A sharp survival threshold for branching Brownian motion in a self-generated resource environment

Let V_t \mathrel{\mathop:}= \eta_t\left(\mathds{R}^d\right)+\theta\mu_t\left(\mathds{R}^d\right), and let N_t^b and N_t^d denote the numbers of births and deaths by time t. Let A_t be the total consumption rate at time t, given by A_t \mathrel{\mathop:}= \int_{\mathds{R}^d}\left(K*\eta_t\right)\left(y\right)\,\mu_t(dy). Between jumps, resource consumption decreases V_t at rate \theta A_t. A birth increases V_t by 1, while a death removes one particle and adds one unit of resource, so it changes V_t by \theta-1. Hence,

By Fubini’s theorem and the symmetry of K we have A_t = \int_{\mathds{R}^d}\left(K*\mu_t\right)\left(x\right)\,\eta_t(dx), so the process \theta A_t is also a compensator for N_t^b. Thus

are local martingales. Substituting these in (Proof) gives

so if \theta\leq1, then V is a nonnegative local supermartingale. Consequently, it converges almost surely.

Every birth produces a jump of size 1 in V, so convergence implies that only finitely many births occur. Hence only finitely many particles ever exist, and since all eventually die, the population becomes extinct in finite time almost surely.

Let T, \gamma, and H_0 be as in Proposition 3, and choose H\geq H_0 so large that

By Lemma 4, there is an r-ball B_0 such that \mathds{P}\left(\eta_T\left(B_0\right)\geq H\right)>0.

On \left\{\eta_T\left(B_0\right)\geq H\right\}, apply Proposition 3 successively at times T,2T,\ldots, using at each step the measurable ball supplied by the proposition. If the first k doublings have succeeded, the kth application starts from an r-ball containing at least 2^kH particles, so its conditional probability of failure is at most e^{-\gamma2^kH}. Hence, by the tower property and a union bound,

where B_k denotes the ball selected after k successful steps. Since the entrance event also has positive probability, with positive probability the population is nonempty at all times (k+1)T. Extinction is absorbing, so on this event \tau_0=\infty. Thus \mathds{P}\left(\tau_0=\infty\right)>0.

Fix k\in\mathds{N}, to be chosen later, and set

where h is as in Lemma 2. Write B\mathrel{\mathop:}= B_k. Our goal is to show that, except with probability exponentially small in H, we have \eta_{S+T}\left(B\right)\gtrsim\lambda^kH. The radius of B grows linearly in k, so B can be covered by O\left(k^d\right) balls of radius r. Since \lambda>1, we can choose k so that one of these r-balls contains at least 2H particles.

We prove the lower bound on \eta_{S+T}\left(B\right) by repeatedly applying Lemma 2 to k successive time intervals of length h. The jth application starts with a dense colony in B_j and implies the existence of a dense colony in B_{j+1}. Thus all relevant particles and corpses remain in B. To iterate, we must ensure that every corpse in the corresponding sets \mathsf C is \alpha-consumed.

Recall that K\left(x\right)\geq k_0 whenever \left|x\right|\leq r. Consider one of the growth intervals \left[t,t+h\right] and a corpse \mathsf{q}\in\mathsf C, as in Lemma 2. Since d_{\mathsf q}\leq t+(1-\kappa)h, the corpse has at least \kappa h time to be consumed before the end of the interval. If at least n particles remain within distance r of X_{d_{\mathsf{q}}}\left(\mathsf{q}\right) throughout \left[t,t+h\right], then

Choose n\in\mathds{N} large enough that e^{-k_0n\kappa h}\leq1-\alpha. Thus n nearby particles are enough to make a corpse \alpha-consumed within time \kappa h.

To apply Lemma 2 repeatedly, we only need to verify that there are at least n particles within distance r of any point in B, throughout \left[S+h,S+T\right]. Choose \mathcal F_S-measurable points z_1,\ldots,z_l such that

where C<\infty depends only on d,r, and R. For m\leq l, define

Thus \mathsf D_m consists of particles from the initial colony which enter the mth covering ball by time S+h, then remain there and alive until S+T. Let \mathcal{D}\mathrel{\mathop:}=\bigcap_{m=1}^l\left\{\left|\mathsf D_m\right|\geq n\right\}. On \mathcal{D}, every point of B has at least n particles within distance r throughout \left[S+h,S+T\right].

For fixed k, there is p>0 such that, conditional on \mathcal F_S, every particle in B_r\left(z\right) belongs to \mathsf D_m with probability at least p, uniformly in m\leq l. Indeed, the centers z_m lie at a bounded distance from z, and a Brownian particle has uniformly positive probability to be in B_{r/2}\left(z_m\right) at time S+h, remain there until S+T, and survive throughout this interval. Conditional on \mathcal F_S, these events are independent for the particles in B_r\left(z\right). Hence, on \left\{\eta_S\left(B_r\left(z\right)\right)\geq H\right\}, each \left|\mathsf D_m\right| stochastically dominates \operatorname{Bin}\left(H,p\right). Since n is fixed, a binomial lower-tail bound and a union bound give \gamma'>0 such that, for all H sufficiently large,

We now set up the k applications of Lemma 2. Set

for j\leq k in the first definition and j\leq k-1 in the last. Thus t_0=S+h, t_k=S+T, and M_j is a lower bound on population size after j growth steps, with M_k=\frac{s_h}{2}\lambda^kH+O(1). If \eta_{t_j}\left(B_j\right)\geq M_j then let \mathsf P_j be an \mathcal F_{t_j}-measurable set of M_j particles in B_j; otherwise let \mathsf P_j=\varnothing. Let \mathsf C_j be the set \mathsf C from Lemma 2 corresponding to \mathsf P_j on the interval \left[t_j,t_{j+1}\right],

These are precisely the corpses whose consumption is required in the jth growth step.

We claim that on \mathcal{D} every corpse in each \mathsf{C}_j is \alpha-consumed by time t_{j+1}. Indeed, if \mathsf{q}\in\mathsf C_j, then its death location belongs to B_{j+1}\subseteq B, so there are at least n particles within distance r of the corpse throughout \left[d_{\mathsf{q}},t_{j+1}\right]. By (?) and the choice of n, the corpse \mathsf{q} is \alpha-consumed by time t_{j+1}.

Now we start the iteration. Conditional on \mathcal F_S, each particle in B_r\left(z\right) survives until S+h and moves by at most r/4 during \left[S,S+h\right] with probability s_h, independently of the other particles. Every such particle belongs to B_0 at time t_0. Hence, on \left\{\eta_S\left(B_r\left(z\right)\right)\geq H\right\}, \eta_{t_0}\left(B_0\right) stochastically dominates \operatorname{Bin}\left(H,s_h\right). Thus a Chernoff bound gives \gamma''>0 such that

For H sufficiently large, M_j\geq M_* for every j\leq k. Whenever \eta_{t_j}\left(B_j\right)\geq M_j, Lemma 2, applied to \mathsf P_j, gives

for some \gamma'''>0. To make the iteration explicit, let \mathcal A_j\mathrel{\mathop:}=\bigcap_{i=0}^j\left\{\eta_{t_i}\left(B_i\right)\geq M_i\right\}. On \mathcal{D}, every corpse in each \mathsf C_j is \alpha-consumed by t_{j+1} and on \mathcal{A}_j the ball B_j contains at least M_j particles. Hence,

so the probability that the jth growth step fails is exponentially small in H. Thus,

for some \gamma''''>0 and all H sufficiently large. We have

Therefore, by (?), (?), and (?),

for some \gamma>0 and all H sufficiently large.

It remains to choose k and z'. Since M_{j+1}\geq\lambda M_j-1, we have M_k \geq \lambda^kM_0-\frac{\lambda^k}{\lambda-1} \geq \frac{s_h}4\lambda^kH for all H sufficiently large. The ball B is covered by the balls B_{r/2}\left(z_i\right)\subseteq B_r\left(z_i\right), i\leq l, with l\leq C(1+k)^d and C depending only on d, r and R. Since \lambda>1, we can fix k, depending only on the model parameters, such that \frac{s_h}{4}\lambda^k \geq 2C(1+k)^d. On \left\{\eta_{S+T}\left(B\right)\geq M_k\right\} some z_i then satisfies

Let z' be z_i for the smallest such i, and z'\mathrel{\mathop:}= z_1 if there is none. Then z' is \mathcal F_{S+T}-measurable and

We first treat the case of one particle \mathsf p and \mu_0\left(\mathds{R}^d\right)>0. Then there exists an r/2-ball B' such that \mu_0\left(B'\right)\geq \varepsilon for some \varepsilon>0. Let B be an r-ball with the same center as B'. For any x,y\in B' we have K\left(x-y\right)\geq k_0. Consider the event that \mathsf p enters B' by time \frac{T}{2} without branching, does not leave B' before T, survives until T, branches exactly H-1 times while in B', and its offspring never leave B', do not branch and do not die before T. Because K is bounded, on this event we have \left(K*\eta_t\right)\left(y\right)\leq H\left\VertK\right\Vert_\infty, so by (?) the resource available in B' stays uniformly bounded below throughout \left[0,T\right]:

Thus the rate at which \mathsf p branches while in B' is at least \theta k_0\varepsilon e^{-H\left\VertK\right\Vert_\infty T}, and it is at most \theta\left\VertK\right\Vert_\infty\mu_0\left(\mathds{R}^d\right), since on this event there are no deaths in \left[0,T\right], so \mu_t\left(\mathds{R}^d\right)\leq \mu_0\left(\mathds{R}^d\right). Since the particle motions are independent of the branching and of each other, the event has positive probability. On it, \eta_T\left(B\right)\geq\eta_T\left(B'\right)=H.

The second case, \eta_0\left(\mathds{R}^d\right)\geq2, reduces to the first with positive probability, via the event that all initial particles but one die before any branching.