Weak order in averaging principle for stochastic differential equations with jumps Hongbo Fu, Li Wan 7 1 Research Centerof Nonlinear Science, College of Mathematics and Computer Science, 0 Wuhan TextileUniversity,Wuhan, 430073, PR China 2 Jicheng Liu∗, Xianming Liu n a School of Mathematics and Statistics, Huazhong Universityof Science and Technology, J Wuhan, 430074, PR China 7 2 ] R P Abstract . h t This article deals with the weak error for averaging principle for two-time- a scale system of jump-diffusion stochastic differential equation. Under suitable m conditions, it is provedthat the rate of weak convergenceto the averagedeffec- [ tive dynamics is of order 1 via an asymptotic expansion approach. 1 Keywords: Jump-diffusion, averagingprinciple, invariant measure, weak v 3 convergence,asymptotic expansion. 8 MSC 2010: primary 60H15, secondary 70K70 9 7 0 1. Introduction . 1 0 Weconsideratwo-time-scalesystemsofjump-diffusionstochasticdifferential 7 equation in form of 1 : v dXǫ =a(Xǫ,Yǫ)dt+b(Xǫ)dB +c(Xǫ )dP , Xǫ =x (1.1) t t t t t t− t 0 i X 1 1 dYǫ = f(Xǫ,Yǫ)dt+ g(Xǫ,Yǫ)dW +h(Xǫ ,Yǫ )dNǫ, Yǫ =y,(1.2) r t ǫ t t √ǫ t t t t− t− t 0 a where Xǫ Rn,Yǫ Rm, the drift functions a(x,y) Rn,f(x,y) Rm, the diffusiontfu∈nctionstb(∈x) Rn×d1,c(x) Rn,g(x,y) R∈m×d2 and h(x∈,y) Rm. ∈ ∈ ∈ ∈ B and W are the vector of d ,d -dimensional independent Brownian motions t t 1 2 on a complete stochastic base (Ω, , ,P), respectively. P is a scalar simple t t F F Poisson process with intensity λ , and Nǫ is a scalar simple Poisson process 1 t ∗Correspondingauthor at: School ofMathematics andStatistics, Huazhong Universityof ScienceandTechnology, Wuhan, 430074, China. Email addresses: [email protected] (HongboFu),[email protected] (LiWan), [email protected] (JichengLiu),[email protected] (XianmingLiu) Preprint submitted toArxiv January 30, 2017 with intensity λ2. The positive parameter ǫ is small and describes the ratio ǫ of time scales between Xǫ and Yǫ. Systems (1.1)-(1.2) with two time scale t t occur frequently in applications including chemical kinetics, signal processing, complex fluids and financial engineering. With the two separationof time scale, we can view the state variable of the system as being divided into two parts: the “slow” variable Xǫ and the “fast” t variable Yǫ. It is often the case that we are interested only in the dynamics t of slow component. Then a simplified equation, which is independent of fast variable and possesses the essential features of the system, is highly desirable. Such a simplified equation is often constructed by averaging procedure as in [2, 20] for deterministic ordinary differential equations, as well as the further development[13,7,8,18,19,14,15,16]forstochasticdifferentialequationswith continuous Gaussian processes. As far as averaging for stochastic dynamical systemsininfinite dimensionalspaceis concerned,itis worthwhileto quote the important work of [4, 5, 6, 26] and also the work of [9, 10]. In order to derive the averaging dynamics of the system (1.1)-(1.2), we in- troduce the fast motionequationwith a frozen slow componentx Rn in form ∈ of dYx =f(x,Yx)dt+g(x,Yx)dW +h(x,Yx )dN , Yx =y, (1.3) t t t t t− t 0 whose solution is denoted by Yǫ(y). Under suitable conditions on f,g and h, t Yǫ(y) induces a unique invariant measure µx(dy) on Rm, which is ergodic and t ensures the averagedequation: dX¯t =a¯(X¯t)dt+b(X¯t)dB˙t+c(X¯t−)dP˙t, X¯0 =x, where the averagingnonlinear is defined by setting a¯(x) = a(x,y)µx(dy) ZRm = lim Ea(x,Yx(y)). t→+∞ t In [12], it is shown that under the above conditions the slow motion Xǫ t convergesstronglytothesolutionX¯ oftheaboveaveragedequationwithjumps. t The order of convergence 1 in strong sense was provided in [17]. To the best 2 knowledge of the authors, the existing literature does not address the weak order in averaging principle for jump diffusion stochastic differential systems. Infact,itisfairtosaythattheweakconvergenceinstochasticaveragingtheory of systems driven by jump noise is not yet fully developed, even though recent years have witnessed some strong approximation results [1, 21, 22, 23, 24, 25]. In the present paper, we are interested in the rate of weak convergence of theaveragingdynamicstothetruesolutionofslowmotionXǫ. Namely,wewill t determine the order, with respect to timescale parameter ǫ, of weak deviation betweenoriginalsolutionofslowequationandthesolutionofthecorresponding averaged equation. The main technique we adapted is to finding an expansion with respect to ǫ of the solutions of the Kolmogorov equations associated with 2 the jump diffusion system. The solvability of the Poisson equation associated with the generatoroffrozenequationprovidesanexpressionfor the coefficients of the expansion. As a result, the boundedness for the coefficients of expansion canbeprovedbysmoothingeffectofthecorrespondingtransitionsemigroupsin the space of bounded and uniformly continuous functions, where some regular conditionsis neededondriftanddiffusionterm. Ourresultshowthatthe weak convergence rate to be 1 even when there are jump components in the system, whichis the mainnoveltyofthis work. We wouldlike to stressthat asymptotic method was first applied by Br´ehier [3] to an averaging result for stochastic reaction-diffusion equations in the case of Gaussian noise of additive type was includedonlyinthefastmotion. However,theextensionofthisargumentisnot straightforward. The method used in the proof of weak order in [3] is strictly relatedto the differentiability in time of averagingprocess. Therefore,once the noiseisintroducedintheslowequation,difficulties willariseandthe procedure becomes more complicated. Our result in this paper bridges such a gap, in which the slow and the fast motions are both perturbed by noise with jumps. The rest of the paper is structured as follows. Section 2 is devoted to no- tation, assumptions and summarize preliminary results. The ergodicity of fast process and the averagingdynamics of system with jumps is introduced in Sec- tion3. Thenthe mainresultofthis article,whichis derivedvia the asymptotic expansions and uniform error estimates, is presented in Section 4. In the fi- nal section, we state and prove technical lemmas to which we appeal in the preceding section. Throughout the paper, the letter C below with or without subscripts will denote generic positive constants independent of ǫ, whose value may change from one line to another. 2. Assumptions and preliminary results Foranyintegerd,thescalarproductandnormond dimensionalEuclidean − space Rd are denoted by , and Rd, respectively. For any integer k, · · Rd k·k wedenote byCk(Rd,R)th(cid:16)e sp(cid:17)aceofallk timesdifferentiable functions onRd, b − which have bounded uniformly continuous derivatives up to the k-th order. In what follows, we shall assume that the drift and diffusion coefficients arising in the system fulfill the following conditions. (A1)Themappingsa(x,y),b(x),c(x),f(x,y),g(x,y) andh(x,y)areofclass C2 and have bounded first and second derivatives. Moreover, we assume that a(x,y),b(x) and c(x) are bounded. (A2) There exists a constant α > 0 such that for any x Rn,y Rm it ∈ ∈ holds yTg(x,y)gT(x,y)y α y Rm. ≥ k k (A3) There exists a constant β > 0 such that for any y ,y Rm and 1 2 x Rn it holds ∈ ∈ y y ,f(x,y ) f(x,y )+λ (h(x,y ) h(x,y )) 1 2 1 2 2 1 2 − − − Rm (cid:16) (cid:17) 3 + g(x,y ) g(x,y ) 2 +λ h(x,y ) h(x,y )2 1 2 Rm 2 1 2 k − k | − | β y y 2 . 1 2 Rm ≤− k − k Remark 2.1. Notice that from (A1) it immediately follows that the following directional derivatives exist and are controlled: Dxa(x,y) k1 Rn L k1 Rn, k · k ≤ k k Dya(x,y) l1 Rn L l1 Rm, k · k ≤ k k kDx2xa(x,y)·(k1,k2)kRn ≤Lkk1kRnkk2kRn, kDy2ya(x,y)·(l1,l2)kRn ≤Lkl1kRmkl2kRm, where L is a constant independent of x,y,k ,k ,l and l . For differentiability 1 2 1 2 of mappings b,c,f,g and h we possess the analogous results. For examples, we have kDx2xb(x)·(k1,k2)kRn ≤Lkk1kRnkk2kRn, k1,k2 ∈Rn, kDy2yf(x,y)·(l1,l2)kRm ≤Lkl1kRmkl2kRmk, l1,l2 ∈Rm. As far as the assumption (A2) is concerned, it is a sort of non-degeneracy condition anditis assumedinorder tohavetheregularizingeffectoftheMarkov transition semigroup associated with the fast dynamics. Assumption (A3) is the dissipative condition which determine how the fast equation converges to its equilibrium state. As assumption (A1) holds, for any ǫ> 0 and any initial conditions x Rn and y Rm, system (1.1)-(1.2) admits a unique solution, which, in ord∈er to ∈ emphasizethedependenceontheinitialdata,isdenotedby(Xǫ(x,y),Yǫ(x,y)). t t Moreover the following lemma holds (for a proof see e.g. [17]). Lemma 2.1. Under the assumption (A1), (A2) and (A3), for any x Rn, y Rm and ǫ>0 we have ∈ ∈ EkXtǫ(x,y)k2Rn ≤CT(1+kxk2Rn +kyk2Rm), t∈[0,T] (2.1) and EkYtǫ(x,y)k2Rn ≤CT(1+kxk2Rn +kyk2Rm), t∈[0,T]. (2.2) 3. Frozen equation and averaging equation Fixing ǫ = 1, we consider the fast equation with frozen slow component x Rn, ∈ dYx(y)=f(x,Yx(y))dt+g(x,Yx(y))dW +h(x,Yx (y))dN , t t t t t− t (3.1) (Y0x =y. 4 Under assumption (A1)-(A3), such a problem has a unique solution, which satisfies [17]: E Yx(y) 2 C(1+ x 2 +e−βt y 2 ), t 0. (3.2) k t kRm ≤ k kRn k kRm ≥ Let Yx(y′) be the solution of problem (3.1) with initial value Yx = y′, the Itoˆ t 0 formula implies that for any t 0, ≥ E Yx(y) Yx(y′) 2 y y′ 2 e−βt. (3.3) k t − t kRm ≤k − kRm Moreover,as discussionin [17] and [12], equation (3.1) admits a unique ergodic invariant measure µx with finite 2 moments: − y 2 µx(dy) C(1+ x 2 ). (3.4) Rm Rn ZRmk k ≤ k k Then, by averaging the coefficient a with respect to the invariant measure µx, we can define an Rn-valued mapping a¯(x):= a(x,y)µx(dy),x Rn. ZRm ∈ Due to assumption (A1), it is easily to check that a¯(x) is 2-times differentiable with bounded derivatives and hence it is Lipschitz-continuous such that a¯(x1) a¯(x2) Rn C x1 x2 Rn, x1,x2 Rn. k − k ≤ k − k ∈ According to the invariantproperty of µx, (3.4) and assumption(A1), we have kEa(x,Ytx(y)−a¯(x)k2Rn = kZRmE a(x,Ytx(y)−a(x,Ytx(z) µx(dz)k2Rn (cid:0) (cid:1) E Yx(y) Yx(z) 2 µx(dz) ≤ ZRm k t − t kRm e−βt y z 2Rmµx(dz) ≤ ZRmk − k Ce−βt 1+ x 2 + y 2 . (3.5) Rn Rm ≤ k k k k (cid:0) (cid:1) Now we can introduce the effective dynamical system dX¯t(x)=a¯(X¯t(x))dt+b(X¯t(x))dB˙t+c(X¯t−(x))dP˙t, (3.6) (X¯0 =x. As the coefficients a¯,b and c are Lipschitz-continuous, this equation admits a unique solution such that E X¯t(x) 2Rn CT(1+ x 2Rn), t [0,T]. (3.7) k k ≤ k k ∈ 5 4. Asymptotic expansion WenowintroducetheKolmogorovoperatorassociatedwiththeslowmotion equation, with frozen fast component y Rm, by setting ∈ 1 Φ(x) = a(x,y),D Φ(x) + Tr D2 Φ(x) b(x)bT(x) L1 x Rn 2 xx · +(cid:16)λ (Φ(x+c(x))(cid:17) Φ(x)), Φ(cid:2) C2(Rn,R). (cid:3) 1 − ∈ b For any frozen slow component x Rm, the Kolmogorovoperator for equation ∈ (3.1) is given by 1 Ψ(y) = f(x,y),D Ψ(y) + Tr D2 Ψ(y) g(x,y)gT(x,y) L2 y Rm 2 yy · (cid:16)+λ (Ψ(y+h(x,y(cid:17))) Ψ(y)),(cid:2)Ψ C2(Rm,R). (cid:3) 2 − ∈ b Let φ C2(Rn,R) and define a function uǫ(t,x,y):[0,T] Rn Rm R by ∈ b × × → uǫ(t,x,y)=Eφ(Xǫ(x,y)). t If we set 1 ǫ := + , L L1 ǫL2 we have that uǫ(t,x,y) is the solutio of the Kolmogorovequation: ∂ uǫ(t,x,y)= ǫuǫ(t,x,y), ∂t L (4.1) (uǫ(0,x,y)=φ(x), Also recall the Kolmogorov operator associated with the averaged equation is defined as 1 ¯Φ(x) = a¯(x),D Φ(x) + Tr D2 Φ(x) b(x)bT(x) L x Rn 2 xx · +(cid:16)λ (Φ(x+c(x(cid:17),y)) Φ(x)(cid:2)), Φ C2(Rn,R). (cid:3) 1 − ∈ b If we set u¯(t,x)=Eφ(X¯ (x)), t we have ∂ u¯(t,x)= ¯u¯(t,x), ∂t L (4.2) (u¯(0,x)=φ(x). Then the weak difference at time T is equal to Eφ(X¯ (x)) Eφ(Xǫ(x,y))=u¯(T,x) uǫ(T,x,y). T − T − With the above assumptions and notation we have the following result, which is a direct consequence of Lemma 4.1, Lemma 4.2 and Lemma 4.5. 6 Theorem 4.1. Assume that x Rn and y Rm, Then, under assumptions (A1), (A2) and (A3), for any ∈any T > 0 a∈nd φ C3(Rn,R), there exists a ∈ b constant C such that T,φ,x,y Eφ(Xǫ(x,y)) Eφ(X¯ (x)) C ǫ. T − T ≤ T,φ,x,y (cid:12) (cid:12) As a consequence(cid:12), it can be claimed that the(cid:12)weak order in averaging principle for jump-diffusion stochastic systems is 1. Inwhatfollows,forthesakeofsimplicity,weoftenomitthetemporalvariable t and spatialvariables x and y. For example, for uǫ(t,x,y), we often write it as uǫ. Our aimis to seek matchedasymptotic expansionsfor the uǫ(T,x,y) of the form uǫ =u +ǫu +rǫ, (4.3) 0 1 where u and u are smooth functions which will be constructed below, and rǫ 0 1 is the remainder term. 4.1. The leading term Let us first determine the leading terms. Now, substituting (4.3) into (4.1) yields ∂u ∂u ∂rǫ 0 +ǫ 1 + = u +ǫ u + rǫ ∂t ∂t ∂t L1 0 L1 1 L1 1 1 + u + u + rǫ ǫL2 0 L2 1 ǫL2 By comparing coefficients of powers of ǫ, we obtain u =0, (4.4) 2 0 L ∂u 0 = u + u . (4.5) ∂t L1 0 L2 1 It follows from (4.4) that u is independent of y, which means 0 u (t,x,y)=u (t,x). 0 0 We also impose the initial condition u (0,x) = φ(x). Since µx is the invariant 0 measure of the Markov process with generator , we have 2 L u (t,x,y)µx(dy)=0. 2 1 ZRmL Thanks to (4.5), this yields ∂u ∂u 0(t,x) = 0(t,x)µx(dy) ∂t ZRm ∂t 7 = u (t,x)µx(dy) 1 0 ZRmL 1 = a(x,y),D u (t,x) µx(dy)+ Tr D2 u (t,x) b(x)bT(x) ZRm(cid:16) x 0 (cid:17)Rn 2 xx 0 · +λ (Φ(x+c(x)) Φ(x)) (cid:2) (cid:3) 1 − = ¯u (t,x), 0 L so that u and u¯ are controlled by the same evolutionary equation. By using a 0 uniqueness argument, such u has to coincide with the solution u¯ and we have 0 the following lemma: Lemma 4.1. Under assumption (A1), (A2) and (A3), for any x Rn, y Rm and T >0, we have u (T,x,y)=u¯(T,x). ∈ ∈ 0 4.2. Construction of u 1 Let us proceed to carry out the construction of u . Thanks to Lemma 4.1 1 and (4.2), the equation (4.5) can be rewritten as ¯u¯= u¯+ u , 1 2 1 L L L and hence we get an elliptic equation for u with form 1 u (t,x,y)= a¯(x) a(x,y),D u¯(t,x) := ρ(t,x,y), (4.6) 2 1 x L − Rn − (cid:16) (cid:17) where ρ is of class C2 with respect to y, with uniformly bounded derivative. Moreover,it satisfies for any t 0 and x Rn, ≥ ∈ ρ(t,x,y)µx(dy)=0. ZRm For any y Rm and s>0 we have ∈ ∂ ρ(t,x,y) = f(x,y),D [ ρ(t,x,y)] s y s ∂sP P Rm (cid:16) 1 (cid:17) + Tr D2 [ ρ(t,x,y)] g(x,y)gT(x,y) 2 yy Ps · +λ2( (cid:2)s[ρ(t,x,y+h(x,y))] s[ρ(t,x,y(cid:3))]), P −P here [ρ(t,x,y)]:=Eρ(t,x,Yx(y)) Ps s satisfying lim Eρ(t,x,Yx(y))= ρ(t,x,z)µx(dz)=0. (4.7) s→+∞ s ZRm Indeed, by the invariant property of µx and Lemma 5.1, we have Eρ(t,x,Yx(y)) ρ(t,x,z)µx(dz) s − (cid:12) ZH (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) 8 = E[ρ(t,x,Yx(y)) ρ(t,x,Yx(z))µx(dz)] (cid:12)ZRm s − s (cid:12) (cid:12) (cid:12) (cid:12) E a(x,Yx(z)) a(x,Yx(y)),D u¯(t,x)(cid:12) µx(dz) ≤Z(cid:12)Rm(cid:12) (cid:16) s − s x (cid:12)(cid:17)Rn(cid:12) ≤CZR(cid:12)(cid:12)mEkYsx(z)−Ysx(y)kRnµx(dz). (cid:12)(cid:12) This, in view of (3.3) and (3.4), yields Eρ(t,x,Yx(y)) ρ(t,x,z)µx(dz) s − (cid:12) ZH (cid:12) (cid:12)(cid:12) C(1+ x Rn + y Rm)e−β2s, (cid:12)(cid:12) ≤(cid:12) k k k k (cid:12) which implies the equality (4.7). Therefore, we get +∞ f(x,y),D [ ρ(t,x,y)]ds y s (cid:16) 1 Z0+∞ P (cid:17)Rm + Tr D2 [ ρ(t,x,y)] g(x,y)gT(x,y)ds 2 yy Ps · (cid:20) Z0 (cid:21) +∞ +∞ +λ [ρ(t,x,y+h(x,y))]ds [ρ(t,x,y)]ds 2 s s P − P (cid:18)Z0 Z0 (cid:19) +∞ d = [ρ(t,x,y)]ds s dsP Z0 = lim Eρ(t,x,Yx(y)) ρ(t,x,y) s→+∞ s − = ρ(t,x,y)µx(dy) ρ(t,x,y) − ZH = ρ(t,x,y), − which implies ( +∞ ρ(t,x,y)ds)= ρ(t,x,y). This means that the L2 0 Ps − R +∞ u (t,x,y):= Eρ(t,x,Yx(y))ds (4.8) 1 s Z0 is the solution of Poissonequation (4.6). Lemma 4.2. Under assumptions (A1), (A2) and (A3), for any x Rn, y Rm and T >0, we have ∈ ∈ u1(t,x,y) CT(1+ x Rn + y Rm), t [0,T]. (4.9) | |≤ k k k k ∈ Proof. As known from (4.8), we have +∞ u (t,x,y)= E a¯(x) a(x,Yx(y)),D u¯(t,x) ds. 1 Z0 (cid:16) − s x (cid:17)Rn 9 This implies that +∞ |u1(t,x,y)| ≤ ka¯(x)−E[a(x,Ysx(y))]kRn ·kDxu¯(t,x)kRnds. Z0 Then, due to Lemma 5.1 and (3.5), this yields: +∞ u1(t,x,y) CT(1+ x Rn + y Rm) e−β2sds | | ≤ k k k k Z0 CT(1+ x Rn + y Rm). ≤ k k k k 4.3. Determination of remainder rǫ Once u andu havebeen determined, we cancarryoutthe constructionof 0 1 the remainder rǫ. It is known that (∂ ǫ)uǫ =0, t −L which, together with (4.4) and (4.5), implies (∂ ǫ)rǫ = (∂ ǫ)u ǫ(∂ ǫ)u t t 0 t 1 −L − −L − −L 1 1 = (∂ )u ǫ(∂ )u − t− ǫL2−L1 0− t− ǫL2−L1 1 = ǫ( u ∂ u ). (4.10) 1 1 t 1 L − Inordertoestimatetheremaindertermrǫweneedthefollowingcruciallemmas. Lemma 4.3. Under assumptions (A1), (A2) and (A3), for any x Rn, y Rm and T >0, we have ∈ ∈ ∂u 1(t,x,y) CT(1+ x Rn + y Rm) ∂t ≤ k k k k (cid:12) (cid:12) (cid:12) (cid:12) Proof. According(cid:12)to (4.8), we(cid:12)have (cid:12) (cid:12) ∂u +∞ ∂ 1(t,x,y)= E a¯(x) a(x,Yx(y)), D u¯(t,x) ds. ∂t − s ∂t x Z0 (cid:18) (cid:19)Rn By Lemma 5.6 introduced in Section 5, we have ∂u +∞ ∂ ∂t1(t,x,y) ≤ E ka¯(x)−a(x,Ysx(y))kRn ·k∂tDxu¯(t,x)kRn ds (cid:12) (cid:12) Z0 (cid:18) (cid:19) (cid:12) (cid:12) +∞ (cid:12)(cid:12) (cid:12)(cid:12) ≤ CT Eka¯(x)−a(x,Ysx(y))kRnds Z0 so that from (3.5) we have ∂u 1(t,x,y) CT(1+ x Rn + y Rm). ∂t ≤ k k k k (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) (cid:12) 10