Topics in Measure Theory

S. Kumaresan
School of Math. and Stat.
University of Hyderabad
Hyderabad 500046
[email protected]

1 Product Measure and Fubini Theorem

Let (X, A, µ) and (Y, B, ν) be two measure spaces. We wish to find a ‘natural’ σ-algebra C
on X × Y and a measure α on C. The obvious requirements on C and α is that if A ∈ A and
B ∈ B, then A × B ∈ C and that α(A × B) = µ(A) × ν(B). Since we “know” how to integrate
with respect to µ and ν, the next requirement is that if f : (X × Y, C) → [0, ∞] is measurable,
then Z Z Z  Z Z 
f (x, y) dα = f (x, y) dµ dν = f (x, y) dν dµ. (1)
X×Y X Y Y X
Note that if we take R := A×B with A ∈ A and B ∈ B, then f := χR ≡ 1R , the characteristic
(or the indicator) function of R, satisfies (1). Note also that for (1) to make sense, we need to
ensure that the functions x → f (x, y) for y fixed and y → f (x, y) for x fixed are measurable on
the respective spaces. One may hope to extend the result in (1) to finite linear combinations of
such 1R ’s and use monotone convergence theorem to extend it to all non-negative measurable
functions.
So, one way of finding C and α could be to use (1). Let R denote the class of rectangles
A × B with A ∈ A and B ∈ B. Note that if we wish to have R ⊂ C, we need to ensure that
(X × Y ) \ (A × B) ∈ C. It is eay to verify that the complement of such a rectangle can be
written as
(X × Y ) \ (A × B) = ((X \ A) × B) ∪ (X × (Y \ B) ,
that is, the complement of a rectangle is a finite disjoint union of rectangles. This suggests
that we should consider the collection S of all such finite disjoint unions of rectangles. It
turns out S is an algebra of sets on X × Y according to the next definition.

Definition 1. Let A ⊂ P (X) be a class of subsets of a set X. We say that A is an algebra

of sets if (1) X ∈ A, (2) if A ∈ A, then X \ A ∈ A and (3) whenever A, B ∈ A, their union
A ∪ B ∈ A.

Since we plan to use monotone convergence theorem, it is necessary that the σ-algebra C
shoudl contain all subsets which are obtainable by taking limts of monotone sequences in S.
This motivates the next definition.

1
Definition 2. Let M ⊂ P (X) be a class of subsets of a set X. We say that A is a monotone
class of sets (i) if An ∈ M with An ↑ A, then A ∈ M and (ii) if Bn ∈ M with Bn ↓ B, then
B ∈ M.

The crucial tool in this article is the following lemma.

Lemma 3. If A is an algebra of subsets of X and M is the smallest monotone class containing

A, then M is a σ-algebra.

Proof. We claim that M is closed under complementation. Let E ∈ M . Consider the class
M0 := {E ∈ M : X \ E ∈ M}. Since A is an algebra, X \ A ∈ A ⊂ M. Hence A ⊂ M0 . It
is easy to check that M0 is a monotone class. Hence we conclude that M ⊂ M0 . Hence the
claim follows.
For A ⊂ X, we let MA := {E ∈ M : E ∩ A ∈ M}. Since A is an algebra, it follows
that, if A ∈ A, then A ⊂ MA . It is again an easy observation that MA is a monotone class.
Consequently, M ⊂ MA . That is, if A ∈ A and E ∈ M, then E ∩ A ∈ M.
Let E ∈ M be arbitrary. Consider ME . By the last paragraph, A ⊂ ME . It is easy
to see that ME is a monotone class and hence M ⊂ ME . That is, finite intersections of
elements of M are again in M. The same is true of union of finite number of elements of
M.

Let M ⊂ P (X ×Y ) be the smallest monotone class containing the algebra S of the class of
all subsets of X × Y which can be written as disjoint union of a finite number of ‘measurable
rectangles’ R = A × V where A ∈ A and B ∈ B. By the monotone class lemma, M is a
σ-algebra on X × Y .

Lemma 4. Assume that µ(X) < ∞ and ν(Y ) < ∞. Let

 Z Z  Z Z  
C := F ⊂ X × Y : 1F (x, y) dµ dν = 1F (x, y) dν dµ .
X Y Y X

Then M ⊂ C.

Proof. The steps are the same. We show that A ⊂ C and then C is a monotone class. Where
do we need the finite measure hypothesis? To deal with the decreasing sequences!

Theorem 5 (Existence of Product Measure). Let (X, A, µ) and (Y, B, ν) be two σ-finite
measure spaces. Then there exists a unique measure α on M with the property that α(A×B) =
µ(A)ν(B) for all A ∈ A and B ∈ B.

Proof. First assume that µ(X) and ν(Y ) are finite. The last lemma shows that for E ∈ M,
the definition
Z Z  Z Z 
α(E) := 1F (x, y) dµ dν = 1F (x, y) dν dµ.
X Y Y X

is well-defined. Also, α is finitely additive on M due to the linearity of the integrals. By

the monotone converegence theorem, α is also countably additive. If β is any measure on M

2
with the required property, then the class of subsets E for which α(E) = α(E) is a monotone
class which contain S. Hence α = β on M.
In the general case, write X = ∪n An and Y = ∪n Bn where the unions are disjoint
µ(An ) < ∞ and ν(Bn ) < ∞ for n ∈ N. Given any E ∈ M, let Emn := E ∩ (Am × Bn ). For
each m, n, thanks to the finite case, we have
Z Z Z Z
1Emn dν dµ = 1Emn dµ dν.
X Y Y X

We can sum these terms over m and n (in any order), to get by countable addivity and
monotone convergence theorem that
Z Z  Z Z 
1F (x, y) dµ dν = 1F (x, y) dν dµ, for E ∈ M.
X Y Y X

If we set α(E) to be one of these integrals, then by the linearity of the integrals, α is finitely
additive. It is countably additive by the monotone convergence theorem. It also has the
required property on R.
If β is any other meaure on M with this property, then writing E = ∪m,n Emn and applying
the uniquenessPpart of the finite
P measure case to conclude α(Emn ) = β(Emn ) for all m, n.
Then α(E) = m,n α(Emn ) = m,n β(Emn ) = β(E) by counatble additivity.

Remark 6. The assumption σ-finiteness in the last theorem is essential. For, take X = Y =
[0, 1] but µ to be counting measure and ν to be the Lebesgue measure. Then the diagonal
D := {(x, x) : x ∈ [0, 1]} is measurable in the product space with
Z Z Z Z
1D dν dµ = 0 whereas 1D dµ dν = 1.
X Y Y X

Note that µ is not σ-finite.

Theorem 7 (Fubini-Tonelli). Let (X, A, µ) and (Y, B, ν) be two σ-finite measure spaces. Let
f : X × Y → [0, ∞] measurable with respect to product σ-algebra M. Then (1) holds true.
The same conclusion is true if we assume f ∈ L1 (X × Y, M, α). Here f (x, y) dµ(x) is
R

defined for almost all y etc.

Proof. For nonnegative simple functions, the result follows from Theorem 5 and the additivity
of the integrals. For nonnegative measurable functions, it follows from MCT.
For f ∈ L1 , apply the result ot each of f + and f − .

The following extension is quite useful in practice.

Theorem 8. Let f : X × Y → R be M-measurable. Assume that one of the iterated integrals
of |f | exists. Then
(i) fR∈ L1 (X × Y, C, α).
(ii) Y f (x, y) dν(y) is defined for all most all x. If it is defined to be 0 on the set of
measure
R zero, then it is measurable and lies in L1 (X, A, µ). The analogous result is also true
for X f (x, y) dµ(x). (iii) The double integral of f exists and is equalt to the each of the
iterated integrals.

3
Proof. Apply Fubini-Tonelli to each of f + and f − .

We now give two typical applications of these results.

Example 9. It is easy to see that if f, f ∈ Cc (R), then the folwoing definition makes sense:
Z
f ∗ g(x) := f (x − y)g(y) dy.
R

It truns out that if f, g ∈ L1 (R) (with respect to the Lebesgue measure), then also it
makes
R R sense almost everywhere.

The only way to prove this is to show that the integral
R R |f (x − y)g(y)| dy dx exists and appeal to Fubini theorem. The details are left to the
reader.

Example 10. Let P : Rn → R be a polynomial function. Let E := {x ∈ Rn : P (x) =

0}. Then E is of Lebegue measure zero in Rn . Apply Fubini to 1E to the n-fold iterated
R fixed, then the polynomial x1 7→
integral and observe that if x2 = a2 , . . . , xn + an are held
P (x1 , a2 . . . , an ) has only finitely many zeros and hence R 1E (x1 , a2 , . . . , an ) dx1 = 0.

2 Radon-Nikodym and Lebesgue Decompostion Theorems

Theorem 11. Let µ and ν be two finite measure on a measurable space (X, B). Then
(i) there exist measures ν1 and ν2 such that ν = ν1 + ν2 where ν1  mu and ν2 ⊥ µ.R
(ii) There exists a nonnegative measurable function ϕ on (X, B) such that ν1 (E) = X ϕ dµ
for E ∈ B.

Proof. Let σ := µ + ν. Consider T : L2 (X, B, σ) → R given by T f := XR f dµ. Then T is a

R

bounded linear functional. Hence there exists g ∈ L2 (σ) such that T f = X f g dµ. It follows
Z Z
f (1 − g) dσ = f dµ, for all f ∈ L2 (σ). (2)

g ≥ 0 almsot everywhere
R σ: If false, take f = 1E where E := {g < 0} in (2) to arrive at
the contradiction σ(E) + E g dσ = µ(E).
g ≤ 1 almsot everywhere σ: If false, take f = 1E where E := {g > 1} in (2) to arrive at
the contradiction µ(E) < 0.
Let A := {0 ≤ g < 1} and B := {g = 1}.
Take f = 1B in (2) to deduce at µ(B) = 0. We define

νA (E) := ν(E ∩ A) and νB (E) := ν(E ∩ B), for E ∈ B.

Then µ(B) = 0 and νB (X \ B) = νB (A) = ν(A ∩ B) = ν(∅) = 0. Hence µ ⊥ νB .

We claim: νA  µ: If µ(E) = 0, we want to prove that νA R(E) = µ(E ∩ A) = 0. So,
assume µ(E) = and E ⊂ A. Taking f = 1E in (2), we get E (1 − g) dσ = 0. Hence
0 = µ(E) = σ(E) = µ(E) + ν(E) so that ν(E) = ν(E ∩ A) = νA (E). Hence the claim.

4
 The following proof of Radon-Nikodym theorem is based on the original argument by John
von Neumann. We suppose that µ and ν are real, nonnegative, and finite. The extension to
the σ-finite case is a standard exercise, as is µ-a.e. uniqueness of Radon-Nikodym derivative.
Having done this, the thesis also holds for signed and complex-valued measures.
Let (X, F) be a measurable space and let µ, ν : F → [0, R] two finite measures on X such
that ν(A) = 0 for every A ∈ F such that µ(A) = 0. Then σ = µ + ν is a finite measure on X
such that σ(A) = 0 if and only if µ(A) = 0.
Consider the linear functional T : L2 (X, F, σ) → R defined by
Z
Tu = u dµ ∀u ∈ L2 (X, F, σ) . (3)
X

T is well-defined because µ is finite and dominated by σ, so that L2 (X, F, σ)p⊆ L2 (X, F, µ) ⊆

L1 (X, F, µ); it is also linear and bounded because |T u| ≤ kukL2 (X,F ,σ) · σ(X). By Riesz
representation theorem, there exists g ∈ L2 (X, F, σ) such that
Z Z
Tu = u dµ = u · g dσ (4)
X X

for every u ∈ L2 (X, F, σ). Then µ(A) = A g dσ for every A ∈ F, so that 0 < g ≤ 1
R

µ- and σ-a.e. (Consider the former with A = {x | g(x) ≤ 0} or A = {x | g(x) > 1}.)
Moreover, the second equality in (4) holds when u = χA for A ∈ F, thus also when u is a
simple measurable function by linearity of integral, and finally when u is a (µ- and σ-a.e.)
nonnegative F-measurable function because of the monotone convergence theorem.
1
Now, 1/g is F-measurable and nonnegative µ- and σ-a.e.; moreover, · g = 1 σ- and
g
µ-a.e. Thus, for every A ∈ F,
Z Z
1
dµ = dσ = σ(A) (5)
A g A

1
Since σ is finite, 1/g ∈ L1 (X, F, µ), and so is f =
− 1. Then for every A ∈ F
g
Z   Z
1
ν(A) = σ(A) − µ(A) = − 1 dµ = f dµ .
A g A

5
3 Complex Measures

Lemma 12. Let T := {zk ∈ C : 1 ≤ k ≤ N } be a finite subset of complex numbers. Then

there exists a subset S ⊂ T such that
N
X 1X
| z| ≥ |zk |. (6)
π
z∈S k=1

Proof. For t P∈ [0, 2π], we let St := {z ∈ T : −π/2 ≤ t − arg (z) ≤ π/2}. (Draw a picture.)
Let f (t) := | z∈St z|. We plan to show that there exists θ ∈ [0, 2π] such that

N
1X
f (θ) ≥ |zk |.
π
k=1

Note that f is piecewise constant on [0, 2π]. Observe that

X X
| z| = |e−it z|
z∈St z∈St
X
= | e−it z|
z∈St
!
X
≥ |Re e−it z |
z∈St
X
= | Re (e−it z)|
z∈St
X
= | |z| cos(t − arg (z))|
z∈St
X
= |z| cos(t − arg (z)).
z∈St

Hence we obtain
Z 2π Z 2π X
f (t) dt = | z| dt
0 0 z∈St
Z 2π X
≥ |z| cos(t − arg (z)) dt
0 z∈St
N Z
X arg (zk )+π/2
= |z| cos(t − arg (z)) dt
k=1 arg (zk )−π/2
N
X
= 2|zk |.
k=1

R π/2
The last is due to the fact that −π/2 cos t dt = 2.

6
 By the extreme value theorem, there exists θ suchthat f (θ) ≥ f (t) for t ∈ [0, 2π]. It
follows that
Z 2π N
X
2πf (θ) ≥ f (t) dt ≥ 2 |zk |.
0 k=1