1: \documentstyle[twocolumn,aps,prl,graphicx,amssymb]{revtex}
2:
3:
4:
5: \begin{document}
6:
7: \draft
8:
9: \title{Complex spatial organization in a simple model of
10: resource allocation}
11:
12: \author{Dami{\'a}n H. Zanette}
13: \address{Consejo Nacional de Investigaciones Cientificas y T\'ecnicas,
14: \\
15: Centro At\'omico Bariloche and Insituto Balseiro, 8400 Bariloche,
16: R\'{\i}o Negro, Argentina}
17:
18: \date{\today}
19:
20: \maketitle
21:
22: \begin{abstract}
23: A dynamical model for the distribution of resources between competing
24: agents is studied. While global competition leads to the
25: accumulation of all the resources by a single agent, local competition
26: allows for a wider resource distribution. Multiplicative processes
27: give rise to almost-ordered spatial structures, thourgh the enhancement
28: of random fluctuations.
29: \end{abstract}
30:
31: \pacs{PACS numbers: 87.23.Ge, 87.23.Cc, 05.65.+b}
32:
33:
34: Though many real systems in the scopes of biology and the social
35: sciences are well described by agent-based models with pair
36: interactions \cite{Murray,Bak,wta}, qualitatively similar to the
37: physical description of interacting particles, a large class of
38: biological and social processes are driven by interactions mediated by
39: an external actor, which generally bears no resemblance with the
40: individual agents. This kind of processes---which, in connection with
41: physics, may be assimilated to the evolution of globally coupled
42: dynamical systems \cite{global}---is typically found in systems where
43: agents compete for resources. A particularly relevant instance of this
44: situation is given by competing biological species. In this case, in
45: fact, the interaction is rarely given by struggling events between
46: individuals of different species, but rather by accessing
47: simultaneously
48: to the limited resources provided by the environment. The
49: performance of each agent, i.e. of each species, is here measured by
50: its ability to get resources, which depends both on the capabilities
51: of each individual and on global features such as the total population
52: of the species. The same scenario is found in some economical systems,
53: for instance, in companies competing for financial resources, usually
54: administrated by banks, or even in scientific research projects
55: competing for funds from a government agency. A key ingredient in the
56: dynamics of these systems is that the ability of each agent to get
57: resources at a given time depends, often strongly, on the resources
58: assigned to the agent at previous stages. This can give rise to a
59: multiplicative process that, in the absence of buffer mechanisms,
60: leads to resource accumulation by a single agent. In the following, we
61: consider a simple dynamical model that incorporates these elements
62: and, in particular, study the effects of competition at local level.
63: We find that local competition softens the process of resource
64: accumulation and gives rise, through the enhancement of spatial
65: fluctuations, to nontrivial structures.
66:
67: Our system consists of an ensemble of $N$ agents, each of them
68: characterized by its productivity $\alpha_i$. For future convenience,
69: we consider the generic situation where all productivities are
70: different. At each time step $t$, each agent is assigned an amount of
71: resources $r_i(t)$, which is used to produce an amount of products
72: $p_i(t)= \alpha_i r_i(t)$. At the next time step, resources are
73: distributed among agents in amounts proportional to their production,
74: namely,
75: \begin{equation} \label{model}
76: r_i(t+1) = \frac{p_i(t)}{\sum_j p_j(t)} R(t+1) =
77: \frac{\alpha_i r_i(t)}{\sum_j \alpha_j r_j(t)} R(t+1),
78: \end{equation}
79: where $R(t+1)$ are the total resources available at time $t+1$. Model
80: (\ref{model}) can be fully solved for arbitrary productivities
81: and initial conditions $r_i(0)$. In the first place, we note that
82: rescaling $r_i(t)/R(t) \to r_i(t)$, Eq. (\ref{model}) becomes
83: independent of the total resources. We thus fix $R(t)=1$ for all $t$,
84: so that $\sum_j r_j(t)=1$. In this situation, the solution to
85: Eq. (\ref{model}) reads
86: \begin{equation} \label{solution}
87: r_i(t) = \frac{\alpha_i^t r_i(0)}{\sum_j \alpha_j^t r_j(0)}.
88: \end{equation}
89: For asymptotically large times, $r_i(t) \to 0$ for all $i$, except for
90: the agent with the maximal productivity, $\alpha_{\max} = \max_i
91: \{\alpha_i \}$, which receives all the available resources. Due to
92: the multiplicative effect of resource allocation according to
93: production, all resources are in the long run accumulated by the agent
94: with the maximal productivity, giving rise to a sort of
95: winner-takes-all state \cite{wta}. By analogy with population
96: dynamics we say that the remaining agents become extinct. In fact, in
97: connection with biological populations, this result is a realization
98: of a well-known principle of ecology, namely, the principle of
99: competitive exclusion \cite{theor,Murray}: in a system of biological
100: species competing for the same resources, only one survives and the
101: others undergo extinction. At moderately large $t$ the resources
102: assigned to each agent are well approximated by
103: $r_i(t)=(\alpha_i/\alpha_{\max})^t r_i(0)/r_{\max}(0)$.
104: Assuming that the
105: productivities $\alpha_i$ are drawn at random from a distribution
106: $P_\alpha (\alpha)$ and that, for simplicity, resources are evenly
107: assigned at the beginning, $r_i(0)=N^{-1}$ for all $i$, the
108: probability distribution for the individual resources at a given time
109: is
110: \begin{equation} \label{P2}
111: P_r(r_i) = \frac{\alpha_{\max}}{r_it} r_i^{1/t} P_\alpha
112: \left( \alpha_{\max} r_i^{1/t} \right) .
113: \end{equation}
114: The dependence of this function on $r_i$ through the power $r_i^{1/t}$
115: is very weak for large $t$, so that on a wide interval of the variable
116: we find $P_r(r_i) \propto r_i^{-1}$. For long times, thus, resources
117: are distributed among agents following a power law with exponent $-1$.
118: Compare this result with Pareto's law of wealth distribution
119: \cite{Pareto,wta}.
120:
121: Model (\ref{model}) admits several variations, that may be of interest
122: in connection with the description of some real systems. For instance,
123: the extinction of all but one agent can be avoided by modifying
124: slightly the allocation method. If, at each time step, a small
125: fraction $\rho$ of the total resources is evenly distributed among the
126: agents whereas the remaining is assigned according to production as
127: above, the individual resources are always larger than $\rho/N$. In
128: this situation, $P_r(r_i)$ becomes stationary for long times and, if
129: $\rho$ is small enough, the power-law dependence quoted above is found
130: again for $r_i>\rho/N$.
131:
132: A second variation consists in assuming that the productivity can in
133: turn depend on the individual resources, for instance, as
134: $\alpha_i(r_i) = \alpha_i^0 A(r_i)$.
135: %This additional nonlinearity is
136: %characterized by the function $A(r)$, which stands for the profile of
137: %the relation between productivity and resources. Taking $A(0)=1$, the
138: %weight $\alpha_i^0$ is the productivity for vanishing resources.
139: If
140: $A(r)$ increases with $r$, the winner-takes-all effect is
141: enhanced. The opposite case, where $A(r)$ is a decreasing function, is
142: more interesting. Lower productivities for higher resources may be the
143: consequence of size effects, crowding, or loss of efficiency due to
144: lower competition. In this case, the system evolves to a stationary
145: situation where the agents whose productivity weights $\alpha_i^0$ are
146: above a certain threshold $\alpha_{\min}^0$ receive nonzero resources,
147: whereas the agents with $\alpha_i^0<\alpha_{\min}^0$ become
148: extinct. In the stationary situation, all the surviving agents have
149: the same productivity $\alpha$. The third instance, where $A(r)$ is a
150: nonmonotonous function, can give origin to multiple stationary
151: nontrivial solutions.
152:
153: In the following we focus the attention on a variation of model
154: (\ref{model}) that introduces a spatial distribution of agents. Though
155: the total resources are still allocated among all the ensemble, the
156: agents compete at a local level only. The set of agents that compete
157: with a given agent $i$ defines its neighborhood ${\cal
158: N}_i$. Resources are assigned according to
159: \begin{equation} \label{local}
160: r_i(t+1) = \frac{1}{Z(t)}
161: \frac{\alpha_i r_i(t)}{\sum_{j \in {\cal N}_i} \alpha_j
162: r_j(t)},
163: \end{equation}
164: where
165: \begin{equation} \label{Z}
166: Z(t) = \sum_i \frac{\alpha_i r_i(t)}{\sum_{j \in {\cal N}_i}
167: \alpha_j r_j(t)}
168: \end{equation}
169: is a normalization factor that insures that $\sum_j r_j(t) =1$ for all
170: $t$. We have performed numerical simulations of this system in
171: ensembles of $N=10^3$ to $10^4$ agents distributed on various
172: geometries. The productivities $\alpha_i$ were drawn at random from a
173: uniform distribution in ($0,1$). Note that, in Eqs. (\ref{model}) and
174: (\ref{local}), a homogeneous rescaling of all productivities leaves
175: the models invariant. Several distributions of initial conditions were
176: tested, but there are no essential differences with the case where,
177: at the first step, resources are evenly allocated,
178: $r_i(0)=N^{-1}$. Therefore, we concentrate our simulations on this
179: simple case.
180:
181: In the first place, we consider a one-dimensional array of agents with
182: periodic boundary conditions. The neighborhood of the $i$th agent
183: consists of its two nearest neighbors. We find that, for long times,
184: approximately half of the agents become extinct and resources are
185: evenly allocated among the surviving agents. Practically everywhere
186: along the array agents are ordered in an alternating sequence of
187: extinct agents and survivors. Occasionally, one finds two neighbor
188: sites where both agents are extinct, but two survivors are never
189: contiguous. In other words, in the neighborhood of a surviving agent
190: no other survivor can be found [cf. model (\ref{model})]. The
191: surprising feature of this asymptotic distribution is that the strong
192: correlation between survival and productivity found for Eq. (\ref{model})
193: is apparently lost in Eq. (\ref{local}). In fact, productivities are
194: distributed completely at random along the array, whereas the
195: resulting distribution of survivors is highly ordered. At first
196: glance, no connection can be established, for instance, between
197: survivor sites and local maxima of the productivity or any spatial
198: pattern caused by fluctuations in the distribution of $\alpha_i$.
199:
200: \begin{figure}
201: \centering
202: \resizebox{\columnwidth}{!}{\includegraphics{2d.eps}}
203: \caption{Asymptotic state on a $100\times 100$-site lattice with
204: nearest-neighbor competition and periodic boundary conditions. Full
205: squares represent survivors.}
206: \label{2D}
207: \end{figure}
208:
209: The same feature is found in a two-dimensional array with periodic
210: boundary conditions, where the neighborhood of each site consists of
211: the four nearest neighbors. Figure \ref{2D} displays the asymptotic
212: state on a $100 \times 100$-site lattice ($N=10^4$), where the full
213: squares represent survivor sites. Regular domains where survivors and
214: extinct agents alternate in both dimensions, separated by worm-like
215: boundaries formed by extinct agents, are apparent. We have also
216: verified that different definitions of the neighborhood of a site,
217: both in one and in two-dimensional lattices, produce similar
218: almost-periodic asymptotic structures. Again, resources are evenly
219: distributed between the surviving agents and, in neighborhood of a
220: survivor, no other survivor can be found. For instance, in a
221: one-dimensional array where the neighborhood consists of four sites
222: (nearest and next-to-nearest neighbors), the resulting structure is a
223: periodic sequence of two extinct agents and one survivor. Occasional
224: defects, with larger zones of extinction, are also found.
225:
226: In order to trace the origin of the almost-periodic structures
227: emerging at asymptotically long times we have inspected the successive
228: states of the system from the earliest evolution stages, for the
229: specific case of a one-dimensional array of agents with two neighbors
230: per site and periodic boundary conditions. As above, productivities
231: are drawn at random from a uniform distribution and $r_i(0)=N^{-1}$
232: for all $i$. We find that even at the first step of the evolution the
233: distribution of resources already exhibits an almost-periodic sequence
234: of relatively high and low values, in spite of the fact that
235: productivities are spatially uncorrelated. Namely, for a substantial
236: fraction of agents, we have $[r_{i+1}(1)-r_i(1)] [r_i(1)-r_{i-1}(1)]
237: <0$. This observation reveals that the simple dynamics of model
238: (\ref{local}) is able to introduce strong spatial correlations to an
239: initially uncorrelated state, even at the level of a single evolution
240: step.
241:
242: To quantify this effect, we take a set of five uncorrelated random
243: numbers $\alpha_1, \dots ,\alpha_5$ drawn from the productivity
244: distribution, and consider the combinations
245: \begin{equation} \label{comb}
246: \sigma_{i}=\frac{\alpha_i}{\alpha_{i-1}+\alpha_i+\alpha_{i+1}}
247: \end{equation}
248: for $i=2,3,4$. The quantities $\sigma_i$ are then proportional to the
249: resources received at $t=1$ by three consecutive agents in the array
250: [cf. Eq. (\ref{local})]. Numerical realizations of these quantities,
251: over $10^6$ independent choices of $\alpha_1,\dots, \alpha_5$, show
252: that $(\sigma_4-\sigma_3)(\sigma_3-\sigma_2)<0$ with probability $p_1
253: = 0.762$. Consequently, slightly more than 76 \% of the agents are
254: expected to belong to a periodic sequence of alternating high and low
255: resources at the first time step. The remaining 24 \% stands for
256: defects in the periodic structure. The same kind of analysis can be
257: performed for successive steps in the evolution. It is found that
258: correlations are further enhanced by the dynamics. At the second
259: evolution step, for instance, the probability for an agent to belong
260: to the periodic structure grows to $p_2 = 0.782$. For later stages,
261: we find $p_5=0.816$, $p_{10}=0.839$, $p_{20}=0.854$, and
262: $p_{50}=0.864$. These probabilities saturate at $p_\infty \approx
263: 0.865$.
264:
265:
266: \begin{figure}
267: \centering
268: \resizebox{\columnwidth}{!}{\includegraphics{prob1.eps}}
269: \caption{Survival probability as a function of the productivity
270: $\alpha$ for one dimensional arrays of $1000$ agents and different
271: neighborhoods. Dots correspond to average measurements over $100$
272: realizations. The line shows the result of an independent calculation
273: method (see text), for the case of two neighbors.}
274: \label{prob1}
275: \end{figure}
276:
277: While the emergence of almost-periodic structures seems to indicate
278: that no correlations subsist in model (\ref{model}) between survival
279: and productivity, the inspection of the evolution equations as well as
280: of Eq. (\ref{comb}) suggests that some remaining correlation should
281: however exist. In fact, low productivities still imply few resources,
282: so that the minima in the periodic structure should preferably
283: coincide with sites with small $\alpha_i$. Moreover, defects in the
284: periodic structures---where extinct agents are found in consecutive
285: sites---should also correspond to low-productivity zones. Such
286: remaining correlation can be characterized by the probability of
287: survival as a function of the productivity, $p_s(\alpha)$. This
288: probability is defined as the fraction of agents with productivity
289: $\alpha$ which survive at asymptotically long times. We have
290: numerically measured $p_s (\alpha)$, in series of $100$ independent
291: realizations of the productivity distribution for ensembles with
292: $N=1000$, for various geometries and neighborhoods. Figure \ref{prob1}
293: shows the survival probability as a function of $\alpha$ for
294: one-dimensional arrays with different numbers of neighbors. Here, the
295: correlation between survival probability and productivity is
296: apparent. Agents with larger productivities are more likely to survive
297: that those with small $\alpha$. Note, however, that for $\alpha=1$ the
298: survival probability is less than unity, so that even with the maximal
299: productivity there are chances of undergoing extinction. Conversely, for
300: $\alpha \to 0$ (but $\alpha\neq 0$), the survival probability is
301: finite. As the number of neighbors grows the probability is more
302: concentrated towards larger productivities, as expected. In the limit
303: where the neighborhood extends to the whole array the original model
304: (\ref{model}) is recovered, and the survival probability must vanish
305: except for $\alpha=1$.
306:
307: The survival probability can be calculated independently of the
308: numerical realization of the model, using expressions as in
309: Eq. (\ref{comb}). For a fixed value of $\alpha_3$ and different random
310: choices of $\alpha_1$, $\alpha_2$, $\alpha_4$ and $\alpha_5$, an
311: estimate at the first evolution step of the survival probability of
312: the agent with productivity $\alpha_3$ is given by the fraction of
313: instances for which $\sigma_3> \sigma_2, \sigma_4$. Generalizing this
314: procedure for successive time steps, the estimate can be improved by
315: considering later stages in the evolution. The line in
316: Fig. \ref{prob1} corresponds to this estimate at $t=50$. It shows an
317: excellent agreement with the numerical realizations.
318:
319: \begin{figure}
320: \centering
321: \resizebox{\columnwidth}{!}{\includegraphics{prob2.eps}}
322: \caption{Survival probability as a function of the productivity
323: $\alpha$ for different geometries with four neighbors per site.}
324: \label{prob2}
325: \end{figure}
326:
327:
328:
329: To compare now several geometries, we have measured the survival
330: probability in systems where each agent has the same number of
331: neighbors---specifically, four---on different kinds of arrays, namely,
332: a one-dimensional array (nearest and next-to-nearest neighbors), a
333: two-dimensional lattice (nearest neighbors) and a random graph with
334: four connections per site. The results, shown in fig. \ref{prob2}
335: corresponds to averages over $100$ realizations in systems with
336: $N=1000$ for one dimension and random graphs. In this latter case,
337: the graph topology is chosen anew at each realization. For two
338: dimensions, the results correspond to $10$ realizations on a $100
339: \times 100$-array. Though the general trend of $p_s(\alpha)$ is
340: qualitatively the same for the three cases, some systematic differences
341: are apparent. For instance, the survival probability in one dimension
342: is appreciably lower than in the other geometries for low
343: productivities. In this range, remarkably, the data for random graphs
344: lies between those for one and two dimensions. For $\alpha >0.8$,
345: instead, the values of $p_s(\alpha)$ are hardly distinguishable within
346: our numerical precision.
347:
348:
349: We may summarize our main results as follows. First, the strong,
350: deterministic correlation between survival and maximal productivity
351: that characterizes the model with global competition (\ref{model}) is
352: replaced, in the model with local competition (\ref{local}), by a much
353: weaker, probabilistic-like correlation. Agents with high productivity
354: are more likely to survive but, even with the maximal productivity,
355: extinction cannot be completely discarded. On the other hand, agents
356: with very low productivity have a chance of survival. This conclusion
357: should be relevant to the possible applications of the present models,
358: both to economy and to biology. Second, the loss of the above
359: mentioned deterministic correlation is accompanied by the formation of
360: an almost-regular spatial structure, with a periodic alternation of
361: survivors and extinct agents. This structure is explained by the
362: emergence of strong spatial correlations out of the fully uncorrelated
363: distribution of productivities, due to the very action of the
364: evolution rules. Nontrivial correlations in the conditional
365: probabilities for certain combinations of uncorrelated variables---a
366: rather simple but counterintuitive phenomenon---have recently been
367: pointed out for series of random numbers \cite{Sornette}. The present
368: model illustrates the occurrence of the same kind of phenomenon in a
369: spatially extended dynamical system.
370:
371: This work has been partially carried out at the Abdus Salam
372: International Centre for Theoretical Physics (Trieste, Italy). The
373: author thanks the Centre for hospitality, and G. Abramson for
374: useful comments.
375:
376: \begin{thebibliography}{99}
377:
378: \bibitem{Murray} J. D. Murray, {\it Mathematical Biology}
379: (Springer, Berlin, 1989).
380:
381:
382: \bibitem{Bak} P. Bak, S. F. Norrellykke, and M. Shubik, Phys. Rev. E
383: {\bf 60}, 2528 (1999).
384:
385:
386: \bibitem{wta} J.-P. Bouchaud and M. M\'ezard, Physica A {\bf 282},
387: 536 (2000).
388:
389: \bibitem{global} See, for instance, K. Kaneko, Physica D {\bf 41},
390: 137 (1990); N. Nakagawa and Y. Kuramoto, Physica D {\bf 75}, 74
391: (1994).
392:
393: \bibitem{theor} G. F. Gause, {\it The Struggle for Existence}
394: (Williams \& Wilkins, Baltimore, 1934).
395:
396:
397: \bibitem{Pareto} B. B. Mandelbrot, {\it Fractals and Scaling in
398: Finance} (Springer, New York, 1997).
399:
400:
401: \bibitem{Sornette} D. Sornette and J. V. Andersen, cond-mat/0001324.
402:
403:
404: \end{thebibliography}
405:
406:
407:
408:
409:
410:
411: \end{document}
412:
413:
414:
415:
416:
417:
418:
419: