A measure is called regular if µ(E) = inf{µ(G) : G open. is compact. then H is also closed under the operation of taking countable intersections. RIESZ REPRESENTATION We see that ∪∞ Ki is contained in ∪∞ Ei and similarly. 1]. i=1 i=1 ∞ µ(∪∞ Ei − ∪∞ Ki ) ≤ i=1 i=1 i=1 µ(Ei − Ki ) < ε/2. being the ﬁnite union of compact sets. i=1 i=1 This proves that ∪i Ei is in H.4 Exercises Exercise 17. Since ∪n Ki increases to ∪∞ Ki . K ⊂ E} for all measurable E.1 Suppose F is a closed subset of [0. 17. 1] and we deﬁne 1 L(f ) = 0 f χF dx for realvalued continuous functions f on [0. B) are regular when X is a compact metric space. Then ∪n Ki . An immediate consequence of what we just proved is that ﬁnite measures on (X.164 CHAPTER 17. . Therefore H is a σalgebra. and i=1 µ(∪∞ Ei − ∪n Ki ) < ε. we can choose n large so that i=1 i=1 µ(∪∞ i=n+1 Ki ) < ε/2. i=1 is contained in ∪∞ Ei . then µ(A) = m(A ∩ F ). where m is Lebesgue measure. Since H is closed under the operations of taking complements c and countable unions and ∩i Ei = (∪i Ei )c . E ⊂ G} and µ(E) = sup{µ(K) : K compact.
1]) such that L(f ) ≤ c1 f + c2 f for all f ∈ C 1 ([0. 1]) be the set of functions whose derivative exists and is continuous on [0. B) and let µ be another ﬁnite measure on (X.3 Let C 1 ([0. prove that there exists a measure µ on the Borel sets of X such that f dν = Y X f ◦ F dµ for all f that are continuous on Y . where ∂A = A − Ao is the boundary of A. .6 Let X be a compact metric space and let B be the Borel σalgebra on X. Show there exists a signed measure µ on the Borel subsets of [0. where c1 and c2 are positive constants and the norm is the supremum norm. (2) lim supn→∞ µn (F ) ≤ µ(F ) for all closed subsets F of X. Exercise 17. EXERCISES 165 Exercise 17.2 Suppose X is a compact metric space and µ is a ﬁnite regular measure on (X. Exercise 17. Prove that if f is a realvalued measurable function and ε > 0. If ν is a ﬁnite measure on the Borel sets of Y . Exercise 17. B). Prove that C(X) has a countable dense subset. 1]). Suppose µn (X) → µ(X).5 Let X be a compact metric space. f ∈ C 1 ([0. (4) limn→∞ µn (A) = µ(A) whenever A is a Borel subset of X such that µ(∂A) = 0. 1]. (3) lim inf n→∞ µn (G) ≥ µ(G) for all open subsets G of X.17. Prove that the following are equivalent: (1) f dµn → f dµ whenever f is a continuous realvalued function on X. where B is the Borel σalgebra. there exists a closed set F such that µ(F c ) < ε and the restriction of f to F is a continuous function on F . Exercise 17.4. Suppose L is a linear functional on C 1 ([0. B). 1] and a constant K such that L(f ) = f dµ + Kf (0). Let µn be a sequence of ﬁnite measures on (X.4 Suppose X and Y are compact metric spaces and F : X → Y is a continuous map from X onto Y . 1]).
(3) With {nj } as in (2). B) and suppose supn µn (X) < ∞. RIESZ REPRESENTATION Exercise 17. Conclude that there exists a measure µ such that f dµnj → for all f ∈ C(X). prove that f dµnj converges for all f ∈ C(X). f dµ . there is a subsequence {nj } such that f dµnj converges.7 Let X be a compact metric space and let B be the Borel σalgebra on X.166 CHAPTER 17. (4) Let L(f ) = limnj →∞ f dµnj . (2) Let A be a countable dense subset of C(X). (1) Prove that if f ∈ C(X). Let µn be a sequence of ﬁnite measures on (X. Prove that L(f ) is a positive linear functional on C(X). Prove that there is a subsequence {nj } such that f dµnj converges for all f ∈ A.
was given in Chapter 1. where every Cauchy sequence converges. Recall that a normed linear space is a metric space if we use the metric d(x. A linear map is a map L from a normed linear space X to a normed linear space Y satisfying L(x + y) = L(x) + L(y) for all x.1 Deﬁnitions The deﬁnition of normed linear space X over a ﬁeld of scalars F . and derive some consequences such as the uniform boundedness theorem and the open mapping theorem. Since L(0) = L(0 + 0) = L(0) + L(0). 18. prove the Baire category theorem. We will give the deﬁnitions. We will sometimes write Lx for L(x). discuss the existence of bounded linear functionals.Chapter 18 Banach spaces Banach spaces are normed linear spaces that are complete. y ∈ X and L(αx) = αL(x) for all x ∈ X and α ∈ F . y) = x − y . then L(0) = 0.1 We deﬁne a Banach space to be a normed linear space that is complete. where F is either the real numbers or the complex numbers. Deﬁnition 18. 167 . that is.
To show (3) implies (1). (3) The linear functional f is continuous at 0. then Y has a maximal element. Proof. there exists a sequence xn ∈ X such that xn = 1 for each n. If we have a set Y with a partial order “≤. a linear map from X to C a complex linear functional.4 If Y is a partially ordered set and every linearly ordered subset of Y has an upper bound. That (2) implies (3) is obvious. if f is not bounded. BANACH SPACES Deﬁnition 18.168 CHAPTER 18. Now we give the HahnBanach theorem for real linear functionals. but f (xn ) → ∞. (2) The linear functional f is continuous. . If we let yn = xn /f (xn ). An element z of Y is maximal if z ≤ y for y ∈ Y implies y = z.2 A linear map f from X to R is a real linear functional. A linearly ordered subset X ⊂ Y has an upper bound if there exists an element z of Y (but it is not necessary that z ∈ X) such that x ≤ z for all x ∈ X. x ≤ 1} < ∞. 18. y ∈ X. but ﬁrst we need Zorn’s lemma.3 The following are equivalent. contradicting (3). then yn → 0 but f (xn ) = 1 → 0. f is a bounded linear functional if f = sup{f (x) : x ∈ X. so (1) implies (2). f (x) − f (y) = f (x − y) ≤ f x − y. (1) The linear functional f is bounded. then either x ≤ y or y ≤ x (or both) holds. Lemma 18. which is equivalent to the axiom of choice.” a linear ordered subset X ⊂ Y is one such that if x. Here is Zorn’s lemma.2 The HahnBanach theorem We want to prove that there are plenty of linear functionals. Proposition 18.
Thus f (x) − α ≤ x − x0 . Let F be the collection of all linear extensions F of f satisfying . Saying that F is an extension of f means that the domain of F contains the domain of f and F (x) = f (x) if x is in the domain of f.18. f (x) − x − x0 ≤ α. If f = 0. Hence f (x) − x − x0 ≤ f (y) + y − x0 for all x. We ﬁrst show that we can extend f by at least one dimension. Replacing x by −x/λ and multiplying by λ. or f (x) − α ≥ − x − x0 . Thus M1 consists of all vectors of the form x + λx0 . This is clearly an extension of f to M1 .2. and α ≤ f (x) + x − x0 . so we may assume that f = 0. By our choice of α. we get λ  − f (x)/λ − α ≤ λ − x/λ − x0 . where x ∈ X and λ is real. or f1 (x + λx0 ) = f (x) + λα ≤ x + λx0 . y ∈ M f (x) − f (y) = f (x − y) ≤ x − y ≤ x − x0 + y − x0 .5 If M is a subspace of a normed linear space X and f is a bounded real linear functional on M . or f (x) − α ≤ x − x0 . Deﬁne f1 (x + λx0 ) = f (x) + λα. We now establish the existence of an extension of f to all of X. Choose x0 ∈ X − M and let M1 be the vector space spanned by M and x0 . Let x ∈ M and λ ∈ R. which is what we wanted to prove. Choose α ∈ R such that f (x) − x − x0 ≤ α ≤ f (y) + y − x0 for all x. We have for all x. then we take F to be identically 0. and then by multiplying by a constant. then f can be extended to a bounded linear functional F on X such that F = f . y ∈ M . We need to verify that the norm of f1 is less than or equal to 1. Proof. THE HAHNBANACH THEOREM 169 Theorem 18. that f = 1. y ∈ M .
Another application is to ﬁx x0 = 0. Thus there exists a linear functional f such that f (x0 ) = x0 and f = 1. we can deﬁne f (x + λx0 ) = λx0 for x ∈ M . then the real part of f . let f (λx0 ) = λ x0 . u = Re f . we say f1 ≤ f2 if M1 ⊂ M2 .6 If M is a subspace of a normed linear space X and f is a bounded complex linear functional on M . F has a maximal element. if the domain of F1 is not all of X. so F ≤ 1. and hence f (x) = u(x) − iu(ix). Let F (x) = U (x) − iU (ix). F (x) = U (eiθ x) ≤ U This holds for all x. Also. given a subspace M and an element x0 not in M . It only remains to show that the norm of F is at most 1. Fix x. is a real valued linear functional. This collection is partially ordered by inclusion. To get a version for complex valued linear functionals is quite easy. Now use the version of the HahnBanach theorem for real linear functionals to ﬁnd a linear functional U that is an extension of u to X such that U ≤ 1. As an application of the HahnBanach theorem. That is. Since the union of any increasing family of subspaces of X is again a subspace. Note that if f (x) = u(x) + iv(x). By Zorn’s lemma. Assume without loss of generality that f = 1. e−iθ x ≤ x . and write F (x) = reiθ . Theorem 18. Then f will be 0 on M but nonzero at x0 . which would be a contradiction to F1 being maximal. Note u(x) ≤ f (x) ≤ x . Then F (x) = r = e−iθ F (x) = F (e−iθ x). namely. then f can be extended to a bounded linear functional F on X such that F = f . so that v(x) = −u(ix). Let u = Re f . u(ix) = Re f (ix) = Re if (x) = −v(x). . then the union of a linearly ordered subfamily of F lies in F. By the construction of the preceding two paragraphs. BANACH SPACES F ≤ 1. if f1 is an extension of f to a subspace M1 and f2 is an extension of f to a subspace M2 . and then extend f to all of X. F1 . Proof. we can ﬁnd an extension.170 CHAPTER 18. and then extend this linear functional to all of X. Therefore F1 is the desired extension. say. Since this quantity is real and nonnegative.
A set A is dense in X if A = X and A is nowhere dense if (A)o = ∅. . then ∩n Gn is dense in X. (2) X cannot be written as the countable union of nowhere dense sets. G2 . r1 ) ⊂ H ∩ G1 and 0 < r1 < 1. and then o c o Fn = ∅ and X = ∪n Fn . Also. and since X is complete. . which is open. then Gn = X. H ∩G1 is nonempty and open. y) < r}. n > N . so there exists xn and rn such that B(xn . X = ∪n En where (En )o = ∅.7 Let X be a complete metric space. Since Fn = ∅. we see that ∅ = ∩n Gn . rn ) ⊂ Gn ∩ B(xn−1 . we use A for the closure of A and Ao for the interior of A. Since Gn is dense. BAIRE’S THEOREM AND CONSEQUENCES 171 18. and hence in each GN . and we can ﬁnd x1 and r1 such that B(x1 . Suppose we have chosen xn−1 and rn−1 for some n ≥ 2.3. Let Gn = Fn . It remains to show that x ∈ H ∩ (∩n Gn ). If m. that is. We must prove (1). We continue and get a sequence xn in X. Therefore x ∈ ∩n Gn . xn ) < 2rN < 2−N +1 . then x lies in each B(xN . Let H be any nonempty open set in X. . rn−1 ) and 0 < rn < 2−n . rn−1 ) ⊂ · · · ⊂ B(x1 . x ∈ B(xn .3 Baire’s theorem and consequences We turn now to the Baire category theorem and some of its consequences. Since xn lies in B(xN . rN ) if n > N . We ﬁrst show that (1) implies (2). r1 ) ⊂ H. are open and dense in X. r) = {y ∈ X : d(z. then xm and xm both lie on B(xN . . rn ) ⊂ B(xn−1 . (1) If Gn are open sets dense in X. Proof. a contradiction to (1). where d is the metric. Theorem 18. which is a closed set. then Gn ∩ B(xn−1 . Suppose we can write X as a countable union of nowhere dense sets. rn−1 ) is open and nonempty. Since G1 is dense in X. The Baire category theorem is the following. We need to show there exists a point in H ∩ (∩n Gn ). We will construct a certain Cauchy sequence {xn } and the limit point. will be the point we seek. Therefore xn is a Cauchy sequence. rN ). and so d(xm . We let Fn = En . Starting with X = ∪n Fn and taking complements.18. Suppose G1 . rN ). Completeness of the metric space is crucial to the proof. xn converges to a point x ∈ X. x. Recall that if A is a set. Let B(z.
also called the uniform boundedness theorem. r) ∩ GN = ∅. We conclude that if y ≤ r and α ∈ A. we let L(A) = {Lx : x ∈ A}. The map x → Lα x is a continuous function for each α since Lα is a bounded linear functional. Theorem 18. This can be rephrased as saying that if x − x0 ≤ r. is that every Gn is dense in X. hence is open. We argue that Gn is open. Let A be an index set and let {Lα : α ∈ A} be a collection of bounded linear maps from X into Y . If y ≤ r. we have y = (x0 + y) − x0 .8 Suppose X is a Banach space and Y is a normed linear space. supα Lα ≤ M with M = 2N/r. x0 − x0 = 0 ≤ r. Since x ∈ Gn if and only if for some α ∈ A we have Lα x > n. An important application of the Baire category theorem is the BanachSteinhaus theorem. Then either there exists a M < ∞ such that Lα ≤ M for all α ∈ A or else supα Lα x = ∞ for some x. For a measurable set A. Also. But (x) = ∞ for every x ∈ ∩n Gn . then Lα (x) ≤ N for all α ∈ A. The other possibility. Lα y = Lα ((x0 + y) − x0 ) ≤ Lα (x0 + y) + Lα x0 ≤ 2N. we conclude Gn is the union of open sets. and hence Lα (x0 + y) ≤ N for all α. A set A ⊂ X is called meager or of the ﬁrst category if it is the countable union of nowhere dense sets. by the Baire category theorem. Let Gn = {x : (x) > n}. Consequently. It is important that L be onto. otherwise it is of the second category. Proof. Then there exists x0 and r such that B(x0 . Suppose there exists N such that GN is not dense in X. and in this case ∩n Gn is dense in X. BANACH SPACES Thus we have found a point x in H ∩ (∩n Gn ).172 CHAPTER 18. Let (x) = supα∈A Lα x . Then (x0 + y) − x0 = y ≤ r. and thus Lα (x0 ) ≤ N for all α. The following theorem is called the open mapping theorem. . A mapping L : X → Y is open if L(U ) is open in Y whenever U is open in X. This implies that for each α. of course. the set {x : Lα x > n} is open.
BAIRE’S THEOREM AND CONSEQUENCES 173 Theorem 18. 4r) ⊂ L(B(0. Since z1 ∈ B(0. then L(B(x. 1) be such that y1 = Lz1 . 1)) ⊂ L(B(0. r2−n ) ⊂ L(B(0. We will show L(B(0. Pick y1 ∈ L(B(0. 1)). L(B(0. Let ε = r2−(n+1) . it suﬃces to show that L(B(0. if y < r.3. It follows by linearity that if y < r2−n . We need to show that if B(x. Step 2. n=1 The Baire category theorem tells us that at least one of the sets L(B(0. then −z1 + B(0. Suppose we have chosen x1 . n)) cannot be nowhere dense. then y ∈ L(B(0. 2). By linearity. . r)) contains a ball in Y . r)) contains a ball centered at 0 in Y . r) ⊂ X. This can be rephrased as saying that if y < r2−n and ε > 0. r)) will contain a ball centered at Lx in Y . xn−1 such that n−1 y− j=1 Lxj < r2−n . Y = ∪∞ L(B(0. 1)).9 Let X and Y be Banach spaces. . 2r). Since L is onto. then y ∈ L(B(0. to show that L(B(0.18. Thus if y < 2r. . then y + y1 ∈ B(y1 . 4r) ⊂ L(B(0. By Step 1. . then there exists x such that x < 2−n and y − Lx < ε. 2)). we can ﬁnd x1 ∈ B(0. We show that there exists r such that B(0. . By the linearity of L. n)). 1)) such that y1 − y0 < 2r and let z1 ∈ B(0. 1)) contains a ball centered at 0 in Y . 2r) ⊂ B(y0 . We will construct a sequence {xj } by ∞ induction such that y = L( j=1 xj ). Suppose y < r/2. 1)). r)) contains a ball centered at 0. A bounded linear map L from X onto Y is open. we can ﬁnd xn such that xn < 2−n and n n−1 y− j=1 Lxj = y− j=1 Lxj − Lxn < r2−(n+1) . 1)) cannot be nowhere dense. 1)). L(B(x. Since L is linear. 1). 2−n )) for each n. and so y = −Lz1 + (y + y1 ) ∈ L(−z1 + B(0. 1/2) such that y − Lx1 < r/4. Step 1. 1) ⊂ B(0. Then B(y1 . 2−n )). Proof. hence y ∈ L(−z1 + B(0. Then using the linearity of L. By Step 1 with ε = r/4. Thus there exist y0 and r such that B(y0 .
Since X is complete. then wn is a Cauchy sequence. But then x < ∞ −j = 1. Let C k ([0.4 Exercises Exercise 18.x=y f (x) − f (y) .5 For positive integers n let 1 An = f ∈ L1 ([0. where f (k) is the k th derivative of f . 1]) be the set of functions f with f C α < ∞. 1)). and a bounded linear functional f on Y with norm 1 such that f has two distinct extensions to L1 (µ) and each of the extensions has norm equal to 1.1 Find a measure space (X. 1]) complete with respect to the norm · C k ? Exercise 18. 1]) complete with respect to the norm · C α ? Exercise 18. 1]) is separable.174 CHAPTER 18. to x. A. that is. 1]) : 0 f (x)2 dx ≤ n . y = Lx. Show that L∞ ([0. . 1).1]. 1]) is not separable. r/2). Exercise 18. deﬁne f Ck = f ∞ + f ∞ + · · · + f (k) ∞.4 Let α ∈ (0. 18. then y ∈ L(B(0. 1]) with empty interior. 1] → R that are k times diﬀerentiable.1] sup x. Let n wn = j=1 xj . wn converges.3 For k ≥ 1 and functions f : [0. Is C k (0. For f a realvalued continuous function on [0. if j=1 2 y ∈ B(0. That is. µ). x − yα Let C ([0. say. Exercise 18. 1] deﬁne f α Cα = sup f (x) + x∈[0.y∈[0. if 1 ≤ p < ∞. and since L is continuous. Show that each An is a closed subset of L1 ([0. a subspace Y of L1 (µ). Is C α ([0. 1]) be the collection of k times diﬀerentiable functions f with f C k < ∞. there is a countable dense subset. BANACH SPACES We continue by induction to construct the sequence {xj }.2 Show that Lp ([0. Since xj < 2−j .
Exercise 18. (1) Let (X. Exercise 18. (2) Prove that the open unit ball in a normed linear space is convex. A.8 The unit ball in a normed linear space X is strictly convex if λx + (1 − λ)y < 1 whenever f = g = 1.7 A set A in a normed linear space is convex if λx + (1 − λ)y ∈ A whenever x. and C(X) are not strictly convex provided X consists of more than one point. and λ ∈ (0. and suppose X is complete with respect to both norms.6 Suppose L is a linear functional on a normed linear space X. Exercise 18. 1). f = g. y) = x + y .10 Suppose · 1 and · 2 are two norms such that x 1 ≤ x 2 for all x in a vector space X. (The open unit ball is the set of x such that x < 1. Prove there exist an interval and a number M such that supn fn  is bounded by M on that interval. y ∈ A and λ ∈ [0.11 Suppose X and Y are Banach spaces. (1) Let X × Y be the set of ordered pairs (x.) Exercise 18. (1) Prove that if A is convex. y1 + y2 ) = (x1 . EXERCISES 175 Exercise 18. y2 ∈ Y and c(x. x2 ∈ X and y1 . 1]. µ) be a measure space. y) = (cx. Prove that there exists a positive constant c such that x for all x ∈ X.4. y2 ) for each x1 . Exercise 18. y) with (x1 + x2 . L∞ (µ). Prove that the unit ball in Lp (µ) is strictly convex. (2) Prove that the unit balls in L1 (µ). Prove that X × Y is a Banach space. cy) if x ∈ R. (2) Let L be a linear map from X into Y such that if xn → x in 2 ≤c x 1 . Prove that L is a bounded linear functional if and only if the set {x ∈ X : L(x) = 0} is closed. then the closure of A is convex.18. Deﬁne (x.9 Let fn be a sequence of continuous functions on R that converge at every point. y1 ) + (x2 .
1]}. and maps G onto X. Prove that A is a closed convex subset of C([0. but there does not exist f ∈ A such that f = inf g . Lx) → x is continuous. then L is a continuous map. 1]). (3) Prove that the function (x. 1] such that 1/2 1 f (x) dx − 0 1/2 f (x) dx = 1. BANACH SPACES X and Lxn → y in Y . . Deﬁne D : X → Y by Df = f . 1] such that f (x) − f (y) ≤ nx − y for all y ∈ [0. 1]). Prove that G is a closed subset of X × Y . 1]. then y = Lx.176 CHAPTER 18. g∈A Exercise 18. which says that if L is a linear map from one Banach space to another that is a closed map. Such a map is called a closed map. hence is complete. Exercise 18. oneone.12 Let X = C 1 ([0.14 Let An be the subset of the realvalued continuous functions on [0. 1]) (deﬁned in Exercise 18. y) : y = Lx}. f (x) does not exist at any point of [0. 1]). 1] given by An = {f : there exists x ∈ [0. Exercise 18. deﬁned by G = {(x. 1]) which are nowhere diﬀerentiable on [0.13 Let A be the set of realvalued continuous functions on [0. 1]. Let G be the graph of L. linear.3) and Y = C([0. (4) Prove the closed graph theorem. (1) Prove that An is nowhere dense in C([0. that is. (2) Prove that there exist functions f in C([0. Show that D is a closed map but not a bounded one.
z for all x. y . We deﬁne x = x.1 Let H be a vector space where the set of scalars F is either the real numbers or the complex numbers. When a is real. y . y for all x.Chapter 19 Hilbert spaces Hilbert spaces are complete normed linear spaces that have an inner product. (5) x. y = α x. z = x. αy = α x. As an application we brieﬂy discuss Fourier series. z + y. so that x. for x. z ∈ H. The proof is the same as the one usually taught in undergraduate linear algebra 177 1/2 .1 Inner products Recall that if a is a complex number. H is an inner product space if there is a map ·. x = 0 if and only if x = 0. y ∈ H and α ∈ F . x . a is just a itself. (2) x + y. · from H × H to F such that (1) y. y = 0 and x. (3) αx. 19. x = x. y. We will give the deﬁnitions and basic properties. Deﬁnition 19. x ≥ 0 for all x ∈ H. x = x 2 . y ∈ H. (4) x. This added structure allows one to talk about orthonormal sets. From the deﬁnitions it follows easily that 0. The following is the CauchySchwarz inequality. then a represents the complex conjugate.
Proof. x + y = x. let α = eiθ . y = 0. x + y. x − rαy = x. y = Reiθ . We have for real r 0 ≤ x − rαy 2 = x − rαy.178 CHAPTER 19. the inequality is obvious. HILBERT SPACES classes. From the CauchySchwarz inequality we get the triangle inequality: Proposition 19. except for some complications due to the fact that we allow the set of scalars to be the complex numbers. Proof. y ∈ H we have x+y ≤ x + y . If x. and then α = 1 and α y. then y = 0. y  ≤ x y . If C = 0. y also equals  x. y + r2 y. Taking square roots of both sides gives the inequality we wanted. we have  x. Therefore we will suppose that C > 0 and B = 0. x + x. y ≤ x 2 +2 x y + y 2 = ( x + y )2 . x − rα x. x =  x. . Theorem 19. y . we may take r = B/C. y + y. x − rα y. we have that α x. hence x.3 For all x. Since B is real.2 For all x. Since we are supposing that C > 0. We write x+y 2 2 − 2r x. and the inequality holds. y  + r2 y 2 . = x + y. and C = y 2 . B =  x. y . y ∈ H. y = x Therefore A − 2Br + Cr2 ≥ 0 for all real numbers r. and we obtain B 2 ≤ AC. Let A = x 2 . y  = B. If B = 0.
If we let µ be counting measure on the natural numbers. By the CauchySchwarz inequality. Proposition 19. . or x − x ≤ x−x .). Proof. b = n=1 an bn . .) such that n=1 an 2 < ∞ and if b = (b1 .1. . Then the functions x → x. by letting µ be counting measure on {1.19. and thus the function x → x is continuous. we get what is known as the space 2 . This is easily seen to be a Hilbert space.  x. which proves that the function x → x. y  ≤ x − x y . and so if we deﬁne the distance between x and y by x − y . let H = L2 (µ). . Deﬁnition 19. . we have a metric space. . 2. a2 . . y and x → x are continuous.5 Let µ be a positive measure on a set X. The triangle inequality implies x−z ≤ x−y + y−z . g = f g dµ. then ∞ a. y  =  x − x . x ≤ x − x + x . By the triangle inequality. . n}. and deﬁne f. so  x − x ≤ x−x . An element of 2 is a sequence ∞ a = (a1 .6 Let y ∈ H be ﬁxed. y) = x − y .4 A Hilbert space H is an inner product space that is complete with respect to the metric d(x. . INNER PRODUCTS as desired. y − x . The same holds with x and x reversed. b2 . ndimensional Euclidean space. Example 19. y is continuous. . . To show the completeness we use Theorem 15.4. We get another common Hilbert space. 179 Therefore · is a norm on H.
.1) Choose yn ∈ E such that yn → δ.7 A subset M of a vector space is a subspace if M is itself a vector space with respect to the same operations of addition and scalar multiplication. For an example of a subspace that is not closed. and we conclude M is not closed because / xn − x as n → ∞. . . A closed subspace is a subspace that is closed relative to the metric given by ·.) and x = (1. HILBERT SPACES 19. y ∈ E. Let δ = inf{ x : x ∈ E}. . . n . . · . M is clearly a subspace. x + y and similarly for x − y 2 . x 2 . Let xn = 1 (1. if x. .2 Subspaces Deﬁnition 19. . 0. Applying (19. .180 CHAPTER 19. Proposition 19. then 1 4 x−y 2 = 1 2 x 2 + 1 2 y 2 − x+y 2 2 . Since E is convex. a simple calculation shows that x+y 2 2 = 1 →0 j2 j=n+1 ∞ + x−y 2 =2 x 2 + y 2. 2 2 3 x ∈ M . Dividing by 4.8 Each nonempty closed convex subset E of H has a unique element of smallest norm. 1 . and we have x−y 2 ≤2 x 2 +2 y 2 − 4δ 2 . we see that yn − ym 2 ≤ 2 yn 2 + 2 ym 2 − 4δ 2 . if x. 1 . Then each xn ∈ M . (19.1) with x replaced by yn and y replaced by ym . . 1 . consider 2 and let M be the collection of sequences for which all but ﬁnitely many elements are zero. Since x + y 2 = x + y.). y ∈ E. 0. A set E ⊂ H is convex if λx + (1 − λx) ∈ E whenever 0 ≤ λ ≤ 1 and x. y ∈ E. Proof. then (x + y)/2 ∈ E. . and y 2 .
” be the set of all y in X that are orthogonal to x. if x. y = lim yn = δ. read “x perp. For bounded linear functionals on H. and the y of the next theorem is the vector associated with the transpose of A. y . z 2 = z. it converges to some y ∈ H. We say x ⊥ y. Let x⊥ . It is clear from the linearity of the inner product that x⊥ is a subspace of H. The following is sometimes called the Riesz representation theorem.2. To motivate the theorem.3. consider the case where H is ndimensional Euclidean space. let M = {x : Lx = 0}. so A is 1 × n. it is easy to see that M ⊥ is a subspace. y . we take y = 0.1) with x replaced by y we have y − y = 0. M ⊥ is closed. Ly = Lz Lz = Lz2 / z.9 If L is a bounded linear functional on H. or x is orthogonal to y. Elements of Rn can be identiﬁed with n × 1 matrices and linear maps from Rn to Rm can be represented by multiplication on the left by a m × n matrix A.19. If Lx = x. and hence y = y . although usually that name is reserved for Theorem 17. y = x.6. z = 0. y ∈ E. If M is a subspace. then by (19. Since the norm is a continuous function. z . Also. and in particular. which is continuous by Proposition 19. Proof. then x. The subspace x⊥ is closed because it is the same as the set f −1 ({0}). Hence yn is a Cauchy sequence. Otherwise. Since yn ∈ E and E is closed. If y is another point with y = δ. then z so z = 0. z = y. z. and since H is complete. y = 0. where f (x) = x. SUBSPACES 181 and the right hand side tends to 0 as m and n tend to inﬁnity. y . then there exists a unique y ∈ H such that Lx = x. and let y = αz where α = Lz/ z. y . . let M ⊥ be the set of all y that are orthogonal to all points in M . take z = 0 in M ⊥ . and since M ⊥ = ∩x∈M x⊥ . If Lx = 0 for all x. Notice y ∈ M ⊥ . when x = y − y . We make the observation that if z ∈ M ∩ M ⊥ . m = 1. We now prove existence. The uniqueness is easy. Theorem 19. y − y = 0 for all x.
uα 2 ≤ x 2 . The GramSchmidt procedure from linear algebra also works in inﬁnitely many dimensions.3 Orthonormal sets A subset {uα }α∈A of H is orthonormal if uα = 1 for all α and uα . . . β ∈ A and α = β. Let F be a ﬁnite subset of A. uα uα . y = x − w.  x. 19. Proof. y = Lx as desired. then for each x ∈ H. . If x ∈ H and CHAPTER 19.e. HILBERT SPACES w =x− Lx y. u N = vN / vN . Let y= α∈F x. ui = 0 if i < N . no ﬁnite linear combination of the xn is 0. Let u1 = x1 / x1 and deﬁne inductively n−1 v N = xN − i=1 xN . uβ = 0 whenever α. .10 If {uα }α∈A is an orthonormal set. Then x.2) α∈A This is called Bessel’s inequality. uN are orthonormal. Suppose {xn }∞ is a linearly inden=1 pendent sequence. y then Lw = 0. so u1 . i. y = 0. so w ∈ M . and hence w. hence only countably many of the summands can be nonzero. .2) can be larger than 1/n for each n. Proposition 19.. This inequality implies that only ﬁnitely many of the summands on the left hand side of (19. y.182 and y = 0. We have vN . (19. ui ui .
uα 2 . ORTHONORMAL SETS Then 0≤ x−y Now y. j=1 which is what we wanted. . x = α∈F x. uα 2 . uβ uα . x. uαj 2 = lim J→∞  x. Proposition 19. y = y.  x.11 Suppose {uα }α∈A is orthonormal. x = α∈F  x. α∈F = where we used the fact that {uα } is an orthonormal set and so uα . uα uα . uαj 2 ≤ x 2 .19. x . then x. (2) x 2 = α∈A  x. x = α∈A x.β∈F x. (1) If x. uβ uβ = α. uα uα . Hence  x. β∈F x. . . uα 2 > 1/n for more than N of the α. uβ  x. (3) For each x ∈ H. Also y 2 = y. Therefore 0≤ y−x Rearranging. uα 2 ≤ x α∈F 2 2 = x 2 − α∈F  x. uα 2 for all x. uα 2 = 0 for only countably many α. y − y. If N is an integer larger than n x 2 . y = α∈F x. Since this is real. x + y 2 . when F is a ﬁnite subset of A. uα uα . it is not possible that  x. uα 2 . x = α∈F 2 183 = x 2 − x. uα x. . uα 2 = α∈A j=1  x.3. then x = 0. uβ equals 0 if α = β and equals 1 if α = β. α2 .. uα = 0 for each α ∈ A. Label those α’s as α1 . Then ∞ J  x. uα uα . Then the following are equivalent.
k=m n x. this implies z − x = 0. 0. This means that (3) in Proposition 19. . . . uαj uαj is a Cauchy sequence. uαj x.10 there can be at most countably many α such that  x.13 Every Hilbert space has an orthonormal basis. When (1) holds. . Then z − x. be an enumeration of those α. x2 . . By Bessel’s inequality. By (1). uαj uαj . Let α1 . . uαj = 0 for each αj .e. 1.. n 2 n x. Thus j=1 x. Then {ei } is a complete orthonormal system. . HILBERT SPACES We make a few remarks. uαk  x. we say the orthonormal set is complete. . Proposition 19. . Let z = j=1 x. uαj uαj j=m = j. We see that (3) implies (2) because n n x 2 − j=1  x. (2) is called Parseval’s identity. uαj uαj → 0. y = i xi yi . uα 2 = 0. and ∞ hence converges. Proof. A collection of elements {eα } is a basis for H if the set of ﬁnite linear combinations of the eα is dense in H.184 CHAPTER 19. 0. where ei = (0. uαj 2 = x − j=1 2 x.). In (3) the convergence is with respect to the norm of H and implies that only countably many of the terms on the right hand side are nonzero. n → ∞. . Let x ∈ H.11 holds. Using that {uα } is an orthonormal set. α2 . That (2) implies (1) is clear. i. 0. First we show (1) implies (3).) : xi 2 < ∞} with x. . uαj 2 → 0 j=m = n as m. uαi 2 converges. . uαk uαj . By Bessel’s inequality and the remarks following the statement of Proposition 19. the series i  x.12 Take H = 2 = {x = (x1 . Example 19. the only nonzero coordinate of ei is the ith one. . .
(n can be negative. then B is a basis that is strictly bigger than B.19. or equivalently. 2π)) and let 1 un = √ einx 2π for n an integer. Choose x ∈ V ⊥ . this maximal orthonormal set must be a basis. The ﬁrst step is to show that the closure of F with respect to the supremum norm is equal to the set of continuous functions f on [0. FOURIER SERIES 185 Proof.4. and if we let B = B ∪ {x/ x }. the closure with respect to the norm in H of the set of ﬁnite linear combinations of elements of B. It is easy to see that {un } is an orthonormal set: 2π 2π einx e−imx dx = 0 0 ei(n−m)x dx = 0 if n = m and equals 2π if n = m. By the preceding paragraph. 2π)). ∞) that take the same value at 0 and 2π with the continuous functions on the .) Recall that 2π f. For our Hilbert space we take H = L2 ([0. If B = {uα } is orthonormal. We will accomplish this by using the StoneWeierstrass theorem. to trigonometric series. We want to show that F is a dense subset of L2 ([0.4 Fourier series An interesting application of Hilbert space techniques is to Fourier series.7. g = 0 f (x)g(x) dx and f 2 = 2π 0 f (x)2 dx. 2π) with f (0) = f (2π). Theorem 1. Let F be the set of ﬁnite linear combinations of the un . and so there is a maximal one by Zorn’s lemma. We identify the set of continuous functions on [0. let V be the closure of the linear span of B. for otherwise we could ﬁnd a larger basis. It is easy to see that the union of an increasing sequence of orthonormal sets is an orthonormal set. that is. but not a basis. 19.
so u1 (θ1 ) = ei(θ1 −θ2 ) = 1. we can ﬁnd fn ∈ F tending to f in L2 . If θ1 . Since the complex conjugate of un is u−n . if g is continuous on [0. By the StoneWeierstrass theorem (Theorem 1. the closure of F with respect to the supremum norm is equal to the set of continuous complexvalued functions on S. If f ∈ L2 ([0. By what we showed above. then θ1 − θ2 is not an integer multiple of 2π. then gm → g in L2 ([0. If f is continuous on [0. 2π) with f (0) = f (2π). Therefore F separates points. The ﬁrst term on the right of the inequality sign is bounded by f − fn f . a continuous function with support in [1/m. The second term on the right of the inequality sign is 0. Note un (eiθ ) = einθ . then F is closed under the operation of taking complex conjugates. 2π − 1/m]) can be approximated in L2 by continuous functions which have support in the interval [1/m. HILBERT SPACES circle. let S = {eiθ : 0 ≤ θ < 2π} be the unit circle in C. 2π)). 2π − 1/m] can be approximated uniformly on [0. It remains to show the completeness of the un . u0 is identically equal to 1. so F vanishes at no point. u1 (θ2 ) or u1 (θ1 ) = u1 (θ2 ). Then f 2 =  f. it follows that F is closed under the operation of multiplication. deﬁne f : S → C by f (eiθ ) = f (θ). 2π). 2π) by elements of F.6 any function in L2 ([1/m.7). f .186 CHAPTER 19. then f − f χ[1/m. Putting all this together proves that F is dense in L2 ([0. to every element of F. f  +  fn . 2π)). Since F is dense in L2 ([0. θ2 ∈ S and θ1 = θ2 . then it is orthogonal to every ﬁnite linear combination. f  ≤  f − fn . Finally. By Corollary 15.2π−1/m] 2 → 0 by dominated convergence as m → ∞. 2π)). If f is orthogonal to each un . 2π − 1/m]. 2π) and gm → g uniformly on [0. S is a compact metric space. 2π)) by dominated convergence. Let F be the set of ﬁnite linear combinations of the un . To do this. That it is closed under scalar multiplication and addition is obvious. that is. Since un ·um = un+m .
or f = 0. 0 the Fourier coeﬃcients of f . Let H = C([0. 1]. hence the {un } are complete. Parseval’s identity says that f 2 = n cn 2 . Is H a Hilbert space with respect to the norm deﬁned in terms of the inner product ·. write 2π 2π cn = f. and Cn = i(cn − c−n ). For any f in L we also have cn un → f n≤N as N → ∞ in the sense that f− n≤N cn un →0 2 as N → ∞. 2π)).5. Thus results involving the un can be transferred to results for series of sines and cosines and vice versa. and this tends to 0 as n → ∞. g = 0 f (x)g(x) dx. un = 0 1 f un dx = √ 2π 2 f (x)e−inx dx. Therefore {un } is a complete orthonormal system. . Conversely. where A0 = c0 . 1]). · ? Justify your answer. let f. 19.5 Exercises 1 Exercise 19. g ∈ L2 ([0. ∞ ∞ ∞ A0 + n=1 Bn cos nx + n=1 Cn sin nx = n=−∞ cn einx if we let c0 = A0 .1 For f. using cos nx = (einx + e−inx )/2 and sin nx = (einx − e−inx )/2i. Therefore f 2 = 0. Using einx = cos nx + i sin nx.19. cn = Bn /2 + Cn /2i for n > 0 and cn = Bn /2 − Cn /2i for n < 0. 1]) be the functions that are continuous on [0. EXERCISES 187 by the CauchySchwarz inequality. Given f in L2 ([0. we have ∞ ∞ ∞ cn einx = A0 + n=−∞ n=1 Bn cos nx + n=1 Cn sin nx. Bn = cn + c−n .
give a counterexample. Exercise 19. Exercise 19.9 that there exists a realvalued measurable function g in L2 (µ) such that L(f ) = f g dµ for all f ∈ L2 (µ).8 The purpose of Exercise 13. y for every y in H. HILBERT SPACES Exercise 19. n Exercise 19. that is. Exercise 19.3 Prove that if M is a closed subset of a Hilbert space H.188 CHAPTER 19. Assume for the current problem that this is the case and that µ and ν are ﬁnite measures. A n→∞ This is known as the RiemannLebesgue lemma. prove that lim einx dx = 0. then (M ⊥ )⊥ = M . deﬁne L(f ) = f dν. (2) Conclude by Theorem 19. Prove that if xn is a sequence in H with supn xn ≤ 1. Prove that xn − x → 0 as n → ∞. Is this necessarily true if M is not closed? If not. Exercise 19. Suppose xn → x as n → ∞ and xn . Prove that ∞ n=1 a2 < ∞. (1) Show that L is a bounded linear functional on L2 (µ). we can assume that ν(A) ≤ µ(A) for all measurable A. For f realvalued and in L2 with respect to µ. . Exercise 19.6 We say xn → x weakly if xn .2 Suppose H is a Hilbert space with a countable basis. it has no ﬁnite basis.4 Prove that if H is inﬁnitedimensional. y → x.4 was to show that in proving the RadonNikodym theorem. y → x. y as n → ∞ for every y ∈ H. then there is a subsequence {nj } and an element x of H with x ≤ 1 such that xnj converges to x weakly. Prove that dν = g dµ. then the closed unit ball in H is not compact. 2π].7 If A is a measurable subset of [0.5 Suppose an is a sequence of real numbers such that ∞ an bn < ∞ n=1 whenever ∞ 2 n=1 bn < ∞. We use this to give an alternative proof of the RadonNikodym theorem.
P is called the projection of x onto M .10 If M is a closed subspace of a Hilbert space. Prove that 1 n→∞ n lim n 1 f (jγ) = j=1 0 f (x) dx.5. then P x = x and Qx = 0. (1) Prove that x + M is a closed convex subset of H. Let γ be an irrational number. (4) Prove that if x ∈ M . (6) Prove that x 2 = P x 2 + Qx 2 . (3) Prove that P and Q are linear mappings. then P x = 0 and Qx = x. (5) Prove that if x ∈ M ⊥ . .19.9 Suppose f is a continuous realvalued function on R such that f (x + 1) = f (x) for every x. Exercise 19. EXERCISES 189 Exercise 19. Prove that P and Q are mappings of H into M and M ⊥ . respectively. (2) Let Qx be the point of x+M of smallest norm and P x = x−Qx. let x + M = {x + y : y ∈ M }.
190
CHAPTER 19. HILBERT SPACES
Bibliography
[1]G.B. Folland. Real Analysis: Modern Techniques and their Applications, 2nd ed. Wiley, New York, 1999. [2]H. Royden. Real Analysis, 3rd ed. Macmillan, New York, 1988. [3]W. Rudin. Principles of Mathematical Analysis, 3rd ed. McGrawHill, New York, 1976. [4]W. Rudin. Real and Complex Analysis, 3rd ed. McGrawHill, New York, 1987.
191
Index
Ac , 1 A − B, 1 A∆B, 1 Ai ↑, 1 Ai ↓, 1 Ao , 3 A, 3 B(x, r), 3 C, 2 C k ([0, 1]), 174 C α ([0, 1]), 174 C(X), 157 Lp , 131 Q, 2 R, 2 x ∨ y, 2 x ∧ y, 2 x+ , 2 x− , 2 z, 2 σ(C), 8 absolutely continuous function, 118 measures, 99 algebra, 7 almost everywhere, 39 antiderivative, 112 Baire category theorem, 171 Banach space, 167 BanachSteinhaus theorem, 172 basis, 184 Bessel’s inequality, 182 Borel σalgebra, 9 Borel measurable, 9 Borel sets, 9 bounded convergence theorem, 50 bounded variation, 117 Cantor function, 28 Cantor set, 28 generalized, 29 Carath´odory extension e theorem, 31 Carath´odory’s theorem, 22 e Cauchy sequence, 3 CauchySchwarz inequality, 177 characteristic function, 41 Chebyshev’s inequality, 76 closed, 3 closed graph theorem, 176 closed map, 176 closure, 3 complete measure space, 15 metric space, 3 orthonormal set, 184 completion, 15 conditional expectation, 105 conjugate exponent, 132 convergence almost everywhere, 75 almost uniform, 78 in Lp , 75
192
INDEX in measure, 75 convex function, 67, 180 set, 175 strictly, 175 convolution, 136 countable additivity, 13 counting measure, 14 covering lemma, 108 decreasing, 2 strictly, 2 density, 101 derivates, 122 derivative, 107 diﬀerentiable, 107 dominated convergence theorem, 55 Egorov’s theorem, 78 equivalence relationship, 4 equivalent measures, 104 essential inﬁmum, 134 essential supremum, 134 Fatou’s lemma, 54 ﬁnitely additive, 13 ﬁrst category, 172 Fourier series, 185 Fubini theorem, 85 FubiniTonelli theorem, 85 generates the σalgebra, 9 graph, 176 Hahn decomposition, 95 HahnBanach theorem, 168 Hardy’s inequality, 144 Heisenberg’s inequality, 156 Hilbert space, 179 increasing, 2 strictly, 2 indeﬁnite integral, 112 inner product space, 177 integrable, 48 integration by parts, 127 interior, 3
193
Jensen’s inequality, 67 jointly measurable, 86 Jordan decomposition theorem, 97 Lebesgue decomposition theorem, 103 Lebesgue measurable, 27 Lebesgue measure, 21, 27 ndimensional, 86 Lebesgue σalgebra, 21, 27 LebesgueStieltjes measure, 21, 27 liminf, 2 limsup, 2 linear functional, 137 bounded, 137, 168 complex, 168 positive, 159 real, 168 linear map, 167 linear space, 4 Lipschitz, 128 locally integrable, 109 lower semicontinuous, 73 maximal function, 109 meager, 172 measurable, 7 measurable rectangle, 81 measurable space, 7 measure, 13 complex, 104 ﬁnite, 15 regular, 164 signed, 93 measure space, 13
175 subspace. 4 Vitali convergence theorem. 14 points of density. 182 outer measure. 22 mutually singular. 109 weak convergence. 118 total variation measure. 143 monotone. 93 positive set. 188 Riesz representation theorem. 180 support. 13 Parseval’s identity. 154 point mass. 137 signed measure. 159. 172 open mapping theorem. 3 open mapping. 81 separable. 172 section. 1 total variation of a function. 184 partial order. 101 RiemannLebesgue lemma. 2 monotone class. 93 premeasure. 3 measure. 172 orthogonal. 73 vanish at no point. 3 Minkowski inequality generalized. 15 σﬁnite. 58 upper semicontinuous. 112 positive measure. 15 product σalgebra. 51 µ∗ measurable. 15. 5 vanishes at inﬁnity. 32 probability. 181 INDEX Schwartz class. 10 monotone convergence theorem. 156 second category. 181 orthonormal. 97 trigonometric series. 10 monotone class theorem. 163 vector space. 189 RadonNikodym theorem. 35 symmetric diﬀerence. 120 Vitali covering lemma. 4 Plancherel theorem. 58 uniformly integrable. 176 null set. 8 σﬁeld. 93 normed linear space. 172 uniformly absolutely continuous. 4 nowhere diﬀerentiable. 180 closed. 81 projection. 101. 188 Zorn’s lemma. 41 StoneWeierstrass theorem. 137 simple function. 93 open. 20. 96 negative set. 174 separate points. 5 σalgebra.194 metric space. 185 uniform boundedness theorem. 3 open ball. 15 sign function. 7 generated by a collection. 20 pairwise disjoint. 59 Vitali cover. 120 weak 11 inequality. 188 RadonNikodym derivative. 168 . 5 strictly convex. 93 signum function.
This action might not be possible to undo. Are you sure you want to continue?
Use one of your book credits to continue reading from where you left off, or restart the preview.