0% found this document useful (0 votes)
137 views17 pages

Measures: Measure Spaces

The document discusses measures and measure spaces. It provides definitions of measures, measure spaces, and terminology related to measures. Examples of measures are given including Dirac measures, counting measures, and an example of a finitely-additive set function that is not a measure. Properties of measures are stated and proved, including finite additivity, monotonicity, continuity, and subadditivity.

Uploaded by

nothard
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
137 views17 pages

Measures: Measure Spaces

The document discusses measures and measure spaces. It provides definitions of measures, measure spaces, and terminology related to measures. Examples of measures are given including Dirac measures, counting measures, and an example of a finitely-additive set function that is not a measure. Properties of measures are stated and proved, including finite additivity, monotonicity, continuity, and subadditivity.

Uploaded by

nothard
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 17

Lecture 2: Measures 1 of 17

Course: Theory of Probability I


Term: Fall 2013
Instructor: Gordan Zitkovic

Lecture 2
Measures

Measure spaces

Definition 2.1 (Measure). Let (S, S) be a measurable space. A map-


ping µ : S → [0, ∞] is called a (positive) measure if

1. µ(∅) = 0, and

2. µ(∪n An ) = ∑n∈N µ( An ), for all pairwise disjoint { An }n∈N in S .

A triple (S, S , µ) consisting of a non-empty set, a σ-algebra S on it and


a measure µ on S is called a measure space.

Remark 2.2.

1. A mapping whose domain is some nonempty set A of subsets of


some set S is sometimes called a set function.

2. If the requirement 2. in the definition of the measure is weakened


so that it is only required that µ( A1 ∪ · · · ∪ An ) = µ( A1 ) + · · · +
µ( An ), for n ∈ N, and pairwise disjoint A1 , . . . , An , we say that the
mapping µ is a finitely-additive measure. If we want to stress that
a mapping µ satisfies the original requirement 2. for sequences of
sets, we say that µ is σ-additive (countably additive).

Definition 2.3 (Terminology). A measure µ on the measurable space


(S, S) is called

1. a probability measure, if µ(S) = 1,

2. a finite measure, if µ(S) < ∞,

3. a σ-finite measure, if there exists a sequence { An }n∈N in S such


that ∪n An = S and µ( An ) < ∞,

4. diffuse or atom-free, if µ({ x }) = 0, whenever x ∈ S and { x } ∈ S .

A set N ∈ S is said to be null if µ( N ) = 0.

Last Updated: September 19, 2013


Lecture 2: Measures 2 of 17

Example 2.4 (Examples of measures). Let S be a non-empty set, and


let S be a σ-algebra on S.

1. Measures on countable sets. Suppose that S is a finite or countable


set. Then each measure µ on S = 2S is of the form

µ( A) = ∑ p ( x ),
x∈ A

for some function p : S → [0, ∞] (why?). In particular, for a finite set


S with N elements, if p( x ) = 1/N then µ is a probability measure
called the uniform measure1 on S. 1
In the finite case, it has the well-known
property that µ( A) = #A #S , where # de-
2. Dirac measure. For x ∈ S, we define the set function δx on S by notes the cardinality (number of ele-
 ments).
1, x ∈ A,
δx ( A) =
0, x 6∈ A.

It is easy to check that δx is indeed a measure on S . Alternatively, δx


is called the point mass at x (or an atom on x, or the Dirac function,
even though it is not really a function). Moreover, δx is a probability
measure and, therefore, a finite and a σ-finite measure. It is atom
free only if { x } 6∈ S .
3. Counting Measure. Define a set function µ : S → [0, ∞] by

#A, A is finite,
µ( A) =
∞, A is infinite,

where, as above, #A denotes the number of elements in the set A.


Again, it is not hard to check that µ is a measure - it is called the
counting measure. Clearly, µ is a finite measure if and only is S
is a finite set. µ could be σ-finite, though, even without S being
finite. Simply take S = N, S = 2N . In that case µ(S) = ∞, but for
An = {n}, n ∈ N, we have µ( An ) = 1, and S = ∪n An . Finally, µ is
never atom-free and it is a probability measure only if #S = 1.

Example 2.5 (A finitely-additive set function which is not a measure).


Let S = N, and S = 2S . For A ∈ S define µ( A) = 0 if A is finite and
µ( A) = ∞, otherwise. For A1 , . . . , An ⊆ S, either

1. all Ai is finite, for i = 1, . . . , n. Then ∪in=1 Ai is also finite and so


n
0 = µ(∪in=1 Ai ) = ∑ µ ( A i ), or
i =1

2. at least one Ai is infinite. Then ∪in=1 Ai is also infinite and so


n
∞ = µ(∪in=1 Ai ) = ∑ µ ( A i ).
i =1

Last Updated: September 19, 2013


Lecture 2: Measures 3 of 17

On the other hand, take Ai = {i }, for i ∈ N. Then µ( Ai ) = 0, for Note: It is possible to construct very
each i ∈ N, and, so, ∑i∈N µ( Ai ) = 0, but µ(∪i Ai ) = µ(N) = ∞. simple-looking finite-additive measures
which are not σ-additive. For example,
there exist {0, 1}-valued finitely-additive
Proposition 2.6 (First properties of measures). Let (S, S , µ) be a measure measures on all subsets of N, which are
not σ-additive. Such objects are called
space. ultrafilters and their existence is equiva-
lent to a certain version of the Axiom of
1. For A1 , . . . , An ∈ S with Ai ∩ A j = ∅, for i 6= j, we have Choice.
n
∑ µ( Ai ) = µ(∪in=1 Ai ) (Finite additivity)
i =1

2. If A, B ∈ S , A ⊆ B, then

µ( A) ≤ µ( B) (Monotonicity of measures)

3. If { An }n∈N in S is increasing, then

µ(∪n An ) = lim µ( An ) = sup µ( An ).


n n

(Continuity with respect to increasing sequences)

4. If { An }n∈N in S is decreasing and µ( A1 ) < ∞, then

µ(∩n An ) = lim µ( An ) = inf µ( An ).


n n

(Continuity with respect to decreasing sequences)

5. For a sequence { An }n∈N in S , we have

µ(∪n An ) ≤ ∑ µ ( A n ). (Subadditivity)
n ∈N

Proof.
1. Note that the sequence A1 , A2 , . . . , An , ∅, ∅, . . . is pairwise disjoint,
and so, by σ-additivity,

µ(∪in=1 Ai ) = µ(∪i∈N Ai ) = ∑ µ ( Ai )
i ∈N
n ∞ n
= ∑ µ ( Ai ) + ∑ µ(∅) = ∑ µ ( A i ).
i =1 i = n +1 i =1

2. Write B as a disjoint union A ∪ ( B \ A) of elements of S . By (1)


above, µ( B) = µ( A) + µ( B \ A) ≥ µ( A).

3. Define B1 = A1 , Bn = An \ An−1 for n > 1. Then { Bn }n∈N is a


pairwise disjoint sequence in S with ∪nk=1 Bk = An for each n ∈ N
(why?). By σ-additivity we have
n
µ(∪n An ) = µ(∪n Bn ) = ∑ µ( Bn ) = lim
n
∑ µ( Bk )
n ∈N k =1
= lim µ(∪nk=1 Bk ) = lim µ( An ).
n n

Last Updated: September 19, 2013


Lecture 2: Measures 4 of 17

4. Consider the increasing sequence { Bn }n∈N in S given by Bn = A1 \


An . By De Morgan laws, finiteness of µ( A1 ) and (3) above, we have

µ( A1 ) − µ(∩n An ) = µ( A1 \ (∩n An )) = µ(∪n Bn ) = lim µ( Bn )


n
= lim µ( A1 \ An ) = µ( A1 ) − lim µ( An ).
n n

Subtracting both sides from µ( A1 ) < ∞ produces the statement.

5. We start from the observation that for A1 , A1 ∈ S the set A1 ∪ A2


can be written as a disjoint union

A1 ∪ A2 = ( A1 \ A2 ) ∪ ( A2 \ A1 ) ∪ ( A1 ∩ A2 ),

so that

µ ( A1 ∪ A2 ) = µ ( A1 \ A2 ) + µ ( A2 \ A1 ) + µ ( A1 ∩ A2 ).

On the other hand,

µ( A1 ) + µ( A2 ) = (µ( A1 \ A2 ) + µ( A1 ∩ A2 ))
 
+ µ ( A2 \ A1 ) + µ ( A1 ∩ A2 )
= µ( A1 \ A2 ) + µ( A2 \ A1 ) + 2µ( A1 ∩ A2 ),

and so

µ( A1 ) + µ( A2 ) − µ( A1 ∪ A2 ) = µ( A1 ∩ A2 ) ≥ 0.

Induction can be used to show that


n
µ ( A1 ∪ · · · ∪ A n ) ≤ ∑ µ ( A k ).
k =1

Since all µ( An ) are nonnegative, we now have

µ( A1 ∪ · · · ∪ An ) ≤ α, for each n ∈ N, where α = ∑ µ ( A n ).


n ∈N

The sequence { Bn }n∈N given by Bn = ∪nk=1 Ak is increasing, so the


continuity of measure with respect to increasing sequences implies
that

µ(∪n An ) = µ(∪n Bn ) = lim µ( Bn ) = lim µ( A1 ∪ · · · ∪ An ) ≤ α.


n n

Remark 2.7. The condition µ( A1 ) < ∞ in the part (4) of Proposition


2.6 cannot be significantly relaxed. Indeed, let µ be the counting mea-
sure on N, and let An = {n, n + 1, . . . }. Then µ( An ) = ∞ and, so
limn µ( An ) = ∞. On the other hand, ∩ An = ∅, so µ(∩n An ) = 0.

Last Updated: September 19, 2013


Lecture 2: Measures 5 of 17

In addition to unions and intersections, one can produce other im-


portant new sets from sequences of old ones. More specifically, let
{ An }n∈N be a sequence of subsets of S. The subset lim infn An of S,
defined by

lim inf An = ∪n Bn , where Bn = ∩k≥n Ak ,


n

is called the limit inferior of the sequence An . It is also denoted by


limn An or { An , ev.} (ev. stands for eventually2 ). 2
the reason for the use of the word even-
Similarly, the subset lim supn An of S, defined by tually is the following: lim infn An is the
set of all x ∈ S which belong to An for all
but finitely many values of the index n, i.e.,
lim sup An = ∩n Bn , where Bn = ∪k≥n Ak , from some value of the index n onwards.
n

is called the limit superior of the sequence An . It is also denoted by


limn An or { An , i.o.} (i.o. stands for infinitely often3 ). Clearly, we have 3
in words, lim supn An is the set of all
x ∈ S which belong An for infinitely
lim inf An ⊆ lim sup An . many values of n.
n n

Problem 2.1. Let (S, S , µ) be a finite measure space. Show that Hint: For the second part, a measure
space with finite (and small) S will do.
µ(lim inf An ) ≤ lim inf µ( An ) ≤ lim sup µ( An ) ≤ µ(lim sup An ),

for any sequence { An }n∈N in S . Give an example of a (single) se-


quence { An }n∈N for which all inequalities above are strict.

Proposition 2.8 (Borel-Cantelli Lemma I). Let (S, S , µ) be a measure


space, and let { An }n∈N be a sequence of sets in S with the property that
∑n∈N µ( An ) < ∞. Then

µ(lim sup An ) = 0.
n

Proof. Set Bn = ∪k≥n Ak , so that { Bn }n∈N is a decreasing sequence of


sets in S with lim supn An = ∩n Bn , and so

µ(lim sup An ) ≤ µ( Bn ), for each n ∈ N.

Using the subadditivity of measures of Proposition 2.6, part 5., we get


µ( Bn ) ≤ ∑ µ ( A n ). (2.1)
k=n

Since ∑n∈N µ( An ) converges, the right-hand side of (2.1) can be made


arbitrarily small by choosing large enough n ∈ N.

Extensions of measures and the coin-toss space

Example 1.19 has introduced the measurable space ({−1, 1}N , S), with
S = ⊗n 2{−1,1} being the product σ-algebra on {−1, 1}N . The purpose

Last Updated: September 19, 2013


Lecture 2: Measures 6 of 17

of the present section is to turn ({−1, 1}N , S) into a measure space,


i.e., to define a suitable measure on it. It is easy to construct just any
measure on {−1, 1}N , but the one we are after is the one which will
justify the name coin-toss space.
The intuition we have about tossing a fair coin infinitely many times
should help us start with the definition of the coin-toss measure - de-
noted by µC - on cylinders. Since the coordinate spaces {−1, 1} are
particularly simple, each product cylinder is of the form C = {−1, 1}N
or C = Cn1 ,...,nk ;b1 ,...,bk , where
n o
Cn1 ,...,nk ;b1 ,...,bk = s = (s1 , s2 , . . . ) ∈ {−1, 1}N : sn1 = b1 , . . . , snk = bk ,

for some k ∈ N, and a choice of 1 ≤ n1 < n2 < · · · < nk ∈ N of


coordinates and the corresponding values b1 , b2 , . . . , bk ∈ {−1, 1}.
In the language of elementary probability, each cylinder corresponds
to the event when the outcome of the ni -th coin is bi ∈ {−1, 1}, for
i = 1, . . . , n. The measure (probability) of this event can only be given
by
µC (Cn1 ,...,nk ;b1 ,...,bk ) = 1
× 12 × · · · × 21 = 2−k .
|2 {z } (2.2)
k times

The hard part is to extend this definition to all elements of S , and not
only cylinders. For example, in order to state the law of large numbers
later on, we will need to be able to compute the measure of the set
n n o
s ∈ {−1, 1}N : lim n1 ∑ sk = 0 ,
n
k =1

which is clearly not a cylinder.


Problem 1.9 states, however, that cylinders form an algebra and gen-
erate the σ-algebra S . Luckily, this puts us close to the conditions of
the following important theorem of Caratheodory.

Theorem 2.9 (Caratheodory’s Extension Theorem). Let S be a non- Note: In words, a σ-additive measure
empty set, let A be an algebra of its subsets and let µ : A → [0, ∞] be a on an algebra A can be extended to
a σ-additive measure on the σ-algebra
set-function with the following properties: generated by A. It is clear that the σ-
additivity requirement of Theorem 2.9 is
1. µ(∅) = 0, and necessary, but it is quite surprising that
it is actually sufficient.
2. µ( A) = ∑∞
n=1 µ ( An ), if { An }n∈N ⊆ A is a partition of A.

Then, there exists a measure µ̃ on σ (A) with the property that µ( A) = µ̃( A)
for A ∈ A.

Of Theorem 2.9. PART I. We start by defining a “measure-like object”, Note: Intuitively, we try all different
called an outer measure, µ∗ : 2S → [0, ∞] in the following way: countable covers of B with elements of
A and minimize the total µ.
n ∞ ∞ o
µ∗ ( B) = inf ∑ An , An ∈ A for all n ∈ N .
[
µ( An ) : B ⊆
n =1 n =1

Last Updated: September 19, 2013


Lecture 2: Measures 7 of 17

Even though we don’t expect the infimum in the definition of µ∗ to be


attained, µ∗ has the following properties:

1. µ∗ (∅) = 0 (nontriviality),

2. for B ⊆ C, µ∗ ( B) ≤ µ∗ (C ) (monotonicity), and

3. µ∗ (∪k Bk ) ≤ ∑∞ ∗
k =1 µ ( Bk ) (subadditivity)

Parts 1. and 2. are immediately clear, while, to show 3., we pick ε > 0
and k ∈ N and find a countable cover { Akn }n∈N with elements of A
such that

µ∗ ( Bk ) ≥ ∑ µ( Akn ) + 2εk .
n =1

Using { Akn }k∈N,n∈N as a candidate cover for ∪k Bk , we conclude that


µ∗ (∪k Bk ) ≤ ∑∞ ∗
k =1 µ ( Bk ) + ε. This being true for each ε > 0 implies 3.
We remark at this point that µ∗ and µ coincide on A. By using
( A, ∅, ∅, . . . ) as a candidate countable cover of A ∈ A, we can con-
clude that µ∗ ( A) ≤ µ( A), for all A ∈ A. Conversely, suppose that
A ⊆ ∪n∈N An with An ∈ A. Given that the elements of A form an al-
gebra, we can assume that An ⊆ A, for all n ∈ N and that { An }n∈N are
pairwise disjoint, as any sequence covering A can be transformed into
such a sequence without increasing ∑n µ( An ). The assumed countable Note: It is, probably, interesting to note
additivity of µ on A now comes into play since, for partition { An }n∈N that this is the only place in the entire
proof where the countable additivity of
of A into elements in A, we necessarily have ∑n µ( An ) = µ( A), and, µ on A is used.
so µ∗ ( A) ≥ µ( A).
PART II. The set-function µ∗ is, in general, not a measure, but comes
with the advantage of being defined on all subsets of S. The central
idea of the proof is to recover countable additivity by restricting its do-
main a little. We say that a subset M ⊆ S is Caratheodory-measurable
or µ∗ -measurable if

µ∗ ( B) = µ∗ ( B ∩ M) + µ∗ ( B ∩ Mc ) for all B ⊆ S, (2.3)

with the family of all µ∗ -measurable subsets of S denoted by M∗ . We


note that, by subadditivity, the equality sign in the definition of the
measurability can be replaced by ≥; this will be used below.
The first thing we need to establish about M∗ is that it is an algebra
and that µ∗ is a finitely-additive measure on M∗ . Clearly ∅ ∈ A∗ and
the complement axiom follows directly from the symmetry in (2.3).
Only the closure under finite unions needs some discussion, and, by
induction, we only need to consider two-element unions; for that, we
pick M, N ∈ M∗ , and introduce the following notation

M00 = Mc ∩ N c , M01 = Mc ∩ N, M10 = M ∩ N c M11 = M ∩ N. (2.4)

Last Updated: September 19, 2013


Lecture 2: Measures 8 of 17

By the measurability of M and N, for any B ⊆ S, we have

µ∗ ( B) = µ∗ ( B ∩ N c ) + µ∗ ( B ∩ N )
= µ∗ ( M00 ∩ B) + µ∗ ( M10 ∩ B) + µ∗ ( M01 ∩ B) + µ∗ ( M11 ∩ B)
On the other hand, M01 ∪ M10 ∪ M11 = M ∪ N, so that, by subadditiv-
ity and (2.4), we have

µ∗ ( M ∪ N )c ∩ B + µ∗ ( M ∪ N ) ∩ B
 
 
= µ∗ ( M00 ∩ B) + µ∗ ( M01 ∩ B) ∪ ( M10 ∩ B) ∪ ( M11 ∩ B)
≤ µ∗ ( M00 ∩ B) + µ∗ ( M10 ∩ B) + µ∗ ( M01 ∩ B) + µ∗ ( M11 ∩ B)
= µ ∗ ( B ),
which implies that that M ∪ N ∈ M∗ . When M ∩ N = ∅ an applica-
tion of measurability of N to B = M ∪ N yields the finite additivity of
µ∗ on M∗ :

µ∗ ( M ∪ N ) = µ∗ (( M ∪ N ) ∩ N ) + µ∗ (( M ∪ N ) ∩ N c ) = µ∗ ( N ) + µ∗ ( M).

PART III. We now turn to the closure of M∗ under countable unions


and the σ-additive property of µ∗ . Since M∗ already known to be an
algebra, it will be enough to show that it is closed under pairwise-
disjoint unions, i.e., that M ∈ M∗ whenever { Mn }n∈N are pairwise
disjoint elements in M∗ with M = ∪n Mn .
For n ∈ N, we set Ln = ∪nk=1 Mk so that, for B ⊆ S, we have

µ∗ ( B) = µ∗ ( B ∩ Ln ) + µ∗ ( B ∩ Lcn )
≥ ∑nk=1 µ∗ ( B ∩ Mk ) + µ∗ ( B ∩ Mc ),
so that
µ ∗ ( B ) ≥ µ ∗ ( B ∩ M c ) + ∑ k ∈N µ ∗ ( B ∩ Mk )
≥ µ∗ ( B ∩ Mc ) + µ∗ (∪k ( B ∩ Mk )) + µ∗ ( B ∩ Mc )
= µ ∗ ( B ∩ M ) + µ ∗ ( B ∩ M c ).
Since all the inequalities above need to be equalities, we immediately
conclude that, with B = S,

µ∗ ( M ) = ∑ µ ∗ ( Mk ) ,
k

i.e., that µ∗
is a countably-additive measure on M∗ . Since A ⊆ M∗ , we
have M∗ ⊇ σ(A) and µ̃ = µ∗ |σ(A) is the required σ-additive extension
of µ.

Back to the coin-toss space. In order to apply Theorem 2.9 in our


situation, we need to check that µ is indeed a countably-additive mea-
sure on the algebra A of all cylinders. The following problem will help
pinpoint the hard part of the argument:

Last Updated: September 19, 2013


Lecture 2: Measures 9 of 17

Problem 2.2. Let A be an algebra on a non-empty set S, and let µ :


A → [0, ∞] be a finite (µ(S) < ∞) and finitely-additive set function on
S with the following, additional, property:

lim µ( An ) = 0, whenever An & ∅. (2.5)


n

Then µ satisfies the conditions of Theorem 2.9.

The part about finite additivity is easy (perhaps a bit messy) and we
leave it to the reader:

Problem 2.3. Show that the set-function µC , defined by (2.2) on the


product cyllinders and extended by additivity to the algebra A of
cylinders, is finitely additive.

Lemma 2.10 (Conditions of Caratheodory’s theorem). The set-function


µC , defined by (2.2), and extenede by additivity to the the algebra A of cylin-
ders, has the property (2.5).

Proof. By Problem 1.10, cylinders are closed sets, and so { An }n∈N is a


sequence of closed sets whose intersection is empty. The same prob-
lem states that {−1, 1}N is compact, so, by the finite-intersection prop-
erty4 , we have An1 ∩ . . . Ank = ∅, for some finite collection n1 , . . . , nk 4
The finite-intersection property refers to
of indices. Since { An }n∈N is decreasing, we must have An = ∅, for all the following fact, familiar from real
analysis: If a family of closed sets of a
n ≥ nk , and, consequently, limn µ( An ) = 0. compact topological space has empty in-
tersection, then it admits a finite subfam-
ily with an empty intersection.
Proposition 2.11 (Existence of the coin-toss measure). There exists a
measure µC on ({−1, 1}N , S) with the property that (2.2) holds for all cylin-
ders.

Proof. Thanks to Lemma 2.10, Theorem 2.9 can now be used.

In order to prove uniqueness, we will need the celebrated π-λ The-


orem of Eugene Dynkin:

Theorem 2.12 (Dynkin’s “π-λ” Theorem). Let P be a π-system on a non-


empty set S, and let Λ be a λ-system which contains P . Then Λ also contains
the σ-algebra σ (P ) generated by P .

Proof. Using the result of part 4. of Problem 1.1, we only need to prove
that λ(P ) (where λ(P ) denotes the λ-system generated by P ) is a π-
system. For A ⊆ S, let G A denote the family of all subsets of S whose
intersections with A are in λ(P ):

G A = {C ⊆ S : C ∩ A ∈ λ(P )}.

Last Updated: September 19, 2013


Lecture 2: Measures 10 of 17

Claim: G A is a λ-system for A ∈ λ(P ).


• Since A ∈ λ(P ), clearly S ∈ G A .

• For an increasing family {Cn }n∈N in G A we have (∪n Cn ) ∩


A = ∪n (Cn ∩ A). Each Cn ∩ A is in Λ, and the family {Cn ∩
A}n∈N is increasing, so (∪n Cn ) ∩ A ∈ Λ.

• Finally, for C1 , C2 ∈ G with C1 ⊆ C2 , we have

(C2 \ C1 ) ∩ A = (C2 ∩ A) \ (C1 ∩ A) ∈ Λ,

because C1 ∩ A ⊆ C2 ∩ A.

P is a π-system, so for any A ∈ P , we have P ⊆ G A . Therefore,


λ(P ) ⊆ G A , because G A is a λ-system. In other words, for A ∈ P and
B ∈ λ(P ), we have A ∩ B ∈ λ(P ).
That means, however, that P ⊆ G B , for any B ∈ λ(P ). Using the fact
that G B is a λ-system we must also have λ(P ) ⊆ G B , for any B ∈ λ(P ),
i.e., A ∩ B ∈ λ(P ), for all A, B ∈ λ(P ), which shows that λ(P ) is
π-system.

Proposition 2.13 (Measures which agree on a π-system). Let (S, S) be


a measurable space, and let P be a π-system which generates S . Suppose that
µ1 and µ2 are two measures on S with the property that µ1 (S) = µ2 (S) < ∞
and
µ1 ( A) = µ2 ( A), for all A ∈ P .
Then µ1 = µ2 , i.e., µ1 ( A) = µ2 ( A), for all A ∈ S .

Proof. Let L be the family of all subsets A of S for which µ1 ( A) =


µ2 ( A). Clearly P ⊆ L, but L is, potentially, bigger. In fact, it follows
easily from the elementary properties of measures (see Proposition 2.6)
and the fact that µ1 (S) = µ2 (S) < ∞ that it necessarily has the struc-
ture of a λ-system5 . By Theorem 2.12 (the π-λ Theorem), L contains 5
It seems that the structure of a λ-system
the σ-algebra generated by P , i.e., S ⊆ L. On the other hand, by is defined so that it would exactly de-
scribe the structure of the family of all
definition, L ⊆ S and so µ1 = µ2 . sets on which two measures (with the
same total mass) agree. The structure
of the π-system corresponds to the min-
Proposition 2.14 (Uniqueness of the coin-toss measure). The measure imal assumption that allows Proposition
µC is the unique measure on ({−1, 1}N , S) with the property that (2.2) 2.13 to hold.
holds for all cylinders.

Proof. The existence is the content of Proposition 2.11. To prove unique-


ness, it suffices to note that algebras are π-systems and use Proposition
2.13.

Problem 2.4. Define D1 , D2 ⊆ {−1, 1}N by

Last Updated: September 19, 2013


Lecture 2: Measures 11 of 17

1. D1 = {s ∈ {−1, 1}N : lim supn sn = 1},

2. D2 = {s ∈ {−1, 1}N : ∃ N ∈ N, s N = s N +1 = s N +2 }.

Show that D1 , D2 ∈ S and compute µ( D1 ), µ( D2 ).

Our next task is to probe the structure of the σ-algebra S on {−1, 1}N
N
a little bit more and show that S 6= 2{−1,1} . It is interesting that such
a result (which deals exclusively with the structure of S ) requires a
use of a measure in its proof.

Example 2.15 (A non-measurable subset of {−1, 1}N (*)). Since σ-


algebras are closed under countable set operations, and since the prod-
uct σ-algebra S for the coin-toss space {−1, 1}N is generated by sets
obtained by restricting finite collections of coordinates, one is tempted
to think that S contains all subsets of {−1, 1}N . That is not the case.
We will use the axiom of choice, together with the fact that a measure
µC can be defined on the whole of {−1, 1}N , to show to “construct”
an example of a non-measurable set.
Let us start by constructing a relation ∼ on {−1, 1}N in the fol-
lowing6 way: we set s1 ∼ s2 if and only if there exists n ∈ N such 6
In words, s1 and s2 are related if they
that s1k = s2k , for k ≥ n (here, as always, si = (s1i , s2i , . . . ), i = 1, 2). only differ in a finite number of coordi-
nates.
It is easy to check that ∼ is an equivalence relation and that it splits
{−1, 1}N into disjoint equivalence classes. One of the many equiva-
lent forms of the axiom of choice states that there exists a subset N of
{−1, 1}N which contains exactly one element from each of the equiv-
alence classes.
Let us suppose that N is an element in S and see if we can reach a
contradiction. For each nonempty n = {n1 , . . . , nk } ∈ 2N N
f in , where 2 f in
denotes the family of all finite subsets of N, let us define the mapping
Tn : {−1, 1}N → {−1, 1}N in the following7 manner: 7
Tn flips the signs of the elements of its
 argument on the positions correspond-
−s , l ∈ n, ing to n.
l
T∅ = Id and ( Tn (s))l = for n ∈ N.
sl , l 6∈ n,

Since n is finite, Tn preserves the ∼-equivalence class of each ele-


ment. Consequently (and using the fact that N contains exactly one
element from each equivalence class) the sets N and Tn ( N ) = { Tn (s) :
s ∈ N } are disjoint. Similarly and more generally, the sets Tn ( N ) and
Tn0 ( N ) are also disjoint whenever n 6= n0 . On the other hand, each
s ∈ {−1, 1}N is equivalent to some ŝ ∈ N, i.e., it can be obtained from
ŝ by flipping a finite number of coordinates. Therefore, the family

N = { Tn ( N ) : n ∈ 2N
f in }

forms a partition of {−1, 1}N .

Last Updated: September 19, 2013


Lecture 2: Measures 12 of 17

The mapping Tn has several other nice properties. First of all, it


is immediate that it is involutory, i.e., Tn ◦ Tn = Id. To show that
it is (S , S)-measurable, we need to prove that its composition with
each projection map πk : S → {−1, 1} is measurable. This follows
immediately from the fact that for k ∈ N

C ,
k;1 k 6∈ n,
(πk ◦ Tn )−1 ({1}) =
Ck;−1 , k ∈ n,

where, for b ∈ {−1, 1}, we recall that Ck;b = {s ∈ {−1, 1}N : sk = b}


is a product cylinder. If we combine the involutivity and measurability
of Tn , we immediately conclude that Tn ( A) ∈ S for each A ∈ S . In
particular, N ⊆ S .
In addition to preserving measurability, the map Tn also preserves
the measure8 the in µC , i.e., µC ( Tn ( A)) = µC ( A), for all A ∈ S . To 8
Actually, we say that a map f from a
prove that, let us pick n ∈ F and consider the set-function µn : S → measure space (S, S , µS ) to the measure
space ( T, T , µ T ) is measure preserving
[0, 1] given by if it is measurable and µS ( f −1 ( A)) =
µn ( A) = µC ( Tn ( A)). µ T ( A), for all A ∈ T . The involutivity
of the map Tn implies that this general
It is a simple matter to show that µn is, in fact, a measure on (S, S) definition agrees with our usage in this
with µn (S) = 1. Moreover, thanks to the simple form (2.2) of the action example.

of the measure µC on cylinders, it is clear that µn = µC on the algebra


of all cylinders. It suffices to invoke Proposition 2.13 to conclude that
µn = µC on the entire S , i.e., that Tn preserves µC .
The above properties of the maps Tn , n ∈ F can imply the following:
N is a partition of S into countably many measurable subsets of equal
measure. Such a partition { N1 , N2 , . . . } cannot exist, however. Indeed
if it did, one of the following two cases would occur:
1. µ( N1 ) = 0. In that case

µ(S) = µ(∪k Nk ) = ∑ µ( Nk ) = ∑ 0 = 0 6= 1 = µ(S).


n n

2. µ( N1 ) = α > 0. In that case

µ(S) = µ(∪k Nk ) = ∑ µ( Nk ) = ∑ α = ∞ 6= 1 = µ(S).


n n

Therefore, the set N cannot be measurable9 in S . 9


Somewhat heavier set-theoretic ma-
chinery can be used to prove that most of
the subsets of S are not in S , in the sense
The Lebesgue measure that the cardinality of the set S is strictly
smaller than the cardinality of the set 2S
of all subsets of S
As we shall see, the coin-toss space can be used as a sort of a universal
measure space in probability theory. We use it here to construct the
Lebesgue measure on [0, 1]. We start with the notion somewhat dual
to the already introduced notion of the pull-back in Definition 1.8. We
leave it as an exercise for the reader to show that the set function f ∗ µ
from Definition 2.16 is indeed a measure.

Last Updated: September 19, 2013


Lecture 2: Measures 13 of 17

Definition 2.16 (Push-forwards). Let (S, S , µ) be a measure space and


let ( T, T ) be a measurable space. The measure f ∗ µ on ( T, T ), defined
by
f ∗ µ( B) = µ( f −1 ( B)), for B ∈ T ,
is called the push-forward of the measure µ by f .

Let f : {−1, 1}N → [0, 1] be the mapping given by


∞  
f (s) = ∑ 1+ s k
2 2−k , s ∈ {−1, 1}N .
k =1

The idea is to use f to establish a correspondence between all real num-


bers in [0, 1] and their expansions in the binary system, with the coding
−1 7→ 0 and 1 7→ 1. It is interesting to note that f is not one-to-one10 10
The reason for this is, poetically speak-
as it, for example, maps s1 = (1, −1, −1, . . . ) and s2 = (−1, 1, 1, . . . ) ing, that [0, 1] is not the Cantor set.

into the same value - namely 21 . Let us show, first, that the map f is
continuous in the metric d defined by part (1.2) of Problem 1.9. Indeed,
we pick s1 and s2 in {−1, 1}N and remember that for d(s1 , s2 ) ≤ 2−n ,
the first n − 1 coordinates of s1 and s2 coincide. Therefore,

| f (s1 ) − f (s2 )| ≤ ∑ 2−k = 2−n+1 = 2d(s1 , s2 ).
k=n

Hence, the map f is Lipschitz and, therefore, continuous.


The continuity of f (together with the fact that S is the Borel σ-
algebra for the topology induced by the metric d) implies that f :
({−1, 1}N , S) → ([0, 1], B([0, 1])) is a measurable mapping. There-
fore, the push-forward λ = f ∗ (µ) is well defined on ([0, 1], B([0, 1])),
and we call it the Lebesgue measure on [0, 1].

Proposition 2.17 (Intuitive properties of the Lebesgue measure). The


Lebesgue measure λ on ([0, 1], B([0, 1])) satisfies

λ([ a, b)) = b − a, λ({ a}) = 0 for 0 ≤ a < b ≤ 1. (2.6)

Proof.

1. Consider a, b of the form b = 2kn and b = k2+n1 , for n ∈ N and k < 2n .


For such a, b we have f −1 ([ a, b)) = C1,...,n;c1 ,c2 ,...,cn , where c1 c2 . . . cn
is the base-2 expansion of k (after the “recoding” −1 7→ 0, 1 7→ 1).
By the very definition of λ and the form (2.2) of the action of the
coin-toss measure µC on cylinders, we have
   
λ [ a, b) = µC f −1 [ a, b) = µC (C1,...,n;c1 ,c2 ,...,cn ) = 2−n = k2+n1 − 2kn .

Therefore, (2.6) holds for a, b of the form b = 2kn and b = 2ln , for
n ∈ N, k < 2n and l = k + 1. Using (finite) additivity of λ, we

Last Updated: September 19, 2013


Lecture 2: Measures 14 of 17

immediately conclude that (2.6) holds for all k, l, i.e., that it holds
for all dyadic rationals. A general a ∈ (0, 1] can be approximated
by an increasing sequence {qn }n∈N of dyadic rationals from the left,
and the continuity of measures with respect to decreasing sequences
implies that
     
λ [ a, p) = λ ∩n [qn , p) = lim λ [qn , p) = lim( p − qn ) = ( p − a),
n n

whenever a ∈ (0, 1] and p is a dyadic rational. In order to remove


the dyadicity requirement from the right limit, we approximate it
from the left by a sequence { pn }n∈N of dyadic rationals with pn > a,
and use the continuity with respect to increasing sequences to get,
for a < b ∈ (0, 1),
     
λ [ a, b) = λ ∪n [ a, pn ) = lim λ [ a, pn ) = lim( pn − a) = (b − a).
n n

The Lebesgue measure has another important property:

Problem 2.5. Show that the Lebesgue measure is translation invari- Hint: Use Proposition 2.13 for the second
ant. More precisely, for B ∈ B([0, 1]) and x ∈ [0, 1), we have part.

1. B +1 x = {b + x (mod 1) : b ∈ B} is in B([0, 1]) and

2. λ( B +1 x ) = λ( B),

where, for a ∈ [0, 2), we define Geometrically, the set x +1 B is obtained


 from B by translating it to the right by x
and then shifting the part that is “stick-
 a, a ≤ 1, ing out” by 1 to the left.
a (mod 1) =
 a − 1, a > 1.

Finally, the notion of the Lebesgue measure is just as useful on the


entire R, as on its compact subset [0, 1]. For a general B ∈ B(R), we
can define the Lebesgue measure of B by measuring its intersections
with all intervals of the form [n, n + 1), and adding them together, i.e.,
∞  


λ( B) = λ B ∩ [n, n + 1) − n .
n=−∞

Note how we are overloading the notation and using the letter λ for
both the Lebesgue measure on [0, 1] and the Lebesgue measure on R.
It is a quite tedious, but does not require any new tools, to show
that many of the properties of λ on [0, 1] transfer to λ on R:

Problem 2.6. Let λ be the Lebesgue measure on (R, B(R)). Show that

1. λ([ a, b)) = b − a, λ({ a}) = 0 for a < b,

Last Updated: September 19, 2013


Lecture 2: Measures 15 of 17

2. λ is σ-finite but not finite,

3. λ( B + x ) = λ( B), for all B ∈ B(R) and x ∈ R, where B + x =


{ b + x : b ∈ B }.

Additional Problems

Problem 2.7 (Local separation by constants). Let (S, S , µ) be a measure


space and let f , g ∈ L0 (S, S , µ) satisfy µ { x ∈ S : f ( x ) < g( x )} > 0.


Prove or construct a counterexample for the following statement:


“There exist constants a, b ∈ R such that

µ { x ∈ S : f ( x ) ≤ a < b ≤ g( x )} > 0.”

Problem 2.8 (A pseudometric on sets). Let (S, S , µ) be a finite measure Note: Let X be a nonempty set. A func-
space. For A, B ∈ S define tion d : X × X → [0, ∞) is called a pseu-
dometric if
d( A, B) = µ( A M B), 1. d( x, y) + d(y, x ) ≥ d( x, z), for all
x, y, z ∈ X,
where M denotes the symmetric difference: A M B = ( A \ B) ∪ ( B \ A). 2. d( x, y) = d(y, x ), for all x, y ∈ X, and
Show that d is a pseudometric on S , and for A ∈ S describe the set of 3. d( x, x ) = 0, for all x ∈ X.
all B ∈ S with d( A, B) = 0. Note how the only difference between a
metric and a pseudometric is that for a
metric d( x, y) = 0 implies x = y, while
Problem 2.9 (Complete measure spaces). A measure space (S, S , µ) is no such requirement is imposed on a
called complete if all subsets of null sets are themselves in S . For a pseudometric.
(possibly incomplete) measure space (S, S , µ) we define the comple-
tion (S, S ∗ , µ∗ ) in the following way: Note: Unfortunately, the same notation
µ∗ is often used for the completion of
S ∗ = { A ∪ N ∗ : A ∈ S and N ∗ ⊆ N for some N ∈ S with µ( N ) = 0}. the measure µ and the outer measure as-
sociated with µ as in the proof of The-
orem 2.9. Fortunately, it can be shown
For B ∈ S ∗ with representation B = A ∪ N ∗ we set µ∗ ( B) = µ( A).
that these two object coincide on the do-
main of the completion.
1. Show that S ∗ is a σ-algebra.

2. Show that the definition µ∗ ( B) = µ( A) above does not depend on


the choice of the decomposition B = A ∪ N ∗ , i.e., that µ( Â) = µ( A)
if B =  ∪ N̂ ∗ is another decomposition of B into a set  in S and
a subset N̂ of a null set in S .

3. Show that µ∗ is a measure on (S, S ∗ ) and that (S, S ∗ , µ∗ ) is a com-


plete measure space with the property that µ∗ ( A) = µ( A), for
A ∈ S.

Problem 2.10 (The Cantor set). The Cantor set is defined as the collec-
tion of all real numbers x in [0, 1] with the representation

x= ∑ c n 3− n , where cn ∈ {0, 2}.
n =1

Show that it is Borel-measurable and compute its Lebesgue measure.

Last Updated: September 19, 2013


Lecture 2: Measures 16 of 17

Problem 2.11 (The uniform measure on a circle). Let S1 be the unit


circle, and let f : [0, 1) → S1 be the “winding map”
 
f ( x ) = cos(2πx ), sin(2πx ) , x ∈ [0, 1).

1. Show that the map f is (B([0, 1)), S 1 )-measurable, where S 1 denotes


the Borel σ-algebra on S1 (with the topology inherited from R2 ).

2. For α ∈ (0, 2π ), let Rα denote the (counter-clockwise) rotation of R2


with center (0, 0) and angle α. Show that Rα ( A) = { Rα ( x ) : x ∈ A}
is in S 1 if and only if A ∈ S 1 .

3. Let µ1 be the push-forward of the Lebesgue measure λ by the map Note: The measure µ1 is called the uni-
f . Show that µ1 is rotation-invariant, i.e., that µ1 ( A) = µ1 Rα ( A) . form measure (or the uniform distribu-

tion) on S1 .

Problem 2.12 (Asymptotic densities). We say that the subset A of N


admits asymptotic density if the limit

#( A ∩ {1, 2, . . . , n})
d( A) = lim ,
n n
exists (remember that # denotes the number of elements of a set). Let
D be the collection of all subsets of N which admit asymptotic density.

1. Is D an algebra? A σ-algebra?

2. Is the map A 7→ d( A) finitely-additive on D ? A measure?

Problem 2.13 (A subset of the coin-toss space). An element in {−1, 1}N


(i.e., a sequence s = (s1 , s2 , . . . ) where sn ∈ {−1, 1} for all n ∈ N) is
said to be eventually periodic if there exists N0 , K ∈ N such that
sn = sn+K for all n ≥ N0 . Let P ⊆ {−1, 1}N be the collection of all
eventually-period sequences. Show that P is measurable in the product
σ-algebra S and compute µC ( P).

Problem 2.14 (Regular measures). The measure space (S, S , µ), where
(S, d) is a metric space and S is a σ-algebra on S which contains the
Borel σ-algebra B(d) on S is called regular if for each A ∈ S and each
ε > 0 there exist a closed set C and an open set O such that C ⊆ A ⊆ O
and µ(O \ C ) < ε.

1. Suppose that (S, S , µ) is a regular measure space, and that the mea-
sure space (S, B(d), µ|B(d) ) is obtained from (S, S , µ) by restrict-
ing the measure µ onto the σ-algebra of Borel sets. Show that
S ⊆ B(d)∗ , where S, B(d)∗ , (µ|B(d) )∗ is the completion (in the


sense of Problem 2.9) of (S, B(d), µ|B(d) )

2. Suppose that (S, d) is a metric space and that µ is a finite measure


on B(d). Show that (S, B(d), µ) is a regular measure space.

Last Updated: September 19, 2013


Lecture 2: Measures 17 of 17

Hint: Consider a collection A of subsets A of S such that for each ε > 0 there exists
a closed set C and an open set O with C ⊆ A ⊆ O and µ(O \ C ) < ε. Argue that A is
a σ-algebra. Then show that each closed set can be written as an intersection of open
sets; use (but prove, first) the fact that the map
x 7→ d( x, C ) = inf{d( x, y) : y ∈ C },

is continuous on S for any nonempty C ⊆ S.

3. Show that (S, B(d), µ) is regular if µ is not necessarily finite, but has
the property that µ( A) < ∞ whenever A ∈ B(d) is bounded, i.e.,
when sup{d( x, y) : x, y ∈ A} < ∞.
Hint: Pick a point x0 ∈ S and, for n ∈ N, define the family { Rn }n∈N of subsets of S
as follows:
R1 = { x ∈ S : d( x, x0 ) < 2}, and
Rn = { x ∈ S : n − 1 < d( x, x0 ) < n + 1}, for n > 1,

as well as a sequence {µn }n∈N of set functions on B(d), given by µn ( A) = µ( A ∩ Rn ),


for A ∈ B(d). Under the right circumstances, even countable unions of closed sets
are closed.

4. Conclude that the Lebesgue measure on Rn , B(Rn ) is regular.




Last Updated: September 19, 2013

You might also like