Introduction To Stochastic Analysis: Lecture Notes

Lecture Notes
Introduction to Stochastic Analysis

Michael Rockner
Universitat Bielefeld
Version: 2. Februar 2012
corrected until and including p. 63
This script has been written by Matthias Stephan on behalf of Prof. M. Rockner and has been
revised by Dr. Judith Dohmann. Please note that despite several checks it may still contain
mistakes and, therefore, there is no guarantee of correctness. A report of any kind of errors
found in these notes is appreciated (mstephan@math.uni-bielefeld.de).
Inhaltsverzeichnis
1. Introduction to Pathwise Ito-Calculus 1
1.1. Preparation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1. Quadratic Variation of Brownian Motion . . . . . . . . . . . . . . . . . . 1
1.2. Quadratic Variation and Itos Formula . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1. Supplement on the Quadratic Variation . . . . . . . . . . . . . . . . . . . 5
1.3. d-Dimensional Ito-Formula and Covariation . . . . . . . . . . . . . . . . . . . . . 7
1.3.1. Important special cases . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4. Ito-Integrals as (Local) Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5. Levys Characterization of Brownian Motion . . . . . . . . . . . . . . . . . . . . 20
2. (Semi-)Martingales and Stochastic Integration 23
2.1. Review of Some Facts from Martingale Theory . . . . . . . . . . . . . . . . . . . 23
2.2. Quadratic Variation and Covariation for Continuous Local Martingales . . . . . . 25
2.3. Construction of stochastic integrals . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.4. Characterization of H
M in /
2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.4.1. Orthogonality in /
2
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.5. Itos Representation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3. Markov Processes 53
3.1. Markov Processes and Semigroups . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2. The Strong Markov Property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.3. Application to Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.4. Sojourn Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4. Girsanov Transformation 65
4.1. Problem Outline (d = 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.2. The General Girsanov Transformation . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3. Girsanov Transform with Brownian Motion . . . . . . . . . . . . . . . . . . . . . 69
4.4. Novikov condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.5. Integration by Parts on Wiener Space . . . . . . . . . . . . . . . . . . . . . . . . 81
5. Brownian motion and potential theory 89
5.1. Brownian motion and Laplace Operator . . . . . . . . . . . . . . . . . . . . . . . 89
5.2. Stochastic solution of the Dirichlet- resp. Poisson-Problem . . . . . . . . . . . . . 90
5.3. Application: Reccurence/transience of the Brownian motion . . . . . . . . . . . . 92
5.4. Regularity of boundary points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.5. Poissin equality and Green function . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.6. Self intersection of the paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.7. Feyman-Kac-Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
5.8. The heat equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6. Stochastic Dierential Equations 109
6.1. Solution denitons, examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
i
Inhaltsverzeichnis
6.2. Construction of solutions via transformation of the state space . . . . . . . . . . 111
6.3. Method of Lamperti-Doss-Sussmann . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.4. Construction of strong solutions on Lipschitz conditions (Picard-Lindelof method) 115
7. The martingale problem 121
7.1. Formulation, motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
A. Time Transformation 129
ii
1. Introduction to Pathwise Ito-Calculus
1.1. Preparation
Denition 1.1.1. X : [0, [R has bounded variation, if for all t 0
var
t
X := sup
t
i
[X (t
i+1
t) X (t
i
t)[ < , (1.1.1)
where is a partition 0 = t
0
< t
1
< . . . < t
N
< .
Notation: Since in Stochastics a process X also depends on , we write X
t
and X
t
() instead
of X(t) and X(t)() respectively.
As a reference see [dB03].
Denition 1.1.2. Suppose X to be right-continuous and of bounded variation.
Let f ((R) and (
n
)
nN
be a sequence of partitions, whose mesh
[
n
[ := sup
1iN
n
_
t
(n)
i+1
t
(n)
i
_
converges to zero as n and t
N
n
n
.
Then, since X is of bounded variation, there exists the Lebesgue-Stieltjes-Integral dened by
_
t
0
f
s
dX
s
:= lim
n
t
(n)
i

n
f
t
(n)
i
_
X
t
(n)
i+1
t
X
t
(n)
i
t
_
. (1.1.2)
Remark 1.1.3. Note that for continuous X this denition is independent of the choice of
(
n
)
nN
(Exercise!).
1.1.1. Quadratic Variation of Brownian Motion
We know (cf. [Roc11]) that a typical path of Brownian motion X on R
1
is of unbounded variation,
since for its quadratic variation we have X
t
= t (see 1.1.4(i) below). Nonetheless, one wants
to dene
_
t
0
f
s
dX
s
for a typical path of the Brownian motion X. More generally, we want to do this for every
continuous path with continuous quadratic variation t X
t
.
(X
t
)
t0
is a (continuous) R-valued Brownian motion on (, T, P) if
(i) The increments X
t
X
s
are independent and N(0, t s) distributed (t > s).
(ii) t X
t
() is continuous for all .
Theorem 1.1.4. Let (X
t
)
t0
be a (continuous) Brownian motion and (
n
)
nN
a sequence of
subdivisions with [
n
[
n
0,
n

n+1
, t
(n)
N
n
n
. Then, for all t 0
1
(i)
t
(n)
i

n
t
(n)
i
t
_
X
t
(n)
i+1
X
t
(n)
i
_
2
n
t P-a.s..
Here the zero set depends on the sequence (
n
)
nN
( and so far it could depend on t.)
(ii) Moreover,
P
_

t
(n)
i

n
t
(n)
i
t
_
X
t
(n)
i+1
X
t
(n)
i
_
2
n
t, t 0
_
= 1.
Proof. (i) See [Roc11, Satz 9.4.5].
(ii) Exercise (sandwich argument, notice that zero set in (i) a-priori depends on t).
1.2. Quadratic Variation and Itos Formula
Fix a continuous and real-valued function t X
t
on [0, [ (in short (X
t
)
t0
) with an existing
sequence of subdivisions (
n
) with [
n
[
n
0 and t
(n)
N
n
n
such that the quadratic variation
(along (
n
))
X
t
:= lim
n
t
(n)
i

n
t
(n)
i
t
_
X
t
(n)
i+1
X
t
(n)
i
_
2
= lim
n
t
(n)
i

n
(X
t
(n)
i+1
t
X
t
(n)
i
t
)
2
, t 0, (1.2.3)
exists for all t 0 and such that t X
t
is continuous on [0, ). Note that X
t
in (1.2.3)
could depend upon the choice of (
n
)
n1
in contrast to the limit in (1.1.2) (cf. [RY99, (2.3)-(2.5)
p. 27/28]). By denition it is obvious that t X
t
is increasing.
Remark 1.2.5. (i) If (X
t
)
t0
is of bounded variation, then
t
(n)
i

n
_
X
t
(n)
i+1
t
X
t
(n)
i
t
_
2
max
t
(n)
i

n
X
t
(n)
i+1
t
X
t
(n)
i
t
. .
< uniformly for large n
t
(n)
i

n
X
t
(n)
i+1
t
X
t
(n)
i
t
. .
<
0,
hence, X 0.
Therefore, X , 0 implies that (X
t
)
t0
is not of bounded variation and the Lebesgue-
Stieltjes-Integral can not be dened in the usual way.
(ii) If t X
t
is increasing, continuous and X
0
= 0, hence, t X
t
is a distribution
function of a measure (i.e. d = dX
t
) on ([0, ), B([0, ))), then (1.2.3) is equivalent
to: The distribution function F
n
of
n
:=
t
(n)
i

n
_
X
t
(n)
i+1
X
t
(n)
i
_
2
(n)
t
i
2
converges pointwise to
F(t) := (] , t]) =
_
1
],t]
d.
But, since for continuous X
t
(t) = lim
n
(] , t])
_
] , t
1
n
_
= lim
n
_
X
t
X
t
1
n
_
= 0,
we have that the pointwise convergence of F
n
to F is equivalent to
n
weakly by
Portemanteau.
(iii) Note that if X is an increasing function, then X is always of bounded variation:
var
t
X = sup
t
i
[X
t
i+1
t
X
t
i
t
[ = sup
t
i
(X
t
i+1
t
X
t
i
t
) = X
t
X
0
< .
Therefore, we can dene the integral with respect to X as a Lebesgue-Stieltjes-Integral.
Lemma 1.2.6 (Calculating integrals with respect to dX
s
). Let g (([0, )). Then
t
(n)
i

n
t
(n)
i
t
g(t
i
)
_
X
t
(n)
i+1
X
t
(n)
i
_
2
_
t
0
g(s) dX
s
.
Proof. The left hand side is equal to
_
g1
[0,t]
d
n
, whereas the right hand side equals
_
g1
[0,t]
d.
But the integrand g1
[0,t]
is -a.e. continuous and bounded. Hence, convergence follows by the
Portemanteau theorem and Remark 1.2.5 (ii).
Theorem 1.2.7 (Pathwise Ito-formula): Let F (
2
(R). Then for all t 0 the Ito-Formula is
given by
F(X
t
) F(X
0
) =
_
t
0
F
t
(X
s
) dX
s
+
1
2
_
t
0
F
tt
(X
s
) dX
s
, (1.2.4)
where
_
t
0
F
t
(X
s
) dX
s
:= lim
n
t
(n)
i

n
t
(n)
i
t
F
t
_
X
t
(n)
i
_
_
X
t
(n)
i+1
X
t
(n)
i
_
.
_
t
0
F
t
(X
s
) dX
s
is called the (pathwise) Ito-Integral and depends on (
n
)
nN
.
Proof. Consider (
n
)
nN
such that X
t
exists along (
n
)
nN
. We apply the Taylor formula to
F. Hence, for all n N there exist
(n)
i
]0, 1[ such that
t
(n)
i

n
t
(n)
i
t
F
_
X
t
(n)
i+1
_
F
_
X
t
(n)
i
__
n
F(X
t
) F(X
0
)
_
=
t
(n)
i

n
t
(n)
i
t
F
t
_
X
t
(n)
i
_
_
X
t
(n)
i+1
X
t
(n)
i
_
+
t
(n)
i

n
t
(n)
i
t
1
2
F
tt
_
X
t
(n)
i
_
_
X
t
(n)
i+1
X
t
(n)
i
_
2
. .
1.2.6
1
2
t
0
F
//
(X
s
) dX)
s
+
t
(n)
i

n
t
(n)
i
t
1
2
_
F
tt
_
X
t
(n)
i
+
(n)
i
_
X
t
(n)
i+1
X
t
(n)
i
__
F
tt
_
X
t
(n)
i
_
_
_
X
t
(n)
i+1
X
t
(n)
i
_
2
. .
=:S
3
Since F
tt
is locally uniformly continuous and X is uniformly continuous,
F
tt
_
X
t
(n)
i
+
(n)
i
_
X
t
(n)
i+1
X
t
(n)
i
__
F
tt
_
X
t
(n)
i
_
<
holds uniformly in i for n big enough. Therefore,
S
t
(n)
i

n
t
(n)
i
t
_
X
t
(n)
i+1
X
t
(n)
i
_
2
. .
n
X)
t
0
0,
which nishes the proof.
Remark 1.2.8. If X
t
0 (e.g. (X
t
)
t0
is of bounded variation), then we are in the classical
case:
F(X
t
) F(X
0
) =
_
t
0
F
t
(X
s
) dX
s
is an ordinary Lebesgue-Stieltjes-Integral. (If X
t
= t, this is the Fundamental Theorem of
Calculus). We introduce a short-hand notation for (1.2.4)
dF(X) = F
t
(X) dX +
1
2
F
tt
(X) dX, (1.2.5)
in contrast to the classical case (i.e. X 0), where
dF(X) = F
t
(X) dX. (1.2.6)
Example 1.2.9. Consider the dierential equation
dX
n
= nX
n1
dX for n N xed.
If X = 0, then a solution is X
n
.
If X , = 0, this is not a solution, since by Ito
dX
n
= nX
n1
dX +
1
2
n(n 1)X
n2
dX.
We would like to nd a function h
n
: R R such that
dh
n
(X) = nh
n1
(X) dX
in the general case X , 0. Later (cf. Example 1.3.25(ii) below), we shall see that the n-th
Hermite polynomial h
n
will provide a solution to this problem.
Remark 1.2.10. We know that, if f (
1
(R), then the Ito-integral
_
t
0
f(X
s
) dX
s
, t 0, is
(well-)dened. (Simply take F as a primitive of f, i.e. F
t
= f, and apply Itos formula.)
Denition 1.2.11 (-Integral). More generally we dene for [0, 1] and f (
1
(R)
-
_
t
0
f(X
s
) dX
s
:= lim
n
t
(n)
i

n
t
(n)
i
t
f
_
X
t
(n)
i
+
_
X
t
(n)
i+1
X
t
(n)
i
__
_
X
t
(n)
i+1
X
t
(n)
i
_
. (1.2.7)
4
Claim: This limit exists and
-
_
t
0
f(X
s
) dX
s
=
_
t
0
f(X
s
) dX
s
+
_
t
0
f
t
(X
s
) dX
s
. (1.2.8)
Proof. Exercise (Compare -sum with 0-sum (Ito-Integral) and use the mean-value-theorem
for f).
Special cases:
= 0: Ito-integral
= 1: Backward Ito-integral.
=
1
2
: Stratonovich-Fisk-Integral
Notation:
_
. . . dX
s
:=
1
2
-
_
. . . dX
s
. Hence
_
t
0
f(X
s
) dX
s
_
=
1
2

_
t
0
f(X
s
) dX
s
_
=
_
t
0
f(X
s
) dX
s
+
1
2
_
t
0
f
t
(X
s
) dX
s
and we have by Ito the Stratonovich-formula
F(X
t
) F(X
0
) =
_
t
0
F
t
(X
s
) dX
s
. (1.2.9)
Remark 1.2.12. (i) An advantage of the Ito-integral is (see Section 1.4 below) that, if X is
a martingale, then, again,
_
f(X
s
) dX
s
is a martingale.
(ii) In the Stratonovich-fromula one only has to deal with derivatives of rst order and, there-
fore, it can be used for manifold-valued X.
1.2.1. Supplement on the Quadratic Variation
Lemma 1.2.13. (i) Let F (
1
(R). Then t F(X
t
) has (nite) quadratic variation (along
xed (
n
)
nN
)
F(X)
t
=
_
t
0
(F
t
(X
s
))
2
dX
s
(automatically continuous in t).
(ii) If M
t
:= X
t
+ A
t
, t 0, for some t A
t
continuous and A 0 (again A calculated
along (
n
)), then
M
t
= X
t
.
(iii) The Ito-integral t
_
t
0
f(X
s
) dX
s
=: M
t
with f (
1
(R), has quadratic variation (along
(
n
)) and
M
t
=
__
t
0
f(X
s
) dX
s
_
t
=
_
t
0
f(X
s
)
2
dX
s
.
5
Proof. (i) We rst apply Taylor up to order 1, then take the square on both sides and nally
apply the Binomial formula to get
t
(n)
i

n
t
(n)
i
t
_
F(X
t
(n)
i+1
) F(X
t
(n)
i
)
_
2
=
t
(n)
i

n
t
(n)
i
t
F
t
(X
t
(n)
i
)
2
_
X
t
(n)
i+1
X
t
(n)
i
_
2
. .
n
t
0
(F
/
(X))
2
dX)
s
by Lemma 1.2.6
+
t
(n)
i

n
t
(n)
i
t
_
F
t
_
X
t
(n)
i
+
(n)
i
_
X
t
(n)
i+1
X
t
(n)
i
__
F
t
_
X
t
(n)
i
_
_
2
. .
< for large n, since F
is uniformly continuous
on the compact set {X
s
s [0, t + 1]}
_
X
t
(n)
i+1
X
t
(n)
i
_
2
+ 2
t
(n)
i

n
t
(n)
i
t
F
t
(X
t
(n)
i
)
_
X
t
(n)
i+1
X
t
(n)
i
_
_
F
t
_
X
t
(n)
i
+
(n)
i
_
X
t
(n)
i+1
X
t
(n)
i
__
F
t
(X
t
(n)
i
)
__
X
t
(n)
i+1
X
t
(n)
i
_
.
Since the second term goes to zero as n , so, by Cauchy-Schwartz, does the third.
(ii)
t
(n)
i

n
t
(n)
i
t
_
M
(n)
t
i+1
M
(n)
t
i
_
2
=
t
(n)
i

n
t
(n)
i
t
_
X
t
(n)
i+1
X
t
(n)
i
_
2
. .
n
X)
t
+
t
(n)
i

n
t
(n)
i
t
_
A
(n)
t
i+1
A
(n)
t
i
_
2
. .
n
0 by assumption
+ 2
t
(n)
i

n
t
(n)
i
t
_
X
t
(n)
i+1
X
t
(n)
i
_
_
A
(n)
t
i+1
A
(n)
t
i
_
. .
n
0 by Cauchy-Schwartz
.
(iii) Let F (
2
(R) such that F
t
= f and apply Ito to get
M
t
= F(X
t
)
_
F(X
0
) +
1
2
_
t
0
F
tt
(X
s
) dX
s
_
. .
=:A
t
.
But A
t
can be written as a dierence of increasing functions. Therefore, A
t
is of bounded
variation, hence, A = 0. Thus, by (ii) and (i)
M
(ii)
= F(X)
t
(i)
=
_
t
0
(F
t
(X
s
))
2
dX
s
.
6
1.3. d-Dimensional Ito-Formula and Covariation
Fix X, Y : [0, ) R continuous with bounded quadratic variation X, Y (along the same
(
n
)
nN
) (cf. [RY99, (2.3)-(2.5) (p.27/28)]).
Denition 1.3.14. If
X, Y
t
:= lim
n
t
(n)
i

n
t
(n)
i
t
_
X
t
(n)
i+1
X
t
(n)
i
__
Y
t
(n)
i+1
Y
t
(n)
i
_
, t 0,
exists, then it is called the covariation of X and Y (along (
n
)).
Lemma 1.3.15. The following assertions are equivalent:
(i) X, Y exists and is continuous.
(ii) X +Y exists and is continuous. In this case the Polarization identity holds:
X, Y =
1
2
(X +Y X Y ) .
In particular, X, Y is the distribution function of a signed measure on R
+
dX, Y =
1
2
dX +Y
1
2
dX
1
2
dY .
Furthermore, if X, Y exists, we have
[X, Y [ X
1
2
Y
1
2
,
i.e. a Cauchy-Schwartz inequality.
Proof. Exercise.
Remark 1.3.16. [X, Y [ X
1
2
Y
1
2
is a special case of the Kunita-Watanabe-inequality (cf.
2.2.16 below).
Example 1.3.17. (i) Let (X
t
)
t0
, (Y
t
)
t0
be independent Brownian motions on (, T, P).
Then there exists X, Y () for P-a.e. and
X, Y () = 0 P-a.e. .
Proof. We know that
Z
t
:=
1
2
(X
t
+Y
t
), t 0,
is a Brownian motion. Hence, (cf. Propostition 1.1.4(i)) Z
t
= t, so there exists X+Y
t
=
2Z
t
= 2t. Then it follows by Lemma 1.3.15 applied to P-a.e. that
X, Y =
1
2
(X +Y X Y ) = 0.
7
(ii) Let f, g ((R) and
Y
t
:=
_
t
0
f(X
s
) dX
s
, Z
t
:=
_
t
0
g(X
s
) dX
s
.
Then (again with respect to our (
n
)) there exists
Y, Z
t
=
_
t
0
f(X
s
)g(X
s
) dX
s
.
Proof. By 1.2.13(iii) the quadratic variation of
Y
t
+Z
t
=
_
t
0
(f +g)(X
s
) dX
s
along our (
n
) exists. Hence, by the polarization identity and Lemma 1.2.13(iii) we get
2Y, Z = Y +Z Y Z
= 2
_
f(X
s
)g(X
s
) dX
s
+
_
f(X
s
)
2
dX
s
+
_
g(X
s
)
2
dX
s
Y Z
= 2
_
f(X
s
)g(X
s
) dX
s
.
Proposition 1.3.18 (Itos product rule). Let X, Y be as above such that there exists X, Y
(with respect to (
n
)) and is continuous in t 0. If there exists either
lim
n
t
(n)
i

n
t
(n)
i
t
X
t
(n)
i
_
Y
t
(n)
i+1
Y
t
(n)
i
_
=:
_
t
0
X
s
dY
s
or
lim
n
t
(n)
i

n
t
(n)
i
t
Y
t
(n)
i
_
X
t
(n)
i+1
X
t
(n)
i
_
=:
_
t
0
Y
s
dX
s
,
then both of these limits exist and
X
t
Y
t
= X
0
Y
0
+
_
t
0
X
s
dY
s
+
_
t
0
Y
s
dX
s
+X, Y
t
.
Proof. We have
X
t
Y
t
=
1
2
_
(X
t
+Y
t
)
2
X
2
t
Y
2
t
_
.
Furthermore, by Lemma 1.3.15 X+Y exists and is continuous, since by Ito with F(X) =
1
2
X
2
we know that
X
t
Y
t
=
1
2
_
(X
0
+Y
0
)
2
X
2
0
Y
2
0
_
. .
=X
0
Y
0
+
_
t
0
(X +Y )
s
d(X +Y )
s
_
t
0
X
s
dX
s
_
t
0
Y
s
dY
s
+
1
2
(X +Y
t
X
t
Y
t
)
. .
=X,Y )
t
.
8
By denition we have
_
t
0
(X +Y )
s
d(X +Y )
s
= lim
n
t
(n)
i

n
t
(n)
i
t
_
X
t
(n)
i
+Y
t
(n)
i
_
_
X
t
(n)
i+1
X
t
(n)
i
+Y
t
(n)
i+1
Y
t
(n)
i
_
. .
=X
t
(n)
i
X
t
(n)
i+1
X
t
(n)
i
+Y
t
(n)
i
Y
t
(n)
i+1
Y
t
(n)
i
+X
t
(n)
i
Y
t
(n)
i+1
Y
t
(n)
i
+Y
t
(n)
i
X
t
(n)
i+1
X
t
(n)
i
Therefore, we get
X
t
Y
t
= X
0
Y
0
+ lim
n
t
(n)
i

n
t
(n)
i
t
_
X
t
(n)
i
_
Y
t
(n)
i+1
Y
t
(n)
i
_
+Y
t
(n)
i
_
X
t
(n)
i+1
Y
t
(n)
i
__
+X, Y
t
,
which implies the assertion.
Remark 1.3.19. If X or Y has bounded variation, e.g. X, then there already exists
_
Y
s
dX
s
and
all assumptions in Proposition 1.3.18 are fullled. In this case X, Y = 0 and, by substituting
dY
s
= Y
t
s
ds, we are in the classical case of integration by parts:
X
t
Y
t
X
0
Y
0
=
_
X
s
Y
t
s
ds +
_
Y
s
X
t
s
ds.
Example 1.3.20. Suppose t Y
t
is of bounded variation, hence, Y 0 and X, Y = 0 (by
Holder). Then by Proposition 1.3.18
_
t
0
Y
s
dX
s
=
_
t
0
X
s
dY
s
+X
t
Y
t
X
0
Y
0
.
Here, we can dene the left hand side by the right hand side since
_
t
0
X
s
dY
s
is a usual Lebesgue-
Stieltjes ingegral. This approach was used by Paley-Wiener to dene stochastic integrals, if X is
a Brownian motion:
Let X() be a typical Brownian path (hence, X
0
() = 0) and h(s)(= Y
s
) continuous, of bounded
variation and independent of with h(1) = 0. Dene
_
1
0
h(s) dX
s
() :=
_
1
0
X
s
() dh(s).
One can show that
E
_
__
1
0
h(s) dX
s
_
2
_
=
_
1
0
h(s)
2
ds,
hence,
L
2
([0, 1], ds) L
2
(, T, P)

h
_
h(s) dX
s
()
is an isometry. It is rst dened for a dense subset of functions h in L
2
([0, 1], ds), and then
extended to the closure, i.e. for all h L
2
([0, 1], ds), by this isometry.
9
Fix now X
t
=
_
X
1
t
, . . . , X
d
t
_
: [0, ) R
d
continuous with continuous X
i
, X
i
, X
j
, i, j
1, . . . , d, i ,= j (along our (
n
)).
Proposition 1.3.21 (d-dimensional Ito-formula). Let F (
2
(R
d
). Then with (, )
t
as Euklidean
inner product on R
F(X
t
) F(X
0
) =
_
t
0
(F(X
s
), dX
s
) +
1
2
_
t
0
d
k,l=1
2
F
x
k
x
l
(X
s
) dX
k
, X
l
s
, (1.3.10)
where
_
t
0
(F(X
s
), dX
s
) := lim
n
t
(n)
i

n
t
(n)
i
t
_
F(X
t
(n)
i
), X
t
(n)
i+1
X
t
(n)
i
_
is called multidimensional Ito-integral.
Proof. Obviously, we have
t
(n)
i

n
t
(n)
i
t
F
_
X
t
(n)
i+1
_
F
_
X
t
(n)
i
_
n
F(X
t
) F(X
0
).
Furthermore, by d-dimensional Taylor forumla we obtain
t
(n)
i

n
t
(n)
i
t
F
_
X
t
(n)
i+1
_
F
_
X
t
(n)
i
_
=
t
(n)
i

n
t
(n)
i
t
_
F
_
X
t
(n)
i
_
, X
t
(n)
i+1
X
t
(n)
i
_
+
1
2
t
(n)
i

n
t
(n)
i
t
_
A
_
X
t
(n)
i
_
_
X
t
(n)
i+1
X
t
(n)
i
_
,
_
X
t
(n)
i+1
X
t
(n)
i
__
. .
=:S
+
1
2
t
(n)
i

n
t
(n)
i
t
_
_
A
_
X
t
(n)
i
+
n
i
_
X
t
(n)
i+1
X
t
(n)
i
__
A
_
X
t
(n)
i
_
_
_
X
t
(n)
i+1
X
t
(n)
i
_
,
_
X
t
(n)
i+1
X
t
(n)
i
_
_
,
where
A(x) =
_

2
x
i
x
j
F(x)
_
i,j
.
The third summand vanishes analogously to the 1-dimensional case. Moreover, since we can
10
interchange the sums, we have by polarization
S =
d
j,k=1
t
(n)
i

n
t
(n)
i
t
2
F
x
j
x
k
_
X
t
(n)
i
_
_
X
(j)
t
(n)
i+1
X
(j)
t
(n)
i
__
X
(k)
t
(n)
i+1
X
(k)
t
(n)
i
_
=
d
j,k=1
t
(n)
i

n
t
(n)
i
t
2
F
x
j
x
k
_
X
t
(n)
i
_
1
2
_
__
X
(j)
t
(n)
i+1
+X
(k)
t
(n)
i+1
_
_
X
(j)
t
(n)
i
+X
(k)
t
(n)
i
__
2
_
X
(j)
t
(n)
i+1
X
(j)
t
(n)
i
_
2
_
X
(k)
t
(n)
i+1
X
(k)
t
(n)
i
_
2
_
n
1
2
d
j,k=1
__
t
0
2
F
x
j
x
k
(X
s
) dX
(j)
+X
(k)
_
t
0
2
F
x
j
x
k
(X
s
) dX
(j)
_
t
0
2
F
x
j
x
k
(X
s
) dX
(k)
s
_
=
_
t
0
d
k,l=1
2
F
x
k
x
l
(X
s
) dX
k
, X
l
s
, P-a.s..
Remark 1.3.22. (i) We have dened
_
t
0
(F(X
s
), dX
s
) := lim
n
d
k=1
t
(n)
i

n
t
(n)
i
t
F
x
k
_
X
t
(n)
i
_
_
X
k
t
(n)
i+1
X
k
t
(n)
i
_
,
but we cannot interchange
d
k=1
with the limit, since we do not know whether for every
k the limit exists.
(ii) For F (
1
(R
d
) the function t F(X
t
) has continuous quadratic variation and
F(X)
t
=
d
k,l=1
_
t
0
F
x
k
(X
s
)
F
x
l
(X
s
) dX
k
, X
l
s
.
11
Proof. By Taylor
t
(n)
i

n
t
(n)
i
t
_
F
_
X
t
(n)
i+1
_
F
_
X
t
(n)
i
_
_
2
=
d
k,l=1
t
(n)
i

n
t
(n)
i
t
F
x
k
_
X
t
(n)
i
_
F
x
l
_
X
t
(n)
i
_
_
X
k
t
(n)
i+1
X
k
t
(n)
i
__
X
l
t
(n)
i+1
X
l
t
(n)
i
_
. .
1.2.6
d
k,l=1
t
0
F
x
k
(X
s
)
F
x
l
(X
s
) dX
k
,X
l
)
s
by polarization
+
d
k,l=1
t
(n)
i

n
t
(n)
i
t
_
F
x
k
_
X
t
(n)
i
+
n
i
_
X
t
(n)
i+1
X
t
(n)
i
__
F
x
k
_
X
t
(n)
i
_
_
. .
< uniformly in i for n big enough
_
F
x
l
_
X
t
(n)
i
+
n
i
_
X
t
(n)
i+1
X
t
(n)
i
__
F
x
l
(X
t
(n)
i
)
_
. .
<
_
X
k
t
(n)
i+1
X
k
t
(n)
i
__
X
l
t
(n)
i+1
X
l
t
(n)
i
_
+ 2
d
k,l=1
t
(n)
i

n
t
(n)
i
t
_
F
x
k
_
X
t
(n)
i
+
n
i
_
X
t
(n)
i+1
X
t
(n)
i
__
F
x
k
_
X
t
(n)
i
_
__
X
k
t
(n)
i+1
X
k
t
(n)
i
_
t
(n)
i

n
t
(n)
i
t
F
x
l
_
X
t
(n)
i
_
_
X
l
t
(n)
i+1
X
l
t
(n)
i
_
.
By Cauchy Schwartz the last two double sums converge to 0 as n goes to .
(iii) Since
1
2
_
t
0
d
k,l=1
2
F
x
k
x
l
(X
s
) dX
k
, X
l
s
is of bounded variation, its quadratic variation is 0. But then, by 5.2.7 and 1.2.13 (ii), we
can conclude that
__

0
(F(X
s
), dX
s
)
_
t
= F(X)F(X
0
)
t
= F(X)
t
(ii)
=
d
k,l=1
_
t
0
F
x
k
(X
s
)
F
x
l
dX
k
, X
l
s
.
1.3.1. Important special cases
Brownian motion in R
d
and Laplace-operator
The components of a d-dimensional Brownian motion are independent, hence, (by 1.3.17(i))
X
k
, X
l
t
() =
kl
t (covariation reects independence),
which implies by Ito for F (
2
(R)
F(X
t
) F(X
0
) =
_
t
0
(F(X
s
), dX
s
) +
1
2
_
t
0
F(X
s
) ds.
12
In particular, if F is harmonic (i.e. F = 0), then
F(X
t
) = F(0) +
_
t
0
(F(X
s
), dX
s
),
which is an Ito-integral of a Brownian motion. Hence, a harmonic function preserves the mar-
tingale property since the Ito-integral again is a martingale (see Section 1.4 below).
Ito-formula for time dependent functions
Proposition 1.3.23. Let F (
2
(R
2
) and X : [0, ) R be continuous with continuous X
(along (
n
)
nN
). Then
F(X
t
, X
t
) = F(X
0
, 0) +
_
t
0
F
x
(X
s
, X
s
) dX
s
+
_
t
0
_
1
2
2
F
x
2
+
F
y
_
(X
s
, X
s
) dX
s
.
Proof. Apply Ito for d = 2 with (X
t
, X
t
) to get
F(X
t
, X
t
) F(X
0
, 0)
=
_
t
0
F
x
(X
s
, X
s
) dX
s
+
_
t
0
F
y
(X
s
, X
s
) dX
s
. .
=:S
+
1
2
_
t
0
2
F
x
2
(X
s
, X
s
) dX
s
+
1
2
_
t
0
2
F
y
2
(X
s
, X
s
) dX
s
, X
s
+
1
2
_
t
0
2
F
xy
(X
s
, X
s
) dX
s
, X
s
. .
=0 since X)
s
is of bounded variation and by 1.3.20
.
The second summand of S exists since X
t
is of bounded variation (along (
n
)
nN
). As the
whole sum S exists by Ito (d = 2), so does the rst summand.
Remark 1.3.24. (i) If X
t
is a Brownian motion, then X
t
represents the time, i.e.
F(X
t
, X
t
) = F(X
t
, t).
Therefore, X
t
is also called inner clock of X
t
.
(ii) If F is a solution to the backward heat equation, i.e.
1
2
2
F
x
2
+
F
t
= 0,
then we have by above (for X as above)
F(X
t
, X
t
) = F(X
0
, 0) +
_
t
0
F
x
(X
s
, X
s
) dX
s
= F(X
0
, 0) + Ito-integral.
Later we shall see that the Ito-integral is a local martingale, if X is one. Therefore, every
solution to the backward heat equation provides a local martingale!
Example 1.3.25. (i) Let F(x, t) := exp
_
x
1
2
2
t
_
. Then F solves the backward heat equa-
tion. Thus, if X
0
= 0, then
G
t
:= F(X
t
, X
t
) = exp
_
X
t
1
2
2
X
t
_
13
solves the dierential equation
G
0
= 1,
dG = G dX,
i.e.
G
t
= 1 +
_
t
0
G
s
dX
s
.
Recall, if X 0, then G
t
= exp(X
t
). Hence, in the Ito-calculus G
t
as above is the right
analogon to the exponential function exp(t) in the usual case.
Application to Brownian motion:
For X with X
t
= t
G
t
() := G
0
e
t
exp
_
X
t
()
1
2
2
t
_
, t 0,
by Proposition 1.3.23 solves the linear SDE
dG = G dX +G dt.
Classical ( = 0):
dG = G dt G
t
= G
0
e
t
.
Here, for > 0 G
t
tends to as t .
Stochastic ( ,= 0):
G
t
= G
0
exp
_
X
t
+
_

1
2
2
_
t
_
.
By law of iterated logarithm
X
t
t
t
0, a.s.,
hence, for large t
X
t
t
<
1
2
_
1
2
_
.
Thus, for <
1
2
2
G
t
G
0
e
(
1
2
)
t
2
t
0, (a.s. pathwise stable).
But G
t
is not uniformly integrable, since by the Residue theorem
E[e
X
t
] = e
1
2
2
t
,
and therefore,
E[G
t
] = G
0
e
t
E
_
e
X
t
1
2
2
t
= G
0
e
t
, if > 0 (unstable in the mean).
(ii) Hermite- polynomials (cf. Example 1.2.9):
Dene h
n
(x, t) by
e
x
1
2
2
t
=
n=0
n
n!
h
n
(x, t), (1.3.11)
14
where the left hand side is analytic in , i.e.
h
n
(x, t) =

n
n
_
e
x
1
2
2
t
_
=0
_
= e
x
n
k=0
_
n
k
_
x
k

nk
nk
_
e
1
2
2
t
_
=0
_
. (1.3.12)
Recall, that for all n N
H
n
(, t) :=
1
n!t
n
h
n
(, t)
is an orthonormal basis of L
2
(R, N(0, t)).
Proof. Clearly, (by monotone classes) spanH
n
, n N is a dense subset of L
2
(R, N(0, t)).
Hence, it is sucient to show that it is an ONS:
n,m
n
n!
h
n
(x, t)
m
m!
h
m
(x, t)
(1.3.11)
= e
(+)x
1
2
(
2
+
2
)t
= exp(( +)x)e
1
2
(+)
2
t
e
t
.
By integration with N(0, t) in x, since sums interchange with integration and
_
e
(+)x
N(0, t)( dx) = e
1
2
(+)
2
t
,
we get by (1.3.12) that
n,m
m
_
1
n!
h
n
(x, t)
1
m!
h
m
(x, t)N(0, t)( dx) = e
t
=
n
t
n
n!
, , .
We have for n 0 by (1.3.12) and by interchanging order of dierentiation
1
2
2
x
2
h
n
+

t
h
n
= 0.
Additionally, for n 1 we have
x
h
n
(, t)
(1.3.11)
= n h
n1
(, t).
Therefore, if X, X are continuous and X
0
= 0, (since h
n
(0, 0) = 0), we get
h
n
(X
t
, X
t
) =
_
t
0
x
h
n
(X
t
1
, X
t
2
) dX
t
1
= n
_
t
0
h
n1
(X
t
1
, X
t
1
) dX
t
1
= . . . = n!
_
t
0
dX
t
1
_
t
1
0
dX
t
2
. . .
_
t
n1
0
dX
t
n
.
15
1.4. Ito-Integrals as (Local) Martingales
Recall: If (T
t
)
t0
is not right-continuous, then dene
T
t
+ :=
s>t
T
s
, t 0.
Then (T
t+
)
t0
is right-continuous.
Let (, T, P) be a probability space equipped with a right-continuous ltration (T
t
)
t0
, i.e.
T
t
=
s>t
T
s
.
Let T : [0, ] and let X : (t, ) ]0, [ : t < T() 0 R be a map
such that X is a continuous local martingale (up to T) i.e. there exists a localizing sequence
T
1
T
2
. . . T and < T on T > 0, i.e. they are (T
t
)-stopping times such that
(i) (X
tT
n
)
t0
is a continuous (T
t
)-martingale for all n N,
(ii) sup
nN
T
n
= T P-a.s..
Assume in addition, that X has a continuous quadratic variation [0, T()[ t X
t
() along
a sequence of partitions (
n
) as above for P-a.e. .
Remark 1.4.26. Later we shall see that for any (
n
) there exists (
n
k
) such that X as above
has continuous quadratic variation X
t
along (
n
k
) for P-a.e. .
Proposition 1.4.27. Let f (
2
(G R
+
), G R
1
open (or f (
1
(G)). Assume there exists
a compact set K G such that
X
0
() K for P-a.e. .
Dene M
0
() = 0 for all and
M
t
() :=
_
t
0
f(X
s
(), X
s
()) dX
s
(), 0 < t < T() for T > 0 . (Ito-integral!)
Then M is a continuous local martingale up to
S := inft > 0[X
t
, G
. .
=:
G
c
T.
Proof. Step 1: Assume that T = , X is a bounded martingale, G = R
1
and f is bounded.
Then M
t
() is dened for all t 0, since S = +.
Claim: (M
t
)
t0
is a continuous martingale.
To see this dene for n N xed
M
(n)
t
:=
t
(n)
i

n
t
(n)
i
t
f
_
X
t
(n)
i
, X
t
(n)
i
_
_
X
t
(n)
i+1
X
t
(n)
i
_
.
To see that (M
(n)
t
)
t0
is a martingale, (but not nessassarily continuous), let s < t and take
t
(n)
k
, t
(n)
l

n
such that t
(n)
k
t < t
(n)
k+1
, t
(n)
l
s < t
(n)
l+1
. Then, we have M
(n)
t
= M
(n)
t
k
and
16
M
(n)
s
= M
(n)
t
l
and therefore,
E[M
(n)
t
M
(n)
s
[T
s
] = E[M
(n)
t
k
M
(n)
t
l
[T
s
]
=
k
i=l+1
E
_
E
_
f
_
X
t
(n)
i
, X
t
(n)
i
_
. .
T
t
(n)
i
measurable
_
X
t
(n)
i+1
X
t
(n)
i
_
T
t
(n)
i
_
T
s
_
=
k
i=l+1
E
_
f
_
X
t
(n)
i
, X
t
(n)
i
_
E
_ _
X
t
(n)
i+1
X
t
(n)
i
_
T
t
(n)
i
_
. .
=0 since X
t
is a martingale
T
s
_
= 0
Hence by Proposition 2.1.4 below it is sucient to show that M
n
t
M
t
in L
1
for all t 0. We
know that by denition,
M
(n)
t
M
t
, P-a.s. t 0.
Furthermore, for all t 0,
E
_
_
M
(n)
t
_
2
_
=
t
(n)
i

n
t
(n)
i
t
E
_
f
2
_
X
t
(n)
i
, X
t
(n)
i
_
_
X
t
(n)
i+1
X
t
(n)
i
_
2
_
,
since all nondiagonal terms vanish by the martingale property of X
t
. Furthermore,
E
_
_
M
(n)
t
_
2
_
sup
RR
+
f
2
t
(n)
i

n
t
(n)
i
t
E
_
E
_
X
2
t
(n)
i+1
2X
t
(n)
i+1
X
t
(n)
i
+X
2
t
(n)
i
T
t
(n)
i
_
. .
=E
_
X
2
t
(n)
i+1
T
t
(n)
i
X
2
t
(n)
i
_
= sup
RR
+
f
2
t
(n)
i

n
t
(n)
i
t
E
_
X
2
t
(n)
i+1
X
2
t
(n)
i
_
= sup
RR
+
f
2
E
_
X
2
t
n
N
n
X
2
0
_
. .
sup
t,
X
2
t
()
.
Hence, sup
n
E
_
_
M
(n)
t
_
2
_
< and thus
_
M
(n)
t
_
nN
is uniformly integrable. By the generalized
Lebesgue dominated convergence theorem the claim follows. Therefore (M
t
)
t0
is a martingale.
It still remains to show that M
t
has P-a.s. continuous sample paths:
In order to see that consider
M
(n)
t
:=
n
i=0
f
_
X
t
(n)
i
, X
t
(n)
i
_
_
X
t
(n)
i+1
t
X
t
(n)
i
t
_
.
Then as above one shows that

M
t
() is a martingale. Note that
_

M
(n)
t
_
t0
has P-a.s. continuous
sample paths and that
M
(n)
t

M
(n)
t
n
0 P-a.s.
and in L
q
(by the same argument as above). Hence,

M
(n)
t
n
M
t
in L
q
for all q [1, 2). Then
by Doobs maximal inequality one can show that (M
t
)
t0
has P-a.s. continuous sample paths
17
(cf. below Proposition 2.1.4).
Step 2 (Localization by stopping times):
Let (T
n
) be a localizing sequence for X. For n N dene
T
n
:= inft > 0[X
t
> n T
and
S
n
:= T
n

G
c
n

T
n
,
where G
n
G, G
n
relatively compact and open and

G
n
G
n+1
for all n N. Without loss
of generality K G
n
n N. Then sup
n
G
c
n
=
G
c and S
n
are stopping times such that
S
n
T
n
T and S
n
T
n
< T on T > 0. Furthermore,

T
n
T, hence, S
n
T
G
c. By
optional stopping (X
tS
n
)
t0
is a continuous martingale taking values in G
n
, hence (X
tS
n
)
t0
is bounded in (t, ). Furthermore, (exercise)
M
tS
n
()
()
!
=
_
t
0
f
_
X
sS
n
()
(), X
sS
n
()
_
dX
sS
n
()
(). (1.4.13)
Take
n
(
2
0
(GR
+
) such that
n
= 1 on G
n
[0, n].
Then we can replace f in (1.4.13) by
n
f (
2
b
(G R
+
). Therefore, the representation for
(M
tS
n
)
t0
in (1.4.13) and Step 1 imply that (M
tS
n
)
t0
is a continuous martingale.
Corollary 1.4.28. Let X be a continuous local martingale up to T with continuous quadratic
variation (later proved to always be the case). Then
(i) X
2
X is a continuous local martingale up to T.
(ii) If X = 0 (which is particularly true if X has bounded variation), then for P-a.s.
X
t
() X
0
() t [0, T()[. (1.4.14)
Proof. Without loss of generality assume that X
0
0. (Otherwise, consider X
t
X
0
, t T.)
(i) By Ito
X
2
t
= 2
_
t
0
X
s
dX
s
+X
t
on t < T.
But the rst term on the right hand side is a continuous local martingale up to T by
Proposition 5.3.9.
(ii) By (i) it also follows that X
2
is a continuous local martingale up to T if X = 0. Hence,
if T
n
T is a localising sequence for X
2
, then
E
_
X
2
tT
n
= E
_
X
2
0
= 0 t 0.
Therefore, 1
t<T
X
t
= 0 P-a.s. t 0, (with zero set depending on t). Hence,
P
_
1
t<T
X
t
= 0 t Q
= 1,
and by P-a.s. continuity in t it follows that X
t
= 0 on t < T for all t 0 P-a.s..
18
Proposition 1.4.29 (d-dimensional version of Proposition 5.3.9). Let X = (X
1
, . . . , X
d
) with
X
1
, . . . , X
d
continuous local martingales up to T such that X
i
, X
j
exist for all 1 i, j d and
are continuous up to T. Let F (
2
(D), D R
d
, D open, with X
0
() K for P-a.e. for
some compact K D. Dene M
0
:= 0 and
M
t
:=
_
t
0
(F(X
s
), dX
s
)
R
d on t < T (d-dimensional It o-integral)
Then M is a continuous local martingale up to
S := T inft > 0[X
t
, D
. .
=:
D
c
.
Proof. Exercise (Proceed as for Proposition 5.3.9).
Remark 1.4.30. By Remark 1.3.22 (iii) for M as in Proposition 1.4.29 we have
M
t
=
d
k,l=1
_
t
0
F
x
k
(X
s
)
F
x
l
(X
s
) dX
k
, X
l
s
on t < S.
Hence, if S , then M is even a martingale provided
F
x
k
, 1 k d, are all bounded and X
is a Brownian motion. This is a consequence of Corollary 1.4.32 below.
Proposition 1.4.31. Let M be a continuous local martingale up to T (with continuous M up
to T) and let T
0
T with T
0
< T on T > 0 be a stopping time such that
E[M
T
0
] < .
Then (M
tT
0
)
t0
is a continuous martingale such that sup
t0
E[M
2
T
0
t
] E[M
T
0
]. Furthermore,
E[M
T
0
] = E[M
0
].
Note that T
0
doesnt need to be bounded.
Proof. Without loss of generality assume M
0
0. (Otherwise, consider M
t
M
0
.) Let T
n
T
be a localizing sequence for M. Then (M
T
0
T
n
t
)
t0
is a continuous martingale. Hence
lim
n
M
T
0
T
n
t
= M
T
0
t
P-a.s. t 0
and
E
_
M
2
T
0
T
n
t
Cor. 5.3.10
= E[M
T
0
T
n
t
] E[M
T
0
] < .
Hence, M
T
0
T
n
t
[n N is uniformly integrable t 0. In particular, by Lebesgue
lim
n
M
T
0
T
n
t
= M
T
0
t
in L
1
t 0.
Therefore, (M
T
0
t
)
t0
is a continuous martingale. In addition,
E
_
M
2
T
0
t
liminf
n
E
_
M
2
T
0
T
n
t
E [M
T
0
] < t 0.
So, by the L
2
-martingale convergence theorem
lim
t
M
T
0
t
= M
T
0
in L
2
,
In particular,
E [M
T
0
] = lim
t
E [M
T
0
t
] = E [M
0
] = 0.
19
Corollary 1.4.32. Let M be a continuous local martingale up to T (with continuous M)
such that
E[M
t
] < t 0.
Then M is a square integrable martingale.
Proposition 1.4.33. Let T be a stopping time and X, A continuous processes with continuous
X and A = 0 up to T. Assume A
0
= 0 and X
0
= 0. Then the following are equivalent:
(i) X is a local martingale up to T with X = A.
(ii) For all 0
G
t
:=
_
1 if t = 0
exp[X
t
1
2
2
A
t
] on t < T
is a local martingale up to T.
Proof. (i) (ii): By Itos formula for time dependent functions (cf. Proposition 1.3.24)
G
t
= G
0
+
_
t
0
G
s
dX
s
+ 0.
But the right hand side is a local martingale up to T by Proposition 5.3.9 for f(X
t
, X
t
) =
exp(X
t
1
2
2
X
t
).
(ii) (i): By Itos formula we get
X
t
=
1
_
log G
t
log G
0
. .
=0
_
+
1
2
A
t
Ito
=
1
__
t
0
1
G
s
dG
s

1
2
_
t
0
1
(G
s
)
2
dG
s
_
+
1
2
A
t
1.3.22(ii)
=
1
__
t
0
1
G
s
dG
s

1
2
_
t
0
1
(G
s
)
2
(G
s
)
2
dX
s
_
+
1
2
A
t
=
1
_
t
0
1
G
s
dG
s
+

2
(A
t
X
t
) R0. (1.4.15)
Here,
1
_
t
0
1
G
s
dG
s
is a local martingale up to T by Proposition 5.3.9 with G =]0, [. Consider
,=
t
and take the dierence of the two corresponding equalities to get
0 = M +

t
2
(A
t
X
t
),
where M is a local martingale up to T. By Corollary 5.3.10(ii) it follows that A
t
X
t
=
A
0
X
0
= 0 on t < T P-a.s., and then (1.4.15) implies that X is a local martingale up to
T.
1.5. Levys Characterization of Brownian Motion
Let X = (X
t
)
t0
be a stochastic process on (, T, P) with continuous sample paths (and conti-
nuous quadratic variation X).
Proposition 1.5.34 (Levy). Assume that X is a continuous local martingale up to with
respect to some ltration (T
t
)
t0
such that X is (T
t
)-adapted. If X
t
= t holds for all t 0,
then X is a Brownian motion.
20
1.5. Levys Characterization of Brownian Motion
Remark 1.5.35. Every continuous martingale can be transformed by a time change into a
Brownian motion provided its quadratic variation is strictly increasing (see later). Therefore,
Brownian motion is the only local martingale, where the quadratic variation is the usual time.
Proof of 1.5.34. Corollary 1.4.32 implies that X is a martingale. For u R we apply the Ito-
formula to
F(x) = e
iux
= cos(ux) +i sin(ux), x R.
Then, for all s < t
e
iuX
tT
n
e
iuX
sT
n
=
_
tT
n
0
iue
iuX
r
dX
r

_
sT
n
0
iue
iuX
r
dX
r
+
1
2
_
tT
n
sT
n
(u
2
)e
iuX
r
dX
r
. .
= dr
,
where T
n
are stopping times such that
__
tT
n
0
iue
iuX
r
dX
r
_
t0
is a martingale for all n. Now take E[[T
s
] of the equality above:
E
_
e
iuX
tT
n
e
iuX
sT
n
T
s
=
1
2
E
__
tT
n
sT
n
(u
2
)e
iuX
r
dr
T
s
_
.
Hence, letting we obtain
1
2
E
__
t
s
(u
2
)e
iuX
r
dr
T
s
_
= E[e
iuX
t
e
iuX
s
[T
s
].
Multiplication by e
iuX
s
yields
E[e
iu(X
t
X
s
)
[T
s
] 1 =
1
2
E
__
t
s
(u
2
)e
iu(X
r
X
s
)
dr
T
s
_
.
Therefore, for all A T
s
E[e
iu(X
t
X
s
)
, A]
. .
=:(t)
P(A)
Fubini
=
1
2
u
2
_
t
s
E[e
iu(X
r
X
s
)
, A]
. .
=(r)
dr.
Then, (
1
(since the right hand side is (
1
) and by dierentiation
(t) =
1
2
u
2
(t) t s.
Solving this equation we get
(t) = C e
1
2
u
2
t
, t s.
Substituting s for t yields
P(A) = (s),
which implies
C = P(A)e
1
2
u
2
s,
hence,
E[e
iu(X
t
X
s
)
; A] = e
1
2
u
2
(ts)
P(A) A T
s
. (1.5.16)
Taking A := implies that X
t
X
s
is N(0, (t s)) distributed by uniqueness of the Fourier-
transform. By a monotone class argument, (1.5.16) implies that X
t
X
s
is independent of T
s
(Exercise!). In particular,
X
t
n
X
t
n1
, . . . , X
t
2
X
t
1
are independent for all 0 t
1
< t
2
< . . . < t
n
< .
21
2. (Semi-)Martingales and Stochastic
Integration
In this chapter we want to dene
_
g(t, ) dM
t
,
where g is dened more general and M
t
is an arbitrary semimartingale. This is more general
than so far, since we could only dene
_
f(t,
t
) dM
t
, f (
1
.
2.1. Review of Some Facts from Martingale Theory
Fix a probability space (, T, P) and let (T
t
)
t0
be a right-continuous ltration. (Sometimes
assume in addition that N T, P(N) = 0 N T
0
.) These conditions are also called usual
conditions.
Proposition 2.1.1. Let (M
t
)
t0
be a martingale. Then there exists a version
(

M)
t0
(i.e. M
t
=

M
t
on N(t)
c
such that P(N(t)) = 0),
such that t

M
t
() is càdlàg (i.e. is right-continuous and has left limits) for all .
Proof. Doobs upcrossing lemma (cf. [vWW90]).
Proposition 2.1.2 (Optional stopping theorem). Let (M
t
)
t0
be a martingale and T be a
stopping time. Then (M
tT
)
t0
is a martingale.
Proof. See [Roc11, Proposition VIII.4.2, Corollary VIII.4.3].
Proposition 2.1.3 (Doobs inequality). Let p > 1. Then
_
_
_
_
sup
st
[M
s
[
_
_
_
_
p
p
p 1
|M
t
|
p
,
where ||
p
= ||
/
p. In particular, if
M
:= sup
t0
[M
t
[,
then
|M
|
p

p
p 1
sup
t0
|M
t
|
p
.
Proposition 2.1.4. Let p 1 and let
_
M
(n)
t
_
t0
, n N, be a sequence of martingales such
that
M
(n)
t
n
M
t
in L
p
, t 0,
Then
23
2. (Semi-)Martingales and Stochastic Integration
(i) (M
t
)
t0
is again a martingale in L
p
.
(ii) If p > 1 and (T)
t0
is such that all P-zero sets in T
t
are in T
0
and each
_
M
(n)
t
_
t0
has
P-a.s. (right-) continuous ((càdlàg)) sample paths, then (M
t
)
t0
has a (right-) continuous
((càdlàg)) (T
t
) -adapted version again denoted by (M
t
)
t0
and
M
(n)
t
n
M
t
in L
p
(, T, P; C ([0, t]; R)) and hence locally uniformly in t and in L
p
and has a locally
uniformly in t P-a.s. convergent subsequence.
Proof. (i) Obvious.
(ii) Fix t > 0. Since M
n
M
m
is a martingale, we have by Doob
_
_
_
_
sup
st
[M
(n)
s
M
(m)
s
[
_
_
_
_
p
p
p 1
_
_
_M
(n)
t
M
(m)
t
_
_
_
p
.
Since they are L
p
-convergent, for some subsequence (n
k
)
kN
k=1
_
_
_
_
sup
st
M
(n
k+1
)
s
M
(n
k
)
s
_
_
_
_
p
p
p 1
k=1
_
_
_M
(n
k+1
)
t
M
(n
k
)
t
_
_
_
p
< .
Therefore,
P
_

k=1
sup
st
M
(n
k+1
)
s
M
(n
k
)
s
<
_
= 1.
Hence, M
(n
k
)
s
k
M
s
uniformly on [0, t] P-a.s. (cf. Proof of Riesz-Fischer!). Then up to
a P-zero set M
s
is T
s
-measurable, since so is each M
(n
k
)
s
. But T
s
contains all P-zero sets
in T
0
. Therefore, M
s
is T
s
-measurable.
Remark 2.1.5 (Localization). Let (M
t
)
t0
be a continuous local martingale (up to ) such
that M
0
= 0. Set
R
n
() := inf
_
t > 0
[M
t
()[ > n
_
_
n

_
.
Claim: For all n N (M
tR
n
)
t0
is a continuous bounded martingale.
Proof. Let (T
k
)
kN
be a localizing sequence for M, such that T
k
k. (Otherwise consider T
k
k.)
Then for s t and n xed
E[M
tR
n
T
k
; A
s
] = E[M
sR
n
T
k
; A
s
] A
s
T
s
.
By Lebesgue, for all A
s
T
s
,
E[M
tR
n
T
k
; A
s
]
k
E[M
tR
n
; A
s
]
and, since [M
tR
n
T
k
[ n for xed n,
E[M
sR
n
T
k
; A
s
]
k
E[M
sR
n
; A
s
].
So, we get E[M
tR
n
; A
s
] = E[M
sR
n
; A
s
].
24
2.2. Quadratic Variation and Covariation for Continuous Local Martingales
2.2. Quadratic Variation and Covariation for Continuous Local
Martingales
Let (M
t
)
t0
be a càdlàg martingale.
Approach to stochastic integration:
Assume initially M
t
L
2
for all t 0. (If M
t
, L
2
, then localize.) By Jensens inequality (M
2
t
)
t0
is a submartingale. Then one can show (by the Doob-Meyer decomposition) that there exists a
unique adapted process (M
t
)
t0
with M
0
= 0, increasing, right-continuous, predictable (see
below) such that
_
M
2
t
M
t
_
t0
()
is a martingale. Then, M is the variance process of M, i.e.
E[(M
t
M
s
=0
..
E[M
t
M
s
])
2
[T
s
] (conditioned variance of M
t
M
s
given T
s
)
=E[M
2
t
M
2
s
[T
s
]
()
= E[M
t
M
s
[T
s
].
In [Roc11] we proved the Doob-Meyer decomposition in discrete time. In continuous time for
càdlàg martingales this is much more dicult (cf. [Kry]).
We are going to prove the Doob-Meyer decomposition only for continuous martingales. In the
stochastic integration theory below, we shall, however, allow càdlàg martingales, simply assuming
Doob-Meyer without proof, and refering to [vWW90, Cor. 6.6.3] instead.
For a continuous martingale M we shall construct the process M such that it pathwise coincides
P-a.s. with the quadratic variation of M along (
n
)
nN
of the previous chapter. Let M be a
continuous local martingale (up to ) and let (
n
)
nN
be a sequence of partitions of [0, ) such
that
[
n
[
n
0
and
t
(n)
N
n
n
.
Denition 2.2.6. Let
V
(n)
t
:=
s
n
(M
s
/
t
M
st
)
2
, n N, t 0
be the quadratic variation of M along
n
on [0, t]. Here, s
t
denotes the successor of s in
n
.
Note that t V
(n)
t
is P-a.s. continuous for all n N.
Remark 2.2.7. Since (a b)
2
= a
2
2b(a b) b
2
,
V
(n)
t
= M
2
t
M
2
0
2
u
n
M
ut
(M
u
/
t
M
ut
)
= M
2
t
M
2
0
2
u
n
, ut
M
u
(M
u
/
t
M
u
)
. .
local martingale
.
Proposition 2.2.8. Assume that T
0
contains all P-zero sets. Let M be a continuous (for all
), local martingale (up to ). Fix (
n
) as above and let (V
(n)
)
nN
be as in Denition
2.2.6. Then there exists a continuous increasing (T
t
)-adapted process M with M
0
= 0 such
that
25
(i) M
t
= lim
n
V
(n)
t
in P-measure locally uniformly in t 0.
(Convergence is even in L
2
(, T, P; (([0, t]; R)) for all t 0, if M is a bounded martin-
gale.)
(ii) M
2
M is a continuous local martingale.
(iii) If M
t
L
2
for all t 0 and if M
t
is a martingale, then M
2
t
M
t
is a martingale. In
particular,
E[M
2
t
] = E[M
2
0
] +E[M
t
] t 0.
(Warning: At this stage M = (M
t
)
t0
depends on (
n
)!)
Proof. (i) Case 1: M is bounded (i.e. sup
t0,
[M
t
()[ < ) and a martingale.
Claim: (V
(n)
t
)
nN
is a Cauchy sequence in L
2
for all t 0.
Take m, n N, N large so that t t
(n)
N
n
, t t
(m)
N
m
. Dene
V
(m,n)
t
:= M
2
t
M
2
0
2
s
n
m
M
s
(M
s
/
t
M
st
)
_
= V
(n,m)
t
_
.
Here, s
t
denotes the successor of s in
n

m
. Then by Remark 2.2.7
V
(m,n)
t
V
(n)
t
= 2
u
n
__

us<u
/
s
n
m
M
s
(M
s
/
t
M
st
)
_
M
u
(M
u
/
t
M
ut
)
_
,
where u
t
is the successor in
n
and s
t
the one in
n

m
. Using the fact that
M
u
/
t
M
ut
=
us<u
/
s
n
m
(M
s
/
t
M
st
)
we conclude that
V
(m,n)
t
V
(n)
t
= 2
u
n
us<u
/
s
n
m
(M
s
M
u
)(M
s
/
t
M
st
).
Hence,
E
_
_
V
(m,n)
t
V
(n)
t
_
2
_
= 4E
_

u
n
us<u
/
s
n
m
(M
s
M
u
)
2
(M
s
/
t
M
st
)
2
_
Note that all occurring mixed terms are zero by the martingale property (exercise!).
Furthermore,
E
_
_
V
(m,n)
t
V
(n)
t
_
2
_
4E
_
s
n
m
(M
s
/
t
M
st
)
2
_
,
where
N
:= sup
nN
sup
m
sup
_
(M
s
M
u
)
2
s
n

m
, u
n
, u s u
t
, s, u t
_
N
0 Pa.s.
26
is bounded in N and , since M is bounded. Then, by Cauchy-Schwarz (since n, m N)
E
_
_
V
(m,n)
t
V
(n)
t
_
2
_
4
_
E
_
2
N
_
1/2
_
_
_
_
_
s
n
m
(M
s
/
t
M
st
)
2
_
_
_
_
_
2
.
By Lebesgues dominated convergence theorem,
_
E
_
2
N
_
1/2
= |
N
|
2
N
0,
since M() is uniformly continuous on [0, t] for all t and (M
t
)
t0
is bounded in (, t). But
for c := sup
t,
[M
t
()[ and for all n, m N
E
_
_

s
n
m
(M
s
/
t
M
st
)
2
_
2
_
=E
_

s
n
m
(M
s
/
t
M
st
)
4
_
+ 2E
_

s
n
u
n
m
s<u
(M
s
/
t
M
st
)
2
(M
u
/
t
M
ut
)
2
_
4c
2
E
_

s
n
m
(M
s
/
t
M
st
)
2
_
. .
=:S
1
+ 2
s
n
u
n
m
s<u
E
_
(M
s
/
t
M
st
)
2
E[(M
u
/
t
M
ut
)
2
[T
ut
]
. .
=:S
2
and
S
1
=
s
n
m
E[M
2
s
/
t
2M
s
/
t
M
st
+M
2
st
]
=
s
n
m
E[M
2
s
/
t
+M
2
st
2E[M
s
/
t
M
st
[T
s
]]
=
s
n
m
E[M
2
s
/
t
+M
2
st
2M
st
E[M
s
/
t
[T
s
]
. .
M
st
]
=
s
n
m
E[M
2
s
/
t
M
2
st
] = E[M
2
t
M
2
0
]
In addition,
S
2
=
u
n
m
s<u
E
_
(M
s
/
t
M
st
)
2
E[M
2
u
/
t
M
2
ut
[T
ut
]
u
n
m
s<u
E
_
(M
s
/
t
M
st
)
2
(M
2
u
/
t
M
2
ut
)
= E[(M
s
/
t
M
st
)
2
(M
2
t
M
2
s
/
t
)]
2c
2
E[E[(M
s
/
t
M
st
)
2
[T
st
]] = 2c
2
E[M
2
s
/
t
M
2
st
].
27
Thus,
E
_
_

s
n
m
(M
s
/
t
M
st
)
2
_
2
_
4c
2
E[M
2
t
M
2
0
] + 4c
2
s
n
m
E[M
2
s
/
t
M
2
st
]
8c
2
E[M
2
t
M
2
0
] 16c
4
.
Alltogether, we obtain for all n, m N
_
E
_
(V
(n)
t
V
(m)
t
)
2
__
2E
_
(V
(n)
t
V
(n,m)
t
)
2
_
+ 2E
_
(V
(m,n)
t
V
(m)
t
)
2
_
8 |
N
|
2
4c
2
+ 8 |
N
|
2
4c
2
N
0
and the claim is proved.
Let V
t
:= L
2
- lim
n
V
(n)
t
, t 0. Dene
Y
(n)
t
:= 2
s
n
M
s
(M
s
/
t
M
st
).
Then each (Y
(n)
t
)
t0
is an (T
t
)-martingale and by Remark 2.2.7 and the above
Y
(n)
t
= M
2
t
M
2
0
V
(n)
t
n
M
2
t
M
2
0
V
t
=: Y
t
in L
2
. ()
Hence, by Proposition 2.1.4 Y
t
is a martinale and has a P-a.s. continuous version and,
therefore, V
t
has a continuous version M and M is (T
t
)-adapted since all P-zero sets in
T are in T
0
, hence in every T
t
. Furthermore, by Proposition 2.1.4(ii), Y
(n)
Y as n
in L
2
(, T, P; C([0, t]; R)), hence V
(n)
M as n in L
2
(, T, P; C([0, t]; R)), i.e.
E
_
sup
st
[V
(n)
s
M
s
[
2
_
n
0.
Case 2: M as in the assertion.
Without loss of generality M
0
0. (Otherwise consider (M
t
M
0
)
t0
.). Let R
k
be as in
2.1.5, i.e.
R
k
:= inft > 0
[M
t
[ > k
and therefore, R
k
is a stopping time. Dene
M
k
t
:= M
tR
k
, t 0.
By Remark 2.1.5 M
k
t
is a bounded martingale. Hence, by Case 1 there exists V
k
t
:=
M
k
t
, t 0. Let (V
k
t
)
(n)
, t > 0, be the corresponding approximations along (
n
). Then
there exists a subsequence (n
l
)
lN
and
0
T
0
, P(
0
) = 1 such that for all k N (cf.
Proposition 2.1.4(ii))
(V
k
s
)
(n
l
)
()
l
M
k
s
()
locally uniformly for s [0, t] and for P-a.e.
0
. But
1
tR
k
V
(n
l
)
t
= 1
tR
k
V
(n
l
)
tR
k
= 1
tR
k
(V
k
t
)
(n
l
)
l
1
tR
k
M
k
t
P a.s.. ( )
28
Hence, we can (well-)dene
M
t
() :=
_
M
k
t
(), with k N such that t R
k

0
,
0, otherwise.
Recall that by the continuity, hence local boundedness of t M
t
() we have that R
k
()
, hence
0
=
kN
t < R
k

0
, so M
t
() is dened by the above
and by ( ) this denition is independent of k. Then M is P-a.s. continuous and
(T
t
)-adapted.
Furthermore, since R
k
as k , for > 0 we can x k = k() large enough, such
that
P[R
k
< t] < .
Then
P
_
sup
0st
V
(n)
s
M
s

_
=P
_
sup
0st
V
(n)
s
M
s
, R
k
< t
_
+P
_
sup
0st
V
(n)
s
M
s
, R
k
t
_
P (R
k
< t) +P
_
sup
0st
(V
k
s
)
(n)
M
k
, R
k
t
_
P (R
k
< t) +P
_
sup
0st
(V
k
s
)
(n)
M
k

_
.
But by the rst part of the proof and Chebychevs inequality
P
_
sup
0st
(V
k
s
)
(n)
M
k

_
n
0.
Therefore,
limsup
n
P
_
sup
0st
V
(n)
s
M
s

_
P[R
k
< t] < .
It remains to show that t M
t
is increasing P-a.s.. But for xed t 0 we have that
V
(n)
t
=
s
n
,st
(M
s
/ M
s
)
2
+N
(n)
t
n
M
t
,
where the sum is increasing in t and
N
(n)
t
:= (M
t
M
(n)
t
)
2
n
0
and
(n)
t
:= sups
n
[s t.
Hence, (i) is completely proved.
(ii) Continuity is clear. By () on p. 28 we know, that P-a.s.
M
2
tR
k
M
tR
k
= (M
k
t
)
2
M
k
t
= Y
k
t
+M
2
0
and the right hand side is a martingale. Hence, M
2
M is a local martingale for T up
to .
29
(iii) Without loss of generality M
0
= 0. We have to show that M
2
T
M
T
L
1
(P) and that
E[M
T
] = E[M
2
T
]
for all bounded stopping times T. By the monotone convergence theorem, we get that
E[M
T
] = lim
n
E[M
TR
n
]
proof of
(ii)
= lim
n
E[M
2
TR
n
]
Fatou
E[M
2
T
].
On the other hand, by the submartingale property, lim
n
E[M
2
TR
n
] E[M
2
T
], so
E[M
T
] = E[M
2
T
].
Finally, by assumption,
E[M
2
T
]
_
M
2
sup
T()
_
< ,
so that
E[M
2
T
M
T
] = 0.
Remark 2.2.9. (i) One can drop the assumption that all P-zero sets in T are in T
0
, but
one only gets that the particular version of M of V is only right-continuous and adapted.
But it is continuous only for N
c
, with a P-zero set N T.
(ii) Since convergence in probability implies P-a.s. convergence of a subsequence, it follows by
Proposition 2.2.8(i) that for some subsequence (n
k
)
kN
P(V
(n
k
)
s
k
M
s
locally uniformly on [0, t], t 0) = 1.
(iii) Since M is exactly the pathwisely dened continuous quadratic variation process of M in
Chapter I, we can apply all results form Chapter I for P-a.e. xed, i.e.
M() = M().
Corollary 2.2.10. Assume M
0
0. Then
M
2
t
M
t
= 2
_
t
0
M
s
dM
s
.
Hence, the continuous local martingale M
2
M is an Ito-integral (cf. Corollary 5.3.10(i)).
Corollary 2.2.11. P-a.s. the paths of a continuous local martinale are either constant or of
unbounded variation.
Proof. Apply Corollary 5.3.10(ii).
Corollary 2.2.12. M is the unique increasing continuous adapted process such that M
0
= 0
and M
2
M is a continuous local martingale. In particular, M is (up to a P-zero set in T)
independent of the chosen sequence of partitions (
n
) in 2.2.6.
Proof. Let A, B be two such processes. Then M
2
A, M
2
B are continuous local martingales.
Hence, BA is a continuous local martingale (with respect to (T
t
) since A, B are (T
t
)-adapted)
of bounded variation. Therefore, by Corollary 2.2.11
B A = c = B
0
A
0
= 0 P-a.s..
30
Denition 2.2.13. (cf. 5.2.3, 1.3.15) Let M, N be continuous local martingales. Then dene
the covariation process of M, N by
M, N
t
:=
1
2
(M +N
t
M
t
N
t
) .
Remark 2.2.14. Since M+N is a continuous local martingale, M+N exists as a continuous
process by Proposition 2.2.8 Therefore, since
1
2
[(a+b)
2
a
2
b
2
] = ab a, b R (
polarization
identity)
M, N = lim
n
s
n
(M
s
/
t
M
st
)(N
s
/
t
N
st
).
(Exercise!)
Proposition 2.2.15. Let M, N be continuous local martingales. M, N is uniquely determined
by the following:
(i) Its paths are (P-a.s.) continuous and of bounded variation and M, N
0
= 0.
(ii) M N M, N is a continuous local martingale. In particular, M, N 0 if and only if
M N is a local martingale.
Proof. Analogous to the case M = N in Corollary 2.2.12 (or use polarization.).
Lemma 2.2.16. Let M, N be a continuous local martingale. Let G(= G
s
(), s 0, ) and
H be B(R
+
) T-measurable. Then
_
t
0
H
s
()G
s
() dM, N
s
()
__
t
0
H
2
s
dM
s
()
_
1
2
__
t
0
G
2
s
dN
s
()
_
1
2
.
In particular, we obtain (by Cauchy) the Kunita-Watanabe inequality
E
_
_
t
0
H
s
()G
s
() dM, N
s
_
E
__
t
0
H
2
s
dM
s
_
1
2
E
__
t
0
G
2
s
dN
s
()
_
1
2
.
Proof. Exercise (cf. [RW87, Vol II, p.50]).
31
2.3. Construction of stochastic integrals
Fix a probability space (, T, P) with right-continuous ltration (T
t
) such that all P-zero sets
in T are in T
0
. We want to dene
H
M :=
_

0
H
s
dM
s
as a martingale
for càdlàg (T
t
)-martingales and most general H, where càdlàg means right-continuous for all
with left limits P-a.s.. (The latter is automatic by [vWW90, p.47, 3.25, 3.26].) M
s
is
called integrator process and H
s
integrand process.
Admissible integrators are given by
/
2
:= /
2
(, T, P) :=
_
M
M is a càdlàg martingale, M
0
= 0, |M|
2
:= sup
t0
E[M
2
t
] <
_
.
We dene
M
2
c
:= M /
2
[ M has P-a.s. continuous sample paths
We know that for each M /
2
the Doob-Meyer decomposition holds, that is:
For all M /
2
there exists a process M such that it is the unique predictable, right continuous
increasing, adapted process and that M
0
0 and M
2
M is a martingale. We only proved
that this M exists and is unique if M /
2
c
(see Proposition 2.2.8 and Corollary 2.2.12). For
the general case see [Kry] or [vWW90, p.130, Cor. 6.6.3]. However, if M /
2
/
2
c
, then M
is NOT equal to the quadratic variation of M.
Remark 2.3.17. (i) Dene T
:=
_
t0
T
t
_
. Clearly, M
t
L
2
(, T
, P) for all M
/
2
and by the L
2
-martingale convergence theorem we have for any càdlàg martingale
such that M
0
= 0:
M /
2
M
:= lim
t
M
t
L
2
(, T
, P).
In this case
M
t
= E[M
[T
t
], t 0.
So (M
t
)
t[0,]
is a martingale.
(ii) Starting with the norm || dene an inner product on /
2
by polarization:
(M, N) :=
1
4
(|M +N|
2
|M N|
2
), M, N /.
It remains to check that it is really an inner product. This could be done by checking that
|| on /
2
satises
|M +N|
2
+|M N|
2
= 2 |M|
2
+ 2 |N|
2
(parallelogram identity)
and using J. von Neumanns theorem. But (iii) implies this more easily.
(iii) Let M /
2
. Since M
2
is a submartingale and because of (i) we have
E[M
2
]
(i)
= lim
t
E[M
2
t
] = sup
t0
E[M
2
t
] = |M|
2
.
Hence, for N /
2
, by polarization
E[M
] =
1
4
_
E[M
+N
]
2
E[M
]
2
_
=
1
4
(|M +N|
2
|M N|
2
) = (M, N).
Therefore, (, ) is really an inner product with corresponding norm ||.
32
(iv) Note that |M|
2
= sup
t0
E[M
2
t
] E[sup
t0
M
2
t
]
Doob
4 sup
t0
E[M
2
t
] 4 |M|
2
. So, on /
2
the
norms || and ||
L
2
(,T,P; C
b
([0,); R))
are equivalent.
Proposition 2.3.18. (i) /
2
is a Hilbert space and
(M, N) = E[M
] = E[M, N
],
where
M, N
:= lim
t
M, N
t
.
(ii) /
2
c
is a closed subspace of /
2
.
Proof. (i) (/
2
, ||) is an inner product space (Pre-Hilbert space) by Remark 2.3.17. Further-
more, it is complete by Proposition 2.1.4(ii) and because
L
2
(, T, P; C
b
([0, ), R)),
is complete.
Finally,
|M|
2
= E[M
2
] = lim
t
E[M
2
t
] = lim
t
E[M
t
]
B.Levi
= E[ lim
t
M
t
. .
=:M)
].
So, by polarization the last assertion follows.
(ii) Apply Proposition 2.1.4(ii).
Now we dene stochastic integrals with M /
2
as integrators, but rst for elementary functi-
ons:
Denition 2.3.19. Dene c to be the set of all processes H on ]0, [ which are of the
following form:
For t > 0,
H
t
() :=
n1
i=0
h
t
i
()1
]t
i
,t
i+1
]
(t), elementary (predictable) adapted processes
where n N, 0 = t
0
< t
1
< . . . < t
n
< and h
t
i
are T
t
i
-measurable bounded. For M /
2
and H as above dene
_
t
0
H
s
dM
s
:=
n1
i=0
h
t
i
(M
t
i+1
t
M
t
i
t
), t 0.
An easy exercise shows that this is independent of the representation of H! Set
(H
M)
t
:=
_
t
0
H
s
dM
s
, t 0.
Then H
M is called the stochastic integral of H with respect to M.

Lemma 2.3.20. Let H c.
33
(i) H
M /
2
and if M /
2
c
, then H
M /
2
c
.
(ii)
H
M
t
=
_
t
0
H
2
s
dM
s
t 0
=
n1
i=0
h
2
t
i
(M
t
i+1
t
M
t
i
t
).
In particular,
|H
M|
2
= E
__

0
H
2
s
dM
s
_
. Isometry
Note that here we always set
_
t
0
H
2
s
dM
s
:=
_
]0,t]
H
2
s
dM
s
.
Proof. (i) By denition we see that:
if M is a cadlag (continuous) martingale, then H
M is càdlàg (continuous),
H
M is adapted,
(H
M)
0
= 0.
Furthermore, since M is a martingale,
sup
t0
E[(H
M)
2
t
] = sup
t0
E
_
n1
i=0
h
2
t
i
_
M
t
i+1
t
M
t
i
t
_
2
_
= sup
t0
n1
i=0
E
_
h
2
t
i
_
M
t
i+1
t
M
t
i
t
_
2
_
sup
0in
_
_
h
2
t
i
_
_
sup
t0
n1
i=0
E
__
M
2
t
i+1
t
M
2
t
i
t
__
sup
0in
_
_
h
2
t
i
_
_
sup
t0
E
_
M
2
t
n
t
sup
0in
_
_
h
2
t
i
_
_
|M|
2
< .
It remains to show the martingale property of H
M:
Let T be a bounded stopping time. Then
E[(H
M)
T
] =
n1
i=0
E
_
h
t
i
_
M
t
i+1
T
M
t
i
T
_
_
=
n1
i=0
E
_
h
t
i
E
_
M
t
i+1
T
M
t
i
T
[T
t
i
. .
=0, since (M
tT
)
t0
is an (T
t
)-martingale
_
= 0.
Thus, H
M is a martingale.
34
(ii) By uniqueness of the Doop-Meyer decomposition (see Corollary 2.2.12 for M /
2
c
) it is
enough to show that
(H
M)
2
t

_
t
0
H
2
s
dM
s
, t 0,
is a martingale. (Then H
M
t
=
_
t
0
H
2
s
dM
s
.) Let T be a bounded stopping time. Then,
by dening
i
M := M
t
i+1
T
M
t
i
T
and since all mixed terms dissappear, we get that
E[(H
M)
2
T
] =
i,j
E
_
h
t
i
h
t
j
(
i
M)(
j
M)
i
E
_
h
2
t
i
E[(
i
M)
2
[T
t
i
]
i
E
_
h
2
t
i
E[M
2
t
i+1
T
M
2
t
i
T
[T
t
i
]
_
=
i
E
_
h
2
t
i
E[M
t
i+1
T
M
t
i
T
[T
t
i
]
= E
__
T
0
H
2
s
dM
s
_
.
In particular,
|H
M|
2
= sup
t0
E[(H
M)
2
t
] = sup
t0
E
__
t
0
H
2
s
dM
s
_
B.Levi
= E
__

0
H
2
s
dM
s
_
.
We now want to consider the above Isometry Property more closely:
Let

:=]0, [, := (t, ),

T = B(]0, [) T. M
t
() is T-measurable for xed t
and for all t M
t
() is right-continuous, nonnegative, increasing, with M
0
() = 0.
Therefore, for xed there exists a unique, positive measure
M(, dt) := dM
t
().
Thus, M(, dt) denes a transition kernel from (, T) to (]0, [, B(]0, [)) and it induces a
measure
P
M
(d ) := P(d) M(, dt)
on (
,

T). Explicitly,
P
M
(A) = E
__

0
1
A
(, t) dM
t
()
_
, A

T.
Note that this is not a probability measure, but it is nite since M /
2
(c)
. In particular,
dening
E
M
[ ] =
_
dP
M
,
we have
E
M
[H
2
] = E
__

0
H
2
s
dM
s
_
, (2.3.1)
35
which is the L
2
(
,

T, P
M
)-norm of H c. (H :

R, H(, t) = H
t
().) Note that the map
c /
2
,
H H
M
is obviously linear and by Lemmas 2.3.20(ii) and (2.3.1), an isometry from c L
2
(
,

T, P
M
)
to (/
2
(c)
, ||). Therefore, there exists a unique isometric extension to the closure of c in
L
2
(
,

T, P
M
) denoted by
c :=

c
M
(which depends on M).
Denition 2.3.21. For H

c
M
, let H
M denote the uniquely determined element in /

2
(c)
with
lim
n
|H
n
M H
M| = 0
for every sequence (H
n
)
nN
c which converges in L
2
(
,

T, P
M
) to H.
Automatically, we have
|H
M|
2
= E
P
M
[H
2
].
But we also have the analogue of 2.3.20(ii), namely:
Proposition 2.3.22. Let H

c
M
, M /
2
(c)
, and therefore, H
M /
2
(c)
. Then
H
M
t
=
_
t
0
H
2
s
dM
s
, t 0,
and hence by polarization for G

c
M
H.M, G.M
t
=
_
t
0
H
s
G
s
M
s
t > 0.
Proof. Let T be a bounded stopping time. Then
E[(H
M)
2
T
] = lim
n
E[(H
n
M)
2
T
],
since
E[((H
M)
T
(H
n
M)
T
)
2
] E
_
sup
t
(H
M H
n
M)
2
t
_
Doob
4 |H
M H
n
M|
2
n
0.
Hence,
E[(H
M)
2
T
] = lim
n
E[(H
n
M)
2
T
]
2.3.20(ii)
= lim
n
E
__
T
0
(H
n
)
2
s
dM
s
_
= E
__
T
0
H
2
s
dM
s
_
,
since H
n
n
H in L
2
(P
M
), so, 1
]0,T]
H
n
1
]0,T]
H in L
2
(P
M
). Therefore, (H
M)
2
t
_
t
0
H
2
s
dM
s
is a martingale and the assertion follows by the uniqueness of the Doob-Meyer decomposition.
36
In our next step we want to determine the size of

c
M
. We want to characterize admissible
integrands in dependence of M.
Denition 2.3.23.
T :=
_
H = (H
t
)
t>0
[H is a (real-valued) left-continuous, adapted process
_
exercise
=
_
H = (H
t
)
t>0
[H is a (real-valued) continuous, adapted process
_
is called -algebra (on ]0, [) of predictable sets. A process H = (H
t
)
t>0
is called predictable,
if it is T-measurable.
Let T
M
:= the completion of T with respect to P
M
in

T.
Remark 2.3.24. Every predictable process (H
t
)
t>0
is progressively measurable, i.e. H
]0,t]
is
B(]0, t]) T
t
-measurable for all t > 0, hence in particular (T
t
)-adapted (cf. [vWW90, Lemma
6.1.6]).
Remark 2.3.25. Let
n
be a partition of [0, ) and
T
n
:= (]t, t
t
[A
t
t
n
, t
t
is the successor of t in
n
and A
t
is T
t
-measurable)
T
n
is called the -algebra of predictable rectangles (with respect to
n
).
(i) If [
n
[
n
0 and t
N
n
n
, then
T =
_
_
n
T
n
_
.
(ii) H is T
n
-measurable and bounded if and only if
H
t
() =
t
i
n
h
t
i
()1
]t
i
,t
i+1
]
(t), t > 0,
where h
t
i
is T
t
i
-measurable and bounded.
Proof. (i) By denition we have T
n
T. Hence, (
n
T
n
) T. To show that T (
n
T
n
),
let (H
t
)
t>0
be a left-continuous adpated process. It suces to show that H is (
n
T
n
)-
mearsurable. H is adapted and left-continuous. Therefore,
H
t
() = lim
n
s
n
H
s
()1
]s,s
/
]
(t)
. .
(1
n
)-measurable
, t > 0,
Hence,
T =
_
_
n
T
n
_
.
(ii) The assertion follows by a monotone class argument.
Proposition 2.3.26.
c
M
= L
2
(
, T
M
, P
M
).
In particular, for all H L
2
(
, T
M
, P
M
) there exists
H
M =
_
H dM.
In particular,
/
2
M
:= H
M[ H L
2
(
, T
M
, P
M
)
is a closed subspace of /
2
.
37
Proof. Clearly,

c
M
L
2
(
, T
M
, P
M
). To prove the dual inclusion, let H L
2
(
, T
M
, P
M
).
Without loss of generality H 0. (Otherwise consider H
+
and H
.) Since
H
(m)
:= (m H)
m
H
in L
2
(
, T
M
, P
M
), we may assume H 0, bounded. Now let
n
be a sequence of partitions as in
Remark 2.3.25(i). Then by the L
1
-martingale convergence theorem (see [Roc11, Korollar 8.5.4
(i)]; ok, since T
M
is a nite measure!)
H
n
:= E
P
M
[H[T
n
]
n
E
P
M
_
H
_
_
n
T
n
__
in L
1
(T
M
).
Since H is bounded, this convergence also holds in L
2
(T
M
). By Remark 2.3.25(i) this implies
that in L
2
(T
M
)
lim
n
H
n
= lim
n
E
P
M
[H[T
n
] = E
P
M
[H[T] = E
P
M
[H[T
M
] = H.
Since H
n
is T
n
-measurable, we have by Remark 2.3.25(ii), that
H
n
c,
and the assertion follows.
Proposition 2.3.27. (i) Let t M
t
be continuous (which is e.g. the case if M /
2
c
).
Then
_
L
2
(
, T
M
, P
M
)
2.3.26
=
_

c
M
= L
2
(
, O
M
, P
M
),
where the optional -algebra O is dened by
O := (H[ H is an adapted càdlàg process on ]0, [)
and O
M
its completion with respect to P
M
in

T.
(ii) If M /
2
c
and dM is absolutely continuous with respect to Lebesgue measure dt, then
c
M
= H L
2
(
,

T, P
M
)[H has an adapted version.
Proof. (i) We have that O T. Hence, also O
M
T
M
. It remains to prove O
M
T
M
. So,
let H be càdlàg and adapted. Consider for
U
:= t > 0[s H
s
() is discontinuous in t.
Then U
is countable, since H is càdlàg (exercise). Dene for

H
t
() := lim
s,t
H
s
() =: H
t
() ( R).
Then H
is left-continuous. Hence, it is T-measurable and

U
= t > 0[H
t
() ,= H
t
().
Note that (t, ) ]0, [[ H
t
() ,= H
t
() O

T. Since M is continuous, we have
_

0
1
U
(t) dM
t
= 0 for P-a.e. .
But then
E
P
M
[1
H,=H
] =
_ _
]0,[
1
U
(t) dM
t
()P(d) = 0.
Hence, H is P-a.e. equal to a predictable process, therefore, H is T
M
-measurable and (i)
is proved.
38
(ii) One has to prove that

T T
M
(cf. [CW90, p.60] or [vWW90, p.124]).
Remark 2.3.28. By a modication of the proof of Proposition 2.3.27(i) one can even show
that every progressively measurable process (H
t
)
t>0
(see Remark 2.3.24 above) has a predictable
T
M
-version (of course still assuming that M is P-a.s. continuous).
Remark 2.3.29 (Extending stochastic integration via localization). Note that, if M is a Brow-
nian motion, then M , /
2
. Therefore, we have to stop. Dene
/
2
loc
:= M[ a sequence T
n
of stopping times such that M
T
n
(:= M
T
n
) /
2
n N.
Then for all H L
2
loc
(
, T
M
, P
M
)
H
M =
_
H dM
is dened via localization. We have to check that
the denition is consistent on T
n
= T
n+1
,
H
M /
2
loc
(We need Lemma 2.4.33 below).
Remark 2.3.30 (Semi-martingales as integrals). If A is predictable and of bounded variation
and M /
2
loc
we can dene for H L
2
loc
(
, T
M
, P
M
)
_
H d(M +A) :=
_
H dM +
_
H dA,
where M +A is a semi-martingale and
_
H dA is a pathwise dened Lebesgue-Stieltjes-integral.
Axiomatic considerations show that, if
_
H d(M +A) is required to have reasonable properties,
then this cannot be generalized. As a consequence, reasonable stochastic integrators are semi-
martingales (Dellacherie-Mokobodzki-Bichteler) (cf. [vWW90]).
39
M in /
2
Fix M /
2
, H L
2
(
, T
M
, P
M
).
Proposition 2.4.31. H
M is the unique element L /

2
such that
(i)
dL, N = H dM, N P-a.s. N /
2
,
that is
L, N
t
=
_
t
0
H
s
dM, N
s
:=
_
1
]0,t]
(s)H
s
dM, N
s
, t 0, P-a.s. N /
2
, t > 0,
respectively, such that the following weaker property holds:
(ii)
E[L
] = E
__

0
H
s
dM, N
s
_ _
:=
_
1
]0,[
(s)H
s
dM, N
s
_
for all N /
2
.
Remark 2.4.32. (i) Because of E[L
] = E[L, N
] (cf 2.3.18(i)), it follows that in

2.4.31, (i) implies (ii).
(ii) By 2.4.31(i) with N replaced by G.N for G L
2
(
, T
M
, P
M
) we particularly have
dH
M, G.N = H dM, G.N = HG dM, N.

Proof of 2.4.31. (a) Uniqueness if (ii) holds:
Let L, L
t
/
2
such that both satisfy 2.4.31(ii). Then
E[(L
L
t
)N
] = 0 N
L
2
(, T
, P).
Hence L
= L
t
P-a.s.. Therefore,
L
t
= E[L
[T
t
] = L
t
t
P-a.s..
Now let us rst prove (ii).
(b) L := H
M satises (ii):
Step 1: Assume for xed s > 0 that
H
t
() := h
s
()1
]s,[
(t),
where h
s
is bounded and T
s
-measurable. Note that H L
2
(
, T
M
, P
M
). Therefore, for
L := H
M we have
L
= (H
M)
=
_

0
h
s
1
]s,[
(t) dM
t
= lim
N
_
N
0
h
s
1
]s,]
(t) dM
t
= lim
N
h
s
(M
N
M
s
) = h
s
(M
M
s
)
40
M in /
2
Thus, for all N /
2
E[L
] =E[h
s
(M
M
s
)N
]
(N
t
)
t[0,]
mart.
= E[h
s
(M
M
s
N
s
)]
=E [h
s
(M
M, N
(M
s
N
s
M, N
s
))]
+E [h
s
(M, N
M, N
s
)]
=E [h
s
(M, N
M, N
s
)]
=E
_
h
s
_
]s,[
dM, N
t
_
=E
__

0
H
t
dM, N
t
_
.
Step 2: Let H c.
In this case (ii) is clear by Step 1 and linearity.
Step 3: (ii) holds for all H L
2
(
, T
M
, P
M
) :
Let H
(n)
c, n N, H
(n)
H in L
2
(
, T
M
, P
M
)(=

c). Then
H
(n)
M H
M in (/
2
, ||).
Therefore, by 2.3.17(iii) for all N /
2
E[(H
M)
] = lim
n
E[(H
(n)
M)
]
Step 2
= lim
n
E
__

0
H
(n)
s
dM, N
s
_
= E
__

0
H
s
dM, N
s
_
,
because by Kunita-Watanabe-inequality (cf. 2.2.16) we have
E
_
_

0
(H
(n)
H) 1 dM, N
_
E
__

0
(H
(n)
H)
2
dM
__1
2
(E[N
])
1
2
n
0.
This proves (ii).
(c) L := H
M satises (i):
By Proposition 2.2.15 we have to show that L
t
N
t
_
t
0
H
s
dM, N
s
, t 0, is a martingale.
Let T be a bounded stopping time. Then, since (L
t
)
t[0,]
is a martingale, for N
T
t
:=
N
Tt
, t 0, we have
E[L
T
N
T
] =E[L
N
T
] = E[L
N
T
]
(ii)
=E
__

0
H
s
dM, N
T
s
_
2.4.33 below
= E
_ _

0
H
s
dM, N
sT
_
=E
_ _
T
0
H
s
dM, N
s
_
(where we used that obviously N
T
/
2
, since N /
2
).
41
Lemma 2.4.33. Let M, N /
2
and T be a stopping time. Then
M, N
T
t
= M, N
Tt
t 0.
Proof. Exercise. (!)
Corollary 2.4.34. Let M /
2
, H
1
, H
2
L
2
(
, T
M
, P
M
) such that H
1
H
2
L
2
(
, T
M
, P
M
).
Then
H
1
(H
2
M) = (H
1
H
2
)
M,
i.e.
_

0
H
1
(s) d(H
2
M)
s
=
_

0
H
1
(s) d
__
s
0
H
2
(r) dM
r
_
=
_

0
H
1
(r)H
2
(r) dM
r
.
Proof. Let L := H
1
(H
2
M). Then by 2.4.31 for all N /
2
dL, N = H
1
dH
2
M, N = H
1
H
2
dM, N,
and
d(H
1
H
2
).M, N = (H
1
H
2
)
M.
Hence (again by 2.4.31)
(H
1
H
2
).M = H
1
(H
2
.M).
Corollary 2.4.35. Let M /
2
and let T be a stopping time.
(i) Then (H
M)
T
= H1
]0,T]
M, i.e.
(H
M)
T
t
=
_
t
0
H
s
1
]0,T]
(s) dM
s
, t 0.
In particular (H
M)
T
/
2
M
, hence, /
2
M
is stopping stable, that is, N /
2
M
implies
N
T
/
2
M
for all stopping times T.
(ii)
(H
M
T
)
t
= (H
M)
T
t
, t 0.
Proof. (i) Let N /
2
. Then
E
_
(H
M)
T
2.3.18(i)
= E
_
(H
M)
T
, N
2.4.33
= E [H
M, N
T
]
2.4.31(i)
= E
__

0
H
s
1
]0,T]
(s) dM, N
s
_
.
Hence, the assertion follows by 2.4.31(ii).
(ii) Since M
T
/
2
, the proof of (ii) is by 2.4.31(i) analogous to (i)
42
M in /
2
2.4.1. Orthogonality in /
2
Denition 2.4.36. Let M, N /
2
.
(i) M, N are called weakly orthogonal if E[M
](= (M, N)) = 0 (i.e. orthogonal in

(/
2
, ||)).
(ii) M, N are called strongly orthogonal, denoted by M N, if M, N = 0, P-a.s..
Remark 2.4.37. (i) M N
2.2.15(i)
M N is a martingale E[M
T
N
T
] = 0 for all bounded
stopping times T. But M
t
t
M
and N
T
t
N
in L
2
(P). Hence if M N, then
E[M
] = 0,
i.e. M, N are weakly orthogonal.
(ii) M N
2.4.31(i)
/
2
M
/
2
N
(since by 2.4.31 dH
M,

H
N = H

H dM, N).
(iii) Since (/
2
, ||) is a Hilbert space and /
2
M
is a closed linear subspace of /
2
(see 2.3.26),
for all N /
2
there exists an H L
2
(
, T
M
, P
M
) and L /
2
such that
N = H
M
. .
/
2
M
+ L
..
/
2
and L is weakly orthogonal to /
2
M
.
But in (iii) we have more because of the following proposition.
Proposition 2.4.38. Let M, L /
2
, L weakly orthogonal to /
2
M
. Then L /
2
M
(since /
2
M
is stopping stable). Hence, L is weakly orthogonal to /
2
M
if and only if L /
2
M
( L M).
Proof. Because of 2.4.37(i), it remains to show that E[L
T
M
T
] = 0 for all bounded stopping
times T. But by 2.4.35 (stopping stability) we have M
T
/
2
M
. Hence,
0 = E[L
M
T
] = E[L
M
T
]
L mart.
= E[L
T
M
T
].
Corollary 2.4.39 (Kunita-Watanabe-decomposition). (i) Let M, N /
2
. Then there exist
a unique L /
2
and a unique H L
2
(
, T
M
, P
M
) such that L /
2
M
and N = H
M+L.
(ii) Suppose L
2
(, T
, P) separable (e.g. true, if T
is countably generated). Then there exist

M
i
/
2
, i N, such that
M
i
M
j
, i ,= j,
and for all N /
2
there exist H
i
L
2
(
, T, P
M
), i N, (uniquely determined by M
i
)
such that
N =
i=1
H
i
M
i
, (the series converges in | |
/
2!)
that is,
/
2
=
i=1
/
2
M
i
and
/
2
M
i
/
2
M
j
, for i ,= j.
43
Proof. (i) 2.4.37(iv) and 2.4.38.
(ii) By assumption and 2.4.37(iii) (/
2
, ||) is separable. So, we can apply (i) and the Gram-
Schmidt orthogonalization procedure (cf. [Wei87],[RS80]).
Corollary 2.4.40. Let F L
2
(, T
, P), M /
2
. Then there exist uniquely determined
H L
2
(
, T
M
, P
M
) and L /
2
, L M, such that,
F = E[F[T
0
] +
_

0
H
s
dM
s
+L
.
Proof. Let F
t
:= E[F[T
t
] E[F[T
0
]. Since F
t
is a martingale, by 2.1.1 there exists a càdlàg
version (

F
t
)
t0
of (F
t
)
t0
(i.e. P[F
t
=

F
t
] = 1 t 0). Then (

F
t
) /
2
. Hence, by 2.4.39(i)
there exist unique L /
2
, L M, and H L
2
(
, T
M
, P
M
) such that P-a.s.
F
t
= (H
M)
t
+L
t
, t 0.
In particular, (since

F
= F E[F[T
0
],) we have P-a.s.
F = E[F[T
0
] +
_

0
H
s
dM
s
+L
.
44
2.5. Itos Representation Theorem
Let (, T, P) be a probability space with ltration (T
t
) and let (W
t
)
t0
be a (real-valued)
Wiener process on (, T, (T
t
), P), that is, (W
t
) is a continuous (T
t
)-adapted process such that,
for all s t, W
t
W
s
is independent of T
s
and N(0, t s) distributed. Let for t 0
T
W
t
:= (W
s
[0 s t)
and T
W
t+
, t 0, the corresponding right-continuous ltration.
Remark 2.5.41. Compared with our denition of Brownian motion for a Wiener process the
increment W
t
W
s
is independent of the larger -algebra T
s
( T
W
s
). In this space of (T
W
t
)-
martingales belonging to L
2
(, T, P) the Kunita-Watanabe decomposition (cf. 4.9(i) with M =
W) has a particularly simple form:
Theorem 2.5.42 (Itos representation theorem). Let M = (M
t
)
t0
L
2
(, T, P) be a right-
continuous martingale with respect to (T
W
t
+
), hence with respect to
(T
W
t+
)
P
:= (T
W
t+
, (N T[P(N) = 0)), t 0.
Then
M
t
= M
0
+
_
t
0
H
s
dW
s
, t 0, P-a.s., (2.5.2)
where H is (T
t
)-adapted and H 1
]0,T]
L
2
(
,

T, P
W
) for all T > 0 (and through this it is
uniquely determined). In particular, M has P-a.s. continuous sample paths and /
2
= /
2
W
(where /
2
is dened as in the previous section with respect to T
t
= (T
W
t
+
)
P
, t 0, which by the
backward martingale convergence theorem is again right-continuous) and
/
2
W
:=
_
__
t
0
H
s
dW
s
_
t0
H 1
]0,t]
L
2
(
,

T, P
W
) t 0, sup
t0
_
_
_
_
_
t
0
H
s
dW
s
_
_
_
_
/
2
(P)
<
_
.
Proof. Without loss of generality M
0
= 0. Fix k N and set
T
W
k+
:= (T
W
k+
)
P
, M
(k)
t
:= M
tk
, W
(k+1)
t
:= W
t(k+1)
, t 0.
Let (M
n
)
nN
L
2
(, T
W
k+
, P), such that |M
n
|
< and
lim
n
M
n
= M
k
= M
(k)
in L
2
(, T
W
k+
, P).
Let M
n
t
:= E[M
n
[T
W
t+
], t 0. Then by the backward martingale convergence theorem (see
[Roc11]) (M
n
t
)
t0
is right continuous. Then for each n N
sup
t,
[M
n
t
()[ < .
In particular, M
n
/
2
for all n. Hence, by the Kunita-Watanabe decomposition (cf. 2.4.39(i))
there exist adapted H
n
L
2
(
,

T, P
W
) and N
n
/
2
, N
n
/
2
W
(k+1)
, such that
M
n
t
=
_
H
n
W
(k+1)
_
t
+N
n
t
, t 0, n N.
We would like to prove that N
n
= 0 for all n N.
Step 1: Claim: Let N /
2
, N W
(k+1)
. If N is bounded, then N
t
= 0 for all t < k + 1.
45
Proof of Claim. Let c := sup
t,
[N
t
()[ < . Dene
D := 1 +
N
2c
.
Then D
1
2
and
E[D] = 1 +
1
2c
E[N
] = 1 +
1
2c
E[N
0
] = 1.
Thus,

P := DP is a probability measure on (, T) equivalent to P.
Claim: (W
t
)
tk+1
is a Brownian motion under

P.
Suppose the claim is true. Then the nite dimensional distributions of (W
t
)
tk+1
under P and
P are the same. Hence, by Radon-Nikodym,

P =

P on T
W
k+1
.
Therefore, P =

P on T
W
t+
for all t < k + 1. Thus, E[D[T
W
t+
] = 1 for all t < k + 1. Hence,
N
t
= E[N
[T
W
t+
] = 0, t < k + 1.
Proof of Claim.
(i) (W
(k+1)
t
) is a martingale under

P, because for all bounded stopping times T we have
E
P
[W
(k+1)
T
] = E
P
[W
(k+1)
T
D] = E
P
[W
(k+1)
T
]
. .
=0
+
1
2c
E
P
[W
(k+1)
T
N
]
. .
=:A
.
But by 2.4.37 W
k+1
T
N
T
is a martingale, since W
k+1
N. Hence,
A = E
P
[W
k+1
T
N
T
] = E
P
[W
(k+1)
0
N
0
] = 0,
and thus
E
P
[W
(k+1)
T
] = 0.
(ii) For P-a.e. we have
W
t
() = t t 0.
Hence, for

P-a.e.
W
t
() = t t 0.
Thus,

P-a.s.
W
(k+1)
t
= t (k + 1) t 0.
Therefore, by (i), (ii) and Levys characterization (cf. Proposition 1.5.34) (W
t
)
tk+1
is a Brow-
nian motion under

P = DP (up to k + 1).
Step 2: Now x n
0
N and set N := N
n
0
. Dene for n N
T
n
:= inf
_
t > 0
_
t
0
H
n
0
s
dW
(k+1)
s
> n
_
(k + 1).
Then T
n
are (T
W
t+
)-stopping times such that T
n
(k + 1) as n . We know that N
T
n
(=
N
T
n
) /
2
W
(k+1)
(cf. 2.4.33, 2.4.37(ii)).
46
Furthermore,
[N
T
n
t
[ = [N
tT
n
[
M
n
0
tT
n
_
tT
n
0
H
n
0
s
dW
(k+1)
s
|M
n
0
|
+n.
Hence,
sup
t,
[N
T
n
t
()[ |M
n
0
|
+n.
So, we can apply Step 1 to conclude that N
T
n
t
= 0, for all t < k + 1 and for all n N. Letting
n , we get N
t
= 0 for all t < k +1 P-a.s. (since N is right-continuous and the zero set does
not depend on t). Therefore, N
n
= 0 for all n and t < k + 1. But for all n, m N
E
P
W
(k+1)
_
(H
n
H
m
)
2
_
= E
__
k+1
0
(H
n
s
H
m
s
)
2
ds
_
= E
P
_
_
(H
n
W
(k+1)
)
k+1
(H
m
W
(k+1)
)
k+1
_
2
_
= E
_
_
M
n
k+1
M
m
k+1
_
2
_
n,m
0.
This is true, since P-a.e.
M
n
k+1
= E
_
M
n
[T
W
(k+1)+
_
=
n
E
_
M
k
[T
W
(k+1)+
_
= M
k
in L
2
(, T, P).
Therefore, (H
n
)
nN
is a Cauchy sequence in L
2
(
,

T, P
W
(k+1) ).
Hence, there exists H L
2
(
,

T, P
W
(k+1) ) such that H
n
H in L
2
(
,

T, P
W
(k+1) ). In particular,
we can assume H to be adapted. Furthermore, for 0 t k + 1,
M
t
= lim
n
M
n
t
= lim
n
(H
n
.W
(k+1)
)
t
= (H
W
(k+1)
)
t
=
_
t
0
H
s
dW
(k+1)
s
P-a.s..
Since k N was arbitrary, the decomposition (2.5.2) follows, since M is right-continuous and
the right hand side of (2.5.2) is continuous, M is P-a.s. continuous.
We now include an independent direct proof that every M /
2
(with underlying ltration
(T
W
t+
)
P
, t 0) is P.a.s. continuous:
Proposition 2.5.43. Every (T
W
t+
)-adapted (M
t
) /
2
loc
is P-a.s. continuous.
Proof. Without loss of generality M /
2
(localization). It suces to consider the case
M
t
:= E[F[T
W
t+
], t 0,
where
F :=
n
i=1
f
i
(W
t
i
), f
i
(
b
(R) uniformly continuous, 0 t
1
< . . . t
n
< ()
and to prove that (M
t
) has a continuous modication. This is enough because M
t
= E[M
[T
W
t+
]
(since M /
2
) and F of type () are dense in L
2
(, T
W
, P).
(Exercise, by a monotone class argument: By 2.1.4 M has a continuous (T
W
t+
)
P
-adapted version.
Dene
V :=
_
G L
2
(, T
W
, P)[F
n
of type () such that F
n
G in T
W
, P)
_
and prove that V = L
(, T
W
, P).)
47
Let t 0 and i 1, . . . , n such that t [t
i
, t
i+1
[ where t
n+1
:= +. Then (cf. section of
Markov property of Brownian motion),
M
t
= E[F[T
W
t+
]
=
ji
f
j
(W
t
j
)E
_
j>i
f
j
(W
t
j
)
T
W
t+
_
=
ji
f
j
(W
t
j
)E
W
t
_
j>i
f
j
(W
t
j
t
)
_
=
ji
f
j
(W
t
j
)p
t
i+1
t
_
f
i+2
(p
t
i+1
t
i+2
f
i+2
. . . p
t
n
t
n1
f
n
) . . .
_
. .
=:gC
0
(R)
(
t
) P-a.s.,
where
p
s
f(x) :=
1
2s
_
f(y)e
1
2s
(xy)
2
dy.
So, we have to prove that p
t
i+1
t
g(
t
) is P-a.s. continuous in t. Set
(
b,u
(R) := (
b
(R)[ uniformly continuous.
(i) Let f (
b,u
(R). Take x, y R. Then
[p
t
f(x) p
t
f(y)[
1
2s
_
[f(x +z) f(y +z)[e
z
2
2s
dz.
Let > 0. Then there exists = () > 0 such that [x y[ < [f(x) f(y)[ < . So,
[p
t
f(x) p
t
f(y)[ and p
s
f C
b,u
(R).
(ii) Let f (
b,u
(R). Then
lim
s0
|p
s
f f|
= 0.
Proof. Let > 0 and > 0 such that [x y[ < [f(x) f(y)[ < . Since
f(x) =
1
2s
_
f(x)e
y
2
2s
dy
we have
[p
s
f(x) f(x)[
2s
_
[f(x +y) f(y)[e
y
2
2s
dy
2s
_

[f(x +y) f(x)[

. .
<
e
y
2
2s
dy +
1
2s
_
[y[>
[f(x +y) f(x)[e
y
2
2s
dy
+ 2 |f|
2s
_
[y[>
e
y
2
2s
dy.
Therefore,
limsup
s0
|p
s
f f|
+ 2 |f|
limsup
s0
1
2s
_
[y[>
e
y
2
2s
dy
y
/
=
y
s
= .
48
(iii)
p
t
(p
s
f) = p
t+s
f,
in short,
p
t
p
s
= p
t+s
.
Proof. Exercise (by Fourier transform and by use of
p
s
f(x) = (
s
f)(x),
where
s
(z) =
1
2s
e
z
2
2s
and recall
s
() = e
s
2
2
).
(iv) For t < t
i+1
we have
lim
st
_
_
p
t
i+1
s
g p
t
i+1
t
g
_
_
= 0 (right-continuity).
Proof. Let h > 0 such that s := t +h < t
i+1
. Then
_
_
p
t
i+1
th
g p
t
i+1
t
g
_
_
(iii)
=
_
_
p
t
i+1
th
(g p
h
g)
_
_
|p
t
i
+1th
|
|g p
h
g|
p
s
1=1
|g p
h
g|
st
0 by (ii).
(v) Assume t
i
< t and let h > 0 such that s := t h > t
i
. Then
lim
s,t
_
_
p
t
i+1
s
g p
t
i+1
t
g
_
_
= 0.
Proof. Let h > 0 such that s := t h > t
i
. Then
_
_
p
t
i+1
t+h
g p
t
i+1
t
g
_
_
(iii)
=
_
_
p
t
i+1
t
(p
h
g g)
_
_
|p
h
g g|
h0
0.
(vi) Consider t := t
i
:
Let h > 0
p
t
i
(t
i
h)
(f
i
p
t
i+1
t
i
f
i+1
. . . p
t
n
t
n1
f
n
) = p
h
(f
i
p
t
i+1
t
i
f
i+1
. . . p
t
n
t
n1
f
n
)
. .
= g(
b,u
(R)
h0
g
uniformly in x. Hence, also continuous in t = t
i
from the left uniformly in x.
Summary: The function G dened by
G
t
(x) :=
ji
f
j
(x)p
t
i+1
t
(f
i+1
p
t
i+2
t
i+1
f
i+2
. . . p
t
n
t
n1
f
n
(x) . . .) for t [t
i
, t
i+1
[
is continuous in t 0 uniformly in x.
(vii) t G(t,
t
) is continuous P-a.s..
49
Proof. Take [G(t,
t
) G(s,
s
)[ [G(t,
t
) G(s,
t
)[ +[G(s,
t
) G(s,
s
)[. Fix s > 0.
Then the right hand side is dominated by
lim
ts
|G(t, ) G(s, )|
. .
=0
+lim
ts
|G(s,
t
) G(s,
s
)|
. .
=0
,
because t
t
is continuous and G is continuous in x.
Corollary 2.5.44 (cf. 2.5.42). Let F L
2
(, T
W
t
0
+
, P), t
0
> 0 xed. Such F are sometimes
called Wiener functional. Then there exists an H L
2
(
,

T, P
W
(t
0
+1) ), (T
W
t+
)-adapted, such
that
F E[F] =
_
t
0
0
H
s
dW
s
.
Proof. (cf. 2.4.40) Let
F
W
t
:= E[F[T
W
t+
] E[F[T
W
0+
], 0 t t
0
.
Then (F
W
t
)
1tt
0
is a martingale, which by the backward martingale convergence theorem (cf.
[Roc11]) is right continous. Hence, by 2.5.42,
F E[F[T
W
0+
] = E[F[T
W
t
0
+
] E[F[T
W
0+
] = F
W
t
0
2.5.42
=
_
t
0
0
H
s
dW
s
.
The uniqueness is also clear by martingale property.
Example 2.5.45 (Special case: the canonical Model). Let (, T, P) be the (classical) Wiener
space, i.e. = (([0, 1], R), T its Borel -algebra and P the Wiener measure. Dene X
t
() :=
(t), t [0, 1], , hence X is Brownian motion,
T
t
:=
s>t
((X
r
[r s 1), P-zero sets in T),
and
T =
_
0<t1
T
t
.
Then, by 2.5.44
F L
2
(, T, P) F = E[F] +
_
1
0
H
s
dX
s
.
Example 2.5.46. Consider the situation of 2.5.45 the Wiener functional
F :=
_
1
0
X
t
dt.
Because of E[F] =
_
1
0
E(X
t
)
. .
=0
dt = 0,
F =
_
1
0
H
s
dX
s
50
for some H L
2
(
,

T, P
W
).
Identication of H:
Let
M
t
:= E[F[T
t
] =
_
t
0
E[X
s
[T
t
] ds +
_
1
t
E[X
s
[T
t
] ds =
_
t
0
X
s
ds +X
t
(1 t)
. .
continuous!
t 0,
and the latter is clearly in /
2
c
.
Claim: H
t
= 1 t.
Proof. In view of 2.4.31(i) it is enough to show
M, X =
_
t
0
H
s
dX
s

_
t
0
(1 s) ds,
since then, by /
2
2.5.42
= /
2
X
, we have M = H
X, in particular,
_
1
0
X
t
dt = F = M
1
=
_
1
0
(1 s) dX
s
.
But by Itos product rule
X
t
t =
_
t
0
X
s
ds +
_
t
0
s dX
s
.
Hence
dM, X
t
= dX
_

0
s dX
s
, X
t
= dt t dt,
thus M, X
t
=
t
_
0
(1 s) ds.
51
3. Markov Processes
3.1. Markov Processes and Semigroups
Denition 3.1.1. A family (p
t
)
t0
of kernels on a measurable space (S, o) is called a semigroup
of kernels if
p
t
p
s
= p
t+s
t, s 0 (Chapman-Kolmogorov equation),
i.e.
p
t
p
s
(x, A) =
_
p
t
(x, dy)p
s
(y, A)
_
:=
_
p
s
(y, A)p
t
(x, dy)
_
.
If for all t, x p
t
(x, S) = 1, (p
t
)
t0
is called Markovian and, if p
t
(x, S) 1 for all x, t, sub-
Markovian, respectively.
Denition 3.1.2. Let (S, o) be a measurable space. A family of stochastic processes
(, T, (X
t
)
t0
, (P
x
)
xS
)
with state space (S, o) is called a Markov process, if
(M1) x P
x
() is o-measurable for all T,
(M2) there exists a ltration (T
t
)
t0
such that each X
t
is T
t
-measurable and
P
x
(X
s+t
B[T
s
) = P
X
s
(X
t
B) P
x
-a.e. s, t 0, B o, x S
(Markov property with respect to (T
t
)
t0
).
Remark 3.1.3. If (M2) is true for (

T
t
)
t0
with

T
t
T
t
t 0, then (M2) is true for (T
t
)
t0
if (X
t
) is (T
t
)-adapted!
The following theorem shows, at least if (S, o) is polish, that Markovian semigroups and Markov
processes are in one-to-one correspondence to each other.
Theorem 3.1.4. (i) Let (S, o) be a measurable space and let M = (, T, (X
t
)
t0
, (P
x
)
xS
)
be a family of stochastic processes with state space (S, o) and T = (X
t
[ t 0).
a) Suppose there exists a Markovian semigroup of kernels (p
t
)
t0
, such that for all 0
t
0
< t
1
< . . . < t
n
, f : S
n+1
R, bounded and o
n+1
measurable, and all x S,
E
x
[f(X
t
0
, . . . , X
t
n
)] =
_
S
p
t
0
(x, dx
0
) . . .
_
p
t
n
t
n1
(x
n1
, dx
n
)f(x
0
, . . . , x
n
) (3.1.1)
Then M is a Markov process with respect to T
t
:= (X
s
[s t), t 0 (and even
(3.1.3) below holds).
b) Suppose M is a Markov process and set
p
t
f(x) := E
x
(f(X
t
)), x S, f : S R bounded , o-measurable, t 0. (3.1.2)
Then (p
t
)
t0
is a Markovian semigroup of kernels on (S, o) and we have (3.1.1).
53
3. Markov Processes
(ii) If (S, o) is polish and (p
t
)
t0
is a Markovian semigroup of kernels on (S, o). Then there
exists a Markov process M with (3.1.2) and = S
[0,]
.
Proof. (i): a) (M1) follows by monotone classes from (3.1.1), since T = (X
t
[t 0).
For (M2) we show an a-priori stronger fact:
E
x
[f(X
t
1
+s
, . . . , X
t
n
+s
), ] = E
x
[E
X
s
[f(X
t
1
, . . . , X
t
m
)], ] (3.1.3)
for all x S, bounded and o
n
-measurable f : S
n
R, 0 s
0
< s
1
< . . . < s
m
= s
and T
s
, s 0. By monotone classes (applied to ) this follows from
E
x
[f(X
t
1
+s
, . . . , X
t
n
+s
)g(X
s
0
, . . . , X
s
m
)]
=E
x
[E
X
s
[f(X
t
1
, . . . , X
t
n
)]g(X
s
0
, . . . , X
s
m
)]
for all bounded o
m+1
-measurable g : S
n+1
R, 0 s
0
< s
1
< . . . < s
m
= s. By
(3.1.1) the left hand side is equal to
_
p
s
0
(x, dx
0
) . . .
_
p
s
m
s
m1
(x
m1
, dx
m
)
_
p
t
1
(x
m
, dy
1
) . . .
_
p
t
n
t
n1
(y
n1
, dy
n
)f(y
1
, . . . , y
n
)
. .
=E
x
m
[f(X
t
1
,...,X
t
n
)]
g(x
0
, x
1
, . . . , x
m
)
= E
x
[E
X
s
m
[f(X
t
1
. . . X
t
n
)]g(X
s
0
, X
s
1
, . . . , X
s
m
)].
b) By (M1) it is clear that (3.1.2) denes a Markovian kernel for all t 0. Furthermore,
p
t+s
f(x) = E
x
[f(X
t+s
)] = E
x
[E
x
[f(X
t+s
)[T
s
]]
M.P.
= E
x
[E
X
s
[f(X
t
)]] = E
x
[p
t
f(X
s
)]
= p
s
(p
t
f)(x) = (p
s
p
t
)f(x)
To prove (3.1.1) let f = f
0
f
1
. . . f
n
, f
i
: S R, bounded and o-measurable.
Then for 0 t
0
< t
1
< < t
n
E
x
[f
0
(X
t
0
)f
1
(X
t
1
) . . . f
n
(X
t
n
)]
=E
x
[f
0
(X
t
0
)f
1
(X
t
1
) . . . f
n1
(X
t
n1
)E
x
[f
n
(X
(t
n
t
n1
)+t
n1
)[T
t
n1
]]
M.P.
= E
x
[f
0
(X
t
0
)f
1
(X
t
1
) . . . f
n1
(X
t
n1
)E
X
t
n1
[f
n
(X
t
n
t
n1
)
)]]
(3.1.2)
= E
x
[f
0
(X
t
0
)f
1
(X
t
1
) . . . f
n1
(X
t
n1
)p
t
n
t
n1
f
n
(X
t
n1
)]
Ind.Hyp.
=
_
. . .
_
p
t
0
(x, dx
0
)p
t
1
t
0
(x
0
, dx
1
) p
t
n1
t
n2
(x
n2
, dx
n1
)
f
0
(x
0
)f
1
(x
1
) . . . f
n1
(x
n1
)p
t
n
t
n1
f
n
(x
n1
)
=
_
. . .
_
p
t
0
(x, dx
0
)p
t
1
t
0
(x
0
, dx
1
) p
t
n1
t
n2
(x
n2
, dx
n1
)p
t
n
t
n1
(x
n1
, dx
n
)
f
0
(x
0
)f
1
(x
1
) . . . f
n1
(x
n1
)f
n
(x
n
).
By monotone classes this equality extends to all f : S
n
R, bounded, measurable.
Hence, (3.1.1) holds.
(ii): see Bauer!
54
3.1. Markov Processes and Semigroups
Now consider the canonical model:
Let S be a topological space, o the Borel--algebra and S
[0,[
(e.g. = S
[0,[
or =
(([0, [, S) or are all bounded continuous paths in S). Dene
X
t
() := (t), t 0, ,
T := (X
t
[t 0),
T
0
t
:= (X
s
[s t) (past)
T
0
t
:= (X
s
[s t) (future).
Let M = (, T, (T
0
t
)
t0
), (X
t
)
t0
, (P
x
)
xS
. .
all information here!
) be a Markov process. Then M is called the
canonical model.
Denition 3.1.5. In the canonical model dene the shift operator
t
: for t 0 by
t
()(s) := (s +t),
i.e.
t
() = ( +t).
It is obvious that
t
: is

T
0
t
/T-measurable and moreover, (exercise)
1
t
(T) =

T
0
t
t 0.
Lemma 3.1.6. (i) is

T
0
t
-measurable if and only if there exists an T-measurable such
that
=
t
.
(ii) Suppose P
x
, x S, are given and that
M:= (, T, (T
0
t
)
t0
, (X
t
)
t0
, (P
x
)
xS
)
satises (M1). Then (M2) holds if and only if
E
x
[
t
[T
0
t
] = E
X
t
[] P
x
-a.s. x S, : R bounded and T-measurable. (3.1.4)
Proof. (i): Exercise (Factorization lemma).
(ii): is clear. (Consider = 1
X
s
B
, B o.)
: By 3.1.4 (i)(b) we know that (3.1.1) holds. But in the proof of 3.1.4(i) we have shown
that (3.1.1) implies (3.1.3). Now Lemma 3.1.6 follows by monotone classes.
It is clear that (3.1.4) implies (elementary Markov property)
E
x
[
t
[T
0
t
] = E
x
[
t
[(X
t
)] t 0. (3.1.5)
Interpretation of (3.1.5): The conditional expectation of a future observable, given the past, only
depends on the present at time t. The equivalent formulations of (3.1.5) in the following lemma
have corresponding interpretations:
Lemma 3.1.7. Fix x S. Then the following statements are equivalent to (3.1.5):
55
3. Markov Processes
(i)
E
x
[
0
t
[
T
0
t
] = E
x
[
0
t
[(X
t
)] P-a.s., t 0,
0
t
: R bounded and T
0
t
-measurable,
(ii)
E
x
[
0
t

0
t
[(X
t
)] = E
x
[
0
t
[(X
t
)]E
x
[
0
t
[(X
t
)] P
x
-a.s.,
for all t 0 and for all
0
t
,
0
t
: R, bounded,
0
t
T
0
t
-measurable and
0
t

T
0
t
-
measurable.
Proof. (3.1.5) (i):
E
x
[
0
t

0
t
] =E
x
[
0
t
E
x
[
0
t
[T
0
t
]]
3.1.6(i),(3.1.5)
= E
x
[
0
t
E
x
[
0
t
[(X
t
)]]
=E
x
[E
x
[
0
t
[(X
t
)]E
x
[
0
t
[(X
t
)]]
=E
x
[
0
t
E
x
[
0
t
[(X
t
)]] bounded

T
0
t
-measurable
0
t
: R.
Therefore, E
x
[
0
t
[(X
t
)] is a P
x
-version of E
x
[
0
t
[
T
0
t
]. Hence, (i) holds.
(i) (ii):
Let f : S R be bounded and o-measurable and
0
t
: R be bounded,

T
0
t
-measurable.
Then
E
x
[
0
t

0
t
f(X
t
)] = E
x
[E
x
[
0
t
[
T
0
t
]
0
t
f(X
t
)]
(i)
= E
x
[E
x
[
0
t
[(X
t
)]
0
t
f(X
t
)]
= E
x
[E
x
[
0
t
[(X
t
)]E
x
[
0
t
[(X
t
)]f(X
t
)].
(ii) (3.1.5):
E
x
[
0
t

0
t
]
(ii)
= E
x
[E
x
[
0
t
[(X
t
)]E[
0
t
[(X
t
)]]
= E
x
[
0
t
E
x
[
0
t
[(X
t
)]] bounded T
0
t
-measurable
0
t
: R.
Hence,
E
x
[
0
t
[(X
t
)] = E
x
[
0
t
[T
0
t
],
and (3.1.5) follows by Lemma 3.1.6(i).
Remark 3.1.8. Consider the situation of 3.1.4(i)(b). Then P
x
(X
0
= x) = 1 if and only if
p
0
(x, ) =
x
(Dirac-measure in x). If this holds for every x S, then M is called normal. In
this case, p
t
(x, A) = P
x
(X
t
A) is the probability to be at time t in A starting in x (transition
probability).
3.2. The Strong Markov Property
Consider the canonical model. ( only needed to have
t
, can be done abstractly!)
Recall: Let (T
t
) be some ltration and T an (T
t
)-stopping time. Dene the -eld of the T-past
by
T
T
:= A T[A T t T
t
t 0.
Denition 3.2.9. Let M = (, T, (X
t
)
t0
, (P
x
)
xS
, (
t
)
t0
) be a canonical Markov process.
Then M satises the strong Markov property (SMP), if there exists a right-continuous ltration
(T
t
)
t0
on (, T) (i.e. T
t
=
>0
T
t+
) such that for all (T
t
)-stopping times T we have that
TT<
is T T < /T- and T
T+s
T < /T
s
-measurable for all s 0 and we have
(SMP)
E
x
[1
T<

T
[T
T
] = 1
T<
E
X
T
[] P-a.s. bounded, T-measurable and x S.
56
3.2. The Strong Markov Property
Remark 3.2.10. (i) It is clear that (SMP) implies the Markov property with respect to (T
t
)
and therefore, with respect to (T
0
t
) by 3.1.6(ii). In general, the converse is false.
(ii) (SMP) holds if and only if for all (T
t
)-stopping times T we have
E
x
[
T
; T < ] = E
x
[E
X
T
[]; T < ] P
x
-a.s. (3.2.6)
for all : R, bounded and T-measurable, and for all x S.
Proof. is clear.
: Let A T
T
and

T := T 1
A
+ 1
A
c. Then

T is a (T
t
)-stopping time because
t 0
T t = A T t T
t
.
Hence
E
x
[
T
; A T < ] = E
x
[
T
;

T < ]
= E
x
[
T
;

T < ]
(3.2.6)
= E
x
[E
X
T
();

T < ]
= E
x
[E
X
T
(); A T < ].
Proposition 3.2.11. Suppose S is a topological space and o Borel -algebra such that o =
((
b
(S)). Let M= (, T, (X
t
)
t0
, (P
x
)
xS
, (
t
)
t0
) be a (canonical) Markov process with
(i) right-continuous paths,
(ii) p
t
((
b
(S)) C
b
(S) (Feller property),
where p
t
is the corresponding semigroup. Then M fullls the (SMP) with respect to T
t
:=
>0
T
0
t+
, t 0 (right-continuous).
Proof. Let T be an (T
t
)-stopping time. By [Roc11] X
T
is T
T
-measurable since X has right-
continuous sample paths. Similarly, one can prove that
T
has the required measurability pro-
perties. Dene
T
n
:=
k=1
k
2
n
1
k1
2
n
T<
k
2
n
+ 1
T=
.
Then T
n
T. We have to show that (3.2.6) holds:
Without loss of generality (otherwise use monotone classes)
= f
0
(X
t
0
)f
1
(X
t
1
) . . . f
n
(X
t
n
), f
i
(
b
(S), t
0
t
1
. . . t
n
.
Furthermore,
T < =

_
kN
_
k 1
2
n
T <
k
2
n
_
(3.2.7)
57
3. Markov Processes
Then
E
x
[
T
, T < ]
(i)
= lim
n
E
x
[
T
n
, T < ]
(3.2.7)
= lim
n
k=1
E
x
_

T
n
..
=
k2
n
,
k 1
2
n
T <
k
2
n
_
M.P.
= lim
n
k=1
E
x
_
E
X
k2
n
[];
k 1
2
n
T <
k
2
n
_
(3.2.7)
= lim
n
E
x
[E
X
T
n
[], T < ]
= E
x
[E
X
T
[]; T < ] (Lebesgue),
since X
T
n
X
T
on [T < ] and
x E
x
[] = E
x
[f
0
(X
t
0
), . . . , f
n
(X
t
n
)] = p
t
0
(f
0
p
t
1
t
0
f
1
p
t
2
t
1
f
2
. . . p
t
n
t
n1
f
n
)(x)
is continuous on S by (ii).
Remark 3.2.12 (Blumenthals 0 1 law). Let M be a canonical normal Markov process with
respect to T
t
:=
s>t
T
0
s
, t 0 (on measurable state space (S, o)). Then P
x
= 0 or P
x
= 1 on
T
0
for every x o.
Proof. Let be T
0
-measurable and bounded. Then for all x S we have
= E
x
[[T
0
] = E
x
[
0
[T
0
]
M.P.
= E
X
0
[] = E
x
[] P
x
-a.s..
Hence, is constant P
x
-a.e..
3.3. Application to Brownian Motion
Let S = R
d
, o = B(R
d
) and
p
t
(x, A) =
_
A
1
(2t)
d
2
e
yx
2
2t
. .
=:g
t
(yx)
dy, x R
d
, t > 0, A B(R
d
),
p
0
(x, ) =
x
.
Lemma 3.3.13. For all t, x 0, we have
p
t
p
s
= p
t+s
.
Proof. Exercise (by Fourier transform).
Let := (([0, [, R
d
), X
t
() := (t), P
0
:= P be the Wiener measure on , T := (X
t
[t 0)
and P
x
the image measure of P
0
under +x. Then we have (cf. [Roc11])
E
x
[f(X
t
0
, . . . , X
t
n
)] =
_
p
t
0
(x, dx
0
)p
t
1
t
0
p
t
n
t
n1
(x
n1
, dx
n
)f(x
0
, . . . , x
n
),
for all x R
d
, for all B((R
d
)
n+1
)-measurable, bounded functions f : (R
d
)
n+1
R and for
0 t
0
t
1
. . . t
n
. Hence,
M:= (, T, (X
t
)
t0
, (P
x
)
xR
d)
58
is by Proposition 3.1.4(i)(a) a Markov Process on (R
d
, B(R
d
)). M has continuous sample paths
and is normal. In particular, Remark 3.2.12 is fullled. But also, by Proposition 3.2.11 M has
SMP with respect to
T
t
=
s>t
T
0
s
, t 0,
because of the following Proposition.
Proposition 3.3.14. (p
t
)
t0
is strong Feller, that is,
p
t
f (
b
(R
d
) f B
b
(R
d
).
(Moreover, we have p
t
f (
for all t > 0 and f B

b
(R
d
).)
Proof. Fix t > 0. Let x R
d
and x
n
x. We have to show that p
t
f(x
n
) p
t
f(x). Let n
0
such that x
n
B
1
(x) for all n n
0
. There exists h L
1
(R) := L
1
(R, dx) (with dx =Lebesgue
measure) such that e
(x
/
y)
2
2t
h(y) for all y, x
t
R such that [x
t
x[ 1.
Dene g( x
1
, . . . , x
d
) = h( x
1
) h( x
d
), x = ( x
1
, . . . , x
d
) R
d
, then
e
[x
/
y[
2
2t
g(y) x
t
B
1
(x) y R
d
.
Hence
lim
n
p
t
f(x
n
) =
1
_
(2t)
d
lim
n
_
f(y)e
yx
n
2
2t
dy =
1
_
(2t)
d
_
f(y)e
yx
2
2t
dy = p
t
f(x)
by Lebesgues dominated convergence theorem since f is bounded and
e
x
/
y
2
2t
g(y) and
g L
1
(R) for all n n
0
.
Corollary 3.3.15. (i) Let M be a normal (canonical) Markov-Process with respect to T
t
:=
s>t
T
0
s
, t 0. Then, by Blumenthal for t not xed,
P
x
[limsup
t0
1
X
t
=x
= 1
. .
T
0
] 0, 1.
For Brownian motion this probability is equal to 1 because of the law of iterated logarithm.
(ii) Let X be Brownian motion starting at x = 0 and let
N() := 0 s 1[ X
s
() = 0.
Then for P
0
-a.e. we have that
a) N() is closed,
b) N() has Lebesgue-measure zero,
c) N() has no isolated points,
that is N() is a Cantor set for P
0
-a.e. , i.e. every point of N() is a cluster point of
N().
(iii) Let U R
d
, U open and bounded and let X be Brownian motion on R
d
. Dene the rst
exit time of U
T := inft > 0[X
t
, U =
U
c.
T is an (T
t
)-stopping time. Then there exists > 0 such that
sup
xU
E
x
[e
T
] < .
59
3. Markov Processes
Proof. (i) Clear.
(ii) a) X is P-a.s. continuous.
b) By Fubini, we get
_ _
1
0
1
N()
(s) dsP
0
( d) = E
0
__
1
0
1
0
X
s
ds
_
=
_
1
0
P
0
X
1
s
(0) ds
=
_
1
0
p
s
(0, 0) ds = 0.
c) (Intuitively clear: by law of (local) iterated logarithm t = 0 is not an isolated point
of N(). By (SMP) this is true for any s N()). Dene
J :=[N() has isolated points
=
_
r,sQ, r<s
[N()]r, s[ contains exactly one (isolated) point
. .
A
r,s
.
Let T
0
:= inft > 0[X
t
= 0. Then T
0
is an (T
t
)
_
=
s>t
T
0
s
_
-stopping time
(exercise). Therefore, for T := T
0

r
+ r (rst hitting time of 0 after time r,
(T
t
)-stopping time!)
P
0
[A
r,s
] P
0
[r +T
0

r
< s, T
0

r+T
{0}
r
> 0]
= E
0
[1
T
{0}
>0

T
; T < s]
= E
0
[E
0
[1
T
{0}
>0

T
[T
T
]; T < s]
SPM
= E
0
[E
X
T
[1
T
{0}
>0
], T < s]
X
T
=0
= E
0
[P
0
[T
0
> 0]
. .
=0
; T < s] = 0,
by (i).
(iii) Let R > 0 such that U B
R
(closed ball with radius R around 0). Dene
T
R
:= inft > 0[ [X
t
[ > R =
B
c
R
.
By Itos formula (or Doob-Meyer), [X
t
[
2
d t is a martingale under P
x
. Hence,
0 [x[
2
= E
x
[[X
0
[
2
d 0] = E
x
[[X
T
R
n
[
2
d T
R
n]
= E
x
[[X
T
R
n
[
2
] d E
x
[T
R
n] R
2
d E
x
[T
R
n], x U,
Therefore, taking n to , we get
E
x
[T] E
x
[T
R
]
R
2
d
< , x U.
Hence, for large t
0
sup
xU
P
x
[T > t
0
] sup
xU
1
t
0
E
x
[T] < 1.
Dene
(t) := sup
xU
P
x
[T > t] (decreasing in t!)
60
Then, we have
(t +s) =sup
xU
E
x
[1
T>t+s
]
=sup
xU
E
x
[1
T>s
1
T>t

s
]
=sup
xU
E
x
[E
x
[1
T>t

s
[T
s
]; T > s] (since T > s T
s
)
MP
= sup
xU
E
x
[P
X
s
[T > t]; T > s]
sup
xU
P
x
[T > t] sup
xU
P
x
[T > s] = (t)(s).
Claim: These two conditions, i.e.
a) there exists a t
0
> 0 such that (t
0
) < 1,
b)
(t +s) (t)(s), ()
imply that is subexponential, i.e. there exist K > 0 and > 0 such that
(t) Ke
t
t t
0
.
Proof. We have
(s) =
_
s
t
0
t
0
_
is decreasing

__
s
t
0
_
t
0
_
Hence, inequality () implies
(s) (t
0
+. . . +t
0
. .
]
s
t
0
|times
) (t
0
) . . . (t
0
)
. .
]
s
t
0
|times
= (t
0
)
]
s
t
0
|
.
From now on, without loss of generality we may assume (t
0
) > 0. Then, since
_
s
t
0
_

s
t
0
1, it follows that
(t
0
)
]
s
t
0
|
= e
]
s
t
0
| ln (t
0
)
exp
_
s
[ ln (t
0
)[
t
0
_
exp([ ln (t
0
)[) = Ke
s
,
where
K := exp([ ln (t
0
)[) and :=
[ ln (t
0
)[
t
0
.
Then for all x U and any > 0
E
x
[e
T
] = E
x
_
_
e
T
0
1 ds
_
= E
x
__

0
1
[0,e
T
[
(s) ds
_
Fubini
=
_

0
P
x
[e
T
> s] ds
=
_
e
t
0
0
P
x
[e
T
> s] ds +
_

e
t
0
P
x
[e
T
> s] ds.
61
3. Markov Processes
For the rst term we have easily
_
e
t
0
0
P
x
[e
T
> s] ds e
t
0
< .
But, we can estimate the second:
_

e
t
0
P
x
[e
T
> s] ds
_

e
t
0
sup
xU
P
x
[e
T
> s] ds
=
_

t
0
e
u
sup
xU
P
x
[e
T
> e
u
]
. .
=P
x
[T>u]
du
=
_

t
0
e
u
(u)du
K
_

t
0
e
u()
du < ,
if < .
3.4. Sojourn Time
Let S be a polish space, o the Borel--algebra and M = (, T, (X
t
)
t0
, (P
x
)
xS
) a normal
Markov process with respect to a right-continuous ltration (T
t
)
t0
with state space S and
continuous paths.
Denition 3.4.16. Fix x S. The time
x
:= inft > 0[X
t
,= x =
x
c
is called sojourn time.
Clearly, by (right) continuity of the sample paths
x
= 0 on X
0
= y, y ,= x.
So,
x
is only non-trivial under P
x
.
Remark 3.4.17.
x
is an (T
t
)
t0
-stopping time because for all t 0
x
t = X
s
= x[0 s t, s Q T
t
,
hence
x
< t T
t
. Since T
t
= T
t+
, also
x
t =
kN
x
< t +
1
k
. .
T
t
+
1
k
T
t+
= T
t
.
Proposition 3.4.18.
P
x
[
x
> t] = exp(c
x
t), t 0,
where
c
x
:= ln P
x
[
x
> 1] [0, ],
i.e.
x
is exponentially distributed with constant c
x
(cf. [KL99]).
62
3.4. Sojourn Time
In particular, if the Markov process started at x does not move for sure before time t = 1, it for
sure never moves. If it moves for sure before time t = 1, it for sure moves immediately.
Proof. (By Markov property.) Dene for t 0
f(t) := P
x
[
x
> t].
Claim: f(t +s) = f(t)f(s) for all t, s 0.
Indeed,
E
x
[1
x
>t+s
[T
s
] = E
x
[1
x
>t

s
[T
s
] 1
x
>s
M.P.
= E
X
s
[1
x
>t
]1
x
>s
X
s
=x on
x
>s
= P
x
[
x
> t] 1
x
>s
.
Hence,
f(t +s) = P
x
[
x
> t +s] = E
x
[E
x
[1
x
>t+s
[T
s
]]
= E
x
[P
x
[
x
> t]1
x
>s
]
= P
x
[
x
> t]P
x
[
x
> s] = f(t)f(s).
In particular,
f(1) = f
_
2
n
1
2
n
_
= f
_
1
2
n
_
2
n
and then
f
_
k
2
n
_
= f
_
1
2
n
_
k
= f(1)
k
2
n
.
Hence, by the right-continuity of f we have f(t) = f(1)
t
= e
t ln f(1)
.
Corollary 3.4.19. Let T
x
= inft > 0[X
t
= x =
x
. Let M be as in Proposition 3.4.18,
but assume in addition, that M has (SMP) with respect to (T
t
)
t0
(right-continuity!). Let y S
such that
P
y
[T
x
< ] > 0.
Then for all u 0, we have
P
y
[X
s
= x, s [T
x
, T
x
+u) ; T
x
< ] = e
c
x
u
P
y
[T
x
< ],
where c
x
is as in Proposition 3.4.18.
Proof. Let := 1
X
s
=x, s[0,u)
. Then
P
y
[X
s
= x, s [T
x
, T
x
+u); T
x
< ] = E
y
[
T
x
; T
x
< ]
= E
y
[E
y
[
T
x
[T
T
x
]; T
x
< ]
= E
y
[E
X
T
x
[]; T
x
< ]
= E
x
[] P
y
[T
x
< ]
= P
x
[X
s
= x, s [0, u) ] P
y
[T
x
< ]
3.4.18
= e
c
x
u
P
y
[T
x
< ].
63
4. Girsanov Transformation
4.1. Problem Outline (d = 1)
We want to construct a process such that it solves (in a weak sense) the following equation
(law of motion for the stochastic dynamics (X
t
)
t0
):
dX
t
= b(X
t
, t) dt + dW
t
,
X
0
= x
0
R
d
,
that is,
X
t
() = x
0
+
_
t
0
b(X
s
(), s) ds +W
t
().
Here, X
t
= x
0
+
_
t
0
b(X
s
, s) ds denotes the deterministic part and W
t
the stochastic perturbation,
i.e. W
t
is a Wiener process.
One possible strategy of solving this equation is to nd a strong solution, that is, for a given
Wiener process (W
t
)
t0
on a given probability space (, T, P) construct the paths (X
t
)
t0
of
the solution by classical methods (e.g. Picard-Lindelof or Euler scheme).
Example: The Ornstein-Uhlenbeck process has the law of motion
dX
t
= X
t
dt + dW
t
, > 0,
X
0
= x
0
R.
Claim: This problem has a strong solution
X
t
:= e
t
x
0
+
_
t
0
e
(ts)
dW
s
= F(x
0
, (W
s
)
st
)(t),
hence, X
t
is adapted to the Wiener ltration.
Proof. We apply Itos product rule to X
t
= e
t
_
x
0
+
_
t
0
e
s
dW
s
_
to get
X
t
= x
0
+
_
t
0
e
s
d
_
x
0
+
_
s
0
e
u
dW
u
_
+
_
t
0
_
x
0
+
_
s
0
e
u
dW
u
_
. .
=X
s
e
s
by denition
()e
s
ds
= x
0
_
t
0
X
s
ds +
_
t
0
1 dW
s
= x
0
_
t
0
X
s
ds +W
t
.
Instead of strong solutions one can construct (probabilistically) weak solutions. We want to
construct a Brownian motion (W
t
)
t0
and a process (X
t
)
t0
on some probability space (, T, P)
such that
X
t
= x
0
+
_
t
0
b(X
s
, s) ds +W
t
65
that is, construct (X
t
)
t0
on a suitable probability space (, T, P) such that
W
t
:= X
t
x
0
_
t
0
b(X
s
, s) ds
is a Brownian motion, e.g. take (, T), (X
t
)
t[0,t]
canonical, i.e.,
:= (([0, 1]),
X
t
() := (t),
T := (X
t
[t 0),
such that
W
t
() := X
t
() x
0
_
t
0
b(X
s
(), s) ds(= G
t
((X
s
()) s t)) (4.1.1)
is a Brownian motion under P.
But we have to identify P!
One technique to nd P is using the Girsanov transformation. This approach has the following
advantages:
One can do this even if the dependence on the past is very complicated (i.e. b(X
s
(), s)
replaced by b
s
).
One can do this for very irregular b
s
.
From now on for simplicity x
0
= X
0
= 0.
Method: For simplicity let x
0
= 0. Let = (([0, 1]), (X
t
)
t0
be the coordinate process, i.e.
X
t
() = (t) and P
0
be the Wiener measure on C([0, 1]) = . Then dene
P := exp
__
1
0
b
t
dX
t
1
2
_
1
0
b
2
t
dt
_
P
0
.
We can check that
W
t
:= X
t
_
t
0
b
s
ds
is a Brownian motion under P, where b
s
denotes the drift. ((X
t
)
t0
P
0
, hence, a martingale under P
0
, but not a martingale under P!)
Catch: We have to check, that P is a probability measure, i.e. we have to check, that
_
e
1
0
b
t
dX
t
1
2
1
0
b
2
t
dt
dP
0
= 1.
This is the hard work in applications.
Relation with the Transformation Rule for Lebesgue Measure
Dene T : (([0, 1]) (([0, 1]) by
T() := X()
. .
=
_

0
b(X
s
(), s) ds.
Then by Girsanov (under certain conditions)
P T
1
=
_
e
1
0
b
s
(X
s
,s) dX
s
1
2
1
0
b(X
s
,s)
2
ds
P
0
_
. .
=[ det DT[
//
T
1
= P
0
.
66
4.2. The General Girsanov Transformation
4.2. The General Girsanov Transformation
Let (, T, P) be a probability space and (T
t
)
t0
be a right-continuous ltration (not necessarily
completed). Then let

P be another probability measure such that
P
loc.
P (i.e.

P[
T
t
P[
T
t
t 0).
Then the Radon-Nikodym densities
Z
t
:=
d
P
dP
T
t
:=
d
P[
T
t
dP[
T
t
exist and (Z
t
)
t0
is a martingale (since for all t > s and F
s
T
s
_
F
s
Z
t
dP
F
s
T
s
T
t
=
_
F
s
d

P =
_
F
s
Z
s
dP,
i.e.
E[Z
t
[T
s
] = Z
s
).
Assumption (from now on in force): Z has continuous sample paths P-a.s.. (This is enough in
many applications. Note that by Itos representation theorem any martingale with respect to a
ltration generated by a Brownian motion has this property!)
Lemma 4.2.1. For every (T
t
)-stopping time T we have
P = Z
T
P on T
T
T < (trace -eld).
In particular,

P P on T
T
if T is nite.
Proof. Let A T
T
. Then, by a lemma in [Roc11] for all t 0
A T t T
Tt
T
t
,
since A T t T t s = A T s T
s
for all s 0. Hence,
P[A T t
. .
T
t
] =
_
ATt
Z
t
dP
ATtT
Tt
=
_
ATt
Z
Tt
dP =
_
ATt
Z
T
dP.
Letting t and applying monotone convergence the assertion follows.
The following lemma describes how martingales behave under change from P to

P. Dene
() := inft 0[Z
t
() = 0 (Then Z
= 0).
() is a stopping time. Recall that then Z
t
() = 0 for all t [(), [, since (Z
t
)
t0
is a
positive (super)martingale (cf. [Roc11]).
For simplicity assume > 0 P-a.s..
Lemma 4.2.2. (i) =

P-a.s. (not necessarily P-a.s.).
(ii) For all s t,
t
T
t
-measurable and positive we have
E
P
[
t
[T
s
] = 1
Z
s
,=0
Z
1
s
E
P
[
t
Z
t
[T
s
]

P a.s..
67
(iii) Let

M := (

M
t
)
t0
be an (T
t
)-adapted continuous process. Then

M is a local

P-martingale
(up to ), if

M Z is a local P-martingale (up to ).
Proof. (i) By Lemma 4.2.1 with T = t, we have
P[ t
. .
T
t
] = E
P
[Z
t
, t] = E
P
[ Z
..
=0
, t] = 0.
Letting t , the assertion follows.
(ii) For all
s
T
s
-measurable and positive,
E
P
[
s
t
] = E
P
[
s
1
Z
s
,=0
t
]
= E
P
[
s
1
Z
s
,=0
Z
t
t
]
= E
P
[
s
1
Z
s
,=0
E
P
[Z
t
t
[T
s
]]
= E
P
[
s
1
Z
s
,=0
Z
1
s
E
P
[
t
Z
t
[T
s
]]
(iii) The assertion directly follows from (ii). But it is also a consequence of Lemma 4.2.1: Let
T
1
T
2
. . . < be a localizing sequence for the local P-martingale

M Z (hence in
particular sup
n
T
n
= ). Then for all bounded stopping times T
E
P
[

M
TT
n
]
4.2.1
= E
P
[

M
TT
n
Z
TT
n
] = E
P
[

M
0
Z
0
] = E
P
[

M
0
].
Hence, (

M
T
n
t
)
t0
is a

P-martingale and the assertion follows, since =

P-a.s. by (i).
Proposition 4.2.3 (General Girsanov transform). Let M be a P-a.s. continuous local P-
martingale (up to ). Then
M := M
_

0
1
Z
s
dM, Z
s
is a continuous local

P-martingale (up to ).
Proof. By Lemma 4.2.2(iii) we have to show that

MZ is a local P-martingale up to .
We have that

MZ is P-a.s. continuous. Let
n
:= inft 0[Z
t

1
n
and T
n
, n N,

T
n
,
be stopping times forming a localizing sequence for the following three local P-martingales
M, MZ M, Z and
_

0
__
r
0
1
Z
s
dM, Z
s
_
dZ
r
.
Dene T
n
:=
n

T
n
. Note that
_
r
0
1
Z
s
dM, Z
s
is predictable, since it is P-a.s. continuous in r
and adapted. We have that 0 T
1
T
2
. . . T
n
< on > 0, since
n
< on > 0
because Z is P-a.s. continuous. Furthermore,
n
, because Z is P-a.s. continuous, hence
T
n
. Then by Itos product rule for all t 0
(

MZ)
tT
n
= (MZ)
tT
n
Z
tT
n
_
tT
n
0
1
Z
s
dM, Z
s
= (MZ)
tT
n

_
tT
n
0
Z
r
d
__
r
0
1
Z
s
dM, Z
s
_
tT
n
0
_
r
0
1
Z
s
dM, Z
s
dZ
r
= (MZ)
tT
n

_
tT
n
0
Z
r
1
Z
r
dM, Z
r

_
tT
n
0
_
r
0
1
Z
s
dM, Z
s
dZ
r
= (MZ)
tT
n
M, Z
tT
n
+M, Z
0
_
tT
n
0
_
r
0
1
Z
s
dM, Z
s
dZ
r
68
4.3. Girsanov Transform with Brownian Motion
Here,
_
tT
n
0
_
r
0
1
Z
s
dM, Z
s
dZ
r
is a P-martingale, and (MZ)
tT
n
M, Z
tT
n
is a P-martingale.
Therefore,

MZ is a local P-martingale up to .
Remark 4.2.4 (Z as an exponential martingale). We have by Ito formula for t < P-a.s.
log Z
t
= log Z
0
+
_
t
0
1
Z
s
dZ
s
. .
=:Y
t
1
2
_
t
0
1
Z
2
s
dZ
s
()
Y
t
is a local P-martingale up to . Since Z is a P-martingale up to and has pathwise quadratic
variation (cf. 5.3.9)
Y
t
=
_
t
0
1
Z
2
s
dZ
s
, t < .
Exponentiating () yields
Z
t
= Z
0
e
Y
t
1
2
Y )
t
(exponential P-martingale).
Then Z solves the following SDE for given Y up to :
dZ = Z dY Z
t
= Z
0
+
_
t
0
Z
s
dY
s
, (Z(0) = Z
0
)
Proof. By the 2-dimensional Itos formula we have
Z
t
= Z
0
+Z
0
_
t
0
e
Y
s
1
2
Y )
s
dY
s
1
2
Z
0
_
t
0
e
Y
s
1
2
Y )
s
dY
s
+
1
2
Z
0
_
t
0
e
Y
s
1
2
Y )
s
dY
s
= Z
0
+
_
t
0
Z
s
dY
s
.
Note that the last terms in Itos formula dont occur, since Y is of bounded variation.
Corollary 4.2.5. For

M as in Proposition 4.2.3 we have

M = M M, Y P-a.s. up to ,
hence

P-a.s. up to .
Proof.
_

0
1
Z
s
dM, Z
s
HN,M)=N,H
M)
=
_
M,
_
r
0
1
Z
s
dZ
s
_
= M, Y .
Hence, the assertion follows by denition of

M.
Let (X
t
)
t[0,1]
be a Brownian motion on a probability space (, T, P), adapted to a right-
continuous ltration (T
t
)
t[0,1]
. (Take for example (, T, P) to be the canonical Wiener space
(C[0, 1])
0
, T, P) with classical Wiener measure P.)
Heuristics: Let Y be a local continuous P-martingale and

P
loc
P with density
Z
t
:= e
Y
t
1
2
Y )
t
on T
t
.
Then we know that

M = X X, Y is a local continuous

P-martingale up to , where X
is a martingale with respect to P (as M above). Since

M
t
= X
t
= t, it follows by Levys
69
characterization theorem of Brownian motion that

M is a Brownian motion under

P. We want
to get that
dX, Y = b
t
dt.
We succeed by the following
Ansatz:
Y
t
:=
_
t
0
b
s
dX
s
.
Remark 4.3.6. In order to make this work, we need that
(i) (b
t
)
t[0,1]
is progressively measurable,
(ii)
P
__
1
0
b
2
s
ds <
_
= 1
(Then, Y is a continuous local P-martingale!) AND!
(iii)
P = e
1
0
b
s
dX
s
1
2
1
0
b
2
s
ds
P
is a probability measure, i.e.
E
P
_
e
1
0
b
s
dX
s
1
2
1
0
b
2
s
ds
_
= 1 (4.3.2)
Theorem 4.3.7 (Girsanov transform for M as a Brownian motion): Assume (i) - (iii) from
Remark 4.3.6. Let
Z
t
:= exp
__
t
0
b
s
dX
s
1
2
_
t
0
b
2
s
ds
_
, t [0, 1],
and
P := Z
1
P.
Then
W
t
:= X
t
_
t
0
b
s
ds, t [0, 1],

P.
Proof. Claim: (Z
t
)
t[0,1]
is a P-martingale up to 1.
Step 1: (Z
t
)
t[0,1]
is a (global) P-supermartingale.
Proof. It is clear, that (Z
t
) is a local continuous P-martingale. Let 0 T
1
. . . T
n
. . . < 1 be
a localizing sequence of stopping times for (Z
t
)
t[0,1]
. Then for 0 s < t 1
E
P
[Z
t
[T
s
] = E
P
[ lim
n
Z
tT
n
[T
s
]
Z
t
0
liminf
n
E
P
[Z
tT
n
[T
s
]
= liminf
n
Z
sT
n
= Z
s
.
Step 2: (Z
t
)
t[0,1]
is a P-martingale.
70
Proof. By (iii) for all s [0, 1]
1 = E[Z
1
] E[Z
s
] E[ Z
0
..
=1
] = 1.
In addition,
0 Z
s
E
P
[Z
1
[T
s
]
and
_
(Z
s
E
P
[Z
1
[T
s
]) dP = 0.
So, Z
s
= E
P
[Z
1
[T
s
] P-a.s.
We have
_
X,
_

0
b
s
dX
s
_
t
=
_
t
0
b
s
ds.
Hence, by Corollary 4.2.5 it follows that W is a continuous local

P-martingale. But W
t
=
X
t
= t P-a.s., hence

P-a.s.. So, by Levy (Proposition 1.5.34) the assertion follows.
Remark 4.3.8. (i) Special case:
b
t
() := b(X
t
(), t),
i.e. depending only on present time, where b : R [0, 1] R is B(R) B([0, 1]) measu-
rable. Then 4.3.6(i) is fullled.
4.3.6(ii) is fullled, if e.g.
E
__
1
0
b
2
t
dt
_
=
_
1
0
E[b
2
(X
t
, t)] dt
E[f(X
t
)]=p
t
f(0)
=
_
1
0
p
t
(b
2
(, t))(0) dt
p
t
(0,dx)=N(0,t)
=
_
1
0
_
R
b
2
(x, t)N(0, t)(dx) dt
=
_
1
0
_
R
b
2
(x, t)
1
2t
e
x
2
2t
dx dt < .
In particular, it is not neccessary that b is bounded.
To satisfy 4.3.6(iii) we have to work harder (see below)!
(ii) In Theorem 4.3.7 (X
t
)
t0
solves the following SDE under

P
dX
t
= dW
t
+b
t
dt.
Here, W
t

P, not under P.
Example 4.3.9. Consider b
t
= R xed for all t. Then, clearly, 4.3.6(i),(ii) hold, but also:
Claim: (iii) holds.
Proof. Since X
1
is normal distributed, we have
E
_
e
X
1
1
2
2
_
= e
1
2
2
_
e
x
N(0, 1)(dx)
= e
1
2
2
_
R
e
x
1
2
e
x
2
2
dx
= e
1
2
2
e
1
2
2
= 1.
71
Here, we have used that for the Laplace-transform L we have
L(N(0,
2
))() =
_
e
x
1
2
2
e
1
2
2
x
2
dx = e
1
2
2
.
Hence, by Theorem 4.3.7 under

P := Z
1
P the process W
t
= X
t
t, t 0, is a Brownian
motion.
Now we consider the canonical model for Brownian motion, i.e. = C([0, 1]), P as the Wiener
measure and X
t
() := (t), t 0. Dene for R
: C([0, 1]) C([0, 1])

by
()(t) = (t) + t, t [0, 1].

Note that by Theorem 4.3.7 we know (Z
1
P) W
1
is Wiener measure! Hence, since W =
and

P = Z
1
P
P
1
= P. (4.3.3)
Proposition 4.3.10.
P
1
= exp(X
1
1
2
2
) P.
Proof. Since by (4.3.3)
P
1
= P,
it follows that
P
1
=

P = exp(X
1
1
2
2
)P.
_
h(s)=s
= exp
__
1
0
h(s) dX
s
1
2
_
1
0
(
h(s))
2
ds
_
P
_
But we have even more:
Theorem 4.3.11 (Cameron-Martin). Let
h E := C([0, 1])
0
:= h C([0, 1])[h(0) = 0
and
X
h
: C([0, 1])
0
dened by
X
h
() = X() +h.
Let P
0
be the law of X on C([0, 1]) (i.e. the classical Wiener measure) and P
h
:= P
0
X
1
h
be
the law of X
h
on C([0, 1]). Then the following assertions are equivalent:
(i) P
h
P
0
.
(ii)
h H :=
_
h C([0, 1])
0
h is absolutely continuous and

h L
2
([0, 1], ds)
_
.
72
In this case
dP
h
dP
0
= exp
__
1
0
h(s) dX(s)
1
2
_
1
0
(
h(s))
2
ds
_
.
H is called Cameron-Martin space. H is a Hilbert space with inner product
h,
h
H
:=
_
1
0
h(s)
h(s) ds, h,
h H.
Note that |h|
2
H
= h, h
H
= 0 implies

h(s) = 0 ds-a.s.. Hence,
h(t) = h(0) +
_
t
0
h(s) ds = h(0) = 0 t 0.
So, ||
H
is not only a semi norm, but a norm. We have that H is dense in E := C([0, 1])
0
with
respect to ||
H
. Then identifying H with its dual H
t
by the Riesz map R we get
E
t
H
t
R
H E
continuously (see below). Before we prove Theorem 4.3.11, we want to characterize E
t
.
Lemma 4.3.12. For
M
0
:=
_
is a signed measure of nite total variation on [0, 1], such that (1) :=
_
1 d = 0
_
,
we have
E
t
= M
0
.
Remark 4.3.13. A signed measure can be written as =
1
2
with positive measures
1
and
2
and the total variation is dened as
sup
AB([0,1])
(A) = ||
Var
< .
Furthermore, for a signed measure , there exist positive measures
+
,
and S B([0, 1]) such

that =
+
and
+
(S
c
) = 0 and
(S) = 0.
Then
||
Var
=
+
([0, 1]) +
([0, 1]).
Proof of 4.3.12. : Dening f (f) :=
_
f d for M
0
we see that any denes an
element in E
t
. Furthermore, if , M
0
such that (f) = (f) f E, then for all f C([0, 1])
_
f d =
_
f d f(0)
_
1 d =
_
(f f(0))
. .
E
d
=
_
(f f(0)) d =
_
f d
_
f(0) d
. .
=0
=
_
f d.
Hence, (since (1) = (1) = 0) M
0
E
t
. (See also [Cho69a] and [Cho69b].)
: Let l E
t
. Then by the Hahn-Banach theorem there exists an

l (C[0, 1])
t
such that

l = l
on E( C([0, 1])). Hence, by Riesz-Markov (cf. [Roc04]) there exists a signed measure on [0, 1]
of nite total variation such that
l(f) =
_
f d f C([0, 1]).
73
Set := (1)
0
, where
0
is the Dirac measure at 0 [0, 1]. Then (1) = 0, so M
0
and
for all f E
_
f d =
_
f d (1) f(0)
..
=0
=
_
f d =

l(f) = l(f).
Proof of Theorem 4.3.11. Without loss of generality consider (, T, P), T
t
and (X
t
)
t0
as the
cannonical model, hence P = P
0
.
(ii) (i): (An alternative proof using Fourier transform can be found in [MR92, Chapter II].)
We rst check the conditions of Remark 4.3.6 (i)-(iii):
(i) and (ii) are obviously satised for
b
s
() :=

h(s) , s [0, 1].
(iii): Dene
Y
t
:=
_
1
0
h(s) dX
s
(Ito-Integral).
Then Y
1
is centered (i.e. E[Y
1
] = 0 for all t (martingale)) and normally distributed as an L
2
-limit
of centered, normally distributed random variables (cf. [Roc11]). Furthermore,
E[Y
2
1
](= E[Y
1
]) = E
__
1
0
(
h(s))
2
ds
_
= |h|
2
H
,
i.e.
Y
1
N(0, |h|
2
H
).
Hence,
E
_
e
Y
1
1
2
|h|
2
H
_
= e
1
2
|h|
2
H
E[e
Y
1
]
. .
=e
1
2
jhj
2
H
= 1
So, 4.3.6(iii) holds. Therefore, applying Girsanov (Theorem 4.3.7) we obtain that under
P := Z
1
P := exp
__
1
0
h(t) dX
t
1
2
_
1
0
h(t)
2
dt
_
P
W := Xh is a Brownian motion, i.e.

PW
1
= P
(0)
, i.e.

P = P
(0)
(W
1
)
1
= P
(0)
X
1
h
= P
h
.
But, since E[Y
2
1
] < , we have Z
1
> 0 P-a.s., therefore, P
h
P.
(i) (ii): We know that P
0
(E) = 1 with E := C([0, 1])
0
. Clearly, we have that H E dense with
respect to the norm ||
(cf. functional analysis).Let h H. Then (by denition of absolute

continuity)
h(t) = h(0)
..
=0
+
_
t
0
h(s) ds t [0, 1],

thus,
[h(t)[
_
t
0
[
h(s)[ ds
_
1
0
[
h(s)[ ds
__
1
0
[
h(s)[
2
ds
_
1
2
= |h|
H
, t [0, 1],
hence,
|h|
|h|
H
.
74
Hence H E continuously w.r.t. the norms ||
H
and ||
. Let R : H
t
H be the Riesz
isomorphism. Then
E
t
H
t
R
H E (continuously).
(Cf. Lemma 4.5.21 below: R() =
_
0
([0, s[) ds, E
t
.)
Claim 1: Let E
t
(= M
0
). The map C([0, 1])
0
() R is Gaussian distributed under
P
0
, more precisely
P
0

1
= N(0, E[
2
]
. .
()
2
P
0
(d)
).
Proof. Recall that = lim
n
N
n
i=1
(n)
i

t
(n)
i
. .
=:
n
,
(n)
i
R, t
(n)
i
[0, 1] weakly (cf. [Bau78]). But
n
: E R, n N, are jointly (centred) Gaussian under P
0
, because
N
l=1
(
l
n
l
)() =
N
l=1
l
N
n
l
i=1
(n
l
)
i

t
(n
l
)
i
() =
N
l=1
l
N
n
l
i=1
(n
l
)
i
(t
(n
l
)
i
)
. .
=X
t
(n
l
)
i
()
f E
i.e.
N
l=1
n
l
=
N
l=1
l
N
n
l
i=1
(n
l
)
i
X
(n
l
)
t
i
. .
(Rvalued) centred Gaussian
N(0, E(
N
i=1
n
l
)
2
)
Hence, since
n
n
weakly, i.e. pointwise on E, we have that
n
n
and is centered
Gaussian with variance E[
2
].
Claim 2: Let M
0
of the form = dt and bounded and
(1) =
_
1
0
dt = 0. ()
Then
E[
2
] =
_
1
0
__
t
0
(s) ds
_
2
dt = |R()|
2
H
. (4.3.4)
75
Proof.
E[
2
]
X
t
():=(t)
= E
__
1
0
X
t
(t) dt
_
1
0
X
t
/ (t
t
) dt
t
_
Fubini
=
_
1
0
_
1
0
(t)(t
t
)E[X
t
X
t
/ ] dt
t
dt
=
_
1
0
_
1
0
(t)(t
t
)(t t
t
) dt
t
dt = 2
_
1
0
_
1
0
t
1
t<t
/
(t)(t
t
) dt dt
t
= 2
_
1
0
t(t)
_
1
t
(t
t
) dt
t
dt
=
_
1
0
t
d
dt
_
__
1
t
(t
t
) dt
t
_
2
_
dt
I.b.p.
=
_
1
0
__
1
t
(t
t
) dt
t
_
2
dt
=
_
1
0
__
t
0
(s) ds
_
2
dt
=
_
1
0
_
d
dt
_
t
0
__
t
/
0
(s) ds
_
dt
t
. .
=:

R()(t)
_
2
dt =
_
1
0
_
d
dt
R()(t)
_
2
dt =
_
_
_
R()
_
_
_
2
H
.
(4.3.5)
For Claim 2 it remains to show that

R() = R(), i.e.
R() =
_

0
_
_
t
/
0
(s) ds
_
dt
t
. (4.3.6)
To this end let

h H. Then
(
h) =
_
1
0
h(t)(t) dt
I.b.p.
=
_
1
0
d
dt
h(t)
_
t
0
(t) ds dt =
_
h,
_

0
_
_
t
/
0
(s) ds
_
dt
t
_
H
(4.3.7)
and Claim 2 is proved.
Since C
1
0
(]0, 1[) is dense in L
2
([0, 1], dt), for all

h H there exists a sequence (v
n
)
nN
in C
1
0
(]0, 1[)
such that v
n

h as n in L
2
([0, 1], dt). Hence
u
n
:=
_

0
v
n
dt
n

h in H.
Since by (4.3.6)
u
n
= R(
n
) for
n
:= v
n
dt E
t
it follows that for
M
0
:=
_
dt
C([0, 1]),
_
1
0
dt = 0
_
that R(

M
0
) is dense in H with respect to ||
H
. R(

M
0
) is also a linear subspace of H.
Let h E(= C([0, 1])
0
) such that P
h
P.
Claim: Let
n

M
0
such that R(
n
)
n
0 in H. Then
n
(h)
n
0.
Suppose the claim is true. Then h H.
76
4.4. Novikov condition
Proof. By the claim the map R(

M
0
) R() (h) R is a linear continuous functional on
(R(

M
0
), ||
H
). Hence, by Riesz (note R(

M
0
) is dense in H, hence Riesz is applicable) there
exists an unique h
0
H such that
(h) = R(), h
0
H

M
0
.
But by 4.3.11 we also have R(), h
0
H
= (h
0
). Therefore,
(h) = (h
0
),

M
0
(4.3.8)
Hence, for all C([0, 1]) and for
h := h
_
1
0
h(t) dt,

h
0
:= h
0
_
1
0
h
0
(t) dt
we have
_

h dt =
_

h
_

_
1
0
ds
_
dt
. .

M
0
=
_
h
_

_
1
0
ds
_
dt
(4.3.8)
=
_
h
0
_

_
1
0
ds
_
dt =
_

h
0
_

_
1
0
ds
_
dt =
_
h
0
dt.
Hence,

h =

h
0
, therefore, h = h
0
, because h(0) = h
0
(0) = 0. Therefore, the assertion is true as
long as the claim holds.
Proof of Claim. Since R(
n
) 0 in H, it follows by (4.3.4) that
E
P
0
[
2
n
] = |R(
n
)|
2
H
n
0,
i.e.
n
0 in L
2
(P), hence, also in P
0
-measure, therefore, because of P
h
P
0
, also in P
h
-
measure. Since
n
, n N is a Gaussian family under P
h
, it follows by [Roc11] that
n
n
0 in L
p
(P
h
) p 1. (4.3.9)
Since we have
E
P
h
[
n
] =
_

n
( +h)
. .
n
()+
n
(h)
P(d) =
_

n
()P(d)
. .
=0
+
n
(h), (4.3.10)
applying (4.3.9) for p = 1, it follows that
limsup
n
[
n
(h)[
(4.3.10)
limsup
n
E
P
h
[[
n
[] = 0.
So, the claim is proved.
Let (, T, P) be a probability space together with a right-continuous complete ltration (T
t
)
and let (Y
t
)
t0
be a P-a.s. continuous local martingale such that Y
0
= 0. Then by 1.4.33
Z
t
:= exp(Y
t
1
2
Y
t
), t 0,
is a (P-a.s.) continuous local martingale. We want to develop a condition for (4.3.2).
77
Lemma 4.4.14. Let t 0 and Y
t
be bounded. Then Z
t
L
p
for all p > 1 and
E(Z
p
t
) exp(
1
2
p(p 1) |Y
t
|
).
Proof.
E(Z
p
t
) = E
_
exp
_
pY
t
1
2
p
2
Y
t
+
_
1
2
p
2
1
2
p
_
Y
t
__
exp
_
1
2
p(p 1) |Y
t
|
_
E
_
exp
_
pY
t
1
2
p
2
Y
t
__
.
Set

Y
t
:= pY
t
. Then
Y
t
= p
2
Y
t
and, therefore,
E(exp(
Y
t
1
2
Y
t
)) E(exp(
Y
0
1
2
Y
0
)) = 1,
since Z is a supermartingale (by Fatou).
Now we come to the condition for (4.3.2).
Theorem 4.4.15 (Novikov). If E(exp(
1
2
Y
t
)) < for all t, then E(Z
t
) = 1 for all t.
Remark 4.4.16. (i) In the examples above (see 4.3.8 and 4.3.9) we had Y
t
=
_
t
0
b
s
dX
s
and
E
_
exp
_
1
2
_
t
0
b
2
s
ds
__
<
was always satised.
(ii) For the proof of Theorem 4.4.15 we need that one can construct a Wiener process W, such
that
Y
t
= W
Y )
t
and (for every xed t) Y
t
is a stopping time with respect to a suitable ltration for which
W is adapted. This means that Y
t
has the form W
T
, with
Y
t
= T.
In other words, any P-a.s. continuous local martingale is the time change of a Brownian
motion. The details are presented in appendix A.
Now Theorem 4.4.15 follows from
Theorem 4.4.17. Let (W
t
)
t0
be a Wiener process on (, T, P) and T be a stopping time. If
E(exp(
1
2
T)) < ,
then the Wald identity holds:
E(exp(W
T

1
2
T)) = 1.
Remark 4.4.18. Set M
t
:= exp(W
t

1
2
t). Then (M
t
) is a continuous martingale, since it
is a continuous positive supermartingale and E[M
t
] = 1 (cf. Example 4.3.9). By the optional
sampling theorem for unbounded stopping times (cf. [Roc11]), we have
E(M
T
) E(M
0
) = 1.
Thus, it is clear that E(exp(W
T

1
2
T)) 1 in 4.4.17. For the proof of in 4.4.17 we will
need two Lemmas.
78
Lemma 4.4.19. Let

P be a probability measure on (, T, P) with
P[
T
t
= exp(W
t
1
2
t) P[
T
t
, t 0,
and T be a stopping time with P(T < ) = 1. Then
E(exp(W
T

1
2
T)) =

P(T < ). (4.4.11)
In particular, the Wald identity (cf. 4.4.17) holds if and only if
P(T < ) = 1.
Proof. Since M
t
:= exp(W
t
1
2
t) is a martingale and T t T
tT
we have
P[T t] = E[1
Tt
M
t
] = E[1
Tt
M
tT
] = E[1
Tt
M
T
].
Letting t we get (since P[T < ] = 1)
P[T < ] = E[M

T
].
Lemma 4.4.20. Let c > 0 and dene the passage time of (W
t
t) by
T
c
:= inft > 0[W
t
= t c.
_
W
T
c
= T c
_
Then
P(T
c
< ) = 1
and, thus, the Wald identity holds for T
c
. Furthermore,
E
_
exp
_
1
2
T
c
__
= e
c
.
Proof. By the law of iterated logarithm we have
P(T
c
< ) = 1.
W
t
= W
t
t is a Brownian motion with respect to

P (cf. Example 4.3.9 with = 1), which
means that T
c
is a passage time of (

W
t
) with respect to

P. Therefore, again by the law of iterated
logarithm
P(T
c
< ) = 1.
Thus, by 4.4.19
1 = E
_
exp
_
W
T
c

1
2
T
c
__
= e
c
E
_
exp
_
1
2
T
c
__
.
Proof of Theorem 4.4.17. It remains to show that holds:
M
t
:= exp(W
t
1
2
t), t 0, is a positive continuous supermartingale. Hence,
1 E(M
T
c
T
) E(M
T
c
)
4.4.20
= 1.
79
But then for all c > 0
1 = E(M
T
c
T
)
W
T
c
=T
c
c
= E
_
exp
_
1
2
T
c
_
exp(c), T
c
T
_
+E
_
exp
_
W
T

1
2
T
_
, T
c
> T
_
e
c
E(e
1
2
T
)
. .
<(Novikov)
+E
_
exp
_
W
T

1
2
T
__
c
E
_
exp
_
W
T

1
2
T
__
80
4.5. Integration by Parts on Wiener Space
4.5. Integration by Parts on Wiener Space: A First Introduction to
the Malliavin Calculus Following J.M. Bismut
Fix the following framework:
Let P C := P
0
be the Wiener measure on = C([0, 1])
0
(= E), (X
t
) the coordinate process,
H :=
_
h C([0, 1])
0
h is absolutely continuous and

_
1
0
h(s)
2
ds <
_
the Cameron-Martin space. H is a Hilbert space with inner product
h, g
H
=
_
1
0
h(s) g(s) ds =
h, g
L
2
([0,1],dt)
,
and F : R be the Wiener functional. We already know
E
t
= [ is a signed measure on [0, 1] of bounded variation such that (1) = 0
and
E
t
H
t
R
H E
continuous and densely, where R denotes the Riesz map.
Lemma 4.5.21.
R() =
_

0
(]s, 1]) ds
_
=
_

0
([0, s[) ds
_
.
(Cf. (4.3.6) as a special case.)
Proof. Let h H. Then
(h) =
_
1
0
h(t)(dt) =
_
1
0
_
t
0
h(s) ds (dt) =
_
1
0
_
1
0
1
[0,t[
(s)
h(s) ds(dt)
Fubini
=
_
1
0
_
1
0
1
]s,1]
(t)(dt)
h(s) ds =
_
1
0
(]s, 1])
h(s) ds
=
_
1
0
d
ds
__
s
0
(]r, 1]) dr
_
h(s) ds =
__

0
(]s, 1]) ds, h
_
H
Hence R() =
_
0
(]s, 1]) ds.
We recall
Denition 4.5.22. F : C([0, 1])
0
R is called Frechet-dierentiable in C([0, 1])
0
, if there
exists F
t
() E
t
such that
F( +) = F() +F
t
()() +o(||), C([0, 1])
0
.
In this case
F() := R(F
t
()) H
is called gradient of F at . Note that F() is in the tangent space of H in .
81
Remark 4.5.23. Because of Lemma 4.5.21 we have for the measure
F
t
()(dt) =: F
t
(, dt)
that
F()() = R(F
t
(, dt)) =
_

0
F
t
(w, ]s, 1]) ds. (4.5.12)
Then by denition of the derivative
lim
0
F( +) F()
= F(),
H
(4.5.12)
=
_
1
0
F
t
(, ]s, 1]) (s) ds H( E := C([0, 1])
0
).
Denition 4.5.24. F L
2
(P) is called H-dierentiable, if for all (T
t
)-adapted real processes
(u
s
)
s[0,1]
, product-measurable, bounded and
U
t
() :=
_
t
0
u
s
() ds, t [0, 1] (H vector eld on E)
(i.e. U() H) there exists a T/B(H)-measurable map F : E H such that
E(|F|
2
H
) <
_
F L
2
( H, P)
_
and
F( +U()) +F()
0
F(), U()
H
in L
2
(P)
or equivalently
U
F := lim
0
F(X +U) +F(X)
= F, U
H
in L
2
(P).
Here, F is called the Malliavin gradient (cf. [Wat84]).
Dene the Malliavin derivative
D
t
F := (F())
(t)
(4.5.12)
= F
t
(, ]s, 1]).
In particular, (D
t
F)
0t1
is a process. (Fact is that this process always has a version that is
product-measurable in (, t)!)
Geometric interpretation of F : H-vector eld on E.
Geometric interpretation of DF : L
2
([0, 1], dt)-vector eld on E.
Remark 4.5.25. Let u and U as in Denition 4.5.24.
(i) We have
F(), U()
H
= D
F(), u
L
2
([0,1],dt)
(ii) Let
Z
t
:= exp
_
_
t
0
u
s
dX
s
1
2
2
_
t
0
u
2
s
ds
_
.
Then Novikovs condition is fullled, since u is bounded. So, Girsanovs theorem implies
that X
:= X U is a Wiener process under P
:= Z
1
P. Hence,
E
P
[F(X
)Z
1
] =
_
F(X
) dP
(X
)
1
=P
=
_
F(X) dP = E
P
[F(X)] (4.5.13)
82
Lemma 4.5.26.
lim
0
Z
1
1
=
_
1
0
u
s
dX
s
in L
2
(P).
In fact this limit exists even in L
p
(P) for all p 1 (exercise).
Proof. By Ito
Z
t
= 1 +
_
t
0
Z
s
u
s
dX
s
. ()
Hence, for t = 1
Z
1
1

_
1
0
u
s
dX
s
=
_
1
0
(Z
s
1)u
s
dX
s
. (4.5.14)
But by Ito-isometry
E
_
__
1
0
(Z
s
1)u
s
dX
s
_
2
_
= E
__
1
0
(Z
s
1)
2
u
2
s
ds
_
|u|
2
E
__
1
0
(Z
s
1)
2
ds
_
.
Z
s
1 is a martingale, since Z
s
is in L
2
(P) (see below), thus, (Z
s
1)
2
is a submartingale.
Therefore,
E
_
__
1
0
(Z
s
1)u
s
dX
s
_
2
_
|u|
2
E
_
(Z
s
1)
2
_
It remains to prove that (Z
1
1) 0 L
2
(P). Clearly, (Z
1
1) 0 P-a.s.. But the set
(Z
1
1)
2
[0 1 is uniformally P-integrable, because
E[(Z
)
p
]
4.4.14
exp
_
1
2
2
p(p 1)
_
_
_
_
_
1
0
u
2
s
ds
_
_
_
_
_
exp
_
1
2
p(p 1)
2
|u|
2
_
,
i.e. Z
1
[0 1 is L
p
bounded for all p 2. Hence, the assertion follows by Lebesgues
dominated convergence theorem.
Proposition 4.5.27 (Bismuts integration by Parts formula on Wiener space). Let u, U be as
in Proposition 4.5.24 and let F : E R be H-dierentiable. Then
_
E
_
D
F, u
L
2
([0,1],dt)
_
= E[F, U
H
] = E
_
F
_
1
0
u
s
dX
s
. .
=div U=D
u
_
. ( u Dom D
)
Remark 4.5.28. Proposition 4.5.27 identies a duality between D and
_
dX (i.e. between the
Malliavin derivative and the Ito-integral). This is the starting point for dening an extension of
the Ito-integral, namely the Skorohod integral. This extension is simply dened as the adjoint
D
of D. Note that domD
contains also non-(T

t
)-adapted processes (cf. Lectures on Malliavin
calculus!) domD is the set of all H-dierentiable functions F : L
2
(E, P
0
),
D : domD L
2
(E, P) L
2
(E L
2
([0, 1], dt), P).
Hence,
D
: domD
L
2
(E L
2
([0, 1], dt), P) L
2
(E, P).
83
Proof of Proposition 4.5.27. (4.5.13) implies
E
(P)
_
F(X
)Z
_
= E
(P)
_
F(X)
_
.
Therefore, (recall X() = )
E
_
F(X
) F(X)
1
_
= E
_
F(X)
Z
1
1
_
0
E
_
F
_
1
0
u
s
dX
s
_
by Lemma 4.5.26. But the left hand side is equal to
E
__
1
(F(X
) F(X)) +F, U
H
_
Z
1
_
E
_
F, U
H
Z
1
_
.
Since F, U
H
L
2
(E, P) and Z
1
0
1 in L
2
(P) (see Proof of 4.5.26), the second summand
converges to E[F, U
H
] as 0. The rst summand converges to 0 by Cauchy-Schwarz as
0.
First application:
Identication of the integrand in Itos representation theorem.
Corollary 4.5.29 (Clark-Formula). Let F L
2
(P) H-dierentiable. Then
F = E[F] +
_
1
0
E[D
t
F[T
t
] dX
t
P-a.s..
Exercise: Show that (t, ) E[D, F[T
t
]() is

P
X
-a.e. equal to a B([0, 1])T-measurable func-
tion. Hint: First consider N
t
() = 1
[a,b[
(t)1
A
() for A T instead of D
t
F (Here,

P
M
(dt, d) =
dM
t
()P(d) = dt P(d)).
Proof of 4.5.29. Without loss of generality E[F] = 0. Let G L
2
(P). Then
G = E[G] +
_
1
0
u
t
dX
t
,
where u L
2
(
,

T,

P) and u adapted. Then by Bismuts Integration by Parts-formula we have
for
u
(n)
:= (u n) (n), n N,
that
E[FG] = lim
n
E
_
F
_
1
0
u
(n)
t
dX
t
_
4.5.27
= lim
n
E
__
1
0
D
t
F u
(n)
t
dt
_
Fubini
= lim
n
_
1
0
E[D
t
F u
(n)
t
] dt =
_
1
0
E[D
t
F u
t
] dt
u adapted
= E
__
1
0
E[D
t
F[T
t
]u
t
dt
_
2.4.31
= E
___
1
0
E[D
t
F[T
t
] dX
t
_
G
_
.
84
Example 4.5.30. (i) Dene
F :=
_
1
0
X
t
dt.
Note that E F() =
_
1
0
X
t
dt is linear and continuous in E, hence F E
t
. Then
E
t
F
t
(, dt) = dt
0
, hence by (4.5.12)
D
t
F() = F
t
(, ]t, 1]) = 1 t.
So, by 4.5.29
_
1
0
X
t
dt = F =
_
1
0
(1 t) dX
t
.
Note that this can be also (in fact much more easily) proved by It os product rule (see
2.5.46).
(ii) F := f(X
1
), where f (
2
(R). Then by the chain rule (which also holds for Frechet-
dierentiable functions on Banach spaces) F is Frechet-dierentiable and
F
t
(, dt) = f
t
(X
1
())
1
f
t
(X
1
())
0
( E
t
).
Thus,
D
t
F() = F
t
(, ]t, 1]) = f
t
(X
1
()).
Hence, by Corollary 4.5.29
f(X
1
) = F = E[f(X
1
)] +
_
1
0
E[f
t
(X
1
)[T
t
] dX
t
. ()
F depends only on time t = 1, but is represented as an integral along paths. The corre-
sponding integrands can be interpreted in terms of solutions h(X
t
(), t) to the following
nal value problem)
h(, 1) = f(),
1
2
h
xx
+h
t
= 0.
The solution has the form h(x, t) = p
1t
f(x) (as easy to see by elementary calculation, cf.
Chapter III). Hence by Itos formula (cf. 1.3.1(iii))
h(X
1
, 1) = h(0, 0) +
_
1
0
h
x
(X
s
, s) dX
s
, ()
where h(X
1
, 1) = f(X
1
) and h(0, 0) = p
1
f(0) = E[f(X
1
)]. By () and () and uniqueness
in Itos representation theorem we can conclude that
E[f
t
(X
1
)[T
s
] = h
x
(X
s
, s) =
x
p
1s
f(X
s
) = (
x
(E[f(X
1s
+x)]))
x=X
s
.
To show that F L
2
(P) is H-dierentiable is rather dicult in general. The following sucient
condition might be useful to check H-dierentiability.
Proposition 4.5.31. The following is a sucient condition for the H-dierentiability of F in
L
2
(P):
There exists a kernel F
t
(, dt) from to B([0, 1]) such that for all U as in 4.5.24
(i)
F(X +U)) F(X)
()
0
_
1
0
F
t
(, dt)U
t
() for P-a.e. E.
85
(ii) For all c > 0
[F(X +U) F(X)[ c |U|
P-a.s..
In this case
(H )F() = R(F
t
(, dt) F
t
(, [0, 1])
0
. .
=:
(dt)E
/
) =
_

0
F
t
(, ]s, 1]) ds.
Proof. By 4.5.21 we have
_
F
t
(, dt)U
t
()
U
0
=0
=
_

(dt)U
t
() = R(
), u()
H
= F(), U()
H
.
Hence, the assertion follows by Lebesgues dominated convergence theorem.
Example 4.5.32. F() := max
0t1
(t) is not Frechet-dierentiable, but dene
F
t
(, dt) :=
T()
(dt),
where
T := inft > 0[X
t
= F.
Exercise: Show that T = sup
n
T
n
, where T
n
:= inft > 0[X
t
> F
1
n
, t Q and that T
n
are stopping times for all n, hence, T is a stopping time, with respect to (T
t
+
), where T
t
:=
(X
s
[s t).
Then it follows by Proposition 4.5.31 that
D
t
F() = F
t
(, ]t, 1])(=
T()
(]t, 1])) = 1
T>t
(). (Exercise!)
Next step: Identify E[D
t
F[T
t
] in order to use the Clark formula.
Dene
M
t
:= max
0st
X
s
.
Then we have for P-a.e. E
E[D
t
F[T
t
]() = P[T > t[T
t
]()
= P
_
max
ts1
X
s
> M
t
T
t
_
()
= P
_
max
0s1t
X
s+t
> M
t
T
t
_
()
Now, we use the superstrong Markov property, i.e. for any TT
t
-measurable positive or bounded
function H : E E R we have
E
x
[H(
t
, )[T
t
]() = E
X
t
()
[H(, )] P
x
-a.s.,
where P
x
is the law of Brownian motion started at x R and
t
() := ( +t).
Hence,
E[D
t
F[T
t
]() = P
X
t
()
_
max
0s1t
X
s
> M
t
()
_
= P
X
t
()
_
max
0s1t
X
s
X
t
() > M
t
() X
t
()
_
86
By the reection principle, this is equal to
2 P
X
t
()
(X
1t
X
t
() > M
t
() X
t
())
2 P
0
(X
1t
> M
t
() X
t
())
=2 N(0, (1 t))(]M
t
() X
t
(), [)
=2 N(0, 1)
_
M
t
() X
t
()
1 t
, [
_
=2
_

M
t
()X
t
()
1t
1
2
e
x
2
2
dx
=2
_
1
_
M
t
() X
t
()
1 t
__
.
87
5. Brownian motion and potential theory
Attention: This chapter has not been proofread yet!
5.1. Brownian motion and Laplace Operator
Let (, T, (X
t
)
t0
, (P
x
)
xR
d) be the canonical Brownian Motion on R
d
, i.e.:
:= C([0, [, R
d
), X
t
() = (t), t 0, T := (X
t
[t 0), T
t
:=
>0
X
s
[s t +,
P
x
:= Pushforward measure of the Wiener measure P
0
under x +
Let (p
t
)
t>0
be the corresponding semi-group of kernels, i.e. for f : R
d
R
+
, B(R
d
)-measurable,
t > 0,
_
f(X
t
)dP
x
= E
x
[f(X
t
)] =
_
f(y)p
t
(x, dy) =
_
f(y)
1
(2t)
d
2
e
xy
2
2t
dy, x R
d
.
Let D R
d
, open.
T
D
:= exit time from D (:= inft > 0[X
t
/ D =
D
c).
x D the following holds for P
x
-a.e. and all t [0, T
D
()[ (cf. I.2.72):
f(X
t
()) f(X
0
())
. .
=f(x)
=
_
t
0
f(X
s
)dX
s
() +
1
2
_
t
0
f(X
s
())ds, if f C
2
(D)
In particular
M
f
t
() := f(X
t
() f(X
0
())
1
2
_
t
0
f(X
s
())ds
is Ito-Integral with respect to Brownian Motion with
M
f
() =
_
t
0
[[f(X
s
)[[
2
ds. ()
Corollary 5.1.1. (i) f C
2
(D) (M
f
t
) is local martingale (until T
D
) under P
x
x D
(cf. I.3.4.), i.e. Brownian Motion is solution for the (so called) martingale problem for
(L :=
1
2
, C
2
(D)) with initial condition x D (cf. [SV06]).
(ii) Preliminary remark: D
0
open rel. kp with

D
0
D E
x
[T
D
0
] < x D
0
(cf.
III.3.3.3). Hence, because of (), E
X
[M
f
T
D
0
] < x D
0
I.3.5.
= Dynkin formula: Let T be stopping time, T
D
0
(< T
D
as

D
0
D), then x D
0
E
x
[f(X
t
)] f(x) =
1
2
E
x
__
T
0
f(X
s
)ds
_
89
Application 5.1.2: (i) T = T
= exit time from open -ball K
(x) around x. Then there

exists for f C
2
(D)
af(x) := lim
0
E
x
[f(X
T
)] f(x)
E
x
[T
]
characteristic (Dynkin) operator (local),
and we have
af(x) =
1
2
f(x), a =
1
2
on C
2
(D) for Brownian Motion (cf. special script)
(ii) D = R
d
, f C
2
0
(R
d
):
Af(x) := lim
t0
p
t
f(x) f(x)
t
= lim
t0
E
x
[f(X
t
)] f(x)
t
=
1
2
f(x)
the innitesimale operator or generator of the semi-group.
A =
1
2
on C
0
(R
d
) for Brownian Motion
Attention: In what way
1
2
really generates (p
t
)
t>0
and with it the Brownian Motion, hence
in particular is uniquely determined, can not be seen from above (cf. special script or
[MR92]).
(iii) It follows from the Dynkin formula with T
r
= exit time from K
r
(x
0
) and f(x) = [[xx
0
[[
2
,
since
1
2
f = d, x K
r
(x
0
) that
E
x
[(X
T
r
x
0
)
2
. .
=r
2
] [[x x
0
[[
2
= dE
x
[T
R
]
= E
x
[T
r
] =
r
2
[[x x
0
[[
2
d
.
5.2. Stochastic solution of the Dirichlet- resp. Poisson-Problem
D R
d
, open, rel. kp.
T = exit time from D, then E
x
[T] < x D.
Possion problem
Have: f C(D), g C
b
(D)
Want: h C
2
(D)
C(

D) with
1
2
h = g on D
h = f on D
If g 0:
Dirichlet problem
Theorem 5.2.3 (Representation Theorem). Let h be a solution of the Poisson problem, then
x D h(x) = E
x
[f(X
T
D
)] E
x
__
T
D
0
g(X
s
)ds
_
90
5.2. Stochastic solution of the Dirichlet- resp. Poisson-Problem
Proof. Let D
n
D,

D
n
D, i.e. in particular

D
n
kp.
=
Dynkin formula
x D
n
n E
x
_
h(X
T
D
n
)
. .
,X
T
D
D
h(x) = E
x
__
T
D
n
0
g(X
s
)ds
_
T
D
n
,T
D
n
Lebesgue
E
x
[f(X
T
D
)] h(x) = E
x
__
T
D
0
g(X
s
)ds
_
We consider now the case g 0, i.e. a Dirichlet problem. For D
0
open, rel. kp in D we we dene
for x D
0
.
D
0
x
(A) := P
x
_
X
T
D
0
A
_
, A B(D
0
)
Then we get
Corollary 5.2.4 (generalized mean value property). Let h be harmonic in D. Then D
0
open, rel. kp in D
h(x) = E
x
_
h(X
T
D
0
)
_
=
_
D
0
h d
D
0
x
, x D
0
.
Denition 5.2.5.
D
0
x
is called the harmonic measure for D
0
in x ( D
0
). (=exit distrubition
P
x
X
1
T
D
0
of the Brownian motion from D
0
with start in x)
Remark 5.2.6. It follows from the deniton of the Brownian motion, that it is translation- and
rotation invariant. Thus,
K
r
(x)
x
is also rotation invariant.
=
K
r
(x)
x
= const
x,r
..
norm. surface
measure
=
K
r
(x)
x
(K
r
(x))=1
K
r
(x)
x
=
x,r
.
More generally, we have y K
r
(x)
K
r
(x)
y
(dz) = r
n2
r
2
[[y x[[
2
[[y z[[
n
. .
Poisson kernel for K
r
(x)
d
x,r
Theorem 5.2.7 (Existence of a generalized solution of the Dirichlet problem). D R
d
, open,
f bounded, measurable on D. Then we have for
h(x) := E
x
[f(X
T
D
)] , x D :
(i) h harmonic on D
(ii) lim
t,T
D
h(X
t
) = f(X
T
D
) P
x
-a.s x D.
Proof. Set T := T
D
and let D
n
D, D
n
open,

D
n
D
n+1
D. Set T
n
:= T
D
n
. x D xed.
Without loss of generality : x D
1
.
91
(i) Harmonicity: n
h(X
T
n
) = E
X
T
n
[f(X
T
)] =
SMP
E
x
[f (X
T

T
n
)
. .
=X
T
n
+T
n
. .
=T, since TT
n
[T
T
n
]
= E
x
[f(X
T
)[T
T
n
]
= (h(X
T
n
))
hN
P
x
-martingale and
h(x) = E
x
[f(X
T
)] = E
x
[E
x
[f(X
T
)[T
T
n
]] =
see above
E
x
[h(X
T
n
)]
= h harmonic. Since K
r
(x) K
r
(x) D, Without loss of generality : D
1
= K
r
(x),
hence h(x) = E
x
[h(X
T
1
)] =
see above
_
h(y)
x,r
(dy), i.e. h satises classic mean value property
=
Analysis
h harmonic.
(ii) (h(X
T
n
))
nN
bounded martingale under P
x
=
mart. conv.
theorem + above
h(X
T
n
) E
x
_
f(X
T
)[(
_
n
T
T
n
)
_
= f(X
T
) P
x
a.s., since X
T
= limX
T
n
(
n
T
T
n
)
measurable
(as T
n
T and continuity)
But according to the upcrossing lemma limh(X
t
) P
x
-a.s., (more detailled: Cor. 5.2, ch.
II, W-Theorie II-Vorlesung), hence it must be the above one.
(Existence: Set h(X
t
) = f(X
T
) for t T, then (h(X
t
))
t0
P
x
-martingale, since
1
t<T
h(X
t
) = 1
t<T
E
X
t
[f(X
T
)] =
MP
E
x
[1
t<T
f(X
T
t
+t
)
. .
=1
{t<T}
f(X
T
)
[T
t
]
= E
x
[f(X
T
)[T
t
] 1
tT
f(X
T
)
. .
T
tT
T
t
measurable
(cf. lemma 4.1 in ch. II, W-Th.II).)
5.3. Application: Reccurence/transience of the Brownian motion
We know (because of the iterated Log) that d: lim
t
[[X
t
[[
R
d = + P
x
-a.s. x R
d
. We want
to show:
Theorem 5.3.8. (i) d 3 = Brownian motion transient, i.e.: lim
t,
[[X
t
[[ = + P
x
-a.s.
x R
d
.
(ii) d = 2 = Brownian motion recurrent, i.e.: Brownian motion hits K
(x
0
) P
x
-a.s. > 0
x
0
R
d
; i.e. P
x
-a.e. path is dense in R
2
.
92
5.3. Application: Reccurence/transience of the Brownian motion
Preparation for the proof
Let A
r,R
:= x R
d
[r < [[x[[ < R and f =
_
1 in x [ [[x[[ = r
0 in x [ [[x[[ = R
be a solution to the
corresponding Dirichlet problem
h
r,R
(x) := E
x
_
f X
T
A
r,R
_
= P
x
_
[[X
T
A
r,R
[[ = r
_
.
One can easily show that analytically (write in spherical coordinates)
h
r,R
(x) =

d
([[x[[)
d
(R)
d
(r)
d
(R)
with
d
() =
_
_
, d = 1
log , d = 2
2d
, d 3
(5.3.1)
(cf. Ruinproblem W.theorie II)
Theorem 5.3.9. d 2 every point in R
d
is polar, i.e.:
P
y
[
x
< ] = 0 y R
d
.
(with
x
:= inft > 0[X
t
= x.)
Proof. Without loss of generality : x = 0 (because auf translation invariance)
(i) y ,= x = 0 :
P
y
[
0
< ] = lim
n
P
y
[
0
<
K
n
(0)
c]
K
n
(0)
c
n
P
y
a.s.
( since forP
y
a.a. X
s
()[0 s t bounded t)
lim
n
lim
r0
r<y
P
y
_
[[X
T
A
r,n
[[ = r
_
= lim
n
lim
r0
r<y
h
r,n
(y) = 0 because of (5.3.1), since d 2.
(ii) y = x :
P
x
[
x
< ] lim
t0
P
x
[
x

t
< ]
=
(MP)
lim
t0
E
x
[P
X
t
[
x
< ]
. .
1)
=0 P
x
a.s.,
since X
t
=x P
x
a.s.
]
Remark 5.3.10. For x / K
r
(0)
P
x
[
K
r
(0)
< ] = lim
n
P
x
[
,n,P
x
a.s.
..
[[X
T
A
r,n
[[ = r]
=
(5.3.1)
_
1, d = 2
[[x[[
2d
r
2d
, d 3.
93
Now:
Proof of Theorem 5.3.8. Let x
o
R
d
, Without loss of generality x
0
= 0. T
n
:= exit time from
K
n
(0), then x R
d
P
x
_
lim
t
[[X
t
[[ < r
_
. .
=
n
_
K
r
(0)

T
n
<
. .
n,
= lim
n
P
x
_
K
r
(0)

T
n
<
=
(SMP)
lim
n
E
x
_
P
X
T
n
[
K
r
(0)
<
=
5.3.10
lim
n
_
1 , d = 2
n
2d
r
2d

n
, d = 3.
5.4. Regularity of boundary points
Let D R
d
, open, rel. kp., D :=

D (R
d
D)
Denition 5.4.11. z D is called regular, if P
z
[
D
c = 0] = 1.
Remark 5.4.12. Since
D
c = 0 T
0
:=
>0
T
, we have acoordingly to Blumenthals zero-one

law a priori P
x
[
D
c = 0] 0, 1.
Example 5.4.13. Let D = U x with U rel. kp. open, z U. Then z D and z is
irregular, if d 2, since z polar. We consider the behaviour of the stochastic solution of the
Dirichlet problem for D with boundary condition f = 1
z
+ 0 1
(D)\z
h(x) = E
x
[f(X
T
D
)]
= 0 P
x
[X
T
D
D z] + 1 P
x
[X
T
D
z]
. .
=0,
if d2, since then z polar.
0
If we understand the above as an inteference of the solution of the Dirichlet problem for
D z, we then obtain stability.
Theorem 5.4.14. z D regular f bounded, measurable on D, continous in z we have:
h(x) := E
x
[f(X
T
D
)]
xz
xD
f(z)
Proof. We only proof : Let in the following always x D.
Claim: lim
xz
P
x
[X
T
D
/ K
(z)] = 0 > 0. Here K
(z) open ball.

If this claim is true, then select = (), such that [f(x) f(z)[ < x K
(z) D. It
follows that
[h(x) f(z)[ E
x
[[f(X
T
D
) f(z)[] + 2[[f[[
P
x
[X
T
D
/ K
(z)]
= lim
xz
[h(x) f(z)[ > 0, hence proved.
94
Proof of the claim. Let > 0 xed. If X
0
() = x K
2
(z), then s > 0:
X
T
D
()
/ K
(z) = T
D
() > s or T
K
2
(x)
s.
So we have x K
2
(z) s > 0
P
x
[X
T
D
/ K
(z)] P
x
[T
D
> s] +P
x
_
T
K
2
(x)
s
_
Because of the translation invariance we get
P
x
_
T
K
2
(x)
s
_
= P
0
_
T
K
2
(0)
s
_
_
_
_
_

s0
P
0
_
T
K
2
(0)
= 0
_
=
because of right
continous pathes
0
_
_
_
_
< for s small
Let s > 0. To show:
y
s
(y) := P
y
[T
D
> s], y R
d
, is not semi-continuous in y := z
Because then 0 lim
xz
s
(x)
s
(z)
z regular
= 0.
But y R
d
s
(y) = P
y
[T
D
> s] = lim
t0
P
y
[
, if t, hence lim=inf
..
t +T
D

t
. .
rst exit time after t
> s]
. .
= P
y
[T
D

t
> s t]
=
MP
E
y
[P
X
t
[T
D
> s t]
. .
st
(X
t
)
]
= lim
t0
p
t
st
(y)
. .
cont. as p
t
strong Feller
_
n.o.h.cont.
Now an analytic criterion for regularity:
Theorem 5.4.15 (Zarembas cone condition). Let z D such that there exixts a cone C
with apex in 0 and h > 0 with z +C
h
D
c
, where C
h
:= C K
n
(0). Then z is regular.
Proof.
P
z
[T
D
t] P
z
[X
t
z +C
h
] t ]0, 1]
= P
0
[X
t
C
h
]
= P
0
_
tX t
(
t)
2
C
h
_
, since
_
cX t
c
2
_
t0
again B.M.; cf. W.th.II, ch. III, Satz 5.1
= P
0
_
X
1
t
1
2
C
h
_
(t1)
P
0
[X
1
C
h
] > 0
95
= P
z
[T
D
= 0] = lim
t
n
0
P
z
[T
D
t
n
] P
0
[X
1
C
h
] > 0 n
=
Blumentahls
0-1 law
P
z
[T
D
= 0] = 1.
Idea: D
c
shouldnt be to small in z.
Exercice: for d = 2 C-line is enough.
In general (z +C
h
) H D
c
is enough, where H hypersurface (cf. [?], p. 276 or 350, analytical
with Wiener theorem).
Example 5.4.16 (Lebesgue-sting). Let
I
n
:= [2
n
, 2
(n1)
]
k
n
:= (x
1
, x
2
, x
3
) R
3
[x
1
I
n
, x
2
= x
3
= 0
F
n
:= (x
1
, x
2
, x
3
) R
3
[x
1
I
n
, x
2
2
+x
2
3

2
n
D = cylinder
(height z)
_
n
F
n
We want to choose
n
such that 0 D irregular.
d = 3 : = P
0
[X
0
never hit K
n
] = 1, since
P
0
[X
0
never goes back to x
2
= x
3
= 0] ( no condition for X
(1)
0
)
= P
d=2
0
[ B.M. in R
2
never goes back to 0] = 1 , 0 polar!
Since Brownian Motion in R
3
is transient for P
0
-a.a. an s() 0 with [[X
t
()[[ 2
t s(). It is obvious that
a
n
() := dist (X
t
()[0 t s(), K
n
) > 0
(as both sets kp. and disjoint). Hence a
n

1
m

m
P-a.s., thus
n
with P
0
[a
n

n
] 3
n
.
Then
P
0
[X
t
F
n
for a t
. .
F
n
<
] P
0
[a
n

n
] 3
n
= P
0
[T
D
= 0]
n
P
0
[X
t
F
n
for a t
. .
F
n
<
]
3
n
< 1
= 0 irregular.
Excursus to Brownian motion in R
1
. Without loss of generality a 0
Dene passage time T
a
:= inft > 0[X
t
> a.
Because X
T
a
= a and continuity of paths, we have:
T
a
= inft > 0[X
t
a = inft > 0[X
t
= a
P := P
0
= Brownian Motion with start in 0.
Theorem 5.4.17 (Reection principle).
1
2
P[T
a
t] = P[X
t
a]
_
=
1
2
P[[X
t
[ a]
_
96
Proof.
P[X
t
a] = P[T
a
t, X
tT
a
(
T
a
)
. .
=X
t
since
canonical model
a]
= E[(,
T
a
()
()); T
a
t],
where (, ) = 1
[a,[
X
tT
a
()
(), (, ) = C([0, [) C([0, [).
is T
T
a
T-measurable. We need:
Lemma 5.4.18 ((very) SMP). Let T be stopping time, : R
+
T
T
T-measurable.
Then it follows for H() := (,
T()
()) and all x R
d
E
x
[H[T
T
]() = E
X
T
()
[(, )] P
x
-a.a. .
Proof. Let F T
T
-measurable, G T-measurable and (, ) = F()G(). Then because of SMP
for P
x
-a.a.
E
x
[F G
T
[T
T
]() = F()E
x
[G
T
[T
t
]() = F()E
X
T()
[G] = E
X
T()
[F()G].
The rest follows with monoton classes.
Continuation of proof 5.4.17.
According to the lemma: for P-a.a.
E[(,
T
a
()
()); T
a
t]()
=
_
T
a
t
P
X
T
a
()
. .
a
[X
t T
a
()
. .
=:s
() a]
. .
=P
s
(a,[a,[)=
1
2
P(d) =
1
2
P[T
a
t].
Corollary 5.4.19 (Distribution of the passage time). P T
1
a
=
a
ds, where
a
(s) =
_
_
_
[a[
2s
3
e
a
2
2s
, s > 0
0 , s 0.
(stable with exponent a, since then E[e
ixT
a
] = e
[x[
a
)
Proof.
P[T
a
t] =
5.4.17
2P[X
t
a]
= 2
_

a
1
2t
e
x
2
2t
dx
=
_
t
0
s
3
2
a
2
e
a
2
2s
. .
a
(s)
ds.
since
a
2
2s
=
x
2
st
= x = a
_
t
s
and
dx
ds
= a
t(
1
2
)s
3
2
97
Corollary 5.4.20. (i) P[T
a
> t]
t
t,
_
2
a. (a 0)
(ii) E[T
a
] = +
Proof. (i) :
P[T
a
> t]
t =
5.4.17
tP [[X
t
[ a] =
1
2
_
a
a
e
x
2
2t
dx
t,
2a
2
.
(ii) :
E[T
a
] =
_

0
P[T
a
> t]dt
(i) small
t
0
=t
0
()big
_

t
0
_
_
2
a
_
t
1
2
dt =
Example 5.4.21 (Solution of the Dirichlet problem on the half-space.).
D := x = (x
1
, ..., x
d
) R
d
[x
d
> 0. Consider D not rel. kp. as always assumed before! But an
essential assumption for the existence of a solution is only P
x
[T
D
< ] = 1 x D. (cf. 5.2.7
and 5.4.15). Since D := z = (y, 0)[y R
d1
no longer kp, we have to assume f C
b
(D).
Stochastic:
Cone condition fullled z D!
Calculation of the harmonic measure
Under P
x
the exit time T := T
D
= inft > 0[X
d
t
= 0 (=passage time) has density
a
from
5.4.19 with a = x
d
. With notation = (
1
, ...,
d
) (C[0, ))
d
, X
T
:= (Y
T
, 0) it follows x D
h(x) =
2.5!
E
x
[f(X
T
)] = E
(x
1
,...,x
d1
)
_
E
x
d
[f(Y
T(
d
)
(
1
, ...,
d1
), 0)]
. .
=
5.4.19
0
f(Y
t
(
1
,...,
d1
),0)
x
d
2t
3
e
x
2
d
2t
. .
x
d
(t)
dt
=
Fubini
_

0
x
d
2t
3
e
x
2
d
2t
_
_
R
d1
f(y, 0)
1
2t
d1
exp
_
[[y (x
1
, ..., x
d1
)[[
2
2t
_
dy
_
dt
=
Fubini
_
R
d1
f(y, 0)
x
d
2
d
_

0
t
(
d
2
+1)
exp
_
[[(y, 0) x[[
2
2t
_
dt
. .
=(
d
2
)[[(y,0)x[[
d
2
d
2
dy
because:
_
0
t
t
dt =
s=
1
_
0
s
2
e
s
ds =
1
( 1)
=
D
x
(d
y
) =
(
d
2
)
d
2
x
d
[[x (y, 0)[[
d

d1
..
Leb.meas.
on R
d1
(dy)
Remark 5.4.22. The stochastic solution of the Dirichlet problem leads us to the harmonic
measure
D
x
= exit distribution of the Brownian Motion. For the Poisson problem we get Green
function = residence density of the BM in D. We want to show this in the next section.
98
5.5. Poissin equality and Green function
5.5. Poissin equality and Green function
Assumption: P
x
[T
D
< ] = 1 x D. (D not necessarily bounded)
Have f C
b
(D). Want u C
2
(D) C(

D) with
1
2
u = f on D
u = 0 on D
_
_
_
_
Poisson problem with homog. boundary condition
Remark 5.5.23. The general case u = g on D can be traced back to the above by adding the
solution of the Dirichlet problem with boundary condition g. We then know, because of 5.2.3,
that
u(x) = E
x
__
T
0
f(X
s
)ds
_
, x D. (5.5.2)
with T := T
D
((T
t
) stopping time with T
t
:=
>0
(X
s
)[s t +).
Remark 5.5.24. p
D
t
(x, A) := P
x
[X
t
A, t < T], A D, measurable, denes a measure on D.
It is obvious that
p
D
t
(x, )
B.sgr.
..
p
t
(x, ) dy (= Lebesgue-measure on R
d
)
= p
D
t
(x, y) : p
D
t
(x, dy) = p
D
t
(x, y)dy.
Denition 5.5.25.
G
D
(x, y) :=
_

0
p
D
t
(x, y)dt , x, y D
is called the Green function of D.
Attention: only if d 3, G
D
(x, y) < x ,= y. If d = 1, 2 and D = R
d
, G
D
(x, y) = x, y.
Aim: Representation of p
D
t
(x, y) and thus of G
D
(x, y). Because then for u as in (5.5.2)
u(x) = E
x
__

0
f(X
s
)1
T>s
ds
_
=
Fubini
_

0
E
x
[f(X
s
); T > s]
. .
f(y)p
D
s
(x,y)dy
ds
=
_
f(y)
_

0
p
D
s
(x, y)ds dy
Thus
Theorem 5.5.26. u(x) =
_
G
D
(x, y)f(y)dy solves Poission problem for D with boundary con-
dition 0.
If G
D
explicit, then u explicit.
Theorem 5.5.27. (i)
p
D
t
(x, y) =
Density of the Br. sgr.
..
p
t
(x, y) E
x
[
cont. in t
..
p
tT
(X
T
, y);
left cont. in t
..
T < t ]
. .
left cont. (Lebesgue)
, x, y D
Note [y X
T
[ d(y, D) > 0 since y D.
(ii) p
D
t
(x, y) is symmetric in x, y.
99
Proof. (i) A B(D)
P
x
[X
t
A, t < T] = P
x
[X
t
A] P
x
[X
t
A, T t]
. .
=P
x
[X
tT
T
A, T t
. .
J
T
]
=
5.4.18
{T<t}
P
X
T()
[X
tT()
A]dP
x
()
=
_
A
p
t
(x, y)dy
_
T<t
_
A
p
tT()
(X
T()
, y)dy dP
x
()
=
_
A
[p
t
(x, y) E
x
[p
tT
(X
T
, y), T < t]] dy
= (i)
(ii) By approximation: Let D
n
D, D
n
D, D
n
kp., T
n
:= T
D
n
. Then T
n
T and hence
T > t =
n
T
n
> t. Let
m
= m-th dyadic decomposition of [0, [.
= P
x
[X
t
A, T > t]
= lim
n
()P
x
[X
t
A, T
n
> t]
= lim
n
() lim
m
() P
x
[X
t
A, X
t
i
D
n
t
i

m
, t
i
t]
. .
=
A
_
D
n
...
_
D
n
p
t
1
..
=
1
2
m
(x, x
1
)p
t
2
t
1
..
=
1
2
m
(x
1
, x
2
)...p
t
t
M
..
=
1
2
m
(x
M
, y)dx
1
...dx
M
. .
=:q
n,m
(x,y) is symm., since p
s
(x
/
,y
/
)=p
s
(y
/
,x
/
).
dy
=
Levi
_
A
lim
n
() lim
m
()q
n,m
(x, y)
. .
=p
D
t
(x,y) symm. as limit of symm. ones
dy
Corollary 5.5.28. Let d 3.
(i) G
D
(x, y) = G(x, y) E
x
[G(X
T
, y)],
where
G(x, y) := 2
1
d
2
(
d
2
1)
(2)
d
2
[[x y[[
2d
=
_

0
p
t
(x, y)dt. Green function on R
d
(ii) G
D
is symm. and G
D
(, y) is harmonic on D y.
Proof. (i) Clear, by 5.5.27 and calculation of
_
0
p
t
(x, y)dt (Excercise!)
(ii) Symm. clear. Also clear that x E
x
[G(X
T
, y)] harmonic on D and G(, y) = 0 on
R
d
y (Excercise!)
100
5.6. Self intersection of the paths
d 2. t 0 x.
P
x
[Returning to X
t
] = P
x
[X
s
= S
t
for one s > t] (5.6.3)
=
_
1
X
t
()<}

t
()dP
x
()
=
5.4.18
X
t
Tmeas.
_
P
X
t
()
[
X
t
()
< ]
. .
=0 , since pt. polar,
as d2
dP
x
() = 0
Nevertheless we have P
x
-a.s.
d 2: n n-times self intersection, i.e.
t
1
() < ... < t
n
() : X
t
1
()
() = ... = X
t
n
()
().
d 3: the double points lie dense on path
d = 3: triple points
d 4: double points.
We want to show now:
Theorem 5.6.29. d = 3 = x R
3
P
x
[ double points lie dense on path ] = 1
Therefor we need some preparations.
Excursus to the capacity
d 3 : G(x, y) :=
1
[[x y[[
d2
(i.e. as before up to constants)
Denition 5.6.30. (i) Let be positive measure on R
d
.
G(x) :=
_
G(x, y)(dy) , x R
d
, ( [0, ])
is called the potential generated by and
e() :=
_
Gd
the energy of .
(ii) For K R
d
kp.,
C(K) := (infe()[ p.measure on K)
1
is the capacity of K.
Remark 5.6.31. (i) C(x) = 0 x R
d
, since e(
x
) = .
(ii) C(K) > 0 p.measure on K : e() <
101
(iii) C(K) = 0 = (K) = 0 ( := Leb. measure),
() because:
(K) > 0 =
_
Gd
. .
=
1
(K)
1
x
d2
1
K
1
K
. .
kp. support
(x)(dx)
< with := (K)
1
1
K

Theorem 5.6.32 (equilibrium principle). C(K) > 0 = nite measure
K
on K with:
G
K
1 , = 1
K
a.e. on K
without proof.
In particular: e(
K
) = [[
K
[[ :=
K
(K) and for
K
:=

K
[[
K
[[
(p.measure):
e(
K
)
1
[[
K
[[
2
e(
K
) =
1
[[
K
[[
= C(K) [[
K
[[.
Theorem 5.6.33. C(K) > 0 = h
K
(x) := P
x
[
=
K
..
T
K
c < ] > 0 on K
c
(and in fact harmonic
on K
c
).
Proof. u := G
K
harmonic on K
c
, since for x B
r
(x
0
) K
c
_
G
K
(z)
B
r
(x
0
)
x
(dz) =
_
K
_
G(z, y)
. .
harm. on
R
d
\y
B
r
(x
0
)
x
(dz)
K
(dy) =
_
G(x, y)
K
(dy).
Clear: 0 < u 1 and u(x) [[[[
K
sup
yK
G(x, y) = [[[[
K
d(x, K)
2d
0 for [[x[[ .
Choose D
n
rel. kp.

D
n
K
c
, D
n
K
c
, thus T
n
:= T
D
n
T := T
K
c. Hence x K
c
0 < u(x) = lim
n
E
x
[u(X
T
n
)] (u harmonice on K
c
)
E
x
_
lim
n
u(X
T
n
)
_
= E
x
_
lim
n
u(X
T
n
); lim
n
T
n
=
_
. .
=0, since on limT
n
=:
[[X
T
n
[[ because of transience
+E
x
_
lim
n
u(X
T
n
); sup
n
T
n
<
_
. .
P
x
[sup T
n
<]
=P
x
[T
K
c<]
= P
x
[T
K
c < ] u(x) > 0.
End of excursus.
Theorem 5.6.34. Let 0 a < b, d = 3. x R
3
, we have for P
x
-a.a. :
C(X
T
()[a t b
. .
=:K
) > 0
102
Proof.
:= Pushforward measure of
[[a,b]
. .
1 dim.
Leb. meas.
under s X
s
().
(occupation measuere for [a, b])
measure on K
and
e(
) =
_ _
G(x, y)
(dx)
(dy) = const.
_
b
a
_
b
a
[[X
s
() X
t
()[[
2d
dsdt
. .
()
< P
x
a.s., if d = 3, since:
E
x
[()] =
_
b
a
_
b
a
E
0
_
[[X
s
X
t
[[
2d
_
dsdt
= E
0
_
[[X
1
[[
2d
_
. .
<
_
b
a
_
b
a
1
[t s[
d2
2
dsdt
. .
()=
< , d = 3
= , d > 3
<
since
() = 2
_
b
a
_
b
t
1
(s t)
d2
2
dsdt = 2
_
b
a
_
bt
0
s
1
d
2
ds
. .
<
1
d
2
> 1
4 > d
dt
and then bounded in t.
Proof of 5.6.29. Let K
:= X
s
()[0 s 1
(, ) T(, ) :=
K
() (:= inft > 0[X

t
() K
)
is T
1
T-measurable,
since T t
!
=
z(2
n
Z)
3
K
2
n1
(z)
1
K
2
n1
(z)
t T
1
T
t
where K
2
n1(z) is a bounded ball around z with radius 2
n1
in max. norm.
= P
x
[X
t
= X
s
for an s [0, 1] and a t ]2, [ ]
= P
x
[
K
0

2
< ]
=
_
E
x
[1
K
0
<

2
[T
2
]()P
x
(d)
5.4.18
=
_
P
X
2
()
[
K
< ]
. .
P
x
a.s. >0, because:
P
x
(d) > 0
103
_
1
K
X
2
()P
x
(d) = E
x
[E
x
[1
K
0
X
2
()[T
1
]]
=
_
E
X
1
()
[1
K
X
1
]dP
x
= 0
as E
y
[1
K
X
1
] = N(y, 1)(K
) =
5.6.34,
5.6.31(iii)
0
and thus, because of 5.6.34 and 5.6.31(iii), 5.6.33 can be used.
= N 2, (large) such that for
T := (2 +
K
0

2
) N ,
(T
t
)-stopping time (with T
t
:=
>0
X
s
[s t +).
T
: = P
x
[X
t
= X
s
for an s [0, 1] and a t ]2, T]]
= P
0
[2 +
K
0

2
N
. .
=:A
o
T
T
] > 0
(independent of x due to translation invariance)
Let A
n
:= X
t
= X
s
for an s [nT, nT + 1] and a t ]nT + 2, (n + 1)T], n N. Then
A
n
=
1
nT
(A
0
), A
n
T
(n+1)T
, thus
P
x
[A
0
... A
n
] = E
x
[ P
x
[A
n
[T
nT
]
. .
=P
X
nT
[A
0
]
=P
x
[A
0
]=
T
, A
0
... A
n1
] =
Induction
n+1
T
= P
x
[A
0
]...P
x
[A
n
] = independent
(as P
x
[A
n
] = E
x
[E
x
[1
A
0

nT
[T
nT
]] =
see
above
T
> 0)
=
Borel-
Cantelli
-many A
n
occur
= P
x
-a.e. path has -many double points =
excercise
density.
5.7. Feyman-Kac-Formula
Let V C(D). Poisson-problem for
1
2
+V :
Have: g C
b
(D), f C(D)
Want: u C
2
(D) C(

D) with
_
1
2
u +V u = g on D
u = f on D
(5.7.4)
Theorem 5.7.35. If E
x
[exp((1 + T)
_
T
0
V
+
(X
s
)ds)] < x D and T := T
D
, then we have
for u with (5.7.4)
u(x) = E
x
_
f(X
T
) exp
__
T
0
V (X
s
)ds
_
_
+E
x
_
_
T
0
g(X
t
) exp
__
T
0
V (X
s
)ds
_
dt
. .
not obvious if /
1
, but as u(x)<, needs to be
_
, x D
104
5.8. The heat equation
Proof. Dene Z
t
:= u(X
t
) A
t
with A
t
:= exp(
_
t
0
V (X
s
)ds). (C
1
!)
=
Itos
product rule
Z
t
= Z
0
+
_
t
0
u(X
s
) dA
s
..
=V (X
s
)A
s
ds
+
_
t
0
A
s
du(X
s
) , t < T
with
_
t
0
A
s
du(X
s
) = lim
n
t
i
n
t
i
t
A
t
i
(u(X
t
i+1
) u(X
t
i
))
=
Ito
lim
n
t
i
m
A
t
i
__
t
i+1
t
i
(u, dX
s
)
_
. .
=:M
n
t
+
1
2
_
t
0
A
s
u(X
s
)ds
Let

D
m
D
m+1
, D
m
open, D
m
D. Then (M
n
t
) is local martingale until T with local sequence
T
m
:= T
D
m
m. Since for s T
D
m
m, A
s
exp(m[[V [[
D
m
), it follows that
M
t
:= lim
n
M
n
t
,
is local martingale with local sequence (T
m
) (cf. poof of I.3.2, II.1.4).
Furthermore
Z
t
= Z
0
+M
t
+
_
t
0
_
1
2
u +V u
_
(X
s
)
. .
g(X
s
)
A
s
ds , t < T
= E
x
[Z
T
m
t
]
. .
m
then t
E
x
[Z
T
]
= E
x
[Z
0
]
. .
u(x)
+E
x
_
_
tT
m
0
g(X
s
)A
s
ds
_
. .
m
then t
E
x
[
T
0
g(X
s
)A
s
ds]
since [Z
T
m
t
[ [[u[[
exp(
_
T
0
v
+
(X
s
)ds) L
1
.
Example 5.7.36.
1
2
u +u = g on D
u 0 on D
At least if < = (D), so small (cf. III 3.3 3)), such that E
x
_
e
T
< , x D. Then
u(x) = E
x
__
T
D
0
e
T
g(X
t
)dt
_
D = R
d
]0, T].
Have: f C(R
d
) with b, c > 0 : [f(x)[ be
c[[x[[
2
x R
d
()
Want: v C
2
(D), continuous on R
d
[0, T]
_
v
t
=
1
2
v on D ( on R
d
)
v(, 0) = f on R
d
(5.8.5)
105
Set v(x, t) := p
t
Br.sgr.
f(x) = E
x
[f(X
t
)] =
1
2t
d
_
f(y)e
yx
2
2t
dy
. .
Exists only if () is fullled
, (x, t) D.
Then v is a solution of (5.8.5) if T <
1
2c
(excercise). = existence
Dual heat equation on R
d
[0, T]:
_
u
t
+
1
2
u = 0 on R
d
[0, T[
u(, T) = f on R
d
_
u C
2
(D), u cont. on R
d
[0, T];
f C(R
d
)
_
(5.8.6)
Solution via time reversal on T, i.e.: Let v be solution of (5.8.5), then
u(x, t) := v(x, T t) solution of (5.8.6)
=
if T<
1
2c
E
x
[f(X
Tt
)] , (x, t) R
d
[0, T]
= existence for 5.8.6
Theorem 5.8.37 (Uniqueness). Let f C(R
d
) with () and T <
1
2c
.
Then ! solution u C
2
(D) of
u
t
+
1
2
u = 0 on R
d
[0, T[ with
(i) lim
t,T
yx
u(y, t) = f(x)
(ii) sup 0 t < T[U(x, t)[ Ke
a[[x[[
2
; K, a > 0.
, and u(x, t) = E
x
[f(X
Tt
)] is this solution.
Proof. u(x, t) := E
x
[f(X
Tt
] = p
Tt
f(x) fullles (i), (ii) (Exercise!)
(with K = b, a =
c
12cT
, [p
Tt
f(x)[ const e
cx
2
12c(Tt)
)
Uniqueness: Let u be solution of (5.8.6) with (i), (ii). Without loss of generality : f = 0.
Need to show: u(x, t) = 0
u(X
x
, t +s) = u(X
0
, t) +
_
s
0
u(X
r
, r)dX
r
. .
=:M
s
(Def.)
but: local martingale (cf. proof of 5.3.9)
+
_
s
0
_
1
2
+

r
_
u(X
r
, r)
. .
=0, since r<s<Tt<T
dr
Dene
R
n
:= infs > 0[ [X
s
[ > n , n N,
and let t
k
T t, t
k
< T t.
=
stopping theorem
+ remark II.1.5
u(x, t) = E
x
[u(X
R
n
t
k
, t +R
n
t
k
)]
= E
x
[u(X
t
k
, t +t
k
); R
n
> t
k
]
. .
=:I
k,n
+ E
x
[u(X
R
n
, t +R
n
); R
n
t
k
]
. .
(ii)
Ke
a[[x
R
n
[[
2
P
x
[R
n
t
k
]
t
k
,Tt
Ke
an
2
P
x
[R
n
< T t]
since 1
R
n
t
k
1
[R
n
<Tt[
106
since I
k,n

t
k
,Tt
E
x
[f(X
Tt
) ; R
n
T t] = 0 n because of (i), (ii) and because of
1
R
n
>t
k
1
R
n
Tt
Let Q
n
:=
_
x R
d
[ max
1id
[x
i
[
n
d
_
K
n
(0)
= P
x
[R
n
< T t]
d
j=1
P
x
_
max
0sTt
X
d
s

n
d
_
+P
x
_
max
0sTt
(X
j
s
)
n
d
_
=
5.4.17
d
j=1
2P
x
_
X
j
Tt

n
d
_
+ 2P
x
_
X
j
Tt

n
d
_
const.
d
j=1
_
e
d
x
j
2
/2(Tt)
+e
d
x
j
2
/2(Tt)
_
where the last inequality holds because of
_
a
e
x
2
2
dx a
1
e
a
2
2
= Ke
an
2
P
x
[R
n
< T t]
n
0, if
1
2d(Tt)
> a.
= u(x, t) = 0, if t > T
1
2ad
.
If we apply the above to T
1
:= T
1
3ad
, we get
u(x, t) = 0 , if t > T
1
1
2ad
= T
1
3ad

1
2ad
It follows with iteration
u(x, t) = 0 t [0, T] , x R
d
.
More general:
D R
d
[0, [. We want solution of
_
(

t
+
1
2
)u = 0 on D
u = f on D
(5.8.7)
Idea: Interpretation as Dirichlet problem for the space-time process.
We dene for r 0,
r
: [0, [ [0, [ by
r
(s) := r +s.
Let

: =
r
[r [0, [
= (,
r
) (sarting point t)
X
s
( ) = (X
s
(), r +s) R
d
[0, [
P
(x,t)
: = P
x
r
.
If we have again everything enough bounded, we get
u(x, t) = E
(x,t)
[f(

X
T
)]
with

T( ) := inft > 0[

X
T
( ) / D.
In the special case above: D := R
d
[0, T[ with f : D = R
d
T R we have
T = T t P
(x,t)
a.s., i.e.
u(x, t) = E
(x,t)
[f(X
Tt
,
0
(T t))]
= E
x
[f(X
Tt
; T)]
Hence as before.
107
6. Stochastic Dierential Equations
6.1. Solution denitons, examples
We consider the following SDE:
_
X
0
=
0
(initial condition)
dX
t
=
t
(X)dW
t
+b
t
(X)dt
(6.1.1)
For the sake of simplicity we will mostly consider only the Markov case:
_
t
(x) = (X
t
, t)
b
t
(X) = b(X
t
, t)
(6.1.2)
and only in R
1
(thus X
t
R), where we assume that
= (x, t) , b = b(x, t) Borel measurable on R [0, [. (6.1.3)
Let the right continuous ltration (T
t
)
t0
be as follows in the following denitions:
T
t
completion of
>0
(
0
, W
s
; s t +) with respect to P in T,
where
0
: R, T-measurable.
Denition 6.1.1. (i) A weak solution of (6.1.1) is a triple (X, W), (, T, P), (T
t
)
t0
with:
W (T
t
)-BM on (, T, P), X (T
t
)-adapted, such that P-a.s.
X
t
=
0
+
_
t
0
s
(X)dW
s
. .
in part.
+
_
t
0
b
s
(X)ds
. .
in part.
t 0 (resp. t < : stopping time). (6.1.4)
(ii) This solution is called weakly unique, if the distribution of
0
(on R) determines uniquely
the distribution of X.
Denition 6.1.2. (i) (6.1.1) has a strong solution, if F : R C([0, [)
0
C([0, [)
0
,
such that F(x, ) is T
0
t
/T
0
t
-measurable x, t and (, T, P), (T
t
), and all (T
t
)-BMs W,
X := F(
0
, W)
fullls (6.1.4) (F is the solution).
Here T
0
t
:= (
s
[s t) with
s
: C([0, [)
0
C([0, [)
0
,
s
() = (s).
(ii) This solution is called strong unique, if (X, W,
0
), which fullls (6.1.4)
X = F(
0
, W)
(in other words F is uniquely determined) ( = weak unique)
109
Remark 6.1.3. T
t
/T
t
-measurability of F(x, ) implies: X adapted
Example 6.1.4. (i) The Ornstein-Uhlenbeck process is a strong (unique) solution of dX =
dW Xdt (cf. solution formula in Section 4.1)
(ii) dX = XdW +Xdt has as strong (unique) solution
X
t
:= X
0
exp
_
W
t
+ (
1
2
2
)t
_
(Heuristically: Ito =
ln X
t
= ln X
0
+
_
t
0
1
X
t
dX
t
1
2
_
t
0
1
X
2
t
=
2
X
2
t
dW)
t
..
dX
t
= ln X
0
+W
t
+t
1
2
2
t
= X
t
= X
0
exp[W
t
+ (
1
2
2
)t]
except for X
t
could become 0 proof of existence.)
Proof. strong solution: Check with Ito X
0
, W.
Uniqueness: Let (X, W, X
0
) with (ii). Show with Ito:
X
t
exp
_
W
t
_

1
2
2
_
t
_
= X
0
(iii) Tanakas example (shows that one is possibly forced to look for a weak solution).
_
X
0
= 0
dX = sign(X)dW
, where sign(x) :=
_
1 , x 0
1 , x < 0
(sign(x) ,= 0, according to our denition, excludes the trivial solution X 0!)
Claim: ! weak solution.
Proof. weak uniqueness: Let X be a solution =
X
t
=
_
t
0
sign(X
s
)dW
s
is martingale
X
t
=
_
t
0
(sign(X
s
))
2
. .
=1
ds = t.
=
Levy
X is a Wiener process.
Existence: Let X be a Wiener process
= W
t
=
_
t
0
sign(X
s
)dX
s
is a Wiener process
= X
t
=
_
t
0
sign(X
s
)sign(X
s
)
. .
=1
dX
s
=
_
t
0
sign(X
s
)dW
s
(by Cor. 2.4.34) is a weak solution.
One can show that strong solution (cf. [?]).
110
6.2. Construction of solutions via transformation of the state space
6.2. Construction of solutions via transformation of the state space
Example 6.2.5. (i) X
t
:= W
2
t
fullls (according to Ito)
dX
t
= 2W
t
dW
t
+dt
Let
W
t
:=
_
t
0
signW
s
dW
s
Wiener process
= 2W
t
dW
t
= 2[W
t
[signW
s
dW
s
= 2[W
t
[d

W
t
= 2
_
X
t
d

W
t
,
i.e. dX
t
= 2
X
t
d

W
t
+dt has as weak solution X
t
:= W
2
t
.
X large: strong diusion! Although its drift is +1, X will be going back to 0 again and
again (since it is the square of BM and therefore the law of the iterated log holds.)
(ii) Bessel process [[W[[
X := [[W[[ =
i=1
(W
i
)
2
, W = (W
1
, ..., W
n
), BM in R
n
with start W
0
= 0.
Let
Y : = X
2
=
n
i=1
(W
i
)
2
( = X =
Y )
=
multdim.
It o
dY =
2W
i
dW
i
+ndt
and dY
t
=
4(W
i
)
2
dt = 4Y dt.
Hence

W
t
:=
_
t
0
1
2
Y
dY
nt
2
Y
t
=
_
t
0
1
Y
W
i
dW
i
is a local martingale with

W
t
=
_
t
0
1
4Y
s
dY
s
. .
4Y
s
ds
= t, hence

W BM.
Apply Ito-formula to X =
Y (ok for n 2, since 0 R

n
polar, i.e. Y
t
> 0 t > 0 P-a.s.)
to get
dX =
1
2
Y
dY +
1
2
_
1
4
Y
3
2
dY
_
= d

W +
1
2
Y
ndt
1
8
Y
3
2
4Y dt
= d

W +
n 1
2X
dt
W BM, since (loc.) martingale.

That is, the Bessel process X := [[W[[ is a weak solution of
dX = d

W +
n 1
2X
dt (strongly drifting away from 0!)
Theorem 6.2.6 (Zvonkin). Let b : R R be Borel measurable, bounded. Then
dX
t
= dW
t
+b(X
t
)dt
has a strong solution.
111
Proof. Idea: Transform X into a local martingale Y := (X) (for some to be found transfor-
mation : R R).
Ansatz: Let X be a solution of dX = dW +b(X) dt. Dene
Y
t
: = (X
t
)
=
It o
dY =
t
(X)dX +
1
2
tt
(X)dt
=
t
(X)dW +
t
(X)b(X) +
1
2
tt
(X)dt
Choose such that
t
(X)b(X) +
1
2
tt
(X) = 0, so Y is a local martingale and we get rid of
the (only) measurable b! I.e.
tt
= 2b
t
=
t
(y) = C exp
_
_
y
0
2b(z)dz
_
(> 0)
= (x) : = C
_
x
0
exp
_
_
y
0
2b(z)dz
_
dy +

C
(strongly increasing, C
1
)
( =
1
: C
1
)
Show: (x) =
_
x
0
exp
_
_
y
0
2b(z)dz
dy is as desired.
Sketch: Then dY =
t
(
1
(Y )
. .
=X
)dW.
Let (y) :=
t
(
1
(Y )), hence
dY = (y)dW (*)
Note that
(y) = exp
_
_

1
(y)
0
2b(z)dz
_
=
t
(y) = (y) 2b(
1
(y))
1
(y)
t
. .
=(
/
(
1
(y)))
1
=(y)
1
= 2b(
1
(y))
Thus
t
globally bounded, hence globally Lipschitz continuous:
[(x) (y)[ 2[[b[[
[x y[
Thus in () we have more regular coecients than in the original equation. According to the
results in Section 6.4 below, () has a unique strong solution Y (as is Lipschitz).
= X =
1
(Y ) (This even gives uniqueness!)
Problem:
1
only C
1
, but not C
2
! Hence attention with Ito, but
t
is still weakly C
1
.
Now rigorously: Solve () and set
X :=
1
(Y )
Then check by Ito that X solves dX
t
= dW
t
+b
t
dt.
Now more general:
dX = (X)dW +b(X)dt (*)
112
6.3. Method of Lamperti-Doss-Sussmann
6.3. Method of Lamperti-Doss-Sussmann
Lamperti ca. 1964
Doss-Sussmann end of the 70th
Surprisingly: One can solve () pathwise!
For simplicity here only for , b suciently smooth and for the Marko case:
dX
t
= (X
t
)dW
t
+b(X
t
)dt. (6.3.5)
We want to rewrite (6.3.5) with respect to the Stratonovich integral:
_
(X) dW.
Denition 6.3.7. Assume X, Y are (local) semimartingales. Then the Stratonovich integral of
X against Y is dened as
_
t
0
X
s
dY
s
:=
_
t
0
X
s
dY
s
+
1
2
X, Y
t
, t 0.
If we apply this to the semimartingales (X) and W we obtain
_
t
0
(X
s
) dW
s
=
_
t
0
(X
s
)dW
s
+
1
2
(X), W
s
=
_
t
0
(X
s
)dW
s
+
1
2
_
t
0
t
(X
s
)dX, W
s
(since d(X) =
t
(X)dX +
1
2
tt
(x)dx)
=
(6.3.5)
_
t
0
(X
s
)dW
s
+
1
2
_
t
0
t
(X
s
)(X
s
)ds.
or in dierential form
(X) dW = (X)dW +
1
2
(
t
)(X)dt.
Recall. In Section 1.2 we looked at another special case of Denition 6.3.7, namely for the
semimartingales f(X) and X (see p. 5) to get
_
t
0
f(X) dX =
_
t
0
f(X)dX +
1
2
f(X), X
t
=
_
t
0
f(X)dX +
1
2
_
t
0
f
t
(X)dX
(since df(X) = f
t
(X)dX +
1
2
f
tt
(X)dX.) Hence
dF(X) = F
t
(X) dX.
Let b
(X
t
) := b(X
t
)
1
2
t
(X
t
)(X
t
), then (6.3.5) can be written as
dX
t
= (X
t
) dW
t
+b
(X
t
)dt (6.3.6)
Let (x, t) be the ow corresponding to the ODE
x = (x)
i.e.
(x, 0) = x
t
:=
d
dt
(x, t) = ((x, t))
=
tt
:=
d
2
dt
2
(x, t) =
t
((x, t))
d
dt
(x, t) = (
t
)((x, t))
113
Ansatz: X
t
= (
t
,
t
) solves (6.3.6) with
0
= X
0
(= X
0
())
d =

dt (in part. t
t
() is C
1
)
() d
t
= dW
t
+

b
X
0
..
to be det.!
(
t
)dt,
0
= 0
=
2 dim. Ito
dX
t
=
x
(
t
,
t
)d
t
+
t
(
t
,
t
)d
t
+
1
2
tt
(
t
,
t
)dt
= (X
t
)dW
t
+
1
2
(
t
)(X
t
)dt
. .
(X
t
)dW
t
+
_
x
(
t
,
t
)

+((
t
,
t
))
b
X
0
(
t
)
_
dt
. .
!
=b
((
t
,
t
)
. .
X
t
)
There are two possibilities to simplify :
= 0 or

b
X
0
= 0
1st possibility: Lamperti
Let
t
X
0
, thus

= 0. Hence it should hold that
b
X
0
(y) =
b
((X
0
, y)) y R.
Then the problem is reduced to solve
d
t
= dW
t
+
b
((X
0
,
t
))dt ()
0
= 0
(ok, if (e.g.) X
0
= x
0
R, and , b are such that
b
((x
0
, )) globally Lipschitz; cf. Section
6.4 below.)
= X
t
= (X
0
,
t
) solves (6.3.6).
(strong solution, if () has strong solution)
2nd possibility: Doss-Sussmann
b
X
0
0, i.e.:
t
= W
t
=

t
=
1
x
(
t
, W
t
())b
((
t
, W
t
()) xed, that is:
= F
(
t
, t)
0
() = X
0
() , xed
Hence
t
=
t
(), but smooth in t for xed. Thus the problem reduces to solve a
family of certain dierential equations. It follows:
X
t
:= (
t
, W
t
) solves (6.3.6)
Special case for both:
dX = (X) dW
(i.e. b
0)
114
6.4. Construction of strong solutions on Lipschitz conditions (Picard-Lindelof method)
1. Lamperti: d = dW, hence X
t
= (X
0
, W
t
)
2. Doss-Sussmann:

t
= 0,
t
X
0
, hence X
t
= (X
0
, W
t
)
Example 6.3.8. (x) = x
2
(and b
0)
= (x, t) =
1
1
x
t
According to the special case X
t
= (X
0
, W
t
) solves
dX
t
= (X
t
) dW
t
(Stratonovich-DE!)
But explosion at time t
0
:= inft > 0[W
t
=
1
x
0
.
Consider now for (x) = x
2
dX
t
= (X)dW (Ito-DE!)
= (X) dW
1
2

t
(X)
. .
2X
3
dt
Then the strong drift X
3
prevents an explosion, because due to Lamperti we get, with initial
condition X
0
= x
0
,
X
t
=
1
1
x
0
t
,
where
d
t
= dW
t
+
b
..
id
((x
0
,
t
))dt
= dW
t
+
1
1
x
0
dt,
and
0
= 0.
Solution: = Bessel process in R
3
at
1
x
0
.
(
1
x
0
polar for BM in R
3
, hence also doesnt hit
1
x
0
, thus (X
t
) doesnt explode in nite time.)
6.4. Construction of strong solutions on Lipschitz conditions
(Picard-Lindelof method)
In this section let d N arbitrarily, G : R
+
R
d
R
d
2
, b : R
+
R
d
R
d
product
measurable and so T
t
-adapted that K > 0 with
[[G(, t, x) G(, t, y)[[ +[[b(, t, x) b(, t, y)[[ K[[x y[[
and [[G(, t, x)[[ +[[b(, t, x)[[ K(1 +[[x[[), for all , t R
+
, x, y R
d
.
(6.4.7)
Let W be a given R
d
-valued Wiener process on (, T, P), (T
t
).
Theorem 6.4.9. For an arbritary initial condition
0
L
2
(; R
d
), T
0
-measurable, there exists
exactly one strong solution X of the stochastic dierential equation
(here: G(, t, x) =: G(t, x), b(, t, x) =: b(t, x))
dX
t
= G(t, X
t
)dW
t
+b(t, X
t
)dt, X
0
=
0
, (6.4.8)
115
i.e., P-a.s.
X
t
=
0
+
_
t
0
G(s, X
s
)dW
x
+
_
t
0
b(s, X
s
)ds t 0
(in particular all integrals on the right hand side exist in the usual sense).
In particular, X has continuous sample paths.
Furthermore we have for every T > 0
E
_
sup
0tT
[[X
t
[[
2
_
C
_
1 +E
_
[[
0
[[
2
_
,
where the constant C only depends on K and T.
For the proof we need the following version of the Banach xed-point theorem:
Theorem 6.4.10. Let (E, d) be a complete metric space, : E E and c
n
[0, [, n 1,
with
n1
c
n
< + xed, such that
d(
n
x,
n
y) c
n
d(x, y) x, y E, n 1. (6.4.9)
Then has exactly one xed-point x and we have for all y E
d(x, y) (1 +
n1
c
n
)d(y, y).
Proof. Let x
0
E, x
n
:=
n
x
0
. Then
d(x
n+1
, x
n
) = d(
n
x
1
,
n
x
0
) c
n
d(x
1
, x
0
)
=
n>m
d(x
n
, x
m
)
n1
k=m
d(x
k+1
, x
k
) (
km
c
k
)d(x
1
, x
0
)
m,
0.
=
E complete
x : = lim
n
x
n
and x = lim
n
n+1
x
0
= x.
If x E is another xed-point, we get
d(x, x) = d(
n
x,
n
x) c
n
d(x, x)
n
0, i.e. x = x.
If y E, we get
d(x, y) = lim
n
d(
n
y, y) lim
n
n1
k=0
d(
k+1
y,
k
y)
lim
n
(1 +
n1
k=1
c
k
)d(y, y) = (1 +
k1
c
k
)d(y, y).
Proof of 6.4.9. Let E := L
2
(; C([0, T]; R
d
)) and [X[
T
:= E[ sup
0tT
[[X(t)[[
2
]
1
2
. Then (E, [ [
T
) is
complete, and for X : C([0, T]; R
d
), dene
(X)(t) :=
0
+
_
t
0
G(s, X(s))dW
s
. .
=:I
t
+
_
t
0
b(s, X(s))ds, t [0, T].
116
1st step: E E.
Since X() C
b
([0, T]; R
d
), E[ sup
0tT
I
2
t
]
Doob
4E[I
2
T
]
isometry
= 4E[
_
T
0
[[G(s, X(s)[[
2
ds],
hence
[X[
T
E
_
[[
0
[[
2
1
2
+ 2E
__
T
0
[[G(s, X(s))[[
2
ds
_
1
2
+E
__
T
0
[[b(s, X(s)[[
2
ds
_
1
2
E
[
0
[[
2
1
2
+ 3KE
__
T
0
(1 +[[X(s)[[)
2
ds
_
1
2
E
_
[[
0
[[
2
1
2
+ 3KE
__
T
0
[[X(s)[[
2
ds
_
1
2
+ 3KT
1
2
E
_
[[
0
[[
2
1
2
+ 3KT
1
2
([X[
T
+ 1) <
2nd step: fullls (6.4.9), because
X, Y E =
[
n+1
X
n+1
Y [
t
= E
_
sup
0rt
_
_
_
_
_
r
0
G(s,
n
X(s)) G(s,
n
Y (s))dW
s
+
_
r
0
b(s,
n
X(s)) b(s,
n
Y (s))ds
_
_
_
_
2
_1
2
2E
__
t
0
[[G(s,
n
X(s)) G(s,
n
Y (s))[[
2
ds
_
1
2
+E
__
t
0
[[b(s,
n
X(s)) b(s,
n
Y (s))[[
2
ds
_
1
2
3KE
__
t
0
[[
n
X(s)
n
Y (s)[[
2
ds
_
1
2
3K
__
t
0
[
n
X
n
Y [
2
s
ds
_
1
2
.
By iteration we get
[
n
X
n
Y [
T
3
n
K
n
__
T
0
_
t
1
0
...
_
t
n
1
0
[X Y [
2
t
n
dt
n
...dt
1
_
1
2
3
n
K
n
_
T
n
n!
_1
2
[X Y [
T
. (6.4.10)
=
6.4.10
! xed-point X E, furthermore
[X[
T

_
_
1 +
n1
3
n
K
n
T
n
2
(n!)
1
2
_
_
_
E
_
[[
0
[[
2
1
2
+ 3KT
1
2
_
C
_
1 +E
_
[[
0
[[
2
_1
2
for a constant C = C(K, T).
(where the rst inequality holds because of the last line of Theorem 6.4.10 with y 0 and
step 1.)
117
3rd step: Let Y be an arbitrary continuous adapted process which fullls (6.4.8), then Y = X.
Let S
n
:= inft 0[ [[X
t
[[ n or [[Y
t
[[ n ( +, n +)
= X
tS
n
Y
tS
n
=
_
tS
n
0
(G(s, X
s
) G(s, Y
s
))dW
s
+
_
tS
n
0
(b(s, X
s
) b(s, Y
s
))ds
= E
_
[[X
tS
n
Y
tS
n
[[
2
isometry
2E
__
tS
n
0
[[G(s, X
s
) G(s, Y
s
)[[
2
ds
_
+ 2E
__
tS
n
0
[[b(s, X
s
) b(s, Y
s
)[[
2
ds
_
4KE
__
tS
n
0
[[X
s
Y
s
[[
2
ds
_
4K
_
t
0
E
_
[[X
sS
n
Y
sS
n
[[
2
ds (6.4.11)
=
Gronwalls
Lemma
E
_
[[X
tS
n
Y
tS
n
[[
2
= 0 tn =
n,
path
continous
X
t
= Y
t
t S
n
P-a.s. n,
hence X = Y .
Lemma 6.4.11 (Gronwall). Let f : [0, T] R continuous. If there exist constants A R, b
]0, [ with
f(t) A+
_
t
0
bf(s)ds, t [0, T], (6.4.12)
then f(t) Ae
bt
, t [0, T].
Proof. Without loss of generality A = 0, otherwise consider f(t) Ae
bt
, which fullls (6.4.12)
because
f(t) Ae
bt
A (1 e
bt
)
. .
=
t
0
be
bs
ds
+
_
t
0
bf(s)ds =
_
t
0
b(f(s) Ae
bs
)ds.
But if A = 0, we have due to (6.4.12) for g(t) := e
bt
_
t
0
f(s)ds
g
t
(t) = bg(t) +e
bt
f(t) bg(t) +bg(t) = 0, g(0) = 0
= g(t) 0 t [0, T]. Hence by (6.4.12) f(t) 0 t [0, T].
Lemma 6.4.12 (generalized Gronwall). Let f, b L
1
([0, T]), such that b is nonnegative, f b
L
1
([0, T]) and let A : [0, T] R be an increasing function, such that
f(t) A(t) +
_
t
0
b(s)f(s)ds for dt a.e. t [0, T]. (6.4.13)
Then
ess sup
s[0,t]
f(s) A(t)e
t
0
b(s)ds
for dt a.e. t [0, T].
118
Proof.
Step 1 Assume A is constant. Then without loss of generality A = 0; otherwise consider
f Ae
0
b(s)ds
, which fullls (6.4.13) because for dt-a.e. t [0, T]
f(t) Ae
t
0
b(s)ds
A
_
1 e
t
0
b(s)ds
_
. .
t
0
b(s)e
s
0
b(r)dr
+
_
t
0
b(s)f(s)ds
=
_
t
0
b(s)
_
f(s) Ae
s
0
b(r)dr
_
ds.
But if A = 0, we have for g(t) := e
t
0
b(s)ds
_
t
0
b(s)f(s)ds and dt-a.e. t [0, T]
g
t
(t) = b(t)g(t) +e
t
0
b(s)ds
b(t) f(t)
..
t
0
b(s)f(s)ds
0.
Since g(0) = 0, it follows that g 0 dt-a.e., hence by (6.4.13) (with A = 0) f 0 dt-a.e.
Step 2 A increasing (not necessarily constant).
Let t [0, T]. Then by (6.4.13)
f(s) A(t) +
_
s
0
b(r)f(r)dr for ds a.e. s [0, t].
So, by Step 1,
f(s) A(t)e
s
0
b(r)dr
for ds-a.e. s [0, t].
Hence A(t)e
t
0
b(r)dr
.
Remark 6.4.13. (i) For the proof of uniqueness (more precisely in (6.4.11)) we only used
that G, b are locally Lipschitz, i.e. Lipschitz continuous on compact sets (e.g. B
n
(0)). In
other words: If G and b are locally Lipschitz, then the stochastic dierential equation (6.4.8)
has at most one strong solution.
(ii) If G 1, then (6.4.8) can be solved directly pathwise deterministic. (Use transformation
x
t
x
t
W
t
, but new b depends explicitly on time and ) If d = 1, then (6.4.8) can be
reduced via Lamperti to the case G = 1 if g, b are enough regular.
Theorem 6.4.14. Let b : R
d
R
d
be locally Lipschitz. Then for x
0
R
d
! solution of X(t) = x
0
+W(t) +
_
t
0
b(X(s))ds
on [0, [, where = explosion time, i.e. lim
t,()
[X(t)()[ = + if () < +. We have that
+ if b is globally Lipschitz.
(Without proof).
(Idea: transformation X
t
X
t
W
t
, then argue pathwise. Then b depends explicitly on time
and )
119
7. The martingale problem
Literature: [Str87]
7.1. Formulation, motivation
:= C([0, ); R
N
) (polish!), / := B
, /
t
:= (x(s)[0 s t), t 0, x(s)() := (s),
M
1
() := all probability-measure on (, /).
S
+
(R
N
) := all non-negative denite symmetric N N matrices.
Let a : [0, [R
N
S
+
(R
N
), b : [0, [R
N
R
N
bounded, measurable. Dene operator-
valued map t [0, [ L
t
by
L
t
:=
1
2
N
i,j=1
a
ij
(t, x)
x
i
x
j
+
N
i=1
b
i
(t, x)
x
i
(7.1.1)
with
x
i
:=

x
i
, 1 i N.
Denition 7.1.1. P M
1
() is called solution of the martingale problem for (L
t
) (with test
function space C
0
(R
N
)) with start in (s, x) [0, [R
N
, denoted by: P M.P.((s, x); (L
t
)), if
(i) P[x(0) = x] = 1 (thus process (s +t, x(t))
t0
starts in (s, x(0)
..
xP-a.s.
)),
(ii) (x(t))
_
t
0
[L
s+u
](x(u))du, t 0, is an (/
t
)-martingale under P C
0
(R
N
).
For C
2
(R
N
) set
X
s,
(t) := (x(t))
_
t
0
[L
s+u
](x(u))du , t 0. (7.1.2)
Example 7.1.2. Suppose there exists a probability space (, T, P) with ltration (T
t
) and an
(T
t
)-BM (W(t))
t0
and a (T
t
)-adapted continuous process X = X(t), t 0, such that for some
x R
N
X(t) = x +
_
t
0
(s +u, X(u))dW(u) +
_
t
0
b(s +u, X(u))du, t 0, P-a.s.
Let C
2
0
(R
N
). Then by the d-dimensional Ito-formula (see Proposition 1.3.21) with ( , )
R
N :=
Euclidean inner product in R
N
and X = (X
1
, . . . , X
N
) for t 0
(X(t)) = (x) +
_
t
0
(
(X(u)), dX(u))
R
N
+
1
2
_
t
0
N
i,j=1
x
i
x
j
(X(u))dX
i
, X
j
u
.
121
But since
X
i
, X
j
t
=
N
=1
_

0
i
(s +u, X(s))dW
(u),
N
/
=1
_

0
i
/
(s +u, X(u))dW
/
(u)
t
=
_
t
0
N
,
/
=1
i
(s +u, X(u))
i
/
(s +u, X(u))d W
, W
w
. .
,
/
du,
we obtain for (a
ij
) :=
T
and t 0
(X(t)) = (x) +
_
t
0
(
(X(u)), (s +u, X(u))dW(u))

R
N
+
_
t
0
(
(X(u)), b(s +u, X(u))

R
N
du
+
1
2
_
t
0
N
i,j=1
a
ij
(s +u, X(u))

2
x
i
x
j
(X(u))du
=
_
t
0
[L
s+u
] (X(u))du
+(x) +
_
t
0
(
(X(u)), (s +u, X(u))dW(u))

R
N
. .
a martingale!
Hence P X
1
M.P.((s, x), (L
t
)), if
t
= (a
ij
).
Let P M.P.((s, x); (L
t
)) be xed and C
0
(R
N
). Then X
s,
Mart
2
c
(P) (:= all (/
t
)-
martingales under P in L
2
with P-a.s. continuous paths).
Lemma 7.1.3. , C
0
(R
N
)
X
s,
, X
s,
t
=
_
t
0
(, a)
R
N(s +u, x(u))du, t 0, P a.s. (7.1.3)
122
Proof. By polarization without loss of generality : = . We have
X
s,
(t)
2
=
(7.1.2)
(x(t))
2
2(x(t))
_
t
0
[L
s+u
](x(u))du +
__
t
0
[L
s+u
](x(u))du
_
2
=
(7.1.2)
2
(x(t)) 2X
s,
(t)
_
t
0
[L
s+u
](x(u))du
__
t
0
[L
s+u
](x(u))du
_
2
=
(7.1.2)+
Itos product
formula
X
s,
2(t)
. .
mart.!
+
_
t
0
[L
s+u
2
](x(u))du
__
t
0
[L
s+u
](x(u))du
_
2
2
_
t
0
_
u
0
[L
s+u
/ ](x(u
t
))du
t
dX
s,
(u)
. .
mart.
2
_
t
0
X
s,
(u)
. .
(7.1.2)
= (x(u))
u
0
[L
s+u
/ ](x(u
/
))du
/
d
__

0
[L
s+u
/ ](x(u
t
))du
t
_
(u)
. .
=
(Itos)
product
formula
2
t
0
[L
s+u
](x(u))du[
t
0
[L
s+u
](x(u))du]
2
=
_
t
0
[L
s+u
2
2L
s+u
]
. .
()
=
a
ij
(s+u, )
x
i
x
j
(x(u))du + martingale
=
_
t
0
(, a)(s +u, x(u))du + martingale
where () holds because
x
i
x
j
2
= 2
x
i
(
x
j
) = 2
x
i
x
j
+ 2(
x
i
)(
x
j
) and
x
i
2
=
2
x
i
.
Lemma 7.1.4. (i) C
2
(R
N
), X
s,
Mart
loc
c
(P) (:=all continuous loc. (/
t
)-martingales
until +) under P with P-a.s. continuous paths), and (7.1.3) holds , C
2
(R
N
).
(ii) If x(t) := x(t)
_
t
0
b(s + u, x(u))du, t 0, then x
i
Mart
2
c
(P), 1 i N (even
x
i
L
q
(P) q ]0, [) and
( x
i
, x
j
(t))
1i,jN
=
_
t
0
a(s +u, x(u))du, t 0 P a.s. (7.1.4)
(iii) If f C
1,2
([0, [R
N
) (i.e. C
1
in t, C
2
in x), then
f(t, x(t)) f(0, x)
_
t
0
[(
u
+L
s+u
)f](u, x(u))du
=
N
i=1
_
t
0
x
i
f(u, x(u))d x
i
(u)
. .
loc. mart.
=:
_
t
0
[
x
f](u, x(u)) d x(u), 0, P a.s.
123
Proof. (i) : Clear (with trivial approximation) X
s,
Mart
2
c
(P) C
2
0
(R
N
). Now let
C
2
(R
N
) and n N. Dene
n
:= inft 0[ [x(t)[ n,
and take
n
C
0
(R
N
) with
n
= 1 on B
n+1
(0). Then
n
and X
s,
(
n
) =
X
s,
n
(
n
) Mart
2
c
(P), hence X
s,
Mart
loc
c
(P) and (7.1.3) is still true.
(ii) : (i) with (x) = x
i
, implies that x
i
Mart
loc
c
(P), 1 i N, and (7.1.3) holds with
(x) = x
j
which implies (7.1.4). By Cor. 1.4.32 it follows that x
i
is a square integrable
martingale 1 i N, since by 7.1.4 x
i
t
is bounded for every t 0.
Now we need
Lemma 7.1.5. Let X Mart
l1oc
c
(P) and stopping times with X()X() A
for some A ]0, [. Then
P[ sup
t
[X(t) X()[ R] 2 exp(R
2
/2A) R 0.
In particular q ]0, [ C
q
[1, [ with
E
p
_
sup
t
[X(t) X()[
q
_
C
q
A
q
2
.
Proof. [Str87, p.59, Lemma (3.10)]
Continuing with proof of 7.1.4 (ii): 7.1.5 = 0 s < T
P
_
sup
stT
[[ x(t) x(s)[[ R
_
2 N exp
_
R
2
2AN(T s)
_
(7.1.5)
with A := sup
t,y
[[a(t, y)[[
oper.norm
. In particular q ]0, [ C(q, N) [1, [ with
[ sup
stT
[[ x(t) x(s)[[ [
L
q
(P)
C(q, N)(A(T s))
1
2
(7.1.6)
In particular x
i
L
q
(P) 1 i N.
(iii) Follows directly with (ii) from Itos formula (for the time-depending case).
Remark 7.1.6. a 0 =
(7.1.4)
& I.3.3 ?
x(t) = x(0) = x(0) = x. Thus
x(t) = x +
_
t
0
b(s +u, x(u))du , t 0 P a.s.
Hence (P-a.s.) deterministic case included.
Theorem 7.1.7. Let a : [0, [R
N
S
+
(R
N
), b : [0, [R
N
R
N
bounded, measurable and
P M.P.((s, x), L
t
)) for an (s, x) [0, [R
N
. Supposing : [0, [R
N
Hom(R
d
; R
N
)
(= d N matrices over R) bounded, measurable with a =
T
.
124
Then ddim. BM (((t))
t0
, (T
t
)
t0
, Q) on a probability space (
, T, Q) and continuous,
(T
t
)
t0
-progressiv measurable map X : [0, [
R
N
, such that
X(t) = x +
_
t
0
(s +u, X(u))d(u) +
_
t
0
b(s +u, X(u))du, t 0, Qa.s.
and
P = Q X
1
.
(i.e. briey: dX(t) = (s +t, X(t))d(t) +b(s +t, X(t))dt, X(0) = x). If particularly a(t, y) > 0
(t, y) [0, [R
N
and := a
1
2
(as pos. s.a. operator uniquely determined!), then one can
take:

:= , Q := P, X() = x() and
(t) :=
_
t
0
1
(s +u, x(u))d x(u), t 0, T
t
:= /
t
, T := /,
with x(t) := x(t)
_
t
0
b(s +u, x(u))du, t 0. (martingale according to 7.1.4(ii)!)
Proof. [Str87, Theorem(2.6), p.91 .]
Exercise 7.1.8: (i) Show: P M.P.((s, x), (L
t
)) P[x(0) = x] = 1 and x
i
Mart
loc
c
(P)
1 i N and (7.1.4) holds.
(ii) Show: What we have done so far, can be generalized to the following situation: Let a :
[0, [R
N
S
+
(R
N
), b : [0, [R
N
R
N
bounded, (/
t
)-progressiv measurable
and (L
t
) as in (7.1.1) with these a, b. Dene M.P(x; (L
t
)) as the set of all P M
1
() with
P[x(0) = x] = 1 and
(x(t))
_
t
0
[L
u
](x(u))du, t 0,
is an (/
t
)-martingale under P C
0
(R
N
).
Dene x(t) := x(t)
_
t
0
b(u)du. Show:
(a) P M.P.(x; (L
t
)) P[x(0) = x] = 1 and x
i
Mart
l
c
oc(P) 1 i N and
( x
i
, x
j
)
i,j
=
_
0
a(u)du) P-a.s.
(b) P M.P.(x; (L
t
)) = 7.1.4(iii), (7.1.5), (7.1.6) hold.
(iii) If a, b in (7.1.1) are independent of t, then the martingale problem in 7.1.1 is time inde-
pendent.
Every time dependent martingale problem becomes time independent via the following
trick (space time homogenization): Set

:= C([0, [; R
N+1
) C([0, ); R
1
) .
Show: P M.P.((s, x), (L
t
))

P M.P.( x,

L), where x := (s, x) ( [0, ) R
N
),
L :=
t
+L
t
,

P :=
s+
P and
s+
M
1
(C[0, [; R
1
) is Dirac measure on path t s+t.
Denition 7.1.9. The martingale problem for (L
t
) in (7.1.1) (with test function space C
0
(R
N
))
is called well-posed, if (s, x) [0, [R
N
#M.P.((s, x), (L
t
)) = 1.
Theorem 7.1.10. Suppose the martingale problem for (L
t
) in (7.1.1) (with C
0
(R
N
)) is well-
posed. Let P
s,x
[(s, x) [0, [R
N
be the corresponding family of solutions. Then:
(i) (s, x) P
s,x
is measurable from [0, [R
N
to M
1
() (with weak topology).
(ii) Suppose we have
x
_
T
0
a(t, x)dt, x
_
T
0
b(t, x)dt continuous T > 0. (7.1.7)
Then (s, x) P
s,x
continuous from [0, [R
N
to M
1
().
125
Proof. [Str87, p.85 .]
Theorem 7.1.11 (Existence). Let (L
t
) be as in (7.1.1), and (7.1.7) holds. Then
#M.P.((s, x), (L
t
)) 1 (s, x) [0, [R
N
.
Proof. [Str87, p. 84 (1.20) & (1.21)]
Theorem 7.1.12 (Uniqueness). Let (L
t
) be as in (7.1.1). Suppose the following condition holds:
(s, x) [0, [R
N
we have
P, Q M.P.((s, x); (L
t
)) = P x(t)
1
= Q s(t)
1
t 0. (7.1.8)
Then
#M.P.((s, x); (L
t
)) 1 (s, x) [0, [R
N
.
7.1.11 is useful for applications, 7.1.12 not. To get a sucient condition for (7.1.8) we need some
preparations:
Denition 7.1.13. Let T be a set of bounded, measurable functions on a measurable space
(E, B). T is called determining set, if , M
1
(E)
_
d =
_
d T = = .
Theorem 7.1.14. Suppose there exisists a determining set T C
b
(R
N
), such that one of the
following conditions holds:
(i) T > 0, T u
T,

()
C
1,2
b
([0, T[R
N
) with
(
t
+L
t
)u
T,
= 0 on [0, T[R
N
and
lim
tT
u
T,
(t, x) = (x) x R
N
,
(automatically unifom in x because of ()).
(ii) T > 0, T u
T,
C
1,2
b
([0, T[R
N
) with
(
t
+L
t
) u
T,
= on [0, T[R
N
and
lim
tT
u
T,
(t, x) = 0 x R
N
,
(automatically uniform in x).
Then #M.P.((s, x); (L
t
)) 1 (s, x) [0, [R
N
.
More precisely, in case (i): (s, x) [0, [R
N
and P M.P.((s, x); (L
t
))
E
P
[(x(T s))] = u
T,
(s, x) T > s (7.1.9)
Thus (7.1.8) is fullled.
More precisely, in case (ii): (s, x) [0, [R
N
and P M.P.((s, x); (L
t
))
E
P
__
Ts
0
(x(u))du
_
= u
T,
(s, x) T > s (7.1.10)
Thus (7.1.8) is fullled.
126
Proof. Suppose (i) holds: Let T > s. Then according to 7.1.4(iii) with f := u
T,
( + s, )
t [0, T s[
E
P
[u
T,
(t +s, x(t))] = u
T,
(s, x).
With t T s (7.1.9) follows , hence (7.1.8) holds and the assertion is implied by 7.1.12. Suppose
(ii) holds: Let T > s. Then according to 7.1.4(iii) with f := u
T,
( +s, ) t [0, T s[
E
P
[ u
T,
(t +s, x(t))] u
T,
(s, x) = E
P
__
t
0
(x(u))du
_
.
With t T s (7.1.10) follows, thus a, b R
+
, a < b
_
b
a
E
P
[(x(u))du = u
b+s,
(s, x) + u
a+s,
(s, x)
Hence by continuity u E
P
[(x(u))] is uniquely determined on R
+
. Thus (7.1.8) holds and
7.1.12 implies the assertion.
Finally we want to prove 7.1.12. Therefore we need:
Theorem 7.1.15. Let P M.P.((s, x), (L
t
)) and a stopping time. Then:
(i) /
= (x(t )[t 0). In particular /
is countably generated.
(ii) Let / be a countably generated sub--algebra of /
, such that
x((), ) /measurable.
Let P
be a regular conditional probability distribution of P given /

_
i.e.: (, A) P
[A], , A /, is stochastic kernel on (, /) such that:

a) A / : P[A[/]() = P
[A] for P a.a. .

b) A / : P
(A) is /-measurable and (regular)

P
[A] = 1
A
() for all , if A /.
(exists, since / is countably generated) i.e.: P
,
=
_
.
Then P-null set / such that:
c
P

1
()
M.P.((s +(), x((), ); (L
t
))
(where for t 0 :
t
: ,
t
() = ( +t).)
Proof. cf. [Str87, III.(1.13), p.78, and III.(1.14), p. 79].
Proof of 7.1.12. Let P, Q M.P.((s, x), (L
t
)).
To show: () P (x(t
1
), ..., x(t
n
))
1
= Q (x(t
1
), ..., x(t
n
))
1
0 t
1
< ... < t
n
and all n N.
Induction over n:
n = 1: Ok, according to assumption.
n n + 1: Suppose () holds for n N and all 0 t
1
< ... < t
n
. Let 0 t
1
< ... < t
n
< t
n+1
.
Let / =: (x(t
1
), ..., x(t
n
)) and P
resp. Q
regular conditional probability distri-

butions of P resp. Q given /.
According to 7.1.14(ii) a P-null set
1
/ resp. a Q-null set
2
/ such that
P

1
t
n
, Q

1
t
n
M.P.((s +t
n
, x(t
n
, )); (L
t
))

c
1

c
2
.
127
Hence for :=
1

2
we have, according to assumption (7.1.8),
c
P
x(t
n+1
)
1
= (P

1
t
n
) x(t
n+1
t
n
)
1
=
(7.1.8)
(Q

1
t
n
) x(t
n+1
t
n
)
1
= Q
x(t
n+1
)
1
. (**)
In addition, due to the induction hypothesis P = Q on /, hence P[] = Q[] = 0, and thus
B
1
, ..., B
n+1
B(R
N
)
P[x(t
1
) B
1
, ..., x(t
n+1
) B
n+1
]
= E
P
[1
B
1
(x(t
1
))...1
B
n
(x(t
n
))
. .
,
E
P
.
[1
B
n+1
(x(t
n+1
))]
. .
,measurable
]
= E
Q
[1
c1
B
1
(x(t
1
))...1
B
n
(x(t
n
)) E
P
.
[1
B
n+1
(x(t
n+1
))]
. .
=
()
E
Q
.
[1
B
n+1
(x(t
n+1
))]
]
=
analog
Q[x(t
1
) B
1
, ..., x(t
n+1
) B
n+1
].
128
A. Time Transformation
Let (, T, P) be a probability space, (T
t
)
t0
a right-contiuous, complete ltration and Y a P-a.s.
continuous locale martingale up to (with respect to (T
t
)) such that Y
0
0. For simplicity let
Y
(:= lim
t
Y
t
) = P-a.s..
Dene the inverse of Y
t
by
C
t
:= inf s > 0[Y
s
> t
. .
,=
.
Theorem A.0.1. Let Y
= P-a.s. and W
t
:= Y
C
t
, t 0. Then, W is a ((
t
)
t0
:= T
C
t
-
Brownian motion (cf. 1.5.34) and Y
t
= W
Y )
t
for all t 0.
For the proof we need the following two lemmas.
Lemma A.0.2. (i) The map t C
t
is increasing and right-continuous.
(ii) C
t
is an (T
s
)-stopping time for all t.
(iii)
Y
C
t
= t P-a.s. t.
(iv) t C
Y )
t
. (Note that in general case t ,= C
Y )
t
!)
Proof. (iii): Without loss of generality Y is continuous everywhere. Then Y
C
t
t, since
t Y
t
is (right-)continuous.
Assumption: Y
C
t
t + for > 0.
Then, since t Y
t
is continuous, there exists a > 0 such that Y
C
t
t +

2
> t. Hence,
C
t
C
t
. Since this is impossible it follows that Y
C
t
< t + for all > 0.
(i): Obviuously t C
t
is increasing. It remains to show that
lim
ut
C
u
C
t
.
( is clear since C
k
is increasing.)
Let > 0. Then Y
C
t
+
> t. Hence, there exists a > 0 such that
Y
C
t
+
> u u [t, t +].
Thus,
C
u
C
t
+ u [t, t +].
(ii): We have
C
t
< u = Y
u
> t u, t, (A.0.1)
because is clear and, if Y
u
> t, then there exists an > 0 such that Y
u
> t and,
therefore, C
t
u < u. But
Y
u
> t T
u
.
Hence, by (A.0.1)
C
t
u T
u+
= T
u
.
(iv): If C
Y )
t
< t, then by (A.0.1) we would get Y
t
> Y
t
. Therefore, C
Y )
t
t.
129
Note that (
t
:= T
C
t
, t 0 (ltration, since C
t
is increasing!).
Lemma A.0.3. (i) ((
t
)
t0
is right-continuous and complete.
(ii) Y
t
is a stopping time with respect to ((
s
)
s0
for all t.
Proof. (ii) is clear by (A.0.1).
(i): Right-continuity: Let A
>0
T
C
t+
(in particular, A T
C
t+
1
n
n). Hence, for all s
A C
t+
1
n
< s T
s
n, s,
therefore,
A
_
n
C
t+
1
n
< s
. .
=C
t
<s
T
s
s,
thus A C
t
s T
s+
= T
s
s, and A T
C
t
.
Completeness: Let A
0
T
C
t
with P[A
0
] = 0 and A A
0
. Then for all s
A C
t
s T
s
as a subset of an T
s
-measurable P-zero set A
0
C
t
s, since T
s
is complete. Hence, A
T
C
t
.
Proof of A.0.1 . Since Y is P-a.s. continuous and (C
t
) is right-continuous, (W
t
) is P-a.s. right-
continuous.
Step 1. t W
t
= Y
C
t
is P-a.s. continuous:
It suce to show that (cf. Exercises)
P[Y
u
= Y
t
for t u
t
t 0] = 1, (A.0.2)
where
t
:= infs > t[Y
s
> Y
t
,
hence,
Y
u
= Y
t
for u [t,
t
].
(A.0.2) is sucient, since then Y is constant on the interval, where Y is constant and by
denition this is the case on [C
t
, C
t
]. Therefore, Y
C
t
= Y
C
t
. For (A.0.2) it remains to show
that for all r Q
+
P[Y
u
= Y
r
for r u
r
r Q
+
] = 1, (A.0.3)
because, if t 0 with t <
t
, then for all r [t,
t
] Q
+
r
=
t
and (A.0.3) implies (A.0.2),
since Y is P-a.s. (right-)continuous.
Let (T
t
n
)
nN
be a localizing sequence for Y . Then
T
n
:= inft > 0[[Y
t
[ > n T
t
n
, n N,
is again a localizing sequence. Fix r Q
+
. For n N set
N
(n)
t
:= Y
(r+t)
r
T
n
Y
rT
n
, t 0,
T
t
:= T
t+r
, t 0.
Then N
(n)
is a continuous bounded martingale with respect to (

T
t
), since by the stopping
theorem s t
E[N
(n)
t
[
T
s
] = E[Y
(r+t)
r
T
n
Y
rT
n
[T
r+s
] = Y
(r+s)
r
T
n
Y
rT
n
= N
(n)
s
,
130
since
r
is a stopping time with respect to (

T
t
). Additionally,
N
(n)
t
= Y
(r+t)
r
T
n
Y
rT
n
= 0 t 0.
Hence,
E[(N
(n)
t
)
2
] = E[N
(n)
t
] = 0 t 0,
thus
N
(n)
t
= 0 P-a.s..
Letting n implies
Y
(r+t)
r
Y
r
= 0 P-a.s. t 0,
which yields (A.0.3).
Step 2. W is a local martingale (up to ) with respect to ((
t
):
By A.0.2(ii) C
t
is an (T
s
)-stopping time. For t 0
E[Y
C
t
s
] = E[Y
C
t
s
] E[Y
C
t
] = t.
Therefore, by Corollary 1.4.32 Y
C
t
is a martingale and
E[(Y
C
t
s
)
2
] liminf
n
E[Y
2
sC
t
T
n
] = liminf
n
E[Y
sC
t
T
n
] t.
Hence, (Y
C
t
s
)
s0
and ((Y
C
t
s
)
2
)
s0
are uniformly integrable.
Moreover, C
Y )
T
n
is a (T
t
)-stopping time, because:
C
Y )
T
n
< u
(A.0.1)
= Y
u
> Y
T
n
=
_
rQ
Y
u
> r r Y
T
n
(A.0.1)
=
_
rQ
C
r
< u C
r
T
n
=
_
rQ
C
r
< u
. .
T
u
C
r
u T
n
. .
T
C
r
uT
n
T
u
Thus,
C
Y )
T
n
u T
u+
= T
u
.
Furthermore, Y
T
n
is a stopping time with respect to ((
t
), since
Y
T
n
u
(A.0.1)
= C
u
T
n
T
T
n
C
u
T
C
u
= (
k
.
Hence, for all t > s (since C
tY )
T
n
= C
t
C
Y )
T
n
, because (C
t
) is increasing)
E[W
tY )
T
n
[(
s
] = E[W
tY )
T
n
[(
s
] = E[Y
C
tY )
T
n
[(
s
]
= E[Y
C
t
C
Y )
T
n
[T
C
s
] = E[Y
C
t
C
t
C
Y )
T
n
[T
C
s
]
Since (Y
C
t
s
)
C
Y )
T
n
is an uniformly integrable martingale (note that C
s
is not necessarily boun-
ded), it follows that
= Y
C
t
C
s
C
Y )
T
n
s<t
= Y
C
s
C
Y )
T
n
= W
sY )
T
n
.
131
Hence, W is a local martingale with localising sequence (Y
T
n
)
nN
(up to ).
Step 3. W
t
= t t. In particular, W is a Brownian motion:
Since by L
2
-martingale convergence theorem ((Y
C
t
s
)
2
)
s0
is uniformly integrable and Y
C
t
s
=
Y
C
t
s
Y
C
t
= t s, it follows that (Y
C
t
)
2
Y
C
t
is an uniformly integrable martingale.
Hence, (though C
t
, C
s
are not bounded) by the stopping theorem we get that for all t > s
E[W
2
t
t[(
s
] = E[Y
2
C
t
Y
C
t
[T
C
s
]
= E[(Y
C
t
C
t
)
2
Y
C
t
C
t
[T
C
s
] = (Y
C
t
C
s
)
2
Y
C
t
C
s
= Y
2
C
s
Y
C
s
= W
2
s
s.
Therefore, (W
2
t
t)
t0
is a continuous martingale with respect to (G)
t0
. By the Doob-Meyer
decomposition it follows that W
t
= t for all t and by 1.4.32 that (W
t
)
t0
is a martingale. By
Levys characterization theorem 1.5.34 W
t
is a Brownian motion.
Step 4. Y
t
= W
Y )
t
for all t 0:
We have
W
Y )
t
= Y
C
Y )
t
and since s Y
s
is increasing and continuous
C
Y )
t
= infs > 0[Y
s
> Y
t
= infs > t[Y
s
> Y
t
=
t
.
Therefore, Y is constant on [t, C
Y )
t
]. By (A.0.2) we get that Y is constant on [t, C
Y )
t
], thus,
W
Y )
t
= Y
C
Y )
t
= Y
t
.
Remark A.0.4. We have supposed that P[Y
= ] = 1. (This has been necessary as the

counter example
:= , Y 0
shows.) Basically, the theorem also holds for the case, where P[Y
< ] > 0. But, one possibly

has to enlarge .
Construction of the enlarged Wiener space:
Let (W
t
t
)
t0
be a Wiener process on (
t
, T
t
, P
t
) with respect to (T
t
t
). Set
:=
t
,
T := T T
t
,
P := P P
t
,
C
t
:=
_
infs[Y
s
> t on Y
> t,
on Y
t.
Let
T
t
:= (T
C
t
s
[s 0),
(
t
:=

T
t
T
t
t
,
W
t
:=
_
Y
C
t
on Y
> t,
W
t
t
W
t
Y )
+Y
on Y
t.
Then W is a Wiener process on (
,

T,

P) with respect to ((
t
)
t0
, Y
t
is a ((
t
)-stopping time
and we have
Y
t
= W
Y )
t
.
For the proof cf. [IW89, Chapter II, Theorem 7,21].
132
Literaturverzeichnis
[Arn06] Vladimir I. Arnold, Ordinary dierential equations, Universitext, Springer-Verlag, Ber-
lin, 2006, Translated from the Russian by Roger Cooke, Second printing of the 1992
edition. MR MR2242407
[Bau78] Heinz Bauer, Wahrscheinlichkeitstheorie und Grundz uge der Masstheorie, revised ed.,
Walter de Gruyter, Berlin-New York, 1978, de Gruyter Lehrbuch. MR MR514208
(80b:60001)
[Cho69a] Gustave Choquet, Lectures on analysis. Vol. I: Integration and topological vector
spaces, Edited by J. Marsden, T. Lance and S. Gelbart, W. A. Benjamin, Inc., New
York-Amsterdam, 1969. MR MR0250011 (40 #3252)
[Cho69b] , Lectures on analysis. Vol. II: Representation theory, Edited by J. Marsden,
T. Lance and S. Gelbart, W. A. Benjamin, Inc., New York-Amsterdam, 1969. MR
MR0250012 (40 #3253)
[CW90] K. L. Chung and R. J. Williams, Introduction to stochastic integration, second ed.,
Probability and its Applications, Birkhauser Boston Inc., Boston, MA, 1990. MR
MR1102676 (92d:60057)
[dB03] G. de Barra, Measure theory and integration, Horwood Publishing Series. Mathematics
and Its Applications, Horwood Publishing Limited, Chichester, 2003, Revised edition
of the 1981 original. MR MR2052644 (2004k:28002)
[IW89] Nobuyuki Ikeda and Shinzo Watanabe, Stochastic dierential equations and diusi-
on processes, second ed., North-Holland Mathematical Library, vol. 24, North-Holland
Publishing Co., Amsterdam, 1989. MR MR1011252 (90m:60069)
[KL99] Claude Kipnis and Claudio Landim, Scaling limits of interacting particle systems,
Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathe-
matical Sciences], vol. 320, Springer-Verlag, Berlin, 1999. MR MR1707314 (2000i:60001)
[Kry] Nicolai V. Krylov, http://www.math.umn.edu/krylov/.
[LS72] R. Sh. Lipster and A. N. Shiryayev, Statistics of conditionally Gaussian random se-
quences, Proceedings of the Sixth Berkeley Symposium on Mathematical Statistics and
Probability (Univ. California, Berkeley, Calif., 1970/1971), Vol. II: Probability theo-
ry (Berkeley, Calif.), Univ. California Press, 1972, pp. 389422. MR MR0400377 (53
#4211)
[MR92] Zhi Ming Ma and Michael Rockner, Introduction to the theory of (nonsymmetric) Di-
richlet forms, Universitext, Springer-Verlag, Berlin, 1992. MR MR1214375 (94d:60119)
[Nua06] David Nualart, The Malliavin calculus and related topics, second ed., Probability and its
Applications (New York), Springer-Verlag, Berlin, 2006. MR MR2200233 (2006j:60004)
[Oks03] Bernt Oksendal, Stochastic dierential equations, sixth ed., Universitext, Springer-
Verlag, Berlin, 2003, An introduction with applications. MR MR2001996 (2004e:60102)
133
Literaturverzeichnis
[Roc04] M. Rockner, Einf uhrung in die lineare Funktionalanalysis (Vorlesungsskript), 2004.
[Roc11] , Wahrscheinlichkeitstheorie (Vorlesungsskript), 2011.
[RS80] Michael Reed and Barry Simon, Methods of modern mathematical physics. I, second
ed., Academic Press Inc. [Harcourt Brace Jovanovich Publishers], New York, 1980,
Functional analysis. MR MR751959 (85e:46002)
[RW87] L. C. G. Rogers and David Williams, Diusions, Markov processes, and martingales.
Vol. 2, Wiley Series in Probability and Mathematical Statistics: Probability and Mathe-
matical Statistics, John Wiley & Sons Inc., New York, 1987, Ito calculus. MR MR921238
(89k:60117)
[RY99] Daniel Revuz and Marc Yor, Continuous martingales and Brownian motion, third
ed., Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of
Mathematical Sciences], vol. 293, Springer-Verlag, Berlin, 1999. MR MR1725357
(2000h:60050)
[Str87] Daniel W. Stroock, Lectures on stochastic analysis: diusion theory, London Mathema-
tical Society Student Texts, vol. 6, Cambridge University Press, Cambridge, 1987. MR
MR884711 (88d:60203)
[SV06] Daniel W. Stroock and S. R. Srinivasa Varadhan, Multidimensional diusion processes,
Classics in Mathematics, Springer-Verlag, Berlin, 2006, Reprint of the 1997 edition. MR
MR2190038 (2006f:60005)
[vWW90] Heinrich von Weizsacker and Gerhard Winkler, Stochastic integrals, Advanced Lec-
tures in Mathematics, Friedr. Vieweg & Sohn, Braunschweig, 1990, An introduction
(Url: http://www.mathematik.uni-kl.de/wwwstoch/literatur.html). MR MR1062600
(91i:60150)
[Wat84] S. Watanabe, Lectures on stochastic dierential equations and Malliavin calculus, Ta-
ta Institute of Fundamental Research Lectures on Mathematics and Physics, vol. 73,
Published for the Tata Institute of Fundamental Research, Bombay, 1984, Notes by M.
Gopalan Nair and B. Rajeev. MR MR742628 (86b:60113)
[Wei87] Joachim Weidmann, Spectral theory of ordinary dierential operators, Lecture Notes in
Mathematics, vol. 1258, Springer-Verlag, Berlin, 1987. MR MR923320 (89b:47070)
134

Introduction To Stochastic Analysis: Lecture Notes

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Introduction To Stochastic Analysis: Lecture Notes

Uploaded by

Copyright:

Available Formats

Lecture Notes

Introduction to Stochastic Analysis

M is called the stochastic integral of H with respect to M.

M denote the uniquely determined element in /

is countable, since H is c`adl`ag (exercise). Dene for

is left-continuous. Hence, it is T-measurable and

M is the unique element L /

] (cf 2.3.18(i)), it follows that in

M, G.N = H dM, G.N = HG dM, N.

](= (M, N)) = 0 (i.e. orthogonal in

, P) separable (e.g. true, if T

is countably generated). Then there exist

P are the same. Hence, by Radon-Nikodym,

[f(x +y) f(x)[

for all t > 0 and f B

: C([0, 1]) C([0, 1])

()(t) = (t) + t, t [0, 1].

h is absolutely continuous and

and S B([0, 1]) such

(cf. functional analysis).Let h H. Then (by denition of absolute

h(s) ds t [0, 1],

P[T < ] = E[M

h is absolutely continuous and

:= X U is a Wiener process under P

of D. Note that domD

contains also non-(T

= exit time from open -ball K

(x) around x. Then there

, we have acoordingly to Blumenthals zero-one

(z)] = 0 > 0. Here K

(z) open ball.

() (:= inft > 0[X

Y (ok for n 2, since 0 R

W BM, since (loc.) martingale.

(X(u)), (s +u, X(u))dW(u))

(X(u)), b(s +u, X(u))

(X(u)), (s +u, X(u))dW(u))

= (x(t )[t 0). In particular /

be a regular conditional probability distribution of P given /

[A], , A /, is stochastic kernel on (, /) such that:

[A] for P a.a. .

(A) is /-measurable and (regular)

regular conditional probability distri-

= ] = 1. (This has been necessary as the

< ] > 0. But, one possibly

You might also like