Solutions

Chapter 2
2.1 If we can cover A by open intervals with prescribed total length, then
we can also cover A by closed intervals with the same endpoints (closed
intervals are bigger), and the total length is the same. The same is true for
any other kind of intervals. For the converse, suppose that A is null in the
sense of covering by closed intervals. Let c > 0, take a cover Cn = [an, bn ]
with Ln (bn - an) < ~, let In = (an - c 2}+2, bn + c 2}+2)' For each n, In
contains Cn, so the In cover A; the total length is less than c. In the same
way we refine the cover by any other kind of intervals.
2.2 Write each element of C in ternary form. Suppose C is countable and so
they can be arranged in a sequence. Define a number which is not in this
sequence but has ternary expansion and so is in C, by exchanging 0 and 2
at the nth position.
2.3 For continuity at x E [0,1] take any c > 0 and find F(x) - c < a < F(x) <
b < F(x) + c of the form a = 2':.' b = :;. By the construction of F, these
numbers are values of F taken on some intervals (aI, a2), (b!, b2), with
ternary ends, and al < x < b2 . Take a 8 such that al < x - 8 < x + 8 < b2
to get the continuity condition.
2.4 m*(B) ~ m*(AUB) ~ m*(A) +m*(B) = m*(B) by monotonicity (Propo-
sition 2.5) and subadditivity (Theorem 2.7). Thus m*(A U B) is squeezed
between m*(B) and m*(B), so has little choice.
2.5 Since A c BU(ALlB), m*(A) ~ m*(B)+m*(ALlB) = m*(B) (monotonic-
ity and subadditivity). Reversing the roles of A and B gives the opposite
inequality.

287
 288 Measure, Integral and Probability

2.6 Since Au B = Au (B \ A) = Au (B \ (A n B)), using additivity and

Proposition 2.15 we have m(AUB) = m(A) +m(B) -m(AnB). Similarly
m(A U B U C) = m(A) +m(B) + m(C) - m(A n B) - m(An C) - m(B n
C) + m(A n B n C). Note, that the second part makes use of the first on
the sets A n C and B n C.
2.7 It is sufficient to note that
00 1 00 1
(a,b) = U(a,b--], (a,b) = U[a+-,b).
n=l
n n=l
n

2.8 If E is Lebesgue-measurable, then existence of 0 and F is given by The-

orems 2.17 and 2.29, respectively. Conversely, let € = ~, find On, Fn , and
then m*(nOn \ E) = 0, m*(E \ UFn) = 0, so E is Borel up to null set.
Hence E is in M.
2.9 We can decompose A into U:I (A n Hi); the components are pairwise
disjoint, so
00 00

i=l i=l

using the definition of conditional probability.

2.10 Since AcnB = (D\A)nB = B\ (AnB), p(ACnB) = P(B) -p(AnB) =
P(B) - P(A)P(B) = P(B)(l - P(A)) = P(B)P(N).
2.11 There are 32 paths altogether. S(5) = 524.88 = 500U 2 D 3 , so there are
m = 10 paths. S(5) > 900 in two cases: S(5) = 1244.16 = 500U 5 or
S(5) = 933.12 = 500U 4 D, so we have 6 paths with probability 0.5 5 =
0.03125 each, so the probability in question is 0.1875.
2.12 There are 2m paths of length m, Fm can be identified with the power set
of the set of all such paths, thus it has 22'" elements.
2.13 Fm c Fm+l since if the first m + 1 elements of a sequence are identical, so
are the first m elements.
2.14 It suffices to note that p(AmnA k ) = i, which is the same as P(Am)P(A k ).

Chapter 3
3.1 If f is monotone (in the sense that Xl ::; X2 implies f (Xl) ::; f (X2)), the
inverse image of interval (a, (0) is either [b, (0) or (b, (0) with b = sup{ X :
f(x) ::; a}, so it is obviously measurable.
Solutions 289

3.2 The level set {x : f(x) = a} is the intersection of f- 1([a, +00)) and
f- 1 ((-00, a]), each measurable by Theorem 3.3.
3.3 If b ~ a, then (r)-l((-oo,b]) =~, and if b < a, then (r)-l((-oo,b]) =
f- 1 ((-00,b]); in each case a measurable set.
3.4 Let A be a non-measurable set and let f(x) = 1 if x E A, f(x) = -1
otherwise. The set f- 1 ([1,00)) is non-measurable (it is A), so f is non-
measurable, but P == 1 which is clearly measurable.
3.5 Let g(x) = lim sup fn(x), h(x) = liminf fn(x); they are measurable by
Theorem 3.9. Their difference is also measurable and the set where fn
converges is the level set of this difference: {x : f n converges} = {x :
(h - g)(x) = O}, hence is measurable.
3.6 If sup f = 00 then there is nothing to prove. If sup f = M, then f(x) ::; M
for all x, so M is one of the z in the definition of ess sup (so the infimum
of that set is at most M). Let f be continuous and suppose ess supf < M
finite. Then we take z between these numbers and by the definition of
ess sup, f(x) ::; z a.e. Hence A = {x : f(x) > z} is null. However, by
continuity A contains the set f- 1 ((z, M)) which is open - a contradiction.
If sup f is infinite, then for each z the condition f (x) ::; z a.e. does not
hold. The infimum of the empty set is +00 so, we are done.
3.7 It is sufficient to notice that if 9 is any O"-field containing the inverse images
of Borel sets, then :Fx c g.
3.8 Let f : ~ -+ ~ be given by f(x) = x 2. The complement of the set f(~)
equal to (-00,0) cannot be ofthe form f(A) since each of these is contained
in [0,00).
3.9 The payoff of a down-and-out European call is f((S(O), S(l), ... , S(n)))
with f(xo, Xl, .•. , XN) = (XN - K)+ . lA, where A = {(xo, Xl,· •. , XN) :
min{xO,x1, ... ,XN} ~ L}.

Chapter 4
4.1 (a) fo101nt(x)dx = Om([O, 1))+lm([1,2))+2m([2,3))+·· ·+9m([9, 10))+
lOm([10, 10]) = 45.
(b) fo2 Int(x 2 ) dx = Om([O, l])+lm([l, V2))+2m([V2, V3))+3m([V3, 2)) =
5- V3- V2.
(c)
-71".
f: 1r
Int(sinx) dx = Om([O, ~))+lm([~, ~])+Om((~,7I"])-lm((7I",271"]) =
290 Measure, Integral and Probability

4.2 We have i(x) = I:~1 k2 k - 1 lAk(X), where Ak is the union of 2 k - 1 inter-

vals of length 31k each, that are removed from [0, 1] at the kth stage. The
convergence is monotone, so

1 [O,lJ
idm =
n 2k-1
nl~~Lk3k = 3 Lk 3
k=l
1
k=l
00 (2)k-1
Since I:~1 a k = 12"" differentiating term by term with respect to a we
have I:~1 ka k - 1 = (1_1",)2' With a = ~ we get fro,lJ i dm = 3.
4.3 The simple function alA is less than i, so its integral, equal to am(A), is
smaller than the integral of f. Next, i :::; bl A , hence the second inequality.
4.4 Let in = nl(o,~J; limin(x) = 0 for all x but fin dm = 1.

4.5 Let a =I=- -1. We have f x'" dx = "'~1 x"'+1 (indefinite integral). First con-
sider E = (0,1):
1 o
1
x'" dx = - - x",+ll
a+1
1
0'
1

which is finite if a> -1. Next E = (1,00):

and for this to be finite we need a < -1.

4.6 The sequence in(x) = 1.tx3
converges to 0 pointwise, ~ l.tx3 : :; : :;
~X-2.5 :::; x- 2 .5 which is integrable on [1,00), so the sequence of integrals
converges to O.
4.7 First a = O. Substitute u = nx:

The sequence of integrands converges to ue- u2 for all u :::: 0; it is dominated

by g(u) = ue- u2 , so

lim 1o00 indm= 100 limindm= 100 ue-udu=-.21

0 0
2

Now a > O. After the same substitution we have

1 00 n2xe-n2x2
2 dx=
1 ue- u2
(u)2 l [na,00)(u)du=
1 in(u)du,
a l+x lR 1+ n lR

say, and in -+ 0, in(u) :::; ue- u2 , so limf in dm = O.

 Solutions 291

4.8 The sequence In(x) converges for x ~ 0 to e- x . We find the dominating

function. Let n > 1. For x E (0, 1), x~ ~ x!, (1 + ~)n ~ 1, so In(x) ::; .Ix
which is integrable over (0,1). For x E [1, 00), x-~ ::; 1, so In(x) ::; (1 +
~)-n. Next

(1 + -X) n = 1+ x + n(n - 1) (X)

- 2+ ... > x 2-
n- 1
- ~
1 2
-x ,
n 2! n 2n 4
so In(x) ::; ~ which is integrable over [1,00).
Therefore, by the dominated convergence theorem,

lim 1 00
In dm = 100
e- X dx = 1.

4.9 (a) f~1 Inaxnl dx = n a tl

Ixl n dx = 2n a fOI xn dx (Ixl n is an even function)
= ~~~. If a < 0, then the series L:n~1 ~~~ converges by comparison with
nl~a, we may apply the Beppo-Levi theorem and the power series in ques-
tion defines an integrable function. If a = 0 the series is L:n>1 xn = l=x'
which is not integrable since f~1 (L:n>1 xn) dx = L::=1 f~1 x;dx = 00. By
comparison the series fails to give an-integrable function if a > O.

e -l = x~
(b) Write _x_ l-e- = '""'
X
fOO xe- nx dx = x(_l)e-nXIOO -
L.Jn>1 xe- nx , JoX n 0
oo nx
(-~) fo e- dx = ~ (integration by parts) and, as is well known,
,"",00 1 71"2
L.Jn=1 ~ = 6'
4.10 We extend I by putting 1(0) = 1, so that I is continuous, hence Riemann-
integrable on any finite interval. Let an = f~:+I)71" I(x) dx. Since I is even,
a_n = an and hence f~oo I (x) dx = 2 L::=o an. The series converges since
an = (-l)nla n l, lanl ::; ;71" (x ~ mr, I f~:+1)71" sinxdxl = 2). However for
I to be in Ll we would need In~.1/1 dm = 2 L::=o bn finite, where bn =
I~:+1)71" I/(x)1 dx. This is impossible due to bn ~ (n';I)71"'
4.11 Denote f OO
-00
2
e- X dx = Ij then

12 = J L2 e-(x 2+y2)dx dy = 1271" 1 00

re- r2 dr do: = 11",

using polar coordinates and Fubini's theorem (Chapter 6).

Substitute x = 0.;
in Ij .,fii = f::O e- x2 dx = vt f~oo e- ~ dz, which
2<7

completes the computation.

4.12 fIR l';x2 dx = arc tanxl:J=: = 11", hence f~oo c(x) dx = 1.
4.13 1000
e->'X dx = -te->,xIOO = t, hence c = A.
 292 Measure, Integral and Probability

4.14 Let an -+ 0, an ~ 0. Then Px({y}) = limn-+oo Px((Y - an,y]) = Fx(y)-

limn-+ oo Fx(y - an), which proves the required equivalence. (Recall that
P x is always right-continuous.)
4.15 (a) Fx(Y) = 1 for y ~ a and zero otherwise.
(b) Fx(Y) °
= for y < 0, Fx(Y) = 1 for y ~ ~, and Fx(y) = 2y otherwise.
(c) Fx(y) = ° for y < 0, Fx(y) = 1 for y ~ ~, and Fx(y) = 1 - (1 - 2y)2
otherwise.
= x 3 , g-l(y) = flY, d~g-l(y) = ~y-~, hence fX3(y) =
4.16 In this case g(x)
I[O,l](fIY)h-~ = h-h[o,l](Y)'
4.17 (a) In adP = aP(fl) = a (constant function is a simple function).
1
(b) Using Exercise 4.15, fx(x) = 2 x l[o,!](x), so JE(X) = Io"2 2xdx = ~.
1
(c) Again by Exercise 4.15, Ix (x) = 4(1-2x) x l[o,!] (x), JE(X) = Io"24x(l-
2x) dx = ~.

4.18 (a) With fx(x) = b~a l[a,b], JE(X) = b~a I: xdx = b~a ~(b2 - a2) = ~(a +
b).
(b) Consider the simple triangle distribution with fx(x) = x + 1 for
x E [-1,0], fx(x) = -x + 1 for x E (0,1] and zero elsewhere. Then
immediately I~l xfx(x) dx = 0. A similar computation for the density fy,
whose triangle's base is [a, b], gives JE(X) = b. at
(c) A Iooo xe - AX dx = ±(integration by parts).
4.19 (a) cpx{t) = (b_1a)it (e ibt - eiat ),

(b) cpx(t) = AIooo e(it-A)xdx = A~it'

. t
(c) cpx(t) = ell'- 1 2 2
-"2<1 t .

4.20 Using call-put parity, the formula for the call and symmetry of the Gaus-
sian distribution we have P = S(0)(N(d 1 ) - 1) - Ke- rT (N(d 2 ) - 1) =
-S(O)N( -dd + Ke- rT N( -d2 )

Chapter 5
5.1 First, f == f as f = f everywhere. Second, if f = 9 a.e., then of course
9 = f a.e. Third, if f = 9 on a full set Fl C E (m(E \ F1 ) = 0) and 9 = h
on a full set F2 C E, then f = h on Fl n F2 which is full as well.
 Solutions 293

5.2 (a) Illn - Iml11 = 2 if m < n, so the sequence is not Cauchy.

(b) Illn - 1m 111 = I::
~ dx = logm -logn (for simplicity assume that n <
m), let c = 1, take any N, let n = N, take m such that logm -logN > 1
(log m --t 00 as m --t 00) - the sequence is not Cauchy.
(c) Illn - Iml11 = I:: tz
dx = -~I~ = ~ - ~ (n < m as before), and for
any c > 0 take N such that -k
< ~ and for n, m ~ N, clearly ~ - ~ < c -
the sequence is Cauchy.
5.3 119n - 9mll~ = I:: ~ dx = --bl~ = !(~ - ~3) - the sequence is Cauchy.
5.4 (a) Illn - Imll~ = m - n if n > m, so the sequence is not Cauchy.
(b) IIln - Imll~ = I:: tz dx = ~ - ~ --t 0 - the sequence is Cauchy.

(c) Illn - Imll~ = I:: ~ dx = (3~3 - 3r!.3) - the sequence is Cauchy.

5.5 III + 911 2 = 4, III - 911 2 = 1, 111112 = 1, 11911 2 = 1, and the parallelogram law
is violated.
5.6 III+ 911~ = 0, III - 911~ = ~, IIIII~ = ~, 11911~ = ~, which contradicts the
parallelogram law.
5.7 Since sinnxcosmx = ~[sin(n + m)x + sin(n - m)x]
and sinnxsinmx =
~[cos(n-m)x+cos(n+m)x], it is easy to compute the indefinite integrals.
They are periodic functions, so the integrals over [-71',71'] are zero (for the
latter we need n !- m).
5.8 No: take any n, m (suppose n < m) and compute
4 11/n
J.1..n (~) =(m-n)~l,
1

119n-9mll!= dx=-x- 1
Tn 11m

so the sequence is not Cauchy.

5.9 Let n = [0,1] with Lebesgue measure, X(w) = Jw, JE(X) = 10 X dm =
1

I; Jx dx = 2, JE(X2) = 101 ~ dx = 00. If we take X(w) = Jw - 2 then

JE(X) = 0 and JE(X2) = 00.

5.10 Var(aX) = JE((aX)2) - (JE(aX))2 = a2(JE(X2) - (JE(X))2) = a2Var(X).

5.11 Let Ix(x) = b~a l[a,bj, JE(X) = at b , VarX = JE(X2) - (a~b)2, JE(X2) =
b~a I: x 2 dx = b~a !(b3 - a3 ) and simple algebra gives the result.
5.12 (a) JE(X) = a, JE((X - a)2) = 0 since X = a a.s.

(b) By Exercise 4.15, Ix (x) = 2x l[o,!j(x) and by Exercise 4.17, JE(X) = ~;

1 1
so Var(X) = 102 2(x - ~)2 dx = 2 I..!! x 2 dx = 418'
4
 294 Measure, Integral and Probability

(c) By Exercise 4.15, fx(x) (4 - 8x)1[0,!J(x), and by Exercise 4.17,

lE(X) = Soi.

5.13 Cov(Y, 2Y + 1) = lE( (Y) (2Y + 1)) -lE(Y)lE(2Y + 1) = 2lE(y2) - 2(lE(y))2 =

2Var(Y), Var(2Y + 1) = Var(2Y) = 4Var(Y), hence p = 1.
5.14 X, Yare uncorrelated by Exercise 5.7. Take a > 0 so small that the sets
A = {w : sin27fw > 1 - a}, B = {w : cos27fw > 1 - a} are disjoint. Then
P(A n B) = 0 but P(A)P(B) =1= O.

Chapter 6
6.1 The function
-!, if 0 < y < x < 1
g(x, y) = { ~lfr if 0 < x < y < 1
o otherwise
is not integrable since the integral of g+ is infinite (the same is true for the
integral of g-). However,

=
ior
1
[1y ~x
1
dx _ ~ r dX] dy = ior
y io
1
[-1 +~-
y
~]
y
dy =-1

while similarly

10 1
[:2 loX dy - 11 :2 dy] dx = 1,

which shows that the iterated integrals may not be equal if Fubini's theorem
condition is violated.

6.2 I I[0,3J x [-1,2J x 2 y dm2 = I~l I: x 2 y dx dy = I~l 9y dy = 2:].

6.3 By symmetry it is sufficient to consider x ~ 0, y ~ 0, and hence the area
Io x
is 4~ a va 2 - 2 dx = ab7f.
Solutions 295

6.4 Fix x E [0,2], I; lA(x, y) dy = m(Ax), hence fx(x) = x for x E [0,1]'

fx(x) = 2 - x for x E (1,2J and zero otherwise (triangle distribution). By
symmetry, the same holds for fy.
6.5 P(X + Y > 4) = P(Y > -X + 4) = I IA fx,Y(x,y) dxdy where A =
{(x,y) : y > 4 - x} n [0,2J x [1,4]' so P(X + Y > 4) = I02 I44_x 510(X 2 +
t
y2)dydx = 510 I;(-4x 2 + x 3 + 16x)dx = 185' Similarly P(Y > X) =
I IA fx,Y(x, y) dxdy where A = {(x, y) : y > x} n [0, 2J x [1,4], so we get

6.6
z<O

fx+Y(z)~ J.fX'Y(X'Z-X)dX~{ ~-z O:=:;z:=:;l

1:=:;z:=:;2
2<z

6.7 By Exercise 6.4, fx,Y(x,y) =I- fx(x)fy(y), so X, Yare not independent.

6.8 fy+(-x)(z) = I~: fy(y)f-x(z - y)dy = I~: ~1[0,2l(y)1[_1,ol(z - y)dy,
so
0 z < -lor 2 < z
f =
() { ~(z + 1) -I:=:; z :=:; 0
y-X z 1
2
0 <
-
z < 1
-
~(2-z) 1:=:;z:=:;2
fx+y(z) = I~: fx(x)fy(z-x)dx = I~: 1[0,1l(x)~1[0,2l(z-x)dx, hence

z < 0 or 3 < z
O:=:;z:=:;l
1:=:;z:=:;2
2:=:;z:=:;3

P(Y > X) = P(Y - X > 0) = Iooo fy_x(z)dz = ~ + I12 ~(2 - z)dz =

~ +~ =~;
P(X + Y > 1) = It fx+y(z) dz = ~ + I23 ~(3 - z) dz = ~ + i = ~.
6.9 fx(x) = Io-~x+l1Ady = 1 - ~x, h(y,x) = 1:~~,~) and lE(YIX = 1) =
2
fl 1
2 J02 xdx = 4'
 296 Measure, Integral and Probability

6.10 Jy(y) = Jol(x + y) dx = ~ + y, h(xly)

fl
Jo
xy+Y
.,+y
dx = !t!Y.
2'+y

Chapter 7
7.1 If f.l(A) = 0 then AI(A) = 0 and A2(A) = 0, hence (AI + A2)(A) = O.
7.2 Let Q be a finite partition of n which refines both PI and P 2 • Each A E
PI and each B E P2 can be written as a disjoint union of in the form
A = Ui$;m Ei , B = Uj$;n Fj , where the Ei and Fj are sets in Q. Thus
A n B = Ui,j (Ei n F j ) is a finite disjoint union of sets in Q, since in
each case, Ei n F j equals either Ei or Fj (sets in Q are disjoint). Thus
Q refines the partition R = {A n B : A E PI, B E P 2 }. On the other
hand, any set in PI or P 2 is obviously a finite union of sets in R (e.g.
A = Ann = An (UBE'P2 B) for A E PI) as the partitions are finite.
7.3 We have to assume first that m(A) =f. O. Then B C A clearly implies
that f.l dominates v. (In fact m(B \ A) = 0 is slightly more general.) Then
consider the partition {B, A \ B, n \ A} to see that h = 1 B. To double
check, v(F) = m(F n B) = JFnB IB dm = JFnB IB df.l.
7.4 Clearly f.l( {w }) ~ v( {w }) is equivalent to f.l dominating v. For each w we
h ave dl/() _ ~lw})
dl-' w -~.

7.5 Since v(E) = JEgdm and we wish to have v(E) = JE ~~ df.l = JE ~~fdm
it is natural to aim at taking ~~ (x) = ~i:~. Then a sufficient condition for
this to work is that if A = {x : f(x) = O} then v(A) = 0, i.e. g(x) = 0
a.e. on A. Then we put ~~ (x) = ~i:~ on AC and 0 on A and we have
v(E) = JEnAc gdm = JEnAc ~~f dm = JE ~~ df.l, as required.
7.6 Clearly v« f.l is equivalent to A = {w: f.l({w}) = O} C {w: v({w}) = O}
and then ~~ (w) = :a~B on AC and zero on A.

7.7 Since v « f.l we may write h = ~~, so that v(F) = JF h df.l. As f.l(F) = 0
if and only if v(F) = 0, the set {h = O} is both f.l-null and v-null. Thus

with A = f.l to conclude that 1 = h-Ih implies *

h- l = (~~)-l is well-defined a.s., and we can use (ii) in Proposition 7.9
= h- l , as required.
7.8 80((0,25]) = 0, but 215ml[o,25]((0,25]) = 1; 215ml[0,25] ({O}) = 0 but
80 ({O}) = 1, so neither PI « P2 nor P2 « Pl. Clearly PI « P3 with
~(x) = 2 x l{o}(x) and P2 «P3 with ~(x) = 2 X 1(0,25] (x).
 Solutions 297

7.9 Aa = ml[2,3], As = Do + miu,2), and h = 1[2,3]'

7.10 Suppose F is non-constant at ai with positive jumps Ci, i = 1,2, .... Take
M -=I- ai, with -M -=I- ai and let 1= {i ; ai E [-M, M]}. Then

mF([-M,M]) = F(M) - F(-M) = I:Ci = LmF({ai}),

iEI iEI
which is finite since F is bounded on a bounded interval. So any A c
[-M, M] \ UiEI{ai} is mF-null, hence measurable. But {ad are mF-
measurable, hence each subset of [-M, M] is mF-measurable. Finally, any
subset E of IR is a union of the sets of the form E n [- M, M], so E is
mF-measurable as well.
7.11 mF has density f(x) = 2 for x E [0,1] and zero otherwise.
7.12 (a) Ixi = 1 + f~l f(y) dy, where f(y) = -1 for y E [-1,0]' and f(y) = 1
for y E (0,1].

(b) Let 1 > c > 0, take D = c 2, I:~=l (Yk - Xk) < D, with Yk :::; Xk+l; then
n
(L 1v'Xk - JYkI)2 :::; (yY: - yX1Y = Yn - 2Y!YnXI + Xl < Yn - Xl
k=l
n
:::; L(Yk - Xk) < c 2 •
k=l
(c) The Lebesgue function f is a.e. differentiable with f' = 0 a.e. If it were
absolutely continuous, it could be written as f(x) = foX f'(y) dy = 0, a
contradiction.

7.13 (a) If F is monotone increasing on [a, b], I:7=IIF(Xi) - F(Xi-I)1 = F(b)-

F(a) for any partition a = Xo < Xl < ... < Xk = b. Hence TF[a, b] =
F(b) - F(a).
(b) If FE BV[a,b] we can write F = F I -F2 where both F I , F2 are mono-
tone increasing, hence have only countably many points of discontinuity.
So F is continuous a.e. and thus Lebesgue-measurable.
(c) f(x) = x2 cos;;; for X -=I- 0 and f(O) = 0 is differentiable but does not
belong to BV[O, 1].

(d) If F is Lipschitz, I:7=1!F(Xi) - F(Xi-dl :::; MI:7=llxi - Xi-II =

M(b - a) for any partition, so TF[a,b] :::; M(b - a) is finite.
7.14 Recall that v+(E) = v(E n B), where B is the positive set in the Hahn
decomposition. As in the hint, if G S;; F, v(G) :::; v+(G n B) :::; v(G n B) +
v((F n B) \ (G n B)) = v(F n B). Since the set (F n B) \ (G n B) S;; B,
 298 Measure, Integral and Probability

its v-measure is non-negative. But F n B <:;:; F, so sup{v(G) : G <:;:; F} is

attained and equals v(FnB) = v+(F). A similar argument shows v-(F) =
sup{ -v(G)} = - infecF{v(G)}.
7.15 For all FE F, v+(F) = IBnFfdp, = sUPecFIe fdp,. If f > 0 on a set
C c AnF with p,(C) > 0, then Ie f dp, > 0, so that Ie f dp, + I BnF f dp, >
SUPeCF Ie f dp,. This is a contradiction since C U (B n F) c F. So f :::; 0-
a.s. (p,) on An F. We can take the set {f = O} into B, since it does not
affect the integrals. Hence {J < O} c A and {f 2: O} c B. But the two
smaller sets partition D, so we have equality in both cases.
Hence f+ = fIB and f- = - flA, and for all F E F

v+(F)=v(BnF)= r fdp,= r f+dp"

-1
iBnF iF
v-(F) = -v(AnF) =
AnF
fdp, =
iF
r f-dp,.
7.16 f E Ll(V) if and only if both I f+ dv and I r
dv are finite. Then
IE f+ 9 dp, and IE f- 9 dp, are well defined and finite, and their difference
is IE fgdp,. So fg E Ll(p,), as IE(f+ - r)lgl dp, < 00. Conversely, if
fg E U(p,) then both IE f+ Igl dp, and IE f-Igl dp, are finite, hence so is
their difference IE f dv.
7.17 (a) lE(XIQ)(w) = ~ if wE [0, ~J, lE(XIQ)(w) = ~ otherwise.

(b) lE(XIQ)(w) = w if wE [0, ~J, lE(XIQ)(w) = ~ otherwise.

7.18 lE(XnIFn-d = lE(Zl Z2 ... ZnlFn-d = Zl Z2 ... Zn-llE(ZnIFn-l), and

since Zn is independent of F n- 1 , lE(ZnIFn-d = lE(Zn) = 1, hence the
result.
7.19 lE(Xn) = np, =I- p, = lE(Xd so Xn is not a martingale. Clearly Yn = Xn -np,
is a martingale.
7.20 Apply Jensen's inequality to Xn = Zl + Z2 + ... + Zn. We obtain

so (Xn) is a submartingale. By (7.5) the compensator (An) satisfies LlAn =

lE((LlX~)IFn-d for each n. But here LlXn = Zn and Z~ == 1. So LlAn = 1
and hence An = n is deterministic.
7.21 For s < t, since the increments are independent and w(t) - w(s) has the
Solutions 299

same distribution as w(t - s),

lE(exp( -O"w(t) - ~a2t)) = e-o-w (s)_!o-2 tlE (exp( -[O"(w(t) - w(s)])l.:Fs )

= e-o- w (sl-!o-2 tlE (exp( -[O"(w(t) - w(s)])
= e-o- w (s)_!o-2 tlE (exp( -O"w(t - s))).
Now O"w(t - s) '-'"' N(O, 0"2(t - s)) so the expectation equals lE(e-o-y't=SZ) =
e-!o-2(t-s) (where Z'-'"' N(O, 1)) and so the result follows.

Chapter 8
8.1 (a) in = l[n,n+~l converges to 0 in LP, pointwise, a.e. but not uniformly.
(b) in = nl[o,~l - nl[_~,Ol converges to 0 pointwise and a.e. It converges
neither in LP nor uniformly.
8.2 We have n [0,1] with Lebesgue measure. The sequences Xn = 1(0,~),
=
Xn = nl(o,~l converge to 0 in probability since P(IXnl > c) ::; ~ and the
same holds for the sequence gn'
8.3 There are endless possibilities, the simplest being Xn(w) == 1 (but this
sequence converges to 1) or, to make sure that it does not converge to
anything, Xn(W) == n.
8.4 Let Xn = 1 indicate the 'heads' and Xn = 0 the 'tails', then ~OOo) is the
average number of 'heads' in 100 tosses. Clearly lE(Xn) =~, lE(~ooJ) =~,
~Tar(X ) - 1 Har(SlOO) - -1-100. 14 -- _1 so
v' n - 4' v' 100 - 100 2 400'

p(IS100_~1>01)< 1
100 2 - . - 0.1 2 400
and
S100 1 1 3
P(1 100 - "21 < 0.1) ~ 1 - 0.1 2 400 4:
8.5 Let Xn be the number shown on the die, lE(Xn) = 3.5, Var(Xn) ~ 2.9167.

P(I ~~~~ - 3.51 < 0.1) ~ 0.2916.

8.6 The union U:=n Am is equal to [0, 1] for all m and so is lim sUPn~oo An-
 300 Measure, Integral and Probability

8.7 Let d = 1. There are (~n) paths that return to 0, so P(S2n = 0) = (;,n) 2;n.
Now
(2n)! (~)2nJ21f2ri 22n
(n!)2 rv (~)2n27rn = vfn7r'

so P(S2n = 0) rv .In
with c = jI.
Hence 2:::=1 P(An) diverges and Borel-
Cantelli applies (as (An) are independent), so that P(S2n = 0 Lo.) = 1.
Same for d = 2 since P(An) rv ~. But for d > 2, P(An) rv nL2' the series
converges and by the first Borel-Cantelli lemma P(S2n = 0 Lo.) = O.
8.8 Write S = SlOOO; P(IS - 5001 < 10) = P( IS~OI < 0.63) = N(0.63) -
N( -0.63) ~ 0.4729, where N is the cumulative distribution function cor-
responding to the standard normal distribution.
8.9 The condition on n is P(I~ - 0.51
n
< 0.005) = P( Is~nl
n/4
< 0.01.jn) :?:
0.99, N(O.Ol.jn) :?: 0.995 = N(2.575835), hence n :?: 66350.
8.10 Write Xn = ea~. Then
1 1 1 (2Rnxn)2 rI. (1 + X~)
-2 (In Un + InDn) = -In(UnDn) = -In ( 2)2 = lne n -In -2-- .
2 2 1 + xn Xn

So it suffices to show that the last term on the right is a;;; + o(~). But
1 +X2 X-I +X ea~ +ea~
__n = n n = = cosh(ay'T/n)
2x n 2 2
a 2T 1
= 1+ - +0(-),
2n n
so that
1+~ ~T 1 ~T 1
In(--n) = In(l + - + 0(-)) = - + 0(-).
2x n 2n n 2n n
 Appendix

Existence of non-measurable and non-Borel sets

In Chapter 2 we defined the a-field B of Borel sets and the larger a-field M
of Lebesgue-measurable sets, and all our subsequent analysis of the Lebesgue
integral and its properties involved these two families of subsets of K The set
inclusions
Be M c P(l~)

are trivial; however, it is not at all obvious at first sight that they are strict,
i.e. that there are sets in ~ which are not Lebesgue-measurable, and as that
there are Lebesgue-measurable sets which are not Borel sets. In this appendix
we construct examples of such sets. Using the fact that A c ~ is measurable
(resp. Borel-measurable) iff its indicator function lA EM (resp. B), it follows
that we will automatically have examples of non-measurable (resp. measurable
but not Borel) functions.
The construction of a non-measurable set requires some set-theoretic prepa-
ration. This takes the form of an axiom which, while not needed for the consis-
tent development of set theory, nevertheless enriches that theory considerably.
Its truth or falsehood cannot be proved from the standard axioms on which
modern set theory is based, but we shall accept its validity as an axiom, without
delving further into foundational matters.

301
302 Measure, Integral and Probal."_ J

The axiom of choice

Suppose that A = {Ac, : a E r} is a non-empty collection, indexed by some
set r, of non-empty disjoint subsets of a fixed set D. There then exists a set
E c D which contains precisely one element from each ofthe sets Ac", i.e. there
is a choice function f : r -t A.

Remark
The Axiom may seem innocuous enough, yet it can be shown to be independent
of the (Zermelo-Fraenkel) axioms of sets theory. If the collection A has only
finitely many members there is no problem in finding a choice function, of
course.

To construct our example of a non-measurable set, first define the following

equivalence relation on [0, 1J: x r-v y if y - x is a rational number (which will be
in [-1, 1]). This relation is easily seen to be reflexive, symmetric and transitive.
Hence it partitions [O,lJ into disjoint equivalence classes (An), where for each
a, any two elements x, y of An differ by a rational, while elements of different
classes will always differ by an irrational. Thus each An is countable, since IQl
is, but there are uncountably many different classes, as [0, 1J is uncountable.
Now use the axiom of choice to construct a new set E c [O,lJ which con-
tains exactly one member an from each of the An. Now enumerate the ra-
tionals in [-1, 1J: there are only count ably many, so we can order them as a
sequence (qn). Define a sequence of translates of E by En = E + qn. If E is
Lebesgue-measurable, then so is each En and their measures are the same, by
Proposition 2.15.
But the (En) are disjoint: to see this, suppose that Z E Em n En for some
m =1= m. Then we can write an + qm = Z = a(3 + qn for some an, a(3 E E, and
their difference an - a(3 = qn - qm is rational. Since E contains only one element
from each class, a = f3 and therefore m = n. Thus U~=l En is a disjoint union
containing [0, IJ.
Thus we have [O,lJ C U~=l En c [-1,2J and m(En) = m(E) for all n. By
countable additivity and monotonicity of m this implies:

L m(En) = m(E) + m(E) + ... :::; 3.

00

1 = m([O, 1]) :::;

n=l
This is clearly impossible, since the sum must be either a or 00. Hence we must
conclude that E is not measurable.
For an example of a measurable set that is not Borel, let C denote the Cantor
set, and define the Cantor function f : [0, 1J -t C as follows: for x E [0, 1J write
x = O.ala2 ... in binary form, i.e. x = L~=l ~,where each an = a or 1 (taking
Appendix 303

non-terminating expansions where the choice exists). The function x H an is

determined by a system of finitely many binary intervals (i.e. the value of an
is fixed by x satisfying finitely many linear inequalities) and so is measurable
- hence so is the function f given by f(x) = 2::=1 ~. Since all the terms of
y = 2::=1 23ann have numerators 0 or 2, it follows that the range Rf of f is a
subset of C. Moreover, the value of y determines the sequence (an) and hence
x, uniquely, so that f is invertible.
Now consider the image in C of the non-measurable set E constructed
above, i.e. let B = f(E). Then B is a subset of the null set C, hence by the
completeness of m it is also measurable and null. On the other hand, E =
f-1(B) is non-measurable. We show that this situation is incompatible with B
being a Borel set.
Given a set B E B and a measurable function g, then g-1(B) must be
measurable. For, by definition of measurable functions, g-1 (1) is measurable
for every interval I, and we have

g-1(U Ai) = Ug-1(A i ),

00 00

i=1 i=1
quite generally for any sets and functions. Hence the collection of sets whose
inverse images under the measurable function 9 are again measurable forms a
a-field containing the intervals, and hence it also contains all Borel sets.
But we have found a measurable function f and a Lebesgue-measurable set
B for which f-1(B) = E is not measurable. Therefore the measurable set B
cannot be a Borel set, i.e. the inclusion B c M is strict.
 References

[1] T.M. Apostol, Mathematical Analysis, Addison-Wesley, Reading, MA,

1974.
[2] P. Billingsley, Probability and Measure, Wiley, New York, 1995.
[3] Z. Brzezniak, T. Zastawniak, Basic Stochastic Processes, Springer, Lon-
don, 1999.
[4] M. Capinski, T. Zastawniak, Mathematics for Finance, An Introduction
to Financial Engineering, Springer, London, 2003.
[5] R.J. Elliott, P.E. Kopp, Mathematics of Financial Markets, Springer, New
York, 1999.
[6] G.R. Grimmett, n.R. Stirzaker, Probability and Random Processes, Cla-
rendon Press, Oxford, 1982.
[7] J. Hull, Options, Futures, and Other Derivatives, Prentice Hall, Upper
Saddle River, NJ, 2000.
[8] P.E. Kopp, Martingales and Stochastic Integrals, Cambridge University
Press. Cambridge, 1984.
[9] P.E. Kopp, Analysis, Edward Arnold, London, 1996.
[10] J. Pitman, Probability, Springer, New York, 1995.
[11] W. Rudin, Real and Complex Analysis, McGraw-Hill, New York, 1966.
[12] G. Smith, Introductory Mathematics: Algebra and Analysis, Springer, Lon-
don, 1998.
[13] n. Williams, Probability with Martingales, Cambridge University Press,
Cambridge, 1991.

305
 Index

a.e., 55 Borel-Cantelli lemma

- convergence, 242 - first, 256
a.s., 56 - second, 258
absolutely continuous, 189 bounded variation, 207
- function, 109, 204 Brownian motion, 234
- measure, 107 BV[a, b], 207
adapted, 223
additive call option, 70
- countably, 27 - down-and-out, 72
additivity call-put parity, 117
- countable, 29 Cantor
- finite, 39 - function, 302
- of measure, 35 - set, 19
almost everywhere, 55 Cauchy
almost surely, 56 - density, 108
American option, 72 - sequence, 11, 128
angle, 138 central limit theorem, 276, 280
central moment, 146
Banach space, 136 centred
Beppo-Levi - random variable, 151
- theorem, 95 characteristic function, 116, 272
Bernstein Chebyshev's inequality, 247
- polynomials, 250 complete
Bernstein-Weierstrass - measure space, 43
- theorem, 250 - space, 128
binomial completion, 43
- tree, 50 concentrated, 197
Black-Scholes conditional
- formula, 119 - expectation, 153, 178, 179, 219
- model, 118 - probability, 47
Borel contingent claim, 71
- function, 57 continuity of measure, 39
- measure, regular, 44 continuous
- set, 40 - absolutely, 204

307
308 Index

convergence - natural, 223

- almost everywhere, 242 first hitting time, 231
- in pth mean, 144 formula
- in LP, 242 - inversion, 180
- in probability, 245 Fourier series, 140
- pointwise, 242 Fubini's theorem, 171
- uniform, 11, 241 function
- weak, 268 - Borel, 57
correlation, 138, 151 - Cantor, 302
countable - characteristic, 116, 272
- additivity, 27, 29 - Dirichlet, 99
covariance, 151 - essentially bounded, 141
cover, 20 - integrable, 86
- Lebesgue, 20
de Moivre-Laplace theorem, 280 - Lebesgue measurable, 57
de Morgan's laws, 3 - simple, 76
density, 107 - step, 102
- Cauchy, 108 fundamental theorem of calculus, 9, 97,
- Gaussian, 107, 174 215
- joint, 173 futures, 71
- normal, 107, 174
- triangle, 107 gamma distribution, 109
derivative Gaussian density, 107, 174
- Radon-Nikodym, 194 geometric distribution, 69
derivative security, 71
- European, 71 Holder inequality, 142
Dirac measure, 68 Hahn-Jordan decomposition, 211, 217
direct sum, 139 Helly's theorem, 270
distance, 126 Hilbert space, 136, 138
distribution
- function, 109, 110, 199 Lo., 255
- gamma, 109 identically distributed, 244
- geometric, 69 independent
- marginal, 174 - events, 48, 49
- Poisson, 69 - random variables, 70, 244
- triangle, 107 - 0'- fields, 49
- uniform, 107 indicator function, 4, 59
dominated convergence theorem, 92 inequality
dominating measure, 190 Chebyshev, 247
Doob decomposition, 227 - Holder, 142
- Jensen, 220
essential - Kolmogorov, 262
- infimum, 66 - Minkowski, 143
- supremum, 66, 140 - Schwarz, 132, 143
essentially bounded, 141 - triangle, 126
event, 47 infimum, 6
eventually, 256 infinitely often, 255
exotic option, 72 inner product, 135, 136
expectation - space, 136
- conditional, 153, 178, 179, 219 integrable function, 86
- of random variable, 114 integral
- improper Riemann, 100
Fatou's lemma, 82 - Lebesgue, 77, 87
filtration, 51, 223 - of a simple function, 76
Index 309

- Riemann, 7 - outer, 20, 45

invariance - probability, 46
- translation, 35 - product, 164
inversion formula, 180 - regular, 44
Ito isometry, 229 - a-finite, 162
- signed, 209, 210
Jensen inequality, 220 - space, 29
joint density, 173 measures
- mutually singular, 197
Kolmogorov inequality, 262 metric, 126
Minkowski inequality, 143
L 1 (E),87 model
L2(E), 131 - binomial, 50
LP convergence, 242 - Black-Scholes, 118
LP(E),140 - CRR,234
LOO(E), 141 moment, 146
law of large numbers monotone class, 164
- strong, 260, 266 - theorem, 165
- weak, 249 monotone convergence theorem, 84
Lebesgue monotonicity
- decomposition, 197 - of integral, 81
- function, 20 - of measure, 21, 35
- integral, 76, 87 Monte Carlo method, 251
- measurable set, 27 mutually singular measures, 197
- measure, 35
Lebesgue-Stieltjes negative part, 63
- measurable, 202 negative variation, 207
- measure, 199 norm, 126
lemma normal density, 107, 174
- Borel-Cantelli, 256 null set, 16
- Fatou, 82
- Riemann-Lebesgue, 104 option
Levy's theorem, 274 - American, 72
liminf,6 - European, 70
limsup,6 - exotic, 72
Lindeberg-Feller theorem, 276 - lookback, 72
lower limit, 256 orthogonal, 137-139
lower sum, 77 orthonormal
- basis, 140
marginal distribution, 174 - set, 140
martingale, 223 outer measure, 20, 45
- transform, 228
mean value theorem, 81 parallelogram law, 136
measurable partition, 190
- function, 57 path, 50
- Lebesgue-Stieltjes, 202 pointwise convergence, 242
- set, 27 Poisson distribution, 69
- space, 189 polarisation identity, 136
measure, 29 portfolio, 183
- absolutely continuous, 107 positive part, 63
- Dirac, 68 positive variation, 207
- F -outer, 201 power set, 2
- Lebesgue, 35 predictable, 226
- Lebesgue-Stieltjes, 199 probability, 46
310 Index

- conditional, 47 - inner product, 136

distribution, 68 - L2(E), 131
measure, 46 - £p(E), 140
- space, 46 - measurable, 189
probability space - measure, 29
- filtered, 223 - probability, 46
process standard normal distribution, 114
- stochastic, 222 step function, 102
- stopped, 231 stochastic
product - integral, discrete, 228
- measure, 164 - process, discrete, 222
- a-field, 160 stopped process, 231
Prokhorov's theorem, 272 stopping time, 230
put option, 72 strong law of large numbers, 266
subadditivity, 24
Radon-Nikodym submartingale, 224
- derivative, 194 summable, 218
- theorem, 190, 195 supermartingale, 224
random time, 230 supremum, 6
random variable, 66 symmetric difference, 36
- centred, 151
rectangle, 3 theorem
refinement, 7, 190 - Beppo-Levi, 95
replication, 233 - Bernstein-Weierstrass approximation,
return, 183 250
Riemann - central limit, 276, 280
- integral, 7 - de Moivre-Laplace, 280
-- improper, 100 - dominated convergence, 92
Riemann's criterion, 8 Fubini, 171
Riemann-Lebesgue lemma, 104 - fundamental of calculus, 9, 97, 215
Helly,270
Schwarz inequality, 132, 143 - intermediate value, 6
section, 162, 169 - Levy, 274
sequence Lindeberg-Feller, 276
- Cauchy, 11, 128 mean value, 81
- tight, 272 - Miller-Modigliani, 118
set - monotone class, 165
- Borel, 40 - monotone convergence, 84
- Cantor, 19 Prokhorov, 272
- Lebesgue-measurable, 27 Radon-Nikodym, 190, 195
- null, 16 - Skorokhod representation, 110, 269
a-field, 29 tight sequence, 272
- generated, 40 total variation, 207
- - by random variable, 67 translation invariance
- product, 160 - of measure, 35
a-finite measure, 162 - of outer measure, 26
signed measure, 209, 210 triangle inequality, 126
simple function, 76 triangular array, 282
Skorokhod representation theorem, 110,
269 uncorrelated random variables, 151
space uniform convergence, 11, 241
- Banach, 136 uniform distribution, 107
- complete, 128 upper limit, 255
- Hilbert, 136, 138 upper sum, 77
Index 311

variance, 147 - total, 207, 212

variation
- bounded, 207 weak
- function, 207 - convergence, 268
- negative, 213 - law of large numbers, 249
- positive, 213 Wiener process, 234