cond-mat0202022/m.tex
1: \documentclass[aps,twocolumn]{revtex4}
2: \usepackage{epsfig}
3: \newcommand{\be}{\begin{equation}}
4: \newcommand{\ee}{\end{equation}}
5: 
6: \begin{document}
7: 
8: \title{Critical Behaviour of the Drossel-Schwabl Forest Fire Model}
9: 
10: \author{Peter Grassberger}
11: 
12: \affiliation{John-von-Neumann Institute for Computing, Forschungszentrum J\"ulich,
13: D-52425 J\"ulich, Germany}
14: 
15: \date{\today}
16: 
17: \begin{abstract}
18: We present high statistics Monte Carlo results for the Drossel-Schwabl 
19: forest fire model in 2 dimensions. They extend to much larger lattices (up to 
20: $65536\times 65536$) than previous simulations and reach much closer to the 
21: critical point (up to $\theta \equiv p/f = 256000$). They are incompatible 
22: with all previous conjectures for the (extrapolated) critical behaviour, 
23: although they in general agree well with previous simulations wherever they 
24: can be directly compared. Instead, they suggest 
25: that scaling laws observed in previous simulations are spurious, and that 
26: the density $\rho$ of trees in the critical state was grossly underestimated.
27: While previous simulations gave $\rho\approx 0.408$, we conjecture that $\rho$ 
28: actually is equal to the critical threshold $p_c = 0.592\ldots$ for site 
29: percolation in $d=2$. This is however still far from the densities reachable 
30: with present day computers, and we estimate that we would need many orders 
31: of magnitude higher CPU times and storage capacities to reach the true critical
32: behaviour -- which might or might not be that of ordinary percolation.
33: \end{abstract}
34: 
35: \maketitle
36: 
37: \section{Introduction}
38: 
39: Empirical analyses suggest that power laws with anomalous exponents are ubiquitous
40: in nature, ranging from $1/f$ noise and earth quake distributions to fractal 
41: coast lines, species extinction rates, weather records, and the statistics of DNA 
42: \cite{mandelbrot,bak,buldyrev,koscielny}. On the other hand, it is well known 
43: that such scaling laws -- most clearly seen by linear relationships in log-log 
44: plots -- can be spurious. Log-log plots have a notorious tendency to suggest 
45: linear curves, even if there are no real power laws. 
46: 
47: It would thus be extremely useful if these empirical observations could be 
48: backed by theoretical models where power laws can be either proven exactly or 
49: at least verified beyond doubt by high statistics simulations. Unfortunately, 
50: equilibrium systems in general show anomalous scaling laws only at critical 
51: points which are codimension one phenomena: One has to fine tune some parameter 
52: (e.g. temperature) to reach them, otherwise no power laws are obtained. Thus, 
53: they cannot be used to justify why such power laws should be seen in nature.
54: 
55: A possible solution of this puzzle was indicated in \cite{btw,bak} where it was 
56: suggested that many {\it non-equilibrium} systems could be driven by their dynamics
57: into a critical state. The main ingredients of this {\it self organized criticality}
58: (SOC) is slow driving towards some instability and a mechanism to relax the 
59: tensions built up by the drive locally and partially. Since the tensions are not 
60: relaxed completely (in many models they are just redistributed), the state becomes 
61: marginally stable and apt to relaxation events on wider and wider length scales,
62: which then lead to anomalous scaling laws. The paradigmatic model is the 
63: ``sand pile" of \cite{btw} which does not describe real sand piles but which 
64: was proven exactly to show anomalous scaling laws at least for some of its 
65: observables \cite{dhar,priezzhev,glkp}.
66: 
67: Another model which was proposed to show SOC is the forest fire model introduced 
68: independently by Henley \cite{henley} and by Drossel and Schwabl \cite{ds}. In this 
69: lattice model with discrete time each site can be either occupied by a tree or empty.
70: New trees are grown with small fixed rate $p$ on empty sites, and with a rate $f\ll
71: p$ sites are hit by lightning strokes which then burn the entire cluster of trees
72: connected to this site. Burning happens infinitely fast, so the only relevant 
73: parameter is the ratio $\theta = p/f$ which also sets the scale for the average 
74: fire size (the number of trees burnt after one lightning). Criticality is observed 
75: in the limit $\theta \to \infty$.
76: 
77: The Drossel-Schwabl model (called DS model in the following)
78: is different from other SOC models in two ways.
79: \begin{itemize}
80: \item It involves not only the separation between 
81: two time scales (the slow build-up of stress and the fast relaxation), but 
82: involves {\it three} time scales: The fast burning of connected clusters of trees,
83: the slow re-growth, and the even slower rate of lightnings.
84: \item The growth of trees does not lead to a state which is inherently 
85: unstable (as does the addition of sand grains to the top of a pile does), but only
86: to a state {\it susceptible} to being burnt. Without lightning, the tree density 
87: in any patch of forest can go far beyond criticality. When the lightning strikes 
88: finally, the surviving trees have a density far above critical.
89: \end{itemize}
90: 
91: Indeed, it was observed in \cite{grass} that the stationary steady state of 
92: the model is not ``critical" in most regions, in the sense that its tree density 
93: is not marginal for the spreading of fire. Rather, it is composed of large
94: patches of roughly uniform density, most of which are either far below or far 
95: above the critical density for spreading. Nevertheless, power laws were observed 
96: for several observables, partly because these patches occur with all sizes, 
97: so that also the fires had a broad spectrum of sizes.
98: 
99: While normal scaling with mean field exponents had been seen in \cite{ds}, all 
100: subsequent simulations \cite{grass,henley,christensen,clar,honecker} showed 
101: clear signs of anomalous scaling:
102: \begin{itemize}
103: \item The fire size distribution scaled, for small $s$ ($s$ is 
104: the number of trees burnt in one fire) and in the limit $\theta\to\infty$,
105: as $P(s) \sim s^{1-\tau}$ with $\tau \approx 2.15$;
106: \item For finite $\theta$, 
107: $P(s)$ is essentially cut off at $s_{\rm max} \sim \theta^\lambda$ with 
108: $\lambda \approx 1.1$; 
109: \item The average rms. radius $R(\theta)= \langle R^2\rangle^{1/2}$ of all 
110: fires scaled as $\theta^\nu$ with $\nu \approx 0.58$. Here, $R^2$ is the 
111: Euclidean distance of a burning tree from the site of lightning,
112: and the average is taken over all trees in all fires at fixed $\theta$; and
113: \item The rms. radius $R(s,\theta)$ of fires of fixed 
114: size $s$ scaled as $s^{1/D}$ where the fractal dimension $D$ of fires is
115: $D \approx 1.96$ \cite{clar} to $2.0$ \cite{grass}. 
116: \end{itemize}
117: Finally, the average 
118: density of trees was found to be $\rho = 0.408 - {\rm const}/\sqrt{\theta}$.
119: 
120: There were however seen already at that time large corrections to this scaling 
121: law and deviations from conventional ansatzes. Thus,
122: \begin{itemize}
123: \item The determination of $\tau$ was subject to large systematic
124: uncertainties \cite{grass};
125: \item The scaling $R \sim s^{1/D}$ was observed in \cite{grass} only for 
126: fires of typical size ($s\sim \theta$). For large $s$, $R$ was significantly 
127: larger; 
128: \item For finite $\theta$, the fire distribution $P(s)$ did not follow the 
129: normal scaling ansatz $P = s^{1-\tau} \phi(s/s_{\rm max})$ \cite{grass};
130: \item There are (at least) two differently divergent length scales 
131: \cite{honecker}: The correlation length evaluated from {\it all} pairs
132: of sites scales differently with $\theta$ than $R(\theta)$ \cite{footnote1};
133: \item Finite size behaviour is abnormal \cite{schenk0}.
134: \end{itemize}
135: 
136: In two recent publications, these problems were taken up again. In 
137: \cite{vespignani} it was claimed that they are due to non-leading 
138: corrections to scaling. In \cite{schenk} a more radical solution was 
139: proposed with two distinct classes of fires which both contribute to the 
140: scaling limit. In the latter paper also a connection to ordinary 
141: percolation was proposed, and the conclusion was backed by a ``coarse 
142: grained" model which supposedly could mimic the DS model at extremely 
143: large $\theta$ and on extremely large lattices.
144: 
145: It is the purpose of this paper to present very large simulations which 
146: show quite unambiguously that none of these describe really the true critical 
147: behaviour of the DS model. While our simulations agree perfectly with 
148: previous ones for the lattice sizes and $\theta$ values used there, we shall 
149: see that the supposed scaling laws are just transients. Even the present 
150: simulations do not reach the true asymptotic regime, but some suggestive 
151: features of the true asymptotics do emerge.
152: 
153: We describe our simulations and their results in the next section. In the 
154: last section we draw our conclusions.
155: 
156: \section{The Simulations}
157: 
158: Our simulations are straightforward, with a few subtleties. They follow 
159: Ref.\cite{grass} in taking $p\to 0$ by making no attempts to grow new trees
160: while a cluster burns. As in \cite{grass} we made {\it exactly} $\theta$
161: growth attempts on randomly chosen sites between two successive lightnings, 
162: instead of letting this number fluctuate. In the large $L$ limit this should 
163: not make any difference.  But in contrast to \cite{grass}, where a depth first
164: algorithm had been used, we used a breadth first algorithm to burn the cluster
165: \cite{footnote2}. 
166: 
167: In order to simulate very large lattices
168: with minimal storage, we use multi-spin coding, i.e. we use only one bit to 
169: store the status of a site. In this way we could simulate lattices of size 
170: $L\times L,\; L = 65536$ on computers with 1 GB main memory. Notice that we do 
171: not need to store for every tree whether it is burning or not, since the
172: burning trees are stored in a separate list which is updated at each time 
173: step. Boundary conditions were helical, i.e. sites are indexed by one scalar
174: index $i$, with neighbours having indices $i\pm 1$ and $i\pm L$, and with
175: $i+L^2\equiv i$. The largest previous simulations \cite{vespignani} had used 
176: $L=19000$.
177: 
178: We were careful to discard sufficiently long transients (between $1.6\time 10^6$ 
179: and $1.2\times 10^7$ lightnings; this is up to one order of magnitude longer 
180: than those in \cite{honecker,vespignani}) before taking statistics. This is 
181: needed since excessively large fires occur during these transients, and thus the 
182: tail of the distribution $P(s)$ is heavily influenced by them. We believe that 
183: previous analyses were affected by this problem which is easily overlooked 
184: since bulk properties (such as $\rho$ or $P(s)$ for typical $s$) show much 
185: faster convergence. The total number of fires in each run used for averaging was 
186: between $10^9$ (for $\theta\le 250$) and $9.3\times 10^6$ (for $\theta = 256000$; 
187: see Table 1). Previous authors \cite{vespignani} went only to $\theta = 32768$ 
188: with $2.5\times 10^7$ fires. Compared to that, our statistics is larger by 
189: roughly 1 order of magnitude.  All simulations were done on fast Alpha 
190: workstations. The CPU times per run varied between 15h and 4 weeks. As random 
191: number generator we used Ziff's four-tap generator with period $2^{9689}-1$ 
192: \cite{ziff}.
193: 
194: We tested for finite size effects by making for the same $\theta$ several 
195: runs on lattices of different sizes. Previously, finite size effects had been 
196: studied systematically in \cite{schenk0}, but only for much smaller lattices 
197: ($L\le 2000$; the authors of \cite{vespignani} called their analysis a 
198: finite size analysis, but they actually made a conventional scaling analysis
199: without checking the finite size behaviour). For $\theta=64000$, e.g., we 
200: made runs with $L=2^{14}, 2^{15},$ and $2^{16}$. We verified that distributions 
201: like $P(s)$, $R(s,\theta)$, or $P(t)$ ($t$ is the burning time of a
202: fire) were independent of $L$ within the statistical accuracy, i.e. any 
203: systematic $L$-dependence was masked by statistical fluctuations. 
204: 
205: Systematic
206: dependencies were seen only for averaged quantities like $\langle s\rangle$,
207: $\rho$, or $R(\theta)$. Indeed, $\rho$ can be measured either immediately
208: before or immediately after a lightning. Since each lightning burns in 
209: average $\langle s\rangle$ trees (here $\langle s\rangle$ is averaged over 
210: {\it all} lightnings, whether they hit a tree or an empty site), we have 
211: $\rho_{\rm before} - \rho_{\rm after} = \langle s\rangle /L^2$. We found
212: that $\rho_{\rm before}$ decreased with $L$, while $\rho_{\rm after}$ 
213: increased with it (see Table 1). More precisely, we have 
214: \be
215:    \rho_{\rm after}(L) \approx \rho(\infty) - 0.6\,\langle s\rangle /L^2 .
216: \ee
217: Notice that $\langle s\rangle$ is given for $L\to\infty$ exactly by 
218: $\langle s\rangle = (1-\rho)\theta$ \cite{ds}. For finite $L$, one has 
219: to replace just $\rho$ by $(\rho_{\rm before}+\rho_{\rm after})/2$, to 
220: leading order in $1/L$. This was verified in the simulations, but it tests
221: just stationarity and the absence of gross mistakes.
222: 
223: The rms. radius of fires depended on $L$ and $\theta$ in a more complicated 
224: way, see Table 1. The data were less clear in that case, but they could be 
225: fitted by 
226: \be
227:    R(\theta)_{L=\infty} - R(\theta)_L \approx 0.004\,\theta^{3/2}/L.
228: \ee
229: 
230: \begin{figure}
231: %Fig 1a,b
232: \psfig{file=fig1a.ps,width=5.8cm,angle=270}
233: \psfig{file=fig1b.ps,width=6.cm,angle=270}
234: \caption{(a) Log-log plot of $P(s)$ versus $s$, for fixed $\theta$ ranging 
235:    from 125 to 256000. Here and in Figs.2 and 5 we used logarithmic binning 
236:    with 2 per cent bin size, to suppress excessive statistical fluctuations.\\
237:    (b) Enlarged part of the same data, multiplied with $s^{1.19}$.}
238: \label{ps.ps}
239: \end{figure}
240: 
241: Results for the fire size distributions $P(s)$ are shown in Fig.1. We see the 
242: approximate power decay $s^{1-\tau}$ for $s < s_{\rm max}$ with $s_{\rm max}$ 
243: roughly proportional to $\theta$. But we see also the strong deviations from 
244: this power law first observed in \cite{grass}. Due to these deviations, $P(s)$ 
245: decreases with $\theta$ in the scaling region (making the effective value of 
246: $\tau$ increase with $\theta$), but has a growing bump at $s\approx s_{\rm 
247: max}$. In \cite{schenk} it was conjectured that the asymptotic value of $\tau$, 
248: estimated from the scaling region $1 \ll s \ll \theta$ in the limit $\theta 
249: \to \infty$, is $\tau=2.45$. This was based on heuristics and on simulations 
250: for small $\theta$ on very small lattices ($L=1300$).
251: In order to test it, we plotted $P(s)$ in Fig.1b after multiplying it with a 
252: suitable power of $s$. It is seen that $s^{1.19}\,P(s)$ becomes flat in a wider 
253: and wider region of $s$ as $\theta$ increases. For very small fires ($s < 300$)
254: $P(s)$ decreases faster than $s^{-1.19}$, but there is no indication that the 
255: slope increases with $\theta$ for $\theta > 8000$. Based on this evidence we 
256: would thus conclude that $\tau = 2.19\pm 0.01$, ruling thereby out the value of 
257: \cite{schenk} -- provided, of course, that there is no change of behaviour as 
258: $\theta$ becomes even larger.
259: 
260: \begin{figure}
261: %Fig 2a,b
262: \psfig{file=fig2a.ps,width=5.8cm,angle=270}
263: \psfig{file=fig2b.ps,width=6.cm,angle=270}
264: \caption{(a) Log-log plot of $P_{\rm int}(s)$ versus $s$, for fixed $\theta$ 
265:    ranging from 125 to 256000. \\
266:    (b) Enlarged part of the same data, multiplied with $s^{0.19}$.}
267: \label{ps_int.ps}
268: \end{figure}
269: 
270: In \cite{grass} it was conjectured that the bump near $s\approx s_{\rm max}$ is 
271: due to the cut-off. If the {\it integrated} distribution $P_{\rm int}(s) = 
272: \sum_{s'=s}^\infty P(s')$ were just a power multiplied by a sharp cut-off, its 
273: derivative $P(s)$ would have a bump where the cut-off sets in. This bump would 
274: consist of those events which would have been in the tail which is cut off. If 
275: this were right and $P_{\rm int}(s)$ would indeed show normal scaling, we should 
276: expect the height of the bump in Fig.1b to be independent of $\theta$. But this 
277: is obviously not the case. Instead, it increases with $\theta$, suggesting a 
278: different scaling law $s^{1-{\tau'}}$ for the envelope of the curves in Fig.1a, 
279: with $\tau' = 2.111\pm 0.006$.
280: 
281: A similar conclusion is reached by looking directly at $P_{\rm int}(s)$.
282: Indeed, log-log plots of $P_{\rm int}(s)$ versus $s$ are much more straight 
283: in the scaling region (Fig.2a). But again a blow-up after multiplication by a 
284: suitable power of $s$ shows that this is misleading (Fig.2b).  Even $P_{\rm int}(s)$ 
285: is not convex for $\theta > 1000$, and it develops an increasingly sharp shoulder 
286: near $s=s_{\rm max}$ as $\theta \to\infty$. We can try to obtain an alternative 
287: estimate of $\tau$ by fitting straight lines such that they touch both maxima in 
288: Fig.2b. Results of this are shown in Fig.3. In contrast to the previous estimate of 
289: $\tau$ they would indicate that the effective $\tau$ decreases with $\theta$ and is 
290: clearly less than 2.19 for $\theta\to\infty$ (also the value 2.159 of \cite{honecker}
291: seems hardly compatible with the extrapolation of Fig.3 to $\theta\to\infty$). But 
292: a convergence to $2.11$ seems quite possible, as we would expect if the bumps in 
293: fig.1b have a width independent of $\theta$ (which is not excluded by the data).
294: 
295: This obviously means that these scaling 
296: violations are not corrections which disappear in the limit $\theta\to\infty$, as 
297: was claimed in \cite{vespignani}. It rather indicates that the conventional scaling 
298: picture, and in particular the usual ansatz 
299: \be 
300:    P(s) = s^{1-\tau}\phi(s/s_{\rm max}),
301: \ee
302: is basically wrong, as was already conjectured in \cite{grass,schenk}.
303: 
304: \begin{figure}
305: %Fig 3
306: \psfig{file=fig3.ps,width=6.cm,angle=270}
307: \caption{Log-log plot of effective $\tau$ values, obtained by drawing lines 
308:    tangent to both maxima in Fig.2b. Notice that this is only defined when 
309:    $P_{\rm int}(s)$ is not convex, i.e. only for $\theta \ge 1000$.
310:    Error bars are roughly of the size of the symbols.}
311: \label{tau.ps}
312: \end{figure}
313: 
314: Previous analyses indicated that $s_{\rm max}$ actually increased faster than 
315: $\theta$, roughly as 
316: \be
317:    s_{\rm max} \sim \theta^\lambda               \label{lambda}
318: \ee
319: with $\lambda\approx 1.08$. We verified this qualitatively, but verified also the 
320: finding of \cite{grass} that $s_{\rm max}$ and thus also $\lambda$ are not well 
321: defined since the sharpness of the cut-off increases with $\theta$. This is already 
322: obvious from Fig.2b, but it persists for larger values of $s$ not shown in this 
323: plot.  It can also be seen by making copies of Fig.2a on transparencies and 
324: overlaying them.
325: 
326: Better scaling than for $P(s)$ was seen in \cite{grass} for $R(\theta)$. Our present 
327: data are fully compatible with those of \cite{grass,clar,honecker,footnote1}, but 
328: involve much higher statistics and cover a wider range of $\theta$. Thus it might not 
329: be too surprising that we now do not see perfect scaling any more. But the observed 
330: deviations from a pure power law (see Fig.4) are much larger than expected from 
331: subasymptotic corrections \cite{footnote3}. They clearly show that the previously 
332: seen power law was spurious and does not describe the asymptotic behaviour. A power 
333: law fit through the last two points would give $R(\theta)\sim \theta^{0.563}$, but 
334: this is obviously not yet the asymptotic behaviour. Again, as for $P(s)$ and 
335: $P_{\rm int}(s)$, we cannot yet say what the correct asymptotic behaviour will be.
336: In any case, the claim of \cite{honecker} that there are two different diverging
337: correlation lengths becomes obsolete.
338: 
339: \begin{figure}
340: %Fig 4
341: \psfig{file=fig4.ps,width=6.cm,angle=270}
342: \caption{Log-linear plot of $\theta^{0.582}\, R(\theta)$ versus $\theta$,
343:    where the power of $\theta$ was chosen such that the data fall roughly on a 
344:    horizontal line. Without the factor $\theta^{0.582}$, the data would be 
345:    hardly distinguishable from a straight line on a log-log plot.}
346: \label{nu.ps}
347: \end{figure}
348: 
349: \begin{figure}
350: %Fig 5
351: \psfig{file=fig5.ps,width=6.cm,angle=270}
352: \caption{Log-linear plot of $R(s,\theta) / \sqrt{s}$ versus $s$, for selected
353:    values of $\theta$.}
354: \label{rs.ps}
355: \end{figure}
356: 
357: \begin{figure}
358: %Fig 6
359: \psfig{file=fig6.ps,width=6.cm,angle=270}
360: \caption{Plot of $\nu_{\rm eff}$ versus $1/\theta$, where $\nu_{\rm eff}$ is 
361:    defined by $R(s,\theta)\sim s^{\nu_{\rm eff}}$ at the inflection points in Fig.5.}
362: \label{nuu.ps}
363: \end{figure}
364: 
365: An indication of the origin of all these puzzles comes from looking at $R(s,\theta)$.
366: This is defined analogously to $R(\theta)$, except that the averaging is now done 
367: over all fires of fixed size $s$. We expect $R(s,\theta) \sim s^{1/D}$ if fires are 
368: fractal with dimension $D$. Previous analyses had given $D\approx 2$ (or slightly
369: less). Thus we do not plot $R(s,\theta)$ versus $s$, but $R(s,\theta)/s^{1/2}$. 
370: From Fig.5 we see that there are 3 distinct regions which by and large coincide
371: with the three regions (left bump, flat middle part, and right peak) in Fig.1b.
372: \begin{itemize}
373: \item Region I: These are very small clusters, of size $s<300$. For them $R(s,\theta)$
374: increases faster than $\sqrt{s}$, which might suggest $D<2$. But we rather interpret
375: this as a finite-cluster artifact. Indeed, even compact clusters (with $D=2$) with 
376: a crumpled boundary will show an effective fractal dimension $<2$. These small 
377: clusters arise from lightnings which hit either very small regions with supercritical
378: tree density which burn off completely, or regions of subcritical density on which 
379: the fires form subcritical percolation clusters. Both mechanisms give compact clusters 
380: with fuzzy boundary. In any case, region I becomes less and less important as $\theta$
381: increases. \\
382: \item Region II: This is roughly equal to the scaling region in fig.1b. Here $R(s)$ is 
383: nearly proportional to $\sqrt{s}$, i.e. $D$ is very close 
384: to 2 -- but not quite. Also, we see a clear decrease of the minimal slope in Fig.5
385: with $\theta$, but it seems not to be sufficient to give $D=2$ in the limit $\theta\to
386: \infty$ (see Fig.6). According to Fig.6 the critical exponent $\nu$ defined by 
387: $R(s,\theta) \sim s^\nu$ in region II seems to converge to 0.505 for $\theta\to\infty$,
388: corresponding to $D = 1.98$. This is half way between the best previous estimates
389: $D=1.96$ \cite{clar} and $D=2.0$ \cite{grass}. It indicates that clusters in Region
390: II are fractal, but more compact than critical percolation clusters ($D= 1.89$).\\
391: \item Region III: In the region of very large fires, corresponding to the increasing 
392: bumps in Fig.1b, the apparent fractal dimension decreases again. Unfortunately, due
393: to the strong curvatures of the curves in Fig.5 in this region, we cannot quote 
394: any definite dimension value. But by plotting $R(s,\theta)/s^\nu$ with suitable 
395: values of $\nu$ on log-log scales, we see that (i) the maximal values of $\nu$ 
396: increase slowly with $\theta$ (except for the very largest $\theta$ where statistics 
397: is poor); and (ii) for our largest $\theta$ and $s$, we have 
398: $\nu \approx 0.64$. Thus the largest fires are definitely more fractal than critical
399: percolation clusters!
400: \end{itemize}
401: 
402: In terms of the scenario with roughly homogeneous patches with constant tree 
403: density \cite{grass,clar,schenk}, fires in region II correspond to single patches 
404: of typical size which are hit by lightning just when their density has reached 
405: about the critical percolation threshold. If lightning would always strike 
406: exactly at criticality, and if all trees in the patch would have burnt during the 
407: last fire so that all trees now are placed randomly, this would give $D=1.89$. Fires
408: in region II are more compact mainly because it will take some time until a 
409: lightning strikes by chance the region, after it has reached the critical density.
410: 
411: The larger a patch is, the bigger will be its perimeter and thus also the chance
412: that a fire ``spills over" and burns also a neighbouring patch, and from this 
413: also a next patch, etc. Since this will happen only along parts of the perimeter,
414: the resulting fires will be rather fuzzy, with effective dimension $<2$. We propose
415: that such fires dominate in region III. Although they are rather few in number, they
416: are very important since they burn large parts of the entire lattice, and they 
417: lead to rearrangements of the global pattern of patches. Notice that fires in region 
418: III burn only small parts of the entire region they cover, leaving behind more 
419: unburnt trees than fires of type II. Since they dominate more and more as 
420: $\theta$ increases, one might suspect that this leads to an increase of $\rho$
421: with $\theta$.
422: 
423: \begin{figure}
424: %Fig 7
425: \psfig{file=fig7.ps,width=6.cm,angle=270}
426: \caption{Plot of $\rho$ versus $1/\sqrt{\theta}$. These values were extrapolated 
427:    to $L\to\infty$ using Eq.(1), but for the largest $L$ the extrapolation shifted
428:    them by less than the error bars (which are indicated {\it inside} the squares).
429:    The dashed line is a fit of the small
430:    $\theta$ values ($\theta \le 8000$) to an ansatz $\rho = a-b/ \sqrt{\theta}$.}
431: \label{rho.ps}
432: \end{figure}
433: 
434: Indeed, a slight increase of $\rho$ with $\theta$ had been seen in all previous 
435: simulations. The best previous estimates were $\rho = 0.4075 - const / \sqrt{\theta}$ 
436: \cite{grass} (unfortunately, the constant multiplying $\theta^{-1/2}$ was 
437: estimated wrongly in \cite{grass} due to a simple mistake) and 
438: $\rho = (0.4084\pm 0.0005)-const/\theta^{0.47}$ \cite{honecker,vespignani}. 
439: The results of our present simulations are shown in Fig.7. For
440: $\theta < 10^4$ we see a perfect agreement with previous results, but for larger
441: $\theta$ there is dramatic disagreement: Our measured values are higher than 
442: predicted by extrapolation from small $\theta$, by up to 100 standard deviations.
443: It is not clear why this was missed in \cite{vespignani}. But we might mention that 
444: no data for $\theta>10000$ are shown in Fig.1 of \cite{vespignani}, although the 
445: authors claimed to have made high statistics measurements up to $\theta=32768$.
446: 
447: When plotted against $\log\theta$, our values of $\rho$ follow roughly a straight 
448: line for $\theta \ge 4000$. We should of course not take this as the asymptotic 
449: behaviour, since $\rho$ can never increase beyond $p_c = 0.5927\ldots$ which is 
450: the critical value for site percolation in 2 dimensions. But we can use it to 
451: obtain a very crude estimate of the order of magnitude when $\rho\approx p_c$ 
452: should be reached. It is $\theta \approx 10^{40}$.
453: 
454: Although any prediction based on such a large extrapolation should be taken with 
455: great care, we conjecture that indeed $\rho$ converges to $p_c$ for $\theta\to
456: \infty$. The main reason is that we see no other plausible scenario compatible with 
457: our present numerics. It is not clear whether fires in this limit correspond to 
458: critical percolation clusters, since we must expect that weak correlations in the 
459: tree densities survive in this limit, sufficiently so to spoil any agreement with 
460: uncorrelated percolation on large scales.
461: 
462: \begin{figure}
463: %Fig 8
464: \psfig{file=fig8.ps,width=6.cm,angle=270}
465: \caption{Rescaled average fire sizes for fixed local density.}
466: \label{s_loc.ps}
467: \end{figure}
468: 
469: A last hint in favour of our claim that large fires asymptotically are 
470: dominantly associated to regions of critical tree densities (i.e., $\rho 
471: \approx p_c$), even though they might not form critical percolation clusters, 
472: is obtained by studying the mean sizes of fires which started in regions 
473: of given tree density. If a lightning hits a tree at site $i$, we define 
474: the local tree density $\rho_{\rm loc}$ at this site as the number of other 
475: trees in the surrounding square of size $9\times 9$ divided by 80. This size 
476: is of course rather arbitrary, but we obtained qualitatively similar results
477: for squares of sizes $7\times 7$ and $5\times 5$. In Fig.8 we plot the average
478: fire size $s(\rho_{\rm loc})$, divided by the overall mean $\langle s\rangle$, 
479: against $\rho_{\rm loc}$. For small $\theta$ we see a monotonic increase 
480: which is easily understood: Since even the largest patches of uniform density 
481: are not much larger than the square, large fires can only result from regions 
482: with high local density. This is no longer true for large $\theta$. There, patches 
483: with very high density are probably small, otherwise they would already have 
484: been burnt down earlier. At values of $\theta$ reached in this work the largest 
485: $s(\rho_{\rm loc})$ are still for large $\rho_{\rm loc}$, but a pronounced peak at
486: $\rho\approx p_c$ develops where $s(\rho_{\rm loc})$ has a local maximum. Since 
487: also the number of sites with $\rho_{\rm loc}\approx p_c$ is much larger than 
488: those with $\rho_{\rm loc}\gg p_c$ (see next paragraph), we see that it is fires 
489: in regions with critical density which play an increasing important role as 
490: $\theta\to\infty$.  In any case, the very strong dependence of $s(\rho_{\rm loc})$ 
491: on $\theta$ shows that we are still far from the asymptotic region where we 
492: expect this dependence to disappear.
493: 
494: \begin{figure}
495: %Fig 9
496: \psfig{file=fig9.ps,width=6.cm,angle=270}
497: \caption{Distribution of local densities.}
498: \label{P_loc.ps}
499: \end{figure}
500: 
501: Finally, we show in Fig.9 the distribution of local densities itself. It depends
502: rather weakly on $\theta$ (less than $s(\rho_{\rm loc})$, at least), but the 
503: precise way it does depend on $\theta$ is rather surprising and not yet fully
504: understood. First of all, it develops an increasingly sharp maximum which slowly 
505: shifts from $\rho_{\rm loc}\approx 0.4$ to $\rho_{\rm loc}\approx 0.6$ as 
506: $\theta$ increases. This is to be expected after our previous observations. What 
507: is unexpected and hard to explain is a shoulder at  $\rho_{\rm loc}\approx 0.8$
508: which develops for the largest $\theta$ values. It is not very large but 
509: statistically highly significant (it was seen in all runs with $\theta \ge 32000$).
510: Presumably related to it is a shoulder at small $\rho_{\rm loc}$ ($\approx 0.1$) 
511: which also increases with $\theta$: If patches with density $> 0.7$ burn, they 
512: leave behind extremely strongly depleted patches. One possible reason why such 
513: high density patches can survive at all is that many large fires are fractal 
514: and leave behind small disconnected regions of fairly high but subcritical 
515: density. These regions then are too small to have a chance to be hit by 
516: lightning until their density has grown far beyond $p_c$.
517: 
518: \section{Discussion}
519: 
520: The simulations reported in this paper leave little doubt that all scenarios 
521: proposed so far for the Drossel-Schwabl forest fire model do not 
522: describe the asymptotic behaviour for $\theta\to\infty$, where the model
523: should show SOC according to the standard folklore. Indeed, we do not see 
524: much indications for {\it any} power laws in this model, as all proposed 
525: scaling laws seem to be just transient. There are a number of observables which 
526: do show straight lines on log-log plots (such as Fig.1a in the central region
527: of $s$ or the envelope to Fig.1a), but it seems more likely that also these are 
528: spurious.
529: 
530: This situation is of course not altogether new. There are a number of models 
531: which were supposed to show anomalous scaling, but closer scrutiny proved 
532: otherwise. A good example is the Bak-Chen-Tang forest fire model \cite{bct}
533: which at least in $d=2$ is not critical \cite{gk}. Other examples include
534: the Newman-Sneppen model \cite{newman} where one can prove semi-numerically 
535: that the observed power laws are transient \cite{grass-unpub} and maybe even the 
536: ``classical" abelian sandpile model in $d=2$. While power laws for {\it waves} 
537: were proven rigorously in that model, it might well be that the observed 
538: deviations from finite size scaling \cite{demenech,tebaldi} do not herald 
539: multifractality but just simply no scaling at all. One indication for the 
540: latter is the fact that some scaling laws show violations which do not seem 
541: do vanish for increasing system sizes \cite{glkp}. Also, some other quantities 
542: in the sandpile model which involve superpositions of many waves depend {\it 
543: qualitatively} on the geometry of the lattice (square {\it vs.} strip)
544: \cite{grass-unpub}. For a system with true scaling one would not expect this.
545: 
546: The situation becomes even worse when going to real life phenomena. It does not 
547: seem unlikely that many of the observed scaling laws are just artifacts or 
548: transients. Problems of spurious scaling in models which can be simulated with 
549: very high precision such as the present one should be warnings that not every 
550: power law supposedly seen in nature is a real power law.
551: 
552: Acknowledgement: I am indebted to Walter Nadler for critically reading the 
553: manuscript. 
554: 
555: Note added: After finishing this paper, I was informed that similar 
556: conclusions had been reached in a recent paper \cite{pruessner}, where 
557: the authors studied $P(s)$ with comparable (or even larger) statistics but 
558: on somewhat smaller lattices and for somewhat smaller $\theta$ than in the 
559: present paper. Unfortunately, neither the spatial sizes of fires nor 
560: precise values of $\rho(\theta)$ were measured in \cite{pruessner}. I am 
561: indebted to Gunnar Pruessner for sending me this paper, for very helpful 
562: correspondence and for pointing out some misprints in my manuscript.
563: 
564: 
565: \begin{table*}
566: \begin{center}
567: %\begin{ruledtabular}
568: \begin{tabular}{|c|c|r|c|r@{ $\pm$ }l|r@{ $\pm$ }l|} \hline
569:  $\qquad\theta\qquad$ & $\;\;\log_2L\;\;$ & $N\qquad$ & $\quad N_{trans}\quad$ 
570:                   &\multicolumn{2}{c|}{density} &\multicolumn{2}{c|}{$R(\theta)$}\\ \hline
571:  256000   &  16 &$   9.3\times 10^6$ & $2.0\times 10^6$ & 0.410667 & 0.000036 & 2635.138&2.59   \\ \hline
572:  128000   &  16 &$   7.5\times 10^6$ & $3.0\times 10^6$ & 0.409321 & 0.000042 & 1789.675&1.96   \\ 
573:           &  15 &$  15.2\times 10^6$ & $1.6\times 10^6$ & 0.409231 & 0.000043 & 1787.024&1.37   \\ \hline
574:   64000   &  16 &$  46.4\times 10^6$ & $6.0\times 10^6$ & 0.407908 & 0.000013 & 1205.893&0.55   \\ 
575:           &  15 &$  42.5\times 10^6$ & $3.0\times 10^6$ & 0.407904 & 0.000037 & 1203.097&0.56   \\ 
576:           &  14 &$  37.7\times 10^6$ & $1.0\times 10^6$ & 0.407832 & 0.000038 & 1202.836&0.59   \\ \hline
577:   32000   &  15 &$  42.3\times 10^6$ & $6.0\times 10^6$ & 0.406623 & 0.000038 &  808.948&0.37   \\ 
578:           &  14 &$  78.3\times 10^6$ & $3.0\times 10^6$ & 0.406633 & 0.000037 &  807.253&0.27   \\ \hline
579:   16000   &  15 &$ 106.8\times 10^6$ & $8.0\times 10^6$ & 0.405428 & 0.000009 &  539.719&0.16   \\ 
580:           &  14 &$ 103.3\times 10^6$ & $5.0\times 10^6$ & 0.405401 & 0.000017 &  539.063&0.16   \\ 
581:           &  13 &$  73.2\times 10^6$ & $2.0\times 10^6$ & 0.405360 & 0.000029 &  538.464&0.20   \\ \hline
582:    8000   &  14 &$ 166.3\times 10^6$ & $6.0\times 10^6$ & 0.404188 & 0.000008 &  359.049&0.084  \\ 
583:           &  13 &$ 110.9\times 10^6$ & $3.0\times 10^6$ & 0.404135 & 0.000028 &  358.703&0.102  \\ \hline
584:    4000   &  14 &$ 207.8\times 10^6$ & $8.0\times 10^6$ & 0.402690 & 0.000007 &  239.213&0.050  \\ 
585:           &  13 &$ 394.2\times 10^6$ & $4.0\times 10^6$ & 0.402691 & 0.000010 &  239.207&0.041  \\ 
586:           &  12 &$ 332.7\times 10^6$ & $2.0\times 10^6$ & 0.402595 & 0.000011 &  238.992&0.044  \\ \hline
587:    2000   &  13 &$ 366.4\times 10^6$ & $6.0\times 10^6$ & 0.400614 & 0.000005 &  159.511&0.026  \\ 
588:           &  12 &$ 525.7\times 10^6$ & $2.0\times 10^6$ & 0.400577 & 0.000012 &  159.441&0.021  \\ \hline
589:    1000   &  13 &$ 319.6\times 10^6$ & $9.0\times 10^6$ & 0.397636 & 0.000005 &  106.647&0.018  \\ 
590:           &  12 &$ 458.4\times 10^6$ & $6.0\times 10^6$ & 0.397624 & 0.000006 &  106.681&0.015  \\ 
591:           &  11 &$ 655.8\times 10^6$ & $4.0\times 10^6$ & 0.397570 & 0.000010 &  106.603&0.013  \\ \hline
592:     500   &  12 &$ 490.8\times 10^6$ & $6.0\times 10^6$ & 0.393419 & 0.000006 &   71.414&0.0097 \\ \hline
593:           &  11 &$ 664.9\times 10^6$ & $3.0\times 10^6$ & 0.393399 & 0.000010 &   71.420&0.0083 \\ \hline
594:     350   &  12 &$ 663.4\times 10^6$ & $8.0\times 10^6$ & 0.390667 & 0.000004 &   58.093&0.0068 \\ \hline
595:     250   &  12 &$1000.0\times 10^6$ & $9.0\times 10^6$ & 0.387643 & 0.000003 &   47.757&0.0045 \\ 
596:           &  11 &$ 664.4\times 10^6$ & $6.0\times 10^6$ & 0.387626 & 0.000007 &   47.768&0.0056 \\ 
597:           &  10 &$1000.0\times 10^6$ & $2.0\times 10^6$ & 0.387574 & 0.000008 &   47.732&0.0045 \\ \hline
598:     175   &  11 &$ 733.6\times 10^6$ & $6.0\times 10^6$ & 0.383909 & 0.000006 &   38.788&0.0043 \\ \hline
599:     125   &  11 &$ 922.7\times 10^6$ & $8.0\times 10^6$ & 0.379837 & 0.000006 &   31.847&0.0032 \\ 
600:           &  10 &$1000.0\times 10^6$ & $3.0\times 10^6$ & 0.379804 & 0.000008 &   31.850&0.0030 \\ \hline
601:      88   &  11 &$1000.0\times 10^6$ &$12.0\times 10^6$ & 0.374935 & 0.000004 &   25.903&0.0025 \\ 
602:           &  10 &$1000.0\times 10^6$ & $8.0\times 10^6$ & 0.374910 & 0.000005 &   25.906&0.0025 \\ \hline
603: \end{tabular}
604: %\end{ruledtabular}
605: \caption{Statistics and main results: $N$ is the number of lightnings used for averaging, 
606: $N_{trans}$ that of lightnings discarded during the transients. The density is measured after
607: fires got extinct and before new trees are grown. Notice that the density has much larger
608: errors in some runs than in others of comparable statistics. This results from the fact that 
609: there are important long-ranged {\it negative} autocorrelations in the density time series, 
610: and I had not written out the information needed to take them into account in these runs.}
611: \end{center}
612: \end{table*}
613: 
614: 
615: \begin{thebibliography}{99}
616: \bibitem{mandelbrot} B. Mandelbrot, {\it The Fractal Geometry of Nature} (W.H. Freeman,
617:    San Francisco 1982).
618: \bibitem{bak} P. Bak, {\it How Nature Works: the science of self-organized criticality}
619:    (Copernicus Press, New York 1996).
620: \bibitem{buldyrev} S.V. Buldyrev {\it et al.}, Physica {\bf A 249}, 430 (1998).
621: \bibitem{koscielny} E. Koscielny-Bunde {\it et al.}, Phys.  Rev.  Lett.  {\bf 81}, 729
622:    (1998).
623: \bibitem{btw} P. Bak, C. Tang, and K. Wiesenfeld, Phys. Rev. Lett. {\bf 59}, 381 (1987)
624: \bibitem{dhar} D. Dhar, Rev. Lett. {\bf 64},  1613  (1990).
625: \bibitem{priezzhev} E.V. Ivashkevich, D.V. Ktitarev, and V.B. Priezzhev, Physica A {\bf
626:    209}, 347 (1994); J. Phys. A {\bf 27}, L585 (1994).
627: \bibitem{glkp} D.V. Ktitarev, S. L\"ubeck, P. Grassberger and V.B. Priezzhev, Phys. Rev.
628:    E {\bf 61}, 81 (2000).
629: \bibitem{henley} C.L. Henley, Bull. Am. Phys. Soc. {\bf 34}, 838 (1989); 
630:    Phys. Rev. Lett. {\bf 71}, 2741 (1993).
631: \bibitem{ds} B. Drossel and F. Schwabl, Phys. Rev. Lett. {\bf 69}, 1629 (1992).
632: %  Physica A {\bf 191}, 47 (1992).
633: \bibitem{grass} P. Grassberger, J. Phys. A {\bf 26}, 2081 (1993).
634: \bibitem{christensen} K. Christensen, H. Flyvbjerg, and Z. Olami, Phys.  Rev.  Lett. 
635:    {\bf 71}, 2737 (1993).
636: \bibitem{clar} S. Clar, B. Drossel and F. Schwabl, Phys. Rev. {\bf E 50}, 1009 (1994).
637: \bibitem{honecker} A. Honecker and I. Peschel, Physica A {\bf 239}, 509 (1997).
638: \bibitem{footnote1} The length $\xi_c$ defined in \cite{honecker} is closely related
639:    to $R(\theta)$. With a slight change of notation compared to \cite{honecker}, it 
640:    is given by $\xi_c^2 = \sum_{y=1}^\infty y^2 K(y{\bf e}_1)/\sum_{y=1}^\infty 
641:    K(y{\bf e}_1)$ with $K({\bf y}) = \langle T({\bf x})T({\bf x}+{\bf y})\rangle_c$,
642:    while $R^2(\theta) = \sum_{{\bf y}\in Z^2} y^2 K({\bf y})/\sum_{{\bf y}\in Z^2}
643:    K({\bf y})$.
644: \bibitem{schenk0}  K. Schenk, B. Drossel, S. Clar, and F. Schwabl, Eur. Phys. J. B 
645:    {\bf 15}, 177 (2000).
646: \bibitem{vespignani} R. Pastor-Satorras and A. Vespignani, Phys. Rev. {\bf E 61}, 
647:    4854 (2000).
648: \bibitem{schenk}  K. Schenk, B. Drossel, and F. Schwabl, preprint cond-mat/0105121
649:    (2001).
650: \bibitem{footnote2} The burning times quoted in \cite{grass} were obtained from the 
651:    stack hights in a recursive (depth-first) implementation. They do not agree with 
652:    the more realistic burning times obtained with breadth first algorithms and should 
653:    thus be considered as obsolete.
654: \bibitem{ziff} R.M. Ziff, Computers in Physics {\bf 12}, 385 (1998) (cond-mat/9710104).
655: \bibitem{footnote3} These deviations from scaling are qualitatively similar to those 
656:    seen on closer inspection in Fig.1 of \cite{honecker} for the correlation length 
657:    $\xi$, but not for $\xi_c$ which should be more closely related to $R(\theta)$.
658: \bibitem{bct} P. Bak, K. Chen, and C. Tang, Phys. Lett. {\bf A 147}, 297 (1990).
659: \bibitem{gk} P. Grassberger and H. Kantz, J. Stat. Phys. {\bf 63}, 685 (1991).
660: \bibitem{newman} M.E.J. Newman and K. Sneppen, Physica D {\bf 110} 209 (1997). 
661: \bibitem{demenech} M. De Menech, A.L. Stella, and C. Tebaldi, Phys. Rev. E {\bf 58}, 
662:    R2677 (1998).
663: \bibitem{tebaldi} C. Tebaldi, M. De Menech, and A.L. Stella, Phys. Rev. Lett. {\bf 83},
664:    3952 (1999).
665: \bibitem{grass-unpub} P. Grassberger, unpublished
666: \bibitem{pruessner} G. Pruessner and H.J. Jensen, preprint cond-mat/0201306 (2002)
667: \end{thebibliography}
668: 
669: \end{document}
670: