physics0603023/vot.tex
1: \documentclass[rmp,twocolumn,showpacs,superscriptaddress]{revtex4}
2: 
3: \usepackage{dcolumn,graphicx,amsmath,amssymb,txfonts}
4: 
5: \newcommand\av[1]{\bar{#1}}
6: \newcommand\defn{\textit}
7: \newcommand{\Ord}{\mathop{\mathrm{O}}\nolimits}
8: 
9: \begin{document}
10: 
11: \title{Nonequilibrium phase transition in the coevolution of networks and
12: opinions}
13: 
14: \author{Petter Holme}
15: \affiliation{Department of Computer Science, University of New Mexico,
16:   Albuquerque, NM 87131, U.S.A.}
17: \affiliation{Department of Physics, University of Michigan,
18:   Ann Arbor, MI 48109, U.S.A.}
19: \author{M. E. J. Newman}
20: \affiliation{Department of Physics, University of Michigan,
21:   Ann Arbor, MI 48109, U.S.A.}
22: 
23: \begin{abstract}
24:   Models of the convergence of opinion in social systems have been the
25:   subject of a considerable amount of recent attention in the physics
26:   literature.  These models divide into two classes, those in which
27:   individuals form their beliefs based on the opinions of their neighbors
28:   in a social network of personal acquaintances, and those in which,
29:   conversely, network connections form between individuals of similar
30:   beliefs.  While both of these processes can give rise to realistic levels
31:   of agreement between acquaintances, practical experience suggests that
32:   opinion formation in the real world is not a result of one process or the
33:   other, but a combination of the two.  Here we present a simple model of
34:   this combination, with a single parameter controlling the balance of the
35:   two processes.  We find that the model undergoes a continuous phase
36:   transition as this parameter is varied, from a regime in which opinions
37:   are arbitrarily diverse to one in which most individuals hold the same
38:   opinion.  We characterize the static and dynamic properties of this
39:   transition.
40: \end{abstract}
41: 
42: \pacs{87.23.Ge, 64.60.Ak, 89.75.Fb, 89.75.Hc}
43: % 87.23.Ge -- Dynamics of social systems
44: % 64.60.Ak -- Renormalization-group, fractal, and percolation studies
45: %               of phase transitions
46: % 89.75.Fb -- Structures and organization in complex systems
47: % 89.75.Hc -- Networks and genealogical trees
48: 
49: \maketitle
50: 
51: \section{Introduction}
52: 
53: Simple mathematical models describing emergent phenomena in human
54: populations~\cite{schelling}, such as voter models and market models, have
55: a long history of study in the social sciences.  It is only relatively
56: recently, however, that physicists have noted the close conceptual and
57: mathematical connections between these models and traditional models in
58: statistical physics such as spin models.  Building on this observation,
59: there have been a number of important advances in the understanding of
60: these models in the last decade or so, most notably in the study of social
61: networks~\cite{AB02,DM02,Newman03d}.  While the physics community has been
62: concerned primarily with studies of network structure, there has also been
63: a substantial line of investigation focusing on dynamical processes on
64: networks.  One example, which has a long history in sociology but is also
65: well suited to study using physics methods, is the dynamics of opinion
66: formation.  This problem highlights one of the fundamental questions in
67: network dynamics, namely whether dynamics controls the structure of a
68: network or the structure controls the dynamics.
69: 
70: It is observed that real social networks tend to divide into groups or
71: communities of like-minded individuals.  An obvious question to ask is
72: whether individuals become like-minded because they are connected via the
73: network~\cite{liggett:vot,cas:vot,sood:vot,sznajd,deuff:opi,cc:axe}, or
74: whether they form network connections because they are
75: like-minded~\cite{mcp:bird}.  Both situations have been studied with
76: physics-style models, the first using opinion formation
77: models~\cite{liggett:vot,cas:vot,sood:vot} and the second using models of
78: ``assortative mixing'' or
79: ``homophily''~\cite{Soderberg02,Newman03c,BPDA04}.  Common sense, however,
80: tells us that the distinction between the two scenarios is not clear-cut.
81: Rather, the real world self-organizes by a combination of the two, the
82: network changing in response to opinion and opinion changing in response to
83: the network.  In this paper we study a simple model---perhaps \emph{the}
84: simplest---that combines opinion dynamics with assortative network
85: formation, revealing an apparent phase transition between regimes in which
86: one process or the other dominates the dynamics.
87: 
88: \section{Model definition}
89: 
90: Consider a network of $N$ vertices, representing individuals, joined in
91: pairs by $M$ edges, representing acquaintance between
92: individuals\footnote{
93:   Although acquaintance networks are typically simple graphs, with
94:   multiedges and self-edges disallowed, we have in the interest of
95:   simplicity, allowed multiedges and self-edges in our calculation.  Since
96:   these form only a small fraction of all edges, we expect that our results
97:   would change little if we were to remove them.}.  Each individual is assumed to hold one of~$G$
98: possible opinions on some topic of interest.  The opinion of individual~$i$
99: is denoted~$g_i$.  In the past, researchers have considered both cases
100: where $G$ is a fixed small number, such as a choice between candidates in
101: an election~\cite{cas:vot,sood:vot,sznajd}, and cases in which the number
102: of possible opinions is essentially unlimited~\cite{deuff:opi}, so that $G$
103: can be arbitrarily large.  An example of the latter might be religious
104: belief (or lack of it)---the number of subtly different religious beliefs
105: appears to be limited only by the number of people available to hold them.
106: 
107: The case of fixed small~$G$ has relatively simple behavior compared to the
108: case of arbitrarily large~$G$, and so it is on the latter that we focus
109: here.  We will assume that the number of possible opinions scales in
110: proportion to the number of individuals, and parameterize this
111: proportionality by the ratio $\gamma=N/G$.  (It is possible that not all
112: opinions will end up existing in the population.  Our model allows for some
113: opinions to become extinct as the dynamics evolves, so that the final
114: number of distinct opinions may be less than~$G$.)
115: 
116: \begin{figure}
117:   \resizebox*{.9\linewidth}{!}{\includegraphics{ill.eps}}
118:   \caption{An illustration of our model, with vertex shapes representing
119:   opinions.  At each time step the system is updated according to the
120:   process illustrated in panel~(a) with probability~$\phi$ or panel~(b)
121:   with probability $1-\phi$.  In (a) a vertex $i$ is selected at random and
122:   one of its edges (in this case the edge $(i,j)$) is rewired to a new
123:   vertex $j'$ holding the same opinion as~$i$.  In (b) vertex~$i$ adopts
124:   the opinion of one of its neighbors~$j$.
125:   }
126:   \label{fig:ill}
127: \end{figure}
128: 
129: The $M$ edges of the network are initially placed uniformly at random
130: between vertex pairs, and opinions are assigned to vertices uniformly at
131: random.  We then study by computer simulation a dynamics in which on each
132: step of the simulation we either move an edge to lie between two
133: individuals whose opinions agree, or we change the opinion of an individual
134: to agree with one of their neighbors.  To be specific, on each step we do
135: the following (see Fig.~\ref{fig:ill}).
136: \begin{enumerate}
137: \item\label{step:rewire} Pick a vertex~$i$ at random.  If the degree
138: $k_i$ of that vertex is zero, do nothing.  Otherwise, with
139: probability~$\phi$, select at random one of the edges attached to $i$ and
140: move the other end of that edge to a vertex chosen randomly from the set of
141: all vertices having opinion~$g_i$.
142: \item\label{step:vote} Otherwise (i.e.,~with probability $1-\phi$) pick a
143: random neighbor~$j$ of $i$ and set $g_i$ equal to~$g_j$.
144: \end{enumerate}
145: Step~\ref{step:rewire} represents the formation of new acquaintances
146: between people of similar opinions. Step~\ref{step:vote} represents the
147: influence of acquaintances on one another, opinions becoming similar as a
148: result of acquaintance.
149: 
150: Note that both the total number of edges~$M$ in our network and the total
151: number of possible opinions~$G$ are fixed.  In the limit of large system
152: size, the model thus has three parameters: the average degree
153: $\av{k}=2M/N$, the mean number of people holding an opinion~$\gamma=N/G$,
154: and the parameter~$\phi$.  In our studies, we primarily keep the first two
155: of these parameters fixed and ask what happens as we vary the third.
156: 
157: \section{Phases and critical scaling of community sizes}
158: 
159: The expected qualitative behavior of the model is clear.  Since both of our
160: update moves tend to decrease the number of nearest-neighbor vertex pairs
161: with different opinions, we should ultimately reach a state in which the
162: network is divided into a set of separate components, disconnected from one
163: another, with all members of a component holding the same opinion.  That
164: is, the model segregates into a set of communities such that no individual
165: has any acquaintances with whom they disagree.  We call this the
166: \defn{consensus state}.  Furthermore, once we reach consensus, all moves in
167: the model involve the random rearrangement of edges within components, and
168: hence, in the limit of long time, the components become random graphs
169: with uniform uncorrelated arrangements of their edges.
170: 
171: The primary interest in our model therefore is in the number and sizes of
172: the communities that form and in the dynamics of the model as it comes to
173: consensus.  Let us consider the distribution $P(s)$ of the sizes~$s$ of the
174: consensus communities.  In the limit $\phi\to1$, only updates that move
175: edges are allowed and hence the consensus state is one in which the
176: communities consist of the sets of initial holders of the individual
177: opinions.  Since the initial assignment of opinions is random, the sizes of
178: these sets follow the multinomial distribution, or the Poisson distribution
179: with mean~$\gamma$ in the limit of large~$N$.  Conversely, in the limit
180: $\phi\to0$, only changes of opinion are allowed and not edge moves, which
181: means that the communities correspond to the initial components in the
182: graph, which are simply the components of a random graph.  Assuming we are
183: in the regime $\av{k}>1$ in which a giant component exists in the random
184: graph, we will then have one giant (extensive) community and an exponential
185: distribution of small communities.  Thus, in varying~$\phi$ we go from a
186: situation in which we have only small communities with constant average
187: size~$\gamma$ to one in which we have a giant community plus a set of small
188: ones.
189: 
190: This is the classic behavior seen in a system undergoing a continuous phase
191: transition and it leads us to conjecture that our model displays a phase
192: transition with decreasing $\phi$ at which a giant community of like-minded
193: individuals forms.  In other words, there is a transition between a regime
194: in which the population holds a broad variety of views and one in which
195: most people believe the same thing.  We now offer a variety of further
196: evidence to support this conjecture.  (Phase transition behavior is also
197: seen in some models of opinion formation on static networks, such as the
198: model of Ref.~\cite{cc:axe}, although the mechanisms at work appear to be
199: different from those considered here.)
200: 
201: 
202: \begin{figure}
203:   \resizebox*{0.9\linewidth}{!}{\includegraphics{hist.eps}}
204:   \caption{Histograms of community sizes in the consensus state for values
205:   of $\phi$ above, at, and below the critical point in panels (a), (b), and
206:   (c) respectively.  Values of the other parameters are $N=3200$, $M=6400$
207:   (giving $\av{k}=4$), and $\gamma=10$.  In panel~(b) the distribution
208:   appears to follow a power law for part of its range with exponent $3.5\pm
209:   0.3$, as indicated by the solid line.  Numerical data are averaged over
210:   $10^4$ realizations for each $\phi$-value and binned logarithmically.
211:   }
212:   \label{fig:hist}
213: \end{figure}
214: 
215: In Fig.~\ref{fig:hist} we show plots of $P(s)$ from simulations of our
216: model for $\av{k}=4$ and $\gamma=10$.  As the figure shows, we do indeed
217: see a qualitative change from a regime with no giant community to one with
218: a giant community.  At an intermediate value of $\phi$ around $0.458$ we
219: find a distribution of community sizes that appears to follow a power law
220: $P(s)\sim s^{-\alpha}$ over a significant part of its range, another
221: typical signature of criticality.  The exponent~$\alpha$ of the power law
222: is measured to be $3.5\pm0.3$, which is incompatible with the value $2.5$
223: of the corresponding exponent for the phase transition at which a giant
224: component forms in a random graph (a transition which belongs to the
225: mean-field percolation universality class).
226: 
227: \begin{figure}
228:   \resizebox*{0.9\linewidth}{!}{\includegraphics{fss.eps}}
229:   \caption{Finite size scaling analysis for $\av{k}=4$ and $\gamma=10$.
230:   (a)~Crossing plot used to determine the critical point~$\phi_c$ and
231:   exponent~$a$.  We find $\phi_c=0.458\pm 0.008$ and $a=0.61\pm 0.05$.  The
232:   inset shows a blow-up of the region around the critical point.
233:   (b)~Scaling collapse used to determine the exponent~$b$ which is found to
234:   take the value $b=0.7\pm 0.1$.  The data are averaged over $10^4$
235:   realizations for $\phi$-value.  Error bars are shown where they are larger
236:   than the symbol size.
237:   }
238:   \label{fig:fss}
239: \end{figure}
240: 
241: To further investigate our transition, we perform a finite size scaling
242: analysis in the critical region.  To do this, we need first to choose an
243: order parameter for the model.  The obvious choice is the size~$S$ of the
244: largest community in the consensus state as a fraction of system size.  The
245: arguments above suggest that this quantity should be of size $\Ord(N^{-1})$
246: for values of $\phi$ above the phase transition (and hence zero in the
247: thermodynamic limit) and $\Ord(1)$ below it.  We assume a scaling relation
248: of the form
249: \begin{equation}\label{eq:fss}
250:   S = N^{-a}\,F\Bigl(N^b(\phi-\phi_c)\Bigr),
251: \end{equation}
252: where $\phi_c$ is the critical value of $\phi$ (which is presumably a
253: function of $\av{k}$ and~$\gamma$), $F$~is a universal scaling function
254: (bounded as its argument tends to $\pm\infty$), and $a$ and $b$ are
255: critical exponents.  To estimate $\phi_c$ we plot $ N^aS$ against $\phi$
256: and tune $a$ such that the results for simulations at different~$N$ but
257: fixed $\av{k}$ and $\gamma$ cross at a single point, which is the critical
258: point.  Such a plot for $\av{k}=4$ and $\gamma=10$ is shown in
259: Fig.~\ref{fig:fss}(a).  With $a=0.61\pm0.05$ we obtain a unique crossing
260: point at $\phi_c=0.458\pm0.008$, which agrees well with the previous rough
261: estimate of $\phi_c$ from Fig.~\ref{fig:hist}.
262: 
263: Using this value we can now determine the exponent $b$ by plotting $N^aS$
264: against $N^b(\phi-\phi_c)$.  Since $F(x)$ is a universal function, we
265: should, for the correct choice of~$b$, find a data collapse in the critical
266: region.  In Fig.~\ref{fig:fss}(b) we show that such a data collapse does
267: indeed occur for $b=0.7\pm0.1$.
268: 
269: \begin{figure}
270:   \resizebox*{0.9\linewidth}{!}{\includegraphics{phd.eps}}
271:   \caption{Values of $\phi_c$ as a function of $\gamma$ for various
272:   $\av{k}$ obtained by finite size scaling analyses using system sizes
273:   $N=200$, $400$, $800$, and $1600$ and $10^4$ realizations for each size
274:   and set of parameter values.  Note that the horizontal axis is
275:   logarithmic.
276:   }
277:   \label{fig:phd}
278: \end{figure}
279: 
280: We have performed similar finite size scaling analyses for a variety of
281: other points $(\av{k},\gamma)$ in the parameter space and, as we would
282: expect, we find that the position $\phi_c$ of the phase transition
283: varies---see Fig.~\ref{fig:phd}---but that good scaling collapses exist at
284: all parameter values for values of the critical exponents consistent with
285: the values $a=0.61$ and $b=0.7$ found above.
286: 
287: Despite the qualitative similarities between the present phase transition
288: and the percolation transition, our exponent values for $a$ and $b$ show
289: that the two transitions are in different universality classes: the
290: corresponding exponents for random graph percolation are $a=b=\frac13$,
291: which are incompatible with the values measured above.
292: 
293: \begin{figure}[b]
294:   \resizebox*{\linewidth}{!}{\includegraphics{time.eps}}
295:   \caption{Scaling of the average time $\tau$ to reach consensus.
296:   (a)~Convergence time as a function of system size for $\phi=1$.  The
297:   straight line is a fit to a logarithmic form and indicates
298:   that~$\tau\sim\log N$.  (b)~Coefficient of variation of the convergence
299:   time as a function of~$\phi$.  The vertical gray line marks the position
300:   of the critical point $\phi_c=0.458$.  (c)~Scaling plot used to determine
301:   the dynamical exponent $z$.  The crossing point falls at $\phi = 0.45\pm
302:   0.02$ consistent with the value of $\phi_c$ obtained above.  The
303:   dynamical exponent is found to take the value $z=0.61\pm 0.15$.  Parameter
304:   values are $\av{k}=4$ and $\gamma=10$ in all panels.  All data points are
305:   averaged over $10^4$ realizations.  Symbols are the same as in
306:   Fig.~\ref{fig:fss}.  For the sake of clarity, system sizes $N=400$ and
307:   $N=1600$ are omitted in~(b).
308:   }
309:   \label{fig:time}
310: \end{figure}
311: 
312: \section{Dynamical critical behavior}
313: 
314: Our model differs from percolation in another important respect also:
315: percolation is a static, geometric phase transition, whereas the present
316: model is fundamentally dynamic, the consensus arising as the limiting fixed
317: point of a converging non-equilibrium dynamics.  It is interesting
318: therefore to explore the way in which our model approaches consensus.
319: 
320: In previous studies of opinion formation models of this type on fixed
321: networks a key quantity of interest is the average convergence time~$\tau$,
322: which is the number of updates per vertex needed to reach consensus.  If
323: $\phi=0$ then $\tau$ is known to scale as $\tau\sim N$ as system size
324: becomes large~\cite{sood:vot}.  In the opposite limit ($\phi=1$), opinions
325: are fixed and convergence to consensus involves moving edges one by one to
326: fall between like-minded pairs of individuals.  This is a standard
327: sampling-with-replacement process in which the number $U$ of unsatisfied
328: edges is expected to decay as $U \sim Me^{-t/M}$ for large times~$t$.  Thus
329: the time to reach a configuration in which $U=\Ord(1)$ is $t\sim M\log M$,
330: and the convergence time is this quantity divided by the system size~$N$.
331: For fixed average degree $\av{k}=2M/N$, this then implies that
332: $\tau\sim\log N$.  This result is confirmed numerically in
333: Fig.~\ref{fig:time}(a).
334: 
335: For $\phi$ close to~$\phi_c$, experience with other phase transitions leads
336: us to expect critical fluctuations and critical slowing down in~$\tau$.
337: Figure~\ref{fig:time}(b) shows that indeed there are large fluctuations in
338: the convergence time in the critical region.  The figure shows the value of
339: the coefficient of variation~$V_\tau$ of the consensus time (i.e.,~the
340: ratio of the standard deviation of $\tau$ to its mean) as a function of
341: $\phi$ and a clear peak is visible around $\phi_c\simeq0.46$.  To
342: characterize the critical slowing down we assume that $\tau$ takes the
343: traditional scaling form $\tau\sim N^z$ at the critical point, where $z$ is
344: a dynamical exponent~\cite{katya:swxy}.  Figure~\ref{fig:time}(c) shows a
345: plot of $\tau N^{-z}$ as a function of~$\phi$.  If the system follows the
346: expected scaling at $\phi_c$ then the resulting curves should cross at the
347: critical point.  Although good numerical results are considerably harder to
348: obtain in this case than for the community sizes presented earlier, we find
349: that the curves cross at a single point if $z=0.61\pm0.15$ and
350: $\phi=0.44\pm0.03$, the latter being consistent with our previous value of
351: $\phi_c=0.46$ for the position of the phase transition.
352: 
353: \section{Summary and conclusions}
354: 
355: To summarize, we have proposed a simple model for the simultaneous
356: formation of opinions and social networks in a situation in which both
357: adapt to the other.  Our model contrasts with earlier models of opinion
358: formation in which social structure is regarded as static and opinions are
359: an outcome of that pre-existing
360: structure~\cite{cc:axe,bik:fads,arthur:contagion,watts:fad,our:fad2}.
361: Our model is a dynamic, non-equilibrium model that reaches a consensus state in
362: finite time on a finite network.  The structure of the consensus state
363: displays clear signatures of a continuous phase transition as the balance
364: between the two processes of opinion change and network rewiring is varied.
365: We have demonstrated a finite size scaling data collapse in the critical
366: region around this phase transition, characterized by universal critical
367: exponents independent of model parameters.  The approach to the consensus
368: state displays critical fluctuations in the time to reach consensus and
369: critical slowing down associated with an additional dynamical exponent.
370: The phase transition in the model is of particular interest in that it
371: provides an example of a simple process in which a fundamental change in
372: the social structure of the community can be produced by only a small
373: change in the parameters of the system.
374: 
375: Finally, we note that for the specific example of opinion formation
376: mentioned in the introduction---that of choice of religion---it is known
377: that the sizes of the communities of adherents of religious sects are in
378: fact distributed, roughly speaking, according to a power law~\cite{ZM01}.
379: This may be a signature of critical behavior in opinion formation, as
380: displayed by the model described here, although other explanations, such as
381: the Yule process~\cite{Yule25,Simon55}, are also possible.
382: 
383: \begin{acknowledgements}
384: The authors thank Paul Krapivsky and Claudio Castellano for helpful
385: suggestions and comments. This work was supported in part by the
386: Wenner-Gren Foundations (PH) and the National Science Foundation under
387: grant number DMS--0234188 (MEJN).
388: \end{acknowledgements}
389: 
390: \begin{thebibliography}{10}
391: 
392: \bibitem{AB02}
393: R.~Albert and A.-L. Barab\'{a}si.
394:  Statistical mechanics of complex networks.
395:  \textit{Rev. Mod. Phys}, 74:47--98, 2002.
396: 
397: \bibitem{arthur:contagion}
398: W.~B. Arthur and D.~A. Lane.
399:  Information contagion.
400:  \textit{Structural Change and Economic Dynamics}, 4:81--104, 1993.
401: 
402: \bibitem{bik:fads}
403: S.~Bikhchandani, D.~Hirshleifer, and I.~Welch.
404:  A theory of fads, fashion, custom, and cultural change as information
405:   cascades.
406:  \textit{Journal of Political Economy}, 100:992--1026, 1992.
407: 
408: \bibitem{BPDA04}
409: M.~Bogu{\~n}\'a, R.~Pastor-Satorras, A.~D\'{\i}az-Guilera, and A.~Arenas.
410:  Models of social networks based on social distance attachment.
411:  \textit{Phys. Rev. E}, 70:056122, 2004.
412: 
413: \bibitem{cas:vot}
414: C.~Castellano, D.~Vilone, and A.~Vespignani.
415:  Incomplete ordering of the voter model on small-world networks.
416:  \textit{Europhys. Lett.}, 63:153--158, 2003.
417: 
418: \bibitem{deuff:opi}
419: G.~Deffuant, F.~Amblard, G.~Weisbuch, and T.~Faure.
420:  How can extremism prevail? {A} study based on the relative agreement
421:   interaction model.
422:  \textit{Journal of Artificial Societies and Social Simulation}, 5:1,
423:   2002.
424: 
425: \bibitem{DM02}
426: S.~N. Dorogovtsev and J.~F.~F. Mendes.
427:  \textit{Evolution of Networks: From Biological Nets to the Internet and
428:   WWW}.
429:  Oxford University Press, Oxford, 2003.
430: 
431: \bibitem{our:fad2}
432: A.~Gr\"{o}nlund and P.~Holme.
433:  A network based threshold model for the spreading of fads in society
434:   and markets.
435:  \textit{Advances in Complex Systems}, 8:261--273, 2005.
436: 
437: \bibitem{liggett:vot}
438: T.~M. Liggett.
439:  \textit{Interacting Particle Systems}.
440:  Springer, New York, 1985.
441: 
442: \bibitem{cc:axe}
443: C. Castellano, M. Marsili, and A. Vespignani,
444: Nonequilibrium phase transition in a model for social influence.
445:  \textit{Phys. Rev. Lett.}, 85:3536--3539, 2000.
446: 
447: \bibitem{mcp:bird}
448: J.~M. McPherson, L.~Smith-Lovin, and J.~Cook.
449:  Birds of a feather: Homophily in social networks.
450:  \textit{Ann. Rev. Sociol.}, 27:415--444, 2001.
451: 
452: \bibitem{katya:swxy}
453: K.~Medvedyeva, P.~Holme, P.~Minnhagen, and B.~J. Kim.
454:  Dynamic critical behavior of the XY model in small-world networks.
455:  \textit{Phys. Rev. E}, 67:036118, 2003.
456: 
457: \bibitem{Newman03c}
458: M.~E.~J. Newman.
459:  Mixing patterns in networks.
460:  \textit{Phys. Rev. E}, 67:026126, 2003.
461: 
462: \bibitem{Newman03d}
463: M.~E.~J. Newman.
464:  The structure and function of complex networks.
465:  \textit{SIAM Review}, 45:167--256, 2003.
466: 
467: \bibitem{schelling}
468: T.~C. Schelling.
469:  \textit{Micromotives and Macrobehavior}.
470:  W.\ W.\ Norton \& Company, New York, 1978.
471: 
472: \bibitem{Simon55}
473: H.~A. Simon.
474:  On a class of skewed distribution functions.
475:  \textit{Biometrika}, 42:425--440, 1955.
476: 
477: \bibitem{Soderberg02}
478: B.~S\"{o}derberg.
479:  General formalism for inhomogeneous random graphs.
480:  \textit{Phys. Rev. E}, 66:066121, 2002.
481: 
482: \bibitem{sood:vot}
483: V.~Sood and S.~Redner.
484:  Voter model on heterogeneous graphs.
485:  \textit{Phys. Rev. Lett.}, 94:178701, 2005.
486: 
487: \bibitem{sznajd}
488: K.~Sznajd-Weron and J.~Sznajd.
489:  Opinion evolution in closed community.
490:  \textit{Int. J. Mod. Phys. C}, 11:1157--1165, 2000.
491: 
492: \bibitem{watts:fad}
493: D.~J. Watts.
494:  A simple model of global cascades on random networks.
495:  \textit{Proc. Natl. Acad. Sci. USA}, 99:5766--5771, 2002.
496: 
497: \bibitem{Yule25}
498: G.~U. Yule.
499:  A mathematical theory of evolution based on the conclusions of Dr.\
500:   J. C. Willis.
501:  \textit{Trans. R. Soc. London B}, 213:21--87, 1925.
502: 
503: \bibitem{ZM01}
504: D.~H. Zanette and S.~C. Manrubia.
505:  Vertical transmission of culture and the distribution of family
506:   names.
507:  \textit{Physica A}, 295:1--8, 2001.
508: 
509: \end{thebibliography}
510: 
511: \end{document}
512: