Density analysis of non-Markovian BSDEs and applications to biology and finance
DDensity analysis of non-Markovian BSDEs and applicationsto biology and finance.
Thibaut Mastrolia ∗ October 8, 2018
Abstract
In this paper, we provide conditions which ensure that stochastic Lipschitz BSDEs admitMalliavin differentiable solutions. We investigate the problem of existence of densities forthe first components of solutions to general path-dependent stochastic Lipschitz BSDEsand obtain results for the second components in particular cases. We apply these resultsto both the study of a gene expression model in biology and to the classical pricingproblems in mathematical finance.
Key words: BSDEs, Malliavin calculus, Nourdin-Viens’ Formula, gene expression, optionpricing.AMS 2010 subject classification:
Primary: 60H10; Secondary: 60H07, 91G30, 92D20.
The problem of existence of densities for random processes, as e.g. solutions of stochasticdifferential equations (SDEs), has been a very active strand of research in the last twodecades, see among others [26, 35]. A very useful criterion to prove that the law of arandom variable admits a density is the criterion of Bouleau and Hirsch, see e.g. [35,Theorem 2.1.2]. The analysis of densities has been the subject of several works dealingwith Stochastic Partial Differential Equations (SPDEs), among which we can mentionthe study of the stochastic heat equation, the stochastic wave equation (see for instance[33], [36], [31]), the Navier-Stokes equation [11] and recently the Landau equation forMaxwellian molecules (see [12]). Besides, most of these papers investigate tails estimatesof the solutions to SPDEs by using the formula of Nourdin and Viens, introduced in [34],to have a better understanding of these processes.Although the problem of existence of densities for S(P)DEs, together with estimates ontheir tails, has been a prosperous field, the corresponding theory for Backward StochasticDifferential Equations (BSDEs) has not received the same attention in the literature.BSDEs were introduced for the first time in 1973 by Bismut in [5], in order to studystochastic control problems and their links to the Pontryagin maximum principle. Thetheory of BSDEs was then formalised and developed in the 90’s, with the seminal papers[38, 39] and [17]. In the last decades, BSDEs have been the object of an ever growing ∗ Université Paris-Dauphine, CEREMADE UMR CNRS 7534, Place du Maréchal De Lattre De Tassigny,75775 Paris Cedex 16, FRANCE, [email protected] a r X i v : . [ m a t h . P R ] F e b nterest, since these equations naturally appear in financial problems, as for instancepricing problems (see [17]) and utility maximisation problems (see [45], [22]).As far as we know, the existence of densities for solutions to BSDEs was studied in threepapers. Conditions ensuring that the first component Y of the solution to a LipschitzBSDE admits a density were provided for the first time in [3]. In this paper, the authorsalso investigated both estimates on the existing density and its smoothness. Then, aresult ensuring existence of a density for the second component Z of the solution to aparticular BSDE, in which the generator is linear with respect to its z variable, wasobtained in [1]. Recently, this problem was studied in [29] for both the Y and the Z components of solutions to BSDEs with a quadratic growth generator. However, [3, 29]only consider Markovian BSDEs, that is the case where the data ξ and ω (cid:55)−→ f ( s, ω, y, z ) of such equations are only random through a Markovian process, and [1] only considersthe semi-Markovian case, that is the case where only ω (cid:55)−→ f ( s, ω, y, z ) is Markovian.Although the previous studies are interesting from a mathematical point of view, theseresults seem to be too restrictive for applications. As an example, consider a pricingproblem which could be reduced to solve the following BSDE (see [17] for more details) dY t = ( r t Y t + θ t Z t ) dt + Z t dW t , Y T = ξ, where r denotes the interest rate of the market, θ is the market price of risk and ξ is theliability. As noticed in [16], assuming that r is bounded, for instance, is not realistic.This remark led the authors of [16] to define a new class of BSDEs satisfying a so-calledstochastic Lipschitz condition for their generator. Existence and uniqueness results havebeen obtained for this class of BSDEs first in [16], and have then been extended in[4, 48, 8] among others.The problem of existence of densities for the laws of components of solutions to stochasticLipschitz BSDEs has not been studied yet, and a fortiori in the non-Markovian frame-work, i.e. when neither the terminal condition ξ nor ω (cid:55)−→ f ( s, ω, y, z ) depend on therandomness through a Markovian process. We give in the present paper conditions on ξ and f to solve these problems. Besides, although it is well-known that under suitableconditions on the data, a non-Markovian stochastic Lipschitz BSDEs admits a uniquesolution (see [16, 48, 4, 8]), the Malliavin differentiability of the solutions to such BSDEshas not been studied yet in the general case. In order to apply Bouleau and Hirsch’s Cri-terion ([35, Theorem 2.1.2]) to solve the problem of existence of densities for the law of Y and Z , we provide also in this paper conditions which ensure that the components Y and Z solutions to non-Markovian stochastic Lipschitz BSDEs are Malliavin differentiable.The structure of this paper is the following. After some preliminaries and notations inSection 2, we provide in Section 3 two approaches to study the Malliavin differentiabilityof solutions to stochastic Lipschitz BSDEs. Indeed, in view of the classical literature, wedistinguish two types of assumptions which provide existence and uniqueness of solutionsto stochastic Lipschitz BSDE. On the one hand, we have assumptions as in [16, 4, 48]dealing with β -spaces (see S p,β and H p,β below), on the other hand, we have assump-tions dealing (mainly) with the BMO-norm of the data, as in [8]. We then reach in thispaper two kind of conditions which ensure that the components of the solution ( Y, Z ) to a stochastic Lipschitz BSDE are Malliavin differentiable. The first one, investigatedin Section 3.1, is based on the papers [16, 4, 48]. Using a priori estimates for solutionsto stochastic Lipschitz BSDE, obtained in [48, Proposition 3.6], we have conditions onthe data of such BSDE which provide the Malliavin differentiability of Y and Z (seeAssumption ( DsL p ,β ) ). The second approach, studied in Section 3.2, is based on thepapers [8, 2]. We give assumptions, similar to those obtained in [30], see Assumptions( sH , ∞ ) and ( sH , ∞ ) below, which ensure that Y and Z are Malliavin differentiable.We then compare these two approaches, and the corresponding conditions, in Section3.3.By taking advantage of the results obtained in Section 3, we deal in Section 4 with theproblem of existence of densities for the laws of solutions to stochastic Lipschitz BSDE in he non-Markovian case. We give in Section 4.1 conditions which ensure the existence ofdensities for the law of the Y component of the solution to stochastic Lipschitz BSDEs,by using Bouleau and Hirsch’s Criterion. We provide weaker conditions in Section 4.2for the Y component of the solution to a non-Markovian Lipschitz BSDE. We then turnto the Z component in Section 5. We first provide in Section 5.1 conditions ensuringthat the law of the Z t component has a density for a particular class of BSDE, extendingthe results of [1]. We then explain in Section 5.2 why we are not able to adapt the proofsof [29] to the non-Markovian framework for the Z components of solutions to generalnon-Markovian BSDEs and we indicate paths for future researches. We finally apply ourstudy in Sections 6 and 7 to biology and finance respectively.In Section 6, we propose to study mathematically a model of synthesis of proteins intro-duced in [46], with the Malliavin calculus. Indeed, in order to validate their model, theauthors of [46] need to compare the law of the protein concentration at time t obtainedby solving a BSDE with the data produced by Gillespie Method (see [20]). However,in [46], the authors assumed implicitly that the law of the first component Y t of theBSDE under consideration admits a density with respect to the Lebesgue measure. Thepresent paper can be seen as a mathematical strengthening of the model developed in[46] by using the so-called Nourdin and Viens’ formula to obtain Gaussian estimates ofthe density. Besides, we propose to extend their model to the non-Markovian setting,which could be quite relevant when we study the synthesis of protein in some models(see for instance [7, 27, 18]).In Section 7, we study classical pricing problems. As showed in [17], this problem can bereduced to solve a stochastic linear BSDE. In this section we aim at applying the resultsobtained in previous sections to Asian and Lookback options in the Vašìček Model toobtain information on both the regularity of the value function and the regularity ofoptimal strategies. We denote by λ the Lebesgue measure on R . Let T > be a time fixed horizon.Let Ω := C ([0 , T ] , R ) be the canonical Wiener space of continuous function ω from [0 , T ] to R such that ω (0) = 0 . We denote by W := ( W t ) t ∈ [0 ,T ] the canonical Wienerprocess, that is, for any time t in [0 , T ] , W t ( ω ) := ω t for any element ω in Ω . Weset F o the natural filtration of W . Under the Wiener measure P , the process W is astandard Brownian motion and we denote by F := ( F t ) t ∈ [0 ,T ] the usual right-continuousand complete augmentation of F o under P . For the sake of simplicity, we denote allexpectations under P by E and we set for any t ∈ [0 , T ] E t [ · ] := E [ ·|F t ] . Besides, allnotions of measurability for elements of Ω will be with respect to the filtration F or the σ -field F T .We set h := L ([0 , T ] , R ) , where B ([0 , T ]) is the Borel σ -algebra on [0 , T ] , and considerthe following inner product on h (cid:104) f, g (cid:105) := (cid:90) T f ( t ) g ( t ) dt, ∀ ( f, g ) ∈ h , with associated norm (cid:107)·(cid:107) h . Let now H be the Cameron-Martin space that is the spaceof functions in Ω which are absolutely continuous with square-integrable derivative andwhich start from at : H := (cid:26) h : [0 , T ] −→ R , ∃ ˙ h ∈ h , h ( t ) = (cid:90) t ˙ h ( x ) dx, ∀ t ∈ [0 , T ] (cid:27) . For any h in H , we will always denote by ˙ h a version of its Radon-Nykodym densitywith respect to the Lebesgue measure. Notice that H is an Hilbert space equipped with he inner product (cid:104) h , h (cid:105) H := (cid:104) ˙ h , ˙ h (cid:105) h , for any ( h , h ) ∈ H × H , and with associatednorm (cid:107) h (cid:107) H := (cid:104) ˙ h, ˙ h (cid:105) h . Let p ≥ . Define L p ( K ) as the set of all F T -measurable randomvariables F which are valued in an Hilbert space K , and such that (cid:107) F (cid:107) L p ( K ) < + ∞ ,where (cid:107) F (cid:107) L p ( K ) := E [ (cid:107) F (cid:107) p K ] /p , where the norm (cid:107) · (cid:107) K is the one canonically induced by the inner product on K . Wedefine L p ([ t, T ]; K ) := (cid:40) f : [ t, T ] −→ K , Borel-measurable, s.t. (cid:90) Tt (cid:107) f ( s ) (cid:107) p K ds < + ∞ (cid:41) . Set
BMO( P ) as the space of square integrable, continuous, R -valued martingales M suchthat (cid:107) M (cid:107) BMO := esssup τ ∈T T (cid:13)(cid:13)(cid:13) E τ (cid:104) ( M T − M τ ) (cid:105)(cid:13)(cid:13)(cid:13) ∞ < + ∞ , where for any t ∈ [0 , T ] , T Tt is the set of F -stopping times taking their values in [ t, T ] .Accordingly, H is the space of R -valued and F -predictable processes Z such that (cid:107) Z (cid:107) H := (cid:13)(cid:13)(cid:13)(cid:13)(cid:90) . Z s dW s (cid:13)(cid:13)(cid:13)(cid:13) BMO < + ∞ . Denoting by E ( M ) the stochastic exponential of a semi-martingale M , we have finallythe following result Theorem 2.1. [25, Theorem 2.3] If M ∈ BMO( P ) then E ( M ) is a uniformly integrablemartingale. For any nonnegative F -adapted process α , we define the following increasing and contin-uous process A αt := (cid:90) t α s ds. Let p > , β > and let α be a nonnegative F -adapted process, we define S p,β,α := (cid:26) Y adapted and càdlàg, (cid:107) Y (cid:107) p S p,β,α := E (cid:20) sup ≤ t ≤ T e pβA αt | Y t | p (cid:21) < + ∞ (cid:27) . H p,β,α := (cid:40) Y progressively measurable, (cid:107) Y (cid:107) p H p,β,α := E (cid:34)(cid:32)(cid:90) T e βA αs | Y s | ds (cid:33) p (cid:35) < + ∞ (cid:41) . H a p,β,α := (cid:40) Y progressively measurable, (cid:107) Y (cid:107) p H α p,β,α := E (cid:34)(cid:90) T α s e βA αs | Y s | p ds (cid:35) < + ∞ (cid:41) . To match with the notations in [8], we define for any real p > the spaces S p and H p by S p := (cid:40) Y adapted and càdlàg processes, (cid:107) Y (cid:107) S p := E (cid:20) sup ≤ t ≤ T | Y t | p (cid:21) ∧ /p < + ∞ (cid:41) H p := Z predictable processes, (cid:107) Z (cid:107) H p := E (cid:32)(cid:90) T | Z s | ds (cid:33) p/ ∧ /p < + ∞ . In particular, for any p > we have S p = S p, ,α and H p = H p, ,α . Notice moreoverthat the following inequality holds for any p > , β > and for any nonnegative F -adapted process α (cid:107) Y (cid:107) p S p + (cid:107) Z (cid:107) p H p ≤ (cid:107) Y (cid:107) p S p,β,α + (cid:107) Z (cid:107) p H p,β,α . (2.1) .2 Elements of Malliavin calculus We give in this section some results on the Malliavin calculus that we will use in thispaper. Let now S be the set of cylindrical functionals, that is the set of random variables F of the form F = f ( W ( h ) , . . . , W ( h n )) , ( h , . . . , h n ) ∈ H n , f ∈ C ∞ b ( R n ) , for some n ≥ , (2.2)where W ( h ) := (cid:82) T ˙ h s dW s for any h in H and where C ∞ b ( R n ) denotes the space ofbounded mappings which are infinitely continuously differentiable with bounded deriva-tives. For any F in S of the form (2.2), the Malliavin derivative ∇ F of F is defined asthe following H -valued random variable: ∇ F := n (cid:88) i =1 f x i ( W ( h ) , . . . , W ( h n )) h i , (2.3)where f x i := dfdx i . It is then customary to identify ∇ F with the stochastic process ( ∇ t F ) t ∈ [0 ,T ] . Denote then by D ,p the closure of S with respect to the Malliavin-Sobolevsemi-norm (cid:107) · (cid:107) ,p , defined as: (cid:107) F (cid:107) ,p := ( E [ | F | p ] + E [ (cid:107)∇ F (cid:107) pH ]) /p . We set D , ∞ := (cid:84) p ≥ D ,p . We make use of the notation DF to represent the derivativeof ∇ F as: ∇ t F = (cid:90) t D s F ds, t ∈ [0 , T ] . To avoid any ambiguity in the non-Markovian case we will consider later on, we need tointroduce immediately some further notations. For any mapping ˜ f from [0 , T ] × Ω × R into R , we let D ˜ f ( t, y ) be the Malliavin derivative, computed at the point ( t, y ) , of ω (cid:55)−→ ˜ f ( t, ω, y ) . If D ˜ f is continuously differentiable with respect to y , we denote by ( D ˜ f ) y its derivative with respect to y . Let now Y be an F -progressively measurable realprocess, with Y t ∈ D , at time t ∈ [0 , T ] . Using the chain rule formula (see for instance[35]), the Malliavin derivative of D ˜ f at ( t, Y t ) , denoted by D ˜ f ( t, Y t ) is given by D v,u ˜ f ( t, Y t ) := D v ( D u ˜ f )( t, Y t ) + ( D u ˜ f ) y ( t, Y t ) D v Y t , ≤ u, v ≤ t. (2.4)Let h be in H and let τ be the following shift operator τ h : Ω −→ Ω defined by τ h ( ω ) := ω + h, ω ∈ Ω . Note that the fact that h belongs to H ensures that τ h is a measurable shift on theWiener space. In the present paper, we will use the characterization of the Malliavin dif-ferentiability, as a convergence of a difference quotient in L p , introduced in [30], recalledbelow. Theorem 2.2 (Theorem 4.1 in [30]) . Let p > and F ∈ L p ( R ) . Then F belongs to D ,p if and only if there exists D F in L p ( H ) and there exists q ∈ [1 , p ) such that for any h in H lim ε → E (cid:20)(cid:12)(cid:12)(cid:12)(cid:12) F ◦ τ εh − Fε − (cid:104)D F, h (cid:105) H (cid:12)(cid:12)(cid:12)(cid:12) q (cid:21) = 0 . In that case, D F = ∇ F . We now recall the criterion that we will use to check the absolute continuity of the lawof a random variable F with respect to the Lebesgue measure. Theorem 2.3 (Bouleau-Hirsch Criterion, see e.g. Theorem 2.1.2 in [35]) . Let F bein D ,p for some p > . Assume that (cid:107) DF (cid:107) h > , P − a.s. Then F has a probabilitydistribution which is absolutely continuous with respect to the Lebesgue measure on R . et F such that (cid:107) DF (cid:107) h > , P − a.s., then the previous criterion implies that the law of F admits a density ρ F with respect to the Lebesgue measure. Assume that there existsin addition a measurable mapping Φ F with Φ F : R h → h , such that DF = Φ F ( W ) , wethen set: g F ( x ) := (cid:90) ∞ e − u E (cid:104) E ∗ [ (cid:104) Φ F ( W ) , (cid:102) Φ uF ( W ) (cid:105) h ] (cid:12)(cid:12)(cid:12) F − E ( F ) = x (cid:105) du, x ∈ R , (2.5)where (cid:102) Φ uF ( W ) := Φ F ( e − u W + √ − e − u W ∗ ) with W ∗ an independent copy of W definedon a probability space (Ω ∗ , F ∗ , P ∗ ) , and E ∗ denotes the expectation under P ∗ ( Φ F beingextended on Ω × Ω ∗ ). We recall the following result from [34]. Theorem 2.4 (Nourdin-Viens’ Formula) . The law of a random variable F has a density ρ F with the respect to the Lebesgue measure if and only if the random variable g F ( F − E [ F ]) is positive a.s. In this case, the support of ρ F , denoted by supp ( ρ F ) , is a closedinterval of R and for all x ∈ supp ( ρ F ) : ρ F ( x ) = E ( | F − E [ F ] | )2 g F ( x − E [ F ]) exp (cid:32) − (cid:90) x − E [ F ]0 udug F ( u ) (cid:33) . The Malliavin differentiability of solutions to non-Markovian Lipschitz BSDE has beenstudied first in [17] and more recently in [30], as well as in [19] for Lévy driven BSDE.In [19], the authors use the well-known characterization of the Malliavin derivative asGâteaux derivative introduced by Sugita in [47] and they obtain similar conditions, for theBrownian part, to those in [17] (see [19, Section 4, ( A f ) ]), while [30] took the advantage ofa new L p characterization of the Malliavin differentiability (see Theorem 2.2) to improveconditions obtained in [17].Here, we extend the results of [30] to the stochastic Lipschitz case. We consider thefollowing non-Markovian BSDE Y t = ξ + (cid:90) Tt f ( s, Y s , Z s ) ds − (cid:90) Tt Z s dW s , ∀ t ∈ [0 , T ] , P − a.s. (3.1)where ξ is an F T -measurable random variable and f : [0 , T ] × Ω × R −→ R is an F -progressively measurable process where as usual the ω -dependence is omitted. (3.1) : an approach inspired by [16, 48] We consider the following assumption for p > and β > , Assumption (sL p,β ). (i) There exists two nonnegative F -adapted processes r and θ such that | f ( t, y, z ) − f ( t, y (cid:48) , z (cid:48) ) | ≤ r t | y − y (cid:48) | + θ t | z − z (cid:48) | , ∀ ( t, y, y (cid:48) , z, z (cid:48) ) ∈ [0 , T ] × R . (ii) Let a t := r t + | θ t | for any t ∈ [0 , T ] . We suppose that a t > , dt ⊗ d P -a.e., E [ A aT ] < + ∞ and f ( t, , a t ∈ H p,β,a . (iii) ξ satisfies E (cid:104) e pβA aT | ξ | p (cid:105) < + ∞ . iv) If p ∈ ( , , there exists a positive constant L such that A aT < L, P -a.s. Remark 3.1.
Notice that the case a ≡ is excluded according to ( ii ) . However, thiscase can be studied easily since a ≡ implies that f is constant with respect to y and z .Then, we can provide an explicit expression for the solution to this kind of BSDE. The main difficulty in this study is that the process a is not bounded and the stochasticintegral of a is not a BMO-martingale under Assumption (sL p,β ) . We recall the followingresult which can be found in [48] and extends the results in [16]. Theorem 3.1 (Theorem 4.1 together with Proposition 3.6 in [48]) . Let p > and β > max { / (2 p − } and assume that Assumption ( sL p ,β ) holds. Then BSDE (3.1) admits a unique solution ( Y, Z ) in ( S p,β ∩ H a p,β ) × H p,β . Moreover, ( i ) if p ≥ , there exists a positive constant C p,β depending only on p and β such that (cid:107) Y (cid:107) p S p,β,a + (cid:107) Y (cid:107) p H a p,β,a + (cid:107) Z (cid:107) p H p,β,a ≤ C p,β (cid:32) E (cid:104) e pβA aT | ξ | p (cid:105) + (cid:13)(cid:13)(cid:13)(cid:13) f ( t, , a t (cid:13)(cid:13)(cid:13)(cid:13) p H p,β,a (cid:33) , (3.2) ( ii ) if p ∈ ( , , there exists a positive constant C p,β,L depending only on p, β, L suchthat Estimate (3.2) holds with C p,β,L . We now turn to the Malliavin differentiability of solutions to BSDE (3.1) under Assump-tion ( sL p,β ) . Such a result requires additional assumptions that we now list. Assumption ( DsL p ,β ) . There exist p > and β > such that for any h ∈ H , ( i ) ξ ∈ D , , lim ε → E (cid:34) e pβA aT (cid:12)(cid:12)(cid:12)(cid:12) ξ ◦ τ εh − ξε − (cid:104)∇ ξ, h (cid:105) H (cid:12)(cid:12)(cid:12)(cid:12) p (cid:35) = 0 , and E (cid:104) e βA aT |(cid:104)∇ ξ, h (cid:105) H | (cid:105) < + ∞ . ( ii ) ω (cid:55)−→ f ( t, ω, y, z ) ∈ D , for any ( t, y, z ) ∈ [0 , T ] × R × R , lim ε → (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) f ( t, ω ◦ τ εh , Y t , Z t ) − f ( t, ω, Y t , Z t ) ε − (cid:104)∇ f ( t, Y t , Z t ) , h (cid:105) H a t (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) H p,β,a = 0 and (cid:13)(cid:13)(cid:13)(cid:13) (cid:104)∇ f ( t, Y t , Z t ) , h (cid:105) H a t (cid:13)(cid:13)(cid:13)(cid:13) H ,β,a < + ∞ . ( iii ) Let ( ε n ) n ∈ N be a sequence in (0 , such that lim n → + ∞ ε n = 0 , and let ( Y n , Z n ) n bea sequence of random variables which converges in S p,β,a × H p,β,a , for any ( p, β ) ∈ ( , × (0 , + ∞ ) , to some ( Y, Z ) . Then there exists η > such that for all h ∈ H , thefollowing convergences hold in probability (cid:107) f y ( · , ω + ε n h, Y n · , Z · ) − f y ( · , ω, Y · , Z · ) (cid:107) L η ([0 ,T ]) −→ n → + ∞ , ess sup t ∈ [0 ,T ] (cid:12)(cid:12)(cid:12)(cid:12) f z ( · , ω + ε n h, Y n · , Z n · ) − f z ( · , ω, Y · , Z · ) a t (cid:12)(cid:12)(cid:12)(cid:12) −→ n → + ∞ (3.3)or (cid:107) f y ( · , ω + ε n h, Y n · , Z n · ) − f y ( · , ω, Y · , Z · ) (cid:107) L η ([0 ,T ]) −→ n → + ∞ , ess sup t ∈ [0 ,T ] (cid:12)(cid:12)(cid:12)(cid:12) f z ( · , ω + ε n h, Y · , Z n · ) − f z ( · , ω, Y · , Z · ) a t (cid:12)(cid:12)(cid:12)(cid:12) −→ n → + ∞ . (3.4) ( iv ) For any q ≥ , E (cid:104)(cid:16)(cid:82) T r s ds (cid:17) q (cid:105) < + ∞ . emark 3.2. Concerning Property ( ii ) of Assumption ( DsL p ,β ) , notice that for fixed ( y, z ) , the process ( s, ω ) (cid:55)−→ Df ( s, ω, y, z ) is defined outside a P -negligible set whichdepends generally on ( y, z ) . Hence, it is not clear that this process is well-defined atthe point ( Y s ( ω ) , Z s ( ω )) . However, under appropriate continuity conditions on the map ( y, z ) (cid:55)−→ Df ( s, · , y, z ) , these negligible sets can actually be aggregated into a universalone, outside of which Df ( s, Y s , Z s ) is indeed well-defined.Nonetheless, let us point out an alternative approach for which no extra conditions on theMalliavin derivative of f is required. The main problem is that the Malliavin derivativeof a random variable is in general only defined P -a.s. (except for instance when it is acylindrical random variable), as a limit in probability of a sequence of random variables(which are defined for every ω , again since they are cylindrical functions). There ex-ists however a notion of limit, called medial limits ( lim med for short ) , which has theparticular property that under very general set theoretic axioms (see below), we have thefollowing result (see e.g. [32]):Let ( Z n ) be a sequence of random variables, then Z ( ω ) := lim med n → + ∞ Z n ( ω ) is universallymeasurable and if Z n converges to some random variable Z P in probability, then Z = Z P , P -a.s.In our case, let F be in D , , there exists a sequence of cylindrical elements F n whichconverges to F in D , . Hence, DF n converges in L ( H ) to the Malliavin derivative of F denoted by DF , defined P -a.s. Let (cid:103) DF be the medial limit of DF n , defined for every ω . By the above result, DF = (cid:103) DF , P − a.s. This approach, which as far as we know has not been considered in the context of Malli-avin calculus before (but see [37] for its use for stochastic integrals), allows to give acomplete pathwise definition of the Malliavin derivative of any random variable in D , .We emphasize nonetheless that the existence of medial limits depends on set-theoreticframework that one is using for instance Zermelo-Fraenkel set theory, plus the axiom ofchoice (ZFC for short), and either the continuum hypothesis or Martin’s axiom (whichis compatible with the negation of the continuum hypothesis). See e.g. the footnote in[42, Remark 4.1] for more explanations and the weakest known conditions ensuring theexistence of medial limits. Before going further, we compare these assumptions with those made in [30]. Assump-tions ( DsL p ,β ) ( i ) and ( ii ) seem quite reasonable in order to prove that the Malliavinderivatives of Y t and Z t are well-defined as the solution in S ,β,a × H ,β,a to the stochasticlinear BSDE (3.5) below, in view of Theorem 3.1. We now turn to Assumption ( DsL p ,β )( iii ) which is less natural and stronger than its equivalent ( H ) in [30]. Indeed, if wecompare for instance (3.3) with its equivalent ( H ) in [30], we first notice that we assumethat there exists η > such that (cid:107) f y ( · , ω + ε n h, Y n · , Z · ) − f y ( · , ω, Y · , Z · ) (cid:107) L η ([ , T ]) −→ n → + ∞ , which provide a condition of order strictly more than 2, unlike Assumption ( H ) in [30]which deals with an L norm. This assumption is necessary for our study and comesin fact directly from the a priori estimates in Theorem 3.1 (see [48, Proposition 3.6])and the definition of H p,β,a . We now turn to the second assumption in (3.3). Thisassumption is quite strong, and is intrinsically linked to the fact Z ∈ H p,β,a . Indeed,to obtain (3.10) in the proof of the Theorem 3.2 below, we are not able to concludewithout this assumption since an Hölder Inequality will provide a term with Z ηs inthe integral and in view of the definition of the space H p,β,a , we can not prove theconvergence. Concerning ( iv ) , this assumption is quite similar to those obtained in thefollowing Section 3.2, and is satisfied as soon as the stochastic integral of r is for instancea BMO-martingale. This gap was pointed by Laurent Denis, during a review of the PhD thesis of the author, concerningAssumption ( D ) in [30] which corresponds to ( DsL p ,β ) . Remark 3.2 has to be also taken into account for thelatter paper. e thus have the following theorem. Theorem 3.2.
Let p be in ∈ (cid:0) , (cid:1) , β > max { / (2 p − } and assume that As-sumptions ( sL ,β ) and ( DsL p ,β ) hold. Then, for any t ∈ [0 , T ] , Y t ∈ D , and Z ∈ L ([ t, T ]; D , ) . Besides, a version of ( D u Y t , D u Z t ) ≤ u ≤ t, ≤ t ≤ T , is given as the solution to the affine BSDE: D u Y t = D u ξ + (cid:90) Tt ( D u f ( s, Y s , Z s ) + f y ( s, Y s , Z s ) D u Y s + f z ( s, Y s , Z s ) D u Z s ) ds − (cid:90) Tt D u Z s dW s . (3.5) Proof.
We only consider the case where (3.3) holds under Assumption ( DsL p ,β ) ( iii ) ,since the other situation can be treated similarly. We aim at applying Theorem 2.2 with F = Y t and F = (cid:82) Tt Z s dW s . The proof is similar to the proof of Theorem 5.1 in [30] andwe recall here the main ideas. Let ε > , h ∈ H and p ∈ (cid:0) , (cid:1) . We have Y s ◦ τ εh = ξ ◦ τ εh + (cid:90) Ts f ( r, Y r , Z r ) ◦ τ εh dr − (cid:90) Ts Z r ◦ τ εh dW r , ∀ s ∈ [ t, T ] , P − a.s. As a consequence, setting for simplicity Y εs := 1 ε ( Y s ◦ τ εh − Y s ) , Z εs := 1 ε ( Z s ◦ τ εh − Z s ) , ξ ε := 1 ε ( ξ ◦ τ εh − ξ ) , s ∈ [ t, T ] , we have that ( Y ε , Z ε ) solves the BSDE: Y εs = ξ ε + (cid:90) Ts (cid:16) ˜ A εr + ˜ A y,εr Y εr + ˜ A z,εr Z εr (cid:17) dr − (cid:90) Ts Z εr dW r , (3.6)with ˜ A y,εr := (cid:90) f y ( r, · + εh, Y r + θ ( Y r ◦ τ εh − Y r ) , Z r ) dθ, ˜ A z,εr := (cid:90) f z ( r, · + εh, Y r ◦ τ εh , Z r + θ ( Z r ◦ τ εh − Z r )) dθ, ˜ A εr := 1 ε ( f ( r, · + εh, Y r , Z r ) − f ( r, · , Y r , Z r )) . Let us now consider the following stochastic affine BSDE on [ t, T ] , which admits a uniquesolution ( ˜ Y h , ˜ Z h ) ∈ ( S ,β,a ∩ H a ,β,a ) × H ,β,a according to Theorem 3.1 under Assumption ( DsL p ,β )˜ Y hs = (cid:104) Dξ, ˙ h (cid:105) L ([0 ,T ]) − (cid:90) Ts ˜ Z hr dW r + (cid:90) Ts (cid:16) (cid:104) Df ( r, Y r , Z r ) , ˙ h (cid:105) L ([0 ,T ]) + ˜ Y hr f y ( r, Y r , Z r ) + ˜ Z hr f z ( r, Y r , Z r ) (cid:17) dr. (3.7)Hence, using Theorem 3.1 together with Inequality (2.1), we obtain (cid:107) Y ε − ˜ Y h (cid:107) p S p + (cid:107) Z ε − ˜ Z h (cid:107) p H p ≤ (cid:107) Y ε − ˜ Y h (cid:107) p S p,β,a + (cid:107) Z ε − ˜ Z h (cid:107) p H p,β,a ≤ C ,β (cid:0) Ξ p,a,βε + X εT + X y,εT + X z,εT (cid:1) here Ξ p,a,βε := E (cid:104) e pβA aT | ξ ε − (cid:104)∇ ξ, h (cid:105) H | p (cid:105) , X εT := (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ˜ A εt − (cid:104)∇ f ( t, Y t , Z t ) , h (cid:105) H a t (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) p H p,β,a ,X y,εT := (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ˜ Y ht ˜ A y,εt − f y ( t, Y t , Z t ) a t (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) p H p,β,a , X z,εT := (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ˜ Z ht ˜ A z,εt − f z ( t, Y t , Z t ) a t (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) p H p,β,a . First notice that under Assumption ( DsL p ,β ) ( i ) and ( ii ) , we have lim ε → (cid:0) Ξ p,a,βε + X εT (cid:1) = 0 . (3.8)We now turn to X y,εT . We have X y,εT = E (cid:34)(cid:32)(cid:90) T e βA at | ˜ Y ht | (cid:12)(cid:12)(cid:12)(cid:12) A y,εt − f y ( t, Y t , Z t ) a t (cid:12)(cid:12)(cid:12)(cid:12) dt (cid:33) p (cid:35) . According to Assumption ( DsL p ,β ) ( iii ) , there exists η > such that (cid:13)(cid:13)(cid:13)(cid:13) A y,εt − f y ( t, Y t , Z t ) a t (cid:13)(cid:13)(cid:13)(cid:13) ηL η ([0 ,T ]) = (cid:90) T (cid:12)(cid:12)(cid:12)(cid:12) A y,εt − f y ( t, Y t , Z t ) a t (cid:12)(cid:12)(cid:12)(cid:12) η dt proba −→ ε → . Hence, using Hölder Inequality with q > such that q = 2 + η and denoting by q itsconjugate and using the fact that ˜ Y h ∈ S ,β,a , we have for some constant C > (cid:90) T e βA at | ˜ Y ht | (cid:12)(cid:12)(cid:12)(cid:12) A y,εt − f y ( t, Y t , Z t ) a t (cid:12)(cid:12)(cid:12)(cid:12) dt ≤ C (cid:32)(cid:90) T e qβA at | ˜ Y ht | q dt (cid:33) /q (cid:13)(cid:13)(cid:13)(cid:13) A y,εt − f y ( t, Y t , Z t ) a t (cid:13)(cid:13)(cid:13)(cid:13) L η ([0 ,T ]) ≤ C (cid:107) ˜ Y h (cid:107) S ,β (cid:13)(cid:13)(cid:13)(cid:13) A y,εt − f y ( t, Y t , Z t ) a t (cid:13)(cid:13)(cid:13)(cid:13) L η ([0 ,T ]) −→ ε → , in probability . Now, let η > small enough such that p + η ) ∈ (1 , . Then, by noticing that thereexists a positive constant c , such that (cid:12)(cid:12)(cid:12) A y,εt − f y ( t,Y t ,Z t ) a t (cid:12)(cid:12)(cid:12) ≤ cr t , since | f y ( t, y, z ) | ≤ r t forany ( t, y, z ) ∈ [0 , T ] × R and from ( iv ) , there exists a positive constant C such that sup ε ∈ (0 , E (cid:32)(cid:90) T e βA at | ˜ Y ht | (cid:12)(cid:12)(cid:12)(cid:12) A y,εt − f y ( t, Y t , Z t ) a t (cid:12)(cid:12)(cid:12)(cid:12) dt (cid:33) p + η ≤ C E (cid:34) sup t ∈ [0 ,T ] e ( p + η ) βA at | ˜ Y ht | p + η ) (cid:35) < + ∞ , since p + η ) < and ˜ Y h ∈ S ,β,a . Hence, using de La Vallée-Poussin Criterion, wededuce that the family of random variables (cid:40)(cid:32)(cid:90) T e βA at | ˜ Y ht | (cid:12)(cid:12)(cid:12)(cid:12) A y,εt − f y ( t, Y t , Z t ) a t (cid:12)(cid:12)(cid:12)(cid:12) dt (cid:33) p (cid:41) , ε ∈ (0 , , is uniformly integrable. Hence, by the dominated convergence theorem, we deduce that X y,εT −→ ε → . (3.9) e now turn to X z,εT . By proceeding similarly, we have X z,εT = E (cid:34)(cid:32)(cid:90) T e βA at | ˜ Z ht | (cid:12)(cid:12)(cid:12)(cid:12) A z,εt − f z ( t, Y t , Z t ) a t (cid:12)(cid:12)(cid:12)(cid:12) dt (cid:33) p (cid:35) . According to Assumption ( DsL p ,β ) ( iii ) and using the fact that ˜ Z h ∈ H ,β,a we knowthat for any t ∈ [0 , T ] (cid:90) T e βA at | ˜ Z ht | (cid:12)(cid:12)(cid:12)(cid:12) A z,εt − f z ( t, Y t , Z t ) a t (cid:12)(cid:12)(cid:12)(cid:12) dt proba −→ ε → . (3.10)Let η > small enough such that p + η ) ∈ (1 , . Then, we can show similarly thatthere exists a positive constant C such that sup ε ∈ (0 , E (cid:32)(cid:90) T e βA at | ˜ Z ht | (cid:12)(cid:12)(cid:12)(cid:12) A z,εt − f z ( t, Y t , Z t ) a t (cid:12)(cid:12)(cid:12)(cid:12) dt (cid:33) p + η ≤ C E (cid:32)(cid:90) T e βA at | ˜ Z ht | dt (cid:33) p + η < + ∞ , since p + η ) < and ˜ Z h ∈ H ,β,a . Hence, using de La Vallée-Poussin Criterion, (3.10)and the dominated convergence theorem, we deduce that X z,εT −→ ε → . (3.11)Finally, from (3.8), (3.9) and (3.11), we thus obtain for p ∈ (cid:0) , (cid:1) (cid:107) Y ε − ˜ Y h (cid:107) p S p + (cid:107) Z ε − ˜ Z h (cid:107) p H p −→ ε → . The rest of the proof is then similar to the proof of Theorem 5.1 in [30] and by applyingTheorem 2.2, we deduce that Y t ∈ D , and using [39, Lemma 2.3], one shows that Z belongs to L ([ t, T ]; D , ) . Besides, we can prove that a version of ( D u Y t , D u Z t ) ≤ u ≤ t, ≤ t ≤ T , is given as the solution to the affine BSDE: D u Y t = D u ξ + (cid:90) Tt ( D u f ( s, Y s , Z s ) + f y ( s, Y s , Z s ) D u Y s + f z ( s, Y s , Z s ) D u Z s ) ds − (cid:90) Tt D u Z s dW s , which admits, from Assumption ( DsL p ,β ) and Theorem 3.1, or [16], a unique solutionin S ,β,a × H ,β,a . (3.1) : an approach inspired by [2, 8] In this section, we will study the Malliavin differentiability of BSDE (3.1) in the stochasticLipschitz case by using the theory developed in [8]. A similar theory, using the BMOtheory was also developed in [2] but for particular stochastic Lipschitz BSDE (see BSDE(16) in [2, Section 4]). We recall Assumptions A1. and A2. from [8]. ( BC ) Assume that there exists a real predictable process K bounded from below by and a constant α ∈ (0 , such that i ) For each t ∈ [0 , T ] , ( y, z ) (cid:55)−→ f ( t, y, z ) is continuous, ( ii ) For any ( t, y, y (cid:48) , z, z (cid:48) ) ∈ [0 , T ] × R × ( L ([0 , T ])) , ( y − y (cid:48) )( f ( t, y, z ) − f ( t, y (cid:48) , z )) ≤ K αt | y − y (cid:48) | , and | f ( t, y, z ) − f ( t, y, z (cid:48) ) | ≤ K t (cid:107) z − z (cid:48) (cid:107) L ([0 ,T ]) . ( iii ) There exists a constant
C > such that for any stopping time τ ≤ T : E (cid:34)(cid:90) Tτ | K s | ds (cid:12)(cid:12)(cid:12) F τ (cid:35) ≤ C . We denote by N the smallest constant C which satisfies this statement.Notice that if the previous Assumption ( BC ) ( iii ) is satisfied then for any u ∈ L ([0 , T ]) with (cid:107) u (cid:107) L ([0 ,T ]) , (cid:16) M t := (cid:82) t K s u s dW s (cid:17) t ∈ [0 ,T ] is a BMO-martingale and (cid:107) M (cid:107) BMO = N . Let now Φ( p ) := (cid:18) p log (cid:18) p − p − (cid:19)(cid:19) − , and q (cid:63) such that Φ( q (cid:63) ) = N . We then defined p (cid:63) the conjugate of q (cid:63) , defined by q (cid:63) + 1 p (cid:63) = 1 . We now recall Assumption A3. and A4. of [8]. ( BC ) There exists p (cid:63) > p (cid:63) > such that E | ξ | p (cid:63) + (cid:32)(cid:90) T | f ( s, , | ds (cid:33) p (cid:63) < + ∞ . ( BC ) There exists a non negative predictable process F such that E (cid:32)(cid:90) T | F s | ds (cid:33) p (cid:63) < + ∞ , and ∀ ( t, y, z ) ∈ [0 , T ] × R × R , | f ( t, y, z ) | ≤ F t + K αt | y | + K t | z | , P − a.s. Then, we have the following a priori estimates for solutions to BSDE (3.1).
Theorem 3.3 (see Corollary 3.4 and Theorem 3.5 in [8]) . Assume that Assumptions ( BC ) , ( BC ) and ( BC ) hold. Then, BSDE (3.1) admits a unique solution ( Y, Z ) ∈S p × H p for any p < p (cid:63) . Besides, for each p ∈ ( p (cid:63) , p (cid:63) ) , (cid:107) Y (cid:107) S p + (cid:107) Z (cid:107) H p ≤ C E | ξ | p (cid:63) + (cid:32)(cid:90) T | f ( s, , | ds (cid:33) p (cid:63) p(cid:63) × E (cid:32)(cid:90) T K αs + K s ds (cid:33) P(cid:63) P(cid:63) , (3.12) where P (cid:63) = p ( p (cid:63) + p ) / ( p (cid:63) − p ) and C is a positive constant. e now set the following assumptions ( sD ∞ ) For any p > , ξ belongs to D ,p and ω (cid:55)−→ f ( t, ω, y, z ) belongs to L ([0 , T ]; D ,p ) .( sH , ∞ ) For any p > and for any h ∈ H lim ε → E (cid:34)(cid:32)(cid:90) T (cid:12)(cid:12)(cid:12)(cid:12) f ( s, · + εh, Y s , Z s ) − f ( s, · , Y s , Z s ) ε − (cid:104) Df ( s, · , Y s , Z s ) , ˙ h (cid:105) h (cid:12)(cid:12)(cid:12)(cid:12) ds (cid:33) p (cid:35) = 0 . ( sH , ∞ ) Let ( ε k ) k ∈ N be a sequence in (0 , such that lim k → + ∞ ε k = 0 , and let ( Y k , Z k ) k be a sequence of random variables which converges in S p × H p for any p < p ∗ to some ( Y, Z ) . Then for all h ∈ H , the following convergences hold in probability (cid:107) f y ( · , ω + ε k h, Y k · , Z · ) − f y ( · , ω, Y · , Z · ) (cid:107) L ([0 ,T ]) −→ k → + ∞ (cid:107) f z ( · , ω + ε k h, Y k · , Z k · ) − f z ( · , ω, Y · , Z · ) (cid:107) L ([0 ,T ]) −→ k → + ∞ , (3.13)or (cid:107) f y ( · , ω + ε k h, Y k · , Z k · ) − f y ( · , ω, Y · , Z · ) (cid:107) L ([0 ,T ]) −→ k → + ∞ (cid:107) f z ( · , ω + ε k h, Y · , Z k · ) − f z ( · , ω, Y · , Z · ) (cid:107) L ([0 ,T ]) −→ k → + ∞ . (3.14) Remark 3.3.
Notice that Assumption ( sH , ∞ ) implies that both ( BC2 ) and ( BC3 ) aretrue for any p ∗ > . Thus, Theorem 3.3 holds under ( BC1 ) and ( sH , ∞ ) and Inequality (3.12) is satisfied for any p > with a corresponding p ∗ which can be chosen greater than p ∗ defined by ( BC1 ) . We can now state the main result of this section.
Theorem 3.4.
Assume that ( BC ) - ( BC ) , ( D , ∞ ) , ( sH , ∞ ) and ( sH , ∞ ) hold. Then,for any p > and t ∈ [0 , T ] , Y t ∈ D ,p and Z ∈ L ([ t, T ]; D ,p ) . Besides, a version of ( D u Y t , D u Z t ) ≤ u ≤ t, ≤ t ≤ T , is given as the solution to the affine BSDE: D u Y t = D u ξ + (cid:90) Tt ( D u f ( s, Y s , Z s ) + f y ( s, Y s , Z s ) D u Y s + f z ( s, Y s , Z s ) D u Z s ) ds − (cid:90) Tt D u Z s dW s . (3.15) Proof.
The proof is similar to that of Theorem 3.2. We only consider the case where(3.13) holds in Assumption ( sH , ∞ ) since the other one can be treated similarly. Firstnotice that under Assumption ( sD ∞ ) , ( BC ) - ( BC ) and according to Theorem 3.3together with Remark 3.3, for any p > , (cid:107) Y (cid:107) S p + (cid:107) Z (cid:107) H p < + ∞ . We aim at applyingTheorem 2.2 (see [30]) with F = Y t and F = (cid:82) Tt Z s dW s . The proof is very close to theproof of Theorem 5.1 in [30] and we recall here the main ideas. Let ε > and h ∈ H .We have Y s ◦ τ εh = ξ ◦ τ εh + (cid:90) Ts f ( r, Y r , Z r ) ◦ τ εh dr − (cid:90) Ts Z r ◦ τ εh dW r , ∀ s ∈ [ t, T ] , P − a.s. (3.16)As a consequence, setting for simplicity Y εs := 1 ε ( Y s ◦ τ εh − Y s ) , Z εs := 1 ε ( Z s ◦ τ εh − Z s ) , ξ ε := 1 ε ( ξ ◦ τ εh − ξ ) , s ∈ [ t, T ] , we have that ( Y ε , Z ε ) solves the BSDE: Y εs = ξ ε + (cid:90) Ts ( ˜ A εr + ˜ A y,εr Y εr + ˜ A z,εr Z εr ) dr − (cid:90) Ts Z εr dW r , (3.17) ith ˜ A y,εr := (cid:90) f y ( r, · + εh, Y r + θ ( Y r ◦ τ εh − Y r ) , Z r ) dθ, ˜ A z,εr := (cid:90) f z ( r, · + εh, Y r ◦ τ εh , Z r + θ ( Z r ◦ τ εh − Z r )) dθ, ˜ A εr := 1 ε ( f ( r, · + εh, Y r , Z r ) − f ( r, · , Y r , Z r )) . Hence, under Assumptions ( BC ) - ( BC ) , ( sD , ∞ ) , according to Theorem 3.3, ( Y ε , Z ε ) is the unique solution of BSDE (3.16) in S p × H p for any p > .Consider now the following stochastic affine BSDE on [ t, T ] , which admits a uniquesolution ( ˜ Y h , ˜ Z h ) ∈ ( S p ×H p ) for any p > according to Theorem 3.3 under Assumption ( BC ) - ( BC ) , ( sD , ∞ ) , ˜ Y hs = (cid:104) Dξ, ˙ h (cid:105) L ([0 ,T ]) − (cid:90) Ts ˜ Z hr dW r + (cid:90) Ts (cid:16) (cid:104) Df ( r, Y r , Z r ) , ˙ h (cid:105) L ([0 ,T ]) + ˜ Y hr f y ( r, Y r , Z r ) + ˜ Z hr f z ( r, Y r , Z r ) (cid:17) dr. (3.18)Hence, using Theorem 3.3 we obtain for any p ∗ > p > (cid:107) Y ε − ˜ Y h (cid:107) S p + (cid:107) Z ε − ˜ Z h (cid:107) H p ≤ C E | ξ ε − (cid:104) Dξ, ˙ h (cid:105) L ([0 ,T ]) | p (cid:63) + (cid:32)(cid:90) T | X εs + X y,εs + X z,εs | ds (cid:33) p (cid:63) p(cid:63) × E (cid:32)(cid:90) T ( K αs + K s ) ds (cid:33) P (cid:63) / /P (cid:63) , where X εs := ˜ A εs − (cid:104) Df ( s, Y s , Z s ) , ˙ h (cid:105) L ([0 ,T ]) X y,εs := ˜ Y hs ( ˜ A y,εs − f y ( s, Y s , Z s )) X z,εs := ˜ Z hs ( ˜ A z,εs − f z ( s, Y r , Z r )) . Notice that under ( BC1 ) ( iii ) we have E (cid:32)(cid:90) T ( K αs + K s ) ds (cid:33) P (cid:63) / < + ∞ . Hence, after the same kind of calculations than those made in the proofs of Theorem 5.1in [30] or Theorem 3.2 above, we deduce that for any t ∈ [0 , T ] and any p > , Y t ∈ D ,p and Z ∈ L ([ t, T ]; D ,p ) and that their Malliavin derivatives are solutions to (3.15). We begin with Assumption ( DsL p ,β ) and the first approach inspired by [48]. Even ifAssumption ( i ) is not too restrictive in view of the theory developed in [48], in practicewe could have some difficulties to verify ( ii ) and ( iii ) . Indeed, in ( ii ) we have to controlthe norm in H p,β,a of /a t , and ( iii ) requires a control of the ess sup of the derivative of f with respect to z . As soon as r and θ are random, these assumptions restrict significantlythe range of possible applications. As explained above, these assumptions are stronglylinked to a priori estimates obtained in [48], which suggests to modify the proofs in [48]to try to obtain weaker conditions, if possible. oncerning the second approach, a priori estimates (3.12) seem to be better, since theyare similar to those obtained in the Lipschitz or quadratic case (see [9]). Notice howeverthat the order of these a priori estimates depends closely on the BMO-norm of thestochastic integral of the Lipschitz constant K , which in practice, could be quite difficultto control. We provide conditions in D ,p for any p > due to the control of the norm of Y and Z at an order depending on this BMO-norm. Assumptions ( sD ∞ ) and ( sH , ∞ ) are not so surprising, since they are similar to conditions obtained in Section 7 in [30]when dealing with quadratic growth BSDEs.From now, we set the following two assumptions. ( EKH p ,β ) Let Assumptions ( sL ,β ) and ( DsL p ,β ) hold. ( BC ) Let Assumptions ( BC ) - ( BC ) , ( sD ∞ ) , ( sH , ∞ ) and ( sH , ∞ ) hold. We now study a particular stochastic Lipschitz BSDE in the non-Markovian case: Y t = ξ + (cid:90) Tt f ( s, Y s , Z s ) ds − (cid:90) Tt Z s dW s , ∀ t ∈ [0 , T ] , P − a.s. (3.19)where f : [0 , T ] × Ω × R × R −→ R ( t, ω, y, z ) (cid:55)−→ λ s ( ω ) + µ s ( ω ) y + ν s ( ω ) z, and where ξ is an F T -measurable random variable and λ, µ, ν : [0 , T ] × Ω −→ R are F -progressively measurable processes. Remark 3.4.
The BSDE (3.19) studied in this section generalizes [2, BSDE (5)] foraffine BSDEs, since the generator of (3.19) is affine in both Y and Z . By adding aLipschitz coefficient with respect to Y and Z in the generator satisfying Assumption ( A in [2], one could show that we strictly extend [2, Section 3]. Besides, we insist onthe fact that λ , µ and ν are not bounded, which also extend the results in [30, Section6.2]. ( A ) There exists a constant
C > such that for any stopping time τ ≤ T : E (cid:34)(cid:90) Tτ | ν s | ds (cid:12)(cid:12)(cid:12) F τ (cid:35) ≤ C . ( A ) For any p > , ( i ) exp (cid:0)(cid:82) · | µ s | ds (cid:1) ∈ S p , and ( ii ) E (cid:104) | ξ | p + (cid:16)(cid:82) T | λ t | dt (cid:17) p (cid:105) < + ∞ . Before going further, notice that ( A ) is equivalent to saying that (cid:82) · ν s dW s is a BMO-martingale, which corresponds to Assumption ( A2 ) in [2] or Assumption A1. in [8].However, we do not assume that the same statement holds for (cid:82) · µ s dW s . Indeed, in ( A ) we just assume that the process (cid:0)(cid:82) · µ s ds (cid:1) t ∈ [0 ,T ] has exponential moments of allorders. Theorem 3.5.
Assume that Assumptions ( A ) and ( A ) hold. Then, BSDE (3.19) admits a unique solution ( Y, Z ) ∈ S p × H p for any p > . Besides, Estimate (3.12) holdsfor any < p < p ∗ .Proof. ( A ) and ( A )( i ) are weaker assumptions than ( BC ) , so we cannot apply di-rectly Corollary 3.4 and Theorem 3.5 in [8]. However, by reproducing the proof ofLemma 3.2, Lemma 3.3, Corollary 3.4 and Theorem 3.5 in [8], one notices that we only eed to have a BMO-property for (cid:82) · ν s dW s , since only Relation (2) in [8], correspondingto ( A )( i ) , is used to deal with terms depending on µ . Hence, for affine BSDE (3.19)we can make replace Assumption ( BC ) with Assumptions ( A ) and ( A ) . We thendeduce that BSDE (3.19) admits a unique solution ( Y, Z ) ∈ S p × H p for any p > and(3.12) holds for any < p < p ∗ .In this particular case, Assumptions ( sD ∞ ) , ( sH , ∞ ) and ( sH , ∞ ) become ( DA ) For any p > , ξ belongs to D ,p and the stochastic processes ( t, ω ) (cid:55)−→ λ t ( ω ) , µ t ( ω ) , ν t ( ω ) belong to L ([0 , T ]; D ,p ) .( DA ) Let ( ε k ) k ∈ N be a sequence in (0 , such that lim k → + ∞ ε k = 0 , and let ( Y k , Z k ) k bea sequence of random variables which converges in S p × H p for any p > to some ( Y, Z ) .Then for all h ∈ H , the following convergences hold in probability (cid:107) µ · ( ω + ε k h ) − µ · ( ω ) (cid:107) L ([0 ,T ]) −→ k → + ∞ , (cid:107) ν · ( ω + ε k h ) − ν · ( ω ) (cid:107) L ([0 ,T ]) −→ k → + ∞ . Theorem 3.6.
Assume that ( A ) , ( A ) , ( DA ) and ( DA ) hold. Then, by denoting ( Y, Z ) the unique solution of (3.19) , for any p > and t ∈ [0 , T ] , Y t ∈ D ,p and Z ∈ L ([ t, T ]; D ,p ) . Besides, a version of ( D u Y t , D u Z t ) ≤ u ≤ t, ≤ t ≤ T , is given as the solution to the affine BSDE: D u Y t = D u ξ + (cid:90) Tt ( D u λ s + D u µ s Y s + D u ν s Z s + µ s D u Y s + ν s D u Z s ) ds − (cid:90) Tt D u Z s dW s , (3.20) Proof.
Under ( A ) and ( A ) and according to Theorem 3.5, BSDE (3.19) admits aunique solution ( Y, Z ) ∈ S p × H p for any p > . Now, Assumptions ( sD ∞ ) , ( sH , ∞ ) and ( sH , ∞ ) are automatically satisfied under ( DA ) and ( DA ) . Hence, by applying The-orem 3.4, we deduce that for any p > and t ∈ [0 , T ] , Y t ∈ D ,p and Z ∈ L ([ t, T ]; D ,p ) and a version of ( DY t , DZ t ) is given by the solution to BSDE (3.20). Y component We now aim at applying Bouleau and Hirsch’s Criterion (see Theorem 2.3) to the Y component of the solution ( Y, Z ) of BSDE (3.1). We set the following assumption (A p,β ) Let p be in ∈ (cid:0) , (cid:1) , β > max { / (2 p − } and let Assumption ( ELK ) p ,β holds and assume moreover that (cid:82) · θ s dW s ∈ BMO ( P ) , where θ is defined in Assumption ( sL ,β ) .Notice that under Assumption (A p,β ) or Assumption (BC) , we have proved that Y t ∈ D ,p and Z ∈ L ([0 , T ]; D ,p ) for some p > and that their Malliavin derivatives ( D r Y t , D r Z t ) satisfy the linear BSDE (3.5) (see Theorems 3.2 and 3.4). We then havethe following theorem which gives conditions ensuring that, given a time t , the law of thefirst component Y t of the solution of the non-Markovian BSDE (3.1) admits a density. Theorem 4.1.
Let ( A p ,β ) or ( BC ) hold. Denote by ( Y, Z ) the unique solution of BSDE (3.1) . If there exists A ⊂ Ω such that P ( A ) > and one of the following two assumptionsholds for t ∈ (0 , T ] and s ∈ [ t, T ] sH+ ) D u ξ ≥ , D u f ( s, Y s , Z s ) ≥ , λ ( du ) − a.e., and D u ξ > , λ ( du ) − a.e. on A ( sH- ) D u ξ ≤ , D u f ( s, Y s , Z s ) ≤ , λ ( du ) − a.e., and D u ξ < , λ ( du ) − a.e. on A, then the law of Y t is absolutely continuous with respect to Lebesgue measure.Proof. Under Assumptions (A p,β ) or (BC) , we know from respectively Theorems 3.1and 3.2 or Theorems 3.3 and 3.4, that BSDE (3.3) admits a unique solution ( Y, Z ) which is Malliavin differentiable, whose Malliavin derivatives ( D u Y t , D u Z t ) ≤ u ≤ t ≤ T aresolutions to the following linear BSDE D u Y t = D u ξ + (cid:90) Tt ( D u f ( s, Y s , Z s ) + f y ( s, Y s , Z s ) D u Y s + f z ( s, Y s , Z s ) D u Z s ) ds − (cid:90) Tt D u Z s dW s , ∀ ≤ u ≤ t ≤ T, P − a.s. Notice that for any ( t, y, z ) ∈ [0 , T ] × R , | f z ( t, y, z ) | ≤ θ t under Assumption (A p,β ) or | f z ( t, y, z ) | ≤ K t under Assumption (BC) . Hence, we can define a probability measure Q by d Q d P := E (cid:32)(cid:90) T f z ( s, Y s , Z s ) dW s (cid:33) = e (cid:82) T f z ( s,Y s ,Z s ) dW s − (cid:82) T | f z ( s,Y s ,Z s ) | ds , where E (cid:0)(cid:82) · f z ( s, Y s , Z s ) dW s (cid:1) is a uniformly integrable martingale according to [25, The-orem 2.3]. Changing the Brownian motion according to Girsanov’s Theorem and usinga linearisation (see [17]), we obtain for any ≤ u ≤ t ≤ TD u Y t = E Q t (cid:34) D u ξe (cid:82) Tt f y ( s,Y s ,Z s ) ds + (cid:90) Tt e (cid:82) st f y ( s,Y s ,Z s ) du D u f ( s, Y s , Z s ) ds (cid:35) ≥ , du ⊗ d P − a.e. Moreover, let A be such that P ( A ) > , and D u ξ > on A . We obtain D u Y t ≥ E Q t (cid:104) A D u ξe (cid:82) Tt f y ( s,Y s ,Z s ) ds (cid:105) > . Thus, (cid:107) DY t (cid:107) L ([0 ,T ]) > , P − a.s. and from Theorem 2.3 the law of Y t is absolutelycontinuous with respect to the Lebesgue measure.The proof under Assumption ( sH- ) is similar. In this section, we study a particular class of stochastic Lipschitz BSDE (3.1), which thegenerator is Lipschitz in its space variables with a nonnegative Lipschitz constant. Weprovide weaker conditions than Conditions ( sH+ ) and ( sH- ) ensuring that the law ofthe component Y t of the solution to the corresponding Lipschitz BSDE has a density.We consider the following non-Markovian Lipschitz BSDE Y t = ξ + (cid:90) Tt f ( s, Y s , Z s ) ds − (cid:90) Tt Z s dW s , ∀ t ∈ [0 , T ] , P − a.s. (4.1)where ξ is an F T -measurable random variable and f : [0 , T ] × Ω × R −→ R is an F -progressively measurable process where as usual the ω -dependence is omitted. We setthe following assumption L ) ( i ) The map ( y, z ) (cid:55)−→ f ( · , y, z ) is differentiable with continuous partial derivativesuniformly bounded by a positive constant m . We denote by f y (resp. f z ) thederivative of f with respect to y (resp. z ). ( ii ) We have E (cid:34) | ξ | + (cid:90) T | f ( s, , | ds (cid:35) < + ∞ . Theorem 4.2 ([17]) . Under Assumption ( L ) , there exists a unique pair of adapted pro-cesses ( Y, Z ) which solves BSDE (4.1) in S × H . We now turn to the Malliavin differentiability of the solution ( Y, Z ) of BSDE (4.1). Thisproblem was studied in [39] in the Markovian case with Lipschitz coefficients ( i.e. whenthe data ξ, f ( t, · , y, z ) are functions of the solution of a Brownian SDE). It was extendedin [17] to the non-Markovian case with Lipschitz coefficients. This question was thenstudied in [19] for Lévy driven BSDEs and in [30] in which the conditions improve thosein [17] (see [30, Section 6.3]). In this section, we recall the results of [30] where a newcriterion ensuring that a random variable is in D , has been proved. Set the followingassumption ( lD ) – ξ ∈ D , , for any ( y, z ) ∈ R , ( t, ω ) (cid:55)−→ f ( t, ω, y, z ) is in L ([0 , T ]; D , ) , f ( · , y, z ) and Df ( · , y, z ) are F -progressively measurable, and E (cid:34)(cid:90) T (cid:107) D · f ( s, Y s , Z s ) (cid:107) h ds (cid:35) < + ∞ . – There exists p ∈ (1 , such that for any h ∈ H lim ε → E (cid:34)(cid:32)(cid:90) T (cid:12)(cid:12)(cid:12)(cid:12) f ( s, · + εh, Y s , Z s ) − f ( s, Y s , Z s ) ε − (cid:104) Df ( s, Y s , Z s ) , ˙ h (cid:105) h (cid:12)(cid:12)(cid:12)(cid:12) ds (cid:33) p (cid:35) = 0 , – Let ( ε n ) n ∈ N be a sequence in (0 , such that lim n → + ∞ ε n = 0 , and let ( Y n , Z n ) n be a sequence of random variables which converges in S × H to some ( Y, Z ) .Then for all h ∈ H , the following convergences hold in probability (cid:107) f y ( · , ω + ε n h, Y n · , Z · ) − f y ( · , ω, Y · , Z · ) (cid:107) h −→ n → + ∞ (cid:107) f z ( · , ω + ε n h, Y n · , Z n · ) − f z ( · , ω, Y · , Z · ) (cid:107) h −→ n → + ∞ , (4.2)or (cid:107) f y ( · , ω + ε n h, Y n · , Z n · ) − f y ( · , ω, Y · , Z · ) (cid:107) h −→ n → + ∞ (cid:107) f z ( · , ω + ε n h, Y · , Z n · ) − f z ( · , ω, Y · , Z · ) (cid:107) h −→ n → + ∞ . (4.3) Theorem 4.3 (Theorem 5.1 in [30]) . Let ( Y, Z ) be the solution of BSDE (4.1) underAssumption ( L ) . Let Assumption ( lD ) be satisfied, then for any t ∈ [0 , T ] , Y t ∈ D , and Z ∈ L ([ t, T ]; D , ) .Besides, by denoting DY t (resp. DZ t ) the Malliavin derivative of Y t (resp. Z t ), the pair ( DY, DZ ) satisfies the following (linear) BSDE D u Y t = D u ξ + (cid:90) Tt ( D u f ( s, Y s , Z s ) + f y ( s, Y s , Z s ) D u Y s + f z ( s, Y s , Z s ) D u Z s ) ds − (cid:90) Tt D u Z s dW s , ≤ u ≤ t ≤ T, P − a.s. (4.4) e now aim at applying Bouleau and Hirsch’s Criterion (see Theorem 2.1.2 in [35])to the Y component of the solution ( Y, Z ) of BSDE (4.1). The existence of a densityfor Y t when t ∈ (0 , T ] when f is Lipschitz in its space variable was solved in [3] in theMarkovian case. We want to extend this result to the non-Markovian case. The followingtheorem gives conditions which ensure that, given a time t , the first component Y t ofthe solution of the non-Markovian BSDE (4.1) admits a density under ( L ) and ( lD ) .These conditions are similar to those of [3, Theorem 3.1] in the Lipschitz Markovian case.Following [3, 1, 29], let A be a subset of Ω such that P ( A ) > . We set dξ := max { M ∈ R , D u ξ ≥ M, du ⊗ P − a.e. } ,df ( t ) := max { M ∈ R , ∀ s ∈ [ t, T ] D u f ( s, Y s , Z s ) ≥ M, du ⊗ P − a.e. } ,dξ A : = max { M ∈ R , D u ξ ≥ M, du − a.e. on A. } ,dξ := min { M ∈ R , D u ξ ≤ M, du ⊗ P − a.e. } ,df ( t ) := min { M ∈ R , ∀ s ∈ [ t, T ] D u f ( s, Y s , Z s ) ≤ M, du ⊗ P − a.e. } ,dξ A := min { M ∈ R , D u ξ ≤ M, du − a.e. on A } . Theorem 4.4.
Let ( Y, Z ) be the solution of BSDE (4.1) under Assumptions ( L ) and ( lD ) . Fix some t ∈ (0 , T ] . If there exists A ⊂ Ω such that P ( A ) > and one of the twofollowing assumptions holds ( H+ ) dξe − sgn ( dξ ) m ( T − t ) + df ( t ) (cid:90) Tt e − sgn ( df ( t )) m ( s − t ) ds ≥ dξ A e − sgn ( dξ A ) m ( T − t ) + df ( t ) (cid:90) Tt e − sgn ( df ( t )) m ( s − t ) ds > H- ) dξe − sgn ( dξ ) m ( T − t ) + df ( t ) (cid:90) Tt e − sgn ( df ( t )) m ( s − t ) ds ≤ dξ A e − sgn ( dξ A ) m ( T − t ) + df ( t ) (cid:90) Tt e − sgn ( df ( t )) m ( s − t ) ds < , then Y t has a law absolutely continuous with respect to the Lebesgue measure.Proof. The proof follows the same line than the one of [3, Theorem 3.1]. Assume that ( H+ ) holds. We aim at applying Bouleau-Hirsch Criterion (Theorem 2.1.2 in [35]).From Theorem 4.3, Y t ∈ D , and Z ∈ L ([ t, T ]; D , ) . Let ≤ u ≤ t ≤ T , using thelinearisation method for BSDE (see [17]) we have D u Y t = D u ξ + (cid:90) Tt ( D u f ( s, Y s , Z s ) + f y ( s, Y s , Z s ) D u Y s + f z ( s, Y s , Z s ) D u Z s ) ds − (cid:90) Tt D u Z s dW s , = E Q t (cid:34) D u ξe (cid:82) Tt f y ( s,Y s ,Z s ) ds + (cid:90) Tt e (cid:82) st f y ( α,Y α ,Z α ) dα D u f ( s, Y s , Z s ) ds (cid:35) ≥ dξ E Q t (cid:104) e (cid:82) Tt f y ( s,Y s ,Z s ) ds (cid:105) + df ( t ) E Q t (cid:34)(cid:90) Tt e (cid:82) st f y ( α,Y α ,Z α ) dα ds (cid:35) ≥ dξe − sgn ( dξ ) m ( T − t ) + df ( t ) (cid:90) Tt e − sgn ( df ( s )) m ( s − t ) ds. Hence, D u Y t ≥ , du ⊗ P − a.e . Moreover, let A be such that P ( A ) > , we obtain u Y t ≥ dξ A E Q t (cid:104) A e (cid:82) Tt f y ( s,Y s ,Z s ) ds (cid:105) + df ( t ) E Q t (cid:34)(cid:90) Tt e (cid:82) st f y ( α,Y α ,Z α ) dα ds (cid:35) > . Thus, (cid:107) DY t (cid:107) L ([0 ,T ]) > , P − a.s. and from Theorem 2.3 the law of Y t is absolutelycontinuous with respect to the Lebesgue measure.The proof under Assumption ( H- ) is similar. We study the existence of a density for the first component of the solution to BSDE(3.19). We set dλ ( t ) := max { M ∈ R , ∀ s ∈ [ t, T ] D u λ s ≥ M, du ⊗ P − a.e. } , and dλ ( t ) := min { M ∈ R , ∀ s ∈ [ t, T ] D u λ s ≤ M, du ⊗ P − a.e. } . We set also the following assumption ( P +) Let ( Y t , Z t ) be the unique solution of BSDE (3.19) for any t ∈ [0 , T ] . Then, for any s ∈ [0 , t ] , D u µ s Y s + D u ν s Z s ≥ , du ⊗ P − a.s. ( P− ) Let ( Y t , Z t ) be the unique solution of BSDE (3.19) for any t ∈ [0 , T ] . Then, for any s ∈ [0 , t ] , D u µ s Y s + D u ν s Z s ≤ , du ⊗ P − a.s. Remark 4.1.
Notice that if ( Y, Z ) is the unique solution to BSDE (3.19) , hence usinga linearisation method (see [17]) Y t := E Q t (cid:34) ξe (cid:82) Tt µ s ds + (cid:90) Tt λ s e (cid:82) ts µ α dα ds (cid:35) , which is non negative as soon as ξ and λ are non negative, where d Q d P := e (cid:82) T ν s dW s − (cid:82) T | ν s | ds , as soon as E (cid:0)(cid:82) · ν s dW s (cid:1) is a martingale (see Assumption ( BMO ) below together with[25, Theorem 2.3]). Thus, as soon as Y is non negative, if µ is a semi-martingale, wecan give conditions which ensures that Dµ t is non negative using a Lamperti transform(see e.g. [1, Step 1 of the proof of Theorem 3.3]).Concerning the Z process, conditions ensuring that Z is non negative has been obtainedin [13] in the Markovian case only. In the non-Markovian case, this problem is still open,as far as we know. However, if ν is deterministic, conditions ( P +) and ( P− ) can besimplified (see Section 7). Theorem 4.5.
Assume that ( A ) , ( A ) , ( DA ) and ( DA ) hold and let ( Y, Z ) be thesolution of BSDE (3.19) . If there exists A ⊂ Ω such that P ( A ) > and one of the twofollowing assumptions holds ( aH +) D u ξ ≥ , λ ( du ) − a.e., D u ξ > , λ ( du ) − a.e. on A, dλ ( t ) ≥ and ( P +) holds, ( aH − ) D u ξ ≤ , λ ( du ) − a.e., D u ξ < , λ ( du ) − a.e. on A, dλ ( t ) ≤ and ( P− ) holds , then Y t has a law absolutely continuous with respect to the Lebesgue measure. roof. We prove the previous theorem under Assumption ( aH+ ) . Let t ∈ (0 , T ] . Weknow from Theorem 3.5 and Theorem 3.6 that BSDE (3.19) admits a unique solution ( Y t , Z t ) t ∈ [0 ,T ] such that Y t ∈ D ,p and Z ∈ L ([ t, T ]; D ,p ) for any p > with derivatives ( DY t , DZ t ) satisfying the following linear BSDE D u Y t = D u ξ + (cid:90) Tt ( D u λ s + D u µ s Y s + D u ν s Z s + µ s D u Y s + ν s D u Z s ) ds − (cid:90) Tt D u Z s dW s , ≤ u ≤ t ≤ T, P − a.s. Set Q a probability measure defined by d Q d P := E (cid:32)(cid:90) T ν s dW s (cid:33) = e (cid:82) T ν s dW s − (cid:82) T | ν s | ds , where E (cid:0)(cid:82) · ν s dW s (cid:1) is a uniformly martingale according to [25, Theorem 2.3]. Changingthe Brownian motion according to Girsanov’s Theorem and using a linearisation, weobtain for any ≤ u ≤ t ≤ TD u Y t = E Q t (cid:34) D u ξe (cid:82) Tt µ s ds + (cid:90) Tt ( D u λ s + D u µ s Y s + D u ν s Z s ) e (cid:82) st µ α dα ds (cid:35) . (4.5)Thus, by reproducing the proof of Theorem 4.4, we show that (cid:107) DY t (cid:107) L ([0 ,T ]) > , P − a.s. and from Theorem 2.3 the law of Y t is absolutely continuous with respect to the Lebesguemeasure.The proof under ( aH- ) is similar. Remark 4.2.
Under the same assumptions than in the previous theorem and assumingthat ( aD+ ) holds (resp. ( aD- ) holds), the proof shows that in fact D u Y t ≥ , du ⊗ P − a.e. (resp. D u Y t ≤ , du ⊗ P − a.e. ). Z -component: a stillopen problem in the non-Markovian case In this section we turn to conditions ensuring the existence of densities for the laws of Z t components of solutions to stochastic Lipschitz BSDEs. We begin to investigate thisproblem for a particular class of stochastic Lipschitz BSDE with a linear generator withrespect to the z component by following the same proofs that in [1] and we explain whywe are not able to extend results obtained in [29] to the non-Markovian case for generalstochastic Lipschitz BSDEs. z Consider the following BSDE Y t = ξ + (cid:90) Tt ( ˜ f ( s, Y s ) + θ s Z s ) ds − (cid:90) Tt Z s dW s , ∀ t ∈ [0 , T ] , P − a.s. (5.1)where θ is a square integrable adapted process. In this case, recall that under ( EKH ) p ,β for any p be in ∈ (cid:0) , (cid:1) , β > max { / (2 p − } or ( BC ) , according to Theorem 3.2 orrespectively Theorem 3.4, BSDE (5.1) admits a unique solution in S p × H p and t ∈ [0 , T ] , Y t ∈ D , and Z ∈ L ([ t, T ]; D , ) . Besides, a version of ( D u Y t , D u Z t ) ≤ u ≤ t, ≤ t ≤ T is given by the solution to the affine BSDE: u Y t = D u ξ + (cid:90) Tt (cid:16) D u ˜ f ( s, Y s ) + f y ( s, Y s ) D u Y s + θ s D u Z s (cid:17) ds − (cid:90) Tt D u Z s dW s . (5.2)As explain in Remark 4.1, in order to obtain a sign for the Malliavin derivative of the Y component of the solution to an affine BSDE with unbounded coefficients when we haveno information on the sign of the Z process, we must assume that θ is deterministic toapply Theorem 4.5. Thus, we set the following assumption ( Θ ) The process θ defined in BSDE (5.1) is deterministic.Let now Y be the first component of the solution to BSDE (5.2). We set for any ≤ v ≤ t ≤ T ( DY +) D v ξ ≥ , D v ˜ f ( t, Y t ) ≥ , P − a.s., ( DY − ) D v ξ ≤ , D v ˜ f ( t, Y t ) ≤ , P − a.s. Remark 5.1.
Similarly to Remark 4.2, Notice that the proof of Theorem 4.1 shows thatfor any ≤ v ≤ s ≤ T :Under Assumption ( DY +) D v Y t ≥ , P − a.s. (5.3) Under Assumption ( DY − ) D v Y t ≤ , P − a.s. (5.4)We have the following theorem which provide conditions on the data ξ and ˜ f ensuringthat the law of Z t has a density with respect to the Lebesgue measure, which can beseen as an extension of [1, Theorem 4.3] to the stochastic Lipschitz case Theorem 5.1.
Let ( Θ ) be hold and let ( Y, Z ) be the unique solution of BSDE (5.1) . Let ξ be in D , , assume moreover that ˜ f is twice continuously differentiable with respect to y . We set the following assumptions for any ≤ t, t (cid:48) ≤ T : ( f +) D t ˜ f y ( t, Y t ) , ( D t ˜ f ) y ( t, Y t ) ≥ , ( f − ) D t ˜ f y ( t, Y t ) , ( D t ˜ f ) y ( t, Y t ) ≤ . Assume that there exists A such that P ( A ) > such that one of the following assumptionsis satisfied ( DZ +) D t (cid:48) ( D t ξ ) ≥ , P − a.e. , D t (cid:48) ( D t ξ ) > on A, D t (cid:48) ( D t ˜ f )( t, Y t ) ≥ (5.5) and Assumptions ( DY +) and ( f +) hold, or Assumptions ( DY − ) and ( f − ) hold, ( DZ − ) D t (cid:48) ( D t ξ ) ≤ , P − a.e. , D t (cid:48) ( D t ξ ) < on A, D t (cid:48) ( D t ˜ f )( t, Y t ) ≤ (5.6) and Assumptions ( DY − ) and ( f +) hold, or Assumptions ( DY +) and ( f − ) hold.Then, the law of Z t is absolutely continuous with respect to Lebesgue’s measure for any t ∈ (0 , T ] . roof. Let ( Y, Z ) be the unique solution to BSDE (5.2) in D , × L ([0 , T ]; D , ) , whichthe Malliavin derivatives are solutions to BSDE (5.2).Let Assumption ( DZ +) be true together with Assumption ( DY +) and ( f +) . We followthe proof of [1, Theorem 4.3] by taking the advantage of the representation of the Z process with Clark-Ocone Formula. Using now a linearization and according to Clark-Ocone Formula, we obtain Z t = E Q t (cid:34) D t ξe (cid:82) Tt ˜ f y ( s,Y s ) ds + (cid:90) Tt D t ˜ f ( s, Y s ) e (cid:82) st ˜ f y ( u,Y u ) du ds (cid:35) , with d Q d P = exp (cid:16)(cid:82) T θ s dW s − (cid:82) T | θ s | ds (cid:17) . Let ≤ v ≤ t , we have D v Z t = E Q t (cid:34) D v ( D t ξ ) e (cid:82) Tt ˜ f y ( s,Y s ) ds + D t ξe (cid:82) Tt ˜ f y ( s,Y s ) ds (cid:90) Tt D v ˜ f y ( s, Y s ) ds + (cid:90) Tt e (cid:82) st ˜ f y ( u,Y u ) du (cid:18) D v,t ˜ f ( s, Y s ) + D t ˜ f ( s, Y s ) (cid:90) st D v ˜ f y ( u, Y u ) du (cid:19) ds (cid:35) . (5.7)Hence, using the definition (2.4) of D v,t ˜ f , Inequality (5.3), Assumption ( f +) and As-sumption (5.5), we deduce that for any ≤ v < t ≤ T , D v Z t > . Thus, the law of Z t has a density for any t ∈ (0 , T ] as a consequence of Theorem 2.3.The proof under Assumptions ( DZ +) , ( DY − ) and ( f − ) is similar, by using (5.7), In-equality (5.4), Assumption ( ii ) and Assumption (5.5).Concerning Assumption ( DZ − ) we follow exactly the same proof and for any ≤ v ≤ t ≤ T , we show that D v Z t < , P − a.s. . Remark 5.2.
Theorem 5.1 extends the results in [1]. In the present paper θ is assumedto be a deterministic map behind the z part of the generator, unlike the model studied in[1] in which the coefficient behind z is constant. Moreover, in our model ˜ f is stochasticLipschitz with respect to its y variable, whereas it is assumed to be Lipschitz in [1].Finally, we deal with the non-Markovian case for both the terminal condition and thegenerator of the BSDE, whereas [1] considers the case where only the terminal conditionis non-Markovian. Existence of density for the Z component has been studied for quadratic growth BSDEsin [29] in the Markovian case. We can in fact adapt this proof to the Markovian stochasticLipschitz case and one could show that conditions ensuring that the law of Z t componenthas a density are similar to those obtained for Markovian quadratic growth BSDE (see[29, Section 4.3]). Although in the latter paper, the authors obtain conditions whichensure that Z t admits a density, we can not reproduce the proof here since it is essentiallybased on Ma-Zhang Representation (see [28, Lemma 2.4]) which holds in the Markoviancase. More precisely, we consider the following forward-backward SDE X t = X + (cid:90) t b ( s, X s ) ds + (cid:90) t σ ( s, X s ) dW s , t ∈ [0 , T ] , P − a.s.Y t = g ( X T ) + (cid:90) Tt f ( s, X s , Y s , Z s ) ds − (cid:90) Tt Z s dW s , t ∈ [0 , T ] , P − a.s. (5.8)Then, under some conditions on the data of such forward-backward system, denotingby ( X, Y, Z ) the solution of (5.8), there exists a version of ( D u X t , D u Y t , D u Z t ) for all < u ≤ t ≤ T which satisfies: D u X t = ∇ X t ( ∇ X u ) − σ ( u, X u ) ,D u Y t = ∇ Y t ( ∇ X u ) − σ ( u, X u ) ,D u Z t = ∇ Z t ( ∇ X u ) − σ ( u, X u ) , here ( ∇ X, ∇ Y, ∇ Z ) is the solution to the following FBSDE: ∇ X t = (cid:90) t b x ( s, X s ) ∇ X s ds + (cid:90) t σ x ( s, X s ) ∇ X s dW s , ∇ Y t = g (cid:48) ( X T ) ∇ X T + (cid:90) Tt ( f x ( s, X s , Y s , Z s ) ∇ X s + f y ( s, X s , Y s , Z s ) ∇ Y s + f z ( s, X s , Y s , Z s ) ∇ Z s ) ds − (cid:90) Tt ∇ Z s dW s . (5.9)As far as we know, the same kind of decomposition is still open for path-dependentBSDEs. However, it seems to be hard to obtain a similar formula in the path-dependentframework. As an example, let Y T = ξ = (cid:82) T B s ds . Hence, Y T ∈ D , and D r Y T = T − r .In order to separate the Malliavin integration variable r and the time variable T as in[28, Lemma 2.4] for Markovian BSDEs, we could similarly compute the gradient in spaceof ξ using a Fréchet derivative, denoted by ∇ F ξ . Let x be in C ([0 , T ]; R ) , that is thespace of R -valued continuous functions of [0 , T ] , and set for any ≤ t, s ≤ TB t,xs := x ( s ) s ≤ t + ( x t + B s − B t ) s ≥ t , where x ( s ) denotes the path of x up to time s . Then, ∇ F B s = 1 for any s ∈ [0 , T ] . Wethus obtain ∇ F ξ = (cid:82) T ∇ F B s ds = T . The relation between ∇ F ξ and D r ξ is not clearand we can not hope to obtain a decomposition as [28, Lemma 2.4] for path-dependentBSDEs using the same method.An other approach to study the Z component could consist in studying the path-dependent PDE associated with the path-dependent BSDE, see e.g. [41, 14, 15, 43].Indeed, it is proved, in the latter papers, that the Z component of the solution to apath-dependent BSDE can be expressed through the Dupire derivative of the solutionto a path-dependent PDE. It will be then interesting to take advantage of this relationtogether with the lifting theorem [10, Theorem 6.1] to study the Z component.Notice nevertheless that in the biological example proposed in Section 6, only the exis-tence of a density for the law of the Y component is relevant to validate the proposedmodel. In the examples in Finance proposed in Section 7, the model of pricing studiedwill be reduced to solve BSDE (5.1), hence we will prove that both the law of Y t and thelaw of Z t have densities with respect to the Lebesgue measure. Stochastic models predicting mRNA and proteins fluctuations were introduced duringthe 70’s (see e.g. [44]). It has become during this last decades a prolific field in thestudying of proteins synthesis known as the "gene expression noise". This section being amathematical study of a biological problem, we consider one active gene which synthesisesone protein and we give a very simplified explanation of the proteins degradation proceed,by focusing on the main step of the mechanism. For more details, see for instance [40].
Step 1: Transcription.
The first step of the synthesis of the protein consists in thetranscription of a gene, made of a piece of DNA, into mRNA. The synthesis of mRNAis catalysed by an enzyme, the RNA polymerase whose the activation rate is denoted by R . Step 2: Translation.
In this step, the mRNA, previously synthesised, is decoded bya ribosome. A transfer RNA brings amino acids to the ribosome to produce an aminoacid chain using the genetic code. The degradation rate of mRNA is denoted by ρ . Atthe end of this step, the protein is synthesised.Here, we assume that the present protein concentration is known and we want to studythe previous protein concentrations which lead to the one observed. As an illustration f this phenomenon, we consider for instance a necrotic cells model, in which we want tocontrol the initial protein concentration. It was showed in [46] that this problem can bereduced to solve the following BSDE Y t = ξ + (cid:90) Tt ( f ( Y s ) − ρY s ) ds − (cid:90) Tt Z s dW s , (6.1)where Y t is the protein concentration at time t , ξ is the terminal protein concentra-tion, which is typically the observed data in a necrotic model, and f is the degrada-tion/syntetization rate of the protein depending on R , ρ and a positive constant a . Inthis study, following [46] we assume that f is the Hill function of the protein with coef-ficient , i.e. f ( Y s ) := R aY s aY s . In biochemistry, f quantifies the fraction of the ligand-binding sites on the receptor pro-tein. The Hill coefficient is , and describes cooperativeness effects. In order to validatetheir model, the authors of [46] need to compare the law of the protein concentration attime t obtained by solving BSDE (6.1) with the data produced by Gillespie Method (see[20]). However, in [46], the authors assumed implicitly that Y t admits a density.We propose in this section to apply the results of Section 4.2 to study mathematicallythe model proposed in [46] when ξ := c + W T , with the Malliavin calculus. It can beseen as a mathematical strengthening of the model developed in [46] by using Nourdinand Viens’ Formula to obtain estimates of the density. Proposition 6.1.
Let ( Y, Z ) be the unique solution of BSDE (6.1) . Assume that ξ ≥ , P − a.s. (resp. ξ ≤ , P − a.s. ), then, for any t ∈ [0 , T ] , Y t ≥ , P − a.s. (resp. Y t ≤ , P − a.s. ).Proof. We reproduce here the linearisation method for BSDE introduced in [17] for BSDE(6.1). Y t = ξ + (cid:90) Tt (cid:18) R aY s aY s − ρ (cid:19) Y s ds − (cid:90) Tt Z s dW s , hence, by setting X t := Y t e (cid:82) t (cid:18) R aYs aY s − ρ (cid:19) ds , we obtain from Ito’s Formula, dX t = dY t e (cid:82) t (cid:18) R aYs aY s − ρ (cid:19) ds + Y t e (cid:82) t (cid:18) R aYs aY s − ρ (cid:19) ds (cid:18) R aY t aY t − ρ (cid:19) dt = − Y t e (cid:82) t (cid:18) R aYs aY s − ρ (cid:19) ds (cid:18) R aY t aY t − ρ (cid:19) dt + Z t e (cid:82) t (cid:18) R aYs aY s − ρ (cid:19) ds dW t + Y t e (cid:82) t (cid:18) R aYs aY s − ρ (cid:19) ds (cid:18) R aY t aY t − ρ (cid:19) dt Thus, Y t = E t (cid:34) ξe (cid:82) Tt (cid:18) R aYs aY s − ρ (cid:19) ds (cid:35) , whose sign is fully determined by the sign of ξ . We extend in this section the model introduced in [46]. We assume that
R, ρ are tworeal constants and that ξ satisfied the following assumption ξ is a Gaussian F T -measurable random variable whose mean is denoted by c andvariance is denoted by σ . • ξ is in D , and there exist < k ≤ k such that for any r ∈ [0 , T ] , < k ≤ D r ξ ≤ k .According to Theorem 4.2 and Theorem 4.3 above, BSDE (6.1) admits a unique solution ( Y, Z ) such that for any t ∈ [0 , T ] , Y t ∈ D , and Z ∈ L ([ t, T ]; D , ) . We then have thefollowing proposition. Proposition 6.2.
The first component Y of the solution of BSDE (6.1) admits a densitydenoted by ρ Y t at any time t ∈ (0 , T ] . Besides, ρ Y t has Gaussian estimates, satisfyingthe following inequalities for any x ∈ R f i ( x ) ≤ ρ Y t ( x ) ≤ f s ( x ) , (6.2) where f i ( x ) = C Y t k t e − C a,R,ρ ( T − t ) e − e − Ca,R,ρ ( T − t ) ( x − E [ Yt ])22 k t ,f s ( x ) = C Y t k t e − C a,R,ρ ( T − t ) e − e − Ca,R,ρ ( T − t ) ( x − E [ Yt ])22 k t , and with C Y t := E [ | Y t − E [ Y t ] | ]2 ,C a,R,ρ := 98 R (cid:114) a − ρ,C a,R,ρ := − R (cid:114) a − ρ. Proof.
Let ( Y, Z ) be the unique solution of (6.1). We deduce from Theorem 4.4 thatfor any t ∈ (0 , T ] , the law of Y t admits a density denoted by ρ Y t . Recall that ( DY, DZ ) satisfies the following linear BSDE D u Y t = D r ξ + (cid:90) Tt R aY s D u Y s (1 + aY s ) − ρD u Y s ds − (cid:90) Tt D u Z s dW s , ≤ u ≤ t ≤ T, P − a.s. By linearisation, we thus obtain D u Y t = E t (cid:34) D u ξe (cid:82) Tt (cid:18) R aYs (1+ aY s )2 − ρ (cid:19) ds (cid:35) . Notice that C a,R,ρ := R (cid:112) a − ρ is the maximum of y (cid:55)−→ R ay (1+ ay ) − ρ and C a,R,ρ := − R (cid:112) a − ρ is its minimum. Hence, for any ≤ u ≤ t ≤ T , ke C a,R,ρ ( T − t ) ≤ D u Y t ≤ ke C a,R,ρ ( T − t ) . Using the definition (2.5) of g Y t , one get for any t ∈ (0 , T ] | k | te C a,R,ρ ( T − t ) ≤ g Y t ( x ) ≤ | k | te C a,R,ρ ( T − t ) , x ∈ R . Thus, according to Theorem 2.4, Relation (6.2) holds.
To validate the method proposed in [46], we have to analyse how close the law of Y t forany t ∈ (0 , T ] is to Gaussian distributions produced by the Gillepsie method (see [46,Section III]). Notice that in [46] the law of Y t is emphasised through a distribution fittingand is not proved rigorously. We propose in this section a more accurate proof in orderto validate the Shamarova-Ramos-Aguiar model. .3.1 Law of Y t by using statistical tests Let ( Y, Z ) be the unique solution to BSDE (6.1) where ξ has a normal distribution. In[46], the authors study their model by assuming that Y t has a normal distribution andcompare the first and second order moments of Y t with those generated by a benchmarkrandom variable, which has a normal distribution. However, it is not clear that thelaw of Y t is normal. Nevertheless, from a statistical point of view, we could validate thisassumption by using a statistical hypothesis test. In this subsection, we set the statisticalhypothesis (H) " Y t has a normal distribution"and we first test it using a Jarque-Bera test with the data of [46, A. Self-regulating gene].Recall that the Jarque-Bera test consists in computing the sample skewness, denotingby S , and the sample kurtosis, denoting by K , of a sample data, such that S := M (cid:80) Mi =1 ( Y it − Y t ) (cid:16) M (cid:80) Mi =1 | Y it − Y t | (cid:17) , K := M (cid:80) Mi =1 ( Y it − Y t ) (cid:16) M (cid:80) Mi =1 | Y it − Y t | (cid:17) , where M denotes the size of the sample, Y it is the ith observated data and Y t is thearithmetical mean of the data. We then define the Jarque-Bera variable denoted by JB ,by the following formula JB := M (cid:18) S K − (cid:19) . Under ( H ) , the law of JB is a chi-squared distribution with two degrees of freedom.Hence, by choosing a risk level α = 5% , the critical region is JB > . , that is to say if JB > . we reject ( H ) . We refer to [23] for more details on this method.We apply this test to Y t , with the data of [46, A. Self-regulating gene]: M = 5000 , R =1 , ρ = 0 . , T = 400 . The results are given in Table 1. Table 1: A Jarque-Bera test for Hypothesis ( H ) with the data of [46, A. Self-regulatinggene]. Time t J B ( H )400 Interpretation
A Jarque-Bera test does not accept the assumption ( H ) with a risklevel α = 0 . . Hence, from a statistical point of view, it is not clear that Y t has agaussian law. The problem comes from the number of simulations which has to be high.We now choose a number of simulation more relevant, by taking M = 100000 . Weuse a Jarque-Bera test together with a Kolmogorov-Smirnov test in order to validatesatistically the model developed in [46]. Recall that if we have a sample ( Y it ) ≤ i ≤ M ofobserved data, we set KS the Kolmogorov-Smirnov statistic corresponding to the sample,defined by KS := √ M sup x { F M ( x ) − F ( x ) } , where F M is the empirical distribution function of the sample of observed data and F is the cumulative distribution function of a normal law with parameters the arithmetic ean and the variance of the sample. Hence, for a level α = 0 . , by using a Kolmogorov-Smirnov test, we reject the Hypothesis ( H ) as soon as KS > . . The results arepresented in Table 2. Table 2: A Jarque-Bera test and a Kolmogorov-Smirnov test for Hypothesis ( H ) with thedata of [46, A. Self-regulating gene] and M = 100000 . We write "Not R." for "not rejected".Statistical tests Time t 400 300 200 100 50Jarque-Bera test J B ( H ) Not R. Not R. Not R. Not R. Not R.Kolmogorov-Smirnoff test KS ( H ) Not R. Not R. Not R. Not R. Not R.
Interpretation
A Jarque-Bera test together with a Kolmogorov-Smirnov test cannotinvalidate the assumption ( H ) with a risk level α = 0 . . Hence, from a statistical pointof view, the model developed in [46] seems to be relevant. However, we propose in thenext section a pure mathematical analyse of this model, by using the Malliavin calculusand by applying results of [29] together with those obtained in Section 4. Assume that ξ = c + σ W T . Then, we can use the result of Section 6.2 and we deducethat BSDE (6.1) admits a unique solution ( Y, Z ) such that for any t ∈ [0 , T ] , Y t ∈ D , and Z ∈ L ([ t, T ]; D , ) . Besides, according to Proposition 6.2, for any t ∈ [0 , T ] , Y t admits a density with respect to the Lebesgue measure denoted by ρ Y t . such that ρ Y t has Gaussian estimates, satisfying the following inequalities for any x ∈ R f i ( x ) ≤ ρ Y t ( x ) ≤ f s ( x ) , where f i ( x ) = C Y t σ t e − C a,R,ρ ( T − t ) e − e − Ca,R,ρ ( T − t ) ( x − E [ Yt ])22 tσ ,f s ( x ) = C Y t σ t e − C a,R,ρ ( T − t ) e − e − Ca,R,ρ ( T − t ) ( x − E [ Yt ])22 tσ . We illustrate these results in Figure 1.
T, a, c, σ = 1 , R = 1 , ρ = 0 . , and 500 000 simulations using a method ofMonte-Carlo (see [6] for instance) to compute the solution of BSDE (6.1). We represent ρ Y t for t = 0 . , . , . , . . We provide in red (resp. in blue) the supremum bound "fs" of ρ (resp. the infimum "fi"), using Nourdin and Viens’ Formula. Time t E [ Y t ] [ Y t ] Interpretation
The closer t is to T , the better the approximation is using Proposition6.2. Besides, this method guarantees Gaussian tails to control extreme events which isfundamental to validate the model developed in [46] by comparing the obtained datawith those induced by Gillepsie Method (see [46] and [20] for more details).Notice finally that the variance of Y t seems to be a decreasing function of the time. Thisis not surprising since Y is deterministic. We now propose to extend the model developed by Shamarova, Ramos and Aguiar(see the previous Example 1) to the non-Markovian setting. This extension might bequite relevant when we study the synthesis of protein in some kind of cells (see forinstance [7, 27, 18]). Assume that there exist α ∈ R , β > and γ ≥ such that ξ = α + βW T + γ (cid:82) T W s ds. Hence, BSDE (6.1)becomes: Y t = α + βW T + γ (cid:90) T W s ds + (cid:90) Tt (cid:18) R aY s aY s − ρY s (cid:19) ds − (cid:90) Tt Z s dW s , (6.3)According to Theorem 4.2 and Theorem 4.3, BSDE (6.3) admits a unique solution ( Y, Z ) such that for any t ∈ [0 , T ] , Y t ∈ D , and Z ∈ L ([ t, T ]; D , ) . According to Proposition .2, for any t ∈ [0 , T ] , Y t admits a density with respect to the Lebesgue measure denotedby ρ Y t such that ρ Y t has Gaussian estimates, satisfying the following inequalities for any x ∈ R f i ( x ) ≤ ρ Y t ( x ) ≤ f s ( x ) , (6.4)where f i ( x ) = C Y t ( β + γT ) t e − C a,R,ρ ( T − t ) e − e − Ca,R,ρ ( T − t ) ( x − E [ Yt ])22 β t ,f s ( x ) = C Y t β t e − C a,R,ρ ( T − t ) e − e − Ca,R,ρ ( T − t ) ( x − E [ Yt ])22( β + γT )2 t . The problem of pricing in finance using BSDE was first developed in [17]. Considera financial market in which an agent invests in a riskless asset, denoted by S , whosethe dynamics is given by the short rate of the market, denoted by r , and a risky asset,denoted by S , whose the dynamic is given through a predictable process, called the riskpremium and denoted by θ . Let now ξ be a contingent claim. The classical pricingproblem consists in finding an hedging strategy Z and a price y such that the terminalwealth of the agent is ξ . It was showed in [17] that this pricing problem can be reducedto solve the following stochastic linear BSDE, when S is a geometric Brownian motion dY t = ( r t Y t + θ t Z t ) dt + Z t dW t , Y T = ξ. (7.1)More generally, we set the following assumption, which enlarge the range of possibleapplications to this study ( S ) Let an asset S such that for any F T -measurable square integrable random variable ξ , the pricing problem can be reduced to study BSDE (7.1), where the process θ depends on the dynamic of S . Remark 7.1.
We provide in this remark two classical examples of process S satisfyingthe previous Assumption ( S ) . (aB) Assume that the asset S is an arithmetic Brownian motion, with the followingdynamic dS t = b t dt + σ t dW t , S = x ∈ R , where b and σ are F -predictable processes with σ t > , P − a.s. . Given an F T -measurablesquare integrable random variable ξ , using the self-financing Property, one can easilyshow that the corresponding pricing problem can be reduced to solve BSDE (7.1) with θ := b − rσ . In this case, the process Y provides the value of the problem and the process Z/σ gives the optimal number of asset owned at time t to solve the pricing problem. (gB) Assume that the dynamic of the asset S is given by dS t = b t S t dt + σ t S t dW t , S = x ∈ R , where b and σ are F -predictable processes with σ t > , P − a.s. Given an F T -measurablesquare integrable random variable ξ , it was showed in [17] that the corresponding pricingproblem can be reduced to solve BSDE (7.1) with θ := b − rσ . In this case, the process Y provides the value of the problem and the process Z/σ gives the optimal quantity ofmoney invested in the risky asset to solve the pricing problem.
Most of models assume that r is bounded to simplify the study. However, as noticed in[16], the assumption on the boundedness of the short rate r rarely holds in a market. Inthis section, we investigate the existence of densities for the laws of the components of he solution to (7.1). In this model, Assumptions ( A ) and ( A ) ( i ) above in Section3.4 become (H1) For any p > , E (cid:104) e p (cid:82) T r s ds (cid:105) < + ∞ and (cid:0)(cid:82) · θ t dW t (cid:1) is a BMO-martingale.We thus have the following lemma Lemma 7.1.
Assume that ( H1 ) holds and that for any p > , E [ | ξ | p ] < + ∞ . Then,BSDE (7.1) admits a unique solution ( Y, Z ) ∈ S p × H p for any p > . Besides, if ξ ≥ , P − a.s. (resp. ξ ≤ , P − a.s. ), then for any t ∈ [0 , T ] , Y t ≥ , P − a.s. (resp. ξ ≤ , P − a.s. )Proof. The proof of the existence of a unique solution ( Y, Z ) in ∈ S p ×H p is a consequenceof Theorem 3.5. Using a linearisation, we get Y t = E Q t (cid:104) ξe − (cid:82) Tt r s ds (cid:105) , where d Q d P := e − (cid:82) T θ s dW s − (cid:82) T | θ s | ds . Thus, we notice that the sign of the Y process is given by the sign of ξ . Let a, b ≥ and (cid:36) > . Assume that the rate of the market r is the solution of thefollowing SDE. dr t = a ( b − r t ) dt + (cid:36)dW t , r ∈ R . (7.2) Lemma 7.2.
Let r := ( r t ) t ∈ [0 ,T ] be the solution to SDE (7.2) . Then, for any p > , q ≥ and for any t ∈ [0 , T ] , r t ∈ D q,p . Besides, for any ≤ u ≤ t , D u r t = (cid:36) ≥ , P − a.s. and for any q > , D q r t = 0 , P − a.s. .Proof. Let r := ( r t ) t ∈ [0 ,T ] be the solution to SDE (7.2). Notice that r t is Malliavindifferentiable (see e.g. [35, Theorem 2.2.1]). Besides, as an Ornstein-Uhlenbeck process, r t can be computed explicitly r t = r e − at + b (1 − e − at ) + (cid:36)e − at (cid:90) t e as dW s . (7.3)Taking the Malliavin derivative, one obtains directly that for any r t ∈ D q,p for any p > , q ≥ . Besides for any ≤ u ≤ t ≤ T , D u r t = (cid:36) ≥ , P − a.s. and for any q > , D q r t = 0 , P − a.s. Since we aim at applying Bouleau and Hirsch Criterion (see Theorem 2.3), we first showthat the components Y t and Z t of the solution to BSDE (7.1) are Malliavin differentiable.In this section we will work under Assumption ( Θ ) (see Section 5.1) since we aim atapplying the results of Section 5.1 to investigate the existence of densities for both the Y t and the Z t components. Although this assumption is really restrictive, we can notdo better as explained in Remark 4.1. However, for the following result dealing withthe Malliavin differentiability of Y t and Z t , one could make weaker Assumption ( Θ ) byconsidering that Assumption ( A ) holds. Proposition 7.1.
Let ξ ∈ D ,p for any p > . Let r be the unique solution to SDE (7.2) and θ satisfying Assumption ( Θ ) . Then, BSDE (7.6) admits a unique solution ( Y, Z ) ∈ S p × H p for any p > . Besides, for any p > and t ∈ [0 , T ] , Y t ∈ D ,p and Z ∈ L ([ t, T ]; D ,p ) .Proof. By noticing that Assumptions ( A ) , ( A ) , ( DA ) and ( DA ) hold and by apply-ing Theorem 3.6, we deduce that BSDE (7.6) admits a unique solution ( Y, Z ) ∈ S p × H p for any p > and that if t ∈ [0 , T ] , Y t ∈ D ,p and Z ∈ L ([ t, T ]; D ,p ) for any p > . n this particular model and as said in Section 5, we provide now conditions on ξ and itsMalliavin derivatives ensuring existence of densities for both the law of the Y t componentand for the law of the Z t component of the solution to BSDE (7.1). Theorem 7.1.
Let ξ ∈ D ,p for any p > . Assume that ( Θ ) holds and that one of thefollowing two assumptions is satisfied for A ⊂ Ω such that P ( A ) > ( ξ +) ξ ≥ , D u ξ ≤ , λ ( du ) − a.e., D u ξ < , λ ( du ) − a.e. on A, ( ξ − ) ξ ≤ , D u ξ ≥ , λ ( du ) − a.e., D u ξ > , λ ( du ) − a.e. on A, then for any t ∈ (0 , T ] , the law of Y t is absolutely continuous with respect to the Lebesguemeasure.Assume now that ξ ∈ D ,p for any p > and assume in addition to ( ξ +) that D v ( D u ξ ) ≥ , ( λ ⊗ λ )( du, dv ) − a.e., D v ( D u ξ ) > , ( λ ⊗ λ )( du, dv ) − a.e. on A (7.4) or in addition to ( ξ − ) that D v ( D u ξ ) ≤ , ( λ ⊗ λ )( du, dv ) − a.e., D v ( D u ξ ) < , ( λ ⊗ λ )( du, dv ) − a.e. on A, (7.5) then the law of Z t has a density with respect to the Lebesgue measure.Proof. We denote by ( Y, Z ) the unique solution in S p × H p for any p > of BSDE (7.1)with for any p > and t ∈ [0 , T ] , Y t ∈ D ,p and Z ∈ L ([ t, T ]; D ,p ) by using Proposition7.1. Let Assumption ( ξ +) be true. Then, according to Theorem 4.5 together withLemmas 7.1 and 7.2, for any t ∈ (0 , T ] the law of Y t has a density with respect tothe Lebesgue measure. Recall that under Assumption ( ξ +) , according to Remark 4.2,we have for any t ∈ (0 , T ] , D u Y t ≤ , λ ( du ) ⊗ P − a.e. By assuming moreover thatConditions (7.4) holds and by applying Theorem 5.1 with ˜ f ( t, Y t ) := − r t Y t , we deducethat ( DZ +) , ( DY − ) and ( f − ) hold. Then D v Z t > for any ≤ v < t ≤ T , P -almostsurely. Thus, according to Theorem 2.3, the law of Z t has a density with respect toLebesgue measure for any t ∈ (0 , T ] .The proof under ( ξ − ) is similar as a consequence of Theorem 5.1 by showing thatAssumptions ( DZ − ) , ( DY +) and ( f − ) hold. In this section, we investigate pricing problems of Asian options, i.e. where the liability ξ is a function of the mean of the risky asset S . We assume that Assumption ( S ) holds,thus the pricing problem is reduced to solve the affine non-Markovian BSDE Y t = ξ − (cid:90) Tt ( r s Y s + θ s Z s ) ds − (cid:90) Tt Z s dW s , ξ = f (cid:32)(cid:90) T g ( W s ) ds (cid:33) , (7.6)where f, g are two continuous maps from R into R . Proposition 7.2.
Assume that ( Θ ) hold. Let r be the unique solution to SDE (7.2) .Assume moreover that f, g are twice differentiable λ ( dx ) -a.e. and one of the followingassumption is satisfied ( A +)( i ) f ≥ , f (cid:48) ≥ , g (cid:48) ≤ and f (cid:48) > , g (cid:48) < on a set A with positive Lebesgue measure, ( ii ) moreover f ” ≥ , g (cid:48)(cid:48) ≥ , and f (cid:48)(cid:48) > or g (cid:48)(cid:48) > on A, ( A +)( i ) f ≥ , f (cid:48) ≤ , g (cid:48) ≥ and f (cid:48) < , g (cid:48) > on a set A with positive Lebesgue measure, ( ii ) moreover f ” ≥ , g (cid:48)(cid:48) ≤ , and f (cid:48)(cid:48) > or g (cid:48)(cid:48) < on A, A − )( i ) f ≤ , f (cid:48) ≥ , g (cid:48) ≥ and f (cid:48) > , g (cid:48) > on a set A with positive Lebesgue measure, ( ii ) moreover f ” ≤ , g (cid:48)(cid:48) ≤ , and f (cid:48)(cid:48) < or g (cid:48)(cid:48) < on A. ( A − )( i ) f ≤ , f (cid:48) ≤ , g (cid:48) ≤ and f (cid:48) < , g (cid:48) < on a set A with positive Lebesgue ( ii ) moreover f ” ≤ , g (cid:48)(cid:48) ≥ , and f (cid:48)(cid:48) < or g (cid:48)(cid:48) > on A.Then, by denoting ( Y, Z ) the unique solution of BSDE (7.6) , for any t ∈ (0 , T ] both thelaw of Y t and the law of Z t are absolutely continuous with respect to Lebesgue measure.Proof. Notice that for any ≤ u ≤ T we have D u ξ = f (cid:48) (cid:32)(cid:90) T g ( W s ) ds (cid:33) (cid:90) Tu g (cid:48) ( W s ) ds, and for any ≤ v ≤ T we have D v ( D u ξ ) = f (cid:48)(cid:48) (cid:32)(cid:90) T g ( W s ) ds (cid:33) (cid:90) Tu g (cid:48) ( W s ) ds (cid:90) Tv g (cid:48) ( W s ) ds + f (cid:48) (cid:32)(cid:90) T g ( W s ) ds (cid:33) (cid:90) Tu ∧ v g (cid:48)(cid:48) ( W s ) ds. Thus, by noticing that Assumption ( A +) ( i ) or ( A +) ( i ) (resp. Assumption ( A − )( i ) or ( A − ) ( i ) ) ensure that ( ξ +) (resp. ( ξ − ) ) is true, we deduce from the first part ofTheorem 7.1 above that the law of Y t has a density with respect to Lebesgue’s measurefor any t ∈ (0 , T ] . Moreover, if Assumption ( A +) ( ii ) or ( A +) ( ii ) (resp. Assumption ( A − ) ( ii ) or ( A − ) ( ii ) ) holds, then Condition (7.4) is satisfied (resp. (7.5)). Byapplying Theorem 7.1 we deduce that Y t and Z t have absolutely continuous law withrespect to Lebesgue measure. In this section, we aim at applying Theorem 7.1 to lookback options. Let Assumption ( S ) be true. Set M := ( M t ) t ∈ [0 ,T ] , where M t = sup s ∈ [0 ,t ] W s . The following lemma is adirect consequence of [21, Lemma 1.1], [24, Remark 8.16 and Problem 8.17]. Lemma 7.3. M t ∈ D , and D r M t = r ≤ τ t , where τ t is almost surely unique such thatdefined by W τ t = M t . More precisely, for any ≤ s ≤ t , P ( τ t ≤ s ) = 2 π arcsin (cid:114) st . We consider the affine non-Markovian BSDE Y t = ξ − (cid:90) Tt ( r s Y s + θ s Z s ) ds − (cid:90) Tt Z s dW s , ξ = f ( M T ) , (7.7)where f is a continuous mapping from R into R . We have the following propositionwhich is a consequence of Lemma 7.3 together with Theorem 4.5. Proposition 7.3.
Let Y be the first component of the solution of BSDE (7.7) , hence forany t ∈ (0 , T ] , if f is differentiable λ ( dx ) -a.e. and one of the following two assumptionsis satisfied ( lb +) f ≥ and f (cid:48) ≤ and f (cid:48) < on a set with positive Lebesgue measure, ( lb − ) f ≤ and f (cid:48) ≥ and f (cid:48) > on a set with positive Lebesgue measure, hen the law of Y t is absolutely continuous with respect to the Lebesgue measure. Remark 7.2.
Since ξ := M T is not twice Malliavin differentiable (see [21]), i.e. ξ doesnot belong to D ,p whatever p ≥ , we cannot reproduce the proof of Proposition 7.2 tostudy the problem of existence of density for Z t . Acknowledgments
The author thanks Dylan Possamaï and Anthony Réveillac for conversations and preciousadvice in the writing of this paper. The author is grateful to Région Ile-De-France forfinancial support.
References [1] O. Aboura and S. Bourguin. Density estimates for solutions to one dimensionalBackward SDE’s.
Potential Analysis , 38(2):573–587, 2013.[2] S. Ankirchner, P. Imkeller, and G. Dos Reis. Classical and variational differen-tiability of BSDEs with quadratic growth.
Electron. J. Probab , 12(53):1418–1453,2007.[3] F. Antonelli and A. Kohatsu-Higa. Densities of one-dimensional backward SDEs.
Potential Analysis , 22(3):263–287, 2005.[4] C. Bender and M. Kohlmann. BSDEs with stochastic Lipschitz condition. Univer-sität Konstanz, Fakultät für Mathematik und Informatik, 2000.[5] JM. Bismut. Conjugate convex functions in optimal stochastic control.
Journal ofMathematical Analysis and Applications , 44(2):384–404, 1973.[6] B. Bouchard and N. Touzi. Discrete-time approximation and Monte-Carlo simu-lation of backward stochastic differential equations.
Stochastic Processes and theirapplications , 111(2):175–206, 2004.[7] D.A Bratsun, D.N Volfson, J. Hasty, and L.S Tsimring. Non-markovian processes ingene regulation (keynote address). In
SPIE Third International Symposium on Fluc-tuations and Noise , pages 210–219. International Society for Optics and Photonics,2005.[8] P. Briand and F. Confortola. BSDEs with stochastic Lipschitz condition andquadratic PDEs in Hilbert spaces.
Stochastic Processes and their Applications ,118(5):818–838, 2008.[9] Ph. Briand, B. Delyon, Y. Hu, E. Pardoux, and L. Stoica. L p solutions of BSDEswith stochastic Lipschitz condition. Stochastic Processes and their Applications ,108(1):109–129, 2003.[10] R. Cont and DA. Fournié. Functional Itô calculus and stochastic integral represen-tation of martingales.
The Annals of Probability , 41(1):109–133, 2013.[11] A. Debussche and M. Romito. Existence of densities for the 3d Navier-Stokes equa-tions driven by Gaussian noise.
Probability Theory and Related Fields , 158(3-4):575–596, 2014.[12] F. Delarue, S. Menozzi, and E. Nualart. The Landau Equation for Maxwellianmolecules and the Brownian Motion on SO N ( R ) . Electronic Journal of Probability ,20:1–39, 2014.[13] G. Dos Reis and Ricardo JN. Dos Reis. A note on comonotonicity and positivity ofthe control components of decoupled quadratic FBSDE.
Stochastics and Dynamics ,13(04):1350005, 2013.[14] I. Ekren, C. Keller, N. Touzi, and J. Zhang. On viscosity solutions of path dependentPDEs.
The Annals of Probability , 42(1):204–236, 2014.
15] I. Ekren, N. Touzi, and J. Zhang. Viscosity solutions of fully nonlinear parabolicpath dependent PDEs: Part i,(2012).
To appear in The Annals of probability, arXivpreprint arXiv:1210.0006 , 2014.[16] N. El Karoui and SJ. Huang. A general result of existence and uniqueness of back-ward stochastic differential equations.
Pitman Research Notes in Mathematics Se-ries , pages 27–38, 1997.[17] N. El Karoui, S. Peng, and MC. Quenez. Backward stochastic differential equationsin finance.
Mathematical finance , 7(1):1–71, 1997.[18] V. Fromion, E. Leoncini, and P. Robert. A stochastic model of the production ofmultiple proteins in cells. arXiv preprint arXiv:1411.1572 , 2014.[19] C. Geiss and A. Steinicke. Malliavin derivative of random functions and applicationsto lévy driven bsdes. arXiv preprint arXiv:1404.4477 , 2014.[20] DT. Gillespie. Exact stochastic simulation of coupled chemical reactions.
Thejournal of physical chemistry , 81(25):2340–2361, 1977.[21] E. Gobet and A. Kohatsu-Higa. Computation of Greeks for barrier and lookbackoptions using Malliavin calculus.
Electron. Comm. Probab , 8:51–62, 2003.[22] Y. Hu, P. Imkeller, and M. Müller. Utility maximization in incomplete markets.
The Annals of Applied Probability , 15(3):1691–1712, 2005.[23] C.M. Jarque and A.K. Bera. A test for normality of observations and regres-sion residuals.
International Statistical Review/Revue Internationale de Statistique ,pages 163–172, 1987.[24] I. Karatzas and S. Shreve.
Brownian motion and stochastic calculus , volume 113.Springer Science, 2012.[25] N. Kazamaki. Continuous exponential martingales and BMO.
Lecture notes inmathematics , 1994.[26] A. Kohatsu-Higa. Lower bounds for densities of uniformly elliptic random variableson Wiener space.
Probability theory and related fields , 126(3):421–457, 2003.[27] E. Leoncini.
Towards a global and systemic understanding of protein production inprokaryotes . PhD thesis, Ecole Polytechnique X, 2013.[28] J. Ma and J. Zhang. Representation theorems for backward stochastic differentialequations.
The Annals of Applied Probability , 12(4):1390–1418, 2002.[29] T. Mastrolia, D. Possamaï, and A. Réveillac. Density analysis of BSDEs.
To appearin Annals of Probability, preprint arXiv:1402.4416 , 2015.[30] T. Mastrolia, D. Possamaï, and A. Réveillac. On the Malliavin differentiabilityof BSDEs.
To appear in Annales de l’Institut Henri Poincaré, arXiv preprintarXiv:1404.1026 , 2015.[31] A. Millet and M. Sanz-Solé. A stochastic wave equation in two space dimension:smoothness of the law.
Annals of Probability , pages 803–844, 1999.[32] G. Mokobodzki. Ultrafiltres rapides sur N.
Construction d’une densité relative dedeux potentiels comparables , 1969.[33] C. Mueller and D. Nualart. Regularity of the density for the stochastic heat equation.
Electron. J. Probab , 13(74):2248–2258, 2008.[34] I. Nourdin and FG. Viens. Density formula and concentration inequalities withMalliavin calculus.
Electron. J. Probab , 14(78):2287–2309, 2009.[35] D. Nualart.
The Malliavin calculus and related topics . Probability and its Applica-tions (New York). Springer-Verlag, Berlin, second edition, 2006.[36] E. Nualart and L. Quer-Sardanyons. Gaussian estimates for the density of the non-linear stochastic heat equation in any space dimension.
Stochastic Processes andtheir Applications , 122(1):418–447, 2012.
37] M. Nutz. Pathwise construction of stochastic integrals.
Electron. Commun. Probab ,17(24):1–7, 2012.[38] E. Pardoux and S. Peng. Adapted solution of a backward stochastic differentialequation.
Systems & Control Letters , 14(1):55–61, 1990.[39] E. Pardoux and S. Peng. Backward stochastic differential equations and quasilinearparabolic partial differential equations. In
Stochastic partial differential equationsand their applications , pages 200–217. Springer, 1992.[40] J. Paulsson. Models of stochastic gene expression.
Physics of life reviews , 2(2):157–175, 2005.[41] S. Peng and F. Wang. BSDE, path-dependent PDE and nonlinear Feynman-Kacformula. arXiv preprint arXiv:1108.4317 , 2011.[42] D. Possamaï, X. Tan, and C. Zhou. Stochastic control for a class of nonlinear kernelsand applications. arXiv preprint arXiv:1510.08439 , 2015.[43] Z. Ren, N. Touzi, and J. Zhang. An overview of viscosity solutions of path-dependentPDEs. In
Stochastic Analysis and Applications 2014 , pages 397–453. Springer, 2014.[44] DR. Rigney and WC. Schieve. Stochastic model of linear, continuous protein syn-thesis in bacterial populations.
Journal of theoretical biology , 69(4):761–766, 1977.[45] R. Rouge and N. El Karoui. Pricing via utility maximization and entropy.
Mathe-matical Finance , 10(2):259–276, 2000.[46] E. Shamarova, AF. Ramos, and P. Aguiar. Backward SDE approach to modellingof gene expression. preprint arXiv:1308.6619 , 2013.[47] H. Sugita. On a characterization of the Sobolev spaces over an abstract Wienerspace.
Journal of Mathematics of Kyoto University , 25(4):717–725, 1985.[48] J. Wang, Q. Ran, and Q. Chen. L p solutions of BSDEs with stochastic Lipschitzcondition. International Journal of Stochastic Analysis , 2007., 2007.