1: \documentstyle[aps,prl,multicol,epsf]{revtex}
2: \begin{document}
3: \title{\bf A Nonconservative Earthquake Model of Self-Organized
4: Criticality on a Random Graph} \author{Stefano Lise and
5: Maya Paczuski}
6:
7: \address{Department of Mathematics, Huxley Building, Imperial College
8: of Science, Technology, and Medicine, London UK SW7 2BZ \\}
9: \date{\today}
10:
11: \maketitle %\parskip 2ex
12:
13: \begin{abstract}
14: We numerically investigate the Olami-Feder-Christensen model on a
15: quenched random graph. Contrary to the case of annealed random
16: neighbors, we find that the quenched model exhibits self-organized
17: criticality deep within the nonconservative regime. The probability
18: distribution for avalanche size obeys finite size scaling, with
19: universal critical exponents. In addition, a power law relation
20: between the size and the duration of an avalanche exists.
21: %For low levels of conservation we observe localised avalanche behaviour.
22: We propose that this may represent the correct mean-field limit of the model
23: rather than the annealed random neighbor version.
24: \end{abstract}
25:
26: \vspace{0.3cm}
27: {PACS numbers: 05.65.+b, 45.70.Ht, 89.75.-k}
28:
29: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
30: \begin{multicols}{2}
31:
32:
33: The idea of self-organized criticality (SOC) was introduced as a
34: possible explanation for the widespread occurrence in nature of long
35: range correlations in space and time \cite{BTW}. The term refers to
36: the intrinsic tendency of a large class of spatially extended
37: dynamical systems to spontaneously organize into a dynamical critical
38: state. In general, SOC systems are driven externally at a very slow
39: rate and relax with bursts of activity, avalanches, on a very fast
40: time scale. One signature of SOC is a scale free, e.g. power law,
41: distribution of avalanche sizes. This is normally related to some
42: long range spatial and temporal correlations within the system.
43: Typical natural realizations of this phenomena include, among others,
44: earthquakes, forest fires, and biological evolution (for
45: reviews, see \cite{bak_book,jen_book}).
46:
47:
48: A problem that has attracted a lot
49: of attention, but is still poorly understood, is that of identifying
50: fundamental mechanisms leading to SOC behavior. In particular, much
51: effort has been directed at understanding how
52: conservation of the transported quantity (e.g. sand) in the avalanche
53: dynamics affects criticality~\cite{hwa-kardar,grinstein}. For
54: instance, it is well known that the Abelian sandpile model \cite{BTW},
55: is subcritical if dissipation is introduced~\cite{manna_1}. On the
56: other hand, nonconservative sandpile models which display criticality
57: have since been introduced~\cite{ali,manna_2}. Although both the
58: analytical and numerical evidence in favor of criticality are quite
59: convincing, the role played in these models by the non-conservative
60: dynamics is not clear. In fact, dissipation is a dynamical variable
61: and does not always occur.
62:
63:
64: A model, which in the context of SOC in nonconservative systems has played an
65: important role, is the Olami-Feder-Christensen (OFC) model of
66: earthquakes~\cite{ofc}. In the OFC model a finite fraction,
67: controlled by a fixed parameter $\alpha$, of the transported
68: quantity is dissipated in each relaxation event. The presence of
69: criticality in the non-conservative OFC model has been controversial
70: since the very introduction of the model~\cite{klein} and it is still
71: debated~\cite{carvalho,kim}. Recent numerical investigations, though,
72: have shown that the OFC model on a square lattice displays scaling
73: behavior, up to lattice sizes presently accessible by
74: computer simulations~\cite{lisepac1,lisepac2}. The avalanche size
75: distribution is described by a power law, characterized by
76: a universal exponent $\tau \simeq 1.8$, independent of the dissipation
77: parameter. This distribution does not display finite size scaling, however.
78:
79:
80: To overcome the limitation of relying almost exclusively on computer
81: simulation results, it has sometimes been useful to consider an
82: annealed random neighbor (RN) version of the
83: model~\cite{lise,chabanol,broker,kinouchi}, where each site interacts
84: with randomly chosen sites instead of its nearest neighbors on the
85: lattice. This considerably simplifies the problem. In the past, RN
86: models have usually been considered as mean-field descriptions of
87: their fixed lattice counterparts, since spatial correlations are
88: absent. Analogous to other RN models, the RN OFC model can be solved
89: analytically~\cite{chabanol,broker}. It displays criticality only in
90: the conservative case, where it becomes equivalent to a critical
91: branching process. As soon as some dissipation is introduced the
92: avalanches become localized, although the mean avalanche size diverges
93: exponentially as dissipation tends to zero. The absence of
94: criticality, together with the exponential divergence has cast some
95: doubt on whether the OFC model on a fixed lattice is critical.
96:
97: However, it is important to point out that the RN model may not
98: describe the behavior of the OFC model on a fixed lattice in any
99: dimension, and thus may not correspond to the mean field limit of the
100: model. Usually, mean field behavior describes the high dimensional
101: behavior of the system (e.g. the behavior above an upper critical
102: dimension); this is not exactly the same limit as a model without any
103: spatial correlations in it.
104:
105:
106: In fact, criticality in the OFC model on a lattice has been ascribed
107: to a mechanism of partial synchronization~\cite{middleton}. The system
108: has a tendency to order into a periodic
109: state~\cite{middleton,socolar,grass2} which is frustrated by the
110: presence of inhomogeneities such as the boundaries. In addition,
111: inhomogeneities induce partial synchronization of the elements of the
112: system building up long range spatial correlations and thereby
113: creating a critical state. The mechanism of synchronization requires
114: an underlying spatial structure and therefore cannot operate in an
115: annealed RN model, where each site is assigned new random neighbors at
116: each update.
117:
118:
119:
120: The main purpose of the present work is the investigation of the OFC
121: model on a quenched random graph. This formulation, which can be
122: handled numerically, is worth analyzing to see if it presents critical
123: or non-critical behavior. Since the largest distance between two
124: sites in the random graph grows only as a logarithm of the number of
125: sites, it can be considered to be a high dimensional limit of a
126: lattice model, and thus may describe the correct mean field limit.
127: Contrary to the RN case, in a random graph the choice
128: of neighbors is not annealed but quenched, so that
129: spatial correlations can develop. Indeed, we show that
130: the OFC model on a random graph displays criticality even in the
131: nonconservative regime.
132:
133:
134:
135:
136:
137:
138: A random graph is defined as a set of $N$ sites connected at random by
139: bonds. Two connected sites are denoted as
140: ``nearest-neighbor''. Formally, the random graph can be constructed by
141: considering all $N(N-1)/2$ possible bonds between sites and occupying
142: a certain number of them with equal probability. A constraint of fixed
143: connectivity can also be imposed by requiring that each site has the
144: same number of neighbors, $q$. We have mainly concentrated on this
145: latter situation but the first case will also be discussed. The model
146: is then defined as follows. To each site of the graph is associated a
147: real variable $F_i$, which initially takes some random value in the
148: interval $(0,F_{th})$. All the forces are increased uniformly and
149: simultaneously at the same speed, until one of them reaches the
150: threshold value $F_{th}$ and becomes unstable $(F_i \geq F_{th})$. The
151: uniform driving is then stopped and an ``earthquake'' (or avalanche)
152: starts:
153: \begin{equation}
154: \label{av_dyn}
155: F_i \geq F_{th} \Rightarrow \left\{ \begin{array}{l}
156: F_i \rightarrow 0 \\
157: F_{nn} \rightarrow F_{nn} + \alpha F_i
158: \end{array} \right.
159: \end{equation}
160: where
161: ``nn'' denotes the set of nearest-neighbor sites of $i$. The parameter $\alpha$
162: controls the level of conservation of the dynamics and, in the case of a
163: graph with fixed connectivity $q$, it takes values between $0$ and $1/q$
164: ($\alpha=1/q$ corresponding to the conservative case).
165: The toppling rule (\ref{av_dyn}) can possibly create new unstable sites,
166: producing a chain reaction. All sites that are above threshold at a
167: given time step in the avalanche relax simultaneously according to
168: (\ref{av_dyn}) and the earthquake is over when there are no more unstable
169: sites in the system ($F_i < F_{th}$, $\forall i$). The uniform growth then
170: starts again. The number of topplings during an earthquake defines its size,
171: $s$, and we will be interested in the probability distribution $P_N (s)$.
172: Another quantity of interest is the duration $t$ of an earthquake which will
173: be identified with the number of time steps needed for the earthquake to
174: finish.
175:
176:
177:
178: We consider first a random graph where all sites have exactly the same number
179: of nearest neighbors $q$. In this case, we have verified (both for $q=4$ and
180: $q=6$) that the system organizes into a subcritical state. This is analogous
181: to what happens in the OFC model on a lattice with periodic boundary
182: conditions, where no critical behavior is
183: observed~\cite{middleton,socolar,grass2}.
184: In order to observe scaling in the avalanche distribution, one has to
185: introduce some inhomogeneities. In the lattice model this is generally
186: achieved by considering open boundary conditions which imply that boundary
187: sites have fewer neighbors and therefore cycle at a different frequency
188: from bulk sites. This is an inhomogeneity with a diverging length scale in
189: the thermodynamic limit. For the OFC model on a random graph, we have
190: found that it suffices to consider just two sites in the system with
191: coordination $q-1$~\cite{note1}. When either of
192: these sites topple according to rule
193: (\ref{av_dyn}), an extra amount $ \alpha F_i$ is simply lost by the system.
194:
195:
196:
197:
198:
199:
200: After a sufficiently long transient time, the system settles into a
201: statistically stationary state. We have verified that the statistical
202: properties of the system (e.g. the avalanche distribution) are
203: independent of the actual realization of the random graph, as long as
204: the coordination number $q$ is the same. As a point of comparison, in
205: figure~1 we report the probability distribution of avalanche sizes for
206: (a) the annealed RN model and (b) for the OFC model on a random graph
207: for various system sizes $N$. The dynamical rule for the annealed RN
208: model are formally similar to (1), where, instead of the
209: nearest-neighbor sites on the graph, $q$ new random sites are chosen
210: at each relaxing event. In both cases of fig.~1, the number of
211: neighbors is $q=4$ and the parameter $\alpha=0.10$. It is clear that
212: no scaling is present in the RN model as the cut-off in the avalanche
213: size distribution does not grow with system size. On the contrary for
214: the model on a random graph, the distribution scales with system size,
215: which is indicative of a critical state. In fact, the largest
216: avalanche roughly coincides with system size. It is important to
217: underline that we are considering a situation far away from the
218: conservative case (60\% of the force in the toppling site is
219: dissipated) and therefore one could expect that if a finite length
220: scale related to conservation existed in the system it should appear
221: for system sizes we have considered.
222:
223:
224: In order to characterize the critical behavior of the model,
225: a finite size scaling (FSS) ansatz is used, i.e.
226: \begin{equation}
227: \label{fss}
228: P_N(s) \simeq N^{-\beta} f(s/N^D)
229: \end{equation}
230: where $f$ is a suitable scaling function and $\beta$ and $D$ are
231: critical exponents describing the scaling of the distribution
232: function. In figure~2, a FSS collapse of $P_N(s)$ for different
233: values of $\alpha$ and for different $q$ is shown. The distribution
234: $P_N(s)$ satisfies the FSS hypothesis reasonably well, with universal
235: critical coefficients. The critical exponent derived from the fit of
236: fig.~2 are $\beta \simeq 1.65$ and $D=1$, independent of the
237: dissipation parameter $\alpha$ and the coordination number of the
238: graph $q$. The FSS hypothesis implies that, for asymptotically large
239: $N$, $P_N(s) \sim s^{-\tau}$ and the value of the exponent is $\tau=
240: \beta/D \simeq 1.65$. Due to the numerical uncertainty on the estimate
241: it is difficult to assert with certainty
242: that $\tau$ is a novel exponent, different
243: from the one for the conservative
244: RN model ($\tau=1.5$) or the lattice model in two dimensions
245: ($\tau \simeq 1.8$).
246:
247: The OFC model on a lattice does not show ordinary FSS~\cite{grass2}. Although
248: the avalanche size distribution converges to a well-defined, universal power
249: law, the cut-off in the distribution due to finite system sizes does not
250: behave according to FSS~\cite{lisepac1}. In particular, the apparent numerical
251: value for the exponent $D$ determined through FSS would violate some exact
252: bounds~\cite{klein}. In fact, in the non-conservative model each site can
253: only discharge a finite number of times during an avalanche, which
254: imposes $D \le 1$ in eq.~(\ref{fss}). In order to recover standard FSS
255: in the two dimensional lattice model one
256: has to consider earthquakes localized within subsystems of linear dimension
257: small compared to the overall system size~\cite{lisepac2}.
258:
259: We have performed numerical simulations of the OFC model on a square two
260: dimensional periodic lattice where just two sites have $3$ neighbors. We
261: have verified that the avalanche size distribution scales with system size
262: but, as for the OFC model with open boundary conditions, FSS appears to be
263: violated in the cut-off region. The power law exponent of the
264: distribution is consistent with the exponent of the OFC model with open
265: boundary conditions, i.e. $\tau \simeq 1.8$. For the range of system sizes
266: we could simulate, the critical behavior of the model on a lattice and on a
267: random graph (with the same number of defects) appear to be different.
268: In particular SOC on the quenched random graph appears to be described
269: by ordinary FSS, consistent with a mean field limit.
270:
271: We now discuss the time properties of the avalanches. In
272: fig.~3 we report the average size of an avalanche stopping at time $t$,
273: $<s>_t$, as a function of the rescaled time $\tilde{t}=t+2$ (as we are mainly
274: interested at large values of $t$, the constant should be irrelevant).
275: The curves for different system sizes overlap (deviations can be attributed
276: to finite-size effects) and we observe that $<s>_t \simeq t ^{\gamma}$, where
277: $\gamma \simeq 2.1$, providing further evidence of criticality in
278: the nonconservative system.
279:
280: An interesting question (of difficult solution though) is whether the model
281: becomes subcritical below a certain non-zero value $\alpha _c$ (for
282: $\alpha = 0$ the system is clearly not critical as sites do not interact).
283: In our simulations we have found that for $\alpha \ge 0.10$ and $q=4$ the
284: model displays scaling behavior with universal critical exponents
285: (see fig.~1 and 2). For lower values of $\alpha$ the analysis is more
286: complicated as the extremely long transient times to stationarity prevent the
287: investigation of large lattices. Nonetheless for very low values of $\alpha$
288: ($\alpha \simeq 0.03$) the cut-off in the avalanche distribution does not
289: appear to vary systematically with system size (even for relatively small
290: systems), suggesting that a non-zero $\alpha _c$ may exist.
291:
292: %From a renormalization group point of view one
293: %can conjecture the following scenario: $\alpha _c$ corresponds to an unstable
294: %fixed point so that for $ \alpha < \alpha _c$ the system renormalizes to
295: %$\alpha=0$ whereas for $ \alpha > \alpha _c$ (and $ \alpha < 1/q$) the system
296: %renormalizes to an attractive fixed point $\alpha ^*$
297: %($\alpha _c < \alpha ^* < 1/q$) which controls the critical behavior in the
298: %nonconservative regime.
299:
300: We have also considered the OFC model on a random
301: graph with variable local connectivity $q_i$. In this case, the toppling
302: rule (\ref{av_dyn}) must be modified to take into account that different
303: sites have a different coordination number $q_i$. Each site consequently
304: has a different $\alpha _i$, which we determined by requiring that the
305: total fraction $\tilde{\alpha}$ of the force transferred
306: from the unstable site to the nearest-neighbors sites is constant in the
307: system, i.e. $\alpha _i=\tilde{\alpha}/q_i$. In particular, we have
308: studied a graph with average connectivity $<q_i> = 4$. We have found that
309: there is no criticality in the system as the cut-off in the probability
310: distribution does not scale with system size. In agreement with previous
311: investigations~\cite{ceva,mousseau}, this result indicates that
312: if the disorder is too strong (as for a completely random graph) the critical
313: state is destroyed. On the other hand inhomogeneities are necessary to break
314: the periodic state the system would otherwise reach. It is a difficult
315: question to establish what is the maximum level of disorder that the system
316: can sustain without loosing its critical properties.
317:
318: % CONCLUSIONS
319:
320: In conclusion, in this paper we have investigated the OFC model on a quenched
321: random graph. We have shown that the model is critical even in the
322: nonconservative regime.
323: %and becomes subcritical only for low conservation levels.
324: This is in contrast to what happens in the annealed
325: RN OFC model which displays criticality only in the conservative case.
326: Contrary to the annealed case, a quenched random graph has an underlying
327: spatial structure so that partial synchronization of the elements of the
328: system can still occur. As a random graph can be regarded as a high
329: dimensional limit of a regular lattice, we propose that this formulation
330: represents the correct mean-field limit of the model rather than the
331: annealed random neighbor version.
332:
333: \medskip
334:
335: This work was supported by the EPSRC (UK),
336: Grant No. GR/M10823/01 and Grant No. GR/R37357/01.
337:
338:
339: \begin{thebibliography}{99}
340:
341: \bibitem{BTW} P. Bak, C. Tang, and K. Wiesenfeld, Phys. Rev. Lett. {\bf59},
342: 381 (1987); Phys. Rev. A. {\bf 38}, 364 (1988).
343:
344: \bibitem{bak_book}
345: P. Bak,
346: {\it How Nature Works: The Science of Self-Organized Criticality}
347: (Copernicus, New York, 1996).
348:
349: \bibitem{jen_book}
350: H. Jensen,
351: {\it Self-Organized Criticality}
352: (Cambridge University Press,New York, 1998).
353:
354: \bibitem{hwa-kardar}
355: T. Hwa and M. Kardar,
356: Phys. Rev. Lett. {\bf 62}, 1813 (1989).
357:
358: \bibitem{grinstein}
359: G. Grinstein, D.-H. Lee, and S. Sachdev,
360: Phys. Rev. Lett. {\bf 64}, 1927 (1990).
361:
362: \bibitem{manna_1}
363: S.S. Manna, L.B. Kiss, and J. Kertesz,
364: J. Stat. Phys. {\bf 61}, 923 (1990).
365:
366: \bibitem{ali}
367: A.A. Ali,
368: Phys. Rev. E {\bf 52}, R4595 (1995)
369:
370: \bibitem{manna_2}
371: S.S. Manna, A.D. Chakrabarti, and R. Cafiero,
372: Phys. Rev. E {\bf 60}, R5005 (1999)
373:
374: \bibitem{ofc}
375: Z. Olami, H.J.S. Feder, and K. Christensen,
376: Phys. Rev. Lett. {\bf 68}, 1244 (1992);
377: K. Christensen and Z. Olami,
378: Phys. Rev. A {\bf 46}, 1829 (1992).
379:
380: \bibitem{klein}
381: W. Klein and J. Rundle, Phys. Rev. Lett. {\bf 71}, 1288 (1993);
382: K. Christensen, Phys. Rev. Lett. {\bf 71}, 1289 (1993).
383:
384: \bibitem{carvalho}
385: J.X. Carvalho and C.P.C. Prado,
386: Phys. Rev. Lett. {\bf 84}, 4006 (2000)
387:
388: \bibitem{kim}
389: K. Christensen, D. Hamon, H.J. Jensen, and S. Lise,
390: Phys. Rev. Lett. {\bf 87}, 039801 (2001);
391: J.X. Carvalho and C.P.C. Prado,
392: Phys. Rev. Lett. {\bf 87}, 039802 (2001).
393:
394: \bibitem{lisepac1}
395: S. Lise and M.Paczuski, Phys. Rev. E, {\bf 63}, 036111 (2001).
396:
397: \bibitem{lisepac2}
398: S. Lise and M.Paczuski, Phys. Rev. E, {\bf 64}, 046111 (2001).
399:
400: \bibitem{lise} S. Lise and H. J. Jensen,
401: Phys. Rev. Lett. {\bf 76}, 2326 (1996).
402:
403: \bibitem{chabanol}
404: M. L. Chabanol and V. Hakim,
405: Phys. Rev. E {\bf 56}, 2343 (1997).
406:
407: \bibitem{broker}
408: H. M. Broker and P. Grassberger,
409: Phys. Rev. E {\bf 56}, 3944 (1997).
410:
411: \bibitem{kinouchi}
412: O. Kinouchi, S.T.R. Pinho, and C.P.C. Prado,
413: Phys. Rev. E {\bf 58}, 3997 (1998).
414:
415:
416: \bibitem{middleton}
417: A. A. Middleton and C. Tang,
418: Phys. Rev. Lett. {\bf 74}, 742 (1995).
419:
420: \bibitem{socolar}
421: J.E.S. Socolar, G. Grinstein, and C. Jayaprakash,
422: Phys. Rev E, {\bf 47}, 2366 (1993).
423:
424: \bibitem{grass2}
425: P. Grassberger,
426: Phys. Rev. E {\bf 49}, 2436 (1994).
427:
428:
429: \bibitem{note1}
430: We have also considered defects whose number scale as $\sqrt{N}$ (analogously
431: to boundary sites in the lattice model). The model still displays scaling but
432: the results appears easier to interpret with just two defects.
433:
434: \bibitem{ceva}
435: H. Ceva,
436: Phys. Rev. E {\bf 52}, 154 (1995).
437:
438: \bibitem{mousseau}
439: M. Mousseau,
440: Phys. Rev. Lett. {\bf 77}, 968 (1996).
441:
442:
443:
444: \end{thebibliography}
445:
446: %\newpage
447: \begin{figure}[hb]
448: \narrowtext
449: \epsfxsize=4in
450: \centerline{\epsffile{fig_1a.eps}}
451: \protect\vspace{0.2cm}
452: \caption[1]{\label{fig_1a}
453: Probability distribution (a) for the RN OFC model and (b) for the
454: OFC model on a random graph. In both cases, $q=4$ and $\alpha=0.10$.
455: System sizes are (a) $N=10^3$, $4\cdot 10^3$ and (b) $N=10^3$, $16 \cdot 10^3$,
456: $256 \cdot 10^3$.
457: }
458: \end{figure}
459:
460:
461: \begin{figure}[hb]
462: \narrowtext
463: \epsfxsize=4in
464: \centerline{\epsffile{fig_2.eps}}
465: \protect\vspace{0.2cm}
466: \caption[1]{\label{fig_2}
467: Finite-size scaling plots for $P_N(s)$ for (a) $q=4$, $\alpha=0.15$,
468: (b) $q=4$, $\alpha=0.20$ and (c) $q=6$, $\alpha=0.10$. System sizes are
469: $N=4\cdot 10^3$, $16 \cdot 10^3$, $64 \cdot 10^3$ and $256 \cdot 10^3$.
470: The critical exponents are $\beta=1.65$ and $D=1$.
471: For visual clarity, curves (a) and (c) have been shifted along the x axis,
472: $x \rightarrow x-1$ and $x \rightarrow x+1$, respectively.
473: }
474: \end{figure}
475:
476:
477:
478: \begin{figure}[hb]
479: \narrowtext
480: \epsfxsize=4in
481: \centerline{\epsffile{fig_3.eps}}
482: \protect\vspace{0.2cm}
483: \caption[1]{\label{fig_3}
484: Average size of an avalanche lasting $t$ time steps as a function of $t$ for
485: $q=4$ and (a) $\alpha=0.15$ and (b) $\alpha=0.20$. Different curves
486: correspond, from bottom to top, to system sizes $N=1\cdot 10^3$,
487: $4\cdot 10^3$, $16 \cdot 10^3$, $64 \cdot 10^3$ and $256 \cdot 10^3$.
488: The slope of the straight line is $\gamma =2.1$. Curve (b) has been shifted
489: along the $x$ axis, $x=x+1$, for visual clarity.
490: }
491: \end{figure}
492:
493:
494: \end{multicols}
495: \end{document}
496:
497: