1: %\documentclass[prl,twocolumn,showpacs]{revtex4}
2: \documentclass[prl,showpacs,twocolumn]{revtex4}
3: %\draft
4: \usepackage{graphicx}
5: \usepackage{psfig}
6:
7: \begin{document}
8:
9: \newcommand {\ee}[1] {\label{#1} \end{equation}}
10: \newcommand{\be}{\begin{equation}}
11: \newcommand {\e} {\varepsilon}
12: \def\reff#1{(\ref{#1})}
13: \def\S{\mbox{\bf S}}
14: \def\P{\mbox{\bf P}}
15: \def\I{\mbox{\bf I}}
16: \def\re{\mbox{Re}}
17: % \draft command makes pacs numbers print
18: % \draft
19: % \tightenlines
20:
21: % \wideabs{
22: \title{Noise-induced dynamics in bistable systems with delay}
23: \date{\today}
24: \author{L. S. Tsimring}
25: \affiliation{Institute for Nonlinear Science, University of California,
26: San Diego, La Jolla, CA 92093-0402}
27: \author{A. Pikovsky}
28: \affiliation{Department of Physics, University of Potsdam,
29: Postfach 601553, D-14415 Potsdam, Germany}
30:
31:
32: \begin{abstract}%
33: Noise-induced dynamics of a prototypical bistable system with delayed
34: feedback is studied theoretically and numerically. For small noise and
35: magnitude of the feedback, the problem is reduced to the analysis of the
36: two-state model with transition rates depending on the
37: earlier state of the system. In this two-state approximation, we found
38: analytical formulae for the autocorrelation function, the power
39: spectrum, and the linear response to a periodic perturbation. They show
40: very good agreement with direct numerical simulations of the original
41: Langevin equation. The power spectrum has a pronounced peak at the
42: frequency corresponding to the inverse delay time, whose amplitude has a
43: maximum at a certain noise level, thus demonstrating coherence
44: resonance. The linear response to the external periodic force also has
45: maxima at the frequencies corresponding to the inverse delay time and
46: its harmonics.
47: \end{abstract}
48:
49: \pacs{05.40.-a, 02.50.Ey}
50: % }
51: %\twocolumn
52: \maketitle
53: The effects of random noise on bistable systems and the related
54: phenomenon of stochastic resonance have received an enormous attention
55: in the last decade. As a result, a comprehensive theory and a whole
56: range of experimental observations have emerged (for a recent review
57: see~\cite{sr}). In many physical as well as biological systems, the
58: time-delayed feedback plays a significant role in the dynamics. These
59: systems in the absence of noise have been well thoroughly investigated
60: using the theory of delay-differential equations (DDE)~\cite{dde}. The
61: theory of stochastic delay-differential equations (SDDE), in which
62: effects of noise and time delay are combined, remains much less studied.
63: Meanwhile, it appears that the combination of these features is
64: ubiquitous in nature as well as in society. Examples include biophysiological
65: dynamics~\cite{longtin90-chen97} and laser dynamics in optical cavities
66: \cite{garcia96-buldu01-masoller01}. It is also believed that the
67: combined effects of noise, bistability, and delay play an important role
68: in gene regulatory networks \cite{collins00}.
69:
70: The delayed stochastic systems have been a subject of several recent
71: papers~\cite{longtin,ohira95-00,ohira99}. In
72: Refs.~\cite{longtin}, a systematic statistical description
73: of a certain class of stochastic delay-differential equations was
74: developed in the limit of small time delay. More interesting, however,
75: is the case of a large time delay which is comparable with the mean
76: Kramers transition time determined by the noise intensity and the
77: potential barrier height. In this case, resonant phenomena may occur
78: which would lead to spontaneous oscillations of the system with a certain
79: preferred frequency. In Refs.~\cite{ohira95-00}, Ohira and
80: co-workers studied the related phenomenon of delayed random walks.
81: In that model, the hopping
82: probability depends on the position of the particle a given number of
83: hops in the past. In certain cases, the particle diffusion is limited,
84: and it exhibits quasi-regular oscillations near the origin.
85: In Ref.~\cite{ohira99}, Ohira and Sato studied a discrete-time
86: two-state system in which the occupancy probabilities of
87: the two states depended on the state of the system some $N$ time steps
88: before. While that model also showed some interesting resonant features, it
89: appears to us somewhat unrealistic, since the states of the system at
90: two consecutive iterations are completely uncoupled, and its dynamics is
91: in fact identical to that of a superposition of $N$ independent
92: one-dimensional maps affected by random noise. In most practically
93: relevant cases, however, the state of the system should be affected in
94: the first place by its immediate past, with additional correction
95: arising from the time-delayed feedback.
96:
97: In this Letter we study the effects of the thermal activation on
98: bistable systems with additional time-delayed feedback.
99: Our prototypical model is the overdamped particle motion in the double-well
100: quartic potential $U(x(t),x(t-T))$, described by the Langevin equation
101: \begin{eqnarray}
102: \frac{dx(t)}{dt}&=&
103: -\frac{\partial U(x(t),x(t-T))}{\partial x(t)}+\sqrt{2D}\xi(t)
104: \nonumber\\
105: &\equiv&
106: x(t)-x^3(t)+\epsilon x(t-T)+\sqrt{2D}\xi(t).
107: \label{sdde}
108: \end{eqnarray}
109: Here $T$ is the delay and $\epsilon$ is the strength of the feedback, and
110: $\xi(t)$ is a Gaussian white noise with $\langle \xi\rangle=0$
111: and $\langle \xi(t)\xi(t')\rangle=\delta(t-t')$.
112:
113: In our analytic description we approximate (\ref{sdde}) with a two-state
114: (dichotomic) system, in
115: which the dynamical variable $s(t)$ takes two values $s=\pm 1$. This
116: reduction has
117: been successfully used in studies of the stochastic resonance \cite{mcnamara89}. The
118: dynamics of $s$ is fully determined by the switching rates, i.e. by the
119: probabilities to switch $s\rightarrow -s$. Because of the delay, we have two
120: switching rates depending on the state $s(t-T)$: $p_1$ if the state at time $t-T$ is
121: the same as at time $t$, and $p_2$ otherwise. Thus, the the switching rate can
122: be written as
123: \begin{equation}
124: p(t)=\frac{p_1+p_2}{2}+\frac{p_1-p_2}{2} s(t)s(t-T)\;.
125: \label{rates}
126: \end{equation}
127: A quantitative relation between the rate process (\ref{rates}) and the original
128: model (\ref{sdde}) can be easily established for small $D$ and $\epsilon$ by
129: virtue of the Kramers formula for the escape rate \cite{kramers}
130: $r_K=(2\pi)^{-1}\sqrt{U''(x_m)U''(x_0)}\exp[-\Delta U/D]$, where $x_m$
131: and $x_0$ are the positions of the minimum and the maximum of the potential,
132: respectively, and $\Delta U$ is the potential barrier to cross over.
133: For small $D$, the switching rates are small compared to the intra-well
134: equilibration rate, and the probability
135: density distribution is close to a narrow Gaussian distribution centered around
136: $x_m$, and so the adiabatic approximation applies.
137: For small $\epsilon$, $|x_m|=1\pm \epsilon/2$ depending on the sign of
138: $x(t)x(t-T)$, $x_0=0$, and in the first order in $\epsilon$ we obtain
139: \begin{equation}
140: p_{1,2}=
141: \frac{\sqrt{2\pm 3\epsilon}}{2\pi}\exp\left[-\frac{1\pm 4\epsilon}{4D}\right]\;.
142: \label{trans}
143: \end{equation}
144:
145: Without loss of generality let us assume that the system is at state
146: $s=1$ at time $0$.
147: We define $n_\pm(t)$ to be the probability of attaining value $\pm 1$
148: at time $t$. The master equations for $n_\pm(t)$ is written in a usual
149: way,
150: \begin{equation}
151: \begin{array}{l}
152: \dot{n}_+(t)=-W_\downarrow (t) n_+(t) + W_\uparrow (t) n_-(t)\;,\\[1ex]
153: \dot{n}_-(t)=-W_\uparrow (t) n_-(t) + W_\downarrow (t) n_+(t)\;,
154: \end{array}
155: \label{master1}
156: \end{equation}
157: where $W_\uparrow (t)\;dt$ is the probability of transition from $-1$ to
158: $+1$ within time interval $(t,t+dt)$
159: and vice versa. In our stochastic model with time-delayed feedback,
160: \begin{equation}
161: \begin{array}{l}
162: W_\downarrow(t)= p_1 n_+(t-T) + p_2 n_-(t-T)\;,\\[1ex]
163: W_\uparrow(t)= p_2 n_+(t-T) + p_1 n_-(t-T)\;.
164: \end{array}
165: \label{wpm}
166: \end{equation}
167: Substituting (\ref{wpm}) in (\ref{master1}) and making use
168: of the normalization condition $n_-(t) + n_+(t) = 1$, we obtain
169: \begin{eqnarray}
170: \dot{n}_+(t)= p_2 n_-(t) - p_1 n_+(t) - (p_2-p_1)n_-(t-T),
171: \label{master2}\\
172: \dot{n}_-(t)= - p_1 n_-(t) + p_2 n_+(t) - (p_2-p_1)n_+(t-T).
173: \label{master3}
174: \end{eqnarray}
175: The correlation function $C(\tau)$ is determined as
176: %the averaged time
177: %response function
178: \begin{equation}
179: C(\tau)=\langle s(\tau)s(0)\rangle=\langle s(\tau)\rangle=n_+(\tau)-n_-(\tau),
180: \end{equation}
181: (we recall that the initial state is $s(0)=1$).
182: Thus, replacing $t$ with $\tau$ and subtracting (\ref{master3}) from
183: (\ref{master2}), we obtain
184: \begin{equation}
185: \frac{dC(\tau)}{d\tau}=-(p_1+p_2)C(\tau)+(p_2-p_1)C(\tau-T).
186: \label{ceq2}
187: \end{equation}
188: This equation should be complemented with the symmetry $C(-\tau)=C(\tau)$ and the
189: normalization $C(0)=1$ conditions.
190:
191: The solution of this linear equation on the interval $(0,T)$
192: can be found using
193: ansatz $C(\tau)=A\exp(-\lambda \tau) + B \exp\lambda(\tau-T)$.
194: Plugging this ansatz in Eq.(\ref{ceq2}) yields
195: $\lambda=2\sqrt{p_1p_2}, \
196: B=A(\sqrt{p_2}-\sqrt{p_1})/(\sqrt{p_2}+\sqrt{p_1})\exp(-2\sqrt{p_1p_2}T)$.
197: Constant $A$ is found from the condition $C(0)=1$, and we
198: obtain
199: \begin{equation}
200: C(\tau)=\frac{(\sqrt{p_1}+\sqrt{p_2})e^{-\lambda\tau}+
201: (\sqrt{p_2}-\sqrt{p_1})e^{\lambda(\tau-T)}}
202: {\sqrt{p_1}+\sqrt{p_2}+(\sqrt{p_2}-\sqrt{p_1})e^{-\lambda T}}\;.
203: \label{ctauf}
204: \end{equation}
205: Using (\ref{ceq2}) and (\ref{ctauf}), one can easily calculate
206: $C(\tau)$ at all $\tau>T$,
207: \begin{eqnarray}
208: &&C(nT+\tau')=e^{-(p_1+p_2)\tau'}C(nT)+(p_2-p_1)\times\nonumber\\
209: &&\times\int_0^{\tau'}
210: C((n-1)T+t)e^{(p_1+p_2)(t-\tau')}dt\;,
211: \label{ctauf1}
212: \end{eqnarray}
213: where $n=1,2,...$ and $0<\tau'<T$.
214:
215: \begin{figure}[!htb]
216: \centerline{\includegraphics[width=0.43\textwidth]%
217: {fig1.ps}}
218: \caption{The autocorrelation function in the two-state model as a
219: function of
220: the time lag $\tau$ and the feedback strength
221: ${p}_2-{p}_1$, for $({p}_1+{p}_2)T=10$.}
222: \label{fig:corf}
223: \end{figure}
224: We present correlation function (\ref{ctauf}),(\ref{ctauf1}) as a function
225: of normalized time $\tau/T$ and dimensionless parameters
226: $p_{1,2}T$ in Fig.~\ref{fig:corf}. Its structure
227: differs depending on whether the feedback is positive ($p_2>p_1$, what
228: corresponds to a positive $\epsilon$ in (\ref{sdde})), or negative ($p_2<p_1$,
229: $\epsilon<0$). For positive feedback the correlation function is positive, and has
230: maxima at $\tau\approx nT$. For negative feedback the peaks at $\tau\approx nT$ have alternating
231: signs. It is interesting to note that the peaks of the correlation
232: function are always delayed with respect to $nT$. For $\lambda
233: T\gg 1$, the time interval corresponding to the first peak is
234: \begin{equation}
235: \tau_1=T+(\sqrt{p_2}-\sqrt{p_1})^{-2}\ln\frac{(\sqrt{p_1}+\sqrt{p_2})^2}{2(p_1+p_2)}\;.
236: \label{offset}
237: \end{equation}
238:
239: From the correlation function we can also determine the power spectrum
240: $S(\omega)=\int_{-\infty}^\infty C(\tau)\cos(\omega\tau)\,d\tau$.
241: It is
242: convenient to derive the expression for $S$ directly from equation (\ref{ceq2}).
243: Denoting $L(\omega)=\int_0^\infty C(\tau)\exp[i\omega\tau]\,d\tau$ and
244: substituting here (\ref{ceq2}) we obtain
245: \begin{equation}
246: -C(0)-i\omega L=-(p_1+p_2)L+(p_2-p_1)e^{i\omega T}
247: \left[L-I(\omega)\right]\;,
248: \label{psp1}
249: \end{equation}
250: where
251: \begin{widetext}
252: \[
253: I(\omega)=\int_0^T C(\tau)e^{-i\omega\tau} d\tau=\frac{(\sqrt{p_1}+\sqrt{p_2})[1-e^{(-i\omega -\lambda)T}]
254: (i\omega+\lambda)^{-1}+
255: (\sqrt{p_2}-\sqrt{p_1})
256: [e^{-i\omega T}-e^{-\lambda T}](i\omega -\lambda)^{-1}}
257: {\sqrt{p_1}+\sqrt{p_2}+(\sqrt{p_2}-\sqrt{p_1})e^{-\lambda T}}\;.
258: %\label{psp2}
259: \]
260: \end{widetext}
261: Using $C(0)=1$ and $S(\omega)=2\re L(\omega)$, we obtain
262: %from (\ref{psp1}) finally
263: \begin{equation}
264: S(\omega)=2\re\frac{1+(p_2-p_1)e^{i\omega
265: T}I(\omega)}{p_1+p_2-(p_2-p_1)e^{i\omega T}-i\omega}\;.
266: \label{pow1}
267: \end{equation}
268:
269: We compare this analytic result with numerical simulations of the bistable
270: oscillator (\ref{sdde}) in Fig.~\ref{fig:spec}.
271:
272:
273: \begin{figure}[!htb]
274: \centerline{\includegraphics[width=0.38\textwidth]%
275: {fig2.ps}}
276: \caption{Comparison of the power spectrum in Eq.~(\ref{sdde}) for $D=0.1$,
277: $T=250$, $\epsilon=0.05$ (solid line) with theory~(\ref{pow1}) (dashed line).
278: }
279: \label{fig:spec}
280: \end{figure}
281: \begin{figure}[!htb]
282: \centerline{\includegraphics[width=0.42\textwidth]%
283: {fig3.ps}}
284: % {spectr_bw.eps}}
285: \caption{Power spectrum in the delay system (\ref{sdde}) calculated in the
286: two-state approximation for $T=250$ and $\epsilon=0.1$.
287: }
288: \label{fig:spec2}
289: \end{figure}
290: In figure~\ref{fig:spec2} we show the dependence of the power spectrum on the
291: noise intensity $D$, while the feedback parameter $\epsilon$ is kept constant
292: (the switching rates $p_{1,2}$ are calculated according to (\ref{trans})). One
293: can see that the peak at the main frequency $\omega\approx 2\pi/T$ has a maximum
294: at a certain noise level. This is a characteristic feature of the coherence
295: resonance \cite{pik97}: the coherence in the noise-driven system attains maximum
296: at a ``resonant'' noise temperature. In the present case the underlying
297: physical mechanism is the resonance between the Kramers rate and the delay.
298: If the Kramers rate is small (for
299: small noise intensity), a characteristic interval between the switches is larger
300: than delay time, and the latter is not displayed in the spectrum because the
301: process is a purely Poissonian one
302: (with a renormalized due to the feedback switching
303: rate). For an intermediate Kramers rate the switchings are highly influenced by
304: the feedback, with a preferable periodicity with the delay time $T$ being
305: manifested as a peak in the spectrum. For large noise intensity the effect of
306: the feedback decreases again, because the relative magnitude of the delayed
307: feedback $(p_2-p_1)/(p_1+p_2)$ is proportional to $D^{-1}$.
308:
309: Let us discuss now the response of the time-delay stochastic system to a periodic
310: external force. Similarly to Ref.\cite{mcnamara89},
311: we assume that the transition rates (\ref{wpm}) are modulated
312: with a frequency $\Omega$ according to the Arrhenius rate law,
313: \begin{eqnarray*}
314: W_\downarrow (t)&=& [p_1 n_+(t-T) + p_2 n_-(t-T)]e^{\gamma(t)}
315: \;,\\
316: W_\uparrow (t)&=& [p_2 n_+(t-T) + p_1 n_-(t-T)]e^{-\gamma(t)}
317: \;,
318: \end{eqnarray*}
319: where $\gamma(t)=\mu D^{-1}\cos(\Omega t +\phi)$.
320: The equation for the quantity $\sigma(t)=n_+(t)-n_-(t)$ (which now is not the
321: autocorrelation function) now reads
322: \begin{eqnarray*}
323: \frac{d\sigma}{dt}&=&-(p_1+p_2)(n_+e^{\gamma(t)}-n_-e^{-\gamma(t)})\nonumber
324: \\&&+
325: (p_2-p_1)(n_+e^{\gamma(t)}+n_-e^{-\gamma(t)})\sigma(t-T)\;.
326: %\label{clr}
327: \end{eqnarray*}
328: In the linear approximation $\mu<<1$ this reduces to
329: \[
330: \frac{d\sigma}{dt}=-(p_1+p_2)(\sigma+\gamma(t))+
331: (p_2-p_1)\sigma(t-T)(1+C(t)\gamma(t))\;.
332: %\label{clr1}
333: \]
334: Now writing $\sigma=\sigma_0+\mu D^{-1} \sigma_1$ where $\sigma_0$ is the solution
335: (\ref{ctauf},\ref{ctauf1}), we obtain for the first-order correction $\sigma_1$
336: \begin{eqnarray}
337: &&\frac{d\sigma_1}{dt}=-(p_1+p_2)\sigma_1+(p_2-p_1)\sigma_1(t-T)\nonumber\\&&
338: +
339: ((p_2-p_1)\sigma_0(t)\sigma_0(t-T)-(p_1+p_2))\cos(\Omega t +\phi)\;.
340: \label{clr2}
341: \end{eqnarray}
342: We are interested in the response at the frequency $\Omega$ for $t\to\infty$,
343: because only this part contributes to the delta-peak in the spectrum at this
344: frequency. For $t\to\infty$ $\sigma_0(t)\to 0$, so we can neglect the corresponding
345: term $(p_2-p_1)\sigma_0(t)\sigma_0(t-T)$ in (\ref{clr2}) and write the solution as
346: \[
347: \sigma_1(t)=
348: \mbox{Re}\frac{(p_1+p_2)e^{i\Omega t +\phi}}{(p_2-p_1)e^{-i\Omega T}-i\Omega
349: -(p_1+p_2)}\;.
350: %\label{clr3}
351: \]
352: This is exactly the periodic component at frequency $\Omega$
353: in the process $s(t)$, and the linear
354: response $\eta$ is
355: \begin{equation}
356: \eta=\frac{1}{2D^2}\frac{(p_1+p_2)^2}{\left|(p_2-p_1)e^{-i\Omega T}-i\Omega
357: -(p_1+p_2)\right|^2}\;.
358: \label{clr5}
359: \end{equation}
360: In the absence of delayed feedback, when $p_1=p_2=r_K$, this expression coincides with
361: that of \cite{mcnamara89} for the stochastic resonance in the two-state
362: model. With feedback, the response demonstrates a resonance-like structure
363: in dependence on the driving frequency (contrary to the classical stochastic
364: resonance), see Fig.~\ref{fig:resp}.
365: \begin{figure}[!htb]
366: \centerline{\includegraphics[width=0.38\textwidth]%
367: {fig4.ps}}
368: \caption{Linear response of model (\ref{sdde}) for $D=0.1$,
369: $T=250$, $\epsilon=0.05$, normalized by the variance of the process
370: (circles), compared with theory~(\ref{clr5}) (line)
371: }
372: \label{fig:resp}
373: \end{figure}
374:
375:
376: In conclusion, we have developed a theory of a noise-driven bistable
377: system with delayed feedback. In general, this problem is very
378: difficult because of the non-Markovian nature of the dynamics. However,
379: for small noise and small magnitude of the feedback, the problem can be
380: greatly simplified by reduction to the two-state (dichotomic) model with
381: certain transition rates which depend on the earlier state of the
382: system. In this two-state approximation, we were able to derive the
383: analytical formulae for the autocorrelation function, the power
384: spectrum, and the linear response to a periodic perturbation. They show
385: very good agreement with direct numerical simulations of the
386: corresponding Langevin equation. The power spectrum has a pronounced
387: peak at the frequency corresponding to the delay time, whose amplitude
388: has a maximum at a certain noise level, thus demonstrating coherence
389: resonance. This level corresponds to the mean switching time comparable
390: to the delay time. The linear response to the external periodic force
391: also has maxima at the frequencies corresponding to the inverse delay
392: time and its harmonics.
393:
394: In a more general context of multistable dynamical systems with memory,
395: the behavior of the system depends on its past through some memory
396: kernel. Such a kernel is equivalent to multiple time delays. A similar
397: analysis of the correlation properties for such systems would be of
398: great interest. Furthermore, in applications, multiple feedback loops
399: with different delay times occur in networks of interacting elements,
400: such as biological neurons, stock traders, or Internet nodes. It is
401: very interesting to study the influence of noise on the dynamics of
402: such networks. We anticipate the emergence of spontaneous oscillations and
403: the coherence resonance features similar to the effects considered in
404: this Letter.
405:
406: We thank A. Neiman and M. Rosenblum for helpful discussions.
407: L.T. acknowledges support from the U.S. Department of Energy,
408: Engineering Research Program of the Office of Basic Energy Sciences
409: under grants DE-FG03-95ER14516 and DE-FG03-96ER14592, and from the
410: U.S. Army Research Office under MURI grant DAAG55-98-1-0269.
411:
412:
413: \begin{thebibliography}{99}
414: \bibitem{sr} L.~Gammaitoni, P.~Hanggi, P.~Jung, F.~Marchesoni, Rev. Mod.
415: Phys., {\bf 70}, 223 (1998).
416: \bibitem{dde}J.~Wiener adn J.~K.~Hale, {\em Ordinary and delay differential
417: equations.} New York, Wiley, 1992.
418: \bibitem{longtin90-chen97} A. Longtin {\it et al.}, Phys. Rev. A,
419: {\bf 41}, 6992 (1990);
420: %\bibitem{chen97}
421: Y. Chen, M. Ding, J. A. Scott Kelso, Phys. Rev. Lett., {\bf
422: 79}, 4501 (1997).
423: \bibitem{garcia96-buldu01-masoller01} J. Garc{\'i}a-Ojalvo, R. Roy, Phys. Lett. A, {\bf 224}, 51
424: (1996);
425: %\bibitem{buldu01}
426: J.~M.~Buld{\'u} {\em et al.}, arXiv:nlin.AO/0104047;
427: %\bibitem{masoller01}
428: C.~Masoller, Phys. Rev. Lett., {\bf 86}, 2782 (2001).
429: \bibitem{collins00}J. Hasty et al. Proc Natl Acad Sci USA, {\bf 97},
430: 2075 (2000).
431: \bibitem{longtin}S.~Guillouzic, I.~L'Heureux, A. Longtin, Phys. Rev. E,
432: {\bf 59}, 3970 (1999); {\bf 61}, 4906 (2000).
433: \bibitem{ohira95-00} T.~Ohira and J.~G.~Milton, Phys. Rev. E, {\bf 52}, 3277
434: (1995);
435: %\bibitem{ohira97}
436: T.~Ohira, Phys. Rev. E, {\bf 55}, R1255
437: (1997);
438: %\bibitem{ohira00}
439: T.~Ohira and T.~Yamane, Phys. Rev. E, {\bf 61}, 1247
440: (2000).
441: \bibitem{ohira99} T.~Ohira and Y.~Sato, Phys. Rev. Lett., {\bf 82}, 2811
442: (1999).
443: \bibitem{mcnamara89} B.~S.~McNamara, K.~Wiesenfeld, Phys. Rev. A, {\bf 39}, 4854
444: (1989).
445: \bibitem{kramers} H. Kramers, Physica (Utrecht), {\bf 7}, 284 (1940).
446: \bibitem{pik97} A. Pikovsky and J. Kurths, Phys. Rev. Lett., {\bf 78}, 775
447: (1997).
448: \end{thebibliography}
449: \end{document}
450: