betaFeedback

WorksMathematics

The Contact Process with Temporary Immunity on Z\mathbb{Z}

Abstract

We develop a regeneration scheme for the one-dimensional contact process with temporary immunity, also known as the spatial SIRS process, a non-attractive interacting particle system. The key ingredient is a one-sided space-time shielding mechanism, constructed from infinite infection paths of an auxiliary process. These paths separate the evolution of the infection front from the configuration and graphical history left behind, allowing us to construct regeneration times without relying on attractiveness. The same shielding mechanism also provides coupling arguments for processes started from different initial configurations. Using this approach, we establish a law of large numbers and a functional central limit theorem for the right edge, existence and uniqueness of an invariant law for the process seen from the edge, and complete convergence. We also prove survival from every nonempty finite infected set for sufficiently large immunity-loss rates and derive an explicit extinction criterion.

Introduction

Regeneration techniques provide a powerful approach to the study of stochastic processes by identifying times at which the relevant future evolution can be separated from the preceding history. In the study of infection fronts, such methods have been developed through break-point constructions and coupling arguments for the classical contact process and related models; see, for instance, [4, 8, 14, 19, 22]. These constructions exploit graphical properties that allow the future evolution of the front to be separated from its past.

The contact process with temporary immunity (CPTI), also known as the spatial SIRS process, provides a natural setting in which constructing such a separation is particularly challenging. In this model, infected individuals transmit infection to susceptible neighbors and, upon recovery, enter a temporary immune state before becoming susceptible again. Thus, previous infections leave behind immune sites that can obstruct subsequent transmission. Although the CPTI is Markovian, the legality of an infection path in its graphical construction depends on immunity intervals generated by earlier infections. Moreover, the model is non-attractive, so the monotone couplings underlying classical arguments are no longer available. The difficulty is therefore to recover a suitable separation from the past without relying on attractiveness.

In this paper, we develop a geometric alternative to monotone coupling based on a one-sided space-time shielding mechanism. Suitable infinite infection paths of an auxiliary process act as barriers that isolate the evolution on one side from modifications of the configuration and graphical history on the other. This allows us to replace the graphical history behind a shielding path without affecting the subsequent evolution of the infection front. By exploiting this property, we construct independent regenerative cycles for the front, even though the full process does not regenerate.

Model and historical background

The CPTI is a continuous-time Markov process on {−1,0,1}Z\{-1,0,1\}^{\mathbb{Z}}, depending on two parameters λ,μ>0\lambda,\mu> 0. The states −1-1, 00, and 11 represent immune, susceptible, and infected individuals, respectively. For a configuration ξ∈{−1,0,1}Z\xi\in\{-1,0,1\}^{\mathbb{Z}}, let

nξ(x):=#{y∈Z:∣y−x∣=1, ξ(y)=1}n_{\xi}(x) := \#\{y \in\mathbb{Z} : |y-x| = 1,\ \xi(y) = 1\}

denote the number of infected neighbors of xx. The transition rates at site xx are

0→1 at rate λnξ(x),1→−1 at rate 1,−1→0 at rate μ.0 \to1 \text{ at rate } \lambda n_{\xi}(x), \qquad1 \to-1 \text{ at rate } 1, \qquad-1 \to0 \text{ at rate } \mu.

Thus, λ\lambda is the infection parameter and μ\mu is the rate at which immunity is lost. In the limiting case μ=∞\mu= \infty, the model reduces to the classical contact process, while μ=0\mu= 0 corresponds to permanent immunity, yielding a spatial epidemic model of SIR type.

The presence of temporary immunity introduces a competition between the propagation of infection and the creation of immune sites. In particular, increasing the infection parameter may facilitate the spread of infection while also producing more immune sites that can obstruct its subsequent propagation. As a consequence, the CPTI lacks the usual attractiveness property, and monotonicity of its survival probability with respect to the parameters is not known.

The CPTI provides a natural setting for studying the interplay between spatial growth and local memory. Infection promotes its own propagation through transmission, but also creates temporary obstacles that may hinder subsequent infections. Understanding how these competing effects influence survival and large-scale growth is a central question for the model. Unlike contact processes in independently evolving random environments, such as those studied in [3, 18, 20], the immune environment is generated by the infection itself. This also distinguishes the CPTI from the renewal and dynamic-edge models considered in [12], where the environmental restrictions evolve independently of the infection. The resulting feedback between infection and immunity introduces a history dependence that is absent from these models.

The CPTI was introduced in [6], where the existence of a nontrivial invariant measure was established in a suitable parameter regime. Subsequently, [1 Theorem 1.2] extended the coexistence results to every dimension. In the same paper, they also established a complete convergence theorem in one dimension for a related but different model, in which the immunity-loss rate is proportional to the number of healthy nearest neighbors [1 Theorem 1.1].

By comparison with the classical contact process, extinction from finite initial infections holds whenever the infection rate is at most the critical value of the classical contact process, independently of the immunity-loss rate. Moreover, [23] showed that, for every fixed infection rate, the one-dimensional CPTI dies out from finite initial infections when the immunity-loss rate is sufficiently small.

A fundamental question, however, remained open in one dimension: does the CPTI survive from a finite initial infection for some choice of parameters? The existence of a nontrivial invariant measure does not by itself imply such survival in the absence of attractiveness. In dimensions d≥2d \ge2, survival from finite infections was established for sufficiently large infection rates in [23]. More recently, [10] showed that weak survival implies strong survival for the one-dimensional CPTI, while [11] obtained survival and extinction criteria in higher dimensions by adapting a block construction originally developed for the contact process with avoidance.

Numerical studies of the CPTI on lattices include simulations and mean-field approximations [13]. On finite graphs, extinction times have been studied for several graph classes [7, 15].

Beyond the existence of nontrivial invariant measures, the long-time behavior of the one-dimensional CPTI has remained largely unexplored. Questions concerning the propagation speed of the infection front, its fluctuations, and convergence to equilibrium require methods capable of controlling the dependence on past infections. In the absence of attractiveness, standard coupling and regeneration arguments cannot be applied directly. The shielding mechanism developed here overcomes this difficulty by isolating the evolution relevant to the infection front from the graphical history behind it.

Main results

For an initial configuration ξ∈{−1,0,1}Z\xi\in\{-1,0,1\}^{\mathbb{Z}}, let Pλ,μξ\mathbb{P}_{\lambda,\mu}^{\xi} denote the law of the CPTI with parameters (λ,μ)(\lambda,\mu) started from ξ\xi, and let Eλ,μξ\mathbb{E}_{\lambda,\mu}^{\xi} denote the corresponding expectation. Throughout § 2-§ 5 and Appendix § 7, we fix λ\lambda and suppress its dependence from the notation. When the initial configuration is clear from the context, we also omit it from the notation.

For A⊂ZA \subset\mathbb{Z}, let ξA\xi^{A} denote the process started with the sites in AA infected and all other sites susceptible, and write

ItA:={x∈Z:ξtA(x)=1}.I_t^A := \{x \in\mathbb{Z} : \xi_t^A(x) = 1\}.

Define the extinction time

τA:=inf⁡{t≥0:ItA=∅}.\tau^A := \inf\{t \ge0 : I_t^A = \varnothing\}.

We say that the process survives from AA if τA=∞\tau^A = \infty, and set

θ(λ,μ):=Pλ,μ(τ{0}=∞).\theta(\lambda,\mu) := \mathbb{P}_{\lambda,\mu}(\tau^{\{0\}} = \infty).

Let λc(∞)\lambda_c(\infty) denote the critical infection rate of the classical contact process. For fixed μ>0\mu> 0, define

λc(μ):=inf⁡{λ>0:θ(λ,μ)>0},\lambda_c(\mu) := \inf\{\lambda> 0 : \theta(\lambda,\mu) > 0\},

with the convention inf⁡∅=∞\inf\varnothing= \infty.

Similarly, for fixed λ>λc(∞)\lambda> \lambda_c(\infty), define

μc(λ):=inf⁡{μ>0:θ(λ,μ)>0}\mu_c(\lambda) := \inf\{\mu> 0 : \theta(\lambda,\mu) > 0\}

and

μ‾c(λ):=sup⁡{μ>0:θ(λ,μ)=0}.\overline{\mu}_c(\lambda) := \sup\{\mu> 0 : \theta(\lambda,\mu) = 0\}.

Since monotonicity of θ(λ,μ)\theta(\lambda,\mu) in the parameters is not known, these quantities are not asserted to define a sharp phase transition curve.

Our first result gives an explicit sufficient condition for extinction from finite initial infections, based on the one-dimensional geometry of infection genealogies.

Theorem 1.1 (Extinction criterion). Let λ,μ>0\lambda,\mu> 0. If

μλ(1+2λ)(1+μ)(1+λ)<1,(1)\frac{\mu\lambda(1+2\lambda)}{(1+\mu)(1+\lambda)} < 1, \tag*{(1)}

then, for every finite A⊂ZA \subset\mathbb{Z},

Pλ,μ(τA<∞)=1.\mathbb{P}_{\lambda,\mu}(\tau^{A}<\infty)=1.

Our next result settles the question of survival from finite initial infections in one dimension provided the immunity-loss rate is sufficiently large.

Theorem 1.2 (Survival for large immunity-loss rate). For every λ>λc(∞)\lambda>\lambda_c(\infty), there exists μ0=μ0(λ)<∞\mu_0=\mu_0(\lambda)<\infty such that, for every μ≥μ0\mu\geq\mu_0 and every nonempty finite A⊂ZA\subset\mathbb{Z},

Pλ,μ(τA=∞)>0.\mathbb{P}_{\lambda,\mu}(\tau^{A}=\infty)>0.

In particular, for every λ>λc(∞)\lambda>\lambda_c(\infty),

μc(λ)<∞.\mu_c(\lambda)<\infty.

Together, these results extend the known one-dimensional phase diagram in two directions: Theorem 1.1 provides an explicit extinction criterion, while Theorem 1.2 establishes a regime of survival from finite initial infections.

Our main technical contribution is a one-sided space-time shielding mechanism that replaces key uses of attractiveness in classical coupling and regeneration arguments. The shielding paths allow us to modify the graphical history on one side of a space-time barrier without affecting the evolution on the other. This separation makes it possible to construct independent regenerative cycles for the infection front without regenerating the full configuration. The same mechanism also yields couplings between processes started from different initial configurations, leading to convergence to equilibrium.

Set

Xe:={ξ∈{−1,0,1}Z:ξ(0)=1, ξ(x)≠1 for all x>0, #{x<0:ξ(x)=1}=∞}.\mathcal{X}^{\mathrm e}:=\left\{\xi\in\{-1,0,1\}^{\mathbb{Z}}:\xi(0)=1,\ \xi(x)\ne1\ \text{for all }x>0,\ \#\{x<0:\xi(x)=1\}=\infty\right\}.

Thus, for ξ∈Xe\xi\in\mathcal{X}^{\mathrm e}, the origin is the rightmost infected site and there are infinitely many infected sites to its left; susceptible and immune sites are otherwise allowed arbitrarily.

For ξ∈Xe\xi\in\mathcal{X}^{\mathrm e}, under Pλ,μξ\mathbb{P}^{\xi}_{\lambda,\mu}, let

rt:=sup⁡{x∈Z:ξt(x)=1}r_t:=\sup\{x\in\mathbb{Z}:\xi_t(x)=1\}

denote the position of the rightmost infected site at time tt.

Our first application of the regenerative structure is a law of large numbers for the position of the right edge.

Theorem 1.3 (Law of large numbers for the right edge). Fix λ>λc(∞)\lambda>\lambda_c(\infty). Then, for every μ≥μ0(λ)\mu\geq\mu_0(\lambda), there exists a deterministic constant v=v(λ,μ)>0∗suchthat,forevery∗v=v(\lambda,\mu)>0 *such that, for every*\xi\in\mathcal{X}^{\mathrm e}$,

rtt⟶vPλ,μξ-almost surely as t→∞.(2)\frac{r_t}{t}\longrightarrow v\qquad\mathbb{P}^{\xi}_{\lambda,\mu}\text{-almost surely as }t\to\infty. \tag*{(2)}

In particular, the asymptotic speed does not depend on the initial configuration ξ\xi.

If ξ\xi is the configuration given by

ξ(x)={1,x≤0,0,x>0,\xi(x)= \begin{cases} 1, & x\leq0,\\ 0, & x>0, \end{cases}

then the convergence in (2) also holds in L1(Pλ,μξ)L^1(\mathbb{P}^{\xi}_{\lambda,\mu}).

As the immunity-loss rate tends to infinity, the asymptotic velocity approaches that of the classical contact process.

Proposition 1.4 (Classical limit of the speed). Fix λ>λc(∞)\lambda> \lambda_c(\infty). Then

lim⁡μ→∞v(λ,μ)=vCP(λ),\lim_{\mu\to\infty} v(\lambda,\mu) = v_{\mathrm{CP}}(\lambda),

where vCP(λ)v_{\mathrm{CP}}(\lambda) is the right-edge speed of the classical contact process with infection rate λ\lambda.

The regenerative structure also yields a functional central limit theorem for the fluctuations of the right edge around its deterministic linear growth.

Theorem 1.5 (Functional central limit theorem for the right edge). Fix λ>λc(∞)\lambda> \lambda_c(\infty) and let μ≥μ0(λ)\mu\ge\mu_0(\lambda). Let v=v(λ,μ)v = v(\lambda,\mu) be the asymptotic speed from Theorem 1.3. Then there exists a deterministic constant

σ2=σ2(λ,μ)∈(0,∞)\sigma^2 = \sigma^2(\lambda,\mu) \in(0,\infty)

such that, for every ξ∈Xe\xi\in\mathcal{X}^{\mathrm{e}}, under Pλ,μξ\mathbb{P}^{\xi}_{\lambda,\mu},

(rnt−vntn)t≥0⟹(σBt)t≥0,n→∞,\left(\frac{r_{nt}-vnt}{\sqrt{n}}\right)_{t\geq0} \Longrightarrow(\sigma B_t)_{t\geq0}, \qquad n \to\infty,

in D([0,∞),R)D([0,\infty),\mathbb{R}) equipped with the Skorokhod J1J_1 topology, where (Bt)t≥0(B_t)_{t\geq0} is a standard Brownian motion.

Beyond the motion of the infection front, we study the configuration of the process as seen from its right edge. The shielding construction also provides a coupling mechanism: infinite infection paths form barriers that prevent differences in the initial configurations from influencing the evolution on the protected side.

For a nonempty configuration ξt\xi_t with a finite rightmost infected site, define

Ξt(x):=ξt(rt+x),x∈Z.\Xi_t(x) := \xi_t(r_t+x), \qquad x \in\mathbb{Z}.

Theorem 1.6 (Invariant law seen from the right edge). Fix λ>λc(∞)\lambda> \lambda_c(\infty) and let μ≥μ0(λ)\mu\geq\mu_0(\lambda). The CPTI seen from its right edge admits a unique invariant probability measure νe\nu^{\mathrm{e}} on Xe\mathcal{X}^{\mathrm{e}}. Moreover, for every ξ∈Xe\xi\in\mathcal{X}^{\mathrm{e}},

Pλ,μξ(Ξt∈⋅)⟹νe,t→∞.(3)\mathbb{P}^{\xi}_{\lambda,\mu}(\Xi_t \in\cdot) \Longrightarrow\nu^{\mathrm{e}}, \qquad t \to\infty. \tag*{(3)}

This result is the CPTI analogue of the edge convergence theorem for the one-dimensional supercritical contact process [8]. Its proof follows the finite-window approach of [4], using the shielding property to show that the distribution in every fixed window behind the edge becomes asymptotically independent of the initial configuration.

Remark 1.1. The methods developed here can be extended to yield a uniform version of the preceding convergence result. More precisely, for every 0<c<v0 < c < v and every bounded local function ff,

lim⁡t→∞sup⁡A⊂Z, 0<∣A∣<∞diam⁡(A)≤ct∣Eλ,μA[f(Ξt)∣τA>t]−∫f dνe∣=0.\lim_{t\to\infty}\sup_{\substack{A\subset\mathbb{Z},\ 0<|A|<\infty\\ \operatorname{diam}(A)\leq ct}} \left|\mathbb{E}^{A}_{\lambda,\mu}\left[f(\Xi_t)\mid\tau^A>t\right]-\int f\,d\nu^{\mathrm{e}}\right|=0.

We refer to Remark 5.1 below for a discussion of a possible strategy to establish this uniform convergence.

Let 0∈{−1,0,1}Z\mathbf{0} \in\{-1,0,1\}^{\mathbb{Z}} denote the all-susceptible configuration,

0(x)=0,x∈Z.\mathbf{0}(x)=0,\qquad x\in\mathbb{Z}.

This configuration is absorbing, and we write δ0\delta_0 for the corresponding point mass.

The shielding mechanism also applies to coupling in fixed spatial windows. By constructing barriers that propagate in opposite directions, we can isolate finite regions from differences in the surrounding configurations and their infection histories. This two-sided coupling provides the key ingredient for our complete convergence theorem.

Theorem 1.7 (Complete convergence). Fix λ>λc(∞)\lambda>\lambda_c(\infty) and let μ≥μ0(λ)\mu\geq\mu_0(\lambda). There exists a nontrivial translation-invariant invariant probability measure ν\nu for the CPTI such that, for every nonempty A⊂ZA\subset\mathbb{Z},

Pλ,μ(ξtA∈⋅)⟹Pλ,μ(τA<∞)δ0+Pλ,μ(τA=∞)ν,t→∞.\mathbb{P}_{\lambda,\mu}(\xi_t^A\in\cdot)\Longrightarrow\mathbb{P}_{\lambda,\mu}(\tau^A<\infty)\delta_0+\mathbb{P}_{\lambda,\mu}(\tau^A=\infty)\nu,\qquad t\to\infty.

Moreover,

Pλ,μ(ξtA∈⋅∣τA=∞)⟹ν,t→∞.\mathbb{P}_{\lambda,\mu}(\xi_t^A\in\cdot\mid\tau^A=\infty)\Longrightarrow\nu,\qquad t\to\infty.

Graphical construction and auxiliary toy process

The shielding construction relies on an auxiliary process introduced in [1], whose infection paths provide the space-time barriers used throughout the paper. We first recall the graphical construction of the CPTI and then introduce this auxiliary process.

We construct the process from a Harris graphical representation on the whole time line. For each ordered nearest-neighbor pair (x,y)(x,y), ∣x−y∣=1|x-y|=1, we put infection arrows from xx to yy at the times of a Poisson process of rate λ\lambda on R\mathbb{R}. For each site xx, we put recovery marks at the times of a Poisson process of rate 11 on R\mathbb{R}. To each recovery mark (x,t)(x,t) we attach an independent exponential random variable I(x,t)I(x,t) with parameter μ\mu. All Poisson processes and all variables I(x,t)I(x,t) are independent.

If the initial configuration contains immune sites, then for each xx with ξ0(x)=−1\xi_0(x)=-1 we assign an independent residual immunity time Ix∼Exp⁡(μ)I_x\sim\operatorname{Exp}(\mu), independent of the Harris system and of the residual immunity times of the other initially immune sites. At time IxI_x, the site xx becomes susceptible.

Given an initial configuration ξ0\xi_0, the process (ξt)t≥0(\xi_t)_{t\geq0} is obtained by reading the graphical construction after time 00 in increasing time order. At a recovery mark (x,t)(x,t), if ξt−(x)=1\xi_{t-}(x)=1, the mark is effective: we set ξt(x)=−1\xi_t(x)=-1, and the site xx remains immune during the interval [t,t+I(x,t))[t,t+I(x,t)). At time t+I(x,t)t+I(x,t), the site returns to state 00. If ξt−(x)≠1\xi_{t-}(x)\ne1, the recovery mark has no effect, and its attached interval is not activated.

We distinguish between two notions of infection path. A graphical arrow path is a piecewise constant nearest-neighbor space-time path that follows infection arrows in their direction and does not cross recovery marks at its current site.

Such a path is a legal CPTI infection path if it starts from an infected space-time point and never enters an immune site. It may enter an already infected site, in which case the paths coalesce. The infected sites at any given time are precisely those reached by legal infection paths from the initial infected set. Unlike graphical arrow paths, legality depends on the immunity intervals activated by the CPTI evolution.

We now introduce a stationary version of the auxiliary process of [1], which we call the toy process. The key idea is to make the blocking mechanism independent of the infection history, thereby recovering an additive structure.

The toy process uses the same infection arrows, recovery marks, and attached exponential variables as the CPTI, but treats immunity intervals differently. In the CPTI, an interval is activated only when its recovery mark corresponds to an actual recovery of an infected site. In the toy process, every recovery mark creates a blocking interval. Thus, a recovery mark at (x,s)(x,s) blocks the site xx during [s,s+I(x,s))[s,s+I(x,s)), regardless of whether the site is infected at time ss.

Since the recovery marks are defined on all of R\mathbb{R}, blocking intervals born before time 00 may still be active at that time. We write

βt(x):=#{s≤t:there is a recovery mark at (x,s) and s+I(x,s)>t},t∈R,\beta_t(x) := \#\{s \le t : \text{there is a recovery mark at }(x,s)\text{ and }s+I(x,s)>t\}, \qquad t \in\mathbb{R},

for the number of active blocking intervals at site xx at time tt. The site is toy-blocked precisely when βt(x)>0\beta_t(x)>0.

For each x∈Zx \in\mathbb{Z}, the process (βt(x))t∈R(\beta_t(x))_{t\in\mathbb{R}} is time-stationary. Moreover, for each t∈Rt \in\mathbb{R}, the random variables (βt(x))x∈Z(\beta_t(x))_{x\in\mathbb{Z}} are i.i.d. with

βt(x)∼Poisson⁡(1/μ)\beta_t(x) \sim\operatorname{Poisson}(1/\mu)

for each fixed tt. Conditional on β0(x)=k\beta_0(x)=k, the kk blocking intervals active at time 00 have independent residual lifetimes with distribution Exp⁡(μ)\operatorname{Exp}(\mu). Equivalently, the blocking field consists of independent immigration–death processes, with immigration rate 11 and death rate μ\mu per blocker.

Throughout the paper, we reserve the letter η\eta for the toy process. Given an initial infected set A⊆ZA \subseteq\mathbb{Z}, let ηtA\eta_t^A denote the toy infected set at time tt, started from AA at time 00. Sites in AA are initially infected regardless of the value of β0\beta_0; active blockers prevent only subsequent infections into blocked sites.

A toy infection is killed at a recovery mark at its current site. An infection arrow from xx to yy at time tt transmits infection if xx is infected immediately before tt and βt−(y)=0\beta_{t-}(y)=0. Equivalently, a toy infection path is a graphical arrow path that never enters a blocked site after its starting time. The path representation immediately implies that the toy process is additive.

The joint process (ηtA,βt)(\eta_t^A,\beta_t) is Markovian, although ηtA\eta_t^A alone is not, since its evolution depends on the blocking field. For survival probabilities, we condition the initial infected seed to be unblocked; in particular, a singleton started at (0,0)(0,0) is considered under β0(0)=0\beta_0(0)=0.

The crucial comparison with the CPTI is that every toy infection path starting from a CPTI-infected site is also a legal CPTI infection path, provided the two processes use the same graphical construction and all immunity intervals inherited from the initial configuration are included among the toy-blocking intervals. Thus, despite the lack of attractiveness of the CPTI, the toy process provides infection paths that remain valid under the original dynamics.

Figure 1 illustrates the CPTI and the toy process under a common graphical construction.

Figure 1. Graphical construction of (a) the CPTI and (b) the toy process. Solid arrows represent infection attempts, red crosses recovery marks, dotted blue segments immunity or blocking intervals, and dashed arrows blocked attempts. Highlighted trajectories are legal infection paths. The intervals at x+3x+3 illustrate blockers inherited from the past and overlapping blocking intervals.

Overview of the approach

The autonomy of the toy-blocking field is crucial for our approach. Unlike CPTI immunity, its evolution does not depend on the infection history, yielding an additive graphical representation. For sufficiently large immunity-loss rates, recent blocking intervals form a small perturbation of the classical contact process, while older intervals constitute a sparse field of obstacles. A block construction then yields survival and growth estimates for toy infection paths. These paths play a dual role: they transfer survival to the CPTI and provide the shielding barriers and quantitative estimates needed for regeneration.

The next difficulty is to turn survival into a regenerative structure. The advance of the CPTI right edge provides opportunities to explore fresh graphical information and search for infinite toy infection paths. By arranging these trials suitably, we eventually obtain such a path, which can be used to separate the evolution of the front from its past. The selection of the leftmost infinite path is essential: its extremal character yields the one-sided shielding property needed to separate the relevant graphical information.

To understand this shielding mechanism, observe that any influence from the left of the barrier must cross it before reaching its right side. Since the barrier is a legal infection path for all CPTI realizations under consideration, its sites are infected in every such realization. Consequently, differences in the configuration or graphical history on the left cannot propagate across the barrier to affect the evolution on its right. Although this does not provide a global comparison principle, it gives a one-sided protection property that is sufficient for both regeneration and coupling.

This separation allows us to replace the graphical history on the left of the barrier by a suitable completion without changing the subsequent evolution of the right edge. The resulting construction yields regeneration times with independent and identically distributed increments after the first cycle. The block estimates yield the moment bounds needed to control both the regeneration increments and the fluctuations within each cycle, leading to the law of large numbers and the functional central limit theorem.

The shielding mechanism also replaces monotone coupling in the study of long-time behavior. Common shielding paths force CPTI evolutions started from different configurations to agree in expanding spatial regions, despite the absence of attractiveness. This leads to convergence to a unique invariant law for the process seen from the edge and, by applying shielding from both spatial directions, to complete convergence.

Finally, our extinction criterion follows from a separate genealogical argument specific to one dimension. The nearest-neighbor geometry allows return infections to be assigned to previously traversed infection edges, whose contributions can be controlled by a branching-process comparison. This yields a subcritical bound on the expected number of infections in successive generations.

Organization of the paper

Section 2 establishes survival and growth properties for the toy process, yielding Theorem 1.2. In Section 3, we develop the shielding and regeneration constructions and establish the renewal structure. Section 4 derives the regenerative limit theorems and the convergence of the edge speed, using exponential moment bounds whose proof is deferred to the Appendix § 7. Section 5 establishes convergence to the invariant law and complete convergence through shielding couplings. Finally, we prove the extinction criterion in Theorem 1.1 in Section 6.

Survival and growth of the toy process

We use a block construction to establish survival and growth of the toy process and obtain the estimates needed for regeneration. The key observation is that, for large immunity-loss rates, recent blocking intervals produce only a small perturbation of the classical contact process, allowing us to approximate them by classical blocks. Although blocking intervals originating in the distant past introduce long-range dependence, they are sufficiently rare to be controlled as sparse obstacles through a Peierls argument.

The cutoff block construction

The toy-blocking field of Section 1.3 is stationary in time, so blocking intervals born before time 0 may affect the process afterward. We first separate recent blocking intervals from older ones.

Fix integers HH, ℓ\ell, KK, with 1≤ℓ≤H1 \le\ell\le H and K>HK > H, and fix a time scale S>0S > 0. For j∈Zj \in\mathbb{Z}, set

Ij:=jK+[0,H].I_j := jK + [0,H].

We say that a configuration contains a seed in IjI_j if every subinterval of IjI_j of length ℓ\ell contains at least one infected site. This condition is increasing in the infected set.

Let

L:={(j,n)∈Z×Z≥0:j+n is even}\mathcal{L} := \{(j,n) \in\mathbb{Z} \times\mathbb{Z}_{\ge0} : j+n \text{ is even}\}

be the oriented lattice, with edges from (j,n)(j,n) to (j±1,n+1)(j \pm1,n+1). A block (j,n)(j,n) propagates a seed in IjI_j at time nSnS to seeds in Ij±1I_{j\pm1} at time (n+1)S(n+1)S.

Choose a finite interval W0⊂ZW_0 \subset\mathbb{Z}, and write Wj:=jK+W0W_j := jK + W_0. The space-time box used by block (j,n)(j,n) is

Bj,n:=Wj×[nS,(n+1)S].B_{j,n} := W_j \times[nS,(n+1)S].

Let Gj,nclG^{\mathrm{cl}}_{j,n} be the event that every seed in IjI_j is connected to both neighboring seed intervals by classical contact paths with parameter λ\lambda inside Bj,nB_{j,n}. Note that the field of block indicators is finite-range dependent. We choose W0W_0 large enough so that Gj,nclG^{\mathrm{cl}}_{j,n} is determined by the graphical construction inside Bj,nB_{j,n}.

Since λ>λc(∞)\lambda> \lambda_c(\infty), the classical contact process admits a supercritical block construction. For every ε>0\varepsilon> 0, the parameters can be chosen so that the block events Gj,nclG^{\mathrm{cl}}_{j,n} satisfy

P(Gj,ncl)≥1−ε.\mathbb{P}(G^{\mathrm{cl}}_{j,n}) \ge1-\varepsilon.

Fix R>0R > 0. Let Dj,n≤RD^{\le R}_{j,n} be the event that an infection arrow in Bj,nB_{j,n} is suppressed by a blocking interval of age at most RR, born at a recovery mark (x,u)(x,u) at the arrow's target $x, with 0≤t−u≤R0 \le t-u \le R, where tt is the arrow time. We define Dj,n>RD^{>R}_{j,n} analogously, but with blocking intervals with age at least RR.

At a fixed site xx and time tt, the expected number of active blocking intervals of age at most RR is

∫0Re−μa da≤1μ.\int_0^R e^{-\mu a}\,\mathrm{d}a \le\frac{1}{\mu}.

Since infection arrows enter each site at rate 2λ2\lambda, we obtain

Pμ(Dj,n≤R)≤2λ∣W0∣Sμ.\mathbf{P}_{\mu}\left(D^{\le R}_{j,n}\right) \le\frac{2\lambda|W_0|S}{\mu}.

Define the cutoff toy-good event by

G~j,ntoy,R:=Gj,ncl∩(Dj,n≤R)c.\widetilde{G}^{\mathrm{toy},R}_{j,n}:=G^{\mathrm{cl}}_{j,n}\cap\left(D^{\le R}_{j,n}\right)^{\mathrm{c}}.

For fixed RR, its indicator field is finite-range dependent, and

Pμ(G~j,ntoy,R)≥1−ε−2λ∣W0∣Sμ.\mathbf{P}_{\mu}\left(\widetilde{G}^{\mathrm{toy},R}_{j,n}\right)\ge1-\varepsilon-\frac{2\lambda|W_0|S}{\mu}.

Thus its marginal probability can be made arbitrarily close to one by choosing ε\varepsilon small and then μ\mu large.

Sparse old-tail obstacles

We now control suppressions caused by blocking intervals older than RR. Rather than track suppressed arrows directly, we introduce a larger obstacle event. Let Oj,nO_{j,n} be the event that some recovery mark (x,u)(x,u), with x∈Wjx\in W_j and I(x,u)>RI(x,u)>R, satisfies

[u+R,u+I(x,u)]∩[nS,(n+1)S]≠∅.[u+R,u+I(x,u)]\cap[nS,(n+1)S]\ne\varnothing.

Thus Oj,nO_{j,n} records old blocking tails intersecting the block. In particular, Dj,n>R⊆Oj,nD^{>R}_{j,n}\subseteq O_{j,n}, so it suffices to control these obstacles.

We need the following uniform estimate.

Lemma 2.1 (Sparse old-tail obstacles). Fix the block parameters HH, ℓ\ell, KK, SS, W0W_0 and the cutoff R>0R>0. There exists ρE=ρE(μ,R)\rho_{\mathrm{E}}=\rho_{\mathrm{E}}(\mu,R), with ρE(μ,R)→0\rho_{\mathrm{E}}(\mu,R)\to0 as μ→∞\mu\to\infty for fixed RR, such that, for every finite A⊂LA\subset\mathcal{L},

Pμ(Oz occurs for every z∈A)≤ρE∣A∣.\mathbf{P}_{\mu}\left(O_z\text{ occurs for every }z\in A\right)\le\rho_{\mathrm{E}}^{|A|}.

Proof. We dominate the old tails by a Poisson stick process. For each recovery mark (x,u)(x,u) with I(x,u)>RI(x,u)>R, set

V:=u+R,L:=I(x,u)−R.V:=u+R,\qquad L:=I(x,u)-R.

By Poisson thinning and the memoryless property, the points (x,V)(x,V) form a Poisson process of intensity e−μRe^{-\mu R} on each vertical line, with independent lengths L∼Exp⁡(μ)L\sim\operatorname{Exp}(\mu).

A stick (x,V,L)(x,V,L) affects the set of block sites

F(x,V,L):={(j,n)∈L:x∈Wj, [V,V+L]∩[nS,(n+1)S]≠∅}.F(x,V,L):=\left\{(j,n)\in\mathcal{L}:x\in W_j,\ [V,V+L]\cap[nS,(n+1)S]\ne\varnothing\right\}.

We call F(x,V,L)F(x,V,L) the block footprint of the stick.

A stick affects only a bounded number of block columns, depending on W0W_0 and KK, and at most 2+L/S2+L/S time levels, up to a change of constants. Since L∼Exp⁡(μ)L\sim\operatorname{Exp}(\mu), its block footprint has an exponential size tail, uniformly for large μ\mu. Let νμ\nu_\mu be the intensity measure of the block footprints. The stick intensity is e−μRe^{-\mu R}, and footprint sizes have uniformly exponential tails. Hence we can choose ρE=ρE(μ,R)∈(0,1)\rho_E=\rho_E(\mu,R)\in(0,1), with ρE→0\rho_E\to0 as μ→∞\mu\to\infty, such that

sup⁡z∈L∫{F:z∈F}ρE−∣F∣ νμ(dF)≤1.\sup_{z\in\mathcal{L}}\int_{\{F:z\in F\}}\rho_E^{-|F|}\,\nu_\mu(\mathrm{d}F)\le1.

We prove the claim by induction on ∣A∣|A|, the case A=∅A=\varnothing being trivial. Fix z∈Az\in A. If every block in AA is hit, some stick has a footprint FF containing zz, while all blocks in A∖FA\setminus F must also be hit. By the Mecke formula and the induction hypothesis,

Pμ(Oy for all y∈A)≤∫{F:z∈F}ρE∣A∖F∣ νμ(dF)≤ρE∣A∣∫{F:z∈F}ρE−∣F∣ νμ(dF)≤ρE∣A∣.\begin{aligned} \mathbb{P}_\mu(\mathcal{O}_y\text{ for all }y\in A) &\le\int_{\{F:z\in F\}}\rho_E^{|A\setminus F|}\,\nu_\mu(\mathrm{d}F)\\ &\le\rho_E^{|A|}\int_{\{F:z\in F\}}\rho_E^{-|F|}\,\nu_\mu(\mathrm{d}F) \le\rho_E^{|A|}. \end{aligned}

This completes the induction. □

Survival and growth after old-tail deletion

We now combine the high-density, finite-range dependent cutoff field with the sparse old-tail obstacles. Deleting blocks hit by old tails recovers the full toy block construction.

Define the toy-good block after old-tail deletion by

Gj,ntoy:=G~j,ntoy,R∩Oj,nc.G^{\mathrm{toy}}_{j,n}:=\widetilde{G}^{\mathrm{toy},R}_{j,n}\cap O^c_{j,n}.

For z=(j,n)∈Lz=(j,n)\in\mathcal{L}, write

ω~zR:=1G~j,ntoy,R,ωztoy:=1Gj,ntoy.\widetilde{\omega}^{R}_z:=\mathbf{1}_{\widetilde{G}^{\mathrm{toy},R}_{j,n}},\qquad \omega^{\mathrm{toy}}_z:=\mathbf{1}_{G^{\mathrm{toy}}_{j,n}}.

We call (ω~zR)z∈L(\widetilde{\omega}^{R}_z)_{z\in\mathcal{L}} the cutoff toy-good block field. This is the field obtained after ignoring blocking intervals whose age is larger than RR. For fixed RR, the cutoff toy-good block field is finite-range dependent, with a fixed dependence range, and its marginal probability tends to one as ε↓0\varepsilon\downarrow0 and then μ→∞\mu\to\infty. Therefore, by [16], it can be coupled with an independent oriented site percolation field (Πz)z∈L(\Pi_z)_{z\in\mathcal{L}} of parameter pp such that

Πz≤ω~zRfor every z∈L,\Pi_z\le\widetilde{\omega}^{R}_z\qquad\text{for every }z\in\mathcal{L},

where pp can be chosen arbitrarily close to one.

The next proposition says that deleting the sparse old-tail obstacles does not destroy the supercritical Peierls estimates, see Figure 2.

Block renormalization with old-tail deletion

Figure 2. Block renormalization with old-tail deletion. The dashed red path is a candidate before deletion. The blue segments represent old-tail obstacles Oz\mathcal{O}_z, which delete the crossed blocks. The solid red path survives deletion.

Proposition 2.2. Suppose the cutoff field ω~R\widetilde{\omega}^{R} dominates independent oriented site percolation with parameter pp, and set q:=1−pq:=1-p. If pp is sufficiently close to one and ρE\rho_E sufficiently small, then the following estimates hold.

For every finite nonempty A0⊂LA_0\subset\mathcal{L}, let

Sblk(A0):={there is an infinite oriented path of Gtoy-blocks from A0}.S^{\mathrm{blk}}(A_0):=\{\text{there is an infinite oriented path of }G^{\mathrm{toy}}\text{-blocks from }A_0\}.

Then

Pμ(Sblk(A0))>0.\mathbb{P}_\mu\bigl(S^{\mathrm{blk}}(A_0)\bigr)>0.

Let Cn(A0)C_n(A_0) be the set of block positions reached at height nn by oriented paths of GtoyG^{\mathrm{toy}}-blocks from A0A_0, and define

Hblk(A0):=sup⁡{n:Cn(A0)≠∅}.H^{\mathrm{blk}}(A_0):=\sup\{n:C_n(A_0)\ne\varnothing\}.

Then there exist C,c>0C,c>0 such that, for all n≥1n\ge1,

Pμ(Hblk(A0)≥n,(Sblk(A0))c)≤Ce−cn.\mathbb{P}_{\mu}\left(H^{\mathrm{blk}}(A_0)\ge n,\left(S^{\mathrm{blk}}(A_0)\right)^c\right)\le Ce^{-cn}.

Define

Rnblk(A0):=sup⁡Cn(A0),R_n^{\mathrm{blk}}(A_0):=\sup C_n(A_0),

with sup⁡∅=−∞\sup\varnothing=-\infty. Then, for some κ,C,c>0\kappa,C,c>0 and all n≥1n\ge1,

Pμ(Sblk(A0),Rnblk(A0)≤max⁡A0+κn)≤Ce−cn.\mathbb{P}_{\mu}\left(S^{\mathrm{blk}}(A_0),R_n^{\mathrm{blk}}(A_0)\le\max A_0+\kappa n\right)\le Ce^{-cn}.

On Sblk(A0)S^{\mathrm{blk}}(A_0), let Γblk(A0)=(Γnblk(A0),n)n≥0\Gamma^{\mathrm{blk}}(A_0)=(\Gamma_n^{\mathrm{blk}}(A_0),n)_{n\ge0} be the leftmost infinite oriented path from A0A_0. Possibly decreasing κ\kappa, there exist C,c>0C,c>0 such that, for all n≥1n\ge1,

Pμ(Sblk(A0),Γnblk(A0)≥min⁡A0−κn)≤Ce−cn.\mathbb{P}_{\mu}\left(S^{\mathrm{blk}}(A_0),\Gamma_n^{\mathrm{blk}}(A_0)\ge\min A_0-\kappa n\right)\le Ce^{-cn}.

Proof. Call a block zz bad if Πz=0\Pi_z=0 or Oz\mathcal{O}_z occurs. Since Πz≤ω~zR\Pi_z\le\widetilde{\omega}_z^{R}, every non-bad block is a GtoyG^{\mathrm{toy}}-block. It therefore suffices to consider oriented paths of non-bad blocks.

Let A⊂LA\subset\mathcal{L} be finite. If every block in AA is bad, then for some C⊆AC\subseteq A, we have that Oz\mathcal{O}_z occurs for all z∈Cz\in C and Πz=0\Pi_z=0 for every z∈A∖Cz\in A\setminus C. By Lemma 2.1 and the product law of (Πz)(\Pi_z),

Pμ(A is entirely bad)≤∑C⊆Amin⁡{ρ∣C∣E, q∣A∣−∣C∣}≤(ρE+q)∣A∣.\begin{aligned} \mathbb{P}_{\mu}(A\text{ is entirely bad})\le\sum_{C\subseteq A}\min\left\{\rho^{|C|}E,\ q^{|A|-|C|}\right\} \\ &\le\left(\sqrt{\rho E}+\sqrt{q}\right)^{|A|}. \end{aligned}

Writing δ:=ρE+q\delta:=\sqrt{\rho E}+\sqrt{q}, we obtain

Pμ(A is entirely bad)≤δ∣A∣.(4)\mathbb{P}_{\mu}(A\text{ is entirely bad})\le\delta^{|A|}. \tag*{(4)}

By choosing pp close to one and then μ\mu large, δ\delta can be made arbitrarily small.

The remaining argument uses standard planar Peierls estimates for oriented percolation; see [5]. The essential point is that (4) applies to every prescribed dual barrier, without requiring independence of the bad-block indicators.

Absence of an infinite oriented path from A0A_0 forces a finite connected dual barrier of bad blocks separating A0A_0 from infinity. The number of such barriers of size mm is at most C0mC_0^m. Hence, for δ\delta sufficiently small,

Pμ((Sblk(A0))c)≤∑m≥m0C0mδm<1,\mathbb{P}_{\mu}\left((S^{\mathrm{blk}}(A_0))^c\right) \leq\sum_{m \geq m_0} C_0^m \delta^m < 1,

where m0≥1m_0 \geq1 is the minimal barrier size. Thus Pμ(Sblk(A0))>0\mathbb{P}_{\mu}(S^{\mathrm{blk}}(A_0)) > 0.

It remains to prove the last assertion. If the reached cluster has height at least nn but is finite, its enclosing bad barrier has size at least c0nc_0n. By (4),

Pμ(Hblk(A0)≥n, (Sblk(A0))c)≤∑m≥c0nC0mδm≤Ce−cn.\mathbb{P}_{\mu}\left(H^{\mathrm{blk}}(A_0) \geq n,\ (S^{\mathrm{blk}}(A_0))^c\right) \leq\sum_{m \geq c_0n} C_0^m \delta^m \leq Ce^{-cn}.

The same dual-barrier argument in a rightward cone gives the lower-speed estimate. For κ>0\kappa> 0 sufficiently small, on Sblk(A0)S^{\mathrm{blk}}(A_0) the event Rnblk(A0)≤max⁡A0+κnR_n^{\mathrm{blk}}(A_0) \leq\max A_0 + \kappa n forces a connected bad barrier of size at least c1nc_1n, separating the reached cluster from the right side of the cone. Hence, by (4),

Pμ(Sblk(A0), Rnblk(A0)≤max⁡A0+κn)≤Ce−cn.(5)\mathbb{P}_{\mu}\left(S^{\mathrm{blk}}(A_0),\ R_n^{\mathrm{blk}}(A_0) \leq\max A_0 + \kappa n\right) \leq Ce^{-cn}. \tag*{(5)}

By reflection, the same argument applies to the leftmost infinite path. On Sblk(A0)S^{\mathrm{blk}}(A_0), the event Γnblk(A0)≥min⁡A0−κn\Gamma_n^{\mathrm{blk}}(A_0) \geq\min A_0 - \kappa n forces a connected bad barrier of size at least c2nc_2n along the leftward cone. Thus

Pμ(Sblk(A0), Γnblk(A0)≥min⁡A0−κn)≤Ce−cn.(6)\mathbb{P}_{\mu}\left(S^{\mathrm{blk}}(A_0),\ \Gamma_n^{\mathrm{blk}}(A_0) \geq\min A_0 - \kappa n\right) \leq Ce^{-cn}. \tag*{(6)}

□

By additivity and translation invariance, positive toy survival from [0,H][0,H] implies positive survival from {0}\{0\}. For the regeneration construction, define

θtoy:=Pμ(the toy process started from {0} survives∣β0(0)=0)>0.\theta^{\mathrm{toy}} := \mathbb{P}_{\mu}\left(\text{the toy process started from }\{0\}\text{ survives}\mid\beta_0(0)=0\right)>0.

Indeed, conditioning on β0(0)=0\beta_0(0)=0 removes only blocking intervals initially active at the origin and cannot destroy any toy infection path.

Corollary 2.3 (One-sided toy survival). For μ\mu sufficiently large,

θ+toy:=Pμ(there exists an infinite toy infection path from (0,0)contained in [0,∞)×[0,∞) ∣ β0(0)=0)>0.\theta^{\mathrm{toy}}_+ := \mathbb{P}_{\mu}\left(\left. \begin{gathered} \text{there exists an infinite toy infection path from }(0,0)\\ \text{contained in }[0,\infty)\times[0,\infty) \end{gathered} \ \right|\ \beta_0(0)=0\right)>0.

Proof. Run the block construction on the right half of the oriented lattice, shifting the initial block so that all boxes lie in the nonnegative half-line. The cutoff estimate, Lemma 2.1, and the Peierls argument of Proposition 2.2 still apply, giving positive probability of an infinite toy infection path from a finite seed.

A finite graphical event in the nonnegative half-line produces such a seed from (0,0)(0,0). Conditional on β0(0)=0\beta_0(0)=0, this event requires only finitely many arrows, the absence of finitely many recovery marks, and the absence of blockers on finitely many space-time segments, so it has positive probability. Concatenating the two constructions proves the claim. □

Proof of Theorem 1.2. Fix a finite nonempty A⊂ZA \subset\mathbb{Z} and x∈Ax \in A. Every toy infection path from {x}\{x\} is a legal CPTI infection path from AA. Hence

Pλ,μ(τA=∞)≥Pμ(the toy process from {x} survives)>0.\mathbb{P}_{\lambda,\mu}(\tau^{A}=\infty) \geq\mathbb{P}_{\mu}(\text{the toy process from }\{x\}\text{ survives}) > 0.

Remark 2.1. Fix ε>0\varepsilon> 0 and p0<1p_0 < 1. By the classical contact-process growth theorem, the block dimensions can be chosen so that

P(Gj,ncl)>p0\mathbb{P}(G^{\mathrm{cl}}_{j,n}) > p_0

and one rightward block step has microscopic slope greater than vCP(λ)−εv_{\mathrm{CP}}(\lambda)-\varepsilon.

Moreover, the right-edge large-deviation estimate for supercritical oriented percolation [5 Theorem 11.1] gives block speed arbitrarily close to one when the percolation parameter is sufficiently large. Thus an infinite path of GclG^{\mathrm{cl}}-blocks following this speed produces a microscopic infection path with asymptotic speed at least vCP(λ)−εv_{\mathrm{CP}}(\lambda)-\varepsilon.

Remark 2.2. On toy survival, the process reaches arbitrarily large levels almost surely. Indeed, by Proposition 2.2, failed block attempts have finite height almost surely, while each new attempt has a uniformly positive probability of producing an infinite block path. A standard restart argument completes the proof.

Regeneration and renewal

The aim of this section is to construct a regenerative structure for the right edge. The key idea is to use successive record times to search for a fresh graphical environment, free of inherited immunities. Within such an environment, we select an infinite toy infection path that acts as a space-time barrier, shielding the new evolution from the information accumulated in the past.

Prepared trials and the one-sided law

3.1. Prepared trials and the one-sided law. Successive record levels of the right edge provide opportunities to explore previously unvisited regions. At such levels, we seek a local configuration free of immunity from which a toy infection path can be launched. If this path survives, it provides a candidate barrier for regeneration.

Fix an initial condition ξ0∈Xe\xi_0 \in\mathcal{X}^{e}. As ξ0\xi_0 is fixed, we will omit the dependence on the initial condition. Let Tx:=inf⁡{t≥0:rt=x}T_x := \inf\{t \geq0 : r_t = x\}. Fix s∗>0s_* > 0. On {Tx<∞}\{T_x < \infty\}, let WxW_x be the event that, during [Tx,Tx+s∗][T_x,T_x+s_*],

  • an infection arrow from xx to x+1x+1 occurs before the first recovery mark at xx after TxT_x;

  • no recovery mark occurs at x+1x+1 during [Tx,Tx+s∗][T_x,T_x+s_*];

  • no infection arrows occur from x+1x+1 to x+2x+2 during [Tx,Tx+s∗][T_x,T_x+s_*];

  • βTx+s∗(x+1)=0\beta_{T_x+s_*}(x+1)=0.

Only after this preparation do we launch the auxiliary toy process from the singleton {x+1}\{x+1\}. The toy-blocking field is not reinitialized at the launch time: the launched process uses the stationary blocking field defined in § 1.3, including any blocking intervals born before the launch time that are still active.

Throughout the regeneration construction, after rooting a launched toy process at time 00, let

τtoy:=inf⁡{t≥0:ηt{0}=∅}\tau^{\mathrm{toy}} := \inf\{t \ge0 : \eta_t^{\{0\}} = \varnothing\}

be the extinction time of the toy process launched from the singleton, with the convention τtoy=∞\tau^{\mathrm{toy}} = \infty on survival, and set

Stoy:={τtoy=∞}.S^{\mathrm{toy}} := \{\tau^{\mathrm{toy}} = \infty\}.

We next fix the reference post-preparation law used throughout the section. Starting from the stationary graphical construction of § 1.3, condition on the preparation event WxW_x, shift the prepared site and launch time to (0,0)(0,0), and retain the rooted graphical environment from which the toy process is launched, including the active toy-blocking intervals crossing time 00 away from the prepared site. We denote this biased post-preparation law by P^\widehat{\mathbb{P}}. By construction,

P^(β0(0)=0)=1.\widehat{\mathbb{P}}(\beta_0(0)=0)=1.

Set

θ^:=P^(Stoy).\widehat{\theta}:=\widehat{\mathbb{P}}(S^{\mathrm{toy}}).

Lemma 3.1 below shows that θ^>0\widehat{\theta}>0 for μ\mu sufficiently large.

Lemma 3.1 (Positive survival under the trial law). Assume that μ\mu is in the regime of Corollary 2.3. Then

θ^=P^(Stoy)≥θ+toy>0.\widehat{\theta}=\widehat{\mathbb{P}}(S^{\mathrm{toy}})\ge\theta^{\mathrm{toy}}_+>0.

Proof. Consider the post-preparation environment under P^\widehat{\mathbb{P}} and root the prepared site and launch time at (0,0)(0,0). By construction, β0(0)=0\beta_0(0)=0.

To prove survival it is enough to restrict the launched toy process to the right half-line [0,∞)[0,\infty).

Before the preparation, the graphical information in this half-line has not been inspected. Apart from the condition β0(0)=0\beta_0(0)=0, the preparation imposes only a fixed finite set of local requirements. Its infection-arrow requirements do not alter the toy-blocking field, while its recovery-mark requirements are absence requirements and are therefore favorable for toy infection paths.

Consequently, the post-preparation toy environment in the right half-line may be coupled with the stationary toy environment conditioned on β0(0)=0\beta_0(0)=0 so that every one-sided toy infection path in the latter environment is also available under P^\widehat{\mathbb{P}}. Therefore Corollary 2.3 gives

θ^=P^(Stoy)≥θ+toy>0.\widehat{\theta}=\widehat{\mathbb{P}}(S^{\mathrm{toy}})\ge\theta^{\mathrm{toy}}_+>0.

Let E\mathcal{E} denote the rooted post-preparation graphical environment. On StoyS^{\mathrm{toy}}, the auxiliary toy construction contains at least one infinite admissible infection path issued from the prepared site. Among all such paths, we choose the leftmost infinite path, with ties resolved according to a fixed measurable deterministic rule, and denote it by

Γ=Γ(E).\Gamma=\Gamma(\mathcal{E}).

This distinguished path will serve as the regeneration barrier.

We call the following procedure a *trial based at level xx. At the record time TxT_x, we attempt the preparation event WxW_x.

If WxW_x fails, the trial is rejected and is resolved at time Tx+s∗T_x+s_*. If WxW_x occurs, then at time Tx+s∗T_x+s_* the site x+1x+1 is an infected, unblocked right edge. From this site we launch the auxiliary toy process in the retained toy-blocking field.

If the launched toy process realizes StoyS^{\mathrm{toy}}, the trial is accepted. Among all infinite admissible toy infection paths issued from x+1x+1, we choose the leftmost infinite path according to the deterministic measurable rule fixed above and denote it by Γx\Gamma_x.

If the launched toy process does not realize StoyS^{\mathrm{toy}}, the trial is rejected. In this case τxtoy<∞\tau_x^{\mathrm{toy}}<\infty, where τxtoy\tau_x^{\mathrm{toy}} denotes the extinction time of the toy process launched from x+1x+1, measured from the launch time. Thus the failure is resolved when the launched toy process becomes empty. In summary, a trial has three possible outcomes:

  • preparation failure,

  • preparation success but toy extinction,

  • preparation success and toy survival.

Only the last outcome is accepted. The two last possibilities are illustrated in Figure 3.

Regeneration trials after successful preparation

Figure 3. Regeneration trials after successful preparation. In (a), the toy process becomes extinct and the trial is rejected. In (b), it survives, and its leftmost infinite infection path Γ\Gamma becomes the regeneration barrier.

The finite graphical pattern entering WxW_x involves only a fixed space-time box. The additional requirement βTx+s∗(x+1)=0\beta_{T_x+s_*}(x+1)=0 is imposed at the prepared site, whose toy-blocking history has not been inspected before the trial. By stationarity of the toy-blocking field,

Pμ(β0(0)=0)=e−1/μ>0.\mathbb{P}_{\mu}\bigl(\beta_0(0)=0\bigr)=e^{-1/\mu}>0.

Since the remaining requirements in Wx\mathcal{W}_x form a fixed finite graphical pattern compatible with this event, there exists qprep>0q_{\mathrm{prep}}>0 such that, uniformly over the information revealed before the trial,

Pμ(Wx∣Gx)≥qprepon {Tx<∞},(7)\mathbb{P}_{\mu}(\mathcal{W}_x\mid\mathcal{G}_x)\ge q_{\mathrm{prep}}\quad\text{on }\{T_x<\infty\}, \tag*{(7)}

where Gx\mathcal{G}_x denotes the information revealed before the preparation at level xx is performed.

By Lemmas 3.1 and 3.4 below, once the preparation succeeds the post-preparation environment has law P^\widehat{\mathbb{P}} and the toy survival probability is θ^>0\widehat{\theta}>0. Hence every trial is accepted with conditional probability at least

pbr:=qprepθ^>0.p_{\mathrm{br}}:=q_{\mathrm{prep}}\widehat{\theta}>0.

An accepted barrier does not regenerate the full CPTI configuration. It regenerates only the future right edge. We therefore introduce a one-sided law QQ. If E\mathcal{E} has law P^\widehat{\mathbb{P}}, condition on StoyS^{\mathrm{toy}}, select the leftmost infinite toy path Γ\Gamma, and retain the marked graphical construction on and to the right of Γ\Gamma. Writing Φ(E)\Phi(\mathcal{E}) for the resulting rooted one-sided state, let Ω+\Omega_+ denote the space of such rooted one-sided states and define

Q:=LP^(Φ(E)∣Stoy).Q:=\mathcal{L}_{\widehat{\mathbb{P}}}\bigl(\Phi(\mathcal{E})\mid S^{\mathrm{toy}}\bigr).

We write PQ\mathbb{P}_Q, EQ\mathbb{E}_Q, and LQ\mathcal{L}_Q for probability, expectation, and law when the initial rooted one-sided state has distribution QQ.

The future right edge can be read from a state with law QQ because the distinguished leftmost toy path is a CPTI infection path and will act as a barrier. We next isolate the deterministic shielding property of an accepted barrier. The geometric mechanism is illustrated in Figure 4.

Shielding mechanism diagram

Figure 4. Shielding mechanism. The infection path γ\gamma separates the retained one-sided graphical state from an admissible completion on its left. Any infection path entering the retained region from the left must first meet γ\gamma, where the paths coalesce.

For a càdlàg nearest-neighbor path γ\gamma, write Dγ+D^+_\gamma for the closed space-time region on and to the right of γ\gamma, including the horizontal edge traversed whenever γ\gamma jumps, and write Dγ−D^-_\gamma for its strict-left complement.

We regard the marked graphical construction in Dγ+D^+_\gamma, together with the distinguished path γ\gamma and the CPTI state on and to the right of γ(0)\gamma(0) at time 00, as a one-sided graphical state.

We now introduce the notion of completion of a one-sided state. Informally, an admissible completion is any way of filling in the discarded strict-left graphical history that is compatible with the retained one-sided state and does not destroy the distinguished path γ\gamma as a legal CPTI infection path.

Definition 3.2. An admissible completion of a one-sided state is any extension to a full marked graphical construction on Z×[0,∞)\mathbb{Z}\times[0,\infty), together with an initial CPTI configuration on the strict-left side, such that:

(1) the extension agrees with the retained graphical construction in Dγ+D^+_\gamma and with the retained CPTI state at time 00 on and to the right of γ(0)\gamma(0); and

(2) in the resulting CPTI construction, γ\gamma is a legal infection path starting from γ(0)\gamma(0) at time 00.

Thus an admissible completion may be arbitrary on the discarded strict-left side, subject only to compatibility with the retained one-sided state and to the requirement that the distinguished path remain a legal CPTI infection path.

Lemma 3.3 (Shielding by an infection path). Fix a one-sided graphical state consisting of an infinite path γ\gamma together with the marked graphical construction in Dγ+D_\gamma^+. Let ξ(1)\xi^{(1)} and ξ(2)\xi^{(2)} be the CPTI constructions obtained from two admissible completions of the same one-sided graphical state.

Then, for every t≥0t \ge0,

ξt(1)(x)=ξt(2)(x)for every x≥γ(t).(8)\xi_t^{(1)}(x) = \xi_t^{(2)}(x) \qquad\text{for every } x \ge\gamma(t). \tag*{(8)}

Consequently their right edges agree for all $t \ge 0.

Proof. We first work in a finite space-time box and argue successively through its graphical event times.

At time 0 the two constructions agree on and to the right of γ(0)\gamma(0). Suppose that this agreement holds immediately before the next relevant graphical event.

Any graphical event occurring strictly to the right of γ\gamma acts identically in the two constructions, since the marked graphical information in Dγ+D_\gamma^+ is common to both. An infection arrow entering the current site of γ\gamma from its strict left cannot create a discrepancy, because the site occupied by γ\gamma is already infected in both constructions.

A recovery mark cannot occur on a vertical segment traversed by γ\gamma, since γ\gamma is a legal infection path in both constructions.

We now record the only model-specific input in the shielding argument. The distinguished barrier γ\gamma is a toy infection path, and under the natural coupling every toy infection path is a legal CPTI infection path. In particular, γ\gamma cannot encounter an active CPTI immunity: if a site visited by γ\gamma were immune at the relevant time, γ\gamma would not be a legal CPTI infection path and hence could not be the barrier. Thus every infection arrow traversed by γ\gamma is legal for the CPTI.

Suppose that at time uu the path γ\gamma jumps from zz to z+1z+1. Immediately before the jump, the two constructions agree at every site x≥zx \ge z. Since the infection arrow traversed by γ\gamma belongs to the common retained graphical construction, it acts identically in both constructions. Hence the two configurations still agree at every site x≥z+1x \ge z+1, which is exactly the region on and to the right of γ\gamma after the jump.

Suppose instead that at time uu the path γ\gamma jumps from z+1z+1 to zz. Immediately before the jump, the source z+1z+1 is infected in both constructions because it lies on γ\gamma. Since both strict-left histories are admissible completions, γ\gamma is a legal infection path in each construction. In particular, the target site zz is available for the arrow traversed by γ\gamma in both constructions. Hence the arrow infects zz in both constructions, or has no effect if zz is already infected. Agreement therefore extends to the newly incorporated boundary site zz.

Thus agreement on and to the right of the current position of γ\gamma is preserved after every graphical event. Induction proves (8) in every finite space-time box. Locality of the Harris construction, followed by an increasing-box argument, gives the assertion for every finite time interval, and hence for all t≥0t \ge0.

Finally, γ(t)\gamma(t) is infected in both constructions. Since the two configurations agree at every site on and to its right, their rightmost infected sites coincide. □

Construction of one cycle from QQ

We now construct one regeneration cycle starting from the one-sided law QQ. The rooted state is shifted so that the prepared site of the already accepted barrier is at the space-time point (0,0)(0,0). Complete the graphical history on the strict-left side in any way compatible with the distinguished barrier. In these coordinates the future CPTI right edge is denoted by rtr_t, with r0=0r_0=0; Lemma 3.3 shows that its future evolution is independent of the chosen compatible completion. We write

rˉt:=sup⁡0≤s≤trs.\bar{r}_t := \sup_{0 \le s \le t} r_s.

By Remark 2.2, on StoyS^{\mathrm{toy}} the toy process reaches arbitrarily large levels almost surely. Since every toy infection path is a legal CPTI infection path under the natural coupling, it follows that rtr_t reaches arbitrarily large levels almost surely under QQ.

A rejected trial may reveal graphical information beyond the current right edge, since the associated toy process can explore far to the right before becoming extinct. To retain fresh graphical information for subsequent trials, we must therefore keep track not only of the advance of the CPTI right edge, but also of the region explored by previous trials.

The revealed frontier. During rejected trials we reveal graphical information in order to decide whether the preparation succeeded and, if it did, to follow the launched toy process until extinction. A finite toy process may explore graphical information far to the right of its starting level before it dies. Set Y0:=0Y_0 := 0 and let YiY_i denote the rightmost spatial point whose graphical information has been revealed after the first ii rejected trials have been resolved, including the fixed local information needed to decide the preparation events and all information inspected while running the rejected toy processes up to extinction.

Candidate record levels. Set ρ0:=0\rho_0 := 0. Suppose that the first i−1i-1 trials have been rejected and resolved at times ρ1,…,ρi−1\rho_1,\ldots,\rho_{i-1}. The next candidate level is chosen to be the first integer strictly to the right of both the current running record and all previously revealed trial information:

ai:=1+max⁡{rˉρi−1,Yi−1}.(9)a_i := 1+\max\{\bar r_{\rho_{i-1}},Y_{i-1}\}. \tag*{(9)}

Let

Ti:=inf⁡{t≥ρi−1:rt=ai}=Tai,T_i := \inf\{t\geq\rho_{i-1}:r_t=a_i\}=T_{a_i},

where the second equality uses the notation of Subsection 3.1. By Remark 2.2, Ti<∞T_i<\infty almost surely under QQ.

At time TiT_i, the half-line strictly to the right of aia_i has not yet been reached by the CPTI right edge. Hence it contains no active CPTI immunities created by the past. Moreover, by the definition of aia_i in (9), this half-line contains no graphical information revealed while resolving the previous rejected trials. Thus the trial at level aia_i is performed in previously uninspected graphical information.

The ii-th trial. At time TiT_i, we perform the trial based at level aia_i, in the sense of Subsection 3.1. If this trial is rejected, let ρi\rho_i be its resolution time. For the ii-th trial, write τitoy:=τaitoy\tau_i^{\mathrm{toy}}:=\tau_{a_i}^{\mathrm{toy}}. Thus ρi=Ti+s∗\rho_i=T_i+s_* in the case of preparation failure, while

ρi=Ti+s∗+τitoy\rho_i=T_i+s_*+\tau_i^{\mathrm{toy}}

in the case where the preparation succeeds but the launched toy process becomes extinct.

For a rejected trial, define the trial overshoot by

Oi:=(max⁡{rˉρi,Yi}−ai)+.O_i:=\bigl(\max\{\bar r_{\rho_i},Y_i\}-a_i\bigr)_+.

Then, on rejection of the ii-th trial,

ai+1=1+max⁡{rˉρi,Yi}≤ai+Oi+1.(10)a_{i+1}=1+\max\{\bar r_{\rho_i},Y_i\}\leq a_i+O_i+1. \tag*{(10)}

If the trial based at aia_i is accepted, the cycle stops. Define

N:=inf⁡{i≥1:the trial based at ai is accepted}.(11)N:=\inf\{i\geq1:\text{the trial based at }a_i\text{ is accepted}\}. \tag*{(11)}

The accepted time, accepted position, and accepted barrier are

τ+:=TN+s∗,X+:=aN+1,Γ+:=ΓaN.\tau_+:=T_N+s_*,\qquad X_+:=a_N+1,\qquad\Gamma_+:=\Gamma_{a_N}.

Let Gi\mathcal G_i denote the information revealed before the preparation at level aia_i is performed. By (9), no graphical information strictly to the right of aia_i has been inspected while resolving the previous rejected trials.

The preceding construction ensures that each candidate trial is performed beyond all graphical information revealed by the earlier rejected trials. The following lemma makes precise the resulting freshness property of the post-preparation environment under this adaptive exploration.

Lemma 3.4 (Invariant trial law under adaptive exploration). Let Ki\mathcal K_i be the σ\sigma-field generated by all the information inspected before the post-preparation environment Ei\mathcal E_i of the ii-th trial is examined, including the information used to resolve the preceding rejected trials and to perform the ii-th preparation. Thus Gi\mathcal G_i contains the information before the ii-th preparation is performed, whereas Ki\mathcal K_i also contains the information used to decide that preparation.

On the event Wai\mathcal W_{a_i} that the ii-th preparation succeeds,

LQ(Ei∣Ki)=P^Pμ-a.s.(12)\mathcal L_Q(\mathcal E_i\mid\mathcal K_i)=\widehat{\mathbb P}\quad\mathbb P_\mu\text{-a.s.} \tag*{(12)}

In particular, the conditional law does not depend on ii or on the realized information in Ki\mathcal{K}_i. Moreover, for every bounded measurable function GG on the one-sided state space, on Wai\mathcal{W}_{a_i},

EQ[1SitoyG(Φ(Ei))∣Ki]=θ^Q(G)Pμ-a.s.,(13)\mathbb{E}_{Q}\left[\mathbf{1}_{\mathcal{S}^{\mathrm{toy}}_i}G\left(\Phi(\mathcal{E}_i)\right)\mid\mathcal{K}_i\right]=\widehat{\theta}Q(G)\quad\mathbb{P}_{\mu}\text{-a.s.}, \tag*{(13)}

where

Sitoy:={Ei∈Stoy}.\mathcal{S}^{\mathrm{toy}}_i:=\{\mathcal{E}_i\in\mathcal{S}^{\mathrm{toy}}\}.

Consequently, on Wai\mathcal{W}_{a_i},

PQ(Sitoy∣Ki)=θ^Pμ-a.s.,\mathbb{P}_{Q}\left(\mathcal{S}^{\mathrm{toy}}_i\mid\mathcal{K}_i\right)=\widehat{\theta}\quad\mathbb{P}_{\mu}\text{-a.s.},

and, conditional on Ki\mathcal{K}_i, Wai\mathcal{W}_{a_i}, and Sitoy\mathcal{S}^{\mathrm{toy}}_i, the accepted one-sided state Φ(Ei)\Phi(\mathcal{E}_i) has law QQ.

Proof. The point is that the adaptive exploration does not inspect the graphical information from which the next post-preparation environment is read.

First consider a one-sided state with law QQ. By definition, it is obtained by conditioning a post-preparation environment with law P^\widehat{\mathbb{P}} on Stoy\mathcal{S}^{\mathrm{toy}}, selecting the leftmost infinite toy infection path, and retaining the marked graphical construction on and to the right of that path. A left-first priority exploration identifies the leftmost infinite toy path by following leftmost continuations and backtracking whenever an alternative dies out. On Stoy\mathcal{S}^{\mathrm{toy}}, this exploration reveals only the selected path and failed alternatives on its left, leaving the strict-right graphical field uninspected. By planarity, a path lying strictly to the left of the selected infinite path cannot cross it without first meeting it. Thus neither toy survival nor the choice of the leftmost surviving path requires inspection of the strict-right marked graphical field. Since the underlying marked Poisson construction is a product field, this uninspected strict-right information retains its original conditional law. This is the unrevealed-field property built into the one-sided law QQ.

Now run the cycle exploration from QQ. Before the ii-th trial, all information revealed while resolving the preceding rejected trials lies at or to the left of the revealed frontier Yi−1Y_{i-1}. By (9), the candidate level aia_i is chosen strictly to the right of both Yi−1Y_{i-1} and the running record r‾ρi−1\overline{r}_{\rho_{i-1}}. Thus, up to the first hitting time TiT_i, no information strictly to the right of aia_i has been inspected by the cycle exploration, and this half-line has not previously been reached by the CPTI right edge. The preparation event then imposes exactly the conditioning used in the definition of P^\widehat{\mathbb{P}}: the fixed local graphical pattern, together with the requirement that the prepared launch site carry no active toy blocker.

The toy-blocking field is part of the retained marked construction, including blocking intervals born before the launch time. Nevertheless, on the uninspected side its pre-launch marks have not been revealed. By the product structure and space-time translation invariance of the stationary marked construction, after translating the prepared site and launch time to (0,0)(0,0), the conditional law of the resulting post-preparation environment is exactly the fixed biased law P^\widehat{\mathbb{P}}. This proves (12).

It remains only to identify the accepted kernel. By the definition of QQ, if E\mathcal{E} has law P^\widehat{\mathbb{P}}, then conditional on Stoy\mathcal{S}^{\mathrm{toy}} the one-sided state Φ(E)\Phi(\mathcal{E}) has law QQ. Therefore, for every bounded measurable GG,

P^[1StoyG(Φ(E))]=θ^Q(G).\widehat{\mathbb{P}}\left[\mathbf{1}_{\mathcal{S}^{\mathrm{toy}}}G\left(\Phi(\mathcal{E})\right)\right]=\widehat{\theta}Q(G).

Combining this identity with (12) gives (13) on Wai\mathcal{W}_{a_i}. Taking G≡1G\equiv1 gives the stated conditional survival probability, and division by θ^\widehat{\theta} gives the conditional law of Φ(Ei)\Phi(\mathcal{E}_i) on Wai∩Sitoy\mathcal{W}_{a_i}\cap\mathcal{S}^{\mathrm{toy}}_i. □\square

Conditional on a successful preparation, Lemma 3.4 implies that the post-preparation environment has law P^\widehat{\mathbb{P}}. Combining this with the uniform preparation bound (7) gives

PQ(the trial based at ai is accepted∣Gi)≥pbr>0.\mathbb{P}_{Q}\left(\text{the trial based at }a_i\text{ is accepted}\mid\mathcal{G}_i\right)\ge p_{\mathrm{br}}>0.

Consequently,

PQ(N>n)≤(1−pbr)n,n≥0.\mathbb{P}_{Q}(N>n)\leq(1-p_{\mathrm{br}})^n,\qquad n\geq0.

Finally, the post-regeneration one-sided state is obtained by shifting (X+,τ+)(X_+,\tau_+) to (0,0)(0,0), retaining the distinguished leftmost barrier Γ+\Gamma_+, and retaining the marked graphical construction on and to the right of Γ+\Gamma_+.

The barrier renewal property

The key to renewal is that an accepted toy barrier protects the fresh graphical environment from the information revealed during the preceding exploration. Consequently, the evolution retained beyond the barrier has the same law at each accepted trial, independently of the completed history.

We can now formulate the renewal statement. Its two ingredients are the invariant trial law established in the preceding subsection and the deterministic shielding property of an accepted toy path established in Lemma 3.3.

Since acceptance is determined by the infinite future, the renewal statement is formulated with respect to the exploration σ\sigma-field H+\mathcal{H}_+ defined below.

For each ii, let Ki\mathcal{K}_i be as in Lemma 3.4. On a successful preparation, write

Sitoy:={Ei∈Stoy}.S_i^{\mathrm{toy}}:=\{\mathcal{E}_i\in S^{\mathrm{toy}}\}.

Recall the accepted-trial index NN from (11).

Let H+\mathcal{H}_+ be the renewal σ\sigma-field characterized as follows: on the event {N=i}\{N=i\} it contains all information in Ki\mathcal{K}_i, the fact that the ii-th preparation succeeded, and the single bit of information Sitoy=1S_i^{\mathrm{toy}}=1, but it contains no further information from Ei\mathcal{E}_i. Equivalently, B∈H+B\in\mathcal{H}_+ if, for each ii, there exists Bi∈KiB_i\in\mathcal{K}_i such that

B∩{N=i}=Bi∩{N=i}.B\cap\{N=i\}=B_i\cap\{N=i\}.

Proposition 3.5 (Barrier renewal property). The post-regeneration one-sided state Φ(EN)\Phi(\mathcal{E}_N) is independent of H+\mathcal{H}_+ and has law QQ. More precisely, for every bounded measurable G:Ω+→RG:\Omega_+\to\mathbb{R},

EQ[G(Φ(EN))∣H+]=Q(G)Pμ-a.s.(14)\mathbb{E}_{Q}\left[G\left(\Phi(\mathcal{E}_N)\right)\mid\mathcal{H}_+\right]=Q(G)\qquad\mathbb{P}_{\mu}\text{-a.s.} \tag*{(14)}

Moreover, the future CPTI right-edge process is determined by Φ(EN)\Phi(\mathcal{E}_N) alone: any two admissible completions of the discarded graphical history to the strict left of the accepted barrier produce the same future right-edge process.

Proof. The distributional assertion is a consequence of the invariant trial-law lemma and does not use the shielding argument.

Fix a bounded measurable GG. Let B∈H+B\in\mathcal{H}_+. On {N=i}\{N=i\} write

B∩{N=i}=Bi∩{N=i},B\cap\{N=i\}=B_i\cap\{N=i\},

where BiB_i is measurable with respect to the information available before the accepted post-preparation environment is exposed.

By (13), on the event that the ii-th preparation succeeds,

EQ[1SitoyG(Φ(Ei))∣Ki]=θ^Q(G),(15)\mathbb{E}_{Q}\left[\mathbf{1}_{S_i^{\mathrm{toy}}}G\left(\Phi(\mathcal{E}_i)\right)\mid\mathcal{K}_i\right]=\widehat{\theta}Q(G), \tag*{(15)}

while, taking G≡1G \equiv1,

PQ(Sitoy∣Ki)=θ^.\mathbb{P}_{Q}\left(S_i^{\mathrm{toy}} \mid\mathcal{K}_i\right)=\hat{\theta}.

Therefore

EQ[1Bi1{N>i−1}1{i-th preparation succeeds}1SitoyG(Φ(Ei))]=Q(G)EQ[1Bi1{N>i−1}1{i-th preparation succeeds}1Sitoy].(16)\mathbb{E}_{Q}\left[\mathbf{1}_{B_i}\mathbf{1}_{\{N>i-1\}}\mathbf{1}_{\{\text{i-th preparation succeeds}\}}\mathbf{1}_{S_i^{\mathrm{toy}}}G\left(\Phi(\mathcal{E}_i)\right)\right]=Q(G)\mathbb{E}_{Q}\left[\mathbf{1}_{B_i}\mathbf{1}_{\{N>i-1\}}\mathbf{1}_{\{\text{i-th preparation succeeds}\}}\mathbf{1}_{S_i^{\mathrm{toy}}}\right]. \tag*{(16)}

The event multiplying GG on the left-hand side is precisely B∩{N=i}B \cap\{N=i\}. Hence

EQ[1B∩{N=i}G(Φ(Ei))]=Q(G)PQ(B∩{N=i}).\mathbb{E}_{Q}\left[\mathbf{1}_{B\cap\{N=i\}}G\left(\Phi(\mathcal{E}_i)\right)\right]=Q(G)\mathbb{P}_{Q}(B\cap\{N=i\}).

Summing over i≥1i \ge1 and using N<∞N<\infty almost surely gives

EQ[1BG(Φ(EN))]=Q(G)PQ(B).\mathbb{E}_{Q}\left[\mathbf{1}_{B}G\left(\Phi(\mathcal{E}_N)\right)\right]=Q(G)\mathbb{P}_{Q}(B).

Since this holds for every B∈H+B \in\mathcal{H}_{+}, (14) follows. Thus Φ(EN)\Phi(\mathcal{E}_N) has law QQ and is independent of H+\mathcal{H}_{+}.

It remains to verify that the retained one-sided state determines the future right edge. Let two graphical histories in the strict-left region be admissible completions of the same realized state Φ(EN)\Phi(\mathcal{E}_N). By definition, the distinguished path contained in Φ(EN)\Phi(\mathcal{E}_N) is a legal CPTI infection path after either completion is glued to the retained graphical construction. The two resulting CPTI constructions have identical marked graphical information on and to the right of this path. Therefore Lemma 3.3 implies that their configurations agree on and to the right of the barrier at every future time. In particular, their right edges coincide. Hence the future right-edge process does not depend on which admissible completion of the discarded strict-left history is used, and is therefore determined by Φ(EN)\Phi(\mathcal{E}_N) alone. □\square

Regeneration increments

We now define the regeneration sequence. Let (X1,τ1)(X_1,\tau_1) be the first accepted barrier constructed from the original half-line initial condition. In other words, τ1\tau_1 is the time at which the first accepted preparation is completed, and X1X_1 is the prepared site. After this time, the distinguished leftmost toy barrier launched from this site is used to define the post-regeneration one-sided state.

Inductively, after (Xn,τn)(X_n,\tau_n) has been constructed, we shift this accepted barrier to (0,0)(0,0) and run the same record-level search from the one-sided state to the right of its distinguished leftmost toy barrier. If the next accepted barrier, in shifted coordinates, has prepared site X+X_{+} and completion time τ+\tau_{+}, we set

Xn+1:=Xn+X+,τn+1:=τn+τ+.X_{n+1}:=X_n+X_{+}, \qquad\tau_{n+1}:=\tau_n+\tau_{+}.

Thus (Xn,τn)(X_n,\tau_n) is the space-time location of the nn-th accepted barrier in the original coordinates. Set X0:=0X_0:=0 and τ0:=0\tau_0:=0. For n≥1n\ge1, define the regeneration increments by

ΔXn:=Xn−Xn−1,Δτn:=τn−τn−1.\Delta X_n:=X_n-X_{n-1}, \qquad\Delta\tau_n:=\tau_n-\tau_{n-1}.

We also define the maximal oscillation during the nn-th cycle by

Mn:=sup⁡τn−1≤t≤τn∣rt−Xn−1∣.M_n:=\sup_{\tau_{n-1}\le t\le\tau_n}\left|r_t-X_{n-1}\right|.

As a consequence of the barrier renewal property, the regeneration increments and maximal oscillations are i.i.d. after the first cycle; see Figure 5. This is made precise in the next proposition.

Regeneration cycles of the right edge

Figure 5. Regeneration cycles of the right edge. The second cycle illustrates the displacement ΔX2\Delta X_2 and duration Δτ2\Delta\tau_2, while the third illustrates the maximal oscillation M3M_3.

Proposition 3.6. After the first accepted barrier, the sequence

(ΔXn,Δτn,Mn)n≥2(17)(\Delta X_n,\Delta\tau_n,M_n)_{n\geq2} \tag*{(17)}

is i.i.d. The first cycle may have a different law.

Proof. Once the first accepted barrier has been found, we translate its prepared site and time to (0,0)(0,0). We then keep the marked graphical construction on and to the right of the distinguished leftmost toy barrier launched at the accepted preparation. The same invariant trial-law argument as in Lemma 3.4, applied to the search from the original half-line initial condition, shows that at the first accepted trial the shifted one-sided state has law QQ, independently of the information revealed before that accepted post-preparation environment is exposed.

Assume that after the nn-th accepted barrier, with n≥1n\geq1, the shifted one-sided state has law QQ and is independent of the completed past cycles. Run the same record-level search from this state. The next cycle variables (ΔXn+1,Δτn+1,Mn+1)(\Delta X_{n+1},\Delta\tau_{n+1},M_{n+1}) are measurable functions of the one-sided state and of the graphical information revealed during the cycle. Since the input law is QQ, their conditional law, given the completed past, is the same fixed law for every n≥1n\geq1. At the accepted barrier, Proposition 3.5 gives a new independent post-regeneration one-sided state with law QQ. Iterating proves that the sequence in (17) is i.i.d. □

Regenerative limit theory for the right edge

The renewal structure reduces the large-scale motion of the right edge to a sequence of independent regeneration cycles. To transfer limit theorems from this regenerative skeleton to the full process, we must also control the fluctuations of the right edge between successive regeneration times.

The regenerative limit theorems rely on the following exponential moment bounds, whose proof is deferred to Appendix § 7.

Proposition 4.1 (Exponential moment bounds). Fix the initial configuration in which every site x≤0x \le0 is infected and every site x>0x > 0 is susceptible. There exists δ>0\delta> 0 such that

Eμ[eδτ1+eδ∣X1∣+eδM1]<∞.(18)\mathbb{E}_{\mu}\left[e^{\delta\tau_1}+e^{\delta|X_1|}+e^{\delta M_1}\right] < \infty. \tag*{(18)}

Moreover, for a generic regeneration cycle (X+,τ+,M+)(X_+,\tau_+,M_+) started from QQ,

EQ[eδτ++eδ∣X+∣+eδM+]<∞.(19)\mathbb{E}_{Q}\left[e^{\delta\tau_+}+e^{\delta|X_+|}+e^{\delta M_+}\right] < \infty. \tag*{(19)}

To extend the regenerative limit theorems to arbitrary ξ∈Xe\xi\in\mathcal{X}^{\mathrm e}, we first show that initial immunities sufficiently far to the right expire before the advancing infection can reach them.

Lemma 4.2 (Entrance past the initial immunities). Fix ξ∈Xe\xi\in\mathcal{X}^{\mathrm e}. For y>0y > 0, let IyI_y be the residual lifetime at time 00 of the immunity at yy if ξ(y)=−1\xi(y)=-1, and set Iy=0I_y=0 otherwise. Let

T~y:=inf⁡{t≥0:there is a graphical arrow path from (−∞,0]×{0} to (y,t)}.\widetilde{T}_y := \inf\{t \ge0 : \text{there is a graphical arrow path from }(-\infty,0] \times\{0\}\text{ to }(y,t)\}.

Then, almost surely, there exists a finite random RR such that

Iy<T~yfor every y≥R.(20)I_y < \widetilde{T}_y \qquad\text{for every } y \ge R. \tag*{(20)}

In particular, beyond the random level RR, no immunity present at time 00 can affect the process before the corresponding site is first reached by the infection.

Proof. Let T~y\widetilde{T}_y be the first arrival time at yy when recovery marks are ignored. Then T~y≤T^y\widetilde{T}_y \le\widehat{T}_y and T~y=dE1+⋯+Ey\widetilde{T}_y \mathrel{\mathop=^{\mathrm d}} E_1+\cdots+E_y, where the EiE_i are i.i.d. Exp⁡(λ)\operatorname{Exp}(\lambda). Hence, for some a,c,C>0a,c,C>0,

Pμ(Iy≥T^y)≤Ce−cy+e−μay.\mathbb{P}_{\mu}(I_y \ge\widehat{T}_y) \le Ce^{-cy}+e^{-\mu ay}.

The bound is summable in yy, so Borel–Cantelli gives (20). □

We now use the disappearance of distant initial immunities to extend the regenerative construction to arbitrary ξ∈Xe\xi\in\mathcal{X}^{\mathrm e}.

Lemma 4.3 (Entrance into the regenerative regime). For every ξ∈Xe\xi\in\mathcal{X}^{\mathrm e}, the process reaches an accepted barrier in finite time almost surely. After translating this barrier to (0,0)(0,0), the retained one-sided state has law QQ, independently of the information revealed before its post-preparation environment is exposed.

Proof. Fix ξ∈Xe\xi\in\mathcal{X}^{\mathrm e} and let Aξ:={x:ξ(x)=1}A_\xi:= \{x : \xi(x)=1\}, which is infinite. By attractiveness of the toy process, positive survival from a singleton, and a spatial restart argument using well-separated sites of AξA_\xi, the toy process started from AξA_\xi survives almost surely. By Remark 2.2, its right edge reaches arbitrarily large levels. Since every toy infection path is a legal CPTI infection path, the same holds for the CPTI started from ξ\xi.

Run the adaptive trial construction of Subsection 3.2, choosing each successive candidate beyond both the running record and all previously revealed graphical information, without reference to the random level RR of Lemma 4.2. Every rejected trial is resolved in finite time and reveals information only up to a finite spatial level. Since the right edge is unbounded, the candidate levels tend to infinity. Almost surely, they therefore eventually exceed RR. By Lemma 4.2, inherited immunities cannot interfere with the graphical construction at these levels. The freshness argument of Lemma 3.4 then gives the uniform conditional acceptance probability pbr=qprepθ^>0p_{\mathrm{br}}=q_{\mathrm{prep}}\hat{\theta}>0 for the unaffected trials. Consequently, an accepted barrier is reached in finite time almost surely.

Finally, Proposition 3.5 shows that selection of the first accepted trial does not bias its post-preparation state. After translation, this state has law QQ and is independent of the information revealed before its post-preparation environment is exposed.

We now derive the limit theorems from the regeneration structure.

Proof of Theorem 1.3. By Propositions 3.6 and 4.1, the post-first cycles are i.i.d. and integrable. The regenerative law of large numbers [2 Theorem VI.3.1] therefore gives

rtt⟶v:=EQ[X+]EQ[τ+]PQ-a.s.\frac{r_t}{t}\longrightarrow v:=\frac{\mathbb{E}_{Q}[X_+]}{\mathbb{E}_{Q}[\tau_+]} \qquad\mathbb{P}_{Q}\text{-a.s.}

Since each cycle reaches a new record level, X+>0X_+>0 almost surely, v>0v>0.

By Lemma 4.3, every process started from ξ∈Xe\xi\in\mathcal{X}^{e} enters the regenerative regime at an almost surely finite time. Thereafter, the regeneration cycles have the same law as under QQ. Since the initial time and spatial displacement are finite almost surely, they do not affect the limiting speed. Thus the law of large numbers holds for every such ξ\xi.

For the half-line initial condition, (18) and (19) give uniform integrability of rt/tr_t/t. Thus the convergence also holds in L1L^1.

Proof of Proposition 1.4. Couple the CPTI and classical contact process, both started from the half-line, using the same infection arrows and recovery marks. Since immunity only suppresses infections, rtμ≤rtclr_t^{\mu}\le r_t^{\mathrm{cl}}, with rtclr_t^{\mathrm{cl}} being the right edge of the classical contact process. By (2) and the classical right-edge law of large numbers,

lim sup⁡μ→∞v(λ,μ)≤vCP(λ).(21)\limsup_{\mu\to\infty}v(\lambda,\mu)\le v_{\mathrm{CP}}(\lambda). \tag*{(21)}

For the reverse inequality, fix ε>0\varepsilon>0 and choose the classical block event Gj,nclG_{j,n}^{\mathrm{cl}} as in Remark 2.1. Choose p⋆<1p_\star<1 sufficiently close to one that the corresponding oriented percolation speed, in microscopic coordinates, is at least vCP(λ)−εv_{\mathrm{CP}}(\lambda)-\varepsilon.

Choose ρ⋆>0\rho_\star>0 so that Proposition 2.2 applies whenever p≥p⋆p\ge p_\star and ρE≤ρ⋆\rho_E\le\rho_\star. By the cutoff approximation and the old-tail estimate of Subsections 2.1 and 2.2, we may choose the cutoff and then μ\mu sufficiently large that both conditions hold.

Proposition 2.2 and Remark 2.1 then give, with positive probability,

lim inf⁡t→∞rtμt≥vCP(λ)−ε.(22)\liminf_{t\to\infty}\frac{r_t^{\mu}}{t}\ge v_{\mathrm{CP}}(\lambda)-\varepsilon. \tag*{(22)}

Using Theorem 1.3 and (22),

lim inf⁡μ→∞v(λ,μ)≥vCP(λ).\liminf_{\mu\to\infty}v(\lambda,\mu)\ge v_{\mathrm{CP}}(\lambda).

Together with (21), this proves the proposition.

Proof of Theorem 1.5. By Proposition 3.6, the post-first cycles are i.i.d. Moreover, (19) gives

u2PQ(M++vτ+>u)⟶0as u⟶∞.u^{2}\mathbb{P}_{Q}(M_{+}+v\tau_{+}>u)\longrightarrow0 \qquad\text{as } u\longrightarrow\infty.

Thus the regenerative functional central limit theorem [9 Theorem 1.(b)] applies. The limiting process is Brownian motion with variance parameter

σ2=EQ[(X+−vτ+)2]EQ[τ+].\sigma^{2}=\frac{\mathbb{E}_{Q}\left[(X_{+}-v\tau_{+})^{2}\right]}{\mathbb{E}_{Q}[\tau_{+}]}.

By Lemma 4.3, every process started from ξ∈Xe\xi\in\mathcal{X}^{e} enters the regenerative regime at an almost surely finite time. After this entrance, the regeneration cycles have the same law as under QQ. The finite initial time and spatial displacement vanish under diffusive scaling, so the functional limit holds for every such ξ\xi.

Finally, σ2>0\sigma^{2}>0. Indeed, on a positive-probability finite graphical event, X+X_{+} is fixed while a Poisson waiting time contributing to τ+\tau_{+} varies over a nondegenerate interval without changing the regeneration construction. Since v>0v>0, X+−vτ+X_{+}-v\tau_{+} is not almost surely constant, proving the claim. □

Shielding and asymptotic laws

We now use shielding to couple CPTI evolutions with different initial configurations and graphical histories. The key idea is that common toy barriers force agreement in expanding space-time regions, despite the lack of attractiveness. Combined with the regenerative structure, this allows us to study the process seen from the right edge and establish complete convergence.

Regenerative coupling and shielding

From now on, we will consider not only configurations with infinitely many infected individuals to the left of the origin, but also configurations with infinitely many infected individuals on both sides of the origin. Once the two processes have entered the regenerative regime, their regeneration structures can be successfully coupled. Shielding then allows the common graphical environment to determine their evolution in expanding regions, despite differences in their previous histories. For convergence seen from the right edge, the regeneration sites are first translated to the origin. For bulk convergence, the coupling is performed in the original spatial coordinates, using barriers expanding in both directions.

Lemma 5.1 (One- and two-sided regenerative coupling). For i=1,2i=1,2, let (ξt(i))t≥0(\xi^{(i)}_{t})_{t\geq0} be a CPTI whose first accepted regeneration time is almost surely finite, and let (Xn(i),τn(i))n≥1(X^{(i)}_{n},\tau^{(i)}_{n})_{n\geq1} be its regeneration sequence.

(i) The two processes may be coupled so that there exist an almost surely finite common regeneration epoch τcpl\tau_{\mathrm{cpl}} and corresponding regeneration sites Xcpl(1)X^{(1)}_{\mathrm{cpl}} and Xcpl(2)X^{(2)}_{\mathrm{cpl}}. After translating these sites to the origin, the two post-regeneration one-sided graphical states are identified and the same future graphical construction is used in the retained region. For s>0s>0, set

ξ~s(i)(x):=ξτcpl+s(i)(Xcpl(i)+x),i=1,2.\widetilde{\xi}^{(i)}_{s}(x):=\xi^{(i)}_{\tau_{\mathrm{cpl}}+s}\left(X^{(i)}_{\mathrm{cpl}}+x\right),\qquad i=1,2.

Let Γ−\Gamma^{-} be the distinguished leftmost infinite toy infection path of the common post-regeneration one-sided graphical state. Then

ξs(1)(x)=ξs(2)(x),x≥Γ−(s),s≥0.(23)\xi^{(1)}_{s}(x) = \xi^{(2)}_{s}(x), \qquad x \ge\Gamma^{-}(s), \quad s \ge0. \tag*{(23)}

Moreover, for some c−>0c_{-} > 0,

lim sup⁡s→∞Γ−(s)s≤−c−almost surely.\limsup_{s\to\infty} \frac{\Gamma^{-}(s)}{s} \le-c_{-} \qquad\textit{almost surely}.

Setting D:=Xcpl(2)−Xcpl(1)D := X^{(2)}_{\mathrm{cpl}} - X^{(1)}_{\mathrm{cpl}}, we have

ξτcpl+s(1)(x)=ξτcpl+s(2)(x+D),x≥Xcpl(1)+Γ−(s).(24)\xi^{(1)}_{\tau_{\mathrm{cpl}}+s}(x) = \xi^{(2)}_{\tau_{\mathrm{cpl}}+s}(x+D), \qquad x \ge X^{(1)}_{\mathrm{cpl}} + \Gamma^{-}(s). \tag*{(24)}

Furthermore, for every ℓ<∞\ell< \infty, there exists an almost surely finite TℓT_{\ell} such that

Ξτcpl+s(1)(x)=Ξτcpl+s(2)(x),−ℓ≤x≤0,\Xi^{(1)}_{\tau_{\mathrm{cpl}}+s}(x) = \Xi^{(2)}_{\tau_{\mathrm{cpl}}+s}(x), \qquad-\ell\le x \le0,

for all s≥Tℓs \ge T_{\ell}.

(ii) The two processes may instead be coupled so that there exist an almost surely finite coupling epoch τcpl\tau_{\mathrm{cpl}} and a site XcplX_{\mathrm{cpl}} such that, at (Xcpl,τcpl)(X_{\mathrm{cpl}},\tau_{\mathrm{cpl}}), their post-regeneration one-sided graphical states are identified in the same spatial coordinates and the same future graphical construction is used.

Let Γ−\Gamma^{-} and Γ+\Gamma^{+} be respectively the leftmost and rightmost infinite toy infection paths issued from the prepared site. For some c−,c+>0c_{-},c_{+}>0,

lim sup⁡s→∞Γ−(s)s≤−c−,lim inf⁡s→∞Γ+(s)s≥c+almost surely.\limsup_{s\to\infty} \frac{\Gamma^{-}(s)}{s} \le-c_{-}, \qquad\liminf_{s\to\infty} \frac{\Gamma^{+}(s)}{s} \ge c_{+} \qquad\textit{almost surely}.

Consequently,

lim inf⁡s→∞Γ+(s)−Γ−(s)s≥c−+c+>0almost surely.\liminf_{s\to\infty} \frac{\Gamma^{+}(s)-\Gamma^{-}(s)}{s} \ge c_{-}+c_{+}>0 \qquad\textit{almost surely}.

The strict-left side of Γ−\Gamma^{-} may be replaced by any admissible completion, and the strict-right side of Γ+\Gamma^{+} may be replaced by any admissible completion in the spatially reflected sense, separately or simultaneously. Under any such replacements, the resulting CPTI evolution agrees with the original one whenever

Xcpl+Γ−(s)≤x≤Xcpl+Γ+(s).X_{\mathrm{cpl}}+\Gamma^{-}(s) \le x \le X_{\mathrm{cpl}}+\Gamma^{+}(s).

In particular, for every ℓ<∞\ell< \infty, there exists an almost surely finite TℓT_{\ell} such that

ξτcpl+s(1)(x)=ξτcpl+s(2)(x),−ℓ≤x≤ℓ,\xi^{(1)}_{\tau_{\mathrm{cpl}}+s}(x) = \xi^{(2)}_{\tau_{\mathrm{cpl}}+s}(x), \qquad-\ell\le x \le\ell,

for all s≥Tℓs \ge T_{\ell}.

Proof. Figure 6 illustrates the two coupling mechanisms used below. We first prove (i). After their first accepted regeneration times, the two processes have the same regenerative law, with finite mean cycle length by Proposition 4.1. Moreover, the inter-regeneration time has a spread-out distribution: on a positive-probability graphical event, one of the Poisson waiting times can vary over a nondegenerate interval without changing the regeneration construction. Hence, by [21 Theorem 1.1], the regeneration sequences can be coupled, after factoring out spatial translations, to have an almost surely finite common regeneration epoch τcpl\tau_{\mathrm{cpl}}. Let Xcpl(1)X_{\mathrm{cpl}}^{(1)} and Xcpl(2)X_{\mathrm{cpl}}^{(2)} be the corresponding regeneration sites. By Proposition 3.5, after translating these sites to the origin, the two post-regeneration one-sided graphical states have law QQ, independently of their respective pasts. We identify these states and use the same future graphical marks in the retained region.

Regenerative couplings of Lemma 5.1

Figure 6. Regenerative couplings of Lemma 5.1. (i) One-sided coupling after translating the regeneration sites to the origin, with agreement to the right of Γ−\Gamma^-. (ii) Two-sided coupling from a common regeneration site, with agreement between Γ−\Gamma^- and Γ+\Gamma^+.

Let Γ−\Gamma^- be the distinguished leftmost infinite toy infection path of this common post-regeneration state. By the shielding property of Lemma 3.3,

ξ~s(1)(x)=ξ~s(2)(x),x≥Γ−(s), s≥0.\widetilde{\xi}^{(1)}_s(x)=\widetilde{\xi}^{(2)}_s(x), \qquad x\geq\Gamma^-(s),\ s\geq0.

This proves (23), and returning to the original coordinates gives (24).

By Proposition 2.2 and Remark 2.2, Γ−\Gamma^- has strictly negative asymptotic speed. Since the two shifted configurations agree to the right of Γ−\Gamma^-, their right edges coincide; denote the common edge by rscplr_s^{\mathrm{cpl}}. By Theorem 1.3, this edge has asymptotic speed v>0v>0. Consequently,

rscpl−Γ−(s)⟶∞almost surely.r_s^{\mathrm{cpl}}-\Gamma^-(s)\longrightarrow\infty\qquad\text{almost surely}.

Thus, every fixed window behind the common right edge eventually lies in the region of agreement, proving (i).

We now prove (ii). The successive regeneration points form a random walk on Z×R\mathbb{Z}\times\mathbb{R} with i.i.d. increments (ΔX,Δτ)(\Delta X,\Delta\tau). The graphical construction shows that every positive spatial increment is possible and that, for consecutive spatial increments, the corresponding regeneration-time distributions have overlapping absolutely continuous components. These overlaps, together with continuous variation of the regeneration times, give the joint aperiodicity and spread-outness needed for the random-walk coupling argument of [17 Proposition 1]. That argument, applied on Z×R\mathbb{Z}\times\mathbb{R}, couples the regeneration sequences at an almost surely finite common space-time point (Xcpl,τcpl)(X_{\mathrm{cpl}},\tau_{\mathrm{cpl}}).

Since the rightmost infinite toy path lies to the right of the distinguished leftmost path, both extremal paths are determined by the common retained graphical state. Thus the same pair of shielding paths is available to both coupled evolutions.

At the common regeneration point, the post-regeneration graphical states can be identified, since both have law QQ independently of their respective completed pasts. The subsequent marked regenerative evolutions can therefore be taken identical.

Let Γ−\Gamma^{-} and Γ+\Gamma^{+} be the leftmost and rightmost infinite toy infection paths issued from the prepared site. The leftward speed estimate above and its spatially reflected counterpart give the asserted linear expansion of the interval between these paths.

By Lemma 3.3, replacing the strict-left graphical history of Γ−\Gamma^{-} by an admissible completion leaves the CPTI evolution unchanged on and to its right. The reflected lemma gives the analogous property to the left of Γ+\Gamma^{+}. These replacements may be performed successively, since each preserves the evolution between the two paths. Thus, identifying the exterior histories of the two coupled processes as admissible completions yields agreement whenever

Xcpl+Γ−(s)≤x≤Xcpl+Γ+(s).X_{\mathrm{cpl}}+\Gamma^{-}(s)\le x\le X_{\mathrm{cpl}}+\Gamma^{+}(s).

Since the barriers expand linearly in opposite directions, this interval eventually contains every fixed finite interval, proving (ii). □

The invariant edge law

The renewal structure allows us to construct a stationary sequence of marked regeneration cycles. By extending this construction backward in time, shielding ensures that the configuration in each fixed window behind the right edge eventually becomes independent of how far in the past the process was started. This yields a stationary edge configuration, while the same coupling mechanism gives convergence from arbitrary initial configurations.

Proof of Theorem 1.6. By Proposition 4.1, the regeneration times have finite mean increments. Together with Proposition 3.5, this allows us to construct a bi-infinite stationary version of the marked regenerative process, whose successive cycles have the prescribed regenerative law. By abuse of notation, denote its law by PQ\mathbf{P}_{Q}.

For n≥1n\ge1, let (X−n,τ−n)(X_{-n},\tau_{-n}) denote the nn-th regeneration point before time 00, and let Γ(−n)\Gamma^{(-n)} be the distinguished barrier issued from this point. Using the common marked graphical construction, concatenate the successive regenerative pieces so that each barrier is preserved by all constructions started at earlier regeneration points. Let Ξ(−n)\Xi^{(-n)} be the process seen from its right edge obtained by starting from the regenerative state at (X−n,τ−n)(X_{-n},\tau_{-n}) and evolving forward to time 00.

By Lemma 5.1 (i), for every ℓ≥1\ell\ge1,

PQ(r0−Γ(−n)(0)≤ℓ)⟶0.\mathbf{P}_{Q}\left(r_{0}-\Gamma^{(-n)}(0)\le\ell\right)\longrightarrow0.

We may therefore choose a subsequence (nk)(n_k) such that the preceding probability, with n=nkn=n_k and ℓ=k\ell=k, is at most 2−k2^{-k}. By Borel–Cantelli,

r0−Γ(−nk)(0)>kr_{0}-\Gamma^{(-n_k)}(0)>k

for all sufficiently large kk, almost surely.

Fix ℓ≥1\ell\ge1. For all sufficiently large kk, the window {r0−ℓ,…,r0}\{r_{0}-\ell,\ldots,r_{0}\} lies on or to the right of Γ(−nk)(0)\Gamma^{(-n_k)}(0). By the shielding property of Lemma 3.3, the restrictions

Ξ0(−nk)∣{−ℓ,…,0}\left.\Xi^{(-n_k)}_{0}\right|_{\{-\ell,\ldots,0\}}

eventually agree. Since these restrictions are compatible as ℓ\ell varies, they define a random edge configuration Ξ0⋆\Xi_0^\star.

Moreover, Ξ0⋆\Xi_0^\star has infected sites arbitrarily far behind its edge. Indeed, fix jj. By the concatenated construction and Lemma 3.3, for every k≥jk \ge j,

Ξ0(−nk)(Γ(−nj)(0)−r0)=1.\Xi_0^{(-n_k)}\left(\Gamma^{(-n_j)}(0)-r_0\right)=1.

Passing to the limit with jj fixed gives

Ξ0⋆(Γ(−nj)(0)−r0)=1.\Xi_0^\star\left(\Gamma^{(-n_j)}(0)-r_0\right)=1.

Since r0−Γ(−nj)(0)→∞r_0-\Gamma^{(-n_j)}(0) \to\infty along the subsequence, these infected sites extend arbitrarily far to the left. Hence Ξ0⋆∈Xe\Xi_0^\star\in\mathcal{X}^{e} a.s.

Let νe\nu^{e} denote the law of Ξ0⋆\Xi_0^\star. By stationarity of the marked regenerative construction, the same backward procedure defines Ξt⋆\Xi_t^\star for every t∈Rt \in\mathbb{R}, with L(Ξt⋆)=νe\mathcal{L}(\Xi_t^\star)=\nu^{e}.

We next prove convergence. Fix ξ∈Xe\xi\in\mathcal{X}^{e}. By Lemma 4.3, the process started from ξ\xi enters the regenerative regime in finite time almost surely. Since (Ξt⋆)t≥0(\Xi_t^\star)_{t\geq0} is constructed from the stationary marked regeneration process with the same regenerative law, the coupling of Lemma 5.1 (i) applies to these two processes. Consequently, for every ℓ<∞\ell< \infty, there is an almost surely finite TℓT_\ell such that

Ξt(x)=Ξt⋆(x),−ℓ≤x≤0,\Xi_t(x)=\Xi_t^\star(x), \qquad-\ell\leq x \leq0,

for all t≥Tℓt \geq T_\ell. If ff is a bounded local function depending only on {−ℓ,…,0}\{-\ell,\ldots,0\}, then f(Ξt)=f(Ξt⋆)f(\Xi_t)=f(\Xi_t^\star) for all sufficiently large tt, almost surely. Since Ξt⋆∼νe\Xi_t^\star\sim\nu^{e} for every tt, bounded convergence gives

lim⁡t→∞Eλ,μξ[f(Ξt)]=∫f dνe.\lim_{t\to\infty}\mathbb{E}^{\xi}_{\lambda,\mu}[f(\Xi_t)]=\int f\,d\nu^{e}.

Hence, the weak convergence in (3) holds.

Since (Ξt⋆)t∈R(\Xi_t^\star)_{t\in\mathbb{R}} is a stationary version of the edge process, its marginal law νe\nu^{e} is invariant. Uniqueness follows from convergence to νe\nu^{e} from every initial configuration in Xe\mathcal{X}^{e} and dominated convergence. □

Remark 5.1 (Uniform convergence from finite initial sets). We briefly indicate how the proof of Theorem 1.6 could be adapted to obtain the uniform convergence stated in Remark 1.1.

The main difference is that the initial infected set is now finite and convergence is required uniformly over sets AA with diam⁡(A)≤ct\operatorname{diam}(A) \leq ct, conditional on survival up to time tt. The first step would be to show that, under this conditioning, the process enters the regenerative regime sufficiently early, uniformly over such sets. This could be achieved using the regeneration trials of § 3 and the corresponding quantitative estimates. Once regeneration has occurred, the coupling of Lemma 5.1 (i), together with the renewal estimates, would provide uniform control of the time needed for agreement with the stationary edge process in any fixed window. The condition c<vc < v would enter in controlling the initial transient relative to the linear growth of the infection front. The shielding argument used in the proof of Theorem 1.6 would then yield the desired convergence. We omit the details.

Complete convergence

We now turn to complete convergence. Unlike convergence seen from the edge, this requires controlling the process in fixed spatial windows, as well as the possibility of extinction. The proof combines entrance into the regenerative regime with shielding couplings from both directions.

Lemma 5.2 (Entrance into the regenerative regime). Let ξ∈{−1,0,1}Z\xi\in\{-1,0,1\}^{\mathbb{Z}} have finite nonempty infected set and finitely many immune sites, and set Sξ={τξ=∞}S_{\xi}=\{\tau^{\xi}=\infty\}. On SξS_{\xi}, the process enters the regenerative regime in finite time almost surely, to the right or to the left.

After translating the accepted regeneration point to (0,0)(0,0) and reflecting space if necessary, the retained one-sided graphical state has law QQ, independently of the previously revealed information.

Proof. On SξS_{\xi}, write

ℓt:=inf⁡Itξ,rt:=sup⁡Itξ.\ell_t:=\inf I_t^{\xi},\qquad r_t:=\sup I_t^{\xi}.

We first observe that survival forces the infection to leave every finite interval. Indeed, the CPTI restricted to [−m,m][-m,m], with infection attempts leaving the interval suppressed, is a finite-state Markov chain that almost surely reaches a configuration without infections. Since the restricted and unrestricted processes agree until the first infection outside [−m,m][-m,m],

Pλ,μ(Sξ,Itξ⊂[−m,m] for all t≥0)=0.\mathbb{P}_{\lambda,\mu}\left(S_{\xi}, I_t^{\xi}\subset[-m,m]\text{ for all }t\ge0\right)=0.

Taking the union over mm gives

Pλ,μ(Sξ∩{sup⁡t≥0rt<∞, inf⁡t≥0ℓt>−∞})=0.(25)\mathbb{P}_{\lambda,\mu}\left(S_{\xi}\cap\left\{\sup_{t\ge0}r_t<\infty,\ \inf_{t\ge0}\ell_t>-\infty\right\}\right)=0. \tag*{(25)}

Suppose first that sup⁡t≥0rt=∞\sup_{t\ge0}r_t=\infty. We apply the adaptive record-level trial construction of Subsection 3.2. After each rejected trial, the next candidate level is chosen beyond the running record and all previously revealed graphical information. By Lemma 3.4, each trial is therefore accepted, conditionally on the past, with probability at least

pbr=qprepθ^>0.p_{\mathrm{br}}=q_{\mathrm{prep}}\widehat{\theta}>0.

Since every rejected trial is resolved in finite time almost surely, the probability that the first nn trials are all rejected is at most (1−pbr)n(1-p_{\mathrm{br}})^n. Hence a regeneration occurs almost surely on the event {sup⁡t≥0rt=∞}\{\sup_{t\ge0}r_t=\infty\}.

The reflected construction applies when the infection escapes to the left. Together with (25), this proves the entrance assertion. The statement concerning the retained graphical law follows from Proposition 3.5, with reflection in the left-hand case. □

We prove complete convergence by comparing finite-start processes with the nontrivial invariant law of [1]. Finite approximations, two-sided shielding, and the regenerative coupling identify this law as the limit conditional on survival.

Proof of Theorem 1.7. Let ν\nu be the nontrivial translation-invariant invariant measure of Andjel and Schinazi [1], normalized to have no absorbing component. By translation invariance, a configuration sampled from ν\nu cannot have a finite nonempty infected set. Stationarity and the absence of an absorbing component then imply that it has infinitely many infected sites almost surely. Let ζ0\zeta_0 have distribution ν\nu, and let ζ0(n)\zeta_0^{(n)} agree with ζ0\zeta_0 on [−n,n][-n,n] and be susceptible elsewhere. Couple their evolutions using the same graphical construction. Since ζ0\zeta_0 has infinitely many infected sites almost surely, we may condition on ζ0\zeta_0 and select arbitrarily many well-separated infected sites. The toy graphical construction is independent of ζ0\zeta_0. Positive toy survival, reflection symmetry, and spatial mixing therefore imply that, with probability tending to one as n→∞n \to\infty, there are toy infection paths starting from infected sites in [−n,n][-n,n] and expanding in opposite directions. These paths may be chosen with the leftward-expanding path starting to the left of the origin and the rightward-expanding path starting to its right, so that their starting points enclose the origin. Since the initial configurations agree between these paths, Lemma 3.3 gives, for every finite Λ⊂Z\Lambda\subset\mathbb{Z},

lim⁡n→∞lim sup⁡t→∞P(ζt(n)∣Λ≠ζt∣Λ)=0.\lim_{n \to\infty}\limsup_{t \to\infty}\mathbb{P}\left(\zeta_t^{(n)}\big|_{\Lambda} \ne\zeta_t\big|_{\Lambda}\right)=0.

In fact, these paths are common legal infection paths for both evolutions. Since the initial configurations agree between their starting points, Lemma 3.3, applied to each path (with spatial reflection for the right-hand barrier), shows that the two evolutions agree at all subsequent times in the region between the paths. By ballisticity, this region eventually contains every fixed finite interval. In particular, the survival probability of ζ(n)\zeta^{(n)} tends to one.

Now fix finite nonempty A⊂ZA \subset\mathbb{Z} and write SA={τA=∞}S_A=\{\tau^A=\infty\}. The proof of Lemma 5.2 applies equally to finite-support initial configurations with immune sites. Thus, for each fixed nn, Lemma 5.1 (ii) couples the surviving processes started from AA and ζ0(n)\zeta_0^{(n)} so that they eventually agree on every fixed finite interval.

Since ζt\zeta_t has distribution ν\nu for every tt, the preceding approximation and coupling imply, by letting first t→∞t \to\infty and then n→∞n \to\infty, that

Pλ,μ(ξtA∈⋅∣SA)⟹ν.\mathbb{P}_{\lambda,\mu}\left(\xi_t^A \in\mathord{\cdot}\mid S_A\right)\Longrightarrow\nu.

On SAcS_A^c, the infection dies out in finite time and the remaining immunity intervals expire. Consequently,

Pλ,μ(ξtA∈⋅)⟹Pλ,μ(SAc)δ0+Pλ,μ(SA)ν.\mathbb{P}_{\lambda,\mu}\left(\xi_t^A \in\mathord{\cdot}\right)\Longrightarrow\mathbb{P}_{\lambda,\mu}(S_A^c)\delta_0+\mathbb{P}_{\lambda,\mu}(S_A)\nu.

Finally, let AA be infinite and choose finite sets An↑AA_n \uparrow A. By Proposition 2.2, reflection symmetry, and FKG, together with spatial mixing, the probability that two toy infection paths expanding in opposite directions arise from sites of AnA_n tends to one. Coupling the processes from AnA_n and AA through the same graphical construction, Lemma 3.3 gives eventual agreement on every fixed finite interval, outside an event of probability tending to zero as n→∞n \to\infty.

These paths also imply that Pλ,μ(SA)=1\mathbb{P}_{\lambda,\mu}(S_A)=1 and Pλ,μ(SAn)→1\mathbb{P}_{\lambda,\mu}(S_{A_n})\to1. Complete convergence for finite initial sets, followed by n→∞n \to\infty, therefore yields

Pλ,μ(ξtA∈⋅)⟹ν.\mathbb{P}_{\lambda,\mu}\left(\xi_t^A \in\mathord{\cdot}\right)\Longrightarrow\nu.

Extinction in the subcritical return regime

The extinction argument relies on the observation that survival from a finite infected set requires repeated infections of previously visited sites. The key is to control these reinfections: each infection of a new site creates a possible route for the infection to return, but such a return requires the former source to recover, lose its immunity, and be reinfected before its neighbor recovers. This mechanism is reminiscent of the Markov-modulated point-process comparison of Broman [3], although our argument uses conditional first-moment estimates rather than stochastic domination.

Proof of Theorem 1.1. An infection event at a site xx is called a return infection if xx was infected at some earlier time, including time zero. Otherwise, it is called a first infection. By [10 Theorem 5], survival from a finite infected set implies infinitely many return infections. Thus, it suffices to show that the number of return infections is almost surely finite.

We organize the return infections into generations according to their infection genealogy. A return infection belongs to generation 1 if its ancestral line contains no earlier return infection. Inductively, a return infection belongs to generation m+1m+1 if the most recent return infection on its ancestral line belongs to generation mm. Let Rm\mathcal{R}_m denote the set of return infections in generation mm, and write Zm:=∣Rm∣Z_m := |\mathcal{R}_m|.

Starting from the space-time points corresponding to Rm\mathcal{R}_m, explore all legal descendants, stopping each branch when the next return infection is encountered. Let Am\mathcal{A}_m be the collection of first-infection edges discovered during this exploration, and GmG_m the number of return gates at generation mm.

We first count the possible gates leading to Rm+1\mathcal{R}_{m+1}. Consider an infection edge u→vu \to v discovered during the exploration. Immediately after this infection, both sites are infected. A return infection may subsequently occur either at uu, through an arrow from vv, or at vv, through an arrow from uu. In either case, the site receiving the infection must first recover and lose its immunity, while the other site must remain infected until the return occurs.

Thus, each infection edge gives rise to two possible directed return gates. Counting also the edges corresponding to Rm\mathcal{R}_m, we obtain

Gm≤2(Zm+∣Am∣).(26)G_m \le2(Z_m + |\mathcal{A}_m|). \tag*{(26)}

Fix an element of Rm\mathcal{R}_m and one spatial direction. Each successive first infection requires an infection arrow before the next recovery mark, which occurs with probability at most λ/(1+λ)\lambda/(1+\lambda). Thus, the expected number of first-infection edges in that direction is at most

∑n≥1(λ1+λ)n=λ.\sum_{n \ge1} \left(\frac{\lambda}{1+\lambda}\right)^n = \lambda.

Summing over both directions and all elements of Rm\mathcal{R}_m yields

Eλ,μ∣Am∣≤2λEλ,μZm.(27)\mathbb{E}_{\lambda,\mu}|\mathcal{A}_m| \le2\lambda\mathbb{E}_{\lambda,\mu}Z_m. \tag*{(27)}

It remains to estimate the probability that a candidate gate succeeds. Consider a candidate gate associated with an infection edge at time ss. Let uu denote the potential recipient of the return infection and vv its potential source. Immediately after the edge is created, both sites are infected. We evaluate the gate at its creation time. Conditional on the graphical history up to time ss, let Ru,Rv∼iidExp⁡(1)R_u, R_v \mathrel{\overset{\mathrm{iid}}{\sim}} \operatorname{Exp}(1) be the residual recovery times of uu and vv, let Iu∼Exp⁡(μ)I_u \sim\operatorname{Exp}(\mu) be the immunity time of uu after its recovery, and let Av,u∼Exp⁡(λ)A_{v,u} \sim\operatorname{Exp}(\lambda) be the waiting time, after the immunity of uu is lost, for the next infection arrow v→uv \to u. These random variables are independent.

For this gate to produce the next return infection, it is necessary that

Ru+Iu+Av,u<Rv.R_u + I_u + A_{v,u} < R_v.

The geometry of the exploration and this reverse-gate mechanism are illustrated in Figure 7. Consequently, by independence and the exponential law, the conditional probability that a candidate gate succeeds is at most

Return-gate exploration diagram

Figure 7. Return-gate exploration. The return infection e0∈Rme_0 \in\mathcal{R}_m generates first-infection edges e1,e2∈Ame_1,e_2 \in\mathcal{A}_m. The figure illustrates a possible return through the reverse gate associated with e1e_1.

q:=Pλ,μ(Ru+Iu+Av,u<Rv)=Eλ,μ[e−(Ru+Iu+Av,u)]=12μ1+μλ1+λ.\begin{aligned} q &:= \mathbb{P}_{\lambda,\mu}(R_u+I_u+A_{v,u}<R_v) \\ &= \mathbb{E}_{\lambda,\mu}\left[e^{-(R_u+I_u+A_{v,u})}\right] \\ &= \frac{1}{2}\frac{\mu}{1+\mu}\frac{\lambda}{1+\lambda}. \end{aligned}

The estimate is made at the creation of the gate, before any information about its descendants is revealed. By symmetry, this bound applies to both orientations of each candidate gate. Charging each candidate gate at its creation time and using the tower property, (6.1) and (6.2) yield

Eλ,μZm+1≤qEλ,μGm≤2qEλ,μ[Zm+∣Am∣]≤μλ(1+2λ)(1+μ)(1+λ)Eλ,μZm.\begin{aligned} \mathbb{E}_{\lambda,\mu}Z_{m+1} &\le q\mathbb{E}_{\lambda,\mu}G_m \\ &\le2q\mathbb{E}_{\lambda,\mu}\left[Z_m+\lvert A_m\rvert\right] \\ &\le\frac{\mu\lambda(1+2\lambda)}{(1+\mu)(1+\lambda)}\mathbb{E}_{\lambda,\mu}Z_m. \end{aligned}

Writing mret=μλ(1+2λ)/[(1+μ)(1+λ)]m_{\mathrm{ret}}=\mu\lambda(1+2\lambda)/[(1+\mu)(1+\lambda)], condition (1) gives mret<1m_{\mathrm{ret}}<1. Since the initial infected set is finite, the same geometric exploration, started from its sites, gives Eλ,μZ1<∞\mathbb{E}_{\lambda,\mu}Z_1<\infty. Consequently,

Pλ,μ(Zm>0)≤Eλ,μZm≤mretm−1Eλ,μZ1⟶0.\mathbb{P}_{\lambda,\mu}(Z_m>0)\le\mathbb{E}_{\lambda,\mu}Z_m\le m_{\mathrm{ret}}^{m-1}\mathbb{E}_{\lambda,\mu}Z_1\longrightarrow0.

Since Eλ,μZm<∞\mathbb{E}_{\lambda,\mu}Z_m<\infty for every mm, each generation is almost surely finite. Consequently, infinitely many return infections would require nonempty generations of arbitrarily large order. But

Pλ,μ(Zm>0)⟶0,\mathbb{P}_{\lambda,\mu}(Z_m>0)\longrightarrow0,

so there are almost surely only finitely many return infections. By [10 Theorem 5], survival from a finite infected set is therefore impossible, completing the proof. □

Moment estimates for regeneration cycles

We establish exponential moment bounds for the duration, displacement, and maximal oscillation of regeneration cycles. The latter controls deviations of the right edge from its regenerative skeleton, allowing limit theorems for the regeneration increments to extend to the full right-edge process.

Proof of Proposition 4.1. We first consider a generic post-first cycle started from QQ, using the notation of Subsection 3.2. Thus NN is the first accepted trial and

τ+=TN+s∗,X+=aN+1,M+=sup⁡0≤t≤τ+∣rt∣.\tau_{+}=T_N+s_{*}, \qquad X_{+}=a_N+1, \qquad M_{+}=\sup_{0\leq t\leq\tau_{+}}|r_t|.

By Proposition 3.6, (X+,τ+,M+)(X_{+},\tau_{+},M_{+}) has the common law of (ΔXn,Δτn,Mn)(\Delta X_n,\Delta\tau_n,M_n) for n≥2n\geq2. We first control the cost of rejected trials. Write

Fi:={the i-th trial is rejected}.F_i:=\{\text{the }i\text{-th trial is rejected}\}.

Conditionally on the information Gi\mathcal{G}_i available before the ii-th trial, Lemma 3.4 gives the fresh-trial law. A preparation failure has a uniformly bounded cost. If the preparation succeeds and the launched toy process dies, we use the standard restart consequence of Proposition 2.2: there exist c,C>0c,C>0 such that

Pμ(t<τtoy<∞)≤Ce−ct,t≥0.(28)\mathbb{P}_{\mu}\left(t<\tau^{\mathrm{toy}}<\infty\right)\leq Ce^{-ct}, \qquad t\geq0. \tag*{(28)}

Indeed, after each finite failed block cluster one restarts the block construction from the surviving toy process, after the fixed local seed-preparation step. The uniform positive survival probability and the exponential tail for the height of a finite cluster give (A.1) by the usual geometric restart argument.

Moreover, the toy process can be coupled in the obvious way with a classical contact process, and the CPTI right edge is bounded above by the right edge of a classical contact process under the common Harris construction. The standard finite-speed estimate for the classical contact process therefore controls both the spatial range revealed while the toy failure is resolved and any simultaneous rightward CPTI displacement. Consequently there exist c,C>0c,C>0, uniformly in ii, such that

PQ(Oi≥t,Fi∣Gi)≤Ce−ct,t≥0.(29)\mathbb{P}_{Q}\left(O_i\geq t, F_i\mid\mathcal{G}_i\right)\leq Ce^{-ct}, \qquad t\geq0. \tag*{(29)}

Since a successful preparation is followed by toy survival with conditional probability θ^>0\hat{\theta}>0, the fresh-trial construction also gives

PQ(Fi∣Gi)≤1−pbra.s.(30)\mathbb{P}_{Q}\left(F_i\mid\mathcal{G}_i\right)\leq1-p_{\mathrm{br}} \qquad\text{a.s.} \tag*{(30)}

We now control the accepted level. By (10),

aN≤N+∑i=1N−1Oi.(31)a_N\leq N+\sum_{i=1}^{N-1}O_i. \tag*{(31)}

Choose α>0\alpha> 0 so small that (A.2) implies

K:=sup⁡i∥EQ[eαOi1Fi∣Gi]∥∞<∞.K := \sup_i \left\lVert\mathbb{E}_Q\left[e^{\alpha O_i}\mathbf{1}_{F_i}\mid\mathcal{G}_i\right]\right\rVert_\infty< \infty.

For 0<δ<α0 < \delta< \alpha, conditional Hölder and (A.3) give

EQ[eδ(Oi+1)1Fi∣Gi]≤eδKδ/α(1−pbr)1−δ/α.\mathbb{E}_Q\left[e^{\delta(O_{i+1})}\mathbf{1}_{F_i}\mid\mathcal{G}_i\right] \le e^{\delta}K^{\delta/\alpha}(1-p_{\mathrm{br}})^{1-\delta/\alpha}.

The right-hand side tends to 1−pbr<11-p_{\mathrm{br}}<1 as δ↓0\delta\downarrow0. Hence, for δ>0\delta>0 sufficiently small, there is χ<1\chi<1 such that

EQ[eδ(Oi+1)1Fi∣Gi]≤χa.s.(32)\mathbb{E}_Q\left[e^{\delta(O_{i+1})}\mathbf{1}_{F_i}\mid\mathcal{G}_i\right]\le\chi\qquad\text{a.s.} \tag*{(32)}

On {N=n}\{N=n\}, (A.4) yields

eδaN≤eδ∏i=1n−1eδ(Oi+1)1Fi.e^{\delta a_N}\le e^\delta\prod_{i=1}^{n-1}e^{\delta(O_{i+1})}\mathbf{1}_{F_i}.

Iterating (A.5) and summing over n≥1n\ge1 gives

EQ[eδaN]<∞.(33)\mathbb{E}_Q[e^{\delta a_N}]<\infty. \tag*{(33)}

We next control the cycle duration. Write

rttoy:=sup⁡ηt{0},rˉttoy:=sup⁡0≤s≤trstoy,r_t^{\mathrm{toy}}:=\sup\eta_t^{\{0\}},\qquad\bar r_t^{\mathrm{toy}}:=\sup_{0\le s\le t}r_s^{\mathrm{toy}},

with the convention that rttoy=−∞r_t^{\mathrm{toy}}=-\infty when ηt{0}=∅\eta_t^{\{0\}}=\varnothing. The same restart construction, together with the linear growth of the infinite good-block path in Proposition 2.2, gives constants v0,c,C>0v_0,c,C>0 such that

Pμ(τtoy=∞, rˉttoy≤v0t−C)≤Ce−ct,t≥0.(34)\mathbb{P}_\mu\left(\tau^{\mathrm{toy}}=\infty,\ \bar r_t^{\mathrm{toy}}\le v_0t-C\right)\le Ce^{-ct},\qquad t\ge0. \tag*{(34)}

Under QQ the distinguished toy process is conditioned on survival. Since every toy infection path is a legal CPTI infection path, (A.7) therefore implies, after changing constants,

PQ(rˉt≤v0t−C)≤Ce−ct,t≥0.(35)\mathbb{P}_Q(\bar r_t\le v_0t-C)\le Ce^{-ct},\qquad t\ge0. \tag*{(35)}

By construction, aN>rˉρN−1a_N>\bar r_{\rho_{N-1}}. Hence aNa_N has not been reached before the NN-th trial and

TN=inf⁡{t≥0:rt=aN}.T_N=\inf\{t\ge0:r_t=a_N\}.

Since the right edge moves by nearest-neighbor jumps,

{TN>t}⊆{aN>v0t−C}∪{rˉt≤v0t−C}.(36)\{T_N>t\}\subseteq\{a_N>v_0t-C\}\cup\{\bar r_t\le v_0t-C\}. \tag*{(36)}

Combining (A.6), (A.8), and (A.9), we obtain

PQ(TN>t)≤Ce−ct,t≥0.(37)\mathbb{P}_Q(T_N>t)\le Ce^{-ct},\qquad t\ge0. \tag*{(37)}

Since τ+=TN+s∗\tau_+=T_N+s_* and X+=aN+1X_+=a_N+1, after decreasing δ\delta if necessary,

EQ[eδτ+]+EQ[eδ∣X+∣]<∞.(38)\mathbb{E}_Q[e^{\delta\tau_+}]+\mathbb{E}_Q[e^{\delta|X_+|}]<\infty. \tag*{(38)}

It remains to control the oscillation during the cycle. Choose ε>0\varepsilon>0 small. For the rightward excursion, the CPTI is dominated by the usual contact process started from the infected half-line, so the standard finite-speed estimate gives an exponential bound. For the leftward excursion, if rs≤−mr_s\le-m, then the distinguished barrier, which is infected at time ss, must also lie at or to the left of −m-m. Since this barrier is a toy infection path and hence a classical graphical infection path, the same finite-speed estimate applies. Conditioning on toy survival only changes the multiplicative constant. Thus there exist c,C>0c,C>0 such that

PQ(sup⁡0≤s≤εm∣rs∣≥m)≤Ce−cm,m≥0,(39)\mathbf{P}_{Q}\left(\sup_{0\leq s\leq\varepsilon m}|r_s|\geq m\right)\leq Ce^{-cm},\qquad m\geq0, \tag*{(39)}

Moreover,

{M+≥m}⊆{τ+>εm}∪{sup⁡0≤s≤εm∣rs∣≥m}.(40)\{M_+\geq m\}\subseteq\{\tau_+>\varepsilon m\}\cup\left\{\sup_{0\leq s\leq\varepsilon m}|r_s|\geq m\right\}. \tag*{(40)}

By (A.10), (A.12), and (A.13),

PQ(M+≥m)≤Ce−cm,m≥0,\mathbf{P}_{Q}(M_+\geq m)\leq Ce^{-cm},\qquad m\geq0,

and hence, after decreasing δ\delta once more,

EQ[eδM+]<∞.(41)\mathbf{E}_{Q}[e^{\delta M_+}]<\infty. \tag*{(41)}

Equations (A.11) and (A.14), together with Proposition 3.6, prove (19). For the first cycle, the rejected-trial estimate (A.2) and the contraction argument remain valid. The block construction from the initially infected half-line and the classical finite-speed comparison give

Pμ(rˉt≤v0t−C)≤Ce−ct,Pμ(sup⁡0≤s≤εm∣rs∣≥m)≤Ce−cm.(42)\mathbf{P}_{\mu}(\bar r_t\leq v_0t-C)\leq Ce^{-ct},\qquad \mathbf{P}_{\mu}\left(\sup_{0\leq s\leq\varepsilon m}|r_s|\geq m\right)\leq Ce^{-cm}. \tag*{(42)}

Writing τ1=T(1)+s∗\tau_1=T^{(1)}+s_* and X1=a(1)+1X_1=a^{(1)}+1, the preceding arguments, with (A.15) replacing the corresponding QQ-estimates, yield

Eμ[eδτ1]+Eμ[eδ∣X1∣]+Eμ[eδM1]<∞,\mathbf{E}_{\mu}[e^{\delta\tau_1}]+\mathbf{E}_{\mu}[e^{\delta|X_1|}]+\mathbf{E}_{\mu}[e^{\delta M_1}]<\infty,

proving (18). □\square

Acknowledgements

R.S. is supported by CNPq grant 403423/2023-6. C.T. is supported by CNPq grant 151432/2024-4. M.E.V. is partially supported by CNPq grant 310734/2021-5 and by FAPERJ grant E-261200.473/2026.

Statement on AI use

Artificial intelligence (AI) tools were used solely for linguistic proofreading and assistance with the code used to produce figures. The authors reviewed and verified all AI-assisted changes and code. AI tools were not employed in the formulation of mathematical results or the development of proofs. The authors take full responsibility for the mathematical content and the final manuscript.

{sup}1` Institute of Mathematics, École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland

{sup}2` Instituto de Matemática, Universidade Federal do Rio de Janeiro (UFRJ), Rio de Janeiro, Brazil

Email address: [email protected]

Email address: [email protected]

Email address: [email protected]

Email address: [email protected]

References

  1. [1]Enrique Andjel and Rinaldo B. Schinazi. A complete convergence theorem for an epidemic model. Journal of Applied Probability, 33(3):741–748, 1996. doi:10.2307/3215355.DOI
  2. [2]Søren Asmussen. Applied Probability and Queues. Springer, 2nd edition, 2003. doi:10.1007/b97236.DOI
  3. [3]Erik I. Broman. Stochastic domination for a hidden Markov chain with application to the contact process in a randomly evolving environment. The Annals of Probability, 35(6):2263–2293, 2007. doi:10.1214/009117906000001187.
  4. [4]John T. Cox, Richard Durrett, and Rinaldo B. Schinazi. The critical contact process seen from the right edge. Probability Theory and Related Fields, 87(3):325–332, 1991. doi:10.1007/BF01321213.DOI
  5. [5]Richard Durrett. Oriented percolation in two dimensions. The Annals of Probability, 12(4):999–1040, 1984. doi:10.1214/aop/1176993140.DOI
  6. [6]Richard Durrett and Claude Neuhauser. Epidemics with recovery in D = 2. The Annals of Applied Probability, 1(2):189–206, 1991. doi:10.1214/aoap/1177005933.DOI
  7. [7]Tobias Friedrich, Andreas Göbel, Nicolas Klodt, Martin S. Krejca, and Marcus Pappik. The irrelevance of influencers: Information diffusion with re-activation and immunity lasts exponentially long on social network models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17389–17397, 2024. doi:10.1609/aaai.v38i16.29687.DOI
  8. [8]Antonio Galves and Errico Presutti. Edge fluctuations for the one-dimensional supercritical contact process. The Annals of Probability, 15(3):1131–1145, 1987. doi:10.1214/aop/1176992086.DOI
  9. [9]Peter W. Glynn and Ward Whitt. Limit theorems for cumulative processes. Stochastic Processes and their Applications, 47(2):299–314, 1993. doi:10.1016/0304-4149(93)90019-Z.DOI
  10. [10]Andrew Heeszel. On the Speed, Fluctuations, and Survival of Non-Attractive Stochastic Growth Models in One Dimension. Doctoral dissertation, The Ohio State University, 2026. URL: http://rave.ohiolink.edu/etdc/view?acc_num=osu1776946712453256.
  11. [11]Andrew Heeszel and Matthew Wascher. Epidemics with avoidance and isolation on Z^d, 2026. doi:10.48550/arXiv.2609.03359.DOI
  12. [12]Marcelo Hilário, Daniel Ungaretti, Daniel Valesin, and Maria E. Vares. Results on the contact process with dynamic edges or under renewals. Electronic Journal of Probability, 27:1–31, 2022. doi:10.1214/22-EJP811.DOI
  13. [13]Jaewook Joo and Joel L. Lebowitz. Pair approximation of the stochastic susceptible-infected-recovered-susceptible epidemic model on the hypercubic lattice. Physical Review E, 70(3):036114, 2004. doi:10.1103/PhysRevE.70.036114.DOI
  14. [14]Thomas Kuczek. The central limit theorem for the right edge of supercritical oriented percolation. The Annals of Probability, 17(4):1322–1332, 1989. doi:10.1214/aop/1176991157.DOI
  15. [15]Phuc Lam, Oanh Nguyen, and Iris Yang. Optimal bound for survival time of the SIRS process on star graphs, 2024. doi:10.48550/arXiv.2412.21138.
  16. [16]Thomas M. Liggett, Roberto H. Schonmann, and Alan M. Stacey. Domination by product measures. The Annals of Probability, 25(1):71–95, 1997. doi:10.1214/aop/1024404279.DOI
  17. [17]Torgny Lindvall and Leonard C. G. Rogers. On coupling of random walks and renewal processes. Journal of Applied Probability, 33(1):122–126, 1996. doi:10.2307/3215269.DOI
  18. [18]Amitai Linker and Daniel Remenik. The contact process with dynamic edges on Z. Electronic Journal of Probability, 25:21–21, 2020. doi:10.1214/20-EJP480.DOI
  19. [19]Thomas S. Mountford and Ted D. Sweet. An extension of Kuczek’s argument to nonnearest neighbor contact processes. Journal of Theoretical Probability, 13(4):1061–1081, 2000. doi:10.1023/A:1007818108889.DOI
  20. [20]Daniel Remenik. The contact process with a dynamic random environment. The Annals of Applied Probability, 18(6):2392–2420, 2008. doi:10.1214/08-AAP528.DOI
  21. [21]Hermann Thorisson. The coupling of regenerative processes. Advances in Applied Probability, 15(3):531–561, 1983. doi:10.2307/1426618.DOI
  22. [22]Achilleas Tzioufas. On the growth of the one-dimensional reverse immunization contact processes. Journal of Applied Probability, 48(3):611–623, 2011. doi:10.1239/jap/1316796902.DOI
  23. [23]Jacob van den Berg, Geoffrey R. Grimmett, and Rinaldo B. Schinazi. Dependent random graphs and spatial epidemics. The Annals of Applied Probability, 8(2):317–336, 1998. doi:10.1214/aoap/1028903529.

Paper details

vlnc:W_3vc6bc8m
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents