Measure Theory

John K. Hunter
Department of Mathematics, University of California at Davis

Abstract. These are some brief notes on measure theory, concentrating on

Lebesgue measure on Rn . Some missing topics I would have liked to have included had time permitted are: the change of variable formula for the Lebesgue
integral on Rn ; absolutely continuous functions and functions of bounded variation of a single variable and their connection with Lebesgue-Stieltjes measures
on R; Radon measures on Rn , and other locally compact Hausdorff topological
spaces, and the Riesz representation theorem for bounded linear functionals
on spaces of continuous functions; and other examples of measures, including
k-dimensional Hausdorff measure in Rn , Wiener measure and Brownian motion, and Haar measure on topological groups. All these topics can be found
in the references.

c John K. Hunter, 2011

Contents
Chapter 1. Measures
1.1. Sets
1.2. Topological spaces
1.3. Extended real numbers
1.4. Outer measures
1.5. -algebras
1.6. Measures
1.7. Sets of measure zero

1
1
2
2
3
4
5
6

Chapter 2. Lebesgue Measure on Rn

2.1. Lebesgue outer measure
2.2. Outer measure of rectangles
2.3. Caratheodory measurability
2.4. Null sets and completeness
2.5. Translational invariance
2.6. Borel sets
2.7. Borel regularity
2.8. Linear transformations
2.9. Lebesgue-Stieltjes measures

9
10
12
14
18
19
20
22
27
30

Chapter 3. Measurable functions

3.1. Measurability
3.2. Real-valued functions
3.3. Pointwise convergence
3.4. Simple functions
3.5. Properties that hold almost everywhere

33
33
34
36
37
38

Chapter 4. Integration
4.1. Simple functions
4.2. Positive functions
4.3. Measurable functions
4.4. Absolute continuity
4.5. Convergence theorems
4.6. Complex-valued functions and a.e. convergence
4.7. L1 spaces
4.8. Riemann integral
4.9. Integrals of vector-valued functions

39
39
40
42
45
47
50
50
52
52

Chapter 5. Product Measures

5.1. Product -algebras

55
55
iii

iv

CONTENTS

5.2.
5.3.
5.4.
5.5.
5.6.
5.7.

Premeasures
Product measures
Measurable functions
Monotone class theorem
Fubinis theorem
Completion of product measures

56
58
60
61
61
61

Chapter 6. Differentiation
6.1. A covering lemma
6.2. Maximal functions
6.3. Weak-L1 spaces
6.4. Hardy-Littlewood theorem
6.5. Lebesgue differentiation theorem
6.6. Signed measures
6.7. Hahn and Jordan decompositions
6.8. Radon-Nikodym theorem
6.9. Complex measures

63
64
65
67
67
68
70
71
74
77

Chapter 7. Lp spaces
7.1. Lp spaces
7.2. Minkowski and H
older inequalities
7.3. Density
7.4. Completeness
7.5. Duality

79
79
80
81
81
83

Bibliography

89

CHAPTER 1

Measures
Measures are a generalization of volume; the fundamental example is Lebesgue
measure on Rn , which we discuss in detail in the next Chapter. Moreover, as
formalized by Kolmogorov (1933), measure theory provides the foundation of probability. Measures are important not only because of their intrinsic geometrical and
probabilistic significance, but because they allow us to define integrals.
This connection, in fact, goes in both directions: we can define an integral
in terms of a measure; or, in the Daniell-Stone approach, we can start with an
integral (a linear functional acting on functions) and use it to define a measure. In
probability theory, this corresponds to taking the expectation of random variables
as the fundamental concept from which the probability of events is derived.
In these notes, we develop the theory of measures first, and then define integrals.
This is (arguably) the more concrete and natural approach; it is also (unarguably)
the original approach of Lebesgue. We begin, in this Chapter, with some preliminary definitions and terminology related to measures on arbitrary sets. See Folland
[4] for further discussion.
1.1. Sets
We use standard definitions and notations from set theory and will assume the
axiom of choice when needed. The words collection and family are synonymous
with set we use them when talking about sets of sets. We denote the collection
of subsets, or power set, of a set X by P(X). The notation 2X is also used.
If E X and the set X is understood, we denote the complement of E in X
by E c = X \ E. De Morgans laws state that
!c
!c

\
[
[
\
c
=
E ,
E
=
Ec .
E
I

We say that a collection

C = {E X : I}
of subsets of a set X, indexed by a set I, covers E X if
[
E E.
I

The collection C is disjoint if E E = for 6= .

The Cartesian product, or product, of sets X, Y is the collection of all ordered
pairs
X Y = {(x, y) : x X, y Y } .
1

1. MEASURES

1.2. Topological spaces

A topological space is a set equipped with a collection of open subsets that
satisfies appropriate conditions.
Definition 1.1. A topological space (X, T ) is a set X and a collection T
P(X) of subsets of X, called open sets, such that
(a) , X T ;
(b) if {U T : I} is an arbitrary collection of open sets, then their
union
[
U T
I

is open;
(c) if {Ui T : i = 1, 2, . . . , N } is a finite collection of open sets, then their
intersection
N
\
Ui T
i=1

is open.
The complement of an open set in X is called a closed set, and T is called a topology
on X.
1.3. Extended real numbers
It is convenient to use the extended real numbers
R = {} R {}.
This allows us, for example, to talk about sets with infinite measure or non-negative
functions with infinite integral. The extended real numbers are totally ordered in
the obvious way: is the largest element, is the smallest element, and real
numbers are ordered as in R. Algebraic operations on R are defined when they are
unambiguous e.g. + x = for every x R except x = , but is
undefined.
We define a topology on R in a natural way, making R homeomorphic to a
compact interval. For example, the function : R [1, 1] defined by

if x =
1
(x) =
x/ 1 + x2 if < x <

1
if x =
is a homeomorphism.
A primary reason to use the extended real numbers is that upper and lower
bounds always exist. Every subset of R has a supremum (equal to if the subset
contains or is not bounded from above in R) and infimum (equal to if the
subset contains or is not bounded from below in R). Every increasing sequence
of extended real numbers converges to its supremum, and every decreasing sequence
converges to its infimum. Similarly, if {an } is a sequence of extended real-numbers
then




lim inf an = sup inf ai
lim sup an = inf sup ai ,
n

nN

in

both exist as extended real numbers.

nN

in

1.4. OUTER MEASURES

P
Every sum i=1 xi with non-negative terms xi 0 converges in R (to if
xi = for some i N or the series diverges in R), where the sum is defined by
(
)

X
X
xi = sup
xi : F N is finite .
i=1

iF

As for non-negative sums of real numbers, non-negative sums of extended real

numbers are unconditionally convergent (the order of the terms does not matter);
we can rearrange sums of non-negative extended real numbers

(xi + yi ) =

xi +

i=1

i=1

yi ;

i=1

and double sums may be evaluated as iterated single sums

X
X
xij : F N N is finite
xij = sup

i,j=1
(i,j)F

X
X

xij
=
=

i=1

j=1

j=1

i=1

xij

Our use of extended real numbers is closely tied to the order and monotonicity
properties of R. In dealing with complex numbers or elements of a vector space,
we will always require that they are strictly finite.
1.4. Outer measures
As stated in the following definition, an outer measure is a monotone, countably
subadditive, non-negative, extended real-valued function defined on all subsets of
a set.
Definition 1.2. An outer measure on a set X is a function
: P(X) [0, ]
such that:
(a) () = 0;
(b) if E F X, then (E) (F );
(c) if {Ei X : i N} is a countable collection of subsets of X, then
!

X
[
(Ei ).
Ei

i=1

i=1

We obtain a statement about finite unions from a statement about infinite

unions by taking all but finitely many sets in the union equal to the empty set.
Note that is not assumed to be additive even if the collection {Ei } is disjoint.

1. MEASURES

1.5. -algebras
A -algebra on a set X is a collection of subsets of a set X that contains and
X, and is closed under complements, finite unions, countable unions, and countable
intersections.
Definition 1.3. A -algebra on a set X is a collection A of subsets of X such
that:
(a) , X A;
(b) if A A then Ac A;
(c) if Ai A for i N then

\
[
Ai A.
Ai A,
i=1

i=1

From de Morgans laws, a collection of subsets is -algebra if it contains and

is closed under the operations of taking complements and countable unions (or,
equivalently, countable intersections).
Example 1.4. If X is a set, then {, X} and P(X) are -algebras on X; they
are the smallest and largest -algebras on X, respectively.
Measurable spaces provide the domain of measures, defined below.

Definition 1.5. A measurable space (X, A) is a non-empty set X equipped

with a -algebra A on X.

It is useful to compare the definition of a -algebra with that of a topology in

Definition 1.1. There are two significant differences. First, the complement of a
measurable set is measurable, but the complement of an open set is not, in general,
open, excluding special cases such as the discrete topology T = P(X). Second,
countable intersections and unions of measurable sets are measurable, but only
finite intersections of open sets are open while arbitrary (even uncountable) unions
of open sets are open. Despite the formal similarities, the properties of measurable
and open sets are very different, and they do not combine in a straightforward way.
If F is any collection of subsets of a set X, then there is a smallest -algebra
on X that contains F , denoted by (F ).
Definition 1.6. If F is any collection of subsets of a set X, then the -algebra
generated by F is
\
(F ) =
{A P(X) : A F and A is a -algebra} .

This intersection is nonempty, since P(X) is a -algebra that contains F , and

an intersection of -algebras is a -algebra. An immediate consequence of the
definition is the following result, which we will use repeatedly.
Proposition 1.7. If F is a collection of subsets of a set X such that F A
where A is a -algebra on X, then (F ) A.

Among the most important -algebras are the Borel -algebras on topological
spaces.
Definition 1.8. Let (X, T ) be a topological space. The Borel -algebra
B(X) = (T )

is the -algebra generated by the collection T of open sets on X.

1.6. MEASURES

1.6. Measures
A measure is a countably additive, non-negative, extended real-valued function
defined on a -algebra.
Definition 1.9. A measure on a measurable space (X, A) is a function
: A [0, ]

such that
(a) () = 0;
(b) if {Ai A : i N} is a countable disjoint collection of sets in A, then
!

X
[
(Ai ).
Ai =

i=1

i=1

In comparison with an outer measure, a measure need not be defined on all

subsets of a set, but it is countably additive rather than countably subadditive.
S
A measure on a set X is finite if (X) < , and -finite if X = n=1 An
is a countable union of measurable sets An with finite measure, (An ) < . A
probability measure is a finite measure with (X) = 1.
A measure space (X, A, ) consists of a set X, a -algebra A on X, and a
measure defined on A. When A and are clear from the context, we will refer to
the measure space X. We define subspaces of measure spaces in the natural way.
Definition 1.10. If (X, A, ) is a measure space and E X is a measurable
subset, then the measure subspace (E, A|E , |E ) is defined by restricting to E:
A|E = {A E : A A} ,

|E (A E) = (A E).

As we will see, the construction of nontrivial measures, such as Lebesgue measure, requires considerable effort. Nevertheless, there is at least one useful example
of a measure that is simple to define.
Example 1.11. Let X be an arbitrary non-empty set. Define : P(X)
[0, ] by
(E) = number of elements in E,
where () = 0 and (E) = if E is not finite. Then is a measure, called counting measure on X. Every subset of X is measurable with respect to . Counting
measure is finite if X is finite and -finite if X is countable.
A useful implication of the countable additivity of a measure is the following
monotonicity result.
Proposition 1.12. If {Ai : i N} is an increasing sequence of measurable
sets, meaning that Ai+1 Ai , then
!

[
Ai = lim (Ai ).
(1.1)

i=1

If {Ai : i N} is a decreasing sequence of measurable sets, meaning that Ai+1 Ai ,

and (A1 ) < , then
!

\
Ai = lim (Ai ).
(1.2)

i=1

1. MEASURES

Proof. If {Ai : i N} is an increasing sequence of sets and Bi = Ai+1 \ Ai ,

then {Bi : i N} is a disjoint sequence with the same union, so by the countable
additivity of
!
!

X
[
[
(Bi ) .
Bi =
Ai =

Moreover, since Aj =

Sj

i=1

i=1

i=1

i=1

Bi ,

(Aj ) =

j
X

(Bi ) ,

i=1

which implies that

(Bi ) = lim (Aj )

j

i=1

and the first result follows.

If (A1 ) < and {Ai } is decreasing, then {Bi = A1 \ Ai } is increasing and
(Bi ) = (A1 ) (Ai ).

It follows from the previous result that

!

[
Bi = lim (Bi ) = (A1 ) lim (Ai ).

i=1

Since

i=1

Bi = A1 \

the result follows.

i=1

Ai ,

i=1

Bi

= (A1 )

i=1

Ai

,


Example 1.13. To illustrate the necessity of the condition (A1 ) < in the
second part of the previous proposition, or more generally (An ) < for some
n N, consider counting measure : P(N) [0, ] on N. If
An = {k N : k n},

then (An ) = for every n N, so (An ) as n , but

!

\
\
An = ,

An = 0.
n=1

n=1

1.7. Sets of measure zero

A set of measure zero, or a null set, is a measurable set N such that (N ) = 0.
A property which holds for all x X \ N where N is a set of measure zero is said
to hold almost everywhere, or a.e. for short. If we want to emphasize the measure,
we say -a.e. In general, a subset of a set of measure zero need not be measurable,
but if it is, it must have measure zero.
It is frequently convenient to use measure spaces which are complete in the
following sense. (This is, of course, a different sense of complete than the one used
in talking about complete metric spaces.)
Definition 1.14. A measure space (X, A, ) is complete if every subset of a
set of measure zero is measurable.

1.7. SETS OF MEASURE ZERO

Note that completeness depends on the measure , not just the -algebra
A. Any measure space (X, A, ) is contained in a uniquely defined completion
(X, A, ), which the smallest complete measure space that contains it and is given
explicitly as follows.
Theorem 1.15. If (X, A, ) is a measure space, define (X, A, ) by

A = {A M : A A, M N where N A satisfies (N ) = 0}

with (A M ) = (A). Then (X, A, ) is a complete measure space such that

A A and is the unique extension of to A.
Proof. The collection A is a -algebra. It is closed under complementation
because, with the notation used in the definition,
Therefore

(A M )c = Ac M c ,

M c = N c (N \ M ).

(A M )c = (Ac N c ) (Ac (N \ M )) A,
since Ac N c A and Ac (N \ M ) N . Moreover, A is closed under countable
unions because if Ai A and Mi Ni where (Ni ) = 0 for each i N, then
!
!

[
[
[
Mi A,
Ai
Ai Mi =

i=1

Ai A,

i=1

i=1

i=1

since

i=1

Mi

i=1

Ni ,

i=1

Ni

= 0.

It is straightforward to check that is well-defined and is the unique extension of

to a measure on A, and that (X, A, ) is complete.


CHAPTER 2

Lebesgue Measure on Rn
Our goal is to construct a notion of the volume, or Lebesgue measure, of rather
general subsets of Rn that reduces to the usual volume of elementary geometrical
sets such as cubes or rectangles.
If L(Rn ) denotes the collection of Lebesgue measurable sets and
: L(Rn ) [0, ]

denotes Lebesgue measure, then we want L(Rn ) to contain all n-dimensional rectangles and (R) should be the usual volume of a rectangle R. Moreover, we want
to be countably additive. That is, if
{Ai L(Rn ) : i N}
is a countable collection of disjoint measurable sets, then their union should be
measurable and
!

X
[
(Ai ) .
Ai =

i=1

i=1

The reason for requiring countable additivity is that finite additivity is too weak
a property to allow the justification of any limiting processes, while uncountable
additivity is too strong; for example, it would imply that if the measure of a set
consisting of a single point is zero, then the measure of every subset of Rn would
be zero.
It is not possible to define the Lebesgue measure of all subsets of Rn in a
geometrically reasonable way. Hausdorff (1914) showed that for any dimension
n 1, there is no countably additive measure defined on all subsets of Rn that is
invariant under isometries (translations and rotations) and assigns measure one to
the unit cube. He further showed that if n 3, there is no such finitely additive
measure. This result is dramatized by the Banach-Tarski paradox: Banach and
Tarski (1924) showed that if n 3, one can cut up a ball in Rn into a finite number
of pieces and use isometries to reassemble the pieces into a ball of any desired volume
e.g. reassemble a pea into the sun. The construction of these pieces requires the
axiom of choice.1 Banach (1923) also showed that if n = 1 or n = 2 there are
finitely additive, isometrically invariant extensions of Lebesgue measure on Rn that
are defined on all subsets of Rn , but these extensions are not countably additive.
For a detailed discussion of the Banach-Tarski paradox and related issues, see [10].
The moral of these results is that some subsets of Rn are too irregular to define
their Lebesgue measure in a way that preserves countable additivity (or even finite
additivity in n 3 dimensions) together with the invariance of the measure under
1Solovay (1970) proved that one has to use the axiom of choice to obtain non-Lebesgue
measurable sets.
9

10

2. LEBESGUE MEASURE ON Rn

isometries. We will show, however, that such a measure can be defined on a algebra L(Rn ) of Lebesgue measurable sets which is large enough to include all set
of practical importance in analysis. Moreover, as we will see, it is possible to define
an isometrically-invariant, countably sub-additive outer measure on all subsets of
Rn .
There are many ways to construct Lebesgue measure, all of which lead to the
same result. We will follow an approach due to Caratheodory, which generalizes
to other measures: We first construct an outer measure on all subsets of Rn by
approximating them from the outside by countable unions of rectangles; we then
restrict this outer measure to a -algebra of measurable subsets on which it is countably additive. This approach is somewhat asymmetrical in that we approximate
sets (and their complements) from the outside by elementary sets, but we do not
approximate them directly from the inside.
Jones [5], Stein and Shakarchi [8], and Wheeler and Zygmund [11] give detailed
introductions to Lebesgue measure on Rn . Cohn [2] gives a similar development to
the one here, and Evans and Gariepy [3] discuss more advanced topics.
2.1. Lebesgue outer measure
We use rectangles as our elementary sets, defined as follows.
Definition 2.1. An n-dimensional, closed rectangle with sides oriented parallel
to the coordinate axes, or rectangle for short, is a subset R Rn of the form
R = [a1 , b1 ] [a2 , b2 ] [an , bn ]
where < ai bi < for i = 1, . . . , n. The volume (R) of R is
(R) = (b1 a1 )(b2 a2 ) . . . (bn an ).
If n = 1 or n = 2, the volume of a rectangle is its length or area, respectively.
We also consider the empty set to be a rectangle with () = 0. We denote the
collection of all n-dimensional rectangles by R(Rn ), or R when n is understood,
and then R 7 (R) defines a map
: R(Rn ) [0, ).
The use of this particular class of elementary sets is for convenience. We could
equally well use open or half-open rectangles, cubes, balls, or other suitable elementary sets; the result would be the same.
Definition 2.2. The outer Lebesgue measure (E) of a subset E Rn , or
outer measure for short, is
)
(
X
S

n
(Ri ) : E i=1 Ri , Ri R(R )
(2.1)
(E) = inf
i=1

where the infimum is taken over all countable collections of rectangles whose union
contains E. The map
: P(Rn ) [0, ],

is called outer Lebesgue measure.

: E 7 (E)

2.1. LEBESGUE OUTER MEASURE

11

P
In this definition, a sum i=1 (Ri ) and (E) may take the value . We do
not require that the rectangles Ri are disjoint, so the same volume may contribute
to multiple terms in the sum on the right-hand side of (2.1); this does not affect
the value of the infimum.
Example 2.3. Let E = Q [0, 1] be the set of rational numbers between 0
and 1. Then E has outer measure zero. To prove this, let {qi : i N} be an
i
enumeration of the points in E.
SGiven > 0, let Ri be an interval of length /2
which contains qi . Then E i=1 (Ri ) so
0 (E)

(Ri ) = .

i=1

Hence (E) = 0 since > 0 is arbitrary. The same argument shows that any
countable set has outer measure zero. Note that if we cover E by a finite collection
of intervals, then the union of the intervals would have to contain [0, 1] since E is
dense in [0, 1] so their lengths sum to at least one.
The previous example illustrates why we need to use countably infinite collections of rectangles, not just finite collections, to define the outer measure.2 The
countable -trick used in the example appears in various forms throughout measure
theory.
Next, we prove that is an outer measure in the sense of Definition 1.2.
Theorem 2.4. Lebesgue outer measure has the following properties.
(a) () = 0;
(b) if E F , then (E) (F );
(c) if {Ei Rn : i N} is a countable collection of subsets of Rn , then
!

X
[

(Ei ) .
Ei

i=1

i=1

Proof. It follows immediately from Definition 2.2 that () = 0, since every

collection of rectangles covers , and that (E) (F ) if E F since any cover
of F covers E.
The main property to prove is the countable subadditivity of . If (Ei ) =
for some i N, there is nothing to prove, so we may assume that (Ei ) is finite
for every i N. If > 0, there is a countable covering {Rij : j N} of Ei by
rectangles Rij such that

X
j=1

(Rij ) (Ei ) +

,
2i

Ei

Rij .

j=1

Then {Rij : i, j N} is a countable covering of

E=

Ei

i=1
2The use of finitely many intervals leads to the notion of the Jordan content of a set, introduced by Peano (1887) and Jordan (1892), which is closely related to the Riemann integral; Borel
(1898) and Lebesgue (1902) generalized Jordans approach to allow for countably many intervals,
leading to Lebesgue measure and the Lebesgue integral.

2. LEBESGUE MEASURE ON Rn

12

and therefore
(E)

i,j=1

(Rij )

n
X

o X
(Ei ) + .
=
2i
i=1

(Ei ) +

i=1

Since > 0 is arbitrary, it follows that

(E)

(Ei )

i=1

which proves the result.

2.2. Outer measure of rectangles

In this section, we prove the geometrically obvious, but not entirely trivial, fact
that the outer measure of a rectangle is equal to its volume. The main point is to
show that the volumes of a countable collection of rectangles that cover a rectangle
R cannot sum to less than the volume of R.3
We begin with some combinatorial facts about finite covers of rectangles [8].
We denote the interior of a rectangle R by R , and we say that rectangles R, S
are almost disjoint if R S = , meaning that they intersect at most along their
boundaries. The proofs of the following results are cumbersome to write out in
detail (its easier to draw a picture) but we briefly explain the argument.
Lemma 2.5. Suppose that
R = I1 I2 In
is an n-dimensional rectangle where each closed, bounded interval Ii R is an
almost disjoint union of closed, bounded intervals {Ii,j R : j = 1, . . . , Ni },
Ii =

Ni
[

Ii,j .

j=1

Define the rectangles

(2.2)

Sj1 j2 ...jn = I1,j1 I2,j2 In,jn .

Then
(R) =

N1
X

j1 =1

Nn
X

(Sj1 j2 ...jn ) .

jn =1

Proof. Denoting the length of an interval I by |I|, using the fact that
|Ii | =

Ni
X
j=1

|Ii,j |,

3As a partial justification of the need to prove this fact, note that it would not be true if we
allowed uncountable covers, since we could cover any rectangle by an uncountable collection of
points all of whose volumes are zero.

2.2. OUTER MEASURE OF RECTANGLES

13

and expanding the resulting product, we get that

(R) = |I1 ||I2 | . . . |In |

Nn
N2
N1
X
X
X
|In,jn |
|I2,j2 | . . .
|I1,j1 |
=
=

j1 =1 j2 =1

N2
N1 X
X

j1 =1 j2 =1

jn =1

j2 =1

j1 =1

N2
N1 X
X

Nn
X

jn =1
Nn
X

|I1,j1 ||I2,j2 | . . . |In,jn |

(Sj1 j2 ...jn ) .

jn =1

Proposition 2.6. If a rectangle R is an almost disjoint, finite union of rectangles {R1 , R2 , . . . , RN }, then
(2.3)

(R) =

N
X

(Ri ).

N
X

(Ri ).

i=1

If R is covered by rectangles {R1 , R2 , . . . , RN }, which need not be disjoint, then

(2.4)

(R)

Proof. Suppose that

i=1

R = [a1 , b1 ] [a2 , b2 ] [an , bn ]

is an almost disjoint union of the rectangles {R1 , R2 , . . . , RN }. Then by extending

the sides of the Ri , we may decompose R into an almost disjoint collection of
rectangles
{Sj1 j2 ...jn : 1 ji Ni for 1 i n}
that is obtained by taking products of subintervals of partitions of the coordinate
intervals [ai , bi ] into unions of almost disjoint, closed subintervals. Explicitly, we
partition [ai , bi ] into
ai = ci,0 ci,1 ci,Ni = bi ,

Ii,j = [ci,j1 , ci,j ].

where the ci,j are obtained by ordering the left and right ith coordinates of all faces
of rectangles in the collection {R1 , R2 , . . . , RN }, and define rectangles Sj1 j2 ...jn as
in (2.2).
Each rectangle Ri in the collection is an almost disjoint union of rectangles
Sj1 j2 ...jn , and their union contains all such products exactly once, so by applying
Lemma 2.5 to each Ri and summing the results we see that
N
X
i=1

(Ri ) =

N1
X

j1 =1

Nn
X

(Sj1 j2 ...jn ) .

jn =1

Similarly, R is an almost disjoint union of all the rectangles Sj1 j2 ...jn , so Lemma 2.5
implies that
Nn
N1
X
X
(Sj1 j2 ...jn ) ,

(R) =
j1 =1

jn =1

2. LEBESGUE MEASURE ON Rn

14

and (2.3) follows.

If a finite collection of rectangles {R1 , R2 , . . . , RN } covers R, then there is a
almost disjoint, finite collection of rectangles {S1 , S2 , . . . , SM } such that
R=

M
[

M
X

Si ,

i=1

i=1

(Si )

N
X

(Ri ).

i=1

To obtain the Si , we replace Ri by the rectangle R Ri , and then decompose

these possibly non-disjoint rectangles into an almost disjoint, finite collection of
sub-rectangles with the same union; we discard overlaps which can only reduce
the sum of the volumes. Then, using (2.3), we get
(R) =

M
X
i=1

which proves (2.4).

(Si )

N
X

(Ri ),

i=1

The outer measure of a rectangle is defined in terms of countable covers. We

reduce these to finite covers by using the topological properties of Rn .
Proposition 2.7. If R is a rectangle in Rn , then (R) = (R).
Proof. Since {R} covers R, we have (R) (R), so we only need to prove
the reverse inequality.
Suppose that {Ri : i N} is a countably infinite collection of rectangles that
covers R. By enlarging Ri slightly we may obtain a rectangle Si whose interior Si
contains Ri such that

(Si ) (Ri ) + i .
2
Then {Si : i N} is an open cover of the compact set R, so it contains a finite

subcover, which we may label as {S1 , S2 , . . . , SN

}. Then {S1 , S2 , . . . , SN } covers
R and, using (2.4), we find that
(R)

N
X
i=1

(Si )

Since > 0 is arbitrary, we have

N n

X
o X
(Ri ) + i
(Ri ) + .
2
i=1
i=1

(R)
and it follows that (R) (R).

(Ri )

i=1

2.3. Carath
eodory measurability
We will obtain Lebesgue measure as the restriction of Lebesgue outer measure
to Lebesgue measurable sets. The construction, due to Caratheodory, works for any
outer measure, as given in Definition 1.2, so we temporarily consider general outer
measures. We will return to Lebesgue measure on Rn at the end of this section.
The following is the Caratheodory definition of measurability.
Definition 2.8. Let be an outer measure on a set X. A subset A X is
Caratheodory measurable with respect to , or measurable for short, if
(2.5)

(E) = (E A) + (E Ac )


2.3. CARATHEODORY
MEASURABILITY

15

for every subset E X.

We also write E Ac as E \ A. Thus, a measurable set A splits any set
E into disjoint pieces whose outer measures add up to the outer measure of E.
Heuristically, this condition means that a set is measurable if it divides other sets
in a nice way. The regularity of the set E being divided is not important here.
Since is subadditive, we always have that
(E) (E A) + (E Ac ).
Thus, in order to prove that A X is measurable, it is sufficient to show that
(E) (E A) + (E Ac )

for every E X, and then we have equality as in (2.5).

Definition 2.8 is perhaps not the most intuitive way to define the measurability
of sets, but it leads directly to the following key result.
Theorem 2.9. The collection of Caratheodory measurable sets with respect to
an outer measure is a -algebra, and the restriction of to the measurable sets
is a measure.
Proof. It follows immediately from (2.5) that is measurable and the complement of a measurable set is measurable, so to prove that the collection of measurable
sets is a -algebra, we only need to show that it is closed under countable unions.
We will prove at the same time that is countably additive on measurable sets;
since () = 0, this will prove that the restriction of to the measurable sets is
a measure.
First, we prove that the union of measurable sets is measurable. Suppose that
A, B are measurable and E X. The measurability of A and B implies that
(2.6)

(E) = (E A) + (E Ac )

= (E A B) + (E A B c )

+ (E Ac B) + (E Ac B c ).

Since A B = (A B) (A B c ) (Ac B) and is subadditive, we have

(E (A B)) (E A B) + (E A B c ) + (E Ac B).

The use of this inequality and the relation Ac B c = (A B)c in (2.6) implies that
(E) (E (A B)) + (E (A B)c )

so A B is measurable.
Moreover, if A is measurable and A B = , then by taking E = A B in
(2.5), we see that
(A B) = (A) + (B).

Thus, the outer measure of the union of disjoint, measurable sets is the sum of
their outer measures. The repeated application of this result implies that the finite
union of measurable sets is measurable and is finitely additive on the collection
of measurable sets.
Next, we we want to show that the countable union of measurable sets is
measurable. It is sufficient to consider disjoint unions. To see this, note that if

2. LEBESGUE MEASURE ON Rn

16

{Ai : i N} is a countably infinite collection of measurable sets, then

j
[

Bj =

Ai ,

for j 1

i=1

form an increasing sequence of measurable sets, and

Cj = Bj \ Bj1

for j 2,

C1 = B1

form a disjoint measurable collection of sets. Moreover

Ai =

Cj .

j=1

i=1

Suppose that {Ai : i N} is a countably infinite, disjoint collection of measurable sets, and define
j

[
[
Ai ,
B=
Ai .
Bj =
i=1

i=1

Let E X. Since Aj is measurable and Bj = Aj Bj1 is a disjoint union (for

j 2),
(E Bj ) = (E Bj Aj ) + (E Bj Acj ),

= (E Aj ) + (E Bj1 ).

Also (E B1 ) = (E A1 ). It follows by induction that

(E Bj ) =

j
X
i=1

(E Ai ).

Since Bj is a finite union of measurable sets, it is measurable, so

(E) = (E Bj ) + (E Bjc ),
and since Bjc B c , we have
(E Bjc ) (E B c ).
It follows that

(E)

j
X
i=1

(E Ai ) + (E B c ).

Taking the limit of this inequality as j and using the subadditivity of , we

get
(E)
(2.7)

X
i=1

(E Ai ) + (E B c )

i=1

E Ai

+ (E B c )

(E B) + (E B c )
(E).


2.3. CARATHEODORY
MEASURABILITY

Therefore, we must have equality in (2.7), which shows that B =

surable. Moreover,
!

X
[

(E Ai ),
E Ai =

17

i=1

Ai is mea-

i=1

i=1

so taking E = X, we see that is countably additive on the -algebra of measurable sets.


Returning to Lebesgue measure on Rn , the preceding theorem shows that we
get a measure on Rn by restricting Lebesgue outer measure to its Caratheodorymeasurable sets, which are the Lebesgue measurable subsets of Rn .
Definition 2.10. A subset A Rn is Lebesgue measurable if
(E) = (E A) + (E Ac )

for every subset E Rn . If L(Rn ) denotes the -algebra of Lebesgue measurable

sets, the restriction of Lebesgue outer measure to the Lebesgue measurable sets
: L(Rn ) [0, ],

= |L(Rn )

is called Lebesgue measure.

From Proposition 2.7, this notation is consistent with our previous use of to
denote the volume of a rectangle. If E Rn is any measurable subset of Rn , then
we define Lebesgue measure on E by restricting Lebesgue measure on Rn to E, as
in Definition 1.10, and denote the corresponding -algebra of Lebesgue measurable
subsets of E by L(E).
Next, we prove that all rectangles are measurable; this implies that L(Rn ) is a
large collection of subsets of Rn . Not all subsets of Rn are Lebesgue measurable,
however; e.g. see Example 2.17 below.
Proposition 2.11. Every rectangle is Lebesgue measurable.
Proof. Let R be an n-dimensional rectangle and E Rn . Given > 0, there
is a cover {Ri : i N} of E by rectangles Ri such that

X
(Ri ).
(E) +
i=1

We can decompose Ri into an almost disjoint, finite union of rectangles

i , Si,1 , . . . , Si,N }
{R
such that

i +
Ri = R

N
[

j=1

From (2.3),

Si,j ,

i = Ri R R,
R

i ) +
(Ri ) = (R

N
X

Si,j Rc .

(Si,j ).

j=1

Using this result in the previous sum, relabeling the Si,j as Si , and rearranging the
resulting sum, we get that

X
X
i) +
(E) +
(R
(Si ).
i=1

i=1

2. LEBESGUE MEASURE ON Rn

18

i : i N} cover E R and the rectangles {Si : i N} cover

Since the rectangles {R
c
E R , we have

X
X
i ),
(Si ).
(R
(E Rc )
(E R)
i=1

i=1

Hence,

(E) + (E R) + (E Rc ).
Since > 0 is arbitrary, it follows that
which proves the result.

(E) (E R) + (E Rc ),


An open rectangle R is a union of an increasing sequence of closed rectangles

whose volumes approach (R); for example
(a1 , b1 ) (a2 , b2 ) (an , bn )
=

[a1 +

k=1

1
1
1
1
1
1
, b1 ] [a2 + , b2 ] [an + , bn ].
k
k
k
k
k
k

Thus, R is measurable and, from Proposition 1.12,

(R ) = (R).
Moreover if R = R \ R denotes the boundary of R, then
(R) = (R) (R ) = 0.

2.4. Null sets and completeness

Sets of measure zero play a particularly important role in measure theory and
integration. First, we show that all sets with outer Lebesgue measure zero are
Lebesgue measurable.
Proposition 2.12. If N Rn and (N ) = 0, then N is Lebesgue measurable,
and the measure space (Rn , L(Rn ), ) is complete.
Proof. If N Rn has outer Lebesgue measure zero and E Rn , then
0 (E N ) (N ) = 0,

so (E N ) = 0. Therefore, since E E N c ,

(E) (E N c ) = (E N ) + (E N c ),

which shows that N is measurable. If N is a measurable set with (N ) = 0 and

M N , then (M ) = 0, since (M ) (N ). Therefore M is measurable and
(Rn , L(Rn ), ) is complete.

In view of the importance of sets of measure zero, we formulate their definition
explicitly.
Definition 2.13. A subset N Rn has Lebesgue measure zero if for every
> 0 there exists a countable collection of rectangles {Ri : i N} such that
N

i=1

Ri ,

X
i=1

(Ri ) < .

2.5. TRANSLATIONAL INVARIANCE

19

The argument in Example 2.3 shows that every countable set has Lebesgue
measure zero, but sets of measure zero may be uncountable; in fact the fine structure
of sets of measure zero is, in general, very intricate.
Example 2.14. The standard Cantor set, obtained by removing middle thirds
from [0, 1], is an uncountable set of zero one-dimensional Lebesgue measure.
Example 2.15. The x-axis in R2


A = (x, 0) R2 : x R

has zero two-dimensional Lebesgue measure. More generally, any linear subspace of
Rn with dimension strictly less than n has zero n-dimensional Lebesgue measure.
2.5. Translational invariance
An important geometric property of Lebesgue measure is its translational invariance. If A Rn and h Rn , let
A + h = {x + h : x A}

denote the translation of A by h.

Proposition 2.16. If A Rn and h Rn , then
(A + h) = (A),

and A + h is measurable if and only if A is measurable.

Proof. The invariance of outer measure result is an immediate consequence
of the definition, since {Ri + h : i N} is a cover of A + h if and only if {Ri :
i N} is a cover of A, and (R + h) = (R) for every rectangle R. Moreover, the
Caratheodory definition of measurability is invariant under translations since
(E + h) (A + h) = (E A) + h.

The space Rn is a locally compact topological (abelian) group with respect to
translation, which is a continuous operation. More generally, there exists a (left or
right) translation-invariant measure, called Haar measure, on any locally compact
topological group; this measure is unique up to a scalar factor.
The following is the standard example of a non-Lebesgue measurable set, due
to Vitali (1905).
Example 2.17. Define an equivalence relation on R by x y if x y Q.
This relation has uncountably many equivalence classes, each of which contains a
countably infinite number of points and is dense in R. Let E [0, 1] be a set that
contains exactly one element from each equivalence class, so that R is the disjoint
union of the countable collection of rational translates of E. Then we claim that E
is not Lebesgue measurable.
To show this, suppose for contradiction that E is measurable. Let {qi : i N}
be an enumeration of the rational numbers in the interval [1, 1] and let Ei = E +qi
denote the translation of E by qi . Then the sets Ei are disjoint and
[0, 1]

i=1

Ei [1, 2].

20

2. LEBESGUE MEASURE ON Rn

The translational invariance of Lebesgue measure implies that each Ei is measurable

with (Ei ) = (E), and the countable additivity of Lebesgue measure implies that
1
But this is impossible, since
(E) = 0 or (E) > 0.

i=1

X
i=1

(Ei ) 3.

(Ei ) is either 0 or , depending on whether if

The above example is geometrically simpler on the circle T = R/Z. When

reduced modulo one, the sets {Ei : i N} partition T into a countable union of
disjoint sets which are translations of each other. If the sets were measurable, their
measures would be equal so they must sum to 0 or , but the measure of T is one.
2.6. Borel sets
The relationship between measure and topology is not a simple one. In this
section, we show that all open and closed sets in Rn , and therefore all Borel sets
(i.e. sets that belong to the -algebra generated by the open sets), are Lebesgue
measurable.
Let T (Rn ) P(Rn ) denote the standard metric topology on Rn consisting of
all open sets. That is, G Rn belongs to T (Rn ) if for every x G there exists
r > 0 such that Br (x) G, where
Br (x) = {y Rn : |x y| < r}

is the open ball of radius r centered at x Rn and | | denotes the Euclidean norm.

Definition 2.18. The Borel -algebra B(Rn ) on Rn is the -algebra generated

by the open sets, B(Rn ) = (T (Rn )). A set that belongs to the Borel -algebra is
called a Borel set.
Since -algebras are closed under complementation, the Borel -algebra is also
generated by the closed sets in Rn . Moreover, since Rn is -compact (i.e. it is a
countable union of compact sets) its Borel -algebra is generated by the compact
sets.

Remark 2.19. This definition is not constructive, since we start with the power
set of Rn and narrow it down until we obtain the smallest -algebra that contains
the open sets. It is surprisingly complicated to obtain B(Rn ) by starting from
the open or closed sets and taking successive complements, countable unions, and
countable intersections. These operations give sequences of collections of sets in Rn
(2.8)

G G G G . . . ,

F F F F . . . ,

where G denotes the open sets, F the closed sets, the operation of countable
unions, and the operation of countable intersections. These collections contain
each other; for example, F G and G F . This process, however, has to
be repeated up to the first uncountable ordinal before we obtain B(Rn ). This is
because if, for example, {Ai : i N} is a countable family of sets such that
A1 G \ G,

A2 G \ G , A3 G \ G , . . .
S
T
and so on, then there is no guarantee that
i=1 Ai or
i=1 Ai belongs to any of
the previously constructed families. In general, one only knows that they belong to
the + 1 iterates G... or G... , respectively, where is the ordinal number

2.6. BOREL SETS

21

of N. A similar argument shows that in order to obtain a family which is closed

under countable intersections or unions, one has to continue this process until one
has constructed an uncountable number of families.
To show that open sets are measurable, we will represent them as countable
unions of rectangles. Every open set in R is a countable disjoint union of open
intervals (one-dimensional open rectangles). When n 2, it is not true that every
open set in Rn is a countable disjoint union of open rectangles, but we have the
following substitute.
Proposition 2.20. Every open set in Rn is a countable union of almost disjoint
rectangles.
Proof. Let G Rn be open. We construct a family of cubes (rectangles of
equal sides) as follows. First, we bisect Rn into almost disjoint cubes {Qi : i N}
of side one with integer coordinates. If Qi G, we include Qi in the family, and
if Qi is disjoint from G, we exclude it. Otherwise, we bisect the sides of Qi to
obtain 2n almost disjoint cubes of side one-half and repeat the procedure. Iterating
this process arbitrarily many times, we obtain a countable family of almost disjoint
cubes.
The union of the cubes in this family is contained in G, since we only include
cubes that are contained in G. Conversely, if x G, then since G is open some sufficiently small cube in the bisection procedure that contains x is entirely contained
in G, and the largest such cube is included in the family. Hence the union of the
family contains G, and is therefore equal to G.

In fact, the proof shows that every open set is an almost disjoint union of dyadic
cubes.
Proposition 2.21. The Borel algebra B(Rn ) is generated by the collection of
rectangles R(Rn ). Every Borel set is Lebesgue measurable.

Proof. Since R is a subset of the closed sets, we have (R) B. Conversely,

by the previous proposition, (R) T , so (R) (T ) = B, and therefore
B = (R). From Proposition 2.11, we have R L. Since L is a -algebra, it
follows that (R) L, so B L.

Note that if
G=

Ri

i=1

is a decomposition of an open set G into an almost disjoint union of closed rectangles, then

[
Ri
G
i=1

is a disjoint union, and therefore

X
X
(Ri ).
(Ri ) (G)
i=1

i=1

Since

(Ri )

= (Ri ), it follows that

(G) =

X
i=1

(Ri )

2. LEBESGUE MEASURE ON Rn

22

for any such decomposition and that the sum is independent of the way in which
G is decomposed into almost disjoint rectangles.
The Borel -algebra B is not complete and is strictly smaller than the Lebesgue
-algebra L. In fact, one can show that the cardinality of B is equal to the cardinality c of the real numbers, whereas the cardinality of L is equal to 2c . For example,
the Cantor set is a set of measure zero with the same cardinality as R and every
subset of the Cantor set is Lebesgue measurable.
We can obtain examples of sets that are Lebesgue measurable but not Borel
measurable by considering subsets of sets of measure zero. In the following example
of such a set in R, we use some properties of measurable functions which will be
proved later.
Example 2.22. Let f : [0, 1] [0, 1] denote the standard Cantor function and
define g : [0, 1] [0, 1] by
g(y) = inf {x [0, 1] : f (x) = y} .

Then g is an increasing, one-to-one function that maps [0, 1] onto the Cantor set
C. Since g is increasing it is Borel measurable, and the inverse image of a Borel
set under g is Borel. Let E [0, 1] be a non-Lebesgue measurable set. Then
F = g(E) C is Lebesgue measurable, since it is a subset of a set of measure zero,
but F is not Borel measurable, since if it was E = g 1 (F ) would be Borel.
Other examples of Lebesgue measurable sets that are not Borel sets arise from
the theory of product measures in Rn for n 2. For example, let N = E {0} R2
where E R is a non-Lebesgue measurable set in R. Then N is a subset of the
x-axis, which has two-dimensional Lebesgue measure zero, so N belongs to L(R2 )
since Lebesgue measure is complete. One can show, however, that if a set belongs
to B(R2 ) then every section with fixed x or y coordinate, belongs to B(R); thus, N
cannot belong to B(R2 ) since the y = 0 section E is not Borel.
As we show below, L(Rn ) is the completion of B(Rn ) with respect to Lebesgue
measure, meaning that we get all Lebesgue measurable sets by adjoining all subsets
of Borel sets of measure zero to the Borel -algebra and taking unions of such sets.
2.7. Borel regularity
Regularity properties of measures refer to the possibility of approximating in
measure one class of sets (for example, nonmeasurable sets) by another class of
sets (for example, measurable sets). Lebesgue measure is Borel regular in the sense
that Lebesgue measurable sets can be approximated in measure from the outside
by open sets and from the inside by closed sets, and they can be approximated
by Borel sets up to sets of measure zero. Moreover, there is a simple criterion for
Lebesgue measurability in terms of open and closed sets.
The following theorem expresses a fundamental approximation property of
Lebesgue measurable sets by open and compact sets. Equations (2.9) and (2.10)
are called outer and inner regularity, respectively.
Theorem 2.23. If A Rn , then

(2.9)

(A) = inf {(G) : A G, G open} ,

and if A is Lebesgue measurable, then

(2.10)

(A) = sup {(K) : K A, K compact} .

2.7. BOREL REGULARITY

23

Proof. First, we prove (2.9). The result is immediate if (A) = , so we

suppose that (A) is finite. If A G, then (A) (G), so
(A) inf {(G) : A G, G open} ,

and we just need to prove the reverse inequality,

(2.11)

(A) inf {(G) : A G, G open} .

Let > 0. There is a cover {Ri : i N} of A by rectangles Ri such that

(Ri ) (A) + .
2
i=1

Let Si be an rectangle whose interior Si contains Ri such that

(Si ) (Ri ) + i+1 .

2
Then the collection of open rectangles {Si : i N} covers A and
G=

Si

i=1

is an open set that contains A. Moreover, since {Si : i N} covers G,

(G)
and therefore
(2.12)
It follows that

X
i=1

(Si )

(Ri ) + ,
2
i=1

(G) (A) + .

inf {(G) : A G, G open} (A) + ,

which proves (2.11) since > 0 is arbitrary.
Next, we prove (2.10). If K A, then (K) (A), so

sup {(K) : K A, K compact} (A).

Therefore, we just need to prove the reverse inequality,

(2.13)

(A) sup {(K) : K A, K compact} .

To do this, we apply the previous result to Ac and use the measurability of A.

First, suppose that A is a bounded measurable set, in which case (A) < .
Let F Rn be a compact set that contains A. By the preceding result, for any
> 0, there is an open set G F \ A such that
(G) (F \ A) + .

Then K = F \ G is a compact set such that K A. Moreover, F K G and

F = A (F \ A), so
It follows that

(F ) (K) + (G),

(F ) = (A) + (F \ A).

(A) = (F ) (F \ A)
(F ) (G) +

(K) + ,

24

2. LEBESGUE MEASURE ON Rn

which implies (2.13) and proves the result for bounded, measurable sets.
Now suppose that A is an unbounded measurable set, and define
(2.14)

Ak = {x A : |x| k} .

Then {Ak : k N} is an increasing sequence of bounded measurable sets whose

union is A, so
(2.15)

(Ak ) (A)

as k .

If (A) = , then (Ak ) as k . By the previous result, we can find a

compact set Kk Ak A such that
(Kk ) + 1 (Ak )
so that (Kk ) . Therefore
sup {(K) : K A, K compact} = ,
which proves the result in this case.
Finally, suppose that A is unbounded and (A) < . From (2.15), for any
> 0 we can choose k N such that

(A) (Ak ) + .
2
Moreover, since Ak is bounded, there is a compact set K Ak such that

(Ak ) (K) + .
2
Therefore, for every > 0 there is a compact set K A such that
(A) (K) + ,
which gives (2.13), and completes the proof.

It follows that we may determine the Lebesgue measure of a measurable set in

terms of the Lebesgue measure of open or compact sets by approximating the set
from the outside by open sets or from the inside by compact sets.
The outer approximation in (2.9) does not require that A is measurable. Thus,
for any set A Rn , given > 0, we can find an open set G A such that
(G) (A) < . If A is measurable, we can strengthen this condition to get
that (G \ A) < ; in fact, this gives a necessary and sufficient condition for
measurability.
Theorem 2.24. A subset A Rn is Lebesgue measurable if and only if for
every > 0 there is an open set G A such that
(2.16)

(G \ A) < .

Proof. First we assume that A is measurable and show that it satisfies the
condition given in the theorem.
Suppose that (A) < and let > 0. From (2.12) there is an open set G A
such that (G) < (A) + . Then, since A is measurable,
(G \ A) = (G) (G A) = (G) (A) < ,
which proves the result when A has finite measure.

2.7. BOREL REGULARITY

25

If (A) = , define Ak A as in (2.14), and let > 0. Since Ak is measurable

with finite measure, the argument above shows that for each k N, there is an
open set Gk Ak such that

(Gk \ Ak ) < k .
2
S
Then G =
k=1 Gk is an open set that contains A, and
!

[
X
X

Gk \ A
(Gk \ A)
(Gk \ Ak ) < .
(G \ A) =
k=1

k=1

k=1

Conversely, suppose that A Rn satisfies the condition in the theorem. Let

> 0, and choose an open set G A such that (G \ A) < . If E Rn , we have
E Ac = (E Gc ) (E (G \ A)).

Hence, by the subadditivity and monotonicity of and the measurability of G,

(E A) + (E Ac ) (E A) + (E Gc ) + (E (G \ A))
(E G) + (E Gc ) + (G \ A)
< (E) + .

Since > 0 is arbitrary, it follows that

(E) (E A) + (E Ac )

which proves that A is measurable.

This theorem states that a set is Lebesgue measurable if and only if it can be
approximated from the outside by an open set in such a way that the difference
has arbitrarily small outer Lebesgue measure. This condition can be adopted as
the definition of Lebesgue measurable sets, rather than the Caratheodory definition
which we have used c.f. [5, 8, 11].
The following theorem gives another characterization of Lebesgue measurable
sets, as ones that can be squeezed between open and closed sets.
Theorem 2.25. A subset A Rn is Lebesgue measurable if and only if for
every > 0 there is an open set G and a closed set F such that G A F and
(2.17)

(G \ F ) < .

If (A) < , then F may be chosen to be compact.

Proof. If A satisfies the condition in the theorem, then it follows from the
monotonicity of that (G \ A) (G \ F ) < , so A is measurable by Theorem 2.24.
Conversely, if A is measurable then Ac is measurable, and by Theorem 2.24
given > 0, there are open sets G A and H Ac such that

(G \ A) < ,
(H \ Ac ) < .
2
2
Then, defining the closed set F = H c , we have G A F and
(G \ F ) (G \ A) + (A \ F ) = (G \ A) + (H \ Ac ) < .

Finally, suppose that (A) < and let > 0. From Theorem 2.23, since A is
measurable, there is a compact set K A such that (A) < (K) + /2 and

(A \ K) = (A) (K) < .

2

26

2. LEBESGUE MEASURE ON Rn

As before, from Theorem 2.24 there is an open set G A such that

(G) < (A) + /2.
It follows that G A K and
(G \ K) = (G \ A) + (A \ K) < ,
which shows that we may take F = K compact when A has finite measure.

From the previous results, we can approximate measurable sets by open or

closed sets, up to sets of arbitrarily small but, in general, nonzero measure. By
taking countable intersections of open sets or countable unions of closed sets, we
can approximate measurable sets by Borel sets, up to sets of measure zero
Definition 2.26. The collection of sets in Rn that are countable intersections
of open sets is denoted by G (Rn ), and the collection of sets in Rn that are countable
unions of closed sets is denoted by F (Rn ).
G and F sets are Borel. Thus, it follows from the next result that every
Lebesgue measurable set can be approximated up to a set of measure zero by a
Borel set. This is the Borel regularity of Lebesgue measure.
Theorem 2.27. Suppose that A Rn is Lebesgue measurable. Then there exist
sets G G (Rn ) and F F (Rn ) such that
G A F,

(G \ A) = (A \ F ) = 0.

Proof. For each k N, choose an open set Gk and a closed set Fk such that
Gk A Fk and
1
(Gk \ Fk )
k
Then

\
[
G=
Gk ,
F =
Fk
k=1

are G and F sets with the required properties.

k=1

In particular, since any measurable set can be approximated up to a set of

measure zero by a G or an F , the complexity of the transfinite construction of
general Borel sets illustrated in (2.8) is hidden inside sets of Lebesgue measure
zero.
As a corollary of this result, we get that the Lebesgue -algebra is the completion of the Borel -algebra with respect to Lebesgue measure.
Theorem 2.28. The Lebesgue -algebra L(Rn ) is the completion of the Borel
-algebra B(Rn ).
Proof. Lebesgue measure is complete from Proposition 2.12. By the previous
theorem, if A Rn is Lebesgue measurable, then there is a F set F A such that
M = A \ F has Lebesgue measure zero. It follows by the approximation theorem
that there is a Borel set N G with (N ) = 0 and M N . Thus, A = F M
where F B and M N B with (N ) = 0, which proves that L(Rn ) is the
completion of B(Rn ) as given in Theorem 1.15.


2.8. LINEAR TRANSFORMATIONS

27

2.8. Linear transformations

The definition of Lebesgue measure is not rotationally invariant, since we used
rectangles whose sides are parallel to the coordinate axes. In this section, we show
that the resulting measure does not, in fact, depend upon the direction of the
coordinate axes and is invariant under orthogonal transformations. We also show
that Lebesgue measure transforms under a linear map by a factor equal to the
absolute value of the determinant of the map.
As before, we use to denote Lebesgue outer measure defined using rectangles
whose sides are parallel to the coordinate axes; a set is Lebesgue measurable if it
satisfies the Caratheodory criterion (2.8) with respect to this outer measure. If
T : Rn Rn is a linear map and E Rn , we denote the image of E under T by
T E = {T x Rn : x E} .

First, we consider the Lebesgue measure of rectangles whose sides are not paral we denote
lel to the coordinate axes. We use a tilde to denote such rectangles by R;
closed rectangles whose sides are parallel to the coordinate axes by R as before.
and R as oblique and parallel rectangles, respectively. We denote
We refer to R
by v(R),
i.e. the product of the lengths of its sides, to
the volume of a rectangle R
We know that (R) = v(R) for
avoid confusion with its Lebesgue measure (R).

parallel rectangles, and that R is measurable since it is closed, but we have not yet
= v(R)
for oblique rectangles.
shown that (R)
More explicitly, we regard Rn as a Euclidean space equipped with the standard
inner product,
(x, y) =

n
X

xi yi ,

x = (x1 , x2 , . . . , xn ),

y = (y1 , y2 , . . . , yn ).

i=1

If {e1 , e2 , . . . , en } is the standard orthonormal basis of Rn ,

e1 = (1, 0, . . . , 0),

e2 = (0, 1, . . . , 0), . . . en = (0, 0, . . . , 1),

and {
e1 , e2 , . . . , en } is another orthonormal basis, then we use R to denote rectangles
to denote rectangles whose sides are parallel
whose sides are parallel to {ei } and R
n
to {
ei }. The linear map Q : R Rn defined by Qei = ei is orthogonal, meaning
that QT = Q1 and
(Qx, Qy) = (x, y)

for all x, y Rn .

= QR
Since Q preserves lengths and angles, it maps a rectangle R to a rectangle R
= v(R).
such that v(R)
We will use the following lemma.
contains a finite almost disjoint collecLemma 2.29. If an oblique rectangle R
tion of parallel rectangles {R1 , R2 , . . . , RN } then
N
X
i=1

v(Ri ) v(R).

This result is geometrically obvious, but a formal proof seems to require a fuller
discussion of the volume function on elementary geometrical sets, which is included
in the theory of valuations in convex geometry. We omit the details.

2. LEBESGUE MEASURE ON Rn

28

is an oblique rectangle, then given any > 0 there is

Proposition 2.30. If R
and satisfies
a collection of parallel rectangles {Ri : i N} that covers R

X
i=1

+ .
v(Ri ) v(R)

in its interior such that

Proof. Let S be an oblique rectangle that contains R
v(R)
+ .
v(S)
Then, from Proposition 2.20, we may decompose the interior of S into an almost
disjoint union of parallel rectangles
S =

Ri .

i=1

It follows from the previous lemma that for every N N

N
X
i=1

which implies that

X
i=1

v(Ri ) v(S),

v(R)
+ .
v(Ri ) v(S)

since its union is S , which contains R.


Moreover, the collection {Ri } covers R
Conversely, by reversing the roles of the axes, we see that if R is a parallel
i : i N}
rectangle and > 0, then there is a cover of R by oblique rectangles {R
such that

X
i ) v(R) + .
(2.18)
v(R
i=1

Theorem 2.31. If E R and Q : Rn Rn is an orthogonal transformation,

then
(QE) = (E),

and E is Lebesgue measurable if an only if QE is Lebesgue measurable.

= QE. Given > 0 there is a cover of E
by parallel rectangles
Proof. Let E
{Ri : i N} such that

X
+ .
v(Ri ) (E)
2
i=1

i,j : j N} of Ri by oblique
From (2.18), for each i N we can choose a cover {R
rectangles such that

X
i,j ) v(Ri ) + .
v(R
2i+1
i=1
i,j : i, j N} is a countable cover of E
by oblique rectangles, and
Then {R

i,j=1

i,j )
v(R

X
i=1

v(Ri ) +

+ .
(E)
2

2.8. LINEAR TRANSFORMATIONS

29

i,j , then {Ri,j : j N} is a cover of E by parallel rectangles, so

If Ri,j = QT R
(E)

v(Ri,j ).

i,j=1

i,j ). It follows that

Moreover, since Q is orthogonal, we have v(Ri,j ) = v(R
(E)

i,j=1

v(Ri,j ) =

i,j=1

i,j ) (E)
+ ,
v(R

and since > 0 is arbitrary, we conclude that

(E) (E).

we get the
By applying the same argument to the inverse mapping E = QT E,

reverse inequality, and it follows that (E) = (E).
Since is invariant under Q, the Caratheodory criterion for measurability is
invariant, and E is measurable if and only if QE is measurable.

It follows from Theorem 2.31 that Lebesgue measure is invariant under rotations
and reflections.4 Since it is also invariant under translations, Lebesgue measure is
invariant under all isometries of Rn .
Next, we consider the effect of dilations on Lebesgue measure. Arbitrary linear
maps may then be analyzed by decomposing them into rotations and dilations.
Proposition 2.32. Suppose that : Rn Rn is the linear transformation
(2.19)

: (x1 , x2 , . . . , xn ) 7 (1 x1 , 2 x2 , . . . , n xn )

where the i > 0 are positive constants. Then

(E) = (det ) (E),
and E is Lebesgue measurable if and only if E is Lebesgue measurable.
Proof. The diagonal map does not change the orientation of a rectangle, so it maps a cover of E by parallel rectangles to a cover of E by parallel rectangles, and conversely. Moreover, multiplies the volume of a rectangle
by det = 1 . . . n , so it immediate from the definition of outer measure that
(E) = (det ) (E), and E satisfies the Caratheodory criterion for measurability if and only if E does.

Theorem 2.33. Suppose that T : Rn Rn is a linear transformation and
E Rn . Then
(T E) = |det T | (E),

and T E is Lebesgue measurable if E is measurable

Proof. If T is singular, then its range is a lower-dimensional subspace of Rn ,

which has Lebesgue measure zero, and its determinant is zero, so the result holds.5
We therefore assume that T is nonsingular.
4Unlike differential volume forms, Lebesgue measure does not depend on the orientation of
Rn ; such measures are sometimes referred to as densities in differential geometry.
5In this case T E, is always Lebesgue measurable, with measure zero, even if E is not
measurable.

2. LEBESGUE MEASURE ON Rn

30

In that case, according to the polar decomposition, the map T may be written
as a composition
T = QU

of a positive definite, symmetric map U = T T T and an orthogonal map Q. Any

positive-definite, symmetric map U may be diagonalized by an orthogonal map O
to get
U = OT O
where : Rn Rn has the form (2.19). From Theorem 2.31, orthogonal mappings
leave the Lebesgue measure of a set invariant, so from Proposition 2.32
(T E) = (E) = (det ) (E).
Since | det Q| = 1 for any orthogonal map Q, we have det = | det T |, and it follows
that (T E) = |det T | (E).
Finally, it is straightforward to see that T E is measurable if E is measurable.

2.9. Lebesgue-Stieltjes measures
We briefly consider a generalization of one-dimensional Lebesgue measure,
called Lebesgue-Stieltjes measures on R. These measures are obtained from an
increasing, right-continuous function F : R R, and assign to a half-open interval
(a, b] the measure
F ((a, b]) = F (b) F (a).

The use of half-open intervals is significant here because a Lebesgue-Stieltjes measure may assign nonzero measure to a single point. Thus, unlike Lebesgue measure,
we need not have F ([a, b]) = F ((a, b]). Half-open intervals are also convenient
because the complement of a half-open interval is a finite union of (possibly infinite) half-open intervals of the same type. Thus, the collection of finite unions of
half-open intervals forms an algebra.
The right-continuity of F is consistent with the use of intervals that are halfopen at the left, since

\
(a, a + 1/i] = ,
i=1

so, from (1.2), if F is to define a measure we need

lim F ((a, a + 1/i]) = 0

or
lim [F (a + 1/i) F (a)] = lim F (x) F (a) = 0.

xa+

Conversely, as we state in the next theorem, any such function F defines a Borel
measure on R.
Theorem 2.34. Suppose that F : R R is an increasing, right-continuous
function. Then there is a unique Borel measure F : B(R) [0, ] such that
F ((a, b]) = F (b) F (a)
for every a < b.

2.9. LEBESGUE-STIELTJES MEASURES

31

The construction of F is similar to the construction of Lebesgue measure on

Rn . We define an outer measure F : P(R) [0, ] by
)
(
X
S

[F (bi ) F (ai )] : E i=1 (ai , bi ] ,

F (E) = inf
i=1

and restrict
to its Caratheodory measurable sets, which include the Borel sets.
See e.g. Section 1.5 of Folland [4] for a detailed proof.
The following examples illustrate the three basic types of Lebesgue-Stieltjes
measures.
Example 2.35. If F (x) = x, then F is Lebesgue measure on R with
F ((a, b]) = b a.
Example 2.36. If

then F

1 if x 0,
0 if x < 0,
is the -measure supported at 0,

1 if 0 A,
F (A) =
0 if 0
/ A.
F (x) =

Example 2.37. If F : R R is the Cantor function, then F assigns measure

one to the Cantor set, which has Lebesgue measure zero, and measure zero to its
complement. Despite the fact that F is supported on a set of Lebesgue measure
zero, the F -measure of any countable set is zero.

CHAPTER 3

Measurable functions
Measurable functions in measure theory are analogous to continuous functions
in topology. A continuous function pulls back open sets to open sets, while a
measurable function pulls back measurable sets to measurable sets.
3.1. Measurability
Most of the theory of measurable functions and integration does not depend
on the specific features of the measure space on which the functions are defined, so
we consider general spaces, although one should keep in mind the case of functions
defined on R or Rn equipped with Lebesgue measure.
Definition 3.1. Let (X, A) and (Y, B) be measurable spaces. A function
f : X Y is measurable if f 1 (B) A for every B B.
Note that the measurability of a function depends only on the -algebras; it is
not necessary that any measures are defined.
In order to show that a function is measurable, it is sufficient to check the
measurability of the inverse images of sets that generate the -algebra on the target
space.
Proposition 3.2. Suppose that (X, A) and (Y, B) are measurable spaces and
B = (G) is generated by a family G P(Y ). Then f : X Y is measurable if
and only if
f 1 (G) A
for every G G.
Proof. Set operations are natural under pull-backs, meaning that
f 1 (Y \ B) = X \ f 1 (B)

and
f

i=1

It follows that

Bi

(Bi ) ,

i=1

i=1

Bi

f 1 (Bi ) .

i=1



M = B Y : f 1 (B) A
is a -algebra on Y . By assumption, M G and therefore M (G) = B, which
implies that f is measurable.

It is worth noting the indirect nature of the proof of containment of -algebras
in the previous proposition; this is required because we typically cannot use an
explicit representation of sets in a -algebra. For example, the proof does not
characterize M, which may be strictly larger than B.
If the target space Y is a topological space, then we always equip it with the
Borel -algebra B(Y ) generated by the open sets (unless stated explicitly otherwise).
33

34

3. MEASURABLE FUNCTIONS

In that case, it follows from Proposition 3.2 that f : X Y is measurable if and

only if f 1 (G) A is a measurable subset of X for every set G that is open in Y . In
particular, every continuous function between topological spaces that are equipped
with their Borel -algebras is measurable. The class of measurable function is,
however, typically much larger than the class of continuous functions, since we only
require that the inverse image of an open set is Borel; it need not be open.
3.2. Real-valued functions
We specialize to the case of real-valued functions
or extended real-valued functions

f :X R
f : X R.

We will consider one case or the other as convenient, and comment on any differences. A positive extended real-valued function is a function
f : X [0, ].

Note that we allow a positive function to take the value zero.

We equip R and R with their Borel -algebras B(R) and B(R). A Borel subset
of R has one of the forms
B,

B {},

B {},

B {, }

where B is a Borel subset of R. As Example 2.22 shows, sets that are Lebesgue
measurable but not Borel measurable need not be well-behaved under the inverse
of even a monotone function, which helps explain why we do not include them in
the range -algebra on R or R.
By contrast, when the domain of a function is a measure space it is often
convenient to use a complete space. For example, if the domain is Rn we typically
equip it with the Lebesgue -algebra, although if completeness is not required
we may use the Borel -algebra. With this understanding, we get the following
definitions. We state them for real-valued functions; the definitions for extended
real-valued functions are completely analogous
Definition 3.3. If (X, A) is a measurable space, then f : X R is measurable
if f 1 (B) A for every Borel set B B(R). A function f : Rn R is Lebesgue
measurable if f 1 (B) is a Lebesgue measurable subset of Rn for every Borel subset
B of R, and it is Borel measurable if f 1 (B) is a Borel measurable subset of Rn
for every Borel subset B of R
This definition ensures that continuous functions f : Rn R are Borel measurable and functions that are equal a.e. to Borel measurable functions are Lebesgue
measurable. If f : R R is Borel measurable and g : Rn R is Lebesgue (or
Borel) measurable, then the composition f g is Lebesgue (or Borel) measurable
since


1
(f g) (B) = g 1 f 1 (B) .
Note that if f is Lebesgue measurable, then f g need not be measurable since
f 1 (B) need not be Borel even if B is Borel.
We can give more easily verifiable conditions for measurability in terms of
generating families for Borel sets.

3.2. REAL-VALUED FUNCTIONS

35

Proposition 3.4. The Borel -algebra on R is generated by any of the following collections of intervals
{(, b) : b R} ,

{(, b] : b R} ,

{(a, ) : a R} ,

{[a, ) : a R} .

Proof. The -algebra generated by intervals of the form (, b) is contained

in the Borel -algebra B(R) since the intervals are open sets. Conversely, the
-algebra contains complementary closed intervals of the form [a, ), half-open
intersections [a, b), and countable intersections
[a, b] =

[a, b +

n=1

1
).
n

From Proposition 2.20, the Borel -algebra B(R) is generated by the collection of
closed rectangles [a, b], so
({(, b) : b R}) = B(R).
The proof for the other collections is similar.

The properties given in the following proposition are sometimes taken as the
definition of a measurable function.
Proposition 3.5. If (X, A) is a measurable space, then f : X R is measurable if and only if one of the following conditions holds:
{x X : f (x) < b} A

for every b R;

{x X : f (x) > a} A

for every a R;

{x X : f (x) b} A

{x X : f (x) a} A

for every b R;

for every a R.

Proof. Note that, for example,

{x X : f (x) < b} = f 1 ((, b))
and the result follows immediately from Propositions 3.2 and 3.4.

If any one of these equivalent conditions holds, then f 1 (B) A for every set
B B(R). We will often use a shorthand notation for sets, such as
{f < b} = {x X : f (x) < b} .
The Borel -algebra on R is generated by intervals of the form [, b), [, b],
(a, ], or [a, ] where a, b R, and exactly the same conditions as the ones
in Proposition 3.5 imply the measurability of an extended real-valued functions
f : X R. In that case, we can allow a, b R to be extended real numbers
in Proposition 3.5, but it is not necessary to do so in order to imply that f is
measurable.
Measurability is well-behaved with respect to algebraic operations.
Proposition 3.6. If f, g : X R are real-valued measurable functions and
k R, then
kf, f + g,
f g,
f /g
are measurable functions, where we assume that g 6= 0 in the case of f /g.

36

3. MEASURABLE FUNCTIONS

Proof. If k > 0, then

{kf < b} = {f < b/k}
so kf is measurable, and similarly if k < 0 or k = 0. We have
[
{f + g < b} =
{f < q} {g < r}
q+r<b;q,rQ

so f + g is measurable. The function f 2 is measurable since if b 0

o
 2
n
f <b = b<f < b .

It follows that

fg =


1
(f + g)2 f 2 g 2
2

is measurable. Finally, if g 6= 0

if b < 0,
{1/b < g < 0}
{1/g < b} =
{ < g < 0}
if b = 0,

{ < g < 0} {1/b < g < } if b > 0,

so 1/g is measurable and therefore f /g is measurable.

An analogous result applies to extended real-valued functions provided that

they are well-defined. For example, f + g is measurable provided that f (x), g(x)
are not simultaneously equal to and , and f g is is measurable provided that
f (x), g(x) are not simultaneously equal to 0 and .
Proposition 3.7. If f, g : X R are extended real-valued measurable functions, then
|f |,
max(f, g),
min(f, g)

are measurable functions.

Proof. We have

{max(f, g) < b} = {f < b} {g < b} ,

{min(f, g) < b} = {f < b} {g < b} ,

and |f | = max(f, 0) min(f, 0), from which the result follows.

3.3. Pointwise convergence

Crucially, measurability is preserved by limiting operations on sequences of
functions. Operations in the following theorem are understood in a pointwise sense;
for example,


sup fn (x) = sup {fn (x)} .

nN

nN

Theorem 3.8. If {fn : n N} is a sequence of measurable functions fn : X

R, then
sup fn ,

nN

inf fn ,

nN

lim sup fn ,
n

are measurable extended real-valued functions on X.

lim inf fn
n

3.4. SIMPLE FUNCTIONS

37

Proof. We have for every b R that




\
sup fn b =
{fn b} ,
nN

inf fn < b

nN

n=1

n=1

{fn < b}

so the supremum and infimum are measurable Moreover, since

lim sup fn = inf sup fk ,
nN kn

lim inf fn = sup inf fk

n

nN kn

it follows that the limsup and liminf are measurable.

Perhaps the most important way in which new functions arise from old ones is
by pointwise convergence.
Definition 3.9. A sequence {fn : n N} of functions fn : X R converges
pointwise to a function f : X R if fn (x) f (x) as n for every x X.
Pointwise convergence preserves measurability (unlike continuity, for example).
This fact explains why the measurable functions form a sufficiently large class for
the needs of analysis.
Theorem 3.10. If {fn : n N} is a sequence of measurable functions fn :
X R and fn f pointwise as n , then f : X R is measurable.
Proof. If fn f pointwise, then
f = lim sup fn = lim inf fn
n

so the result follows from the previous proposition.

3.4. Simple functions

The characteristic function (or indicator function) of a subset E X is the
function E : X R defined by

1 if x E,
E (x) =
0 if x
/ E.
The function E is measurable if and only if E is a measurable set.
Definition 3.11. A simple function : X R on a measurable space (X, A)
is a function of the form
(3.1)

(x) =

N
X

cn En (x)

n=1

where c1 , . . . , cN R and E1 , . . . , EN A.
Note that, according to this definition, a simple function is measurable. The
representation of in (3.1) is not unique; we call it a standard representation if the
constants cn are distinct and the sets En are disjoint.

38

3. MEASURABLE FUNCTIONS

Theorem 3.12. If f : X [0, ] is a positive measurable function, then there

is a monotone increasing sequence of positive simple functions n : X [0, )
with 1 2 n . . . such that n f pointwise as n . If f is
bounded, then n f uniformly.
Proof. For each n N, we divide the interval [0, 2n ] in the range of f into
22n subintervals of width 2n ,
Ik,n = (k2n , (k + 1)2n ],

k = 0, 1, . . . , 22n 1,

let Jn = (2n , ] be the remaining part of the range, and define

Ek,n = f 1 (Ik,n ),

Fn = f 1 (Jn ).

Then the sequence of simple functions given by

n =

22n
1
X

k2n Ek,n + 2n Fn

k=0

has the required properties.

In defining the Lebesgue integral of a measurable function, we will approximate

it by simple functions. By contrast, in defining the Riemann integral of a function
f : [a, b] R, we partition the domain [a, b] into subintervals and approximate f by
step functions that are constant on these subintervals. This difference is sometime
expressed by saying that in the Lebesgue integral we partition the range, and in
the Riemann integral we partition the domain.
3.5. Properties that hold almost everywhere
Often, we want to consider functions or limits which are defined outside a set of
measure zero. In that case, it is convenient to deal with complete measure spaces.
Proposition 3.13. Let (X, A, ) be a complete measure space and f, g : X
R. If f = g pointwise -a.e. and f is measurable, then g is measurable.
Proof. Suppose that f = g on N c where N is a set of measure zero. Then
{g < b} = ({f < b} N c ) ({g < b} N ) .

Each of these sets is measurable: {f < b} is measurable since f is measurable; and

{g < b} N is measurable since it is a subset of a set of measure zero and X is
complete.

The completeness of X is essential in this proposition. For example, if X is not
complete and E N is a non-measurable subset of a set N of measure zero, then
the functions 0 and E are equal almost everywhere, but 0 is measurable and E
is not.
Proposition 3.14. Let (X, A, ) be a complete measure space. If {fn : n N}
is a sequence of measurable functions fn : X R and fn f as n pointwise
-a.e., then f is measurable.
Proof. Since fn is measurable, g = lim supn fn is measurable and f = g
pointwise a.e., so the result follows from the previous proposition.


CHAPTER 4

Integration
In this Chapter, we define the integral of real-valued functions on an arbitrary
measure space and derive some of its basic properties. We refer to this integral as
the Lebesgue integral, whether or not the domain of the functions is subset of Rn
equipped with Lebesgue measure. The Lebesgue integral applies to a much wider
class of functions than the Riemann integral and is better behaved with respect to
pointwise convergence. We carry out the definition in three steps: first for positive
simple functions, then for positive measurable functions, and finally for extended
real-valued measurable functions.

4.1. Simple functions

Suppose that (X, A, ) is a measure space.
Definition 4.1. If : X [0, ) is a positive simple function, given by
=

N
X

ci E i

i=1

where ci 0 and Ei A, then the integral of with respect to is

(4.1)

d =

N
X

ci (Ei ) .

i=1

In (4.1), we use the convention that if ci = 0 and (Ei ) = , then 0 = 0,

meaning that the integral of 0 over a set of measure is equal to 0. The integral
may take the value (if ci > 0 and (Ei ) = for some 1 i N ). One
can verify that the value of the integral in (4.1) is independent of how the simple
function is represented as a linear combination of characteristic functions.
Example 4.2. The characteristic function Q : R R of the rationals is not
Riemann integrable on any compact interval of non-zero length, but it is Lebesgue
integrable with
Z
Q d = 1 (Q) = 0.
The integral of simple functions has the usual properties of an integral. In
particular, it is linear, positive, and monotone.
39

40

4. INTEGRATION

Proposition 4.3. If , : X [0, ) are positive simple functions on a

measure space X, then:
Z
Z
k d = k d
if k [0, );
Z
Z
Z
( + ) d = d + d;
Z
Z
0 d d
if 0 .
Proof. These follow immediately from the definition.

4.2. Positive functions

We define the integral of a measurable function by splitting it into positive and
negative parts, so we begin by defining the integral of a positive function.
Definition 4.4. If f : X [0, ] is a positive, measurable, extended realvalued function on a measure space X, then
Z

Z
f d = sup
d : 0 f , simple .

A positive function f : X [0, ] is integrable if it is measurable and

Z
f d < .

In this definition, we approximate the function f from below by simple functions. In contrast with the definition of the Riemann integral, it is not necessary to
approximate a measurable function from both above and below in order to define
its integral.
If A X is a measurable set and f : X [0, ] is measurable, we define
Z
Z
f d = f A d.
A

Unlike the Riemann integral, where the definition of the integral over non-rectangular
subsets of R2 already presents problems, it is trivial to define the Lebesgue integral
over arbitrary measurable subsets of a set on which it is already defined.
The following properties are an immediate consequence of the definition and
the corresponding properties of simple functions.
Proposition 4.5. If f, g : X [0, ] are positive, measurable, extended realvalued function on a measure space X, then:
Z
Z
kf d = k f d
if k [0, );
Z
Z
0 f d g d
if 0 f g.

The integral is also linear, but this is not immediately obvious from the definition and it depends on the measurability of the functions. To show the linearity,
we will first derive one of the fundamental convergence theorem for the Lebesgue
integral, the monotone convergence theorem. We discuss this theorem and its applications in greater detail in Section 4.5.

4.2. POSITIVE FUNCTIONS

41

Theorem 4.6 (Monotone Convergence Theorem). If {fn : n N} is a monotone increasing sequence

0 f1 f2 fn fn+1 . . .
of positive, measurable, extended real-valued functions fn : X [0, ] and
f = lim fn ,
n

then
lim

fn d =

f d.

Proof. The pointwise limit f : X [0, ] exists since the sequence {fn }
is increasing. Moreover, by the monotonicity of the integral, the integrals are
increasing, and
Z
Z
Z
fn d fn+1 d f d,

so the limit of the integrals exists, and

Z
Z
lim
fn d f d.
n

To prove the reverse inequality, let : X [0, ) be a simple function with

0 f . Fix 0 < t < 1. Then
An = {x X : fn (x) t(x)}
is an increasing sequence of measurable sets A1 A2 An . . . whose
union is X. It follows that
Z
Z
Z
d.
fn d t
(4.2)
fn d
An

An

Moreover, if
=

N
X

ci E i

i=1

we have from the monotonicity of in Proposition 1.12 that

Z
Z
N
N
X
X
ci (Ei ) = d
ci (Ei An )
d =
An

i=1

i=1

as n . Taking the limit as n in (4.2), we therefore get

Z
Z
lim
fn d t d.
n

Since 0 < t < 1 is arbitrary, we conclude that

Z
Z
lim
fn d d,
n

and since f is an arbitrary simple function, we get by taking the supremum

over that
Z
Z
lim
fn d f d.
n

This proves the theorem.

42

4. INTEGRATION

In particular, this theorem implies that we can obtain the integral of a positive
measurable function f as a limit of integrals of an increasing sequence of simple
functions, not just as a supremum over all simple functions dominated by f as in
Definition 4.4. As shown in Theorem 3.12, such a sequence of simple functions
always exists.
Proposition 4.7. If f, g : X [0, ] are positive, measurable functions on a
measure space X, then
Z
Z
Z
(f + g) d = f d + g d.

Proof. Let {n : n N} and {n : n N} be increasing sequences of positive

simple functions such that n f and n g pointwise as n . Then n + n
is an increasing sequence of positive simple functions such that n + n f + g.
It follows from the monotone convergence theorem (Theorem 4.6) and the linearity
of the integral on simple functions that
Z
Z
(f + g) d = lim
(n + n ) d
n
Z

Z
= lim
n d + n d
n
Z
Z
= lim
n d + lim
n d
n
n
Z
Z
= f d + g d,
which proves the result.

4.3. Measurable functions

If f : X R is an extended real-valued function, we define the positive and
negative parts f + , f : X [0, ] of f by
(4.3)

f = f + f ,

f + = max{f, 0},

f = max{f, 0}.

For this decomposition,

|f | = f + + f .
Note that f is measurable if and only if f + and f are measurable.
Definition 4.8. If f : X R is a measurable function, then
Z
Z
Z
f d = f + d f d,
R
R
provided that at least Rone of theR integrals f + d, f d is finite. The function
f is integrable if both f + d, f d are finite, which is the case if and only if
Z
|f | d < .

Note that, according to Definition

4.8,R the integral may take the values or
R
, but it is not defined if both f + d, f d are infinite. Thus, although the
integral of a positive measurable function always exists as an extended real number,
the integral of a general, non-integrable real-valued measurable function may not
exist.

4.3. MEASURABLE FUNCTIONS

43

This Lebesgue integral has all the usual properties of an integral. We restrict
attention to integrable functions to avoid undefined expressions involving extended
real numbers such as .
Proposition 4.9. If f, g : X R are integrable functions, then:
Z
Z
kf d = k f d
if k R;
Z
Z
Z
(f + g) d = f d + g d;
Z
Z
f d g d
if f g;
Z
 Z


 f d |f | d.



Proof. These results follow by writing functions into their positive and negative parts, as in (4.3), and using the results for positive functions.
If f = f + f and k 0, then (kf )+ = kf + and (kf ) = kf , so
Z
Z
Z
Z
Z
Z
kf d = kf + d kf d = k f + d k f d = k f d.

Similarly, (f )+ = f and (f ) = f + , so
Z
Z
Z
Z
(f ) d = f d f + d = f d.
If h = f + g and

f = f + f ,

g = g+ g,

h = h+ h

are the decompositions of f, g, h into their positive and negative parts, then
h+ h = f + f + g + g .

It need not be true that h+ = f + + g + , but we have

f + g + h+ = f + + g + + h .
The linearity of the integral on positive functions gives
Z
Z
Z
Z
Z
Z
f d + g d + h+ d = f + d + g + d + h d,

which implies that

Z
Z
Z
Z
Z
Z
+

+
h d h d = f d f d + g d g d,

R
R
R
or (f + g) d = f d + g d.
It follows that if f g, then
Z
Z
Z
0 (g f ) d = g d f d,

R
R
so f d g d. The last result is then a consequence of the previous results
and |f | f |f |.

Let us give two basic examples of the Lebesgue integral.

44

4. INTEGRATION

Example 4.10. Suppose that X = N and : P(N) [0, ] is counting

measure on N. If f : N R and f (n) = xn , then
Z

X
f d =
xn ,
N

n=1

where the integral is finite if and only if the series is absolutely convergent. Thus,
the theory of absolutely convergent series is a special case of the Lebesgue integral.
Note that a conditionally convergent series, such as the alternating harmonic series,
does not correspond to a Lebesgue integral, since both its positive and negative
parts diverge.

Example 4.11. Suppose that X = [a, b] is a compact interval and : L([a, b])
R is Lesbegue measure on [a, b]. We note in Section 4.8 that any Riemann integrable function f : [a, b] R is integrable with respect to Lebesgue measure , and
its Riemann integral is equal to the Lebesgue integral,
Z b
Z
f d.
f (x) dx =
[a,b]

Thus, all of the usual integrals from elementary calculus remain valid for the
Lebesgue integral on R. We will write an integral with respect to Lebesgue measure
on R, or Rn , as
Z
f dx.

Even though the class of Lebesgue integrable functions on an interval is wider

than the class of Riemann integrable functions, some improper Riemann integrals
may exist even though the Lebesegue integral does not.
Example 4.12. The integral
Z 1
0

1
1
1
sin + cos
x
x
x

dx

is not defined as a Lebesgue integral, although the improper Riemann integral




Z 1
Z 1
1
1
d
1
1
lim
dx = lim+
x cos
dx = cos 1
sin + cos
x
x
x
x
0+
0
dx
exists.

Example 4.13. The integral

Z

1
dx
x

is not defined as a Lebesgue integral, although the principal value integral

Z

Z 1
Z 1
1
1
1
p.v.
dx = lim+
dx +
dx = 0
0
1 x
1 x
x
exists. Note, however, that the Lebesgue integrals
Z 1
Z 0
1
1
dx = ,
dx =
0 x
1 x

are well-defined as extended real numbers.

4.4. ABSOLUTE CONTINUITY

45

The inability of the Lebesgue integral to deal directly with the cancelation
between large positive and negative parts in oscillatory or singular integrals, such
as the ones in the previous examples, is sometimes viewed as a defect (although the
integrals above can still be defined as an appropriate limit of Lebesgue integrals).
Other definitions of the integral such as the Henstock-Kurzweil integral, which is a
generalization of the Riemann integral, avoid this defect but they have not proved
to be as useful as the Lebesgue integral. Similar issues arise in connection with
Feynman path integrals in quantum theory, where one would like to define the
integral of highly oscillatory functionals on an infinite-dimensional function-space.
4.4. Absolute continuity
The following results show that a function with finite integral is finite a.e. and
that the integral depends only on the pointwise a.e. values of a function.
R

Proposition 4.14. If f : X R is an integrable function, meaning that

|f | d < , then f is finite -a.e. on X.

Proof. We may assume that f is positive without loss of generality. Suppose

that
E = {x X : f = }
has nonzero measure. Then for any t > 0, we have f > tE , so
Z
Z
f d tE d = t(E),
R
which implies that f d = .

Proposition 4.15.R Suppose that f : X R is an extended real-valued measurable function. Then |f | d = 0 if and only if f = 0 -a.e.

Proof. By replacing f with |f |, we can assume that f is positive without loss

of generality. Suppose that f = 0 a.e. If 0 f is a simple function,
=

N
X

ci E i ,

i=1

then = 0 a.e., so ci = 0 unless (Ei ) = 0. It follows that

Z
N
X
ci (Ei ) = 0,
d =
i=1

R
and Definition 4.4 implies thatR f d = 0.
Conversely, suppose that f d = 0. For n N, let
En = {x X : f (x) 1/n} .

Then 0 (1/n)En f , so that

Z
Z
1
1
0 (En ) =
En d f d = 0,
n
n
and hence (En ) = 0. Now observe that

[
{x X : f (x) > 0} =
En ,
n=1

so it follows from the countable additivity of that f = 0 a.e.

46

4. INTEGRATION

In particular, it follows that if f : X R is any measurable function, then

Z
(4.4)
f d = 0
if (A) = 0.
A

For integrable functions we can strengthen the previous result to get the following property, which is called the absolute continuity of the integral.
Proposition
4.16. Suppose that f : X R is an integrable function, meaning
R
|f | d < . Then, given any > 0, there exists > 0 such that
Z
(4.5)
0
|f | d <

that

whenever A is a measurable set with (A) < .

Proof. Again, we can assume that f is positive. For n N, define fn : X
[0, ] by

n
if f (x) n,
fn (x) =
f (x) if 0 f (x) < n.

Then {fn } is an increasing sequence of positive measurable functions that converges

pointwise to f . We estimate the integral of f over A as follows:
Z
Z
Z
f d = (f fn ) d +
fn d
A
A
ZA

(f fn ) d + n(A).
X

By the monotone convergence theorem,

Z
Z
f d <
fn d
X

as n . Therefore, given > 0, we can choose n such that

Z

0
(f fn ) d < ,
2
X
and then choose

.
2n
If (A) < , we get (4.5), which proves the result.
=

Proposition 4.16 may fail if f is not integrable.

Example 4.17. Define : B((0, 1)) [0, ] by
Z
1
dx,
(A) =
A x
where the integral is taken with respect to Lebesgue measure . Then (A) = 0 if
(A) = 0, but (4.5) does not hold.
There is a converse to this theorem concerning the representation of absolutely
continuous measures as integrals (the Radon-Nikodym theorem, stated in Theorem 6.27).

4.5. CONVERGENCE THEOREMS

47

4.5. Convergence theorems

One of the most basic questions in integration theory is the following: If fn f
pointwise, when can one say that
Z
Z
(4.6)
fn d f d?

The Riemann integral is not sufficiently general to permit a satisfactory answer to

this question.
Perhaps the simplest condition that guarantees the convergence of the integrals
is that the functions fn : X R converge uniformly to f : X R and X has finite
measure. In that case
Z
 Z
Z


 fn d f d |fn f | d (X) sup |fn f | 0


X

as n . The assumption of uniform convergence is too strong for many purposes,

and the Lebesgue integral allows the formulation of simple and widely applicable
theorems for the convergence of integrals. The most important of these are the
monotone convergence theorem (Theorem 4.6) and the Lebesgue dominated convergence theorem (Theorem 4.24). The utility of these results accounts, in large
part, for the success of the Lebesgue integral.
Some conditions on the functions fn in (4.6) are, however, necessary to ensure
the convergence of the integrals, as can be seen from very simple examples. Roughly
speaking, the convergence may fail because mass can leak out to infinity in the
limit.
Example 4.18. Define fn : R R by

n if 0 < x < 1/n,
fn (x) =
0 otherwise.
Then fn 0 as n pointwise on R, but
Z
fn dx = 1
for every n N.

By modifying this example, and the following ones, we can obtain a sequence fn that
converges pointwise to zero but whose integrals converge to infinity; for example
 2
n if 0 < x < 1/n,
fn (x) =
0
otherwise.
Example 4.19. Define fn : R R by

1/n if 0 < x < n,
fn (x) =
0
otherwise.
Then fn 0 as n pointwise on R, and even uniformly, but
Z
fn dx = 1
for every n N.
Example 4.20. Define fn : R R by

1 if n < x < n + 1,
fn (x) =
0 otherwise.

48

4. INTEGRATION

Then fn 0 as n pointwise on R, but

Z
fn dx = 1
for every n N.
The monotone convergence theorem implies that a similar failure of convergence
of the integrals cannot occur in an increasing sequence of functions, even if the
convergence is not uniform or the domain space does not have finite measure. Note
that the monotone convergence theorem does not hold for the Riemann integral;
indeed, the pointwise limit of a monotone increasing, bounded sequence of Riemann
integrable functions need not even be Riemann integrable.
Example 4.21. Let {qi : i N} be an enumeration of the rational numbers in
the interval [0, 1], and define fn : [0, 1] [0, ) by

1 if x = qi for some 1 i n,
fn (x) =
0 otherwise.
Then {fn } is a monotone increasing sequence of bounded, positive, Riemann integrable functions, each of which has zero integral. Nevertheless, as n the
sequence converges pointwise to the characteristic function of the rationals in [0, 1],
which is not Riemann integrable.
A useful generalization of the monotone convergence theorem is the following
result, called Fatous lemma.
Theorem 4.22. Suppose that {fn : n N} is sequence of positive measurable
functions fn : X [0, ]. Then
Z
Z
(4.7)
lim inf fn d lim inf fn d.
n

Proof. For each n N, let

gn = inf fk .
kn

Then {gn } is a monotone increasing sequence which converges pointwise to lim inf fn
as n , so by the monotone convergence theorem
Z
Z
(4.8)
lim
gn d = lim inf fn d.
n

Moreover, since gn fk for every k n, we have

Z
Z
gn d inf
fk d,
kn

so that
lim

gn d lim inf
n

Using (4.8) in this inequality, we get the result.

fn d.


We may have strict inequality in (4.7), as in the previous examples. The monotone convergence theorem and Fatous Lemma enable us to determine the integrability of functions.

4.5. CONVERGENCE THEOREMS

49

Example 4.23. For R, consider the function f : [0, 1] [0, ] defined by


x
if 0 < x 1,
f (x) =

if x = 0.

For n N, let

x if 1/n x 1,
n
if 0 x < 1/n.
Then {fn } is an increasing sequence of Lebesgue measurable functions (e.g since
fn is continuous) that converges pointwise to f . We denote the integral of f with
R1
respect to Lebesgue measure on [0, 1] by 0 f (x) dx. Then, by the monotone convergence theorem,
Z 1
Z 1
f (x) dx = lim
fn (x) dx.
fn (x) =

From elementary calculus,

fn (x) dx

1
1

as n if < 1, and to if 1. Thus, f is integrable on [0, 1] if and only if

< 1.
Perhaps the most frequently used convergence result is the following dominated
convergence theorem, in which all the integrals are necessarily finite.
Theorem 4.24 (Lebesgue Dominated Convergence Theorem). If {fn : n N}
is a sequence of measurable functions fn : X R such that fn f pointwise,
and
R
|fn | g where g : X [0, ] is an integrable function, meaning that g d < ,
then
Z
Z
fn d

f d

as n .

Proof. Since g + fn 0 for every n N, Fatous lemma implies that

Z
Z
Z
Z
(g + f ) d lim inf (g + fn ) d g d + lim inf fn d,
n

which gives

f d lim inf
n

fn d.

Similarly, g fn 0, so
Z
Z
Z
Z
(g f ) d lim inf (g fn ) d g d lim sup fn d,
n

which gives

and the result follows.

f d lim sup
n

fn d,


An alternative, and perhaps more illuminating, proof of the dominated convergence theorem may be obtained from Egoroffs theorem and the absolute continuity
of the integral. Egoroffs theorem states that if a sequence {fn } of measurable functions, defined on a finite measure space (X, A, ), converges pointwise to a function
f , then for every > 0 there exists a measurable set A X such that {fn } converges
uniformly to f on A and (X \ A) < . The uniform integrability of the functions

50

4. INTEGRATION

and the absolute continuity of the integral imply that what happens off the set A
may be made to have arbitrarily small effect on the integrals. Thus, the convergence
theorems hold because of this almost uniform convergence of pointwise-convergent
sequences of measurable functions.
4.6. Complex-valued functions and a.e. convergence
In this section, we briefly indicate the generalization of the above results to
complex-valued functions and sequences that converge pointwise almost everywhere. The required modifications are straightforward.
If f : X C is a complex valued function f = g + ih, then we say that f is
measurable if and only if its real and imaginary parts g, h : X R are measurable,
and integrable if and only if g, h are integrable. In that case, we define
Z
Z
Z
f d = g d + i h d.

Note that we do not allow extended real-valued functions or infinite integrals here.
It follows from the discussion of product measures that f : X C, where C is
equipped with its Borel -algebra B(C), is measurable if and only if its real and
imaginary parts are measurable, so this definition is consistent with our previous
one.
The integral of complex-valued functions satisfies the properties given in Proposition 4.9, where we allow k C and the condition
f Rg is only relevant for
R
real-valued functions. For example, to show that | f d| |f | d, we let
Z

Z



f d =  f d ei
for a suitable argument , and then
Z

Z
Z
Z
Z


 f d = ei f d = [ei f ] d |[ei f ]| d |f | d.



Complex-valued functions also satisfy the properties given in Section 4.4.

The monotone convergence theorem holds for extended real-valued functions
if fn f pointwise a.e., and the Lebesgue dominated convergence theorem holds
for complex-valued functions if fn f pointwise a.e. and |fn | g pointwise a.e.
where g is an integrable extended real-valued function. If the measure space is not
complete, then we also need to assume that f is measurable. To prove these results,
we replace the functions fn , for example, by fn N c where N is a null set off which
pointwise convergence holds, and apply the previous theorems; the values of any
integrals are unaffected.
4.7. L1 spaces
We introduce here the space L1 (X) of integrable functions on a measure space
X; we will study its properties, and the properties of the closely related Lp spaces,
in more detail later on.
Definition 4.25. If (X, A, ) is a measure space, then the space L1 (X) consists
of the integrable functions f : X R with norm
Z
kf kL1 = |f | d < ,

4.7. L1 SPACES

51

where we identify functions that are equal a.e. A sequence of functions

{fn L1 (X)}

converges in L1 , or in mean, to f L1 (X) if

Z
kf fn kL1 = |f fn | d 0

as n .

We also denote the space of integrable complex-valued functions f : X C by

L1 (X). For definiteness, we consider real-valued functions unless stated otherwise;
in most cases, the results generalize in an obvious way to complex-valued functions
Convergence in mean is not equivalent to pointwise a.e.-convergence. The sequences in Examples 4.184.20 converges to zero pointwise, but they do not converge in mean. The following is an example of a sequence that converges in mean
but not pointwise a.e.
Example 4.26. Define fn : [0, 1] R by


1 if 0 x 1/2,
0 if 0 x < 1/2,
f1 (x) = 1, f2 (x) =
f3 (x) =
0 if 1/2 < x 1,
1 if 1/2 x 1,


0 if 0 x < 1/4,
1 if 0 x 1/4,
1 if 1/4 x 1/2,
f4 (x) =
f5 (x) =
1 if 1/4 < x 1,

0 if 1/2x < x 1,

and so on. That is, for 2m n 2m 1, the function fn consists of a spike of

height one and width 2m that sweeps across the interval [0, 1] as n increases from
2m to 2m 1. This sequence converges to zero in mean, but it does not converge
pointwise as any point x [0, 1].
We will show, however, that a sequence which converges sufficiently rapidly in
mean does converge pointwise a.e.; as a result, every sequence that converges in
mean has a subsequence that converges pointwise a.e. (see Lemma 7.9 and Corollary 7.11).
Let us consider the particular case of L1 (Rn ). As an application of the Borel
regularity of Lebesgue measure, we prove that integrable functions on Rn may be
approximated by continuous functions with compact support. This result means
that L1 (Rn ) is a concrete realization of the completion of Cc (Rn ) with respect to the
L1 (Rn )-norm, where Cc (Rn ) denotes the space of continuous functions f : Rn R
with compact support. The support of f is defined by
suppf = {x Rn : f (x) 6= 0}.

Thus, f has compact support if and only if it vanishes outside a bounded set.
Theorem 4.27. The space Cc (Rn ) is dense in L1 (Rn ). Explicitly, if f
L (Rn ), then for any > 0 there exists a function g Cc (Rn ) such that
1

kf gkL1 < .

Proof. Note first that by the dominated convergence theorem

f f B (0) 1 0
as R ,
R
L

so we can assume that f L1 (Rn ) has compact support. Decomposing f = f + f

into positive and negative parts, we can also assume that f is positive. Then there
is an increasing sequence of compactly supported simple functions that converges

52

4. INTEGRATION

to f pointwise and hence, by the monotone (or dominated) convergence theorem,

in mean. Since every simple function is a finite linear combination of characteristic
functions, it is sufficient to prove the result for the characteristic function A of a
bounded, measurable set A Rn .
Given > 0, by the Borel regularity of Lebesgue measure, there exists a
bounded open set G and a compact set K such that K A G and (G \ K) < .
Let g Cc (Rn ) be a Urysohn function such that g = 1 on K, g = 0 on Gc , and
0 g 1. For example, we can define g explicitly by
g(x) =

d(x, Gc )
d(x, K) + d(x, Gc )

where the distance function d(, F ) : Rn R from a subset F Rn is defined by

d(x, F ) = inf {|x y| : y F } .
If F is closed, then d(, F ) is continuous, so g is continuous.
We then have that
Z
kA gkL1 =
|A g| dx (G \ K) < ,
G\K

which proves the result.


4.8. Riemann integral

Any Riemann integrable function f : [a, b] R is Lebesgue measurable, and in

fact integrable since it must be bounded, but a Lebesgue integrable function need
not be Riemann integrable. Using Lebesgue measure, we can give a necessary and
sufficient condition for Riemann integrability.
Theorem 4.28. If f : [a, b] R is Riemann integrable, then f is Lebesgue
integrable on [a, b] and its Riemann integral is equal to its Lebesgue integral. A
Lebesgue measurable function f : [a, b] R is Riemann integrable if and only if
it is bounded and the set of discontinuities {x [a, b] : f is discontinuous at x} has
Lebesgue measure zero.
For the proof, see e.g. Folland [4].
Example 4.29. The characteristic function of the rationals Q[0,1] is discontinuous at every point and it is not Riemann integrable on [0, 1]. This function is,
however, equal a.e. to the zero function which is continuous at every point and is
Riemann integrable. (Note that being continuous a.e. is not the same thing as being
equal a.e. to a continuous function.) Any function that is bounded and continuous
except at countably many points is Riemann integrable, but these are not the only
Riemann integrable functions. For example, the characteristic function of a Cantor
set with zero measure is a Riemann integrable function that is discontinuous at
uncountable many points.
4.9. Integrals of vector-valued functions
In Definition 4.4, we use the ordering properties of R to define real-valued
integrals as a supremum of integrals of simple functions. Finite integrals of complexvalued functions or vector-valued functions that take values in a finite-dimensional
vector space are then defined componentwise.

4.9. INTEGRALS OF VECTOR-VALUED FUNCTIONS

53

An alternative method is to define the integral of a vector-valued function

f : X Y from a measure space X to a Banach space Y as a limit in norm of
integrals of vector-valued simple functions. The integral of vector-valued simple
functions is defined as in (4.1), assuming that (En ) < ; linear combinations of
the values cn Y make sense since Y is a vector space. A function f : X Y is
integrable if there is a sequence of integrable simple functions {n : X Y } such
that n f pointwise, where the convergence is with respect to the norm k k on
Y , and
Z
kf n k d 0
as n .
Then we define

f d = lim

n d,

where the limit is the norm-limit in Y .

This definition of the integral agrees with the one used above for real-valued,
integrable functions, and amounts to defining the integral of an integrable function
by completion in the L1 -norm. We will not develop this definition here (see [6],
for example, for a detailed account), but it is useful in defining the integral of
functions that take values in an infinite-dimensional Banach space, when it leads to
the Bochner integral. An alternative approach is to reduce vector-valued integrals
to scalar-valued integrals by the use of continuous linear functionals belonging to
the dual space of the Banach space.

CHAPTER 5

Product Measures
Given two measure spaces, we may construct a natural measure on their Cartesian product; the prototype is the construction of Lebesgue measure on R2 as the
product of Lebesgue measures on R. The integral of a measurable function on
the product space may be evaluated as iterated integrals on the individual spaces
provided that the function is positive or integrable (and the measure spaces are
-finite). This result, called Fubinis theorem, is another one of the basic and most
useful properties of the Lebesgue integral. We will not give complete proofs of all
the results in this Chapter.
5.1. Product -algebras
We begin by describing product -algebras. If (X, A) and (Y, B) are measurable
spaces, then a measurable rectangle is a subset A B of X Y where A A and
B B are measurable subsets of X and Y , respectively. For example, if R is
equipped with its Borel -algebra, then Q Q is a measurable rectangle in R R.
(Note that the sides A, B of a measurable rectangle A B R R can be
arbitrary measurable sets; they are not required to be intervals.)
Definition 5.1. Suppose that (X, A) and (Y, B) are measurable spaces. The
product -algebra A B is the -algebra on X Y generated by the collection of
all measurable rectangles,
A B = ({A B : A A, B B}) .

The product of (X, A) and (Y, B) is the measurable space (X Y, A B).

Suppose that E X Y . For any x X and y Y , we define the x-section
Ex Y and the y-section E y X of E by
E y = {x X : (x, y) E} .

Ex = {y Y : (x, y) E} ,

As stated in the next proposition, all sections of a measurable set are measurable.
Proposition 5.2. If (X, A) and (Y, B) are measurable spaces and E A B,
then Ex B for every x X and E y A for every y Y .
Proof. Let
M = {E X Y : Ex B for every x X and E y A for every y Y } .

Then M contains all measurable rectangles, since the x-sections of A B are either
or B and the y-sections are either or A. Moreover, M is a -algebra since, for
example, if E, Ei X Y and x X, then
!

[
[
(Ei )x .
Ei
=
(E c )x = (Ex )c ,
i=1

55

i=1

56

5. PRODUCT MEASURES

It follows that M A B, which proves the proposition.

As an example, we consider the product of Borel -algebras on Rn .

Proposition 5.3. Suppose that Rm , Rn are equipped with their Borel -algebras
B(Rm ), B(Rn ) and let Rm+n = Rm Rn . Then
B(Rm+n ) = B(Rm ) B(Rn ).

Proof. Every (m + n)-dimensional rectangle, in the sense of Definition 2.1, is

a product of an m-dimensional and an n-dimensional rectangle. Therefore
B(Rm ) B(Rn ) R(Rm+n )

where R(Rm+n ) denotes the collection of rectangles in Rm+n . From Proposition 2.21, the rectangles generate the Borel -algebra, and therefore
B(Rm ) B(Rn ) B(Rm+n ).

To prove the the reverse inclusion, let

M = {A Rm : A Rn B(Rm+n )} .

Then M is a -algebra, since B(Rm+n ) is a -algebra and

!

[
[
c
n
n c
(Ai Rn ) .
Ai Rn =
A R = (A R ) ,
i=1

i=1

Moreover, M contains all open sets, since G R is open in Rm+n if G is open

in Rm . It follows that M B(Rm ), so A Rn B(Rm+n ) for every A B(Rm ),
meaning that
B(Rm+n ) {A Rn : A B(Rm )} .
Similarly, we have
B(Rm+n ) {Rn B : B B(Rn )} .
Therefore, since B(Rm+n ) is closed under intersections,
B(Rm+n ) {A B : A B(Rm ), B B(Rn )} ,

which implies that

B(Rm+n ) B(Rm ) B(Rn ).

By the repeated application of this result, we see that the Borel -algebra on
Rn is the n-fold product of the Borel -algebra on R. This leads to an alternative
method of constructing Lebesgue measure on Rn as a product of Lebesgue measures
on R, instead of the direct construction we gave earlier.
5.2. Premeasures
Premeasures provide a useful way to generate outer measures and measures,
and we will use them to construct product measures. In this section, we derive
some general results about premeasures and their associated measures that we use
below. Premeasures are defined on algebras, rather than -algebras, but they are
consistent with countable additivity.
Definition 5.4. An algebra on a set X is a collection of subsets of X that
contains and X and is closed under complements, finite unions, and finite intersections.

5.2. PREMEASURES

57

If F P(X) is a family of subsets of a set X, then the algebra generated by F is

the smallest algebra that contains F . It is much easier to give an explicit description
of the algebra generated by a family of sets F than the -algebra generated by F .
For example, if F has the property that for A, B F , the intersection A B F
and the complement Ac is a finite union of sets belonging to F , then the algebra
generated by F is the collection of all finite unions of sets in F .
Definition 5.5. Suppose that E is an algebra of subsets of a set X. A premeasure on E, or on X if the algebra is understood, is a function : E [0, ]
such that:
(a) () = 0;
(b) if {Ai E : i N} is a countable collection of disjoint sets in E such that

[
Ai E,
i=1

then

Ai

i=1

(Ai ) .

i=1

Note that a premeasure is finitely additive, since we may take Ai = for

i N , and monotone, since if A B, then (A) = (A \ B) + (B) (B).
To define the outer measure associated with a premeasure, we use countable
coverings by sets in the algebra.
Definition 5.6. Suppose that E is an algebra on a set X and : E [0, ] is
a premeasure. The outer measure : P(X) [0, ] associated with is defined
for E X by
)
(
X
S

(Ai ) : E i=1 Ai where Ai E .

(E) = inf
i=1

As we observe next, the set-function is an outer measure. Moreover, every

set belonging to E is -measurable and its outer measure is equal to its premeasure.

Proposition 5.7. The set function : P(X) [0, ] given by Definition 5.6.
is an outer measure on X. Every set A E is Caratheodory measurable and
(A) = (A).

Proof. The proof that is an outer measure is identical to the proof of

Theorem 2.4 for outer Lebesgue measure.
If A E, then (A) (A) since A covers itself. To prove the reverse
inequality, suppose that {Ai : i N} is a countable cover of A by sets Ai E. Let
B1 = A A1 and
!
j1
[
for j 2.
Ai
Bj = A Aj \
i=1

Then Bj A and A is the disjoint union of {Bj : j N}. Since Bj Aj , it follows

that

X
X
(Aj ),
(Bj )
(A) =
j=1

j=1

which implies that (A) (A). Hence, (A) = (A).

58

5. PRODUCT MEASURES

If E X, A E, and > 0, then there is a cover {Bi E : i N} of E such

that

X
(Bi ).
(E) +
i=1

Since is countably additive on E,

(E) +

X
i=1

(Bi A) +

X
i=1

(Bi Ac ) (E A) + (E Ac ),

and since > 0 is arbitrary, it follows that (E) (E A) + (E Ac ), which

implies that A is measurable.

Using Theorem 2.9, we see from the preceding results that every premeasure on
an algebra E may be extended to a measure on (E). A natural question is whether
such an extension is unique. In general, the answer is no, but if the measure space
is not too big, in the following sense, then we do have uniqueness.
Definition 5.8. Let S
X be a set and a premeasure on an algebra E P(X).

Then is -finite if X = i=1 Ai where Ai E and (Ai ) < .

Theorem 5.9. If : E [0, ] is a -finite premeasure on an algebra E and

A is the -algebra generated by E, then there is a unique measure : A [0, ]
such that (A) = (A) for every A E.
5.3. Product measures

Next, we construct a product measure on the product of measure spaces that

satisfies the natural condition that the measure of a measurable rectangle is the
product of the measures of its sides. To do this, we will use the Caratheodory
method and first define an outer measure on the product of the measure spaces in
terms of the natural premeasure defined on measurable rectangles. The procedure
is essentially the same as the one we used to construct Lebesgue measure on Rn .
Suppose that (X, A) and (Y, B) are measurable spaces. The intersection of
measurable rectangles is a measurable rectangle
(A B) (C D) = (A C) (B D),
and the complement of a measurable rectangle is a finite union of measurable rectangles
(A B)c = (Ac B) (A B c ) (Ac B c ).

Thus, the collection of finite unions of measurable rectangles in X Y forms an

algebra, which we denote by E. This algebra is not, in general, a -algebra, but
obviously it generates the same product -algebra as the measurable rectangles.
Every set E E may be represented as a finite disjoint union of measurable
rectangles, though not necessarily in a unique way. To define a premeasure on E,
we first define the premeasure of measurable rectangles.
Definition 5.10. If (X, A, ) and (Y, B, ) are measure spaces, then the product premeasure (A B) of a measurable rectangle A B X Y is given by
(A B) = (A)(B)
where 0 = 0.

5.3. PRODUCT MEASURES

59

The premeasure is countably additive on rectangles. The simplest way to

show this is to integrate the characteristic functions of the rectangles, which allows
us to use the monotone convergence theorem.
Proposition 5.11. If a measurable rectangle A B is a countable disjoint
union of measurable rectangles {Ai Bi : i N}, then
(A B) =
Proof. If
AB =

X
i=1

(Ai Bi ).

(Ai Bi )

i=1

is a disjoint union, then the characteristic function AB : X Y [0, ) satisfies

AB (x, y) =

Ai Bi (x, y).

i=1

Therefore, since AB (x, y) = A (x)B (y),

A (x)B (y) =

Ai (x)Bi (y).

Ai (x)(Bi ).

(Ai )(Bi ),

i=1

Integrating this equation over Y for fixed x X and using the monotone convergence theorem, we get
A (x)(B) =

i=1

Integrating again with respect to x, we get

(A)(B) =

i=1

which proves the result.

In particular, it follows that is finitely additive on rectangles and therefore

may be extended to a well-defined function on E. To see this, note that any two
representations of the same set as a finite disjoint union of rectangles may be decomposed into a common refinement such that each of the original rectangles is a
disjoint union of rectangles in the refinement. The following definition therefore
makes sense.
Definition 5.12. Suppose that (X, A, ) and (Y, B, ) are measure spaces and
E is the algebra generated by the measurable rectangles. The product premeasure
: E [0, ] is given by
(E) =

N
X
i=1

(Ai )(Bi ),

E=

N
[

i=1

Ai Bi

SN
where E = i=1 Ai Bi is any representation of E E as a disjoint union of
measurable rectangles.

60

5. PRODUCT MEASURES

Proposition 5.11 implies that is countably additive on E, since we may decompose any countable disjoint union of sets in E into a countable common disjoint
refinement of rectangles, so is a premeasure as claimed. The outer product measure associated with , which we write as ( ) , is defined in terms of countable
coverings by measurable rectangles. This gives the following.
Definition 5.13. Suppose that (X, A, ) and (Y, B, ) are measure spaces.
Then the product outer measure
( ) : P(X Y ) [0, ]

on X Y is defined for E X Y by
)
(
X
S

(Ai )(Bi ) : E i=1 (Ai Bi ) where Ai A, Bi B .

( ) (E) = inf
i=1

The product measure

( ) : A B [0, ],

( ) = ( ) |AB

is the restriction of the product outer measure to the product -algebra.

It follows from Proposition 5.7 that ( ) is an outer measure and every
measurable rectangle is ( ) -measurable with measure equal to its product
premeasure. We summarize the conclusions of the Caratheodory theorem and Theorem 5.9 in the case of product measures as the following result.
Theorem 5.14. If (X, A, ) and (Y, B, ) are measure spaces, then
( ) : A B [0, ]

is a measure on X Y such that

( )(A B) = (A)(B)

for every A A, B B.

Moreover, if (X, A, ) and (Y, B, ) are -finite measure spaces, then ( ) is the
unique measure on A B with this property.
Note that, in general, the -algebra of Caratheodory measurable sets associated
with ( ) is strictly larger than the product -algebra. For example, if Rm and
Rn are equipped with Lebesgue measure defined on their Borel -algebras, then the
Caratheodory -algebra on the product Rm+n = Rm Rn is the Lebesgue -algebra
L(Rm+n ), whereas the product -algebra is the Borel -algebra B(Rm+n ).
5.4. Measurable functions
If f : X Y C is a function of (x, y) X Y , then for each x X we define
the x-section fx : Y C and for each y Y we define the y-section f y : Y C
by
fx (y) = f (x, y),
f y (x) = f (x, y).
Theorem 5.15. If (X, A, ), (Y, B, ) are measure spaces and f : X Y C
is a measurable function, then fx : Y C, f y : X C are measurable for every
x X, y Y . Moreover, if (X, A, ), (Y, B, ) are -finite, then the functions
g : X C, h : Y C defined by
Z
Z
g(x) = fx d,
h(y) = f y d
are measurable.

5.7. COMPLETION OF PRODUCT MEASURES

61

5.5. Monotone class theorem

We prove a general result about -algebras, called the monotone class theorem,
which we will use in proving Fubinis theorem. A collection of subsets of a set
is called a monotone class if it is closed under countable increasing unions and
countable decreasing intersections.
Definition 5.16. A monotone class on a set X is a collection C P(X) of
subsets of X such that if Ei , Fi C and
E1 E2 Ei . . . ,

F1 F2 Fi . . . ,

then

i=1

Ei C,

i=1

Fi C.

Obviously, every -algebra is a monotone class, but not conversely. As with

-algebras, every family F P(X) of subsets of a set X is contained in a smallest
monotone class, called the monotone class generated by F , which is the intersection
of all monotone classes on X that contain F . As stated in the next theorem, if F
is an algebra, then this monotone class is, in fact, a -algebra.
Theorem 5.17 (Monotone Class Theorem). If F is an algebra of sets, the
monotone class generated by F coincides with the -algebra generated by F .
5.6. Fubinis theorem
Theorem 5.18 (Fubinis Theorem). Suppose that (X, A, ) and (Y, B, ) are
-finite measure spaces. A measurable function f : X Y C is integrable if and
only if either one of the iterated integrals


Z Z
Z Z
|f y | d d,
|fx | d d
is finite. In that case


Z
Z Z
Z Z
f d d =
f y d d =
fx d d.
Example 5.19. An application of Fubinis theorem to counting measure on
NN implies that if {amn C | m, n N} is a doubly-indexed sequence of complex
numbers such that
!

X
X
|amn | <
m=1

n=1

then

m=1

n=1

amn

n=1

m=1

amn

5.7. Completion of product measures

The product of complete measure spaces is not necessarily complete.

62

5. PRODUCT MEASURES

Example 5.20. If N R is a non-Lebesgue measurable subset of R, then

{0} N does not belong to the product -algebra L(R) L(R) on R2 = R R,
since every section of a set in the product -algebra is measurable. It does, however,
belong to L(R2 ), since it is a subset of the set {0} R of two-dimensional Lebesgue
measure zero, and Lebesgue measure is complete. Instead one can show that the
Lebesgue -algebra on Rm+n is the completion with respect to Lebesgue measure
of the product of the Lebesgue -algebras on Rm and Rn :
L(Rm+n ) = L(Rm ) L(Rn ).
We state a version of Fubinis theorem for Lebesgue measurable functions on
Rn .
Theorem 5.21. A Lebesgue measurable function f : Rm+n C is integrable,
meaning that
Z
|f (x, y)| dxdy < ,
Rm+n

if and only if either one of the iterated integrals


Z Z
Z Z
|f (x, y)| dx dy,
Rn

Rm

Rm

is finite. In that case,

Z
Z
f (x, y) dxdy =
Rm+n

Rn

Z

Rm

Rn

|f (x, y)| dy


Z
f (x, y) dx dy =

Rm

where all of the integrals are well-defined and finite a.e.

Z

dx

f (x, y) dy
Rn

dx,

CHAPTER 6

Differentiation
The generalization from elementary calculus of differentiation in measure theory
is less obvious than that of integration, and the methods of treating it are somewhat
involved.
Consider the fundamental theorem of calculus (FTC) for smooth functions of
a single variable. In one direction (FTC-I, say) it states that the derivative of the
integral is the original function, meaning that
Z x
d
f (y) dy.
(6.1)
f (x) =
dx a
In the other direction (FTC-II, say) it states that we recover the original function
by integrating its derivative
Z x
(6.2)
F (x) = F (a) +
f (y) dy,
f = F .
a

As we will see, (6.1) holds pointwise a.e. provided that f is locally integrable, which
is needed to ensure that the right-hand side is well-defined. Equation (6.2), however,
does not hold for all continuous functions F whose pointwise derivative is defined
a.e. and integrable; we also need to require that F is absolutely continuous. The
Cantor function is a counter-example.
First, we consider a generalization of (6.1) to locally integrable functions on
Rn , which leads to the Lebesgue differentiation theorem. We say that a function
f : Rn R is locally integrable if it is Lebesgue measurable and
Z
|f | dx <
K

for every compact subset K R ; we denote the space of locally integrable functions by L1loc (Rn ).
Let
(6.3)

Br (x) = {y Rn : |y x| < r}

denote the open ball of radius r and center x Rn . We denote Lebesgue measure
on Rn by and the Lebesgue measure of a ball B by (B) = |B|.
To motivate the statement of the Lebesgue differentiation theorem, observe
that (6.1) may be written in terms of symmetric differences as
Z x+r
1
f (y) dy.
(6.4)
f (x) = lim+
r0 2r xr
In other words, the value of f at a point x is the limit of local averages of f over
intervals centered at x as their lengths approach zero. An n-dimensional version of
63

64

6. DIFFERENTIATION

(6.4) is
(6.5)

f (x) = lim+
r0

1
|Br (x)|

f (y) dy

Br (x)

where the integral is with respect n-dimensional Lebesgue measure. The Lebesgue
differentiation theorem states that (6.5) holds pointwise -a.e. for any locally integrable function f .
To prove the theorem, we will introduce the maximal function of an integrable
function, whose key property is that it is weak-L1, as stated in the Hardy-Littlewood
theorem. This property may be shown by the use of a simple covering lemma, which
we begin by proving.
Second, we consider a generalization of (6.2) on the representation of a function
as an integral. In defining integrals on a general measure space, it is natural to
think of them as defined on sets rather than real numbers. For example, in (6.2),
we would write F (x) = ([a, x]) where : B([a, b]) R is a signed measure. This
interpretation leads to the following question: if , are measures on a measurable
space X is there a function f : X [0, ] such that
Z
f d.
(A) =
A

If so, we regard f = d/d as the (Radon-Nikodym) derivative of with respect to

. More generally, we may consider signed (or complex) measures, whose values are
not restricted to positive numbers. The Radon-Nikodym theorem gives a necessary
and sufficient condition for the differentiability of with respect to , subject to a
-finiteness assumption: namely, that is absolutely continuous with respect to .
6.1. A covering lemma
We need only the following simple form of a covering lemma; there are many
more sophisticated versions, such as the Vitali and Besicovitch covering theorems,
which we do not consider here.
Lemma 6.1. Let {B1 , B2 , . . . , BN } be a finite collection of open balls in Rn .

} where Bj = Bij , such that

There is a disjoint subcollection {B1 , B2 , . . . , BM
!
M
N
X
[
 
 Bj  .
Bi 3 n

i=1

i=1

b denote the open ball with the same center

Proof. If B is an open ball, let B
as B and three times the radius. Then
b = 3n |B|.
|B|

Moreover, if B1 , B2 are nondisjoint open balls and the radius of B1 is greater than
b1 B2 .
or equal to the radius of B2 , then B

We obtain the subfamily {Bj } by an iterative procedure. Choose B1 to be a

ball with the largest radius from the collection {B1 , B2 , . . . , BN }. Delete from the
collection all balls Bi that intersect B1 , and choose B2 to be a ball with the largest
radius from the remaining balls. Repeat this process until the balls are exhausted,
which gives M N balls, say.

6.2. MAXIMAL FUNCTIONS

65

By construction, the balls {B1 , B2 , . . . , BM

} are disjoint and
N
[

i=1

It follows that

N
[

i=1

which proves the result.

Bi

Bi

M
[

b .
B
j

j=1

M
M 

X
X
 
 b 
 Bj  ,
Bj  = 3n
i=1

i=1

6.2. Maximal functions

The maximal function of a locally integrable function is obtained by taking the
supremum of averages of the absolute value of the function about a point. Maximal
functions were introduced by Hardy and Littlewood (1930), and they are the key
to proving pointwise properties of integrable functions. They also play a central
role in harmonic analysis.
by

Definition 6.2. If f L1loc (Rn ), then the maximal function M f of f is defined

M f (x) = sup
r>0

1
|Br (x)|

Br (x)

|f (y)| dy.

The use of centered open balls to define the maximal function is for convenience.
We could use non-centered balls or other sets, such as cubes, to define the maximal
function. Some restriction on the shapes on the sets is, however, required; for
example, we cannot use arbitrary rectangles, since averages over progressively longer
and thinner rectangles about a point whose volumes shrink to zero do not, in
general, converge to the value of the function at the point, even if the function is
continuous.
Note that any two functions that are equal a.e. have the same maximal function.
Example 6.3. If f : R R is the step function

1 if x 0,
f (x) =
0 if x < 0,
then
M f (x) =

1
if x > 0,
1/2 if x 0.

This example illustrates the following result.

Proposition 6.4. If f L1loc (Rn ), then the maximal function M f is lower
semi-continuous and therefore Borel measurable.
Proof. The function M f 0 is lower semi-continuous if
Et = {x : M f (x) > t}
is open for every 0 < t < . To prove that Et is open, let x Et . Then there
exists r > 0 such that
Z
1
|f (y)| dy > t.
|Br (x)| Br (x)

66

6. DIFFERENTIATION

Choose r > r such that we still have

Z
1
|f (y)| dy > t.
|Br (x)| Br (x)
If |x x| < r r, then Br (x) Br (x ), so
Z
Z
1
1
t<
|f (y)| dy
|f (y)| dy M f (x ),
|Br (x)| Br (x)
|Br (x )| Br (x )
It follows that x Et , which proves that Et is open.

The maximal function of a non-zero function f L1 (Rn ) is not in L1 (Rn )

because it decays too slowly at infinity for its integral to converge. To show this,
let a > 0 and suppose that |x| a. Then, by considering the average of |f | at x
over a ball of radius r = 2|x| and using the fact that B2|x| (x) Ba (0), we see that
Z
1
|f (y)| dy
M f (x)
|B2|x| (x)| B2|x| (x)
Z
C
|f (y)| dy,

|x|n Ba (0)
where C > 0. The function 1/|x|n is not integrable on Rn \ Ba (0), so if M f is
integrable then we must have
Z
|f (y)| dy = 0
Ba (0)

for every a > 0, which implies that f = 0 a.e. in Rn .

Moreover, as the following example shows, the maximal function of an integrable function need not even be locally integrable.
Example 6.5. Define f : R R by

1/(x log2 x)
f (x) =
0

if 0 < x < 1/2,

otherwise.

The change of variable u = log x implies that

Z 1/2
1
dx
x|
log
x|n
0
is finite if and only if n > 1. Thus f L1 (R) and for 0 < x < 1/2
Z 2x
1
M f (x)
|f (y)| dy
2x 0
Z x
1
1
dy

2x 0 y log2 y
1

2x| log x|
so M f
/ L1loc (R).

6.4. HARDY-LITTLEWOOD THEOREM

67

6.3. Weak-L1 spaces

Although the maximal function of an integrable function is not integrable, it
is not much worse than an integrable function. As we show in the next section, it
belongs to the space weak-L1, which is defined as follows
Definition 6.6. The space weak-L1(Rn ) consists of measurable functions
f : Rn R
such that there exists a constant C, depending on f but not on t, with the property
that for every 0 < t <
{x Rn : |f (x)| > t}

C
.
t

An estimate of this form arises for integrable function from the following, almost
trivial, Chebyshev inequality.
Theorem 6.7 (Chebyshevs inequality). Suppose that (X, A, ) is a measure
space. If f : X R is integrable and 0 < t < , then
(6.6)

({x X : |f (x)| > t})

1
kf kL1 .
t

Proof. Let Et = {x X : |f (x)| > t}. Then

Z
Z
|f | d t(Et ),
|f | d
Et

which proves the result.

Chebyshevs inequality implies immediately that if f belongs to L1 (Rn ), then

f belongs to weak-L1 (Rn ). The converse statement is, however, false.
Example 6.8. The function f : R R defined by
f (x) =

1
x

for x 6= 0 satisfies

2
,
t
so f belongs to weak-L1 (R), but f is not integrable or even locally integrable.
{x R : |f (x)| > t} =

6.4. Hardy-Littlewood theorem

The following Hardy-Littlewood theorem states that the maximal function of
an integrable function is weak-L1.
Theorem 6.9 (Hardy-Littlewood). If f L1 (Rn ), there is a constant C such
that for every 0 < t <
({x Rn : M f (x) > t})
where C = 3n depends only on n.

C
kf kL1
t

68

6. DIFFERENTIATION

Proof. Fix t > 0 and let

Et = {x Rn : M f (x) > t} .

By the inner regularity of Lebesgue measure

(Et ) = sup {(K) : K Et is compact}

so it is enough to prove that

(K)

C
t

Rn

|f (y)| dy.

for every compact subset K of Et .

If x K, then there is an open ball Bx centered at x such that
Z
1
|f (y)| dy > t.
|Bx | Bx

Since K is compact, we may extract a finite subcover {B1 , B2 , . . . , BN } from the

open cover {Bx : x K}. By Lemma 6.1, there is a finite subfamily of disjoint

balls {B1 , B2 , . . . , BM
} such that
(K)

N
X
i=1

3n

|Bi |

M
X
j=1

|Bj |

M Z
n X

3
t

3n

t
n
which proves the result with C = 3 .

j=1

Bj

|f | dx

|f | dx,


6.5. Lebesgue differentiation theorem

The maximal function provides the crucial estimate in the following proof.
Theorem 6.10. If f L1loc (Rn ), then for a.e. x Rn
"
#
Z
1
lim
f (y) dy = f (x).
r0+ |Br (x)| Br (x)

Moreover, for a.e. x Rn

"
lim

r0+

1
|Br (x)|

Br (x)

|f (y) f (x)| dy = 0.

Proof. Since


Z
Z
 1

1


f (y) f (x) dy 
|f (y) f (x)| dy,

 |Br (x)| Br (x)
 |Br (x)| Br (x)

we just need to prove the second result. We define f : Rn [0, ] by

"
#
Z
1

|f (y) f (x)| dy .
f (x) = lim sup
|Br (x)| Br (x)
r0+

6.5. LEBESGUE DIFFERENTIATION THEOREM

69

We want to show that f = 0 pointwise a.e.

If g Cc (Rn ) is continuous, then given any > 0 there exists > 0 such that
|g(x) g(y)| < whenever |x y| < . Hence if r <
Z
1
|f (y) f (x)| dy < ,
|Br (x)| Br (x)
which implies that g = 0. We prove the result for general f by approximation
with a continuous function.
First, note that we can assume that f L1 (Rn ) is integrable without loss of
generality; for example, if the result holds for f Bk (0) L1 (Rn ) for each k N
except
on a set Ek of measure zero, then it holds for f L1loc (Rn ) except on
S
k=1 Ek , which has measure zero.
Next, observe that since
|f (y) + g(y) [f (x) + g(x)]| |f (y) f (x)| + |g(y) g(x)|
and lim sup(A + B) lim sup A + lim sup B, we have
(f + g) f + g .

Thus, if f L1 (Rn ) and g Cc (Rn ), we have

(f g) f + g = f ,

f = (f g + g) (f g) + g = (f g) ,
which shows that (f g) = f .
If f L1 (Rn ), then we claim that there is a constant C, depending only on n,
such that for every 0 < t <
({x Rn : f (x) > t})

(6.7)

C
kf kL1 .
t

To show this, we estimate

"

#
Z
1
f (x) sup
|f (y) f (x)| dy
r>0 |Br (x)| Br (x)
#
"
Z
1
|f (y)| dy + |f (x)|
sup
r>0 |Br (x)| Br (x)

M f (x) + |f (x)|.
It follows that
{f > t} {M f + |f | > t} {M f > t/2} {|f | > t/2} .
By the Hardy-Littlewood theorem,
({x Rn : M f (x) > t/2})

2 3n
kf kL1 ,
t

and by the Chebyshev inequality

({x Rn : |f (x)| > t/2})

2
kf kL1 .
t

Combining these estimates, we conclude that (6.7) holds with C = 2 (3n + 1).

70

6. DIFFERENTIATION

Finally suppose that f L1 (Rn ) and 0 < t < . From Theorem 4.27, for any
> 0, there exists g Cc (Rn ) such that kf g||L1 < . Then
({x Rn : f (x) > t}) = ({x Rn : (f g) (x) > t})
C
kf gkL1
t
C
.

t
Since > 0 is arbitrary, it follows that
({x Rn : f (x) > t}) = 0,

and hence since

{x Rn : f (x) > 0} =
that
This proves the result.

k=1

{x Rn : f (x) > 1/k}

({x Rn : f (x) > 0}) = 0.

The set of points x for which the limits in Theorem 6.10 exist for a suitable
definition of f (x) is called the Lebesgue set of f .
Definition 6.11. If f L1loc (Rn ), then a point x Rn belongs to the Lebesgue
set of f if there exists a constant c R such that
"
#
Z
1
lim
|f (y) c| dy = 0.
r0+ |Br (x)| Br (x)
If such a constant c exists, then it is unique. Moreover, its value depends only
on the equivalence class of f with respect to pointwise a.e. equality. Thus, we can
use this definition to give a canonical pointwise a.e. representative of a function
f L1loc (Rn ) that is defined on its Lebesgue set.
Example 6.12. The Lebesgue set of the step function f in Example 6.3 is
R \ {0}. The point 0 does not belong to the Lebesgue set, since

 Z r
1
1
|f (y) c| dy = (|c| + |1 c|)
lim
+
2r r
2
r0
is nonzero for every c R. Note that the existence of the limit
 Z r

1
1
lim
f (y) dy =
2
r0+ 2r r

is not sufficient to imply that 0 belongs to the Lebesgue set of f .

6.6. Signed measures

A signed measure is a countably additive, extended real-valued set function

whose values are not required to be positive. Measures may be thought of as a
generalization of volume or mass, and signed measures may be thought of as a
generalization of charge, or a similar quantity. We allow a signed measure to take
infinite values, but to avoid undefined expressions of the form , it should not
take both positive and negative infinite values.

6.7. HAHN AND JORDAN DECOMPOSITIONS

71

Definition 6.13. Let (X, A) be a measurable space. A signed measure on

X is a function : A R such that:
(a) () = 0;
(b) attains at most one of the values , ;
(c) if {Ai A : i N} is a disjoint collection of measurable sets, then
!

X
[
(Ai ).
Ai =

i=1

i=1

We say
that
Sthat a signed measure is finite if it takes only finite values.PNote

(A
since ( i=1 Ai ) does not depend on the order of the Ai , the sum
i)
i=1
converges unconditionally if it is finite, and therefore it is absolutely convergent.
Signed measures have the same monotonicity property (1.1) as measures, with
essentially the same proof. We will always refer to signed measures explicitly, and
measure will always refer to a positive measure.
Example 6.14. If (X, A, ) is a measure space and + , : A [0, ] are
measures, one of which is finite, then = + is a signed measure.
Example 6.15. If (X, A, ) is a measure space and f : X R is an Ameasurable function whose integral with respect to is defined as an extended real
number, then : A R defined by
Z
(6.8)
(A) =
f d
A

is a signed measure on X. As we describe below, we interpret f as the derivative

d/d of with respect to . If f = f + f is the decomposition of f into positive
and negative parts then = + , where the measures + , : A [0, ] are
defined by
Z
Z
f d.
f + d,
(A) =
+ (A) =
A

We will show that any signed measure can be decomposed into a difference of
singular measures, called its Jordan decomposition. Thus, Example 6.14 includes all
signed measures. Not all signed measures have the form given in Example 6.15. As
we discuss this further in connection with the Radon-Nikodym theorem, a signed
measure of the form (6.8) must be absolutely continuous with respect to the
measure .
6.7. Hahn and Jordan decompositions
To prove the Jordan decomposition of a signed measure, we first show that a
measure space can be decomposed into disjoint subsets on which a signed measure
is positive or negative, respectively. This is called the Hahn decomposition.
Definition 6.16. Suppose that is a signed measure on a measurable space
X. A set A X is positive for if it is measurable and (B) 0 for every
measurable subset B A. Similarly, A is negative for if it is measurable and
(B) 0 for every measurable subset B A, and null for if it is measurable and
(B) = 0 for every measurable subset B A.

72

6. DIFFERENTIATION

Because of the possible cancelation between the positive and negative signed
measure of subsets, (A) > 0 does not imply that A is positive for , nor does
(A) = 0 imply that A is null for . Nevertheless, as we show in the next result, if
(A) > 0, then A contains a subset that is positive for . The idea of the (slightly
tricky) proof is to remove subsets of A with negative signed measure until only a
positive subset is left.
Lemma 6.17. Suppose that is a signed measure on a measurable space (X, A).
If A A and 0 < (A) < , then there exists a positive subset P A such that
(P ) > 0.
Proof. First, we show that if A A is a measurable set with |(A)| < ,
then |(B)| < for every measurable subset B A. This is because takes
at most one infinite value, so there is no possibility of canceling an infinite signed
measure to give a finite measure. In more detail, we may suppose without loss of
generality that : A [, ) does not take the value . (Otherwise, consider
.) Then (B) 6= ; and if B A, then the additivity of implies that
(B) = (A) (A \ B) 6=
since (A) is finite and (A \ B) 6= .
Now suppose that 0 < (A) < . Let
1 = inf {(E) : E A and E A} .
Then 1 0, since A. Choose A1 A such that 1 (A1 ) 1 /2
if 1 is finite, or (A1 ) 1 if 1 = . Define a disjoint sequence of subsets
{Ai A : i N} inductively by setting
o
n
S
i1
A
i = inf (E) : E A and E A \
j
j=1

S
i1
and choosing Ai A \
j=1 Aj such that
i (Ai )

1
i
2

if < i 0, or (Ai ) 1 if i = .
Let

[
Ai ,
P = A \ B.
B=
i=1

Then, since the Ai are disjoint, we have

(B) =

(Ai ).

i=1

As proved above, (B) is finite, so this negative sum must converge. It follows that
(Ai ) 1 for only finitely many i, and therefore i is infinite for at most finitely
many i. For the remaining i, we have
X
1X
(Ai )
i 0,
2
P
so
i converges and therefore i 0 as i .

6.7. HAHN AND JORDAN DECOMPOSITIONS

73

If E P , then by construction (E) i for every sufficiently large i N.

Hence, taking the limit as i , we see that (E) 0, which implies that P is
positive. The proof also shows that, since (B) 0, we have
(P ) = (A) (B) (A) > 0,
which proves that P has strictly positive signed measure.

The Hahn decomposition follows from this result in a straightforward way.

Theorem 6.18 (Hahn decomposition). If is a signed measure on a measurable
space (X, A), then there is a positive set P and a negative set N for such that
P N = X and P N = . These sets are unique up to -null sets.
Proof. Suppose, without loss of generality, that (A) < for every A A.
(Otherwise, consider .) Let
m = sup{(A) : A A such that A is positive for },
and choose a sequence {Ai : i N} of positive sets such that (Ai ) m as i .
Then, since the union of positive sets is positive,
P =

Ai

i=1

is a positive set. Moreover, by the monotonicity of of , we have (P ) = m. Since

(P ) 6= , it follows that m 0 is finite.
Let N = X \P . Then we claim that N is negative for . If not, there is a subset
A N such that (A ) > 0, so by Lemma 6.17 there is a positive set P A with
(P ) > 0. But then P P is a positive set with (P P ) > m, which contradicts
the definition of m.
Finally, if P , N is another such pair of positive and negative sets, then
P \ P P N ,

so P \ P is both positive and negative for and therefore null, and similarly for
P \ P . Thus, the decomposition is unique up to -null sets.

To describe the corresponding decomposition of the signed measure into the
difference of measures, we introduce the notion of singular measures, which are
measures that are supported on disjoint sets.
Definition 6.19. Two measures , on a measurable space (X, A) are singular, written , if there exist sets M, N A such that M N = , M N = X
and (M ) = 0, (N ) = 0.
Example 6.20. The -measure in Example 2.36 and the Cantor measure in
Example 2.37 are singular with respect to Lebesgue measure on R (and conversely,
since the relation is symmetric).
Theorem 6.21 (Jordan decomposition). If is a signed measure on a measurable space (X, A), then there exist unique measures + , : A [0, ], one of
which is finite, such that
= +

and + .

74

6. DIFFERENTIATION

Proof. Let X = P N where P , N are positive, negative sets for . Then

+ (A) = (A P ),

(A) = (A N )

is the required decomposition. The values of are independent of the choice of

P , N up to a -null set, so the decomposition is unique.

We call + and the positive and negative parts of , respectively. The total
variation || of is the measure
|| = + + .

We say that the signed measure is -finite if || is -finite.

6.8. Radon-Nikodym theorem
The absolute continuity of measures is in some sense the opposite relationship
to the singularity of measures. If a measure singular with respect to a measure
, then it is supported on different sets from , while if is absolutely continuous
with respect to , then it supported on on the same sets as .
Definition 6.22. Let be a signed measure and a measure on a measurable
space (X, A). Then is absolutely continuous with respect to , written , if
(A) = 0 for every set A A such that (A) = 0.
Equivalently, if every -null set is a -null set. Unlike singularity,
absolute continuity is not symmetric.
Example 6.23. If is Lebesgue measure and is counting measure on B(R),
then , but 6 .
Example 6.24. If f : X R is a measurable function on a measure space
(X, A, ) whose integral with respect is well-defined as an extended real number
and the signed measure : A R is defined by
Z
f d,
(A) =
A

then (4.4) shows that is absolutely continuous with respect to .

The next result clarifies the relation between Definition 6.22 and the absolute
continuity property of integrable functions proved in Proposition 4.16.
Proposition 6.25. If is a finite signed measure and is a measure, then
if and only if for every > 0, there exists > 0 such that |(A)| <
whenever (A) < .
Proof. Suppose that the given condition holds. If (A) = 0, then |(A)| <
for every > 0, so (A) = 0, which shows that .
Conversely, suppose that the given condition does not hold. Then there exists
> 0 such that for every k N there exists a measurable set Ak with ||(Ak )
and (Ak ) < 1/2k . Defining
[

\
B=
Aj ,
k=1 j=k

we see that (B) = 0 but ||(B) , so is not absolutely continuous with respect
to .


6.8. RADON-NIKODYM THEOREM

75

The Radon-Nikodym theorem provides a converse to Example 6.24 for absolutely continuous, -finite measures. As part of the proof, from [4], we also show
that any signed measure can be decomposed into an absolutely continuous and
singular part with respect to a measure (the Lebesgue decomposition of ). In
the proof of the theorem, we will use the following lemma.
Lemma 6.26. Suppose that , are finite measures on a measurable space
(X, A). Then either , or there exists > 0 and a set P such that (P ) > 0
and P is a positive set for the signed measure .
Proof. For each n N, let X = Pn Nn be a Hahn decomposition of X for
the signed measure n1 . If

P =

Pn

N=

n=1

Nn ,

n=1

then X = P N is a disjoint union, and

1
(N )
n
for every n N, so (N ) = 0. Thus, either (P ) = 0, when , or (Pn ) > 0
for some n N, which proves the result with = 1/n.

0 (N )

Theorem 6.27 (Lebesgue-Radon-Nikodym theorem). Let be a -finite signed

measure and a -finite measure on a measurable space (X, A). Then there exist
unique -finite signed measures a , s such that
= a + s

where a and s .

Moreover, there exists a measurable function f : X R, uniquely defined up to

-a.e. equivalence, such that
Z
f d
a (A) =
A

for every A A, where the integral is well-defined as an extended real number.

Proof. It is enough to prove the result when is a measure, since we may

decompose a signed measure into its positive and negative parts and apply the
result to each part.
First, we assume that , are finite. We will construct a function f and an
absolutely continuous signed measure a such that
Z
f d
for all A A.
a (A) =
A

We write this equation as da = f d for short. The remainder s = a is the

singular part of .
Let F be the set of all A-measurable functions g : X [0, ] such that
Z
g d (A)
for every A A.
A

We obtain f by taking a supremum of functions from F . If g, h F , then

max{g, h} F . To see this, note that if A A, then we may write A = B C
where
B = A {x X : g(x) > h(x)} ,

C = A {x X : g(x) h(x)} ,

76

6. DIFFERENTIATION

and therefore
Z
A

max {g, h} d =

Let

g d +

m = sup

Z

h d (B) + (C) = (A).

g d : g F

(X).

Choose a sequence {gn F : n N} such that

Z
gn d = m.
lim
n

By replacing gn with max{g1 , g2 , . . . , gn }, we may assume that {gn } is an increasing

sequence of functions in F . Let
f = lim gn .
n

Then, by the monotone convergence theorem, for every A A we have

Z
Z
gn d (A),
f d = lim
n

so f F and

f d = m.

Define s : A [0, ) by

s (A) = (A)

f d.
A

Then s is a positive measure on X. We claim that s , which proves the result

in this case. Suppose not. Then, by Lemma 6.26, there exists > 0 and a set P
with (P ) > 0 such that s on P . It follows that for any A A
Z
f d + s (A)
(A) =
A
Z
f d + s (A P )

ZA
f d + (A P )

ZA
(f + P ) d.

It follows that f + P F but

Z
(f + P ) d = m + (P ) > m,
X

which contradicts the definition of m. Hence s .

If = a +s and = a +s are two such decompositions, then a a = s s
is both absolutely continuous and singular with respect to which implies that it
is zero. Moreover, f is determined uniquely by a up to pointwise a.e. equivalence.
Finally, if , are -finite measures, then we may decompose
X=

i=1

Ai

6.9. COMPLEX MEASURES

77

into a countable disjoint union of sets with (Ai ) < and (Ai ) < . We
decompose the finite measure i = |Ai as
i = ia + is

where ia i and is i .

Then = a + s is the required decomposition with

a =

ia ,

s =

ia

i=1

i=1

is the required decomposition.

The decomposition = a + s is called the Lebesgue decomposition of , and

the representation of an absolutely continuous signed measure as d = f d
is the Radon-Nikodym theorem. We call the function f here the Radon-Nikodym
derivative of with respect to , and denote it by
f=

d
.
d

Some hypothesis of -finiteness is essential in the theorem, as the following

example shows.
Example 6.28. Let B be the Borel -algebra on [0, 1], Lebesgue measure,
and counting measure on B. Then is finite and , but is not -finite.
There is no function f : [0, 1] [0, ] such that
Z
X
(A) =
f d =
f (x).
A

xA

There are generalizations of the Radon-Nikodym theorem which apply to measures that are not -finite, but we will not consider them here.
6.9. Complex measures
Complex measures are defined analogously to signed measures, except that they
are only permitted to take finite complex values.
Definition 6.29. Let (X, A) be a measurable space. A complex measure on
X is a function : A C such that:
(a) () = 0;
(b) if {Ai A : i N} is a disjoint collection of measurable sets, then
!

X
[
(Ai ).
Ai =

i=1

i=1

There is an analogous Radon-Nikodym theorems for complex measures. The

Radon-Nikodym derivative of a complex measure is necessarily integrable, since the
measure is finite.
Theorem 6.30 (Lebesgue-Radon-Nikodym theorem). Let be a complex measure and a -finite measure on a measurable space (X, A). Then there exist
unique complex measures a , s such that
= a + s

where a and s .

78

6. DIFFERENTIATION

Moreover, there exists an integrable function f : X C, uniquely defined up to

-a.e. equivalence, such that
Z
f d
a (A) =
A

for every A A.

To prove the result, we decompose a complex measure into its real and imaginary parts, which are finite signed measures, and apply the corresponding theorem
for signed measures.

CHAPTER 7

Lp spaces
In this Chapter we consider Lp -spaces of functions whose pth powers are integrable. We will not develop the full theory of such spaces here, but consider only
those properties that are directly related to measure theory in particular, density, completeness, and duality results. The fact that spaces of Lebesgue integrable
functions are complete, and therefore Banach spaces, is another crucial reason for
the success of the Lebesgue integral. The Lp -spaces are perhaps the most useful
and important examples of Banach spaces.
7.1. Lp spaces
For definiteness, we consider real-valued functions. Analogous results apply to
complex-valued functions.
Definition 7.1. Let (X, A, ) be a measure space and 1 p < . The space
Lp (X) consists of equivalence classes of measurable functions f : X R such that
Z
|f |p d < ,

where two measurable functions are equivalent if they are equal -a.e. The Lp -norm
of f Lp (X) is defined by
Z
1/p
p
.
kf kLp =
|f | d

The notation Lp (X) assumes that the measure on X is understood. We say

that fn f in Lp if kf fn kLp 0. The reason to regard functions that are
equal a.e. as equivalent is so that kf kLp = 0 implies that f = 0. For example, the
characteristic function Q of the rationals on R is equivalent to 0 in Lp (R). We will
not worry about the distinction between a function and its equivalence class, except
when the precise pointwise values of a representative function are significant.
Example 7.2. If N is equipped with counting measure, then Lp (N) consists of
all sequences {xn R : n N} such that

X
|xn |p < .
n=1

We write this sequence space as p (N), with norm

!1/p

X
p
k{xn }kp =
.
|xn |
n=1

The space L (X) is defined in a slightly different way. First, we introduce the
notion of esssential supremum.
79

7. Lp SPACES

80

Definition 7.3. Let f : X R be a measurable function on a measure space

(X, A, ). The essential supremum of f on X is
ess sup f = inf {a R : {x X : f (x) > a} = 0} .
X

Equivalently,



ess sup f = inf sup g : g = f pointwise a.e. .
X

Thus, the essential supremum of a function depends only on its -a.e. equivalence
class. We say that f is essentially bounded on X if
ess sup |f | < .
X

Definition 7.4. Let (X, A, ) be a measure space. The space L (X) consists
of pointwise a.e.-equivalence classes of essentially bounded measurable functions
f : X R with norm
kf kL = ess sup |f |.
X

In future, we will write

ess sup f = sup f.
We rarely want to use the supremum instead of the essential supremum when the
two have different values, so this notation should not lead to any confusion.
7.2. Minkowski and H
older inequalities
We state without proof two fundamental inequalities.
Theorem 7.5 (Minkowski inequality). If f, g Lp (X), where 1 p , then
f + g Lp (X) and
kf + gkLp kf kLp + kf kLp .

This inequality means, as stated previously, that k kLp is a norm on Lp (X)

for 1 p . If 0 < p < 1, then the reverse inequality holds
kf kLp + kgkLp kf + gkLp ,

so k kLp is not a norm in that case. Nevertheless, for 0 < p < 1 we have
|f + g|p |f |p + |g|p ,

so Lp (X) is a linear space in that case also.

To state the second inequality, we define the H
older conjugate of an exponent.
Definition 7.6. Let 1 p . The H
older conjugate p of p is defined by
1
1
=1
if 1 < p < ,
+
p p
and 1 = , = 1.
Note that 1 p , and the H
older conjugate of p is p.

Theorem 7.7 (Holders inequality). Suppose that (X, A, ) is a measure space

and 1 p . If f Lp (X) and g Lp (X), then f g L1 (X) and

Z
|f g| d kf kLp kgkLp .
For p = p = 2, this is the Cauchy-Schwartz inequality.

7.4. COMPLETENESS

81

7.3. Density
Density theorems enable us to prove properties of Lp functions by proving them
for functions in a dense subspace and then extending the result by continuity. For
general measure spaces, the simple functions are dense in Lp .
Theorem 7.8. Suppose that (X, A, ) is a measure space and 1 p .
Then the simple functions that belong to Lp (X) are dense in Lp (X).
Proof. It is sufficient to prove that we can approximate a positive function
f : X [0, ) by simple functions, since a general function may be decomposed
into its positive and negative parts.
First suppose that f Lp (X) where 1 p < . Then, from Theorem 3.12,
there is an increasing sequence of simple functions {n } such that n f pointwise.
These simple functions belong to Lp , and
p

|f n | |f |p L1 (X).

Hence, the dominated convergence theorem implies that

Z
p
|f n | d 0
as n ,

which proves the result in this case.

If f L (X), then we may choose a representative of f that is bounded.
According to Theorem 3.12, there is a sequence of simple functions that converges
uniformly to f , and therefore in L (X).

Note that a simple function
=

n
X

c i Ai

i=1

belongs to Lp for 1 p < if and only if (Ai ) < for every Ai such that
ci 6= 0, meaning that its support has finite measure. On the other hand, every
simple function belongs to L .
For suitable measures defined on topological spaces, Theorem 7.8 can be used to
prove the density of continuous functions in Lp for 1 p < , as in Theorem 4.27
for Lebesgue measure on Rn . We will not consider extensions of that result to more
general measures or topological spaces here.
7.4. Completeness
In proving the completeness of Lp (X), we will use the following Lemma.
Lemma 7.9. Suppose that X is a measure space and 1 p < . If
{gk Lp (X) : k N}

is a sequence of Lp -functions such that

X
kgk kLp < ,
k=1
p

then there exists a function f L (X) such that

X
gk = f
k=1

7. Lp SPACES

82

where the sum converges pointwise a.e. and in Lp .

Proof. Define hn , h : X [0, ] by
hn =

n
X

k=1

|gk | ,

h=

k=1

|gk | .

Then {hn } is an increasing sequence of functions that converges pointwise to h, so

the monotone convergence theorem implies that
Z
Z
p
h d = lim
hpn d.
n

By Minkowskis inequality, we have for each n N that

khn kLp

n
X

k=1

kgk kLp M

P
where k=1 kgk kLp = M . It follows that h Lp (X) withPkhkLp M , and in

particular that h is finite pointwise a.e. Moreover, the sum k=1 gk is absolutely
convergent pointwise a.e., so it converges pointwise a.e. to a function f Lp (X)
with |f | h. Since
p

!p
n
n


X
X


|gk |
(2h)p L1 (X),
gk  |f | +
f


k=1

k=1

the dominated convergence theorem implies that

p
Z 
n

X


as n ,
gk  d 0
f


k=1

meaning that

k=1

gk converges to f in Lp .

The following theorem implies that Lp (X) equipped with the Lp -norm is a
Banach space.
Theorem 7.10 (Riesz-Fischer theorem). If X is a measure space and 1 p
, then Lp (X) is complete.
Proof. First, suppose that 1 p < . If {fk : k N} is a Cauchy sequence
in Lp (X), then we can choose a subsequence {fkj : j N} such that


fkj+1 fkj p 1 .
L
2j

Writing gj = fkj+1 fkj , we have

X
j=1

kgj kLp < ,

so by Lemma 7.9, the sum

fk1 +

X
j=1

gj

7.5. DUALITY

83

converges pointwise a.e. and in Lp to a function f Lp . Hence, the limit of the

subsequence
!
j1

X
X
gj = f
gi = fk1 +
lim fkj = lim fk1 +
j

j=1

i=1

exists in L . Since the original sequence is Cauchy, it follows that

lim fk = f

in Lp . Therefore every Cauchy sequence converges, and Lp (X) is complete when

1 p < .
Second, suppose that p = . If {fk } is Cauchy in L , then for every m N
there exists an integer n N such that we have
(7.1)

|fj (x) fk (x)| <

1
m

c
for all j, k n and x Nj,k,m

where Nj,k,m is a null set. Let

N=

Nj,k,m .

j,k,mN

Then N is a null set, and for every x N c the sequence {fk (x) : k N} is Cauchy
in R. We define a measurable function f : X R, unique up to pointwise a.e.
equivalence, by
f (x) = lim fk (x)
for x N c .
k

Letting k in (7.1), we find that for every m N there exists an integer n N

such that
1
|fj (x) f (x)|
for j n and x N c .
m
It follows that f is essentially bounded and fj f in L as j . This proves
that L is complete.

One useful consequence of this proof is worth stating explicitly.
Corollary 7.11. Suppose that X is a measure space and 1 p < . If {fk }
is a sequence in Lp (X) that converges in Lp to f , then there is a subsequence {fkj }
that converges pointwise a.e. to f .
As Example 4.26 shows, the full sequence need not converge pointwise a.e.
7.5. Duality
The dual space of a Banach space consists of all bounded linear functionals on
the space.
Definition 7.12. If X is a real Banach space, the dual space of X consists
of all bounded linear functionals F : X R, with norm


|F (x)|
< .
kF kX = sup
xX\{0} kxkX

7. Lp SPACES

84

A linear functional is bounded if and only if it is continuous. For Lp spaces,

we will use the Radon-Nikodym theorem to show that Lp (X) may be identified

with Lp (X) for 1 < p < . Under a -finiteness assumption, it is also true that
1

L (X) = L (X), but in general L (X) 6= L1 (X).

H
olders inequality implies that functions in Lp define bounded linear funcp
tionals on L with the same norm, as stated in the following proposition.
Proposition 7.13. Suppose that (X, A, ) is a measure space and 1 < p .

If f Lp (X), then
Z
F (g) = f g d

defines a bounded linear functional F : Lp (X) R, and

kF kLp = kf kLp .

If X is -finite, then the same result holds for p = 1.

Proof. From H
olders inequality, we have for 1 p that
|F (g)| kf kLp kgkLp ,

which implies that F is a bounded linear functional on Lp with

kF kLp kf kLp .

In proving the reverse inequality, we may assume that f 6= 0 (otherwise the result
is trivial).
First, suppose that 1 < p < . Let

p /p
|f |
g = (sgn f )
.
kf kLp

Then g Lp , since f Lp , and kgkLp = 1. Also, since p /p = p 1,

p 1

Z
|f |
d
F (g) = (sgn f )f
kf kLp
= kf kLp .
Since kgkLp = 1, we have kF kLp |F (g)|, so that
kF kLp kf kLp .

If p = , we get the same conclusion by taking g = sgn f L . Thus, in these

cases the supremum defining kF kLp is actually attained for a suitable function g.
Second, suppose that p = 1 and X is -finite. For > 0, let
A = {x X : |f (x)| > kf kL } .

Then 0 < (A) . Moreover, since X is -finite, there is an increasing sequence

of sets An of finite measure whose union is A such that (An ) (A), so we can
find a subset B A such that 0 < (B) < . Let
B
g = (sgn f )
.
(B)
Then g L1 (X) with kgkL1 = 1, and
Z
1
|f | d kf kL .
F (g) =
(B) B

7.5. DUALITY

85

It follows that
kF kL1 kf kL ,

and therefore kF kL1 kf kL since > 0 is arbitrary.

This proposition shows that the map F = J(f ) defined by

Z
p
p

(7.2)
J : L (X) L (X) ,
J(f ) : g 7 f g d,

is an isometry from Lp into Lp . The main part of the following result is that J is
onto when 1 < p < , meaning that every bounded linear functional on Lp arises

in this way from an Lp -function.

The proof is based on the idea that if F : Lp (X) R is a bounded linear
functional on Lp (X), then (E) = F (E ) defines an absolutely continuous measure

on (X, A, ), and its Radon-Nikodym derivative f = d/d is the element of Lp

corresponding to F .
Theorem 7.14 (Dual space of Lp ). Let (X, A, ) be a measure space. If 1 <

p < , then (7.2) defines an isometric isomorphism of Lp (X) onto the dual space
p
of L (X).
Proof. We just have to show that the map J defined in (7.2) is onto, meaning

that every F Lp (X) is given by J(f ) for some f Lp (X).

First, suppose that X has finite measure, and let
F : Lp (X) R

be a bounded linear functional on Lp (X). If A A, then A Lp (X), since X has

finite measure, and we may define : A R by
(A) = F (A ) .
If A =

i=1

Ai is a disjoint union of measurable sets, then

A =

Ai ,

i=1

and the dominated convergence theorem implies that

n


X


Ai 0
A
p

i=1

as n . Hence, since F is a continuous linear functional on Lp ,

!

X
X
X
(Ai ),
F (Ai ) =
Ai =
(A) = F (A ) = F
i=1

i=1

i=1

meaning that is a signed measure on (X, A).

If (A) = 0, then A is equivalent to 0 in Lp and therefore (A) = 0 by
the linearity of F . Thus, is absolutely continuous with respect to . By the
Radon-Nikodym theorem, there is a function f : X R such that d = f d and
Z
F (A ) = f A d
for everyA A.

7. Lp SPACES

86

Hence, by the linearity and boundedness of F ,

Z
F () = f d
for all simple functions , and
Z



 f d M kkLp



where M = kF kLp .
Taking = sgn f , which is a simple function, we see that f L1 (X). We may
then extend the integral of f against bounded functions by continuity. Explicitly,
if g L (X), then from Theorem 7.8 there is a sequence of simple functions {n }
with |n | |g| such that n g in L , and therefore also in Lp . Since
|f n | kgkL |f | L1 (X),
the dominated convergence theorem and the continuity of F imply that
Z
Z
F (g) = lim F (n ) = lim
f n d = f g d,
n

and that
(7.3)

Z



 f g d M kgkLp



for every g L (X).

Next we prove that f Lp (X). We will estimate the Lp norm of f by a

similar argument to the one used in the proof of Proposition 7.13. However, we
need to apply the argument to a suitable approximation of f , since we do not know

a priori that f Lp .
Let {n } be a sequence of simple functions such that
n f

pointwise a.e. as n

and |n | |f |. Define
gn = (sgn f )

|n |
kn kLp

p /p

Then gn L (X) and kgn kLp = 1. Moreover, f gn = |f gn | and

Z
|n gn | d = kn kLp .
It follows from these equalities, Fatous lemma, the inequality |n | |f |, and (7.3)
that
kf kLp lim inf kn kLp
n
Z
lim inf |n gn | d
n
Z
lim inf |f gn | d
n

M.

7.5. DUALITY

87

Thus, f Lp . Since the simple functions are dense in Lp and g

7
f g d is a
R

continuous functional on Lp when f Lp , it follows that F (g) = f g d for every

g Lp (X). Proposition 7.13 then implies that
kF kLp = kf kLp ,

which proves the result when X has finite measure.

The extension to non-finite measure spaces is straightforward, and we only
outline the proof. If X is -finite, then there is an increasing sequence {An } of sets
with finite measure whose union is X. By the previous result, there is a unique

function fn Lp (An ) such that

Z
fn g d
for all g Lp (An ).
F (g) =
An

If m n, the functions fm , fn are equal pointwise a.e. on An , and the dominated

convergence theorem implies that f = limn fn Lp (X) is the required function.

Finally, if X is not -finite, then

X, let fA Lp (A)
R for each -finite subset A
be the function such that F (g) = A fA g d for every g Lp (A). Define
n
o
M = sup kfA kLp (A) : A X is -finite kF kLp (X) ,
and choose an increasing sequence of sets An such that

Defining B =

kfAn kLp (An ) M

n=1

as n .

An , one may verify that fB is the required function.

A Banach space X is reflexive if its bi-dual X is equal to the original space

X under the natural identification
: X X

where (x)(F ) = F (x) for every F X ,

meaning that x acting on F is equal to F acting on x. Reflexive Banach spaces

are generally better-behaved than non-reflexive ones, and an immediate corollary
of Theorem 7.14 is the following.
Corollary 7.15. If X is a measure space and 1 < p < , then Lp (X) is
reflexive.
Theorem 7.14 also holds if p = 1 provided that X is -finite, but we omit a
detailed proof. On the other hand, the theorem does not hold if p = . Thus L1
and L are not reflexive Banach spaces, except in trivial cases.
The following example illustrates a bounded linear functional on an L -space
that does not arise from an element of L1 .
Example 7.16. Consider the sequence space (N). For
x = {xi : i N} (N),

kxk = sup |xi | < ,

iN

define Fn (N) by

Fn (x) =

1X
xi ,
n i=1

meaning that Fn maps a sequence to the mean of its first n terms. Then
kFn k = 1

7. Lp SPACES

88

for every n N, so by the Alaoglu theorem on the weak- compactness of the unit
ball, there exists a subsequence {Fnj : j N} and an element F (N) with

kF k 1 such that Fnj F in the weak- topology on .

If u is the unit sequence with ui = 1 for every i N, then Fn (u) = 1 for
every n N, and hence
F (u) = lim Fnj (u) = 1,
j

so F 6= 0; in fact, kF k = 1. Now suppose that there were y = {yi } 1 (N) such

that

X
xi yi
for every x .
F (x) =
i=1

Then, denoting by ek the sequence with kth component equal to 1 and all
other components equal to 0, we have
1
=0
yk = F (ek ) = lim Fnj (ek ) = lim
j nj
j

so y = 0, which is a contradiction. Thus, (N) is strictly larger than 1 (N).

We remark that if a sequence x = {xi } has a limit L = limi xi , then
F (x) = L, so F defines a generalized limit of arbitrary bounded sequences in terms
of their Ces`
aro sums. Such bounded linear functionals on (N) are called Banach
limits.
It is possible to characterize the dual of L (X) as a space ba(X) of bounded,
finitely additive, signed measures that are absolutely continuous with respect to
the measure on X. This result is rarely useful, however, since finitely additive
measures are not easy to work with. Thus, for example, instead of using the weak
topology on L (X), we can regard L (X) as the dual space of L1 (X) and use the
corresponding weak- topology.

Bibliography
[1] V. I. Bogachev, Measure Theory, Vol I. and II., Springer-Verlag, Heidelberg, 2007.
[2] D. L. Cohn, Measure Theory, Birkh
auser, Boston, 1980.
[3] L. C. Evans and R. F. Gariepy, Measure Theory and Fine Properties of Functions, CRC
Press, Boca Raton, 1992.
[4] G. B. Folland, Real Analysis, 2nd ed., Wiley, New York, 1999.
[5] F. Jones, Lebesgue Integration on Euclidean Space, Revised Ed., Jones and Bartlett, Sudberry, 2001.
[6] S. Lang, Real and Functional Analysis, Springer-Verlag, 1993.
[7] E. H. Lieb and M. Loss, Analysis, AMS 1997.
[8] E. M. Stein and R. Shakarchi, Real Analysis, Princeton University Press, 2005.
[9] M. E. Taylor, Measure Theory and Integration, American Mathematical Society, Providence,
2006.
[10] S. Wagon, The Banach-Tarski Paradox, Encyclopedia of Mathematics and its Applications,
Vol. 24 Cambridge University Press, 1985.
[11] R. L. Wheeler and Z. Zygmund, Measure and Integral, Marcel Dekker, 1977.

89