1: \documentclass[prl,twocolumn,amssymb,showpacs]{revtex4}
2: \usepackage[T1]{fontenc}
3: \usepackage[latin1]{inputenc}
4: \usepackage{graphics}
5:
6: \makeatletter
7:
8: \makeatother
9:
10: \begin{document}
11:
12:
13: \title{Entropy Analysis of Stochastic Processes at Finite Resolution}
14: \author{D.M. Tavares}
15: \author{L.S. Lucena}
16: \affiliation{International Center for Complex Systems}
17: \address{Departamento de Física Teórica e Experimental - UFRN \\
18: Natal-RN 59078-970 Brazil}
19:
20: \begin{abstract}
21: The time evolution of complex systems usually can be described
22: through stochastic processes. These processes are measured at
23: finite resolution, what necessarily reduces them to finite
24: sequences of real numbers. In order to relate these data sets to
25: realizations of the original stochastic processes (to any
26: functions, indeed) it is obligatory to choose an interpolation
27: space (for example, the space of band-limited functions). Clearly,
28: this choice is crucial if the intent is to approximate optimally
29: the original processes inside the interval of measurement. Here we
30: argue that discrete wavelets are suitable to this end. The wavelet
31: approximations of stochastic processes allow us to define an
32: entropy measure for the order-disorder balance of evolution
33: regimes of complex systems, where order is understood as
34: confinement of energy in simple local modes. We calculate exact
35: results for the fractional Brownian motion (fBm), with application
36: to Kolmogorov K41 theory for fully developed turbulence.
37: \end{abstract}
38:
39: \pacs{05.45.Tp, 05.65.+b, 89.75.Fb, 02.50.-r}
40:
41: \maketitle
42:
43: Many physical systems investigated at present have a complex
44: evolution in time. Frequently, the major information we can obtain
45: on their dynamics comes from time series of noisy
46: appearance~\cite{consolini,politzer,santi,sreenivasan,gafarov}.
47: These series are samples at finite resolution of subjacent
48: stochastic processes, whose properties are better investigated
49: through a multi-resolution approach, because the realizations of
50: the mentioned stochastic processes are singular everywhere
51: functions. This is, for example, the case of \(1/f^{\alpha}\)
52: noises from self-organized
53: systems~\cite{bak,durin,urbach,cote,carreras}. Thus, singularities
54: are common, and should be interpreted as details that influence
55: function variation at all scales~\cite{muzy,mallat}. In this paper
56: we propose a method for entropy analysis of arbitrary complex time
57: series that takes advantage of this theoretical standpoint.
58: Moreover, we account for the fact that measurements are made at
59: finite resolution, considering the consequences of sampling, what
60: is not fully accomplished in previous approaches. In general,
61: other formulations are strongly influenced by
62: information-theoretical arguments, applied to the analysis of
63: chaotic behavior~\cite{bandt}. We see difficulties in two main
64: aspects. First, the notion of complexity that is employed in these
65: formulations is based on entropy rates, like the Kolmogorov-Sinai
66: entropy, that measure the degree in which information on initial
67: conditions is lost when the systems evolve. This gives a scale of
68: complexity ranging from zero (non-chaotic deterministic case) to
69: infinite (stochastic case). In this scale, finite values quantify
70: the complexity of deterministic chaos. As a consequence, the
71: problem of defining a proper complexity estimator for stochastic
72: processes is substituted by the statistical investigation of
73: chaotic deterministic dynamical systems~\cite{eckmann}. Second,
74: these formulations do not consider the relationship that scaling
75: has with disorder~\cite{madalena}. Our approach is tailored to
76: face these difficulties. Here, we begin by assuming that the
77: stochastic processes are supported on the real axis, and measured
78: at a discrete set of points with a sampling interval \( \tau \),
79: during a time \(T\). The resolution of the measurement is \(
80: N=T/\tau \), and for convenience we make \( T=1 \), and \( \tau
81: =1/N=2^{-J} \) (we will work in these units). Such measurement
82: results in a loss of information, which depends on two factors:
83: The resolution, and the interpolation space in which the
84: stochastic processes will be projected~\cite{unser}. The idea is
85: that the mere sampled values say almost nothing about a function.
86: Much more information is conveyed through the hypothesis on how
87: the function varies between the sampled values. This regularity
88: hypothesis, usually implicit when we draw smooth curves between
89: data points, is an essential element of the theory. Without this
90: hypothesis, no information found on the discrete and finite data
91: sets can be attributed to the subjacent model, that is assumed to
92: hold on the continuous support. Now, in the time-scale
93: (time-frequency) domain, it makes sense to search a representation
94: \emph{i}) that is minimally affected by the created end
95: singularities, \emph{ii}) that provides an interpolation based on
96: multi-scale approximation of the actual singularities, and
97: \emph{iii}) that deals with the resolution \(N\) as a direct
98: experimental parameter. The first and second requirements are the
99: most crucial, and establish the way in which the energy
100: (\(L^2\)-norm squared) is assumed to be distributed inside the
101: interval of measurement, so that this distribution corresponds to
102: the singularity structure of the process, seen at a finite
103: resolution. As a supplementary condition, the best is that the
104: algorithmic complexity grows only linearly with the length of the
105: time series. These requirements are met if we project the
106: stochastic processes in a discrete wavelet space.
107:
108: Wavelets are associated with time-scale representations. Let \(
109: \left| F\right\rangle \) be a vector in a Hilbert space. There
110: are bases \( \left\{ \left| \psi _{jn}\right\rangle \right\} \)
111: such that \( \left| F\right\rangle \) can be expanded as
112:
113:
114: \begin{equation}
115: \label{eq:waveletexpansion1} \left| F\right\rangle =\sum
116: ^{j=\infty }_{j=-\infty }\sum ^{n=\infty }_{n=-\infty
117: }\left\langle \psi _{jn}|F\right\rangle \left| \psi
118: _{jn}\right\rangle ,
119: \end{equation}
120: the form of the basis functions being \( \psi _{jn}\left( t\right)
121: =\frac{1}{\sqrt{2^{j}}}\psi \left( \frac{t-2^{j}n}{2^{j}}\right) .
122: \) These functions are called discrete wavelets, and are square
123: integrable with zero mean, and may have \(p\) vanishing moments.
124: The discrete wavelet transform of \( F \) to this basis is written
125: \( D_{jn}F\equiv\left\langle \psi _{jn}|F\right\rangle ,\) and
126: gives information on the behavior of the function \( F \) at scale
127: \( 2^{j} \) and time \( 2^{j}n \)~\cite{mallat}.
128:
129: \begin{figure}
130: {\par\centering \resizebox*{6cm}{6cm}{\includegraphics{fig1.eps}}
131: \par}
132: \caption{The Haar wavelet decomposition of a stochastic process
133: sampled at the finite resolution \(N=2^{15}\). The details are
134: presented from scale the \(2^{-14}\) to the scale \(2^{-6}\). Note
135: how the energy accumulates at the small scale details. This causes
136: the high entropy \(S=12.52\) as compared to the maximum
137: \(S_{\rm{max}}=15\).\label{fig:coefstable}}
138: \end{figure}
139:
140: Confining the function in expression~(\ref{eq:waveletexpansion1})
141: to a finite observation window, and measuring at finite
142: resolution, we obtain the approximation
143: \begin{equation}
144: \label{eq:waveletexpansion2} \left| f_J\right\rangle =\left\langle
145: \phi |f\right\rangle \left| \phi \right\rangle +\sum
146: ^{j=0}_{j=-J+1}\sum ^{n=2^{-j}-1}_{n=0}\left\langle \psi
147: _{jn}|f\right\rangle \left| \psi _{jn}\right\rangle ,
148: \end{equation}
149: where \( f \) is the restriction of \( F \) to \( [0,1] \).
150: The function \( \phi \) is called the scaling function, having
151: the properties of a smoothing kernel~\cite{mallat}. We are
152: interested in the fluctuations, so, without any loss, we employ in
153: our analysis the transform \(f\rightarrow f-\left\langle
154: f\right\rangle\), getting from \(f\) a version with null
155: time-average. At this point we are ready to express an energy and
156: an entropy for the time series, that are also resolution dependent
157: approximations of the energy density and entropy density for the
158: ideal stochastic process. The energy of \( f_J \) is defined by
159: the ensemble average of the \(L^2\)-norm squared,
160: \begin{equation}
161: \label{eq:energydensity} E\left(J\right)
162: \equiv\overline{\left\langle f_{J}|f_{J}\right\rangle },
163: \end{equation}
164: that is directly determined with the simple formula \(
165: E\left(J\right) =\overline{\sum ^{i=N}_{i=1}f^{2}_{i}}, \) where
166: \( f_{i} \) are the sampled values of \( f \). The entropy has the
167: definition
168: \begin{equation}
169: \label{eq:entropydensity} S\left(J\right) \equiv -\sum
170: ^{j=0}_{j=-J+1}\sum ^{2^{-j}-1}_{n=0}P_{jn}\log
171: _{2}P_{jn}-P_{\phi}\log_2P_{\phi},
172: \end{equation}
173: with \(P_{jn}\equiv \overline{\ |D_{jn}F|^2}/E\left(J\right)\),
174: and \(P_{\phi}\equiv \overline{\ |\left\langle
175: \phi|f\right\rangle|^2}/E\left(J\right)\). This expression has
176: similarities with the basis entropy cost, using the language of
177: reference~\cite{mallat}. In the limit of a single pulse localized
178: around \(t\) and with duration \(\delta t\sim k\tau\), with \(k\)
179: a natural number, only very few wavelet coefficients around the
180: scale \(k\tau\) and localized near \(t\) will be appreciably
181: different from zero. This characterizes low entropy processes. On
182: the other hand, the white noises~(WN) in which the energy is, in
183: average, equally distributed on all wavelets (the same power
184: dissipation at all times and scales) have the maximum entropy
185: \(S_{\rm{max}}=J\). This entropy functional is useful to quantify
186: the balance between order and disorder (complexity) for different
187: regimes of evolution of complex systems. Actually, we expect that
188: the level of disorder in a process mainly depends on the energy
189: distributed into the details. Consider, for example, the
190: interaction between two surfaces that slip one over the other.
191: When there is no friction we expect no sound. But if we introduce
192: roughness, the energy will be dissipated in microscopic
193: collisions, originating a sound that is classified as a noise, due
194: to the excess of details. Indeed, the intervals between such
195: collisions are so small that no measurement will capture them. The
196: process will continue a noise for \lq high\rq\ resolution
197: instruments, like it is for our \lq low\rq\ resolution ears. In
198: FIG.~\ref{fig:coefstable} we see the wavelet decomposition of the
199: sound caused by a piece of plastic pressed and pushed on a rough
200: table. The calculated entropy was \(S=12.52\) for a maximum of
201: \(S_{\rm{max}}=15\). The random small scale details have
202: amplitudes comparable to that of the function during the whole
203: observation, constituting important components. The great amount
204: of energy distributed on the multitude of details causes this high
205: entropy. This is why we suggest that the proposed entropy is
206: physically appropriate.
207:
208: A paradigmatic model for the self-affine processes is the
209: fractional Brownian motion (fBm), introduced by Mandelbrot and Van
210: Ness in the realm of stochastic processes with \( 1/f^{\alpha } \)
211: spectra~\cite{mandelbrot}. The fBm is a mono-parametric stochastic
212: process \( B_{H} \) with null ensemble average, Gaussian
213: increments, and whose correlation function has the general form:
214: \begin{equation}
215: \label{eq:correlation} \overline{B_{H}\left( t\right) B_{H}\left(
216: s\right) }=\frac{\sigma ^{2}_{H}}{2}\left( \left| t\right|
217: ^{2H}+\left| s\right| ^{2H}-\left| t-s\right| ^{2H}\right) ,
218: \end{equation}
219: with \(\sigma ^{2}_{H}\equiv \overline{B^{2}_{H}\left( 1\right)
220: }\). The parameter of the process is the Hurst exponent \( H
221: \in(0,1)\), which controls the nature of the correlations. The
222: variance of \( B_{H} \) depends on time according to
223: \(\overline{B^{2}_{H}}=\sigma ^{2}_{H}\left| t\right| ^{2H},\)
224: that reduces to the Brownian motion form when \( H=1/2 \),
225: representing the uncorrelated case. In this process, for \( H<1/2
226: \) the increments present infinite range anti-correlations, while
227: for \( H>1/2 \) there are infinite range correlations. The Fourier
228: spectrum of fBm follows the power law
229: \begin{equation}
230: \label{eq:spectrum} S\left( f \right) \sim \frac{1}{f ^{2H+1}}.
231: \end{equation}
232:
233: \begin{figure}
234: {\par\centering \resizebox*{6cm}{6cm}{\includegraphics{fig2.eps}}
235: \par}
236: \caption{The solid line is the theoretical fBm entropy at
237: resolution \protect\( N=2048\protect \). We see that the SRA
238: algorithm gives a very good approximation to the theoretical
239: prediction. Two lines represent the effect of Gaussian white noise
240: superposition. The (visually) straight dotted line at the upper
241: part gives the entropy of simulated pure Gaussian white
242: noise.\label{fig:SversusH}}
243: \end{figure}
244:
245: In the Haar wavelet basis one finds exact and simple expressions
246: for the resolution dependent energy and entropy densities of fBm.
247: The Haar wavelets are periods of square waves written as
248: \begin{equation}
249: \label{eq:haarwavelets} h_{jn}\left( t\right) =\left\{
250: \begin{array}{cc}
251: \frac{1}{\sqrt{2^{j}}}, & 2^{j}n\leq t<2^{j}\left( n+1/2\right) \\
252: -\frac{1}{\sqrt{2^{j}}}, & 2^{j}\left( n+1/2\right) \leq
253: t<2^{j}\left( n+1\right)
254: \end{array}\right. .
255: \end{equation}
256: For Haar wavelets one proves, using the expression~(\ref{eq:correlation}), that
257: \begin{equation}
258: \label{eq:expectedvalueofWTsquared} \overline{\left| \Delta
259: _{jn}b_{H,J}\right| ^{2}}=\frac{\sigma ^{2}_{H}\left(
260: 1-2^{-2H}\right) }{\left( 2H+1\right) \left( 2H+2\right)
261: }2^{\left( 2H+1\right) j},
262: \end{equation}
263: where \( \Delta _{jn}f\equiv \left\langle h_{jn}|f\right\rangle \), and \(b_H\) indicates restriction of
264: \(B_H\) to \([0,1]\). This expression gives the scale dependent
265: power spectrum in the Haar basis, from which one can calculate the
266: energy density of \( B_{H} \)
267: \begin{equation}
268: \label{eq:energydensity} E\left( H,J\right)
269: =\overline{\left\langle b_{H,J}|b_{H,J}\right\rangle
270: }=\frac{\sigma ^{2}_{H}\left( 1-2^{-2HJ }\right) }{\left(
271: 2H+1\right) \left( 2H+2\right) },
272: \end{equation}
273: depending on \(\sigma_H\), that must be determined experimentally
274: . The entropy density has the final form
275: \begin{widetext}
276: \begin{equation}
277: \label{eq:shannonresult} S\left(H,J\right )=2^{-2H}\left(
278: 2H+1\right) \frac{1-2^{-2H(J-1)}J +2^{-2HJ}(J-1)}{\left(
279: 1-2^{-2H}\right) \left[ 1-2^{-2HJ }\right] }+\log _{2}C\left(
280: H,J\right) ,
281: \end{equation}
282: \end{widetext}
283: where \(C\left(H,J\right)=\left[1-2^{-2HJ}\right]/
284: \left(1-2^{-2H}\right)\).
285:
286: This entropy density is universal and exact for the
287: fBm processes sampled without rounding error to a given finite
288: resolution. It is very useful since for a great number of
289: practical applications there are just low resolution data, what
290: makes mandatory that the theoretical predictions take resolution
291: into account explicitly. In the limit \(J\rightarrow\infty\) the
292: entropy diverges at \(H=0\) as \(S(H,\infty)\sim 1/2H\ln 2\). For
293: other values of \(H\) the entropy is finite at infinite
294: resolution. At finite resolution \(S(0,J)=(J-1)/2+\log_2(J)\).
295:
296: The FIG.~\ref{fig:SversusH} shows a plot of fBm theoretical
297: entropy for the resolution \( N=2048 \). The theoretical result is
298: compared with \( S\left( H,J\right) \) calculated from samples
299: simulated with the Successive Random Additions (SRA)
300: algorithm~\cite{peitgen}. It shows also how Gaussian white
301: noise~(WN) affects the entropy, clarifying what is to be expected
302: from noisy data. For comparison, the dots at the upper part of the
303: figure show the entropy for the simulated pure Gaussian white
304: noise, whose average value is \(S=10.98\sim 11\). In the dotted
305: line we see the effect of a superposed noise with 10\% of the
306: amplitude of the fBm. The dash-dotted line shows this effect when
307: the percentage grows to 50\%. The entropies for all simulated
308: processes were calculated from 100 samples. The drastic diminution
309: of entropy when \(H\) increases happens because the details
310: (roughness) become weaker. There is very good agreement between
311: theoretical prediction and simulation.
312:
313: In conclusion, we propose that the entropy
314: measure~(\ref{eq:entropydensity}) is suitable to estimate the
315: complexity of time series by revealing the degree in which
316: \emph{detail modes} are excited in a process. These detail modes
317: are easily represented as localized oscillations or wavelets,
318: which in turn are very effective mathematical instruments for
319: time-scale analysis. Considering such phenomena as fluid or scalar
320: turbulence~\cite{zsshe,shraiman}, that are characterized by
321: fluctuating cascades from high inertial scales to low dissipative
322: scales, or plasma turbulence~\cite{krommes}, in which some scales
323: are most important for the energy transfer, the proposed entropy
324: seems a natural measure of the intrinsic order-disorder balance.
325: In such phenomena, many structures (vortices, convective cells)
326: appear at intermediate scales, that should be detected as
327: excitations of intermediate scale modes, lowering the entropy, as
328: compared with noises that are completely determined by microscopic
329: dissipation (whose characteristics are indicated in
330: FIG.~\ref{fig:coefstable}). For example, if a theory with the same
331: correlations of Kolmogorov K41 theory~\cite{kolmogorov41} is
332: valid, implying \(H=1/3\), the time series of velocity increments
333: would lead to the entropy \(S(1/3,\infty)=4.27\dots\), a number
334: significantly small, compared to the case in
335: FIG.~\ref{fig:coefstable}. What we learn from the fBm analysis
336: presented here suggests that the measurement of the entropy is a
337: strong method of characterization. Furthermore, it is interesting
338: to obtain experimental curves of entropy for stochastic processes
339: depending on parameters, or in situations where the entropy may
340: change with time. This method has given results (which will be
341: published soon) in three situations: In the analysis of the sound
342: of lungs, in the analysis of heart beat series, and in the
343: analysis of global positioning system errors caused by equatorial
344: spread F in ionospheric plasma. We observe that such entropy
345: measurements could give important information on the
346: non-equilibrium approach to self-organized regimes~\cite{corral},
347: as well as, on self-organized regimes made unstable, for example,
348: by diffusion~\cite{newman}.
349:
350: We acknowledge the CAPES, CNPq, FINEP and CTPETRO for financial
351: support.
352:
353: \begin{thebibliography}{99}
354: \bibitem{consolini}G. Consolini, T. Chang,
355: \emph{Complexity, magnetic field topology, criticality, and
356: metastability in magnetotail dynamics}: Journal of Atmospheric and
357: Solar-Terrestrial Physics \textbf{64}, 541-549 (2002).
358: \bibitem{politzer}P.A. Politzer,
359: \emph{Observation of Avalanchelike Phenomena in a Magnetically
360: Confined Plasma}: Phys. Rev. Lett. \textbf{84}, 1192-1195 (2000).
361: \bibitem{santi}L. Santi, J.C. Denardin, M.R. Dotto, L.F. Schelp,
362: and R. Sommer, \emph{Barkhousen noise measurements in materials
363: with vanishing magnetoelastic anisotropies}: J. Applied Phys.
364: \textbf{91}, 8201-8203 (2002).
365: \bibitem{sreenivasan}K.R. Sreenivasan, \emph{On Local Isotropy
366: of Passive Scalars in Turbulent Shear Flows}: Proc. Roy. Soc.
367: London A, \textbf{434}, Issue 1890, 165-182 (1991).
368: \bibitem{gafarov}R. Yulrnetyev, P. Hänggi, and F. Gafarov
369: \emph{Quantification of heart rate variability by discrete
370: nonstationary non-Markov stochastic processes}: Phys. Rev. E
371: \textbf{65}, 046107 (2002).
372: \bibitem{bak}P. Bak, C. Tang, K. Wiesenfeld, \emph{Self-Organized Criticality: An explanation of 1/f Noise}: Phys. Rev. Lett. \textbf{59},
373: 381-384 (1987).
374: \bibitem{durin}G. Durin, and S. Zapperi,
375: \emph{Scaling Exponents for Barkhousen Avalanches in
376: Polycrystalline and Amorphous Ferromagnets}: Phys. Rev. Lett.
377: \textbf{84}, 4705-4708 (2002).
378: \bibitem{urbach}J.S. Urbach, R.C. Madison, and J.T. Markert,
379: \emph{Interface Depinning, Self-Organized Criticality, and the
380: Barkhousen Effect}: Phys. Rev. Lett. \textbf{75}, 276-279 (1995).
381: \bibitem{cote}P.J. Cote, and L.V. Meisel,
382: \emph{Self-Organized Criticality and the Barkhausen Effect}: Phys.
383: Rev. Lett. \textbf{67}, 1334-1337 (1991).
384: \bibitem{carreras}B.A. Carreras, B. van Milligen, M.A. Pedrosa,
385: R. Balbín, C. Hidalgo, D.E. Newman, E. Sánchez, M. Frances, I.
386: García-Cortés, J. Bleuel, M. Endler, S. Davies, and G.F. Matthews,
387: \emph{Long-Range Time Correlations in Plasma Edge Turbulence}:
388: Phys. Rev. Lett. \textbf{80}, 4438-4441 (1998).
389: \bibitem{muzy}J.F. Muzy, E. Bacry, and A. Arnéodo, \emph{The multifractal formalism revisited with wavelets}: Int. J. of Bifurcation and Chaos \textbf{4},
390: 245-302 (1994).
391: \bibitem{mallat}S. Mallat, \emph{A Wavelet Tour of
392: Signal Processing} 2nd. ed. (Academic Press, Boston, 1998).
393: \bibitem{bandt}C. Bandt and B. Pompe,
394: \emph{Permutation Entropy: A Natural Complexity Measure for Time
395: Series}: Phys. Rev. Lett. \textbf{88}, 174102 (2002).
396: \bibitem{eckmann}J.P. Eckmann and D. Ruelle,
397: \emph{Ergodic theory of chaos and strange attractors}: Rev. Mod.
398: Phys. \textbf{57}, 617-656 (1985).
399: \bibitem{madalena}M. Costa, A.L. Goldberger, and C.K. Peng,
400: \emph{Multiscale Entropy Analysis of Complex Physiologic Time
401: Series}: Phys. Rev. Lett. \textbf{89}, 068102-1 (2002).
402: \bibitem{unser}M. Unser,
403: \emph{Sampling - 50 Years After Shannon}: Proc. of the IEEE
404: \textbf{88}, 569-587 (2000).
405: \bibitem{mandelbrot}B.B. Mandelbrot and J.W. Van Ness, \emph{Fractal Brownian motions, fractional noises and applications}: SIAM Rev. 10, 422-437 (1968).
406: \bibitem{peitgen}\emph{The Science of Fractal Images}, edited by H.O. Peitgen and D. Saupe (Springer-Verlag, New York, 1987).
407: \bibitem{zsshe}Z.S. She, E. Jackson, and S.A. Orszag,
408: \emph{Structure and Dynamics of Homogeneous Turbulence: Models and
409: Simulations}: Proc. Roy. Soc. London A \textbf{434}, Issue 1890,
410: 101-124 (1991).
411: \bibitem{shraiman}B.L. Shraiman and E.D. Siggia,
412: \emph{Scalar turbulence}: Nature
413: \textbf{405}, 639-646 (2000).
414: \bibitem{krommes}J.A. Krommes,
415: \emph{Fundamental statistical descriptions of plasma turbulence in
416: magnetic fields}: Phys. Rep. \textbf{360}, 1-352 (2002).
417: \bibitem{kolmogorov41}A.N. Kolmogorov, \emph{The Local Structure of Turbulence in Incompressible Viscous Fluid for Very Large Reynolds Numbers}: C.R. Acad. Sci. USSR \textbf{30}, 301-305 (1941).
418: \bibitem{corral}A. Corral, and M. Paczuski,
419: \emph{Avalanche Merging and Continuous Flow in a Sandpile Model}:
420: Phys. Rev. Lett. \textbf{83}, 572-575 (1999).
421: \bibitem{newman}D.E. Newman, R. Sánchez, B.A. Carreras, and W. Farenbaugh,
422: \emph{Transition in the Dynamics of a Diffusive Running Sandpile}:
423: Phys. Rev. Lett. \textbf{88}, 204304-1 (2002).
424: \end{thebibliography}
425: \end{document}
426: