J o u a rn l o f P c r i ob n o abil r t ity Elec Vol. 8 (2003) Paper no. 14, pages 1–31. Journal URL http://www.math.washington.edu/˜ejpecp/ Itô Formula and Local Time for the Fractional Brownian Sheet Ciprian A. Tudor1 and Frederi G. Viens2 1 Laboratoire de Probabilités, Université de Paris 6 4, place Jussieu, 75252 Paris, France Email: tudor@ccr.jussieu.fr and 2 Dept. Statistics and Dept. Mathematics, Purdue University 150 N. University St., West Lafayette, IN 47907-2067, USA Email: viens@purdue.edu Abstract. Using the techniques of the stochastic calculus of variations for Gaussian processes, we derive an Itô formula for the fractional Brownian sheet with Hurst parameters bigger than 1/2. As an application, we give a stochastic integral representation for the local time of the fractional Brownian sheet. Key words and phrases: fractional Brownian sheet, Itô formula, local time, Tanaka formula, Malliavin calculus. AMS subject classification (2000): primary 60H07; seconday 60G18, 60G15, 60J55. Submitted to EJP on February 18, 2003. Final version accepted on August 20, 2003. 1 1 Introduction Let (Btα )t∈[0,1] be the fractional Brownian (fBm) motion with Hurst parameter α ∈ (0, 1). When α = 1/2, B 1/2 is the standard Brownian motion. When α 6= 1/2, this process is not a semimartingale and in this case the powerful theory and tools of the classical Itô stochastic calculus cannot be applied to it. However, its special properties such as self-similarity, stationarity of increments, and long range dependence make this process a good candidate as a model for different phenomena. Therefore, a development of the stochastic calculus with respect to the fBm was needed. We refer to [1], [6], [9] and [11] for some approaches of this stochastic calculus. The aim of this paper is to develop a stochastic calculus for the two-parameter fBm introduced in [3], also known as the fractional Brownian sheet. As the one-parameter fBm, the fractional Brownian sheet is not a semimartingale. But the Malliavin calculus for Gaussian processes can be adapted to it, Skorohod stochastic integrals can be defined, and it will allow us to derive an Itô formula when the Hurst parameters are bigger than 1/2 which extends the Itô formula for the two parameters martingales (see [10], [16] and [13] for a complete exposition of this topic). As an application we study the representation of the local time of the fractional Brownian sheet in terms of Skorohod stochastic integrals. Our paper is organized as follows: Section 2 contains the basic notions of the Malliavin calculus for Gaussian processes and the definition of the fractional Brownian sheet. In Section 3 we give an analogue of the Doob-Meyer decomposition for the square of a martingale, and an Itô formula for the fractional Brownian sheet. The interpretation of a specific new term appearing in the Itô formula is given in Section 4. Finally, as an application of this formula, Section 5 contains the study of the local time of the fractional Brownian sheet using stochastic integrals and an occupation time formula. 2 2.1 Preliminaries The Malliavin calculus for Gaussian processes Let T = [0, 1]. Consider (Bt )t∈T a centered Gaussian process, with covariance E(Bt Bs ) = R(t, s). We consider the canonical Hilbert space H associated with the process B, defined as the closure of the linear space generated by the function {1[0,t] , t ∈ T } with respect to the scalar product h1[0,t] , 1[0,s] i = R(t, s). Then the mapping 1[0,t] → Bt gives an isometry between H and the first chaos generated by {Bt , t ∈ T } and B(φ) denotes the image of a element φ ∈ H. We can develop a stochastic calculus of variations, or a Malliavin calculus, with respect to B. For a smooth H-valued functional F = f (B(ϕ1 ), · · · , B(ϕn )) with n ≥ 1 , f ∈ Cb∞ (Rn ) and ϕ1 , · · · , ϕn ∈ H we put DB (F ) = n X ∂f (B(ϕ1 ), · · · , B(ϕn ))ϕj ∂xj j=1 2 and DB F will be closable from Lp (Ω) into Lp (Ω; H). Therefore, we can extend D B to the closure of smooth functionals on the Sobolev space D k,p defined by the norm kF kpk,p = E|F |p + k X j=1 kDB,j F kpLp (Ω;H⊗j ) . Similarly, for a given Hilbert space V we can define Sobolev spaces of V -valued random variables D k,p (V ) (see [17]). Consider the adjoint δ B of DB . Then its domain is the class of u ∈ L2 (Ω; H) such that E|hDB F, uiH | ≤ CkF k2 and δ B (u) is the element of L2 (Ω) given by E(δ B (u)F ) = EhD B F, uiH for every F smooth. It is well-known that D 1,2 (H) is included in the domain of δ B . Note that Eδ B (u) = 0 and the variance of the divergence operator δ B is given by Eδ B (u)2 = Ekuk2H + EhD B u, (DB u)∗ iH⊗H (1) where (D B u)∗ is the adjoint of D B u in the Hilbert space H ⊗ H. This last identity implies that Eδ B (u)2 ≤ Ekuk2H + EkD B uk2H⊗H . (2) We will need the property F δ B (u) = δ B (F u) + hD B F, uiH (3) if F ∈ D1,2 and u ∈ Dom(δ B ) such that F u ∈ Dom(δ B ). 2.2 Representation of fBm Let now B = B α be the fractional Brownian motion with parameter α ∈ (0, 1). This process is centered Gaussian, B0 = 0 and ´ 1 ³ 2α (4) s + t2α − |t − s|2α . R(t, s) = 2 We know that B admits a representation as Wiener integral of the form Z t Bt = K(t, s)dWs , (5) 0 where W = {Wt , t ∈ T } is a Wiener process, and K(t, s) is the kernel µ ¶ t α− 21 α− 12 K(t, s) = dα (t − s) +s F1 , s 3 (6) dα being a constant and F1 (z) = dα µ 1 −α 2 ¶Z z−1 0 This kernel satisfies the condition (see [15]): ³ ´ 1 3 θα− 2 1 − (θ + 1)α− 2 dθ. 3 ∂K 1 s 1 (t, s) = dα (α − )( ) 2 −α (t − s)α− 2 . ∂t 2 t (7) (8) α α We will denote by Hα its canonical Hilbert space and by D B and δ B the corresponding Malliavin derivative and Skorohod integral with respect to B α . Fix now α ∈ ( 21 , 1) and define the operator K ? from the set of step functions on T (denoted by E) to L2 (T ) by Z 1 ∂K α (r, s)dr. (K ? f )(s) = f (r) ∂r s It was proved in [1] that the space Hα of the fBm can be represented as the closure of E with respect to the norm kf kHα = kK ? f kL2 (T ) and K ? is an isometry between Hα and a closed subspace of L2 (T ). This fact implies the following relation between the Skorohod integration with respect to B α and W , the Wiener process: α δ B (u) = δ W (K ? u) α (9) α for any Hα -valued random variable u in Dom(δ B ). We will call δ B (u) the Skorohod integral of the process u with respect to B. Since α > 1/2, we have the following representation of the scalar product in Hα : Z 1Z 1 hf, giHα = α (2α − 1) f (a) g (b) |a − b|2α−2 dadb. (10) 0 0 This formula holds as long as the integral above is finite when f and g are replaced by |f | and |g|. 2.3 Fractional Brownian sheet α,β Consider (Ws,t )s,t∈[0,1] a fractional Brownian sheet with parameters α, β ∈ (0, 1). This process is Gaussian, it starts from 0 and its covariance is ³ ´ α,β α,β E Wt,s Wu,v = Rα (t, u)Rβ (s, v) ´ ¢ 1 ³ 2β 1 ¡ 2α s + v 2β − |s − v|2β . t + u2α − |t − u|2α 2 2 This process self similar and with stationary increments and it admits a continuous version. We refer to [3] for the basic properties of W α,β . The fractional Brownian sheet with Hurst parameters α, β ∈ (0, 1) can be defined as (see [4]) Z tZ s α,β Ws,t = K α (t, u)K β (s, v)dWu,v = 0 0 4 where (Wu,v )u,v∈[0,1] is the Brownian sheet. The canonical Hilbert Space Hα,β of the fractional Brownian sheet is the closure of linear space generated by the indicator functions on [0, 1]2 with respect to the scalar product h1[0,t]×[0,s] , 1[0,u]×[0,v] iHα,β = Rα (t, u)Rβ (s, v). Fix α, β > 21 . Notice that in this case, by tensorization of (10), we have for every f, g ∈ H α,β Z 1Z 1Z 1Z 1 f (a, b)g(m, n) |a − m|2α−2 |b − n|2β−n dadbdmdn hf, giHα,β = c(α)c(β) 0 0 0 0 ∗,2 and c(α) = α(2α − 1). Define the operator Kα,β on the space of step function on [0, 1]2 to L2 ([0, 1]2 ) given by Z 1Z 1 ´ ³ ∂K α ∂K β 0 ∗,2 f (r, r0 ) (r, t) (r , s)drdr 0 . (11) Kα,β f (s, t) = ∂r ∂r0 s t We have ∗,2 ∗,2 giL2 ([0,1]2 ) = hf, giHα,β . f, Kα,β hKα,β Indeed, (12) ∗,2 ∗,2 hKα,β f, Kα,β giL2 ([0,1]2 ) ¶ Z 1 Z 1 µZ 1 Z 1 ∂K β ∂K α = (a, u) (b, v)dadb f (a, b) ∂a ∂b 0 0 u v µZ 1 Z 1 ¶ ∂K β ∂K α × (m, u) (n, v)dmdn dudv g(m, n) ∂m ∂n u v ¶ µZ a∧m Z b∧n Z 1Z 1Z 1Z 1 ∂K β ∂K α (m, u) (n, v)dmdndudv dadbdmdn = f (a, b)g(m, n) ∂m ∂n 0 0 0 0 0 0 Z 1Z 1Z 1Z 1 ∂ 2 Rα ∂ 2 Rβ f (a, b)g(m, n) = (a, m) (b, n) ∂a∂m ∂b∂n 0 0 0 0 Z 1Z 1Z 1Z 1 f (a, b)g(m, n) |a − m|2α−2 |b − n|2β−n dadbdmdn = c(α)c(β) 0 0 0 0 = hf, giHα,β . One can develop a Malliavin stochastic calculus with respect to the fractional Brownian sheet following the method of Section 2.1, using rectangles in [0, 1]2 instead of intervals α,β α,β its associated Skorohod integral and Malliavin derivaand DW in [0, 1]. Denote by δ W tive respectively. A consequence of the identity (12) is that the following expression of the α,β α,β Skorohod integral δ W holds: For f ∈ Dom(δ W ) ⊂ L2 (Ω, Hα,β ), Z 1Z 1 Z 1Z 1 ∗,2 α,β f (u, v)dWu,v = (Kα,β f )(u, v)dWu,v 0 0 0 0 ¶ Z 1 Z 1 µZ 1 Z 1 ∂K α ∂K β f (a, b) = (a, u) (b, v)dadb dWu,v . ∂a ∂b v u 0 0 5 All properties of the anticipating Skorohod integration with respect to Gaussian processes will hold in this case. α,β α,β . These processes are real and t → Ws,t Consider now the processes s → Ws,t β α α fractional Brownian motions with the same law as t W and s W β respectively and with covariances ´ ¢ 1¡ 1 ³ 2β 2α 2α 2β R1 (s1 , s2 ) = t2β s2α and R2 (t1 , t2 ) = s2α t1 + t2β − |t − t | 1 2 1 + s2 − |s1 − s2 | 2 2 2 (13) respectively. Since these processes are fBm’s themselves, we have the Malliavin calculus with respect to them. We denote by δ1t the Skorohod integral with respect to the fractional α,β Brownian motion s → Ws,t and by δ2s the Skorohod integral with respect to the fractional α,β Brownian motion t → Ws,t . We will only use the integration by parts formulas ³ ´ α,β t α,β α,β f (Ws,t )δ1 (u) = δ1t f (Ws,t (14) )u + t2β f 0 (Ws,t )hu, 1[0,s] iHα and 3 ³ ´ α,β s α,β α,β f (Ws,t )δ2 (u) = δ2s f (Ws,t )u + s2α f 0 (Ws,t )hu, 1[0,t] iHβ . (15) The Itô formula for the fractional Brownian sheet We will start this section with three technical Lemmas that will be used below, and that rely on the following exponential quadratic growth assumption on f and/or its derivatives: (H) We say that a function g on R satisfies Condition (H) if, for |x| large enough, we have |g (x)| ≤ M exp ax2 where a is a constant that depends only on α and/or β, and M is an arbitrary constant. Lemma 1 Let s, t ∈ [0, 1] and π 1 : 0 = s0 < s1 < . . . < sn = s and π 2 : 0 = t0 < t1 < . . . < tm = t be two partitions of the intervals [0, s] and [0, t] respectively. Assume f , f 0 , and f 00 satisfy Condition (H). Then we have the following convergences in L 2 (Ω) when the partition meshes both tend to zero: 1 If (W α )t∈[0,1] is a fBm with parameter α ∈ ( 21 , 1), then for any f ∈ Cb2 (R), Z t m−1 ´ m−1 X ³ X Z tj+1 α 2α α f (Wuα )u2α dWuα . f (W )t dW → δ f (Wtαj )t2α 1 (·) = [tj ,tj+1 ] tj j u j j=0 j=0 0 tj (16) α,β 2 If (Ws,t )s,t∈[0,1] is a fractional Brownian sheet with parameters α, β ∈ ( 12 , 1), then for any f ∈ Cb2 (R), n−1 ´ n−1 X m−1 X ³ X m−1 X Z si+1 Z tj+1 α,β α,β f (Wsα,β )dWu,v (17) δ 1[si ,si+1 ]×[tj ,tj+1 ] (·)f (Wsi ,tj ) = i ,tj i=0 j=0 Z tZ s → 0 0 i=0 j=0 α,β α,β f (Wu,v )dWu,v . si tj (18) 6 Proof: We use the notation ∆j = [tj , tj+1 ]. Concerning point 1, we prove first the convergence in L2 (Ω; Hα ) of the sum m−1 X³ f (Wtαj )t2α j 1∆j (u) j=0 ´ (19) to the process f (Wuα )u2α 1[0,t] (u) = m−1 X f (Wuα )u2α 1∆j (u). (20) j=0 The L2 -norm of the difference D between the two terms can be estimated as follows ¯ ¯2 ¯m−1 ¯ ³ ´ X ¯ ¯ 2α α α 2α ¯ D = E¯ f (Wtj )tj − f (W· ) (·) 1∆j ¯¯ ¯ j=0 ¯ α H ¯ ¯2 ¯m−1 ¯ ¯ ¯ X ¯¯ ¯ ¯ ¯ ≤ 2E ¯¯ ¯f (Wtαj ) − f (W·α )¯ t2α j 1 ∆j ¯ ¯ j=0 ¯ α H ¯2 ¯ ¯ ¯m−1 ¯ ¯X ¯ |f (W·α )| | (·)2α − t2α + 2E ¯¯ j |1∆j ¯ ¯ α ¯ j=0 H We can write that k P aj k2Hα ≤ P |aj |Hα |ak |Hα . Also we have the inequality ¯ ¯ ¯ ¯ 0 ¯¯ ¯ α¯ α ¯ α ¯¯ α α ¯ − W W ≤ sup ) − f (W ) f (W ) f (W ¯ tj ¯ u ¯. u ¯ x tj j j,k x∈[0,1] This follows from the mean value theorem and the ¯fact¯ that W is almost-surely continuous. Since here we will have u ∈ ∆j , so that |u − tj | ≤ ¯π 2 ¯, it also follows that ¯ ¯ ¯ ¯ ¯ ¯ sup |Waα − Wbα | . ¯f (Wtαj ) − f (Wuα )¯ ≤ sup ¯f 0 (Wxα )¯ a,b:|a−b|≤|π 2 | x∈[0,1] Therefore, using Hölder’s inequality, # " " ¯ ¯ 0 4 D ≤ 2E sup ¯f (Wxα )¯ E x∈[0,1] + 2E " sup |f x∈[0,1] (Wxα )|2 # sup a,b:|a−b|≤|π 2 | |π 2 |4α |Waα − Wbα |4 # X j,k h1∆j , 1∆k iHα X h1∆j , 1∆k iHα . j,k Since W is Gaussian and almost-surely continuous, the general Dudley-Fernique theory h i 4 α α guarantees that E supa,b:|a−b|≤r E |Wa − Wb | converges to 0 as r tends to 0. To guarantee that D goes to zero as |π 2 | → 0, we only need to assume that condition (H) holds for f and 7 f 0 with a < 8−1 V ar(Z)−1 where Z = supx∈[0,1] |Wxα |. Indeed, since Z is the supremum of a £ ¤ centered continuous Gaussian process, a classical result of Fernique implies that E exp aZ 2 is finite for a < V ar(Z)−1 2−1 . Using the inequality (2), the proof of point 1 will be finished if we show that the derivative of (19) converges to the derivatives of (20). It holds that Ds f (Wtαj ) = f 0 (Wtαj )1[0,tj ] (s) and Ds f (Wuα ) = f 0 (Wuα )1[0,u] (s) and ¯ ¯ ¯m−1 ¯ ³ ´ X ¯ ¯ α 2α α 2α ¯ Ds f (Wtj )tj − Ds f (Wu )u 1∆j (u)¯¯ ¯ ¯ j=0 ¯ ¯ ¯ ¯ ¯m−1 ´ ¯ ¯X ³ 0 α 0 α 2α ¯ f (Wtj ) − f (Wu ) 1[0,tj ] (s)tj 1∆j (u)¯¯ ≤ E¯ ¯ ¯ j=0 ¯ ¯ ¯m−1 ¯ ¯X 0 ¯ 2α ¯ + E ¯¯ f (Wuα )(t2α − u )1 (s)1 (u) ∆j [0,tj ] j ¯ ¯ j=0 ¯ ¯ ¯ ¯ ¯m−1 ¯ ¯X 0 α 2α f (Wu )u 1[tj ,u] (s)1∆j (u)¯¯ . + E ¯¯ ¯ ¯ j=0 We seek convergence of these three terms as functions of (s, u) in the space (H α )⊗2 . The convergence of the first two terms to zero can be done using the bounds presented above in this proof, using Condition (H) on f 0 and f 00 . Finally, using the definition of the scalar product in (Hα )⊗2 , including the fact that we can use the inequality 1[tj ,u] (s) ≤ 1∆j (s), we obtain ¯2 ¯ ¯ ¯m−1 ¯ ¯X 0 2α α ¯ f (W· ) (·) 1[tj ,·] (?)1∆j (·)¯¯ E¯ ¯ α ⊗2 ¯ j=0 (H ) # " ¯ ¯2 X ® ­ h1∆ (·) , 1∆ (·)iHα 1∆ (?) , 1∆ (?) α ≤ E sup ¯f 0 (Wxα )¯ j x∈[0,1] ≤E " k j k H j,k 2 # X ¯ 0 ¯ ° ° 2 °1 ∆ ° 2 α . sup ¯f (Wxα )¯ j H x∈[0,1] j h i Condition (H) on f 0 guarantees that E supx∈[0,1] |f 0 (Wxα )|2 is finite, and the other factor ¯ ¯P ¯ ¯ P converges to 0 with ¯π 2 ¯ since j k1∆j k2Hα ≤ ¯π 2 ¯ j k1∆j kHα converges to zero. Concerning point 2, we need first to prove that n−1 X m−1 X i=0 j=0 ´ ³ α,β ) − f (W ) 1∆i ×∆j (u, v) f (Wsα,β u,v i ,tj 8 (21) ¡ ¢ goes to zero in L2 Ω; Hα,β . We have, as before, ¯2 ¯ ¯n−1 ³ ´¯¯ X ¯X m−1 α,β ¯ 1∆i ×∆j (u, v) f (Wsα,β ) − f (Wu,v ) ¯ E ¯¯ i ,tj ¯ α,β ¯ i=0 j=0 H # " " ¯ ³ ´¯4 ¯ α,β ¯ sup ≤ E1/2 sup ¯f 0 Wx,y ¯ E1/2 |a1 −b1 |≤|π 1 |,|a2 −b2 |≤|π 2 | x,y∈[0,1] · X i,i0 ,j,j 0 h1∆i , 1∆i0 iHα h1∆j , 1∆j 0 iHβ ¯2 ¯ ¯ α,β α,β ¯ ¯Wa1 ,a2 − Wb1 ,b2 ¯ # ¯ ¯ ¯ ¯ which, assuming condition (H) on f 0 , converges to zero as ¯π 1 ¯ and ¯π 2 ¯ tend to zero, due to the almost-sure continuity of W α,β as a centered Gaussian process on [0, 1]2 . The convergence of the Malliavin derivatives of (21) is no more difficult than the other proofs above; we leave it to the reader. The proof of the lemma is completed. ¤ The next lemma is trivial. Lemma 2 Let s, t ∈ [0, 1], π 1 , π 2 as in Lemma 1 and A, B two finite variation processes. Then, for every f ∈ C(R) and g ∈ C(R2 ) the following convergences hold in L1 (Ω). m−1 X j=0 and n−1 X m−1 X i=0 j=0 ¡ ¢ f (tj ) Btj+1 − Btj → Z t f (u)dBu (22) 0 ¡ ¢¡ ¢ g(si , tj ) Asi+1 − Asi Btj+1 − Btj → Z 0 sZ t g(u, v)dAu dBv (23) 0 The next lemma is useful for dealing with certain specific double Skorohod integrals that we will encounter. Lemma 3 For s, t ∈ [0, 1] and for fixed n and m, let the partitions π 1 , π 2 be again ¢ ¡ 1 integers 2 . Also let as in Lemma 1. Denote π = πn,m = πn , πm aπ(u1 ,v1 );(u2 ,v2 ) = n−1 X m−1 X 1[si ,si+1 ]×[0,tj ] ((u1 , v1 ))1[0,si ]×[tj ,tj+1 ] ((u2 , v2 )) (24) i=0 j=0 ¡ ¢⊗2 for every u = (u1 , u2 ), v = (v1 , v2 ) ∈ [0, 1]2 . Then the sequence (aπ )π converges in Hα,β as the mesh of π tends to 0 (when n and m both tend to infinity) to the function a defined for every (u1 , v1 ) , (u2 , v2 ) ∈ [0, 1]2 by: a ((u1 , v1 ) ; (u2 , v2 )) = 1[0,s] (u1 ) 1[0,t] (v2 ) 1[0,u1 ] (u2 ) 1[0,v2 ] (v1 ) . 9 Proof: We denote s0 = t0 = 0 and ∆i = ∆ni = [si ; si+1 ], ∆j = ∆m j = [tj ; tj+1 ], as well as Di = Din = [0, si ] and Dj = Djm = [0, tj ]. Although the names of these intervals are ambiguous, the names of their indices allow us to identify which variables they refer to, which lightens the notation. To further lighten notation, we allow D j (t) to denote the indicator function 1Dj (t), and similarly for the other intervals. We only need to show that the doubly indexed sequence of functions [an,m ((u1 , v1 ) , (u2 , v2 )) − a ((u1 , v1 ) , (u2 , v2 ))] £ ¡¡ ¢ ¡ ¢¢ ¡¡ ¢ ¡ ¢¢¤ · an,m u01 , v10 , u02 , v20 − a u01 , v10 , u02 , v20 n−1 X X m−1 Dj (v1 ) ∆i (u1 ) Di (u2 ) ∆j (v2 ) − a ((u1 , v1 ) , (u2 , v2 )) = i=0 j=0 · n−1 X X m−1 i=0 j=0 ¡ 0¢ ¡ 0¢ ¡ 0¢ ¡ 0¢ ¡¡ 0 0 ¢ ¡ 0 0 ¢¢ D j v1 ∆ i u 1 D i u 2 ∆ j v2 − a u 1 , v1 , u 2 , v 2 (25) ¢2 ¡ defined on [0, 1]2 × [0, 1]2 converges to 0 as m and n tend to infinity in the space L1 (µ) where µ is the measure defined by ¡ ¢ µ du1 du2 dv1 dv2 du01 du02 dv10 dv20 ¯ ¯2α−2 ¯ ¯ ¯ ¯ ¯ ¯ ¯u2 − u02 ¯2α−2 ¯v1 − v10 ¯2β−2 ¯v2 − v20 ¯2β−2 du1 du2 dv1 dv2 du01 du02 dv10 dv20 . = ¯u1 − u01 ¯ We first note that, by virtue of the partitions, for every pair of points (u1 , v1 ) , (u2 , v2 ) ∈ [0, 1]2 , except for the countably many points in the partitions π 1 and π 2 , we have convergence of an,m ((u1 , v1 ) ; (u2 , v2 )) to a ((u1 , v1 ) ; (u2 , v2 )). In other words, an,m converges to a Lebesgue-almost everywhere in [0, 1]2 × [0, 1]2 . Thus the quantity in (25) converges ¡ ¢2 in [0, 1]2 × [0, 1]2 almost everywhere to 0 with respect to the Lebesgue measure, and also with respect to the measure µ since µ is Lebesgue-absolutely continuous. Since the sequence in (25) is bounded by 1, and since 1 is integrable with respect to µ, by dominated convergence, the lemma is proved. ¤ We will also need the basic lemma for the convergence of the Skorohod integral. ¡ ¢ Lemma 4 Let un be a sequence of elements in Dom(δ) which converges to u in L2 Ω; Hα,β . Suppose that δ(un ) converges in L2 (Ω) to some square integrable random variable G. Then u belongs to the domain of δ and δ(u) = G. This result also holds for sequences in H α , or in Hβ , and their respective Skorohod integrals. 3.1 α,β 2 A decomposition for (Ws,t ) Let, from now on, s, t ∈ [0, 1] and π 1 : 0 = s0 < s1 < . . . < sn = s and π 2 : 0 = t0 < t1 < . . . < tm = t be two partitions of the intervals [0, s] and [0, t] respectively. Without loss of generality, we can choose the dyadic partitions. 10 Proposition 1 Let ( 21 , 1). Then it holds ³ ³ α,β Ws,t α,β Ws,t ´ s,t∈[0,1] ´2 =2 a fractional Brownian sheet with parameters α, β ∈ Z tZ 0 s 0 α,β α,β Wu,v dWu,v + 2M̃s,t + s2α t2β (26) where, with a and aπ as in Lemma 3, and with δ (2) the double Skorohod integral with respect to W α,β on [0, 1]2 × [0, 1]2 , n−1 X X m−1 M̃s,t := lim δ (2) 1[si ,si+1 ]×[0,tj ] (·)1[0,si ]×[tj ,tj+1] (?) (27) |π|→0 = lim δ i=0 j=0 (2) |π|→0 π (a ) = δ (2) (a) . Although this result is contained in the Itô formula that we prove independently in ¡ ¢2 the next section, we give a detailed self-contained proof of this decomposition for W α,β as a dydactic tool that enables us to introduce some of the essential ingredients and techniques used in proving Itô’s formula, including the random field M̃ . This section helps us make the next section more concise and easier to read. Proof of the proposition: First write Taylor’s formula α,β 2 (Ws,t ) =2 n−1 X i=0 ´ ³ ¯ α,β α,β W − W Wsα,β s ,t s ,t ,t i+1 i i ¯ where Wsα,β is an arbitrary point located between Wsα,β and Wsα,β . Observe first that the i ,t i ,t i+1 ,t 2 1 L (Ω)-limit as |π | → 0 of the difference n−1 X³ i=0 ¯ Wsα,β − Wsα,β i ,t i ,t ´³ Wsα,β − Wsα,β i+1 ,t i ,t ´ is equal to 0. Indeed, by Hölder inequalities, ¯n−1 ¯ ¯X ³ ´³ ´¯2 ¯ ¯ ¯ Wsα,β − Wsα,β Wsα,β − Wsα,β E¯ ¯ i ,t i ,t i+1 ,t i ,t ¯ ¯ i=0 n−1 X ≤E ≤ ≤ i,l=0 sup ¯³ ¯¯ ¯ ´³ ´ ¯ α,β α,β α,β α,β α,β ¯ α,β ¯ ¯ α,β Wsα,β − W ¯Wsi+1 ,t − Wsi ,t ¯ ¯Wsl+1 ,t − Wsl ,t ¯ Wsi+1 ,t − Wsi ,t ,t s ,t l+1 l |a−b|≤|π 1 | sup |a−b|≤|π 1 | µ ¯ ¯2 ¶ 12 ¯2 ¶ 12 n−1 X µ ¯¯ α,β ¯ α,β α,β α,β α,β ¯ α,β ¯ E ¯(Wsi+1 ,t − Wsi ,t )(Wsl+1 ,t − Wsl ,t )¯ E ¯Wa,t − Wb,t ¯ i,l=0 à ! µ ¯ µ ¯ ¯2 ¶ 12 n−1 ¯4 ¶ 41 2 X ¯ ¯ α,β ¯ ¯ α,β E ¯Wa,t − Wb,t ≤ |π 1 |4α−2 − Wsα,β E ¯Wsα,β ¯ ¯ i+1 ,t i ,t i=0 11 and this goes to 0 when α > 21 . Therefore it suffices to study the limit of the sum T =2 n−1 X i=0 Using that g(t) = g(0) + T =2 n−1 Xh X m−1 i=0 j=0 =2 n−1 X m−1 X i=0 j=0 i=0 j=0 =2 n−1 X m−1 X i=0 j=0 =I +J j=0 (g(tj+1 ) − g(tj )) we will obtain ³ ´ ³ ´i α,β α,β α,β α,β α,β Wsα,β W − W − W W − W si+1 ,tj+1 si ,tj+1 si ,tj si+1 ,tj si ,tj i ,tj+1 ´ ³ α,β α,β α,β α,β + W − W − W W Wsα,β si ,tj si+1 ,tj si ,tj+1 si+1 ,tj+1 i ,tj+1 n−1 X³ X m−1 +2 Pm−1 ³ ´ α,β α,β Wsα,β W − W si+1 ,t si ,t . i ,t Wsα,β − Wsα,β i ,tj+1 i ,tj ´³ ¢ ¡ Wsα,β δ 1∆i ×∆j i ,tj+1 Wsα,β − Wsα,β i+1 ,tj i ,tj +2 n−1 X m−1 X³ i=0 j=0 ´ − Wsα,β Wsα,β i ,tj+1 i ,tj ´³ Wsα,β − Wsα,β i ,tj i+1 ,tj ´ where we denoted by ∆i = [si , si+1 ] and ∆j = [tj , tj+1 ]. Now, by property (3), the summand I becomes I=2 n−1 X m−1 X i=0 j=0 =2 n−1 X m−1 X i=0 j=0 n−1 ³ ´ X m−1 X α,β δ Wsi ,tj+1 1∆i ×∆j + 2 h1[0,si ]×[0,tj+1 ] , 1∆i ×∆j iHα,β i=0 j=0 ³ ´ δ Wsα,β +2 1 ∆ ×∆ ,t i j i j+1 n−1 X m−1 X i=0 j=0 h1[0,si ] , 1∆i iHα h1[0,tj+1 ] , 1∆j iHβ . By Lemma 1 point 2, and Lemma 2, we have the following convergences in L2 (Ω) n−1 X m−1 X i=0 j=0 Z tZ ³ ´ δ Wsα,β 1 → i ,tj+1 ∆i ×∆j 0 s 0 α,β α,β Wu,v dWu,v (28) and n−1 X X m−1 i=0 j=0 = h1[0,si ] , 1∆i iHα h1[0,tj+1 ] , 1∆j iHβ ´ ¢ X ³ 2β 1 1 X ¡ 2α 2β 2α tj+1 − t2β → s2α t2β . si+1 − s2α i − (si+1 − si ) j − (tj+1 − tj ) 4 4 i j 12 (29) The term J admits the following decomposition J =2 n−1 X X m−1 δ i=0 j=0 = 2M (π) + 2 h³ n−1 ´ i X X m−1 α,β Wsα,β − W 1 + 2 h1[0,si ] , 1∆i iHα h1[0,tj ] , 1∆j iHβ ∆i ×[0,tj ] si ,tj i ,tj+1 i=0 j=0 n−1 X X m−1 i=0 j=0 h1[0,si ] , 1∆i iHα h1[0,tj ] , 1∆j iHβ where M (π) = n−1 X X m−1 i=0 j=0 = δ (2) δ h³ ´ i α,β Wsα,β − W 1 ∆ ×[0,t ] ,t s ,t i j i j+1 i j n−1 X X m−1 i=0 j=0 1∆i ×[0,tj ] (·)1[0,si ]×∆j (?) = δ (2) (aπ ) (30) where δ (2) denotes the double Skorohod integral and the process aπ is defined in Lemma 3. As before, we can show that n−1 X X m−1 i=0 j=0 h1[0,si ] , 1∆i iHα h1[0,tj ] , 1∆j iHβ →|π|→0 1 2α 2β s t in L2 . 4 (31) Lemma 4 can be used in combination with Lemma 3 to assert immediately that δ (2) (aπ ) converges in L2 (Ω) to δ (2) (a); indeed the convergence of the deterministic integrands in ¡ α,β ¢⊗2 H is in Lemma 3, and since the corresponding Skorohod integrals are equal to α,β 2 (Ws,t ) minus other terms that were already proved to converge in L2 (Ω), they converge in L2 (Ω). Note that Lemma 4 is not stated for the convergence of the double Skorohod integral, but the simple one; however it can be still applied because a double integral can be written as a simple one in which the integrand is convergent in L2 (Ω; Hα,β )). Combining this convergence and (28), (29), (31), we find the result of the proposition. ¤ 3.2 Itô formula α,β In the sequel of this article we will make use of the notation du Wu,v to denote the differential α,β α,β α,β . As before, dWu,v of the f Bm given by u 7→ Wu,v for fixed v, and similarly for dv Wu,v denote the differential of the fractional Brownian sheet with respect to both parameters. α,β Theorem 1 Let f ∈ C 4 (R) and (Ws,t )s,t∈[0,1] a fractional Brownian sheet with α, β ∈ 13 ( 12 , 1). Also assume that f, f 0 , f 00 , f 000 and f (iv) satisfy Condition (H). Then it holds Z tZ s α,β α,β α,β f (Ws,t ) = f (0) + f 0 (Wu,v )dWu,v 0 0 Z tZ s Z tZ s α,β 00 α,β 2α−1 2β−1 f 00 (Wu,v )dM̃u,v f (Wu,v )u v dudv + + 2αβ 0 0 0 0 Z tZ s Z tZ s α,β 2α 2β−1 α,β 000 α,β 2α−1 2β α,β f 000 (Wu,v )u v du Wu,v dv f (Wu,v )u v dudv Wu,v + β +α 0 0 0 0 Z tZ s α,β 4α−1 4β−1 f iv (Wu,v )u v dudv (32) + αβ 0 0 where, by definition, Z tZ 0 s 0 α,β f 00 (Wu,v )dM̃u,v = lim δ (2) |π|→0 n−1 X X m−1 i=0 j=0 ³ ´ f 00 Wsα,β 1[0,si ]×[tj ,tj+1 ] (·)1[si ,si+1 ]×[0,tj ] (?) i ,tj where the last limit exists in L2 (Ω). More specifically, we have the following formula, which can be taken as a definition of the integral with respect to M̃ which would otherwise only be a formal notation: Z Z t 0 s 0 α,β f 00 (Wu,v )dM̃u,v := δ (2) (N ) where for every (u1 , v1 ) , (u2 , v2 ) ∈ [0, 1]2 the function N is defined by: ³ ´ N ((u1 , v1 ) ; (u2 , v2 )) = f 00 Wuα,β 1[0,s] (u1 ) 1[0,t] (v2 ) 1[0,u1 ] (u2 ) 1[0,v2 ] (v1 ) . 1 ,v2 Proof: Let s, t ∈ [0, 1] and write Taylor’s formula with fixed t. It holds that f ³ α,β Ws,t ´ = f (0) + n−1 X i=0 + n−1 X i=0 ´ ´³ ³ α,β α,β W − W f 0 Wsα,β si+1 ,t si ,t i ,t ´2 ³ ´³ ¯ α,β α,β f 00 Wsα,β W − W si+1 ,t si ,t i ,t 14 (33) (34) ¯ where Wsα,β is a point on the segment from Wsα,β to Wsα,β . First, we can prove that the i ,t i+1 ,t i ,t 2 last term goes to zero in L (Ω). Indeed, by Hölder’s inequality, ¯ ¯ ¯X ³ ´³ ´2 ¯2 ¯α,β ¯ ¯ α,β α,β 00 f Wsi ,t E¯ Wsi+1 ,t − Wsi ,t ¯ ¯ ¯ i ≤E n−1 X i,l=0 ≤ E1/2 ≤ CK " ¯ ³ ´ ³ ´¯ ³ ´2 ³ ´2 ¯α,β ¯α,β ¯ ¯ 00 α,β α,β α,β Wsα,β − W ¯f Wsi ,t f 00 Wsl ,t ¯ Wsi+1 ,t − Wsi ,t sl ,t l+1 ,t # ·¯ ·¯ ¯8 ¸ ¯8 ¸ ¯ ³ ´¯4 X ¯ 00 ¯ α,β ¯ α,β ¯ 1/4 ¯ α,β 1/4 ¯ α,β α,β E sup ¯f Wx ¯Wsl+1 ,t − Wsl ,t ¯ ¯Wsi+1 ,t − Wsi ,t ¯ E ¯ x∈[0,1]2 X i,l |si+1 − si | i,l 2α |sl+1 − sl |2α where K is a constant obtained by using Condition (H) on f 00 , C is a universal constant; this all tends clearly to zero since 2α > 1. It remains to study the limit of the term T = n−1 X i=0 ´ ´³ ³ α,β α,β . W − W f 0 Wsα,β s ,t s ,t ,t i+1 i i Writing the fact that g(t) = g(0) + T = n−1 Xh X m−1 i=0 j=0 = n−1 X m−1 X i=0 j=0 + j=0 (g(tj+1 ) − g(tj )) we get ³ ´³ ´i ³ ´³ ´ α,β α,β α,β α,β α,β 0 W W − W W − W − f f 0 Wsα,β si ,tj si+1 ,tj si ,tj si+1 ,tj+1 si ,tj+1 i ,tj+1 ³ ´³ ´ α,β α,β α,β α,β f 0 Wsα,β W − W − W + W ,t s ,t s ,t s ,t s ,t i j+1 i+1 j+1 i j+1 i+1 j i j n−1 X m−1 X³ i=0 j=0 Pm−1 ´ ³ ´ ´´ ³ ³ α,β α,β α,β 0 f 0 Wsα,β − f W W − W ,t s ,t s ,t s ,t i j+1 i+1 j+1 i j+1 i j = A + B. 3.2.1 The estimation of the term A Note that ³ ´ ¢ ¡ α,β α,β α,β (·) Wsα,β − W − W + W = δ 1 ∆ ×∆ ,t s ,t s ,t s ,t i j i+1 j+1 i j+1 i+1 j i j 15 and by the properties of the Skorohod integral we obtain A= n−1 X X m−1 i=0 j=0 + n−1 X m−1 X i=0 j=0 h ³ ´ i δ f 0 Wsα,β 1 (·) ∆ ×∆ i j i ,tj+1 ³ ´ h1[0,si ]×[0,tj+1 ] , 1∆i ×∆j iHα,β f 00 Wsα,β ,t i j+1 = A1 + A2 . The convergence of A1 follows from Lemma 1 point 2, assuming condition (H) for f 0 , f 00 and f 000 , yielding n−1 X X m−1 h δ f i=0 j=0 0 ³ Wsα,β i ,tj+1 ´ i 1∆i ×∆j (·) →L2 (Ω) Z tZ 0 s 0 α,β α,β f 0 (Wu,v )dWu,v . (35) Concerning the term A2 , we can write A2 = n−1 X m−1 X i=0 j=0 = n−1 X m−1 X i=0 j=0 ³ ´ f 00 Wsα,β h1[0,si ] , 1∆i iHα h1[0,tj+1 ] , 1∆j iHβ i ,tj+1 ´ ³ ´1¡ ¢ ³ 2β 2β 2β 2α 2α 2α t − t − (t − t ) s − s − (s − s ) f 00 Wsα,β j+1 j i+1 i i j+1 j i ,tj+1 4 i+1 and as above this converges, due to Lemma 2, to Z t ³ Z t ³ ´ ´ α,β α,β f 00 Wu,v f 00 Wu,v u2α−1 v 2β−1 dudv dRuα dRvβ = αβ (36) 0 0 where Ruα = Rα (u, u) = u2α and Rvβ = Rβ (v, v) = v 2β . due to Lemma 2. 3.2.2 Since The estimation of the term B ³ ´ ³ ´ ´³ ´ ³ ¯α,β α,β α,β α,β 0 00 f 0 Wsα,β − f W = f W W − W si ,tj si ,tj si ,tj+1 si ,tj i ,tj+1 ¯ where Wsα,β is a point located between Wsα,β and Wsα,β , the term B becomes i ,tj i ,tj i ,tj+1 B= X i,j ´³ ´³ ´ ³ ¯ α,β α,β α,β α,β f 00 Wsα,β W − W W − W si ,tj+1 si ,tj si+1 ,tj si ,tj i ,tj and using the argument as above, assuming Condition (H) for f 000 , we can prove that the limit of B is the same as the limit of the term ´³ ´ X ³ α,β ´ ³ α,β α,β α,β B0 = f 00 Wsi ,tj Wsi ,tj+1 − Wsα,β W − W si+1 ,tj si ,tj . i ,tj i,j 16 It holds that B0 = X i,j + X i,j ´i ³ ´ h ³ α,β α,β − W W δ 1 f 00 Wsα,β [0,si ]×∆j si ,tj si+1 ,tj i ,tj f 00 ³ Wsα,β i ,tj ´ h1[0,si ] , 1∆i iHα h1[0,tj+1 ] , 1∆j iHβ = B1 + B2 . The second term is similar with the term A2 studied before and we have its convergence to Z t ³ Z t ³ ´ ´ α,β α,β dRuα dRvβ = αβ u2α−1 v 2β−1 dudv. (37) f 00 Wu,v f 00 Wu,v 0 0 For the term B1 we can write ´i ³ X ³ α,β ´ s h α,β − W B1 = f 00 Wsi ,tj δ2i 1∆j (·) Wsα,β s ,t ,t i j i+1 j i,j = X i,j ´i ´ ³ h ³ α,β α,β − W 1 (·) W δ2si f 00 Wsα,β ∆ si ,tj si+1 ,tj j i ,tj ´ ³ ´³ ³ ´ α,β α,β 2β 2β 2α 1 2β f 000 Wsα,β s W − W t − t − (t − t ) j+1 j i si+1 ,tj si ,tj j i ,tj 2 j+1 i,j ´i ³ X s h ³ α,β ´ tj α,β − W = δ2i f 00 Wsi ,tj δ1 (1∆i (?)) 1∆j (·) Wsα,β si ,tj i+1 ,tj + X + X i,j i,j ³ ´ ³ ´ tj 2β 2β 2α 1 2β f 000 Wsα,β δ (1 (?)) s − (t − t ) − t t j+1 j ∆ i i 1 j i ,tj 2 j+1 and by applying the integration by parts several times for the Skorohod integral with respect to the corresponding fractional Brownian motion, we obtain i X s tj h ³ α,β ´ B1 = δ2i δ1 f 00 Wsi ,tj 1∆j (·)1∆i (?) i,j h ³ ´ i1¡ ¢ 2α 2α 2β s2α tj δ2si f 000 Wsα,β 1 (·) ∆ i+1 − si − (si+1 − si ) j i ,tj 2 i,j ³ ´¸ X tj · ³ α,β ´ 2β 2β 000 2α 1 2β + δ1 f Wsi ,tj 1∆i (?)si t − tj − (tj+1 − tj ) 2 j+1 i,j ´ ¡ ³ X ¢ 2β 2β 2β 1 2α 2α 2β 2α 1 − (t − t ) − t s2α tj t + f (iv) (Wsα,β )s j+1 j i+1 − si − (si+1 − si ) i j j+1 i ,tj 2 2 + X i,j = B11 + B12 + B13 + B14 . For the summand B12 we have i¡ ¢ 1 X si h 000 ³ α,β ´ 2β 2α B12 = δ2 f Wsi ,tj tj 1∆j (·) s2α i+1 − si 2 i,j i h ³ ´ X 1 2β + (·) (si+1 − si )2α . δ2si f 000 Wsα,β t 1 ∆ ,t j j i j 2 i,j 17 By Lemma 1 point 1, assuming condition (H) for f 000 and f (iv) , it holds that, for every si m−1 X j=0 δ2si h f 000 ³ Wsα,β i ,tj ´ t2β j 1∆j (·) i →L2 (Ω) Z t 0 ³ ´ f 000 Wsα,β v 2β dv Wsα,β i ,v i ,v and this fact together with Lemma 2 implies that the first part of B12 goes to the integral R s R t 000 ³ α,β ´ 2α−1 α,β α 0 0f Wu,v u duv 2β dv Wu,v . Using Condition (H) on f 000 , since 2α > 1 it is easy to observe that the last part of B12 converges to zero and therefore we have the convergence Z sZ t ³ ´ α,β α,β f 000 Wu,v B12 → α u2α−1 duv 2β dv Wu,v . (38) 0 In the same way Z B13 → β 0 0 sZ t 0 ³ ´ α,β α,β u2α v 2β−1 du Wu,v dv. f 000 Wu,v (39) Lemma 2 also gives that B1,4 converges to 1 4 Z 0 sZ t f (iv) 0 α,β 2α 2β (Wu,v )u v dRα (u)dRβ (v) = αβ Z 0 sZ t 0 α,β 4α−1 4β−1 f (iv) (Wu,v )u v dudv. (40) Now, note that B11 = X i,j h ³ ´ i δ (2) f 00 Wsα,β 1 (·)1 (?) [0,s ]×∆ ∆ ×[0,t ] i j i j i ,tj where δ (2) denotes again the double Skorohod integral, and specifically, h ³ ´ i Z si Z tj+1 Z si+1 Z tj α,β (2) 00 δ f Wsi ,tj 1[0,si ]×∆j (·)1∆i ×[0,tj ] (?) = f 00 (Wsα,β )dWuα,β dWuα,β . 1 ,v1 2 ,v2 i ,tj 0 tj si 0 We now show Lemma 5 Using the same partitions as in Lemma 3, define the process X ³ α,β ´ π N(u = f 00 Wsi ,tj 1∆i ×[0,tj ] ((u1 , v1 ))1[0,si ]×∆j ((u2 , v2 )). 1 ,v1 );(u2 ,v2 ) (41) i,j ³ ¡ ¢⊗2 ´ to Then assuming condition (H) for f 00 the sequence (N π )π converges in L2 Ω; Hα,β the function N defined in the statement of Theorem 1. Proof: The proof uses Condition (H) combined with the proof of Lemma 3. We ¡ ¢⊗2 only need to establish convergence in Hα,β for fixed randomness ω ∈ Ω. Indeed, the ´ ³ the ¡ α,β ¢⊗2 2 follows by dominated convergence under condition (H) convergence in L Ω; H 18 ¯ ¯¢−1 ¡ for f 00 , as long as a < 2−1 V ar max[0,1]2 ¯W α,β ¯ for instance. Now, using the almosteverywhere convergence established in proving Lemma 3, and noticing that we can rewrite N π as X ³ α,β ´ π N(u = f 00 Wsi ,tj 1∆i (u1 )1∆j (v2 ) 1 ,v1 );(u2 ,v2 ) i,j X · 1∆i ×[0,tj ] ((u1 , v1 ))1[0,si ]×∆j ((u2 , v2 )) , i,j we only need to show that for almost every ω ∈ Ω, for fixed u1 , v2 , the quantity ´ X ³ α,β 00 f Wsi ,tj (ω) 1∆i (u1 )1∆j (v2 ) i,j ³ ´ α,β is almost-surely continuous on [0, 1]2 , this conconverges to f 00 Wuα,β 1 ,v2 (ω) . Since W vergence is trivial. The lemma is proved. ¤ To guarantee that Lemma 4 can be invoked to conclude that the term n−1 ´ ³ X X m−1 1[0,si ]×∆j (·)1∆i ×[0,tj ] (?) (42) f 00 Wsα,β δ (2) (N π ) := δ (2) i ,tj i=0 j=0 is convergent in L2 (Ω) integral, since we already have the convergence of ³ to¡ a Skorohod ´ ¢ ⊗2 from Lemma 5, we only need to guarantee that the the integrands in L2 Ω; Hα,β Skorohod integrals in (42) converge to a random variable in L2 (Ω). However this is trivial since the term in (42) is the difference of all the other terms that have been estimated in α,β this proof, and have been proved to converge in L2 (Ω) to f (Ws,t ) minus the sum of all terms on the right hand side of (32) except for the one denoted by Z tZ 0 s 0 ³ ´ α,β f 00 Wu,v dM̃u,v . Therefore, the Skorohod integral in (42) converges to the Skorohod integral δ (2) (N ), and (32) is established, so Theorem 1 is proved. ¤ 4 Interpretation of the integral with respect to M̃ As we mentioned before in the statement of the Itô formula (Theorem 1), the notation M̃ is defined using a double Skorohod integral. However, the rationale for using a specific 19 differential notation for M̃ is the following. We can write Z Z M̃ (s, t) = dWuα,β dWuα,β 1 (u1 ) 1[0,t] (v2 ) 1[0,u1 ] (u2 ) 1[0,v2 ] (v1 ) 1 ,v1 2 ,v2 [0,s] = = = Z u1 ,v1 s u2 ,v2 t Z Z u1 u1 =0 v2 =0 u2 =0 Z s Z t µ Z v2 u1 =0 v2 =0 Z s Z t u1 =0 v2 =0 Z v1 =0 v2 dWuα,β dWuα,β 1 ,v1 2 ,v2 v1 =0 ¶ µZ u1 ¶ α,β α,β dWu1 ,v1 dWu2 ,v2 (43) (44) u2 =0 du1 Wuα,β · dv2 Wuα,β 1 ,v2 1 ,v2 (45) which shows that dM̃ (s, t) should be interpreted as α,β α,β ds Ws,t · dt Ws,t . (46) To make the string of equalities above rigorous, the Fubini-type property used to go from (43) to (44) must be justified, and the double stochastic integral in (45) must be properly defined. This can be done without needing to invoke a new Skorohod integration theory, by simply noting that the integral in (45) can be defined as a random variable in the second Gaussian chaos of W α,β with an appropriate distribution, and checking that it equals M̃ (s, t). We omit these details. Defining stochastic integration with respect to M̃ is also trivial if the integrands are deterministic. However, in connection with the general Itô formula (Theorem 1), we are much more interested in random integrands. A proper theory of general Skorohod integration with respect to M̃ can be given. We will not present this theory here, since it constitutes only a tangent to the Itô formula that is the subject of this article, and especially since it is not clear that it brings any more computational information than formulas (33) and (34) which can be taken as a definition of general Skorohod integration against M̃ . What we will do instead is to present briefly a somewhat formal calculation that shows exactly how formula (33) is related to integration against M̃ . The main point in this calculation is that, although the integrands are non-deterministic, the familiar Malliavinderivative-type corrections that come from pulling random non-time-dependent terms out of Skorohod integrals do not appear in the final formula. We recall the general formula for Skorohod integration with respect to an arbitrary Gaussian field, and a fixed L 2 random variable F : F δ (u) = δ (uF ) + hu; DF iH where h·; ·iH is the inner product in the canonical Hilbert space of the underlying Gaussian field. We also recall that the stochastic differential of W α,β can be formally understood as follows, using a standard Brownian sheet W : Z u Z v ∂K β ∂K α α,β (u, a) (v, b) dWa,b . dWu,v = dudv ∂v a=0 b=0 ∂u 20 Therefore we have I := = Z Z u u0 =0 u u0 =0 Z Z v v 0 =0 α,β f (u, v) dWuα,β 0 ,v dWu,v 0 v f (u, v) v 0 =0 u0 Z v ∂K α ¡ 0 ¢ ∂K β u ,a (v, b) dWa,b ∂v a=0 b=0 ∂u Z u Z v0 ¢ ∂K β ¡ 0 0 ¢ ∂K α ¡ · dudv 0 u, a0 v , b dWa0 ,b0 . ∂v a0 =0 b0 =0 ∂u 0 · du dv Z Now we pull the constant L2 (Ω)-random variable f (u, v) into the integral with respect to dWa,b , modulo a correction term, then use Fubini to pull the Riemann integral with respect to u0 all the way in (which requires no correction), yielding ·Z u Z v Z u ¸ Z v α ¡ ¢ ∂K β 0 ∂K 0 du ·dv u ,a (v, b) f (u, v) dWa,b − correction term I= ∂u ∂v a=0 b=0 u0 =a v 0 =0 Z u Z v0 ¢ ∂K β ¡ 0 0 ¢ ∂K α ¡ 0 u, a0 v , b dWa0 ,b0 · dudv ∂v a0 =0 b0 =0 ∂u ¸ Z v ·Z u Z v ∂K β α α = dudv (v, b) f (u, v) dWa,b − correction term (K (u, a) − K (a, a)) ∂v v 0 =0 a=0 b=0 Z u Z v0 ¢ ∂K β ¡ 0 0 ¢ ∂K α ¡ 0 u, a0 v , b dWa0 ,b0 · dv ∂v a0 =0 b0 =0 ∂u ·Z u Z v ¸ ∂K β α K (u, a) = dudv (v, b) dWa,b ∂v a=0 b=0 Z u Z v0 Z v ¢ ∂K β ¡ 0 0 ¢ ∂K α ¡ dv 0 f (u, v) u, a0 v , b dWa0 ,b0 · ∂v a0 =0 b0 =0 ∂u v 0 =0 Now we perform the same operation on f and the integral with respect to v 0 , yielding ¸ ·Z u Z v ∂K β α (v, b) dWa,b I = dudv K (u, a) ∂v a=0 b=0 # " Z v Z u Z v0 α ¡ β ¡ ¢ ¢ ∂K ∂K u, a0 v 0 , b0 f (u, v) dWa0 ,b0 − correction term · dv 0 ∂u ∂v 0 0 0 v =0 a =0 b =0 ·Z u Z v ¸ ∂K β α K (u, a) = dudv (v, b) dWa,b ∂v ·Z u Z va=0 b=0 ¸ ¢ β ¡ 0¢ ∂K α ¡ 0 · u, a K v, b f (u, v) dWa0 ,b0 − correction term a0 =0 b0 =0 ∂u 21 u v ¸ ∂K β K (u, a) (v, b) dWa,b = dudv ∂v a=0 b=0 ¸ ·Z u Z v ¢ ¡ ¢ ∂K α ¡ u, a0 K β v, b0 dWa0 ,b0 · f (u, v) a0 =0 b0 =0 ∂u ¸ ¸· Z u · Z v ¢ ∂K α ¡ ∂K β 1/2,β α,1/2 0 du u, a da0 Wa0 ,v (v, b) db Wu,b = f (u, v) dv a0 =0 ∂u b=0 ∂v ·Z Z α α,β α,β = f (u, v) dv Wu,v · du Wu,v . α,β α,β Modulo the definition of Skorohod integration with respect to dv Wu,v · du Wu,v , this calculation justifies the formula Z tZ s f (u, v)dM̃u,v 0 0 Z Z := dWuα,β dWuα,β f (u1 , v2 ) 1[0,s] (u1 ) 1[0,t] (v2 ) 1[0,u1 ] (u2 ) 1[0,v2 ] (v1 ) 1 ,v1 2 ,v2 [0,1]2 [0,1]2 s = Z tZ 0 0 α,β α,β f (u, v) dv Wu,v · du Wu,v , which can be summarized by α,β α,β dM̃u,v = dv Wu,v · du Wu,v as announced in (46). 5 ¤ Application to the local time of the fractional Brownian sheet As an application of the Itô formula, we will establish in this section the existence and the stochastic integral representation of the local time of the fractional Brownian sheet. 5.1 Known results and motivation We start with a short summary of the results on the local times for one and multiparameter fractional Brownian motion. motion B H was introduced • The local time λat of the one parameter fractional Brownian Rt in [5] as the density of the occupation measure Γ → 0 1Γ (BsH )ds. The author proved that λat has a jointly continuous version in the variables a and t. Moreover λat has Hölder-continuous paths of order δ < 1 − H in time and of order γ < 1−H 2H in the space variable a provided that H ≥ 31 . As a density of an occupation measure, the local time is increasing in t and the measure La (dt) is concentrated on the level set {s : BsH = a}. A chaos expansion of λat is given in [14], along with an L2 estimate. 22 • Another version of the local time of the fractional BrownianR motion was introduced in [8] t as the density of the occupation measure mt (Γ) = 2H 0 1Γ (BsH )s2H−1 ds. If H ≥ 31 , this local time satisfy the Tanaka formula Z t H |Bt − a| = | − a| + sign(Bsa )dBsH + Lat 0 where the stochastic integral is defined in the Skorohod sense. This formula was recently extended by [7] to H ∈ (0, 1). The regularity properties of λat can be transferred to Lat by integrating by parts. • In [12] the local time of the d-dimensional fractional sheet with R N -parameters was studied. This local time can be formally defined as Lat = [0,t] δa (WsH )ds, where t = (t1 , . . . , tN ) ∈ T N , H = (H1 , . . . , HN ) and δa denotes the Dirac function. The smoothness of the local time in the Sobolev-Watanabe spaces and its asymptotic behavior were studied. • Recently, in [18], using the techniques of the Fourier analysis the authors proved the existence of the local time of the N -parameters d-dimensional fractional Brownian sheet satisfying the occupation density formula Z Z ´ ³ f W Ht d t = f (x)L(x, A)dx d A R for any Borel set A ⊂ RN and for any measurable function f : Rd → R. The existence of a jointly continuous version of the local time is proved. Obviously, this local time coincides with the one used in [12]. ³ In this section we define the local time Las,t , a ∈ R of the fractional Brownian sheet ´ α,β Ws,t as the density of the occupation measure s,t∈[0,1] ms,t (Γ) = αβ Z tZ 0 s 0 α,β 4α−1 4β−1 1Γ (Wu,v )u v dudv. (47) Note that in the case α = β = 21 this local time coincide with the local time introduced in [13] and [16] for the standard Wiener process with two parameters. Remark 1 Using the integration by parts we can transfer the joint continuity of the local time defined in [18] to Las,t . The details are left to the reader. 5.2 Let Stochastic representation of local time 1 2 e−x /2ε pε (x) = √ 2πε 23 be the heat kernel with variance ε > 0. The following results gives an approximation in L 2 of the local time of the fractional Brownian sheet. It can be deduced from [12], where the authors used the Wiener chaos decomposition of the local time. Proposition 2 Let a ∈ R and s, t ∈ [0, 1] and define La,ε s,t = αβ Z tZ 0 s 0 α,β 4α−1 4β−1 pε (Wu,v )u v dudv. a 2 Then the random variable La,ε s,t converges to Ls,t in L (Ω) as ε tends to zero. The next result gives the stochastic integral representation of the local time. Theorem 2 For any a ∈ R and s, t ∈ [0, 1], we have ¯³ ¯³ ´2 1 Z t Z s ¯ ´ 1 ¯¯ α,β ¯ α,β ¯ ¯ α,β α,β α,β − a dWu,v ¯Ws,t − a¯ Ws,t − a − ¯Wu,v − a¯ Wu,v 6 2 0 0 Z tZ s¯ ¯ ¯ α,β ¯ − 2αβ ¯Wu,v − a¯ u2α−1 v 2β−1 dudv 0 0 Z tZ s¯ ¯ ¯ α,β ¯ − ¯Wu,v − a¯ dM̃u,v 0 0 Z tZ s α,β α,β sign(Wu,v − a)u2α−1 v 2β dv Wu,v du −α 0 0 Z tZ s α,β α,β sign(Wu,v − a)u2α v 2β−1 du Wu,v dv. −β Las,t = 0 0 Proof: Let us introduce the functions gεIV 00 gε (x) = Z x 0 (x) = pε (x), 000 gε (y)dy, 000 gε (x) = 2 0 gε (x) = Clearly, as ε tends to zero, we have 000 Z x gε (x) → g (x) = sign(x), 1 0 0 gε (x) → g (x) = x|x|, 2 x 0 pε (y)dy − 1, gε (y)dy, 0 000 00 Z 00 gε (x) = x 0 00 gε (x) → g (x) = |x|, gε (x) → g(x) = 24 Z 1 2 x |x|. 6 0 gε (y)dy. Let’s write Itô’s formula for the function gε (x − a), where a is a real number. We get α,β gε (Ws,t − a) Z tZ s α,β α,β gε0 (Wu,v − a)dWu,v = gε (0) + 0 0 Z tZ s Z tZ s α,β 00 α,β 2α−1 2β−1 gε00 (Wu,v − a)dM̃u,v gε (Wu,v − a)u v dudv + + 2αβ 0 0 0 0 Z tZ s Z tZ s 000 α,β 2α−1 2β α,β α,β α,β +α gε (Wu,v − a)u v dudv Wu,v + β gε000 (Wu,v − a)u2α v 2β−1 dvdu Wu,v 0 0 0 0 Z tZ s α,β + αβ gεIV (Wu,v − a)u4α−1 v 4β−1 dudv. 0 0 We will decompose the proof into several steps. Step 1. Given the existence of moments of all orders for the supremum of continuous Gaussian fields, dominated convergence implies the following convergences hold in L 2 (Ω): 1 α,β α,β (Ws,t − a)2 |Ws,t − a|, 6 Z tZ s¯ ¯ ¯ α,β ¯ 2α−1 2β−1 − a)u v dudv → 2αβ ¯Wu,v − a¯ u2α−1 v 2β−1 dudv, α,β gε (Ws,t − a) → 2αβ Z tZ 0 s 0 α,β gε00 (Wu,v and by Proposition 2, 0 αβ Z tZ 0 s 0 0 α,β 4α−1 4β−1 pε (Wu,v )u v dudv → Las,t . Step 2. We will show that the integrand of the stochastic integral with respect to the fractional Brownian sheet is convergent. More precisely, consider the process α,β gε0 (Wu,v )1[0,t]×[0,s] (u, v) ¡ ¢ and let us show its convergence in L2 Ω; Hα,β as ε → 0 to α,β g 0 (Wu,v )1[0,t]×[0,s] (u, v) where g 0 (x) = 21 x|x|. Equivalently, we will show that K ∗,2 (g 0 (W α,β )1[0,t]×[0,s] )(u, v) converges to K ∗,2 (g 0 (W α,β )1[0,t]×[0,s] )(u, v) in L2 ([0, 1]2 × Ω). We have, by Jensen’s inequality ¯2 Z t Z s ¯Z t Z s ³ ´ ∂K α ¯ ¯ ∂K β α,β α,β 0 0 ¯ dudv ¯ (a, u) (b, v)dadb E ) ) − g (W g (W ε a,b a,b ¯ ¯ ∂a ∂b 0 0 u v ¶ Z t Z s µZ t Z s ¯ ¯2 ¯ 0 α,β α,β ¯ H− 32 0 H− 32 (b − v) dadb dudv ≤ cE ¯gε (Wa,b ) − g (Wa,b )¯ (a − u) u v 0 0 Z tZ s¯ ¯2 ¯ 0 α,β α,β ¯ ≤ cE ¯gε (Wa,b ) − g 0 (Wa,b )¯ dadb 0 0 25 which goes to zero as ε → 0. Step 3. Lemma 5 together with the proof of Theorem 1 shows that the limit in L 2 (Ω) i X h ³ α,β ´ f 00 Wsi ,tj 1[0,si ]×∆j (·)1∆i ×[0,tj ] (?) lim δ (2) |π|→0 i,j exists for every f satisfying Condition (H). Consequently, for every ε > 0, denote by A ε the following random variable in L2 (Ω): i X h ³ α,β ´ (2) 00 Aε = lim δ gε Wsi ,tj 1[0,si ]×∆j (·)1∆i ×[0,tj ] (?) . |π|→0 i,j Since gε00 converges to g(x) = 12 |x|x, we obtain by Lemma 4 (with the observation made 2 at the end of the proof of Proposition 1) theh convergence ³ ´ in L (Ω) of Aε , as iε goes to P zero, to the random variable lim|π|→0 δ (2) i,j g Wsα,β 1[0,si ]×∆j (·)1∆i ×[0,tj ] (?) which is i ,tj in L2 (Ω). Step 4. We consider now the term Z tZ s α,β α,β β gε000 (Wu,v − a)u2α v 2β−1 dvdu Wu,v 0 0 and we will prove its convergence to Z tZ s α,β α,β − a)u2α v 2β−1 dvdu Wu,v . sign(Wu,v β 0 0 An argument from the one-parameter case can be used to conclude this fact. The proof of this step follows from the next lemma and the dominated convergence theorem. Lemma B α a one-dimensional fBm with α ∈ ( 12 , 1). Then, for every s ∈ T , R s 000 α6 Consider R s 2α α 2 α 2α α 0 gε (Bu − a)u dWu converges in L (Ω) to 0 sign(Bu − a)u dWu . Proof: Let us write Itô’s formula with hε (t, x) = t2α gε00 (x) (this is a straightforward application of Theorem 1 of [1]). We get Z s Z s 00 α 2α−1 2α 00 α gε000 (Buα − a)u2α dBuα gε (Bu − a)u du + s gε (Bs − a) = 2α 0 0 Z s + 2α pε (Buα − a)u4α−1 du 0 The following convergences in L2 (Ω) can be proved exactly as in [8]: s2α gε00 (Bsα − a) → s2α |Bsα − a| 26 2α Z s 0 gε00 (Buα − a)u2α−1 du → 2α 2α Lau Z s 0 pε (Buα − a)u 4α−1 Z s 0 du → |Buα − a| u2α−1 du Z s 0 Lau u2α du (here is the local time of the one-parameter fBm defined in [8] and presented at the beginning of this section), and 1[0,s] (u)gε000 (Buα − a)u4α−1 → 1[0,s] (u)sign(Buα − a)u2α , the last convergence being in L2 (Ω; H). That gives the conclusion using Lemma 4. Lemma 4 again and Steps 1 to 4 above finish the proof of the theorem. ¤ ¤ Remark 2 Using standard arguments, additional versions of the stochastic integral representation of the local time of the fractional Brownian sheet can be obtained by using, instead 0 00 000 0 00 of the functions g, g , g and g appearing in the proof of Theorem 2, the functions j, j , j 000 and j , where 1 1 0 j(x) = [(x − a)+ ]3 , j (x) = [(x − a)+ ]2 6 2 and 00 000 j (x) = (x − a)+ , j (x) = 1(a,∞) (x). More precisely, we have Z tZ s ³ ´ 1 a α,β α,β α,β Ls,t = j(Ws,t ) − j 0 Wu,v − a dWu,v 2 0 0 Z tZ s ³ ´ α,β − 2αβ j 00 Wu,v − a u2α−1 v 2β−1 dudv 0 0 Z tZ s ³ ´ α,β − j 00 Wu,v − a dM̃u,v 0 0 Z tZ s ³ ´ α,β α,β −α j 000 Wu,v − a u2α−1 v 2β dv Wu,v du 0 0 Z tZ s ³ ´ α,β α,β −β j 000 Wu,v − a u2α v 2β−1 du Wu,v dv. 0 0 Negative part functions can be also used to express the local time. Our last result is an occupation time formula for the local time of the fractional Brownian sheet, for which the previous remark is useful. Proposition 3 For any Borel function f : R → R and and for any a ∈ R, s, t ∈ [0, 1], it holds that Z tZ s Z α,β 4α−1 4β−1 2αβ f (Wu,v )u v dudv = f (a)Las,t da. (48) 0 0 R 27 Proof: We use the same idea as in [16]. Let us introduce the function hIV x (y) = 1(0,x) (y) and its anti-derivatives. Notice that 000 hx (y) = and 0 hx (y) = Z x 0 Z x 0 1(y>x0 ) dx0 , 00 hx (y) = 1 [(y − x0 )+ ]2 dx0 , 2 Z hx (y) = x 0 Z (y − x0 )+ dx0 , x 0 1 [(y − x0 )+ ]3 dx0 . 6 On the one hand, we can write Itô’s formula for hx (y) (although the fourth derivative is not continuous, is it not difficult to see that Itô’s formula still holds, by approximating hx by smooth functions) and we get Z tZ s α,β 4α−1 4β−1 1(0,x) (Wu,v )u v dudv Z tZ s 0 α,β α,β α,β ) − hx (0) − hx (Wu,v = hx (Ws,t )dWu,v 0 0 Z tZ s Z tZ s 00 00 α,β α,β 2α−1 2β−1 hx (Wu,v hx (Wu,v )u )dM̃u,v v dudv − − 2αβ 0 0 0 0 Z tZ s Z tZ s 000 000 α,β 2α−1 2β α,β α,β 2α 2β−1 α,β −α )u v dudv Wu,v −β hx (Wu,v )u v du Wu,v dv. hx (Wu,v αβ 0 0 0 0 0 0 On the third hand, using the integral expression of the local time Lys,t given in Remark 2, multiplying both sides by 1(0,x) (y), integrating over the real line and changing the order of integration (see exercise 3.2.8 of [17] for the Fubini anticipating theorem), we will obtain that Z Z tZ s 1 y α,β 4α−1 4β−1 1(0,x) (Wu,v )u v dudv. 1 (y)Ls,t dy = αβ 2 R (0,x) 0 0 The formula (48) is proved for f (y) = 1(0,x) (y). A monotone class argument is sufficient to conclude the proof. ¤ 5.3 Relation with the variation of M̃ When α = β = 21 the process M̃ is a martingale and in this case (see [16]) the formula (48) can be written as Z Z Z 1 t s α,β f (Wu,v )dhM̃ iu,v = f (a)Las,t da. (49) 2 0 0 R This can be seen immediately by formally taking the square of the differential dM̃u,v = ³ ´2 du M̃u,v · dv M̃u,v which yields dM̃u,v = uvdudv. If α or β are not 12 , then M̃ is not a semimartingale. To investigate whether a formula such as (49) still holds, the results of Section 4 can be invoked. They show that the quadratic variation of M̃ is obviously not the right quantity to look at, since it is 0. 28 As with one-parameter fBm, there is a specific non-quadratic variation of M̃ that will be non-zero and non-infinite. One can easily prove that, for small h, k > 0, and p, q > 1 ´q i ³ ³ ´p ³ h³ ´´ α,β α,β α,β α,β − Wu,v = v βp hαp uαq k βq + o max hαp , k βq . (50) − Wu,v Wu,v+k E Wu+h,v In order for the increment ³ ´p ³ ´ α,β α,β α,β α,β α,β α,β dM̃u,v = dv Wu,v · du Wu,v = Wu+du,v − Wu,v Wu,v+dv − Wu,v to yield the differential of a non-trivial bounded variation process, one must choose the powers p and q above in order to get the product dudv. Therefore we define the (p, q)variation of M̃ by the formula D E(p) D E(q) D E(p,q) α,β α,β = d W·,v d M̃ d Wu,· (51) u,v u v where for any number r > 1, the r-variation of a one-parameter process X is defined as usual using partitions π = {t1 , · · · , tn } of [0, t] by hXi (r) n X ¯ ¯ ¯Xt − X t ¯r . (t) = lim i i−1 |π|→0 (52) i=1 D E(α−1 ,β −1 ) exists and is non-zero, and in fact Formula (50) proves that M̃ u,v D E(α−1 ,β −1 ) = uα/β v β/α dudv. d M̃ u,v D Despite the somewhat formal nature of the above development, the definition of E(α−1 ,β −1 ) M̃ can be made rigorous because of the tensor-type nature of M̃ , and the last u,v formula can be established rigorously by only referring to the definition of M̃ as a double Skorohod integral in the statement of Proposition 1 or Theorem 1. By comparing with formula (48) in Proposition 3, we can state the following, which answers the issue of the validity D E of formula (49) negatively, leaving a rigorous proof, and a rigorous definition of M̃ , to the reader. Proposition 4 The (p, q)-variation of the fractional Brownian sheet (u, v) 7→ D E(p,q) , M̃ u,v defined in equations (51) and (52), is a non-trivial bounded-variation process on [0, 1] 2 if and only if p = 1/α and q = 1/β. In that case D E(α−1 ,β −1 ) Z s Z t M̃ = dudvuα/β v β/α . s,t 0 0 The formula (49) does not hold unless α = β = 1/2. 29 References [1] E. Alos, O. Mazet, D. Nualart (2001): Stochatic calculus with respect to Gaussian processes. Annals of probability, 29: 766-801. [2] E. Alos, D. Nualart (2001): Stochastic integration with respect to the fractional Brownian motion. Preprint. [3] A. Ayache, S. Leger and M. Pontier (2002): Analysis, 17(1): 31-43. Drap Brownien fractionnaire. Potential [4] X. Bardina, M. Jolis and C.A. Tudor (2002): Weak convergence to the fractional Brownian sheet. Preprint núm. 06/2002, Universitat Autònoma de Barcelona. [5] S. Berman (1973): Local nondeterminism and local times of Gaussian processes. Indiana Univ. Math. J., 23: 69-94. [6] P. Carmona, L.Coutin (1998): Stochastic integration with respect to fractional Brownian motion. Preprint. [7] P. Cheridito, D. Nualart (2002): Stochastic integral of divergence type with respect to fBm with H ∈ (0, 12 ). Preprint. [8] L. Coutin, D. Nualart and C.A. Tudor (2001): The Tanaka formula for the fractional Brownian motion. Stoc. Proc. Appl., 94(2):301-315. [9] L. Decreusefond, A. Ustunel (1998): Stochastic analysis of the fractional Brownian motion. Potential Analysis, 10:177-214. [10] M. Dozzi (1989): Stochastic processes with a multidimensional parameter. Longman Scientific and Technical. [11] T. E. Duncan, Y. Hu and B. Pasik-Duncan (2000): Stochastic calculus for fractional Brownian motion I. Theory. Siam J. Control Optim., 38(2):582-612. [12] M. Eddahbi, R. Lacayo, J.L. Sole, C.A. Tudor, J. Vives (2002): Regularity and asymptotic behaviour of the local time for the d-dimensional fractional Brownian motion with N -parameters. Preprint. [13] P. Imkeller (1984): Stochastic analysis and local time for (N.d) -Wiener process. Ann. Inst. Henri Poincaré, 20(1): 75-101. [14] Y. Hu, B. Oksendal (2002): Chaos expansion of local time of fractional Brownian motions. Stoch. Analy. Appl., 20 (4): 815-837. [15] B.B. Mandelbrot, J.W. Van Ness: Fractional Brownian motion, fractional noises and application SIAM Review, 10(4):422-437 30 [16] D. Nualart (1981): Une formule d’Itô pour les martingales continues à deux indices et quelques applications. Ann. Inst. Henri Poincaré, 20(3):251-275. [17] D. Nualart (1995): Malliavin Calculus and Related Topics. Springer. [18] Y. Xiao, T. Zhang (2002) : Local times of fractional Brownian sheets. Probab. Theory Relat. Fields,to appear. 31