The Power Light Cone of the Discrete Bak-Sneppen, Contact and other local processes

We consider a class of random processes on graphs that include the discrete Bak-Sneppen (DBS) process and the several versions of the contact process (CP), with a focus on the former. These processes are parametrized by a probability $0\leq p \leq 1$ that controls a local update rule. Numerical simulations reveal a phase transition when $p$ goes from 0 to 1. Analytically little is known about the phase transition threshold, even for one-dimensional chains. In this article we consider a power-series approach based on representing certain quantities, such as the survival probability or the expected number of steps per site to reach the steady state, as a power-series in $p$. We prove that the coefficients of those power series stabilize as the length $n$ of the chain grows. This is a phenomenon that has been used in the physics community but was not yet proven. We show that for local events $A,B$ of which the support is a distance $d$ apart we have $\mathrm{cor}(A,B) = \mathcal{O}(p^d)$. The stabilization allows for the (exact) computation of coefficients for arbitrary large systems which can then be analyzed using the wide range of existing methods of power series analysis.


Introduction
In physics, critical behaviour involves systems in which correlations decay as a power law with distance. It is an important topic in many areas of physics and can also be found in stochastic processes on graphs. Often, such systems have a parameter (e.g. temperature) and when it is set to a critical value, the system exhibits critical behaviour. Power series expansion techniques have been used in the physics literature to numerically approximate critical values and associated exponents. It was often observed that the coefficients of such power series stabilize when the system size grows, and we provide a rigorous proof of this for a large class of stochastic processes.
Self-organized criticality is a name common to models where the critical behaviour is present but without the need of tuning a parameter. A simple model for evolution and selforganized criticality was proposed by Bak and Sneppen [BS93] in 1993. In this random process there are vertices on a cycle each representing a species. Every vertex has a fitness value in [0 1] and the dynamics is defined as follows. Every time step, the vertex with the lowest fitness value is chosen and that vertex together with its two neighbors get replaced by three independent uniform random samples from [0 1]. The model exhibits self-organized criticality, as most of the fitness values automatically become distributed uniformly in [ 1] for some critical value 0 < < 1. This process has received a lot of attention [dBDF + 94,Mar94,Bak96,MDLRM98], and a discrete version of the process has been introduced in [BK01]. In the discrete Bak-Sneppen (DBS) process, the fitness values can only be 0 or 1. At every time step, choose a uniform random vertex with value 0 and replace it and its two neighbors by three independent values, which are 0 with probability and 1 with probability 1 − . The DBS process has a phase transition with associated critical value [MZ02,Ban05]. The Bak-Sneppen process was originally described in the context of evolutionary biology but its study has much broader consequences, e.g., the process was rediscovered in the setting of theoretical computer science [CCS + 17]. To study the limits of a randomized algorithm for solving satisfiability, the discrete Bak-Sneppen process turned out to be a natural process to analyze.
The DBS process is closely related to the so-called contact process (CP), originally introduced in [Har74]. Sometimes referred to as the basic contact process, this process models the spreading of an epidemic on a graph where each vertex (an individual) can be healthy or infected. Infected individuals can become healthy (probability 1 − ), or infect a random neighbor (probability ). The contact process has also been studied in the context of interacting particle systems and many variants of it exist, such as a parity-preserving version [Inu95] and a contact process that only infects in one direction [TIK97]. Depending on the particular flavor of the processes, the CP and DBS processes are closely related [Ban05] and in certain cases have the same critical values. The processes are similar in the sense that vertices can be active (fitness 0 or infected) or inactive (fitness 1 or healthy). The dynamics only update the state in the neighborhood of active vertices with a simple local update rule. In this article we consider a wide class of processes that fit this description, and our proofs are valid in this general setting. We will, however, focus on the DBS process when we present explicit examples.
In this paper we take a power-series approach and represent several probabilities and expectation values as a power series in the parameter . There is a wealth of physics literature on series analysis in the theory of critical phenomena, see for example [HB73,BH73,HB79] for an overview. Processes typically only have a critical point when the system size is infinite, but numerical simulations often only allow for probing of finite systems. Our main theorem proves, for our general class of processes, that one can extract coefficients of the power series for an arbitrary large system by computing quantities in only a finite system. One can then apply series analysis techniques to these coefficients of the large system. Series expansion techniques are not new, they have been extensively used for variants of the contact process as well as for closely related directed percolation models [Dic89,JD93,Inu95,IK96,TIK97,Inu98,KTIK99] in order to extract information about critical values and exponents. For example, in [TIK97] the contact process on a line is studied where infection only happens in one direction. In [Inu95] a process is studied where the parity of the number of active vertices is preserved. In both articles, the power series of the survival probability is computed up to 12 terms and used to find estimates for the critical values and exponents. However, in all this work the stabilization of coefficients has been observed 1 but not proven.
Our main contribution is a definition of a general class of processes that encapsulates most of the above processes (Definition 2) and an in-depth understanding of the stabilization phenomenon, complete with a rigorous proof (Lemma 9, Theorem 13). The results are illustrated with examples.
Road map. In Section 1.1 we will provide two example power series that exhibit the stabilization phenomenon. In Section 1.2 we will sketch our results without going into technicalities and explain the intuition behind them, something that we call the Power Light Cone. In Section 2 we define our general class of processes in more detail and provide our theorems with their proofs. In Section 3 we apply our result to the DBS process, and we compute powerseries coefficients for several quantities. As an application, we apply the method of Padé approximants to extract an estimate for and we estimate a critical exponent that suggests that the DBS process is in the directed percolation universality class.

Stabilization of coefficients
There are different ways of defining the DBS process. These definitions map to each other in straightforward ways, and only slightly differ in the notion of a "time step". For example, one can define a time step as picking a random vertex (not necessarily with value 0) and then not do anything when it has value 1 but still count it as a time step. To study the process for an infinite-sized system, one can consider a continuous-time version of this process with exponential clocks at every vertex. Resampling of a vertex and its neighbors happens when the clock of the vertex rings and the current value of the vertex is zero.
In this section we consider the discrete-time process on a finite system, where a 0-vertex is picked in every time step, and the process terminates when an all-1 state is reached. Furthermore, we will always refer to the vertex-state 0 as active and 1 as inactive.
From numerical simulations it is apparent that there is a phase transition in the DBS process when goes from 0 to 1. There is some critical probability such that for < the active vertices quickly die out and the system is pushed toward a state with no active vertices. However for > , the active vertices have the overhand and dominate the system. This phase transition can clearly be seen in Figure 1 from two different perspectives: (a) We have computed the expected number of steps per vertex, on a cycle of length , to reach the all-inactive state, where the vertices are initialized i.i.d. to active with probability . (b) We have computed the probability that the end of a (non-periodic) chain is activated when the process is started with only one active vertex on the other end.
Let us write these functions as a power-series in and in = 1 − respectively.
We will define these functions in more detail in Section 3, where we show, amongst other things, that they are rational functions for each . For example R (4) ( ) = (6 − 12 + 10 2 − 3 3 ) 6(1 − ) 4 = (1 − )(1 + + 2 + 3 3 ) 6 4 Although they only have an operational meaning for ∈ [0 1], we give a plot of such a function over the complex plane, see Figure 2. The plot shows the poles of S (6) ( ), which seem to approach the value on the real line (for larger see Figure 5). Something similar happens for partition functions in statistical physics. The partition function is usually in the denominator of observable physical quantities, so that its zeroes are the poles of such quantities. A classic  (2), the probability to 'reach' the other side of the system: the DBS process on a non-periodic chain of size is started with a single active vertex at position 1 (denoted by start {1}) and we plot the probability that vertex ever becomes active (denoted BA ( ) ) before the all-inactive state is reached. For = 5000 the result was obtained with a Monte Carlo simulation. For the lower , the results were computed symbolically. The inset shows a zoomed in version of the Monte Carlo data, showing that ≈ 0 635.    result on the partition function for certain gasses [YL52] shows that when an open region around the real axis is free of zeroes, then many physical quantities are analytic in that region and therefore there is no phase transition. In [PR18] the hardcore model on graphs with bounded degree is studied, and it is proven that the partition function has zeroes in the complex plane arbitrary close to the critical point. For now, we simply want to highlight the behaviour of the coefficients ( ) and [ ] .
We now know the first few terms of the power series for arbitrary large systems and we can proceed to use methods of series analysis. By applying the method of Padé approximants, we can estimate ≈ 0 6352. More details on this can be found in Section 3.

Locality of update rule implies stabilization
Our proofs are based on an observation that we call the Power Light Cone. Let X be a set of vertices, and let L X be an event that is local on X , meaning that the event depends only on what happens to the vertices in X . For example, when X = { 0 } and L X is the event that vertex 0 is picked at least times, then L X is local on X . In Section 2 we will give a more precise definition of local events. We now wish to compare the probability P(L X ) when the 2 At first sight one is tempted to conjecture that the coefficients ( ) are all non-negative and are monotone increasing with . Unfortunately neither of these conjectures hold since (10) 1114 < 0. We found this counterexample by observing that the radius of convergence for R 10 ( ) is less than 0 96, which considering that R 10 ( ) is bounded on [0 0 96] implies that there must be a negative coefficient in its power series. process is initialized in two different starting states, A and A . When A and A differ only on vertices that are at least a distance away from X , then we have P(L X | start in A) − P(L X | start in A ) = ( ) By the notation ( ) we mean that when this quantity is written as power series in , then the first − 1 terms of the series are zero. It only has non-zero terms of order and higher, i.e. the two probabilities agree on at least the first − 1 terms of their power series. This is the essence of the Power Light Cone. A vertex that is a distance away from the set X will only influence probabilities and expectation values of X -local events with terms of order or higher. The intuition behind this is that the probability of a single activation is ( ) and in order for such a vertex to influence the state of a vertex in X , it needs to form a chain of activations of size to reach X . This observation will also allow us to compare the process on systems of different sizes.

Lemma 1 (Informal version of Lemma 9) Let G and G be two graphs and let X be a set of vertices present in both graphs such that the -neighborhood of X and the local update process (a single update may only affect a vertex and its neighbors) on it is the same in both graphs. Then for any event L X that is local on X we have
This idea applies to expectation values as well. Consider the expected number of steps per vertex on a cycle. By translation invariance, we have 1 E(total steps) = E(#times vertex 1 was picked) making it a {1}-local quantity. If we add an extra vertex to the cycle, the expectation value only changes by a term of order ( /2 ) since the new vertex has distance /2 to vertex 1.

Parametrized local-update processes
The class of parametrized (discrete) local-update processes, introduced in this section, includes the DBS, the CP and many other natural processes. We prove a general 'stabilization of the coefficients theorem' for them, suggesting the usefulness of the power-series approach for members of the class. Let G = (V E) be an undirected graph with vertex set V and edge set E. We consider processes where every vertex of G is either active or inactive. A state is a configuration of active/inactive vertices, denoted by the subset of active vertices A ⊆ V . For ∈ V let us denote by Γ( ) the neighbors of in G including itself. A local update process in each discrete time step picks a random active vertex ∈ A and resamples the state of its neighbors Γ( ). If the state is ∅ (there are no active vertices) then the process stops and all vertices remain inactive afterwards.

Definition 2 (PLUP -Parametrized local-update process) We say that M G is a parametrized local-update process on the graph G
The probability P R is a polynomial in such that for = 0 we get A A with probability 1, i.e., when any previously inactive vertex becomes active ( |A \ A| > 0) or when A = A then the constant term in P R must be zero. 5

(iv) Termination. The process terminates when the all-inactive state ∅ is reached.
We write P G and E G for probabilities and expectation values associated to the PLUP M G .

Definition 3 (Local events) Let G = (V E) be a graph and let M G be a PLUP. Let S ⊆ V be any subset of vertices, and let ∈ V be any vertex.
• Let II (S) be the event that all vertices in S get initialized as inactive.
• Let RI (S) be the event that all vertices in S remain inactive during the entire process.
• Define BA (S) as the complement of RI (S) : the event that there exists a vertex in S that becomes active at some point during the process, including initialization.
• Let #A ( ) be the number of times that was selected while it was active.
• Let # ( ) be the number of times that the value of was changed.
• We say an event L is local on the vertex set S if it is in the sigma algebra 6 generated by the events

Lemma 4 (Time equivalence) The three versions of the selection dynamics of a PLUP, desribed in property (ii) of Definition 2, are equivalent for local events. That is, for any local event L the probability P(L) is independent of the chosen selection dynamics in property (ii)
. 4 The properties of the selection dynamics are used in the proof of Lemma 7 5 The condition |A \ A| > 0 =⇒ P R = ( ) is used in the proof of Lemma 8: a fresh activation is at least one power of so you need to cover a distance . The extra condition A = A =⇒ P R = ( ) is used for absolute convergence in Claim 18 because without it you can have infinitely many paths with a finite power of . 6 If < 1, then the process terminates with probability 1. If we remove the 0 probability event of not terminating, then the discrete nature of the process implies that the probability space has a discrete set of atoms see, e.g., Definition 16 and Equation (13), therefore one does not need to worry about difficulties arising from σ -additivity.
Proof. The three selection dynamics only differ in the counting of time, and the presence of self loops in the Markov Chain. The definition of local events only includes events that are independent of the way time is counted. They only depend on which active vertices are selected and the changes to the state of the graph.
It is easy to see that (ii)b implements the dynamics of (ii)a via rejection sampling, therefore they give rise to the same probabilities. One can also see that on a finite graph the selection rule (ii)c induces the same selection rule as (ii)b. This is because the exponential clocks induce a Poisson process at each vertex. The independent Poisson processes with rates are equivalent to one single Poisson process with rate W = ∈V but where each point of the single process is of type with probability /W . One can simulate (ii)c by sampling a time value from an exponential distribution with parameter W and then sampling a random vertex with probability /W (as in (ii)b). Since the time is not relevant for local events we can ignore the sampled time value and this gives rise to the same probabilities.
Our lemmas and theorems only concern local events and therefore we can use any one of the three selection dynamics when proving them.
Definition 5 (Induced process) Suppose that V ⊆ V , then we define the induced process M G on the induced subgraph G = (V E ) such that we run the process on G and after each step we deactivate all vertices in V \ V . We can then view this as a process on G . Let L be a local event on V . We denote the probability of L under the induced process M G with P G (L).

Similarly we use the notation E G for expectation values induced by the process M G .
It is easy to see that the induced process of a PLUP is also a PLUP.

Definition 6 (Graph definitions) Let G = (V E) be a graph, S ⊆ V be any subset of vertices and ∈ V be any vertex.
• Define G \ S as the induced subgraph on V \ S and G ∩ S as the induced subgraph on S.
• Define the -neighbourhood Γ(S ) of S as the set of vertices that are connected to S with a path of length at most . In particular Γ({ }; 1) = Γ( ).
The following lemma says that if a set S splits the graph into two parts, then those two parts become independent under the condition that the vertices in S never become active.

S X Y
Lemma 7 (Splitting lemma) Let M G be a parametrized local-update process on the graph G = (V E). Let S X Y ⊆ V be a partition of the vertices, such that X and Y are disconnected in the graph G \ S. Furthermore, let L X and L Y be local events on X and Y respectively. Then we have The condition of initializing S to inactive is present only to prevent counting the initialization probabilities twice. Equivalently we could write the condition only once: and by Bayes rule we also have Proof. We will use the 'continuous-time clocks' version of selection dynamics (PLUP property (ii)c). By Lemma 4 the statement will then hold for all versions. We proceed with a coupling argument. There are three processes, one on G and the induced ones on G \ Y and G \ X . We couple them by letting all three processes use the same source of randomness. Every vertex in G has an exponential clock that is shared by all three processes, and the randomness used for the local updates for each vertex will also come from the same source. This means that when the clock of a vertex rings, and the neighborhood Γ( ) is equal in different processes, then the update result will also be equal. Now we simply observe that L X ∩ L Y ∩ RI (S) holds in the G-process if and only if L X ∩ RI (S) holds in the (G \ Y )-process and L Y ∩ RI (S) holds in the (G \ X )-process. This is because all vertices in S are initialized as inactive (all three probabilities are conditioned on this), so a vertex in S can only be activated by an update from a vertex in X or Y . To check if the event RI (S) holds, it is sufficient to trace the process up to the first activation of a vertex in S. Before this first activation, anything that happends to the vertices in X only depends on the clocks and updates of vertices in X , and similar for Y . Since S splits X and Y in disconnected parts, these parts can not influence each other unless a vertex in S is activated. Because of the coupling, the evolution of the X vertices in G \ Y will be exactly the same as the evolution in G, and similar for Y . Once a vertex in S does get activated, the evolution of the three processes is no longer the same but in that case the event RI (S) does not hold, regardless of any further updates in any system. The clocks and updates of each vertex are independent sources of randomness, and when RI (S) holds then all the randomness of the S vertices is ignored. Therefore the probability of RI (S) in the (G \ Y )-process and (G \ X )-process depends only on independent random variables and we get the required equality.

Power Light Cone results
Now we present the results that exhibit the power light cone. The intuition is that if two vertices have distance in the graph, then the only way they can affect each other is that an interaction chain is forming between them, meaning that every vertex gets activated at least once in between them. When we write ( ) = ( ) for some function then we mean the following: when ( ) is written as a power-series in , i.e., ( ) = ∞ =0 α , then α = 0 for 0 ≤ ≤ − 1. Proof. When E holds, all vertices in X become active. By PLUP property (i) any activation in the initial state is ( ) and by property (iii) any subsequent activation is also ( ). Therefore, for any path ξ of the Markov Chain with ξ ∈ E we have P(ξ) = ( |X | ), where P(ξ) is a polynomial in . We have P(E) = ξ∈E P(ξ) by definition. This is a sum over infinitely many polynomials, and by considering P(E) as a power series in we are effectively regrouping terms in this sum. In Appendix B we prove that these regroupings are allowed by proving the absolute convergence of certain series.
Lemma 9 (Graph surgery) Let M G be a parametrized local-update process on the graph G = (V E). If X Y ⊆ V , X ∩ Y = ∅ and L X is a local event on X , then Proof. We can assume without loss of generality, that X = ∅ = Y , otherwise the statement is trivial. Also we can assume without loss of generality that (X Y ) ≤ ∞, i.e., X Y are in the same connected component of G, otherwise we can use Lemma 7 with S = ∅. The proof goes by induction on (X Y ). For the base case, (X Y ) = 1, first note that when = 0, the process initializes everything to inactive by property (i). Depending on whether this atomic event is included in L X , the probability P(L X ) for = 0 (i.e. the constant term) is either 0 or 1 and independent of the graph. Now we show the inductive step, assuming we know the statement for , and that (X Y ) = + 1. First we assume, that RI (X ) ⊆ L X , i.e., L X ⊆ BA (X ) . Define When L X holds, all vertices at distance remain inactive, but for all ≤ − 1 there exists a vertex at distance that become active. These events form a partition L X =˙ ∈[ +1] L X . Below we depict L X graphically: It is easy to see that for all ∈ [ + 1] we have L X ⊆ BA (X ) ∩ ∈[ −1] BA (∂(X )) , and therefore by Lemma 8 we get P G (L X | II (∂(X )) ) = ( ) Now we use, for all ∈ [ ], the Splitting lemma 7 with S = ∂(X ) to split Γ(X − 1) from G \ Γ(X ). We get P G (L X ) = P Γ(X ) (L X | II (∂(X )) ) · P G\Γ(X −1) (RI (∂(X )) ) (by Lemma 7) = P Γ(X ) (L X | II (∂(X )) ) · P G\Y \Γ(X −1) (RI (∂(X )) ) + ( +1− ) (by induction) = P Γ(X ) (L X | II (∂(X )) ) · P G\Y \Γ(X −1) (RI (∂(X )) ) + ( +1 ) (by equation (3)) = P G\Y (L X ) + ( +1 ) (by Lemma 7) Therefore We finish the proof by observing that by RI (X ) is an atomic event of the sigma algebra of the local events of X , so if RI (X ) L X , then we necessarily have RI (X ) ⊆ L X . Therefore we can use the above proof with C X := L X and use that P(L X ) = 1 − P(C X ).
Corollary 10 (Decay of correlations) Let M G be a parametrized local-update process on the graph G = (V E). If X Y ⊆ V and L X L Y are local events on X and Y respectively, then and The proof of this lemma is analogous to the proof of Lemma 9 and can be found in Appendix A.
In order to state our general result about the stabilization of the coefficients in the power series we define a notion of isomorphism between different PLUPs.

Definition 11 (PLUP isomorphism) We say that the PLUPs M G and M G are isomorphic with the fixed sets X X if there is a graph isomorphism : G → G such that (X ) = X , moreover the probability of transitioning in one step from a state A to A is preserved under the isomorphism:
and similarly the probability of initialising to a particular state A is preserved: We denote such an isomorphism relation by M G X X M G Now we define convergent families of PLUPs. Our requirements for such a family of processes imply that the underlying graphs converge to a common graph limit, also called graphing, therefore justifying the term "convergent". Examples of convergent families of PLUPs include DBS and CP on toruses of any dimension, when the limit graphing is just the infinite grid. Less regular examples are also included, such as toroid ladder graphs or discrete Möbius strips of fixed width.

Definition 12 (Convergent family of PLUPs) We say that family {(M G ) : ∈ N} of rooted PLUPs is convergent, if for all ∈ N and for all ≥ we have M
We are ready to state our generic result about the stabilization of coefficients phenomena. In Claim 20 in Appendix B, we prove that these types of sums are absolutely convergent for small enough . Therefore the equality holds when the left-and right-hand side are considered as a power series in .

The discrete Bak-Sneppen process
In Section 1.1 we introduced two quantities that exhibit a phase transition in the DBS process. We saw that the coefficients of their power series stabilize. In this section we will look at them in more detail.

Notation
We denote by M G the DBS process on the graph G = (V E). With a slight abuse of notation we also denote by M G the leaking transition matrix of this time-independent Markov Chain, where the row and column that correspond to the all-inactive configuration is set to zero. We will index vectors (and matrices) by sets A ⊆ V , where A is the set of active vertices, as in Section 2. We will denote probability row vectors by ρ ∈ R 2 so that ρ · M G is the state of the system after one time step. Setting the all-inactive row and column to zero corresponds to the property that for every A ⊆ V we have (M G ) ∅ A = (M G ) A ∅ = 0. We will use notation M ( ) for the process on the cycle of length and M [ ] for the process on the chain (not periodic) of length . In both case we identify vertices with V :

Expected number of resamples per site
The first quantity of interest is the expected number of steps per vertex to reach the all-inactive state. Consider the DBS process on the cycle of length . We start the process by letting each vertex be active with probability and inactive with probability 1 − , independently for each vertex. Denote this initial state by ρ (0) , so ρ (0) Let J be a vector with all entries equal to 1, except for the entry of the all-inactive state which is zero. Then ρ (0) ·M ( ) ·J T is the probability that the all-inactive state has not been reached in steps, starting from ρ (0) . Now define R ( ) ( ) as the expected number of steps per vertex, before reaching the all-inactive state: where P ( ) P ( ) are polynomials as can be seen by using Cramer's rule for matrix inversion. Therefore we can conclude that R ( ) ( ) is a rational function. For small we can compute the functions R ( ) ( ) by symbolic matrix inversion, which is how we obtained the coefficients in Table 1. For ≥ 9 we computed the matrix inversion for rational values of exactly, and then computed the rational function using Thiele's interpolation formula. As we have seen in the previous subsection, R ( ) ( ) is a rational function. Since a rational function is analytic, and R ( ) ( ) has no pole at = 0 (it actually takes value 0), we can write it as

The power-series of R ( ) ( )
where the (non-zero) radius of convergence of the above power series equals the absolute value of the closest pole of R ( ) ( ) to 0. In order to get some intuition about the radius of convergence we plotted the location of the poles of R ( ) ( ) on the complex plane in Figure 3. For = 10 there is a pole at a point with absolute value ≈ 0 9598, hence R (10) ( ) has a radius of convergence strictly smaller than 1 even though the rational function R ( ) ( ) is well-defined for all ∈ [0 1).  As was shown in Section 1.1, Table 1, the coefficients ( ) stabilize as grows. This is proven by Theorem 13, since the family of DBS processes on the cycles, indexed by , is a convergent familiy of PLUPs. The theorem only guarantees the stabilization for > 2 since going from a cycle of size to + 1 adds a vertex at a distance /2 to any fixed vertex. In the table, however, we saw that the stabilization already holds for ≥ + 1. In Appendix D we prove this more precise version of the stabilization that holds for cycles. We define the 'stabilized' coefficients and make the following conjecture.

Conjecture 14 (Radius of convergence) The radius of convergence of R (∞) ( ) is equal to the critical probability of the DBS process.
In Appendix B we explain a method to compute coefficients of the R (∞) ( ) power series (see the text below Claim 19). As an application, we can apply known methods of series analysis. For example, Figure 4 shows estimates for using the ratio method and the Padé approximant method. For details on these methods, see for example [HB73]. The ratio method can be used to estimate the critical value when the singularity that determines the radius of convergence is at , i.e. there are no other singularities closer to the origin, which is what we assume in Conjecture 14. The figure also shows estimates based on the power-series coefficients of the functions T N and S Z . The function T N is the expected number of total steps on a chain with one end, with a single active vertex at that end as a starting state. This series is included because we can compute more terms for it. The function S Z is the probability of survival on the infinite line with a single active vertex as a starting state. This is a series in = 1 − and it is included because other work studies the equivalent function for the contact process and this allows for comparison of critical exponents [Dic89]. The Padé approximant method suggests that the critical value is ≈ 0 63523 ± 0 00005 and that the critical exponent for , which suggests that it is in the directed-percolation (DP) universality class alongside several variants of the contact process [Dic89, Inu95, TIK97].

Reaching one end of the chain from the other
Another quantity we considered in Section 1.1 is the probability of ever activating one end point of a finite chain, when we start the process with only a single active vertex at the other end. Let us consider the length-chain, and suppose we start the DBS process with a single active vertex at site 1. As in Equation (2), we consider Note that in order to satisfy property (i) of the PLUP definition, the initial state needs to be {1} with probability and ∅ with probability 1 − . To get the above definition of S [ ] ( ) with a deterministic starting state one can then simply divide by . The power-series coefficients of S [ ] ( ) stabilize, which follows from Lemma 9 by letting X = { } and Y = {1}. However, as suggested by Figure 1, the limiting power series around = 0 will become the zero function and it is therefore not so interesting. Instead, we can take the power series centered around = 1 and it turns out that also there the coefficients stabilize. We prove this below. Define = 1 − . Similarly to what we did for R ( ) ( ) we can write S [ ] ( ) using a matrix inverse. We will start the process in the (deterministic) state with a single active vertex at location 1, denoted by the probability vector δ {1}  Note that for the event (BA ({ −1}) ∩RI ({ }) ) to hold, all vertices 1 −1 must have been active. Since the process terminates with probability 1, this means all those vertices must also have been deactivated at least once. In the DBS process a deactivation is ( ), so every terminating path of the Markov Chain that is in this set has a factor of at least −1 associated to it, hence Here we use the absolute convergence of certain power series in , which we prove in Claim 23 in Appendix C. We see that S [ ] ( )−S [ −1] ( ) = ( −1 ) so the coefficients stabilize.

A Decay of correlations
Recall Corollary 10.

Corollary 10 (Decay of correlations) Let M G be a parametrized local-update process on the graph G = (V E). If X Y ⊆ V and L X L Y are local events on X and Y respectively, then
and Proof. First observe that if (X Y ) = ∞, it means that either X and Y are in different connected components of G, or one of them is the empty set, therefore L X and L Y are independent events, so the statement holds.
Note that due to Property (i) the only path which has a non-zero constant term is the trivial path, when every vertex is initialized as inactive, thus the constant term of the probability of any local event is either 0 or 1. Also the constant term of P G (L X ∩ L Y ) is 1 if and only if the constant terms of both P G (L X ) and P G (L Y ) are 1, which concludes the (X Y ) = 0 case.
Note that by De Morgan's law, (6) is equivalent with Now we proceed by induction on (X Y ). Assume (5)-(6) hold for (X Y ) = − 1. We will prove the statement for (X Y ) = . We apply a similar idea as in the proof of Lemma 9. Define When L X holds, everything at distance remains inactive, but for all distances ≤ − 1 there exist vertices that become active at that distance. These events form a partition L X =˙ ∈[ ] L X , and similarly for L Y . Below we depict L X ∩ L Y graphically.

B Absolute convergence
Recall Lemma 8. The proof requires the regrouping of terms in an infinite sum. In this section we prove the absolute convergence of certain series that allows for this regrouping. We start with some notation.
Definition 16 (Paths) Define a path of length as an initialization and sequence of updates, where we only count steps in which an active vertex was selected. We write a path ξ as Here denotes the vertex that was selected in the -th step and R ⊆ Γ( ) is the result of the corresponding update that happened afterwards. After steps, the state of the process For a general PLUP we have P(ξ) = P(A 0 )P(( 1 R 1 ) | A 0 )P(( 2 R 2 ) | A 1 ) · · · P(( R ) | A −1 ) where the polynomial P(A 0 ) is the probability of starting in state A 0 and P(( R ) | A −1 ) are polynomials satisfying property (iii) of the PLUP definition. For the DBS process these polynomials take the specific form P(ξ) = P( Here 0 < < are constants depending on the particular process and P(A) is the probability of starting in state A (a polynomial).
Proof. Note that is a sum over finitely many polynomials. It is sufficient to prove the statement for each ξ and it then follows for the sum. Since A ⊆ A it also follows for P( A ). Let ξ be a path as described in Definition 16. As stated in the text below Definition 16 we have P(ξ) = P(A 0 )P(( 1 R 1 ) | A 0 )P(( 2 R 2 ) | A 1 ) · · · P(( R ) | A −1 ) where A ⊆ [ ] is the state after steps and A 0 = A. Let be the degree of the highest order term of any possible local-update step of this process (finitely many possibilities) than maxdeg(P(ξ)) ≤ · + maxdeg(P(A)).
For the DBS process, in the context of Section 3.2, we can slightly refine the above claim. Proof. We repeat the proof of Claim 18. Note that mindeg(P(( R ) | A −1 )) ≥ 1 + |A | − |A −1 | for the DBS process, so = 1. For DBS, = 3 which is the maximum degree of the local update rule ( 3 occurs when all three resampled vertices become active). The claim then follows by noting that P(A) = |A| (1 − ) −|A| in the starting state ρ (0) . This claim is convenient for the computation of the R ( ) ( ) power series. It implies that the term is only present in those polynomials ρ (0) · M ( ) · J T for which − 3 ≤ ≤ . To compute the power-series coefficient ( ) it is sufficient to consider this finite set of polynomials. In other words, in order to compute R ( ) ( ) up to -th order in , it suffices to consider only the first steps of the DBS process. We use this observation to compute the coefficients of the ≥ 18 series by computing matrix powers symbolically in , see Table 1.
Claim 20 There is a constant δ > 0 such that, for any polynomial ( ), the following series is absolutely convergent for ∈ [0 δ]: Note that the sum is over all paths, not only the terminating ones. Proof. We have P(ξ) = P(A 0 )P(( 1 R 1 ) | A 0 )P(( 2 R 2 ) | A 1 ) · · · P(( R ) | A −1 ). The polynomials P := P(( R ) | A −1 ) come from a finite set of polynomials: for each vertex there are at most 2 |Γ( )| possible updates and there are at most vertices. Therefore there is a constant C such that for all these polynomials P abs ≤ C mindeg(P ) By Claim 18 there is a such that P(ξ) abs ≤ P(A 0 ) abs P 1 abs · · · P abs ≤ P(A 0 ) abs C mindeg(P 1 )+···mindeg(P ) There are at most (2 max ) paths of length for a fixed starting state so we have Since there are finitely many (2 ) starting states A, the whole expression is absolutely convergent for < 2 max C −1/ . Since any local event is a subset of the set of all terminating paths, the powerseries for any event is also absolutely convergent. This yields the following corollaries.
Proof. For ∈ [0 δ] we have P(E) = ∞ = by Claim 20. By uniqueness of power series, this equality holds for all up to the radius of convergence.

Corollary 22
Let A 0 ⊆ [ ] be any state and let E be an event such that for all paths ξ ∈ E we have P(ξ | start A 0 ) = ( ). Then P(E | start A 0 ) = ( ).
Proof. The proofs of Claim 20 and Corollary 21 also hold when everything is conditioned on a fixed starting state A 0 ⊆ [ ]. The sum over A and the factors P(A) abs are simply removed from the equations.

C Absolute convergence for the power series
We now turn our attention to the S [ ] ( ) series defined in Equation (10). This process starts with a single active vertex at position 1, i.e. A = {1}, and we look at the probability that vertex is never activated, P(RI ({ }) | start A), as a function of = 1 − . To prove the absolute convergence of such series for general PLUPs we require more properties than the definition of PLUP that we have. We now consider the update polynomials as a function of = 1 − . The update rule for a single time step should satisfy the following two properties.
• When = 0 then the probability that an active vertex becomes inactive is zero.
This implies that any inactivation is ( ).
• There is a > 0 such that if = 0 then: for all vertices with an active neighbor, the probability of activating when that neighbor is selected is at least .
These properties are satisfied by the CP and DBS process.  By the second additional property there is a such that any neighbor of an active vertex can be activated with probability at least when = 0, so the constant term in the corresponding transition polynomials is at least . When taking the absolute polynomials Q abs ( ) this also holds for non zero . The functionsQ ( ) are then at least ( + ( ))/(1 + ( )) so there is a 0 > 0 such thatQ ( ) ≥ = /2 for all 0 ≤ ≤ 0 .
Define the following random variables. Let I B ∈ {0 1} be 1 if any active vertex got inactivated in step or if the process has terminated or if a vertex in B has been active at some point before step . Let G B ∈ {0 1} be 1 if any inactive vertex got activated in step or if the process has terminated or if a vertex in B has been active at some point before step . Let Proof. By definition, if there are no active or no inactive vertices in step then G B = 1. If there are, then an active vertex with inactive neighbor is picked with probability at least (this holds for both the original process and the process withQ ( ) transition functions) and the neighbor is then activated with probability at least . Therefore we haveP(G B = 1) ≥ · > 0. Define i.i.d. Bernoulli variables C with success probability · and C = =1 C . The expectation of C is E(C ) = and using the Chernoff bound we can bound the probability that C deviates far from its mean: We use a coupling argument to compare the G B variables with the C 's. Let U be i.i.d. uniform [0 1] variables. Define C to be 1 if U < so the C 's are indeed i.i.d. Bernoulli variables with the correct distribution.
The probability P(G B = 1) depends on the history of the process. Run the process and in each step of the process, first compute the true probability = P(G B = 1 | history), so > . Now instead of running the process normally, define G B = 1 if and only if U < . Then continue the process conditioned on the value of G B so it follows the normal distribution. This way, the G B variables come from the correct distribution but they are coupled to the C 's. We see C = 1 implies G B = 1 so G B ≤ ≥ C and therefore the Chernoff bound applies to G B ≤ as well. Now we continue the proof of Claim 23.
By the first property every inactivation has an update step that is ( ) (even with theQ functions). Every path of length has I B ≤ inactivations so has probability ( ) I B ≤ . Let T be the event that the process takes exactly steps. For the "tilde process" we havẽ In the main text we looked at R ( ) ( ) and saw that the coefficients of this series stabilize. Our main theorem, however, only shows that this stabilization happens for > 2 since any new vertex added by going from to + 1 is at distance at most /2 from all existing vertices. In this section we prove the stronger statement, namely that the coefficients stabilize for > . Let P C := RI (∂C ) ∩ ∈C BA ({ }) be the event that every vertex in C becomes active, and the boundary remains inactive. If P C holds, we say C is a patch of the activations. The intuition of the following theorem is similar to that of Corollary 10. A site can only realize the length of the cycle after an interaction chain was formed around the cycle, implying that every vertex was activated at least once.