STOCHASTIC INTEGRATION WITH RESPECT TO

FRACTIONAL BROWNIAN MOTION

LAURENT DECREUSEFOND

This is a tutorial on stochastic integration with respect

to fractional Brownian motion and how one can construct a sort of
stochastic calculus ' la It'.
Abstract.

1. Introduction
In the past few years, fractional Brownian motion has been the subject
of numerous investigations. Its potential applications to telecommunications
and mathematical nance are the practical reasons for which this process
is so much studied.

On the theoretical point of view, it is an interesting

process because it is neither a Markov process nor a semi-martingale so that

stochastic calculus with respect to it is challenging.
attempts have been made to dene a

good

In particular, several

stochastic integral with respect

to fBm. This paper aims to describe the main approaches developed up to

now and to show how they relate to each other. This work is organized as
follows: in Section 3, we give and compare the dierent denitions, existing
in the current literature, of what could be a stochastic integral with respect
to fBm. Section 4 is devoted to the usage we can make of these integrals to
develop a stochastic calculus ' la It', namely we focus on the basic tools
such as Girsanov theorem, Clark representation formula and It formula. At
last, in Section 5, we illustrate how these tools can be applied to two concrete
problems: the parametric estimation of the drift of a 'fractional diusion'
and the non-linear ltering problem. Notations and concepts of deterministic
fractional calculus and Malliavin calculus are given in the appendices.

2. Preliminaries

Denition 2.1.

H in (0, 1), fractional Brownian motion of index

H , {BH (t); t [0, 1]} is the centered Gaussian process

For any

(Hurst parameter)

whose covariance kernel is given by

def

RH (s, t) = E [BH (s)BH (t)] =


VH  2H
s + t2H |t s|2H
2

AMS Subject classication: Primary 60H05; secondary 60H07 60H07.

Keywords and phrases: fractional Brownian motion, Malliavin calculus,
integral

stochastic

L. Decreusefond

where

(2 2H) cos(H)
.
H(1 2H)

def

VH =

Its main useful properties are (see [9, 11] and references therein for their
proofs) :
i  The sample-paths of

BH

are almost surely Hlder continuous of any

H.
1/H -variation on [0, t]

order less than

ii  Its

lim

n 1
2X

n+

j=0

is nite :

0
n
n
p
|BH ((j + 1)2 t) BH (j2 t)| =

VH .t

if
if
if

pH > 1,
pH < 1,
pH = 1.

BH is a Dirichlet process (i.e., its quadratic variation is

H > 1/2 and has an innite quadratic variation for H < 1/2.
exists a standard Brownian motion {B(s), s [0, 1]} such that
Z t
BH (t) =
KH (t, r) dB(r), where

In particular,
null) for
iii  There
(2.1)

0
H 1

(t r)+ 2
1
1
1
t
KH (t, r) =
1 F ( 2 H, H 2 , H + 2 , 1 r ).
(H + 2 )

(2.2)

F (, , , z) (see [21]) is the analytic continuation on C C C\{1, 2, . . .} {z C, Arg|1z| < }

The Gauss hyper-geometric function

of the power series

+
X
()k ()k
k=0
Here

()k

()k k!

zk .

denotes the Pochhammer symbol dened by

()0 = 1

and

def

()k =

Remark 2.1. Actually,

KH

( + k)
= ( + 1) . . . ( + k 1).
()

is just the function which is such that

tr

KH (t, s)KH (r, s) ds = RH (t, r).

(2.3)

0
Using known results on hyper-geometric functions, one can show (this is done
in [11] that

KH

is the product of a bicontinuous function with the singular

factor
1

s|H /2| (t s)H /2 .

It is thus not surprising that it has the following graphical appearance:
It is also crucial to note that (2.3) means that the linear integral operator

KH is a quasi-nilpotent square root of the integral operator of

RH . Hence such a procedure can be done for any Gaussian process

of kernel
kernel

since such a square-root always exists  see Davidson [8].

Stochastic Integration w.r.t. fBm

3.5
3.5
3
3
2.5
2.5
2
2
1.5
1.5
1
1
0.5
s
00.5

0.2

0.4

0.6

0.8

Figure 1. Graph of

KH (1, .)

We x once for all a value of

the index

and

H (1, .)
K

H 6= 1/2

0.2

0.4

for

0.6

0.8

H = 0.2, H = 0.8.

so that, otherwise stated, we omit

everywhere for the sake of notational simplicity.

3. Constructions
The dierent denitions of stochastic integrals with respect to

BH

can

be sorted in two main groups : those which are relying on the sample-path
properties of
3.1.

BH

and those which are based on its Gaussiannity.

Sample-path dened integrals.

The very rst idea which comes to

mind when we try to construct a stochastic integral with respect to

BH

is

to consider the so-called Riemann sums :

(3.1)

def

RSn (u) =

n 1
2X

ui2n (BH ((i + 1)2n ) BH (i2n ))

i=0
and then to identify the conditions on

which are sucient to ensure the

convergence (at least in probability) of this quantity.

Since for

H > 1/2,

the fBm is a Dirichlet process , (3.1) can be given a

sense using the approach developed by Fllmer [14].

since fBm has

1/H

On the other hand,

bounded variation, one can use the work of Bertoin [3]

RSn (u) converges whenever u

+ H > 1 and 2. In the same vein,

1/ -bounded

in which it is proved that

has

variation with

one can also cite

the papers Dai et al. [7, 19] which consider more specically the case of the
fractional Brownian motion. Using the fact that for

and

g two continuously

dierentiable functions, we have

Z
|

f dg| ckf kHol() kgkHol()

(see Appendix A for the denitions of the notations such as

+ 1, it is proved in
that u Hol+ (1 H) 

Feyel et al. [13] that

RSn (u)

where

for similar approaches, see Ciesielski et al.

Afterwards, we set

Z
(rs)-

Hol())

converges provided

u(s) dBH (s) = lim RSn (u),

n

[5].

L. Decreusefond

whichever hypothesis are chosen to ensure its convergence. A slightly dierent approach appeared in Zhle [32].

Denition 3.1.
of

If

with respect to

u belongs to B,1 for some > 1H, a stochastic integral

BH can be dened according to the fractional integration

by parts formula :

Z
(3.2)

(fip)-

u(s) dBH (s) =

Z 1
1

1
+

(1)
I0
+ (u0+ )(s)I1 (BH )(s) ds + u(0 )BH (1 ),
0

+
+
where u0+ (s) = u(s) u(0 ) and u(0 ) = lim0 u(). Similarly,
BH (1 ) BH (s) and BH (1 ) = lim0 BH (1 ).

1 (s) =
BH

R
u Hol+ (1 RH), the integral (fip)- u(s) dBH (s) is well
dened and the process t 7 (fip)- u(s)1[0,t] (s) dBH (s) belongs to Hol (H).
Hence even though this denition is valid for any value of H in (0, 1), one
In particular, if

can iterate it (i.e., consider the stochastic integral of a process dened itself
as a stochastic integral) only for
Moreover, in the case

Theorem 3.1

H > /2
1

H > 1/2,

the value for which

R . If u is in Hol+ (1 H),
(fip)- u(s) dBH (s).

(cf. [32])

and coincides with

H > 1 H.

we have :

(rs)-

u(s) dBH (s)

exists

In all the denitions of this section, one did not care of the adaptedness
of the integrand. Since all these denitions are made pathwise, it is only the
sample-path regularity of the integrand that counts. It is thus not surprising
that these stochastic integrals are not very suitable to a real stochastic calculus. Except for very specic case, it is in fact impossible to compute even
the expectation of any of these integrals. More probabilistic approaches are
given by the next methods.
3.2.

Wiener integrals.

Instead of using the sample-paths properties of

BH ,

we focus now on its Gaussiannity. Stochastic integrals of deterministic functions with respect to a Gaussian process are well known since the early fties
(see Neveu [20] an references therein) and are called Wiener integrals. Their
appealing property is that in the case of the usual Brownian motion, this
integral coincides, for deterministic integrands, with the celebrated It integral. As we shall show below, this is no longer the case for other Gaussian
processes. Nevertheless, there is still a strong relationship between the different kinds of approaches.
As a preliminary to the denition of Wiener integrals, we rst x a few
notations and denitions: We denote by
the space of continuous functions from

the canonical space which is here

[0, 1] into R, vanishing at time 0,

equipped with its strong topology :

kf k = sup |f (t)|.
t[0,1]

Stochastic Integration w.r.t. fBm

, is such that the canonical process, {BH (, t) =

(t), t [0, 1]}, is a fractional Brownian motion of Hurst index H. As for any

The probability measure P on

Gaussian processes, the so-called Reproducing Kernel Hilbert Space (RKHS

for short), denoted by

H,

is of paramount importance. It is dened as the

closure of the vector space spanned by the set of functions

{R(t, .), t [0, 1]}

equipped with the scalar product :

< R(t, .), R(s, .) >= R(t, s)

for all

s.

and

Unfortunately, this general denition of the RKHS is clearly not a satisfying

one. Actually, according to this denition, deciding whether a given function
belongs to

H,

is almost impossible.

Meanwhile, it is well known that for

H is nothing but the space of absolutely

0, whose derivative belongs to L2 ([0, 1]).
it has been proved (in [9] for any H after a result of
Poor [2] for the case H > 1/2) that :

the standard Brownian motion,

continuous functions, vanishing at

In the case of fBm,
Barton and Vincent

Theorem 3.2.

For any

written as

with

f belonging

H (0, 1), H is the set of

Z t
f (t) =
K(t, s)f(s)ds,

0
2
to L ([0, 1]). By denition,

functions

which can be

kf kH = kfkL2 .

It turns out that as a vector space (without any consideration about

norms),

coincides with the space

H+1/2

I0+

(L2 )

(see appendix A for the def-

(H + 1/2)
2
dierentiable, vanishing at 0, whose (H + /2)-th derivative belongs to L .
This way, one can see the similarity between the standard case (H = 1/2)
and the case (H 6= 1/2).

inition of this space), which is the space of functions which are

1

Denition 3.2.

As for any Gaussian process, the (abstract) Wiener integral

with respect to fractional Brownian motion is dened as the linear extension

2
from H in L (, P) of the isometric map :

H L2 (, P)
R(t, .) 7 BH (t).
This means that :

Z X
n

(Wiener)-

i RH (ti , s) dBH (s) =

i=1

in

H.

un

i BH (ti ).

i=1

For any (deterministic) function

where each

n
X

u H,

there exists a sequence

is a nite linear combination of the

R(t, .),

(un , n 1),

which converge to

By denition,

Z
(Wiener)-

u(s) dBH (s) = L (, P) lim

(Wiener)-

un (s) dBH (s).

If we apply this construction to Brownian motion, because of the denition of

the RKHS, we see that we can only integrate functions which are absolutely

L. Decreusefond

continuous and whose derivative is square integrable on

[0, 1].

On the other

hand, it is well known that it is sucient to have the square integrability of

the integrand to give a sense to a stochastic integral with respect to

B.

This

diculty is due to a simple and very often implicit identication. Actually, it

does not change anything to the properties of Wiener integrals if we replace

by an isometrically isomorphic Hilbert space. Thus in the Brownian case,

it is customary to identify

L2 ([0, 1])

and

since the map

isometry from the rst space onto the other: if

identify

and

f (t) =

Rt

0 u ds =

(I01+ u)(t).

I01+

belongs to

is a bijective

L2 ,

then we

It is only when the Wiener

integral is dened this way that it is appears as the restriction of the It

integral.
Let us now do the same sort of work for fractional Brownian motion.
By a representation of

H,

we mean a pair composed of a functional space

and a bijective isometry between this space and

H.

By a concrete version

of the Wiener integral with respect to fBm, we mean a pair composed of a

representation of

together with the family of functions

which will be mapped to the

BH (t)'s

{j(t), t [0, 1]}

in order to conserve the necessary

isometry property, which is



kj(t)kH = E Wt2 .
As a trivial consequence of Theorem 3.2, we have :

Theorem 3.3. There exists a canonical isometric bijection between L2 ([0, 1])
and

given by :

i1 : L2 ([0, 1]) H
t

Z
h 7 f (t) =

K(t, s)h(s)ds.
0

Hence

(L2 ([0, 1]), i1 )

is a representation of

H.

H = 1/2, i1 is nothing but I01+ , the quadrature operator

and that i1 (1[0,t] ) = K1 (t, .). More generally, according to (2.3), for any H,
/2
the predecessor of R(t, .) by i1 is K(t, .).
Another representation, well dened only for H > 1/2, is given by the
Note that when

following theorem :

Theorem 3.4

For any

([2, 12, 22])

H > 1/2,

consider

with the twisted scalar product :

ZZ
< f, g >=

f (s)g(t)|t s|2H2 ds dt.

[0,1]2
First dene the linear map i2 on step functions by :

i2 : (L2 ([0, 1]), <, >) H

1[0,t] 7 R(t, .).

L2 ([0, 1])

equipped

Stochastic Integration w.r.t. fBm

def

H2 =

Denote by i2 , the extension of this map to the whole of

, >). Then (H2 , i2 ) is a representation of H.

2
closure(L ([0, 1]), <

Remark 3.1. Other characterizations of the RKHS of fBm are given in [24].
2

(L ([0, 1]), <, >)

In particular, it is proved there that

is not complete and

that explains that we need to take its closure to obtain a Hilbert space.
To see the connection between this paper and [24], we may transform the
expression of

K(t, s).

After [29, Table 10], we know that for

H 1/2

and

L2 ,
H1/2

Kf = I01+ xH /2 I0+

(3.3)
We introduce

x /2H f.

K , the adjoint of K, i.e., for any f

g two square integrable

and

functions, K g is dened by the relation:

f (s)K g(s) ds.

Kf (s)g(s) ds =
0

By Fubini's Theorem, it turns out that

K g(t) =

K(s, t)g(s) ds
t

and at least formally,

at time

t.

where

t

denotes the Dirac mass

On the other hand, we deduce from (3.3) that

H1/2 H1/2 1
I1 .
x

K = x /2H I1

(3.4)
Since

K(t, s) = K (t )(s),

I11 (t ) = 1[0,t] ,

it follows that

H1/2 H1/2

K(t, s) = s /2H I1

1[0,t] (s).

This is precisely the expression used in [24] to prove that

i3 : L2 ([0, 1]) H
H1/2 H1/2

h 7 f (t) = t /2H I1
with

kf kH = khkL2 ,

is a representation of

H,

even for

h (t),

H < 1/2.

Actually,

rephrased in the setting of the current chapter, this representation is the

'dual' representation of

i1 .

of continuous functions on

Recall that

[0, 1],

is densely included in the space

hence its dual space contains the linear

combination of Dirac masses as a dense subspace. Henceforth, in virtue of

H, since this
H and since we work up to isomorphisms,
representation of H.

(3.4), i3 is strictly speaking a representation of the dual space of

space is isometrically isomorphic to

i3

can also be viewed as a

Now, one can consider the Wiener integrals as the extensions of the following isometries :

Z
(3.5)

(W1)-

: L2 ([0, 1]) L2 (, P)
K(t, .) 7 BH (t)

Z
or

(W2)-

: H2 L2 (, P)

1[0,t] (.) 7 BH (t).

L. Decreusefond

These denitions require a few remarks :

- Note that mapping (W1)-

1[0,t] (s) dBH (s)

to

BH (t)

would be inconsistent

with the isometry property required by the abstract scheme of the Wiener
integral since
(3.6)
E



Z


|(W1)- 1[0,t] (s) dBH (s)|2 = k1[0,t] k2L2 = t 6= E |BH (t)|2 = VH t2H .

- As a consequence of the denition of a Wiener integral and of (3.6), the

1[0,t] (s) dBH (s), t 0 is a centered Gaussian process whose

min(t, s), hence it is a standard Brownian motion which

we denote by B. Moreover, for the very same reasons, for any u determin2
istic belonging to L ,
Z
Z
s ,
(W1)u(s) dBH (s) = u(s) dB
process (W1)-

covariance kernel is

where the right-hand-side integral has to be understood in the It sense.

L2 ([0, 1])

- So either we keep the original scalar product on

change the pre-image of

L2 ([0, 1])

on

BH (t)

to

so that we can keep

K(t, .)

1[0,t]

and we have to

or we change the scalar-product

as the predecessor of

BH (t).

This is

the main point where the situation for fBm diers from the standard case.
All the following diculties come in fact from this Gordian knot.
- A consequence of changing the scalar-product appears in the computation
of the expectation of

BH (t)

given the past

(BH (u), u r)

for some

r < t.

In both cases, one has to express the orthogonality relations:

E [BH (u)(BH (t)

E [BH (t) | BH (u), u r])] = 0,

in terms of orthogonality equations in

L2 ([0, 1]),

respectively

computations turn to be straightforward with (W1)-

H2 .

Explicit

and we easily obtain

that :

Lemma 3.1.

For any

0 r t 1,

E [BH (t) | BH (u), u

Z
r] = (W1)- K(t, s)1[0,r] (s) dBH (s).

The corresponding formula (for

H > 1/2

deduced from the following identity (the

Rexplicit

conditional expectation in terms of (W2)-

Theorem 3.5.

can be

expression of the previous

is given in [22].) :

u L2 ([0, 1]),
Z
u(s) dBH (s) = (W2)- u(s) dBH (s),
K K11
/

For

Z
(3.7)

H > 1/2,

only) in terms of (W2)-

(W1)-

for any function

where K denotes the adjoint of the integral operator of kernel

f and g in L2 ([0, 1]),

Z
(Kf )(s)g(s) ds =

f (s)(K g)(s) ds.

K,

i.e., for

Stochastic Integration w.r.t. fBm

Remark 3.2. We denote by

I01+ ,

we get

K=

the map

K11 K,
/2

I01
+ K that is to say

since

K1/2

coincides with

Kf (t) = (Kf )0 (t).

For

H > 1/2,

according to (3.3), we get

H1/2

Ku(t) = tH /2 I0+
thus

(x /2H u)(t),

H1/2

K u(t) = t /2H I1

(xH /2 u)(t),

which according to the denition of fractional integrals turns out to be

1

t /2H
K u(t) =
(H 1/2)

For

H = /2, K
1

(r t)H /2 rH /2 u(r) dr.

is the identity map. For

H < 1/2, K

is a densely dened,
1

/ H

2
closable operator from L into itself which admits

I0+2

(L2 )

as a core

(i.e., as a dense subset of its domain)  see [10].

Proof of (3.7). By their very denition,

K (t )(s) = K(t, s)

where

t

and

and

are such that

K1/ (t )(s) = I11 (t )(s) = 1[0,t] (s),

2

t.

denotes the Dirac mass at time

L2 ([0, 1])

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

i=1 ui 1[0,ti ] . Since K

Hence

(3.7) is true for step functions of the form

continuous map from

K1/2

Pn

and
is a

into itself, the general case follows by a

trivial limiting procedure.

- The consequence of changing the pre-image of
from

1[0,t]

BH (t) to something dierent

appears in the Riemann-sums approach of the denition of a

H = 1/2, K1/2 (t, .) = 1[0,t]

Z
Z
(W1)u(s) dBH (s) = (W2)- u(s) dBH (s) =

stochastic integral. When

lim

n+
for any continuous

n 1
2X

(W2)-

u(i2n )(BH ((i + 1)2n ) BH (i2n )),

i=0

R
u. From (3.7),R it is clear that (W1)-

from (3.5), it is clear that (W1)-

so that we have :

R
R dier and
(rs)- , whereas

and (W2)-

no longer coincides with

does coincide with (rs)- .

At this point, we thus have two diculties: we have only used Wiener integrals and that means we can integrate only deterministic integrands. The

second point is that we have two integrals at our disposal, namely (W1)-

and (W2)-

and none is strictly better than the other.

The latter may be

more appealing because it coincides with the Riemann sums and thus has
some reminiscences of Riemann-Stieltjes framework of integration. However,

it would be a great mistake to neglect (W1)-

since as we'll see in section 5,

10

L. Decreusefond

it is the convenient tool to express and prove crucial theorems such as Girsanov Theorem and Clark Formula. In virtue of (3.7), it is always possible

to rewrite a result expressed in terms of (W1)-

in terms of (W2)-

but the

result is intractable.
We will see in the next section that the rst diculty (which is only of
technical order) can be easily overcome whereas the second point is really
unavoidable. Actually, for any Gaussian process distinct from the standard
Brownian motion, the situation will be the same as it is for fBm.

It is

Brownian motion which is a singularity and the other processes which follow
the common rules.

3.3.

Skohorod-like integrals.

The Malliavin calculus provides the conve-

nient framework to extend the Wiener integral to random, even non-adapted,

integrands.

For the sake of simplicity, we won't go into the details of the

Malliavin calculus but we will only summarize the useful results.

The basic tool for the following is in fact the Skohorod integral (see Appendix B) with respect to the standard Brownian motion

mentioned in

property (iii) of section 2. The Skohorod integral of a process

denoted by

u(s) dB(s)

is usually

and its denition can be found in several text-

books (see [30]) and is recalled at the end of this chapter. Two of its main
properties are that :

- The expectation of

u(s) dB(s)

is null.

u(s) dB(s) coincides with the

u. The space of processes which are Skohorod integrable,
denoted by Dom is much wider than the set of adapted and square integrable processes. For instance, any functional f (B(t1 ), . . . , B(tn )) where
ti belongs to [0, 1] for any i {1, . . . , n} and f is a Lipschitz continuous
n
function from R into R is Skohorod integrable but not It integrable. For

- For

adapted and square integrable,

It integral of

all these reasons, one can say that the Skohorod integral is an extension of

the It integral (which itself is an extension of the Wiener integral (W1)- ).

Moreover, it is proved in [11] that for deterministic processes, the Skohorod

integral coincides with (W1)-

We thus have solved half of our problem by

extending the rst Wiener integral by the Skohorod integral, which we will

denote from now on by (Sko1)-. To extend (W2)-

Denition 3.3.

u
Z

For

(Sko2)-

(3.8)

such that

K u

belongs to

def

u(s) dBH (s) =

we simply use (3.7):

Dom ,

we consider :

K u(s) dB(s),
K u.
s > t), we

where the right-hand-side denotes the Skohorod integral of

Since

is lower-triangular (i.e.,

K(t, s) = 0

Kt u K1 (u1[0,t] ),

when

have:

Stochastic Integration w.r.t. fBm

11

Kt denotes, by analogy with the notation Ib , the

L ([0, t]), i.e.,
Z t
Z t
Kf (s)g(s) ds =
f (s)Kt g(s) ds
where
2

0
for suciently regular

adjoint of

in

and

g.

Henceforth, there is no ambiguity to set

Kt u(s) dB(s).

u(s) dBH (s) =

( Sko2)0

Remark 3.3. This integral rst appeared in [11] for

H > 1/2

and coincides

with the integral dened in [12] by means of Wick products.

Remark 3.4. It is also remarkable that (Sko2)-

is the integral which pops

up when we try to dene heuristically a stochastic integral with respect to

R
BH . Say we want to dene something like u(s)BH (s) ds where BH aims to
be the derivative of BH  which rigorously speaking does not exist. Using
1
relation (2.1) and the fact that the derivative operator is nothing but K1 ,
/2

we can say that :

u(s) BH (s) ds =

u(s) K11 K(B)(s)

ds.
/2

1
By taking the adjoint map of K1
/2

K, which is in fact K,
Z

u(s)BH (s) ds = K u(s)B(s)

ds,

we obtain

which is the empirical version of (3.8).

Remark 3.5. Note that (Sko2)-

H > 1/2,

is an anticipative integral. For instance, for

since we have (see [29]) :

1

H1/2

Kf = I01+ xH /2 I0+

(3.9)
the computation of

K f (t)

needs the knowledge of

t /2H
(H 1/2)

K f (t) =

x /2H f,

between time

and

xH /2 (x t)H /2 f (x) dx.

H < 1/2, the problem is the same. However, note that for u adapted,
Rt

(Sko2)0 u(s) dBH (s) belongs to Ft even if Kt u is an anticipative process.

The key point here is that by taking the adjoint map of K, we proceed to

For

the analog of an integration by parts and this explains how an anticipative

Rx
0

R1

0 wv dx with v(x) =
v 0 (y) dy. We obtain by a classical integration by parts (including the trace

integral appears. Namely, when one wants to compute

terms in the integral) or by Fubini's theorem,

Z
wv dx =

v (x)

w(y) dy dx,
x

so that we actually have a sort of an anticipative integral of

u.

12

L. Decreusefond

Proposition 3.1

(see [11])

H,

For any

provided that

belongs to the do-

is square integrable for H > 1/2 and which is

1/2H 2
satised whenever u belongs to I +
(L ) for H < 1/2), the stochastic inte0
R
gral ( Sko2)- coincides with the stochastic integral dened by Riemann sums

main of

(which is true if

for deterministic processes. We have the following identity provided that

is deterministic and both sides exist :

When

u(s) dBH (s) = lim

(Sko2)-

|n |0

u(ti )(BH (ti+1 ) BH (ti )).

ti n

is a regular random process, we have :

Z
(3.10)

u(s) dBH (s) =

(Sko2)-

lim

|n |0

u(ti )(BH (ti+1 ) BH (ti ))

Ds u(s) ds.
0

ti n

The term regular and the operator

are dened in appendix B, denition

B.2.
Equation (3.10) leads naturally to dene :

Denition 3.4.

For

regular, we dene :

Z
u(s) dBH (s) = (Sko2)-

(Sko3)-

Z
u(s) dBH (s) +

Ds u(s) ds
0

and

Z
( Sko3)-

Z t
Z t
u(s) dBH (s) = ( Sko2)- u(s) dBH (s) +
Ds u(s) ds.

The similarity of (Sko3)-

with a Stratonovitch-Skohorod type integral (see

[23]) leads us to look at the limit of

LIn (u) =

n1
X

(i+1)/n

n
i/n

i=0


i+1
i 
) BH ( ) .
u(s) ds BH (
n
n

That is to say we look at the limit of the sum obtained by substituting the
linear interpolation of fBm to the dierential element

Theorem 3.6 (see [10]).

dBH (s).

H > 1/2, for u L2 ( [0, 1], dP dt) Dom

p
p
2
and such that u belongs to L (; L ([0, 1] )) for p(H 1/2) > 1; then
2
u is regular
and the sequence {LIn (u), n 1} converges in L (, dP) to
R
(Sko3)- u(s) dBH (s).

Theorem 3.7

H < 1/2, for u D2,1 (B3/2H,2 ), u is regular

R
{LIn (u), n 1} converges in L2 (, dP) to (Sko3)- u(s)

(see [10])

and the sequence

dBH (s).

For

For

Stochastic Integration w.r.t. fBm

3.4.

Other constructions.

At least two other approaches are used to con-

struct a stochastic integral with respect to

- In [1],

BH

13

BH .

is expressed as the limit in probability of a sequence of semi-

martingales. The stochastic integrals with respect to these semi-martingales

being well-dened, the stochastic integral with respect to

BH

is dened as

their limit when it exists. Up to a slight change of notations, the expression

of this integral coincides with that of (Sko3)-

- Another kind of limiting procedure is used in Privault [25].

First dene

the following two spaces :

C0 ([0, 1]2 ) = {k C , k(1, r) = k(0, r) = 0, r [0, 1]},

and let

C0 ([0, 1]2 ) under the norm

Z 1 Z 1
= kkk2L2 ([0,1]2 ) +
s
k 2 (s, r) dr ds.

be the completion of

kkk2W

W -integrable whenever for any

2
sequence (kn , n N) whose elements belong to C0 ([0, 1] ), converging to
R1
k in W, we have 0 u(s)s kn (s, .) ds Dom for any n and the limit of
Z 1 Z 1

k
u(s)s kn (s, .) ds dB(s)
Wn (u) =
A square integrable process is said to be

0
0
2
exists in L (, dP) and is independent of the choice of (kn , n 0). AcR1

tually, considering
0 u(s)s kn (s, .) ds is a way to approach K u (by reg-

K so that Ku becomes
dierentiable). Thus when a process is
R
W -integrable and its (Sko2)- integral is well dened, the latter coincides
k
with the limit of Wn (u).
ularizing

4. Stochastic calculus
One step is to construct a stochastic integral, the other is to be able to
state some theorems using it. The situation for

H 6= 1/2,

will be again very

dierent from the standard one in the sense that we will have to choose one
construction or the other according to the theorem we want to state. For

the Girsanov theorem, (Sko1)-

Theorem 4.1

is the most adequate tool :

(Girsanov theorem, see [11])

Let

be a square integrable

adapted process such that

u
E [1 ]

=1

u
where t

Z
= exp((Sko1)0

1
u(s) dBH (s)
2

u(s)2 ds).

Let Pu be the probability dened by

dPu 


= ut .
d P F H
t
Rt
The law of the process {BH (t) 0 K(t, s)u(s) ds, t [0, 1]} under Pu is the
same as the law of the process BH under P. More generally, for any v square

14

L. Decreusefond

Rt
integrable and adapted process, the law of the process { 0 K(t, s)v(s) dB(s)
Rt
s)u(s)v(s) ds, t [0, 1]} under Pu is the same as the law of the pro0 K(t,
Rt
cess { 0 K(t, s)v(s)u(s) dB(s), t [0, 1]} under P.
Remark 4.1. For non-adapted shifts, Girsanov theorem still holds under con-

venient hypothesis (see [31]) but it is always expressed in terms of (Sko1)-

According to the relation (3.7), this theorem can be expressed in terms of

but the Radon-Nykodym derivative is much less easily computed 

(Sko2)-

see [22] for such an approach. The point is that since

the same ltration, every

(; F, P)

BH

and

generates

martingale is in fact a martingale with

B
B.

respect to the standard Brownian motion

an It stochastic integral with respect to

and hence can be expressed as

For the very same reason, the It-Clark formula is easily expressed as :

Theorem 4.2 (ItClark representation formula).

random variable

F,

For any square integrable

there exists a Skohorod integrable process such that

Z
F E [F ] = (Sko1)-

u(s) dBH (s)

0

If

belongs to

D2,1 ,

then

can be computed by :

u(s) = E [s F | Fs ] .
The It formula is usually one of the most useful tools in stochastic calculus.

Several proofs of a It formula for processes dened as stochastic

integrals with respect to

BH

do exist but they do not seem to be as useful

as the classical formula (see section 5.2).

Assume in a rst part that

is greater than

1/2.

In this case, fractional

Brownian motion has a zero quadratic variation hence it is a Dirichlet process. It is well known that for such processes there exists an It formula (see
[14, 27, 28, 32]) of the form :

Z
where

F 0 (BH (s))1[0,t] (s) dBH (s),

F (BH (t)) = F (BH (0)) + (RS)-

(4.1)

is of class

C 1 . In fact, in our situation, the results known for Dirichlet

processes can be made more explicit. One of the most important properties
of semi-martingales is that a semi-martingale under a probability P is still
a semi-martingale under any probability absolutely continuous with respect
to P. Theorem 4.1 states that processes of the form

Z
(4.2)

K(t, s)(s) dB(s),

K(t, s)(s) ds +

Xt = x0 +

satisfy the above property so they could be taken here as the analog of
semi-martingales for fBm.
Note that for

constant, the second term reduces to

a classical drift term in the form

Rt
0

(s) ds,

BH .

One could had

but technically speaking, it

behaves classically in the sense that in the It formula, there is no cross

Stochastic Integration w.r.t. fBm

15

term between it and the terms involving the Brownian motion and moreover,
in view of the Girsanov, this form of drift is non logical.
Let

(s) = s12H

and denote set

L2 ([0, 1])

h2 (s)(s) ds < +}.

= {h : [0, 1] R,
0

H > 1/2, L2 ([0, 1]) L2 ([0, 1]) because (s) 1 for any
s [0, 1]. The weighted Sobolev space D2,k, is the set of elements of D2,k
such that for any n k, the n-th Gross-Sobolev derivative of belongs to
L2 ([0, 1])n and the norm of in this space is dened by

Note that when

kk22,k,H, = kk2L2 (P) +

k
X

i
k(i) k2L2 ([0,1])i .

i=1

Theorem 4.3

(cf.

[11])

H > 1/2,

Suppose

let

F : R 7 R

be twice

dierentiable with bounded derivatives and X be a process of the form given

2
2
by (4.2), where is in D2,1, (L ([0, 1])) and is in D2,2, (L ([0, 1])). We
have

(4.3)

F (Xt ) F (x0 ) =
F 0 (Xs )(K)(s) ds
0
Z t
+
(s)Kt (F 00 X KH (s ))(s) ds
0
Z t
+
(s)Kt (F 0 X)(s) dB(s)
0
Z t
+
Kt (F 00 XK(., s))(s) (s)2 ds
0
Z t
Z 1



00
+
Kt F X s (r)KH (., r) dB(r) (s)(s) ds.
0

Remark 4.2. As a consequence, it is apparent that the family of processes

of the form (4.2) is not stable by a non-linear transformation. Actually, it
is an open question to nd such a family of processes stable by non-linear
transformations.
One can also wonder what the non-linear transformation of a Skohorod
integral looks like. The answer is given by the following formula.

Theorem 4.4 (cf.

[11])

Assume that for any

Z
(s) ds + (Sko2)-

Zt = z0 +

t [0, 1],

(s) dBH (s),

0
where

belongs to

D2,1, (L2 ([0, 1]))

and

D2,2, (L2 ([0, 1])).

twice dierentiable with bounded derivatives, we have for any

For any

t,

Palmost

16

L. Decreusefond

surely :

Z
F (Zt ) = F (z0 ) +

(4.4)

Z t
F 0 (Zs )(s) ds + (Sko2)- F 0 (Zs )(s) dBH (s)
0
Z t
+
F 00 (Zs )Ds Zs (s) ds.
0

by

This can nicely be expressed in terms of (Sko3)-

Z
F (Zt ) = F (z0 ) +

F 0 (Zs )(s) ds + (Sko3)-

0
Remark 4.3. For

F 0 (Zs )(s) dBH (s).

X(t) = BH (t),

formula (4.3)as well as formula (4.4) read

as

Z t
F (BH (t)) = F (0) + (Sko2)- F 0 (BH (s)) dBH (s)
0
Z t
+ HVH
F 00 (BH (s))s2H1 ds.
0
For
of

/4 H < 1/2,

BH

there exists a formula only for non-linear transformation

itself (see [4, 25, 26]) and it involves several other terms.

5. Applications
We now give two applications of the preceding theorems.
5.1.

An estimation problem.

Consider the process

dened by

Xt = t + BH (t)
and let PX denote its law. We aim to estimate
of a sample-path of

over

[0, t].

Let

through the observation

be such that

(K)(t) = t,

by the

Girsanov Theorem


Z t
Z
dPX 
2 t
= exp( (W1)- (s) dBH (s)
(s)2 ds) or equivalently,
d P  Ft
2 0
0

Z t
Z
2 t
dP 
= exp( (W1)- (s) dBH (s) +
(s)2 ds), i.e.,
d PX  F t
2 0
0


Z t
Z
2 t
2
(s) ds) .
E [F ] = E F (X(w)) exp( (W1)(s) dBH (s) +
2 0
0
Hence the P maximum likelihood ratio estimate

t = R t
0

1
(s)2 ds

Z
(W1)-

(s) dBH (s) = R t

2
0 (s) ds

of

is
Z t

(W2)-

K1 (s) dBH (s).

Note that despite its appearance, the last stochastic integral can be computed
with the knowledge of the sample-path of

by Proposition 3.1 The functions

Stochastic Integration w.r.t. fBm

17

and K1 can be computed, using (3.9), by searching a monomial solution

and we nd:

(s) =

(3/2 H) 1/2H
s
(2 2H)
K1 (s) =

and
5.2.

Non-linear ltering.


1/2H
(3/2 H)
s(1

s)
.
(2 2H)(H + 1/2)

This application is here to show that the It

formula is not that useful here and that an innite dimensional approach
may be more fruitful.
Assume that we are given the observation

of a signal

corrupted by

a fractional Brownian motion noise. Assume furthermore that the original

X,

signal

satises an evolution equation involving an other fBm with a pos-

sibly dierent Hurst index.

conditional law of

The ltering problem consists of nding the

with respect to

Y.

The problem in its whole generality

has been addressed in [6]  some simpler ltering problems have also been
studied in [15, 16, 18]. The usual strategy can be replicated here with some
additional technical diculties due to the lack of martingale properties. In
order to show the main ideas, we will explicitly handle the following simple
case :

Xt = BH1 (t)
Z t
Yt =
KH2 (t, s)h(Xs ) ds + BH2 (t),
0
where

BH1

and

BH2

are two independent fractional Brownian motion of

respective Hurst index

Theorem 5.1.

dQ 
dP 

def

Ft

H1

and

H2 .

Let Q be the probability measure dened by

ht

Z t
Z


1 t
= exp (Sko1)
h(Xs )dBH2 (s)
h(Xs )2 ds
2 0
0

Under Q, the processes :

H (t) = BH (t)
B
1
1

and

H (t) = BH (t)
B
2
2

KH2 (t, s)h(Xs )ds,

0

are two independent fractional Brownian motion of Hurst index

H1

and

We give the proof to illustrate the role of the innite dimension.

Proof. Let

Bi

be the standard Brownian motion

Z
B i (t) = (w1)- 1[0,t] (s) dBH (s).
(t1 , . . . , tn ) [0, 1]n , consider the 2n-dimensional P martingale
Z r
Z r
Zrt1 ,...,tn = (
KH1 (tj , s) dB 1 (s),
KH2 (tj , s) dB 2 (s))j=1,...,n

For any

H2 .

18

L. Decreusefond

and the bounded variation process

Atr1 ,...,tn = (0,

KH2 (tj , s) h(Xs ) ds)j=1,...,n .

0

f,




t ,...,tn
EQ f (Zr1
Atr1 ,...,tn ) = E f (Zrt1 ,...,tn )) .

The classical Girsanov theorem states that for any bounded

Taking
E

r = supi ti

yields to

i
H (tj ), i = 1, 2, 1 j n) = E [f (BH (tj ), i = 1, 2, 1 j n)] .
f (B
1
i

Hence the nite dimensional Qlaws of

(BH1 , BH2 )

H , B
H )
(B
1
2

coincide with those of

under P.

The same reasoning shows that under Q, the couple

H , B
H ).
(B
1
2

law as

(X, Y )

At this point, it is classical to apply the It formula

(provided that it exists hence we need to assume here that

the product

has the same

f (Xt )ht

H1 > 1/2)

to

and then compute the conditional expectations to ob-

tain the socalled unnormalized lter

def

t (f ) =



f (Xt )ht | Ys , s t . The

moral of the part of this work is that all the technical diculties can be overcome but the nal result is not satisfactory. Actually, one nds (see [6]) that
the equation satised by

t (f )

is of the form :

dt (f ) = t (f 00 (Xt] )) dt + t (f.h) dBH2 (t), ,

(Xt] ) is
time t. This

where

a measurable function of the whole samplepath of

to

diculty can be seen as a consequence of the longrange

dependence of the fractional Brownian motion

X,

the innite dimensional process

where

up

BH1 . To overcome it, we lter

W = C0 ([0, 1], R)valued

is the

process

def

Xt =

KH1 (., s) dB 1 (s),

0
which is related to

BH1

by the identity

Xt (t) = BH1 (t).

Using the It formula of Kuo [17] for Banach valued processes, for su-

F  : W R,

h
EQ F (Xt )t | Yt solves the

ciently regular

Z
(5.1)

t (F ) = F (x0 ) +
0

we nd that the unnormalized lter

def

t (F ) =

dierential equation :

s (F.h ps ) dB 2 (s)


Z
1 t
2

s (KH1 (., s)) D F (.)(KH1 (., s)) ds,

+
2 0

ps (x) = x(s) for any x W and D2 F is the second order Frchet

drivative of F in the direction B,2 . As an example of the eciency of this
method, for h = Id, choose F (x) = exp(ix(1)) where x(1) is the value
where

Stochastic Integration w.r.t. fBm

at time

of the element

of

W,

and set

19

X(t, ) =
t (F ),

we obtain the

following linear partial stochastic dierential equation :

dX(t, ) = 2 KH (1, t)2 X(t, ) dt i

X
(t, ) dBH2 (t),

X(0, ) = x0 .
It is then tricky but possible to nd an explicit Gaussian solution of this
equation and thus to obtain the KalmanBucy lter of
(see [15] for a related result in which

BH2

with respect to

is a standard Brownian motion).

20

L. Decreusefond

Appendix A. Deterministic fractional calculus

f L1 ([0, 1]),

For

the left and right fractional integrals of

are dened

by :

1
()

1
(Ib f )(x) =
()

def
(I0+ f )(x) =
def

f (t)(x t)1 dt , x 0,

0
b

f (t)(t x)1 dt , x b,

> 0 and I 0 = Id . For any 0, any f Lp ([0, 1]) and g Lq ([0, 1])
1 + q 1 , we have :
where p
Z 1
Z 1

f (s)(I0+ g)(s) ds =
(I1 f )(s)g(s) ds.
(A.1)

where

p def
The Besov space I + (L ) =
0

B,p

is usually equipped with the norm :

kf kB,p = kI0
+ f kLp .
We then have the following continuity results (see [13, 29]) :

Proposition A.1.
0 < < 1, 1 < p < 1/, then I0+ is a bounded operator from Lp ([0, 1])
q
1
into L ([0, 1]) with q = p(1 p) .
For any 0 < < 1 and any p 1, B,p is continuously embedded
in Hol( 1/p) provided that 1/p > 0. Hol() denotes the space
of Hldercontinuous functions, null at time 0, equipped with the usual

i. If
ii.

norm :

kf kHol() = sup
t6=s

|f (t) f (s)|
.
|t s|

Hol () the intersection of the spaces Hol()

< and by Hol+ () the union of the spaces Hol() for > .
For any 0 < < < 1, Hol() is compactly embedded in B, .

We formally denote by

for

all
iii.

By

I1 .

I0
+ ,

respectively

I1
,

we mean the inverse map of

I0+ ,

respectively

Appendix B. Malliavin calculus

We give here a very sketchy introduction to the Malliavin calculus. We

denote by

Cn

the simplex of

Rn ,

i.e.,

Cn = {(t1 , . . . , tn ) : 0 t1 t2 . . . tn 1}.
For a deterministic function

Cn

such that :

f (t1 , . . . , tn )2 dt1 . . . dtn < +,

Stochastic Integration w.r.t. fBm

the

n-th

iterated It integral of

1Z

tn

is well dened and given by :

t2

Z
...

21

f (t1 , . . . , tn ) dBt1 dBt2 . . . dBtn .

0

L2s ([0, 1]n ) the space of symmetric elements of

2
n
Ls ([0, 1] ), by denition we have :

We denote by

1 Z tn

Z
In (f ) = n!

For

t2

...
0

L2 ([0, 1]n ).

f (t1 , . . . , tn ) dBt1 dBt2 . . . dBtn .

0

L2 ([0, 1]n ), In (f ) is dened to be equal to In (f s )

f. It is well known that any square inteF on (, P) has a chaos decomposition in L2 (, P) :

For a general element of

s
where f is the symmetrization of
grable random variable

F =

(B.1)

+
X

In (fn ),

n=0
where for each

n, fn L2s ([0, 1]n ).

Denition B.1.

Let

be a square integrable random variable with a devel-

opment of the form (B.1). The functional

is said to belong to

D2,1

if and

only if

(B.2)

nn!kfn k2L2 ([0,1]n ) < +,

n=1
and in this case we set :

t F =

nIn1 (fn (., t)),

n=1
and the sum of the series in (B.2) coincides with E

hR

1
2
0 t F dt

k > 1, the k -th iterated Gross-Sobolev derivative of F, denoted (k) F,

is dened analogously by induction on k. We denote by D2,k the set of func(k) F is well dened.
tionals for which
For

The following denition is specic to our context since we need to use a

damped derivative which we will denote by

Denition B.2.
linear map

Ku

A process

D.

u L2 ( [0, 1])

Recall that

Kf = (Kf )0 .

will be said to be regular if the

is of trace class, i.e.,


 Z 1
(K )s u(s) ds| < .
E |
0

is a closed, unbounded and densely dened operator from L2 (, P) in

L2 ([0, 1]). It thus admits an adjoint, denoted by , which is an unbounded
2
2
operator on L ([0, 1]) with values in L (, P). is the so-called Skohorod
integral or the divergence operator.

22

L. Decreusefond

Denition B.3.

The domain of

such that :

Z

is the set of processes

u L2 ( [0, 1])

|E

t F ut dt | ckF kL2 () ,
0
for all F D2,1 , where c is some constant depending on u. If
Dom , then (u) is the element of L2 (, P) characterized by :
1

Z
E [F (u)]

belongs to


t F ut dt .

=E
0

Moreover, it can be shown that the chaos expansion can be generalized to

processes in the following sense (see [30]) :

Lemma B.1.

For

u L2 ( [0, 1]),

there exists a family of deterministic

fm (t1 , . . . , tm , t), for each

variables and

measurable and square integrable kernels

symmetric with respect to its rst

ut =

(B.3)

m, fm

is

In (fn (., t)),

n=0

L2 ( [0, 1]) and



1
X
u2t dt =
n!kfn kL2 ([0,1]n+1 ) .

where the convergence holds in

Z
E

n=0

We can now express the action of

Theorem B.1.
belongs to

Dom

Let

in terms of the Wiener chaos expansion.

u L2 ( [0, 1])

with the expansion (B.3), then

if and only if

(u) =

In+1 (fn )

n=0
2
converges in L (, P), where

fn (t1 , . . . , tn , t) =

1
n+1

fn

is the symmetrization of

fn ,

i.e.,


fn (t1 , . . . , tn , t)
+

n
X


fn (t1 , . . . , ti1 , t, ti+1 , . . . , tn ) .

i=1
Since the Skohorod integral in an extension of the It integral to a class
of non-adapted processes, it is customary to choose a similar notation for

Rthe two integrals.

u(s) dB(s).

Denition B.4.

That explains why we chose to denote in this text

We dene the damped gradient

(u) by

by:

DF (s) = K(. F )(s),

provided that the right-hand-side is meaningful  see the denition of the
domain of

in Remark 3.2.

Stochastic Integration w.r.t. fBm

It is a common problem, in the standard case, to dene properly

R1
0

s u s

for a process

u.

23

trace(u) =

Actually the gradient of a random variable is only

a square integrable function and also is

us ;

hence, in general, we don't have

any information on the regularity of the function

s 7 s us . Not surprisingly,

the sucient conditions known for the trace to exist impose the continuity
of this function (see [23]). Here the situation is on one side (H
and on the other side, worse.
that it is sucient that

all the more stringent as

the unboundedness of
of

u,

> 1/2) better

For H > /2, the regularizing eect of K entails
1

satises a (strong) integrability (see Theorem 3.6,

is closer to

/2.

On the other hand, for

H < 1/2,

K entails a stronger condition on u than the continuity

. u to be (1 H)-Hlder (see Theorem 3.7).

namely we expect

References

[1] E. Als, O. Mazet, and D. Nualart, Stochastic Calculus with respect to Fractional
Brownian Motion with Hurst Parameter less than 1/2, Stochastic Processes and their
Applications 86 (2000), no. 1, 121139.
[2] R. J. Barton and H. Vincent Poor, Signal Detection in Fractional Gaussian Noise,
IEEE trans. on information theory 34 (1988), no. 5, 943959.
[3] J. Bertoin, Sur une intgrale pour les processus variation borne, The Annals of
Probability 17 (1989), no. 4, 15211535.
[4] P. Carmona and L. Coutin, Intgrale stochastique pour le mouvement brownien fractionnaire, Compte-rendus de l'Acadmie des Sciences 330 2000, 231236.
[5] Z. Ciesielski, G. Kerkyacharian, and B. Roynette, Quelques espaces fonctionnels associs des processus gaussiens, Studia Mathematica 107 (1993), no. 2, 171204.
[6] L. Coutin and L. Decreusefond, Abstract Nonlinear Filtering Theory in Presence of
Fractional Brownian Motion, Annals of Applied Probability 9 (1999), no. 4, 1058
1090.
[7] W. Dai and C.C. Heyde, It's Formula with respect to Fractional Brownian Motion
and its Application, Journal of applied and stochastic analysis 9 (1996), 439458.
[8] K.R. Davidson, Nest Algebras, Triangular Forms for Operator Algebras on Hilbert
Space, Longman Scientic & Technical, 1993.
[9] L. Decreusefond and A.S. stnel, Application du calcul des variations stochastiques
au mouvement brownien fractionnaire, Compte-Rendus de l'Acadmie des Sciences,
321 1995, 16051608.
[10] L. Decreusefond, A Skohorod-Stratonovitch integral for the fractional Brownian motion, Proceedings of the 7-th Workshop on stochastic analysis and related elds, 2000.
Birkhuser.
[11] L. Decreusefond and A.S. stnel, Stochastic Analysis of the Fractional Brownian
Motion, Potential Analysis 10 (1999), no. 2, 177214.
[12] T.E. Duncan, Y. Hu, and B. Pasik-Duncan, Stochastic Calculus for Fractional Brownian Motion I, Theory, SIAM Journal of Control and Optimisation (To appear, 2000).
[13] D. Feyel and A. de La Pradelle, On Fractional Brownian Processes, Potential Analysis
10 (1999), no. 3, 273288.
[14] H. Fllmer, Calcul d'It sans probabibilit, Sminaire de probabilits XV, SpringerVerlag, 1980, pp. 143150.
[15] M.L. Kleptsyna, Kloeden P.E., and V.V. Anh, Linear ltering with fractional Brownian motion, Stochastic Analysis and its Applications 16 (1998), no. 5, 907914.
[16]
, Nonlinear ltering with fractional Brownian motion, Information Transmission Problems 34 (1998), no. 2.

24

L. Decreusefond

[17] H.H. Kuo, Gaussian Measures in Banach Spaces, Lect. Notes in Mathematics, vol.
463, SpringerVerlag, 1975.
[18] A. Le Breton, Filtering and parameter estimation in a simple linear model driven by
a fractional Brownian motion, Statistics and Probability Letters 38 (1998), no. 3,
263274.
[19] S.J. Lin, Stochastic analysis of fractional Brownian motions, Stochastics and Stochastics Reports 55 (1995), no. 1-2, 121140.
[20] J. Neveu, Processus alatoires gaussiens, Presses de l'Universit de Montral, 1968.
[21] A.F. Nikiforov and V.B. Uvarov, Special Functions of Mathematical Physics,
Birkhuser, 1988.
[22] I. Norros, E. Valkeila, and J. Virtamo, A Girsanovtype Formula for the Fractional
Brownian Motion, Proceedings of the rst NordicRussian Symposium on Stochastics,
1996.
[23] D. Nualart, The Malliavin Calculus and Related Topics, SpringerVerlag, 1995.
[24] V. Pipiras and M.S. Taqqu, Fractional calculus and its connections to fractional Brownian motion, In this volume.
[25] N. Privault, Skorohod stochastic integration with respect to non-adapted processes on
Wiener space, Stochastics and Stochastics Reports 65 (1998), 1339.
[26] M. Errami and F. Russo, n-covariation and symetric SDE's driven by nite cubic
variation processes, 2000, Preprint.
[27] F. Russo and P. Vallois, The generalized covariation process and It formula., Stochastic Processes Appl. 59 (1995), no. 1, 81104.
, It formula for C 1 -functions of semimartingales., Probab. Theory Relat.
[28]
Fields 104 (1996), no. 1, 2741.
[29] S.G. Samko, A.A. Kilbas, and O.I. Marichev, Fractional Integrals and Derivatives,
Gordon & Breach Science, 1993.
[30] A.S. stnel, An Introduction to Analysis on Wiener Space, Lectures Notes in Mathematics, vol. 1610, Springer-Verlag, 1995.
[31] A.S. stnel and M. Zakai, Transformations of measure on Wiener spaces, Monographs in mathematics, Springer-Verlag, 2000.
[32] M. Zhle, Integration with respect to Fractal Functions and Stochastic Calculus, Probability Theory and Related Fields 111 (1998).