cs0412027/v5.tex
1: %\documentclass[aps,preprint,showpacs,superscriptaddress]{revtex4}
2: 
3: 
4: \documentclass[aps,prl,twocolumn,showpacs,superscriptaddress]{revtex4}
5: \usepackage{graphicx}
6: \usepackage{amssymb}
7: \newcommand{\maya}[1]{{\huge{$\bullet$}}{\em #1}}
8: %%%%
9: \newcommand{\beq}{\begin{equation}}
10: \newcommand{\eeq}{\end{equation}}
11: %%%%
12: 
13: \begin{document}
14: 
15: \title{Correlated dynamics in human  printing behavior}
16: 
17: \author{Uli Harder}
18: \email{uh@doc.ic.ac.uk}
19: \affiliation{Department of Computing,  Imperial College London, London
20:   UK SW7 2AZ.} 
21: \author{Maya Paczuski} 
22: \email{maya@ic.ac.uk}
23: \affiliation{John-von-Neumann
24: Institute for Computing, Forschungszentrum J\"{u}lich, D-52425 J\"{u}lich,
25: Germany}
26: \affiliation{Department of Mathematics,  
27: Imperial College London, London UK SW7 2AZ.}
28: 
29: \begin{abstract}
30:   {Arrival times of requests to print in a student laboratory were
31: analyzed.  Inter-arrival times between subsequent requests follow a
32: universal scaling law relating time intervals and the size of the
33: request, indicating a scale invariant dynamics with respect to the size.
34: The cumulative distribution of file sizes is well-described by 
35: a modified power law often seen in non-equilibrium critical systems.
36: For each user, waiting 
37: times between their individual requests show long range dependence and
38: are broadly distributed from seconds to weeks. All results are incompatible
39: with Poisson models, and may provide
40: evidence of critical dynamics associated with
41: voluntary thought processes in the brain.
42: }
43: 
44: \end{abstract}
45: 
46: \pacs{89.75.-k, 89.75.Da, 02.50.E, 05.40.-a}
47: 
48: \maketitle
49: 
50: 
51: %%%%%%
52: 
53: Since the early work of Berger and Mandelbrot~\cite{berger} examining
54: error clustering in telephone circuits, it has been recognized that
55: standard Poisson models may be inadequate to describe electronic
56: information networks.  This was confirmed, for instance by Leland et
57: al.~\cite{LTWW}, who studied network traffic and found that packet
58: traces show scaling behavior.  
59: Observations of scaling behavior raise a number of
60: questions about how to model these systems, optimize performance, or
61: improve design. Significant effects include  an increase in response
62: times, required buffer sizes, etc. In Ref.~\cite{AESOP:mascots-2002-pe} 
63: the authors show how the file size distribution of a web server effects the
64: resulting network traffic.
65: Large fluctuations (which are inherent in critical systems)
66:  in packet traffic or
67: demand for resources in computer networks can significantly degrade
68: worst case performance~\cite{huberman}.
69: Scaling behavior has  been found not only in the size
70: distribution of files stored in computer systems~\cite{UFSD}, and the sizes
71:  of web server requests~\cite{C}, but also in the physical
72: structure of the internet~\cite{faloutsos} and the hyper-link
73: structure of the world-wide web
74: \cite{albert,bornholdt02:_book}.  So far, no definitive causes
75: have been established for the complexity of the modern information network.
76: Of course, humans interact when they build the internet, make hyper-link
77: connections, and send and receive information.  Like
78: traffic jams \cite{nagel_pac} on roads, internet jams are produced by
79: humans who act and react, often in response to
80: information originating within the network or outside it.  Various
81: parts of the information network/user system are themselves
82: complex systems, and one of the problems in modeling modern
83: information networks  is how to disentangle these effects.
84: 
85: One recognizes that psychological experiments have demonstrated that correlated
86: dynamics occurs in  individual human behavior
87: \cite{KMT93,CC94,DA96,gottschalk,paulus}, even in situations where
88: interactions with other humans are minimal~\cite{CDK,GTM,OHT03}. For
89: instance, Ref.~\cite{GTM} describes an experiment where
90: subjects had to estimate the duration of time intervals from
91: memory. The time series of errors in the estimates exhibits a $1/f$
92: power spectrum, showing that the errors are correlated in time. In
93: contrast, the sequence of reaction times to an event showed no long
94: range correlation. The authors proposed that  long range
95: dependence is associated with voluntary thought processes in the
96: brain~\cite{GTM}.  Similar observations were made for the dynamics of moods
97: \cite{gottschalk} and psychotic states.  For instance the
98: distribution of time intervals between subsequent hospitalizations for
99: schizophrenia is approximately power-law~\cite{DA96}. 
100:  A physical basis for these
101: behaviors may be related to scale-free functional networks in the brain, which
102: have recently been observed in situ~\cite{dante2}.
103: 
104: 
105: 
106: In order to better describe individual human behavior in a networked
107: computing environment, we study a simple case where the use
108: or demand is primarily associated with individual choice rather than with
109: group dynamics.  The particular quantity we focus on is the
110: inter-arrival times between subsequent print requests made by users in a computing
111: laboratory for university students.  We find evidence of long range
112: correlations in the inter-arrival times for individual users to send
113: requests, as well as a broad distribution of inter-arrival times. The totality
114: of print requests from all users reveals a scaling law relating
115: inter-arrival times and the sizes of the print request. This law indicates
116: that the same (re-scaled)
117:  dynamics is responsible for requests to print small and large
118: documents. This law is
119:  similar to that
120: recently observed for waiting times between successive
121: earthquakes~\cite{bak02:_unified,alvaro,davidsen_goltz} or solar
122: flares~\cite{bps}.  The
123: scaling function for the re-scaled inter-arrival times is approximately
124: log-normal.  The cumulative distribution of the sizes
125: of print requests is well-described by a modified power law, which is referred
126: to  as the $\chi^2$ distribution of
127: superstatistics~\cite{beck_cohen,beck}, or the q-exponential of non extensive
128: statistical mechanics
129: \cite{tsallis,wilk}.  An elementary stochastic
130: process is studied that reproduces some, but not all, of the observed
131: features. Our results are supportive of the hypothesis that the brain
132: operates at or near a self-organized critical state \cite{bak_brain}.
133: It also suggests the possibility of using data collected via  the modern information network
134: to systematically investigate models of human behavior. 
135: 
136: 
137: 
138: 
139: The Department of Computing at Imperial College London maintains a
140: networked printing system for staff and students. The student labs
141: offer about 300 computer work spaces, and are divided into different
142: rooms, the largest one accommodating up to 150 students. The printers
143: are networked and accessible from any machine in the department.  A
144: user selects a printer and submits her print job to a central
145: server. The server records the time a request is submitted with a resolution
146: of one second.  It also records the size of the request,
147: the user name and the intended printer. This investigation focuses on
148: requests sent to the printer, {\em chrome}, that is located in the
149: largest room. The labs are closed between 23:00 and 7:00, but users can
150: print during closure times when logged in remotely.  The data used here include the entire
151: year of 2003 and closure
152: times have been included in the analysis.  Table 1 gives relevant
153: parameters for the data set studied, which can be accessed
154: at~\cite{data}.
155: \begin{table}[htbp]
156:   \centering
157:   \begin{tabular}{|l|l|}
158:    \hline
159:     number of users & 1122\\ \hline
160:     number of users issuing & \\ 
161:     more than three requests & 1001\\ \hline
162:     number of requests per year & 73853\\ \hline
163:     mean document size & 1.2 Mbytes\\
164:     mean time between requests & 7.1 min\\ \hline
165:     minimum time resolution & 1.0 sec\\ \hline
166:   \end{tabular}
167:   \caption{Parameters of the user and printing system in 2003.}
168:   \label{tab:key}
169: \end{table}
170: 
171: %525600
172: 
173: %\section{Analysis}
174: 
175: We first analyze the distribution of  inter-arrival times
176:  between subsequent print
177: requests for the entire year. Time
178: differences from the logged event times $T^S_i$ are measured as
179: \begin{equation}
180:   \label{eq:diff}
181:   t_i^S  = T^S_{i+1} - T^S_i, \text{where } 0 \leq i
182:   \leq N^S \quad .
183: \end{equation}
184: The superscript $S$ refers to the size of the print request in bytes
185: and indicates that this set of times only includes requests that
186: are larger than $S$. The quantity $N^S$ is the number of print requests that are
187: larger
188: than $S$. Time intervals of length zero are
189: neglected from the analysis. For each chosen threshold $S$ we estimate
190:  $P_S(t)$, which is the probability
191: of a certain time interval $t$ between subsequent requests
192: of size $S$ or larger.  To display this distribution we count the
193: number of time differences in exponentially growing bins and normalize
194: the count by the bin size. Fig. \ref{fig:unscaled-iat} shows that the 
195: shape of the waiting time distribution depends on the size threshold,
196: $S$, of the documents. This could indicate  different dynamical processes
197: responsible for the small and large documents.  However, all distributions are  broad and show an
198: anomaly near one day. The anomaly is related to the overnight
199: closure of the  labs.
200: \begin{figure}[htb]
201:     \centering
202:      \includegraphics[width=6cm,angle=270]{figures/unscaled-iat.ps}
203:      \caption{Distribution of inter-arrival between subsequent requests to
204:        the printer "chrome" in 2003.  Different curves are for
205:        different threshold sizes of the requests. }
206:      \label{fig:unscaled-iat}
207:  \end{figure}
208: 
209: To determine if a different dynamics is responsible for requests of different
210: sizes, we implement
211:  a scaling argument similar to one recently put forward by Bak {\it et
212:    al}~\cite{bak02:_unified} to describe the waiting time statistics
213:    of earthquakes.     The
214:    average time between requests $\langle t \rangle_S$ may provide a 
215:     rescaling factor for the inter-arrival times, so that
216:    the distributions measured with different
217:    size thresholds, $S$, collapse onto a single scaling function.
218:    Of course, $\langle t\rangle
219:    _S = \frac{T}{N_{>S}}= \frac{1}{R(S)}$.  Here $T$ is the time span
220:    of the record and $R(S)$ is the rate of requests larger than $S$.  $N(>S)$ is the cumulative number of requests
221:    larger than size $S$. As
222:    shown in the inset of Fig.  \ref{fig:scaled-iat},  $N(>S)$ is
223: well described by a modified power law~\cite{beck_cohen,beck,tsallis,wilk}:
224: \begin{equation}
225:   \label{eq:S}
226:   N(>S) \sim \frac{1}{(1+(S/S^*))^{\gamma -1}}
227: \end{equation}
228:   where
229:     $S^* = (7.9 \pm 0.5) \times 10^{5}$ and
230:    $\gamma-1=0.76 \pm 0.03$.  
231: 
232: \begin{figure}[htb]
233:     \centering
234:      \includegraphics[width=6cm,angle=270]{figures/scaled-iat-with-rates.ps}
235:      \caption{Universal scaling law for the inter-arrival
236: times between requests larger than size $S$, according to Eq. \ref{eq:ansatz}.
237: The solid line is a fit of to a log-normal function as described in the
238: text. Data from the numerical simulation is also shown. The inset displays the cumulative distribution of requests sizes.}
239:      \label{fig:scaled-iat}
240:  \end{figure}
241: 
242:  We test
243:   the {\it ansatz} 
244: \begin{equation}
245:   \label{eq:ansatz}
246:  P_S(t)\sim R(S)g(t R(S)) \quad ,  
247: \end{equation}
248:  where $g(x)$ is a scaling function and $x=tR(S)$ is a scaling variable.
249: Fig.~\ref{fig:scaled-iat} shows the results of rescaling the different
250:  curves in Fig.~\ref{fig:unscaled-iat} by their average rate.  We see that the scaling
251: {\it ansatz} of Eq. \ref{eq:ansatz} appears to hold over a wide range, about seven orders of
252: magnitude in the scaling variable. This indicates that the same scale invariant
253: dynamics operates when users send requests of any size.
254:  The slight
255:  deviation from data collapse at short times is due to the finite
256:  temporal resolution of our data (one second). There is an additional
257:  deviation due to the  diurnal period. 
258: The scaling function $g$ is close to  a log-normal distribution: 
259: \begin{equation}
260:   \label{eq:lognormal}
261:   g(x) = \frac{ 1 }{ \sqrt{2 \pi} \sigma x   }  
262: \exp(-\frac{ (\ln(x) - m)^2}{2 \sigma^2 } ) 
263: \end{equation}
264: with $m= -3.41 \pm 0.07$ and $\sigma = 2.16 \pm 0.04$, as also shown
265:  in Fig.~\ref{fig:scaled-iat}. This feature is also found in numerical simulations
266: of a stochastic process described later.
267: 
268: 
269: 
270: 
271: 
272: The inter-arrival times for all users do not
273:  necessarily give a good estimate for the times that pass between subsequent
274: requests issued by a {\em single} user. To this end we study
275: the inter-arrival times $t_i^u$ for each user $u$ printing more than three
276: documents over the one year period. In the discussion below we set the threshold
277: $S=0$.
278: \begin{equation}
279:   \label{eq:user}
280:   t_i^u  = T^u_{i+1} - T^u_i, \text{where } 0 \leq i
281:   \leq N^u .
282: \end{equation}
283: Each user's list of inter-arrival times is concatenated to determine
284: the probability  $P_{\rm ind}(t)$ of single user inter-arrival times,
285: shown in Fig.~\ref{fig:single-iat}.  This distribution is
286: approximately a power law over several decades ranging from one minute
287: to about a day, with an exponent $\alpha
288: \approx 1.3$. We also analyze the inter-arrival times for the
289: busiest single user, which is similar.  For comparison we
290: show in Fig.~\ref{fig:single-iat} an exponential distribution for a
291: Poisson event process that has the same average rate,
292: $\lambda=3.4\times 10^{-5}/{\rm sec}$, as the process of the busiest
293: single user.  A critical system with a  power-law distribution of intervals
294:  is a more
295: accurate description of the data than a Poisson model of print requests.
296: 
297: 
298: \begin{figure}[htb]
299:     \centering
300:      \includegraphics[width=6cm,angle=270]{figures/single-iat.ps}
301:      \caption{Single user inter-arrival time distribution, averaged over all
302: users and for the single busiest user. The solid, straight line indicates a power law
303: distribution, $P_{\rm ind}(t) \sim t^{-\alpha}$ with $\alpha=1.3$. For comparison,
304: an exponential distribution with the same rate as the busiest user is 
305: shown as a dashed curve.}
306:      \label{fig:single-iat}
307:  \end{figure}
308: 
309: To decide if inter-arrival times  are correlated, we measured
310: the auto correlation function of waiting times
311: for  single users.
312: The autocorrelation $a_u(\tau)$ at lag step $\tau$ is defined as
313: \begin{equation}
314:   \label{eq:auto}
315:   a_u(\tau) = \frac{1}{N_u - \tau} \sum_{i=1}^{N_u - \tau} s^u_i s^u_{i+\tau}
316: \end{equation}
317: where $s^u_i = t_i^u - \frac{1}{N_u} \sum_{j=1}^{N_u} t_j^u$.  If the
318: inter-arrival times are uncorrelated and independent, the arrival process of individual
319: requests to print can be modeled as a fractal renewal process
320: \cite{pmb,LT193}.  Analyzing data separately for the three most busy users, we
321: find that the auto correlation function decays as $1/\tau^{\delta}$ with
322: $\delta \approx 0.6$. When the order of the inter-arrival times
323: for an individual user are shuffled randomly this power law
324: disappears, and the waiting times become uncorrelated, with $a_u(\tau)$ independent
325: of $\tau$ for $\tau \geq 1$.   The sequence of inter-arrival times
326:  for individual users are
327: correlated over the entire time span of our data set.
328: 
329: Our data shows that models of criticality are relevant for describing individual
330: human
331: behavior in the modern information network.  Lacking, at present, a microscopic
332: dynamical model, we compare our observations with results from a 
333: simple stochastic process. Consider $N$ arrival streams of print
334: requests. In each stream,  time intervals between subsequent
335: requests are independent random variables chosen from a truncated
336: Pareto distribution.  We neglect correlations between
337: intervals.  All intervals have
338: the same probability distribution
339: \begin{equation}
340:   \label{eq:pareto}
341:   P_{\rm ind}(x) = \frac{1}{C} k x^{-1-k} \text{ where } 1 \leq a \leq x \leq b
342: \end{equation}
343: where $a$ and $b$ are the points where the Pareto distribution is
344: truncated and $C$ is a normalization constant.  We choose the
345: parameter $k=0.3$ motivated by the  results in Fig.
346: \ref{fig:single-iat}. The short time cut-off $a=2.5$ sec is set to
347: reflect the fact that in some application users must wait
348:  before a subsequent print job can be sent off.  Most students
349:  leave after  at most 8 years, so $b = 8$ years appears to
350: be a reasonable choice. Generating approximately 73,000 requests in
351: a year fixes the number of users close to $N = 1000$.
352:  
353: At the start of the numerical
354:  simulation we schedule an arrival event for each stream
355: according to  Eq. \ref{eq:pareto}. Upon each arrival, the next arrival time
356: is scheduled using the same distribution. The system
357: takes  about 5 years with the above parameters to reach a statistically
358: stationary state.
359:  As shown in Fig. 2, the inter arrival times measured in the simulation
360: compare fairly well with the real data. 
361: However the real data has significantly larger variance.
362:  
363:  We also examined the  time series defined by the number of print requests
364:  in each second.  We calculated the power spectrum $S(f)$ of this time series
365:  and find $1/f^{\alpha}$ behavior,
366:  as shown in Fig. 4.
367: The exponent $\alpha$ observed in the numerical simulation is fixed by the
368: value of $k$
369:  in Eq. \ref{eq:pareto}, and is $\alpha =
370:  0.3$ \cite{PMB96,LT193}. The real data 
371: show instead a larger value $\alpha \approx 0.5$,
372:   which indicates, just as
373: the auto-correlation function $a_u(\tau)$, that the real
374: arrival process is more complicated than a fractal renewal process. A more
375: accurate model of individual user behavior in a computing network may be that of Davidsen and Schuster~\cite{joern1}.
376: \begin{figure}[htb]
377:     \centering
378:      \includegraphics[width=6cm,angle=270]{figures/power.ps}
379:      \caption{Power spectrum of the time series defined by the print
380:      requests per second  
381:      based on the 
382:      real arrival data and the 
383:      simulated arrivals in the fifth 
384:      year. The solid line is a fit for the real data,
385:      the dashed one for the simulation results, see text.} 
386:      \label{fig:sim-power}
387:  \end{figure}
388: 
389: 
390: The authors  thank the Computer Support Group of the
391: Department of Computing at Imperial College, especially T.
392: Southerwood, for making the data available to us.
393: M.P. thanks J. Davidsen, A.L. Stella, and P. Grassberger for  conversations.
394: U.H. would like to thank A. Argent-Katwala, T. Field and W.
395: Knottenbelt for interesting discussions and suggestions. U.H. is
396: funded by  EPSRC (research grant PASTRAMI, GR/S24961/01)
397: %\bibliographystyle{apsrev}
398: %\bibliography{queue_prl}
399: 
400: \begin{thebibliography}{33}
401: \expandafter\ifx\csname natexlab\endcsname\relax\def\natexlab#1{#1}\fi
402: \expandafter\ifx\csname bibnamefont\endcsname\relax
403:   \def\bibnamefont#1{#1}\fi
404: \expandafter\ifx\csname bibfnamefont\endcsname\relax
405:   \def\bibfnamefont#1{#1}\fi
406: \expandafter\ifx\csname citenamefont\endcsname\relax
407:   \def\citenamefont#1{#1}\fi
408: \expandafter\ifx\csname url\endcsname\relax
409:   \def\url#1{\texttt{#1}}\fi
410: \expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi
411: \providecommand{\bibinfo}[2]{#2}
412: \providecommand{\eprint}[2][]{\url{#2}}
413: 
414: \bibitem[{\citenamefont{Berger and Mandelbrot}(1963)}]{berger}
415: \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Berger}} \bibnamefont{and}
416:   \bibinfo{author}{\bibfnamefont{B.}~\bibnamefont{Mandelbrot}},
417:   \bibinfo{journal}{IBM J. Res. Dev.} \textbf{\bibinfo{volume}{7}},
418:   \bibinfo{pages}{224} (\bibinfo{year}{1963}).
419: 
420: \bibitem[{\citenamefont{E.~Leland et~al.}(1994)\citenamefont{E.~Leland, Taqqu,
421:   Willinger, and Wilson}}]{LTWW}
422: \bibinfo{author}{\bibfnamefont{W.}~\bibnamefont{E.~Leland}},
423:   \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Taqqu}},
424:   \bibinfo{author}{\bibfnamefont{W.}~\bibnamefont{Willinger}},
425:   \bibnamefont{and} \bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Wilson}},
426:   \bibinfo{journal}{IEEE/ACM Transactions on Networking}
427:   \textbf{\bibinfo{volume}{2}}, \bibinfo{pages}{1} (\bibinfo{year}{1994}).
428: 
429: \bibitem[{\citenamefont{Field et~al.}(2004)\citenamefont{Field, Harder, and
430:   Harrison}}]{AESOP:mascots-2002-pe}
431: \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Field}},
432:   \bibinfo{author}{\bibfnamefont{U.}~\bibnamefont{Harder}}, \bibnamefont{and}
433:   \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Harrison}},
434:   \bibinfo{journal}{Performance Evaluation} \textbf{\bibinfo{volume}{58}},
435:   \bibinfo{pages}{243} (\bibinfo{year}{2004}).
436: 
437: \bibitem[{\citenamefont{Huberman and Lukose}(1997)}]{huberman}
438: \bibinfo{author}{\bibfnamefont{B.~A.} \bibnamefont{Huberman}} \bibnamefont{and}
439:   \bibinfo{author}{\bibfnamefont{R.~M.} \bibnamefont{Lukose}},
440:   \bibinfo{journal}{Science} \textbf{\bibinfo{volume}{277}},
441:   \bibinfo{pages}{535} (\bibinfo{year}{1997}).
442: 
443: \bibitem[{\citenamefont{Irlam}()}]{UFSD}
444: \bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Irlam}},
445:   \bibinfo{note}{\url{http://www.base.com/gordoni/ufs93.html}}.
446: 
447: \bibitem[{\citenamefont{Crovella and Bestavros}(1997)}]{C}
448: \bibinfo{author}{\bibfnamefont{M.~E.} \bibnamefont{Crovella}} \bibnamefont{and}
449:   \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Bestavros}},
450:   \bibinfo{journal}{IEEE\slash ACM Transactions on Networking}
451:   \textbf{\bibinfo{volume}{5}}, \bibinfo{pages}{835} (\bibinfo{year}{1997}).
452: 
453: \bibitem[{\citenamefont{Faloutsos et~al.}(1999)\citenamefont{Faloutsos,
454:   Faloutsos, and Faloutsos}}]{faloutsos}
455: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Faloutsos}},
456:   \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Faloutsos}},
457:   \bibnamefont{and}
458:   \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Faloutsos}},
459:   \bibinfo{journal}{Proc.\ ACM SIGCOMM, Comput.\ Commun.\ Rev.}
460:   \textbf{\bibinfo{volume}{29}}, \bibinfo{pages}{251} (\bibinfo{year}{1999}).
461: 
462: \bibitem[{\citenamefont{Albert and Barabasi}(2002)}]{albert}
463: \bibinfo{author}{\bibfnamefont{R.}~\bibnamefont{Albert}} \bibnamefont{and}
464:   \bibinfo{author}{\bibfnamefont{A.-L.} \bibnamefont{Barabasi}},
465:   \bibinfo{journal}{Rev. Mod. Phys.} \textbf{\bibinfo{volume}{74}},
466:   \bibinfo{eid}{47} (\bibinfo{year}{2002}).
467: 
468: \bibitem[{\citenamefont{Bornholdt and Schuster}(2002)}]{bornholdt02:_book}
469: \bibinfo{editor}{\bibfnamefont{S.}~\bibnamefont{Bornholdt}} \bibnamefont{and}
470:   \bibinfo{editor}{\bibfnamefont{H.~G.} \bibnamefont{Schuster}}, eds.,
471:   \emph{\bibinfo{title}{Handbook of Graphs and Networks}}
472:   (\bibinfo{publisher}{Wiley-VCH}, \bibinfo{year}{2002}).
473: 
474: \bibitem[{\citenamefont{Nagel and Paczuski}(1995)}]{nagel_pac}
475: \bibinfo{author}{\bibfnamefont{K.}~\bibnamefont{Nagel}} \bibnamefont{and}
476:   \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Paczuski}},
477:   \bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{51}},
478:   \bibinfo{pages}{2909} (\bibinfo{year}{1995}).
479: 
480: \bibitem[{\citenamefont{Kniffki et~al.}(1993)\citenamefont{Kniffki, Mandel, and
481:   Tran-Gia}}]{KMT93}
482: \bibinfo{author}{\bibfnamefont{K.-D.} \bibnamefont{Kniffki}},
483:   \bibinfo{author}{\bibfnamefont{W.}~\bibnamefont{Mandel}}, \bibnamefont{and}
484:   \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Tran-Gia}},
485:   \bibinfo{journal}{Fractals} \textbf{\bibinfo{volume}{1:380-387}}
486:   (\bibinfo{year}{1993}).
487: 
488: \bibitem[{\citenamefont{Canessa and Calmetta}(1994)}]{CC94}
489: \bibinfo{author}{\bibfnamefont{E.}~\bibnamefont{Canessa}} \bibnamefont{and}
490:   \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Calmetta}},
491:   \bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{50}}
492:   (\bibinfo{year}{1994}).
493: 
494: \bibitem[{\citenamefont{D\"unki and Amb\"uhl}(1996)}]{DA96}
495: \bibinfo{author}{\bibfnamefont{R.~M.} \bibnamefont{D\"unki}} \bibnamefont{and}
496:   \bibinfo{author}{\bibfnamefont{B.}~\bibnamefont{Amb\"uhl}},
497:   \bibinfo{journal}{Physica A} \textbf{\bibinfo{volume}{230}},
498:   \bibinfo{pages}{544} (\bibinfo{year}{1996}).
499: 
500: \bibitem[{\citenamefont{Gottschalk et~al.}(1995)\citenamefont{Gottschalk,
501:   Bauer, and Whybrow}}]{gottschalk}
502: \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Gottschalk}},
503:   \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Bauer}}, \bibnamefont{and}
504:   \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Whybrow}},
505:   \bibinfo{journal}{Arch. Gen. Psychiatry} \textbf{\bibinfo{volume}{52}},
506:   \bibinfo{pages}{947} (\bibinfo{year}{1995}).
507: 
508: \bibitem[{\citenamefont{Paulus and Braff}(2003)}]{paulus}
509: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Paulus}} \bibnamefont{and}
510:   \bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Braff}},
511:   \bibinfo{journal}{Biol. Psychiatry} \textbf{\bibinfo{volume}{54}},
512:   \bibinfo{pages}{3} (\bibinfo{year}{2003}).
513: 
514: \bibitem[{\citenamefont{Chen et~al.}(1997)\citenamefont{Chen, Ding, and
515:   Kelso}}]{CDK}
516: \bibinfo{author}{\bibfnamefont{Y.}~\bibnamefont{Chen}},
517:   \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Ding}}, \bibnamefont{and}
518:   \bibinfo{author}{\bibfnamefont{J.~A.~S.} \bibnamefont{Kelso}},
519:   \bibinfo{journal}{Phys.\ Rev.\ Lett.} \textbf{\bibinfo{volume}{79}},
520:   \bibinfo{pages}{4501} (\bibinfo{year}{1997}).
521: 
522: \bibitem[{\citenamefont{Gilden et~al.}(1995)\citenamefont{Gilden, Thornton, and
523:   Mallon}}]{GTM}
524: \bibinfo{author}{\bibfnamefont{D.~L.} \bibnamefont{Gilden}},
525:   \bibinfo{author}{\bibfnamefont{T.}~\bibnamefont{Thornton}}, \bibnamefont{and}
526:   \bibinfo{author}{\bibfnamefont{M.~W.} \bibnamefont{Mallon}},
527:   \bibinfo{journal}{Science} \textbf{\bibinfo{volume}{267}},
528:   \bibinfo{pages}{1837} (\bibinfo{year}{1995}).
529: 
530: \bibitem[{\citenamefont{van Orden et~al.}(2003)\citenamefont{van Orden, Holden,
531:   and Turvey}}]{OHT03}
532: \bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{van Orden}},
533:   \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Holden}}, \bibnamefont{and}
534:   \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Turvey}},
535:   \bibinfo{journal}{Journal of Experimental Psychology - General}
536:   \textbf{\bibinfo{volume}{3}}, \bibinfo{pages}{331} (\bibinfo{year}{2003}).
537: 
538: \bibitem[{\citenamefont{Sporns et~al.}(2004)\citenamefont{Sporns, Chialvo,
539:   Kaiser, and Hilgetag}}]{dante2}
540: \bibinfo{author}{\bibfnamefont{O.}~\bibnamefont{Sporns}},
541:   \bibinfo{author}{\bibfnamefont{D.}~\bibnamefont{Chialvo}},
542:   \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Kaiser}}, \bibnamefont{and}
543:   \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Hilgetag}},
544:   \bibinfo{journal}{Trends Cogn. Sci.} \textbf{\bibinfo{volume}{8}},
545:   \bibinfo{pages}{418} (\bibinfo{year}{2004}).
546: 
547: \bibitem[{\citenamefont{Bak et~al.}(2002)\citenamefont{Bak, Christensen, Danon,
548:   and Scanlon}}]{bak02:_unified}
549: \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Bak}},
550:   \bibinfo{author}{\bibfnamefont{K.}~\bibnamefont{Christensen}},
551:   \bibinfo{author}{\bibfnamefont{L.}~\bibnamefont{Danon}}, \bibnamefont{and}
552:   \bibinfo{author}{\bibfnamefont{T.}~\bibnamefont{Scanlon}},
553:   \bibinfo{journal}{Phys.\ Rev.\ Lett.} \textbf{\bibinfo{volume}{88}},
554:   \bibinfo{pages}{178501} (\bibinfo{year}{2002}).
555: 
556: \bibitem[{\citenamefont{Corral}(2003)}]{alvaro}
557: \bibinfo{author}{\bibfnamefont{A.}~\bibnamefont{Corral}},
558:   \bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{68}},
559:   \bibinfo{pages}{035102(R)} (\bibinfo{year}{2003}).
560: 
561: \bibitem[{\citenamefont{Davidsen and Goltz}(2004)}]{davidsen_goltz}
562: \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Davidsen}} \bibnamefont{and}
563:   \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Goltz}},
564:   \bibinfo{journal}{Geophys. Res. Lett.} \textbf{\bibinfo{volume}{31}},
565:   \bibinfo{pages}{L21612} (\bibinfo{year}{2004}).
566: 
567: \bibitem[{\citenamefont{Baiesi et~al.}(2004)\citenamefont{Baiesi, Paczuski, and
568:   Stella}}]{bps}
569: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Baiesi}},
570:   \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Paczuski}}, \bibnamefont{and}
571:   \bibinfo{author}{\bibfnamefont{A.~L.} \bibnamefont{Stella}}
572:   (\bibinfo{year}{2004}), \bibinfo{note}{e-print cond-mat/0411342}.
573: 
574: \bibitem[{\citenamefont{Beck and Cohen}(2003)}]{beck_cohen}
575: \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Beck}} \bibnamefont{and}
576:   \bibinfo{author}{\bibfnamefont{E.}~\bibnamefont{Cohen}},
577:   \bibinfo{journal}{Physica \ A} \textbf{\bibinfo{volume}{322}},
578:   \bibinfo{pages}{267} (\bibinfo{year}{2003}).
579: 
580: \bibitem[{\citenamefont{Beck}(2001)}]{beck}
581: \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Beck}},
582:   \bibinfo{journal}{Phys.\ Rev.\ Lett.} \textbf{\bibinfo{volume}{87}},
583:   \bibinfo{pages}{180601} (\bibinfo{year}{2001}).
584: 
585: \bibitem[{\citenamefont{Tsallis}(1998)}]{tsallis}
586: \bibinfo{author}{\bibfnamefont{C.}~\bibnamefont{Tsallis}},
587:   \bibinfo{journal}{J.\ Stat.\ Phys.} \textbf{\bibinfo{volume}{52}},
588:   \bibinfo{pages}{479} (\bibinfo{year}{1998}).
589: 
590: \bibitem[{\citenamefont{Wilk and Wlodarczyk}(2000)}]{wilk}
591: \bibinfo{author}{\bibfnamefont{G.}~\bibnamefont{Wilk}} \bibnamefont{and}
592:   \bibinfo{author}{\bibfnamefont{Z.}~\bibnamefont{Wlodarczyk}},
593:   \bibinfo{journal}{Phys.\ Rev.\ Lett.} \textbf{\bibinfo{volume}{84}},
594:   \bibinfo{pages}{2770} (\bibinfo{year}{2000}).
595: 
596: \bibitem[{\citenamefont{Chialvo and Bak}(1999)}]{bak_brain}
597: \bibinfo{author}{\bibfnamefont{D.~R.} \bibnamefont{Chialvo}} \bibnamefont{and}
598:   \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Bak}},
599:   \bibinfo{journal}{Neuroscience} \textbf{\bibinfo{volume}{90}},
600:   \bibinfo{pages}{1137} (\bibinfo{year}{1999}).
601: 
602: \bibitem[{\citenamefont{Harder and Paczuski}(2003)}]{data}
603: \bibinfo{author}{\bibfnamefont{U.}~\bibnamefont{Harder}} \bibnamefont{and}
604:   \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Paczuski}},
605:   \bibinfo{journal}{\url{http://www.doc.ic.ac.uk/~uh/PASTRAMI/Printer/data}}
606:   (\bibinfo{year}{2003}).
607: 
608: \bibitem[{\citenamefont{Paczuski et~al.}(1994)\citenamefont{Paczuski, Maslov,
609:   and Bak}}]{pmb}
610: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Paczuski}},
611:   \bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Maslov}}, \bibnamefont{and}
612:   \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Bak}},
613:   \bibinfo{journal}{Europhys.\ Lett.} \textbf{\bibinfo{volume}{28}},
614:   \bibinfo{pages}{295} (\bibinfo{year}{1994}).
615: 
616: \bibitem[{\citenamefont{Lowen and Teich}(1993)}]{LT193}
617: \bibinfo{author}{\bibfnamefont{S.~B.} \bibnamefont{Lowen}} \bibnamefont{and}
618:   \bibinfo{author}{\bibfnamefont{M.~C.} \bibnamefont{Teich}},
619:   \bibinfo{journal}{IEEE Trans. Inform. Theory} \textbf{\bibinfo{volume}{39}},
620:   \bibinfo{pages}{1669} (\bibinfo{year}{1993}).
621: 
622: \bibitem[{\citenamefont{Paczuski et~al.}(1996)\citenamefont{Paczuski, Maslov,
623:   and Bak}}]{PMB96}
624: \bibinfo{author}{\bibfnamefont{M.}~\bibnamefont{Paczuski}},
625:   \bibinfo{author}{\bibfnamefont{S.}~\bibnamefont{Maslov}}, \bibnamefont{and}
626:   \bibinfo{author}{\bibfnamefont{P.}~\bibnamefont{Bak}},
627:   \bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{53}},
628:   \bibinfo{pages}{414} (\bibinfo{year}{1996}).
629: 
630: \bibitem[{\citenamefont{Davidsen and Schuster}(2002)}]{joern1}
631: \bibinfo{author}{\bibfnamefont{J.}~\bibnamefont{Davidsen}} \bibnamefont{and}
632:   \bibinfo{author}{\bibfnamefont{H.}~\bibnamefont{Schuster}},
633:   \bibinfo{journal}{Phys.\ Rev.\ E} \textbf{\bibinfo{volume}{65}},
634:   \bibinfo{pages}{026120} (\bibinfo{year}{2002}).
635: 
636: \end{thebibliography}
637: 
638: 
639: 
640: 
641: 
642: \end{document}
643: