AN INTRODUCTION TO

MALLIAVIN CALCULUS
WITH APPLICATIONS TO ECONOMICS
Bernt ksendal
Dept. of Mathematics, University of Oslo,
Box 1053 Blindern, N0316 Oslo, Norway
Institute of Finance and Management Science,
Norwegian School of Economics and Business Administration,
Helleveien 30, N5035 Bergen-Sandviken, Norway.
Email: oksendal@math.uio.no
May 1997

Preface
These are unpolished lecture notes from the course BF 05 Malliavin calculus with appli-
cations to economics, which I gave at the Norwegian School of Economics and Business
Administration (NHH), Bergen, in the Spring semester 1996. The application I had in
mind was mainly the use of the Clark-Ocone formula and its generalization to nance,
especially portfolio analysis, option pricing and hedging. This and other applications are
described in the impressive paper by Karatzas and Ocone [KO] (see reference list in the
end of Chapter 5). To be able to understand these applications, we had to work through
the theory and methods of the underlying mathematical machinery, usually called the
Malliavin calculus. The main literature we used for this part of the course are the books
by Ustunel [U] and Nualart [N] regarding the analysis on the Wiener space, and the
forthcoming book by Holden, ksendal, Ube and Zhang [HUZ] regarding the related
white noise analysis (Chapter 3). The prerequisites for the course are some basic knowl-
edge of stochastic analysis, including Ito integrals, the Ito representation theorem and the
Girsanov theorem, which can be found in e.g. [1].
The course was followed by an inspiring group of (about a dozen) students and employees
at HNN. I am indebted to them all for their active participation and useful comments. In
particular, I would like to thank Knut Aase for his help in getting the course started and
his constant encouragement. I am also grateful to Kerry Back, Darrell Due, Yaozhong
Hu, Monique Jeanblanc-Picque and Dan Ocone for their useful comments and to Dina
Haraldsson for her procient typing.
Oslo, May 1997
Bernt ksendal
i

Contents
1 The Wiener-Ito chaos expansion . . . . . . . . . . . . . . . . . . . . . . . . 1.1
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.8
2 The Skorohod integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1
The Skorohod integral is an extension of the Ito integral . . . . . . . . . . 2.4
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.6
3 White noise, the Wick product and stochastic integration . . . . . . . . . . 3.1
The Wiener-Ito chaos expansion revisited . . . . . . . . . . . . . . . . . . . 3.3
Singular (pointwise) white noise . . . . . . . . . . . . . . . . . . . . . . . . 3.6
The Wick product in terms of iterated Ito integrals . . . . . . . . . . . . . 3.9
Some properties of the Wick product . . . . . . . . . . . . . . . . . . . . . 3.9
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.12
4 Dierentiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1
Closability of the derivative operator . . . . . . . . . . . . . . . . . . . . . 4.7
Integration by parts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.8
Dierentiation in terms of the chaos expansion . . . . . . . . . . . . . . . . 4.11
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.13
5 The Clark-Ocone formula and its generalization. Application to nance . . 5.1
The Clark-Ocone formula . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5
The generalized Clark-Ocone formula . . . . . . . . . . . . . . . . . . . . . 5.5
Application to nance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.10
The Black-Scholes option pricing formula and generalizations . . . . . . . . 5.13
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.15
ii

6 Solutions to the exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1
iii

1 The Wiener-Ito chaos expansion
The celebrated Wiener-Ito chaos expansion is fundamental in stochastic analysis. In
particular, it plays a crucial role in the Malliavin calculus. We therefore give a detailed
proof.
The rst version of this theorem was proved by Wiener in 1938. Later Ito (1951) showed
that in the Wiener space setting the expansion could be expressed in terms of iterated Ito
integrals (see below).
Before we state the theorem we introduce some useful notation and give some auxiliary
results.
Let W(t) = W(t, ); t 0, be a 1-dimensional Wiener process (Brownian motion)
on the probability space (, T, P) such that W(0, ) = 0 a.s. P.
For t 0 let T
t
be the -algebra generated by W(s, ); 0 s t. Fix T > 0 (constant).
A real function g : [0, T]
n
R is called symmetric if
(1.1) g(x

1
, . . . , x

n
) = g(x
1
, . . . , x
n
)
for all permutations of (1, 2, . . . , n). If in addition
(1.2) |g|
2
L
2
([0,T]
n
)
: =
_
[0,T]
n
g
2
(x
1
, . . . , x
n
)dx
1
dx
n
<
we say that g

L
2
([0, T]
n
), the space of symmetric square integrable functions on [0, T]
n
.
Let
(1.3) S
n
= (x
1
, . . . , x
n
) [0, T]
n
; 0 x
1
x
2
x
n
T.
The set S
n
occupies the fraction
1
n!
of the whole n-dimensional box [0, T]
n
. Therefore, if
g

L
2
([0, T]
n
) then
(1.4) |g|
2
L
2
([0,T]
n
)
= n!
_
S
n
g
2
(x
1
, . . . , x
n
)dx
1
. . . dx
n
= n!|g|
2
L
2
(S
n
)
If f is any real function dened on [0, T]
n
, then the symmetrization

f of f is dened by
(1.5)

f(x
1
, . . . , x
n
) =
1
n!

f(x

1
, . . . , x

n
)
where the sum is taken over all permutations of (1, . . . , n). Note that

f = f if and only
if f is symmetric. For example if
f(x
1
, x
2
) = x
2
1
+x
2
sin x
1
then

f(x
1
, x
2
) =
1
2
[x
2
1
+x
2
2
+x
2
sin x
1
+x
1
sin x
2
].
1.1

Note that if f is a deterministic function dened on S
n
(n 1) such that
|f|
2
L
2
(S
n
)
: =
_
S
n
f
2
(t
1
, . . . , t
n
)dt
1
dt
n
< ,
then we can form the (n-fold) iterated Ito integral
(1.6) J
n
(f): =
T
_
0
t
n
_
0

t
3
_
0
(
t
2
_
0
f(t
1
, . . . , t
n
)dW(t
1
))dW(t
2
) dW(t
n1
)dW(t
n
),
because at each Ito integration with respect to dW(t
i
) the integrand is T
t
-adapted and
square integrable with respect to dP dt
i
, 1 i n.
Moreover, applying the Ito isometry iteratively we get
E[J
2
n
(h)] = E[
T
_
0
(
t
n
_
0

t
2
_
0
h(t
1
, . . . , t
n
)dW(t
1
) )dW(t
n
)
2
]
=
T
_
0
E[(
t
n
_
0

t
2
_
0
h(t
1
, . . . , t
n
)dW(t
1
) dW(t
n1
))
2
]dt
n
= =
T
_
0
t
n
_
0

t
2
_
0
h
2
(t
1
, . . . , t
n
)dt
1
dt
n
= |h|
2
L
2
(S
n
)
. (1.7)
Similarly, if g L
2
(S
m
) and h L
2
(S
n
) with m < n, then by the Ito isometry applied
iteratively we see that
E[J
m
(g)J
n
(h)]
= E[
T
_
0
(
s
m
_
0

s
2
_
0
g(s
1
, . . . , s
m
)dW(s
1
) dW(s
m
)

T
_
0
(
s
m
_
0

t
2
_
0
h(t
1
, . . . , t
nm
, s
1
, . . . , s
m
)dW(t
1
) )dW(s
m
)]
=
T
_
0
E[
s
m
_
0

s
2
_
0
g(s
1
, . . . , s
m1
, s
m
)dW(s
1
) dW(s
m1
)

s
m
_
0

t
2
_
0
h(t
1
, . . . , s
m1
, s
m
)dW(t
1
) dW(s
m1
)]ds
m
=
T
_
0
s
m
_
0

s
2
_
0
E[g(s
1
, s
2
, . . . , s
m
)
s
1
_
0

t
2
_
0
h(t
1
, . . . , t
nm
, s
1
, . . . , s
m
)
dW(t
1
) dW(t
nm
)]ds
1
, ds
m
= 0 (1.8)
because the expected value of an Ito integral is zero.
1.2

We summarize these results as follows:
(1.9) E[J
m
(g)J
n
(h)] =
_
0 if n ,= m
(g, h)
L
2
(S
n
)
if n = m
where
(1.10) (g, h)
L
2
(S
n
)
=
_
S
n
g(x
1
, . . . , x
n
)h(x
1
, . . . , x
n
)dx
1
dx
n
is the inner product of L
2
(S
n
).
Note that (1.9) also holds for n = 0 or m = 0 if we dene
J
0
(g) = g if g is a constant
and
(g, h)
L
2
(S
0
)
= gh if g, h are constants.
If g

L
2
([0, T]
n
) we dene
(1.11) I
n
(g): =
_
[0,T]
n
g(t
1
, . . . , t
n
)dW
n
(t): = n!J
n
(g)
Note that from (1.7) and (1.11) we have
(1.12) E[I
2
n
(g)] = E[(n!)
2
J
2
n
(g)] = (n!)
2
|g|
2
L
2
(S
n
)
= n!|g|
2
L
2
([0,T]
n
)
for all g

L
2
([0, T]
n
).
Recall that the Hermite polynomials h
n
(x); n = 0, 1, 2, . . . are dened by
(1.13) h
n
(x) = (1)
n
e
1
2
x
2 d
n
dx
n
(e

1
2
x
2
); n = 0, 1, 2, . . .
Thus the rst Hermite polynomials are
h
0
(x) = 1, h
1
(x) = x, h
2
(x) = x
2
1, h
3
(x) = x
3
3x,
h
4
(x) = x
4
6x
2
+ 3, h
5
(x) = x
5
10x
3
+ 15x, . . .
There is a useful formula due to Ito [I] for the iterated Ito integral in the special case
when the integrand is the tensor power of a function g L
2
([0, T]):
(1.14) n!
T
_
0
t
n
_
0

t
2
_
0
g(t
1
)g(t
2
) g(t
n
)dW(t
1
) dW(t
n
) = |g|
n
h
n
(

|g|
),
where
|g| = |g|
L
2
([0,T])
and =
T
_
0
g(t)dW(t).
1.3

For example, choosing g 1 and n = 3 we get
6
T
_
0
t
3
_
0
t
2
_
0
dW(t
1
)dW(t
2
)dW(t
3
) = T
3/2
h
3
(
W(T)
T
1/2
) = W
3
(T) 3T W(T).
THEOREM 1.1. (The Wiener-Ito chaos expansion) Let be an T
T
-measurable
random variable such that
||
2
L
2
()
: = ||
2
L
2
(P)
: = E
P
[
2
] < .
Then there exists a (unique) sequence f
n

n=0
of (deterministic) functions f
n


L
2
([0, T]
n
)
such that
(1.15) () =

n=0
I
n
(f
n
) (convergence in L
2
(P)).
Moreover, we have the isometry
(1.16) ||
2
L
2
(P)
=

n=0
n!|f
n
|
2
L
2
([0,T]
n
)
Proof. By the Ito representation theorem there exists an T
t
-adapted process
1
(s
1
, ),
0 s
1
T such that
(1.17) E[
T
_
0

2
1
(s
1
, )ds
1
] ||
2
L
2
(P)
and
(1.18) () = E[] +
T
_
0

1
(s
1
, )dW(s
1
)
Dene
(1.19) g
0
= E[] (constant).
For a.a. s
1
T we apply the Ito representation theorem to
1
(s
1
, ) to conclude that
there exists an T
t
-adapted process
2
(s
2
, s
1
, ); 0 s
2
s
1
such that
(1.20) E[
s
1
_
0

2
2
(s
2
, s
1
, )ds
2
] E[
2
1
(s
1
)] <
and
(1.21)
1
(s
1
, ) = E[
1
(s
1
)] +
s
1
_
0

2
(s
2
, s
1
, )dW(s
2
).
1.4

Substituting (1.21) in (1.18) we get
(1.22) () = g
0
+
T
_
0
g
1
(s
1
)dW(s
1
) +
T
_
0
(
s
1
_
0

2
(s
2
, s
1
, )dW(s
2
)dW(s
1
)
where
(1.23) g
1
(s
1
) = E[
1
(s
1
)].
Note that by the Ito isometry, (1.17) and (1.20) we have
(1.24) E[
T
_
0
(
s
1
_
0

2
(s
1
, s
2
, )dW(s
2
))dW(s
1
)
2
] =
T
_
0
(
s
1
_
0
E[
2
2
(s
1
, s
2
, )]ds
2
)ds
1
||
2
L
2
(P)
.
Similarly, for a.a. s
2
s
1
T we apply the Ito representation theorem to
2
(s
2
, s
1
, ) to
get an T
t
-adapted process
3
(s
3
, s
2
, s
1
, ); 0 s
3
s
2
such that
(1.25) E[
s
2
_
0

2
3
(s
3
, s
2
, s
1
, )ds
3
] E[
2
2
(s
2
, s
1
)] <
and
(1.26)
2
(s
2
, s
1
, ) = E[
2
(s
2
, s
1
, )] +
s
2
_
0

3
(s
3
, s
2
, s
1
, )dW(s
3
).
Substituting (1.26) in (1.22) we get
() = g
0
+
T
_
0
g
1
(s
1
)dW(s
1
) +
T
_
0
(
s
1
_
0
g
2
(s
2
, s
1
)dW(s
2
))dW(s
1
)
+
T
_
0
(
s
1
_
0
(
s
2
_
0

3
(s
3
, s
2
, s
1
, )dW(s
3
))dW(s
2
))dW(s
1
), (1.27)
where
(1.28) g
2
(s
2
, s
1
) = E[
2
(s
2
, s
1
)]; 0 s
2
s
1
T.
By the Ito isometry, (1.17), (1.20) and (1.25) we have
(1.29) E[
T
_
0
s
1
_
0
s
2
_
0

3
(s
3
, s
2
, s
1
, )dW(s
3
)dW(s
2
)dW(s
3
)
2
] ||
2
L
2
(P)
.
By iterating this procedure we obtain by induction after n steps a process
n+1
(t
1
, t
2
, . . .,
t
n+1
, ); 0 t
1
t
2
t
n+1
T and n + 1 deterministic functions g
0
, g
1
, . . . , g
n
with g
0
constant and g
k
dened on S
k
for 1 k n, such that
(1.30) () =
n

k=0
J
k
(g
k
) +
_
S
n+1

n+1
dW
(n+1)
,
1.5

where
(1.31)
_
S
n+1

n+1
dW
(n+1)
=
T
_
0
t
n+1
_
0

t
2
_
0

n+1
(t
1
, . . . , t
n+1
, )dW(t
1
) dW(t
n+1
)
is the (n + 1)-fold iterated integral of
n+1
. Moreover,
(1.32) E[
_
S
n+1

n+1
dW
(n+1)

2
] ||
2
L
2
()
.
In particular, the family

n+1
: =
_
S
n+1

n+1
dW
(n+1)
; n = 1, 2, . . .
is bounded in L
2
(P). Moreover
(1.33) (
n+1
, J
k
(f
k
))
L
2
()
= 0 for k n, f
k
L
2
([0, T]
k
).
Hence by the Pythagorean theorem
(1.34) ||
2
L
2
()
=
n

k=0
|J
k
(g
k
)|
2
L
2
()
+|
n+1
|
2
L
2
()
In particular,
n

k=0
|J
k
(g
k
)|
2
L
2
()
<
and therefore

k=0
J
k
(g
k
) is strongly convergent in L
2
(). Hence
lim
n

n+1
=: exists (limit in L
2
())
But by (1.33) we have
(1.35) (J
k
(f
k
), )
L
2
()
= 0 for all k and all f
k
L
2
([0, T]
k
)
In particular, by (1.14) this implies that
E[h
k
(

|g|
) ] = 0 for all g L
2
([0, T]), all k 0
where =
T
_
0
g(t)dW(t).
But then, from the denition of the Hermite polynomials,
E[
k
] = 0 for all k 0
1.6

which again implies that
E[exp ] =

k=0
1
k!
E[
k
] = 0.
Since the family
exp ; g L
2
([0, T])
is dense in L
2
() (see [1], Lemma 4.9), we conclude that
= 0.
Hence
(1.36) () =

k=0
J
k
(g
k
) (convergence in L
2
())
and
(1.37) ||
2
L
2
()
=
n

k=0
|J
k
(g
k
)|
2
L
2
()
.
Finally, to obtain (1.15)(1.16) we proceed as follows:
The function g
n
is only dened on S
n
, but we can extend g
n
to [0, T]
n
by putting
(1.38) g
n
(t
1
, . . . , t
n
) = 0 if (t
1
, . . . , t
n
) [0, T]
n
S
n
.
Now dene
f
n
= g
n
, the symmetrization of g.
Then
I
n
(f
n
) = n!J
n
(f
n
) = n!J
n
( g
n
) = J
n
(g
n
)
and (1.15)(1.16) follow from (1.36) and (1.37). .
Examples
1) What is the Wiener-Ito expansion of
() = W
2
(T, ) ?
From (1.14) we get
2
T
_
0
(
t
2
_
0
dW(t
1
))dW(t
2
) = Th
2
(
W(T)
T
1/2
) = W
2
(T) T,
and therefore
W
2
(T) = T +I
2
(1).
1.7

2) Note that for t (0, T) we have
T
_
0
(
t
2
_
0
A
t
1
<t<t
2

(t
1
, t
2
)dW(t
1
))dW(t
2
)
=
T
_
t
W(t)dW(t
2
) = W(t)(W(T) W(t)).
Hence, if we put
() = W(t)(W(T) W(t)), g(t
1
, t
2
) = A
t
1
<t<t
2

we see that
() = J
2
(g) = 2J
2
( g) = I
2
(f
2
),
where
f
2
(t
1
, t
2
) = g(t
1
, t
2
) =
1
2
(A
t
1
<t<t
2

+A
t
2
<t<t
1

).
Exercises
1.1 a) Let h
n
(x); n = 0, 1, 2, . . . be the Hermite polynomials, dened in (1.13). Prove
that
exp(tx
t
2
2
) =

n=0
t
n
n!
h
n
(x) for all t, x.
(Hint: Write
exp(tx
t
2
2
) = exp(
1
2
x
2
) exp(
1
2
(x t)
2
)
and apply the Taylor formula on the last factor.)
b) Show that if > 0 then
exp(tx
t
2

2
) =

n=0
t
n

n
2
n!
h
n
(
x

).
c) Let g L
2
([0, T]) be deterministic. Put
= () =
T
_
0
g(s)dW(s)
and
|g| = |g|
L
2
([0,T])
= (
T
_
0
g
2
(s)ds)
1/2
.
Show that
exp(
T
_
0
g(s)dW(s)
1
2
|g|
2
) =

n=0
|g|
n
n!
h
n
(

|g|
)
1.8

d) Let t [0, T]. Show that
exp(W(t)
1
2
t) =

n=0
t
n/2
n!
h
n
(
W(t)

t
).
1.2 Find the Wiener-Ito chaos expansion of the following random variables:
a) () = W(t, ) (t [0, T] xed)
b) () =
T
_
0
g(s)dW(s) (g L
2
([0, T]) deterministic)
c) () = W
2
(t, ) (t [0, T] xed)
d) () = exp(
T
_
0
g(s)dW(s)) (g L
2
([0, T]) deterministic)
(Hint: Use (1.14).)
1.3 The Ito representation theorem states that if F L
2
() is T
T
-measurable, then
there exists a unique T
t
-adapted process (t, ) such that
(1.40) F() = E[F] +
T
_
0
(t, )dW(t).
(See e.g. [1], Theorem 4.10.)
As we will show in Chapter 5, this result is important in mathematical nance. Moreover,
it is important to be able to nd more explicitly the integrand (t, ). This is achieved
by the Clark-Ocone formula, which says that (under some extra conditions)
(1.41) (t, ) = E[D
t
F[T
t
](),
where D
t
F is the (Malliavin) derivative of F. We will return to this in Chapters 4 and 5.
For special functions F() it is possible to nd (t, ) directly, by using the Ito formula.
For example, nd (t, ) when
a) F() = W
2
(T)
b) F() = exp W(T)
c) F() =
T
_
0
W(t)dt
d) F() = W
3
(T)
e) F() = cos W(T)
(Hint: Check that N(t): = e
1
2
t
cos W(t) is a martingale.)
1.9

1.4. [Hu] Suppose the function F of Exercise 1.3 has the form
(1.42) F() = f(X(T))
where X(t) = X(t, ) is an Ito diusion given by
(1.43) dX(t) = b(X(t))dt +(X(t))dW(t); X(0) = x R.
Here b: R R and : R R are given Lipschitz continuous functions of at most linear
growth, so (1.43) has a unique solution X(t); t 0. Then there is a useful formula for
the process (t, ) in (1.40). This is described as follows:
If g is a real function with the property
(1.44) E
x
[[g(X(t))[] < for all t 0, x R
(where E
x
denotes expectation w.r.t. the law of X(t) when X(0) = x) then we dene
(1.45) u(t, x): = P
t
g(x): = E
x
[g(X(t))]; t 0, x R
Suppose that there exists > 0 such that
(1.46) [(x)[ for all x R
Then u(t, x) C
1,2
(R
+
R) and
(1.47)
u
t
= b(x)
u
x
+
1
2

2
(x)

2
u
x
2
for all t 0, x R
(Kolmogorovs backward equation).
See e.g. [D2], Theorem 13.18 p. 53 and [D1], Theorem 5.11 p. 162 and [1], Theorem 8.1.
a) Use Itos formula for the process
Y (t) = g(t, X(t)) with g(t, x) = P
Tt
f(x)
to show that
(1.48) f(X(T)) = P
T
f(x) +
T
_
0
[()

P
Tt
f()]
=X(t)
dW(t)
for all f C
2
(R).
In other words, with the notation of Exercise 1.3 we have shown that if F() =
f(X(T)) then
(1.49) E[F] = P
T
f(x) and (t, ) = [()

P
Tt
f()]
=X(t)
.
Use (1.49) to nd E[F] and (t, ) when
1.10

b) F() = W
2
(T)
c) F() = W
3
(T)
d) F() = X(T, )
where
dX(t) = X(t)dt +X(t)dW(t) (, constants)
i.e. X(t) is geometric Brownian motion.
e) Extend formula (1.48) to the case when X(t) R
n
and f: R
n
R. In this case
condition (1.46) must be replaced by the condition
(1.50)
T

T
(x)(x) [[
2
for all x, R
n
where
T
(x) denotes the transposed of the mn-matrix (x).
1.11

2 The Skorohod integral
The Wiener-Ito chaos expansion is a convenient starting point for the introduction of
several important stochastic concepts, including the Skorohod integral. This integral may
be regarded as an extenstion of the Ito integral to integrands which are not necessarily
T
t
-adapted. It is also connected to the Malliavin derivative. We rst introduce some
convenient notation.
Let u(t, ), , t [0, T] be a stochastic process (always assumed to be (t, )-
measurable), such that
(2.1) u(t, ) is T
T
-measurable for all t [0, T]
and
(2.2) E[u
2
(t, )] < for all t [0, T].
Then for each t [0, T] we can apply the Wiener-Ito chaos expansion to the random
variable u(t, ) and obtain functions f
n,t
(t
1
, . . . , t
n
)

L
2
(R
n
) such that
(2.3) u(t, ) =

n=0
I
n
(f
n,t
()).
The functions f
n,t
() depend on the parameter t, so we can write
(2.4) f
n,t
(t
1
, . . . , t
n
) = f
n
(t
1
, . . . , t
n
, t)
Hence we may regard f
n
as a function of n + 1 variables t
1
, . . . , t
n
, t. Since this function
is symmetric with respect to its rst n variables, its symmetrization

f
n
as a function of
n + 1 variables t
1
, . . . , t
n
, t is given by, with t
n+1
= t,
(2.5)

f
n
(t
1
, . . . , t
n+1
) =
1
n + 1
[f
n
(t
1
, . . . , t
n+1
) + +f
n
(t
1
, . . . , t
i1
, t
i+1
, . . . , t
n+1
, t
i
) + +f
n
(t
2
, . . . , t
n+1
, t
1
)],
where we only sum over those permutations of the indices (1, . . . , n + 1) which inter-
change the last component with one of the others and leave the rest in place.
EXAMPLE 2.1. Suppose
f
2,t
(t
1
, t
2
) = f
2
(t
1
, t
2
, t) =
1
2
[A
t
1
<t<t
2

+A
t
2
<t<t
1

].
Then the symmetrization

f
2
(t
1
, t
2
, t
3
) of f
2
as a function of 3 variables is given by

f
2
(t
1
, t
2
, t
3
) =
1
3
[
1
2
(A
t
1
<t
3
<t
2

+A
t
2
<t
3
<t
1

) +
1
2
(A
t
1
<t
2
<t
3

+A
t
3
<t
2
<t
1

) +
1
2
(A
t
2
<t
1
<t
3

+A
t
3
<t
1
<t
2

)]
2.1

This sum is
1
6
except on the set where some of the variables coincide, but this set has
measure zero, so we have
(2.6)

f
2
(t
1
, t
2
, t
3
) =
1
6
a.e.
DEFINITION 2.2. Suppose u(t, ) is a stochastic process satisfying (2.1), (2.2) and
with Wiener-Ito chaos expansion
(2.7) u(t, ) =

n=0
I
n
(f
n
(, t)).
Then we dene the Skorohod integral of u by
(2.8) (u): =
T
_
0
u(t, )W(t) :=

n=0
I
n+1
(

f
n
) (when convergent)
where

f
n
is the symmetrization of f
n
(t
1
, . . . , t
n
, t) as a function of n+1 variables t
1
, . . . , t
n
, t.
We say u is Skorohod-integrable and write u Dom() if the series in (2.8) converges in
L
2
(P). By (1.16) this occurs i
(2.9) E[(u)
2
] =

n=0
(n + 1)!|

f
n
|
2
L
2
([0,T]
n+1
)
< .
EXAMPLE 2.3. Let us compute the Skorohod integral
T
_
0
W(T, )W(t).
Here u(t, ) = W(T, ) =
T
_
0
1 dW(t), so
f
0
= 0, f
1
= 1 and f
n
= 0 for all n 2.
Hence
(u) = I
2
(

f
1
) = I
2
(1) = 2
T
_
0
(
t
2
_
0
dW(t
1
))dW(t
2
) = W
2
(T, ) T.
Note that even if W(T, ) does not depend on t, we have
T
_
0
W(T, )W(t) ,= W(T, )
T
_
0
W(t) (but see (3.64)).
EXAMPLE 2.4. What is
T
_
0
W(t, )[W(T, ) W(t, )]W(t) ?
2.2

Note that
T
_
0
(
t
2
_
0
A
t
1
<t<t
2

(t
1
, t
2
)dW(t
1
))dW(t
2
)
=
T
_
0
W(t, )A
t<t
2

(t
2
)dW(t
2
)
= W(t, )
T
_
t
dW(t
2
) = W(t, )[W(T, ) W(t, )].
Hence
u(t, ): = W(t, )[W(T, ) W(t, )] = J
2
(A
t
1
<t<t
2

(t
1
, t
2
))
= I
2
(f
2
(, t)),
where
f
2
(t
1
, t
2
, t) =
1
2
(A
t
1
<t<t
2

+A
t
2
<t<t
1

).
Hence by Example 2.1 and (1.14)
(u) = I
3
(

f
2
) = I
3
(
1
6
) = (
1
6
)I
3
(1)
=
1
6
[W
3
(T, ) 3T W(T, )].
As mentioned earlier the Skorohod integral is an extension of the Ito integral. More
precisely, if the integrand u(t, ) is T
t
-adapted, then the two integrals coincide. To prove
this, we need a characterization of T
t
-adaptedness in terms of the functions f
n
(, t) in the
chaos expansion:
LEMMA 2.5. Let u(t, ) be a stochastic process satisfying (2.1), (2.2) and let
u(t, ) =

n=0
I
n
(f
n
(, t))
be the Wiener-Ito chaos expansion of u(t, ), for each t [0, T]. Then u(t, ) is T
t
-adapted
if and only if
(2.10) f
n
(t
1
, . . . , t
n
, t) = 0 if t < max
1in
t
i
.
REMARK The statement (2.10) should as most statements about L
2
-functions be
regarded as an almost everywhere (a.e.) statement. More precisely, (2.10) means that for
each t [0, T] we have
f
n
(t
1
, . . . , t
n
, t) = 0 for a.a. (t
1
, . . . , t
n
) H,
where H = (t
1
, . . . , t
n
) [0, T]
n
; t < max
1in
t
i
.
2.3

Proof of Lemma 2.5. First note that for any g

L
2
([0, T]
n
) we have
E[I
n
(g)[T
t
] = n!E[J
n
(g)[T
t
]
= n!E[
T
_
0

t
n
_
0

t
2
_
0
g(t
1
, . . . , t
n
)dW(t
1
) dW(t
n
)[T
t
]
= n!
t
_
0

t
n
_
0

t
2
_
0
g(t
1
, . . . , t
n
)dW(t
1
) dW(t
n
)
= n!J
n
(g(t
1
, . . . , t
n
) A
max t
i
<t
)
= I
n
(g(t
1
, . . . , t
n
) A
max t
i
<t
). (2.11)
Hence
u(t, ) is T
t
-adpted
E[u(t, )[T
t
] = u(t, )

n=0
E[I
n
(f
n
(, t))[T
t
] =

n=0
I
n
(f
n
(, t))

n=0
I
n
(f
n
(, t) A
max t
i
<t
) =

n=0
I
n
(f
n
(, t))
f
n
(t
1
, . . . , t
n
, t) A
max t
i
<t
= f
n
(t
1
, . . . , t
n
, t) a.e.,
by uniqueness of the Wiener-Ito expansion. Since the last identity is equivalent to (2.10),
the Lemma is proved. .
THEOREM 2.6. (The Skorohod integral is an extension of the Ito integral)
Let u(t, ) be a stochastic process such that
(2.12) E[
T
_
0
u
2
(t, )dt] <
and suppose that
(2.13) u(t, ) is T
t
-adapted for t [0, T].
Then u Dom() and
(2.14)
T
_
0
u(t, )W(t) =
T
_
0
u(t, )dW(t)
Proof. First note that by (2.5) and Lemma 2.5 we have
(2.15)

f
n
(t
1
, . . . , t
n
, t
n+1
) =
1
n + 1
f
n
( , t
j1
, t
j+1
, . . . , t
j
)
where
t
j
= max
1in+1
t
i
.
2.4

Hence
|

f
n
|
2
L
2
([0,T]
n+1
)
= (n + 1)!
_
S
n+1

f
2
n
(x
1
, . . . , x
n+1
)dx
1
dx
n+1
=
(n + 1)!
(n + 1)
2
_
S
n+1
f
2
n
(x
1
, . . . , x
n+1
)dx
1
dx
n
=
n!
n + 1
T
_
0
(
t
_
0
x
n
_
0

x
2
_
0
f
2
n
(x
1
, . . . , x
n
, t)dx
1
dx
n
)dt
=
n!
n + 1
T
_
0
(
T
_
0
x
n
_
0

x
2
_
0
f
2
n
(x
1
, . . . , x
n
, t)dx
1
dx
n
)dt
=
1
n + 1
T
_
0
|f
n
(, t)|
2
L
2
([0,T]
n
)
dt,
again by using Lemma 2.5.
Hence, by (1.16),

n=0
(n + 1)!|

f
n
|
2
L
2
([0,T]
n+1
)
=

n=0
n!
T
_
0
|f
n
(, t)|
2
L
2
([0,T]
n
)
dt
=
T
_
0
(

n=0
n!|f
n
(, t)|
2
L
2
([0,T]
n
)
)dt
= E[
T
_
0
u
2
(t, )dt] < by assumption. (2.16)
This proves that u Dom().
Finally, to prove (2.14) we again apply (2.15):
T
_
0
u(t, )dW(t) =

n=0
T
_
0
I
n
(f
n
(, t))dW(t)
=

n=0
T
_
0
n!
_
0t
1
t
n
t
f
n
(t
1
, . . . , t
n
, t)dW(t
1
) dW(t
n
)dW(t)
=

n=0
T
_
0
n!(n + 1)
_
0t
1
t
n
t
n+1

f
n
(t
1
, . . . , t
n
, t
n+1
)dW(t
1
) dW(t
n
)dW(t
n+1
)
=

n=0
(n + 1)!J
n+1
(

f
n
) =

n=0
I
n+1
(

f
n
) =
T
_
0
u(t, )W(t).
.
2.5

Exercises
2.1 Compute the following Skorohod integrals:
a)
T
_
0
W(t)W(t)
b)
T
_
0
(
T
_
0
g(s)dW(s))W(t) (g L
2
([0, T]) deterministic)
c)
T
_
0
W
2
(t
0
)W(t) (t
0
[0, T] xed)
d)
T
_
0
exp(W(T))W(t)
(Hint: Use Exercise 1.2.)
2.6

3 White noise, the Wick product and stochastic in-
tegration
This chapter gives an introduction to the white noise analysis and its relation to the
analysis on Wiener spaces discussed in the rst two chapters. Although it is not strictly
necessary for the following chapters, it gives a useful alternative approach. Moreover, it
provides a natural platform for the Wick product, which is closely related to Skorohod
integration (see (3.22)). For example, we shall see that the Wick calculus can be used to
simplify the computation of these integrals considerably.
The Wick product was introduced by C. G. Wick in 1950 as a renormalization technique
in quantum physics. This concept (or rather a relative of it) was introduced by T. Hida
and N. Ikeda in 1965. In 1989 P. A. Meyer and J. A. Yan extended the construction to
cover Wick products of stochastic distributions (Hida distributions), including the white
noise.
The Wick product has turned out to be a very useful tool in stochastic analysis in general.
For example, it can be used to facilitate both the theory and the explicit calculations in
stochastic integration and stochastic dierential equations. For this reason we include a
brief introduction in this course. It remains to be seen if the Wick product also has more
direct applications in economics.
General references for this section are [H], [HKPS], [HUZ], [HP], [LU 1-3], [1], [2]
and [GHLUZ].
We start with the construction of the white noise probability space (o
t
, B, ):
Let o = o(R) be the Schwartz space of rapidly decreasing smooth functions on R with the
usual topology and let o
t
= o
t
(R) be its dual (the space of tempered distributions). Let
B denote the family of all Borel subsets of o
t
(R) (equipped with the weak-star topology).
If o
t
and o we let
(3.1) () = , )
denote the action of on . (For example, if is a measure m on R then
, ) =
_
R
(x)dm(x)
and if is evaluation at x
0
R then
, ) = (x
0
) etc.)
By the Minlos theorem [GV] there exists a probability meaure on o
t
such that
(3.2)
_
S

e
i,)
d() = e

1
2
||
2
; o
where
(3.3) ||
2
=
_
R
[(x)[
2
dx = ||
2
L
2
(R)
.
3.1

is called the white noise probability measure and (o
t
, B, ) is called the white noise
probability space.
DEFINITION 3.1 The (smoothed) white noise process is the map
w : o o
t
R
given by
(3.4) w(, ) = w

() = , ) ; o, o
t
From w

we can construct a Wiener process (Brownian motion) W

t
as follows:
STEP 1. (The Ito isometry)
(3.5) E

[, )
2
] = ||
2
; o
where E

denotes expectation w.r.t. , so that

E

[, )
2
] =
_
S

, )
2
d().
STEP 2. Use Step 1 to dene, for arbitrary L
2
(R),
, ) := lim,
n
),
where
n
o and
n
in L
2
(R) (3.6)
STEP 3. Use Step 2 to dene
(3.7)

W
t
() =

W(t, ) := ,
[0,t]
()) for t 0
by choosing
(s) =
[0,t]
(s) =
_
1 if s [0, t]
0 if s , [0, t]
which belongs to L
2
(R) for all t 0.
STEP 4. Prove that

W
t
has a continuous modication W
t
= W
t
(), i.e.
P[

W
t
() = W
t
()] = 1 for all t.
This continuous process W
t
= W
t
() = W(t, ) = W(t) is a Wiener process.
Note that when the Wiener process W
t
() is constructed this way, then each is inter-
preted as an element of : = o
t
(R), i.e. as a tempered distribution.
From the above it follows that the relation between smoothed white noise w

() and the
Wiener process W
t
() is
(3.8) w

() =
_
R
(t)dW
t
() ; o
where the integral on the right is the Wiener-Ito integral.
3.2

The Wiener-Ito chaos expansion revisited
As before let the Hermite polynomials h
n
(x) be dened by
(3.9) h
n
(x) = (1)
n
e
x
2
2
d
n
dx
n
(e

x
2
2
) ; n = 0, 1, 2,
This gives for example
h
0
(x) = 1, h
1
(x) = x, h
2
(x) = x
2
1, h
3
(x) = x
3
3x
h
4
(x) = x
4
6x
2
+ 3, h
5
(x) = x
5
10x
3
+ 15x,
Let e
k
be the kth Hermite function dened by
(3.10) e
k
(x) =

1
4
((k 1)!)

1
2
e

x
2
2
h
k1
(

2x); k = 1, 2,
Then e
k

k1
constitutes an orthonormal basis for L
2
(R) and e
k
o for all k.
Dene
(3.11)
k
() = , e
k
) = W
e
k
() =
_
R
e
k
(x)dW
x
()
Let denote the set of all nite multi-indices = (
1
,
2
, . . . ,
m
) (m = 1, 2, . . .) of
non-negative integers
i
. If = (
1
, ,
m
) we put
(3.12) H

() =
m

j=1
h

j
(
j
)
For example, if =
k
= (0, 0, , 1) with 1 on kth place, then
H

k
() = h
1
(
k
) = , e
k
),
while
H
3,0,2
() = h
3
(
1
)h
0
(
2
)h
2
(
3
) = (
3
1
3
1
) (
2
3
1).
The family H

()

is an orthogonal basis for the Hilbert space

(3.13) L
2
() = X : o
t
R such that |X|
2
L
2
()
:=
_
S

X()
2
d() < .
In fact, we have
THEOREM 3.2 (The Wiener-Ito chaos expansion theorem II)
For all X L
2
() there exist (uniquely determined) numbers c

R such that
(3.14) X() =

().
Moreover, we have
(3.15) |X|
2
L
2
()
=

!c
2

3.3

where ! =
1
!
2
!
m
! if = (
1
,
2
,
m
).
Let us compare with the equivalent formulation of this theorem in terms of multiple Ito
integrals: (See Chapter 1)
If (t
1
, t
2
, , t
n
) is a real symmetric function in its n (real) variables t
1
, , t
n
and
L
2
(R
n
), i.e.
(3.16) ||
L
2
(R
n
)
:= [
_
R
n
[(t
1
, t
2
, , t
n
)[
2
dt
1
dt
2
dt
n
]
1/2
<
then its n-tuple Ito integral is dened by
I
n
(): =
_
R
n
dW
n
:=
n!
_

(
_
t
n

(
_
t
n1

(
_
t
2

(t
1
, t
2
, , t
n
)dW
t
1
)dW
t
2
)dW
t
n
(3.17)
where the integral on the right consists of n iterated Ito integrals (note that in each
step the corresponding integrand is adapted because of the upper limits of the preceding
integrals). Applying the Ito isometry n times we see that
(3.18) E[(
_
R
n
dW
n
)
2
] = n!||
2
L
2
(R
n
)
; n 1
For n = 0 we adopt the convention that
(3.19) I
0
(): =
_
R
0
dW
0
= = ||
L
2
(R
0
)
when is constant
Let

L
2
(R
n
) denote the set of symmetric real functions (on R
n
) which are square integrable
with respect to Lebesque measure. Then we have (see Theorem 1.1):
THEOREM 3.3 (The Wiener-Ito chaos expansion theorem I)
For all X L
2
() there exist (uniquely determined) functions f
n


L
2
(R
n
) such that
(3.20) X() =

n=0
_
R
n
f
n
dW
n
() =

n=0
I
n
(f
n
)
Moreover, we have
(3.21) |X|
2
L
2
()
=

n=0
n!|f
n
|
2
L
2
(R
n
)
REMARK The connection between these two expansions in Theorem 3.2 and Theorem
3.3 is given by
(3.22) f
n
=

J
||=n
c

1
1

e

2
2



e

m
m
; n = 0, 1, 2,
3.4

where [[ =
1
+ +
m
if = (
1
, ,
m
) (m = 1, 2, ). The functions e
1
, e
2
,
are dened in (3.10) and and

denote tensor product and symmetrized tensor product,
respectively. For example, if f and g are real functions on R then
(f g)(x
1
, x
2
) = f(x
1
)g(x
2
)
and
(f

g)(x
1
, x
2
) =
1
2
[f(x
1
)g(x
2
) +f(x
2
)g(x
1
)] ; (x
1
, x
2
) R
2
.
Analogous to the test functions o(R) and the tempered distributions o
t
(R) on the real
line R, there is a useful space of stochastic test functions (o) and a space of stochastic
distributions (o)

on the white noise probability space:

DEFINITION 3.4 ([Z])
a) We say that f =

L
2
() belongs to the Hida test function space (o) if
(3.23)

!a
2

j=1
(2j)

k
< for all k <
b) A formal sum F =

belongs to the Hida distribution space (o)

if
(3.24) there exists q < s.t.

!c
2

j=1
(2j)

q
<
(o)

is the dual of (o). The action of F =

(o)

on f =

(o) is given
by
F, f) =

!a

We have the inclusions

(o) L
2
() (o)

.
EXAMPLE 3.5
a) The smoothed white noise w

() belongs to (o) if o, because if =

j
c
j
e
j
we
have
(3.25) w

j
c
j
H

j
so w

(o) if and only if (using (3.23))

j
c
2
j
(2j)
k
< for all k,
which holds because o. (See e.g. [RS]).
3.5

b) The singular (pointwise) white noise

W
t
() is dened as follows:
(3.26)

W
t
() =

k
e
k
(t)H

k
()
Using (3.24) one can verify that

W
t
() (o)

for all t. This is the precise

denition of singular/pointwise white noise!
The Wick product
In addition to a canonical vector space structure, the spaces (o) and (o)

also have a
natural multiplication:
DEFINITION 3.6 If X =

(o)

, Y =

(o)

then the Wick product,

X Y , of X and Y is dened by
(3.27) X Y =

,
a

H
+
=

+=
a

)H

Using (3.24) and (3.23) one can now verify the following:
(3.28) X, Y (o)

X Y (o)

(3.29) X, Y (o) X Y (o)

(Note, however, that X, Y L
2
() , X Y L
2
() in general)
EXAMPLE 3.7
(i) The Wick square of white noise is
(singular case)

W
t
2
= (

W
t
)
2
=

k,m
e
k
(t)e
m
(t)H

k
+
m
(smoothed case) w
2

k,m
c
k
c
m
H

k
+
m
if =

c
k
e
k
o
Since
H

k
+
m
=
_
H

k
H

m
if k ,= m
H
2

k
1 if k = m
we see that
w
2

= w
2

k
c
2
k
= w
2

||
2
.
Note, in particular, that w
2

is not positive. In fact, E[w

2

] = 0 by (2.5).
3.6

(ii) The Wick exponential of smoothed white noise is dened by
exp

n=0
1
n!
w
n

; o.
It can be shown that (see Exercise 1.1)
(3.30) exp

= exp(w

1
2
||
2
)
so exp

is positive. Moreover, we have

(3.31) E

[exp

] = 1 for all o.
Why the Wick product?
We list some reasons that the Wick product is natural to use in stochastic calculus:
1) First, note that if (at least) one of the factors X, Y is deterministic, then
X Y = X Y
Therefore the two types of products, the Wick product and the ordinary (-pointwise)
product, coincide in the deterministic calculus. So when one extends a deterministic
model to a stochastic model by introducing noise, it is not obvious which interpreta-
tion to choose for the products involved. The choice should be based on additional
modelling and mathematical considerations.
2) The Wick product is the only product which is dened for singular white noise

W
t
.
Pointwise product X Y does not make sense in (o)

!
3) The Wick product has been used for 40 years already in quantum physics as a
renormalization procedure.
4) Last, but not least: There is a fundamental relation between Ito/Skorohod integrals
and Wick products, given by
(3.32)
_
Y
t
()W
t
() =
_
Y
t

W
t
dt
(see [LU 2], [B]).
Here the integral on the right is interpreted as a Pettis integral with values in (o)

.
In view of (3.32) one could say that the Wick product is the core of Ito integration, hence
it is natural to use in stochastic calculus in general.
Finally we recall the denition of a pair of dual spaces, ( and (

, which are sometimes

useful. See [PT] and the references therein for more information.
3.7

DEFINITION 3.8
a) Let R. Then the space (

consists of all formal expansions

(3.33) X =

n=0
_
R
n
f
n
dW
n
such that
(3.34) |X|

:= [

n=0
n!e
2n
|f
n
|
2
L
2
(R
n
)
]
1
2
<
For each R the space (

is a Hilbert space with inner product

(X, Y )

n=0
n!e
2n
(f
n
, g
n
)
L
2
(R
n
)
if X =

n=0
_
R
n
f
n
dW
n
, Y =

m=0
_
R
m
g
m
dW
m
(3.35)
Note that
1

2
(

2
(

1
. Dene
(3.36) ( =

R
(

, with projective limit topology.

b) (

is dened to be the dual of (. Hence

(3.37) (

=
_
R
(

, with inductive limit topology.

REMARK. Note that an element Y (

can be represented as a formal sum

(3.38) Y =

n=0
_
R
n
g
n
dW
n
where g
n


L
2
(R
n
) and |Y |

< for some R, while an X ( satises |X|

<
for all R.
If X ( and Y (

have the representations (3.33), (3.38), respectively, then the action

of Y on X, Y, X), is given by
(3.39) Y, X) =

n=0
n!(f
n
, g
n
)
L
2
(R
n
)
where
(3.40) (f
n
, g
n
)
L
2
(R
n
)
=
_
R
n
f(x)g(x)dx
One can show that
(3.41) (o) ( L
2
() (

(o)

.
3.8

The space (

is not big enough to contain the singular white noise W

t
. However, it does
often contain the solution X
t
of stochastic dierential equations. This fact allows one to
deduce some useful properties of X
t
.
Like (o) and (o)

the spaces ( and (

are closed under Wick product ([PT, Theorem

2.7]):
(3.42) X
1
, X
2
( X
1
X
2
(
(3.43) Y
1
, Y
2
(

Y
1
Y
2
(

The Wick product in terms of iterated Ito integrals

The denition we have given of the Ito product is based on the chaos expansion II, because
only this is general enough to include the singular white noise. However, it is useful to
know how the Wick product is expressed in terms of chaos expansion I for L
2
()-functions
or, more generally, for elements of (

:
THEOREM 3.9 Suppose X =

n=0
I
n
(f
n
) (

, Y =

m=0
I
m
(g
m
) (

. Then the Wick

product of X and Y can be expressed by
(3.44) X Y =

n,m=0
I
n+m
(f
n

g
m
) =

k=0
(

n+m=k
I
k
(f
n

g
m
)).
For example, integration by parts gives that
(
_
R
f(x)dW
x
) (
_
R
g(y)dW
y
) =
_
R
2
(f

g)(x, y)dW
2
=
_
R
(
y
_

(f(x)g(y) +f(y)g(x))dW
x
)dW
y
=
_
R
g(y)(
y
_

f(x)dW
x
)dW
y
+
_
R
f(y)(
y
_

g(x)dW
x
)dW
y
= (
_
R
g(y)dW
y
)(
_
R
f(x)dW
x
)
_
R
f(t)g(t)dt . (3.45)
Some properties of the Wick product
We list below some useful properties of the Wick product. Some are easy to prove, others
harder. For complete proofs see [HUZ].
For arbitrary X, Y, Z (

we have
X Y = Y X (commutative law) (3.46)
X (Y Z) = (X Y ) Z (associative law) (3.47)
X (Y +Z) = (X Y ) + (X Z) (distributive law) (3.48)
3.9

Thus the Wick algebra obeys the same rules as the ordinary algebra. For example,
(3.49) (X +Y )
2
= X
2
+ 2X Y +Y
2
(no Ito formula!)
and
(3.50) exp

(X +Y ) = exp

(X) exp

(Y ).
Note, however, that combinations of ordinary products and Wick products requires cau-
tion. For example, in general we have
X (Y Z) ,= (X Y ) Z .
A remarkable property of the Wick product is that
(3.51) E

[X Y ] = E

[X] E

[Y ]
whenever X, Y and X Y are -integrable. (Note that it is not required that X and Y
are independent!)
A reformulation of (3.45) is that
w

= w

1
2
_
R
(t)(t)dt; , .
(See also Example 3.7(i))
In particular,
(3.52) W
2
t
= W
2
t
t; t 0
and
(3.53) if supp supp = , then w

= w

Hence
If 0 t
1
t
2
t
3
t
4
then (3.54)
(W
t
4
W
t
3
) (W
t
2
W
t
1
) = (W
t
4
W
t
3
) (W
t
2
W
t
1
)
More generally, it can be proved that if F() is T
t
-measurable and h > 0, then
(3.55) F (W
t+h
W
t
) = F (W
t+h
W
t
)
(For a proof see e.g. Exercise 2.22 in [HUZ].)
Note that from (3.44) we have that
(3.56) (
_
R
g(t)dW
t
)
n
= n!
_
R
n
g
n
(x
1
, . . . , x
n
)dW
n
; g L
2
(R).
3.10

Combining this with (1.14) we get, with =
_
R
gdW,
(3.57)
n
= |g|
n
h
n
(

|g|
).
In particular,
(3.58) W
n
t
= t
n/2
h
n
(
W
t

t
), n = 0, 1, 2, . . .
Moreover, combining (3.57) with the generating formula for Hermite polynomials
(3.59) exp(tx
t
2
2
) =

n=0
t
n
n!
h
n
(x) (see Exercise 1.1)
we get (see Example 3.7 (ii))
exp(
_
R
g(t)dW
t

1
2
|g|
2
) =

n=0
|g|
n
n!
h
n
(

|g|
)
=

n=0
1
n!

n
= exp

. (3.60)
Hence
(3.61) exp

(
_
R
gdW) = exp(
_
R
gdW
1
2
|g|
2
); g L
2
(R).
In particular,
(3.62) exp

(W
t
) = exp(W
t

1
2
t); t 0.
Combining the properties above with the fundamental relation (3.32) for Skorohod inte-
gration, we get a powerful calculation technique for stochastic integration. First of all,
note that, by (3.32),
(3.63)
T
_
0

W
t
dt = W
T
.
Moreover, using (3.48) one can deduce that
(3.64)
T
_
0
X Y
t

W
t
dt = X
T
_
0
Y
t

W
t
dt
if X does not depend on t.
(Compare this with the fact that for Skorohod integrals we generally have
(3.65)
T
_
0
X Y
t
W
t
,= X
T
_
0
Y
t
W
t
,
3.11

even if X does not depend on t.)
To illustrate the use of Wick calculus, let us again consider Example 2.4:
T
_
0
W
t
[W
T
W
t
]W
t
=
T
_
0
W
t
(W
T
W
t
)

W
t
dt
=
T
_
0
W
t
W
T

W
t
dt
T
_
0
W
2
t

W
t
dt
= W
T

T
_
0
W
t

W
t
dt
1
3
W
3
T
=
1
6
W
3
T
=
1
6
[W
3
T
3TW
T
],
where we have used (3.54), (3.48), (3.64) and (3.58).
Exercises
3.1 Use the identity (3.32) and Wick calculus to compute the following Skorohod inte-
grals
a)
T
_
0
W(T)W(t) =
T
_
0
W(T)

W
(t)dt
b)
T
_
0
(
T
_
0
g(s)dW(s))W(t) (g L
2
([0, T]) deterministic)
c)
T
_
0
W
2
(t
0
)W(t) (t
0
[0, T] xed)
d)
T
_
0
exp(W(T))W(t)
Compare with your calculations in Exercise 2.1!
3.12

4 Dierentiation
Let us rst recall some basic concepts from classical analysis:
DEFINITION 4.1. Let U be an open subset of R
n
and let f be a function from U
into R
m
.
a) We say that f has a directional derivative at the point x U in the direction y R
n
if
(4.1) D
y
f(x): = lim
0
f(x +y) f(x)

=
d
d
[f(x +y)]
=0
exists. If this is the case we call the vector D
y
f(x) R
m
the directional derivative
(at x in direction y). In particular, if we choose y to be the jth unit vector e
j
=
(0, . . . , 1, . . . , 0), with 1 on jth place, we get
D

j
f(x) =
f
x
j
,
the jth partial derivative of f.
b) We say that f is dierentiable at x U if there exists a matrix A R
mn
such that
(4.2) lim
h0
hR
n
1
[h[
[f(x +h) f(x) Ah[ = 0
If this is the case we call A the derivative of f at x and we write
A = f
t
(x).
The following relations between the two concepts are well-known:
PROPOSITION 4.2.
(i) If f is dierentiable at x U then f has a directional derivative in all directions
y R
n
and
(4.3) D
y
f(x) = f
t
(x)y
(ii) Conversely, if f has a directional derivative at all x U in all the directions y = e
j
;
1 j n and all the partial derivatives
D
e
j
f(x) =
f
x
j
(x)
are continuous functions of x, then f is dierentiable at all x U and
(4.4) f
t
(x) = [
f
i
x
j
] 1im
1jn
R
mn
,
4.1

where f
i
is component number i of f, i.e.
f =
_

_
f
1
.
.
.
f
m
_

_
We will now dene similar operations in a more general context. First let us recall some
basic concepts from functional analysis:
DEFINITION 4.3. Let X be a Banach space, i.e. a complete, normed vector space
(over R), and let |x| denote the norm of the element x X. A linear functional on X
is a linear map
T: X R
(T is called linear if T(ax+y) = aT(x)+T(y) for all a R, x, y X). A linear functional
T is called bounded (or continuous) if
[|T|[: = sup
|x|1
[T(x)[ <
Sometimes we write T, x) or Tx instead of T(x) and call T, x) the action of T on x.
The set of all bounded linear functionals is called the dual of X and is denoted by X

.
Equipped with the norm [| |[ the space X

becomes a Banach space also.

EXAMPLE 4.4.
(i) X = R
n
with the usual Euclidean norm [x[ =
_
x
2
1
+ +x
2
n
is a Banach space. In
this case it is easy to see that we can identify X

with R
n
.
(ii) Let X = C
0
([0, T]), the space of continuous, real functions on [0, T] such that
(0) = 0. Then
||

: = sup
t[0,T]
[(t)[
is a norm on X called the uniform norm. This norm makes X into a Banach space
and its dual X

can be identied with the space /([0, T]) of all signed measures
on [0, T], with norm
[||[ = sup
[f[1
T
_
0
f(t)d(t) = [[([0, T])
(iii) If X = L
p
([0, T]) = f: [0, T] R;
T
_
0
[f(t)]
p
dt < equipped with the norm
|f|
p
= [
T
_
0
[f(t)[
p
dt]
1/p
(1 p < )
then X is a Banach space, whose dual can be identied with L
q
([0, T]), where
1
p
+
1
q
= 1 .
In particular, if p = 2 then q = 2 so L
2
([0, T]) is its own dual.
4.2

We now extend the denitions we had for R
n
to arbitrary Banach spaces:
DEFINITION 4.5. Let U be an open subset of a Banach space X and let f be a
function from U into R
m
.
a) We say that f has a directional derivative (or Gateaux derivative) at x U in the
direction y X if
(4.5) D
y
f(x): =
d
d
[f(x +y)]
=0
R
m
exists. If this is the case we call D
y
f(x) the directional (or Gateaux) derivative of
f (at x in the direction y).
b) We say that f is Frechet-dierentiable at x U if there exists a bounded linear map
A: X R
m
(i.e. A =
_

_
A
1
.
.
.
A
m
_

_ with A
i
X

for 1 i m) such that

(4.6) lim
h0
hX
1
|h|
[f(x +h) f(x) A(h)[ = 0
If this is the case we call A the Frechet derivative of f at x and we write
(4.7) A = f
t
(x) =
_

_
f
t
(x)
1
.
.
.
f
t
(x)
m
_

_ (X

)
m
.
Similar to the Euclidean case (Proposition 4.2) we have
PROPOSITION 4.6.
(i) If f is Frechet-dierentiable at x U X then f has a directional derivative at x
in all directions y X and
(4.8) D
y
f(x) = f
t
(x), y) R
m
where
f
t
(x), y) = (f
t
(x)
1
, y), . . . , f
t
(x)
m
, y))
is the m-vector whose ith component is the action of the ith component f
t
(x)
i
of
f
t
(x) on y.
(ii) Conversely, if f has a directional derivative at all x U in all directions y X and
the (linear) map
y D
y
f(x) ; y X
4.3

is continuous for all x U, then there exists an element f(x) (X

)
m
such that
D
y
f(x) = f(x), y) .
If this map x f(x) (X

)
m
is continuous on U, then f is Frechet dierentiable
and
(4.9) f
t
(x) = f(x) .
We now apply these operations to the Banach space = C
0
([0, T]) considered in Example
4.4 (ii) above. This space is called the Wiener space, because we can regard each path
t W(t, )
of the Wiener process starting at 0 as an element of C
0
([0, 1]). Thus we may identify
W(t, ) with the value (t) at time t of an element C
0
([0, T]):
W(t, ) = (t)
With this identication the Wiener process simply becomes the space = C
0
([0, T])
and the probability law P of the Wiener process becomes the measure dened on the
cylinder sets of by
(; (t
1
) F
1
, . . . , (t
k
) F
k
) = P[W(t
1
) F
1
, . . . , W(t
k
) F
k
]
=
_
F
1
F
k
(t
1
, x, x
1
)(t
1
t
1
, x, x
2
) (t
k
t
k1
, x
k1
, x
k
)dx
1
, dx
k
where F
i
R; 0 t
1
< t
2
< < t
k
and
(t, x, y) = (2t)
1/2
exp(
1
2
[x y[
2
); t > 0; x, y R.
The measure is called the Wiener measure on . In the following we will write L
2
()
for L
2
() and L
2
([0, T] ) for L
2
() etc., where is the Lebesgue measure on [0, T].
Just as for Banach spaces in general we now dene
DEFINITION 4.6. As before let L
2
([0, T]) be the space of (deterministic) square
integrable functions with respect to Lebesgue measure (dt) = dt on [0, T]. Let F: R
be a random variable, choose g L
2
([0, T]) and put
(4.10) (t) =
t
_
0
g(s)ds .
Then we dene the directional derivative of F at the point in direction by
(4.11) D

F() =
d
d
[F( +)]
=0
,
if the derivative exists in some sense (to be made precise below).
4.4

Note that we only consider the derivative in special directions, namely in the directions of
elements of the form (4.10). The set of which can be written on the form (4.10)
for some g L
2
([0, T]) is called the Cameron-Martin space and denoted by H. It turns
out that it is dicult to obtain a tractable theory involving derivatives in all directions.
However, the derivatives in the directions H are sucient for our purposes.
DEFINITION 4.7. Assume that F: Rhas a directional derivative in all directions
of the form
(t) =
t
_
0
g(s)ds with g L
2
([0, T])
in the strong sense that
(4.12) D

F(): = lim
0
1

[F( +) F()]
exists in L
2
(). Assume in addition that there exists (t, ) L
2
([0, T] ) such that
(4.13) D

F() =
T
_
0
(t, )g(t)dt .
Then we say that F is dierentiable and we set
(4.14) D
t
F(): = (t, ).
We call D
t
F() L
2
([0, T] ) the derivative of F.
The set of all dierentiable random variables is denoted by T
1,2
.
EXAMPLE 4.8. Suppose F() =
T
_
0
f(s)dW(s) =
T
_
0
f(s)d(s), where f(s) L
2
([0, T]).
Then if (t) =
t
_
0
g(s)ds we have
F( +) =
T
_
0
f(s)(d(s) +d(s))
=
T
_
0
f(s)d(s) +
T
_
0
f(s)g(s)ds,
and hence
1

[F( +) F()] =
T
_
0
f(s)g(s)ds for all > 0.
Comparing with (4.13) we see that F T
1,2
and
(4.15) D
t
F() = f(t); t [0, T], .
In particular, choose
f(t) = A
[0,t
1
]
(t)
4.5

we get
F() =
T
_
0
A
[0,t
1
]
(s)dW(s) = W(t
1
, )
and hence
(4.16) D
t
(W(t
1
, )) = A
[0,t
1
]
(t).
Let P denote the family of all random variables F: R of the form
F() = (
1
, . . . ,
n
)
where (x
1
, . . . , x
n
) =

is a polynomial in n variables x
1
, . . . , x
n
and

i
=
T
_
0
f
i
(t)dW(t) for some f
i
L
2
([0, T]) (deterministic).
Such random variables are called Wiener polynomials. Note that P is dense in L
2
().
By combining (4.16) with the chain rule we get that P T
1,2
:
LEMMA 4.9. Let F() = (
1
, . . . ,
n
) P. Then F T
1,2
and
(4.17) D
t
F() =
n

i=1

x
i
(
1
, . . . ,
n
) f
i
(t).
Proof. Let (t, ) denote the right hand side of (4.17). Since
sup
s[0,T]
E[[W(s)[
N
] < for all N N,
we see that
1

[F( +) F()] =
1

[(
1
+(f
1
, g), . . . ,
n
+(f
n
, g) (
1
, . . . ,
n
)]

i=1

x
i
(
1
, . . . ,
n
) D

(
i
) in L
2
() as 0
Hence F has a directional derivative in direction (in the strong sense) and by (4.15) we
have
D

F() =
T
_
0
(t, )g(t)dt.
.
We now introduce the following norm, | |
1,2
, on T
1,2
:
(4.18) |F|
1,2
= |F|
L
2
()
+|D
t
F|
L
2
([0,T])
; F T
1,2
.
Unfortunately, it is not clear if T
1,2
is closed under this norm, i.e. if any | |
1,2
-Cauchy
sequence in T
1,2
converges to an element of T
1,2
. To avoid this diculty we work with
the following family:
4.6

DEFINITION 4.10. We dene D
1,2
to be the closure of the family P with respect to
the norm | |
1,2
.
Thus D
1,2
consists of all F L
2
() such that there exists F
n
P with the property that
(4.19) F
n
F in L
2
() as n
and
(4.20) D
t
F
n

n=1
is convergent in L
2
([0, T] ).
If this is the case, it is tempting to dene
D
t
F: = lim
n
D
t
F
n
.
However, for this to work we need to know that this denes D
t
F uniquely. In other words,
if there is another sequence G
n
P such that
(4.21) G
n
F in L
2
() as n
and
(4.22) D
t
G
n

n=1
is convergent in L
2
([0, T] ),
does it follow that
(4.23) lim
n
D
t
F
n
= lim
n
D
t
G
n
?
By considering the dierence H
n
= F
n
G
n
we see that the answer to this question is
yes, in virtue of the following theorem:
THEOREM 4.11. (Closability of the operator D
t
)
Suppose H
n

n=1
P has the properties
(4.26) H
n
0 in L
2
() as n
and
(4.27) D
t
H
n

n=1
converges in L
2
([0, T] ) as n
Then
lim
n
D
t
H
n
= 0.
The proof is based on the following useful result:
4.7

LEMMA 4.12. (Integration by parts)
Suppose F T
1,2
, T
1,2
and (t) =
t
_
0
g(s)ds with g L
2
([0, T]). Then
(4.28) E[D

F ] = E[F
T
_
0
gdW] E[F D

]
Proof. By the Girsanov theorem we have
E[F( +) ()] = E[F()( ) exp(
T
_
0
gdW
1
2

2
T
_
0
g
2
ds)]
and this gives
E[D

F() () = E[lim
0
1

[F( +) F()] ()]

= lim
0
1

E[F( +)() F()()]

= lim
0
1

E[F()[( ) exp(
T
_
0
gdW
1
2

2
T
_
0
g
2
ds) ()]]
= E[F()
d
d
[( ) exp(
T
_
0
gdW
1
2

2
T
_
0
g
2
ds)]
=0
]
= E[F()()
T
_
0
gdW] E[F()D

()].
.
Proof of Theorem 4.11. By Lemma 4.12 we get
E[D

H
n
] = E[H
n

T
_
0
gdW] E[H
n
D

]
0 as n for all P.
Since D

H
n

n=1
converges in L
2
() and P is dense in L
2
() we conclude that D

H
n
0
in L
2
() as n . Since this holds for all =

_
0
gds, we obtain that D
t
H
n
0 in
L
2
([0, T] ). .
In view of Theorem 4.11 and the discussion preceding it, we can now make the following
(unambiguous) denition:
DEFINITION 4.13. Let F D
1,2
, so that there exists F
n
P such that
F
n
F in L
2
()
and
D
t
F
n

n=1
is convergent in L
2
([0, T] ).
4.8

Then we dene
(4.29) D
t
F = lim
n
D
t
F
n
and
D

F =
T
_
0
D
t
F g(t)dt
for all (t) =
t
_
0
g(s)ds with g L
2
([0, T]).
We will call D
t
F the Malliavin derivative of F.
REMARK Strictly speaking we now have two apparently dierent denitions of the
derivative of F:
1) The derivative D
t
F of F T
1,2
given by Denition 4.7.
2) The Malliavin derivative D
t
F of F D
1,2
given by Denition 4.13.
However, the next result shows that if F T
1,2
D
1,2
then the two derivatives coincide:
LEMMA 4.14. Let F T
1,2
D
1,2
and suppose that F
n
P has the properties
(4.30) F
n
F in L
2
() and D
t
F
n
converges in L
2
([0, T] ).
Then
(4.31) D
t
F = lim
n
D
t
F
n
.
Hence
(4.32) D
t
F = D
t
F for F T
1,2
D
1,2
.
Proof. By (4.30) we get that D

F
n
converges in L
2
() for each (t) =
t
_
0
g(s)ds;
g L
2
([0, T]). By Lemma 4.12 we get
E[(D

F
n
D

F) ]
= E[(F
n
F)
t
_
0
gdW] E[(F
n
F) D

]
0 for all P by (4.30).
Hence D

F
n
D

F in L
2
() and (4.31) follows. .
4.9

In view of Lemma 4.14 we will now use the same symbol D
t
F and D

F for the derivative

and directional derivative, respectively, of elements F T
1,2
D
1,2
.
REMARK 4.15. Note that it follows from the denition of D
1,2
that if F
n
D
1,2
for
n = 1, 2, . . . and F
n
F in L
2
() and
D
t
F
n

n
is convergent in L
2
([0, T] )
then
F D
1,2
and D
t
F = lim
n
D
t
F
n
.
Since an arbitrary F L
2
() can be represented by its chaos expansion
F() =

n=0
I
n
(f
n
); f
n


L
2
([0, T]
n
)
it is natural to ask if we can express the derivative of F (if it exists) by means of this.
We rst consider a special case:
LEMMA 4.16. Suppose F() = I
n
(f
n
) for some f
n


L
2
([0, T]
n
). Then F D
1,2
and
(4.33) D
t
F() = nI
n1
(f
n
(, t)),
where the notation I
n1
(f
n
(, t)) means that the (n1)-iterated Ito integral is taken with
respect to the n1 rst variables t
1
, . . . , t
n1
of f
n
(t
1
, . . . , t
n1
, t) (i.e. t is xed and kept
outside the integration).
Proof. First consider the special case when
f
n
= f
n
for some f L
2
([0, T]), i.e. when
f
n
(t
1
, . . . , t
n
) = f(t
1
) . . . f(t
n
); (t
1
, . . . , t
n
) [0, T]
n
.
Then by (1.14)
(4.34) I
n
(f
n
) = |f|
n
h
n
(

|f|
),
where =
T
_
0
fdW and h
n
is the Hermite polynomial of order n.
Hence by the chain rule
D
t
I
n
(f
n
) = |f|
n
h
t
n
(

|f|
)
f(t)
|f|
A basic property of the Hermite polynomials is that
(4.35) h
t
n
(x) = nh
n1
(x).
4.10

This gives, again by (1.14),
D
t
I
n
(f
n
) = n|f|
n1
h
n1
(

|f|
)f(t) = nI
n1
(f
(n1)
)f(t) = nI
n1
(f
n
(, t)).
Next, suppose f
n
has the form
(4.36) f
n
=
n
1
1

n
2
2

n
k
k
; n
1
+ +n
k
= n
where

denotes symmetrized tensor product and
j
is an orthonormal basis for
L
2
([0, T]). Then by an extension of (1.14) we have (see [I])
(4.37) I
n
(f
n
) = h
n
1
(
1
) h
n
k
(
k
)
with

j
=
T
_
0

j
dW
and again (4.33) follows by the chain rule. Since any f
n


L
2
([0, T]
n
) can be approximated
in L
2
([0, T]
n
) by linear combinations of functions of the form given by the right hand side
of (4.36), the general result follows. .
LEMMA 4.17 Let P
0
denote the set of Wiener polynomials of the form
p
k
(
T
_
0
e
1
dW, . . . ,
T
_
0
e
k
dW)
where p
k
(x
1
, . . . , x
k
) is an arbitrary polynomial in k variables and e
1
, e
2
, . . . is a given
orthonormal basis for L
2
([0, T]). Then P
0
is dense in P in the norm | |
1,2
.
Proof. If q: = p(
T
_
0
f
1
dW, . . . ,
T
_
0
f
k
dW) P we approximate q by
q
(m)
: = p(
T
_
0
m

j=0
(f
1
, e
j
)e
j
dW, . . . ,
T
_
0
m

j=0
(f
k
, e
j
)e
j
dW)
Then q
(m)
q in L
2
() and
D
t
q
(m)
=
k

i=1
p
x
i

j=1
(f
i
, e
j
)e
j
(t)
k

i=1
p
x
i
f
i
(t)
in L
2
([0, T] ) as m . .
THEOREM 4.18. Let F =

n=0
I
n
(f
n
) L
2
(). Then F D
1,2
if and only if
(4.38)

n=1
nn!|f
n
|
2
L
2
([0,T]
n
)
<
4.11

and if this is the case we have
(4.39) D
t
F =

n=0
nI
n1
(f
n
(, t)).
Proof. Dene F
m
=
m

n=0
I
n
(f
n
). Then F
m
D
1,2
and F
m
F in L
2
(). Moreover, if
m > k we have by Lemma 4.15,
|D
t
F
m
D
t
F
k
|
2
L
2
([0,T])
= |
m

n=k+1
nI
n1
(f
n
(, t))|
2
L
2
([0,T])
=
T
_
0
E[
m

n=k+1
nI
n1
(f
n
(, t))
2
]dt
=
_
T
0
m

n=k+1
n
2
(n 1)!|f
n
(, t)|
2
L
2
([0,T]
n1
)
dt
=
m

n=k+1
nn!|f
n
|
2
L
2
([0,T]
n
)
. (4.40)
Hence if (4.38) holds then D
t
F
n

n=1
is convergent in L
2
([0, T] ) and hence F D
1,2
and
D
t
F = lim
m
D
t
F
m
=

n=0
nI
n1
(f
n
(, t)).
Conversely, if F D
1,2
then there exist polynomials p
k
(x
1
, . . . , x
n
k
) of degree k and
1
, . . .,

n
k
0 as in (4.36) such that if we put F
k
= p
k
(
1
, . . . ,
n
k
) =

m
i
:

m
i
k
a
m
1
,...,m
n
k
n
k

i=1
h
m
i
(
i
)
(for some a
m
1
,...,m
n
k
R) then F
k
P and F
k
F in L
2
() and
D
t
F
k
D
t
F in L
2
([0, T] ), as k .
By applying (4.37) we see that there exist f
(k)
j


L
2
([0, T]
j
); 1 j k such that
F
k
=
k

j=0
I
j
(f
(k)
j
).
Since F
k
F in L
2
() we have
k

j=0
j!|f
(k)
j
f
j
|
2
L
2
([0,T]
j
)
|F
k
F|
2
L
2
()
0 as k .
Therefore |f
(k)
j
f
j
|
L
2
([0,T]
j
)
0 as k , for all j. This implies that
(4.41) |f
(k)
j
|
L
2
([0,T]
j
)
|f
j
|
L
2
([0,T]
j
)
as k , for all j.
Similarly, since D
t
F
k
D
t
F in L
2
([0, T] ) we get by the Fatou lemma combined with
the calculation leading to (4.40) that

j=0
j j!|f
j
|
2
L
2
([0,T]
j
)
=

j=0
lim
k
(j j!|f
(k)
j
|
2
L
2
([0,T]
j
)
)
4.12

lim
k

j=0
j j!|f
(k)
j
|
2
L
2
([0,T]
j
)
= lim
k
|D
t
F
k
|
2
L
2
([0,T])
= |D
t
F|
2
L
2
([0,T])
< ,
where we have put f
(k)
j
= 0 for j > k. Hence (4.38) holds and the proof is complete. .
Exercises
4.1 Find the Malliavin derivative D
t
F() of the following random variables:
a) F() = W(T)
b) F() = exp(W(t
0
)) (t
0
[0, T])
c) F() =
T
_
0
s
2
dW(s)
d) F() =
T
_
0
(
t
2
_
0
cos(t
1
+t
2
)dW(t
1
))dW(t
2
)
e) F() = 3W(s
0
)W
2
(t
0
) + ln(1 +W
2
(s
0
)) (s
0
, t
0
[0, T])
f) F() =
T
_
0
W(t
0
)W(t) (t
0
[0, T])
(Hint: Use Exercise 2.1b).)
4.2 a) Find the Malliavin derivative D
t
F() when
F() = exp(
T
_
0
g(s)dW(s)) (g L
2
([0, T]))
by using that (see Exercise 1.2d))
F() =

n=0
I
n
[f
n
],
with
f
n
(t
1
, . . . , t
n
) =
1
n!
exp(
1
2
|g|
2
L
2
([0,T])
)g(t
1
) . . . g(t
n
)
b) Verify the result by using the chain rule.
4.3 Verify the integration by parts formula (4.28) in the following case:
F() =
T
_
0
(s)dW(s) with L
2
([0, T]) deterministic,
1.
4.13

5 The Clark-Ocone formula and its generalization.
Application to nance
In this section we look at the connection between dierentiation D and Skorohod inte-
gration and apply this to prove the Clark-Ocone formula and its generalization needed
for e.g. portfolio applications.
First we establish some useful results about conditional expectation:
DEFINITION 5.1. Let G be a Borel subset of [0, T]. Then we dene T
G
to be the
-algebra generated by all random variables of the form
_
A
dW(t): =
T
_
0
A
A
(t)dW(t); A G Borel set (5.1)
Thus if G = [0, t] we have, with this notation
T
[0,t]
= T
t
for t 0.
LEMMA 5.2. Let g L
2
([0, T]) be deterministic. Then
E[
_
T
0
g(t)dW(t)[T
G
] =
T
_
0
A
G
(t)g(t)dW(t).
Proof. By denition of conditional expectation, it suces to verify that
(5.2)
T
_
0
A
G
(t)g(t)dW(t) is T
G
-measurable
and
(5.3) E[F()
T
_
0
g(t)dW(t) = E[F()
T
_
0
A
G
(t)g(t)dW(t)]
for all bounded T
G
-measurable random variables F.
To prove (5.2) we may assume that g is continuous, because the continuous functions are
dense in L
2
([0, T]). If g is continuous, then
T
_
0
A
G
(t)g(t)dW(t) = lim
t
i
0

i
g(t
i
)
t
i
+1
_
t
i
A
G
(t)dW(t)
(limit in L
2
()) and since each term in the sum is T
G
-measurable the sum and its limit
is.
5.1

To prove (5.2) we may assume that
F() =
T
_
0
A
A
(t)dW(t) for some A G.
This gives that the left hand side of (5.3) becomes
E[
T
_
0
A
A
(t)dW(t)
T
_
0
g(t)dW(t)] = E[
T
_
0
A
A
(t)g(t)dt],
by the Ito isometry. Similarly, the right hand side becomes
E[
T
_
0
A
A
(t)dW(t)
T
_
0
A
G
(t)g(t)dW(t)] = E[
T
_
0
A
A
(t)A
G
(t)g(t)dt]
= E[
T
_
0
A
A
(t)g(t)dt] since A G.
.
LEMMA 5.3. Let v(t, ) R be a stochastic process such that
(i) v(t, ) is T
t
T
G
-measurable for all t and
(ii) E[
T
_
0
v
2
(t, )dt] < .
Then _
G
v(t, )dW(t) is T
G
-measurable.
Proof. By a standard approximation procedure we see that we may assume that v(t, )
is an elementary process of the form
v(t, ) =

i
v
i
()A
[t
i
,t
i+1
)
(t)
where 0 = t
0
< t
1
< < t
n
= T and v
i
() is T
t
i
T
G
-measurable. For such v we have
_
G
v(t, )dW(t) =

i
v
i
()
_
G[t
i
,t
i+1
)
dW(t),
which is a sum of products of T
G
-measurable functions and hence T
G
-measurable. .
LEMMA 5.4. Let u(t, ) be an T
t
-adapted process such that
E[
T
_
0
u
2
(t, )dt] < .
5.2

Then
E[
T
_
0
u(t, )dW(t)[T
G
] =
_
G
E[u(t, )[T
G
]dW(t)
Proof. By Lemma 5.3 it suces to verify that
(5.4) E[f()
T
_
0
u(t, )dW(t)] = E[f()
_
G
E[u(t, )[T
G
]dW(t)]
for all f() of the form
f() =
_
A
dW(t); A G
For such f we obtain by the Ito isometry that the left hand side of (5.4) is equal to
E[
T
_
0
A
A
(t)u(t, )dt] =
_
A
E[u(t, )]dt
while the right hand side is equal to
E[
T
_
0
A
A
(t)A
G
(t)E[u(t, )[T
G
]dt] =
T
_
0
A
A
(t)E[E[u(t, )[T
G
]]dt
=
_
A
E[u(t, )]dt .
.
PROPOSITION 5.5. Let f
n


L
2
([0, T]
n
). Then
(5.5) E[I
n
(f
n
)[T
G
] = I
n
[f
n
A
n
G
],
where
(f
n
A
n
G
)(t
1
, . . . , t
n
) = f
n
(t
1
, . . . , t
n
)A
G
(t
1
) A
G
(t
n
).
Proof. We proceed by induction on n. For n = 1 we have by Lemma 5.4
E[I
1
(f
1
)[T
G
] = E[
T
_
0
f
1
(t
1
)dW(t
1
)[T
G
] =
T
_
0
f
1
(t
1
)A
G
(t
1
)dW(t).
Assume that (5.5) holds for n = k. Then by Lemma 5.4
E[I
k+1
(f
k+1
)[T
G
]
= (k + 1)!E[
T
_
0

t
k
_
0

t
2
_
0
f
k+1
(t
1
, . . . , t
k+1
)dW(t
1
) dW(t
k+1
)[T
G
]
= (k + 1)!
T
_
0
E[
t
k
_
0

t
2
_
0
f
k+1
(t
1
, . . . , t
k+1
)dW(t
1
) dW(t
k
)[T
G
] A
G
(t
k+1
)dW(t
k+1
)
5.3

= (k + 1)!
T
_
0
t
k
_
0

t
2
_
0
f
k+1
(t
1
, . . . , t
k+1
)A
G
(t
1
) A
G
(t
k+1
)dW(t
1
) dW(t
k+1
)
= I
k+1
[f
k+1
A
(k+1)
G
],
and the proof is complete. .
PROPOSITION 5.6. If F D
1,2
then E[F[T
G
] D
1,2
and
D
t
(E[F[T
G
]) = E[D
t
F[T
G
] A
G
(t).
Proof. First asume that F = I
n
(f
n
) for some f
n


L
2
([0, T]
n
). Then by Proposition 5.5
D
t
(E[F[T
G
]) = D
t
E[I
n
(f
n
)[T
G
]
= D
t
[I
n
(f
n
A
n
G
)] = nI
n1
[f
n
(, t)A
(n1)
G
() A
G
(t)]
= nI
n1
[f
n
(, t)A
(n1)
G
()] A
G
(t)
= E[D
t
F[T
G
] A
G
(t). (5.6)
Next, suppose F D
1,2
is arbitrary. Then as in the proof of Theorem 4.16 we see that we
can nd F
k
P such that
F
k
F in L
2
() and D
t
F
k
D
t
F in L
2
( [0, T])
as k , and there exists f
(k)
j


L
2
([0, T]
j
) such that
F
k
=
k

j=0
I
j
(f
(k)
j
) for all k.
By (5.6) we have
D
t
(E[F
k
[T
G
]) = E[D
t
F
k
[T
G
] A
G
(t) for all k
and taking the limit of this as k we obtain the result. .
COROLLARY 5.7. Let u(s, ) be an T
s
-adapted stochastic process and assume that
u(s, ) D
1,2
for all s. Then
(i) D
t
u(s, ) is T
s
-adapted for all t
and
(ii) D
t
u(s, ) = 0 for t > s.
Proof. By Proposition 5.6 we have that
D
t
u(s, ) = D
t
(E[u(s, )[T
s
]) = E[D
t
u(s, )[T
s
] A
[0,s]
(t),
from which (i) and (ii) follow immediately. .
We now have all the necessary ingredients for our rst main result in this section:
5.4

THEOREM 5.8. (The Clark-Ocone formula) Let F D
1,2
be T
T
-measurable.
Then
(5.7) F() = E[F] +
T
_
0
E[D
t
F[T
t
]()dW(t).
Proof. Write F =

n=0
I
n
(f
n
) with f
n


L
2
([0, T]
n
). Then by Theorem 4.16, Proposition
5.5 and Denition 2.2
T
_
0
E[D
t
F[T
t
]dW(t) =
T
_
0
E[

n=1
nI
n1
(f
n
(, t))[T
t
]dW(t) =
T
_
0

n=1
nE[I
n1
(f
n
(, t))[T
t
]dW(t)
=
T
_
0

n=1
nI
n1
[f
n
(, t) A
(n1)
[0,t]
()]dW(t) =
T
_
0

n=1
n(n 1)!J
n1
[f
n
(, t)A
(n1)
[0,t]
]dW(t)
=

n=1
n!J
n
[f
n
()] =

n=1
I
n
[f
n
] =

n=0
I
n
[f
n
] I
0
[f
0
] = F E[F].
.
The generalized Clark-Ocone formula
We proceed to prove the generalized Clark-Ocone formula. This formula expresses an
T
T
-measurable random variable F() as a stochastic integral with respect to a process of
the form
(5.8)

W(t, ) =
t
_
0
(s, )ds +W(t, ); 0 t T
where (s, ) is a given T
s
-adapted stochastic process satisfying some additional condi-
tions. By the Girsanov theorem (see Exercise 5.1) the process

W(t) =

W(t, ) is a Wiener
process under the new probability measure Q dened on T
T
by
(5.9) dQ() = Z(T, )dP()
where
(5.10) Z(t) = Z(t, ) = exp
t
_
0
(s, )dW(s)
1
2
t
_
0

2
(s, )ds; 0 t T
We let E
Q
denote expectation w.r.t. Q, while E
P
= E denotes expectation w.r.t. P.
THEOREM 5.9. (The generalized Clark-Ocone formula [KO])
Suppose F D
1,2
is T
t
-measurable and that
E
Q
[[F[] < (5.11)
E
Q
[
T
_
0
[D
t
F[
2
dt] < (5.12)
E
Q
[[F[
T
_
0
(
T
_
0
D
t
(s, )dW(s) +
T
_
0
D
t
(s, )(s, )ds)
2
dt] < (5.13)
5.5

Then
F() = E
Q
[F] +
T
_
0
E
Q
[(D
t
F F
T
_
t
D
t
(s, )d

W(s))[T
t
]d

W(t). (5.14)
Remark. Note that we cannot obtain a representation as an integral w.r.t.

W simply by
applying the Clark-Ocone formula to our new Wiener process (

W(t), Q), because F is only

assumed to be T
T
-measurable, not

T
T
-measurable, where

T
T
is the -algebra generated
by

W(t, ); t T. In general we have

T
T
T
T
and usually

T
T
,= T
T
. Nevertheless, the
Ito integral w.r.t.

W in (5.14) does make sense, because

W(t) is a martingale w.r.t. T
t
and Q (see Exercise 5.1)).
The proof of Theorem 5.9 is split up into several useful results of independent interest:
LEMMA 5.10. Let and be two probability measures on a measurable space (, ()
such that d() = f()d() for some f L
1
(). Let X be a random variable on (, ()
such that X L
1
(). Let H ( be a -algebra. Then
(5.15) E

[X[H] E

[f[H] = E

[fX[H]
Proof. See e.g. [, Lemma 8.24].
COROLLARY 5.11. Suppose G L
1
(Q). Then
(5.16) E
Q
[G[T
t
] =
E[Z(T)G[T
t
]
Z(t)
The next result gives a useful connection between dierentiation and Skorohod integration:
THEOREM 5.12. Let u(s, ) be a stochastic process such that
(5.17) E[
T
_
0
u
2
(s, )ds] <
and assume that u(s, ) D
1,2
for all s [0, T], that D
t
u Dom() for all t [0, T], and
that
(5.18) E[
T
_
0
((D
t
u))
2
dt] <
Then
T
_
0
u(s, )W(s) D
1,2
and
(5.19) D
t
(
_
T
0
u(s, )W(s)) =
T
_
0
D
t
u(s, )W(s) +u(t, ).
Proof. First assume that
u(s, ) = I
n
(f
n
(, s)),
5.6

where f
n
(t
1
, . . . , t
n
, s) is symmetric with respect to t
1
, . . . , t
n
. Then
T
_
0
u(s, )W(s) = I
n+1
[

f
n
],
where

f
n
(x
1
, . . . , x
n+1
) =
1
n + 1
[f
n
(, x
1
) + +f
n
(, x
n+1
)]
is the symmetrization of f
n
as a function of all its n + 1 variables. Hence
(5.20) D
t
(
T
_
0
u(s, )W(s)) = (n + 1)I
n
[

f
n
(, t)],
where
(5.21)

f
n
(, t) =
1
n + 1
[f
n
(t
,
, x
1
) + +f
n
(t, , x
n
) +f
n
(, t)]
(since f
n
is symmetric w.r.t. its rst n variables, we may choose t to be the rst of them,
in the rst n terms on the right hand side). Combining (5.20) with (5.21) we get
(5.22) D
t
(
T
_
0
u(s, )W(s)) = I
n
[f
n
(t, , x
1
) + +f
n
(t, , x
n
) +f
n
(, t)]
(integration in I
n
is w.r.t. x
1
, . . . , x
n
).
To compare this with the right hand side of (5.19) we consider
(5.23) (D
t
u) =
T
_
0
D
t
u(s, )W(S) =
T
_
0
nI
n1
[f
n
(, t, s)]W(s) = nI
n
[

f
n
(, t, )],
where
(5.24)

f
n
(x
1
, . . . , x
n1
, t, x
n
) =
1
n
[f
n
(t, , x
1
) + +f
n
(t, , x
n
)]
is the symmetrization of f
n
(x
1
, . . . , x
n1
, t, x
n
) w.r.t. x
1
, . . . , x
n
.
From (5.23) and (5.24) we get
(5.25)
T
_
0
D
t
u(s, )W(s) = I
n
[f
n
(t, , x
1
) + +f
n
(t, , x
n
)].
Comparing (5.22) and (5.25) we obtain (5.19).
Next, consider the general case when
u(s, ) =

n=0
I
n
[f
n
(, s)].
5.7

Dene
(5.26) u
m
(s, ) =
m

n=0
I
n
[f
n
(, s)]; m = 1, 2, . . .
Then by the above we have
(5.27) D
t
((u
m
)) = (D
t
u
m
) +u
m
(t) for all m.
By (5.23) we see that (5.18) is equivalent to saying that
E[
T
_
0
((D
t
u))
2
dt] =

n=0
n
2
n!
T
_
0
|

f
n
(, t, )|
2
L
2
([0,T]
n
)
dt
=

n=0
n
2
n!|

f
n
|
2
L
2
([0,T]
n+1
)
< since D
t
u Dom(). (5.28)
Hence
(5.29) |(D
t
u
m
) (D
t
u)|
2
L
2
([0,T])
=

n=m+1
n
2
n!|

f
n
|
2
L
2
([0,T]
n+1
)
0 as m .
Therefore, by (5.27)
(5.30) D
t
((u
m
)) (D
t
u) +u(t) in L
2
([0, T] )
as m . Note that from (5.21) and (5.24) we have
(n + 1)

f
n
(, t) = n

f
n
(, t, ) +f
n
(, t)
and hence
(n + 1)!|

f
n
|
2
L
2
([0,T]
n+1
)

2n
2
n!
n + 1
|

f
n
|
2
L
2
([0,T]
n+1
)
+
2n!
n + 1
|f
n
|
2
L
2
([0,T]
n+1
)
Therefore,
(5.31) E[((u
m
) (u))
2
] =

n=m+1
(n + 1)!|

f
n
|
2
L
2
([0,T]
n+1
)
0
as m . From (5.30) and (5.31) we conclude that (u) D
1,2
and that
D
t
((u)) = (D
t
u) +u(t), which is (5.19).
.
COROLLARY 5.13. Let u(s, ) be as in Theorem 5.12 and assume in addition that
u(s, ) is T
s
-adapted.
Then
(5.32) D
t
(
T
_
0
u(s, )dW(s)) =
T
_
t
D
t
u(s, )dW(s) +u(t, ).
5.8

Proof. This is an immediate consequence of Theorem 5.12 and Corollary 5.7. .
LEMMA 5.14. Let F, be as in Theorem 5.9 and let Q and Z(t) be as in (5.9), (5.10).
Then
(5.33) D
t
(Z(T)F) = Z(T)[D
t
F F(t, ) +
T
_
t
D
t
(s, )d

W(s)].
Proof. By the chain rule we have, using Corollary 5.13,
D
t
(Z(T)F) = Z(T)D
t
F +F D
t
Z(T) and
D
t
Z(T) = Z(T)D
t
(
T
_
0
(s, )dW(s))
1
2
D
t
(
T
_
0

2
(s, )ds)
= Z(T)
T
_
t
D
t
(s, )dW(s) (t, )
T
_
0
(s, )D
t
(s, )ds
= Z(T)
T
_
t
D
t
(s, )d

W(s) (t, ).
.
Proof of Theorem 5.9: Suppose that (5.11)(5.13) hold and put
(5.34) Y (t) = E
Q
[F[T
t
]
and
(5.35) (t) = Z
1
(t) = exp
t
_
0
(s, )dW(s) +
1
2
t
_
0

2
(s, )ds.
Note that
(5.36) (t) = exp
t
_
0
(s, )d

W(s)
1
2
t
_
0

2
(s, )ds.
By Corollary 5.11, Theorem 5.8 and Corollary 5.7 we can write
Y
t
= (t)E[Z(T)F[T
t
]
= (t)E[E[Z(T)F[T
t
]] +
T
_
0
E[D
s
E[Z(T)F[T
t
][T
s
]dW(s)
= (t)E[Z(T)F] +
t
_
0
E[D
s
(Z(T)F)[T
s
]dW(s)
=: (t)U(t). (5.37)
By (5.36) and the Ito formula we have
d(t) = (t)(t)d

W(t) (5.38)
5.9

Combining (5.37), (5.38) and (5.33) we get
dY (t) = (t) E[D
t
(Z(T)F)[T
t
]dW(t)
+(t)(t)U(t)d

W(t)
+(t)(t)E[D
t
(Z(T)F)[T
t
]dW(t)d

W(t)
= (t)E[D
t
(Z(T)F)[T
t
]d

W(t) +(t)Y (t)d

W(t)
= (t)E[Z(T)D
t
F[T
t
] E[Z(T)F(t)[T
t
]
E[Z(T)F
T
_
t
D
t
(s)d

W(s)[T
t
]d

W(t) +(t)Y (t)d

W(t)
Hence
dY (t) = E
Q
[D
t
F[T
t
] E
Q
[F(t)[T
t
] E
Q
[F
T
_
t
D
t
(s)d

W(s)[T
t
]d

W(t)
+(t)E
Q
[F[T
t
]d

W(t)
= E
Q
[(D
t
F F
T
_
t
D
t
(s)d

W(s))[T
t
]d

W(t). (5.39)
Since
Y (T) = E
Q
[F[T
T
] = F
and
Y (0) = E
Q
[F[T
0
] = E
Q
[F],
we see that Theorem 5.9 follows from (5.39). The conditions (5.11)(5.13) are needed to
make all the above operations valid. We omit the details. .
Application to nance
We end this section by explaining how the generalized Clark-Ocone theorem can be applied
in portfolio analysis:
Suppose we have two possible investments:
a) A safe investment (e.g. a bond), with price dynamics
(5.40) dA(t) = (t) A(t)dt
b) A risky investment (e.g. a stock), with price dynamics
(5.41) dS(t) = (t)S(t)dt +(t)S(t)dW(t)
5.10

Here (t) = (t, ), (t) = (t, ) and (t) = (t, ) are T
t
-adapted processes. In the
following we will not specify further conditions, but simply assume that these processes
are suciently nice to make the operations convergent and well-dened.
Let (t) = (t, ), (t) = (t, ) denote the number of units invested at time t in in-
vestments a), b), respectively. Then the value at time t, V (t) = V (t, ), of this portfolio
((t), (t)) is given by
(5.42) V (t) = (t)A(t) +(t)S(t)
The portfolio ((t), (t) is called self-nancing if
(5.43) dV (t) = (t)dA(t) +(t)dS(t).
Assume from now on that ((t), (t)) is self-nancing. Then by substituting
(5.44) (t) =
V (t) (t)S(t)
A(t)
from (5.42) in (5.43) and using (5.40) we get
(5.45) dV (t) = (t)(V (t) (t)S(t))dt +(t)dS(t).
Then by (5.41) this can be written
(5.46) dV (t) = [(t)V (t) + ((t) (t))(t)S(t)]dt +(t)(t)S(t)dW(t)
Suppose now that we are required to nd a portfolio ((t), (t)) which leads to a given
value
(5.47) V (T, ) = F() a.s.
at a given (deterministic) future time T, where the given F() is T
T
-measurable. Then
the problem is:
What initial fortune V (0) is needed to achieve this, and what portfolio ((t), (t)) should
we use? Is V (0) and ((t), (t)) unique?
This type of question appears in option pricing.
For example, in the classical Black-Scholes model we have
F() = (S(T, ) K)
+
where K is the exercise price and then V (0) is the price of the option.
Because of the relation (5.44) we see that we might as well consider (V (t), (t)) to be
the unknown T
t
-adapted processes. Then (5.46)(5.47) constitutes what is known as a
stochastic backward dierential equation (SBDE): The nal value V (T, ) is given and
one seeks the value of V (t), (t) for 0 t T. Note that since V (t) is T
t
-adapted, we
5.11

have that V (0) is T
0
-measurable and therefore a constant. The general theory of SBDE
gives that (under reasonable conditions on F, , and ) equation (5.46)(5.47) has a
unique solution of T
t
-adapted processe V (t), (t). See e.g. [PP]. However, this general
theory says little about how to nd this solution explicitly. This is where the generalized
Clark-Ocone theorem enters the scene:
Dene
(5.48) (t) = (t, ) =
(t) (t)
(t)
and put
(5.49)

W(t) =
_
t
0
(s)ds +W(t).
Then

W(t) is a Wiener process w.r.t. the measure Q dened by (5.9), (5.10). In terms of

W(t) equation (5.46) gets the form

dV (t) = [(t)V (t) + ((t) (t))(t)S(t)]dt +(t)(t)S(t)d

W(t)
(t)(t)S(t)
1
(t)((t) (t))dt
i.e.
(5.50) dV (t) = (t)V (t)dt +(t)(t)S(t)d

W(t).
Dene
(5.51) U(t) = e

_
t
0
(s,)ds
V (t).
Then, substituting in (5.50), we get
(5.52) dU(t) = e

_
t
0
ds
(t)(t)S(t)d

W(t)
or
(5.53) e

_
T
0
ds
V (T) = V (0) +
T
_
0
e

_
t
0
ds
(t)(t)S(t)d

W(t).
By the generalized Clark-Ocone theorem applied to
(5.54) G(): = e

_
T
0
(s,)ds
F()
we get
(5.55) G() = E
Q
[G] +
T
_
0
E
Q
[(D
t
GG
T
_
t
D
t
(s, )d

W(s))[T
t
]d

W,
By uniqueness we conclude from (5.53) and (5.55) that
(5.56) V (0) = E
Q
[G]
5.12

and the required risky investment at time t is
(5.57) (t) = e
_
t
0
ds

1
(t)S
1
(t)E
Q
[(D
t
GG
T
_
t
D
t
(s)d

W(s))[T
t
].
EXAMPLE 5.14. Suppose (t, ) = , (t, ) = and (t, ) = ,= 0 are constants.
Then
(t, ) = =

is constant also and hence D

t
= 0. Therefore by (5.57)
(t) = e
(tT)

1
S
1
(t)E
Q
[D
t
F[T
t
].
For example, if the payo function is
F() = exp(W(T)) ( ,= 0 constant)
then by the chain rule we get
(t) = e
(tT)

1
S
1
(t)E
Q
[exp(W(T))[T
t
]
= e
(tT)

1
S
1
(t)Z
1
(t)E[Z(T) exp(W(T))[T
t
]. (5.58)
Note that
Z(T) exp(W(T)) = M(T) exp(
1
2
( )
2
T)
where M(t): = exp( )W(t)
1
2
( )
2
t) is a martingale. This gives
(t) = e
(tT)

1
S
1
(t)Z
1
(t)M(t) exp(
1
2
( )
2
T)
= e
(tT)

1
exp( )W(t) + (
1
2

2
+
1
2

2
)t +
1
2
( )
2
(T t).
EXAMPLE 5.15 (The Black and Scholes formula)
Finally, let us illustrate the method above by using it to prove the celebrated Black and
Scholes formula (see e.g. [Du]). As in Example 5.14 let us assume that (t, ) = ,
(t, ) = and (t, ) = ,= 0 are constants. Then
=

is constant and hence D

t
= 0. Hence
(5.59) (t) = e
(tT)

1
S
1
(t)E
Q
[D
t
F [ T
t
]
as in Example 5.14. However, in this case F() represents the payo at time T (xed)
of a (European call) option which gives the owner the right to buy the stock with value
5.13

S(T, ) at a xed exercise price K, say. Thus if S(T, ) > K the owner of the option gets
the prot S(T, ) K and if S(T, ) K the owner does not exercise the option and the
prot is 0. Hence in this case
(5.60) F() = (S(T, ) K)
+
.
Thus we may write
(5.61) F() = f(S(T, ))
where
(5.62) f(x) = (x K)
+
.
The function f is not dierentiable at x = K, so we cannot use the chain rule directly to
evaluate D
t
G from (5.61). However, we can approximate f by C
1
functions f
n
with the
property that
(5.63) f
n
(x) = f(x) for [x K[
1
n
and
(5.64) 0 f
t
n
(x) 1 for all x.
Putting
F
n
() = f
n
(S(T, ))
we then see that
D
t
F() = lim
n
D
t
F
n
() = A
[K,]
(S(T, ))D
t
S(T, )
= A
[K,]
(S(T, )) S(T, ) (5.65)
Hence by (5.59)
(5.66) (t) = e
(tT)
S
1
(t)E
Q
[S(T) A
[K,]
(S(T))[T
t
]
By the Markov property of S(t) this is the same as
(5.66) (t) = e
(tT)
S
1
(t)E
y
Q
[S(T t) A
[K,]
(S(T t))]
y=S(t)
where E
y
Q
is the expectation when S(0) = y. Since
dS(t) = S(t)dt +S(t)dW(t)
= ( )S(t)dt +S(t)d

W(t)
= S(t)dt +S(t)d

W(t),
we have
(5.67) S(t) = S(0) exp((
1
2

2
)t +

W(t))
5.14

and hence
(5.68) (t) = e
(tT)
S
1
(t)E
y
[Y (T t)A
[K,]
(Y (T t))]
y=S(t)
,
where
(5.69) Y (t) = S(0) exp((
1
2

2
)t +W(t)) .
Since the distribution of W(t) is well-known, we can express the solution (5.68) explicitly
in terms of quantities involving S(t) and the normal distribution function.
In this model (t) represents the number of units we must invest in the risky investment
at times t T in order to be guaranteed to get the payo F() = (S(T, )K)
+
(a.s.) at
time T. The constant V (0) represents the corresponding initial fortune needed to achieve
this. Thus V (0) is the (unique) initial fortune which makes it possible to establish a
(self-nancing) portfolio with the same payo at time T as the option gives. Hence V (0)
deserves to be called the right price for such an option. By (5.56) this is given by
V (0) = E
Q
[e
T
F()] = e
T
E
Q
[(S(T) K)
+
]
= e
T
E[(Y (T) K)
+
] , (5.70)
which again can be expressed explicitly by the normal distribution function.
Final remarks In the Markovian case, i.e. when the price S(t) is given by a stochastic
dierential equation of the form
dS(t) = (S(t))S(t)dt +(S(t))S(t)dW(t)
where : R R and : R R are given functions, then there is a well-known alternative
method for nding the option price V (0) and the corresponding replicating portfolio (t):
One assumes that the value process has the form
V (t, ) = f(t, S(t, ))
for some function f: R
2
R and deduces a (deterministic) partial dierential equation
which determines f. Then is given by
(t, ) =
_
f(t, x)
x
_
x=S(t,)
.
However, the method does not work in the non-markovian case. The method based on the
Clark-Ocone formula has the advantage that it does not depend on a Markovian setup.
Exercises
5.1 Recall the Girsanov theorem (see e.g. [1], Th. 8.26): Let Y (t) R
n
be an Ito
process of the form
(5.71) dY (t) = (t, )dt +(t, )dW(t); t T
5.15

where (t, ) R
n
, (t, ) R
nm
are T
t
-adapted and W(t) is m-dimensional. Suppose
there exist T
t
-adapted processes (t, ) R
m
and (t, ) R
n
such that
(5.72) (t, )(t, ) = (t, ) (t, )
and such that Novikovs condition
(5.73) E[exp(
1
2
T
_
0

2
(s, )ds)] <
holds. Put
(5.74) Z(t, ) = exp(
t
_
0
(s, )dW(s)
1
2
t
_
0

2
(s, )ds); t T
and dene a measure Q on T
T
by
(5.75) dQ() = Z(T, )dP() on T
T
Then
(5.76)

W(t, ): =
t
_
0
(s, ) +W(t, ); 0 t T
is a Wiener process w.r.t. Q, and in terms of

W(t, ) the process Y (t, ) has the stochastic
integral representation
(5.77) dY (t) = (t, )dt +(t, )d

W(t).
a) Show that

W(t) is an T
t
-martingale w.r.t. Q.
(Hint: Apply Itos formula to Y (t): = Z(t)

W(t).)
b) Suppose X(t) = at + W(t) R, t T, where a R is a constant. Find a
probability measure Q on T
T
such that X(t) is a Wiener process w.r.t. Q.
c) Let a, b, c ,= 0 be real constants and dene
dY (t) = bY (t)dt +cY (t)dW(t).
Find a probability measure Q and a Wiener process

W(t) w.r.t. Q such that
dY (t) = aY (t)dt +cY (t)d

W(t).
5.2 Verify the Clark-Ocone formula
F() = E[F] +
T
_
0
E[D
t
F [ T
t
]dW(t)
for the following T
T
-measurable random variables F
5.16

a) F() = W(T)
b) F() =
T
_
0
W(s)ds
c) F() = W
2
(T)
d) F() = W
3
(T)
e) F() = exp W(T)
f) F() = (W(T) +T) exp(W(T)
1
2
T)
5.3 Let

W(t) =
t
_
0
(s, )ds + W(t) and Q be as in Exercise 5.1. Use the generalized
Clark-Ocone formula to nd the T
t
-adapted process (t, ) such that
F() = E
Q
[F] +
T
_
0
(t, )d

W(t)
in the following cases:
a) F() = W
2
(T), (s, ) = (s) is deterministic
b) F() = exp(
T
_
0
(s)dW(s)), (s) and (s) are deterministic.
c) F() like in b), (s, ) = W(s).
5.4 Suppose we have the choice between the investments (5.40), (5.41). Find the initial
fortune V (0) and the number of units (t, ) which must be invested at time t in the
risky investment in order to produce the terminal value V (T, ) = F() = W(T, ) when
(t, ) = > 0 (constant) and the price S(t) of the risky investment is given by
a) dS(t) = S(t)dt +S(t)dW(t); , constants ( ,= 0)
b) dS(t) = cdW(t); c ,= 0 constant
c) dS(t) = S(t)dt +cdW(t); , c constants (the Ornstein-Uhlenbeck process).
Hint:
S(t) = e
t
S(s) +c
T
_
0
e
(ts)
dW(s).
5.17

Bibliography
[B] F.E. Benth. Integrals in the Hida distribution space (o)

. In T. Lindstrm,
B. ksendal and A. S. Ustunel (editors). Stochastic Analysis and Related Top-
ics. Gordon & Breach 1993, pp. 89-99.
[D1] E. B. Dynkin: Markov Processes I. Springer 1965.
[D2] E. B. Dynkin: Markov Processes II. Springer 1965.
[Du] D. Due: Dynamic Asset Pricing Theory. Princeton University Press 1992.
[GHLUZ] H. Gjessing, H. Holden, T. Lindstrm, J. Ube and T. Zhang. The Wick
product. In H. Niemi, G. H ogn as, A.N. Shiryaev and A. Melnikov (editors).
Frontiers in Pure and Applied Probability, Vol. 1. TVP Publishers, Moscow,
1993, pp. 29-67.
[GV] I.M. Gelfand and N.Y. Vilenkin. Generalized Functions, Vol. 4. Applications of
Harmonic Analysis. Academic Press 1964 (English translation).
[H] T. Hida. Brownian Motion. Springer-Verlag 1980.
[Hu] Y. Hu. Ito-Wiener chaos expansion with exact residual and correlation, variance
inequalities. Manuscript 1995.
[HKPS] T. Hida, H.-H. Kuo, J. Pottho and L. Streit. White Noise Analysis. Kluwer
1993.
[HLUZ] H. Holden, T. Lindstrm, B. ksendal, J. Ube and T. Zhang. The pressure
equation for uid ow in a stochastic medium. Potential Analysis 4 1995, 655
674.
[HP] T. Hida and J. Pottho. White noise analysis an overview. In T. Hida, H.-H.
Kuo, J. Pottho and L. Streit (eds.). White Noise Analysis. World Scientic
1990.
[HUZ] H. Holden, B. ksendal, J. Ube and T. Zhang. Stochastic Partial Dierential
Equations. Birkhauser 1996.
[I] K. Ito: Multiple Wiener integral. J. Math. Soc. Japan 3 (1951), 157169.
5.18

[IW] N. Ikeda and S. Watanable. Stochastic Dierential Equations and Diusion
Processes (Second edition). North-Holland/Kodansha 1989.
[KO] I. Karatzas and D. Ocone: A generalized Clark representation formula, with
application to optimal portfolios. Stochastics and Stochastics Reports 34 (1991),
187220.
[LU 1] T. Lindstrm, B. ksendal and J. Ube. Stochastic dierential equations involv-
ing positive noise. In M. Barlow and N. Bingham (editors). Stochastic Analysis.
Cambridge Univ. Press 1991, pp. 261303.
[LU 2] T. Lindstrm, B. ksendal and J. Ube. Wick multiplication and Ito-Skorohod
stochastic dierential equations. In S. Albeverio et al (editors). Ideas and Meth-
ods in Mathematical Analysis, Stochastics, and Applications. Cambridge Univ.
Press 1992, pp. 183206.
[LU 3] T. Lindstrm, B. ksendal and J. Ube. Stochastic modelling of uid ow in
porous media. In S. Chen and J. Yong (editors). Control Theory, Stochastic
Analysis and Applications. World Scientic 1991, pp. 156172.
[N] D. Nualart: The Malliavin Calculus and Related Topics. Springer 1995.
[PP] E. Pardoux and S. Peng: Adapted solution of a backward stochastic dierential
equation. Systems & Control Letters 14 (1990), 5561.
[PT] J. Pottho and M. Timpel. On a dual pair of spaces of smooth and generalized
random variables. Preprint University of Mannheim 1993.
[RS] M. Reed and B. Simon. Methods of Modern Mathematical Physics, Vol. 1.
Academic Press 1972.
[U] A. S. Ustunel: An Introduction to Analysis on Wiener Space. Springer LNM
1610, 1995.
[1] B. ksendal. Stochastic Dierential Equations. Springer-Verlag 1995 (Fourth
edition).
[2] B. ksendal. Stochastic Partial Dierential Equations. A mathematical con-
nection between macrocosmos and microcosmos. In M. Gyllenberg and L.E.
Person (editors). Analysis, Algebra, and Computers in Mathematical Research.
Proceedings of the 21st Nordic Congress of Mathematicians. Marcel Dekker
1994, pp. 365385.
[Z] B. ksendal and T. Zhang. The stochastic Volterra equation. In D. Nualart
and M. Sanz Sole (editors). The Barcelona Seminar on Stochastic Analysis.
Birkh auser 1993, pp. 168202.
[Z] T. Zhang. Characterizations of white noise test functions and Hida distribu-
tions. Stochastics 41 (1992), 7187.
5.19

6 Solutions to the exercises
1.1. a) exp(tx
t
2
2
) = exp(
1
2
x
2
) exp(
1
2
(x t)
2
)
= exp(
1
2
x
2
)

n=0

1
n!
d
n
dt
n
(exp(
1
2
(x t)
2
))
t=0
t
n

(u = x t) = exp(
1
2
x
2
)

n=0

1
n!
d
n
du
n
(exp(
1
2
u
2
))
u=x
(1)
n
t
n

= exp(
1
2
x
2
)

n=0

1
n!
(1)
n
d
n
dx
n
(exp(
1
2
x
2
))t
n

n=0
t
n
n!
h
n
(x).
1.1. b) u = t

gives
exp(tx
t
2

2
) = exp(u
x

u
2
2
)
(by a)) =

n=0
u
n
n!
h
n
(
x

) =

n=0
t
n

n/2
n!
h
n
(
x

).
1.1. c) If we choose x = , = |g|
2
and t = 1 in b), we get
exp(
T
_
0
gdW
1
2
|g|
2
) =

n=0
|g|
n
n!
h
n
(

|g|
).
1.1. d) In particular, if we choose g(s) = A
[0,t]
(s), we get
exp(W(t)
1
2
t) =

n=0
t
n/2
n!
h
n
(
W(t)

t
).
1.2. a) () = W(t, ) =
T
_
0
A
[0,t]
(s)dW(s), so f
0
= 0, f
1
= A
[0,t]
and f
n
= 0 for n 2.
1.2. b) () =
T
_
0
g(s)dW(s) f
0
= 0, f
1
= g, f
n
= 0 for n 2
1.2. c) Since
t
_
0
(
t
2
_
0
1 dW(t
1
))dW(t
2
) =
t
_
0
W(t
2
)dW(t
2
) =
1
2
W
2
(t)
1
2
t ,
6.1

we get that
W
2
(t) = t + 2
t
_
0
t
2
_
0
dW(t
1
)dW(t
2
)
= t + 2
T
_
0
t
2
_
0
A
[0,t]
(t
1
)A
[0,t]
(t
2
)dW(t
1
)dW(t
2
) = t +I
2
[f
2
],
where
f
2
(t
1
, t
2
) = A
[0,t]
(t
1
)A
[0,t]
(t
2
) = A
2
[0,t]
.
So f
0
= t and f
n
= 0 for n ,= 2.
1.2. d) By Exercise 1.1c) and (1.14) we have
() = exp(
T
_
0
g(s)dW(s))
= exp(
1
2
|g|
2
)

n=0
|g|
n
n!
h
n
(

|g|
)
= exp(
1
2
|g|
2
)

n=0
J
n
[g
n
] =

n=0
1
n!
exp(
1
2
|g|
2
)I
n
[g
n
].
Hence
f
n
=
1
n!
exp(
1
2
|g|
2
)g
n
; n = 0, 1, 2, . . .
where
g
n
(x
1
, . . . , x
n
) = g(x
1
)g(x
2
) g(x
n
).
1.3. a) Since
T
_
0
W(t)dW(t) =
1
2
W
2
(T)
1
2
T, we have
F() = W
2
(T) = T + 2
T
_
0
W(t)dW(t).
Hence
E[F] = T and (t, ) = 2W(t).
1.3. b) Dene M(t) = exp(W(t)
1
2
t). Then by the Ito formula
dM(t) = M(t)dW(t)
and therefore
M(T) = 1 +
T
_
0
M(t)dW(t)
6.2

or
F() = exp W(T) = exp
1
2
T + exp(
1
2
T)
T
_
0
exp(W(t)
1
2
t)dW(t).
Hence
E[F] = exp(
1
2
T) and (t, ) = exp(W(t) +
1
2
(T t)).
1.3. c) Integration by parts gives
F() =
T
_
0
W(t)dt = TW(T)
T
_
0
tdW(t) =
T
_
0
(T t)dW(t).
Hence E[F] = 0 and (t, ) = T t.
1.3. d) By the Ito formula
d(W
3
(t)) = 3W
2
(t)dW(t) + 3W(t)dt
Hence
F() = W
3
(T) = 3
T
_
0
W
2
(t)dW(t) + 3
T
_
0
W(t)dt
Therefore, by 1.3.c) we get
E[F] = 0 and (t, ) = 3W
2
(t) + 3T(1 t).
1.3. e) Put X(t) = e
1
2
t
, Y (t) = cos W(t), N(t) = X(t)Y (t). Then
dN(t) = X(t)dY (t) +Y (t)dX(t) +dX(t)dY (t)
= e
1
2
t
[sin W(t)dW(t)
1
2
cos W(t)dt] + cos W(t) e
1
2
t

1
2
dt
= e
1
2
t
sin W(t)dW(t).
Hence
e
1
2
T
cos W(T) = 1
T
_
0
e
1
2
t
sin W(t)dW(t)
or
F() = cos W(T) = e

1
2
T
e

1
2
T
T
_
0
e
1
2
t
sin W(t)dW(t).
Hence E[F] = e

1
2
T
and (t, ) = e
1
2
(tT)
sin W(t).
6.3

1.4. a) By Itos formula and Kolmogorovs backward equation we have
dY (t) =
g
t
(t, X(t))dt +
g
x
(t, X(t))dX(t) +
1
2

2
g
x
2
(t, X(t))(dX(t))
2
=

t
[P
Tt
f()]
=X(t)
dt +(X(t))

[P
Tt
f()]
=X(t)
dW(t)
+b(X(t))

[P
Tt
f()]
=X(t)
+
1
2

2
(X(t))

2

2
[P
Tt
f()]
=X(t)
dt
=

t
[P
Tt
f()]
=X(t)
dt +(X(t))

[P
Tt
f()]
=X(t)
dW(t)
+

u
[P
u
f()] =X(t)
u=Tt
dt
= (X(t))

[P
Tt
f()]
=X(t)
dW(t).
Hence
Y (T) = Y (0) +
T
_
0
[(x)

P
Tt
f()]
=X(t)
dW(t).
Since Y (T) = g(T, X(T)) = [P
0
f()]
=X(T)
= f(X(T)) and Y (0) = g(0, X(0)) = P
T
f(X),
(1.48) follows.
1.4. b) If F() = W
2
(T) we apply a) to the case when f() =
2
and X(t) = x+W(t)
(assuming W(0) = 0 as before). This gives
P
s
f() = E

[f(X(x))] = E

[X
2
(s)] =
2
+s
and hence
E[F] = P
T
f(x) = x
2
+T
and
(t, ) = [

(
2
+s)]
=x+W(t)
= 2W(t) + 2x.
1.4. c) If F() = W
3
(T) we choose f() =
3
and X(t) = x +W(t) and get
P
s
f() = E

[X
3
(s)] =
3
+ 3s .
Hence
E[F] = P
T
f(x) = x
3
+ 3Tx
and
(t, ) = [

(
3
+ 3(T t))]
=x+W(t)
= 3(x +W(t))
2
+ 3(T t)
6.4

1.4. d) In this case f() = so
P
s
f() = E

[X(s)] = exp(s)
so
E[F] = P
T
f(x) = x exp(T)
and
(t, ) = [

exp((T t))]
=X(t)
= X(t) exp((T t)) = x exp(T
1
2

2
t +W(t)).
1.4. e) We proceed as in a) and put
Y (t) = g(t, X(t)) with g(t, x) = P
Tt
f(x)
and
dX(t) = b(X(t))dt +(X(t))dW(t) ; X
0
= x R
n
where
b: R
n
R
n
, : R
n
R
nm
and W(t) = (W
1
(t), . . . , W
m
(t))
is the m-dimensional Wiener process.
Then by Itos formula and (1.50) we have
dY (t) =
g
t
(t, X(t))dt +
n

i=1
g
x
i
(t, X(t))dX
i
(t)
+
1
2

i,j

2
g
x
i
x
j
(t, X(t))dX
i
(t)dX
j
(t)
=

t
[P
Tt
f()]
=X(t)
dt + [
T
()

(P
Tt
f())]
=X(t)
dW(t)
+[L

(P
Tt
f())]
=X(t)
dt
where
L

=
n

i=1
b
i
()

i
+
1
2
n

i,j=1
(
T
)
ij
()

2

j
is the generator of the Ito diusion X(t). So by the Kolmogorov backward equation we
get
dY (t) = [
T
()

(P
Tt
f()]
=X(t)
dW(t)
and hence, as in a),
Y (T) = f(X(T)) = P
T
f(x) +
T
_
0
[
T
()

(P
Tt
f())]
=X(t)
dW(t),
which gives, with F = f(X(T)),
E[F] = P
T
f(x) and (t, ) = [
T
()

(P
Tt
f())]
=X(t)
.
6.5

2.1. a) Since W(t) is T
t
-adapted, we have
T
_
0
W(t)W(t) =
T
_
0
W(t)dW(t) =
1
2
W
2
(T)
1
2
T.
2.1. b)
T
_
0
(
T
_
0
gdW)(W) =
T
_
0
I
1
[f
1
(t
1
, t)]W(t
1
) = I
2
[

f
1
] where
f
1
(t
1
, t) = g(t
1
).
This gives

f
1
(t
1
, t) =
1
2
(g(t
1
) +g(t))
and hence
I
2
[

f
1
] = 2
T
_
0
t
2
_
0

f
1
(t
1
, t
2
)dW(t
1
)dW(t
2
)
=
T
_
0
t
2
_
0
g(t
1
)dW(t
1
)dW(t
2
) +
T
_
0
t
2
_
0
g(t
2
)dW(t
1
)dW(t
2
). (6.1)
Using integration by parts (i.e. the Ito formula) we see that
T
_
0
t
2
_
0
g(t
1
)dW(t
1
)dW(t
2
) =
= (
T
_
0
gdW)W(T)
T
_
0
g(t)W(t)dW(t)
T
_
0
g(t)dt. (6.2)
Combining (6.1) and (6.2) we get
T
_
0
(
T
_
0
gdW)W = (
T
_
0
gdW) W(T)
T
_
0
g(s)ds.
2.1. c) By Exercise 1.2. c) we have
T
_
0
W
2
(t
0
)W(t) =
T
_
0
(t
0
+I
2
[f
2
])W(t), (6.3)
where
f
2
(t
1
, t
2
, t) = A
[0,t
0
]
(t
1
)A
[0,t
0
]
(t
2
).
6.6

Now

f
2
(t
1
, t
2
, t) =
1
3
[f
2
(t
1
, t
2
, t) +f
2
(t, t
2
, t
1
) +f
2
(t
1
, t, t
2
)]
=
1
3
[A
[0,t
0
]
(t
1
)A
[0,t
0
]
(t
2
) +A
[0,t
0
]
(t)A
[0,t
0
]
(t
2
) +A
[0,t
0
]
(t
1
)A
[0,t
0
]
(t)]
=
1
3
[A
t
1
,t
2
<t
0

+A
t,t
2
<t
0

+A
t
1
,t<t
0

]
= A
t,t
1
,t
2
<t
0

+
1
3
A
t
1
,t
2
<t
0
<t
+
1
3
A
t,t
2
<t
0
<t
1

+
1
3
A
t,t
1
<t
0
<t
2

(6.4)
and hence, using (1.14),
T
_
0
W
2
(t
0
)W(t) = t
0
W(T) +
T
_
0
I
2
[f
2
]W(t)
= t
0
W(T) +I
3
[

f
2
]
= t
0
W(T) + 6J
3
[

f
2
]
= t
0
W(T) + 6
T
_
0
t
3
_
0
t
2
_
0
A
3
[0,t
0
]
(t
1
, t
2
, t
3
)dW(t
1
)dW(t
2
)dW(t
3
)
+6
T
_
0
t
3
_
0
t
2
_
0
1
3
A
t
1
,t
2
<t
0
<t
3

dW(t
1
)dW(t
2
)dW(t
3
)
= t
0
W(T) +t
3/2
0
h
3
(
W(t
0
)

t
0
) + 2
T
_
t
0
t
0
_
0
t
2
_
0
dW(t
1
)dW(t
2
)dW(t
3
)
= t
0
W(T) +t
3/2
0
(
W
3
(t
0
)
t
3/2
0
3
W(t
0
)

t
0
) + 2
T
_
t
0
(
1
2
W
2
(t
0
)
1
2
t
0
)dW(t
3
)
= t
0
W(T) +W
3
(t
0
) 3t
0
W(t
0
) + (W
2
(t
0
) t
0
)(W(T) W(t
0
))
= W
2
(t
0
)W(T) 2t
0
W(t
0
).
2.1. d) By Exercise 1.2. d) and (1.14) we get
T
_
0
exp(W(T))W(t) =
T
_
0
(

n=0
1
n!
exp(
1
2
T)I
n
[1])W(t)
=

n=0
1
n!
exp(
1
2
T)I
n+1
[1] = exp(
1
2
T)

n=0
1
n!
T
n+1
2
h
n+1
(
W(T)

T
)
3.1. a)
T
_
0
W(T)W(t) =
T
_
0
W(T)

W
(t)dt = W(T)
T
_
0

W
(t)dt = W(T) W(T) =
W
2
(T) T, by (3.58).
6.7

3.1. b)
T
_
0
(
T
_
0
gdW)

W
(t)dt = (
T
_
0
gdW)
T
_
0

W
(t)dt = (
T
_
0
gdW) W(T)
= (
T
_
0
gdW)W(T)
T
_
0
g(s)ds, by (3.45).
3.1. c)
T
_
0
W
2
(t
0
)W(t) =
T
_
0
(W
2
(t
0
) +t
0
)W(t)
= W
2
(t
0
) W(T) +t
0
W(T)
= W
2
(t
0
(W(T) W(t
0
)) +W
2
(t
0
) W(t
0
) +t
0
W(T)
= W
2
(t
0
) (W(T) W(t
0
)) +W
3
(t
0
) +t
0
W(T)
= (W
2
(t
0
) t
0
) (W(T) W(t
0
)) +W
3
(t
0
) 3t
0
W(t
0
) +t
0
W(T)
= W
2
(t
0
)W(T) 2t
0
W(t
0
),
where we have used (3.55) and (3.58).
3.1. d)
T
_
0
exp(W(T))W(t) = exp(W(T))
T
_
0

W
(t)dt = exp(W(T)) W(T) =
exp

(W(T) +
1
2
T) W(T) = exp(
1
2
T)

n=0
1
n!
W(T)
(n+1)
= exp(
1
2
T)

n=0
T
n+1
2
n!
h
n+1
_
W(T)

T
_
.
4.1. a) D
t
W(T) = A
[0,T]
(t) = 1 (for t [0, T]), by (4.16).
4.1. b) By the chain rule (Lemma 4.9) we get
D
t
(exp W(t
0
)) = exp W(t
0
) A
[0,t
0
]
(t).
4.1. c) By (4.15) we get
D
t
(
T
_
0
s
2
dW(s)) = t
2
.
4.1. d) By Theorem 4.16 we have
D
t
(
T
_
0
(
t
2
_
0
cos(t
1
+t
2
)dW(t
1
))dW(t
2
)) = D
t
(
1
2
I
2
[cos(t
1
+t
2
)])
=
1
2
2 I
1
[cos( +t)] =
T
_
0
cos(t
1
+t)dW(t
1
).
6.8

4.1. e) D
t
(3W(s
0
)W
2
(t
0
) + ln(1 +W
2
(s
0
)))
= [3W
2
(t
0
) +
2W(s
0
)
1 +W
2
(s
0
)
] A
[0,s
0
]
(t)
+6W(s
0
)W(t
0
)A
[0,t
0
]
(t)
4.1. f) By Exercise 2.1. b) we have
D
t
(
T
_
0
W(t
0
)W(t)) = D
t
(W(t
0
)W(T) t
0
)
= W(t
0
) A
[0,T]
(t) +W(T)A
[0,t
0
]
(t) = W(t
0
) +W(T)A
[0,t
0
]
(t).
4.2. a) D
t
(exp(
T
_
0
g(s)dW(s))) = D
t
(

n=0
I
n
[f
n
])
=

n=1
nI
n1
[f
n
(, t)]
=

n=1
n
1
n!
exp(
1
2
|g|
2
)I
n1
[g(t
1
) . . . g(t
n1
)g(t)]
= g(t)

n=1
1
(n 1)!
exp(
1
2
|g|
2
)I
n1
[g
(n1)
]
= g(t) exp(
T
_
0
g(s)dW(s)),
where we have used Theorem 4.16.
4.2. b) The chain rule gives
D
t
(exp(
T
_
0
g(s)dW(s))) = exp(
T
_
0
g(s)dW(s))D
t
(
T
_
0
g(s)dW(t))
= g(t) exp(
T
_
0
g(s)dW(s)) by (4.15).
4.3 With the given F and the left hand side of (4.28) becomes
E[D

F ] = E[D

F] = E[
T
_
0
D
t
F g(t)dt] =
T
_
0
(t)g(t)dt,
6.9

while the right hand side becomes
E[F
T
_
0
gdW] E[F D

]
= E[(
T
_
0
dW) (
T
_
0
gdW)],
which is the same as the left hand side by the Ito isometry.
5.1. a) If s > t we have
E
Q
[

W(s)[T
t
] =
E[Z(T)

W(s)[T
t
]
E[Z(T)[T
t
]
=
E[Z(T)

W(s)[T
t
]
Z(t)
= Z
1
(t)E[E[Z(T)

W(s)[T
s
][T
t
]
= Z
1
(t)E[

W(s)E[Z(T)[T
s
][T
t
]
= Z
1
(t)E[

W(s)Z(s)[T
t
]. (6.5)
Applying Itos formula to Y (t): = Z(t)

W(t) we get
dY (t) = Z(t)d

W(t) +

W(t)dZ(t) +d

W(t)dZ(t)
= Z(t)[(t)dt +dW(t)] +

W(t)[(t)Z(t)dW(t)] (t)Z(t)dt
= Z(t)[1 (t)

W(t)]dW(t),
and hence Y (t) is an T
t
-martingale (w.r.t. P). Therefore, by (6.5),
E
Q
[

W(s)[T
t
] = Z
1
(t)E[Y (s)[T
t
] = Z
1
(t)Y (t) =

W(t).

5.1. b) We apply the Girsanov theorem to the case with (t, ) = a. Then X(t) is a
Wiener process w.r.t. the measure Q dened by
dQ() = Z(T, )dP() on T
T
,
where
Z(t) = exp(aW(t)
1
2
a
2
t) ; 0 t T.
5.1. c) In this case we have
(t, ) = bY (t, ), (t, ) = aY (t, ), (t, ) = cY (t, )
and hence we put
=
(t, ) (t, )
(t, )
=
b a
c
6.10

and
Z(t) = exp(W(t)
1
2

2
t); 0 t T.
Then

W(t): = t +W(t)
is a Wiener process w.r.t. the measure Q dened by dQ() = Z(T, )dP() on T
T
and
dY (t) = bY (t)dt +cY (t)[d

W(t) dt] = aY (t)dt +cY (t)d

W(t).

5.2. a) F() = W(T) D

t
F() = A
[0,T]
(t) = 1 for t [0, T] and hence
E[F] +
T
_
0
E[D
t
F[T
t
]dW(t) =
T
_
0
1 dW(t) = W(T) = F.

5.2. b) F() =
T
_
0
W(s)ds D
t
F() =
T
_
0
D
t
W(s)ds =
T
_
0
A
[0,s]
(t)ds =
T
_
t
ds = T t,
which gives
E[F] +
T
_
0
E[D
t
F[T
t
]dW(t) =
T
_
0
(T t)dW(t)
=
T
_
0
W(s)dW(s) = F,
using integration by parts.
5.2. c) F() = W
2
(T) D
t
F() = 2W(T) D
t
W(T) = 2W(T). Hence
E[F] +
T
_
0
E[D
t
F[T
t
]dW(t) = T +
T
_
0
E[2W(T)[T
t
]dW(t)
= T + 2
T
_
0
W(t)dW(t) = T +W
2
(T) T = W
2
(T) = F.

5.2. d) F() = W
3
(T) D
t
F() = 2W
2
(T). Hence
E[F] +
T
_
0
E[D
t
F[T
t
]dW(t) =
T
_
0
E[3W
2
(T)[T
t
]dW(t)
6.11

= 3
T
_
9
E[(W(T) W(t))
2
+ 2W(t)W(T) W
2
(t)[T
t
]dW(t)
= 3
T
_
0
(T t)dW(t) + 6
T
_
0
W
2
(t)dW(t) 3
T
_
0
W
2
(t)dW(t)
= 3
T
_
0
W
2
(t)dW(t) 3
T
_
0
W(t)dt = W
3
(T),
by Itos formula.
5.2. e) F() = exp W(T) D
t
F() = exp W(T). Hence
RHS = E[F] +
T
_
0
E[D
t
F[T
t
]dW(t)
= exp(
1
2
T) +
T
_
0
E[exp W(T)[T
t
]dW(t)
= exp(
1
2
T) +
T
_
0
E[exp(W(T)
1
2
T) exp(
1
2
T)[T
t
]dW(t)
= exp(
1
2
T) + exp(
1
2
T)
T
_
0
exp(W(t)
1
2
t)dW(t). (6.6)
Here we have used that
M(t): = exp(W(t)
1
2
t)
is a martingale. In fact, by Itos formula we have dM(t) = M(t)dW(t). Combined with
(6.6) this gives
RHS = exp(
1
2
T) + exp(
1
2
T)(M(T) M(0)) = exp W(T) = F .

5.2. f) F() = (W(T) +T) exp(W(T)

1
2
T)
D
t
F = exp(W(T)
1
2
T)[1 W(T) T]. Note that
Y (t): = (W(t) +t)N(t), with N(t) = exp(W(t)
1
2
t)
is a martingale, since
dY (t) = (W(t)+t)N(t)(dW(t))+N(t)(dW(t)+dt)N(t)dt = N(t)[1tW(t)]dW(t)
6.12

Therefore
E[F] +
T
_
0
E[D
t
F[T
t
]dW(t)
=
T
_
0
E[N(T)(1 (W(T) +T))[T
t
]dW(t)
=
T
_
0
N(t)(1 (W(t) +t))dW(t) =
T
_
0
dY (t) = Y (T) Y (0)
= (W(T) +T) exp(W(T)
1
2
T) = F.

5.3. a) (t, ) = E
Q
[D
t
F F
T
_
t
D
t
(s, )d

W(s)[T
t
]
If (s, ) = (s) is deterministic, then D
t
= 0 and hence
(t, ) = E
Q
[D
t
F[T
t
] = E
Q
[2W(T)[T
t
]
= E
Q
[2

W(T) 2
T
_
0
(s)ds[T
t
] = 2

W(t) 2
T
_
0
(s)ds = 2W(t) 2
T
_
t
(s)ds.
5.3. b) (t, ) = E
Q
[D
t
F[T
t
]
= E
Q
[exp(
T
_
0
(s)dW(s))(t)[T
t
]
= (t)E
Q
[exp(
T
_
0
(s)d

W(s)
T
_
0
(s)(s)ds)[T
t
]
= (t) exp(
T
_
0
(
1
2

2
(s) (s)(s))ds)E
Q
[exp(
T
_
0
(s)d

W(s)
1
2
T
_
0

2
(s)ds)[T
t
]
= (t) exp(
T
_
0
(s)(
1
2
(s) (s))ds) exp(
t
_
0
(s)d

W(s)
1
2
t
_
0

2
(s)ds)
= (t) exp(
t
_
0
(s)dW(s) +
T
_
t
(s)(
1
2
(s) (s))ds).
5.3. c) (t, ) = E
Q
[D
t
F F
T
_
t
D
t
(s, )d

W(s)[T
t
]
= E
Q
[(t)F[T
t
] E
Q
[F
T
_
t
d

W(s)[T
t
]
6.13

= A B, say. (6.7)
Now

W(t) = W(t) +
t
_
0
(s, )ds = W(t) +
t
_
0
W(s)ds or
dW(t) +W(t)dt = d

W(t).
We solve this equation for W(t) by multiplying by the integrating factor e
t
and get
d(e
t
W(t)) = e
t
d

W(t)
Hence
W(u) = e
u
u
_
0
e
s
d

W(s). (6.8)
or
dW(u) = e
u
u
_
0
e
s
d

W(s)du +d

W(u) (6.9)
Using (6.9) we may rewrite F() as follows:
F() = exp(
T
_
0
(s)dW(s))
= exp(
T
_
0
(s)d

W(s)
T
_
0
(u)e
u
(
u
_
0
e
s
d

W(s))du)
= exp(
T
_
0
(s)d

W(s)
T
_
0
(
T
_
0
(u)e
u
du)e
s
d

W(s))
= K(T) exp(
1
2
T
_
0

2
(s)ds),
where
(s) = (s) e
s
T
_
s
(u)e
u
du (6.10)
and
K(t) = exp(
t
_
0
(s)d

W(s)
1
2
t
_
0

2
(s)ds); 0 t T. (6.11)
Hence
A = E
Q
[(t)F[T
t
] = (t) exp(
1
2
T
_
0

2
(s)ds)E[K(T)[T
t
]
= (t) exp(
1
2
T
_
0

2
(s)ds)K(t). (6.12)
6.14

Moreover, if we put
H = exp(
1
2
T
_
0

2
(s)ds), (6.13)
we get
B = E
Q
[F(

W(T)

W(t))[T
t
]
= H E
Q
[K(T)(

W(T)

W(t))[T
t
]
= H E
Q
[K(t) exp(
T
_
t
(s)d

W(s)
1
2
T
_
t

2
(s)ds)(

W(T)

W(t))[T
t
]
= H K(t)E
Q
[exp(
T
_
t
(s)d

W(s)
1
2
T
_
t

2
(s)ds)(

W(T)

W(t))]
= H K(t)E[exp(
T
_
t
(s)dW(s)
1
2
T
_
t

2
(s)ds)(W(T) W(t))]. (6.14)
This last expectation can be evaluated by using Itos formula: Put
X(t) = exp(
t
_
t
0
(s)dW(s)
1
2
t
_
t
0

2
(s)ds)
and
Y (t) = X(t) (W(t) W(t
0
)).
Then
dY (t) = X(t)dW(t) + (W(t) W(t
0
))dX(t) +dX(t)dW(t)
= X(t)[1 + (W(t) W(t
0
))(t)]dW(t) +(t)X(t)dt
and hence
E[Y (T)] = E[Y (t
0
)] +E[
T
_
t
0
(s)X(s)ds]
=
T
_
t
0
(s)E[X(s)]ds =
T
_
t
0
(s)ds (6.15)
Combining (6.7), (6.10)(6.15) we conclude that
(t, ) = (t)HK(t) H K(t)
T
_
t
(s)ds
= exp(
1
2
T
_
0

2
(s)ds) exp(
t
_
0
(s)d

W(s)
1
2
t
_
0

2
(s)ds)[(t)
T
_
t
(s)ds]
6.15

5.4. a) Since =

is constant we get by (5.57)

(t) = e
t

1
S
1
(t)E
Q
[e
T
D
t
W(T)[T
t
] = e
(tT)

1
S
1
(t).
5.4. b) Here = 0, (s, ) = c S
1
(s) and hence
(s, ) =

c
S(s) = (W(s) +S(0)).
Hence
T
_
t
D
t
(s, )d

W(s) = [

W(t)

W(T)].
Therefore
B: = E
Q
[F
T
_
t
D
t
(s)d

W(s)[T
t
] = E
Q
[e
T
W(T)(

W(t)

W(T))[T
t
]. (6.16)
To proceed further, we need to express W(t) in terms of

W(t): Since

W(t) = W(t) +
t
_
0
(s, )ds = W(t) S(0)t
t
_
0
W(s)ds
we have
d

W(t) = dW(t) W(t)dt S(0)dt

or
e
t
dW(t) e
t
W(t)dt = e
t
(d

W(t) +S(0)dt)
or
d(e
t
W(t)) = e
t
d

W(t) +e
t
S(0)dt
Hence
W(t) = S(0)[e
t
1] +e
t
t
_
0
e
s
d

W(s). (6.17)
Substituting this into (6.16) we get
B = E
Q
[
T
_
0
e
s
d

W(s) (

W(t)

W(T))[T
t
]
= E
Q
[
t
_
0
e
s
d

W(s) (

W(t)

W(T))[T
t
]
+E
Q
[
T
_
t
e
s
d

W(s) (

W(t)

W(T))[T
t
]
= E
Q
[
T
_
t
e
s
d

W(s) (

W(t)

W(T))]
=
T
_
t
e
s
(1)ds = e
T
e
t
.
6.16

Hence
(t) = e
t
c
1
(E
Q
[D
t
(e
T
W(T))[T
t
] B)
= e
t
c
1
(e
T
e
T
+e
t
) = c
1
(as expected).
5.4. c) Here = c S
1
(t) and hence
(s, ) =

c
S(s) =

c
[e
s
S(0) +c
s
_
0
e
(sr)
dW(r)]
So
D
t
(s, ) = ( )e
(st)
A
[0,s]
(t)
Hence
(t, ) = e
t
c
1
E
Q
[(D
t
(e
T
W(T)) e
T
W(T)
T
_
t
D
t
(s, )d

W(s)[T
t
]
= e
(tT)
c
1
(1 ( )E
Q
[W(T)
T
_
t
e
(st)
d

W(s)[T
t
]). (6.18)
Again we try to express W(t) in terms of

W(t): Since
d

W(t) = dW(t) +(t, )dt

= dW(t) +

c
[e
t
S(0) +c
t
_
0
e
(tr)
dW(r)]dt
we have
e
t
d

W(t) = e
t
dW(t) + [

c
S(0) + ( )
t
_
0
e
r
dW(r)]dt (6.19)
If we put
X(t) =
t
_
0
e
r
dW(r),

X(t) =
t
_
0
e
r
d

W(r),
(6.19) can be written
d

X(t) = dX(t) +

c
S(0)dt + ( )X(t)dt
or
d(e
()t
X(t)) = e
()t
d

X(t) +

c
S(0)e
[)t
dt
or
X(t) = e
()t
t
_
0
e
s
d

W(s) +

c
S(0)e
()t
t
_
0
e
()s
ds
= e
()t
t
_
0
e
s
d

W(s) +
S(0)
c
[1 e
()t
]
6.17

From this we get
e
t
dW(t) = e
()t
e
t
d

W(t) + ( )e
()t
(
t
_
0
e
s
d

W(s))dt

S(0)
c
( )e
()t
dt
or
dW(t) = d

W(t) + ( )e
t
(
t
_
0
e
s
d

W(s))dt
S(0)
c
( )e
t
dt
In particular,
W(T) =

W(T) + ( )
T
_
0
e
s
(
s
_
0
e
r
d

W(r))ds
S(0)
c
( )(e
T
1) (6.20)
Substituted in (6.18) this gives
(t, ) = e
(tT)
c
1
1 ( )E
Q
[

W(T)
T
_
t
e
(st)
d

W(s)[T
t
]
+( )
2
E
Q
[
T
_
0
e
s
(
s
_
0
e
r
d

W(r))ds
T
_
t
e
(st)
d

W(s)[T
t
]
= e
(tT)
c
1
1 ( )
T
_
t
e
(st)
ds
+( )
2
T
_
t
e
s
E
Q
[(
s
_
t
e
r
d

W(r)) (
T
_
t
e
(rt)
d

W(r))[T
t
]ds
= e
(tT)
c
1
1

(e
(Tt)
1) + ( )
2
T
_
t
e
r
(
s
_
t
e
r
e
(rt)
dr)ds
= e
(tT)
c
1
1

(e
(Tt)
1).
6.18