This paper aims to establish two compensation formulas for branching Lévy processes. The first expresses expected sums over spatial jumps as integrals against the Lévy measure, with the future family of the jumping particle started from its position immediately after the jump. The second expresses expected sums over reproduction events as time integrals weighted by the branching rate and the offspring distribution, with each event marked by the ordered tuple of branching processes initiated by its offspring. Combined with the fixed-time many-to-one formula, it recovers the whole-tree many-to-one identity for functionals of ancestral spatial paths.
Introduction
Consider a one-dimensional branching Lévy process with local reproduction defined on a probability space (Ω,F,P). The system starts with a single particle that moves according to a Lévy process with triplet (β,σ2,ν). Its lifetime is exponentially distributed with rate γ>0 and is independent of its motion. At the end of its lifetime, the particle is replaced by k offspring with probability pk, where p={pk}k≥0 is the offspring distribution. All offspring are born at the parent’s position at death and evolve independently according to the same motion and reproduction rules. We assume that only finitely many branching events occur in each bounded time interval almost surely.
This branching particle system can also be constructed by equipping a continuous-time Galton–Watson tree with independent Lévy motions along the lifetimes of its individuals; see [1] and § 2. We label the initial particle by ∅ and other particles using the Ulam–Harris system. The collection T of all labels forms a subset of U defined by
U:=n=0⋃∞N+n.
with N+={1,2,…} and the convention N+0={∅}; see [1, 3] for details. Let Nf(R) denote the space of finite counting measures on R, equipped with the topology of weak convergence. Let Df be the space of Nf(R)-valued càdlàg paths, equipped with the Skorokhod J1-topology. For z∈R, write δz for the unit point mass at z. For t≥0, denote by Vt the set of particles alive at time t. For u∈Vt and 0≤s≤t, let Xu(s) be the position of u or its ancestor at time s. The Nf(R)-valued process Z={Zt}t≥0 defined by
Zt:=u∈Vt∑δXu(t)(1)
has sample paths in Df and records the positions of all living particles over time. It is usually referred to as a branching Lévy process. We denote its natural filtration after the usual completion by {Ft}t≥0. For x∈R, we write Px and Ex for the law and expectation of this process with the initial particle started from x∈R, and abbreviate P=P0 and E=E0.
Our first result extends the Lévy process compensation formula to the spatial jumps of all particles in the branching Lévy process. For t≥0 and u∈Vt, let ΔXu(s):=Xu(s)−Xu(s−) denotes the jump of Xu at time s∈(0,t], and Zt,u={Zst,u}s≥0 denotes the future family of u, with its time origin at t and its positions in the original spatial coordinates, i.e.,
where u⪯v means that u=v or u is an ancestor of v. In particular, Zgu,u is the sub branching Lévy process initiated by u at its birth.
Theorem 1.1. Given a (Ft)-predictable process F={Ft:t≥0} taking values in the space of non-negative measurable functions on U×R such that Ft(u,0)=0 for all t≥0 and u∈U, and a deterministic measurable function G={Gt:t≥0} taking values in the space of non-negative measurable functions on U×R×Df. For any x∈R, we have
Remark 1.2. When p1=1, exactly one particle is alive at each time, and its position evolves as a Lévy process. When p0=1, the initial particle dies without offspring, so its motion is killed at an independent exponential time of rate γ. In these two cases, the formula (2) with G≡1 recovers the compensation formulas for a Lévy process and for a Lévy process killed at rate γ, respectively.
For n≥1, let Dfn be the space of ordered n-tuples of paths in Df endowed with the product Borel σ-algebra. Set
Df:=n=0⋃∞DfnwithDf0:={()},
and equip Df with the disjoint-union σ-algebra. Here () denotes the empty tuple. For x∈R and k≥1, let Px⊗k denote the joint law on Dfk of k independent copies Z1,…,Zk of the branching Lévy process started from a single particle at x. Write Ex⊗k for the corresponding expectation. For k=0, let Px⊗0 be the unit mass at ().
Our second result gives a compensation formula for the reproduction events of the branching Lévy process. For each individual u∈T, let gu, du, and ℓu denote its birth time, death time, and lifetime, respectively. We make the convention that g∅=0. At its death time du, particle u produces ku offspring, where ku has law p. Under the Ulam–Harris labeling, the i-th child of u is labeled ui for 1≤i≤ku.
Theorem 1.3. Given a (Ft)-predictable process F={Ft:t≥0} taking values in the space of non-negative measurable functions on U×N, and a deterministic measurable function G={Gt:t≥0} taking values in the space of non-negative measurable functions on U×Df. For any x∈R, we have
Remark 1.4 The proof of Theorem 1.3 remains valid when the Lévy motion is replaced by a strong Markov motion with a Borel measurable path law as a function of the starting point. In particular, let f be a non-negative measurable function on R+×D([0,∞),R) such that f(t,ω) depends only on ω∣[0,t). We write f(t,ω∣[0,t)) for its value. Taking G≡1, Ft(u,k)=f(t,Xu∣[0,t)) for u∈Vt−, k≥0, and setting Ft(u,k)=0 elsewhere, we obtain from (3) that
where Exξ denotes expectation for a single copy ξ of the spatial motion started at x. This is the whole-tree many-to-one formula of [1 Proposition 3.5] specialized to local reproduction and functionals of the spatial path.
Remark 1.5 For any x0∈R, consider a branching Lévy process Z with particles being killed as they enter (−∞,x0). The two compensation formulas (2) and (3) remain valid for Z, since Z can be constructed as a functional of Z, i.e., for any t and u∈Vt, we have Xu(t)=Xu(t) and
Let Ft≥0 denote the natural filtration of Z, completed in the usual way. Obviously, it is a subfiltration of {Ft}t≥0. As an illustration, let F and G be two functionals satisfying the assumptions of Theorem 1.1, with Ft replaced by Ft. Then
Before proving the two main theorems, we first describe in detail the construction of the branching Lévy process Z started from a single particle at x∈R by assigning independent Lévy motions to the individuals of a continuous-time Galton–Watson tree. To each potential label u∈U, we assign an exponential lifetime ℓu of rate γ, an offspring count ku with law p, and a Lévy process Yu=(Yu(s))s≥0 started at zero with triplet (β,σ2,ν). The parameter s denotes the age of the particle. All these random elements are mutually independent, both within and across labels.
The realized tree T contains the root ∅. For a label u=(u1,…,un) with n≥1, it belongs to T if and only if ui≤ku(i−1) for all i=1,…,n, where u(0)=∅ and u(i−1)=(u1,…,ui−1) for i≥2. For each u∈T, its birth and death times are
gu:=v≺u∑ℓvanddu:=gu+ℓu,
where v≺u means that v is a strict ancestor of u. The empty sum gives g∅=0. Thus the set of particles alive at time t≥0 is Vt={u∈T:gu≤t<du}. For u∈Vt, its position at time t can be given by
Xu(t)=x+v≺u∑Yv(ℓv)+Yu(t−gu).
Each strict ancestor contributes its displacement over its entire lifetime, and u contributes its displacement up to age t−gu. Substituting Vt and Xu(t) into (1) gives a realization of Z.
Proof of Theorem 1.1. Let T={(u,ℓu)}u∈T be the whole genealogy of all labels and lifetimes. Almost surely, no jump occurs at gu for all u∈T, since the lifetime is independent of the motion and a Lévy process has no jump at a fixed age. Reindexing all jumps gives
For ϵ>0 and u∈T, set τu,0ϵ=gu and define recursively τu,jϵ:=inf{s∈(τu,j−1ϵ,du):∣ΔXu(s)∣≥ϵ} for j≥1 with the convention inf∅=∞. The finite terms of {τu,jϵ}j≥1 are the successive times at which particle u makes an upward or downward jump of size at least ϵ during its lifetime. Let
The monotone convergence theorem yields that SuT(ϵ)a.s.ϵ→0+SuT as ϵ→0+. For t≥0, let Gt:=Ft∨σ(T) reveal T at time zero and each particle’s spatial motion up to time t. It is obvious that {τu,jϵ}j≥1 is a sequence of (Gt)-stopping times. Since σ(T)⊂Gτu,1ϵ, we have
Conditioned on a given genealogy T, for any particle u and t∈[gu,du), let Qt,uT(x,dz) be the law on Df of branching Lévy process started at x∈R at time t, using the branching schedule of T from time t and fresh independent Lévy motions. Define
Conditioned on T, the process Xu is a Lévy process on [0,du] started at x with Lévy measure ν. By the compensation formula for jumps of Lévy process; see e.g. [2 p.7],
By the exponential memorylessness and the independent future motions and reproduction, the branching property at time t yields that conditional on Ft, the process Zt,u is a branching Lévy process started at Xu(t−)+y. This along with (7) induces that for u∈Vt,
The desired formula (2) follows directly by plugging this into the right-hand side of the second equality in (8) and using the stochastic continuity of Lévy processes. This completes the proof. □
Proof of Theorem 1.3. The first equality in (3) follows directly from the fact that each particle contributes to the sum only at its death time. For the second equality, we first have
For i=1,…,k∅, let Ti collects labels of the i-th child of ∅ and all its descendants. By the tower property and the mutual independence among descendants, the second expectation on the right-hand side of (9),
For each u∈U, let Au:=σ((Yv)v∈U,(ℓv,kv)v∈U∖{u}). Hence both the event {u∈T} and the time gu are Au-measurable, since they depend only on the strict ancestors of u. Conditionally on Au, the lifetime ℓu and the offspring count ku are still independent. For 0≤a<b<∞, A∈Fa and k≥0, we first consider the contribution of particle u when A occurs and a<du≤b. On the event {u∈T,du>a}, particle u has not died by time a, so the history up to a is the same if we replace ℓu by a+1 and ku by zero while keeping all other variables fixed. Consequently, on this event, 1A=1Au for some Au∈Au, obtained by evaluating A in that modified history. Since du=gu+ℓu on {u∈T}, the density of ℓu and the law of ku give
When u∈/T, both conditional expectations in (12) are zero. Taking expectations on both sides of (12) yields the desired identity for a predictable rectangle (a,b]×A with mark k. For A∈Fa, we apply the same identity on (a+1/n,b] using an event in Fa+1/n that agrees with A almost surely, and then let n→∞. By the monotone class theorem and the monotone convergence theorem, we obtain
The second equality in (3) follows directly by substituting the definition of Ht(u,k) into the right-hand side of the last equality, then using the stochastic continuity of V and Lévy processes. This completes the proof. □
References
[1]V. Bansaye, J.F. Delmas, L. Marsalle, and V.C. Tran. Limit theorems for Markov processes indexed by continuous time Galton–Watson trees. Ann. Appl. Probab., 21(6):2263–2314, 2011.
[2]J. Bertoin. Lévy Processes. Cambridge University Press, Cambridge, 1996.
[3]J.F. Le Gall. Random trees and applications. Probab. Surv., 2:245–311, 2005.DOI