Forum. Math. Galley Proof, 41 pages Forum Mathematicum DOI 10.1515/FORM.2011.100 © de Gruyter 2011
Interacting superprocesses with discontinuous spatial motion
Zhen-Qing Chen, Hao Wang and Jie Xiong Communicated by Ichiro Shigekawa
Abstract. A class of interacting superprocesses arising from branching particle systems with continuous spatial motions, called superprocesses with dependent spatial motion (SDSMs), has been introduced and studied by Wang and by Dawson, Li and Wang. In this paper, we extend the model to allow discontinuous spatial motions. Under Lipschitz con- dition for coefficients, we show that under a proper rescaling, branching particle systems with jump-diffusion underlying motions in a random medium converge to a measure- valued process, called stable SDSM. We further characterize this stable SDSM as a unique solution of a well-posed martingale problem. To prove the uniqueness of the martingale 2 problem, we establish the C C -regularity for the transition semigroup of a class of jump- diffusion processes, which may be of independent interest. Keywords. Superprocess, Brownian sheet, interaction, symmetric stable process, frac- 2 tional Laplacian, C C -regularity of transition semigroup, branching mechanism, scaling limit, duality method, martingale problem. 2010 Mathematics Subject Classification. Primary 60J68, 60K37; secondary 60G52, 60J80.
1 Introduction
Dawson–Watanabe superprocesses arise naturally as the scaling limit of branching processes where each particle moves independently of each other. In this paper, we are concerned with the scaling limit of an interacting branching particle system where the motion of each particle is governed by the following equation: for each i N, 2 Z t Z tZ Z t i i i i i i i zt z0 c.zs /dBs h.y zs /W.dy; ds/ b.zs / dSs ; D 0 C 0 R C 0 (1.1)
The research is supported in part by NSF Grants DMS-0600206 and DMS-0906743 (first author), by a research Grant of University of Oregon (second author) and by NSF Grant DMS-0906907 (third author). 2 Z.-Q.Chen, H. Wang and J. Xiong where Bi i 1 are independent standard Brownian motions on R, W is a ¹ I º Brownian sheet on R2 (see definition below) and S i i 1 are independent ¹ I º 1-dimensional symmetric ˛-stable processes with ˛ .0; 2/. The processes W , 2 Bi i 1 and S i i 1 are assumed to be independent of each other. Equa- ¹ I º ¹ I º tion (1.1) says that each particle is subject to a common random medium force, which is described by the Brownian sheet W , and each particle has its own jump- diffusion dynamics, which are described by an individual Brownian motion and an individual symmetric stable process. We will show in this paper that under suit- able conditions and scaling, the branching particle system converges weakly to a measure-valued process, which we call an interacting superprocess. We will fur- ther show that this measure-valued process is the unique solution to a related mar- tingale problem. When b 0 (that is, there is no stable-noise motion S i in the par- D ticle system), such a model has been studied in Wang [32,33], which has been fur- ther investigated in subsequent papers by Dawson et al. [10] and Li et al. [24, 25]. However, many physical, economic and biological systems are better modeled by using discontinuous stochastic processes because the jumps exhibited by discon- tinuous processes reveal sudden big changes when viewed in a long time scale. Thus it is natural and meaningful to study the scaling limit of interacting branch- ing processes with discontinuous spatial motion such as those described by (1.1). The presence of jumps introduces many new challenges. For example, to estab- lish the uniqueness of the martingale problem of the interacting superprocess, one needs to establish the C 2-regularity for the semigroup of the n-particle system of (1.1). However unlike the diffusion case, due to the presence of (discontinuous) stable Lévy motion, the infinitesimal of such a semigroup is a non-local opera- tor. Establishing regularity results for pseudo-differential operators is typically a challenging problem and is, in fact, of current research interest (see, e.g., [2], es- pecially item 6 of its Section 7.7). One of the main results of this paper is to show that under certain conditions there is some .0; 1/ so that for every n 1, 2 2 n the semigroup of the n-particle system of (1.1) maps Cb C .R / into itself. The following is a more detailed description on the content of this paper. Let L2.R/ be the Hilbert space of square-integrable functions on R and let m m Bb.R / be the space of bounded Lebesgue measurable functions on R . Denote by C.Rm/ and C j .Rm/ the space of continuous functions on Rm and the space of continuous functions on Rm with continuous derivatives up to and including m k m order j , respectively. Denote by Cb.R / and Cb .R / the space of bounded continuous functions on Rm and the space of bounded continuous functions that have bounded derivatives up to and including order k, respectively. We use Lip.R/ to denote the space of Lipschitz functions on R; that is, f Lip.R/ if there is a 2 constant M > 0 such that f .x/ f .y/ M x y for every x; y R. The class j j Ä j j 2 of bounded Lipschitz functions on R will be denoted by Lipb.R/. Interacting superprocesses with discontinuous spatial motion 3
Under the condition that b; c Lip .R/ and h L2.R/ Lip.R/, we will 2 b 2 \ show that the interacting stochastic differential equation (SDE) (1.1) has a unique strong solution. From the strong solution of (1.1), we can construct a family of interacting branching particle systems. Let > 0 and > 1 be fixed constants. .n/ i For n 1, suppose that initially there are m0 number of particles located at z0, .n/ 1 i m , and each has mass n. These particles evolve according to Ä Ä 0 (1.1) and branch independently at rate n. The branching mechanism for each particle is assumed to be state independent, independent to each other and iden- tically distributed. After branching, the offsprings of each particle evolves in- dependently according to (1.1) and then branches again. The common nth-stage branching mechanism q.n/ q.n/ k 0; 1; : : : is assumed to be critical (that WD ¹ k I D º is, the average number of offspring is 1), and it cannot produce 1 or more than n n Pm.n/ number of children. Under the assumption that the initial state ı k of k 1 z0 D.n/ the particles converges to a measure 0 and that the branching function q con- verges uniformly to a limiting branching function having finite second moment n P and some additional conditions, we show that the empirical process ı k k 1 zt converges to a measure-valued process. By Itô’s formula and the independence of motions and branching, we will show that the limiting measure-valued process has the following formal generators (usually called pregenerators):
LF ./ AF ./ BF ./; (1.2) WD C Z 2 1 2 ı F ./ BF ./ 2 .dx/ (1.3) WD 2 R ı.x/ and
Z Â 2 Ã 1 d ˛ ˛=2 ıF ./ AF ./ 2 a.x/ 2 b.x/ .dx/ WD R dx C j j ı.x/ (1.4) 1 Z Z @2 Â ı2F ./ Ã .x y/ .dx/.dy/ C 2 R R @x@y ı.x/ı.y/ for F ./ D.L/ C.MF .R//. Here > 0 is related to the branching rate of 2 the particle system and 2 > 0 is the variance of the limiting offspring distribution; MF .R/ denotes the Polish space of all finite measures on R with weak topology, C.MF .R// is the space of all continuous functions on MF .R/. The variational derivative is defined by
ıF ./ F . "ıx/ F ./ lim C (1.5) ı.x/ WD " 0 " I # 4 Z.-Q.Chen, H. Wang and J. Xiong the fractional Laplacian ˛=2 on C 2.R/ is defined for C 2.R/, b 2 b Z ˛=2 c˛ .x/ .x / .x/ 0.x/1 1 1 ˛ d (1.6) WD R 0 C ¹j jÄ º C n¹ º j j and Z .x/ h.y x/h.y/dy; a.x/ c2.x/ .0/; (1.7) WD R WD C and D.L/, the domain of the pregenerator L, consists of functions of the form
F ./ f . 1; ;:::; ; /; D h i h k i 2 2 k with i C .R/, f C .R /, k N . In the usual models (for example, 2 c 2 b 2 .˛; d; ˇ/-superprocesses, see Dawson [8], Dynkin [13] or Dynkin et al. [14]) in which the motions of particles are independent and the motions are independent of branching, the particle systems have the following multiplicative property. If two branching Markov processes evolve independently, with initial state m1 and m2 respectively, then their sum has the same distribution as the branching process with initial state m1 m2. It is well known that the log-Laplace functional (or C evolution equation) technique can be applied to these models in order to construct the limiting measure-valued process. However, for our model and pregenerator, it is obvious that the motions of particles are not independent and this destroys the multiplicative property. Thus, just as in Wang [33] and Dawson et al. [11], the usual log-Laplace functional method is not applicable to our new model. In order to construct the branching particle system, we show that under Lipschitz condi- tion on functions c, h and b, SDE (1.1) has a strong solution and the solution is pathwise unique. Since the symmetric ˛-stable process S i is not square-integrable and is a martingale only when ˛ .1; 2/, the Picard’s successive approximation 2 method is not directly applicable to (1.1) and a truncation procedure on the jumps of S i is needed. To prove the uniqueness of the martingale problem for L for the measure-valued interacting process, we use a duality method due to Dawson and Kurtz [9]. To ap- ply Dawson and Kurtz’s duality method in our context, we need to find, for every 2 m m integer m 1, a linear subspace D of Cb .R / so that Pt D D for every m t > 0. Here Pt t 0 is the transition semigroup of the underlying motion of ¹ I º 2 m m-particles given by (1.1). We show in this paper that we can take C C .R / for such D for some .0; 1/. That is, we show that there is some .0; 1/ so m 2 2 m 2 m 2 that P f C C .R / for each fixed t > 0 and f C C .R /. This result t 2 b 2 b has its own independent interest and is one of the main results of this paper. For earlier regularity results for non-local operators, we refer the reader to [2, 6, 28] Interacting superprocesses with discontinuous spatial motion 5 and the references therein. Note that the infinitesimal generator of m-particles .z1; : : : ; zm/ of (1.1) is given by (see the proof of (3.4) below)
.m/ L f .x1; : : : ; xm/ m 1 X @2 ij .x1; : : : ; xm/ f .x1; : : : ; xm/ WD 2 @xi @xj i;j 1 D m X ˛ ˛=2 b.xj / f .x1; : : : ; xm/ C j j xj j 1 D m (1.8) 1 X @2 ij .x1; : : : ; xm/ f .x1; : : : ; xm/ D 2 @xi @xj i;j 1 D m X Z Â f .x1; : : : ; xj w; : : : ; xm/ f .x1; : : : ; xm/ C C j 1 R 0 D X¹ º Ã ˛ @ c˛ b.xj / f .x1; : : : ; xm/w1 w 1 j 1 ˛j dw; @xj ¹j jÄ º w j j C where ´ a.xi / if i j , ij .x1; : : : ; xm/ D (1.9) WD .xi xj / if i j . ¤ ˛=2 Here xj f is the fractional Laplacian defined in (1.6) applied to the function
xj f .x1; : : : ; xj ; : : : ; xm/: 7! 2 m The C C -regularity of Pt will be proved by a perturbation method, treating Pm ˛ ˛=2 .m/ the non-local operator part j 1 b.xj / xj of L as a perturbation of the 1 PDmj j @2 elliptic differential operator 2 i;j 1 ij .x/ @x @x . D i j In this paper, to better illustrate our approach to the study of interacting super- processes whose underlying spatial motions are discontinuous, we restrict our- selves to a specific type of jump-diffusions for the spatial motions given by (1.1). However the method of this paper works for a more general class of jump-diffu- R t i sions. For example, there can be a drift term 0 r.zs/ds in (1.1) and the symmetric stable process S i can be replaced by some other Lévy process of pure jump type such as mixed stable processes. Moreover, S i i 1 can even be replaced by ¹ I º a certain family of independent random measures (see [20]). Furthermore, the spatial motions zi in (1.1) can also be multidimensional. 6 Z.-Q.Chen, H. Wang and J. Xiong
While the component of ˛-stable noise in the spatial motion is captured by the fractional Laplacian ˛=2 in the expression (1.4) of A for the pregenerator L of the interacting superprocess, the superprocess itself has continuous trajectories. This is because in this paper, we have critical branching at every stage and the branching function q.n/ converges to a limiting branching function in which the offspring distribution has a finite second moment. These assumptions imply that the limiting interacting measure-valued process has null Lévy (jumping) measure for its branching mechanism and therefore the superprocess is continuous. This is in analog to the case of Dawson–Watanabe process that it has continuous paths if and only if the Lévy (jumping) measure for the branching mechanism is zero (see Fitzsimmons [18] and El Karoui–Roelly [15]). If the offspring distribution determined by the limiting branching function of q.n/ has infinite second moment, for example, a stable branching mechanism or a general branching mechanism, then the limiting measure-valued process is discon- tinuous. For the latter case, we believe that the conditional log-Laplace functional would be a good tool to construct the limiting measure-valued process. See [8] and [25] for related work in this direction on interacting superprocesses with con- tinuous spatial motion. The remainder of this paper is organized as follows. Section 2 is devoted to establish the strong existence and pathwise uniqueness for solutions of (1.1). This is the basis for the construction of the branching particle systems. In Section 3, we construct branching particle system. The tightness of the corresponding empir- n ical measure-valued processes t t 0 is established in Section 4. We prove 2 ¹ I º n the L -convergence of each term in the the decomposition of ; t and then 0 h i derive the corresponding decomposition of ; t for any weak sequential limit 0 n h i t t 0 of t t 0 n 1. The latter implies that the measure-valued pro- ¹ I º ¹ I º cess 0 t 0 solves the martingale problem for the generator .L; D.L// ¹ t I º of (1.2). Moreover, we show that 0 t 0 is a continuous process taking ¹ t I º values in MF .R/. We then use Dawson–Kurtz’s duality method to show that the martingale problem for .L; D.L// is well-posed. This in particular implies n that t 0 converges weakly in the Skorokhod space D.Œ0; /; MF .R// ¹ t I º 1 to a continuous measure-value process that solves the martingale problem for L. 2 .m/ The proof of the C C -regularity property of the semigroup Pt , which is used in a crucial way in the proof of well-posedness of the aforementioned martingale problem, is given in Section 5. For the reader’s convenience, let us recall the definition of the Brownian sheet W on R. For a D Rm or D R R , let B.D/ be the Borel -field D D C on D. By abusing the notation, the Lebesgue measure on R and on R2 will all be denoted by m. Let .; F ; Ft t 0; P/ be a filtered probability space with a right ¹ º Interacting superprocesses with discontinuous spatial motion 7
continuous filtration Ft t 0. A random set function W on B.R R / defined ¹ º C on .; F ; Ft t 0; P/ is called a Brownian sheet or a space-time white noise on ¹ º R if (i) For A B.R R / having finite Lebesgue measure, W.A/ is a Gaussian 2 C random variable with mean zero and variance m.A/.
(ii) If Ai B.R R /, i 1; 2, have finite Lebesgue measure and satisfy 2 C D A1 A2 , then W.A1/ and W.A2/ are independent and \ D;
W.A1 A2/ W.A1/ W.A2/ P-a.s. [ D C (iii) For every A B.R/ having finite Lebesgue measures, 2
M.A/t W.A Œ0; t/; WD
as a process in t 0, is a square-integrable Ft -martingales. ¹ º In fact, it follows from (i) and (ii) that for every A B.R/ having finite 2 Lebesgue measures, t M.A/t is a centered Gaussian process with indepen- 7! dent stationary increments with EŒM.A/2 m.A/t. Thus it has a continuous t D modification and is a Brownian motion in R. Condition (iii) above puts in an additional requirement that M.A/ be an Ft -martingale. It is a consequence of ¹ º above (ii) and (iii) that for every A; B B.R/ with finite Lebesgue measures, the 2 covariance process for martingales M.A/ and M.B/ satisfies
M.A/; M.B/ t m.A B/ t; t 0: h i D \ For more detailed information on Brownian sheet, the reader is referred to Walsh [31, Chapter 2] and Dawson [8, Section 7.1].
2 Strong existence and pathwise uniqueness of underlying motions
Recall that a one-dimensional symmetric stable process of index ˛ .0; 2 is a 2 Lévy process S St t 0 such that WD ¹ I º h i ˛ E ei.St S0/ e t for every R: D j j 2 Note that when ˛ 2, S is just a Brownian motion with speed running twice D fast as the standard Brownian motion. More generally, given a filtered probability space .; F ; Ft t 0; P/, a process S St t 0 is said to be a Ft -sym- ¹ º WD ¹ I º ¹ º 8 Z.-Q.Chen, H. Wang and J. Xiong
metric ˛-stable process if St is Ft -measurable for every t 0 and for every t > r > 0, St Sr is independent of Fr and h i ˛ E ei.St Sr / e .t r/ for every R: D j j 2
c˛ The Lévy measure for such a process is given by w 1 ˛ dw, where c˛ is a con- stant depending only on ˛. Its role can be understoodj j C in the following Lévy sys- tem formula for a one-dimensional Ft -symmetric ˛-stable process S. For any ¹ º Ft -predictable process H Ht t 0 and any Borel measurable function F ¹ º 2 R t R WD ¹ I º c˛ on R satisfying 0 R HuF .Su;Su w/ w 1 ˛ dw du < , we have for every j C j C 1 t > 0, j j Z tZ X c˛ t HuF .Su ;Su/ HuF .Su;Su w/ dw du (2.1) w 1 ˛ 7! u t 0 R C C Ä j j is a local martingale with respect to the filtration Ft t 0. It is a martingale with ¹ º respect to the filtration Ft t 0 if in addition H is a bounded process and ¹ º Ä t Z Z c˛ E HuF .Su;Su w/ 1 ˛ dwdu < R j C j w 1 0 j j C for every t > 0. Here for any process with jumps we use Xs limt s Xt to WD " denote the left hand limit and Xs Xs Xs to denote the jump at time s. See, 4 WD for example, [4], [26], [20] and [30] for more information on the stochastic anal- ysis of processes with jumps. The L2-infinitesimal generator K of a symmetric ˛-stable process is the fractional Laplacian . /˛=2. It can be defined in terms of Fourier transform as follows: ° ± Dom.K/ f L2.R/ ˛fb ./ L2.R/ ; WD 2 W j j 2 Kf ./ ˛fb ./ for f Dom.K/: WD j j 2 2 R ix Here for f L .R/, fb ./ e f .x/dx. It can be shown (see, e.g., [3]) 2 WD R that C 2.R/ Dom.K/ and Kf ˛=2f for f C 2.R/, where the latter is c D 2 c defined by (1.6). b Theorem 2.1. Let h L2.R/ Lip.R/, b; c Lip .R/. Let W be a Brown- 2 \ 2 b ian sheet, B Bt t 0 be a standard one-dimensional Brownian motion WD ¹ I º and S St t 0 be a one-dimensional symmetric ˛-stable process with WD ¹ I º ˛ .0; 2/. Assume that W , B and S are defined on a common filtered probabil- 2 ity space .; F ; Ft t 0; P/ and are independent of each other. Then for every ¹ º Interacting superprocesses with discontinuous spatial motion 9
F0-measurable random starting point Z0 R the stochastic integral equation 2 Z t Z t Z tZ Zt Z0 c.Zs /dBs b.Zs / dSs h.y Zs /W.dy; ds/ D C 0 C 0 C 0 R (2.2) has a unique strong solution.
Proof. Note that the symmetric ˛-stable process S is not square-integrable and is a martingale only when ˛ .1; 2/. So the standard Picard’s successive approxi- 2 mation method does not work here. However, it can be made to work with a truncation argument. Such a truncation method is known to experts. However, for the reader’s convenience, we will spell out the details of the proof. We make the convention S0 S0. With the above notations, it is well known D that X bS t St Ss1 Ss >1 ; t 0; (2.3) WD s t 4 ¹j j º Ä c˛ is a Lévy process with Lévy measure w 1 ˛ 1 w 1 dw and that bS and S bS are C ¹j jÄ º independent (cf. [4, pp. 13–15]). Notej thatj the process bS is a pure jump, square- integrable martingale and
bS t St for t < T1 inf r > 0 Sr > 1 : (2.4) D WD ¹ W j4 j º We now use Picard’s successive approximation method to construct a strong solu- tion Y for (2.2) but with bS in place of S. Later on, we will use Y to construct the strong solution for (2.2). .1/ For any given real-valued F0-measurable random variable Z0, let Y Z0 t WD for t 0, and define for k 2 and t 0, Z t Z t .k/ .k 1/ .k 1/ Y Z0 c.Y /dBs b.Y / dbS s t WD C s C s 0 0 (2.5) Z tZ .k 1/ h.y Ys /W.dy; ds/: C 0 R Thus for k 3 and t 0, we have Z t .k/ .k 1/ .k 1/ .k 2/ Á Yt Yt c.Ys / c.Ys / dBs D 0 Z t .k 1/ .k 2/ Á b.Ys / b.Ys / dbS s C 0 Z .k 1/ .k 2/ Á h.y Ys / h.y Ys / W.dy; ds/: C R 10 Z.-Q.Chen, H. Wang and J. Xiong
As the three terms on the right hand side are square-integrable martingales that are independent to each other, by Doob’s maximal inequality we have
Ä ˇ ˇ2 ˇ .k/ .k 1/ˇ fk.t/ E sup ˇYs Ys ˇ Á 0 s t Ä Ä Ä 2 ˇ .k/ .k 1/ˇ 12E ˇY Y ˇ Ä ˇ t t ˇ ÄZ t 2 .k 1/ .k 2/ Á 12E c.Ys / c.Ys / ds D 0 ÄZ tZ 2 2 .k 1/ .k 2/ Á c˛ w 12E b.Ys / b.Ys / j1 j˛ 1 w 1 dwds C R w ¹j jÄ º 0 j j C ÄZ tZ 2 .k 1/ .k 2/Á 4E h.y Ys / h.y Ys dy ds C 0 R ÄZ t 2 2 2 2 .k 1/ .k 2/Á c c0 b0 h0 E Ys / Ys ds Ä k k1 C k k1 C k k1 0 Z t c1 fk 1.s/ds; (2.6) Ä 0 where is the supremum norm. A similar calculation shows that k k1 Ä ˇ ˇ2 ġ ˇ2 ˇ .2/ .1/ˇ ˇ .2/ .1/ˇ E sup ˇYs Ys ˇ 4E ˇYt Yt ˇ 0 s t Ä Ä Ä ÄÂZ t Z t 4E c.Z0/dBs b.Z0/dbS s D 0 C 0 Z tZ Ã2 h.y Z0/W.dy; ds/ C 0 R Ä Z 2 2 2 c˛ w 4E c.Z0/ t b.Z0/ j1 j˛ 1 w 1 dw D C R w ¹j jÄ º j j C Z tZ 2 h.y Z0/ dyds C 0 R 2 2 c2 . c b .0// t c3 t: Ä k k1 C k k1 C DW It then follows from above that, for n 3, Ä ˇ ˇ2 t n 1 ˇ .n/ .n 1/ˇ n 1 E sup ˇYs Ys ˇ c4 ; 0 s t Ä .n 1/Š Ä Ä Interacting superprocesses with discontinuous spatial motion 11
where c4 c1 c3. This implies that for every t > 0, D _ v u Ä Ã2 u X1 ˇ .n/ .n 1/ˇ tE sup ˇYs Ys ˇ ˇ ˇ n 1 0 s t D Ä Ä s Ä 2 X1 ˇ .n/ .n 1/ˇ E sup ˇYs Ys ˇ < : Ä ˇ ˇ 1 n 1 0 s t D Ä Ä P ˇ .n/ .n 1/ˇ n Therefore 1 sup ˇYs Ys ˇ converges a.s. and consequently Y n 1 0 s t converges uniformlyD onÄ eachÄ Œ0; t to a càdlàg process Y a.s. Clearly we also have from above that for every t > 0,
Ä ˇ ˇ2 ˇ .n/ ˇ lim E sup ˇYs Ysˇ 0: n 0 s t D !1 Ä Ä Thus we deduce by the same argument as that for (2.6) that
Z t Z t Z tZ .k 1/ .k 1/ .k 1/ c.Ys /dBs b.Ys / dbS s h.y Ys /W.dy; ds/ 0 C 0 C 0 R converges in L2 as k to ! 1 Z t Z t Z tZ c.Ys /dBs b.Ys / dbS s h.y Ys /W.dy; ds/: 0 C 0 C 0 R This implies by (2.5) that
Z t Z t Z tZ Yt z c.Ys /dBs b.Ys / dbS s h.y Ys /W.dy; ds/: (2.7) D C 0 C 0 C 0 R
Moreover, Y is the unique strong solution for (2.7). Indeed, suppose Yb is another strong solution of (2.7). Then for every t 0, Z t Á Z t Á Yt Ybt c.Ys / c.Ybs / dBs b.Ys / b.Ybs / dbS t D 0 C 0 Z Á h.y Ys / h.y Ybs / W.dy; ds/: C R By a similar argument as that for (2.6), we have
Ä ˇ ˇ2 Ä ˇ ˇ2 ˇ ˇ ˇ ˇ E sup ˇYs Ybsˇ c5t E sup ˇYs Ybsˇ : 0 s t Ä 0 s t Ä Ä Ä Ä 12 Z.-Q.Chen, H. Wang and J. Xiong
ˇ ˇ It follows that sup0 s t ˇYs Ybsˇ 0 a.s. for every t Œ0; 1=c5/, and therefore Ä Ä D 2 for every t Œk=c5; .k 1/=c5/. This proves that Y Yb a.s. 2 C D Recall the stopping time T1 in (2.4). Define ´ Yt if t < T1; Zt WD YT1 b.YT1 /.ST1 ST1 / if t T1: C D Then, in view of (2.3), Z is the unique strong solution for SDE (2.2) over the random time interval Œ0; T1. For k 1, define Tk Tk 1 T1 Tk 1 ; WD C ı where t t > 0 are shift operators on the canonical path space . In other ¹ I ºth words, Tk is the k time that the symmetric ˛ stable process S has a jump of size larger than 1. Since t St is right continuous and has left hand limit, S can only 7! have finite many jumps of size larger than 1 over any compact time intervals. So limk Tk almost surely. !1 D 1 Note that BT1 t BT1 t 0 is a Brownian motion, ST1 t ST1 t 0 ¹ C I º ¹ C I º is symmetric ˛-stable process and W ..y1; y2; ŒT1;T1 t/ t 0 is a Brow- ¹ C I º nian sheet that are independent to each other and are independent of FT1 . Sup- 2 pose Y t 0 is the strong solution of (2.7) but with ZT in place of Z0, and ¹ t I º 1 BT1 t BT1 t 0 , ST1 t ST1 t 0 and W ..y1; y2; ŒT1;T1 t/ in ¹ C I º ¹ C I º C place of B, S and W ..y1; y2; Œ0; t/ there, respectively. Define
´ 2 Yt if t < T1 T1 ; ZT1 t 2 2 ı C WD Y b.Y /.S S / if t T1 : T1 T T1 T T2 T2 T1 ı 1 C ı 1 D ı
It is easy to see that such defined Z solves SDE (2.2) for t T2. Iterating this Ä procedure, we can define Z on Œ0; T for every k 1 and hence on Œ0; /. Thus k 1 we have obtained a strong solution Z for (2.2) for t Œ0; /. Such a strong 2 1 solution is unique as it is unique on each time interval ŒTk 1;Tk. Here we take T0 0. D
3 Branching particle systems
In order to construct branching particle systems, first we need to introduce an index set to identify each particle in the branching tree structure. Let be the < set of all multi-indices, i.e., strings of the form n1n2 : : : n , where the ni ’s D k are non-negative integers. Let denote the length of . We provide with j j < the arboreal ordering: m1m2 : : : mp n1n2 : : : nq if and only if p q and Ä Interacting superprocesses with discontinuous spatial motion 13
m1 n1; : : : ; mp np. If p, then has exactly p 1 predecessors, which D D j j D we shall denote respectively by 1, 2; : : : ; 1. For example, with j j C 6879, we get 1 687, 2 68 and 3 6. D D D D Let B be an independent family of standard Brownian motions in R, ¹ I 2 <º S be an independent family of one-dimensional symmetric ˛-stable ¹ I 2 <º processes with ˛ .0; 2/ and W be a Brownian sheet. Assume further that W , 2 B and S are defined on a common filtered probability space ¹ I 2 <º ¹ I 2 <º .; F ; Ft t 0; P/, and independent of each other. For every index and ¹ º 2 < initial data z0, by Theorem 2.1 there is a unique strong solution zt for Z t Z t z z c.z /dB b.z / dS t D 0 C s s C s s 0 0 (3.1) Z tZ h.y zs /W.dy; ds/: C 0 R
Since the strong solution of (3.1) only depends on the initial state z0, the Brownian motion B Bt t 0 , the symmetric ˛-stable process S St t 0 WD ¹ I º WD ¹ I º and the common white noise W , we can write z ˆ.z ;B ;S ; t/ for some t D 0 measurable real-valued map ˆ (omitting W in the notation as it is selected and fixed once and for all).
Lemma 3.1. There is a constant c > 0 such that for every C 2.R/, 2 b ˛=2 c 00 : k k1 Ä k k1 C k k1 Proof. By the mean-value theorem, we have ˇZ ˇ ˛=2 ˇ c˛ ˇ .x/ supˇ .x w/ .x/ 0.x/w1 w 1 1 ˛ dwˇ k k1 D x Rˇ R 0 C ¹j jÄ º w C ˇ X¹ º j j Â2 Z Z Ã 1 ˛ 1 ˛ c 00 w dw w dw Ä k k1 w 1 j j C k k1 w >1 j j j jÄ j j < : 1 This proves the lemma.
Define X S S S 1 ; bt t s Ss >1 WD s t 4 ¹j4 j º Ä which is a Lévy process independent of S bS . Note that zt b.zt / St . 4 D 4 14 Z.-Q.Chen, H. Wang and J. Xiong
So by the property of Lévy measure (see (2.1)), we have for every C 2.R/, 2 b ; X h i M .z / .z / 0b .z / S 1 t s s s s Ss 1 WD s t 4 ¹j4 jÄ º Ä Z t ÂZ Á zs b.zs /w .zs / 0 R 0 C X¹ º Ã Á c˛ 0b .zs / w1 w 1 1 ˛ dw ds ¹j jÄ º w j j C is a square-integrable martingale with
Ä 2 ÄZ tZ ; Á E Mt E .zs b.zs /w/ .zs / D 0 R C 2 Á c˛ .0b/.zs /w1 w 1 1 ˛ dwds ¹j jÄ º w j j C ÄZ t ÂZ 1 2 2 (3.2) E 1 00b w 1 w 1 zs @ Ä 0 ¹ 6D º R 2k k1 j j ¹j jÄ º à 2 Á c˛ 2 1 w >1 1 ˛ dw ds C k k1 ¹j j º w j j C 2 c b 00 t: Ä k k1 C k k1 1 ˛ Observe that due to the symmetry of the kernel c˛= w , j j C Z ˛=2 c˛ .z/ .z w/ .z/ 0.z/w1 w h 1 ˛ dw D R 0 C ¹j jÄ º w C X¹ º j j for every h > 0. So by a change of variable, we have Z c˛ .x b.x/w/ .x/ .0b/.x/w1 w 1 1 ˛ dw R 0 C ¹j jÄ º w C X¹ º j j b.x/ ˛˛=2.x/: D j j It follows that
; X Á M .z / .z / 0b .z / S 1 t s s s s Ss 1 D s t 4 ¹j4 jÄ º Ä (3.3) Z t ˛ ˛=2 Á b .zs /ds: 0 j j Interacting superprocesses with discontinuous spatial motion 15
With C 2.R/, we have by Itô’s formula that for every t > 0, 2 b Z t Z tZ .zt / .z0/ 0c .zs /dBs 0.zs /h.y zs /W.dy; ds/ D 0 C 0 R Z t Z t 1 2 0b .zs / dSs ..0/ c .zs //00.zs /ds C 0 C 2 0 C X h i .zs / .zs / 0b .zs / Ss C s t 4 Ä Z t Z tZ 0c .zs /dBs 0.zs /h.y zs /W.dy; ds/ D 0 C 0 R Z t Z t 1 0b .zs / dbS s a00 .zs /ds C 0 C 2 0 X h i .z / .z / 0b .z / S 1 s s s s Ss 1 C s t 4 ¹j4 jÄ º Ä Z t Z tZ 0c .zs /dBs 0.zs /h.y zs /W.dy; ds/ D 0 C 0 R Z t ; 0b .zs / dbS s Mt C 0 C Z t 1 ˛ ˛=2 Á 2 a00 b .zs /ds: (3.4) C 0 C j j We now consider the branching particle systems in which each particle’s spatial motion is modeled by the SDE (3.1). For every positive integer n 1, there is an n n initial system of m0 particles. Each particle has mass and branches indepen- dently at rate n. The branching mechanism is assumed to be state independent. n Let qk denote the probability of having k offsprings. We assume that qn 0 if k 1 or k n 1; k D D C and n X n n kq 1 and lim sup q pk 0; k D n j k j D k 0 !1 k 0 D where p ; k 0; 1; 2; : : : is the limiting offspring distribution which is assumed ¹ k D º to satisfy following conditions:
X1 X1 2 p1 0; kp 1 and m2 k p < : D k D WD k 1 k 0 k 0 D D 16 Z.-Q.Chen, H. Wang and J. Xiong
Let mn Pn .k 1/4qn. Here mn n 1 may be unbounded, but we c WD k 0 k ¹ c I º assume that D mn lim c 0 for any 2. n 2n D !1 We will see that the limiting offspring distribution is the offspring distribution of the Stable SDSM, the limiting measure-valued process we will construct. We n n assume that the initial number of particles m0 , where > 0 and 2 are n Pn 2 n Ä„2 n „ 2 fixed constants. Define m2 k 0 k qk , n m2 1 and m2 1. 2 2 WD D WD WD Note that n and are the variance of the n-stage and the limiting offspring 2 2 2 distribution, respectively. We have n < and limn n . 1 !1 D Let C be a family of i.i.d. real-valued exponential random variables ¹ I 2 <º with parameter n, which serve as lifetimes of the particles, and D be ¹ I 2 <º a family of i.i.d. random variables with P.D k/ qn for k 0; 1; 2; : : : , D D k D where qn k 0; 1; : : : is the offspring distribution satisfying above conditions. ¹ k I D º We assume W , B , S , C and D are all ¹ I 2 <º ¹ I 2 <º ¹ I 2 <º ¹ I 2 <º independent. The birth time ˇ./ of the particle x is given by
´P 1 j j jj j 1 C if D 2 for every j 1; : : : ; 1, ˇ./ D D j j WD ; otherwise: 1
The death time of x is given by ./ ˇ./ C and the indicator function of D C the lifespan of x is denoted by h 1 .t/. t WD Œˇ./;.// Recall that @ denotes the cemetery point. Define x @ if either t < ˇ./ or t D t ./. We make a convention that any function f defined on R is automatically extended to R @ by setting f .@/ 0. This convention allows us to keep track [ ¹ º D of only those particles that is alive at any given time. n n n Pm0 n Given 0 MF .R/, let 0 1 ıx be such that 0 0 as 2 WD 0 ) n (see [33]). We are thus provided withD collection of initial starting points ! 1 x for each n 1. ¹ 0º Let N n 1; 2; : : : ; mn be the set of indices for the first generation of parti- 1 WD ¹ 0º cles. For any N n , define 2 1 \ < ´ ˆ.x ;B ;S ; t/ if 0 t < C ; x 0 Ä (3.5) t WD @ if t C ; and x @ for any .N N n/ and t 0: t Á 2 n 1 \ < Interacting superprocesses with discontinuous spatial motion 17
n If 0 N and D 0 .!/ k 2, define for every 0i i 1; 2; : : : ; k , 2 1 D 2 ¹ I D º ´ ˆ.x0 ;B ;S ; t ˇ.// if t Œˇ./; .//; .0/ xt 2 (3.6) WD @ otherwise.
0 If D .!/ 0, define xt @ for 0 t < and for every 0i i 1 . D Á Ä 1 n 2 ¹ I º More generally for an integer m 1, let Nm be the set of all indices for th n < the living particles in m -generation. If 0 N and D 0 .!/ k 2, define 2 m D for 0i i 1; 2; : : : ; k 2 ¹ I D º ´ ˆ.x0 ;B ;S ; t ˇ.// if t Œˇ./; .//; .0/ xt 2 (3.7) WD @ otherwise:
If D0 .!/ 0, define D x @ for 0 t < and for every 0i i 1 : t Á Ä 1 2 ¹ I º Continuing in this way, we get a branching tree of particles for any given ! with n ® 1 2 m0 ¯ random initial state taking values in x0 ; x0 ; : : : ; x0 . This gives us our branch- ing particle systems on the real line where particles undergo a finite-variance branching at independent exponential times and have interacting spatial motion powered by jump-diffusions and a common white noise.
4 Tightness and the limiting martingale problem
Recall the branching particle system constructed in the last section. Define its associated empirical process by
n 1 X t .A/ ı .A/ for A B.R/: (4.1) WD n xt 2 2< n In the following, we will show that t 0 converges in D.Œ0; /; MF .R// ¹ t I º 1 weakly as n and its weak limit is the stable SDSM. ! 1 For any t > 0, define
n X .D 1/ M .A .0; t/ n 1 x A; ./ t ; (4.2) WD ./ ¹ 2 Ä º 2< which gives the total mass along each survived trajectory in set A up to time t. 18 Z.-Q.Chen, H. Wang and J. Xiong
Note that (3.4) implies that for every C 2.R/, 2 b n n 1 n 1 n n n ; t ; 0 Ut ./ Ht ./ Xt ./ Yt ./ h i h i D p n C p n C C (4.3) 1 n n Jt ./ Mt ./; C p n C where, recalling hs 1 .s/, D Œˇ./; .// Z t n 1 X Ut ./ hs 0.xs /c.xs /dBs ; WD p n 0 2< Z tZ n n Xt ./ h.y /0. /; s W.dy; ds/ ; WD 0 Rh i Z t n 1 ˛ ˛=2 n Yt ./ 2 a00 b ; s ds; WD 0 h C j j i Z t n 1 X Ht ./ hs 0.xs /b.xs /dbS s ; WD p n 0 2< 1 X ; J n./ M t WD p n t 2< Z tZ n n Mt ./ .x/M .dx; ds/ WD 0 R X .D 1/ n .x./ / 1 ./ t : D ¹ Ä º 2< The six terms in (4.3) represent the respective components of the overall mo- n tion of the finite particle systems ; t contributed by the individual Brownian n h i n motions (Ut ./), the random medium (Xt ./), the mean effect of jump-diffusive n n motions (Yt ./), the branching mechanism (Mt ./), the individual stable mo- n tions (Ht ./) with jump size bounded by one and the individual compensated n jump sums (Jt ./). Using a result of Dynkin [12, Theorem 10.25 on p. 325], we get at once the following theorem.
Theorem 4.1. For any n N, n defined by (4.1) is a right continuous strong 2 t Markov process which is the unique strong solution of (4.3) in the sense that it is a unique solution of (4.3) for a given probability space .; F ; P/ and given W , B , S , C , D defined on .; F ; P/. Furthermore, n t 0 are all ¹ º ¹ º ¹ º ¹ º ¹ t I º defined on the common probability space .; F ; P/. Interacting superprocesses with discontinuous spatial motion 19
For each t 0, let F n denote the -algebra generated by the collection of t processes ° ± n./; U n./; X n./; Y n./; M n./; H n./; J n./; C 2.R/; t 0 : t t t t t t t 2 b
Note that according to our assumption, the fourth moment of D ,
mn E .D 1/4 ; c WD ¹ º n 2n is finite and limn mc 0 for any 2. !1 D Lemma 4.2. Based on the above notations, we have the following:
2 n n (i) For every Cb .R/, M ./ Mt ./ t 0 is a purely discontinuous square-integrable2 martingale withWD ¹ I º
Z t n 2 2 n M ./ t n ; u du for every t 0: h i D 0 h i
(ii) For any t 0 and n 1, we have h n 2i n 2 2 n E sup 1; s 2 1; 0 8n t 1; 0 : 0 s th i Ä h i C h i Ä Ä There is a constant > 0 such that for every t 0, h n 4i 3 6 3 n 2 4 2 n 2 E sup 1; s n t 1; 0 n t 1; 0 0 s th i Ä h i C h i Ä Ä mn Á c t 1; n 1; n 4 : C 2n h 0i C h 0i
(iii) Let n t 0 be defined by (4.1). Then n t 0 is tight as a family of ¹ t I º ¹ t I º processes with sample paths in D.Œ0; /; MF .R//. 1 Proof. (i) Recall that C are i.i.d. exponential random variables with ¹ I 2 <º parameter n, D 1 are i.i.d. random variables with zero mean and ¹ I 2 <º these two families are independent. Thus E M n./ 0 for every t > 0 and ¹ t º D S.R/. Since this holds for any initial state n, by the Markov property of 2 0 n t 0 , we have for every t; s > 0, ¹ t I º n n ˇ n n n E M ./ M ./ E n M ./ M ./ 0: t s t ˇ F t t s 0 C D D 20 Z.-Q.Chen, H. Wang and J. Xiong
n This shows that M ./ is a martingale. Clearly, it is purely discontinuous. Note 1 Ä add some E 2.x / ./ t wording ./ I Ä or replace period by Ä colon? E 1 .ˇ./ C /2.x / Œ0;t .ˇ./ C / D C C ÄZ 1 2 n nu E 1Œ0;t.ˇ./ u/ .x.ˇ./ u/ / e du D 0 C C ÄZ 1 2 n E 1Œ0;t.ˇ./ u/ .x.ˇ./ u/ / 1 C >u du by independence D 0 C C ¹ º ÄZ 1 2 n E 1Œ0;t.ˇ./ u/ .xˇ./ u/ du by the definition of x D 0 C C ÄZ 1 2 n E 1Œ0;t.v/ .xv/ dv D ˇ./ ÄZ t n 2 E 1Œˇ./;.//.v/ .xv/dv : D 0 As C and D are all independent and ED 1, we conclude ¹ I 2 <º ¹ I 2 <º D that Ä n 2 X 2n h 2i 2 E Mt ./ E .D 1/ E .x./ /1../ t/ D Ä 2< ÄZ t 2n 2 X n 2 n E 1Œˇ./;.//.v/ .xv/dv (4.4) D 0 2< ÄZ t 2 2 n n E ; v dv : D 0 h i n Note that identity (4.4) holds for any initial distribution 0. Again by the Markov property of n t 0 , we have for every t; s > 0, ¹ t I º Ä Z t s ˇ n 2 n 2 2 C 2 n ˇ n E Mt s./ Mt ./ n ; v dv ˇ F t C t h i ˇ Ä Z s n 2 2 2 n E n M ./ ; dv 0: t s n v D 0 h i D This shows that M n./2 2 R t 2; n dv is a martingale. Hence we conclude t n 0 h vi that M n./ is a purely discontinuous square-integrable martingale with Z t n 2 2 n M ./ t n ; u du for every t 0: h i D 0 h i Interacting superprocesses with discontinuous spatial motion 21
(ii) Since 1; n n M n.1/ is a zero-mean martingale, by Doob’s maximal h t 0i D t inequality, we have Ä Ä n 2 n 2 n 2 E sup 1; s 2E sup Ms .1/ 2 1; 0 0 s th i Ä 0 s t C h i Ä Ä Ä Ä 8E M n.1/2 2 1; n 2 Ä t C h 0i 8 2t 1; n 2 1; n 2: Ä n h 0i C h 0i n P D 1 Note that Mt .1/ n 1 ./ t is a purely discontinuous martingale D ¹ Ä º and D 1 are i.i.d2< random variables with zero mean and are independent ¹ I 2 <º of C . Thus ¹ I 2 <º Ä Â 2 2 à h n 4i X .D 1/ .D 1/ E Mt .1/ E 2n 1 ./ t 2n 1 ./ t D ¹ Ä º ¹ Ä º ; ; 2< ¤ ÄX .D 1/4 E 4n 1 ./ t C ¹ Ä º 2< 4 Ä n Ä n X mc X 4n E 1 ./ t 1 ./ t 4n E 1 ./ t D ¹ Ä º ¹ Ä º C ¹ Ä º ; ; 2< ¤ 2< 4 n ÄZ t n X mc n 4n E 1 ./ t 1 ./ t 2n E 1; v dv : D ¹ Ä º ¹ Ä º C 0 h i ; ; 2< ¤ For ; with , C and C are independent and so 2 < ¤ E 1 ./ t 1 ./ t ¹ Ä Äº ¹ Ä º E 1 .ˇ./ C / 1 .ˇ./ C / D Œ0;t C Œ0;t C ÄZ Z 2 2n 1 1 nu nv E 1Œ0;t.ˇ./ u/1Œ0;t.ˇ./ v/e e dudv D 0 0 C C ÄZ Z 2 2n 1 1 E 1Œ0;t.ˇ./ u/1Œ0;t.ˇ./ v/ Ä 0 0 C C Á 1 x @ 1 x @ dudv ˇ./ u ˇ.Á/ u ¹ C 6D º ¹ C 6D º ÄÂZ ÃÂZ à 2 2n 1 1 E 1Œ0;t.r/1 dr 1Œ0;t.s/1 Á ds xr @ xs @ D ˇ./ ¹ 6D º ˇ./ ¹ 6D º ÄÂZ t ÃÂZ t à 2 2n E 1Œˇ./; .//.r/dr 1Œˇ./; .//.s/ds : D 0 0 22 Z.-Q.Chen, H. Wang and J. Xiong
Therefore X E 1 ./ t 1 ./ t ¹ Ä º ¹ Ä º ; ; 2< ¤ ÄÂZ t ÃÂZ t à X 2 2n E 1Œˇ./; .//.r/dr 1Œˇ./; .//.s/ds Ä 0 0 ; 2< 2 3  Z t Ã2 2 2n X E 4 1Œˇ./; .//.r/dr 5 D 0 2< ÄÂZ t Ã2 2 4n n E 1; r dr : D 0 h i It then follows 4 ÄÂZ t Ã2 n ÄZ t h n 4i n 2 4n n mc n E Mt .1/ 4n E 1; r dr 2n E 1; v dv Ä 0 h i C 0 h i n 2 4 2 h n 2i mc n n t E sup 1; r 2n 1; 0 t D r Œ0;th i C h i 2 mn 2 4t 2 8 2t 1; n 2 1; n 2 c 1; n t Ä n n h 0i C h 0i C 2n h 0i mn 83 6t 3 1; n 22 4t 2 1; n 2 c t 1; n : (4.5) D n h 0i C n h 0i C 2n h 0i By Doob’s maximal inequality,
h n 4i h n n n 4i E sup 1; s E sup 1; s 0 1; 0 0 s th i Ä 0 s t h i C h i Ä Ä Ä Ä h n 4i n 4 8E sup Ms .1/ 8 1; 0 Ä 0 s t j j C h i Ä Ä h 4i 8.4=3/4 E M n.1/ 8 1; n 4 Ä t C h 0i 3 6t 3 1; n 2 4t 2 1; n 2 Ä n h 0i C n h 0i mn Á c t 1; n 1; n 4 : C 2n h 0i C h 0i n (iii) We first prove the tightness of in D.Œ0; /; MF .R//, where R is the ¹ º 1 O O one-point compactification of R. Note that (ii) above implies the compact con- tainment property for n . By Theorems 4.5.4 and 4.6.1 in Dawson [7], we then ¹ º only need to prove that, for any given " > 0, > 0, T > 0, C 2.R/ and 2 b Interacting superprocesses with discontinuous spatial motion 23
any stopping time n bounded by T , there exist ı > 0 and n0 1 such that sup sup P n ./ n ./ > " . n n0 t Œ0;ı ¹j n t n j º Ä For this, first2 observe thatC by (4.3), n n P. n t ./ n ./ > "/ j C j 1 h 2i E n ./ n ./ 2 n t n Ä " C Ä 6 1 2 1 2 E U n ./ U n ./ H n ./ H n ./ 2 n n t n n n t n t Ä " C C C C n n 2 n n 2 Xn t ./ Xn ./ Yn t ./ Yn ./ C C C C 1 2 2 J n ./ J n ./ M n ./ M n ./ : n n t n t n t n C C C C C Note that by the independence of B , ¹ I 2 <º ÄZ n t h 2i 1 X E U n ./ U n ./ E C h . c/2.x /ds n t n n s 0 s C D n 2< ÄZ n t C ˝ 2 n˛ E .0c/ ; s ds D n h ˝ 2 n˛i t E sup .0c/ ; s Ä s T t Ä C 2 h n i t 0c E sup 1; s : Ä k k1 s T t h i Ä C By the independence of bS and the Lévy system formula (2.1), ¹ I 2 <º h n n 2i E Hn t ./ Hn ./ C Ä 1 X X 2 2 n E hs .0b/ .xs /. bS s/ D 4 n