Probab. Theory Relat.

Fields 97, 403421 (1993)

Probability
Theory a.d Related Fields

9 Springer-Verlag 1993

Forward, backward
and symmetric stochastic integration
Francesco Russo 1,* and Pierre Vallois 2
1 Universit~ de Provence, UFR-MIM, 3 Place Victor Hugo, F-13331 Marseille C~dex 3, France
2 Universit6 Pierre et Marie Curie, URA 224, CNRS, Laboratoire de Probabilit6s, Tour 56,
3e 6tage, 4 Place Jussieu, F-75252 Paris C6dex 05, France

Received December 20, 1991; in revised form April 26, 1993

Summary. We define three types of non causal stochastic integrals: forward,

backward and symmetric. Our approach consists in approximating the integra-
tor. Two optics are considered: the first one is based on traditional usual stochas-
tic calculus and the second one on Wiener distributions.

Mathematics Subject Classification: 60H05, 60H07, 60H30, 60J65

0 Introduction and notations

Many authors have examined extensions of classical stochastic integrals to a

certain class of anticipating integrands. A good list of references for this purpose
is contained in IN]. Among the generalisations we find the classical backward
stochastic integration [K 1], the enlargement of filtrations [Je], the extension
of Stratonovich integral [O], [Z], [NP], the Skorohod integration IN], and
finally the forward integration [-KR], IBM], [-AP]. Let us remark that in the
three last approaches the only integrator involved is essentially the Brownian
motion.
In this paper we define forward, backward and symmetric integrals by a
limit procedure. These integrals are respectively extensions of Ito, backward
and Stratonovich integrals. We have focused our attention especially on the
forward integration; the backward anticipating integral is defined in an analo-
gous way and the symmetric integral is obtained by averaging. Our work has
the following features:
i) through the definition, we make explicite the " f o r w a r d " nature of the integral:
in fact the integrator must operate with an infinitesimal anticipation with respect
to the integrand,
ii) we introduce an effective non-causal stochastic integration with respect to
more general integrators than the Brownian motion,

* Permanent address: Universit/it Bielefeld,BIBOS, D-4800 Bielefeld,Germany

404 F. Russo and P. Vallois

iii) one of the straightforward consequence of the definition is an integration

by parts formula,
iv) we relate the main anticipating stochastic integrals; in particular we connect
Skorohod integral with the one defined through enlargement of filtration.
We introduce now some useful notations. Let e >0. If X--(X(t); 0___<t < 1)
and Y=(Y(t); 0<t__<l) are two stochastic processes, continuous at 0 and 1,
we set,

t y((t§
(1)_ l - ( e , X , d Y ) = ~ X(t)
0

1 Y ( t ) - Y ( ( t - ~) v O) dt.
(1)+ I + (e, X, d Y ) = ~ X(t)
8
0

1
W e also define by symmetry, I ~ + I - ) / 2 . We denote by ~ X d - Y (resp.
1 1 0
Xcl + Y, ~ X d ~ Y) as the limit in probability of I - (~, X, d Y) (resp. I + (e, X, d Y),
0 0
I~ X, d Y)) when e ~ 0 + ; the first limit is called forward (resp. backward, sym-
metric) integral of X with respect to Y.
An obvious relation between (1)_ and (1)§ is given by,

(2) I - ( e , X , dY)+I+(e, Y, d X ) = Y(1) ~ X(s)ds -X(O) ~ Y(s)ds .

1-~ 0

If we take the limit when e goes to 0 + , we get the following integration by

parts formula,

1 1
(3) X d - Y + ~ Yd + X =(XY) (1)-(XY)(0).
0 0

By symmetry we obtain,
1 1
(4) ~ X d o y + ~ Yd o X = (X Y) (1) - (X Y) (0).
0 0

In our framework it is easy to see that the regularisation of the integrator

with a mollifier naturally leads to the study of the objects I - and I +. IT]
has approached the symmetric integral with respect to the Brownian sheet by
making use of convolution procedure.
The article is organised in two sections. The first one is inspired by the
tools of the classical stochastic calculus and the second one by the modern
technics of the Wiener functionals analysis.
In Sect. 1, Theorem 1.1 gives a substitution formula for the forward integral
with respect to a martingale; this result allows us easily to construct a solution
to a system of stochastic differential equations with anticipating initial condition
(Theorem 1.2). In the second part, we concentrate on the case where the integra-
Forward, backward and symmetric stochastic integration 405

tor is a one-dimensional Brownian motion. Theorem 2.1 is the main result and
relates our integrals to Skorohod integral and the trace of the Malliavin deriva-
tive of the integrand, the traces being defined in a weak sense.

1 A substitution theorem. Application to stochastic differential equations

We begin by showing that the forward integral of a previsible and bounded

process with respect to a semimartingale equals the Ito integral.
Notations. 1~ (f2, ~,~= ( 4 ; 0 < t < 1), P) will denote a classical filtered probability
space, ~ satisfying the usual conditions. N is the ~r-algebra generated by the
previsible processes. Let (B(t); 0 < t < 1) be a usual ~ - B r o w n i a n motion.
2 ~ If Y is a continuous Y-semimartingale and H a previsible process, such
1 1
that ~ H(s)2d(Y,, Y ) (s)< 0% a.s., then ~ H(s) dY(s) is usual stochastic integral
0 0
with respect to Y
3~ If f is a locally integrable function on N + , we denote by ~ ( f ) the set

of all t > 0 such that lim -1 i f(s) ds=l=f(t).

*~0 5
t-8

We recall the theorem of Stein (IS], Theorem 1, p5); for every p > 1, there
exists an universal constant cp such that,
1 1
M(s)Pds <=Cp ~ If(s)] p ds,
0 0

where,
1
M(s)= 0<e_-<l~s
sup J r ~ If(t)[ 1[~ 11dt, 0 <- s _ < l ,

and V~= [ - 5, e] or V~= [ - ~, 0] or V~= [0, e].

4 ~ Let (X(t); 0<t__<l) be a stochastic process and ue[0, 1]. X u is the process
defined by: X " = X 1Eo,ur
Lemma 1.1 ([RY], exercise 5.17, p 165) Assume M is a continuous square integra-
ble ~-martingale, H a mapping defined on ~2xN+ x lR+, bounded and
| N(lR+)-measurable. Then for every s>O and t>O we get,
t
(1.1)

Proposition 1.1 Assume Y is a continuous ~-semimartingale, Y= M + V is the

canonical decomposition of Y, where M (resp. V) is a square integrable martingale
(resp. V is a bounded variation process) defined on [0, 13. Suppose X is a W-
previsible and bounded process such that:
1

(1.2) S l~x)~ (dig[ ( s ) + d ( M , M>s) = 0 a.s.

0
406 F. Russo and P. Vallois

Then for every te[O, 1], we have a.s.:

1 t

x d- Y= X(s) d r(s).
0 0

Remarks. 1~ The assumption (1.2) is realized if X is a left-continuous process,

or V and (M, M ) are two absolutely continuous processes with respect to the
Lebesgue measure.

2 ~ When X(t,o))= ~ cq(t) Gi(co), where for every l < i < n , o:~ (resp. Gi) is a
i=1
bounded function defined on ~ + with compact support (resp. r.v.) and
1
~ l{~s~(~)}(d[V(s)[ + d ( M , M ) s ) = 0 a.s., then
i=10

X d - Y= o~i(s) dY(s) Gi.

0 i=1

3~ Assume X and Y are two continuous square integrable semimartingales,

1
such that X ( 0 ) = 0 or Y(0)=0. Let ~ X(s)odY(s) be the Stratonovich integral
1 0 1
of X with respect to Y, then S X(s)odY(s)= ~ X d ~ Y
0 0
4 ~ If Z is a process, we denote by 2 the process defined by: 2 ( t ) = Z ( 1 - t ) ,
0 _ < t _ l . From the identity, I+(e,X, d Y ) = - I - ( ~ , X , dY), we deduce that
1
X d +B is the backward integral [K 1].
o
5~ Assume Y is a continuous square integrable ~--martingale, and (g=((gt; 0
_<t_<l) is a filtration containing ~ (i.e. ~ c ( g z V t e [ 0 , 1]). When X is only
1
(g-adapted, ~ X ( t ) d Y(t) has no meaning. However if Y is a N-continuous semi-
o
martingale, then the theory of enlargement of filtrations [Je], allows us to define
1
X d g Y as the usual (g-stochastic integral of X with respect to Y Clearly we
0
have:
1 1
x-dr= xd r.
0 0

Proof of Proposition 1.1 1~ Let e > 0 and te [0, 1]. We have,

I - (e, X t, d Y) = I - (e, X t, dM) + I - (~, X t, d V).

Forward, backward and symmetric stochastic integration 407

i) By Fubini theorem we get,

I - (8, X t, dV) = Xt(u) leo , ll(u) du 1[o, 1 ] ( S ) dV(s).

0 s

1
By (1.2), dlV] almost surely, lim - j Xt(u) du=Xt(s).
z~O+ 8 s-~

The process X t is bounded, by the dominated convergence theorem we

obtain,
1 t
x ' d - v= X(u) dv(u).
0 0

ii) Recall that,

t
I - (8, X t, dM) = 1 ~ X(u) (M((u + 8)/x 1 ) - M(u)) du.
8 0

We notice that X (u) (M ((u + 8) A 1)-M(u)) is the stochastic integral of the ele-
mentary process (X(u) l~,<s=<,+,~; O < s < 1), with respect to the martingale (M(t);
0<t<l).
By Lemma 1.1, we have,

I - (8, X t, dM) = Xt(u) lto ' ll(U) du 1[o' tl(s) dM(s).

0 s

But,

1
and lira - j Xt(u) 1Eo,l~(u)du=Xt(s), d(M,M)~ a.s.
a~O+ 8 s--~

To process X ~ is bounded; using again the dominated convergence theorem

1 1
we show that I-(8, X t, dM) converges in L2 to ~ Xt(u)dM(u); then ~ X ~ d - M
t 0 0
= ~X(sldM(s). []
0

We give now a partial reciprocity for Proposition 1.1. A "cfidlfig" (resp.

"cfiglfid') process is a process whose paths are a.s., right-continuous (resp. left-
continuous) on [0, 1[ (resp. ]0, 1]), with limit from the left (resp. right) on ]0, 1]
(resp. [0, 1 [).
408 F. Russo and P. Vallois

1
Proposition 1.2 Let Y be a ~-adapted c~d l~g process and suppose that ~ X d - Y
o
exists for any bounded, c~glfid and ~-previsible process X. Then Y is a ~ -
semimartingale.
Remark (1.3) If O reduces to a single point, processes are deterministic functions
1

defined on [0, 1]. In particular Proposition 1.2 tells us that if ~ f d - g exists

o
for every bounded and cfigl/td function f then g is a bounded variation function.
The same conclusion holds if the - i n t e g r a l is replaced by +integral We can
1

prove that if ~ f d - g exists for every bounded and Borel function f then g
0
is absolutely continuous on [0, I].
Proof of Proposition 1.2 1~ We introduce two metric spaces ~r I and f 2 defined
as follows:
(i) f l is the set of bounded, ~--previsible and cfidl~ig processes defined on [0, 1],
and dl(X, Y)= sup IX(t, o))- Y(t, o))l, where X and Ybelong to X 1.
0_<t<l
r

(ii) ~2 is equal to the set of ~-l-measurable r.v., and d z ( X , Y ) = E ( I X - Y ) [ / x 1)

where X and Y are two elements of 3f 2.
(~1, dl) and (:Y2, dE) are two complete metric spaces which moreover satisfy,
(a) di(X, Y ) : d i ( X - Y, 0), i = 1 or 2.
(b) for every X in ~ i (resp. a in IR), t ~ t X (resp. Y ~ a Y ) is a continuous
function from N (resp. (Y'~, d~)) to (Y'z, d~), i = 1 or 2.
According to ([DS], Chap. II), (~Y1, dl) and ( f 2 , dE) are two F-spaces. ( f 2 , d2)
is equal to X2 equipped with the topology defined by the convergence in proba-
bility.
We introduce a family {#~; e > 0} of linear operators from ( f :, dO to (Y'2, dz)-
Let e > 0. We set #~(Z) = I - (e, Z, d Y). We have,

11
I~(z)l~ sup ]Z(t, co)l- jlY((t+e)/x 1)-Y(t)ldt.
0 -<t-<l 8 0

Then/~ is continuous. According to our assumptions, tt~(Z) converges in (Y'2, d2)

1
to ~ Z d - Y
0
In particular for every fixed Z in (~ 1, d 1), the family {#,(Z), ~ > 0} is bounded
1
in (Y'2, d2). Theorem 18 p. 55 of [DS] proves that Z ~ ~ Z d - Y is a linear and
o
continuous map from (4"1, d0 to (5~2, d2).
2 ~ Let :g"t be the set of elementary previsible processes Z of the type: Z =
~Zilm,,.11 where 0__<tl<...<tn<t,+l=l, and Z~ is bounded and
i=1
Forward, backward and symmetric stochastic integration 409
1
~,-measurable, for every ie{1, ..., n}. Y"I is a subspace of f l and Z ~ ~ Zd" Y
is continuous map from (Y":, dl) to ( f 2 , d2). o
We fix an element Z of the previous type. We have,

I-(GZ, dY)= ~ I-(e, Zillt,,t,+ll, dY)= ~ ZiI-(e, llt~.t,+,l,dY ).

i=1 i=1
Since
i ti+l+g ~ ti-k-g
I-(~,ll...+,l,dY)= f Y(tA1)dt-- ~ Y(t/xl)dt,
ti + 1 ti

and Yis a right-continuous process, we get,

1
f Z d - Y= ~ Zi(Y(ti+l)- Y(ti)).
0 i=1

From the Bichteler-Jacod theorem ([DM], Chap. VIII, p. 400), Y is a ~-semi-

martingale. []

Theorem 1.1 Assume G is lRd-valued r.v., c~>l, q>2c~/(~-l), ~5>d(2~+q)/2a,

and X aNORd)| ~-measurable map from IRd x s x [0, 1] to N such that for every
positive N, we have:
1
(i, ,,qdsl
(1.4)
(ii) E
[i [X(a,s)-X(b,s)lqd
,1<=cN[a-bla;Vlal<=N,Vlbl<=N.

Let (M(t); 0 < t =<1) be a continuous local martingale such that d ( M ) (t)= h(t)dt
and
(1.5) E [ i h(t)~ dt]< oo.

1
7hen a.s., for every re[0, 1], ~ Xt(G, .) d- M exists and
0

(1.6)
1
~ Xt(G,')d- M=
(/ Xt(a,u)dM(u) )
0 a=G"

Remarks(1.7). 1~ By (1.5) we have E[(M,M)(1)~]<oo, then M is a

L2%martingale.
2 ~ In the Brownian case, if X satisfies (1.4) with q > 2 and 6 > d then (1.6)
holds (choose c~> 1 large enough such that 6 > d(2 a + q)/2 ~ and q > 2 e/(cr
Proof of Theorem 1.1 1) Let E be the space cg([0, 1]) equipped with the uniform
norm: IlfIl= sup [f(u)l. (E; I[-II) is a Banach space. Let ((U(a,t); O<t_<l);
aeNd) o_<u_<~
410 F. Russo and P. Vallois

and ((V(e, a, t); O< t < 1); aMRde>O) be two families of E-valued random pro-
cesses defined by,

1 t

V(a, t)= ~ Xt(a, s) dM(s)= ~ X(a, s) dM(s),

0 0

V(e,a,t)=I-(e,X'(a,.),dM)=~ ~s X(a'u) l~ur176 dM(s),

0 e

for every t e [0, 1].

We set: p=2Y)c~/(2c~+q),l~=qc~/(2c~+q); then p>d and 1<i~<c~. We use
successively Doob and Burkholder-Davis-Gundy inequalities,

E [ll U (a,,)llZT]~ CE [(i X (a, s)2 h(s) ds) ~]< CE [i X (a, s)Z~h(s)~ds].

Let p' = c~/7; q' = (2 e + q)/2c~is the conjugate exponent of p', then,

E[l[U(a,.)][ze]<C'{E[~o X(a,s)qds]} 1/q',

( I-I I ) I/p"
whereC'=C~Et!h(s)~ds]) 9
Using (1.4) and also the maximal inequality in Lq[O,1] (IS], Theorem 1,
p. 5), we get,

(1.8) E[llg(a,.)[12']< oo, g[l]g(e,a,.)ll2']< oo.

(1.9) g[[lg(e,a,.)-g(e,b,.)[lZ~]+E[llU(a,.)-U(b,.)[12~]<ku[a-b[ p.
where kN is a constant.
2) By (1.9) and Kolmogorov lemma, U admits a continuous version Uo, and
Uo also verifies:

(1.10) E[l[Uo(a,.)-Uo(b,.)lla~]<__kNla-b]P; Vlal<N and ]b]<N.

3) We recall Garsia, Rodemich and Rumsey lemma (G.R.R.) stated by Barlow
and Yor ([BY]).
Fix p>d, 7>0, O<m<p-d, g > 0 , KN={aelRd;]al<--_g} and {V(a),a~KN}
a family of E-valued random variables such that,

E[llV(a)-V(b)llZ~]<kla-b[ p, VaeKN, beKN.

Then, there exist two constants C1 and C2, independent of the process V and
a r.v. F such that,
(i) r[V(a)-- g(b)rl2, < C1 ]a-b[mr Va~KN, b~Ks
(ii) E(F)<C2 k.
Forward, backward and symmetric stochastic integration 411

4) There exist two constants C~ and C2 and a r.v. F(~) such that,

(1.11) [IV(~, a,.)-- V(e, b,.)]1z r =<_C1 l a _ blm F(e),

(1.12) E (F(@ < C 3 ,

with C3 = C2 kN.
5) Let ~ > 0 and GN=G 1(I~i~ ~.

P(II V(~, G , . ) - Uo (G,.)[[ > ~ ) ~ P(IG[ > N ) + P(II V(~, GN,.) - Uo (GN,.)II >~).

Even if it means replacing G by GN, we can assume that [G[ being bounded
by N. Fix n > N and G, a discrete r.v. (with finite values) such that G, e K N
and 1[G - G , ]l~ < 1/n. We deduce from (1.1 1) and (l.12) the inequalities,

(1.13) E[IIV(~,G,.)--V(~,G~,.)IIzq~CaE[IG--G~I~I~(~)]~C~Cg(1/nF.

A similar proof using (1.10) shows that,

(1.14) EEllgo(G,.)- No(G~,.)II~q ~ C4(1/n) m.

For every fixed a > 0 , we deduce from (1.8) and the maximal inequality that
the limit in the L2~(E)-sense of V(e, a,.) is equal to Uo(a,.)=U(a,.). Moreover
it is clear that if G is a discrete r.v. which takes its values in KN, V(e, G,.)
converges in L2>'(E) to Uo(G,.).
It is sufficient to use now the two uniform inequalities (1.13) and (1.14). []
Let us discuss now the application of Theorem 1.1 to the systems of stochastic
differential equations (S.D.E.'s) with an initial non-adapted value. We work in
the general outline defined by Jacod ([Ja], Chap. XIV). We introduce
O=cg([0, 1],IR), ~ = ( ~ ; 0 < t _ _ < l ) the natural filtration on O, O ' = O x ~ , ~ '
=(~ (~| 0<t__< 1) and ~ ' the a algebra generated by the previsible pro-
g<=t
cesses defined on t2'. Let (B~ ..... /3,) be a n-dimensional Brownian motion
defined on (s ~ = ( ~ ; 0 < t < l ) , P). We assume that o-=(o-1, ...,o-,) and b to
satisfy,

(1.15) b and ai for every 1 < i < n, maps [0, 1] x f2' to IRa and N'-measurable.

There exists two constants K~ > 0 and K2 > 0 such that,

(1.16) ~ Io-i(s, ~o, 05)- oi(s, ~o, 051)1+ Ib(s, o~, 05)-b(s, ~, ~1)] <K1 1105--051lit,
i=l

(1.17) ~ lai(s, ~o, 05)12+ [b(s, (9, o5)12< K2 (1 + 1105112),

i=l
412 F. Russo and P. Vallois

for every 0<s__< 1, cocO, (5 and (51 in c~([0, s]), where 11(511~= sup 1(5(u)[.
O<u<_s

When assumptions (1.15), (1.16) and (1.17) are realized, we know ([Ja], (14.50))
that the systems of S.D.E.'s:

(1.18) X ( t ) = x + ~ iai(s, co, X.)dBi(s)+ ib(s, co, X.)ds,

i=10 0

has a unique solution, where x e l R a, and X. is the element of O defined by,

X. (s)= X (s, co). We denote by X (x,') the unique solution of (1.18).
Theorem 1.2 Assume ~ is a r.v. f f l-measurable, rs[0, 1], and Y is the process:
Y(t) = X (4, t); 0 _<t < 1. Then a.s.,

(1.19) Y(z)=(+ ~ ai(s, co, Y(.)) lto.,~d- Bi(s)+ ~ b(s, co, Y(.))ds.
i=1 0 0

Proof. We set Ui(t, x, co)=ai( t, o~, X (x,', co)) lto ' ~1(t), O<_i <_n. It is clear from (1.16)
that:

(1.20) ~, IVi(t, x, co)- a~(t, y, co)[~K1 [IX(x,., co)-X(y,., co)H.

i=1

for every t~[0, 1].

According to an easy adaptation of [K2], Theorem 2.1, p. 211 there exist two
constants Ca and C2 such that,

(1.21) g [l[X (x,.)-- X (y,.)[?] <-_C11x-yl ~,

(1.22) g [llX (O,.)llq] < C2,
where q > 0. We chose q > max (2, d). By (1.16), (1.17) a straightforward calculation
shows that Ui satisfies (1.4) for every i~{1, ...,n}; we can apply Theorem 1.1,
and (1.19) follows immediately. []

2 Wiener analysis approach (Malliavin calculus)

In this section, we would like to examine the convergence of the 8-integrals

defined in the Introduction through a Wiener interpretation; [W] and [BH]
will be the basic references on this subject.
All along this section, r will be a fixed real number and 1 < q __<2.
Let (O=c~[0, 1], H = L 2 [ 0 , 1)], P) be the canonical Wiener space with the
usual Brownian motion (B(t); 0 < t < l ) . E will be a separable Hilbert space,
with inner product (.,.). ~ o (E) will be the algebra of E-valued Wiener functionals
and the space of E-valued Wiener distributions will be the dual space ~ _ ~ (E).
We recall that

9 _ . (E) = U (E) = G,s(E),

p>l p>l
s~ s~
Forward, backward and symmetric stochastic integration 413

where (Np, s(E); p > 1, seN.) is the family of Sobolev-Watanabe-Kr6e spaces. (.,. }
will stand for the duality between N_~ (E) and N~ (E). If E = P,., we will simply
drop E. Lq(E) stands for U(O;E); we recall that Nq, o(E)=Lq(E). The gradient
operator D maps continuously Nq,~ into Nq,~_ a (H). The divergence operator
(Skorohod integral) maps continuously Nq,~(H) into N0, s_ ~. D and 5 are dual
operators.
Let denote H , = H | ... | (n times). If TeNq,~(H,), aeH,, we denote by
T~ the Wiener distribution in Nq,~ defined by, (T~, Y} = (T, c~| Y}. If TeNq,~(H,)
there exists a unique ~'I'eLq([O, 1]; Nq,~) such that,

(2.1) T~= 5 ~(t) T(t)dt,

[0, lln

where the integral is understood in Pettis sense; this follows from a slight exten-
sion of [BH], Proposition III 1.1.8. Twill be the Nq,r-decomposition of T. Clearly,
we can replace Nq, r by N-oo. If TeN_oo, then DTe@_oo(H), (D~ T;0_<s<l)
will denote its N_ ~o-decomposition, and D, T stands for (DT)~. Since N~ | H,
is dense in Noo(H,), then (2.1) can be extended to,

(2.2) (T,Z>= ~ (Z(t),T(t))dt,

[0, 1]~

for every Z in @o~(Hn).

The Wick product of two Wiener distributions is defined in [BH], exercise
1.6 p. 133, with the help of the notion of characteristic functionals and will
be symbolized by: However if U = I~ (f,), V= Im(g,,) are Wiener iterated Ito inte-
grals where f~eL2([0, 1] ~) and g,,eLa([-0, 1Ira), then U: V=I,+,,(f,| If hell,
YeU then the definition of the Wick product implies that 6h: YeNq,_ 1 and

(2.3) ah: Y = a ( h | Y).

From now on A will be an element of { - , +, 0}. For simplicity we will

write Ia(e, X) instead of Ia(e, X, d B). Let X = (X(t); 0 < t__<1) be a process in
U(H). By standard arguments it is easy to check that Ia(e, X)eLP'(H), for some
p'>l.
The following three functions which are elements of H2 will play an impor-
tant role in the sequel,

_i1~ 1
0{e, -- - - ~ - {( , t)e[0, l]2lt<=s<=t+e}, ~e, + = 7 1{(8, l')e[0, 1121t--8=$</'),

1
~,0 =~ l{(s, t)e[O, ll21t-e<-s<t+e} 9

X~,a will be the process defined by

1
X~,~(s)= ~ x(t)~,~(s, t) dr, se[0, iI.
0
414 F. Russo and P. Vallois

In the definition of I~(~, X) we make use of the Wick product instead of the
ordinary product of random variables, more precisely,

(2.4)
l
1,7 (e, X ) = ~ X(t):
[B((t+~)~l)--B(t)]dt.
0

Similarly to the definition of 1 + (e, X) and I~ X), we can define I + (e, X) and
t~ X).
By (2.3), we get

B((t + ~) A 1)- B(t)]: X (t) = a (~, _ (., t)): X(t) = a r~, _ (., t) x(t):l,
8

(2.5) law(e, X) = a (X,, a).

Lemma 2.1 Let XsLq(H). Then X,,asLq(H) and

(2.6) I a (e, x)=I,,(~, X ) + D .... X.

Proof.For simplicity c~ will stand for e~,a; since e is bounded by l/e, the first
statement is obvious. Therefore Ia~(e,X)s~q,_ 1. We set,
a(~;) = (Iw(e, X), Y>, Ye~.
Using duality we get,

a(e)=(X~,a,DY)=E ~D~Y c~(r,t)X(t)dt dr

kO

1 1

]
According to [W], Proposition 1.11, p. 51, we have,

D~(.,O Y= Y6(c~(', t))-6(c~(', t) Y).

Therefore, we can write

(2.7) a(e)=E(IJ(e, X) Y)-al(e ),

where,
1

Using once more duality we obtain,

al(e)=E ~oe(r,t) YDrX(t)drdt =E(YD~X).

This and (2.7) show (2.6). []

Forward, backward and symmetric stochastic integration 415

At this stage, we need a suitable notion of trace for Ts ~_ ~ (H2). Let Te @q,r(H2);
we will say that Thas a ~q,r A-trace if lira T~.~ weakly exists in Nq, r. Obviously
e-r 0 +
~q,r can be replaced by ~ _ ~ ; in this case, we will speak about ~ _ ~ trace.
Let F be a topological vector space, and f : [0, 1] 2 --, F, Pettis-integrable.
We will say that f has a FA-trace or simply a A-trace if lim j c~,,A(U) f (U) d u
e ~ 0 + [ 0 , 1] 2

weakly exists in F. This limit is denoted by Tra(f). This definition has been
performed by [RV]. It is clear that Te@0,r(H2) (resp. TsN-oo(H2)) has a @q,~
(resp. ~_ o~) A-trace iff Thas a Nq, ~ (resp. ~ _ ~o) A-trace.
Theorem 2.1 Let us suppose X~Lq(H). Then
a) lim I A ( e , X ) = 6 X in ~q,_t.
e~O+

b) lira IA(e,X) weakly exists in @q,-1 (resp. @-o~) iff OX has a @q,-1 (resp.
~0+

~ - o~) A-trace.
In this case lim IA(e,X) is equal to the sum of fiX and the A-trace of
DX. ~-~o+
Remark 2.1 The theorem is still true if we replace Lq(H) by ~q, a (H) and ~q,_
by L~
Proof Using part a) and (2.6) we establish b).
In order to prove a) we observe that it is enough to check that,

(2.8) limX~,a = X in Lq(H).

e~0

After this we can use the continuity of 6 from Lq(H) into @q,_ 1. As for the
proof of (2.8) we need the following lemma.
Lemma2.2 Let Y belonging to Lq(H), V ~ = [ - e , e ] (resp. V ~ = [ - ~ , 0 ] or V~=
i
[0, el). Then ~ ] ~ Y(t) lt0 ' 11(0 dt converges in U(H) to Y, when e goes to O.
s+V~

Proof We set,

1
M(s)= sup ~,,., j [Y(t)[l[0,11(t)dt, 0<s<l._
_
O<e<l Ivd s+Ve

Using Stein inequality we obtain,

By the dominated convergence theorem we can conclude. []

Remark 2.2 We would like now to answer the following question: are the two
approaches of Sects. 1 and 2 compatible? In other words, let X be a process
416 F. Russo and P. Vallois

in Lq(H); let us suppose that I~(e, X) converges in probability to an integrable

v.a. Z and also weakly converges in @_~ to an integrable r.v. Z'. Do we have
Z = Z' a.s.? The answer is not known; however, if we replace the convergence
in probability with the L1 convergence, it is positive. [Hint: Let J f be the
set of bounded r.v. q~ such that qo=f(B(tl), ..., B(t,)), where 0 < t l < ... <tn < 1
and f is a C ~ bounded function such that any derivative is also bounded.
If ~oe ~ , we have E [(pZ] = E [q~Z']. This equality can be extended to ~0e L~ c~~
by density of W into L~176 c~ ~oo].
Theorem 2.1 generalizes in some sense the results of [2;] and [SU] concerning
the relation between Ogawa symmetric integral and traces of derivatives. [Z]
introduces the traces in a functional analysis framework and [SU] through
a limiting technique. The different traces are not really comparable, see [HM].
Let us relate our concept of trace with the classical traces coming out from
functional analysis. Let g be a E-valued function defined on [0, 1]. We say
that g has a E classical trace if g is a square integrable Bochner function and
there exists an element denoted by J-~l(g)~E, such that for any complete orhonor-
real sequence (q~,, n > 0) in H,

(2.9) ~ y cp,(u) cp,(v) g(u, v) dudv,

n = O [0, 112

converges, in E, to ~l(g)-
Remarks 2.3 1~ Let denote ~(u, v)=(g(u, v)+g(v, u))/2. In the definition, g does
not need to be symmetric; however obviously, g has a E classical trace if and
only if ~ has one.
2 ~ This concept of trace appears in [Rl. Let e ~ F and Keg2[o, 1]-*/,210, 1],
the linear operator defined by

1
(K~ ~o)(u)= y ~(v)(~(u, O, e) dr.
0

Rosinski ([R] Proposition 2.1 and Corollary 2.2) has proved, that the sum (2.9)
converges for every (qon,n > 0 ) iff K e is a nuclear operator. In this case (2.9)
does not depend on (~0n,n > 0).
We begin with the scalar case, i.e. E=IR.
Proposition 2.1 Let g be a symmetric function of H 2. Assume g has a IR classical
trace then for any A ~{--, + , 0} g has a A-trace and T r ~ ( g ) = ~ l ( g ) .
Remark2.4 It is clear that if g is continuous on {(s,t);O<_t<_s<_l} (resp.
{(s, t); 0 <__s _<t_< 1}) then g has a - trace (resp. + trace). Recall that Balakrishnan
([B], p. 126) gives an example of a real valued function g such that g is continu-
ous and has no classical trace, this shows that Proposition 2.1 admits no con-
verse.
P r o o f Let consider the operator G: H ~ H, defined by

1
GO(t)= ~ g(t, u) 4,(u) du, t~[0, 1].
0
Forward, backwardand symmetricstochasticintegration 417

Since G is a symmetric operator. So by ([B], Theorem 3.4.3, p. 115), G is nuclear

iff g has a classical trace. Moreover G is diagonalisable; let consider a complete
orthonormal basis ((p., n=>0) of H, such that for every n, q~. is an eigenvector
of G associated with the eigenvalue 2.. (~o=| (p,.; n > 0, m > 0)) is an orthonormal
basis of HE. Then
g= E a.,,.@.|
?t, m

where a=,,== ~ (o.(u)%=(v)g(u,v)dudv.

[0, 1] 2

We have
1
a=.,. = ~ (p=(u) G(~o,=)(u) du = ,I., c~....
0

where 6.,,. is the Kronecker symbol. Consequently

Since G is nuclear then

~t2.l<oo and ~1(g)=~2..

n n

g(u,v) 1[o,,l(v) d v : Z 2. Io.(u ) S (o.(v) l[o. ,j(v) dv, uE[0, 1].

We introduce

S.(u)= sup (o.(v) lto, l~(v)dv , ue[0, 1].

O<e<l U

It is clear that

ua[O, 1].

By Stein theorem we get,

lira ( 1 ~ (p.(v)leo ,lj(v) dv)=~0"(u), fora.e. u~E0, I],

e-~O+

1 1

(s,(vl) 2 d~<__C I q',(v) 2 d,~.

0 0

Using Cauchy-Schwarz inequality, we have

1

l (Y.I;0.1 I~o.(u)l s,(u)) du =<lfCY. I;4 < o0.

O n n
418 F. Russo and P. Vallois

We can apply the dominated convergence theorem,

T r - ( g ) = lim ~ g(u,v)c%,_(u,v)dudv=~2,=J-~l(g).
g--+ 0 + [0, 1] 2 n

Since g is symmetric, it is obvious that all the A-traces are equal. []

Suppose now that f is a F-valued function defined on [0, 1] z. We introduce
fA a F-valued function defined on l-0, 1] 2 by

I f (s/x t, s v t) if A = +
fJ(s,t)=Jf(svt, sAt) if A = - ' ,
[~(s, t) if A =0.

The function f a is symmetric and f 0 = ~ = ( f + + f - ) / 2 .

Proposition 2.2 Let g be an element of L2([0, lJ2, E). I f gA (resp. g) has a E
classical trace then g has a A-trace and ~ l ( g a ) = T r a ( g ) (resp. ~ l ( g ) = ~ l ( g ~
=Yr~(g~ for every A ~ { + , --,0}).
Proof Assume g is a symmetric function and g admits a E classical trace. Let
yEE. We denote by go the element of H 2 defined by

go(S, t)--(g(s, t), y), (s, t)~ [-0, 1] 2.

go is a symmetric function, admits a N classical trace and

~'-el ( g o ) = ( ~ c l (g), Y)"

Therefore by Proposition 2.1, go has a A-trace. We set

Tr"'a(f)= j" f(u,v) e~,A(U,v)dudv.

[0, 1] 2

Then,

(TRY,a (g), y) = Tr ., a (go).

Consequently, (Tff' a (g), y) converges to ( ~ ( g ) , y); this means that g has a A-trace
and ~ l ( g ) = Tra(g) 9
We study now the general case. We suppose that ga has a E classical trace.
By the previous step, ga has a A-trace and J-d(ga)= Tra(ga). An easy calculation
shows that Tr ~'A(ga)= TP' A(g). In particular, g has a A-trace and

Tra (ga)= Tra (g)= ~ []

Let us write now some consequences for stochastic processes. Let X be an

element of L2(H) (resp. 92,1(H)), D X = ( D s X ( t ) ; (s, t)e[0, lJ 2) the 92.-1 (resp.
L2)-decomposition of DX and
D a X = (DX) A.

Theorem 2.2 I f Da X has a L2 (resp. 92,_ 0 classical trace then D X admits

a L2 (resp. 92, - 1) A-trace and
Forward, backward and symmetric stochastic integration 419

(i) ~ ( D a X) = Tr ~ (DX),
(ii) I ~(e, X) weakly converges to 6X + 3-~l(DAX).
Remarks2.5 1~ Let X be an element of ~2,t(H) and assume that a.s. DX
1 1
admits a A-trace, then J X d ~ B exists and ~ X d ~ B = 6 X + T r a D X , a.s. This is
0 0
an easy consequence of (2.6) and Remark 2.1.
2~ Let 9 be the family of all complete orthonormal basis of H. Then the weak
limit of Ia(e,X) is equal to the symmetric type integral introduced by Zakai
([Z]). Some authors call it also, Ogawa integral ([Nu], [NZ]).
3~ Previous result extends some considerations of [NP] Sect. 7, (see also [N],
Sect. 6). The authors define a class g2 of stochastic processes XeN2,1 (H) such
that (/SX) is "continuous". If X belongs to F~2, then the forward (resp. backward,
symmetric) integral of X is equal to the sum of 6X and the - t r a c e (resp.
+ trace, 0trace) of DX.
Moreover the weak limit of I - ( e , X ) (resp. I+(e,X); I~ is equal to
the forward (resp. backward; symmetric) integral defined by Nualart.
The final result relates the Skorohod integral and the integral defined through
the enlargement of filtration. Assume N=(N;0__<t<I) is a filtration on (O,~-
=(o~,;0<t<l),P), satisfying the usual conditions, and "bigger" than ~- (i.e.
~N for every re[0, 1]). We suppose that
(i) B is a N-semimartingale, B = / 3 + V is the N-canonical decomposition of B,

where Vt = i H (s) d s, 0 < t < 1,

0

(ii) there exists p > 1 such that (H(s); 0<s=< 1) belongs to LP([0, 1] x O, IR).
We know that/3 is a N Brownian motion (see for instance [Je]).
Proposition2.3 Assume X is a N-previsible process, belonging to Lq([0, 1]
x F2,IR)c~Lq(H) where q is the conjugate exponent of p, Tr-(DX) exists in ~ _ ~
and belongs to 12. Then

1 1
(2.10) ~Xd-B= ~XdgB=aX+Tr-(DX) a.s.
0 0

Proof. Through a slight modification in the proof of Proposition 1.1, it is easy

1
to check that ~ X d - B is the limit in 12 of I-(e, X). We have already noticed
0
1 1
that ~ X d - B= ~ XdgB. Theorem 2.l and Remark 2.2 implie the second equality
0 0
in (2.10). U]

Example, Let ~=(N~;0__<t_<i) be the smallest filtration such that, ~ , for

every 0_< t < 1, B(1) is No-measurable. By ([Je], Theorem 3.23 p. 46) we have

B(1)--B(s)
H(s)= --, 0<s<l.
1--s
420 F. Russo and P. Vallois

Since E(IH(s)IP)=(1-s)-P/2E(IB(1)IO, we can a p p l y P r o p o s i t i o n 2.3 with 1 < p

<2.

Acknowledgements. Part of this work has been done during the stay of the first author at
the Ecole Nationale Sup&ieure des T~ldcommunications (D~partement R6seaux). We would like
to thank Professors H. Korezlioglu and S. Ustunel for stimulating discussions. The work of
Russo has been partially supported by a grant of the "Fonds National Suisse de la Recherche
Scientifique'.

References
lAP] Asch, J, Potthoff, J.: Ito lemma without non-anticipatory conditions. Probab. Theory
Relat. Fields 88, 17-46 (1991)
[B] Balakrishnan, A.V.: Applied functional analysis. 2nd edn. Berlin Heidelberg New York:
Springer 1981
[BY] Barlow, M., Yor, M. : Semi-martingale inequalities via Garsia-Rodemich Rumsey lem-
ma. Application to local times. J. Funct. Anal. 49, (2) (1982)
[BH] Bouleau, N., Hirsch, F.: Dirichlet forms and analysis on Wiener space. Berlin New
York: Walter de Gruyter 1991
[BM] Berger, M.A., Mizel, V.J.: An extension of the stochastic integral. Ann. Probab. 10,
(2) 435-450 (1982)
[DM] Dellacharie, C., Meyer, P.A.: Probabilit+s et Potentiel, Chapitres V /t VIII, Th6orie
des martingales. Paris: Hermann 1975
[-DS] Dunford, N., Schwartz, J.T.: Linear Operators. Part I, General Theory. New York:
Wiley-Intersciene 1967
[HM] Hu, Y.Z., Meyer, P.A.: Sur l'approximation des int6grales multiples de Stratonovich.
(Preprint)
[Ja] Jacod, J.: Calcul stochastique et probl+mes de martingales. (Lect. Notes Math., vol. 714)
Berlin Heidelberg New York: Springer 1979
[Je] Jeulin, T.: Semi-martingales et grossissement d'une filtration. (Lect. Notes Math.,
vol. 833) Berlin Heidelberg New York: Springer 1980
[JK] Johnson, G.W., Kallianpur, G.: Some remarks on Hu and Meyer's paper and infinite
dimensional calculus on finite additive canonical Hilbert space. Theory Probab. Appl.
(SIAM) 34, 679-689 (1989)
[K 1] Kunita, H.: On backward stochastic differential equations. Stochastics 6, 293-313 (1982)
[K2] Kunita, H.: Stochastic differential equations and stochastic flow of diffeomorphisms.
Ecole d'6t6 de Saint-Flour XII. (Lect. Notes Math., vol. 1097) Heidelberg New York:
Springer 1982
[KR] Kuo, H.H., Russek, A.: White noise approach to stochastic integration. J. Multivariate
Anal. 24, 218-236 (1988)
[N] Nualart, D. : Non causal stochastic integrals and calculus. Stochastic analysis and relat-
ed topics (Proceedings Silivri 1986). Korzelioglu, H., Ustunel, A.S. (eds.) (Lect. Notes
Math., vol. 1316, pp. 80-129) Berlin Heidelberg New York: Springer 1986
[NP] Nualart, D., Pardoux, E.: Stochastic calculus with anticipating integrands. Probab.
Theory Relat. Fields 78, 535-581 (1988)
[NZ] Nualart, D., Zaka'i, M.: Generalized stochastic integrals and the Malliavin calculus.
Probab. Theory Retat. Fields 73, 255~80 (1986)
[0] Ogawa, S. : Une remarque sur l'approximation de l'int6grale stochastique du type non-
causal par une suite d'int6grales de Stieltjes. Tohoku Math. J. 36, 41-48 (1984)
[RY] Revuz, D., Yor, M.: Continuous martingales and Brownian motion. Berlin Heidelberg
New York: Springer 1991
JR] Rosinski, J.: On stochastic integration by series of Wiener integrals. Technical report
n~ 112. Chapel Hill (1985)
[RV] Russo, F., Vallois, P.: Int6grales progressive, r6trograde et sym6trique de processus
non adapt6s. Note C.R. Acad. Sci. S&. I 312, 615-618 (1991)
[S] Stein, E.M.: Singular integrals and differentiability properties of functions. Princeton:
Princeton University Press 1970
Forward, backward and symmetric stochastic integration 421

[SU] Sold,J.L., Utzet, F.: Stratonovich integral and trace. Stochastics 29, 203-220 (1990)
IT] Thieullen, M.: Calcul stochastique non adatd pour des processus g deux param+tres:
formules de ehangement de variables de type Stratonovitch et de type Skorohod. Pro-
bab. Theory Relat. Fields 89, 457-485 (1991)
[W] Watanabe, S.: Lectures on stochastic differential equations and Malliavin calculus.
Bombay: Tata Institute of Fundamental Research. Berlin Heidelberg New York:
Springer 1984
[Z] Zakai, M.: Stochastic integration, trace and skeleton of Wiener functionals. Stochastics
33, 93-108 (1990)