1: \documentclass[figures]{epl}
2: \usepackage{amsmath}
3: \usepackage{amssymb}
4:
5: \title{Synchronization and frustration in oscillator
6: networks with attractive and repulsive interactions}
7: \shorttitle{Synchronization and frustration in oscillator
8: networks}
9: \author{Dami\'an H. Zanette}
10: \shortauthor{D. H. Zanette} \institute{Consejo Nacional de
11: Investigaciones Cient\'{\i}ficas y T\'ecnicas, Centro At\'omico
12: Bariloche and Instituto Balseiro, 8400 Bariloche, R\'{\i}o Negro,
13: Argentina }
14:
15: \pacs{05.65.+b}{Self-organized systems}
16: \pacs{05.45.Xt}{Synchronization; coupled oscillators}
17:
18: \begin{document}
19:
20: \maketitle
21:
22: \begin{abstract}
23: We study, numerically and analytically, the stability of
24: synchronization for an ensemble of coupled phase oscillators with
25: attractive and repulsive interactions, as a function of the number
26: of repulsive couplings and their intensity. Scaling properties of
27: the desynchronization transition are disclosed, and it is shown
28: that unsynchronized configurations have different symmetries
29: depending on the intensity of the repulsive interaction. The
30: concept of frustration minimization helps understanding the main
31: features observed in the collective dynamics of the oscillator
32: network.
33: \end{abstract}
34:
35: Synchronization is a well-known form of collective dynamics in
36: large ensembles of interacting dynamical systems. Under the action
37: of attractive coupling, strong correlations develop in time
38: between the states of individual elements, and the ensemble is
39: able to generate signals of macroscopic amplitude. This coherent
40: behaviour is found in a broad class of natural systems
41: --especially, in the realm of life sciences \cite{Win}-- and has
42: been reproduced by means of a variety of mathematical models
43: \cite{Kura,pik,nos}.
44:
45: The strongest manifestation of synchronized dynamics, which can be
46: realized in an ensemble of identical periodic oscillators subject
47: to attractive coupling, is full synchronization. In a fully
48: synchronized ensemble, the individual motions of all the
49: oscillators coincide. Heterogeneities, chaotic individual
50: dynamics, and repulsive interaction, on the other hand, may
51: inhibit the development of synchronization \cite{nos}.
52:
53: In this Letter, we explore the competing effect of attractive and
54: repulsive interaction in oscillator ensembles. Specifically, we
55: study the influence of repulsive couplings on the stability of
56: full synchronization. The combined action of attractive and
57: repulsive couplings has already been considered by Daido
58: \cite{d1,d2,d3}, who disclosed the possibility of weaker forms of
59: synchronized dynamics, glass-like behaviour, and anomalous
60: relaxation. These results, however, typically apply to cases where
61: the number of attractive and repulsive couplings and their
62: relative intensity are equal on the average. Here, instead, we are
63: interested at the effect of repulsive couplings as their abundance
64: and intensity are varied gradually. As discussed below, this leads
65: naturally to the consideration of heterogeneous random networks of
66: coupled oscillators.
67:
68: We consider an ensemble of $N$ identical oscillators, whose
69: individual states are given by the phases $\phi_i \in [0,2\pi)$
70: ($i=1,\dots,N$). In the absence of interaction, the evolution of
71: the phases is given by $\dot \phi=\omega$ for all $i$. Coupling is
72: introduced following Kuramoto's scheme \cite{Kura},
73: \begin{equation} \label{osc}
74: \dot \phi_i = \omega + \frac {1}{N} \sum_{j=1}^N W_{ij} \sin
75: (\phi_j-\phi_i),
76: \end{equation}
77: where $W_{ij}$ weights the interaction between oscillators $i$ and
78: $j$. By transforming $\phi_i \to \phi_i + \omega t$ for all $i$,
79: we fix, without generality loss, $\omega=0$. In the following, we
80: study the case of symmetric coupling, $W_{ij}=W_{ji}$. For each
81: realization of the ensemble, the interaction weights are chosen
82: at random as
83: \begin{equation}
84: W_{ij} = \left\{
85: \begin{array}{rl}
86: 1 & \mbox{with probability $1-p$} \\ -w & \mbox{with probability
87: $p$}
88: \end{array}
89: \right.
90: \end{equation}
91: with $w>0$. Thus, on the average, a fraction $p$ of the $N(N-1)/2$
92: couplings correspond to repulsive interaction, while the remaining
93: couplings are attractive. The relative intensity of the repulsive
94: interaction is $w$. Equations (\ref{osc}) can be rewritten as
95: \begin{equation} \label{osc1}
96: \dot \phi_i=\frac{1}{N} \sum_{j=1}^N \sin (\phi_j-\phi_i) -
97: \frac{1+w}{N}\sum_{j\in {\cal R}_i} \sin (\phi_j-\phi_i).
98: \end{equation}
99: In this representation, the first term in the r.h.s.~of the
100: equation corresponds to global attractive coupling, where all
101: oscillator pairs interact with the same intensity. The second term
102: corresponds to repulsive interaction of relative intensity $1+w$.
103: Repulsive couplings are restricted to a random network determined
104: by the sets ${\cal R}_i$. An oscillator $j$ belongs to ${\cal
105: R}_i$ if it interacts repulsively with oscillator $i$. Calling
106: $z_i$ the number of oscillators in ${\cal R}_i$, the average
107: connectivity of the network of repulsive couplings is $\langle z_i
108: \rangle = p(N-1)$.
109:
110: Our aim here is to analyze the long-time solutions of Eqs.
111: (\ref{osc1}) as the parameters $p$ and $w$ are varied, averaging
112: over different realizations of the network of repulsive couplings.
113: Specifically, we are interested at the destabilization of full
114: synchronization as the number and intensity of repulsive
115: couplings grow. The state of full synchronization,
116: $\phi_i(t)=\phi^*$ for all $i$ and constant $\phi^*$, is in fact a
117: solution to Eqs. (\ref{osc1}) for any $w$ and arbitrary ${\cal
118: R}_i$. Its stability, however, is expected to depend on the
119: intensity of the repulsive interaction and on the network
120: topology. For $p=0$, which corresponds to the case of pure global
121: attractive coupling, repulsive interaction is absent and full
122: synchronization is stable. For $p=1$, coupling is also global but
123: purely repulsive; the sets ${\cal R}_i$ are extended to the whole
124: ensemble. In this situation, full synchronization is unstable and,
125: for asymptotically large times, the system reaches one of the
126: infinitely many states given by the equations $\sum_j \sin
127: (\phi_j-\phi_i)=0$ ($i=1,\dots,N$). In such states, typically, the
128: phases $\phi_i$ are more or less uniformly distributed over
129: $[0,2\pi)$ \cite{nos}. Destabilization of the fully synchronized
130: state is thus expected to occur at intermediate values of $p$.
131:
132: Linear stability of full synchronization can be assessed by
133: studying the eigenvalue spectrum of the $N\times N$ matrix with
134: elements
135: \begin{equation} \label{Sij}
136: S_{ij}=\frac{1}{N}+ \left( -1+\frac{1+w}{N} z_i \right)
137: \delta_{ij} - \frac{1+w}{N} R_{ij},
138: \end{equation}
139: where $\delta_{ij}$ are the elements of the identity matrix, and
140: $R_{ij}$ are the elements of the (symmetric) adjacency matrix
141: corresponding to the network of repulsive couplings: $R_{ij}=1$ if
142: $j\in {\cal R}_i$, and $0$ otherwise. The state of full
143: synchronization is linearly stable if {\it all} the eigenvalues of
144: $S_{ij}$ are negative.
145:
146: \begin{figure}
147: \onefigure[width=8cm]{fig1.eps} \caption{Data collapse of the
148: fraction $f$ of realizations with unstable full synchronization,
149: as a function of the fraction of repulsive couplings $p$, for
150: different system sizes $N$. The intensity of the repulsive
151: interaction is $w=1$. Inset: the transition point $p_T$ and the
152: transition width $\delta$ as functions of $N$. The analytical
153: result for the transition point, $p_A$, is also shown. Dotted
154: lines are spline interpolations, added for clarity.} \label{fig1}
155: \end{figure}
156:
157:
158: Using a standard numerical routine, we have calculated the
159: eigenvalues of $S_{ij}$ for different values of the parameters $p$
160: and $w$, and system sizes $N$ between $10$ and $10^3$. For each
161: parameter set, $10^2$ to $10^4$ realizations of the network of
162: repulsive couplings have been considered. As a measure of the
163: degree of instability of full synchronization, we have recorded
164: the fraction $f$ of realizations for which at least one of the
165: eigenvalues is positive --i.e., for which full synchronization is
166: unstable. For fixed $N$ and $w$, as expected, this fraction
167: changes from $f\approx 0$ for small $p$ to $f\approx 1$ for large
168: $p$. The transition takes place in a rather narrow range of $p$.
169: As illustrated in Fig. \ref{fig1} for $w=1$, numerical results for
170: $f$ as a function of $p$ and different system sizes admit a fairly
171: good collapse as a function of the transformed parameter
172: $(p-p_T)/\delta$. Here, $p_T$ has been chosen as the value of $p$
173: for which $f=1/2$, which we identify as the transition point. The
174: coefficient $\delta$, in turn, measures the width of the
175: transition range. For each size $N$, this coefficient is adjusted
176: to achieve the data collapse, taking as a reference --without
177: generality loss-- $\delta=1$ for $N=100$. The inset of Fig.
178: \ref{fig1} shows a plot of $p_T$ and $\delta$ as functions of $N$.
179: As $N$ grows, the transition point seems to approach an asymptotic
180: value $p_T\approx 0.5$. The width $\delta$, on the other hand,
181: decreases steadily, approximately following a power law
182: $N^{-0.48\pm 0.01}$. This indicates that the transition range
183: becomes narrower, and the transition itself becomes better
184: defined, as the system size is increased.
185:
186: An independent, analytical evaluation of the transition point can
187: be achieved under the hypothesis that the network of repulsive
188: couplings is regular, i.e. when all the sets ${\cal R}_i$ have
189: exactly the same number of elements, $z_i=z=p(N-1)$ for all $i$.
190: In these conditions, the eigenvalues of $S_{ij}$ in Eq.
191: (\ref{Sij}) are directly related to the eigenvalues $\rho_k$
192: ($k=1,\dots,N$) of $R_{ij}$. Consequently, the stability condition
193: for full synchronization can be given in terms of $\rho_k$. The
194: transition point turns out to be determined by the minimum
195: eigenvalue of $R_{ij}$, $\rho_{\min}= \min \{ \rho_k\}$. For
196: sufficiently large $N$, the typical value of this eigenvalue can
197: be estimated from the so-called semicircle law \cite{semic} which,
198: in the present framework, establishes that the spectral density of
199: $R_{ij}$ has the shape of a semicircle of radius $2\sqrt{Np(1-p)}$
200: centered at the origin. The instability threshold for full
201: synchronization is given by the condition \cite{fut}
202: \begin{equation} \label{anal}
203: p\left( 1-\frac{1}{N}\right)+ 2 \sqrt{\frac{p(1-p)}{N}}
204: =\frac{1}{1+w}.
205: \end{equation}
206: The inset of Fig. \ref{fig1} shows that the value $p_A$ obtained
207: from this equation for $w=1$, as a function of the system size, is
208: in very good qualitative agreement with $p_T$. Note that, for $N
209: \to \infty$, our analytical evaluation predicts a transition point
210: at $p_A=(1+w)^{-1}$; for $w=1$, we get $p_A=1/2$. Evaluation of
211: the spectral density of $R_{ij}$ beyond the semicircle law
212: \cite{Bronk, Farkas} shows that the fraction of eigenvalues
213: outside the interval $(-2\sqrt{Np(1-p)},2\sqrt{Np(1-p)})$
214: decreases with the system size as $N^{-1}$. This is consistent
215: with our observation that the transition becomes sharper as $N$
216: grows.
217:
218: \begin{figure}
219: \onefigure[width=8cm]{fig2.eps} \caption{Data collapse of the
220: fraction $f$ of realizations with unstable full synchronization,
221: as a function of the fraction of repulsive couplings $p$, for
222: different intensities $w$. The system size is $N=100$. Inset:
223: the numerical and the analytical transition points, $p_T$ and
224: $p_A$, and the transition width $\delta$, as functions of $w$.
225: Dotted lines are spline interpolations, added for clarity.}
226: \label{fig2}
227: \end{figure}
228:
229: Figure \ref{fig2} shows numerical results for $f$ in systems of
230: fixed size, $N=100$, for various values of $w$. As in Fig.
231: \ref{fig1}, the data collapse was achieved by plotting $f$ as a
232: function of $(p-p_T)/\delta$, where $p_T$ is the fraction of
233: repulsive couplings at which $f=1/2$. The inset shows $p_T$ and
234: $\delta$ as functions of $w$. Since for both large and small $w$
235: the transition range is narrower than in between, $\delta$ reaches
236: a maximum for an intermediate value of the intensity of the
237: repulsive interaction. The analytical threshold $p_A$, obtained
238: from Eq. (\ref{anal}), is again in very good qualitative agreement
239: with the numerical results.
240:
241: What is the nature of the stationary state of the oscillator
242: ensemble just beyond the destabilization of full synchronization?
243: To answer this question, we have inspected the long-time
244: distribution of phases for networks of $N=100$ oscillators, where
245: repulsive couplings were successively added. This has made it
246: possible to detect, for each network, the exact point at which
247: full synchronization becomes unstable, and to analyze the
248: resulting unsynchronized asymptotic state. It turns out that this
249: unsynchronized state changes qualitatively when the intensity of
250: the repulsive interaction grows. For small and moderate $w$,
251: destabilization of full synchronization gives place to a state
252: where oscillator phases are irregularly spread over a small
253: interval. Typically, the oscillator with maximal number of
254: repulsive couplings has a substantially different phase. As
255: expected, the distribution of phases widens as the fraction of
256: repulsive couplings $p$ increases. These features are illustrated
257: by the upper panels of Fig. \ref{fig3}, for $w=1$. In this
258: particular realization of the oscillator network, full
259: synchronization becomes unstable for $p=0.3527\dots$. The figure
260: shows snapshots of the phase distribution in the plane $(\cos
261: \phi, \sin \phi)$. Lines join oscillator pairs with repulsive
262: couplings.
263:
264: \begin{figure}
265: \onefigure[width=8cm]{fig3.eps} \caption{Long-time snapshots of
266: the distribution of oscillator phases for different values of the
267: intensity $w$ and fraction $p$ of repulsive couplings, in a system
268: of size $N=100$. Individual phases are represented by dots on the
269: unit circle (dotted line). Straight lines join oscillators which
270: interact repulsively.} \label{fig3}
271: \end{figure}
272:
273: The situation changes substantially if the intensity of the
274: repulsive interaction grows enough. For large $w$, when relatively
275: few repulsive couplings are needed to destabilize full
276: synchronization, unsynchronized states just beyond the transition
277: are characterized by a symmetric distribution of phases. Some of
278: the oscillator pairs with repulsive interaction abandon the
279: synchronized cluster and their two elements reach symmetric
280: positions at each side of the cluster. The remaining oscillators
281: stay synchronized at a common phase. As the fraction of repulsive
282: couplings keeps growing, the number of unsynchronized pairs
283: increases, still preserving their symmetric configuration.
284: Eventually, however, symmetry is broken, and the ensemble reaches
285: a more or less irregular phase distribution, similar to that
286: observed for lower $w$. The second line of panels in Fig.
287: \ref{fig3} illustrates these features for $w=30$. In this
288: realization of the network, the transition takes place at
289: $p=0.0048$, corresponding to just $24$ repulsive couplings. For
290: that value of $p$ and for $p=0.006$ we find symmetric
291: configurations of phases, while for $p=0.01$ the distribution is
292: already asymmetric. In the symmetric cases, the cluster of
293: remaining synchronized oscillators is the isolated dot in the
294: centre of the configuration.
295:
296: Our numerical findings on the destabilization of full
297: synchronization and the resulting unsynchronized states can be
298: understood, at least qualitatively, by resorting to a concept
299: borrowed from the study of disordered spin systems, namely, {\it
300: frustration} \cite{frustration,d2}. In spin systems, where
301: interactions are usually local, frustration is associated with the
302: impossibility that three mutually interacting spins with
303: interaction weights of different sign achieve simultaneously a
304: state of minimal energy. In our system, where (attractive or
305: repulsive) interaction affects every oscillator pair, frustration
306: must be defined for each pair in relation to the state of the
307: whole ensemble. For instance, an oscillator pair with repulsive
308: coupling is frustrated if the system is in a collective
309: synchronized state, so that the individual phases of the two
310: oscillators are forced to coincide. Conversely, a pair with
311: attractive coupling is frustrated if the interactions with the
312: rest of the ensemble inhibits the synchronization of the two
313: oscillators. Frustration is quantitatively characterized by the
314: function
315: \begin{equation} \label{Frust}
316: F=-\frac{1}{N}\sum_{i,j=1}^N W_{ij} \cos (\phi_j-\phi_i).
317: \end{equation}
318: Pairs with attractive coupling, $W_{ij}>0$, contribute to decrease
319: the frustration $F$ if $\phi_j\approx \phi_i$. {\it Vice versa},
320: for $W_{ij}<0$, $F$ decreases as the difference of the two phases
321: grows. The relevance of the frustration $F$ with respect to the
322: oscillator dynamics resides in the fact that it plays the role of
323: a nonequilibrium potential. Specifically, for $\omega=0$, Eqs.
324: (\ref{osc}) can be written as $\dot \phi_i =-\partial F/\partial
325: \phi_i$. Therefore,
326: \begin{equation}
327: \dot F = \sum_{i=1}^n \frac{\partial F}{\partial \phi_i} \dot
328: \phi_i= -\sum_{i=1}^n \left( \frac{\partial F}{\partial \phi_i}
329: \right)^2 \le 0 .
330: \end{equation}
331: In other words, the dynamics convey a steady decrease of the
332: frustration.
333:
334: Bearing these considerations in mind, we can argue as follows. As
335: the fraction of repulsive couplings grows from $p=0$, full
336: synchronization is initially stable. Frustration, nevertheless,
337: increases because of the increasing number of oscillators with
338: repulsive interactions that are forced to stay into the
339: synchronized cluster. The collective state of the ensemble is thus
340: able to bear a certain degree of frustration, in order to maintain
341: synchronization. Eventually, however, frustration reaches too high
342: levels, and synchronization is no longer a ``convenient''
343: collective configuration --though it still is a possible
344: stationary state for the ensemble. Consequently, the synchronized
345: cluster breaks down, the individual oscillator phases spread out,
346: and the level of frustration is alleviated. Since, now,
347: oscillators are not locked to the synchronized cluster, further
348: increase of $p$ will lead to the gradual widening of the
349: distribution of phases, in the search for the state of minimal
350: frustration.
351:
352: When the repulsive interaction is strong enough, just a few
353: repulsive couplings suffice to make full synchronization
354: ``inconvenient'' from the viewpoint of frustration, and the
355: synchronized cluster may break down for very small values of $p$.
356: Under these conditions, the network of repulsive couplings is very
357: simple: it typically consists of isolated oscillator pairs. If
358: full synchronization is unstable, the two oscillators of each pair
359: abandon the synchronized cluster, trying to maximize their mutual
360: distance. At the same time, both of them tend to minimize the
361: distance to the cluster, since their interaction with the
362: oscillators there is attractive. The resulting configuration is
363: thus symmetric around the cluster, as illustrated in Fig.
364: \ref{fig3} for $w=30$.
365:
366: In summary, we have studied the desynchronization transition in a
367: network of coupled identical phase oscillators with attractive and
368: repulsive interactions, as the number and intensity of repulsive
369: couplings is increased. For finite system sizes, the transition is
370: gradual: the probability that the state of full synchronization is
371: unstable for a given realization of the network of couplings grows
372: smoothly with the number of repulsive couplings. Numerical
373: finite-size analysis as well as analytical results, however,
374: suggest that the desynchronization transition would become abrupt
375: for infinitely large systems. Beyond the transition, the nature of
376: the unsynchronized long-time state depends sensibly on the
377: intensity of the repulsive interaction. For small and moderate
378: intensities, individual oscillator phases reach a more or less
379: disordered configuration, reminiscent of the disordered
380: equilibrium state of an oscillator ensemble with global repulsive
381: coupling \cite{nos}. On the other hand, for strong repulsive
382: interaction, desynchronization gives rise to highly symmetric
383: configurations, where the phase of unsynchronized oscillators are
384: situated at both sides of the remaining synchronized cluster. The
385: possibility of observing these symmetric configuration depends on
386: their existence and stability, which we analyze in detail
387: elsewhere \cite{fut}. The notion of frustration --already
388: addressed, in the frame of coupled oscillator ensembles, by Daido
389: \cite{d2}-- is useful to construct a variational-like formulation
390: for the dynamics of the system studied here. In this framework, we
391: have given a qualitative description of our main results. Such
392: formulation, however, should also play a role in a quantitatively
393: detailed analysis of the system. In fact, as defined in Eq.
394: (\ref{Frust}) and under appropriate conditions at its singular
395: points, frustration could be used as a Lyapunov function for the
396: dynamics.
397:
398: \begin{thebibliography}{0}
399:
400: \bibitem{Win}
401: \Name{Winfree A. T.} \Book{The Geometry of Biological Time}
402: \Publ{Springer, New York} \Year{2001}.
403:
404: \bibitem{Kura}
405: \Name{Kuramoto Y.} \Book{Chemical Oscillations, Waves, and
406: Turbulence} \Publ{Springer, Berlin} \Year{1984}.
407:
408: \bibitem{pik}
409: \Name{Pikovsky A., Rosemblum M. \and Kurths J.}
410: \Book{Synchronization. A Universal Concept in Non-linear
411: Sciences.} \Publ{Cambridge University Press, Cambridge}
412: \Year{2001}, and references therein.
413:
414: \bibitem{nos}
415: \Name{Manrubia S. C., Mikhailov A. S. \and Zanette D. H.}
416: \Book{Emergence of Dynamical Order. Synchronization Phenomena in
417: Complex Systems} \Publ{World Scientific, Singapore} \Year{2004},
418: and references therein.
419:
420: \bibitem{d1}\Name{Daido H.} \REVIEW{Prog. Theor. Phys.}{77}{1987}{622}.
421:
422: \bibitem{d2}\Name{Daido H.} \REVIEW{Phys. Rev. Lett.}{68}{1992}{1073}.
423:
424: \bibitem{d3}\Name{Daido H.} \REVIEW{Phys. Rev. E}{61}{2000}{2145}.
425:
426: \bibitem{semic}\Name{Mehta M. L.}\Book{Random Matrices} \Publ{Academic Press, New York}
427: \Year{1991}, 2nd edition.
428:
429: \bibitem{fut}\Name{Zanette D. H.} in preparation.
430:
431: \bibitem{Bronk}\Name{Bronk B. V.} \REVIEW{J. Math. Phys.}{5}{1964}{215}.
432:
433: \bibitem{Farkas}\Name{Farkas I. J., Der\'enyi I., Barab\'asi A.-L. \and Vicsek T.}
434: \REVIEW{Phys. Rev. E}{64}{2001}{026704}.
435:
436:
437: \bibitem{frustration}\Name{M\'ezard M., Parisi G. \and Virasoro M.
438: A.} \Book{Spin Glass Theory and Beyond} \Publ{World Scientific,
439: Singapore} \Year{1987}.
440:
441:
442:
443: \end{thebibliography}
444:
445: \end{document}
446:
447: The possibility of finding the oscillators in a symmetric
448: configuration depends on its existence and stability.
449:
450: \begin{figure}
451: \onefigure[width=8cm]{frus3.eps} \caption{Stability diagram in
452: parameter space, for the $3$-cluster unsynchronized state analyzed
453: in the text. The $3$-cluster state does not exist in the dark
454: zone. The full curve is the boundary between the regions of
455: instability and stability. The dotted curve corresponds to the
456: stability boundary for full synchronization evaluated from Eq.
457: (\ref{anal}).} \label{fig4}
458: \end{figure}
459:
460: %\acknowledgments Valuable discussions with G. Abramson, H. Kori,
461: %M. Kuperman, and A. S. Mikhailov and are gratefully acknowledged.
462: