Penalization methods for the Skorokhod problem and reflecting SDEs with jumps

# Penalization methods for the Skorokhod problem and reflecting SDEs with jumps

\fnmsWeronika \snmŁaukajtys\thanksrefe1label=e1 [    mark]vera@mat.uni.torun.pl    \fnmsLeszek \snmSłomiński\corref\thanksrefe2label=e2 [    mark]leszeks@mat.uni.torun.pl Faculty of Mathematics and Computer Science, Nicolaus Copernicus University, ul. Chopina 12/18 87–100 Toruń, Poland. \printeade1; \printead*e2
\smonth5 \syear2011\smonth10 \syear2011
\smonth5 \syear2011\smonth10 \syear2011
\smonth5 \syear2011\smonth10 \syear2011
###### Abstract

We study the problem of approximation of solutions of the Skorokhod problem and reflecting stochastic differential equations (SDEs) with jumps by sequences of solutions of equations with penalization terms. Applications to discrete approximation of weak and strong solutions of reflecting SDEs are given. Our proofs are based on new estimates for solutions of equations with penalization terms and the theory of convergence in the Jakubowski -topology.

\kwd
\aid

0 \volume19 \issue5A 2013 \firstpage1750 \lastpage1775 \doi10.3150/12-BEJ428 \newremarkremarkRemark[section] \newremarkexample[remark]Example

\runtitle

Penalization methods

{aug}

and

penalization methods \kwdreflecting stochastic differential equation \kwd-topology \kwdSkorokhod problem

## 1 Introduction

Let be a convex open set in . Consider a -dimensional reflecting stochastic differential equation (SDE),

 Xt=Ht+∫t0f(Xs−)dZs+Kt,t∈R+, (1)

where is a -dimensional semimartingale with , is an adapted process with , and is a continuous function such that

 ∥∥f(x)∥∥≤L(1+|x|),x∈Rd (2)

(for the precise definition, see Section 3). Our main purpose is to study the problem of approximation of weak solution of (1) by solutions of nonreflecting SDEs of the form

 Xnt=Hnt+∫t0f(Xns−)dZns−n∫t0(Xns−Π(Xns))ds,t∈R+,n∈N, (3)

where and are perturbations of and , respectively, and denotes projection of on . Because for large , the drift term forces to stay close to , it is called the penalization term, and the SDE (3) is called the SDE with penalization term.

The foregoing problem was intensively investigated in the case where is a Lipschitz continuous function, , and is a continuous semimartingale. In particular, Lions et al. lms () and Menaldi me () have proven that for , provided that is a -dimensional standard Wiener process. In the case where has jumps, to the best of our knowledge, such a problem has been considered previously only by Menaldi and Robin mr () and Łaukajtys and Słomiński laus (). Menaldi and Robin studied the case where is a diffusion with Poissonian jumps and . However, they imposed a very restrictive condition on the Poissonian measure coefficient, and consequently, is a process with continuous trajectories. In this case, earlier methods of approximation remain in force. In earlier work, we considered in detail the case where and is a general semimartingale. Because the approximating sequence might not be relatively compact in the Skorokhod topology , we proved our convergence results in the -topology introduced by Jakubowski ja (). It is worth pointing out that in both of the aforementioned papers, the initial process is constant (i.e., ), and is a Lipschitz continuous function.

The purpose of the present paper is to investigate the problem of approximation of by in the case of arbitrary initial process and arbitrary continuous coefficient satisfying the linear growth condition (2). Our proofs are based on new estimates for solutions of equations with penalization terms.

The paper is organized as follows. In Section 2 we consider a deterministic problem of approximating a solution of the Skorokhod problem , on domain associated with a given function such that (for precise definition, see Section 2). The penalization method involves approximating by solutions of equations of the form

 (4)

where in the Skorokhod topology . Lions and Sznitman ls () and Cépa ce () proved that tends to if is continuous. We omit the latter assumption and consider arbitrary function . In this general case, we prove that the variation of the penalization term of the SDE (4) is locally uniformly bounded and for fixed , provided that , which implies in particular that tends toward in the -topology. It is noteworthy that, similar to ce (), here we do not assume that the domain satisfies the so-called condition () introduced by Tanaka ta ().

In Section 3 we present new estimates on solutions of equations with penalization terms associated with a given process such that , that is, solutions of SDEs,

 (5)

In particular, we prove that if is a process admitting the decomposition , where is an adapted process, is an adapted local martingale with and is an adapted processes of bounded variation with , then, for every , there exist constants such that for every ,

 P(supt≤q∣∣Xnt−a∣∣≥η) ≤ P(ω′Hn(δ,q)≥da/2)+P(supt≤q∣∣Hnt−a∣∣≥C1η) +C2η−2E([Mn]q+∣∣Vn∣∣2q)

and

 P(n∫q0∣∣Xns−Π(Xns)∣∣ds≥η2) ≤ P(ω′Hn(δ,q)≥da/2)+7P(supt≤q∣∣Hnt−a∣∣≥C1η) +C2η−2E([Mn]q+∣∣Vn∣∣2q),

where denotes the usual modulus of continuity and .

In Section 4, we use estimates derived in Section 3 to prove our main results on the approximation of by . We assume that is a sequence of semimartingales satisfying the so-called condition (UT), and we prove that if converges weakly to in the topology, then converges weakly in the topology to . Moreover, we prove convergence of finite-dimensional distributions of to the corresponding finite-dimensional distributions of outside the set of discontinuity points of and . Consequently, using discrete approximations constructed in a manner analogous to Euler’s formula, we prove the existence of a weak solution of the SDE (1), provided that is continuous and satisfies (2). Moreover, if the SDE (1) has the weak uniqueness property, then our approximations computed by simple recurrent formulas allows us to obtain numerical solution of the SDE (1). In the case of reflected diffusion processes, similar approximation schemes have been considered previously (see, e.g., Liu li (), Pettersson p2 (), Słomiński s4 ()). In this section we also present some natural conditions ensuring convergence of to in probability provided that (1) has the so-called pathwise uniqueness property. Related results concerning diffusion processes have been given by, for instance, Kaneko and Nakao kn (), Gyöngy and Krylov gk (), Bahlali, Mezerdi and Ouknine bmo (), Alibert and Bahlali ab () and Słomiński s4 ().

We note that we consider the space equipped with two different topologies,  and . Definitions and required results for the Skorokhod topology have been given by, for example, Billingsley bil () and Jacod and Shiryayev js (). For the convenience of the reader, we have collected basic definitions and properties of the -topology in the Appendix. More details have been provided in Jakubowski ja ().

In this paper, we use the following notation. Every process appearing in the sequel is assumed to have trajectories in the space . If is a semimartingale, then represents and represents the quadratic variation process of , . Similarly, , and represents the predictable compensator of , . If is the process with locally finite variation, then , where is a total variation of on . In general, we let and denote convergence in law and in probability, respectively. To avoid ambiguity, we write () in if converges weakly (in probability) to in the space equipped with . Following ja (), we write () in when we consider the topology. For , , , we let and denote classical moduli of continuity of on , that is, , and where . We also use the modulus introduced in Jakubowski ja (). We recall that for , , .

## 2 A deterministic case

Let be a nonempty convex (possibly unbounded) open set in , and let denote the set of inward normal unit vectors at ( if and only if for every , where denotes the usual inner product in ). The following remark also can be found in Menaldi me () or Storm sto ().

{remark}

(i) If , then there exists a unique such that . Moreover, .

(ii) For every ,

where .

Let be a function with initial value in . We recall that a pair of functions is called a solution of the Skorokhod problem associated with if

• ,

• is -valued,

• is a function with locally bounded variation such that and

 kt=∫t0nsd|k|s,|k|t=∫t01{xs∈∂D}d|k|s,t∈R+,

where if .

The problem of existence of solutions of the Skorokhod problem and its approximation by solutions of equations with penalization terms has been discussed by many authors. Tanaka ta () proved existence and uniqueness of solutions in the case of continuous and domains also satisfying the following condition:

1. []

2. there exist constants and such that for every , we can find such that and .

Tanaka also observed that holds true in dimensions 1 and 2 or if is a bounded set. On the other hand, in dimension 2, one can construct examples of nonbounded convex domains not satisfying (). For instance, the cone with the basis and peak at , that is, the set

 C={(λx,λy,1−λ)∈R3;x2≤y,0≤λ≤1}, (6)

does not satisfy (). Cépa ce () omitted the assumption and proved the existence and uniqueness of the solution to the Skorokhod problem in the case of continuous function . In addition, Cépa proved convergence , of solutions of equations (4) for every sequence such that , .

The case of functions with jumps was considered for the first time by Anulova and Liptser al (), who proved the existence and uniqueness of solutions under condition (). Their result was generalized to the case of arbitrary convex by Łaukajtys lau (). In an earlier work laus (), we considered the problem of approximating noncontinuous by solutions of equations with penalization terms only in this very special case. We now consider the problem of approximating noncontinuous by solutions of equations with penalization terms in the general case of arbitrary sequences such that in . Our main tools are the following estimates on the solution of (4):

###### Lemma \theremark

Let , and let be a solution of the equation (4). Then for any and such that

 ω′yn(δ,q)

we have {longlist}

where , and denotes the largest integer less or equal to .

{pf}

We follow the proof of Theorem 3.2 in ce (). Let . Because is a continuous function such that ,

 ∣∣xnt−a∣∣2 = = ∣∣ynt−a∣∣2+2∫t0⟨xnu−a,dknu⟩+2∫t0⟨ynt−ynu,dknu⟩.

Therefore, for any ,

 ∣∣xnt−a∣∣2−∣∣xns−a∣∣2 = ∣∣ynt−a∣∣2−∣∣yns−a∣∣2+2∫ts⟨xnu−a,dknu⟩ −2∫ts⟨ynu−yns,dknu⟩+2⟨knt,ynt−yns⟩.

By Remark 2(ii),

 2∫ts⟨xnu−a,dknu⟩ = −2n∫ts⟨xnu−a,xnu−Π(xnu)⟩du ≤ −2dan∫ts∣∣xnu−Π(xnu)∣∣du=−2da∣∣kn∣∣ts,

and, consequently,

 ∣∣xnt−a∣∣2−∣∣xns−a∣∣2 ≤ ∣∣ynt−a∣∣2−∣∣yns−a∣∣2−2da∣∣kn∣∣ts−2∫ts⟨ynu−yns,dknu⟩ −2⟨ynt−a,ynt−yns⟩−2⟨a−xnt,ynt−yns⟩ ≤ −2∫ts⟨ynu−yns,dknu⟩.

By (7), there exists a subdivision of such that , , , where and . Thus, in particular,

 ∫sksk−1⟨ynu−ynsk−1,dku⟩≤∣∣∣∫(sk−1,sk)⟨ynu−ynsk−1,dknu⟩∣∣∣≤da2∣∣kn∣∣sksk−1.

Therefore,

 2(−∫sksk−1⟨ynu−ynsk−1,dknu⟩−da∣∣kn∣∣sksk−1)≤2(da2∣∣kn∣∣sksk−1−da∣∣kn∣∣sksk−1)=−da∣∣kn∣∣sksk−1,

which implies that

 ∣∣xnsk−a∣∣2−∣∣xnsk−1−a∣∣2≤5supt≤q∣∣ynt−a∣∣2+4supt≤q∣∣ynt−a∣∣⋅supt≤q∣∣xnt−a∣∣−da∣∣kn∣∣sksk−1. (8)

From (8), it follows immediately that

 ∣∣xnsk−a∣∣2−∣∣xnsk−1−a∣∣2 ≤

Set . Then

 ∣∣xnt−a∣∣2 = ≤ r(5supt≤q∣∣ynt−a∣∣2+4supt≤q∣∣ynt−a∣∣⋅supt≤q∣∣xnt−a∣∣)+supt≤q∣∣ynt−a∣∣2,

which implies that

 supt≤q∣∣xnt−a∣∣2≤14r2supt≤q∣∣ynt−a∣∣2+supt≤q∣∣xnt−a∣∣2/2.

Thus, , and the proof of (i) is complete.

(ii) Using (8) and (i) gives

 da∣∣kn∣∣sksk−1 ≤ ≤ 13supt≤q∣∣ynt−a∣∣2+32supt≤q∣∣xnt−a∣∣2≤55r2supt≤q∣∣ynt−a∣∣2

for . Thus, , which completes the proof.

###### Theorem \theremark

Assume that , , and let denote the solution of the equation (4), . If in , then {longlist}

and , ,

, provided that ,

, where denotes the solution of the Skorokhod problem associated with .

{pf}

(i) Because is relatively compact in , for any , and for any , , there exists such that . Therefore, the first conclusion follows from Lemma 2.

(ii) Let be a sequence of constants such that and . Set , , , where is an array of constants satisfying and . Now, for every , set , , , , . Observe that for every ,

 limi→∞supt≤q∣∣y(i)t−yt∣∣=0and% limi→∞limsupn→∞supt≤q∣∣yn,(i)t−ynt∣∣=0.

Let be a solution of an equation with a penalization term of the form

 xn,(i)t=yn,(i)t−n∫t0(xn,(i)s−Π(xn,(i)s))ds=yn,(i)t+kn,(i)t,t∈R+.

Fix and consider the decomposition , where denotes a solution of the Skorokhod problem associated with . Due to laus (), Lemma 2.2(i),

where variations , are bounded uniformly by Lemma 2(ii). Therefore,

 limi→∞limsupn→∞∣∣xnt−xn,(i)t∣∣=0.

Moreover, if , then . Because, by laus (), Lemma 3.3, , , it follows that . On the other hand, by ta (), Lemma 2.2, , and (ii) follows.

(iii) The sequence is relatively compact in , and consequently it is relatively compact. Because by part (i), , , the sequence is also relatively compact, and thus is relatively compact as well. In view of Corollary Appendix: The topology (Appendix), this proves (iii).

Recall that if in , then for every there exists a sequence such that

 yntn⟶yt,yntn−⟶yt−andΔyntn⟶Δyt. (9)

Moreover, for arbitrary sequences , such that , , and , we have

 ynt′n⟶yt−andynt′′n⟶yt (10)

(see, e.g., js (), Chapter VI, Proposition 2.1).

###### Corollary \theremark

Under the assumptions of Theorem 2, {longlist}

for every , if is a sequence satisfying (9), then

 xntn⟶xt−+Δyt

and for arbitrary sequences , such that , and , we have

 xnt′n⟶xt−andxnt′′n⟶xt,

moreover, if is continuous, then

 supt≤q∣∣xnt−xt∣∣⟶0,q∈R+.
{pf}

(i) If , are step functions, then the result follows from laus (), Lemma 3.3. In the general case, it is sufficient to use (9), (10) and repeat the approximation procedure from the proof of Theorem 2.

(ii) By Theorem 2(ii), , provided that . Therefore, it suffices to prove that is relatively compact in . Because , , , it is sufficient to show that

 limδ↓0limsupn→∞ω(Π(xn),yn)([0,δ])=0 (11)

and

 limδ↓0limsupn→∞ω′′(Π(xn),yn)(δ,q)=0,q∈R+. (12)

To prove (11), first observe that and , which implies that (11) is equivalent to the following:

• for every sequence such that ,

 Π(xnsn0)⟶x0,ynsn0⟶y0. (13)

Next, note that (13) is implied by (i) (it is sufficient to put and observe that in this case, ).

Similarly, (12) is equivalent to the condition

• for every and every sequence , such that , , if , , , , then

 a1=a2andb1=b2ora2=a3andb2=b3. (14)

Because is continuous and , it follows from (i) that for arbitrary sequences , such that and , we have

 Π(xnt′n)⟶xt−andΠ(xnt′′n)⟶xt. (15)

Combining (10) with (15), we see that there are only four possibilities:

 a1 = a2=a3=xt−andb1=b2=b3=yt−, a1 = a2=xt−,a3=xtandb1=b2=yt−,b3=yt, a1 = xt−,a2=a3=xtandb1=yt−,b2=b3=yt, a1