Chase-Escape Percolation on the 2D Square Lattice

Aanjaneya Kumar∗ Department of Physics, Indian Institute of Science Education and Research, Dr. Homi Bhabha Road, Pune 411008, India

Peter Grassberger† JSC, Forschungszentrum J¨ulich, D-52425 J¨ulich,Germany

Deepak Dhar‡ Department of Physics, Indian Institute of Science Education and Research, Dr. Homi Bhabha Road, Pune 411008, India (Dated: February 15, 2021) Chase-escape percolation is a variation of the standard epidemic spread models. In this model, each site can be in one of three states: unoccupied, occupied by a single prey, or occupied by a single predator. Prey particles spread to neighboring empty sites at rate p, and predator particles spread only to neighboring sites occupied by prey particles at rate 1, killing the prey particle that existed at that site. It was found that the prey can survive with non-zero probability, if p > pc with pc < 1. Using Monte Carlo simulations on the square lattice, we estimate the value of pc = 0.49451±0.00001, and the critical exponents are consistent with the undirected percolation universality class. We define a discrete-time parallel-update version of the model, which brings out the relation between chase- escape and undirected bond percolation. For all p < pc in D-dimensions, the number of predators in the absorbing configuration has a stretched-exponential distribution in contrast to the exponential distribution in the standard percolation theory. We also study the problem starting from the line initial condition with predator particles on all lattice points of the line y = 0 and prey particles on the line y = 1. In this case, for pc < p < 1, the center of mass of the fluctuating prey and predator fronts travel at the same speed. This speed is strictly smaller than the speed of an Eden front with the same value of p, but with no predators. At p = 1, the fronts undergo a depinning transition. The fluctuations of the front follow Kardar-Parisi-Zhang scaling both above and below this depinning transition.

I. INTRODUCTION to better understand the course of the pandemic, and study the effects of different intervention strategies [26]. Growth and spreading processes are ubiquitous, e.g. An important issue in ecology and conservation stud- growing crystals from supersaturated solutions [1, 2], di- ies is understanding how much predation would result in electric breakdown [3, 4], bacterial colonies [5], growth of prey populations becoming extinct. Similarly, on social tumors [6, 7] and spreading of rumours in social networks networks, one would like to know how efficient the rumor- [8–10]. Spreading processes and stochastic growth mod- scotching process has to be, in order to stop the spread els have attracted a lot of attention in statistical physics, of misinformation and rumors [8]. A simple model of owing to their wide applicability and their beauty [11– this, called Chase-Escape percolation was introduced re- 13, 18]. Simple mathematical models of these have pro- cently, where it was shown that indeed a slowly growing vided us with much insight about their universal critical prey population can coexist with relatively fast grow- behavior, and more generally into nonequilibrium phe- ing predators. This is the subject of this paper. This nomena [19–22]. One of the earliest of such models was model was first studied in [27] as a prey-predator model the Eden model of tumour growth [12]. Many ques- on trees where the prey tries to escape predators that are tions about this model have been studied, such as the chasing it. Later, it was extended to lattices and graphs velocity of its growth in different directions, the asymp- with cycles where the model displays very rich features totic shape [14–16, 23], and fluctuations of the inter- [29, 31]. It was shown in Ref. [29] that Chase-Escape face [18]. Several variants have been studied, starting (CE) percolation on the 2D square lattice undergoes a from the model of skin cancer by Williams and Bjerknes survival-extinction phase transition for the prey, and in- [24] to the SIR (Susceptible-Infected-Removed) and SIS terestingly, the critical point for this phase transition was (Susceptible-Infected-Susceptible) models of epidemics arXiv:2010.05310v2 [cond-mat.stat-mech] 12 Feb 2021 shown to be very close to 1/2 [29], the well-known critical [25]. The importance of such modelling has been under- threshold of bond percolation on the square lattice [30]. scored during the current COVID-19 pandemic, where In fact, in [31], it was suggested that the critical value for specialized SIR-type models have been extensively used CE percolation on the square lattice may also be exactly 1/2. While there is no real argument in favor of the criti- ∗ [email protected] cal value for CE percolation on the square lattice being † [email protected] exactly 1/2, it suggests a subtle connection to the well- ‡ [email protected] studied (Bernoulli) bond percolation process (hereafter 2 referred to as the ‘standard’ percolation process). One ing configuration is greater than s is bounded from below of the aims of our study was to determine the critical by exp(−Kp−1s1/d), where K is some p-independent con- threshold in this model more precisely numerically by stant. This is in contrast to the usual percolation prob- simulations. We find pc = 0.49451 ± 0.00001, which rules lem, where such distributions decay exponentially with out an exact correspondence between these two problems. s, for large s. In Section VII, we simulate clusters start- However, we will describe a discrete-time two-parameter ing with point seeds, which allows us to estimate critical generalization of the CE process, which brings out the exponents at p = pc. Also, from these and by means connection to the standard percolation process. We fur- of growth starting from tilted line seeds, we study if the ther show that the entire subcritical phase has a stretched critical threshold is direction dependent, as is known to exponential decay of the cluster size distribution, as com- happen in the well-studied cases of directed percolation pared to the exponential distribution obtained in the case and Eden growth. We present evidence that it is not, of standard percolation. Nevertheless we also show that again in agreement with standard undirected percolation. the critical behavior is completely consistent with being Section VIII is devoted to the second phase transition in in the standard percolation universality class. In par- this model, which is like a pinning-depinning transition ticular, the correlation length exponent ν obtained from of the predator and prey fronts, and occurs at p = 1, a finite size scaling analysis and dynamical exponents but has not been discussed much earlier. For p < 1, the obtained from cluster growth exactly at pc give values prey and predator fronts move together at the same ve- in complete agreement with the known 2D percolation locity. Our simulations show that these pinned-together values, and, using two different methods, we show that fronts travel slower than the Eden front moving at the rotational invariance is recovered at the critical point pc. same rate p, in contrast with the results of Owen and Lewis for some continuum descriptions of prey-predator In the range pc < p < 1, we find that the velocity of the prey front is less than what it would be in the ab- models [32]. We identify the mechanism for this slowing sence of predators. This is different from the behavior down of the prey front, and verify it using numerical sim- found in some continuum predator-prey evolution mod- ulations. Finally, we use finite-size scaling to show that els, where predators are unable to affect the front speed both above and below the depinning transition, moving [32]. We explain the difference in terms of the discrete fronts are in the KPZ universality class. In Section IX, nature of the front, which is not well-captured in the we summarize our results and discuss some directions for continuum description. Finally, we study the depinning further work. transition between the predator and prey fronts at p = 1, and estimate the roughness and dynamical exponents to be α = 1/2, z = 1.5 in the entire range p > pc, which is II. DEFINITION OF THE MODEL consistent with the KPZ universality class. The plan of the paper is as follows: in the next section, CE percolation on D-dimensional hypercubical lattices we define the model precisely, and review earlier results. is defined as follows: each site can either be occupied by a In Section III, we discuss our Monte Carlo simulation predator species particle (denoted by blue color), a prey procedure and the analysis – based on ‘seeds’ where we species particle (denoted by red color) or it can be va- start with entire lines of prey and predators – which al- cant. In this paper, we shall consider two types of initial lows us to estimate the critical value pc. Finite size scal- conditions: (i) Point seeds, i.e. at time t = 0, the ori- ing analysis for different lattice sizes is used to determine gin O is occupied by a blue particle. All the neighbors the critical exponent ν corresponding to the divergence of of the origin are red, and all other sites are vacant. (ii) correlation length away from pc. Simulations with point Line seeds, i.e. an entire line (on a lattice with peri- seeds are discussed in Sec. IV, which give an even more odic boundary conditions) is blue at t = 0, and the line precise estimate of pc and estimates for critical exponents immediately above it is red. The evolution is a continu- at p = pc. In Sec. V, we discuss a discrete-time version ous time Markov process in which a red particle can give of the model, and determine the qualitative phase di- birth to another red particle at a vacant neighbouring site agram. We find that one end-point of the line of CE at rate p, and a blue particle can eat up a red particle transitions is exactly the standard bond percolation crit- at a neighbouring site at rate 1. When a predator eats ical point on the square lattice. If the critical behavior is up a prey particle on a neighbouring site, it gives birth the same all along this line, this implies that it is in the to a predator at the site earlier occupied by the eaten percolation universality class. Some other phase bound- prey. Note that the preys (the reds) can only reproduce aries corresponding to pinning-depinning transitions in at neighbouring vacant sites, and predators (the blues) this generalized model, that are in the directed perco- reproduce only by eating a prey at a nearest neighbour lation universality class are also determined exactly. In site. Note also that neither prey nor predators can die Section VI, we discuss, for the general D-dimensional CE spontaneously. If all the prey particles are eaten up, the percolation, the subcritical regime with p < pc, where all predators cannot grow, and since they do not die, the the prey particles eventually die out and the system en- system goes into one of its many absorbing states. If ters one of its many absorbing states. We argue that the there are no predators initially, then the model reduces probability that of the number of predators in the absorb- to the Eden model. 3

250 (a) 1 200 0.9 0.8 150 0.7 0.6

100 0.5 0.4 L=256

percolation prob. 0.3 50 L=512 0.2 L=1024 25 L=2048 0 50 100 150 200 250 0.1 L=4096 L=8192 0 0.48 0.485 0.49 0.495 0.5 0.505 0.51 p

FIG. 2. Survival (‘percolation’) probability as a function of p is plotted for L × L lattices, with L ranging from 256 to 8192. Observe the confluence of curves away from p = 0.5. From these simulations, the critical probability is determined to be 0.4945 ± 0.0001.

already mentioned, these authors noted that in this case, FIG. 1. Snapshots of realizations of the CE process with dif- pc is very close to 1/2. ferent initial conditions, with blue representing sites occupied by the predators, and red, by prey. (a) for p = 0.495 on a 250 × 250 square lattice with a point initial condition (b) III. IMPROVED DETERMINATION OF a close-up view of the front with a line initial condition at CRITICAL PROBABILITY pc: LINE SEEDS p = 0.65 for L = 500 at time 500. We used Monte Carlo simulations to determine more precisely the critical value pc on the square lattice. In the In Fig. 1, we show results of a typical evolution in the first set of runs, we simulated the process on an L×L lat- CE model, starting from a point initial condition in panel tice with periodic boundary conditions along the x−axis (a), and starting from a line initial condition in panel (b). and initial conditions such that each site on the line y = 0 This model has been studied on k−ary trees with ini- is occupied by a predator particle and each site on the tial conditions such that the predator is located at the line y = 1 by a prey particle. With this ‘line’ initial con- root and the prey occupy the k daughters of the root dition, the boundaries of the clusters are statistically sta- [27, 28]. It was shown that the critical value for coexis- tionary for translations along the x-direction, and there tence in this case is given by is greater self-averaging in the evolving clusters, and one avoids the very slowly-relaxing fluctuations in point seed p 2 pc(k) = 2k − 1 − 2 k − k. (1) simulations. In a simulation on a L×L lattice, we say that the prey 1 Note that in large k limit, this gives pc ∼ 4k . Clearly, it survives (or ‘percolates’) if any of the prey particles are is seen that the prey can survive on such trees even if it able to reach the top boundary y = L. Survival prob- moves at a much slower rate than the predators and that ability is then calculated by performing the simulation the critical value pc goes down as k increases. The CE multiple times and computing the fraction of realizations model was also studied on the ladder graph where the in which the prey survives. We used between 50000 and 5 critical rate pc is 1 [31]. 3 × 10 realizations to estimate the survival probability The CE model shows rich behaviour on 2D lattices and for each value of p and L. The plot of survival probabil- it was numerically explored in Ref. [29]. They showed ity as function of p, for L ranging from 256 to 8192. is convincingly that even on the square lattice, coexistence shown in Fig. 2. between the prey and predator is possible when the rate We see that the curves for different L seem to meet at of spreading of prey p is strictly less than that of the a common point at p ≈ 0.4945. From finite-size scaling predator. This is expected as the average number of va- theory, this point is identified as the critical probability. cant neighbors per prey particle is greater than the cor- From the spread of the region of intersection, we estimate responding number of prey neighbors of a predator. As the error bar in this estimate to be 0.0001. We thus get 4

IV. SPREADING SIMULATIONS FROM POINT SEEDS 1 0.9 Simulations starting from point initial conditions for 0.8 which we observe percolation to one or more than one 0.7 boundary lines would give less precise results. Much more 0.6 precise results are obtained if we observe the growth as a 0.5 function of time and stop it, before the boundaries of the 0.4 L=256 lattice are reached. This is indeed the preferred strategy 0.3 L=512 percolation prob. for directed percolation [33], but it also gives very good L=1024 0.2 L=2048 results for standard percolation, both in low and high di- 0.1 L=4096 mensions [34]. In the present simulations we used lattices L=8192 0 of size up to 32768 × 32768. -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 In addition, one can simulate in this way also subcriti- 3/4 (p-pc)*L , pc = 0.4945 cal CE percolation, if one uses p sufficiently small so that all clusters die before they reach the lattice boundary. Mean cluster sizes obtained in the latter way, i.e. aver- age number of predators as t → ∞ for p < pc, are plotted FIG. 3. Scaling collapse for the survival probability is plotted as a function of pc − p in Fig. 4. For p near pc, we expect 1/ν −γ against (p − pc)L , with ν = 1.333 and pc = 0.4945. that the average cluster size would scale as (pc − p) . We find a good agreement with γ = 43/18, which is the critical exponent for standard 2D percolation. By finding the value of p that gives the best slope, this also allows us to refine our estimate of pc to be 0.49451 ± 0.00002. At the critical point one expects that the prey survival probability P (t), which is the probability that there ex- ists at least one prey particle at time t, scales as

P (t) ∼ t−δ, (2)

while the average number of prey particles (averaged over all runs, even those which had already died) and the av- erage r.m.s. distances of prey particles from the origin scale as

µ 1/z Nred(t) ∼ t ,R(t) ∼ t . (3)

In order to test these we plot, for each of the three ob- servables X = P,Nred, and R and for several values of p α near pc, the ratios X(t)t against t, with suitably cho- sen exponents α. If α = −δ, µ, and 1/z, respectively, we then expect the curves for p = pc to be asymptotically FIG. 4. Average number of predators in the absorbing state horizontal. Such plots are shown in Fig. 5. They show as a function of pc − p. Near pc, the cluster size scales as again that CE percolation is in the standard percolation −γ (pc − p) with γ = 43/18, and pc = 0.49451. universality class, and linear interpolation of the slopes in all three panels of Fig. 5 to get the p value corre- sponding to zero slope give our best and final estimate p = 0.49451(1). estimate for the critical probability to be pc = 0.4945 ± 0.0001. Thus, our simulations clearly show that pc is V. THE DISCRETE-TIME PARALLEL-UPDATE 1 VARIATION strictly less than 2 .

In Figure 3, we show the results of a finite-size scaling We define below a discrete-time parallel-update vari- analysis of the data. We plot the survival probability ation of the CE percolation, which brings out its rela- 1/ν against (p − pc)L , and obtain a good data collapse for tionship to the standard percolation. In this variation, the values pc = 0.4945 and ν = 1.333. This numerical as before, there are three states per site: unoccupied, red estimate of ν is in excellent agreement with its value in or blue. The time evolution is discrete. It is specified in the standard 2D percolation universality class. terms of two parameters p1 and p2. In one time step, a 5

0.86 p = 0.4947 0.84 p = 0.4946 p = 0.4945 0.82 p = 0.4944 p = 0.4943 0.8

0.09212 0.78

0.76 P(t) x t 0.74 (a) 0.72

0.7 1 10 100 1000 10000 t

8.5 p = 0.4947 p = 0.4946 8 p = 0.4945 p = 0.4944 7.5 p = 0.4943

7 0.584466

6.5 FIG. 6. A schematic depicting different phases in the pa- (t) / t

red rameter space (p1, p2) of the generalized Chase-Escape model N 6 with parallel updates. The parameter space is divided into (b) six phases, where Q denotes the absorbing phase, and Q to 5.5 0 1 Q5 denote the five different active phases discussed text. 5 1 10 100 1000 10000 t red site invades every empty neighboring site with proba- 3.2 bility p1, and a blue site invades red sites with probability p = 0.4947 p . If a site that is unoccupied, but has r red neighbors 3 p = 0.4946 2 p = 0.4945 at time t, it thus becomes red at time (t + 1) with proba- 2.8 p = 0.4944 bility [1 − (1 − p )r], and remains unoccupied otherwise. p = 0.4943 1 2.6 A red site with r0 blue neighbors at time t, becomes blue 2.4 r0 (c) at time (t + 1) with probability [1 − (1 − p2) ]. A blue 0.88435 2.2 site at time t remains blue at all subsequent times. 2 The parameter space here is (p1, p2), both of which lie R(t) / t 1.8 in the interval [0, 1]. This is shown as a square ABCD 1.6 in Fig. 6. It is easy to determine the behavior of the 1.4 model along the boundaries of this square, say starting with point initial conditions. Also, the continuous time 1.2 1 10 100 1000 10000 model studied in previous sections is recovered by setting t p1 = pδt, p2 = δt, in the limit δt tends to zero. Along the line AD, we have p1 = 0, and the system will reach an absorbing state, with exactly five blue sites, for all p2 > 0. Along the p2 = 0 line AB, the blue cluster does

FIG. 5. Log-linear plots of growth observables X = P,Nred, not grow, and the red cluster grows as in the discrete- and R at fixed values of p ≈ pc, plotted against time. For time Eden model (also called the Richardson model [36]). better visual discrimination, the actual variables plotted are The radius of the red cluster increases linearly with time. α t X(t) with α chosen such that the curves for p = pc be- For p1 = 1 ( the line BC), the red cluster grows deter- come flat asymptotically. Panel (a): X(t) = P (t), α = −δ = ministically, as a diamond, with the maximum velocity −0.09212; panel (b): X(t) = Nred(t), α = µ = 0.584466; 1. All the perimeter sites of the growing blue cluster are panel (c): X(t) = R(t), α = 1/z = 0.88435. The values cho- red, and it grows as the Richardson model with growth sen for δ, µ, and z are indeed those for standard 2D percolation parameter p . [35]. The fact that in all three panels the curves become flat 2 for p = 0.49451(1) shows convincingly that CE percolation Along the line CD, with p2 = 1, it is easily seen that at is in the standard percolation universality class, and gives at any time, a red site has always at least one blue neighbor. the same time the best estimate of pc. Thus it survives for only one unit of time. We assign to each bond i.i.d. random variables ‘occupied’ with proba- bility p1, and ’unoccupied’ with probability (1 − p1). We 6

FIG. 7. Snapshots of realizations of the generalized discrete time-parallel update CE process with point initial conditions for the six different phases at selected representative points. make a red site at time t blue at time (t + 1) iff it had a and the red front moves faster than the blue one in all di- blue neighbor at time t connected by an occupied bond. rections. In Q3, the fronts are also fully depinned, but for Then clearly, the stochastic evolution of this model is as p1 > pc,DP , where pc,DP is the critical probability for di- defined above. Also, the probability that the blue sites rected percolation, the red front has four linear segments, eventually constitute a finite cluster C is the same as the where the front velocity, in the L1 norm, reaches the max- probability of that cluster C in standard bond percolation imum possible value 1. For the square lattice, pc,DP is with occupied bond density p2. known rather accurately: pc,DP ≈ 0.644700185(5) [38]. In region Q , the blue front also reaches the maximum Since the percolation probability for bond percolation 4 possible velocity in some directions, and then in these on the square lattice is well-known to be 1/2, we see directions, the red and blue fronts are again pinned to- that the point E, which marks the end of the absorbing gether, and in Q , the fronts are fully pinned again. phase in Fig. 6 has coordinates (1/2, 1), and the critical 5 Clearly, the boundary between Q and Q is a vertical exponents at this point are the same as in the standard 2 3 line, with equation p = p , whereas Q and Q are bond percolation. If we assume that critical behavior is 1 c,DP 3 4 separated by the line p = p . The line p = p forms same all along the line AE, as seems reasonable, then 2 c,DP 1 2 the boundary between Q and Q , and also between Q the universality class of CE-percolation is the same as 1 2 4 and Q . standard percolation. Also, the line AE has average slope 5 2, and the fact our estimated pc in the continuum case was near 1/2 indicates that its slope differs from 2 only by small amount all along the curve. VI. LOWER BOUND TO THE CLUSTER SIZE IN THE ABSORBING STATE The region in the (p1, p2) plane to the right of line AE is the region where prey can survive indefinitely with a non-zero probability. Using the result of Durrett and Consider the CE percolation on the D-dimensional hy- Liggett about the existence of linear segments on the lim- percubical lattice starting from point seed initial condi- iting front shape [37], we can divide this region into five tions (a single predator at the origin, surrounded by 2D sub-regions, denoted by Q1 to Q5 in Fig. 6. These are prey particles on its neighbouring sites). We consider defined as follows: In Q1, both red and blue fronts are the case p < pc, when the system eventually reaches one pinned together in all directions. The common velocity of its absorbing states and we are left with a cluster of of the front is weakly direction dependent, but no front predators. has a linear part. In Q2, the fronts are fully depinned, Let CumProb(s) be the cumulative probability that, in 7

percolation, where CumProb(s) decreases exponentially with s for large s, for values of p below the critical thresh- old. To accurately capture the tail of the subcritical clus- ter size distribution in numerical simulations, we adopt a rare event sampling approach, where we delay any action of the first predator by a holding time T , for various val- ues of T , and let the prey grow freely till then, without facing any predation. After time T , the process follows its usual evolution and we note down the final cluster size and assign it a probability weight of exp[−4T ]. In Fig. 6, we plot the mean size of the final predator cluster as a function of the holding time T , for p = 0.1, 0.2, 0.3 and 0.4, and to these plots, we fit a function y = aT 2 + bT . The fit shows very good agreement with our data. For FIG. 8. Mean cluster size of predators in the absorbing state p = 0.2, the best values of a and b are 1.13 and 15.98 of CE percolation as a function of holding times T , for p = respectively, whereas for p = 0.4, we have a = 7.96 and 0.1, 0.2, 0.3 and 0.4. To each plot, we fit a function Y = b = 124.68. This suggests that clear indications of the 2 aT + bT and obtain an excellent fit. stretched exponential tail would come at large cluster sizes (corresponding to large holding times). In Fig. 9, we plot the full cumulative distribution, as defined in Eq. (4), of the cluster size distribution for p = 0.1, 0.2 and 0.3 using the importance sampling method. To ob- tain the tail behaviour and sample the full distribution, we plot the weighted distribution obtained by merging the cluster size distributions for holding times T rang- ing from 1 to 10 in spaces of 1 unit time, and 20 to 100, in steps of 10 units of time. For each holding time, we perform 20000 realizations to obtain the cluster size distribution. The plot shows a clear deviation from expo- nential behaviour, and provides evidence in favor of our stretched exponential lower bound. This stretched exponential decay seems to be inconsis- tent with the exponential decay for standard percolation. If CE percolation is in the universality class of the usual FIG. 9. Cumulative probability as defined in Eq. (4) for 2D percolation problem, then not only the critical expo- p = 0.1, 0.2 and 0.3, computed using the rare event sampling nents, but also the scaling functions in the critical region technique as described in the text. The x-axis is linear, where as the y-axis is plotted on logarithmic scale. The plot shows should be universal. In the standard percolation, the a clear deviation from exponential behaviour, and provides probability distribution of cluster size s has the scaling evidence in favour of our stretched exponential lower bound. form

−τ φ the absorbing state, the number of predators is greater Prob(s) ∼ s f(s ), (5) than s. It is clear that this probability will be bounded from below by the probability that the prey population φ as  → 0, with s fixed. Here,  = pc − p, and f(x) is reaches s before the predator present at the origin eats known to decay as exp(−x), for large x. However, we can up its first prey. The probability that the predator at the have origin does not eat any of its adjacent prey up to time T is exp(−2DT ), but in that much time, the number of D −τ φ 1/D prey particles grows to order of (pT ) . It is clear that Prob(s) ∼ s exp(−B1s ) + exp(−B2s ). (6) for the prey population to reach size s, the time required would be of order s1/D/p. This immediately gives us Clearly, the stretched exponential decay of cluster size  C s1/D  distribution dominates only for s > −φD/(D−1), but then CumProb(s) ≥ C exp − 2 (4) 1 p the scaling variable x in the scaling function would tend to infinity. Thus, the asymptotic stretched exponential where C1 and C2 are constants, which may be chosen decay of Prob(s) for large s is consistent with the expo- to be independent of p. This is in contrast to standard nential decay of the scaling function f(x) for large x. 8

VII. DIRECTION DEPENDENCE OF SURVIVAL PROBABILITY 1 The qualitative reason why the faster predators can 0.9 have the same front velocity as the prey in the regime 0.8 p ≤ pc (and, as we shall see later, also for pc < p ≤ 1) 0.7 is clearly due to the fact that the predators can only 0.6 move along the sites occupied by the prey. But the paths 0.5 created by the prey along different directions are quali- 0.4 L=256 tatively different. This is well known in the Eden model, L=512 which is just the model where the prey spreads without percolation prob. 0.3 L=1024 interference of predators [14, 15]. The first hints that 0.2 L=2048 L=4096 the spreading velocities in the 2D Eden model are dif- 0.1 L=8192 L=16384 ferent along the diagonal and parallel to the axes were 0 obtained in [17], while the most precise simulations, by 0.49 0.492 0.494 0.496 0.498 0.5 Alm and Deijfen, show that the velocity along the diag- p onal is smaller than the velocity along one of the axes by roughly 1.3% [40]. Also, in most of the regions of the (p1, p2) plane of FIG. 10. Survival (‘percolation’) probability as a function of p the discrete-time problem, the velocities of front are di- is plotted for L×L lattices, both with the original orientation rection dependent. The rotational invariance is expected (data are from Fig. 2) and for lattices tilted by 45 degrees to emerge only at the critical point, in the scaling limit. so that the spreading occurs effectively along one of the two Clearly, the geometrical structures of paths with different diagonals of the untilted lattice. For each pair of curves with average orientations are somewhat different. If the front the same color, the shorter and less steep one is for the tilted velocity in different directions is different, even the crit- lattice. ical thresholds could be different in different directions. We recall that in the well-studied case of directed per- Sec. III. colation, there is a orientation dependence of the critical Results are shown in Fig. 10, together with those from threshold p (θ), and infinite directed paths in the direc- c Fig. 2. We see that both sets of curves cross (in the limit tion θ appear only if p > p (θ). c of large L) at the same value of p. Thus p is indeed Since the isotropic scaling limit is an important pre- c independent of the orientation. We can also try a data diction of the conformal field theory that gives the exact collapse in the way of Fig.3. When doing this, we just values of critical exponents of the percolation theory, it is have to take into account that the process now takes useful to have a direct test of the restoration of isotropy, effectively place on a tilted square lattice with lattice at the critical point, in the scaling limit. Rigorous proofs √ constant√ being 2, thus we should use on the x-axis (p − of this are not easy, but at least for bond percolation and 3/4 several other models such a proof was given recently in pc)[L/ 2] . This is shown in Fig. 11. We see that Ref. [39]. In the following we shall present two different the scaling curves in the two different directions show sets of simulations for directly verifying this prediction. excellent collapse, verifying that the scaling function is independent of direction.

A. Line seed spreading from a tilted line B. Anisotropy of fourth order moments in clusters grown from point seeds In Sec. III we had measured pc by observing the spread- ing from a line seed that was oriented parallel to the For any radially symmetric distribution ρ(x, y) one has x-axis. In complete analogy, we can also use a line seed parallel to one of the two diagonals. Equivalently, we can Z 1 Z dxdyρ(x, y)x2y2 = dxdyρ(x, y)x4 . (7) replace the neighborhood {(x±1, y), (x, y±1)} of any site 3 (x, y) by the neighborhood {(x ± 1, y ± −1), (x ± −1, y ± 1)}, and start again from an initial condition where even Therefore, if there were no lattice anisotropy, clusters sites on the x-axis are occupied by predators, and odd grown from standard 2D percolation, from CE percola- sites on the line above by prey. Again we use periodic tion, and from the Eden model would all satisfy boundary conditions laterally, and measure the fraction 1 of percolating runs where we say that a run percolates 2hx2y2i = hx4 + y4i . (8) 3 if at least one prey particle reaches the top line of an L × L lattice. Notice that now only half of all lattice Any deviation from this must be a consequence of lattice sites (where x + y is even) participate in the dynamics, anisotropy. For short times, this anisotropy is of course thus we can simulate somewhat larger lattices than in unavoidable, but exactly at the critical point, one expects 9

that isotropy is restored at large length scales and Eq. (8) holds in the limit t → ∞. 1 In Fig. 12 we plotted the ratio 2hx2y2i/hx4+y4i against 0.9 time t, for several models: The Eden model (with p equal 0.8 to pc of CE percolation), critical bond percolation, and 0.7 CE percolation. For the latter we show both moments of 0.6 prey and moments of all (prey + predator) particles. The 0.5 latter two are indistinguishable on this plot. The Eden 2 2 4 4 0.4 model is clearly anisotropic (with 2hx y i/hx + y i → 0.3246(1) for t → ∞), but the others are clearly isotropic. 0.3 percolation prob. The rate of convergence towards 1/3 seems to be different 0.2 L=8192, axis-aligned from that in standard percolation, but this could also be 0.1 L=8192, diagonal L=16384, diagonal a consequence of the fact that we needed to start from a 0 larger seed in the latter. -2 -1.5 -1 -0.5 0 0.5 1 1.5 3/4 √ 3/4 (p-pc)*L resp. (p-pc)*[L/ 2] , pc = 0.4945

VIII. DYNAMICS OF THE FRONTS FOR p > pc

FIG. 11. Testing the rotational invariance at the critical A picture of the interface for p = 0.65 was shown point: The scaling function for the survival probability as in Fig. 1(b). We can see that the red sites form a 3/4 a function of the normalized scaling variable (p − pc)L for rather large number of disconnected clusters. The propagation along the axis shown for two values of L, col- front does become smoother for larger p. There are lapses to that calculated along the diagonal, if the diagonal several possibilities to define such fronts and their value of the lattice size L ( which√ is an integer, measured positions precisely. For our discussion it is convenient in lattice units) is scaled by 1/ 2. Curves for other L are to work with single-valued height functions h (x, t) omitted for visual clarity. red and hblue(x, t), called the red and blue front heights at . time t, that specify the largest y−coordinate among all sites ever reached with that value of x−coordinate by that color up to time t. We will study here the mean values and variances of hred(x, t) and hblue(x, t), and their dependence on t. Note that if for some x, there is no red site at time t, then that particular value of x is disregarded during the computation of the height 0.34 statistics. Alternatively, we can – instead of defining > 4 explicitly the prey front and the corresponding hred(x, t) 0.32

+ y – define only the predator front and the front of all 4 non-white sites, with heights hblue(x, t) and htot(x, t) 0.3 respectively. In this way we do not have to discard any values of x. It is easy to see that for p > 1, the mean > / < x 2 0.28

y distance w(t) between the prey and predator fronts 2 1/3 increases linearly with time, as does also the average of chase-escape, prey only δh(x, t) = h (x, t) − h (x, t) and the density of red 2 < x 0.26 chase-escape, prey + predator tot blue bond percolation sites ρred(t) per unit x. For p < pc, the system goes into Eden 0.24 one of its absorbing states and for large t, ρred(t) and 1 10 100 1000 10000 δh(x, t) are exactly zero. Time We studied the dynamics of the prey and predator fronts by numerical simulations on a lattices of dimen- FIG. 12. Log-linear plot of the ratio 2hx2y2i/hx4 + y4i of sions up to 65536 × 32678 with periodic boundary condi- fourth moments, for clusters grown from point seeds. For tions along the x−axis, unless explicitly stated otherwise. isotropic clusters this ratio is 1/3, for any radial distribu- For pc < p < 1, we find that the centre of mass of the tion. For the Eden model this ratio is clearly smaller than prey and predator fronts move together at the same ve- 1/3, showing again its anisotropy. For the other three cases locity. In this regime, the two fronts are pinned together (bond percolation, prey in critical CE percolation, and prey and at p = 1, a depinning transition [41, 42] occurs. + predators in critical CE percolation) isotropy is violated for In Fig. 13, we plot the average velocities of the prey small times, but is restored in the limit t → ∞ and predator fronts (measured over a sliding window of . size 100 units) at p = 0.6, and also of the Eden front using the same value of p. The dashed green line is 10

5000 1.5 Eden Chase-Escape 4000

p=0.8 p=1.0 3000 1.4 p=0.6 avg Y 2.5 Velocity 2000 2

Slope 1.5 Eden 1000 1.3 Chase-Escape (Red) 1 0.5 0.6 0.7 0.8 0.9 1 Chase-Escape (Blue) p 0 0 1000 2000 3000 4000 0 1000 2000 3000 4000 5000 6000 Time Time

FIG. 13. Average velocities for p = 0.6 of the three fronts over FIG. 14. CE and Eden fronts at values p = 0.6, 0.8 and 1. We a sliding window of 100 units as a function of time, to check can see that the Eden and CE fronts coincide at p = 1. As for convergence of velocity. The dashed green line depicts the pointed out in Fig. 13, the CE front is significantly slower than Eden velocity from the simulations of Alm and Deijfen [40]. the Eden front at p = 0.6. The inset is a plot of the velocities of the CE and Eden fronts as a function of p. Clearly, the velocity of the CE front in the pinned regime is not linear in the velocity of the Eden front as obtained by Alm and p. Deijfen [40]. Our simulation of the Eden front is in agreement with their result. It is clear that the front velocity in CE is strictly less than what it would be if predators were absent.

This is somewhat surprising. It was shown by Owen and Lewis [32] in a rather general continuum model of predator-prey dynamics that the prey front velocity is unaffected by the presence of a predator, except if some special conditions are satisfied. The dynamics of Owen and Lewis model is described by the equations ∂u ∂2u = D 2 + ruf(u) − φvh(u) ∂t ∂x (9) ∂v ∂2v = D + γvh(u) − δv ∂t ∂x2 FIG. 15. Two trajectories of the maximum extent of infection where u and v denote the population densities of prey at p = 0.6 for a single realization of the Eden front and the and predators respectively and f(u) and h(u) are arbi- red front in CE percolation are plotted. The inset shows a trary positive functions [32]. Here f(u) is a effective re- zoomed-in picture of the trajectory of the CE front, where production rate per individual. Usually, we expect f(u) the maximum position of a red infection is seen to drop quite to be a decreasing function of u. However, sometimes, often. for very small u, the reproduction rate may decrease, say df(u) because of difficulty of finding a mate. If du > 0 in some range of u, and an increase in density of prey leads that the velocity of the fronts in the pinned regime is not to an increase in reproduction rate per individual, this linear in p. Anyhow, for p ≤ pc the CE fronts cannot is called the Allee effect. Owen and Lewis showed that, propagate any more, and thus the front speed must con- if the dynamics is described by Eq. (9), and there is no verge to zero for p → pc. What could be the reason for Allee effect, the velocity of prey front is not affected by the this discrepancy, and the inapplicability of the Owen the presence of predators. For the CE problem, increas- and Lewis analysis for CE percolation? ing density of prey can not increase the reproductive rate, A clear difference between the CE and Eden fronts is and there is no Allee effect. Hence, the surprise. that the CE front consists of inactive parts (sites occu- Figure 14 is a plot of CE and Eden fronts at values pied by predators) that can only grow if there is a prey p = 0.6, 0.8 and 1. We can see that the centres of mass of adjacent to it. In the CE front, it is also possible that the Eden and CE fronts coincide at p = 1, and move at the prey at a local maximum of the front is eaten up by a the same velocity. However, at p = 0.6, the CE front is predator, and the front thereby recedes. This contributes significantly slower. The inset of Fig. 14 further reveals to the CE front velocity becoming less than the Eden 11

290 1090

1088 288 1086

286 1084

1082 10 284 1080 198 200 202 204 786 788 790 792 794 = 1/3 max β Y 3504 3624

3502 3622 (t) ,

3500 3620 prey ρ

β - 3498 3618 t 1 3496 3616 2580 2585 2674 2676 2678 2680 Time -6 -4 -2 0 2 4 6 ν t1/ t (p-1) , ν = 3/2 FIG. 16. Four instances from the trajectory in Fig. 15 where t there is a drop in the maximum position of the red front.

FIG. 18. Scaling collapse for the data shown in Fig. 17. We 1 β 1 νt use the scaling ansatz ρred(t) = t G[(p − 1)t ] where β = 3 . 3 p = 1.1 This suggests that νt = 2 , and we demonstrate a good scaling p = 1.05 collapse with these exponents. To reduce small-t corrections, p = 1.02 only data for t ≥ 7 are plotted. p = 1.01 p = 1.005 p = 1.0 100 p = 0.995 p = 0.99

(t) is no rightmost particle. But in the discrete model there p = 0.98

prey p = 0.95 is, and there is a non-zero probability that this gets eaten ρ p = 0.9 up by the predators. These events lead to an overall de- crease in the velocity of the prey front. Thus the discrete 10 nature of the fluctuations of the front leads to a macro- scopic observable consequence, which has also been noted in the context of the Fisher-KPP fronts [43]. The pres- ence of predators leads to a further decrease in the front 10 100 1000 10000 velocity. The continuum description fails to capture this t effect. In regions Q2, Q3 and Q4 in Fig. 6, where the predator and prey fronts are far from each other, the Owen and Lewis analysis remains valid. FIG. 17. Density of red sites, ρred(t), per unit x, for values In Fig. 17, we plot the mean density of prey ρred(t) of p ranging from 0.9 to 1.1. For pc < p < 1, ρred(t) tends to per unit x, as a function of time, for the values ranging a constant, and for p > 1, it grows linearly with time. At the 0.333(3) from p = 0.9 to and p = 1.1. As expected, at p < 1, the critical point p = 1 (black curve), ρred(t) ∼ t . process is in the pinned regime and ρred(t) tends to a fi- nite constant. At p > 1, ρred(t) grows linearly with time. However, at the critical point p = 1, by continuity argu- front velocity. In Fig. 15, we plot the maximum y co- β ments, one would expect that ρred(t) ∼ t with β < 1 . ordinate of the red particles (Ymax) at p = 0.6 for Eden This expectation is verified in our numerical simulations, front and the prey front in CE. Inset shows a zoomed and we find β = 0.333(3). To further explore the critical in picture of the trajectory of the prey front where it is behavior near p = 1, we use a scaling ansatz for ρred(t) clearly visible that the maximum position of does drop of the form quite often. Figure 16 shows four other instances from β 1/νt the trajectory of the maximum prey front where the drop ρred(t) = t G[(p − 1)t ], (10) is clearly visible. At p = 0.6, the backward motion occurs roughly 1% of the time on a lattice of size 500 × 5000. with a scaling function G(z) which is a finite constant We believe that the reason for this is the discrete na- at z = 0, and has a power law scaling as z → ±∞. We ture of the model, which is not correctly captured in the know that, as t → ∞, ρred(t) tends to a constant for coarse-grained continuum equations of the type described pc < p < 1, and ρred(t) ∼ (p − 1)t for p > 1. This would 3 in Eq. (9). In the coarse-grained continuum description, thus give us that 1/νt = 1 − β, implying that νt = 2 . In u(x) has a rapidly decaying tail of the distribution for Fig. 18, we show a scaling collapse for p near 1, with the 1 3 large x, but u(x) is always non-zero at finite x, and there exponents β = 3 , and νt = 2 . 12

FIG. 19. Scaled variance as a function of time at p = 1. (Left) Scaling of the variance by L2α, with α = 1/2. leads to a good data collapse for the asymptotic variance of the prey front. (Right) Scaling of time by Lz, with dynamical exponent z = 1.5. The exponents α and z are consistent with the KPZ universality class.

Furthermore, for both regimes – p > 1 and p < 1, one distribution is not an exponentially-tailed distribution for would expect that the fluctuations in the CE fronts are any non-zero p. captured by KPZ theory, and the process falls in the KPZ For p just above pc, we studied the direction depen- universality class [19, 20]. For p > 1 this is certainly true, dence of the survival probabilities with the point initial because in this regime, fluctuations of both fronts scale condition, and our data is fully consistent with full ro- as interfaces of Eden clusters. tational invariance at the critical point. Our simulations For p = 1 this is not so obvious. We thus studied the show that in the regime pc < p < 1, the prey and preda- fluctuations of the CE front at the critical value p = 1. In tor fronts move with a common velocity, that is strictly Fig. 19, we plot the scaled variance of the prey front as a less than than the velocity prey would have if predators function of time for lattice sizes L×10000, with L taking were absent. We note that this slowing is explicitly due values ranging from 16 to 4096. The left panel is a plot to the predators, and in addition, is due to the decrease of the scaling of the variance by L2α, with α = 1/2 which in front velocity that occurs due to fluctuations caused leads to a good data collapse for the asymptotic variance by the discrete nature of particles, which is also not cor- of the prey front. For the dynamical scaling, we plot in rectly captured in the Owen-Lewis type partial differen- the right panel, the scaled variance of the red front, as tial evolution equations without noise. We suggest that a function of scaled time where time has been scaled by this effect depends crucially on the discrete nature of fluc- a factor of Lz with z = 1.5 ± 0.05. The exponents thus tuations of the advancing front. Sometimes the leading obtained are consistent with the KPZ universality class prey on the front get eaten by predators, and thus the (α = 1/2, z = 1.5), thus leading to the conclusion that maximum position of the prey front is a non-monotonic though the depinning transition takes place at p = 1, the function of time. critical exponents corresponding to the surface fluctua- We defined a generalization of the model with discrete tions are not affected by this transition. time, characterized by two parameters (p1, p2). In this generalization, the CE-transition is seen along a line in the (p1, p2) plane. One end point of the line is exactly IX. SUMMARY mapped to the standard bond-percolation problem, and the other end corresponds to the original CE-model. The In summary, we studied the CE percolation on the 2- critical behavior is expected to be in the standard perco- D square lattice. We estimated the critical probability lation universality class along the entire line. Looking for the survival-extinction phase transition in this model at the direction-dependent properties, we can identify from Monte Carlo simulations, and our final estimate is other critical lines in this generalized parameter space that show criticality corresponding the directed percola- that pc = 0.49451 ± 0.00001 and we can thus say with tion universality class. fair confidence that pc 6= 1/2. From the finite size scaling analysis near this transition, we find that the correlation There are several questions that still remain unan- length exponent ν is equal to the 2-D undirected perco- swered about CE percolation. It would be nice to extend lation value 4/3 within our error bars. We also point out the results about the ladder to all graphs with a finite that the subcritical phase of CE percolation differs from width, which are essentially one-dimensional, and prove the usual percolation problem, as in CE, the cluster size that here pc is always 1. Also, it would be nice to have a 13 rigorous proof that for a regular d-dimensional lattice pc where the predators can die with a small rate δ, then for CE percolation, pc is strictly less than 1. More gener- eventually, we have a unique absorbing state for p be- ∗ ally, determination of the exact front velocity or critical low some pc (possibly different from the pc determined ∗ probability for some non-trivial non-tree graph is an open in this paper) and for p > pc , we would have coexis- problem. For the generalized two-parameter model, sev- tence between the prey and predators. We would expect ∗ eral questions seem interesting to explore further. One the survival-extinction transition at this pc to be in the would like to understand why the phase boundary AE in universality class of (2+1) dimensional directed perco- Fig. 6 is nearly straight. Is it everywhere convex? The lation, according to the Janssen-Grassberger conjecture model shows rich interplay of percolation, directed per- [22, 45, 46]. Such a crossover is also seen in the SIRS colation, pinning-depinning and active-absorbing transi- model on the square lattice, where recovered nodes can tions, and seems interesting for further study. again become susceptible at a finite rate [47]. We also CE percolation is somewhat similar to the SIR model, expect a similar crossover if we allow the predators to which belongs to the same universality class as undirected diffuse on to empty sites at a small rate, as the extinc- percolation [44]. In the SIR model, each site is either Sus- tion of the prey would stop the predator population from ceptible, Infected or “Removed” (i.e., dead or recovered growing, and the existing predators would diffuse to infin- and immune). Infected sites can infect neighbouring sites ity. Developing a better understanding of this crossover that are susceptible and each infected site can die/recover seems like an interesting avenue for future research. independently. In some variants of this process, recovered sites become susceptible again with partial immunity (if they do not acquire any immunity, then the resulting SIS ACKNOWLEDGEMENTS model is in a different universality class). In CE perco- lation, the vacant sites can be thought of as susceptible, sites occupied by prey particles as infected and the ones Aanjaneya Kumar would like to acknowledge the occupied by predators as removed. The key difference Prime Minister’s Research Fellowship of the government between the two models however is that in the SIR pro- of India for financial support. cess, the death/recovery of each infected particle is spon- This paper is dedicated to the memory of Dietrich taneous and independent of others whereas in CE, the Stauffer, who was a friend, and collaborator of two of us. recovery can occur only at a site neighboring a recovered Both DD and PG fondly remember his wit and wisdom, site, and the predators always form a single connected his dedication to physics, and his ever helpful nature. It cluster. If one considers a variant of the CE problem, has been a privilege knowing him.

[1] Nancollas, G.H., 1979. The growth of crystals in solution. [11] J. M. Hammersley and D. J. A. Welsh, in Bernoulli 1713 Advances in Colloid and Interface Science 10, 215–252. Bayes 1763 Laplace 1813: Anniversary Volume, edited [2] Nielsen, A.E., 1969. Nucleation and Growth of Crystals by J. Neyman and L. M. Le Cam (Springer Berlin Hei- at High Supersaturation. Kristall und Technik 4, 17–38. delberg, Berlin, Heidelberg, 1965), pp. 61–110. [3] Li, X., Tung, C.H., Pey, K.L., 2008. The nature of dielec- [12] Eden, M., 1961. A Two-dimensional Growth Process. In tric breakdown. Appl. Phys. Lett. 93, 072903. Proceedings of the Fourth Berkeley Symposium on Math- [4] Niemeyer, L., Pietronero, L., Wiesmann, H.J., 1984. ematical Statistics and Probability, edited by F. Neyman Fractal Dimension of Dielectric Breakdown. Phys. Rev. (Univ. of California Press, Berkeley, 1961), Vol.IV, p.233 Lett. 52, 1033–1036. [13] Hammersley J.M., Welsh D.J.A., 1965. First-Passage [5] Allen, R.J., Waclaw, B., 2019. Bacterial growth: a sta- Percolation, Subadditive Processes, Stochastic Net- tistical physicist’s guide. Reports on Progress in Physics works, and Generalized Renewal Theory. In: Ney- 82, 016601. man J., Le Cam L.M. (eds) Bernoulli 1713, Bayes [6] Waclaw, B., Bozic, I., Pittman, M.E., Hruban, R.H., Vo- 1763, Laplace 1813. Springer, Berlin, Heidelberg. gelstein, B., Nowak, M.A., 2015. A spatial model predicts https://doi.org/10.1007/978-3-642-49749-0 that dispersal and cell turnover limit intratumour hetero- [14] Dhar, D. , 1986. Asymptotic Shape of Eden Clusters. In: geneity. Nature 525, 261–264. Stanley H.E., Ostrowsky N. (eds.) On Growth and Form. [7] Br´u, A., Albertos, S., Luis Subiza, J., Garc´ıa-Asenjo, NATO ASI Series (Series E: Applied Sciences), vol. 100, J.L., Br´u,I., 2003. The Universal Dynamics of Tumor [Springer, Dordrecht, 1986]. Growth. Biophysical Journal 85, 2948–2961. [15] Wolf, D. E., 1987. Wulff construction and anisotropic sur- [8] Nekovee, M., Moreno, Y., Bianconi, G., Marsili, M., face properties of two-dimensional Eden clusters, J. Phys. 2007. Theory of rumour spreading in complex social net- A: Math. Gen. 20, 1251. works. Physica A 374, 457–470. [16] Bertrand, Q. and Pertinand, J., Dimension improvement [9] O’Connor, C. and J. O. Weatherall, The Misinformation in Dhar’s refutation of the Eden conjecture, Phys. Lett. Age: How False Beliefs Spread (Yale University Press, A 382, 761-765. 2019). [17] Plischke, M., R´acz,Z., 1984. Active Zone of Growing [10] LaCroix, T., Geil, A., and O’Connor, C. The dynamics Clusters: Diffusion-Limited Aggregation and the Eden of retraction in epistemic networks. Preprint. (2019) Model, Phys. Rev. Lett. 53, 415. 14

[18] Auffinger, A., Damron, M. and Hanson, J., 2017. 50 years [33] Grassberger, P., de la Torre, A., 1979. Reggeon field the- of first-passage percolation (Univ. Lecture Series, Vol. 68, ory (Schl¨ogl’sfirst model) on a lattice: Monte Carlo cal- American Mathematical Soc.) culations of critical behaviour. Annals of Physics 122, [19] Kardar, M., Parisi, G., Zhang, Y.-C., 1986. Dynamic 373 (1979). Scaling of Growing Interfaces. Phys. Rev. Lett. 56, [34] Grassberger, P., 2003. Critical percolation in high dimen- 889–892. sions. Phys. Rev. E 67, 036101 (2003). [20] Corwin, I., 2012. The Kardar–Parisi–Zhang equation and [35] Wikipedia, Percolation critical exponents, 2021. universality class. Random Matrices: Theory Appl. 01, https://en.wikipedia.org. 1130001. [36] Richardson, D., Random growth in a tesselation, Proc. [21] Ara´ujo,N., Grassberger, P., Kahng, B., Schrenk, K.J., Cambridge Phil. Soc. 74, 515-528 (1973). Ziff, R.M., 2014. Recent advances and open challenges in [37] Durrett, R., Liggett, T. M., 1981. The shape of the limit percolation. Eur. Phys. J. Spec. Top. 223, 2307–2321. set in the Richardson Growth model, Ann. Prob. 9, 186- [22] Hinrichsen, H., 2000. Non-equilibrium critical phenom- 193. ena and phase transitions into absorbing states. Advances [38] Jensen, I., 1999 Low-density series expansions for di- in physics, 49(7), 815-958. rected percolation: I. A new efficient algorithm with ap- [23] Kumar, A., Dhar, D., 2020. Improved Upper Bounds plications to the square lattice, J. Phys. A: Math. Gen. on the Asymptotic Growth Velocity of Eden Clusters. 32, 5233–5249. J. Stat. Phys. 180, 710–720. [39] Duminil-Copin, H., Kozlowski, K. K., Krachun, [24] Williams, T., Bjerknes, R., 1972. Stochastic model for D., Manolescu, I., Oulamara, M., 2020. Rota- abnormal clone spread through epithelial basal layer. Na- tional invariance in critical planar lattice models. ture 236, 19–21. arXiv:2012.11672[math.PR] [25] Wang, Z., Bauch, C.T., Bhattacharyya, S., d’Onofrio, A., [40] Alm, S.E., Deijfen, M., 2015. First Passage Percolation 2 Manfredi, P., Perc, M., Perra, N., Salath´e,M., Zhao, D., on Z : A Simulation Study. J. Stat. Phys. 161, 657–678. 2016. Statistical physics of vaccination. Physics Reports [41] Reis, Fabio D. A. Aarao. (2003). Depinning transitions 664, 1–113. in interface growth models. Brazilian Journal of Physics, [26] Das, A., Dhar, A., Goyal, S., Kundu, A., 2020. Covid- 33, 501 -513. 19: analysis of a modified SEIR model, a comparison of [42] Derrida, B., Retaux, M. The Depinning Transition in different intervention strategies and projections for India. Presence of Disorder: A Toy Model. J. Stat. Phys. 156, arXiv:2005.11511 [physics, q-bio]. 268–290 (2014) [27] Kordzakhia, G., 2005. The Escape Model on a Homoge- [43] Brunet, E., Derrida, B., Mueller, A.H., Munier, S., 2006. neous Tree. Electron. Commun. Probab. 10, 113–124. Phenomenological Theory Giving the Full Statistics of [28] Bordenave, C., 2014. Extinction probability and total the Position of Fluctuating Pulled Fronts, Phys. Rev. E progeny of predator-prey dynamics on infinite trees. Elec- 73, 056126. tron. J. Probab. 19, paper no. 20. [44] Tom´e, T., Ziff, R.M., 2010. Critical behavior of the [29] Tang, S., Kordzakhia, G., Lalley, S.P., 2018. Phase susceptible-infected-recovered model on a square lattice. Transition for the Chase-Escape Model on 2D Lattices. Phys. Rev. E 82, 051921. arXiv:1807.08387 [cond-mat, q-bio]. [45] Janssen, H.K., 1981. On the nonequilibrium phase transi- [30] Kesten, H., 1980. The critical probability of bond perco- tion in reaction-diffusion systems with an absorbing sta- lation on the square lattice equals 1/2. Commun.Math. tionary state, Z. Phys. B 42, 151–154. Phys. 74, 41–59. [46] Grassberger, P., 1982 On phase transitions in Schl¨ogl’s [31] Durrett, R., Junge, M., Tang, S., 2020. Coexistence in second model, Z. Phys. B 47, 365–374. chase-escape. Electron. Commun. Probab. 25, paper no. [47] de Souza, D.R., Tom´e,T., 2010. Stochastic Lattice Gas 22 . Model Describing the Dynamics of the SIRS Epidemic [32] Owen, M.R., Lewis, M.A., 2001. How Predation can Process, Physica A 389, 1142. Slow, Stop or Reverse a Prey Invasion. Bulletin of Math- ematical Biology 63, 655–684.