STOCHASTIC INTEGRATION WITH RESPECT TO

FRACTIONAL BROWNIAN MOTION

PHILIPPE CARMONA, LAURE COUTIN, AND GERARD MONTSENY

A BSTRACT. For every value of the Hurst index H (0, 1) we define a

stochastic integral with respect to fractional Brownian motion of index
H. We do so by approximating fractional Brownian motion by semi-
martingales.
Then, for H > 1/6, we establish an Itos change of variables formula,
which is more precise than Privaults Ito formula [24] (established for
every H > 0), since it only involves anticipating integrals with respect to
a driving Brownian motion.

R E SUM E Pour tout H (0, 1), nous construisons une integrale stochas-
tique par rapport au mouvement Brownien fractionnaire de param`etre de
Hurst H. Cette integrale est basee sur lapproximation du mouvement
Brownien fractionnaire par une suite de semi-martingales.
Ensuite, pour H > 1/6, nous e tablissons une formule dIto qui precise
celle obtenue par Privault [24], au sens o`u elle ne comporte que des
integrales anticipantes par rapport a` un mouvement Brownien directeur.

1. I NTRODUCTION
Fractional Brownian motion was originally defined and studied by Kol-
mogorov [14] within a Hilbert space framework. Fractional Brownian mo-
tion of Hurst index H (0, 1) is a centered Gaussian process W H with co-
variance

 1
E WtH WsH = (t 2H + s2H |t s|2H )

(s,t 0)
2
(for H = 12 we obtain standard Brownian motion).
Fractional Brownian motion has stationary increments

E (W H (t) W H (s))2 = |t s|2H

 
(s,t 0),
and is H-self similar
1 d
( H W H (ct) ;t 0 ) = (W H (t) ;t 0 ) (for all c > 0).
c

Date: December 6, 2001.

1991 Mathematics Subject Classification. Primary 60G15, 60H07, 60H05 ; Secondary
60J65, 60F25 .
Key words and phrases. Gaussian processes, Stochastic Integrals, Malliavin Calculus,
fractional integration.
1
2 P. CARMONA, L. COUTIN, AND G. MONTSENY

The Hurst parameter H accounts not only for the sign of the correlation
of the increments, but also for the regularity of the sample paths. Indeed,
for H > 12 , the increments are positively correlated, and for H < 12 they are
negatively correlated. Furthermore, for every (0, H), its sample paths
are almost surely Holder continuous with index . Finally, it is worthy of
note that for H > 12 , according to Berans definition [3], it is a long memory
process: the covariance of increments at distance u decrease as u2H2 .
These significant properties make fractional Brownian motion a natural can-
didate as a model of noise in mathematical finance (see Comte and Re-
nault [5], Rogers [26]), and in communication networks (see, for instance,
Leland, Taqqu and Willinger [16]).
Recently, there has been numerous attempts at defining a stochastic integral
with respect to fractional Brownian motion. Indeed, for H 6= 12 W H is not a
semi-martingale (see, e.g., example 2 of section 4.9.13 of Liptser and Shyr-
iaev [19]), and usual Itos stochastic calculus may not be applied. However,
the integral
Z t
(1.1) a(s) dW H (s)
0

may be defined for suitable a. On the one hand, since W H has almost its
sample paths Holder continuous of index , for any < H, the integral (1.1)
exists in the Riemann-Stieljes sense (path by path) if almost every sample
path of a has finite p-variation with 1p + > 1 (see Young [32]): this is the
approach used by Dai and Heyde [8] and Lin [18] when H > 12 . Let us recall
that the p-variation of a function f over an interval [0,t] is the least upper
bound of sums i | f (xi ) f (xi1 )| p over all partitions 0 = x0 < x1 < . . . <
xn = T . A recent survey of the important properties of Riemann-Stieltjes in-
tegral is the concentrated advanced course of Dudley and Norvaisa [11]. An
extension of Riemann-Stieltjes integral has been defined by Zahle [33], see
also the recent work of Ruzmaikina [29], by means of composition formu-
las, integration by parts formula, Weyl derivative formula concerning frac-
tional integration/differentiation, and the generalized quadratic variation of
Russo and Vallois [27, 28].
On the other hand, W H is a Gaussian process, and (1.1) can be defined for
deterministic processes a by way of an L2 isometry: see, for example, Nor-
ros, Valkeila and Virtamo [21] or Pipiras and Taqqu [23]. With the help
of stochastic calculus of variations (see [22]) this integral may be extended
to random processes a. In this case, the stochastic integral (1.1) is a di-
vergence operator, that is the adjoint of a stochastic gradient operator (see
the pioneering paper of Decreusefond and Ustunel [9]). It must be noted
that Duncan, Hu and Pasik-Duncan [12] have defined the stochastic integral
in a similar way by using Wick product. Feyel and de La Pradelle [13],
Ciesielski, Kerkyacharian and Roynette [4] also used the Gaussian property
of W H to prove that W H belongs to suitable function spaces and construct a
stochastic integral.
 3

Eventually, Alos, Mazet and Nualart [1] have established, following the
ideas introduced in a previous version of this paper, very sharp sufficient
conditions that ensures existence of the stochastic integral (1.1).
The construction of the stochastic integral. The starting point of our ap-
proach is the following representation of fractional Brownian motion given
by Decreusefond and Ustunel [9]
Z t
H
W (t) = K H (t, s) dB(s) ,
0
where
1
H (t s)H 2 1 1 1 t
(1.2) K (t, s) = F(H , H, H + , 1 ) (s < t) ,
(H + 2 )
1 2 2 2 s
where F denotes Gauss hypergeometric function and (B(t) ;t 0 ) is a stan-
dard Brownian motion.
The first step we take is to define the integral of deterministic functions
(section 4). Observing that K H (t, s) = ItH 1[0,t] (s) where ItH is the integral
operator
Z t
(1.3) ItH H
f (s) = K (t, s) f (s) + ( f (u) f (s))1 K H (u, s) du ,
s
we define
Z t Z t
H
(1.4) f (s) dW (s) = ItH f (u) dBu
0 0
for suitable deterministic functions.
In order to extend formula (1.4) to random processes f , we introduce the
key idea of this paper, which is to approximate W H by processes of the type
Z t
(1.5) WK (t) = K(t, s) dB(s) (t [0, T ])
0
with a kernel K smooth enough to ensure that WK is a semi-martingale
(Proposition 2.5).
Then, for a such a semi-martingale kernel K and a good integrand a, we
prove the following generalization of (1.4)
Z t Z t Z t
(1.6) a(s) dWK (s) = It a(s) dB(s) + It Ds a(s) ds ,
0 0 0
where Ds a denotes the stochastic gradient and It is defined as ItH .
In sections 2 and 3 topologies are defined on the space of kernels and the
space of integrands. These are not natural topologies but ad hoc ones,
whose sole aim is to fulfill the following requirements:
1. The mapping (a, K) 0t a(s) dWK (s) is bilinear continuous (Proposi-
R

tions 6.1).
2. Rough kernels such as K H are in the closure of the space of semi-
martingale kernels.
4 P. CARMONA, L. COUTIN, AND G. MONTSENY

3. For nice integrands a the process t 0t a(s) dWK (s) has a continuous
R

modification (Theorem 7.1).

Eventually, we are able to take limits in the Itos formula established for
semi-martingale kernels in Proposition 8.1. Let Cbn be the set of functions
f whose derivatives, up to the order n N, are continuous and bounded.
Privault proved that for f Cb2 ,
Z t Z t
1
f (W H (t)) = f (0)+ f 0 (W H (s)). dW H (s)+ f 00 (W H (s))(cH 2Hs2H1 ) ds ,
0 2 0
where 0t f 0 (W H (s)). dW H (s) is the L2 -limit of divergences (Skorokhod in-
R

tegrals). For H > 1/4 and f Cb3 , we show that this integral, which is not
in general the limit of Riemann sums, is a Skorokhod integral with respect
to the driving Brownian motion:
Z t
H
(1.7) f (W (t)) = f (0) + ItH f 0 (W H )(s) dBs
0
1 t 00 H
Z
+ f (W (s))(cH 2Hs2H1 ) ds .
2 0
By the same procedure, we have been able to prove an Itos formula for
H > 1/6, in Proposition 8.11. We do not give here this formula since it
is complex (5 lines) and does not seem to be an easy starting point for a
generalization to every H (0, 1).

To end this rather lengthy introduction we try to give an answer to a question

that nearly everyone involved with stochastic integration with respect to
Fractional Brownian motion has asked us.
What is the difference between the stochastic approach and the pathwise
approach to integration with respect to Fractional Brownian motion?
On the one hand we prove in the Appendix that for W1H ,W2H two indepen-
dent fractional Brownian motions of index 1/4 < H < 1/2, the integral
Z t
W1H (s) dW2H (s)
0
cannot be defined for the classical and generalized pathwise integrals.
On the other hand, a slight generalization of our results
R t Henabled HCoutin and
1 1
Qian [7, 6] to show that for 4 < H < 2 the integral 0 W1 (s) dW2 (s) can be
defined as a process with a continuous modification.
Acknowledgement We want to thank an anonymous referee, not only for
correcting some mistakes, but also for detailed and useful comments.
Our main source of inspiration for Itos formula is T. Lyons paper [20],
and in particular the idea that in order to integrate with respect to rough
signals, you may sometimes have to replace approximating Riemann sums,
of order 1, with Taylor sums of higher order. More precisely, you may want
to replace f (W H (ti+1 )) f (W H (ti )) by
f (W H (ti+1 )) f (W H (ti )) W H (ti+1 ) W H (ti ) f 0 (W H (ti )) + . . .


 5

2. T HE VECTOR SPACES OF KERNELS

The spaces of integrands and kernels we shall work with depend on pa-
rameters (p, , ). We shall assume from now on that the set (p, , ) of
parameters is admissible, that is

1 1 2 2p
(2.1) 1 < p < 2, + = 1, > > 1/q, =
p q q 2 p
1 2 1 1
(2.2) h = <= , + = 1.
q /2 /2
Observe that /2 and q/2 are conjugated Holder exponents, and that 1 > 2q .
Let K : R2 R be a measurable kernel such that
Z t
(2.3) K(t, s)2 ds < + for every t > 0
0
(2.4) K(t, s) = 0 if s > t .
We consider the adapted Gaussian process
Z t
WK (t) = K(t, s) dBs ,
0
where (Bt ;t 0 ) is a standard Brownian motion.
Definition 2.1. We say that K is a rough kernel if there exists a measurable
function (t, s) 1 K(t, s), such that u 1 K(u, s) is integrable on every
[t,t 0 ] (s, +) and satisfies
Z t0
(2.5) 0
K(t , s) K(t, s) = 1 K(u, s) du (s < t < t 0 ) .
t
We say that K is a smooth kernel if K is a rough kernel that satisfies
Z t
(2.6) K(t, s) K(s, s) = 1 K(u, s) du (s < t) .
s
The topology on the space of rough kernels is defined via mixed Lebesgue
spaces (see Stroock [31], section 6.2). Let (E1 , B1 , 1 ) and (E2 , B2 , 2 ) be a
pair of -finite measure spaces and let p1 , p2 [1, ). Given a measurable
function f on (E1 E2 , B1 B2 ), define
"Z Z  p2 /p1 #1
p2

k f k p1 ,p2 = | f (x1 , x2 )| p1 1 (dx1 ) 2 (dx2 ) ,

E2 E1

and let L(p1 ,p2 ) (1 , 2 ) denote the mixed Lebesgue space of R valued, B1
B2 measurable f for which k f k p1 ,p2 < .
For i = 1, 2 we let (Ei , Bi , i ) = ((0,t), B (0,t), ds) be the space (0,t) en-
dowed with the -field of Borel sets and Lebesgue measure, p1 = p and
def
p2 = , f (u, s) = 1 K(u, s)(us) 1(0<s<u) and kKk,,p,t = k f k p, +kK(t, .)k ,
that is
6 P. CARMONA, L. COUTIN, AND G. MONTSENY

Z t Z t /p !1/
def
(2.7) kKk,,p,t = |1 K(u, s)(u s) | p du ds
0 s
1
t
Z

+ |K(t, s)| ds .
0

Let E,,p,t be the space of rough kernels K such that kKk,,p,t < +.
Lemma 2.2. E,,p,t is a Banach space.
Proof. Assume that (Kn )nN is a Cauchy sequence in E,,p,t . Then the
sequence of functions fn (u, s) = 1 Kn (u, s)(u s) 1(0<s<u) is a Cauchy se-
quence in L(p,2p/2p) . Therefore it converges in L(p,2p/2p) to a function
f and we let z(u, s) = f (u, s)(u s) . Since (Kn (t, .))nN is a Cauchy se-
quence in L (0,t) it converges in L to a function L (0,t). Finally,
letting K be the kernel
(
(s) vt z(u, s) du
R
if 0 < s < v < t
K(v, s) =
0 if v s t
we conclude that K is a rough kernel for which kK Kn k,,p,t 0 since
by construction :
kK Kn k,,p,t = k f fn k p, + kKn (t, .) (.)k .

Remark 2.3. It is quite straightforward to prove, via Holders inequality,

that for t T , the space E,,p,T can be continuously embedded in E,,p,t .
Indeed, we only need to prove that kK(T, ) K(t, )kL (0,t) < +. We first
observe that fors < t:
Z T  Z T 1/p
1

1 K(u, s) du (t s) q p
 
|K(T, s) K(t, s)| =  |1 K(u, s)(u s) | du .
t t

Therefore, applying Holders inequality to the pair of conjugated exponents

(r = /, r0 )
Z t Z t 1/r0
( q1 )r0
|K(T, s) K(t, s)| ds C (t s) ds kKk,,p,T .
0 0

The first factor may be majorized by a constant C, since 2q > 1

implies
( 1q )r0 > 1.
Let us recall that our main results are proved by approximating the frac-
tional Brownian motion kernel by smooth kernels for which WK is a semi-
martingale.
 7

Definition 2.4. We say that the smooth kernel K is a semi-martingale ker-

nel if
Z u  Zt
(2.8) sup 1 K(u, s) ds + K(s, s)2 ds < + (t > 0).
2
ut 0 0

Proposition 2.5. (1) If K is a semi-martingale kernel, then WK is a semi-

martingale with decomposition
Z t Z t
WK (t) = K(s, s) dBs + W1 K (s) ds .
0 0
(2) The vector space F,,p,t of semi-martingale kernels in E,,p,t is dense
in E,,p,t .

Proof. (1) The definition of a smooth kernel is more than what we need to
apply Fubinis stochastic Theorem (see, e.g., Protter [25] Theorem 46):
Z t Z t
WK (t) = (K(s, s) + 1 K(u, s) du) dBs
0 s
Z t Z t Z u
= K(s, s)dBs + ( 1 K(u, s) dBs ) du .
0 0 0
(2) Let > 0 be given. We perform the change of variables u = v + s in the
definition of kKk,,p,t , and we let g(v, s) = 1 K(v + s, s)v 1(0<v<ts) . Then
g is in L(p,) ((0,t) (0,t)) and by Lemma 6.2.11 of Stroock [31] can be
approached at distance less than by a function h(v, s) having the form
n
h(v, s) = i (v) j (s) ,
i=1
where the i , i are in L (0,t) and the i have disjoint support. Therefore,
if we let
Z t
(r, s) = K(t, s) h(u s, s)(u s) du (s < r < t)
r
then is a kernel such that kK k,,p,t . Furthermore since 1 (r, s) =
h(r s, s)(r s) for 0 < s < r < t, and h is bounded, we see that is a
smooth kernel. However, is not a semi-martingale kernel if > 21 . Hence
for > 0 we consider
Z t
(2.9) (r, s) = K(t, s) h(u s, s)(u + s) du (s < r < t)
r
The is a semi-martingale kernel, and
Z t Z t    /p
u s
k k,,p,t (const) ds

1du 0
u+s

0 s

as 0 by dominated convergence.
8 P. CARMONA, L. COUTIN, AND G. MONTSENY

Definition 2.6. In order to prove the continuity of our stochastic integral

we introduce A,,p,T the subset of E,,p,T of kernels K such that
Z t Z t+ 2

2
kKk = sup (h +)
(u s) 1 K(u, s) du ds < +
0<t<t+T 0 t

and endow it with the norm

k|K|k = kKk,,p,T + kKk .
Since Itos formula is proved by approximating with smooth kernels, we

introduce B,,p,T the closure of F,,p,T in A,,p,T .
2.1. The case of fractional Brownian motion. We recall here the ba-
sic properties of hypergeometric functions required throughout this paper
(see e.g. Chapter 9 of Lebedev [15]). Gauss hypergeometric function
F(, , , z) is defined for every , , , |z| < 1 and 6= 0, 1, . . . by

()k ()k k
F(, , , z) = z ,
k0 ()k k!
( + k)
where ()0 = 1 and ()k = = (+1) . . . (+k 1) is the Pocham-
()
mer index. The convergence radius of this series is 1 and, as soon as
Re( ) > 0, limz1 F(, , , z) exists and is finite. Obviously we
have

F(, , , z) = F(, , , z) , F(, , , 0) = 1 .

Furthermore, if |arg(1 z)| < and Re() > Re() > 0,
()
Z 1
(2.10) F(, , , z) = u1 (1 u)1 (1 zu) du .
()( ) 0

The hypergeometric function F(0 , 0 , 0 , .) is said to be contiguous to F(, , , .)

if | 0 | = 1 or | 0 | = 1 or | 0 | = 1. If F1 and F2 are hypergeometric
functions contiguous to F, then there exists a relation of the type
P(z)F(z) + P1 (z)F1 (z) + P2 (z)F2 (z) = 0 (|arg(1 z)| < ) ,
with P(z), P1 (z), P2 (z) polynomials in the variable z. These relations ensure
that there exists an analytical continuation of F(, , , z) to the domain

C C (C\{0, 1, 2, . . . }) {z : |arg(1 z)| < } .

We shall also use the asymptotic estimate, for |arg(1 z)| < , Re() >
Re() > 0, Re() < 0:
()( )
F(, , , z) (z) (|z| ) .
()( )
 9

Lemma 2.7. Given s > 0 and H 6= 12 , the function t K H (t, s) is differen-

tiable on (s, +) with derivative
1
(s/t) 2 H
1 K (t, s) =
H
(t s)H3/2 (0 < s < t) .
(H 2 )1

Furthermore, K H A,,p,T for admissible parameters (p, q, , . . . ) such that
 
1 1 
+ > 3/2 H , H  + 1 > 0 and + H < 1.


p 2
In particular, for 0 < < H < 12 , there exists a set of admissible parameters
for which h = H .
Remark 2.8. Observe that for H < 12 , we have 1q > 1/2 H > 0.
Proof. Let us state three facts:
(1) From the analyticity of the Gauss function, we deduce that the function
t K H (t, s) is differentiable on (s, +)
 with a derivative f (H,t, s) such
that H f (H,t, s) is holomorphic on H C : Re(H) > 12 .
(2) The integral representation (2.10) implies that for H (1/2, 1) the func-
tion t K H (t, s) is differentiable on (s, +) with derivative 1 K H (t, s).
(3) The function H 1 K H (t, s) is holomorphic in the region
 
1
= {H C : 0 < Re(H) < 1}\ .
2
Therefore, by analytic continuation: for 0 < s < t, 1 K H (t, s) = f (H,t, s)
on .
To determine the admissible set of parameters for which K H E,,p,T , we
just use the fact that (s,t) K H (t, s) is continuous on {0 < s < t}, com-
bined with the following consequence of the asymptotic estimate of F,
1
K H (t, s) (const) s|H 2 |
s0+
1
K H (t, s) st (const) (t s)H 2
Therefore, given 0 < < H < 12 , we can set 1q = 12 3 and h = 1 = H ,
q = H /2 to get q > 2, 2/q > , and q = q (H /2) > 2 H >
2 1 1 1

0.
It remains to show that kKk < + (indeed, we can then approximate K
by the sequence Kn (t, s) = K(t + 1n , s) of smooth kernels in F,,p,T ) : for
0 < t < t + T,
Z t Z t+ 2

(u s) 1 K(u, s) du ds
0 t
Z t Z t+
2
1+2 +H1
(t s) (u s) du ds
0 t

C2(+H)
10 P. CARMONA, L. COUTIN, AND G. MONTSENY

2.2. Properties of Integral Operators associated to Kernels. Let X be

a separable Hilbert space with norm ||X , and K be a rough kernel. To a
measurable a : [0, T ] X we associate
Z t
(2.11) ItK (a)(s) = K(t, s)a(s) + (a(u) a(s)) 1 K(u, s) du ,
0

as soon as the integral in the right hand side makes sense for almost every
s [0, T ].
Similarly we shall denote by JtK the integral operator defined on measurable
f : [0, T ]2 X by
Z t
(2.12) JtK ( f )(s) = f (u, s)1 K(u, s) du ,
s

as soon as the integral in the right hand side makes sense for almost every
s [0, T ].
Observe that if 0s,u a = a(u) a(s), then

ItK (a)(s) = K(t, s)a(s) + JtK (0 a)(s) .

Since K and 1 K need not be positive, we introduce a dominating positive
kernel

T
 Z 
K+ (t, s) = |K(T, s)| + |1 K(u, s)| du 1(s<t) .
t
For 0 s < t t + T we have

(2.13) |K(t + , s) K(t, s)| K+ (t, s) K+ (t + , s)

|K(t, s)| K+ (t, s) .

For R, 6= 0, we let w be the weight

w (u, s) = max(u s, 0) .

Then if du ds denotes the Lebesgue measure on [0, T ]2 , we introduce the

Lebesgue space

Lwq (X) = Lq ([0, T ]2 ; X; w (u, s)du ds) .

Eventually we denote by H (X) the set of Holder continuous functions

taking their values in X:
 11

( )
| f (u) f (s)|X
H (X) = f : [0, T ] X, k f kH (X) = sup < +
0u<sT |u s|
For functions of two variables f (u, s) we introduce

( )
| f (u, s)|X
H (X) = f : [0, T ]2 X, k f kH (X) = sup < +
0u<sT |u s|
Through the rest of this paper (, , p) denotes a set of admissible param-
eters, T > 0 and R . The constant C may vary from line to line but
depends only on (, , p, T, ) ; to ease notation, we shall omit this depen-
dency when the context allows it.

Proposition 2.9. For every fixed t [0, T ]

K
1. The mappings J : ( f , K) J K ( f )(t) and J + : ( f , K) Jt + ( f ) are
q
bilinear continuous from Lwq (X) E,,p,T to L2 ([0, T ]2 ; ds, X).
2. The mapping It : (a, K)  ItK (a) is bilinear continuous
q
L ([0, T ]; ds; X) a : a Lwq (X) E,,p,T
0

from to
2 2
L ([0, T ] ; ds, X).

  K
Proof. Since JtK ( f )X Jt + (| f |X ), we can suppose, without any loss in
generality, that X = R and study J K+ .
First, apply Holders inequality to the pair of conjugated exponents (p, q):
 Z t  1 Z t 1
 q p
 K+ q p
Jt ( f )(s) wq (u, s)| f | (u, s) du wp (u, s)|1 K(u, s)| du .

s s

Then, apply Holders inequality to (q/2, /2) to get

K+
Jt ( f )(.) 2 k f kLwq (R) kKk,,p,T .

L ([0,T ],ds) q

The second assertion is clearly a consequence of the first one, another

Holders inequality for (/2, /2), and Remark 2.3:

kK(t, )a()kL2 ([0,T ],ds) kakL ([0,T ];R) kKk,,p,t CkakL ([0,T ];R) kKk,,p,T .

The next proposition establish the Holder regularity of a rough Kernel K(t, s)
with respect to the first variable t.
12 P. CARMONA, L. COUTIN, AND G. MONTSENY

Proposition 2.10. The space E,,p,T may be continuously embedded into

H h (L2 ([0, T ]; ds)).
Furthermore, the mapping K K+ is continuous from E,,p,T into
H h (L2 ([0, T ]; ds)).

Proof. According to the inequalities (2.13), we only need to prove the sec-
ond point. The embedding is the following : we identify K+ with the
function F : [0, T ] X = L2 ([0, T ]; ds) defined by F(t) = K+ (t, ). Let
0 t t + T ; we have, for s 6= t:
K
K+ (t + , s) K+ (t, s) = |K(T, s)|1(t,t+) (s) + JT + ( f + g)(s)
with f (u, s) = 1(t,t+) (s)1(t+,T ) (u) , g(u, s) = 1(0,t) (s)1(t,t+) (u)
According to Lemma B.1, there exists a constant C such that:
k f kLwq + kgkLwq C||2/q
q q

and

C1/ .

1(t,t+)
L ([0,T ],ds)

Therefore, since h = 1 < 2/q , we only need to apply Proposition 2.9

to f , g, 1(t,t+) to obtain the continuity of K K+ .

Similarly, we can establish Holder regularity for the operators J and J + .

Proposition 2.11. The operators J and J + are bilinear continuous from

H (X) A,,p,T to H +h (L2 ([0, T ]); ds; X) as soon as 0 < h < h .

Proof. Let 0 t t + T and K be one of K, K+ . Then,

(2.14) t,t+
0
J K (a)(s) = Jt+
K
(a)(s) JtK (a)(s) = Jt+
K
(a)(max(s,t)) .
We plug in the estimate |a(u, s)| kakH w (u, s) for u > s, and get
 

t,t+ J K (a)(s) t,t+
 0 0
J K (w )(s)kakH


and thus the Proposition is an easy consequence of the following Lemma.

Lemma 2.12. The mapping K J K (w ) is linear continuous from A,,p,T
to H +h (L2 ([0, T ]); ds) as soon as 0 < h < h .
 13

Proof. For 0 < t < t + T

K+ K K K
Jt+ (w )(s) Jt + (w )(s) = Jt+
+ +
(1(t,t+) (s)w )(s) + Jt+ (1(0,t) (s)w )(s) .
q
According to Lemma B.1, the norm of (u, s) 1(t,t+) (s)w (u, s) in Lwq
is dominated by C+h . Therefore, Proposition 2.9

C+h kKk,,p,T .
K+
Jt+ (1(t,t+) (s)w )(s) 2

L ([0,T ],ds)

Since K+ A,,p,T , then for h < h ,

C+h kKk
K+
Jt+ (1(0,t) (s)w )(s)

L2 ([0,T ],ds)
Combining these two upper bounds yields the desired result.

The following technical results now involve pairs of rough kernels.

Lemma 2.13. The mapping (K1 , K2 ) hK1 (t, ), K2 (t, )iL2 ([0,T ]) is bilinear
2
continuous from E,,p,T to the set of absolutely continuous functions, van-
ishing at the origin, endowed with the norm
Z T 
kfk =  f(s) ds
0
where f is equal almost everywhere to the derivative of f .
Proof. Assume that the kernels can be written, as in the proof of Proposi-
tion 2.5,
(2.15)
Z T
Ki (t, s) = Ki (T, s) hi (u s, s)w (u + , s) du for s t, i = 1, 2
t
where hi (v, s) = nl=1 l,i (v)l,i (s) and the l,i , l,i L (0, T ) have disjoint
supports. Then the mapping
: t hK1 (t, ), K2 (t, )iL2 ([0,T ])
is differentiable with derivative
Z t
(t)
= K1 (t,t)K2 (t,t) + (1 K1 (t, s)K2 (t, s) + K1 (t, s)1 K2 (t, s)) ds .
0
From the domination relation (2.13) we get
 d hK1,+ (t, ), K2,+ (t, )i 2
(t)
 
(2.16) L ([0,T ])
dt
Integrating with respect to t yields, taking into account Proposition 2.10,

Z T
(t)
dt hK1,+ (t, ), K2,+ (t, )i 2

L ([0,T ]) CkK1 k,,p,T kK2 k,,p,T
0
14 P. CARMONA, L. COUTIN, AND G. MONTSENY

By Proposition 2.5, the space of kernels that can be written as in (2.15) is

dense in E,,p,T . This concludes our proof since the space of absolutely
continuous functions is closed.

We now associate to a pair (K, K 0 ) of rough kernels an integral operator

Z t 
K,K 0 0
(2.17) J2,t (a)(s, s1 ) = a(s, s1 ) K(u, s)1 K (u, s1()) du 1(s1 <s)
s

defined for measurable a : [0, T ]2 X.

Lemma 2.14. If 0 < h < h , then the mapping J2,t is bilinear continuous
+h
from H (X) E A into H +2h (L2 ([0, T ]2 ; X; ds ds )).
,,p,T ,,p,T 1

Proof. According to equality (2.14) applied to

s1 ) = K(u, s)a(s, s1 ) 1(s1 <s)
a(u,
we get
0 0
1(s1 <s) t,t+
0
J2K,K (a)(s, s1 ) = Jt+
K
(a)(max(s,t))
.
Using again the dominating kernels, a H (X) and that
(s s1 ) (u s1 ) for 0 < s1 < s < u
we obtain
0
K,K 0
 
K+
t,t+ J2 (a)(s, s1 ) kakH Jt+
 0
(K+ (u, s)w (u, s1 ))(s1 )1(s>s1 )


According to Proposition 2.10 there exists a constant C such that:

(2.18) (u, s) K+ (u, s)w (u, s1 ) +h 2 CkKk,,p,T .
H (L ([0,T ]))

Therefore, thanks to Proposition 2.11 we have the upper bound

K,K 0

t,t+ J2 (a)(s, s1 ) CkKk,,p,T K 0 kakH +2h .
0
+h
L2 ([0,T ],X,ds) A,,p,T

Lemma 2.15. The mapping (K, K 0 )

hK(t + , ) K(t, ) , K 0 (t, )iL2 ([0,T ]) is bilinear continuous from
2
A,,p,T E,,p,T to R with norm bounded by C+h for any h (0, h ) (we
extend K so that K(u, s) = 0 if u < s).
 15

0 . we have:
Proof. Thanks to (2.13) we can restrict ourselves to K+ , K+


 K+ (t + , ) K+ (t, ) , K+
0

(t, ) L2 ([0,T ]) 
 
Z t Z t+
= 0
ds K+ (t, s) du 1 K+ (u, s)
0 t
Z t Z t+
= 0
ds K+ (t, s) du (u s) (u s) 1 K+ (u, s)
0 t
Z t Z t+
0
ds K+ (t, s)(t s) du (u s) 1 K+ (u, s)
0 t
Z t
= 0
ds K+ (t, s)(t s) t,t+
0
J K+ (w )(s)
0
Z t
 1 Z t
1
2
2 2
0
ds K+ (t, s)2 (t s) 2
ds t,t+
0
J K+ (w )(s)
0 0
+h
(C1 K 0 A ) (C2

kKk,,p,T ) .
,,p,T

The second inequality is Cauchy-Schwarzs inequality, and the last one is

due to Holders inequality (/2, /2) for the first factor (since < 1)
and Lemma 2.12 for the second factor. We can now conclude with the help
of Remark 2.3.

3. T HE SPACE OF INTEGRANDS
3.1. A review of basic notions of Malliavin Calculus. A nice introduc-
tion to Malliavin Calculus can be found in Nualarts book [22], but for the
sake of completeness, we state here the few definitions and properties we
use in this paper.
Let denote the space C(I, R), I = [0, T ], equipped with the topology of
uniform convergence on the compact sets, F the Borel field on , P
the standard Wiener measure, and let {Bt () = (t), 0 t T }. For any
t 0, we define Ft = ((s), s t) N , where N denotes the class of the
elements in F which have zero P measure. For h L2 (I, R), we denote by
B(h) the Wiener integral
Z
B(h) = (h(t), dBt ).
I

Let X be a separable Hilbert space with norm ||X . (Usually X = L2 ([0, T ]i ; R)).
Let S denote the dense subset of L2 (, F , P) consisting of those classes of
random variables of the form :

(3.1) F = f (B(h1 ), ..., B(hn )),

16 P. CARMONA, L. COUTIN, AND G. MONTSENY

where n N, f Cb (Rn ; X), h1 , .., hn L2 (I, R). If F has the form (3.1),
def
we define its derivative as the process DF = {Dt F, t I} given by
n
f
Dt F = xk (B(h1), ..., B(hn))hk (t).
k=1

We shall denote by D1, the closure of S with respect to the norm

kFk1, = kFk + kkDFkL2 (I) k .
" Z
t
/2 #1/
 1/
= E |F| +E |Du F|2 du
0

The higher order derivative Dn F are defined inductively, and the space Dn,
is the closure of S under the norm

n
kFkn, = kFk + Di F L2 (I i ) .


i=1


Then, define , the Skorokhod integral with respect to W , as the adjoint of
D, i.e. , Dom() is the set of u L2 ( I) such that there exists a constant
c with  Z 
 
E Dt Fut dt  ckFk2 , F S .
I

If u Dom, (u) is defined as the unique element of L2 () which satisfies
Z 
E [(u)F] = E Dt Fut dt , F S .
I
In order to prevent a confusion between , an admissible parameter, and
the Skorokhod integral, we shall from now on use the same notations for
the Ito and the Skorokhod integral, that is:
Z T
(u) = u(s) dBs .
0
Let L1,2 be the Hilbert space L2 (I; D1,2 )
endowed with the norm
Z Z 
2 2 2
kukL1,2 = E ut dt + (Dv ut ) dv dt .
I II

We have L1,2 Dom, and for u L1,2

(3.2)
" Z 2 #
T
Z  Z Z 
E u(s) dBs =E 2
ut dt + E Dt us Ds ut ds dt kuk2L1,2 .
0 I I I

Note that {u L2 ( I); u is Ft progressively measurable } Dom, and

for such a u, (u) coincide with the usual Ito integral. Note that when u is
progressively measurable, Ds ut = 0 for s > t, so (3.2) is consistent with the
formula in the adapted case.
 17

Let L2,2 be the Hilbert space L2 (I 2 , D2,2 ) endowed with the norm
(3.3)
Z Z Z 
kuk2L2,2 =E 2
u (t, s) dtds + 2
(Dv u(t, s)) dvdtds + (D2v,r u(t, s))2 dtdsdvdr .
I2 I3 I 4

We have L2,2 Dom( ) and for u L2,2

Z T Z s  
(3.4) E u(s, s1 ) dBs1 dBs 2kuk2L2,2 .
0 0

Property P : Integration by parts formula Suppose thatu belong to L1,2 .

Let F be a random variable belonging to D1,2 such that E F 2 I ut2 dt < ,
R

then
Z Z Z
(3.5) Fut dBt = F ut dBt Dt Fut dt,
I I I
in the sense that Fu belong to Dom() if and only if the right hand side of
(3.5) belongs to L2 ().

For m = 1, 2, let C Lm,2 (0,t) = H (Dm,2 ) Lm,2 be the space of processes

a Lm,2 (0,t) such that for a constant C
(3.6) ka(u) a(v)kDm,2 C|u v| (0 u, v t).

3.2. The space of good integrands. Recall that (p, , ) is a set of admis-
sible parameters, that is

1 1 2 2p
(3.7) 1 < p < 2, + = 1, > > 1/q, =
p q q 2 p
1 2 1 1
(3.8) h = < , + = 1.
q /2 /2
The space GI,,q,t is the space of adapted processes a L1,2 (0,t) such that
kak,,q,t < where

Z 1/q
def q
kak,,q,t = (ka(u) a(s)kD1,2 (u s) ) du ds
0<s<u<t
Z t
1/
ka(s)k2 ds

+
0

(we hope that the use of the same notations form norms of kernels and
norms of integrands will not confuse the reader). With the notations intro-
duced in the previous section,
kak,,q,t = 0 a Lq (D1,2 ) + kakL (L2 ()) .

w q
18 P. CARMONA, L. COUTIN, AND G. MONTSENY

Observe that if a C L1,2 (0,t) is adapted and bounded, and if > q1 ,

then a GI,,q,t .
Finally, we need to introduce the subset CGI,,q,t of processes a GI,,q,t
such that Z tZ t  q/2
E (Ds a(u))2 (u s)q duds < +
0 s
which we endow with the norm

|kak|,,q,t = kak,,q,t + kDakLwq 2 ()) .

q (L

4. I NTEGRATION OF DETERMINISTIC FUNCTIONS

We recall that to a rough kernel K we associate the integral operator defined
for a measurable a on (0,t)
Z t
(4.1) It a(s) = K(t, s)a(s) + (a(u) a(s)) 1 K(u, s) du (0 < s < t)
s

as soon as the integral on the right hand side makes sense for almost every
s in (0,t).
When K is a smooth kernel, we have
Z t
It a(s) = K(s, s)a(s) + a(u)1 K(u, s) du (0 < s < t) .
s
We shall also consider the integral operator
Z t
(4.2) Jt a(s) = a(u, s) 1 K(u, s) du (0 < s < t) .
s

We end this section by the definition of the integral of a deterministic func-

tion. Since K(t, s) = It 1[0,t] (s), we have, for s < t,

t Z s
Z 
Cov(WK (s),WK (t)) = E K(t, u) dBu K(s, v) dBv
0 0


= It 1[0,t] , It 1[0,s] L2
where h,iL2 denotes the inner product of L2 (R+ ).
Accordingly, we define LK2 (0,t) = It1 (L2 (0,t)) and endow it with the inner
product
def
(4.3) h f , giL2 (0,t) = hIt f , It giL2 .
K

LK2 (0,t) is the closure of LK2 (0,t) with respect to h, iL2 (0,t) . Given f
K
LK2 (0,t), it is natural to define
Z t Z t
(4.4) f (s) dWK (s) = It f (u) dBu .
0 0
 19

This is clearly an isometry between LK2 (0,t) and the Gaussian space gener-
ated by (WK (s), 0 s t). It is interesting to note that for f 1 we obtain
Z t
dWK (s) = WK (t) .
0

5. R EGULARITY OF SAMPLE PATHS

Proposition 5.1. Assume that K E,,p,t . Then, almost surely, the sample
paths of WK are Holder continuous of index (0, h ). More precisely, if
> h1 and [0, h 1 [ then for some constant C

|WK (v) WK (u)|
sup CkKk,,p,t .


0u<vt |v u|

Proof. Since WK is a Gaussian process with covariance

Z min(u,v)
(5.1) RK (u, v) = K(v, s)K(u, s) ds
0
Proposition 2.10 yields for 0 u < v < t

kWK (v) WK (u)k2 = kK(v, ) K(u, )kL2 (0,t) C|v u|h kKk,,p,t .
Then, we use Kolmogorovs continuity Lemma in the special case of Gauss-
ian processes, as described by the next Lemma.
Lemma 5.2. Let (Z(t), 0 t T ) be a centered Gaussian process such
that for some > 0
E (Z(t) Z(s))2 C(T )|t s| (0 s,t T ) .
 

Then Z has a modification (denoted by the same letter Z) with Holder con-
tinuous paths of order for each [0, /2[.
Furthermore, for every p sup(1, 2/) we have

21+1/p
T /2 m p
1
sup |Z(t) Z(s)| C(T ) 2

1
/2
0s,tT
p 12 p

|Z(t) Z(s)| 1
sup C0 (, , p, T )C(T ) 2 (0 < /2 1/p).

0s<tT (t s)
p
p
Where mp
is the p-th moment of the absolute value of a standard Gaussian
random variable.
Since the proof of this Lemma uses the explicit constant of Kolmogorovs
Lemma, we feel compelled to state here the version of this lemma that we
use, which can be found in [10], section XXIII, numbers 19 and 20.
20 P. CARMONA, L. COUTIN, AND G. MONTSENY

Lemma 5.3. Let (X(t), 0 t T ) be a real process such that for some
> 0 and some p 1

E [|Xt Xs | p ] C|t s|1+ (0 s,t T ) .

Then for every [0, /p[ the random variable
( )
|Xt Xs |
M = sup : s,t Q, s 6= t, 0 s,t T
|t s|
is almost surely finite. More precisely,
2+1+1/p 1+
kM k p C1/p T p .
1 2/p
Proof of Lemma 5.2. For every p 1, since Z(t)Z(s) is a centered Gauss-
ian random variable, we have

E [|Z(t) Z(s)| p ] m ppC(T ) p/2 |t s|p/2 (0 s,t, T ) .

If p > sup(1, 2/), then we obtain the upper bound on the p-th moment, and
the fact that Z has Holder continuous paths for every [0, /2 1/p[.

Corollary 5.4. Assume that for some p 1, K E,,p,t . Then there exists
a sequence of kernels (Kn , n N) in F,,p,t such that

sup |WK (s) WKn (s)| 0 in L2 and a.s.

st

Proof. There exists a sequence of kernels (Kn , n N) in F,,p,t such that

kKn Kk,,p,t 0. We let

Mn = sup |WK (s) WKn (s)| .

st

Proposition 5.1 yields, for > 1

h and = 0,

kMn k CkK Kn k,,p,t 0 .

By taking a subsequence, we can ensure that kMn k < +, and thus ob-
tain, by a Borel-Cantelli argument, the almost sure convergence of Mn to
0.

6. C ONSTRUCTION OF THE S TOCHASTIC INTEGRAL

The
R first step of our construction is to identify the semi-martingale integral
a(s)dWK (s), for a adapted and WK a semi-martingale, with a sum of (Sko-
rokhod) integrals of a and its stochastic gradient with respect to the driving
Brownian motion B and time.
 21

Proposition 6.1. Assume that a GI,,q,t and K is a semi-martingale ker-

nel in E,,p,t . Then
(i) The process It (a)() is in L1,2 (0,t) and
Z t

It (a)(s) dBs kIt (a)k 1,2
0 L (0,t) kak,,q,t kKk,,p,t .

(ii) We have the decomposition

Z t Z t Z t Z t 
(6.1) a(s)dWK (s) = It a(s) dBs + Ds a(u)1 K(u, s) du ds .
0 0 0 s
(iii) If furthermore a CGI,,q,t then the process
Z t
s Jt (D (a)())(s) = Ds a(u)1 K(u, s) du
s
satisfies
Z t Z t

Jt (D (a)())(s)
0 kJt (D (a)())(s)kL2 () ds
0

L2 ()
C|kak|,,q,t kKk,,p,t .
and we have the upper bound
Z t

(6.2) a(s)dWK (s)
0 C|kak|,,q,t kKk,,p,t .
L2 ()

Proof. (i) On the one hand, Lemma 6.2 (ii) implies that Jt (a)() L1,2 (0,t)
and

kJt (a)kL1,2 (0,t) kakLwq 1,2 ) kKk,,p,t .

q (D

On the other hand, the second part of Proposition 2.9 implies that
kK(t, )a()kL2 ([0,T ],L2 (,P)) CkakL (L2 ()) kKk,,p,T .

We conclude by combining these two results.

(ii) Recall from Proposition 2.5 that WK is a semi-martingale with decom-
position
dWK (s) = K(s, s) dBs +W1 K (s) ds .
Therefore, since a is adapted and in L1,2 (0,t), we have that
Z t Z t Z t
a(s)dWK (s) = a(s)K(s, s) dBs +
a(s)W1 K (s) ds
0 0 0
Z t Z t Z u 
= a(s)K(s, s) dBs + a(u) 1 K(u, s) dBs du .
0 0 0
22 P. CARMONA, L. COUTIN, AND G. MONTSENY

We now apply the integration by parts formula of Malliavin Calculus

(6.3)
Z t Z t Z t Z u 
a(s)dWK (s) = a(s)K(s, s) dBs + a(u)1 K(u, s) dBs du
0 0 0 0
Z t Z u 
+ Ds a(u)1 K(u, s) ds du.
0 0

def
Since M 2 = supxt 0x 1 K(x, u)2 du < +, we can apply Fubinis Theo-
R

rem to the third term on the right hand side of (6.3) to obtain
Z t Z u  Z t Z t 
Ds a(u)1 K(u, s) ds du = Ds a(u)1 K(u, s) du ds .
0 0 0 s
Indeed,
Z t Z u
 
E |Ds a(u)||1 K(u, s)| ds du
0 0
" Z Z
t u
1 #
2
2
ME |Ds a(u)| ds du
0 0

MkakL1,2 .
The anticipating Fubinis Stochastic Theorem (see Theorem 3.1 of Leon [17])
yields that
Z t Z s  Z t Z t 
a(u)1 K(u, s) dBs du = a(u)1 K(u, s) du dBs .
0 0 0 s

Indeed, all we need to do is to consider the finite measure (dx) = dx 1[0,t] (x)
on X = [0,t] and the measurable function
(x, u, ) = a(x, )1 K(x, u) (x, u t) .
It is clear that for fixed x [0,t], (x) : u (x, u, .) is in L1,2 (0,t), with
Dv (x, u) = Dv a(x)1 K(x, u) and
Z t  Z t  
2
k(x)kL1,2 = E (x, u) + (Dv (x, u)) dv du
2 2
0 0
Z x  Z x 
= du 1 K(x, u) E a(x) + (Dv a(x)) dv .
2 2 2
0 0

Therefore the Bochner integral X (x) d(x) is well defined since

R

Z 2 Z
(dx)k(x)kL1,2 (X) (dx)k(x)k2L1,2
Z t Z x Z x 
(X)E 2
dx(a(x) + (Dv a(x)) dv) 2
1 K(x, u) du
2
0 0 0
 Z x 
(X) sup 1 K(x, u) du kakL1,2 (0,t) < + .
2
xt 0
 23

(iii) The first part is a consequence of the first part of Proposition 2.9 ap-
s) = Ds a(u). We conclude by combining this upper
plied to the process a(u,
bound with (i).

Rt
Lemma 6.2. (i) For every a L1,2 (0,t) such that s s a(v)1 K(v, s) dv is
in L1,2 we have the commutation relation
Z t Z t
Du a(v)1 K(v, s) dv = Du a(v)1 K(v, s) dv .
s s
(ii) The application (a, K) JtK (a) is bilinear continuous from
q
L1,2 (0,t) Lwq (Dm,2 ) E,,p,T to Lm,2 with m = 0, 1, 2.

Proof. (i) The simplest way to see this property is to use the Wiener chaos
expansion of a
a(s) = m( fm(., s)),
m0

where m denotes the m-th multiple Wiener integral and fm (s1 , . . . , sm , s)

are square integrable kernels symmetric in the first m variables. By linearity,
we can restrict ourselves to the case a = m ( fm (., s)) for some m 1.
On the one hand, Fubinis Stochastic Theorem implies that
Z t Z t
a(v)1 K(v, s) dv = dv 1 K(v, s)m ( fm (, v))
s s
 Z t 
= m 1 K(v, s) fm (, v) dv .
s

Therefore,
Z t  Z t 
Du a(v)1 K(v, s) dv = mm1 1 K(v, s) fm (, u, v) dv .
s s

On the other hand,

Z t Z t
Du a(v)1 K(v, s) dv = 1 K(v, s)mm1 ( fm (, u, v)) dv
s s
 Z t 
= mm1 1 K(v, s) fm (, u, v) dv .
s

(ii) For t [0, T ], Proposition 2.9 implies

(6.4)
s JtK (a(., s))(s) 2 + (r, s) JtK (Dr a(., s))(s) L2 ([0,T ]2 )

L ([0,T ])
CkakLwq 1,2 kKk,,p,T .
q (D )
24 P. CARMONA, L. COUTIN, AND G. MONTSENY

We shall now use a by product of the proof of Proposition 2.5, namely the
existence of a sequence of semi-martingale kernels (n )nN given by
Z T
1
n (r, s) = K(T, s) hn (u s, s)(u + s) du
r n
N
hn (v, s) = ni (v)ni (s) ,
i=0

where ni , ni are in L , the i ni have disjoint support and limn kK n k,,p,T =

0.
It is clear (see (i)) that for every n, the random variable
Z t
1
Jtn (a(, s))(s) = a(u, s)hn (u s, s) (u + s) du
s n
belongs to D1,2 and that

(6.5) Dr Jtn (a(, s))(s) = Jtn (Dr a(, s))(s) .

From the inequality (6.4) applied to the kernels n K, n m we de-
duce that the sequence (Jtn (a(, s))(s), s [0, T ]) is a Cauchy sequence in
L1,2 (0, T ) that converges to (JtK (a(, s))(s), s [0, T ]) in L2 ( [0, T ]).
Thanks to (6.5), we obtain also that (Dr Jtn (a(, s))(s), (s, r) [0, T ]2 ) con-
verges to (J K (Dr a(., s))(s), (s, r) [0, T ]2 ) in L2 ( [0, T ]2 ).
Therefore, (JtK (a(, s))(s), s [0, T ]) is in L1,2 (0, T ) and we have the com-
mutation relation.
(6.6) Dr JtK (a(, s))(s) = JtK (Dr a(, s))(s) .
It is now clear that (6.4) gives the proof of the second part of the Lemma
for m = 1. For m = 2, we only need to replace a by Da.

Combining the density of F,,p,t in E,,p,t (see Proposition 2.5) with Propo-
sition 6.1 yields that the continuous bilinear operator t : CGI,,q,t F,,p,t
L2 :
Z t
t (a, K) = a(s) dWK (s)
0

can be uniquely extended to an operator t : CGI,,q,t E,,p,t L2 , de-

fined on the closure of this product space.

Definition 6.3. For a CGI,,q,t and K E,,p,t the stochastic integral of

a with respect to WK is defined to be
t (a, K) and denoted by R t a(s) dWK (s).
0
By construction we have the decomposition (6.1) and the upper bound (6.2).
 25

7. T HE STOCHASTIC INTEGRAL AS A PROCESS

Rt
We shall exhibit assumptions that ensure that the process t 0 a(s) dWK (s)
has a continuous modification.

Theorem 7.1. Assume that for some > 0 such that + h > 1
2 :

K A,,p,T .
the adapted integrand a is in C L1,2 (0, T ).
supsT |a| L , for h > 1.
Then the process t 0t It a(s) dBs has a continuous
R
modification on [0, T ].
If, furthermore, CGI,,q,T then t 0 a(s) dWK (s) has a continuous
Rt

modification on [0, T ].

Proof. Recall that the stochastic integral may be represented as

Z t
It (a)(s) dBs = X(t) +Y (t)
0
Z t
X(t) = K(t, s)a(s) dBs
0
Z t
Y (t) = Jt (0 a)(s) dBs .
0

The continuity of X is established via Kolmogorovs continuity criterion.

More precisely for > 0 big enough, there exists > 0 and C = C(T, )
such that
E |X(t + ) X(t)| C1+ (0 t < t + T ) .
 
(7.1)
First we write
Z t Z t+
X(t +)X(t) = (K(t +, s)K(t, s)) a(s)dBs + K(t +, s)a(s)dBs
0 t
Since a is adapted, we can apply Burkholder-Davis-Gundy inequalities to
the martingales
Z
r [K(t + , s) K(t, s)]1(0,t) (s) a(s) dBs
0
Z
r K(t + , s)1(t,t+) (s) a(s) dBs
0
to obtain, for a constant C , the upper bound
" Z
t
/2 #1/
E |X(t + ) X(t)|
1/
[K(t + , s) K(t, s)]2 a(s)2 ds
 
C E
0
" Z
t+
/2 #1/
+C E K(t + , s)2 a(s)2 ds
t
26 P. CARMONA, L. COUTIN, AND G. MONTSENY

The integrability assumptions on a, and Proposition 5.1 implies, for > 2,

kX(t + ) X(t)kL () C ksupsT |a(s)|kL () kK(t + , ) K(t, )kL2 ([0,T ])

CksupsT |a(s)|kL () kKk,,p,T h .

Therefore we only need to assume also h > 1 to obtain (7.1). The exis-

tence of a continuous modification for the process Y is given by the Propo-

sition 7.2.

If, furthermore, CGI,,q,T , then the stochastic integral may be repre-

sented as

Z t Z t Z t
a(s) dWK (s) = It (a)(s) dBs + Z(t) , Z(t) = Jt (D (a))(s)ds
0 0 0

The process Z is obviously a continuous process so that ends the proof.

Proposition 7.2. Let > 0 such that +h > 12 . Assume that a H (D1,2 )

and K A,,p,T . Then the process t 0t JtK (a(, s))(s) dBs has a continu-
R

ous modification.
In particular, if a C L1,2 (0, T ) then a(u,
s) = 0s,u a = a(u) a(s) is in
H (D1,2 ) and t 0t JtK (0 a)(s) dBs has a continuous modification.
R

Proof. According to Lemma 6.2 (ii) we have Dr J K (a)(s) = J K (Dr a)(s) and

+h .
K
Jt+ (a(, s))(s) JtK (a(, s))(s) 1,2 Ckak 1,2 kKk

L H (D ) A ,,p,T

Since 2( + h ) > 1 we can conclude by Kolmogorovs continuity criterion.

8. I T O S FORMULA
Itos formula for fractional Brownian motion of Hurst parameter H > 12 is
now well known: see, for instance, Decreusefond and Ustunel [9], Dai and
Heyde [8]. Here, we show how to obtain it for a wide family of kernels.
The first step of our method is to write Itos formula for a semi-martingale
kernel in a suitable way, that is a way that will be easily extended to more
general kernels.
 27

Proposition 8.1. Let K be a semi-martingale kernel and f Cb2 . Then,

almost surely, for every t > 0:
Z t
(8.1) f (WK (t)) = f (0) + It ( f 0 (WK ))(s) dBs
0
Z t
1 d 
f 00 (WK (s)) E WK (s)2 ds .

+
2 0 ds

Proof. We let (s) = E WK (s)2 = 0s K(s, r)2 dr . Since K is a semi-martingale

  R
kernel, is differentiable with
Z s
(s) = K(s, s) + 2
0 2
K(s, r)1 K(s, r) dr .
0

By the chain rule,

Ds f 0 (WK (u)) = f 00 (WK (u))Ds (WK (u)) = f 00 (WK (u))K(u, s) ,

and therefore the process f 0 (WK (s)) is in L1,2 (0,t) and we may use Propo-
sition 6.1. Furthermore,
Z t Z t Z t
0
Jt (Ds f (WK ))(s) ds = ds dv 1 K(v, s)Ds f 0 (WK (v))
0 0 s
Z t Z t
= ds dv 1 K(v, s) f 00 (WK (v))K(v, s) .
0 s

Recall that the semi-martingale decomposition of WK is

dWK (s) = K(s, s) dBs +W1 K (s) ds .

Therefore, Itos formula yields

Z t Z t
1
f (WK (t)) = f (0) + f 0 (WK (s)) dWK (s) + f 00 (WK (s))K(s, s)2 ds
0 2 0
Z t Z t
= f (0) + It ( f 0 (WK ))(s) dBs + Jt (Ds f 0 (WK ))(s) ds
0 0
1
Z t
+ f 00 (WK (s))K(s, s)2 ds
2 0
Z t Z t
1
= f (0) + It ( f 0 (WK ))(s) dBs + f 00 (WK (s))0 (s) ds .
0 2 0

We shall make use of B,,p,T defined in 2.6.
28 P. CARMONA, L. COUTIN, AND G. MONTSENY

Theorem 8.2. Assume that for some < h the kernel K is in B,,p,T for
every T > 0, that f Cb4 , and that h > 41 . Then, almost surely, for every
t > 0,
Z t
(8.2) f (WK (t)) = f (0) + It ( f 0 (WK ))(s) dBs
0
Z t
1 d 
f 00 (WK (s)) E WK (s)2 ds .

+
2 0 ds

Proof. Our first step will be to prove that the processes on both sides have
a continuous modification. Then, we shall establish this formula for fixed
t > 0.
Since WK has a continuous modification, see Proposition 5.1, f (WK ) also
has a continuous modification. It is obvious that
1 t 00 d 
Z
f (WK (s)) E WK (s)2 ds ,

t f (0) +
2 0 ds
also has a continuous modification.
In fact, Proposition 5.1 ensures that a.s. and in L2 , sample paths of the
Gaussian process WK are Holder continuous with index < h . Since f 0 , f 00
and f (3) are bounded, Proposition 8.4 ensures that a = f 0 (WK ) belongs to
C L1,2 (0, T ) L ([0, T ] ) as soon as < h . We furthermore impose
that > 12 h > 1/q, which is possible since h > 1/4; then, accord-
ing to Theorem 7.1, the process t 0t It ( f 0 (WK ))(s) dBs has a continuous
R

modification.
Let us now establish formula (8.2) for a fixed t > 0. Assume that Kn is

a family of semi-martingale kernels such that Kn K in B,,p,T . Then
formula (8.2) is valid for Kn .
We shall first handle the term
Z t
d 
f 00 (WKn (s)) E WKn (s)2 ds .

0 ds
According to the proof of Corollary 5.4, by taking a subsequence, we may
assume that supst |WKn (s) WK (s)| 0 both in L2 and almost surely. There-
fore according to Lemma 2.13:
Z t Z t
00 d  d 
2
f 00 (WK (s)) E WK (s)2 ds .
 
f (WKn (s)) E WKn (s) ds
0 ds 0 ds
To prove the convergence
Z t Z t
0
It f (WKn )(s) dBs It f 0 (WK (s)) dBs
0 0

we need to prove that ItKn an

It a in L1,2 (0,t) where a(s) = f 0 (WK (s))
0
and an (s) = f (WKn (s)). Thanks to Proposition 6.1, and to the fact that
 29

Kn K in B,,p,T , it is enough to prove that ka an k,,q,t 7 0, and this
n
we achieve in the following Proposition 8.4.
Definition 8.3. From now on, 0 F,1 F will denote the Taylor expansion
of the function F:
0x,y F = F(y) F(x) , 1x,y F = F(y) F(x) (y x)F 0 (x) .

Proposition 8.4. Assume f Cbm+i+2 with i = 0, 1 and m = 0, 1, 2. Then the

mapping K W i
K (s),WK (u)
f is continuous from E,,p,T to H (i+1)h (Dm,2 ).

Proof. Let X be a separable Hilbert space.

Definition 8.5. A function G : R2 [0, T ]2 X is said to satisfy assump-
if G is m-times differentiable with respect to its first two
tion HH (m, k, k)
variables, and if there exists a constant C(G) = C(G, m, k, k) such that for
every real numbers (x, y) R2 , 0 u, s T , integers i, j such that 0
i + j m, we have
j
i, j G(x, y, u, s) C(G) |y|
 
 i+ j max(kl,0)
|u s|k .

X l=0

Observe that if f CbM+m , then G(x, y, u, s) = M1 1 1 (M) R

x,x+y f = M! 0 f (x +
y)y d satisfies HH (m, M, 0). Therefore Proposition 8.4 is an immediate
M

consequence of the following Lemma.

Lemma 8.6. Assume 1 and G satisfies HH (m, k; k).
Then for every
K E,,p,T , 0 u, s T ,
m
G(WK (s), 0s,uWK , u, s) n, C C(G) kKkk+ j |u s|kh +k .

D (X) ,,p,T
j=0

Proof. The proof is an induction on the integer m.

Assume that G satisfies HH (0, k, k). Then,
G(WK (s); 0s,uWK ; u, s) C(G)0s,uWK k |u s|k .
   
X
Since the random variable s,uWK is centered with covariance 0s,u K(, r) L2 ([0,T ],dr)
0

with the convention K(u, r) = 0 for u < r, we deduce from Proposition 2.10
the upper bound
s,u K(, r) 2 CkKk,,p,T |u s|h .
0
(8.3) L ([0,T ],dr)
We plug in the expression of the k-th order moment of a centered Gaussian
random variable, to obtain that the Lemma is satisfied for a constant
k
C0 = C N (0, 1) kh .

30 P. CARMONA, L. COUTIN, AND G. MONTSENY

We now assume the Lemma satisfied for m and let G verify assumption
Then G is differentiable with respect to the first two vari-
HH (m + 1, k, k).
ables up to the order m + 1, and G together with its partial derivatives have
at most polynomial growth. According to the chain rules of Malliavin Cal-
culus, the random variable G(WK (s), 0s,uWK ; u, s) belongs to Dm+1, and
Dm+1 G(WK (s), 0s,uWK ; u, s) = Gi, j, (WK (s), 0s,uWK ; u, s) ,
 
Sm+1 ,i+ j=m+1
where Sn denotes the group of order n permutations, the tensor product
and
(8.4)
1 i+ j
G(i, j, )(x, y, u, s)(r1 , . . . , rm+1 ) = G(x, y, u, s)i K(s, )(r(1) , . . . , r(i) )
i! j! i, j
j 0s,u K(, )(r(i+1) , . . . , r(m+1) ) .
Injecting the upper bound 8.3, we get that G(i, j, ) satisfies HH (0, k
j, k + jh ) on X = L2 ([0, T ]m+1 ). We can therefore apply the Lemma for
m = 0 to every G(i, j, ) to obtain
(8.5) Dm+1 G(WK (s), 0s,uWK ; u, s) L (X)
 

CC(G, m + 1) kKk,,p,T |u s|kh +k .

k+ j

We conclude by using the definition of the norm in D,m+1 and the induction
assumption.
Lemma 8.7. Assume that G satisfies HH (m, k, k) with m 1, k > 0. Then

the mapping K G(WK (s), 0s,uWK , u, s) is continuous from E,,p,T to H kh +k (Lm1,2 )
(with the convention L0,2 = L2 ()).
Proof. Using Taylor integral expansion we have for K, K 0 E,,p,T
Z 1
(8.6) G(WK (s); 0s,uWK ; u; s) G(WK 0 (s); 0s,uWK 0 ; u; s) = B d ,
0
with
t
B = G(WK+(KK 0 ) (s); 0s,uWK+(KK 0 ) ; u, s) WKK 0 (s); 0u,sWKK 0


where xt denotes the transpose of x.

For i = 1, 2, the mapping i G(x, y, u, s) satisfies HH (m 1, k + 1 i, k).

Therefore Lemma 8.6 implies

(8.7) i G(WK+(KK 0 ) ; 0s,uWK+(KK 0 ) ; s, u) Lm1,2

m1
(kKk,,p,T + K 0 ,,p,T ) .
k+ j k+ j
C|u s|(k+1i)h +k
j=0
The same Lemma 8.6 yields
 31

kWKK 0 (s)kLm1,2 C K K 0 ,,p,T

(8.8)
s,uWKK 0 m1,2 C K K 0 |u s|h .
0
L ,,p,T
Combining (8.7) and (8.8) yields,

m1

k+ j 0 k+ j
0
|u s|kh +k

kB kLm1,2 C (kKk ,,p,T + K
,,p,T
) K K
j=1
and injecting this into (8.6) yields the desired result.

8.1. Application to fractional Brownian motion. According to Lemma 2.7

we can take h = H with > 0 small enough so that h > 41 and then use
Theorem 8.2.
8.2. Ito formula for H > 61 . We first derive a new expression for the Ito
formula for the Gaussian processes associated to semi-martingale kernels.
Proposition 8.8. Let K a semi-martingale kernel and f Cb4 . Then, for
t [0, T ] almost everywhere,
Z t
1 d
f (WK (t)) = f (0) + f 00 (WK (s)) RK (s, s)ds
2 0 ds
Z t
K(t, s) f 0 (WK (s)) + f 00 (WK (s))E [WK (t) WK (s) | Fs ] dBs


+
0
Z t
K(t, s) f 000 (WK (s))E [(WK (t) WK (s))WK (s)] dBs
0
Z t h i
+ ItK W 1
K (s),WK ()
f 0
+ f 000
(WK (s))E [(WK () W K (s))W K (s)] (s) dBs
0
Z t Z s 
Zt
 
00 00
f (WK (s)) f (WK (s1 )) 1 K(u, s1 )K(u, s) du dBs1 dBs
0 0 s
with RK the covariance function of the Gaussian process WK
RK (x, y) = E [WK (x)WK (y)] ,
Proof. Let K a semi-martingale kernel and f Cb4 . According to Proposi-
tion 8.1, for t [0, T ], we have
Z t
f (WK (t)) = f (0) + ItK ( f 0 (WK ))(s)dB(s)
0
1 t 00 d 
Z
f (WK (s)) E WK (s)2 ds,

+
2 0 ds
0
A Taylor expansion of f between WK (u) and WK (s) yields
ItK ( f 0 (WK ))(s) = K(t, s) f 0 (WK (s)) + ItK (W
1
K (s),WK ()
f 0 )(s)
+ f 00 (WK (s))ItK (0s,WK )(s).
32 P. CARMONA, L. COUTIN, AND G. MONTSENY

Since f Cb4 , the process (a(s) = f 00 (WK (s)), 0 s T ) belongs to L2,2

and
Z s Z t Z t
d
Ds1 a(s) K(u, s)1 K(u, s1 )duds1 = f 000 (WK (s)) K(u, s) E [WK (u)WK (s)] du.
0 s s ds
Using the integration by parts formula,
Z s Z t
Ds1 a(s) K(u, s)1 K(u, s1 )duds1
0 s

= f 000 (WK (s)) K(t, s)E 0s,t WK WK (s)
 

JtK (E 0s,WK WK (s) )(s) .
 

Then Lemma 8.9 yields,

1 t 00 d
Z
f (WK (t)) = f (0) + f (WK (s)) RK (s, s)ds
2 0 ds
Z th i
f 0 (WK (s)) + f 00 (WK (s))E 0s,t WK | Fs K(t, s) dBs

+
0
Z th i
f (WK (s))E s,t WK WK (s) K(t, s) dBs
000
 0
+
0
Z t h i
+ ItK W 1 0 000

 0
K (s),WK ()
f + f (WK (s))E W W
s, K K (s) (s) dBs
0
Z t Z s 
+ WK (s),WK (s1 ) f Jt (K(, s))(s1 ) dBs1 dBs
0 00 K
0 0

Lemma 8.9. Let K be a semi-martingale kernel and a L2,2 an adapted

process. Then
Z t Z t
K 0
a(s)E 0s,t WK | Fs K(t, s) dBs
 
a(s)It (s,WK )(s) dBs =
0 0
Z t Z s  Z t  
+ s,s1 a
0
K(u, s)1 K(u, s1 ) du dBs1 dBs
0 0 s
Z t Z s  Z t  
Ds1 a(s) K(u, s)1 K(u, s1 ) du ds1 dBs .
0 0 0

Proof. We split WK (u) WK (s) = 0s,uWK in two parts:

0s,uWK = E 0s,uWK | Fs + 0s,uWK E 0s,uWK | Fs
   

Observe that WKE = (E 0s,t WK | Fs , s t T ) is a centered Gaussian pro-

 
cess that can be written, thanks to Fubinis Stochastic Theorem,
 33

Z s
WKE (t) = 0s,t K(, s1 ) dBs1
0
Z t Z us 
= du 1 K(u, s1 ) dBs1
s 0
The integration by parts formula (applied to the semi-martingale WKE and
the deterministic kernel K(, s)) yields
Z t Z t
K(t, s)WKE (t) = K(s, s)WKE (s)+ WKE (u)1 K(u, s) du+ K(u, s)dWKE (u)
s s
Since a is adapted and WKE (s) = 0, we obtain
Z t
a(s)E 0s,uWK | Fs 1 K(u, s) du
 
s
Z t
= a(s)E s,t WK | Fs K(t, s) a(s)1(s,t) (u)K(u, s)dWKE (u)
 0 
0
= a(s)E s,t WK | Fs K(t, s)
 0 
Z t Z s 
a(s)K(u, s) 1 K(u, s1 ) dBs1 du .
s 0
The Integration by parts formula yields then, as in the proof of Proposi-
tion 6.1
Z t
a(s)E 0s,uWK | Fs 1 K(u, s) du = a(s)E [WK (t) WK (s) | Fs ]K(t, s)
 
s
Z s Z t 
a(s) K(u, s)1 K(u, s1 ) du dBs1
0 s
Z s Z t 
Ds1 a(s)K(u, s)1 K(u, s1 ) du ds1 .
0 s

Integrating with respect to dB(s) each term of this equality we obtain,

Z t Z t 
(8.9) a(s) E [WK (u) WK (s) | Fs ]1 K(u, s) du dBs =
0 s
Z t
a(s)E [WK (t) WK (s) | Fs ]K(t, s) dBs
0
Z t Z s  Z t  
a(s) K(u, s)1 K(u, s1 ) du dBs1 dBs
0 0 s
Z t Z s Z t  
Ds1 a(s)K(u, s)1 K(u, s1 ) du ds1 dBs
0 0 s
Observe that for s t,
Z t
WK (u) WK (s) E [WK (u) WK (s) | Fs ] = K(u, s1 )dB(s1 )
s
34 P. CARMONA, L. COUTIN, AND G. MONTSENY

and using again stochastic integration by parts and Fubinis Stochastic The-
orem for deterministic integrands, we obtain that
Z tZ t
a(s)(WK (u) WK (s) E [WK (u) WK (s) | Fs ])1 K(u, s)du dBs =
0s
Z t Z s1 Z t 
a(s) K(u, s1 )1 K(u, s) du dBs dBs1 .
0 0 s
Summing up this equality with (8.9) yields, after one substitution s s1 ,
Z t
a(s)ItK (WK () WK (s))(s) dBs = a(s)E [WK (t) WK (s) | Fs ]K(t, s)
0
Z t Z s  Z t  
(a(s) a(s1 )) K(u, s)1 K(u, s1 ) du dBs1 dBs
0 0 s
Z t Z s  Z t  
Ds1 a(s) K(u, s)1 K(u, s1 ) du ds1 dBs .
0 0 0

Proposition 8.10. Assume + 2h > 12 . Then the mapping

Z t
0
(K, K , a) a(s, s1 )JtK (K 0 (, s))(s1 )1(s1 <s) = a(s, s1 )1(s1 <s) K 0 (u, s)1 K(u, s1 ) du
s

is multilinear continuous from A,,p,T E,,p,T Lw (D2,2 ) to L2,2 . In
particular, there exists a constant C such that

Z t Z s
K 0


dB(s)
0 a(s, s1 )Jt (K (, s))(s1 ) dB(s1 )
0

L2 0
CkakH kKkA K
,,p,T
.
,,p,T

Moreover, the process

Z t Z s Z t 
0
(8.10) t dB(s) a(s, s1 ) K (u, s)1 K(u, s) du dBs1
0 0 s

has a continuous modification on [0, T ].

Proof. According to the definition (2.17) of J2,t we have

K,K 0
(8.11) a(s, s1 )JtK (K 0 (, s))(s1 )1(s1 <s) = J2,t (a)(s, s1 ) .
Hence, Lemma 2.14 implies that
0
K ,K 0
J2,t (a)(s, s1 ) 2,2 CkakH kKkA K .

L ,,p,T
,,p,T

Moreover, thanks to the same Lemma 2.14, if

 35

Z t Z s
a(s, s1 )J K (K 0 (, s))(s)1(s1 <s) dBs1


X(t) = dB(s)
0 0
then we have, for 0 t t + T ,

kX(t + ) X(t)kL2 CkakH kKkA +2h

0
K
,,p,T
,,p,T

Since 2( + 2h ) > 1, Kolmogorovs continuity criterion ensures the exis-

tence of a continuous modification.

We can now state and prove an Ito Formula that applies to the family of
Gaussian processes WK such that h > 16 and therefore, by Lemma 2.7 to
fractional Brownian motion with Hurst index H > 16 .

Proposition 8.11. Let K B,,p,T with /2 > h > 61 and let f Cb6 . Then
almost surely, for 0 < t T ,
1 t 00 d 
Z
f (WK (s)) E WK (s)2 ds

f (WK (t)) = f (0) +
2 0 ds
Z t
f 0 (WK (s)) + f 00 (WK (s)E [WK (t) WK (s) | Fs ] K(t, s) dB(s)

+
0
Z t
f 000 (WK (s))E [(WK (t) WK (s))WK (s)] K(t, s) dB(s)


0
Z t  
+ ItK W
1
K (s),WK ()
f 0
+ f 000
(WK (s))E [(WK () WK (s))WK (s)] (s) dBs
0
Z t Z s
+ dBs 0s,s1 f 00 (WK ) ItK (K(, s))(s1 ) dBs1
0 0

Proof. The first step of the proof consists in showing that every process
involved in this formula has a continuous modification. Then we shall prove
in a second step that the formula is valid for each fixed t by establishing
the continuity with respect to K F,,p,T for the norm kk,,p,T and using

Proposition 8.8. We conclude by using the density of F,,p,T into B,,p,T .

First Step : existence of a continuous modification.

1. Since f (4) is bounded, Proposition 8.4 entails that W1
K (s),WK (u)
f 0 be-
longs to H 2h (D1,2 ). Hence, we infer from Proposition 7.2 that since
3h > 12 , the process
Z t
t ItK (W
1
K (s),WK ()
f 0 )(s) dBs
0
has a continuous modification.
36 P. CARMONA, L. COUTIN, AND G. MONTSENY

2. Similarly, since f Cb4 , Lemma 2.15 implies that we may apply Propo-
sition 7.2 to the process ( f 000 (WK (s))E [(WK (u) WK (s))WK (s)] ; 0
u, s T ) to obtain the existence of a continuous modification for
Z t
t ItK ( f 000 (WK (s))E [(WK (u) WK (s))WK (s)])(s) dBs .
0
3. Since f Cb5 we can apply Proposition 8.10 to W 0
K (s),WK (s1 )
f 00 to ob-
tain the existence of a continuous modification for
Z t Z s
t dBs 0s,s1 f 00 (WK ) ItK (K(, s))(s1 ) dBs1 .
0 0
4. The existence of a continuous modification for
Z t
t f 00 (WK (s))E [WK (t) WK (s) | Fs ]K(t, s) dBs
0
is a consequence of Burkholder-Davis-Gundy inequalities, Proposi-
tion 5.1, and the fact that f 00 is bounded.
5. Eventually, since f Cb3 the existence of a continuous modification
for
Z t
t f 000 (WK (s))E [WK (s)(WK (t) WK (s))]K(t, s) dBs
0
is a consequence of Burkholder-Davis-Gundy inequalities and Lemma 2.15.
Second Step : Itos formula for a fixed time. According to the defini-

tion 2.6, page 8, of B,,p,T , the mapping K 0t f 00 (WK (s)) ds
d
E WK (s)2 ds
R  

is continuous on B,,p,T .
In order to establish the continuity of
Z t
f 0 (WK (s)) + f 00 (WK (s)E [WK (t) WK (s) | Fs ] K(t, s) dB(s)

K
0
Z t
f 000 (WK (s))E [(WK (t) WK (s))WK (s)] K(t, s) dB(s)


0
we only need to prove the continuity from E,,p,T into the space L , for
some /2, of the integrand mapping
K f 0 (WK (s))+ f 00 (WK (s)E [WK (t) WK (s) | Fs ] f 000 (WK (s))E [(WK (t) WK (s))WK (s)] .
This can be obtained by Taylor expansions for f 0 , f 00 and f 000 , Proposi-
tion 5.1, Lemma 2.15 and the following upper bound, proved for any adapted
process b bounded in L /2 via Burkholder-Davis-Gundy and Holder (/2, /2)

inequalities
Z t Z T h i 1/
/2

K(t, s)b(s) dBs CkKk
0 ,,p,T E |b(s)| ds .
0
2
L
Thanks to Lemma 6.2, the continuity of
Z t
K ItK (W
1
K (s),WK ()
f 0 )(s) dBs
0
 37

is a consequence of the continuity of

K E,,p,T W
1
K (s),WK (u)
f 0 H (D2,2 )
for some > h 1q . This latter continuity is an easy consequence of Propo-
sition 8.4 applied to = 2h .
Since h > 1/6, combining Lemmas 6.2 and 2.15 with a Taylor expansion
of f 000 , shows that the mapping
Z t
K ItK ( f 000 (WK (s))E [(WK () WK (s))WK (s)])(s) dBs
0
h
is continuous from A,,p,T into L2 (, P).
Eventually, thanks to Proposition 8.4, we can apply Proposition 8.10 to
a(s, s1 ) = f 0 (WK (s)) f 0 (WK (s1 )) to get the continuity of
Z t Z s
K dBs 0s,s1 f 00 (WK ) ItK (K(, s))(s1 ) dBs1
0 0
2h
from A,,p,T into L2 (, P).

A PPENDIX A. PATHWISE VERSUS STOCHASTIC INTEGRATION

We shall establish that for W1H , W2H two independent fractional Brownian
motions of index 0 < H < 12 ,
Z t
W1H (s) dW2H (s)
0
cannot be defined as a classical pathwise integral. To this end we recall
that if C [0, T ] denotes the space of Holder continuous functions on [0, T ]
of index then:
(A.1) for every (0, H), a.s W H C [0, T ] .

P W H CH [0, T ] = 0 .


(A.2)
((A.1) can be proved with Kolmogorovs continuity criterion, and (A.2) is
an immediate consequence of the law of the iterated logarithm, Theorem 1.7
of [2]). Let v( f ) be the variation index of the function f :

v( f ) = inf p > 0 : v p ( f ) < + .
Then, see Theorem 5.3 of [11],
1
(A.3) almost surely v(W H ) = .
H
(1) 0t W1H (s) dW2H (s) cannot be defined as a p-variation integral. Indeed,
R

see [11], this requires that for p, q > 0 such that 1p + 1q > 1,
W1H W p ([0, T ]), W2H Wq ([0, T ]) .
1
From (A.3) we deduce that p, q H and thus H > 12 .
38 P. CARMONA, L. COUTIN, AND G. MONTSENY
Rt
(2) 0 W1H (s) dW2H (s) cannot be defined as a generalized stochastic inte-
gral of Russo-Vallois [28] and Zahle [34]. Indeed, this integral requires
that the integrand W1H admits a generalized quadratic variation, and this im-
1

plies by Proposition 5.1 [34], that W1H is in the function space W2,
2
. Since,
see Theorem 1.1 [34], for > 0 small enough
1
1 1
W2,
2
I 2 2 (L2 ) C 2 3 ,
1
Hence, we need that W1H C 2 3 for arbitrary small , and this implies
H 12 .

(3) 0t W1H (s) dW2H (s) cannot be defined as a generalized integral of Ciesiel-
R

ski and al [4]. If this were the case, then section V.B implies that
1H 1 1 1
W1H B p,1 for < H < 0 = 1 .
p p p
Since, for > 0 small enough, B p,1
1H 1H = C1H , this implies
B,
again H 12 .

A PPENDIX B. A N ELEMENTARY CALCULUS L EMMA

This Lemma is used in section 2 to evaluate the Holder exponent of some
integral processes.
Lemma B.1. Let, for 0 s t T and (0, 1) (1, 2),
Z
f (s,t) = dudv (v u) .
0<u<s<v<t
Then,
| f (s,t)| C(T, ) (t s)h (0 s t T ) ,
with h = 1 if (0, 1) and h = 2 if (1, 2). Similarly the function
defined for 0 s t T by ,
Z
g(s,t) = (u s) dudv
s<u<t<v<T
h
satisfies |g(s,t)| C|T s| .
Proof. Since 6 {1, 2}, then
1
f (s,t) = (t 2 (t s)2 s2 ) .
(1 )(2 )
If 1 < < 2, this implies | f (s,t)| C(T, )(t s)2 since the function
x x is locally Holder of index if (0, 1).
If 0 < < 1, then Taylors formula yields t 2 s2 C(t, )(t s) and
this implies | f (s,t)| C(T, )(t s).
 39

R EFERENCES
[1] E. A LOS , O. M AZET , AND D. N UALART, Stochastic calculus with respect to frac-
tional Brownian motion with Hurst parameter less than 1/2, Stochastic Process.
Appl., 86 (2000), pp. 121139.
[2] A. B ENASSI , S. JAFFARD , AND D. ROUX, Elliptic Gaussian random processes,
Revista Matematica Iberoamericana, 13 (1997), pp. 1990.
[3] J. B ERAN AND N. T ERRIN, Testing for a change of the long-memory parameter,
Biometrika, 83 (1996), pp. 627638.
[4] Z. C IESIELSKI , G. K ERKYACHARIAN , AND B. ROYNETTE, Quelques espaces fonc-
tionnels associes a` des processus gaussiens. (Some function spaces associated with
gaussian processes)., Stud. Math., 107 (1993), pp. 171204.
[5] F. C OMTE AND E. R ENAULT, Long memory continuous time models, Journal of
Econometrics, 73 (1996), pp. 101150.
[6] L. C OUTIN AND Z. Q IAN, Stochastic analysis, rough path analysis and fractional
Brownian motions. To be published in PTRF, 2000.
[7] , Stochastic differential equations for fractional Brownian motions, C. R. Acad.
Sci. Paris Ser. I Math., 331 (2000), pp. 7580.
[8] W. DAI AND C. H EYDE, Itos formula with respect to fractional Brownian motion
and its application., J. Appl. Math. Stochastic Anal., 9 (1996), pp. 439448.
[9] L. D ECREUSEFOND AND A. U STUNEL, Stochastic analysis of the fractional Brow-
nian motion, Potential Analysis, 10 (1997), pp. 177214.
[10] C. D ELLACHERIE , B. M AISONNEUVE , AND P. M EYER, Probabilites et potentiel.
Chapitres XVII a` XXIV: Processus de Markov (fin) Complements de calcul stochas-
tique, Hermann, Paris, 1992.
[11] R. D UDLEY AND R. N ORVAISA, An introduction to p-variation and Young integrals,
Tech. Rep. 1, Maphysto, Centre for Mathematical Physics and Stochastics, University
of Aarhus, Ny Munkegade, DK-8000 Aarhus C, Denmark, dec 1998. Concentrated
advanced course.
[12] T. D UNCAN , Y. H U , AND B. PASIK -D UNCAN, Stochastic calculus for fractional
Brownian motion I. Theory., SIAM J. Control optim., 38 (2000), pp. 582612.
[13] D. F EYEL AND A. DE L A P RADELLE, On the approximate solution of the
Stratonovitch equation, Electronic Journal of Probability, 3 (1998).
[14] A. KOLMOGOROV, Wienersche spiralen und einige andere interessante kurven im
Hilbertschen raum, S.R. (Doklady) Acad. Sci. USSR (N.S.), 26 (1940), pp. 115118.
[15] N. L EBEDEV, Special functions and their applications, Dover Publications, New
York, 1972. Translated and edited by Richard A. Silverman.
[16] W. L ELAND , M. TAQQU , W. W ILLINGER , AND D. W ILSON, On the self-similar
nature of Ethernet traffic, IEEE/ACM Trans. Networking, 2 (1994), pp. 115.
[17] J. L EON, Fubini theorem for anticipating stochastic integrals in Hilbert space, Ap-
plied Mathematics and Optimization, 27 (1993), pp. 313327.
[18] S. L IN, Stochastic analysis of fractional Brownian motions, Stochastics Stochastics
Rep., 55 (1995), pp. 121140.
[19] R. L IPTSER AND A. S HYRIAEV, Theory of Martingales, Mathematics and its Appli-
cations, Kluwer Academic Publishers, 1989.
[20] T. LYONS, Differential equations driven by rough signals, Revista Matematica Ibero-
Americana, 14 (1998), pp. 215310.
[21] I. N ORROS , E. VALKEILA , AND J. V IRTAMO, An elementary approach to a
Girsanov formula and other analytical results on fractional Brownian motions,
Bernoulli, 5 (1999), pp. 571588.
[22] D. N UALART, The Malliavin calculus and related topics., Probability and Its Appli-
cations, Springer-Verlag, New York, NY, 1995. [ISBN 0-387-94432-X].
40 P. CARMONA, L. COUTIN, AND G. MONTSENY

[23] V. P IPIRAS AND M. S. TAQQU, Integration questions related to fractional Brownian

motion, Probab. Theory Related Fields, (2000), pp. 251291.
[24] N. P RIVAULT, Skorokhod stochastic integration with respect to non-adapted pro-
cesses on Wiener space, Stochastics Stochastics Rep., 65 (1998), pp. 1339.
[25] P. P ROTTER, Stochastic Integration and Differential Equations, vol. 21 of Applica-
tions of Mathematics, SpringerVerlag, 1992.
[26] L. ROGERS, Arbitrage with fractional Brownian motion, Math. Finance, 7 (1997),
pp. 95105.
[27] F. RUSSO AND P. VALLOIS, Forward, backward and symmetric stochastic integra-
tion, Probab. Theory Related Fields, 97 (1993), pp. 403421.
[28] , The generalized covariation process and Ito formula, Stochastic Process.
Appl., 59 (1995), pp. 81104.
[29] A. A. RUZMAIKINA, Stieltjes integrals of Holder continuous functions with applica-
tions to fractional Brownian motion, J. Statist. Phys., 100 (2000), pp. 10491069.
[30] S. S AMKO , A. K ILBAS , AND O. M ARICHEV, Fractional integrals and derivatives,
Gordon& Breach Science, 1993.
[31] D. S TROOCK, A Concise Introduction to the Theory of Integration, Birkhauser, 1994.
second edition.
[32] L. YOUNG, An inequality of Holder type, connected with Stieltjes integration, Acta
Math., 67 (1936), pp. 251282.

[33] M. Z AHLE , Integration with respect to fractal functions and Stochastic Calculus,
Probability Theory and Related Fields, 111 (1998), pp. 333374.
[34] , On the link between fractional and stochastic calculus., in Stochastic dynam-
ics. Conference on Random dynamical systems, Bremen, Germany, April 28 - May
2, 1997., H. e. e. a. Crauel, ed., Springer, 1999, pp. 305325. Dedicated to Ludwig
Arnold on the occasion of his 60th birthday.

P HILIPPE C ARMONA , L ABORATOIRE DE S TATISTIQUE ET P ROBABILIT E S , U NIVER -

PAUL S ABATIER , 118 ROUTE DE NARBONNE , 31062 T OULOUSE C EDEX 4
SIT E
E-mail address: carmona@cict.fr

L AURE C OUTIN , L ABORATOIRE DE S TATISTIQUE ET P ROBABILIT E S , U NIVERSIT E PAUL

S ABATIER , 118 ROUTE DE NARBONNE , 31062 T OULOUSE C EDEX 4
E-mail address: coutin@cict.fr

G E RARD M ONTSENY , L.A.A.S, 7, AVENUE DU C OLONEL ROCHE , F-31077 T OULOUSE

C EDEX 4
E-mail address: montseny@laas.fr