Académique Documents
Professionnel Documents
Culture Documents
Oscar Blasco
Contents
1 Abstract Measure
1.1 Basic notions on sets . . . . . .
1.2 Basic notions on set functions. .
1.3 Outer measures. . . . . . . . . .
1.4 Extension of measures. . . . . .
1.5 Borel-Stieltjes measures on R. .
1.6 Measurable and non-measurable
1.7 Exercises . . . . . . . . . . . . .
. . .
. . .
. . .
. . .
. . .
sets.
. . .
4 The
4.1
4.2
4.3
4.4
Radon-Nikodym Theorem
Complex and real measures. .
The theorem and its proof. . .
Applications . . . . . . . . . .
Exercises . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
theorem
. . . . . .
. . . . . .
. . . . . .
. . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5
5
9
11
15
17
23
26
.
.
.
.
29
29
35
38
50
.
.
.
.
57
57
63
64
68
.
.
.
.
73
73
80
84
88
Chapter 1
Abstract Measure
1.1
Remark 1.1.1 Let P(X). is a -algebra if and only if is a monotone class and an algebra.
Let A P(X). A is an algebra if and only if A is a ring containing X.
Example 1.1.1 (1) The trivial -algebras are P(X) and = {, X}.
(2) Let M = {An : n N} X where A1 = , An An+1 for all n N
and n An = X. This is a monotone class but not necessarily a ring.
(3) Let X = [0, 1) and A = {finite unions of intervals [a, b), 0 a b
1} is an algebra but not -algebra.
If the intervals in the previous family are assumed to have 0 a b < 1
then it is a ring but not an algebra.
(4) Let X be non empty and numerable. R = {A X : card(A) < } is
a ring but not algebra.
Definition 1.1.5 Let An P(X) for n N. The upper limit (respect. lower
limit) of the sequence is defined by
lim sup An =
n=1 k=n Ak
(respect.
lim inf An =
n=1 k=n Ak .)
A sequence is said to have limit if lim sup An = lim inf An . Such a set is
called lim An .
Remark 1.1.2 Any monotone sequence has a limit. If An is increasing
(respect. decreasing) then lim An = n An (respect. lim An = n An ).
Proposition 1.1.6 Let R be a ring. If A, B R then A M B R and
A B R.
Proof: Note that A M B = ((A B) \ A) ((A B) \ B) and A B =
A B \ (A M B).
Proposition 1.1.7 Let be a -algebra. If An for all n N then
n An , lim sup An and lim inf An
Proof: Write n An = X \ n (X \ An ) and apply the properties of an
-algebra.
Proof: Notice that any open set is a numerable union of open intervals.
Hence Remark 1.1.3 shows that B(R) = (E1 ).
Let us observe that (a, b] = n (a, b + n1 ), [a, b) = n (a n1 , b) and [a, b] =
n (a n1 , b + n1 ) to get the other cases.
Proposition 1.1.14 Let n N and X = Rn and consider E = {(a1 , b1 ]
(a2 , b2 ] ...(an , bn ] : ai bi , i = 1, ..., n}. Then B(Rn ) = (E).
Proof: We sketch the proof for the case n = 2.
Consider J0 = {(n, n + 1] (m, m + 1] : n, m Z} and Jn the collection
of intervals resulting of dividing each square of the previous family into four
of the same area.
Now for each x R2 there exists a unique sequence of intervals Ik (x) Jk
such that x Ik (x) for all k N, Ik+1 (x) Ik (x) and Area(Ik (x)) 0 as
k .
Given an open set G we can consider the family F = {J k Jk : J G}.
Of course G = JF J and with a little effort it can be seen that only a
numerable number of sets is needed. Since F E then the proof is finished.
Definition 1.1.15 Let E P(X). M(E) stands for the smallest monotone
class containing E, which is called the monotone class generated by E.
Theorem 1.1.16 (The monotone class theorem) Let A be an algebra over
X. Then (A) = M(A).
Proof: It suffices to see that M(A) is -algebra . Since M(A) is a monotone
class we have only to show that it is an algebra.
Clearly A M(A). Now given A M(A), to see that X \A M(A)
let us define = {A M(A) : X \ A M(A)} and show that = M(A).
For such a purpose we shall see that is a monotone class and contains A.
Indeed, if An is a monotone sequence then lim An . Clearly if A A
then X \ A A M(A). Therefore A .
Given now A, B M(A), we need to show that A B M(A). Let us
define A = {B M(A) : A B M(A)}. Note that A is a monotone class
(since (Bn )n monotone sequence in A has limit in A ).
Now let us deal first with the case A A. In this case A contains A
(since B A implies A B A M(A)). Therefore A = M(A) for all
A A.
1.2
X
n=1
(Bn )
(An ).
n=1
n
X
k=1
(Bk )
X
n=1
(Bn ) = (n Bn ) = (n An ).
10
and
N N. Then (X, ,
) is a complete measure space such that
11
X \ A = (X \ (A B)) (B \ (N A)) .
If An = An Nn where An and Nn N for all n N. Hence
n An = (n An ) (n Nn ) .
Indeed, if A N = A0 N 0
Let us now show that
is well defined on .
0
0
where A, A and N, N N where N B and N 0 B 0 for B, B 0
and (B 0 ) = (B) = 0 then
(A) (A0 B 0 ) (A0 ) and (A0 ) (A B) (A).
If H A N with
It is elementary to see that
is a measure over .
1.3
Outer measures.
12
(En )
(An ) + .
(Ak ) = (nk=1 Ak ) (n An )
(An ),
k=1
(Ei ) : A i Ei , Ei A, i N}.
i=1
X
n,k
(En,k )
X
n
(An ) + .
13
X
i
(A Ei )
(Ei ).
14
((E A1 ) (E A2 )
(E 0 A1 ) + (E 0 (X \ A1 ))
(E A1 ) + (E A2 )
E (A1 ) + E (A2 )
N
X
E (An ).
k=1
(N
k=1 (E Ak )) + (E (X \ k=1 Ak ))
N
X
(E Ak ) + (E (X \
k=1 Ak ))
k=1
(E Ak ) + (E (X \
k=1 Ak ))
k=1
= (E (
k=1 Ak )) + (E (X \ k=1 Ak )).
15
1.4
Extension of measures.
(Ai ) : A i Ai , Ai A, i N}.
i=1
(An ) (E) + .
Hence
(E A) + (E (X \ A)) (n (An A)) + (n (An (X \ A))
X
X
(An A) +
(An (X \ A))
n
X
n
(An A) +
(An (X \ A))
(An ) (E) + .
16
Using that i,n are finite we have that Mn is a monotone class (see (5) and
(6) in Proposition 1.2.2) which clearly contains A. Therefore 1,n = 2,n and
hence 1 = 2 on (A).
Now
is -finite because it extends .
Remark 1.4.1 Take X = [0, 1) and the algebra A given by finite unions
of intervals [a, b), 0 a b 1. Consider the measure on A such that
() = 0 and (A) = for any 6= A A.
We can get two different extensions to (A). The counting measure and
the measure 1 defined by 1 () = 0 and 1 (A) = for any 6= A (A).
They are different since ({1/2}) = 1 and 1 ({1/2}) = .
Theorem 1.4.2 Let A be an algebra on X and : (A) [0, ) a measure
such that there exists Xn A of finite measure such that X = n Xn . Then
for each > 0 and each A (A) with (A) < there exists B A such
that (A4B) < .
Proof: Using the uniqueness in Hahns theorem we have that
(A) = inf{
(An ) : A n An , An A}
N
Take now N N such that
k=N +1 (Ak ) 2 and define B = k=1 Ak A.
Clearly (A4B) (
k=N +1 Ak ) + (n An \ A) .
Lemma 1.4.3 Let be a measure on and (A) = inf{(E) : A E, E
}. If A P(X) then there exists E such that A E and (E) = (A).
Proof: In the case (A) = take E = X. If (A) < we can select,
for each n N, En such that (En ) < (A) + n1 . Then E = n En
and (E) = (A).
Theorem 1.4.4 Let (X, , ) be a finite measure space. Then the com
pletion (X, ,
) coincides with (X, , ).
17
1.5
Borel-Stieltjes measures on R.
Throughout this section I denotes an interval in R with not empty interior, int(I) 6= , and let x0 int(T ).
Definition 1.5.1 Given a measure on B(I) which is finite on bounded
intervals contained in I we define F : I R as the function F (x) =
((x0 , x]) for x I, x x0 and F (x) = ((x, x0 ]) for x I, x < x0 .
Proposition 1.5.2 If : B(I) [0, ] is a measure such that (J) <
for all bounded interval J I, then F is increasing, right continuous and
F (x0 ) = 0.
Moreover F is continuous at x if and only if ({x}) = 0.
18
k=1
Our aim is, given a distribution function F to find a measure such that
F = F . For such a purpose we first need the following lemma.
Proposition 1.5.4 Let F : I R an increasing function. If we have(a, b]
nk=1 (ak , bk ] I where a1 a2 ...an and a, b, ak I for k = 1, ..., n, then
F (b) F (a)
n
X
F (bk ) F (ak ).
k=1
Proof: We may assume that (a, b] (ak , bk ] 6= for all k and,since (a, b]
must be contained in one of its connected components, also assume that
nk=1 (ak , bk ] = (ak0 , bk0 ], .
19
F (an1 ) F (bn0 )
F (bk ) F (ak )
k6=n0 ,n1
and therefore
F (b) F (a) F (bn1 ) F (an0 )
F (bn1 ) F (an1 ) + F (an1 ) F (bn0 ) + F (bn0 ) F (an0 )
n
X
F (bk ) F (ak ).
k=1
k=1
Hence
F (n An )
X
k,n=1
F (bnk ) F (ank )
X
n=1
.
2n
F (An ) + .
20
(ii) Clearly if A = (a, b] then F (A) F (b) F (a). Assume now that
(a, b] k (ak , bk ] where ak , bk int(I).
Given > 0, since F is right continuous there exists a0 > a so that
F (a0 ) F (a) < 2 and there exist b0k > bk such that F (b0k ) F (bk ) < 2k+1
for
all k .N.
0
Since [a0 , b] k (ak , b0k ), using compactness, we have [a0 , b] N
k=1 (ak , bk )
for some N N. Now we can apply Proposition 1.5.4 to get
F (b) F (a0 )
N
X
k=1
F (b0k ) F (ak )
N
X
F (bk ) F (ak ) + .
2
k=1
Ak .
Observe now that if (a, b] (ak , bk ] 6= then coincides with the interval
(max(a, ak ), min(b, bk )]. We have four situations, namely
Case a ak and b bk .
Hence (a, b] (ak , bk ] = (ak , b] and (I \ (a, b]) (ak , bk ] = (b, bk ] what
implies that
Ak = F (b) F (ak ) + F (bk ) F (b) = F (bk ) F (ak ),
Case a ak and bk < b.
Hence (a, b] (ak , bk ] = (ak , bk ] and (I \ (a, b]) (ak , bk ] = what implies
that
Ak = F (bk ) F (ak )
Case ak < a and b bk .
Hence (a, b] (ak , bk ] = (a, b] and (I \ (a, b]) (ak , bk ] = (ak , a] (b, bk ]
what implies that
Ak = F (b) F (a) + (F (a) F (ak ) + F (bk ) F (b)) = F (bk ) F (ak ),
Case ak < a and bk < b.
21
Hence (a, b] (ak , bk ] = (a, bk ] and (I \ (a, b]) (ak , bk ] = (ak , a] what
implies that
Ak = F (bk ) F (a) + F (a) F (ak ) = F (bk ) F (ak ).
Therefore
F ((a, b] E) + F ((I \ (a, b]) E)
X
F (bk ) F (ak )
k
< F (E) +
22
Let us now characterize the Legesgue measure on R by means of its invariance under translations.
Theorem 1.5.8 Let be a Borel measure over R which is invariant under
translations and finite over bounded intervals in R. Then = cm where
c = (0, 1]) and m is the Lebesgue measure restricted to B.
Proof: Let us first see that ((a, b]) = c(b a) where c = ((0, 1]).
Since (a, b] = a + (0, b a] it suffices to see that ((0, x]) = cx for any
x > 0. Since x = limn qn where qn Q, and (qn ) is an increasing sequence,
then it is enough to prove that ((0, q]) = cq for any q > 0 and q Q.
where m, n N and observing that
Now writing q = m
n
(0,
m
1
1 2
m1 m
] = (0, ] ( , ] ... (
, ]
n
n
n n
n
n
k1 k
1
we have that ((0, m
]) = m
k=1 (( n , n ]) = m(0, n ]). Finally to see that
n
(0, n1 ]) = nc simply note that 1 = nn and then ((0, 1]) = n(0, n1 ]).
Now let A be the algebra generated by {(a, b] : a b} which is easily
seen to coincide with finite unions of intervals (a, b] or (a, ) where
a b < . Using that R = (n, n] and (n, n] = 2nc we can apply
the uniqueness of the Hanh theorem to conclude that = cm (because both
measures coincide on A).
).
2k
23
Hence
(G) = m(G)
X
k
b k ak +
< m(E) + .
2
1.6
24
m([0, 1] \ C) =
X
k
m(Jk ) =
X 2X
k
l=1
m(Ik,l ) =
X
k
2k1
1
= 1.
3k
25
Let us now show that E is not Lebesgue measurable. In case it is measurable one has
X
1
m(E + r) 3,
rQ[1,1]
m1
X
k=n 0x1
m1
X
1
k=n 0x 3k
m1
X
1
2k+1
k=n
1
1
.
6 k=n 2k
3
1
( )k k+1 |
2 3
26
= 2
X 2X
k
m(ck,l + Ik,l )
l=1
k1
= 2
X 2X
k
m(Ik,l )
l=1
= 2 m([0, 1] \ C) = 1.
We then have that g(C) is a Borel set (using the fact, to be proved in
Exercise 1.7.5, that g(B([0, 1])) B([0, 2])) and m(g(C)) > 0. So we can find
E
/ M([0, 2]) with E g(C). Now consider A = g 1 (E)). This is a Lebesgue
measurable set (since A C) but A
/ B([0, 1]) (since g(A) = E
/ B([0, 2])).
1.7
Exercises
Exercise 1.7.1 (i) Let R be a ring. Show that (R, 4, ) is a ring in the
algebraic sense.
(ii) Let M be a non-empty family of sets in X and R(M) the ring generated by M. Show that any set in R(M) can be covered by a finite union of
sets in M.
Exercise 1.7.2 Let f : X Y be a function, A an algebra over Y . Let
(A) denote the -algebra generated by A and f 1 (A) = {f 1 (A) : A A}.
Then (f 1 (A)) coincides with f 1 ((A)).
Exercise 1.7.3 Let (X, , ) be a measure space and let (An ) be a sequence
in such that Aj intersects at most one other set in the sequence. Show that
(
j=1 Aj )
X
n=1
(Aj ) 2(
j=1 Aj ).
1.7. Exercises
27
Exercise 1.7.4 Study whether or not the following set functions are outer
measures.
(i) Let X = (ai,j ) be a matrix in M10 and define (A) = card{j : aij A}.
.
(ii) Let X = N and define (A) = limsup card(A{1,2,...,k})
k
(iii) Let X = Z and define (A) = 0 for A = ,
(A) =
a
if A is finite, where a = sup{|n| : n A} ,
a+1
(A) = 1 if A is infinite.
(iv) Let X be a metric space with distance d and let > 0. Define
= sup>0 inf{
k=1
+ (log(x) 1)[1,) .
Exercise 1.7.8 Let F (x) = 1
x (0,1)
(i) Find an unbounded Borel set A with 0 < mF (A) < .
(ii) Find an open set G such that 0 G0 and mF (G) < .
Exercise 1.7.9 Let mF be the Lebesgue-Stieltjes measure (0, ) associated
n2 +1
to F (x) = x for some > 0 and let A =
n=1 (n, n ).
Find the values of so that mF (A) < .
28
Chapter 2
Measurable and Integrable
functions
2.1
Measurable functions
30
n +1
n2X
k=1
k1
En,k .
2n
31
Let us show first that sn (x) is increasing for all x X. Fix n N and
x X.
If f (x) n + 1 then sn (x) = n sn+1 (x) = n + 1.
If n f (x) < n + 1 then sn (x) = n and sn+1 (x) = 2k1
n+1 for some
n+1
k n2
+ 1. Hence sn (x) sn+1 (x).
j
, k ) [ 2j1
If f (x) < n then there exist k, j so that f (x) [ k1
n+1 , 2n+1 ).
2n 2n
Hence either j1
= k 1 or j1
= k. In the first case sn (x) = k1
= 2j1
n+1 =
2
2
2n
j3
j1
k1
sn+1 (x) and in the second case sn (x) = 2n = 2n+1 2n+1 = sn+1 (x) .
Let us now show that lim sn (x) = f (x).
n
The case f (x) = we have sn (x) = n for all n N.
Assume f (x) < . Take n0 = n0 (x) N so that f (x) n0 . For all
n > n0 we have that there exists k {1, 2, .., n2n } such that f (x) En,k .
Hence for all n n0 we have
f (x) sn (x) = f (x)
1
k1
<
.
2n
2n
Note that if f is bounded n0 is the same for all x and the previous estimate
holds uniformly in x X. This finishes the proof.
Definition 2.1.5 Let (X1 , 1 ) and (X2 , 2 ) be measurable spaces. A function f : X1 X2 is called (1 , 2 )-measurable if f 1 (A) 1 for all A 2 .
For 2 = B(X2 ) where (X2 , ) is a topological space, a function f : X1
X2 is (1 , B(X2 ))-measurable (usually called simply measurable) if and only
if f 1 (A) 1 for all A .
If X and Y are topological spaces, f : X Y is said to be Borel measurable if f 1 (G) B(X) for all open set G X.
For X = Y = R a function f : R R is called Lebesgue measurable if it
is (M, B)-measurable, or equivalently f 1 ((a, b]) M for all a < b.
Remark 2.1.2 Given a measurable space (X, ) and a function f : X
[0, ]. Considering [0, ] as the compactification of the topological space
R+ then f is measurable (according to Definition 2.1.1) if and only if f is
(, B([0, ])-measurable (according to Definition 2.1.5).
Equivalently f is measurable if and only if X = {x X : f (x) = }
and f |X\X is (, B(R+ ))-measurable.
Let us establish some simple results about measurability on composition
of functions.
32
Lemma 2.1.6 (i) Let (X1 , 1 ), (X2 , 2 ) and (X3 , 3 ) be measurable spaces
and functions f : X1 X2 and g : X2 X3 . If f is (1 , 2 )-measurable
and g is (2 , 3 )-measurable then g f is (1 , 3 )-measurable.
(ii) Let (X1 , 1 ) be a measurable space,let X2 and X3 be topological spaces.
If f : X1 X2 is measurable and g : X2 X3 is continuous then g f is
measurable.
(iii) Let (X1 , 1 ) and (X2 , 2 ) be measurable spaces and let Y be a topological space. If f1 : X1 Rn and f2 : X1 Rm are measurable and :
Rn Rm Y is continuous then h : X Y given by h(x) = (f1 (x), f2 (x))
is measurable.
Proof: (i) and (ii) are immediate.
(iii) Consider : X Rn Rm given by (x) = (f1 (x), f2 (x)) and note
that
1 ((a1 , b1 ] ... (an , bn ] (a01 , b01 ] ... (a0m , b0m ])
= (f1 )1 ((a1 , b1 ] ... (an , bn ]) (f2 )1 ((a01 , b01 ] ... (a0m , b0m ]).
Since any open set in Rn Rm = Rn+m is numerable union of sets (a1 , b1 ]
... (an , bn ] (a01 , b01 ] ... (a0m , b0m ] we have that is measurable and then
h = too.
Let us recollect several operations on functions which are stable under
measurability.
Proposition 2.1.7 Let (X; ) be a measurable space and let 0 and
f, g, {fn } be measurable functions.
Then (with the convections 0 = 0, 1/0 = and 1/ = 0) f , f + g,
f.g, max{f, g}, min{f, g} 1/f , supn fn , inf n fn , lim supn fn and lim inf n fn
are measurable functions.
Proof: Write X = f 1 ({}), X 0 = g 1 ({}), f1 = f |X\X and
g1 = g|X\X 0 Hence
0
{f + g > } = X X
{f1 + g1 > },
{f > } = {f > }.
33
1
1
> } = {f = 0} {0 < f < }.
f
{max(f, g) } = {f } {g }.
{min(f, g) > } = {f > } {g > }.
{sup fn } = n {fn }.
n
34
(rn tn )2 + (pn qn )2
(rn + tn )2 + (pn + qn )2
2.2
35
(u+ + u )2 + (v + + v )2
u2 + v 2 = |f |.
36
X
n=1
(sn sn1 ) =
ci Ei
i=1
for all i N.
where ci > 0 and Ei
We now choose Ai , Bi so that Ai Ei Bi and (Bi \ Ai ) = 0.
P
Define g =
i=1 ci Ai . It is measurable with respect to . If N = i (Bi \ Ai )
we have that f (x) = g(x) for x X \ N (since x
/ N implies that for each i
one gets x
/ Bi or x Ai ) and (N ) = 0.
The case f : X C follows from the previous one in the usual way and
it is left to the reader.
Assume now f = g -a.e. for some g : X [0, ] (respect. g : X C)
measurable with respect to . Using that g is also measurable with respect
and Proposition 2.2.2 we get that f is measurable with respect .
37
n xB
n xA
/
Remark 2.2.3 The almost uniform convergence is weaker than the uniform
convergence.
It suffices to take a sequence uniformly convergent to a function and modify the limit function in a set of measure cero to get an almost uniform limit
which is not uniform.
Remark 2.2.4 The pointwise convergence is weaker that the almost uniform
convergence.
Take fn = [n,n+1] in (R, M, m). Of course fn (x) converges pointwise to
zero, but for any set A M with m(A) = 0 we have that for each n N there
is xn [n, n + 1]
/ A (otherwise there exists k N such that [k, k + 1] A).
Hence supxA
/ |fn (x)| supn |fn (xn )| = 1.
Theorem 2.2.6 (Egorovs theorem) Let (X, , ) be a finite measure space.
Let fn : X [0, ] be finite a.e. measurable functions (i.e. ({fn = }) =
0) and let f (x) = limn fn (x) a.e is measurable and finite a.e. (i.e.({f =
}) = 0 and there exists C , (C) = 0 such that limn fn (x) = f (x), x
/
C). Then for any > 0 there exists A with (A) < such that fn
converges to f uniformly in X \ A.
Proof: We can assume that fn , f take values in R+ and that fn (x)
converges to f (x) for all x X. In other case, we put An = {fn = },
B = {f = } and C with (C) = 0 such that lim fn (x) = f (x) if x
/ C.
We work in X1 = X \ N where N = (n An ) B C.
Let us first prove that , > 0 there exists A, and n, N such
that (A, ) < and |fm (x) f (x)| < , for all x
/ A, and m n, .
Indeed, write
An () = {x X : |fm (x) f (x)| for some m n}.
38
2.3
Integrable functions.
sd =
n
X
i (Ei )
i=1
sd =
m
X
j=1
j (Fj ).
39
Proof: Assume first that Fj are pairwise disjoint. For each i 1, ..., n,
consider Mi = {j {1, ..., m} : j = i } Hence Mi are pairwise disjoint,
{1, ..., m} = ni=1 Mi and Ei = jMi Fj . Therefore
n
X
i (Ei ) =
i=1
=
=
n
X
i=1
n
X
(Fj )
jMi
j (Fj )
i=1 jMi
m
X
j (Fj )
j=1
sd = (1 + 2 + 3 )(F1 F2 F3 )
+
+
+
+
=
(1 + 2 )((F1 F2 ) \ F3 )
(1 + 3 )((F1 F3 ) \ F2 )
(2 + 3 )((F2 F3 ) \ F1 )
1 (F1 \ (F2 F3 )) + 2 (F2 \ (F1 F3 )) + 3 (F3 \ (F1 F2 ))
1 (F1 ) + 2 (F2 ) + 3 (F3 ).
40
n
X
s =
i Fi \(j6=i Fj )
i=1
i6=j
F0 0
1 ,i2 ,i3 ) i1 ,i2 ,i3
+ ....
n
X
+ (
i )F1 ...Fn .
i=1
Now using the result for disjoint sets and adding up the measures the
result is complete.
Definition 2.3.3 Let (X, , ) a measure space and let f : X [0, ] be a
measurable function. We define
Z
f d = sup{
sd : 0 s f, s simple }.
sd =
(iv)
R
X
n
X
i (E Ei ) = sup{
i=1
f d =
R
X
f d.
td : 0 t s : t simple }.
41
f d = lim
n X
fn d.
fn d
Z
X
fn+1 d
f d.
An
An
and lim
(An Ei ) = (Ei ) and therefore limn An csd = X csd we get
R n
that
X csd M . Since this holds for all s f and 0 < c < 1 we have
R
X f d M .
R
(lim
inf fn )d =
n
( lim gn )d = n
lim
X n
Z
= lim inf
n
gn d
XZ
gn d lim inf
n
fn d.
42
R
(v) IfR (E) = 0 then E f = 0.
(vi) X f d = 0 if and only if f = 0 -a.e.
(vii) If f is -integrable then ({x X : f (x) = }) = 0.
Proof: (i) follows from Proposition 2.3.2 for simple functions. The general
case then follows by combining Theorem 2.1.4 and the Lebesgue monotone
convergence theorem.
(ii) Note that f EF = f E + f F and apply (i).
(iii) For each s simple function with 0 s f one has 0 s g. Then
the result follows from the definition.
(iv) Note that E f E where E = {x X : f (x) > } and
integrate.
R
P
(v)R Let s simple with 0 s f . Then E sd = ni=1 i (E Ei ) = 0.
Then E f d = 0.
(vi) Assume f = 0 -a.e. Hence ({x : f (x) > 0} = 0 and then, using
(iv) we get
Z
Z
Z
f d = 0
f d +
f d =
{f >0}
{f =0}
f d
Z
{f > n1 }
0
f d >
1
1
({x : f (x) > }) > 0.
n0
n0
43
X n=1
fn )d =
Z
X
n=1 X
Pn
k=1
fn d.
fk .
(i) is a measure on .
(ii) is finite if and only if f is -integrable.
(iii) If (E) = 0 then (E) = 0.
(iv) If is finite then lim (E) = 0.
(E)0
Z
X
Z
X
f En d =
n=1 X
f En d =
(En ).
n=1
(ii) It is obvious.
(iii) follows from (iv) in Proposition 2.3.7.
(iv) Note that lim (E) = 0 means that for all > 0 there exists > 0
(E)0
Now for s =
Pn
i=1
Z
E
Take < 2 Pn
i=1 i
f d
sd +
i Ei we have that
sd =
n
X
n
X
i (Ei E) (
i=1
i )(E).
i=1
R
E
f d < .
44
Definition 2.3.10
Let f : X C be a measurable function. f is said to be
R
-integrable if X |f |d < .
Moreover if f = u + iv is -integrable we define
Z
f d =
u+ d
u d + i
v + d i
v d.
f d =
ud + i
vd.
f d =
f1 d
Z
X
f2 d.
f + d +
Therefore
f2 d =
f d =
f d +
Z
X
f1 d
Z
X
f1 d.
f2 d.
(f + g)d =
f d +
gd.
|f + g|d ||
|f |d + ||
|g|d < .
It suffices to show the result for real valued functions. The case of complex
values follows immediately from the previous one using that if f = u + iv
then
Z
Z
Z
f d =
ud + i vd.
X
45
(f + g)d =
+ i
(au bv)d +
X
Z
(av + bu)d + i
(a0 v 0 + b0 u0 )d
ud b
ZX
+ ia
(a0 u0 b0 v 0 )d
XZ
= a
vd + ib
vd + a
u d b
ud + ia
= (a + ib)
v 0 d
vd + ib
ud
(u0 + iv 0 )d.
(f ) = + f + f + .
These give
Z
(f )d =
+ f + + f d
ZX
+ f + f + d
= +
(f + f )d
(f + f )d
f d.
(f + g)d =
f d +
gd.
R
X
f d|
R
X
|f |d.
46
f d| = |
f d
f d|
f d| =
f d =
f d +
f d =
|f |d.
X
|z|
z
f d = z C \ {0}. Take =
f d =
<(f )d
|f |d =
and
|f |d.
n X
In particular limn
R
X
fn d =
|fn f |d = 0.
R
X
f d.
Z
X
hn d =
Z
X
|fn f |d.
47
Definition 2.3.16 We define the equivalence relation f g if f = g a.e. and we denote by L1 () the set of equivalence classes of -integrable
functions.
R
Let us denote ||f ||1 = X |f |d.
Proposition 2.3.17 (L1 (), ||.||1 ) is a Banach space.
Proof: Let us first show that L1 () is a vector space.
This actually shows that ||f + g||1 ||f ||1 + ||g||1 .
Obviously ||f ||1 = ||||f ||1 .
R
The other property to get a norm is that X |f |d = 0 implies that f = 0
-a.e.
P
Let us show the completeness. It is equivalent to show that
n=1 ||fn ||1 <
P
implies that the series n=1 fn converges in L1 ().
R P
P
Now, using that
n=1 |fn |d < we get that the series
n=1 ||fn ||1 = X
P
|f
|
<
is
convergent
-a.e,
and
hence there exists A such that
n=1 n
P
for all x
/ A.
n=1 fn (x) is convergent
P
/ A andR f (x) = 0 for x A. Observe
Define f (x) = n=1 fn (x) forR x
P
first that f is -integrable, since X |f |d X
n=1 |fn |d < .
Now
lim
(
n
n
X
fk (x) f (x)) = 0, x
/ A and
k=1
n
X
k=1
n X
n
X
k=1
fk f |d = lim
n
X
n X\A
k=1
fk f |d = 0.
48
implies E |f |d < .
R
(ii) For all > 0 there exists B with (B) < and X\B |f |d < .
Proof:
(i) Note that, using the dominated Lebesgue theorem, for each > 0
there exist a simple function s such that
|f s|d .
2
X
Now for s =
Pn
i=1
i Ei we have that
|s|d
n
X
i=1
|f |d
X\B
Z
X\B
|f sn0 |d < .
Definition 2.3.20 A sequence {fn } of -integrable functions is called equiintegrable if for every > 0
(a) there exist > 0 such that implies that
(E) < = sup
nN
Z
E
|fn |d <
Z
X\B
|fn |d < .
49
X\B
|fn |d < ,
sup
Z
B
|f fn |d
Z
B\A
Z
B\A
|f fn |d +
Z
A
|f fn |d
|f fn |d + sup
n
Z
A
|f fn |d
Applying Fatous Lemma again and (A) < we can say that
Z
B
|f fn |d
Z
B\A
|f fn |d + 2.
Using that limn supxB\A |fn (x) f (x)| = 0 we have |fn f | M B\A
for some constant
M > 0. So we can use the bounded convergence theorem
R
and get limn B\A |f fn |d = 0.R
Finally this shows
that limRn X |f fn |d =
0. This also gives that f is
R
R
-integrable Rsince X |f |d X |f fn |d + X |fn |d < after choosing
n such that X |f fn |d < 1.
(ii) =(i) Note that
|
Z
E
fn d|
Z
E
|fn f |d + |
f d|
for all n
N and E . Hence for each > 0 there
exists n0 N, 0 > 0 for
R
R
which X |fn f |d < /2 for all n n0 and | E f d| < /2 if (E) < 0 .
50
sup
nn0
Z
X\B0
|fn |d sup
nn0
and
sup
n
2.4
Z
X\B
Z
X\B0
|f fn |d + /2 <
|fn |d < .
Exercises
2.4. Exercises
51
52
g d =
gd().
fk d n
Z
X
(supkN fk )d.
xR\Q} .
2.4. Exercises
53
e n x 2 log xdx.
n 0
Z 1
n
sen(
))dx.
d) lim
n 0
2n + x
Z n
x
e) lim
eax (1 + )n dx, (a > 0).
n 0
n
1
t
1X
f ) lim+
arctag .
t0 t
n
n=1 n
c) lim
1
f
1
),
|f |2 + |g|2 , 1+|g|
,
f 2 , f 3 , arctag f , |f | + |g|, f.g, f g, sen( 1+|f
|
|f | ( R).
Exercise 2.4.21 Let (X, , ) be a measure space and let f : X [0, ] be
integrable. Show that {x X : f (x) > 0} is a countable union of sets of
finite measure.
Exercise 2.4.22 Give an example where the inequality of Fatous lemma is
strict.
Exercise 2.4.23 Let ((, ], B((, ]), m) be the Lebesgue measure space
over B((, ]). Consider T = {z C : |z| = 1}, B(T) and the measure
= (m) where (t) = eit . Show that
54
X
1
1 Z
xs1
=
dx.
(s) =
s
(s) (0,) ex 1
n=1 n
2
x dx =
nn .
n=1
Exercise 2.4.29 Show, using the image measure, the following well-known
change of variable result. Let X : R R be of class C 1 , strictly increasing
and bijective and let g : R R be integrable. Then
Z
R
(g X)(t)dt =
Z
R
2.4. Exercises
55
Chapter 3
The product measure and
Fubinis theorem
3.1
58
59
Z
X
Case 2.- E =
nk=1 Ak
(Ex ) =
(Ex )d =
Z
Y
n
X
(E y ) =
n
X
n
X
k=1
k=1
and
(E y )d.
(Ak )(Bk ) =
k=1
Z
X
(Ex )d =
(E y )d.
In general, define
= {E 1 2 : satisfies (i) and (ii)}.
We shall prove that is a monotone class. Since it contains A we will have
M(A) = = 1 2 and the proof will be finished.
Let {En } be an increasing sequence in . Taking into account that
((En )x ) (respect. ((En )y )) are increasing sequences of 2 -measurable (respect. 1 -measurable) functions and converges to ((n En )x ) (respect. to
((n En )y )) we get that n En verifies (i).
Using also the monotone convergence theorem we have that
lim
n
Z
X
(Ex )d =
Z
X
((n En )x )d,
60
(E )d =
((n En )y )d.
lim
n
lim
n
Hence
R
Y
(Ex )d =
y
(E )d =
Z
X
((n En )y )d =
R
X
((n En )x )d,
((n En )y )d.
((n En )x )d and n En .
Corollary 3.1.7 Let (X, 1 , ) and (Y, 2 , ) be -finite measure spaces and
let E 1 2 . Then
(i) xR (Ex ) is R1 -measurable and y (E y ) is 2 -measurable.
(ii) X (Ex )d = X (E y )d.
Proof: Write X =
n=1 Xn with Xn 1 for all n N, (Xn ) < and
0
Xn Xn = if n 6= n0 , and Y =
m=1 Ym with Ym 2 for all m N,
(Ym ) < and Ym Ym0 = if m 6= m0 .
Let n, m N and define n (A) = (A Xn ) and m (B) = (B Ym ).
Since (X, 1 , n ) and (Y, 2 , m ) are finite measure spaces, then x m (Ex )
is 1 -measurable and y n (E y ) is 2 -measurable and also
Z
Xn
(Ex Ym )d =
Z
Ym
(E y Xn )d.
y
y
Now (Ex ) =
n=1 n (E ) and hence x
m=1 m (Ex ) and (E ) =
y
(Ex ) is 1 -measurable and y (E ) is 2 -measurable.
Moreover
Z
X
(Ex )d =
=
Z
X
m=1 X
X
m (Ex )d
Z
m=1 n=1 Xn
(Ex Ym )d
61
X
Z
X
n=1 m=1 Ym
Z
X
y
n=1 Y
(E y Xn )d
(E Xn )d
n (E y )d.
Z
X
(Ex )d =
(E y )d
for all E 1 2 .
Then (X Y, 1 2 , ) is a -finite measure space.
Remark 3.1.1 (A B) = (A)(B) for all A 1 and B 2 .
Remark 3.1.2 Let E 1 2 . Then
Z
XY
E (x, y)d =
Z
X
Ex d)d(x) =
Z
Y
E y d)d(y).
Remark 3.1.3 The -finiteness is necessary for the iterated integrals to coincide.
Indeed, let X = Y = [0, 1], = m be the Lebesgue measure and let be
the counting measure. Take E = {(x, x) : 0 x 1}. Note that Ex = {x}
and E y = {y},R and hence (E
) = 1 and (E y ) = 0, for all x, y [0, 1].
R x
Therefore 1 = X (Ex )d 6= Y (E y )d = 0.
Definition 3.1.9 In the case X = Y = R, 1 = 2 = B(R) and =
= m the Lebesgue measure on B(R), we define the Lebesgue measure on
B(R) B(R) = B(R2 ) as the product measure m m and it is denoted m2 .
Inductively mn = m mn1 is defined on B(Rn ) for all n N. It follows
from the unicity of the Hahn theorem that mn = mk ml for any n = k + l.
Remark 3.1.4 The space (R2 , B(R2 ), m2 ) is a -finite not complete measure
space.
Indeed, let E and C be the Vitali and the Cantor sets in [0, 1] respectively.
Then E C
/ B(R2 ) (due to the fact (E C)y = E for any y C) but
E C [0, 1] C and m2 ([0, 1] C) = 0.
62
(3.1)
Any automorphism in Rn can be decomposed into products of automorphisms of the following types:
(i) (T (e1 ), ..., T (en )) is a permutation of (e1 , ..., en ),
(ii)T (e1 ) = e1 and T (ei ) = ei for i = 2, ..., n, or
(iii) T (e1 ) = e1 + e2 and T (ei ) = ei for i = 2, ..., n.
Hence it suffices to show (3.1) for these cases.
In the first one T (Qn ) = Qn and det(T ) = 1, in the second case T (Qn ) =
[0, ) Qn1 and det(T ) = and in the third case T (Qn ) = {(x1 , x1 +
x2 , x3 , ..., xn ) : 0 xi < 1} and det(T ) = 1. Since T (Qn ) = A2 Qn2 where
clearly m2 (A2 ) = 1 we get the result.
Case 2: T is not bijection and A B(Rn ).
The image T (A) would lie in a proper subspace. Hence the result follows
by showing that any subspace has measure zero. Actually, due to the previous
case, by composing with automorphims we may assume that T (A) {x
Rn : x1 = 0} = {0} Rn1 which is clearly of measure zero.
63
Case 3: A M(Rn ). Now the measure and |det(T )|mn are -finite
measures such that coincide on the generating family and then coincide on
B(Rn ). Now their complections must coincide and the result extends then
to Lebesgue measurable sets.
3.2
Fubini theorem
Theorem 3.2.1 Let (X, 1 , ) and (Y, 2 , ) be -finite measure spaces and
let f : X YR [0, ] be 1 2 -measurable.R Then
(i) xR Y fx d is 1R-measurable
and Ry
f y d is 2 -measurable.
R
R Xy
(ii) XY f d = X ( Y fx d)d = Y ( X f d)d.
Proof: For f = E this coincides with Corollary 3.1.7.
P
P
For simple functions f = ni=1 i Ei we have fx = ni=1 i ((Ei )x ) and
P
f y = ni=1 i ((Ei )y ). Hence f y and fx are 1 and 2 -measurable respectively. Moreover
Z
f d =
XY
coincides with
n
X
and
n
X
((Ei )x )d =
i=1
i (Ei )
i=1
i=1
n
X
((Ei ) )d =
fx d)d
f y d)d.
Z
Y
(sn )x d)d =
y
(sn ) d)d =
Z
Y
fx d)d,
f y d)d,
64
Z
XY
sn d =
f d .
XY
f d =
XY
fx d)d =
f y d)d < .
Hence
f d < -a.e. and X f y d < -a.e. and x Y fx d and
R Yy x
y X f d are and -integrable respectively.
The general case follows from the previous one applied to the decompositon f = u+ u + iv + iv .
R
Corollary 3.2.3 Let (X, 1 , ) and (Y, 2 , ) be -finite measure spaces and
let f : X Y C be 1 2 -measurable.
R R
If X ( Y |f (x, y)|d(y))d(x) < then f is -integrable and
Z
f d =
XY
Z
X
fx d)d =
Z
Y
f y d)d.
Proof: Apply Theorem 3.2.1 to the function |f | to get that it is integrable and then apply Theorem 3.2.2.
3.3
Applications
q
3.3. Applications
65
f (x)dmn (x) =
(0,)
Sn1
r2
= (
r1
66
f dmn =
Rn \{0}
m
X Z
i=1
m
X
i=1
Z
Pm
i=1
i Ei
f dmn
Ei dmn
Rn \{0}
n1
Sn1
n1
Ei (ru)dn1 (u))dr
f (ru)dn1 (u))dr
Sn1
f dmn = nvn
(r)rn1 dr
where vn = mn (Bn ).
Proof: Applying the previous theorem we get
Z
Rn
f dmn =
n1
Z
Sn1
Proposition 3.3.4 ( 12 ) =
R
R
t2
f (ru)dn1 = nm(Bn )
et dm1 (t) =
R
0
dm1 (t))
(r)rn1 dr
R
0
R2
= 2v2
= v2
t1/2 et dt = 2
Z
Z
0
rer dr
er dr = v2 =
3.3. Applications
67
n/2
.
( n
+1)
2
Proof: Write
Bn = {x Rn : ||x|| < 1} = {(t, x0 ) RRn1 : ||x0 || < 1, |t| <
1 ||x0 ||2 }.
Hence
vn =
=
Z
||x0 ||<1
||x0 ||<1
= (n 1)vn1
= (n 1)vn1
rn2 2 1 r2 dr
n3
2
1 tdt
n1 3
, )
2
2
3
)(
)
( n1
2
2
= (n 1)vn1
n
( 2 + 1)
= (n 1)vn1 B(
)( 32 )
)( 32 ) ( n2
)( 32 )
( n1
( n1
2
2
2
=
(n
1)(n
2)v
.
n2
( n2 + 1)
( n2 + 1) ( n1
+ 1)
2
)k ( 32 )
( nk
2
,
( n2 + 1)
n/2
.
( n
+1)
2
Definition 3.3.6 Let (X, , ) be a -finite measure space. Given a measurable function f : X [0, ] we define the distribution function of f by
f (t) = ({x X : f (x) > t}).
68
(f (x))d(x) =
Z
0
0 (t)f (t)dt.
ZX
(0,)
Hence
Z
(f (x))d(x) =
Z
X
(0,f (x))
(t)dm(t))d =
Z
0
f (t)0 (t)dt.
Corollary 3.3.8 LetR (X, , ) beR a -finite measure space, 1 p < and
f measurable. Then X |f |p d = 0 ptp1 f (t)dt.
3.4
Exercises
Exercise 3.4.1 Let X, Y be non empty sets, and let M P(X) y R P(Y )
such that X M y Y R. Show that (M R) = (M) (R).
Exercise 3.4.2 Let (X, 1 , ) and (Y, 2 , ) be -finite measure spaces. De e Y their complections and by X Y
the completion of X Y with
note by X
Y = X Y
.?
respect the product measure. Does it hold that X
Exercise 3.4.3 Let (X, 1 , ) and (Y, 2 , ) be -finite complet meausure
2 , )
Show that if A 1
= 0,then (Ax ) = 0 -a. e. and
(Ay ) = 0 a.e.
1
2Deduce then that Fubinis theorem holds true for non negative
1
2 -integrable functions.
measurable functions or for
3.4. Exercises
69
[a,x]
f (x)G(x)d(x) = F (b)G(b)
[a,b]
F (x )g(x)d(x).
[a,b]
0 y < f (x)}
and F (y) = ({x E : f (x) > y}) , y > 0 (called the distribution function
f over E).
Show that, for the Lebesgue measure m over R, we have
Z
f d = ( m)(R(f, E)) =
F (y)dm(y)
f p d =
ptp1 F (t)dm(t).
(f )d =
0 (t)F (t)dt.
(f )d =
F (t)dm (t).
Exercise 3.4.7 Let (N, P(N), ) be the measure space for the counting measure . Let (X, , ) be a measure space. Define, for E P(N), the
measure
(E) =
X
n=1
(En ).
70
Z
X
f d( ) =
n=1 X
NX
fn d =
X n=1
fn d
1
2
and g(x) =
y 2 sin2 (x)
x2 (x2 +y 2 )(x2 +y 2 +m2 )
y 2 sen2 x
)1
||x||
Exercise 3.4.12 Let f : Rk \ {0} R be given by f (x) = ||x||k (1||x||)
.
R
Show thatR it does not exist {||x||<1} f (x)dx but it does exist the principal
value lim0 {<||x||<1} f (x)dx. Compute such a value.
Exercise 3.4.13 Find the values of for the following functions to ve integrable Rand compute the value of their integrals.
dx
(i) Rk (1+||x||
2 ) .
R
(ii) {||x||<r} ||x|| dx.
x21 x22 +x23 +...+(1)k+1 x2k
dx.
{||x||<1}
||x||
R
|x1 |+...+|xk |
(iv) {||x||<1}
dx.
||x||
(iii)
3.4. Exercises
71
Exercise
3.4.14 Find out the integrals Ik = Bk |x1 ...xk |dmk (x) and Jk =
R
k
Sk1 |u1 ...uk |dk1 (u), where Bk is the closed unit ball of R and Sk1 the
sphere ||x|| = 1.
R
Chapter 4
The Radon-Nikodym Theorem
4.1
Let us start with the following result to motivate the next definitions.
Proposition 4.1.1 Let (X, , ) be a measure space and f : X C be
-integrable.
R
Then (E) = E f d for E defines a complex-valued set function on
with the following properties:
P
(i) (
n=1 (En ) for any sequence {En } of pairwise disjoint
n=1 En ) =
measurable sets.
(ii) If (E) = 0 then (E) = 0.
(iii) lim (E) = 0, i.e. for all > 0 there is > 0 such that (E) <
(E)0
Z
X
f En d =
Z
X
n=1 X
f En d =
(En ).
n=1
(ii) It is obvious.
R
(iii) Given > 0 take a simple function s such that X |f s|d 2 .
73
74
Now for s =
|
i=1
i Ei we have that
sd|
Hence if <
n
X
i=1
2 supi=1,...n |i |
R
E
f d| < .
(
n=1 En ) =
(En ).
n=1
Let us see first that this last remark holds true for any complex measure,
that is there exists always a non negative measure such that |(E)| (E)
for all E .
Definition 4.1.3 Let : C be a complex measure and E , we
define the variation of on E, as
||(E) = sup{
n
X
|(En )| : E =
n=1 En , En , En Em = for n 6= m}.
n=1
75
|(En )|
n=1
X
n=1 m=1
X
|(En Am )|
|(En Am )|
m=1 n=1
||(Am )|
m=1
||(Ak ) <
|(En,k )| +
k=1 n=1
k=1
k=1
X
n=1
|(En )|
(En ) = (E).
n=1
76
n
X
|s (Ai E)| =
i=1
n
X
|i |(Ai E) =
i=1
Z
E
|s|d |s |(E).
Z
ZE
|f s|d +
|f s|d +
|s f |d +
|f |d
|s|d
ZE
|f |d
f ()d =
n
X
k=1
n
X
|zk |
|zk |
k=1
n
X
= 2
k=1
|zk |.
cos+ (k )d
cos+ ()d
77
f ()d 2f () = 2|
zk |,
kS()
|(En )| = . Let
n=1
X
kS
(Ek )| >
N
1X
|(Ek )| > 1 + |(E)|.
k=1
+
can be decomposed as = where (E) = E f + d and (E) =
R
E f d.
We shall see that this decomposition is actually true for any real measure.
R
78
(E) =
f d,
(E) =
f d.
X
n=1
|(An )| + =
X
nS
(An )
(An ) + = (F ) (G) + .
nS
/
79
= (Fn ) +
(Fn+k \ k1
j=0 Fn+j ).
k=1
1
2n+k
+ + (E) (Fn+k )
k1
= (Fn+k \ j=0
Fn+j ) + (k1
j=0 Fn+j )
k1
(Fn+k \ j=0
Fn+j ) + + (E)
1
Therefore 2n+k
(Fn+k \ k1
j=0 Fn+j ) and then
(
k=n Fk )
X
1
1
1
= (E + ) n1 .
(E ) n
n+k
2
2
k=1 2
+
80
4.2
k=n
Z
E
gd (E), E }.
81
Z
E
fn d =
Z
E{fn1 gn }
gn d +
Z
E{fn1 <gn }
fn1 d
g
n d X fn d X f d M for all n N, we
X
R
obtain M = X f d.
R
R
Our aim is to show that E f d = (E) for all E . Since (E) E f d
is a non-negative measure, it suffices to see that (X) = M .
Assume
that (X) > M . Using that (X) <R we get > 0 so that
R
(X) > X (f + )d. Define now (E) = (E) E (f + )d. We have that
is a real measure. Consider
A, B the Hahn decompositionRof the measure
R
. Therefore (E) E (f + )d, for E A and (E) E (f + )d for
E B.
Define g = f B + (f + )A . We have that g G. Indeed,
Z
E
gd =
f d +
EB
(f + )d (E B) + (E A) = (E).
EA
Remark 4.2.2 The Radon-Nikodym theorem does not hold without assumptions on the measures and .
Take the counting measure on [0, 1] and the Lebesgue measure onR [0, 1].
Since there are no -null sets besidesR we have << , but m(E) = E f d
for all E B([0, 1]) leads to f (x) = {x} f d = m({x}) = 0 for all x [0, 1].
Theorem 4.2.4 Let (X, , ) and (X, , ) be -finite spaces. Then <<
if and only if there exists aR measurable function f : X [0, ] (and -a.e.
unique) such that (E) = E f d for all E .
82
(n,m)N2
Z
E
fn,m dn,m .
(E) =
(E Xn,m )
(n,m)N2
(n,m)N2
Z
E
fn,m dn,m
fn,m Xn,m d
(n,m)N2
f d.
The uniqueness follows from the fact that f = g -a.e. in Xn,m for all
(n, m) N2 and hence f = g -a.e.
Theorem 4.2.5 Let (X, , ) be a -finite space and (X, , ) any measure
space. Then << if and only if there exists aR measurable function f :
X [0, ] (and -a.e. unique) such that (E) = E f d for all E .
Proof: We may assume that there is E0 such that (E0 ) < , otherwise f = . We first deal with (X) < . Define
C = {E : : E [0, ] -finite},
where E = {E A : A }. Observe that C 6= since E0 C.
Let S = sup{(E) : E C}. Note that S (X). Take En C with
limn (En ) = S. Consider X1 = n En . Clearly X1 C, hence (X1 ) = S
83
since (En ) (X1 ) S for all n N. Using Theorem 4.2.4 there exists
f1 : X1 [0, ] measurable with respect X1 such that
(A X1 ) =
Z
AX1
f1 d
for all A .
Let us define f (x) = f1 (x) for x X1 and f (x) = for x X \ X1 . We
have that f is -measurable and
(E) =
Z
EX1
f1 d + (E (X \ X1 )).
(E (X \ X1 )) =
f d =
E(X\X1 )
gd.
E(X\X1 )
Z
X\X1
f d =
gd.
X\X1
84
4.3
Applications
4.3. Applications
85
Theorem 4.3.3 (The Lebesgue decomposition theorem) Let (X, ) be a measurable space. If and are -finite measures then there exist two unique
measures a and s such that = a + s where a << and s .
Proof: Let us begin assuming (X) < . Using that << + we can
apply the Radon -Nikodym theorem and find a -integrable function f such
that, for all E ,
Z
Z
(E) =
f d + f d.
E
f d +
EB
f d =
EB
f d.
EB
(A) =
Z
A
f d +
f d (A) + (A).
86
4.3. Applications
87
Therefore, taking limits as k , one gets |fn (x) f (x)| /2 < for
x
/ n,kn0 Nn,k and n n0 .
This implies that ||fn f || < for n n0 and the proof is complete.
Theorem 4.3.7 Let (X, , ) be a finite measure space. Then the dual of
L1 () is isometrically isomorphic to L ().
Proof:
Let : L () (L1 ()) be defined by (f ) = f where f (g) =
R
X f gd.
Observe first that if f L () and g L1 () then f g L1 () since
|f g| ||f || |g|, which says that f is well defined. From the properties of
the integral, it is linear and verifies |f (g)| ||f || ||g||1 . Therefore f
(L1 ()) .
Also it follows from the properties of the integral that is linear and
from the previous estimate ||(f )|| ||f || .
Therefore is a continuous injective linear map with |||| 1.
Let us show that it is surjective and |||| = 1.
Assume first that (X) < . Given (L1 ()) we can define (E) =
(E ) for any E (since E L1 ()).
We first see that is a complex measure. Indeed, if {Ek } are pairwise
P
disjoint sets in then k Ek converges in L1 (). Hence
X
(k Ek ) = (
Ek ) =
X
k
(Ek ) =
(Ek ).
On the other hand | (E)| ||||(E), which implies that << . Using
now
the Radon-Nikodym theorem we obtain f L1 () such that (E) =
R
X E f d = (E ).
R
Note that | (E)| = | E f d| ||||(E) for all E . This gives
| |(E) ||||(E) for all E .
Let us define E0 = {x RX : |f (x)| > ||||}. Observe that (E0 ) = 0,
since otherwise | |(E0 ) = E0 |f |d > ||||(E0 ). This shows that f
L () and ||f || ||||.
R
Since is linear we get (s) = X sfR d for all simple functions s and by
approximation we actually
get (g) = X f gd for all g L1 (), due to the
R
facts that and g X f gd are both continuous maps. This shows that
(f ) = f = and that ||f || = ||f || = ||||.
88
4.4
Exercises
R
A
|x|dx.
Exercise 4.4.2 Let identify Q with (rn )nN and R define, for n N, fn :
R R the non-negative Borel function such that fn dx = 1 which vanishes
in the exterior of the closed interval of length 21n centered on rn . Let (A) =
R P
fn dx for a Borel set A.
A
P
i) Show that fn (x) < m-a.e. x R.
ii) Show that is -finite, << m and that every non empty open set A
verifies that (A) = .
Exercise 4.4.3 Let and be -finite measures over (X, A), such that
<< and let g be the Radon-Nikodym derivative of with respect to and
let f be A-measurable.
R
Show
that
f
is
-integrable
if
and
only
if
f
g
es
-integrable
and
f d =
R
f gd.
Exercise 4.4.4 Let X be a non numerable set, M the class of all numerable
or co-numerable sets in X and let be the counting measure. Let (E) = 0
for numerable sets E and (E) = otherwise.
Show that, although << , one cannot define the Radon-Nikodym
derivative in this case.
d
, g =
Exercise 4.4.5 Let , , be -finite measures and let f = d(+)
d
d
, F = d(++) .
d(+)
Justify their existence and get an expression of F in terms of f and g.
4.4. Exercises
89
90
Exercise 4.4.12 Let , be complex measures which are absolutely continuous with respect to a -finite measure . Show that for all a, b C one
has
d
d
d(a + b)
=a
+b .
d
d
d
Exercise 4.4.13 Let , , be -finite measures over (X, ) such that <<
and << . Show the following chain rule
d
d d
=
. .
d
d d
Exercise 4.4.14 Let , be -finite measures over (X, ) such that <<
y << . Show that
d
6= 0 a.e.
d
d
1
=
d
d/d
a.e.
4.4. Exercises
91
(i) Show that there exists a real measure over R such that
(A B) = (A)(B) for A and B R.
(ii) Find the Hahn decomposition of from the Hanh descompositions
of each factor.
(iii) Compute ( )+ , ( ) and | | in terms of those of y .
Exercise 4.4.18 Let = B([0, 1]) and (E) = m(E) + im(E [0, 21 ]).
(i) Describe || in terms of m.
(ii) Show that
(E) (Re)+ (E) + (Re) (E) + (Im)+ (E) + (Im) (E)
and that the inequality can be strict.
R
(iii) Find a Borel function h such that |h| = 1 and (E) = E hd|| for
all E .
Exercise 4.4.19 For each Borel set in R define
(E) =
Z
E(0,)
Z
sen3 t
sen3 t
dt
dt.
t3
t3
E(,0)