You are on page 1of 12
JOURNAL OF MULTIVARIATE ANALY 44, 179-190 (1993) Integration by Parts for Poisson Processes Rosert J. Euiotr* anp ALLaNus H. Tsor’ Department of Statistics and Applied Probable, University of Alberta, Ednonton, Alberto, Canada TG 2G1 Using perturbation of the rate of Poisson process and an inverse time change an integration by parts formula is obtained. This enables a new form of the integrand in & martingale cepresentaion result to be obtained. — + 1} Acer 1. INtRopuCTION In his paper [3] on the Malliavin calculus Bismut obtains an integration by parts formula for a diffusion by considering a small perturbation of the trajectories and then compensating for this by using a Girsanov change of measure. That is, suppose ¢ denotes the original trajectory and ¢* the perturbation. Let £ (resp. E") denote expectation with respect to. the original measure (resp. the measure afier the Girsanov transformation) Then for any bounded, differentiable function «, itis the case that ElelE)] = ELEN) aay The left side of this equation is independent of ¢ and Bismut obtains his integration by parts formula by differentiating in c and putting «=0. Integration by parts formulae for Markov jump processes have been obtained by Bass and Cranston [1]. and Bichteler er al. [2]. Again, the variation of the trajectories considered by these authors consists of perturbing the size of the jumps. A Poisson process is a counting process, and all jumps are of unit size Consequently, a perturbation of the trajectories of the kind considered in Received Januaty 25, 1990; revised July 8, 1992, [AMS classifications: 60}H07, 60175. Key words and phrases: Poisson process, integration by parts, time change, martingale representation * Research partially supported by grants from NSERC A-7964, the Air Force Office of Scientific Research, United States Air Force. under Contract AFOSR-8640332, and the US. Army Research OMfie, under Contract DAALO3-87-0102, "Research partially supported by NSERC A-7968 and a University of Alberta Dissertation Fellowship 179 (0047-25993 $5.00 180, FLIOTE AND SOI [1, 2] does not make sense. Instead we consider below a Girsanov change of measure which alters the rate of the Poisson process by a small amount. This is then compensated by considering a time change of the process under the new measure, An identity analogous to (1.1) is obtained and the integration by parts formula follows by differentiating with respect to a parameter ¢ and putting ©=0. The case where the function depends only On finitely many jumps is discussed first, and the general case, for a func tional of the Poisson process over the time interval [0, 1], is then deduced. There is a close relation between integration by parts formulae and martingale representation results. It is well known that any uniformly integrable martingale on the sigma fields generated by a Poisson process can be represented as a stochastic integral with respect to the associated martingale, The integrand can be obtained by considering one jump at @ time (though the precise form given in Eq. (2.6) does not appear to be in the literature). What is interesting is that the integration by parts method gives an alternative expression for this integrand, which does involve a derivative of the functional of the process. The equality of these two expressions is verified in the Appendix when the functional depends on finitely many jump times. This expression for the integrand is similar to that obtained by Clark [5] for functionals of Brownian motion. Of course, the identities for the Poisson process given below are casily verified using the exponential distribution of the jump times. However, our derivations using perturbations of the jump rate rather than jump size are new and, hopefully, of interest 2. MARTINGALE REPRESENTATION AND Tis: CHANGE Let N be a Poisson process on (2,%,(¥).P) with jump times Tyne Tyrone We shall write Ty=0. Let GUT). 0 Tyr :) be an integrable function ‘of Ty... Ty... Consider the martingale M defined by M, = ELGUT, ) ay For n> 1, write PT, =My,— My, , FGI F,.]- ELG|F, 1 (22) From Theorem 79, Chapter 3 of [4], the martingale M defined by (2.1) has the representation M,=F(G1+{" 9.40, 23) INTEGRATION RY PARIS 181 where Q,=N a= Ris) Hfs-T)-e 8 { fuye" du (24) on (71<5<7,,,), and J. Since Fy=o{T\T)}, Tiay~T, is independent of Ty... 7). and exponentially distributed, so P1S)= EMI Tis Fo Tra FT, EU Ma | Fr] “du Puyo" du (25) From (2.4) and (2.5), we have gs) =ETGIT, From (25) MEU GF] (2.6) ELMy, Fi) Siluye™* de (27) Also, /' can be written as Us)= f's—T))— My, (28) Using (24), (2.7), and (28), we have Hor roy+ | Le 29) Throughout the rest of this paper we let {u,, 10} be a real predictable process satisfying: i) {u,, 120) is positive and as. pounded, |u,| 0, consider the martingale "au, dO, E uan-[ (2.10) Define the family of exponentials AwexplX LNA) TL Uae” Sil (14a, v,yerp( [em 48) Quy Then | A!, 10} satisfies the equation va flay ae, a1 fas add, (2.12) and {45,120} is a martingale, (See [6}.) Using the fact that |u,) < Bas. and w vanishes outside the integral (0, 6] it is easily checked that 4; is a square integrable martingale for 1 >0. Hence 4‘, exists and a new probability measure P* can be defined by ap dP Then the process {Q/} defined by y,-[/ Ut aads is an (%) martingale under P* (see [7]. Now define, with ¢<1/B, ddty= | (1+ eu.) ds (2.13) Let y()= 6, (1), Then ¥.19,(1))= 1 80 (214) WTA INTEGRATION BY PARTS, 183 If we let F) =F then from Theorem T16, Chapter It of [4], the process {.N', 20} defined by Nos ean (245) is a Poisson process on (2, F,(F 7). P’). 3, INTEGRATION nY PARTS Suppose G is a function of the first » jump times T;.... 7, of a Poisson process N. Since 9,(1)=¥, '(1) if 7, is the ith jump time of {N,}, then 4,(T,)is the ith jump time of the process {N,,.)}. Changing the rate of the point process by a Girsanov transformation, and then changing the time scale of the process, we have the following result: THEOREM 3.1, Let G(Tyy. Ty) be bounded with bounded first partial derivatives, Then el ([f «a0.) errs, a --2[% £ GT. 7.) wa] any eh Proof. By the results in Section 2, because Nya is a Poisson process under P’ with jump times 4,(T), consequently. ELG(T yon TD = ETC AT so BATON S ELAS GOAT) on OAT VI 2) where E*[ ] denotes that expectation is taken with respect to P'. Differen- tiating (3.2) with respect to c, and then setting ¢=0, we obtain “ete=0 ge “(de From (2.12) and the definition of 4°, . e[ay £ GIdAT hon (Teno GIT) OT Die ol (3) a4 de 184 ELLIOTT AND TSOL Noting the definition of ¢,. (33) becomes (3.1) and the proof is complete, I Remarks 3.2. For suitable functions w and G (3.1) can be checked using the exponential distribution of the jump times. However, this proof is no shorter and does not exhibit the above motivation. Consider a function H of the form H(T, A 1,u,T, 1), where H is bounded and hhas bounded first derivatives. Applying Theorem 3.1 to G(T, T,)= HT) A Aon T, 4 1) and noting that GOUT on TAS MT) boone Ty Tres 1, we have the following: COROLLARY 3.3. If H(T, A Wen Ty 1) is bounded and has bounded first derivatives, then af (ff) mr nara] ee eee 34) Remark 34, Recall the martingale representation (2.3) and (2.4), or GUT Ty) = ECG) + fg. dQ. G5) where gees) for T, .T, Also, if we consider the measure j« defined by not all” noted, 6) x a AT, nn Ty) 8p (dO. ade INTEGRATION BY PARTS, 185 Then the right-hand side of (3.1) can be written Sarto) a “aff fon] = HE [ff ute 9nd] ~e[ Et. Sin, Taina | 7) Let C= D0) Jy, 2ACG/P1.MT ys To). Then there exists @ predictable projection C*'of C such that for each ', CI=EIC IA 1 as. Also for any predictable process {w,,s>0}, Elu,C.1= Ele LLC FV Elu,CP] (8) Let be the family of subsets of [0, 0¢) x @ of the form {0} x F, and (1)xF, where Foe A and Fe F, for s0. Then el (a2, orn, Alou Tyas | HEelZananatanafiabtah em Proof. First note that for each M > 0, the partial sum ¥ Uyeda LPN 2 I<4e so that by hypothesis, the right-hand side of (3.12) is finite. For each n> 1, define HT oy Ty) = LGTY Va ioe Ty A eo Fh Then HT oT, FELGT Noon Te A Ne (Tet Spe iA Te (Tyt Snir tt Sp DA Neo Fe] SENGT, A byes Ta A WT t Sua tA Deon (Ty 4Sy arto + Spa) A bods (3.13) where S,=T,—T, , for K>1, and the last expectation £* in (3.13) is INTRGRATION BY PARTS, 187 taken only over the random variables S, .y.Sy.cj.4 and the Ty, Ty are given. From (3.1), Hea) [Pua raf na G14) | GAT, A Vpn Ty A Ue (Tt Soe) A booed Py ] GT, Von Ty AL (Tat Seed Ae MnoitFe] - eL. C u, ds z E 0 as nm (G.15) eon Te ATs MneilFa] by the hypotheses on {u,} and G Also for 1 20 in (3.14), because of (3.15) and (3.16), we obtain (3.12). 188 ELLIOTT AND 1901 We conclude with the following theorem: THEOREM 3.7. Suppose M is the right continuous martingale My=ELG(T, 0 1 THA NF Then M,=ELGI+{ @,dQu where re, AL NF as. (3.17) Proof. The proof is similar to the one for Proposition 3.5. APPENDIX We now give a direct proof that the integrands ¢ obtained in Sections 2 and 3 are equal, The idea is to first establish the equivalence on {11 without any more calculation. First we need a preparatory lemma, Given a bounded measurable function R(x), .. %,) define Fv Med Mees Mad HR, TYAN pretty ah Note for later use that for j>i (ay Finally, for é fixed, let 7,=7,—T,_, for ji and note that (7... 7,) has the same probability distribution as (T),..T, 4) Lema 41 Uf (Ty, T,) are the successive jump times of a Poisson process, then (a) ELR(T).- TMF, J CPP oe Tete 8) Mego, vee INTEGRATION BY PARTS, 189 (b) ELR(T yy. Ta Tie = ELEY, Testo, Menon weet fe) ECR (Ty. TMF Vir, vere May peg ELE non Byte Md To tH, Maayan vettnt 0 ‘The proof of this lemma is elementary and relies mainly on the independence of (7... 7,) from Ty... T, y, and the equality Fn 17, 11}. A ditect calculation and an integration by parts shows that Vip 2 MECGUT son TMT) =e EU 7,2: AT ys Tadd ne e[ESu, Tait] By part (c) of the lemma and by using the facts that 1y),., is % measurable and that Vij.o1; Ijn2.= lyn =n almost surely for j>1, we find that this last expression equals (Ties Ta) a] This completes Step 1 Step 2. We prove the equivalence on {T, ,<¢

You might also like