Arxiv:1908.08163V2 [Cond-Mat.Stat-Mech] 8 Oct 2020 Bedsrpnishv Perdi Eetyas There’S Years

Home , Directed percolation, Universality class

arXiv:1908.08163v2 [cond-mat.stat-mech] 8 Oct 2020 bedsrpnishv perdi eetyas There’s years. recent in appeared have unavoid- discrepancies successful, able quite proven has processes branching generic [18]. a networks are training neural they of of have course suggesting property the networks, distributions in neural power-law naturally artificial emerge Equivalent to and process- found hypothesis, been information 15]. optimal brain [13, dynam- with critical ing brain connectivity the in and principles of ics avalanches self-organizing neuronal core together These link the 17]. 16, at 11, are 9, [8, imag- whole-brain (fMRI) resonant to magnetic ing from neurons, functional magnitude, and hundred imaging of few calcium orders a many of by slices have dif- size that percolation in systems directed neural fer in of experimentally observed that been similar to with exponents avalanches or critical cascades spreading power-laws. characteristic Neuronal exhibiting statistics scale-free a passes spreading process) of threshold. probability branching critical the as typically (i.e. transition processes phase directed-percolation Such a network. a exhibit pro- spreading on all neu- unfolding are and [8–15] cesses 7], brain power the [6, in [5], networks avalanches ronal networks electrical on social net- cascading on computer failures shared rumours through [4], tunnelling works malware and worms rtclt nsraigpoesswtottm-cl sep time-scale without processes spreading in Criticality lhuhtempigo ernlaaace to avalanches neuronal of mapping the Although shared exhibit systems these point, critical this At [1–3], networks contact human on spreading Diseases 1 eateto hsc n srnm,Uiest fCalgary of University Astronomy, and Physics of Department iiiy alt sals rtclt nteasneo ti of bra absence brain. the the the as in in such criticality criticality measures, establish scali used to appropriate fail commonly tibility, that an show identify we and governed Finally, is activity, exponents both of these of between propagation classes transition universality the that the this find univer from limits, the exponents two to critical these correspond of Outside line critical respectively. scale- this which percolation, of along limits exam line two for critical The the epileptic, giving turning limit activity mean-field neural or epidemic ing udmnal hne.Ti hnemnfssisl natr a in itself manifests of scen change i a classes cascades such This in spiking that changes. or here fundamentally diseases show we zoonotic th th topologies, — as network of relevant process such initiation branching case, separatio the a the or clear and not percolation a cascade directed exists spreading as there single modelled a cases of some co gation the in from While interest; broad dynamics. of is behavior their derstanding pedn rcse nntok r bqiosi ohhuma both in ubiquitous are networks on processes Spreading ailJ Korchinski, J. Daniel .INTRODUCTION I. 2 3 eateto ple hsc,HnagUiest,Ansan, University, Hanyang Physics, Applied of Department ocks ri nttt,Uiest fClay Calgary Calgary, of University Institute, Brain Hotchkiss critical pedn,wihdtrie h ne n h rprisof properties the and onset the determines which spreading, 1 airG Orlandi, G. Javier Dtd coe 2 2020) 12, October (Dated: nvitro in nvivo in hypothesis 1 en-o Son, Seung-Woo o“ii”o u oetra esr nu.Frsmaller For input. sensory external to due so or due not “minis” activate are spontaneously to can systems neurons the Neural – however than a simple avalanches. faster new words, much of timescales other initiation on In terminate prop- avalanches and so assumes typi-agate node. do description processes to “root” process branching branching induced one and when presuppose node, excited cally antecedent be an to by only nodes process for branching proper allows classical show rarely a systems separation; real time-scale that fact overlooked ten by truncated effects. those finite-size from Fur- avalanches dis- to sub-critical challenging [27]. tinguish it make coarse-graining constraints experimental and limi- ther, sub-sampling experimental These the of by tations [23–26]. complicated transition are percolation challenged, observations a tran- been of oscillation-synchronization instead also a sition has proposing be- some percolation appeared even with exponents also directed critical have to the with 2.5 long Whether avalanches to 22–24]. 21], 14, 1.2 16, [10, from [11, ranging literature exponents the in reported with e.Atog naaacesz distribution size val- avalanche an mean-field are Although predicted avalanches ues. the in observed following least power-laws whether param- at truly determining order stems suitable and debate a eter identifying This challenges from defini- 20]. part really different [19, are a altogether systems or tion these sub-critical, whether quasi-critical, on critical, debate ongoing an esaesprto aln o enlssof reanalysis a for calling separation me-scale oeo hs sushwvr ih edet h of- the to due be might however, issues these of Some ult aiet tefi h appearance the in itself manifests duality τ resraigbhvorcnb observed. be can behaviour spreading free to feieist nesadn brain understanding to epidemics of ntrol ietdadudrce eclto.We percolation. undirected and directed grltosi o h rniinpoint. transition the for relationship ng l.W rsn nltclrslsi the in results analytical present We ple. ≈ aiycasso ietdadundirected and directed of classes sality erlntok.Fralrecasof class large a For networks. neural n yacmeiinbtenmrigand merging between competition a by niinbtendffrn universality different between ansition rotentr fteoealspreading overall the of nature the ario et—sc htsraigcnbe can spreading that such — next e r r lopoessfrwihti is this which for processes also are ere 1 ftm clsbtentepropa- the between scales time of n agr,AbraTN14 Canada 1N4, T2N Alberta Calgary, , . cnitn ihma-ed a enwidely been has mean-field) with (consistent 5 let 2 N,Canada 4N1, T2N Alberta , -aeadntrlsses Un- systems. natural and n-made cigrtoaddnmcsuscep- dynamic and ratio nching ,2 1, 58,Rpbi fKorea of Republic 15588, rto n h rtclbrain critical the and aration n Jörn Davidsen and naaaceturn- avalanche an ,3 1, p ( S ) ∼ S − τ 2 and intermediate systems, where multiple independent 4 cascades are rare, the assumption of time-scale separation is not particularly limiting. If the rate of sponta- 3 neous activity is not too high and all activity is aggre- gated into a single avalanche, then τ changes from the 2 expected 1.5 to 1.25 [28]. However, in large neural systems, there are no global quiet periods with which to 1 delimit neuronal avalanches as in smaller systems. Some attempts at defining avalanches by activity exceeding a 0 threshold exist, but it is not clear whether these thresh- olds identify a genuine critical point [25, 26]. This reflects t=0 t=1 t=2 t=3 t=4 t=5 t=6 the general challenge of defining criticality in strongly driven systems. However, some progress has been made. FIG. 1. Example dynamics of the branching process with To disentangle neural avalanches in whole-brain fMRI spontaneous activations. Multiple spontaneous activations and study their statistics, Tagliazucchi et al. used phys- initiate on a simple linear bidirectional network (left). The ical proximity (i.e. nearest neighbour connections) to dynamics here exhibit two independent avalanches, one with delimit avalanches, instead of binning all activity to- two roots (node 1 at time t = 1 and node 4 at time t = 2), gether [11]. The same method was necessary to assess and one with a single root (node 0 at time t = 3). With spontaneous activations, spatially distinct events can overlap criticality in whole-brain zebra-fish data [29]. The ap- in time (e.g. t = 3 and t = 4) and initially distinct cascades proach of using network topology to identify avalanches of activity can overlap to form larger avalanches. makes sense, as information processing can only occur between connected elements of the network. To this end, neuronal avalanches have been generalized by “causal II. RESULTS webs”, which uses network structure to separate out independent avalanches [30]. A. Model It has remained an open question whether a genuine critical point with scale-free activity can exist alongside To study the effect of spontaneous activations, we spontaneous activity, particularly as other markers of a consider here a discrete time ε-SIS process [31] on di- directed percolation transition (such as a unity branching rected networks equipped with spontaneous activations ratio and the appearance of an active fraction) are all (see Fig. 1). At each time step, a given node can be acti- affected in different ways by spontaneous activity. We vated by means of a spontaneous activation with proba- address the issue here using a minimal spreading process bility p or through an incoming link by an infected parent with a spontaneous activation rate, making it a discrete- with probability q. More precisely, the probability that time ε-SIS (susceptible-infected-susceptible) model, and node i is activated at time t + 1 is given by: study it on a variety of complex networks. Without spontaneous activations, there are no con- P (i,t +1)=1 (1 p)(1 q)m(i,t), (1) current independent cascades of activity, and a directed − − − percolation phase transition is present. We show that where m(i,t) counts how many parents of node i were the introduction of spontaneous activations means that active at time t. Therefore spreading events occurs on the macroscopic markers used to identify the directed a timescale of 1, while spontaneous activations occur on 1 percolation transition, (i.e. the appearance of an active a timescale of p− . This model be considered a type of fraction, or a branching ratio of one) no longer identify Domany-Kinzel cellular automaton [32], or akin to some a phase transition. Nonetheless, by using the network form of mixed-percolation [33]. In our model, nodes do structure to disentangle causally unrelated avalanches, not remain activated, but recover and so may be re- we can define a phase transition even in the presence of activated the following time step. In the Appendix H, spontaneous activations, with scaling-relations between we also consider a variant of this model that includes exponents and finite-size scaling. We perform an exten- the immunization of nodes, i.e., a susceptible-infected- sive study of the critical properties at this phase tran- recovered (SIR) process, with multiple initial spreaders. sition on a variety of network structures and show that Large systems with spontaneous activations will have the presence of any spontaneous activation changes the concurrent and possibly unrelated avalanches. We em- underlying universality class from that of directed per- ploy the network structure itself to identify indepen- colation to that of undirected percolation, while preserv- dent avalanches, using the “causal-webs” approach de- ing some features of directed percolation. To explain scribed in [30]. To summarize the approach, we identify these results, we derive an analytical mean-field theory nodes with no active parents (i.e., no possible source of for branching processes with spontaneous activation and network-borne activation) as “roots” of newly-initiated show that the appearance of undirected percolation ex- avalanches (see Fig. 1). Nodes with active parents inherit ponents is a direct result of the merging of initially inde- the avalanche labels of their parents. As avalanches over- pendent avalanches. lap, they are merged together, so that nodes only ever 3

a. 100 b. have one label. This merging reﬂects that true causal in- Supercritical formation is often obscured in real systems, or that con- Critical -3 Subcritical tributions from both streams of activity are necessary for 10 Giants activation. This description of avalanches maps naturally 10-6 Probability P(s)

to the clusters of traditional percolation. Giant Fraction (g) The model has two limiting cases: i) For p = 0, this 10-9 100 102 104 106 108 model is a pure branching process with branching pa- Size (s) Spreading Probability (q) rameter q, which belongs to the directed-percolation universality class. This case corresponds to neural activ- FIG. 2. Approach to the phase transition in directed perco- ity where avalanches are infrequent and a single lead- lation. a. Avalanche distributions for with different spreading neuron can be positively identified in each avalanche. ing parameters q with spontaneous activation rate p → 0 on ii) q = 0 corresponds to the ordinary percolation model random 10-regular networks. Sub-critical avalanches are dis- on a directed network with probability p. This corre- tributed with an exponentially-truncated power law p(s) ∼ −3/2 sponds to neural activity that is entirely driven by exter- s exp[−s/sc]. At criticality, avalanches are scale-free −3/2 nal sources or spontaneous activation and do not spread sc → ∞ such that p(s) ∼ s . The absolute size of the giants in the super-critical phase scales linearly with the net- on the network, such as in the retina. Both limits ex- 4 5 hibit continuous phase transitions, but fall into different work size (empty symbols N = 10 , filled N = 10 ). b. The universality classes, and are characterized by power laws giant fraction g of nodes participating in the largest avalanche for the same simulations as in a. exhibiting different critical exponents. Establishing that a phase transition exists for p and q that are simultaneously nonzero and the universality class of this transition avalanche exponent has been reported for directed small- are the principle efforts of this paper. world networks. To check the consistency of our findings for these networks, we can lower the density of long- range connections. Indeed, we find that the directed- B. Numerical results percolation exponent for small-world networks tends to- ward the (1+1)-dimensional directed-percolation limit of The most experimentally accessible indicator of criti- 1.108 expected of a circulant graph (see Fig. 12 and cality in systems with activity spreading is the size dis- associated≈ text). For the power-law network, the cor- tribution of clusters, which shows a different critical ex- responding degree exponent was chosen such that the ponent in directed and undirected percolation. For all p, undirected-percolation exponent would change (from the at some critical qc(p) we observe a transition that defines 5/2 mean-field value to 8/3 as predicted in [37]) while a critical line. Below the critical point, with q < qc(p), leaving the directed-percolation exponent at 3/2 (as pre- avalanches are limited in size (see Fig. 2a), while above dicted in [36]). the critical point, a permanent giant component affecting In hierarchical modular networks, we observe non- a non-zero fraction of the network appears. At the crit- scale-free behaviour for avalanches below the base mod- ical point, the exponential cut-off that characterizes the ule size M (s < M), a p dependent power law for 2/3 sub-critical phase diverges and the avalanche distribution intermediate-size avalanches (M 0, the critical point belongs We find that in all cases the first power-law exponent to the same universality class for that network topology. is consistent with the directed-percolation exponent for In the super-critical regime, a giant component appears, that system, while the latter exponent is indistinguish- just as in directed and undirected percolation. The prob- able from the pure percolation exponent (see Table I). ability that a randomly-selected node is in the giant com- Yet, to the best of our knowledge, no directed-percolation ponent is the giant component fraction g, which exhibits 4

TABLE I. Calculated critical exponents for diﬀerent network structures. In each column, we report both the theoretical value from percolation theory, and the value determined in this work, either analytically (by application of generating functions) or from numerical simulations (where they are reported with decimal values). Errors in 1/ν indicate the range over which an acceptable curve-collapse was obtained. A number of additional critical exponents were determined for the k-regular network, these are summarized in Table III. Exponent τDP τ β 1/ν −τDP −τ β −1/ν Quantity P (s) ∼ s for ssm g ∼ (q − qc) qc(N) − qc ∼ N Small-World - ≈ 1.35 5/2 [35] ≈ 2.5 1 [35] ≈ 1.0 - 0.35(5) Power-law 3/2 [36] ≈ 1.5 8/3 [37] ≈ 2.67 2 [37] ≈ 2.0 1/5 [37] 0.25(5) Hierarchical Modular Network Varies [22, 38] Varies - ≈ 2.1 - ≈ 0.8 - 0.15(5) k-Regular Network 3/2 [39] 3/2 5/2 [40] ≈ 2.5 1 [40] 1 1/3 [37] 0.36(2)

a. 100 b. a. b. p=10-4 s-1.35 10-1 101 100 p=10-3 Increasing N 2/3 1 p (q-qc) 10-5 -3 -1 5/2 10 10 1 P(s) (q-q ) 0.36 c -2 10 s1.15 g N s-5/2 P(s) S -5 p=10 -5 -3 Giant Fraction (g) 10 10 p=10-2 Increasing N 10-10 10-4 100 102 104 106 -4 0 4 10 10 10 10-7 10-5 Size (s) s p2/3 10-3 10-2 10-1 10-2 100 0 c. 10 d. 0.36 q - qc (q - qc)N

-3/2 s 10-2

-1 c. d.

-4 p 10 5 8/3 10 Increasing N ) 7/6 2/3 P(s) s 0 -8/3 10 s 10-4

2 -8 P(s)(s p 10 p=10-7 (q-qc) 0.50 -0.36 102

-3 N g N

p=10 χ 10-6 100 103 106 10-4 10-1 102 10-2 Size (s) s p2/3 10-1 e. f. 104 10-1 10-2 100 10-2 10-1 100 2/3 0.33 p s

2/3 1 -9 -3 0.36 0.25 s p=10 p=10 (q - qc)N (q - qc) N p 2.10 2 -3 5/2 10 p=10 10-4 s0.8 P(s) s

P(s)/P(M) s FIG. 4. Finite size scaling effects on giant component size p=10-6 10-7 100 and susceptibility. a. Giant components on 10-regular net- -4 0 4 -6 -1 4 10 10 10 10 10 10 works of varying sizes. Circles and triangles are p = 10−3 s p2/3 s/M p2/3 and p = 10−4, respectively, while solid lines are the analytical calculations for infinite lattices. b. As in a, but re-scaled to produce a finite-size scaling curve collapse. c. Symbols are FIG. 3. Critical avalanche statistics on critical line for various 2 network topologies. a. Small-world networks with N = 105 as in a, but studying the susceptibility χ = hs ic for finite and average degree hki = 10. b. As in a, but re-scaled clusters, which shares the same finite-size scaling exponent. d. Curve collapse for power-law networks (degree distribution to produce a curve collapse. c. Uncorrelated power-law in- −3.5 −4 β and out-degree distributions, p(k) ∼ k−3.5, with N = 107 p(k) ∼ k ) with p = 10 . Solid line is g ∼ (q − qc) for nodes. d. Asin c, but rescaled to produce a curve-collapse. e. β = 2, consistent with theoretical prediction. Re-scaled avalanche distributions from 10-regular networks. Lines are simulations on infinite networks, while transparent circles are from finite simulations with N = 107; both are G β a power-law scaling, g = NT (q qc) , where G denotes at q = qc(p), the analytically-determined critical point. f. the size of the largest cluster,∼N −the number of nodes in Re-scaled avalanche distributions from a hierarchical modular the system, and T the simulation duration. However, g network, with N = M × 215 nodes, on a 15-layer hierarchy 2 exhibits a strong finite-size effect, with smaller systems with base module size M = 10 . Solid lines are a guide for the eye. Unscaled panels e-f may be found in Figs. 16a and having a larger effective critical point. Below the effective 16b. The critical lines arising in each network topology are critical point, the largest cluster G does not scale with presented in Fig. 6. the simulation duration and is not percolating. For this reason, we resort to finite-size scaling to reveal the critical behaviour of g on finite networks (Fig. 4). As N , → ∞ the effective critical point qc(N) tends towards the true 1/ν¯ critical point, with (qc(N) qc) N − . The correct- ness of our finite-size scaling− is confirmed∼ by considering the finite cluster susceptibility χ s2 (where the de- ≡ h ic c 5 notes an average over all clusters), which should obey the each neighbours connects to is distributed according to γ/ν¯ same scaling collapse with χN − (Fig. 4c). For mean the original PGF. Our approach, detailed in Appendix B, field, 1/ν¯ = 1/3 and for pure percolation on power-law follows that same spirit, except that a system of two 3.5 networks (with degree distribution p(k) k− ), we ex- self-consistently coupled PGFs are required to describe pect that 1/ν¯ = 1/5 [37]. Here, we find∼ that the best the total cluster distribution. One PGF corresponds to 1 scaling collapse occurs near these values, with ν¯ 0.36 the sizes of clusters reached from a direct descendent, 1 ≈ while the other describes the cluster size reached when for 10-regular networks and ν¯ 0.25 for power-law networks. ≈ two independent cascades merge. On a tree-like net- Above the effective critical point, the giant component work, merging occurs when spontaneous activations meet agrees with our analytical predictions for the infinite- and become a larger avalanche. Therefore, the directed- size limit (Fig. 4a). As expected, the giant components percolation-like behaviour is entirely contained within emerge with β = 1 for the mean-field case of random 10- the first PGF, while the second PGF captures the effect regular networks (Figs. 4a and 4b). As can be seen in of new spontaneous activations. Fig. 4d, the giant component grows with β = 2 for the This pair of PGFs combines to define the PGF 3.5 s given power-law networks. Since p(k) k− and there H (x) = ∞ P (s)x which corresponds to the ∼ 0 s=1 n are no correlations between the indegree and outdegree, avalanche-size distribution Pn(s) obtained from sam- it is known that β = 2 for undirected percolation [37] pling randomP active nodes (denoted with the sub-script and β = 1 for directed percolation [36]. This implies n). Derivatives of H0(x) evaluated at x = 0 directly that the emerging giant component in our system is in yield Pn(s). The average cluster size is just given 2 the universality class of undirected percolation. by s n = H0′ (1) and the susceptibility χ = s c = h i 1 h i H0′ (1)/ 0 H0(x)dx, where the sub-script c indicates sampling over clusters as opposed to active nodes. Since the C. Analytical results giant componentR is the unique infinite-size avalanche, the probability a random active node belongs to it is just In this section we establish analytically that the uni- 1 H0(1), and so the giant component is also deter- − versality class of the phase transition is lifted from di- mined by H0. For sufficiently small p and q, the giant rected to undirected percolation by the addition of spon- component is zero and all clusters are finite, but as the taneous activations. To understand the transition be- critical line is approached the susceptibility χ diverges as 1 1 tween directed- and undirected-percolation exponents, χ qc q − (for fixed p) or χ pc p − (for fixed ∼ | − | ∼ | − | we consider the analytically tractable k-regular network. q) as derived in Appendix B. This divergence defines the Using the generating function formalism, we can explic- critical line, which can be simply expressed as itly derive scaling exponents related to the avalanche 2 size and emergence of the giant component, as well as 0= k(1 σ) (k 1)σσm , (2) derive a critical line. By studying the sizes of singly- − − − rooted avalanches on this critical line, we can identify where σ(p, q) is the reproduction number or branching the size at which the merging of independent clusters of activity becomes the predominate mechanism for clus- ratio, and σm(p, q) is the merging number, corresponding to the number of cascades of activity leading to a ter growth, and thereby explain the scaling collapse effected by p2/3 observed in Figs. 3c–3f. Additionally, the randomly-selected active node with at least one parent. Clearly then, the critical line has σ < 1 for all σ > 0, directed-percolation universality class exhibits two dis- m tinct diverging correlation lengths at the critical point. meaning that giants can occur even before an average reproduction number of 1 is attained. However, only one correlation length diverges in our model, reinforcing that this is an undirected-percolation As shown in Fig. 5a, σ = 1 only in the p 0 and q 1 limit of directed percolation for a k→-ary transition. → k The avalanche distribution of our model is akin to the tree [40], where it agrees with the derived critical line. cluster-size distribution of percolation and directed per- At the directed-percolation critical point, the active frac- colation; this distribution has been analytically deter- tion of nodes Φ = Φ(t) exhibits a divergence in its h i ∂Φ mined on a variety of infinite random networks for both dynamic susceptibility χ0, with χ0 ∂p diverging as 1 1 ≡ types of percolation by using probability generating func- χ0 k q − . In the context of neural systems with tions (PGFs) [1, 36, 37, 42]. The technique’s key assump- mixed∼ | time-scales− | and a fixed level of spontaneous activation is that there are no loops and that all nodes are tion, the maximum of this dynamic susceptibility defines equivalent, in the sense that their network properties are a “Widom” line (cf. Fig. 10 and associated text) and has independent of the properties of their neighbours. Al- been proposed as a quasi-critical line [44]. Although all though this is only an approximation, this tree-like ap- three of these measures identify the directed-percolation 1 proximation can still perform well in cases where loops critical point p = 0 and q = k , they disagree as soon as are prevalent [43]. This assumption lets one write down spontaneous activation is introduced (p = 0) and exhibit a self-consistent equation for the PGF in terms of the distinct scaling (cf. Fig. 11). In the p6 1 limit, the number of connected neighbours where the cluster that Widom line scales as p ( 1 q), the σ ≪= 1 line scales ∼ k − 6

a. 0.06 b. 1010

1 # Roots = 1 avalanches are described by a branching process, on a

/ S §¨© 2£ ¤¥ s-3/2 # Roots > 1 k-ary tree, with a branching probability Pd1 = (1 100 0.04 Widom k 1 − Line -5/2 Φ) − (1 (1 p)(1 q)) corresponding to the probability sm s p − − − P(s) / p that a daughter branch activates with exactly one par- σ= 10-10 0.02 -3/2 s exp[-s/sm] ent. The probability distribution for the size of the singly 1

¦ 3/2 k 10-20 rooted avalanche is Pmergeless(s) s− exp[ s/sm], 0 10-8 10-4 100 104 108 ∼ k 1 −

0 0.04 0 ¡¢ 0.12 where sm = 1/ ln[kPd1(Pd/(1 1/k)) − ] (see Eq. (C3)) s / p-2/3 − − q denotes the characteristic scale above which avalanches c. d. 102 merge and Pd is the probability a site does not activate, δq-1 104 despite having an active parent. Hence, we expect that 10-1 3/2 〉 δq-3 the exponent s− should be exponentially suppressed 1 p

s χ

R(s) -4 102 10 〈 at s . In the limit of p 0 on the critical line (see m → 10-7 δq < 0 Eq. (C4)) sm scales as: δq > 0 100

-5 0 5 -4 -2 0 2 2 10 10 10 10 10 10 10 2(k 1) 1 − 1/3 s / sm |δq| / p s − q . (4) m ≈ 3k3 k − e. f. T-2 102 103 Combining Eqs. (3) and (4) shows that the characteristic s1/4

-1/3 2/3 /p -2/3 -4 size before merging scales as sm p− on the critical 10 -7 〉 T ∼ 100

T(s) line. P(T) p 〈 s1/2 10-11 Simulations conﬁrm that the smallest avalanches typ-

10-2 10-18 ically only have one root (Fig. 5b), while the largest 10-4 100 104 108 10-2 100 102 avalanches have a number of roots that scale with the -1/3 s/sm T/p avalanche size (Fig. 5c). This means that there are two competing processes at play in these avalanches, both FIG. 5. Power-law transitions are governed by merging. a. the propagation of the avalanche, which belongs in the Phase diagram for the k-regular network, with k = 10. Points directed-percolation universality class, and the merging on the critical line correspond to the (p, qc(p))in the other of initially-independent events, which falls into the perco- panels of this figure, with the p ranging from 10−2 to 10−9. lation universality class. This explains the appearance of b. Re-scaled avalanche size distribution for various p simu- the two power-laws and the associated curve collapse in lated on an infinite 10-regular network, partitioned into those Fig. 3. The first power-law is governed by the spread- avalanches with a single initiation site (empty circles) and ing of activity from a single initiation site, while the those with multiple initiation sits (crosses). The theoretical second power-law is governed by the merging of activ- distribution of mergeless avalanches is indicated with the solid ity springing from multiple sites. This 2 scaling is a line (cf. Eq. (C2)). c. Re-scaled average number of roots R − 3 for avalanches of a given size for simulations on an infinite good approximation for random graphs that are close to 10-regular network. d. Re-scaled susceptibility near to the mean-field. In Appendix F, we consider small-world net- critical point, where δq = qc − q, calculated by the generat- works with a low shortcut density. These networks are ing function H0. Sub-critical values δq < 0 are shown with locally one-dimensional, and as a consequence exhibit a 0.75 empty circles and exhibit two power-laws, while super-critical different scaling, sm p− (cf. Fig. 12) due to the values δq > 0 show only one. e. Average avalanche duration directed-percolation phase∼ being (1+1)-dimensional infor simulations of a given size collapse onto a single curve. f. stead of mean-field. Avalanche duration distribution collapses onto a single curve This transition between the directed- and undirected- with two power-laws. Unscaled data for panels c-f are found exponents also manifests itself in the approach to the in Fig. 17. critical point. For instance, for q < qc the susceptibility can be approximated by

1 2 s as p ( k q) and the critical line scales as m 2 τDP ∞ 2 τ ∼ − χ s s− ds + Θ(s s ) s s− F (s/s )ds , ≈ ξ − m ξ 1 3 (k 1)2(2k 1) Z1 Zsm qc − 5 − pc . (3) k − ≈ k where Θ is the Heaviside step function, and sξ is the 1/σ size cut-oﬀ of the pure-percolation tail, s q q − As for the q = 0 endpoint to the critical line, χ diverges ξ ∼ | c − | 1 and F is a universal scaling function. Then, using that when p = and q = 0, the pure percolation critical 1 1/3 (2k 1) qc p from Eq. (3), χ is (up to arbitrary multi- point for the− Bethe lattice of coordination number 2k. k − ∼ plicative constants C1, C2), Hence, the critical line contains members belonging to 1 3 3 (3 τ)/σ two distinct universality classes. χ = C1p− (1 + δq/√p) + C2Θ(sξ sm)δq − . 2/3 To understand the appearance of the p− scaling of − the transition point shown in Fig. 3, we can consider This suggests that we see a transition between exponents the distribution of avalanches with only one root. These when δq √3 p, precisely as observed in Fig. 5d. Now, ≈ 7

γ 2 2 since χ δq defines γ, we have arrived at the usual (1 σ) = σ β, i.e. on the critical line where s n ∼ 3 τ − h i scaling relation γ = −σ . This scaling relation holds diverges. The divergence of ξ and the decay of g(2d, 0) 1 3 is compared to its analytical⊥ form in Fig. 13. The for both the directed (γDP = 3, σDP = 2 , τDP = 2 ) 1 5 non-divergence of ξ on the critical line suggests that the and undirected (γ = 1, σ = 2 , and τ = 2 ) percolation k regimes. critical line is not a directed-percolation transition. One might object that the non-divergence of ξ is a problem A transition from directed-percolation exponents also with its construction. In the Appendix Gk, we consider appears in the dynamical exponents relating the size of also an isotropic correlation length that diverges on the avalanches to their duration (cf. Fig. 5e), where the ex- 1 critical line, and show that an alternative definition ponent transitions from s T σνz= 2 to a power-law 1 ∼ h i of ξ based on the typical number of generations of consistent with s T 4 . The onset of this transition k ∼ 2 direct descendants to an active node only diverges again occurs with avalanches of size sm p− 3 , which de- ∼ 1 on the σ = 1 line (cf. Fig. 5a). Additionally, the fines a characteristic time to merging, Tm √sm p− 3 . ∼ ∼ fact that ξ diverges with ν = 1, as in undirected The scaling of this characteristic time captures an expo- ⊥ 1 ⊥ percolation, instead of ν = 2 as expected in mean-field nent transition in the distribution of the avalanche dura- directed percolation reinforces⊥ that the critical line is an tions (cf. Fig. 3f, with P (T ) T α, with α = 2 and ∼ DP undirected-percolation transition. a new asymptotic α 7.0. Intriguingly, the directed- In summary, the critical line is an undirected (as op- ≈ τ 1 percolation scaling relation α−1 = σνz is satisfied even posed to directed) percolation transition except at a − in the merging regime, assuming α = 7, τ = 5/2, and singular point. This is supported by avalanche distri- 1 σνz = 4 . bution exponents, exponents of the order parameter g, The existence of robust scaling relations and of curve undirected-percolation scaling relations, and the diver- collapses (Figs. 5b–5f) that appear universal indicate gence of a single correlation length. Many critical ex- that the critical line (for p> 0) belongs to a single univer- ponents of the directed percolation remain observable on 1 sality class. Since this includes the point p = 2k 1 and small scales, such as in the beginning of the avalanche size q = 0, which we know is exactly undirected percolation,− distribution or in the susceptibility χ. These exponents it suggests that the entire critical line (save for p = 0, then shift to the undirected exponents when the merg- q = 1/k) belongs to the universality class of undirected ing of initially-independent avalanches becomes preva- percolation. lent. Meanwhile, other measures of criticality that hold We can further strengthen the argument that the crit- for directed percolation, such as the divergence of the dy- ical line is an undirected-percolation transition by study- namical susceptibility χ0 and a reproduction number of ing the correlation lengths of the system. Undirected one, no longer capture critical behaviour. Instead, they percolation exhibits a single isotropic diverging correla- predict phase-curves that agree only in the p = 0 limit, tion length ξ δq ν , while directed percolation ex- and scale with different power-laws near the directed- − 1 a ∼ | | ν⊥ percolation limit. Specifically, q p , with a = 1 hibits two diverging correlation lengths ξ δq − and k − ∼ νk ⊥ ∼ | | for the Widom line, a = 2 for the σ = 1 line, and a = 3 ξ δq − corresponding to spatial and temporal correlationk ∼ | | lengths respectively. We consider the correla- for the critical line (c.f. Fig. 11). Thus, the directed- tion lengths corresponding to ξ and ξ for our system, percolation transition is not robust with respect to the and will show that only one diverges⊥ onk the critical line, introduction of spontaneous activation – any level of ex- precluding a directed-percolation transition. The two- ogenous driving will introduce independent outbreaks, which on the largest scales will begin to merge. This point connectedness function, γ(i,ti,j,tj ) measures the is perhaps surprising, because the undirected-percolation probability that node i at time ti and node j at time limit q = 0 obeys detailed balance, while for the remain- tj belong to the same cluster over the ensemble average. If we denote the shortest path connecting nodes i der of the critical line with q > 0 detailed balance is not respected. and j as dij then we expect that the average connectedness function should decay with dij . This can be seen by studying the exponential decay of the average connected- III. DISCUSSION ness function, g(d, t) = γ(i,ti,j,tj ) dij =d, t=tj ti, i active which measures the decayh of activityi away from− an active node. A. General model observations Typically, activity decays exponentially, with g(2d, 0) exp[ d/ξ ] and g(0,t) exp[ t/ξ ] We have described a two-parameter spreading pro- defining the∼ two correlation− ⊥ lengths ξ and∼ ξ [32].− Ink cess that includes spontaneous activations and exhibits ⊥ k the loop-less (large N) approximation, g(0,t) = δt0, as a phase line along which the critical exponents and be- in the absence of loops activity can never return to the haviour of both directed and undirected percolation ap- same site, meaning the correlation length ξ vanishes. pear. When there is no spontaneous activation, the Meanwhile, the perpendicular correlationk length is model exhibits a directed-percolation transition, marked given by ξ = 1/ ln[σ2(1 + √β)2], which implies that by a divergence in the dynamical susceptibility, a re- the perpendicular⊥ − correlation length diverges when production number of 1, power-law distributed outbreak 8 sizes and the appearance of a giant component. However, avalanches have only been reported for waking states, the introduction of spontaneous activation means that when animals presumably are stimulated by sensory in- the dynamical susceptibility no longer diverges, and that put [50, 54, 58]. In light of our work, this might indicate the reproduction number is shifted. Nonetheless, by con- that criticality is only attained with the help of external sidering the cluster-size distribution and statistics related drive. to the cluster size, a critical line — exhibiting universal Our work addresses both (i) and (ii). For (i), we show curve collapses and finite-size scaling — can be defined. that a susceptibility χ based on causal webs accurately This means that even in the presence of spontaneous ac- identifies the critical line, with the excellent finite-size tivity, a genuine phase transition exists. The introduc- scaling one expects from a true phase-transition. This tion of spontaneous activity destroys the transition to the offers a well-defined observable that is coherent in the absorbing state, and shifts the phase line into the univer- present of spontaneous activity, unlike the branching ra- sality class of undirected percolation. We showed numer- tio or a global measure like the dynamic susceptibil- ically, on a variety of relevant network topologies, that ity. As for (ii), we show that power-laws (with expo- in the largest clusters as merging becomes increasingly nents that vary based on network structure) are present dominant it is the undirected-percolation exponents that at the critical point for any level of spontaneous acti- dominate. Although all the networks we considered were vation. While power-law statistics can also appear in nominally directed, we expect that our results survive on non-critical systems with simpler dynamics as a recent undirected networks, as our small-world networks were critique showed [59], the presence of scaling relations be- comprised predominately of bidirectional connections. tween the critical exponents [9] is only true in pure critical systems. We demonstrate such scaling relations in our model with coexisting power-law regimes, showing B. Critical brain hypothesis for the first time that neural networks with spontaneous activity can still be genuinely critical. Additionally, our Our results have repercussions for the critical brain hy- prediction of two power-laws may help to explain the va- pothesis. Although the hypothesis itself is not new [45, riety of exponents fitted to power-laws in the literature. 46], it gained traction with the seminal work on neuronal Only recently other works have started to pay atten- avalanches [16, 21] and with the development of large- tion to the role of spontaneous activity for the critical scale brain recording techniques. It is still a highly de- brain hypothesis [28, 30, 44, 60–63]. Most of the pre- bated topic [19, 20], and recent work has focused mostly vious definitions and requirements for criticality cannot on (i) the appropriate definition of an order parameter be satisfied in the presence of spontaneous activity since and its tuning and (ii) whether neural activity distribu- time-scale separation is barely satisfied for any realis- tions show critical power-law statistics. tic activity rate. Several attempts have been made to Regarding (i) the main objective is to find a plausi- recover the original definition of criticality and power- ble mechanism by which the brain is able to tune its law statistics by accurately choosing and exploring the own activity to a critical point; ongoing research focuses appropriate temporal bin size for the avalanche defini- on self-organized criticality [47, 48], excitatory-inhibitory tion [14]; but it is still not clear whether that approach activity balance [49], up and down states [50, 51], adap- really recovers the same underlying dynamics, and might tive mechanisms [52], and learning [53] amongst others. only hold if the system is exactly at the critical point Many of these concepts are deeply related, and all of them and for intermediate size systems. What is clear how- might play a role. Regarding (ii), early work focused on ever, is that in the thermodynamic limit, with a fixed whether the measured statistics followed a real power- spontaneous activation rate, there will always be unre- law or just an approximation, and whether the activity lated avalanches occurring, and so avalanches defined by was sub- or super-critical instead [54]. However, this is a global observables (such as fluctuations in Φ or those de- challenging issue to solve experimentally due to the role limited by global quiescence) are not well-defined. As of finite-size and sub-sampling effects [55–57], and due shown in the present work, in the absence of time-scale to the real lack of separation of time scales as we report separation, it is essential to know the network structure here. The initial reports on neuronal avalanches reported to resolve the underlying dynamics. There is currently an exponent close to τ 1.5 for the size distributions, no way to recover the correct exponents without access consistent with a mean-field≈ branching process [9, 16], to the network structure, but approaches that first try but experiments on a variety of neural systems have re- to infer the structure from the dynamics appear to be ported a variety of exponents, usually in the range 1.2 promising [64]. - 2.5. While this discrepancy might be partly explained Another key point that will have to be tackled in the by the technical challenges in probing the tails of power- future in relation to the critical brain hypothesis in the law distributions, or attributed to to variation in network presence of spontaneous activity has to do with infor- topology between studies and heterogeneous dynamical mation transmission. Can it still be optimal in this properties [14], there yet remain reasons to think sponta- regime? Spontaneous activity can indeed enhance in- neous activation has a significant bearing on the critical formation transmission from a sensory system [65], or brain hypothesis. For instance, critical in vivo neuronal shift an inherently sub-critical system closer to the criti- 9 cal point (see Fig. 5). However, a read-out of this activity Finally, our SIS-based model describes a population might require an error-correction code to be present. that can be readily reinfected by a disease, i.e. with no acquired immunity. Although this is not representa- tive of all diseases, the inclusion of immunization (as in C. Aspects of disease spreading the SIR model) does not impact our conclusions about the universality class of disease outbreaks with sponta- Much study of spreading on complex networks has neous infection. In the appendix (cf. Fig. 15 and associ- been driven by a desire to study disease spreading in ated text), we consider a variant of our model that does human contact networks, and we have borrowed heavily not allow for the reinfection of nodes, and find outbreak from that discipline in this paper. We would be remiss distributions that exhibit a transition between directed- not to mention that a small number of studies have re- and undirected-percolation power-laws. This is a conse- laxed the patient-zero assumption of the typical spread- quence of the fact that immunization only plays a role ing process, by including multiple initial spreaders [66– when a node would be reactivated i.e. after activity tra- 72]. A limited few have reflect a disease reservoir that can verses a loop in the network. As we have shown with our cause new outbreaks even as old ones spread [31, 73, 74]. tree-like (loop-free) analysis of the SIS case, these loops To the best of our knowledge however, the question of do not play a significant role in the mean-field limit. Im- the universality class of spreading as new spreaders are munization may play a stronger role in undirected net- introduced has not yet been addressed in the literature. works or directed networks with many short-range loops. By endowing a spreading process with a spontaneous infection rate, we describe disease with an off-network reservoir. This might be an appropriate description for D. Other applications diseases like Zika virus, which can spread via human sex- ual networks, but also “off-network” via mosquito [75]. The generality of our model makes it applicable to Other zoonotic diseases lack an obvious patient zero, as other systems where the timescales of spontaneous acti- they exhibit periodic reintroduction from animal reser- vation and propagation of activity are comparable, e.g., voirs [76]. For example, two-thirds of new cases in the rumour spreading on social networks [5] or the distribu- 2018 Democratic Republic of the Congo Ebola outbreak tion and propagation of computer viruses [4]. It remains could not be linked to existing cases, potentially repre- to be seen how our findings translate to self-organized senting hidden network links or new outbreaks originat- systems and in particular to those that are known to ex- ing from contaminated bush meat [77, 78]. In this case, hibit a self-organized critical (SOC) regime under time- the spontaneous infection rate could represent either in- scale separation [79–81]. Previous studies have shown fection from the environment, or a first order approxima- that a sufficiently high driving rate can induce a transition for a failure in contact-tracing. tion from avalanche dynamics to continuous flow in SOC Our model predicts that a non-zero spontaneous in- systems [82]. Within the context of zoonotic diseases, a fection rate lowers the epidemic threshold. One con- multilayer network approach that incorporates human- sequence is that the local reproduction number can be animal interactions directly [83] or including cooperative less than one even in the epidemic phase. Further, for diseases [84] might open the door for novel dynamics in zoonotic diseases, the average outbreak size near the epi- the absence of time-scale separation. It would also be demic threshold will not follow the typical directed perco- interesting to see which metrics, such as k-shell decom- lation scaling, but rather the isotropic percolation scal- position [85] or a local analysis [86], identify significant ing exponents. Perhaps more importantly, close to the spreaders in our model and if network interventions, such directed-percolation limit, a small change in the sponta- as those connected to explosive percolation [87–89] and neous infection rate can have a drastic impact on the av- others [90], could be used to stymie or promote epidemics erage outbreak size. This suggests that a dual approach in the presence of spontaneous activation. These remain to epidemics, targeting both the transmission between exciting challenges for the future. individuals and the initial routes through which diseases enter the population, may represent a more efficient allocation of epidemiological intervention. For example, one IV. CONCLUSIONS can imagine disease control as a constrained optimization problem attempting to minimize Φ, with finite financial Spreading processes on networks frequently appear in resources for disease interventions that affect p and q. natural and human systems. The inclusion of sponta- To compute Lagrange multipliers and find an optimal neous activity changes the phase transition in these sys- ∂Φ ∂Φ allocation requires calculating ∂p and ∂q . We analyt- tems from directed percolation to isotropic percolation, ically do this for k-regular networks, but our approach because previously independent streams of activity can can be directly adopted to networks more pertinent for merge together. These universality classes have differing disease spreading. By mapping real diseases onto our critical exponents, meaning that diseases with sponta- “normal-form”, a concrete optimization problem can be neous infections (e.g. zoonotic diseases) will show dif- constructed. ferent growth profiles near the critical point. This also 10

2 (k 10)/(2 0.5 ) has several implications for the critical brain hypothesis. out-degree distribution p(k) e− − × and con- Global quantities – such as the active fraction and its nect to other uniformly drawn∼ nodes in the same module. susceptibility, the branching ratio, or avalanches defined For the next αnc nodes, we draw from the same out- by global periods of quiescence – do not capture criti- degree distribution, but connect to the first 102 nodes cal behaviour when spontaneous activity is considered. in the other modules, according to the module backbone As such, criticality in the brain should be re-assessed wiring. using measures that tolerate spontaneous activity. This requires that the use of network structure (e.g. tractogra- phy) be paired with dynamics measurement (e.g. fMRI). Simulation of model on finite networks Proximity to criticality should be assessed using an order parameter based on causal webs, such as susceptibil- Networks are initiated with no active nodes. Each ity (χ). Other measures of criticality, like the branching time-step, the number of nodes that will activate is drawn ratio, might lead to critical behaviour being interpreted from the binomial distribution, with activation proba- as sub-critical. If the brain as a whole is critical, then bility p. That number of nodes are randomly selected the largest avalanches will have isotropic, rather than di- with uniform weighting, re-drawing duplicates. Nodes rected, percolation critical exponents as merging becomes that activate spontaneously and had no active parents the dominant growth mechanism. in the preceding time-step initiate a new cluster. Then, all nodes that had active parents in the preceding time step that were not already activated spontaneously are ACKNOWLEDGMENTS checked for activation. Each node with m active parents in the previous time-step is activated with probability 1 (1 q)m. Nodes inherit the cluster label of their par- JD, DJK and JGO acknowledge helpful discussions − − with Rashid Williams-Garcia. This project was finan- ents. If a node would inherit more than one cluster label, cially supported by NSERC (JD, DJK, JGO), the Eyes then those clusters are merged into a single cluster by re- High Initiative of the University of Calgary (JGO, JD), labelling all nodes belonging to the smaller cluster with Alberta Innovates (DJK) and the Korea-Canada Coop- the label of the larger cluster. Clusters that are found to erative Development Program through the National Re- have no active nodes in a given time-step are terminated, search Foundation of Korea (NRF) funded by the Min- and their size, duration, and number of roots recorded. istry of Science and ICT, NRF-2018K1A3A1A74065535 (S-WS, JD) and Basic Science Research Program through NRF-2017R1D1A1B03032864 (S-WS). S-WS also ac- Simulations on infinite k-regular networks knowledges the support and hospitality of the Depart- ment of Physics & Astronomy at the University of Cal- For the infinite networks, we begin at a randomly se- gary during his visit to Canada. lected active node. We then check its immediate neighbours to see whether they are part of the same cluster. For those that are included, we then check their neigh- Appendix A: Methods bours for inclusion. We can perform this process such that we need only count the number of unexplored neighbours, of which there are two types: (I) daughters that Network generation haven’t been checked for inclusion and (II) parents that are known to be included, but whose neighbours haven’t For the finite networks, we generate finite directed k- been checked. If we’re beginning from a root node, there regular networks via the configuration model, shuffling are initially k unchecked daughter branches (type I). If connections to avoid self-links and multi-links. To gen- we’re beginning from a randomly-active node, we begin erate power-law networks we employed a variation of with one type (I) neighbour and one type (II) neighbour. the configuration model described in [91], with a de- The algorithm proceeds to check each unevaluated con- in/out 3.5 gree distribution p (k) = k− with a domain k nection (of type-I or type-II), possibly adding more as ∈ [5, ..., 1000], with rejection parameters κ =0.5, δ =0.05. it goes, until none remain or the cluster exceeds a given We generate the small-world networks using a directed size (typically 1010). Each type of connection is added network generalization of the Watts-Strogatz model [34], as follows: 2 using rewire probability 10− and with average degree 3 • 10 (a rewire probability of 10− is also considered in Ap- (Type I): We check each type-I, by assuming it pendix F). We generate the hierarchical modular net- has md other active parents (drawn from a bino- k 1 md k 1 md works described as “HMN-2” in [22] as a backbone for mial distribution P (md)= − Φ (1 Φ) − − md − our modular networks. Within each base module with nc of k 1 other parents, activated with probability 2 − connections in the module backbone, we place 10 + αnc Φ, the active fraction, given by Eq. (B11)), and in- nodes with α = 4 being the four inter-modular connect- clude each type-I with probability 1 (1 p)(1 ing nodes. The first M = 102 nodes we draw from the q)md . If it is included, then we add −k type-I− con-− 11

nections from this daughter and m type-II connec- d TABLE II. Table of fit parameters for the critical lines of tions. a Fig. 6. Fits are power-law fits of the form p = c(qc,DP − q) . • (Type II): Each type II is included with probabil- Entries for the k-regular correspond to the low p approximation given in Eq. (3). ity 1. It adds k 1 additional type-I connections, and m type-II− parents, with m drawn from the Fit parameters p p Network c a q distribution in Eq. (A1): C,DP Small-world, Rewire = 10−2 6.00 2.97 0.112665 Small-world, Rewire = 10−3 7.99 3.09 0.117104 k mp k mp Φ (1 Φ) − mp Hierarchical Modular Network 2.78 1.94 0.10482 p(m )= − (1 (1 p)(1 q)mp ). p Power-law network 2.29 2.94 0.131118 Φ − − − 5 (A1) k k-Regular Network (2k−1)(k−1)2 3 1/k The probabilities of adding a daughter or parent are as derived in the appendix. For the purposes of measur- ing the two-point connectedness function, the above algo- appearance of the giant component or the diverging sus- rithm can be easily extended to also include the number ceptibility. of time-steps, by simply tracking how many times each Instead, we employ the method of de Souza et al. [92] 2 s c active front has followed a daughter branch or a parent and consider the quantity B = g h i2 , which was shown s c branch. to have no finite-size dependenceh ati the critical point.

Therefore, for ﬁxed p, the critical point qc can be found P"#$ %&'( )

10-2 as the intersection point of the B(q) curves for diﬀerent H*+

-3 ,. 1345 678 9:; <> ?@AB C D E

-2 values of N (see Fig. 12a). This enables the numerical FG IJKL MNO QRT UV WXYZ [ \ ] 10-3 ^_`a b cde f determination of the p, q critical lines for the networks discussed in the main text (cf. Fig. 6 and Table II). 10-4 p

10-5

10-6 Appendix B: Generating Functions

10-! Before deriving the generating functions associated 0.08 0.085 0.09 0.095 0.1 0.105 0.11 0.115 0.12 0.125 0.13 with the average cluster size, we begin with a brief re- q view of probability generating functions (PGFs). For a discrete random variable X drawn from the probability FIG. 6. Numerically determined critical lines for various networks. Points correspond to the avalanche simulations plot- mass function p(x), the probability generating function ted in Fig. 3 in the main article, except for the small-world can be deﬁned as network with a re-wire of 10−3, which is only studied in the appendix. Solid lines are approximate ﬁts of the form X ∞ x a gX (z)= E z = p(x)z . p(q)= c(qc,DP − q) for constant c, a, and qc,DP , whose val- x=1 ues are summarized in Table II. k-regular phase curve is exact. X

gX generates the probability p(x) in the sense that th 1 (n) gX (0) = p(0), and the n derivative yields: n! gX (0) = p(n). The probability generating function can be used to Critical point determination obtain the moments of X, as X = g′ (1), X(X 1) = h i X h − i gX′′ (1), and so on. The final property of probability gen- Accurate determination of the critical point is neces- erating functions we will use is perhaps its most use- sary to effect accurate finite-size scaling. In the case of ful: when a family of independent and identically distributed variables X1,X2,...XN generated by gX (z) the k-regular network, the critical point can be deter- { N } mined analytically. However, for the power-law, small- are summed Y = Xi, with N also being a random Y world, and hierarchical modular networks, determination variable generated by gN (z), then gY (z) = E(z ) = NX P X n of the critical point can be done in two ways. The naive E(z )= n∞=1 p(n = N) E z = gY (gX (z)). Al- approach is to simply tune q for fixed N and p until though this may seem esoteric, it means that the sum of P power-laws appear in the avalanche distribution. How- a collection of some random number of random variables ever, this is prone to finite-size effects: for fixed p and can be concisely expressed using generating functions. N, the largest power-laws in the P (s) distribution will We will derive the PGFs corresponding to the cluster appear at the pseudo-critical point, corresponding to a size distribution beginning from randomly selected active larger qc(N) value than the true qc(N ). Finite- sites on a random network. If there are no loops in the size effects similarly cripple approaches based→ ∞ on just the network (the tree-like approximation), we can express the 12

a. b. c.

0 0 0 sp,1 1 1 1 Hd

of k parents Bi(x) np active

I 1 x

of k daughters Ao(x) nd active

0 0 sd,1 sd,2 1 1 Hp Hp

nd np np nd s =1+ Pl=1 sd,l + Pm=1 sp,m H0(x)= x Ql=1 Hd(x) Qm=1 Hp(x)

= xBi(Hd(x))Ao(Hp(x))

FIG. 7. A ﬁring pattern example represented both as the sum of variables and the product of generating functions. a. An example activation pattern beginning from a randomly selected initial active node, I, on a 4-regular network. Thick edges indicated connected active nodes. b. The number of parent and daughter edges contributing to the cluster are labelled by np and nd, random variables that could vary from 0 to k. Nodes are labelled with their size contribution to the cluster – nodes that do not activate contribute zero, the initially considered node contributes one activation, while active parents and daughter contribute a random variable. c. The same activity pattern labelled with the the probability generating functions corresponding to each random variable.

cluster size s starting from a random active as the following self-consistent equation:

nd np Hp(x)= xAo(Hp(x))Ai(Hd(x)) . (B3) s =1+ sd,l + sp,m , (B1) Similarly, an active parent of I, here labelled X, will have m=1 Xl=1 X one fewer daughter branch to consider, so its cluster size contribution is where nd is the number of active daughters of the initial site, np is the number of active parents of the initial site, np n˜d th sd,l is the size of the cluster reached from the l active sp =1+ sp,l + sd,m , daughter, and s is the size of the cluster reached from p,m l=1 m=1 the mth active parent. This means that the PGF for the X X total cluster size s is wheren ˜ ranges from 0 to k 1, and counts the daughters d − other than I. sp is generated by Hd, which obeys the H0(x)= xAo(Hp(x))Bi(Hd(x)) , (B2) following self-consistent equation: for the PGFs Ao (generating nd), Bi (generating np), Hd(x)= xBo(Hp(x))Bi(Hd(x)) . (B4) Hp(x) (generating the sd,l) and Hd(x) (generating the sp,m). The connection between the activation pattern The relationship between the three size generating func- and Eqs. (B1,B2) is illustrated in Fig. 7. tions, H , H , and H are illustrated in Fig. 8. An active daughter of I, here labelled Y , will have 0 d p To summarize, H corresponds to the cluster size the number of parent branches that can be considered p reached when arriving at a node from one of its parent reduced by one, so, branches and Hd corresponds to the cluster size reached

n˜p nd when arriving at a node from one of its daughter branches. The two pairs of generating functions (Ai, Ao) sd =1+ sp,l + sd,m , l=1 m=1 and (Bi, Bo) describe the number of active neighbours X X for Hp and Hd respectively. In terms of the nodes la- wheren ˜ ranges from 0 to k 1 and counts the parents belled in Fig. 8, the neighbour generating functions and p − other than I. sd is therefore generated by Hp which obeys their corresponding probability mass functions are: 13

a. b. c.

X I

of k parents of k parents of k − 1 parents np active np active n˜p active

I X Y

of k daughters of k − 1 daughters of k daughters nd active n˜d active nd active

Y I

nd np n˜d np nd n˜p s =1+ Pl=1 sd,l + Pm=1 sp,m sp =1+ Pl=1 sd,l + Pm=1 sp,m sd =1+ Pl=1 sd,l + Pm=1 sp,m H0(x)= xAo(Hp(x))Bi(Hd(x)) Hd(x)= xBo(Hp(x))Bi(Hd(x)) Hp(x)= xAo(Hp(x))Ai(Hd(x))

FIG. 8. An example of the three size generating functions. To illustrate the difference between H0, Hd, and Hp we consider the example firing pattern of Fig. 7, but centred on three different nodes, X, I, and Y . a. Corresponding to H0, and beginning from a randomly selected initial active node, I, on a 4-regular network. Thick edges indicated connected active nodes. Two neighbouring active nodes, a parent and daughter (X and Y respectively) are highlighted as corresponding to the other two generating functions. b. An example activation pattern near X. Since one daughter connection leads to I, only k − 1 are available for other connections. This restriction on daughters is why Hd differs from H0. c. An example activation pattern near Y . Since one of the parents of Y is I, only k − 1 other parents need to be considered. In part due to the restrictions on parents, Hp differs from H0.

A P (˜n active parents of Y excluding I Y active & I active) (B5) i ←→ p | A P (n active daughters of Y I active) (B6) o ←→ d | B P (n parents of X X active) (B7) i ←→ p | B P (˜n active daughters excluding I X active) (B8) o ←→ d |

As the giant component appears, the average cluster Ao′ (1))(1 Bi′(1)) Bo′ (1)Ai′ (1). This condition will yield size diverges. Therefore, identifying the conditions under the critical− line, when− supplied with the PGFs for A and which the average cluster size diverges is a natural way B. to identify the critical line. For q q , when H (1) = ≤ c 0 Hp(1) = Hd(1) = 1, the average cluster size is given by 1. Neighbour generating functions for k-regular s = H′ (1)= 1+ A′ (1)H′ (1) + B′(1)H′ (1) , (B9) h in 0 o p i d networks where the subscript n denotes an average conducted by sampling randomly selected nodes instead of averaging So far, we’ve been quite generic in developing the gen- over clusters. Now, since Ao′ (1) and Bi′(1) correspond to erating function H0. To proceed further, we must sup- the mean number of daughters and parents of the initial ply Ai/o and Bi/o for a given network. For simplicity, randomly selected node, quantities that are necessarily we focus on the k-regular network. This will allow us bounded above by the mean (in/out)degrees, Ao′ (1) and to develop expressions for Φ, the active fraction, and B′(1) cannot diverge. Therefore, s n can only diverge P , the probability that the daughter of an active site i h i d if Hp′ (1) or Hd′ (1) do. Using Eqs. (B3) and (B4), the activates in the next time step. The first quantity we following self-consistency relation for Hp′ (1) and Hd′ (1) will need is the active fraction – the proportion of nodes (with q qc) can be obtained. activated in each time step. A randomly selected (not ≤ necessarily active) node will have m active parents with 1 Ao′ (1) Ai′ (1) Hp′ (1) 1 k m k m − − = . (B10) probability Φ (1 Φ) − , as each parent is inde- B′ (1) 1 B′(1) H′ (1) 1 m − o − i d pendent. With m parents,− the probability of activating m Therefore Hp′ (1) and Hd′ (1) diverge when the determi- is 1 (1 p)(1 q) . Now since the probability of acti- nant of the above matrix is zero, i.e., when 0 = (1 vation− for− a random− node is also Φ, we can write (using − 14 the notation p =1 p to denote complementary proba- Note that σ = kP defines the branching ratio. − d bilities) the self-consistent equation: Now that we have both Pd and Φ, we can derive Ai/o and Bi/o. The simplest to derive are Ao(x) and k k k m Bo(x), because they describe the number of activated Φ= ΦmΦ − (1 p qm) daughters, and the activation of each daughter is in- m − · m=0 dependent of the others. Considering a single daugh- X k =1 p qΦ . (B11) ter, whose activation can be described by a single ran- − · dom variable m 0, 1 , with m = 1 only if the sin- ∈ { } It will be useful, when performing asymptotic analysis in gle daughter activates. The PGF corresponding to m is C(x)= E(xm)=(1 P )x0+P x1 = P +P x. If n is the the limit that p 0, to have a closed-form approximation − d d d d for Φ. If we assume→ that Φ 1, we can truncate the number of activated daughters for a site with l available l k≪(k 1) 2 2 expression Φ= p(1 kqΦ+ − q Φ + ...) to first or daughters, then n = i=1 mi for ml being independent − 2 and identically distributed (iid) Bernoulli variables gen- second order in Φ and solve for Φ, from which we obtain P erated by C(x). Then, taking l = k for Ao, we have the first order approximation l n P =1 mi k mi k Ao(x)= E(x )= E(x i )= E(x )= C(x) , p i=1 Φ (B12) so ≈ 1 kq Q − k Ao(x)= Pd + Pdx . (B15) and the second order approximation (choosing the posi- tive root, since Φ > 0) For Bo, we have one fewer daughter from which to choose, because we arrived at the node in question by means of kpq 1+ 1 k2p2q2 2kpq(1 pq) one active daughter, so we take l = k 1 to find − Φ − − −2 − . (B13) ≈ p (k 1)kpq k 1 − Bo(x)= Pd + Pdx − . (B16) For P , we have one active parent, and k 1 parents that Now, for Ai and Bi, we cannot treat the parents’ acti- d are independently active with probability− Φ. Hence, vation as independent. This is because we must condition on the knowledge that their daughter must activate and, k 1 in the case of A , also on the presence of other active − k 1 k 1 m i P = − ΦmΦ − − 1 p qm+1 parents. d m − · m=0 Treating Bi(x) first, we are considering an active site X k 1 =1 p q qΦ − (labelled X) that we arrived at by means of an active − · · daughter (in Fig. 8, I). Therefore, we have no knowledge and simplifying using Eq. (B11) about the number of active parents, save for the fact that they successfully activated the node in question. Consid- q Φ ering the probability mass function in Eq. (B7), Bayes’ Pd =1 · . (B14) theorem allows us to write − qΦ

P (X active n active parents)P (n active parents) P (n active parents of X X active) = | p p . p | P (X active)

np However, P (X active np parents) = 1 p q by deﬁ- responding to Bi is therefore given by nition of the model (Eq.| (1) of the main− text), while k the probability of np active parents, unconditioned on B (x)= P (n active parents of X X active)xnp anything else is just given by P (np active parents) = i p | k np np=0 k Φnp Φ − . Lastly, the probability that X is active, X np conditioned on nothing else, is just the active fraction Φ. and can therefore be expressed as Thus, 1 k k Bi(x)= Φ+Φx p Φ+Φqx (B18) P (np active parents of X X active) Φ − | k np h i (1 p qnp ) k Φnp Φ − np For Ai(x), we are considering a node, Y , that we ar- = − , (B17) rived at from an active node (labelled I in Fig. 8) that is Φ one of Y ’s parent branches. Ai(x) is the generating func- which is exactly Eq. (A1). The generating function cor- tion for the number of additional active parents of Y . 15

Considering the probability mass function in Eq. (B5), and applying Bayes’ theorem,

P (˜n active parents of Y excluding I Y active & I active) p | P (˜n of the k 1 parents other than I active ) = P ( Y active I active &n ˜ other active parents of Y ) p − . | p × P (Y active I active) | Each of these probabilities are known.

P ( Y active I active &n ˜ other parents of Y active) = 1 p qn˜p+1 (B19) | p − by deﬁnition of the model (Eq. (1) of main text),

k 1 k 1 n˜p P (˜n of k 1 parents other than I active ) = − Φn˜p Φ − − , (B20) p − n˜ p and P (Y active I active) = P . Hence, | d

1 k 1 k 1 n˜p P (˜n parents of Y other than I active Y active & I active) = 1 p qn˜p+1 − Φ n˜p Φ − − . (B21) p | P − n˜ d p Now, the generating function Ai is given by

k 1 − A (x)= xn˜p P (˜n active parents of Y excluding I Y active & I active) , i p | n˜Xp=0

so after some algebra we have tibility χ, and cluster distribution Pc(s) from the generating function H (x). Practically speaking, we solve 1 k 1 k 1 0 Ai(x)= Φ+Φx − p q Φ+Φqx − . Eqs. (B3) and (B4) self-consistently for H (x) and H (x) P − d p d via a Newton-Raphson scheme for a given set of model h i This concludes the calculation of the four generating parameters p, q, and x. With Hd(x) and Hp(x) in hand, functions Ai/o and Bi/o for the k-regular network. These we can insert these into Eq. (B2) and obtain H0(x). calculations can also be conducted for other random networks, although the calculation is more technically involved when the in-degree can vary or correlations exist between the in- and out-degrees. The first quantity we can obtain from H0(x) is the In summary, and in terms of Φ and Pd, the PGFs Ai/o fraction of nodes involved in finite clusters, which is just and Bi/o for the k-regular network may be expressed as H0(1) = p(s)s = 1 P . So the giant component s − ∞ k fraction g, the fraction of all nodes at all times that are Ao(x) = (Pd + Pdx) , (B22) P part of the infinite cluster, is just g = Φ(1 H0(1)). For the susceptibility, χ = s2 = s2p (s),− we must B (x) = (P + P x)k 1 , (B23) c o d d − make the distinction betweenh thei cluster size distribution p (s) (for numerical simulations,P reported simply as 1 k 1 k 1 c Ai(x)= Φ+Φx − p q Φ+Φqx − , P (s)) and the per-node cluster size distribution Pn(s). Pd − The latter describes the cluster sizes observed by sam- h i(B24) pling random active nodes, and is directly calculated and by the generating function approach, or accessed by 1 k 1 k Bi(X)= Φ+Φx − p Φ+Φqx . (B25) simulating avalanches on the infinite lattice. Clearly, Φ − Pn(s) = AsP (s), for a normalization factor A. Since h i 1 1 P (s) = 1, A = 0 H0(x)dx. So, χ = A sPn(s)s = ′ 2. Observables from the generating function s n H0(1) h i = 1 . Of course, we can directly access Pn(s) PA R0 H0(x)dx R P s 1 d H0(x) by using Pn(s) = ! d s . As was pointed out in Here, we summarize how to extract observables, such s x x=0 as the size fraction of the giant component g, suscep- [35], numerically evaluating this derivative for large s is

16 most easily accomplished via a contour integral Solving 0 = (1 σ)2 σ2β for q, and assuming Φ 1 − − ≪ (as occurs in the p 1 limit with q < qc) yields s 1 k 1 ≪ d H0(x) 1 H0(z)dz p = −2 √2k 1√Φ. Inserting the ﬁrst-order closed = , (B26) k − 1 k − dxs x=0 2πi zs+1 form approximation for Φ 1 (Eq. (B12)) into the solu- H tion for q yields the small p≪expansion for the phase-curve iφ d on the circle z = e for φ [0, 2π]. z becomes highly (Eq. (3) in the main text) oscillatory at large d, so convergence∈ of this integral can be improved via standard numerical techniques for oscil- 3 1 (k 1)2(2k 1) latory integrals [93]. The cluster probability distribution q = − − p . (B29) 1 k − k5 can then be accessed as P (s)= As Pn(s). The average cluster size s (and therefore suscepti- h in 3. Phase-diagram for the k-regular network bility χ) diverges for (p, q) near to points on the criti- γ cal line (pc, qc) as s n pc p − (for q = qc) and s q q γ (forh ip =∼p |) with− γ| = 1. This is a direct We can study the divergence of χ s by solving n c − c n consequenceh i ∼ | − of| the fact that the numerator and denom- Eq. (B10), and inserting the solution∼ into h i Eq. (B9) to inator of Eq. (B27) cannot both be simultaneously zero obtain (except for the degenerate q = 1 case). Hence, the be- σ 1 (σ σm) haviour near the critical line will depend only on how the s = − k − (B27) 2 k 1 n 2 k 1 denominator f(p, q)=(1 σ) − σσm scales near its h i (1 σ) − σσm − − k − − k zero p , q . As ∂f = 0 and ∂f =0at(p , q ), the Taylor c c ∂p 6 ∂q 6 c c where σ = A (1) = (k 1)P = (k ∂f ∂f m i′ p1 series approximation f(p, q) ∂p (p pc)+ ∂q (q qc). 2 − − ≈ − − Φ (1 Pd) 1) 1 − is the expected number of other ac- Choosing p = pc or q = qc immediately yields the power- Pd 1 Φ − − law scaling exponent γ = 1. This divergence can be tive parents, to an active node with one already known visualized in Fig. 9. parent. That is, σm describes the rate of merging of initially independent clusters. Clearly, s n diverges when 2 h i k(1 σ) (k 1)σσm = 0 (Eq. (2) of the main text). This− result− could− also have been arrived at by setting the 4. The giant component determinant of Eq. (B10) to zero. A reparameterization that will be convenient when considering the correlation The giant component fraction g is given by g = Φ(1 (k 1)σm length is to replace σ with β = − , meaning that − m kσ H0(1)) = Φ 1 Hd(1) Pd + PdHp(1) . At the critical the critical line diverges when − point, Hp(1) = Hd(1) = 1. Sofor δ = q qc 1, we have − ≪ ∂Hp ∂H 2 2 d (1 σ) = σ β . (B28) that g ΦPd ∂q +Φ ∂q δ. Since both Hd(1) and − ≈ Hp(1) are strictly decreasing functions of q, g (q qc) ∼ − The set of (pc, qc) that cause this divergence deﬁne a identifying the critical exponent β = 1. critical line (see Fig. 5a in the main text).

6 10 Appendix C: Mergeless Avalanches

105 The mergeless clusters are exactly those clusters with 4 10 one root. The number of conﬁgurations of singly rooted

3 -1 10 (q-qc) clusters of size s is given by the Fuss-Catalan numbers n 〉 (k) ks s 1

〈 Cs = (k 1)s+1 s , which count the number of incom- 102 − p=10ij plete k-ary trees with s vertices [94]. Such a tree has 101 perimeter (unoccupied branches) of length t = (k 1)s + − p=10-2 1. Nodes are included in the tree with probability 100 k 1 − 10-1 Pd1 = Φ (1 p q) , (C1)

10-7 10g h 10-5 10-4 10-3 10-2 10-1 − qc - q denoting the probability that a given daughter node is activated while having exactly one parent. The excluded FIG. 9. Average cluster size for k-regular networks approach- nodes on the perimeter occur with probability Pd, which ing critciality. Numerical simulations (points represent mean is the probability of not activating, despite having an ac- of 106 realizations) on an inﬁnite 10-regular graph yield good tive parent. Hence, the probability of observing a merge- agreement with analytical predictions (lines) for hsin (k) s 1 t less cluster of size s is given by P (s) = Cs Pd1− Pd . 17

Applying Stirling’s approximation we find In equilibrium critical points, divergence in the correlation length is associated with a divergence in the sus- s ceptibility of the order parameter to an infinitesimal ap- 3/2 Pd1k 3/2 s/sm P (s) s− = s− e− , plication of an external field. In directed percolation, the ∼ ((1 1 )/P )k 1 − k d − ! order parameter is Φ. The susceptibility measures activ- (C2) ity of the system in response to an external stimuli. We where the characteristic mergeless size is can imagine that the external stimuli is an infinitesimal increase in the average spontaneous activity of the sys- k 1 Pd − tem, and hence we can define the dynamic susceptibility sm = 1/log kPd1 . ∂Φ − 1 1 as χ0 . So, using Eq. (B11) we find: " − k # ≡ ∂p k k 1 1 χ0 = Φq + qkp Φq − χ0 , In the limit of p 1 and k q 1, expressing Pd ≪ − ≪ k+1 (Eq. (B14)) and Pd1 (Eq. (C1)) in p, q, and Φ, and again p Φq using the closed form approximation for Φ (Eq. (B12)), = χ0 = k , we find to lowest order in p and ( 1 q): ⇒ p(qΦ qkp qΦ ) k − − and simplifying with Eq. (B11) we obtain 1 3 2 1 (2k 1)(k 1) 1 − k 1 s− − − p q + q , Φ qΦ m ≈ k2 k − 2(k 1) k − χ0 = . (D1) p(qΦ qkΦ) − (C3) − and if we apply Eq. (3) to observe how the cut-off scales In the limit p 0 with Φ 0 and q 1/k, χ0 is, asymp- → →1 → on the critical line, we find: 1 1 − totically, χ0 k k q , and therefore diverges at the ∼ − 1 2 directed percolation critical point q = and p = 0. This 2(k 1) 1 − k s − q , (C4) susceptibility has been studied in the context of neural m ≈ 3k3 k − systems, where the mixing of initiation and spreading time-scales means χ0 no longer diverges (cf. Fig. 10), 1 which was Eq. (4) from the main text. Since qc,DP = k but instead is maximized on a quasi-critical “Widom” 1/σDP line, where the fluctuations Var(Φ(t)) are also maximized and the relation sm (qc,DP q)− defines the directed percolation exponent∼ σDP− , we have also recovered [44]. DP 1 the usual directed percolation exponent σ = 2 (cf. Table III). Appendix E: Phase curve scaling

Appendix D: Widom line 100

10-1 1 1 r st q /k

-2 103 10

10-3

1 2 v wx u /k

-4 p 10

0 2 -5 1 3

y z {| χ 10 10 /k

10-6

10n o

1 lm 10 p=10 10-8 Dynamic Susceptibility, 10-4 10-3 10-2 10-1 1 /k - q

p=10-2 〈S〉n }~ σ = 1

100

0.075 0.08 0.085 0.09 0.095 0.1 0.105 0.11 0.115 0.12 0.125 FIG. 11. Power-law scaling of the critical and quasi-critical q lines near the directed percolation limit. From bottom to top, the critical line, σ = 1, and Widom line in the limit p → 0 FIG. 10. The dynamic susceptibility for various p and q as limit. calculated for an inﬁnite 10-regular graph. For each p there is a corresponding q that maximizes the susceptibility. These In directed percolation, there are several indicators of maxima are labelled by the squares, and fall on the Widom- the critical point. The mean cluster size diverges, the line. branching ratio is one, and the dynamic susceptibility 18

TABLE III. Critical exponents for the k-regular network, beyond those reported in Table I of the main text, where δq = |qc − q| and δqDP = |qc,DP − q|, where qc,DP denotes the directed percolation critical point (i.e. at p = 0). Exponents related to temporal dynamics (i.e. α and 1/σνz) reﬂect numerical observations from Fig. 17, while all other exponents are analytically determined. Exponent Quantity This work Literature 2 −γDP δq 3−τ γDP χ = hs ic ∼ δq for p ≫ 1 3 3 [39] (using γ = σ ) −γ δq γ χ ∼ δq for p ≪ 1 1 1 [40] −αDP αDP P (T ) ∼ T for T Tm ≈ 7 - 1 1/σνzDP T ∼ s for ssm ≈ 4 - ν ξ ∼ δq−ν 1 1 [40] 1 −1/σDP sm ∼ δq 2 2 [39] σDP DP −τ 1 P (s) ∼ s G(s/sξ) for s>sm 1 σ − 2 2 [40] and G(x ≫ 1) → 0 then sξ ∼ δq σ

diverges. However, with the introduction of spontaneous a. 102 b. 10-1 101 10-2 activation, it is clear that these indicators no longer agree 100 10-3 10-1 10-4 -2 -5 (cf. Fig. 5a of the main text, or Fig. 11). In fact, the 10 p 10 Ê -3 -6 3.0907 10 10 pc ~ (0.117104 - qc) ¶·¸¹º »¼½¾ qc branching ratio is no longer a clear signal, because in- 10-4 10-7 10-5 10-8

dependent streams of activity can merge together and 10-6 10-9 ¿ÀÁÂÃÄÅÆÇ ÈÉ 10

nodes can spontaneously activate. Meanwhile, the dy- 10-4 10-3 10-2

¡¢ £¤¥¦§ ¨©ª«¬ ®¯° ±²³´µ

q -q ËÌ ÍÎ namic susceptibility no longer diverges, but instead at- q tains a maximum one what is referred to as the Widom c. 104 d. 106 s-1.108 # Roots = 1 104 # Roots > 1 line. Ú 103 2

Ù 10

Ø ×

1.1 0

Ö 10 2 1 Although the Widom line, unity branching ratio (σ = Õ 10 (s) ~ s

Ô 10-2 -2.5

Ó s Ò

Ñ -4 P(s) / p 10 1) line, and line of diverging cluster size all agree as p Ð 101 → Ï 10-6 0, they obey different power laws in their approach to 100 10-8 that point (Fig. 11). In this section, we will derive the 10-3 10-2 10-1 100 101 102 103 10-4 10-2 100 102 104 -0.75 -0.75 different scalings associated with these critical and quasi- ÛÜÝÞßàá â ãä åæç è é ê Rescaled Size, s / p critical lines. 1 2 The scaling for the σ = 1 line is given by k q p, FIG. 12. Critical line and avalanche scaling of small-world which can be seen by solving Eq. (B14) for q −and using∼ network with a low rewire probability. The phase diagram the closed form for Φ (Eq. (B12)), which immediately and curve collapse along that phase line for N = 106 small- 2 2 yields k 1 q p on the σ = 1 line. world networks with the re-wire probability 10−3. a The ratio k 2 k 2 − ≈ hs ic 5 − ∂χ0 B = g hsi2 for p = 10 , identifies the critical point qc = As for the Widom line, by setting ∂q = 0 and applying c some simple algebraic manipulation, the Widom line can 0.115333 ± 0.000010. b The numerically derived critical line be found to consist of the q and p satisfying 0 = 1 is represented with symbols, the black line is the non-linear least-squares fit to the data. c The average number of roots, 2 2 − kqΦ 2Φ + qΦ . The first order approximation for Φ −0.75 d − exhibits a transition that scales with sm ∼ p . This given by Eq. (B12) is poor in the vicinity of the Widom same sm effects a curve-collapse in the avalanche distribution, line (after all, it is in the vicinity of the point of maximum which we see separates the mergeless and merging avalanches. susceptibility in Φ) and so a second order approximation for Φ (Eq. (B13)) is necessary. With this approximation, the Widom line becomes (upon expansion around p =0 and q = 1 ): p k ( 1 q). world networks. In the main-text, we showed that the k ≈ 2 k − The critical line was previously shown (Eq. (3)) to obey avalanche distribution P (s) exhibits a scaling collapse 2/3 3 when assuming s p , for a small-world network the scaling 1 q p. These scalings are illustrated m − k with a re-wire probability∼ of 10 2. This was somewhat in Fig. 11. − ∼ − surprising, as the small-world network still exhibits a large number of recurrent connections and is very nearly a circulant graph. However, for a lower re-wire proba- Appendix F: Phase diagram and scaling collapse on 3 bility, 10− , recurrent connections play an even larger small-world networks 2/3 role, and p− does not provide such a robust scal- 0.75 ing. Instead, we find the collapse is best for p− By numerically determining the critical point for dif- (cf. Fig. 12). This can be understood by considering ferent p values, we can build a critical line for the small- the root-size distribution at the critical point, as is done 19 in Fig. 12c. Clearly, the characteristic merging size sm 2d, because when a daughter branch is followed, t ad- 0.75 scales as sm p− . vances by one, while a parent branch decreases t by one. Interestingly,∼ this also allows us to estimate the di- However, ∆t = 0, so the number of parent and daughter rected percolation 1/σDP exponent for the small-world branches must both be equal. network. Given a phase-line that scales as p (qc,DP That being said, not all routes with d daughters and a a ∼b − q) = δDP and a curve collapse effected by p , and that d parent connections are are equally likely. For instance, we expect the curve collapse to scale as s δ1/σDP , whenever a parental connection follows a daughter con- m ∼ DP we have the scaling relation 1/σDP = ab. The phase nection, the route requires that two initially independent diagram for this small-world network (Fig. 12b) relates avalanches merge at that point. It turns out that the 3.0908 p (qc,DP q) for qc,DP 0.11533, which im- most convenient way to compute g(2d, 0) is to sum over ∼ − 1 ≈ plies σDP 3.09 0.75 = 0.43 for the small-world net- collections of routes that have a fixed number of merges. ≈ × 3 work with a re-wire probability 10− . When the short- The number of routes with m merges is given by cut density is low, the small-world network is approxi- mately a 1-dimensional circulant graph, which suggests 2 2m d 1 d 1 2d 1 k 1 we should use the 1 + 1-dimensional directed percolation + − k − − . (G1) m k 1 m 1 k exponents. Our result of σDP 0.43 compares reason- ≈ − − ably well with the σDP =0.391 reported in the literature [39]. Similarly, we should identify the avalanche exponent Given a sequence of parent / daughter network hops, the d τDP =1.108, which matches well with our numerical re- combinatorial factor m counts the number of ways that sults (cf. Fig. 12). the parent / daughter network hops could be rearranged without altering the number of merges. A network hop that follows a network hop of the same kind (i.e. a parent Appendix G: Correlation length hop following a parent hop, or a daughter hop following a daughter), contributes k possible paths. Every time In the main text, we introduce the pair connectedness a change in direction occurs (i.e. a daughter followed function g(d, t). By consider the (typically exponential) by parent or parent by daughter), only k 1 links are − decay of this pair-connectedness function, we can define available, because one was taken to arrive at the node 2m correlation lengths ξ of the form g(d) exp[ d/ξ]. We in question. The factor of k2d 1 k 1 captures the ∼ − − −k will show analytically that the perpendicular correlation number of possible paths, given a sequence of daugh- 1 d 1 length ξ corresponding to the decay of g(d, 0) diverges ter / parent network hops. A correction of − ⊥ k 1 m 1 on the critical line. is required, to account for those paths with one− or two− fewer change in direction (a boundary condition effect). A daughter or parental connection occurs with weight Pd 1. Divergence of perpendicular correlation length or Pp = Pd, except when a parental connection follows a daughter connection, when it instead contributes Pp1. We can derive the divergence of the perpendicular cor- So, since the number of merges could range from 0 to d relation length ξ by computing g(2d, 0). We consider we can write: ⊥

d 2 2m 1 m d d 1 k 1 − (kP )2d P g(2d, 0) = (k 1) + − − d p1 . − m m 1 × k k2 P m=0 d ! X − This expression is compared to simulations on the inﬁnite lattice in Fig. 13, and has the closed form expansion:

kσ (k 1)2P g(2d, 0) − − p1 = 2(σ (k 1)P ) F (1 d, d; 1; β) + ((k + 1)σ + kP ) F (1 d, 1 d; 1; β) , σd − − − p1 2 1 − − p1 2 1 − −

where 2F1 denotes the Gauss hypergeometric functions [95], we have that 2F1(1 d, d; 1; β) = (1 2 (k 1) Pp1 d − − − tion, and σ = kP and β = −2 as before. β) 2F1(d, d; 1; β/(1 β)). Using an identity owing to d k P d d − − (1+√β) 2F1(1 d, d;1;β) Since in the limit of large d, − − = Wilson [96] we have 2F1(1 d, d; 1; β) √ , and 2F1(1 (d 1), (d 1);1;β) − − ∝ d − − − − 2F1(1 d,1 d;1;β) − − 2F1(1 (d 1),1 (d 1);1;β) , we need only treat one of them− asymptotically.− − − By way of Kummer’s 24 solu- 20

a. 101 b.105

4 -1

0 J KL

10 10 HI c

C ¨ -1 g 10

§ 10

] 2

P 10

-2 \

ÿ 10

[

ù T

ø 101 S

÷ -3 R

ö 10

ñ ð

ï 100 î

í -4

ì 10 ë

10-1

10-5 F G 10-7 10D E 10-5 10-4 10 10-2 10-1

0 10000 20000 30000 40000 50000 n o

m c

N !" #

-4 -5 AB -2 -4

q / 0 1 2 34 ; < = > ?@ p rs y z{ c $ % & '( c c p0 p0 p0 p0

-4.5 -5.5 wx -5 -7

* + , -. 5 6 7 8 9: t uv | }~ ) c c p0 p0 p0 ‘

FIG. 13. Perpendicular correlation length near criticality on inﬁnite k-regular networks. a The simultaneous (perpendicular) path connectedness function has an exponentially decaying tail, with the exact form predicted analytically. Solid lines are analytical results while symbols are numerical simulations on inﬁnite 10-regular networks (averaged over 20,000,000 clusters), −3 −1 simulated at p0 = 10 . b The isotropic correlation length diverges with a power-law of (qc − q) .

so in the limit of large d, a.101

g(2(d + 1), 0) 100

® σ(1 + β) , (G2) ¬ iso

g(2d, 0) ≈ « ª

§ 10 ¦

p ¥

1 ¤ 1 £ which yields ξ = − (q q) , yielding ¢

c − ¡

2 log(σ(1+√β))

⊥ ∼ − -2

the scaling exponent ν = 1. Finally, it’s clear that the

⊥

exp(-x/1700) divergence in correlation length occurs precisely when -3

2 10

σ 1+ √β = 1. Simple algebra shows that this is

2 2 equivalent to σ β = (1 σ ), which is the critical line -4 10 derived by considering the− divergence of the average clus-

ter size (see Eq. B28). 10-5

0 10000 20000 30000 40000 50000 ¯°±²³ ´µ ¶· ¸¹º »¼½¾ ¿

-4 / 2 ÒÓ Ô Õ -8 / 2 -10 / 2

Á Â Ã ÄÅ Ì Í Î Ï ÐÑ Ü Ý Þ ß àá ì í î ï ð ñ À c c c c

-5 / 2 -7 / 2 èé ê ë -11 / 2

Ç È É ÊË Ö × Ø Ù ÚÛ â ã ä å æç ò ó ô õ ö ÷ 2. Alternative correlation lengths Æ c c c c b.105

An alternative parallel correlation length ξ for random graphs is the length characterizing the decayk of di- 104

rect descendants of an active site. This can be measured

t ø with g(t,t)= σ = exp[ t/ξ ] where ξ = 1/ ln(σ) de- 10

d d

− −

notes the descendant correlation length. This correlation ¨

ý þ ÿ -

§ c

¤ £ length is given strictly by the branching ratio, σ, and di- ¢ 2 ¡ 10 1 C verges as ξ q(σ = 1,p) q − . However, the σ = 1 d ∼ | − | line on which ξd diverges is well into the super-critical 1 101 regime, save for the singular point p = 0 and q = k . We’ve considered a correlation length ξ that corre- ⊥

sponds to the directed percolation perpendicular corre- 100 û ü

lation length. ξ is anisotropic, which may appear to 10ùú 10-5 10-4 10 10-2 10-1 make it unrelated⊥ to the isotropic correlation length of q c undirected percolation. However, the presence of a diverging anistropic correlation length implies that any FIG. 14. Isotropic correlation length near criticality foron in- isotropic correlation length will also diverge. To illus- finite k-regular networks. a The istropic path connectedness trate this, consider the correlation length, ξ defined by function giso(d) for all paths of length d decays exponentially. the mean number of sites active after d (isotropic) net- This is the result of simulations of infinite 10-regular networks, simulated at p = 10−3. b The corresponding isotropic work hops away from an active node, i.e. giso(d) = −1 d correlation length ξ diverges with a power-law of (qc − q) . t= d g(d, t) exp[ d/ξ]. Since this sum contains − ∼ − P 21 g(d, 0) exp[ d/(2ξ )], which tends to a constant as In contrast to our results with the SIS model, we antic- ∼ − ⊥ q qc, we know therefore that ξ also diverges in the ipate that in SIR this transition is more sensitive to the same→ limit, as can be seen in Fig. 14. details of the network. Immunization plays a role when activity traverses a loop in the network. Percolation on directed k-regular networks is mean-field, and so loops Appendix H: Static model play little role. However, for networks where loops are prevalent (such as small-world networks, or undirected Many diseases exhibit immunity after spreading. In networks) immunization may have a stronger effect. It this paper, we have predominately considered a model is known, for instance, that the SIR model falls into with re-excitable nodes, corresponding to the SIS model. the dynamical percolation universality class and shows It is, however, also of interest to consider diseases that an undirected percolation phase transition on undirected might have multiple initiation points, but which cannot networks [1, 97]. The directed percolation exponents reinfect an individual after they have contracted the dis- are only demonstrated with SIR on directed networks ease. This is a modification of the canonical SIR model. [36]. It would be interesting to see which aspects of the To consider such diseases, we initially infect some frac- spreading-merging transition remain, if the SIR model tion p of the network. Each connection between nodes were to be simulated on undirected networks. transmits the infection with probability q. The analogy to our initial model is not exact because we have dis- posed of the temporal aspect of the model, i.e. de novo Appendix I: Additional figures infections are not repeatedly introduced to the system. Nonetheless, there are two modes by which the disease In this section we include several figures to supple- grows – a period of spreading followed by the merging ment the main text. Fig. 16 contains un-scaled avalanche of large clusters. Unsurprisingly, this model also shows distributions for the hierarchical modular networks and two power-laws in the cluster size, with an exponent of 10-regular networks corresponding to Fig. 3. Similarly, 2.5 for q 0 and an exponent of 1.5 for p 0 (cf. Fig. 17 consists of the exponent transitions for 10-regular −Fig. (15)) on→ directed k-regular networks− in the→ percola- graphs without the exponent transitions presented in tion and directed percolation limits respectively. Fig. 5.

[1] M. E. Newman, Spread of epidemic disease on networks, (2012). Physical Review E 66, 016128 (2002). [10] J. G. Orlandi, J. Soriano, E. Alvarez-Lacalle, S. Teller, [2] E. Kenah and J. M. Robins, Second look at the spread and J. Casademunt, Noise focusing and the emergence of of epidemics on networks, Physical Review E 76, 036113 coherent activity in neuronal cultures, Nature Physics 9, (2007). 582 (2013). [3] R. Pastor-Satorras, C. Castellano, P. Van Mieghem, and [11] E. Tagliazucchi, P. Balenzuela, D. Fraiman, and D. R. A. Vespignani, Epidemic processes in complex networks, Chialvo, Criticality in large-scale brain fmri dynamics Reviews of Modern Physics 87, 925 (2015). unveiled by a novel point process analysis, Frontiers in [4] J. C. Wierman and D. J. Marchette, Modeling computer physiology 3, 15 (2012). virus prevalence with a susceptible-infected-susceptible [12] S. di Santo, P. Villegas, R. Burioni, and M. A. Muñoz, model with reintroduction, Computational Statistics & Landauginzburg theory of cortex dynamics: Scale-free Data Analysis 45, 3 (2004). avalanches emerge at the edge of synchronization, Pro- [5] M. Nekovee, Y. Moreno, G. Bianconi, and M. Marsili, ceedings of the National Academy of Sciences of the Theory of rumour spreading in complex social networks, United States of America 115, E1356 (2018). Physica A: Statistical Mechanics and its Applications [13] O. Kinouchi and M. Copelli, Optimal dynamical range 374, 457 (2007). of excitable networks at criticality, Nature Physics 2, 348 [6] P. Crucitti, V. Latora, and M. Marchiori, A topological (2006). analysis of the italian electric power grid, Physica A: Sta- [14] M. Yaghoubi, T. de Graaf, J. G. Orlandi, F. Girotto, tistical mechanics and its applications 338, 92 (2004). M. A. Colicos, and J. Davidsen, Neuronal avalanche dy- [7] X. Wang, Y. Ko¸c, R. E. Kooij, and P. Van Mieghem, namics indicates different universality classes in neuronal A network approach for power grid robustness against cultures, Scientific Reports 8, 3417 (2018). cascading failures, in 2015 7th international workshop on [15] D. Dahmen, S. Grün, M. Diesmann, and M. Helias, Sec- reliable networks design and modeling (RNDM) (IEEE, ond type of criticality in the brain uncovers rich multiple- 2015) pp. 208–214. neuron dynamics, Proceedings of the National Academy [8] D. R. Chialvo, Emergent complex neural dynamics, Na- of Sciences of the United States of America 116, 13051 ture Physics 6, 744 (2010). (2019). [9] N. Friedman, S. Ito, B. A. Brinkman, M. Shimono, R. L. [16] J. M. Beggs and D. Plenz, Neu- DeVille, K. A. Dahmen, J. M. Beggs, and T. C. But- ronal avalanches in neocortical circuits., ler, Universal critical dynamics in high resolution neu- The Journal of neuroscience : the official journal of the Society for Neuroscience 23, 11167 (2003). ronal avalanche data, Physical review letters 108, 208102 22

100 avalanches emerge at the edge of synchronization, Pro- ceedings of the National Academy of Sciences 115, E1356 -1 10 (2018).

10-2 [27] J. Pinheiro Neto, F. P. Spitzner, and V. Priesemann, A uniﬁed picture of neuronal avalanches arises from the un- s-3/2 10-3 derstanding of sampling eﬀects, arXiv , arXiv (2019). [28] A. Das and A. Levina, Critical neuronal models with re- 10-4 P(s) laxed timescale separation, Physical Review X 9, 021062 s-5/2

10-5 (2019). [29] A. Ponce-Alvarez, A. Jouary, M. Privat, G. Deco, and 10-6 G. Sumbre, Whole-Brain Neuronal Activity Displays Crackling Noise Dynamics, Neuron 100, 1446 (2018).

10 q=0 q = 0.09 [30] R. V. Williams-Garcia, J. M. Beggs, and G. Ortiz, Un-

p ! veiling causal activity of complex networks, EPL (Euro- 10-8

100 101 102 103 104 physics Letters) 119, 18003 (2017).

S [31] P. Van Mieghem and E. Cator, Epidemics in networks with nodal self-infection and the epidemic threshold, FIG. 15. Cluster size distributions on undirected 10-regular Physical Review E 86, 016116 (2012). networks with N = 216. p fraction of nodes are infected and [32] H. Hinrichsen, Non-equilibrium critical phe- disease spreads with the probability q through the links, which nomena and phase transitions into absorb- corresponds to SIR model. Only when p = 0, do we consider ing states, Advances in Physics 49, 815 (2000), a single patient zero. Data points are the result of 1000 real- https://doi.org/10.1080/00018730050198152. izations, except p = 0, q = 0.1 which has 100,000 realizations. [33] M. Yanuka, The mixed bond-site percolation problem and its application to capillary phenomena in porous media, Journal of Colloid and Interface Science 134, 198 (1990). [34] H. F. Song and X.-J. Wang, Simple, distance-dependent [17] A. Ponce-Alvarez, A. Jouary, M. Privat, G. Deco, and formulation of the watts-strogatz model for directed and G. Sumbre, Whole-Brain Neuronal Activity Displays undirected small-world networks, Physical Review E 90, Crackling Noise Dynamics., Neuron 100, 1446 (2018). 062801 (2014). [18] B. Del Papa, V. Priesemann, and J. Triesch, Criti- [35] C. Moore and M. E. Newman, Exact solution of site and cality meets learning: Criticality signatures in a self- bond percolation on small-world networks, Physical Re- organizing recurrent neural network, PLOS ONE 12, view E 62, 7059 (2000). e0178683 (2017). [36] N. Schwartz, R. Cohen, D. Ben-Avraham, A.-L. [19] J. M. Beggs and N. Timme, Being Critical of Criticality Barabási, and S. Havlin, Percolation in directed scale- in the Brain, Frontiers in Physiology 3, 163 (2012). free networks, Physical Review E 66, 015104 (2002). [20] L. Cocchi, L. L. Gollo, A. Zalesky, and M. Breakspear, [37] R. Cohen, D. Ben-Avraham, and S. Havlin, Percolation Criticality in the brain: A synthesis of neurobiology, critical exponents in scale-free networks, Physical Review models and cognition, Progress in Neurobiology 158, 132 E 66, 036113 (2002). (2017). [38] M. A. Muñoz, R. Juhász, C. Castellano, and G. Odor,´ [21] J. M. Beggs and D. Plenz, Neuronal Avalanches Griffiths phases on complex networks, Physical Review Are Diverse and Precise Activity Patterns That Are Letters 105, 128701 (2010). Stable for Many Hours in Cortical Slice Cultures, [39] M. A. Munoz, R. Dickman, A. Vespignani, and S. Zap- The Journal of Neuroscience 24, 5216 (2004). peri, Avalanche and spreading exponents in systems with [22] P. Moretti and M. A. Muñoz, Griffiths phases and the absorbing states, Physical Review E 59, 6175 (1999). stretching of criticality in brain networks, Nature Com- [40] K. Christensen and N. R. Moloney, Complexity and Crit- munications 4, 2521 (2013). icality, Vol. 1 (World Scientific Publishing Company, [23] A. J. Fontenele, N. A. P. de Vasconcelos, T. Feli- 2005). ciano, L. A. A. Aguiar, C. Soares-Cunha, B. Coimbra, [41] W. Cota, G. Odor,´ and S. C. Ferreira, Griffiths phases in L. Dalla Porta, S. Ribeiro, A. J. a. Rodrigues, N. Sousa, infinite-dimensional, non-hierarchical modular networks, P. V. Carelli, and M. Copelli, Criticality between cortical Scientific Reports 8, 9144 (2018). states, Phys. Rev. Lett. 122, 208101 (2019). [42] D. S. Callaway, M. E. Newman, S. H. Strogatz, and D. J. [24] L. Dalla Porta and M. Copelli, Modeling neuronal Watts, Network robustness and fragility: Percolation on avalanches and long-range temporal correlations at the random graphs, Physical Review Letters 85, 5468 (2000). emergence of collective oscillations: Continuously vary- [43] S. Melnik, A. Hackett, M. A. Porter, P. J. Mucha, and ing exponents mimic m/eeg results, PLoS computational J. P. Gleeson, The unreasonable effectiveness of tree- biology 15, e1006924 (2019). based theory for networks with clustering, Physical Re- [25] S.-S. Poil, R. Hardstone, H. D. Mansvelder, and view E 83, 036112 (2011). K. Linkenkaer-Hansen, Critical-state dynamics of [44] R. V. Williams-Garc´ıa, M. Moore, J. M. Beggs, and avalanches and oscillations jointly emerge from balanced G. Ortiz, Quasicritical brain dynamics on a nonequilib- excitation/inhibition in neuronal networks, Journal of rium widom line, Physical Review E 90, 062714 (2014). Neuroscience 32, 9817 (2012). [45] P. Bak, C. Tang, and K. Wiesenfeld, Self-organized crit- [26] S. Di Santo, P. Villegas, R. Burioni, and M. A. Muñoz, icality, Physical review A 38, 364 (1988). Landau–ginzburg theory of cortex dynamics: Scale-free 23

a. 100 b. 100

S1.3

-2 10 s-3/2 10-2

10-4 s-5/2

-4

10 2.1

8 S

7 6

-6 -7 5 4

10 &'() 3

P(s) 2

1 -6 P p=10 10-6

10-8

-8 10 p=10-3 10-10

-2 "#$% q=0 10-12 10-10

100 101 102 103 104 105 100 101 102 103 104 105

*+, ./0 s9:

FIG. 16. Unscaled critical avalanche distributions. a Avalanche distributions for 10-regular networks as in Fig. 3e. Solid lines are analytical P (s) determined from the generating function, while the crosses and circles are simulations on inﬁnite and ﬁnite (N = 107) networks. b Avalanche distributions for HMN networks as in Fig. 3f of the main text.

10 a. 10 b. δq < 0 10.1103/physrevlett.108.228703 (2012). 107 δq > 0 108 [50] D. Millman, S. Mihalas, A. Kirkwood, and

6 105 〉 10 E. Niebur, Self-organized criticality occurs in non- χ

R(s) 4 〈 10 103 δq-3 conservative neuronal networks during Up states., s1 102 Nature physics 6, 801 (2010). 101 δq-1 100 [51] S. Scarpetta, I. Apicella, L. Minati, and A. d. Can- 100 101 102 103 104 105 106 107 108 109 1010 10-4 10-3 10-2 10-1 dia, Hysteresis, neural avalanches, and critical behavior s |δq| near a ﬁrst-order transition of a spiking neural network, c. 105 d. 100 T-2 Physical Review E 97, 062305 (2018). 104 [52] A. Levina, J. M. Herrmann, and T. Geisel, Dynamical 10-6 3 〉 10 synapses causing self-organized criticality in neural net- T(s) 2 P(T) -7 〈 10 1/2 s -12 T works, Nature Physics 3, nphys758 (2007), 0712.1003. s1/4 10 101 [53] B. D. Papa, V. Priesemann, and J. Triesch,

100 10-18 Criticality meets learning: Criticality signatures 100 102 104 106 108 1010 100 101 102 103 104 105 in a self-organizing recurrent neural network, s T PLOS ONE 12, e0178683 (2017). [54] V. Priesemann, M. Valderrama, M. Wibral, FIG. 17. Exponent transitions for critical avalanches, with- and M. L. V. Quyen, Neuronal Avalanches Dif- out the curve collapse presented in the main text. Exponent fer from Wakefulness to Deep Sleep Evidence transitions for 10-regular networks, without the rescaling pre- from Intracranial Depth Recordings in Humans, sented in Fig. 5 of the main text. Data are for p = 10−2 to PLoS Computational Biology 9, e1002985 (2013). p = 10−8. a The average number of roots for avalanches of a [55] V. Priesemann, M. Wibral, M. Valderrama, R. Pröpper, given size, as simulated on infinite networks. b The analyti- M. Le Van Quyen, T. Geisel, J. Triesch, D. Nikolić, and cally determined susceptibility χ. c The avalanche duration- M. H. Munk, Spike avalanches in vivo suggest a driven, size relation. d The avalanche duration distribution. slightly subcritical brain state, Frontiers in systems neuroscience 8, 108 (2014). [56] A. Levina and V. Priesemann, Subsampling scaling, Nature Communications 8, ncomms15140 (2017), [46] E. Bienenstock, A model of neocortex, 1701.04277. 6 Network: Computation in Neural Systems , 179 (1995). [57] M. Nonnenmacher, C. Behrens, P. Berens, M. Bethge, [47] P. Massobrio, V. Pasquale, and S. Martinoia, Self- and J. H. Macke, Signatures of criticality arise organized criticality in cortical assemblies occurs from random subsampling in simple population mod- in concurrent scale-free and small-world networks, els, PLOS Computational Biology 13, e1005718 (2017), 5 Scientific Reports , 10578 (2015). 1603.00097. [48] C. Tetzlaff, S. Okujeni, U. Egert, F. Wrgt- [58] G. Scott, E. D. Fagerholm, H. Mutoh, R. Leech, ter, and M. Butz, Self-Organized Criti- D. J. Sharp, W. L. Shew, and T. Knpfel, cality in Developing Neuronal Networks, Voltage Imaging of Waking Mouse Cortex Re- 6 PLoS Computational Biology , e1001013 (2010). veals Emergence of Critical Neuronal Dynamics, [49] F. Lombardi, H. Herrmann, C. Perrone-Capano, The Journal of Neuroscience 34, 16611 (2014). D. Plenz, and L. d. Arcangelis, Balance between Excita- [59] J. Touboul and A. Destexhe, Power-law statistics tion and Inhibition Controls the Temporal Organization and universal scaling in the absence of criticality, of Neuronal Avalanches, Physical Review Letters 108, 24

Physical Review E 95, 012413 (2017). [78] A. Maxmen, Ebola detectives race to identify hidden [60] V. Priesemann and O. Shriki, Can a time varying ex- sources of infection as outbreak spreads., Nature 564, ternal drive give rise to apparent criticality in neural 174 (2018). systems?, PLOS Computational Biology 14, e1006081 [79] H. J. Jensen, Self-organized criticality: emergent com- (2018). plex behavior in physical and biological systems, Vol. 10 [61] P. Villegas, S. di Santo, R. Burioni, and M. A. Muñoz, (Cambridge university press, 1998). Time-series thresholding and the definition of avalanche [80] G. Pruessner, Self-organised criticality: theory, mod- size, Physical Review E 100, 012133 (2019). els and characterisation (Cambridge University Press, [62] M. Girardi-Schappo and M. H. R. Tragtenberg, Measur- 2012). ing neuronal avalanches in disordered systems with ab- [81] P. Bak, How nature works: the science of self-organized sorbing states, Physical Review E 97, 042415 (2018). criticality (Springer Science & Business Media, 2013). [63] T. Bolt, M. L. Anderson, and L. Q. Uddin, Be- [82] A.´ Corral and M. Paczuski, Avalanche Merg- yond the evoked/intrinsic neural process dichotomy, ing and Continuous Flow in a Sandpile Model, Network Neuroscience 2, 1 (2018). Physical Review Letters 83, 572 (1999). [64] R. V. Williams-Garca, J. M. Beggs, and G. Or- [83] M. De Domenico, C. Granell, M. A. Porter, and A. Are- tiz, Unveiling causal activity of complex net- nas, The physics of spreading processes in multilayer networks, EPL (Europhysics Letters) 119, 18003 (2017), works, Nature Physics 12, 901 (2016). 1603.05659. [84] J. Sanz, C.-Y. Xia, S. Meloni, and Y. Moreno, Dynam- [65] J. Zierenberg, J. Wilting, V. Priesemann, and ics of interacting diseases, Physical Review X 4, 041005 A. Levina, Tailored ensembles of neural net- (2014). works optimize sensitivity to stimulus statistics, [85] M. Kitsak, L. K. Gallos, S. Havlin, F. Liljeros, L. Much- Physical Review Research 2, 013115 (2020). nik, H. E. Stanley, and H. A. Makse, Identification of in- [66] Z.-L. Hu, J.-G. Liu, G.-Y. Yang, and Z.-M. Ren, Effects of fluential spreaders in complex networks, Nature Physics the distance among multiple spreaders on the spreading, 6, 888 (2010). EPL (Europhysics Letters) 106, 18002 (2014). [86] Y. Hu, S. Ji, Y. Jin, L. Feng, H. E. Stanley, and [67] J. C. Miller, Epidemics on networks with large initial S. Havlin, Local structure can identify and quantify in- conditions or changing structure, PLOS ONE 9, e101421 fluential global spreaders in large scale social networks, (2014). Proceedings of the National Academy of Sciences of the [68] M. Zheng, C. Wang, J. Zhou, M. Zhao, S. Guan, Y. Zou, United States of America 115, 7468 (2018). and Z. Liu, Non-periodic outbreaks of recurrent epi- [87] J. Nagler, A. Levina, and M. Timme, Impact of single demics and its network modelling, Scientific Reports 5, links in competitive percolation, Nature Physics 7, 265 16010 (2015). (2011). [69] T. Hasegawa and K. Nemoto, Outbreaks in susceptible- [88] P. Grassberger, C. Christensen, G. Bizhani, S.-W. Son, infected-removed epidemics with multiple seeds, Physical and M. Paczuski, Explosive percolation is continuous, but Review E 93, 032324 (2016). with unusual finite size behavior, Physical Review Letters [70] W. Choi, D. Lee, and B. Kahng, Critical behavior of a 106, 225701 (2011). two-step contagion model with multiple seeds, Physical [89] R. M. DSouza and J. Nagler, Anomalous critical and su- Review E 95, 062115 (2017). percritical phenomena in explosive percolation, Nature [71] T. Hasegawa and K. Nemoto, Sudden spreading of infec- Physics 11, 531 (2015). tions in an epidemic model with a finite seed fraction, [90] F. Morone, K. Roth, B. Min, H. E. Stanley, and H. A. The European Physical Journal B 91, 58 (2018). Makse, Model of brain activation predicts the neural col- [72] M. Di Muro, L. Alvarez-Zuzek, S. Havlin, and L. Braun- lective influence map of the brain, Proceedings of the stein, Multiple outbreaks in epidemic spreading with lo- National Academy of Sciences of the United States of cal vaccination and limited vaccines, New Journal of America 114, 3849 (2017). Physics 20, 083025 (2018). [91] N. Chen and M. Olvera-Cravioto, Directed random [73] E. Cator and P. Van Mieghem, Susceptible-infected- graphs with given degree distributions, Stochastic Sys- susceptible epidemics on the complete graph and the star tems 3, 147 (2013). graph: Exact analysis, Physical Review E 87, 012811 [92] D. R. De Souza, T. Tomé, and R. M. Ziff, A new scale- (2013). invariant ratio and finite-size scaling for the stochastic [74] J. Zhang, J. M. Moura, and J. Zhang, Contact process susceptible–infected–recovered model, Journal of Statis- with exogenous infection and the scaled sis process, Jour- tical Mechanics: Theory and Experiment 2011, P03006 nal of Complex Networks 5, 712 (2017). (2011). [75] A. J. Rodriguez-Morales, A. C. Bandeira, and C. Franco- [93] G. Evans and J. Webster, A comparison of some methods Paredes, The expanding spectrum of modes of transmis- for the evaluation of highly oscillatory integrals, Jour- sion of zika virus: a global concern, Annals of Clinical nal of Computational and Applied Mathematics 112, 55 Microbiology and Antimicrobials 15, 13 (2016). (1999). [76] E. M. Leroy, B. Kumulungui, X. Pourrut, P. Rouquet, [94] R. L. Graham, D. E. Knuth, and O. Patashnik, Con- A. Hassanin, P. Yaba, A. Délicat, J. T. Paweska, J.-P. crete Mathematics: A Foundation for Computer Science Gonzalez, and R. Swanepoel, Fruit bats as reservoirs of (Addison-Wesley: New York, 1989). ebola virus, Nature 438, 575 (2005). [95] Y. L. Luke, Special Functions and Their Approximations, [77] N. D. Wolfe, P. Daszak, A. M. Kilpatrick, and D. S. Vol. 1 (Academic Press, 1969). Burke, Bushmeat hunting, deforestation, and prediction [96] W. N. Watson, Asymptotic expansions of hypergeometric of zoonotic disease, Emerging Infectious Diseases 11, functions, Transactions of the Cambridge Philosophical 1822 (2005). 25

Society 22, 277 (1918). Physical Review E 82, 051921 (2010). [97] T. Tom´e and R. M. Ziﬀ, Critical behavior of the susceptible-infected-recovered model on a square lattice,