You are on page 1of 142

INTRODUCTION TO STOCHASTIC

ANALYSIS
Giuseppe Da Prato
June 22, 2009
Contents
1 Gaussian measures in Hilbert spaces 3
1.1 Some concepts of Probability . . . . . . . . . . . . . . . . . . 3
1.1.1 Random variables . . . . . . . . . . . . . . . . . . . . . 3
1.1.2 Product measures . . . . . . . . . . . . . . . . . . . . . 5
1.2 Probability measures in Hilbert spaces . . . . . . . . . . . . . 5
1.2.1 Mean and covariance . . . . . . . . . . . . . . . . . . . 5
1.2.2 Finite dimensional projections of measures . . . . . . . 7
1.3 Gaussian probability measures . . . . . . . . . . . . . . . . . . 9
1.3.1 Gaussian probability measures in R . . . . . . . . . . . 9
1.3.2 Gaussian probability measures in R
n
. . . . . . . . . . 10
1.3.3 Gaussian probability measures in H . . . . . . . . . . . 11
1.3.4 Computation of some Gaussian integrals . . . . . . . . 11
1.3.5 The CameronMartin space . . . . . . . . . . . . . . . 13
2 Gaussian random variables 17
2.1 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.2.1 Independent real variables . . . . . . . . . . . . . . . . 18
2.2.2 Independent Gaussian random variables . . . . . . . . 21
2.3 Gaussian random variables dened in a Hilbert space . . . . . 21
2.3.1 Ane changes of variables . . . . . . . . . . . . . . . . 22
2.4 The white noise function . . . . . . . . . . . . . . . . . . . . . 23
2.4.1 Equivalence classes of random variables . . . . . . . . . 23
2.4.2 Denition of the white noise function . . . . . . . . . . 25
3 Brownian Motion 27
3.1 Stochastic Processes . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Brownian motion . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2.1 Construction of a Brownian motion . . . . . . . . . . . 29
3.2.2 Some properties of a Brownian motion . . . . . . . . . 29
3.3 Wiener integral . . . . . . . . . . . . . . . . . . . . . . . . . . 31
i
ii
3.4 Continuity of Brownian motion . . . . . . . . . . . . . . . . . 35
3.5 The standard Brownian motion . . . . . . . . . . . . . . . . . 36
3.5.1 Some properties of C
0
. . . . . . . . . . . . . . . . . . 37
3.5.2 The Wiener measure and the standard Brownian motion 37
3.6 Quadratic variation of the Brownian motion . . . . . . . . . . 39
3.7 Multidimensional Brownian motions . . . . . . . . . . . . . . . 41
4 Markov property of the Brownian motion 43
4.1 Filtration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.1.1 F
t
-measurable random variables . . . . . . . . . . . . . 44
4.2 Stopping times . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.3 The Brownian motion W(t + ) W() . . . . . . . . . . . . 49
4.4 Transition semigroup . . . . . . . . . . . . . . . . . . . . . . . 50
4.5 Markov property . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.5.1 Strong Markov property . . . . . . . . . . . . . . . . . 52
4.6 Some consequences of the strong Markov property . . . . . . . 53
4.7 Application to partial dierential equations . . . . . . . . . . . 56
4.7.1 The Dirichlet problem in the half-line . . . . . . . . . . 57
4.7.2 The Neumann problem . . . . . . . . . . . . . . . . . . 58
4.7.3 The Ventzell problem . . . . . . . . . . . . . . . . . . . 59
5 The Ito integral 61
5.1 Denition of Itos integral . . . . . . . . . . . . . . . . . . . . 61
5.1.1 Itos integral for elementary processes . . . . . . . . . . 61
5.1.2 General denition of Itos integral . . . . . . . . . . . . 63
5.2 It o integral for mean square continuous processes . . . . . . . 66
5.3 The Ito integral as a stochastic process . . . . . . . . . . . . . 67
5.4 It o integral with stopping times . . . . . . . . . . . . . . . . . 70
5.4.1 Stopping times . . . . . . . . . . . . . . . . . . . . . . 70
5.4.2 Itos integral with stopping times . . . . . . . . . . . . 71
5.5 Multidimensional It o integrals . . . . . . . . . . . . . . . . . . 72
6 The Ito formula 75
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
6.1.1 The It o formula for unbounded functions . . . . . . . . 82
6.2 It o formula for a vector valued process . . . . . . . . . . . . . 84
7 Stochastic evolution equations 89
7.1 Existence and uniqueness . . . . . . . . . . . . . . . . . . . . . 90
7.1.1 Solution of the stochastic dierential equation in the
space C
B
([s, T]; L
2m
(; R
d
)). . . . . . . . . . . . . . . 94
1
7.1.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . 94
7.1.3 Dierential stochastic equations with random coecients 96
7.2 Continuous dependence on data . . . . . . . . . . . . . . . . . 97
7.2.1 Continuous dependence on mean square . . . . . . . . 97
7.3 Almost sure continuity and holderianity of trajectories . . . . 100
7.4 Dierentiability of X(t, s, x) with respect to x . . . . . . . . . 101
7.4.1 Existence of X
x
(t, s, x) . . . . . . . . . . . . . . . . . . 101
7.4.2 Existence of X
xx
(t, s, x) . . . . . . . . . . . . . . . . . 102
7.5 It o Dierentiability of X(t, s, x) with respect to s. . . . . . . . 105
7.5.1 The deterministic case . . . . . . . . . . . . . . . . . . 105
7.5.2 The stochastic case . . . . . . . . . . . . . . . . . . . . 106
7.5.3 Backward Itos formula . . . . . . . . . . . . . . . . . . 107
8 Kolmogorov equations 111
8.1 The deterministic case . . . . . . . . . . . . . . . . . . . . . . 111
8.1.1 The autonomous case . . . . . . . . . . . . . . . . . . . 113
8.2 Stochastic case . . . . . . . . . . . . . . . . . . . . . . . . . . 114
8.3 Basic properties of transition operators . . . . . . . . . . . . . 115
8.4 Parabolic equations . . . . . . . . . . . . . . . . . . . . . . . . 116
8.4.1 Autonomous case . . . . . . . . . . . . . . . . . . . . . 117
8.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
A -systems and -systems 121
B Conditional expectation 123
B.1 Denition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
B.2 Basic properties . . . . . . . . . . . . . . . . . . . . . . . . . . 124
C Martingales 127
C.1 Denitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
C.2 The basic inequality for martingales . . . . . . . . . . . . . . . 128
C.3 Square integrable martingales . . . . . . . . . . . . . . . . . . 129
D Fixed points depending on parameters 133
D.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
D.2 G ateaux dierentiable mappings . . . . . . . . . . . . . . . . . 134
D.3 The main result . . . . . . . . . . . . . . . . . . . . . . . . . . 135
E Fractional Sobolev spaces and regularity of processes 137
E.1 Fractional Sobolev spaces on [0, 1] . . . . . . . . . . . . . . . . 137
E.2 Processes belonging to W
,2m
(0, T) . . . . . . . . . . . . . . . 138
2
E.3 Multi dimensional Sobolev spaces and regularity of random
elds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Chapter 1
Gaussian measures in Hilbert
spaces
We shall denote by H a real separable Hilbert space (with inner product
, ) and norm [ [), and by L(H) the Banach algebra of all linear bounded
operators T : H H, endowed with the norm
|T| = sup
xH, |x|=1
[Tx[.
We recall that T L(H) is said to be symmetric if Tx, y) = x, Ty) for all
x, y H, positive if Tx, x) 0 for all x H. The set of all symmetric and
positive elements of L(H) will be denoted by L
+
(H).
Finally, we shall denote by C
b
(H) the space of all functions : H R
which are continuous and bounded. C
b
(H), endowed with the norm
||
0
: = sup
xH
[(x)[,
is a Banach space.
Next section is devoted to some basic facts from Measure Theory and
Probability needed in what follows.
1.1 Some concepts of Probability
1.1.1 Random variables
Let (, F, P) be a probabilty space and let E be a Polish (complete separable
metric) space; we shall denote by B(E) the algebra generated by all closed
(or equivalently open) subsets of E. The elements of B(E) are called Borel
sets.
3
4 Chapter 1
By an E-valued random variable in (, F) we mean a mapping
X: E, X(),
such that
I B(E) X
1
(I) F.
The law (or image measure or push-forward measure) of X is the probability
measure X
#
P on (E, B(E)) dened as
(X
#
P)(I) = P(X
1
(I)), I B(E).
Sometimes we shall use the notation X
#
P = P
X
.
Let us prove the following basic change of variables formula.
Theorem 1.1 Let X be an E-valued random variable in (, F, P). Let
moreover : E R be a nonnegative Borel function. Then we have
_

(X())P(d) =
_
E
(x)(X
#
P)(dx). (1.1)
Proof. Let rst = 1l
I
with I B(E)
(1)
. In this case we have
(X()) = 1l
X
1
(I)
(), .
So,
_

(X())P(d) = P(X
1
(I)) = X
#
P(I) =
_
E
(x)X
#
P(dx).
Consequently, (1.1) holds for all simple functions of the form
=
n

i=1
c
i
1l
I
i
,
with n N, c
1
, ..., c
n
0 and I
1
, ..., I
n
B(E). Since any positive Borel
functions is the limit of an increasing sequence of positive simple functions,
the conclusion follows from the monotone convergence theorem.
(1)
1l
I
() is the characteristic function of I; it is equal to 1 if I to 0 if / I.
Gaussian measures 5
1.1.2 Product measures
Let (
i
, F
i
, P
i
), i = 1, ..., n, be probability spaces. Set =

n
i=1

i
. A mea-
surable rectangle of is, by denition, a set of the form R =

n
i=1
A
i
where
A
i
F
i
, i = 1, 2, ..., n. The -algebra generated by all measurable rectangles
is called the product -algebra of F
i
, ..., F
n
; it is denoted by

n
i=1
F
i
.
For any R =

n
i=1
A
i
we dene
P(R) :=
n

i=1
P
i
(A
i
).
One can show that P can be uniquely extended to a probability measure on
(, F) which is called the product probability of P
1
, P
2
, ..., P
n
.
1.2 Probability measures in Hilbert spaces
1.2.1 Mean and covariance
Let be a probability measure on (H, B(H)). Assume that has nite rst
momentum,
_
H
[x[(dx) < +.
Then the linear functional F : H R dened as
F(h) =
_
H
x, h)(dx), h H,
is continuous since
[F(h)[
_
H
[x[(dx) [h[, h H.
By the Riesz representation theorem there exists m H such that
m, h) =
_
H
x, h)(dx), h H.
m is called the mean of . We shall write
m =
_
H
x(dx).
Assume now that the second moment of is nite,
_
H
[x[
2
(dx) < +,
6 Chapter 1
(so that the rst one is nite as well). Let us consider the bilinear form
G : H H R dened as
G(h, k) =
_
H
h, x m)k, x m)(dx), h, k H.
G is continuous since
[G(h, k)[
_
H
[x m[
2
(dx) [h[ [k[, h, k H.
Therefore there is a unique linear bounded operator Q L(H) such that
Qh, k) =
_
H
h, x m)k, x m)(dx), h, k H.
Q is called the covariance of .
In order to state the next result we need the concept of trace class op-
erator. A symmetric and positive operator Q L(H) is said to be of trace
class if
Tr Q: =

k=1
Qe
k
, e
k
) < +
for one (and consequently for any) complete orthonormal system (e
k
). One
can show that any trace class operator Q is compact and that Tr Q is the
sum of its eigenvalues repeated according to their multiplicity, see e. g. N.
Dunford and J.T. Schwartz, Linear Operators. Part II, Interscience, 1964.
(2)
Proposition 1.2 The covariance operator Q of is symmetric, positive and
of trace class.
Proof. Symmetry and positivity of Q are clear. To prove that Q is of trace
class choose a complete orthonormal system (e
k
) in H. Then we have
Qe
k
, e
k
) =
_
H
[x m, e
k
)[
2
(dx), k N.
Therefore, by the monotone convergence theorem and the Parseval identity,
we nd that
Tr Q =

k=1
_
H
[x m, e
k
)[
2
(dx) =
_
H
[x m[
2
(dx) < +.
(2)
It is also possible to dene trace-class operators which are not symmetric, but we shall
not need in what follows.
Gaussian measures 7

We shall denote by L
+
1
(H) the set of all positive, symmetric operators in
H of trace class.
We nally dene the Fourier transform of a probability measure
setting
(h) =
_
H
e
ix,h
(dx), h H. (1.2)
One checks easily that : H C is continuous.
1.2.2 Finite dimensional projections of measures
We are given a probability measure P(H). Let (e
k
) be a complete or-
thonormal system in H. For any n N we consider the projection P
n
: H
P
n
(H) dened as
P
n
x =
n

k=1
x, e
k
)e
k
, x H. (1.3)
We have lim
n
P
n
x = x for all x H.
For any n N we consider the measure
n
:= (P
n
)
#
dened by
_
H
(P
n
x)(dx) =
_
H
n
(y)
n
(dy),
for all C
b
(R).
Thus
n
is a probability measure on (P
n
(H), B(P
n
(H)),
n
). We shall
also consider
n
as a probability measure on (H, B(H), ), setting

n
(I) =
n
(I P
n
(H)), I B(H).
We want now to show that is determined by the sequence (
n
). For this
we rst need the following result.
Proposition 1.3 Let , P(H) be such that
_
H
(x)(dx) =
_
H
(x)(dx), C
b
(H). (1.4)
Then = .
Proof. Let C H be closed and let (
n
) C
b
(H) be such that
(i) lim
n

n
(x) = 1l
C
(x) for all x H.
8 Chapter 1
(ii) |
n
|
0
1 for all N.
A sequence (
n
) C
b
(H) fullling (i) and (ii) is provided by,

n
(x) =
_
_
_
1 if x C,
1 n d(x, C) if d(x, C)
1
n
0 if d(x, C)
1
n
.
Now, by the dominate convergence theorem it follows that
lim
n
_
H

n
d = lim
n
_
H

n
d = (C) = (C).
Since closed sets generate the Borel algebra of H this implies that = .

We can now prove the announced result.


Proposition 1.4 Let , P(H). If (P
n
)
#
= (P
n
)
#
for any n N we
have = .
Proof. Let C
b
(H). Then, using the dominated convergence theorem and
the change of variables formula, we have
_
H
(x)(dx) = lim
n
_
H
(P
n
x)(dx) = lim
n
_
P
n
(H)
()((P
n
)
#
)(d)
and
_
H
(x)(dx) = lim
n
_
H
(P
n
x)(dx) = lim
n
_
P
n
(H)
()((P
n
)
#
)(d).
Since (P
n
)
#
= (P
n
)
#
by assumption, we conclude that
_
H
(x)(dx) =
_
H
(x)(dx)
for all C
b
(H). Therefore, in view of Proposition 1.3 we have = .
As an application of Proposition 1.4 we prove that the Fourier transform
of determines .
Proposition 1.5 Let , P(H) be such that (h) = (h) for all h H.
Then = .
Gaussian measures 9
Proof. We assume as granted the result when H is nite-dimensional
(3)
. In
the general case we have by (1.1) for any h H and n N,
(P
n
h) =
_
H
e
ix,P
n
h
(dx) =
_
P
n
(H)
e
iP
n
,P
n
h
(P
n
)
#
(d) =

(P
n
)
#
(P
n
h)
and
(P
n
h) =
_
H
e
ix,P
n
h
(dx) =
_
P
n
(H)
e
iP
n
,P
n
h
(P
n
)
#
(d) =

(P
n
)
#
(P
n
h).
Therefore measures (P
n
)
#
and (P
n
)
#
have the same Fourier tranforms and
so they coincide. The conclusion follows from Proposition 1.4.
1.3 Gaussian probability measures
We rst recall the denition of Gaussian measure on (R, B(R)), then we go
to the general case.
1.3.1 Gaussian probability measures in R
For any pair of real numbers (m, q) with m R and q 0 we dene a
probability measure N
m,q
on (R, B(R)) as follows. If q = 0 we set
N
m,0
=
m
,
where
m
is the Dirac measure at m, dened for all B B(R) by

m
(B) =
_
_
_
1 if m B,
0 if m / B.
If q > 0 we set
N
m,q
(B) =
1

2q
_
B
e

(xm)
2
2q
dx, for all B B(R).
N
m,q
is a probability measure since
N
m,q
(R) =
1

2q
_
+

(xm)
2
2q
dx =
1

2
_
+

x
2
2
dx = 1.
(3)
See e.g. M. Metivier, Notions fondamentales de la theorie des probabilitees, Dunod
Universite, 1968.
10 Chapter 1
If q > 0, N
m,q
is absolutely continuous with respect to the Lebesgue measure

1
(dx) = dx in (R, B(R) and
N
m,q
(dx) =
1

2q
e

(xm)
2
2q
dx.
When m = 0 we shall write for short N
q
instead N
0,q
.
It is easy to see that m is the mean and q the covariance of N
m,q
. Moreover,
its Fourier transform is given by

N
m,q
(h) :=
_
R
e
ihx
N
m,q
(dx) = e
imh
1
2
qh
2
, h R. (1.5)
1.3.2 Gaussian probability measures in R
n
We are going to dene a Gaussian measure N
m,Q
for any m = (m
1
, ..., m
n
)
R
n
and any Q L
+
(R
n
).
Let Q L
+
(R
n
) and let (e
1
, ..., e
n
) be an orthonormal basis on R
n
such
that Qe
k
=
k
e
k
, k = 1, ..., n, for some
k
0. Then we dene a probability
measure N
a,Q
on (R
n
, B(R
n
)) by setting
N
m,Q
=
n

k=1
N
m
k
,
k
.
When m = 0 we shall write N
Q
instead of N
m,Q
for short.
The proof of the following proposition is easy; it is left to the reader.
Proposition 1.6 Let m R
n
, Q L
+
(R
n
) and = N
m,Q
. Then we have
_
R
n
x(dx) = m,
_
R
n
y, x a)z, x a)(dx) = Qy, z), y, z R
n
.
Moreover the Fourier tranform of N
a,Q
is given by

N
a,Q
(h) :=
_
R
n
e
ih,x
(dx) = e
ia,h
1
2
Qh,h
, h R
n
.
Finally, if the determinant of Q is positive, N
a,Q
is absolutely continuous
with respect to the Lebesgue measure in R
n
and we have
N
a,Q
(dx) =
1
_
(2)
d
det Q
e

1
2
Q
1
(xa),xa
dx.
Therefore m is the mean and Q the covariance operator of N
a,Q
.
Gaussian measures 11
1.3.3 Gaussian probability measures in H
Let m H and Q L
+
1
(H). We denote by N
m,Q
the probability measure on
(H, B(H)) of mean m, covariance Q and Fourier transform given by

N
m,Q
(h) = e
im,h
1
2
Qh,h
, h H. (1.6)
One can show that such a measure does exist
(4)
; it is unique thanks to
Proposition 1.5.
1.3.4 Computation of some Gaussian integrals
To compute some integrals with respect to a Gaussian measure = N
m,Q
in
an innite dimensional Hilbert space H it is useful to reduce the computation
to integrals on a sequence (H
n
) of nite dimensional vector spaces convergent
to H and then to let n .
More precisely, given = N
m,Q
P(H), we shall proceed as follows.
Since Q is compact there exists an orthonormal complete system (e
k
) in H
and a sequence of nonnegative numbers (
k
) such that
Qe
k
=
k
e
k
, k N.
For any n N we set m
n
:= m, e
n
),
P
n
x =
n

k=1
x, e
k
)e
k
, x H
and identify P
n
(H) with R
n
through the isomorphism,
P
n
(H) R
n
, x =
n

k=1
x, e
k
)e
k
(x, e
1
), ..., x, e
n
)).
Exercise 1.7 Prove that

n
= (P
n
)
#
=
n

i=1
N
m
k
,
k
.
Hint. Show that the Fourier transform of
n
is given by

n
(h) = e
i
P
n
k=1
m
k
h
k
e

1
2
P
n
k=1

k
h
2
k
.
(4)
see e.g. G. Da Prato, An introduction to innite-dimensional analysis. Springer-
Verlag, Berlin, 2006.
12 Chapter 1
We shall assume (which is always true after a rearrangement) that
1

2

n
.
To formulate the next result notice that for any <
1

1
, the linear operator
1 Q is invertible and (1 Q)
1
is bounded. We have in fact, as easily
checked,
(1 Q)
1
x =

k=1
1
1
k
x, e
k
)e
k
, x H.
In this case we can dene the determinant of (1 Q) by setting
det(1 Q): = lim
n
n

k=1
(1
k
) :=

k=1
(1
k
).
Exercise 1.8 Prove that

k=1
(1
k
) > 0.
Hint. Write
log
_

k=1
(1
k
)
_
=

k=1
log(1
k
)
and show that the series is convergent since

k=1

k
< +.
Proposition 1.9 Let R. Then we have
_
H
e

2
|x|
2
(dx) =
_
_
_
[det(1 Q)]
1/2
e

2
(1Q)
1
m,m
, if <
1

1
,
+, otherwise.
(1.7)
Proof. For any n N we have, taking into account Exercise 1.7
_
H
e

2
|P
n
x|
2
(dx) =
_
P
n
(H)
e

2
|P
n
|
2

n
(d) =
n

k=1
_
R
e

2
k
N
m
k
,
k
(d
k
).
Since [P
n
x[
2
[x[
2
as n and, by an elementary computation,
_
R
e

2
x
2
k
N
m
k
,
k
(dx
k
) =
1

1
k
e

2
m
2
k
1
k
,
the conclusion follows from the monotone convergence theorem.
Gaussian measures 13
Exercise 1.10 Prove that for all m N
J
m
:=
_
H
[x[
2m
(dx) <
and compute J
m
.
Hint. Notice that J
m
= 2
m
F
(m)
(0), where
F() =
_
H
e

2
|x|
2
(dx), > 0.
Proposition 1.11 We have
_
H
e
h,x
(dx) = e
a,h
e
1
2
Qh,h
, h H. (1.8)
Proof. For any > 0 we have
e
h,x
e
|x| |h|
e
|x|
2
e
1

|h|
2
.
Choosing <
1

1
, we have, by the dominated convergence theorem, that
_
H
e
h,x
(dx) = lim
n
_
H
e
h,P
n
x
(dx) = lim
n
_
P
n
(H)
e
h,P
n

n
(dx)
= lim
n
e
P
n
m,h
e
1
2
P
n
Qh,h
= e
m,h
e
1
2
Qh,h
.

1.3.5 The CameronMartin space


We are given a Gaussian measure = N
Q
, where Q L
+
1
(H). We say that
is non degenerate if Ker Q := x H : Qx = 0 = 0. Thus, if H is
nite-dimensional is non degenerate if and only if det Q > 0.
Assume now that H is innite-dimensional and that is non degenerate.
We denote by (e
k
) a complete orthonormal system in H such that Qe
k
=

k
e
k
, k N, where (
k
) are the eigenvalues of Q and we set x
k
= x, e
k
), k
N.
We notice that the inverse Q
1
of Q (which is well dened since Ker
Q = 0) is not continuous because,
Q
1
e
k
=
1

k
e
k
, k N
and
k
0 as k . Consequently, recalling the closed graph theorem,
we see that the range Q(H) does not coincide with H. However, it is dense
in H as the following lemma shows.
14 Chapter 1
Lemma 1.12 Q(H) is a dense subspace of H.
Proof. In fact if x
0
is an element of H orthogonal to Q(H), we have
Qx, x
0
) = x, Qx
0
) = 0, x H,
which yields Qx
0
= 0, and so x
0
= 0 because Ker(Q) = 0.
It is useful to introduce the operator Q
1/2
dened as
Q
1/2
x =

k=1
_

k
x, e
k
)e
k
, x H.
Its range Q
1/2
(H) is called the CameronMartin space of the measure .
Arguing as before we see that Q
1/2
(H) is a subspace of H dierent of H and
dense in H. Moreover it is clear that x Q
1/2
(H) if and only if,

k=1

1
k
x
2
k
< +.
It is important to notice that the measure of the CameronMartin space
is zero.
Proposition 1.13 We have (Q
1/2
(H)) = 0.
Proof. For any n, k N set
U
n
=
_
y H :

h=1

1
h
y
2
h
< n
2
_
= y Q
1/2
(H) : [Q
1/2
y[ < n,
and
U
n,k
=
_
y H :
2k

h=1

1
h
y
2
h
< n
2
_
.
Clearly U
n
Q
1/2
(H) as n , and for any n N, U
n,k
U
n
as k .
So, it is enough to show that
(U
n
) = lim
k
(U
n,k
) = 0. (1.9)
We have in fact
(U
n,k
) =
_

yH:
P
2k
h=1

1
h
y
2
h
<n
2

2k

h=1
N

k
(dy
k
),
Gaussian measures 15
which, setting z
h
=
1/2
h
y
h
is equivalent to
(U
n,k
) =
_

zR
2k
:|z|<n

N
I
2k
(dz),
where I
2k
is the identity in R
2k
. Let us compute (U
n,k
). We have
(U
n,k
) =
(U
n,k
)
(H)
=
_
n
0
e

r
2
2
r
2k1
dr
_
+
0
e

r
2
2
r
2k1
dr
=
_
n
2
/2
0
e

k1
d
_
+
0
e

k1
d
.
Therefore
(U
n,k
) =
1
(k 1)!
_
n
2
/2
0
e

k1
d
1
(k 1)!
_
n
2
/2
0

k1
d =
1
k!
_
n
2
2
_
k
,
and (1.9) follows.
16 Chapter 1
Chapter 2
Gaussian random variables
2.1 Notations
Let (, F, P) be a probability space, H a separable Hilbert space, X: H
a random variable such that
_

[X()[
2
P(d) < .
We denote by X
#
P the law of X, by m(X) the mean of X
#
P and by Q(X)
the covariance of X
#
P.
By the change of variables formula it follows that the Fourier transform
of X
#
P is given by

X
#
P(h) =
_

e
iX(),h
P(d), h H
and that
m(X), h) =
_

X(), h)P(d), h H,
and
Q(X)h, k) =
_

X() m(X), h) X() m(X), k)P(d), h, k H.


Denition 2.1 We say that X
#
P is a Gaussian random variable if X
#
P is
a Gaussian measure, that is if

X
#
P(h) = e
im(X),h
e

1
2
Q(X)h,h
, h H.
In this case we call m(X) the mean and Q(X) the covariance of X.
17
18 Chapter 2
Example 2.2 Let n N, X
1
, ..., X
n
be real random variables on (, F, P).
Then X = (X
1
, ..., X
n
) is a R
n
-valued random variable. So, m(X) is a vector
of R
n
denoted by (m(X)
1
, ..., m(X)
n
) and Q(X) is a n n matrix denoted
Q(X)
i,j
, i, j = 1, ..., n.
More precisely, let (e
1
, ..., e
n
) be the canonical basis in R
n
. Then for any
k = 1, ..., n we have
m(X)
k
= m(X), e
k
) =
_

X
k
()P(d) = m(X
k
)
and for any j, k = 1, ..., n we have
Q(X)
j,k
= Q(X)e
j
, e
k
) =
_

(X
j
() m
j
(X
j
))(X
k
() m
k
(X
k
))P(d).
In particular, if j = k we nd
Q(X)
k,k
= Q(X
k
), k = 1, ..., n.
Example 2.3 Assume that X = (X
1
, ..., X
n
) is a n-dimensional Gaussian
random variable. Then X
1
, ..., X
n
are real Gaussian random variables. In
fact if k = 1, ..., n and a R we have
_

e
iaX
k
()
P(d) =
_

e
iae
k
,X()
P(d)
= e
iae
k
,m(X)
e

1
2
a
2
Q(X)e
k
,e
k

= e
iam(X
k
)
e

1
2
a
2
Q(X
k
)
.
Notice that, if conversely X
1
, ..., X
n
are real Gaussian random variables, then
X = (X
1
, ..., X
n
) is not necessarily Gaussian.
2.2 Independence
In this section we introduce the basic concept of independence.
2.2.1 Independent real variables
Denition 2.4 Let n N and let X
1
, ..., X
n
be real random variables in
(, F, P). Consider the R
n
-valued random variable
X() = (X
1
(), ..., X
n
()), .
random variables 19
We say that X
1
, ..., X
n
are independent if
X
#
P =
n

j=1
(X
j
)
#
P.
Let (X
i
) be a sequence of real random variables. They are called independent
if X
i
1
, . . . , X
i
n
are independent for any choice of n and of positive integers
i
1
< i
2
< < i
n
.
A necessary and sucient condition for the independence is provided by
the following proposition.
Proposition 2.5 Let X
1
, ..., X
n
, n N, be real independent random vari-
ables in (, F, P). Let moreover
1
, ...,
n
be Borel positive functions. Then
we have
_

1
(X
1
())
n
(X
n
())P(d)
=
_

1
(X
1
())P(d)
_

n
(X
n
())P(d).
(2.1)
Conversely, if (2.1) holds for any choice of positive Borel functions
1
, ...,
n
,
then X
1
, ..., X
n
are independent.
Proof. Set X = (X
1
, ..., X
n
) and let : R
n
R be dened as
(
1
, ...,
n
) =
1
(
1
)
k
(
n
), (
1
, ...,
n
) R
n
.
Then by the change of variable formula we have, taking into account the
independence of X
1
, ..., X
n
,
_

1
(X
1
())
n
(X
n
())P(d) =
_

(X())P(d)
=
_
R
n
()(X
#
P)(d) =
_
R

1
(
1
)((X
1
)
#
P)(d
1
)
_
R

k
(
n
)((X
n
)
#
P)(d
n
)
=
_

1
(X
1
())P(d)
_

n
(X
n
())P(d).
Assume conversely that (2.1) holds for any choice of functions
1
, ...,
n
positive Borel. To prove independence of X
1
, ..., X
n
it is enough to show that
(X
#
P)(I
1
I
n
) = ((X
1
)
#
P)(I
1
) ((X
n
)
#
P)(I
n
), I
1
, ..., I
n
B(R).
But this follows immediately setting in (2.1)

i
= 1l
I
i
, i = 1, ..., n.

20 Chapter 2
Exercise 2.6 Let X
1
, ..., X
n
be real independent random variables in (, F, P).
Show that
_

X
1
X
n
dP =
_

X
1
dP
_

X
n
dP
and
V (X
1
+ + X
n
) = V (X
1
) + + V (X
n
).
The following useful result is left to the reader as an exercise.
Proposition 2.7 Let X
1
, ..., X
n
be real random variables in (, F, P) and
let X = (X
1
, ..., X
n
). Then X
1
, ..., X
n
are independent if and only if

X
#
P(h) =
n

k=1

(X
k
)
#
P(h
k
), h = (h
1
, ..., h
n
) R
n
.
Denition 2.8 Let (, F, P) be a probability space and A
1
, ..., A
n
F.
We say that the sets A
1
, ..., A
n
are independent if the random variables
1l
A
1
, ..., 1l
A
n
are so.
Exercise 2.9 Show that sets A
1
, ..., A
n
are independent if and only if
P(A
j
1
A
j
k
) = P(A
j
1
) P(A
j
k
),
for all k = 1, ..., n and k dierent positive integer j
1
, ..., j
k
less or equal to n.
Proposition 2.10 Let X
1
, ..., X
n
be real independent random variables in
(, F, P) and let X = (X
1
, ..., X
n
). Then the covariance matrix Q(X) is
diagonal.
Proof. We have in fact (by Exercise 2.6) for i, j = 1, ..., n
Q(X)
i,j
=
_

(X
i
() m
i
(X))(X
j
() m
j
(X))P(d)
=
_

(X
i
() m
i
(X))P(d)
_

(X
j
() m
j
(X))P(d) = 0.
The converse of Proposition 2.10 does not hold in general.
random variables 21
2.2.2 Independent Gaussian random variables
Let X
1
, ..., X
n
be real random variables in (, F, P) and let X = (X
1
, ..., X
n
).
Proposition 2.11 Assume that X
1
, ..., X
n
are independent Gaussian ran-
dom variables. Then X = (X
1
, ..., X
n
) is Gaussian.
Proof. In fact, let h = (h
1
, ..., h
n
) R
n
. Then, taking into account the
independence of (X
1
, ..., X
n
),

X
#
P(h) =
_

e
i(X
1
()h
1
++X
1
()h
n
)
P(d) =
n

k=1
_

e
iX
k
()h
k
P(d)
= e
i(m(X
1
)h
1
++m(X
n
)h
n
)
e

1
2
(Q(X
1
)h
2
1
++Q(X
n
)h
2
n
)
.

Proposition 2.12 Assume that X


1
, ..., X
n
are real random variables and
that X = (X
1
, ..., X
n
) is Gaussian. Then X
1
, ..., X
n
are independent if and
only if Q(X) is diagonal.
Proof. If X
1
, ..., X
n
are independent the conclusion follows from Proposition
2.11. Assume now that Q(X) is diagonal. By Proposition 2.7 it is enough to
show that

X
#
P(h) =
n

i=1

(X
k
)
#
P(h),
for each h = (h
1
, ..., h
n
) H.
We have in fact

X
#
P(h) = e
im(X),h
e

1
2
Q(X)h,h
= e
im(X),h
e

1
2
P
n
k=1
Q(X)
k,k
h
2
k
= e
im(X),h
e

1
2
P
n
k=1
Q(X
k
)h
2
k
=
n

i=1

(X
k
)
#
P(h).

2.3 Gaussian random variables dened in a


Hilbert space
We now consider the case when (, F, P) coincides with (H, B(H), ), where
H is a separable Hilbert space and = N
m,Q
with m H and Q L
+
1
(H).
22 Chapter 2
2.3.1 Ane changes of variables
Let b K and A L(H, K) where K is another separable Hilbert space.
Let us consider the ane transformation
T(x) = Ax + b, x H.
Proposition 2.13 T is a Gaussian random variable and its law T
#
is given
by N
Aa+b,AQA
, where A

is the transpose of A.
Proof. We have in fact
_
K
e
ik,y
T
#
(dy) =
_
H
e
ik,T(x)
(dx) =
_
H
e
ik,Ax+b
(dx)
= e
ik,b
_
H
e
iA

k,x
(dx) = e
ik,Aa+b
e

1
2
AQA

k,k
, k K.

Example 2.14 Let = N


m,Q
and n N, f
1
, ..., f
n
H. Let F : H R
n
be dened as
F(x) := (x, f
1
), ..., x, f
n
)), x H.
Then by Proposition 2.13 F is a Gaussian random variable with mean m(F)
and covariance Q(F) given by,
m(F) = F(m) = (m, f
1
), ..., m, f
n
))
and
Q(F) = FQF

.
On the other hand, the linear operator F

: R
n
H is given by
F

() =
n

k=1
f
k

k
, = (
1
, ...,
n
) R
n
.
Therefore
QF

() =
n

k=1
Qf
k

k
, = (
1
, ...,
n
) R
n
and
FQF

() =
__
n

k=1
Qf
k

k
, f
1
_
, ...,
_
n

k=1
Qf
k

k
, f
n
__
random variables 23
so that
Q(F)
h,k
= Qf
h
, f
k
). (2.2)
Therefore, F
1
, ..., F
n
are independent if and only if
Qf
h
, f
k
) = 0, h, k = 1, ..., n,
if h ,= k.
2.4 The white noise function
In order to dene the white noise function (which will play an important role
in what follows), we shall deal with equivalence class of random variables
(rather than random variables), which we briey discuss in the next sub-
section.
2.4.1 Equivalence classes of random variables
Let (, F, P) be a probability space and let H be a separable Hilbert space.
We denote by R(H) the set of all H-valued random variables.
Denition 2.15 We say that X, Y R(H) are equivalent (and write X
Y ) if
P( : X() = Y ()) = 1.
One can easily check that X Y, X, Y R(H) is an equivalence relation,
so that the set R(H) is disjoint union of equivalences classes.
We notice that if X Y then the laws of X and Y coincide. In fact set
K = : X() ,= Y (),
so that P(K) = 0. Since for any I B(H) we have
X
1
(I) Y
1
(I) K,
it follows that P(X
1
(I)) P(Y
1
(I)) and, exchanging X and Y we see that
P(X
1
(I)) = P(Y
1
(I)).
Consequently, all random variables belonging to a xed equivalence class

X have the same law, which is called the law of



X.
In the following we shall not distinguish between a random variable X
and the equivalence class

X including X, except when needed.
24 Chapter 2
By L
p
(, F, P; H), p 1, we mean the space of all equivalence class of
random variables X: H such that
_

[X()[
p
P(d) < +.
L
p
(, F, P; H), endowed with the norm
|X|
L
p
(,F,P;H)
=
__

[X()[
p
P(d)
_
1/p
,
is a Banach space. We shall write L
p
(, F, P; H) = L
p
(, P; H) for brevity.
We prove now that the limit of a convergent sequence in L
2
(, P; H) of
Gaussian random variables is Gaussian.
Proposition 2.16 Let (X
n
) L
2
(, P; H) be a sequence of Gaussian ran-
dom variables convergent to X in L
2
(, P; H). Then X is a Gaussian random
variable and
m(X), h) = lim
n
m(X
n
), h), h H,
and
Q(X)h, k) = lim
n
Q(X
n
)h, k), h, k H.
Proof. Since X
n
X in L
2
(, P; H) we have
lim
n
m(X
n
), h) = lim
n
_

X
n
(), h)P(d) =
_

X(), h)P(d) = m(X), h)


and
lim
n
Q(X
n
)h, k) = lim
n
_

X
n
() m(X
n
), h) X
n
() m(X
n
), k)P(d)
=
_

X() m(X), h) X() m(X), k)P(d) = Q(X)h, k).


Let us show now that X is a Gaussian random variable. We have in fact
_
H
e
ix,h
(X
#
)P(dy) =
_

e
iX(),h
P(d) = lim
n
_

e
iX
n
(),h
P(d)
= lim
n
e
im(X
n
),h
e

1
2
Q(X
n
)h,h
= e
im(X),k
e

1
2
Q(X)h,h
.

random variables 25
2.4.2 Denition of the white noise function
In this section we assume that the Hilbert space H is innite dimensional and
consider a non degenerate Gaussian measure = N
Q
in H (Ker (Q) = 0).
Since Q is compact there exists a complete orthonormal basis (e
k
) on H and
a sequence of positive numbers (
k
) such that
Qe
k
=
k
e
k
, k N.
Let us dene a mapping
W : Q
1/2
(H) C(H), z W
z
where
W
z
(x) = x, Q
1/2
z), x H.
Here Q
1/2
(H) is the CameronMartin space and C(H) the space of all real
continuous functions on H.
Lemma 2.17 For all z
1
, z
2
Q
1/2
(H) we have
_
H
W
z
1
(x)W
z
2
(x)(dx) = z
1
, z
2
). (2.3)
Proof. We have in fact
_
H
W
z
1
(x)W
z
2
(x)(dx) =
_
H
x, Q
1/2
z
1
)x, Q
1/2
z
2
)(dx)
= QQ
1/2
z
1
, QQ
1/2
z
2
) = z
1
, z
2
).

Since Q
1/2
(H) is dense in H, the mapping W can be uniquely extended
as a mapping from H into L
2
(H, ) which we denote still by W and call the
white noise function.
W
f
is linear in the sense that for all , R we have
W
f
(x + y) = W
f
(x) + W
f
(y), x, y a.e..
Remark 2.18 Given z H (not belonging to Q
1/2
(H)) it would be tempt-
ing to dene the random variable W
z
by setting,
W
z
(x) = Q
1/2
x, z), x Q
1/2
(H).
However this denition is meaningless because (Q
1/2
(H)) = 0, by Proposi-
tion 1.13
26 Chapter 2
Proposition 2.19 Let z H. Then W
z
is a real Gaussian random variable
with mean 0 and covariance [z[
2
.
Proof. We have to show that
_
H
e
iW
z
(x)
(dx) = e

1
2

2
|z|
2
, R.
Let (z
n
) Q
1/2
(H) be a sequence such that z
n
z in H. Then, by the
dominated convergence theorem, we have
_
H
e
iW
z
(x)
(dx) = lim
n
_
H
e
iQ
1/2
z
n
,x
(dx) = lim
n
e

1
2

2
|z
n
|
2
= e

1
2

2
|z|
2
.
So, the conclusion follows.
The following generalization of Proposition 2.19 is important.
Proposition 2.20 Let n N, z
1
, ..., z
n
H. Then (W
z
1
, ..., W
z
n
) is an n-
dimensional Gaussian random variable with mean 0 and covariance operator
Q
z
given by
(Q
z
)
h,k
= z
h
, z
k
), h, k = 1, ..., n. (2.4)
The random variables W
z
1
, ..., W
z
n
are independent if and only if z
1
, ..., z
n
are
mutually orthogonal.
Proof. Let (z
1
j
), ..., (z
n
j
) be n sequences in Q
1/2
(H) convergent respectively to
z
1
, ..., z
n
in H. Then we have by the dominated convergence theorem, that
_
H
e
i(
1
W
z
1
(x)++
n
W
z
n
(x))
(dx) = lim
j
_
H
e
i(
1
Q
1/2
z
j
1
,x++
n
Q
1/2
z
j
n
,x)
(dx)
= lim
j
_
H
e
ix,Q
1/2
(
1
z
j
1
++
n
z
j
n
)
(dx)
= lim
j
e

1
2
|
1
z
j
1
++
n
z
j
n
|
2
= e

1
2
|
1
z
1
++
n
z
n
|
2
= e

1
2
P
n
j,k=1
z
j
,z
k
|
j

k
.

Chapter 3
Brownian Motion
3.1 Stochastic Processes
We are given a probability space (, F, P). We denote by P

the outer
measure of P. We recall that a null set of is a set of outer measure zero.
For any integrable real random variable F we note
E(F) =
_

F()P(d).
So, in particular we have
F
#
P(I) = E(1l
I
(F)), I B(R).
We say that a property concerning elements of holds P-a.s. if the set
where does not hold is a null set.
Denition 3.1 A family X = (X(t))
t0
of real random variables in (, F, P)
is called a real stochastic process in [0, +). For any , X(, ) is called
a trajectory of X.
X is Gaussian if for any n N and any 0 t
1
< < t
n
the n-
dimensional random variable (X(t
1
), ..., X(t
n
)) is Gaussian.
X is continuous if X(, ) is continuous P-a.s.
X is p-mean continuous, p 1, if
(i) X(t) is p-integrable for any t 0.
(ii) We have
lim
tt
0
E[[X(t) X(t
0
)[
p
] = 0, t
0
0. (3.1)
27
28 Chapter 3
We notice that a p-mean continuous process is not continuous in general.
We say that two stochastic processes X and Y are equivalent if for all
t 0 we have
X(t, ) = Y (t, ), P-a.s..
When X and Y are equivalent we also say that Y is a version of X (or that
X is a version of Y ).
3.2 Brownian motion
Denition 3.2 A real Brownian motion B = (B(t))
t0
on (, F, P) is a
real stochastic process such that
(i) B(0) = 0 and if 0 s < t, B(t) B(s) is a real Gaussian random
variable with law N
ts
.
(ii) If 0 < t
1
< ... < t
n
, the random variables,
B(t
1
), B(t
2
) B(t
1
), , B(t
n
) B(t
n1
)
are independent.
We express condition (ii) by saying that B is a process with independent
increments.
Lemma 3.3 Let t, s > 0. Then
E[B(t)(B(s)] = mint, s. (3.2)
Proof. Let for instance t > s. Then we have
E[B(t)B(s)] = E[(B(t) B(s))B(s)] +E[B
2
(s)].
On the other hand, B(t) B(s) is independent of B(s) so that
E[(B(t) B(s))B(s)] = E[B(t) B(s)]E[B(s)] = 0.
Since the law of B(s) is N
s
we conclude that E[B(t)B(s)] = s as required.
Brownian motion 29
3.2.1 Construction of a Brownian motion
Consider the probability space (H, B(H), ), where H = L
2
(0, +) and
= N
Q
, Q being an arbitrary (but xed) non degenerate Gaussian measure
in H.
Dene
B(t) = W
1l
[0,t]
, t 0, (3.3)
where
1l
[0,t]
(s) =
_
_
_
1 if s [0, t],
0 otherwise,
and W is the white noise function dened in Chapter 2.
More precisely, for any t 0 we choose an arbitrary element in the
equivalence class of B(t) which we still denote by B(t).
Clearly, for any t 0, B(t) is a Gaussian random variable N
t
and for any
t > s 0, B(t) B(s) = W
1l
(s,t]
is a Gaussian random variable N
ts
. So, B
fullls Denition 3.2(i). Let us prove (ii). Since the system of elements of H,
(1l
[0,t
1
]
, 1l
(t
1
,t
2
]
, ..., 1l
(t
n1
,t
n
]
),
is orthogonal, we have by Proposition 2.20 that the random variables
B(t
1
), B(t
2
) B(t
1
), , B(t
n
) B(t
n1
)
are independent. Thus (ii) is proved as well.
3.2.2 Some properties of a Brownian motion
Proposition 3.4 Let B(t), t 0, be a Brownian motion on (, F, P).
Then B is a Gaussian process. Moreover, if 0 < t
1
< ... < t
n
the law of
(B(t
1
), ..., B(t
n
)) is given by
P((B(t
1
), ..., B(t
n
)) I)
= (2)
n/2
(t
1
(t
2
t
1
) (t
n
t
n1
))
1/2
_
I
e


2
1
2t
1

(
2

1
)
2
2(t
2
t
1
)

(
n

n1
)
2
2(t
n
t
n1
)
d,
(3.4)
for all I B(R
n
).
Proof. Let 0 < t
1
< ... < t
n
and set
X := (B(t
1
), B(t
2
) B(t
1
), ..., B(t
n
) B(t
n1
))
Z := (B(t
1
), ..., B(t
n
)).
30 Chapter 3
Since random variables B(t
1
), B(t
2
) B(t
1
), ..., B(t
n
) B(t
n1
) are inde-
pendent, by Proposition 2.11 it follows that X is a n-dimensional Gaussian
random variable with mean 0 and covariance operator
Q(X) = diag (t
1
, t
2
t
1
, ..., t
n
t
n1
).
Now, consider the linear mapping T L(R
n
) dened by,
T(x
1
, ..., x
n
) = (x
1
, x
1
+ x
2
, ..., x
1
+ + x
n
), (x
1
, ..., x
n
) R
n
.
It is clear that Z = T(X). Therefore by Proposition 2.13 Z is Gaussian with
mean 0 and covariance Q(Z) = TQ(X)T

where T

is the transpose of T.
It remain to show (3.4). If I B(R
n
) we have
P(Z I) = (2)
n/2
(det Q(Z))
1/2
_
I
e

1
2
(Q(Z))
1
,
d.
Since det T = det T

= 1, as easily checked, we have


det Q(Z) = det Q(X) = t
1
(t
2
t
1
) (t
n
t
n1
).
Moreover, since
T
1
= (
1
,
2

1
, ...,
n

n1
),
we have
(Q(Z))
1
, ) = Q
1
T
1
, T
1
) =

2
1
t
1

(
2

1
)
2
(t
2
t
1
)

(
n

n1
)
2
(t
n
t
n1
)
and so, the conclusion follows.
Proposition 3.5 Let B(t), t 0, be a Brownian motion on (, F, P).
Then B is p-mean square continuous for all p 1.
Proof. It is enough to show the result for p = 2m, m N. Let t > t
0
0.
Since B(t) B(t
0
) is a Gaussian random variable N
tt
0
, we have
E([B(t) B(t
0
)[
2m
) =
_
R
[[
2m
N
tt
0
(d) =
(2m)!
m!2
m
(t t
0
)
m
.
Therefore
lim
t0
E([B(t) B(t
0
)[
2m
) = 0
and the conclusion follows.
Exercise 3.6 Let B(t) be a Brownian motion in a probability space (, F, P).
Prove that the following are Brownian motions.
Brownian motion 31
(i) B
1
(t) = B(t + h) B(h), t 0, where h > 0 is given.
(ii) B
2
(t) = B(
2
t), t 0, where > 0 is given.
(iii) B
3
(t) = tB(1/t), t > 0, B
3
(0) = 0.
(iv) B
4
(t) = B(t), t 0.
3.3 Wiener integral
Let B(t), t 0, be a Brownian motion in (, F, P) and let f L
2
(0, T)
with T > 0. We want to dene the stochastic integral:
_
T
0
f(s)dB(s).
We start with step functions. Let 0 = t
0
< t
1
< < t
n
= T, f
0
, f
1
, ..., f
n1

R and set
f =
n

j=1
t
j1
1l
(t
j
t
j1
]
.
Then dene
_
T
0
f(s)dB(s) :=
n

j=1
f
t
j1
(B(t
j
) B(t
j1
)).
Let us prove two basic identities.
Lemma 3.7 We have
E
__
T
0
f(s)dB(s)
_
= 0 (3.5)
and
E
_
__
T
0
f(s)dB(s)
_
2
_
=
n

j=1
[f(t
j1
)[
2
(t
j
t
j1
) =
_
t
0
f
2
(s)ds. (3.6)
Proof. Identity (3.5) is obvious. Let us prove (3.6). We have
E([I

(f)[
2
) = E
_
n

j=1
[f(t
j1
)[
2
[B(t
j
) B(t
j1
)]
2
_
+2E
_
n

j<k
f(t
j1
)f(t
k1
)[B(t
j
) B(t
j1
)][B(t
k
) B(t
k1
)]
_
.
(3.7)
32 Chapter 3
Now the conclusion follows taking into account that B(t
j
) B(t
j1
) is a real
Gaussian random variable N
t
j1
t
j
and that B(t
j
) B(t
j1
) is independent
of B(t
k
) B(t
k1
) for k ,= j.
Denote by S(0, T) the linear space of all step functions. By (3.6) it follows
that the linear mapping I
S(0, T) L
2
(0, T) L
2
(, F, P), f I(f) =
_
T
0
f(s)dB(s),
is continuous. Since S(0, T) is dense in L
2
(0, T) it can be uniquely extended
to the whole L
2
(0, T). We still denote by I(f) =
_
T
0
f(s)dB(s) this estension.
It is clear that for any f L
2
(0, T) we have
E
__
T
0
f(s)dB(s)
_
= 0, (3.8)
and
E
_
__
T
0
f(s)dB(s)
_
2
_
=
_
t
0
f
2
(s)ds. (3.9)
The random variable (more precisely, the equivalence class of random
variables)
_
T
0
f(s)dB(s), which belongs to L
2
(, F, P), is called the Wiener
integral of f in [0, T].
We dene in an obvious way the Wiener integral
_
b
a
f(s)dB(s) for any
a, b 0. It is easy to see that if a, b, c 0 we have
_
b
a
f(s)dB(s) +
_
c
b
f(s)dB(s) =
_
c
a
f(s)dB(s).
Exercise 3.8 Let f, g L
2
(0, T). Show that
E
__
T
0
f(s)dB(s)
_
T
0
g(s)dB(s)
_
=
_
T
0
f(s)g(s)ds.
Proposition 3.9 Let f L
2
(0, T). Then I(f) =
_
T
0
f(s)dB(s) is a real
Gaussian random variable N
q
with q =
_
T
0
[f(s)[
2
ds.
Proof. It is enough to prove the result for f of the form
f =
n

i=1
f
t
i1
(t
i
t
i1
),
Brownian motion 33
where n N, 0 = t
0
< t
1
< ... < t
n1
= T, so that
I(f) =
n

i=1
f
t
i1
(B(t
i
) B(t
i1
)).
Since random variables
B(t
1
), B(t
2
) B(t
1
), , B(t
n
) B(t
n1
),
are independent, we have that I(f) is a real Gaussian random variable N
q
with
q =
n

i=1
f
2
(t
i1
)(t
i
t
i1
).

We now show a relation between the white noise function and the Wiener
integral.
Example 3.10 We use here notations of Section 3.2.1. Let f L
2
(0, ).
Then we have
W
f
=
_

0
f(s)dB(s). (3.10)
It is enough to show (3.10) when
f =
n

k=1
f
t
k1
1l
(t
k1
,t
k
]
,
where 0 t
0
< < t
n
. In this case we have in fact
_

0
f(s)dB(s) =
n

k=1
f
t
k1
W
1l
(t
k1
,t
k
]
= W
P
n
k=1
f
t
k1
1l
(t
k1
,t
k
]
= W
f
.
Let f : [0, ) R such that it is integrable in all interval [0, T], T0. Let
us introduce a stochastic process setting
F(t) =
_
t
0
f(s)ds, t 0.
Proposition 3.11 The process F(t), t 0 is p-mean continuous for any
p 1.
34 Chapter 3
Proof. Let p = 2m, m N and t > t
0
0. Then by Proposition 3.9 we
have that
F(t) F(t
0
) =
_
t
t
0
f(s)dB(s)
is a real Gaussian random variable with mean 0 and covariance
_
t
t
0
f
2
(s)ds.
Therefore
E[F(t) F(t
0
)[
2m
=
(2m)!
m!2
m
__
t
t
0
f
2
(s)ds
_
q
,
so that
lim
tt
0
E[F(t) F(t
0
)[
2m
= 0.
We note nally, that if f C
1
([0, T]) then it is possible to express the
Wiener integral
_
T
0
f(s)dB(s) in terms of a Riemann integral as the following
integration by parts formula shows.
Proposition 3.12 If f C
1
([0, T]) we have
_
T
0
f(s)dB(s) = f(T)B(T)
_
T
0
f

(s)B(s)ds, P-a.e. . (3.11)


Proof. Let = t
0
, t
1
, , t
n
. Then we have
I

(f) =
n

k=1
f(t
k1
)(B(t
k
) B(t
k1
))
=
n

k=1
(f(t
k
)B(t
k
) f(t
k1
)B(t
k1
))

k=1
(f(t
k
) f(t
k1
))B(t
k
)
= f(T)B(T)
n

k=1
(f(t
k
) f(t
k1
))B(t
k
)
= f(T)B(T)
n

k=1
f

(
k
)B(t
k
)(t
k
t
k1
),
where
k
are suitable numbers in the interval [t
k1
, t
k
], k = 1, ..., n. It follows
that
lim
||0
I

(f) = f(T)B(T)
_
T
0
f

(s)dB(s)ds, P-a.s..

Brownian motion 35
3.4 Continuity of Brownian motion
Let B(t), t 0, be a Brownian motion on a probability space (, F, P). We
are going to show that B possesses a continuous version. To this purpose we
shall use a representation formula for B proved in the next proposition.
Proposition 3.13 For any (0, 1/2) we have
B(t) =
sin

_
t
0
(t )
1
Y

()d, (3.12)
where
Y

() =
_

0
( s)

dB(s). (3.13)
Notice that the Wiener integral Y

is meaningful since (0, 1/2).


Proof. We start from the following elementary identity which is valid for
any (0, 1).
_
t
s
(t )
1
( s)

d =

sin
, 0 s t, (3.14)
where (0, 1). To check (3.14) it is enough to set = r(t s) +s so that
(3.14) becomes
_
1
0
(1 r)
1
r

dr = (, 1 ) =

sin
.
Now since, obviously, B(t) =
_
s
0
dB(s) we can write
B(t) =
sin

_
t
0
__
t
s
(t )
1
( s)

d
_
dB(s).
Exchanging integrals
(1)
, yields
B(t) =
sin

_
t
0
d(t )
1
__

0
( s)

dB(s)
_
.

We can now prove the result.


Theorem 3.14 Let B(t), t 0, be a Brownian motion on a probability
space (, F, P). Then B possesses a continuous version.
(1)
This requires a proof which is left to the reader.
36 Chapter 3
Proof. Choose a version Y

(, ) of the stochastic process Y

which is 2m-
integrable with 2m > 1/. This is possible in view of Proposition 3.11. Now
set
B(t, ) =
sin

_
t
0
(t )
1
Y

(, )d, t 0.
Then B(, ) is a continuous version of B thanks to the following analytic
lemma.
Lemma 3.15 Let (0, 1/2), m N with 2m > 1/ and f L
2m
(0, T).
Set
F(t) =
_
t
0
(t )
1
f()d, t [0, T].
Then F C([0, T]; H).
Proof. By H olders inequality we have
[F(t)[
__
t
0
(t )
(1)
2m
2m1
d
_
2m1
2m
[f[
L
2m
(0,T;H)
. (3.15)
(Notice that ( 1)
2m
2m1
> 1.) Therefore F L

(0, T; H) and F is con-


tinuous at 0. Let us prove that F is continuous on [
t
0
2
, T] for any t
0
(0, T].
Let us set for <
t
0
2
,
F

(t) =
_
t
0
(t )
1
f()d, t [0, T].
F

is obviously continuous on [
t
0
2
, 1]. Moreover, using again H olders inequal-
ity, we nd
[F(t) F

(t)[ M
_
2m1
2m 1
_2m1
2m

1
2m
[f[
L
2m
(0,T;H)
.
Thus lim
0
F

(t) = F(t), uniformly on [


t
0
2
, T], and F is continuous as re-
quired.
Exercise 3.16 Prove that B possesses an H older continuous version with
any exponent < 1/2.
3.5 The standard Brownian motion
Let us consider a Brownian motion B(t), t 0, in a probability space
(, F, P) such that B(, ) is continuous for all . We denote by B
the mapping
B : C
0
, B(, ),
where C
0
= C([0, +)) : (0) = 0.
Brownian motion 37
3.5.1 Some properties of C
0
First we notice that, as easily checked, C
0
, endowed with the metric,
d(
1
,
2
) :=

k=1
|
1

2
|
k
2
k
(1 +|
1

2
|
k
)
,
is a complete metric space. We have set for any k N,
||
k
= sup[(t)[ : t [0, k], C
0
.
Let us now consider the -algebra B(C
0
). It is important to notice that
B(C
0
) is generated by the cylindrical subsets of C
0
that we shall introduce
now.
For n N, 0 < t
1
< < t
n
and A B(R
n
) we dene
C
t
1
,t
2
,...,t
n
;A
:= C
0
: ((t
1
), ..., (t
n
)) A .
Note that
C
t
1
,t
2
,...,t
n
;A
= C
t
1
,t
2
,...,t
n
,t
n+1
,...,t
n+k
;AR
k, k, n N.
Using this identity one can easily see that C is an algebra. Moreover, the
-algebra generated by C coincides with B(C
0
) since any ball (with respect
to the metric of C
0
) is a countable intersection of cylindrical sets.
3.5.2 The Wiener measure and the standard Brownian
motion
We come back to the mapping B
B : C
0
, B(, )
and we denote by Q its law (which is a probability measure on (C
0
, B(C
0
)).
Q is called the Wiener measure on (C
0
, B(C
0
)).
So, for any nonnegative Borel mapping
F : C
0
R, F(),
we have
E[F(B())] =
_

F(B(, ))P(d) =
_
C
0
F()Q(d). (3.16)
Some examples of mappings F are the following.
38 Chapter 3
(i) F() = g((t
0
)), for all C
0
, where g : R R is nonnegative Borel
and t
0
> 0 is given.
(ii) F() = G((t
1
), ..., (t
n
)), for all C
0
, where G : R
n
R is nonneg-
ative Borel and t
1
, ..., t
n
> 0 are given.
(iii) F() = sup
t[0,1]
[(t)[, for all C
0
.
Now we dene a stochastic process W(t), t 0, in (C
0
, B(C
0
), Q) setting
W(t)() = (t), C
0
, t 0.
Proposition 3.17 W is a Brownian motion in (C
0
, B(C
0
), Q), called the
standard Brownian motion.
Proof. The proof is straightforward. Let us show for instance that for
t > s 0, W(t) W(s) is a Gaussian random variable N
ts
. For this it is
enough to show that the Fourier transform of W(t) W(s)
(h) :=
_
C
0
e
i((t)(s))h
Q(d), h R,
is given by e

1
2
(ts)h
2
, h R.
In fact by (3.16) we have
_
C
0
e
i((t)(s))h
Q(d) =
_

e
i(B(t,)B(s,))h
P(d)
= E[e
i(B(t)B(s))
] = e

1
2
(ts)h
2
, h R.
In an analogous way one can prove that W(t), t 0, has independent incre-
ments.
Let us compute the Wiener measure of a cylindrical set.
Proposition 3.18 Let C
t
1
,t
2
,...,t
n
;A
be a cylindrical set. Then we have
Q(C
t
1
,t
2
,...,t
n
;A
)
=
1
_
(2)
n
t
1
(t
2
t
1
) (t
n
t
n1
)
_
A
e


2
1
2t
1

(
2

1
)
2
2(t
2
t
1
)

(
n

n1
)
2
2(t
n
t
n1
)
d.
Proof. We simply note that, thanks to (3.16), we have
Q(C
t
1
,t
2
,...,t
n
;A
) = P((B(t
1
), ..., B(t
n
)) A),
so that the conclusion follows from Proposition 3.4.
Brownian motion 39
3.6 Quadratic variation of the Brownian mo-
tion
In this section we are given a real continuous Brownian motion B(t), t 0,
on a probability space (, F, P). For any T > 0 we denote by (0, T) the
set of all decompositions of [0, T]
= 0 = t
0
< t
1
< < t
n
= T.
Then for any = 0 = t
0
< t
1
< < t
n
= T (0, T) we set
[[ := mint
k
t
k1
: k = 1, ...n 1.
We introduce a partial ordering on (0, T), setting

1

2
if and only if [
1
[ [
2
[.
Let us now introduce the quadratic variation of Brownian motion B in
[0, T]. For any = 0 = t
0
< t
1
< < t
n
= T (0, T) we dene
J

:=
n

k=1
[B(t
k
) B(t
k1
)[
2
.
Then we prove
Theorem 3.19 We have
lim
||0
J

= T in L
2
(, F, P).
We say that T is the quadratic variation of B in [0, T].
Proof. Since B
t
k
B
t
k1
is a real Gaussian random variable with law N
t
k
t
k1
,
we have E(J

) = T, and so,
E([J

T[
2
) = E(J
2

) 2TE(J

) + T
2
= E(J
2

) T
2
. (3.17)
Moreover
E[J

[
2
= E

k=1
[B(t
k
) B(t
k1
)[
2

2
= E
n

k=1
[B(t
k
) B(t
k1
)[
4
+ 2
n

h<k=1
E[B(t
h
) B(t
h1
)[
2
[B(t
k
) B(t
k1
)[
2
.
40 Chapter 3
But we have
E
n

k=1
[B(t
k
) B(t
k1
)[
4
= 3
n

k=1
(t
k
t
k1
)
2
, (3.18)
and, since B(t
h
) B(t
h1
) and B(t
k
) B(t
k1
) are independent, we have
n

h<k=1
E[B(t
h
) B(t
h1
)[
2
[B(t
k
) B(t
k1
)[
2
=
n

h<k=1
(t
h
t
h1
)(t
k
t
k1
).
(3.19)
Therefore
E[J

[
2
= 3
n

k=1
(t
k
t
k1
)
2
+ 2
n

h<k=1
(t
h
t
h1
)(t
k
t
k1
)
= 2
n

k=1
(t
k
t
k1
)
2
+
_
n

k=1
(t
k
t
k1
)
_
2
.
= 2
n

k=1
(t
k
t
k1
)
2
+ T
2
.
(3.20)
Now, substituting (3.20) on (3.17), we obtain
E
_
[J

T[
2
_
= 2
n

k=1
(t
k
t
k1
)
2
0,
as [[ 0.
An important consequence of Theorem 3.19 is that almost all trajectories
of the Brownian motion B have not bounded variation
(2)
. In other terms
the set
V
T
:= : B(, ) BV (0, T)
has outer probability zero.
In fact the following result holds.
Proposition 3.20 We have P

(V
T
) = 0.
(2)
Let f : [0, T] R. Then for any = 0 = t
0
< t
1
< < ..., t
n
= T (0, T) we
set V

(f) =

n
k=1
[f(t
k
) f(t
k1
)[ and dene V (f) := sup

(f), V (f) is called the


variation of f. BV (0, T) is the set of all functions f : [0, T] R of nite variation.
Brownian motion 41
Proof. Set
:= : B(, ) is continuous ,
so that P() = 1 because B is continuous.
Since lim
||0
J

= T in L
2
(, F, P) there exists a sequence (
n
)
(0, T) such that [
n
[ 0 and a set
1
F such that
(i) P(
1
) = 1.
(ii) lim
n
J

n
() = T for all
1
.
We claim that
V
T

c
1
. (3.21)
By the claim the conclusion will follow since P(
c
1
) = 0.
Let us prove the claim. Let V
T
. Since B(, ) is uniformly
continuous in [0, T], for any > 0 there exists

> 0 such that


t, s [0, T], [t s[ <

= [B(t, ) B(s, )[ < .


Consequently, if n is so large that [
n
[ <

we have J

n
() V (B(, )).
Since is arbitrary cannot belong to
1
. The claim is proved.
3.7 Multidimensional Brownian motions
Denition 3.21 Let n N and let X
1
, ..., X
n
be stochastic processes on a
probability space (, F, P). Then X(t) := (X
1
(t), ..., X
n
(t)), t 0, is called
an n-dimensional stochastic process.
X
1
, ..., X
n
are said to be independent if for any t
1
, ..., t
n
[0, +) the
random variables X
i
(t
i
) are independent.
A n-dimensional Brownian motion is a n-dimensional stochastic process
B(t) := (B
1
(t), ..., B
n
(t)), t 0, such that B
1
, ..., B
n
are independent Brow-
nian motions.
Example 3.22 Let us construct an n-dimensional Brownian motion. Let
(e
1
, ..., e
n
) be the canonical basis in R
n
and choose = H = L
2
(0, +; R
n
),
F = B(H) and P = N
Q
, where Q is any operator in L
+
1
(H) such that Ker
Q = 0.
Then set
B
i
(t) = W
e
i
1l
[0,t]
, t 0, i = 1, ..., n.
Then one can check easily that B(t) = (B
1
(t), ..., B
n
(t)) is an n-dimensional
Brownian motion.
42 Chapter 3
Let B be a Brownian motion in R
n
. Then the following properties are
easily checked.
(i) If t > s, B(t) B(s) is a Gaussian random variable with law N
(ts)I
n
,
t 0, where I
n
represents the identity in R
n
,
(ii) E[B
i
(t)B
j
(t)] = 0 if i ,= j.
(iii) We have
E
_
[B(t) B(s)[
2

= n(t s). (3.22)


Let us check (iii). We have
E
_
[B(t) B(s)[
2

=
n

k=1
E
_
[B
k
(t) B
k
(s)[
2

= n(t s).
Exercise 3.23 Prove that for 0 s < t we have
E
_
[B(t) B(s)[
4

= (2n + n
2
)(t s)
2
. (3.23)
Exercise 3.24 Let A, C L(R
d
) and set
Z(t) = e
tA
x +
_
t
0
e
(ts)A
CdB(s), t 0.
Prove that the law of Z(t) in R
d
is given by
N
e
tA
x,Q
t
, (3.24)
where
Q
t
=
_
t
0
e
sA
CC

e
sA

ds, (3.25)
where A

and C

are the adjoint of A and C respectively.


Chapter 4
Markov property of the
Brownian motion
Let us consider the probability space (C
0
, B(C
0
), Q) where C
0
is the complete
metric space of all continuous functions : [0, +) R introduced in
Chapter 3 and Q is the Wiener measure. Moreover, let W(t), t 0, the
standard Brownian motion in (C
0
, B(C
0
), Q) dened by
W(t)() = (t), t 0, C
0
.
This chapter is devoted to some sharp properties of the Brownian motion,
in particular the Markov and strong Markov property and the reexion prin-
ciple. To this purpose we shall introduce some basic concepts as ltration,
stopping time and transition semigroup.
4.1 Filtration
For any t > 0 we denote by C
t
the algebra of all cylindrical sets
C
t
1
, ,t
n
;A
= C
0
: ((t
1
), ..., (t
n
)) A
= C
0
: (W(t
1
), ..., W(t
n
)) A
where 0 t
1
< ... < t
n
, t
n
t and A B(R
n
). Moreover, we denote by F
t
the -algebra generated by C
t
. Obviously F
0
= , .
The family of algebras (F
t
)
t0
is increasing; it is called the natural
ltration of W. For any t > 0 we dene
F
t
= F
t
: (0, t)
43
44 Chapter 4
where
_

(0,t)
F
t
_
is the -algebra generated by F
t
for (0, t) and
F
t
+ : =

>0
F
t+
, t 0.
Proposition 4.1 For all t > 0 we have F
t
= F
t
.
Due to Proposition 4.1 we say that the natural ltration (F
t
)
t0
is left con-
tinuous.
Proof. Let t > 0. It is clear that
F
t

_
(0,t)
F
t
,
so that F
t
F
t
. To prove the converse inclusion it is enough to show that
C
t
F
t
.
Let in fact I = C
t
1
, ,t
n
;A
C
t
so that t
n
t. If t
n
< t then I belongs to F
t

whereas if t
n
= t we have
I = lim
k
C
t
1
, ,t
t
1
k
;A
F
t
,
so that I F
t
as well.
Remark 4.2 The ltration (F
t
)
t0
is not right continuous, that is F
t
+ ,= F
t
for all t 0. Let for instance t = 0 and consider the sets
A
n
= : [(1/n)[ 1/n, n N.
Then A
n
F
1/n
and A =

nN
A
n
F
0
+. Notice that
A = : [

(0)[ = 0,
so that F
0
+ ,= F
0
.
4.1.1 F
t
-measurable random variables
We say that a real random variable X is F
t
-measurable if
I B(R) X
1
(I) F
t
.
In this case we say also that X depends from the story of the Brownian
motion only up to t.
The following lemma will be frequently used.
Markov property 45
Lemma 4.3 Let s
2
> s
1
t > 0, and let be a real random variable
F
t
measurable. Then W(s
2
) W(s
1
) and are independent.
Proof. It is enough to show that for any A F
t
, W(s
2
) W(s
1
) and 1l
A
are
independent; in other words that F
t
coincides with the set D dened below.
D = A F
t
: 1l
A
is independent of W(s
2
) W(s
1
).
Since W is a process with independent increments, D contains the algebra
of all cylindrical set belonging to C
t
(which is a -system). Moreover, D
is a -system. In fact if A D it is obvious that A
c
D. Moreover, if
(A
n
) is a sequence in D consisting of disjoint sets, one can show easily that

n=1
A
n
D. Now the claim follows from Dynkins theorem (Theorem A.1
in Appendix A).
Next result shows that F
0
+ contains only trivial sets.
Proposition 4.4 (one-zero law) Assume that A F
0
+. Then either P(A) =
1 or P(A) = 0.
Proof. Let A F
0
+. Denote by G the -algebra generated by all sets of
the form
D
t
1
,...,t
n
,h;I
= : ((t
1
+ h) (h), ..., (t
n
+ h) (h)) I,
where n N, 0 < t
1
< < t
n
, h > 0, I B(R
n
). It is clear that A is
independent of G, since it belongs to all F
t
, t > 0, and W has independent
increments. Then we have
P(A G) = P(A)P(G), G G. (4.1)
On the other hand, we claim that G = B(C
0
). To prove the claim it is
enough to show that any cylindrical set C
t
1
,...,t
n
,h;I
belongs to G; but this
follows from the identity
lim
j
D
t
1

1
j
,...,t
n

1
j
,
1
j
;I
= lim
j
: ((t
1
) (1/j), ..., (t
n
) (1/j)) I = C
t
1
,...,t
n
;I
.
Since G = B(C
0
) we can set in (4.1) G = A, so that P
2
(A) = P(A) which
yields P(A) equal to zero or one.
Remark 4.5 For any t 0 denote by F
t
the -algebra generated by F
t
and all null sets of (called the completion of F
t
). By using Proposition 4.4
one can easily show that (F
t
)
t0
is both right and left continuous.
46 Chapter 4
4.2 Stopping times
A nonnegative extended (that is with values in [0, +]) random variable in
(C
0
, B(C
0
), Q) is called a stopping time with respect to the ltration (F
t
)
t0
if
t F
t
for all t 0.
To any stopping time we associate the -algebra
F

: = A F : A t F
t
for all t 0.
Let us describe the -algebra F

, For 0 < t
1
< ... < t
n
and IB(R) we
dene
C
()
t
1
,...,t
n
;I
= : t
n
() < , ((t
1
), ..., (t
n
)) I = C
t
1
,...,t
n
;I
t
n
< .
We claim that C
()
t
1
,...,t
n
;I
is F

-measurable.
In fact
C
()
t
1
,...,t
n
;I
t = C
t
1
,...,t
n
;I
t
n
< t
So, the -algebra generated by all C
()
t
1
,...,t
n
;I
in included in F

and one can


show that it coincides with F

.
If is stopping time, then > t and = t belong obviously to F
t
for all t 0.
Moreover, is F

-measurable. In fact, if A = s we have


A t = t s F
ts
F
t
.
In other words we have
F

(),
where () is the -algebra generated by .
Remark 4.6 Let be an extended random variable such that
< t F
t
, for all t 0.
Then is not in general a stopping time with respect to (F
t
)
t0
, but it is a
stopping time with respect to the ltration (F
t
+)
t0
. In fact
t =

k=1
_
t +
1
k
_
F
t
+.
Markov property 47
Exercise 4.7 Assume that the nonnegative random variable is discrete,
that is that () = (
k
)
kN
where
k
is an increasing sequence of positive
numbers. Show that is a stopping time if and only if =
k
F

k
for
all k N. Show that in this case F

is the algebra
F

: = A F : A =
k
F

k
for all k N.
Proposition 4.8 Let be a stopping time. Then there exists a decreasing
sequence (
n
) of discrete stopping times convergent pointwise to such that
F

n
F

for all n N.
Proof. Dene for any n N and

n
() =
k
2
n
if
k 1
2
n
() <
k
2
n
, k N. (4.2)
It is clear that the sequence (
n
) is decreasing. Moreover,
n
is a stopping
time. In fact, if t =
k
2
n
with k N we have

n
= t =
_
k 1
2
n
<
k
2
n
_
F
t
. (4.3)
Finally, let A F

, that is
A t F
t
, t 0.
Then we have
A
_

n
=
k
2
n
_
= A
_
k 1
2
n
<
k
2
n
_
F k
2
n
, k N,
so that A F

n
.
We want to extend several properties concerning time t to general stop-
ping times . We start by showing that W

is F

-measurable.
Proposition 4.9 Let be a stopping time and set
W

() = W((), ), .
Then W

is F

-measurable.
Proof. Assume rst discrete,
() = t
k
, 0 < t
1
< < t
k
<
48 Chapter 4
and set A
k
= = t
k
, k N. Then we have
W

() = W(t
k
)(), A
k
, k N.
Let I B(R). Then
W

I t =

k=1
[W

I t A
k
]
=

k=1
[W
t
k
I t A
k
]
=

{kN: t
k
t}
[W
t
k
I t A
k
] F
t
.
So, the conclusion holds in this case.
Let now be arbitrary, let
n
be dened by (4.2) and set
W

n
() = W(
n
(), ), .
Since W is continuous we have
lim
n
W

n
() = W

(), .
Fix t 0. By the previous argument we have
W

n
I
n
t F
t
for all I B(R). (4.4)
Now the conclusion follows letting n .
Example 4.10 Let a R and set
(1)

a
= inft 0 : W(t) = a.
Then

a
> t =

s[0,t]
W(s) < a =

s[0,t]Q
W(s) < a F
t
.
So,
a
is a stopping time with respect to the ltration (F
t
)
t0
.
Let now
= inft 0 : W(t) > a.
Then we have
t =

s[0,t]
W(s) a =

s[0,t]Q
W(s) a F
t
.
Consequently, by Remark 4.6, is a stopping time with respect to ltration
F
t
+
t0
.
(1)
We use the convention that the inmum of the empty set is +.
Markov property 49
4.3 The Brownian motion W(t + ) W()
We recall that W(t +h) W(t), t 0, is a Brownian motion for any h > 0.
We want now to show that the same holds when h is replaced by a stopping
time.
Proposition 4.11 Let be a stopping time. Then
C(t) := W(t + ) W(), t 0,
is a Brownian motion.
Proof. Let us rst prove that the law of C(t) is N
t
. For this it is enough to
show that for any R we have
E
_
e
iC(t)
_
= E
_
e
i(W(t+)W())
_
= e

1
2

2
t
, R. (4.5)
Assume rst that is discrete, () = (t
k
) and set
A
i
= = t
i
F
t
i
, i N.
Then we have
E
_
e
i(W(t+)W())
_
=

i=1
_
A
i
e
i(W(t+t
i
)W(t
i
))
dP =

i=1
E
_
1l
A
i
e
i(W(t+t
i
)W(t
i
))
_
.
Since 1l
A
i
and W(t + t
i
) W(t
i
) are independent, it follows that
E
_
e
i(W(t+)W())
_
=

i=1
P(A
i
)E
_
e
i(W(t+t
i
)W(t
i
))
_
= e

1
2

2
t
and so (4.5) is proved.
Let now be general and let (
n
) be the sequence of nite stoppping
times dened by (4.2). We have just proved that
E
_
e
i(W(t+
n
)W(
n
))
_
= e

1
2

2
t
, R.
Now (4.5) follows letting n tend to innity. By (4.5) it follows that C(t) is a
Gaussian random variable N
t
. Proceeding similarly one can prove that the
law of C(t) C(s) with t > s > 0 is N
ts
and that C(t) has independent
increments. Continuity of C(t) is obvious.
50 Chapter 4
4.4 Transition semigroup
We shall denote by B
b
(R) the set of all real, bounded and Borel functions
and by C
b
(R) the subspace of B
b
(R) of those functions which are uniformly
continuous and bounded on R.
Given B
b
(R) we want to study the evolution in time of (W(t) +x).
To this purpose, we dene the transition semigroup
P
t
(x) = E[(W(t) + x)], t 0, x R, B
b
(R), (4.6)
Since the law of W(t) + x is N
x,t
we have
P
t
(x) = E[(W(t) + x)]
=
1

2t
_
+

1
2t
(xy)
2
(y)dy
=
_
+

g
t
(x y)(y)dy,
(4.7)
where
g
t
() =
1

2t
e

2
2t
, t > 0, R. (4.8)
We deduce, by an explicit computation, that P
t
, t 0, is a semigroup of
linear operators in B
b
(R), that is P
0
= I and
P
t+s
= P
t
P
s
, t, s 0.
Notice that P
t
coincides with the heat semigroup in R. In fact one checks
easily that if C
b
(R) then the function u : [0, +) R R, u(t, x) =
P
t
(x) is continuous, innitely dierentiable and fullls
_

_
u
t
(t, x) =
1
2
u
xx
(t, x), t > 0, x R,
u(0, x) = (x), x R.
Remark 4.12 One can show that u(t, x) = P
t
(x), t 0, x R, is the
unique solution of the Dirichlet problem above.
There is a simple deterministic proof based on maximum principle and a
stochastic proof, which we will present later, based on Itos formula.
Exercise 4.13 Prove that for t > s 0,
P
ts
(x) = E[(W(t) W(s) + x)], B
b
(H), x R. (4.9)
Markov property 51
4.5 Markov property
In this section we shall use several properties of conditional expectation, they
are recalled in Appendix A.
We are here concerned with the stochastic process
X(t) = X(t, x) = W(t) + x, t 0,
where x R.
Proposition 4.14 For any t > s > 0 and any B
b
(H) we have
E[(X(t))[F
s
] = (P
ts
)(X(s)). (4.10)
Equivalently
_
A
(X(t))dP =
_
A
(P
ts
)(X(s))dP, A F
s
. (4.11)
Moreover X() is a Markov process.
Proof. Set
X(t) = W(t) + x = (W(s) + x) + (W(t) W(s)) =: U + V.
Notice that U is F
s
-measurable and V is independent of F
s
. By Proposition
B.6 it follows that
E[(X(t))[F
s
] = E[(U + V )[F
s
] = h(U),
where (recall Exercise 4.13)
h(u) = E[(u + V )] = E[(u + W(t) W(s))] = P
ts
(u).
So, (4.10) is proved.
To prove the last statement notice that by Proposition B.3 we have
E[(X(t))[X(s)] = E[E[(X(t))[F
s
][X(s)]
= E[P
ts
(X(s))[X(s)]
= P
ts
(X(s)) = E[(X(t))[F
s
].

Exercise 4.15 Let s > 0, a F


s
-measurable random variable and
B
b
(R). Show that
E[(W(t) + [F
s
] = (P
ts
()).
52 Chapter 4
4.5.1 Strong Markov property
We now consider conditional expectation with respect to F

where is a
stopping time.
Proposition 4.16 Let be a stopping time and let t and B
b
(H).Then
we have
E[(X(t))[F

] = (P
t
)(X()). (4.12)
Equivalently
_
A
(X(t))dP =
_
A
(P
t
)(X())dP, A F

. (4.13)
Proof. We set x = 0 for simplicity, so that X(t) = W(t). Assume rst that
is of the form
() = (t
k
)
kN
.
Let A F

. Then we have
_
A
(P
t
)(W())dP =

i=1
_
A{=t
i
}
(P
t
)(W())dP
=

i=1
_
A{=t
i
}
(P
tt
i
)(W(t
i
))dP.
Therefore, by (4.10) and taking into account that by the denition of F

we
have
A = t
i
F
t
i
, i = 1, ..., n,
we can write,
_
A
(P
t
)(W())dP =

i=1
_
A{=t
i
}
(P
tt
i
)(W(t
i
))dP
=

i=1
_
A{=t
i
}
E[(W(t))[F
t
i
]dP
=

i=1
_
A{=t
i
}
(W(t))dP =
_
A
(W(t))dP.
Therefore, (4.13) is proved.
Markov property 53
Let now be an arbitrary stopping time and let (
n
) be dened by (4.2).
Recall that (Proposition 4.8)
F

n
for all n N.
Let A F

. Then by (4.13) it follows that


_
A
(W(t))dP =
_
A
(P
t
n
)(W(
n
))dP for all A F

.
Now the conclusion follows letting n .
Property (4.12) is called the strong Markov property of W.
4.6 Some consequences of the strong Markov
property
In this section we want to determine the laws of the following important
random variables.
T
b
= inft 0 : B(t) = b, b R.
M(t) = max
s[0,t]
B(s), t 0.
m(t) = min
s[0,t]
B(s), t 0.
Notice that
T
a
t = M(t) a, t 0, a 0 (4.14)
and
T
a
t = m(t) a, t 0, a 0. (4.15)
To nd the laws of T
a
with a 0 and M(t) the following lemma is useful.
Lemma 4.17 Let a 0 and t 0. Then we have
P(B(t) a, M(t) a) = P(B(t) a). (4.16)
Proof. We have, taking into account that
T
a
t = M(t) a
54 Chapter 4
P(W(t) a, M(t) a) = P(W(t) a, T
a
t)
=
_
{T
a
t}
1l
(,a]
(W(t))dP
=
_
{T
a
t}
E[1l
(,a]
(W(t))[F
T
a
]dP,
since T
a
t F
T
a
. By the strong Markov property it follows that
P(W(t) a, M(t) a) =
_
{T
a
t}
E[1l
(,a]
(W(t))[F
T
a
]dP
=
_
{T
a
t}
E[P
tT
a
1l
(,a]
(W(T
a
))]dP
=
_
{T
a
t}
E[P
tT
a
1l
(,a]
(a)]dP.
On the other hand, we have, as easily checked,
P
s
1l
(,a]
(a) = P
s
1l
[a,+)
(a), s > 0, a > 0.
Therefore
P(W(t) a, M(t) a) =
_
{T
a
t}
E[P
tT
a
1l
(,a]
(a)]dP
=
_
{T
a
t}
E[P
tT
a
1l
[a,+)
(a)]dP
=
_
{T
a
t}
E[1l
[a,+)
(W(t))[F
T
a
]dP
= P(W(t) a, M(t) a)
= P(W(t) a).
Proposition 4.18 (Reection principle) For all a 0 we have
P(M(t) a) = 2P(W(t) a), (4.17)
Proof. Write
P(M(t) a) = P(M(t) a, W(t) a) +P(M(t) a, W(t) a).
Markov property 55
Now, by Lemma 4.17 we have P(M(t) a, W(t) a) = P(W(t) a).
Moreover, it is clear that P(M(t) a, W(t) a) = P(W(t) a) so, the
conclusion follows.
By Proposition 4.18 we can easily deduce the expressions of the laws of
M(t) and T
a
for all a R.
Corollary 4.19 (Law of M(t)) For all t 0 we have
(M(t)
#
P)(d) =
2

2t
e

2
2t
1l
[0,+)
()d. (4.18)
Proof. We have in fact by Proposition 4.18 for any a 0
P(M(t) a) = 2P(W(t)[ a) =
2

2t
_
+
a
e

2
2t
d
= P([W(t)[ a).

Remark 4.20 From Corollary 4.19 it follows that at xed time t the law
of M(t) coincides with that of [W(t)[, though random variables M(t) and
[W(t)[ are dierent; in particular M(t) is increasing whereas [W(t)[ is not.
Obviously the laws of M() and [W()[ on C
0
([0, +)) are dierent.
Corollary 4.21 (Law of T
a
) Let a 0 and t 0. Then we have
((T
a
)
#
P)(dt) =
a

2t
3
e

a
2
2t
dt. (4.19)
Proof. By (4.14) and Proposition 4.18 we have
P(T
a
t) = P(M(t) a) =
2

2t
_
+
a
e

2
2t
d
=
2

2
_
+
at
1/2
e

2
2
d.
Therefore
d
dt
P(T
a
t) =
a

2t
3
e

a
2
2t
dt,
which implies the conclusion.
The following results can be proved similarly.
56 Chapter 4
Lemma 4.22 Let a 0 and t 0. Then we have
P(W(t) a, m(t) a) = P(W(t) a). (4.20)
Proposition 4.23 (Reection principle) For all a 0 we have
P(m(t) a) = 2P(W(t) a). (4.21)
Corollary 4.24 (Law of m(t)) For all t 0 we have
(m(t)
#
P)(d) =
2

2t
e

2
2t
1
(,a]
()d. (4.22)
Corollary 4.25 (Law of T
a
) Let a R and t 0. Then we have
((T
a
)
#
P)(dt) =
[a[

2t
3
e

a
2
2t
dt. (4.23)
4.7 Application to partial dierential equa-
tions
For any x 0 we set in this section

x
= inft 0 : W(t) + x = 0 = T
x
.
Moreover we consider the following processes which take values in [0, +).
(i) Y (t) = W(t) + x, t [0,
x
].
Y (t) is called the Brownian motion killed in 0.
(ii) U(t) = [W(t) + x[, x 0, t 0.
U(t) is called the Brownian motion reected in 0
(iii) V (t) = W(t
x
) + x, t 0. V (t) is called the Brownian motion
absorbed in 0
Markov property 57
4.7.1 The Dirichlet problem in the half-line
We are here concerned with the process Y (t) = W(t) + x, t [0,
x
].
Dene for any B
b
([0, +))
U
t
(x) := u(t, x) := E[(W(t) + x)1l
t
x
], t 0, x H. (4.24)
We are going to show that u(t, x) is the solution of the Dirichlet problem in
[0, +),
_

_
u
t
(t, x) =
1
2
u
xx
(t, x), x > 0, t > 0
u(t, 0) = 0, t > 0,
u(0, x) = (x), x 0.
(4.25)
Proposition 4.26 We have
u(t, x) =
_
+
0
[g
t
(x y) g
t
(x + y)](y)dy, x 0, t 0, (4.26)
where g is dened by (4.8).
Proof. We have
u(t, x) = E[(W(t) + x)1l
t
x
]
= P
t
(x) E[(W(t) + x)1l
t>
x
],
where is extended to R by setting
(x) = (x), x 0.
Write
E[(W(t) + x)1l
t>
x
] = E[E[1l
t>
x
(W(t) + x)[F

x
]]
= E[1l
t>
x
E[(W(t) + x)[F

x
]]
Now, using the strong Markov property we nd that,
E[(W(t) + x)1l
t>
x
] = E[1l
t>
x
(P
t
x
)(0)] =: E[(
x
)],
where
() = 1l
t>
1
_
2(t )
_
R
e


2
2(t)
()d, > 0.
58 Chapter 4
Next, recalling the law of
x
(see (4.23)) it follows that
E[(W(t) + x)1l
t>
x
] =
_
t
0
__
R
g
ts
(y)(y)dy
_
x

2s
3
e

x
2
2s
ds
=

x
_
t
0
__
R
g
ts
(y)(y)dy
_
g
s
(x)ds
=
_
R
g
t
(x y)(y)dy +

x
_
R
G
x,y
(y)dy,
where
(2)
G
x,y
=
_
t
0
g
ts
(y)g
s
(x)ds =
1
2
Erfc
_
[x[ +[y[

2t
_
.
Since, for x > 0,

x
G
x,y
=
1

2t
e

(x+|y|)
2
2t
= g
t
(x +[y[)
we get
u(t, x) =
_
R
g
t
(x y)(y)dy
_
R
g
t
(x +[y[)(y)dy,
and the conclusion follows.
It is easy to check, by a direct computation, that if C
b
([0, +)),
U
t
(x) = u(t, x) is the solution of the Dirichlet problem (4.25). Moreover
U
0
= I and U
t+s
= U(t)U(s) for all t, s 0.
4.7.2 The Neumann problem
We consider the process
U(t) = [W(t) + x[, x 0, t 0.
For any B
b
([0, +)) we set
Q
t
(x) = E[([W(t) + x[)] = (2t)
1/2
_
R
e

|xy|
2
2t
([y[)dy.
Replacing in the last integral y with y, we see that
Q
t
(x) =
_
+
0
[g
t
(x y) + g
t
(x + y)](y),
(2)
We recall that Erfc (a) =
2

_
+
a
e
r
2
dr.
Markov property 59
where g
t
is dened by (4.8).
Now it is easy to check that if C
b
([0, +)) then u(t, x) = Q
t
(x) is
continuous in [0, ) [0, ), innitely dierentiable in (0, ) [0, ) and
solves the following Neumann problem
_

_
u
t
(t, x) =
1
2
u
xx
(t, x), x 0, t > 0,
u
x
(t, 0) = 0, t > 0,
u(0, x) = (x), x 0.
Moreover Q
0
= I and Q
t+s
= Q(t)Q(s) for all t, s 0.
4.7.3 The Ventzell problem
Let us consider the stochastic process,
V (t) = W(t
x
) + x, t 0,
where x 0.
Set
Z
t
(x) = E[(W(t
x
) + x)], B
b
([0, +)), x 0.
So,
Z
t
(x) =
_

(B(t
x
) + x)dP
=
_
{t<
x
}
(W(t) + x)dP +
_
{t
x
}
(0)dP,
since W(
x
) + x = 0. Therefore
Z
t
(x) = U
t
(x) + (0) P(T
x
t),
where U
t
is dened by (4.24). So
Z
t
(x) =
_
+
0
[g
t
(x y) g
t
(x + y)](y)dy +
(0)

2t
_
x

y
2
2t
dy.
If C
b
([0, +)), setting u(t, x) = Z
t
(x) we see that u is the solution to
the Ventzell problem,
_

_
u
t
(t, x) =
1
2
u
xx
(t, x), x 0, t 0
u
xx
(t, 0) = 0, t 0,
u(0, x) = (x), x 0.
60 Chapter 4
Moreover Z
0
= I and Z
t+s
= Z(t)Z(s) for all t, s 0.
Chapter 5
The It o integral
In all this chapter B represents a Brownian motion in a probability space
(, F, P).
Similarly as in Chapter 4, for any t > 0 we denote by C
t
the algebra of
all cylindrical sets
C
t
1
, ,t
n
;A
= C
0
: (B(t
1
), ..., B(t
n
)) A
where 0 t
1
< ... < t
n
, t
n
t and A B(R
n
).
Moreover, we denote by F
t
the -algebra generated by C
t
and all P-null
sets of . We call F
t
, t 0 the natural ltration of B augmented with the
null sets of P.
The family of algebras (F
t
)
t0
is increasing; it is called the natural
ltration of B.
We denote by (F
t
)
t0
the completion of the natural ltration of B with
all P-null sets of .
We say that a stochastic process F(t), t [0, T], is adapted to the Brow-
nian motion B if F(t) is F
t
-measurable for any t [0, T].
5.1 Denition of Itos integral
5.1.1 It os integral for elementary processes
Denition 5.1 Let T > 0. An elementary process F(t), t [0, T], in
(, F, P) is a stochastic process of the form
F =
n

i=1
F
i1
1l
[t
i1
,t
i
)
, (5.1)
61
62 The Ito integral
where n N, 0 = t
0
< t
1
< < t
n
= T and F
i
is F
t
i
-measurable for any
i = 0, 1, ..., n 1.
For any elementary process F(t), t [0, T], we dene the Ito integral
setting
I(F): =
_
T
0
F(s)dB(s) =
n

i=1
F
i1
(B(t
i
) B(t
i1
)). (5.2)
Obviously any elementary process is adapted. This property is needed to
prove some basic identities (similar to those obtained for the Wiener integral)
which allow to extend the integral to more general processes.
Proposition 5.2 Assume that F E
2
B
(0, T). Then I(F) L
2
(, F, P) and
we have
E
__
T
0
F(s)dB(s)
_
= 0 (5.3)
E
_
__
T
0
F(s)dB(s)
_
2
_
=
_
T
0
E([F(s)[
2
)ds. (5.4)
Proof. Let us prove (5.3). We have
E[I(F)] =
n

j=1
E[F
j1
(B(t
j
) B(t
j1
))].
Since F
j1
is F
j1
measurable, it is independent of B(t
j
)B(t
j1
), by Lemma
4.3. Therefore we have
E[I(F)] =
n

j=1
E[F
j1
]E[B(t
j
) B(t
j1
)] = 0
and (5.3) is proved.
Let us prove (5.4). We have
E[[I(F)[
2
] = E
_
n

j=1
[F
j1
[
2
[B(t
j
) B(t
j1
)]
2
_
+2E
_

j<k
F
j1
F
k1
[B(t
j
) B(t
j1
)] [B(t
k
) B(t
k1
)]
_
.
Notice now that for j < k the random variable
F
j1
F
k1
[B(t
j
) B(t
j1
)],
Chapter 5 63
is F
k1
measurable and consequently is independent of B(t
k
) B(t
k1
).
Therefore, taking the expectation, we have
E[F
j1
F
k1
[B(t
j
) B(t
j1
)][B(t
k
) B(t
k1
)]]
= E[F
j1
F
j1
[B(t
j
) B(t
j1
)]] E[B(t
k
) B(t
k1
)] = 0.
It follows that
E[[I(F)[
2
] =
n

j=1
E[[F
j1
[
2
](t
j
t
j1
),
as required.
Exercise 5.3 Let F, G E
2
B
(0, T). Prove that
E
__
T
0
F(s)dB(s)
_
T
0
G(s)dB(s)
_
=
_
T
0
E[F(s)G(s)]ds.
Hint: Use the identity
ab =
1
2
(a + b)
2

1
2
a
2

1
2
b
2
, a, b R.
5.1.2 General denition of It os integral
Let us denote by
Z
T
:= L
2
([0, T] , B(0, T) F, dt P)
the Hilbert space of all (equivalence classes of) functions
F : [0, T] , (t, ) F(t, ),
which are measurable with respect to the product -algebra, B(0, T) F
and such that
|F|
Z
T
:= E
_
T
0
[F(t, )[
2
dt < .
The scalar product on Z is dened by
F, F
1
) = E
_
T
0
F(t, )F
1
(t, )dt.
Obviously any elementary process F belongs to Z.
64 The Ito integral
In view of (5.4), the mapping
E
2
B
(0, T) Z
T
L
2
(, F
T
, P)F
_
T
0
F(s)dB(s),
is an isometry. Therefore it can be uniquely extended to the closure E
2
B
(0, T)
of E
2
B
(0, T) in Z
T
.
Processes belonging to E
2
B
(0, T) are called predictable.
So, the It o integral can be uniquely dened by extension for any pre-
dictable square integrable process F(t), t 0 and the following properties
are fullled.
E
__
T
0
F(s)dB(s)
_
= 0 (5.5)
E
_
__
T
0
F(s)dB(s)
_
2
_
=
_
T
0
E([F(s)[
2
)ds. (5.6)
Moreover, from Exercise 5.3 it follows that if F and G are predictable square
integrable processes we have
E
__
T
0
F(s)G(s)dB(s)
_
=
_
T
0
E[F(s)G(s)]ds. (5.7)
We can dene in an obvious way the Ito integral
_
b
a
F(s)dB(s) in any
interval [a, b] [0, T]. We have
E
__
b
a
F(s)dB(s)
_
= 0,
and
E
_
__
b
a
F(s)dB(s)
_
2
_
=
_
b
a
(E[F(s)[
2
)ds.
Moreover, for any a, b, c [0, T] we have
_
c
a
F(s)dB(s) =
_
b
a
F(s)dB(s) +
_
c
b
F(s)dB(s).
Let us now present a characterization of predictable processes (that is of
space E
2
B
(0, T)). Note rst that an elementary process is a linear combination
of processes of the form
F1l
[a,b)
, with F F
a
-measurable.
Chapter 5 65
In turn each F can be approximated by linear combinations of characteristic
functions of F
a
-measurable sets. So, it is natural to approximate a general
predictable process by linear combinations of functions of the form
1l
A[a,b)
, with A F
a
measurable.
We call A [a, b) a predictable rectangle. We denote by R the family of all
predictable rectangles and by P the -algebra generated by R. P is called
the -algebra of all predictable events.
Denition 5.4 A real predictable process in [0, T] is a real random variable
in the probability space
([0, T] , P, dt P).
Proposition 5.5 The closure E
2
B
([0, T]) is precisely L
2
([0, T], P, dtP).
Proof. Denote by
T
the closure of E
2
B
([0, T]) in L
2
([0, T] , P, dt P).
Since any element of L
2
([0, T] , P, dt P) can be approximated by a
monotonic sequence of simple functions, it is enough to show that 1l
A

T
for any A P. For this we shall use the Dynkin Theorem, see Appendix A.
We rst note that R is a -system. Then we set
D = A P : 1l
A

T
.
We claim that D is a -system, i.e. that it fullls (A.1). Properties (B.1)-
(i)-(ii) are clear, let us show (A.1)-(iii). Let (A
n
) D be mutually disjoint
sets and set

n
=
n

k=1
1l
A
k
.
Then, by the monotone convergence theorem,
n
= 1l
A
in L
2
([0, T]
, P, dt P) where A =

k=1
A
k
. So, A D and (A.1)-(iii) is fullled. Now
the conclusion follows by Theorem A.1.
Exercise 5.6 Let F L
2
([0, T] , P, dt P), [s, t] [0, T] and let
L

(, F
s
, P). Prove that

_
t
s
F(r)dB(r) =
_
t
s
F(r)dB(r). (5.8)
Exercise 5.7 Let F L
2
([0, T] , P, dt P) such that
_
T
0
F(s)dB(s) = 0.
Show that F = 0.
66 The Ito integral
5.2 It o integral for mean square continuous
processes
We shall denote by C
B
([0, T]; L
2
()) the space of all stochastic processes
which are mean square continuous and adapted. We recall that if F
C
B
([0, T]; L
2
()) then F(t) is F
t
-measurable for all t [0, T] and the map-
ping
[0, T] L
2
(, F, P), t F(t),
is continuous.
For any decomposition = t
0
, t
1
, , t
n
(0, T) consider the ele-
mentary process
F

:=
n

j=1
F(t
j1
)1l
[t
j1
,t
j
)
and set
I

(F) :=
_
T
0
F

(s)dB(s) =
n

j=1
F(t
j1
)(B(t
j
) B(t
j1
)).
Clearly F

E
2
B
(0, T) and, using the continuity of F one can check easily
that
lim
||0
F

= F, in L
2
([0, T] , P, dt P). (5.9)
Consequently we have
lim
||0
I

(F) =
_
T
0
F(s)dB(s) in L
2
(, F, P). (5.10)
Example 5.8 Let us prove that
_
T
0
B(t)dB(t) =
1
2
(B
2
(T) T). (5.11)
Let = t
0
, t
1
, ..., t
n
(0, T). Write
B(t
k1
)(B(t
k
) B(t
k1
)) = B(t
k1
)B(t
k
) B
2
(t
k1
))
=
1
2
B
2
(t
k
) + B(t
k1
)B(t
k
)
1
2
B
2
(t
k1
) +
1
2
B
2
(t
k
)
1
2
B
2
(t
k1
)
=
1
2
B
2
(t
k
)
1
2
B
2
(t
k1
)
1
2
(B(t
k
) B(t
k1
))
2
.
Chapter 5 67
Then we have
I

(B) =
1
2
B
2
(T)
1
2
n

k=1
(B(t
k
) B(t
k1
))
2
.
Recalling that the quadratic variation of B is T (Theorem 3.19), we deduce
that
_
T
0
B(t)dB(t) = lim
||0
I

(B) =
1
2
(B
2
(T) T).
Exercise 5.9 Prove that
lim
||0
n

k=1
B(t
k
)(B(t
k
) B(t
k1
)) =
1
2
(B
2
(T) + T), in L
2
(, F, P),
and
lim
||0
n

k=1
B
_
t
k
+ t
k1
2
_
(B(t
k
) B(t
k1
)) =
1
2
B
2
(T), in L
2
(, F, P).
Therefore the denition of the Ito integral depends on the particular form of
the integral sums.
5.3 The It o integral as a stochastic process
Let F L
2
([0, T] , P, dt P and set
X(t) =
_
t
0
F(s)dB(s), t [0, T].
We rst notice that X(t), t 0, is not a process with independent increments
in general (unless f is deterministic); take for instance
X(t) =
_
t
0
B(s)dB(s) =
1
2
(B
2
(t) t), t 0.
However, X(t), t 0, has orthogonal increments (in the sense of L
2
(, F, P))
as the following result shows.
Proposition 5.10 Let 0 t
1
t
2
t
3
t
4
T. Then we have
E[(X(t
2
) X(t
1
))(X(t
4
) X(t
3
))] = 0
68 The Ito integral
Proof. We have in fact, taking into account (5.7)
E[(X(t
2
) X(t
1
))(X(t
4
) X(t
3
))]
= E
__
t
2
t
1
F(s)dB(s)
_
t
4
t
3
F(s)dB(s)
_
= E
__
T
0
1l
[t
1
,t
2
]
F(s)dB(s)
_
T
0
1l
[t
3
,t
4
]
F(s)dB(s)
_
=
_
T
0
1l
[t
1
,t
2
]
1l
[t
3
,t
4
]
E(F
2
(s))ds = 0.

We are going to show that X(t), t 0, is mean square continuous, then


that it is a continuous process.
Proposition 5.11 Let F L
2
([0, T], P, dtP). Then X C
B
([0, T]; L
2
()).
Proof. We know that for any t [0, T], X(t) L
2
(, F
t
, P). Moreover, for
any t, t
0
[0, T] we have
E([X(t) X(t
0
)[
2
) =

_
t
t
0
E([F(r)[
2
)dr

,
so that
lim
tt
0
E([X(t) X(t
0
)[
2
) = 0.
The conclusion follows.
We show now that X(t), t 0, is a continuous process. For this we rst
prove that it is a martingale with respect to the ltration (F
t
) (see Appendix
C).
Proposition 5.12 X(t), t [0, T], is a F
t
martingale
Proof. Let t > s. Since
X(t) X(s) =
_
t
s
F(r)dB(r),
we have
E[X(t)[F
s
] = X(s) +E
__
t
s
F(r)dB(r)[F
s
_
.
Chapter 5 69
So, it remains to prove that
E
__
t
s
F(r)dB(r)[F
s
_
= 0. (5.12)
Notice that this is not obvious since
_
t
s
F(r)dB(r) is not independent of F
s
in general
(1)
. It is enough to prove (5.12) when F is an elementary process,
F =
n

i=1
F
i1
1l
[t
i1
,t
i
)
,
where s = t
1
, , t
n
= t and F
i1
L
2
(, F, P). In this case, taking into
account that F
s
F
i1
, we write
E
__
t
s
F(r)dB(r)[F
s
_
=
n

i=1
E[F
i1
(B(t
i
) B(t
i1
))[F
s
]
=
n

i=1
EE[F
i1
(B(t
i
) B(t
i1
))[F
i1
][F
s
= 0,
since F
i1
is F
i1
measurable and B(t
i
) B(t
i1
) is independent of F
i1
.
So, (5.12) is proved and the conclusion follows.
We are now ready to prove the continuity of X.
Theorem 5.13 Let F L
2
([0, T] , P, dt P) and let
X(t) =
_
t
0
F(s)dB(s), t [0, T].
Then X has a continuous version and
E
_
sup
t[0,T]
[X(t)[
2
_
4
_
T
0
E[F(s)[
2
ds. (5.13)
Proof. Let (F
n
) E
2
B
(0, T) such that
F
n
F in L
2
([0, T] , P, dt P)
and set
X
n
(t) =
_
t
0
F
n
(s)dB(s), n N, t [0, T].
(1)
because F(r) contains in general the story of the Brownian motion from 0 to r.
70 The Ito integral
Since B(t) is continuous it is clear that X
n
(t) is continuous for all n
N. Taking into account Proposition 5.12 we see that X(t), t [0, T], is
a continuous F
t
martingale. Then by Corollary C.6 it follows that for any
n, m N
E
_
sup
t[0,T]
[X
n
(t) X
m
(t)[
2
_
4E([X
n
(T) X
m
(T)[
2
)
= 4E
__
T
0
[F
n
(s) F
m
(s)[
2
ds
_
.
Consequently (X
n
)() is Cauchy in C([0, T]) for almost all and its limit,
which coincides with X() is continuous.
5.4 It o integral with stopping times
5.4.1 Stopping times
We proceed here as in Section 4.2.
A nonnegative extended random variable in (, F, P) is called a stopping
time with respect to the ltration (F
t
)
t0
if
t F
t
for all t 0.
To any stopping time we associate the -algebra
F

: = A F : A t F
t
for all t 0.
The proofs of the two following propositions are completely similar to that
of Proposition 4.8 and 4.8. So, they will be omitted.
Proposition 5.14 Let be a stopping time. Then there exists a decreasing
sequence (
n
) of discrete stopping times convergent pointwise to such that
F

n
F

for all n N.
Proposition 5.15 Let be a stopping time and set
W()() = W(())(), .
Then W() is F

-measurable and W(t + ) W(), t 0 is a Brownian


motion in (, F, P).
Chapter 5 71
5.4.2 It os integral with stopping times
Let F L
2
([0, T] , P, P) and set
X(t) =
_
t
0
F(s)dB(s), t [0, T].
Let moreover T be a stopping time. Dene
_

0
F(s)dB(s): = X(),
where
X(, ) = X((), ), .
Arguing as in Proposition 5.15 and using the fact that X(t), t [0, T], has
a continuous version, one can see that X() is F

measurable.
The following result reduces a It os integral with a stopping time to a
usual one between 0 to T.
Proposition 5.16 Let F L
2
([0, T] , P, dt P) and let T be a
stopping time. Then we have
_

0
F(s)dB(s) =
_
T
0
1l
{s<}
F(s)dB(s). (5.14)
Proof. It is enough to prove the result when is of the form,
() = (t
1
, t
2
, ..., t
n
),
with 0 < t
1
< t
2
< < t
n
T.
Set
A
i
:= = t
i
, i = 1, ..., n.
T

hen A
i
F
t
i
, i = 1, ..., n.
Consider now the stochastic process
h(s) = 1l
{s}
, s [0, T].
We have
h(s) = 1, s [0, t
1
).
If s [t
1
, t
2
) we have
h(s)() = 1 if A
2
A
n
,
72 The Ito integral
so that
h(s) = 1l
A
2
A
n
= 1l
A
c
1
.
Similarly, if s [t
k1
, t
k
) with k n we have
h(s) = 1l
(A
k
...A
n
)
c.
Then h is predictable and
_
T
0
1l
{t<}
F(s)dB(s) =
_
t
1
0
F(s)dB(s) + 1l
(A
1
)
c
_
t
2
t
1
F(s)dB(s)
+ + 1l
(A
1
A
2
A
n1
)
c
_
t
n
t
n1
F(s)dB(s)
= X(t
1
) + 1l
(A
1
)
c(X(t
2
) X(t
1
))
+ + 1l
(A
1
A
2
A
n1
)
c(X(t
n
) X(t
n1
) = X().

5.5 Multidimensional It o integrals


Let m N be xed and consider a standard m-dimensional Brownian motion
B(t) = (B
1
(t), ..., B
m
(t)), t 0
in the probability space (, F, P). Let (F
t
)
t[0,T]
be the natural ltration of
B (augmented with all P-null sets of ) .
We shall dene the It o integral for predictable processes with values
in L(R
m
, R
d
) (that is such that any matrix element belongs to L
2
([0, T]
, P, dtP)). We shall denote this space by L
2
([0, T], P, dtP; L(R
m
, R
d
))).
First we need a lemma whose simple proof is left to the reader.
Lemma 5.17 Let f, g L
2
([0, T] , P, dt P). Then we have
E
__
T
0
f(s)dB
i
(s)
_
T
0
g(s)dB
j
(s)
_
=
i,j
_
T
0
E[f(s)g(s)]ds, i, j = 1, ..., m.
(5.15)
Let now F L
2
([0, T] , P, dt P; L(R
m
, R
d
)). We dene the Ito
integral of F as the d-dimensional process
__
T
0
F(t)dB(t)
_
i
=
m

j=1
_
T
0
F
i,j
(t)dB
j
(t), i = 1, ..., d.
Chapter 5 73
Proposition 5.18 Let F L
2
([0, T] , P, dt P; L(R
m
, R
d
)). Then we
have
E

_
T
0
F(t)dB(t)

2
=
_
T
0
E[Tr (F(t)F

(t))]dt, (5.16)
where Tr denotes the trace.
Proof. Set I(F) =
_
T
0
F(t)dB(t). Then we have
(I(F))
i
=
m

j=1
_
T
0
F
i,j
(t)dB
j
(t), i = 1, ..., d.
It follows that
E[I(F)[
2
=
d

i=1
E
_
m

j=1
_
T
0
F
i,j
(t)dB
j
(t)
_
2
and, taking into account (5.15),
E[I(F)[
2
=
d

i=1
m

j=1
_
T
0
E[F
i,j
(t)
2
]dt,
which yields (5.16).
Remark 5.19 Assume that d = 1 so that L(R
d
; R
m
) is isomorphic to R
m
and F becomes a vector F = (F
1
, , F
m
).
In this case we shall write the Ito integral of F as
_
T
0
F(s), dB(s))
and formula (5.16) reduces to
E

_
T
0
F(t), dB(t))

2
=
_
T
0
E[F(t)[
2
dt. (5.17)
74 The Ito integral
Chapter 6
The It o formula
6.1 Introduction
Let (, F, P) be a probability space, B a real Brownian motion, (F
t
)
t0
the
natural ltration of B augmented with the null sets of P and P the -algebra
of all predictable events (also augmented with the null sets of P).
We are given two stochastic processes b, L
2
([0, T] , P, dt P) and
consider the stochastic process
X(t) = x +
_
t
0
b(s)ds +
_
t
0
(s)dB(s), t 0, (6.1)
where x R. X is adapted, continuous and continuous in mean square.
We set
dX(t) = b(t)dt + (t)dB(t)
and call dX(t) the Ito dierential of X.
Given a regular real function , we are going to give a meaning to the
Itos dierential

(X(t)).
We need some notations. For any k N we denote by C
k
b
(R) the linear
space of all real mappings which are uniformly continuous and bounded to-
gether with their derivatives of order less or equal to k. If C
k
b
(R) we
set
||
0
= sup
xR
[(x)[,
and
||
k
= ||
0
+
k

j=1
sup
xR
[D
j
(x)[.
75
76 Chapter 6
We shall prove the following Itos formula,
(X(t)) = (x) +
_
t
0

(X(s))(s)dB(s)
+
_
t
0
_
1
2

2
(s)

(X(s)) + b(s)

(X(s))
_
ds, t 0.
(6.2)
We shall write (6.2) in the dierential form, setting

(X(t)) =

(X(t))(t)dB(t),
+
_
1
2

2
(t)

(X(t)) + b(t)

(X(t))
_
dt, t 0,
(6.3)
or, also as

(X(t)) =

(X(t))dX(t) +
1
2

2
(t)

(X(t))dt, t 0. (6.4)
Remark 6.1 One can deduce formally Itos formula by proceeding as fol-
lows. Write dX = b(t)dt + (t)dB and
d(X) = (X + dX) (X) =

(X)dX +
1
2

(X)(dX)
2
=

(X)dX +
1
2

(X)b
2
(t)(dt)
2
+ 2b(t)(t)dt dB +
2
(t)(dB)
2
.
Put (dB)
2
= dt and neglet the terms of order greater than dt, that is terms
with (dt)
2
and dt dB(t).
Writing (dB)
2
= dt is justied by Lemma 6.2 below.
Tthe following result on quadratic sums of a process is a generalization of
Theorem 3.19.
Lemma 6.2 Let F C
B
([0, T]; L
2
(, F, P)) and let = 0 = t
0
< t
1
<
< t
n
= T (0, T). Then we have
lim
||0
n

k=1
F(t
k1
)(B(t
k
) B(t
k1
))
2
=
_
T
0
F(s)ds in L
2
(, F, P) (6.5)
Proof. Set
J

:=
n

k=1
F(t
k1
)(B(t
k
) B(t
k1
))
2
.
The Ito formula 77
It is enough to prove that
lim
||0
E
_
_
_
J

k=1
F(t
k1
)(t
k
t
k1
)
_
2
_
_
= 0, (6.6)
since, obviously
lim
||0
n

k=1
F(t
k1
)(t
k
t
k1
) =
_
T
0
F(s)ds in L
2
(, F, P).
To prove (6.6) write
E
_
_
_
J

k=1
F(t
k1
)(t
k
t
k1
)
_
2
_
_
= E
_
_
_
n

k=1
F(t
k1
)
_
[B(t
k
) B(t
k1
)[
2
(t
k
t
k1
)
_
_
2
_
_
=
n

k=1
E
_
[F(t
k1
)[
2
_
[B(t
k
) B(t
k1
)[
2
(t
k
t
k1
)

2
_
+2
n

j<k=1
E
_
F(t
j1
)[[B(t
j
) B(t
j1
)[
2
(t
j
t
j1
)]
F(t
k1
)[[B(t
k
) B(t
k1
)[
2
(t
k
t
k1
)]
_
Since the Brownian motion has independent increments, the last sum van-
ishes, so that
E
_
_
_
J

k=1
F(t
k1
)(t
k
t
k1
)
_
2
_
_
=
n

k=1
E
_
[F(t
k1
)[
2
_
[B(t
k
) B(t
k1
)[
2
(t
k
t
k1
)

2
_
=
n

k=1
E[F(t
k1
)[
2
E
_
_
[B(t
k
) B(t
k1
)[
2
(t
k
t
k1
)

2
_
,
(6.7)
78 Chapter 6
since F(t
k1
) and B(t
k
) B(t
k1
) are independent.
Now, taking into account that
E[[B(t
k
) B(t
k1
)[
2
] = (t
k
t
k1
),
E[[B(t
k
) B(t
k1
)[
4
] = 3(t
k
t
k1
)
2
,
we have
E
_
_
_
J

k=1
F(t
k1
)(t
k
t
k1
)
_
2
_
_
= 2
n

k=1
E[[F(t
k1
)[
2
](t
k
t
k1
)
2
2[[
n

k=1
E[[F(t
k1
)[
2
(t
k
t
k1
)] 0,
as [[ 0. The conclusion follows.
Now we are in position to prove It os formula. First we assume that b
and are elementary processes,
b =
p

i=1
b
i1
1l
[
i1
,
i
)
, =
p

i=1

i1
1l
[
i1
,
i
)
, (6.8)
where p N, 0 =
0
<
1
< <
p
and b
i
,
i
are F
t
i
-measurable for any
i = 0, 1, ..., p 1.
Lemma 6.3 Let C
2
b
(R), x R, b and given by (6.8) and X by (6.1).
Then identity (6.2) holds.
Proof. Since C
3
b
(R) is dense in C
2
b
(R) it is enough to show (6.2) when
C
3
b
(R). We start by proving (6.2) in [0, t] with t
1
. In this case we
have
b(t) = b
0
, (t) =
0
, t [0,
1
]
and
X(t) = b
0
t +
0
B(t), t [0,
1
].
Let = t
0
= 0 < t
1
< < t
N
= t. Then we obviously have
(X(t)) (x) =
N

k=1
[(X(t
k
)) (X(t
k1
))].
The Ito formula 79
On the other hand, using Taylors formula we can write
(X(t)) (x) =
N

k=1

(X(t
k1
))(X(t
k
) X(t
k1
))
+
1
2
N

k=1

(X(t
k1
))(X(t
k
) X(t
k1
))
2
+ R

=: I
1
+ I
2
+ I
3
. (6.9)
Concerning I
1
we have
I
1
=
N

k=1

(X(t
k1
))(b
0
(t
k
t
k1
) +
0
(B(t
k
) B(t
k1
)).
So,
lim
||0
I
1
=
_
t
0

(X(s))b(s)ds +
_
t
0

(X(s))(s)dB(s) in L
2
(, F, P).
(6.10)
Concerning I
2
we write
2I
2
=
N

k=1

(X(t
k1
))b
2
0
(t
k
t
k1
)
2
+ 2
N

k=1

(X(t
k1
))b
0

0
(t
k

k1
)(B(t
k
) B(t
k1
))
+
N

k=1

(X(t
k1
))
2
0
(B(t
k
) B(t
k1
))
2
=: I
2,1
+ I
2,2
+ I
2,3
. (6.11)
It is easy to check that
lim
||0
I
2,1
= lim
||0
I
2,2
= 0 in L
1
(, F, P) (6.12)
In fact
[I
2,1
[
1
2
||
2
[b
0
[
2
N

k=1
(t
k
t
k1
)
2
0 as [[ 0
80 Chapter 6
and
(1)
E[I
2,2
[ ||
2
[b
0
[ [
0
[
N

k=1
(t
k
t
k1
)E[B(t
k
) B(t
k1
)[
||
2
[b
0
[ [
0
[
N

k=1
(t
k
t
k1
)
3/2
0 as [[ 0.
Moreover, by Lemma 6.2 it follows that
lim
||0
2I
2,3
=
_
t
0

(X(s))
2
(s)ds in L
2
(, F, P). (6.13)
So, the conclusion will follow provided
lim
||0
E[R

[ = 0, (6.14)
Let us prove (6.14). We have
R

=
N

k=1
_
1
0
(1 )[

(
k
)

(X(t
k1
))](X(t
k
) X(t
k1
))
2
d,
where

k
= (1 )X(t
k1
) + X(t
k
).
Since C
3
b
(R) we have by the mean value theorem,
[

(
k
)

(X(t
k1
))[ ||
0
(1 )[X(t
k
) X(t
k1
)[,
so that, we deduce setting 1 1,
[R

[ ||
3
N

k=1
[X(t
k
) X(t
k1
)[
3
.
Consequently
[R

[ 3||
3
[b
0
[
3
N

k=1
[t
k
t
k1
[
3
+ 3||
3
[
0
[
3
N

k=1
[B(t
k
) B(t
k1
)[
3
(1)
since E[B(t)[ [E[B
2
(t)[]
1/2
= t
1/2
.
The Ito formula 81
and so
(2)
,
E([R

[) 3||
3
[b
0
[
3
N

k=1
[t
k
t
k1
[
3
+ 3||
3
[
0
[
3

15
N

k=1
[t
k
t
k1
[
3/2
0,
as [[ 0. The proof is complete when t
1
. The general case can be
treated in the same way taking into account that b
k1
and
k1
are indepen-
dent of B(t
k
) B(t
k1
).
We nally prove
Theorem 6.4 Let x R, b, L
2
([0, T] , P, dt P) and C
2
b
(R).
Then identity (6.2) holds for all t [0, T].
Proof. Let (b
j
) and (
j
) be sequences of elementary processes such that
lim
j
b
j
= b, lim
j

j
= in L
2
([0, T] , P, dt P).
Set, for any j N,
X
j
(t) = x +
_
t
0
b
j
(s)ds +
_
t
0

j
(s)dB(s), s [0, T]. (6.15)
Then we have (see (5.10))
lim
j
X
j
= X in C
B
([0, T]; L
2
()).
Moreover by (6.2) we have
(X
j
(t)) = (x) +
_
t
0

(X
j
(s))
j
(s)dB(s),
+
_
t
0
_
1
2

j
(s)

(X
j
(s)) + b
j
(s)

(X
j
(s))
_
ds.
(6.16)
Now the conclusion follows by the dominated convergence theorem letting
j .
Taking expectation in the Ito formula we nd a useful identity which
allows to estimate the expectation of (X(t)).
(2)
Since E[B(t)[
3
) [E(B(t)
6
)]
1/2
=

15.
82 Chapter 6
Proposition 6.5 Assume that x R, b, L
2
([0, T] , P, dt P) and
C
2
b
(R). Let
X(t) = x +
_
t
0
b(s)ds +
_
t
0
(s)dB(s), t [0, T].
Then
E[(X(t))] = (x) +
1
2
E
_
t
0
[

(X(s))
2
(s) + 2

(X(s))b(s)]ds. (6.17)
6.1.1 The It o formula for unbounded functions
We want now to show that formula (6.17) also holds without the assumption
that is bounded, provided the integrand in the right hand side is summable.
Proposition 6.6 Assume that x R, b, L
2
([0, T] , P, dt P) and
C
2
(R). Set
X(t) = x +
_
t
0
b(s)ds +
_
t
0
(s)dB(s), t [0, T]. (6.18)
and assume in addition that
E
_
t
0
[

(X(s))
2
(s) + 2

(X(s))b(s)[ds < +. (6.19)


Then E[(X(t))] < + and (6.17) holds.
Example 6.7 Take (x) = x
2
. Then condition (6.19) becomes
E
_
t
0
[
2
(s) + 2X(s)b(s)[ds < +
which is clearly fullled. Then
E([X(t)[
2
) = [x[
2
+E
_
t
0
(
2
(s) + 2X(s)b(s))ds.
Proof of Proposition 6.6. For any R > 0 consider a function
R

C
2
b
(R) such that

R
(x) =
_
(x) if [x[ R,
0 if [x[ R + 1.
The Ito formula 83
Then, applying Itos formula (6.2) to
R
(X(t)), yields for any R > 0

R
(X(t)) (x) =
1
2
_
t
0
[

R
(X(s))
2
(s) + 2

R
(X(s)b(s)]ds
+
_
t
0

R
(X(s)))(s)dB(s).
(6.20)
Let now
R
be the stopping time

R
=
_

_
inft [0, T] : [X(t)[ R if sup
t[0,T]
[X(t)[ R,
T if sup
t[0,T]
[X(t)[ < R.
It is clear that
R
is increasing and bounded by T. We know that X(, ) is
continuous for almost all . For such a , X(, ) attains the maximum,
say M(). Then we have
R
() = T for all R > M(). So,
lim
R

R
= T Pa.s.. (6.21)
Now, in view of Proposition 5.16 we can write
(X(t
R
)) (x) =
1
2
_
t
0
1l
s<(t
R
)
[

(X(s))
2
(s) + 2

(X(s)b(s)]ds
+
_
t
0
1l
s<(t
R
)

(X(s)))(s)dB(s).
(6.22)
Taking expectation we obtain
E[(X(t
R
))] (x)
=
1
2
E
_
t
0
1l
s<(t
R
)
[

(X(s))
2
(s) + 2

(X(s)b(s)]ds.
(6.23)
Now, by the assumption (6.19), (6.21) and the dominated convergence theo-
rem, we can let R obtaining the conclusion.
As an application of Proposition 6.6 let us estimate E
_
_
T
0
F(s)dB(s)
_
2m
where F is predictable and m N, m > 1.
84 Chapter 6
Proposition 6.8 Assume that F L
2m
([0, T] ; P, dt P), m N, and
set
X(t) =
_
t
0
F(s)dB(s), t [0, T].
Then X L
2m
([0, T] ; P, dt P) and we have
E[[X(T)[
2m
] [m(2m1)]
m
T
m1
_
T
0
E
_
[F(t)[
2m

dt. (6.24)
Proof. It is enough to prove (6.24) when F is bounded (because L

([0, T]
; P, dt P) is dense in L
2m
([0, T] ; P, dt P)).
We start from the case m = 2, setting (x) = x
4
. Then (6.19) holds so
that, by Proposition 6.6 we have
E[[X(t)[
4
] = 6E
__
t
0
[X(s)[
2
[F(s)[
2
ds
_
.
By Holders inequality it follows that
E[[X(t)[
4
] 6
_
E
_
t
0
[X(s)[
4
ds
_
1/2
_
E
_
t
0
[F(s)[
4
ds
_
1/2
. (6.25)
Integrating between 0 and T, yields
_
T
0
E[X(t)[
4
dt 6T
_
E
_
T
0
[X(t)[
4
dt
_
1/2
_
E
_
T
0
[F(t)[
4
dt
_
1/2
. (6.26)
From which
_
T
0
E[X(t)[
4
dt 36T
2
_
T
0
E[F(t)[
4
dt.
Substituting this in (6.25) yields
E[[X(t)[
4
] 36TE
_
T
0
[F(t)[
4
dt.
So, (6.24) is proved for m = 2. We can now easily iterate the previous
argument taking successively m = 3, 4 and so on.
6.2 It o formula for a vector valued process
Let d, m N. Assume that x R
d
, b L
2
([0, T] ; P, dt P; R
d
) and
L
2
([0, T] ; P, dt P; L(R
m
; R
d
)). Set
X(t) = x +
_
t
0
b(s)ds +
_
t
0
(s)dW(s), t [0, T]
The Ito formula 85
We are going to prove the following Itos formula,
(X(t)) = (x) +
_
t
0
D(X(s)), (s)dB(s)),
+
_
t
0
_
1
2
Tr[(

)(s)D
2
(X(s))] +b(s), D(X(s)))
_
ds,
(6.27)
for all t [0, T]. We shall write (6.27) in the dierential form

(X(t)) = D(X(t)), (t)dB(t))


+
_
1
2
Tr[(

)(t)D
2
(X(t))] +b(t), D(X(t)))
_
dt, t 0,
(6.28)
The proof is similar to that of the one-dimensional case seen before. So, we
shall only sketch some points of the proof. Let us start with a preliminary
lemma.
Lemma 6.9 Let f C
B
([0, T]; L
2
()) and let i, j 1, 2..., m. Then we
have
lim
||0
n

k=1
f(t
k1
)(B
i
(t
k
) B
i
(t
k1
))(B
j
(t
k
) B
j
(t
k1
))
=
i,j
_
T
0
f(s)ds, in L
2
(, F, P).
(6.29)
Proof. Let = 0 = t
0
< t
1
< < t
n
= T be a decomposition of [0, T].
If i = j, (6.29) follows from Lemma 6.2. Let i ,= j and set
I

i,j
:=
n

k=1
f(t
k1
)(B
i
(t
k
) B
i
(t
k1
))(B
j
(t
k
) B
j
(t
k1
)).
Then we have
E[(I

i,j
)
2
] = E
n

h,k=1
f(t
h1
)f(t
k1
)(B
i
(t
h
) B
i
(t
h1
))(B
j
(t
h
) B
j
(t
h1
))
(B
i
(t
k
) B
i
(t
k1
))(B
j
(t
k
) B
j
(t
k1
))
= E
n

h=1
f
2
(t
h1
)(B
i
(t
h
) B
i
(t
h1
))
2
(B
j
(t
h
) B
j
(t
h1
))
2
=
n

h=1
E(f
2
(t
h1
))(t
h
t
h1
)
2
0,
86 Chapter 6
as [[ 0.
Now we prove Itos formula when b and are elementary processes as,
b =
p

i=1
b
i1
1l
[
i1
,
i
)
, =
p

i=1

i1
1l
[
i1
,
i
)
, (6.30)
where p N, 0 =
0
<
1
< <
p
b
i
L
2
(, F
t
i
, P; R
d
) and
i

L
2
(, F
t
i
, P; L(R
m
; R
d
)) i = 0, 1, ..., p 1.
Lemma 6.10 Let C
2
b
(R
d
), x R
d
and let b and given by (6.30).
Then identity (6.27) holds.
Proof. We proceed as in the proof of Lemma 6.3, taking C
3
b
(R
d
) and
proving (6.6) in [0, t] with t
1
. We have
b(t) = b
0
, (t) =
0
, t [0,
1
]
and
X(t) = b
0
t +
0
B(t), t [0,
1
].
Let = t
0
= 0 < t
1
< < t
N
= t. Then we obviously have
(X(t)) (x) =
N

k=1
[(X(t
k
)) (X(t
k1
))].
On the other hand, by Taylors formula we can write
(3)
(X(t)) (x) =
N

k=1
D(X(t
k1
)), X(t
k
) X(t
k1
))
+
1
2
N

k=1
D
2
(X(t
k1
))(X(t
k
) X(t
k1
)), X(t
k
) X(t
k1
)) + R

=: I
1
+ I
2
+ I
3
. (6.31)
Concerning I
1
we have
I
1
=
N

k=1
D(X(t
k1
)), b
0
(t
k
t
k1
) +
0
(B(t
k
) B(t
k1
)).
(3)
We use the notations D(x)h = D(x), h) and D
2
(x)(h, k) = D
2
(x)h, k) for all
x, h, k R
d
.
The Ito formula 87
So,
lim
||0
I
1
=
_
t
0
D(X(s)), b(s))ds+
_
t
0
D(X(s)), (s)dB(s)) in L
2
(, F, P).
(6.32)
Concerning I
2
we write
2I
2
=
N

k=1
D
2
(X(t
k1
))b
0
, b
0
)(t
k
t
k1
)
2
+ 2
N

k=1
D
2
(X(t
k1
))b
0
,
0
(B(t
k
) B(t
k1
)))(t
k
t
k1
)
+
N

k=1
D
2
(X(t
k1
))
0
(B(t
k
)B(t
k1
)),
0
(B(t
k
)B(t
k1
))) =: I
2,1
+I
2,2
+I
2,3
.
(6.33)
It is easy to check that
lim
||0
I
2,1
= lim
||0
I
2,2
= 0 in L
1
(, F, P) (6.34)
Moreover, we have
2I
2,3
=
N

k=1
D
2
(X(t
k1
))((B(t
k
) B(t
k1
))), (B(t
k
) B(t
k1
)))
=
N

k=1
d

i,j=1
m

,=1
D
2
i,j

i,
(B

(t
k
) B

(t
k1
))
i,
(B

(t
k
) B

(t
k1
)).
Therefore, taking into account Lemma 6.9 we have
lim
||0
2I
2,3
=
_
t
0
d

i,j=1
m

=1
D
2
i,j
(X(s))
i,
(s)
i,
(s)ds
=
_
t
0
Tr [D
2
(X(s))(

(s))]ds.
Now, proceeding as before, we see that
lim
||0
E[R

[ = 0, (6.35)
88 Chapter 6
The proof is complete when t
1
. The general case can be treated in
the same way taking into account that b
k1
and
k1
are independent of
B(t
k
) B(t
k1
).
Finally, proceeding as we did for the proof of Theorem 6.4 we obtain the
result
Theorem 6.11 Let b L
2
([0, T] , P, dt P : R
d
), L
2
([0, T]
, P, dt P : L(R
m
; R
d
)), x R
d
and C
2
b
(R
d
). Then identity (6.27)
holds for any t [0, T].
Exercise 6.12 Let d = 1, m N, b,
k
L
2
([0, T] , P, dt P), k =
1, ..., m.
Set
X(t) =
_
t
0
b(s)ds +
m

k=1
_
t
0

k
(s)dB
k
(s).
Let C
2
b
(R). Prove that
d(X(t)) =

(X(t))dX(t) +
1
2

(X(t))[(t)[
2
dt, (6.36)
where (t) = (
1
(t), ...,
m
(t)).
Exercise 6.13 Let d N, m = 1 b
i
,
i
L
2
([0, T] , P, dt P), i = 1, 2 =
..., d. Set
X(t) = b(t)dt + dB(t), i = 1, 2,
where = (
1
, ...,
d
). Let moreover C
2
b
(R
d
). Prove that
d(X(t)) = D(X(t)), dX(t)) +
1
2
D
2
(X(t))(t), (t))dt. (6.37)
Chapter 7
Stochastic evolution equations
We are given two positive integers r, d and an r-dimensional standard Brow-
nian motion B(t), t 0, in a probability space (, F, P). We denote by
(F
t
)
t0
the natural ltration of B(t) (augmented with all P-null sets of ).
Let us consider the following integral equation
X(t) = +
_
t
s
b(u, X(u))du +
_
t
s
(u, X(u))dB(u), t [s, T], (7.1)
where s [0, T), L
2
(, F
s
, P; R
d
), b: [0, T] R
d
R
d
and : [0, T]
R
d
L(R
r
, R
d
). b is called the drift and the diusion coecient of the
equation.
We shall write (7.1) in dierential form as
_
_
_
dX(t) = b(t, X(t))dt + (t, X(t))dB(t),
X(s) = .
(7.2)
By a solution of equation (7.1) on the interval [s, T] we mean a function
X C
B
([s, T]; L
2
(; R
d
)) that fullls equation (7.1).
In order to solve (7.1) we shall use a xed point argument, based on the
identity
E

_
b
a
G(t)dB(t)

2
=
_
b
a
E[Tr (G(t)G

(t))] dt.
for all G C
B
([0, T]; L
2
(, L(R
r
, R
d
))) and 0 a < b T. This suggests to
endow L(R
r
, R
d
) with the HilbertSchmidt norm, setting
|S|
HS
: = [Tr(SS

)]
1/2
, S L(R
r
, R
d
)
and to write
E

_
b
a
G(t)dB(t)

2
=
_
b
a
E
_
|G(t)|
2
HS
_
dt. (7.3)
89
90 Chapter 7
7.1 Existence and uniqueness
The standard assumptions for the well-posedness of problem (7.1) are the
following.
Hypothesis 7.1
(i) b and are continuous on [0, T] R
d
.
(ii) There exists M > 0 such that for all t [0, T], x, y R
d
, we have
[b(t, x) b(t, y)[
2
+|(t, x) (t, y)|
2
HS
M
2
[x y[
2
(7.4)
and
[b(t, x)[
2
+|(t, x)|
2
HS
M
2
(1 +[x[
2
). (7.5)
Notice that, after possibly changing the constant M, (7.5) is a consequence
of (7.4).
Theorem 7.1 Assume that Hypothesis 7.1 holds and let s [0, T),
L
2
(, F
s
, P; R
d
). Then problem (7.1) has a unique solution
X C
B
([s, T]; L
2
(; R
d
)).
Proof. We are going to solve (7.1) by a xed point argument in the space
C
B
:= C
B
([s, T]; L
2
(; R
d
)).
Dene

1
(X)(t) :=
_
t
s
b(u, X(u))du, X C
B
, t [s, T],

2
(X)(t) :=
_
t
s
(u, X(u))dB(u), X C
B
, t [s, T]
and set
(X) := +
1
(X) +
2
(X), X C
B
.
Then equation (7.1) is equivalent to the following,
X = +
1
(X) +
2
(X) = (X). (7.6)
Step 1.
1
and
2
map C
B
into itself.
Stochastic evolution equations 91
Concerning
1
we have, using the H older inequality and taking into ac-
count (7.5),
[
1
(X)(t)[
2
(t s)
_
t
s
[b(u, X(u))[
2
du M
2
(t s)
_
t
s
(1 +[X(u)[
2
)du
M
2
(t s)
2
(1 +|X|
2
C
B
).
Since
1
(X)(t) is F
t
measurable for all t [s, T],
1
maps C
B
into itself and
|
1
(X)|
C
B
M(t s)(1 +|X|
C
B
).
Concerning
2
we have taking into account (7.3) and (7.5),
E[
2
(X)(t)[
2
=
_
t
s
E(|(u, X(u))|
2
HS
)du
M
2
_
t
s
(1 +[X(u)[
2
)du M
2
(t s)(1 +|X|
2
C
B
)
So, we see that
2
maps C
B
into itself.
Step 2. is Lipschitz continuous.
Let X, Y C
B
. We have, using again the Holder inequality and taking
into account (7.4),
[
1
(X)(t)
1
(Y )(t)[
2
(t s)
_
t
s
[b(u, X(u)) b(u, Y (u))[
2
du
(t s)M
2
_
t
s
[X(u) Y (u)[
2
du (t s)
2
M
2
|X Y |
2
C
B
du.
Consequently
|
1
(X)
1
(Y )|
C
B
M (T s) |X Y |
C
B
, X, Y C
B
(7.7)
Furthermore
E[
2
(X)(t)
2
(Y )(t)[
2
=
_
t
s
E(|(u, X(u)) (u, Y (u))|
2
HS
)du
M
2
(t s)|X Y |
2
C
B
,
92 Chapter 7
and so,
|
2
(X)
2
(Y )|
C
B
M

T s |X Y |
C
B
, X, Y C
B
. (7.8)
By (7.7) and (7.8) it follows that maps C
B
into itself and
|(X) (Y )|
C
B
M(T s +

T s )|X Y [|
C
B
,
for all X, Y C
B
. Now if T s is such that
M
_
T s +

T s
_
1/2, (7.9)
is a 1/2contraction on C
B
, and so, it possesses a unique xed point. If
(7.9) does not hold we choose T
1
(s, T] such that
M
_
T
1
s +
_
T
1
s
_
1/2.
Then by the previous argument there is a unique solution to (7.1) on [s, T
1
].
Now we repeat the proof with T
1
replacing s and in a nite number of steps
we arrive to the conclusion.
Remark 7.2 By Theorem 5.13 it follows that there exists a version of the
solution X(, s, ) which belongs to L
2
(, C([s, T])) and so it is a continuous
process.
In the following we shall denote by X(, s, ) the solution of problem (7.1).
Whe shall use greek letters for stochastic initial data and latin letters for
deterministic ones.
Let us prove the co-cycle law.
Proposition 7.3 Assume that Hypothesis 7.1 holds and let L
2
(, F
s
, P; R
d
).
Then
X(t, s, ) = X(t, r, X(r, s, )), 0 s r t T. (7.10)
Proof. Dene Z(t) = X(t, s, ), t [s, T]. Then Z solves the problem
_
_
_
dZ(t) = b(t, Z(t))dt + (t, Z(t))dB(t),
Z(r) = X(r, s, ).
By the uniqueness part of Theorem 7.1 it follows that
Z(t) = X(t, s, ) = X(t, r, X(r, s, )),
as required.
Stochastic evolution equations 93
Remark 7.4 By the contraction principle it follows that the solution X(t, s, )
of problem (7.1) can be obtained as a limit of successive approximations.
More precisely, dene X
0
(t, s, ) = and for any N N,
X
N+1
(t, s, ) = +
_
t
s
b(u, X
N
(u, s, ))du +
_
t
s
(u, X
N
(u, s, ))dB(u).
(7.11)
Then we have
lim
N
X
N
(, s, ) = X(, s, ) in C
B
([s, T]; L
2
(; R
d
)). (7.12)
Next result, which as we shall see plays an important role in proving that
X(, s, x) is a Markov process, gives some information about the relationship
between X(t, s, ), L
2
(, F
s
, P; R
d
) and X(t, s, x), x R
d
.
Proposition 7.5 Assume that Hypothesis 7.1 holds and that
=
n

k=1
x
k
1l
A
k
, (7.13)
where x
1
, ..., x
n
R
d
, and A
1
, ..., A
n
are mutually disjoints sets in F
s
such
that
=
n
_
k=1
A
k
.
Then we have
X(t, s, ) =
n

k=1
X(t, s, x
k
)1l
A
k
. (7.14)
Proof. Let X
N
be dened by (7.11). We claim that
X
N
(t, s, ) =
n

k=1
X
N
(t, s, x
k
)1l
A
k
, N N. (7.15)
Once (7.15) is proved, the conclusion follows letting N tend to innity. Let
us proceed by recurrence. Equality (7.15) is clear for N = 0. Assume that it
holds for a given N N, so that
X
N
(t, s, ) = X
N
(t, s, x
k
) in A
k
, k = 1, ..., n.
Then we have
b(u, X
N
(u, s, )) = b(u, X
N
(u, s, x
k
)) in A
k
, k = 1, ..., n,
(u, X
N
(u, s, )) = (u, X
N
(u, s, x
k
)) in A
k
, k = 1, ..., n,
94 Chapter 7
so that
b(u, X
N
(u, s, )) =
n

k=1
1l
A
k
b(u, X
N
(u, s, x
k
)),
(u, X
N
(u, s, )) =
n

k=1
1l
A
k
(u, X
N
(u, s, x
k
)).
Consequently
X
N+1
(t, s, ) =
n

k=1
1l
A
k
_
X
0
(t, s, x
k
) +
_
t
s
b(u, X
N
(u, s, x
k
)du
+
_
t
s
(u, X
N
(u, s, x
k
))dB(u)
_
=
n

k=1
1l
A
k
X
N+1
(t, s, x
k
)
and (7.15) holds for N + 1. So, the conclusion follows.
7.1.1 Solution of the stochastic dierential equation in
the space C
B
([s, T]; L
2m
(; R
d
)).
Theorem 7.6 Assume that Hypothesis 7.1 holds and let m N, s [0, T),
L
2m
(, F
s
, P; R
d
). Then problem (7.1) has a unique solution
X(, s, ) C
B
([s, T]; L
2m
(; R
d
)).
In particular
X(, s, x) C
B
([s, T]; L
2m
(; R
d
)), x R
d
.
Proof. We proceed as in the proof of Theorem 7.1 by a xed point argument
in the space
C
m
B
:= C
B
([s, T]; L
2m
(; R
d
)),
using inequality (6.24) proved in Proposition 6.8.
7.1.2 Examples
Example 7.7 Consider the stochastic dierential equation
dX = AXdt + CdB(t), X(0) = x, (7.16)
where A L(R
d
), C L(R
r
; R
d
) and x R
d
.
Stochastic evolution equations 95
Clearly Theorem 7.1 applies so that (7.16) has a unique solution X(t)
which fullls the integral equation
X(t) = x + A
_
t
0
X(s)ds + CB(t). (7.17)
Setting
Y (t) =
_
t
0
X(s)ds, t [0, T],
Y fullls the equation
Y

(t) = AY (t) + x + CB(t), Y (0) = 0, t [0, T],


which can be easily solved by the method of variation of constants. We
obtain
Y (t) =
_
t
0
e
(ts)A
(x + CB(s))ds, t [0, T].
By substituting Y (t) in (7.17) yields
X(t) = A
_
t
0
e
(ts)A
(x + CB(s))ds + x + CB(t).
Taking into account that, thanks to Proposition 3.12,
_
t
0
e
(ts)A
CdB(s) = CB(t) + A
_
t
0
e
(ts)A
CB(s)ds,
we nd
X(t) = e
tA
x +
_
t
0
e
(ts)A
CdB(s). (7.18)
Example 7.8 Let r = d = 1 and consider the stochastic dierential equation
dX = aXdt + cXdB(t), X(0) = x, (7.19)
where a, c, x R. Again Theorem 7.1 applies. We want to show that the
solution of (7.19) is given by
X(t) = e
t
(
a
1
2
c
2
)
e
cB(t)
x, t 0. (7.20)
For this we check that X(t) given by (7.20) solves (7.19).
Write X(t) = e
F(t)
where F(t) = t
_
a
1
2
c
2
_
+ cB(t). Then we have
dF(t) =
_
a
1
2
c
2
_
dt + cdB(t)
96 Chapter 7
and, by Itos formula,
dX(t) = e
F(t)
dF(t) +
1
2
c
2
e
F(t)
dt
= e
F(t)
_
a
1
2
c
2
_
dt + cdB(t) +
1
2
c
2
e
F(t)
dt
= aX(t)dt + cX(t)dB(t).
Exercise 7.9 Let r = 1 and consider the dierential stochastic equation
dX = AXdt + CXdB(t), X(0) = x, (7.21)
where A, C L(R
d
), x R
d
and AC = CA. Show that the solution of (7.21)
is given by
X(t) = e
t(AC
2
/2)
e
CB(t)
x. (7.22)
7.1.3 Dierential stochastic equations with random co-
ecients
In some situations (see Subsections 7.3 and 7.4) one deals with stochastic
dierential equations having random coecients,
X(t, ) = () +
_
t
s
b(u, X(u, ), )du +
_
t
s
(u, X(u, ), )dB(u). (7.23)
Here L
2
(, F
s
, R
d
), b: [0, T] R
d
R
d
and : [0, T] L(R
r
, R
d
)
R
d
are such that:
Hypothesis 7.2
(i) There exists M > 0 such that for all t [0, T], x, y R
d
,
[b(t, x, )b(t, y, )[
2
+|(t, x, )(t, y, )|
2
HS
M
2
[xy[
2
(7.24)
and
[b(t, x, )[
2
+|(t, x, )|
2
HS
M
2
(1 +[x[
2
). (7.25)
(ii) For any Y C
B
([0, T]; L
2
(, R
d
)) we have U C
B
([0, T]; L
2
(, R
d
))
and V C
B
([0, T]; L
2
(, L(R
r
, R
d
))) where, for all t [0, T], ,
U(t, ) = b(t, Y (t, ), )), V (t, ) = (t, Y (t, ), )).
The following result can be proved as Theorem 7.1.
Stochastic evolution equations 97
Theorem 7.10 Assume that Hypothesis 7.2 holds. Let s [0, T) and
L
2
(, F
s
, R
d
). Then problem (7.23) has a unique solution
X C
B
([s, T]; L
2
(; R
d
)).
Example 7.11 Let d = 1 and consider the stochastic dierential equation
_
_
_
dX(t) = X(t)F(t), dB(t)), t [0, T],
X(0) = x,
(7.26)
where F C
B
(0, T; L

(; R
d
)). Now it is easy to check that Theorem 7.10
applies and so there exists a solution X of (7.26). Let us show that
X(t) = e

1
2
R
t
0
|F(s)|
2
ds+
R
t
0
F(s),dB(s)
x, t 0. (7.27)
For this we check that X(t) given by (7.27) solves (7.26).
Write X(t) = e
H(t)
where
H(t) =
1
2
_
t
0
[F(s)[
2
ds +
_
t
0
F(s), dB(s)).
Then we have
dH(t) =
1
2
[F(t)[
2
dt +F(t), dB(t)), t 0.
Now by Itos formula we nd
dX(t) = e
H(t)
dH(t) +
1
2
e
H(t)
[F(t)[
2
dt
= e
H(t)
F(t), dB(t)) = X(t)F(t), dB(t)), t 0.
So, (7.27) is proved.
7.2 Continuous dependence on data
7.2.1 Continuous dependence on mean square
We assume here that Hypothesis 7.1 holds. We are going to prove that
the solution X(t, s, ) to (7.1) is H older continuous on t, s and Lipschitz
continuous on in mean square. First we show that E[X(t, s, )[
2
is bounded.
98 Chapter 7
Lemma 7.12 Assume that Hypothesis 7.1 holds. Then for all s [0, T] and
L
2
(, F
s
, P; R
d
) we have
E
_
[X(t, s, )[
2
_
3[E([[
2
) + M
2
((T s)
2
+ (T s)]e
3M
2
(Ts+1)
. (7.28)
Proof. Writing for short X(t, s, ) = X(t), we have
E([X(t)[
2
) 3E([[
2
) + 3E
_

_
t
s
b(u, X(u))du

2
_
+3
_
t
s
E(|(u, X(u))|
2
HS
)du.
By Hypothesis 7.1(ii) and the H older inequality we deduce that
E([X(t)[
2
) 3E([[
2
) + 3M
2
(t s)
_
t
s
(1 +E
_
[X(u)[
2
_
)du
+3M
2
_
t
s
(1 +E
_
[X(u)[
2
_
)du.
Consequently
E([X(t)[
2
) 3E([[
2
) + 3M
2
((T s)
2
+ (T s))
+3M
2
((T s) + 1)
_
t
s
E
_
[X(u)[
2
_
du.
The conclusion follows from the Gronwall lemma.
We now study the regularity of X(t, s, ) with respect to t, s, . We note
that, by Lemma 7.12, there exists a constant C(T, E([[
2
)) such that
E
_
[X(t, s, )[
2
_
C(T, E([[
2
)), 0 s < t T. (7.29)
We start with the regularity of X(t, s, ) with respect to t.
Proposition 7.13 Assume that Hypothesis 7.1 holds. Let 0 s t
1
< t
T and L
2
(, F
s
, R
d
). Then there exists a constant C
1
(T, E([[
2
)) such
that we have
E
_
[X(t, s, ) X(t
1
, s, )[
2
_
C
1
(T, E([[
2
))(t t
1
). (7.30)
Stochastic evolution equations 99
Proof. We have
E
_
[X(t, s, ) X(t
1
, s, )[
2
_
2M
2
(t t
1
)
_
t
t
1
(1 +E
_
[X(u, s, )[
2
_
du
+ 2M
2
_
t
t
1
(1 +E
_
[X(u, s, )[
2
_
)du.
Consequently,
E
_
[X(t, s, ) X(t
1
, s, )[
2
_
2M
2
((t t
1
)
2
+ t t
1
)(1 + C
2
(T, E([[
2
)))
and the conclusion follows.
Let us study the regularity of X(t, s, ) with respect to .
Proposition 7.14 Assume that Hypothesis 7.1 holds, let 0 s < t T and
, L
2
(, F
s
, R
d
). Then
E
_
[X(t, s, ) X(t, s, )[
2
_
3e
3M
2
(Ts+1)(ts)
E([ [
2
). (7.31)
Proof. We have
[X(t, s, ) X(t, s, )[
2
3[ [
2
+ 3

_
t
s
(b(u, X(u, s, ) b(u, X(u, s, ))du

2
+ 3

_
t
s
((u, X(u, s, ) (u, X(u, s, ))dB(u)

2
.
Taking expectation and using (7.4) we obtain
E([X(t, s, ) X(t, s, )[
2
) 3E([ [
2
) + 3M
2
(T s + 1)

_
t
s
E
_
[X(u, s, ) X(u, s, )[
2
_
du
and the conclusion follows from the Gronwall lemma.
We nally study the regularity of X(t, s, ) with respect to s.
Proposition 7.15 Assume that Hypothesis 7.1 holds, let 0 < s < s
1
< t
T, and L
2
(, F
s
, P; R
d
). Then there exists a constant C

T,
> 0 such that
E
_
[X(t, s, ) X(t, s
1
, )[
2
_
C

T,
[s s
1
[. (7.32)
100 Chapter 7
Proof. Taking into account the co-cycle law (7.10), we can write
X(t, s, ) X(t, s
1
, ) = X(t, s
1
, X(s
1
, s, )) X(t, s
1
, ).
By (7.31) there exists C
T
> 0 such that
E([X(t, s, ) X(t, s
1
, )[
2
) C
2
T
E([X(s
1
, s, ) [
2
)
= C
2
T
E([X(s
1
, s, ) X(s, s, )[
2
) .
The conclusion follows now from (7.30).
7.3 Almost sure continuity and holderianity
of trajectories
In this section we show that X(, s, x) belongs to a suitable Sobolev space,
whose denition is recalled in Appendix E below. Then the Sobolev embed-
ding theorem (also stated in Appendix E) will imply that X(, s, x) is H older
continuous almost surely.
First we need a lemma, which can be proved as Proposition 7.13 using
(6.24).
Lemma 7.16 Assume that Hypothesis 7.1 holds. Let 0 s t
1
< t
T, x R
d
and m N. Then there exists a constant C
1
(T, [x[) such that
E
_
[X(t, s, x) X(t
1
, s, x)[
2m
_
C
1
(T, [x[
2
))(t t
1
)
m
. (7.33)
Now from Proposition E.3 and the Sobolev embedding theorem E.1 it
follows that
Proposition 7.17 Assume that Hypothesis 7.1 holds. Let x R
d
, 0 s
t T, m N and (0, 1/2). Then we have
E
_
[X(, s, x)[
2m
,2m

< +. (7.34)
Moreover, X(, s, x) belongs to C
1/(2m)
([s, T]) almost surely.
Finally, we consider almost sure regularity of X(t, s, ). First, arguing as
in the proof of Proposition 7.14 we have
Lemma 7.18 Assume that Hypothesis 7.1 holds, let 0 s < t T and
x, y R
d
. Then there is a constant C(T) > 0 such that
E
_
[X(t, s, x) X(t, s, y)[
2m
_
C(T)[x y[
2m
. (7.35)
Stochastic evolution equations 101
Now from Proposition E.3 it follows that
Proposition 7.19 Assume that Hypothesis 7.1 holds, let 0 s < t T and
x, y [0, 1]
d
. Then for any m > 1 and (0, 1) we have
E
_
[X(t, s, )[
2m
,2m

< +. (7.36)
Moreover, X(t, s, ) belongs to C
d/(2m)
([0, 1]
d
) almost surely.
7.4 Dierentiability of X(t, s, x) with respect
to x
In this section we assume, besides Hypothesis 7.1, that
Hypothesis 7.3
(i) D
x
b, D
2
x
b, D
x
and D
2
x
are continuous on [0, T] R
d
.
(ii) We have
(1)
sup
t[0,T]
([b(t, )]
2
+ [(t, )]
2
) < . (7.37)
We set
C
B
= C
B
([s, T]) =: C
B
([s, T]; L
2
(; R
d
)).
7.4.1 Existence of X
x
(t, s, x)
Theorem 7.20 Assume that Hypotheses 7.1 and 7.3 hold. Then for any
s [0, T] the mapping
R
d
C
B
, x X(, s, x),
is continuously Gateaux dierentiable and its Gateaux derivative is given by
X
x
(t, s, x) h =
h
(t, s, x), x, h R
d
, (7.38)
where
h
(t, s, x) is the solution to the stochastic dierential equation with
random coecients,
_

_
d
h
(t, s, x) = b
x
(t, X(t, s, x))
h
(t, s, x)dt
+
x
(t, X(t, s, x))(
h
(t, s, x), dB(t))

h
(s, s, x) = h.
(7.39)
(1)
Recall the notations given at the beginning of Chapter 6.
102 Chapter 7
Proof. Note that the coecients of equation (7.39) fulll Hypothesis 7.3, so
it possesses a unique solution by Theorem 7.10.
To prove the theorem we use Theorem D.6 from Appendix D (with = R
d
and E = C
B
). We set C
B
= C
B
([s, T
1
]) and dene a mapping
F : R
d
C
B
C
B
,
setting
[F(x, X)](t): = x +
_
t
s
b(r, X(r))dr +
_
t
s
(r, X(r))dB(r), t [s, T
1
],
(7.40)
where T
1
> s is chosen such that
|F(x, X
1
) F(x, X
2
)|
C
B

1
2
|X
1
X
2
|
C
B
for all X
1
, X
2
C
B
, x R
d
.
(7.41)
Then F fullls Hypothesis D.1 so that it possesses a unique xed point
X(x) C
B
, that is
F(x, X(x)) = X(x), x R
d
,
which depends continuously on x. X(x) coincides with the solution X(, s, x)
of (7.2).
It is not dicult to check that F is Gateaux continuously dierentiable,
(the straightforward proof is left to the reader) and that for each x R
d
,
X, Y C
B
we have
F
x
(x, X) = I,
[F
X
(x, X)Y ](t) =
_
t
s
b
x
(r, X(r))Y (r)dr+
_
t
s

x
(r, X(r))Y (r)dB(r), t [s, T
1
],
So, the conclusion follows from Theorem D.6.
7.4.2 Existence of X
xx
(t, s, x)
We now prove the existence of the second derivative of X(t, s, x) with respect
to x.
Theorem 7.21 Assume that Hypotheses 7.1 and 7.3 hold. Then the mapping
R
d
C
B
, x X(, s, x),
is twice dierentiable with respect to x in any couple of directions (h, k) in
R
d
. Moreover, setting
X
xx
(t, s, x)(h, k) =
h,k
(t, s, x), x, h R
d
, (7.42)
Stochastic evolution equations 103

h,k
(t, s, x) is the solution to the stochastic dierential equation (with random
coecients)
_

_
d
h,k
(t, s, x) = b
x
(t, X(t, s, x))
h,k
(t, s, x)dt
+b
xx
(t, X(t, s, x))(
h
(t, s, x),
k
(t, s, x))dt
+
x
(t, X(t, s, x))(
h,k
(t, s, x), dB(s))
+
xx
(t, X(t, s, x)) (
h
(t, s, x),
k
(t, s, x), dB(t))

h,k
(s, s, x) = 0.
(7.43)
We shall prove the theorem when n = r = 1 for simplicity. We rst prove a
lemma.
Lemma 7.22 Let (, s, x) C
B
([s, T]; L
2
()) be the solution of the equa-
tion
(t, s, x) = 1 +
_
t
s
b
x
(r, X(r, s, x))(r, s, x)dr
+
_
t
s

x
(r, X(r, s, x))(r, s, x)dB(r).
(7.44)
Then (, s, x) C
B
([s, T]; L
4
()) and there exists C > 0 such that
E[(, s, x)[
4
C, s [0, T), x R
d
. (7.45)
Proof. We have,
[(t, s, x)[
4
27 + 27

_
t
s
b
x
(r, X(r, s, x))(r, s, x)dr

4
+27

_
t
s

x
(r, X(r, s, x))(r, s, x)dB(r)

4
.
By using (7.37) and the Holder inequality we see that there exists a constant
C
1
such that
[(t, s, x)[
4
27 + C
1
_
t
s
[(r, s, x)[
4
dr
+C
1

_
t
s

x
(r, X(r, s, x))(r, s, x)dB(r)

4
.
104 Chapter 7
Now, taking expectation on both sides of this inequality and using Corollary
6.8, we nd that
E[(t, s, x)[
4
C
2
(1 +
_
t
s
E[(r, s, x)[
4
dr), 0 s t T, x R,
where C
2
is another constant. The conclusion follows from the Gronwall
lemma.
Proof of Theorem 7.21. We choose T
1
as in (7.41) and C
B
= C
B
([s, T
1
])
as before. By Theorem 7.20 we know that X(t, s, x) is dierentiable with
respect to x and that its derivative (, s, x) = X
x
(, s, x) belongs to C
B
and
fullls equation (7.44). For any x R we dene a linear bounded operator
T(x) from C
B
into C
B
setting for all t [s, T
1
],
(T(x)Z)(t) =
_
t
s
b
x
(r, X(r, s, x))Z(r)dr
_
t
s

x
(r, X(r, s, x))Z(r)dB(r).
(7.46)
Notice that, since (, s, x) C
B
([s, T]; L
4
()), T(x)Z is dierentiable with
respect to x for any Z C
B
([s, T]; L
4
()) and it results
(T

(x)Z)(t) =
_
t
s
b
xx
(r, X(r, s, x))Z(r)(, r, x)dr

_
t
s

xx
(r, X(r, s, x))Z(r)(, r, x)dB(r).
(7.47)
Now we write equation (7.44) as
(, s, x) = 1 + T(x)(, s, x) (7.48)
By (7.41) it follows that
|T(x)|
L(C
B
)
1/2, x R.
Thus the solution of (7.48) is given by
(, s, x) = (1 T(x))
1
(1). (7.49)
From this identity it is easy to show the existence of
x
(, s, x) := (, s, x).
We have in fact, by a straightforward computation

x
(, s, x) = (1 T(x))
1
(T

(x)(, s, x)), (7.50)


Stochastic evolution equations 105
where
T

(x)(, s, x)(t) =
_
t
s
b
xx
(r, X(r, s, x))
2
(, s, x)dr
+
_
t
s

xx
(r, X(r, s, x))
2
(, s, x)dB(r).
(7.51)
Now by (7.50) it follows that

x
(t, s, x) T(x)
x
(, s, x)(t) =
_
t
s
b
xx
(r, X(r, s, x))
2
(, s, x)dr
+
_
t
s

xx
(r, X(r, s, x))
2
(, s, x)dB(r),
and the conclusion follows.
7.5 It o Dierentiability of X(t, s, x) with re-
spect to s.
It is useful to recall rst some results in the deterministic case.
7.5.1 The deterministic case
Let us consider the problem
_
_
_
X

(t) = b(t, X(t)), t [s, T],


X(s) = x,
(7.52)
under Hypotheses 7.1 and 7.3 with = 0. Denote by X(t, s, x) the solution
of (7.52). Let us compute X
s
(t, s, x) (it is well known that X(t, s, x) is C
1
in all variables).
Write
X(t, s, x) = X(t, r, X(r, s, x)), t r s. (7.53)
Dierentiating (7.53) with respect to r yields
0 = X
s
(t, r, X(r, s, x)) + X
x
(t, r, X(r, s, x)) X
t
(r, s, x).
Setting r = s we nd
X
s
(t, s, x) = X
x
(t, s, x)b(s, x),
106 Chapter 7
which is equivalent to
X(t, s, x) = x +
_
t
s
X
x
(t, r, x)b(r, x)dr, 0 s t T. (7.54)
In the next subsection we are going to generalize this formula for the solution
X(t, s, x) of (7.2).
7.5.2 The stochastic case
Here we want to study the dierentiability of X(t, s, x) with respect to s in
a sense to be precised. A diculty arises since the process s X(t, s, x) is
not adapted, because X(t, s, x) is not F
s
-measurable. It happens, however,
that for any s [0, T], X(t, s, x) is measurable with respect to the algebra
F
+
s
generated by all sets of the form
: (B(s
1
()) B(s()), ..., B(s
n
()) B(s())) A ,
where n N, 0 s s
1
< ... < s
n
T and A B(R
n
). The family
(F
+
s
)
s[0,T]
is called the future ltration of B.
Proposition 7.23 Assume that Hypotheses 7.1 holds. Let x R
d
, s [0, T].
Then X(t, s, x) is F
+
s
-measurable.
Proof. Let X
N
(t, s, x) be dened by (7.11), N N. Then X
1
(t, s, x) is
F
+
s
measurable. We have in fact
X
1
(t, s, x) = x +
_
t
s
b(u, x)du +
_
t
s
(u, x)dB(u).
Since
_
t
s
(u, x)dB(u) = lim
||0
n

k=1
(t
k1
, x)(B(t
k
) B(t
k1
)),
where = s = t
0
< t
1
< < t
n
= t, then X
1
(t, s, x) is F
+
s
-measurable.
We end the proof by recurrence.
Now we introduce the backward Ito integral for a process wich is adapted
to the future ltration. For this we need the following result which can be
proved as Lemma 4.3.
Lemma 7.24 Let t
1
< t
2
s, and let L
2
(, F
+
s
, P). Then B(t
2
)B(t
1
)
and are independent.
Stochastic evolution equations 107
We dene C
B
+([0, T]; L
2
(; L(R
r
; R
d
))) by a straightforward generaliza-
tion of the space C
B
([0, T]; L
2
(; L(R
r
; R
d
))) dened in Chapter 5.
The elements of C
B
+([0, T]; L
2
(; L(R
r
; R
d
)) are called stochastic pro-
cesses adapted to the future ltration (F
+
t
) and continuous in quadratic
mean.
Let F C
B
+([0, T]; L
2
(; L(R
r
; R
d
))). For any with = 0 =
s
0
< s
1
< < s
n
= T we set
I

(F) =
n

k=1
F(t
k
)(B(t
k
) B(t
k1
))
The proof of next theorem is completely similar to that of equation (5.10).
Theorem 7.25 For any F C
B
+([0, T]; L
2
(; L(R
r
; R
d
))) there exists the
limit
lim
||0
I

(F) =:
_
T
0
F(s)dB(s), (7.55)
in L
2
(). Moreover we have
E
_
T
0
F(s)dB(s) = 0, (7.56)
and
E

_
T
0
F(s)dB(s)

2
=
_
T
0
E
_
|F(s)|
2
HS

ds. (7.57)
_
T
0
F(s)dB(s) is called the backward Ito integral of the function F in [0, T].
Exercise 7.26 Let t > s. Prove that
_
t
s
B(r)dB(r) =
1
2
(B(t)
2
B(s)
2
+ (t s)).
7.5.3 Backward Itos formula
Theorem 7.27 Assume that Hypotheses 7.1 and 7.3 hold. Then we have
X(t, s, x) x =
_
t
s
X
x
(t, r, x) b(r, x)dr
+
1
2
_
t
s
TR [X
xx
(t, r, x)((r, x), (r, x))]dr
+
_
t
s
X
x
(t, r, x)((r, x), dB(r))),
(7.58)
108 Chapter 7
where
TR [X
xx
(t, r, x)((r, x), (r, x))] =
d

k=1
X
xx
(t, r, x)((r, x)e
k
, (r, x)e
k
)
and (e
k
) is any orthonormal basis in R
d
.
Proof. We take d = r = 1 for simplicity. For any (s, t) we set
[[ = max
k=1,...,n
(t
k
t
k1
).
If (s, t) we have
X(t, s, x) x =
n

k=1
[X(t, s
k
, x) X(t, s
k1
, x)]
=
n

k=1
[X(t, s
k
, x) X(t, s
k
, X(s
k
, s
k1
, x))]
=
n

k=1
X
x
(t, s
k
, x)(x X(s
k
, s
k1
, x))

1
2
n

k=1
X
xx
(t, s
k
, x)(x X(s
k
, s
k1
, x))
2
+ o([[).
(7.59)
Arguing as in the proof of It os formula one can show, after some tedious but
straighforward computations, that
lim
||0
o([[) = 0, P-a.s..
On the other hand we have
X(s
k
, s
k1
, x) x =
_
s
k
s
k1
b(r, X(r, s
k1
, x))dr
+
_
s
k
s
k1
(r, X(r, s
k1
, x))dB(r)
= b(s
k
, x)(s
k
s
k1
) + (s
k
, x)(B(s
k
) B(s
k1
)) + o(s
k
s
k1
).
(7.60)
Stochastic evolution equations 109
(Notice that, since b is deterministic, one can replace in (7.60) b(s
k
, x) with
b(
k
, x) where
k
is any point in [s
k1
, s
k
].) Substituting (7.60) in (7.59) we
nd that
X(t, s, x) x =
n

k=1
X
x
(t, s
k
, x)b(s
k
, x)(s
k
s
k1
)
+
n

k=1
X
x
(t, s
k
, x)(s
k
, x)(B(s
k
) B(s
k1
))
+
1
2
n

k=1
X
xx
(t, s
k
, x)
2
(s
k
, x)(B(s
k
) B(s
k1
))
2
+I
1
() + I
2
() + I
3
() + o
1
([[).
(7.61)
Obviously
lim
||0
I
1
() =
_
t
s
X
x
(r, x)b(r, x)dr.
Concerning I
2
(), we note that it is an integral sum corresponding to the
backward It o integral since X
x
(t, s
k
, x) is F
+
s
k
measurable by Proposition
7.23. Therefore we have
lim
||0
I
2
() =
_
t
s
X
x
(r, x)(r, x)dB(r).
The other terms I
3
() and o
1
([[) can be handled as in the proof of Itos
formula.
In a similar way one can prove the following backward Ito formula.
Theorem 7.28 Let C
2
b
(R
d
). Then for any 0 s < t T, we have
(X(t, s, x)) (x) =
_
t
s
D
x
[(X(t, r, x))], b(r, x))dr
+
1
2
_
t
s
Tr [D
2
x
[(X(t, r, x))](r, x)

(r, x)]dr
+
_
t
s
D
x
[(X(t, r, x))], (r, x)dB(r).
(7.62)
110 Chapter 7
Chapter 8
Kolmogorov equations
8.1 The deterministic case
We consider here the problem
_
_
_
X

(t) = b(t, X(t)), t [s, T],


X(s) = x R
n
,
(8.1)
where s [0, T) and b : [0, T] R
n
R
n
fullls the following hypothesis.
Hypothesis 8.1
(i) b is continuous on [0, T] R
n
.
(ii) There exists M > 0 such that
[b(t, x) b(t, y)[ M[x y[, x, y R
n
, t [0, T].
(iii) b is dierentiable with respect to x and b
x
is continuous on [0, T] R
n
.
As well known, under Hypothesis 8.1 problem (8.1) has a unique solution
X() = X(, s, x) C
1
([s, T]; R
n
), and it holds
X(t, s, x) = X(t, u, X(u, s, x)), 0 s u t T, x R
n
. (8.2)
Morever, dierentiating (8.2) with respect to u and setting u = s we nd
X
s
(t, s, x) + X
x
(t, s, x) b(s, x) = 0, 0 s t T, x R
n
. (8.3)
Of great interest for the applications is the transition evolution operator
P
s,t
, s, t [0, T], dened on the space C
b
(R
n
) by
P
s,t
(x) = (X(t, s, x)), x R
n
, s, t [0, T]. (8.4)
111
112 Kolmogorov equations
As easily checked, P
s,t
is a linear bounded operator on C
b
(R
n
). Moreover for
any C
b
(R
n
) the mapping
[0, T] [0, T] R
n
R
n
, (s, t, x) P
s,t
(x),
is continuous. From (8.2) it follows immediately the cocycle property
P
s,t
= P
s,u
P
u,t
, s, t, u [0, T]. (8.5)
Proposition 8.1 For any C
1
b
(R
n
) we have
d
dt
P
s,t
= P
s,t
L(t), t s (8.6)
and
d
ds
P
s,t
= L(s)P
s,t
, t s, (8.7)
where
L(t)(x) = b(t, x),
x
(x)), C
1
b
(R
n
), x R
n
. (8.8)
Proof. We have
d
dt
P
s,t
(x) =
d
dt
(X(t, s, x)) = b(t, X(t, s, x)),
x
(X(t, s, x)))
and
P
s,t
L(t)(x) = b(t, X(t, s, x)),
x
(X(t, s, x))),
so that (8.6) follows.
Let us prove (8.7). We have, taking into acccount (8.3),
d
ds
P
s,t
(x) =
d
ds
(X(t, s, x)) =
x
(X(t, s, x)), X
x
(t, s, x) b(s, x))
= L(s)P
s,t
(x).

Let us now consider the following partial dierential equation called trans-
port equation
_
_
_
z
s
(s, x) +b(s, x), z
x
(s, x)) = 0, s [0, T]
z(T, x) = (x),
(8.9)
where C
1
b
(R
n
) and T > 0 is xed.
Chapter 8 113
Theorem 8.2 Assume that b : [0, T] R
n
R
n
fullls Hypothesis 8.1 and
let C
1
b
(R
n
). Then problem (8.9) has a unique solution z. z is given by
z(s, x) = P
s,T
(x) = (X(T, s, x)), s [0, T], x R
n
. (8.10)
Proof Existence. It is enough to notice that z, given by (8.10), is a
solution of (8.9) by (8.6).
Uniqueness. If z is a solution of problem (8.9) we have
d
ds
z(s, X(s, u, x))
= z
t
(s, X(s, u, x)) +z
x
(s, X(s, u, x)), X
t
(s, u, x))
= z
t
(s, X(s, u, x)) +z
x
(s, X(s, u, x)), b(s, X(s, u, x))) = 0.
Therefore z(s, X(s, u, x)) is constant in s. Setting s = T and s = u we nd
that z(T, X(T, u, x)) = z(u, X(u, u, x)) which implies z(u, x) = (X(T, s, x))
as required.
8.1.1 The autonomous case
We assume here that b(t, x) = b(x) and consider the problem
_
_
_
X

(t) = b(X(t)), t 0,
X(0) = x R
n
,
(8.11)
whose solution we denote by X(, x). In this case it is easy to check that for
any t > s 0, we have P
s,t
= P
0,ts
.
Dene
P
t
(x) = (X(t, x)), C
b
(R
n
), t 0, x R
n
, (8.12)
so that by (8.5) it follows the semigroup law
P
t+s
= P
t
P
s
, t, s 0. (8.13)
P
t
is called the transition semigroup associated with (8.11). By Proposition
8.1 we deduce
Proposition 8.3 For any C
1
b
(R
n
) we have
D
t
P
t
= P
t
L = LP
t
, t 0 (8.14)
where
L(x) = b(x),
x
(x)), C
1
b
(R
n
), x R
n
. (8.15)
114 Kolmogorov equations
Finally, by Theorem 8.2 we have
Theorem 8.4 Assume that b C
1
b
(R
n
) and let C
1
b
(R). Then problem
_
_
_
u
t
(t, x) = b(x), u
x
(t, x)), t 0, x R
n
u(0, x) = (x), x R
n
.
(8.16)
has a unique solution given by
u(t, x) = P
t
(x) = (X(t, x)), t 0, x R
n
. (8.17)
8.2 Stochastic case
We consider the stochastic evolution equation
_
_
_
dX(t) = b(t, X(t))dt + (t, X(t))dB(t)
X(s) = x R
d
(8.18)
and assume that the following hypothesis holds.
Hypothesis 8.2 (i) b : [0, T] R
n
R
n
and : [0, T] R
n
L(R
r
, R
n
)
are continuous.
(ii) There exists M > 0 such that
[b(t, x)b(t, y)[+|(t, x)(t, y)|
HS
M[xy[, x, y R
n
, t [0, T].
(iii) b and have rst and second partial derivatives with respect to x con-
tinuous and bounded in [0, T] R
n
.
We denote as before by X(, s, x) the solution of (8.18) corresponding to
= x R
n
. For all t, s with 0 s t T and for all function C
b
(R
n
)
we set
P
s,t
(x) = E[(X(t, s, x))], x R
n
, 0 s t T. (8.19)
As easily checked, P
s,t
is a linear bounded operator on C
b
(R
n
).
P
s,t
, 0 < s t T, is called the transition evolution operator associated with
(8.18). By Chapter 6 we know that the mapping
(s, t, x) P
s,t
(x),
is continuous for all C
b
(R
n
).
Chapter 8 115
8.3 Basic properties of transition operators
Let us introduce the Kolmogorov operator
(L(s))(x) =
1
2
Tr [
xx
(x)(s, x)

(s, x)] +b(s, x),


x
(x)), C
2
b
(R
n
).
(8.20)
The rst basic identity is the following.
Proposition 8.5 Assume that Hypothesis 8.2 holds and let C
2
b
(R
n
).
Then P
s,t
is dierentiable in t and we have
d
dt
P
s,t
= P
s,t
L(t), t 0. (8.21)
Proof. By the It o formula we have that
d
t
(X(t, s, x)) = (L(t))(X(t, s, x)) +
x
(X(t, s, x)), (t, X(t, s, x))dB(t)).
Integrating with respect to t and taking expectation, yields
E[(X(t, s, x))] = (x) +
_
t
s
E[(L(r))(X(r, s, x))]dr,
that is
P
s,t
(x) = (x) +
_
t
s
P
r,t
(L(r))(x)dr,
which coincides with (8.21).
The second basic identity is the following,
Proposition 8.6 Assume that Hypothesis 8.2 holds and let C
2
b
(R
n
).
Then P
s,t
is dierentiable in s and we have
d
ds
P
s,t
= L(s)P
s,t
, t 0. (8.22)
Proof. Taking expectation in the backward It o formula (7.62) we nd
P
s,t
(x) (x) =
_
t
s
L(r)P
s,r
(x)dr,
which yields (8.22).
116 Kolmogorov equations
8.4 Parabolic equations
We consider here the parabolic equation
_
_
_
z
s
(s, x) + (L(s)(z(s, )))(x) = 0, 0 s < T,
z(T, x) = (x), x R
n
,
(8.23)
We say that a function z : [0, T] R
n
R is a solution to (8.23) if z is
continuous and bounded together with its partial derivatives z
t
, z
x
, z
xx
, and
fullls (8.23).
Theorem 8.7 Assume that Hypothesis 8.2 holds and let C
2
b
(R
n
). Then
there exists a unique solution z of problem (8.23). z is given by
z(s, x) = E[(X(T, s, x))], 0 < s T, C
2
b
(R
n
). (8.24)
Proof. Existence. By (8.22) it follows that
z(s, x) = P
s,T
(x), s [0, T], x R
n
,
fullls (8.23).
Uniqueness. Let z be a solution to (8.23), and let 0 u s T. Let us
compute the It o dierential of z(s, X(s, u, x)). We have
d
s
z(s, X(s, u, x)) = z
s
(s, X(s, u, x))ds + (L(s)z(s, X(s, u, )))(x)
+z
x
(s, X(s, u, x)), (s, X(s, u, x))dB(s))
= z
x
(s, X(s, u, x)), (s, X(s, u, x))dB(s)).
since z fullls (8.23). Integrating in s between u and T yields
z(T, X(T, u, x)) z(u, X(u, u, x)) = (X(t, u, x)) z(u, x)
=
_
t
u
z
x
(s, X(s, u, x))(s, X(s, u, x))dB(s).
Now, taking expectation we nd
z(u, x) = E[(X(t, u, x))].

Exercise 8.8 Prove the cocycle law


P
s,r
P
r,t
= P
s,t
(8.25)
for 0 s t t T.
Chapter 8 117
8.4.1 Autonomous case
Assume that b and are independent of t :
b(t, x) = b(x), (t, x) = (x), x R
n
.
Then we have L(s) = L where
L(x) =
1
2
Tr [
xx
(x)(x)

(x)] +b(x),
x
(x)), C
2
b
(R
n
).
Proposition 8.9 Let X(t, s, x) be the solution of the stochastic evolution
equation
_
_
_
dX(t) = b(X(t))dt + (X(t))dB(t)
X(s) = x R
n
.
(8.26)
Then for any and a > 0 the laws of X(t, s, x) and X(t +a, s +a, x) coincide.
Proof. Set Y (t) = X(t + a, s + a, x). The we have
X(t+a, s+a, x) = x+
_
t+a
s+a
b(X(r, s+a, x))dr+
_
t+a
s+a
(X(r, s+a, x))dB(r).
Setting r a = yields
Y (t) = x +
_
t
s
b(Y ())d +
_
t
s
(Y ())d[B( + a) B(a)].
Setting B
1
(t) = B(t + a) B(a) we see that Y (t) fullls equation (8.26)
but with the Brownian motion B(t) replaced by B
1
(t). Now the conclusion
follows.
By the proposition and the cocycle law (8.25)it follows that, setting
P
t
= P
0,t
, t 0,
we have
P
t+s
= P
t
P
s
, t, s 0, P
0
= 1.
Thus P
t
, t 0 is a semgroup of linear operators in C
b
(R
d
).
Setting
v(s, x) = u(t, t s, x), t 0, s [0, t], x R
n
,
problem (8.23) becomes
_
_
_
v
s
(s, x) = Lv(s, x), s [0, t], x R
n
,
v(0, x) = (x), x R
(8.27)
Then by Theorem 8.7 we nd the result
118 Kolmogorov equations
Theorem 8.10 Assume that b, : R R are Lipschitz continuous and of
class C
2
. Then, for any C
2
b
(R), problem (8.27) has a unique solution
given by
v(s, x) = P
ts,t
(x) = P
t
(x), t 0, s [0, t], x R. (8.28)
8.5 Examples
Example 8.11 Consider the parabolic equation in R
n
_
_
_
u
t
(t, x) =
1
2
Tr [Qu
xx
(t, x)] +Ax + u
x
(t, x))
u(0, x) = (x),
(8.29)
where A, Q L(R
n
), Q is symmetric and Qx, x) 0 for all x R
n
.
The corresponding stochastic dierential equation is
_
_
_
dX(t) = AX(t)dt +

Q dB(t),
X(0) = x,
(8.30)
where B is a standard Brownian motion in a probability space (, G, P)
taking values in R
n
. The solution of (8.30) is given by the variation of
constants formula
X(t, x) = e
tA
x +
_
t
0
e
(ts)A
_
QdB(s). (8.31)
Therefore the law of X(t, x) is given by
X(t, x)
#
P = N
e
tA
x,Q
t
, (8.32)
where
Q
t
=
_
t
0
e
sA
Qe
sA

ds, t 0, (8.33)
where A

is the adjoint of A.
Consequently, the transition semigroup P
t
looks like
P
t
(x) =
_
R
n
(y)N
e
tA
x,Q
t
(dy). (8.34)
So, the solution of (8.29) is given by
u(t, x) = P
t
(x).
Chapter 8 119
If, in particular, det Q
t
> 0 we have
u(t, x) = (2)
n/2
[det Q
t
]
1/2
_
R
n
e

1
2
Q
1
t
(ye
tA
x),(ye
tA
x)
(y)dy. (8.35)
Example 8.12 Consider the parabolic equation in R
_
_
_
u
t
(t, x) =
1
2
qx
2
u
xx
(t, x) + axu
x
(t, x)
u(0, x) = (x),
(8.36)
where q > 0 and a R.
The corresponding stochastic dierential equation is
_
_
_
dX(t) = aX(t)dt +

q X(t)dB(t),
X(0) = x,
(8.37)
where B is a real Brownian motion in is a real Brownian motion in some
probability space (, F, P).
The solution of (8.37) is given by
X(t, x) = e
(aq/2)t+

q B(t)
x. (8.38)
Therefore
P
t
(x) =
1

2t
_
+

y
2
2t
(e
(aq/2)t+

q y
x)dy. (8.39)
120 Kolmogorov equations
Appendix A
-systems and -systems
Let be a non empty set. A non empty family R of parts of is called a
-system if
A, B R = A B R,
a -system if
_

_
(i) , D.
(ii) A D = A
c
D.
(iii) (A
i
) D mutually disjoint =

i=1
A
i
D.
(A.1)
Obviously any algebra is a -system. Moreover, if D is a -system such that
A, B D = A B D then it is algebra. In fact if (A
i
) is a sequence
in D of not necessarily disjoint sets we have

_
i=1
A
i
= A
1
(A
2
A
1
) (A
3
A
2
A
1
) D
and so

i=1
A
i
D by (ii) and (iii).
Let us prove the following Dynkin theorem.
Theorem A.1 Let R be a -system and let D be a -system including R.
Then we have (R) D, where (R) is the algebra generated by R. If in
particular, D (R) we have (R) = D.
Proof. Let D
0
be the minimal -system including R. We are going to show
that D
0
is a algebra, which will imply the theorem. For this it is enough
to show, as remarked before, that the following inclusion holds
A, B D
0
= A B D
0
. (A.2)
121
122 -systems and -systems
For any B D
0
we set
H (B) = F D
0
: B F D
0
.
We claim that H (B) is a -system. In fact properties (i) and (iii) are clear.
It remains to show that if F B D
0
then F
c
B D
0
or, equivalently,
that F B
c
D
0
. In fact, since F B
c
= (F B
c
) B
c
= (F B) B
c
and
F B and B
c
are disjoint, we have that F B
c
D
0
as required.
If we show that
H (B) R, B D
0
(A.3)
then we conclude that H (B) = D
0
by the minimality of D
0
and (A.2) is
proved.
On the other hand it is clear that if R R we have R H (R) since R is
a -system. Therefore H (R) = D
0
by the minimality of D
0
. Consequently,
the following implication holds
R R, B D
0
R B D
0
,
which yields R H (B) and (A.3) is fullled.
Example A.2 Let A be an algebra of subsets of and let F be the -
algebra generated by A. Let P
1
and P
2
be probability measures on (, F)
such that
P
1
(I) = P
2
(I), I A.
Using the Dynkin theorem we can show that P
1
= P
2
. It is clear in fact that
A is a -system. Dene
D = B F : P
1
(B) = P
2
(B).
It is easy to see that D is a -system which contains D. So, by Corollary
A.1 it follows that P
1
= P
2
.
Appendix B
Conditional expectation
B.1 Denition
We are given a probability space (, F, P) and a -algebra G included in F.
Let X : R be a real random variable on (, F, P)
(1)
.
We say that X is G-measurable if
I B(R) X
1
(I) F.
It is clear that X is not G-measurable in general.
Let us consider the signed measure
(G) =
_
G
XdP, G G.
It is clear that is absolutely continuous with respect to the restriction of
P to G. Therefore, by the Radon-Nikodym Theorem there exists a unique
Y L
1
(, G, P) such that
(G) =
_
G
XdP =
_
G
Y dP, G G. (B.1)
The G-measurable random variable Y is called the conditional expectation of
X given G; it is denoted by E(X[G).
In view of (B.1) E(X[G) is characterized by
_
G
XdP =
_
G
E(X[G)dP, G G. (B.2)
Exercise B.1 Assume that X L
2
(, F, P). Show that E(X[G) coincides
with the orthogonal projection of X into the closed subspace L
2
(, G, P) of
L
2
(, F, P).
(1)
In all this appendix by random variable we mean an equivalence class of random
variables with respect to the usual equivalence relation.
123
124 Conditional expectation
B.2 Basic properties
Let X, Y L
1
(, F, P) and let G be -algebra included in F. It is obvious
that if X is G-measurable, we have E(X[G) = X. Setting G = in (B.2)
yields
E[E(X[G)] = E(X). (B.3)
Moreover, one can check easily the linearity of conditional expectation,
E(X + Y [G) = E(X[G) + E(Y [G), (B.4)
for all , R and all X, Y L
1
(, F, P). Also if X 0, P-a.s., one has
E(X[G) 0, P-a.s. From this one deduces the inequality
[E(X[G)[ E([X[ [G). (B.5)
Proposition B.2 Assume that X is independent of G. Then we have
E(X[G) = E(X). (B.6)
Proof. Let A G. Then 1l
A
and X are independent so that
_
A
XdP =
_

1l
A
XdP = P(A)E(X) =
_
A
E(X[G)dP.
Proposition B.3 Let H be a -algebra included in G. Then we have
E(X[H ) = E
_
E(X[G)

. (B.7)
Proof. Let A H . Then we have
_
A
XdP =
_
A
E(X[H )dP (B.8)
and
_
A
XdP =
_
A
E(X[G)dP =
_
A
E
_
E(X[G)

dP. (B.9)
So, comparing (B.8) and (B.9) we see that
_
A
E(X[H )dP =
_
A
XdP =
_
A
E
_
E(X[G)

dP.

Proposition B.4 Let X, Y, XY L


1
(, F, P). Assume that X is G-measurable.
Then we have
E(XY [G) = XE(Y [G). (B.10)
Appendix B 125
Proof. It is enough to show (B.10) for X = 1l
A
where A G. Let now
G G, then since G A G we have
_
G
E(1l
A
Y [G)dP =
_
G
1l
A
Y dP =
_
GA
Y dP
=
_
GA
E(Y [G)dP =
_
G
1l
A
E(Y [G)dP,
for any G G.
Recalling Proposition B.2 we nd.
Corollary B.5 Let X, Y, XY L
1
(, F, P). Assume that X is G-measurable
and that Y is independent of G. Then we have
E(XY [G) = XE(Y ). (B.11)
Let us prove now a useful generalization of this Corollary.
Proposition B.6 Let X, Y L
1
(, F, P) and let : R
2
R be bounded
and Borel. Assume that X is G-measurable and Y is independent of G. Then
we have
E((X, Y )[G) = h(X), (B.12)
where
h(x) = E[(x, Y )], x R. (B.13)
Proof. We have to show that
_
G
(X, Y )dP =
_
G
h(X)dP, G G.
This is clearly equivalent to
E(Z(X, Y )) = E(Zh(X)), Z L
1
(, G, P). (B.14)
Denote by the law of the random variable (X, Y, Z) with values in R
3
= (X, Y, Z)
#
P.
So,
E(Z(X, Y )) =
_
R
3
z(x, y)(dx, dy, dz). (B.15)
126 Conditional expectation
Since X and Z are G-measurable and Y is independent of G, the random
variables (X, Z) and Y are independent so that
(dx, dy, dz) = (dx, dz)(dy),
where
(dx, dz) = (X, Z)
#
P(dx, dz), (dy) = Y
#
P(dy).
Therefore we can write (B.15) as
E(Z(X, Y )) =
_
R
3
z(x, y)(dx, dz)(dy).
Using the Fubini Theorem we get nally
E(Z(X, Y )) =
_
R
2
z
__
R
(x, y)(dy)
_
(dx, dz)
=
_
R
2
zh(x)(dx, dz) = E(Zh(X)),
as required.
Exercise B.7 Let F, H, FH L
1
(, G, P) and Z = E(H[G). Prove that
E(FH) = E(FZ). (B.16)
Exercise B.8 Let g : R R be convex and let F, g(F) L
1
(, F, P).
Prove the Jensen inequality
E(g(F)[G) g(E(F[G)). (B.17)
Appendix C
Martingales
C.1 Denitions
Let (, F, P) be a probability space, (F
t
)
t0
an increasing family of -
algebras included in F and (M(t))
t[0,T]
with M(t) L
1
(, F
t
, P), t [0, T],
a stochastic process.
(M(t))
t[0,T]
is said to be a martingale (with respect to the ltration
(F
t
)
t0
) if
E[M(t)[F
s
] = M(s), 0 s < t T,
a submartingale if
E[M(t)[F
s
] M(s), 0 s < t T,
a supermartingale if
E[M(t)[F
s
] M(s), 0 s < t T.
Thus (M(t))
t[0,T]
is a martingale if and only if
_
A
M(s)dP =
_
A
M(t)dP, 0 s < t T, A F
s
,
a submartingale if and only if
_
A
M(s)dP
_
A
M(t)dP, 0 s < t T, A F
s
,
and a supermartingale if and only if
_
A
M(s)dP
_
A
M(t)dP, 0 s < t T, A F
s
.
127
128 Martingales
Proposition C.1 If M is a martingale then [M[ is a submartingale.
Proof. Let 0 s < t T, A F
s
. Set
A
+
= : M(s)() > 0, A

= : M(s)() 0.
Clearly A
+
and A

belong to F
s
. Consequently we have
_
A
[M(s)[dP =
_
A
+
M(s)dP
_
A

M(s)dP
=
_
A
+
M(t)dP
_
A

M(t)dP
_
A
[M(t)[dP.
This shows that [M[ is a submartingale.
Example C.2 The Brownian motion B is a martingale. In fact, let t > s
and A F
s
. Since B(t) B(s) and 1l
A
are independent we have
_
A
(B(t) B(s))dP = E(1l
A
(B(t) B(s))) = 0,
so that
_
A
B(t)dP =
_
A
B(s)dP.
Exercise C.3 Using Jensens inequality prove that any convex function of
a martingale is a submartingale. (See Exercise B.8).
C.2 The basic inequality for martingales
Let M(t) be a martingale, let 0 < t
1
< t
2
< ... < t
n
T and set
S = sup
1in
[M(t
i
)[.
We are going to prove an important estimate (due to Kolmogorov) of S in
terms of M(t
n
).
Proposition C.4 For all > 0 we have
P(S )
1

_
{S}
[M(t
n
)[dP. (C.1)
Appendix C 129
Proof. Set
A
1
= [M(t
1
)[ ,
A
2
= [M(t
1
)[ < , [M(t
2
)[ ,

A
n
= [M(t
1
)[ < , ..., [M(t
n
)[ .
Clearly, sets A
1
, ..., A
n
are mutually disjoint. Moreover A
i
F
t
i
, i = 1, ..., n,
and we have
S =
n
_
i=1
A
i
.
Let us estimate
_
{S}
[M(t
n
)[dP. We have obviously
_
A
n
[M(t
n
)[dP P(A
n
).
Now we estimate
_
A
n1
[X(t
n
)[dP. We have, recalling that [M(t)[ is a sub
martingale,
P(A
n1
)
_
A
n1
[M(t
n1
)[dP
_
A
n1
[M(t
n
)[dP.
Therefore
_
A
n1
[M(t
n
)[dP P(A
n1
).
Proceeding in a similar way we obtain
_
A
k
[M(t
n
)[dP P(A
k
), k = 1, . . . , n. (C.2)
Summing up on k from 1 to n the conclusion follows.
C.3 Square integrable martingales
In this section we are given a martingale M(t) such that M(t) L
2
(, F, P)
for all t [0, T].
Let 0 < t
1
< t
2
< ... < t
n
T and set as before
S = sup
1in
[M(t
i
)[.
We are going to estimate of E[S
2
] in terms of E[M
2
(t
n
)].
130 Martingales
Proposition C.5 We have
E
_
sup
1in
[M(t
i
)[
2
_
4E([M(t
n
)[
2
). (C.3)
Proof. Set
F(t) = P(S > t), t 0.
By (C.1) we have
F(t)
1
t
_
{St}
[M(t
n
)[dP. (C.4)
Consequently
E(S
2
) =
_

0
P(S
2
> t)dt =
_

0
P(S >

t)dt.
So, by (C.1) and the Fubini Theorem we have
E(S
2
)
_

0
_
1

t
_
{S

t}
[M(t
n
)[dP
_
dt
=
_
[0,+)
1

t
[M(t
n
)[1l
{S

t}
P(d)dt
=
_

[M(t
n
)[P(d)
_

0
1

t
1l
{S

t}
dt
=
_

[M(t
n
)[P(d)
_
S
2
0
1

t
dt
= 2
_

[M(t
n
)SP(d) 2
__

[M(t
n
)[
2
dP
_
1/2
__

S
2
dP
_
1/2
.
Now the conclusion follows easily.
Corollary C.6 Let M be a square integrable continuous martingale. Then
for any T > 0 we have
E
_
sup
t[0,T]
[M(t)[
2
_
4E[M
2
(T)]. (C.5)
Appendix C 131
Proof. Let 0 < s
1
< s
2
< < s
m
= T. By Proposition C.5 it follows that
E
_
sup
1im
[M(s
i
)[
2
_
4E
_
[M(T)[
2

.
Since M is continuous it follows, by the arbitrariness of the sequence s
1
, s
2
, . . . , s
m
,
that
E
_
sup
s[0,T]
[M(s)[
2
_
4E
_
[M(T)[
2

,
as required.
132 Martingales
Appendix D
Fixed points depending on
parameters
D.1 Introduction
Let , E be Banach spaces (norms [ [). We are given a continuous mapping
F : E E, (, x) F(, x)
and assume that
Hypothesis D.1 There exists [0, 1) such that
[F(, x) F(, y)[ [x y[, , x, y E.
The following result (contraction principle) is classical.
Theorem D.1 (i). There exists a unique continuous mapping
x : E, x(),
such that
x() = F(, x()), . (D.1)
(ii). If in addition F is of class C
1
, then x is of class C
1
and
x

() = F

(, x()) + F
x
(, x())x

(). (D.2)
We want to generalize the second part of this result to mappings F(, x)
which are only continuously Gateaux dierentiable.
133
134 Fixed points
D.2 Gateaux dierentiable mappings
Let A and B be Banach spaces and let : A B be a continuous mapping
from A into B.
Denition D.2 We say that is G ateaux dierentiable if there exists a
mapping
D : A L(A, B), a D(a),
such that
lim
0
1

((a + c) (a)) = D(a)c, a, c A.


If in addition for all c A the mapping A B, a D(a)c is continuous
we say that is continuously G ateaux dierentiable.
Remark D.3 It is well known that if the mapping A L(A, B), a
D(a) is continuous then is dierentiable.
(1)
Example D.4 Let A, B = L
2
(0, 1) and (x) = sin x. Then one can check
easily that is continuously G ateaux dierentiable and
D(x)y = y cos x, x, y L
2
(0, 1).
However, (as one can see) is not dierentiable in any point.
We shall need the following result.
Proposition D.5 Let : A B be continuously Gateaux dierentiable.
Then the following identity holds
(c) (a) =
_
1
0
D((1 )a + c)(c a)d. (D.3)
Proof. Set
F() = ((1 )a + c), [0, 1].
Then we have
F

() = D((1 )a + c)(c a)d,


and the conclusion follows just integrating this identity between 0 and 1.
(1)
One also says that is Frechet dierentiable.
Appendix D 135
D.3 The main result
We can back to the notations of the introduction and consider two Banach
spaces and E and a continuous mapping
F : E E, (, x) F(, x).
We assume that Hypothesis D.1 is fullled and denote by x the mapping
x : E, x(),
such that
x() = F(, x()), . (D.4)
Theorem D.6 Assume that Hypotheses D.1 is fullled and that F is con-
tinuously Gateaux dierentiable. Then x() is continuously Gateaux dieren-
tiable as well and we have
x

() = (1 F
x
(, x()))
1
F

(, x()) , (D.5)
equivalently
x

() = F

(, x()) + F
x
(, x())(x

() ). (D.6)
Proof. Let , and h R. From (D.4) and (D.3) it follows that
x( + h) x() = F( + h, x( + h)) F(, x())
= h
_
1
0
F

( + h, x() + (x( + h) x())) d


+
_
1
0
F
x
( + h, x() + (x( + h) x())) (x( + h) x())d.
(D.7)
Set now
G(, x, , h)z = Gz :=
_
1
0
F
x
(+h, x()+(x(+h)x()))zd, z E.
Then G L(E) and by Hypothesis D.1
[Gz[ [z[, z E.
136 Fixed points
Then from equation (D.7) we have
(1 G(, x, , h))(x( + h) x())
= h
_
1
0
F

( + h, x() + (x( + h) x())) d,


which implies
1
h
x( + h) x()) = (1 G(, x, , h))
1

_
1
0
F

( + h, x() + (x( + h) x())) d.


Letting h 0 we nd
x

() = (1 F
x
(, x()))
1
F

(, x()).
Therefore
x

() F
x
(, x())(x

() ) = F

(, x()).

Appendix E
Fractional Sobolev spaces and
regularity of processes
E.1 Fractional Sobolev spaces on [0, 1]
Let (0, 1), m N. Dene
|f|
2m
,2m
:=
_
[0,T]
2
[f(t) f(s)[
2m
[t s[
1+2m
dt ds
W
,2m
(0, T) is by denition the space of all f : [0, T] R such that |f|
,2m
<
+.
Theorem E.1 (Sobolev embedding) Assume that > 1/(2m). Then the
following inclusion holds with continuous embedding.
W
,2m
(0, T) C
1/(2m)
([0, T]). (E.1)
Example E.2 (The Brownian motion) Let > 0 and let p 1. We ask
the question whether B() belongs to W
,p
(0, T) or not.
Let us compute
E(|B|
p
W
,p) = E
_
[0,T]
2
[B(t) B(s)[
p
[t s[
1+p
dt ds
Take for simplicity p = 2m, then
E
_
|B|
2m
W
,2m
_
= E
_
[0,T]
2
[B(t) B(s)[
2m
[t s[
1+2m
dt ds
= c
m
_
[0,T]
2
[t s[
m
[t s[
1+2m
dt ds = c
m
_
[0,T]
2
[t s[
m12m
dt ds
137
138 Fractional Sobolev spaces
The integral is nite if and only if <
1
2
.
For instance taking m = 1 we conclude that B() W
,2
(0, T) for <
1
2
.
This does not imply that B() is continuous.
But if we take m = 2 we have B() W
,4
(0, T) again for <
1
2
. Therefore
if
1
4
< <
1
2
we conclude by the Sobolev embedding that B() C

1
4
(0, T).
Arguing similarly taking larger m we conclude that B() C

(0, T) for
any (0, 1/2).
E.2 Processes belonging to W
,2m
(0, T)
Let (, F, P) be probability space and let X(t), t [0, T], be a real stochastic
process on (, F, P). One situation often encountered is when the following
estimate holds for some m > 1, (0, 1/2), and c
m
> 0
E[[X(t) X(s)[
2m
] c
m
[t s[
m
, t, s [0, T]. (E.2)
This estimate (provided m > 1) allows us to conclude that trajectories of X
are Holder continuous almost surely, as the next proposition shows.
Proposition E.3 Assume that there is m > 1, (0, 1/2), and c
m
> 0
such that (E.2) is fullled. Then we have
E
_
[X[
2m
,2m

< +. (E.3)
Moreover, X(, ) belongs to C
1/(2m)
([0, T]) for almost .
Proof. We have in fact
E
_
|X|
2m
,2m
_
c
m
_
[0,T]
2
[t s[
m12m
dt ds < ,
since (0, 1/2) and m 1 2m > 1. The last statement follows from
the Sobolev embedding theorem.
Remark E.4 Kolomogorov test It is a generalization Proposition E.3. As-
sume that there is a > 0, b > 0 such that
E[[X(t) X(s)[
1+a
] c
m
[t s[
1+b
t, s [0, T]. (E.4)
Then X has -H older continuous trajectories with <
1+b
a
.
Appendix F 139
E.3 Multi dimensional Sobolev spaces and reg-
ularity of random elds
Let (0, 1), m N, d N. Dene
|f|
2m
,2m
:=
_
[0,T]
2d
[f(x) f(y)[
2m
[x y[
d+2m
dx dy.
W
,2m
([0, T]
d
) is by denition the space of all f : [0, T]
d
R such that
|f|
,2m
< +.
Theorem E.5 (Sobolev embedding) Assume that > d/(2m). Then the
following inclusion holds with continuous embedding.
W
,2m
([0, T]
d
) C
d/(2m)
([0, T]
d
). (E.5)
Let (, F, P) be probability space and let X(x), x [0, T]
d
, be a random
eld on (, F, P).
Assume that there is m > 1, (0, 1), and c
m
> 0
E[[X(x) X(y)[
2m
] c
m
[t s[
2m
, t, s [0, T]. (E.6)
This estimate implies that almost all trajectories of X are H older continuous
almost surely.
Proposition E.6 Assume that there is m > 1, (0, 1), and c
m
> 0 such
that (E.2) is fullled. Then we have
E
_
[X[
2m
,2m

< +. (E.7)
Moreover, X(, ) belongs to C
d/(2m)
([0, T]) for almost .
Proof. We have in fact
E(|X|
2m
,2m
) c
m
_
[0,T]
2
[t s[
m12m
dt ds < ,
since (0, 1/2) and m 1 2m > 1. The last statement follows from
the Sobolev embedding theorem.

You might also like