PHYSICAL REVIEW E 75, 056103 共2007兲 Seed size strongly affects cascades on random networks 1 James P. Gleeson1 and Diarmuid J. Cahalane2 Applied Mathematics, University College Cork, Cork, Ireland 2 MACSI, University College Cork, Cork, Ireland 共Received 16 February 2007; published 3 May 2007兲 The average avalanche size in the Watts model of threshold dynamics on random networks of arbitrary degree distribution is determined analytically. Existence criteria for global cascades are shown to depend sensitively on the size of the initial seed disturbance. The dependence of cascade size upon the mean degree z of the network is known to exhibit several transitions—these are typically continuous at low z and discontinuous at high z; here it is demonstrated that the low-z transition may in fact be discontinuous in certain parameter regimes. Connections between these results and the zero-temperature random-field Ising model on random graphs are discussed. DOI: 10.1103/PhysRevE.75.056103 PACS number共s兲: 89.75.Hc Dynamical models on networks have attracted a great deal of recent research interest; see the reviews 关1–3兴 and references therein. Of particular interest are the effects of the topological structure of the network upon the time evolution of dynamical quantities. Under certain circumstances a change of state at a small number of network nodes may cause a cascade of changes over the whole network—such cascades may occur as, for example, overload failures 关4–9兴, avalanches in sandpile models 关10,11兴, evolution of species 关12兴, or the spread of rumors and fads in social networks 关13–15兴. In this paper, we analyze a model of dynamics on random networks from the class of problems known as binary decisions with externalities 关16兴. Each node of an undirected network represents an agent, and may be in one of two states, called active and inactive. The network is chosen from an ensemble of graphs with a specified degree distribution pk, i.e., the probability that a node has degree k is pk, with 兺pk = 1 关1,17兴. Each agent also has a 共frozen兲 random threshold r, chosen from a distribution with F共r兲 denoting the probability that an agent has threshold less than r. Initially the network is seeded by activating a randomly chosen fraction 0 of the N nodes. When updated, an inactive agent calculates the fraction of its neighbors who are currently active 共m / k, where k is the degree of the node and m is the number of active neighbors兲, and it becomes active if this fraction exceeds his threshold r. Once active, an agent cannot become inactive; this will be referred to as the permanently active property 共PAP兲. Isolated nodes 共degree k = 0兲 are never updated. It can be shown that 共unlike some other sequential spin-flip models 关18兴兲 the steady state of the system does not depend on whether the updating of all agents is performed synchronously, or in a random asynchronous fashion. Watts introduced the above model 关13兴 to demonstrate the interaction between network topology and individual thresholds in the spreading of rumors and fads. He used percolation methods to determine whether a single activated initial agent can induce network-spanning cascades of activation. We focus here on calculating , the average fraction of active nodes in the steady state, where the ensemble average is over realizations of networks and threshold values, and in the limit of infinite network size: N → ⬁. Watts’ criterion for glo1539-3755/2007/75共5兲/056103共4兲 bal cascades is recovered in the 0 → 0 limit 共since 0 = 1 / N for single-node initialization兲 but important differences are found when the seed size is finite. The quantity measures the average avalanche size, and has recently been studied for dynamics on directed random graphs, with applications to random Boolean networks 关19兴, and on small-world topologies 关20兴. The magnetization in the zero-temperature random-field Ising model 共RFIM兲 关21,22兴 is also closely analogous to ; the agents’ threshold values correspond to the local quenched random fields of the RFIM. There are, however, two important differences: 共i兲 the RFIM does not exhibit the PAP, since spins may flip either up or down depending on their local field; 共ii兲 the RFIM has not been studied on random graphs of arbitrary degree distribution. Analytical solutions for the RFIM have been developed in the mean-field case 关21兴 and on Bethe lattices 关22,23兴. Our approach reduces to the latter in the case of regular random graphs, where each node has exactly z neighbors 共pk = ␦kz for integer z兲, provided that 0 is zero. Our main result is that the average final fraction of active nodes is given by ⬁ k = 0 + 共1 − 0兲 兺 pk 兺 k=1 m=0 冉冊 k m q⬁m共1 − q⬁兲k−mF 冉冊 m , 共1兲 k where q⬁ is the fixed point of the recursion relation qn+1 = 0 + 共1 − 0兲G共qn兲 for n = 0,1,2 . . . , 共2兲 with q0 = 0, and the nonlinear function G is defined by ⬁ k−1 冉 冊 冉冊 k−1 m k m G共q兲 = 兺 pk 兺 q 共1 − q兲k−1−mF . k m k=1 z m=0 共3兲 Here z is the mean degree of the network, z = 兺kpk. We postpone the derivation of 共1兲 to the end of the paper, and first examine some of its consequences. Figure 1 shows in the case of uniform thresholds 共all nodes have identical threshold r = R兲 on Poisson 共ErdősRényi兲 random graphs with degree distribution pk = e−zzk / k!. In Fig. 1共a兲 is color coded on the R , z parameter plane for seed fraction 0 = 10−2; Fig. 1共b兲 compares at R = 0.18 with numerical simulations for several 0 values, as a function of 056103-1 ©2007 The American Physical Society PHYSICAL REVIEW E 75, 056103 共2007兲 JAMES P. GLEESON AND DIARMUID J. CAHALANE 1 15 (a) 10 10 0.5 z z 0.5 5 5 0 0 0.1 0.2 R 0 0.3 0 0 1 0.1 0.2 R 2 4 0.3 0 0.4 1 (b) (b) ρ ρ 1 15 (a) 0.5 0.5 0 0 2 4 z 6 8 0 0 10 FIG. 1. 共Color online兲 Average density of active nodes in a Poisson random graph of mean degree z and uniform threshold value R. 共a兲 Color-coded values of from Eq. 共1兲 on the R , z plane with seed fraction 0 = 10−2. Lines show approximations to the global cascade boundaries: the solid line encloses the region where 共5兲 is satisfied; the dashed line represents the extended cascade boundary from Eq. 共6兲. 共b兲 Values of at R = 0.18 from Eq. 共1兲 共lines兲 and numerical simulations 共symbols兲, averaged over 100 realizations with N = 105. Seed fractions are 0 = 10−3 共solid兲, 5 ⫻ 10−3 共dashed兲, and 10−2 共dot-dashed兲. The arrow marks the cascade boundary given by Eq. 共5兲 in the limit of zero seed fraction. z. The theory shows excellent agreement with simulations. Note the smooth transition from ⬇ 0 to ⬇ 1 near z = 1, and the discontinuous transition back to ⬇ 0 at larger z values. The z location of the upper transition clearly depends sensitively on the initial activated fraction 0. A global cascade occurs with high probability when a small seed 0 results in a large value of . Writing G共q兲 as ⬁ 兺ᐉ=0 Cᐉqᐉ with coefficients ⬁ Cᐉ = ᐉ 兺 兺 k=ᐉ+1 n=0 冉 冊冉 冊 k−1 ᐉ 冉冊 ᐉ k n 共− 1兲ᐉ+n pkF , z n k 共4兲 and linearizing Eq. 共2兲 near q = 0 gives a 共first-order兲 condition for global cascades to occur: 共1 − 0兲C1 ⬎ 1, since this guarantees that qn increases with n, at least initially. This cascade condition may also be written as ⬁ 兺 k=1 冋冉冊 册 k共k − 1兲 1 1 pk F − F共0兲 ⬎ . z k 1 − 0 6 8 10 FIG. 2. 共Color online兲 As Fig. 1, but threshold distribution is Gaussian with mean R and standard deviation 0.2, and seed fraction 0 = 0. 共a兲 Cascade boundaries as in Fig. 1; the arrow marks the critical point 共Rc , zc兲 described in the text; 共b兲 theory and numerical simulation results 共N = 105, average over 100 realizations兲 at R = 0.2 共solid兲, 0.362 共dashed兲, and 0.38 共dash-dotted兲. 0 = 10−3 and 10−2 共close to the z value marked by the arrow兲, but the simulations and full theory show that the respective transitions are actually near z ⬇ 6.4 and 9.2. Condition 共5兲 clearly does not accurately represent the effects of finite seed size 关nor of nonzero F共0兲; see below兴, because the function G is not well approximated by a straight line near the critical parameters for cascade transitions. We seek an improved cascade condition by extending the series for G共q兲 to order q2. This approximation results in a quadratic equation for the fixed point q⬁ which we represent as aq⬁2 + bq⬁ + c = 0. Under the approximation of small q values, we interpret the existence of a positive root of this quadratic equation to imply q⬁ Ⰶ 1, and hence the impossibility of global cascades. Note that the first-order approximation 共5兲 requires b ⬎ 0 for global cascades. However, when the nonlinear behavior of G is important, it is still possible for global cascades to occur when b is negative, provided that the full solution of the quadratic equation precludes positive real roots. We therefore extend the cascade boundary to include regions where either b ⬎ 0 共as in 共5兲兲 or b2 − 4ac ⬍ 0, the latter giving 共to first order in 0兲 共5兲 In the limit 0 → 0 and with F共0兲 = 0, this reduces to the condition derived by Watts using percolation arguments 关13兴. On the R , z plane of Fig. 1共a兲 the condition 共5兲 is satisfied inside the solid line; for R = 0.18 in Fig. 1共b兲, 共5兲 predicts cascade transitions at z values between 5.7 and 5.8 for both z 共C1 − 1兲2 − 4C0C2 + 20共C1 − C21 − 2C2 + 4C0C2兲 ⬍ 0. 共6兲 It is straightforward to plot this extended cascade condition: the dashed lines in Figs. 1共a兲 and 2共a兲 clearly give much improved approximations to the actual boundaries for global cascades. 056103-2 PHYSICAL REVIEW E 75, 056103 共2007兲 SEED SIZE STRONGLY AFFECTS CASCADES ON RANDOM… 1 (a) q 0.5 0 0 2 4 2 4 2 4 z 6 8 10 6 8 10 6 8 10 1 (b) q 0.5 0 0 z 1 (c) q 0.5 0 0 z FIG. 3. 共Color online兲 Bifurcation diagrams as described in text for dependence of q⬁ on z for = 0.2 and 0 = 0 at R = 共a兲 0.35, 共b兲 0.371, and 共c兲 0.375. Figure 2共a兲 shows on Poisson random graphs with thresholds drawn from a Gaussian distribution of mean R and standard deviation = 0.2. Unlike the assumption in 关13兴 that F共0兲 = 0, a Gaussian distribution necessarily implies the presence of negative-valued thresholds among the population, so F共0兲 ⬎ 0. Negative-threshold agents act as a natural seed, since they activate regardless of the states of their neighbors. The presence of these innovators 关13兴 allows us to set 0 = 0 in this case. The extended cascade condition 共6兲 again gives a good approximation to the discontinuous transition at high z values. Figure 2共b兲 focuses on the low-z transition and highlights the existence of a discontinuous transition in z for certain threshold distributions. This is qualitatively different from the previously-studied case 关13兴 where only continuous low-z transitions were found. Bifurcation analysis of Eq. 共2兲 elucidates this result. In Fig. 3 we plot the roots of the fixed-point equations G共q兲 − q = 0 共recall that 0 = 0 here; extension to nonzero 0 is straightforward兲 as functions of z, for different values of the mean threshold R. Thick solid and dashed lines denote stable and unstable fixed points respectively 关24兴. The PAP means the value of q⬁ achieved at a given z is that of the lowest stable branch above q = 0. The occurrence of triple roots as R is increased causes the smooth low-z transition seen in Fig. 3共a兲 to become discontinuous 关as shown by the thin solid line in Fig. 3共b兲兴, as previously seen in the numerical simulations of Fig. 2共b兲. The discontinuous low-z transition occurs for R ⬎ Rc, where the critical coordinates 共Rc , zc兲 and the value q = qc where the triple root appears are found by numerical root finding for the system of three equations q = 0 + 共1 − 0兲G共q兲, 共1 − 0兲G⬘共q兲 = 1, and G⬙共q兲 = 0. For = 0.2 this yields 共Rc , zc兲 = 共0.3543, 3.136兲; this point is marked with an arrow in Fig. 2共a兲. We remark that the discontinuous transition from q⬁ ⬇ 1 to q⬁ ⬇ 0 关which induces a similar transition in through Eq. 共1兲兴 occurs due to a saddle-node bifurcation 关24兴. This behavior is quite generic, occurring for a wide variety of parameters, with the exception of the special case studied by Watts. For 0 = 0 and F共0兲 = 0 as in 关13兴, the coefficient C0 is zero and the fixed-point equation always has a root at q = 0, with transcritical bifurcations on the q = 0 line giving rise to the observed transitions. However, any nonzero seed size replaces the transcritical bifurcations with saddlenode bifurcations as described above. We have confirmed the accuracy of these results 关and Eq. 共1兲兴 against numerical simulations on other configuration model network topologies 关1兴, including power-law degree distributions 共with exponential cutoff兲: pk ⬀ k− exp共−k / 兲 关17兴. We turn now to the derivation of Eqs. 共1兲–共3兲. Our analytical approach is based on methods introduced by Dhar et al. to study the zero-temperature random-field Ising model on Bethe lattices 关22兴. The RFIM is a spin-based model of magnetic materials, and its zero-temperature limit has been extensively studied as a model for systems exhibiting hysteresis and Barkhausen noise 关21兴. A Bethe lattice of coordination number z 共for integer z兲 is an infinite tree where every node has exactly z neighbors. Dhar et al. derive analytical results valid on Bethe lattices, but their numerical simulations show that the theory also applies very accurately to random graphs where every node has exactly z neighbors, provided that short-distance loops are rare. To analyze Watts’ model we extend the approach of 关22兴 in two ways. First, we consider treelike random graphs with arbitrary degree distributions, rather than the Bethe lattices of 关22兴. Second, we account for the PAP, which is the essential difference between Watts’ update rule and standard RFIM dynamics. This difference between the update rules is crucial to our derivation of the 0 dependence of the activated fraction . We begin by replacing the given random graph 共with degree distribution pk兲 by a tree structure. The top level of the tree is a single node with degree k, and this is connected to its k neighbors at the next lower level of the tree. Each of these nodes is in turn connected to ki − 1 neighbors at the next lower level, where ki is the degree of node i. The degree distribution of the nodes in the tree is given by p̃k = 共k / z兲pk, which is the distribution for the number of nearest neighbors in a connected graph 关1,25兴. To find the final density of active nodes, we label the levels of the tree from n = 0 at the bottom, with the top node at an infinitely high level 共n → ⬁兲. Define qn as the conditional probability that a node on level n is active, conditioned on its parent 共on level n + 1兲 being inactive. Consider updating a node on level n + 1, assuming that the nodes on all lower levels have already been updated. With probability p̃k the chosen node has k neighbors: one of these is its parent 共on level n + 2兲, and the remaining k − 1 are its children 共on level n兲. Since a fraction 0 of nodes were initially set to be active, there is a probability 0 that we have chosen one of these nodes. In this case the state of the node remains unchanged. On the other hand, with 056103-3 PHYSICAL REVIEW E 75, 056103 共2007兲 JAMES P. GLEESON AND DIARMUID J. CAHALANE probability 共1 − 0兲 the node in question is inactive. In this case we must consider its neighbors. Each of the k − 1 children is active with probability qn. We also assume its parent is inactive. Thus the node has m active children 共and therefore k − 1 − m inactive children兲 with probability 共 km−1 兲qm n 共1 − qn兲k−1−m. The probability that its threshold ri is less than its fraction of active neighbors m / k is given by F共m / k兲, and combining the independent probabilities yields Eq. 共2兲 for qn+1. The probability that the single node at the top of the tree is active is given by adding the probabilities for two independent cases: either it is already active as part of the originally activated fraction 共with probability 0兲, or it was initially inactive 共with probability 1 − 0兲. In the latter case, it will become active if sufficiently many of its k neighbors 共all on the next lower level兲 are active. Noting that the top node has degree k with probability pk, we conclude that the total probability of it being active is given by Eq. 共1兲. Although the theory is defined in terms of level-by-level updating on a tree, these results also apply to the randomgraph Watts model, provided that 共i兲 the network structure is locally treelike, 共ii兲 the state of each node is altered at most once, and 共iii兲 the steady state is independent of the order in which nodes are updated 关23兴. We note that condition 共i兲 is true in configuration model networks, where the clustering coefficient scales as 1 / N as N → ⬁ 关1兴. It does not, however, hold in small-world 关20,26兴 and other real-world networks where clustering and loops play an important role. Condition 共ii兲 is guaranteed by the PAP, and 共iii兲 holds for this and related unordered binary avalanche models 关19兴. In summary, our main result is Eq. 共1兲 for determining the expected fraction of active nodes 共mean avalanche size兲 due to the activation of a seed fraction 0 of random nodes. The simple conditions for global cascades 共5兲 and 共6兲 follow; these substantially extend Watts’ original result to include nonzero initial seed sizes. Values of 0 as low as 0.1% have dramatic effects on the location of cascade transition points 关see Fig. 1共b兲兴. Transitions in the dependence of upon the mean network degree z are explained by the appearance of saddle-node bifurcation points. Of particular note is the fact that in certain parameter regimes the low-z transition may be discontinuous. 关1兴 M. E. J. Newman, SIAM Rev. 45, 167 共2003兲. 关2兴 L. A. N. Amaral and J. M. Ottino, Eur. Phys. J. B 38, 147 共2004兲. 关3兴 S. Boccaletti et al., Phys. Rep. 424, 175 共2006兲. 关4兴 E. J. Lee, K. I. Goh, B. Kahng, and D. Kim, Phys. Rev. E 71, 056108 共2005兲. 关5兴 Y. C. Lai et al., Pramana, J. Phys. 64, 483 共2005兲. 关6兴 P. Crucitti, V. Latora, and M. Marchiori, Phys. Rev. E 69, 045104共R兲 共2004兲. 关7兴 A. E. Motter and Y. C. Lai, Phys. Rev. E 66, 065102共R兲 共2002兲. 关8兴 Y. Moreno et al., Europhys. Lett. 62, 292 共2003兲. 关9兴 P. Holme and B. J. Kim, Phys. Rev. E 65, 066109 共2002兲. 关10兴 K. I. Goh et al., Physica A 346, 93 共2005兲. 关11兴 K. I. Goh, D. S. Lee, B. Kahng, and D. Kim, Phys. Rev. Lett. 91, 148701 共2003兲. 关12兴 T. Antal, S. Redner, and V. Sood, Phys. Rev. Lett. 96, 188104 共2006兲. 关13兴 D. J. Watts, Proc. Natl. Acad. Sci. U.S.A. 99, 5766 共2002兲. 关14兴 P. S. Dodds and D. J. Watts, Phys. Rev. Lett. 92, 218701 共2004兲. 关15兴 W. Q. Duan, Z. Chen, Z. Liu, and W. Jin, Phys. Rev. E 72, 026133 共2005兲. 关16兴 T. C. Schelling, J. Conflict Resolut. 17, 381 共1973兲. 关17兴 M. E. J. Newman, S. H. Strogatz, and D. J. Watts, Phys. Rev. E 64, 026118 共2001兲. 关18兴 P. R. Eastham, R. A. Blythe, A. J. Bray, and M. A. Moore, Phys. Rev. B 74, 020406共R兲 共2006兲. 关19兴 B. Samuelsson and J. E. S. Socolar, Phys. Rev. E 74, 036113 共2006兲. 关20兴 D. Centola, V. M. Eguiluz, and M. W. Macy, Physica A 374, 449 共2007兲. 关21兴 J. P. Sethna, K. Dahmen, S. Kartha, J. A. Krumhansl, B. W. Roberts, and J. D. Shore, Phys. Rev. Lett. 70, 3347 共1993兲. 关22兴 D. Dhar, P. Shukla, and J. P. Sethna, J. Phys. A 30, 5259 共1997兲. 关23兴 P. Shukla, Int. J. Mod. Phys. B 17, 5583 共2003兲. 关24兴 S. H. Strogatz Nonlinear Dynamics and Chaos 共Perseus Books, Cambridge, MA, 1994兲. 关25兴 S. N. Dorogovtsev and J. F. F. Mendes, Evolution of Networks: From Biological Nets to the Internet and WWW 共Oxford University Press, Oxford, 2003兲. 关26兴 D. J. Watts and S. H. Strogatz, Nature 共London兲 393, 440 共1998兲. This work is funded by Science Foundation Ireland under Investigator Grant No. 02/IN.1/IM062 and the Mathematics Applications Consortium for Science and Industry in Ireland 共MACSI兲, and by the Irish Research Council for Science, Engineering and Technology under the Embark Initiative, Grant No. RS/2005/16. 056103-4
© Copyright 2026 Paperzz