nlin0509020/dp2.tex
1: %\documentclass[preprint,showpacs,preprintnumbers,amsmath,pre]{revtex4}
2: \documentclass[twocolumn,showpacs,preprintnumbers,amsmath,pre]{revtex4}
3: 
4: 
5: \usepackage{graphicx}% Include figure files
6: \usepackage{dcolumn}% Align table columns on decimal \varepsilonpoint
7: \usepackage{bm}% bold math
8: \usepackage{amssymb}
9: 
10: \begin{document}
11: 
12: \title{Thermodynamic limit of the 
13: first-order phase transition in the Kuramoto model}
14: \author{Diego Paz\'o}
15: \email{pazo@pks.mpg.de}
16: \affiliation{Max-Planck-Institut f{\"u}r Physik komplexer Systeme, 
17: N{\"o}thnitzer Stra{\ss}e 38, 01187 Dresden, Germany}
18: 
19: 
20: 
21: \date{\today}
22: 
23: \begin{abstract}
24: 
25: In the Kuramoto model, a uniform distribution of the natural
26: frequencies leads to a first-order (i.e., discontinuous)
27: phase transition from incoherence to synchronization, 
28: at the critical coupling parameter $K_c$.
29: We obtain the asymptotic dependence 
30: of the order parameter above criticality: 
31: $r-r_c \propto (K-K_c)^{2/3}$.
32: For a finite population, we demonstrate
33: that the population size $N$ may be included into
34: a self-consistency
35: equation relating $r$ and $K$
36: in the synchronized state. 
37: {We analyze the convergence to the thermodynamic limit 
38: of two alternative schemes to set the natural frequencies.
39: Other frequency distributions different
40: from the uniform one are also considered.}
41: 
42: 
43: 
44: \end{abstract}
45: \pacs{05.45.Xt}
46: \maketitle
47: \section{Introduction}
48: 
49: Synchronization is a universal phenomenon
50: that plays 
51: an important role in all natural sciences as
52: well as in technology \cite{PRK,Blekhman}. 
53: In particular, the synchronization 
54: of populations of globally coupled 
55: oscillators with distributed natural frequencies 
56: has been an object of study since very early times,
57: mainly in a biological context~\cite{Winfree}.
58: Later, it has found application in other areas, 
59: such as Josephson junctions~\cite{wiesenfeld}, 
60: nanomechanics~\cite{alex}, etc. 
61: When increasing the coupling parameter,
62: these systems undergo transitions 
63: from a totally incoherent state to 
64: a partially coherent state where part of the population
65: becomes entrained sharing the same frequency. 
66: Interestingly, there
67: are several analogies with the phase 
68: transitions in statistical mechanics~\cite{daido90}.
69: Thus, one may define an order parameter, that usually
70: grows continuously from zero (the incoherent state) 
71: when the coupling parameter exceeds a threshold value, 
72: analogously to a second-order phase transition.
73: Nonetheless, in some situations~\cite{inertia,inertiad,bonilla92},
74: mutual entrainment occurs in an abrupt way (a 
75: first-order phase transition). After an infinitesimal 
76: variation of the coupling strength 
77: a macroscopic (i.e.~order $N$) part of the population becomes 
78: synchronized. One may speculate that 
79: first-order phase 
80: transitions may be of 
81: interest for practical applications, if one pursues  
82: a system exhibiting an abrupt off/on switch.
83: 
84: The most simple example 
85: of a first-order phase transition
86: is found in the Kuramoto model~\cite{Kuramoto} when the
87: natural frequencies are uniformly distributed. In this
88: case, it is known that all the population becomes 
89: synchronized in a single step~\cite{vanhemmen,inertiad}.
90: We obtain here the asymptotic dependence of the order parameter
91: after criticality, which exhibits a critical exponent $2/3$. 
92: 
93: Still, one of the open problems in the Kuramoto model 
94: is to fully understand the finite-$N$ behavior. As we show below, 
95: the simplicity of the uniform frequency distribution allows
96: to  cope with finite-size effects in an original way,
97: providing analytic and numerical results. Some of 
98: these results apply to other frequency
99: distributions with compact support.
100: 
101: 
102: This paper is organized as follows.
103: In Sec.~\ref{km}, we present the Kuramoto model,
104: and show some numerical results that motivated
105: this work. In Sec.~\ref{cont} an infinite population
106: is considered, finding 
107: the asymptotic dependence of the order parameter
108: after criticality. 
109: In Sec.~\ref{discr}, we study 
110: finite ensembles, finding: a) an $N$-dependent formula for
111: the order parameter, b) different convergence rates to the
112: thermodynamic limit for alternative sampling schemes of the natural frequencies.
113: Section \ref{other} is devoted to analyzing
114: frequency distributions different from the uniform one.
115: Finally in Sec.~\ref{concl} 
116: the main conclusions of this work are summarized. 
117: 
118: \section{The Kuramoto Model}
119: \label{km}
120: The Kuramoto model is probably the most studied model of
121: synchronization in a population of oscillators 
122: with all-to-all coupling~\cite{rmp_kuramoto}. 
123: The state of each oscillator
124: is described only by a phase variable (this stems from
125: the the fact that, at small coupling,
126: only the phase of a self-sustained oscillator is affected by the 
127: interaction). The phase $\theta_j$ of each oscillator 
128: satisfies the following ODE:
129: \begin{equation}
130: \dot \theta_j = \omega_j + \frac{K}{N} \sum_{l=1}^N \sin( \theta_l- \theta_j)
131: \label{kuramoto_model}
132: \end{equation}
133: where $\omega_j$ are the natural frequencies, and $K$ 
134: is the parameter controlling the coupling strength. 
135: 
136: To quantify the state of synchronization,
137: Kuramoto proposed to use a complex-valued quantity (so-called
138: order parameter to emphasize the relation with phase transitions):
139: \begin{equation}
140: r e^{i\psi}= \frac{1}{N} \sum_{l=1}^{N} e^{i\theta_l} .
141: \label{op}
142: \end{equation}
143: It allows us to set the governing equation (\ref{kuramoto_model})
144: in the form:
145: \begin{equation}
146: \dot \theta_j = \omega_j + K r \sin( \psi - \theta_j) .
147: \label{mean_field}
148: \end{equation}
149: If the natural frequencies are
150: distributed (i.e.~$\omega_j \neq \omega_{j'}$), synchronization only appears above 
151: some coupling threshold.
152: Here, we consider the case of evenly spaced natural frequencies:
153: \begin{equation}
154: \omega_j=-\gamma+\frac{\gamma}{N}(2j-1)
155: \label{unif}
156: \end{equation}
157: which, like all symmetric frequency distributions 
158: can be assumed centered at zero
159: (by going into a rotating frame if necessary).
160: Throughout this paper, the numerical integration of the Kuramoto model 
161: [Eq.~(\ref{kuramoto_model})]
162: is carried out by means of a fourth-order Runge-Kutta method with time
163: step $\Delta t=0.1$.
164: 
165: Recently, Maistrenko {\em et al.}~\cite{maistrenko} 
166: studied the Kuramoto model with a small number of oscillators
167: and natural frequencies distributed uniformly (or close to that).
168: They found that the synchronized state
169: robustly splits into several clusters with 
170: different average frequencies.
171: It is shown there, and in Fig.~\ref{figN}, that for 
172: $N=3$ and $5$, the synchronized state splits directly into $N$ clusters.
173: But, for values of $N$ other
174: than 3 and 5 the scenario is not so simple.
175: As $N$ increases,
176: the number of splittings for going from one to $N$ clusters,
177: increases as well. Hereafter, we denote 
178: by $K_s$ the coupling at the frequency-splitting
179: from the synchronized state.
180: In congruence 
181: with the thermodynamic limit (see below), 
182: all the splittings [including the first one at $K=K_s(N)$] 
183: must accumulate at $K_c$ as $N\rightarrow \infty$.
184: $K_c$ is the 
185: abrupt transition point for an infinite population. 
186: 
187: 
188: 
189: \begin{figure}
190: \includegraphics[width=3.1in]{fig1.eps}
191: \caption[]{Average frequencies ($\bar\omega_j$)
192: as a function of the coupling 
193: for different population sizes.
194: The natural frequencies are taken according
195: to (\ref{unif}) with $\gamma=1/2$. 
196: By $K_s$ we denote the value of $K$
197: for the first splitting bifurcation. The
198: critical point $K_c$ in the thermodynamic
199: limit ($N \rightarrow \infty$) is located at 
200: the dotted line.} \label{figN}
201: \end{figure}
202: 
203: \section{Infinite population}
204: \label{cont}
205: 
206: In this section, we briefly study the 
207: Kuramoto model for a uniform frequency distribution.
208: Some of the formulas will be later compared
209: to those obtained for a finite population 
210: in Sec.~\ref{discr}. We also go one step further,
211: and deduce an explicit formula, with critical exponent 2/3, 
212: for the dependence of the order parameter after criticality.
213: 
214: In correspondence to the finite case, Eq.~(\ref{unif}), 
215: we consider a uniform density of the natural frequencies:
216: \begin{eqnarray}
217: g(\omega)=
218: \left\{
219: \begin{array}{cc}
220: \frac{1}{2\gamma} \mbox{ for } |\omega| \le \gamma \\
221: 0 \mbox{ for } |\omega| > \gamma
222: \end{array}
223: \right. 
224: \label{g}
225: \end{eqnarray}
226: 
227: We first note that due to the invariance
228: under global rotation, for stationary solutions,
229: we can set a vanishing phase
230: for the order parameter in Eq. (\ref{op}), $\psi=0$,
231: without lack of generality. 
232: Kuramoto's classical analysis   
233: gives the order parameter equation:
234: \begin{equation}
235: r=\left< e^{i\theta} \right> = \left< \cos \theta  \right> 
236: \equiv \int_{-\infty}^{\infty} \cos \theta(\omega) \mbox{ } 
237: g(\omega)\mbox{ } d\omega  \mbox{ }.
238: \label{rcont}
239: \end{equation} 
240: 
241: In the totally locked regime, we obtain then:
242: \begin{eqnarray}
243: r&=& \int_{-\gamma}^\gamma g(\omega)
244: \sqrt{1-\frac{\omega^2}{K^2r^2}} d\omega  \mbox{ } \Rightarrow \label{rintegral} \\
245: r&=& \frac{1}{2} \sqrt{1-\frac{\gamma^2}{K^2r^2}}+\frac{Kr}{2\gamma}
246: \arcsin\left(\frac{\gamma}{Kr}\right) \mbox{ }.
247: \label{solvnum}
248: \end{eqnarray}
249: 
250: Eq.~(\ref{solvnum}) gives implicitely the dependence of $r$ on $K$.
251: A solution exists only for $Kr \ge \gamma$. 
252: At the critical point $K_cr_c=\gamma$,
253: the locked solution disappears with $r_c=\pi/4$. The corresponding value
254: of the  coupling is $K_c= 4\gamma/\pi$, that is precisely
255: the value where the incoherent solution $r=0$ becomes unstable
256: according to the classical result~\cite{strogatz91} for all unimodal 
257: distributions:
258: $K_c=2/(\pi g(0))$. Two remarks are in order. First, in contrast with
259: strictly unimodal distributions, the
260: transition is of first-order type: the order parameter ``jumps"
261: from zero to $r_c$. Second, at $K_c$ {\em all} the population becomes
262: entrained. This last
263: remark is quite important because it simplifies both numerical
264: and theoretical analyses. Also, note that when synchronized,
265: the oscillators' phases are spanned
266: along an interval (of length $\pi$ at $K=K_c$). When $K$
267: is increased $r$ grows from $r_c$ to $1$ in the 
268: $K\rightarrow \infty$ limit.
269: 
270: \begin{figure}
271: \includegraphics[width=3.1in]{fig2.eps}
272: \caption[]{Log-log dependence of $r$ on $K$ in a neighborhood
273: of criticality for $\gamma=1/2$. The circles correspond to
274: numerical solutions of the self-consistency condition (\ref{solvnum}), 
275: the solid line depicts Eq.~(\ref{2/3}), and the dotted straight line
276: arises taking the leading $\delta K^{2/3}$ term only.} \label{scaling}
277: \end{figure}
278: 
279: The first result of this paper is the dependence of $r$
280: on $K$, just above the phase transition. 
281: First of all, we make a change of variables onto Eq.~(\ref{rcont})
282: as usual (see e.g.~\cite{PRK,strogatz2000}):
283: \begin{equation}
284: r=Kr  
285: \int_{\theta_{min}}^{\theta_{max}} 
286: \cos^2\theta \mbox{ }
287: g(Kr\sin\theta) \mbox{ } d\theta .
288: \end{equation}
289: where $\theta_{max}$ ($\theta_{min}$) is the phase of the oscillator
290: with frequency $\gamma$ ($-\gamma$).
291: After an 
292: expansion above criticality:  
293: \begin{eqnarray}
294: K&=&K_c + \delta K \\
295: r&=&r_c + \delta r \\
296: \theta_{max} &=& -\theta_{min}=\pi/2 - \delta \theta .
297: \end{eqnarray}
298: and discarding the trivial incoherent solution $r=0$,
299: we get:
300: \begin{equation}
301: 1=
302: \frac{K_c+ \delta K}{2\gamma} \left[ \frac{\pi}{2} -\delta \theta
303: +\frac{1}{2}\sin(\pi-2\delta\theta)\right]
304: \end{equation}
305: An expansion of the sine function up to the cubic term yields:
306: \begin{equation}
307: 0=\frac{\pi}{2}\delta K - \frac{8\gamma}{3 \pi} \delta\theta^3
308: \label{dkdt}
309: \end{equation}
310: %
311: %
312: Thus the problem reduces to 
313: finding $\delta \theta$, 
314: from Eq.~(\ref{mean_field}):
315: \begin{eqnarray}
316: \gamma&=&(K_c+\delta K) (r_c + \delta r) \sin(\pi/2 -\delta \theta) \\
317: &\Rightarrow& \delta \theta \approx 
318:  \sqrt{\frac{8}{\pi}\delta r + \frac{\pi}{2\gamma}\delta K}
319: \label{dteta}
320: \end{eqnarray}
321: Introducing this 
322: expression into Eq.~(\ref{dkdt}), 
323: a formula for the asymptotic dependence of 
324: $\delta r$ on $\delta K$ is obtained:
325: \begin{equation}
326: \delta r = \left[\frac{9\pi^7}{2^{17} \gamma^2}\right]^{1/3}\delta K^{2/3} -
327: \frac{\pi^2}{16\gamma}\delta K 
328: \label{2/3}
329: \end{equation}
330: The result is compared in
331: Fig.~\ref{scaling} to the exact solution, arising
332: from numerically solving Eq.~(\ref{solvnum}). It confirms that the order parameter
333: grows from $r_c$ with a power of $K-K_c$ with exponent 2/3. 
334: To our knowledge, this exponent and
335: expression (\ref{2/3}) 
336: have not been reported before.
337: 
338: \section{Finite population}
339: \label{discr}
340: 
341: 
342: Finite-size effects in the Kuramoto model have been previously 
343: considered in the literature. Among the different 
344: approaches, we may list: the  investigation of the divergence 
345: of fluctuations around criticality~\cite{daido87,daido90},
346: the observation of ephemeral coherent structures
347: in the incoherent state~\cite{balmforth2000}, 
348: and the  reduction to a normal form in
349: the case of identical natural frequencies with 
350: additive noise~\cite{pikarufo}. Also, very recently,
351: Mirollo and Strogatz~\cite{mirollo2005} have analyzed 
352: the (local) stability of the fully locked state for a finite
353: population, finding that the locked solution is 
354: stable and disappears
355: in a saddle-node bifurcation~\footnote{In the presence of 
356: an irrelevant degenerate eigenvalue at zero due to global
357: phase-shift invariance.} 
358: (as already observed in~\cite{maistrenko} for small $N$).
359: 
360: In this section, we show that for
361: the uniform frequency distribution finite-size effects 
362: ---on the order parameter and on the loss of the synchronization---
363: can be studied in a novel way. 
364: 
365: 
366: 
367: \subsection{Dependence of the order parameter on $K$}
368: 
369: For a finite population in the synchronized state ($K\ge K_s$), 
370: the order parameter is expressed [in correspondence with the integral
371: form (\ref{rintegral})] by:
372: \begin{equation}
373: r=\frac{1}{N} \sum_{j=1}^N \sqrt{1-\frac{\omega_j^2}{K^2r^2}}
374: \label{riemann}
375: \end{equation}
376: where the order parameter $r$
377: is a time independent quantity. 
378: We devote the following 
379: lines to deducing an $N$-dependent self-consistency equation 
380: [that reduces to (\ref{solvnum}) in the $N\rightarrow\infty$ limit]. 
381: It is accurate provided
382: that $K$ is not too close to $K_s$.    
383: 
384: According to (\ref{unif}), the natural frequencies of the 
385: finite population are taken with step $\Delta \omega =\omega_{j+1}-\omega_j=2\gamma/N$. 
386: Therefore, the order parameter (\ref{riemann})
387: is equivalent to a Riemann sum with 
388: constant step of the integral found
389: in the continuum limit (\ref{rintegral}). The discrepancy
390: between both cases is, at the leading order,
391: dependent on $f''$  (where $f(\omega) \equiv \sqrt{1-(\omega/Kr)^2}$). 
392: Hence, for one Riemann box centered at $\omega_j$: 
393: \begin{equation}
394: \int_{\Delta\omega}f(\omega) d\omega = f(\omega_j) \Delta\omega +
395: \frac{f^{''}(\omega_j)}{24}  \Delta \omega^3 + {\cal O}(\Delta\omega^5) .
396: \label{serie1}
397: \end{equation}
398: %
399: Therefore, for the finite $N$ case we may approximate (\ref{riemann}) by:
400: \begin{equation}
401: r \approx  \frac{1}{2\gamma}\int_{-\gamma}^\gamma \sqrt{1-\frac{\omega^2}{K^2r^2}} d\omega
402: -\frac{\gamma^2}{6N^3}\sum_{j=1}^N f''(\omega_j)
403: \end{equation}
404: %
405: In this expression, the sum may be approximated
406: by its corresponding integral 
407: %(recall $l$ is even):
408: \begin{eqnarray}
409: \sum_{j=1}^N f''(\omega_j)&=&
410: \frac{N}{2\gamma}\left[ \int_{-\gamma}^\gamma f''(\omega) d\omega
411: + {\cal O}(N^{-2}) \right] \nonumber \\ &\approx& \frac{N}{2\gamma} 
412: \left. f'(\omega) \right|_{-\gamma}^\gamma
413: =-\frac{N}{\gamma Kr\sqrt{\frac{K^2r^2}{\gamma^2}-1}}
414: \label{serie2}
415: \end{eqnarray}
416: %
417: And we obtain a self-consistency equation for a
418: finite population of $N$ oscillators:
419: \begin{equation}
420: r= \frac{1}{2} \sqrt{1-\frac{\gamma^2}{K^2r^2}}+
421: \frac{Kr}{2\gamma}\arcsin\left(\frac{\gamma}{Kr}\right)
422: + \frac{\gamma N^{-2}}{6Kr\sqrt{\frac{K^2r^2}{\gamma^2}-1}}
423: \label{solvnumn}
424: \end{equation}
425: With respect to  the equation for the thermodynamic 
426: limit [Eq.~(\ref{solvnum})], there is an additional term 
427: depending on $N$. 
428: Plots of the numerical solutions for $N$ = 10 and 100 
429: are shown in Fig.~\ref{selfcheck}. 
430: Eq.~(\ref{solvnumn})
431: is deduced from the continuum equation, and therefore 
432: cannot intersect the line $r=\gamma / K$, because several
433: terms explode. Note that adding more terms of the series
434: in Eqs.~(\ref{serie1},\ref{serie2}) does not 
435: overcome this problem. This suggests that, unfortunately, 
436: the behavior very close to $K_s$ cannot be deduced
437: by simply manipulating the equations for an infinite population.
438: Nonetheless, out of that region,
439: the solution of (\ref{solvnumn}) reproduces
440: the numerical results, even for such a relatively
441: small number of oscillators as $N=10$, Fig.~\ref{selfcheck}(a).
442: 
443: 
444: 
445: \begin{figure}
446: \includegraphics[width=3.1in]{fig3.eps}
447: \caption[]{Dependence of $r$ on $K$ for finite populations:  
448: (a) $N=10$, and (b) $N=100$; $\gamma=1/2$ in both cases. Values obtained from
449: a direct computation of the Kuramoto model are marked by $\times$.
450: The first frequency splitting is observed at $K_s$
451: (for $K<K_s$, $r$ is no longer constant in time). 
452: The solid line is obtained solving 
453: Eq.~(\ref{solvnumn}) numerically. It matches the 
454: computed values, but fails when approaching the line $r= \gamma /K$. 
455: As a reference, the solution for an infinite population is shown with 
456: circles [after numerically solving Eq.~(\ref{solvnum})]; 
457: the critical point $(K_c,r_c)$ is marked
458: with a $\bullet$ symbol.} \label{selfcheck}
459: \end{figure}
460: 
461: 
462: 
463: \subsection{Thermodynamic limit of the first frequency-splitting}
464: 
465: 
466: The 
467: arrangement of the natural frequencies 
468: in Eq.~(\ref{unif}) converges to
469: the uniform frequency distribution (\ref{g}). 
470: But, as explained above, this limit is non-trivial:
471: a first-order phase transition is substituted, when $N$ becomes
472: finite,
473: by a set of frequency-splitting bifurcations
474: accumulating at $K_c$. Numerically, the study of these
475: bifurcations is quite involved. Nonetheless, the
476: point where, as $K$ decreases, 
477: the first splitting occurs ($K=K_s$), can be accurately computed.
478: This is possible because above $K_s$ the system is in a fixed point state.
479: Our simulations, Fig.~\ref{asymp}, indicate
480: that for the arrangement in (\ref{unif}) $K_s$ converges to ${K_c}^-$ according to a power law: 
481: \begin{equation}
482: K_c-K_s(N) \propto N^{-\mu} .
483: \label{powerlaw}
484: \end{equation} 
485: Note that $K_s$ and $K_c$ are both
486: proportional to $\gamma$, so $\mu \approx 1.5$ is independent of $\gamma$.
487: 
488: 
489: \begin{figure}
490: \includegraphics[width=3.1in]{fig4.eps}
491: \caption[]{Log-log dependence of the distance from $K_s$ 
492: to $K_c$ on the population
493: size ($\gamma=1/2$). Squares and diamonds correspond
494: to  different arrangements of the natural frequencies, 
495: Eqs.~(\ref{unif}) and (\ref{unif2}), respectively.
496: In the first case, the last two decades were 
497: fitted with a straight line, yielding a slope $-\mu=-1.502$.
498: The dashed straight line $\log_{10}(4\gamma/\pi)-x$ 
499: arises from Eq.~(\ref{-1}).} \label{asymp}
500: \end{figure}
501: 
502: The recipe followed to mimic the thermodynamic limit 
503: was to divide the frequency distribution $g(\omega)$ 
504: in $N$ parts of equal area taking each $\omega_j$ 
505: at the center of 
506: each block. 
507: %$\int_{\omega_j}^{\omega_{j+1}} g(\omega) d\omega= 
508: %2 \int_{-\infty}^{\omega_1} g(\omega) d\omega =1/N$. 
509: %In this way, arrangement in Eq.~(\ref{unif})
510: %stems from the uniform frequency distribution (\ref{g}).
511: 
512: Nonetheless, there are many (infinite in fact) possible 
513: discrete arrangements with the same continuum limit, 
514: %but that in (\ref{unif}) appears to be optimal. 
515:  but for arrangement (\ref{unif}) $\mu$ is large
516: enough to deduce the decay for other distributions, 
517: by just considering the variation of the effective $\gamma$.
518: For instance, if the natural frequencies
519: are taken as in \cite{matthews91,inertiad}:
520: \begin{equation}
521: \omega_j=-\gamma+\frac{\gamma}{N-1}2(j-1)
522: \label{unif2}
523: \end{equation}
524: one observes that $K_s$ converges
525: to ${K_c}^+$  with a power-law, see $\diamond$'s in Fig.~\ref{asymp}, 
526: but more slowly than 
527: for the arrangement in (\ref{unif}).
528: As the extrema of (\ref{unif2}) are fixed,
529: a change in $N$ varies the effective width of 
530: the equivalent continuous distribution: 
531: $\gamma_{eff}=\gamma + \gamma/(N-1)$.
532: Hence, from (\ref{powerlaw}) we get:
533: \begin{equation}
534: K_s(N)=\frac{4\gamma_{eff}}{\pi}-\alpha N^{-\mu}
535: %=\frac{4}{\pi}\left(\gamma+\frac{\gamma}{N-1} \right)
536: %=K_c+\frac{4\gamma}{\pi(N-1)}+{\cal O}(N^{-\mu})
537: \approx K_c + \frac{4 \gamma}{\pi N} +{\cal O}(N^{-\mu})
538: \label{-1}
539: \end{equation}
540: which agrees with the observed result 
541: (see the dashed line in Fig.~\ref{asymp}).
542: 
543: \section{Other frequency distributions}
544: \label{other}
545: 
546: In this section, we briefly discuss the extension of the previous 
547: results to other 
548: frequency distributions supported on a finite interval $[-\gamma,\gamma]$.
549: We first note that for non-uniform 
550: distributions the lost of complete synchronization and 
551: the critical point where the incoherent solution becomes unstable 
552: do not coincide: 
553: $K_s^\infty \equiv K_s (N\rightarrow \infty) > K_c$. In other words,
554: there exists an intermediate range of partial entrainment in which
555: one part of the population is synchronized whereas the remaining oscillators
556: drift.
557: 
558: As model distributions we considered 
559: %the 
560: three  unimodal distributions listed in Table~\ref{tabla}.
561: As for the uniform distribution, the desynchronization point $K_s^\infty$ 
562: may be computed analytically using Eq.~(\ref{rintegral}).  
563: 
564: \begin{table}
565: \centering
566: \begin{tabular}{l  c  c} \hline\hline
567:  & $g(\omega)$ & $K_s^\infty$  \\ \hline
568: Parabolic &  $\frac{3}{4 \gamma^3}(\gamma^2-\omega^2)$  & $\frac{32\gamma}{9 \pi}$ \\ \hline
569: Triangular & $\frac{\gamma - |\omega|}{\gamma^2}$ & $\frac{6 \gamma}{3\pi - 4}$ \\ \hline 
570: Hat-shaped & 
571: $ \begin{array}{l} \frac{2}{3\gamma} \quad (|\omega| \leq \frac{\gamma}{2})\\ 
572: \frac{1}{3\gamma} \, (\frac{\gamma}{2} < |\omega| \le \gamma) \end{array}$&  
573: $\frac{36 \gamma}{ 8\pi +3\sqrt{3}}$ \\ \hline\hline 
574: \end{tabular}
575: \caption{Three frequency distributions considered in 
576: Sec.~\ref{other}.} 
577: \label{tabla}
578: \end{table}
579: 
580: We focused on two simple sampling schemes
581: to discretize distributions supported on a bounded interval:
582: \begin{enumerate}
583: 
584: \item[(i)] $\int_{\tilde\omega_j}^{\tilde\omega_{j+1}} g(\omega) d\omega= 2\gamma /N ,\tilde\omega_1=-\gamma, 
585: \omega_j=\frac{\tilde\omega_j+\tilde\omega_{j+1}}{2}$.
586: \item[(ii)] $\int_{\omega_j}^{\omega_{j+1}} g(\omega) d\omega= 2\gamma /(N-1) , \omega_N=-\omega_1=\gamma$.
587: \end{enumerate} 
588: Applying (i) and (ii) to a uniform distribution one gets the 
589: arrangements in Eqs.~(\ref{unif}) and (\ref{unif2}), respectively.
590: 
591: For both schemes and the three frequency distributions
592: in Table~\ref{tabla}, 
593: the approach of the first frequency splitting
594: to the thermodynamic limit satisfies 
595: a power law: $|K_s^\infty-K_s(N)|\propto N^{-\mu}$,
596: as occurred for the uniform distribution 
597: ($K_s^\infty=K_c$ in this case).
598: For sampling (ii) we find that the value of the exponent is
599: always $\mu \approx 1$. However for (i), different exponents
600: arise, in contrast to $\mu \approx 1.5$ obtained for the uniform distribution:
601: $\mu\approx 0.5$ for triangular and parabolic distributions,
602: and $\mu\approx 1$ for the hat-shaped one. We have checked that this exponent
603: arises for other 
604: distributions with an abrupt boundary ($g(\pm \gamma)>0$). 
605: 
606: Another interesting power law is the shift
607: of the order parameter in the synchronized state:
608: $|r(K,N)-r(K,N=\infty)| \sim  N^{-\nu}$.
609: For sampling scheme (i) 
610: theoretical results may be obtained, using again arguments based
611: on the Riemann sum.
612: If $g(\pm \gamma) > 0$  ---e.g.~uniform [see Eq.~(\ref{solvnumn})] 
613: or hat-shaped distributions--- one may obtain 
614: a formal solution \footnote{$r= \int_{-\gamma}^\gamma g \, f \,  d\omega 
615: - \frac{N^{-2}}{24} \int_{-\gamma}^\gamma \frac{g  f'' \:   
616: + 2 g' f'}{g^2} d\omega$.} that yields
617: $\nu =2$. 
618: However for distributions that approach zero 
619: at $\omega=\pm\gamma$, the ``Riemann-sum approach'' is not
620: valid due to divergences at $\omega=\pm \gamma$. One must,
621: therefore, analyze these points separately. In particular,
622: one obtains $\nu=3/2$ for linearly decaying
623: distributions [$g(\omega \rightarrow \pm \gamma) \sim (\gamma \mp \omega)$, 
624: e.g. parabolic and triangular distributions].  Also, our simulations
625: indicate that for sampling (ii) $\nu \approx 1$, irrespective of the 
626: frequency distribution (for the uniform distribution $\nu=1$ is straightforward
627: due to the ${\cal O}(N^{-1})$ effective shift of $\gamma$).
628: 
629: Finally, we note that a recipe similar to (i) consisting in
630: taking
631: the frequencies at the median (instead of the center) of each 
632: block~\footnote{$\int_{\omega_j}^{\omega_{j+1}} g(\omega) d\omega= 2\gamma /N= 
633: 2\int_{-\gamma}^{\omega_1} g(\omega) d\omega$.}
634: exhibits the same exponents $\mu, \nu$ that scheme (i).
635: 
636: 
637: \section{Conclusions}
638: \label{concl}
639: In the present paper, the first-order phase transition arising
640: when imposing a uniform frequency distribution on  
641: the Kuramoto model has been studied.
642: In the case of an infinite population, 
643: we have found an explicit asymptotic 
644: dependence of the order parameter after criticality, Eq.~(\ref{2/3}).
645: 
646: 
647: For a {\em finite} population, our first conclusion
648: is that in contrast to strictly unimodal distributions
649: (Cauchy, Gaussian, parabolic,...) of the natural frequencies, 
650: which exhibit transitions
651: of second-order type,
652: the thermodynamic limit is non-trivial for a uniform 
653: distribution. In the finite-$N$ case, 
654: the synchronized state does not split directly into $N$ clusters,
655: but through a cascade of frequency splittings. To be 
656: congruent with the first-order phase transition
657: predicted in the thermodynamic limit all the splittings
658: must accumulate at $K_c$ as $N\rightarrow\infty$.
659: 
660: The dependence of the order parameter $r$ 
661: on the coupling $K$, has been 
662: expressed in an easy-to-compute formula, Eq.~(\ref{solvnumn}).
663: In this formula, the population size $N$ enters explicitly,
664: and it allows, except very close to the desynchronization
665: point, 
666: an accurate computation of the order parameter,
667: even for small $N$.  
668: 
669: {Two sampling schemes to set the natural
670: frequencies have been compared. 
671: The scheme we propose in (\ref{unif}) converges to the 
672: thermodynamic limit faster than another used in the literature [Eq.~(\ref{unif2})]. 
673: The comparison is
674: based on the different exponent of the power-law convergence 
675: for the point of the first frequency splitting ($\mu\approx 1.5$ vs. $\mu =1$) and the 
676: shift of the order parameter ($\nu=2$ vs. $\nu=1$). 
677: Other
678: frequency distributions (with compact support) different from the uniform one
679: have been discussed. Among the infinite
680: possible sampling schemes, those studied here appear the most 
681: natural ones to us. Nonetheless, further investigation is needed 
682: to assess the existence of an optimal
683: sampling scheme to mimic the thermodynamic limit
684: with a finite population.}
685: 
686: %Also, we have found that the coupling at the 
687: %first frequency-splitting ($K_s$) approaches the critical point ($K_c$) 
688: %according to a power of $N$. The exponent of this power law
689: %depends crucially on the arrangement of the
690: %natural frequencies.
691: 
692: %Finally, 
693: %we have studied other frequency distributions with compact support, finding
694: %also power-law dependences on $N$ for the coupling of 
695: %first frequency splitting, and for the order parameter $r$ 
696: %in the fully synchronized state.
697: 
698: 
699: In spite of the lack of structural stability under perturbations 
700: of the uniform frequency distribution (in
701: the thermodynamic limit, but not in the finite case 
702: as proved in~\cite{maistrenko}), the results in this paper 
703: could be useful in order to
704: understand more complicated
705: schemes, like oscillator networks~\cite{Watts}. The 
706: use of a frequency distribution with an
707: abrupt transition seems more suited to
708: better resolve critical points
709: in this kind of systems. 
710: 
711: 
712: 
713: \acknowledgments
714: 
715: The author thanks Ernest~Montbri\'o for fruitful discussions and critical 
716: reading of the manuscript. Useful comments from Eduardo G.~Altmann 
717: and Luis G.~Morelli are also gratefully acknowledged. 
718:  
719: 
720: 
721: 
722: \bibliographystyle{prsty}
723: 
724: \begin{thebibliography}{10}
725: 
726: \bibitem{PRK}
727: A. Pikovsky, M. Rosenblum, and J. Kurths, {\em Synchronization: A {U}niversal
728:   {C}oncept in {N}onlinear {S}ciences} (Cambridge University Press, Cambridge,
729:   2001).
730: 
731: \bibitem{Blekhman}
732: I. Blekhman, {\it Synchronization in {S}cience and {T}echnology}. Nauka,
733:   Moscow, 1981 (in Russian); English translation: ASME Press, New York, 1988.
734: 
735: \bibitem{Winfree}
736: A.~T. Winfree, {\em The geometry of biological time} (Springer, New York,
737:   1980).
738: 
739: \bibitem{wiesenfeld}
740: K. Wiesenfeld, P. Colet, and S.~H. Strogatz, Phys. Rev. Lett. {\bf 76},  404
741:   (1996).
742: 
743: \bibitem{alex}
744: M.~C. Cross, A. Zumdieck, R. Lifshitz, and J.~L. Rogers, Phys. Rev. Lett. {\bf
745:   93},  224101  (2004).
746: 
747: \bibitem{daido90}
748: H. Daido, J. Stat. Phys. {\bf 60},  753  (1990).
749: 
750: \bibitem{inertia}
751: H.-A. Tanaka, A.~J. Lichtenberg, and S. Oishi, Phys. Rev. Lett. {\bf 78},  2104
752:    (1997).
753: 
754: \bibitem{bonilla92}
755: L.~L. Bonilla, J.~C. Neu, and R. Spigler, J. Stat. Phys. {\bf 67},  313
756:   (1992).
757: 
758: \bibitem{inertiad}
759: H.-A. Tanaka, A.~J. Lichtenberg, and S. Oishi, Physica D {\bf 100},  279
760:   (1997).
761: 
762: \bibitem{Kuramoto}
763: Y. Kuramoto, {\em Chemical Oscillations, Waves, and Turbulence}
764:   (Springer-Verlag, Berlin, 1984).
765: 
766: \bibitem{vanhemmen}
767: J.~L. van Hemmen and W.~F. Wreszinski, J. Stat. Phys. {\bf 72},  145  (1993).
768: 
769: \bibitem{rmp_kuramoto}
770: J.~A. Acebr\'on {\it et~al.}, Rev. Mod. Phys. {\bf 77},  137  (2005).
771: 
772: \bibitem{maistrenko}
773: Yu. Maistrenko, O. Popovych, O. Burylko, and P.~A. Tass, Phys. Rev. Lett. {\bf
774:   93},  084102  (2004).
775: 
776: \bibitem{strogatz91}
777: S.~H. Strogatz and R.~E. Mirollo, J. Stat. Phys. {\bf 63},  613  (1991).
778: 
779: 
780: \bibitem{strogatz2000}
781: S.~H. Strogatz, Physica D {\bf 143},  1  (2000).
782: 
783: \bibitem{daido87}
784: H. Daido, J. Phys. A {\bf 20},  L629  (1987).
785: 
786: \bibitem{balmforth2000}
787: N.~J. Balmforth and R. Sassi, Physica D {\bf 143},  21  (2000).
788: 
789: \bibitem{pikarufo}
790: A. Pikovsky and S. Ruffo, Phys. Rev. E {\bf 59},  1633  (1999).
791: 
792: \bibitem{mirollo2005}
793: R.~E. Mirollo and S.~H. Strogatz, Physica D {\bf 205},  249  (2005).
794: 
795: \bibitem{matthews91}
796: P.~C. Matthews, R.~E. Mirollo, and S.~H. Strogatz, Physica D {\bf 52},  293
797:   (1991). O.~V. Popovych, Yu.~L. Maistrenko, and P.~A. Tass, Phys. Rev. E {\bf 71},  065201(R)  (2005).
798: 
799: \bibitem{Watts}
800: D.~J. Watts, {\em Small Worlds}, Chap.~9 (Princeton University Press, Princeton, 1999).
801: 
802: \end{thebibliography}
803: 
804: 
805: %\bibliography{references}
806: 
807: \end{document}
808: