1:
2: %\documentstyle[preprint,aps]{revtex}
3: \documentstyle[aps,prl,multicol,epsf,latexsym]{revtex}
4: %\documentstyle[aps,prl,twocolumn]{revtex}
5:
6: %\topmargin -.5in \headheight 0pt \headsep 0pt
7: %\oddsidemargin -5mm
8: %\textheight 25cm \textwidth 17cm
9: %\renewcommand{\baselinestretch}{1}
10:
11: \begin{document}
12: \draft
13:
14: \title{Nonequilibrium Phase Transitions in
15: Directed Small-World Networks}
16:
17: \author{Alejandro D. S\'anchez\cite{mail1},
18: Juan M. L\'opez\cite{mail2}
19: and Miguel A. Rodr\'{\i}guez\cite{mail3}}
20: \address{Instituto de F\'{\i}sica de Cantabria,
21: CSIC-UC, E-39005 Santander, Spain}
22:
23: \maketitle
24:
25: \begin{abstract}
26: Many social, biological, and economic systems can be
27: approached by complex networks of interacting
28: units. The behaviour of several
29: models on small-world networks
30: has recently been studied. These models are expected to
31: capture the essential features of the complex
32: processes taking place on real networks like
33: disease spreading, formation of public opinion,
34: distribution of wealth, etc. In many
35: of these systems relations are {\em directed}, in the sense
36: that links only act in one direction (outwards or inwards).
37: We investigate the effect of directed links on the behaviour
38: of a simple spin-like model evolving on a small-world network.
39: We show that directed networks may lead to a highly nontrivial
40: phase diagram including first and second-order phase transitions
41: out of equilibrium.
42: \end{abstract}
43:
44: \pacs{89.75.Hc, 64.60.Ht, 05.70.Fh, 05.70.Ln} %Revised
45:
46: \begin{multicols}{2}
47: \narrowtext
48: Complex networks have recently attracted an increasing
49: interest among physicists. The main reason being that
50: they seem to be exceedingly simple model systems of complex
51: behaviour in real world networks \cite{watts-book,amaral},
52: including chemical reaction
53: networks \cite{alon}, food webs \cite{pimm,paine,cann},
54: the Internet \cite{huberman,baraba} and the
55: World-Wide-Web \cite{guido},
56: metabolic \cite{jeong1} and
57: protein networks\cite{jeong2},
58: scientific collaboration networks \cite{newman1}, etc. The hope
59: is that the ideas and techniques developed in the last
60: fifty years in the field of statistical physics to deal with
61: cooperative phenomena in many body systems,
62: may be useful to understand emergent complex behaviour
63: in systems outside the traditional realm of physics.
64: In particular, small-world (SW) networks, recently
65: introduced by Watts and Strogatz \cite{watts}, have been very
66: much studied because they constitute an interesting attempt
67: to translate the complex topology of social, economic and
68: physical networks into a simple model. SW networks result from
69: randomly replacing a fraction $p$ of links of a $d$-dimensional regular
70: lattice with new random links. As a result of this random rewiring,
71: SW networks interpolate between the two limiting cases of a regular
72: lattice ($p=0$) and completely random graphs ($p=1$).
73: Studies of real network data have shown
74: that SW-like topologies are found in situations
75: as diverse as the network of movie actors
76: collaboration, the electric power grid of Southern California,
77: the network of world airports, the acquaintance network
78: of Mormons, etc \cite{watts-book,amaral}.
79:
80: Many topological properties of the SW model
81: have recently been investigated, as for instance
82: the shortest-path distance and clustering
83: coefficient \cite{watts,barrat},
84: the crossover from regular to SW behaviour
85: occurring at $p=0$ \cite{barthelemy},
86: a mean-field solution \cite{newman}, and percolation on
87: SW networks \cite{moore}
88: among others. Specifically, SW models are expected to play
89: an important role in understanding the interplay between
90: the underlying disordered
91: network and the dynamics of many social or economic processes,
92: such as distribution of wealth, disease
93: spreading, transmission of cultural traits,
94: and formation of public opinion \cite{watts-book}.
95:
96: In the language of social network analysis \cite{social-book},
97: sites are referred to as {\em actors}. Actors may represent
98: individuals, companies, airports, countries, etc., depending
99: on the social or economic process we are interested in.
100: Actors are linked to one another by a relational, social
101: or physical {\em tie} like
102: for instance, friendship, business transactions,
103: flight connections, kinship, or
104: scientific collaboration, among many others.
105: Some of those relational links are {\em symmetric},
106: in the sense that if Alice is tied to Bob, then Bob must
107: also be tied to Alice, as occurs for instance in
108: the authorship of scientific papers.
109: However, many other networks are {\em directed}
110: and exhibit links that are definitely {\em asymmetric},
111: like for instance in the case of networks of
112: import-export of goods, world-wide-web pages links,
113: lending transactions, food webs, cultural influences, etc.
114: In directed networks then,
115: when Alice is tied to Bob, Bob may not be linked to
116: Alice but to someone else instead. Asymmetric synaptic
117: strengths have already been shown to be very important
118: in trying to describe the process of learning in realistic
119: neural network model approaches to brain
120: function\cite{parisi,crisanti}.
121:
122: Several models have recently been studied
123: in order to understand the effect of SW topology
124: on classical systems like the Ising
125: model \cite{barrat} or the
126: spread of infections and epidemics \cite{chule,romu}.
127: Such a simple models are expected to capture the essential
128: features of the more complicated processes taking place on
129: real networks. However, as mentioned earlier, many social,
130: commercial or biological relations are asymmetric and the following
131: question naturally arises: What is the effect of directed
132: links on a simple model that evolves on the network?
133:
134: In this Letter we investigate the effect of directed SW topology
135: on the behaviour of a simple model. In case
136: only undirected links are used, our model becomes identical to
137: the classical Ising model on a standard (undirected) SW
138: network \cite{barrat}.
139: This allows us to study, in a systematic way,
140: the effect of directed ties on this classical model.
141: We find that the existence of directed links
142: completely changes the behaviour of the system from
143: mean-field behaviour (for undirected networks) to
144: a highly nontrivial and rich phase diagram in the case of
145: directed networks.
146: By means of extensive numerical simulations
147: we find that, for rewiring probabilities in
148: the range $0 < p < p_c$,
149: the model exhibits a line of continuous phase
150: transitions from an ordered to a disordered state. Those phase
151: transitions occur at a critical value of the temperature $T_c(p)$,
152: which depends on $p$. However, for higher disorder
153: densities $p_c < p \leq 1$, the phase transition becomes
154: first-order.
155: Our results show that, in order to model biological, social
156: or economic processes on complex networks,
157: it is crucial to take into account the character,
158: directed or undirected, of the corresponding relational links.
159:
160: {\em The model.-}
161: We have studied directed networks in $d=1$ and 2.
162: For simplicity we focus here on $d=2$ and further results
163: in $d=1$ will be published elsewhere.
164: In order to construct a directed SW network,
165: we start from a 2-dimensional square lattice consisting
166: of sites linked to their four nearest-neighbours by
167: both, outgoing and incoming links.
168: Then, with probability $p$ we reconnect nearest-neighbour
169: outgoing links to a different site chosen at random.
170: After repeating this process for every outgoing link we are left
171: with a network with a density $p$ of SW directed links,
172: as shown in Figure 1.
173: Note that, by this procedure, every site will have exactly
174: four outgoing links and a varying (random) number of
175: incoming links. Generalization to higher dimensions is
176: straightforward.
177:
178: Adopting social network nomenclature,
179: actors are then placed at the network sites.
180: Any given actor is connected by four outgoing links to
181: other actors, which we call {\em mates}.
182: We allow every actor to be in one of two possible
183: states, so that, at any given time, the state of an
184: actor is described by a binary spin-like variable
185: $s_i \in \{+1,-1\}$. Depending on the state of their
186: mates, an actor may change its state
187: according to a majority (ferromagnetic) rule:
188: Actors prefer to be in the same state as their mates are.
189: In order to implement this, we introduce the pay-off function
190: \begin{equation}
191: G(i)=2 s_i\sum _{{\rm mates \: of}\: i} s_j,
192: \label{pay-off}
193: \end{equation}
194: where the sum is carried out over the four mates of
195: actor $i$. Note that this pay-off function is positive
196: whenever $s_i$ points in the same direction as the majority of
197: its four mates.
198: External noise is included to allow
199: some degree of randomness in the time evolution by means of
200: a temperature-like parameter, $T$, which we shall call
201: temperature for short from now on.
202: For a given value of the external temperature,
203: the update of the model is then performed as follows:
204: At each time step, an actor (network site)
205: is randomly chosen and
206: its corresponding pay-off function $G(i)$ is computed according
207: to Eq. (\ref{pay-off}). If $G(i) < 0$, actor $i$ is opposing
208: its mates' majority and the change $s_i \to -s_i$ is accepted.
209: Unfavorable changes, {\it i.e.} when $G(i) > 0$, are accepted with
210: probability $\exp[-G(i)/T]$, which
211: depends on temperature in the usual fashion.
212:
213: Concerning the physics of the above defined model,
214: there are two interesting points that should be
215: explicitly mentioned. On the one hand,
216: the model is nonequilibrium since detailed balance
217: is not satisfied. On the other hand,
218: the model is not simply the asymmetric
219: counterpart of the Ising model, since the pay-off
220: function $G(i)$ in Eq. (\ref{pay-off}) does not include
221: the corresponding interaction
222: terms coming from the ingoing links (needed
223: in order to identify $G(i)$ with the energy change
224: after a spin update in the asymmetric Ising model). In fact,
225: one can easily see that the pay-off function $G(i)$
226: of our model cannot be written as variation of
227: any Hamiltonian.
228: However, we would like to remark that if only symmetric
229: links are allowed, our model becomes
230: exactly equal to the (equilibrium) Ising model in
231: an undirected SW network that was studied
232: in Ref.\cite{barrat}.
233: This can be seen by a simple comparison of
234: the pay-off function Eq. (\ref{pay-off}) with the
235: change of energy after a spin update in the
236: standard (symmetric) Ising
237: ferromagnet. In this case, it is known that
238: the system presents mean-field behaviour for
239: any value of the disorder $p>0$ \cite{barrat}.
240:
241: {\em Results.-}
242: We have carried out extensive numerical simulations
243: of the model for different values of the
244: density of SW directed ties $p$ and temperature.
245: Our results are qualitatively the same for
246: directed networks generated from regular lattices
247: in $d=1$ and $2$. In the following we focus on
248: $d=2$.
249: We have simulated the model in directed SW networks
250: generated from $L \times L$ square lattices for sizes
251: ranging from $L=8$ to $100$ and different
252: rewiring probabilities $p \in [0,1]$.
253: The system is left to evolve until, after some transient,
254: a stationary nonequilibrium state is reached.
255: The stationary state can be described by the
256: appropriate
257: order parameter, which can be defined in a
258: natural way
259: by means of the 'magnetization' per site
260: \begin{equation}
261: \label{m}
262: m = {1 \over L^2} \sum_{i=1}^{L^2} s_i.
263: \end{equation}
264: We find that the system becomes ordered,
265: {\it i.e.} $\langle |m| \rangle \neq 0$,
266: below a critical temperature $T_c(p)$, so that
267: most actors are, in average, in the same state.
268: In Figure 2 the average absolute value of the order
269: parameter is plotted {\it vs.} temperature for two
270: different values of the disorder $p=0.1$ and 0.9,
271: calculated in systems of different sizes.
272: For every system size $L^2$,
273: results were averaged over both, $10$ runs
274: of the dynamics for each network
275: realization and $n$ different realizations of
276: the network, in such a way that
277: $n \times L^2 \approx 1.5 \times 10^5$.
278: Figure 2 shows that the order-disorder transition
279: is continuous (top panel) for a low disorder
280: density, while it becomes discontinuous for a
281: higher concentration of directed SW links (bottom panel).
282: A more systematic study of the phase diagram,
283: as shown in Figure 3, reveals that there is a line of
284: continuous phase transitions for disorder
285: densities below some critical value $p_c$.
286: Very interestingly, the transition becomes first-order
287: above $p_c$, indicating that there exists a
288: {\em nonequilibrium tricritical
289: point} at $p_c$. We estimate $p_c$ to be roughly
290: at $p_c=0.65(5)$. The character, continuous or discontinuous,
291: of the phase transition is better realized when looking at
292: the probability density function (PDF) of the order parameter.
293: For sake of illustration,
294: the insets of Fig. 3 show typical PDFs at points of
295: the phase diagram (p,T), all near the critical line.
296: From those PDFs, one can see that
297: the phase transition is second-order from Fig. 3a to 3b, in the
298: region $p<p_c$. In contrast, for $p>p_c$
299: the transition is discontinuous, from Fig. 3c to 3d.
300: The most probable values of $m$, which correspond to the
301: two equally highest symmetric peaks in Fig. 3c, become unstable
302: in favor of $m=0$ as the transition line is crossed
303: towards Fig. 3d. The transition occurs in such a way
304: that the order parameter exhibits a finite jump at the critical line.
305:
306: The critical behaviour of the model
307: in the region $p<p_c$, where transitions are continuous,
308: can be studied in detail. We find that the order parameter
309: exhibits finite-size scaling with exponents that depend on
310: the disorder density $p$. Close to the critical point,
311: $|t| \to 0$,
312: we have $\langle |m| \rangle \sim |t|^\beta$, where
313: $t = (T - T_c)/T_c$ is the reduced temperature.
314: At the critical point, $t=0$,
315: the order parameter scales with system size as
316: $\langle |m| \rangle \sim L^{-\beta/\nu}$, where $\nu$
317: is the correlation length exponent. Figure 4 displays
318: the behaviour of $\langle |m| \rangle$ {\it vs.} $L$ for
319: two disorder densities $p=0.1$ and 0.5 below $p_c \approx 0.65$.
320: Only for $T=T_c(p)$ a power-law is obtained and the slope
321: of the straight line in a log-log plot gives an estimation
322: of the ratio $\beta/\nu$ between critical exponents. From
323: Fig. 4 we obtain that for $p=0.1$ and $p=0.5$,
324: $\beta/\nu=0.53(2)$ and $0.40(3)$, respectively.
325: Moreover, from data collapse analysis (not shown) at the
326: corresponding $T_c(p)$ we have $\beta=0.50(3)$,
327: $\nu=0.94(6)$, and $\beta=0.30(3)$,
328: $\nu=0.80(3)$ for $p=0.1$ and $p=0.5$, respectively.
329: These critical exponents
330: are different from both mean-field, ($\beta = \nu = 1/2$)
331: and exact values ($\beta = 1/8$ and $\nu = 1$) for the
332: Ising model in $d=2$.
333:
334: {\em Conclusions.-}
335: Many social, economic and biological networks
336: in the real world exhibit directed ties or
337: relations. This may be modeled by including
338: directed links in the corresponding complex
339: network model. In addition, spin-like models,
340: borrowed from statistical physics, have recently
341: been proposed as toy models to understand
342: some social processes, like for instance
343: conflict vs. cooperation
344: among coalitions \cite{galam,galam2}
345: or formation of cultural domains \cite{claudio}.
346: We claim that the directedness
347: of the network may strongly affect the behaviour
348: of simple processes evolving on complex networks.
349: We studied the effect of a directed small-world
350: topology on a very simple spin model.
351: Our model becomes equal to the Ising model when
352: all links are undirected. From numerical simulations
353: we showed that, when directed links exist,
354: the phase diagram of
355: the model is nontrivial. We found that the system exhibits
356: continuous phase transitions for disorder densities
357: below a critical threshold $p_c \approx 0.65$.
358: For stronger disorder, the transition is
359: first-order. At this stage we can only speculate
360: that the competition among weekly coupled
361: clusters may be related
362: to the existence of first-order transitions.
363:
364: We believe that the effect of directed
365: links may be relevant in
366: other types of disordered networks, like free-scale
367: networks, and different dynamical models.
368: In trying to model real systems, directed links
369: may play an important and unforeseen role.
370:
371: {\em Acknowledgement.-}
372: We would like to thank J. G\"uemez for collaboration
373: at the early stages of this work. We are also grateful to
374: J.J. Ramasco, L. Pesquera and X. Guardiola
375: for stimulating discussions and comments.
376: A.D.S. acknowledges postdoctoral fellowship from
377: CONICET (Argentina).
378: J.M.L. acknowledges financial support from the
379: European Commission (contract HPMF-CT-1999-00133).
380: Partial financial support from Ministerio de Ciencia y
381: Tecnolog{\'\i}a (Spain) through projects
382: BFM2000-0628-C03-02 and BFM2000-1105 is acknowledged.
383:
384:
385: \begin{thebibliography}{99}
386:
387: \bibitem[*] {mail1}
388: Electronic address: asanchez@ifca.unican.es
389:
390: \bibitem[\dag] {mail2}
391: Electronic address: lopez@ifca.unican.es
392:
393: \bibitem[\ddag] {mail3}
394: Electronic address: rodrigma@ifca.unican.es
395:
396: \bibitem{watts-book} D. J. Watts, {\em Small Worlds: The Dynamics of
397: Networks Between Order and Randomness}, Princeton Univ. Press,
398: Princeton, New Jersey (1999).
399:
400: \bibitem{amaral} L. A. N. Amaral {\it et. al.},
401: Proc. Natl. Acad. Sci. {\bf 97}, 11149 (2000).
402:
403: \bibitem{alon} U. Alon {\it et. al.},
404: Nature (London) {\bf 397}, 168 (1999).
405:
406: \bibitem{pimm} S. L. Pimm, J. H. Lawton, and J. E. Cohen,
407: Nature (London) {\bf 350}, 669 (1991).
408:
409: \bibitem{paine} R. T. Paine,
410: Nature (London) {\bf 355}, 73 (1992).
411:
412: \bibitem{cann} K. McCann, A. Hastings, and G. R. Huxel,
413: Nature (London) {\bf 395}, 794 (1998).
414:
415: \bibitem{huberman} B. A. Huberman and L. A. Adamic,
416: Nature (London) {\bf 401}, 131 (1999).
417:
418: \bibitem{baraba} R. Albert, H. Jeong, and A.-L. Barab\'asi,
419: Nature (London) {\bf 401}, 130 (1999).
420:
421: \bibitem{guido} G. Caldarelli, R. Marchetti and L. Pietronero,
422: Europhys. Lett. {\bf 52}, 386 (2000).
423:
424: \bibitem{jeong1} H. Jeong {\it et. al.},
425: Nature (London) {\bf 407}, 651 (2000).
426:
427: \bibitem{jeong2} H. Jeong {\it et. al.},
428: Nature (London) {\bf 411}, 42 (2001).
429:
430: \bibitem{newman1} M. E. J. Newman,
431: Proc. Natl. Acad. Sci. {\bf 98}, 404 (2001).
432:
433: \bibitem{watts} D. J. Watts and S. H. Strogatz,
434: Nature (London) {\bf 393}, 440 (1998).
435:
436: \bibitem{barrat} A. Barrat and M. Weigt,
437: Eur. Phys. J. B {\bf 13}, 547 (2000).
438:
439: \bibitem{barthelemy} M. Barth\'el\'emy, and L. A. N. Amaral,
440: Phys. Rev. Lett. {\bf 82}, 3180 (1999).
441:
442: \bibitem{newman} M. E. J. Newman, C. Moore, and D. J. Watts,
443: Phys. Rev. Lett. {\bf 84}, 3201 (2000).
444:
445: \bibitem{moore} C. Moore and M. E. J. Newman,
446: Phys. Rev. E {\bf 61}, 5678 (2000).
447:
448: \bibitem{social-book} S. Wasserman and K. Faust,
449: {\em Social network analysis} (Cambridge University
450: Press, Cambridge, 1994).
451:
452: \bibitem{parisi} G. Parisi,
453: J. Phys. A: Math. Gen. {\bf 19}, L675 (1986).
454:
455: \bibitem{crisanti} A. Crisanti, M. Falcioni, and A. Vulpiani,
456: J. Phys. A {\bf 26}, 3441 (1993).
457:
458: \bibitem{chule} M. Kuperman and G. Abramson,
459: Phys. Rev. Lett. {\bf 86}, 2909 (2001).
460:
461: \bibitem{romu} R. Pastor-Satorras and A. Vespignani,
462: Phys. Rev. Lett. {\bf 86}, 3200 (2001).
463:
464: \bibitem{galam} S. Galam, Physica A {\bf 230}, 174 (1996).
465:
466: \bibitem{galam2} R. Florian and S. Galam,
467: Eur. Phys. J. B {\bf 16}, 189 (2000).
468:
469: \bibitem{claudio} C. Castellano, M. Marsili, and A. Vespignani,
470: Phys. Rev. Lett. {\bf 85}, 3536 (2000).
471:
472: \end{thebibliography}
473:
474: \begin{figure}
475: \centerline{
476: \epsfxsize=5.0cm
477: \epsfbox{fig1.eps}}
478: \caption{Sketch of a directed small-world network constructed from
479: a square regular lattice in $d=2$. For sake of clarity only a few links
480: have been reconnected. Arrows indicate the direction
481: of the corresponding link. Dotted lines represent rewired links.
482: Note that every site always has four outgoing links.}
483: \end{figure}
484:
485: \begin{figure}
486: \centerline{
487: \epsfxsize=7.5cm
488: \epsfbox{fig2.eps}}
489: \caption{Order parameter {\it vs.} $T$ for different system
490: sizes. For $p=0.1$ (top panel) the transition is continuous.
491: For a higher disorder density, $p=0.9$, the transition becomes
492: first-order (bottom panel).}
493: \end{figure}
494:
495: \begin{figure}
496: \centerline{
497: \epsfxsize=7.5cm
498: \epsfbox{fig3.eps}}
499: \caption{Phase diagram of the model. The system
500: is in the ordered state below the line. Points
501: are numerical determinations of the critical
502: temperatures $T_c(p)$ for different degrees of disorder.
503: The transition is continuous for small values of
504: $p$ (circles), while it becomes discontinuous
505: for $p$ larger than $p_c \approx 0.65$
506: (filled circles). The insets show the
507: PDFs of $m$ for: a) $p=0.1$, $T=2.68$;
508: b) $p=0.1$, $T=2.70$; c) $p=0.9$, $T=2.498$; d) $p=0.9$,
509: $T=2.500$. Simulations were performed
510: in a $100 \times 100$ sites network.}
511: \end{figure}
512:
513: \begin{figure}
514: \centerline{
515: \epsfxsize=7.5cm
516: \epsfbox{fig4.eps}}
517: \caption{Finite-size scaling of the order parameter for
518: $p=0.1$ (hollow symbols) and $p=0.5$ (filled symbols),
519: both below $p_c$. Power-law behaviour is obtained for
520: $T=2.700(2)$ and $T=2.635(3)$ for $p=0.1$ (squares)
521: and $p=0.5$ (filled squares), respectively.}
522: \end{figure}
523:
524: \end{multicols}
525: \end{document}
526:
527: