betaFeedback

WorksMathematics

Avalanche Sharpness and the Stationary State of the Bak–Sneppen Model on Zd\mathbb Z^d

Abstract

We establish the stationary picture for the nearest-neighbour Bak–Sneppen model on Zd\mathbb Z^d, for every d≥1d\ge1. There exists a critical level tct_c such that, as N→∞N\to\infty, every fixed family of stationary fitnesses on the torus (Z/NZ)d(\mathbb Z/N\mathbb Z)^d converges, in exponential coordinates, to independent random variables with law tc+Exp⁡(1)t_c+\operatorname{Exp}(1). The proof relies on three estimates for avalanches. For d≥2d\ge2, these replace the interval structure available in dimension one. We show that the threshold at which the mean avalanche range diverges coincides with the threshold for the appearance of an infinite avalanche, using the hierarchical structure of avalanches and an estimate on the creation of new sites under translation. We then prove that an avalanche reaching a macroscopic distance on the torus can, after any fixed positive increase of the level, be amplified into one covering the whole torus. Finally, below criticality, we obtain exponential integrability of the avalanche range and give a finite–infinite volume comparison. These three ingredients allow us to implement the Meester and Znamenski program that connects these avalanche estimates to the stationary law of the finite-volume process.

Introduction and main results

The Bak–Sneppen model is a simple model of self-organized criticality [2]. Each vertex of a finite graph carries a fitness, and at every step the vertex of smallest fitness and its neighbours receive fresh independent values. Despite the simplicity of this rule, the stationary fitness distribution is expected to develop, as the system grows, a sharp threshold: fitnesses typically lie above this level, while the dynamics produces bursts of updates involving smaller values. These bursts are called avalanches. Fixing a level tt and following the updates until all fitnesses are again above tt provides a natural way to isolate and study them [10, 14].

Meester and Znamenski [11, 12] proposed a rigorous program connecting these avalanches to the stationary large-volume limit; see also [6, 7, 15, 17]. The key is to show that three critical levels coincide: the level at which the mean avalanche duration becomes infinite, the level at which the mean number of sites visited becomes infinite, and the level at which an avalanche can last forever. Together with suitable finite-volume estimates, this identifies the stationary limit as a product measure supported above a common fitness threshold. We recall this program in Section 6.

For the one-dimensional one-sided model, the range size admits a closed scalar evolution [8]. In the isotropic one-dimensional model, the range remains an interval, but the scalar description based on the graft length alone no longer closes, since the locations of the two endpoints matter. For d≥2d \ge2, avalanche geometry is even more complex, largely unknown, and no comparable scalar description is available. Figure 1 illustrates this nested geometry in dimension two. Our aim is to replace these one-dimensional reductions by estimates that remain effective in arbitrary dimension. A key quantity is the overlap of independent avalanche ranges.

A near-critical avalanche on $\mathbb{Z}^2$ at level $t=0.398$

Figure 1. A near-critical avalanche on Z2\mathbb{Z}^2 at level t=0.398t = 0.398 (tc≈0.3988t_c \approx0.3988). Sites are coloured by their entry level τ(x)\tau(x) in the monotone coupling, revealing the nested avalanche ranges. Of the 133 grafted avalanches, only 43 reach new sites, accounting for the 44 colours. The + marks the origin. This realization is the largest among approximately 9000 independent avalanches at this level. (Most avalanches are still small at criticality).

We consider the model on the discrete torus

ΛN=(Z/NZ)d,d≥1,\Lambda_N = (\mathbb{Z}/N\mathbb{Z})^d,\qquad d \ge1,

and use exponential coordinates for the fitnesses. Thus every refreshed fitness has law Exp⁡(1)\operatorname{Exp}(1), and an update at xx refreshes the sites in

x+Nd,Nd:={0,±e1,…,±ed}.x + \mathcal{N}_d,\qquad\mathcal{N}_d := \{0,\pm e_1,\ldots,\pm e_d\}.

Let πN\pi_N be the stationary law of this finite-volume Markov chain, and, for x∈Zdx \in\mathbb{Z}^d, let FN,xF_{N,x} denote the stationary fitness at the image of xx in ΛN\Lambda_N. Our main result identifies the large-volume stationary state.

Theorem 1.1 (Stationary limit). There exists tc∈(0,∞)t_c \in(0,\infty) such that, for every fixed collection of distinct sites x1,…,xk∈Zdx_1,\ldots,x_k \in\mathbb{Z}^d,

(FN,x1,…,FN,xk)→dN→∞(tc+E1,…,tc+Ek),(F_{N,x_1},\ldots,F_{N,x_k}) \xrightarrow[d]{N\to\infty} (t_c+E_1,\ldots,t_c+E_k),

where E1,…,EkE_1,\ldots,E_k are independent Exp⁡(1)\operatorname{Exp}(1) random variables.

For y≥0y \ge0, let νy\nu_y denote the law of y+Ey+E, where E∼Exp⁡(1)E \sim\operatorname{Exp}(1). Under the increasing change of variables

u=1−e−f,u = 1-e^{-f},

Theorem 1.1 is equivalent to the usual uniform representation: the limiting fitnesses are independent and uniform on

[1−e−tc,1].[1-e^{-t_c},1].

The critical level tct_c is characterized through infinite-volume avalanches. Fix t≥0t \ge0, start from a configuration in which all fitnesses are larger than tt, and force an update at the origin. Thereafter, as long as some current fitness is at most tt, update the site carrying the smallest fitness and its neighbours. The procedure stops when all current fitnesses are again larger than tt. The set of sites refreshed during this procedure (including the forced initial update) is the range of the tt-avalanche.

Let ξ(t)\xi(t) denote this range on Zd\mathbb{Z}^d, let η(t)\eta(t) be its duration, including the forced update and with value ∞\infty if the avalanche does not terminate, and put

r(t):=∣ξ(t)∣,R(t):=Er(t),D(t):=Eη(t),θ(t):=P(r(t)=∞).r(t):=|\xi(t)|,\qquad R(t):=\mathbb{E}r(t),\qquad D(t):=\mathbb{E}\eta(t),\qquad\theta(t):=\mathbb{P}(r(t)=\infty).

We use throughout the natural monotone coupling in the level parameter, under which ξ(s)⊆ξ(t)\xi(s) \subseteq\xi(t) for s≤ts \le t. Avalanche ranges are connected, since every newly refreshed site is a nearest neighbour of a site already reached by the avalanche.

Following Meester and Znamenski, introduce the three critical levels

td:=sup⁡{t:D(t)<∞},tsus:=sup⁡{t:R(t)<∞},t∞:=inf⁡{t:θ(t)>0}.t_d:=\sup\{t:D(t)<\infty\},\qquad t_{\mathrm{sus}}:=\sup\{t:R(t)<\infty\},\qquad t_\infty:=\inf\{t:\theta(t)>0\}.

Every update refreshes at most 2d+12d+1 sites, hence r(t)≤(2d+1)η(t)r(t)\le(2d+1)\eta(t), while positive mass at infinity forces infinite mean range. Therefore

td≤tsus≤t∞.t_d\le t_{\mathrm{sus}}\le t_\infty.

Both inequalities could a priori be strict. Our first result identifies all three levels.

Theorem 1.2 (Avalanche sharpness). For the nearest-neighbour Bak–Sneppen model on Zd\mathbb{Z}^d, d≥1d \ge1,

td=tsus=t∞=:tc.t_d=t_{\mathrm{sus}}=t_\infty=:t_c.

Moreover, for every t<tct<t_c,

tc≤t+10R(t).(1)t_c\le t+\frac{10}{R(t)}. \tag*{(1)}

Remark 1.3. The equality

td=tsust_d=t_{\mathrm{sus}}

was already derived by Meester and Znamenski [12 Theorem 3.3 and Section 6], through the range–duration relation, stated here as Lemma 2.2. The novelty here is the identification tsus=t∞t_{\mathrm{sus}}=t_\infty, together with the quantitative bound (1).

Remark 1.4. Finiteness of tct_c follows directly from (1), taking t=0t=0, since R(0)=2d+1R(0)=2d+1. Positivity of tct_c is elementary: in the exploration of a tt-avalanche, each update produces 2d+12d+1 fresh fitnesses, each of which is at most tt with probability 1−e−t1-e^{-t}. Ignoring overlaps therefore gives a domination by a Galton–Watson process with offspring distribution

Bin⁡(2d+1,1−e−t).\operatorname{Bin}(2d+1,1-e^{-t}).

This process is subcritical whenever (2d+1)(1−e−t)<1(2d+1)(1-e^{-t})<1, which implies finite mean avalanche duration for all sufficiently small t>0t>0, and hence tc>0t_c>0.

For d=1d=1, Theorem 1.2 already suffices to complete the Meester–Znamenski program [11], recalled in § 6, and derive the stationary-limit result. In dimensions d≥2d \ge2, two additional finite-volume ingredients are required.

The first ingredient concerns avalanches on the finite torus. Let ξN(t)\xi_N(t) denote the range of the tt-avalanche on ΛN\Lambda_N, and set

rN(t):=∣ξN(t)∣,RN(t):=ErN(t),PN(t):=P(ξN(t)=ΛN).r_N(t) := |\xi_N(t)|,\qquad R_N(t) := \mathbb{E}r_N(t),\qquad P_N(t) := \mathbb{P}(\xi_N(t)=\Lambda_N).

The following result says that once a supercritical avalanche has reached a macroscopic distance, any further fixed increase of the level amplifies it into complete coverage of the torus.

Theorem 1.5 (Macroscopic amplification). Fix tc<s<tt_c < s < t and 0<ρ<120 < \rho< \frac{1}{2}.

Let

EN(s):={ξN(s)∩{x:dΛN(0,x)=⌊ρN⌋}≠∅}.E_N(s) := \{\xi_N(s) \cap\{x : d_{\Lambda_N}(0,x)=\lfloor\rho N\rfloor\} \ne\varnothing\}.

Then

P(ξN(t)=ΛN∣EN(s))⟶1.\mathbb{P}(\xi_N(t)=\Lambda_N \mid E_N(s)) \longrightarrow1.

Consequently,

lim inf⁡N→∞PN(t)≥θ(s)>0.\liminf_{N\to\infty} P_N(t) \ge\theta(s)>0.

The complementary statement below shows that, in the subcritical regime, an avalanche is typically too small to detect the periodic geometry of the torus.

Theorem 1.6 (Subcritical exponential moments and locality). For every t<tct<t_c, there exists λt>0\lambda_t>0 such that

E[eλtr(t)]<∞.\mathbb{E}[e^{\lambda_t r(t)}] < \infty.

In particular, all polynomial moments of r(t)r(t) are finite. Moreover,

sup⁡0≤u≤t∣RN(u)−R(u)∣⟶0.\sup_{0\le u\le t}|R_N(u)-R(u)| \longrightarrow0.

Proof strategy

The proof for general d≥1d \ge1 is organized around the following three mechanisms.

1. Infinite-volume sharpness: creation under translation. The level parameter gives avalanches a hierarchical structure. Conditionally on a finite current range AA, each site x∈Ax \in A produces marks at rate one, each grafting an independent avalanche translated by xx. The relevant geometric quantity is the part of this translated range lying outside AA.

The deterministic input is

∑x∈A∣(x+B)∖A∣≥c∣A∣∣B∣,0∈B,∣B∣≥∣A∣≥3,\sum_{x\in A}|(x+B)\setminus A| \ge c|A||B|,\qquad0\in B,\qquad|B|\ge|A|\ge3,

which follows from a translation-invariant total order on Zd\mathbb{Z}^d. Applied to

ψ(t,λ):=E[1−e−λr(t)],\psi(t,\lambda):=\mathbb{E}[1-e^{-\lambda r(t)}],

the hierarchical decomposition gives the Burgers-type inequality

∂tψ(t,λ)≥110ψ(t,λ)∂λψ(t,λ).\partial_t\psi(t,\lambda)\ge\frac{1}{10}\psi(t,\lambda)\partial_\lambda\psi(t,\lambda).

Since ψ(t,0+)=θ(t)\psi(t,0+)=\theta(t), while ∂λψ(t,0+)=R(t)\partial_\lambda\psi(t,0+)=R(t) whenever r(t)<∞r(t)<\infty almost surely, the characteristic flow prevents the boundary slope from becoming infinite strictly before positive boundary mass appears. This yields Theorem 1.2.

2. Supercritical amplification: translation expansion. An avalanche reaching distance of order NN need not resemble a ball or be well mixed on the torus. We retain only one nearest-neighbour edge crossing each successive radius.

For A⊂ΛNA\subset\Lambda_N, let

bA(z):=∣{x∈A:x+z∉A}∣b_A(z):=\lvert\{x\in A:x+z\notin A\}\rvert

be its translation boundary. It satisfies

bA(z+w)≤bA(z)+bA(w),∑z∈ΛNbA(z)=∣A∣(Nd−∣A∣).b_A(z+w)\le b_A(z)+b_A(w),\qquad\sum_{z\in\Lambda_N}b_A(z)=\lvert A\rvert(N^d-\lvert A\rvert).

Fix 0<ρ<1/20<\rho<1/2, and let EE be the event that an independent avalanche reaches distance ρN\rho N. On EE, select uniformly one crossing edge at each radius 1≤j≤L:=⌊ρN⌋1\le j\le L:=\lfloor\rho N\rfloor. Let κ\kappa be their expected endpoint counting measure, with zero contribution off EE. The selected edges need not form a connected path.

At a fixed supercritical level, P(E)\mathbb{P}(E) is bounded away from zero uniformly in NN, so κ(ΛN)=2LP(E)\kappa(\Lambda_N)=2L\mathbb{P}(E) is of order NN. Hyperoctahedral symmetry, subadditivity of bAb_A, and the crossing edges then imply

∑zκ(z)bA(z)≥cN∣A∣(Nd−∣A∣)Nd.\sum_z\kappa(z)b_A(z)\ge cN\frac{\lvert A\rvert(N^d-\lvert A\rvert)}{N^d}.

Thus the averaged endpoint measure supplies enough translations to expand any nontrivial set AA.

To turn expansion into covering, we use the Lyapunov function

Φ(A):=Nd∣A∣−1.\Phi(A):=\frac{N^d}{\lvert A\rvert}-1.

Conditionally on macroscopic reach at the initial level, the preceding estimate implies exponential decay of the mean of Φ\Phi at rate of order NN. Since every noncovering range satisfies Φ≥1/(Nd−1)\Phi\ge1/(N^d-1), any fixed positive sprinkling then forces complete coverage with high (conditional) probability.

3. Subcritical locality: exponential moments replace exact geometry. Below tct_c, the hierarchy gives a transport inequality for the truncated exponential transform of the avalanche range. Finite susceptibility can thus be bootstrapped to exponential integrability of r(t)r(t), making ranges of order NN exponentially unlikely.

Every connected subset of Zd\mathbb{Z}^d with fewer than NN vertices projects injectively onto ΛN\Lambda_N. A subcritical avalanche is therefore overwhelmingly unlikely to detect the periodic geometry. This gives the finite–infinite volume comparison of Theorem 1.6.

These mechanisms provide the inputs for the locking-threshold representation; see (25) and [11 Section 2]. Sharpness identifies the candidate threshold, while amplification and subcritical locality control the locking thresholds from above and below. The stationary limit then follows from a dimension-free finite-volume argument.

Choice of the underlying graph. Several parts of the argument extend beyond Zd\mathbb{Z}^d, for instance to suitable Cayley graphs and their finite quotients. One could formulate the proof in such an abstract setting, replacing translations and the amplification argument by appropriate group-theoretic and spectral assumptions. We have chosen not to do so: the lattice Zd\mathbb{Z}^d already contains the essential difficulties, and its geometry makes the amplification mechanism particularly transparent.

Methodological perspective

Two analogies help place the proof. The first is with coagulation and gelation. For solvable Smoluchowski coagulation equations, Laplace transforms lead to first-order nonlinear equations related to inviscid Burgers, while gelation corresponds to mass escaping from the finite-cluster distribution; see [1, 13]. Spatial overlaps prevent a closed equation for the avalanche law, but the translation estimate still yields

∂tψ≥cψ∂λψ.\partial_t\psi\ge c\psi\partial_\lambda\psi.

This one-sided inequality is enough for the characteristic argument: divergence of the susceptibility is transported into positive boundary mass at a higher level, signalling an infinite avalanche.

The second analogy concerns amplification. The measure κ\kappa may be viewed as the jump measure of a translation walk on ΛN\Lambda_N, and the expansion estimate is equivalent, up to constants, to a spectral-gap bound for this auxiliary walk; see Remark 4.4 and compare [5]. Here the estimate follows directly from subadditivity of the translation boundary and the selected crossing edges, without any shape or mixing assumption on the macroscopic range.

Disclosure on the use of AI

Disclosure on the use of AI. Human was used at the level of an annoyingly slow co-author in this work (see Appendix § 7 for a more detailed account of our interactions with AI). In view of the extent of the AI contribution, we do not presently intend to submit this work for publication.

Relation with [8]. The present paper supersedes the first version of [8], whose main goal was to establish sharpness in dimension 1. Accordingly, that work has been refocused on quantitative bounds for the critical tail of the avalanche range for the anisotropic model. This is a genuinely different question, on which the present method, despite its strength for proving sharpness, has little to say.

Organization of the paper

Section 2 gives the avalanche hierarchy and its basic identities. Section 3 proves infinite-volume sharpness. Section 4 establishes the finite-volume expansion inequality and the amplification theorem. Section 5 proves exponential moments of the subcritical range and the finite–infinite volume comparison. The stationary limit is deduced in Section 6.

Avalanche hierarchy and basic identities

We collect here the structural facts about avalanches which are common to all three parts of the proof.

Recall that ξ(t)\xi(t) and ξN(t)\xi_N(t) denote the ranges of the tt-avalanches rooted at the origin in Zd\mathbb{Z}^d and ΛN\Lambda_N, respectively. We use the standard monotone graphical construction in the level parameter, so that

ξ(s)⊆ξ(t),s≤t.\xi(s) \subseteq\xi(t), \qquad s \le t.

The key property is regeneration in the level variable. We use the censored exploration filtration (Hu)u≥0(\mathcal{H}_u)_{u\ge0}: during the exploration of a uu-avalanche, refreshed values at most uu are revealed exactly, whereas values larger than uu are revealed only through the event that they exceed uu. The σ\sigma-field Hu\mathcal{H}_u records the explored genealogy, the identities of the refresh variables encountered, and this partially revealed information. The monotone graphical construction makes this filtration increasing.

Proposition 2.1 (Hierarchical avalanche decomposition). *Consider either the infinite lattice or the finite torus. On the event that the range AA at level uu is finite, conditionally on Hu\mathcal{H}_u, the terminal fitnesses at its sites are independent with law νu\nu_u. Each site x∈Ax \in A

therefore carries an independent rate-one clock in the level parameter. At a mark rooted at xx and level uu, an independent uu-avalanche ξ′(u)\xi'(u) is grafted onto the current range:

A⟼A∪(x+ξ′(u)).A \longmapsto A \cup(x+\xi'(u)).

The graft is independent of the censored past. In infinite volume this description is used on finite-range states, up to explosion, or with the process stopped when the range reaches a prescribed finite size.

Proof. We first verify that finite range implies finite duration almost surely, so that the terminal fitness at every reached site is well defined. Indeed, while the range has at most KK sites, there are at most KK active fitnesses at or below uu. A block of KK updates in which all fresh values exceed uu removes these active fitnesses without creating new ones, and therefore forces termination within the block. Such a block has conditional probability at least

e−(2d+1)uK>0.e^{-(2d+1)uK}>0.

The probability of completing mm blocks without either terminating or reaching more than KK sites is consequently at most (1−e−(2d+1)uK)m(1-e^{-(2d+1)uK})^m. Letting m→∞m\to\infty, and then taking the union over KK, proves the assertion.

Since the avalanche has finite duration, each reached site has a last refresh. The censored exploration determines the identities of these last refresh variables, without revealing their values beyond the fact that they exceed uu. As these variables are distinct, independence and the memoryless property imply that, conditionally on Hu\mathcal{H}_u, the terminal fitnesses are independent with law u+Exp⁡(1)u+\operatorname{Exp}(1).

At the next triggering level, fresh refresh variables generate an independent avalanche at that level, translated to the triggering site. Iterating this argument gives the stated grafting description.

Equivalently, on the torus over any finite level interval, and in infinite volume over [0,T][0,T] with T<t∞T<t_\infty, independently for each site xx, take Poisson marks (u,B)(u,B) with intensity du P(ξN(u)∈dB)du\,\mathbb{P}(\xi_N(u)\in dB) on the torus and du P(ξ(u)∈dB)du\,\mathbb{P}(\xi(u)\in dB) in infinite volume. Starting from Ξ0=Nd\Xi_0=\mathbb{N}^d, retain a mark (u,B)(u,B) whenever xx belongs to the current range, in which case

Ξu=Ξu−∪(x+B).\Xi_u=\Xi_{u-}\cup(x+B).

This realizes the preceding level-by-level description. On the torus it is directly nonexplosive. In infinite volume, monotonicity gives

r(u)≤r(T)<∞a.s.,u≤T,r(u)\le r(T)<\infty\quad\text{a.s.}, \qquad u\le T,

and one may equivalently stop when the range reaches KK sites and then let K→∞K\to\infty. □

Thus, conditionally on a current range AA, marks occur at total rate ∣A∣|A|. This observation is the common source of the compensation identity of Section 3, the growth estimates of Section 4, and the transport inequality of Section 5. We shall also use the same construction with each graft decorated by the duration of the corresponding independent avalanche; this gives the range–duration identity below.

Let ηN(t)\eta_N(t) denote the duration of the finite-volume tt-avalanche. With η(t)\eta(t) and D(t)D(t) as above, set

DN(t):=EηN(t).D_N(t):=\mathbb{E}\eta_N(t).

Lemma 2.2 (Range–duration identity). For every t<tsust<t_{\mathrm{sus}}, D(t)<∞D(t)<\infty, and on the torus DN(t)<∞D_N(t)<\infty for every finite tt. Moreover, RR is continuous on [0,tsus)[0,t_{\mathrm{sus}}), RNR_N is continuous on [0,∞)[0,\infty), and DD and DND_N are C1C^1 on these respective domains, with

D′(t)=R(t)D(t),DN′(t)=RN(t)DN(t),D'(t)=R(t)D(t), \qquad D_N'(t)=R_N(t)D_N(t),

and

D(0)=DN(0)=1.D(0)=D_N(0)=1.

Consequently,

D(t)=exp⁡{∫0tR(u) du},DN(t)=exp⁡{∫0tRN(u) du}.(2)D(t)=\exp\left\{\int_0^t R(u)\,\mathrm{d}u\right\},\qquad D_N(t)=\exp\left\{\int_0^t R_N(u)\,\mathrm{d}u\right\}. \tag*{(2)}

Proof. Fix T<tsusT<t_{\mathrm{sus}}, so that R(u)≤R(T)<∞R(u)\le R(T)<\infty for u≤Tu\le T. In the decorated hierarchical construction a mark at level uu adds exactly the duration η′(u)\eta'(u) of an independent uu-avalanche. For L≥1L\ge1, put DL(u):=E[η(u)∧L]D_L(u):=\mathbb{E}[\eta(u)\wedge L]. Since

(η+η′)∧L−η∧L≤η′∧L,(\eta+\eta')\wedge L-\eta\wedge L\le\eta'\wedge L,

compensation, first stopped when the range reaches KK sites, gives after letting $K\to\infty

DL(t)≤1+∫0tR(u)DL(u) du,0≤t≤T.D_L(t)\le1+\int_0^t R(u)D_L(u)\,\mathrm{d}u,\qquad0\le t\le T.

Here the rate of marks is the current range size, and the graft is independent of the censored past. The stopping can be removed because r(T)<∞r(T)<\infty almost surely, the observable is bounded by LL, and its expected compensator is bounded by LR(u)LR(u). Gronwall’s inequality yields

DL(t)≤exp⁡{∫0tR(u) du}.D_L(t)\le\exp\left\{\int_0^t R(u)\,\mathrm{d}u\right\}.

Monotone convergence as L→∞L\to\infty proves D(t)<∞D(t)<\infty, uniformly on [0,T][0,T]. The same argument on the torus uses RN(u)≤NdR_N(u)\le N^d and proves DN(t)≤eNdt<∞D_N(t)\le e^{N^dt}<\infty on every bounded level interval.

We next record the continuity of the mean ranges. Couple the avalanches ξ(u)\xi(u), 0≤u≤T0\le u\le T, through the same hierarchical construction. Since r(T)<∞r(T)<\infty almost surely, only finitely many relevant marks occur below level TT, and no deterministic level carries a mark. Hence, for any deterministic tn→tt_n\to t in [0,T][0,T], one has ξ(tn)=ξ(t)\xi(t_n)=\xi(t), and therefore r(tn)=r(t)r(t_n)=r(t), eventually almost surely. Since r(tn)≤r(T)r(t_n)\le r(T) and Er(T)=R(T)<∞\mathbb{E}r(T)=R(T)<\infty, dominated convergence yields R(tn)→R(t)R(t_n)\to R(t). Thus RR is continuous on [0,tsus)[0,t_{\mathrm{sus}}). The same argument on the torus is immediate from rN(T)≤Ndr_N(T)\le N^d, and shows that RNR_N is continuous on every bounded level interval.

We may now apply compensation to the full duration. Tonelli’s theorem gives

D(t)=1+∫0tR(u)D(u) du,DN(t)=1+∫0tRN(u)DN(u) du.D(t)=1+\int_0^t R(u)D(u)\,\mathrm{d}u,\qquad D_N(t)=1+\int_0^t R_N(u)D_N(u)\,\mathrm{d}u.

In infinite volume the stopping by range size is again removed by monotone convergence of the accumulated duration and compensator. The integral identities first give continuity of DD and DND_N. Together with the continuity of RR and RNR_N, the fundamental theorem of calculus then yields the stated C1C^1 identities. Since D,DN≥1D,D_N\ge1, integrating (log⁡D)′=R(\log D)'=R and (log⁡DN)′=RN(\log D_N)'=R_N gives (2). □

Infinite-volume sharpness

We prove Theorem 1.2. The hierarchical avalanche decomposition provides an exact compensation formula for a bounded transform of the range, while a deterministic translation estimate ensures that an independent graft creates sufficiently many new sites on average. Their combination yields a Burgers-type differential inequality.

For λ>0\lambda>0, put

ψ(t,λ):=E[1−e−λr(t)]\psi(t,\lambda):=\mathbb{E}\left[1-e^{-\lambda r(t)}\right]

with the convention

e−λ∞=0.e^{-\lambda\infty}=0.

Lemma 3.1 (Avalanche compensation). For every λ>0\lambda> 0 and 0≤s<t<t∞0 \le s < t < t_\infty,

ψ(t,λ)−ψ(s,λ)=∫stH(u,λ) du,(3)\psi(t,\lambda)-\psi(s,\lambda)=\int_s^t H(u,\lambda)\,\mathrm{d}u, \tag*{(3)}

where, if ξ′(u)\xi'(u) is an independent copy of ξ(u)\xi(u),

H(u,λ):=E[e−λr(u)∑x∈ξ(u)(1−e−λ∣(x+ξ′(u))∖ξ(u)∣)].(4)H(u,\lambda):=\mathbb{E}\left[e^{-\lambda r(u)}\sum_{x\in\xi(u)}\left(1-e^{-\lambda\lvert(x+\xi'(u))\setminus\xi(u)\rvert}\right)\right]. \tag*{(4)}

In particular, t↦ψ(t,λ)t\mapsto\psi(t,\lambda) is locally Lipschitz on [0,t∞)[0,t_\infty). The functions HH and ∂λψ\partial_\lambda\psi are jointly continuous on [0,t∞)×(0,∞)[0,t_\infty)\times(0,\infty), and

∂tψ(t,λ)=H(t,λ),∂λψ(t,λ)=E[r(t)e−λr(t)].\partial_t\psi(t,\lambda)=H(t,\lambda),\qquad\partial_\lambda\psi(t,\lambda)=\mathbb{E}[r(t)e^{-\lambda r(t)}].

The tt-derivative at t=0t=0 is understood from the right.

Proof. Fix T<t∞T<t_\infty, and stop the hierarchical construction when the range reaches KK sites. Before this stopping time the total rate of relevant marks is at most KK, so compensation applies without any integrability assumption.

For

Fλ(A):=1−e−λ∣A∣,F_\lambda(A):=1-e^{-\lambda\lvert A\rvert},

a graft BB rooted at x∈Ax\in A produces the increment

Fλ(A∪(x+B))−Fλ(A)=e−λ∣A∣(1−e−λ∣(x+B)∖A∣).F_\lambda(A\cup(x+B))-F_\lambda(A)=e^{-\lambda\lvert A\rvert}\left(1-e^{-\lambda\lvert(x+B)\setminus A\rvert}\right).

Compensation for the stopped construction therefore gives the stopped version of (3). The compensator is uniformly bounded in the stopping level, since

∣A∣e−λ∣A∣≤1eλ.\lvert A\rvert e^{-\lambda\lvert A\rvert}\le\frac{1}{e\lambda}.

Moreover,

P(r(T)≥K)⟶0.\mathbb{P}(r(T)\ge K)\longrightarrow0.

Dominated convergence as K→∞K\to\infty proves (3) and (4). The same bound gives

∣ψ(t,λ)−ψ(s,λ)∣≤∣t−s∣eλ,\lvert\psi(t,\lambda)-\psi(s,\lambda)\rvert\le\frac{\lvert t-s\rvert}{e\lambda},

hence local Lipschitz continuity.

For joint continuity, couple two independent range processes on [0,T][0,T], with T<t∞T<t_\infty. The pathwise local-constancy argument established in the proof of Lemma 2.2 applies verbatim: since T<t∞T<t_\infty, the terminal ranges are finite almost surely, and for any deterministic tn→tt_n\to t in [0,T][0,T], both ranges are eventually unchanged almost surely. No moment assumption is needed here. If also λn→λ>0\lambda_n\to\lambda>0, the integrand defining HH converges almost surely and is bounded by 1/(eλ∗)1/(e\lambda_*), where λ∗>0\lambda_*>0 is a common lower bound for the exponents. Dominated convergence gives joint continuity of HH. The same bound justifies differentiation of ψ\psi under the expectation and joint continuity of ∂λψ\partial_\lambda\psi; the integrand of ψ\psi itself is bounded by one. Finally, (3) and continuity of HH yield ∂tψ=H\partial_t\psi=H at every level in the stated domain. □

The geometric input is the following elementary inequality.

Lemma 3.2 (Average creation under translation). Let A,B⊂ZdA,B \subset\mathbb{Z}^d be finite, with 0∈B0 \in B, and write

a:=∣A∣,b:=∣B∣.a := |A|,\qquad b := |B|.

If b≥a≥3b \ge a \ge3, then

∑x∈A∣(x+B)∖A∣≥15ab.(5)\sum_{x\in A} |(x+B)\setminus A| \ge\frac{1}{5}ab. \tag*{(5)}

Proof. Write C(A,B):=∑x∈A∣(x+B)∖A∣C(A,B):=\sum_{x\in A}|(x+B)\setminus A|, and first assume b=ab=a. Equip Zd\mathbb{Z}^d with the translation-invariant lexicographic order, and write

A={a1<⋯<aa},p:=∣{z∈B:z>0}∣,q:=∣{z∈B:z<0}∣.A=\{a_1<\cdots<a_a\},\qquad p:=|\{z\in B:z>0\}|,\qquad q:=|\{z\in B:z<0\}|.

Then p+q=a−1p+q=a-1. From aia_i, at most a−ia-i positive displacements and i−1i-1 negative displacements can land in AA. Hence

C(A,B)≥∑i=1a((p−a+i)++(q−i+1)+)=p(p+1)+q(q+1)2≥⌊a24⌋.C(A,B)\ge\sum_{i=1}^{a}\bigl((p-a+i)_+ +(q-i+1)_+\bigr)=\frac{p(p+1)+q(q+1)}{2}\ge\left\lfloor\frac{a^2}{4}\right\rfloor.

Here v+:=max⁡{v,0}v_+:=\max\{v,0\}.

For general b≥ab\ge a, choose uniformly a subset B0⊂BB_0\subset B of size aa containing zero. The zero displacement contributes nothing, while every other displacement is retained with probability (a−1)/(b−1)(a-1)/(b-1). Applying the equal-cardinality case gives

a−1b−1C(A,B)=E[C(A,B0)]≥⌊a24⌋.\frac{a-1}{b-1}C(A,B)=\mathbb{E}[C(A,B_0)]\ge\left\lfloor\frac{a^2}{4}\right\rfloor.

Since (b−1)/(a−1)≥b/a(b-1)/(a-1)\ge b/a and a≥3a\ge3,

C(A,B)≥ba⌊a24⌋≥15ab,C(A,B)\ge\frac{b}{a}\left\lfloor\frac{a^2}{4}\right\rfloor\ge\frac{1}{5}ab,

as claimed. □

Remark 3.3. The proof of Lemma 3.2 uses only the existence of a translation-invariant total order and the cardinalities of AA and BB. In particular, neither isoperimetry nor any metric property of Zd\mathbb{Z}^d enters the infinite-volume sharpness argument.

We now combine the two lemmas.

Proposition 3.4 (Burgers comparison). For every 0≤s<t<t∞0\le s<t<t_\infty and λ>0\lambda>0,

ψ(t,λ)−ψ(s,λ)≥110∫stψ(u,λ) ∂λψ(u,λ) du.(6)\psi(t,\lambda)-\psi(s,\lambda)\ge\frac{1}{10}\int_s^t\psi(u,\lambda)\,\partial_\lambda\psi(u,\lambda)\,\mathrm{d}u. \tag*{(6)}

Equivalently,

∂tψ(t,λ)≥110ψ(t,λ)∂λψ(t,λ)\partial_t\psi(t,\lambda)\ge\frac{1}{10}\psi(t,\lambda)\partial_\lambda\psi(t,\lambda)

for every 0≤t<t∞0\le t<t_\infty.

Proof. Fix u<t∞u<t_\infty, and let

A:=ξ(u),B:=ξ′(u),X:=∣A∣,Y:=∣B∣.A:=\xi(u),\qquad B:=\xi'(u),\qquad X:=|A|,\qquad Y:=|B|.

Then XX and YY are independent and identically distributed, finite almost surely, and

X,Y≥2d+1≥3.X,Y\ge2d+1\ge3.

For x∈Ax\in A, put

dx:=∣(x+B)∖A∣.d_x:=|(x+B)\setminus A|.

On {Y≥X}\{Y \ge X\}, Lemma 3.2 gives

∑x∈Adx≥15XY.\sum_{x\in A} d_x \ge\frac{1}{5}XY.

Since r↦1−e−λrr \mapsto1-e^{-\lambda r} is concave, vanishes at zero, and 0≤dx≤Y0 \le d_x \le Y,

1−e−λdx≥dxY(1−e−λY).1-e^{-\lambda d_x} \ge\frac{d_x}{Y}(1-e^{-\lambda Y}).

Consequently, on {Y≥X}\{Y \ge X\},

∑x∈A(1−e−λdx)≥X5(1−e−λY).\sum_{x\in A}(1-e^{-\lambda d_x}) \ge\frac{X}{5}(1-e^{-\lambda Y}).

Lemma 3.1 therefore gives

H(u,λ)≥15E[Xe−λX(1−e−λY)1{Y≥X}].(7)H(u,\lambda) \ge\frac{1}{5}\mathbb{E}\left[Xe^{-\lambda X}(1-e^{-\lambda Y})\mathbf{1}_{\{Y\ge X\}}\right]. \tag*{(7)}

Set

f(r):=re−λr,g(r):=1−e−λr.f(r):=re^{-\lambda r}, \qquad g(r):=1-e^{-\lambda r}.

The ratio

f(r)g(r)=reλr−1\frac{f(r)}{g(r)}=\frac{r}{e^{\lambda r}-1}

is decreasing on (0,∞)(0,\infty). Hence, on {X<Y}\{X<Y\},

f(Y)g(X)≤f(X)g(Y).f(Y)g(X) \le f(X)g(Y).

By exchangeability,

E[f(X)g(Y)1{Y<X}]=E[f(Y)g(X)1{X<Y}]≤E[f(X)g(Y)1{X<Y}].\mathbb{E}[f(X)g(Y)\mathbf{1}_{\{Y<X\}}]=\mathbb{E}[f(Y)g(X)\mathbf{1}_{\{X<Y\}}] \le\mathbb{E}[f(X)g(Y)\mathbf{1}_{\{X<Y\}}].

Therefore

E[f(X)g(Y)1{Y≥X}]≥12E[f(X)g(Y)].\mathbb{E}[f(X)g(Y)\mathbf{1}_{\{Y\ge X\}}]\ge\frac{1}{2}\mathbb{E}[f(X)g(Y)].

Independence gives

E[f(X)g(Y)]=E[Xe−λX]E[1−e−λX]=∂λψ(u,λ)ψ(u,λ),\mathbb{E}[f(X)g(Y)]=\mathbb{E}[Xe^{-\lambda X}]\mathbb{E}[1-e^{-\lambda X}] =\partial_\lambda\psi(u,\lambda)\psi(u,\lambda),

where differentiation under the expectation is justified by

re−λr≤1eλ.re^{-\lambda r}\le\frac{1}{e\lambda}.

Combining this with (7),

H(u,λ)≥110ψ(u,λ)∂λψ(u,λ).H(u,\lambda)\ge\frac{1}{10}\psi(u,\lambda)\partial_\lambda\psi(u,\lambda).

Integrating in (3) proves (6). □

The next lemma records the characteristic consequence of the Burgers comparison.

Lemma 3.5 (Comparison of the characteristics). Fix t0<t∞t_0<t_\infty and λ0>0\lambda_0>0, and put q:=ψ(t0,λ0)q:=\psi(t_0,\lambda_0).

Define

λ(s):=λ0−q10(s−t0)\lambda(s):=\lambda_0-\frac{q}{10}(s-t_0)

as long as λ(s)>0\lambda(s)>0. Then

ψ(s,λ(s))≥q\psi(s,\lambda(s))\ge q

for every s≥t0s \ge t_0 such that

s<t∞,λ(s)>0.s < t_\infty,\qquad\lambda(s)>0.

Proof. Fix t1>t0t_1>t_0 with t1<t∞t_1<t_\infty and λ(t1)>0\lambda(t_1)>0, and put v(s):=ψ(s,λ(s))v(s):=\psi(s,\lambda(s)) for s∈[t0,t1]s\in[t_0,t_1]. Lemma 3.1 permits the ordinary chain rule. Proposition 3.4 therefore gives

v′(s)=∂tψ(s,λ(s))−q10∂λψ(s,λ(s))≥110∂λψ(s,λ(s))(v(s)−q).v'(s)=\partial_t\psi(s,\lambda(s))-\frac{q}{10}\partial_\lambda\psi(s,\lambda(s))\ge\frac{1}{10}\partial_\lambda\psi(s,\lambda(s))(v(s)-q).

Set a(s):=110∂λψ(s,λ(s))a(s):=\frac{1}{10}\partial_\lambda\psi(s,\lambda(s)), which is continuous on this interval. Multiplication by the integrating factor gives

dds[e−∫t0sa(u) du(v(s)−q)]≥0.\frac{d}{ds}\left[e^{-\int_{t_0}^{s}a(u)\,du}(v(s)-q)\right]\ge0.

Since v(t0)=qv(t_0)=q, it follows that v(s)≥qv(s)\ge q on [t0,t1][t_0,t_1]. The choice of t1t_1 was arbitrary. □

Proof of Theorem 1.2. Fix t<t∞t<t_\infty and λ>0\lambda>0, and set

q:=ψ(t,λ)>0.q:=\psi(t,\lambda)>0.

Consider

λ(s)=λ−q10(s−t).\lambda(s)=\lambda-\frac{q}{10}(s-t).

Its hitting time of zero is

s∗=t+10λq.s_* = t+\frac{10\lambda}{q}.

We claim that

s∗≥t∞.s_*\ge t_\infty.

Indeed, if s∗<t∞s_*<t_\infty, then Lemma 3.5 gives

ψ(s,λ(s))≥q,t≤s<s∗.\psi(s,\lambda(s))\ge q,\qquad t\le s<s_*.

Since r(s∗)<∞r(s_*)<\infty almost surely and r(s)≤r(s∗)r(s)\le r(s_*),

0≤ψ(s,λ(s))≤E[1−e−λ(s)r(s∗)]⟶0(s↑s∗),0\le\psi(s,\lambda(s))\le\mathbb{E}[1-e^{-\lambda(s)r(s_*)}]\longrightarrow0\qquad(s\uparrow s_*),

by dominated convergence, a contradiction. Therefore

t+10λψ(t,λ)≥t∞,t+\frac{10\lambda}{\psi(t,\lambda)}\ge t_\infty,

or equivalently

ψ(t,λ)λ≤10t∞−t.\frac{\psi(t,\lambda)}{\lambda}\le\frac{10}{t_\infty-t}.

Since r(t)<∞r(t)<\infty almost surely,

1−e−λr(t)λ↑r(t)(λ↓0).\frac{1-e^{-\lambda r(t)}}{\lambda}\uparrow r(t)\qquad(\lambda\downarrow0).

Monotone convergence yields

R(t)≤10t∞−t.R(t)\le\frac{10}{t_\infty-t}.

Thus R(t)<∞R(t)<\infty for every t<t∞t<t_\infty, and hence

tsus≥t∞.t_{\mathrm{sus}}\ge t_\infty.

The reverse inequality is immediate. Therefore

tsus=t∞=:tc.t_{\mathrm{sus}}=t_\infty=:t_c.

It remains to identify the duration threshold. Since r(t)≤(2d+1)η(t)r(t) \le(2d+1)\eta(t), finite mean duration implies finite mean range, so td≤tsust_{\mathrm d} \le t_{\mathrm{sus}}. Conversely, Lemma 2.2 gives D(t)<∞D(t) < \infty for every t<tsust < t_{\mathrm{sus}}, hence td≥tsust_{\mathrm d} \ge t_{\mathrm{sus}}. Thus

td=tsus=t∞=tc,t_{\mathrm d}=t_{\mathrm{sus}}=t_{\infty}=t_{\mathrm c},

and (1) follows. □

Remark 3.6. The quantity

ψ(t,λ)=E[1−e−λr(t)]\psi(t,\lambda)=\mathbb{E}[1-e^{-\lambda r(t)}]

may be viewed as the probability that the avalanche range is hit by an independent Poisson field of intensity λ\lambda. Infinite susceptibility corresponds to an infinite right slope at λ=0\lambda=0, whereas an infinite avalanche corresponds to positive mass at the boundary:

ψ(t,0+)=θ(t).\psi(t,0+)=\theta(t).

The Burgers comparison forces positive values of ψ\psi to propagate towards this boundary. Estimate (1) expresses quantitatively that the boundary slope cannot diverge strictly before positive boundary mass appears.

Finite-volume amplification

We prove Theorem 1.5. Fix

d≥1,tc<s<t,0<ρ<12,d \ge1,\qquad t_{\mathrm c}<s<t,\qquad0<\rho<\frac{1}{2},

and put

ΛN:=(Z/NZ)d,V:=Nd,L:=⌊ρN⌋.\Lambda_N := (\mathbb{Z}/N\mathbb{Z})^d,\qquad V:=N^d,\qquad L:=\lfloor\rho N\rfloor.

For j≤Lj\le L, identify

Sj:={x∈Zd:∥x∥1=j}S_j:=\{x\in\mathbb{Z}^d:\|x\|_1=j\}

with its canonical image in ΛN\Lambda_N. Since 2L<N2L<N for all sufficiently large NN, this identification is unambiguous. Let

EN(u):={ξN(u)∩SL≠∅}.E_N(u):=\{\xi_N(u)\cap S_L\ne\varnothing\}.

The finite- and infinite-volume graphical constructions agree until the range first reaches SLS_L. Therefore

P(EN(u))=P(ξ(u)∩SL≠∅)≥θ(u).(8)\mathbb{P}(E_N(u))=\mathbb{P}(\xi(u)\cap S_L\ne\varnothing)\ge\theta(u). \tag*{(8)}

In particular,

P(EN(u))≥θ(s)>0,u≥s.\mathbb{P}(E_N(u))\ge\theta(s)>0,\qquad u\ge s.

For A⊂ΛNA\subset\Lambda_N, write a:=∣A∣a:=|A| and define its translation boundary:

bA(z):=∣{x∈A:x+z∉A}∣,z∈ΛN.b_A(z):=|\{x\in A:x+z\notin A\}|,\qquad z\in\Lambda_N.

Since AA and A−zA-z have equal cardinality,

bA(z)=12∣A△(A−z)∣,bA(−z)=bA(z),bA(z+w)≤bA(z)+bA(w).(9)b_A(z)=\frac{1}{2}|A\mathbin{\triangle}(A-z)|,\qquad b_A(-z)=b_A(z),\qquad b_A(z+w)\le b_A(z)+b_A(w). \tag*{(9)}

Moreover,

∑z∈ΛNbA(z)=a(V−a),(10)\sum_{z\in\Lambda_N}b_A(z)=a(V-a), \tag*{(10)}

since both sides count pairs in A×AcA\times A^c according to their displacement.

Lemma 4.1 (Short translations control all translations). Let f:Z/NZ→[0,∞)f:\mathbb{Z}/N\mathbb{Z}\to[0,\infty) satisfy f(0)=0f(0)=0 and

f(r+r′)≤f(r)+f(r′).f(r+r')\le f(r)+f(r').

For L=⌊ρN⌋L=\lfloor\rho N\rfloor, with 0<ρ<1/20<\rho<1/2 fixed, there exists Cρ<∞C_\rho<\infty such that, for all sufficiently large NN,

∑r∈Z/NZf(r)≤Cρ∑j=1Lf(j).\sum_{r\in\mathbb{Z}/N\mathbb{Z}}f(r)\le C_\rho\sum_{j=1}^{L}f(j).

Proof. Set

S:=∑j=1Lf(j).S:=\sum_{j=1}^{L}f(j).

For 1≤q<L1\le q<L, subadditivity gives

f(L)≤f(q)+f(L−q).f(L)\le f(q)+f(L-q).

Summing over qq yields

(L−1)f(L)≤2S.(L-1)f(L)\le2S.

Since L−1≥ρN/2L-1\ge\rho N/2 for all sufficiently large NN, it follows that f(L)≤4S/(ρN)f(L)\le4S/(\rho N). For 0≤r<N0\le r<N, write r=mL+qr=mL+q, where 0≤q<L0\le q<L. Then

f(r)≤mf(L)+f(q).f(r)\le mf(L)+f(q).

Also L≥ρN/2L\ge\rho N/2, hence m≤N/L≤2/ρm\le N/L\le2/\rho, and each value of qq occurs for at most 2/ρ+12/\rho+1 values of rr. Therefore

∑r∈Z/NZf(r)≤2Nρf(L)+(2ρ+1)S≤(8ρ2+2ρ+1)S.\sum_{r\in\mathbb{Z}/N\mathbb{Z}}f(r)\le\frac{2N}{\rho}f(L)+\left(\frac{2}{\rho}+1\right)S\le\left(\frac{8}{\rho^2}+\frac{2}{\rho}+1\right)S.

Let

Hd=(Z/2Z)d⋊SdH_d=(\mathbb{Z}/2\mathbb{Z})^d\rtimes\mathfrak{S}_d

be the hyperoctahedral group, acting on ΛN\Lambda_N by sign changes and coordinate permutations.

Proposition 4.2 (Translation expansion). Let EE be an event with positive probability. Suppose that on EE, for every 1≤j≤L1\le j\le L, we are given a nearest-neighbour edge (Wj,Zj)(W_j,Z_j), Wj∈Sj−1W_j\in S_{j-1}, Zj∈SjZ_j\in S_j, whose joint conditional law given EE is HdH_d-invariant. Define

κ:=E[1E∑j=1L(δWj+δZj)].\kappa:=\mathbb{E}\left[\mathbf{1}_E\sum_{j=1}^{L}(\delta_{W_j}+\delta_{Z_j})\right].

Then

κ(ΛN)=2LP(E),\kappa(\Lambda_N)=2L\mathbb{P}(E),

and there exists cd,ρ>0c_{d,\rho}>0 such that, for every A⊂ΛNA\subset\Lambda_N, writing a=∣A∣a=|A|, and all sufficiently large NN,

∑x∈A,z∈ΛNκ(z)1{x+z∉A}≥cd,ρκ(ΛN)a(V−a)V.(11)\sum_{\substack{x\in A,z\in\Lambda_N}}\kappa(z)\mathbf{1}_{\{x+z\notin A\}}\ge c_{d,\rho}\kappa(\Lambda_N)\frac{a(V-a)}{V}. \tag*{(11)}

The constant is uniform in NN, AA, and in the law of the selected edges.

Proof. Symmetrize the translation boundary by putting

f(z):=1∣Hd∣∑g∈HdbA(gz),I:=∑zκ(z)bA(z).f(z) := \frac{1}{\lvert H_d\rvert}\sum_{g\in H_d} b_A(gz), \qquad I := \sum_z \kappa(z)b_A(z).

The function ff is subadditive and HdH_d-invariant, with f(0)=0f(0)=0 and f(−z)=f(z)f(-z)=f(z). Invariance of κ\kappa and (10) give

I=∑zκ(z)f(z),∑z∈ΛNf(z)=a(V−a).I=\sum_z\kappa(z)f(z), \qquad\sum_{z\in\Lambda_N}f(z)=a(V-a).

Fix z∈Sjz\in S_j, and let a1,…,ada_1,\ldots,a_d be its absolute coordinates, so that ∑iai=j\sum_i a_i=j. Permute the coordinates to put aia_i first, then reflect that coordinate. The difference of the two resulting vectors is 2aie12a_i e_1, and subadditivity and invariance give

f(2aie1)≤2f(z).f(2a_i e_1)\le2f(z).

Consequently,

f(2je1)≤∑i=1df(2aie1)≤2df(z),f(2j e_1)\le\sum_{i=1}^{d}f(2a_i e_1)\le2d f(z),

and the endpoints ZjZ_j yield

I≥P(E)2d∑j=1Lf(2je1).I\ge\frac{\mathbb{P}(E)}{2d}\sum_{j=1}^{L}f(2j e_1).

For each selected edge, Zj−Wj=±eℓZ_j-W_j=\pm e_\ell, so

f(e1)=f(Zj−Wj)≤f(Zj)+f(Wj).f(e_1)=f(Z_j-W_j)\le f(Z_j)+f(W_j).

Summing gives I≥P(E)Lf(e1)I\ge\mathbb{P}(E)Lf(e_1). Together the two estimates imply

I≥P(E)4d{∑j=1Lf(2je1)+Lf(e1)}.(12)I\ge\frac{\mathbb{P}(E)}{4d}\left\{\sum_{j=1}^{L}f(2j e_1)+Lf(e_1)\right\}. \tag*{(12)}

To control all translations, decompose each vector into its coordinates. By subadditivity and coordinate symmetry,

∑z∈ΛNf(z)≤dNd−1∑r∈Z/NZf(re1).\sum_{z\in\Lambda_N}f(z)\le dN^{d-1}\sum_{r\in\mathbb{Z}/N\mathbb{Z}}f(re_1).

If NN is odd, multiplication by two permutes Z/NZ\mathbb{Z}/N\mathbb{Z}, so the last sum equals ∑rf(2re1)\sum_r f(2re_1). If NN is even, writing each odd translation as an even translation followed by e1e_1 gives

∑r∈Z/NZf(re1)≤∑r∈Z/NZf(2re1)+N2f(e1).\sum_{r\in\mathbb{Z}/N\mathbb{Z}}f(re_1)\le\sum_{r\in\mathbb{Z}/N\mathbb{Z}}f(2re_1)+\frac{N}{2}f(e_1).

Lemma 4.1, applied to r↦f(2re1)r\mapsto f(2re_1), gives

∑r∈Z/NZf(2re1)≤Cρ∑j=1Lf(2je1).\sum_{r\in\mathbb{Z}/N\mathbb{Z}}f(2re_1)\le C_\rho\sum_{j=1}^{L}f(2je_1).

Since L≥ρN/2L\ge\rho N/2 for all sufficiently large NN, the extra term in the even case satisfies

N2f(e1)≤1ρLf(e1).\frac{N}{2}f(e_1)\le\frac{1}{\rho}Lf(e_1).

Thus in both parity cases we conclude that

a(V−a)≤Cd,ρNd−1{∑j=1Lf(2je1)+Lf(e1)}.(13)a(V-a) \le C_{d,\rho}N^{d-1}\left\{\sum_{j=1}^{L} f(2je_1)+Lf(e_1)\right\}. \tag*{(13)}

Combining (12) and (13) yields

I≥cd,ρP(E)Na(V−a)V.I \ge c_{d,\rho}\mathbb{P}(E)N\frac{a(V-a)}{V}.

Since

κ(ΛN)=2LP(E)≤2ρNP(E),\kappa(\Lambda_N)=2L\mathbb{P}(E)\le2\rho N\mathbb{P}(E),

the preceding estimate, after changing the constant, gives

I≥cd,ρκ(ΛN)a(V−a)V.I \ge c_{d,\rho}\kappa(\Lambda_N)\frac{a(V-a)}{V}.

Finally,

I=∑x∈A,z∈ΛNκ(z)1{x+z∉A},I=\sum_{x\in A,z\in\Lambda_N}\kappa(z)\mathbf{1}_{\{x+z\notin A\}},

which proves (11).

Remark 4.3. We need both endpoints of the selected edges to handle parity. The vertices ZjZ_j alone control the even translations 2je12je_1. When NN is even, these do not control odd translations. Using also WjW_j, the nearest-neighbour relation Zj−Wj=±eℓZ_j-W_j=\pm e_\ell yields f(e1)≤f(Zj)+f(Wj)f(e_1)\le f(Z_j)+f(W_j), which bridges the two parity classes. For odd NN, multiplication by 22 is invertible modulo NN, and this extra ingredient is unnecessary.

Remark 4.4. Proposition 4.2 also gives a spectral-gap bound for the auxiliary translation walk. Put m:=κ(ΛN)m:=\kappa(\Lambda_N) and μ:=κ/m\mu:=\kappa/m, and let Pμ(x,y):=μ(y−x)P_\mu(x,y):=\mu(y-x). By symmetry of κ\kappa, this kernel is reversible for the uniform law. The corresponding continuous-time generator is

Lκf(x):=∑zκ(z)(f(x+z)−f(x))=m(Pμ−I)f(x).L_\kappa f(x):=\sum_z\kappa(z)\bigl(f(x+z)-f(x)\bigr)=m(P_\mu-I)f(x).

The proposition gives the conductance bound

inf⁡0<∣A∣≤V/2∑zμ(z)bA(z)∣A∣≥cd,ρ2.\inf_{0<|A|\le V/2}\frac{\sum_z\mu(z)b_A(z)}{|A|}\ge\frac{c_{d,\rho}}{2}.

Hence Cheeger’s inequality [9 Theorem 13.10] implies that the spectral gap λκ\lambda_\kappa of −Lκ-L_\kappa satisfies

λκ=mgap⁡(I−Pμ)≥cd,ρ28m.\lambda_\kappa=m\operatorname{gap}(I-P_\mu)\ge\frac{c_{d,\rho}^{2}}{8}m.

Conversely, a spectral-gap estimate of the form λκ≥cd,ρκ(ΛN)\lambda_\kappa\ge c_{d,\rho}\kappa(\Lambda_N) automatically provides an alternative proof of Proposition 4.2. Indeed, writing π\pi for the uniform law, the Dirichlet form associated with LκL_\kappa is

Eκ(f,f):=−⟨f,Lκf⟩π=12V∑x,zκ(z)(f(x+z)−f(x))2.\mathcal{E}_\kappa(f,f):=-\langle f,L_\kappa f\rangle_\pi=\frac{1}{2V}\sum_{x,z}\kappa(z)\bigl(f(x+z)-f(x)\bigr)^2.

Thus

Eκ(1A,1A)=1V∑zκ(z)bA(z),Var⁡π(1A)=∣A∣(V−∣A∣)V2.\mathcal{E}_\kappa(\mathbf{1}_A,\mathbf{1}_A)=\frac{1}{V}\sum_z\kappa(z)b_A(z),\qquad\operatorname{Var}_\pi(\mathbf{1}_A)=\frac{|A|(V-|A|)}{V^2}.

Poincaré’s inequality [9 Lemma 13.7 and Remark 13.8],

λκVar⁡π(f)≤Eκ(f,f),\lambda_\kappa\operatorname{Var}_\pi(f)\le\mathcal{E}_\kappa(f,f),

then implies (11), up to constants.

We now apply Proposition 4.2 to the avalanche. On EN(u)E_N(u), connectedness of ξN(u)\xi_N(u) implies that, for every 1≤j≤L1 \le j \le L, the range contains an edge

(Wj,Zj),Wj∈Sj−1,Zj∈Sj.(W_j,Z_j), \qquad W_j \in S_{j-1}, \quad Z_j \in S_j.

Choose one uniformly among all such edges, independently for each jj conditional on the range. This selection is equivariant under HdH_d, so its conditional law given EN(u)E_N(u) is HdH_d-invariant. Set

κu:=E[1EN(u)∑j=1L(δWj+δZj)].\kappa_u := \mathbb{E}\left[1_{E_N(u)}\sum_{j=1}^{L}(\delta_{W_j}+\delta_{Z_j})\right].

Since

κu(ΛN)=2LP(EN(u)),\kappa_u(\Lambda_N)=2L\mathbb{P}(E_N(u)),

while (8) gives

P(EN(u))≥θ(s),u≥s,\mathbb{P}(E_N(u))\ge\theta(s), \qquad u\ge s,

and, since L≥ρN/2L\ge\rho N/2 for all sufficiently large NN,

κu(ΛN)≥ρθ(s)N.\kappa_u(\Lambda_N)\ge\rho\theta(s)N.

Hence Proposition 4.2 gives cd,s,ρ>0c_{d,s,\rho}>0 such that, uniformly in u≥su\ge s, A⊂ΛNA\subset\Lambda_N, and large NN,

∑x∈A,zκu(z)1{x+z∉A}≥cd,s,ρN∣A∣(V−∣A∣)V.(14)\sum_{x\in A,z}\kappa_u(z)1_{\{x+z\notin A\}}\ge c_{d,s,\rho}N\frac{|A|(V-|A|)}{V}. \tag*{(14)}

By Proposition 2.1, conditionally on the current range AA at level uu, every x∈Ax\in A carries marks at rate one, and a mark rooted at xx grafts an independent uu-avalanche translated by xx.

For such a graft, the event EN(u)E_N(u) and the selected endpoints refer to the independent avalanche. Let DA(x)D_A(x) be the number of distinct sites outside AA among the selected translated endpoints, with DA(x)=0D_A(x)=0 when EN(u)E_N(u) fails. A vertex on SjS_j can occur only as ZjZ_j, and possibly also as Wj+1W_{j+1}; hence every selected endpoint has multiplicity at most two. Therefore

0≤DA(x)≤2L,DA(x)≥121EN(u)∑j=1L(1{x+Wj∉A}+1{x+Zj∉A}).0\le D_A(x)\le2L,\qquad D_A(x)\ge\frac{1}{2}1_{E_N(u)}\sum_{j=1}^{L}\left(1_{\{x+W_j\notin A\}}+1_{\{x+Z_j\notin A\}}\right).

Summing over the possible roots, taking expectations, and using the definition of κu\kappa_u and (14) gives, after decreasing the constant,

∑x∈AE[DA(x)]≥cd,s,ρN∣A∣(V−∣A∣)V.(15)\sum_{x\in A}\mathbb{E}[D_A(x)]\ge c_{d,s,\rho}N\frac{|A|(V-|A|)}{V}. \tag*{(15)}

The expectation is over the independent graft and the auxiliary selection of its crossing edges, with AA fixed. The actual number of new sites created by a graft rooted at xx dominates DA(x)D_A(x).

Proposition 4.5 (Supercritical covering). For every tc<s<tt_c<s<t and every fixed 0<ρ<1/20<\rho<1/2,

P(ξN(t)=ΛN∣EN(s))⟶1.\mathbb{P}(\xi_N(t)=\Lambda_N\mid E_N(s))\longrightarrow1.

Consequently,

lim inf⁡N→∞PN(t)≥θ(s)>0.\liminf_{N\to\infty}P_N(t)\ge\theta(s)>0.

Proof. Condition throughout on EN(s)E_N(s). Connectedness gives rN(s)≥L+1r_N(s) \ge L + 1, and the range can only increase. For a nonempty A⊂ΛNA \subset\Lambda_N, put

Φ(A):=V∣A∣−1.\Phi(A) := \frac{V}{|A|} - 1.

Let a:=∣A∣≥La := |A| \ge L. A graft rooted at x∈Ax \in A adds Δ:=∣(x+ξN′(u))∖A∣\Delta:= |(x + \xi'_N(u)) \setminus A| new sites, where ξN′(u)\xi'_N(u) is an independent uu-avalanche. Since Δ≥DA(x)\Delta\ge D_A(x), DA(x)≤2L≤2aD_A(x) \le2L \le2a, and v↦v/(a+v)v \mapsto v/(a+v) is increasing,

Φ(A)−Φ(A∪(x+ξN′(u)))=VΔa(a+Δ)≥VDA(x)a(a+DA(x))≥VDA(x)3a2.\Phi(A) - \Phi(A \cup(x + \xi'_N(u))) = \frac{V\Delta}{a(a+\Delta)} \ge\frac{VD_A(x)}{a(a+D_A(x))} \ge\frac{VD_A(x)}{3a^2}.

Let Lu\mathcal{L}_u denote the generator of the range process. Each root has rate one, so (15) gives, uniformly for u≥su \ge s and a≥La \ge L,

−LuΦ(A)≥V3a2∑x∈AE[DA(x)]≥cd,s,ρN3V−aa=cd,s,ρN3Φ(A).-\mathcal{L}_u\Phi(A) \ge\frac{V}{3a^2}\sum_{x\in A}\mathbb{E}[D_A(x)] \ge\frac{c_{d,s,\rho}N}{3}\frac{V-a}{a} = \frac{c_{d,s,\rho}N}{3}\Phi(A).

The conditioning event belongs to Hs\mathcal{H}_s, and all subsequent grafts retain their independent laws. Dynkin’s formula and Gronwall’s inequality therefore yield

E[Φ(ξN(t))∣EN(s)]≤VLexp⁡{−cd,s,ρN3(t−s)}.(16)\mathbb{E}[\Phi(\xi_N(t))\mid E_N(s)] \le\frac{V}{L}\exp\left\{-\frac{c_{d,s,\rho}N}{3}(t-s)\right\}. \tag*{(16)}

If A≠ΛNA \ne\Lambda_N, then Φ(A)≥1/(V−1)\Phi(A) \ge1/(V-1). Consequently,

P(ξN(t)≠ΛN∣EN(s))≤V(V−1)Lexp⁡{−cd,s,ρN3(t−s)}→0.(17)\mathbb{P}(\xi_N(t) \ne\Lambda_N\mid E_N(s)) \le\frac{V(V-1)}{L}\exp\left\{-\frac{c_{d,s,\rho}N}{3}(t-s)\right\} \to0. \tag*{(17)}

Finally,

PN(t)≥P(EN(s))P(ξN(t)=ΛN∣EN(s)),P_N(t) \ge\mathbb{P}(E_N(s))\mathbb{P}(\xi_N(t)=\Lambda_N\mid E_N(s)),

and (8) gives lim inf⁡N→∞PN(t)≥θ(s)>0\liminf_{N\to\infty}P_N(t)\ge\theta(s)>0.

Proof of Theorem 1.5. This is Proposition 4.5.

Remark 4.6. The estimate (17) controls covering throughout the growth from a macroscopic arm to the whole torus. Since V=NdV=N^d and L≥ρN/2L\ge\rho N/2 for all sufficiently large NN, its prefactor satisfies

V(V−1)L≤2ρN2d−1.\frac{V(V-1)}{L} \le\frac{2}{\rho}N^{2d-1}.

Thus the comparison also supplies a covering bound after a sprinkling C(log⁡N)/NC(\log N)/N, with CC sufficiently large depending on dd, ss, ρ\rho.

Subcritical locality

We prove Theorem 1.6. Recall that

R(t)=Er(t),tc=tsus,R(t)=\mathbb{E}r(t),\qquad t_c=t_{\mathrm{sus}},

so that R(t)<∞R(t)<\infty for every t<tct<t_c. The hierarchical decomposition first upgrades this finite susceptibility to exponential integrability of the avalanche range. Local injectivity then compares infinite- and finite-volume ranges, while Lemma 2.2 transfers the resulting convergence of mean ranges to mean durations. The following proof is inspired by [16].

Proposition 5.1 (Subcritical exponential moments). For every t<tct<t_c, there exists λt>0\lambda_t>0 such that

E[eλtr(t)]<∞.\mathbb{E}[e^{\lambda_t r(t)}] < \infty.

In particular, E[r(t)p]<∞\mathbb{E}[r(t)^p]<\infty for every integer p≥1p\ge1.

Proof. Fix L≥1L \ge1 and, for λ≥0\lambda\ge0, set

FL(u,λ):=E[eλ(r(u)∧L)].F_L(u,\lambda):=\mathbb{E}\left[e^{\lambda(r(u)\wedge L)}\right].

By Proposition 2.1, conditionally on a current range AA of size xx, marks occur at total rate xx, and each mark grafts an independent avalanche BB with ∣B∣=dr(u)|B| \overset{\mathrm{d}}{=} r(u). Since

∣A∪(z+B)∣≤x+∣B∣,|A\cup(z+B)|\le x+|B|,

if x<Lx<L, then

eλ((x+∣B∣)∧L)−eλx≤eλx(eλ(∣B∣∧L)−1),e^{\lambda((x+|B|)\wedge L)}-e^{\lambda x}\le e^{\lambda x}\left(e^{\lambda(|B|\wedge L)}-1\right),

whereas the truncated observable is constant when x≥Lx\ge L. Conditioning on the current range and using the independence of the graft therefore gives, for almost every u<tcu<t_c,

∂uFL(u,λ)≤(FL(u,λ)−1)∂λFL(u,λ).(18)\partial_uF_L(u,\lambda)\le(F_L(u,\lambda)-1)\partial_\lambda F_L(u,\lambda). \tag*{(18)}

Indeed, the first factor produced by the rate is bounded by

E[r(u)eλr(u)1{r(u)<L}]≤∂λFL(u,λ).\mathbb{E}\left[r(u)e^{\lambda r(u)}\mathbf{1}_{\{r(u)<L\}}\right]\le\partial_\lambda F_L(u,\lambda).

We claim that exponential integrability propagates over every interval [s,s+h][s,s+h] such that hR(s)<1hR(s)<1. Suppose E[eα0r(s)]<∞\mathbb{E}[e^{\alpha_0r(s)}]<\infty for some α0>0\alpha_0>0. Since

E[eαr(s)]−1α⟶R(s)(α↓0),\frac{\mathbb{E}[e^{\alpha r(s)}]-1}{\alpha}\longrightarrow R(s)\qquad(\alpha\downarrow0),

we may choose 0<α<α00<\alpha<\alpha_0 such that, writing q:=E[eαr(s)]q:=\mathbb{E}[e^{\alpha r(s)}],

h(q−1)<α.h(q-1)<\alpha.

For s≤u≤s+hs\le u\le s+h, set

λ(u):=α−(q−1)(u−s),GL(u):=FL(u,λ(u)).\lambda(u):=\alpha-(q-1)(u-s),\qquad G_L(u):=F_L(u,\lambda(u)).

Then λ(u)>0\lambda(u)>0, and (18) yields

GL′(u)≤∂λFL(u,λ(u))(GL(u)−q),GL(s)≤q.G_L'(u)\le\partial_\lambda F_L(u,\lambda(u))(G_L(u)-q),\qquad G_L(s)\le q.

Hence GL(u)≤qG_L(u)\le q on [s,s+h][s,s+h]. In particular,

FL(s+h,α−h(q−1))≤q,α−h(q−1)>0.F_L(s+h,\alpha-h(q-1))\le q,\qquad\alpha-h(q-1)>0.

Letting L→∞L\to\infty gives

E[e(α−h(q−1))r(s+h)]≤q<∞.\mathbb{E}\left[e^{(\alpha-h(q-1))r(s+h)}\right]\le q<\infty.

Now fix t<tct<t_c. Since R(t)<∞R(t)<\infty, choose a partition 0=s0<⋯<sm=t0=s_0<\cdots<s_m=t with mesh smaller than R(t)−1R(t)^{-1}. The initial range r(0)=∣Nd∣=2d+1r(0)=|\mathcal{N}_d|=2d+1 is deterministic. If r(sj)r(s_j) has an exponential moment, then, by monotonicity of RR,

(sj+1−sj)R(sj)≤(sj+1−sj)R(t)<1,(s_{j+1}-s_j)R(s_j)\le(s_{j+1}-s_j)R(t)<1,

so the preceding propagation argument gives an exponential moment for r(sj+1)r(s_{j+1}). Iterating finitely many times yields

E[eλtr(t)]<∞\mathbb{E}[e^{\lambda_t r(t)}]<\infty

for some λt>0\lambda_t>0. □\square

The preceding argument is independent of the geometry of Zd\mathbb{Z}^d. Geometry enters only when these range moments are used to compare the infinite lattice with the torus.

Lemma 5.2 (Local injectivity). Let

ϖN:Zd→ΛN\varpi_N:\mathbb{Z}^d\to\Lambda_N

be the canonical projection. If A⊂ZdA\subset\mathbb{Z}^d is connected and

∣A∣<N,|A|<N,

then ϖN\varpi_N is injective on AA.

Consequently, for every KK with

K+2d<N,K+2d<N,

the infinite- and finite-volume avalanche constructions can be coupled in physical update time, using the same refresh variables, so as to agree, including their durations, until their common range first contains at least KK sites.

Proof. A connected set with kk vertices has graph diameter at most k−1k-1. Thus, if ∣A∣<N|A|<N,

diam⁡ℓ1(A)≤∣A∣−1<N.\operatorname{diam}_{\ell^1}(A)\leq|A|-1<N.

On the other hand, distinct x,y∈Zdx,y\in\mathbb{Z}^d satisfying

ϖN(x)=ϖN(y)\varpi_N(x)=\varpi_N(y)

have

∥x−y∥1≥N.\lVert x-y\rVert_1\geq N.

Hence ϖN\varpi_N is injective on AA.

For the second assertion, couple the elementary Bak–Sneppen updates in Zd\mathbb{Z}^d and on the torus using the same refresh variables, as long as the explored range projects injectively. Avalanche ranges are connected, and one elementary Bak–Sneppen update can add at most 2d2d new sites to the range. Hence, while the current range contains fewer than KK sites, after the next update it contains fewer than K+2d<NK+2d<N sites and still projects injectively. The two explorations therefore remain identical until the claimed stopping time. □

Proposition 5.3 (Subcritical finite-volume approximation). For every t<tct<t_c,

sup⁡0≤u≤t∣RN(u)−R(u)∣⟶0.(19)\sup_{0\leq u\leq t}|R_N(u)-R(u)|\longrightarrow0. \tag*{(19)}

Moreover,

DN(t)⟶D(t).(20)D_N(t)\longrightarrow D(t). \tag*{(20)}

Finally, for every fixed KK, for all sufficiently large NN and every 1≤k≤K1\leq k\leq K,

E[ηN(t)1{rN(t)=k}]=E[η(t)1{r(t)=k}].(21)\mathbb{E}\left[\eta_N(t)1_{\{r_N(t)=k\}}\right]=\mathbb{E}\left[\eta(t)1_{\{r(t)=k\}}\right]. \tag*{(21)}

Proof. Take

KN:=⌊N/3⌋.K_N:=\lfloor N/3\rfloor.

For large NN,

KN+2d<N,K_N+2d<N,

so Lemma 5.2 couples the two avalanches identically until their common range reaches KNK_N. In particular, on {r(u)<KN}\{r(u)<K_N\},

rN(u)=r(u).r_N(u)=r(u).

Since

rN(u)≤Nd,r_N(u)\leq N^d,

we obtain

∣RN(u)−R(u)∣≤NdP(r(u)≥KN)+E[r(u)1{r(u)≥KN}].|R_N(u)-R(u)|\leq N^d\mathbb{P}(r(u)\geq K_N)+\mathbb{E}[r(u)1_{\{r(u)\geq K_N\}}].

Fix an integer p>dp>d. By monotonicity in the level parameter and Proposition 5.1,

P(r(u)≥KN)≤E[r(t)p]KNp,\mathbb{P}(r(u)\ge K_N)\le\frac{\mathbb{E}[r(t)^p]}{K_N^p},

and

E[r(u)1{r(u)≥KN}]≤E[r(t)p]KNp−1,\mathbb{E}[r(u)\mathbf{1}_{\{r(u)\ge K_N\}}]\le\frac{\mathbb{E}[r(t)^p]}{K_N^{p-1}},

uniformly for u≤tu\le t. Hence

sup⁡u≤t∣RN(u)−R(u)∣≤Cp,t(Nd−p+N1−p)⟶0,\sup_{u\le t}|R_N(u)-R(u)|\le C_{p,t}\left(N^{d-p}+N^{1-p}\right)\longrightarrow0,

which proves (19).

By Lemma 2.2,

DN(t)=exp⁡{∫0tRN(u) du},D(t)=exp⁡{∫0tR(u) du}.D_N(t)=\exp\left\{\int_0^t R_N(u)\,du\right\},\qquad D(t)=\exp\left\{\int_0^t R(u)\,du\right\}.

Therefore (19) implies (20).

Finally fix KK, and take NN large enough that

K+1+2d<N.K+1+2d<N.

Apply Lemma 5.2 with K+1K+1. The two avalanches then agree until their common range first contains at least K+1K+1 sites. On the event that the final range has size k≤Kk\le K, this stopping level is never reached. Hence the entire finite-volume avalanche is identical to the infinite-volume avalanche, including its duration, and conversely. This proves (21). □

Proof of Theorem 1.6. The exponential-moment statement is Proposition 5.1, and the convergence of the mean ranges is (19). □

From avalanches to the stationary state

We now complete the passage from avalanche criticality to the stationary state. The argument is the finite-volume mechanism introduced by Meester and Znamenski [11, 12], but we give it in full because it is dimension-free and is an essential part of the present theorem. Write

V:=∣ΛN∣=Nd,νh:=L(h+E),E∼Exp⁡(1).V:=|\Lambda_N|=N^d,\qquad\nu_h:=\mathcal{L}(h+E),\qquad E\sim\operatorname{Exp}(1).

All neighbourhoods on the torus are understood with repetitions removed.

Proposition 6.1 (Finite-volume ergodicity). For every NN, the Bak–Sneppen chain on ΛN\Lambda_N has a unique stationary law πN\pi_N. From every deterministic initial configuration, its law converges to πN\pi_N in total variation. Moreover πN\pi_N is translation invariant.

Proof. If one update refreshes the whole torus, then after one step the configuration consists of independent Exp⁡(1)\operatorname{Exp}(1) variables, independently of the initial state, and the conclusion is immediate. Assume henceforth that one update does not cover ΛN\Lambda_N.

Fix h>0h>0, write PP for the transition kernel and V=NdV=N^d, and set

Ch:={f:fx>h for every x∈ΛN}.C_h:=\{f:f_x>h\text{ for every }x\in\Lambda_N\}.

With probability at least qh:=e−(2d+1)hV>0q_h:=e^{-(2d+1)hV}>0, every fitness sampled in the next VV updates exceeds hh, uniformly in the current state. On this event no value at most hh is created. As long as such values remain, the selected minimum is at most hh and is refreshed, so their number decreases at each update. After at most VV updates none remains. Hence

PV(f,Ch)≥qh,f∈[0,∞)ΛN.(22)P^{V}(f,C_h) \ge q_h,\qquad f \in[0,\infty)^{\Lambda_N}. \tag*{(22)}

Iteration gives a geometric tail for the hitting time of ChC_h, uniformly in the initial state.

For the minorization on ChC_h, choose a finite nearest-neighbour walk starting at the first updated site and visiting every vertex. At each intermediate update, require that the prescribed next vertex receive a fitness at most hh, while every other refreshed vertex receives a fitness above hh; at the final update require all fresh values to exceed hh. At each intermediate step the prescribed next vertex is then the unique site with fitness at most hh, so the avalanche follows the walk, covers the torus, and terminates. These finitely many requirements have positive probability, so PN(h)>0P_N(h)>0. Since the avalanche duration is integer-valued, there exists m≥1m\ge1 such that

ph:=P(rN(h)=V, ηN(h)=m)>0.p_h := \mathbb{P}\bigl(r_N(h)=V,\ \eta_N(h)=m\bigr)>0.

This range–duration event is measurable with respect to the censored level-hh exploration, whose law, up to translation of its starting point, does not depend on f∈Chf\in C_h. On this event every site has been refreshed, so Proposition 2.1 implies that the terminal fitnesses are independent with common law νh\nu_h, independently of the censored exploration. Thus, setting

μN,h:=νh⊗ΛN,\mu_{N,h}:=\nu_h^{\otimes\Lambda_N},

we obtain

Pm(f,⋅)≥phμN,h(⋅),f∈Ch.(23)P^m(f,\cdot)\ge p_h\mu_{N,h}(\cdot),\qquad f\in C_h. \tag*{(23)}

Combining (22) and (23) by the Markov property gives

PV+m(f,⋅)≥qhphμN,h(⋅)for every f.(24)P^{V+m}(f,\cdot)\ge q_hp_h\mu_{N,h}(\cdot)\qquad\text{for every }f. \tag*{(24)}

Hence λ↦λPV+m\lambda\mapsto\lambda P^{V+m} is a contraction in total variation with factor 1−qhph1-q_hp_h. The space of probability measures is complete for this metric, so Banach’s fixed-point theorem gives a unique fixed point πN\pi_N. Since πNP\pi_NP is also PV+mP^{V+m}-invariant, uniqueness gives πNP=πN\pi_NP=\pi_N, so πN\pi_N is the unique invariant law of PP. Iterating the contraction and using that Markov kernels are nonexpansive in total variation yields

∥λPn−πN∥TV≤(1−qhph)⌊n/(V+m)⌋∥λ−πN∥TV⟶0\lVert\lambda P^n-\pi_N\rVert_{\mathrm{TV}} \le (1-q_hp_h)^{\left\lfloor n/(V+m)\right\rfloor} \lVert\lambda-\pi_N\rVert_{\mathrm{TV}} \longrightarrow0

for every initial law λ\lambda.

Finally, start the chain from independent Exp⁡(1)\operatorname{Exp}(1) fitnesses. Its law is translation invariant, ties have probability zero at every time, and the Bak–Sneppen dynamics preserves translation invariance. The total-variation convergence just proved therefore implies that its limit πN\pi_N is translation invariant. □

We next introduce the locking variables. Start the finite-volume chain from independent Exp⁡(1)\operatorname{Exp}(1) fitnesses and set YN,x(0)=0Y_{N,x}(0)=0. If iN(n)i_N(n) is the site of minimum fitness at time nn and

MN(n):=FN,iN(n)(n),M_N(n):=F_{N,i_N(n)}(n),

define

YN,x(n+1)={0,x∈iN(n)+Nd,max⁡{YN,x(n),MN(n)},x∉iN(n)+Nd.(25)Y_{N,x}(n+1)= \begin{cases} 0, & x\in i_N(n)+\mathcal{N}_d,\\ \max\{Y_{N,x}(n),M_N(n)\}, & x\notin i_N(n)+\mathcal{N}_d. \end{cases} \tag*{(25)}

For y,z≥0y,z\ge0, set

Hy(z):={0,z<y,1−e−(z−y),z≥y.H_y(z):= \begin{cases} 0, & z<y,\\ 1-e^{-(z-y)}, & z\ge y. \end{cases}

Proposition 6.2 (Conditional exponential representation). Conditionally on YN(n)=(YN,x(n))x∈ΛNY_N(n)=(Y_{N,x}(n))_{x\in\Lambda_N}, the fitnesses (FN,x(n))x∈ΛN(F_{N,x}(n))_{x\in\Lambda_N} are independent and

FN,x(n)∼νYN,x(n).F_{N,x}(n)\sim\nu_{Y_{N,x}(n)}.

Consequently, for distinct x1,…,xk∈ΛNx_1,\ldots,x_k\in\Lambda_N and z1,…,zk≥0z_1,\ldots,z_k\ge0,

P(FN,xj(n)≤zj, 1≤j≤k)=E[∏j=1kHYN,xj(n)(zj)].(26)\mathbb{P}\left(F_{N,x_j}(n)\le z_j,\ 1\le j\le k\right)=\mathbb{E}\left[\prod_{j=1}^{k}H_{Y_{N,x_j}(n)}(z_j)\right]. \tag*{(26)}

Proof. The assertion holds at time zero. Assume it holds at time nn, and condition on YN(n)Y_N(n), the location iN(n)i_N(n), and the value MN(n)M_N(n) of the minimum. By the lack-of-memory property, the unrefreshed coordinates remain independent and can be written as

max⁡{YN,x(n),MN(n)}+EN,x′,\max\{Y_{N,x}(n),M_N(n)\}+E'_{N,x},

where the EN,x′E'_{N,x} are independent Exp⁡(1)\operatorname{Exp}(1) variables. The refreshed coordinates are fresh independent exponentials. Thus, conditionally on these variables, the fitnesses at time n+1n+1 are independent exponentials above the thresholds in (25). Since this conditional law depends on the conditioning variables only through YN(n+1)Y_N(n+1), the same statement holds conditionally on YN(n+1)Y_N(n+1) alone. This proves the claim by induction. □\square

Put

GN(n,t):=P(YN,0(n)≤t).G_N(n,t):=\mathbb{P}(Y_{N,0}(n)\le t).

Translation invariance of the initial law and of the dynamics makes the choice of the origin immaterial.

Proposition 6.3 (Subcritical locking thresholds). For every t<tct<t_c,

lim⁡N→∞lim sup⁡n→∞GN(n,t)=0.(27)\lim_{N\to\infty}\limsup_{n\to\infty}G_N(n,t)=0. \tag*{(27)}

Proof. Let

KN(n,t):=∑x∈ΛN1{YN,x(n)≤t}.K_N(n,t):=\sum_{x\in\Lambda_N}\mathbf{1}_{\{Y_{N,x}(n)\le t\}}.

Let CtC_t be the set of configurations with all fitnesses above tt. By (22), its hitting time has finite mean. Discard the initial segment up to the first visit to CtC_t, and from then on decompose the trajectory into successive tt-avalanches. If TjT_j is the starting time of the jj-th one and (rj,ηj)(r_j,\eta_j) its range–duration pair, then Tj+1−Tj=ηjT_{j+1}-T_j=\eta_j, and these pairs are i.i.d. with the law of (rN(t),ηN(t))(r_N(t),\eta_N(t)). At the first update of the avalanche the current minimum exceeds tt, so every site outside its eventual range has its locking threshold raised above tt. During the remaining updates the selected minima are at most tt, and only sites in the avalanche range are refreshed. Hence

KN(n,t)≤rj,Tj<n≤Tj+1.K_N(n,t)\le r_j,\qquad T_j<n\le T_{j+1}.

For all sufficiently large NN, a one-step noncovering tt-avalanche has positive probability, so the cycle length has span one. The delayed renewal–reward theorem and translation invariance therefore give

lim sup⁡n→∞GN(n,t)≤E[ηN(t)rN(t)]VDN(t).\limsup_{n\to\infty}G_N(n,t)\le\frac{\mathbb{E}[\eta_N(t)r_N(t)]}{V D_N(t)}.

For every fixed K≥1K\ge1, using rN(t)≤Vr_N(t)\le V,

lim sup⁡n→∞GN(n,t)≤KV+E[ηN(t)1{rN(t)>K}]DN(t).(28)\limsup_{n\to\infty}G_N(n,t)\le\frac{K}{V}+\frac{\mathbb{E}[\eta_N(t)\mathbf{1}_{\{r_N(t)>K\}}]}{D_N(t)}. \tag*{(28)}

Now

E[ηN(t)1{rN(t)>K}]=DN(t)−∑k=1KE[ηN(t)1{rN(t)=k}].\mathbb{E}[\eta_N(t)\mathbf{1}_{\{r_N(t)>K\}}]=D_N(t)-\sum_{k=1}^{K}\mathbb{E}[\eta_N(t)\mathbf{1}_{\{r_N(t)=k\}}].

By Proposition 5.3, as N→∞N \to\infty this converges to E[η(t)1{r(t)>K}]\mathbb{E}[\eta(t)\mathbf{1}_{\{r(t)>K\}}], while DN(t)→D(t)<∞D_N(t) \to D(t)<\infty. Letting first N→∞N \to\infty and then K→∞K \to\infty in (28) proves (27). □

Proposition 6.4 (Supercritical locking thresholds). For every b>tcb>t_c,

lim⁡N→∞lim inf⁡n→∞GN(n,b)=1.(29)\lim_{N\to\infty}\liminf_{n\to\infty}G_N(n,b)=1. \tag*{(29)}

Proof. Fix tc<t<bt_c<t<b. Starting from the process defining GNG_N, ignore the initial segment up to the first visit to CtC_t. Thereafter run successive tt-avalanches until the first one covering ΛN\Lambda_N, and then continue until the next visit to CbC_b.

The initial delay and the final recovery have finite mean by (22). Moreover PN(t)>0P_N(t)>0 by Proposition 6.1, and DN(t)<∞D_N(t)<\infty by Lemma 2.2, so the waiting time for the first covering tt-avalanche also has finite mean.

At the end of this covering avalanche, every site has been refreshed. Conditionally on its tt-censored exploration, the terminal fitnesses are independent with law νt\nu_t, by Proposition 2.1. Censor the subsequent recovery at level bb: reveal fitnesses at most bb exactly and reveal larger fitnesses only through their exceedance of bb. At the designated return to CbC_b, each terminal value has been observed only to exceed bb. Thus, conditionally on the censored cycle history, the terminal vector has law

μN,b=νb⊗ΛN.\mu_{N,b}=\nu_b^{\otimes\Lambda_N}.

Treat the trajectory up to the first such return as the initial delay, and repeat the construction from that point. Fresh future refresh variables and this censored regeneration make the successive pairs of waiting and recovery lengths independent and identically distributed. Let ANA_N be the time spent in successive tt-avalanches up to and including the first covering one in a cycle started with law μN,b\mu_{N,b}, and let BNB_N be the subsequent recovery time to CbC_b. For large NN, the cycle length has span one, since a one-step noncovering tt-avalanche may be inserted before the first covering one with positive probability.

At the end of the covering avalanche all locking thresholds are at most tt, and throughout the recovery every selected minimum is at most bb. Thus

KN(n,b)=VK_N(n,b)=V

during the whole recovery phase. By the delayed renewal–reward theorem,

lim inf⁡n→∞GN(n,b)≥EBNEAN+EBN.(30)\liminf_{n\to\infty}G_N(n,b)\ge\frac{\mathbb{E}B_N}{\mathbb{E}A_N+\mathbb{E}B_N}. \tag*{(30)}

Since the successive tt-avalanches are i.i.d. until the first covering one, Wald’s identity gives

EAN=DN(t)PN(t).(31)\mathbb{E}A_N=\frac{D_N(t)}{P_N(t)}. \tag*{(31)}

To bound BNB_N from below, use the decorated construction of Proposition 2.1. Since the covering tt-avalanche has range ΛN\Lambda_N, every site is active in the level extension from tt to bb. For each x∈ΛNx\in\Lambda_N, let τx\tau_x be the first mark above tt on the vertical line through xx. Then τx−t\tau_x-t are i.i.d. Exp⁡(1)\operatorname{Exp}(1), and, conditionally on τx=u≤b\tau_x=u\le b, the corresponding graft contributes a duration of mean DN(u)D_N(u). Hence, by Tonelli and monotonicity of DND_N,

EBN≥V∫tbe−(u−t)DN(u) du≥V(1−e−(b−t))DN(t).(32)\mathbb{E}B_N \ge V\int_t^b e^{-(u-t)}D_N(u)\,\mathrm{d}u \ge V(1-e^{-(b-t)})D_N(t). \tag*{(32)}

Combining (30)–(32),

1−lim inf⁡n→∞GN(n,b)≤1V(1−e−(b−t))PN(t).(33)1-\liminf_{n\to\infty}G_N(n,b)\le\frac{1}{V(1-e^{-(b-t)})P_N(t)}. \tag*{(33)}

Choose s∈(tc,t)s\in(t_c,t). By Theorem 1.5,

lim inf⁡N→∞PN(t)≥θ(s)>0.\liminf_{N\to\infty}P_N(t)\ge\theta(s)>0.

Since V=Nd→∞V=N^d\to\infty, the right-hand side of (33) tends to zero, proving (29).

Proposition 6.5 (Locking-threshold concentration and product limit). For every ε>0\varepsilon>0,

lim⁡N→∞lim sup⁡n→∞P(∣YN,0(n)−tc∣>ε)=0.(34)\lim_{N\to\infty}\limsup_{n\to\infty}\mathbb{P}(|Y_{N,0}(n)-t_c|>\varepsilon)=0. \tag*{(34)}

Consequently, for every fixed collection of distinct sites x1,…,xk∈Zdx_1,\ldots,x_k\in\mathbb{Z}^d, the corresponding marginal of πN\pi_N converges to νtc⊗k\nu_{t_c}^{\otimes k}.

Proof. For 0<ε<tc0<\varepsilon<t_c, Propositions 6.3 and 6.4, applied at tc−εt_c-\varepsilon and tc+εt_c+\varepsilon, give (34); for ε≥tc\varepsilon\ge t_c, the lower-tail event is empty and the same conclusion follows from the supercritical bound. For fixed x1,…,xkx_1,\ldots,x_k, and NN large enough that their images in ΛN\Lambda_N are distinct, translation invariance and the union bound imply

lim sup⁡N→∞lim sup⁡n→∞P(max⁡1≤j≤k∣YN,xj(n)−tc∣>ε)=0.\limsup_{N\to\infty}\limsup_{n\to\infty}\mathbb{P}\left(\max_{1\le j\le k}|Y_{N,x_j}(n)-t_c|>\varepsilon\right)=0.

Fix z1,…,zk≥0z_1,\ldots,z_k\ge0 and define

Ψ(y1,…,yk):=∏j=1kHyj(zj).\Psi(y_1,\ldots,y_k):=\prod_{j=1}^k H_{y_j}(z_j).

The function Ψ\Psi is bounded and continuous. If

ωε:=sup⁡max⁡j∣yj−tc∣≤ε∣Ψ(y1,…,yk)−Ψ(tc,…,tc)∣,\omega_\varepsilon:=\sup_{\max_j|y_j-t_c|\le\varepsilon}|\Psi(y_1,\ldots,y_k)-\Psi(t_c,\ldots,t_c)|,

then ωε→0\omega_\varepsilon\to0 as ε↓0\varepsilon\downarrow0. By Proposition 6.2,

∣P(FN,xj(n)≤zj, 1≤j≤k)−∏j=1kHtc(zj)∣≤ωε+P(max⁡1≤j≤k∣YN,xj(n)−tc∣>ε).\left|\mathbb{P}(F_{N,x_j}(n)\le z_j,\ 1\le j\le k)-\prod_{j=1}^k H_{t_c}(z_j)\right| \le\omega_\varepsilon+\mathbb{P}\left(\max_{1\le j\le k}|Y_{N,x_j}(n)-t_c|>\varepsilon\right).

For fixed NN, Proposition 6.1 lets n→∞n\to\infty and replaces the law of the chain by πN\pi_N. Letting next N→∞N\to\infty and finally ε↓0\varepsilon\downarrow0 proves convergence of the joint distribution functions to ∏jHtc(zj)\prod_j H_{t_c}(z_j), the distribution function of νtc⊗k\nu_{t_c}^{\otimes k}. □

Proof of Theorem 1.1. Proposition 6.5 gives the asserted convergence in exponential coordinates. Under the inverse transformation u=1−e−fu=1-e^{-f}, a random variable with law tc+Exp⁡(1)t_c+\operatorname{Exp}(1) becomes uniform on [1−e−tc,1][1-e^{-t_c},1]. This proves the theorem. □

Use of AI

The (compensated) Laplace transform of the nonlinear additive process in which XtX_t is replaced by Xt+Xt′X_t + X'_t, with Xt′X'_t an independent copy of XtX_t, at rate XtX_t, satisfies an inviscid Burgers equation. Starting from X0=1X_0 = 1, its solution is given by the Borel(tt) distribution, the law of the total progeny of a Poisson(tt) Galton–Watson tree. The phase transition at t=1t = 1 can then be identified by the method of characteristics; see [13, 4].

This mechanism turns out to be closely related to the branching-process description of Bak–Sneppen avalanches in the mean-field setting [3], and for two years, we had been looking for a way to relate it to the evolution of the avalanche range in our spatial setting. During that period, we sought a suitable comparison or domination by solutions of the Burgers equation, both through our own efforts and by prompting earlier AI models. These attempts were unsuccessful and, in particular, repeatedly led to incorrect boundary-layer arguments by AI.

In August 2026, ChatGPT 5.6 Sol produced the key argument used in this paper: re-strict attention to grafts involving large avalanches, derive a first-moment lower bound on the number of newly created vertices, and combine this estimate with concavity to obtain the Burgers-type comparison (Lemma 3.2 and Proposition 3.4).

The amplification argument was subsequently obtained by the same model through a spectral-gap approach, along the lines of Remark 4.4. We then suggested looking for a purely combinatorial proof, leading to the argument retained in the paper.

Finally, finiteness of all polynomial moments in the subcritical regime was first established by the model through a recursive argument. We then suggested seeking a recursive proof of exponential moments inspired by [16]; the resulting argument was developed with AI assistance.

Claude did the simulation leading to the picture.

References

Institut de mathématique d’Orsay, Université Paris-Saclay, 91405 Orsay, France

Email address: [email protected]

References

  1. [1]David J. Aldous, Deterministic and stochastic models for coalescence (aggregation and coagulation): a review of the mean-field theory for probabilists, Bernoulli 5 (1999), no. 1, 3–48.DOI
  2. [2]Per Bak and Kim Sneppen, Punctuated equilibrium and criticality in a simple model of evolution, Physical Review Letters 71 (1993), 4083–4086.pubmed.ncbi.nlm.nih.gov/10055149
  3. [3]Jan de Boer, Bernard Derrida, Henrik Flyvbjerg, Andrew D. Jackson, and Tilo Wettig, Simple model of self-organized biological evolution, Physical Review Letters 73 (1994), no. 6, 906–909.pubmed.ncbi.nlm.nih.gov/10057569
  4. [4]Madalina Deaconu and Etienne Tanré, Smoluchowski’s coagulation equation: Probabilistic interpretation of solutions for constant, additive and multiplicative kernels, Annali della Scuola Normale Superiore di Pisa, Classe di Scienze 29 (2000), no. 3, 549–579.
  5. [5]Persi Diaconis and Laurent Saloff-Coste, Comparison techniques for random walk on finite groups, Annals of Probability 21 (1993), no. 4, 2131–2156.DOI
  6. [6]Andrew Gillett, Ronald Meester, and Misja Nuyens, Bounds for avalanche critical values of the Bak–Sneppen model, Markov Processes and Related Fields 12 (2006), no. 4, 679–694.DOI
  7. [7]Andrew J. Gillett, Ronald W. J. Meester, and Peter van der Wal, Maximal avalanches in the Bak–Sneppen model, Journal of Applied Probability 43 (2006), no. 3, 840–851.DOI
  8. [8]Olivier Hénard, Sharpness of the avalanche phase transition in the Bak–Sneppen model, arXiv preprint, arXiv:2608.10992 (2026).
  9. [9]David A. Levin and Yuval Peres, Markov chains and mixing times, 2 ed., American Mathematical Society, Providence, RI, 2017.
  10. [10]Sergei Maslov, Infinite series of exact equations in the Bak–Sneppen model of biological evolution, Physical Review Letters 77 (1996), 1182–1185.pubmed.ncbi.nlm.nih.gov/10063011
  11. [11]Ronald Meester and Dmitri Znamenski, Limit behavior of the Bak–Sneppen evolution model, Annals of Probability 31 (2003), 1986–2002.DOI
  12. [12]———, Critical thresholds and the limit distribution in the Bak–Sneppen model, Communications in Mathematical Physics 246 (2004), 63–86.DOI
  13. [13]Govind Menon and Robert L. Pego, Approach to self-similarity in Smoluchowski’s coagulation equations, Communications on Pure and Applied Mathematics 57 (2004), no. 9, 1197–1232.DOI
  14. [14]Maya Paczuski, Sergei Maslov, and Per Bak, Avalanche dynamics in evolution, growth, and de-pinning models, Physical Review E 53 (1996), 414–443.
  15. [15]Serguei Popov and Stanislav Volkov, On extinction and survival of the Bak–Sneppen model on arbitrary graphs, arXiv preprint arXiv:2506.15500 (2026).arxiv.org/abs/2506.15500
  16. [16]Oliver Riordan and Lutz Warnke, The evolution of subcritical achlioptas processes, Random Structures & Algorithms 47 (2015), no. 1, 174–203.DOI
  17. [17]Stanislav Volkov, Rigorous upper bound for the discrete Bak–Sneppen model, Journal of Statistical Physics 186 (2022), 8.arxiv.org/abs/2003.00222

Paper details

vlnc:W_nbn3fh4d
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

GPT-5.6 Sol (bare), Claude

View statement
Looking for related papers…
Contents