2

The pleasures of counting.

Additional Material

Problems
2.22. For A0j , A1j X, j N, we have
[
(A0j A1j ) =
jN

i(k)

Ak .

i=(i(k))kN {0,1}N kN

Solution to 2.22: Since for A, A0 , B, B 0 X we have the multiplication

rule
(A B) (A0 B 0 ) = (A A0 ) (A B 0 ) (B A0 ) (B B 0 )
and since this rule carries over to the infinite case, we get the formula
from the problem by multiplying out the countable union
(A01 A11 ) (A02 A12 ) (A03 A13 ) (A04 A14 ) .
More formally, one argues as follows:
[
x
(A0j A1j ) j0 : x A0j0 A1j0

(*)

jN

while
x

i(k)

Ak

i=(i(k))kN {0,1}N kN

i = (i(k))kN {0, 1}N : x

i(k)

Ak

kN
N

i(k )

i = (i(k))kN {0, 1} k0 N : x Ak0 0

(**)

Clearly, (*) implies (**). On the other hand, assume that (**) holds
but that (*) is wrong, i.e. suppose that for every j we have that either
x A0j or x A1j or x is in neither of A0j , A1j . Thus we can construct a
uniquely defined sequence i(j) {0, 1}, j N, by setting

0 if x Aj ;
i(j) = 1 if x A1j ;

0 if x 6 A0j and x 6 A1j .

1

Schilling: Measures, Integrals & Martingales

Define by i0 (j) := 1 i(j) the complementary 0-1-sequence. Then
[ i(j)
[ i0 (j)
x
Aj
but x 6
Aj
j

contradicting our assumption (**).

Measures.
Additional Material

Problems
4.16. Let (X, A, ) be a measure space, F A be a sub--algebra and denote by N =
{N A : (N ) = 0} the collection of all -null sets. Then

(F, N) = F 4 N : F F, N N
where F 4 N := (F \N )(N \F ) denotes the symmetric difference, cf. also Problems
2.2, 2.6.
4.17. Let A denote the completion of A as in Problem 4.13 and write N := {N X :
M A, N M, (M ) = 0} for the family of all subsets of A-measurable null
sets. Show that

A = (A, N) = A 4 N : A A, N N}.
Conclude that for every set A A there is some A A such that A 4 A N.
4.18. Consider on (Rn , B(Rn )) the Dirac measure x for some fixed x Rn . Find the
completion of B(Rn ) with respect to x .
4.19. Regularity. Let X be a metric space and be a finite measure on the Borel sets
B = B(X) and denote the open sets by O and the closed sets by F. Define a family
of sets
:= {A X : > 0 U O, F F s.t. F A U, (U \ F ) < }.
(i) Show that A = Ac and that F .
(ii) Show that is stable under finite intersections.
(iii) Show that is a -algebra containing the Borel sets B.
(iv) Conclude that is regular, i.e. for all Borel sets B B
(B) =

sup

(F ) =

F B, F F

inf

U B, U O

(U ).

(v) Assume that there exists an increasing sequence of compact (cpt, for short)
sets Kj such that Kj X. Show that satisfies
(B) =

sup

(K).

F B, K cpt

(vi) Extend (B) = supF B, F F (F ) to a -finite measure .

Schilling: Measures, Integrals & Martingales

4.20. Regularity on Polish spaces. A Polish space X is a complete metric space which
has a countable dense subset D X.
Let be a finite measure on (X, B(X)). Then is regular in the sense that
(B) =

sup
KB, K compact

(K) =

inf

U B, U open

(U ).

T Sk(n)
[Hint: Use Problem 4.19 and show that the set K := n j=1 K1/n (dk(n) ), K (x) is
a closed ball of radius and centre x and {dk }k is an enumeration of D, is compact.
Choosing k(n) sufficiently large, we can achieve that (X \ K) < .]
4.21. Find a measure on (R, B1 ) which is -finite but assigns to every interval [a, b)
with b a > 2, finite mass.

Solution to 4.16: Set

:= F 4 N : F F, N N .

and denote, without further mentioning, by F, Fj resp. N, Nj sets from

F resp. N. Since F 4 = F , 4 N = N and F 4 N (F, N) we get
F, N (F, N)

(*)

and the first assertion follows if we can show that is a -algebra. In

this case, we can apply the -operation to the inclusions (*) and get
(F, N) () ((F, N))
which is just
(F, N) (F, N).
To see that is a -algebra, we check conditions (1 )(3 ).
(1 ): Clearly, X F and N N so that X = X 4 ;
(2 ): We have, using de Morgans identities over and over again:
[F 4 N ]c = [(F \ N ) (N \ F )]c
= (F N c )c (N F c )c
= (F c N ) (N c F )
= (F c N c ) (Gc G) (N N c ) (N F )
= (F c N c ) (N F )
= (F c \ N ) (N \ F c )
= |{z}
Fc 4N
F

Chapter 4. Additional material. Last update: June 23, 2008

(3 ): We begin by a few simple observations, namely that for all F F

and N, N 0 N
F N = F 4 (N \ F ) ;
| {z }

(a)

F \ N = F 4 (N F ) ;
| {z }

(b)

N \ F = N 4 (F N ) ;
| {z }
N

0
(F 4 N ) N = F 4 N 4 N \ (F 4 N )

= F 4 N 4 (N 0 \ (F 4 N )) ,
|
{z
}

(c)

(d)

where we used Problem 2.6 and part (a) for (d).

S
Now let (Fj )jN F and (Nj )jN N and set F :=S j Fj F
and, because of -subadditivity of measures N := j Nj N.
Then
[
[
[
F \N =
(Fj \ N )
(Fj \ Nj )
Fj = F
jN

jN

jN

as well as

(Nj \ Fj )

jN

Nj = N

jN

which shows that

F \N

(Fj 4 Nj ) F N.

(**)

jN

S
Since F, N A, and consequently jN (Fj 4 Nj ) A, and since
A-measurable subsets of null sets are again in N, the inclusions
(**) show that there exists some N 0 N so that
[
jN

(Fj 4 Nj ) = (F \ N ) N 0
| {z }
, cf. (b)

where we used (d) for the last inclusion.

Schilling: Measures, Integrals & Martingales

Solution to 4.17: By definition,

A = A N : A A, N N .
Since
A N = A 4 (N \ A)
| {z }
N

and since by an application of Problem 4.16 to (X, A ,

), A, N (instead
of (X, A, ), G, N) we get

(A, N) = A 4 N : A A, N N
and we conclude that
A (A, N).
On the other hand,
A A

and N A

so that, since A is a -algebra,

(A, N) (A ) = A (A, N).
Finally, assume that A A and A A. Then A = A 4 N and we
get
A 4 A = A 4 N 4 A = (A 4 A) 4 N = N.
Note that this result would also follow directly from 4.13 since we know
from there that A = A N so that
A 4 A = (A N ) 4 A = A 4 (N \ A) 4 A = N \ A
Solution to 4.18: Denote the completion by B and write Nx for all subsets
of Borel null sets of x . Clearly,
Nx = {A Rn : x 6 A}.
Recall from Problem 4.13(i) that B contains all sets of the form B N
with B B and N Nx . Now let C Rn be any set. If x C, then
write
C = {x} (C \ {x}) B ;
|{z} | {z }
Nx

Otherwise, x 6 C and
C = C \ {x} = |{z}
(C \ {x}) B .
| {z }
B

Nx

This shows that B = P(Rn ) is the power set of Rn .

Chapter 4. Additional material. Last update: June 23, 2008

Solution to 4.19: (i) Fix > 0 and choose for A sets U O, F F

such that F A U and (U \ F ) < . Set U 0 := F c O and
F 0 := U c F. Then we have
F 0 Ac U 0 and U 0 \ F 0 = F c \ U c = F c U = U \ F
and so (U 0 \ F 0 ) = (U \ F ) < . This means that Ac .
Denote by d(x, y) the distance of two points x, y X and write
B1/n (0) for the open ball {y X : d(y, 0) < n1 }. As in the solution
of Problem 3.12(ii) we see that Un := F + B1/n (0) is a sequence of
open sets such that Un F . Because of the continuity of measures
n
we get (Un \ F ) 0 and since F 3 F F Un O, this
means that F .
(ii) Fix > 0 and pick for Aj , j = 1, 2, open sets Uj and closed
sets Fj such that Fj A Uj and (Uj \ Fj ) < . Then F1 F2
and U1 U2 are again closed resp. open, satisfy F1 F2 A1 A2
U1 U2 as well as

(U1 U2 ) \ (F1 F2 ) = (U1 U2 ) (F1c F2c )

= [(U1 U2 ) \ F1 ] [(U1 U2 ) \ F2 ]

6 (U1 U2 ) \ F1 + (U1 U2 ) \ F2
< 2.
This shows that is -stable.
(iii) Fix and pick for a given sequence (Aj )jN open sets Uj and
closed sets Fj such that
Fj Aj Uj and (Uj \ Fj ) < 2j .
S
S
Set AS:= j Aj . Then U := j Uj A is an open set wile
F := j Fj is contained in A but it is only an increasing limit of
closed sets n := F1 . . . Fn . Using Problem 4.9 we get
X
X
2j 6 .
(Uj \ Fj ) 6
(U \ F ) 6
j

Since n A U and U \ n U \ F , we can use the continuity

of measures to conclude that inf n (U \ n ) = (U \ F ) 6 , i.e.
(U \ N ) 6 2 if N = N is sufficiently large. This shows that
contains all countable unions of its members. Because of part (i)
it is also stable under complementation and contains the empty
set. Thus, is a -algebra.
As F and B = (F), we have B .

Schilling: Measures, Integrals & Martingales

(iv) For any Borel set B and any > 0 we can find open and
closed sets U and F , respectively, such that F B U and
(B \ F ) 6 (U \ F ) < = (B) 6 + (F ),
(U \ B) 6 (U \ F ) < = (B) > (U ) .
Thus,
sup
F B, F F

inf

U B, U O

(F ) 6 (B) 6 + (F ) 6 +
(U ) 6 (U ) 6 (b) 6

sup

(F )

F B, F F

inf

U B, U O

(U ).

(v) For every closed F F the intersections Kj F , j N, will be

compact and Kj F F . By the continuity of measures we get
(F ) = sup (Kj F ) 6
j

sup

(K) 6 (F ).

KF,K cpt

Thus,
(F ) =

sup

(K)

F F.

(*)

KF,K cpt

Combining (iv) and (*) we get

(iv)

(B) =

sup

(F )

F B, F F
(*)

sup

sup

(K)

F B, F F KF,K cpt

sup

sup

F B, F F

KB,K cpt

(K)
{z
}

note: independent of F B

sup

(K)

KB,K cpt

and since (K) 6 (B) for K B and supKB,K cpt (K) 6 (B)
are obvious, we are finished.
(vi) Assume now that is -finite. Let (Bn )nN B be an exhausting sequence for X such that (Bn ) < . Then the measures
n (B) := (B Bn ) defined on B are finite and regular according
to part (iv). Since we may interchange any two suprema (cf. the
solution of Problem 4.6) we get
(B) = sup n (B) = sup
n

sup
F B, F F

n (F )

Chapter 4. Additional material. Last update: June 23, 2008

=

sup
F B, F F

sup

sup n (F )
n

(F ).

F B, F F

Solution to 4.20: First of all, Problem 4.19 (iv) shows that

(B) =

sup

(F ).

(*)

F B,F closed

Let (dk )k be an enumeration of the dense set D X and write for

the metric in X and Kr (x) := {y X : (x, y) 6 r} for the closed ball
with centre x and radius r.
Since, for any fixed n N the sets
K1/n (d1 ) K1/n (dm ) X for m
we get from (*)
> 0 k(n) N : (Fn ) +

> (X)
2n

if Fn := K1/n (d1 ) K1/n (dk(n) ). Setting

\
K := K :=
Fn
n

it is clear that K is closed. Moreover, since K is, for every 1/n, covered
by finitely many balls of radius 1/n, to wit,
K K1/n (d1 ) K1/n (dk(n) ),
we see that K is compact. Indeed, if (xj )j K is a sequence, there is
a subsequence (xnj )j which is completely contained in one of the balls
K1/n (d1 ), . . . , K1/n (dk(n) ). Passing iteratively to sub-sub-etc. sequences
we find a subsequence (yj )j (xj )j which is contained in a sequence
of closed balls K1/n (cn ) (cn is a suitable element from D). Thus (yj )j
is a Cauchy sequence and converges, because of completeness, to an
element x which is, as the Fn are closed, in every Fn , hence in K.
Thus K is (sequentially) compact.
Since
(X \ K) =

[
n

X \ Fn

X
n

(X \ Fn ) 6

X
= ,
n
2
n

10

Schilling: Measures, Integrals & Martingales

we have found a sequence of compact sets Kn such that (Kn ) (X)
(note that the Kn need not converge X as a set!). Obviously, Kn F
is compact for every closed F and we have (Kn F ) (F ), hence
(F ) =

sup

(K) F F.

KF,K cpt

Now we can use the argument from the proof of Problem 4.20(v).
Solution to 4.21: Define a measure which assigns every point n
1
:
n Z, k N the mass 2k
=

1
,
2k

XX 1
n 1 .
2k
2k
nZ kN

(Since Z N is countable, Problem 4.6 shows that this object is indeed

a measure!) Obviously, any interval [a, b) of length b a > 2 contains
some integer, say m [a, b) so that [m 1/2, m) [a, b), thus
[a, b) > [m 1/2, m) =

X 1
= .
2k
kN

On the other hand, the sequence of sets

Bn :=

n
[

k 1, k

k=n

satisfies (Bn ) < and

S
n

Bn = R.

1
2n

Measurable mappings.
Additional Material

Problems
7.12. Let E, F P(X) be two families of subsets of X. One usually uses the notation (as
we do in this book)

E F = A : A E or A F
and E F = A : A E and A F .
Let us, for this problem, also introduce the families

E d F = E F : E E, F F
and E e F = E F : E E, F F .
Assume now that E and F are -Algebras.
(i) Show that E d F E F and E e F E F;
(ii) Show that, in general, we have no equality in (i);
(iii) Show that (E d F) = (E e F) = (E F).

Solution to 7.12:

(i) Since E and F we get

E E : E E d F = E E d F

and
F F : F E d F = F E d F
so that E F E d F. A similar argument, using that X E and
X F, shows E F E e F.
(ii) Let A, B X such that A B 6= , A B 6= X and that
A 6 B, B 6 A. Then we find for E := {, A, Ac , X} and
F := {, B, B c , X} that
E F = {, A, B, Ac , B c , X}
while
E d F = {, A, B, Ac , B c , A B, Ac B c , A B c , Ac B, X}.
A similar example works for E e F.
11

12

Schilling: Measures, Integrals & Martingales

(iii) Part (i) shows immediately that
(E d F) (E F) and (E e F) (E F).
Conversely, it is obvious that
E d F (E F) and E e F (E F)
so that
(E d F) (E F) and (E e F) (E F)
which proves
(E d F) = (E F) = (E e F).

Integration of positive functions.

Additional Material

Problems
9.13. Let (X, A, ) be a measure space and A1 , . . . , AN A such that (Aj ) < . Then
[
X
N
N
X
Aj >

(Aj )
j=1

j=1

(Aj Ak ).

16j<k6N

[Hint: show first an inequality for indicator functions.]

9.14. Let (X, A, ) be a measure space. Show the following variant of Theorem 9.6: If
uj > 0 are measurable functions such that for some u we have
K N : x : uj+K (x) u(x) as j ,
R
R
then u > 0 is measurable and uj d u d.
Show that we cannot replace the above condition with
x K N uj (x) u(x) as j .

Solution to 9.13: We use indicator functions. Note that any fixed x can
be contained in k {0, 1, . . . , N} of the sets Aj . Then x is contained
in A1 AN as well as in k2 of the pairs Aj Ak where j < k; as
usual: m
= 0 if m < n. This gives
n

X
X
k
1Aj 1Ak
1Aj = k 6 1 +
= 1A1 AN +
2
j
j<k
X
= 1A1 AN +
1Aj Ak .
j<k

Integrating this inequality w.r.t. yields the result.

Solution to 9.14: The first part is trivial since it just says that the sequence
becomes only from index K onwards. This K does not depend on x but
is uniform for the whole sequence. Since we are anyway only interested
in u = limj uj = supj>K uj , we can neglect the elements u1 , . . . , uK
and consider only the then increasing sequence (uj+K )j . Then we can
directly apply Beppo Levis theorem, Theorem 9.6.
13

14

Schilling: Measures, Integrals & Martingales

The other condition says that the sequence uj+K (x) is increasing for
some K = K(x). But since K may depend on x, we will never get
some overall increasing behaviour of the sequence of functions. Take,
for example, on (R, B(R), := 1 ),
uj (x) = j 2 (x + 1j )1(1/j,0) (x) j 2 (x 1j )1(0,1/j) (x).
This is a sequence of symmetric tent-like functions of tents with base
(1/j, 1/j) and tip at j 2 (which we take out and replace by the value
0). Clearly:
Z
j
uj (x) 0 and
uj (x) dx = 1 j.
Moreover, if j > K = K(x) with K(x) defined to be the smallest
integer > 1/|x|,
is clearly
R then uj (x) = 0 so that the second
R condition
R
satisfied, but uj (x) dx = 1 cannot converge to 0 dx = u(x) dx = 0.

10

Integrals of measurable functions and

null sets.
Additional Material

Problems
10.17. Prove Lemma 10.8.

Solution to 10.17: Clearly, is defined on A and takes values in [0, ].

Since 1 0 we have
Z
Z
() = 1 u d = 0 d = 0.
If (Aj )jN A are pairwise disjoint measurable sets, we get

[
Z

u d

Aj = 1S
j=1 Aj
j=1

Z X

1Aj u d

j=1

Z
X
j=1

1Aj u d =

X
j=1

which proves -additivity.

15

(Aj )

13

Product measures and Fubinis

theorem
Additional Material

Problems
13.15. Let (, A, P ) be a probability space, i.e. a measure space such that P () = 1.
Show that for a measurable function T : [0, ) and every > 0 the following
formula holds:
Z
Z
eT dP = 1

es P (T > s) ds.

What happens if we also allow negative values of ?

Solution to 13.15: Assume first that > 0. The point here is that Corollary 13.13 does not apply to the function s 7 es since this function is decreasing and has the value 1 for s = 0. Consider therefore
(s) := 1 es . This is admissible in 13.13 and we get
Z
Z
Z

T
es P (T > s) ds.
(T ) dP =
1e
dP =
0

Rearranging this equality then yields

Z
Z
T
e
dP = 1
es P (T > s) ds.
0

If < 0 the formula remains valid if we understand it in the sense that

either both sides are finite or both sides are infinite. The above argument needs some small changes, though. First, es is now increasing
(which is fine) but still takes the value 1 if s = 0. So we should change
to (s) := es 1. Now the same calculation as above goes through.
If one side is finite, so is the other; and if one side is infinite, then the
other is infinite, too. The last statement follows from Theorem 13.11
or Corollary 13.13.

16

14

Integrals with respect to image measures

Additional Material

Problems
14.12. A general Young inequality. Generalize Youngs inequality given in Problem
14.9 and show that
kf1 ? f2 ? ? fN kr 6

N
Y

kfj kp ,

p=

j=1

Nr
,
(N 1)r + 1

for all N N, r [1, ) and fj Lp (n ).

Solution to 14.12: For N = 1 the inequality is trivial, for N = 2 it is in

line with Problem 14.9 with p = q.
Let us, first of all, give a heuristic derivation of this result which explains how one arrives at the particular form for the value of p =
p(r, N ). We may assume that N > 2. Set Fj := fj ? . . . ? fN for
j = 1, 2, . . . N 1. Then
kf1 ? ? fN kr
6 kf1 kp kF2 kq2 = kf1 kp kf2 ? F3 kq2
by Pr. 14.9 where

1
r

+1=

1
p

1
q2

1
p

1 +

1
q2

+1

6 kf1 kp kf2 kp kF3 kq3 = kf1 kp kf2 kp kf3 ? F4 kq3

by Pr. 14.9 where 1r + 1 = p1 1 + p1 + q13 = 2 p1 1 + 1 +

| {z }
=

1
+1
q2

and repeating this procedure N 2 times we arrive at

kf1 ? ? fN kr 6 kf1 kp kfN 2 kp kfN 1 ? fN kqN 1
6 kf1 kp kfN 2 kp kfN 1 kp kfN kqN
with the condition
1

1
1
1
1
1
+ 1 = (N 2)
1 +1+
= (N 2)
1 + +
r
p
qN 1
p
p qN
17

1
q3

18

Schilling: Measures, Integrals & Martingales

and since we need qN = p we get
1
2
1
N
+ 1 = (N 2)
1 + =
N +2
r
p
p
p
and rearranging this identity yields
p=

Nr
.
(N 1)r + 1

If you do not like this derivation of if you got lost counting the repetitions, heres the formal proof using inductionbut with the drawback
that one needs a good educated guess what p = p(N, r) should look
like. The start of the induction N = 2 is done in Problem 14.9 (starting
at N = 1 wont help much as we need Youngs inequality for N = 2
anyway...).
The induction hypothesis is, of course,
kf1 ? ? fM kt 6

M
Y

kfj k

for all M = 1, 2, . . . , N 1

j=1

where t > 0 is arbitrary and =

Mt
.
(M 1)t+1

The induction step uses Youngs inequality:

kf1 ? f2 ? ? fN kr 6 kf1 kp kf2 ? ? fN kq
where p =

Nr
(N 1)r+1

and q is given by

1
1 1
(N 1)r + 1 1
1
1
1
+1= + =
+ =1+
+
r
q q
Nr
q
q N
Nr
so that

Nr
.
N +r1
Using the induction hypothesis we now get
q=

kf1 ? ? fN kr 6 kf1 kp kf2 ? ? fN kq 6 kf1 kp kf2 ks kfN ks

where s is, because of the induction assumption, given by
s=

(N 1)q
(N 2)q + 1

Chapter 14. Additional material. Last update: June 23, 2008

=
=
=
=
=
and we are done.

Nr
(N 1) N +r1
Nr
+1
(N 2) N +r1
(N 1)N r
(N 2)N r + N + r 1
(N 1)N r
2
N r 2N r + r + (N 1)
(N 1)N r
(N 1)2 r + (N 1)
Nr
=p
(N 1)r + 1

19

16

Uniform integrability and Vitalis

convergence theorem.
Additional Material

This is an extended version of Lemma 16.4, page 164 of the printed text
16.4 Lemma. Let (uj )jN Lp (), p [1, ), and (wk )kN M(A).
Then
(i)
(ii)

lim kuj ukp = 0 implies uj u;

lim wk (x) = w(x) a.e. implies wk w.

Let f : R R be a continuous function. If ({|u| = }) = 0, then

(iii)

uj u implies f uj f u.

Proof. (i) follows immediately from the Markov inequality P10.12,

({|uj u| > } A) 6 ({|uj u| > }) = ({|uj u|p > p })
1
6 p kuj ukpp .

(ii) Observe that for all > 0

{|wk w| > } { |wk w| > }.
An application of the Markov inequality P10.12 yields
({|wk w| > } A) 6 ({ |wk w| > } A)
Z
Z

1
1
6
|wk w| 1A d.
|wk w| d =
A

If (A) < , the function 1A L1+ () is integrable, dominates the integrand |wk w| 1RA , and Lebesgues dominated convergence theorem 11.2
implies that limk A ( |wk w|) d = 0.
(iii) Let R > 0 and , (0, 1). Clearly
{|f u f uj | > }

{|f u f uj | > } {|u uj | 6 } {|u| 6 R}

20

Appendix 16. Additional material. Last update: June 23, 2008

21

{|u uj | > } {|u| > R}.

Since {|u uj | 6 } {|u| 6 R} {|uj | 6 R + 1} {|u| 6 R}, and since f
is uniformly continuous on the closed interval [R 1, R + 1], we can choose
= so small that the first set on the right hand side is empty, i.e.
{|f u f uj | > } {|u uj | > } {|u| > R}.
This shows that for all A A with (A) <

{|f u f uj | > } A 6 {|u uj | > } A + {|u| > R} A .

Using uj u, we find

R
lim sup {|f u f uj | > } A 6 {|u| > R} A 0
j

since ({|u| = }) = 0.
This is an extended version of the Proof of (iii)(i) of Theorem 16.8, pages
165-6 of the printed text
Proof (of Theorem 16.8). (iii)(ii): Fix (0, 1). We have
Z
Z
Z
p

p
p
|uj | d 6
|u| d +
|uj | |u|p d
{|uj |>|u|}

{|uj |>|u|}

|u| d +

{|uj |>|u|}

|uj | |u|

{|uj |>|u|}

d +

|uj |p |u|p d.

{|uj |6|u|}

Denote the three integrals I1 , I2 and I3 , respectively. Because of assumption

(iii) we know that
I2 6 p , for all j > N
R
for some N N. Moreover, (iii) shows that supjN |uj | d 6 C < which
means that
Z
Z
Z
p
p
p
p
I1 6
|u| d 6
|uj | d 6
|uj |p d 6 C p .
{|uj |>|u|}

{|uj |>|u|}

Finally, the inclusion

|uj | 6 |u| ||uj |p |u|p | 6 (p + 1)|u|p

22

Schilling: Measures, Integrals & Martingales

yields with := p + 1 and pj := |uj |p |u|p

Z
pj d.
|I3 | 6
{pj 6|u|p }

Splitting the integration domain into three parts,

p

j 6 (|u|p ) ,
6 pj 6 |u|p |u|p > R ,

and
6 pj 6 |u|p |u|p 6 R ,
we get
Z
Z
p
|I3 | 6 (|u| ) d +

|u|p d + R

{|u|p >R}

pj > |u|p > .

By dominated convergence, the two integrals on the right hand side tend to
0 as 0 resp. R . Thus, we may choose = and R = R such that

|I3 | 6 p + R pj > |u|p > .

Since the set {|u|p > /} has finite -measure[X] and since |uj |p |u|p ,

i.e. pj 0, see Lemma 16.4(iii), we can find some M > N such that
|I3 | 6 2p

for all j > M .

Setting w := max{|u1 |, . . . , |uM |, |u|} we have w Lp+ ()[X] and see

Z
Z
sup
|uj | d 6 (C + 3) p .
|uj | d 6 sup
jN

|uj |> 1 w

j>M

{|uj |6|u|}

Since
w Lp+ () wp L1+ () and

|uj | > 1 w = |uj |p >

1
p

wp ,
(16.3)

we have established the uniform integrability of (|uj |p )jN .

This is a streamlined version of the proof of Theorem 16.8(vi)(vii), pages
171-3 of the printed text
(vi)(vii): For u F we set n := n (u) := ({|u| > n}) and define
Z

X
(t) :=
(s) (ds),
(s) :=
n 1[n,n+1) (s).
[0,t)

n=1

Appendix 16. Additional material. Last update: June 23, 2008

23

We will now determine the numbers 1 , 2 , 3 , . . .. Clearly,

Z

X
X
(t) =
n
1[n,n+1) (s) (ds) =
n [(t n)+ 1]
[0,t)

n=1

and
Z
(|u|) d =

Z
n

n=1

(|u| n)+ 1 d 6
n (|u| > n).

n=1

(16.6)

n=1

If we can construct (n )nN such that it increases to and (16.6) is finite

(uniformly for all u F), then we are done: (s) will increase to , (t)
will be convex1 and satisfy
Z
Z

1
(t)
1
(s) (ds) >
(s) (ds) > 12 2t .
=
t
t [0,t)
t [t/2,t)
By assumption we
R can find an increasing sequence (rj )jN R such that
limj rj = and {rj <|u|} |u| d 6 2j . Now2

1{k<|u|} 6

[|u|]
X

1{k<|u|} 6

k=rj

k=rj

[|u|]
X

1{rj <|u|} 6 |u| 1{rj <|u|} .

k=rj

If we integrate both sides of this inequality and use Beppo Levis theorem in
the form of Corollary 9.9 on the left-hand side, we get
Z

Z
X
X
({|u| > k}) =
1{k<|u|} d 6 |u| 1{rj <|u|} d 6 2j .
k=rj

k=rj

Usually one argues that 00 > 0 a.e., but for this we need to know that the monotone function = 0 is almost everywhere differentiableand this requires Lebesgues
differentiation theorem 19.20. Here is an alternative elementary argument: it is not hard
6 (z)(x)
holds for
to see that : (a, b) R is convex if, and only, if (y)(x)
yx
zx
all a < Rx < y < z < b, use e.g. the technique of the proof of Lemma 12.13. Since
x
(x) = 0 (s) ds (by L13.12 and T11.8), this is the same as
Z y
Z z
Z y
Z z
1
1
1
1
(s) ds 6
(s) ds
(s) ds 6
(s) ds
yx x
zx x
yx x
zy y
Z 1
Z 1

(s(y x) + x) ds 6
(s(z y) + y) ds.
0

The latter inequality follows from the fact that is increasing and s(y x) + x [x, y]
while s(z y) + y [y, z] for 0 6 s 6 1.
2
for x R we will denote by [x] the Gau bracket, i.e. the largest integer 6 x.

24

Schilling: Measures, Integrals & Martingales

Summing over j N finally yields

X

({|u| > k}) 6

j=1 k=rj

2j = 1.

j=1

and a simple interchange in the order of the summation gives

!
X

X
X
X
({|u| > k}) =
1[1,k] (rj ) ({|u| > k}) 6 1.
j=1 k=rj

k=1

j=1

{z
=: k

This finishes the construction of the sequence (k )kN .

Problems
16.15. Let (fi )iI be a family of uniformly integrable functions and let (ui )iI L1 () be
some further family such that |ui | 6 |fi | for every i I. Then (ui )iI is uniformly
integrable. In particular, every family of functions (ui )iI with |ui | 6 g for some
g L1+ is uniformly integrable.
16.16. Let (X, A, ) be an arbitrary measure space. Show that a family F L1 () is
uniformly integrable if, and only if the following condition holds:
Z
1
> 0 g L+ () : sup (|u| g |u|) d < .
uF

Give a simplified version of this equivalence for finite measure spaces.

Solution to 16.15: Fix > 0. By assumption there is some w = w L1+

such that
Z
|fi | d 6 .
sup
i

{|fi |>w}

Consider now
Z
Z
|ui | d 6
|fi | d
{|ui |>2w}
{|ui |>2w}
Z
Z
=
|fi | d +
|fi | d
{|ui |>2w}{|fi |6w}
{|ui |>2w}{|fi |>w}
Z
Z
1
6
|ui | d +
|fi | d
{|ui |>2w}{|fi |6w} 2
{|ui |>2w}{|fi |>w}
Z
Z
1
6
|ui | d +
|fi | d
{|ui |>2w} 2
{|fi |>w}

Appendix 16. Additional material. Last update: June 23, 2008

25

Thus,
1
2

Z
|ui | d 6

{|ui |>2w}

|fi | d
{|fi |>w}

6 sup
i

|fi | d 6
{|fi |>w}

uniformly for all i I.

Solution to 16.16: Let g L1+ (). Then
Z
Z
Z
0 6 (|u| g |u|) d =
(|u| g) d 6
{|u|>g}

|u| d.
{|u|>g}

This implies that uniform integrability of the family F implies that the
condition of Problem 16.16 holds. On the other hand,
Z
Z
|u| d =
(2|u| |u|) d
{|u|>g}
{|u|>g}
Z
6
(2|u| g) d
{|u|>g}
Z
6
(2|u| g) d
{2|u|>g}
Z

=2
|u| 12 g d
{|u|> 12 g}
Z

|u| 12 g |u| d
=2
{|u|> 12 g}

and since g L1 if, and only if, 12 g L1 , we see that the condition
given in Problem 16.16 entails uniform integrability.
In finite measure spaces this conditions is simpler: constants are integrable functions in finite measure spaces; thus we can replace the
condition given in Problem 16.16 by
Z
lim sup (|u| R |u|) d = 0.
R uF

18

Martingale convergence theorems.

Additional Material

This is a streamlined version of (i)(ii) of the proof of 18.6, pages 194-5

of the printed text
Proof. (i)(ii): Since |A0 is -finite, we can fix an exhausting sequence
(Ak )kN A0 with
Ak X and (Ak ) < . It is not hard to see that
P
the function w := k=1 2k (1 + (Ak ))1 1Ak is strictly positive w > 0 and
integrable w L1 (A0 , ). Example 17.3(iv) shows that (u+
j )jN{} is still a
submartingale, so that for every L > 0
Z
Z
+
uj d 6
u+
d
{u+
j >Lw}

{u+
j >Lw}

Z
u+

+
d +
6

u d


+
+
+
+
1
1
uj >Lw u 6 2 Lw
u >Lw u > 2 Lw
Z j
Z
1
+
+
6
u d +

uj d.

1
2 u+j >Lw u+ 6 12 Lw
u+
> 2 Lw
1
+
For the last estimate we used that on the set {u+
j > Lw} {u 6 2 Lw}
1 +
1
the integrand
of the second integral satisfies u+
6 2 Lw < 2 uj . Now we
R
subtract 21 {u+ >Lw} u+
j d on both sides and get, uniformly for all j N,
j
Z
Z
1
L
+
+
u d 6
u d 0.
+
1
2 {u+j >Lw} j
u > 2 Lw

This follows from the dominated convergence theorem, Theorem 11.2, since
L
|u | L1 (A , ) dominates u+
0. Thus, (u+
1
1{u+
j )jN is uni > 2 Lw}
+
+
formly integrable. From limj uj = u a.e., we conclude
j uRj = u ,
R lim
+
+
and Vitalis convergence theorem 16.6 shows that limj uj d = u d.
Thus
Z
Z
Z
Z
j
+
+
|uj | d = (2uj uj ) d (2u u ) d = |u | d,
and another application of Vitalis theorem proves that (uj )jN is uniformly
integrable.
This is a streamlined version of the second part of Example 18.8, page 199
of the printed text
26

Appendix 18. Additional material. Last update: June 23, 2008

27

18.8 Example (continued). We can now continue with the proof of the
necessity part of Kolmogorovs strong law of large numbers. Since the a.e.
limit exists, we get
Xn
Sn n 1 Sn1 n
=

0,
n
n
n n1
which shows that
An := {|Xn | > n} happens only for finitely many n.
P

In other words, P ( j=1 1Aj = ) = 0; since the An are all independent, the
P
1
Borel-Cantelli lemma T18.9 shows that
j=1 P (Aj ) < . Thus

|X1 | 6 |X1 | + 1 = 1 +

|X1 |
X

1=1+

n=1

1{n6|X1 |} .

n=1

If we integrate this inequality w.r.t. P and use Beppo Levis theorem, we

arrive at
Z

X
X
|X1 | dP 6 1 +
P ({|X1 | > n}) = 1 +
P ({|Xn | > n})
n=1

n=1

since X1 and Xn have the same distribution.

We will see more applications of the martingale convergence theorems in
the following chapters.

for x R we will denote by [x] the Gau bracket, i.e. the largest integer 6 x.

19

The Radon-Nikod
ym Theorem and
Other Applications of Martingales.
Additional Material

Problems
19.19. Let and be measures on the measurable space (X, A). Show that the absolute
continuity condition (19.1) is equivalent to
(A 4 B) = 0

(A) = (B)

for all A, B A.

Solution to 19.19: : Assume first that (19.1) holds, i.e. that . If

(A 4 B) = 0 for some A, B A we get (A 4 B) = 0. By definition,
(A 4 B) = (A \ B) + (B \ A) = (A \ (A B)) + (B \ (A B)) = 0
so that
(A \ (A B)) = (B \ (A B)) = 0.
Assume that (A) < . Then (A B) 6 (A) < and we see that
(A) = (A B) and (B) = (A B)
which means that (A) = (B).
If (A) = the condition (A\(AB)) = 0 shows that (AB) = ,
otherwise 0 = (A\(AB)) = (A)(AB) = which is impossible.
Again we have (A) = = (B).
: Assume now that the condition stated in the problem is satisfied.
If N A is any -null set, we choose A := N and B := and observe
that A 4 B = N . Thus,
(N ) = (A 4 B) = 0

(A) = (B)

but this is just (N ) = (A) = () = 0. Condition (19.1) follows.

28

22

Conditional expectations in L2
Additional Material

Problems
22.4. Let (X, A, ) be a measure space and G A be a sub--algebra. such that |G
is -finite.
T Let p, q (1, ) be conjugate numbers, i.e. 1/p + 1/q = 1, and let
u, w p[1,] Lp (A, ). Show that
G

E (uw) 6 EG (|u|p ) 1/p EG (|w|q ) 1/q .

1/p

1/q
[Hint: use Youngs inequality 12.1 with A = |u|/ EG (|u|p )
and B = |w|/ EG (|w|q )
whenever the numerator is not 0.]

Solution to 22.4: Let Gu := {EG |u|p > 0}, Gw := {EG |w|q > 0} and G :=
Gu Gw . Following the hint we get

|u|

|w|

1/p G
1/q
EG (|u|p )
E (|w|q )

1G 6

|u|p
|u|q
1
+
1G
G
p EG (|u|p )
q EG (|w|q )

Since 1G is bounded and G-measurable, we can apply EG on both sides

of the above inequality and get
EG (|u||w|)
EG (|u|p )
EG (|u|q )
1G +
1G = 1G
G
1/p G
1/q 1G 6
p EG (|u|p )
q EG (|w|q )
E (|u|p )
E (|w|q )
or

1/p G
1/q
EG (|u||w|) 1G 6 EG (|u|p )
E (|w|q )
1G
G

1/p
1/q
6 E (|u|p )
EG (|w|q )
.
Denote by Gn an exhaustion of X such that Gn G, Gn X and
(Gn ) < . Then
Z
Z
p
|u| d = sup
|u|p d
Gcu

Gcu Gn

= suph1Gcu Gn , |u|p i
n

= suphEG 1Gcu Gn , |u|p i

n

29

30

Schilling: Measures, Integrals & Martingales

= suph1Gcu Gn , EG (|u|p )i
n

=0
which means that 1Gu u = u almost everywhere. Thus,
EG (|u||w|) 1G = EG (|u||w|1G ) = EG (|u|1Gu |w|1Gw ) = EG (|u||w|)
and the inequality follows since
G

E (uw) 6 EG (|uw|).

23

Conditional expectations in Lp
Additional Material

Problems
23.13. Let (X, A, Aj , ) be a finite filtered measure space, let uj L1 (Aj ) be a sequence
of measurable functions such that for some f L1 (A)
|uj | 6 E Aj f
holds true. Show that the family (uj )jN is uniformly integrable.

Solution to 23.13: (Compare this problem with Problem 16.15.) Recall

that in finite measure spaces uniform integrability follows from (and is
actually equivalent to)
Z
|uj | d = 0;
lim sup
R

{|uj |>R}

this is true since in a finite measure space the constant function w R

is integrable.
Observe now that
Z
Z
|uj | d 6
E Aj f d
{|uj |>R}
{|u |>R}
Z j
=
f d
{|uj |>R}
Z
Z
=
f d +
f d
{|uj |>R}{f 6R/2}
{|uj |>R}{f >R/2}
Z
Z
1
|uj | d +
f d
6
{|uj |>R}{f >R/2}
{|uj |>R}{f 6R/2} 2
Z
Z
1
6
|uj | d +
f d
{|uj |>R} 2
{f >R/2}
This shows that
Z
Z
1
R
|uj | d 6
f d 0.
uniformly for all j
2 {|uj |>R}
{f >R/2}

31

24

Orthonormal systems and their convergence behaviour

Additional Material

This is a streamlined and corrected version of the Epilogue 24.29, pages

309-12 of the printed text
24.29 Epilogue. The combination of martingale methods and orthogonal
expansions opens up a whole new world. Let us illustrate this by a rapid
construction of one of the most prominent stochastic processes: the Wiener
process or Brownian motion.
Choose in Theorem 24.27 (X, A, P ) = ([0, 1], B[0, 1], ) where is onedimensional Lebesgue measure on [0, 1]; denoting points in [0, 1] by , we will
often write d instead of (d). Assume that the independent, identically
distributed random variables ej are all standard normal Gaussian random
variables, i.e.
Z
1
2
ex /2 dx,
B B(R),
P (ej B) =
2 B
and consider the series expansion
Wt () :=

en ()h1[0,t] , Hn i,

[0, 1].

n=0

R1
Here t [0, 1] is a parameter, hu, vi = 0 u(x)v(x) dx, and Hn , n = 2k + j,
0 6 j < 2k , denote the lexicographically ordered Haar functions (24.16). A
short calculation confirms for n > 1
Z t
Z t
k/2
h1[0,t] , Hn i =
Hn (x) dx = 2
H1 (2k x j) dx = Fn (t),
0

Rt
where F1 (t) = 0 H1 (x) dx 1[0,1] (t) = t1[0, 1 ) (t) (t 1)1[ 1 ,1] (t) and Fn (t) =
2
2
2k/2 F1 (2k t j) are tent-functions which have the same supports as the Haar
functions. Since the Haar functions are a complete orthonormal system, cf.
Theorem 24.17, we may apply Bessels inequality, 21.11(iii), to get

X
h1[0,t] , Hn i2 6 h1[0,t] , 1[0,t] i = t 6 1.
n=0

32

Chapter 24. Additional material. Last update: June 23, 2008

33

Thus, Theorem 24.27(ii) guarantees that Wt () exists, for each t [0, 1],
both in L2 (d)-sense and (d)-almost everywhere.
More is true. Since the en are independent Gaussian random variables,
so are their finite linear combinations (e.g. Bauer [?, 24]) and, in particular,
the partial sums
N
X
SN (t; ) :=
en ()h1[0,t] , Hn i.
n=0

Gaussianity is preserved under L2 -limits;6 we conclude that Wt () has a

Gaussian distribution for each t. The mean is given by
Z

Wt () d =
0

Z
X
n=0

en () d h1[0,t] , Hn i = 0
0

(to change integration and summation use that L2 (d)-convergence

entails
R
L1 (d)-convergence on a finite measure space). Since en em d = 0 or 1
according to n 6= m or n = m, we can calculate for 0 6 s < t 6 1 the
variance by
Z 1
(Wt () Ws ())2 d
0

=
=

Z
X
n,m=0

en ()em () d h1[0,t] 1[0,s] , Hn ih1[0,t] 1[0,s] , Hm i

0

h1(s,t] , Hn i2

24.17,21.13

h1(s,t] , 1(s,t] i = t s.

n=0

In particular, the increment Wt Ws has the same probability distribution

as Wts . In the same vein we find for 0 6 s < t 6 u < v 6 1 that
Z 1
(Wt () Ws ())(Wv () Wu ()) d = h1(s,t] , 1(u,v] i = 0.
0

Since Wt Ws is Gaussian, this proves already the independence of the two

increments Wt Ws and Wv Wu , cf. [?, 24]. By induction, we conclude
that
Wtn Wtn1 , . . . , Wt1 Wt0 ,
6

(cf. [?, 23, 24]) if Xn is normal distributed with mean 0 and variance n2 , its Fourier
R
2 2
n
transform is eiXn dP = en /2 . If Xn X in L2 -sense, we have n2 2 and,
R iX
R
2 2
2 2
by dominated convergence, e
dP = limn eiXn dP = limn en /2 = e /2 ; the
claim follows from the uniqueness of the Fourier transform.

34

Schilling: Measures, Integrals & Martingales

are independent for all 0 6 t0 6 6 tn 6 1.

Let us finally turn to the dependence of Wt () on t. Note that for m < n
Z 1

4
sup S2n 1 (t; ) S2m 1 (t; ) d
0 t[0,1]

sup

k 1
X
n1 2X

0 t[0,1]

sup

sup
0 t[0,1]

e2k +j ()h1[0,t] , H2k +j i

k=m j=0

X
n1

0 t[0,1]

k8

k 1
2X

2 e2k +j ()h1[0,t] , H2k +j i

j=0

k=m

X
n1

k
8

k 1
X
4
n1 2X
k
k6

2 8 e2k +j ()h1[0,t] , H2k +j i d

2

| k=m{z

j=0

k=m

6 10

where we used Holders inequality for the outer sum with p = 34 and q = 4.
Since the functions F2k +j (t) = h1[0,t] , H2k +j i with 0 6 j < 2k have disjoint
supports and are bounded by 2k/2 , we find
Z 1

4
sup S2n 1 (t; ) S2m 1 (t; ) d
0 t[0,1]

6 10

sup

1
n1 2X
X

0 t[0,1] k=m j=0

6 10

1
n1 2X
X
k=m j=0

k
2

|0

e42k +j () d 22k
{z
}

= (2)1/2

=C

n1
X

2 2 e42k +j ()h1[0,t] , H2k +j i4 d

y 4 ey

2 /2

dy j,k
m

2 2 2k 22k 6 2C 2 2

k=m

which means that the partial sums S2n 1 (t; ) of Wt () converge in L4 (d)
uniformly for all t [0, 1]. By C12.8 we can extract a subsequence, which
converges (uniformly in t) for (d)-almost all to Wt (); since for fixed
the partial sums t 7 S2n 1 (t; ) are continuous functions of t, this property
is inherited by the a.e. limit Wt ().
The above construction is a variation of a theme by Levy [?, Chap. I.1,
pp. 1520] and Ciesielski [?]. In one or another form it can be found in many
probability textbooks, e.g. Bass [?, pp. 1113] or Steele [?, pp. 3539]. A

Chapter 24. Additional material. Last update: June 23, 2008

35

related construction of Wiener, see Paley and Wiener [?, Chapter XI], using
random Fourier series, is discussed in Kahane [?, 16.13].

Appendix E
A summary of the Riemann integral.
Additional Material
This is a streamlined and extended version of Theorem E.11, pages 344-5
of the printed text
E.11 Theorem. The Riemann integral is a positive linear form on the vector lattice R[a, b], that is, for all , R and u, w R[a, b] one has
Z b
Z b
Z b
w dt;
u dt +
(u + w) dt =
(i)
u + w R[a, b] and
(ii)

u6w

(iii)

u dt 6
a
+

w dt;
a

u w, u w, u , u , |u| R[a, b]

and

Z b
Z b

6
u
dt
|u| dt;

(iv)
(v)

|u|p , u w R[a, b], 1 6 p < .

If : R2 R is a Lipschitz continuous function1 then (u, w)
R[a, b].

Proof. (i) follows immediately from the linearity of the limit criterion in
Theorem E.5(iv).
Rb
a

(ii): In view of (i) it is enough to show that v := w u > 0 entails

v dt > 0. This, however is clear since v R[a, b] and
Z b
Rb
v dt.
0 6 v =
a

(iii), (iv) follow at once from (v) since the functions (x, y) 7 x y,
(x, y) 7 x y, x 7 x 0, x 7 (x) 0 and x 7 |x| are clearly Lipschitz
continuous.
1
i.e. for every ball Br (0) R2 there is a (so-called Lipschitz ) constant L = L(r) <
such that

|(x, x0 ) (y, y 0 )| 6 L |x y| + |x0 y 0 |

holds for all (x, x0 ), (y, y 0 ) Br (0).

36

Appendix E. Additional material. Last update: June 23, 2008

37

The estimate in part (iii) can be derived from (i) and

R (ii) by
observing
Rb
Rb
b
Rb
that u 6 |u| entails a u dt 6 a |u| dt which implies a u dt 6 a |u| dt.
(v): Let u, w R[a, b]. Since any Riemann integrable function is, by
definition, bounded, there is some r < such that r 6 u(x), w(x) 6 r;
write L = L(r) for the corresponding Lipschitz constant of |Br (0) . Observe
that for any partition = {a = t0 < t1 < < tk = b} of [a, b] we have

S [ (u, w) S [ (u, w)]

k h
i
X
=
sup (u(s), w(s)) inf (u(t), w(t)) (tj tj1 )
j=1

t[tj1 ,tj ]

s[tj1 ,tj ]

k
X

sup

(u(s), w(s)) (u(t), w(t)) (tj tj1 )

j=1 s,t[tj1 ,tj ]

6L

k
X

sup

|u(s) u(t)| + |w(s) w(t)| (tj tj1 ).

j=1 s,t[tj1 ,tj ]

Because of the symmetric roles of s and t,

sup |u(s) u(t)| = sup max{u(s) u(t), u(t) u(s)} = sup(u(s) u(t)),
s,t

s,t

s,t

and this proves

S [ (u, w) S [ (u, w)]

6L

k
X

sup

u(s) u(t) (tj tj1 )

j=1 s,t[tj1 ,tj ]

+L

k
X

sup

w(s) w(t) (tj tj1 )

j=1 s,t[tj1 ,tj ]

= L S [u] S [u] + L S [w] S [w] .

Thus, if both u and w are Riemann integrable, so is (u, w).