Lecture Notes On Probability Theory
Lecture Notes On Probability Theory
CONTENTS
Contents
4
1 Preliminaries
4
18
24
35
38
47
57
62
66
8.1
8.2
8.3
8.4
8.5
8.6
8.7
..
..
..
..
..
..
..
9.1
9.2
9.3
9.4
10 Martingales
10.1
10.2
10.3
10.4
10.5
10.6
A.BENHARI
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
66
74
81
.81.
.82.
.87.
.89.
.91.
.94.
.94.
96
.96.
.102.
.104.
.109.
110
.110.
.113.
.115.
.117.
.120.
.126.
11.1
11.2
11.3
11.4
.
.
.
.
.
.
.
.
..
..
..
..
.
.
.
.
..
..
..
..
.
.
.
.
..
..
..
..
128
128
132
136
139
141
13 Exer ises
146
BNibliography
156
A.BENHARI
1 Preliminaries
1.1
2. P(
i=1
Ai ) =
1
P
i=1
3. P(A2 n A1 ) = P(A2 )
4. P(A) = 1
P(A)
1
S
5. P(
i=1
1
T
P(Ai ); if Ai Aj = ; 8 i; j with i 6= j
P(A1 ), if A1 A2
Ai ) = ilim
P(Ai ); if Ai Ai+1
!1
8i
Ai ) = ilim
P(Ai ); if Ai Ai+1 8 i
!1
Denition 1.2 A random variable (r.v.) on a probability spa
e f
; A; Pg
is a A-measurable real fun
tion of the elementary events ! : (! ).
6. P(
i=1
Remark 1.1
1. Let A be a -algebra of
; E 2 A.
A real valued fun
tion f : E ! R is
alled measurable on E if and only if
f 1 (G) 2 A 8 open subsets G R:
2. f : E ! R is measurable on E if and only if 8 a 2 R : f 1 (( 1; a)) 2 A:
lim
n!1 n
= lim
a.s.,
lim
n n
n n
see A =
= lim
a.s.
lim
n n
n n
9g is A-measurable
1
1 S
T
f! : jm
m
j < k1 g
Remark 1.2 h(t1 ; : : : ; tn) is
alled Borel - fun
tion, if it is measurable with
respe
t to the -algebra of Rn , whi
h is generated by the open subsets of R n .
Denition 1.4 A r.v. is
alled dis
rete, if there are at most
ountable many
values:
P
=
k IAk (! ) Ak 2 A;
k
S
k
Ak =
; Ai \ Aj = ;
8 i 6= j
Theorem 1.3 Let be an arbitrary r.v., then there exists a sequen
e of dis
rete
r.v. (n )n1 with nite number of values, su
h that
lim (! ) = (! ) 8 ! 2
:
n!1 n
Proof:
n =
n
nP1 P
i= n k=1
for
< i + nk g
) j nj = j (i + k n 1 )j < n1
n
-1
i i+1
n-1 n
assertion
Theorem 1.4 Let be a nonneagtive r.v.. Then there exists a monotone in
reasing sequen
e of dis
rete r.v.'s (with
ountable many values), whi
h
onverges
n=2
uniformly to .
Proof:
n :=
0
1
P
i
2n IAin
i=0
1 2
= 252
with Ain := f! :
i
2n
<
i+1
2n
If R is a metri
spa
e, then B denotes the -algebra of Borel sets (if not
denoted otherwise).
n
Q
k=1
n
Q
Rk
and C = Bk
(produ
t of -algebras B1 ; : : : ; Bn ).
k=1
The r.e. is
alled dire
t produ
t of the r.e.'s 1 ; : : : ; n :
Remark 1.3
1.
n
Q
k=1
Rk := fy : y = (x1 ; : : : ; xn ); xi 2 Ri 8 ig
2. proje
tion: i :
3.
n
Q
k=1
n
Q
k=1
Rk ! Ri : i (y) = xi
Bk := ( Q Bk ; Bk 2 Bk )
k=1
Q
2
Bk with Bk 2 Bk .
R = fy = y( ) : ! ; 2 R ; 2 g
k=1
Generalization:
S=
n
Q
B = Q B
2
Example 1.1
sto k pri e
JJ
bb""
LL
!!
Let (! ) be a r.e. on
time
daily values ) N [ f0g
(time series, random sequen
e)
R = (0; 1)
) v = (vx; vy ; vz )T
) R3 [0; 1); R R3
) [0; 1);
time
= [0; 1);
6
) R2 ; R R
(
(( ""ll(
(((XXX
!
" !X
"
(
(
X
(
( XXXX(%%(((HH%%3
(2) ((
H((
(1)
f
; A; Pg
with values in
fS ; Bg
S
i
Ai 2 C
(
; A)
S
(
; A )
(!)
! (S ; B)
#g
.
! (R; C )
1
1 (Bn0 ) = An n
nS1
Ai = An
i=1
1
1
1( S B0 ) = S A
n
n
i=1
i=1
) (
) S Bn0
i=1
) = g( )
!1 n
due to 1. 9 gn with n = gn( ) and gn (x) B - measurable
Theorem 1.2 4. ) S := fx : nlim
g (x) 9g 2 B (
onsidered more general)
!1 n
it holds nlim
(! ) = nlim
g ( (! )) = (! ) ) (
) S
!1 n
!1 n
(see nlim
= 8 ! 2
)
!1 n
) for x 2 (
) : nlim
g (x) = nlim
g ( (! )) = nlim
(! ) = (! )
!1 n
!1 n
!1 n
hen
e
(
lim gn (x) for x 2 S
g (x) := n!1
) = g( )
0
for x 62 S
(see g ( (! )) = nlim
g ( (! )) = (! ))
) assertion
!1 n
Denition 1.7 The expe
tation (or mean value) of the r.v. (! ) is the integral
of (! ) with respe
t to the probability measure P :
R
R
E = (! )P(d! ) = dP:
(Lebesgue-integral,
f -probability density)
P
E = xk P( = xk ) (see measure theory, se
tion 2)
1
dis rete :
We have introdu
ed
onvergen
e with probability 1 (a.s.
onvergen
e) for sequen
es of r.v.'s (see Def. 1.3).
Denition 1.8 The sequen
e of r.v.'s (n )n1
onverges in probability to the
r.v. , if
lim P(jn j > ") = 0 for every " > 0:
n!1
A.BENHARI
notation:
= Pn!1lim n
interpretation: For every " > 0 the distribution of the "probability mass" of n
for in
reasing n
on
entrates in a "-ball around zero.
then
= mod P.
( = a.s.)
Theorem 1.7
1
X
n=1
A.BENHARI
Denition 1.9
Let for p 1
Lp = Lp(
; A; P) = f r.v. on f
; A; Pg : E j jp < 1g
A sequen
e (n)n1 is
alled Lp-
onvergent (or
onvergent in mean of order
p), if
lim E j njp = 0:
n!1
Remark 1.5 For p 1 Lp is a
omplete spa
e (i.e. Cau
hy sequen
es are
onvergent) and it
an be equipped with the norm : k k = (E j jp) p : ) Lp is a
1
E j njp
for
"p
p 1:
l.i.m.
= (limit in mean)
n!1 n
B (t1 ; t2 ) := E t t
1
t1 ; t2 2 T
is alled ovarian e.
Denition 1.11 Let (t) be a nonegative real fun
tion with D( ) = T whi
h
may have arbitrarily small values. For a family ft ; t 2 T g with t 2 L2 the
r.v. 2 L2 is
alled L2 -limit or mean-square limit for (t) ! 0 if for all
" > 0 there exists a > 0 su
h that
E jt j2 " 8 t with 0 (t) < .
Example 1.2 T = R 1 ; (t) = jt t0 j;
is the mean-square limit of t for t ! t0
Theorem 1.9 There exists the mean-square limit of the family ft ; t 2 T g
for (t) ! 0 if and only if there exists the limit of the
ovarian
e B (t1 ; t2 ) =
E t t for
(t1 )+ (t2 ) ! 0. In this
ase it holds E j j2 = lim B (t; t):
1
(t)!0
A.BENHARI
Proof:
ne
essity: to prove:
9 -l.i.m.
t = ) 9
lim
B (t; t2 ) (= [; )
(t)!0
(t )+ (t )!0
k t k k t k j[ t ; t j
(Cau
hy-inequality)
jj
jj
= j[; [; t [t ; + [t ; t j
jj
jj
= j [; [; t [t ; + [t ; t j
j
j
j
j
(t1 ) ! 0
(t2 ) ! 0
(t2 ) ! 0
(t1 ) ! 0
#
#
#
#
1
0
0
0
(t )+ (t )!0
[t ; t = E t t = B (t1 ; t2 )
! [;
(follows from
ontinuity of s
alar produ
t)
1
su ien y: it
E jt
lim
!
(t)!0
2
2
) kk ktk = kk ktk(kk + ktk)
k tk kk + k t kktk
) kk2
(t)!0
t )(t t )
= E t t E t t E t t + E t t
1
L2 omplete
= E j j2 =
(t)!0
! 0
(see jkf k kg kj kf g k)
lim B (t; t) (see kt k = [t ; t = E t t = B (t; t))
(t)!0
) assertion
denotes the Hilbert spa
e of random ve
tors with values in the
omplex
spa
e Cm
m
P
2 Lm2 if
= ( 1 (! ); : : : ; m (! )) 2 Cm and E j i j2 < 1
; 2 Lm2
B (t1 ; t2 ) := [t ; t = E
1
m
P
ti1 ti2
i=1
m
P
ii;
i=1
i=1
m
P
i=1
j ij2
10
f ( (! ))P(d! )) =
f (x)(dx)
Remark 1.6 The above theorem states that arbitrary sto
hasti
hara
teristi
s
E f ( )
an be
omputed as integrals over S using the distribution of .
restri
tion to distributions in Rm (
an be generalized to general metri
spa
es)
Bm : -algeba of Borel sets of Rm
a = (a1 ; a2 ; : : : ; am ) 2 Rm ;
b = (b1 ; b2 ; : : : ; bm ) 2 Rm
a < b (resp a b)
! ai < bi (resp. ai bi); i = 1; 2; : : : ; m
Ia = fx : x ag;
I (a; b = fx : a < x bg
omputation of P( 2 I (a; b) = (I (a; b) using the distribution fun
tion:
let G be an arbitrary fun
tion on Rm , ; 2 R 6
(b1 ; b2 )
(k)
1
k
1
k
+1
m
(; G(x) := G(x ; : : : ; x ; ; x ; : : : ; x )
(a1 ; a2 )
G(x1 ; : : : ; xk 1 ; ; xk+1; : : : ; xm )
(k)
) (; F (x) =
P( 1 x1 ; : : : ; k 1 xk 1 ; < k ; k+1 xk+1 ; : : : ; m xm )
m
()
(am ;bm F (x)
Using the subsequent properties of distribution fun
tions it is also possible to
onsider open intervals I (a; b) and
losed intervals I [a; b:
1
Theorem 1.11 (ne
essary and su
ient
onditions for distribution fun
tions)
1. A distribution fun
tion F (x) = (Ix) = P( x) posseses the following
properties
1. 0 F (x) 1
2. F (x) F (y ) for x y
11
5. F (x) ! 0, if xk !
6.
2. For an arbitrary fun
tiuon F (x) on Rm satisfying
onditions 1.-6. there exists
a unique measure on Bm with a distribution fun
tion whi
h
oin
ides with F (x).
Proof: see basi
ourse Theorem 2.6 and 2.7
Rm
1: n (Rm ) C
2: 8 " > 0 9 an interval I (a; b with sup n(Rm n I (a; b) < ":
n
J (u) =
E ei(u;)
ei(u;x) (dx);
u = (u1 ; : : : ; um )
Rm
m
P
uj xj )
j =1
Theorem 1.13 Chara
teristi
fun
tions possess the following properties:
1. J (0) = 1; jJ (u)j 1,
2. J (u) is uniformly
ontinous 8 u 2 Rm ,
3. J (u) is positive semi-denite, i.e. 8 n 2 N; 8 zk 2 C; 8 uk 2 Rm it holds
n
P
J (uk ul )zk zl 0.
k;l=1
Every fun
tion satisfying properties 1.-3. is a
hara
teristi
fun
tion of a
ertain
distribution.
Proof: see basi
ourse Theorem 6.2, 6.6
Question: Is the distribution uniquely determined by its
hara
teristi
fun
tion ?
A.BENHARI
12
1 = 2 :
Rm
Relation between weak onvergen e of distributions and onvergen e of hara teristi fun tions:
Proof:
1. 1 ; 2 independent ) E ei(u; + ) = E ei(u; ) E ei(u; )
2. ne
essitiy: follows from 1.
su
ien
y: follows form the unique relation between the
hara
teristi
fun
tion and the distribution (see Theorem 1.14).
1
A.BENHARI
13
i)q
q J (u)
where q = j1 + : : : + js for q p:
uj1 : : : ujss u=0
1
Conversely, if for an even p 2 N the
hara
teristi
fun
tion J (u1 ; : : : ; us) is ptimes dierentiable at u = 0 then there exist all moments of order q p and it
holds the above relation.
Proof:
part 1: follows from J (u) = E ei(u;) and the inter
hangeability of dierentation
and expe
tation (integation w.r.t.P)
part 2: see basi
ourse, Theorem 6.2, property 6
Stopping times
Example 1.3 sto
k pri
e
(t ; t 2 [0; T )
t
a reasonable question:
when does the pri
e rea
h a
ertain
level for the rst time
no reasonable question:
when does the pri
e rea
h
its maximum value in [0; T
(
an not be answered before T )
JJ
bb""
a reasonable question:
at whi
h time a
ertain set B has
been
rossed for the rst time
no reasonable question:
at whi
h time a
ertain set B has
been
rossed for the last time
t
!! LL
#
"!
B
AA BB AA
A
JJ
B
J
B
A.BENHARI
14
f! : tg 2 Ft 8 t 2 T:
Remark 1.9
1. f! : tg 2 Ft ,
"stopping time
t"
an be de ided
2.
: domain of denition of the stopping time with values in T
3. f tg 2 Ft
T
ountable:
() f > tg 2 Ft
f tg 2 Ft () f = tg 2 Ft 8 t 2 T
F := fB 2 A : B \ f tg 2 Ft 8 t 2 T g ) F is a -algebra
F is the -algebra whi
h is generated by the stopping time .
The stopping time is F -measurable (B =
).
Example 1.5 = t0 ; t0 2 T is a stopping time
;
for
t
<
t
0
) f tg =
for t t0
) B \ f tg = ;B tt < tt00
) F = Ft see
t=t
1. B 2 F ; i.e. 8 t B \ f tg 2 Ft =
) B \ f t0g = B 2 Ft
i.e. F Ft
2. B 2 Ft : t < t0 ) B \ f tg = ; 2 Ft ;
t t0 ) B \f tg = B 2 Ft Ft , i.e. B 2 F ) F Ft
0
15
2 K g is Ft-measurable.
Let (t) : T ! T be a Borel fun
tion with (t) t 8 t 2 T then ( ) is a
x2K
stopping time.
3. Let 1 ; 2 be stopping times then min(1 ; 2 ); max(1 ; 2 ) are stopping times
espe
ially: if is a stopping time then it is min(; t0 ); t0 2 T:
4. Let 1 ; 2 be stopping times then 1 2 implies F F .
1
Proof:
1. f
sup x t;
x2Alk
pl
1 T
f 2 K g = S
f 2 Alk g Ft-measurable
l=1 k=1
2Ft
4.
| {z }
2Ft
k t
S
(f! : k < xg \ f! : = k g)
k t |
) f! : < xg 2 F
A.BENHARI
16
{z
2F k
{z
2F k
2 Ft
assertion
1.2
Independen e
4. P(A \
n
S
i=1
n
S
n
P
i=1
i=1
Bi ) = P( (A \ Bi )) =
= P(A)P(
n
S
i=1
Bi )
P(A \ Bi ) =
n
P
i=1
P(A)P(Bi )
(see (A \ Bk ) \ (A \ Bh ) = ; for k 6= h)
For n = 2; 3; : : : the random events A1 ; : : : ; An 2 A are alled (jointly) independent if for k = 1; : : : ; n and 1 i1 < i2 : : : < ik n it holds
17
Remark 1.10
1. Note, that the pairwise independen
e of events A1 ; : : : ; An does not imply
their jointly independen
e.
2. In the above denition the set I may have an innite number of elements.
for any n = 2; 3; : : : ; and pairwise dierent i1 ; : : : ; in 2 I; the joint distribution fun
tion of i ; : : : ; in is equal to the produ
t of the distribution
fun
tions of ik :
1
P(i a1 ; i a2 ; : : : ; in an ) =
1
n
Y
k=1
P(ik ak ); a1 ; : : : ; an 2 R:
are independent.
f! : i a1 ; : : : ; in an g; n = 1; 2; : : : ; ik 2 I ; ak 2 R
is
alled the -algebra generated by the
lass of r.v.'s fi ; i 2 I g.
Remark 1.11 fi ; i 2 I g is the minimal -algebra with respe
t to whi
h all
1
the i are random variables, i.e. all the fun
tions i (! ) are measurable.
Spe
ial
ase: a single r.v. (I
ontains one element only)
) f g = f f! : ag; a 2 Rg
A.BENHARI
18
Remark 1.12
on
ept of
losure: see measure theory, Theorem 1.6 shows that
the
losures ~ f: : :g are -algebras
Theorem 1.21 Let g (t1 ; : : : ; ts); 2 J; be a set of Borel fun
tions of s real
variables (g : Rs ! R). If the
lasses of r.v.'s f1; : : : ; s g; 2 J; are independent, then the r.v.'s := g (1 ; : : : ; s ); 2 J; are also independent.
Proof: Denition 1.24 ) the
lasses of events
M
N
n
\
k=1
fik 2 Bik g
n
Y
k=1
P (ik 2 Bik ) :
k=1
is the minimal
n
Q
k=1
Sk ; B
(n)
19
n
Q
k=1
Sk ; B
(n)
Conversely, if m12:::n
oin
ides with the produ
t of the measures m1 ; : : : ; mn then
the r.e.'s 1 ; : : : ; n are independent. )
Theorem 1.22 The r.e.'s 1 ; : : : ; n are independent if and only if the measure
m1:::n indu
ed by the sequen
e (1 ; : : : ; n) on the -algebra B(n) is the produ
t of
the measures mk ; k = 1; 2; : : : ; n; indu
ed by the elements k on Bk .
Theorem 1.23 Let g (x1 ; x2 ) be a B(2) -measurable nite real fun
tion and 1 ; 2
be independent r.e.'s with values in fSi ; Bi g; i = 1; 2; and let, moreover,
E jg (1; 2 )j < 1: Then (x1 ) = E g (x1 ; 2 ) is a B1 -measurable fun
tion and
or
E g (1 ; 2) =
S1
m1 (dx1 )
S2
Proof:
g (x1 ; x2 ) is a B(2) -measurable fun
tion, i.e. g 1 (G) 2 B(2)
E g (1; 2 )
=
independen
e
Fubini
S2
S1 S
Z 2
SZ1 S2
S1
8 open subsets G R
m1 (dx1 )
S2
Corollary 1.1 If 1 and 2 are independent r.v.'s with nite expe
tations, then
E 1 2 = E 1 E 2
(see g (x1 ; x2 ) = x1 x2 ):
0-1-laws
Theorem 1.24 Let (An )n1 be a sequen
e of events.
1
P
If
P(An ) < 1; then the event lim
An is of probability 0.
n
n=1
Proof: it holds
An = f! 2
: ! 2 An for innite many ng =
lim
n
1 S
1
T
1
S
1
P
1 S
1
T
m=1 n=m
An :
An ) = mlim
P( A ) mlim
P(An ) = 0
!1 n=m n
!1 n=m
m=1 n=m
(see Theroem 1.1, 6. and subadditivity of P)
The following stronger result is valid in
ase of independent events.
P(lim
An ) = P(
n
A.BENHARI
20
for
(ii) P(lim
An ) = 1; for
n
1
P
k=1
1
P
k=1
P(Ak ) < 1
P(Ak ) = 1:
Proof: Theorem 1.24 ) (i), for (ii) see basi ourse Theorem 5.7.
Ck =: lim
Dn is also a -algebra (tail or terminal -algebra)
n
Dn remains un
hanged if a nite number of -algebras Dk is
hanged.
lim
n
C=
k=1
1
T
A.BENHARI
21
Remark 1.13 Examples for fun
tions f (1; 2 ; : : :) in 3. are fun
tions whi
h do
not depend on a nite number of variables:
= I 1
1
; lim ; I P
lim
P
n n n n
n <1
n <1
n
n=k
=1
not:
1
I P
n=1
A.BENHARI
;
n <
22
If = =:::=0g
1
1.3
P(AjB ) =
P(A \ B )
P(B )
dis rete:
Ak = f! : = xk g
E f jB g =
P(d! jB ) =
xk P(Ak jB )
1 X
1
xk P(Ak \ B ) =
=
P(B ) k
P(B )
espe
ially for = IA (! )
(+)
X
k
P(d! )
1
E fIAjB g =
P(B )
IA dP =
P(A \ B )
= P(AjB )
P(B )
i.e.
onditional probability is a spe
ial
ase of the
onditional expe
tation
How
an the above denitions be extended to the
ase P(B ) = 0 ?
1
S
Let M = fBi 2 A :
Bi =
; Bi \ Bj = ;; i 6= j g be a
ountable
lass of
i=1
disjoint events
) E f jMg(!) := E f jBig for ! 2 Bi is a random variable
A.BENHARI
23
see:
8 B 2 fMg 9 a representation B =
Z
1 Z
X
E f jMgdP =
k=1Bj
k
|
1 zZ
X
dP =
k=1Bj
1
S
k=1
Bjk ; Bjk 2 M
E f jMgdP
{z
}
= (o)
}|
{
dP
sin e: if P(! : 1
2 > 0) > 0
ontradi
ition
analogously P(! : 1
f!: 1 2 >0g
(1
2 )dP > 0
2 < 0) > 0
E f jBgdP =
R
B
dP
8 B 2 B.
Remark 1.14
A)
2. Existen
e
and uniqueness follows dire
tly from the Radon-Nikodym theorem:
R
dP is a -nite,
ountable-additive set fun
tion on B whi
h is absolutely
B
ontinuous w.r.t. P
R
R
) 9 B-measurable fun
tion (!) su
h that dP = dP and
(! )
is unique (mod P)
24
) = E f jBg
PfAjBg dP = P(A \ B ) 8 B 2 B:
A = P (
);
ard(A) = 236
M = fB2; : : : ; B12 g
with Bk = f! = (i; j ) : i + j = kg
8B 2 B :
B = Bk :
2k7:
f jBgdP
E f jBk g P
(B )
| {z k}
BE
E f jBk g k361
a1 =a2
1
2
3
4
5
6
1
B2
B3
B4
B5
B6
B7
2
B3
B4
B5
B6
B7
B8
3
B4
B5
B6
B7
B8
B9
1 R dP = k
P(Bk )
2
R Bk
= B dP
R
= Bk dP
6
P
=
i P(Ai{z Bk})
i=1 |
1
or 0
36
4
5
6
B5 B6 B7
B6 B7 B8
B7 B8 B9
B8 B9 B10
B9 B10 B11
B10 B11 B12
on Bk
kP1
i 361
i=1
k(k 1) 1
2 36
Ai = f! : (! ) = ig = f(i; j ) : 1 j 6g
Ai \ Bk = f(i; j ) : 1 i; j 6; i + j = kg
) E f jBk g = k2
8 k 12 :
analogously
E f jBg is
B-measurable (it is
onstant on Bk ; k = 2; : : : ; 12)
is not B-measurable (it is not
onstant on Bk ; k = 2; : : : ; 12)
A = P (
),
M = fB1; : : : ; B12g
) E f jBg
A.BENHARI
!2Bk
R
1
P(Bk ) B
dP
P (M)
=
^ monthly average
25
6
%
%
e
(
h
(
h
h(
e
""
XX(((
%
! !%
!!B1
B2
E f jBg
ee
Z
B3
ZZPP
Jb bb
B12
(in memoriam to the 2000-ele tions between of G.W. Bush and A. Gore)
n
m
r
=
A=
A=
P(A) =
=
Bk
assume P(Bk ) =
k
k
k=1
per
entage of ele
tors voting in favour of 1 is
r
X
k=1
in general it holds
) if 1 has the majority of votes then 1 gets not ne
essarily the majority of ele
tors
Theorem 1.28 (properties of
onditional expe
tation)
26
1. 0
2.
3.
4.
5.
) E f jBg 0; = 0 ) E f jBg = 0;
is B-measurable ) E f jBg = ;
espe
ially: B 2 B ) PfB jBg = IB ;
E E f jBg = E ;
E j1 j; E j2 j < 1; a; b 2 R ) E fa1 + b2 jBg = aE f1 jBg + bE f2 jBg;
espe
ially: C1 \ C2 = ; ) PfC1 [ C2 jBg = PfC1 jBg + PfC2 jBg;
(n)n1 ; n 0 monotoni
ally in
reasing sequen
e of r.v.'s
) nlim
E fnjBg = E fnlim
jBg;
!1
!1 n
espe
ially:
(a): (Bn )n1
(b): (Cn)n1
n=1
sequen
e of events Ci \ Cj = ;; i 6= j
1
1
) P PfCnjBg = Pf S CnjBg;
n=1
6.
7.
8.
9.
) E f jBg = E ;
) PfAjBg = P(A);
is B - measurable ) E f jBg = E f jBg;
B1 B2 ) E fE f jB1gjB2g = E f jB1g;
B2 B1 ) E fE f jB1gjB2g = E f jB2g;
(! ) : f
; Ag ! fS ; Cg-measurable, (! ) : f
; Ag ! fZ ; Dg-measurable;
g (x; z ) fun
tion on S Z ; measurable w.r.t. fC Dg; E jg (; )j < 1;
)if is B-measurable (B A), then E fg(; )jBg = E fg(; z)jBgz= :
;
espe
ially: A;
B
B
n=1
independent
independent
) E f jg is
Bn
1 (C )
PfAj g = PfAjF g
27
Theorem 1.29 The
onditional expe
tation given the random element is a
C -measurable fun
tion, i.e. there exists a C -measurable fun
tion s(y) su
h that
E f j g = s( ):
fS ; C ; P g
Proof:
R
f
; A; P g f
; F ; P g
(! )
fR; Bg
s( )
E f j g
Borel sets
R
1 (C )
E f j gdP
1 (C )
(see rule of
hange of variables for integrals, basi
ourse, Theorem 3.25 )
R
R
=)
E f j gdP =
s( )dP
8 C2C
1 (C )
1 (C )
=) E f j g = s( )
s( (! )) is F - measurable sin
e s it is C -measurable
uniqueness
assertion
1 (C )
1 (C )
28
Proof:
1. if ; independent, i.e. ; f g = F independent
)
E f j g = E f jF g = E
(produ t spa e)
h(!; A) = p(!; A)
Theorem 1.31 If PfAjBg is a regular
onditional probability and (! ) a random variable then it holds R
E f jBg = (! )Pfd! jBg (mod P)
()
A
) () is valid for the indi
ator of random events IA;
) linearity of E and integral implies () is valid for dis
rete r.v.
let 0, Theorem 1.4 ) 9 monotoni
ally in
reasing sequen
e (n)n1 of
dis
rete r.v.'s with
ountable many values su
h that nlim
= uniformly 8 ! 2
!1 n
Theorem 1.28, 5, and Lebesgue's Theorem on monotone
onvergen
e imply
E f jBg = nlim
E fn jBg = nlim
!1
!1
A.BENHARI
29
P(d! jB);
assertion
Remark 1.15 It is possiblie to emphasize that PfAjBg is a fun
tion of elementary events ! by:
PfAjBg = PB (!; A)
or = P(!; A); if B is xed.
analogously:
PfAj g = P (!; A)
Denition 1.31 Let fS ; Cg be a measurable spa
e, : f
; A; Pg ! fS ; Cg
be a random element and B be a -algebra with B A.
The fun
tion Q(!; C ) dened on
C is
alled the regular
onditional distribution of the r.e. given a -algebra B, if
1. for a xed C 2 C Q(!; C ) is B - measurable
2. for a xed ! Q(!; C ) is a probability measure on C with probability 1
3. for ea
h C 2 C it holds Q(!; C ) = Pf 2 C gjBg (mod P)
R
B
S:
C:
Z
C
A.BENHARI
(x)dm =
(x)m(dx);
30
8 C 2 C:
f ( (! )dP =
f (x)(x)dm
) 9 (x) 0
with P (C ) = (x)dm
C
) m(C ) = 0 implies P (C ) = 0 ) P
w.r.t. m
is absolutely ontinuous
) assertion
S2
R
C1
S1
R R
C1 S2
assertion
S2
31
Proof:
Z
8 C1 2 C1
it holds:
Z
1 1 (C1 )
f (2 )dP
1 1 (C1 )
Z Z
(see 1 1 (C1 ) = f! : 1 (! ) 2 C1 g)
S1 S2
Z Z
S2
C1
(x ; x )
f (x2 ) 1 2 m2 (dx2 ) (x1 )m1 (dx1 )
(x1 )
{z
f(x1 )
Remark 1.18
C1
) E ff (2)j1g
= f(1 ) =
f(1 )dP
1 1 (C1 )
f (x2 )
S2
(1 ; x2 )
m (dx )
(1 ) 2 2
assertion
S2
f (x2 )
(x1 ; x2 )
m (dx ):
(x1 ) 2 2
1
Z2
x2
0
2Z 2x1
(x1 ; x2 )
dx
(x1 ) 2
x2
2 R2x1
0
x22
=
2(1 x1 ) 2
1
A.BENHARI
1dx2
dx2 =
2 2x1
0
32
2Z 2x1
x2
=1
x1
2(1
x1 )
dx2
6
E (2 j1 = x1 ) = 1
2
HH D
HH
1
HHH
HHH
HHH
H
x1
A.BENHARI
33
x1
(instead of x we use t)
X = f0; 1; 2; : : :g
or X = f: : : ; n; : : : ; 1; 0; 1; : : : ; n; : : :g
random fun
tion is
alled dis
rete time random pro
ess or time series
Theorem 2.1 The family of nite dimensional distributions of a random fun
tion (x) possesses the following
onditions of
ompatibility or
onsisten
e:
1. Px :::xn xn :::xn m (C n S m ) = Px :::xn (C n ) 8 C n 2 C n
2. let : permutation of (1; 2; : : : ; n),
(x1 ; : : : ; xn ) := (x(1) ; : : : ; x(n) );
:
orresponding permutation of sets in S n ; i.e.
for any any subset C n of points (z1 ; : : : ; zn ) 2 S n with C n 2 C n
it holds (C n ) = f(z(1) ; : : : z(n) ) 2 S n : (z1 ; : : : ; zn ) 2 C n )g
1
+1
8 C n 2 C n; 8 n 2 N
Proof: obvious
From the pra
ti
al point of view two random fun
tions are dierent only if the
orresponding f.d.d`s are dierent.
A.BENHARI
34
Denition 2.4 Two random fun
tions 1 (x) and 2 (x); x 2 X dened on the
same probability spa
e f
; A; Pg are
alled sto
hasti
ally equivalent, if
P(! : 1 (x; ! ) 6= 2 (x; ! )) = 0
8 x2X
and are said to be modi
ations or versions of ea
h other.
They are
alled indistinguishable if
P(! : 1 (x; ! ) = 2 (x; ! ); 8 x 2 X ) = 1:
Theorem 2.2 If 1 (x) and 2 (x) are sto
hasti
ally equivalent, then they are
sto
hasti
ally equivalent in the wide sense.
with
n
[
f! : 1(xi) 6= 2(xi )g
i=1
Example 2.1
= [0; 1; A : Borel sets of [0; 1; P : Lebesgue measure;
S = R;
C : Borel sets of R; X = [0; 1;
1 (x; ! ) = 0 8 (x; ! ) 2 X
= [0; 12
1 (x; ! ) 2 D = f(x; ! ) : x = ! g (D diagonal)
2 (x; ! ) =
0 (x; ! ) 2 [0; 12 n D
paths of 1 are
ontinuous while the paths of 2 are not
ontinuous
8 x 2 X : P(! : 1(x; !) 6= 2(x; !)) = P(! : ! = x) = 0
A.BENHARI
35
distributions
usually random fun
tions are only given in the wide sense, sin
e
1. in many pra
ti
al problems random fun
tions are
hara
terized by their
f.d.d.'s while the underlying probability spa
e is not given
2. it is simpler to dene the f.d.d's than the
orresponding probability spa
e
and the mapping (x; ! );
3. for solving many pra
ti
al problems it is su
ient to know only the f.d.d's
(1)
::: xn (C
n );
8C n 2 C n
then f
; A; Pg and (x; ! ) are
alled representation of the family of the
distributions (1).
Theorem 2.3
Kolmogorov
A.BENHARI
36
joint normal distribution, i.e. for its
hara
teristi
fun
tion it holds
1
J (x1 ; : : : ; xn ; u1 ; : : : ; un) = exp iu a
u Bu ;
2
where u = (u1 ; : : : ; un )
a = (a(x1 ); : : : ; a(xn )) with a(x) = E (x);
B = (b(xi ; xj ))1i;j n with b(x; y ) = E ( (x)
The fun tion a(x) is alled mean value fun tion and b(x; y ) orrelation fun tion.
Theorem 3.1 The
orrelation fun
tion b(x; y ) possesses the following properties
1. b(x; y ) = b(y; x);
2. b(x; y ) is a positive semi-denite fun
tion:
n
P
b(xi ; xj )ui uj 0 8 n 2 N; 8 ui 2 R; 8 xi 2 X ;
i;j =1
i.e. the matrix B = (b(xi ; xj ))1i;j n is symmetri
and positive semi-denite.
Proof:
1. obvious n
n
P
P
( (xi ) a(xi ))ui 2 =
E ( (xi ) a(xi ))( (xj ) a(xj ))ui uj
2. 0 E
i=1
n
P
i;j =1
i;j =1
b(xi ; xj )ui uj
assertion
Theorem 3.2 Let (x) be a Gaussian random fun
tion with a (stri
tly) positve
denite
orrelation fun
tion b, then for the distribution of ( (x1 ); : : : ; (xn)) it
holds for C n 2 Bn
P(( (x1 ); : : : ; (xn )) 2
C n)
f (y)dy;
Cn
a) B 1 (y
a))
(Q
A.BENHARI
Q ;
1
2
L
B 1 = (LL ) 1 = (L ) 1L 1 = L L 1
det B = det Q det det Q = det = 1 : : : n
p
p
p
p
det
det L = det
Q
=
:
:
:
=
det
=
det B
1
n
| {z }
()
1
2
=1
R
Z
exp
iu (y
Rn
(2 )n
p1
a)
1
dv;
det B
u = L v; du = det L dv = p
u Bu
1
u Bu du
2
det B
v L 1
B L
|{z}
LL
exp iv L 1 (y
{z
Rn
v
()
= v v
1 A
v v dv
2
a})
(+)
Set := L 1 (y
(+); (++) ) f (y ) =
=
(2 )n det B
(2 )n det B
exp
1
(y
2
exp
Rn
a) B 1 (y
A.BENHARI
1
(y
2
Z
a)
exp
Rn
a) B 1 (y
38
a)
1
((v + i
) (v + i
) +
) dv
2
n
Q
1
= p
exp
(2 )n det B
(++)
a)
k=1
1
R
e
1
1
w w dw
2
{z
2
wk
2
dwk = 2 n
assertion
Remark 3.1 Without the assumption of a stri
tly positive denite
orrelation
fun
tion b the matrix B
an be singular, i.e. the density f (y ) is not well dened.
This
ase
orresponds to a degenerated random ve
tor. Let rank B = m < n
then the distribution of ( (x1 ; : : : ; (xn )) is
on
entrated on a m-dimensional
submanifold of Rn . A well dened density exists only for proje
tions on this
submanifold.
In
ase of n = 1 the matrix B
orresponds to the varian
e of (x1 ). If B = 0 then
P( (x1 ) = a(x1 )) = 1.
X:
be an arbitrary set,
a(x) a real fun
tion on X ,
b(x; y ) : a positive semi-denite, symmetri
, real fun
tion on X 2 .
Then there exists a Gaussian random fun
tion (x) for whi
h a(x) is its mean
value and b(x; y ) its
orrelation fun
tion.
Proof:
onsider the family of distributions from Denition 3.1;
it satises the
onditions of
ompatibility;
Theorem of Kolmogoro (Theorem 2.3)
assertion
Denition 3.2 The random ve
tor fun
tion (x) = (1 (x); : : : ; m (x)) 2
Rm ; m 2 N; x 2 X is
alled Gaussian, if 8 n 2 N and x1 ; : : : ; xn 2 X the joint
J (x1 ; : : : ; xn ; u1 ; : : : ; un ) = exp i
n
X
uk a(xk )
k=1
Rm ;
n
1X
uj b(xj ; xk )uk
2 j;k=1
)
and
uk a(x) =
m
P
p=1
m
P
p;q=1
j;k=1
Proof: evaluate E
A.BENHARI
0; x: 2 X ; u: 2 Rm; 8 n 2 N:
p=1 k=1
39
2
Theorem 3.4
X:
Let
n
P
k=1
Theorem 3.5 For the moments of a s
alar Gaussian random fun
tion it holds
(
mj ;:::;jn (x1 ; : : : ; xn ) =
1
0
1
2s
s)
(
(
Bu;
u
)
u1 j1 ::: un jn 2s s!
u=0
order is odd
order is even (= 2s).
=)
Theorem 3.2
=)
Theorem 1.17
n
X
J (x; u) = exp
1
X
1
1
(Bu; u) =
2
p!
p=0
q J (x; u)
mj ::: jn (x) = ( i)q j
u1 : : : un jn u=0
1
jk = 2s +1 : mj
k=1
n
X
1
2
k;l=1
p
(q =
::: jn (x)
n
P
bkl uk ul ;
(Bu; u)p
n
X
k=1
jk )
2s
1 1 s
jk = 2s : mj ::: jn (x) = (
(Bu; u)s
j
j
n
u
:
:
:
u
s
!
2
1
n
k=1
2
s
1
s
=
(
Bu;
u
)
u1 j : : : un jn 2ss!
u=0
) assertion
i)2s
u=0
!
b(xj ; xk )uj uk = b(x1 ; x2 ) = E (x1 ) (x2 )
j;k=1
u=0
1 2
(b(x; x)u21 ) = b(x; x) = m11 (x; x) = E 2(x)
2 u1 2
u=0
1 2
m11 (x1 ; x2 ) =
2 u1 u2
m2 (x) =
A.BENHARI
2
X
40
2s = 4:
1
4
(Bu; u)2
m1111 (x1 ; x2 ; x3 ; x4 ) =
2
u1 : : : u4 2 2!
where (Bu; u)2 =
4
X
i;j =1
!2
bij ui uj
4
X
i;j;k;l=1
u=0
bij bkl ui uj uk ul )
PQ
j1
times
j2
jn
times
times
41
r2
r
f (r) = 2 exp
; u > 0; 2 (x) = D2 (x) = b11 (x; x)
2
(x)
2 (x)
(x) is uniformly distributed on ( ; ), i.e. its density is
1
g (v ) = ; v 2 ( ; ):
2
Proof: (x) = (x) + i (x) ) (x) = (x)
os (x); (x) = (x) sin (x);
(x) is Gaussian Denition
=) 3.3 ( (x); (x)) is a Gaussian ve
tor fun
tion
) density of ( (x); (x)) is
1
1
1 u2 v 2
p
+
(sin
e b12 (x; x) = 0)
f (u; v ) =
exp
2 b11 b22
2 b11 b22
1
u2 + v 2
=
exp
(sin
e b11 = b22 )
2b11
2b11
transformation of r.v.'s: density of ((x); (x))
(u; v )
f (r; s) = f (u(r; s); v (r; s))
(u = r
os s; v = r sin s)
(r; s)
1
r2
os s r sin s 1 r
r2
=
=
exp
exp
2b11
2b11 sin s r
os s 2 b11
2b11
|
{z
=r
) f (r; s)
A.BENHARI
Z
Z1
0
r
f (r; s)ds = 2 exp
(x)
f (r; s)dr =
= f (r)f (s)
42
r2
; r > 0; b11 (x; x) = 2 (x)
2
2 (x)
1
; s 2 ( ; )
2
assertion
Proof:
1. b11 (x; y ) = E (x) (y )
=)
Theorem 3.1 2.
assertion
2
n
P
(
x
)
z
k
k
k=1
=E
n
P
k;l=1
n
P
k;l=1
assertion
Theorem 3.8 Let b(x; y ); x; y 2 X be an arbitrary positive semi-denite fun
tion. Then there exists a
omplex Gauusian random fun
tion (x) su
h that
E (x) = 0
and
E (x) (y ) = b(x; y ).
Proof: plan of the proof:
(i):
onsider the real 2 2 matrix fun
tion
1 Re b(x; y )
b
11 (x; y ) b12 (x; y )
B (x; y ) = b (x; y ) b (x; y ) :=
2 Im b(x; y )
21
22
Im b(x; y )
Re b(x; y )
n
P
p;q=1
n
P
p;q=1
n
P
p;q=1
b(xp ; xq )zp zq =
n
P
p;q=1
b(xp ; xq )zp zq
b(xq ; xp )zp zq
) b(x; y) = b(y; x)
(
oe
ients of identi
al polynomials)
1
1
) b11 (x; y) = 2 Re b(x; y) = 2 Re b(y; x) = b11 (y; x);
) b12 (x; y) =
A.BENHARI
1
2 Im b(x; y )
1 Im b(y; x)
2
43
(b) 0
n
P
=2
=2
p;q=1
n
P
{z
}|
{z }
| {z } |
1
z
p
Re b(xp ;xq )
Im
b
(
x
;x
)
p
q
2
(p) (q)
b11 (xp ; xq )u1 u1 + b11 (xp ; xq ) u(2p) u(2q)
| {z }
=b22 (xp ;xq )
(p) (q)
b12 (xp ; xq ) u1 u2 + b12 (xp ; xq )u(1q) u(2p) +
|
{z
}
=b21 (xp ;xq )
p;q=1
n
P
p;q=1
n P
2
P
p;=1 k;l=1
(q)
iu
)
{z 2 }
zq
1
2
n
P
p;q=1
i0
(sin e it is real)
B (xp ; xq )zp ; zq
(iv):
=0
=0
assertion
Why do Gaussian random fun
tions play an important role in pra
ti
al problems ?
The answer gives a multivariate generalization of the
entral limit theorem (theorem on normal
orrelation).
Under very general
onditions the sum of a large number of small (in magnitude)
random fun
tions is approximately a Gaussian random fun
tion independently of
the probabilisti
nature of the summands.
Consider a double sequen
e of random fun
tions:
nk (x); x 2 X ; k = 1; 2; : : : ; mn ; n = 1; 2; : : : ; nlim
m =1
!1 n
Set n (x) :=
" (x)
nk
m
Pn
k=1
";") (x)
1 x 2 ( "; ")
0 else
Theorem 3.9 Let the random fun
tions n1 (x); : : : ; nmn (x) be mutually independent for ea
h n and satisfy the
onditions
1. for any " > 0 it holds:
lim
n!1
mn
P
k=1
k=1
A.BENHARI
k=1
44
Then for n ! 1 the nite dimensional distributions of the random fun
tion
n (x)
onverge weakly (in the sense of weak
onvergen
e of measures) to the
orresponding distributions of a Gaussian random fun
tion with mean value a(x) and
orrelation fun
tion b(x; y ).
Proof: not given
p1n
it
k=1
"0
holds nk
=0
nk )
1
1
b"nk (x; y ) = bnk (x; y ) = E nk (x)nk (y ) = E k (x)k (y ) = b(x; y )
n
n
The
orrelation fun
tion of k is
0
n
n
1X
1X
E n (x)n (y ) =
E (x)l (y ) =
b(x; y ) = b(x; y );
n k;l=1 k
n k=1
2.
8n 2 N
) limit distribution of (n)n1 is Gaussain with mean 0 and
orrelation fun
tion
b(x; y )
A.BENHARI
45
t2 < : : : < tn , the random ve
tors (a); (t1 ) (a); : : : ; (tn) (tn 1 ) are
mutually independent.
The ve
tor (a) is
alled initial state or initial value of the pro
ess and its
distribution the initial distribution.
A pro
ess with independent in
rements is dened in the wide sense if
P0 (B )
the initial distribution and
P (t; h; B ) the distribution of the ve
tor (t + h) (t)
for h > 0; t a; B 2 Bm (Borel sets of Rm )
are given.
to prove: the nite dimensional distributions are uniquely dened
(Denition 2.5 ) pro
ess is given in the wide sense)
(tn )
(tn 1 )
(t1 )
n
2
1
0
(t2 )
= (t0 )
a = t0
t1
t2
tn
:::
tn
dene:
0 1
0 1
Z Z
B~0 B~1
A.BENHARI
:::
t0 ; dy1) : : : P (tn 1 ; tn
P0 (dy0 )P (t0 ; t1
B~n
46
tn 1 ; dyn)
P0 (dy0)
B0
P (t0 ; t1
B1 y0
t0 ; dy1) : : :
:::
P (tn 1 ; tn
()
tn 1 ; dyn)
Bn (y0 +:::yn 1 )
Qt t t (B0 R B2 ) =
0 2
P0 (dy0)
0 1 2
B0
P (t0 ; t1
t0 ; dy1)
R y0
|{z}
=R
|
P(! : 1 + 2 2 C ) =
=
Z
C
Z
P (t0 ; t2
Z
P (t1 ; t2
t1 ; dy2 )
B2 (y0 +y1 )
{z
=H
t0 ; dz )
t0 ; dy1) P (t1 ; t2
P (t0 ; t1
t1 ; dy2)
R C y1
set C = B2
Z
B2 y0
A.BENHARI
P (t0 ; t2
t0 ; dz ) =
P (t0 ; t1
t0 ; dy1 )
B2 (y0 +y1 )
47
P (t1 ; t2
t1 ; dy2) = H
) Qt t t (B0 R B2 )
0 1 2
P0 (dy0 )
B0
P (t0 ; t2
t0 ; dz )
B2 y0
= Qt t (B0 B2 )
(+)
0 2
Theorem 2.3 of Kolmogoro implies the existen
e of a random pro
ess with
f.d.d.'s given in ()
stru
ture of these f.d.d.'s ) possesses independent in
rements
It is
onvenient to study pro
esses with independent in
rements using
hara
teristi
fun
tions
Denition 4.2
J (t; h; u) :=
Rm
is
alled
hara
teristi
fun
tion of the pro
ess with independent in
rements.
J (t; h; u) is the
hara
teristi
fun
tion of the in
rement (t + h) (t):
Theorem 4.1 A pro
ess with independent in
rements is given in the wide sense
if and only if
P0 (B )
the initial distribution and
J (t; h; u) the
hara
teristi
fun
tion of (t + h) (t)
are given where J (t; h; u) satises the
ondition
J (t; h1 + h2 ; u) = J (t; h1 ; u) J (t + h1 ; h2 ; u):
()
J (t1 ; : : : ; tn ; u1 ; : : : ; un ) =
n
Q
k=1
J (tk 1 ; tk
tk 1 ; uk )
()
Chara
teristi
fun
tion denes uniquely the distribution ) f.d.d.'s of are given.
ondition () :
onsider t0 = t; t1 = t + h1 ; t2 = t + h1 + h2 ;
1 = (t1 ) (t0 ); 2 = (t2 ) (t1 );
r.v. distribution
hara
terristi
fun
tion
1
P (t0 ; h1 ; B )
, J (t0 ; h1; u)
2
P (t1 ; h2 ; B )
, J (t1 ; h2; u)
1 + 2 P (t0 ; h1 + h2 ; B ) , J (t0 ; h1 + h2 ; u)
) J (t0 ; h1 + h2 ; u)
1 ;2
=
E exp(i(u; 1 + 2 ))
independent
=
E exp(i(u; 1 ))E exp(i(u; 2 ))
=
J (t0 ; h1 ; u)J (t1 ; h2 ; u)
)
A.BENHARI
48
ondition ()
Denition 4.3 A random pro
ess (t); t 2 T; is
alled sto
hasti
ally
ontinuous in t0 ; if for any " > 0 it holds
lim P( j (t; ! ) (t0 ; ! )j ") = 0:
t!t0
) A homogeneous pro
ess with independent in
rements is sto
hasti
ally
ontinuous, if
lim P (h; U") = 0
h!0
where
(t0 )j ")
Denition 4.5 A sto
hasti
ally
ontinuous and homogeneous pro
ess with independent in
rements is
alled Levy pro
ess.
Theorem 4.2
1. Let (t); t 2 T be a Levy pro
ess. Then for h ! 0
(a) the in
rement (t + h) (t)
onverges to zero in probability,
(b) the distribution of (t + h) (t)
onverges weakly to zero (the measure
on
entrated in zero).
2. (t) is sto
hasti
ally
ontinuous if and only if lim J (h; u) = 1 uniformly in
h!0
any bounded region juj K:
Proof:
1. (a) (t) sto
hasti
ally
ontinuous Denition
) 4.3
lim P( j( (t + h) (t)) 0j ") = 0
h!0
) (t + h) (t)
onverges in probability to 0
(b)
onvergen
e in probability implies weak
onvergene (see basi
ourse
Theorem 5.8.),
i.e.
for any
ontinuous and bounded fun
tion f it holds
R
f (x)P (h; dx) h!!0 f (0)
Rm
2. ()): J (h; u) =
R
Rm
1(b).
! 1;
h!0
Theorem 1.15
A.BENHARI
49
Theorem 4.3 The
hara
teristi
fun
tion of a Levy pro
ess possesses the following properties:
1. J (h1 + h2 ; u) = J (h1 ; u)J (h2 ; u); espe
ially J (nh; u) = [J (h; u)n 8 n 2 N
2. J (h; u) 6= 0
3. J (h; u) = exp(h g (u)) with a uniquely dened
ontinuous fun
tion g (u);
() property
ompletely
hara
terizes the dependen
e of J (h; u) on h)
Proof:
1. follows from Theorem 4.1 applied to homogeneous pro
esses
2. Theorem 4.2 ) lim J (h; u) = 1 8 juj K
h!0
) 9 h0 > 0 : jJ (h; u)j 21 8 h 2 [0; h0
let h be arbitrary ) h = h0 (n + ) where 0 < 1
1. ) J (h; u) = J (h0 n; u)J (h0 ; u) = [J (h0 ; u)nJ (h0 ; u)
) jJ (h; u)j ( 12 )n+1
) assertion
3. sin
e lim J (h; u) = 1 uniformly 8 juj K there exists a uniquely dened
h!0
fun
tion g1 (h; u) = ln J (h; u) 8 juj K ; h 2 [0; s where s = s(K )
g1 (h; u) is
ontinuous in [0; s fjuj K g
be
ause of 1. it holds :
g (u)
an be an arbitrary
ontinuous fun
tion provided that exp(hg (u)) is the
hara
teristi
fun
tion of a
ertain distribution for ea
h h
Theorem 4.4
Levy-Chin hin-representation
Let J (h; u); h > 0; u 2 Rm be a family of
hara
teristi
fun
tions su
h that the
limit
1 (J (h; u) 1) for juj K (K > 0 arbitrarily);
g (u) = hlim
!0 h
uniformly exists. Then there exists a nite measure on fRm ; Bm g with
(f0g) = 0, a non-negative denite matrix B 2 Rmm and a ve
tor a 2 Rm ;
su
h that
g (u) = i(a; u)
Z
1
(B u; u) +
2
ei(u;z)
Rm
A.BENHARI
50
i(u; z )
1 + jz j2
1 + jz j2
jzj2 (dz):
If is a Levy pro
ess then J (h; u) = exp(hg (u)), hen
e g (u) = lim h1 (J (h; u) 1),
h!0
onvergen
e is uniform in every bounded domain juj K; 0 < K < 1
see h1 (J (h; u) 1) = g (u) + 2!h g 2 (u) exp(hg (u)); 2 [0; 1, (Taylor expansion)
g (u) is
ontinuous for juj K
) jg(u)j C = C (K )
1
) h (J (h; u) 1) g(u) h2 C 2 ehC independent of u
) uniform
onvergen
e
Theorem 4.5 If (t); t > 0 is a Levy pro
ess with values in Rm , then the
hara
teristi
fun
tion J (h; u) of the in
rement (t + h) (t) is of the form
J (h; u) = exp(h g (u));
where g (u) is given in Theorem 4.4.
6
(t0 )
t0
1 h(B u; u))
2
Denition 4.6 A (s
alar) sto
hasti
pro
ess W (t); t 2 [0; 1) is
alled Wiener
pro
ess, if
1. P(W (0) = 0) = 1,
2. for any 0 t0 < t1 < : : : < tn the in
rements
W (t1 ) W (t0 ); : : : ; W (tn) W (tn 1 ) are mutually independent,
A.BENHARI
51
E W (t1 )W (t2 )
= E W (t1 ) W (0) W (t2 ) W (t1 ) + W (t1 ) W (0) (W (0) = 0 a.s.)
= E W (t1 ) W (0) W (t2 ) W (t1 ) + E W (t1 ) W (0) 2
= E W (t1 ) W (0) E W (t2 ) W (t1 ) + E W (t1 ) W (0) 2
|
{z
0
}|
{z
0
{z
2 t1
(independen
e of in
rements)
= 2 t1
f (z )(dz ) = qf (z0 )
Rm
Theorem 4.4 )
(1 + jz0 j2 ) i(u;z )
e
J (h; u) = exp hq
2
j
z
0j
| {z }
0
i(u; z0 )
1 + jz0 j2
=:
=
it
an be shown: (t)
q
i
(
u;z
0)
exp(
h(e{z
1))} exp h 2 i(u; z0 )
|
z0
{z
}
Poisson distribution |
degenerated distribution
j j
(0) = z0 N (t)
t=h
qt
jz0 j2 ;
where N (t) is a so-
alled Poisson pro
ess with the parameter = q (1+jzjzj j )
0
0
2
Denition 4.7 A (s
alar) sto
hasti
pro
ess N (t); t 2 [0; 1) with homogeneous
and independent in
rements is
alled Poisson pro
ess with the parameter > 0,
if
1. P(N (0) = 0) = 1,
2. for any 0 t0 < t1 < : : : < tn the in
rements
N (t1 ) N (t0 ); : : : ; N (tn ) N (tn 1 ) are mutually independent,
3. for all t and h > 0 the in
rement N (t + h)
with the parameter h,
A.BENHARI
52
P( = k) = k e
k!
it holds:
for
k = 0; 1; 2; : : : :
E = D2 =
J (u) = exp (eiu 1)
1
(h)k e h
k!
1
P N (t) = k = P N (t) N (0) = k = (t)k e
k!
) E N (t) = t
D2 N (t) = E N (t) E N (t) 2 = t
) P N (t + h) N (t) = k
t
independen
e of in
rements )
= E N (t1 ) N (0) E N (t2 ) N (t1 ) + E N (t1 )2
|
{z
t1
= t1 (t2
(4) B = 0; su h that
Rm
}|
{z
(t2 t1 )
|
{z
}
(EN (t1 ))2 +D2 N (t1 )
1
jzj2 (dz ) < 1 (property of with respe
t to zero)
m
X
k=1
1
with 0 (B ) :=
Z
B
uk ak
|
Rm
R
zk
jzj2 (dz) +
{z
a^k
1 + jz j2
jzj2 (dz)
and
(ei(u;z)
53
1)0 (dz )
Rm
:=
Z
Rm
Chemnitz/Fakult
at
A.BENHARI
1 + jz j2
jzj2 (dz)
) J (h; u)
= exp ih(^a; u)
k=0
h
1
(h)k
k!
0 (dz )
Rm
Z
ei(u;z)
h
k
ei(u;z)
0 (dz )
(+)
Rm
) J (h; u) is the
hara
teristi
fun
tion of the sum ha^ + 1 + + N (h) where
1. 1 ; 2 ; : : : 2 Rm are independent random ve
tors with the distribution 0 ;
2. a^ is a
onstant ve
tor and
3. N (h) is a random variable whi
h is independent of fk ; k = 0; 1; : : :g
possessing a Poisson-distribution with the parameter h (N (:) is a Poisson
pro
ess).
Let (0) = 0 a.s. then
N
(t)
X
k=1
k
1
X
k=0
1
X
k=0
1
X
k=0
hq
}
EIk =P (N (h)=k)
ei(u;ha^)
1
X
k=0
) (+)
ei(u;ha^)
k
Y
E ei(u;p )
p=1
{z
}
|
1 ;:::;k are i.i.d
Z
k
hq 1 (hq )k
i
(
u;z
)
e 0 (dz )
k!
Rm
(5) a; as in (4), B 6= 0
Theorem 4.4 )
A.BENHARI
: : : (dz )
{z
ha^+1 +:::N (h)
.f. of
54
|exp(h({zBu; u))}
(2)
.f. of
N
(t)
X
k=1
k
A.BENHARI
55
present
past
(t) depends only on the present value (tn ) but is independent of the past
values (ti ) with i < n;
Theorem 1.29: ) E f j g = d( ); where
E f j g is -measurable
i.e. f
; F g measurable
! fR; Bg
measurable
is a random element in fS ; Cg i.e. f
; Ag
! fS ; Cg
measurable
d
is a C -measurable fun
tion
i.e. fS ; Cg
! fR; Bg
56
The fun
tions P (s; y; t; C ) are
alled transition probabilities of the Markov
pro
ess.
ee ts
ondition 2a (Chapman-Kolmogorov equation) is a ondition of ompatibility or onsisten e to P (s; y; t; C ) for the times t1 < t2 < t3
ondition 2b means that Pf (t) 2 C j (s)g is a regular
onditional distribution (see Denition 1.30), i.e.
{ P (s; (s); t; C ) is for xed ! a probability on C ( (s) = (s; ! ) !)
{ P (s; (s); t; C ) = Pf (t) 2 C j (s)g a.s. for any C 2 C
1 for y 2 C
0 else
For a Markov pro ess also holds the generalized Markov property:
()
Theorem 5.1 Let (t); t 2 T; be a Markov pro
ess. Moreover, let Ft = f (s),
s t; s 2 T g and Ft = f (s); s > t; s 2 T g: Then it holds
PfC j Ftg = PfC j (t)g (mod P)
8 C 2 Ft:
Remark 5.1 The theorem shows that the dependen
e of an event in the future
C 2 Ft on the past and present Ft is
ompletely determined by the dependen
e
on the present state (t).
Denition 5.2 Let T = [0; b or T = [0; 1) and fS ; Cg be a measurable spa
e.
A family
onsiting of
1. an initial distribution P0 on fS ; Cg and
2. transition probabilities P (s; y; t; C ) with t > s; t; s 2 T; C 2 C ; satisfying
the
onditions of Denition 5.1 2.
is
alled Markov pro
ess in the wide sense with values in fS ; Cg
TU Chemnitz/Fakult
at f
ur Mathematik
A.BENHARI
57
:::
S C1
Cn
P0 (dy0)
:::
P( (t1 ) 2 C1 ) =
see
C1
P (t1 ; y1 ; t2 ; dy2 )
C2
P
(0; y0{z; t1 ; C1}) P0 (dy0)
|
S Pf(t )2C j(0)=y g
(see formula
of total probability)
Z
Z
1
Z
C1
Cn
P0 (dy0)
C1
| R
C1
{z
}
P f (t2 )2C2 j (t1 )=y1 gP (0;y0 ;t1 ;dy1 )
Theorem 5.3 Let S be a omplete, separable and metri spa e, then for every
Markov pro
ess in the wide sense there exists a representation (in the sense of
Denition 2.6).
Proof:
onstru
tion of f.d.d.'s (see Theorem 5.2);
they satisfy the
onditions of
ompatibility
Theorem of Kolmogorov
) assertion
P (y; h1 + h2 ; C ) =
It
an be seen that
A.BENHARI
58
(3)
it is enough to know P (y; h; C ) for all h ", where " is arbitrarily small,
be
ause for h~ > " P (y; ~h; C )
an be determined by (3).
1 ))
s; C )
S2
) E ff (2)j1 = x1 g
the Wiener pro
ess is a Gaussian pro
ess, Theorem 3.2 ) density of ( (s); (t))
is given by
1
1 1
p
%(x1 ; x2 ) =
exp
xB x
with x = (x1 ; x2 )
2
2 det B
E 2 (s) E (s) (t)
s
min(s; t)
s s
and B =
=
=
E (t) (s) E 2(t)
min(t; s)
t
s t
det B = st
s2
= s(t
s);
s(t s)
t
s
s s
A.BENHARI
59
it follows %(x1 ; x2 ) =
p1
2 s(t
x2
2s
exp
s)
x21
2s
(x1 x2 )2
2(t s)
it follows
%(x ; x )
1
exp
%(x2 j x1 ) = 1 2 = p
% (x1 )
2 (t s)
1
assertion
A.BENHARI
60
(x1 x2 )2
2(t s)
Denition 6.1 Let " > 0 be arbitrarily, I = f1; 2; : : : ; ng be an index set. A set
of points Q = fxi : xi 2 Rm ; i 2 I g = fx1 ; : : : ; xn g is
alled "-adjoining, if it
an not be separated into two subsets with a distan
e larger than ", i.e. for any
two subsets I1 ; I2Sof I with I1 \ I2 = ; and I1 [ I2 = I it holds J (I1 ) \ J (I2 ) 6= ;,
where J (Ip ) := fj 2 I :j xi xj j "g; p = 1; 2.
i2Ip
x4
x1 x3
x2
x6
x5
"
Rm
f5;6g
61
Denition 6.2 A random fun
tion " (x; ! ) on D Rm and values in R with
Ef" (x)g = 0 is
alled weakly
orrelated with the
orrelation length " if it
holds
Y
i2I
"(xi ) =
p
Y
j =1
8
<Y
E:
i2Ij
9
=
"(xi )
p
[
j =1
Ij = I
K" (x1 ; x2 ) =
8
<
:
K"(x1 ; x2 )
for jx1 x2 j "
E
"(x1 ) E "(x2 ) = 0
else
|
{z } | {z }
=0
=0
x
1 x2
2
1
"
while the denition of weakly orrelated fun tion is based on a property of moments of any order.
Theorem 6.1 Let "(x; ! ) be a weakly orrelated with the orrelation length "
and
1 E
X
k=0
j "(x) jk < 1 8 x 2 D:
k!
62
Remark 6.2 If a random fun
tion is weakly
orrelated with
orrelation length
" then the assumption of the above theorem implies its "-dependen
e.
But a (
entered)"-dependent
fun
tion needs not to be weakly
orrelated be
ause
Q
" (xi ) may not exist.
the moments E
i2I
Theorem 6.2 For any " > 0 there exists Gaussian random fun
tion whi
h is
weakly
orrelated with the
orrelation length ".
A
entered Gaussian random fun
tion (x) with
E (x1 ) (x2 ) = 0 for jx1 x2 j > "
is weakly
orrelated with a
orrelation length "0 ".
The probabilisti
analysis of solutions to equations
ontaining weakly
orrelated
fun
tions often leads to integral fun
tionals of the type
" (x) =
Example 6.2 Initial value problem for an ODE with a random inhomogeneous
term, D = [0; 1);
> 0
f 0 (x) =
solution
f (x; ! ) =
=
Zx
Z0
e|
(x y) (y; ! )dy
{z } "
F (x;y)
G = Gx = [0; x
1
with the intensity of ": a(x) := lim m
"#0 "
E " (x)"(x + y ) dy
fjyj"g
2. For " # 0 the f.d.d.'s of p1"m " (x)
onverge weakly to the
orresponding
f.d.d.'s of a
entered Gaussian random fun
tion (x) with
E (x1 ) (x2 ) =
results
an be extended to
A.BENHARI
63
A.BENHARI
64
L2 = L2(
; A; P)
:= f r.v. on f
; A; Pg : E j j2 < 1g
(; ) := E
is alled
G(xp ; xq )zp zq 0
8 xp 2 X ; 8 zp 2 ; p = 1; : : : ; n; 8 n 2 N:
C
Theorem 7.1 The
ovarian
e- and
orrelation fun
tions are positive semi-denite kernels on X 2 :
Proof:
0 Ej
0 Ej
n
P
p=1
n
P
p=1
(xp)zp j2
=
(xp) E (xp) zp j2 =
n
P
p;q=1
n
P
p;q=1
B (xp ; xq )zpzq
R(xp ; xq )zp zq
Theorem 7.2 Positive semi-denite kernels G(x1 ; x2 ) possess the following properties:
1. G(x; x) 0
2. G(x1 ; x2 ) = G(x2 ; x1 )
3. jG(x1 ; x2 )j2 G(x1 ; x1 )G(x2 ; x2 )
4. jG(x1 ; x3 ) G(x2 ; x3 )j2 G(x3 ; x3 ) G(x1 ; x1 ) + G(x2 ; x2 ) 2Re G(x1 ; x2 )
Proof:
G(x1 ; x1 ) jz1 j2 0
65
assertion 1.
{z
) real
{z
real
) Ai + B ( i) = 0 ) A = B
) A(z1 z2 + z2 z1 ) = 0 ) A = B = 0 )
assertion 2.
det
G(x1 ; x1 ) G(x1 ; x2 )
G(x2 ; x1 ) G(x2 ; x2 )
) G(x1 ; x1)G(x2 ; x2)
0
G(x1 ; x2)G(x2 ; x1)
2.
=
G(x1 ; x2 )G(x1 ; x2 )
= jG(x1 ; x2 )j2 ) assertion 3.
{z
a11
a22
+ (G13
|
with y1 := z; y2 := z3 it follows
)
hen
e det
assertion 4.
2
P
ij =1
{z
a21
a12
aij yi yj 0
0
0
Denition 7.3
1. Let 1 (x); 2 (x) be Hilbert random fun
tions with E p(x) = ap (x); p = 1; 2:
Then
A.BENHARI
66
where
Remark 7.2 If in 1. it holds 1 = 2 = then R (x1 ; x2 ) sometimes is
alled auto
orrelation fun
tion. The matrix
orrelation fun
tion R(x; y ) in 2.
ontains
the auto
orrelation fun
tions of k as diagonal entries and the
ross
orrelation
fun
tions of k and l as non diagonal entries, k; l = 1; : : : ; m; k 6= l.
Denition 7.4 A matrix fun
tion G(x; y ) = (Gpq (x; y ))1p;qm is
alled a positive semi-denite matrix kernel on X 2 , if
n
X
p;q=1
0; 8 xp 2 X ; 8zp 2
m
P
j;k=1
m; p
= 1; : : : ; n; 8n 2 N
Theorem 7.3 The matrix orrelation fun tion is a positive semi-denite matrix
kernel.
Proof:
n
X
E zp ( (xp)
p=1
X
n
= E
=
n
X
p;q=1
p;q=1
zp ( (xp)
2
a(xp ))
a(xq )) zq
Theorem 7.4 There hold the following properties of positive semi-denite matrix
kernels G(x; y ):
1. the matrix G(x; x) is positive semi-denite 8 x 2 X , i.e.
m
P
Gkl (x; x)zk zl 0 8 z = (z1 ; : : : ; zm ) 2 C m ;
z G(x; x)z =
k;l=1
67
Proof:
1. Denition 7.4 with n = 1 :
z1 G(x1 ; x1 )z1 0
) assertion 1.
2. Denition 7.4 with n = 2 :
) z1 G(x1 ; x1)z1 + z2 G(x2 ; x2)z2 + z2G(x2 ; x1 )z1 + z1 G(x1 ; x2)z2 0
|
{z
0
{z
0
{z
real
z2 G(x2 ; x1 )z1 + z1 G(x1 ; x2 )z2 = z1 G(x1 ; x2 )z2 + z2 G(x2 ; x1 )z1
Proof:
ne
essity: follows from Theorem 7.1 and Theorem 7.3
su
ien
y: let R(x; y ) be positve semi-denite, Theorems 3.3, 3.4, 3.8 )
9 Gaussian random fun
tion with R as
orrelation fun
tion
assume now that X is a metri
spa
e with the metri
68
nition 4.3).
Remark 7.4 The m.-s. ontinuity on X (i.e. for all points in X ) does not imply
) P (t + h) (t) = k
E j (t + h) (t)j2
1
(hq )k e hq
k!
1
X
1
=
k2 (hq )k e
k!
k=0
=
hq
= hq + (hq )2
h!0
2 R1 with + tk 2 X
X = ( 1; +1) then
t2 ).
69
f (xp
xq )zp zq 0;
8 n 2 N; xp 2 X ; zp 2 ; p = 1; : : : ; n:
C
Example 7.1 The orrelation fun tion of a wide-sense stationary random pro ess is positve semi-denite.
Proof:
G(x1 ; x2 ) = f (x1
x2 )
E k (t) = ak =
onst
E (k (t1 ) ak )(l (t2 ) al ) = Rkl (t1 ; t2 ) = Rkl (t1 t2 )
i.e. the
omponents of are also wide-sense stationary (k = l)
Denition 7.9 If 1 (t) and 2 (t) are wide-sense stationary and the ve
tor pro
ess
(1 (t); 2 (t)) is also wide-sense stationary then the pro
esses 1 and 2 are
alled
wide-sense stationarily related or stationarily
orrelated.
Example 7.2 Let (t) be
entered wide-sense stationary, then 1 (t) = (t) and
2 (t) = (2t) are wide-sense stationary but not stationarily related sin
e
R12 (t1 ; t2 ) = E 1 (t1 )2 (t2 ) = E (t1 ) (2t2 ) = R (t1
2t2 ):
zp G(xp
xq )zq
0;
8 n 2 N; 8xp 2 X ; 8 zp 2
m;
p = 1; : : : ; n:
Example 7.3 Matrix orrelation fun tions of wide-sense stationary ve tor pro esses are positive semi-denite.
Theorem 7.8 Let G be a positive semi-denite matrix fun
tion, then it holds
1. G(0) is a positive semi-denite matrix,
2. Gkl (x) = Gkl ( x); i.e. G(x) = G ( x)
A.BENHARI
70
3. jGkl (x)j2
x2 )
Examples of wide-sense stationary random pro
esses
m
X
Bkl
|
p;q=1
8 m
< PB
kk;pp
p=1
:
see
k (n)k
p;q=1
k=0
Apq Apq =
m
X
p;q=1
()
for k 6= l
jApq j2
(Frobenius norm)
kAk k2 < 1
1 1
P
P
k
l
k=0
l=0
1
() P
=
Ak 2 <
k=0
= E (n) (n) = E
=
1
P
k;l=0
k k
E k l
(n)
is wide-sense stationary
1
P
see E (n + p) (n) = E
Ak (n + p
=
=
A.BENHARI
m
X
1
P
series
onverges in Lm
2 if
{z
pqkl
1
P
k=0
k;l=0
1
P
l=0
1
P
k)
(n
l=0
Ak E
(n + p {zk) (n l}) Al
|
n+p
k;n l
I =p+l;k I
l) A
l
X
k
k eiuk t =
X
k
) R(t1; t2 )
= E (t1 ) (t2 ) = E
=
X
k
X
j;k
2k exp(iuk (t1
j k exp(iuj t1
t2 )) =: R(t1
iuk t2 )
t2 )
average power
arried by the harmoni
s of (t) with frequen
ies smaller than u
F (u)
ompletely
hara
terizes the average powers E j
k j2 =
2k
P
sin
e
2k = F (uk + 0) F (uk );
2k = F (u2 ) F (u1 )
u1 u<u2
orrelation fun tion of an be written in terms of the spe tral fun tion:
R(t) =
A.BENHARI
X
k
Z1
2k eiuk t =
eitu F (du)
{z
}
|
Lebesgue-Stieltjes integral
72
7.2
x2 ).
The
orrelation fun
tion of a homogeneous random fun
tion depends only on the
ve
tor x1 x2 and it holds:
n
P
R(xj xk )zj zk 0 8 n; 8 xj 2 R m ; 8 zj 2 C; j = 1; : : : ; n:
j;k=1
Theorem 7.9 Let be a homogeneous random fun
tion on Rm with the
orrelation fun
tion R(x). If R(x) is
ontinuous at x = 0 then (x) is mean-square
ontinuous for all x 2 Rm :
Proof: E j (x + h)
= 2(R(0) Re R(h))
assertion
Theorem 7.10 The fun
tion R(x); x 2 Rm , is the
orrelation fun
tion of a
mean-square
ontinuous homogeneous random fun
tion f (x); x 2 R m g if and
only if R(x) admits a representation
R(x) =
()
ei(x;u) F (du)
Rm
Bm of Rm .
Proof:
su
ien
y ((): let R(x) be given by (), then
R(xj
xk )zj zk =
=
Z X
n
Rm
Z
Rm
h!0
h!0 Rm
ei(xj
xk ;u) z z F (du)
j k
j;k=1
2
n
X
i
(
x
;u
)
k
e
zk F (du)
k=1
ei(h;u) F (du) =
R
Rm
1 F (du) = R(0)
73
Theorem 3.8 )
9
omplex-valued Gaussian random fun
tion (x)
possessing R(x) as
orreletion fun
tion
is m.-s.-
ontinuous sin
e R(x) is
ontinuous at x = 0 (see Theorem 7.9)
ne
essity ()): follows from the next Theorem 7.11
if (x) is a m.-s.-
ontinuous homogeneous random fun
tion then its
orrelation fun
tion R(x) is positive semi-denite and
ontinuous at x = 0
) R(x) is
ontinuous 8 x(exer
ise) Theorem 7.11 ) representation ()
(x) :=
F0 ei [(;x)+ :
Then it follows
E (x) =
F0
E ei(;x) E ei
= 0 sin e
E ei
y)
1
=
2
Z
eiv dv = 0
Rm
assertion
Theorem 7.11 Every
ontinuous and positive semi-denite fun
tion R on Rm
admits a representation () (with a uniquely determined F ).
Proof: idea:
onstru
t the measure F using the Fourier transform of
R(x)
holds:
!1
jxj2 ;
2N
{z
!1 8x
Z1
Z1
exp
A.BENHARI
may be not absolutely integrable, but for its " orre tion" RN (x) it
74
x2
dx < 1
2N
R
Rm
Rm
f (x)e i(x;z) dx
f~(z )ei(x;z) dz
R~ N (z ) =
p1 m
( 2 )
R(x) exp
Rm
(1),(2)
) FN (A)
(Rm ; Bm )
R
:= (p21)m R~ N (z )dz; A
2 Bm
is a nite measure on
FN (A) := FFNN(R(Am)) is a probability measure with the
hara
teristi
fun
tion:
Z
JN (u) =
Rm
1
1
p
m
FN (R ) ( 2 )m
| {z }
1
R(0)
ei(u;z)
m
ZR
1
1 ~
p
R (z )dz
m
FN (R ) ( 2 )m N
ei(u;z) R~ N (z )dz
Rm
{z
2
RN (u)=R(u) exp j2uNj
FN
(Rm )
1
= p m
( 2)
Rm
1 R(u) e
) JN (u) = R(0)
juj2
) 8 u 2 Rm it holds:
u)
JN (u) = RR((0)
e
J (u) is ontinuous at u = 0;
2N
juj2
2N
N !1
R(u)
R(0)
=: J (u);
Theorem 1.15 ) J (u) is the
hara
teristi
fun
tion of a probability measure
F , moreover FN
onverges weakly to F
) 9 nite measure F = R(0)F
R R
Rm Rm
Rm , sin
e
75
Rm R m
R R
Rm Rm
p
R( 2 u) exp( juj N+jvj
2
i( 2u; z ))dudv
y = 2 u; x + y = 2 v; x = p12 (u + v ); y = p12 (v
;:::;xm ;y ;:::;ym )
jxj2 + jyj2 = juj2 + jvj2; j ((ux ;:::;u
j=1
m ;v ;:::;vm )
see x
R
Rm
R( 2 u) exp( juNj
vgl.
= N
= N
R
Rm
m R
R1
exp( jvNj )dv = ( exp(
Rm
1
j
uj2
R( 2 u) exp( N
1
2 p m
2
m
Rm
s2 )ds)m
N
p21 m
i( 2 u; z ))du
= N
i( 2 u; z ))du
R(w) exp(
Rm
u);
{z
R~ N (z )
(w = 2u)
p mp
= N m R~ N (z )
(2) to prove : R~N (z ) is integrable
~ g~)L = (f; g )L
Parseval equation for Fourier integrals: (f;
f (x) = RN (x);
f~(z ) = R~N (z );
p
g (x) = " m exp( jx2"j ); g~(z ) = exp( 12 "jz j2 )
R
) j(f;~ g~)j = j R~N (z) exp( 12 "jzj2 )dzj
2
Rm
= j(f; g )j = j
Rm
R
R(0) p"1m
Rm
integrability
see m = 1 :
R~N (z ) =
1
(R
~
h N (z + h)
R1
H (x)e
ixz dx
ixz
R~ N (z ))
H (x)( ix)e dx
1
1
R
ihx 1
e
ixz
= H (x)e ( h + ix)dx
R1
A.BENHARI
76
1
R
=
h2
1
ihx
h (e
see
ixz 1 hx2 e #ixh dx
2
H (x)e
1
R1
1 hx2 e #ixh
2
1) = ix
jH (x)j x2 1 dx
R1
h
p
= 2 2 jR(x)je
1
R1 2
xe
dx h!!0 0
1
) assertion
Denition 7.12 The measure F in representation () of Theorem 7.10 is
alled
spe
tral measure (to the
ontinuous and positive semi-denite fun
tion R(x)).
The
orresponding distribution fun
tion
F (u) = F (Iu)
x2
2N
x2 dx
ph
2 2 jR(0)j
x2
2N
Iu = fx 2 Rm : x < ug
where
is
alled spe
tral fun
tion. IfR F (A) is absolutely
ontinuous with respe
t to
Lebesgue measure, i.e. F (A) = f (u)du, then f (u) is
alled spe
tral density.
A
Theorem 7.12 If R(x) is absolutely integrable (i.e. jR(x)jdx < 1) then the
Rm
spe
tral density f (u) exists and it holds
Z
R(x) =
ei(x;u) f (u)du:
Rm
1
g (u) = p m
2
Z
Rm
m
1 Y
= p m
2 k=1
p1 m
1
= p m
2
m
Y
yZ
k +hk
ei(u;z) IK (z )dz
Rm
eiuk zk dzk
yk hk
1
exp(i(yk + hk )uk ) exp( i(yk
2 k=1 iuk
R~ N (z )dz =
Z
Rm
Z
Rm |
{z
2N
Rm
77
g (u)
du
| {z }
} j:jp2m 2m h :::h
p 1 m V (K ) jR(u)jdu
2
hk )uk )
RN (u)g (u)dz
Rm
u2
jj
j:jjR(u)j1
Z
A.BENHARI
V (K ) =
m
Q
ei(yk
hk )uk )
2hk
k=1
FN (K )
R
dene
:= p21 m R~ N (z )dz as in the proof of Theorem 7.11
K
) FN is absolutely
ontinuous
w.r.t. Lebesgue measure 8 N
N ! 1 ) F is absolutely
ontinuous w.r.t. Lebesgue measure
R
) 9 f su
h that F (A) = f (z)dz
A
) R(x) =
Rm
ei(x;z) F (dz ) =
Rm
ei(x;z) f (z )dz
assertion
Corollary 7.1 A fun
tion R(t); t 2 ( 1; 1); is the
orrelation fun
tion of a
m.-s.-
ontinuous wide-sense stationary pro
ess if and only if it admits a repreR1
sentation R(t) = eitu F (du); where F (:) is a nite measure on B1 .
1
Denition 7.13 A random fun
tion is
alled isotropi
if its
orrelation fun
tion
satises the
ondition
v
u m
uX
x2 = t (x1;p
p=1
x2;p )2 ;
i.e. it depends only on x1 and the distan
e between x1 and x2 (but not on the
dire
tion of x1 x2 ).
R(x1 ; x2 ) = R(x1
x2 ) = R((x1 ; x2 )) = R(jx1
x2 j):
Theorem 7.13 The fun
tion R(); 0 < 1; is the
orrelation fun
tion of
a homogeneneous, isotropi
, m.-s.-
ontinuous random fun
tion if and only if it
admits a representation
R() = 2
m Z1
I m (r)
2
1
G(dr);
m
(r)
2
where G is a nite measure on [0; 1) and Ip (x) is the Bessel fun
tions of rst
kind, i.e.
1
x p+2k
X
1
k
Ip(x) = ( 1)
k! (p + k + 1) 2
k=0
with the Gamma fun
tion
A.BENHARI
R1
(x) = e t tx 1 dt:
0
78
Remark 7.7 For the Gamma fun
tion and the Bessel fun
tion their hold the
following relations
(1) = 1; ( 12 ) = p;
I (x) =
1
2
2
x
os x; I (x) =
1
2
R1
(n + 1) = n!
2
x sin x
(x + 1) = x (x)
1
(r)
()
p
x Ip
R1
G(dr) =
1
2
os(r) G(dr)
R1 ir
e F (dr)
m = 2 : R() =
m = 3 : R() =
R1
0
I0 (r) G(dr)
R1
) R()
see ( 23 ) =
1
2
p
( 21 ) =
R1 sin(r)
r G(dr )
0
q
= x2 sin x;
; I (x)
1
2
O = s(dx) =
S
2
( m2 )
m 1
m
2
R(x)s(dx) = R() O
Z Z
S
Fubini
S Rm
Z Z
Rm
(+)
S
it holds
Z
S
d
f (x)s(dx) =
d
Z
B
f (x)dx and
Z
B
ei(x;u) dx
2
=
juj
m
2
I m (juj)
2
79
ei(x;u) s(dx)
S
d
=
d
m 2
=
2 juj
(+)
2
=
juj
Z
1
R() =
O m
R
it holds
Z
2
juj
m 1
2
m
2
I m (juj)
2
2 m
2
I (juj) +
juj
juj
m
2
m
I 0m (juj) juj
| 2 {z }
()
= I m2 1 2mjuj I m2
jujI m 1(juj)
2
juj
Rm
= nlim
!1
m
n
X
k=1
n
X
k=1
f (rk ) G([rk 1; rk )) =
Z1
f (r)G(dr)
dene
) R() = 2
m Z1
I m (r)
2
1
m
(r)
G(dr) ) assertion
R(t; ) =
Z1 Z1
10
m
2
m (r )G(dv dr ) where
m (y ) =
y
eitv
m
R(t; x)
=
proof Theorem 7.13
A.BENHARI
Z1
1
Z1
1
eitv
eitv
Rm
Z1
I m (y ):
2
1; +1)g:
80
8. Brownian motion
8.1
A sto
hasti
pro
ess is a phenomenon whi
h evolves in time in a random manner. There are
many examples of phenomena whi
h
an be thought of as a fun
tion both of time and of a
random (or un
ertainty) fa
tor, think of the pri
e of shares, the size of some populations, or the
number of parti
les registered by a Geiger
ounter.
In the last
hapter we have
alled any sequen
e of random variables dened on the same probability spa
e a (dis
rete time) sto
hasti
pro
ess. We know already mathemati
al examples,
of su
h dis
rete pro
esses: Sequen
es of independent, identi
ally distributed random variables
Y1 ; Y2 ; : : :, their partial sums
Xn = Y1 + + Yn ;
whi
h are the random walks based on a sequen
e Y1; Y2 ; : : : and, dened in the last
hapter, the
Galton-Watson pro
ess with ospring distribution given by (p0 ; p1 ; p2 ; : : :). In this
hapter we
start the dis
ussion of
ontinuous time sto
hasti
pro
esses. They are not so easy to
onstru
t,
but of great pra
ti
al and theoreti
al importan
e. In this
hapter we get to know the most
important example, the Brownian motion.
Denition: Suppose I IR is an interval. A (
ontinuous time) sto
hasti
pro
ess (with values
in IRd) is a family fX (t) : t 2 I g of random variables with values in IRd dened on the same
probability spa
e (
; A; IP).
If we also
in
lude the variable ! in our
onsideration, we now write it in square bra
kets X (t) :
d
A.BENHARI
81
8.2
In 1827 the English botanist Brown observed that pollen parti
les suspended in a liquid perform irregular random movements. This is due to the hitting of pollen by the mu
h smaller
mole
ules of the liquid. These hits o
ur a large number of times in any small interval of time,
independently of ea
h other and hen
e a simple mathemati
al model of this pro
ess should be
a sto
hasti
pro
ess fB (t) : t 0g with the following features:
(1) for all times 0 t1 t2 : : : tn the displa
ements or in
rements B (tn)
B (tn 1 ); B (tn 1 ) B (tn 2 ); : : : ; B (t2 ) B (t1 ) are independent, we say that B (t) is a
pro
ess with independent in
rements,
(2) the distribution of the displa
ement B (t + h) B (t) does not depend on t, we say that the
pro
ess is stationary or has stationary in
rements.
(3) the pro
ess fB (t) : t 0g has almost surely
ontinuous paths.
A real valued pro
ess with the rst two properties is sometimes
alled a Levy Pro
ess. It should
be
lear that pro
esses with these features do
ome up naturally as a model in many other
situations in nature, s
ien
e and other elds of appli
ations. The rst question however must
be: Do nontrivial pro
esses with su
h features exist (in a mathemati
al sense)? In parti
ular,
it is not
lear whether the randomness required in the rst two
onditions does not
ontradi
t
A.BENHARI
82
the
ontinuity required in the last
ondition. We approa
h this question rst in the
ase of
dimension d = 1.
A good starting point for an answer to this question is to nd ne
essary
onsequen
es of these
assumptions. Surprisingly, it turns out that the distributions of the displa
ements are almost
ompletely determined by the three
onditions. Here is our rst major probabilisti
theorem.
Theorem 8.2 Suppose that fB (t) : t 0g is a real valued stationary pro
ess with independent
in
rements and almost surely
ontinuous paths. Then there are and 0 su
h that, for ea
h
t 0 and h 0, the in
rement B (t + h) B (t) is normally distributed with expe
tation h and
varian
e h2 .
We now
ome to the proof of Theorem 8.2. The proof uses the
entral limit theorem. We
re
all a formulation of the
entral limit theorem, whi
h was proved (in a slightly more general
form) in the le
ture \Sto
hastis
he Methoden", see Satz 4.2.
Central Limit Theorem: For ea
h n 2 IN let X1n ; : : : ; Xnn be independent and identi
ally
distributed random variables with expe
tation n, positive varian
e n2 and nite (
entred) third
moment
n = IEjXin IEXinj3 su
h that
lim
pn = 0 :
n!1 n3 n
Then we have, for all a b 2 [ 1; 1,
n Xn
1
1 b e x =2 dx :
n
i
p
lim
IP p
2
(
a;
b
)
=
n!1
n i=1 n
2 a
n
A.BENHARI
83
Proof of Theorem 8.2: We rst x t = 0 and h 0 and show that the in
rement B (h) B (0)
is normally distributed, before showing that expe
tation and varian
e have the given stru
ture.
To determine the distribution of the in
rements we x h > 0. For ea
h n 2 IN and 1 k n
we dene
hk
B h(kn 1) if this is less than 1 in absolute value,
Ykn = B n
0
otherwise.
(
Step 1:
Proof:
We now establish some fa
ts about the distribution of the Ykn. By our assumptions, for a given
n, all Ykn have the same distribution and are independent.
Step 2: For all > 0, we have limn!1 nIPfjY1n j g = 0.
Proof: Be
ause almost sure
onvergen
e implies
onvergen
e in probability we infer from (3.1)
h(k 1)
hk
n
1 = nlim
IP max B
B
<
!1
k=1
n
n
n
hk
h(k 1)
= lim IP
B
B
<
n!1
k=1
n
o
n
= nlim
!1 IP jB (h=n) B (0)j <
nIPfjB (h=n) B (0)j g n
= nlim
1
!1
n
= exp nlim
!1 nIPfjB (h=n) B (0)j g ;
using Euler's formula. Hen
e limn!1 nIPfjB (h=n) B (0)j g = 0. As jY1nj jB (h=n)
the statement follows.
n
IPfjY1n j g d
lim IPfjY1nj g d = 0:
0 n!1
Var
A.BENHARI
X
k=1
Ykn
B (0)j
= nVar(Y1n) ;
84
(3.2)
B (0) is
onstant.
k=1
IEYkn ;
k=1
In order to apply the
entral limit theorem to this expression we have to
he
k the moment
onditions. Let n be the expe
tation, n2 be the varian
e
and
n the third (
entred) moment
of X1n. They all exist be
ause jXkn j is bounded by pn.
n
Step 4: nlim
!1 n3 pn = 0.
Proof: We have, by Step 3,
liminf
2 = liminf
nVar(Ykn ) > 0 ;
n!1 n
n!1
Choose a small > 0. Let Z1n = X1n IEX1n. Be
ause jY1n IEY1nj is bounded by 2 we infer from
(3.2),
n = IE[jZ1n j3
n3=2 IE jY1n IEY1nj3 1fjY n EY n jg + 8IPfjY1n IEY1nj > g
pnnIE jY1n IEY1nj2 + o(pn)
pnIE[(Z1n)2 + o(pn) :
Hen
e
n
1
p
limsup :
limsup
3
n!1 n n
n!1 n
As
an be
hosen arbitrarily small and limsup1=n < 1 the statement follows.
h
lim IP
n!1
At the same time, by (3.3),
n Xn
1
i
lim IP p
n
n!1
A.BENHARI
n i=1
n
p1
Xin n
n i=1 n
n
2 (a; b) = 21
2 (a; b) = nlim
!1 IP
85
B (h) B (0)
n
x2 =2 dx :
pn
n
2 (a; b)
(3.4)
p
p
= nlim
IP B (h) B (0) 2 (an + nn ; bn + nn ) :
!1
This
an only hold true if (n ) and (pnn) stay bounded. Then we
an pi
k
onvergent subsequen
e with limits and and infer that the distributions of B (h) B (0) and X +
oin
ide
for a standard normally distributed random variable X . As X + is normally distributed with
expe
tation and varian
e 2 the proof is nished.
n
We
an now nish the proof of Theorem 8.2 by showing the spe
ial stru
ture of the expe
tations and varian
es.
Step 6: There is and 0 su
h that, for all t; h 0, IE(B (t + h) B (t)) = h and
Var(B (t + h) B (t)) = h2 .
Proof: By stationarity it su
es to show this for t = 0. Dene the fun
tion f; g : [0; 1) ! IR
by f (h) = Var(B (h) B (0)) and g(h) = IE(B (h) B (0)). For all h; k 0 we have by stationarity,
g(h + k) = IE B (h + k) B (h) + IE B (h) B (0) = g(k) + g(h) :
and, by Bienayme's equality and stationarity,
f (h + k) = Var(B (h + k) B (0)) = Var(B (h + k) B (k)) + Var(B (k) B (0)) = f (h) + f (k) :
Hen
e there is a with g(h) = h and a 2 0 su
h that f (h) = 2h.
We nish this se
tion with an example, whi
h shows that there are other pro
esses, whi
h have
stationary, independent in
rements but fail to have
ontinuous paths.
Example Consider X (t) the number of
ustomers arriving at a store by time t 0. Intuitively
the pro
ess fX (t) : t 0g has to satisfy the following assumptions
The number of
ustomers arriving during one time interval does not ae
t the number of
ustomers arriving in another disjoint time interval. Mathemati
ally, this means that the
pro
ess has independent in
rements.
The rate at whi
h
ustomers arrive should be
onstant, more pre
isely, there is some 0
su
h that IE[X (t) = t.
Customers arrive one at a time. To make this pre
ise we assume that X (t) is in
reasing,
takes values in IN and we have
IPfX (t + h) = X (t) + 1g = h + o(h);
IPfX (t + h) X (t) + 2g = o(h):
Brownian motion satises the rst two, but not the last assumption. A sto
hasti
pro
ess
fullling these assumptions is
alled a Poisson pro
ess with rate (or intensity) . We shall see
as an exer
ise that it is uniquely determined up to equivalen
e. Here is one way to
onstru
t it:
Let S be a Poisson distributed random variable with parameter and Y1; Y2 ; Y3; : : : independent
random variables with uniform distribution on [0; 1). For 0 t 1 let
X (t) = #fYi : Yi t and i S g :
Then X satises the assumptions on the interval [0; 1) and we extend X to [0; 1) by glueing
together independent
opies of X .
A.BENHARI
86
8.3
In this se
tion we prepare the existen
e proof of Brownian motion with some lemmas about
Gaussian random variables. This and the following two se
tions are essentially taken from Peres
(1998).
We have seen that the normal distribution
omes up naturally in the study of sto
hasti
pro
esses. One of the drawba
ks of this distribution is that its distribution fun
tion
an not be
expressed in terms of
lassi
al fun
tions. Therefore the following (quite pre
ise) estimate will
later be useful.
Lemma 8.4 Suppose X is standard normally distributed. Then, for all x 0,
x
p1 e x =2 IPfX > xg x1 p1 e x =2:
x2 + 1 2
2
Proof: The right inequality is obtained by the estimate
1 1 u e u =2 du = 1 p1 e x =2 :
IPfX > xg p
x 2
2 x x
2
u2 =2 du :
2x x e u =2 du = 2x x e u =2 du e
whi
h is positive for x 0, by the rst part. Hen
e f (x) 0, proving the lemma.
f 0(x) = (1 x2 + x2 + 1)e
x2 =2
Z
x2 =2
We now look more
losely at random ve
tors with normally distributed
omponents. Our motivation is that they arise, for example, as ve
tors
onsisting of the in
rements of a Brownian
motion. Let us
larify some terminology.
Denition: A random variable X = (X1 ; : : : ; Xd )T with values in IRd has the d-dimensional
standard Gaussian distribution if its d
oordinates are standard normally distributed and independent (independent whether we write it as row{ or
olumn). A random variable Y with values
in IRn is
alled Gaussian if there exists an n d matrix A and an n dimensional ve
tor b su
h
that Y T = AX + b. The
ovarian
e matrix of the ve
tor Y is given by
Cov(Y ) = IE (Y IEY )(Y IEY )T = AAT ;
here the expe
tations are dened
omponentwise.
Our rst lemma shows that applying an orthogonal d d matrix does not
hange the distribution
of a standard Gaussian random ve
tor.
Lemma 8.5 If A is an orthogonal d d matrix, i.e. AAT = Id , and X is a d-dimensional
h
A.BENHARI
87
Proof: As the
oordinates of X are independent, standard normally distributed, X has a density
f (x1 ; : : : ; xd ) =
p1 e
x2i =2
= (21)d=2 e
kxk2 =2 ;
2
where kp k is the Eu
lidean norm. The density of AX is (by the transformation rule)
f (AT x) AAT . The determinant is 1 and hen
e, sin
e orthogonal matri
es preserve the Eu
lidean norm, the density of X is invariant under A.
i=1
Corollary 8.6 Let X1 and X2 be independent and normally distributed with expe
tation 0 and
varian
e 2 > 0. Then X1 + X2 and X1 X2 are independent and normally distributed with
expe
tation 0 and varian
e 22 .
Proof: (X1 =; X2 =)T
p1
2
p12
p1
2
p12
This isp an orthogonal matrix and applying it to our ve
tor yields ((X1 + X2 )=(p2); (X1
X2 )=( 2)), whi
h thus must have independent standard normal
oordinates.
The next lemma shows that the distribution of a Gaussian random ve
tor is determined by its
expe
tation and
ovarian
e.
Lemma 8.7 If X and Y are d-dimensional Gaussian ve
tors with IEX
Cov(Y ), then X and Y have the same distribution.
= IEY
and
Cov(X ) =
It is su
ient to
onsider the
ase IEX = IEY = 0. By denition, there are standard
Gaussian random ve
tors X1 and X2 and matri
es A and B with X = AX1 and Y = BX2 .
By adding
olumns of zeroes to A or B , if ne
essary, we
an assume that X1 and Xk2 are both
k-ve
tors and A; B are both d k matri
es. Let A and B be the ve
tor subspa
es of IR generated
by the row ve
tors of A resp. B . To simplify notation assume that the rst l row ve
tors of A
form a basis of A. Dene the linear map L : A ! B by
L(Ai ) = Bi for i = 1; : : : ; l:
Here Ai is the ith row ve
tor of A. Our aim is to show that L is an orthogonal isomorphism
and then use the previous lemma. Let us rst show that L is an isomorphism. Our
ovarian
e
assumption gives that AAT = BB T . Assume there is a ve
tor v1A1 + : : : vl Al whose image is 0.
Then the d-ve
tor
v = (v1 ; : : : ; vl ; 0; : : : ; 0)
satises vB = 0. Hen
e
kvAk2 = vAAT vT = vBB T vT = 0 :
We
on
lude that vA = 0. Hen
e L is inje
tive and dim A dim B. Inter
hanging the roles
of A and B gives that L is an isomorphism. As the entry (i; j ) of AAT = BB T is the s
alar
Proof:
A.BENHARI
88
We end this se
tion with the denition of an important
lass of pro
esses.
Denition
A sto
hasti
pro
ess fY (t) : t 2 I g is
alled a Gaussian pro
ess, if for all t1 t2 : : : tn the
ve
tor (Y (t1 ); : : : ; Y (tn )) is a Gaussian random ve
tor.
Show as an exer
ise that every Brownian motion is a Gaussian pro
ess, i.e. given times
t1 : : : tn nd a matrix A and a ve
tor b su
h for a standard Gaussian ve
tor X , we have
(B (t1 ); : : : ; B (tn)) = AX +
8.4
The
entral result of this se
tion is the following theorem, whi
h establishes the existen
e of
Brownian motions. Re
all that we need only
onstru
t a standard Brownian motion B , as
X (t) = x + B (t) + t is a Brownian motion with arbitrary starting point, drift and diusion
onstant.
Theorem 8.9 (Wiener 1923) Standard Brownian motion exists.
We prove this using a
onstru
tion of Paul Levy (1948). We will use the existen
e theorem
for
ountable produ
ts. We rst
onstru
t Brownian motion on the interval [0; 1 as a random
element on the spa
e C [0; 1 of
ontinuous fun
tions. The idea is to
onstru
t the right values of
Brownian motion step by step on the nite sets
Dn = 2kn : 0 k 2n
of dyadi
points. If the values on Dn are
onstru
ted we interpolate them linearly and later we
dene Brownian motion as the uniform limit of these
ontinuous fun
tions, whi
h is automati
ally
ontinuous.
To do this let D = 1n=0 Dn and let (
; A; IP) be a
ountable produ
t spa
e su
h that a
olle
tion
fZd : d 2 Dg of independent, standard normally distributed random variables
an be dened
on the spa
e. Let B (0) = 0 and B (1) = Z1 . For ea
h n 2 IN we
onstru
t random variables
B (d), d 2 Dn su
h that
for all r < s < t in Dn the random variable B (t) B (s) has N (0; t s)-distribution and
is independent of B (s) B (r),
n
A.BENHARI
89
2 n) + B (d + 2 n) + Zd :
2
2(n 1)=2
The values of B used in this denition are both in Dn 1 and the se
ond property is
learly
fullled. Sin
e (1=2)[B (d + 2 n) B (d 2 n ) is (by indu
tion) normally distributed with
expe
tation 0 and varian
e 1=2n+1 and Zd =2(n+1)=2 is independent with the same distribution,
their sum B (d) B (d 2 n) and their dieren
e B (d + 2n) B (d) are independent with the
same distribution by Corollary 3.6. Now if d 2 Dj for some 1 j n every in
rement of length
2 n with right endpoint a dyadi
point in (d 2 j ; d is independent from all in
rements of
length 2 n with left endpoint a dyadi
point in [d; d + 2 j ), by
onstru
tion and independen
e
of the Zd for dierent d. The rst of our properties follows from this.
Having thus
hosen the values of the pro
ess on all dyadi
points, we interpolate between them.
Formally, dene
Z1
for x = 1
F0 (x) = 0
for x = 0
linear in between.
and, for ea
h n 0,
2 (n+1)=2 Zx for x 2 Dn n Dn 1
Fn (x) = 0
for x 2 Dn 1
linear
between
onse
utive points in Dn:
These fun
tions are
ontinuous on [0; 1 and for all n and d 2 Dn
B (d) =
8
>
<
>
:
8
>
<
>
:
B (d) =
i=0
Fi (d) =
i=0
Fi (d):
(3.5)
This
an be seen by indu
tion. It holds for n = 0. Suppose that it holds for n 1. Let
Sin
e for 0 i n 1 the fun
tion Fi is linear on [d 2 n; d + 2 n, we get
n 1
n 1 F (d 2 n ) + F (d + 2 n )
B (d 2 n ) + B (d + 2 n )
i
i
Fi (d) =
=
:
2
2
i=0
i=1
Sin
e Fn(d) = 1=2(n 1)=2 Zd , this gives (3.5).
On the other hand, we have, by denition of Zd and by Lemma 3.4, for large n,
d 2 Dn n Dn 1 .
X
png exp
IPfjZd j
n=0
A.BENHARI
IPf9d 2 Dn
2 n
X X
n=0 d2Dn
90
png
IPfjZd j
n=0
Fn (t)
is uniformly
onvergent on [0; 1 and we denote the
ontinuous limit by fB (t) : t 2 [0; 1g.
It remains to
he
k that the in
rements of this pro
ess have the right nite-dimensional joint
distributions. This follows dire
tly from the properties of B on the dense set D [0; 1 and the
ontinuity of the paths. Indeed, suppose that t1 > t2 > t3 are in [0; 1. We nd sequen
es ti;n in
D
onverging to ti and infer from the
ontinuity of B that
B (t3 ) B (t2 ) = nlim
!1 B (t3;n ) B (t2;n ) :
As a limit of normally distributed random variables, this in
rement is itself normally distributed
with mean 0 and varian
e
lim t t2;n = t3 t2 :
n!1 3;n
The analogous fa
t holds for B (t2) B (t1). Moreover, the random variables entering in the
limits are independent if n is large enough that t1;n > t2;n > t3;n. Hen
e, the in
rements must
be independent, too. From Corollary 3.8 we infer that the pro
ess has independent in
rements
on [0; 1.
We have thus
onstru
ted a Brownian motion B on [0; 1. By the existen
e of produ
t spa
es
there exists a probability spa
e on whi
h a sequen
e B1; B2 ; : : : of independent C [0; 1-valued
random variables with the properties of our B
an be dened. We glue them together by letting
B (t) = B[t (t
[t) +
[X
t 1
i=0
Bi (1) :
This denes a
ontinuous random fun
tion in C [0; 1) and one
an see easily from what we have
shown so far that the requirements of a standard Brownian motion are fullled.
We end this
hapter by noting that our pro
edure denes a natural probability measure W on
the spa
e ff 2 C [0; 1) : f (0) = 0g of
ontinuous real valued fun
tions starting in 0. This
probability measure is
alled the Wiener measure
8.5
The Brownian motion has two very useful invarian
e properties. The rst of them is the s
aling
invarian
e, the se
ond is the invarian
e under time-inversion. In ea
h
ase there is a transformation on the spa
e of fun
tions, whi
h
hanges the individual Brownian random fun
tions but
leaves their distribution un
hanged.
A.BENHARI
91
B (a2 t)
a
Continuity of the paths, independen
e and stationarity of the in
rements remain un
hanged under the res
aling. It remains to observe that
1
X (t) X (s) = B (a2 t) B (a2 s)
a
is normally distributed with expe
tation 0 and varian
e (1=a2 )(a2 t a2s) = t s.
Proof:
: t 0g
0
t=0
tB (1=t) t > 0
Proof: Like Brownian motion fX (t)g is a Gaussian pro
ess su
h that the Gaussian random
ve
tors (X (t1 ); : : : ; X (tn )) have expe
tation 0. The
ovarian
es, for t; h 0, are given by
Cov(X (t + h); X (t)) = (t + h)t Cov(B (1=t + h); B (1=t)) = t(t + h) t +1 h = t :
Hen
e X is a variant of Brownian motion. Its paths are
learly
ontinuous for all t > 0 and in
t = 0 we use the following two fa
ts: First, the distribution of X on the rationals Q is the same
as for a Brownian motion, hen
e
lim X (t) = 0 almost surely.
t!0;t2Q
Using the time-inversion we see that limt!1 B (t)=t = limt!1 X (1=t) = X (0) = 0.
We now prove two theorems that make the degree of
ontinuity of the paths of Brownian motion
more pre
ise. Pay attention to the order of the almost surely and the for ea
h in the following
theorem and note that a
hange of this order would give a (
orre
t, but) mu
h weaker statement.
A.BENHARI
92
Theorem 8.13 There exists a random variable C su
h that, almost surely, for ea
h
t + h 1,
q
B
(
t
+
h
)
B
(
t
)
C
h log(1=h) :
Proof: We go ba
k to the
onstru
tion of the Brownian motion.
motion as a series
B (t) =
n=0
0t
Fn (t) ;
where ea
h Fn is a pie
ewise linear fun
tion. Its derivative exists almost everywhere, and by
denition and (3.6),
kFn0 k1 2k2Fnnk1 C1(!) + 2
pn2n=2 :
The random
onstant C1 is here to deal with the nitely many ex
eptions to (3.6). Now for ea
h
t; t + h 2 [0; 1,
jB (t + h) B (t)j
n=0
jFn(t + h) Fn(t)j
n=0
hkFn0 k1 +
n=l+1
2kFn k1 :
X
n=0
C1 (!) + 2 n2n=2
+2
n=l+1
n2
n=2
p
p
C2(!)h l2l=2 + C3(!) l2
l=2 :
Choosing l = log2(1=h) and C (!) su
iently large to take
are of the
ases l N we get
h log(1=h)
Theorem 8.14 There exists a
onstant
> 0 su
h that, almost surely, for every " > 0 there
exist t 0 and 0 < h < " with
q
p
IP(Ak;n ) = IPfB (2 )) >
n2
n
A.BENHARI
n=2 g
93
B (k2
) > pn2
n=2
p
p
n
= IPfB (1) >
ng
2n + 1 e
2 n=2 :
xe
IP
2n 1
\
k=0
for all x,
A k;n
= (1
2n IP(Ak;n) ! 1 :
(Ak;n))2n exp( 2nIP(Ak;n)) ! 0 :
IP
= 0:
and, for
We shall
ome ba
k to a longer dis
ussion of the properties of Brownian motions later in the
le
ture, in parti
ular we will then prove that Brownian motion is nowhere
8.6
In order to model a time homogenous sto
k pri
e in the simplest manner one would require a
pro
ess fX (t) : t 0g with the following features:
the pro
ess X is almost surely
ontinuous and has positive values,
for all 0 t1 t2 : : : tn the returns X (tn )=X (tn 1 ); X (tn 1 )=X (tn 2 );
: : : ; X (t2 )=X (t1 ) are independent.
the distribution of the returns X (t + h)=X (t) does not depend on t.
These assumption already determine the model up to two
onstants:
A.BENHARI
94
Theorem 8.16 Suppose that the pro
ess fX (t) : t 0g satises the assumptions above and
X (0) = x > 0 is the initial pri
e. Then there is a and a su
h that log X (t) is a Brownian
motion with start in log x, drift and varian
e 2 . The pro
ess X is
alled geometri
Brownian
motion with trend parameter and volatility parameter .
Proof:
log X satises the onditions of the Chara terization Theorem for Brownian motion.
Of
ourse, this model is not the nal word about modelling a sto
k pri
e. For example, in
pra
ti
e, trend and volatility parameters will depend on t. The denition of su
h more
ompli
ated pro
esses has to be deferred to the se
ond part of the le
ture, when methods of sto
hasti
analysis are available.
A.BENHARI
95
We start the dis
ussion of the Markov property with a thorough introdu
tion of the term
onditional expe
tation, whi
h is essential to the
on
epts of the Markov property and also of
martingales.
9.1
Suppose that (
; A; IP) is a probability spa
e and X and Z are random variables on this spa
e.
We assume that both random variables take on nitely many real values fx1 ; : : : ; xm g resp.
fz1 ; : : : ; zn g ea
h with positive probability. We suppose that X and Z are not independent. How
does knowledge about the out
ome of Z in
uen
e the out
ome of X ? This
an be des
ribed by
means of
onditional probabilities
IP X = xi and Z = zj
:
IP X = xi j Z = zj :=
IPfZ = zj g
One way to look at this is the following: if we have observed the event fZ = zj g this
hanges
our per
eption of the random variable X , it is now dened on a dierent probability spa
e, on
whi
h only those !
an o
ur, whi
h satisfy Z (!) = zj . The new spa
e still
onsists of the set
and the -eld A, but the probability measure is now
on
entrated on the set
Z 1(zj ) = f! 2
: Z (!) = zj g :
The new probability measure IPfjZ = zj g is given by
IP A and Z = zj
IP A j Z = zj :=
:
IPfZ = zj g
The random variable X may still be dened on the spa
e, but its distribution has
hanged and
is now the
onditional distribution given Z = zj . Its expe
tation is now
n
IE
A.BENHARI
X j Z = zj
:=
k=1
96
xk IP X = xk jZ = zj :
This expe tation depends on zj and an be interpreted as a mapping (or random variable) on
namely by
IE X j Z :
! IR ;
IE X j Z (! ) = IE X j Z = Z (! ) :
The mapping IEfX jZ g has the following properties :
1) it is measurable with respe
t to the -eld generated by the sets Z 1fzj g for j = 1; : : : ; n.
2) for every A fz1 ; : : : ; zn g we have Z (A) IEfX jZ g(!)dIP(!) = Z (A) X (!)dIP(!).
n
The rst property just means that IEfX jZ g is
onstant on ea
h set Z 1fzj g and the se
ond
property follows from the
al
ulation
IEfX jZ g(! ) dIP (! ) =
IEfX jZ = z gIPfZ = z g
Z (A)
Z
z 2A
X X
z 2A k=1
xk IP X = xk jZ = z
= zg ;
IPfZ
X X
z 2A k=1
xk IP X = xk
X (!)
m
X
k=1
and Z = z =
1fX (!)=xk g
z 2A
X X
z 2A k=1
(!) =
Z 1 (A)
X (!) dIP(!) :
One
ould say that we have de
omposed the probability spa
e a
ording to the information
oming from Z , ea
h probability measure IPfjZ = zj g is
on
entrated on a dierent region of
the spa
e
, be
ause the additional information we had
onsisted of the knowledge of the
ell
in whi
h the random ! was to be found.
What
an we say if additional information is
oming into our experiment from other sour
es
than the observation of the values of a dis
rete random variable su
h as Z ? Can we get a similar
de
omposition? How
an we des
ribe information?
Information from our point of view is a sub
olle
tion B A of events, about whi
h we have
already the knowledge whether they o
ur or not. Su
h a sub
olle
tion, by simple
onsiderations
is ne
essarily a -eld. Hen
e we adopt the point of view that information
an be modelled by
means of (sub)--elds B A. Let us give some examples:
Suppose we have observed the value of a random variable Z 0 : (
; A; IP) ! (
0 ; A0). Then
we have knowledge about all the events fZ 2 Ag if A runs through the -eld A0. These
sets form a -eld Z 1(A0). In the
ase that Z takes on only nitely many values (as
before) the -eld
onsists of all nite unions of the
ells (or atoms) Z 1(zj ).
Suppose we have observed the path of a sto
hasti
pro
ess fX (t) : t 0g up to time T .
The information gained by this is en
oded in the -eld FT generated by all the in
rements
X (t) X (s) for s t T , equivalently generated by
fX (tn ) X (tn 1 ) 2 An; : : : ; X (t1 ) X (t0 ) 2 A1 g
A.BENHARI
97
2 F,
X dIP :
Any two random variables Y satisfying these two
onditions
oin
ide almost surely. If F =
Z 1 (A0 ) is the -eld of preimages of all Borel sets under a random variable Z , one also writes
IEfX jZ g for IEfX jFg and says this is the
onditional expe
tation of X given Z .
Remark:
By
hosing F =
in the last property, we see that
IE IEfX jFg = IEX :
n
We start with the proof of the uniqueness. Suppose Y1 and Y2 both satisfy the denition
of a
onditional expe
tation. Suppose further that IPfY1 > Y2 g > 0. Then there is a natural
number n su
h that IPfY1 Y2 > 1=ng > 1=n. Note that the event F = fY1 Y2 > 1=ng is in
F . Hen
e
1=n2 F (Y1 Y2) dIP = F X dIP F X dIP = 0 ;
a
ontradi
tion. Therefore IPfY1 > Y2 g = 0. Ex
hanging the role of Y1 and Y2 yields
IPfY1 = Y2 g = 1.
Z
Before we prove existen
e, we formulate some properties of
onditional expe
tations. Note rst
that uniqueness implies the following property
alled linearity: For all a; b real, almost surely,
IE aX + bY jFg = aIEfX jFg + bIEfY jFg :
Pay attention to the order of the quantors and of the almost surely. The following theorem is to
be proved as an exer
ise. It des
ribes the
onvergen
e properties of
onditional expe
tations.
n
A.BENHARI
98
Monotone Convergen e If 0 Xn " X , then IEfXn jFg " IEfX jFg almost surely.
liminf
X j Fg liminf
IEfXn j Fg almost surely.
n!1 n
n!1
IEf
Dominated Convergen
e If there is a random variable Z su
h that IEZ < 1 and jXn j Z
for all n, and if Xn ! X almost surely, then IEfXn jFg ! IEfX jFg.
The next set of properties is of more probabilisti
nature. If F and G are -elds we denote by
F _ G the -eld generated by their union.
Theorem 9.3 Let (
; A; IP) be a probability spa
e, F A a sub--eld and X a random variable
with IEjX j < 1. Then every
onditional probability IEfX jFg has the following properties.
IEfX jFgG
H and
= IEfX jGg :
A.BENHARI
99
We show the more general statement that, for all h : IR ! [0; 1),
Z
Note that, if h = 1A , the left hand side is IP(F \ fX 2 Ag) and the right hand side is
IP(F )IPfX 2 Ag, so that our statement is the denition of independen
e of the events fX 2 Ag
and F . By the standard measure theory ma
hinery this
an be extended to arbitrary h and
plugging in h(x) = x gives the desired result. The last property is similar and an exer
ise.
The proof of the existen
e illustrates that IEfX jFg is in a sense a proje
tion of X on the
spa
e of F -measurable random variables. Our strategy is to show this rst for random variables
X satisfying IEjX j2 < 1. The idea is that then X is an element of the Hilbert spa
e
L2(
; A; IP) = X : (
; A) ! IR; A-measurable with IEjX j2 < 1
of square integrable fun
tions on the probability spa
e with two fun
tions X and Y identied if
IPfX = Y g = 1. The s
alar produ
t on this spa
e is
hX; Y i = IE XY :
Hen
e the norm is kX k2 = IEjX j2 . If you do not know that this denes a Hilbert spa
e, then see
the exer
ises, where the
ompleteness of this spa
e is proved, by
he
king that every Cau
hy
sequen
e
onverges:
Lemma (Completeness): If fXn g is a sequen
e of random variables on a probability spa
e
(
; A; IP) su
h that IEjXnj2 < 1 for all n and limk!1 supn;mk IEjXn Xm j2 = 0 ; then there is
a random variable X on (
; A; IP) su
h that limn!1 IEjXn X j2 = 0 :
Hilbert spa
es are the mathemati
al stru
ture that allows to make sense of the notion of proje
tions. For our purpose we need the following.
n
kX Y k = I := inf fkX W k : W 2 V g,
hX Y; Z i = 0 for all Z 2 V .
Remark: Y
surely.
is alled the proje tion of X onto Y and any two proje tions are equal almost
Proof: Re
all that
ompleteness of V means that every Cau
hy-sequen
e of random ve
tors Yn
in V
onverges in norm to a random variable Y 2 V . We rst
hoose a sequen
e fYng in V su
h
that
kX Ynk ! I :
A.BENHARI
100
Ym k2
Be
ause 12 (Yn Ym ) 2 V , the subtra
ted term is at least I 2 and hen
e fYng is a Cau
hy sequen
e
and
onverges to some Random variable Y 2 V . By the triangle inequality
I kX Y k kX Yn k + kY Yn k ! I ;
and thus kX Y k = I . For every Z 2 V we have Y + tZ 2 for ea
h t 2 IR and hen
e
kX Y tZ k2 kX Y k2 ;
whi
h implies 2thZ; X Y i + t2kZ k2 0. As a fun
tion of t this has a minimum in t = 0 and
We now show the existen
e of
onditional expe
tation rst for a random variable X 2
L2 (
; A; IP). We proje
t X onto the
omplete subspa
e
V := L2 (
; F ; IP) L2 (
; A; IP) ;
and
all the result Y . As Y 2 V it is F -measurable. Moreover,
hX Y; Z i = 0 for all Z 2 V .
If F 2 F we may
hoose Z = 1F 2 V there and get
X dIP = hX; 1F i = hY; 1F i = Y dIP :
F
F
Hen
e Y is a
onditional expe
tation of X given F . It remains to show the same if X just
fullls IEjX j < 1. It su
es to
onsider the
ase X 0. Choose Xn = X ^ n and observe that
Xn 2 L2 (
; A; IP) and there are
onditional expe
tations IEfXn jFg. By Theorem 4.2 (Positivity)
Z
= nlim
!1 IEfXn jFg :
Clearly, Y is F -measurable and, as Xn in
reases to X , we
an use monotone
onvergen
e to see
that, for every F 2 F ,
Y dIP(!) = nlim
!1 F IEfXn jFg dIP(!) = nlim
!1 F Xn dIP(!) = F X dIP(!) :
F
This also implies IEY = IEX < 1 and hen
e Y is a
onditional expe
tation of X given F .
Z
As an exer
ise we investigate the situation when two real valued random variables X and Z
have a joint density f (x; z). Then,
learly, the density of X is fX : x 7! f (x; z) dz and the
density of Z is fZ : z 7! f (x; z) dx. Assume that
IEjX j = jxjfX (x) dx = jxj f (x; z ) dz dx < 1 :
One
an dene a fun
tion
if fZ (z) 6= 0 ;
fX jZ (xjz ) = 0f (x; z )=fZ (z ) otherwise
.
With this fun
tion dene a random variable
Y = xfX jZ (xjZ ) dx :
Then Y is a
onditional expe
tation of X given Z .
R
A.BENHARI
101
9.2
01-Law
Suppose that fX (t) : t 0g is a sto
hasti
pro
ess. Intuitively, the Markov property of a pro
ess
says that if we know the pro
ess fX (t)g on the interval [0; s, this is as useful as just knowing
the endpoint X (s). We will see that Brownian motion is a pro
ess having this property.
We now work on the spa
e
= C ([0; 1)) of
ontinuous fun
tions with the Borel -eld A
generated by all in
rements and the probability measures IPx, whi
h are the distributions of
Brownian motion started in x. The
orresponding expe
tations and
onditional expe
tations
are denoted IEx and IExfjg. On
we have for every s > 0 shift transformations s; s :
!
dened by
sB (t) = B (s + t)
and sB (t) = B (s + t) B (s) :
Note that one
an
onsider s and s as
-valued random variables and thus as sto
hasti
pro
esses by
s (t) = s (t)[B = sB (t) = B (t + s) ;
and analogously for s. Dene the -eld F 0 (t) to be the -eld generated by the in
rements
B (tn ) B (tn 1 ); : : : ; B (t2 ) B (t1 ) for all 0 t1 tn t. Note that
F 0(t) F 0 (s) for all t s :
A family fF (t) : t 0g of -elds su
h that F (t) F (s) for all t s is
alled a ltration.
Theorem 9.5 (Weak Markov property) For every s 0 and every bounded random variable Y : (
; A) ! IR we have
In order to prove the result in this generality we need one more measure theoreti
tool, the
monotone
lass theorem. This is an improvement of our standard measure theory ma
hinery.
Lemma 9.6 (Monotone
lass theorem) Let B be a \-stable system that
ontains
and
generates the -eld A. Let H be a
olle
tion of real valued fun
tions
ontaing all indi
ators
1A ; A 2 B su
h that
The proof of this
an be found in the book of Durrett, Chapter 5, Theorem (1.5) or in most
texts on measure theory.
A.BENHARI
102
Proof: The parti
ular statement at the end of the theorem is evident from the independen
e of
the in
rements of Brownian motion, be
ause F 0 (s) is generated by the in
rements up to time
s and independen
e of F 0 (s) from the given pro
ess starting in 0 is just independen
e from its
in
rements. In the formal language of
onditional expe
tations we get, for ea
h Y of the spe
ial
form Y (B ) = X (B (0))Z (0 B ), using 0s = s,
ages of) B (s). The argument
an be extended from these spe
ial Y to general bounded Y by
applying the monotone
lass theorem to the
olle
tion H of all fun
tions Y for whi
h the theorem
holds and using the sets of the form fB (0) 2 A1 g \ f0B 2 A2g, A1 IR Borel, A2 2 A, as
\-stable
olle
tion B.
We now dis
uss a formal generalization of the weak Markov property, whi
h has surprising
appli
ations. For this we make ea
h -eld a bit larger by allowing an innitesimal glan
e into
the future. Dene
F + (s) = F 0 (t) :
\
t>s
fF + (s)g
Then
is a slightly larger ltration, for whi
h the weak Markov property still holds.
Re
all that we are dealing with Brownian motion dened on C ([0; 1)).
Theorem 9.7 For every s 0 and every bounded random variable Y
IEx fY
: (
; A) ! IR we have
We now look at the germ eld F +(0), whi
h heuristi
ally
omprises all events dened in terms
of Brownian motion on an innitesimal small interval to the right of the origin.
Theorem 9.8 (Blumenthal's 01-law.) Let x 2 IR and A 2 F + (0). Then IPx fAg 2 f0; 1g.
Proof:
A.BENHARI
103
n=1
is
learly in F +(0). Hen
e we just have to show that this event has positive probability. This
follows, as, for t > 0,
IP0 f tg IP0 fB (t) > 0g = 1=2 :
Hen
e IP0f = 0g 1=2 and we have shown the rst part. The same argument works repla
ing
B (t) > 0 by B (t) < 0 and from these two fa
ts IP0 f = 0g = 1 follows, using the intermediate
value property of
ontinuous fun
tions.
As an exer
ise we prove another 01-law. Dene G (t) to be the -eld dened by all in
rements
t1. G (t) des
ribes the future at time t. Let T = t0 G (t) be the
Heuristi
ally, the weak Markov property states that Brownian motion is started anew at ea
h
deterministi
time instan
e. It is a
ru
ial property of Brownian motion that this holds also for
an important
lass of random times. These random times are
alled stopping times, they are of
vital importan
e.
The basi
idea is that a random time T is a stopping time if we
an de
ide whether fT < tg by
just knowing the path of the sto
hasti
pro
ess up to time t. Think of the situation that T is
the moment where some random event related to the pro
ess happens.
Denition
A random variable T with values in IR [ f1g is
alled a stopping time with respe
t to the
ltration fF (t) : t 0g if, for every t 0, fT < tg 2 F (t). It is
alled a stri
t stopping time if,
for every t 0, fT tg 2 F (t). Every stri
t stopping time is also a stopping time, be
ause
fT < tg =
A.BENHARI
fT t 1=ng 2 F (t) :
n=1
104
For
ertain ni
e ltrations stri
t stoppping times and stopping times agree. In order to
ome
into this situation we are going to work with the ltration fF + (t)g in the
ase of Brownian
motion and refer the notions of stopping time, et
. always to this ltration. As this ltration
is larger than fF 0 (t)g, there are more stopping times. The
ru
ial property whi
h distinguishes
fF + (t)g from fF 0 (t)g is right-
ontinuity, whi
h means that
F +(t + ") = F + (t) :
\
">0
">0
F +(t + ") =
1 1
\ \
n=1 k=1
Theorem 9.11 Every stopping time T with respe
t to the ltration fF + (t)g is automati
ally a
stri
t stopping time.
Proof:
fT < t + 1=kg 2
k=1
n=1
s2Q
fT tg =
n=1
fTn tg 2 F +(t) :
stopping time.
Proof: Let G(n) be an open neighbourhood of K with K = G(n). Then Tn = inf ft
0 : B (t) 2 G(n)g are stopping times, whi
h are in
reasing to T .
T
A.BENHARI
105
In other words, we stop at the rst time of the form k2 n after T . It is easy to see that
Tn is a stopping time. We use it later as a dis
rete approximation to T .
We now want to prove that Brownian motion starts anew at ea
h stopping time. For this purpose
we dene, for every stopping time T , the -eld
F (T ) = fA 2 A : A \ fT < tg 2 F +(t) for all t 0g :
This means that the part of A that lies in fT < tg should be measurable with respe
t to the
information available at time t. Heuristi
ally, this is the
olle
tion of events that happened
before the stopping time T . As in the proof of the last theorem we
an infer that fT tg may
repla
e fT < tg without
hanging the denition.
We need three lemmas, the third of whi
h is proved as an exer
ise.
Lemma 9.12 If S T are stopping times, then F (S ) F (T ).
Proof:
T
Lemma 9.13 If Tn # T are stopping times, then F (T ) = 1
n=1 F (Tn ).
Proof:
By the last lemma, F (Tn )
T1
A 2 n=1 F (Tn ), then for all t 0,
A \ fT < tg =
Hen e A 2 F (T ).
n=1
This result
an be used in the proof of the following lemma. Use the dis
rete approximation of
T by a sequen
e Tn # T , see the last example.
Lemma 9.14 If T is a stopping time, then the random variable B (T ) is F (T )-measurable.
Theorem 9.15 (Strong Markov property) For every almost surely nite stopping time T
0 and every bounded random variable Y :
IR ! IR we have
IEx fY
fB (T + t) B (T ) : t 0g
is a standard Brownian motion independent of F (T ).
A.BENHARI
106
Proof: We show our statement rst for stopping times S su
h that only
ountably many values
s1 < s2 < s3 < : : : are taken with positive probability and we
an use this to
ondition with
respe
t to the value of S . Let A 2 F (S ), then
Z
Y (S B; S ) dIPx (B ) =
Now,
Y (S ; S ) dIPx =
Y (sn ; sn ) dIPx :
A \ fS = sn g = (A \ fS sng) n (A \ fS sn 1 g) 2 F + (sn ) ;
so by the denition of
onditional expe
tation and the weak Markov property, the sum is equal
to
1
IEB (S ) fY
(; S )g dIPx =
IEB (S ) fY
IEB (sn ) fY
(; sn )g dIPx
(; S )g dIPx :
Y (T ; T ) dIPx =
IEB (T ) fY
(; T )g dIPx :
A.BENHARI
107
Now, in order to pass to more general Y , one
an rst pass to an indi
ator of the form
Y (f; T ) = 1A (f (tn ) f (tn 1); : : : ; f (t2 ) f (t1 ); T ) ;
for A IRn open, by applying our result to a monotonously in
reasing sequen
e of
ontinuous
fun
tions Yk
onverging to Y and using the theorem of monotone
onvergen
e. Then we
an use
the monotone
lass theorem to the
olle
tion H nof all fun
tions Y for whi
h the theorem holds,
using the inverse images of the open sets A IR under
(f; T ) 7! (f (tn) f (tn 1); : : : ; f (t2) f (t1); T ) 0 t1 : : : tn :
as \-stable
olle
tion B. Finally, note that the parti
ular statement at the end follows immediately from the formula just proven.
Let T = inf ft 0 : B (t) = max0s1 B (s)g. It is intuitively
lear that T is not a
stopping time. To prove it, observe that almost surely T < 1. B (t + T ) B (T ) does not take
negative values in a small neighbourhood to the right of 0, whi
h
ontradi
ts the strong Markov
property and Theorem 4.9.
We will see many appli
ations of the strong Markov property later, the next subse
tion names
an interesting one.
Remark:
9.4
The re
e
tion prin
iple states that Brownian motion re
e
ted at some stopping time T is still
a Brownian motion. More formally:
Theorem 9.16 (Re
e
tion prin
iple) If T is a stopping time and fB (t)g is a standard
Brownian motion, then Brownian motion re
e
ted at T dened by
B (t) = B (t)1ftT g + (2B (T ) B (t))1ft>T g
is also a standard Brownian motion.
Let M (t) = max0st B (s). A priori it is not at all
lear what the distribution of this random
variable is, but we
an determine it as a
onsequen
e of the re
e
tion prin
iple. Re
all that B (t)
has distribution N (0; t).
A.BENHARI
108
Theorem 9.17 If a > 0 then IP0 fM (t) > ag = 2IP0 fB (t) > ag = IP0 fjB (t)j > ag.
A.BENHARI
109
10.Martingales
In this
hapter we get to know the pro
esses whi
h
orrespond to fair games, the martingales.
We rst study the theory mainly for dis
rete time and later (in Probability II ) extend the
theory to
ontinuous time. Symmetri
random walk and the
riti
al Galton-Watson pro
ess
(whose ospring distribution has expe
ted value one) turn out to be important examples of
dis
rete time martingales. The word martingales originally denotes a spe
ial gambling strategy,
indi
ating the
onne
tion to fair games.
10.1
We now look at dis
rete time pro
esses fXn g and a ltration, i.e. an in
reasing sequen
e
F (0) F (1) F (2) : : :
of -elds. We always assume that the pro
ess fXn g is fF (n)g-adapted, whi
h means that
Xn is F (n)-measurable,
but in most
ases we even
onsider the ltration F (n) generated by the (preimages of) random
variables X0 ; : : : ; Xn . This ltration is
alled the natural ltration for fXng.
Denition
A dis
rete time pro
ess fXn : n 0g is
alled a martingale relative to the ltration fF (n)g if
fXn g is fF (n)g-adapted,
IEfjXn jg < 1 for all n, and
IEfXn j F (n 1)g = Xn 1 almost surely, for all n 1.
If we have just in the last
ondition, then fXn g is
alled a supermartingale.
Let us
onvin
e ourselves, before starting the dis
ussion of martingales, that there are plenty of
interesting examples.
A.BENHARI
110
1) Suppose
that X1 ; X2 ; : : : are independent random variables with IEjXn j < 1 and IEXn = 0.
Let Sn = Pnk=1 Xk be the partial sums and F (n) be the natural ltration of the fSng. Observe
that this is also the natural ltration for the fXn g. Then
IEfSn j F
(n 1)g = IE
n 1
nX
k=1
Xk + Xn F (n
1) =
nX1
k=1
Xk + IEXn = Sn 1 :
4) Re
all the denition of the Galton Watson pro
ess fXn g with ospring distribution
(p0 ; p1 ; p2; : : :). There are independent random variables Yk , k 2 A, su
h that
Xn =
Yk :
X
k2T;jkj=n 1
The Yk are independent with the distribution given by the sequen
e (p0 ; p1; : : :) and
Xn 1 = #fk 2 T; jkj = n 1g :
Let F (n) be the -eld on
generated by the Yk with jkj n 1. Note that Xn is F (n)measurable. We see that
IEfXn jF (n 1)g =
IE 1fk2T g Yk F (n 1) =
1fk2T g IE Yk F (n 1) ;
k2A;jkj=n 1
k2A;jkj=n 1
k2A;jkj=n 1
k2A;jkj=n 1
Hen
e fXng is a martingale if and only if the expe
ted number of ospring is
IEfYk g
A.BENHARI
n=0
npn = 1
111
Let us dis
uss some
onsequen
es of the denition. First note that, for every martingale fXn g,
from the tower property of
onditional expe
tation, for all m < n,
IEfXn jF (m)g = IEfIEfXn jF (n 1)g j F (m)g = IEfXn 1 jF (m)g = : : : = Xm :
Taking expe
ted values gives,
IEfXn g = IEfX0 g for all n:
It is immediate from the denition that, for every martingale fXn g,
IEfXn Xn 1 jF (n 1)g = 0 :
(5.1)
Considering fXn g to be the
apital of a gambler at time n, this
an be interpreted as saying
that the game is fair, the expe
ted prot in ea
h step is 0. In the supermartingale
ase this is
0 and the game is unfavourable. Let us explore this interpretation a bit more.
Suppose that fCn : n 1g is your stake on game n. You have to base your de
ision on Cn on
the history of the game up to time n 1. Formally, Cn has to be F (n 1)-measurable. We use
this to dene:
A pro
ess fCn : n 1g is
alled previsible if Cn is F (n 1)-measurable.
Your winnings in game n are then Cn(Xn Xn 1) and the total winnings up to time n are given
by
n
Yn = Ck (Xk Xk 1 ) =: (C X )n :
X
k=1
The pro
ess (C X ) is
alled the martingale transform of X by C , it is the dis
rete analogue of
the sto
hasti
integrals. The big question is now:
an you
hoose fCn g su
h that your expe
ted
total winnings are positive? A positive answer to this question would be the most useful result
of this le
ture, however, we
an prove:
Theorem 10.1 (You
an't beat the system) Let C be a bounded, previsible pro
ess (i.e.
su
h that jCn (!)j C for all n 1 and ! 2
). Then, if fXn g is a martingale, so is
f(C X )n g. Moreover, (C X )0 = 0 and hen
e IEf(C X )g = 0
Proof: It is
lear that (C X )n is integrable (as Cn is bounded) and by denition the pro
ess
fYng = f(C X )n g is adapted to the ltration fF (n)g and starts in 0. We
al
ulate
IEfYn jF
=
This proves it all.
nX1
k=1
(n 1)g = IE
Ck (Xk
nX
k=1
Ck (Xk
Xk 1 ) + Cn IEfXn
Xk 1 )F (n
Xn 1 jF (n
1)
1)g = Yn
1:
A.BENHARI
112
10.2
We now study stopping times for martingales. The intuition is very similar to the stopping
times we have dis
ussed for Brownian motions, but some te
hni
al points are easier.
Suppose that fXn g is a martingale for the ltration fF (n)g on a probability spa
e (
; A; IP).
To avoid te
hni
alities assume that the union of all F (n) generates A. A map T :
! IN [f1g
is
alled a stopping time if
fT ng = f! : T (!) ng 2 F (n) for all n < 1 :
This
orresponds to stri
t stopping times in the
ontinuous time setting, repla
ing fT ng by
fT < ng would
hange the denition, but as in dis
rete time fT < ng = fT n 1g a stopping
time in this weaker sense is a stopping time for the shifted ltration F 0 (n) := F (n + 1) so that
no new theory would evolve from su
h a
hange. However one
an require
fT = ng = f! : T (!) = ng 2 F (n) for all n < 1 ;
and this denition is equivalent to our stopping time denition. This is easy to see, as the rst
denition implies
fT = ng = fT ng n fT n 1g 2 F (n) ;
and the se
ond denition implies
fT ng =
k=0
fT = kg 2 F (n) :
Interpreting fXn g as a game we interpret the stopping times as those instan
es when we
an
quit playing (and obtain our winnings or pay our losses). If we follow that strategy and play
unit stakes up to time T and then quit playing, the stake pro
ess C is
Cn = 1fnT g = 1 1fT n 1g ;
whi
h is F (n 1)-measurable and hen
e fCn g is previsible. The winnings pro
ess is
(C X )n =
k=1
Ck (Xk
Xk 1 ) = XT ^n X0 ;
A.BENHARI
for all n :
113
k=1
the simple random walk. We have seen that this pro
ess is a martingale (with respe
t to the
natural ltration) and we have studied this pro
ess a little in the le
ture Sto
hastis
he Methoden.
Let
T = inf fn 0 : Xn = 1g :
Clearly, T is a stopping time. It is not straightforward to see that
IPfT < 1g = 1 ;
but this follows from the fa
ts proved in Sto
hastis
he Methoden and should be a
epted for
the moment. Our theorem then states that
IEfXT ^n g = IEfX0 g for all n:
However, we have XT = 1 almost surely and hen
e
1 = IEXT 6= IEX0 = 0 :
In some sense this means that you
an beat the system if you have an innite amount of time
and
redit.
After this example one would very mu
h like to see
onditions whi
h make sure that for ni
e
martingales and stopping times we have IEXT = IEX0. This is the
ontent of Doob's optional
stopping theorem.
Theorem 10.3(Doob's optional stopping theorem) Let T be a stopping time and X a
martingale. Then XT is integrable and
IEfXT g
= IEfX0 g ;
(1)
(2)
T is almost surely nite and X is bounded, (i.e. there is a real K su
h that jXn (!)j < K
for all n and !),
(3) IEfT g < 1 and there is K > 0 su h that, for all n and !, jXn(!)
Xn 1 (!)j K .
A.BENHARI
114
Proof:
jXT ^n X0j =
T ^n
k=1
(Xk
Xk 1 ) KT :
By assumption KT is an integrable fun
tion and we
an use dominated
onvergen
e again. For
(4) we use Fatou's lemma to see
IEfXT g = IEfliminf XT ^n g liminf IEfXT ^n g = IEfX0 g :
n!1
n!1
The statement for martingales follows by applying the previous to the supermartingales fXng
and f Xn g separately.
In the situation of our example these
onditions must fail. A glan
e at the third
ondition gives
a striking
orollary:
Corollary 10.4For a simple random walk the expe
ted rst hitting time of level 1 is innite.
10.3
Suppose (a; b) is an interval
ontaining the origin. At whi
h end does Brownian motion leave the
interval? In this se
tion we use martingales to
al
ulate the probability that Brownian motion
leaves su
h an interval at the upper or lower end. We use the strong Markov property to embed
a martingale into Brownian motion. Let us rst state the main result of this se
tion.
Theorem 10.5 Let fB (t) : t 0g be standard Brownian motion and suppose that a < 0 < b
and T = inf ft 0 : B (t) 62 (a; b)g. Then T is an almost surely nite stopping time and
IP
Proof:
B (T ) = a
= b b a and IP B (T ) = b = b
n
a
:
a
The fa
t that T is almost surely nite follows from the fa
t that, almost surely,
p
limsup
jB (t)= tj = 1 ;
t!1
A.BENHARI
115
Then we study the dis
rete time pro
ess Xk = B (Tk ). This pro
ess takes values in the set
D = fl=n : l integer g:
Denote by Yk = Xk Xk 1 the in
rements of the pro
ess. Then, by the strong Markov property, the Yk are independent and identi
ally distributed. The individual distributions are, by
symmetry, given by
1
IPfYk = 1=ng = = IPfYk = 1=ng :
2
In parti
ular, as the expe
ted value is zero, the pro
ess
Xk =
i=1
Yi
(T ) = ag = IPfXS = p=ng = q q p = b b a :
Now look at the stopped martingale fXnS g dened by XnS = XS^n and apply Doob's optional
stopping theorem. Condition (2) is satised be
ause X S is bounded from above by q=n and
from below by p=n and S is almost surely nite. Hen
e,
IEfXS g = IEfXSS g = IEfX0S g = IEfX0 g = 0 :
Abbreviating P = IPfXS = p=ng and Q = IPfXS = q=ng we thus get
0 = np P + nq Q and,
learly, 1 = P + Q :
Solving this system of equations gives the required result for X and hen
e also for Brownian
motion.
It remains to look at irrational values a < 0 < b. This
an be dealt with by approximation. Let
an " a be a sequen
e of rationals in
reasing to a and bn " b be a sequen
e of rationals in
reasing
to b and Tn the exit time from the interval (an; bn). Then
IPfB
The other inequality
an be obtained by approximating a and b from above by rationals in the
same manner. This nishes the proof.
A.BENHARI
116
A similar theorem
an be proved for simple random walk fXng. We
hoose integers a < 0 < b.
What is the probability that fXn g hits a before b? Look at the stopping times
S a = inf n 0 : Xn = a and S b = inf n 0 : Xn = b :
Then show as an exer
ise
b
IPfS a < S b g =
:
b a
We now look at an amusing example of Keeler and Spen
er:
Optimal doubling strategy for ba
kgammon: In our idealization we need a
ontinuous time
fun
tion B (t) : [0; T ! [0; 1, whi
h models your
urrent
han
es of winning. For ba
kgammon
it is best to
hoose Brownian motion fB (t) : t 0g started at 1=2 and stopped upon leaving
the interval [0; 1. At the beginning the stake is, say, one. You and the opponent both have the
right to announ
e a doubling of the stake. If player A does this, player B has the right to either
a
ept this | then the game goes on with doubled stakes and only B has the right to announ
e
the next doubling | or player B gives up and loses the
urrent stake.
A doubling strategy
onsists of two numbers 0 < b < 1=2 < a < 1 and you announ
e a double
if B (t) a and give up if the opponent announ
es a double and B (t) < b. An optimal strategy
a ; b must satisfy:
when B (t) = b a
epting and giving up must have the same expe
ted payo.
Denoting by v your expe
ted winnings from unit stake when you announ
e a doubling at B (t) =
a , this means, by the previous theorem,
1 = ab 2v + a a b ( 2) :
Now,
learly, 1=2 < a 1 b . The expe
ted winnings per unit stake are v = 1 if you announ
e
a doubling at 1 b, at the instant when the opponent starts giving up. If a < 1 b your
expe
ted winnings are lower (otherwise he would give up), so that ne
essarily a = 1 b and
v = 1. Hen
e,
1 = 1 b b 2 + 11 2bb ( 2) :
Solving the system we obtain b = 1=5 and a = 4=5.
Result: If your opponent announ
es a double you should give up if you feel your
han
es of
winning are below 20%. You should announ
e a double if your
han
es are above 80%.
n
10.4
Doob's famous forward
onvergen
e theorem gives a su
ient
ondition for the almost sure
onvergen
e of martingales fXn g to a limiting random variable.
Theorem 10.6 (Martingale Convergen
e Theorem) Let fXn g be a supermartingale, whi
h
is bounded in L1 , i.e. there is K > 0 su
h that IEjXn j K for all n. Then there exists a realvalued random variable X on the same probability spa
e su
h that
lim X = X almost surely.
n!1 n
A.BENHARI
117
Yn = (C X )n =
k=1
Ck (Xk
Xk 1 ) :
A.BENHARI
118
a)
Lemma 10.8Let fXn g be a supermartingale, whi
h is bounded in L1 . Then, for the in
reasing
limit
U [a; b := lim UN [a; b
N !1
we have
(b
Proof:
and observe that if Xn does neither
onverges nor diverges to 1, there are rationals a < b
su
h that event M [a; b takes pla
e. But if M [a; b takes pla
e, then U [a; b = 1, whi
h has
probability zero. Taking the union over the
ountable
olle
tion of rationals a < b we obtain
that almost surely fXn g
onverges to a possibly innite random variable X . But, by Fatou's
lemma,
IEjX j = IEfliminf jXn jg liminf IEjXn j K < 1 ;
n!1
n!1
hen
e jX j < 1 almost surely. This nishes the proof of Doob's martingale
onvergen
e theorem.
Examples. a) Let fXn g be a simple random walk. Then fXn g is a martingale, whi
h does
not
onverge. Namely, by the re
urren
e of simple random walks, both level 0 and level 1 are
rea
hed innitely often,
ontradi
ting
onvergen
e. Clearly, fXn g is not L1 -bounded.
b) If F (0) F (1) : : : is a ltration of a measurable spa
e whose -eld is generated by the
union of the F (n) and X is a nonnegative random variable, then we infer from the martingale
onvergen
e theorem that
lim IEfX jF (n)g
n!1
exists almost surely. It is a natural suspi
ion (and we will show in the next se
tion) that this
limit is almost surely equal to X .
) We now look at the Galton-Watson pro
ess fXn g ospring distribution given by (p0 ; p1 ; : : :).
Assume that the mean ospring number is
1
n=0
npn = 1 :
A.BENHARI
119
Now assume p1 < 1 (in the
ase p1 = 1 trivially Xn = 1 for all n). For every k > 0 we then
have, for any K ,
IPfXn
n=K +1
IPfXn
= k j Xn 1 = kg = 0 ;
hen
e X = 0 almost surely. In other words, the
riti
al Galton-Watson pro
ess be
omes extin
t
in nite time. However, re
all that IEfXn g = 1 for all n. The
onvergen
e in the martingale
onvergen
e theorem does not hold for the expe
ted values.
10.5
The key to the dieren
e between the two examples is the question when the almost sure
onvergen
e in the martingale
onvergen
e theorem
an be repla
ed by L1 -
onvergen
e. We
rst study a general
riterion for L1-
onvergen
e of an arbitrary sequen
e of random variables.
Re
all that a sequen
e fXn g of random variables
onverges in L1 to X i IEjXn X j ! 0. This
also implies that IEfXn g ! IEfX g.
Denition: A sequen
e fXn g of random variables is
alled uniformly integrable if, for every
" > 0 there is K 0 su
h that
jXn j dIP < " for all n :
fjX j>K g
Z
Theorem 10.9 (a) Every bounded sequen
e of random variables is uniformly integrable.
(b) If a sequen
e of random variables is dominated by an integrable, nonnegative random variable Y , i.e. if jXn j Y for all n, then the sequen
e is uniformly integrable.
(
) Let p 1. A sequen
e is Lp -bounded if supn IEfjXn jp g <
Lp -bounded for some p > 1 is uniformly integrable.
1.
Proof:
Clearly, (a) follows from (b), whi
h we now prove. Observe that,
Z
fjXn j>K g
jXnj dIP
fjY j>K g
jY j dIP :
= liminf
1fjY jK gjY j dIP :
K !1
Hen
e limK !1 fjY j>K g jY j dIP = 0 and this implies uniform integrability of our family.
We
ome to (
) and suppose that supn IEfjXn jpg < C for some p > 1. Observe that, for every
K,
IEjY j
fjXn j>K g
A.BENHARI
jXnj dIP K 1
120
fjXn j>K g
jXn jp dIP K 1 pC :
for suitably
hosen K . Finally, suppose that U is uniformly distributed on (0; 1) and let
Xn = n1fU 1=ng 0 :
Then IEjXnj = 1 and hen
e the family is L1-bounded, but
jXn j dIP = 1 for all n > K ;
fjX j>K g
Z
Re
all that every almost surely
onvergent sequen
e also
onverges in probability, whi
h means
that
lim IPfjXn X j > "g = 0 for all " > 0 :
n!1
We
an now state the key theorem of this se
tion.
Theorem 10.10(Uniform Integrability Theorem) Suppose that fXn g is a uniformly integrable sequen
e of random variables, whi
h
onverges in probability to a random variable X . Then
the sequen
e
onverges also in L1 .
8
>
<
>
:
A.BENHARI
121
Note that, by Theorem 5.9(b), the uniform integrability theorem is a generalization of the
dominated
onvergen
e theorem.
We go ba
k to the study of martingales. Let fF (n)g be a ltration and fXng a martingale with
respe
t to this ltration. The next theorem shows that every uniformly integrable martingale is
of the form that data is a
umulated about a random variable X .
Theorem 10.11(Convergen
e for uniformly integrable martingales) Suppose that the
martingale fXn g is uniformly integrable. Then there is an almost surely nite random variable X su
h that
lim X = X almost surely and in L1 :
n!1 n
Moreover, for every n, Xn = IEfX jF (n)g.
Be
ause fXn g is uniformly integrable, it is in parti
ular L1-bounded and thus, by the
martingale
onvergen
e theorem, almost surely
onvergent to a real-valued random variable X .
By the previous theorem this
onvergen
e holds also in the L1 -sense. To
he
k the last assertion,
we verify the two properties of
onditional expe
tation. F (n)-measurability of Xn is
lear by
denition, so let F 2 F (n). For all m n we have, by the martingale property,
Proof:
We let m ! 1. Then
Xm dIP =
Xm dIP
F
Z
F
Z
X dIP
X dIP =
Xn dIP :
jXm X j dIP ! 0 ;
Xn dIP :
The previous theorem shows that every uniformly integrable martingale is of the type that
data about some (hidden) random variable is a
umulated (see example 3). Conversely, every
martingale of this type is uniformly integrable and
onvergent to the hidden variable. This is
the
ontent of Levy's upward theorem.
Theorem 10.12(Levy's upward theorem) Let X be an integrable random variable on
(
; A; P ) and fF (n)g be a ltration su
h that the union of the F (n) generates A. Dene
Xn = IEfX jF (n)g. Then fXn g is a uniformly integrable martingale and
lim
X
n!1 n
A.BENHARI
122
implies
This is possible, sin
e otherwise we
ould nd a sequen
e F (n) 2 A of events with IP(F (n)) < 2 n
and F (n) jX jdIP ". Then look at the event H that innitely many of the events F (n) happen.
By Borel-Cantelli IP(H ) = 0 but at the same time
R
jX j dIP = limsup
1F (n) jX j dIP limsup
n!1
n!1
1F (n) jX j dIP
";
=
This nishes our proof.
fjIEfX jF (n)gj>K g
A.BENHARI
123
IEfX g A
IEfX g A
Y dIP :
Now F := F (n) is a \-stable system whi
h generates A. If A 2 F (n), then, for all m n,
1
1
P (A) =
X dIP =
Xm dIP :
S
IEfX g A
Also,
IEfX g A
1
1
lim
X
d
IP =
lim Xm dIP = Q(A) ;
m
m!1 IEfX g A
IEfX g A m!1
where the rst equality follows, as before, from L1 -
onvergen
e. The uniqueness theorem now
yields P (A) = Q(A) for all A 2 A. Let A = fX > Y g 2 A. Then IP(A) > 0 would imply,
by denition, P (A) > Q(A), whi
h is a
ontradi
tion. Hen
e, IP(A) = 0, whi
h means X Y
almost surely. In the same manner one
an show X Y almost surely, and this nishes the
proof of the Upward Theorem.
Z
G ( 1) :=
k=0
G ( k) G ( n) G ( 2) G ( 1) :
X
Then
= IEfX jG ( n)g :
n!1
= IE X G ((n 1 N ) ^ ( 1)) = Yn 1 ;
this is indeed a martingale. We obtain, from the up
rossing lemma,
(b a)IEfUN [a; bg IEf(X 1 a) g :
Letting N ! 1 shows that, almost surely, the total number of down
rossings of [a; b by the
pro
ess fX n g is nite. Now one has this simultaneously for all rationals a < b and we
an
argue for fX n g as in the martingale
onvergen
e theorem. Hen
e, lim X n = X 1 exists almost
A.BENHARI
124
surely. By Lemma 5.13 the sequen
e is even uniformly integrable and hen
e
onvergen
e holds in
L1 . To see that X 1 = IEfX jG ( 1)g, rst observe that, for all m and G 2 G ( 1) G ( m),
Z
X dIP =
m dIP ;
We rst show that Mn = Xn=n denes a martingale. To see this re
all that there are independent random variables Yk , k 2 A, su
h that
Xn =
1fk2T g Yk :
X
k2A;jkj=n 1
The Yk are independent with the distribution given by the sequen
e (p0 ; p1; : : :) and
Xn 1 = #fk 2 T; jkj = n 1g :
Let F (n) be the -eld on
generated by the Yk with jkj n 1. Note that Xn is F (n)measurable. We see that
IE
Xn
F (n
n
1) =
X
nXn 1 IEfYk g = nn 11 :
k2A;jkj=n 1
X
k2A;jkj=n 1
(n 1)
1)
IE 1fk2T g Yk F
1fk2T g IE
Yk F (n
We want to show that IPfM > 0g > 0. For this purpose we show that fMng is uniformly
integrable. If this holds we have
IEM = lim IEMn = 1 ;
n!1
A.BENHARI
125
hen
e M 0
annot be identi
ally zero so that IPfM > 0g > 0. To
he
k uniform integrability
it su
es to
he
k L2-boundedness (see Theorem 5.9
). We have
IEfMn2 jF (n 1)g = IEfMn2 1 jF (n 1)g + IEf2Mn 1 (Mn Mn 1 )jF (n 1)g
+ IEf(Mn Mn 1)2 jF (n 1)g
= Mn2 1 + IEf(Mn Mn 1)2 jF (n 1)g :
To
ompute the se
ond term observe
IEf(Mn Mn 1 )2 jF (n 1)g = 2n IEf(Xn Xn 1 )2 jF (n 1)g
and that, on fXn 1 = N g,
IE
(Xn
Xn 1 )2 jF (n
1) = IE
n
k2T;jkj=n 1
N
F (n 1) = N2 = Xn 12 :
IEfMn2 g
1 IEfMn
IEfMn2 g
Yk
= IEfMn2 1 g + n+1 ;
1
X
1
IEfMn2 g 1 + 2
k+1 < 1 ;
n=1
martingale fMng is L2-bounded. Altogether
Xn = Mn n
2 ;
so that if M > 0 the Galton Watson pro
ess in
reases exponentially and the event fM > 0g has
positive probability.
10.6
In this se
tion we give a very short proof of Kolmogorov's strong law of large numbers under
minimal moment
onditions. Martingale theory will serve as the major tool.
Theorem 10.15Suppose that fXn g is a sequen
e of independent and identi
ally distributed integrable random variables. Then
lim 1
n!1 n
k=1
A.BENHARI
126
Proof: Abbreviate Sn = nk=1 Xk . Let G ( n) be the -eld generated by the random variables
Sn ; Sn+1 ; Sn+2 ; : : : and G ( 1) the interse
tion of all these elds. From the exer
ises we know
P
that
IEfX1 jG
k=1
Now the limit is measurable with respe
t to the tail -eld T of the sequen
e fXng and, as this
has only events of probability zero or one (by Kolmogorov's 0-1-law), we have IEfX1 jG ( 1)g =
almost surely for some
onstant value . Be
ause we have L1-
onvergen
e we also have
= IE
A.BENHARI
Sk
IEfSk g
lim
=
lim
= IEXn for all n;
k!1 k
k!1 k
o
127
Suppose we have a high pre
ision ma
hine whi
h produ
es units of given length. Many very
small ee
ts
hange the adjustment of the ma
hine slightly, so that after some time the small
perturbations have summed up and the ma
hine needs readjustment. Indeed ea
h of these small
perturbations is again the sum of many small ee
ts, so that the total perturbation after time
t is
[nt
P (t) = Xk
X
k=1
where Xk are the independent perturbations o
urring in the time interval [k=n; (k + 1)=n)
satisfying IEfXk g = 0 and varian
e of order 1=n. If we
onsider in
reasingly small ee
ts, we
in
rease the n and the question is whether this leads to
onvergen
e of the pro
ess fP (t)g to a
limit pro
ess, whi
h would then be a good model for the perturbing pro
ess. This is the problem
of a fun
tional
entral limit theorem, whi
h we address in this
hapter. The main result is that
all random walks whose in
rements have nite varian
e
an be res
aled so that they
onverge
in distribution to Brownian motion.
We start with a proper denition of the
onvergen
e of distributions on fun
tion (and more
general metri
) spa
es.
11.1 Convergen
e of distributions on metri
spa
es
A.BENHARI
128
shape of the distributions of Xn for large n is like the shape of the distribution of X . Sample
values from Xn allow no inferen
e towards sample values from X and, indeed, there is no need
to dene Xn and X on the same probability spa
e. In fa
t,
onvergen
e in distribution is only
related to the
onvergen
e of the distributions of the random variables and not to the random
variables themselves.
We start by giving a denition of
onvergen
e in distributions for random variables in metri
spa
es, explore some of its properties and then show that the
on
ept of
onvergen
e in distribution for real-valued random variables known from the previous
ourse is
onsistent with our
denition.
Denition: Suppose (M; d) is a metri
spa
e and A the Borel--eld on M . Suppose that Xn
and X are M -valued random variables. Then we say that Xn
onverges in distribution to X , if,
for every bounded
ontinuous g : M ! IR,
lim IEfg(Xn )g = IEfg(X )g:
n!1
We write Xn ) X for
onvergen
e in distribution.
Warning: If M = IR and Xn ) X this does not imply that IEfXn g
onverges to IEfX g. Note
that g(x) = x is not a bounded fun
tion on IR.
Here is an alternative approa
h, whi
h shows that
onvergen
e in distribution is in fa
t a
onvergen
e of the distributions. The statement of the following proposition is trivial.
Proposition 11.1Let Prob(M ) be the set of probability measures on (M; A). A sequen
e fPn g
Prob(M )
onverges weakly to a limit P 2 Prob if, for every
ontinuous, bounded fun
tion g :
M ! IR,
Z
Z
lim g dPn = g dP :
n!1
Then the limit of a
onvergent sequen
e is uniquely determined. Suppose that Xn and X are M valued random variables. Then Xn
onverges in distribution to X , if and only if the distributions
of Xn
onverge weakly to the distribution of X .
Proof: Only the uniqueness of the limit needs proof. If P and Q are two limits of the same
sequen
e, then f dP = f dQ for all bounded
ontinuous f : M ! IR. For every open set
G M we may
hoose an in
reasing sequen
e fn(x) = nd(x; G
) ^ 1 of
ontinuous fun
tions
onverging to 1G and infer from monotone
onvergen
e that P (G) = Q(G). Now P = Q follows
from the Uniqueness Theorem.
R
In
omplete, separable metri
spa
es M weak
onvergen
e stems from a suitably dened metri
on the spa
e Prob(M ). The
ase M = IR will be dis
ussed as an exer
ise.
Examples:
A.BENHARI
129
Theorem 11.2 Suppose a sequen
e fXn g of random variables
onverges almost surely to a random variable X (of
ourse, all on the same probability spa
e). Then Xn
onverges in distribution
to X .
Suppose g is bounded and
ontinuous. The g(Xn )
onverges almost surely to g(X ).
As the sequen
e is bounded it is also uniformly integrable, hen
e
onvergen
e holds also in the
L1 -sense and this implies
onvergen
e of the expe
tations, i.e. IEfg(Xn )g ! IEfg(X )g.
Proof:
(iii) For all open sets G M , liminf n!1 IPfXn 2 Gg IPfX 2 Gg.
(iv) For all Borel sets A M with IPfX 2 Ag = 0, limn!1 IPfXn 2 Ag = IPfX 2 Ag.
(v) For all bounded measurable fun
tions g : M
have IEfg(Xn )g ! IEfg(X )g.
Proof:
(i))(ii)
k!1
k!1
Let n ! 1. The integrand on the right hand side is bounded by 1 and
onverges pointwise
and hen
e in the L1 -sense to 1K (X ).
(ii))(iii) Follows from 1G = 1 1K for the
losed set K = G
.
(iii))(iv) Let G be the interior and K the
losure of A. Then, by assumption, IPfX 2 Gg =
IPfX 2 K g = IPfX 2 Ag and we may use (iii) and (ii) (whi
h follows immediately from
(iii)) to get
limsup
IPfXn 2 Ag limsup IPfXn 2 K g IPfX 2 K g = IPfX 2 Ag;
n!1
n!1
liminf
IPfXn 2 Ag liminf IPfXn 2 Gg IPfX 2 Gg = IPfX 2 Ag:
n!1
n!1
From (iv) we infer that the
onvergen
e holds for g of the form g(x) = Nn=1 an1An
where An satises IPfX 2 Ang = 0. Let us
all su
h fun
tions elementary. Given g as in
(v) we observe that for every a < b with possibly a
ountable set of ex
eptions
IP X 2 fx : g (x) 2 (a; bg = 0 :
(iv))(v)
A.BENHARI
130
and
liminf
IEfg (Xn )g liminf IEfgk (Xn )g = IEfgk (X )g :
n!1
n!1
and the right sides
onverge, as k ! 1, by bounded
onvergen
e, to IEfg(X )g.
(v))(i) This is trivial.
To remember the dire
tions of the inequalities in the Portmanteau Theorem it is useful to re
all
the last example Xn = 1=n ! 0 and
hoose G = (0; 1) and K = f0g to obtain
ases where the
opposite inequalities fail.
The notion of
onvergen
e in distribution for real valued random variables is already known
from the previous
ourse. We
an now see, that the
on
epts agree for the
ase of real random
variables.
Theorem 11.4(Helly-Bray Theorem) Let Xn and X be real valued random variables and
dene the asso
iated distribution fun
tions Fn (x) = IPfXn xg and F (x) = IPfX xg. Then
the following assertions are equivalent.
(a) Xn
onverges in distribution to X ,
(b) nlim
!1 Fn (x) = F (x) for all x su
h that F is
ontinuous in x.
Proof:
(a))(b)
liminf
IPfXn 2 Gg
n!1
k=1
liminf
IPfXn 2 Ik g =
n!1
A.BENHARI
131
k=1
IPfX
2 Ik g ;
We nish this se
tion with an easy observation, whi
h follows dire
tly from the denition:
Lemma 11.5 If Xn ) X and g : M ! IR is
ontinuous, then g(Xn ) ) g(X ).
11.2 The Donsker Invarian
e Prin
iple: Statement and Appli
ations
Let fXn g be a sequen
e of independent and identi
ally distributed random variables and assume
that they are normalized, so that IEfXn g = 0 and Var(Xn ) = 1. This assumption is no loss of
generality, be
ause if Xn has nite varian
e we
an always
onsider the normalization
Xpn
IEfXn g
Sn =
Var(Xn ) :
We look at the random walk generated by the sequen
e
n
k=1
Xk ;
The theorem is also
alled fun
tional
entral limit theorem. Before dis
ussing the proof, we
shall dis
uss
onsequen
es and appli
ations of the theorem. Using the normalization indi
ated
before one obtains that fXn g is any sequen
e of independent and identi
ally distributed random
variables with expe
tation and nite, positive varian
e 2, then fS ng
onverges in distribution
to a Brownian motion with drift parameter and diusion parameter 2.
As a rst appli
ation we prove a
entral limit theorem with minimal moment
onditions.
Theorem 11.7(Central Limit Theorem) Suppose that fXk g is a sequen
e of independent,
identi
ally distributed random variables with IEfXk g = 0 and Var(Xk ) = 1. Then
p1
n k=1
Xk ) X ;
A.BENHARI
132
Proof:
Consider the
ontinuous fun
tion g : C [0; 1 ! IR dened by g(f ) = f (1). Then
n
p1 X = g(S n )
X
n k=1
and g(B ) is standard normally distributed. Hen
e the statement follows from Donsker's theorem
by means of Lemma 6.5.
The next two theorems give examples why the name invarian
e prin
iple is justied. The limits
we obtain are invariant under (i.e. do not depend on) the
hoi
e of the exa
t distribution of the
random variables Xn. A spe
ial
ase of interest is IPfXn = 1g = 1=2 = IPfXn = 1g in whi
h
ase the asso
iated random walk is the symmetri
random walk.
Theorem 11.8 Suppose that fXk g is a sequen
e of independent, identi
ally distributed random
variables with IEfXk g = 0 and Var(Xk ) = 1. Let fSn g be the asso
iated random walk and
Mn = maxfSk
: 0 k ng
p
2
lim
IPfMn x ng = p
n!1
2
1
x
y2 =2 dy :
Then, by denition,
IE
G(S n )
= IE
and
IE
n
g 0max
t1
G(B )
n
o
S (tn) o
pn = IE g max0pknn Sk ;
n
= IE
n
o
g 0max
B (t) :
t1
o
n
o
M
pnn = IE g 0max
B
(
t
)
:
t1
From the Portmanteau Theorem and the re
e
tion prin
iple we infer
lim IPfMn xpng = IPf0max
B (t) xg = 2IPfB (1) xg ;
n!1
t1
and the latter probability is the given integral.
A.BENHARI
133
Theorem 11.9(Ar
-sine law for the last sign-
hange) Suppose that fXk g is a sequen
e of
independent, identi
ally distributed random variables with IEfXk g = 0 and Var(Xk ) = 1. Let
fSng be the asso
iated random walk and
Nn = maxf1 k n : Sk Sk 1 0g
the last time the random walk hanges its sign before time n. Then Nn =n onverges in distribution to a random variable with density
1
x(1
x)
p
2
lim
IPfNn xng = ar
sin( x) :
n!1
Remark: Note that this is surprising: the probability is high that Nn =n is near 0 or 1.
Proof: Step 1: Dene a bounded fun
tion g on C [0; 1 by
g(f ) = maxft 1 : f (t) = 0g:
It is
lear that g(S n ) diers from Nn=n by a term, whi
h is bounded by 1=n and hen
e vanishes
asymptoti
ally. Hen
e Donsker would imply
onvergen
e of Nn=n in distribution to
g(B ) = supft 1 : B (t) = 0g
if g was
ontinuous. g is not
ontinuous, but we shall see that g is
ontinuous on the set C of all
f 2 C [0; 1 su
h that f takes positive and negative values in every neighbourhood of every zero
and f (1) 6= 0. We shall also see that Brownian motion is almost surely in C . From property (v)
in the Portmanteau Theorem we
an infer that, for every
ontinuous bounded h : IR ! IR,
lim IE
n!1
n
n
o
n
o
n
o
Nn o
=
lim
IE h g (S n ) = IE h g (B ) = IE h(supft 1 : B (t) = 0g) :
n!1
n
Step 2: g is
ontinuous on C .
Let " > 0 is given and f 2 C . Let
0 =
min jf (t)j ;
t2[g(f )+";1
A.BENHARI
134
that fB (t + tq ) : t 0g is a Brownian motion and by Theorem 4.9, almost surely, there exist
positive and negative values in every small interval to the right of tq . Taking the union over all
rationals gives that, almost surely, all the zeroes of the form tq have positive and negative values
in every neighbourhood. But if t is a zero whi
h is not equal to tq for any rational q there exists,
for every rational q < t a zero tq in the interval [q; t), hen
e t has positive and negative values
in every small interval to its left.
Step 4: Cal
ulate the distribution of the random variable L = supft 1 : B (t) = 0g.
This will be done with the help of the Markov property and the re
e
tion prin
iple. Write
Ta = inf ft > 0 : B (t) = ag ;
whi
h is a stopping time. We get the distribution of Ta from the re
e
tion prin
iple, let a 0,
1 1
p exp( x2=2t) dx ;
IPfTa tg = IPf sup B (s) ag = 2IPfB (t) ag = 2
2t
a
0st
Z
1
Z
sr3 0
p1 3 r rs
dr
1 s sr + s
Z 1
(r + s)2 1=2 s dr;
1 s
rs
(r + s)2
and substitute t = s=(r + s) with dt = s=(r + s)2dr to see that this equals
= 1
= 1
0
1 dt = 2 ar
sin(px) :
t(1 t)
For the proof of the Donsker invarian
e prin
iple we have essentially two possibilities:
Suppose that a subsequen
e of fS ng
onverges in distribution to a limit X . This limit
is a
ontinuous random fun
tion, whi
h is easily seen to have stationary, independent
in
rements. Hen
e it is a Brownian motion and one
an
he
k that it must have drift 0
and varian
e 1. So Brownian motion is the only possible limit point of the sequen
e fS ng.
The di
ult part of this proof is to show that every subsequen
e of fS n g has a
onvergent
subsubsequen
e (the tightness property).
A.BENHARI
135
We will follow the idea behind Theorem 6.2 and
onstru
t the random variables
X1 ; X2 ; X3 ; : : : on the same probability spa
e as the Brownian motion in su
h a way that
fS n g is with high probability
lose to a Brownian motion. This is
alled embedding the
random walk in to Brownian motion and
an be done with the famous embedding theorem
of Skorokhod (1965).
As the rst approa
h is more measure theoreti
, we will prefer the se
ond approa
h, whi
h
represents a typi
al te
hnique in probability theory. The proof uses almost all the te
hniques
we have developed in the le
ture so far.
11.3 The Skorokhod Embedding Theorem
Theorem11.10 (Skorokhod Embedding Theorem) Suppose that X is a real valued random
variable with IEfX g = 0 and IEfX 2 g < 1. Then X and a Brownian motion fB (t) : t 0g
an be dened on a joint probability spa
e su
h that there exists a stopping time T su
h that
IEfT g = IEfX 2 g < 1 and, almost surely, B (T ) = X .
Examples:
A.BENHARI
136
'(x) dP (x)
+
=
Z
Z
( u) dP (u)
1
Z
'(u) dP (u)
dP (v)
Z
'(v) dP (v)
1
Z
dP (u)(v
v dP (v)
dP (u)(v'(u) u'(v)):
v
u)
v u
'(u) +
o
u
'(v) :
v u
Note that the expression in the
urly bra
kets is IEf'(B (T ))g, where T is the rst exit time of
the interval (u; v). We now let (
1; A1; IP1) be a probability spa
e on whi
h a pair (U; V ) of
random variables
an be dened with U 0 and V 0 and distribution given by
IP1 f(U; V ) = (0; 0)g = P fX = 0g
and, for Borel sets A ( 1; 0) (0; 1),
1
IP1 f(U; V ) 2 Ag =
dP (u) dP (v)(v u) :
ZZ
(u;v)2A
Note that this properly denes the distribution of the random ve
tor (U; V ). We formally dene
probability measures u;v on fu; vg by 0;0 f0g = 1 and
and u;v fvg = v uu for u < 0 < v:
With this denition the key formula
an be written as follows. For all bounded measurable
fun
tions ',
'(x) dP (x) = IE1 '(x)U;V (dx) :
Now dene the produ
t spa
e
(
; A; IP) = (
1; A1 ; IP1)
(C [0; 1); A0 ; IP0 )
of
1 and Wiener spa
e. Observe that (U; V ) and fB (t) : t 0g are independent random
variables on this spa
e. Dene
T = inf ft 0 : B (t) 62 (U; V )g:
This
an be seen either as a stopping time with respe
t to F + (t) A
onditional on U = u and
V = v or, on the whole spa
e, as a stopping time with respe
t to the ltration F (t) = A
F + (t).
Our key formula gives (denoting expe
tations with respe
t to IP0 by IE0 and expe
tations with
respe
t to IP1 by IE1), by Fubini's Theorem,
u;v fug =
v
v u
IEf'
A.BENHARI
nZ
nZ
'(x)U;V (dx)
137
Hen
e B (T ) and X have the same distribution. It remains to show that IEfT g < 1. First re
all
that
IEfT g = IE1 fIE0 fT gg = IE1 f UV g:
Now we use the distribution of (U; V ), using the alternative expressions for
= 01 v dP (v) =
0 ( u) dP (u),
1
1 0 dP (u)( u) 1 dP (v) v(v u)
IE1 f UV g =
R
=
=
=
dP (u)( u)
u+
dP (v)
1 2
v dP (v)
( u)2 dP (u) + 0
E f(X )2 g + E f(X + )2 g
1
v2 o
E fX 2 g:
Note that we had to enlarge Wiener spa
e in order to dene T . This
an be avoided (Dubin's
For the purpose of our proof the Skorokhod representation is satisfa
tory, it
allows, by means of the Markov property, to embed the whole random walk in the Brownian
motion.
stopping rule).
Corollary11.11 Suppose that X1 ; X2 ; X3 ; : : : are independent, identi
ally distributed real valued
random variables with mean zero and varian
e 1. Then there exists a sequen
e of stopping
times 0 = T0 T1 T2 T3 : : : su
h that the in
rements Tn Tn 1 are independent,
identi
ally distributed, IEfTn g = n < 1 and, almost surely, the sequen
e fB (Tn ) : n 1g has
the distribution of the random walk fSn g asso
iated with fXn g.
Proof: We dene the spa
e (
1 ; A1 ; IP1 ) su
h that a sequen
e of independent variables (Un ; Vn )
an be dened for Xn as in the proof before. Then we
an dene
T1 = inf ft 0 : B (t) 62 (U1 ; V1 )g
and obtain B (T1) = X1 in distribution and IEfT1 g = 1. By the strong Markov property
fB2(t) : t 0g = fB (T1 + t) B (T1 ) : t 0g
is again a Brownian motion and independent of F (T ) and, in parti
ular, of (T1 ; B (T1 )). Hen
e
we
an dene a stopping time
T2 = T1 + inf ft 0 : B2 (t) 62 (U2 ; V2 )g
and observe that T2 T1 is independent of T1 with the same distribution and B (T2 ) B (T1) =
B2 (T2 T1 ) has the same distribution at X2 and is independent of X1 . Furthermore IEfT2 g = 2.
We
an pro
eed indu
tively to get the
orollary.
We have thus embedded the random walk fSn g into Brownian motion. In the next se
tion we
use this to prove Donsker's Theorem.
A.BENHARI
138
From the s
aling property of Brownian motion we
an see that all the random fun
tions W n are
standard Brownian motions. We show that, for all " > 0,
lim IP kW n S nksup > "g = 0 :
(6.1)
n!1
Let us rst see why (6.1) implies the theorem. Suppose that K C [0; 1 is
losed and dene
K [" = ff 2 C [0; 1 : kf gksup " for some g 2 K g:
Then
IPfS n 2 K g IPfW n 2 K ["g + IPfkS n W n ksup > " :
By (6.1) the se
ond term goes to 0 and the rst term is equal to IPfB 2 K ["g for a Brownian
motion B , independently of n. As K is
losed we have
lim IPfB 2 K ["g = IP B 2 K [" = IPfB 2 K g:
"!0
n
">0
Altogether,
limsup
IPfS n 2 K g IPfB 2 K g ;
n!1
whi
h is
ondition (iii) in the Portmanteau Theorem. It just remains to prove (6.1).
Let An be the event that there exists t 2 [0; 1 su
h that jS n(t) W n(t)j > ". We have to show
IP(An ) ! 0. Let k = k (t) be the unique integer with (k 1)=n t < k=n. Be
ause S n is linear
on su
h an interval we have
S
S
An 9t : pk W n (t) > " [ 9t : pk 1 W n (t) > " :
n
n
9t : W n Tnk
[ 9t : W n Tkn 1
n
For given 0 < < 1 the event on the right implies that either
9s; t 2 [0; 2 : js tj < ; jW n(s) W n(t)j > "
or
9t 2 [0; 1 : jTk =n tj _ jTk 1=n tj :
n
A.BENHARI
139
Note that the probability of the rst event does not depend on n. Choosing > 0 small, we
an
make the rst probability as small as we wish, sin
e Brownian motion is uniformly
ontinuous
on [0; 2. It remains to show that for arbitrary, xed > 0,
lim IP 9t : jTk =n tj _ jTk 1=n tj = 0 :
(6.2)
n!1
To prove this we use that
n
lim Tn = lim 1 (T T ) = 1 almost surely.
n
n!1
n!1 n
k=1
k 1
This is Kolmogorov's law of large numbers for the sequen
e fTk Tk 1g of independent identi
ally
distributed random variables with mean 1. Observe that for every sequen
e fan g of reals one
has
lim an = 1 ) nlim
jak kj=n = 0 :
n!1 n
!1 0sup
kn
This is a matter of plain (deterministi
) arithmeti
. Hen
e we have,
lim IP sup jTk n kj = 0 :
n!1
n
(6.3)
0kn
Now re
all that t 2 [(k 1)=n; k=n and let n > 2=. Then
IP 9t : jTk =n tj _ jTk 1 =n tj
IP sup (Tk (k 1))n _ (k Tk 1)
0kn
IP sup Tk k =2 + IP sup (k 1)
n
0kn
1kn
Tk 1
=2 ;
whi h by (6.3) onverges to 0. Hen e (6.2) and Donsker's invarian e prin iple are proved.
A.BENHARI
140
Looking ba
k through the material of this
ourse one
an see that the emphasis was on
ontinuous
time pro
esses, and in parti
ular Brownian motion, rather than dis
rete time pro
esses. The
only notable ex
eption was the
hapter on martingales, where we were dealing ex
lusively with
dis
rete time martingales (and, in fa
t, some extra eort was needed to apply the results to
Brownian motion). It is quite easy to dene martingales also in
ontinuous time.
Denition: Let (
; A; IP) be a probability spa
e and fF (t) : t 0g a
olle
tion of sub--elds
with F (s) F (t) for all s t, in other words a ltration. A pro
ess fX (t) : t 0g is
alled a
martingale if
IE X (t) F (s) = X (s) almost surely, for all s t:
Choose the ltration F (t) = F +(t) and use the weak Markov property, to see
IE X (t) F (s) = IEX (s) fB (t s)g = X (s) ;
for all s t.
Proof:
The rst major step in the dis
rete time martingale theory and the key step to all later theorems
was the You-
an't-beat-the-system-Theorem. Finding a
ontinuous time analogue of this is a
major problem for us. Re
all that we were using a bounded, previsible pro
ess fCn g, to dene
the martingale transform
n
(C X )n = Ck (Xk Xk 1):
X
k=1
Cn , the stake at the game Xn Xn 1 , had to be measurable with respe
t to F (n 1). All
major theorems about martingales were derived from the fa
t that the martingale transform
again denes a martingale. Re
all that (C X )k
an be interpreted as our total prot at time
k if Cn is the number of items of a good we possess at time n and fXn g is the pri
e pro
ess of
one item of the good. It is
lear that su
h a pro
ess would be of great pra
ti
al and theoreti
al
interest in a
ontinuous time setting, however it is un
lear how it
ould be dened and also what
the appropriate
ondition for the stake pro
ess fC (t)g
ould be.
A.BENHARI
141
Instead of solving this problem now, we rst try a pathwise approa
h for integration with respe
t
to a standard Brownian motion fB (t) : t 2 [0; 1g. The most natural idea would be to dene a
Stieltjes integral
t
f (s) dB (s):
0
This would mean we let, for n 2,
n = f0 = tn0 < tn1 < tn2 < < tnn = 1g
be a
olle
tion of partitions with n n+1 su
h that the mesh of the partition
n n
n = max
(t tni 1)
i=1 i
Z
k=1
onverges almost surely to a limit, whi
h would be a reasonable generalization of the martingale
transform. However, we shall see that su
h a limit may fail to exist.
Theorem12.2 Suppose fn g is a sequen
e of partitions as above with mesh n ! 0. Then,
almost surely, there exists a measurable,
ontinuous fun
tion f : [0; 1 ! [ 1; 1 su
h that
limsup
n!1
k=1
To prove this, we rst prove a positive result of independent interest. We show that Brownian
motion has nite quadrati
variation.
Theorem12.3(Quadrati
variation) Suppose fn g is a sequen
e of partitions
n!1
k=1
(B (tnk)
Proof:
A.BENHARI
n X
k=1
(B (tnk)
B (tnk 1))2
(t
142
o
s)
= 0:
Using the independen
e of the in
rements of Brownian motion, one
an get that,
IE
n X
(B (tnk)
k=1
=
=
=
IE
n X
B (tnk 1))2
(B (tnk)
k=1 l=1
n n
X
IE
k=1
n n
X
IE
k=1
s)
B (tnk 1))2
k=1
n n
IE
(B (tnk)
XX
o
(t
o
2(t
s)IE
nX
(B (tnk)
k=1
(B (tnk)
B (tnk 1 ))4
(B (tnk)
B (tnk 1 ))4 :
XX
k=1 ll6==1k
(tnk
tnk 1)(tnl
B (tnk 1 ))2
(t
tnl 1 )
+ (t
s)2
s)2
(t
s)2
For every normally distributed random variable X with = 0 we have, by partial integration,
2 1 x4 exp( x2=22 ) dx
IEfX 4 g = p
22 01
= p 2 2 0 3x2 2 exp( x2=22 ) dx
2
= 32 IEfX 2 g = 3IEfX 2 g2 :
We infer that
Z
k=1
IE
(B ( )
tnk
B(
)) = 3 (tnk
tnk 1 4
k=1
tnk 1 )2 3n (t s) ! 0 ;
whi
h proves the L2 -
onvergen
e. It is
lear that this implies the existen
e of an almost surely
onvergent subsequen
e.
fng is a sequen
e of partitions
n = fs = tn0 < tn1 < tn2 < < tnn = tg
of a nondegenerate interval [s; t with n n+1 and mesh n ! 0. Then, almost surely,
Corollary12.4(Unbounded variation) Suppose
limsup
n!1
k=1
Proof: By the Holder property we
an nd, for any 2 (0; 1=2), an n su
h that jB (a) B (b)j
ja bj for all a; b 2 [s; t with ja bj n . Using the quadrati
variation of Brownian motion
A.BENHARI
k=1
1
jB (tnk) B (tnk 1)j limsup
n!1
(B (tnk)
n k=1
143
(t s)
B (tnk 1))2 nlim
!1 n
= 1:
n=1 n is
ountable and let
1
be a set with IP(
1 ) = 1 su
h that the statement of the
orollary
holds simultaneously for the partitions indu
ed by n on all intervals bounded by a xed pair
of partition points. We have to
onstru
t the fun
tion f so that
limsup
n!1
l=1
and jB (tnn) B (tnn 1)j < 1=2. Then dene f to be
onstant 1 or 1 on [tnl 1; tnl ), a
ording to
the sign of the in
rement B (tnl) B (tnl 1). Then
n
l=1
nX1
l=1
if jf (tnn 1)j 1. Now
hoose n = n(1) so large that, in the remaining interval I = [tnn(0)
(0)
n
l=m
n(0)
1 ; tn(0) ,
n
n
where m is the index with tnm 1 = tnn(0)
(0) 1 , and jB (tn ) B (tn 1 )j < 1=2. On ea
h interval
[tnl 1 ; tnl), for m l < n,
hoose f
onstant equal to 1=2 or 1=2 a
ording to the sign of the
in
rement. Pro
eed like this indu
tively, always rening until the variation in the remaining
interval ex
eeds 2k + 1=2 and
hoosing the value of f on the partition sets from 2 k . This
denes a measurable fun
tion f , bounded by 1, be
ause after k steps the values of f are all
determined up to distan
e 2 k in the sup-norm. Also, for every k the sum
n
l=1
A.BENHARI
144
motion is of unbounded variation destroys many typi
al properties of the integral | most notably the
hange of variables formula. But they are repla
ed by new ones, so that the sto
hasti
integrals provides a new type of
al
ulus, the sto
hasti
al
ulus. In the next
ourse we shall
present the
ornerstones of this theory. Sto
hasti
al
ulus will enable us to obtain deeper
insight in the properties of Brownian motion and also dene interesting new pro
esses, the diffusions. Sto
hasti
al
ulus allows elegant proofs of nontrivial probabilisti
theorems and opens
the door to some nontrivial appli
ations, for example in the mathemati
s of nan
ial markets
or sto
hasti
dierential equations.
A.BENHARI
145
13.Exercises
1st Question:
2nd
3rd Question:
A.BENHARI
146
5th
6th
Chara
terize
(a) those sets
, for whi
h X and Y are independent,
(b) those sets
, for whi
h X and Y are independent and identi
ally distributed.
Question: Suppose that X : (
1
2 ; A1
A2 ) ! (
0 ; A0 ) is measurable. Then, for all
!1 2
1 , the mappings
X! : (
2 ; A2 ) ! (
0 ; A0 );
X! (!2 ) = X (!1 ; !2 ) ;
and, for all !2 2
2, the mappings
X ! : (
1 ; A1 ) ! (
0 ; A0 );
X ! (!1 ) = X (!1 ; !2 ) ;
are measurable.
Hint: First redu
e the problem to the
ase that X is the indi
ator fun
tion of E 2 A1
A2.
Question: Suppose that X is a random point uniformly
hosen from the sphere
S 2 (r) = f(x; y; z ) 2 IR3 : x2 + y2 + z 2 = r2 g
and P 2 IR3 n S 2(r) a xed point. Dene a random variable Z as the inverse of the
Eu
lidean distan
e of P and X . Find IEZ and limP !1 kP kIEZ and interpret the results.
Hint: By rotation one
an assume that P = (0; 0; p) for p 6= r. Re
all from higher dimensional
al
ulus that the
ontribution of the upper hemisphere S+2 (r) parametrized by some
f : A ! S+2 (r) is given by
d`(y)
1 dIPX (x) = 1
det
Df (y)T Df (y)
2
4r A
kP f (y)k :
f (A) kP xk
1
8th Question:
and let G =
n=1
A.BENHARI
147
IP
X1 2 A1 ; : : : ; Xn 2 An
j =1
IP
Xj 2 Aj :
(a) Show
that on a produ
t spa
e
= 1
i equipped with a produ
t measure IP =
1 Pi the sequen
e of proje
tions Xji=1
(!1; !2; : : :) = !j is independent.
i=1
(b) For all disjoint sets A IN and B IN dene the random variables XA = (Xa : a 2 A)
and XB = (Xb : b 2 B ) with values in the produ
t spa
es (
A; AA) and (
B ; AB ),
where
A = a2A
a and AA = a2A Aa . Show that XA and XB are independent.
(
) Give an example of a sequen
e X1; X2 ; : : : su
h that for ea
h i 6= j the random
variables Xi and Xj are independent, but the sequen
e is not independent.
10th Question: Suppose fXn g is a Galton-Watson pro
ess with ospring distribution given by
the sequen
e (p0; p1 ; : : :). Dene a fun
tion
Q
G : [0; 1 ! [0; 1 ;
G(x) =
n=0
pn xn :
A.BENHARI
148
Suppose that fX (t) : t 0g is a sto
hasti
pro
ess modelling the number of
ustomers in a shop. fX (t)g should satisfy the following assumptions:
The number of
ustomers arriving during one time interval does not ae
t the number
of
ustomers arriving in another disjoint time interval. Mathemati
ally, this means
that the pro
ess has independent in
rements.
The rate at whi
h
ustomers arrive should be
onstant, more pre
isely, there is some
0 su
h that IE[X (t) = t.
Customers arrive one at a time. To make this pre
ise we assume that X (t) takes
values in IN, is in
reasing, and we have
IPfX (t + h) = X (t) + 1g = h + o(h);
IPfX (t + h) X (t) + 2g = o(h):
13th Question:
Show that for every t > s the in
rements of the pro
ess X (t) X (s) are Poisson distributed
with parameter (t s), i.e.
IPfX
14th
15th
16th
(t)
((t s))k :
X (s) = kg = e (t s)
k!
A.BENHARI
149
Let (
; A; IP) be a probability spa
e, F A a sub--eld and X a random
variable with IEjX j < 1. Then every
onditional probability IEfX jFg has the following
properties.
Positivity If X 0 almost surely, then IEfX jFg 0 almost surely.
Monotone Convergen
e If 0 Xn " X , then IEfXn jFg " IEfX jFg almost surely.
Fatou If 0 Xn and IEfXn jFg < 1, then
IEfliminf Xn j Fg liminf IEfXn j Fg almost surely.
n!1
n!1
21st Question:
surely.
We investigate
onditional expe
tations in the situation when two real valued
random variables X and Z have a jont density f (x; z).
22nd Question:
A.BENHARI
150
23rd
k!1 n;mk
25th
Xm jp = 0 ;
26th Question:
IEfX1 jSn g
b) Use the answer to the 25th question to show that this implies
IEfX1 jSn ; Sn+1 ; : : :g
A.BENHARI
151
= Snn :
n=1
n
m=1 Am
(1
Hint: Show that Xn =
IPfAm jF (m 1)g) is a martingale satisfying the
requirement of Question 32. Then look at the two
ases separately.
P
A.BENHARI
152
Suppose a box
ontains initially one bla
k and one white ball. At ea
h step
we pi
k a ball at random from the box and return it together with another ball of the same
olour. After n steps the box
ontains n + 2 balls and we denote by Mn the proportion of
white balls in the box.
a) Show that fMn : n 2 INg is a martingale.
b) Show that Mn
onverges almost surely and determine the distribution of the limiting
random variable.
Question (Optional Sampling Theorem): Suppose that fXn g is a uniformly integrable martingale and T an almost surely nite stopping time with IEjXT j < 1. Show that
IEfXT g = IEfX0 g.
Question (Hitting probabilities for biased random walks): Let X1 ; X2 ; : : : be a
sequen
e of independent, identi
ally distributed random variables with
IPfXn = 1g = p and IPfXn = 1g = 1 p ;
for some 0 < p < 1=2. Let N be a positive integer and a 2 f0; : : : ; N g. Dene the biased
random walk with start in a as
n
Sn = a + Xk :
34th Question:
35th
36th
k=1
37th
nX
n=1
38th Question (Doob de
omposition): Let fXn g be an adapted pro
ess with IEjXn j < 1
for all n. Then there is a martingale fMn g and a previsible pro
ess fCn g su
h that
Xn = Mn + Cn .
39th Question:
Let fMng be a martingale with IEfMn2 g < 1 for all n. Show that if
1
n=1
(Mn
IEf
Mn 1 )2 g < 1;
2
there is a random variable M su
h that nlim
!1 Mn = M almost surely and in L .
A.BENHARI
153
(a) Show that k2 < 1 implies that the random series Xk
onverges almost surely.
k=1
k=1
(b) Suppose that 1there is a K > 0 su
h that jXn j K 1for all n.
Show that if Xk
onverges almost surely, then k2 < 1.
k=1
k=1
Hint: Show that Nn = ( nk=1 Xk )2 nk=1 k2 denes a martingale.
Question: Suppose that fB (t) : t 0g is a standard Brownian motion.
(a) For a < 0 < b dene the stopping time T = inf ft 0 : B (t) 62 (a; b)g and show that
IEfT g = ab.
(b) For a > 0 dene the stopping time T = inf ft 0 : B (t) > ag and show that
IEfT g = 1.
Question: For two probability distributions P and Q on IR with the asso
iated distribtuion fun
tions F and G dene the Levy distan
e as
d(P; Q) = inf f > 0 : F (x ) G(x) F (x + ) + for all xg:
(a) Show that d denes a metri
on the set Prob(IR).
(b) Show that nlim
!1 d(Pn ; P ) = 0 if and only if fPn g
onverges weakly to P .
X
41st
42nd
A.BENHARI
154
Bibliography:
1.P.Billingsley: Probability and Measure, Wiley and sons, New-York, 1979
2.L.Breiman: Probability, Addison-Wesley, Reading, 1968
3.W. B. Davenport, Jr. and W. L. Root,An Introduction to the Theory of Random Signals and Noise.
New York, NY, USA: McGraw-Hill Inc., 1958.
4.W. B. Davenport, Jr., Probability and Random Processes: An Introduction for Applied
Scientists and Engineers. New York, NY, USA: McGraw-Hill Inc., 1970.
5.J. L. Doob, Stochastic Processes. New York, NY, USA: John Wiley & Sons Inc.,1958.
6.R.Durrett.Probability:Theory and examples.Duxbury press, 1996
7.W. Feller An introduction to probability theory and its application Tome I&II. Wiley(1966)
8. J. E. Freund,Mathematical Statistics. Englewood Cliffs, NJ, USA: Prentice-Hall, Inc.,
16th printing, 1962.
9.J. E. Freund and G. A. Simon, Modern Elementary Statistics. Englewood Cliffs, NJ,
USA: Prentice-Hall, Inc., 8th ed., 1992.
10.W. A. Gardner, Introduction to Random Processes with Applications to Signals and
Systems. London, UK: Collier Macmillan Publishers, 1986.
11.Peter Galko, ELG 5119/92.519 Stochastic Processes Course Notes, Faculty of
Engineering, University of Ottawa, Ottawa, ON, Canada, Fall 1987.
12.W. A. Gardner, Introduction to Random Processes with Applications to Signals and
Systems. New York, NY, USA: McGraw-Hill Publishing Company, 2nd ed., 1990.
13. B. V. Gnedenko, Theory of Probability. New York, NY, USA: Chelsea Publishing
Co., 1962. Library of Congress Card No. 61-13496.
14.R. M. Gray and L. D. Davisson,Random Processes: A Mathematical Approach for
Engineers. Englewood Cliffs, NJ, USA: Prentice-Hall, Inc., 1986.
15.H. P. Hsu,Schaum's outline of theory and problems of probability, random variables,
and random processes. New York, NY, USA: McGraw-Hill Inc., 1997.
16.A. N. Kolmogorov,Foundations of the Theory of Probability. New York, NY, USA:
Chelsea Publishing Co., english translation of 1933 german edition, 2nd english ed.,1956.
17.Alberto Leon-Garcia,Probability and Random Processes for Electrical Engineering
Reading, MA, USA: Addison Wesley Publishing Co. Inc., 2nd ed., 1994. ISBN 0-20150037-X.
18.Alberto Leon-Garcia, Student Solutions Manual: Probability and Random Processes
for Electrical Engineering. Reading, MA, USA: Addison-Wesley Publishing Co. Inc.,
2nd ed., 1994. ISBN 0-201-55738-X.
19.M. Love,Probability Theory. Princeton, NJ, USA: D. Van Nostrand Co., Inc., 2nd
ed., 1960.
20.M. Love, Probability Theory, vol. I. New York, NY, USA: Springer, 4th ed., 1977.
21.M. Love, Probability Theory, vol. II. New York, NY, USA: Springer, 4th ed., 1978.
22.I. Miller and J. E. Freund,Probability and Statistics for Engineers. Englewood Cliffs,
NJ, USA: Prentice-Hall, Inc., 2nd ed., 1977.
A.BENHARI
155
A.BENHARI
156