S R S Varadhan Tata Institute of Fundamental Research Lectures On Diffusion Problems and Partial Differential Equations 1981

Lectures on
Diusion Problems and Partial Dierential

Equations
By
S. R. S. Varadhan
Notes by
P. Muthuramalingam
Tara R. Nanda
Tata Institute of Fundamental Research, Bombay
1989
Author
S. R. S. Varadhan
Courant Institute of Mathematical Sciences
251, Mercer Street
New York. N. Y. 10012.
U.S.A.
c Tata Institute of Fundamental Research, 1989
ISBN 3-540-08773-7. Springer-Verlag, Berlin, Heidelberg. New York
ISBN 0-387-08773-7. Springer-Verlag, New York. Heidelberg. Berlin
No part of this book may be reproduced in any
form by print, microlm or any other means with-
out written permission from the Tata Institute of
Fundamental Research, Bombay 400 005
Printed by N. S. Ray at the Book Centre Limited
Sion East, Bombay 400 022 and published by H. Goetze
Springer-Vertal, Heidelberg, West Germany
PRINTED IN INDIA
Contents
1 The Heat Equation 1
2 Kolmogorovs Theorem 11
3 The One Dimensional Random Walk 15
4 Construction of Wiener Measure 19
5 Generalised Brownian Motion 31
6 Markov Properties of Brownian Motion 33
7 Reection Principle 39
8 Blumenthals Zero-One Law 53
9 Properties of Brownian Motion in One Dimension 57
10 Dirichlet Problem and Brownian Motion 63
11 Stochastic Integration 71
12 Change of Variable Formula 91
13 Extension to Vector-Valued It o Processes 95
14 Brownian Motion as a Gaussian Process 101
iii
iv Contents
15 Equivalent For of It o Process 105
16 It os Formula 117
17 Solution of Poissons Equations 129
18 The Feynman-Kac Formula 133
19 An Application of the Feynman-Kac Formula.... 139
20 Brownian Motion with Drift 147
21 Integral Equations 155
22 Large Deviations 161
23 Stochastic Integral for a Wider Class of Functions 187
24 Explosions 195
25 Construction of a Diusion Process 201
26 Uniqueness of Diusion Process 211
27 On Lipschitz Square Roots 223
28 Random Time Changes 229
29 Cameron - Martin - Girsanov Formula 237
30 Behaviour of Diusions for Large Times 245
31 Invariant Probability Distributions 253
32 Ergodic Theorem 275
33 Application of Stochastic Integral 281
Appendix 284
Contents v
Language of Probability 284
Kolmogorovs Theorem 288
Martingales 290
Uniform Integrability 305
Up Crossings and Down Crossings 307
Bibliography 317
vi Contents
Preface
THESE ARE NOTES based on the lectures given at the T.I.F.R.
Centre, Indian Institute of Science, Bangalore, during July and August
of 1977. Starting from Brownian Motion, the lectures quickly got into
the areas of Stochastic Dierential Equations and Diusion Theory. An
attempt was made to introduce to the students diverse aspects of the
theory. The last section on Martingales is based on some additional
lectures given by K. Ramamurthy of the Indian Institute of Science. The
author would like to express his appreciation of the eorts by Tara R.
Nanda and PL. Muthuramalingam whose dedication and perseverance
has made these notes possible.
S.R.S. Varadhan
1. The Heat Equation
LET US CONSIDER the equation 1
(1) u
t

1
2
u = 0
which describes (in a suitable system of units) the temperature distribu-
tion of a certain homogeneous, isotropic body in the absence of any heat
sources within the body. Here
u u(x
1
, . . . , x
d
, t); u
t

u
t
; u =
d
i=1
2
u
x
2
i
,
t represents the time ranging over [0, ) or [0, T] and x (x
1
. . . x
d
)
belongs to R
d
.
We rst consider the initial value problem. It consists in integrating
equation (1) subject to the initial condition
(2) u(0, x) = f (x).
The relation (2) is to be understood in the sense that
Lt
t0
u(t, x) = f (x).
Physically (2) means that the distribution of temperature throughout
the body is known at the initial moment of time.
We assume that the solution u has continuous derivatives, in the
space coordinates upto second order inclusive and rst order derivative
in time.
1
2 1. The Heat Equation
It is easily veried that
(3) u(t, x) =
1
(2t)
d/2
exp
_
|x|
2
2t
_
; |x|
2
=
d
i=1
x
2
i
,
satises (1) and 2
(4) u(0, x) = Lt
t0
u(t, x) = (x)
Equation (4) gives us a very nice physical interpretation. The so-
lution (3) can be interpreted as the temperature distribution within the
body due to a unit sourse of head specied at t = 0 at the space point
x = 0. The linearity of the equation (1) now tells us that (by superpo-
sition) the solution of the initial value problem may be expected in the
form
(5) u(t, x) =
_
R
d
f (y)p(t, x y)dy,
where
p(t, x) =
1
(2t)
d/2
exp
|x|
2
2t
.
Exercise 1. Let f (x) be any bounded continuous function. Verify that
p(t, x) satises (1) and show that
(a)
_
p(t, x)dx = 1, t > 0;
(b) Lt
t0+
_
p(t, x) f (x)dx = f (0);
(c) using (b) justify (4). Also show that (5) solves the initial value
problem.
(Hints: For (a) use
e
x
2
dx =
. For part (b) make the substitution

y =
x
and apply Lebesgue dominated convergence theorem).

3
Since equation (1) is linear with constant coecients it is invariant
under time as well as space translations. This means that translates of
solutions are also solutions. Further, for s 0, t > 0 and y R
d
, 3
(6) u(t, x) =
1
[2(t + s)]
d/2
exp
|x y|
2
2(t + s)
and for t > s, y R
d
,
(7) u(t, x) =
1
[2(t s)]
d/2
exp
|x y|
2
2(t s)
are also solutions of the heat equation (1).
The above method of solving the initial value problem is a sort of
trial method, viz. we pick out a solution and verify that it satises (1).
But one may ask, how does one obtain the solution? A partial clue to this
is provided by the method of Fourier transforms. We pretend as if our
solution u(t, x) is going to be very well behaved and allow all operations
performed on u to be legitimate.
Put v(t, x) = u(t, x) where stands for the Fourier transform in the
space variables only (in this case), i.e.
v(t, x) =
_
R
d
u(t, y)e
i xy
dy.
Using equation (1), one easily veries that
(8) v
t
(t, x) =
1
2
|x|
2
v(t, x)
with
(9) v(0, x) =

f (x).
The solution of equation (8) is given by
(10) v(t, x) =

f (x)e
t|x|
2
/2
.
We have used (9) in obtaining (10).
Exercise 2. Verify that 4
p(t, x) = exp
_
t|x|
2
2
_
.
Using Exercise 2, (10) can be written as
(11) v(t, x) = u(t, x) =

f (x) p(t, x).
The right hand side above is the product of two Fourier transforms
and we know that the Fourier transform of the convolution of two fun-
tions is given by the product of the Fourier transforms. Hence u(t, x) is
expected to be of the form (5).
Observe that if f is non-negative, then u is nonnegative and if f
is bounded by M then u is also bounded by M in view of part (a) of
Exercise 1.
The Inhomogeneous Equation. Consider the equation
v
t

v
2
= g, with v(0, x) = 0,
which describes the temperature within a homogeneous isotropic body
in the presence of heat sources, specied as a function of time and space
by g(t, x). For t > s,
u(t, x) =
1
[2(t s)]
d/2
exp
|x y|
2
2(t s)
is a solution of u
t
(t, x)
1
2
u(t, x) = 0 corresponding to a unit source at
t = s, x = y. Consequently, a solution of the inhomogeneous problem is
obtained by superposition.
Let
v(t, x) =
_
R
d
t
_
0
g(s, y)
1
[2(t s)]
d/2
exp
_
|x y|
2
2(t s)
_
dy ds
i.e. 5
5
v(t, x) =
t
_
0
w(t, x, s)ds
where
w(t, x, s) =
_
R
d
g(s, y)
1
[2(t, s)]
d/2
exp
_
|x y|
2
2(t s)
_
dy.
Exercise 3. Show that v(t, x) dened above solves the inhomogeneous
heat equation and satises v(0, x) = 0. Assume that g is suciently
smooth and has compact support. v
t

1
2
v = Lt
st
w(t, x, s) and now use
part (b) of Exercise (1).
Remark 1. We can assume g has compact support because in evaluating
v
t

1
2
v the contribution to the integral is mainly from a small neigh-
bourhood of the point (t, x). Outside this neighbourhood
1
[2(t s)]
d/2
exp
_
|x y|
2
2(t s)
_
satises
u
t

1
2
u = 0.
2. If we put g(s, y) = 0 for s < 0, we recognize that v(t, x) = g p.
Taking spatial Fourier transforms this can be written as
v(t, ) =
t
_
0
g(s, ) exp
1
2
(t s)||
2
d,
or
v
t
=
v
t
= g(t, ) +
1
2
v =
_
g(t, ) +
1
2
v
_
.
Therefore
v
t

1
2
v = g.
Exercise 4. Solve w
t

1
2
w = g on [0, ) R
d
with w = f on {0} R
d
6
(Cauchy problem for the heat equation).
Uniqueness. The solution of the Cauchy problem is unique provided the
class of solutions is suitably restricted. The uniqueness of the solution
is a consequence of the Maximum Principle.
Maximum Principle. Let u be smooth and bounded on [0, T] R
d
sat-
isfying
u
t

u
2
0 in (0, T] R
d
and u(0, x) 0, x R
d
.
Then
u(t, x) 0 , t [0, T] and x R
d
.
Proof. The idea is to nd minima for u or for an auxillary function.
Step 1. Let v be any function satisfying
v
t

v
2
> 0 in (0, T] R
d
.
Claim. v cannot attain a minimum for t
0
(0, T]. Assume (to get a
contradiction) that v(t
0
, x
0
) v(t, x) for some t
0
> 0 and for all t
[0, T], x R
d
. At a minimum v
t
(t
0
, x
0
) 0, (since t
0
0) v(t
0
, x
0
)
0. Therefore
_
v
t

v
2
_
(t
0
, x
0
) 0.
Thus, if v has any minimum it should occur at t
0
= 0.
Step 2. Let > 0 be arbitrary. Choose such that
h(t, x) = |x|
2
+ t
satises 7
h
t

h
2
= d > 0 (say = 2d).
7
Put v
= u + h. Then
v
t

1
2
v
> 0.
As u is bounded, v
+as |x| +, v
must attain a minimum.

This minimum occurs at t = 0 by Step 1. Therefore,
v
(t, x) v
(0, x
0
) for some x
0
R
d
,
i.e.
v
(t, x) u(0, x
0
) + |x
0
|
2
> 0,
i.e.
u(t, x) + h(t, x) > 0, .
This gives
u(t, x) 0.
This completes the proof.
Exercise 5. (a) Let L be a linear dierential operator satisfying Lu =
g on (open in R
d
) and u = f on . Show that u is uniquely
determined by f and g if and only if Lu = 0 on and u = 0 on
imply u = 0 on .
(b) Let u be a bounded solution of the heat equation u
t

1
2
u = g
with u(0, x) = f (x). Use the maximum principle and part (a) to
show that u is unique in the class of all bounded functions.
(c) Let
g(t) =
_
_
e
1/t
2
, if t > 0,
0, if t 0,
u(t, x) =
k=0
g
(k)
(t/2)x
2
k
(2k)!
, on R R.
8
Then
u(0, x) = 0, u
t
=
u
2
, u 0,
i.e. u satises
u
t

1
2
2
u
x
2
= 0, with u(0, x) = 0.
This example shows that the solution is not unique because, u is
not bounded. (This example is due to Tychono).
Lemma 1. Let p(t, x) =
1
(2t)
d/2
exp
|x|
2
2t
for t > 0. Then
p(t, ) p(s, ) = p(t + s, ).
Proof. Let f be any bounded continuous function and put
u(t, x) =
_
R
d
f (y)p(t, x y)dy.
Then u satises
u
t

1
2
u = 0, u(0, x) = f (x).
Let
v(t, x) = u(t + s, x).
Then
v
t

1
2
v = 0, v(0, x) = u(s, x).
This has the unique solution
v(t, x) =
_
u(s, y)p(t, x y)dy.
Thus
_
R
d
f (y)p(t + s, x y)dy =
f (z)p(s, y z)p(t, x y)dz dy.

9
9
This is true for all f bounded and continuous. We conclude, there-
fore, that
p(t, ) p(s, ) = p(t + s, ).
Exercise 6. Prove Lemma 1 directly using Fourier transforms.

It will be convenient to make a small change in notation which will
be useful later on. We shall write p(s, x, t, y) = p(ts, yx) for every x, y
and t > s. p(s, x, t, y) is called the transition probability, in dealing with
Brownian motion. It represents the probability density that a Brownian
particle located at space point x at time s moves to the space point y at
a later time t.
Note . We use the same symbol p for the transition probability; it is
function of four variables and there will not be any ambiguity in using
the same symbol p.
Exercise 7. Verify that
_
R
d
p(s, x, t, y)p(t, y, , z)dy = p(s, x, , z), s < t < .
(Use Exercise 6).
Remark . The signicance of this result is obvious. The probability
that the particle goes from x at time s to z at time is the sum total
of the probabilities, that the particle moves from x at s to y at some
intermediate time t and then to z at time .
10
In this section we have introduced Brownian Motion corresponding
to the operator
1
2
. Later on we shall introduce a more general diusion
process which corresponds to the operator
1
2
_
a
i j
2
x
i
x
j
+
_
b
j
x
j
.
2. Kolmogorovs Theorem
Denition. LET (, B, P) BE A probability space. Astochastic process 11
in R
d
is a collection {X
t
: t I} of R
d
-valued random variables dened
on (, B).
Note 1. I will always denote a subset of R
+
= [0, ).
2. X
t
is also denoted by X(t).
Let {X
t
: t I} be a stochastic process. For any collection t
1
,
t
2
, . . . , t
k
such that t
i
I and 0 t
1
< t
2
< . . . < t
k
and any Borel
set , in R
d
R
d
R
d
(k times),. dene
F
t
1
. . . t
k
() = P(w : (X
t
1
(w), . . . , X
t
k
(w)) ).
If
{t
1
, . . . , t
k
} {s
1
, . . . , s
} I, with l k
such that
s
(0)
1
< . . . < s
(0)
n
0
< t
1
< s
(1)
1
. . . < s
(1)
n
< t
2
. . . < t
k
< s
(k)
1
. . . < s
(k)
n
k
,
let then
: R
d
R
d
(1 times) R
d
R
d
(k times)
be the canonical projection. If E
t
i
R
d
is any Borel set in R
d
, i =
1, 2, . . . , k, then
1
(E
t
1
E
t
k
) = R
d
E
t
1
R
d
E
t
2
R
d
11
12 2. Kolmogorovs Theorem
(l times). The following condition always holds.
(*) E
t
1
. . . t
k
(E
t
1
E
t
k
) = F
s
1
. . . s
1
(
1
(E
t
1
E
t
k
)).
If () holds for an arbitrary collection {F
t
1
. . . t
k
: 0 t
1
< t
2
. . . < t
k
} 12
(k = 1, 2, 3 . . .) of distributions then it is said to satisfy the consistency
condition.
Exercise 1. (a) Verify that F
t
1
. . . t
k
is a probability measure on R
d
R
d
(k times).
(b) Verify (). (If B
m
denotes the Borel eld of R
m
, B
m+n
= B
m

B
n
).
The following theorem is a converse of Exercise 1 and is often used
to identify a stochastic process with a family of distributions satisfying
the consistency condition.
Kolmogorovs Theorem.
Let {F
t
1
,t
2
,...t
k
0 t
1
< t
2
< . . . < t
k
< } be a family of probability
distributions (on R
d
R
d
, k times, k = 1, 2, . . .) satisfying the
consistency condition. Then there exists a measurable space (
k
, B),
a unique probability measure P an (
k
, B) and a stochastic process
{X
t
: 0 t < } such that the family of probability measures associated
with it is precisely
{F
t
1
,t
2
,...t
k
: 0 t
1
< t
2
< . . . < t
k
< }, k = 1, 2, . . . .
A proof can be found in the APPENDIX. We mention a few points
about the proof which prove to be very useful and should be observed
carefully.
1. The space
k
is the set of all R
d
-valued functions dened on 13
[0, ):
K
=
_
t[0,)
R
d
2. The random variable X
t
is the t
th
-projection of
K
onto R
d
t
.
13
3. B is the smallest -algebra with respect to which all the projec-
tions are measurable.
4. P given by
P(w : X
t
1
(w) A
1
, . . . X
t
k
(w) A
k
) = F
t
1
...t
k
(A
1
A
k
)
where A
i
is a Borel set in R
d
, is a measure on the algebra generated
by {X
t
1
, . . . X
t
k
}(k = 1, 2, 3 . . .) and extends uniquely to B.
Remark . Although the proof of Kolmogorovs theorem is very con-
structive the space
K
is too large and the -algebra B too small
for practical purposes. In applications one needs a nice collection
of R
d
-valued functions (for example continuous, or dierentiable func-
tions), a large -algebra on this collection and a probability measure
concentrated on this family.
3. The One Dimensional
Random Walk
BEFORE WE TAKE up Brownian motion, we describe a one dimen- 14
sional random walk which in a certain limiting case possesses the prop-
erties of Brownian motion.
Imagine a person at the position x = 0 at time t = 0. Assume that at
equal intervals of time t = he takes a step h either along the positive
x axis or the negative x axis and reaches the point x(t) = x(t ) + h or
x(t) = x(t ) h respectively. The probability that he takes a step in
either direction is assumed to be 1/2. Denote by f (x, t) the probability
that after the time t = n (n intervals of time ) he reaches the position x.
If he takes m steps to the right (positive x-axis) in reaching x then there
are
n
C
m
possible ways in which he can achieve these m steps. Therefore,
the probability f (x, t) is
n
C
m
(
1
2
)
n
.
f (x, t) satises the dierence equation
(1) f (x, t + ) =
1
2
f (x h, t) +
1
2
f (x + h, t)
and
(2) x = h(m (n m)) = (2m n)h.
To see this one need only observe that to reach (x, t + ) there are
two ways possible, viz. (x h, t) (x, t +) or (x +h, t) (x, t +) and
the probability for each one of these is 1/2. Also note that by denition
15
16 3. The One Dimensional Random Walk
of f ,
(3) f (h, ) =
1
2
= f (h, ),
so that 15
(4) f (x, t + ) = f (h, ) f (x h, t) + f (h, ) f (x + h, t).
The reader can identify (4) as a discrete version of convolution.
By our assumption,
(5) f (0, 0) = 1, f (x, 0) = 0 if x 0.
We examine equation (1) in the limit h 0, 0. To obtain
reasonable results we cannot let h and tend to zero arbitratily. Instead
we assume that
(6)
h
1 as h 0 and 0.
The physical nature of the problem suggests that (6) should hold. To
see this we argue as follows. Since the person is equally likely to go in
either direction the average value of x will be 0. Therefore a reasonable
measure of the progress made by the person is either |x| or x
2
. Indeed,
since x is a random variable (since m is one) one gets, using (2),
E(x) = 2E(m) n = 0, E(x
2
) = h
2
E((2m n)
2
) = h
2
n.
(Use
n
_
m=0
m
n
C
m
_
1
2
_
n
=
n
2
,
n
_
m=0
n
C
m
_
1
2
_
n
=
n(n + 1)
4
)
Thus
E
_
x
2
t
_
=
1
t
E(x
2
) =
h
2
n
n
=
h
2
,
and as t becomes large we expect that the average distance covered per
unit time remains constant. (This constant is chosen to be 1 for reasons 16
that will become apparent later). This justies (6). In fact, a simple
argument shows that if
h
2
0 or +, x may approach + in a nite

time which is physically untenable.
17
(1) now gives
f (x, t + ) f (x, t) =
1
2
{ f (x h, t) f (x, t) + f (x, h, t) f (x, t)}.
Assuming sucient smoothness on f , we get in the limit as h, 0
and in view of (6),
(7)
f
t
=
1
2
2
f
x
2
(to get the factor 1/2 we choose
h
2
1). This is the equation satised

by the probability density f . The particle in this limit performs what is
known as Brownian motion to which we now turn our attention.
References.
[1] GNEDENKO: The theory of probability, Ch. 10.
[2] The Feynman Lectures on physics, Vol. 1, Ch. 6.
4. Construction of Wiener
Measure
ONEEXAMPLEWHEREthe Kolmogorov construction yields a proba- 17
bility measure concentrated on a nice class is the Brownian motion.
Denition . A Brownian motion with starting point x is an R
d
-valued
stochastic process {X(t) : 0 t < } where
(i) X(0) = x = constant;
(ii) the family of distribution is specied by
F
t
1
. . . t
k
(A) =
_
A
p(0, x, t
1
, x
1
)p(t
1
, x
1
, t
2
, x
2
) . . .
p(t
k1
, x
k1
, t
k
, x
k
)dx
1
. . . dx
k
for every Borel set A in R
d
R
d
(k times).
N.B. The stochastic process appearing in the denition above is the one
given by the Kolmogorov construction.
It may be useful to have the following picture of a Brownian motion.
The space
k
may be thought of as representing particles performing
Brownian movement; {X
t
: 0 t < } then represents the trajectories
of these particles in the space R
d
as functions of time and Bcan be con-
sidered as a representation of the observations made on these particles.
Exercise 2. (a) Show that F
t
1
...t
k
dened above is a probability mea-
sure on R
d
R
d
(k times).
19
20 4. Construction of Wiener Measure
(b) {F
t
1
...t
k
: 0 t
1
< t
2
< . . . t
k
< } satises the consistency
condition. (Use Fubinis theorem).
(c) X
t
1
x, X
t
2
X
t
1
, . . . , X
t
k
X
t
k1
are independent random variables 18
and if t > s, then X
t
X
s
is a random variable whose distribution
density is given by
p(t s, y) =
1
[2(t s)]
d/2
exp
_
1
2
(t s)
1
|y|
2
_
.
(Hint: Try to show that X
t
1
x, X
t
2
X
t
1
, . . . , X
t
k
X
t
k1
have a
joint distribution given by a product measure. For this let be
any bounded real measurable function on R
d
R
d
(k times).
Then
E((Z
1
, . . . , Z
k
))
X
t
1
x,X
t
2
X
t
1
,...,X
t
k
X
t
k1
=
E
X
t
1
,...,X
t
k
((Z
1
x, . . . , Z
k
Z
k1
))
where E()
X
t
1
...X
t
k
is the expectation of with respect to the joint dis-
tribution of (X
t
1
. . . , X
t
k
). You may also require the change of vari-
able formula).
Problem. Given a Brownian motion with starting point x our aim is to
nd a probability P
x
on the space = C([0, ); R
d
) of all continuous
funcitons from [0, ) R
d
which induces the Brownian motion. We
will thus have a continuous realisation to Brownian motion. To achieve
this goal we will work with the collection {F
t
1
,...,t
k
: 0 t
1
< t
2
< . . . <
t
k
} where t
i
D, a countable dense subset of [0, ).
Step 1. The rst step is to nd a probability measure on a smaller 19
space and lift it to C([0, ); R
d
). Let
= C([0, ); R
d
),
D a countable dense subset of [0, ); (D) = {F : D R
d
} where f
is uniformly continuous on [0, N] D for N = 1, 2, . . .. We equip
with the topology of uniform convergence on compact sets and (D)
with the topology of uniform convergence on sets of the form D K
21
where K [0, ) is compact; and (D) are separable metric spaces
isometric to each other.
Exercise 3. Let
p
n
( f , g) = sup
0tn
| f (t) g(t)| for f , g
and
p
n,D
( f , g) = sup
0tn
tD
| f (t) g(t)| for f , g (D).
Dene
( f , g) =
n=1
1
2
n
p
n
( f , g)
1 + p
n
( f , g)
, f , g ,
D
( f , g) =
n=1
1
2
n
p
n,D
( f , g)
1 + p
n,D
( f , g)
, f , g (D).
Show that
(i) { f
n
} converges to f if and only if f
n
f uniformly on
compact subsets of [0, );
(ii) { f
n
} (D) converges to f if and only if f
n|DK
f
|DK|
uni-
formly for every compact subset K of [0, );
(iii) {(P
1
, . . . , P
d
)} where P
i
is a polynomial with rational coecients 20
is a countable dense subset of ;
(iv) {(P
1D
, . . . , P
dD
)} is a countable dense subset of (D);
(v) : (D) where ( f ) = f
|D
is a (,
D
)-isometry of onto
(D);
(vi) if V( f , , n) = {g : p
n
( f , g) < } for f , > 0 and
V
D
( f , , n) = {g (D) : p
n,D
( f , g) < } for f (D), > 0,
then
{V( f , , n) : f , > 0, n = 1, 2 . . .}
is a base for the topology of and
{V
D
( f , , n) : f (D), > 0, n = 1, 2, . . .}
is a base for the topology of (D).
Remark. By Exercise 3(v) any Borel probability measure on (D) can
be lifted to a Borel probability measure on .
2nd Step. Dene the modulus of continuity
T,
D
( f ) of a function f on
D in the interval [0, T] by
T,
D
( f ) = sup{| f (t) f (s)| : |t s| < t, s D [0, T]}
As D is countable one has
Exercise 4. (a) Show that f :
N,
1
j
D
( f )
1
k
} is measurable in the -
algebra generated by the projections
t
: {R
d
t
: t D} R
d
t
Proof. The lemma is equivalent to showing that B = (E ). As each of 21
the projection
t
1
...t
k
is continuous, (E ) B. To show that B (E ),
it is enough to show that V
D
( f , , n) E because (D) is separable. (Cf.
Exercise 3(iv) and 3(vi)). By denition
V
D
( f , , n) = {g (D) : P
n,D
( f , g) < }
=
_
m=1
_
g (D) : p
n,D
( f , g)
1
m
_
=
_
m=1
{g (D) : |g(t
i
) f (t
i
)|
1
m
, t
i
D [0, n]}.
The result follows if one observes that each
t
i
is continuous.
Remark 1. The lemma together with Exercise 4(b) signies that the
Kolmogorov probability P
x
on {R
d
t
: t D} is dened on the topo-
logical Borel -eld of (D).
2. The proof of the lemma goes through if (D) is replaced by .
23
Step 3. We want to show that P
x
((D)) = 1. By Exercise 4(b) this is
equivalent to showing that Lt
j
P(
N,1/ j
D
( f )
1
k
) = 1 for all N and k.
The lemmas which follow will give the desired result.
Lemma (L evy). Let X
1
, . . . X
n
be independent random variables, > 0
and > 0 arbitrary. If
P(|X
r
+ X
r+1
+ + X
| )
r, such that 1 r n, then
P( sup
1jn
|X
1
+ + X
j
| 2) 2.
(see Kolmogorovs theorem) for every j = 1, 2, . . . , for every N = 22
1, 2, . . . and for every k = 1, 2, . . .. (Hint: Use the fact that the pro-
jections are continuous).
(b) Show that (D) =
_
N=1
_
k=1
_
j=1
{
N,
1
j
D
( f )
1
k
} and hence (D) is
measurable in {R
d
t
: t D}.
Let
t
1
...t
k
: (D) R
d
R
d
R
d
(k times) be the projections
and let
E
t
1
...t
k
=
1
t
1
...t
k
(B(R
d
)
k times
B(R
d
)).
Put
E = {E
t
1
...t
k
: 0 t
1
< t
2
< . . . < t
k
< ; t
i
D}.
Then, as
E
t
1
...t
k
E
s
1
...s
1
E
1
...
m
,
where
{t
1
. . . t
k
, s
1
. . . s
1
} {
1
. . .
m
},
E is an algebra. Let (E ) be the -algebra generated by E .
Lemma . Let B be the (topological) Borel -eld of (D). Then B is
the -algebra generated by all the projections
{
t
i
...t
k
: 0 t
1
< t
2
< . . . < t
k
, t
i
D}.
Remark. By subadditivity it is clear that 23
P
_
_
sup
1jn
|X
1
+ + X
j
| 2
_
_
n.
Ultimately, we shall let n and this estimate is of no utility. The
importance of the lemma is that it gives an estimate independent of n.
Proof. Let S
j
= X
1
+ + X
j
, E = { sup
1jn
|S
j
| 2}. Put
E
1
= {|S
1
| 2},
E
2
= {|S
1
| < 2, |S
2
| 2},
. . . . . . . . . . . .
. . . . . . . . . . . .
E
n
= {|S
j
| < 2, 1 j n 1, |S
n
| 2}.
Then
E =
n
_
j=1
E
j
, E
j
E
i
= if j i;
P{E (|S
n
| ) = P
_
_
n
_
j=1
(E
j
(|S
n
| ))
_
_
P
_
_
(E
i
(|S
n
S
j
| ))
_
j=1
P(E
j
)P(|S
n
S
j
| ) (by independence)
P(E) (by data).
= P{E (|S
n
| > )} P(|S
n
| > ) (by data).
Combining the two estimates above, we get
P(E) + P(E).
If >
1
2
, 2 > 1. If <
1
2
,

1
2. In either case P(E) 2. 24
25
Lemma . Let {X(t)} be a Brownian motion, I [0, ) be a nite inter-
val, F I D be nite. Then
P
x
_
Sup
t,F
|X(t) X()| 4
_
C(d)
|I|
2
4
,
where |I| is the length of the interval and C(d) a constant depending only
on d.
Remark. Observe that the estimate is independent of the nite set F.
Proof. Let F = {t
i
: 0 t
1
< t
2
< . . . < t
k
< }.
Put
X
1
= X(t
2
) X(t
1
), . . . , X
k1
= X(t
k
) X(t
k1
).
Then X
1
. . . X
k1
are independent (Cf. Exercise 2(c)). Let
= sup
1r1k1
P
x
(|X
r
+ X
r+1
+ + X
1
| ).
Note that
P
x
(|X
r
+ + X
1
| ) = P(|X(t
) X(t
)| ) for some t
, t
in F
E(|X(t
) X(t
)|
4
)
4
(see Tchebysheys inequality in Appendix)
(t
4
(C
= constant)
|I|
2
4
.
(*)
Therefore
C
|I|
2
4
. Now
P
x
( sup
t,P
|X(t) X()| 4)
P
x
( sup
1ik
|X(t
i
) X(t
1
)| 2)
= P
x
( sup
ijk1
|X
1
+ + X
j
| 2) 2 (by previous lemma)
2C
|I|
2
4
=
C|I|
2
4
.
25
Exercise 5. Verify ().
(Hint: Use the density function obtained in Exercise 2(c) to evaluate the
expectation and go over to popular coordinates. (The value of C
is
d(2d + 1))).
Lemma .
P
x
_
_
sup
|ts|h
t,s[0,t]D
|X(t) X(s)| >
_
_
= P
x
(
T,h
D
> )
(T, , h) = C
h
4
__
T
h
_
+ 1
_
.
Note that (T, , h) 0 as h 0 for every xed T and .
Proof. Dene the intervals I
1
, I
2
, . . . by
I
k
= [(k 1)h, (k + 1)h] (0, T], k = 1, 2, . . . .
Let I
1
, I
2
, . . . I
r
be those intervals for which
I
j
[0, T] ( j = 1, 2, . . . , r).
Clearly there are [
T
h
] +1 of them. If |t s| h then t, s I
j
for some 26
j, 1 j r. Write D =
_
n=1
F
n
where F
n
F
n+1
and F
n
is nite. Then
P
x
_
_
sup
|ts|h
t,s[0,T]D
|X(t) X(s)| >
_
_
= P
x
_
_
n=1
_
_
sup
|ts|h
t,sDF
n
|X(t) X(s)| >
_
_
_
_
= sup
n
P
x
_
_
sup
j
sup
t,sF
n
(|X
I
j
(t) X(s)| > )
_
_
27
sup
n
r
j=1
P
x
_
sup
t,sF
n
(|X
I
j
(t) X(s)| > )
_
sup
n
__
T
h
_
+ 1
_
C(2h)
2
(/4)
4
by the last lemma
(T, , h).
Theorem . P
x
((D)) = 1.
Proof. It is enough to show that
Lt
j
P
x
_
N,
1
j
D
( f )
1
k
_
= 1 (See Exercise 4(b)).
But this is guaranteed by the previous lemma.
Remark. 1. It can be shown that the outer measure of is 1.
2. is not measurable in

t0
R
d
t
.
Let

P
x
be the measure on induced by P
x
on (D). We have al- 27
ready remarked that P
x
is dened on the (topological Borel eld of
(D). As P
x
is a probability measure,

P
x
is also a probability mea-
sure. It should now be veried that

P
x
is really the probability measure
consistent with the given distribution.
Theorem .

P
x
is a required probability measure for a continuous real-
ization of the Brownian motion.
Proof. We have to show that
F
t
1
...t
k
=

P
x
1
t
1
...t
k
for all t
1
, t
2
. . . t
k
in [0, ).
Step 1. Let t
1
, . . . , t
k
D. Then
P
x
(
1
t
1
...t
k
(A
1
A
k
)) = P
x
(
t
1
...t
k
(A
1
A
k
))
for every A
i
Borel in R
d
. The right side above is
P
x
(
1
t
1
...t
k
(A
1
A
k
)) = F
t
1
...t
k
(A
1
A
k
)
(by denition of P
x
).
Step 2. We know that T
t
1
,t
2
...t
k
=

P
x
t
1
,t
2
,...,t
k
provided that t
1
, t
2
, . . . , t
k

D. Let us pick t
(n)
1
, . . . , t
(n)
k
, such that each t
(n)
i
D and t
(n)
k
t
k
as
n . For each n and for each xed f : R
d
R which is bounded
and continuous,
E
F
(n)
1
,...t
(n)
k
[ f (x
1
, . . . , x
k
)] = E

P
x
[ f (x(t
(n)
1
, . . . , x(t
(n)
k
)))].
Letting n we see that 28
E
F
t
1
,...,t
k
[ f (x
1
, . . . , x
k
)] = E
P
x
[ f (x(t
1
), . . . , x(t
k
))]
for all t
1
, . . . , t
k
. This completes the proof.
The denition of the Brownian motion given earlier has a built-in
constraint that all trajectories start from X(0) = x. This result is given
by
Theorem .

P
0
{ f : f (0) = 0} = 1.
Proof. Obvious; because E

P
x
[(x(0))] = (x).
Note. In future

P
x
will be replaced by P
x
and

P
0
= P
0
will be denoted
by P.
Let T
x
: be the map given by (T
x
f )(t) = x + f (t). T
x
translates every trajectory through the vector x.
Time
Let us conceive a real Brownian motion of a system of particles. The
operation T
x
means that the system is translated in space (along with
29
everything else that aects it) through a vector x. The symmetry of the
physicl laws governing this motion tells us that any property exhibited 29
by the rst process should be exhibited by the second process and vice
versa. Mathematically this is expressed by
Theorem . P
x
= PT
1
x
.
Proof. It is enough to show that
P
x
(T
x
1
t
1
...t
k
(A
1
A
k
)) = P(
1
t
1
...t
k
(A
1
A
k
))
for every A
i
Borel in R
d
. Clearly,
T
x
1
t
1
...t
k
(A
1
A
k
) =
1
t
1
...t
k
(A
1
x A
k
x).
Thus we have only to show that
_
A
1
x
_
. . .
_
A
k
x
p(0, x, t
1
, x
1
) . . . p(t
k1
, x
k1
, t
k
, x
k
)dx
1
. . . dx
k
=
_
A
1
. . .
_
A
k
p(0, 0, t
1
, x
1
) . . . p(t
k1
, x
k1
, t
k
, x
k
)dx
1
. . . dx
k
,
which is obvious.
Exercise. (a) If (t, ) is a Brownian motion s tarting at (0, 0) then
1
(t) is a Brownian motion starting at (0, 0) for every > 0.

(b) If X is a d-dimensional Brownian motion and Y is a d
-dimensio-
nal Brownian motion then (X, Y) is a d+d
dimensional Brownian
motion provided that X and Y are independent.
(c) If X
t
= (X
1
t
, . . . , X
d
t
) is a d-dimensional Brownian motion, then X
j
t
is a one-dimensional Brownian motion. ( j = 1, 2, . . . d).
(w) = inf{t : |X
t
(w)| +1}
= inf{t : |w(t)| 1}
(w) is the rst time the particle hits either of the horizontal lines 30
+1 or 1.
6.4
2. Let {X
t
} be a d-dimensional Brownian motion, G any closed set in
R
d
. Dene
(w) = inf{t : w(t) G}.
This is a generalization of Example 1. To see that is a stopping
time use
{ s} =
_
n=1
lim
[0,s]
rational
{w : w() G
n
},
where
G
n
=
_
x R
d
: d(x, G)
1
n
_
.
3. Let (X
t
) be a d-dimensional Brownian motion, C and D disjoint
closed sets in R
d
. Dene
(w) = inf{t; w(t) C and for some s t, w(s) D}.
(w) is the rst time that w hits C after visiting D.
5. Generalised Brownian
Motion
LET BE ANY space, F a -eld and (F
t
) an increasing family of 31
sub -elds such that (F
t
) = F. Let P be a measure on (, F).
X(t, w) : [0, w) R
d
is called a Brownian motion relative to (, F
t
, P) if
(i) X(t, w) is progressively measurable with respect to F
t
;
(ii) X(t, w) is a.e. continuous in t;
(iii) X(t, w) X(s, w) for t > s is independent of F
s
and is distributed
normally with mean 0 and variance t s, i.e.
P(X(t, ) X(s, ) A|F
s
) =
_
A
1
[2(t s)]
d/s
exp
|y|
2
2(t s)
dy.
Note. 1. The Brownian motion constructed previously was concen-
trated on = C([0, ); R
d
), F was the Borel eld of , X(t, w) =
w(t) and F
t
= {X(s) : 0 s t}. The measure P so obtained is
often called the Wiener measure.
2. The above denition is more general because
{X(s) : 0 s t} F
t
.
31
32 5. Generalised Brownian Motion
Exercise. (Brownian motion starting at time s). Let = C([s, ); R
d
),
B =Borel eld of . Show that for each x R
d
a probability measure
P
s
x
on such that
(i) P
s
x
{w : w(s) = x} = 1; 32
(ii) P
s
x
(X
t
1
A
1
, . . . , X
t
k
A
k
)
=
_
A
1
_
A
2
. . .
_
A
k
p(s, x, t
1
, x
1
)p(t
1
, x
1
, t
2
, x
2
) . . .
. . . p(t
k1
x
k1
, t
k
, x
k
)dx
1
. . . dx
k
,
s < t
1
< . . . < t
k
.
For reasons which will become clear later, we would like to shift the
measure P
s
x
to a measure on C([0, ); R
d
). To do this we dene
T : C([s, ); R
d
) C([0, ); R
d
)
by
(Tw)(t) =
_
_
w(t), if t s,
w(s), if t s.
Clearly, T is continuous. Put
P
s,x
= P
s
x
T
1
.
Then
(i) P
s,x
is a probability measure on the Borel eld of C([0, ); R
d
);
(ii) P
s,x
{w : w(s) = x} = 1.
6. Markov Properties of
Brownian Motion
Notation. 1. Arandom variable of a stochastic process {X(t)}
tI
shall 33
be denoted by X
t
or X(t). 0 t < .
2. F
s
will denote the -algebra generated by {X
t
: 0 t s};
F
s+
= {F
a
: a > s}; F
s
will be the -algebra generated by
{F
a
: a < s}s > 0. It is clear that {F
t
} is an increasing family.
3. For the Brownian motion, B = the -algebra generated by {F
t
:
t < } will be denoted by F.
Theorem . Let {X
t
: 0 t < } be a Brownian motion. Then X
t
X
s
is
independent of F
s
.
Proof. Let
0 t
1
< t
2
< t
3
< . . . < t
k
s.
Then the -algebra generated by X
t
1
, . . . , X
t
k
is the same as the -
algebra generated by
X
t
1
, X
t
2
X
t
1
, . . . , X
t
k
X
t
k1
.
Since X
t
X
s
is independent of these increments, it is independent
of {X
t
1
, . . . , X
t
k
}. This is true for every nite set t
1
, . . . , t
k
and therefore
X
t
X
s
is independent of F
s
.
Let us carry out the following calculation very formally.
P[X
t
A | F
s
](w) = P[X
t
X
s
B | F
s
](w), B = A X
s
(w),
33
34 6. Markov Properties of Brownian Motion
= P[X
t
X
s
B], by independence,
i.e. 34
P[X
t
A | F
s
](w) =
_
A
1
(2t)
d/2
exp
|y X
s
(w)|
2
2(t s)
dy.
This formal calculation leads us to

Theorem .
P[X
t
A | F
s
](w) =
_
A
1
(2t)
d/2
exp
|y X
s
(w)|
2
2(t s)
dy.
where A is Borel in R
d
, t > s.
Remark. It may be useful to note that p(s, X
s
(w), t, y) can be thought of
as a conditional probability density.
Proof. (i) We show that
f
A
(w) =
_
A
1
(2t)
d/2
exp
|y X
s
(w)|
2
2(t s)
dy
is F
S
-measurable. Assume rst that A is bounded and Borel in
R
d
. If
n
, then f
A
(
n
) f
A
(), i.e. f
A
is continuous and
hence F
s
-measurable. The general case follows if we note that
any Borel set can be got as an increasing union of a countable
number of bounded Borel sets.
(ii) For any C F
s
we show that
(*)
_
X
C
X
1
t
(A)dP() =
_
C
_
A
exp |y X
s
()|
2
(2(t s))
d/2
dy dP().
It is enough to verify () for C of the form
C =
_
: (X
t
1
(), . . . , X
t
k
()) A
1
A
k
; 0 t
1
< . . . < t
k
s
_
,
35
where A
i
is Borel in R
d
for i = 1, 2 . . . k. The left side of () is
then
_
A
i
A
k
A
p(0, 0, t
1
, x
t
1
)p(t
1
, x
t
1
, t
2
, x
t
2
) . . . p(t
k
, x
t
k
, t, x
t
)dx
t
1
. . . dx
t
.
35
To compute the right side dene
f : R
(k+1)d
B
by
f (u
1
, . . . , u
k
, u) = X
A
1
(u
1
) . . . X
A
k
(u
k
)p(s, u, t, y).
Clearly f is Borel measurable. An application of Fubinis theorem
to the right side of () yields
_
A
dy
_
X
A
1
(X
t
1
()) . . . X
A
k
(X
t
k
())p(s, X
s
(), t, y)dP()
=
_
A
dy
_
R
d
R
d
(k+1) times
f (x
1
. . . x
k
, x
s
)dF
t
1
...t
k
, s
=
_
A
dy
_
A
1
A
k
R
d
p(0, 0, t
1
, x
1
) . . . p(t
k1
, t
k
, x
k
)
p(t
k
, x
k
, s, x
s
)p(s, x
s
, t, y)dx
1
. . . dx
k
dx
s
=
_
A
1
A
k
A
p(0, 0, t
1
, x
1
) . . . p(t
k1
, x
k1
, t
k
, x
k
)p(t
k
, x
k
, t, y)
dx
1
. . . , dx
k
dy
(by the convolution rule)
= left side.
Examples of Stopping Times.

36 6. Markov Properties of Brownian Motion
1. Let (X
t
) be a one-dimensional Brownian motion. Dene by
{ s} =
_
n=1
lim
1
,
2
1
,
2
rational in [0,s]
{w : w(
1
) D
n
, w(
2
) C
n
},
where 36
D
n
=
_
x R
d
: d(x, D)
1
n
_
, C
n
=
_
x R
d
: d(x, C)
1
n
_
Exercise 1. Let be as in Example 1.
(a) If A = {w : X
1
(w) } show that A F
.
(Hint: A { 0} F
0
). This shows that F
F
0
.
(b) P
0
{w : (w) = } = 0.
(Hint: P
0
{w : |w(t)| < 1}
_
|y|t
1/2
e
1/2|y|
2
dy t).
Theorem . (Strong Markov Property of Brownian Motion). Let be any
nite stopping time, i.e. < a.e. Let Y
t
= X
+t
X
. Then
1. P[(Y
t
1
A
1
, . . . , Y
t
k
A
k
) A] = P(X
t
1
A
1
, . . . X
t
k
A
k
) P(A),
A F
and for every A

i
Borel in R
d
. Consequently,
2. (Y
t
) is a Brownian motion.
3. (Y
t
) is independent of F
.
The assertion is that a Brownian motion starts afresh at every stop-
ping time.
Proof.
Step 1. Let take only countably many values, say s
1
, s
2
, s
3
. . .. Put
E
j
=
1
{s
j
}. Then each E
j
is F
-measure and
=
_
j=1
E
j
, E
j
E
i
= j i.
37
37
Fix A F
.
P[(Y
t
1
A
1
, . . . , Y
t
k
A
k
) A]
=
j=1
P[(Y
t
1
A
1
, . . . , Y
t
k
A
k
) A E
j
]
=
j=1
P[(X
t
1
+s
j
X
s
j
) A
1
, . . . , X
t
k
+s
j
X
s
j
A
k
) A E
j
]
=
j=1
P[(X
t
1
A
1
), . . . , (X
t
k
A
k
)]P(A E
j
)
(by the Markov property)
= P(X
t
1
A
1
, . . . , X
t
k
A
k
) P(A)
Step 2. Let be any stopping time; put
n
=
[n] + 1
n
. A simple cal-
culation shows that
n
is a stopping time taking only countably many
values. As
n
, F
n
. Let Y
(n)
t
= X
n
+t
X
n
.
By Step 1,
P[(Y
(n)
t
1
< x
1
, . . . , Y
(n)
t
k
< x
k
) A]
= P(X
t
1
< x
1
, . . . , X
t
k
< x
k
) P(A)
(where x < y means x
i
< y
i
i = 1, 2, . . . , d) for every A F
. As all
the Brownian paths are continuous, Y
(n)
t
Y
t
a.e. Thus, if x
1
, . . . , x
k
is
a point of continuity of the joint distribution of X
t
1
, . . . , X
t
k
, we have
P[(Y
t
1
< x
1
, . . . , Y
t
k
< x
k
) A] = P(X
t
1
< x
1
, . . . , X
t
k
< x
k
)P(A)
A F
. Now assertion (1) follows easily. 38

For (2), put A = in (1), and (3) is a consequence of (1) and (2).
7. Reection Principle
LET (X
t
) BE A one-dimensional Brownian motion. Then P( sup
0st
X
s
39
a) = 2P(X
t
a) with a > 0. This gives us the probability of a Brownian
particle hitting the line x = a some time less than or equal to t. The
intuitive idea of the proof is as follows.
Time,
Among all the paths that hit a before time t exactly half of them end
up below a at time t. This is due to the reection symmetry. If X
s
= a
for some s < t, reection about the horizontal line at a gives a one -
one correspondence between paths with X
t
> a and paths with X
t
< a.
Therefore
P
_
max
0st
X
s
a, X
t
> a
_
=
_
max
0st
X
s
a, X
t
< a
_
39
40 7. Reection Principle
Since P{X
t
= a} = 0, we obtain
P
_
sup
0st
X
s
a
_
= P
_
sup
0st
X
s
a, X
t
> a
_
+ P
_
sup
0st
X
s
a, X
t
> a
_
= 2P{X
t
a}
We shall now give a precise argument. We need a few elementary 40
results.
Lemma 1. Let X
n
=
n
_
k=1
Y
k
where the Y
k
are independent random vari-
ables such that P{Y
k
B} = P{Y
k
B} Borel set B R (i.e. Y
k
are
symmetric). Then for any real number a,
P
_
max
1in
X
i
> a
_
2P{X
n
> a}
Proof. It is easy to verify that a random variable is symmetric if and
only if its characteristic function is real. Dene
A
i
= {X
1
a, . . . X
i1
a, X
i
> a}, i = 1, 2, . . . , n;
B = {X
n
> a}
Then A
i
A
j
= if i j. Now,
P(A
i
B) P(A
i
{X
n
X
i
})
= P(A
i
)P(X
n
X
i
), by independence.
= P(A
i
)P(Y
i+1
+ + Y
n
0).
As Y
i+1
, . . . , Y
n
are independent, the characteristic function of Y
i+1
+
+ Y
n
is the product of the characteristic functions of Y
i+1
+ + Y
n
,
so that Y
i+1
+ + Y
n
is symmetric. Therefore
P(Y
i+1
+ + Y
n
0)
1
2
.
Thus P(A
i
B)
1
2
P(A
i
) and
P(B)
n
i=1
P(A
i
B)
1
2
P(A
i
)
1
2
P
_
_
n
_
i=1
A
i
_
_
,
41
i.e. 41
2P(B) P
_
_
n
_
i=1
A
i
_
_
,
or
P
_
max
1in
X
i
> a
_
2P{X
n
> a}
Lemma 2. Let Y
i
, . . . , Y
n
be independent random variables. Put X
n
=
n
_
k=1
Y
k
and let = min{i : X
i
> a}, a > 0 and = if there is no such i.
Then for each > 0,
(a) P{ n1, X
n
X
} P{ n1, X
n
a} +
n1
_
j=1
P(Y
j
> ).
(b) P{ n1, X
n
> a+2} P{ n1, X
n
X
> }+
n1
_
j=1
P{Y
j
> }
(c) P{X
n
> a + 2} P{ n 1, X
n
> a + 2} + P{Y
n
> 2}.
If, further, Y
1
, . . . , Y
n
are symmetric, then
(d) P{max
1in
X
i
> a, X
n
a} P{X
n
> a + 2} P{Y
n
2}
2
n1
_
j=1
P{Y
j
> }
(e) P{max
1in
X
i
> a} 2P{X
n
> a + 2} 2
n
_
j=1
P{Y
j
> }
Proof. (a) Suppose w { n 1, X
n
X
} and w {
n 1, X
n
a}. Then X
n
(w) > a and X
n
(w) + X
(w)
(w) or,
X
(w)
(w) > a + .
By denition of (w), X
(w)1
(w) a and therefore,
Y
(w)
(w) = X
(w)
(w) X
(w)1
(w) > a + a =
if (w) > 1; if (w) = 1, Y
(w)
(w) = X
(w)
(w) > a + > .
Thus Y
j
(w) > for some j n 1, i.e. 42
w
n1
_
j=1
{Y
j
> }.
Therefore
{ n 1, X
n
X
} { n 1, X
n
a}
n1
_
j=1
{Y
j
> }
and (q) follows.
(b) Suppose w { n1, X
n
> a+2} but w { n1, X
n
X
>
}. Then
X
n
(w) X
(w)
(w) ,
or, X
(w)
(w) > a+ so that Y
(w)
(w) > as in (a); hence Y
j
(w) >
for some j n 1. This proves (b).
(c) If w {X
n
> a + 2}, then (w) n; if w { n 1, X
n
>
a + 2}, then (w) = n so that X
n1
(w) a; therefore Y
n
(w) =
X
n
(w) X
n1
(w) > 2. i.e. w {Y
n
> 2}. Therefore
{X
n
> a + 2} { n 1, X
n
> a + 2} {Y
n
> 2}.
This establishes (c).
(d) P{max
1in
X
i
> a, X
n
a} = P{ n 1, X
n
a}
P{ n 1, X
n
X
}
n1
_
j=1
P(Y
j
> ), by (a),
P
_
n1
_
k=1
{ = k, X
n
X
k
}
_
n1
_
j=1
P(Y
j
> )
=
n1
_
k=1
P{ = k, X
n
X
k
}
n1
_
j=1
P(Y
j
> )
=
n1
_
k=1
P{ = k}P{X
n
X
k
}
n1
_
j=1
P(Y
j
> )
(by independence)
43
=
n
_
k=1
P{ = k}P{X
n
X
k
> }
n1
_
j=1
P(Y
j
> ) (by symmetry)
= P{ n 1, X
n
X
}
n1
_
j=1
P(Y
j
> )
P{ n 1, X
n
X
> }
n1
_
j=1
P(Y
j
> )
P{ n 1, X
n
> a + 2} 2
n1
_
j=1
P{Y
j
> } (by (b))
P{X
n
> a + 2} P{Y
n
> 2} 2
n1
_
j=1
P{Y
j
> } (by (c)) 43
This proves (d).
(e) P{max
1in
X
i
> a} = P{max
1in
X
i
> a, X
n
a} + P{max
1in
X
i
> a, X
n
> a}
= P{max
1in
X
i
> a, X
n
a} + P{X
n
> a}
= P{X
n
> a + 2} P{Y
n
> 2} + P{X
n
> a}
2
n1
_
j=1
P{Y
j
> } (by (d))
Since P{X
n
> a + 2} P{X
n
> a} and
P{Y
n
> 2} P{Y
n
> } 2P{Y
n
> },
we get
P{max
1in
X
i
> a} 2P{X
n
> a + 2} 2
n
j=1
P(Y
j
> )
This completes the proof. 44
Proof of the reection principle.

By Lemma 1
p = P
_
max
1jn
X
_
jt
n
_
>
_
2P(X(t) > a).
By Lemma 2(e),
p 2P(X(t) > a + 2) 2
n
j=1
P
__
X
_
jt
n
_
X
_
( j 1)t
n
__
>
_
.
Since X
_
jt
n
_
X
_
( j 1)t
n
_
are independent and identically distribu-
ted normal random variables with mean zero and variance
t
n
(in partic-
ular they are symmetric),
P
__
X
_
jt
n
_
X
_
( j 1)t
n
__
>
_
= P
__
X
_
t
n
_
X(0)
_
>
_
= P
_
X
_
t
n
_
>
_
.
Therefore
p 2P(X(t) > a + 2) 2n P
_
X
_
t
n
_
>
_
.
P(X(t/n) > ) =
(2t/n)
e
x
2
/
2t
n
dx
=
n/
(2t)
e
x
2
x
dx
n/
(2t)
1
_
n/
(2t)
xe
x
2
dx
or
P(X(t/n) > )
1
2
2
n/2t
(2t)
n
.
Therefore
nP(X(t/n) > )
n
2
(2t)/
(n)e
2
n/2t
0 as n +.
45
By continuity,
P
_
max
1n
X( jt/n) > a
_
P
_
max
0st
X(s) > a
_
.
45
We let n tend to through values 2, 2
2
, 2
3
, . . . so that we get
2P{X(t) > a + 2} 2n P{X(t/n) > }
P
_
max
1jn
X(t/n) > a
_
2P{X(t) > a},
or
2P{X(t) > a} 2P{X(t) a} P
_
max
0st
X(t) > a
_
2P{X(t) > a},
on letting n +rst and then letting 0. Therefore,
P
_
max
0st
X(s) > a
_
= 2P{X(t) > a}
= 2
_
a
1/
(2t)e
x
2
/2t
dx.
AN APPLICATION. Consider a one-dimensional Brownian motion. A
particle starts at 0. What can we say about the behaviour of the particle
in a small interval of time [0, )? The answer is given by the following
result.
P(A) P{w : > 0, t, s in [0, ) such that X
t
(w) > 0 and
X
s
(w) < 0} = 1.
INTERPRETATION. Near zero all the particles oscillate about their 46
starting point. Let
A
+
= {w : > 0t [0, ) such that X
t
(w) > 0},
A
= {w : > 0 s [0, ) such that X

s
(w) < 0}.
We show that P(A
+
) = P(A
) = 1 and therefore P(A) = P(A

+
) = 1.
A
+
_
n=1
_
_
sup
0t1/n
w > 0
_
_
=
_
n=1
_
m=1
_
_
sup
0t1/n
w(t) 1/m
_
_
Therefore
P(A
+
) Lt
n
sup
m
P
_
_
sup
0t1/n
w(t) 1/m
_
_
2 Lt
nm
sup P(w(1/n) 1/m) (by the reection principle)
1.
Similarly P(A
) = 1.
Theorem . Let {X
t
} be a one-dimensional Brownian motion, A (, a)
(a > 0) and Borel subset of R. Then
P
0
{X
t
A, X
s
< a s such that 0 s t}
=
_
A
1/
(2t)e
y
2
/2t
dy
_
A
1/
(2t)e
(2ay)
2
/2t
dy
Proof. Let (w) = inf{t : w(t) a}. By the strong Markov property of
Brownian motion,
P
0
{B(X( + s) X() A)} = P
0
(B)P
0
(X(s) A)
for every set B in F
t
. This can be written as
E(X
(X(+s)X()A)
|F
) = P
0
(X(s) A)
Therefore 47
E(X
(X(+(w))X()A)
|F
) = P
0
(X((w)) A)
for every function (w) which is F
-measurable. Therefore,
P
0
(( t) ((X( + (w)) X()) A) =
_
{t}
P
0
(X((w)) A)dP(w)
In particular, take (w) = t (w), clearly (w) is F
-measurable.
Therefore,
P
0
(( t)((X(t) X()) A)) =
_
{t}
P
0
(X((w) A)dP(w)).
47
Now X((w)) = a. Replace A by A a to get
(*) P
0
(( t) (X(t) A)) =
_
{t}
P
0
(X((w) A a)dP(w))
Consider now
P
2a
(X(t) A) = P
0
(X(t) A 2a)
= P
0
(X(t) 2a A) (by symmetry of x)
= P
0
(( t) (X(t) 2a A)).
The last step follows from the face that A (, a) and the conti-
nuity of the Brownina paths. Therefore
P
2a
(X(t) A) =
_
{t}
P
0
(X((w)) a A)dP(w), (using )
= P
0
(( t) (X(t) A)).
Now the required probability
P
0
{X
t
A, X
s
< a s 0 s t} = P
0
{X
t
A} P
0
{( t) (X
t
A)}
=
_
A
1/
(2t)e
y
2
/2t
dy
_
A
1/
(2t)e
(2ay)
2
/2t
dy.
The intuitive idea of the previous theorem is quite clear. To obtain 48
the paths that reach A at time t without hitting the horizontal line x = a,
we consider all paths that reach A at time t and subtract those paths that
hit the horizontal line x = a before time t and then reach A at time t. To
see exactly which paths reach A at time t after hitting x = a we consider
a typical path X(w).
The reection principle (or the strong Markov property) allows us

to replace this path by the dotted path (see Fig.). The symmetry of the
Brownian motion can then be used to reect this path about the line
x = a and obtain the path shown in dark. Thus we have the following
result:
the probability that a Brownian particle starts from x = 0 at t = 0
and reaches A at time t after it has hit x = a at some time t is the
same as if the particle started at time t = 0 at x = 2a and reached A at
time t. (The continuity of the path ensures that at some time t, this
particle has to hit x = a).
We shall use the intuitive approach in what follows, the mathemati- 49
cal analysis being clear, thorugh lengthy.
Theorem . Let X(t) be a one-dimensional Brownian motion, A (1, 1)
any Borel subset of R. Then
P
0
_
sup
0st
|X(s)| < 1, X(t) A
_
=
_
A
(t, y)dy,
where
(t, y) =
n=
(1)
n
/
(2t)e
(y2n)
2
/2t
.
49
Proof.
Let E
n
be the set of those trajections which (i) start at x = 0 at time
t = 0 (ii) hit x = +1 at some time
1
< t (iii) hit x = 1 at some later
time
2
< t (iv) hit x = 1 again at a later time
3
< t . . . and nally reach
A at time t. The number of s should be equal to n at least, i.e.
E
n
= {w : there exists a sequence
1
, . . .
n
of
stopping times such that 0 <
1
<
2
< . . . < t
n
< t, X(
j
) = (1)
j1
,
X(t) A}. Similarly, let
F
n
= {w : there exists a sequence
1
, . . . ,
n
of stopping times
0 <
1
<
2
< . . . <
n
< t, X(
j
) = (1)
j
, X(t) A}
Note that 50
E
1
E
2
E
3
. . . ,
F
1
F
2
F
3
. . . ,
F
n
E
n+1
; E
n
F
n+1
,
E
n
F
n
= E
n+1
F
n+1
.
Let
(t, A) = P
_
sup
0st
|X(s)| < 1, X(t) A
_
.
Therefore
(t, A) = P[X(t) A] P
_
sup
0st
|X(s)| 1, X(t) A
_
=
_
A
1/
(2t)e
y
2
/2t
dy P[(E
1
F
1
) A
0
],
where
A
0
= {X(t) A} =
_
A
1/
(2t)e
y
2
/2t
dy P[(E
1
A
0
) (F
1
A
0
)].
Use the fact that P[A B] = P(A) + P(B) P(A B) to get
(t, A) =
_
A
1/
(2t)e
y
2
/2t
dyP[E
1
A
0
]P[F
1
A
0
]+P[E
1
F
1
A
0
],
as E
1
F
1
= E
2
F
2
. Proceeding successively we nally get
(t, A) =
_
A
1/
(2t)e
y
2
/2t
dy+
n=1
(1)
n
P[E
n
A
0
]+
n=1
(1)
n
P[F
n
A
0
]
We shall obtain the expression for P(E
1
A
0
) and P[E
2
A
0
], the
other terms can be obtained similarly.
E
1
A
0
consists of those trajectries that hit x = 1 at some time
t and then reach A at time t. Thus P[E
1
A
0
] is given by the
previous theorem by
_
A
1/
(2t)e
(y2)
2
/2t
dy.
E
2
A
0
consists of those trajectories that hit x = 1 at time
1
, hit 51
x = 1 at time
2
and reach A at time t(
1
<
2
< t).
According to the previous theorem we can reect the trajectory upto
2
about x = 1 so that P(E
2
A
0
) is the same as if the particle starts at
x = 2 at time t = 0, hits x = 3 at time
1
and ends up in A at time t.
We can now reect the trajectory
51
upto time
1
(the dotted curve should be reected) about x = 3 to
obtain the required probability as if the trajectory started at x = 4.
Thus,
P(E
2
A
0
) =
_
A
e
(y+4)
2
/2t/
(2t)
dy.
Thus
(t, A) =
n=
(1)
n
_
A
1/
2te
(y2n)
2
/2t
dy
=
_
A
(t, y)dy.
The previous theorem leads to an interesting result: 52
P
_
sup
0st
|X(s)| < 1
_
=
1
_
1
(t, y)dy
Therefore
P
_
sup
0st
|X(s)| 1
_
= 1 P
_
sup
0st
|X(s)| < 1
_
= 1
1
_
1
(t, y)dy,
(t, y) =
n=
(1)
n
/
(2t)e
(y2n)
2
/2t
Case (i). t is very small.
In this case it is enough to consider the terms corresponding to n = 0,
1 (the higher order terms are very small). As y varies from 1 to 1,
(t, y) 1/
(2t)
_
e
y
2
/2t
e
(y2)
2
/2t
e
(y+2)
2
/2t
_
.
Therefore
1
_
1
(t, y)dy 4/
(2t)e
1/2t
.
Case (ii). t is large. In this case we use Poissons summation formula
for (t, y):
(t, y) =
k=0
e
(2k+1)
2
2
t/8
Cos{(k + 1)/2y},
to get
1
_
1
(t, y)dy 4/e
2
t/8
for large t. Thus, P( > t) = 4/e
2
t/8
. 53
This result says that for large values of t the probability of paths
which stay between 1 and +1 is very very small and the decay rate is
governed by the factor e
2
t/8
. This is connected with the solution of a
certain dierential equation as shall be seen later on.
8. Blumenthals Zero-One
Law
LET X
t
BE A d-dimensional Brownian motion. If A F
0+
=
_
t>0
F
t
, 54
then P(A) = 0 or P(A) = 1.
Interpretation. If an event is observable in every interval [0, t] of time
then either it always happens or it never happens.
We shall need the following two lemmas.
Lemma 1. Let (, B, P) be any probability space, C
a
sub-algebra of
B. Then
(a) L
2
(, C, P) is a closed subspace of L
2
(, B, P).
(b) If : L
2
(, B, P) L
2
(, C, P) is the projection map then f =
E( f |C).
Proof. Refer appendix.
Lemma 2. Let = C([0, ); R
d
), P
0
the probability corresponding to
the Brownian motion. Then the set {(
t
1
, . . . , t
k
) is continuous,
bounded on R
d
R
d
(k times),
t
1
, . . . , t
k
the canonical projection)
is dense in L
2
(, B, P).
Proof. Functions of the form (x(t
1
), . . . , x(t
k
) where runs over con-
tinuous functions is clearly dense in L
2
(, F
t
1
,t
2
,...,t
k
, P) and
_
k
_
t
1
,...,t
k
L
2
(, F
t
1
,...,t
k
, P)
53
54 8. Blumenthals Zero-One Law
is clearly dense in L
2
(, B, P).
Proof of zero-one law. Let 55
H
t
= L
2
(, F
t
, P), H = L
2
(, B, P), H
0+
=
_
t>0
H
t
.
Clearly H
0+
= L
2
(, F
0+
, P).
Let
t
: H H
t
be the projection. Then
t
f
0+
f
f
in H.
To prove the law it is enough to show that H
0+
contains only constants,
which is equivalent to
0+
f = constant f in H. As
0+
is continuous
and linear it is enough to show that
0+
= const of the Lemma 2:
0+
= Lt
t0
t
= Lt
t0
E(|
t
) by Lemma 1,
= Lt
t0
E((t
1
, . . . , t
k
)|F
t
).
We can assume without loss of generality that t < t
1
< t
2
< . . . < t
k
.
E((t
1
, . . . , t
k
)|F
t
) =
_
(y
1
, . . . , y
k
)1/
(2(t
1
t))e
|y
1
X
t
(w)|
2
/2(t
1
t)
. . .
. . . 1/
_
(2(t
k
t
k1
))e
|y
k
y
k1
|
2
2(t
k
t
k1
)
dy
1
. . . dy
k
.
Since X
0
(w) = 0 we get, as t 0,
0+
= constant.
APPLICATION. Let 1 A = {w :
1
_
0
|w(t)|/t
< }. Then A F
0+
.
For, if 0 < s < 1, then
1
_
s
|w(t)|/t
< . Therefore w A or not according

as
s
_
0
|w(t)|/t
dt converges or not. But this convergence can be asserted 56

55
by knowing the history of w upto time s. Hence A F
s
. Blumenthals
law implies that
1
_
0
|w(t)|/t
dt < a.e.w., or,

1
_
0
|w(t)|/t
dt = a.e.w.
A precise argument can be given along the following lines. If 0 <
s < 1,
A = {w :
s
_
0
|w(t)|/t
< }
= {w : sup I
n,s
(w) < }
where I
n,s
(w) is the lower Riemannian sum of |w(t)
n
|/t
corresponding
to the partition {0, s/n, . . . , s} and each I
n,s
F
s
.
9. Properties of Brownian
Motion in One Dimension
WE NOW PROVE the following. 57
Lemma . Let (X
t
) be a one-dimensional Brownian motion. Then
(a) P(limX
t
= ) = 1; consequently P(lim X
t
< ) = 0.
(b) P(limX
t
= ) = 1; consequently P(lim X
t
> ) = 0.
(c) P(limX
t
= ); lim X
t
= ) = 1.
SIGNIFICANCE. By (c) almost every Brownian path assumes each
value innitely often.
Proof.
{limX
t
= } =
_
n=1
(lim X
t
> n)
=
_
n=1
( lim
rational
X
> n) (by continuity of Brownian paths)

First, note that
P
0
_
sup
0st
X(s) n
_
= 1 P
0
_
sup
0st
X(s) > n
_
57
58 9. Properties of Brownian Motion in One Dimension
= 1 21/
(2t)
_
n
e
y
2
/2t
dy
=
(2/t)
n
_
0
e
y
2
/2t
dy.
Therefore, for any x
0
and t,
P
_
sup
t
0
st
X(s) n|X(t
0
) = x
0
_
= P
0
_
_
sup
0stt
0
X(s) n x
0
_
_
(independent increments) which tends to 1 as t . Consequently,
P
0
_
sup
tt
0
X(t) n
_
= EP
_
sup
tt
0
X(t) n|X(t
0
)
_
= E1 = 1.
58
In other words,
P
0
_
limsup
t
X(t) n
_
= 1
for every n. Thus
P(limX
t
= ) = 1.
(b) is clear if one notes that w w leaves the probability invariant.
(c) P(lim X
t
= , limX
t
= )
= P(limX
t
= ) P(lim X
t
> , limX
t
= ).
1 P(lim X
t
> )
= 1.
Corollary . Let (X
t
) be a d-dimensional Brownian motion. Then
P(lim|X
t
| = ) = 1.
59
Remark . If d 3 we shall see later that P( Lt
t
|X
t
| = ) = 1. i.e.
almost every Brownian path wanders o to .
Theorem . Almost all Brownian paths are of unbounded variation in
any interval.
Proof. Let I be any interval [a, b] with a < b. For n = 1, 2, . . . dene
V
n
(wQ
n
) =
n
i=1
|w(t
i
) w(t
i1
)| (t
i
= a + (b a)i/n, i = 0, 1, 2, . . . n),
The variation corresponding to the partioin Q
n
dividing [a, b] into n 59
equal parts. Let
U
n
(w, Q
n
) =
n
i=1
|(w(t
i
) w(t
i1
)|
2
.
If
A
n
(w, Q
n
) sup
1in
|w(t
i
) w(t
i1
)|,
then
A
n
(w, Q
n
)V
n
(w, Q
n
) U
n
(w, Q
n
).
By continuity Lt
n
A
n
(w, Q
n
) = 0.
Claim. Lt
n
E[(U
n
(w, Q
n
) (b a))
2
] = 0.
Proof.
E[(U
n
(b a))
2
]
= E
_
_
n
j=1
[(X
t
j
X
t
j1
)
2
(b a/n)]
_
_
2
E[(
(Z
2
j
b a/n))
2
], Z
j
= X
t
j
X
t
j1
,
= nE[(Z
2
1
b a/n)
2
]
60 9. Properties of Brownian Motion in One Dimension
(because Z
j
are independent and identically distributed).
= n[E(Z
4
1
) (b a/n)
2
] = 2(b a/n)
2
0.
Thus a subsequence U
n
i
ba almost everywhere. Since A
n
i
0
it follows that V
n
i
(w, Q
n
) almost everywhere. This completes the
proof.
Note . {w : w is of bounded variation on [a, b]} can be shown to be 60
measurable if one proves
Exercise. Let f be continuous on [a, b] and dene V
n
( f , Q
n
) as above.
Show that f is of bounded variation on [a, b] i sup
n=1,2,...
V
n
( f , Q
n
) < .
Theorem . Let t be any xed real number in [0, ), D
t
= {w : w is
dierentiable at t}. Then P(D
t
) = 0.
Proof. The measurability of D
t
follows from the following observation:
if f is continuous then f is dierentiable at t if and only if
Lt
r0
r rational
f (t + r) f (t)
r
.
exists. Now
D
t
=
_
m=1
w : |
w(t + h) w(t)
h
| M, for all h 0, rational}
and
P
_
w : |
X
t+h
X
t
h
| M h Q, h 0
_
2 inf
h
M
h
_
0
1
(2)
e
|y|
2/2
dy = 0
Remark. A stronger result holds:

P
_
_
_
t0
D
t
_
_
= 0.
61
Hint:
_
0t1
D
t
_
=1
_
m=1
n+2
_
nm i=1 k=i+1,i+2,i+3
_
w : w
_
k
n
_
w
_
k 1
n
_
|
71
n
_
and
P
_
_
n+2
_
i=1 k=i+1,...,i+3
_
w : w
_
k
n
_
w
_
k 1
n
_
|
71
n
_
const/
n
_
_
This construction is due to A. Dvoretski, P. Erdos & S. Kakutani.
10. Dirichlet Problem and
Brownian Motion
LET G BE ANY bounded open set in R
d
. Dene the exit time
G
(w) as 61
follows:
G
(w) = {inf t : w(t) G}.
If w(0) G,
G
(w) = 0; if w(0) G,
G
(w) is the rst time w escapes
G or, equivalently, it is the rst time that w hits the boundary G of G.
Clearly
G
(w) is a stopping time. By denition X
G
(w) G, w and
X
G
is a random variable. We can dene a Borel probability measure on
G by
G
(x, ) = P
x
(X
G
)
= probability that w hits .
If f is a bounded, real-valued measurable funciton dened on G,
we dene
u(x) = E
x
( f (X
G
)) =
_
G
f (y)
G
(x, dy)
where
E
x
= E
P
x
.
In case G is a sphere centred around x, the exact form of
G
(x, ) is
computable.
Theorem . Let S = S (0; r) = {y R
d
: |y| < r}. Then
S
(0, r) =
surface area of
surface area of S
.
63
64 10. Dirichlet Problem and Brownian Motion
Proof. The distributions {F
t
1
,...,t
k
} dening Brownian motion are invari- 62
ant under rotations. Thus
S
(0, ) is a rotationally invariant probability
measure. The result follows from the fact that the only probability mea-
sure (on the surface of a sphere) that is invariant under rotations is the
normalised surface area.
Theorem . Let G be any bounded region, f a bounded measurable real
valued function dened on G. Dene u(x) = E
x
( f (X
G
)). Then
(i) u is measurable and bounded;
(ii) u has the mean value property; consequently,
(iii) u is harmonic in G.
Proof. (i) To prove this, it is enough to show that the mapping x
P
x
(A) is measurable for every Borel set A.
Let C = {A B : x P
x
(A) is measurable}
It is clear that
1
t
1
,...,t
k
(B) C, Borel set B in R
d
R
d
. As
C is a monotone class C = B.
(ii) Let S be any sphere with centre at x, and S G. Let =
S
denote the exit time through S . Clearly
G
. By the strong
Markov property,
u(X
) = E( f (X
G
)|F
).
Now
u(x) = E
x
( f (X
G
)) = E
x
(E( f (X
G
))|F
)
= E
x
(u(X
)) =
_
S
u(y)
S
(x, dy)
=
1
|S |
_
S
u(y)dS ; |S | = surface area of S.
63
65
(iii) is a consequence of (i) and (ii). (See exercise below).
Exercise

. Let u be a bounded measurable function in a region G satis-
fying the mean value property, i.e.
u(x) =
1
|S |
_
S
u(y)dS
for every sphere S G. Then
(i) u(x) =
1
v lS
_
S
u(y)dy.
(ii) Using (i) show that u is continuous.
(iii) Using (i) and (ii) show that u is harmonic.
We shall now solve the boundary value problem under suitable con-
ditions on the region G.
Theorem . Let G, f , u be as in the previous theorem. Further suppose
that
(i) f is continuous;
(ii) G satises the exterior cone condition at every point of G, i.e.
for each y G there exists a cone C
h
with vertex at the point y
of height h and such that C
h
{y} exterior of G. Then
lim
xy,xG
u(x) = f (y), y G.
64
Proof.
Step 1. P
y
{w : w(0) = y, w remains in G for some positive time} = 0.
Let A
n
= {w : w(0) = y, w(s) G for 0 s 1/n},
B
n
= A
n
, A =
_
n=1
A
n
, B =
_
n=1
B
n
.
As A
n
s are increasing, B
n
s are decreasing and B
n
F
1/n
; so that
B F
0+
. We show that P(B) > 0, so that by Bluementhals zero-one
law, P(B) = 1, i.e. P(A) = 0.
P
y
(B) = lim
n
P
y
(B
n
) lim
n
P
y
{w : w(0) = y, w(
1
2n
) C
h
{y}}
Thus
P
y
(b) lim
_
C
h
{y}
1/
(2/2n)
d
exp(|z y|
2
/2/2n)dz
=
_
C
1/
(2)e
|y|
2
/2
dy,
where C
is the cone of innite height obtained from C

h
. Thus P
y
(B) >
0.
Step 2. If C is closed then the mapping x P
x
(C) is upper semi-
continuous.
For, denote by X
C
the indicator function of C. As C is closed (in
a metric space) a sequence of continuous functions f
n
decreasing to
X
C
such that 0 f
n
1. Thus E
x
( f
n
) decreases to E
x
(X
C
) = P
x
(C).
Clearly x E
x
(F
n
) is continuous. The result follows from the fact that
the inmum of any collection of continuous functions is upper semi-
continuous.
Step 3. Let > 0, 65
N(y; ) = {z G : |z y| < },
B
= {w : w(0) G, X
G
(w) G N(y; )},
i.e. B
consists of trajectories which start at a point of G and escape for

the rst time through G at a point not in N(y; ). If C
= B
, then
C
{w : w(0) = y} A {w : w(0) = y}
where A is as in Step 1.
67
For, suppose w C
{w : w(0) = y}. Then there exists w

n
B
such
that w
n
w uniformaly on compact sets. If w A{w : w(0) = y} there
exists > 0 such that w(t) Gt in (0, ]. Let
= inf
0t
d(w(t), G
N(y, )). Then
> 0. If t
n
=
G
(w
n
) and t
n
does not converge to 0,
then there exists a subsequence, again denoted by t
n
, such that t
n
k >
0 for some k (0, 1). Since w
n
(k) G and w
n
(k), w(k) G, a
contradiction. Thus we can assume that t
n
converges to 0 and also that
t
n
n, But then
(*) |w
n
(t
n
) w(t
n
)|
.
However, as w
n
converges to w uniformly on [0, ],
w
n
(t
n
) w(t
n
) w(0) w(0) = 0
contradicting (*). Thus w A{w : w(0) = y}.
Step 4. lim
xy,xG
P
x
(B
) = 0.
For, 66
lim
xy
P
x
(B
) lim
xy
P
x
(C
) P
y
(C
) (by Step 2)
= P
y
(C
{w : w(0) = y})
P
y
(A) (by Step 3)
= 0.
Step 5.
|u(x) f (y)| = |
_
f (X
G
(w))dP
x
(w)
_
f (y)dP
x
(w)|
_
B
| f (X
G
(w)) f (y)|dP
x
(w) + |
_
B
( f (X
G
(w)) f (y))dP
x
(w)|
_
B
| f (X
G
(w)) f (y)|dP
x
(w) + 2|| f ||
P
x
(B
)
and the right hand side converges to 0 as x y (by Step 4 and the fact
that f is continuous). This proves the theorem.
Remark. The theorem is local.

Theorem . Let G = {y R
d
: < |y| < R}, f any continuous function
on G = {|y| = } {|y| = R}. If u is any harmonic function in G with
boundary values f , then u(x) = E
x
( f (X
G
)).
Proof. Clearly G has the exterior cone property. Thus, if
v(x) = E
x
( f (X
G
)),
then v is harmonic in G and has boundary values f (by the previous
theorem). The result follows from the uniqueness of the solution of the
Dirichlet problem for the Laplacian operator.
The function f = 0 on |y| = R and f = 1 on |y| = is of spe-
cial interest. Denote by
R,0
,1
the corresponding solution of the Dirichlet
problem.
Exercise. (i) If d = 2 then 67
U
R,0
,1
(x) =
log R log |x|
log R log
, x G.
(ii) If d 3 then
U
R,0
,1
(x) =
|x|
n+2
R
n+2
n+2
R
n+2
.
Case (i): d = 2. Then
log R log |x|
log R log
= U
R,0
,1
(x).
Now,
E
x
( f (X
G
)) =
_
|y|=
G
(x, dy) = P
x
(|X
G
| = ),
i.e.
log R log |x|
log R log
= P
x
(|X
G
| = )
69
P
x
(the particle hits |y| = before it hits |y| = R).
Fix R and let 0; then 0 = P
x
(the particle hits 0 before hitting
|y| = R).
Let R take values 1, 2, 3, . . ., then 0 = P
x
(the particle hits 0 before
hitting any of the circles |y| = N). Recalling that
P
x
(lim|X
t
| = ) = 1,
we get
Proposition . A two-dimensional Brownian motion does not visit a
point.
Next, keep xed and let R , then,
1 = P
x
(|w(t)| = for some time t > 0).
Since any time t can be taken as the starting time for the Brownian 68
motion, we have
Proposition . Two-dimensional Brownian motion has the recurrence
property.
Case (ii): d 3. In this case
P
x
(w : w hits |y| = before it hits |y| = R)
= (1/|x|
n2
1/R
n2
)/(1/
n2
1/R
n2
).
Letting R we get
P
x
(w : w hits |y| = ) = (/|x|)
n2
which lies strictly between 0 and 1. Fixing and letting |x| , we
have
Proposition . If the particle start at a point for away from 0 then it has
very little chance of hitting the circle |y| = .
If |x| , then
P(w hits S
) = 1 where S
= {y R
d
: |y| = }.
Let
V
(x) = (/|x|)
n2
for |x| .
In view of the above result it is natural to extend V
to all space
by putting V
(x) = 1 for |x| . As Brownian motion has the Markov

property
P
x
{w : w hits S
after time t}
=
_
V
(y)1/
(2t)
d
exp |y|
2
/2t dy 0 as t +.
Thus P(w hits S
for arbitrarily large t) = 0. In other words, P(w : 69

lim
t
|w(t)| ) = 1. As this is true > 0, we get the following
important result.
Proposition . P( lim
t
|w(t)| = ) = 1,
i.e. for d 3, the Brownian particle wander away to +.
11. Stochastic Integration
LET {X
t
: t 0} BE A one-dimensional Brownian motion. We want 70
rst to dene integrals of the type
_
0
f (s)dX(s) for real functions f
L
1
[0, ). If X(s, w) is of bounded variation almost everywhere then we
can give a meaning to
_
0
f (s)dX(s, w) = g(w). However, since X(s, w)
is not bounded variation almost everywhere, g(w) is not dened in the
usual sense.
In order to dene g(w) =
_
0
f (s)dX(s, w) proceed as follows.
Let f be a step function of the following type:
f =
n
i=1
a
i
X
[t
i
,t
i+1
)
, 0 t
1
< t
2
< . . . < t
n+1
.
We naturally dene
g(w) =
_
0
f (s)dX(s, w) =
n
i=1
a
i
(X
t
i+1
(w) X
t
i
(w))
=
n
i=1
a
i
(w(t
i+1
) w(t
i
)).
g satises the following properties:
(i) g is a random variable;
71
72 11. Stochastic Integration
(ii) E(g) = 0; E(g
2
) =
_
a
2
i
(t
i+1
t
i
) = || f ||
2
.
This follows from the facts that (a) X
t
i+1
X
t
i
is a normal random
variable with mean 0 and variance (t
i+1
t
i
) and (b) X
t
i+1
X
t
i
are inde-
pendent increments, i.e. we have
E
_
_
0
f dX
_
_
= 0, E
_
_
|
_
0
f dX|
2
_
_
= || f ||
2
2
.
71
Exercise 1. If
f =
n
i=1
a
i
X
[t
i
,t
i+1
)
, 0 t
1
< . . . < t
n+1
,
g =
m
i=1
b
i
X
[s
i
,s
i+1
)
, 0 s
1
< . . . < s
m+1
,
Show that
_
0
( f + g)dX(s, w) =
_
0
f dX(s, w) +
_
0
gdX(s, w)
and
_
0
(f )dX(s, w) =
_
0
f dX(s, w), R.
Remark. The mapping f
_
0
f dX is therefore a linear L
2
R
-isometry of
the space S of all simple functions of the type
n
i=1
a
i
X
[t
i
,t
i+1
)
, (0 t
1
< . . . < t
n+1
)
into L
2
(, B, P).
73
Exercise 2. Show that S is a dense subspace of L
2
[0, ).
Hint: C
c
[0, ), i.e. the set of all continuous functions with compact sup-
port, is dense in L
2
[0, ). Show that S contains the closure of C
c
[0, ).
Remark. The mapping f
_
0
f dX can now be uniquely extended as
an isometry of L
2
[0, ) into L
2
(, B, P).
Next we dene integrals fo the type 72
g(w) =
t
_
0
X(s, w)dX(s, w)
Put t = 1 (the general case can be dealt with similarly). It seems
natural to dene
(*)
1
_
0
X(s, w)dX(s) = Lt
sup |t
j
t
j1
|0
n
j=1
X( j)(X(t
j
) X(t
j1
))
where 0 = t
0
< t
1
< . . . < t
n
= 1 is a partion of [0, 1] with t
j1

j
t
j
.
In general the limit on the right hand side may not exist. Even if it
exists it may happen that depending on the choice of
j
, we may obtain
dierent limits. To consider an example we choose
j
= t
j
and then
j
= t
j1
and compute the right hand side of (). If
j
= t
j1
,
n
j=1
X
j
(X
t
j
X
t
j1
) =
n
j=1
X
t
j1
(X
t
j
X
t
j1
)
=
1
2
n
j=1
(X
t
j
) (X
t
j1
)
1
2
n
j=1
(X
t
j
X
t
j1
)
1
2
[X
2
(1) X
2
(0)]
1
2
as n , and sup |t
j
t
j1
| 0,
arguing as in the proof of the result that Brownian motion is not of
bounded variation. If
j
= t
j
,
Lt
n
Sup |t
j
t
j1
|0
n
j=1
X
t
j
(X
t
j
X
t
j1
) = 1/2X(1) 1/2X(0) + 1/2.
Thus we get dierent answers depending on the choice of
j
and 73
hence one has to be very careful in dening the integral. It turns out
that the choice of
j
= t
j1
is more appropriate in the denition of the
integral and gives better results.
Remark. The limit in () should be understood in the sense of conver-
gence probability.
Exercise 3. Let 0 a < b. Show that the left integral (
j
= t
j1
) is
given by
L
b
_
a
X(s)dX(s) =
X
2
(b) X
2
(a) (b a)
2
and the right integral (
j
= t
j
) is given by
R
b
_
s
X(s)dX(s) =
X
2
(b) X
2
(a) + (b a)
2
.
We now take up the general theory of stochastic integration. To
motivate the denitions which follow let us consider a d-dimensional
Brownian motion {(t) : t 0}. We have
E[(t + s) (t) A|F
t
] =
_
A
1/
(2s)e
|y|
2
/2s
dy.
Thus
E( f ((t + s) (t))|F
t
] =
_
f (y)1/
(2s)e
|y|
2/2s
dy.
In particular, if f (x) = e
ix.u
,
E[e
iu
((t + s) (t))|F
t
] =
_
e
iu.y
1/
(2s)e
|y|
2
/2s
dy
= e
s|u|
2
2
.
Thus 74
75
E[e
iu.(t+s)
|F
t
] = e
iu.(t)
e
s|u|
2
/2
,
or,
E[e
iu.(t+s)+(t+s)|u|
2
/2
|F
t
] = e
iu.(t)+t|u|
2
/2
.
Replacing iu by we get
E[e
.(s)|s|
2
/2
| F
t
] = e
.(t)t||
2/2
, s > t, .
It is clear that e
.(t)t||
2
/2
is F
t
-measurable and a simple calculation
gives
E(e
.(t)||
2
t/2|
) < .
We thus have
Theorem . If {(t) : t 0} is a d-dimensional Brownian motion then
exp[.(t) ||
2
t/2] is a Martingale relative to F
t
, the -algebra gener-
ated by ((s) : s t).
Denition. Let (, B, P) be a probability space (F
t
)
t0
and increasing
family of sub--algebras of F with F = (
_
t0
F
t
).
Let
(i) a : [0, ) [0, ) be bounded and progressively measurable;
(ii) b : [0, ) R be bounded and progressively measurable;
(iii) X : [0, ) R be progressively measurable, right continuous
on [0, ), w , and continous on [0, ) almost everywhere
on ;
(iv) Z
t
(w) = e
X(t,w)
t
_
0
b(s,w)ds
2
2
t
_
0
a(s,w)ds
75
be a Martingale relative to (F
t
)
t0
.
Then X(t, w) is called an Ito process corresponding to the parameters
b and a and we write X
t
I[b, a].
N.B. The progressive measurability of X X
t
is F
t
-measurable.
Example. If {(t) : t 0} is a Brownian motion, then X(t, w) =
t
(w)
is an Ito process corresponding to parameters 0 and 1. (i) and (ii) are
obvious. (iii) follows by right continuity of
t
and measurability of
t
relative to F
t
and (iv) is proved in the previous theorem.
Exercise 4. Show that Z
t
(w) dened in (iv) is F
t
-measurable and pro-
gressively measurable.
[Hint:
(i) Z
t
is right continuous.
(ii) Use Fubinis theorem to prove measurability].
Remark. If we put Y(t, w) = X(t, w)
t
_
0
b(s, w)ds then Y(t, w) is pro-
gressively measurable and Y(t, w) is an Ito process corresponding to
the parameters 0, a. Thus we need only consider integrals of the type
t
_
0
f (s, w)dY(s, w) and dene
t
_
0
f (s, w)dX(s, w) =
t
_
0
f (s, w)dY(s, w) +
t
_
0
f (s, w)b(s, w)ds.
(Note that formally we have dY = dX dbt).
Lemma . If Y(t, w) I[0, a], then 76
Y(t, w) and Y
2
(t, w)
t
_
0
a(s, w)ds
are Martingales relative to (F
t
).
Proof. To motivate the arguments which follow, we rst give a formal
proof. Let
Y
(t) = e
Y(t,w)
2
2
t
_
0
a(s,w)ds
.
77
Then Y
(t) is a martingale, . Therefore

Y
is a Martingale, .
Hence (formally),
lim
0
Y
= Y
|
=0
is a Martingale.
Step 1. Y(t, ) L
k
(, F, P), k = 0, 1, 2, . . . and t. In fact, for every
real , Y
(t) is a Martingale and hence E(Y
) < . Since a is bounded

this means that
E(e
Y(t,)
) < , .
Taking = 1 and 1 we conclude that E(e
|Y|
) < and hence
E(|Y|
k
) < , k = 0, 1, 2, . . .. Since Y is an Ito process we also get
sup
||
E
_
_
_
_
e
Y(t,)
2
2
t
_
0
ads
_
_
k
_
_
<
k and for every > 0.
Step 2. Let X
(t) = [Y(t, )
t
_
0
ads]Y(t) =
d
d
Y
(t, ). 77
Dene
A
() =
_
A
(X
(t, ) X
(s, ))dP(w)
where t > s, A F
s
. Then
2
_
A
()d =
2
_
1
_
A
[X
(t, ) X
(S, )]dP(w)d.
Since a is bounded, sup
||
E([Y
(t, )]
k
) < , and E(|Y|
k
) < , k; we
can use Fubinis theorem to get
2
_
A
()d =
_
A
2
_
1
[X
(t, ) X
(s, )]d dP(w).

or
2
_
A
()d =
_
A
Y
2
(t, ) Y
1
(t, )dP(w)
_
A
Y
1
(s, ) Y
1
(s, )dP(w).
Let A F
s
and t > s; then, since Y is a Martingale,
2
_
A
()d = 0.
This is true
1
<
2
and since
A
() is a continuous function of ,
we conclude that
A
() = 0, .
In particular,
A
() = 0 which means that
_
A
Y(t, )dP(w) =
_
A
Y(s, )dP(w), A F
s
, t > s,
i.e., Y(t) is a Martingale relative to (, F
t
, P). 78
To prove the second part we put
Z
(t, ) =
d
2
d
2
Y
(t)
and
A
() =
_
A
{Z
(t, ) Z
(s, )}dP(w).
Then, by Fubini,
2
_
A
()d =
_
A
2
_
1
Z
(t, ) Z
(s, )d dP(w).
79
or,
2
_
A
()d =
A
(
2
)
A
(
1
)
= 0 if A F
s
, t > s.
Therefore
A
() = 0, .
In particular,
A
() = 0 implies that
Y
2
(t, w)
t
_
0
a(s, w)ds
is an (, F
t
, P) Martingale. This completes the proof of lemma 1.
Denition. A function : [0, ) R is called simple if there exist
reals s
0
, s
1
, . . . , s
n
, . . .
0 s
0
< s
1
< . . . < s
n
. . . < ,
s
n
increasing to + and
(s, w) =
j
(w)
if s [s
j
, s
j+1
), where
j
(w) is F
s
j
-measurable and bounded. 79
Denition. Let : [0, ) R be a simple function and Y(t, w)
I[0, a]. We dene the stochastic integral of with respect to Y, denoted
t
_
0
(s, w)dY(s, w)),
by
(t, w) =
t
_
0
(s, w)dY(s, w)
=
k
j=1
j1
(w)[Y(s
j
, w) Y(s
j1
, w)] +
k
(w)[Y(t, w) Y(s
k
, w)].
Lemma 2. Let : [0, ) R be a simple function and Y(t, w)
I[0, a]. Then
(t, w) =
t
_
0
(s, w)dY(s, w) I[0, a
2
].
Proof. (i) By denition, is right continuous and (t, w) is F
t
-
measurable; hence it is progressively measurable. Since a is pro-
gressively measurable and bounded
a
2
: [0, ) [0, )
is progressively measurable and bounded.
(ii) From the denition of it is clear that (t, ) is right continuous, 80
continous almost everywhere and F
t
-measurable therefore is
progressively measurable.
(iii) Z
t
(w) = e
[(t,w)
2
2
t
_
0
a
2
ds]
is clearly F
t
-measurable . We show that
E(Z
t
) < , t and E(Z
t
2
|F
t
1
) = Z
t
1
if t
1
< t
2
.
We can assume without loss of generality that = 1 (if 1 we
replace by ). Therefore
Z
t
(w) = e
[(t,w)
t
_
0
a
2
ds]
.
81
Since a and are bounded uniformly on [0, t], it is enough to show that
E(e
(t,w)
) < . By denition,
(t, w) =
k
j=1
j1
(w)[Y(s
j
, w) Y(s
j1
, w)] +
k
(w)(Y(t, w) Y(s
k
, w)).
The result E(e
(t,w)
) < will follow from the generalised Holders in-
equality provided we show that
E(e
(w)[Y(t,w)Y(s,w)]
) <
for every bounded function which is F
s
-measurable. Now
E(e
[Y(t,)Y(s,)]
|F
s
) =
constant for every constant , since Y I[0, a]. Therefore
E(e
(w)[Y(t,)Y(s,)]
|F
s
) = constant
for every which is bounded and F
s
-measurable. Thus 81
E(e
(w)[Y(t,)Y(s,)]
) < .
This proves that E(Z
t
(w)) .
Finally we show that
E(Z
t
2
|F
t
1
) = Z
t
1
(w), if t
1
< t
2
.
Consider rst the case when t
1
and t
2
are in the same interval
[s
k
, s
k+1
).
Then
(t
2
, w) = (t
1
, w) +
k
(w)[Y(t
2
, w) Y(t
1
, w)] (see denition),
t
2
_
0
a
2
(s, w)ds =
t
1
_
0
a
2
(s, w)ds +
t
2
_
t
1
a
2
(s, w)ds.
Therefore
E(Z
t
2
(w)|F
t
1
) = Z
t
1
(w)E(exp[
k
(w)[Y(t
2
, w) Y(t
1
, w)]
2
2
t
2
_
t
1
a
2
ds)|F
t
1
)
as Y I[0, a].
(*) E(exp[(Y(t
2
, w) T(t
1
, w))
2
2
t
2
_
t
1
a(s, w)ds]|F
t
1
) = 1
and since
k
(w) is F
t
1
-measurable () remains valid if is replaced by
k
. Thus
E(Z
t
2
|F
t
1
) = Z
t
1
(w).
The general case follows if we use the identity
E(E(X|C
1
)|C
2
) = E(X|C
2
) for C
2
C
1
.
Thus Z
t
is a Martingale and (t, w) I[0, a
2
].
Corollary . (i) (t, w) is a martingale; E((t, w)) = 0; 82
(ii)
2
(t, w)
t
_
0
a
2
ds
is a Martingale with
E(
2
(t, w)) = E(
t
_
0
a
2
(s, w)ds.
Proof. Follows from Lemma 1.
Lemma 3. Let (s, w) be progressively measurable such that for each
t,
E(
t
_
0
2
(s, w)ds) < .
83
Then there exists a sequence
n
(s, w) of simple functions such that
lim
n
E
_
_
t
_
0
|
n
(s, w) (s, w)|
2
ds
_
_
= 0.
Proof. We may assume that is bounded, for if
N
= for || N
and 0 if || > N, then
n
, (s, w) [0, t] .
N
is progressively
measurable and |
n
|
2
4||
2
. By hypothesis L([0, t] : ).
Therefore E(
t
_
0
|
n
|ds) 0, by dominated convergence. Further,
we can also assume that is continuous. For, if is bounded, dene
h
(t, w) = 1/h
t
_
(th)v0
(s, w)ds.
n
is continuous in t and F
t
-measurable and hence progressively mea-
surable. Also by Lebesgues theorem
h
(t, w) (t, w), as h 0, t, w.
Since is bounded by C,
h
is also bounded by C. Thus 83
E(
t
_
0
|
h
(s, w) (s, w)|
2
ds) 0.
(by dominated convergence). If is continuous, bounded and progres-
sively measurable, then
n
(s, w) =
_
[ns]
n
, w
_
is progressively measurable, bounded and simple. But
Lt
n
n
(s, w) = (s, w).
Thus by dominated convergence
E
_
_
t
_
0
|
n
|
2
ds
_
_
0 as n .
Theorem . Let (s, w) be progressively measurable, such that

E(
t
_
0
2
(s, w)ds) <
for each t > 0. Let (
n
) be simple approximations to as in Lemma 3.
Put
n
(t, w) =
t
_
0
n
(s, w)dY(s, w)
where Y I[0, a]. Then
(i) Lt
n
n
(t, w) exists uniformly in probability, i.e. there exists an al-
most surely continuous (t, w) such that
Lt
n
P
_
sup
0tT
|
n
(t, w) (t, w)|
_
= 0
for each > 0 and for each T. Moreover, is independent of the
sequence (
0
).
(ii) The map is linear. 84
(iii) (t, w) and
2
(t, w)
t
_
0
a
2
ds are Martingales.
(iv) If is bounded, I[0, a
2
].
85
Proof. (i) It is easily seen that for simple functions the stochastic
integral is linear. Therefore
(
n

m
)(t, w) =
t
_
0
(
n

m
)(s, w)dY(s, w).
Since
n

m
is an almost surely continuous martingale
P
_
sup
0tT
|
n
(t, w)
m
(t, w)|
_

1
2
E[(
n

m
)
2
(T, w)].
This is a consequence of Kolmogorov inequality (See Appendix).
Since
(
n

m
)
2
t
_
0
a(
n

m
)
2
ds
is a Martingale, and a is bounded,
E[(
n

m
)
2
(T, w)] = E
_
_
T
_
0
(
n

m
)
2
a ds
_
_
. (*)
const
1
2
E
_
_
T
_
0
(
n

m
)
2
ds
_
_
.
Therefore
Lt
n,m
E[(
n

m
)
2
(T, w)] = 0.
Thus (
n
m
) is uniformly Cauchy in probability. Therefore there
exists a progressively measurable such that
Lt
n
P
_
sup
0tT
|
n
(t, w) (t, w)|
_
= 0, > 0, T.
It can be shown that is almost surely continuous.
If (
n
) and (
n
) are two sequences of simple functions approxi- 85
mating , then () shows that
E[(
n

n
)
2
(T, w)] 0.
Thus
Lt
n

n
= Lt
n

n
,
i.e. is independent of (
n
).
(ii) is obvious.
(iii) (*) shows that
n
in L and therefore
n
(t, ) (t, ) in L
1
for
each xed t. Since
n
(t, w) is a martingale for each n, (t, w) is a
martingale.
(iv)
2
n
(t, w)
t
_
0
a
2
n
is a martingale for each n.
Since
n
(t, w) (t, w) in L
2
for each xed t and
2
n
(t, w)
2
(t, w) in L
1
for each xed t.
For
2
n
(t, w)
2
(t, w) = (
n
)(
n
+ ) and using H olders in-
equality, we get the result.
Similarly, since
n
in L
2
([0, t] ),
2
n

2
in L
1
([0, t] ),
and because a is bounded a
2
n
a
2
in L
1
([0, t]). This shows
that
2
n
(t, w)
t
_
0
a
2
n
ds converges to
2
(t, w)
t
_
0
a
2
ds
87
for each t in L
1
. Therefore
2
(t, w)
t
_
0
a
2
ds
is a martingale. 86
(v) Let be bounded. To show that I[0,
2
] it is enough to show
that
e
(t,w)
2
2
t
_
0
a
2
ds
is a martingale for each , the other conditions being trivially sat-
ised. Let
Z
n
(t, w) = e
n
(t,w)
2
2
t
_
0
a
2
n
ds
We can assume that |
n
| C if || C (see the proof of Lemma
3).
Z
n
= exp
_
_
2
n
(t, w)
(2)
2
2
t
_
0
a
2
n
ds +
2
t
_
0
a
2
n
ds
_
_
.
Thus
(**) E(Z
n
) const E
_
_
e
2
n
(t,w)
(2)
2
2
t
_
0
a
2
n
ds
_
_
= const
since Z
n
is a martingale for each . A subsequence Z
n
i
converges
to
e
(t,w)
2
2
t
_
0
a
2
ds
almost everywhere (P). This together with (**) ensures uniform
integrability of (Z
n
) and therefore
e
(t,w)
2
2
t
_
0
a
2
ds
is a martingale. Thus is an Ito process, I[0, a
2
].
Denition. With the hypothesis as in the above theorem we dene the

stochastic integral
(t, w) =
t
_
0
(s, w)dY(s, w).
87
Exercise. Show that d(X + Y) = dX + dY.
Remark. If is bounded, then satises the hypothesis of the previ-
ous theorem and so one can dene the integral of with respect to Y.
Further, since itself is It o, we can also dene stochastic integrals with
respect to.
Examples. 1. Let {(t) : t 0} be a Brownian motion; then (t, w) is
progressively measurable (being continuous and F
t
-measurable).
Also,
_
t
_
0
2
(s)ds dP =
t
_
0
_
2
(s)dP ds =
t
_
0
sds =
t
2
Hence
t
_
0
(s, w)d(s, w)
is well dened.
2. Similarly
t
_
0
(s/2)d(s) is well dened.
3. However
t
_
0
(2s)d(s)
is not well dened, the reason being that (2s) is not progressively
measurable.
89
Exercise 5. Show that (2s) is not progressively measurable. 88
(Hint: Try to show that (2s) is not F
s
-measurable for every s. To show
this prove that F
s
F
2s
).
Exercise 6. Showthat for a Brownian motion (t), the stochastic integral
1
_
0
(s, )d(s, )
is the same as the left integral
L
1
_
0
(s, )d(s, )
dened earlier.
12. Change of Variable
Formula
WE SHALL PROVE the 89
Theorem . Let be any bounded progressively measurable function
and Y be an Ito process. If is any progressively measurable function
such that
E
_
_
t
_
0
2
ds
_
_
< , t,
then
(*)
t
_
0
d(s, w) =
t
_
0
(s, w)(s, w)dY(s, w),
where
(t, w) =
t
_
0
(s, w)dY(s, w).
Proof.
Step 1. Let and be both simple, with bounded. By a renement
of the partition, if necessary, we may assume that there exist reals 0 =
s
0
, s
1
, . . . , s
n
, . . . increasing to + such that and are constant on
[s
j
, s
j+1
), say =
j
(w), =
j
(w), where
j
(w) and
j
(w) are F
s
j
-
measurable. In this case (*) is a direct consequence of the denition.
91
92 12. Change of Variable Formula
Step 2. Let be simple and bounded. Let (
n
) be a sequence of simple
bounded functions as in Lemma 3. Put
n
(t, w) =
t
_
0
n
(s, w)dY(s, w)
By Step 1,
(**)
t
_
0
d
n
=
t
_
0
n
dY(s, w).
90
Since is bounded,
n
converges to in L
2
([0, t] ). Hence,
by denition,
t
_
0
n
dY(s, w) converges to
t
_
0
dY in probability.
Further,
t
_
0
d
n
(s, w) = (s
0
, w)[
n
(s
1
, w)
n
(s
0
, w)] +
+ + (s
k
, w)[
n
(t, w)
n
(s
k1
, w)],
where s
0
< s
1
< . . . . . . is a partition for , and
n
(t, w) converges to
(t, w) in probability for every t. Therefore
t
_
0
d
n
(s, w)
converges in probability to
t
_
0
d(s, w).
Taking limit as n in (**) we get
t
_
0
d(s, w) =
t
_
0
dY(s, w).
93
Step 3. Let be any progressively measurable function with
E(
t
_
0
2
ds) < , t.
Let
n
be a simple approximation to as in Lemma 3. Then, by Step
2,
(***)
t
_
0
n
(s, w)d(s, w) =
t
_
0
n
(s, w)(s, w)dY(s, w).
By denition, the left side above converges to
t
_
0
(s, w)d(s, w)
in probability. As is bounded
n
converges to in L
2
([0, t] ). 91
Therefore
P
_
_
sup
0tT
|
t
_
0
n
dY(s, w)
t
_
0
dy(s, w)|
_
_
||a||
1/
2
E
_
_
t
_
0
(
n
)
2
ds
_
_
(see proof of the main theorem leading to the denition of the stochastic
integral). Thus
t
_
0
n
dY(s, w)
converges to
t
_
0
dY(s, w)
in probability. Let n tend to + in (***) to conclude the proof.
94 12. Change of Variable Formula
13. Extension to
Vector-Valued It o Processes
Denition. Let (, F, P) be a probability space and (F
t
) an increasing 92
family of sub -algebras of F. Suppose further that
(i) a : [0, ) S
d
+
is a probability measurable, bounded function taking values in the class
of all symmetric positive semi-denite d d matrices, with real entries;
(ii) b : [0, ) R
d
is a progressively measurable, bounded function;
(iii) X : [0, ) R
d
is progressively measurable, right continuous for every w and continu-
ous almost everywhere (P);
Z(t, ) = exp[, X(t, )
t
_
0
, b(s, )ds
1
2
t
_
0
, a(s, )ds] (iv)
is a martingale for each R
d
, where
x, y = x
1
y
1
+ + x
d
y
d
, x, y R
d
.
95
96 13. Extension to Vector-Valued It o Processes
Then X is called an It o process corresponding to the parameters b
and a, and we write X I[b, a]
Note. 1. Z(t, w) is a real valued function.
2. b is progressively measurable if and only if each b
i
is progres-
sively measurable.
3. a is progressively measurable if and only if each a
i j
is so. 93
Exercise 1. If X I[b, a], then show that
X
i
I[b
i
, a
ii
], (i)
Y =
d
i=1
i
X
i
I[, b, , a], (ii)
where
= (
1
, . . . ,
d
).
(Hint: (ii) (i). To prove (ii) appeal to the denition).
Remark. If X has a multivariate normal distribution with mean and
covariance (
i j
), then Y = , X has also a normal distribution with
mean , and variance , . Note the analogy with the above exer-
cise. This analogy explains why at times b is referred to as the mean
and a as the covariance.
Exercise 2. If {(t) : t 0} is a d-dimensional Brownian motion, then
(t, w) I[0, I] where I = d d identity matrix.
As before one can show that Y(t, ) = X(t, )
t
_
0
b(s, w)ds is an It o
process with parameters 0 and a.
Denition. Let X be a d-dimensional Ito process. = (
1
, . . . ,
d
) a
d-dimensional progressively measurable function such that
E
_
_
t
_
0
(s, ), (s, ) > ds
_
_
97
is nite or, equivalently,
E
_
_
t
_
0
2
i
(s, )ds
_
_
< , (i = 1, 2, . . . d).
94
Then by denition
t
_
0
(s, ), dX(s, ) =
d
i=1
t
_
0
i
(s, )dX
i
(s, ).
Proposition . Let X be a d-dimensional It o process X I[b, a] and let
be progressively measurable and bounded. If
i
(t, ) =
t
_
0
i
dX
i
(s, ),
then
= (
1
, . . . ,
d
) I[B, A],
where
B = (
1
b
1
, . . . ,
d
b
d
) and A
i j
=
i
j
a
i j
.
Proof. (i) Clearly A
i j
is progressively measurable and bounded.
Since a S
d
+
, A S
d
+
.
(ii) Again B is progressively measurable and bounded.
(iii) Since is bounded, each
i
(t, ) is an It o process; hence is pro-
gressively measurable, right continuous, continuous almost ev-
erywhere (P). It only remains to verify the martingale condition.
Step 1. Let = (
1
, . . . ,
d
) R
d
. By hypothesis,
E(exp[(
1
X
1
+ +
d
X
d
)|
t
s

t
_
s
(
1
b
1
+ +
d
b
d
)du (*)
1
2
t
_
0
j
a
i j
ds]|F
s
) = 1.
Assume that each
i
is constant on [s, t],
i
=
i
(w) and F
s
- 95
measurable. Then (*) remains true if
i
are replaced by
i
i
(w) and since
i
s are constant over [s, t], we get
E(exp[
t
_
0
d
i=1
i
(s, )dX
i
(s, )
t
_
0
i
b
i
i
(s, )ds
1
2
t
_
0
i
(s, )
j
(s, )a
i j
ds]|
s
)
exp
_
_
s
_
0
d
i=1
i
(s, )dX
i
(s, )
s
_
0
, Bdu 1
s
_
0
, Adu
_
_
.
Step 2. Let each
i
be a simple function.
By considering ner partitions we may assume that each
i
is a step
function,
finest partition
i.e. there exist points s
0
, s
1
, s
2
, . . . , s
n
, s = s
0
< s
1
< . . . < s
n+1
= t,
such that on [s
j
, s
j+1
) each
i
is a constant and
s
j
-measurable. Then (**)
holds if we use the fact that if C
1
C
2
.
E(E( f |C
1
)|C
2
) = E( f |C
2
).
99
Step 3. Let be bounded, || C. Let (
(n)
) be a sequence of sim-
ple functions approximating as in Lemma 3. (**) is true if
i
is re- 96
placed by
(n)
i
for each n. A simple verication shows that the expres-
sion Z
n
(t, ), in the parenthes is on the left side of (**) with
i
replaced
by
(n)
i
, converges to
Z(t, ) =
= Exp
_
t
_
0
i
(s, )ds
t
_
0
i
b
i
i
(s, )ds
1
2
t
_
0
i, j
j
a
i j
ds
_
as n in probability. Since Z
n
(t, ) is a martingale and the functions
i
,
j
, a are all bounded,
sup
n
E(Z
n
(t, )) < .
This proves that Z(t, ) is a martingale.
Corollary . With the hypothesis as in the above proposition dene
Z(t) =
t
_
0
(s, ), dX(s, ).
Then
Z(t, ) I[, b, a
]
where
is the transpose of .
Proof. Z(t, ) =
1
(t, ) + +
d
(t, ).
Denition. Let (s, w) = (
i j
(s, w)) be a n d matrix of progressively
measurable functions with
E
_
_
t
_
0
2
i j
(s, )ds
_
_
< .
If X is a d-dimensional It o process, we dene 97
_
_
t
_
0
(s, )dX(s, )
_
_
i
=
d
j=1
t
_
0
i j
(s, )dX
j
(s, ).
Exercise 3. Let
Z(t, w) =
t
_
0
(s, )dY(s, ),
where Y I[0, a] is a d-dimensional It o process and is as in the above
denition. Show that
Z(t, ) I[0, a
]
is an n-dimensional Ito process, (assume that is bounded).
Exercise 4. Verify that
E(|Z(t)|
2
) = E
_
_
t
_
0
tr(a
)ds
_
_
.
Exercise 5. Do exercise 3 with the assumption that a
is bounded.
Exercise 6. State and prove a change of variable formula for stochastic
integrals in the case of several dimensions.
(Hint: For the proof, use the change of variable formula in the one di-
mensional case and d(X + Y) = dX + dY).
14. Brownian Motion as a
Gaussian Process
SO FAR WE have been considering Brownian motion as a Markov pro- 98
cess. We shall now show that Brownian motion can be considered as a
Gaussian process.
Denition. Let X (X
1
, . . . , X
N
) be an N-dimensional random variable.
It is called an N-variate normal (or Gaussian) distribution with mean
(
1
, . . . ,
N
) and covariance A if the density function is
1
(2)
N/2
1
(det A)
1/2
exp
_
1
2
[(X )A
1
(X )
]
_
where A is an N N positive denite symmetric matrix.
Note. 1. E(X
i
) =
i
.
2. Cov(X
i
, X
j
) = (A)
i j
.
Theorem . X (X
1
, . . . , X
N
) is a multivariate normal distribution if and
only if for every R
N
, , X is a one-dimensional Gaussian random
variable.
We omit the proof.
Denition. A stochastic process {X
t
: t I} is called a Gaussian process
if t
1
, t
2
, . . . , t
N
I, (X
t
1
, . . . , X
t
N
) is an N-variate normal distribution.
101
102 14. Brownian Motion as a Gaussian Process
Exercise 1. Let {X
t
: t 0} be a one dimensional Brownian motion.
Then show that
(a) X
t
is a Gaussian process. 99
(Hint: Use the previous theorem and the fact that increments are
independent)
(b) E(X
t
) = 0, t, E(X(t)X(s)) = s t.
Let : [0, 1] = [0, 1] R be dened by
(s, t) = s t.
Dene K : L
2
R
[0, 1] L
2
R
[0, 1] by
K f (s) =
1
_
0
(s, t) f (t)dt.
Theorem . K is a symmetric, compact operator. It has only a countable
number of eigenvalues and has a complete set of eigenvectors.
We omit the proof.
Exercise 2. Let be any eigenvalue of K and f an eigenvector belonging
to . Show that
(a) f
+ f = 0 with f (0) = 0 = f
(1).
(b) Using (a) deduce that the eigenvalues are given by
n
= 4/(2n +
1)
2
2
and the corresponding eigenvectors are given by
f
n
=
2 Sin 1/2[(2n + 1)t]n = 0, 1, 2, . . . .

Let Z
0
, Z
1
, . . . , Z
n
. . . be identically distributed, independent, normal
random variables with mean 0 and variance 1. Then we have
Proposition . Y(t, w) =
_
n=0
Z
n
(w) f
n
(t)
n
converges in mean for every real t.
103
Proof. Let Y
m
(t, w) =
m
_
i=0
Z
i
(w) f
i
(t)
i
. Therefore 100
E{(Y
n+m
(t, ) Y
n
(t, ))
2
} =
n+m
n+1
f
2
i
(t)
i
,
E(||Y
n+m
() Y
n
()||
2
n+m
n+1
i
0.
Remark. As each Y
n
(t, ) is a normal random variable with mean 0 and
variance
n
_
i=0
i
f
2
i
(t), Y(t, ) is also a normal random variable with mean
zero and variance

_
i=0
i
f
2
i
. To see this one need only observe that the
limit of a sequence of normal random variables is a normal random vari-
able.
Theorem (Mercer).
(s, t) =
i=0
i
f
i
(t) f
i
(s), (s, t) [0, 1] [0, 1].
The convergence is uniform.
We omit the proof.
Exercise 3. Using Mercers theorem show that {X
t
: 0 t 1} is a
Brownian motion, where
X(t, w) =
n=0
Z
n
(w) f
n
(t)
n
.
This exercise now implies that
1
_
0
X
2
(s, w)ds = (L
2
norm of X)
2
104 14. Brownian Motion as a Gaussian Process
=
n
Z
2
n
(w),
since f
n
(t) are orthonormal. Therefore 101
E(e
1
_
0
X
2
(s,)ds
) = E(e
_
n=0
n
Z
2
n
(w)
) =
_
n=0
E(e
n
Z
2
n
)
(by independence of Z
n
)
=
_
n=0
E(e
n
Z
2
0
)
as Z
0
, Z
n
. . . are identically distributed. Therefore
E(e
1
_
0
X
2
(s,)ds
) =
_
n=0
1/
(1 + 2
n
)
=
_
n=0
1/
_
1 +
8 8
(2n + 1)
2
2
_
= 1/
(cosh)
(2).
APPLICATION. If F(a) = P(
1
_
0
X
2
(s)ds < a), then
_
0
e
a
dF(a) =
e
a
dF(a)
= E(e
_
1
0
X
2
(s)ds
) = 1/
(cosh)
(2).
15. Equivalent For of It o
Process
LET (, F, P) BE A probability space with (F
t
)
t0
and increasing fam- 102
ily of sub -algebras of F such that (U F
t
t0
) = F. Let
(i) a : [0, ) S
+
d
be a progressively measurable, bounded func-
tion taking values in S
+
d
, the class of all dd positive semidenite
matrices with real entries;
(ii) b : [0, ) R
d
be a bounded, progressively measurable
function;
(iii) X : [0, ) R
d
be progressively measurable, right continu-
ous and continuous a.s. (s, w) [0, ) .
For (s, w) [0, ) dene the operator
L
s,w
=
1
2
d
i, j=1
a
i j
(s, w)

2
x
i
x
j
+
d
j=1
b
j
(s, w)

x
j
.
For f , u, h belonging to C
0
(R
d
), C
0
([0, ) R
d
) and C
1,2
b
([0, )
R
d
) respectively we dene Y
f
(t, w), Z
u
(t, w), P
h
(t, w) as follows:
Y
f
(t, w) = f (X(t, w))
t
_
0
(L
s,w
( f )(X(s, w))ds,
105
106 15. Equivalent For of It o Process
Z
u
(t, w) = u(t, X(t, w))
t
_
0
_
u
s
+ L
s,w
u
_
(s, X(s, w))ds,
P
h
(t, w) = exp[h(t, X(t, w))
t
_
0
_
h
s
+ L
s,w
h
_
(s, X(s, w)ds
1
2
_
t
0
a(s, w)
x
h(s, X(s, w)),
x
h(s, X(s, w))ds].
Theorem . The following conditions are equivalent. 103
(i) X
(t, w) = exp[, X(t, w)

t
_
0
, b(s, w)ds
t
_
0
, a(s, w)ds]
is a martingale relative to (, F
t
, P), R
d
.
(ii) X
(t, w) is a martingale
in R
d
. In particular X
i
(t, w) is a mar-
tingale R
d
.
(iii) Y
f
(t, w) is a martingale for every f C
0
(R
d
)
(iv) Z
u
(t, w) is a martingale for every u C
0
([0, ) R
d
).
(v) P
h
(t, w) is a martingale for every h C
1,2
b
[(0, ) R
d
).
(vi) The result (v) is true for functions h C
1,2
([0, ) R
d
) with
linear growth, i.e. there exist constants A and B such that |h(x)|
A|x| + B.
The functions
h
t
,
h
x
i
, and

2
h
x
i
x
j
which occur under the integral
sign in the exponent also grow linearly.
Remark. The above theorem enables one to replace the martingale con-
dition in the denition of an It o process by any of the six equivalent
conditions given above.
Proof. (i) (ii). X
(t, ) is F
t
-measurable because it is progressively mea-
surable. That E(|X
(t, w)|) < is a consequence of (i) and the fact that

a is bounded.
107
The function
(t, w)
X
(s, w)
is continuous for xed t, s, w, (t > s).
Moreras theorem shows that is analytic. Let A F
s
. Then
_
A
X
(t, w)
X
(s, w)
dP(w)
is analytic. By hypothesis, 104
_
A
X
(t, w)
X
(s, w)
dP(w) = 1, R
d
.
Thus
_
A
X
(t, w)
X
(s, w)
dP(w) = 1, complex . Therefore
E(X
(t, w)|F
s
) = X
(s, w),
proving (ii). (ii) (iii). Let
A(t, w) = exp
_
_
i
t
_
0
, b(s, w)ds +
1
2
t
_
0
, a(s, w)ds
_
_
, R
d
.
By denition, A is progressively measurable and continuous. Also
|
dA
dt
(t, w)| is bounded on every compact set in R and the bound is in-
dependent of w. Therefore A(t, w) is of bounded variation on every in-
terval [0, T] with the variation ||A||
[0,T]
bounded uniformly in w. Let
M(t, w) = X
i
(t, w). Therefore
sup
0tT
|M(t, w)| e
1/2 T
sup
0tT
|, a|.
By (ii) M(t, ) is a martingale and since
E
_
sup
0tT
|M(t, w)| ||A||
[0,T]
(w)
_
< , T,
M(t, )A(t, )
1
2
t
_
0
M(s, )dA(s, )
is a martingale (for a proof see Appendix), i.e. Y
f
(t, w) is a martingale
when f (x) = e
i,x
.
Let f C
0
(R
d
). Then f F(R
d
) the Schwartz-space. Therefore
by the Fourier inversion theorem
f (x) =
_
R
d
f ()e
i,x
d.
105
On simplication we get
Y
f
(t, w) =
_
R
d
f ()Y
(t, w)d
where Y
Y
e
i, x. Clearly Y
f
(t, ) is progressively measurable and
hence F
t
-measurable.
Using the fact that
E(|Y
(t, w)|) 1 + t d|| ||b||
+
d
2
2
||
2
||a||
,
the fact that F(R
d
) L
1
(R
d
) and that F(R
d
) is closed under multi-
plication by polynomials, we get E(|Y
f
(t, w)|) < . An application of
Fubinis theorem gives E(Y
f
(t, w)|F
s
) = Y
f
(s, w), if t > s. This proves
(iii).
(iii) (iv). Let u C
0
([0, ) R
d
).
Clearly Z
u
(t, ) is progressively measurable. Since Z
u
(t, w) is boun-
ded for every w, E(|Z
u
(t, w)|) < . Let t > s. Then
E(Z
u
(t, w) Z
u
(s, w)|F
s
) =
= E(u(t, X(t, w) u(s, X(s, w)|F
s
) E(
t
_
s
(
u
+ L
,w
u)(, X(, w)d|F
s
)
= E(u(t, X(t, w) u(t, X(s, w))|F
s
) + E(u(t, X(s, w) u(s, X(s, w))|F
s
)
E(
t
_
s
(
u
+ L
u
w
)(, X(, w))d|F
s
)
109
= E(
t
_
s
(L
,w
u)(t, X(, w))d|F
s
)+
+ E(
t
_
s
(
u
(, X(s, w))d|F
s
)
E(
t
_
s
(
u
+ L
u
, w)(, X(, w))d|F
s
), by (iii)
= E(
t
_
s
[L
,w
u(t, X(, w)) L
,w
u(, X(, w))]d|F
s
)
+ E(
t
_
s
[
u
(, X(s, w))
u
(, X(, w))]d|F
s
)
= E(
t
_
s
(L
,w
u(t, X(, w)) L
,w
u(, X(, w))]d|F
s
)
E(
t
_
s
d
_
s
L
,w
u
(, X(, w))d|F
s
)
106
The last step follows from (iii) (the fact that > s gives a minus
sign).
= E(
t
_
0
d
t
_
L
,w
u(, X(, w))d|F
s
)
E(
t
_
s
d
_
s
L
,w
u
(, X(, w))d|F
s
)
= 0
(by Fubini). Therefore Z
u
(t, w) is a martingale.
Before proving (iv) (v) we show that (iv) is true if u C
1,2
b
([0, )
R
d
. Let u C
1,2
b
.
(*) Assume that there exists a sequence (u
n
) C
0
[[0, ) R
d
] such
that
u
n
u,
u
n
t

u
t
,
u
n
x
i
u
x
i
,
u
n
x
i
x
j

2
u
x
i
x
j
uniformly on compact sets. 107
Then Z
u
n
Z
u
pointwise and sup
n
(|Z
u
n
(t, w)|) < .
Therefore Z
u
is a martingale. Hence it is enough to justify (*).
For every u C
1,2
b
([0, ) R
d
) we construct a u C
1,2
b
((, )
R
d
) C
1,2
b
(R R
d
) as follows. Put
u(t, x) =
_
_
u(t, x), if t 0,
C
1
u(t, x) + C
2
u(
t
2
, x), if t < 0;
matching
u
t
,
u
t
at t = 0 and u(t, x) and u(t, x) at t = 0 and u(t, x) and
u(t, x) at t = 0 yields the desired constants C
1
and C
2
. In fact C
1
= 3,
C
2
= 4. (*) will be proved if we obtain an approximating sequence for
u. Let S : R be any C function such that if |x| 1,
S (x) =
_
_
1, if |x| 1,
0, if |x| 2.
Let S
n
(x) = S
_
|x|
2
n
_
where |x|
2
= x
2
1
+ + x
2
d+1
. Pur u
n
= S
n
u.
This satises (*).
(iv) (v). Let
h C
1,2
b
([0, ) R
d
).
Put u = exp(h(t, x)) in (iv) to conclude that
M(t, w) = e
h(t,X(t,w))
t
_
0
e
h(s,X(s,w))
_
h
s
+ L
s,w
h +
1
2
x
h, a
x
hds
_
is a martingale.
Put 108
111
A(t, w) = exp
_
_
t
_
0
h
s
(s, w) + L
s,w
(s, w) +
1
2
a(s, w)
x
h,
x
hds
_
_
.
A((t, w)) is progressively measurable, continuous everywhere and
||A||
[0,T]
(w) C
1
C
2
T
where C
1
and C
2
are constants. This follows from the fact that |
dA
dt
| is
uniformly bounded in w. Also sup
0tT
|M(t, w)| is uniformly bounded in
w. Therefore
E( sup
0tT
|M(t, w)| ||A||
[0,T]
(w)) < .
Hence M(t, )A
t
_
0
M(s, )dA(s, ) is a martingale. Now
dA(s, w)
A(s, w)
=
_
h
s
(s, w) + L
s,w
h(s, w) +
1
2
a
x
h,
x
h
_
Therefore
M(t, w) = e
h(t,X(t,w))
+
t
_
0
e
h(s,X(s,w))
dA(s, w)
A(s, w)
.
M(t, w)A(t, w) = P
h
(t, w) + A(t, w)
t
_
0
e
h(s,X(s,w))
dA(s, w)
A(s, w)
t
_
0
M(s, )dA(s, ) =
t
_
0
e
h(s,X(s,w))
dA(s, w)
+
t
_
0
dA(s, w)
s
_
0
e
h(,X(,w))
dA(, w)
A(, w)
Use Fubinis theorem to evaluate the second integral on the right
above and conclude that P
h
(t, w) is a martingale.
(vi) (i) is clear if we take h(t, x) = , x. It only remains to prove
that (v) (vi).
(v) (vi). The technique used to prove this is an important one and
we shall have occasion to use it again.
Step 1. 0 Let h(t, x) =
1
x
1
+
2
x
2
+
d
x
d
= , x for every (t, x) 109
[0, ) R
d
, is some xed element of R
d
. Let
Z(t) = exp
_
_
, X
t

t
_
0
, bds
1
2
t
_
0
, ads
_
_
We claim that Z(t, ) is a supermartingale.
Let f : R R be a C
function with compact support such that

f (x) = x in |x| 1/2 and | f (x)| 1, x. Put f
n
(x) = nf (x/n). Therefore
| f
n
(x)| C|x| for some C independent of n and x and f
n
(x) converges to
x.
Let h
n
(x) =
d
_
i=1
i
f
n
(x
i
). Then h
n
(x) converges to , x and |h
n
(x)|
C
|x| where C is also independent of n and x. By (v),

Z
n
(t) = exp
_
_
h
n
(t, X
t
)
t
_
0
_
h
n
s
+ L
s,w
h
_
ds
1
2
t
_
0
a
x
h
n
,
x
h
n
ds
_
_
is a martingale. As h
n
(x) converges to , x, Z
n
(t, ) converges to Z(t, )
pointwise. Consequently
E(Z(t)) = E(limZ
n
(t)) lim E(Z
n
(t)) = 1
and Z(t) is a supermartingale.
Step 2. E(exp B sup
0st
|X(s, w)|) < for each t and B. For, let Y(w) =
sup
0st
|X(s, w)|, Y
i
(w) = sup
0st
|X
i
(s, w)| where X = (X
1
, . . . , X
d
). Clearly
Y Y
1
+ + Y
d
. Therefore
E(e
BY
) E(e
BY
1e
BY
2 . . . e
BY
d).
113
The right hand side above is nite provided E(e
BY
i) < for each i 110
as can be seen by the generalised Holders inequality. Thus to prove the
assertion it is enough to show E(e
BY
i) < for each i = 1, 2, . . . d with
aB
dierent from B; more specically for B
bounded.
Put
2
= 0 =
3
= . . . =
d
in Step 1 to get
u(t) = exp[
1
X
1
(t)
t
_
0
1
b
1
(s, )ds
1
2
2
1
t
_
0
a
11
(s, )ds]
is a supermartingale. Therefore
P
_
sup
0st
u(s, )
_

1
E(u(t)) =
1
, > 0.
(Refer section on Martingales). Let c be a common bound for both b
1
and a
11
and let
1
> 0. Then () reads
P
_
sup
0st
exp
1
X
1
(s) exp(
1
ct +
1
2
2
1
ct)
_

1
.
Replacing by
e
1
e
ct
1
1/2ct
2
1
we get
P
_
sup
0st
exp
1
X
1
(s) exp
1
_
e
1
+
1
ct+1/2
2
1
ct
,
i.e.
P
_
sup
0st
X
1
(s)
_
e
1
+
1
ct+1/2
2
1
ct
,
1
> 0.
Similarly
P
_
sup
0st
X
1
(s)
_
e
1
+
1
ct+1/2
2
1
tc
,
1
> 0.
As
{Y
1
(w) }
_
sup
0st
X
1
(s)
_
_
sup
0st
X
1
(s)
_
,
we get 111
P{Y
1
} 2e
1
+
1
ct+1/2
2
1
ct
,
1
> 0.
Now we get
E(exp BY
1
) =
1
B
_
0
exp(Bx)P(Y
1
x)dx (since Y
1
0)
2
B
_
0
exp(Bx x
1
+
1
ct +
1
2
2
1
ct)dx
< , if B <
1
This completes the proof of step 2.
Step 3. Z(t, w) is a martingale. For
|Z
n
(t, w)| = Z
n
(t, w)
= exp
_
_
h
n
(X
t
)
t
_
0
_
h
n
s
+ L
s,w
h
n
_
dx
1
2
t
_
0
a
x
h
n
,
x
h
n
ds
_
_
exp
_
_
h
n
(X
t
)
t
_
0
L
s,w
h
n
_
_
(since a is positive semidenite and h
n
/s = 0).
Therefore |Z
n
(t, w)| Aexp(Bsup
0 s t
|X(s, w)|) (use the fact that
|h
n
(s)| C|x| and
h
n
x
i
,

2
h
n
x
i
x
j
are bounded by the same constant). The
result now follows from the dominated convergence theorem and Step
2.
Remark. In Steps 1, 2 and 3 we have proved that (v) (i). The idea of
the proof was to express Z(t, ) as a limit of a sequence of martingales 112
proving rst that Z(t, ) is a supermartingale. Using the supermartingale
inequality it was then shown that (Z
n
) is a uniformly integrable family
proving thereby that Z(t, ) is a martingale.
115
Step 4. Let h(t, x) C
1,2
([0, )R
d
) such that h(t, x),
h
s
(t, x),
h
x
i
(t, x),
2
h
x
i
x
j
(t, x) are all dominated by |x| + for some suitable scalars and
. Let
n
be a sequence of real valued C
functions dened on R
d
such
that
n
=
_
_
1 on |x| n
0 on |x| 2n
and suppose there exists a common bound C for
n
,

n
x
i
,

2
n
x
i
x
j
(n).
Let h
n
(t, x) = h(t, x)
n
(x). By (v) Z
h
n
(t, w) is a martingale. The
conditions on the function h and
n
s show that
|Z
h
n
(t, w)| Aexp
_
B sup
0st
|X(s, w)|
_
where A and B are constants. By Step 2, (Z
h
n
) are uniformly integrable.
Also Z
h
n
(t, ) converges pointwise to P
h
(t, ) (since h
n
h pointwise).
By the dominated convergence theorem P
h
(t, ) is a martingale, proving
(vi).
16. It os Formula
Motivation. Let (t) be a one-dimensional Brownian motion. We have 113
seen that the left integral
(*) L
_
_
2
t
_
0
(s, )d
_
_
= [
2
(t, )
2
(0, ) t]
Formally (*) can be written as
d
2
(t) = 2(t)d(t) + dt.
For, on integrating we recover (*).
Newtonian calculus gives the result:
d f ((t)) = f
((t))d(t) +
1
2
f
((t))d
2
(t) +
for reasonably smooth functions f and . If is of bounded variation,
only the rst term contributes something if we integrate the above equa-
tion. This is because
_
d
2
= 0 for a function of bounded variation.
For the Brownian motion we have seen that
_
d
2
a non zero value,
but one can prove that
_
d
3
, . . . converge to 0. We therefore expect the
following result to hold:
d f ((t)) f
((t))d(t) +
1
2
f
((t))d
2
(t).
We show that for a one-dimensional Brownian motion
(d)
3
, sup(d)
4
, . . .
117
118 16. It os Formula
all vanish.
E(
(d)
3
) = E
_
_
n
i=0
[(t
i+1
) (t
i
)]
3
_
_
=
i=0
E[((t
i+1
) (t
i
)]
3
=
n
i=1
0 = 0,
because (t
i+1
) (t
i
) is a normal random variable with mean zero and 114
variance t
i+1
t
i
. Similarly the higher odd moments vanish. Even mo-
ments of a normal random variable with mean 0 and variance
2
are
connected by the formula
2k+2
=
2
(2k + 1)
2k
, k > 1.
So
(d)
4
=
n
i=0
((t
i+1
) (t
i
))
4
.
Therefore
E(
(d)
4
) =
n
i=0
E([(t
i+1
) (t
i
))
4
]
= 3
n
i=0
(t
i+1
t
i
)
2
;
the right hand side converges to 0 as the mesh of the partition goes to 0.
Similarly the higher order even moments vanish.
More generally, if (t, ) is a d-dimensional Brownian motion then
we expect
d f ((t)) f ((t)) d(t) +
1
2
i, j
2
f
x
i
x
j
d
i
d
j
.
119
However
_
d
i
d
j
= 0 if i j (see exercise below). Therefore
d f ((t)) f ((t)) d(t) +
1
2
f ((t))d
2
(t).
The appearance of on the right side above is related to the heat
equation.
Exercise 4. Check that d
i
d
j
=
i j
dt. 115
(Hint: For i = j, the result was proved earlier. If i j, consider a
partition 0 = t
0
< t
1
< . . . < t
n
= t. Let
k
i
=
i
(t
k
)
i
(t
k1
). Then
E
_
_
n
k=1
j
_
_
2
=
k
E(
2
k
2
k
j
) + 2
kl
E[(
k
i
)(
1
j
)];
the right side converges to 0 as n because
k
i
and
j
are inde-
pendent for k ).
Before stating It os formula, we prove a few preliminary results.
Lemma 1. Let X(t, ) I[b, 0] be a one-dimensional It o process. Then
X(t, ) X(0, ) =
t
_
0
b(s, )ds a.e.
Proof. exp[X(t, ) X(0, )
t
_
0
b(s, )ds] is a martingale for each .
Therefore
E(exp[(X(t, ) X(0, ))
t
_
0
b(s, )ds]) = constant = 1, t.
Let
W(t, ) = X(t, ) X(0, )
t
_
0
b(s, )ds.
Then
E(exp W(t, )) = Moment generating function of w = 1, t.
Therefore (t, ) = 0 a.e.
Remark . If X(t, ) I[0, 0] then X(t, ) = X(0, ) a.e.; i.e. X(t, ) is a 116
trivial process.
We now state a theorem, which is a particular case of the theorem
on page 103.
Theorem . If h C
1,2
([0, ) R
d
) such that (i) |h(x)| A|x| + B,
x [0, ) R
d
, for constants A and B (ii)
h
t
,
h
x
i
,

2
h
x
i
x
j
also grow
linearly, then
exp[h(t, (t, )
t
_
0
_
h
s
+
1
2
h
_
(s, (s, )
1
2
t
_
0
|h|
2
(s, (s, ))ds]
is a martingale.
Itos Formula. Let f C
1,2
0
([0, ) R
d
) and let (t, ) be a d-dimensio-
nal Brownian motion. Then
f (t, (t)) f (0, (0)) =
t
_
0
f
s
(s, (s, ))ds+
+
t
_
0
f (s, (s, )), d(s, ) +
1
2
t
_
0
f (s, (s, ))ds.
where

2
x
2
1
+ +
2
x
2
d
.
Proof.
Step 1. Consider a (d + 1)-dimensional process dened by
X
0
(t, ) = f (t, (t, )),
X
j
(t, ) =
j
(t, ).
121
We claim that X(t, ) (X
0
, X
1
, . . . , X
d
) is a (d +1)-dimensional It o-
process with parameters
b =
__
f
s
+
1
2
f
_
(s, (s, )), 0, 0, . . . 0
_
d terms
and 117
a =
_
_
a
00
a
01
. . . a
0d
a
10
.
.
.
I
dd
a
d0
_
_
where
a
00
= |
x
f |
2
(s, (s, )),
a
0j
=
_

x
j
f
_
(s, (s, )),
a
j0
= a
0j
.
For, put h = f (t, x) + , x, x = (x
1
, . . . , x
d
) R
d
, in the previous
theorem. Then
h
s
=
f
s
, h = f ,
h
x
j
=
f
x
j
+
j
.
Therefore we seen that
exp[f (t, (t, )) + , (t, )
t
_
0
_
f
s
+
1
2
x
f
_
(s, (s, )ds
1
2
2
t
_
0
|f |
2
(s, (s, ))ds
1
2
||
2
t ,
t
_
0
( f (s, (s, )))ds]
is a martingale.
Consider (, )a
_
_
. We have
a
_
_
=
_
a
00
+
_
d
j=1
a
0j
j
+
_
, =
_
_
a
10
.
.
.
a
d0
_
_
.
Therefore
(, )a
_
_
= a
00
2
+
d
j=1
a
0j
j
+
d
j=1
a
0j
j
+
=
2
|f |
2
+ 2
f
x
j
j
+ ||
2
.
118
Thus (*) reads
exp[f (t, (t, ) + , (t, )
t
_
0
b
0
(s, )ds
1
2
t
_
0
, ads]
is a martingale where = (, ) R
d+1
. This proves the claim made
above.
Step 2. Derine (s, ) = (1,
x
f (s, (s, ))) and let
Z(t, ) =
t
_
0
(s, ), dX(s, ) where X (X
0
, X
1
, . . . , X
d
)
is the (d+1)-dimensional It o process obtained in Step 1. Since f C
1,2
b
,
Z(t, ) is an It o process with parameters , b and a
:
, b =
f
s
+
1
2
f ,
a
=
_
a
00

I
_ _
1
x
f
_
=
_
a
00
,
x
f

x
f
_
=
_
_
0
0
.
.
.
0
_
_
Therefore a
= 0. Hence by Lemma 1,
Z(t, ) Z(0, )
t
_
0
, bds = 0(a.e.),
123
Z(t, ) =
t
_
0
dX
0
(s)
t
_
0
f (s, (s, )), d(s, )
= f (t, (t)) f (0, (0))
t
_
0
f ( f , (s, )), d(s, )
119
Hence Z(0) = 0. Thus
f (t, (t)) f (0, (0))
t
_
0
f (s, (s, ))d(s, )
t
_
0
_
f
s
+
1
2
x
f
_
(s, (s, ))ds = 0 a.e.
This estabilished It os formula.
Exercise. (It os formula for the general case). Let

(t, x) C
1,2
b
([0, ) R
d
).
If X(t, ) is a d-dimensional It o process corresponding to the param-
eters b and a, then the following formula holds:
(t, X(t, w)) (0, X(0, w))
=
t
_
0
s
(s, X(s, x))ds +
t
_
0
x
, dX +
1
2
t
_
0
a
i j
2
x
i
x
j
ds.
This is also written as
d(s, X(s, w)) =
s
ds +
x
, dX +
1
2
a
i j
x
i
x
j
dx.
To prove this formula proceed as follows.
(i) Take h(t, x) = (t, x) + , x in (vi) of the theorem on the equiv-
alence of It o process to conclude that
Y(t, ) = ((t, X(t, )), X(t, ))
is a (d + 1)-dimensional Ito process with parameters
b
=
_
t
+ L
s,w
, b
_
and 120
A =
a
x
,
x
, a
x
1 1 1 d
a
x
a
d 1 d d
(ii) Let (t, x) = (1,

x
(t, x)) and
Z(t, ) =
t
_
0
(s, X(s, )), dY(s, ).
The assumptions on imply that Z is an It o process corresponding
to
(, b
, A
)
_
t
+
1
2
a
i j
x
i
x
j
, 0
_
.
(iii) Use (ii) to conclude that
Z(t, ) =
t
_
0
, b
ds a.e.
This is It os formula.
(iv) (Exercise) Verify that It os formula agrees with the formula ob-
tained for the case of Brownian motion.
Note. Observe that It os formula does not depend on b.
125
Examples. 1. Let (t) be a one-dimensional Brownian motion. Then
d(e
t
((t)) = e
t
d((t)) + ((t))d(e
t
)
= e
t
((t))d(t) + ((t))e
t
dt+
+
1
2
((t))e
t
dt.
2. To evaluate d(e
_
t
e
V((s,))ds
u(t, (t))) where V is a smooth function, 121
put
X
2
(t, ) =
t
_
0
V((s, ))ds, b =
_
0
V((t, ))
_
=
_
b
1
b
2
_
a =
_
1 0
0 0
_
. Let X
1
(t, ) = (t, ), X = (X
1
, X
2
).
Then
exp[
1
X
1
(t, )
2
X
2
(t, )
1
t
_
0
b
1
(X(s, ))ds

2
t
_
0
b
2
(X(s, ))ds
1
2
t
_
0
a, ds]
exp[
1
X
1
(t, )
2
1
2
2
t];
the right side is a martingale. Therefore (X
1
, X
2
) is a 2-dimensio-
nal It o process with parameters b and a and one can use It os
formula to write
d
_
e
_
t
0
V((s,))ds
u(t, (t))
_
= d
_
e
X
2
(t)
u(t, (t))
_
= e
_
t
0
V((s,))

t
u(t, (t))dt+
+e
_
t
0
V((s,))dt

x
u(t, (t))d(t) + e
_
t
0
V((s,))ds
u(t, (t))dX
2
+
1
2
e
_
t
0
V((s,))ds

t
u(t, (t))dt
= e
_
t
0
V((s,))dt

t
u(t, (t))dt +
t
u(t, (s))d(t) +
1
2
2
x
2
u(t, (t))dt
+u(t, (t))V((t, )dt].
3. Let
i
, f
i
, i = 1, 2, . . . k be bounded and progressively measurable 122
relative to a one-dimensional Brownian motion (, F
t
, P). Write
X
i
(t, ) =
t
_
0
i
(s, )d(s, ) +
t
_
0
f
i
(s, )ds.
Then X
i
(t, ) is an It o process with parameters (
t
_
0
f
i
(s, )ds,
2
i
) and
(X
1
, . . . , X
k
) is an It o process with parameters
B =
_
_
t
_
0
f
1
(s, )ds, . . . ,
t
_
0
f
k
(s, )ds
_
_
,
A = (A
i j
) where A
i j
=
i
i j
.
If (t, X
1
(t) . . . , X
k
(t)), then by It os formula
d =
s
ds +
x
1
dX
1
+ +
x
k
dX
k
+
1
2
i j
x
i
x
j
ds
=
s
ds +
x
1
dX
1
+ +
x
k
dX
k
+
1
2
k
i=1
2
i
x
i
ds
Exercise. Take
1
= 1,
2
= 0, f
1
= f
2
= 0 above and verify that if
= e
X
2
(t,)
u(t, (t)),
then one gets the result obtained in Example 2 above.
127
We give below a set of rules which can be used to calculate d in
practice, where is as in Example 3 above.
1. With each d associate a term

(dt)
2. If = (t, X
1
, . . . , X
k
), formally dierentiate using ordinary 123
calculus retaining terms upto the second order to get
(*) d =
t
dt +
x
1
dX
1
+ +
x
k
X
k
+
1
2
x
i
x
j
dX
i
dX
j
3. Formally write dX
i
= f
i
dt +
i
d
i
, dX
j
=
j
d
j
+ f
j
dt.
4. Multiply dX
i
dX
j
and retain only the rst order term in dt., For
d
i
d
j
substitute
i j
dt. Substitute in (*) to get the desired formula.
Illustration of the use of It o Calculus. We refer the reader to the sec-
tion on Dirichlet problem. There it was shown that
u(x) =
_
G
u(y)(x, dy) = E(u(X()))
satises u = 0 in a region G with u = u(X()) on the boundry of G
(here is the rst hitting time).
The form of the solution is given directly by It os formula (without
having recourse to the mean value property). If u = u(X(t)) satises
u = 0 then by It os formula
du(X(t)) = u, dX.
Therefore
u(X(t)) = u(X(0)) +
t
_
0
u(X(s)), dX(s)
124
Assuming u to be bounded, we see that u(X(t)) is a martingale. By
the optional stopping theorem
E(u(X()) = u(X(0)) = u(x).
Thus It os formula connects solutions of certain dierential equa-
tions with the hitting probabilities.
17. Solution of Poissons
Equations
LET X(t, ) BE A d-dimensional Brownian motion with (, F
t
, P) as 125
usual. Let u(x) : R
d
R be such that
1
2
u = f . Assume u C
2
b
(R
d
).
Then L
s,w
u
1
2
u = f and we know that u(X(t, ))
t
_
0
f (X(s, ))ds is
a (, F
t
, P)-martingale. Suppose now that u(x) is dened only on an
open subset G R
d
and
1
2
u = f on G. We would like to consider
Z(t, ) = u(X(t, ))
t
_
0
f (X(s, ))ds
and ask whether Z(t, ) is still a martingale relative to (, F
t
, P). Let
(w) = inf{t, X(t, w) G}. Put this way, the question is not well-posed
because Z(t, ) is dened only upto time (w) for u is not dened outside
G. Even if at a time t > (w)X(t, w) G, one needs to know the values
of f for t > (w) to compute the integral.
To answer the question we therefore proceed as follows. Let A
t
=
[w : (w) > t]. As t increases, A
t
have decreasing measures. We shall
give a meaning to the statement Z(t, ) is a martingale on A
t
. Dene
Z(t, ) = u(X( t, ))
t
_
0
f (X(s, ))ds.
129
130 17. Solution of Poissons Equations
Therefore
Z(t, ) =
_
_
Z(t), on A
t
Z(, ), on (A
t
)
c
.
Since Z(t, ) is progressively measurable upto time , Z(t, ) is F
t
- 126
measurable.
Theorem . Z(t, ) is a martingale.
Proof. Let G
n
be a sequence of compact sets increasing to G such that
G
n
G
0
n+1
. Choose a C
function
n
such that
n
= 1 on G
n
and
support
n
G. Put u
n
=
n
u and f
n
=
1
2
u
n
. Then
Z
n
(t, ) = u
n
(X(t, ))
t
_
0
f
n
(X(s, ))ds
is a martingale for each n. Put
n
= inf{t : X(t, ) G
n
}
Then Z
n
(
n
t, ) is also a martingale (See exercise below). But
Z
n
(
n
t) = Z(
n
t).
Therefore M
n
(t, ) = Z(
n
t, ) is a martingale. Observe that
n

n+1
and since G
n
G we have
n
. Therefore Z(
n
t) Z(t) (by
continuity); also |M
n
(t, )| ||u||
+ || f ||
t. Therefore Z( t) = Z(t, )
is a martingale.
Exercise. If M(t, ) is a (, F
t
, P)-martingale, show that for my stopping
time , M( t, ) is also a martingale relative to (F
t
).
[Hint: One has to show that if t
2
> t
1
,
_
A
M( t
2
, w)dP(w) =
_
A
M( t
1
, w)dP(w), A F
t
1
.
131
The right side =
_
A(>t
1
)
M(t
1
, w)dP(w) +
_
A(<t
1
)
M(, w)dP(w).
The left side 127
=
_
A(>t
1
)
M(t
2
, w)dP(w) +
_
A(<t
1
)
M(, w)dP(w).
Now use optional stopping theorem].
Lemma . Let G be a bounded region and be as above. Then E
x
() <
, x G, where E
x
= E
P
x
.
Proof. Without loss of generality we assume that G is a sphere of radius
R. The function u(x) =
R
2
|x|
2
d
0 and satises
1
2
u = 1 in G. By
the previous theorem
u(X( t, ) +
t
_
0
ds
is a martingale. Therefore
E
x
(u(X( t, ))) + E
x
( t) = u(X(0)) = u(x).
Therefore E
x
( t) u(x) (since u 0). By Fatous lemma, on
letting t , we obtain E
x
() u(x) < . Thus the mere existence of
a u satisfying
1
2
u = 1 helps in concluding that E
x
() < .
Theorem . Let u C
2
b
(G) and suppose that u satises
1
2
u = f in G, (*)
u = g on G.
Then u(x) = E
x
[g] E
x
[
_
0
f (X(s, ))ds] solves (*).
132 17. Solution of Poissons Equations
Remark. The rst part of the solutin u(x) is the solution of the homo-
geneous equation, and the second part accounts for the inhomogeneous
term.
Proof. Dene Z(t, ) = u(X( t))
t
_
0
f (X(s, ))ds. Then Z is a mar- 128
tingale. Also |Z| ||u||
+ || f ||
. Therefore, by the previous Lemma,

Z(t, ) is a uniformly integrable martingale. Therefore we can equate the
expectations at time t = 0 and at time t = to get
u(x) = E
x
(g) E
x
[
_
0
f (X(s, ))ds].
18. The Feynman-Kac

Formula
WE NOW CONSIDER the modied heat equation 129
(*)
u
t
+
1
2
u + V(x)u(t, x) = 0, 0 t T,
where u(T, x) = f (x). The Feynman-Kac formula says that the solution
for s T is given by
(**) u(s, x) = E
s,x
(e
_
T
s
V(X(s))dS
f (X(T))).
Observe that the solution at time s depends on the expectation with
respect to the process starting at time s.
Note . (**) is to be understood in the following sense. If (*) admits a
smooth solution then it must be given by (**). We shall not go into the
conditions under which the solution exists. Let
Z(t, ) = u(t, X(t, ))e
_
t
s
V(X(,)d)
, t s.
By Itos formula (see Example 2 of section 16), we get
Z(t, ) = Z(s, ) +
t
_
s
e
_
t
s
V(X(,)d)
u(, X()), dX(),
provided that u satises (*). Assume tentatively that V and u are
bounded and progressively measurable. Then Z(t, ) is a martingale.
Therefore
E
s,x
(Z(T, )) = E
z,x
(Z(s, )),
133
134 18. The Feynman-Kac Formula
or
E
s,x
(u(T, X(T)))e
_
T
s
V(X(,))d
= u(s, x).
This proves the result.
We shall now remove the condition
u
is bounded and prove the 130
uniqueness of the solution corresponding to (*) under the assumption
that V is bounded above and
|u(t, x)| e
A+|x|
, < 2, on [s, T).

In particular, the Feynman-Kac formula extends the uniqueness the-
orem for the heat equation to the class of unbounded functions satisfying
a growth condition of the form given above.
Let be a C
function such that = 1 on |X| R, and = 0 outside

|x| > R + 1. Put u
R
(t, x) = u(t, x),
Z
R
(t, x) = u
R
(t, x)e
_
t
s
V(X()d)
.
By what we have already proved, Z
R
(t, ) is a martingale. Let
R
() = inf{t : t s(t) S (0; R) = {|x| R}}.
Then Z
R
(t
R
, )) is also a martingale, i.e.
u
R
(t
R
, X(t
R
, ))e
_
t
R
s
V(X())d
is a martingale. Equating the expectations at time t = s and time t = T
and using the fact that
u
R
(t
R
, X(t
R
, )) = u(t
R
, X(t
R
, ))),
we conclude that
u(s, x) = E
s,x
[u(
R
T, X(
R
T, ))e
_

R
T
s
V(X(s))ds
]
= E
s,x
[X
(
R
T)
f (X(T))e
_
T
s
V(X(s))ds
]+
+ E
s,x
[X
(
R
T)
u(
R
, X
(
R
)
e
R
_
s
V(X(s))ds
]
135
131
Consider the second term on the right:
|E
s,x
[X
(
R
T)
u(
R
, X(
R
))e
_

R
s
V(X(s))ds
]|
A
e
R
P[
R
T]
(where A
is a constant given in terms of the constants A and T and the

bound of V)
= A
e
R
P[ sup
sT
|X()| R].
P[ sup
sT
|X()| R] is of the order of e
c(T)R
2
and since < 2, the
second term on the right side above tends to 0 as R . Hence, on
letting R +, we get, by the bounded convergence theorem,
u(s, x) = E
s,x
[ f (X(T))e
_
T
s
V(X(s))ds
]
Application. Let (t, ) be a one-dimensional Brownian motion. Recall
(Cf. Reection principle) that P{ sup
0st
|(s)| 1} is of the order of
4
2
t
8
. The Feynman-Kac formula will be used to explain the occurance
of the factor

2
8
in the exponent. First observe that

2
8
=
2
2
where is
the rst positive root of Cos = 0. Let
(w) = inf{t : |(t)| 1}.
Then
P{ sup
0st
|(s, )| 1} = P{ t}.
Let (x) = E
x
[e
], < 0. We claim that satises 132

1
2
+ = 0, |x| < 1,
= 1, |x| = 1.
(*)
136 18. The Feynman-Kac Formula
Assume to be suciently smooth. Using It os formula we get
d(e
t
((t)) = e
t
((t))d(t) + [((t)) +
1
2
((t))]e
t
dt.
Therefore
e
t
((t)) ((0)) =
t
_
0
e
s
((s))d(s)+
+
t
_
0
[((s)) +
1
2
((s))]e
s
ds,
i.e.
e
t
((t)) ((0))
t
_
0
[((s)) +
1
2
((s))]e
s
ds
is a martingale. By Doobs optional sampling theorem we can stop this
martingale at time , i.e.
e
(t)
((t )) ((0))
t
_
0
[((s)) +
1
2
((s))]e
s
ds
is also a martingale. But for s t ,
+
1
2
= 0.
Thus we conclude that ((t ))e
(t)
is a martingale. Since <
0 and ((t )) is bounded, this martingale is uniformly integrable.
Therefore equating the expectation at t = 0 and t = we get (since
(()) = 1)
(x) = E
x
[e
].
By uniqueness property this must be the solution. However (*) has 133
a solution given by
(x) =
Cos(
(2x))
Cos(
(2))
.
137
Therefore
(1) E
0
[e
] =
1
Cos(
(2))
( < 0),
If F(t) = P( t), then
_
0
e
t
dF(t) = E
0
(e
).
A theorem on Laplace transforms now tells us that (1) is valid till
we cross the rst singularity of
1
Cos(
(2))
. This occurs at =
2
8
. By
the monotone convergence theorem
E
0
[e
2
/8
] = +
Hence
_
0
e
t
dF(t) converges for <

2
8
and diverges for

2
8
.
Thus

2
8
is the supremum of for which
_
0
e
t
dF(t) converges, i.e. sup
[ : E
0
(e
)] exists, i.e. the decay rate is connected to the existence or

the non existence of the solution of the system (*). This is a general
feature and prevails even in higher dimensions.
19. An Application of the
Feynman-Kac Formula. The
Arc Sine Law.
LET (t, ) BE THE one-dimensional Brownian motion with (0, ) = 0. 134
Dene
t
(w) =
1
t
t
_
0
X
[0,)
((s, w))ds.
t
(w) is a random variable and denotes the fraction of the time that a
Brownian particle stays above the x-axis during the time interval [0, t].
We shall calculate
P[w :
t
(w) a] = F
t
(a)
Brownian Scaling. Let X
t
(s) =
1
t
(ts). Then X
t
is also a Brownian
motion with same distribution as that of (s). We can write
t
(w) =
1
_
0
X
[0,)
(X
t
(s, w))ds.
The
t
(w) = time spent above the x-axis by the Brownian motion
X
t
(s) in [0, 1]. Hence F
t
(a) is independent of t and is therefore denoted
139
140 19. An Application of the Feynman-Kac Formula....
by F(a). Suppose we succeed in solving for F(a); if, now,
t
(w) =
t
_
0
X
[0,)
((s))ds = t
t
,
then the amount of time (s, ) > 0 in [0, t] is t (amount of time X
t
(s) > 0
in [0, 1]). Hence we can solve for P[
t
(w) a] = G
t
(a). Clearly the
solution of G
t
is given by
G
t
(a) = F(a/t).
It is clear that 135
F(a) =
_
_
0 if a 0,
1 if a 1.
Hence it is enough to solve for F(a) in 0 a 1. Let
u
(t, x) = E
x
[e
(
_
t
0
X
[0,)
((s,w))ds)
]
Then
u
1
(t, 0) = E[e
(w))
] =
1
_
0
e
tx
dF(x).
Also note that u
(t, x) is bounded by 1, if 0. By the Feynman-

Kac formula (appropriately modied in case
1
2
is replaced by
1
2
)
u
1
(t, x) satises
u
t
=
1
2
2
u
x
2
u, x > 0,
=
1
2
2
u
x
2
, x < 0,
(*)
and u(0, x) = 1. Let
(x) =
_
0
u(t, x)e
t
dt, > 0, where u = u
1
,
141
=
_
0
u(t, x)(de
t
).
A simple integration by parts together with (*) gives the following
system of ordinary dierential equations for
1
2
+ ( + 1)
= , x > 0,
1
2
= , x < 0.
These have a solution
(x) =
+ 1
+ Ae
x
(2(+1))
+ Be
x
(2(+1))
, x > 0,
= 1 + Ce
x
(2)
+ De
x
(2)
, x < 0.
136
However u is bounded by 1 (see denition of u
(t, x)). Therefore
is also bounded by 1. This forces A = D = 0. We demand that
and
d
dx
should match at x = 0. (Some justication for this will be given
later on).
Lt
x0+
(x) = Lt
x0
(x)
gives
1 +
+ B = C + 1.
Similarly we get B
(2( +1)) = C
(2) by matching
d
dx
. Solv-
ing for B and C we get
B =
(1 + )(
( + 1))
,
C =
1
(1 + )(
( + 1))
.
Therefore
(0) =
+ 1
+ B =
( + 1)
,
i.e.
_
0
E[e
t
t
e
t
]dt =
( + 1)
.
Using Fubinis theorem this gives
E[

+
] =
(

+ 1
),
or
E[
1
1 + /a
] =
(
1
1 + (1/)
), i.e.
1
_
0
dF(x)
1 + x
=
1
(1 + )
.
137
This can be inverted to get
dF(x) =
2
dx
(x(1 x))
.
(Refer tables on transforms or check directly that
2
1
_
0
1
1 + x
dx
(x(1 x))
=
1
(1 + )
by expanding the left side in powers of (). Therefore
F(a) =
2
arcsin (
a), 0 a 1.
Hence G
t
(a) =
2
arcsin (
(
a
t
)), 0 a t, i.e.
P[
t
a] =
2
arcsin (
(
a
t
)), 0 a t.
This result goes by the name of arc sine law for obvious reasons.
We now give some justication regarding the matching conditions
used above.
143
The equation we solved was

1
2
+ V = f
where was bounded V 0. Suppose we formally use It os formula to
calculate
d
_
((t)e
_
t
0
(+V)((s)ds)
)
= e
_
t
0
(+V)((s,))ds
[f ((s, ))dt +
d
dX
((t))d(t)]
(see Example 2 of It os formula). Therefore
(Z(t, ) = ((t, ))e
_
t
0
(+V)((s,))ds
+
_
t
0
f (s, ) exp(
s
_
0
( + V)d)ds
is a martingale. Since , f are bounded and V 0, 138
|Z(t, )| ||||
+ || f ||
_
0
e
s
ds ||||
+ C|| f ||
.
Therefore Z(t, ) is uniformly integrable. Equating the expectations
at time 0 and gives
(*) (0) = E
0
_
0
[ f ((s, ))e
s
_
s
0
V(()d)
]ds.
This is exactly the form obtained by solving the dierential equa-
tions. In order to use It os formula one has to justify it. If we show
that It os formula is valid for functions having a discontinuity in the
second derivative, (*) will be a legitimate solution and in general there
is no reason why the second derivatives (or higher derivatives) should be
matched. This partially explains the need for matching and
d
dx
only.
Proposition . Let (t, ) denote a one-dimensional Brownian motion.
Suppose C
1
b
and satised

1
2
+ V = f ,
Then
((t))
t
_
0
f ((s))ds
is a martingale.
Proof. Let (
) C
2
b
such that

1
2
+ V
+ V
= f
and such
that (i)
converges to uniformly on compact sets, (ii)
converges to
uniformly on compact sets, (iii)
converges pointwise to
except
at 0. We may suppose that the convergence is bounded.
Claim.
t
_
0
f
((s))ds converges to
t
_
0
f ((s))ds a.e. As f
((s)) converges 139

to f ((s)) except when (s) = 0, it is enough to prove that
(*) P[w: Lebesgue measure (s : (s) = 0) > 0]= 0. Let X
{0}
denote
the indicator function of {0}. Then
E
t
_
0
X
{0}
((s))ds =
t
_
0
EX
{0}
((s))ds = 0.
Thus (*) holds and establishes the claim. Now
((t))
t
_
0
f
((s))ds
is a uniformly bounded martingale converging to
((t))
t
_
0
f ((s))ds.
145
Therefore
((t))
t
_
0
f ((s))ds
is a martingale.
20. Brownian Motion with
Drift
LET = C[0, ; R
d
], F = BOREL -FIELD of , {X(t, )} Brow- 140
nian motion, F
t
= [X(s, ) : 0 s t], P
x
probability measure
on corresponding to the Brownian motion starting at time 0 at x.
F = (
U
t0
F
t
). Let b : R
d
R
d
be any bounded measurable funci-
ton. Then the map (s, w)| b(w(s)) is progressively measurable and
Z(t, ) = exp[
t
_
0
b(X(s, )), dX(s, )
1
2
t
_
0
|b(X(s, ))|
2
ds]
t
, P
x
). Dene Q
t
x
on F
t
by
Q
t
x
(A) =
_
A
Z(t, )dP
x
,
i.e. Z(t, ) is the Radon-Nikodym derivative of Q
t
x
with respect to P
x
on
F
t
.
Proposition . (i) Q
t
x
is a probability measure.
(ii) {Q
t
x
: t 0} is a consistent family on
t0
F
t
, i.e. if A F
t
1
and
t
2
t
1
then Q
t
x
1(A) = Q
t
x
2(A).
Proof. Q
t
x
being an indenite integral, is a measure. Since Z(t, ) 0,
Q
t
x
is a positive measure. Q
t
x
() = E
x
(Z(t, )) = E
x
(Z(0, )) = 1. This
proves (i). (ii) follows from the fact that Z(t, ) is a martingale.
147
148 20. Brownian Motion with Drift
If A F
t
, we dene
Q
x
(A) = Q
t(A)
x
.
The above proposition shows that Q
x
is well dened and since (F
t
) 141
is an increasing family, Q
x
is nitely additive on
_
t0
F
t
.
Exercise. Show that Q
x
is countably additive on the algebra
_
t0
F
t
.
Then Q
x
extends as a measure to F = (
_
t0
F
t
). Thus we get a
family of measures {Q
x
: x R
d
} dened on (, F).
Proposition . If s < t then
Q
x
(X
t
A|F
s
) = Q
X(s)
(X(t s) A) a.e.
Denition. If a family of measures {Q
x
} satises the above property it
is called a homogeneous Markov family.
Proof. Let B F
s
. Therefore B X
1
t
(A) F
t
and by denition,
Q
x
((X(t) A) B)) =
_
BX
1
t
(A)
Z(t, )dP
x
E
P
x
(Z(t, )
(w)
B

A
(X(t, )))
= E
P
x
(E
P
x
(
Z(t, )
Z(s, )
Z(s, )
(w)
B

A
(X(t, ))|F
s
)
= E
P
x
([
B
Z(s, ))E
P
x
(
Z(t, )
Z(s, )
A
((t, ))]|F
s
)
(since B F
s
and Z(s, ) is F
s
-measurable)
= E
Q
x
[
B
E
P
x
(
Z(t, )
Z(s, )
A
((t, ))|F
s
)] . . . (1)
= E
Q
x
[
B
E
P
x
(exp[
t
_
s
b, dX
1
2
t
_
0
|b|
2
]
A
(X(t, ))|F
s
)]
149
= E
Q
x
[
B
E
P
X(s)
(exp[
ts
_
0
b, dX
1
2
ts
_
0
|b|
2
]X
A
(X(t s))]
(by Markov property of Brownian motion)
= E
Q
x
(
B
E
Q
ts
X(s)
(
A
((t s))) (since
dQ
ts
X(s)
dP
X(s)
= Z(t s, )
= E
Q
x
(X
B
E
Q
X(s)
A
(X(t s, ))
142
The result follows from denition.
Let b : [0, ] R
d
R
d
be a bounded measurable function, P
s,x
the probability measure corresponding to the Brownian motion starting
at time s at the point x. Dene, for t s,
Z
s,t
(w) = exp[
t
_
s
b(, X(, w)), dX(, w)
1
2
t
_
s
|b(, X(, w))|
2
d]
Exercise. (i) Z
s,t
is a martingale relative to (F
s
t
, P
s,x
).
(ii) Dene Q
t
s,x
by Q
t
s,x
(A) =
_
A
Z
s,t
dP
s,x
, A F
s
t
.
Show that Q
t
s,x
is a probability measure on F
s
t
.
(iii) Q
t
s,x
is a consistent family.
(iv) Q
s,x
dened on
U
ts
F
s
t
by Q
s,x
|F
s
t
= Q
t
s,x
is a nitely additive set
function which is countably additive.
(v) The family {Q
s,x
: 0 s < , x R
d
} is an inhomogeneous
Markov family, i.e.
Q
s,x
(X(t, ) A|F
s
) = Q
,X(,)
(X(t, ) A), s < < t, A F
s
t
.
[Hint: Repeat the arguments of the previous section with obvious 143
modications].
Proposition . Let be a stopping time, s. Then
Q
s,x
[X(t, ) A|F
s
] = Q
,X
()
(X(t, ) A) on (w) < t,
=
A
(X(t, )) on (w) t.
Proof. Let B F
s
and B { < t} so that B F

s
t
.
E
Q
s,x
(
B
A
(X
t
)) = E
P
s,x
(Z
s,t
A
(X
t
))
= E
P
s,x
[E
P
s,x
(Z
,t
Z
s,
A
(X
t
)|F
s
)]
(since Z satises the multiplicative property)
= E
P
s,x
(Z
s,
B
E
P
s,x
(Z
,t
A
(X
t
)|F
s
)]
(since Z
s,
is F
s
-measurable)
(*) = E
P
s,x
[Z
s,
X
B
E
P
,X
(Z
,t
A
(X
t
))]
(by strong Markov property).
Now
dQ
s,x
dP
s,x
F
s
t
= Z
s,t
.
so that the optional stopping theorem,
dQ
s,x
dP
s,x
F
s
= Z
s,
on { < t}, x.
Putting this in (*) we get 144
E
Q
s,x
[X
B
A
(X
t
)] = E
P
s,x
[Z
s,
B
E
Q
,X
(
A
X
t
)
].
Observe that
B
E
Q
,X
(
A
(X
t
))
is F
s
-measurable to conclude the rst part of the proof. For part (ii)
observe that
X
1
t
(A) { t} { k}
151
is in F
s
k
if k > s, so that
X
1
t
(A) { t} F
s
.
Therefore
E
Q
s,x
(X
t
A ( t))|F
s
) =
A
(X
t
)
{t}
,
or
E
Q
s,x
[(X
t
A)|F
s
] =
A
(X
t
) if t.
Proposition . Let b : [0, ) R

d
R
d
be a bounded measurable func-
tion, f : R
d
R any continuous bounded function. If
u
s
+
1
2
u + b(s, x), u = 0, 0 s t,
u(t, x) = f (x)
has a solution u, then
u(s, x) =
_
f (X
t
)dQ
s,x
.
Remark. b is called the drift. If b = 0 and s = 0 then we recover the 145
result obtained earlier. With the presence of the drift term, the result is
the same except that instead of P
s,x
one has to use Q
s,x
to evaluate the
expectation.
Proof. Let
Y(, ) =
_
s
b(, X(, )), dX(, )
1
2
_
s
|b(, X(, )|
2
d.
Step 1. (X(, )X(s, ), Y(, )) is a (d+1)-dimensional It o process with
parameters
(0, 0, . . . , 0,
1
2
|b(, X(, ))|
2
) and
d terms
a =
_
I
dd
b
d1
b
1d
b
_
Let = (
1
, . . . ,
d
). We have to show that
exp[Y +
_
s
|b(, X(, )|
2
d + , X(, ) X(s, )
1
2
_
s
, + 2, b +
2
|b(, )|
2
d]
is a martingale, i.e. that
exp[, X(, ) X(s, ) +
_
s
b, dX
1
2
_
s
| + b|
2
d].
is a martingale; in other words that
exp[
_
s
+ b, dX
1
2
_
s
| + b|
2
d]
is a martingale. But this is obvious because 146
Z(, ) =
_
s
+ b, dX
is a stochastic integral and hence an It o process with parameters (0, | +
b|
2
). (Refer to the section on vector-valued It o process).
Step 2. Put (, X(, ), Y(, )) = u(, X(, ))e
Y(,)
. By It o formula,
d = e
Y
u
t
dt + e
Y
u, dX + ue
Y
dY +
1
2
a
i j
z
i
z
j
,
where z = (x, y), or
d = e
Y
[
u
t
dt + u, dX + ub, dX
2
|b|
2
dt +
1
2
udt + b, udt+
153
+
1
2
u|b|
2
dt]
= e
Y
[u, dX + ub, dX].
Therefore is an It o process and hence a martingale. Therefore
E((t, )) = E((s, ))
u(s, x) = E
P
s,x
[( f (X(t))e
_
t
s
b,dX
1
2
t
_
0
|b|
2
d
]
= E
Q
s,x
[ f (X(t))],
which proves the theorem.
Alternate Proof. 147

Exercise . Let Y(, ) be progressively measurable for s. Then
Y(, ) is a martingale relative to (Q
s,x
, F
s
t
) if and only if Y()Z
s,
is
a martingale relative to (P
s,x
, F
s
t
).
Nowfor any function which is progressively measurable and boun-
ded,
exp[
t
_
s
, dX
1
2
t
_
s
||
2
d]
s
t
, P
s,x
). In particular let be replaced
by + b(, w()). After some rearrangement one nds that X
t
is an It o
process with parameters b, I relative to Q
s,x
. Therefore
u(t, X
t
)
t
_
s
_
u
+ b, u +
1
2
u
_
d
is a martingale relative to Q
s,x
. But
u
+ b, u +
1
2
u = 0.
Therefore
E
Q
s,x
(u(t, X(t)) = u(s, X).
We have dened Q
s,x
by using the notion of the Radon-Nikodym
derivative. We give one more relation between P and Q.
Theorem . Let T : C([s, ), R
d
) C([s, ), R
d
) be given by
TX = Y where Y(t) = X(t)
t
_
s
b(, X())d.
(b is as before). Then
Q
s,x
T
1
= P
s,x
.
Proof. Dene Y(t, w) = X(t, Tw) where X is a Brownian motion. We 148
prove that Y is a Brownian motion with respect to Q
s,x
. Clearly Y is
progressively measurable because T is (F
t
F
t
)-measurable for every
t, i.e. T
1
(F
t
) F
t
and X is progressively measurable. Clearly Y(t, w)
is continuous w. We have only to show that Y(t
2
) Y(t
1
) is Q
s,x
-
independent of F
s
t
1
and has distribution N(0; (t
2
t
1
)I) for each t
2
>
t
1
s. But we have checked that
exp[, X
t
x
1
2
||
2
(t s)
t
_
s
, bd]
is a martingale relative to Q
s,x
. Therefore
E
Q
s,x
(exp, Y
t
2
Y
t
1
|F
s
t
1
) = exp(
1
2
||
2
(t
2
t
1
)),
showing that Y
t
2
Y
t
1
is independent of F
s
t
1
and has normal distribution
N(0; (t
2
t
1
)I). Thus Y is a Brownian motion relative to Q
s,x
. Therefore
Q
s,x
T
1
= P
s,x
.
21. Integral Equations

Denition. A function b : R
d
R
d
is said to be locally Lipschitz if 149
given any x
0
R
d
there exists an open set U
0
containing x
0
such that
b|
U
0
is Lipschitz.
Exercise 1. b is locally Lipschitz i b|
K
is Lipschitz for every compact
set K i.e. i b|
K
is Lipschitz for every closed sphere K.
Exercise 2. Every locally Lipschitz function is continuous.
Theorem . Let b : R
d
R
d
be locally Lipschitz and X : [0, ) R
d
continuous. Then
(i) the equation
Y(t) = X(t) +
t
_
0
b(Y(s))ds ()
has a continuous solution near 0, i.e. there exists an > 0 and a
continuous function Y : [0, ] R
d
such that the above equation
is satised for all t in [0, ].
(ii) (Uniqueness) If Y
1
, Y
2
are continuous solutions of the above equa-
tion in [0, T], then
Y
1
= Y
2
on [0, T].
Proof. (ii) (Uniqueness) Let f (t) = |Y
1
(t) Y
2
(t)|. As Y
1
, Y
2
are con-
tinuous, there exists a k > 0 such that |Y
1
(t)|, |Y
2
(t)| k for all t in
155
156 21. Integral Equations
[0, T]. Choose C such that |b(x) b(y)| C|x y| for |x|, |y| k and
C sup
0tT
f (t). Then f (t) C and f (t) C
t
_
0
f (s)ds so that f (t)
(ct)
n
n!
for 150
all n = 1, 2, 3, . . .. Thus Y
1
(t) = Y
2
(t), proving uniqueness.
(i) (Existence) We can very well assume that X(0) = 0. Let a =
inf{t : |X(t)| 1},
M > sup{|b(x)| : |x| 2}, = inf{a,
1
M
},
C 0, a Lipschitz constant, so that |b(x) b(y)| C|x y| for all |x|,
|y| 2. Dene the iterations Y
0
, Y
1
, . . . by
Y
0
(t) = X(t), Y
n+1
(t) = X(t) +
t
_
0
b(Y
n
(s))ds
for all t 0. By induction, each Y
n
is continuous. By induction again,
|Y
n
(t) X(t)| Mt for all n, 0 t . Again, by induction |Y
n+1
(t)
Y
n
(t)|
M
C
(Ct)
n+1
(n+1)!
for 0 t . Again, by induciton |Y
n+1
(t) Y
n
(t)|
M
C
(Ct)
n+1
(n+1)!
for 0 t . Thus Y
n
(t) converges uniformly on [0, ] to a
continuous function Y(t) which is seen to satisfy the integral equation.
Remark. Let X : [, ) R
d
be continuous where > 0. Then a
similar proof guarantees that the equation (*) has a solution in [, ]
for some > 0.
Dene B(X) = sup{t : () has a solution in [0, t]}. The theorem
above implies that 0 < B(X) . B(X) is called the exploding time.
Remark . If b is, in addition, either bounded or globally Lipschitz,
B(X) = for every continuous X : [0, ) R
d
.
Example. Let b(y) = y
2
, X(t) = 0. The equation 151
Y(t) = x
0
+
t
_
0
b(y(s))ds
157
with x
0
> 0 has a solution
Y(t) =
1
1
x
0
t
, t <
1
x
0
;
the solution explodes at t = x
1
0
.
Proposition . If
B(w) < , then Lt
tB(w)
|y(t)| = +.
Proof. Suppose that lim
tB(w)
|y| = R < . Let (t
n
) be a sequence increas-
ing to B(w) such that |y(t
n
)| R + 1, n. Let
n
= inf{t t
n
: |y(t) y(
n
)| 1}.
Then
1 = |y(
n
) y(t
n
)|
w(
n
) w(t
n
)| + (
n
t
n
) sup |b()| . . . , (1)
S (y(t
n
), 1).
Since (t
n
) is bounded, we can choose a constant M such that
|w(t
n
) w(t)| <
1
2
if |t t
n
| M.
Then using (1),
n
t
n
inf{M, (2 sup |b()|)
1
where S (Y(t
n
); 1)
Therefore
n
t
n
inf(M, (2 sup |b()|)
1
, S (0; R + 2)) = (say) n.
152
158 21. Integral Equations
Chose n such that
n
> B(w) > t
n
. Then y is bounded in [t
n
, B(w)]
and hence it is bounded in [0, B(w)). From the equation
y(t) = X(t) +
t
_
0
b(y(s))ds
one then gets that Lt
tB(w)
y(t) exists. But this is clearly a contradiction
since in such a case the solution exists in [0, B(w) + ) for suitable ,
contradicting the denition of B(w). Thus
lim
tB(w)
|y(t)| = +
and hence
lim
tB(w)
|y(t)| = +.
Corollary . If b is locally Lipschitz and bounded, then B(X) = for all

X in C([0, ), R
d
).
Proof. Left as an exercise.
Proposition . Let b : R
d
R
d
be locally Lipschitz and bounded. Dene
T : C([0, ), R
d
) C([0, ), R
d
) by TX = Y where
Y(t) = X(t) +
t
_
0
b(Y(s))ds.
Then T is continuous.
Proof. Let X, X
: [0, ) R
d
be continuous, K > 0 be given.
Let Y
0
, Y
1
, . . . , Y
0
, Y
1
, . . . be the iterations for X, X
respectively. Then
|Y
n
(t) X(t)| K||b||
for 0 t K, n = 0, 1, 2, 3, . . ., so that we can

nd R such that |Y
n
(t)|, Y
n
(t)| R for 0 t k, n = 0, 1, 2, . . ., Let 153
C 1 be any Lipschitz constant for the function b on |x| R. Then
|Y
n
(t) Y
n
(t)| sup
0tK
|X(t) X
(t)| (1 + Ct +
(Ct)
2
2!
+
159
+ +
(Ct)
n
n!
) for 0 t K, n = 0, 1, 2, 3, . . . .
A b is bounded, Y
n
converges uniformly to Y on [0, K]. Letting
n , we get
sup
0tK
|(TX)t TX
)t| e
ck
sup
0tK
|X(t) X
(t)|, . . . (2)
where c depends on sup
0tK
|X(t)|, sup
0tK
|X
(t)|. The proof follows by (2).
22. Large Deviations

LET P
BE THE Brownian motion starting from zero scaled to Brown- 154

ian motion corresponding to the operator
2
. More precisely, let
P
(A) = P
_
A
_
where P is the Brownian motion starting at time 0 at the point 0.
Interpretation 1. Let {X
t
: t 0} be Brownian motion with X(0) = x.
Let Y(t) = X(t), t 0. Then P
is the measure induced by the process

Y(t). This amounts to stretching the time or scaling time.
Interpretation 2. Let Y(t, ) =
X(t, ). In this case also P
is the
measure induced by the process Y(t, ). This amounts to looking at the
process from a distance or scaling the length.
Exercise. Make the interpretations given above precise.
(Hint: Calculate (i) the probability that X(t) A, and (ii) the probability
that

X(t, ) A).
Problem. Let
I(w) =
1
2
1
_
0
| w(t)|
2
dt
if w(0) = 0, w absolutely continuous on [0, 1]. Put I(w) = otherwise.
161
162 22. Large Deviations
We would like to evaluate
_
e
F(w)
dP
(w)
for small values of . Here F(w) : C[0, 1] R is assumed to the a 155
bounded and continuous function.
Theorem . Let C be any closed set in C[0, 1] and let G be any open set
in C[0, 1]. Then
limsup
0
log P
(C)
_
wC
I(w),
liminf
0
log P
(G) inf
wG
I(w).
Here P
(G) = P
(
1
G) where : C[0, ) C[0, 1] is the canoni-
cal projection.
Signicance of the theorem . If
1.
dP
= e
I(w)
,
then
P
(A)
=
_
A
e
I(w)
dP
is asymptotically equivalent to
exp[
1
inf
wA
I(w)].
2. If A is any set such that
inf
wA
0
I(w) = inf
wA
I(w),
then by the theorem
Lt
0
log P
(A) = inf
wA
I(w).
163
Proof of the theorem. 156
Lemma 1. Let w
0
with I(w
0
) = < . If S = S (w
0
; ) is any
sphere of radius with centre at w
0
then lim
0
log P
(S ) I(w
0
).
Proof.
P(S ) =
_

(w)
S (w
0
,)
dP
=
_

(ww
0
)
S (0;)
dP
=
_

S (0;)
(w)dP
_
w
_
, where (w) = w w
0
,
=
_

S (0;)
((
w))dP(w)
=
_

S (0,)
(
w) exp
_
_
1
_
0
w
0
, dX
1
2
1
_
0
| w
0
|
2
d
_
_
dP(w)
=
_

S (0;)
(
w) exp
_
_
1
_
0
w
0
, dX I(w
0
)
_
_
dP(w)
=
_

S (0;)
(w) exp
_
1
_
0
w
0
, dX
1
I(w
0
)
_
_
dP
(w)
= e
I(w
0
)
(S (0; ))
1
P
(S (0; ))
_
S (0;)
exp
_
1
_
0
w
0
, dX
_
_
dP
e
I(w
0
)
(S (0; ))e
_
1
P
(S (0; ))
_
S (0;)
1
_
0
w
0
, dXdP
_
by Jensens inequality,
= e
I(w
0
)
(S (0, ))e
0
(use P
(w) = P
(w) if w S (0; ))
= e
I(w
0
)
(S (0, )).
157
Therefore
P
(S (w
0
; )) e
I(w
0
)
P(S (0;

))
or,
log P
(S (w
0
; )) I(w
0
) + log P(S (0;

));
let 0 to get the result. Note that the Lemma is trivially satised if
I(w
0
) = +.
Proof of Part 2 of the theorem.
Let G be open, w
0
G; then there exists > 0 with S (w
0
, ) G.
By Lemma 1
lim
0
log P
(G) lim
0
log P
(S (w
0
; )) I(w
0
).
Since w
0
is arbitrary, we get
lim log P
(G) inf{I(w
0
) : w
0
G}.
For part 1 we need some more preliminaries.
Lemma 2. Let (w
n
) C[0, 1] be such that w
n
w uniformly on [0, 1],
I(w
n
) < . Then I(w) < , i.e. I is lower semi-continuous.
Proof.
Step 1. w is absolutely continuous. Let {(x
i
, x
i
)}
n
i=1
be a collection of
mutually disjoint intervals in [0, 1]. Then
n
i=1
|w
m
(x
i
) w
m
(x
i
)|
n
i=1|
|x
i
x
i
|
1/2
[
x
i
_
x
i
|w
m
|
2
]
1/2
(by H olders inequality)
165
_
n
i=1
x
i
_
x
i
|w
m
|
2
_
_
1/2
_
_
n
i=1
|x
i
x
i
|
_
_
1/2
(again by H older)
(2)(
|x
i
x
i
|)
1/2
.
158
Letting m we get the result.
Step 2. Observe that w
m
(0) = 0S
0
w(0) = 0. Therefore
|
w
n
(x + h) w
n
(x)
h
|
2
= |
1
h
x+h
_
x
w
n
dt|
2
1
h
2
_
_
x+h
_
x
|w
n
|dt
_
_
2
1
h
x+h
_
x
|w
n
|
2
dt.
Hence
1h
_
0
|
w
n
(x + h) w
n
(x)
h
|
2
dx
1
h
1h
_
0
dx
h
_
0
|( w
n
(x + t))|
2
dt
1
h
h
_
0
dt
1h
_
0
| w
n
(x + t)|
2
dx
1
2
2
h
_
0
dt = 2
letting n , we get
1h
_
0
|
w(x + h) w(x)
h
|
2
dx 2.
Let h 0 to get I(w) , completing the proof.
Lemma 3. Let C be closed and put C
=
_
wC
S (w; ); then 159
lim
0
( inf
wC
I(w)) = inf
wC
I(w).
Proof. If
1
<
2
, then C
1
C
2
so that inf
wC
I(w) is decreasing. As
C
C for each ,
lim
0
( inf
wC
I(w)) inf
wC
I(w)
Let = lim
0
( inf
wC
I(w)). Then there exists w
such that I(w
)
, and therefore (I(w
)) is a bounded set bounded by (say).

Claim.
|w
(t
1
) w
(t
2
)| = |
t
2
_
t
1
w
dt|
|(|t
1
t
2
|)(
_
|w
|
2
)
1/2
(2|t
1
t
2
|).
The family (w
) is therefore equicontinuous which, in view of the

fact that w
(0) = 0, implies that it is uniformly bounded and the claim

follows from Ascolis theorem. Hence every subfamily of (w
) is equi-
continuous. By Ascolis theorem there exists a sequence
n
0 such
that w
n
w uniformly on [0, 1]. It is clear that w C. By lower
semicontinuity of I(w),
lim
0
inf
wC
I(w) inf
wC
completing the proof.
Proof of Part 1 of the theorem. Let X be continuous in [0, 1]. For each 160
n let X
n
be a piecewise linear version of X based on n equal intervals, i.e.
167
X
n
is a polygonal function joining the points (0, X(0)), (1/n, X(1/n)), . . . ,
(1, X(1)).
P
(||X
n
X|| ), (|| || = || ||
)
P
_
_
_
n
sup
ijn
sup
j1
n
t
j
n
|X
r
(t) X
r
_
j 1
n
_
|

2
d
_
_
,
where X = (X
1
, . . . , X
d
).
ndP
( sup
0t1/n
|X(t) X(0)|

2
d
(Markov property; here X is one-
dimensional).
ndP
_
sup
0t1/n
|X
t
|

2
d
_
_
(since X(0) = 0)
2nd P
( sup
0t1/n
X
t

d
)
= 2dn P( sup
0t1/n
X
t

d
)
= 4dn P(X(1/n)

2
d
) (by the reection principle)
= 4dn
n/2
d
1
2/n
e
ny
2
/2
dy
= 4d
n/2
d
1
2
e
x
2
/2
dx
Now, for every a > 0,
a
_
a
e
x
2
/2
dx
_
a
xe
x
2
/2
dx = e
a
2
/2
.
Thus 161
P
(||X
n
X|| )
4dn e
n
2
/(8d)
n/2
d
= C
1
(n)
e
n
2
/(8d)
,
where C
1
depends only on n. We have now
P
(X
n
C
) P
(I(X
n
)
) where
= inf{I(w)w C
}.
= P
_
_
1
2
n1
j=0
n|X
_
j + 1
n
_
X
_
j
n
_
|
2

_
= P
_
Y
2
1
+ Y
2
2
+ + Y
2
nd

2
_
,
where Y
1
=
n(X
1
(1/n) X
1
(0)) etc. are independent normal random
variables with mean 0 and variance 1. Therefore,
P(Y
2
1
+ + Y
2
nd

2
)
=
_
y
2
1
++y
2
nd
2
e
(y
2
1
++y
2
nd
)
1/2
dy
1
. . . dy
nd
.
= C(n)
(2
/)
e
r
2
/2
r
nd1
dr,
using polar coordinates, i.e.
P(Y
2
1
+ Y
2
2
+ + Y
2
nd

2
) = C
(n)
_
(
/)
e
s
s
nd
2
1
ds
(change the variable from r to s =
r
2
2
). An integration by parts gives
e
s
s
k
ds = e
(
k
+
k!
(k 1)!
k2
+ ).
Using this estimate (for n even) we get
P((Y
2
1
+ + Y
2
nd
)
2
) C
2
(n)e
/
(
)
nd
2
1
,
169
where C
2
depends only on n. Thus,
P
(C) P
(||X
n
X|| ) + P
(X
n
C
)
C
1
(n)
_
e
n
2
/(8d)
+ C
2
(n)e
/
_
_
nd
2
1
2 max[C
1
(n)
_
e
n
2
/(8d)
, C
2
(n)e
/
_
_
nd
2
1
log P
(C) log 2 + max[log(C

1
(n)
_
e
n
2
/(8d)
logC
2
(n)e
/
(
)
nd
2
1
]
162
Let 0 to get
lim log P
(C) max
_
n
2
8d
,
1
_
.
Fix and let n through even values to get
lim log P
(C)
.
Now let 0 and use the previous lemma to get
lim
0
log P
(C)
_
wC
I(w).
Proposition . Let be nite; then {w : I(w) } is compact in .
Proof. Let (w
n
) be any sequence, I(w
n
) . Then
|w
n
(t
1
) w
n
(t
2
)|
(|t
1
t
2
|)
and since w
n
(0) = 0, we conclude that {w
n
} is equicontinuous and uni-
formly bounded.
Assumptions. Let be any separable metric space, F = Borel -eld
on . For every > 0 let P
be a probability measure. Let I :

[0, ] be any function such that
(i) I is lower semi-continuous.
(ii) nite , {w : I(w) } is compact. 163
(iii) For every closed set C in ,
lim
0
sup log P
(C) inf
wC
I(w).
(iv) For every open set G in
lim
0
inf log P
(G) inf
wG
I(w).
Remark. Let = C[0, 1], P
the Brownian measure corresponding to

the scaling . If I(w) =
1
2
1
_
0
|w|
2
dt if w(0) = 0 and otherwise, then all
the above assumptions are satised.
Theorem . Let F : R be bounded and continuous. Under the
above assumptions the following results hold.
(i) For every closed set C in
lim
0
sup log
_
C
exp
F(w)
dP
sup
wC
(F(w) I(w)).
(ii) For every open set G in
lim
0
inf log
_
G
exp
F(w)
dP
sup
wG
(F(w) I(w)).
In particular, if G = = C, then
lim
0
log
_
exp
F(w)
dP
= sup
w
(F(w) I(w)).
171
Proof. Let G be open, w
0
G. Let 0 be given. Then there exists a
neighbourhood N of w
0
, F(w) F(w
0
) , w in N. Therefore
_
G
exp
F(w)
dP

_
N
exp
F(w)
dP
e
F(w
0
)
(N).
164
Therefore
log
_
G
exp
F(w)
dP
F(w
0
) + log P
(N).
Thus
limlog
_
G
exp
F(w)
dP
F(w
0
) + lim log P
(N).
F(w
0
) inf
wN
I(w) F(w
0
) I(w
0
) .
Since and w
0
are arbitrary (w
0
G) we get
lim log
_
G
exp
F(w)
dP
sup
wG
(F(w) I(w)).
This proves Part (ii) of the theorem.
Proof of Part (i).
Step 1. Let C be compact; L = sup
wG
(F(w) I(w)). If L = it follows
easily that
lim
0
sup log
_
C
e
F/
dP
.
(Use the fact that F is bounded). Thus without any loss, we may assume
L to be nite. Let w
0
C; then there exists a neighbourhood N of w
0
such that F(w) F(w
0
) + and by lower semi-continuity of I,
I(w) I(w
0
) , w N(w
0
).
By regularity, there exists an open set G
w
0
containing w
0
such that
G
w
0
G
w
0
N(w
0
). Therefore 165
_
G
w
0
exp
F(w)
dP
exp
_
F(w
0
) +
_
P
(G
w
0
).
Therefore
lim
0
sup log
_
G
w
0
exp
F(w)
dP
F(w
0
) + + lim
0
P
(G
w
0
)
F(w
0
) + inf
wG
w
0
I(w)
F(w
0
) + I(w
0
) +
L + 2.
Let K
= {w : I(w) }. By assumption, K
is compact. Therefore,
for each > 0, there exists an open set G
containing K
C such that
lim
0
sup log
_
G
e
F(w)
dP
L + 2.
Therefore
lim
0
sup log
_
G
C
e
F(w)
dP
L + 2,
_
G
c
C
e
F(w)
dP
e
M/
P(C G
c
).
Therefore
lim
0
sup log
_
G
c
C
e
F(w)
dP
M + lim
0
sup log P
(C
c
C)
M inf
wCG
c
I(w).
173
Now 166
G
c
K
c
C
c
.
Therefore
C G
c
C K
c
if w C G
c
, w K
. Therefore I(w) > . Thus

lim
0
sup log
_
G
c
C
e
F(w)/
dP
M L L + 2.
This proves that
lim
0
sup log
_
C
exp
F(w)
dP
L + 2.
Since C is compact there exists a nite number of points w
1
, . . . , w
n
in C such that
C
n
_
i=1
G
w
i
Therefore
lim log
_
C
exp
F(w)
dP
lim log
_
_
n
i=1
G
w
i
e
F(w)/
dP
lim( log n Max

1i
_
G
w
i
exp
F(w)
dP
)
L + 2.
Since is arbitrary.
lim log
_
C
exp
F(w)
dP
sup
wC
(F(w) I(w)).
The above proof shows that given a compact set C, and > 0 there 167
exists an open set G containing C such that
lim log
_
G
exp
F(w)
dP
L + 2.
Step 2. Let C be any arbitrary closed set in . Let
L = sup
wC
(F(w) I(w)).
Since F is bounded there exists an M such that |F(w)| M for all
w. Choose so large that M L. Since is arbitrary
lim
0
sup log
_
C
exp
F(w)
dP
sup
wC
(F(w) I(w))
We now prove the above theorem when P
is replaced by Q
x
. Let
P
x
be the Brownian motion starting at time t = 0 at the space point x
corresponding to the scaling . Precisely stated, if
: C([0, ); R
d
) C([0, ); R
d
)
is the map given by (
w)(t) = w(t), then P
x
=
def
P
x
. Note T
= T
and T
is given by
T
w = y where y(t) = w(t) +

t
_
0
b(y(s))ds.
Hence
P
x
T
1
= P
x
(T
1
,
)
1
= P
x
T
1
1
= P
x
T
1
1
;
either of these probability measures is denoted by Q
x
.
Theorem . Let b : R
d
R
d
be bounded measurable and locally Lips-
chitz. Dene
I(w) :
1
2
1
_
0
|X(t) b(X(t))|
2
dt
168
If w C([0, ); R
d
), w(0) = x and x absolutely continuous. Put
I(w) = otherwise. If C is closed in C[(0, 1]; R
d
), then
lim
0
log Q
x
(C) inf
wC
I(w).
175
If G is open in C([0, 1]; R
d
), then
lim
0
log Q
x
(G) inf
wG
I(w).
As usual Q
x
(C) = Q
1
(C) where
: C([0, ); R
d
) C([0, 1]; R
d
)
is the canonical projection.
Remark. If b = 0 we have the previous case.
Proof. Let T be the map x() y() where
y(t) = x(t) +
t
_
0
b(y(s))ds.
Then
Q
x
= P
x
(T
1
).
If C is closed
Q
x
(C) = P
x
(T
1
C).
The map T is continuous. Therefore T
1
(C) is closed. Thus
lim
0
sup log Q
x
(C) = lim
0
sup log P
x
(T
1
C)
inf
wT
1
(C)
1
2
1
_
0
|X|
2
dt (see Exercise 1 below)
= inf
wC
1
2
1
_
0
|T
1
w|
2
dt.
(*)
Now 169
y()
T
1
y(t)
t
_
0
b(y(s))ds.
Therefore
(T
1
y) = y b(y(s)).
Therefore
lim
0
sup log Q
x
(C) inf
wC
I(w).
The proof when G is one is similar.
Exercise 1. Replace P
by P
x
and I by I
x
where
I
x
(w) =
1
2
1
_
0
|w|
2
, w(0) = x, w absolutely continuous,
= otherwise.
Check that (*) holds, i.e.
lim
0
sup log P
x
(C) inf
wC
I
x
(w), if C is closed,
and
lim
0
inf log P
x
(G) inf
wG
I
x
(w).
Let G be a bounded open set in R
n
, with a smooth boundary = G.
Let b : R
d
R
d
be a smooth C
function such that

(i) b(x), n(x)0, x G where n(x) is the unit inward normal.
(ii) there exists a point x
0
G with b(x
0
) = 0 and |b(x)| > 0, x in
G {x
0
}.
(iii) for any x in G the solution 170
(t) = x +
t
_
0
b((s))ds,
of the vector eld starting from x converges to x
0
as t +.
Remark. (a) (iii) is usually interpreted by saying that x
0
is stable.
177
(b) By (i) and (ii) every solution of (iii) takes all its values in G and
ultimately stays close to x
0
.
Let > 0 be given; f : G R be any continuous bounded func-
tion. Consider the system
L
=
1
2
u
+ b(x) u
= 0 in G
u
= f on G.
We want to study lim
0
u
(x). Dene
I
T
0
(X(t)) =
1
2
T
_
0
|X(t) b(X(t))|
2
dt; X : [0, T] R
d
whenever X is absolutely continuous, = otherwise.
Remark. Any solution of (iii) is called an integral curve. For any curve
X on [0, T], I
T
0
gives a measure of the deviation of X from being an
integral curve. Let
V
T
(x, y) = inf{I
T
0
(X) : X(0) = x; X(T) = y}
and
V(x, y) = inf{V
T
(x, y) : T > 0}.
V has the following properties.
(i) V(x, y) V(x, z) + V(z, y) x, y, z. 171
(ii) Given any x, 0 and C > 0 such that for all y with |x y| .
V(x, y) C|x y|
Proof. Let X(t) =
t(y x)
|y x|
+ x.
Put
T = |y x|, X(0) = x, X(T) = y,
I
T
0
(X(t)) =
1
2
T
_
0
y x
T
b(X +
S
T
(y x))
2
ds.
Then
I
T
0

1
2
T
_
0
2
_
|y x|
2
T
2
+ ||b||
2
_
ds,
where
||b||
= sup
|x||yx|
b(),
or,
I
T
0
(1 + ||b||
2
)|y x|.
As a consequence of (ii) we conclude that
V(x, y)
_
_
1 + sup
|x|yx|
|b()|
2
_
_
|y x|,
i.e. V is locally Lipschitz.
The answer to the problem raised is given by the following.
Theorem .
lim
0
u
(x) = f (y
0
)
where y
0
is assumed to be such that y
0
G and
V(x
0
, y
0
) < V(x, y), y G, y y
0
.
172
We rst proceed to get an equivalent statement of the theorem. Let
P
x
be the Brownian measure corresponding to the starting point x, and
corresponding to the scaling . Then there exists a probability measure
Q
x
such that
dQ
x
dP
F
t
= Z(t)
179
where
Z(t, ) = exp
t
_
0
b
(X(s)), dX(s)
1
2
t
_
0
b
(X(s))ds;
b
is any bounded smooth function such that b
= b on G. Further we
have the integral representation
u
(x) =
_
G
f (X())dQ
x
where is the exit time of G, i.e.
(w) = inf{t : w(t) G}.
|u
(x) f (y
0
)| = |
_
G
( f (X()) f (y
0
))dQ
x
|
|
_
NG
( f (X()) f (Y
0
))dQ
x
|+
+|
_
N
c
G
( f (X()) f (Y
0
))dQ
x
|
(N is any neighbourhood of y
0
).
Q
x
(X() N G) sup
NG
| f () f (y
0
)|+
+2|| f ||
x
(X() N
c
G).
173
Since f is continuous, to prove the theorem it is sucient to prove
the
Theorem .
lim
0
Q
x
(X() N
c
G) = 0
for every neighbourhood N of y
0
.
Let N be any neighbourhood of y
0
. Let
V = V(x
0
, y
0
), V
= inf
yN
c
G
V(x, y).
By denition of y
0
and the fact that N
c
G is compact, we conclude
that V
> V. Choose = (N) > 0 such that V
= V + . For any > 0

let D = S (x
0
; ) = {y : |y x
0
| < }, D = {y : |y x
0
| = }.
Claim. We can choose a
2
such that
(i) V(x, y) V +
3
4
, x D
2
, y N
c
G.
(ii) V(x, y
0
) V +
4
, x D
2
.
Proof. (i) V(x
0
, y) V + , y N
c
G. Therefore
V + V(x
0
, y) V(x
0
, x) + V(x, y)
C|x x
0
| + V(x, y).
Choose C such that C|x x
0
|

4
. Thus
V +
3
4
V(x, y) if C|x x
0
|

4
, y N
c
G.
C depends only on x
0
. This proves (i). 174
(ii) |V(x
0
, y
0
) V(x, y
0
)| V(x
0
, x) C|x
0
x|

4
if x is close to x
0
.
Thus
V(x, y
0
) V(x
0
, y
0
) +
4
= V +
4
if x is close to x
0
. This can be achieved by choosing
2
very small.
Claim (iii). We can choose

1
<
2
such that for points x
1
, x
2
in D
1
there is a path X() joining x
1
, x
2
with X() D
2
D
1
, i.e. it never
penetrates D
1
; and
I(X)

8
.
181
Proof. Let C = sup{|b()|
2
: | x
0
|
2
}. Choose X() to be any path
on [0, T], taking values in D
2
with X(0) = x
1
; X(T) = x
2
and such that
|X| = 1 (i.e. the path has unit speed). Then
I
T
0
(X)
T
_
0
(|X|
2
+ C)dt CT +
T
_
0
|X|dt
= (c + 1)T = (C + 1)|x
2
x
1
|.
Choose
1
small such that (C + 1)|x
2
x
1
|
8
.
Let
1
= {w : w(t) G D
1
, t 0}, i.e.
1
consists of all
trajectories in G that avoid D
1
.
Claim (iv).
inf
T>0
inf
X
1
,X(0)D
2
I
T
0
(X()) V +
3
4
X(T) N
c
G
175
Proof. Follows from Claim (i) and (ii).
Claim (v).
inf
T>0
inf
X
1
,X(0)D
2
X(T)=y
0
I
T
0
(X()) V +
3
8
.
Proof. By (ii) V(x, y
0
) V +
4
x D
2
, i.e.
inf
T>0
inf
X(0)=x,X(T)=y
0
I
T
0
(X()) V +
4
.
Let > 0 be arbitrary. Choose T and X() such that I
T
0
(X) V+
4
+
with X(0) = x, X(T) = y
0
, X() G. If X
1
dene Y = X. If X
1
dene Y as follows:
Let t
1
be the rst time that X enters D
1
and t
2
the last time that it
gets out of D
1
. Then 0 < t
1
t
2
< T. Let X
be a path on [0, s] such

that (by Claim (iii)) I
s
0
(X
)

8
with X
(0) = X(t
1
) and X
(s) = X(t
2
).
Dene Y on [0, T (t
2
t
1
) + s] [T (t
2
t
1
) + s, ) by
Y(t) = X(t) on [0, t
1
] = X
(t t
1
) on [t
1
, t
1
+ s]
= X(t t
1
s + t
1
), on [t
1
+ s, T (t
2
t
1
) + s]
= X(t
2
), for t T (t
2
t
1
) + s.
Then
I
Tt
2
+t
1
+s
0
=
1
2
t
1
_
0
|X b(X(s))|
2
ds +
1
2
s
_
0
|X
b(X
(s))|
2
ds
+
1
2
T
_
t
2
|X(s) b(X(s))|
2
ds
V +
4
+ +
8
by choice of X and X
. As Y
1
, we have shown that 176
inf
T>0
inf
X
1
,X(0)D
1
X(T)=y
0
I
T
0
(X()) V +
3
8
+ .
Since is arbitrary we have proved (v).
Lemma . I
T
0
is lower semi-continuous for every nite T.
Proof. This is left as an exercise as it involves a repetiti on of an argu-
ment used earlier.
Lemma . Let X
n

1
. If T
n
then I
T
n
0
(X
n
) .
This result says that we cannot have a trajectory which starts outside
of a deleted ball for which I remains nite for arbitrary long lengths of
time.
183
Proof. Assume the contrary. Then there exists a constant M such that
I
T
n
0
(X
0
) M, n. Let T < , so that M
T
= sup
n
I
T
0
(X
n
) < .
Dene X
T
n
= X
n
|
[0,T]
.
Claim. {X
T
n
}
n=1
is an equicontinuous family. 177
Proof.
|X
T
n
(x
2
) X
T
n
(x
1
)|
2
= |
x
2
_
x
1
X
T
n
(t)dt|
2
|x
2
x
1
|
2
x
2
_
x
1
|X
T
n
|
2
dt
2|x
2
x
1
|
2
x
2
_
|X
T
n
b(X
T
n
)|
2
ds +
T
_
0
b(X
T
n
)
2
ds]
2|x
2
x
1
|
2
[2M
T
+ T||b||
2
].
Thus, {X
T
n
}
n
is an equicontinuous family. Since G is bounded, {X
T
n
}
n
is uniformly bounded. By Arzela-Ascoli theorem and a diagonal pro-
cedure there exists a subsequence X
n
k
and a continuous function to X
uniformly on compact subsets of [0, ). As X
n
k
() GD
1
, X GD
1
.
Let m n. I
T
n
0
(X
m
) M. X
n
X uniformly on [0, T
n
]. By lower semi-
continuity I
T
0
(X) M. Since this is true for every T we get on letting T
tend to , that
1
2
_
0
|X b(X(s))|
2
ds M.
Thus we can nd a sequence a
1
< b
1
< a
2
< b
2
< . . . such that
I
b
n
a
n
(X()) =
1
2
b
n
_
a
n
|X(t) b(X(t))|
2
dt
converges to zero with b
n
a
n
. Let Y
n
(t) = X(t + a
n
). Then
I
b
n
a
n
0
(Y
n
) 0 with b
n
a
n
+, Y
n

1
.
Just as X was constructed from X
n
, we can construct Y from Y
n
such 178
that Y
n
Y uniformly on compact subsets of [0, ).
I
b
n
a
n
0
(Y) inf
mn
I
b
n
a
n
0
(Y
m
) = 0
(by lower semi-conrmity of I
T
0
). Thus I
b
n
a
n
0
(Y) = 0, n, showing that
_
0
Y(t) b(Y(t))|
2
dt = 0
Thus Y satises Y() = x +
t
_
0
b(Y(s))ds with Y(t) G D
1
, t.
Case (i). Y(t
0
) G for some t
0
. Let Z(t) = Y(t+t
0
) so that Z is an integral
curve starting at a point of G and remaining away from D
1
contradicting
the stability condition.
Case (ii). Y(t
0
) G for any t
0
, i.e. Y(t) G for all t. Since Y(t) =
b(Y(t))Y(t), n(Y(t)) is strictly positive. But Y(t) G and hence
Y(t), n(Y(t)) = 0 which leads to a contradiction. Thus our assump-
tion is incorrect and hence the lemma follows.
Lemma . Let x D
2
and dene
E = {X(t) exits from G before hitting D
1
and it exits from N}
F = {X(t) exists from G before hitting D
1
and it exits from N
c
}
Then
Q
x
(F)
Q
x
(E)
exp
_
3
8
+ 0
_
1
__
0 uniformly in x(x D
2
).
Signicance. Q
x
(E) and Q
x
(F) are both small because diusion is small 179
and the drift is large. The lemma says that Q
x
(E) is relatively much
larger than Q
x
(F).
185
Proof. Q
x
(E) Q
x
{X(t) exists from G before hitting D
1
, and exists in
N before time T}, = Q
x
(B) exp[
1
inf I
T
0
(X())] where the inmum
is taken over the interior of B,
exp
_
_
V +
3
8
_
+ 0
_
1
__
.
Similarly,
Q
x
(F) exp
_
_
V +
3
4
_
+ 0
_
1
__
.
Therefore
Q
x
(F)
Q
x
(E)
exp
_
3
8
+ 0
_
1
__
0 as 0.
We now proceed to prove the main theorem. Let
0
= 0,
1
= rst time D
1
is hit,
2
= next time D
2
is hit,
3
= next time D
1
is hit,
. . . . . . . . . . . . . . . . . .
and so on. Observe that the particle can get out of G only between the
time intervals
2n
and
2n+1
. Let E
n
= {between
2n
and
2n+1
the path
exits from G for the rst time and that it exits in N}, F
n
= {between
2n
180
and
2n+1
the path exits from G for the rst time and that it exists in N
c
}.
Q
x
(X() N) + Q
x
(X() N
c
) = 1.
Also
Q
x
(X() N
c
) =
n=1
Q
x
(F
n
),
Q
x
(X() N) =
n=1
Q
x
(E
n
),
n=1
Q
x
(F
n
) =
n=1
E
Q
x
(Q
x
(F
n
|F
2n
))
n=1
E
Q
x
[
(>
2n
)
sup
xD
2
Q
x
(F)] (by the Strong Markov property)
0()
n=1
E
Q
x
[
(>
2n
)
inf
xD
2
Q
x
(E)] (as
Q
x
(F)
Q
x
(E)
0)
0()
n=1
Q
x
(E
n
) = 0()Q
x
(X() N
c
).
Therefore
Q
x
(() N) 1, Q
x
(X() N
c
) 0.
Exercise. Suppose b(x) = u(x) for some u C

1
(G G, R). Assume
that u(x
0
) = 0 and u(x) < 0 for x x
0
. Show that
V(x
0
, y) = 2u(y).
[Hint: For any trajectory X with X(0) = x
0
,
X(T) = y, I
T
0
(X) =
1
2
T
_
0
|X + u(X)|
2
dt 2
T
_
0
u(X) X(t)dt 2u(y)
so that V(x
0
, y) 2u(y). For the other inequality, let u be a solution of 181
X(t) +u(X(t) = 0 on [0, ) with X(0) = y. Show that because
duX(s)
ds
0
for X(s) 0 and x
0
is the only zero of u, limit
t
X(t) = x
0
. Now conclude
that V(x
0
, y) u(y)].
23. Stochastic Integral for a
Wider Class of Functions
WE SHALL NOW dene the stochastic integral for a wider class of 182
functions.
Let : [0, ) R
d
be any progressively measurable function
such that for every t
t
_
0
|(s, w)|
2
ds < , a.e.
Dene, for every nite L 0,
L
(s, w) =
_
_
(s, w), if
s
_
0
|(t, w)|
2
dt < L < ,
0, if
s
_
0
|(t, w)|
2
dt L.
We can write
L
(s, w) = (s, w)
[0,L)
((s, w)) where
(s, w) =
s
_
0
|(t, w)|
2
dt
is progressively measurable. Hence
L
(s, w) is progressively measur-
187
188 23. Stochastic Integral for a Wider Class of Functions
able. It is clear that
T
_
0
|
L
(s, w)|
2
ds L, a.e. T. Therefore
E(
T
_
0
|
L
(s, w)|
2
ds) L.
Thus the stochastic integral
L
(t, w) =
t
_
0
L
(s, w), dX(s, w) is well de-
ned.
The proofs of the next three lemmas follow closely the treatment of
Stochastic integration given earlier.
Lemma 1. Let be a bounded, progressively measurable, continuous
function. Let be any nite stopping time. If (s, w) = 0, (s, w) such 183
that 0 s (w) then
t
_
0
(s, w), dX(s, w) = 0 for 0 t (w).
Proof. Dene
n
(s, w) = (
[ns]
n
, w).
n
is progressively measurable and
by denition of the stochastic integral of
n
,
t
_
0
n
(s, w), dX(s, w) = 0, t, 0 t (w).
Now
E
_
_
t
_
0
|
n
(s, w) (s, w)|
2
ds
_
_
= E
_
_
t
_
0
|
_
[ns]
n
w
_
(s, w)|
2
ds
_
_
0 as n
and
t
_
0
n
(s, w), dX(s, w)
t
_
0
(s, w), dX(s, w)
189
in probability. Therefore
t
_
0
, dX = 0 if 0 t (w).
Lemma 2. If is progressively measurable and bounded the assertion

of Lemma 1 still holds.
Proof. Let
n
(t, w) =
1
n
t
_
(t1/n)V
0
(s, w)ds.
Then
E
_
_
T
_
0
|
n
(t, w) (t, w)|
2
dt
_
_
0 (Lebesgues theorem).
n
is continuous and boundd,
n
(t, w) = 0 for 0 t (w). By lemma 1 184
t
_
0
n
(s, w), dX(s, w) = 0
if 0 t (w). This proves the result.
Lemma 3. Let be progressively measurable such that, for all t,
E
_
_
t
_
0
|(s, w)|
2
ds
_
_
< .
If (s, w) = 0 for 0 s (w), then
t
_
0
(s, w), dX(s, w) = 0 for 0 t (w).
Proof. Dene
n
=
_
_
, if || n,
0, if || > n.
Then
c
_
0
n
, dX = 0, if 0 t (w), (Lemma 2) and
E(
t
_
0
|
n
|
2
ds) 0. The result follows.
Lemma 4. Let be progressively measurable such that t,
t
_
0
|(s, w)|
2
ds < a.e.
Then Lt
L
L
(t, w) exists a.e.
Proof. Dene 185
L
(w) = inf
_
_
s :
s
_
0
|(, w)|
2
d L
_
_
;
clearly
L
is a stopping time. If L
1
L
2
,
L
1
(w)
L
2
(w) and by
assumptions of the lemma
L
as L . If
L
1
L
2
,
L
1
(s, w) =
L
2
(s, w) for 0 s
L
1
(w).
Therefore by Lemma 3,
L
2
(t, w) =
L
1
(t, w)
if 0 t
L
1
(w). Therefore as soon as L is large enough such that
t
L
(w),
L
(t, w) remains constant (as a function of L). Therefore
Lt
L
L
(t, w) exists a.e.
191
Denition. The stochastic integral of is dened by
t
_
0
(s, w), dX(s, w) = Lt
L
L
(t, w).
Exercise. Check that the denition of the stochastic integral given above
coincides with the previous denition in case
E
_
_
t
_
0
|(s, w)|
2
ds
_
_
< , t.
Lemma . Let be a progressively measurable function, such that
t
_
0
|(s, w)|
2
ds < , t.
If (t, w) denotes the stochastic integral of , then 186
P
_
sup
0tT
|(t, )|
_
P
_
_
T
_
0
||
2
ds L
_
_
+
L
2
2
.
Proof. Let
L
= inf{t :
t
_
0
||
2
ds L}. If T <
L
(w), then
L
(s, w) =
(s, w). Also
L
(t, w) = (t, w) for 0 t T.
Claim.
_
w : sup
0tT
|(t, w)|
_
_
w : sup
0tT
|
L
(t, w)|
_
{w :
L
(w) T}
For, if w is not contained in the right side, then
sup
0tT
|
L
(t, w)|
2
< and |
L
(w)| > T.
If
L
> T,
L
(t, w) = (t, w) t T. Therefore
sup
0tT
|
L
(t, w)| = sup
0tT
|(t, w)|
Therefore w left side. Since
{w :
L
(w) > T} =
_
_
w :
T
_
0
||
2
ds L
_
_
we get
P
_
sup
0tT
|(t, )|
_
P
_
_
T
_
0
||
2
ds L
_
_
+ P
_
sup
0tT
|
L
(t, )|
_
P
_
_
T
_
0
||
2
ds L
_
_
+
L
2
2
,
by Kolmogorovs inequality. This proves the result. 187
Corollary . Let
n
and be progressively measurable functions such
that
(a)
t
_
0
|
n
(s, w)|
2
ds < ,
t
_
0
|(s, w)|
2
ds < , t;
(b) Lt
n
t
_
0
|
n
(s, w) (s, w)|
2
ds = 0 in probability.
If
n
(t, w) and (t, w) denote, respectively the stochastic integrals of
n
and , then sup
0tT
|
n
(t, w) (t, w)| converges to zero in probability.
193
Proof. Let
n,L
(w) = inf{t :
t
_
0
|
n
|
2
ds L}; replacing by
n
and
by
n
in the previous lemma, we get
P
_
sup
0tT
|
n
(t, ) (t, )|
_
L
2
2
+ P
_
_
T
_
0
|
n
(s, ) (s, )|
2
ds L
_
_
.
Letting n , we get
lim
n
P
_
sup
0tT
|
n
(t, ) (t, )|
_
L
2
2
.
As L is arbitrary we get the desired result.
Proposition . Let be progressively measurable such that
t
_
0
|(s, w)|
2
ds < , t and w.
Then 188
(*) Z(t, ) = exp
_
_
t
_
0
(s, ), dX(s, )
1
2
t
_
0
|(s, )|
2
ds
_
_
is a super martingale satisfying
(a) E(Z(t, )) 1;
(b) Lt
t0
E(Z(t, )) = 1.
Proof. Let (
n
) be a sequence of bounded progressively measurable
functions such that
t
_
0
|
n
|
2
ds 0 t, w.
(For example we may take
n
= if || < n, = 0 otherwise). Then
() is a martingale when is replaced by
n
. This martingale satises
E(Z
n
(t, )) = 1, and Z
n
(t, ) Z(t, ) pointwise (a) now follows from
Fatous lemma:
lim
t0
E(Z(t, )) E
_
lim
t0
Z(t, )
_
= E(1) = 1.
Therefore Lt
t0
E(Z(t, )) = 1. This proves (b).
24. Explosions
Exercise. Let R > 0 be given. Put b
R
= b
R
where
R
= 1 on |x| R, 189
R
= 0 if |x| R + 1;
R
is C
. Show that b
R
= b on |x| R, b
R
is
bounded on R
d
and b
R
is globally Lipschitz.
Let
T
= {w : B(w) > T}. Let S
T
=
T
C[0, T] be the map
S
T
w = y() where y(t) = w(t) +
t
_
0
b(y(s))ds on [0, T]. Unless otherwise
specie b : R
d
R
d
is assumed to be locally Lipschitz. Dene the
measure Q
T
x
on (, T) by
Q
T
x
(A) = P
x
{w : S
T
w A, B(w) > T},
where P
x
is the probability measure corresponding to Brownian motion.
Theorem .
Q
T
x
(A) =
_
A
Z(T, )dP
x
, A F
T
,
where
Z(T, ) = exp
_
_
T
_
0
b, dX
1
2
T
_
0
|b(X(s, ))|
2
ds
_
_
.
Remark. If b is bounded or if b satises a global Lipschitz condition
then B(w) = , so that
T
= and Q
T
x
are probability measures.
Proof. Let 0 R < . For any w in , let y be given by
y(t) = w(t) +
t
_
0
b(y())d.
195
196 24. Explosions
Dene
R
(w) = inf{t : |y(t)| R and let b
R
be as in the Exercise.
Then the equation
y
R
(t) = w(t) +
t
_
0
b
R
(y
R
())d
has a global solution. Denote by S
R
: the map w y
R
. If Q
R,x
190
is the measure induced by S
R
, then
dQ
R,x
dP
x
F
t
= Z
r
(t) = exp
_
_
t
_
0
b
R
, dX
1
2
t
_
0
|b
R
|
2
ds
_
_
.
Let
R
(w) = inf{t : |w(t)| > R}.
R
is a stopping time satisfying
R
S
R
=
R
. By the optional stopping theorem.
(1)
dQ
R,x
dP
x
R
T
= Z
R
(
R
T) = Z(
R
T).
Claim. Q
R,x
((
R
> T) A) = Q
T
x
((
R
> T) A), A in F
T
.
Proof.
Right side = P
x
{w : B(w) > T, S
T
(w) A (
R
> T)}
= P
x
{w : B(w) > T, y A, sup
0tT
|y(t)| < R}
= P
x
{w : y is dened at least upto time T,
y A, sup
0tT
|y(t)| > R}
= P
x
{w : y
R
A, sup
0tT
|y
R
(t)| < R}
= P
x
{w : S
R
(w) A,
R
S
R
(w) > T}
= Q
R,x
{(
R
> T) A}
(by denition). As is an increasing union of {
R
> T} for R increasing,
Q
T
x
(A) = lt
R+
Q
T
x
((
R
> T) A), A in F
T
,
197
= lt
R
Q
R,x
((
R
> T) A) (by claim)
= lt
R
_
(
R
T)A
exp
_
R
T
_
0
b, dX
1
2
R
T
_
0
|b|
2
ds
_
_
dP
x
(by (1))
=
_
A
exp
_
_
T
_
0
b, dX
1
2
T
_
0
|b|
2
ds
_
_
dP
x
=
_
A
Z(T)dP
x
.
191
Theorem . Suppose b : R
d
R
d
is locally Lipschitz; let L =
2
+ b..
(i) If there exists a smooth function u : R
d
(0, ) such that u(x)
as |x| and Lu cu for some c > 0 then P
x
{w : B(w) <
} = 0, i.e. for almost all w there is no explosion.
(ii) If there exists a smooth bounded function u : R
d
(0, ) such
that Lu cu for some c > 0, then P
x
{w : B(w) < } > 0, i.e.
there is explosion.
Corollary . Suppose, in particular, b satises |b(x), x| A + B|x|
2
for
some constants A and B; then P
x
(w : B(w) < ) = 0.
Proof. Take u(x) = 1 + |x|
2
and use part (1) of the theorem.
Proof of theorem. Let b
R
be as in the Exercise and let L
R
=
2
+ b
R
;
then L
R
u(x) cu(x) if |x| R.
Claim. u(X(t))e
ct
is a supermartingale upto time
R
relative to Q
R
x
,
d
_
_
u(X(t))e
ct
exp
_
_
t
_
0
b
R
, dX
1
2
t
_
0
|b
R
|
2
ds
_
_
_
_
198 24. Explosions
e
ct
t
_
0
b
R
,dX
1
2
t
_
0
|b
R
|
2
ds
{cudt + u, dX + u(x)[b
R
, dX
|b
R
|
2
2
dt] + b
R
udt +
1
2
|b
R
|
2
udt}
= exp(ct +
t
_
0
b
R
, dX
1
2
t
_
0
|b
R
|
2
ds)
[L
R
c)u + u, dX + ub
R
, dX].
192
Therefore
u(X(t))e
ct
E
t
_
e
0
b
R
,dX
1
2
t
_
0
|b
R
|
2
ds
t
_
0
exp
_
_
cs +
s
_
0
b
R
, dX
s
_
0
|b
R
|
2
ds
_
_
(L
R
c)u(X(s))ds
is a Brownian stochastic integral. Therefore
u(X(
R
t)) exp
_
_
c(
R
t) +
R
t
_
0
b
R
, dX
1
2
R
t
_
0
|b
R
|
2
ds
_
R
t
_
0
exp
_
_
cs +
s
_
0
b
R
, dX
1
2
s
_
0
|b
R
|
2
ds
_
_
(L
R
c)u(X(s))ds
is a martingale relative to P
x
, F
R
t
. But b
R
(x) = b(x) if |x| R. There-
fore
u(X(
R
t)) exp
_
_
c(
R
t) +
R
t
_
0
b, dX
1
2
R
t
_
0
|b|
2
ds
_
R
t
_
0
exp
_
_
cs +
s
_
b, dX
s
_
0
|b|
2
ds
_
_
(L
R
c)u(X(s))ds
199
is a martingale relative to F
R
t
. But (L
R
c)u 0 in [0,
R
].
Therefore 193
u(X(
R
t)) exp(c(
R
t) +
R
t
_
0
b, dX
1
2
R
t
_
0
|b|
2
ds)
is a supermartingale relative to F
R
t
, P
x
. Therefore u(X(
R
t)e
c(
R
t)
is a supermartingale relative to Q
R
x
(optional sampling theorem). There-
fore
E
Q
R
x
(u(X(t
R
t))e
c(
R
t)
u(x);
letting t , we get, using Fatous lemma,
E
Q
R
x
(u(X(
R
)e
c
R
) u(x).
Therefore
E
Q
R
x
(e
c
R
)
u(x)
inf
|y|=R
|u(y)|
.
Thus
E
P
x
(e
c
R
)
u(x)
inf
|y|=R
|u(y)|
(by change of variable). Let R to get Lt
R
_
e
c
R
dP
x
= 0, i.e.
P
x
{w : B(w) < } = 0.
Sketch of proof for Part (ii).
By using the same technique as for Part (i), show that u(X(t))e
ct
is
a submartingale upto time
R
relative to Q
R
x
, so that
E
P
x
(e
c
R
)
u(x)
sup
|y|=R
|u(y)|
u(x)
||u||
> 0;
let R to get the result. 194
Exercise. Show that if L =
1
2
2
x
+ x
3

x
, there is explosion. (Hint: take
u = e
tan
1
(x
2
)
and show that Lu u).
25. Construction of a
Diusion Process
Problem. Given a : [0, ) R
d
S
+
d
, bounded measurable and b : 195
[0, ) R
d
R
d
bounded measurable, to nd (, F
t
, P, X) where is
a space, {F
t
}
t0
an increasing family of -algebras on , P a probability
measure on the smallest -algebra containing all the F
t
s. and X :
[0, t) R
d
, a progressively measurable function such that X(t, w)
I[b(t, X
t
), a(t, X
t
)].
Let = C[0, ); R
n
), (t, ) = n-dimensional Brownian motion,
F
t
= {(s) : 0 s t}, P the Brownian measure on and a and b as
given in the problem. We shall show that thee problem has a solution,
under some special conditions on a and b.
Theorem . Assume that there exists : [0, ) R
d
M
dn
(M
dn
=
set of all d n matrices over the reals) such that
= a. Further let
i, j
|
i j
(t, x)| C,
j
|b
j
(t, x)| C,
i, j
|
i j
(t, x
1
)
i j
(t, x
2
)| A|x
1
x
2
|,
j
|b
j
(t, x
1
) b
j
(t, x
2
)| A|x
1
x
2
|.
201
202 25. Construction of a Diusion Process
Then the equation
(1) (t, ) = x +
t
_
0
(s, (s, )), d(s, ) +
t
_
0
b(s, (s, ))ds
has a solution. The solution (t, w) : [0, ) R
d
can be taken 196
to be such that (t, ) is progressively measurable and such that (t, ) is
continuous for a, a.e. If , are progressively measurable, continuous
(for a.a.e) solutions of equation (l), then = n for a.a.w.
Proof. The proof proceeds in several steps.
Lemma 1. Let be any space with (F
t
)
t0
an increasing family of -
algebras. If 0 T then there exists a -algebra A
0
A =
B[0, T) F
T
such that a function f : [0, T] R is progressively
measurable if and only if f is measurable with respect to A
0
.
Proof. Let A
0
= {A A :
A
is progressively measurable}. Clearly
[0, T] A
0
, and if A A
0
, A
c
A
c
. Thus A
0
is an algebra. As
increasing limits (decreasing limits) of progressively measurable func-
tions are progressively measurable, A
0
is a monotone class and hence a
-algebra.
Let f : [0, T] R be progressively measurable; in fact, f
+
=
f + 1 = f |
2
, f
=
f | f |
2
. Let g = f
+
. Then
g
n
=
n2
n
i=1
i 1
2
n

g
1
[
i1
2
n
,
i
2
n
)
+ n
g
1
[n,)
is progressively measurable. Hence nVg
n
is progressively measurable,
i.e. n
g
1
[n,)
is progressively measurable. Similarly
g
1
[
i1
2
n
,
i
2
n
)
is pro- 197
gressively measurable, etc. Therefore, by denition, g
n
is measurable
with respect to A
0
. As g = f
+
is the pointwise limit of g
n
, f
+
is mea-
surable with respect to A
0
. Similarly f
is A
0
-measurable. Thus f is
A
0
-measurable.
203
Let f : [0, T] R be measurable with respect to A
0
. Again, if
g = f
+
g
n
=
n2
n
i=1
i 1
2
n

g
1
[
i1
2
n
,
i
2
n
)
+ n
g
1
[n,)
is A
0
-measurable. Since g
1
[n, ), . . . g
1
[
i 1
2
n
,
i
2
n
) A
0
. So g
n
is progressively measurable. Therefore g is progressively measurable.
Hence f is progressively measurable. This completes the proof of the
Lemma.
To solve (1) we use the standard iteration technique.
Step 1. Let
0
(t, w) = x,
n
(t, w) = x +
t
_
0
(s,
n1
(s, w)), d(s, w) +
t
_
0
b(s,
n1
(s, w))ds.
By induction, it follows that
n
(t, w) is progressively measurable.
Step 2. Let
n
(t) = E(|
n+1
(t)
n
(t)|
2
). If 0 t T,
n
(t) C
t
_
0
n1
(s)ds and
0
(t) C
t, where C
is a constant depending only on T.

Proof.
0
(t) = E(|(t) x|
2
)
= E
_
_
|
t
_
0
(s, x), d(s, x) +
t
_
0
b(s, x)ds|
2
_
_
2E
_
_
|
t
_
0
(s, x), d(s, x)|
2
_
_
+
+ 2E
_
_
|
t
_
0
b(s, x)ds|
2
_
_
(use the fact that |x + y|
2
2(|x|
2
+ |y|
2
) x, y R
d
)
= 2E
_
_
t
_
0
Tr
ds
_
_
= 2E
_
_
|
t
_
0
b(s, x)ds|
2
_
_
.
or 198
0
(t) 2E
_
_
t
_
0
tr
ds
_
_
+ 2E
_
_
t
t
_
0
|b(s, x)|
2
ds
_
_
(Cauchy-Schwarz inequality)
2nd C
2
(1 + T)t.
n
(t) = E(|
n+1
(t)
n
(t)|
2
)
= E
_
_
|
t
_
0
(s,
n
(s, w)) (s,
n1
(s, w))d+
+
t
_
0
b(s,
n
(s, w)) b(s,
n1
(s, w))ds|
2
_
_
2E
_
_
|
t
_
0
(s,
n
(s, w)) (s,
n1
(s, w)), d(s, w)|
2
_
_
+
+ 2E(|
t
_
0
(b(s,
n
(s, w)) b(s,
n1
(s, w))ds|
2
)
2E(
t
_
0
tr[((s,
n
(s, w)) (s,
n1
(s, w))]
[
(s,
n
(s, w))
(s,
n1
(s, w))]ds]+
+ 2E
_
_
t
t
_
0
|b(s,
n
(s, w)) b(s,
n1
(s, w))|
2
ds
_
_
2dn A
2
t
_
0
n1
(s)ds + 2tA
2
n
t
_
0
n1
(s)ds
205
2dn A
2
(1 + T)
T
_
0
n1
(s)ds.
199
This proves the result.
Step 3.
n
(t)
(C
t)
n+1
(n + 1)!
n in 0 t T, where
C
= max{2nd C
2
(1 + T), and A
2
(1 + T)}.
Proof follows by induction on n.
Step 4.
n
|
[0,T]
is Cauchy in L
2
([0, T] , B([0, T] ), P), where
is the Lebesgue measure on [0, T].
Proof.
n
(t)
(C
t)
n+1
(n + 1)!
implies that
||
n+1

n
||
2
2

(C
T)
n+2
(n + 2)!
.
Here || ||
2
is the norm in L
2
([0, T] ). Thus
n=1
||
n+1

n
||
2
< , proving Step (4).
Step 5. (4) implies that

n
|
[0,T]
is Cauchy in L
2
([0, T] , A
0
,
P) where A
0
is as in Lemma 1. Thus
n
|
[0,T]
converges to
T
in
L
2
([0, T] ) where
T
is progressively measurable.
Step 6. If
n
|
[0,T
2
]
T
2
in L
2
([0, T
2
] ) and 200
n
|
[0,T
1
]
T
1
in L
2
([0, T
1
] ),
then
T
2
|
[0,T
1
]
=
T
1
a.e. on [0, T
1
] , T
1
< T
2
.
This follows from the fact that if
n
in L
2
, a subsequence of
(
n
) converges pointwise a.e. to .
Step 7. Let be dened on [0, ) by |
[0,T]
=
T
. We now show
that
(t, w) = x +
t
_
0
(s, (s, )), d(s, ) +
t
_
0
b(s, (s, ))ds.
Proof. Let 0 t T. By denition,
n
(t, w) = x +
t
_
0
(s,
n1
(s, )), d(s, ) +
t
_
0
b(s,
n1
(s, ))ds.
E
_
_
_
_
t
_
0
((s,
n
(s, )) (s, (s, w))), d(s, w)
_
_
2
_
_
= E(
T
_
0
tr[((s,
n
(s, w)) (s, (s, w)))((s,
n
(s, w)) (s, (s, w)))
ds
dn A
T
_
0
_
|
n
(s, w) (s, w)|
2
ds 0 as n
(by Lipschitz condition on ).
Therefore
t
_
0
(s,
n1
(s, w), d(s, w)
t
_
0
(s, (s, w)), d(s, w)
in L
2
(, P). Similarly, 201
t
_
0
b(s,
n
(s, w))ds
t
_
0
b(s, (s, w))ds, in L
2
.
Thus we get
(t, w) = x +
t
_
0
(s, (s, w)), d(s, w) + (*)
207
+
t
_
0
b(s, (s, w))ds a.e. in t, w.
Step 8. Let (t, w) the right hand side of () above. Then (t, w) is
almost surely continuous because the stochastic integral of a bounded
progressively measurable function is almost surely continuous. The re-
sult follows by noting that [0, ) =
_
n=1
[0, n] and a function on [0, ) is
continuous i it is continuous on [0, n], n.
Step 9. Replacing by in the right side of () we get a solution
(t, w) = x +
t
_
0
(s, ), d +
t
_
0
b(s, (s, w))ds
that is a.s. continuous t and a.e.
Uniqueness. Let and be two progressively measurable a.s. continu-
ous functions satisfying (1). As in Step 3,
E(|(t, w) x|
2
) 2(E(
t
_
0
tr
ds) + 2E(t
t
_
0
b|
2
ds)
2E(
T
_
0
tr
ds + 2E(T
T
_
0
|b|
2
ds), if 0 t T
< .
202
Thus E(|(t, w)|
2
) is bounded in 0 t T. Therefore
(t) = E(|(t, w) (t, w)|
2
)
2E(|(t, w)|
2
) + 2E(|(t, w)|
2
)
and so (t) is bounded in 0 t T. But
(t) 2dn A
2
(1 + T)
t
_
0
(s)ds
as in Step 2; using boundedness of (t) in 0 t T we can nd a
constant C such that
(t) Ct and (t) C
t
_
0
(s)ds, 0 t T.
By iteration (t)
(Ct)
n
n!

(CT)
n
n!
. Therefore
= 0 on [0, T],
i.e. (t, w) = (t, w) a.e. in [0, T]. But rationals being dense in R we
have
= a.e. and t.
It is now clear that I[b, a].
Remark. The above theorem is valid for the equation
(t, w) = x
0
+
t
_
t
0
(s, ), d +
t
_
t
0
b(s, )ds, t t
0
.
This solution will be denoted by
t
0
,x
0
.
Proposition . Let : C[(0, ); R
n
) C([t
0
, ); R
d
) be the map send- 203
ing w to
t
0
,x
0
, P the Brownian measure on C([0, ); R
n
). Let P
t
0
,x
0
=
P
1
be the measure induced on C([t
0
, ); R
d
). Dene X : [t
0
, )
C)[t
0
, ); R
d
) by X(t, w) = w(t). Then X is an It o process relative to
(C([t
0
, ); R
d
),
t
0
, P
t
0
,x
0
) with parameters
[b(t, X
t
), a(t, X
t
)].
The proof of the proposition follows from
209
Exercise. Let (, F
t
, P), (, F
t
, P) be any two measure spaces with X,
Y progressively measurable on , respectively. Suppose :
is such that is (F
t
, F
t
)-measurable for all t, and P
1
= P. Let
X(t, w) = Y(t, w), w . Show that
(a) If X is a martingale, so is Y.
(b) If X I[b(t, X
t
), a(t, X
t
)] then
Y I[b(t, Y
t
), a(t, Y
t
)].
Lemma . Let f : R
2
R be (, P)-measurable,
_
a sub - - algebra
of F. Let X : R and Y : R be such that X is
_
-measurable
and Y is
_
-independent. If g(w) = f (X(w), Y(w)) with E(g(w)) < ,
then
E(g|
)(w) = E( f (x, Y)|

x=X(w)
,
i.e.
E( f (X, Y)|
_
)(w) =
_
f (X(w), Y(w
))dP(w
).
Proof. Let A and B be measurable subsets in R. The result is trivially 204
veried if f = X
AB
. The set
A = {F R : the result is true for X
F
}
is a monotone class containing all measurable rectangles. Thus the
Lemma is true for all characteristic functions. The general result fol-
lows by limiting procedures.
26. Uniqueness of Diusion
Process
IN THE LAST section we proved that 205
(t, w) = x
0
+
t
_
t
0
(s, (s, w)), d(s, w) +
t
_
t
0
b(s, (s, w)ds
has a solution under certain conditions on b and where
= a. The
measure P
t
0
,x
0
= P
1
t
0
,x
0
was constructed on (C([t
0
, ); R
d
), F
t
0
) so that
the map X(t, w) = w(t) is an It o process with parameters b and a. We
now settle the uniqueness question, about the diusion process.
Theorem . Let
(i) a : [0, ) R
d
S
+
d
and b : [0, ) R
d
R
d
be bounded
measurable functions;
(ii) = C([0, ); R
d
);
(iii) X : [0, ) R
d
be dened by X(t, w) = w(t);
(iv) X
t
= {X(s) : 0 s t};
(v) P be any probability measure on
=
_
_
_
t0
X
t
_
_
211
212 26. Uniqueness of Diusion Process
such that P{X(0) = x
0
} = 1 and X is an It o process relative to
(, X
t
, P) with parameters b(t, X
t
) and a(t, X
t
);
(vi) : [0, ) R
d
M
dn
be a bounded measurable map into the
set of all real d n matrices such that
= a on [0, ) R
d
.
Then there exists a generalised n-dimensional Brownian motion 206
on (,
_
t
, Q) and a progressively measurable a.s. continuous map :
[0, ) R
d
satisfying the equation
(1) (t, w) = x
0
+
t
_
0
(s,
_
(s, w)), d(s, w) +
t
_
0
b(s, (s, w))ds
with Q
1
= P, where : is given by ((w))(t) = (t, w).
Roughly speaking, any It o process can be realised by means of a
diusion process governed by equation (1) with
= a.
Proof. Case (i). Assume that there exist constants m, M > 0 such that
mI a(t, x) MI and is a d d matrix satisfying
= a. In this
case we can identify (,
_
t
, Q) with (, F
t
, P). Since D(t, ) is an It o
process,
exp, X(t)
t
_
0
, b(s, X(s, ))ds
1
2
t
_
0
, a(s, X(s, ))ds
is a (, F
t
, P)-martingale. Put
Y(t, w) = X(t, w)
t
_
0
b(s, X(s, w))ds x
0
.
Clearly Y(t, w) is an It o process corresponding to the parameters
[0, a(s, X
s
)],
213
so that
exp, Y(t, w)
1
2
t
_
0
, a(s, X(s, ))ds
is a (, F
t
, P)-martingale. The conditions m a M imply that
1
exists and is bounded. Let
(t) =
t
_
0
1
dY =
t
_
0
1
(s, X(s, ))dY(s, ),
so that (by denition of a stochastic integral) is a (, F
t
, P)-It o pro- 207
cess with parameters zero and
1
a(
1
)
= 1. Thus is a Brownian
motion relative to (, F
t
, P). Now by change of variable formula for
stochastic integrals,
t
_
0
d =
t
_
0
1
dY
= Y(t) Y(0) = Y((t),
since Y(0) = 0. Thus
X(t) = x
0
+
t
_
0
(s, X(s, ))d +
t
_
0
b(s, X(s, ))ds.
Taking Q = P we get the result.
Case (ii). a = 0, b = 0, x
0
= 0, = 0 where M
dn
. Let (
, F
t
, P
)
be an n-dimensional Brownian motion. Dene
(,
t
, Q) = (
, F
t
F
t
, P P
).
If is the n-dimensional Brownian motion on (
, F
t
, P
), we de-
ne on by (t, w, w
) = (t, w
). It is easy to verify that is an

n-dimensional Brownian motion on (,
_
t
, Q). Taking (t, w, w
) = x
0
we get the result.
Before we take up the general case we prove a few Lemmas.
Lemma 1. Let : R
n
R
d
be linear
= a : R
d
R
d
; then 208
there exists a linear map which we denote by
1
: R
d
R
n
such that
1
a
1
=
N
, where denotes the projection and N
null space of
.
Proof. Let R
= range of . Clearly : N
R is an isomorphism.
Let : R
be the inverse. We put
1
= 0 : R
Lemma 2. Let X, Y be martingales relative to (, F

t
, P) and (, F
t
, P)
respectively. Then Z given by
Z(t, w, w) = X(t, w)Y(t, w)
is a martingale relative to
( , F
t
F
t
, P P).
Proof. From the denition it is clear that for every t > s
_
AA
Z(t, w, w)d(P P)|
F
s
F
s
=
_
AA
Z(s, w, w)d(P P)
if A F
s
and A F
s
. The general case follows easily.
As a corrollary to Lemma 2, we have
Lemma 3. Let X be a d-dimensional It o process with parameters b and
a relative to (, F
t
, P) and let Y be a d-dimensional It o process relative
to (, F
t
, P) relative to b and a. Then Z(t, w, w) = (X(t, w), Y(t, w)) is a
(d + d)-dimensional It o process with parameters B = (b, b), A =
_
a 0
0 a
_
relative to ( , F
t
F
t
, P P).
215
Lemma 4. Let X be an It o process relative to (, F
t
, P) with parame- 209
ters 0 and a. If is progressively measurable such that E(
t
_
0
||
2
, ds) <
, t and a
is bounded, then
t
_
0
, dX I[0, a
].
Proof. Let
n
be dened by
i
n
=
_
i
, if || n,
0, otherwise;
Then
t
_
0
n
, dX I[0,
n
a
n
]. Therefore
X
n
(t) = exp(
t
_
0
n
, dX

2
2
t
_
0
n
, a
n
ds
is a martingale converging pointwise to
X(t) = exp
_
t
_
0
, dX

2
2
t
_
0
, ads
_
_
.
To prove that
t
_
0
, dX is an It o process we have only to show that
X
n
(t) is uniformly integrable. Without loss of generality we may assume
that = 1. Let [0, T] be given
E(X
2
n
(t, w)) = E
_
_
exp
_
_
2
t
_
0
n
, dX
t
_
0
n
, a
n
ds
_
_
_
_
= E
_
_
exp
_
_
2
t
_
0
n
, dX 2
t
_
0
n
, a
n
ds
+
t
_
0
n
, a
n
ds
_
_
_
_
.
e
T
sup
0tT
n
, a
n
.
But , a
is bounded and therefore

n
, a
n
is uniformly bounded 210
in n. Therefore (X
n
) are uniformly integrable. Thus X(t, ) is a martin-
gale.
Case (iii). Take d = 1, and assume that
t
_
0
a
1
(a>0)
ds < , t
with a > 0; let = +ve squareroot of a. Dene 1/ = 1/ if > 0,
and 1/ = 0 if = 0. Let
Y(t) = X(t) x
0

t
_
0
b(s, X(s))ds.
Denote by Z the one-dimensional Brownian motion on (
, F
t
, P
)
where
= C([0, ), R). Now

Y I[0, a(s, X(s, ))], Z I[0, 1].
By Lemma 3,
(Y, Z) I
_
(0, 0);
_
a 0
0 1
__
.
If
(t, w, w
) =
t
_
0
_
1
(s, X(s, ))
(>0)
,
=0
, d(Y, Z)
_
then Lemma 4 shows that
I[0, 1].
Therefore is a one-dimensional Brownian motion on = (
, F
t
F
t
, P P
). Put
Y(t, w, w
) = Y(t, w) and X(t, w, w
) = X(t, w);
217
then 211
t
_
0
d =
t
_
0
(>0)
dY +
t
_
0
(=0)
dZ
=
t
_
0
(>0)
dY.
Since
E
_
_
_
_
t
_
0
(=0)
dY
_
_
2
_
_
= E
_
_
t
_
0
(=0)
ds
_
_
= 0,
it follows that
t
_
0
d =
t
_
0
dY = Y(t) = Y(t, w, w
).
Thus,
X(t, w, w
) = x
0
+
t
_
0
(s, X(s, w, w
)d+
+
t
_
0
b(s, X(s, w, w
)ds
with X(t, w, w
) = X(t, w). Now

(P P
)X
1
(A) = (P P
)(A
) = P(A).
Therefore
(P P
)X
1
= P or QX
1
= P.
Case (iv). (General Case). Dene
Y(t, ) = X(t, ) x
0

t
_
0
b(s, X(s, ))ds.
Therefore Y I[0, a(s, X(s, ))] relative to (, F
t
, P). Let Z be the 212
n-dimensional Brownian motion on (
, F
t
, P
) where
= C([0, ); R
n
).
(Y, Z) I
_
(0, 0);
_
a(s, X
s
), 0
0 I
__
Let be a d n matrix such that
= a on [0, ) R
d
. Let
(t, w, w
) =
t
_
0
1
(s, X(s, w))dY(s, w) +
t
_
0
r
N
(s, Z(s, w
))dZ(s, w
)
=
t
_
0
(
1
(s, X(s, w)),
N
(s, Z(s, w
))), d(Y, Z).

Therefore is an Ito process with parameters zero and
A = (
1
,
N
)
_
a 0
0 I
_ _
_
=
1
a(
1
)
+
N
.
=
N
+
N
(for any projection PP
= PP = P)
= I
R
n
.
Therefore is n-dimensional Brownian motion on
(, F
t
, P) = (
, F
t
F
t
, P P
).
t
_
0
(s, X(s, w))d(s, w, w
)
219
=
t
_
0
(s, X(s, w, w
))d(s, w, w
), where X(s, w, w
) = X(s, w),
=
t
_
0
1
dY +
t
_
0
dZ.
=
t
_
0
dY, since
1
=
R
and
N
= 0,
=
t
_
0
(1
R
)dY.
213
Claim.
t
_
0
dY = 0.
For
E
_
_
_
_
t
_
0
dY
_
_
2
_
_
=
t
_
0
a
R
ds =
t
_
0
ds
=
t
_
0
(0)ds = 0.
Therefore we obtain
t
_
0
(s, X(s, w))d(s, w, w
) =
t
_
0
dY = Y(t) Y(0) = Y(t)
putting Y(t, w, w
) = Y(t, w), one gets

X(t, w, w
) = x
0
+
t
_
0
(s, X(s, w, w
))d(s, w, w
)
+
t
_
0
b(s, X(s, w, w
))ds.
As in Case (iii) one shows easily that
(P P
)X
1
= P.
This completes the proof of the theorem.
Corollary . Let a : [0, ) R
d
S
+
d
, and b : [0, ) R
d
R
d
be bounded, progressively measurable functions. If for some choice of
a Lipschitz function : [0, ) R
d
M
dn
,
= a then the It o
process corresponding to [b, a) is unique.
To state the result precisely, let P
1
and P
2
be two probability mea- 214
sures on C([0, ); R
d
) such that X(t, w) = w(t) is an It o process with
parameters b and a. Then P
1
= P
2
.
Proof. By the theorem, there exists a generalised n-dimensional Brow-
nian motion
i
on (
i
,
_
i
t
, Q
i
) and a map
i
:
i
satisfying (for
i = 1, 2)
i
(t,w)
= x
0
+
t
_
0
(s,
i
(s, w))d
i
(s, w) +
t
_
0
b(s,
i
(s, w))ds.
and P
i
= Q
i
1
i
.
Now is Lipschitz so that
i
is unique but we know that the iter-
ations converge to a solution. As the solution is unique the iterations
converge to
i
. Each iteration is progressively measurable with respect
to
i
t
= {
i
(s); 0 s t} so that
i
is also progressively
measurable with respect to F
i
t
. Thus we can restate the result as follows:
There exists (
i
, F
i
t
, Q
i
) and a map
i
:
i
satisfying
i
(t, w) = x
0
+
t
_
0
(s,
i
(s, w))d
i
(s, w)
221
+
t
_
0
b(s,
i
(s, w))ds,
and P
i
= Q
i
1
i
.
(
i
, F
i
t
, Q
i
,
i
) can be identied with the standard Brownian motion
(
, F
t
, Q, ). Thus P
1
= Q
1
= P
2
, completing the proof.
27. On Lipschitz Square
Roots
Lemma . Let f : R R be such that f (x) 0, f (x) C
2
and | f
(x)| 215
A on (, ); then
| f
(x)|
_
f (x)
2A.
Proof.
0 f (y) = f (x) + (y x) f
(x) +
(y x)
2
2
f
()
f (x) + Z f
(x) +
Z
2
2
f
()
where Z = y x, or f (y) f (x) + Z f
(x) +
AZ
2
2
. Therefore
AZ
2
2
+ Z f
(x) + f (x) 0, Z R
| f
(x)|
2
2A f (x).
So
| f
(x)|
2Af (x).
Note. If we take f (x) = x

2
, we note that the constant is the best possible.
Corollary . If f 0, | f
| A, then
|
( f (x
1
))
( f (x
2
))
(A/2)|x
1
x
2
|.
223
224 27. On Lipschitz Square Roots
Proof. Let > 0, then

( f (x) + ) is a smooth function.
(
( f (x) + ))
=
f
(x)
2
( f (x) + )
=
( f (x) + )
( f (x) + )
.
Therefore
|(
( f (x) + ))
(2A/2)
(A/2),
or
|
( f (x
1
) + )
( f (x
2
) + )|
(A/2)|x
1
x
2
|.
216
Let 0 to get the result.
We now consider the general case and give conditions on the matrix
a so that dened by
= a is Lipschitz.
Theorem . Let a : R
n
S
+
d
be continuous and bounded C
2
-function
such that the second derivative is uniformly bounded, i.e. ||D
s
D
r
a
i j
||
M, where M is independent of i, j, r, s; (D
r

d
dx
r
). If : R
n
S
+
d
is
the unique positive square root of a, then
||(x
1
) (x
2
)|| A|x
1
x
2
|, x
1
, x
2
, A = A(M, d).
Proof.
Step 1. Let A S
+
d
be strictly positive such that ||I A|| < 1. Then
A =
(I (I A))
=
r=0
C
r
r!
(I A)
r
,
so that on the set {A : ||I A|| < 1} the map A
A is C
(in fact
analytic).
Now assume that A is any positive denite matrix. Let
1
, . . . ,
n
be the eigen values so that
j
> 0, j = 1, 2 . . . n. Therefore I A is
225
aymmetric with eigen values 1
j
. By choosing suciently small
we can make
||I A|| = max{1
1
, . . . , 1
n
} < 1.
Fixing such an we observe that
A =
1
(A) =
1
(I (I A)).
217
So the map A
A is smooth on the set of symmetric positive

denite matrices.
Step 2. Let n = 1, (t
0
) =
(a(t
0
) where a(t
0
) is positive denite. As-
sume a(t
0
) to be diagonal so that (t
0
) is also diagonal.
i j
(t)
jk
(t) = a
ik
(t).
Dierentiating with respect to t at t = t
0
we get
i j
(t
0
)
jk
(t
0
) +
i j
(t
0
)
jk
(t
0
) = a
ik
(t
0
)
or
a
ii
(t
0
)
ik
(t
0
) +
a
kk
(t
0
)
ik
(t
0
) = a
ik
(t
0
)
or
ik
(t
0
) =
a
ik
(t
0
)
(a
ii
(t
0
)) +
(a
kk
(t
0
))
.
Since the second derivatives are bounded by 4M and a
ii
2a
i j
+a
j j

0, we get
|a
ii
(t) + 2a
i j
(t) + a
j j
(t)|
(8M)
(a
ii
(t) + 2a
i j
(t) + a
j j
(t))
(8M)
(a
ii
+ a
j j
)(t)
or
(1) |a
ii
(t) + 2a
i j
(t) + a
j j
(t)| 4
M(
a
ii
+
a
j j
).
226 27. On Lipschitz Square Roots
Since a is non-negative denite,
|a
ii
(t)|
(2M)
(a
ii
(t)), i.
substituting this in (1), we get 218
|a
i j
(t)| 4
M(
a
ii
+
a
j j
),
and hence
|
i j
(t
0
)| 4
M.
Step 3. Let a(t
0
) be positive denite and its positive denite square
root. There exists a constant unitary matrix such that a(t
0
)
1
= b(t
0
)
is a diagonal positive denite matrix. Let (t
0
) be the positive square
root of b(t
0
) so that
(t
0
) = (t
0
)
1
.
Therefore
(t
0
) = (
1
)(t
0
) where (
(t
0
))
i j
=
i j
(t
0
) and
a
(t
0
) = (
1
b
)(t
0
).
Since is unitary.
|||| = ||||, ||a
|| = ||b
||, ||
|| = ||
||.
By hypothesis, ||b
|| = ||a
|| C(d) M. Therefore
||
|| 4
(MC(d)),
i.e.
||
|| 4
(MC(d)).
Thus ||(t
1
) (t
2
)|| |t
1
t
2
|C(M, d).
Step 4. Let a : R S
+
d
and be the unique non-negative denite
square root of a. For each > 0 let a
= a + I,
= unique positive
square root of a
. Then by step 3,
||
(t
1
)
(t
2
)|| C(M, d)|t
1
t
2
|.
227
If a is diagonal then it is obvious that
as 0. In the
general case reduce a to the diagonal form and conclude that
.
Thus 219
||(t
1
) (t
2
)|| C(M, d)|t
1
t
2
|.
Step 5. Let a : R
n
S
+
d
and
2
= a, with ||D
r
D
s
a
i j
|| M, x, i, j; r,
s R
n
. Choose x
1
, x
2
R
n
. Let x
1
= y
1
, y
2
, . . . , y
n+1
= x
2
be (n + 1)
points such that y
i
and y
i+1
dier almost in one coordinate. By Step 4,
we have
(*) ||(y
i
) (y
i+1
)|| C|y
i
y
i+1
|.
The result follows easily from the fact that
||x||
1
=
n
i=1
|x
i
| and ||x||
2
= (x
1
+ + x
n
)
1/2
are equivalent norms.
This completes the proof of the theorem.
28. Random Time Changes

LET 220
L =
1
2
i, j
a
i j
2
x
i
x
j
+
j
b
j
x
j
with a : R
b
S
+
d
and b : R
d
R
d
bounded measurable funcitons. Let
X(t, ), given by X(t, w) = w(t) for (t, w) in [0, ) C([0, ) : R
d
) be an
It o process corresponding to (, F
t
, Q) with parameters b and a where
= C([0, ); R
d
). For every constant c > 0 dene
L
c
c
_
_
1
2
i, j
a
i j
2
x
i
x
j
+
j
b
j
x
j
_
_
Dene Q
c
by Q
c
= PT
1
c
where (T
c
w)(t) = w(ct). Then one can
show that X is an It o process corresponding to (, F
t
, Q
c
) with param-
eters cb and ca [Note: We have done this in the case where a
i j
=
i j
].
Consider the equation
u
t
= L
c
u with u(0, x) = f (x).
This can be written as
u
= Lu with u(0, x) = f (x) when = ct.

Thus changing time in the dierential equation is equivalent to stretch-
ing time in probablistic language.
So far we have assumed that c is a constant. Now we shall allow c
to be a function of x.
229
230 28. Random Time Changes
Let : R
d
R be any bounded measurable function such that
0 < C
1
(x) < C
2
< , x R
d
and suitable constants C
1
and C
2
. If
L
_
1
2
a
i j
2
x
i
x
j
+
b
j
x
j
_
we dene 221
L
= L
_
1
2
a
i j
2
x
i
x
j
+
b
j
x
j
_
.
In this case we can say that the manner which time changes depends
on the position of the particle.
Dene T
: by
(T
w)(t) = w(
t
(w))
where
t
(w) is the solution of the equation
t
_
0
ds
(w(s))
= t.
As C
1
C
2
it is clear that
t
1
C
1
t
t
1
C
2
. When c a
constant, then
t
= ct and T
coincides with T
c
.
As
0 < C
1
C
2
< ,
_
0
1
(w(s))
ds
is continuous and increases strictly from 0 to as increases, so that
t
exists, is unique, and is a continuous function of t for each xed w.
Some properties of T
.
(i) If l is the constant function taking the value 1 then it is clear that
T
l
= identity.
231
(ii) Let and be two measurable funcitons such that 0 < a (x), 222
(x) b < , x R
d
. Then T
T = T
= T
.
Proof. Fix w. Let
t
be given by
t
_
0
1
(w(s))
ds = t.
Let w
(t) = w(
t
) and let
t
be given by
t
_
0
1
(w
(s))
ds = t.
Let w
(t) = w
(
t
) = w(
t
). Therefore
((T
)w)(t) = (T
)(t) = w
(
t
)
= w
(t) = w(
t
).
Hence to prove the property (ii) we need only show that
t
_
0
1
(w(s))
1
(w(s))
ds = t.
Since
t
_
0
1
(w(s))
ds = t,
dt
d
t
=
1
(w(
t
))
and
dt
d
t
=
1
(w
(
t
))
=
1
(w(
t
))
Therefore
d
t
dt
=
d
t
d
t
t
dt
= (w(
t
))(w
(
t
))
= (w(
t
)(w(
t
))
= ()(w(
t
)).
223
Thus
t
_
0
1
()(w(s))
ds = t.
(iii) From (i) and (ii) it is clear that T
1
= T
1
where
1
=
1
.
(iv) (
t
) is a stopping time relative to
t
. i.e.
_
_
w :
_
0
1
(w(s))
ds r
_
_
for each 0.
(v) T
(w)(t) = w(
t
w) = X
t
(w).
Thus T
is (F
t
F
t
)-measurable, i.e. T
1
(F
t
) F
t
.
Since X(t) is an It o process, with parameters b, a, f C
0
(R
d
),
f (X(t))
t
_
0
(Lf )(X(s))ds is a martingale relative to (, F
t
, P). By the
optional sampling theorem
f (X
t
)
t
_
0
(Lf )(X(s))ds
t
, P), i.e.
f (X
t
)
t
_
0
(Lf )(X(
s
))d
s
233
t
, P). But
d
s
dt
= . Therefore
f (X(
t
))
t
_
0
(Lf )(X
s
)(X
s
)ds
is a martingale. 224
Put Y(t) = X
t
and appeal to the denition of L
to conclude that
f (Y(t))
t
_
0
(L
f )(Y(s))ds
is a martingale. Y(t, w) = X
t
(w) = (T
w)(t). Let F
t
= {Y(s) : 0 s
t}. Then F
t
= T
1
(F
t
) F
t
. Thus
f (Y(t))
t
_
0
(L
f )(Y(s))ds
t
, P). Dene Q = PT
1
so that
f (X(t))
t
_
0
(L
f )(X(s))ds
is an (, F
t
, Q)-martingale, i.e. Q is an It o process that corresponds
to the operator L. Or, PT
1
is an It o process that corresponds to the

operator L.
We have now proved the following theorem.
Theorem . Let = C([0, ); R
d
); X(t, w) = w(t);
L =
1
2
i, j
a
i j
2
x
i
x
j
+
b
j
x
j
.
Suppose that X(t) is an It o process relative to (, F
t
, P) that corre-
sponds to the operator L. Let 0 C
1
C
2
where : R
d
R 225
is measurable. If Q = PT
1
, then X(t) is an It o process relative to

(, F
t
, Q) that corresponds to the operator L.
As 0 < C
1
C
2
, we get 0 < 1/C
2
1/ < 1/C
1
with
T
1
T
= I. We have thus an obvious corollary.

Corollary . There exists a probability measure P on such that X is an
It o process relative to (, F
t
, P) that corresponds to the operator L if
and only if there exists a probability measure Q on such that X is an
Ito process relative to (, F
t
, Q) that corresponds to the operator L.
Remark. If C
2
C
1
> 0 then we have shown that existence and
uniqueness of an It o process for the operator L guarantees existence and
uniqueness of the It o process for the operator L. The solution is no
longer unique if we relax the strict positivity on C
1
as is illustrated by
the following example.
Let a(x) = |x|
1 where 0 < < 1 and let L =

1
2
a
2
x
. Dene
0
on {C([0, ); R)} by
0
(A) =
_
_
1, if A, A,
0, if A,
where is the zero function on [0, ).
Claim.
0
is an It o process with parameters 0 and a. For this it is enough
to show that, f C
0
(R)
f (X(t))
t
_
0
(Lf )(X(s))ds
is a martingale, using a(0) = 0, it follows easily that 226
_
A
t
_
0
(Lf )(X())dd
0
= 0
Borel set A of C([0, ); R) and
_
A
f (X(t))d
0
= 0 if A and
_
A
f (X(t))d
0
= f (0)
235
if A, and this is true t, showing that X(t, w) = w(t) is an It o process
relative to
0
corresponding to the operator L.
Next we shall dene T
a
(as in the theorem); we note that T
a
cannot
be dened everywhere (for example T
a
() is not dened). However T
a
is dened a.e. P where P = P
0
is the Brownian motion.
E
P
_
_
t
_
0
1
|X(s)|
ds
_
_
=
t
_
0
_
0
1
y
(2s)
e
y
2s
dy ds <
since 0 < < 1. Thus by Fubinis theorem,
t
_
0
1
|w(s)|
ds < a.e.
Taking t = 1, 2, 3 . . ., there exists a set
such that P(
) = 1 and
t
_
0
1
|w(s)|
ds < , t, w
Observe that
t
_
0
1
|w(s)|
ds <
implies that
t
_
0
1
|w(s)|
1
ds < ,
for 227
t
_
0
ds
|w(s)|
1
=
=
_
[0,t]{|w(s)|
>1}
ds
|w(s)|
1
+
_
{|w(s)|
1}[0,t]
ds
|w(s)|
, . . .
m{(|w(s)|
> 1)[0, t]} +

t
_
0
1
|w(s)|
ds <
(m = Lebesgue measure)
Thus T
a
is dened on the whole of
. Using the same argument

as in the theorem, it can now be proved that X is an It o process relative
to Q corresponding to the operator L. Finally, we show that Q{} = 0.
Q{} = PT
1
a
{}. Now: T
1
a
{} = empty. For, let w T
1
a
{}. Then
w(
t
) = 0, t, w
. Since |
t

s
| |t s|, one nds that
t
is a
continuous function of t. Further
1
> 0, and w = 0 on [0,
1
] gives
1
_
0
1
|w(s)|
1
ds = .
This is false unless T
1
a
{} = empty. Thus Q{} = 0 and Q is dier-
ent from
0
.
29. Cameron - Martin -
Girsanov Formula
LET US REVIEW what we did in Brownian motion with drift. 228
Let (, F
t
, P) be a d-dimensional Brownian motion with
P{w : w(0) = x} = 1.
Let b : R
d
R
d
be a bounded measurable function and dene
Z(t) = exp
_
_
t
_
0
b, dx
1
2
t
_
0
|b|
2
ds
_
_
.
Then we see that Z(t, ) is an (, F
t
, P)-martingale. We then had a
probability measure Q given by the formula
dQ
dP
F
t
= Z(t, ).
We leave it as an exercise to check that in eect X is an It o process
relative to Q with parameters b and I. In other words we had made a
transition from the operator /2 to /2 + b . We now see whether
such a relation also exists for the more general operator L.
Theorem . Let a : R
d
S
+
d
be bounded and measurable such that
a CI for some C > 0. Let b : R
d
R
d
be bounded, = ([0, ); R
d
),
X(t, w) = w(t), P any probability measure on such that X is an It o
237
238 29. Cameron - Martin - Girsanov Formula
process relative to (, F
t
, P) with parameters [0, a]. Dene Q
t
on F
t
by the rule
dQ
t
dP
F
t
= Z(t, ) = exp
_
_
t
_
0
a
1
b, dX
1
2
t
_
0
b, a
1
bds
_
_
.
229
Then
(i) {0
t
}t 0 is a consistent family.
(ii) there exists a measure Q on (||F
t
):
Q
F
t
= Q
t
.
(iii) X(t) is an It o process relative to (, F
t
, Q) with parameters [b, a],
i.e. it corresponds to the operator
1
2
i, j
a
i j
2
x
i
x
j
+
j
b
j
x
j
.
Proof. (i) Let A(t) =
t
_
0
a
1
b, dX. Then A I[0, b, a
1
b].
Therefore Z(t) is a martingale relative to (, F
t
, P) hence {Q
t
}
t0
is a consistent family.
(ii) Proof as in the case of Brownian motion.
(iii) We have to show that
exp[, X(t, ) ,
t
_
0
bds
1
2
t
_
0
, ads]
t
, Q).
239
Nowfor any function which is progressively measurable and boun-
ded
exp[
t
_
0
, dX
1
2
t
_
0
, ads]
is an (, F
t
, P)-martingale. Replace by (w) = + (a
1
b)((s, w)), 230
where now is a constant vector. Then
exp[
t
_
0
+ a
1
b, dX
1
2
t
_
0
+ a
1
b, ads
is an (, F
t
, P)-martingale, i.e.
exp[, X(t)
1
2
t
_
0
, ads
1
2
t
_
0
a
1
b, a
1
2
t
_
0
, b]
is an (, F
t
, Q)-martingale, and
a
1
b, a = a
a
1
b,
= aa
1
b, (since a = a
)
= b, .
Thus
exp[, X(t)
1
2
t
_
0
, ads
t
_
0
, bds]
is an (, F
t
, Q)-martingale, i.e. X is an It o process relative to (, F
t
, Q)
with parameters [b, a]. This proves the theorem.
We now prove the converse part.
Theorem . Let
L
1
=
1
2
i, ja
i j
2
x
i
x
j
and
L
2

1
2
i, j
a
i j
2
x
i
x
j
+
j
b
j
x
j
,
where a : R
d
S
+
d
is bounded measurable such that a CI for some
C > 0; b : R
d
R
d
is bounded and measurable. Let = C([0, ); R
d
)
with F
t
as usual. Let be a probability measure on (F
t
) and X a 231
progressively measurable function such that X is an It o process relative
to (, F
t
, Q) with parameters [b, a] i.e. X corresponds to the operator
L
2
. Let
Z(t) = exp[
t
_
0
a
1
b, dX +
1
2
t
_
0
b, a
1
bds].
Then
(i) Z(t) is an (, F
t
, Q)-martingale.
(ii) If P
t
is dened on F
t
by
dP
t
dQ
F
t
= Z(t),
Then there exists a probability measure P on (F
t
) such that
P
F
t
= P
t
(iii) X is an It o process relative to (, F
t
, P) corresponding to param-
eters [0, a], i.e. X corresponds to the operator L
1
.
Proof. (i) Let
A(t) =
t
_
0
a
1
b, dX.
Then A(t) is an It o process with parameters [a
1
b, b, a
1
b, b].
241
Thus
exp[A(t)
t
_
0
a
1
b, bds
1
2
t
_
0
a
1
b, bds]
is an (, F
t
, Q)-martingale, i.e. Z(t) is an (, F
t
, Q) martingale.
(ii) By (i), P
t
is a consistent family. The proof that there exists a 232
probability measure P is same as before.
Since X is an It o process relative to Q with parameters b and a,
exp[
t
_
0
, dX
t
_
0
, bds
1
2
t
_
0
, ads]
is a martingale relative to Q for every bounded measurable . Replace
by (w) = (a
1
b)(X(s, w)) where now is a constant vector to get
exp[, X(t)
t
_
0
a
1
b, dX
t
_
0
, b +
t
_
0
a
1
b, bds
1
2
t
_
0
a
1
b, a bds]
is an (, F
t
, Q) martingale, i.e.
exp[, X
t
_
0
a
1
b, dX
t
_
0
, bds +
t
_
0
a
1
b, bds
1
2
t
_
0
, ads
1
2
t
_
0
a
1
b, bds +
1
2
t
_
0
, bds+
+
t
_
0
a
1
b, ads]
is an (, F
t
, Q) martingale. Let R
d
, so that
exp[, X
1
2
t
_
0
, bds
1
2
t
_
0
, ads +
1
2
t
_
0
a
1
b, ads]Z(t)
is an (, F
t
, Q)-matringale and
a
1
b, a = b, (since a = a
).
Therefore 233
exp[, X
1
2
t
_
0
, ads]Z(t)
is an (, F
t
, Q) martingale.
Using the fact that
dP
dQ
F
t
= Z(t), we conclude that
exp[, X
1
2
t
_
0
, ads]
t
, P), i.e. X I[0, a] relative to
(, F
t
, P).
This proves the theorem.
Summary. We have the following situation
L
1
, , F
t
, = C([0, ); R
d
), L
2
, , F
t
.
P a probability measure
such that X is an It o Pro-
cess relative to P corre-
sponding to the operator
L
1
.
_
_
=
_
_
X is an It o process relative
to a probability measure Q
corresponding to L
2
. Q is
given by
dQ
dP
F
t
= Z(t, )
to P corresponding to L
1
where
dP
dQ
F
t
=
1
Z(t, )
_
_
=
_
_
to Q corresponding to L
2
.
243
Thus existence and uniqueness for any system guarantees the exis-
tence and uniqueness for the other system.
Application. (Exercise). 234
Take d = 1, a : R R bounded and measurable with 0 < C
1

a < C
2
< . Let L =
a
2
2
x
2
+ b

x
. Show that there exists a unique
probability masure P on = C([0, ); R) such that X(t) is It o relative
to P corresponding to L. (X(t, w) w(t)) for any given starting point.
30. Behaviour of Diusions
for Large Times
LET L
2
= /2 + b WITH b : R
d
R
d
measurable and bounded 235
on each compact set. We assume that there is no explosion. If P
x
is the
d-dimensional Brownian measure on = C([0, ); R
d
) we know that
there exists a probability measure Q
x
on such that
dQ
x
dP
x
t
= exp
_
_
t
_
0
b, dX
1
2
t
_
0
|b|
2
ds
_
_
Let K be any compact set in R
d
with non-empty interior. We are
interested in nding out how often the trajectories visit K and whether
this frequency depends on the starting point of the trajectory and the
compact set K.
Theorem . Let K be any compact set in R
d
having a non-empty interior.
Let
E
K
= {w : w revisits K for arbitrarily large times}

= {w : there exists a sequence t
1
< t
2
< <
with t
n
such that w(t
n
) K}
Then,
either Q
x
(E
K
) = 0, x, and K,
or Q
x
(E
K
) = 1, x, and K.
245
246 30. Behaviour of Diusions for Large Times
Remark. 1. In the rst case lim
t+
|X(t)| = +a.e. Q
x
, x, i.e. almost
all trajectories stay within K only for a short period.
These trajectories are called transient. In the second case almost 236
all trajectories visit K for arbitrary large times. Such trajectories
are called recurrent.
2. If b = 0 then Q
x
= P
x
. For the case d = 1 or d = 2 we know
that the trajectories are recurrent. If d 3 the trajectories are
transient.
Proof.
Step 1. We introduce the following sets.
E
K
0
= {w : X(t, w) K for some t 0},
E
K
t
0
= {w : X(t, w) K for some t t
0
}, 0 t
0
< .
Then clearly
E
K
_
n=1
E
K
n
=
_
t
0
0
E
K
t
0
.
Let
(x) = Q
x
(E
K
), F =
E
K
.
E
Q
x
(F|F
t
) = E
Q
x
(
E
K
|F
t
) = Q
X(t)
(E
K
)
by the Markov property,
= (X(t)) a.e. Q
x
.
Next we show that (X(t)) is a martingale relative to Q
x
. For, if
s < t,
E
Q
x
((X(t))|F
s
)
= E
Q
x
(E
Q
x
(F|F
t
)|F
s
)
= E
Q
x
(F|F
s
)
= (X(s)).
247
Equating the expectations at time t = 0 and time t one gets 237
(x) =
_
(X(t))dQ
x
=
_
(y)q(t, x, y)dy,
where q(t, x, A) = Q
x
(X
t
A), A Borel in R
d
.
We assume for the present that (x) is continuous (This will be
shown in Lemma 4 in the next section). By denition 0 1.
Step 2. (x) = 1, x or (x) = 0, x.
Suppose that (x
0
) = 0 for some x
0
. Then
0 = (x
0
) =
_
(y)q(t, x
0
, y)dy.
As q > 0 a.e. and 0 we conclude that (y) = 0 a.e. (with respect to
Lebesgue measure). Since is continuous must vanish identically.
If (x
0
) = 1 for some x
0
, we apply the above argument to 1
to conclude that = 1, x. We now show that the third possibility
0 < (x) < 1 can never occur.
Since K is compact and is continuous,
0 < a = inf
yK
(y) sup
yK
(y) = b < 1.
From an Exercise in the section on martingales it follows that
(X(t))
E
K
a.e. Q
x
as t +.
Therefore lim
t
(X(t))(1 (X(t))) = 0 a.e. Q
x
. Now
(x
0
) = Q
x
0
(E
K
) = Q
x
0
{w : w(t) K for arbitrary large time}
Q
x
0
{w : a (X(t, w)) b for arbitrarily large times}
Q
x
0
{w : (1 b)a (X(t))[1 (X(t)] b(1 a)
for arbitrarily large times}
= 0.
238
Thus (x) = 0 identically, which is a contradiction. Thus for the
given compact set K,
either Q
x
(E
K
) = 0, x,
or Q
x
(E
K
) = 1, x.
Step 3. If Q
x
(E
K
0
) = 1 for some compact set K

0
(
K
0
) and x, then
Q
x
(E
K
) = 1, compact set K with non-empty interior.

We rst given an intuitive argument. Suppose Q
x
(E
K
0
) = 1, i.e.
almost all trajectories visit K
0
for arbitrarily large times. Each time a
trajectory hits K
0
, it has some chance of hitting K. Since the trajectory
visits K
0
for arbitrarily large times it will visit K for arbitrarily large
times. We now give a precise arguent. Let
0
= inf{t : X(t) K
0
}
1
= inf{t t
0
+ 1 X(t) K
0
}
. . . . . . . . . . . .
n
= inf{t t
n1
+ 1 : X(t) K
0
}
. . . . . . . . . . . .
Clearly
0
<
1
< . . . < and
n
n.
Q
x
(E
K
n
) Q
x
{X(t) K for t
n
}
Q
x
{X(t) K for t
_
j=n
[
j
,
j
+ 1]}
= 1 Q
x
_
_
_
jn
X(t) K for t [
j
,
j
+ 1]
_
_
1 Q
x
_
_
_
jn
X(
j
+ 1) K
_
_
249
239
We claim that
Q
x
(
_
jn
X(
j
+ 1) K) = 0,
so that Q
x
(E
K
n
) = 1 for every n and hence Q
x
(E
K
) = 1, completing the
proof of the theorem.
Now
q(1, x, K) q(1, x,
K) > 0, x,
K interior of K.
It is clear that if x
n
x, then
limq(1, x
n
,
K) q(1, x,
K).
Let d = inf
xK
0
q(1, x,
K). Then there exists a sequence x

n
in K
0
such
that d = Lt
n
q(1, x
n
,
K). K
0
being compact, there exists a subsequence
y
n
of x
n
with y
n
x in K
0
, so that
d = lim
n
q(1, x,
K) = limq(1, y
n
,
K) q(1, x,
K) > 0.
Thus
inf
xK
0
q(1, x, K) d > 0.
Now
Q
x
_
_
N
_
j=n
X(
j
+ 1) K|F
N
_
_
=
N1
_
j=n
(X(
j
+ 1) K)Q
x
(X(
N
+ 1) K|F
N
) because
j
+ 1
N
for j < N,
=
N1
_
j=n
(
X(
j
+1)
K)Q
X(
N
)
(X(1) K) by the strong
Markov property,
=
N1
_
j=1
q(1, X(
N
), K
c
)
(X(
j
+1)K)
.
240
Therefore
Q
x
_
_
N
_
j=n
X(
j
+ 1) K
_
_
= E
Q
x
(Q
x
(
N
_
j=n
X(
j
+ 1) K|
N
)))
= E
Q
x
_
_
N1
_
j=n
(
[X(
j
+1)K]
)q(1, X(
N
), K
c
)
_
_
Since K
0
is compact and X(
N
) K
0
,
q(1, X(
N
), K
c
) = 1 q(1, X(
N
), K) 1 d
Hence
Q
x
_
_
N
_
j=n
X(
j
+ 1) K
_
_
(1 d)Q
x
_
_
N1
_
j=n
X(
j
+ 1) K
_
_
.
Iterating, we get
Q
x
_
_
N
_
j=n
X(
j
+ 1) K
_
_
(1 d)
Nn+1
, N.
Let N to get 241
Q
x
_
_
_
j=n
X(
j
+ 1) K
_
_
= 0,
since 0 1 d < 1. Thus the claim is proved and so is the theorem.
251
Corollary . Let K be compact,

K . Then Q
x
(E
K
) = 1 if and only if
Q
x
(E
K
0
) = 1, x.
Proof. Suppose Q
x
(E
K
) = 1; then Q
x
(E
K
0
) = 1 because E
K
E
K
0
. Sup-
pose Q
x
(E
K
0
) = 1, then
Q
x
(E
K
n
) = E
Q
x
(E
Q
x
(
E
K
n
|F
n
))
= E
Q
x
(Q
X(n)
(E
K
0
))
= E
Q
x
(1)
= 1, n.
Therefore Q
x
(E
K
) = 1.
Remark. If Q
x
(E
K
) = 0 then it need not imply that

Q
x
(E
K
0
) = 0.
Example . Take b = 0 and d = 3. LEt K = S
1
= {x R
3
such that
|x| 1}. Dene
(n) =
_
_
1, for |x| 1,
1
|x|
, for |x| 1.
P
x
(E
K
0
) constant but P
x
(E
K
) = 0. In fact, P
x
(E
K
0
) = (x) (Refer
Dirichlet Problem).
31. Invariant Probability
Distributions
Denition. Let {P
x
}
xR
d be a family of Markov process on 242
= C([0, ); R
d
)
indexed by the starting points x, with homogeneous transition probabil-
ity p(t, x, A) = P
x
(X
t
A) for every Borel set A in R
d
. A probability
measure on the Borel eld of R
d
is called an invariant distribution if,
A Borel in R
d
.
_
R
d
p(t, x, A)d(x) = (A).
We shall denote dp(t, x, y) by p(t, x, dy) or p(t, x, y)dy if it has a den-
sity.
Proposition . Let L
2
= /2 + b with no explosion. Let Q
x
be the
associated measure. If {Q
x
} has an invariant measure then the process
is recurrent.
Proof. It is enough to show that if K is a compact set with non-empty
interior then
Q
x
(E
K
) = 1
for some x. Also Q
x
(E
K
t
) Q
x
(X
t
K) = q(t, x, K). Therefore
(K) =
_
q(t, x, K)d(x)
_
Q
x
(E
K
t
)d(x).
253
254 31. Invariant Probability Distributions
Now, 0 Q
x
(E
K
t
) 1 and Q
x
(E
K
t
) decreases to Q
x
(E
K
) as t .
Therefore by the dominated convergence theorem
(K)
_
Q
x
(E
K
)d(x).
243
If the process were transient, then Q
x
(E
S
n
) = 0, n, where S
n
=
{x R
d
: |x| n}, i.e. (S
n
) = 0, n. Therefore (R
d
) = 0, which is
false. Thus the process is recurrent.
The converse of this proposition is not true as is seen by the follow-
ing example.
Let L =
1
2
2
x
2
so that we are in a one-dimensional situation (Brow-
nian motion). Then
p(t, x, K) =
_
K
1
(2t)
e
(xy)
2
2t
dy
1
(2t)
(K),
where denotes the Lebesgue measure on R. If there exists an invariant
distribution , then
(K) =
_
p(t, x, K)d(x)
1
(2t)
(K)
_
d(x) =
(K)
(2t)
Letting t , we get (K) = 0 compact K, giving = 0, which
is false.
Theorem . Let L = /2 + b with no explosion. Assume b to be
C
. Dene the formal adjoint L
of L by L
= /2 b (i.e. L
u =
1
2
u (bu)). Suppose there exists a smooth function (C
2
- would
do) such that L
= 0, 0, inf dx = 1. If one denes by the rule

(A) =
_
A
(y)dy, then is an invariant distribution relative to the family
{Q
x
}.
Proof. We assume the following result from the theory of partial dier- 244
ential equations.
255
If f C
0
(G) where G is a bounded open set with a smooth bound-
ary G and f 0, then there exists a smooth function U
G
: [0, )G
[0, ) such that
U
G
t
= LU
G
on (0, ) G,
U
G
(0, x) = f (x) on {0} G,
U
G
(t, x) = 0, x G.
Let t > 0. As U
G
, are smooth and G is bounded, we have
t
_
G
U
G
(t, x)(x)dx =
_
G
t
U
G
ds =
_
G
LU
G
dx
Using Greens formula this can be written as
t
_
G
U
G
(t, x)(x)dx =
_
G
U
G
L

1
2
_
G
_
U
G
n
U
G
n
_
dS +
+
_
G
b nU
G
(t, x)(x)dS
Here n is assumed to be the inward normal to G. So,
t
_
G
U
G
(t, x)(x)dx =
1
2
_
G
(x)
U
G
n
(t, x)dS
(Use the equation satised by and the conditions on U
G
). NowU
G
(t, x)
0, x G, U
G
(t, x) = 0, x in G, so that
U
G
n
(t, x) 0.
This means that 245
t
_
G
U
G
(t, x)(x)dx 0, t > 0,
i.e.
_
G
U
G
(t, x)(x)dx is a monotonically decresing function of t. There-
fore
_
G
U
G
(t, x)(x)dx
_
G
U
G
(0, x)(x)dx
=
_
G
f (x)(x)dx
=
_
R
d
f (x)(x)dx.
Next we prove that if U : [0, )R
d
[0, ) is such that
U
t
= LU,
t > 0 and U(0, x) = f (x), then
_
R
d
U(t, x)(x)dx
_
R
d
f (x)(x)dx.
The solution U
G
(t, x) can be obtained by using It o calculus and is
given by
U
G
(t, x) =
_
f (X(t))
{
G>t
}
dQ
x
.
We already know that
U(t, x) =
_
f (X(t))dQ
x
.
Therefore
_
U(t, x)(x)dx =
f (X(t))(x)DQ
x
dx.
Now
f (X(t))
{
G
>t}
dQ
x
(x)dx
_
U
G
(t, x)(x)dx
_
R
d
f (x)(x)dx.
257
246
Letting G increase to R
d
and using Fatous lemma, we get
f (X(t))(x)dQ
x
dx
_
f (x)(x)dx
This proves the assertion made above.

Let
(A) =
_
A
(X)dx,
(A) =
_
Q
x
(X
t
A)d(x) =
_
q(t, x, A)d(x).
Let f C
0
(G), f 0, where G is a bounded open set with smooth
boundary. Now
_
f (y)d(y) =
f (y)q(t, x, y)d(x)dy
=
f (X(t))dQ
x
d(x)
=
_
U(t, x)d(x)
=
_
U(t, x)(x)dx
_
f (x)(x)dx =
_
f (x)d(x).
Thus, f 0 such that f C
0
,
_
f (x)d(x)
_
f (x)d(x).
This implies that (A) (A) for every Borel set A. (Use mollier 247
s and the dominated convergence theorem to prove the above inequality
for
A
when A is bounded). Therefore (A
c
) (A
c
), or 1 (A) 1
(A), since , are both probability measures. This gives (A) = (A),
i.e.
(A) =
_
q(t, x, A)d(x), t,
i.e. is an invariant distribution.
We now see whether the converse result is true or not. Suppose there
exists a probability measure on R
d
such that
_
Q
x
(X
t
A)d(x) = (A), A Borel in R
d
and t.
The question we have in mind is whether (A) =
_
A
dx for some
smooth satisfying L
= 0, 0,
_
(x)dx = 1. To answer this we
proceed as follows.
By denition (A) =
_
q(t, x, A)d(x). Therefore
f (X(t))dQ
x
d(x)
=
f (y)q(t, x, y)dy d(x)

=
_
f (y)d(y)f C
0
(R
d
)|| f ||
1. (1)
Since X is an It o process relative to Q
x
with parameters b and I,
f (X(t))
t
_
0
(Lf )(X(s))ds
is a martingale. Equating the expectations at time t = 0 and time t we 248
obtain
E
Q
x
( f (X(t)) = f (x) + E
Q
x
_
_
t
_
0
(Lf )(X(s))ds
_
_
Integrating this expression with respect to gives
f (X(t))dQ
x
d(x) =
_
f (x)d(x)
R
d
t
_
0
(Lf )(X(s))ds dQ
x
d.
259
Using (1), we get
0 =
_
R
d
_
t
_
0
(Lf )(X(s))dQ
x
ds d(x)
Applying equation (1) to the function Lf we then get
0 =
_
R
d
t
_
0
(Lf )(y)d(y)ds
= t
_
R
d
(Lf )(y)d(y), t > 0.
Thus
0 =
_
R
d
(Lf )(y)d(y), f C
0
(R
d
).
In the language of distributions this just means that L
= 0.
From the theory of partial dierential equations it then follows that
there exists a smooth function such that A Borel in R
d
,
(A) =
_
A
(y)dy
with L
= 0. As 0, 0 and since
(R
d
) = 1,
_
R
d
(x)dx = 1.
249
We have thus proved the following (converse of the previous) theo-
rem.
Theorem . Let be an invariant distribution with respect to the family
{Q
x
} with b : R
d
R
d
being C
. Then there exists a L
(R
d
), 0,
smooth such that
L
= 0,
_
(y)dy = 1
and such that
(A) =
_
A
(y)dy, A Borel in R
d
.
Theorem (Uniqueness). Let
1
,
2
be smooth on R
d
such that
1
,
2
0, 1 =
_
R
d
1
dy =
_
R
d
2
dy, L
1
= 0 = L
2
.
Then
1
=
2
.
Proof. Let f (x) =
1
(x)
2
(x),
i
(A) =
_
A
i
(x)dx, i = 1, 2.
Then
1
, and
2
are invariant distributions. Therefore
_
q(t, x, y)
i
(x)dx =
_
q(t, x, y)d
i
(x)
=
i
(y), (a.e.), i = 1, 2.
Taking the dierence we obtain
_
q(t, x, y) f (x)dx = f (y), a.e.
Now
_
| f (y) dy =
_
|
_
q(t, x, y) f (x)dx|dy
q(t, x, y)| f (x)|dx dy

=
_
| f (x)|dx
_
q(t, x, y)dy
=
_
| f (x)|dx.
261
250
Thus
(*)
| f (x)|q(t, x, y)dx dy =
_
|
_
q(t, x, y) f (x)dx|dy t.
We show that f does not change sign, i.e. f 0 a.e. or f 0 a.e.
The result then follows from the fact that
_
f (x)dx = 0. Now
|
_
q(1, x, y) f (x)dx|
_
q(1, x, y)| f (x)|dx
and () above gives
_
|
_
q(1, x, y) f (x)dx|dy =
q(1, x, y)| f (x)|dx dy.

Thus
|
_
q(1, x, y) f (x)dx| =
_
q(1, x, y)| f (x)|dx a.e. y,
i.e.
|
_
E
q(1, x, y) f (x)dx +
_
E
q(1, x, y) f (x)dx|
=
_
E
+
q(1, x, y) f (x)dx
_
E
q(1, x, y) f (x)dx a.e. y,

where
E
+
= {x : f (x) > 0}, E
= {x : f (x) < 0}, E

0
= {x : f (x) = 0}.
Squaring both sides of the above equality, we obtain
(**)
_
_
_
E
+
q(1, x, y) f (x)dx
_
_
_
_
_
E
q(1, x, y) f (x)dx
_
_
= 0, a.e. y.
251
Let A be a set of positive Lebesgue measure; then p(1, x, A) =
P
x
(X(1) A) > 0. Since Q
x
is equivalent to P
x
on we have Q
x
(X(1)
A) = q(1, x, A) > 0. Therefore q(1, x, y) > 0 a.e. y for each x. By
Fubinis theorem q(1, x, y) > 0 a.e. x, y. Therefore for almost all y,
q(1, x, y) > 0 for almost all x. Now pick a y such that () holds for
which q(1, x, y) > 0 a.e. x.
We therefore conclude from () that either
_
E
+
q(1, x, y) f (x)dx = 0, in which case f 0 a.e.,
or
_
E
q(1, x, y) f (x)dx = 0, in which case f 0 a.e.

Thus f does not change its sign, which completes the proof.
Remark. The only property of the operator L we used was to conclude
q > 0. We may therefore expect a similar result for more general opera-
tors.
Theorem . Let L
= 0 where 0 is smooth and

_
(x)dx = 1. Let K
be any compact set. Then
sup
xK
_
|q(t, x, y) (y)|dy 0 as t +.
Lemma 1. Let b be bounded and smooth. For every f : R
d
R
d
that is
bounded and measurable let u(t, x) = E
Q
x
( f (X(t)). Then for every xed
t, u(t, x) is a continuous function of x. Further, for t > 0,
|u(t, x)
_
u(t , y)
1
(2)
d
exp
|x y|
2
2
dy|
|| f ||
(e
ct
(e
c
1)),
where c is a constant depending only on ||b||
. 252
263
Proof. Let
(T
t
f )(x) = E
Q
x
( f (X(t)) = E
P
x
( f (X(t))Z(, t))+
+ E
P
x
( f (X(t))(Z(t) Z(, t))),
(1)
where
Z(t) = exp
_
_
t
_
0
b
2
, dx
1
2
t
_
0
|b|
2
ds
_
_
,
Z(, t) = exp
_
_
t
_
b, dx
1
2
t
_
|b(X(s))|
2
ds
_
_
.
E
P
x
( f (X(t))Z(, t)) = E
P
x
(E
P
X
( f (X(t))Z(, t)|
))
= E
P
x
(E
P
X
)( f (X(t ))Z(t )))
(by Markov property),
= E
P
x
(u(t , X()).
(2) =
_
u(t , y)
1
(
(2))
d
exp
_
|(x y)|
2
2
_
dy.
Now
(E
P
x
(|Z(t) Z(, t)|)) =
= E
P
x
(|Z()Z(, t) Z(, t)|))
2
= E
P
x
(Z(, t)Z() 1|))
2
(E
P
x
((Z() 1)
2
))(E
P
x
(Z
2
(, t)))
(by Cauchy Schwarz inequality),
E
P
x
(Z
2
() 2Z() + 1)E
P
x
(Z
2
(, t))
E
P
x
(Z
2
() 1)E
P
x
(Z
2
(, t)), (since E
P
x
(Z()) = 1),
E
P
x
(Z
2
() 1)E
P
x
(exp(2
t
_
b, dX
2
2
t
_
|b|
2
ds +
t
_
|b|
2
ds))
E
P
x
(Z
2
() 1)e
ct
,
using Cauchy Schwarz inequality and the fact that 253
E
P
x
(exp(2
t
_
b, dX
2
2
2
t
_
|b|
2
ds)) = 1.
Thus
E
P
x
(|Z(t) Z(, t)||
2
(e
c
1)e
ct
where c depends only on ||b||
. Hence
|E
P
x
( f (X(t))(Z(t) Z(, t))| || f ||
E
P
x
(|Z(t) Z(, t)|)
|| f ||
((e
c
1)e
ct
). (3)
Substituting (2) and (3) in (1) we get
|u(t, x)
_
u(t y)
1
(
(2)
d
)
exp
_
|x y|
2
2
_
dy
|| f ||
((e
c
1)e
ct
)
Note that the right hand side is independent of x and as 0 the
right hand side converges to 0. Thus to show that u(t, x) is a continuous
function of x (t xed), it is enough to show that
_
u(t , y)
1
(
(2)
d
exp
_
|x y|
2
2
_
dy
is a continuous function of x; but this is clear since u is bounded. Thus 254
for any xed tu(t, x) is continuous.
Lemma 2. For any compact set K R
d
, for r large enough so that
K {x : |x| < r}, x Q
x
(
r
t)
is continuous on K for each t 0, where
r
(w) = inf{s : |w(s)| r}.
265
Proof. Q
x
(
r
t) depends only on the coecient b(x) on |x| r. So
modifying, if necessary, outside |x| r, we can very well assume that
|b(x)| M for all x. Let
r
= inf{s : s , |w(s)| r}.
Q
x
(
r
t) = E
Q
x
(u(X())),
where
u(x) = Q
x
(
r
t )
As b and u are bounded, for every xed > 0, by Lemma 1, Q
x
(
r

t) is a continuous function of x. As
|Q
x
(
r
t) Q
x
(
r
t)| Q
x
(
r
),
to prove the lemma we have only to show that
limit
0
sup
xK
Q
x
(
r
) = 0
Now
Q
x
(
r
) =
_
{
r
}
Z()dP
x
(
_
(Z())
2
dP
x
)
1/2
P
x
(
r
),
by Cauchy-Schwarz inequality. The rst factor is bounded because b is 255
bounded. The second factor tends to zero uniformly on K because
sup
xK
P
x
(
r
) P( sup
0s
|w(s)| > )
where
= inf
yK
|x|=r.
|(x y)|.
Lemma 3. Let K be compact in R

d
. Then for xed t, Q
x
(
r
t) monot-
ically decreses to zero as r and the convergence is uniform on
K.
Proof. Let f
r
(x) = Q
x
(
r
t). As {
r
t} decreases to the null set, f
r
(x)
decreases to zero. As K is compact, there exists an r
0
such that for r
r
0
, f
r
(x) is continuous on K, by Lemma 2. Lemma 3 is a consequence
of Dinis theorem.
Lemma 4. Let b : R
r
R
d
be smooth (not necessarily bounded). Then
E
Q
x
( f (X(t))) is continuous in x for every xed t, f being any bounded
measurable function.
Proof. Let b
r
be any bounded smooth function on R
d
such that b
r
b
on |x| r and Q
r
x
the measure corresponding to b
r
. Then by Lemma 1,
E
Q
x
( f (X(t))) is continuous in x for all r. Further,
|E
Q
r
x
( f (X(t))) E
Q
x
( f (X(t)))| 2|| f ||
Q
x
(
r
t).
The result follows by Lemma 3. 256
Lemma 5. With the hypothesis as the same as in Lemma 1, (S
1
) is an
equicontinuous family, where
S
1
= { f : R
d
R, f bounded measurable, || f ||
1}
Proof. For any f in S
1
, let U(x) = U(t, x) E
Q
x
( f (X(t))) and
U
(x) = U
(t, x) =
_
U(t , y)
1
(
(2)
d
)
exp
_
|x y|
2
2
_
dy.
By Lemma 1,
|U(x) U
(x)| (((e
c
1)
ct
))
1/2
|U(x) U(y)| |U(x) U
(x)| + |U
(y) U(y)| + |U
(x) U
(y)|
2
((e
c
1)e
ct
) + |U
(x) U
(y)|.
The family {U
: f S
1
} is equicontinuous because every U occur-
ing in the expression for U
is bounded by 1, and the exponential factor

is uniformly continuous. Thus the right hand side is very small if is
small and |x y| is small. This proves the lemma.
267
Lemma 6. Let b be smooth and assume that there is no explosion (b
is not necessarily bounded). Then (S
1
) is an equi-continuous family
t > 0.
Proof. Let r > 0 be given. Dene b
r
C
such that b
r
= 0 on |x| > r+1,
b
r
= b on |x| r, b
r
: R
d
R. By Lemma 2, we have that
{E
Q
r
x
( f (X(t))) : f S
1
}
is equicontinuous, where Q
r
x
is the probability measure corresponding 257
to the function b
r
.
(1) E
Q
x
( f (X(t))
{
r
>t}
)E
Q
r
x
( f (X(t))
{
r
>t}
).
Therefore
|E
Q
x
( f (X(t))) E
Q
r
x
( f (X(t))|
= |E
Q
x
( f (X(t))
{
r
>t}
) + E
Q
x
( f (X(t))
{
r
t}
)
E
Q
r
x
( f (X(t))
{
r
>t}
) E
Q
r
x
( f (X(t)))
{
r
t}
)
= |E
Q
x
( f (X(t)
{
r
t}
) E
Q
r
x
( f (X(t))
{
r
t}
)|
|| f ||
(E
Q
x
(
{
r
t}
) + E
Q
r
x
(
{
r
t}
)
l[E
Q
x
(
(
r
t)
) + E
Q
x
(
(
r
t)
)](use (1) with f = 1)
= 2E
Q
x
(
(
r
t)
).
Thus
sup
xK
sup
|| f ||
1
|E
Q
x
( f (X(t)) E
Q
r
x
( f (X(t))| 2 sup
xK
(
{
r
t}
).
By Lemma 3,
sup
xK
E
0
x
(
r
t) 0
for every compact set K as n , for every xed t.
The equicontinuity of the family (S
1
) now follows easily. For xed 258
x
0
, put u
r
(x) = E
Q
r
x
( f (X(t))) and u(x) = E
Q
x
( f (X(t))) and let K =
s[x
0
, 1] = {x : |x x
0
| 1}. Then
|u(x) u(x
0
)| |u(x) u
r
(x)| + |u(x
0
) u
r
(x
0
)| + |u
r
(x) u
r
(x
0
)|
2 sup
yK
E
Q
y
(
(
r
|t)
) + |u
r
(x) u
r
(x
0
)|
By the previous lemma {u
r
} is an equicontinuous family and since
sup
yK
E
Q
y
(
(
r
1)
) 0, {u : || f ||
1} is equicontinuous at x
0
. This
proves the Lemma.
Lemma 7. T
r
T
s
= T
t+s
, s, t 0.
Remark. This property is called the semigroup property.
Proof.
T
r
(T
s
f )(x)
=
f (z)q(s, y, z)q(t, x, y)dy dz.

Thus we have only to show that
_
q(t, x, y)q(s, y, A)dy = q(t + s, x, A).
q(t + s, x, A) = E
Q
x
(X(t + s) A)
= E
Q
x
(X(t + s) A|
t
))
= E
Q
x
(E
Q
x
X(t)(X(s) A))),
by Markov property
= E
Q
x
(q(s, X(t), A))
=
_
q(t, x, y)q(s, y, A)dy,
which proves the result.
As a trivial consequence we have the following. 259
Lemma 8. Let > 0 and let S
1
be the unit ball in B(R
d
). Then
_
t
T
t
(S
1
)
is equicontinuous.
269
Proof.
_
t>0
T
t
(S
1
) = T(
_
t0
T
t
(S
1
)) (by Lemma 7) T
(S
1
).
The result follows by Lemma 6.
Lemma 9. Let u(ttx) = E
Q
x
( f (X(t))) with || f ||
1. Let > 0 be given

and K any compact set. Then there exists a T
0
= T
0
(, K) such that
T T
0
and x
1
, x
2
K,
|u(T, x
1
) u(T, x
2
)| .
Proof. Dene q
(t, x
1
, x
2
, y
1
, y
2
) = q(t, x
1
, y
1
)q(t, x
2
, y
2
) and let Q
(x
1
,x
2
)
be the measure corresponding to the operator
L =
1
2
(
x
1
+
x
2
) + b(x
1
)
x
1
+ b(x
2
)
x
2
i.e., for any u : R
d
R
d
R,
Lu =
1
2
2d
i=1
2
u
x
2
i
+
t
i=1
b
i
(x
1
, . . . , x
d
)
u
x
2
i
+
+
d
i=1
b
i
(x
d+1,...,x
2d
)
u
x
i+d
.
Then Q
(x
1
,x
2
)
will be a measure on C([0, ); R
d
R
d
). We claim that
Q
(x
1
,x
2
)
= Q
x
1
Q
x
2
. Note that
C([0, ); R
d
R
d
) = C([0, ); R
d
) C[(0, ); R
d
)
and since C([0, ); R
d
) is a second countable metric space, the Borel 260
eld of C([0, )R
d
R
d
) is the -algebra generated by
B = (C([0, ); R
d
)) B(C[0, ); R
d
).
By going to the nite-dimensional distributions one can check that
P
(x
1
,x
2
)
= P
x
1
P
x
2
.
dQ
(x
1
,x
2
)
dP
(x
1
,x
2
)
F
t
= exp
_
_
t
_
0
b
(1)
, dX
1

1
2
t
_
0
|b
(1)
|
2
ds
_

exp
_
_
t
_
0
b
(2)
, dX
2

1
2
t
_
0
|b
(2)
|
2
ds
_
_
,
where
b
(1)
(x
1
. . . x
d
) = b(x
1
. . . , x
d
) b
(2)
(x
d+1
. . . x
2d
) = b(x
d+1
, . . . , x
2d
),
so that Q
(x
1
,x
2
)
= Q
x
1
Q
x
2
.
It is clear that if dened an invariant measure for the process Q
x
,
i.e.
_
A
(x)dx =
_
(y)Qy(X
t
A)dy,
then (y
1
)(y
2
) denes an invariant measure for the process Q
(x
1
,x
2
)
.
Thus the process Q
(x
1
,x
2
)
is recurrent.
Next we show that u(T t, X
1
(t)) is a martingale (0 t T) for any
xed T on C([0, T]; R
d
).
E
Q
x
(u(T t, X(t)|F
s
))
= [
_
u(T t, y)q(t s, x, dy)]
x=X(s)
= [
f (z)q(T t, y, dz)q(t s, x, dy)]

x=X(s)
= [
_
f (z)q(T s, x, dz)]
x=X(s)
= u(T s, X(s)), s < t.
261
It now follows that u(T t, X
1
(t)) is a martingale on C([0, ); R
d
)
C([0, ); R
d
). Hence u(T t, X
1
(t)) u(T t, X
2
(t)) is a martingale
relative to Q
(X
1
,x
2
)
.
Let V = S (0, /2) R
d
R
d
with < 1/4. If (x
1
, x
2
) V, then
|x
1
x
2
| |(x
1
, 0) (0, 0)| + |(0, 0) (0, x
2
)| < .
271
Claim 1. Q
(x
1
,x
2
)
(
V
T) 1 as T , where
V
is the exit time
from R
d
V.
Proof. If w is any trajectory starting at some point in V, then
V
=
0 T, T. If w starts at some point outside V then, by the recurrence
property, w has to visit a ball with centre 0 and radius /2; hence it must
get into V at some nite time. Thus {
V
T} to the whole space as
T . Next we show that the convergence is uniform on compact sets.
If x
1
, x
2
K, (x
1
, x
2
) K K (a compact set). Put g
T
(x
1
, x
2
) =
Q
(x
1
,x
2
)
(
V
T). Then g
T
(x
1
, x
2
) 0 and g
T
(x
1
, x
2
) increases to 1 as T
tends to .
g
T
(x
1
, x
2
) = Q
(x
1
,x
2
)
(
V
T)
Q
(x
1
,x
2
)
(
1
V
T),
where 262
1
V
= inf{t 1 : (x
1
, x
2
) V}.
Therefore
g
T
(x
1
, x
2
) E
Q
(x
1
, x
2
)(E
Q
(x
1
, x
2
)((
1
V
T)|
1
))
= E
Q
(x
1
, x
2
)(Q
(X
1
(1),X
2
(1))
{
1
V
T)})
= E
Q
(x
1
, x
2
)(
T
(X
1
(1), X
2
(1))),
where
T
is a bounded non-negative function. Thus, if
h
T
(x
1
, x
2
) = Q
(x
1
,x
2
)
(
1
V
T) =
= E
Q
(x
1
, x
2
)(
T
(X
1
(1), X
2
(1))),
then by Lemma 4, h
T
is continuous for each T, g
T
h
T
and h
T
increases
to 1 as T . Therefore, h
T
converges uniformly (and so does g
T
) on
compact sets.
Thus given > 0 chose T
0
= T
0
(, K) such that if T T
0
,
sup
x
2
K
sup
x
1
K
Q
(x
1
,x
2
)
(
V
T 1) .
By Doobs optional stopping theorem and the fact that
u(T t, X
1
(t)) u(t t, X
2
(t))
is a martingale, we get, on equating expectations,
|u(T, x
1
) u(T, x
2
)|
= |E
Q
(x
1
,x
2
)
[u(T 0, X
1
(0) u(T 0, X
2
(0)]|
= |E
Q
(x
1
,x
2
)
[u(T (
v
(T 1)), X
1
(T (
v
(T 1))
u(T (
v
T(1)), X
2
(T (
v
(T 1)]|
|
_
{
v
T1}
[u(1, X
1
(1)) u(1, X
2
(1))]dQ
(x
1
,x
2
)
+
+
_
{
v
<(T1)}
[u(T
v
, X
1
(T
v
)) u(T
v
), X
2
(T
v
))dQ
(x
1
,x
2
)
|.
263
Therefore
|u(T, x
1
) u(T, x
2
)|
_
{
v
(T1)}
|[u(1, X
1
(1)) u(1, X
2
(1))]|dQ
(x
1
,x
2
)
+
+ |
_
{
v
<(T1)}
[u(T
v
, X
1
(T
v
)) u(T
v
, X
2
(T
v
))dQ
(x
1
,x
2
)
|
2 + |
_
{
v
<(T1)}
[u(T
v
, X
1
(T
v
)) u(T
v
, X
2
(T
v
))]dQ
(x
1
,x
2
)
|,
since u is bounded by 1.
The second integration is to be carried out on the set {T v 1}.
Since
_
t1
T
t
(S
1
) is equicontinuous we can choose a > 0 such that
whenever x
1
, x
2
K such that |x
1
x
2
| <
|u(t, x
1
) u(t, x
2
)| , t 1.
264
Thus |u(T, x
1
) u(T, x
2
)| 3 whenever x
1
, x
2
K and T T
0
.
This proves the Lemma.
273
Corollary to Lemma 9. sup
x
1
,x
2
K
_
|q(t, x
1
, y)|dy converges to 0 as t .
Proof. Since the dual of L
1
is L
, we have
_
|q(t, x
1
, y) q(t, x
2
, y)|dy
= sup
|| f ||
1
|
_
[q(t, x
1
, y) q(t, x
2
, y)] f (y)dy|
and the right side converges to 0 as t , by Lemma 9.
We now come to the proof of the main theorem stated before Lemma
1. Now
_
|q(t, x, y) (y)|dy
=
_
|q(t, x, y)
_
(x
1
)q(t, x
1
, y)dx
1
|dy
(by invariance property)
=
_
|
_
q(t, x, y)(x
1
)dx
1
_
(x
1
)q(t, x
1
, y)dx
1
|dy
(since
_
(x
1
)dx
1
= 1)
|q(t, x, y) q(t, x
1
, y)|(x
1
)dx
1
dy (since 0)
=
_
(x
1
)dx
1
_
|q(t, x, y) q(t, x
1
, y)|dy
Since
_
(x
1
)dx
1
= Lt
n
_
|x
1
|n
(x
1
)dx
1
,
choose a compact set L K such that
_
R
d
L
(x
1
)dx
1
< . Then 265
_
(x
1
)dx
1
_
|q(t, x, y) q(t, x
1
, y)|dy
=
_
L
(x
1
)dx
1
_
|q(t, x, y) q(t, x
1
, y)|dy+
+
_
R
d
L
(x
1
)dx
1
_
|q(t, x, y) q(t, x
1
, y)|dy
L
(x
1
)dx
1
_
|q(t, x, y) q(t, x
1
, y)|dv + 2.
Chose t
0
such that whenever t t
0
,
_
|q(t, x, y) q(t, x
1
, y)|dy
1 +
_
L
(x
1
)dx
1
x, x
1
in L. (Corollary to Lemma 9). Then
_
|q(t, x, y) (y)|dy 3,
if t t
0
x K completing the proof of the theorem.
32. Ergodic Theorem
Theorem . Let f : R
d
R be bounded and measurable with || f ||
1. 266
If is an invariant distribution for the family {Q
x
}, x R
d
then
lim
t
1
0t
2
t
1
E
Q
x
( f (X(t
1
)) f (X(t
2
))) = [
_
f (y)(y)dy]
2
Proof.
E
Q
x
[( f (X(t
1
) f (X(t
2
))]
= E
Q
x
(E
Q
x
[ f (X(t
1
)) f (X(t
2
))|F
t
1
])
= E
Q
x
( f (X(t
1
))(E
Q
x
[ f (X(t
2
))|F
t
1
]))
= E
Q
x
( f (X(t
1
))
_
f (y)q(t
2
t
1
, X(t
1
), y))dy), t
2
> t
1
(by Markov property),
=
_
f (z)q(t
1
, x, z)dz
_
f (y)q(t
2
t
1
, z, y)dy (1)
does any bounded an measurable f . By theorem of 31,
sup
xK
|
_
f (y)[q(t, x, y) (y)]dy| 0
as t +. We can therefore write (1) in the form
E
Q
x
[( f (X(t
1
)) f (X(t
2
))] =
= (
_
f (z)q(t
1
, x, x)dz)
_
f +
_
f (z)q(t
1
, x, z)A(t
2
t
1
, z)dz,
275
276 32. Ergodic Theorem
where A(t
2
t
1
, z) converges to 0 (uniformly on compact sets as) t
2
t
1

+.
To prove the theorem we have therefore only to show that 267
_
f (z)q(t
1
, x, z)A(t
2
t
1
, z)dz 0
as t
1
+and t
2
t
1
(because
_
f (z)q(t
1
, x, z)dz
_
f ). Now
|
_
f (z)q(t
1
, x, z)A(t
2
t
1
, z)dz|
|| f ||
_
q(t
1
, x, z)|A(t
2
t
1
, z)|dz
_
q(t
1
, x, z)|A(t
2
t
1
, z)|dz (2)
Let K be any compact set, then
_
K
q(t
1
, x, z)dz =
_

K
q(t
1
, x, z)dz
_

K
(z)dz
at t
1
. Given > 0, let K be compact so that
|
_

K
c
(z)dz| ;
then |
_

K
c
q(t
1
, x, z)dz| 2 if t
1
0. Using (2) we therefore get
|
_
f (z)q(t
1
x, z)A(t
2
t
1
, z)dz|
_
K
q(t
1
, x, z)|A(t
2
t
1
, z)|dz +
_
K
c
q(t
1
, x, z)|A(t
2
t
1
, z)|dz
_
K
q(t
1
, x, z)|A(t
2
t
1
, z)|dz + 2
_
K
c
q(t
1
, x, z)dz,
since |A(t
2
t
1
, z)| 2,
_
K
q(t
1
, x, z)|A(t
2
t
1
, z)|dz + 2, if t
1
0.
277
The theorem now follows from the fact that 268
lim
t
2
t
1
sup
zK
|A(t
2
t
1
, z) = 0.
Weak Ergodic Theorem.

lim
t
E
Q
x
_
_
|
1
t
t
_
0
f (X(s))ds f (x)(x)dx| >
_
_
= 0.
Proof.
E
Q
x
_
_
|
1
t
t
_
0
f (X(s))ds
_
f (x)(x)dx| >
_
2
E
Q
x
_
_
|
1
t
t
_
0
f (X(s))ds
_
f (y)(y)dy|
2
_
_
,
by Tchebychevs inequality. We show that the right side 0 as t .
Now
E
Q
x
_
_
|
1
t
t
_
0
f (X(s))ds
_
f |
2
_
_
= E
Q
x
[|
1
t
2
t
_
0
t
_
0
f (X(
1
)) f (X(
2
))d
1
d
2
+ (
_
f dy)
2
1
t
t
_
0
f (X())d
_
f dy]
Also
sup
xK
|E
Q
x
[ f (X(t))
_
f (y)(y)dy]|
= sup
xK
|
_
q(t, x, y) f (y)dy
_
f (y)(y)dy|
|| f ||
sup
xK
_
|q(t, x, y) (y)|dy;
the right hand side tends to 0 as t tends to +. Consider 269
|E
Q
x
(
1
t
t
_
0
f (X())d
_
f (y)(y)dy)|
= |E
Q
x
_
_
1
t
T
_
0
f (X())d
_
f (y)(y)dy +
1
t
t
_
0
f (X())d
_
_
, 0 T t,
1
t
|
T
_
0
E
Q
x
f (X())d T
_
f (y)(y)dy|+
+ |E
Q
x
_
_
1
t
t
_
T
f (X())d
_
t T
t
_
_
f (y)(y)dy
_
_
|.
Given > 0 choose T large so that
|E
Q
x
( f (X())
_
f (y)(y)dy| , ( T).
Then
|E
Q
x
_
1
t
t
_
0
f (X())d
_
f (y)(y)dy|
|
1
t
T
_
0
E
Q
x
[ f (X())]
T
t
_
f (y)(y)dy]| +
t T
t

2
provided t is large. Thus
lim
t+
E
Q
x
[
1
t
t
_
0
f (X())d] =
_
f dy.
279
To prove the result we have therefore only to show that
lim
t+
E
Q
x
_
_
1
t
2
t
_
0
t
_
0
f (X(
1
)) f (X(
2
))d
1
d
2
_
_
=
__
f (y)(y)dy
_
2
270
POR is the region
2
t
0
,
1

2
t
0
.
Let I
= E
Q
x
_
_
1
t
t
_
0
t
_
0
f (X(
1
)) f (X(
2
))d
1
d
2
_
__
f dy
_
=
2
t
2
_
_
_
E
Q
x
( f (X(
1
)) f (X(
2
)))
__
f (y)(y)dy
_
2
_
_
d
1
d
2
0
2

1
t.
Then
|I|
2
t
2
_
PQR
|E
Q
x
( f (X(
1
)) f (X(
2
)))
__
f (y)(y)dy
_
2
|d
1
d
2
+
2
t
2
2|| f ||
2
[area of OAB area of PQR]

By the Ergodic theorem the integrand of the rst term on the right
can be made less than /2 provided t
0
is large (see diagram). Therefore
|I|
2

2
t
2
area of PQR +
4
t
2
|| f ||
2
_
t
2
2

_
(t 2t
0
)
2
_
2
_
2
+
2|| f ||
2
t
2
[4tt
0
4t
2
0
].
<
if t is large. This completes the proof of the theorem. 271
33. Application of Stochastic
Integral
LET b BE A bounded function. For every Brownian measure P
x
on 272
= C([0, ); R
d
) we have a probability measure Q
x
on (, F).
Problem. Let q(t, x, A) = Q
x
(X
t
A) q(t, x, ) is a probability measure
on R
d
. We would like to know if q(t, x, ) is given by a density function
on R
d
and study its properties.
Step (i). q(t, x, ) is absolutely continuous with respect to the Lebesgue
measure.
For, p(t, x, A) = P
x
(X(t) A) is given by a density function. There-
fore p(t, x, ) m
d
(Lebesgue measure). Since
Q
x
P
x
on F
t
,
q(t, z, ) M
d
on F
t
.
Step (ii). Let q(t, x, y) 0 be the density function of q(t, x, ) and write
p(t, x, y) for the density of p(t, x, ). Let 1 < < . Put
r(t, x, y) =
q(t, x, y)
p(t, x, y)
.
_
R
d
q
dy =
_
r
dy
=
_
r
p
1/
P
1
dy
__
r
2
pdy
_
1/
281
282 33. Application of Stochastic Integral
__
p
+1
dy
_
1/
Step (iii). 273
Q
x
(X(t) A) =
_
q(t, x, y)dy
=
_
r(t, x, y)p(t, x, y)dy
=
_
r(t, x, y)P
x
(X
t
dy).
Therefore
dQ
x
dP
x
t
t
= r(t, x, y)
Therefore
__
r
2
pdy
_
1/
=
dQ
x
dP
x
t
t
2
,P
x
dQ
x
dP
x
2
,P
x
, since F
t
t
F
t
,
{E
P
x
[Z(t)
2
]}
1/
= {E
P
x
[exp(
2
t
_
0
b, dX
2
2
t
_
0
|b|
2
ds)]}
1/
= {E
P
x
[exp(
2
t
_
0
b, dX

4
2
t
_
0
|b|
2
ds +
4

2
2
t
_
0
|b|
2
ds)]}
1/
,
i.e.,
__
r
2
pdy
_
1/
_
E
P
x
_
_
exp
_
4
a
2
2
ct +
2
t
_
0
b, dX
4
2
t
_
0
|b|
2
ds
_
_
_
_
_
_
1/
where c is such that |b|
2
c. Using Schwarz inequality we then get
__
r
2
pdy
_
1/
_
exp
_
4

2
2
ct
__
1/
.
283
Hence 274
_
q
dy
_
exp
_
4

2
2
ct
__
1/
__
P
+1
dy
_
1/
Signicance. Pure analytical objects like q(t, x, y) can be studied using
stochastic integrals.
Appendix
Language of Probability
Denition. A probability space is a measure space (, B, P) with P() 275
= 1. P is called a probability measure or simply a probability. Elements
of B are called events. A measurable function X : (, B) R
d
is
called d-dimensional random variable. Given the random variable X,
dene F : R
d
R by
F((a
1
, . . . a
n
)) = P{w : X
i
(w) < a
i
, for i = 1, 2, . . . , d}
where X = (X
1
, X
2
, . . . , X
d
). Then F is called the distribution func-
tion of the random variable X. For any random variable X,
_
X dP =
(
_
X
1
dP, . . . ,
_
X
d
dP), if it exists, is called mean of X or expectation of
X and is denoted by E(X). Thus E(X) =
_
XdP = . E(X
n
), where
X
n
= (X
n
1
, X
n
2
, . . . , X
n
d
) is called the n
th
moment about zero. E((X )
n
)
is called the n
th
central moment. The 2nd central moment is called vari-
ance and is denoted by
2
we have the following.
Tchebyshevs Inequality.
Let X be a one-dimensional random variable with mean and variance
. Then for every > 0, P{w : |X(w) | }
2
/
2
.
Generalised Tchebyshevs Inequality. Let f : R R be measurable
such that f (u) = f (u), f is strictly positive and increasing on (0, ). 276
Then for any random variable X : R,
P(w : |X(w)| > )
E( f (X))
f ()
for every > 0.
For any random variable X : R
d
, (t) = E(e
itX
) : R
d
C is
called the characteristic function of X. Here t = (t
1
, . . . , t
d
) and tX =
t
1
X
1
+ t
2
X
2
+ + t
d
X
d
.
285
Independence. Events E
1
, . . . , E
n
are called independent if for every
{i
1
, . . . , i
k
} {1, 2, . . . , n} we have
P(E
i
1
. . . E
i
k
) = P(E
i
1
)P(E
i
2
) . . . P(E
i
k
).
An arbitrary collection of events {E
: I} is called independent
if every nite sub-collection is independent. Let {F
: I} be a
collection of sub--algebras of B. This collection is said to be inde-
pendent if for every collection {E
: I}, where E
, of events
is independent. A collection of random variables {X
: I} is said
to be independent if {(X
) : I} is independent where (X
) is the
-algebra generated by X
.
Theorem . Let X
1
, X
2
, . . . , X
n
be random variables with F
X
1
, . . . , F
X
n
as their distribution functions and let F be distribution function of X =
(X
1
, . . . , X
n
),
X
1
, . . . ,
X
n
the characteristic functions of X
1
, . . . , X
n
and
that of X = (X
1
, . . . , X
n
). X
1
, . . . , X
n
are independent if and only if
F((a
1
, . . . , a
n
)) = F
X
1
(a
1
) . . . F
X
n
(a
n
) for all a
1
, . . . , a
n
, i ((t
1
, . . . , t
n
))
=
X
1
(t
1
) . . .
X
n
(t
n
) for all t
1
, . . . , t
n
. 277
Conditioning.
Theorem . Let X : (, B, P) R
d
be a random variable, with E(X)
nite, i.e. if X = (X
1
, . . . , X
d
), E(X
i
) is nite for each i. Let C be a sub-
-algebra of B. Then there exists a random variable Y : (, C) R
d
such that
_
C
YdP =
_
C
XdP for every C in C.
If Z is any random variable with the same properties then Y = Z
almost everywhere (P).
Denition. Any such Y is called the conditional expectation of X with
respect to C and is denoted by E(X|C).
If X =
A
, the characteristic function of A in B, then E(
A
|C) is
also denoted by P(A|C).
Properties of conditional expectation.
1. E(1|C) = 1.
2. E(aX+bY|C) = aE(X|C) +bE(Y|C) for all real numbers a, b and
random variables X, Y.
3. If X is a one-dimensional random variable and X 0, then
E(X|C) 0.
4. If Y is a bounded C-measurable real valued random variable and
X is a one-dimensional random variable, then
E(YX|C) = YE(X|C).
5. If D C B are -algebras, then
E(E(X|C)|D) = E(X|D).
6.
_
|E(X|D)|d(P|D)
_
E(X|C)d(P|C). 278
Exercise 1. Let (, B, P) be a probability space, C a sub--algebra of
B. Let X(t, )Y(t, ) : R be measurable with respect to B and C
respectively where t ranges over the real line. Further let E(X(t, )|C) =
Y(t, ) for each t. If f is a simple C-measurable function then show that
_
C
X( f (w), w)d(P|C) =
_
C
Y( f (w)w)dP
for every C in C.
[Hint. Let A
1
, . . . , A
n
be a C-measurable partition such that f is con-
stant on each A
i
. Verify the equality when C is replaced by C A
i
.]
Exercise 2. Give conditions on X, Y such that exercise 1 is valid for all
bounded C-measurable functions and prove your claim.
The next lemma exhibits conditioning as a projection on a Hilbert
space.
Lemma . Let (, B, P) be any probability space C a sub--algebra of
B. Then
287
(a) L
2
(, C, P) is a closed subspace of L
2
(, B, P).
(b) If : L
2
(, B, P) L
2
(, C, P) is the projection,then ( f ) =
E( f |C).
Proof. (a) is clear, because for any f L
1
(, C, P)
_
f d(P|C) =
_
f dP
(use simple function 0 s
1
. . . f , if f 0) and L
2
(, C, P) 279
is complete.
(b) To prove this it is enough to verify it for characteristic functions
because both and f E( f |C) are linear and continuous.
Let A B, C C then (
C
) =
C
. As is a projection
_
(
A
)
C
d(P|B) =
_

A
(
C
)d(P|B),
i.e.
_
C
(
A
)d(P|B) =
_
C
X
A
d(P|B).
Since (
A
) is C-measurable,
_
C
(
A
)d(P|B) =
_
C
(
A
)d(P|C)
Therefore
_
C
(
A
)d(P|C) =
_
C
A
d(P|B), C in C.
Hence
(
A
) = E(
A
|C).

Kolmogorovs Theorem.
Statement. Let A be any nonempty set and for each nite ordered subset
(t
1
, t
2
, . . . , t
n
) of A [i.e. (t
1
, . . . , t
n
) an ordered n-tuple with t
i
in A], let
P
(t
1
,...,t
n
)
be a probability on the Borel sets in R
dn
= R
d
R
d
R
d
. 280
Assume that the family P
(t
1
,...,t
n
)
satises the following two conditions
(i) Let : {1, 2, . . . , n} {1, 2, . . . , n} be any permutation and f
:
R
dn
R
dn
be given by
f
((x
1
, . . . , x
n
)) = (x
(1)
, . . . , x
(n)
).
We have
P (E)
(t
(1)
,...,t
(n)
)
= P
(t
1
,...,t
n
)
( f
1
(E))
for every Borel set E of R
dn
. In short, we write this condition as
P
t
= P
t
1
.
(ii) P (E)
(t
1
,...,t
n
)
= P
(t
1
,t
2
,...,t
n
,t
n+1
,...t
n+m
)
(E R
dm
) for all Borel sets E of R
dn
and this is true for all t
1
, . . . , t
n
, t
n+1
, . . . , t
n+m
of A.
Then, there exists a probability space (, B, P) and a collection of
random variable {X
t
: t A} : (, B) R
d
such that
P (E)
(t
1
,...,t
n
)
= P{w : (X
t
1
(w), . . . , X
t
n
(w)) E}
for all Borel sets E of R
dn
.
Proof. Let = {R
d
t
: t A} where R
d
t
= R
d
for each t. Dene
X
t
: R
d
to be the projection given by X
t
(w) = w(t). Let B
0
be
the algebra generated by {X
t
: t A} and B the -algebra generated by
{X
t
: t A}. Having got and B we have to construct a probability P
on (, B) satisfying the conditions of the theorem.
Given t
1
, . . . , t
n
dene 281
(t
1
,...,t
n
)
: R
d
R
d
R
d
(n times)
289
by
(t
1
,...,t
n
)
(w) = (w(t
1
), . . . , w(t
n
)).
It is easy to see that every element of B
0
is
1
(t
1
,...,t
n
)
(E) for suitable
t
1
, . . . , t
n
in A and a suitable Borel set E of R
dn
. Dene P on B
0
by
P(
1
(t
1
,...,t
n
)
(E)) = P
(t
1
,...,t
n
)
(E). Conditions (1) and (2) ensure that P is
a well-dened function on B
0
and that, as P
(t
1
,...,t
n
)
are measures, P is
nitely additive on B
0
.
Claim. Let C
1
C
2
. . . C
n
. . . be a decreasing sequence in B
0
with limit
n
P(C
n
) > 0. Then C
n
is non-empty. Once the claim is
proved, by Kolmogorovs theorem on extension of measures, the nitely
additive set function P can be extended to a measure P on B. One easily
sees that P is a required probability measure.
Proof of the Claim. As C
n
B
0
, we have
C
n
=
1
(t
(n)
1
,...,t
(n)
k(n)
)
(E
n
) for suitable t
(n)
i
in A
and Borel set E
n
in R
dk(n)
. Let
T
n
= (t
(n)
1
, . . . , t
(n)
k(n)
) and A
n
= {t
(n)
, . . . , t
(n)
k(n)
}.
We can very well assume that A
n
is increasing with n. Choose a compact 282
subset E
n
of E
n
such that
P
T
n
(E
n
E
n
) /2
n+1
.
If C
n
=
1
T
n
(E
n
), then P(C
n
C
n
) /2
n+1
. If C
n
= C
1
C
2
. . .C
n
then C
n
C
n
C
n
, C
n
is decreasing and
P(C
n
) P(C
n
)
n
i=1
P(C
i
C
i
) /2.
We prove C
n
is not empty, which proves the claim.
Choose w
n
in C
n
. As
T
1
(w
n
) is in the compact set E
1
for all n,
choose a subsequence
n
(1)
1
, n
(1)
2
, . . . of 1, 2, . . . such that
T
1
(w
n
k
(1))
converges as k . But for nitely many n
(1)
k
s,
T
2
(
n
m
(1)) is in
the compact set E
2
. As before choose a subsequence n
(2)
k
of n
(1)
k
such
that
T
1
(
n
k
(2)) converges as k . By the diagonal process obtain a
subsequence, w
n
of w
n
such that
T
m
(w
n
) converges as n for all m.
Thus, if t is in
_
m=1
A
m
, then limit
n
w
n
(t) = x
t
exists. Dene w by w(t) = 0 if t A

_
m=1
A
m
, w(t) = x
t
if t
_
m=1
A
m
. One
easily sees that w
_
n=1
C
n
, completing the proof of the theorem.
Martingales. 283
Denition. Let (, F, P) be a probability space, (T, ) a totally ordered
set. Let (F
t
)
tT
be an increasing family of sub--algebras of F. A
collection (X
t
)
tT
of random variables on is called a martingale with
respect to the family (F
t
)
tT
if
(i) E(|X
t
|) < , t T;
(ii) X
t
is F
t
-measurable for each t T;
(iii) E(X
t
|F
s
) = X
s
a.s. for each s, t in T with t s. (Markov prop-
erty).
If instead of (iii) one has
(iii)
E(X
t
|F
s
) ()X
s
a.s.,
then (X
t
)
tT
is called a submartingale (respectively supermartin-
gale).
From the denition it is clear that (X
t
)
tT
is a submartingale if
and only if (X
t
)
tT
is a supermartingale, hence it is sucient to
study the properties of only one of these. T is usually any one of
the following sets
[0, ), N, Z, {1, 2, . . . , n}, [0, ] or N {}.
291
Examples. (1) Let (X
n
)
n=1,2...
be a sequence of independent random
variables with
E(X
n
) = 0.
Then Y
n
= X
1
+ +X
n
is a martingale with respect to (F
n
)
n=1,2,...
284
where
F
n
= {Y
1
, . . . , Y
n
} = {X
1
, . . . , X
n
}.
Proof. By denition, each Y
n
is F
n
-measurable.
E(Y
n
) = 0.
E((X
1
+ + X
n
+ X
n+1
+ + X
n+m
)|{X
1
, . . . , X
n
})
= X
1
+ + X
n
+ E((X
n+1
+ + X
n+m
)|{X
1
, . . . , X
n
})
= Y
n
+ E(X
n+1
+ + X
n+m
) = Y
n
.
(2) Let (, F, P) be a probability space, Y a random variable with

E(|Y|) < . Let F
t
F be a -algebra such that t [0, )
F
t
F
s
if t s.
If X
t
= E(Y|F
t
), X
t
is a martingale with respect to (F
t
).
Proof. (i) By denition, X
t
is F
t
-measurable.
(ii) E(X
t
) = E(Y) (by denition) < .
(iii) if t s,
E(X
t
|F
s
) = E(E(Y|F
t
)|F
s
) = E(Y|F
s
) = X
s
Exercise 1. = [0, 1], F = -algebra of all Borel sub sets of , P =

Lebesgue measure.
Let F
n
=-algebra generated by the sets
_
0,
1
2
n
__
1
2
n
,
2
2
n
_
; . . . ,
_
2
n
1
2
n
, 1
_
.
Let f L
[0, 1] and dene 285

X
n
(w) = 2
n
_
_
2
n
1
j=1
[
j1
2
n
,
j
2
n
)
j/2
n
_
j1/2
n
f dy +
[
2
n
1
2
n
,1]
1
_
2
n
1/2
n
f dy
_
_
Show that (X
n
) is a martingale relative to (F
n
).
Exercise. Show that a submartingale or a supermartingale {X
s
} is a mar-
tingale i E(X
s
) = constant.
Theorem . If (X
t
)
tT
, (Y
t
)
tT
are supermartingales then
(i) (aX
t
+ bY
t
)
tT
is a supermartingale, a, b R
+
= [0, ).
(ii) (X
t
Y
t
)
tT
is a supermartingale.
Proof. (i) Clearly Z
t
= aX
t
+ bY
t
is F
t
-measurable and E(|Z
t
|)
aE(|X
t
|) + bE(|Y
t
|) < .
E(aX
t
+ bY
t
|F
s
) = aE(X
t
|F
s
) + bE(Y
t
|F
s
)
aX
s
+ bY
s
= Z
s
, if t s.
(ii) Again X
t
Y
t
is F
t
-measurable and E(|X
t
Y
t
|) < ,
E(X
t
Y
t
|F
s
) E(X
t
|F
s
) X
s
.
Similarly 286
E(X
t
Y
t
|F
s
) E(Y
t
|F
s
) Y
s
, if t s.
Therefore
E(X
t
Y
t
|F
s
) X
s
Y
s
.
Jensens Inequality. Let X be a random variable in (, B, P) with

E(|X|) < and let (x) be a convex function dened on the real line
such that E(|
0
X|) < . Then
(E(X|C)) E(
0
X|C) a.e.
where C is any sub--algebra of B.
293
Proof. The function being convex, there exist sequences a
1
, a
2
, . . . a
n
,
. . . , b
1
, b
2
, . . . of real numbers such that (x) = sup
n
(a
n
x +b
n
) for each x.
Let L
n
(x) = a
n
x + b
n
. Then
L
n
(E(X|C)) = E(L
n
(X)|C) E((X)|C)
for all n so that
(E(X|C)) E((X)|C).
Exercise. (a) If {X
t
: t T} is a martingale with respect to {F
t
: t
T} and is a convex function on the real line such that E(|(X
t
)|) <
for every t, then {(X
t
)} is a sub martingale.
(b) If (X
t
)
tT
is a submartingale and (x) is a convex function and
nondecreasing and if E(|
0
X
t
|) < , t then {(X
t
)} is a sub-
martingale. (Hint: Use Jensens inequality).
Denition. Let (, B, P) be a probability space and (F
t
)
t[0,)
an in- 287
creasing family of sub--algebras of F. Let (X
t
)
t[0,)
be a family of
random variables on such that X
t
is F
t
-measurable for each t 0.
(X
t
) is said to be progressively measurable if
X : [0, t] R dened by X(s, w) = X
s
(w)
is measurable with respect to the -algebra B[0, t] F
t
for every t.
Stopping times. Let us suppose we are playing a game of chance, say,
tossing a coin. The two possible outcomes of a toss are H (Heads) and
T (Tails). We assume that the coin is unbiased so that the probability
of getting a head is the same as the probability of getting a tail. Further
suppose that we gain +1 for every head and lose 1 for every tail. A game
of chance of this sort has the following features.
1. A person starts playing with an initial amount N and nishes with
a certain amount M.
2. Certain rules are specied which allow one to decide when to stop
playing the game. For example, a person may not have sucient
money to play all the games, in which case he may decide to play
only a certain number of games.
It is obvious that such a game of chance is fair in that it is neither ad-
vantageous nor disadvantageous to play such a game and on the average 288
M will equal N, the initial amount. Furthermore, the stopping rules that
are permissible have to be reasonable. The following type of stopping
rule is obviously unreasonable.
Rule. If the rst toss is a tail the person quits at time 0 and if the rst
toss is a head the person quits at time t = 1.
This rule is unreasonable because the decision to quit is made on
the basis of a future event, whereas if the game is fair this decision
should depend only on the events that have already occured. Suppose,
for example, 10 games are played, then the quitting times can be 0,
1, 2, . . . , 10. If
1
, . . . ,
10
are the outcomes (
i
= +1 for H,
i
= 1 for
T) then the quitting time at the 5th stage (say) should depend only on
1
, . . . ,
4
and not any of
5
, . . . ,
10
. If we denote = (
1
, . . . ,
10
) and
the quitting time as a function of then we can say that { : = 5
depends only
1
, . . . ,
4
}. This leads us to the notion of stopping times.
Denition. Let (, F, P) be a probability space, (F
t
)
t[0,)
an increas-
ing family of sub--algebras of F. : [0, ] is called a stopping
time or Markov time (or a random variable independent of the future) if
{w : (w) t} F
t
for each t 0.
Observe that a stopping time is a measurable function with respect
to (F
t
) F.
Examples. 1. = constant is a stopping time. 289
2. For a Brownian motion (X
t
), the hitting time of a closed set is
stopping time.
Exercise 2. Let F
t+

_
Def s>t
F
s
F
t
.
295
[If this is satised for every t 0, F
t
is said to be right continuous]. If
{ < t} F
t
for each t 0, then is a stopping time. (Hint: { t =
_
n=k
{ < t + 1/n} for every k).
We shall denote by F
the -algebra generated by

_
tT
F
t
. If is a
stopping time, we dene
F
= {A F
: A { t} F
t
, t 0}
Exercise 3. (a) Show that F
is a -algebra. (If A F
,
A
c
{ t} = {t t} A { t}).
(b) If = t (constant) show that F
= F
t
.
Theorem . Let and be stopping times. Then
(i) + , v, are all stopping times.
(ii) If , then F
.
(iii) is F
-measurable.
(iv) If A F
, then A { = } and A { } are in F
. In particular, { < }, { = }, { > } are all in 290

F
.
(v) If
is F
-measurable and
, then
is a stopping time.
(vi) If {
n
} is a sequence of stopping times, then lim
n
. lim
n
are also
stopping times provided that F
t+
= F
t
, t 0.
(vii) If
n
, then F
_
n=1
F
n
provided that F
t+
= F
t
, t 0.
Proof. (i)
{ + } > t} = { + > t, t, t} { > t} { > t};
{ + > t, t} = A
=
_
rQ
0rt
{ > r > t , t, t}
(Q = set of rationals)
{ > r > t , t, t} = {t > r} {t > t r}
= { t} { r}
c
{ t} { t r}
c
.
The right side is in F
t
. Therefore + is a scopping time.
{V t} = { t} { t}
{ > t} = { > t} { > t}
(ii) Follows from (iv).
(iii) { t}{ s} = { t s} F
ts
F
s
, s 0.
(iv) A { < } { t} = [A { t < }]
U[A
U
rQ
0rt
{ < } { t}] F
t
.
A { } { t} = A { } { t}. 291
It is now enough to show that ( ) F
; but this is obvious

because ( < ) = ( )
c
is in F
. Therefore A{
} F
and (iv) is proved.

(v) {
t} = {
t} { t} F
t
as (
t) F
. Therefore
is
a stopping time.
(vi) lim
n
sup
n
inf
kn
k
= sup
n
inf
inf{
n
,
n+1
, . . . ,
n+
}.
By (i), inf{
n
,
n+1
, . . . ,
n+
} is a stopping time. Thus we have
only to prove that if
n
or
n
where
n
are stopping times,
then is a stopping time. Let
n
. Then { t} =
_
n=1
{
n
t}
so that is a stopping time. Let
n
. Then
{ t} =
_
n=1
{
n
t}.
297
By Exercise 3, is a stopping time. That lim
n
is a stopping time
is proved similarly.
(vii) Since
n
, n, F
_
n=1
F
n
. Let A
_
n=1
F
n
. Therefore
A (
n
< t) F
t
, n. A ( < t) =
_
m=1
(A (
m
< t)) F
t
.
Therefore A F
Optional Sampling Theorem. (Discrete case). Let {X

1
, . . . , X
k
} be a
martingale relative to {F
1
, . . . , F
k
}. Let {
1
, . . . ,
p
} be a collection of
stopping times relative to {F
1
, . . . , F
k
} such that
1

2
. . .
p
292
a.s. and each
i
takes values in {1, 2, . . . , k}. Then {X
1
, . . . , X
p
} is a
martingale relative to {F
1
, . . . , F
p
} where for any stopping time ,
X
() = X
(w)
().
Proof. It is easy to see that each X
i
is a random variable. In fact X
m
=
k
_
i=1
X
i
{
m
=i}
. Let {1, 2, . . . , k}. Then
E(|X
|)
k
j=1
_
|X
j
|dP < .
Consider
(X
j
t) (
j
s) =
_
s
(X
t) F
s
.
Then (X
j
t) is in F
j
, i.e. X
j
is F
j
-measurable. Next we show
that
(*) E(X
j
|F
k
) X
k
, if j k.
(*) is true if and only if
_
A
X
j
dP
_
A
X
k
dP for every A F
k
.
The theorem is therefore a consequence of the following
Lemma . Let {X
1
, . . . , X
k
} be a supermartingale relative to
{F
1
, . . . , F
k
}.
If and are stopping times relative to {F
1
, . . . , F
k
} taking values in
{1, 2, . . . , k} such that then
_
A
X
dP
_
A
X
dP for every A F
.
Proof. Assume rst that 1. Then 293
_
A
(X
)dP =
k
j=1
_
[A(=j)(<)]
(X
)dP
=
k
j=1
_
[A(=j)]
(X
j
X
j+1
)dP
A F
i
. Therefore A ( = j) F
j
. By supermartingale property
_
[A(=j)]
(X
j
X
j+1
)dP 0.
Therefore
_
A
(X
)dP 0.
Consider now the general case . Dene
n
= ( + n).
Therefore
n
.
n
is a stopping time taking values in {1, 2, . . . , k},
n+1

n
,
n+1

n
1,
k
= .
Therefore
_
A
X
n
dP
_
A
X
n+1
dP, A F
n
. If A F
then A
F
n
, n. Therefore
_
A
X
1
dP
_
A
X
2
dP . . .
_
A
X
k
dP, A F
.
299
Now
1
1.
1
. Therefore
_
A
X
dP
_
A
X
1
dP
_
A
X
dP.
N.B. The equality in (*) follows by applying the argument to 294
{X
1
, . . . , X
k
}.
Corollary 1. Let {X
1
, X
2
, . . . , X
k
} be a super-martingale relative to
{F
1
, . . . , F
k
}.
If is any stopping time, then
E(X
k
) E(X
) E(X
1
).
Proof. Follows from the fact that {X
1
, X
, X
k
} is a supermartingale rela-
tive to {F
1
, F
, F
k
}.
Corollary 2. If {X
1
, X
2
, . . . , X
k
} is a super-martingale relative to
{F
1
, . . . , F
k
}
and is any stopping time, then
E(X
) E(|X
1
|) + 2E(X
k
) 3 sup
1nk
E(|X
n
|)
where for any real x, x
=
|x| x
2
.
Proof. X
k
=
|X
k
| X
k
2
, so 2E(X
k
) = E(|X
k
|) E(X
k
).
By theorem{X
0, X
k
0} is a super-martingale relative to {F
, F
k
}.
Therefore E(X
k
0|F
) E(X
0). Hence
E(X
k
) E(X
) =
E(|X
|) E(X
)
2
.
Therefore
E(|X
|) 2E(X
k
) + E(X
)
2E(X
k
) + E(X
1
) 3 sup
1nk
E(|X
n
|).
Theorem . Let (, F, P) be a probability space and (F

t
)
t0
on increas-
ing family of sub--algebras of F. Let be a nite stopping time, and 295
(X
t
)
t0
a progressively measurable family (i.e. X : [0, ) R
dened by X(t, w) = X
t
(w) is progressively measurable). If X
(w) =
X
(w)
(w), then X
is F
-measurable.
Proof. We show that {w : X((w), w) t, (w) s} F
t
for every t.
Let
s
= {w : (w) s};
s
F
s
and hence the -algebra induced by
F
s
on
s
is precisely
{A
s
: A F
s
} = {A F
s
: A
s
}.
Since (w) is measurable,
w ((w), w) of
s
[0, s]
s
is (F
s
, B[0, s] F
s
)-measurable. Since X is progressively measurable,
[0, s]
s
X
R is measurable.
Therefore {w : X((w), w) t, (w) s} -algebra on
s
. There-
fore X
is F
measurable.
The next theorem gives a condition under which (X
t
)
t0
is progres-
sively measurable.
Theorem . If X
t
is right continuous in t, w and X
t
is F
t
-measurable,
t 0 then (X
t
)
t0
is progressively measurable.
Proof. Dene
X
n
(t, w) = X
_
[nt] + 1
n
, w
_
[nt] + 1
n
t.
301
Then
Lt
n
X
n
(t, w) = X(t, w) (by right continuity)
296
Step 1. Suppose T is rational, T = m/n where m 0 is an integer. Then
{(t, w) : 0 t < T, X
n
(t, w) }
=
_
0im1
_
_
_
i
n
,
i + 1
n
_
X
X
1
i+1
n
(, ]
_
_
Thus if T = m/n, X
n
|
[0,T)
is B[0, T] F
T
-measurable. Now
T =
km
kn
. Letting k , by right continuity of X(t) one gets X|
[0,)
is
[0, T] F
T
-measurable. As X(T) is F
T
-measurable, one gets X|
[0,T]
is [0, T] F
T
-measurable.
Step 2. Let T be irrational. Choose a sequence of rationals S
n
increasing
to T.
{(t, w) : 0 t T, X(t, w) }
=
_
n=1
{(t, w) : 0 t S
n
, X(t, w) } {T} X
1
T
(, ]
The countable union is in B[0, T]F
T
by Step 1. The second mem-
ber is also in B[0, T] F
T
as X(T) is F
T
-measurable. Thus X|
[0,T]
is B
[0,T]
F
T
-measurable when T is irrational also.
Remark. The technique used above is similar to the one used for prov-
ing that a right continuous function f : R R is Borel measurable.
Theorem . Let {X
1
, . . . , X
k
} be a supermartingale and 0. Then 297
(1) P( sup
1nk
X
n
) E(X
1
)
_
_
sup X
n
<
1nk
_
X
k
dP
E(X
1
) + E(X
k
).
(2) P( inf
1nk
X
n
)
_
{inf X
n
}
X
k
dP
E(X
k
).
Proof. Dene
(w) = inf{n : X
n
} if sup X
n
,
= k, if sup
n
X
n
< .
Clearly 0 and is a stopping time. If < k, then X
(w) for
each w.
E(X
) =
_
(sup X
n
)
X
dP +
_
(sup X
n
<)
X
dP
P(sup X
n
) +
_
(sup X
n
<)
X
k
dP.
Therefore
E(X
1
) P(sup X
n
) +
_
(sup X
n
<)
X
k
dP,
P(sup X
n
) E(X
1
)
_
(sup X
n
<)
X
k
dP E(X
1
) + E(X
k
)
The proof of (2) is similar if we dene
(w) =
_
_
inf{n : X
n
}, if inf X
n
,
k, if inf X
n
> .
Kolmogorovs Inequality (Discrete Case). Let {X

1
, . . . , X
k
} be a nite 298
sequence of independent random variables with mean 0. Then
P
_
sup
1nk
(|X
1
+ + X
n
| )
1
2
E((X
1
+ X
2
+ + X
k
)
2
)
_
303
Proof. If S
n
= X
1
+ + X
n
, n = 1, 2, . . . , k, then {S
1
, . . . , S
k
} is a
martingale with respect to {F
1
, . . . , F
k
} where F
n
= {X
1
, . . . , X
n
}.
Therefore S
2
1
, . . . , S
2
k
is a submartingale (since x x
2
is convex). By
the previous theorem,
2
P{inf S
2
n

2
} E((S
2
K
)
)
Therefore
P{sup |S
n
| }
E((S
2
k
)
2
=
E(S
2
k
)
2
=
1
2
E((X
1
+ X
2
+ + X
k
)
2
).
Kolmogorovs Inequality (Continuous case). Let {X(t) : t 0} be a

continuous martingale with E(X(0)) = 0. If 0 < T < , then for any
> 0
P
_
w : sup
0sT
|X(s, w)|
_

1
1
E((X(T))
2
).
Proof. For any positive integer k dene Y
0
= X(0),
Y
1
= X
_
T
2
k
_
X(0), Y
2
= X
_
2T
2
k
_
X
_
T
2
k
_
, . . . , Y
2
k
= X
_
2
k
T
2
k
_
X
_
(2
k
1)
2
k
T
_
.
By Kolmogorov inequality for the discrete case, for any > 0.
P
_
_
sup
0n2
k
|X
_
nT
2
k
_
| >
_
_

1
2
E((X(T))
2
).
By continuity of X(t), A
k
= {w : sup
0n2
k
|X
_
nT
2
k
_
| > } increases to 299
{ sup
0sT
|X(s)| > } so that one gets
(1) P
_
sup
0sT
|X(s)| >
_

1
2
E((X(T))
2
).
Now
P
_
sup
0sT
|X(s)|
_
limit
m
P
_
sup
0sT
|X(s)| >
1
m
_
limit
m
1
( 1/m)
2
E((X(T))
2
), by (1).
= 1/
2
E((X(T))
2
).
Optional Sampling Theorem (Countable case). Let {X
n
: n 1} be
a supermartingale relative to {F
n
: n 1}. Assume that for some
X
L
1
, X
n
E(X
|F
n
). Let , be stopping times taking values in
N {}, with . Dene X
= X
on { = } and X
= X
on
{ = }. Then E(X
|F
) X
.
Proof. We prove the theorem in three steps.
Step 1. Let X
= 0 so that X
n
0. Let
k
= k,
k
= k. By
optional sampling theorem for discrete case E(X
k
) E(X
k
) E(X
1
).
By Fatous lemma, E(X
) < . Again by optional sampling theorem

for the discrete case,
E(X
k
|F
k
) X
k
. . . , (0).
Let A F
. Then A { k} F
k
, and by (0) 300
_
A{k}
X
dP
_
A(k)
X
k
dP
_
A(k)
X
k
dP
_
A(k)
X
dP.
Letting k ,
(1)
_
A()
X
dP
_
A()
X
dP.
Clearly
(2)
_
A(=)
X
dP =
_
A
X
dP =
_
A(=)
X
dP
305
By (1) and (2), inf
A
X dP
_
A
X
dP, proving that

E(X
|F
) X
.
Step 2. Suppose X
n
= E(X
|F
n
). In this case we show that X
=
E(X
|F
) for every stopping time so that E(X
|F
) = X
. If A F
,
then
_
(k)
X
dP =
_
A(K)
X
dP for every k.
Letting k ,
_
A()
X
dP =
_
A()
X
dP, (1)
_
A(=)
X
dP =
_
A
X
dP =
_
A(=)
X
dP (2)
The assertion follows from (1) and (2).
Step 3. Let X
n
be general. Then 301
X
n
= X
n
E(X
|F
n
) + E(X
|F
n
).
Apply Step (1) to Y
n
= X
n
E(X
|F
n
) and Step (2) to
Z
n
= E(X
|F
n
)
to complete the proof.
Uniform Integrability.
Denition. Let (, B, P) be any probability space, L
1
= L
1
(, B, P).
A family H L
1
is called uniformly integrable if for every > 0 there
exists a > 0 such that
_
(|X|)
|X|dP < for all X in H.
Note. Every uniformly integrable family is a bounded family.
Proposition . Let X
n
be a sequence in L
1
and let X
n
X a.e. Then
X
n
X in L
1
i {X
n
: n 1} is uniformly integrable.
Proof. is left as an exercise.
As {X
n
: n 1} is a bounded family, by Fatous lemma X L
1
.
Let > 0 be given. By Egoros theorem there exists a set F such that
P(F) < and X
n
X uniformly on F.
_
|X
n
X|dP ||X
n
X||
,F
+
_
F
|X
n
X|dP
||X
n
X||
,F
+
_
F
|X
n
|dP +
_
F
|X|dP
||X
n
X||
,F
+
_
F(|X
n
|)
|X
n
|dP +
_
F(|X|)
|X|dP+
+
_
F{|X
n
|}
|X
n
|dP +
_
F(|X|)
XdP
||X
n
X||
,F
+
_
(|X
n
|)
|X
n
|dP +
_
(|X|)
|X|dP + 2
302
The result follows by uniform integrability of {X, X
n
: n 1}.
Corollary . Let C be any sub--algebra of B. If X
n
X a.e. and X
n
is
uniformly integrable, then E(X
n
|C) E(X|C) in L
1
(, C, P).
Proposition . Let H L
1
. Suppose there exists an increasing convex
function G : [0, ) [0, ) such that
limit
t
G(t)
t
= and sup
XH
E(G(|X|)) < .
Then the family H is uniformly integrable.
307
Example. G(t) = t
2
is a function satisfying the conditions of the theo-
rem.
Proof. (of the proposition). Let
M = sup
XH
E(G(|X|)).
Let > 0 be given. Choose > 0 such that
G(t)
t

M
for t .
Then for X in H
_
(|X|)
|X|dP
M
_
(|X|)
G(|X|)dP
M
_
G
G(|X|)dP
Remark. The converse of the theorem is also true.

Exercise. Let H be a bounded set in L
, i.e. there exists a constant M

such that ||X||
M for all X in H. Then H is uniformly integrable.

Up Crossings and Down Crossings. 303
Denition. Let a < b be real numbers; let s
1
, s
2
, . . . , s
k
be also given
reals. Dene i
1
, i
2
, . . . , i
k
as follows.
i
1
=
_
_
inf{n : s
n
< a},
k, if no s
i
< a;
i
2
=
_
_
inf{n > i
1
: s
n
> b},
k, if s
n
b for each n > i
1
;
i
3
=
_
_
inf{n > i
2
: s
n
< a},
k, if s
n
a for each n > i
2
;
and so on
Let t
1
= s
i
1
, t
2
= s
i
2
, . . .. If (t
1
, t
2
), (t
3
, t
4
), . . ., (t
2p1
, t
2p
) are the
only non-empty intervals and (t
2p+1
, t
2p+2
), . . . are all empty, then p is
called the ******** of the sequence s
1
, . . . , s
k
for the interval [a, b] and
is denoted by U(s
1
, . . . , s
k
; [a, b]).
Note. U (the up crossing) always takes values in {0, 1, 2, 3, . . .}.
Denition. For any subset S of reals dene
U(S ; [a, b]) = sup{U(F; [a, b]) : F is a nite subset of S }
The number of down crossings is dened by
D(S ; [a, b]) = U(S ; [b, a]).
For any real valued function f on any set S we dene
U( f , S, [a, b]) = U( f (S ), [a, b]).
If the domain of S is known, we usually suppress it.
Proposition . Let a
1
, a
2
, . . . be any sequence of real numbers and S = 304
{a
1
, a
2
, . . .}. If U(S, [a, b]) < for all a < b, then these sequence {a
n
}
is a convergent sequence.
Proof. It is clear that if T S then U(T, [a, b]) U(S, [a, b]). If
the sequence were not convergent, then we can nd a and b such that
liminf a
n
< a < b < limsup a
n
. Choose n
1
< n
2
< n
3
. . .; m
1
< m
2
<
. . . such that a
n
i
< a and a
m
i
> b for all i. If T = {a
n
1
, a
m
1
, a
n
2
, a
m
2
, . . .},
then U(S ; [a, b]) U(T; [a, b]) = ; a contradiction.
Remark. The converse of the proposition is also true.
Theorem . (Doobs inequalities for up crossings and down crossings).
Let {X
1
, . . . , X
k
} be a submartingale relative to {F
1
, . . . , F
k
} a < b.
Dene U(w, [a, b]) = U(X
1
(w), . . . , X
k
(w); [a, b]) and similarly dene
D(w, [a, b]). Then
(i) U, D are measurable functions;
309
(ii) E(U(, [a, b]))
E((X
k
a) + 1) E((X
1
a)
+
)
b a
;
(iii) E(D(, [a b])) E((X
k
b)
+
)/(b a).
Proof. (i) is left as an exercise.
(ii) Dene Y
n
= (X
n
a)
+
; there are submartingales. Then clearly
Y
n
0 if and only if X
n
a and Y
n
b a i X
n
b, so that
UY
1
(w), . . . , Y
k
(w); [0, b a]) = U(X
1
(w), . . . , X
k
(w); [a, b])
305
Dene
1
= 1
2
=
_
_
inf{n : Y
n
= 0}
k, if each Y
n
= 0
3
=
_
_
inf{n >
2
: Y
n
> b a,
k, if Y
n
< b a for each n >
2
;
k+1
= k.
As {Y
1
, . . . , Y
k
} is a submartingale, by optional sampling theorem
Y
1
, . . . , Y
k+1
is also a submartingale. Thus
(1) E(Y
2
Y
1
) + E(Y
4
Y
3
) + 0.
Clearly
[(Y
3
Y
2
) + (Y
5
Y
4
) + ](w) (b a) (Y
1
(w), . . . Y
k
(w);
[0, b a]) = (b a) (w, [a, b]).
Therefore
(2) E(Y
3
Y
2
) + E(Y
5
Y
4
) + (b a)E(U(, [a, b])).
By (1) and (2),
E(Y
k
Y
1
) (b a)E(U(, [a, b]))
giving the result.
(iii) Let Y
n
= (X
n
a)
+
so that
D(Y
1
(w), . . . Y
k
(w); [0, b a]) = D(X
1
(w), . . . , X
k
(w); [a, b])
Dene 306
1
= 1;
2
=
_
_
inf{n : Y
n
b a},
k, if each Y
n
< b a;
3
=
_
_
inf{n >
2
: Y
n
= 0},
k, if each Y
n
> 0 for each n >
2
;
k+1
= k.
By optional sampling theorem we get
0 E(Y
2
Y
3
) + E(Y
4
Y
5
) + .
Therefore
0 (b a)E(D(Y
1
, . . . , Y
k
; [0, b a])) + E((b a) Y
k
).
Hence
E(D(, [a, b])) E((X
k
a)
+
(b a))/(b a)
E((X
k
b)
+
)
(b a)
, for (c a)
+
(b a) (c b)
+
for all c.
311
Corollary . Let {X
1
, . . . , X
k
} be a supermartingale. U, D as in theorem.
Then
(i) E(D(, [a, b]))
E(X
1
b) E(X
k
b)
b a
.
(ii) E(U(, [a, b]))
E((X
k
b)
)
b a
.
Proof. (i) E(D(, [a, b])) = E(U(X
1
(w), . . . , X
k
(w), [b, a])
E((X
k
+ b)
+
(X
1
+ b)
+
)
b a
, by above theorem,
E((b X
k
) (b X
1
))
b a
, since for
since for all a, b, c, (b c)
+
(b a)
+
(b a) (b c).
(ii) E(U(, [a, b])) =
= E(D(X
1
(w), . . . , X
k
(w); [b, a]))
E((X
k
+ a)
+
)
b a
, by theorem,
E((X
k
b)
)
b a
, 307
(since (X
k
+ a)
+
(X
k
b)
Theorem . Let {X
n
: n = 1, 2, . . .} be a supermartingale relative to {F
n
:
n = 1, 2, . . .}. Let (, F, P) be complete.
(i) If sup
n
E(X
n
) < , then X
n
converges a.e. to a random variable
denoted by X
.
(ii) if {X
n
: n 1} is uniformly integrable, then also X
exists. Fur-
ther, {X
n
: n = 1, 2, . . . , n = } is a supermartingale with the
natural order.
(iii) if {X
n
: n 1} is a martingale, then {X
n
: n 1, n = } is a
martingale.
Proof. (i) Let U(w[a, b]) = U(X
1
(w), X
2
(w), . . . , [a, b]). By the coro-
llary to Doobs inequalities theorem,
E(U(, [a, b]) sup
n
E((X
n
b)
) <
for all a < b. Allowing a, b to vary over the rationals alone we 308
nd that the sequence X
n
is convergent a.e.
(ii) Sup
n
E(X
n
) sup
n
E(|X
n
|) < so that X
exists. As X
n
X
in
L
1
we get that {X
n
: n 1, n = } is a supermartingale.
(iii) follows from (ii).
Proposition . Let {X
t
: t 0} be a supermartingale relative to {F
t
: t
0}. I = [r, s], a < b and S any countable dense subset. Let U(w, S
I, [a, b]) = U(, {X
t
(w) : t S I}, [a, b]). Then
E(U(, S I, [a, b))
E((X
s
b)
)
b a
.
Proof. Let S I be an increasing union of nite sets F
n
: then
E(U(, F
n
, [a, b]))
E((X
max F
n
b)
)
b a

E((X
s
b)
)
b a
.
The result follows by Fatous lemma.
Exercise . If further X
t
is continuous i.e. t X
t
(w) is continuous for
each w, then prove that
E(U(, I, [a, b]))
E((X
s
b)
)
b a
Theorem . Let (, F, P) be complete and {X
t
: t 0} a continuous
supermartingale.
(i) If sup
t0
E(X
t
) < , then X
t
converges a.e. to a random variable 309
X
.
313
(ii) If {X
t
: t 0} is uniformly integrable then also X
exists and
{X
t
: t 0, t = } is a supermartingale.
Proof. (i) E(U(, [0, n], [a, b])) E((X
n
b)
)/(b a) so that
limit
n
E(U(, [0, n], [a, b])) sup
0s
E((X
s
b)
)
b a
for all a < b. Thus {X
t
(w) : t > 0} converges a.e. whose limit in
denoted by X
which is measurable.
(ii) As E(X
t
) E(|X
t
|) by (i) X
exists, the other assertion is a con-

sequence of uniform integrability.
Corollary . Let {X
t
: t 0} be a continuous uniformly integrable mar-
tingale. Then {X
t
: 0 t } is also a martingale.
Exercise. Let {X
t
: t 0} be a continuous martingale such that for some
Y with 0 Y 1 E(Y|F
t
) = X
t
show that X
t
Y a.e.
Lemma . Let (, F, P) be a probability space, F
1
F
2
F
3
. . . be
sub--algebras. Let X
1
, X
2
, . . . be a real valued functions measurable
with respect to F
1
, . . . , F
n
, . . . respectively. Let
(i) E(X
n1
|F
n
) X
n
(ii) sup
n
E(X
n
) < .
Then {X
n
: n 1} is uniformly integrable. 310
Proof. By (i) E(X
n
) is increasing. By (ii) given > 0, we can nd n
0
such that if n n
0
then E(X
n
) E(X
n
0
) + . For and > 0,
n n
0
_
(|X
n
|)
|X
n
|dP
= E(X
n
) +
_
(X
n
)
X
n
dP
_
(X
n
<)
X
n
dP
_
(X
n
)
X
n
0
dP
_
(X
n
<)
X
n
0
dP + E(X
n
) by (i)
+
_
(X
n
)
X
n
0
dP
_
(X
n
)
X
n
0
dP (because E(X
n
) E(X
n
0
) + )
+
_
(|X
n
|)
|X
n
0
|dP
Thus to show uniform integrability we have only to show P(|X
n
|
) 0 uniformly in n as . Now
E(|X
n
|) = E(X
n
+ 2X
n
)
E(X
n
) + 2E(|X
1
|) by (i)
M < for all n by (ii)
The result follows as P(|X
n
| ) M/.
Optional Sampling Theorem. (Continuous case).
Let {X
t
: t 0} be a right continuous supermartingale relative to
{F
t
: t 0}. Assume there exists an X
(, F, P) such that X
t

E(X
|F
t
) for t 0. For any stopping time taking values in [0, ], let 311
X
= X
on { = }. Then
(i) X
is integrable.
(ii) If are stopping times, then
E(X
|F
) X
.
Proof. Dene
n
=
[2
n
] + 1
2
n
,
n
=
[2
n
] + 1
2
n
.
Then
n
,
n
are stopping times,
n

n
,
n
,
n
.
n
,
n
take values in D
n
= {, 0, 1/2
n
, 2/2
n
, . . . , 1/2
n
, . . .} so that we have
E(X
n
|F
n
) X
n
. Thus if, A F
n
, then
(*)
_
A
X
n
dP
_
A
X
n
dP
315
As
1

2
. . ., by optional sampling theorem for the countable
case, we have
E(X
n1
|F
n
) X
n
.
Further
F
1
F
2
. . . ; E(X
n
|F
0
) X
0
.
By the lemma {X
n
}, {X
n
} are uniformly integrable families. By
right continuity X
n
X
pointwise and X
n
X
pointwise. Letting
n in (*) we get the required result.
Lemma (Integration by Parts). Let M(t, ) be a continuous progres- 312
sively measurable martingale and A(t, w) : [0, ) R be of
bounded variation for each w. Further, assume that A(t, w) is F
t
-measu-
rable for each t. Then
Y(t, ) = M(t, )A(t, ) =
t
_
0
M(s, )dA(s, )
is a martingale if
E( sup
0st
|M(s, )| ||A()||
t
) <
for each t, where ||A(w)||
t
is the total variation of A(s, w) in [0, t].
Proof. By hypothesis,
n
i=0
M(s, )(A(s
i+1
, ) A(s
i
, ))
converges to
t
_
s
M(u, )dA(u, ) in L
1
as n
and as the norm of the partition s = s
0
< s
1
< . . . < s
n+1
= t converges
to zero. Hence it is enough to show that
E([M(t, )A(t, )
n
i=0
M(s
i+1
, )(A(s
i+1
, ) A(s
i
, ))]|F
s
)
= M(s, )A(s, ).
But the left side = E(M(s
n+1
, )A(s
n+1
, )
i=0
M(s
i+1
, )(A(s
i+1
, ) A(s
i
, ))|F
s
)
= M(s, )A(s, ).
Taking limits as n and observing that 313
sup
0in
|(s
i+1
s
i
)| 0,
we get
E(M(t, )A(t, )
t
_
0
M(u, )dA(u, )|F
s
)
= M(s, )A(s, )
s
_
0
M(u, )dA(u, ).
Bibliography
[1] BILLINGSLEY, P. Convergence of probability measures, John Wi- 314
ley, New York, (1968).
[2] DOOB, J.L. Stochastic Processes, John Wiley, New York (1952).
[3] DYNKIN, E.B. Markov Processes and Problems in Analysis, Proc.
Int. Cong of Math. Stockholm, (1962) 36 - 58.
[4] DYNKIN, E.B. Markov Processes, Vols. 1, 2, Springer-Verlag,
Berlin (1965).
[5] FRIEDMAN, AVNER. Stochastic Dierential Equations and Ap-
plications, vols. 1, 2, Academic Press, New York (1976).
[6] FRIEDMAN, AVNER. Partial Dierential Equations of Parabolic
type, Prentice Hall, Englewood Clis (1964).
[7] GIHMAN, I.I. and A.V. SKOROHOD. Stochastic Dierential
Equations, Springer-Verlag, New York (1973).
[8] IT

O, K. On Stochastic Dierential Equations, Memoirs Amer. Math.
Soc. 4 (1951).
[9] IT

O, K. and H.P. McKEAN, Jr. Diusion Processes and their sam-
ple paths, Springer-Verlag, Berlin (1965).
[10] KUNITA, H. and S. WATANABE. On Square Integrable Martin-
gales, Nagoya Math. Journal. 30 (1967) 209 - 245.
317
318 BIBLIOGRAPHY
[11] L
EVY, P. Processus Stochastic et Mouvement Brownien, Gaulhier-

Villars, Paris (1948).
[12] MEYER, P.A. Probability and Potentials, Blaisdell, Wallham 315
(1966).
[13] NEVEU, J. Discrete Parameter Martingales, North Holland, Am-
sterdam (1975).
[14] PARTHASARATHY, K.R. Probability Measures on Metric
Spaces, Academic Press, New York (1967).
[15] PROTTER, M.H. and H.F. WEINBERGET. Maximum Principles
in Dierential Equations, Printice Hall, Englewood Clis (1967).
[16] STROOCK, D.W. and S.R.S. VARADHAN. Diusion Process in
Several Dimension, Springer-Verlag, New York (to appear)
[17] WILLIAMS, D. Brownian motions and diusions as Markov Pro-
cesses. Bull. Lond. Math. Soc. 6 (1974) 257 - 303.

S R S Varadhan Tata Institute of Fundamental Research Lectures On Diffusion Problems and Partial Differential Equations 1981

Transféré par

Informations du document

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

S R S Varadhan Tata Institute of Fundamental Research Lectures On Diffusion Problems and Partial Differential Equations 1981

Transféré par

Droits d'auteur :

Formats disponibles

Lectures on

Diusion Problems and Partial Dierential

. For part (b) make the substitution

and apply Lebesgue dominated convergence theorem).

must attain a minimum.

f (z)p(s, y z)p(t, x y)dz dy.

Exercise 6. Prove Lemma 1 directly using Fourier transforms.

0 or +, x may approach + in a nite

1). This is the equation satised

(t) is a Brownian motion starting at (0, 0) for every > 0.

This formal calculation leads us to

Examples of Stopping Times.

and for every A

. Now assertion (1) follows easily. 38

Proof of the reection principle.

= {w : > 0 s [0, ) such that X

) = 1 and therefore P(A) = P(A

The reection principle (or the strong Markov property) allows us

< . Therefore w A or not according

dt converges or not. But this convergence can be asserted 56

dt < a.e.w., or,

> n) (by continuity of Brownian paths)

Remark. A stronger result holds:

is the cone of innite height obtained from C

consists of trajectories which start at a point of G and escape for

{w : w(0) = y}. Then there exists w

Remark. The theorem is local.

(x) = 1 for |x| . As Brownian motion has the Markov

for arbitrarily large t) = 0. In other words, P(w : 69

(t) is a martingale, . Therefore

(t) is a Martingale and hence E(Y

) < . Since a is bounded

(s, )]d dP(w).

Theorem . Let (s, w) be progressively measurable, such that

Denition. With the hypothesis as in the above theorem we dene the

2 Sin 1/2[(2n + 1)t]n = 0, 1, 2, . . . .

(t, w) = exp[, X(t, w)

(t, w)|) < is a consequence of (i) and the fact that

(t, w)|) 1 + t d|| ||b||

function with compact support such that

|x| where C is also independent of n and x. By (v),

dierent from B; more specically for B

Exercise. (It os formula for the general case). Let

(ii) Let (t, x) = (1,

. Therefore, by the previous Lemma,

18. The Feynman-Kac

, < 2, on [s, T).

function such that = 1 on |X| R, and = 0 outside

is a constant given in terms of the constants A and T and the

], < 0. We claim that satises 132

)] exists, i.e. the decay rate is connected to the existence or

(t, x) is bounded by 1, if 0. By the Feynman-

(t, x)). Therefore

is also bounded by 1. This forces A = D = 0. We demand that

converges to uniformly on compact sets, (ii)

uniformly on compact sets, (iii)

((s)) converges 139

and B { < t} so that B F

Proposition . Let b : [0, ) R

Alternate Proof. 147

21. Integral Equations

Corollary . If b is locally Lipschitz and bounded, then B(X) = for all

for 0 t K, n = 0, 1, 2, 3, . . ., so that we can