Physics

Procedía

An extraordinary transition in a minimal adaptive network of introverts and

extroverts

R.K.P. Zia, Wenjia Liu, and B. Schmittmann

Physics Department, Virginia Tech, Blacksburg, VA, 24061-0435 and Department of Physics and Astronomy, Iowa State University, Ames, Iowa

50011-3160

Abstract

We study a minimal adaptive network involving two populations, modeling the behavior of extreme introverts (I) and extroverts (E). When chosen to update, an I simply cuts one of its links at random while an E adds a link to any other yet-to-be-connected individual (node). In the steady state, the active links in the system are obviously only the cross-links between the I's and the E's. With no free parameters other than the numbers of each population (Ni, NE), this minimal model displays remarkable properties: Through simulations using 0(10)-0(1000) nodes, we find that the typical number of cross-links (X) fluctuates surprisingly close to the minimum or the maximum allowed values, depending on whether Ni > NE or otherwise. At the transition point (i.e., Ni = NE), the fraction X/(NiNe) wanders across a substantial part of the unit interval, much like a pure random walk confined between two walls. Since this system can be mapped to a NiNE Ising model with spin flip dynamics, we note that such fluctuations are far greater than those in the standard Ising model (at either first or second order transitions). Thus, we refer to the case here as an "extraordinary transition." Thanks to the restoration of detailed balance and the existence of a "Hamiltonian," several qualitative aspects of these remarkable phenomena can be understood analytically.

Keywords: Adaptive network, Co-evolving network, Social network, Phase transition

introduction. In a recent Workshop, we introduced a model of adaptive networks with preferred degrees and the role they play in the evolution of epidemics [1]. Motivated by a society with individuals who prefer to have some specific number, k, of contacts (or 'friends'), we studied the statistics of a network with a single, homogeneous k, as well as a network involving two communities with different k's interacting in a variety of ways. Returning to a single network, we also endowed the nodes with a simple binary degree of freedom: susceptible or infected. In contrast to the numerous studies of such contact processes (i.e., the SIS model) on regular lattices or scale-free networks [3, 4], we attempt to model human beings' natural response to the presence of an epidemic: by letting k decrease as the number of infected individuals increases [5]. Many interesting behaviors are found by simulation techniques. While some of the phenomena can be qualitatively understood through rough approximations and heuristic arguments, very little can be predicted exactly. Being a stochastic process that violates detailed balance, there are few analytic tools available to find even the stationary distribution (the equivalent of a Boltzmann factor for systems in thermal equilibrium), let alone compute averages and correlations.

In this note, we report some progress on this analytic front - for a dynamic network involving two communities, but with no degrees of freedom associated with the nodes. Letting the k's take extreme values (0, ro - a population we refer to as extreme introverts and extroverts), we show that the dynamics now obeys detailed balance. Consequently, "simple integration" can be exploited to compute the exact steady state distribution, P*. Indeed, we are able to obtain a simple closed form expression for our model. Interpretating P* as a Boltzmann factor in an equilibrium system,

Available online at www.sciencedirect.cor

SciVerse ScienceDirect

Physics Procedía 34 (2012) 124 - 127

1875-3892 © 2012 Published by Elsevier B.V. Selection and/or peer review under responsibility of the Organising Committee of the CSP 2012 Conference.

doi:10.1016/j.phpro.2012.05.020

we find the associated "Hamiltonian." The full description of our minimal model of an adaptive network for two communities can be regarded as an Ising model, evolving according to spin-flip dynamics, with only two control parameters: NI and NE, the number of introverts and extroverts. The associated "Hamiltonian" is characterized by peculiar anisotropic long-range multispin interactions. Through simulations, we find an "extraordinary" transition as NI - Ne crosses zero, involving fluctuations much larger than those found in the Lenz-Ising model (i.e., the standard model with nearest neighbor interactions) at criticality. Several qualitative aspects of our system can be understood, thanks to the existence of a Hamiltonian. In the remainder of this note, we provide a few highlights of both simulation and analytic results.

Networks with degree preference and a minimal system of extreme individuals. Focusing first on the dynamics of the links of a single network, we consider N nodes (individuals) with the same behavior. In an attempt, a random node is chosen and its degree, k, is examined. In the simplest of this class of "minimal" models, it adds/cuts a randomly chosen new/existing link if k is less/more than a fixed parameter k. To avoid ambiguity, let k be a half integer. It is easy to demonstrate that this dynamics does not satistfy detailed balance so that, when the system settles into a stationarity state, it will be in a non-equilibrium steady state, with non-trivial time-independent probability current loops[2]. The degree distribution in this state was found to be a Laplacian: ^ 3 |k k| [1], which can be understood from a simple minded argument. This picture becomes considerably more complex if we allow the population to be inhomogeneous. For example, for a system with just two "communities" (associated with k1 < k2), several new and puzzling features emerge and will be reported elsewhere [6]. Here, we restrict ourselves to a population of extreme introverts (I) and extroverts (E), i.e., ki = 0 and ke = With the I's/E's always cutting/adding links, our system quickly evolves into a state where there are no I-I links while all the E-E links are present. Only the cross-links (I-E) are actively evolving, depending on whether the associated I node or the E node is chosen.

Simulation and theoretical results. It is clear that the full configuration space for this system consists of the set of variables: {a^-}, with Aij = 0/1 representing the absence/presence of the i- j link and i = 1,..., NI; j = 1,..., NE. Of course, A here is just the active sector of the adjacency matrix. In this form, our system is precisely a NI x NE Ising model (with s- = 2A- - 1), evolving with a peculiar spin-flip dynamics. The master equation for the probability of finding the system in a particular configuration t attempts from some initial state, P (A, t), can be easily written [6]. We observe that, similar to the Lenz-Ising model, there is an exact symmetry associated with this dynamics, namely, Aij & 1 - Aji © Ni & NE.

Using simple techniques, we simulated a system of NI introverts and NE extroverts. For our system, it is more natural to define a Monte Carlo step (MCS) as NI + NE attempts, so that each node will have one chance of being selected, on the average, in a MCS. Here, we report our findings for a minimal model with NI + NE = 200, for various NI - Ne . Such a small system is chosen after we discovered an extraordinary "critical" point at NI = NE, for which relaxation times are extremely long. By contrast, the system settles quickly otherwise, even for (NI, NE) = (101,99). In particular, with 106 MCS runs, we find good statistics for typical quantities of interest, such as the degree distributions, for all NI + NE. For the NI = NE case however, the system is still evolving at such time scales. A good representation of this remarkable scenario is the time trace of X, the total number of cross-links in the system. Note that, being Ait plays precisely the role of the total magnetisation in the Lenz-Ising model (i.e., M = 2X -N; with N = NI x Ne). Fig. 1 shows three traces (green, red, and blue) corresponding to (101,99), (100,100), and (99,101). While the green and blue traces are reminiscent of M (t) in a non-interacting Ising model with an external field, the red trace is extraordinary indeed. Let us remind the reader that, if the temperature is set low enough for the spontaneous magnetisation in the Lenz-Ising model to be say, 0.75, M (t) would be mostly localized around ±0.75, with occasional quick excursions from one to the other (for a system with finite N). Instead, X seems to perform a random walk throughout its "allowed range!" Indeed, we have measured the power spectrum of this trace and found that it is essentially 1/f2 (for large frequencies, i.e., times small compared to the traversal of the allowed range), a behavior entirely consistent with pure random walks. Since X changes typically by ±1 in each attempt, it would take O (108) attempts to traverse the range of O (104). Now, a MCS is 200 attempts here, providing a consistent picture

with the O (106) MCS scale in Fig. 1. From the time traces, we compile histograms which exhibit the stationary distributions: P (X). These are displayed in Fig. 2, showing that the distribution across the "plateau" in the (100,100) case is indeed uniform! By contrast, let us consider its counterpart in a (finite) Ising model: P (M). With T < Tc and H = 0, P (M) has two sharp peaks centered on the spontaneous magnetisation, not a flat plateau from one cliff-edge to the other. Even at criticality, where P (M) has more interesting structures, its width is not extensive! The results of our

simulations show that we are dealing with a very unusual Ising system which displays extremely large fluctuations. Below, we provide some insight on this astonishing phenomenon.

t (in 106 MCS)

0 2000 4000 6000 8000 10000

Figure 1: Time traces of X in systems with (Nj, NE) = (101,99), (100,100), and (99,101), indicated by the colors green, red, and blue respectively. Note that the allowed range of X is [0, NJNE].

Figure 2: Stationary distributions P (X), obtained by compiling histograms of X from the time traces in Fig. 1, with the same color codes.

On the theoretical front, detailed balance is, remarkably, restored in this extreme limit. Furthermore, we are able to find the stationary distribution, P* (A), in closed form. Defining

kt = ZjAij; pj = Nj - ZiAij

as the degree of node i and the complement of the degree of node j respectively, we find

1 Nj Ne

(A) = Q n (k'!) 0 (j

i=1 j=1

where Q = 2{A}n (ki!) n (pjl) is the "partition function." Note that the exact symmetry is manifest here.

Despite the availability of an explicit P*, we are unable to compute Q, averages, fluctuations, and correlations of quantities of interest. On the other hand, if we regard - ln P* as a "Hamiltonian," we can appreciate why this analytic task is gargantuan. Unlike the standard Ising model, this system consists of peculiarly anisotropic, long ranged, multispin interactions: Each "spin" interacts with all others in the same row and column! Meanwhile, even P (M) is not known in closed form analytically (for the 2-d Lenz-Ising model). Thus, it would be surprising if P (X) can be found. Nevertheless, some analytic insight would be valuable. To this end, we turn to a standard first step: the mean-field approximation. In this spirit, we start with

P (X) = £ 5 (X, 2ijAijP* (A)

and simply replace all the stochastic variables by their means: ki ^ X/Nj and pj ^ Nj - X/NE. Now, the constrained sum provides just a binomial (X). Exploiting Stirling's formula with large N with finite p = X/N, we construct a "Landau free energy", F (p), defined as the negative log of this approximation for P (X). The result is

r T 1/ln p ln T1 - pl\ /1

F(p)-plnNJ +11 -plnne - 2\Ne + ) + 0 NN

As expected, the difference NJ -NE plays the role of a external field, mainly via ln (NJ/NE). At the leading order of the N's, there is no other term besides the one linear in p (which plays the role of the magnetisation m = 2p -1 € [-1, +1] in the Lenz-Ising model). The only "restoring force" comes from the ln terms, with their 0 (1/N) coefficients. Thus,

the system experiences an effectively flat landscape when Nj = NE. At generic points, the minima occur within O (1/Nln[N//Ne]) of the boundaries! This picture fits the simulation data qualitatively. Unfortunately, quantitave understanding will be, given our non-trivial "Hamiltonian" here, quite challenging.

Summary and Outlook. We report some results of a model for a social network in which individuals prefer to have some number of contacts (in contrast to say, random graphs or scale free networks). In particular, we are motivated by the presence of the two temperaments, introverts and extroverts, in our society. While a realistic model would involve preferences from a broad (but bimodal, say) spectrum, we choose to consider a system with a minimal number of parameters, specifically, the extreme case where the introverts prefer no contacts while the extroverts would prefer to be friends with all others. The only free parameters here are the numbers of individuals in each community: N¡, NE. The simplicity of this model is deceptive, as we may expect X to depend on N¡ - NE in a mundane manner. Instead, we discovered that the fraction X/(N¡NE) lies mainly near zero or unity (especially for large populations), when N¡ > NE or N¡ < Ne . More remarkable is the symmetric case N¡ = NE, in which this fraction ranges over almost the entire unit interval! The fluctuations are gigantic along with exceedingly long relaxation times. Fortunately, the stationary (microscopic) distribution can be obtained exactly in closed form. Nonetheless, few observables can be computed analytically, as we can map our system to a N¡ x NE Ising model with anisotropic, long ranged, multispin interactions.

While many interesting issues about this extreme system can be pursued, many other questions naturally arise, especially for more realistic models. For example, is there a simple way to predict the degree distributions? are there surprises associated with clustering? what if new links are not chosen at random for creation but biased towards a friend of a friend? [6]. Beyond the study of dynamics of such networks, we have started to explore the effects on the variables associated with nodes, e.g., the spread of epidemics. In particular, how is the epidemics slowed down if links are cut not at random but biased against infected individuals [5]. Needless to say, the node variables are not restricted to health, they can be wealth, opinion, information, etc. Comparing model behavior with data from real societies will be obviously interesting and important. Further into the future, investigations of a wide variety of real networks, as well as their interactions, form natural (though highly complex) extensions of the simple, baseline study presented here.

Acknowledgements We thank K. Bassler, B. Intoy, S. Jolad, H. Kim, W. Kob, M. Pleimling, and Z. Toroczkai for illuminating discussions. One of us (RKPZ) is grateful to W. Kob for suggesting the mean-field approach. This research is supported in part by a grant from the US NSF: DMR-1005417.

[1] R. K. P. Zia, Wenjia Liu, Shivakumar Jolad, and B. Schmittmann, Physics Procedia, 15, 102-105 (2011).

[2] R. K. P. Zia and B. Schmittmann, J. Stat. Mech. Theory Exp. P07012 (2007).

[3] R.M. Anderson and R.M. May, Infectious Diseases of Humans (Oxford University Press, New York, 1991).

[4] S. N. Dorogovtsev, A. V. Goltsev, and J. F. F. Mendes, Rev. Mod. Phys. 80 1275 (2008).

[5] Shivakumar Jolad, B. Schmittmann and R.K.P. Zia, to be published. See also arXiv:1109.5440.

[6] Wenjia Liu, B. Schmittmann and R.K.P. Zia, to be published.