WorksMathematics
Avalanche Sharpness and the Stationary State of the Bak–Sneppen Model on
Abstract
We establish the stationary picture for the nearest-neighbour Bak–Sneppen model on , for every . There exists a critical level such that, as , every fixed family of stationary fitnesses on the torus converges, in exponential coordinates, to independent random variables with law . The proof relies on three estimates for avalanches. For , 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 and following the updates until all fitnesses are again above 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 , 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.

Figure 1. A near-critical avalanche on at level (). Sites are coloured by their entry level 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
and use exponential coordinates for the fitnesses. Thus every refreshed fitness has law , and an update at refreshes the sites in
Let be the stationary law of this finite-volume Markov chain, and, for , let denote the stationary fitness at the image of in . Our main result identifies the large-volume stationary state.
Theorem 1.1 (Stationary limit). There exists such that, for every fixed collection of distinct sites ,
where are independent random variables.
For , let denote the law of , where . Under the increasing change of variables
Theorem 1.1 is equivalent to the usual uniform representation: the limiting fitnesses are independent and uniform on
The critical level is characterized through infinite-volume avalanches. Fix , start from a configuration in which all fitnesses are larger than , and force an update at the origin. Thereafter, as long as some current fitness is at most , update the site carrying the smallest fitness and its neighbours. The procedure stops when all current fitnesses are again larger than . The set of sites refreshed during this procedure (including the forced initial update) is the range of the -avalanche.
Let denote this range on , let be its duration, including the forced update and with value if the avalanche does not terminate, and put
We use throughout the natural monotone coupling in the level parameter, under which for . 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
Every update refreshes at most sites, hence , while positive mass at infinity forces infinite mean range. Therefore
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 , ,
Moreover, for every ,
Remark 1.3. The equality
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 , together with the quantitative bound (1).
Remark 1.4. Finiteness of follows directly from (1), taking , since . Positivity of is elementary: in the exploration of a -avalanche, each update produces fresh fitnesses, each of which is at most with probability . Ignoring overlaps therefore gives a domination by a Galton–Watson process with offspring distribution
This process is subcritical whenever , which implies finite mean avalanche duration for all sufficiently small , and hence .
For , Theorem 1.2 already suffices to complete the Meester–Znamenski program [11], recalled in § 6, and derive the stationary-limit result. In dimensions , two additional finite-volume ingredients are required.
The first ingredient concerns avalanches on the finite torus. Let denote the range of the -avalanche on , and set
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 and .
Let
Then
Consequently,
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 , there exists such that
In particular, all polynomial moments of are finite. Moreover,
Proof strategy
The proof for general 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 , each site produces marks at rate one, each grafting an independent avalanche translated by . The relevant geometric quantity is the part of this translated range lying outside .
The deterministic input is
which follows from a translation-invariant total order on . Applied to
the hierarchical decomposition gives the Burgers-type inequality
Since , while whenever 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 need not resemble a ball or be well mixed on the torus. We retain only one nearest-neighbour edge crossing each successive radius.
For , let
be its translation boundary. It satisfies
Fix , and let be the event that an independent avalanche reaches distance . On , select uniformly one crossing edge at each radius . Let be their expected endpoint counting measure, with zero contribution off . The selected edges need not form a connected path.
At a fixed supercritical level, is bounded away from zero uniformly in , so is of order . Hyperoctahedral symmetry, subadditivity of , and the crossing edges then imply
Thus the averaged endpoint measure supplies enough translations to expand any nontrivial set .
To turn expansion into covering, we use the Lyapunov function
Conditionally on macroscopic reach at the initial level, the preceding estimate implies exponential decay of the mean of at rate of order . Since every noncovering range satisfies , any fixed positive sprinkling then forces complete coverage with high (conditional) probability.
3. Subcritical locality: exponential moments replace exact geometry. Below , 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 , making ranges of order exponentially unlikely.
Every connected subset of with fewer than vertices projects injectively onto . 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 , 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 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
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 may be viewed as the jump measure of a translation walk on , 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 and denote the ranges of the -avalanches rooted at the origin in and , respectively. We use the standard monotone graphical construction in the level parameter, so that
The key property is regeneration in the level variable. We use the censored exploration filtration : during the exploration of a -avalanche, refreshed values at most are revealed exactly, whereas values larger than are revealed only through the event that they exceed . The -field 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 at level is finite, conditionally on , the terminal fitnesses at its sites are independent with law . Each site
therefore carries an independent rate-one clock in the level parameter. At a mark rooted at and level , an independent -avalanche is grafted onto the current range:
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 sites, there are at most active fitnesses at or below . A block of updates in which all fresh values exceed removes these active fitnesses without creating new ones, and therefore forces termination within the block. Such a block has conditional probability at least
The probability of completing blocks without either terminating or reaching more than sites is consequently at most . Letting , and then taking the union over , 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 . As these variables are distinct, independence and the memoryless property imply that, conditionally on , the terminal fitnesses are independent with law .
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 with , independently for each site , take Poisson marks with intensity on the torus and in infinite volume. Starting from , retain a mark whenever belongs to the current range, in which case
This realizes the preceding level-by-level description. On the torus it is directly nonexplosive. In infinite volume, monotonicity gives
and one may equivalently stop when the range reaches sites and then let . □
Thus, conditionally on a current range , marks occur at total rate . 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 denote the duration of the finite-volume -avalanche. With and as above, set
Lemma 2.2 (Range–duration identity). For every , , and on the torus for every finite . Moreover, is continuous on , is continuous on , and and are on these respective domains, with
and
Consequently,
Proof. Fix , so that for . In the decorated hierarchical construction a mark at level adds exactly the duration of an independent -avalanche. For , put . Since
compensation, first stopped when the range reaches sites, gives after letting $K\to\infty
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 almost surely, the observable is bounded by , and its expected compensator is bounded by . Gronwall’s inequality yields
Monotone convergence as proves , uniformly on . The same argument on the torus uses and proves on every bounded level interval.
We next record the continuity of the mean ranges. Couple the avalanches , , through the same hierarchical construction. Since almost surely, only finitely many relevant marks occur below level , and no deterministic level carries a mark. Hence, for any deterministic in , one has , and therefore , eventually almost surely. Since and , dominated convergence yields . Thus is continuous on . The same argument on the torus is immediate from , and shows that is continuous on every bounded level interval.
We may now apply compensation to the full duration. Tonelli’s theorem gives
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 and . Together with the continuity of and , the fundamental theorem of calculus then yields the stated identities. Since , integrating and 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 , put
with the convention
Lemma 3.1 (Avalanche compensation). For every and ,
where, if is an independent copy of ,
In particular, is locally Lipschitz on . The functions and are jointly continuous on , and
The -derivative at is understood from the right.
Proof. Fix , and stop the hierarchical construction when the range reaches sites. Before this stopping time the total rate of relevant marks is at most , so compensation applies without any integrability assumption.
For
a graft rooted at produces the increment
Compensation for the stopped construction therefore gives the stopped version of (3). The compensator is uniformly bounded in the stopping level, since
Moreover,
Dominated convergence as proves (3) and (4). The same bound gives
hence local Lipschitz continuity.
For joint continuity, couple two independent range processes on , with . The pathwise local-constancy argument established in the proof of Lemma 2.2 applies verbatim: since , the terminal ranges are finite almost surely, and for any deterministic in , both ranges are eventually unchanged almost surely. No moment assumption is needed here. If also , the integrand defining converges almost surely and is bounded by , where is a common lower bound for the exponents. Dominated convergence gives joint continuity of . The same bound justifies differentiation of under the expectation and joint continuity of ; the integrand of itself is bounded by one. Finally, (3) and continuity of yield at every level in the stated domain. □
The geometric input is the following elementary inequality.
Lemma 3.2 (Average creation under translation). Let be finite, with , and write
If , then
Proof. Write , and first assume . Equip with the translation-invariant lexicographic order, and write
Then . From , at most positive displacements and negative displacements can land in . Hence
Here .
For general , choose uniformly a subset of size containing zero. The zero displacement contributes nothing, while every other displacement is retained with probability . Applying the equal-cardinality case gives
Since and ,
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 and . In particular, neither isoperimetry nor any metric property of enters the infinite-volume sharpness argument.
We now combine the two lemmas.
Proposition 3.4 (Burgers comparison). For every and ,
Equivalently,
for every .
Proof. Fix , and let
Then and are independent and identically distributed, finite almost surely, and
For , put
On , Lemma 3.2 gives
Since is concave, vanishes at zero, and ,
Consequently, on ,
Lemma 3.1 therefore gives
Set
The ratio
is decreasing on . Hence, on ,
By exchangeability,
Therefore
Independence gives
where differentiation under the expectation is justified by
Combining this with (7),
Integrating in (3) proves (6). □
The next lemma records the characteristic consequence of the Burgers comparison.
Lemma 3.5 (Comparison of the characteristics). Fix and , and put .
Define
as long as . Then
for every such that
Proof. Fix with and , and put for . Lemma 3.1 permits the ordinary chain rule. Proposition 3.4 therefore gives
Set , which is continuous on this interval. Multiplication by the integrating factor gives
Since , it follows that on . The choice of was arbitrary. □
Proof of Theorem 1.2. Fix and , and set
Consider
Its hitting time of zero is
We claim that
Indeed, if , then Lemma 3.5 gives
Since almost surely and ,
by dominated convergence, a contradiction. Therefore
or equivalently
Since almost surely,
Monotone convergence yields
Thus for every , and hence
The reverse inequality is immediate. Therefore
It remains to identify the duration threshold. Since , finite mean duration implies finite mean range, so . Conversely, Lemma 2.2 gives for every , hence . Thus
and (1) follows. □
Remark 3.6. The quantity
may be viewed as the probability that the avalanche range is hit by an independent Poisson field of intensity . Infinite susceptibility corresponds to an infinite right slope at , whereas an infinite avalanche corresponds to positive mass at the boundary:
The Burgers comparison forces positive values of 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
and put
For , identify
with its canonical image in . Since for all sufficiently large , this identification is unambiguous. Let
The finite- and infinite-volume graphical constructions agree until the range first reaches . Therefore
In particular,
For , write and define its translation boundary:
Since and have equal cardinality,
Moreover,
since both sides count pairs in according to their displacement.
Lemma 4.1 (Short translations control all translations). Let satisfy and
For , with fixed, there exists such that, for all sufficiently large ,
Proof. Set
For , subadditivity gives
Summing over yields
Since for all sufficiently large , it follows that . For , write , where . Then
Also , hence , and each value of occurs for at most values of . Therefore
Let
be the hyperoctahedral group, acting on by sign changes and coordinate permutations.
Proposition 4.2 (Translation expansion). Let be an event with positive probability. Suppose that on , for every , we are given a nearest-neighbour edge , , , whose joint conditional law given is -invariant. Define
Then
and there exists such that, for every , writing , and all sufficiently large ,
The constant is uniform in , , and in the law of the selected edges.
Proof. Symmetrize the translation boundary by putting
The function is subadditive and -invariant, with and . Invariance of and (10) give
Fix , and let be its absolute coordinates, so that . Permute the coordinates to put first, then reflect that coordinate. The difference of the two resulting vectors is , and subadditivity and invariance give
Consequently,
and the endpoints yield
For each selected edge, , so
Summing gives . Together the two estimates imply
To control all translations, decompose each vector into its coordinates. By subadditivity and coordinate symmetry,
If is odd, multiplication by two permutes , so the last sum equals . If is even, writing each odd translation as an even translation followed by gives
Lemma 4.1, applied to , gives
Since for all sufficiently large , the extra term in the even case satisfies
Thus in both parity cases we conclude that
Combining (12) and (13) yields
Since
the preceding estimate, after changing the constant, gives
Finally,
which proves (11).
Remark 4.3. We need both endpoints of the selected edges to handle parity. The vertices alone control the even translations . When is even, these do not control odd translations. Using also , the nearest-neighbour relation yields , which bridges the two parity classes. For odd , multiplication by is invertible modulo , and this extra ingredient is unnecessary.
Remark 4.4. Proposition 4.2 also gives a spectral-gap bound for the auxiliary translation walk. Put and , and let . By symmetry of , this kernel is reversible for the uniform law. The corresponding continuous-time generator is
The proposition gives the conductance bound
Hence Cheeger’s inequality [9 Theorem 13.10] implies that the spectral gap of satisfies
Conversely, a spectral-gap estimate of the form automatically provides an alternative proof of Proposition 4.2. Indeed, writing for the uniform law, the Dirichlet form associated with is
Thus
Poincaré’s inequality [9 Lemma 13.7 and Remark 13.8],
then implies (11), up to constants.
We now apply Proposition 4.2 to the avalanche. On , connectedness of implies that, for every , the range contains an edge
Choose one uniformly among all such edges, independently for each conditional on the range. This selection is equivariant under , so its conditional law given is -invariant. Set
Since
while (8) gives
and, since for all sufficiently large ,
Hence Proposition 4.2 gives such that, uniformly in , , and large ,
By Proposition 2.1, conditionally on the current range at level , every carries marks at rate one, and a mark rooted at grafts an independent -avalanche translated by .
For such a graft, the event and the selected endpoints refer to the independent avalanche. Let be the number of distinct sites outside among the selected translated endpoints, with when fails. A vertex on can occur only as , and possibly also as ; hence every selected endpoint has multiplicity at most two. Therefore
Summing over the possible roots, taking expectations, and using the definition of and (14) gives, after decreasing the constant,
The expectation is over the independent graft and the auxiliary selection of its crossing edges, with fixed. The actual number of new sites created by a graft rooted at dominates .
Proposition 4.5 (Supercritical covering). For every and every fixed ,
Consequently,
Proof. Condition throughout on . Connectedness gives , and the range can only increase. For a nonempty , put
Let . A graft rooted at adds new sites, where is an independent -avalanche. Since , , and is increasing,
Let denote the generator of the range process. Each root has rate one, so (15) gives, uniformly for and ,
The conditioning event belongs to , and all subsequent grafts retain their independent laws. Dynkin’s formula and Gronwall’s inequality therefore yield
If , then . Consequently,
Finally,
and (8) gives .
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 and for all sufficiently large , its prefactor satisfies
Thus the comparison also supplies a covering bound after a sprinkling , with sufficiently large depending on , , .
Subcritical locality
We prove Theorem 1.6. Recall that
so that for every . 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 , there exists such that
In particular, for every integer .
Proof. Fix and, for , set
By Proposition 2.1, conditionally on a current range of size , marks occur at total rate , and each mark grafts an independent avalanche with . Since
if , then
whereas the truncated observable is constant when . Conditioning on the current range and using the independence of the graft therefore gives, for almost every ,
Indeed, the first factor produced by the rate is bounded by
We claim that exponential integrability propagates over every interval such that . Suppose for some . Since
we may choose such that, writing ,
For , set
Then , and (18) yields
Hence on . In particular,
Letting gives
Now fix . Since , choose a partition with mesh smaller than . The initial range is deterministic. If has an exponential moment, then, by monotonicity of ,
so the preceding propagation argument gives an exponential moment for . Iterating finitely many times yields
for some .
The preceding argument is independent of the geometry of . Geometry enters only when these range moments are used to compare the infinite lattice with the torus.
Lemma 5.2 (Local injectivity). Let
be the canonical projection. If is connected and
then is injective on .
Consequently, for every with
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 sites.
Proof. A connected set with vertices has graph diameter at most . Thus, if ,
On the other hand, distinct satisfying
have
Hence is injective on .
For the second assertion, couple the elementary Bak–Sneppen updates in 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 new sites to the range. Hence, while the current range contains fewer than sites, after the next update it contains fewer than 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 ,
Moreover,
Finally, for every fixed , for all sufficiently large and every ,
Proof. Take
For large ,
so Lemma 5.2 couples the two avalanches identically until their common range reaches . In particular, on ,
Since
we obtain
Fix an integer . By monotonicity in the level parameter and Proposition 5.1,
and
uniformly for . Hence
which proves (19).
By Lemma 2.2,
Therefore (19) implies (20).
Finally fix , and take large enough that
Apply Lemma 5.2 with . The two avalanches then agree until their common range first contains at least sites. On the event that the final range has size , 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
All neighbourhoods on the torus are understood with repetitions removed.
Proposition 6.1 (Finite-volume ergodicity). For every , the Bak–Sneppen chain on has a unique stationary law . From every deterministic initial configuration, its law converges to in total variation. Moreover is translation invariant.
Proof. If one update refreshes the whole torus, then after one step the configuration consists of independent variables, independently of the initial state, and the conclusion is immediate. Assume henceforth that one update does not cover .
Fix , write for the transition kernel and , and set
With probability at least , every fitness sampled in the next updates exceeds , uniformly in the current state. On this event no value at most is created. As long as such values remain, the selected minimum is at most and is refreshed, so their number decreases at each update. After at most updates none remains. Hence
Iteration gives a geometric tail for the hitting time of , uniformly in the initial state.
For the minorization on , 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 , while every other refreshed vertex receives a fitness above ; at the final update require all fresh values to exceed . At each intermediate step the prescribed next vertex is then the unique site with fitness at most , so the avalanche follows the walk, covers the torus, and terminates. These finitely many requirements have positive probability, so . Since the avalanche duration is integer-valued, there exists such that
This range–duration event is measurable with respect to the censored level- exploration, whose law, up to translation of its starting point, does not depend on . On this event every site has been refreshed, so Proposition 2.1 implies that the terminal fitnesses are independent with common law , independently of the censored exploration. Thus, setting
we obtain
Combining (22) and (23) by the Markov property gives
Hence is a contraction in total variation with factor . The space of probability measures is complete for this metric, so Banach’s fixed-point theorem gives a unique fixed point . Since is also -invariant, uniqueness gives , so is the unique invariant law of . Iterating the contraction and using that Markov kernels are nonexpansive in total variation yields
for every initial law .
Finally, start the chain from independent 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 is translation invariant. □
We next introduce the locking variables. Start the finite-volume chain from independent fitnesses and set . If is the site of minimum fitness at time and
define
For , set
Proposition 6.2 (Conditional exponential representation). Conditionally on , the fitnesses are independent and
Consequently, for distinct and ,
Proof. The assertion holds at time zero. Assume it holds at time , and condition on , the location , and the value of the minimum. By the lack-of-memory property, the unrefreshed coordinates remain independent and can be written as
where the are independent variables. The refreshed coordinates are fresh independent exponentials. Thus, conditionally on these variables, the fitnesses at time are independent exponentials above the thresholds in (25). Since this conditional law depends on the conditioning variables only through , the same statement holds conditionally on alone. This proves the claim by induction.
Put
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 ,
Proof. Let
Let be the set of configurations with all fitnesses above . By (22), its hitting time has finite mean. Discard the initial segment up to the first visit to , and from then on decompose the trajectory into successive -avalanches. If is the starting time of the -th one and its range–duration pair, then , and these pairs are i.i.d. with the law of . At the first update of the avalanche the current minimum exceeds , so every site outside its eventual range has its locking threshold raised above . During the remaining updates the selected minima are at most , and only sites in the avalanche range are refreshed. Hence
For all sufficiently large , a one-step noncovering -avalanche has positive probability, so the cycle length has span one. The delayed renewal–reward theorem and translation invariance therefore give
For every fixed , using ,
Now
By Proposition 5.3, as this converges to , while . Letting first and then in (28) proves (27). □
Proposition 6.4 (Supercritical locking thresholds). For every ,
Proof. Fix . Starting from the process defining , ignore the initial segment up to the first visit to . Thereafter run successive -avalanches until the first one covering , and then continue until the next visit to .
The initial delay and the final recovery have finite mean by (22). Moreover by Proposition 6.1, and by Lemma 2.2, so the waiting time for the first covering -avalanche also has finite mean.
At the end of this covering avalanche, every site has been refreshed. Conditionally on its -censored exploration, the terminal fitnesses are independent with law , by Proposition 2.1. Censor the subsequent recovery at level : reveal fitnesses at most exactly and reveal larger fitnesses only through their exceedance of . At the designated return to , each terminal value has been observed only to exceed . Thus, conditionally on the censored cycle history, the terminal vector has law
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 be the time spent in successive -avalanches up to and including the first covering one in a cycle started with law , and let be the subsequent recovery time to . For large , the cycle length has span one, since a one-step noncovering -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 , and throughout the recovery every selected minimum is at most . Thus
during the whole recovery phase. By the delayed renewal–reward theorem,
Since the successive -avalanches are i.i.d. until the first covering one, Wald’s identity gives
To bound from below, use the decorated construction of Proposition 2.1. Since the covering -avalanche has range , every site is active in the level extension from to . For each , let be the first mark above on the vertical line through . Then are i.i.d. , and, conditionally on , the corresponding graft contributes a duration of mean . Hence, by Tonelli and monotonicity of ,
Combining (30)–(32),
Choose . By Theorem 1.5,
Since , the right-hand side of (33) tends to zero, proving (29).
Proposition 6.5 (Locking-threshold concentration and product limit). For every ,
Consequently, for every fixed collection of distinct sites , the corresponding marginal of converges to .
Proof. For , Propositions 6.3 and 6.4, applied at and , give (34); for , the lower-tail event is empty and the same conclusion follows from the supercritical bound. For fixed , and large enough that their images in are distinct, translation invariance and the union bound imply
Fix and define
The function is bounded and continuous. If
then as . By Proposition 6.2,
For fixed , Proposition 6.1 lets and replaces the law of the chain by . Letting next and finally proves convergence of the joint distribution functions to , the distribution function of . □
Proof of Theorem 1.1. Proposition 6.5 gives the asserted convergence in exponential coordinates. Under the inverse transformation , a random variable with law becomes uniform on . This proves the theorem. □
Use of AI
The (compensated) Laplace transform of the nonlinear additive process in which is replaced by , with an independent copy of , at rate , satisfies an inviscid Burgers equation. Starting from , its solution is given by the Borel() distribution, the law of the total progeny of a Poisson() Galton–Watson tree. The phase transition at 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]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]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]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]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]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]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]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]Olivier Hénard, Sharpness of the avalanche phase transition in the Bak–Sneppen model, arXiv preprint, arXiv:2608.10992 (2026).
- [9]David A. Levin and Yuval Peres, Markov chains and mixing times, 2 ed., American Mathematical Society, Providence, RI, 2017.
- [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]Ronald Meester and Dmitri Znamenski, Limit behavior of the Bak–Sneppen evolution model, Annals of Probability 31 (2003), 1986–2002.DOI
- [12]———, Critical thresholds and the limit distribution in the Bak–Sneppen model, Communications in Mathematical Physics 246 (2004), 63–86.DOI
- [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]Maya Paczuski, Sergei Maslov, and Per Bak, Avalanche dynamics in evolution, growth, and de-pinning models, Physical Review E 53 (1996), 414–443.
- [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]Oliver Riordan and Lutz Warnke, The evolution of subcritical achlioptas processes, Random Structures & Algorithms 47 (2015), no. 1, 174–203.DOI
- [17]Stanislav Volkov, Rigorous upper bound for the discrete Bak–Sneppen model, Journal of Statistical Physics 186 (2022), 8.arxiv.org/abs/2003.00222