Abstract

Notes for MAA-5616 Measure Theory I, taught by Dr. Alexander

Reznikov in Spring 2019. In these notes, AnB := A \ B c . Also, the
statement “A B” does not necessarily mean that A is a proper subset
of B. Homeworks have been included in these notes and are labelled as
“Problem”. Knowledge of basic set theory is assumed, as is familiarity
with an introductory course on Real Analysis.
I accept sole responsibility for errors, of which I believe there will be
many. Please feel free to o¤er feedback at amalik at math dot fsu dot edu
 INTRODUCTION TO MEASURE
THEORY

July 31, 2019

Contents
1 Measure 1
1.1 General Measure Spaces . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Completion of Measure . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 Construction of Measures . . . . . . . . . . . . . . . . . . . . . . 11
1.3.1 Lebesgue-Carathéodory Theorem . . . . . . . . . . . . . . 16
1.4 Lebesgue Measure . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1.4.1 A non-measurable Set . . . . . . . . . . . . . . . . . . . . 28
1.4.2 Cantor Set . . . . . . . . . . . . . . . . . . . . . . . . . . 30

2 Functions on Measure Spaces 33

2.1 Measurable Functions . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2 Simple Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.2.1 Approximation of Measurable Functions . . . . . . . . . . 40

3 Integration 46
3.1 Integral of Simple and Measurable Functions . . . . . . . . . . . 46
3.2 Integral of Sequence of Functions . . . . . . . . . . . . . . . . . . 52
3.3 Integral as a Measure . . . . . . . . . . . . . . . . . . . . . . . . 59
3.4 Integral of Continuous Functions . . . . . . . . . . . . . . . . . . 63
3.5 Convergence in Measure . . . . . . . . . . . . . . . . . . . . . . . 71
3.6 Riemann Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.7 Product Measures . . . . . . . . . . . . . . . . . . . . . . . . . . 78

4 Classi…cation of Functions 89
4.1 Di¤erentiability . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.2 Functions of Bounded Variation . . . . . . . . . . . . . . . . . . . 100
4.3 A Pathological Function . . . . . . . . . . . . . . . . . . . . . . . 115

5 Di¤erentiation of Measures 116

5.1 Signed Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
 Syllabus
The purpose of this course is to introduce the notion of measure spaces
and integrals against general measures; in particular, to construct the Lebesgue
measure and Lebesgue Integral
Text book: Real Analysis by Royden and Fitzpatrick
Course Objectives: The purpose of this course is to introduce the no-
tion of measure spaces and integrals against general measures; in particular, to
construct the Lebesgue measure and Lebesgue integral.
Homeworks: Weekly graded homeworks will be given. Students are en-
couraged to collaborate on homeworks, but every student should write the …nal
solutions on his or her own. It is crucial for passing the qualifying exam to
understand every homework problem in all details.
Grading: There will be a midterm, a …nal exam, and weekly homeworks
University Attendance Policy: Excused absences include documented
illness, deaths in the family and other documented crises, call to active military
duty or jury duty, religious holy days, and o¢cial University activities. These
absences will be accommodated in a way that does not arbitrarily penalize
students who have a valid excuse. Consideration will also be given to students
whose dependent children experience serious illness.
Academic Honor Policy: The Florida State University Academic Honor
Policy outlines the University’s expectations for the integrity of students’ acad-
emic work, the procedures for resolving alleged violations of those expectations,
and the rights and responsibilities of students and faculty members throughout
the process. Students are responsible for reading the Academic Honor Policy
and for living up to their pledge to “...be honest and truthful and ... [to] strive for
personal and institutional integrity at Florida State University." (Florida State
University Academic Honor Policy, found at http://fda.fsu.edu/Academics/AcademicHonor-
Policy.) To summarize, violations of these policies will result in a rather messy
a¤air for you and me, so just don’t do it.
Americans with Disabilities Act: Students with disabilities needing aca-
demic accommodation should: (1) register with and provide documentation to
the Student Disability Resource Center; and (2) bring a letter to the instruc-
tor indicating the need for accommodation and what type. Please note that
instructors are not allowed to provide classroom accommodation to a student
until appropriate veri…cation from the Student Disability Resource Center has
been provided.
This syllabus and other class materials are available in alternative format
upon request. For more information about services available to FSU students
with disabilities, contact the Student Disability Resource Center
874 Traditions Way
108 Student Services Building Florida State University
Tallahassee, FL 32306-4167
(850) 644-9566 (voice)
(850) 644-8504 (TDD)
 [email protected]
http://www.disabilitycenter.fsu.edu/
Syllabus Change Policy: Except for changes that substantially a¤ect
implementation of the evaluation (grading) statement, this syllabus is a guide
for the course and is subject to change with advance notice.
1 Measure
1.1 General Measure Spaces
In general, a measure is a function applied on a collection of sets. Both the
function and the domain have special properties. The ultimate idea is to de…ne
a function that somehow “measures” a decent set. This measurement may
be in the form of length, volume, number of elements or other appropriate
sense of measure. In general, the function itself must follow some commonsense
properties, one of them being that the measure of union of disjoint sets must be
the sum of measures.
To get other commonsense properties, the collection on which the function
acts must have some structure to them. This is de…ned as follows: for any non-
empty set X, the collection A of subsets of X is called a -algebra, denoted
by (X; A), a measure space, if

S1 X 2 A
S2 A; B 2 A =) AnB 2 A
1
[
S3 fAi : i 2 Ng A =) Ai 2 A
i=1

Elements of a -algebra are called measurable.

For any -algebra A, ? 2 A and A is also closed under countable intersec-
tions.
Proof. For the …rst claim, let A 2 A. Then, AnA 2 A by S2 so that ? 2 A.
Let fAi : i 2 Ng A. Since X 2 A by S1, for each i, XnAi = Aci 2 A. We
also have
[1
Aci 2 A
i=1

by S3. Since X 2 A, again by S1, we have

1
[ 1
\
Xn Aci 2 A () Ai 2 A
i=1 i=1

De…nition 1 Let (X; A) be a -algebra. A measure on (X; A) is a function

: A ! [0; 1] such that

M1 (?) = 0, and
M2 For a disjoint family fAi : i 2 Ng A,
1
! 1
[ X
Ai = (Ai )
i=1 i=1

1
 The triplet (X; A; ) is then known as a measurable space. The property
M2 is known as the countably additive property.

Example 2 Let X be any (countable) set. De…ne : 2X ! [0; 1] as (A) =

jAj.

Such a is called the counting measure. For our second example, we

will need to be a little more precise since that concerns what’s called the Borel
-algebra on R, L. L is the smallest -algebra that contains all open sets of R.
Obviously, L cannot contain only open sets A of R since, by de…nition, we must
have A 2 L =) Ac 2 L, where Ac “should” be open as well, but that is not
the case. The -algebra is generated by adding all countable unions, countable
intersections, and relative complements of all open (under the usual topology)
subsets of R. A process, called completion of a measurable space which we
will see in §1.2, turns (R; L; m1 ), the measurable space using Borel -algebra
with the Lebesgue Measure, to (R; M1 ; m1 ), the Lebesgue Measurable Space.
This fact will not be proved in these notes but the machinery to do so will be
developed as we move on, as will many details of the latter space. And we have
already moved ahead of ourselves. At any rate, a measure has the following
properties:

Lemma 3 Let X be a non-empty set and let be a measure on (X; A). Then,

Finite Additivity If fAi : 1 i ng A is a collection of pairwise disjoint sets, then

n
! n
[ X
Ai = (Ai )
i=1 i=1

Monotonicity A; B 2 A with A B =) (A) (B)

Complementation A; B 2 A with A B and (A) < 1 =) (BnA) = (B) (A)
Subadditivity If fAi : i 2 Ng A, then
1
! 1
[ X
Ai (Ai )
i=1 i=1

Proof. Finite additivity

Let An+k = ? for k = 1; 2; :::. Then, fAi : i 2 Ng A is pairwise disjoint
so that, ! !
n
[ 1
[ X1
Ai = Ai = (Ai )
i=1 i=1 i=1

by M2 and
n
! n n
[ X X
Ai = (Ai ) + 0 + 0 + ::: = (Ai )
i=1 i=1 i=1

2
by M1
Monotonicity
Given that A B, we can write B = A [ BnA. Then, by previous part,
(B) = (A [ BnA) = (A) + (BnA). Since BnA 2 A by S2, (BnA) 0.
Thus, (A) (B)
Complementation
Proceeding as above, we have (B) = (A) + (BnA) but now we can
subtract the real number (A) on both sides as we have (A) < 1.
Subadditivity
Let B1 = A1 and
n[1
Bn = An n Ai
i=1

Then, the family fBi : i 2 Ng is pairwise disjoint: assume that i < j (the argu-
ment for i > j is similar). Then,
i[1
! j[1
!
Bi \ Bj = Ai n Ak \ Aj n Am
k=1 m=1
i\1 j\1
= Ai \ Ack \ Aj \ Acm = ?
k=1 m=1

since Ai \ Acm0 = ? for some m0 with 1 m0 = i < j (and, obviously,

Ai \ Aci = ?). Thus,
1
! 1 1
[ X X
Bi = (Bi ) (Ai )
i=1 i=1 i=1

since Bi Ai for each i. We will be done if we can prove that

1
[ 1
[
Bi = Ai
i=1 i=1

Since Bi Ai for each i, we must have

1
[ 1
[
Bi Ai
i=1 i=1

For the converse, let

1
[
x2 Ai
i=1

De…ne I = fj : x 2 Aj g N. By the well-ordering principle, I has a least

element, say m. Then,
m
[1
x 2 Bm = Am n Ai
i=1

3
so that both inclusions hold.
The following properties mimic continuity properties.
Theorem 4 Let fAi : i 2 Ng be a family of measurable sets. If this family is
increasing and nested, that is, we have a sequence of sets A1 A2 :::, then
1
!
[
Ai = lim (An )
n!1
i=1

In this case, is said to be continuous from below. On the other hand, if

we have a nested decreasing sequence A1 A2 ::: and (A1 ) < 1, then
1
!
\
Ai = lim (An )
n!1
i=1

In this case, is said to be continuous from above.

Proof. For …rst part, let E1 = A1 and En = An nAn 1 . This family is mea-
surable by S2 and disjoint: WLOG assume that i < j. Then, Ei \ Ej =
Ai \ Aci 1 \ Aj \ Acj 1
c
= Ai \ Aj \ (Ai 1 [ Aj 1 )
= Ai \ Acj 1 = ? because i j 1
Moreover,
1
[ 1
[
Ai = Ei
i=1 i=1
One side of the inclusion is obtained by observing that Ei Ai for each i, which
gives us
1
[ 1
[
Ai Ei
i=1 i=1
For the reverse inclusion, let
1
[
x2 Ai
i=1
De…ne I = fj : x 2 Aj g N. By the well-ordering principle, I has a least
element, say m. Then, x 2 Em = Am nAm 1 with A0 := ?, so that
1
[
x2 Ei
i=1

Then,
1
! 1
! 1
[ [ X
Ai = Ei = (Ei )
i=1 i=1 i=1
Xn n
X
= lim (Ei ) = lim (Ai nAi 1)
n!1 n!1
i=1 i=1
Xn
= lim ( (Ai ) (Ai 1 ))
n!1
i=1

4
 The last step is possible if we assume that (Ai ) < 1 for each i, otherwise
we would have 1 = 1, so that equality still holds and the proof withstands
scrutiny. Thus, !
1
[
Ai = lim (An ) (A0 )
n!1
i=1
Since we’ve assumed that A0 = ?, we have (A0 ) = ? and we’re done.
For the second part, let B1 = ?, B2 = A1 nA2 , B3 = A1 nA3 , ... . Then,
Bi is an increasing sequence of sets: Bi = A1 nAi and Bi+1 = A1 nAi+1 . Since
Ai+1 Ai , we have A1 nAi A1 nAi+1 , or that Bi Bi+1 . Thus,
1
!
[
Bi = lim (Bn ) = lim ( (A1 ) (An )) (1)
n!1 n!1
i=1

To …nish the proof, we …rst show that

1
[ 1
\
Bi = A1 n Ai
i=1 i=1

To see this, note that B1 = A1 nA1 = ?. B2 = B1 [B2 = A1 nA2 = A1 n (A1 \ A2 ).

This takes care of the base step for induction. Now, let
k
[ k
\
Bi = A1 n Ai
i=1 i=1

Then,
k+1
[
Bi = Bk+1 = A1 nAk+1
i=1
k+1
\
= A1 n Ai
i=1

We can therefore have

1
! 1
!
[ \
Bi = (A1 ) Ai (2)
i=1 i=1

Equating (1) and (2), we get

1
!
\
(A1 ) lim (An ) = (A1 ) Ai
n!1
i=1

Since (A1 ) < 1, then we can cancel this real number from both sides to get
1
!
\
Ai = lim (An )
n!1
i=1

The converse also holds under di¤erent assumptions:

5
Problem 5 Let X be a non-empty set and let (X; A) be a -algebra. Let :
A ! [0; 1] be …nitely additive and, for every nested increasing sequence A1
A2 ::: in A, !
1
[
Ai = lim (An )
n!1
i=1

Then, is a measure.

Solution 6 M1 Let A 2 A such that (A) < 1. Then, (A [ ?) = (A) =

(A) + (?) since A and ? are disjoint, so that (?) = 0.
M2 Let fBi : i 2 Ng be a sequence of disjoint sets. We need to show that
1
! 1
[ X
Bi = (Bi )
i=1 i=1

De…ne
n
[
An = Bi
i=1

with A1 = B1 . Then,
n
[ n+1
[
An = Bi Bi = An+1
i=1 i=1

gives us a family fAi : i 2 Ng of an increasing sequence in A so that

1
! n
! n
[ [ X
Ai = lim (An ) = lim Bi = lim (Bi )
n!1 n!1 n!1
i=1 i=1 i=1

by …nite additivity of so that

1
! 1
[ X
Ai = (Bi )
i=1 i=1

It remains to show that

1
[ 1
[
Ai = Bi
i=1 i=1

Let
1
[
x2 Ai
i=1

Then, 9j such that

j
[ 1
[
x 2 Aj = Bi =) x 2 Bi
i=1 i=1

6
Thus,
1
[ 1
[
Ai Bi
i=1 i=1
Conversely, since Bi Ai for each i, it follows that
1
[ 1
[
Bi Ai
i=1 i=1

and, therefore ! !
1
[ 1
[ 1
X
Bi = Ai = (Bi )
i=1 i=1 i=1

Problem 7 Let X be a non-empty set and let (X; A) be a -algebra. Let

: A ! [0; 1] be …nitely additive. Then, is a measure if for every nested
decreasing sequence A1 A2 ::: in A with (A1 ) < 1,
1
!
\
Ai = lim (An )
n!1
i=1

Solution 8 Let fBi : i 2 Ng be family of disjoint sets. Let

n
[
An = Bi
i=1

Then, An An+1 so that we have a nested decreasing sequence Acn+1 Acn .

Let us …rst show that
[1 [1
Ai = Bi
i=1 i=1
To this end, we …rst note that
1
[ j
[ 1
[
8x 2 Ai ; 9j such that x 2 Aj = Bi Bi
i=1 i=1 i=1

Conversely,
1
[ j
[ 1
[
x2 Bi =) x 2 Bj for some j and Bj Bi = Aj Ai
i=1 i=1 i=1

It follows that
1 1 1
! 1
!
\ \ \ \
Aci = Bic =) Bic = Aci = lim (Acn )
n!1
i=1 i=1 i=1 i=1

Since is …nitely additive, we must have (X) = (C [ C c ) = (C) + (C c ).

That is, (C) = (X) (C c ). Of course we can only say this, provided that

7
 (C c ) < 1, which we will take the liberty of assuming (in the other case, the
equality still holds). Moreover, we can denote
1
\
Aci = Ac
i=1

Then,
1
! 1
!
[ \
c
(A) = Ai = (X) (A ) = (X) Aci = (X) lim (Acn )
n!1
i=1 i=1

That is,
1
!c ! 1
!c ! n
[ [ X
Ai = Bi = (X) lim (An ) = (X) lim (Bi )
n!1 n!1
i=1 i=1 i=1

Thus, !
1
[ n
X
Bi = lim (Bi )
n!1
i=1 i=1

De…nition 9 We say that a property holds almost everywhere or for almost

all x 2 X if the measure of the set for which the property does not hold is zero.

Lemma 10 (Borel-Cantelli) If we have a family of measurable sets fEi : i 2 Ng

and their union has …nite measure, i.e.,
1
X
(Ei ) < 1,
i=1

then almost every x 2 X belongs to at most …nitely many Ej0 s.

Proof. We need to show that the set F = fx 2 X : x belongs to in…nitely many Ej ’sg
is measurable and has zero measure. Let x 2 F . Then, for every N 2 N,
1
[
x2 En
n=N

Since this is true for every N , then

1 [
\ 1
x2 En
N =1n=N

Conversely,
1 [
\ 1
x2 En =) x 2 F
N =1n=N

8
Thus,
1 [
\ 1
F = En
N =1n=N

so that F is measurable. Now, note that

1
[ 1
[ 1
[
En En En :::
n=1 n=2 n=3

Then, by continuity from above, we have that

1 [ 1
! 1
! 1
\ [ X
(F ) = Ei = lim Ei lim (Ei ) = 0
N !1 N !1
N =1i=N i=N i=N

because the union of fEi : i 2 Ng has …nite measure.

1.2 Completion of Measure

Let (X; A; ) be a measure space and let

M = fE X : E = A [ B s.t. A 2 A ^ B C: (C) = 0g

That is, we add sets to a collection with measure zero. Then, M is a -algebra.
Proof. We need to show that (a) X 2 M, (b) if E; F 2 M, then EnF 2 M and
(c) M is closed under countable unions
(a) Note that for each A 2 A, A = A [ ? and ? ? and (?) = 0, so that
A 2 M. Thus, A M so that X; ? 2 M.
(b) We will …rst show that for any E 2 M, E c 2 M and then show that
M is closed under intersection. These two facts together will show that for any
E; F in M, EnF = E \ F c 2 M.
Let E 2 M where E = A [ B with A 2 A and B C such that (C) = 0.
Note that C B implies C c B c so that C c [ B c = B c and that
B c = X \ B c = (C c [ C) \ (C c [ B c ) = C c [ (C \ B c ).
Then, E c = Ac \ B c = Ac \ (C c [ (C \ B c ))
= (Ac \ C c ) [ (Ac \ C \ B c )
Now, since A; C 2 A, then Ac ; C c 2 A and so Ac \ C c 2 A. Also, by
de…nition, Ac \ C \ B c C and by assumption (C) = 0 so that (Ac \ C c ) [
(Ac \ C \ B c ) = E c 2 M.
Next, let Let E; F 2 M. Then, 9A1 ; A2 ; C1 ; C2 2 A such that E = A1 [ B1 ,
F = A2 [ B2 and Bi Ci such that (Ci ) = 0 for i = 1; 2. Then, E \ F =
(A1 [ B1 ) \ (A2 [ B2 )
= (A1 \ (A2 [ B2 )) [ (B1 \ (A2 [ B2 ))
= (A1 \ A2 ) [ (A1 \ B2 ) [ (B1 \ A2 ) [ (B1 \ B2 )
Now, A1 \ A2 2 A and A1 \ B2 B2 C2 , B1 \ A2 B1 C1 and B1 \ B2
is a subset of C1 (and C2 ). Thus, (A1 \ B2 ) [ (B1 \ A2 ) [ (B1 \ B2 ) C1 [ C2 .
Note that (C1 [ C2 ) (C1 ) + (C2 ) = 0 + 0 = 0 so that (C1 [ C2 ) = 0
since (A) 0 for all A 2 A. Thus, E \ F 2 M.

9
 We can now assert that for any E; F 2 M, EnF = E \ F c 2 M.
(c) Assume that fEi : i 2 Ng M is a family of disjoint sets, where, for
each i, Ei = Ai [ Bi where Ai 2 A and Bi Ci such that (Ci ) = 0. Then,
1
[ 1
[ 1
[ 1
[
Ei = (Ai [ Bi ) = Ai [ Bi
i=1 i=1 i=1 i=1

Again,
1
[ 1
[ 1
[
Ai 2 A and Bi Ci
i=1 i=1 i=1

and that
1
! 1 1
! 1
[ X [ [
Ci (Ci ) = 0 =) Ci = 0 =) Ei 2 M
i=1 i=1 i=1 i=1

furnishing a proof of the third claim.

For such E, de…ne : M ! [0; 1] by (E) = (A). This is a well-de…ned
measure on M.
Proof. To show that is well de…ned, assume that E1 = A1 [ B1 and E2 =
A2 [ B2 with B1 C1 , B2 C2 and (Ci ) = 0 for i = 1; 2. We need to show
that E1 = E2 =) (E1 ) = (E2 ). Now, E1 = E2
=) A1 [ B1 = A2 [ B2
=) A1 [ B1 [ C1 [ C2 = A2 [ B2 [ C1 [ C2
=) A1 [ C1 [ C2 = A2 [ C1 [ C2 since Bi Ci for i = 1; 2.
Now, since is measure on A, A1 [ C1 [ C2 = A2 [ C1 [ C2 2 A, Ai 2 A
and A1 A1 [ C1 [ C2 , then (A1 ) (A1 [ C1 [ C2 ) = (A2 [ C1 [ C2 )
(A2 ) + (C1 ) + (C2 ) = (A2 ) + 0 + 0 = (A2 ).
Similarly, (A2 ) (A2 [ C1 [ C2 ) = (A1 [ C1 [ C2 ) (A1 ). Thus,
(A1 ) = (A2 ), which means that (E1 ) = (E2 )
We now show that is a measure.
M1 Since A M by (a) above, then jA = , by de…nition. Thus, (?) =
(?) = 0.
M2 Let fEi : i 2 Ng M, where Ei = Ai [ Bi for each i, be a pairwise
disjoint family of M-measurable subsets. Then,
1
! 1
! 1 1
[ [ X X
Ei = Ai = (Ai ) = (Ei )
i=1 i=1 i=1 i=1

making a bona …de measure.

Problem 11 Let (X; A; ) be a measure space, E 2 A and (E) > 0. De-

…ne AE := fA E : A 2 Ag. Show that that AE is a -algebra on E and the
restriction of to E is a measure.

10
Solution 12 Since E E, E 2 AE . Let A; B 2 AE . Then, A E and
A; B 2 A (so AnB 2 A). Since AnB A E, we have that AnB E. Thus,
AnB 2 AE . Finally, let fEi : i 2 Ng AE . Then, for each i, Ei E and
fEi : i 2 Ng A. These two facts imply that
1
[
Ei E
i=1

by properties of sets and, respectively,

1
[
Ei 2 A
i=1

since A is a -algebra. Thus

1
[
Ei 2 AE
i=1

and AE is a -algebra. Let jE = . Then, (?) = jE (?) = 0 = (?) = 0

since ? 2 A. Now, let fEi : i 2 Ng AE be pairwise disjoint. Then,
1
[
Ei 2 AE A
i=1

so that
1
! 1
! 1
!
[ [ [
Ei = jE Ei = Ei =
i=1 i=1 i=1
1
X 1
X 1
X
= (Ei ) = jE (Ei ) = (Ei )
i=1 i=1 i=1

1.3 Construction of Measures

There is a canonical way to build measures. For that, we need a little machinery.

De…nition 13 Let S 2X and : S ! [0; 1] be a function. is said to be

countably monotone if for every E 2 S, and for every fEi : i 2 Ng S, we
have
[1 X1
E Ei =) (E) (Ei )
i=1 i=1

De…nition 14 Let S 2X and : S ! [0; 1] be a function. is said to be

countably subadditive if for every fAi : i 2 Ng S,
1
! 1
[ X
Ai (Ai )
i=1 i=1

11
De…nition 15 Let S 2X and : S ! [0; 1] be a function. is said to be
a pre-measure if

P1 (?) = 0;
P2 is …nitely additive; and
P3 is countably monotone

De…nition 16 An outer measure, : 2X ! [0; 1], is a function de…nable

on every subset of a non-empty set X such that

O1 (?) = 0
O2 is countably monotone.

Lemma 17 is countably subadditive and, therefore, …nitely subadditive.

Proof. Let fAi : i 2 Ng 2X . Then,

1
[ 1
[
A= Ai =) A Ai
i=1 i=1

so that !
1
[ 1
X
(A) = Ai (Ai )
i=1 i=1

Example 18 The trivial measure (A) = 0 8A. The counting measure is

clearly another example.

Here is another interesting example.

Lemma 19 Let S 2X and let : S ! [0; 1] be a function (we do not even

need to be a pre-measure!). De…ne : 2X ! [0; 1] by (?) = 0 and
(1 1
)
X [
(E) = inf (Ei ) : E Ei ^ Ei 2 S8i
i=1 i=1

Then, is an outer measure.

Proof. We already have (?) = 0 so that O1 is trivially satis…ed. Thus,

to prove that is an outer measure, we only need to prove O2 – countable
monotonicity from covering of E by fEi : i 2 Ng:
1
[ 1
X
E Ei =) (E) (Ei )
i=1 i=1

12
If for some i, (Ei ) = 1, we are done. Assume (Ei ) < 1 for all i. Let
> 0. For every i, by properties of in…mum, we can cover Ei by a family
Eik : k 2 N S such that
1
X 1
[ 1
[
(Ei ) Eik =) E Ei Eik
2i i=1
k=1 k;i=1

That is, E is covered by Eik : k 2 N , a countable cover. By de…nition of ,

1
X 1
X
(E) = inf Eik Eik
Ek
i;k=1 i;k=1

Since all the terms of the in…nite sum are positive, we can re-arrange the terms
to get
1
X 1 X
X 1 1
X 1
X
Eik = Eik (Ei ) + = (Ei ) +
i=1 k=1 i=1
2j i=1
i;k=1

Now, we can pass the limit to to get

1
X
(E) (Ei )
i=1

The idea over here is to take a covering and bring it down to a di¤erent
family within S.
Notice that the de…nition of the outer measure does not assume that the
domain is a -algebra. With slight modi…cations, however, a subdomain of
forms a -algebra: let us …rst call a set E X measurable with respect
to if, for any A X, we have (A) = (A \ E) + (A \ E c ). Note
c
that (A) (A \ E) + (A \ E ) is always true because A = A \ X =
A \ (E [ E c ) = (A \ E) [ (A \ E c ) and because is …nitely subadditive.

Proposition 20 Let E X. If (E) = 0, then E is measurable with respect

to

Proof. Let A X. Since (A \ E) (E), we must have (A \ E) = 0.

Again, by monotonicity, (A) (A \ E c ) = (A \ E c ) + (A \ E).
Such a set, which we will call M, of -measurable sets form a -algebra. The
proof of this fact is broken into a bunch of lemmas. First, note that (A) =
(A \ ?)+ (A \ ?c ) so that (?) = 0 so that ? is measurable with respect
to and hence ? 2 M. Also note that if E is measurable with respect to ,
then E c is measurable with respect to , by simple commutativity of the real
numbers under addition.

Lemma 21 If E1 and E2 are measurable with respect to , then E1 [ E2 is

measurable.

13
Proof. If E1 is measurable, then for any A, (A) = (A \ E1 )+ (A \ E1c ).
Also, again, since E2 is measurable, then letting B = A \ E1c , we have, for all B,
(B) = (B \ E2 ) + (B \ E2c ). Thus, (A) = (A \ E1 ) + (A \ E1c )
c c c
= (A \ E1 ) + (A \ E1 \ E2 ) + (A \ E1 \ E2 )
c
= (A \ E1 ) + (A \ E1c \ E2 ) + (A \ (E1 [ E2 ) )
If A, E1 and E2 have a non-trivial intersection, then A \ (E1 [ E2 ) = A \
X \ (E1 [ E2 ) = A \ (E1 [ E1c ) \ (E1 [ E2 )
= (A \ E1 ) [ (A \ E1c ) \ (E1 [ E2 )
= (A \ E1 ) [ (A \ E1c \ E1 ) [ (A \ E1c \ E2 )
= (A \ E1 ) [ (A \ E1c \ E2 ) so that (A \ (E1 [ E2 )) (A \ E1 ) +
(A \ E1c \ E2 ) by …nite subadditivity, with equality if E1 and E2 cover A.
c
Thus, (A \ (E1 [ E2 )) + (A \ (E1 [ E2 ) )
c c
(A \ E1 ) + (A \ E1 \ E2 ) + (A \ (E1 [ E2 ) )
= (A)
c
Thus, (A) = (A \ (E1 [ E2 )) + (A \ (E1 [ E2 ) ).

Lemma 22 If E1 and E2 are measurable with respect to , then E1 nE2 is

measurable with respect to .

Proof. We’ve already proved that for E1 ; E2 2 M, E1 [ E2 2 M and that

c
E1c ; E2c 2 M. Since E1 nE2 = E1 \ E2c = (E1c [ E2 ) , we have that E1 nE2 2 M

Lemma 23 If E1 and E2 are measurable with respect to and disjoint, then,

for every A, (A \ (E1 [ E2 )) = (A \ E1 ) + (A \ E2 )

Proof. Since E2 is measurable, we have (A \ (E1 [ E2 )) = (A \ (E1 [ E2 ) \ E2 )+

(A \ (E1 [ E2 ) \ E2c )
= (A \ ((E1 \ E2 ) [ (E2 \ E2 ))) + (A \ ((E1 \ E2c ) [ (E2 \ E2c )))
Since E1 \ E2 = ?, we have (A \ (? [ E2 )) + (A \ ((E1 \ E2c ) [ ?))
= (A \ E2 ) + (A \ (E1 \ E2c ))
Again, since E1 \ E2 = ?, we must have that E1 E2c . Thus, E1 \ E2c = E1
so that (A \ (E1 [ E2 )) = (A \ E2 ) + (A \ E1 )
By induction, the above holds for n sets.

Lemma 24 If fEi : i 2 Ng are measurable with respect to , then so is their

union.

Proof. Let
k[1 n
[ n
[
ek = Ek n
E Ei and Fn = ei =
E Ei
i=1 i=1 i=1

Fn is measurable because …nite unions are measurable. Now set

1
[
E= Ei
i=1

14
We need to prove that (A) = (A \ E)+ (A \ E c ) i.e., E is -measurable.
We know that Fn E for any n so that E c Fnc . Therefore,

(A) = (A \ Fn ) + (A \ Fnc )
(A \ Fn ) + (A \ E c )
n
!!
[
= A\ ei
E + (A \ E c )
i=1

ei ’s are disjoint, we then have, by Lemma 23,

Since E
n
X
= ei +
A\E (A \ E c )
i=1

Now, we also know that

1
[ 1
[ 1
[
ei =
E Ei = E =) A \ E = ei
A\E
i=1 i=1 i=1

so that, by countable subadditivity, we have

1
X 1
X
(A \ E) ei
A\E =) (A \ E)+ (A \ E c ) ei +
A\E (A \ E c )
i=1 i=1

Thus,
n
!!
[
(A) A\ Ei + (A \ E c )
i=1
n
X
= (A \ Ei ) + (A \ E c )
i=1
(A \ E) + (A \ E c )

The other inequality holds trivially.

Thus, the collection of all measurable sets M under forms a -algebra. It
can now be shown that = jM forms a (complete) measure on M. For this
to be valid, we need to prove that a family of pairwise disjoint sets fEi : i 2 Ng,
1
! 1
[ X
Ei = (Ei )
i=1 i=1

One inequality of this holds always because of countably monotone property.

For the other side, by Lemma 23, for a (…nite) family of pairwise disjoint sets
fEi g, for any A, !
[n Xn
A\ Ei = (A \ Ei )
i=1 i=1

15
Substituting A = X, we have
1
! n
[ X
Ei = (Ei )
i=1 i=1

so that we can replace n with 1, which is what is needed. Thus, is a measure.

To prove that is complete (given a measurable subset E of measure zero
and a subset F of E, the measure of F is zero), note that (E) = 0, so that
(F ) = 0 by countable monotonicity. For a …xed A X, (A \ F ) = 0 and
(A \ F c ) (A) so that (A) (A \ F c ) + (A \ F ). Thus, F is
measurable with respect to
To summarise, we have a function de…ned on S. From this, we constructed
and outer measure over 2X and from this, we constructed de…ned on M.
Do and always agree on S? Not necessarily so, as we are forcing even
de…nition of the empty set.
Think of S as a collection of intervals, with measuring length of the interval.
It would be really nice if did the same and even better when S M (i.e.
jS = ). When does this happen? The answer to this is given in the Lebesgue-
Carathéodory Theorem.

1.3.1 Lebesgue-Carathéodory Theorem

Call S a semi-ring if

1. A; B 2 S =) A \ B 2 S
2. A; B 2 S =) 9 a …nite disjoint family fCi : 1 i ng in S such that
n
[
AnB = Ci
i=1

The last requirement is much weaker than requiring AnB to be in S. Also,

note that empty unions are allowed.

Example 25 Let S be the collection of all intervals of the form [a; b). Half
open intervals are needed to ensure the 2nd condition holds. For example, for
[c; d) [a; b) with c < a and d < b, then [a; b) n [c; d) is another half open
interval. However, if we had only open intervals and an interval B was properly
contained in another A, then AnB would have been half open so that S would
not have been closed under relative complements.

Problem 26 Let X and Y be two non-empty sets and let SX be a semiring on

X and SY be a semiring on Y . Denote Z = X Y and

SZ = fA B : A 2 SX ; B 2 SY g

Prove that SZ is a semiring on Z

16
Solution 27 Let A; B 2 SZ . Then, we can write A = A1 B1 and B = A2 B2 ,
where A1 ; A2 2 SX and B1 ; B2 2 SY . We know that

(A1 B1 ) \ (A2 B2 ) = (A1 \ A2 ) (B1 \ B2 )

Since SX and SY are semi-rings, therefore A1 \ A2 2 SX and B1 \ B2 2 SY .

Let A = A1 \ A2 and B =B1 \ B2 . Then, A \ B = A B, where A 2 SX and
B 2 SY . Thus, A \ B 2 SZ .
Next, let A; B 2 SZ and so, A = A1 B1 and B = A2 B2 , where A1 ; A2 2
SX and B1 ; B2 2 SY =) 9 a …nite disjoint family fCi : 1 i ng SX and
fDi : 1 i mg SY such that
n
[ m
[
A1 nA2 = Ci and B1 nB2 = Di
i=1 i=1

Now,

AnB = (A1 B1 ) n (A2 B2 ) = ((A1 nA2 ) B1 ) [ (A1 (B1 nB2 ))

n
! ! m
!!
[ [
= Ci B1 [ A1 Di
i=1 i=1

Note that A1 = (A1 nA2 ) [ (A1 \ A2 ) and B1 = (B1 nB2 ) [ (B1 \ B2 ) so that
n
! ! n
!!
[ [
AnB = Ci ((B1 nB2 ) [ (B1 \ B2 )) [ ((A1 nA2 ) [ (A1 \ A2 )) Di
i=1 i=1

The …rst term becomes

n
! ! n
! !
[ [
Ci (B1 nB2 ) [ Ci (B1 \ B2 )
i=1 i=1

The second term becomes

m
!! m
!!
[ [
(A1 nA2 ) Di [ (A1 \ A2 ) Di
i=1 i=1

For the …rst term, note that B1 nB2 is disjoint from B1 \ B2 so that the …rst
term, which comprises of sets each from SX and SY , is a disjoint union of two
sets in SZ . Similarly, the other term is a disjoint union of two sets in SZ . All
in all,
n
! m
!! n
! ! m
!!
[ [ [ [
AnB = Ci Di [ Ci (B1 \ B2 ) [ (A1 \ A2 ) Di
i=1 i=1 i=1 i=1

which is a union of disjoint sets, so that the second axiom holds. If we allow
for empty unions, then ? 2 SZ

17
 If S is a semi-ring, we can de…ne RS to be the collection of all …nite disjoint
unions of sets from S. Then, RS is a ring i.e. a collection of sets closed under
relative complements and unions.
Proof. A 2 R =) AnA 2 RS =) ? 2 RS . Let A; B 2 RS . Then,
9 fAi : 1 i ng ; fBi : 1 i mg S such that
n
[ m
[
A= Ai and B = Bi
i=1 i=1

and Ai \ Aj = ? = Bi \ Bj for i 6= j. Then,

n
[ n
[
AnB = A \ B c = (Ai \ B c ) = Ci
i=1 i=1

Note that Ci \ Cj = Ai \ B c \ Aj \ B c = ? so that AnB 2 RS

Generally, the union of two collections of disjoint sets may not be disjoint.
In order to make sure that A [ B is in RS , note that A [ B = (AnB) [ (A \ B) [
(BnA) is a disjoint union. Thus, to show that A [ B 2 RS , we need to show
that A \ B 2 RS for A; B 2 RS :
A \ B = ? [ (A \ B)
= (A \ Ac ) [ (A \ B)
c
= A \ (Ac [ B) = A \ (A \ B c )
= An (AnB), thus A \ B 2 RS

Lemma 28 Let S be a semiring and : S ! [0; 1] be a pre-measure. Then,

extends to RS

Proof. Let : RS ! [0; 1] be the extension of . If A 2 RS , then

n
[
A= Ai
i=1

for disjoint Ai ’s in S. De…ne (A) = (A1 ) + ::: + (An ). To show that this is
well-de…ned, let
n
[ m
[
A= Ai = Bi
i=1 i=1

for disjoint Ai ’s and Bi ’s in S. Note that

n
[ m
[
Ai = Bi
i=1 i=1
n
[ m
[
=) Aj \ Ai = Aj \ Bi
i=1 i=1
m
[
=) Aj = (Aj \ Bi ) .
i=1

18
Then Aj \ Bk for every k is in S. That is, fAj \ Bk : 1 k mg S. This
collection is also pairwise disjoint. Then, (Aj ) = (Aj \ B1 ) + (Aj \ B2 ) +
::: + (Aj \ Bm ) so that
n
X n X
X m
(Aj ) = (Aj \ Bk ) .
j=1 j=1 k=1

The same argument can be repeated with B replaced with A and we can get
the same for
Xm
(Bk ) ;
k=1
except with the summation reversed. Let us do this.
n
[ m
[
Ai = Bi
i=1 i=1
n
[ m
[
=) Bi \ Ai = Bi \ Bi
i=1 i=1
[n
=) Bi = (Bi \ Ak )
i=1

Then, Ak \ Bi for every k is in S. That is, fAk \ Bi : 1 k ng S. This

collection is also pairwise disjoint. Then, (Bi ) = (Bi \ A1 ) + (Bi \ A2 ) +
::: + (Bi \ An ) so that
m
X m X
X n
(Bi ) = (Ak \ Bi )
i=1 i=1 k=1

This and the previous double sum tells us that

m
X n
X
(Bi ) = (Ai )
i=1 i=1

so that is well-de…ned.
P1 Since is a pre-measure, (?) = (?) = 0.
P2 We need to show that is …nitely additive. Let fAi : 1 i ng RS
(i)
be a pairwise disjoint collection. Since Ai 2 RS , 9Bj 2 S such that
mi
[ (i)
Ai = Bj
j=1

with Bi \ Bj = ? for i 6= j. Then,

! 0 1
[n [ mi
n [ mi
n X
X
(i) (i)
Ai = @ Bj A = Bj
i=1 i=1j=1 i=1 j=1

19
 (i)
since each Bj is disjoint. Thus,
n
! n
[ X
Ai = (Ai )
i=1 i=1

Before the next property is established, we note that is monotone: let

A B with A; B 2 RS . Then, 9Ai ; Bj 2 S such that
n
[ m
[
A= Ai and B = Bj
i=1 j=1

Now, we have that

! 0 1
n
X n
[ m
X m
[
(A) = (Ai ) = Ai and (B) = (Bj ) = @ Bj A
i=1 i=1 j=1 j=1

where the …rst equalities follow from de…nition of and the other by …nite addi-
tivity of , a pre-measure. By countable monotonicity (hence …nite monotonicity
of ) and …nite additivity
n
[ m
[ n
X m
X
Ai Bj =) (Ai ) (Bj )
i=1 j=1 i=1 j=1

so that (A) (B).

P3 Let A 2 RS ,
1
[
A Ai
i=1
and Ai 2 RS . We need to show that
1
X
(A) (Aj )
k=1

For this, de…ne

n[1 1
[
en := An n
A Ai 2 RS , A := ei which gives us
A en
A (An )
i=1 i=1

en 2 RS so we can let
Next, note so that A
ni
[
ei = (i)
A Ck
k=1

(i) (i)
where Ck are pairwise disjoint and Ck 2 S for each k. Then,
1 nj
1 [
[ [ (i)
A ei
A Ck
i=1 j=1k=1

20
The latter is still countable. Thus, WLOG, we can say that Aj 2 S is disjoint.
Since A 2 RS , we must have
m
[
A= Bi
i=1

where Bi ’s are from S and are pairwise disjoint. Then

1
[
A Ai
i=1
[m 1
[
=) Bi Ai
i=1 i=1
m
[ 1
[
=) Bk \ Bi Bk \ Ai
i=1 i=1
1
[
=) Bk (Ai \ Bk )
i=1
1
X
=) (Bk ) (Ai \ Bk ) .
k=1

where the last line follows since is a pre-measure. By de…nition of ,

m
X X 1
m X X m
1 X 1
X
(A) = (Bk ) (Ai \ Bk ) = (Ai \ Bk ) = (Ai )
k=1 k=1 i=1 i=1 k=1 i=1

where the last interchange of summations is possible since every summand is

positive and …nite.
With all this machinery, we can …nally answer the question we set before
the beginning of this subsection.

Theorem 29 (Lebesgue-Carathéodory) Let S be a semiring and : S !

[0; 1] be a pre-measure. Then, if we de…ne , outer measure, A, a -algebra
and = jM , we will have (a) S A and (b) and agree on S.

Proof. By Lemma 28, instead of the semi-ring S, we can just as well use
the ring RS . Let A 2 RS . We want A 2 A. That is, (E) = (A \ E) +
(Ac \ E) for every E X. Fix E X. By de…nition of outer measure
(properties of in…mum), for any > 0, we can have
1
[ 1
X
E Ei =) (E) (Ei )
i=1 i=1

21
with Ei 2 RS . Since Ei ; A 2 RS , we can have Ei nA = Ei \ Ac 2 RS and
Ei \ A 2 RS . Then, (Ei ) = (Ei nA) + (Ei \ A) so that
1
X 1
X 1
X
(Ei ) = (Ei \ Ac ) + (Ei \ A)
i=1 i=1 i=1
X1
(Ei \ Ac ) + (E \ A)
i=1
(EnA) + (E \ A)
That is,
1
X
(E) + (Ej ) (EnA) + (E \ A)
i=1
Since this inequality holds for any , we can let it go to zero and we will then
be done (the other inequality holds, as well). Thus, A is -measurable. That
is, A 2 A.
(b) If (A) = 1, then (A) = (A), trivially. Assume otherwise. We
divide the proof into two steps: for any A 2 A, (A) (A) and, conversely,
(A) (A). The …rst is immediate: to recall,
1
X
(E) = inf (Ei )
i=1

where inf is taken over fEi : i 2 Ng S such that this family covers E. Since
A A is a covering of itself and since (A) = (A) for A 2 A, then (A)
(A). For the second inequality, let
1
[
A Ai
i=1

with Ai 2 S for each i and let

i[1
Bi = Ai n An
n=1

Then,
i[1
! j[1
!
Bi \ Bj = Ai n Ak \ Aj n Am
k=1 m=1
i\1 j\1
= Ai \ Ack \ Aj \ Acm = ?
k=1 m=1

since the index Ai \ Acm0 = ? for some 1 m0 < j. Thus,

1
! 1 1
[ X X
Bi (Bi ) (Ai )
i=1 i=1 i=1

22
since Bi Ai for each i. We will be done if we can prove that
1
[ 1
[
Bi = Ai
i=1 i=1

Since Bi Ai for each i, we must have

1
[ 1
[
Bi Ai
i=1 i=1

For the converse, let

1
[
x2 Ai
i=1
De…ne I = fj : x 2 Aj g N. By the well-ordering principle, I has a least
element, say m. Then,
1
[
x 2 Bm =) x 2 Bi .
i=1
so that both sides of the inclusion hold.
For example, consider the collection S = f[a; b) : a; b 2 Rg for subsets of R
and de…ne ([a; b)) = b a. Then, we can have complete measure de…ned
on some -algebra M and m1 ([a; b)) = ([a; b)) = b a and m is the outer
measure, the intermediate step. Such a measure is given a special name, called
the Lebesgue Measure and is covered in the next section.

1.4 Lebesgue Measure

Let X = Rp for some integer p. De…ne a half-open rectangle R = [a1 ; b1 ) :::
[ap ; bp ). Such a collection of R’s forms a semi-ring, as already been discussed
for p = 1. The pre-measure on such a semi-ring may be de…ned as the volume
of each half-open rectangle : S ! [0; 1] by (R) = (b1 a1 ) ::: (bp ap ).
It is easy to see that (?) = 0 and that is …nitely additive. Countable
monotonicity requires some work. Assume p = 1. Our goal will be to extract a
…nite cover.
Let
[1
[a; b) [ai ; bi )
i=1

and let > 0. There exists a j such that b = bj . De…ne ebj = bj + for only
this j and leave the others …xed. Similarly, there exists an i such that a+ > ai .
De…ne e ai = ai . Then, the collection
n o
ak ; ebk : k 2 N
e

covers [a; b] so that we can have a …nite subcover

n o
ak ; ebk : 1 i N
e

23
and so
N
X N
X 1
X
b a ebk ak =
e bk ak + 2 bk ak + 2
k=1 k=1 k=1

Since this is true for every , we can let it tend to zero. Thus, from this pre-
measure on the semiring S, we can construct an outer measure on 2R and use
that to form a measure on a -algebra M.
This argument can be generalised.

Problem 30 Let f : I ! R be continuous for any interval I R. Prove that

if f is increasing and continuous, then ([a; b)) := f (b) f (a) is a pre-measure
on the semiring S of half-open intervals [a; b)

Solution 31 Let a; b 2 R and a b. Then, f (a) f (b) so that ([a; b)) 0.

Hence the function is well-de…ned. Also, since [a; a) = ? is an interval for any
a 2 R, then ([a; a)) = f (a) f (a) = 0. Thus, (?) = 0.
Now, let
n
[ [n
E = [a; b) = Ei = [ i; i)
i=1 i=1

where fEi : i 2 Ng S and Ei \ Ej = ?. If the Ei ’s are already ordered, then,

since the union of two disjoint intervals I1 and I2 is an interval if inf I2 = sup I2
we can have a = 1 1 = 2 2 = 3 ::: = n n = b. If Ei ’s are
n n
not ordered, then for k = 1; :::; n, let ak = min i , bk = min i . It follows that
i=k i=k
we have the chain a = a1 b1 = a2 b2 = a3 ::: = an bn = b and
n
X n
X
(E) = f (b) f (a) = f (bn ) + f (ai ) f (ai ) f (a1 )
i=2 i=2
n
X n
X
= f (bn ) + f (bi ) f (ai ) f (a1 )
i=2 i=2
n
X n
X n
X n
X
= f (bi ) f (ai ) = (f (bi ) f (ai )) = (Ei )
i=1 i=1 i=1 i=1

Thus, is …nitely additive. We prove that is monotone: let [a; b) [c; d) with
c a < b d. Then, by monotonicity of f , f (b) f (a) f (d) f (a)
f (d) f (c). Thus, ([a; b)) ([c; d)).
These two combined tell us that is …nitely monotone. We now prove that
is countably monotone.
Let
[1 1
[
E = [a; b) Ei = [ai ; bi )
i=1 i=1

Let > 0 so that, by continuity of f , for every > 0, we can have we can
have i > 0 and > 0 such that ai < i + ai =) f (ai ) f (ai i ) < 2i and
b < b =) f (b) f (b )< 2

24
 Now,
1
[ 1
[
[a; b ] [a; b) [ai ; bi ) (ai i ; bi )
i=1 i=1

That is,
1
[
[a; b ] (ai i ; bi )
i=1

Since [a; b ] is compact and

1
[
(ai i ; bi )
i=1

is an open cover, we can extract a …nite subcover, say

n
[
[a; b ] (ai i ; bi )
i=1

By properties of already established (monotonicity and, therefore, …nite sub-

additivity), we have
n
X
f (b ) f (a) f (bi ) f (ai i)
i=1

Now,

f (b) f (a) < f (b ) f (a) +

2
n
X
(f (bi ) f (ai i )) +
i=1
2
Xn
< f (bi ) f (ai ) + +
i=1
2i+1 2
Xn n
X
= f (bi ) f (ai ) + +
i=1 i=1
2i+1 2

Letting n ! 1 gives us
1
X
f (b) f (a) f (bi ) f (ai ) +
i=1

We can now let ! 0 to get the required countable monotonicity.

For f (t) = t, this de…nes the Lebesgue measure.

As of now, the pre-measure de…ned in the beginning of this section is
unde…ned on single points. However, we can extend this to the Lebesgue

25
measure mp on the -algebra Mp , in which case mp (fag) = 0 and mp (A) = 0
for any countable A. The measure mp is also translation invariant: mp (a + A) =
mp (A) if A 2 Mp . In particular, m1 (Q) = 0 and, therefore, m1 (Qc \ [0; 1]) = 1.
This example shows that everywhere dense sets have zero measure.

Problem 32 Let X be any non-empty set, S 2X and : S ! [0; 1] be a

function such that there exists a -algebra A and a measure on A such that
S A and for any A 2 S, we have (A) = (A). Is it true that is a
pre-measure?

Solution 33 No. We are not guaranteed that the empty set is in S, so we

cannot even say that (?) = 0. It is true that is …nite additive and countably
monotone by de…nition (since every measure is) but the very …rst requirement
for a pre-measure may fail to hold.

De…nition 34 Let (X; ) be a topological space. We call G a G set if

1
\
G= Oi
i=1

where Oi is open for each i. We call E a F -set if

1
[
E= Fi
i=1

where Fi is closed for each i.

These sets are complements of each other.

Theorem 35 The following are equivalent

1. E 2 Mp
2. 8 > 0, there exists an open O with E O and mp (OnE) <
3. There exists a G set G with E G and mp (GnE) = 0
4. 8 > 0, there exists a closed set F with F E and mp (EnF ) <
5. There exists an F set F with F E and mp (EnF ) = 0

Proof. (1 =) 2)
Let E 2 Mp with mp (E) < 1. Then, by Lebesgue-Carathéodory Theorem,
mp (E) = mp (E) so that
1
X
mp (E) = inf Volume (Ri )
i=1

26
By the properties of in…mum, we can choose a family fRi : i 2 Ng with
1
X
mp (E) > Volume (Ri )
i=1
2

ei so that
Enlarge each Ri to an open R

ei = Volume (Ri ) +
Volume R
2i+1
Then,
1
[
O= ei
R
i=1

is a countable union of open sets and is, therefore, open with

1
X 1
X 1
X
mp (O) ei =
mp R mp (Ri ) + = + mp (Ri ) < mp (E) +
i=1 i=1
2i+1 2 i=1

That is, mp (OnE) = mp (O) mp (E) <

If mp (E) = 1, we can take Rj = 2j ; 2j ::: 2j ; 2j and let Ej =
E \ Rj so that mp (Ej ) < 1. Now take open sets Oj with Ej Oj such that
mp (Oj nEj ) < 2j . With this, we can have an open set
1
[ 1
[ 1
X
O= Oi with E O =) OnE (Oi nEi ) =) mp (OnE) < mp (Oi nEi ) <
i=1 i=1 i=1

(2 =) 3)
For = k1 , we can have E Ok for each k so that
1
\ 1
G= Ok E =) mp (GnE) mp (Ok nE) < =) mp (GnE) = 0
i=1
k

(3 =) 1)
We know that there is a G set G with E G and mp (GnE) = 0 (note the
use of star here!). Thus, GnE is measurable with respect to mp by Proposition
20 and so, GnE 2 Mp . Since G is the union of open sets, clearly G 2 Mp . Since
E G, we have E = Gn (GnE). Thus, E is measurable.
(1 =) 4)
Let E 2 Mp and let > 0. Since Mp is a -algebra, we must have E c 2 Mp .
Then, there is an open set O with E c O and mp (OnE c ) < . Now, E c
c
O =) O E = E. That is, mp (OnE ) = mp (O) mp (E c ). Now, note that
cc c

mp (X) = mp (O) + mp (Oc ) and, similarly, mp (X) = mp (E) + mp (E c ). We,

therefore have mp (X) mp (X) = mp (O) + mp (Oc ) mp (E) mp (E c ) = 0 so
that mp (OnE c ) = mp (O) mp (E c ) = mp (E) mp (Oc ) = mp (EnOc ). Since
Oc is closed, by (2), the proof is done.
(4 =) 5)

27
 Let = n1 . Then, there exists a closed set Fn with Fn E and mp (EnFn ) <
1
n. By de…nition,
1
[
F = Fn
n=1

tells us that F is a F -set. Also, since Fn F so that EnF EnFn and so,
by monotonicity of mp , mp (EnF ) mp (EnFn ) < n1 . Letting n ! 1 gives us
mp (EnF ) 0. That is, mp (EnF ) = 0.
(5 =) 1)
We know that there is a F set F with F E and mp (EnF ) = 0. That is,
EnF is mp -measurable and so EnF 2 Mp . Moreover, F 2 Mp because it is the
intersection of open sets. Since F E, we have E = E \ X = E \ (F [ F c ) =
c c
(F [ E) \ (F [ E) = F [ (E \ F ) = F [ (EnF ). Since EnF; F 2 Mp , it follows
that E = F [ (EnF ) 2 Mp , since any -algebra is closed under unions. Thus,
E is measurable.
This tells us something very important:

Theorem 36 (Regularity of Lebesgue Measure) Assume that E 2 Mp .

(1)
Then, mp (E) = inf fmp (O) : O is open and E Og
(2)
= sup fmp (K) : K is compact and K Eg

Proof. By properties of in…mum and (2) in Theorem 35, The …rst equality
follows directly. For the second equality, by (4) in Theorem 35, we need this
only for closed sets E. Let Rj = 2j ; 2j ::: 2j ; 2j and let Kj = E \ Rj .
Then, Kj is compact. Note that K1 K2 :::. Then,
1
[
E= Ki =) mp (E) = lim mp (Kn ) = sup mp (Kn )
n!1
i=1

which establishes the second equality.

1.4.1 A non-measurable Set

And now, we will construct a non-measurable set E [0; 1] such that E 2 Mp .
Call x; y 2 [0; 1] equivalent if x y 2 Q. Thus, this relation partitions [0; 1].
That is, [
[0; 1] = A

where is a representative of each equivalence class. That is, if x; y 2 A , then

x y 2 Q. In this indexing, the Axiom of Choice is needed (there exists a set
C [0; 1] such that for every , C has exactly one element from A ). This C
is not measurable. Assume that it is.
If q; r 2 Q with q 6= r, then C + q and C + r are disjoint because if z 2
(C + q) \ (C + r), then z = x + q and z = y + r for some x; y 2 C. Then, x y
but that would mean x = y so that q = r, a contradiction.

28
 Thus, C + q and C + r are two distinct measurable sets. That is, C + q 2 Mp
for every q. It is easy to prove that
[
(C + q) [0; 2]
q2Q\[0;1]

which is a countable union of disjoint sets. Thus,

0 1
[
2 = mp ([0; 2]) mp @ (C + q)A
q2Q
X X
= mp (C + q) = mp (C)
q2Q q2Q

which is either 0 or in…nity. The latter is a contradiction so that mp (C) = 0.

However,
[ X
[0; 1] (C + q) =) 1 = mp ([0; 1]) mp (C) = 0
q2Q q2Q

another contradiction.
Thus, C is not measurable.
Had it been mp instead of mp , we would not have been able to use countable
additivity of measure.
In fact, every set of positive measure contains a non-measurable set.
Proof. As above, let E R be a measurable subset of E 2 M1 . Call x; y 2 E
equivalent if x y 2 Q. Let C be the set of representatives for this equivalence
relation. C is not measurable. Assume that it is.
As above, C + q and C + r for q; r 2 Q are two distinct measurable sets.
That is, C + q 2 M1 for every q. Consider C + q mod 1 for every q 2 Q \ [0; 1).
Then, [
(C + q mod 1) = [0; 1]
q2Q\(0;1]

so that
X 1
X
m1 ([0; 1]) = m1 (C + q mod 1) = m1 (C)
q2Q\(0;1] i=1

If C were measurable, then m1 (C) 0. However, the in…nite sum of repeated

non-negative numbers is either 0 or in…nity, neither of which is equal to 1 =
m1 ([0; 1]).
Moreover, the Axiom of Choice is, in fact, equivalent to existence of non-
measurable sets.
p
All in all, this says that Mp 6= 2R

29
1.4.2 Cantor Set
There is a unique creature called the Cantor set. It is a set with zero measure
and is yet uncountable. It is constructed via an iterative process in which at the
n-th step, we delete intervals Cn . The countable intersection of such sets Cn is
the Cantor Set C. Usually, the Cn ’s are formed by deleting the middle third
interval at each n-th step but this is allowed to vary. For example, if we consider
deleting middle thirds, then C1 = [0; 1=3][[2=3; 1] and C2 = [0; 1=9][[2=9; 1=3][
[2=3; 7=9] [ [8=9; 1] and so on. The Cantor set C 6= ? because 0 2 C. Moreover,
C is measurable since we are taking countable intersection of measurable sets.
For each n, it is obvious that m1 (Cn ) = 2n =3n (see Problem 37) where the
denominator is length of one interval in Cn and the numerator follows from the
number of intervals. Moreover, C Cn . As n ! 1, by continuity of Lebesgue
measure, m1 (C) = 0.
Problem 37 C is a perfect.
Recall that a set is said to be perfect if it has no isolated points. Thus, we
have to prove that 8x 2 C, any neighborhood of x contains another point from
C.
Solution 38 Let > 0. For x 2 C, de…ne N (x) = fy : jx yj < g. We need
to show C \ N (x) n fxg = 6 ?.
Since > 0, by the Archimedean property of real numbers, 9N such 1=3N <
. Since C1 = [0; 1=3] [ [2=3; 1]. C2 = [0; 1=9] [ [2=9; 3=9] [ [6=9; 7=9] [ [8=9; 9=9]
and, in general, for the N -th step, we have
1
3N[ 1
3k 3k + 1 3k + 2 3k + 3
CN = ; [ ;
3N 3N 3N 3N
k=0

and C is formed by the countably in…nite intersection of such Ci ’s. It follows

that x 2 33k 3k+1
N ; 3N = Ek (say) or x 2 3k+2
3N
; 3k+3
3N
= Ek0 (say). Moreover
3k + 1 3k 1 3k + 3 3k + 2
m1 (Ek ) = = N = = m1 (Ek0 ) <
3N 3N 3 3N 3N
Thus, if x 2 Ek , then Ek N (x). If x 2 Ek0 , then Ek0 N (x).
Now in both cases, in the N + 1 step, Ek is split into
32 k 32 k + 1 32 k + 2 32 k + 3
; [ ; N +1
3N +1 3N +1 3N +1 3
3j 3j + 1 3j + 2 3j + 3
= ; [ ; = Ek0 [ Ek00 (say)
3N +1 3N +1 3N +1 3N +1
3k+2 3k+3
and 3N
; 3N is split into
32 k + 6 32 k + 7 32 k + 8 32 k + 9
; [ ; N +1
3N +1 3N +1 3N +1 3
3 (j + 2) 3 (j + 2) + 1 3 (j + 2) + 8 3 (j + 2) + 3
= ; [ ; = Ek0 0 [ Ek000 (say),
3N +1 3N +1 3N +1 3N +1

30
for j = 3k. That is, replacing the dummy variable, we can re-write CN +1 as
3N
[1 3k 3k + 1 3k + 2 3k + 3
CN +1 = ; [ ;
3N +1 3N +1 3N +1 3N +1
k=0

Again, note the

1
m1 (Ek0 0 ) = = m1 (Ek000 ) = m1 (Ek0 ) = m1 (Ek00 )
3N +1
If x 2 Ek , then either x 2 Ek0 or x 2 Ek00 . If x 2 Ek0 , then since C Ek00 Ek
N (x), we must have C\ N (x) n fxg = 6 ?. If x 2 Ek00 , then since C Ek0
Ek N (x), we must have C\ N (x) n fxg 6= ?. Similarly, x 2 Ek0 , then
either x 2 Ek0 0 or x 2 Ek000 . If x 2 Ek0 0 , then since C Ek000 Ek N (x), we
must have C\ N (x) n fxg = 00
6 ?. If x 2 Ek0 , then since C Ek0 0 E2 N (x),
we must have C\ N (x) n fxg = 6 ?. In either case, we have that, for any x 2 C
and any > 0, C\ N (x) n fxg = 6 ?.
Theorem 39 C is uncountable
Proof. Assume otherwise. Then, we can let C = fck : k 2 Ng. C1 is the union
of two disjoint intervals. One of these does not contain c1 . Let C \ C1 = F1
be the set which does not contain c1 . F1 is a union of two parts, one of which
does not contain c2 . Call this part (which does not contain c2 ) F2 and so on.
Eventually, we end up with a sequence Fn+1 Fn C where ck 62 Fk . If we
have countably many compact sets and all …nite intersections are non-empty,
then the total intersection is non-empty.
Since we have a decreasing sequence of sets, the intersection
1
\
F = Fn 6= ?
n=1

Moreover, F C. That is, if x 2 F , then x 2 C. Hence x = cN for some N , a

contradiction, since cN 62 FN and F FN C.
Recall that the Borel -algebra L is the minimal -algebra that contains all
open subsets of R, under the usual topology. Since L is the smallest -algebra
on R, we can claim that L M1 . However, the converse does not hold. By
completeness of M1 , any subset of the Cantor set is in M1 . It can shown that
there is such a subset not in L. Moreover, the completion of L gives M1 but
the proof of this fact will have to wait for now, until we get into a discussion of
functions acting on Measure Spaces.
Problem 40 C is totally disconnected set.
Solution 41 We need to show that for every x 2 C, any neighborhood of x
contains a point from the complement of C. That is, we need to show that
8x 2 C and any > 0, C c \ N (x) 6= ?. To begin the proof, we observe that
n 1 n 1
1 3 [
[ 1 1 3 [
[ 1
c 3k + 1 3k + 2
C = ; = Ik
n=1 k=0
3n 3n n=1 k=0

31
in [0; 1]. Again, since > 0, by the Archimedean property of real numbers, 9N
such that 1=3N +1 < . Note that, at the N -th step, m1 (Ik ) = 3 N < 3m1 (Ik ) =
3 N +1 < . Moreover, x 2 Ek or x 2 Ek0 . In either case, Ik \ N (x) 6= ? since
m1 (N (x)) = 2 > > m1 (Ik ) and Ik is in between Ek and Ek0 . That is, C c \
N (x) 6= ?.
Problem 42 Suppose that in the construction of a Cantor set, on the n-th
step, instead of throwing away the middle third of every interval, we throw away
middle intervals of length an > 0 (an depends only on n). Show that, by choosing
the right sequence fan : n 2 Ng, we can make the measure of the resulting set
to be any number between 0 and 1. Deduce that there exists an open subset of
[0; 1] whose boundary has positive measure.
Solution 43 Let x 2 [0; 1]. We want to construct a Cantor set of Lebesgue
measure x. Since x 2 [0; 1], 9 a sequence fbn : n 2 Ng such that
1
X
x= bn
n=1

Let
1
X
1 x= an
n=1

where an is some sequence dependent on the sequence fbn : n 2 Ng, clearly con-
vergent. Now, begin with [0; 1] and throw away
1 a1 1 + a1
;
2 2
which has Lebesgue measure a1 . The resulting C1 is
1 a1 1 + a1
C1 = 0; [ ;1 :
2 2
The middle point of
1 a1
0;
2
is
1 a1
4
so for the left interval, we throw away
1 a1 a2 1 a1 + a2
; :
4 4
This set has Lebesgue measure a2 =2. Similarly, the middle point for the right
interval is 3+a
4 , we throw away
1

3 + a1 a2 3 + a1 a2
; + :
4 4 4 4

32
This set has Lebesgue measure a2 =2, so in total, we have thrown away intervals
of length a2 . We are left with
1
a1 a2 1 a1 + a2 2 2a1
C2 = 0; [ ;
4 4 4
2 + 2a1 3 + a1 a2 3 + a1 + a2
[ ; [ ;1 :
4 4 4

For the …rst set in C2 , we can throw away 1 a1 8a2 a3 ; 1 a1 8a2 +a3 . For the
second, throw away a2 +3 83a1 a3 ; a2 +3 83a1 +a3 ; for the third, 5+3a1 8 a2 a3 ; 5+3a1 8 a2 +a3 ;
and for the fourth, 3+a1 +a82 +4 a3 ; 3+a1 +a82 +4+a3 . This gives us

1 a1 a2 a3 1 a1 a2 + a3 2 2a1 2a2
C3 = 0; [ ; [
8 8 8
1 a1 + a2 3 3a1 + a2 a3 3 3a1 + a2 + a3 1 a1
; [ ; [
4 8 8 2
2 + 2a1 5 + 3a1 a2 a3 5 + 3a1 a2 + a3 3 + a1 a2
; [ ; [
4 8 8 4
3 + a1 + a2 3 + a1 + a2 + 4 a3 3 + a1 + a2 + 4 + a3
; [ ;1
4 8 8
Continuing this way, we observe that we can throw away intervals of total length
an at the n-th step. Since [0; 1] is the disjoint union of all these open sets and
the intersections of Cn , we get that
1
X
1 = m1 (C) + an
n=1

or that m1 (C) = x.
In this construction, we note that C is countable intersection of …nite union
of closed sets Cn . Thus, C is closed so that C c is open. Since [0; 1] = C [ C c
and @C c = Cl (C c ) Int (C c ) = Cl (C c ) C c = Cl (C c ) \ C = (Cn f0; 1g) \
C = Cn f0; 1g. Thus, m1 (@C c ) = m1 (C) m1 (f0; 1g) = m1 (C). The crucial
point here is realizing that Cl (C c ) = Cn f0; 1g and this is because for each n,
(n)
@Cn n f0; 1g = Cl (In ) where In is the disjoint union of the open sets Ik for
which m1 (In ) = an .

2 Functions on Measure Spaces

2.1 Measurable Functions
Let (X; AX ; X ) ; (Y; AY ; Y ) be a measure space and let f : X ! Y be a
function. f is said to be a measurable function (with respect to A) if the
pre-image of any Y -measurable set is X -measurable. Therefore, for Y = R
and AY = M1 , f is a measurable function if f 1 ([a; b)) 2 AX for all a; b 2 R.

33
This will be shortened to saying that f is X -measurable or even that f is
measurable, if the underlying measure space is clear. In probability theory,
such an f is called a random variable.
We will only be concerned with real-valued functions.

Theorem 44 Let (X; A; ) be a measure space and let f : X ! R be a mea-

surable function. The following are equivalent
1
1. f ([a; b)) 2 A
1
2. f ((a; 1)) 2 A
1
3. f ((a; b)) 2 A
1
4. f ([a; b]) 2 A
1
5. f (( 1; a)) 2 A
1
6. f (( 1; a]) 2 A

Proof. (1 =) 2)
To see this, note that
1
[ \ 1
1 [ 1 [
\ 1
1 1 1 1
(a; 1) = (a; n) = a ;n =) f ((a; 1)) = f a ;n
n=1 m=1n=1
m m=1n=1
m

The set on the right hand-side is a countable intersection of a countable union

of measurable sets, hence measurable. Therefore, f 1 ((a; 1)) is measurable.
That is, f 1 ((a; 1)) 2 A. The rest of the statements are veri…ed in a similar
manner. We only mention their decomposition.
(2 =) 3)
(a; b) = (a; 1) \ (b; 1)
(3 =) 4)
1
\
[a; b] = a + n1 ; b + n1
n=1
(4 =) 5)
1
[ 1 [
\ 1
( 1; a) = ( n; a) = [1=m n; a 1=m]
n=1 m=1n=1
(5 =) 6)
1
\
1
( 1; a] = 1; a + n
n=1
(6 =) 1)
We can use (6 =) 5) because
1
[ 1
( 1; a) = 1; a
n=1
n

34
Taking complement on both sides gives us
1
\ 1
[a; 1) = a ;1 :
n=1
n

Thus, assuming 6, we can have ( 1; b) for some b > a and ( 1; b) \ [a; 1) =

[a; b) 2 A.

Example 45 Let (X; A; ) be a measure space and let f : X ! R be de…ned

by
1 if x 2 E
f (x) = E (x) =
0 if x 62 E
In this case, f is measurable function if and only if E 2 A: if f is measurable,
then f 1 (0; 1) = E 2 A. Conversely, if E is measurable, then for any (a; b),
if both 0 and 1 belong to (a; b), then f 1 (a; b) = X. If either belongs to (a; b),
then f 1 (a; b) = E c or E, respectively. In either case, f 1 (a; b) 2 A.

Problem 46 Let (X; A; ) be a measure space, E 2 A and (E) > 0. De…ne

AE := fA E : A 2 Ag. Recall from Problem 11 that AE is a -algebra and
that jE is a measure on AE .

1. Prove that if f is measurable on X, then the restriction of f to E is

measurable on E
2. Assume (XnE) = 0 and f : X ! R is a function such that its restric-
tion to E is measurable with respect to AE . Is it true that f is measurable
on X? If no, what extra condition do we need for this to be true?

Solution 47 1. Let f jE = g and I R be an interval. Then, g 1 (I) =

f (I)\E. Since f (I) ; E 2 A (so that f 1 (I)\E 2 A) and f 1 (I)\E E,
1 1

hence f 1 (I) \ E 2 AE . That is, g 1 (I) 2 AE . Thus, g is measurable.

2. No. It may happen that f 1 (A) XnE but f 1 (A) 62 A. For example,
consider X = f1; 2; 3g with A = f?; f1; 2g ; f3g ; f1; 2; 3gg and E = f3g. De…ne
(f1; 2g) = 0 and (f3g) = 1. Now de…ne f (x) = x. Then, f restricted
to E is a measurable function. However, f 1 (1=2; 3=2) = f1g 62 A. For the
extension of a measurable function to be measurable, we must have A complete.
That is, when every subset of a null set (a set of measure zero) is measurable.
If m1 (XnE) = 0, then every subset of XnE is of measure zero and not included
c
in AE . Now, extend f by letting f (x) 2 f (E) if x 2 XnE. Then, either
f (a; b) E or f (a; b) XnE. In the former case, f 1 (a; b) 2 AE
1 1
A
so that f is measurable. In the latter, m1 f 1 (a; b) = 0 so that f 1 (a; b) 2 A,
making f again measurable.

When is the composition of the two functions measurable? Continuous func-

tions are always measurable so that gives us a lot but the following lemma shows
we can generalise this a bit.

35
Lemma 48 Let X be a topological space, F : X ! Rp and g : F (X) ! R
be functions. If g is continuous and components of F are measurable, then
f = g F is measurable.

Proof. Let
1 1
f ((a; 1)) = fx : f (x) 2 (a; 1)g = x : F (x) 2 g (a; 1)

Then, g 1 (a; 1) is relatively open in F (X). Thus, there exists an open set
O Rp such that O \ F (X) = g 1 (a; 1). Since O is open, we can cover O by
disjoint, open (even half-open work) rectangles fRi : i 2 Ng in Rp . Thus,
1
f (a; 1) = fF (x) 2 F (X) \ Og
( 1
)
[
= F (x) 2 F (X) \ Ri
i=1
( 1
!)
[
1
= x2F Ri
i=1
1
[
1
= x2F (Ri )
i=1

Let us look at the individual pre-image of a rectangle, i.e. at x 2 F 1 (Ri ) .

Since Ri = I1i ::: Ipi and F = '1 ; :::; 'p , with each component measurable
by hypothesis, then x 2 F 1 (Ri ) () x 2 'k 1 (Ik ) for each k so that
p
\ p
\
x2 'k 1 (Ik ) =) f 1
((a; 1)) = 'k 1 (Ik )
k=1 k=1

1
Thus, f ((a; 1)) 2 A.

Corollary 49 If f; g are measurable functions, then f +g; f; f:g; jf j ; max (f; g)

and min (f; g) are measurable.

Proof. Let f and g be two measurable functions de…ned on the same domain,
X be a topological space, F : X ! R2 de…ned by F (x) = (f (x) ; g (x)),
A : F (X) F (X) ! R de…ned by A (x; y) = x + y, P : F (X) F (X) ! R
de…ned by P (x; y) = xy, G : X ! R de…ned by G (x) = f (x), S : F (X) !
R de…ned by S (x) = x, M : F (X) ! R de…ned by M (x) = jxj, Mx :
F (X) F (X) ! R2 de…ned by Mx (x; y) = max (f (x) ; g (y)) and Mn :
F (X) F (X) ! R2 de…ned by Mn (x; y) = min (f (x) ; g (y)). Then, F and
G have measurable functions as components and A, P , S, M , Mx and Mn are
continuous functions. By above lemma, A F = f + g, P F = f:g, S G = f ,
M G = jf j, Mn F = min (f; g) and Mx F = max (f; g) are all measurable.

36
Theorem 50 Let ffn : n 2 Ng be a sequence of measurable functions. Then,
the following are measurable:

inf fn (x) ; supfn (x) ; lim inf fn (x) and lim sup fn (x)
n2N n2N N !1n N N !1n N

are all measurable.

Proof. First, let us call

f (x) = supfn (x)
n2N

Consider,

1
f (( 1; a)) = fx : f (x) < ag = x : supfn (x) < a
n

Thus, the set becomes

1
\
fx : fi (x) < a; 8i 2 Ng = fx : fi (x) < ag
i=1

1
Now, fx : fi (x) < ag = fi (( 1; a)) is measurable, so we must have
1
\
fx : fi (x) < ag 2 A
i=1

Thus, f 1 (( 1; a)) is a measurable set and so f is a measurable function.

Similarly, let
g (x) = inf fn (x)
n2N

Consider the set

1
g ((a; 1)) = fx : g (x) > ag = x : inf fn (x) > a
n2N
1
\
= fx : fi (x) > a; 8i 2 Ng = fx : fi (x) > ag
i=1

1
Since fx : fi (x)g = fi ((a; 1)) is measurable, we must have
1
\
fx : fi (x) > ag 2 A
i=1

1
Thus, g ((a; 1)) is a measurable set so that

g = inf fn
n2N

is a measurable function. Since

inf fn (x)
n N

37
is an increasing sequence, we can have
lim inf fn (x) = sup inf fn (x)
N !1n N N n N

Since fi (x) is measurable for each i, then

gN (x) = inf fn (x)
n N

is measurable so that
sup gN (x)
N 2N
is measurable. Now, recall that
lim sup fn (x) = inf sup fn (x)
N !1n N N 0n N

Then, fn (x) is measurable

=) sup fn (x) is measurable
n N
=) inf sup fn (x) is measurable
N 0n N

Corollary 51 If ffn : n 2 Ng is a sequence of measurable functions, and fn !

f pointwise (i.e. fn (x) ! f (x) for every x 2 X), then f (x) is measurable.
Proof. Since fn (x) ! f (x), we have
f (x) = lim fn (x) = lim sup fn (x) = lim inf fn (x)
n!1 N !1n N N !1n N

The latter two are measurable.

What if fn (x) ! f (x) almost everywhere? Under certain conditions, f (x)
is measurable:
Problem 52 Let ffn : n 2 Ng be a sequence of real-valued measurable functions
on X. For every natural n, de…ne
n
En := x 2 X : jfn (x) fn+1 (x)j > 2
n
Show that if (En ) < 2 for every n, then fn is pointwise convergent almost
everywhere on X.
Solution 53 Since (En ) < 2 n for every n, we have a family of measurable
sets fEn : n 2 Ng and their union has …nite measure,
1
X X1
1
(En ) < n
=1<1
n=1 n=1
2

Then, by Borel-Cantelli Lemma, almost every x 2 X belongs to at most …nitely

many En ’s. In other words, almost no x 2 X belongs to in…nitely many En ’s.
By de…nition of En , fn does not converge pointwise on En since for = 2 n ,
there exists no N such that jfn (x) fn+1 (x)j < 2 n for N n. Thus, for
almost every x 2 X, fn is not pointwise convergent on at most …nitely many
En ’s, which means that fn is pointwise convergent almost everywhere.

38
2.2 Simple Functions
Let us now approximate! For measurable functions, there are many canonical
ways that work but the one with a simple function is what we will study.

De…nition 54 A function f : X ! R is called simple if there exist measur-

able sets E1 ; :::; En such that
n
X
f (x) = cj : EJ (x)
j=1

where cj are scalars.

These are also called step functions. The measurable sets are not required
to be disjoint. These are called simple because they’re easy to integrate.

Problem 55 Show that a sum, a product, a min and a max of two simple
functions is a simple function.

Solution 56 Let f : X ! R and g : X ! R be simple. Then, 9 measurable

sets E1 ; :::; En and F1 ; :::; Fm such that
n
X m
X
f= cj : Ej and g = di : Fi
j=1 i=1

where Ej = fx 2 X : f (x) = cj g and Fi = fx 2 X : g (x) = di g. We can as-

sume that ci 6= cj for i 6= j so that Ei 6= Ej . Similarly for g.
Since the intersection of measurable sets is measurable, we have Ej \ Fi is
measurable for 1 j n and 1 k m.
Recall that the domain for the sum and product of two functions is the inter-
section of the domains. Also recall that Ei \Fj = Ei + Fj Ei Fj . Then, if
we let ck + dk = ak and ck = 0 and Ek = ? for k > n = min fm; ng or dk = 0
and Fk = ? for k > m = min fm; ng, then
maxfm;ng maxfm;ng maxfm;ng maxfm;ng
X X X X
f +g = ai Ei \Fi = ai Ei + ai Fi ai Ei Fi <1
i=1 i=1 i=1 i=1

Moreover
maxfm;ng
X
max (f; g) = max (ci ; di ) Ei \Fj
i=1

and,
maxfm;ng
X
min (f; g) = min (ci ; di ) Ei \Fj
i=1

are both …nite so that the sum, product, min and maximum of simple functions
is simple. Similarly, for f:g where (f:g) (x) = f (x) g (x), if we let ck dk = bk

39
and ck = 1 and Ek = ? for k > n = min fm; ng or dk = 1 and Fk = ? for
k > m = min fm; ng, then
maxfm;ng
X
f:g = bi Ei \Fj <1
i=1

2.2.1 Approximation of Measurable Functions

Lemma 57 Assume that f : X ! R is measurable and bounded (i.e. jf (x)j
M for every x 2 X). Then, 8 > 0, 9 simple '; with ' f such that
j' (x) (x)j < 8x

Proof. Let > 0 and 8x, we know that f (x) 2 [ M; M + 1). Denote c = M
and d = M + 1. Take y0 = c < y1 < ::: < yn = d such that yk yk 1 <
for each k. Then, each Xk = f 1 [yk 1 ; yk ) is measurable. Note that Xk ’s are
disjoint. De…ne
Xn
' (x) = yk 1 : Xk (x)
k=1

and
n
X
(x) = yk : Xk (x)
k=1

Since f (x) 2 [c; d), 9!k such that yk 1 f (x) yk . That is, if x 2 Xk ,
then ' (x) = yk 1 f (x) yk = (x). Thus, ' f . Moreover, since
jyk 1 yk j < , we must have j' (x) (x)j < 8x.

Theorem 58 Let f : X ! R be measurable. Then, there is a sequence f ng

of simple functions such that

1. j n (x)j jf (x)j
2. n (x) ! f (x) for every x 2 X
3. If f 0, then n (x) % f (x). That is, for a …xed x, the sequence n is
increasing and converges to f (x).

Proof. Let En = fx 2 X : jf (x)j ng. Then, En is measurable. By Lemma

57, 9gn ; hn in En with gn f hn and jgn hn j < n1 . For x 2 En , de…ne
8
< 0 f (x) = 0
n (x) = max (h n (x) ; 0) f (x) > 0
:
min (gn (x) ; 0) f (x) < 0

and for x 62 En , set n (x) = 0. This is a simple function since it has …nitely
many values and is absorbed in the characteristic function. The values of n (x)
are obtained by measurable functions; thus n is measurable, for each n.

40
 Now we prove (1). Fix an x 2 X. If x 2 En and f (x) > 0, then hn (x) > 0
so that hn f =) j n (x)j jf (x)j. If x 2 En and f (x) < 0, then gn (x) < 0
so that f gn < 0 =) j n (x)j jf (x)j. In all other cases, n (x) = 0 so
j n (x)j jf (x)j.
For (2), there are three cases to consider.
If 0 < f (x) < 1 for every x, then 9N such that jf (x)j N so that for
n N and x 2 En , 0 n (x) = max (h n (x) ; 0) jf (x)j so that from
0 hn (x) = n (x) f (x) gn (x) and 0 gn (x) hn (x) < n1 , we get 0
f (x) n (x) g (x) hn (x) < n1 . Letting n ! 1, we get f (x) n (x) ! 0
so that n (x) ! f (x).
If 0 > f (x) > 1 for every x, then 9N such that jf (x)j N so that
for n N and x 2 En , 0 n (x) = min (g n (x) ; 0) jf (x)j so that 0
1
f (x) n (x) f (x) g n (x) < n . Thus, f (x) n (x) ! 0.
If f (x) = 1, then de…ne
8
< n (x) x 2 En
e (x) = n f (x) > n
n
:
n f (x) < n

Then, by de…nition,
e (x) jf (x)j
n

and e n ! f .
For (3), we need special consideration since the sequence n is not necessarily
increasing. This is because the approximations hn and gn , as constructed in
Lemma 57, rely on the nature of f itself. Thus, hn ’s and gn ’s are not necessarily
increasing since f is not. We get our way around: if f 0, then de…ne

n (x) = max e j (x)

1 j n

Clearly, n is an increasing sequence of functions. Moreover, since

e (x) jf (x)j
n

for each n, we must have

n (x) jf (x)j

Finally, it is easy to see that 0 f (x) n (x) f (x) e n (x) ! 0.

Since hn and gn are simple, then so is e
n , and so is j , and, therefore, n
is also a simple function for each n.

Lemma 59 Let f : Rp ! R be a (measurable) function. Let E 2 Mp ,

mp (E) < 1. Assume that ffn : n 2 Ng is a sequence of measurable functions
with fn (x) ! f (x) 8x 2 E (i.e. pointwise). Then, 8 > 0, > 0, there exists
A E with A 2 Mp such that jfn f j < on A 8n N and mp (EnA) <

41
 A is not necessarily closed. Hence it di¤ers from what we have established in
Theorem 35. That is, the existence of a closed set A such that mp (EnA) < .
Although this doesn’t necessarily matter, as will be clear in the proof.
Proof. Note that fn is measurable and

f = lim fn = lim sup fn = lim inf fn

n!1 N !1n N n!1n N

so that f is also measurable, hence the brackets in the hypothesis. De…ne

En = fx 2 Rp : jf (x) fk (x)j < 8k ng

This set is measurable because both fn and f are measurable. That is, En 2 Mp .
Note that En En+1 for each n 1. De…ne
1
[
Ej = E
j=1

so that
mp (E) = lim mp (En )
n!1

Hence 9N such that 8n > N , mp (E) < mp (En ) + for n N and, so,
mp (EnEn ) < for n N . Setting A = EN +1 establishes the lemma.

Theorem 60 (Egoro¤ ) Let f : Rp ! R be a measurable function. Let

E 2 Mp , mp (E) < 1. Assume that ffn : n 2 Ng is a sequence of measur-
able functions such that 8x 2 E, fn (x) ! f (x) pointwise. Then, 9 a closed
F E with mp (EnF ) < and fn ! f uniformly on F .

This theorem holds in more generality for measure spaces. Thus, up to a

measure , pointwise convergence implies uniform convergence.
Proof. By Lemma 59, for a …xed n, we can …nd An E and N (n) such that
"
mp (EnAn ) < 2n+1 for n N so that 8k N (n), jf fk j < n1 on An . Now let
1
\
A= Aj
j=1

A may not be closed! Then,

1
!
[
mp (EnA) = mp (EnAn )
n=1
1
X
= mp (EnAn )
n=1
X1
" "
< n+1
=
n=1
2 2

42
Our goal now is to show that fn ! f on A uniformly. This is written as
fn f on A. That is, 8 > 0, 9M such that 8k > M , jfk (x) f (x)j <
for every x 2 A. Take n such that n1 < . Then, for every x 2 An , we have
jfk (x) f (x)j < n1 < as soon as k N (n) = N ( ). Let M = N (n) =
N ( ). Then, for a …xed > 0, we found an M such that 8k M , 8x 2 A, we
have jfk (x) f (x)j < , i.e. fn f.
Now, recall that mp (EnA) < 2" . By the Regularity of the Lebesgue Measure
(i.e. mp (EnA) = sup fmp (K) : K is compact and K EnAg) we can …nd a
closed F A with mp (AnF ) < 2" . Then, mp (EnF ) < ".
Problem 61 Show that the conclusion of Egoro¤ ’s Theorem can fail if we drop
the assumption that the domain has …nite measure.
Solution 62 Consider the sequence of indicator functions [ n;n] := fn de…ned
on R. Then, for x 2 R,
lim sup fn = lim inf fn = 1
N !1n N n!1n N

so that fn converges to the constant function 1. That is, to a function 1 :R !R

such that 1 (x) = 1 for all x, pointwise. However, assume that fn 1 on some
closed subset F of R. Choose = 1, then for k 2 N, j1 (x) fk (x)j < 1 holds
when fk (x) = 1 so that, for the N we should be able to …nd, given our = 1,
for n N ,
1
\
En = fx 2 R : j1 fk (x)j < 1 8k ng = [ k; k] = [ N; N ] = F
k=N

However, m1 (RnF ) = m1 (( 1; N ) [ (N; 1)) 6< 1.

Lemma 63 Let E Rp and f : E ! R be a simple function. Then, 8 >
0, 9 a continuous function g : Rp ! R and a closed set F E such that
mp (EnF ) < , f (x) = g (x) if x 2 F
Proof. Let
n
X
f (x) = ak Ek (x)
k=1
where we can assume WLOG that Ek are disjoint. Then, we can …nd a closed
set Fk Ek such that mp (Ek nFk ) < 2k . Set
n
[
F = Fk
k=1

Note the …nite union: this ensures that F is closed. Moreover, mp (EnF ) < .
Set g (x) = ak when x 2 Fk . g is de…ned on F and it is continuous on F : if
x 2 F , then 9!k 0 such that x 2 Fk0 . That is, some neighborhood of x does not
intersect Fj for j 6= k 0 . Now, the continuity of g is easy to prove. We can now
extend g to a continuous function on Rp by setting constant values for the “end
points” of F .

43
Problem 64 De…ne
8
>
> 2 if x 2 [0; 1)
<
5 if x 2 [1; 2)
f (x) :=
>
> 1 if x 2 [3; 4)
:
0 otherwise
For every > 0, construct a continuous function ' : R ! R such that
m1 (fx : f (x) 6= ' (x)g) < .
Solution 65 Let E1 = [0; 1), E2 = [1; 2) and E3 = [3; 4) and E4 = [4; 1).
Then, f = 2 E1 + 5 E2 E3 + 0 E4 is simple and, therefore, measurable.
Thus, f : R ! R is a simple function, so we are guaranteed the existence of
such a continuous function by Lemma 63. We can choose the closed sets Fk
to be
h i h i h i h
F1 = ; 1 ; F2 = 1; 2 ; F 3 = 3 + ; 4 and F 4 = 4 + 4;1
4 4 22 24 24 2
Then, for F = F1 [ F2 [ F3 [ F4 and E = E1 [ E2 [ E3 [ E4 ,
15
m1 (EnF ) = + + + = <
2 22 23 24 16
Now, de…ne ' : R ! R
8 8
>
> 2 x if x 2 0; 4
>
>
>
> 2 if x 2 F1
>
> 12
>
> x+2 if x 2 1 4 ; 1
>
>
< 5 if x 2 F2
40 20
' (x) = x if x 2 2 4 ; 2
>
> 48 16
>
> x if x 2 3; 3 + 16
>
>
>
> 1 if x 2 F3
>
> 8
>
> x 21 32 if x 2 4 16 ; 4 + 16
:
0 otherwise
By construction, ' is continuous and EnF = fx : f (x) 6= ' (x)g
Theorem 66 (Lusin) Let E Rp and f : E ! R be a measurable function.
Then, 8 > 0, 9 a continuous function g : Rp ! R and a closed set F E
such that mp (EnF ) < , f (x) = g (x) 8x 2 F
That is, a measurable function can be approximated by a continuous func-
tion.
Proof. There are two cases to consider, with E having …nite and in…nite mea-
sure.
Assume that mp (E) < 1.
Since f is measurable, 9 a sequence of simple functions f'n : n 2 Ng on E
such that 'n ! f . As shown in Problem 64 and proved in Lemma 63, for
each n, we can …nd a continuous function gn such that Fn E is closed and

mp (EnFn ) <
2n+1

44
Note that gn = 'n on Fn . By Egoro¤, there is a closed F0 E where 'n f
and
mp (EnF0 ) <
2
Now, let
1
\
F = Fn
n=0
x2F
where Fn is closed and mp (EnF ) < . On F , we have gn (x) = n 'n (x)
f (x). That is, on F , fgn : n 2 Ng converges uniformly to f . Then, f is contin-
uous on F . We can then set g = f jF . Then, g is continuous on F . We can now
extend g to Rp .
If mp (E) = 1, we can split E into boxes of side-length 1. Let
1
[
E= Bi
i=0

and let Ei = E \ Bi . Then,

1
[
E= Ei
i=0
and Ei for each i is measurable, m1 (Ei ) 1 and by Problem 46, f is measur-
able on Ei and, therefore, 9 a sequence of simple functions f'n : n 2 Ng on Ei
such that 'n ! f . Again, by Lemma 63, for each n, we can …nd a continuous
(i)
function gn such that Fn Ei is closed and

mp Ei nFn(i) <
2i+n+1
(i) (i)
with gn = 'n on Fn . By Egoro¤, there is a closed F0 Ei where 'n f
and
(i)
mp Ei nF0 < i+1
2
Now, let
\1
(i)
F = Fn(i)
n=0

(i) x2F (i)

where Fn is closed and mp Ei nF (i) < 2i+1 . On F (i) , we have gn (x) =n
'n (x) f (x). That is, on F (i) , fgn : n 2 Ng converges uniformly to f . Then,
f is continuous on F (i) . We can then set gi = f jF (i) . Then, gi is continuous on
F (i) . Now let
1
X 1
[ 1
[
h (x) = gi (x) F (i) (x) and F 0 = F (i) Ei
i=1 i=0 i=0

Then, g = f jF . This union, however, may not be closed but is, however,
measurable. Thus, we can …nd a closed set F F 0 and by Egoro¤, a continuous
function g such that m1 (F 0 nF ) < =2 so that m1 (EnF 0 ) < and g (x) = h (x)
on F .

45
3 Integration
3.1 Integral of Simple and Measurable Functions
The integral is naturally de…ned using simple functions over a set of …nite mea-
sure. To do this, however, we need to ensure that the integral does not depend
on the representation of the simple function:

Lemma 67 Let (X; A; ) be a measure space and let ' : X ! R be a simple

function. Assume that
N
X M
X
' (x) = aj Aj (x) = bk Bk (x)
j=1 k=1

where fAj : 1 j N g and fBj : 1 j M g are both pair-wise disjoint and

measurable, and that
[N M
[
Aj = X = Bj
j=0 j=0

Then, for every E 2 A,

N
X M
X
aj (Aj \ E) = bk (Bk \ E)
j=1 k=1

Proof. Note that aj (Aj \ E \ Bk ) = bk (Bk \ E \ Aj ) for each j; k. This is

because if Aj and Bk are disjoint, then we trivially have equality on both sides
since we then have 0 = 0. If x 2 Aj \ Bk , then let ' (x) = cj . Since x 2 Aj , we
have ' (x) = aj and ' (x) = bk because x 2 Bk . Thus, cj = bk = aj . Now, 8k,
0 1
N
[ N
[
Bk = Bk \ X = Bk \ @ Aj A = (Aj \ Bk )
j=1 j=1

46
and
0 1 0 1
M
X M
X N
[ M
X N
[
bk (Bk \ E) = bk @E \ (Aj \ Bk )A = bk @ (E \ Aj \ Bk )A
k=1 k=1 j=1 k=1 j=1
M
X N
X M X
X N
= bk (E \ Aj \ Bk ) = bk (E \ Aj \ Bk )
k=1 j=1 k=1 j=1

X N
M X N
X M
X
= aj (E \ Aj \ Bk ) = aj (E \ Aj \ Bk )
k=1 j=1 j=1 k=1
N M
! N M
!
X [ X [
= aj (E \ Aj \ Bk ) = aj E\ (Aj \ Bk )
j=1 k=1 j=1 k=1
N M
!! N
X [ X
= aj E\ Aj \ Bk = aj (E \ (Aj \ X))
j=1 k=1 j=1
N
X
= aj (E \ Aj )
j=1

Now, assume that we have a simple, positive function ', on X with

N
X
' (x) = ai Ai (x)
i=1

where fAi : 1 i N g is pair-wise disjoint and measurable. Then for every

E 2 A, we de…ne the integral of ' over E with respect to as follows:
Z N
X
'd = aj (Ai \ E)
E i=1

Lemma 68 Let ' be a simple, non-negative function with

N
X
' (x) = ai Ai (x)
i=1

Let be a measure and let E 2 A. Then,

Z
'd 2 [0; 1]
E

Proof. Since
Z N
X
'd = aj (Ai \ E)
E i=1

and each aj > 0 by hypothesis ( 0 by default)

47
Lemma 69 Let ' be a constant simple, non-negative function. That is, ' (x) =
c 8x 2 E where E 2 A. Let be a measure. Then,
Z
'd = c (E)
E

Proof. If ' (x) = c, then ' (x) = c E so that

Z
'd = c (E \ E) = c (E)
E

Lemma 70 Let ' and be simple, non-negative functions with

N
X M
X
' (x) = ai Ai (x) and (x) = bi Bi (x)
i=1 i=1

Let be a measure, ; be scalars and E 2 A. Then,

Z Z Z
( '+ )d = 'd + d
E E E

Proof. The result will be proven in two steps: …rst, we will show that the sum
gets distributed and later on, that the scalar product can be pulled out.
We may choose a …nite, disjoint collection Ci such that
K
X K
X
' (x) = i Ci (x) and (x) = i Ci (x)
i=1 i=1

Then,
K
X
(' + ) (x) = ( i + i) Ci (x)
i=1
so that
Z K
X
(' + ) d = ( i + i) (Ci \ E)
E i=1
K
X K
X
= i (Ci \ E) + i (Ci \ E)
i=1 i=1
Z Z
= 'd + d
E E

Now,
Z K
X K
X Z
'd = ( i (Ci \ E)) = ( i (Ci \ E)) = 'd
E i=1 i=1 E

Combining these two facts readily allows for the required result.

48
Lemma 71 Let ' and be simple, non-negative functions with
N
X M
X
' (x) = ai Ai (x) and (x) = bi Bi (x)
i=1 i=1

Let be a measure, ' and let E 2 A. Then,

Z Z
'd d
E E

Proof. Again, we may choose a …nite, disjoint collection Ci such that

K
X K
X
' (x) = i Ci and (x) = i Ci
i=1 i=1

with i i so that
Z K
X K
X Z
'd = i (Ci \ E) i (Ci \ E) = d
E i=1 i=1 E

Lemma 72 Let ' be a simple, non-negative function with

N
X
' (x) = ai Ai (x)
i=1

Let be a measure and let E 2 A. Then,

Z Z
'd = ' Ed
E X

Proof. Note that ' E is simple and is equal to aj on E \ Aj and 0 on E c .

Also, note that E \ A1 ; E \ A2 ; :::; E \ Ak ; E c form a partition of X. Then,
Z N
X Z Z
' Ed = aj (Aj \ E) + 0 (E c ) = 'd + 0 = 'd
X j=1 E E

Lemma 73 Let ' be a simple, non-negative function with

N
X
' (x) = ai Ai
i=1

Let be a measure and let E; F 2 A with E F . Then,

Z Z
'd 'd
E F

49
Proof. Follows from Lemmas 71 and 72 because ' E F on X.
Lemma 74 Let ' be a simple, non-negative function with
N
X
' (x) = ai Ai (x)
i=1

Let be a measure and let E 2 A. Then

Z Z
'd = sup d : simple, 0; ' on E
E E

Proof. By Lemma 71
Z Z
d 'd 80 '
E E
hence Z Z
sup d : simple, 0; ' on E 'd
E E
But ' = is a candidate, as well. Hence
Z Z
sup d : simple, 0; ' on E 'd
E E

establishing equality.
Lemma 74 hints at what the de…nition of the integral of a measurable func-
tion should be, since we can approximate a measurable function by a sequence
of simple functions.
De…nition 75 Let f be a measurable and non-negative function. Then, the
integral of f is de…ned as
Z Z
f d = sup d : simple, 0 f on E
E E

Proposition 76 Let (X; A; ) be a measure space, E 2 A and let f : E ! R

be a measurable and non-negative function. Then
Z Z
fd = f Ed
E X
Proof. Let be a simple, non-negative function such that f on E. Then,
E is simple, non-negative function and E f E on X. Hence
Z Z
fd = sup d : simple, 0 f on E
E E
Z
= sup Ed : simple, 0 f on E
X
Z
sup d : simple, 0 f E on X
X
Z
= f Ed
X

50
 Conversely, if f on X, then (x) = 0 if x 2 E c and hence = E so
that
Z Z
f Ed = sup d : simple, 0 f E on X
X X
Z Z
sup d + d : simple, 0 f E on X
X Ec
Z
= sup d : simple, 0 f E on X
X
Z
= sup Ed : simple, 0 f E on X
X
Z Z
= sup d : simple, 0 f on E = fd
E E

which gives us the other equality.

In other words, we have just proved that, for a measure space (X; A; ) ,
E 2 A and f : X ! R measurable and non-negative, if f is integrable over X,
then it is integrable over any E.

Lemma 77 Let (X; A; ) be a measure space, E 2 A and let f; g : E ! R be

measurable and non-negative functions such that f g. Then,
Z Z
fd gd
E E

Proof.
Z Z
fd = sup d : simple, 0 f on E
E E
Z
sup d : simple, 0 g on E
E
Z
= gd
E

where the inequality follows because we are expanding the set over which supre-
mum is taken.

Lemma 78 Let (X; A; ) be a measure space, E; F 2 A such that E F and

let f : E ! R be a measurable and non-negative function. Then,
Z Z
fd fd
E F

Proof. Note that f E f F so that

Z Z Z Z
fd = f Ed f Fd = fd
E X X F

by Lemma 77.

51
Lemma 79 Let (X; A; ) be a measure space and let f : E ! R be a constant,
non-negative function, de…ned by f (x) = c. Then, 8x 2 E,
Z
f d = c (E)
E

Proof. f on E implies cidE on E for each where idE is the identity

function on E. This implies
Z
d c (E)
E

by Lemma 69. This tells us that

Z Z
f d = sup d : simple, 0 f on E c (E)
E E

But = cidE on X is another candidate so

Z Z
c (E) sup d : simple, 0 f on E = fd
E E

3.2 Integral of Sequence of Functions

Theorem 80 (Monotone Convergence Theorem) Let ffn (x) : n 2 Ng be
a sequence of measurable functions. Assume that fn (x) ! f (x) and that fn is
an increasing sequence (i.e., 8x, 0 fn (x) fn+1 (x)), then
Z Z
lim fn d = fd
n!1 E E

Proof. Note that fn (x) f (x) for every n so that

Z Z
fn d fd
E E

Also note that Z

fn d
E
is an increasing sequence of numbers so that
Z Z
lim fn d fd
n!1 E E

For the other inequality, let ' be a simple function with f ' 0. By
de…nition, Z Z
f d = sup d : is simple, f 0
E E

52
Then, the reverse inequality will follow if we can show that
Z Z
lim fn d 'd
n!1 E E

Take a number t 2 (0; 1). Then, En = fx 2 E; fn (x) t' (x)g En+1 . Let
1
[
E= Ei
i=1

If x 2 E, then fn (x) ! f (x) and, for some large n, we have fn (x) tf (x). To
prove this, assume, for the sake of contradiction, that fn (x) < tf (x) for every
n. Then f (x) < tf (x) so that f (x) = 0 and, therefore, fn (x) < tf (x) = 0, a
contradiction to the fact that fn (x) 0 for all x. As an aside, by continuity of
the measure, 8A 2 A,

lim (A \ En ) = (A \ E)
n!1

Now, since ' is a simple function, it can be represented by

N
X
' (x) = aj Aj (x)
j=1

and, for every j,

lim (Aj \ En ) = (Aj \ E)
n!1

so that
Z Z Z N
X
fn d fn d t'd = taj (Aj \ En )
E En En j=1

Passing to the limit, Z

lim fn d
n!1 En

is bounded, so is
N
X
taj (Aj \ En )
j=1

Also, limit can be passed inside summation. So that

Z N
X Z
lim fn d t lim aj (Aj \ En ) = t 'd
n!1 E n!1 E
j=1

which tells us that Z Z

lim fn d t 'd
n!1 E E
Now, we can let t ! 1.

53
Problem 81 Let ffn : n 2 Ng be a sequence of measurable, non-negative func-
tions. Show that
1
X
f (x) = fn (x)
n=1

is a measurable function and

Z 1 Z
X
fd = fn d
X n=1 X

Solution 82 Let k 2 N. Since the …nite sum of (non-negative) measurable

functions is measurable, we must have
k
X
fi (x)
i=1

measurable. Now, since fn 0, we must have

k
X k
X
lim sup fi (x) = lim fi (x)
k!1 k k!1
i=1 i=1

Thus,
k
X
lim fi (x)
k!1
i=1

is measurable. That is,

k
X
lim fi (x) = f (x)
k!1
i=1

is measurable. Again, since the …nite sum of (nonnegative), measurable func-

tions is measurable, we have that
k
X
fi (x)
i=1

is measurable. We can then have

Z X
k
fi (x) d
X i=1

Since each fi is non-negative, we have

Z X
k k Z
X
fi (x) d = fi (x) d
X i=1 i=1

54
by linearity of the integral. Let
k
X
gk (x) = fi (x)
i=1

Then, fgn (x) : n 2 Ng is a sequence of measurable functions, gn (x) ! f (x)

and 0 gn (x) gn+1 (x). Thus, by the Monotone Convergence Theorem,
Z Z Z X
n n Z
X 1 Z
X
f d = lim gn d = lim fi (x) d = lim fi (x) d = fi (x) d
X n!1 X n!1 X n!1
i=1 i=1 X i=1 X

Problem 83 Show that the converse of the Monotone Convergence Theorem

fails.

Solution 84 The following is a modi…cation of a classic example, called March-

ing Intervals. For n 2 N (excludes zero!), de…ne fn : [0; 1] ! R as follows

1 x 2 0; 21n
f2n 1 (x) =
0 otherwise

1 x 2 21n ; 1
f2n (x) =
0 otherwise
Then, Z
lim fn dm1 = 0
n!1

since the intervals are shrinking. However, fn 6! 0 since the sequence is oscil-
lating.

Corollary 85 Let (X; A; ) be a measure space and let f; g : X ! R be non-

negative, measurable functions. Then,
Z Z Z
(f + g) d = fd + gd
X X X

It is possible to prove this corollary using the de…nition directly but let’s use
the Monotone Convergence Theorem.
Proof. Let 'n and n be an increasing sequences of simple functions such that
'n % f and n % g. Then, 'n + n % f + g so that
Z Z Z
(f + g) d = lim ('n + n ) d = lim ('n + n ) d
X X n!1 n!1 X
Z Z Z Z
= lim 'n d + lim nd = fd + gd
n!1 X n!1 X X X

55
Corollary 86 Let (X; A; ) be a measure space, A; B 2 A and let f : X ! R
be a non-negative, measurable function. If A \ B = ?, then
Z Z Z
fd = fd + fd
A[B A B

Proof. Z Z Z
fd = f A[B d = (f A +f B) d
A[B X X
since A \ B = ?. Thus,
Z Z Z Z Z
fd = f Ad + f Bd = fd + fd
A[B X X A B

Lemma 87 (Fatou) If fn 0 is measurable, then

Z Z
lim inf fn d lim inf fn d
N !1n N X X N !1n N

Notice that there is no mention of the limit of f .

Proof. Let
gN (x) = inf fn (x)
n N

Then, gN % and
lim gN (x) = lim inf fn (x)
N !1 N !1n N

By the Monotone Convergence theorem,

Z Z Z
lim gN d = lim gN d = lim inf fn d
N !1 X X N !1 X N !1n N

so that Z Z
gN d fn d
X X
which implies Z Z
lim inf gN d lim inf fn d
n!1n N X n!1n N X
and hence
Z Z Z
lim gN d = lim gN d lim inf fn d
N !1 X X N !1 N !1n N X

That is, Z Z
lim inf fn d lim inf fn d
n!1n N X X n!1n N

56
Exercise 88 Give an example where equality does not hold and another where
the inequality does not hold.
Solution 89 Let E = R and consider fn = (n;n+1) for n 2 N. Then, fn ! 0
pointwise, where 0 is the zero function. However,
Z Z Z
lim inf fn d = (0) d = 0 < 1 = lim inf fn d
E N !1n N E N !1n N E

If we have a sequence of increasing functions ffn : n 2 Ng such that fn ! f ,

then Z Z
lim inf fn d = fd
N !1n N X X

Theorem 90 (Chebyshev Inequality) Let f 0 be a measurable function

and > 0 be a constant. Then,
Z
1
(X ) fd
X
1
where X = fx 2 X : f (x) g=f [ ; 1] is a measurable set.
Proof. Since X X, we must have
Z Z Z
fd fd d = (X )
X X X

Problem 91 Let f be a nonnegative measurable function on X such that

Z
fd = 0
X

Show that f = 0 almost everywhere on X.

Solution 92 Let > n1 and Yn = x : f (x) n1 . Since f is measurable, we
must have Yn 2 A and that (Yn ) 0. Then, by Chebyshev’s inequality,
Z
1
(Yn ) fd = 0
X
1
Thus, (Yn ) = 0 for every > n. Now let
1
[
Y = Yn
n=1

Then, Y = fx : f (x) > 0g and, by countable subadditivity of ,

1
X
(Y ) (Yn ) = 0
n=1

Thus, (Y ) = 0. The result follows from Borel-Cantelli’s Lemma.

57
Problem 93 Show that Fatou’s Lemma implies Monotone Convergence Theo-
rem.

Solution 94 Let ffn (x) : n 2 Ng be a sequence of non-negative measurable func-

tions with fn (x) ! f (x) and 0 fn (x) fn+1 (x). We need to show that
Z Z
lim fn d = fd
n!1 E E

We can view Z
fn d
E
as a sequence of positive, real numbers. In general, we have that
Z Z
lim sup fk d lim inf fk d (3)
n!1k n E n!1k n E

Since fn fn+1 and fn ! f , it must be that fn f for all n. Thus,

Z Z
fn d fd
E E

for all n. Since this is valid for each n, we must have

Z Z
sup fk d fd
k n E E

Letting n ! 1 on both sides gives us

Z Z
lim sup fk d fd (4)
n!1k n E E

Since fn (x) ! f (x), we must have

lim inf fk = lim fn = f

n!1k n n!1

so that Z Z
lim sup fk d lim inf fk d
n!1k n E E n!1k n

by Eq (4). This and Eq (3) together imply that

Z Z
lim inf fk d = lim sup fk d
n!1k n E n!1k n E

and, therefore
Z Z Z
lim fn d = lim inf fk d = lim sup fk d
n!1 E n!1k n E n!1k n E

58
Then, by Fatou’s lemma, we have the chain
Z Z Z Z
lim inf fk d lim sup fk d lim inf fk d lim inf fk d
n!1k n E n!1k n E E n!1k n n!1k n E
(5)
That is, Z Z Z
lim sup fk d = lim inf fk d = lim fn d (6)
n!1k n E n!1k n E n!1 E

(3) and (6) together imply

Z Z Z Z
lim sup fk d f d = lim inf fk d = lim fn d
n!1k n E E n!1k n E n!1 E

Thus, (5) can be re-written as

Z Z Z
lim fn d fd lim fn d
n!1 E E n!1 E

so that Z Z
lim fn d = fd
n!1 E E

3.3 Integral as a Measure

Can we de…ne the integral di¤erently? No. There is, in a certain sense, unique-
ness.

Theorem 95 Let + (X) be the set of all measurable non-negative functions

f : X ! R. Assume J : + (X) A ! R+ [ f1g such that

1. J (f; A) 0 for all f and A

2. A; B 2 A with A \ B = ? =) J (f; A [ B) = J (f; A) + J (f; B)

3. f (x) = c for some constant c on A =) J (f; A) = cJ (id; A)

4. fn % f =) lim J (fn ; A) = J (f; A)
n!1

Then, J is unique and Z

J (f; A) = fd
A

Before we enter a proof of Theorem 95, we demonstrate the following.

These can indeed be proven using the fact that J (f; A) has an integral repre-
sentation but let us show that this follows from the 4 properties above.

Corollary 96 J ( A ; X) = J (id; A)

59
Proof. J ( A ; X) = J ( A ; A [ Ac ) = J ( A ; A) +J( A; A
c
) by (2)
= 1J (id; A) + 0J (id; Ac ) by (3)
= J (id; A)

Corollary 97 J (f; ?) = 0.

Proof. J (f; ?) = J (f; ? [ ?) = J (f; ?) + J (f; ?) by (2) so that J (f; ?) = 0

Corollary 98 Let f and g be measurable functions. Then, J ( f + g; A) =

J (f; A) + J (g; A)

Proof. Let us start with simple functions:

n
X n
X
= ci Ai and ' = di Ai
i=1 i=1

where Ai is a partition of X. Then, by using induction on (2), we get

n
!! n
[ X
J ( + '; A) = J + '; A \ Ai = J ( + '; A \ Ai )
i=1 i=1

By (3), we get
n
X
J ( + '; A) = (ci + di ) J (id; A \ Ak )
i=1
Xn n
X
= ci J (id; A \ Ak ) + di J (id; A \ Ak )
i=1 i=1
= J ( ; A) + J ('; A)

Now since f; g are measurable, 9 simple fn ; gn such that fn % f and gn % g.

Then, (4) gives us the required result. Similarly, we can show that J ( f; A) =
J (f; A).
And now, for a proof of Theorem 95.
Proof. Let (A) = J (id; A). We show that is a measure. (?) = 0
follows immediately from Corollary 97. To show countable additivity, we show
continuity from below. Let An An+1 be an increasing sequence of measurable
sets and let
1
[
Ai = A
i=1

Then, An An+1 and An ! A pointwise on X. By (4), lim J An ; X =

n!1
J( A ; X). By Corollary 96, (A) = J (id; A) = J ( A ; X) = lim J An ; X =
n!1
lim J (id; An ) = lim (An ).
n!1 n!1

60
 Now, let
n
X
f= ci Ai
i=1

be a simple function. Then,

n
! n
[ X
J (f; A) = J f; A \ Ak = J (f; A \ Ak ) by (2)
i=1 i=1
n
X
= ci J (id; A \ Ak ) by (3)
i=1
Xn Z
= ci (A \ Ak ) = fd
i=1 A

Now, for an arbitrary measurable function, we can always come up with an

increasing sequence of simple functions which converge to our arbitrary function.

What follows in this section will be a general treatment of obtaining a mea-

sure from another.

Theorem 99 Let (X; A; ) be a measurable space. Fix a measurable function

f 0. De…ne Z
'f (A) = fd
A
Then, 'f is a measure on A.

Proof. It is clear that 'f (?) = 0. It is also easy to show that 'f is …nitely
additive and monotone, by de…nition. To prove that 'f is countably additive,
we can show 'f is continuous from below: that is, for any increasing sequence
A1 A2 ::: of measurable sets,

lim 'f (An ) = 'f (A)

n!1

where
1
[
A= Ai
i=1

De…ne gn (x) = f (x) An (x). Then, gn is an increasing sequence and converges

to f: A , that is gn % f: A . By Monotone Convergence Theorem,
Z Z Z
lim 'f (An ) = lim gn d = f Ad = f d = 'f (A)
n!1 n!1 X X A

Lemma 100 (Absolute continuity) (A) = 0 =) 'f (A) = 0

61
Proof. If f is simple, we get that
N
X
'f (A) = ak (A \ Ak ) (A) = 0
k=1

If f is not simple, then take fn % f with fn simple. By the Monotone

Convergence Theorem, Z Z
lim fn d = fd
n!1 A A
Since all the integrals on the left side are 0, we must have
Z
fd = 0
A

Lemma 101 Assume that Z

fd < 1
X
Then, f is …nite almost everywhere. Moreover,
Z
lim fd = 0
(E)!0 E

Proof. We can let Z

fd c
X
for some constant c. Now, note that
1
\ 1
\
A = fx : f (x) = 1g = An = ff (x) ng
n=1 n=1

so that (A) (An ) and by Chebyshev’s Inequality,

Z
1 c
(An ) fd
n X n

Since c < 1, we can pass to the limit to get (A) = 0.

For the second part, we need to prove that 8 > 0, there is a > 0 such that
Z
(E) < =) fd <
E

In other words, (E) < =) 'f (E) < .

If f is simple, then
Z N
X
fd = ak (Ak )
X k=1

62
where we can assume that ak < 1, since otherwise we have equality trivially.
Then,
Z XN XN
fd = ak (E \ Ak ) (E) ak
E k=1 k=1

Thus, given > 0, we can set

N
X
= = ak
k=1

If f is not simple, then 9 a sequence of simple functions fgn : n 2 Ng such

that gn ! f . Take a simple function g < f from the sequence of simple
functions such that
Z Z Z
gd fd < gd +
X X X 2
where g , because it is integrable, has …nitely many values and they are all …nite
so that g is bounded (this is di¤erent from 1=x, say, which does not have …nitely
many values). Thus, g (x) c for any x 2 X and so,
Z Z Z Z
fd = (f g ) d + gd (f g ) d + c (X)
X X X X

Also, Z
(f g )d <
X 2
and now take = 2c . Then,
Z
(E) < =) 'f (E) = fd
2c E
Z Z
= f Ed < g Ed +
X X 2
c (E) + =
2

3.4 Integral of Continuous Functions

Let (X; A; ) be a measure space and let f (x) be a -measurable function and
let g (x) = f (x) if f (x) 0 and = 0 otherwise. Then, g (x) is integrable. For
the negative part, let h (x) = f (x). Then, h (x) is also integrable and we
de…ne Z Z
fd = hd
X X
for f (x) 0. In general, provided that f is …nite, we can decompose f (x)
as g (x) + h (x). There is a cleaner, more standard way of doing this: we

63
let f + = max (f; 0) and f = max ( f; 0). f + and f are -measurable by
Corollary 49. Moreover, both functions are non-negative and f = f + f .
Denote jf j = f + + f . If
Z Z
+
f d ; f d <1
E E

then we say that f is -summable (i.e. …nite). The standard terminology

is “integrable”, i.e. the integral is well-de…ned but we will prefer summable.
Thus, we can de…ne
Z Z Z
+
fd = f d f d
E E E

for any E 2 A.

Example 102 Let X = N = f1; 2; :::g, (A) = jAj (counting measure), A

X. Let g : N ! R be a sequence g (n) = gn < 1. Then,
Z X
gd = gn
E n2E

To show this, note that we can write g as

1
X
g= cn fng
n=1

where cn = g (n) and gk = ck fkg . Note that

Z
gk d = ck
N

Furthermore, gk is a sequence of measurable functions such that

1
X
g= gn
n=1

Thus, by Problem 81,

Z 1 Z
X 1
X
gd = gk d = g (n)
N n=1 N n=1

Things are, however, not always this clean. The above representation rests on
the assumption that g (n) < 1 for each n. If we do not have this guarantee, we
can run across 1 1. Moreover, we have convergence issues to deal with, as
well. Therefore, it is not always true that
Z 1
X
f d 6= fn
X n=1

64
 ( 1)n
What if fn = n ? The sequence converges and so does the series

X1 n
( 1)
n=1
n

On the other hand,

1
X
fn+
n=1

diverges. Same for fn , so Z

fd
X
is not well-de…ned!

Problem 103 Let x 2 X and let X; 2X ; x be a measure space where x is

de…ned as
1 x2E
x (E) =
0 x2 6 E
If f is a nonnegative function such that
Z
f d x < 1;

What can we say about f ?

Solution 104 f is measurable since for any measurable subset of the codomain
of f , the pre-image is a subset of X and hence in the -algebra – measurable.
We also know that f is summable because the integral is …nite. By Lemma
101, f is …nite almost everywhere. That is, the set
1
\
fy 2 X : jf (y)j > ng
n=0

has x -measure zero. Hence

1
\
x 62 fy 2 X : jf (y)j > ng
n=0

That is, 8n, x 62 fy 2 X : jf (y)j ng. In particular, jf (x)j 0. Since x is

arbitrary, we can conclude that f is the zero function.

Lemma 105 Let f be a -measurable function, de…ned on a measure space

(X; A; ). Then, f is -summable if and only if jf j is -summable. Moreover,
Z Z
fd jf j d
X X

65
 This follows easily from the basic de…nition. This feature is super important
and is what makes Lebesgue integral di¤erent from Riemann integral, which is
discussed in the next section.
Proof. f is -summable ()
Z Z
+
both f d < 1 and f d <1
E E

Then, Z Z Z
+
f d + f d = jf j d < 1
E E E
and conversely. Furthermore,
Z Z Z Z Z
fd = f +d f d f +d + f d
E E E E E
Z Z Z
= f +d + f d = jf j d
E E E

Problem 106 Show that the converse doesn’t hold

Solution 107 Let (X; A; ) be a measure space with (X) < 1 and let A X
such that A 62 A. De…ne f = 1 + 2 A . Then, jf (x)j = 1 for all x so that the
jf j is summable. However, f is not measurable since A is not measurable.

Problem 108 Let fn be a sequence of summable functions and fn ! f for a

summable f . Show that
Z Z Z
jfn f j d ! 0 () jfn j d ! jf j d
X X X

Solution 109 Let fn = fn+

fn and f = f +
f , and fn f = gn = gn+ gn .
+ + +
Rearranging gives us gn + fn + f = fn + f + gn so that
Z Z Z Z Z Z
gn+ d + fn d + f +d = fn+ d + f d + gn d (7)
X X X X X X

Now, the …rst condition is equivalent to

Z Z Z Z
jfn f j d ! 0 () jgn j d ! 0 () gn+ d + gn d ! 0
X X X X
Z Z
() gn+ d ! 0 and gn d ! 0
X X

66
so that applying limit to Eq (7) gives us
Z Z Z Z
lim fn d + f +d = lim fn+ d + f d
n!1 X X n!1 X X
Z Z
() lim fn+ d fn d
n!1 X
Z Z X Z Z
+
= lim fn d = f d f d = fd
n!1 X X
Z Z Z X ZX
() fn d ! f d () jfn j d ! jf j d
X X X X

Now for some notation: if f is -summable, then write f 2 L1 ( ). The 1 in

the superscript is immaterial for now but its importance will become relevant
when we cover Lp spaces.

Theorem 110 Let f; g 2 L1 ( ) and t 2 R. Then, tf 2 L1 ( ), f + g 2 L1 ( )

and Z Z Z Z Z
tf d = t f d and (f + g) d = fd + gd
E E E E E
R 1
Proof. Assume that t = 0. Then, E tf d = 0 so that tf 2 L ( ). If t 6= 0, we
know that tf is measurable by Corollary 49. We show that jtf j is summable.
Z Z Z Z Z
jtf j d = jtj jf j d = jtj f + + jtj f d = jtj f + d + jtj f d < 1
E E E E E
_
since the latter two are …nite, because f + and f are both summable.
Next, let h = f +g so that h+ h = f + f +g + g . Then, h+ +f +g =
f + g + + h so we can use linearity since we are working with non-negative
+

functions. This gives us

Z Z Z Z Z Z
+ + +
h d + f d + g d = f d + g d + h d
E E E E E E

Rearranging this gives us

Z Z Z Z Z Z
h+ d h d = f +d f d + g+ d g d
E E E E E E
Z Z
= fd + gd
E E

We know that h is measurable but is it summable? As shown in Problem 106,

it does not follow that h 62 L. Note that since the two terms on the RHS are

67
…nite, it follows that the term on the LHS is …nite.
Z Z Z Z
h+ d h d h+ d + h d
E E
ZE ZE Z Z
+ +
h d + h d = h d + h d
ZE Z E Z E Z E
=) h+ d h d h+ d + h d
E E E E
Z Z Z
=) hd fd + gd
E E E

because h+ = f + + g + and h = f + g . Therefore, h = f + g 2 L1 ( ) and

the Lebesgue Integral is linear.
We are assuming that
Z Z
f d < 1 and g d <1
E E

so, by Lemma 101, the four functions f and g are …nite almost everywhere.
Therefore, f + g is well de…ned and …nite on XnN where (N ) = 0 and we
can extend f + g to be (honestly) …nite by de…ning (f + g) (n) = 0 for n 2 N .
The product of two summable functions is not summable,
p however. For a simple
example, take I = (0; 1] and f (x) = g (x) = 1= x and = m1 .

Theorem 111 Let f; g 2 L1 ( ) with f g. Then, for any E 2 A,

Z Z
fd gd :
E E

Proof. Again, f + f g+ g so that f + + g f + g + . We can then

use linearity of the integral.

Theorem 112 (Lebesgue Dominated Convergence Theorem) Let fn 2

L1 ( ), fn (x) ! f (x) for every x and jfn (x)j F (x) for every x and n for
some F 2 L1 ( ). Then,
Z Z
lim fn d = fd :
n!1 E E

Proof. Case 1. f (x) = 0 for all x 2 X and fn 0.

Then, Z
fn d 0
E
so that Z Z
lim fn d = lim inf fn d 0:
n!1 E n!1n N E

68
Let us consider F (x) fn (x) 0. By Fatou’s lemma,
Z Z
lim inf (F fn ) d lim inf (F fn ) d :
n!1n N E E n!1n N

The lower limit is not linear! It is monotone in one direction but not linear.
Recall that

lim inf (A + an ) = A + lim inf an and lim inf ( an ) = lim sup an

n!1n N n!1n N n!1n N n!1n N

Thus, Z Z
lim inf (F fn ) d F lim sup fn d
n!1n N E E n!1n N

Since fn ! 0, we have
Z Z Z Z
lim inf (F fn ) d = lim inf Fd fn d Fd
n!1n N E n!1n N E E E
Z Z Z
=) Fd lim sup fn d Fd
E n!1n N E E
Z
=) lim sup fn d 0
n!1n N E
Z Z
=) lim fn d = 0 = fd
n!1 E E

The rest of the cases are left as an exercise.

Lemma 113 Let (X; A; ) be a measurable space with (X) < 1. Let f be a
summable function. Then, f is …nite almost everywhere.

Proof. Consider fn = n En and let En = fx 2 X : f (x) ng. Then, we have

a decreasing sequence of sets En En 1 . Now for all x in En we have that
n En (x) = n jf (x)j. From the monotonicity of the integral,
Z Z Z
n (En ) = n En d jf jd jf jd = C < 1
X En X

for some C 2 R. That is, (En ) C=n. Now set

1
\
E= En
n=1

i.e., x belongs to E i¤ jf (x)j = 1. Since (E1 ) < 1 (because (X) < 1),
then by continuity of measure,

0 (E) (En ) C=n =) (E) = 0

69
Corollary 114 Let (X; A; ) be a measurable space with (X) < 1 and let
fn ,f be functions which are …nite almost everywhere for each n. Assume that
jfn j C for each n and fn ! f . Then
Z Z
lim fn d = fd
n!1 X X

Proof. We can let C (x) = C be a constant function. Then C 2 L1 ( ).

Moreover, note that, for each n,
Z Z
fn d jfn j d C (X) < 1
X X

hence fn 2 L1 ( ). Hence by Lebesgue’s Dominated Convergence Theorem, the

result follows.

Problem 115 Assume ffn : n 2 Ng is a decreasing sequence of positive mea-

surable functions with fn (x) ! f (x) for every x 2 X. If f1 is summable, show
that Z Z
lim fn d = fd
n!1 X X
but this might not be true if f1 is not summable.

Solution 116
R Since we
R have f1 fn for all n, and in particular, f1 f.
Moreover, X fn d f
X 1
d . By Fatou’s Lemma,
Z Z Z Z
1> f1 d lim inf fn d lim inf fn d = fd
X N !1n N X X N !1n N X

Hence f is summable and positive almost everywhere. The result then follows by
Lebesgue’s Dominated
n Convergence. o
Consider fn = n (0;1=n 1) : n 2 N . Then, f1 is not summable, fn+1 fn
and fn ! 0 pointwise but
Z Z
n
lim fn d = lim = 1 6= fd = 0
n!1 R n!1 n 1 R

Problem 117 Suppose is a positive measure on X, f is a non-negative mea-

surable function and Z
fd = c < 1
X
Show that
8
Z < 1 for 2 (0; 1)
lim n log (1 + (f =n) ) d = c if = 1
n!1 X :
0 if > 1

70
Solution 118 Let fn (x) = n log (1 + (f (x) =n) ). Then, fn 0. Consider the
case that 0 < < 1. Then, (f (x) =n) ! 0. By L’Hospital’s Rule,
f (x) =n +1
log (1 + (f (x) =n) ) 1+(f (x)=n) f (x)
lim 1 = lim 1 = lim =1
n!1 n!1 n!1 n 2 + f (x) n 2
n n2
R R R
Hence
R by Fatou’s lemma, lim X fn d X
(limfn ) d = 1 so that lim X fn d =
lim X fn d = 1. Next, suppose that = 1 and recall that x 0 =)
log (1 + x) x. Thus,
f f
fn = log 1 +
n n
and so fn f . Thus, each fn is summable and by the Lebesgue Dominated
Convergence Theorem,
Z Z
lim fn d = f d = c
n!1

Finally, consider > 1. Then,

f (x) =n +1
log (1 + (f (x) =n) ) 1+(f (x)=n) f (x)
lim 1 = lim 1 = lim 1
=0
n!1 n!1 n!1 n + f (x) =n
n n2

By Fatou’s Lemma, the result follows.

3.5 Convergence in Measure

De…nition 119 Let L0 ( ) be the set of -measurable functions which are …nite
almost everywhere. If ffn : n 2 Ng[ff g L0 ( ), we say that fn converges to f
in measure, written as fn =) f , if 8 > 0, (fx : jfn (x) f (x)j g) ! 0
as n ! 1.

Problem 120 If ffn : n 2 Ng [ ff; gg L0 ( ), and fn =) f and fn =) g,

then f = g almost everywhere

Solution 121 Let > 0, Ef g := fx : jf (x) g (x)j g, Ef = x : jf (x) fn (x)j 2

and Eg = x : jg (x) fn (x)j 2 . Then, note that Ef g Ef [ Eg . To show
this, let x 62 Eg \ Ef . Then, jfn (x) f (x)j < =2 and jfn (x) g (x)j < =2
so that jf (x) g (x)j = jf (x) fn (x) + fn (x) g (x)j jf (x) fn (x)j +
jfn (x) g (x)j < so that x 62 Ef g . Then, by monotonicity of , it fol-
lows that (Ef g ) (Ef ) + (Eg ). Since (Ef ) ; (Eg ) ! 0, it follows that
(Ef g ) ! 0 so that g (x) = f (x) almost everywhere.

Theorem 122 (Lebesgue) Let (X; A; ) be a measurable space with (X) <
1, ffn : n 2 Ng [ ff g L0 ( ) such that fn ! f almost everywhere. Then
fn =) f .

71
 That is, on a set of …nite measure, almost every convergence implies conver-
gence in measure.
Proof. We will prove the theorem for the case fn ! f everywhere where fn ; f
are …nite everywhere. The result will then follow directly from this fact as
“everywhere =) almost everywhere”.
Consider the error set Ek ( ) = fx : jfk f j g. Our goal is to show that
(En ( )) ! 0 where
1
[
En ( ) = Ek ( )
k=n

De…ning En ( ) this way gives us a decreasing sequence E1 ( ) E2 ( ) :::.

Since we have convergence of fn , we have (E1 ( )) < 1 so that
1
!
\
(En ( )) ! En ( ) :
n=1

by continuity of measure. We now claim that

1
\
En ( )
n=1

is empty: if
1
\
x2 En ( )
n=1

then for every n, 9N n such that jfn (x) f (x)j . Thus, fn does not con-
verge, a contradiction. Thus, (En ( )) ! 0. Since 0 (En ( )) (En ( )),
it follows the fn =) f .
Converse is not true!

Example 123 Consider X = [0; 1], the Lebesgue Measure and sequence fn =
k
[ 2jk ; j+1
2k
] where k = blog2 nc and j = n 2 . This sequence is called the Type
Writer sequence. The …rst …ve terms of the sequence are [0; 1 ] , [ 1 ;1] , [0;1=4] ,
2 2

[ 1 ; 1 ] . As n increases, the intervals shrink further. Thus, for > 0, the measure
4 2
of the set En = fx : jfn (x)j g approaches zero so that fn =) 0, the zero
function, so we have convergence in measure. Moreover, each fn is …nite, hence
is …nite almost everywhere. However, for any x 2 [0; 1], fn does not converge
to any function since the sequence is oscillating. Hence fn does not converge
anywhere.

The …niteness condition is necessary, as well. Consider fn = (n;n+1) so that

fn ! f 0 almost everywhere. However, for = 1=n, and n > 1, m1 (En ) =
m1 (fx : jfn (x)j 1=ng) = m1 (n; n + 1) = 1 and this does not converge to zero.
An even stronger statement to Lebesgue is the following:

Theorem 124 (Reisz) Every sequence that converges in measure contains a

subsequence which converges almost everywhere to the same limit.

72
Theorem 125 Let (X; A; ) be a measurable space, ffn : n 2 Ng [ ff g ; fF g
L1 ( ) such fn =) f and jfn j F . Then
Z Z
lim fn d = fd
n!1 X X

This is not Lebesgue’s Dominated Convergence Theorem since we have con-

vergence in measure, not the usual point-wise convergence. That is, for a
bounded sequence which converges in measure, the integrals coverge.
Proof. We …rst claim that that jf j F almost everywhere
To see this, we note that jfn j F for all n tells us that
sup jfn j F =) lim sup jfn j = jf j F
n N N !1n N

Case I
Assume that (X) < 1.
Since jf j F , we can assert, by triangle inequality, that jfn f j 2F is
valid for all x. Now let En ( ) = fx 2 X : jfn (x) f (x)j g. Such a collection
is measurable. Then,
Z Z Z
jfn f j d = jfn f j d + jfn f j d
X En ( ) XnEn ( )

The …rst sum is sort of a global estimate (i.e. 2F ) whereas the second
sum is a local one (i.e. < ). Since XnEn ( ) = fx 2 X : jfn (x) f (x)j < g,
we must have Z Z
jfn f j d 2F d + (X) :
X En ( )

For a …xed , by Lemma 101, (En ( )) ! 0 tells us that

Z
jfn f j d (X) :
X

We do not know if the limits exist, yet we can apply lim sup on both sides to
get. Z
lim sup jfn f j d (X) :
n!1k n X

Notice that in the equality, nothing depends on n on the right side and every-
thing on the left is true for every . Thus, we can let ! 0 to get
Z
lim sup jfk (x) f (x)j d = 0
n!1k n X

Now,
Z
0 jfn (x) f (x)j d
X
Z Z
=) 0 lim inf jfn (x) f (x)j d lim sup jfk (x) f (x)j d = 0
n!1k n X n!1k n X

73
so that Z
lim jfn (x) f (x)j d = 0
n!1 X
Since Z Z
(fn (x) f (x)) d jfn (x) f (x)j d
X X
we can apply limit on both sides to conclude that
Z Z
lim fn (x) d = f (x) d
n!1 X X

Case II
1
Now assume that (X) = 1. It R can be shown that 8g 2 L ( ) with g 0,
9A 2 A such that (A) < 1 and XnA gd < . We use this without proof.
Then, jfn f j 0 and jfn f j 2 L1 ( ). Use this and the R continuity of
'
Rf (:) to prove the
R second case. The plan
R would be to use R X jfn f j d =
jf
A n
f j d + jf
XnA n
f j d . Thus, jf
XnA n
f j d XnA
2F d . Now
use g = 2F .
We cannot hammer uniform continuity in here because (XnA) = 1.
The assumption of boundedness of the sequence is essential, even if the space
has …nite measure.

Example 126 Consider X = (0; 1), fn = n (0;1=n) . We have convergence in

measure to the zero function. However,
Z
fn d = n (0; 1=n) = 1
X

for all n so that we do not have convergence of integrals.

Problem 127 Let (X; A; ) be a measurable space, with (X) < 1. Show that
the sequence ffn : n 2 Ng converges in measure to f if and only if
Z
jfn f j
lim d =0
n!1 X 1 + jfn fj
Solution 128 Let > 0 and En = fx : jfn (x) f (x)j g. Since X = En [
Enc , we have
Z Z Z
jfn f j jfn f j jfn f j
d = d + d
X 1 + jfn f j E 1 + jfn f j Ec 1 + jfn f j
By second part of Lemma 101, the …rst limit of the …rst integral on the right
is zero. For the second integral, we have Enc = fx : jfn (x) f (x)j < g. Note
jfn f j
that 1+jf n fj
< jfn f j < so that
Z Z Z
jfn f j jfn f j
d d < d = (Enc ) < (X)
c
En 1 + jfn f j c
En 1 + jfn f j c
En

74
Thus, in the limit, this integal is zero, as well.
Conversely, assume that the sequence does not converge in measure. That
is, 9 > 0 such that lim (En ) 6= 0 where En = fx : jf (x) fn (x)j g. That
n!1
is, 9 > 0 such that 8N , j (En )j for some n < N . Since En X, we have
Z Z Z
jfn f j jfn f j
d d d = (En )
X 1 + jfn fj En 1 + jfn fj En 1 + 1+ 1+
so that the sequence of integrals does not converge, a contradiction.

3.6 Riemann Integral

De…nition 129 The Riemann integral of a function f is de…ned as the number
Z b X
f (x) dx = lim f k;n (xk;n xk 1;n )
a n!1

where a = x1;n < x2;n < ::: < xn;n = b and k;n 2 [xk 1;n ; xk;n ]

The …rst condition is that size of every partitions tends to zero and that the
limit does not depend on k;n . A very important example is as follows

1 x2Q
f (x) =
6 Q
0 x2
in which case the integral does depend on k;n .
If f is continuous, then the integral exists. The axioms of Lebesgue measure
can be given as follows: de…ne
Z b
f ( )= f (x) dx
a

where is a compact interval. Then, the following hold

1. f ( 1) + f ( 2) = f ( ) where 1 = [a; c] and 2 = [c; b]. This is

where the trouble comes in (the end points of intervals should match).
2. c ( ) = c (b a) where c is a constant function taking value c every-
where.
3. If f g, then f ( ) g ( )
These three properties uniquely de…ne the Riemann integral. To show the
integral exists (assuming that f is continuous) on closed intervals, by uniform
continuity, we can have a that does not depend on x. We can then split the
closed interval [a; b] by to get f ( ) f (x) f ( ) + , which leads to the
de…nition of the Riemann integral
Thus, instead of using the de…nition of Riemann integral, we could use the
three axioms. These three already hold for the Lebesgue integral, thus the
mantra “Riemann Integral =) Lebesgue Integral” for a compact domain.

75
Theorem 130 Let f 2 C ([a; b]). Then, f 2 L1 (m1 ) and
Z Z b
f dm1 = f (x) dx
[a:b] a

Proof. f 1 (a; 1), by continuity, is open and every open set is measurable.
That is, f 1 (a; 1) 2 M1 . Now, jf (x)j M for every x 2 [a; b] and m1 ([a; b]) =
b a. Note that adding or subtracting c from an interval (to make the two
intervals disjoint) doesn’t make much a di¤erence since it is just erasing or
adding a set of measure 0. Thus, f 2 L1 (m1 ). This essentially means
Z
'f ( ) = f dm1 :

By uniqueness, 'f ( ) = f ( ).
The converse, of course fails. Recall the Dirichlet criterion for improper
integrals: if the Riemann integral of a function f is uniformly bounded over
all intervals, and g is a monotonically decreasing non-negative function, then
the Riemann integral of f g is a convergent improper integral. Therefore, the
Riemann integral is de…ned in
Z 1
sin x
dx:
1 x
if we take g (x) = 1=x and f (x) = sin x. However, it has no chance of being
summable because Z
sin x
dm1 = 1:
(1;1) x
We will prove rigorously the above when we discuss functions of bounded vari-
ation, but for now, let us take that on faith.
There are ways to work around this limitation:
Theorem 131 Let f 2 C ([a; b)) where b 2 R [ f1g. Then, f 2 L1 (m1 ) ()
Z b Z b
jf (x)j dx = lim jf (x)j dx < 1
a t!b a

In this case,
Z b Z
jf (x)j dx = f dm1
a [a;b)

Proof. (=))
f 2 L1 (m1 ) =) jf j 2 L1 (m1 ). If t < b, then f 2 C ([a; t]). By the previous
result,
Z t Z
jf (x)j dx = jf j dm1
a [a;t]
Z Z
= jf j dm1 jf j dm1
[a;b) (t;b)

76
Our goal is to show that Z
lim jf j dm1 = 0:
t!b (t;b)

If b 2 R, then it follows trivially since m1 ((t; b)) ! 0. If b = 1, consider a

monotonically increasing, unbounded sequence fan : n 2 Ng with a1 = t. Then,
lim t = lim an
t!b n!1

We can then consider the decreasing sequence of intervals (an ; b) (an+1 ; b).
1
\
Let Ek = (an ; b). Then, Ek Ek+1 so that by Absolute Continuity of the
n=k
Lebesgue measure,
Z 1
!
\
lim jf j dm1 = 'jf j (an ; b) = lim 'jf j (En ) = 0
t!b (t;b) n!1
n=1

( (= ) If the Riemann integral exists, then we need to prove that the

Lebesgue integral exists, provided that
Z b Z b
jf (x)j dx = lim jf (x)j dx < 1
a t!b a

If t < b, by Theorem 130,

Z Z t
'jf j ([a; t]) = jf j dm1 = jf (x)j dx
[a;t] a

Choose ftn : n 2 Ng such that tn % b. Then, [a; t1 ] [a; t2 ] ::: . Since 'jf j is
a measure, then, by continuity of measure,
Z b Z t Z
jf (x)j dx = lim jf (x)j dx = lim jf j dm1 = lim 'jf j ([a; tn ]) < 1:
a t!b a t!b [a;t] n!1

Problem 132 Let f (x) = x 1=2 for x 2 (0; 1] with f (0) = 1. Without refer-
ring to Riemann integration, prove that f is summable with respect to m1
Solution 133 Let
n
X
'= ci Ai f
i=1
be a simple function with Ai being the partition of (0; 1]. Then,
Z Xn
dm1 = ci m1 (Ai \ (0; 1]) < 1
(0;1] i=1

and so
Z Z
sup d : simple, 0 f on E = fd < 1
E E

77
3.7 Product Measures
The goal of this section is to de…ne product of measures, and, therefore, multiple
integrals. That is, dA = dxdy = dm1 dm1 = dm2 where the product commutes.
This is necessary since not in all cases do we get an iterated integral. Proving
this is the main motivation, so we can ultimately compute
Z 1
2
e x dx
1

Some notation: we will have two measure spaces (X; AX ; X ) and (Y; AY ; Y )
and de…ne P = fA B : A 2 AX and B 2 AY g. This is at least a semiring
but, sadly, at most a semiring. The fact that P is a semiring follows from
Problem 26. To show that the product of two measures is not a measure, even
if the measure is complete, consider X = [0; 1] = Y and let X = Y = m1 .
Let A be a nonmeasurable susbet of [0; 1]. Then, A f0g 62 P. Note that
A f0g [0; 1] f0g and the latter has measure zero. Now take E = 0; 14
3
and F = 4 ; 1 . Then, clearly E [ F 2 AX = AY . Moreover, E E; F F 2 P.
However, (E E) [ (F F ) 62 P.
Since we’ve seen that we can construct a measure using an ordinary set
function by Lebesgue-Carathéodory theorem, it would not be outlandish to
propose a set function 0 (A B) = X (A) : Y (B), which is naturally well-
de…ned on P. Let us prove that it satis…es the hypothesis of the Lebesgue-
Carathéodory theorem.

Lemma 134 0 is a pre-measure.

Proof. P1 0 (?) = X (A) : Y (?) or X (?) : Y (B) for any A2AX or B 2

AY , and both are zero.
P2 To show that 0 is …nitely additive, let fAj : 1 j ng AX and
fBk : 1 k mg AY be collections of disjoint sets and let
n
[ m
[
A= Aj and B = Bk
j=1 k=1

Then,
n
! m
! m
n X
X X X
0 (A B) = X (A) Y (B) = (Ai ) (Bk ) = (Ai ) (Bk )
i=1 k=1 i=1 k=1

On the other hand, since

0 1 !
n m (n;m)
[ [ [
A B=@ Aj A Bk = (Ai Bk )
j=1 k=1 (j;k)

we therefore have equality in both sides.

78
 P3 For countable monotonicity, let
1
[ 1
[
E=A B Aj Bj = Ej
j=1 j=1

where Ej = Aj Bj 2 P. We need to show that, for E,

1
X 1
X 1
X
0 (E) = X (A) Y (B) X (Aj ) Y (Bj ) = 0 (Aj Bj ) = 0 (Ej )
j=1 j=1 j=1

De…ne
B x2A
Ex =
? x 62 A
Then,
1
[
Ex Ejx
j=1

where
Bj x 2 Aj
Ejx =
? x 62 Aj
It follows that, for x 2 X,
1
X
Y (E x ) Y Ejx :
j=1

This is because x 2 A implies

1
X
Y (B) Y Ejx
j=1

and if x 62 A,
1
X
Y (?) = 0 Y Ejx :
j=1
x
Also, Y (E ) = 0 if, for each j, Aj (x) = 0, assuming that no Bj is null. It
follows that
1
X
X (A) Y (B) X (Aj ) Y (Bj ) :
j=1

Now, by Lebesgue-Carathéodory theorem, we can extend 0 to a measure

= X Y , on some -algebra on X Y.

Example 135 Consider mr on Rr and ms on Rs . Then, mr ms is a measure

on Rr+s .

79
 This is also the same as the canonical measure on Rr+s since they are de…ned
on the same -algebra and agree on each element of the -algebra. A rigorous
proof of this fact will be skipped.

De…nition 136 A measure is called -…nite if X can be written as a count-

able union of measurable sets, each of …nite measure.

Example 137 m1 on R, where Xj = 2j ; 2j

Example 138 The counting measure is not -…nite on R since any in…nite set
has measure in…nity yet may not cover R.

For the rest of this section, we assume that all measures are -…nite and
complete. Let (X; AX ; X ) and (Y; AY ; Y ) be two such measure spaces. Let
(Z; AZ ; Z ) where Z = X Y be the completion of the product spaces.

Proposition 139 (Z; AZ ; Z) is -…nite.

Proof. Let
1
[ 1
[
X= Ai and Y = Bi
i=1 i=1

Then, Ai ; Bi 2 P =AX AY and since 0 = jP , the result follows.

Lemma 140 If (Z; AZ ; Z)is the complete measurable space obtained from pre-
measure 0 on P =AX AY , then for any E 2 A, 9G with E G such that
(GnE) = 0 and
1 [
\ 1
G= Ak;n ;
n=1k=1

where Ak;n 2 P.

Proof. That fact that we can always …nd a G with the given measure holds true
because of completeness. If E is a bounded region, then we can have rectangles
Ak;n that determine this region. If E is unbounded, then the result is trivially
true.
We now prove Fubini’s theorem for product of characteristics functions,
which is basically a variant of Cavalieri’s Principle (which we state without
proof):

Theorem 141 (Cavalieri’s Principle) Let (X; AX ; X ) and (Y; AY ; Y ) be

two -…nite measure spaces. Let (Z; AZ ; Z ) where Z = X Y be the com-
pletion of the product spaces. Let C be a measurable set in AZ and C x =
fy 2 Y : (x; y) 2 Cg. Then

1. For almost every x, C x 2 Ay

2. The function x 7 ! Y (C x ) X -is measurable

80
 3. The integral over C can be calculated by iterations:
Z
x
(C) = (mr ms ) (C) = Y (C ) d X
X

The generalization of Cavalieri’s Principle to non-negative functions allows

functions to be integrated in the product space, even if they are not summable.
This allows us to show that 0 , the pre-measure obtained by the product of
measure, is countably additive.
Proof. We know that 0 is countably subadditive by Lemma 134. Let
fAj Bj : j 2 Ng be a pairwise disjoint and let
1
[
E=A B= Ej
n=1

where Ej = Aj Bj . Then, for each (x; y) 2 X Y , note that Aj Bj (x; y) =

Aj (x) Bj (y). Thus,

1
X
E (x; y) = Aj (x) Bj (y)
n=1

Since we have Bj (y) Y -measurable, the integral

Z X
1

Aj (x) Bj (y) d Y
Y j=1

is well-de…ned. Moreover, by Monotone Convergence Theorem,

Z n
X Z Xn
lim Aj (x) Bj (y) d Y = lim Aj (x) Bj (y) d Y
n!1 n!1
Y j=1 Y j=1

so that by linearity of measure,

Z X n n Z
X
lim Aj (x) Bj (y) d Y = lim Aj (x) Bj (y) d Y
n!1 n!1
Y n=1 n=1 Y

Xn Z 1
X Z
= lim Aj (x) Bj (y) d Y = Aj (x) Bj (y) d Y
n!1
n=1 Y n=1 Y

Note that Z
Bj (y) d Y = Y (Bj )
Y

so that we have
1
X
Aj (x) Y (Bj )
n=1

81
Now, Aj (x) is X -measurable so that
Z X
1

Aj (x) Y (Bj ) d X
X n=1

Applying the same steps as above gives us

1
X Z 1
X 1
X
Y (Bj ) Aj (x) d X = Y (Bj ) X (Aj ) = X (Aj ) Y (Bj )
n=1 X n=1 n=1

Now,
0 1
Z Z Z
@ Ad y
E (x; y) d Y X = Y (E ) Ex (x) d X
X Y X
= Y (E y ) X (E x ) = X (E x ) Y (E y ) = X (A) Y (B) = 0 (E)

where
B x2A
Ex =
? x 62 A
and
A x2B
Ey =
? x2
6 B
Thus,
1
X 1
X
0 (E) = X (Aj ) Y (Bj ) = 0 (Ej )
n=1 n=1

Let (X; AX ; X ) and (Y; AY ; Y ) be measurable spaces, (Z; AZ ; Z ) be a

completion of the product space Z = X Y and let f : Z ! [0; 1] be
a non-negative function (not necessarily Z -measurable) De…ne fy such that
fy
x 7 ! f (x; y) with the integral
R
X
fy d X if it exists
I (y) =
0 otherwise

Then, I (y) exists for almost every y. Assuming that f is Z -measurable, then
Z Z
I (y) d Y = fd
Y Z

Theorem 142 (Tonelli) Let (X; AX ; X ) and (Y; AY ; Y ) be two -…nite mea-
sure spaces. Let (Z; AZ ; Z ) where Z = X Y be the completion of the product
spaces and let f : Z ! [0; 1] be Z -measurable, then

82
 1. For almost every y 2 Y , the following integral exists
Z
fy d X
X

2. Moreover, we have the iterated integral

Z Z Z
fy d X d Y = fd
Y X Z

fx
We can equivalently have fx such that y 7 ! f (x; y), giving us
Z Z Z Z
fy (x) d X d Y = fx (y) d Y d X :
Y X X Y

Theorem 143 (Fubini) Let (X; AX ; X ) and (Y; AY ; Y ) be two -…nite mea-
sure spaces. Let (Z; AZ ; Z ) where Z = X Y be the completion of the product
spaces. Let f : Z ! R be Z -summable. Then,

1. almost every y 2 Y , the following integral exists

Z
fy d X
X

2. Moreover, we have the iterated integral

Z Z Z
fy d X d Y = fd
Y X Z

Corollary 144 If g (x) is summable on X and h (y) is summable on Y , then

f (x; y) = g (x) h (y) is summable on Z. Moreover,
Z Z Z
f (x; y) d = gd X hd Y
Z X Y

However, we may still not have measurability! To get that, set g1 (x; y) =
g (x) and h1 (x; y) = h (y). In this case, g1 and h1 are measurable, then their
product is measurable.

Problem 145 Let (X; AX ; X ) and (Y; AY ; Y ) be -…nite and complete mea-
sure spaces, g be summable on X and h be summable on Y . Show that f (x; y) =
g (x) h (y) is summable on X Y and that
Z Z Z
f (x; y) d = gd X hd Y
Z X Y

83
Solution 146 Note that X is AX -measurable and Y is AY -measurable implies
X Y is measurable on the -algebra generated by AX AY , say AZ . Let
g1 (x; y) = g (x) and h1 (x; y) = h (y). By Cavalieri’s Principle, g1 and h1 are
AZ -measurable. Since the product of measurable functions is measurable, f (x; y)
is AZ -measurable. We will show that jf j is summable. By Toneli’s Theorem,
Z Z Z
jf (x; y)j d = jg1 (x; y) h1 (x; y)j d Y d X
Z X Y

Now, g1 (x; y) = g (x) and h1 (x; y) = h (y) so that

Z Z Z Z
= jg (x) h (y)j d Y d X = jg (x)j jh (y)j d Y d X
ZX Y Z X Y
Z Z
= jg (x)j jh (y)j d Y d X = jg (x)j jh (y)j d Y d X
X Y X Y
Z Z
= jg (x)j d X jh (y)j d Y
X Y

Now, since g and h are summable, then so is jgj and jhj. Thus, the multiplicands
in the last line are two …nite numbers and, therefore, the product is …nite. Thus,
jf j is summable so that f is summable. Thus, by Fubini’s theorem,
Z Z Z
f (x; y) d = gd X hd Y
Z X Y

Problem 147 Let (X; AX ; X ) and (Y; AY ; Y ) be measure spaces with X =

Y = N with X and Y counting measures. De…ne
8
< 2 2 x x=y
f (x; y) = 2+2 x x=y+1
:
0 otherwise

Show that f is measurable with respect to X Y . Does it contradict the

Fubini’s Theorem?

Solution 148 We have that AZ = 2N 2N since AX = AY = 2N so that every

subset of N N is measurable. Thus, the preimage of any measurable subset of

84
R under f will naturally be measurable. Now,
Z Z Z Z
f (x; y) d X d Y = f (x; y) d X d Y
N N N N
Z 1
X Z 1
X
= f (k; y) d Y = fk (y) d Y
N k=1 N k=1

XZ
1 1
XX 1
= fk (y) d Y = f (k; j)
k=1 N k=1 j=1
X1
= (f (k; 1) + f (k; 2) + :::)
k=1
8
>
> f (1; 1) + f (2; 1) + f (3; 1) + :::+
>
>
< (1; 2) +
> f f (2; 2) + f (3; 2) + :::+
= f (1; 3) f (2; 3) f (3; 3) + :::+
>
> f (1; 4) f (2; 4) f (3; 4) + :::+
>
>
>
: .. ..
. .
= f (1; 1) + f (2; 1) + f (2; 2) + f (3; 2) + :::
= 2 2 1 + 2+2 2 + 2 2 2 + 2+2 3
+ :::
= 3=2 7=4 + 7=4 15=8 + 15=8 + ::: = 3=2
On the other hand,
Z Z 1 X
X 1
f (x; y) d Y d X = f (j; k)
N N k=1 j=1
1
X
= (f (1; k) + f (2; k) + :::)
k=1
8
>
> f (1; 1) + f (1; 2) + f (1; 3) + :::+
>
>
< (2; 1) + f (2; 2) +
> f f (2; 3) + :::+
= f (3; 1) f (2; 3) f (3; 3) + :::+
>
> f (4; 1) f (2; 4) f (3; 3) + :::+
>
>
>
: .
.. ..
.
1
X 1
X
k k 1 k 1
= 2 +2 = 2 ( 2 + 1)
k=1 k=1
X1 1
k 1 1X k 1
= 2 = 2 =
2 2
k=1 k=1

Thus, Z Z Z Z
f (x; y) d Xd Y 6= f (x; y) d Y d X
N N N N
This is because f is not summable: f + (x; y) = 2 2 x for x = y. Thus,
f + (x; y) = g (n) = 2 2 n . On the other hand, f (x; y) = 2 + 2 x for

85
 n 1
x = y + 1. That is, f (x; y) = h (n) = 2 + 2 . Note that
Z 1
X
n
g (n) d = 2 2 =1
N n=1

whereas Z 1
X
n 1
h (n) d = 2+2 = 1
N n=1
Thus, Fubini’s theorem is not contradicted, since the hypothesis of Fubini’s the-
orem are not satis…ed.
Problem 149 Let X = Y = [0; 1], with X being the Lebesge measure and Y
being the counting measure, AX and AY both the Lebesgue -algebras. Show
that D = f(x; y) : x = yg X Y = Z is a measurable set with respect to
= X Y . Show that
Z Z Z
D d =
6 D (x; y) d Y d X
Z X Y

Does it contradict the Fubini Theorem and why not?

Solution 150 Let n 2 N and consider the square [0; 1] [0; 1]. Add k 1
equidistant points (with 0 k n 1) on the diagonal. Using these, form k
rectangles with heights n1 ; n2 ; ::::; 1 and widths n1 . Mathematically, form the sets
k k+1 k k+1
n; n n ; n . Let
n[1
k k+1 k k+1
Dn = ; ;
n n n n
k=0

Then,Dn 2 AZ for each n. Moreover, by construction

\
D= Dn
n2N

so that D 2 AZ . Next, it su¢ces to show that

Z Z Z Z
D (x; y)d Y d X 6= D (x; y)d X d Y
X Y Y X

since Fubini’s theorem implies

Z Z Z
D (x; y)d X d Y = Dd
Y X Z

and, therefore, the above are equal. We are essentially looking at the contrapos-
itive of this statement. We evaluate each side of the above.
Z Z Z
D (x; y)d Y d X = Y (fy 2 Y j y = xg) d X
ZX Y
Z X

= Y (fyg) d X = 1d X = 1 X ([0; 1]) = 1: (1 0) = 1

X X

86
whereas
Z Z
D (x; y)d X d Y
ZY X

= X (fx 2 X j x = yg) d Y
Y
Z Z
= X (fxg) d X = 0d Y =0
Y Y

Again, this does not contradict Fubini’s theorem because the space ([0; 1] ; AY ; Y )
is not -…nite: assume that we can split the uncountable set [0; 1] into a count-
able number of countable sets. Let
1
[
[0; 1] = An
n=1

1
where the collection fAn gn=1 is pairwise disjoint and each An is countable.
Since the countable union of countable sets is countable, we have that j[0; 1]j =
jNj. That is, @0 = c, a contradiction.

Problem 151 Let f; g be two increasing functions on [0; 1], measurable with
respect to m1 . Prove that
Z Z ! Z !
f gdm1 f dm1 f dm1
[0;1] [0;1] [0;1]

R R
Solution 152 If f and g are not summable, then [0;1] f dm1 = [0;1] gdm1 =
R
[0;1]
f gdm1 = 1 and so the inequality holds. If either one of these, say f , is
R R
not summable, then [0;1] f dm1 = 1 =) [0;1] f gdm1 = 1 so that again the
inequality holds. Assume that both are summable. Let Z = [0; 1] [0; 1] and
(x; y) 2 [0; 1] [0; 1]. Note that
Z
(f (x) f (y)) (g (x) g (y)) dm2
Z
Z Z
= (f (x) f (y)) (g (x) g (y)) dm1 (x) dm1 (y)
[0;1] [0;1]
Z Z
= f (x) g (x) f (y) g (x) f (x) g (y) + f (y) g (y) dm1 (x) dm1 (y)
[0;1] [0;1]

87
 2 R R 3
Z f (x) g (x) dm1 (x) f (y) g (x) dm1 (x)
4 [0;1] [0;1]
5 dm1 (y)
= R R
[0;1] [0;1]
f (x) g (y) dm1 (x) +
[0;1]
f (y) g (y) dm 1 (x)
2 R R 3
Z f (x) g (x) dm1 (x) f (y) [0;1] g (x) dm1 (x)
4 [0;1] 5 dm1 (y)
= R R
[0;1] g (y) [0;1] f (x) dm1 (x) + f (y) g (y) [0;1] dm1 (x)
2 R R 3
Z f (x) g (x) dm1 (x) f (y) [0;1] g (x) dm1 (x)
4 [0;1] 5 dm1 (y)
= R
[0;1] g (y) [0;1] f (x) dm1 (x) + (f (y) g (y) (1 0))
Z Z ! Z Z !
= f (x) g (x) dm1 (x) dm1 (y) f (y) g (x) dm1 (x) dm1 (y)
[0;1] [0;1] [0;1] [0;1]
Z Z ! Z
g (y) f (x) dm1 (x) dm1 (y) + (f (y) g (y)) dm1 (y)
[0;1] [0;1] [0;1]
Z !Z Z Z !
= f (x) g (x) dm1 (x) dm1 (y) f (y) g (x) dm1 (x) dm1 (y)
[0;1] [0;1] [0;1] [0;1]
Z Z ! Z
g (y) f (x) dm1 (x) dm1 (y) + (f (y) g (y)) dm1 (y)
[0;1] [0;1] [0;1]

Z ! Z ! Z !
= f (x) g (x) dm1 (x) ((1 0)) f (y) dm1 (y) g (x) dm1 (x)
[0;1] [0;1] [0;1]
Z ! Z ! Z
g (y) dm1 (y) f (x) dm1 (x) + (f (y) g (y)) dm1 (y)
[0;1] [0;1] [0;1]
Z ! Z ! Z !
= f (x) g (x) dm1 (x) f (x) dm1 (x) g (x) dm1 (x)
[0;1] [0;1] [0;1]
Z ! Z ! Z
g (x) dm1 (x) f (x) dm1 (x) + (f (x) g (x)) dm1 (x)
[0;1] [0;1] [0;1]
Z ! Z ! Z !
= 2 f (x) g (x) dm1 (x) 2 f (x) dm1 (x) g (x) dm1 (x)
[0;1] [0;1] [0;1]

Now, if x < y, we have f (x) f (y) 0 and similarly g (x) g (y) 0. Thus,
(f (x) f (y)) (g (x) g (y)) 0. If x = y, then (f (x) f (y)) (g (x) g (y)) =
0. If x > y, then f (x) f (y) 0 and similarly g (x) g (y) 0. Thus, in any
case, (f (x) f (y)) (g (x) g (y)) 0. Thus,
Z
(f (x) f (y)) (g (x) g (y)) 0 =) (f (x) f (y)) (g (x) g (y)) dm2 0
Z

88
That is,
Z ! Z ! Z !
2 f (x) g (x) dm1 (x) 2 f (x) dm1 (x) g (x) dm1 (x) 0
[0;1] [0;1] [0;1]

In other words,
Z Z ! Z !
f gdm1 f dm1 gdm1
[0;1] [0;1] [0;1]

4 Classi…cation of Functions
4.1 Di¤erentiability
Our ultimate goal for di¤erentiability is to come up with nice properties of a
function f such that the fundamental theorem of calculus holds. That is,
Z x
F (x) = f (t) dt =) f = F 0
a

for a nice enough f . Moreover, if f is summable, then we would want F to be

continuous. The natural question we should be asking is under what conditions
is F di¤erentiable? This would help answering when F 0 = f .

Problem 153 If Z
F (x) = f dm1 ,
( 1;x)

is it true that F is continuous?

Problem 154 Let xn ! x. We need to show that F (xn ) ! F (x). Note that
Z Z
f dm1 = ( 1;xn ] f dm1
( 1;xn ) R

Thus, we can let ( 1;xn ] f = fn . Note that fn is summable because f is,

fn ! f pointwise and jfn j f . Thus, by Lebesgue’s Dominated Convergence,
Z Z x
lim F (xn ) = lim ( 1;xn ] f dm1 = f dm1 = F (x)
n!1 n!1 R 1

Before we get into this business, we prove the following useful lemma.

De…nition 155 Let E be a measurable subset of R. Let F be a family of closed,

non-degenerate covers such that 8 > 0 and 8x 2 E, there exists I 2 F such
that m1 (I) < and x 2 I. F is said to be a Vitali cover of E.

89
Theorem 156 (Vitali’s Lemma) Let E be a measurable set such that m1 (E) <
1 and let F be a Vitali Cover of E. Then, we can …nd disjoint intervals
I1 ; :::; In 2 F such that !
[n
m1 En Ik < :
k=1

The proof is very geometric in nature.

Proof. We can …nd an open G with E G such that m1 (G) < 1. We can
always make I’s from F smaller to put them all inside G. That is, I G. This
does not spoil our condition. This is particularly true since G is open.
If there exists a disjoint subfamily fIk : 1 k ng F such that
n
[
E Ik :
k=1

In this case, we trivially have

n
!
[
m1 En Ik = 0:
k=1

If there is no such family, then proceed by induction: take any I1 2 F. We need

to choose the second one. From another subfamily F1 = fI 2 F : I \ I1 = ?g.
This is non-empty by assumption in this case. Now, for every I 2 F in general
and in particular, in F1 , since I G, we must have m1 (I) m1 (G). Since
m1 (G) < 1, we must have s1 = sup fm1 (I) : I 2 F1 g < 1. Take any I2 2 F1
with m1 (I) > s1 =2. On the n-th step, we have I1 ; :::; In chosen. By procedure,
these are pairwise disjoint. Let Fn = fI 2 F : I \ I1 = ?; :::; I \ In = ?g. This
is again possible for the same reason as above with sn = sup fm1 (I) : I 2 Fn g <
1 so that we can choose m1 (In+1 ) > sn =2. The crucial step here relies on the
fact that
n
[ [1
En Ik 5Ik
k=1 k=n+1

where 5Ik is the interval with the same midpoint as Ik but 5 times wider and this
makes easy applications in higher dimensions. To prove this, we need to prove
that 1) m1 (In ) ! 0. This is because the pairwise disjoint family fIn : n 2 Ng
is covered by G. That is,
1
[
Ik G
k=1
so that
1
X
m1 (In ) m1 (G) < 1:
n=1
We also need to prove that
n
[
x 2 En Ik
k=1

90
so that 9I 2 Fn such that x 2 I. However, 9IN such that I \ IN 6= ?.
The overall strategy is to take this x and show that it is covered 5 times an
interval. Assume that this is false. Then, 8N , I \ IN = ?, which means that
I 2 FN for all N but this means that sN m1 (I) =) m1 (IN ) m1 (I) =2,
a contradiction since I is …xed and m1 (I) 6! 0. 3) N n. This holds because
I 2 Fn and so it is disjoint with I1 ; :::; In . Take the …rst N for which IN \I 6= ?.
In particular, I \ I1 = ?; :::; I \ IN 1 = ?. So, I 2 Fn so that, the crucial thing,
m1 (IN ) > m1 (I) =2, which justi…es the 5 scale. If x 2 I, then distance from x
to the middle of IN is m1 (I) + m1 (IN ) =2 which is less than 5=2m1 (IN ) so
that x 2 5IN .
Now, choose n such that
1
X
m1 (Ij ) < =5:
j=n+1

We can do this since the tail goes to zero. Now, the outer measure is monotonic
hence ! 0 1
n
[ [1 1
X
m1 En Ik m1 @ 5Ij A 5m1 (5Ij ) < :
k=1 j=n+1 j=n+1

The fact that Vitali’s Lemma needs closed, nondegenerate intervals is crucial.
Example 157 Let E be any …nite interval and > 0. For each x 2 E, de…ne
Ix = [x; x]. Clearly, [
E Ix
x2E

Moreover, m1 (Ix ) = 0 for each x. In particular, for any > 0 and for any
x 2 E, we have 0 = m1 (Ix ) < and x 2 Ix , by construction. Then, for any n,
n
!
[
m1 En Ik = m1 (E)
k=1

so that for m1 (E), the conclusion of Vitali’s lemma fails.

In fact, the Vitali lemma does not even extend to the case in which the
covering collection consists of non-degenerate general (not necessarily closed)
intervals.
Example 158 Consider the set E = (0; j) for some …xed j 2 N. Then,
m1 (E) = j < 1. For each x 2 E, de…ne intervals B (x; rx ) centered at x
of radius
min (j; j x)
rx =
2
Now, consider the collection F = fB (x; rx ) : x 2 Eg. Then, by construction,
[
E B (x; rx )
x2E

91
so that F is an open cover of E. Moreover, for each x 2 E, by construction,
we can …nd an open ball B (x; rx ) such that x 2 B (x; rx ). Since B (x; rx ) is
an interval, any x 2 B (x; rx ) is a limit point so that for any > 0; we can
thus …nd a neighborhood I of x such that I B (x; rx ) and m1 (I) < . Thus,
the collection of the open balls B (x; rx ) for x 2 E (and such associated neigh-
borhoods I) form a Vitali cover of E. The paranthetical remark is unnecessary
as B (x; rx ) [ I = B (x; rx ). Now, for any …nite number n < j, and disjoint
intervals I1 ; :::; In 2 F with m1 (Ii ) = 21 min (i; i x), we observe that
n
X n
X
0 m1 (Ik ) = min (k; k x) kn
k=1 k=1

By construction, j kn j so that for <j kn, the conclusion of Vitali’s

lemma fails to hold.

Before we embark on the main topic of this section, the following problem
is a useful reminder of why Lebesgue Measure’s is better suited than the Borel
measure.

Problem 159 Show that any union of any collection of closed, bounded, non-
degenerate intervals is measurable.

Solution 160 The …nite and countable case is easy, since we can just ap-
peal to the properties of the -algebra M1 . For the uncountable case, we use
Vitali’s lemma. Let F be a family of closed, bounded nondegenerate inter-
vals I , with as the index for some uncountable indexing set J. Let E =
fI : 9 2 J; I I g. Essentially, we are picking the “small” intervals from
F. Note that it is still true that
[ [
F = I= I
I2F I2E

Moreover, by construction, E is a Vitali Cover of F so that by Vitali’s lemma,

we can …nd disjoint intervals I1 ; :::Ik 2 E such that
n
!
[
m1 F n Ik < :
k=1

In fact, by Regularity of Measure, we can pass to the limit to extend to the

countable case to get fIi : i 2 Ng E such that
1
X
m1 (F ) = m1 (Ik )
k=1

Thus, F is measurable.

Theorem 161 If f : (a; b) ! R is monotone, where the endpoints are allowed

to be in…nite, then f is continuous almost everywhere

92
Proof. WLOG, we can assume that f is increasing, a; b are …nite and we can
consider the interval [a; b] instead of (a; b). We can write
1
[ 1 1
(a; b) = a+ ;b
k k
k=1

For a …xed x0 2 (a; b), de…ne

f x0 = sup ff (x) : x < x0 g and f x+

0 = inf ff (x) : x > x0 g :

Then, f x+ 0 f x0 0 because f is increasing. f is continuous at x0 if

and only if f x+ 0 = f x 0 . In other words, f is not continuous at x0 if
and only if f x+ 0 < f x 0 . Let J (x0 ) = y : f x0 < y < f x0
+
. This
open set consists of points at which the function jumps, hence the choice of
letter. It is obvious that J (x0 ) [f (a) ; f (b)]. Moreover, if x0 6= x00 , then
0
J (x0 ) \ J (x0 ) = ? and so
[
[f (a) ; f (b)] J (x0 )
x0 2[a;b]

is an open cover, for which we can …nd …nitely many subcovers. Therefore, for
each n, the cardinality of the set x0 : m1 (J (x0 )) > n1 is …nite, which implies
that there are only …nitely many discontinuities.
Let f : [a; b] ! R and x 2 (a; b). We can de…ne upper derivative

f (x + t) f (x)
Df (x) = lim sup
h!0+ jtj h t

and lower derivative as

f (x + t) f (x)
Df (x) = lim inf :
h!0+ jtj h t

Clearly, Df (x) Df (x). Also, f 0 (x) < 1 () Df (x) = Df (x) < 1.

Problem 162 Find Df (0) and Df (0) where f (x) is de…ned as

x sin (1=x) x 6= 0
f (x) =
0 x=0

Solution 163 Note that

" # " #
f (t) f (0) t sin 1t
Df (0) = lim sup = lim sup
h!0 0<jtj h t h!0 0<jtj h t
" #
1
= lim sup sin =1
h!0 0<jtj h t

93
This is because sin 1t attains a supremum of 1 at, say, t = 2
h, which satis…es
jtj h.

f (t) f (0) t sin 1t

Df (0) = lim inf = lim inf
h!0 0<jtj h t h!0 0<jtj h t
1
= lim inf sin = 1
h!0 0<jtj h t
Again, this is because sin 1t attains an in…mum of 1 at, say, t = 2
3
h, which
satis…es jtj h.
Recall that if a continuous function f is increasing, then Df 0. This
converse holds true for the upper derivative, as well.
Proof. Let f be a continuous function on [a; b] with Df 0 on (a; b). Let
[c; d] [a; b] with a 6= c d 6= b and > 0. Let
f (x) f (c)
E= x 2 [c; d] :
x c
We need to show that sup E = d for every so that we can let it go to zero and
get the desired result. E is well de…ned since
f (c) + c f (c) + c
so that c 2 E. Since f is continuous and [c; d] is compact, f ([c; d]) is also
compact and closed. Therefore E is closed and so, sup E 2 E. Let sup E = .
If > d, then 62 E, a contradiction. If < d, let
f ( + t) f( )
g (h) = sup
0<t h t
where h 2 (0; d ]. Then,
lim g (h) = Df ( )
h!0+

By hypothesis, Df ( ) 0 so that
lim g (h) 0
h!0+

By de…nition of this limit, 8 0 > 0, 9 : we have that g (h) > 0

whenever
0
h = jhj < . For = , since h 2 (0; d ], 9 1 with < 1 d such that
f( 1) f( )
>
1

That is,
f( 1) + 1 > f( ) + f (c) + c
so that 1 2 E and 1 > sup E, a contradiction.
Thus, = d. In summary, for any [c; d] [a; b] with c d and for any > 0,
f (d) f (c) + (d c). Thus, we can let ! 0 to get f (d) f (c).

94
Problem 164 Show that if the upper and lower derivatives of f are bounded
on (a; b), then there exists a constant C > 0 such that for every x; y 2 [a; b],
jf (x) f (y)j C jx yj

Solution 165 We are given the existence of two constants C1 and C2 such that
" #
f (x + t) f (x)
Df (x) = lim sup = C1 < 1
h!0 0<jtj h t

and
f (x + t) f (x)
jDf (x)j = lim inf = C2 < 1
h!0 0<jtj h t
Let C = 2 max fC1 ; C2 g. Then, Df (x) C=2 and jDf (x)j C=2 so that

Df (x) + Df (x) Df (x) + jDf (x)j C

Let x; y 2 (a; b) with x < y. Then, 9t > 0 such that y x = t. Then,

f (y) f (x) f (x + t) f (x)

=
t t
so that

f (x + t) f (x) f (x + t) f (x) f (x + t) f (x)

inf sup
0<jtj h t t 0<jtj h t

and so
f (y) f (x)
jDf (x)j Df (x) Df (x) + jDf (x)j C
y x

That is,
f (y) f (x)
C
y x

Lemma 166 Let f be an increasing function on [a; b]. Then,

1
m1 x 2 (a; b) : Df (x) > (f (b) f (a))

and
m1 x 2 (a; b) : Df (x) = 1 =0

Proof. If f were continuous on [c; d] [a; b] and di¤erentiable on (c; d), then by
the Mean Value Theorem, we have a point x0 2 (c; d) such that f (d) f (c) =
f 0 (x0 ) (d c). If f 0 (x) on (c; d), then f (d) f (c) (d c) so the result
holds if f is di¤erentiable in (c; d).
In the general case, de…ne E = x 2 (a; b) : Df (x) > and let 0 2 (0; ).
0
Now, de…ne the family F = fI = [c; d] (a; b) : f (d) f (c) (d c)g. Is

95
this non-empty? Yes. It actually covers E as in the Vitali’s Lemma. Indeed,
take x 2 E . Then, Df (x) > so that 8 > 0, 9 : 8h < , 9t 2 [ h; h], giving
us
f (x + t) f (x)
> :
t
0
Take = . Then, 8h < , 9t 2 [ h; h] : f (x + t) f (x) > t 0 ,
giving us I = [x; x + t] or I = [x + t; x]. We then have a disjoint collection
fIk = [ck ; dk ] : 1 k ng F with
0 1
n
[
m1 @E n Ij A <
j=1

Note that
0 1 0 1
n
[ n
[
m1 (E ) m1 @E \ Ij A + m1 @E n Ij A
j=1 j=1
0 1
n
[
m1 @ Ij A +
j=1
n
X
(dj cj ) +
j=1
n
1X
0
(f (dj ) f (cj )) +
j=1
1
0
(f (b) f (a)) +

Since this is true for any 0 , we can let 0 ! ( ! 0) to get what we need.
Using this, we complete the proof as follows:
1
m1 x 2 (a; b) : Df (x) = 1 m1 x 2 (a; b) : Df (x) > (f (b) f (a))

and let ! 1.
And now, for a main theorem

Theorem 167 (Lebesgue’s Theorem) Let f : (a; b) ! R be an increasing

function. Then, f 0 (x) exists for almost every x 2 (a; b)

Proof. The set fx : f 0 (x) does not existg = x : Df (x) = 1 [ x : Df (x) Df (x)
The measure of the …rst set is determined by Lemma 166. For the second
one, note that
[ [
x : Df (x) > Df (x) = x : Df (x) > > > Df (x) = E ; (say)
; 2Q ; 2Q

96
Our goal is to show that m1 (E ; ) = 0. Take an open G such that, for, E ;
G, we have m1 (G) < m1 (E ; ) + . This does not follow from the regularity of
the measure! In fact, the outer measure is not even additive.
Note that X
m1 (E ; ) = inf m1 ((aj ; bj )) :

Now, let F = fI = [c; d] G : f (d) f (c) > (d c)g. F covers E ; , as in

Vitali’s Lemma. To show this, let x 2 E and > 0. Then, Df (x) > Df (x).
That is,

f (x + t) f (x) f (x + t) f (x)
lim sup > lim inf
h!0+ jtj h t h!0+ jtj h t

By de…nition of limit and choice of , we can …nd > 0 : 8h < ; 9t 2 [ h; h]

such that
f (x + t) f (x)
> Df (x)
t
Pick such that Df (x) is a rational number p. Then,

f (x + t) f (x) > tp

We, therefore, have found our [c; d] = [x; x + t] or I = [x + t; x]. Moreover,

by choice of , m1 ([c; d]) = t < . Thus, we can take the disjoint collection
fIk : 1 k ng F so that
0 1
[n
m1 @E ; n Ij A <
j=1

where Ik = [ck ; dk ], giving us

n
X
m1 (E ; ) m1 (E ; \ Ij ) +
j=1
Xn
m1 x 2 (cj ; dj ) : Df (x) > +
j=1
Xn
1
(f (dj ) f (cj )) + (by Lemma 166)
j=1
n
X
=) (f (dj ) f (cj )) m1 (E ; )
j=1

On the other hand,

n
X n
X n
X
(f (dj ) f (cj )) < (dj cj ) = m1 (Ij ) < m1 (G) < m1 (E ; )+
j=1 j=1 j=1

97
giving us m1 (E ; ) m1 (E ; )+ . That is,
+
m1 (E ; ) :

Letting ! 0, we get the desired result.

Corollary 168 Let f : [0; b] ! R be an increasing function. Then,

Z
f 0 dm1 f (b) f (a)
[a;b]

Proof. Set f (x) = f (b) for x b. Then, f 0 exists on E [a; b + 1] and

0
m1 ([a; b + 1] nE) = 0 so that f (x) = Df (x) 0 for x 2 E. For this proof and
onwards, we introduce the following notation: let
f (x + h) f (x)
Dh f (x) =
h
For a …xed h 1, Dh f is measurable on [a; b]. If x 2 E, we have D1=n f (x) !
f 0 (x) (we can only take the limit when we have a countable set, as in a sequence).
Thus, f 0 is measurable on E. Since m1 is complete, by Problem 46, f 0 is
measurable on [a; b]. Therefore,
Z
f 0 dm1
[a;b]

is legal. Now, by Fatou’s Lemma,

Z Z Z
lim D1=n f dm1 lim D1=n f dm1 = f 0 dm1 :
n!1 n!1
[a;b] [a;b] [a;b]

For h = 1=n, we therefore have

Z Z
f (x + h) f (x)
Dh f dm1 = dm1 :
h
[a;b] [a;b]

We can now use linearity of the measure, which applies only when f is summable,
to give
0 1
Z Z Z
f (x + h) f (x) B C
dm1 = 1=h @ f (x + h) dm1 f (x) dm1 A : (8)
h
[a;b] [a;b] [a;b]

First note that

Z Z
g (x + h) dm1 (x) = g (x) dm1 (x)
[a;b] [a+h;b+h]

98
for any measurable g : X ! R and [a; b] X. To see this, let t = x + h. Then,
x = a =) t = a + h and, similarly, x = b =) t = b + h. Replacing the
dummy variable t with x gives us the required result and, therefore, modi…es
our integral in Eq (8) to
0 1
Z Z
B C
1=h @ f (x) dm1 f (x) dm1 A
[a+h;b+h] [a;b]
0 1
Z Z Z
B C
= 1=h @ f (x) dm1 + f (x) dm1 f (x) dm1 A
[a+h;b] [b;b+h] [a;b]
0 1
Z Z Z Z Z
B C
= 1=h @ f (x) dm1 + f (x) dm1 + f (x) dm1 f (x) dm1 f (x) dm1 A
[a+h;b] [b;b+h] [a;a+h] [a;a+h] [a;b]
00 1 1
Z Z Z Z Z
BB C C
= 1=h @@ f (x) dm1 + f (x) dm1 A f (x) dm1 + f (x) dm1 f (x) dm1 A
[a;a+h] [a+h;b] [a;b] [b;b+h] [a;a+h]
0 1
Z Z
B C
= 1=h @ f (x) dm1 f (x) dm1 A
[b;b+h] [a;a+h]

because [a; b + h] = [a; a + h] [ [a + h; b] [ [b; b + h]. Now, we can use f (x) =

f (b) to give
0 1
Z
B C 1
1=h @hf (b) f dm1 A (hf (b) hf (a)) = f (b) f (a) :
h
[a;a+h]

Thus, Z
D1=n f dm1 f (b) f (a)
[a;b]

so that Z
lim D1=n f dm1 f (b) f (a) :
n!1
[a;b]

This inequality is very sharp!

Example 169 Consider

1
x2 sin x2 x 6= 0
f (x) =
0 x=0

99
This function is continuous and …nite but not monotone. Moreover,
Z
f 0 dm1 = 1:
[0;1]

0
Thus, f is not summable so that monotonicity is crucial, even when the function
above is di¤erentiable at x 6= 0.

4.2 Functions of Bounded Variation

Let f : [a; b] ! R be a function and P = fxj : 0 j ng be a partition of the
interval [a; b]. De…ne
n
X
V (f; P ) = jf (xj ) f (xj 1 )j
j=1

Without the absolute value, this is just a telescopic sum. With it, the bounded
variation gives a sense of how oscillating the function is. Some partitions do
that and some don’t. To …nd out how bad these oscillations can get, we take
the supremum over partitions. This is called the total variation T V (f ):
T V (f ) = supV (f; P ) ,
P

which may be denoted by T V f[a;b] since the total variation is de…ned on

a particular domain for a function f . If T V (f ) < 1, then f is of bounded
variation.
Problem 170 T V (f ) = T V ( f )
Solution 171 Let P be a partition of the largest closed subset of domain of f .
Then,
n
X n
X n
X
V ( f; P ) = j( f ) (xj ) ( f ) (xj 1 )j = j f (xj ) + f (xj 1 )j = jf (xj ) f (xj 1 )j = V ( f; P )
j=1 j=1 j=1

Since supremum is taken over all partitions, the result follows.

Example 172 Let f be a monotonic increasing function. Then, T V f[a;b] =
f (b) f (a). This is because f (xj ) f (xj 1 ) > 0 for xj 1 < xj so that
jf (xj ) f (xj 1 )j = f (xj ) f (xj 1 ) and the sum becomes telescoping.
Example 173 Another good family of functions called Lipschitz are de…ned by
a constant C such that, for any x; y 2 [a; b] : jf (x) f (y)j C jx yj. The
absolute value function is one example. Lipschitz functions are all continuous
and di¤erentiable almost everywhere. Note that
n
X n
X
jf (xj ) f (xj 1 )j C jxj xj 1j = C (b a)
j=1 j=1

hence Lipshcitz functions are of bounded variation over a bounded domain.

100
Example 174 We have seen in Problem 164 that if functions with bounded
upper and lower derivatives are Lipschitz. Thus, functions with bounded upper
and lower derivatives are also of bounded variation.
Example 175 Consider the function
x cos 2x 0<x 1
f (x) =
0 x=0
1 1
with partition PN = 0; 2N ; 3N ; :::; 13 ; 21 ; 1 . Computing the variation tells us
that each consecutive point applied on f give us zero and 1 on each subsequent
point for the cosine part. Thus, V (f; PN ) = N1 + ::: + 1. Therefore, the function
f is not of bounded variation.
Proposition 176 T V f[a;b] = T V f[a;c] + T V f[c;b] if a c b
Proof. Let P = fxj : 0 j ng be a partition of the interval [a; b]. Then, if
9k such that xk = c, then by triangular inequality,
n
X k
X n
X
V f[a;b] ; P = jf (xj ) f (xj 1 )j jf (xj ) f (xj 1 )j + jf (xj ) f (xj 1 )j
j=1 j=1 j=k

= V f[a;c] ; P 0 + V f[c;b] ; P 00
for some partition P 0 of [a; c] and P 00 of [c; b]. Since this holds for any P , we have
T V f[a;b] T V f[a;c] + T V f[c;b] . Conversely, note that any paritition P of
[a; c] and P 0 of [c; b] gives rise to a partition P [ P 0 of [a; b]. Thus, V f[a;c] ; P +
V f[c;b] ; P V f[a;b] ; P [ P 0 . It follows that T V f[a;b] T V f[a;c] +
T V f[c;b] .
Theorem 177 (Jordan Decomposition) T V (f ) < 1 () f (x) = f1 (x)
f2 (x) where f1 ; f2 are both increasing
Proof. ( =) ) Let ' (x) = T V f[a;x] . If x > y, then, T V f[a;x] T V f[a;y] =
T V f[y;x] 0. That is, ' (x) ' (y). That is, ' is an increasing func-
tion. Now let (x) = f (x) + ' (x). (x) is also monotone, regardless of
the behaviour of f : if x > y , then consider the partition P = fx; yg. Then,
f (x) f (y) jf (x) f (y)j T V f[y;x] = T V f[a;x] T V f[a;y] . Simi-
larly for f (y) f (x) jf (x) f (y)j T V f[y;x] = T V f[a;x] T V f[a;y] .
In summary, f (y) + T V f[a;y] f (x) + T V f[a;x] so that (y) (x).
That is, we have found ourselves an expression f (x) = (x) ' (x), where
and ' are both increasing.
( (= ) For any partition P = fxj : 0 j ng of the interval [a; b],
n
X n
X
jf (xj ) f (xj 1 )j = jf1 (xj ) f1 (xj 1) + f2 (xj 1) f2 (xj )j
j=1 j=1
Xn n
X
jf1 (xj ) f1 (xj 1 )j + jf2 (xj ) f2 (xj 1 )j
j=1 j=1
= f1 (b) f1 (a) + f2 (b) f2 (a) < 1

101
Corollary 178 T V (f ) < 1 =) f 0 exists almost everywhere

Proof. By Jordan’s Decomposition, f = f1 f2 where f1 and f2 are both

monotone. By Lebesgue’s Theorem, both are di¤erentiable almost everywhere,
as is their di¤erence.

Corollary 179 T V (f ) < 1 =) f 0 is summable

Proof. Follows directly by Jordan Decomposition and Corollary 168.

Problem 180 Let T V (f ) < 1 on [a; b] and de…ne ' (x) = T V f[a;x] . Show
that jf 0 j '0 almost everywhere on [a; b]. Deduce that
Z
jf 0 j dm1 T V (f )
[a;b]

Solution 181 Since f is bounded variation, then f 0 exists almost everywhere.

Moreover, since ' is increasing, '0 exists almost everywhere on (a; b). Let y
be such a point where f 0 (y) and '0 (y) are de…ned. Let x 2 [a; b] with x > y.
Consider the partition P = fx; yg. Then, f (x) f (y) jf (x) f (y)j
T V f[y;x] = T V f[a;x] T V f[a;y] . Since x > y, there exists h > 0 such
that x = h + y. Then,

jf (h + y) f (y)j T V f[a;y+h] T V f[a;y] ' (y + h) ' (y)

=
h h h
Since this holds for any such x, we can let h ! 0 to get jf 0 (y)j '0 (y). Since
y was arbitrary, therefore jf 0 j '0 almost everywhere.
Now, since ' is continuous and increasing on [a; b], as a real-valued function.
Then by Corollary 168,
Z
'0 dm1 ' (b) ' (a)
[a;b]

From jf 0 j '0 , we have

Z Z
0
jf j dm1 '0 dm1 ' (b) ' (a)
[a;b] [a;b]

= T V f[a;b] T V f[a;a] = T V f[a;b] = T V (f )

De…nition 182 Let f : [a; b] ! R be a function. Then, f is said to be ab-

solutely continuous if 8 > 0, 9 > 0 such that, if f(ai ; bi ) : 1 i ng is a
family of disjoint intervals, then
n
X n
X
bj aj < =) jf (bj ) f (aj )j <
j=1 j=1

102
 These intervals may not be a partition of [a; b]. Note that continuity on [a; b]
may imply uniform continuity but uniform continuity does not imply absolute
continuity. Also note that the sum of two absolutely continuous functions is
continuous. Skipping this, we move to
Proposition 183 Let f; g be absolutely continuous on [a; b]. Show that the
function f:g de…ned by (f:g) (x) = f (x) g (x) is absolutely continuous.
Proof. First, we prove that f:g is continuous. Let x0 2 [a; b]. Since f; g are
both continuous, for any > 0, we have 1 , 2 and 3 such that
jx x0 j < 1 =) jf (x) f (x0 )j <
2 (jg (x0 )j + )
jx x0 j < 2 =) jg(x) g(x0 )j <
2 jf (x0 )j
and
jx x0 j < 3 =) jg(x) g(x0 )j < jg(x0 )j +
Take = min ( 1 ; 2 ; 3 ). Then,
j(f:g) (x) (f:g) (x0 )j
= jf (x) g (x) f (x0 ) g (x0 )j = jf (x) g (x) f (x0 ) g (x) + f (x0 ) g (x) f (x0 ) g (x0 )j
jf (x) g (x) f (x0 ) g (x)j + jf (x0 ) g (x) f (x0 ) g (x0 )j
= jf (x) f (x0 )j jg (x)j + jf (x0 )j jg (x) g (x0 )j < jg (x)j + jf (x0 )j
2 (jg (x0 )j + ) 2 jf (x0 )j
< (jg(x0 )j + ) + jf (x0 )j = =2 + =2 =
2 (jg (x0 )j + ) 2 (jf (x0 )j)

Since the function f:g is continuous on a compact domain [a; b], it achives
its maximum and minimum by Extreme Value Theorem. Thus, f:g is bounded.
Let jf:gj M
Now, given > 0, we know that 9 1 ; 2 (distinct from above proof!) such
that f(ai ; bi ) : 1 i ng is a family of disjoint intervals with
n
X n
X
bj aj < 1 =) jf (bj ) f (aj )j < =2nM
j=1 j=1

and
n
X n
X
bj aj < 2 =) jg (bj ) g (aj )j < =2nM
j=1 j=1

Let = min f 1 ; 2 g. Then, for each j,

jf (bj ) g (bj ) f (aj ) g (aj )j
= jf (bj ) g (bj ) f (bj ) g (aj ) + f (bj ) g (aj ) f (aj ) g (aj )j
jf (bj ) g (bj ) f (bj ) g (aj )j + jf (bj ) g (aj ) f (aj ) g (aj )j
= jf (bj )j jg (bj ) g (aj )j + jf (bj ) f (aj )j jg (aj )j
jf (bj )j jg (aj )j M M
< + + = =n
2nM 2nM 2nM 2nM

103
so that
n
X n
jf (bj ) g (bj ) f (aj ) g (aj )j < =
i=1
n
Thus, f:g is absolutely continuous.

Example 184 A good example is a Lipschitz function. Let jf (x) f (y)j

C jx yj. Then,
Xn Xn
jf (bj ) f (aj )j C jbj aj j
j=1 j=1

so that we can take = =C.

Example 185 An absolutely continuous function is Lipschitz if and only if jf 0 j

is bounded. To show this, let jf 0 j c and x; y 2 D (f ), domain of f . Consider
the interval [x; y]. Then, f is absolutely continuous on [x; y] so that
Z y Z y
f 0 dm1 = jf (y) f (x)j jf 0 j dm1 cm1 [x; y] = c jy xj
x x

Converse follows from Example 174.

Theorem 186 Let f be absolutely continuous. Then, f is of bounded variation.

Moreover, f = f1 f2 where f1 ; f2 are both increasing and absolutely continuous.
b a
Proof. Take for = 1. Split [a; b] into N intervals of length < . N .
These intervals are [ak ; bk ]. Then,
N
X
T V (f ) = T V f[ak ;bk ]
k=1

If P = fxk : 0 k ng is a partition of [aj ; bj ], then f(xk ; xk+1 ) : 0 k n 1g

is a disjoint collection and
n
X
jxk 1 xk j = jbj aj j < =) V (f; P ) < 1 =) T V f[aj ;bj ] 1 =) T V (f ) N
k=1

Now, denote f2 (x) = ' (x) = T V f[a;x] and f1 (x) = f (x) + T V f[a;x] =
(x). It su¢ces to show that ' (x) is absolutely continuous. Take = =2.
Assume that
Xn
bk ak <
k=1

where f(ak ; bk ) : 1 k ng is a disjoint collection of intervals. Now, consider

the partition Pk of (ak ; bk ). Then,
n
X n
X
jf (bk ) f (ak )j = V (f; Pk ) < =2
k=1 k=1

104
so that
n
X
T V f[ak ;bk ] =2 <
k=1

Now, by the linearity T V f[ak ;bk ] = T V f[a;bk ] f[a;ak ] = ' (bk ) ' (ak )
so that
X n
j' (bk ) ' (ak )j <
k=1

Thus, in addition to being absolutely continuous, we can always assume that

our function is increasing. However, note that there exists a function of bounded
variation which is not absolutely increasing. For example, any monotone func-
tion.

Problem 187 Let f be continuous on [0; 1] and absolutely continuous on [ ; 1]

for each 0 < < 1. Show that f may not be absolutely continuous on [0; 1], but
it is if f in increasing.

Solution 188 For the …rst part, we want to have a function continuous on
[0; 1], absolutely continuous on [ ; 1] but not absolutely continuous on [0; 1]. Con-
sider the function
x sin x1 x 2 (0; 1]
f (x) =
0 x=0
f is continuous on (0; 1] clearly, as it is the product of two continuous functions
on (0; 1]. f is also continuous at 0: note that

1
jf (x) f (0)j = x sin jxj
x

since sin x1 1. Then,

lim jf (x) f (0)j 0 =) lim f (x) = f (0)

x!0 x!0

Moreover, f is di¤erentiable on [ ; 1] for any > 0 and, by the product rule,

1 1 1 1 1 1
f 0 (x) = sin cos =) jf 0 (x)j = sin cos
x x x x x x
1 1 1 1 1
sin + cos 1+ 1+
x x x x

By Example 185, f is Lipschitz on [ ; 1]. Therefore, by Example 184, f

is absolutely continuous on [ ; 1]. It remains to show that f is not absolutely
continuous on [0; 1]. We can accomplish this by showing that f is not of bounded
variation. Let > 0. Then, 9n such that > 1=n. Consider the partition P

105
 1 1 1
de…ned by xn+1 = 0 n + =2 (n 1) + =2 :::: + =2 1 = x0 . Let
xj = 1= (j + =2) for j 2 f1; :::; ng = J. Then,
8
>
> 1 j 2 J is even
<
1 j 2 J is odd
sin xj =
>
> sin 1 j=0
:
0 j =n+1
so that 8
>
> xj j is even and j 6= 0
<
xj j is odd and j 6= n + 1
f (xj ) =
>
> sin 1 j=0
:
0 j =n+1
Then,
n+1
X
jf (xj ) f (xj 1 )j = jx1 0j + jx2 x3 j + ::: + jsin 1 xn j
j=1
= x1 + x2 + x3 + ::: + xn + sin 1
n
X
= sin 1 + 1= (j + =2)
j=1

The sum on the right is the (displaced) harmonic series, which diverges as n !
1. Hence the function is not of bounded variation.

Problem 189 Let f be continuous and increasing on [0; 1] and absolutely con-
tinuous on [ ; 1] for each 0 < < 1. Show that f is absolutely continuous.

Solution 190 Since f is increasing, given > 0, we must have f ( ) f (0) > 0.
Since f is continuous on [0; 1], 8 > 0 9 0 such that j 0j = < 0 =)
jf ( ) f (0)j = f ( ) f (0) < =2. Let n 2 N and I = f1; 2; :::; ng. Since
=2 > 0 and f is absolutely continuous on [ ; 1], 9 00 > 0 such that, for a family
of disjoint intervals f(ai ; bi ) : i 2 Ig, each a subset of [ ; 1],
n
X n
X
00
jbj aj j < =) jf (bj ) f (aj )j < =2
j=1 j=1

Consider a …nite family f(ai ; bi ) : 1 i mg of disjoint subintervals of [0; 1].

Let I = fi : (ai ; bi ) [ ; 1]g and J = fi : (ai ; bi ) [0; ]g. Note that if 2
(ai ; bi ), then ai < < bi so that f (ai ) < f ( ) < f (bi ). Moreover, jf (bi ) f ( )j =
f (bi ) f ( ) and jf ( ) f (ai )j = f ( ) f (ai ) so that jf (bi ) f ( )j+jf ( ) f (ai )j =
jf (bi ) f (ai )j. Thus, the presence of in an interval makes no di¤erence in
the sum we are looking for, so we can assume that I [ J = f1; 2; ::::; mg. Now,
because f is increasing, we must have
X 0 X
jbj aj j < =) jf (bj ) f (aj )j f( ) f (0) < =2
2
j2J j2J

106
Moreover, WLOG, we can assume that jIj = jIj. Thus, we can re-arrange terms
in J to get I = f1; 2; ::::; n0 g. Now let 2 = max 0 ; 00 . Then, since aj ; bj
for each j 2 J, we must have
X X
jbj aj j < and jbi ai j <
2 2
j2J i2I

so that
m
X m
X
jbj aj j < =) jf (bj ) f (aj )j < + =
i=1 i=1
2 2

That is, for any > 0 we can …nd a such that for any …nite family f(ai ; bi ) : 1 i mg
of disjoint subintervals of [0; 1], the above holds.

If f is absolutely continuous on [a; b], we can always extend the domain to

[a; b + c] for some constant c and de…ne f (x) = f (b) for x > b.

De…nition 191 Let F be a family of functions. F is said to be uniformly

summable if 8 > 0, 9 > 0 such that
Z
m1 (A) < =) jgj dm1 <
A

for all g 2 F

Theorem 192 If f is absolutely continuous on [a; b], then fDh f : 0 h 1g

is uniformly summable. That is, 8 > 0, 9 > 0 such that
Z
m1 (A) < =) jDh f j dm1 < 8h
A

Proof. Observe that we can assume that f is increasing so that jDh f j = Dh f .

Let > 0. We need to …nd a > 0 such that
Z
Dh f dm1 <
A

whenever m1 (A) < . Observe further that we have a G set G with A G and
m1 (GnA) = 0. Therefore, WLOG, we can assume that A is a G set. Thus,
A is the countable intersection of open sets fGi : i 2 Ng, with Gn+1 Gn and
m1 (G1 ) < 1 and Gn is a …nite disjoint union of open intervals. This follows
from the construction in the proof of regularity of Lebesgue Measure because
m1 (Gn ) ! m1 (G) and Z Z
Dh f ! Dh f
Gn G

as n ! 1. Now,
n
[
A= (ai ; bi )
i=1

107
where intervals are pairwise disjoint and
n
X
(bi ai ) <
i=1

To calculate Z bj
Dh f dm1 ;
aj

we observe that
Z h
1
gj (t) = f (bj + t) f (aj + t) = gj dm1
h 0

so that Z Z !
h n
X
1
Dh f dm1 = gi dm1
A h 0 i=1

and for every t,

n
X n
X
gi = f (bj + t) f (aj + t) < ;
i=1 i=1

by absolute continuity of f , where does not depend on t.

Lemma 193 Let ffi : i 2 Ng be a family of real-valued, uniformly summable

functions. If fn ! f almost everywhere, then for every A with m1 (A) < 1, f
is summable on A and Z Z
fn dm1 ! f dm1
A A

Proof. Since the measure of A is …nite, we can split A into a union of N Ak ’s,
where m1 (Ak ) < and comes from the uniform summability of = 1. Then,
by Fatou’s Lemma,
Z Z N Z
X
jf j dm1 lim jfn j dm1 lim jfn j dm1 :
A A k=1 Ak

Note that Z Z
jfn j dm1 1 =) jf j dm1 N
Ak A

so that f is summable on A, f 2 L
Now, WLOG, assume that f = 0 and fn 0 (otherwise take jfn f j). By
Egoro¤’s Theorem, 9 a set A0 A such that m1 (A0 ) < and fn uniformly
converges to 0, fn 0 on AnA0 . Thus, for n N0 ,
Z Z Z Z
fn dm1 = fn dm1 + fn dm1 < fn dm1 +
A AnA0 A0 AnA0

108
so that Z Z
lim fn dm1 lim fn dm1 + lim :
A AnA0

does not depend on n so

Z Z
lim fn dm1 =) lim fn dm1 = 0:
A A

Rb
Theorem 194 If f is absolutely continuous on [a; b], then a
f 0 dm1 = f (b)
f (a)

Proof. Since f is absolutely continuous, it is of bounded variation. By Corol-

lary 178, D1=n f ! f 0 almost everywhere so that
Z b Z b Z b
f 0 dm1 = lim D1=n f dm1 = lim D1=n f dm1
a a n!1 n!1 a

by the previous lemma. Set h = n1 . Then,

Z b " Z a+h !# " Z #
a+h
1 1
lim (Dh f ) dm1 = lim hf (b) f dm1 = lim f (b) f dm1
h!0 a h!0 h a h!0 h a

where the …rst equality follows similar line of reasoning as in the proof of Corol-
lary 168. The proof can then be completed by observing that since f is con-
tinuous,
Z a+h
lim f dm1 = f (a) :
h!0 a

Theorem 195 (Fundamental Theorem R x of Calculus) If f is absolutely con-

tinuous on [a; x], then f (x) = f (a) + a f 0 dm1 for every x.

Theorem 196 If f is absolutely continuous

R x on [a; b] if and only if there exists
a g 2 L1 (m1 ) such that f (x) = f (a) + a gdm1 for every x

Proof. ((=) Take g = f 0

(=)) Let g 2 L1 (m1 ). Then, jgj 2 L1 (m1 ) so that 8 > 0, 9 > 0 such that
Z
m1 (A) < =) gdm1 < :
A

If f(aj ; bj ) : 1 k ng is a disjoint collection of intervals, then

n
X
(bj aj ) <
j=1

109
so that
n
X n Z
X
jf (bj ) f (aj )j = gdm1
j=1 j=1 (aj ;bj )

Xn Z
jgj dm1 <
j=1 (aj ;bj )

Corollary 197 If f is increasing on [a; b] and

Z b
f 0 dm1 = f (b) f (a)
a

then f is absolutely continuous.

Proof. Assume for the sake of contradiction that f is not absolutely continuous.
Then, from Theorem 196, we may assume that
Z x Z b
0
f dm1 < f (x) f (a) but f 0 dm1 f (b) f (x)
a x

Then,
Z x Z b
f 0 dm1 + f 0 dm1 < f (b) f (a)
a x
a contradiction.

Problem 198 Let f be absolutely continuous on [a; b] and f 0 0 almost every-

where. Prove that f is constant.

Solution 199 Let f (a) = c and x 2 [a; b]. Since f is absolutely continuous on
[a; b], we must have Z x
f 0 dm1 = f (x) f (a)
a

However, since f 0 0 a.e., the LHS of the above equation is zero. Hence f (x) =
f (a) = c. Since x was arbitrary, we are done.

Proposition 200 For a summable function f ,

Z y
f 0 a.e. () f dm1 = 0 8x; y 2 [a; b]
x

Proof. Observe that the forward direction (=)) is trivial. For the reverse
direction, Z
f dm1 = 0
I

110
for every I tells us that Z
f dm1 = 0
G
for any open G so that the integral has the same value for any G set and so,
we have the same value for any measurable set and for set of measure 0. Take
E + = fx : f (x) 0g. This set is measurable. Then,
Z
f dm1 = 0
E+

but note that Z Z

f dm1 = f + dm1 = 0
E+ [a;b]

so that f + is zero almost everywhere. Similarly, for E = fx : f (x) 0g. Then,

Z
f dm1 = 0
E

but note that Z Z

f dm1 = f dm1 = 0
E [a;b]

so that f is zero almost everywhere. Since f = f + f , we have that f is

zero almost everywhere.

Theorem 201 If f is a summable function on [a; b], then for almost every
x 2 (a; b), we have Z x
d
f dm1 = f (x)
dx a
Proof. Let Z x
F (x) = f dm1
a

Then, f is summable so that F is absolutely continuous and F (a) = 0. This

tells us that Z x
F (x) = F 0 dm1
a
Subtracting both integrals gives us
Z x
(F 0 f ) dm1 = 0
a

for every x 2 [a; b] so that

Z y Z y
0
(F f ) dm1 = 0 =) (F 0 f ) dm1 = 0 for any x; y
a x

=) F 0 = f almost everywhere by Proposition 200.

111
Problem 202 Let f be continuous on [a; b], di¤erentiable almost everywhere.
Then Z b
f 0 dm1 = f (b) f (a)
a
if and only if
Z b Z b
lim D1=n f dm1 = lim D1=n f dm1
a n!1 n!1 a

Solution 203 (=)) Referring back to the reasoning in Corollary 168, we

know that Z b Z b+h Z a+h !
1
Dh dm1 = f dm1 f dm1
a h b a

so that
Z b Z b
lim D1=n dm1 = lim Dh dm1
n!1 a h!0 a
Z Z !
b+h a+h
1
= lim f dm1 f dm1
h!0 h b a

= f (b) f (a)

Now, since f is di¤erentiable almost everywhere, then

f (x + 1=n) f (x)
lim D1=n f = lim = f0
n!1 n!1 h
Then,
Z b Z b
lim D1=n f dm1 = f 0 dm1 = f (b) f (a)
a n!1 a
Thus,
Z b Z b
lim D1=n dm1 = lim D1=n dm1
n!1 a a n!1

( (= ) Since
Z b Z b
lim D1=n dm1 = f 0 dm1
a n!1 a
and
Z b Z b
lim D1=n dm1 = lim Dh dm1
n!1 a h!0 a
Z Z !
b+h a+h
1
= lim f dm1 f dm1
h!0 h b a

= f (b) f (a)

112
Given that Z Z
b b
lim D1=n dm1 = lim D1=n dm1
a n!1 n!1 a
it follows that
Z b
f 0 dm1 = f (b) f (a)
a

We even the famed product rule!

Problem 204 Let f and g be absolutely continuous on [a; b]. Show that
Z b Z b
f:g 0 dm1 = f (b) g (b) f (a) g (a) f 0 :gdm1
a a

Solution 205 By Proposition 183, f:g is absolutely continuous on [a; b]. Re-
0
call that (f:g) = f g 0 + f 0 g. Also recall that absolute continuity on [a; b] of f
(and g) tells us that
Z b Z b
0
f dm1 = f (b) f (a) and g 0 dm1 = g (b) g (a)
a a

Then,
Z b Z b Z b
0 0 0
f :gdm1 + f:g dm1 = (f:g) dm1 = f (b) g (b) f (a) g (a)
a a a

Rearranging this gives us the desired result.

Problem 206 Let f be strictly increasing and absolutely continuous on [0; 1].
Show that

1. If G (0; 1) is open, then

Z
f 0 dm1 = m1 (f (G))
G

2. Show the same when G is a G set

3. Show the same when m1 (G) = 0.
4. Deduce that the same is true for any measurable set G.

Solution 207 1. Since G is open, G is a disjoint union of half-open intervals

fIi : i 2 Ng, where Ii are allowed to be empty for some i and that m1 (Ii ) 2 [0; 1)
for each i. Thus, we can have
Z Z 1 Z
X
f 0 dm1 = f 0 dm1 = f 0 dm1 (9)
G [Ii i=1 Ii

113
Let end-points of interval Ii be ai and bi , where ai bi (half open degenerate
intervals are empty). Since f is strictly increasing and absolutely continuous on
[0; 1], we must have, for each i,
Z
f 0 dm1 = f (bi ) f (ai ) .
Ii

Since
1
X 1
X
m1 (G) = m1 (Ii ) = bi ai .
i=1 i=1

Moreover, since f is increasing and bi ai , by Problem 30, we must have

m1 (f (Ii )) = f (bi ) f (ai ). It follows that
1
X 1
X
m1 (f (Ii )) = f (bi ) f (ai ) = m1 (f (G)) .
i=1 i=1

By Eq. (9), we are done.

2.
3.
4. Let E be a measurable set. Then, E can be written as the disjoint union
of countable number of open, half-open, degenerate intervals and G sets. Let
1
[ 1
[
E= Ei =) f (E) = f (Ei )
i=1 i=1

Since f is strictly increasing, f (Ei ) is pairwise disjoint. Moreover, if ai and bi

are (not necessarily distinct) endpoints of Ei , then m1 (f (Ei )) = f (bi ) f (ai )
so that
X1 1
X
m1 (f (Ei )) = f (bi ) f (ai ) = m1 (f (E))
i=1 i=1

It follows that
Z 1 Z
X 1
X
0 0
f dm1 = f dm1 = m1 (f (Ei )) = m1 (f (E))
E i=1 Ei i=1

As seen above, the notion of absolute continuity is closely tied the the Fun-
damental Theorem of Calculus for the Lebesgue integral and …lters the idea of
bounded variation, crucially relying on compatibility with arbitrary but disjoint
sums. Can we do better than disjoint sums? Let us agree to call a function
f : [a; b] ! R super absolutely continuous if for any > 0, there ex-
ists a > 0 such that for a …nite family of intervals (not necessarily disjoint)
f(ai ; bi ) : 1 i ng, we have
n
X n
X
(bj aj ) < =) jf (bj ) f (aj )j <
j=1 j=1

114
 Is it true that f is absolutely continuous implies that f is super absolutely
p
continuous? No. Let > 0 and consider f : [0; 1] ! [0; 1] de…ned by x 7 ! x.
Then, by Archimedean’s Principle, 9n such that > n1 . Now, consider the family
f(0; 1=i) : 1 i ng. Then,
n
X n
X p
jf (1=j) f (0)j = 1= j < n 6<
j=1 j=1
p p
for any . However, x is absolutely continuous. To show this, observe that x
is increasing on [0; 1]. Moreover, since f 0 is continuous on (0; 1), it is Riemann
Integrable, so that
Z Z x
1 1 p
p dm1 = p dx = f (x) f (0) = x
[0;x] 2 x 0 2 x

Hence f is absolutely continuous on [0; 1].

4.3 A Pathological Function

When we were building the Cantor set C, at the k-th step, we threw away, in
total, 2k 1 open intervals (this includes those thrown away at k-th step!). Let
Ok = I1k [ I2k [ ::: [ I2kk 1 be the union of these open intervals at the k-th step.
Note that Ok Ok+1 . De…ne ' (x) = m=2k for x 2 Im k
. To show that ' is well-
de…ned, let us explore some of its values. At k = 1, ' (x) = 1=2 in the middle
erased interval and at k = 2, ' (x) = 2=22 in the same interval. Moreover, in the
second erased interval (1=9; 2=9), ' (x) = 1=4 and at (7=9; 8=9), ' (x) = 3=4.
Now, to show that ' is indeed well-de…ned, assume I Ok and I Ok+1 .
k+1 k
Then, I = I2m and I = Im so that in Ok , ' (x) = 2m=2k+1 = m=2k .
' is increasing on each Ok . Let
1
[
O= Ok
k=1

and de…ne ' (0) = 0. Then, ' (x) = sup f' (t) : t < x : t 2 O; x 2 [0; 1] nO = Cg
' is continuous on O. To show this, note that ' is constant on each open
interval. However, ' is not constant on [0; 1]. Moreover, m1 (O) = 1.
Let x0 2 O, then ' (x0 ) is constant around x0 . Assume that x0 2 C where
k k
1 6= x0 6= 0. If k is any number, then x0 lies between Im and Im+1 . The idea
is that if ' is not continuous at x0 , it only has the choice of jumping at x0
k
since ' is increasing on a closed and bounded interval. Take bk 2 Im+1 and
ak 2 Imk
. Then, ak < x0 < bk and that ' (ak ) = 2mk and ' (bk ) = m+1 2k
. Thus,
' (bk ) ' (ak ) = 1=2k . As k increases, the jump vanishes: if ' has a jump at
x0 , we always have, for some > 0, ' (b) ' (a) if a < x0 < b where
' (b) ' x+ 0 and ' (a) ' x0 so that ' (b) ' (a) ' x+ 0 ' x0 , a
contradiction.
' is not absolutely continuous since the fundamental theorem does not hold.

115
Proof. Let O = [0; 1] nC. Let us look at ' (O). Since
1
[
O= Ik
k=1

where Ik are disjoint intervals, not necessarily in increasing order. By construc-

tion, jIk is constant for each k. Note that ' (Ik ) = Ik + c, where ' (Ik ) = c.
Thus, ' translates Ik and so m1 (' (Ik )) = m1 (Ik ). Moreover, f' (Ik ) : k 2 Ng
is a disjoint family since ' is increasing. Thus,
1
X 1
X
m1 (' (O)) = m1 (' (Ik )) = m1 (Ik ) = m1 (O) = 1
k=1 k=1

Note that ' (O) \ ' (C) = ? and that m1 (' ([0; 1])) = 2 so that m1 (' (C)) =
2 m1 (' (O)) = 1.

Corollary 208 9 a subset of a Cantor set A, measurable by completeness of

Lebesgue measure, such that ' (A) 62 M1 .

Proof. We will use the fact that [0; 1] has a non-measurable set. Any subset
E [0; 1] (not necessarily with …nite out measure) with m1 (E) > 0 contains a
non-measurable subset (exercise). Therefore, 9B ' (C) such that B 62 M1 .
Now, de…ne A = ' 1 (B). Then, A C but B = ' (A) 62 M1
As a remark, we state without proof that if A 2 L1 and f is a strictly
increasing and continuous function, then f (A) 2 L1 . To show that, one may
proceed by observing that A may be split into intervals. We can now show that
L1 is not complete.
Proof. We need to show that 9A 2 M1 but A 62 L1 . Take A C. Then,
(A) 62 M1 . Since m1 (C) = 0, A 2 M1 . If A was in L1 , then by previous fact,
(A) 2 L1 M1 , a contradiction. Thus, (R; L1 ; m1 ) is not complete.

5 Di¤erentiation of Measures
In this section, our big goal is to split our master set X into a positive and
negative set, where positive and negative have a di¤erent technical name.

5.1 Signed Measures

De…nition 209 Let A be a -algebra on X, and : A ! [ 1; 1]. Then, v
is called a signed measure or a charge if

1. (?) = 0
2. assumes either +1 or 1 for any A 2 A

116
 3. If fAk : k 2 Ng A is a family of disjoint sets, then
1
[ 1
X
A= Ak =) (A) = (Ak )
k=1 k=1

and the series converges absolutely if j (A)j < 1.

We call a set A 2 A positive if for every measurable E A, we have

(E) 0 and null if for every measurable E A, we have (E) = 0

Example 210 Assume that we have two signed measures, 1 ; 2 . Then, 1 2

is not necessarily subadditive hence not even countably additive. However, it is
additive and gives zero at the empty set.

Example 211 Let f be a -measurable function, then

Z
v (A) = fd
A

is a signed measure. Notice that if we let E + = fx : f (x) 0g and E =

fx : f (x) 0g, then X = E [ E + and
Z Z Z Z
fd = f + d and fd = f d
E+ X E X

Proposition 212 If A is positive, E A is measurable, then E is positive.

Proof. Obvious

Proposition 213 If fAi : i 2 Ng is positive for each i, then so is the union of

the family.

Proof. Let
1
[
E Ak
k=1

and E1 = E \ A1 , E2 = (E \ A2 ) nE1 ; :::; En = (E \ An ) n (E1 [ ::: [ En 1) and

pairwise disjoint. Then, (Ek ) 0 for each k and
1
X
(E) = (Ek ) 0
k=1

Lemma 214 (Hahn) Let 0 < (A) < 1. Then, there exists a positive set
E A with (E) > 0.

117
Proof. If A is positive, we are done, since A has strictly positive measure. If
e A;
A is not positive, then 9A Ae < 0. Let
n o
e
m1 = min m 2 N : 9A A; e <
A 1=m

and take A1 with (A1 ) < 1=m1 . If AnA1 is positive, then we are done.
Otherwise, let
n o
e
m2 = min m 2 N : 9A AnA1 ; e <
A 1=m1 ; (A2 ) < 1=m2

Assume that we have A1 ; :::; An . That is, at the n-th step, if

n
[
An Ak
k=1

is not positive, we have

( n[1
)
e
mn = min m 2 N : 9A An Ak ; e <
A 1=m1 ; (An ) < 1=mn
k=1

Now we have An+1 An (A1 [ A2 [ ::: [ An ) with (An+1 ) < 1=mn . We now
claim that
1
[
E = An An
n=1

is positive. Note that the A is the union of two disjoint sets:

1 1
!
[ [
A=E[ An =) (A) = (E) + An
n=1 n=1

(A) < 1, then the right hand side is also …nite. In particular, for the in…nite
union is …nite. Thus,
1
X 1
X 1
X 1
j (An )j < 1 =) 1< (An ) <
n=1 n=1 n=1
mn
X1
1
=) < 1 =) lim mn = 1
m
n=1 n
n!1

Now let B E. Our goal is to show that (B) 0. Note that we can have an
N 2 N such that
1
[ [N
B An An An An
n=1 n=1

Then, because mN was the minimum, we have (B) 1= (mN 1). This is
true for every N so we can apply limit and get (B) 0.

118
 Now, …nally
1
!
[
(A) 0 =) (AnE) = (A) (E) = An <0
n=1

so that (E) > 0.

Essentially, for any given set X, we can split X into two disjoint sets, one
for positive and the other for negative

Theorem 215 (Hahn Decomposition) Let be a signed measure on (X; A).

Then, 9A X with A positive and 9B X with B negative such that X = A[B
and A \ B = ?

Proof. WLOG, we can assume that is never 1. If (E) 0 for every E X,

then set A = ? and B = X. If there exists an E with (E) > 0, then, by Hahn’s
Lemma, we can have a positive subset. Let = sup f (E) : E is positiveg.
Note that 0. By de…nition of supremum, there exists a sequence of sets
fAk : k 2 Ng such that (Ak ) ! as k ! 1.
Set
[1
A= Ak
k=1

since union of positive sets is positive, A is positive. Thus, (A). On

the other hand, AnAk A so that (AnAk ) 0. Since both sets have …nite
measure, we have (A) (Ak ) for each k. Taking limit on both sides, we get
(A). Thus, = (A).
Now let B = XnA. We need to show that B is negative. If not, then there
exists an E B with (E) > 0. Thus, there is a positive subset E e E such
that 0 < Ee . Then, A [ E e is also positive. Moreover, (A) + Ee
(A) = , a contradiction to the fact that is the supremum. Thus, B is
indeed negative.
We then get the following useful corollary, which we state without proof.

Corollary 216 (Jordan Decomposition) If is a signed measure, then there

exist two measures + and (not signed!) such that = + and there
exist sets A and B such that A [ B = X, A \ B = ?, + (B) = (A) = 0.

Problem 217 This Jordan Decomposition in unique

Solution 218 Assume that 9 measures + and (not signed!), positive A 2 A

and negative B 2 A such that X = A [ B and A \ B = ? such that = +
and + (B) = (A) = 0. Let = + be another Jordan decomposition of
with positive and negative A0 and B 0 . We …rst need to show that consideration
of di¤erent sets is immaterial.
Consider A4A0 = A[A0 n (A \ A0 ). Let E (A [ A0 ) n (A \ A0 ) be measur-
able. In particular, E A [ A0 . Then, (E) 0. Since (A [ A0 ) n (A \ A0 ) =

119
(A [ A0 ) \ (Ac [ A0c ) = (A [ A0 ) \ (B [ B 0 ), then, in particular, E B [ B0.
0
Hence (E) 0 so that (E) = 0. Thus, A 4 A is a null-set with respect to
. Similarly, B 4 B 0 is a null set with respect to .
Now let us focus on = + . By above, we can work with same A and B.
Let E 2 A. Then, E\A A so that (E \ A) = + (E \ A) (E \ A). Since
E \ A is positive, (E \ A) = 0. Thus, (E \ A) = + (E \ A). Now, since
+
is a measure and A is + -measurable, for any set E, + (E) = + (E \ A) +
+
(E \ Ac ) = + (E \ A)+ + (E \ B). Since E \B is negative, + (E \ B) =
0.
Thus, + (E) = + (E \ A) = (E \ A). Since (A 4 A0 ) = 0, we must
have (E \ A) = (E \ A0 ). Also since = + , we can have (E \ A0 ) =
+ 0 + 0
( ) (E \ A ) = (E \ A ) (E \ A ). Since E \ A0 is positive,
0
0 0 +
(E \ A ) = 0. Hence (E \ A ) = (E \ A0 ).
+ +
Thus far, we have that (E) = (E \ A) = + (E \ A0 ). Now, again,
+ + 0 + 0c +
(E) = (E \ A )+ (E \ A ) = (E \ A0 )+ + (E \ B 0 ) = + (E \ A0 )+
0.
In summary, for any E, + (E) = + (E). Hence + = + . Combining
this with the assumption that + = + gives us = . Thus, the
Jordan decomposition is unique.

Problem 219 For a signed measure, de…ne

jvj (E) = v + (E) + v (E) :

Show that jvj (E) is equal to

n
X
sup jv (Ek )j
k=1

where supremum is taken over all …nite disjoint families fEk : 1 k ng of

measurable subsets of X.

Solution 220 Let E 2 A. Then, v (E) = v + (E) v (E) v + (E) v + (E) +

v (E) = jvj (E). Conversely, v (E) = v (E) v + (E) v (E) v + (E) +
v (E) = jvj (E). That is, jvj (E) v (E). Together, these imply that jv (E)j
jvj (E) for any measurable E. Let fEk : 1 k ng be a disjoint family of E.
Then, note that
n
X n
X n
X
+ +
jvj (E) = v (E) + v (E) = v (Ek ) + v (Ek ) = jvj (Ek )
k=1 k=1 k=1

Moreover, we have that jv (Ek )j jvj (Ek ) 8k. Thus,

n
X n
X
jv (Ek )j jvj (Ek ) = jvj (E)
k=1 k=1

120
Since this holds for an arbitrary decomposition of E, we must have
n
X
sup jv (Ek )j jvj (E)
k=1

On the other hand, by Hahn Decomposition, 9 positive A and negative B such

that X = A [ B. Then, jvj (X) = v + (A) + v (B) = jv + (A)j + jv (B)j
with A \ B = ?. Now, note that for any measurable E can be partitioned
using these positive and negative sets (which is a candidate for the supremum)
since E = E \ X = E \ (A [ B) = (E \ A) [ (E \ B) and that jvj (E) =
v + (E) + v (E) = v + (E \ A) + v (E \ A) + v + (E \ B) + v (E \ B)
= v + (E \ A)+v (E \ B) v (E \ A)+v (E \ B) jv (E \ A)j+jv (E \ B)j.
Thus, the other side of the inequality holds.

Assuming that f is positive and measurable and that

Z
fd = 0
E

whenever (E) = 0, then for any measurable set A, we can de…ne the measure
Z
(A) = fd
A

Thus, if and are de…ned on (X; A), we say that is absolutely continuous
with respect to if (E) = 0 =) (E) = 0. The notation in this case is
.

Theorem 221 Assume that (X) < 1. Then, () 8 > 0, 9 > 0 :

8E 2 A; (E) < =) (E) <

Proof. ((=) Fix any ; take E with (E) = 0. Then, (E) = 0 < =)
(E) < . Since this is true for any , then (E) = 0.
( =) ) For the sake of contradiction, assume that the conclusion fails. Then,
9 0 > 0 : 8 > 0; 9E 2 A, which depends on , (E) < but (E) .
Let En be subsets of E such that (En ) = 1=2n and let

[ 1
X 1 1
AN = Ek =) (AN ) k
= N 1
2 2
k N k=N

Now let
1
\
A= AN
N =1

Then, (A) (AN ) 1=2N 1 for every N so that (A) = 0. However,

(A) > 0: to see this, note that (AN ) (EN ) 0 . Since (A1 ) (X) <
1, then (A) = lim (AN ) 0 . That is, (A) > 0.
N !1

121
Example 222 Take = m1 and = 0 measure. The latter works like a charge
at 0. Then, it is not true that because (f0g) = 0 but (f0g) = 1.

Theorem 223 (Radon-Nikodym) If ; are both -…nite measures on the

same -algebra A and , then 9 a nonnegative,
R measurable (not necessarily
summable) function f such that (E) = E f d .

Proof. Assume ; are both …nite. Also assume that (X) > 0, for otherwise
f 0 works and gives us the trivial case. For t > 0, denote t = t . t is
a signed measure but not necessarily nonnegative. We can have sets Pt and Nt
which are, respectively, positive for t and negative for t with Pt \ Nt = ? and
Pt [ Nt = X by the Hahn Decomposition. We need to …x t. We claim that 9t
with (Pt ) > 0. Assume that is not true. Then, for 8E 2 A, we have t (E) 0
since t (E) = t (E \ Pt ) + t (E \ Ptc ) = t (E \ Pt ) + t (E \ Nt )
= t (E \ Pt ) = ( t ) (E \ Pt ) = (E \ Pt ) t (E \ Pt ) 0. Thus,
(X) t (X) for every t. But then, (X) = 0, a contradiction.
Now, de…ne the family
Z
F = f 0: fd (E) 8E 2 A
E

This set is non-empty since f 0 2 F. Is there a non-trivial function? De…ne

f = t: Pt where (Pt ) > 0. Then,
Z
f d = t (E \ Pt )
E

Since t (E \ Pt ) 0, we must have t (E \ Pt ) (E \ Pt ) (E)

We need to …nd the “biggest” possible f . Let us consider the “average”.
Z
M = sup fd
f 2F X

Then, M > 0. Our next goal is to show that the supremum is actually obtained
by showing that 9f 2 F with
Z
M= fd
X

Observe that if f; g 2 F, then h = max (f; g) 2 F. To show this, if E 2 A, take

E1 = fx 2 E : f gg and E2 = fx 2 E : f < gg. Then,
Z Z Z Z Z
hd = hd + hd = fd + gd (E1 ) + (E2 ) (E)
E E1 E2 E1 E2

Now let ffn : n 2 Ng F such that

Z
fn d ! M
X

122
Take gn = max (f1 ; :::; fn ). Then, gn is an increasing sequence and g (x) =
lim gn (x) is well-de…ned. By Monotone Convergence Theorem,
n!1
Z Z
gd = lim gn d
E n!1 E

Since each gn 2 F, then, for each n, the left hand side is less than the limit of
(E). That is, (E) itself. On the other hand,
Z Z Z
gd gn d ! M =) gd = M
X X X

Now assume that for the measure space (X; A), and are -…nite on X.
Then, 9 fXn : n 2 Ng A such that
1
[
X= Xn
n=1

where m1 (Xn ) < 1 for each n. WLOG, we can assume that Xn Xn+1 .
Restricting on the subspace -algebra AXn = fE \ Xn : E 2 Ag gives us a
…nite measures jA X := n and jA X := n on Xn for each n since m1 (Xn ) <
n n
1.
This enables us to use the previous case of and being …nite measures;
for each n, we can …nd a unique function fn measurable with respect to AXn
such that Z
n (E) = fn d n for all E 2 AXn
E
Now, if n m, then Xn Xm so that AXn AXm . Thus, n is the restriction
of m such that n = m for E 2 AXn . That is, n can be extended to m . It
follows that
Z Z
for all E 2 AXn , n (E) = fn d n = fm d m = m (E)
E E

so that n is the restriction of m . Since fn is unique, fn = fm for all Xn nA

where n (A) = 0. Therefore, if F 2 A, then
Z
n (F \ Xn ) = fn d n for all F 2 A
F

Note that F \ Xn F \ Xn+1 since Xn Xn+1 and that

1
[ 1
[
F =F \X =F \ Xn = (F \ Xn )
n=1 n=1

Thus, by continuity from above,

1
!
[
(F ) = (F \ Xn ) = lim n (F \ Xn )
n!1
i=1

123
 Let gn = max ff1 ; :::; fn g. Then, fgn : n 2 Ng is a monotone increasing se-
quence of sets in X. Let g = lim gn (we don’t necessarily need g to be sum-
n!1
mable). Since for each n, gn is -measurable, then g is also -measurable.
Moreover, by Monotone Convergence Theorem,
Z Z Z
lim gn d n = lim gn d n = gd
F \Xn n!1 n!1 F \Xn F

so that there exists a -measurable function g such that, for any F 2 A

Z
(F ) = gd
F

Problem 224 (a) Characterize the measure spaces (X; A; ) for which the count-
ing measure M is absolutely continuous with respect to ; and (b) characterize
the measure spaces for which, given x 2 X, the measure x is absolutely contin-
uous with respect to .

Solution 225 (a) We know that for any E 2 A, (E) = 0 =) M (E) = 0. In

other words, M (E) 6= 0 =) (E) 6= 0. If M (E) 6= 0, then E is non-empty.
Since M is de…ned on 2X , we must have de…ned on A = 2X with (E) 6= 0
(that is, (E) > 0) for any non-empty set E. Thus the spaces (X; A; ) we are
looking for are precisely those for which there are no non-empty sets of measure
0.
(b) Let x 2 X be …xed. Then, for any E 2 A, (E) = 0 =) x (E) = 0.
In other words, x (E) 6= 0 =) (E) 6= 0. That is, x (E) = 1 =) (E) > 0.
That is, x 2 E =) (E) > 0. Thus, given x 2 X, the spaces (X; A; ) we are
looking for are precisely those for which there are no non-empty sets containing
x of measure 0.

124