1: %\documentclass[aps,pre,preprint]{revtex4}
2: %\documentclass[twocolumn,showpacs,preprintnumbers,amsmath,amssymb]{revtex4}
3: \documentclass[aps,pre,preprint,showpacs,preprintnumbers,amsmath,amssymb]
4: {revtex4}
5: \usepackage{graphicx}
6: %\usepackage{showtags}
7: %\usepackage{showlabels}
8: %\usepackage{dcolumn}% Align table columns on decimal point
9: \usepackage{bm}% bold math
10: %\font\bbb=msbm10 at 12pt
11: %\documentstyle[eqsecnum,aps]{revtex}
12: %\font\bbb=msbm10
13: %\let\labelm\label\renewcommand{\label}[1]{\fbox{\tt #1}\labelm{#1}}
14: \def\btt#1{{\tt$\backslash$#1}}
15: %\renewcommand{\textfraction}{0}
16: \def\be{\begin{equation}}
17: \def\ba{\begin{eqnarray}}
18: \def\a{\alpha}
19: \def\b{\beta}
20: \def\g{\gamma}
21: \def\G{\Gamma}
22: \def\d{\delta}
23: \def\D{\Delta}
24: \def\e{\epsilon}
25: \def\z{\zeta}
26: \def\h{\eta}
27: \def\th{\theta}
28: \def\Th{\Theta}
29: \def\k{\kappa}
30: \def\l{\lambda}
31: \def\La{\Lambda}
32: \def\m{\mu}
33: \def\n{\nu}
34: \def\x{\xi}
35: \def\p{\pi}
36: \def\r{\rho}
37: \def\s{\sigma}
38: \def\S{\Sigma}
39: \def\t{\tau}
40: \def\f{\varphi}
41: \def\F{\Phi}
42: \def\c{\chi}
43: \def\ps{\psi}
44: \def\o{\omega}
45: \def\O{\Omega}
46: \def\i{\int}
47: \def\lg{\langle \g \rangle}
48: \def\bA{{\mathbf A}}
49: \def\bfa{{\mathbf a}}
50: \def\bB{{\mathbf B}}
51: \def\bx{{\mathbf x}}
52: \def\bX{{\mathbf X}}
53: \def\by{{\mathbf y}}
54: \def\bv{{\mathbf v}}
55: \def\bD{{\mathbf D}}
56: \def\bK{{\mathbf K}}
57: \def\bM{{\mathbf M}}
58: \def\bb{{\mathbf b}}
59: \def\bff{{\mathbf f}}
60: \def\bG{{\mathbf G}}
61: \def\bV{{\mathbf V}}
62: \def\br{\mbox{\boldmath{$\r$}}}
63: \def\bfi{\mbox{\boldmath{$\f$}}}
64: \def\bg{\mbox{\boldmath{$\g$}}}
65: \def\bps{\mbox{\boldmath{$\ps$}}}
66: \def\bz{\mbox{\boldmath{$\z$}}}
67: \def\bLa{\mbox{\boldmath{$\La$}}}
68: \def\ee#1{\label{#1}\end{equation}}
69: \def\ea#1{\label{#1}\end{eqnarray}}
70:
71: \begin{document}
72: \title{Rate Description of Fokker-Planck Processes
73: with Time Dependent Parameters }
74: \author{Peter Talkner}
75: \affiliation{Universit\"at Augsburg, Institut f\"ur Physik,
76: Universit\"atsstrasse 1, D-86135 Augsburg, Germany}
77: \author{Jerzy \L uczka}
78: \affiliation{Institute of Physics, University of Silesia, 40-007
79: Katowice, Poland}
80: \date{\today}
81: \begin{abstract}
82: The reduction of a continuous Markov process with multiple metastable
83: states to a discrete rate process is investigated in the presence of
84: slow
85: time dependent parameters such as periodic external forces or
86: slowly fluctuating barrier heights.
87: A quantitative criterion is provided
88: under which condition a kinetic description with time dependent frozen
89: rates applies. Finally it is shown how the long time behavior of the
90: underlying continuous process can be retrieved from the knowledge of
91: the discrete process by means of an appropriate random decoration of
92: the discrete states. As a particular example of the presented theory an
93: over-damped bistable Brownian oscillator with periodic driving is
94: discussed.
95: \end{abstract}
96: \pacs{02.50.Ga, 82.20.Uv, 82.37.-j, 82.39.Rt}
97:
98: \maketitle
99:
100: \section{Introduction}\label{I}
101: The separation of time scales is a frequently met dynamical feature of
102: physical,
103: chemical and biological
104: processes. In
105: practically all cases the parameters characterizing a particular system
106: are only constant on a limited time scale. On longer time scales they
107: may change their values because the surrounding environment slowly
108: changes its state. We just note that even the most fundamental ``constants''
109: of nature like e.g. the fine structure constant possibly undergo a very slow
110: change in time \cite{D}. On the other hand,
111: most natural processes also display very fast changes which usually will be
112: disregarded in a phenomenological description. They only will show up in a
113: microscopic description in which the effective degrees of freedom of the
114: phenomenological model are
115: represented as functions of microscopic degrees of freedom.
116: The time evolution of
117: these microscopic degrees of freedom
118: typically takes place on much faster time scales than those of the
119: macroscopic observables whereas, as already mentioned, global aspects of the
120: environment may
121: undergo temporal variations that are much slower than the dynamics of the
122: considered macroscopic system.
123:
124: It is this separation of time scales that allows
125: one to describe a system in terms of a comparatively simple model that
126: is
127: closed in the sense that only a few macroscopic observables are
128: subject to the dynamics.
129: On the slow time scale
130: the key technical notion
131: here is that of a
132: {\it constraint equilibrium} that is reached by the fast part of a system
133: before the slow constituents change their values on the slower time scale.
134: This constraint equilibrium then immediately adapts to
135: the changes of the slow system.
136:
137: This approach has been applied
138: in many branches of physics and
139: other sciences and is vital for various methods like
140: adiabatic elimination, averaging,
141: synergetics, subdynamics, chemical kinetics and hierarchical computational
142: schemes \cite{ae}. In the present paper we consider Markovian
143: processes the dynamics of
144: which can be subdivided into a fast and a slow component.
145: In particular, we have in mind the case where the slow motion is caused by
146: transitions between
147: metastable states \cite{htb}. The relaxation toward these states
148: constitutes the fast part of the dynamics.
149: Additionally, some
150: of the system's parameters may undergo slow variations in time that may be
151: caused by a slow drift, or fluctuations of the parameters, or by an
152: external driving of the system.
153: The dynamics of the considered system then has two slow
154: components: The transitions and the driving.
155: About their relation we do not make further assumptions:
156: The slow intrinsic dynamics can be faster or slower than the
157: external driving that must only be slow compared to the fast dynamics of the
158: system. In Ref.~\cite{t99} we have called this type of situation the
159: semiadiabatic limit, in contrast to the adiabatic limit in which the driving
160: is slow compared to the total intrinsic dynamics including that of the
161: transitions. Clearly, the
162: adiabatic limit is
163: contained in the semiadiabatic limit as a special case.
164:
165:
166: Stochastic resonance is just one example for which the semiadiabatic limit
167: is of relevance when the period of the signal is large
168: \cite{j,ghjm}.
169: Ratchets, or Brownian motors, provide another class of systems
170: which also operate in the semiadiabatic regime \cite{r}.
171: Atomic force microscopes and optical tweezers often act on time
172: scales that are slow compared to the vibrational and other intrinsic
173: time-scales of the molecules that are manipulated by them
174: \cite{erm}.
175:
176:
177:
178: The paper is organized as follows. In Sect.~\ref{II}
179: the reduction of
180: a Fokker-Planck process with multiple mestastable states to a master
181: equation is reviewed as it was developed in
182: Refs.~\cite{Ry,Mo}. This method is generalized to processes with
183: slowly time dependent parameters in Sect.~\ref{III}. The time rate of
184: change of the probability now consists of two contributions. One
185: is given by the rate as it would result if the parameters of the
186: process were frozen and the other contribution mainly takes into
187: account the changes of the geometry of the domains of attraction of
188: the metastable states. This geometric contribution is proportional to
189: the time rate of change of the parameters and therefore is negligible
190: compared to the frozen rates if the time dependent parameters change
191: sufficiently
192: slowly. In contrast to previous investigations,
193: \cite{Ast} we find that the fastest time dependence
194: of the parameters for which one may neglect the geometric contributions
195: is not only determined by the deterministic time scales of the process
196: but in general also depends on the strength of the noise. The way how
197: the noise here enters depends on the particular process.
198:
199: The present investigation is complementary to previous work that is restricted
200: to periodic driving with an intermediate regime of external driving
201: frequencies \cite{lrh}.
202:
203: In Sect.~\ref{IV} we assume the validity of the master equation with
204: frozen rates and determine a time dependent
205: decoration of the discrete states such that the dynamics of the
206: continuous variables of the underlying Fokker-Planck process are
207: recovered on the long time scale. The archetypical example for
208: stochastic resonance namely that of a
209: periodically damped bistable Brownian oscillator is discussed in
210: Sect.~\ref{V} and the validity of the McNamara-Wiesenfeld model of
211: stochastic resonance
212: \cite{McNW}
213: is discussed.
214: The paper ends with a summary in Sect.~\ref{VI}.
215:
216:
217:
218:
219:
220:
221:
222:
223:
224: \section{Multistable systems at small noise}\label{II}
225: In this section we review the reduction of a Fokker-Planck equation
226: describing a {\it time-homogeneous} process with multiple metastable states
227: to a master equation that gives the dynamics
228: on the time scales of the
229: transitions between the metastable states.
230: For this purpose we consider a dynamical system under the
231: influence of weak random
232: forces and assume that these forces can be modeled by Gaussian white noise.
233: Then the stochastic dynamics is described by a Fokker-Planck equation
234: \cite{Ri}:
235: \be
236: \frac{\partial }{\partial t}\r(\bx,t) = L \r(\bx,t)
237: \ee{FP}
238: where $L$ denotes the Fokker-Planck operator:
239: \be
240: L = -\sum_i \frac{\partial}{\partial x_i} K_i(\bx) +
241: \sum_{i,j} \frac{\partial^2}{\partial x_i
242: \partial x_j} D_{ij}(\bx)
243: \ee{FPO}
244: Here, $\bx = (x_1, x_2, ... x_n)$ denotes a point with coordinates $x_i$ in
245: the $n$-dimensional state space $\S$; further,
246: $\bK(\bx) =(K_i(\bx))$ is the drift vector and $\bD(\bx) = (D_{i,j}(\bx))$
247: the diffusion
248: matrix resulting from the random forces. We further restrict ourselves
249: to systems with a uniquely defined stationary
250: probability density $\r_0(\bx)$ with respect to which detailed balance
251: holds. Consequently, the Fokker-Planck operator satisfies the relation
252: \cite{Ri}:
253: \be
254: L \hat{\r}_0 = \hat{\r}_0 \tilde{L}^+.
255: \ee{db}
256: where $L^+$
257: is the backward, or adjoint Fokker-Planck operator:
258: \be
259: L^+ = K_i(\bx) \frac{\partial}{\partial x_i} +
260: D_{i,j}(\bx) \frac{\partial^2}{\partial x_i \partial x_j}.
261: \ee{BFP}
262: The tilde denotes the operation of time reversal, i.e.
263: $\tilde{x}_i =\e_i x_i$ with parity $\e_i = \pm 1$ and
264: $\hat{\r}_0$ is the multiplication operator with the
265: probability density, i.e. $\hat{\r}_0 f(\bx) = \r_0(\bx)f(\bx)$ where $f(\bx)$
266: is an arbitrary
267: state space function.
268:
269: In the deterministic limit the
270: diffusion matrix goes to zero and the drift vector approaches the
271: deterministic vector field $\bK^{(0)}(\bx)$ governing the deterministic motion
272: of the system:
273: \be
274: \dot{\bx}(t) = \bK^{(0)}(\bx(t))
275: \ee{dem}
276: In the presence of weak random perturbations $\bK(\bx)$
277: may differ from $\bK^{0}(\bx)$
278: by small noise induced contributions.
279:
280: We here are interested in cases where the deterministic system
281: (\ref{dem}) has a number $m$ of
282: coexisting attractors labeled by $\a$. To each attractor
283: $\a$ there belongs
284: a domain of attraction ${\cal D}_\a$.
285: The domains are disjoint and partition the total available state space:
286: \ba
287: {\cal D}_\a \bigcap {\cal D}_{\a'} & = & \emptyset \quad \mbox{for}
288: \; \a \neq {\a'} \noindent \\
289: \bigcup_{\a =1}^m {\cal D}_\a & = & \S
290: \ea{da}
291: \begin{figure}
292: \includegraphics[width=10cm]{semifig1.eps}
293: \caption{The overdamped deterministic motion $\dot{x}(t) = - V'(x(t))$
294: in the above shown potential induces a
295: partitioning of the state space $\S = \mathbb{R}_1$ into the
296: domains of attraction $\mathcal{D}_\a$ of the three
297: attractors $\a =1,2,3$ at the potential minima. Here the locations of
298: the local potential maxima coincide with the boundaries between the
299: domains of attraction.}
300: \label{f1}
301: \end{figure}
302: For weak noise the stationary probability density is almost zero
303: everywhere except close to the attractors. If the noise vanishes the
304: stationary density shrinks to Dirac $\d$-functions sitting at the attractors
305: of the deterministic system. Because no transitions between the different
306: attractors may occur in the noiseless case, each of the delta functions is a
307: stationary solution of the Fokker-Planck equation
308: $\partial/ \partial x_i K_i^0(\bx) \r_0(\bx) = 0$
309: belonging to a $m$-fold degenerate
310: eigenvalue $0$.
311:
312: The degeneracy also follows from the backward equation
313: which is governed by the adjoint Fokker-Planck operator $L^+$ defined in
314: eq.~(\ref{BFP}).
315: In the deterministic case the diffusion matrix vanishes, so
316: the second derivatives disappear and the
317: stationary solutions of the backward operator are constant along the
318: trajectories of the deterministic system. Because within a domain of
319: attraction all trajectories approach
320: the same attractor the stationary solutions of the noiseless
321: backward equation are constant on the domain of attraction
322: ${\cal D}_\a$. One then can choose the
323: characteristic functions $\c^{(0)}_\a(\bx)$
324: of the domains of attraction as $m$ independent
325: solutions of the stationary deterministic backward equation
326: $K^{(0)}_i(\bx) \partial \c^{(0)}(\bx)/\partial x_i = 0$:
327: \be
328: \c^{(0)}_\a(\bx) = \left \{ \begin{array}{ll}
329: 1 \qquad &\mbox{for} \; \bx \in {\cal D}_\a \\
330: 0 & \mbox{else}
331: \end{array}
332: \right.
333: \ee{chi}
334: In the presence of weak noise these
335: functions can be modified slightly
336: such that they still solve the stationary backward equation up to
337: corrections that are exponentially small in the noise strength: The
338: step-like
339: discontinuities then are smeared out on layers located at the
340: boundaries of the domains of attraction
341: ${\cal D}_\a$. The functions can be constructed by means of the method
342: of matched
343: asymptotic expansions \cite{MS}, see also Refs. \cite{Ry,Mo}.
344: Outside the boundary layers
345: no modification is required and hence they retain their
346: property of localizing the different domains of attraction.
347: In the following, these localizing functions will be
348: denoted by $\c_\a(\bx)$. We summarize their main properties: They solve
349: $L^+\c_\a(\bx) =0$ up to exponentially small corrections in the noise strength,
350: have the values one within and zero outside ${\cal D}_\a$ and
351: interpolate
352: between these values smoothly on a thin boundary layer \cite{t87,t94}.
353: As solutions of these particular boundary value problems, the functions
354: $\c_\a(\bx)$ are the {\it splitting
355: probabilities} that give the relative frequencies with which a trajectory
356: starting at $\bx$ first reaches the attractor $\a$ before it visits any other
357: attractor \cite{Ma}.
358:
359: The splitting probabilities in general are not eigenfunctions of the backward
360: operator. However, they span the subspace of eigenfunctions that emerges
361: from the eigenspace of the
362: $m$-fold degenerate eigenvalue $0$ of the deterministic dynamics
363: if weak noise perturbes the
364: system and lifts the degeneracy.
365: The resulting finite eigenvalues are still small compared to all
366: other finite eigenvalues. Hence, this subspace describes the slow dynamics of
367: the system at weak noise. In particular, it contains the constant function
368: which is the exact eigenfunction of the backward equation belonging to the
369: eigenvalue $0$. At weak noise this eigenfunction
370: can be represented as the sum over the characteristic
371: functions $\c_\a(\bx)$:
372: \be
373: \sum_\a \c_\a(\bx) = 1.
374: \ee{s1}
375:
376:
377:
378:
379:
380:
381:
382: Because of detailed
383: balance, the slow subspace of the forward operator is spanned by the basis
384: set $\{ \tilde{\c}_\a(\bx) \r_0(\bx) \}$, where $\tilde{\c}_\a(\bx) =
385: \c_\a(\tilde{\bx})$ denotes
386: the image of $\c_a(\bx)$ under time
387: reversal. Hence, the probability density describing the slow
388: dynamics can be represented as:
389: \be
390: \r(\bx,t) = \sum_{\a =1}^m c_\a(t) \tilde{\c}_\a(\bx) \r_0(\bx)
391: \ee{sr}
392:
393: As already noted the characteristic functions $\c_\a(\bx)$ and, therefore also
394: their time reversed partners $\tilde{\c}_\a(\bx)$, are well localized functions
395: as long as the noise does not become too large.
396: As a consequence, one may
397: use the functions $\c_\a(\bx)$ in order to determine the probability $p_\a(t)$
398: that the system resides in the domain of attraction of the attractor $\a$
399: \cite{Mo}:
400: \be
401: p_\a(t) = \left ( \c_\a, \r(t) \right )
402: \ee{pat}
403: where the scalar product is defined as the integral of the product of its
404: arguments over the state space:
405: \be
406: \left ( f, g \right ) = \i_\S d^n\bx f(\bx) g(\bx).
407: \ee{sp}
408: If we assume that the probability density $\r(t)$ is of the form (\ref{sr})
409: we obtain
410: \be
411: p_\a(t) = c_\a(t) n_\a
412: \ee{pn}
413: where we have made use of the localization of the characteristic
414: functions at weak noise:
415: \be
416: \left ( \c_\a,\tilde{\c}_{\a'} \r_0 \right ) = \d_{\a,\a'} n_\a
417: \ee{cab}
418: and have introduced the function
419: $n_\a = \i dx \c_\a(x)\tilde{\c}_{\a}(x) \r_0(x)
420: \approx \i dx \c_\a(x) \r_0(x)$
421: giving the
422: population of the metastable state $\a$ in the equilibrium distribution
423: $\r_0(x)$.
424:
425: The time evolution of the probabilities $p_\a(t)$ now follows from the
426: projection of the Fokker-Planck equation (\ref{FP})
427: onto the slow subspace, i.e. we
428: assume the probability density to take the form (\ref{sr}) and determine
429: the scalar product with $\c_\a(\bx)$ on both
430: sides of the Fokker-Planck equation.
431: This
432: yields
433: \be
434: \frac{d}{d t} p_\a(t) = \sum_{\a'}
435: \frac{ \left ( \c_a,L \tilde{\c}_{\a'} \r_0 \right ) } {n_{\a'}} p_{\a'}(t)
436: \ee{pme}
437: Under the sum on the right hand side
438: we expressed the coefficients $c_{\a'}(t)$ by means of
439: eq.~(\ref{pn}) in terms of the
440: probabilities $p_{\a'}(t)$ leading to the equilibrium weights $n_{\a'}$ in the
441: denominator of the respective coefficient.
442: For $\a \neq \a'$ the coefficient $\left ( \c_\a, L \tilde{\c}_{\a'}
443: \r_0 \right )/n_{\a'} $ coincides with the Rayleigh-quotient
444: expressions for the rate from $\a'$ to $\a$ \cite{t94}.
445: The integral is dominated by a
446: neighborhood of the saddle
447: point that connects the domains of attraction $\mathcal{D}_{\a'}$ and
448: $\mathcal{D}_\a$ \cite{c1}. Within this neighborhood, for small noise,
449: the localizing function $\c_\a(\bx)$ can be
450: approximated by the characteristic function of $\cal{D}_\a $,
451: $\c_\a^{0}(x)$ \cite{t95}.
452: The resulting coefficient then
453: coincides with the better known flux-over-population expression for
454: the respective rate \cite{htb}. In particular,
455: one can show that for sufficiently weak noise the expression
456: $\left ( \c_\a, L \tilde{\c}_{\a'}\r_0 \right )/n_{\a'} $ is positive
457: for $\a \neq \a'$.
458: From eq.~(\ref{s1}) it
459: follows that the sum of the coefficients over
460: $\a$ vanishes. Hence, eq.~(\ref{pme}) has the proper form of a master
461: equation:
462: \be
463: \frac{d}{d t} p_\a(t) = \sum_{\a' \neq \a} r_{\a,\a'} p_{\a'}(t) -
464: \sum_{\a' \neq \a} r_{\a'\!,\a}\: p_\a(t)
465: \ee{me}
466: where
467: \be
468: r_{\a,\a' } = \frac{ \left ( \c_\a, L \tilde{\c}_{\a'} \r_0 \right ) }{n_{\a'}}
469: \ee{rab}
470: is the transition probability per unit time from $\a'$ to $\a$.
471:
472: The regime of validity of this approximation, of course, depends on the
473: particular system and the degree of accuracy one would like to achieve.
474: By its very nature this master equation
475: is only appropriate for the long time dynamics of the
476: Fokker-Planck equation and therefore requires a clear separation of time
477: scales dividing the process into fast relaxations within the domains of
478: attraction and slow transitions between the attractors. We note that various
479: definitions of transition rates exist
480: which are different from a physical point of
481: view
482: but which still yield the same result if a master equation provides a
483: correct description of the long time dynamics. It turns out that the actual
484: differences between these rate expressions are exponentially small in the
485: noise strength, i.e. of the form $\exp \left \{ -\D V/k_B T \right \}$ where
486: $\D V $ is a relevant barrier hight and $k_B T$ the noise strength. So if we
487: are ready to accept an error of, say $1 \%$, all barriers in the system must
488: be higher than approximately $4.5 k_B T$.
489:
490:
491: \section{Slow driving}\label{III}
492: After these preparatory considerations we turn to our main topic and
493: consider a Fokker-Planck process of the kind of the previous section with the
494: only difference that we now allow for slow changes of the parameters of the
495: system. These could result from a slow drift of the environmental parameters
496: of the system like e.g. the temperature, a slow increase of an external field,
497: or a slow periodic force driving the system, to name but a few
498: realizations.
499:
500: We emphasize that the change of the parameters must be slow compared to
501: the fast relaxation within each domain of attraction but no assumption is
502: made how it compares to the slow intrinsic dynamics of the system describing
503: the transitions between different domains of attraction.
504: We assume that in the presence of these time dependent
505: parameters one still can
506: identify domains of attraction which do not merge or split as time changes
507: but retain their identity. In other words, we restrict ourselves to changes
508: of the parameters that do not lead to topological changes of the system.
509:
510: The resulting process is still Markovian and is governed by a
511: Fokker-Planck equation:
512: \be
513: \frac{\partial}{\partial t} \r(t) = L(t) \r(t)
514: \ee{FPT}
515: where the Fokker-Planck operator $L(t)$ is of the form given in eq.~(\ref{FPO})
516: with time dependent drift $\bK(\bx,t)$ and
517: possibly also a time dependent diffusion
518: matrix $\bD(\bx,t)$.
519: In particular, we assume that all parameter values which the system
520: experiences in the course of time correspond to equilibrium systems, i.e.
521: that for any frozen parameter value the system will reach an equilibrium
522: distribution
523: relative to which the system obeys detailed balance:
524: \be
525: L(t) \hat{\r}_0(t) = \hat{\r}_0(t) \tilde{L}^+(t)
526: \ee{fdb}
527: where the tilde denotes time reversal and $\hat{\r}_0(t)$ is the
528: multiplication operator with the equilibrium distribution
529: $\r_0(\bx,t)$
530: that would be attained if the parameter values were frozen at the
531: values that they assume at time $t$:
532: \be
533: L(t) \r_0(\bx,t) =0.
534: \ee{L0}
535:
536:
537: For vanishing noise the system will move toward the nearest attractor
538: as defined by the momentary values of the parameters. During this relaxation
539: time, the
540: parameters of the system are supposed to change by such a small amount
541: that they can be
542: considered as kept constant. Only when the system has reached the attractor
543: and stays there for a much longer time,
544: the change of parameters will become noticeable and the
545: system will follow the slow
546: motion of the attractor as it results from the changing parameters.
547: This means that the time rate of change of the deterministic drift is
548: characterized by a small parameter $\e$:
549: \be
550: \frac{|\dot{\bK}(\bx,t)|}{| \bK(\bx,t)|} \t_r= {\cal O}(\e)
551: \ee{e}
552: where $\t_r$ is a typical relaxation time of the deterministic motion.
553: Under this very condition many of the properties of the time homogeneous
554: process carry over to the case with time dependent parameters.
555:
556: To each attractor $\a$ then there belongs a domain of attraction
557: ${\cal D}_\a(t)$ together with its characteristic function
558: $\c^{(0)}_\a(\bx,t)$ which is
559: $1$ on ${\cal D}_\a(t)$ and $0$ elsewhere in $\S$. The domains of attractions
560: also slowly depend on time and so do the characteristic functions.
561: If the system
562: is weakly perturbed by noise the characteristic functions will change into
563: smooth functions $\c_\a(\bx,t)$
564: that coincide with the noiseless functions everywhere
565: except in a thin boundary layer at $\partial {\cal D}_\a(t)$, where a steep
566: but smooth transition from $1$ to $0$ takes place. As in the time
567: independent case, they are solutions of the homogeneous backward equation that
568: approach $1$ in the interior of ${\cal D}_\a(t)$ and vanish on all other
569: attractors $\a' \neq \a$.
570: Further one also can represent the slow
571: dynamics of the probability
572: density in terms of linear
573: combinations of ${\tilde \c}_\a(\bx,t) \r_0(\bx,t)$:
574: \be
575: \r(\bx,t) = \sum_{\a =1}^m c_\a(t) \tilde{\c}_\a(\bx,t) \r_0(\bx,t)
576: \ee{srt}
577: The probabilities to find the system in the state $\a$ is
578: analogously defined as in the case with constant parameters:
579: \be
580: p_\a(t) = \left (\c_\a(t), \r(t) \right )
581: \ee{patt}
582:
583: The time rate of change of $p_\a(t)$ consequently has two contributions,
584: resulting from the derivative of $\c_\a(\bx,t)$ and of $\r_0(\bx,t)$:
585: \be
586: \frac{d}{dt} p_\a(t)= \left ( \frac{\partial}{\partial} \c_\a(t), \r(t) \right ) +
587: \left ( \c_\a(t), \frac{\partial}{\partial t}\r(t) \right ).
588: \ee{dpa}
589: The latter probabilistic contribution
590: can be expressed in terms of the Fokker-Planck equation yielding
591: \be
592: \left ( \c_{\a}, \frac{ \partial }{ \partial t} \r(t) \right )
593: = \sum_{\a'} r_{\a,\a' }(t) p_\a (t)
594: \ee{pt1}
595: where we used eq.~(\ref{srt}) and introduced the time dependent
596: transition
597: rates $r_{\a,\a' }(t)$ in an analogous way as in the time homogeneous
598: case, see eq.~(\ref{rab}):
599: \be
600: r_{\a,\a' }(t) = \frac{ \left ( \c_\a(t), L(t)
601: \tilde{\c}_{\a'} (t) \r_0(t) \right )}{
602: n_{\a'} (t) }.
603: \ee{rt}
604: It determines the rate of change of the probability $p_\a(t)$ caused by
605: transitions from $\a'$ to $\a$. Here $n_{\a'}(t)$ is the population of the
606: metastable state $\a'$ in the frozen equilibrium distribution $\r_0(\bx,t)$. It
607: is given by:
608: \be
609: n_\a(t) = \left ( \c_\a(t), \tilde{\c}_\a(t) \r_0(t) \right ) \approx
610: \left ( \c_\a(t), \r_0(t) \right )
611: \ee{nat}
612: where the term $\tilde{\c}_\a(\bx,t)$ is neglected in the second
613: equation. This approximation
614: holds up to exponentially small terms in the noise
615: strength.
616:
617: The other contribution, $\left (\partial \c_\a (t) / \partial t,\r(t) \right
618: )$,
619: describes the change of the probability caused by the geometric
620: change of the domain ${\cal D}_\a(t)$. It also is linear in the probabilities
621: $p_{\a'} (t)$ and can be written as
622: \be
623: \left (\frac{\partial}{ \partial t} \c_\a(t), \r(t) \right ) = \sum_{\a'}
624: g_{\a,\a' }(t) p_{\a'} (t)
625: \ee{pt2}
626: where
627: \be
628: g_{\a,\a' }(t) = \frac{ \left ( \frac{\partial}{ \partial t} \c_\a(t),
629: \tilde{\c}_{\a'} (t) \r_0(t) \right )}{ n_{\a'} (t) }
630: \ee{g}
631: The coefficients $g_{\a,\a' }(t)$ are proportional to the smallness parameter
632: $\e$ because they contain the time derivative of $\c_\a(x,t)$. The proportionality
633: factor roughly is of the same order of magnitude as the transition rate
634: $r_{\a,\a'}(t)$:
635: \be
636: \frac{ g_{\a,\a' }(t)}{ r_{\a,\a' }(t)} = {\cal O}(\e)
637: \ee{gr}
638: The dependence of this ratio on the noise strength is discussed for a
639: particular example in section~\ref{V}.
640: For a sufficiently slow parameter change, i.e a small value of $\e$,
641: the geometric contribution to the time rate of change of the
642: probability $p_\a(t)$ can be neglected and a master equation results that
643: is completely determined by the instantaneous rates $r_{\a,\a'}(t)$:
644: \be
645: \frac{d}{d t} p_\a(t) =
646: \sum_{\a'\neq \a} r_{\a,\a'}(t) p_{\a'}(t)-\sum_{\a'\neq \a}
647: r_{\a',\a}(t)
648: p_{\a}(t)
649: \ee{met}
650: where we used that the time dependent rates $r_{\a,\a'}(t)$ (\ref{rt})
651: have the same
652: general form and therefore
653: the same formal properties
654: as the time independent ones (\ref{rab}):
655: Those for $\a \neq \a'$ are positive and the
656: sum over the first index vanishes: $\sum_\a r_{\a,\a'}(t)= 0$.
657: This master equation with time dependent rates is the central result
658: of the present paper.
659: For faster driving the geometric contributions of the rates have to be taken
660: into account. However, they may become negative, and therefore formally
661: negative rates may result if the driving is too fast. This indicates that then
662: the instananeous eigenfunctions of the Fokker Planck equation no longer
663: provide a good basis.
664:
665:
666:
667:
668:
669:
670: \section{Decorating the meta-stable states}\label{IV}
671: Often the knowledge of the probabilities $p_\a(t)$ to find the system at time
672: $t$ in the meta-stable state $\a$ is not sufficient. For example one
673: may be interested in the average position $\bx$ of the system or its
674: single- or multi-time statistical properties.
675: We here show how the time dependence of these quantities on the slow time
676: scales of the transitions
677: and the external driving can be retrieved from exactly the same information
678: that is necessary to determine the transition rates from the Fokker-Planck
679: equation. For that purpose we first
680: consider the average of an arbitrary function
681: $f(\bx)$
682: of the position:
683: \ba
684: \langle f(t) \rangle & = & \i d^n \bx f(\bx) \r(\bx,t) \nonumber \\
685: & = & \sum_\a \langle f(t) |\a \rangle p_\a(t)
686: \ea{fa}
687: where $\langle f(t) |\a \rangle$ denotes the expectation value of $f(x)$ under
688: the condition that the system resides at $t$ in the domain of attraction ${\cal
689: D}_\a(t)$. Using the long-time behavior of the probability density $\r(t)$ as
690: given by eq.~(\ref{srt}) one finds:
691: \be
692: \langle f(t) |\a \rangle = \i d^n \bx f(\bx) \r(\bx,t|\a)
693: \ee{rxta}
694: where the conditional probability $\r(\bx,t|\a)$ to find the system in the
695: continuous state $ \bx \in {\cal D}_\a(t)$ is given by
696: \be
697: \r(\bx,t|\a) = \frac{\tilde{\c}_\a(\bx,t)}{ n_\a(t)} \r_0(\bx,t)
698: \ee{rxtae}
699: Similarly, one finds for the correlation of two functions $f(\bx)$ and $g(\bx)$
700: at times $t$ and $t +\t$ being separated by a positive time $\t$ that is long
701: compared to the
702: fast time scale of relaxations within the domains of attraction the
703: following expression:
704: \be
705: \langle f(t+ \t) g(t) \rangle = \sum_{\a,\a' } \langle f(t+\t)|\a \rangle\:
706: \langle g(t) |\a' \rangle\: p(\a, t+\t | \a', t )\: p_{\a'}(t).
707: \ee{fg}
708: Here $p(\a,t+\t|\a',t)$ is the conditional probability to find the
709: system in the metastable state $\a$ at time $t+\t$ provided it was in the
710: state $\a'$ at the earlier time time $t$. This conditional probability is the
711: solution of the master equation (\ref{met}) subject to the initial
712: conditions $p(\a, t| \a', t) = \d_{\a,\a'}$. Eq.~(\ref{fg}) holds for
713: arbitrary functions $f(\bx)$ and $g(\bx)$. Therefore
714: the conditional probability of
715: finding the system at time $t+\t$ at the continuous state $x$ if it was at
716: the earlier time $t$ at $y$ takes the form
717: \be
718: \r(\bx,t+\t|\by,t) = \sum_{\a,\a'} \r(\bx,t+\t|\a) \:\c_{\a'}(\by,t) \:
719: p(\a,t+\t|\a',t).
720: \ee{cpx}
721: For an independent derivation of this result see the Appendix~\ref{B}.
722:
723: Yet another analogous way exists
724: to characterize the continuous process on the long time scale:
725: One takes the discrete process $z(t)$ assuming the values
726: $\a = 1\ldots m$
727: according to the master equation (\ref{met}) and decorates the
728: states with a random point in the state space,
729: $\bX(z(t),t)$, depending on time and the
730: particular state $\a=z(t)$ that is realized at the time $t$.
731: If the probability density of the decoration $\bX(\a,t)$ is chosen as the
732: conditional
733: probability density $\r(\bx,t|\a)$ given in eq.~(\ref{rxtae}) the mean values
734: and the correlation functions of the process
735: $\bX(\bx,t)$
736: coincide with the expressions (\ref{fa})
737: and (\ref{fg}), respectively, and accordingly are characterized by the
738: conditional probability (\ref{cpx}). The decorated process is
739: Markovian and has the same single and two time probability density as
740: the continuous process $\bx(t)$ on the long time scale and hence
741: coincides there
742: with it:
743: \be
744: \bx(t)= \bX(z(t),t)
745: \ee{xXz}
746:
747: \section{Bistable overdamped oscillator}\label{V}
748: We here consider the example of an overdamped bistable Brownian oscillator
749: that is periodically driven at a frequency $\O$ which is slow compared to the
750: typical relaxation rates. Its dynamics is given by the Smoluchowski operator
751: \be
752: L_S(t) = \frac{\partial}{ \partial x} \frac{\partial V(x,t)}{ \partial x} +
753: \b^{-1} \frac{\partial^2}{ \partial x^2}
754: \ee{LS}
755: where $\b^{-1}$ is the noise strength which is proportional to the
756: temperature $\th$ of the fluid surrounding the oscillator,
757: and $V(x,t)$ denotes the time-dependent potential:
758: \be
759: V(x,t) = \frac{1 }{ 4} x^4 - \frac{1 }{ 2} x^2 - A x \sin(\O t).
760: \ee{Vxt}
761: Here $A$ denotes the strength of the periodic force. Throughout dimensionless
762: units are used.
763: It is supposed to stay within the limits $|A| < 2/(3 \sqrt{3) }$. Actually
764: it must keep some finite distance from these limits in order that the
765: following asymptotic theory applies. Under this condition on the strength of
766: the external force, the potential
767: has three stationary points that are solutions of the algebraic equation:
768: \be
769: x^3 -x = A \sin(\O t)
770: \ee{V'}
771: As time varies they form three branches: $x_1(t)$ and $x_2(t)$ are those
772: tracing the two local minima that at $t=0$ assume the values $x_1(0) = -1$
773: and $x_2(0) = 1$, respectively, and $x_b(t)$ gives the location of the local
774: potential maximum between the minima. At $t=0$ it is located at $x_b(0) =0$.
775: The corresponding extreme values of the potential $V(x,t)$
776: are denoted by:
777: \ba
778: V_\a(t) & = & V(x_\a(t),t) \qquad \mbox{for} \;\a=1,2 \nonumber \\
779: V_b(t) & = & V(x_b(t),t)
780: \ea{Vpmb}
781: The instantaneous well and barrier frequencies are defined accordingly as:
782: \ba
783: \o_\a(t) = & \sqrt{\frac{ \partial^2 V(x_\a(t),t)}{ \partial x}} = & \sqrt{3
784: x_\a^2(t) -1} \qquad \mbox{for} \; \a=1,2\nonumber \\
785: \o_b(t) = & \sqrt{\frac{ -\partial^2 V(x_b(t),t)}{ \partial x}} = & \sqrt{1-3
786: x_b^2(t)}
787: \ea{opmb}
788: For later use we give the time derivatives of the barrier position and the
789: barrier
790: frequencies. For the position one finds from eq.~(\ref{V'}):
791: \be
792: \dot{x}_b (t) = -\frac{ \O A}{ \o_b(t)^2} \cos(\O t)
793: \ee{dx}
794: and for the frequency it follows:
795: \be
796: \dot{\o}_b(t) = -3\frac{ x_b(t)}{ \o_b(t)} \dot{x}_b(t)
797: \ee{do}
798: The instantaneous stationary solution of the Smoluchowski equation
799: $L_S(t) \r_0(x,t) = 0$ is given by the Boltzmann distribution:
800: \be
801: \r_0(x,t) = Z^{-1}(t) e^{-\b V(x,t)}
802: \ee{rB}
803: where for small noise the partition function is given by
804: \be
805: Z(t) = \sqrt{\frac{2 \p }{ \b}} \left \{
806: \frac{ \exp \left \{-\b V_1(t)\right \} }{
807: \o_1} +
808: \frac{\exp \left \{-\b V_2(t)\right \}
809: }{ \o_2} \right \}
810: \ee{Z}
811: Here algebraic correction terms of the order ${\cal O}(1/(\b \Delta
812: V_\a(t)))$ were neglected.
813: The domains of attraction of the instantaneous locally stable states
814: $x_\a(t)$, $\a = 1,2$,
815: extend from minus infinity to the instantaneous barrier and
816: from there to infinity: ${\cal D}_1(t) = ( -\infty,x_b(t))$, and
817: ${\cal D}_2(t) = ( x_b(t), \infty) $.
818:
819: The corresponding localizing functions $\c_\a(x,t)$
820: are the solutions of the backward equation
821: \be
822: \left \{\o_b(t)^2 (x-x_b(t)) \frac{\partial}{ \partial x} + \b^{-1}
823: \frac{\partial^2}{ \partial x^2} \right \} \c_\a(x,t) = 0
824: \ee{ca}
825: with boundary conditions:
826: \be \begin{array}{c}
827: \c_1(x,t) = \left \{\begin{array}{ll}
828: 1 \quad &\mbox{for}\;\; x \to - \infty \\
829: 0 &\mbox{for}\; \; x \to \infty
830: \end{array} \right .\\
831: \\
832: \c_2(x,t) = \left \{\begin{array}{ll}
833: 0 \quad &\mbox{for}\;\; x \to - \infty \\
834: 1 &\mbox{for}\; \; x \to \infty
835: \end{array} \right .
836: \end{array}
837: \ee{c12}
838:
839: The solutions readily are found:
840: \ba
841: \c_1(x,t)& = & \frac{1 }{ 2} \mbox{erfc}(\o_b(t)\sqrt{\b/2}(x-x_b(t)))
842: \nonumber \\
843: \c_2(x,t)& = & 1 - \c_1(x,t)
844: \ea{cc}
845:
846: In the present model there is only a single variable, the coordinate
847: $x$
848: which transforms evenly under time reversal, hence, $\tilde{\c}_\a(x,t)
849: = \c_\a(x,t)$.
850: Using the eq.~(\ref{cc}), the scalar products
851: $\left ( \c_\a(t), L_S(t) \c_{\a'}(t) \r_0(t) \right )$, $\a,\a' = 1,2$ can be
852: expressed by a single one that we denote by $q(t)$:
853: \ba
854: \left (\c_1(t), L_S(t) \c_2(t) \r_0(t) \right )& = &
855: \left (\c_2(t), L_S(t) \c_1(t) \r_0(t) \right ) = \nonumber \\
856: -\left (\c_1(t), L_S(t) \c_1(t) \r_0(t) \right )& = &
857: -\left (\c_2(t), L_S(t) \c_2(t) \r_0(t) \right ) = q(t)
858: \ea{cq}
859: where $q(t)$ takes the form:
860: \be
861: q(t) = \b^{-1} \i dx \left ( \frac{\partial \c_1(x,t)}{ \partial x}
862: \right )^2 \r_0(x,t)
863: \ee{qt}
864: Using eq.~(\ref{cc}) for weak noise it can further be simplified as
865: \be
866: q(t) = \frac{\o_b(t) e^{-\b V_b(t)}}{ Z(t) \sqrt{2 \p \b}}
867: \ee{q}
868:
869: The populations $n_\a(t)$ of the metastable states
870: $\a=1,2$ in the frozen equilibrium distribution become at weak noise:
871: \be
872: n_\a(t) = \sqrt{\frac{2 \p }{ \b}} \frac{ e^{-V_\a(t) \b}}{ \o_\a(t) Z(t)}
873: \ee{n12t}
874: Using eq.~(\ref{rt}) this gives for the rates the instantaneous
875: expressions \cite{htb}:
876: \ba
877: r_{2,1}(t) = & - r_{1,1}(t) = & \frac{\o_1(t) \o_b(t)}{ 2 \p} e^{-\b \D V_1(t)}
878: \nonumber \\
879: r_{1,2}(t) = & - r_{2,2}(t) = & \frac{\o_2(t) \o_b(t)}{ 2 \p} e^{-\b \D V_2(t)}
880: \ea{r12t}
881: \subsection{The geometric correction to the rate}\label{Va}
882: Now we come to the discussion of the corrections of the instantaneous rates
883: that are determined by the
884: geometric contributions $\left ( \partial \c_\a(t)/\partial t,
885: \c_{\a'}(t) \r_0(t) \right )$, $\a, \a' = 1,2$.
886: For the relative magnitude of the geometric correction to the
887: instantaneous rate $r_{2,1}(t)$ one finds after some algebra:
888: \ba
889: \e(t)& \equiv & \frac{g_{2,1}(t)}{r_{2,1}(t)} =
890: \frac{\left ( \partial \c_2(t)/ \partial t,\c_1(t) \r_0(t) \right
891: )}{\left (\c_2(t), L_S(t) \c_1(t) \r_0(t)\right )} \nonumber \\
892: & = & \ - \frac{\b}{2}
893: \dot{x}_b(t)
894: \i_{-\infty}^\infty dy \left (\frac{3 x_b(t)}{\o_b^2(t)} y +1 \right )
895: \nonumber \\ & &\times
896: \mbox{erfc} \left ( \o_b(t) \sqrt{\b/2} y \right )
897: \exp \left \{ -\b \left ( y^4/4 +
898: x_b(t) y^3 \right ) \right \}
899: \ea{g21}
900: As expected, the error is proportional to the driving frequency
901: $\O$ via the time derivative $\dot{x}_b(t)$.
902: The remaining $\O$-dependence may be absorbed in a rescaled time
903: $\t= \O t$.
904: An analytic solution of the integral in eq.~(\ref{g21}) is not known.
905: In Fig.~\ref{f2}
906: the error divided by the driving frequency
907: is shown as it results from a numerical evaluation for different values
908: of the driving strength $A$ and inverse temperature as a function of
909: time. Both increasing driving strength $A$ and inverse temperature
910: $\b$ lead to an increase of the extrema of the error that must be
911: compensated by a smaller driving frequency in order that the master
912: equation with the instantaneous rates provides a valid description.
913: \begin{figure}
914: \begin{center}
915: \includegraphics[width=8cm]{semifig2a.eps}
916: \hfill
917: \includegraphics[width=8cm]{semifig2b.eps}
918: \end{center}
919: \caption{The relative error of the rate oscillates as a function of
920: time $\t = \O t$. It vanishes at the extrema of the driving force. Panel
921: (a) shows $\e(\t)/\O$ for different inverse temperatures $\b = 30,\:
922: 60, \: 90$ at the
923: driving strength $A = 0.15$ and panel (b) for different driving
924: strength $A = 0.1,\: 0.15,\: 0.2$ at $\b = 60$. The parameter
925: values are indicated close to the respective curves.}
926: \label{f2}
927: \end{figure}
928:
929: As an average measure we introduce the root mean square error
930: $E= \left ( \i_0^T dt \e^2(t)/ T \right )^{1/2}$ which is strictly
931: proportional to the driving frequency $\O$.
932: The proportionality factor $E/\O$ increases as a function of the
933: inverse temperature first as a power law and changes to an exponential
934: growth for large values of $\b$, see Fig~\ref{f3}.
935: \begin{figure}
936: \includegraphics[width=8cm]{semifig3.eps}
937: \caption{The asymptotic exponential growth of the frequency
938: independent ratio $E/\O$ is shown for the driving strength
939: $A=0.12$. The straight line corresponds to a barrier height $0.97 \times
940: 10^{-3}$.}
941: \label{f3}
942: \end{figure}
943:
944:
945: We propose the
946: following physical
947: explanation of this effect: At large $\beta$ the noise is very small
948: such that a particle that has reached at some time the top of the
949: barrier has a very small chance to be pushed to either side by the
950: noise. During the same time when it resides there, the potential moves
951: such that the particle no longer will sit on top of the
952: barrier. Depending on the direction of the motion of the potential the
953: particle may now be on the opposite side of the barrier or on the
954: original side where it came from. In the former case the geometric
955: correction leads to an increase of the rate and in the latter case to
956: a decrease of the transition rate. Because it then has again to
957: overcome a barrier the increase of the rate is exponentially large in
958: the inverse temperature.
959:
960: Finally we note that the deviation from the frozen
961: rate on the noise strength depends on the particular model.
962: For example in symmetric systems
963: with varying barrier height the deviation only grows with the square
964: root of
965: the inverse temperature rather than exponential as in the above case
966: where both the location and the height of the barrier changes in time.
967:
968:
969: \subsection{Average motion}\label{Vb}
970: We have projected the dynamics of a
971: a process with multiple metastable states
972: and slowly time dependent parameters onto the slow subspace defined by
973: the eigenfunctions of Fokker-Planck operator
974:
975: \begin{figure}[b]
976: \includegraphics[width=10cm]{semifig4.eps}
977: \caption{The response of the average position on the driving force for
978: the driving strength $A=0.05$ and driving frequency $\O = 10^{-5}$
979: strongly depends on the magnitude of the noise. For $\beta = 35$ the
980: noise is relatively large such that the average position closely
981: follows its adiabatic value and therefore shows almost no
982: hysteresis;
983: for $\beta = 48$ it stays behind the
984: force, and for $\beta = 75$ it is hardly influenced by the force. The
985: hysteresis curves are traced in the mathematically positive sense.}
986: \label{f4}
987: \end{figure}
988: As a simple example of the decoration of the two metastable states we
989: come back to the bistable oscillator and
990: consider the asymptotic motion of its average position. From
991: eq.~(\ref{fa}) it results as
992: \be
993: \langle x(t) \rangle = \langle x(t)|1 \rangle p_1(t) +
994: \langle x(t)|2 \rangle p_2(t)
995: \ee{xt}
996: where $\langle x(t) | \a \rangle $ denotes the conditional average of
997: the position in the well corresponding to the metastable state $\a$:
998: \ba
999: \langle x(t)|\a \rangle & = & \i dx \:x \frac{\c_\a(x,t)}{n_\a(t)}
1000: \r_0(x,t) \nonumber \\
1001: & \approx & x_\a(t)
1002: \ea{xta}
1003: In the last equation we disregarded small contributions of the order
1004: $\mathcal{O}(1/(\o_\a(t)\sqrt{\b}))$. The probabilities $p_\a(t)$
1005: follow as the asymptotic solutions of the master equation:
1006: \ba
1007: p_1(t)& = & \frac{\i_0^T ds \:e^{(K(T)-K(s))} r_{1,2}(t)}{1- e^{K(T)}}\:
1008: e^{-K(t)} +
1009: \i_0^t ds \:e^{-(K(t) -K(s) )} r_{1,2}(s) \nonumber \\
1010: p_2(t) & = & 1 - p_2(t)
1011: \ea{p1}
1012: where
1013: \be
1014: K(t) = \i_0^t ds \left ( r_{1,2}(t) + r_{2,1}(t) \right )
1015: \ee{K}
1016:
1017:
1018:
1019: Rather than considering the time dependence we here show in
1020: Fig.~\ref{f4} the dependence of
1021: the position on the driving force $F(t) = A \cos (\O t)$. In general
1022: the position lacks behind the force and therefore the dependence is
1023: hysteretic. Only if the driving is either very slow or very fast
1024: compared to the average rates there is no hysteretic behavior.
1025: In the slow case many transitions
1026: occur before a change of the potential becomes sensible. The
1027: asymptotic probabilities $p_\a(t)$ take the form of the stationary
1028: probabilities for the frozen rates.
1029: They then depend on the force but not on its rate of change.
1030: In the
1031: opposite limit the asymptotic probabilities see only the average rates
1032: and hence are time independent.
1033:
1034: \section{Summary}\label{VI}
1035: In this paper we studied the dynamics of externally driven systems
1036: with multiple metastable states in the semiadiabatic limit. The
1037: full dynamics of the system comprising the fast and the slow
1038: time scales of the intrawell
1039: relaxations and of the external driving as well as of
1040: the transitions between the metastable states, respectively,
1041: is supposed to be Markovian and continuous and, hence,
1042: described by a Fokker-Planck equation. We showed that the long
1043: time dynamics of this Fokker-Planck equation is equivalent to the
1044: dynamics of the master
1045: equation for the transitions between the discrete metastable states.
1046: The rates that determine this master equation are time dependent
1047: taking values as if the parameters were frozen at their instantaneous
1048: values. This kind of kinetic description is not new and
1049: has long been used in the literature \cite{McNW,ki}. The present work,
1050: however, provides three new aspects. First the master equation does
1051: not result from an educated guess.
1052: It
1053: rather is
1054: obtained as the result of the projection of the dynamics onto the slow
1055: subspace of the Fokker-Planck equation. This systematic approach
1056: additionally gives a quantitative criterion when deviations from
1057: the master equation with frozen rates must be expected. This criterion
1058: does explicitly take into account the time change of the parameters
1059: and in contrast to other works \cite{Ast} is not
1060: based on a comparison of time scales of the
1061: frozen dynamics with the rate of change of the parameters. As a
1062: particular result we found for the dynamics of a periodically driven
1063: overdamped Brownian oscillator that the maximal driving frequency
1064: depends on the noise strength and decreases with decreasing noise
1065: strength. Asymptotically this dependence even is exponential. Future work
1066: still has to show whether one can achieve an improvement of the
1067: dynamics if the so called geometric corrections to the rates are taken
1068: into account in
1069: cases when the time scale separation between the fast intrawell
1070: relaxation and the changes of the parameters becomes less pronounced.
1071: The third main outcome of the present approach is a partial retrieval
1072: of the underlying dynamics of the continuous variables by means of a
1073: proper decoration of the discrete states with time dependent random
1074: variables.
1075:
1076: The necessary time scale separation forced us to
1077: exclude the occurrence of bifurcations of the deterministic dynamics
1078: caused by the change of the parameters. Such bifurcations are
1079: accompanied by a slowing down of the deterministic dynamics
1080: and at the same time by a lowering of a barrier height and
1081: consequently by a decreasing time scale of transitions.
1082: Hence, for time dependent parameters,
1083: the time scale separation between interwell and intrawell
1084: dynamics is violated within a time window around the instant of bifurcation.
1085: It seems plausible that it should be possible to cut out this time
1086: window and to bridge it by some connection condition for the
1087: probabilities of those states that are involved in the bifurcation.
1088:
1089:
1090:
1091: Apart from the separation of the time scales, another
1092: assumption was made in the present work. First, we assumed that
1093: for fixed values of the parameters the system reaches a state of
1094: thermal equilibrium and hence it obeys detailed balance. We think
1095: that this assumption is not really essential for our results but it
1096: simplifies the analysis considerably. Without detailed balance more
1097: general limit sets of the dynamics can occur apart from limit points such as
1098: limit cycles and chaotic states. In general, the stationary
1099: densities are not known in absence of detailed balance, the
1100: weak noise asymptotics is plagued by notorious nonanalyticities
1101: \cite{gra}, and rather little is known about transition rates in
1102: nonequilibrium systems. On the other hand there are many important
1103: systems that are driven out of equilibrium by time dependent
1104: parameters.
1105:
1106: \acknowledgments
1107: The authors thank Igor Goychuk, Peter H\"anggi, Sigmund Kohler,
1108: Marcin Kostur and Michael Schindler
1109: for valuable discussions and hints. This work was supported by the
1110: Deutsche Forschungsgemeinschaft (SFB438).
1111:
1112:
1113:
1114:
1115:
1116: \appendix
1117: \section{The conditional probability at long times}\label{B}
1118: For a sufficiently long time lag $\t$ the conditional
1119: probability $\r(x,t+\t|y,t)$ of the
1120: continuous process $x(t)$ can be expressed in terms of the basis functions
1121: $\c_\a(y,t)$ and $\tilde{\c}_\a(x,t+\t) \r_0(x,t +\t)$
1122: spanning the slow subspaces of the
1123: backward operator at time $t$ and the forward operator at time $t+\t$,
1124: respectively. One therefore can write:
1125: \be
1126: \r(x,t+\t|y,t) = \sum_{\b,\b'} d(\b,\b';t,\t) \r(x,t+\t|\b) \c_{\b'}(y,t)
1127: \ee{qrc}
1128: where we expressed the basis function $\tilde{\c}_\a(x,t+\t) \r_0(x,t +\t)$
1129: by $\r(x,t+\t|\a) n_\a(t+\t)$, see eq.~(\ref{rxtae}) and introduced the yet
1130: undetermined
1131: coefficients $d(\a,\a';t,\t)$ into which the time dependent factor $n_\a(t)$
1132: is absorbed. We now multiply the conditional probability of
1133: the continuous process $\r(x,t+\t|y,t)$ with
1134: $\r(y,t|\a')$, integrate over the domain of attraction ${\cal D}_\a(t+\t)$
1135: and obtain the conditional probability of the discrete process
1136: $p(\a,t+\t|\a',t)$. Using the ansatz (\ref{qrc}) for $\r(x,t+\t|t)$
1137: we then find for the coefficients
1138: $d(\a,\a';t,\t) = p(\a,t+\t|,\a',t)$. Together with eq.~(\ref{qrc}) this
1139: yields the expression eq.~(\ref{cpx}) for the conditional probability
1140: of the continuous
1141: process as claimed in Sect.~\ref{IV}
1142:
1143:
1144:
1145:
1146:
1147:
1148: \begin{thebibliography}{99}
1149: \bibitem{D} P.A.M. Dirac, Nature {\bf 192}, 325 (1937); J.K. Webb et
1150: al., Phys. Rev. Lett. {\bf 87}, 091301 (2001).
1151: \bibitem{ae} C.M. Bender, S.A. Orszag, Advanced Mathematical Methods for
1152: Scientists and Engineers, McGraw-Hill, New York, 1978;
1153: A.H. Nayfeh, Perturbation Methods, John Wiley, New York, 1973;
1154: N. Krylov, N.N. Bogoliubov, Introduction to Nonlinear Mechanics,
1155: Princeton University Press, Princeton, 1947;
1156: H. Haken, Synergetics, An Introduction, Springer Verlag, Berlin
1157: 1983.
1158: \bibitem{htb} P. H\"anggi, P. Talkner, M. Borkovec,
1159: Rev. Mod. Phys. {\bf 62}, 251 (1990).
1160: \bibitem{t99} P. Talkner, New J. Phys. {\bf 1}, 4 (1999),
1161: (http://www.njp.org/).
1162: \bibitem{j} P. Jung, Phys. Rep. {\bf 234}, 175 (1993).
1163: \bibitem{ghjm} L. Gammaitoni, P. H\"anggi, P. Jung, F. Marchesoni,
1164: Rev. Mod. Phys. {\bf 70}, 223 (1998).
1165: \bibitem{r} P. Reimann, Phys. Rep. {\bf 361}, 57 (2002); R.D. Astumian,
1166: P. H\"anggi, Phys. Today {\bf 55}(11), 33 (2002).
1167: \bibitem{erm}E. Evans, K. Ritchie, Biophysics J. {\bf 72}, 1541
1168: (1997);
1169: R. Merkel, Phys. Rep. {\bf 346}, 343 (2001).
1170: \bibitem{Ry} D. Ryter, H. Meyer, Physica A {\bf 142}, 122 (1987).
1171: \bibitem{Mo} G. Moro, J. Chem. Phys. {\bf 103}, 7514 (1995).
1172: \bibitem{Ast} M. Bier, R.D. Astumian, Phys. Lett. A {\bf 247}, 385
1173: (1998);
1174: I. Derenyi, R.D. Astumian, Phys. Rev. Lett. {\bf 82}, 2623 (1999);
1175: M. Bier, I. Derenyi, M. Kostur, R.D. Astumian, Phys. Rev. E {\bf 59},
1176: 6422 (1999).
1177: \bibitem{lrh}
1178: J. Lehmann, P. Reimann, P. H\"anggi, Phys. Stat. Solodi B {\bf 237}, 53 (2003);
1179: J. Lehmann, P. Reimann, P. H\"anggi, Phys. Rev. E {\bf 62}, 6282 (2000);
1180: J. Lehmann, P. Reimann, P. H\"anggi, Phys. Rev. Lett. {\bf 84}, 1639 (2000).
1181: \bibitem{McNW} B. McNamara, K. Wiesenfeld, Phys. Rev. A {\bf 39}, 4854 (1989).
1182: \bibitem{Ri} H. Risken, The Fokker-Planck Equation, Springer Verlag, Berlin,
1183: 1989.
1184: \bibitem{MS} B.J. Matkowsky, Z. Schuss, SIAM J. Appl. Math. {\bf
1185: 42},822 (1982).
1186: \bibitem{t87} P. Talkner, Z. Phys. B {\bf 68}, 201 (1987).
1187: \bibitem{t94} P. Talkner, Chem. Phys. {\bf 180}, 199 (1994).
1188: \bibitem{Ma} M. Mangel, SIAM J. Appl. Math. {\bf 36}, 544 (1979);
1189: D. Ryter, J. Stat. Phys. {\bf 49}, 122 (1987).
1190: \bibitem{c1} If there is more than one isolated saddle point on the common
1191: separatrix of the domains of attraction $\mathcal{D}_{\a'}$ and
1192: $\mathcal{D}_\a$ the contribution of each saddle point to the
1193: matrixelement in eq.~(\ref{pme}) can be
1194: determined separately. More complicated cases need an extra
1195: treatment, see Refs. \cite{htb,btez}.
1196: \bibitem{t95} P. Talkner, in P. Talkner, P. H\"anggi (ed.),
1197: {\it New Trends in Kramers' Reaction Rate
1198: Theory}, Kluwer Academic
1199: Publishers, Dordrecht, 1995, p. 47.
1200: \bibitem{btez} A.M. Berezhkovskii, P. Talkner, J. Emmerich,
1201: V.Y. Zitserman, J. Chem. Phys. {\bf105}, 10890, (1996).
1202: \bibitem{ki} R.D Astumian, B. Robertson, J. Chem. Phys. {\bf 91}, 4891
1203: (1989); U. Z\"urcher, C.R. Doering, Phys. Rev. E {\bf 47}, 3862
1204: (1993); I.A. Goychuk, E.G. Petrov, V. May, J. Chem. Phys. {\bf103},
1205: 4937 (1995);
1206: P. Talkner, Ann. Phys. (Leipzig), {\bf 9}, 741 (2000);
1207: R. Rozenfeld, J.A. Freund, A. Neiman,
1208: L. Schimansky-Geier, Phys. Rev. E {\bf 64}, 051107 (2001);
1209: R.D. Astumian, Appl. Phys. A {\bf 75}, 193 (2002); P. Talkner,
1210: Physica A, in press.
1211: \bibitem{gra} R. Graham, T. Tel, J. Stat. Phys. {\bf 35}, 729 (1984);
1212: H.R. Jauslin, Physica A {\bf 144}, 179 (1987).
1213:
1214:
1215: \end{thebibliography}
1216:
1217:
1218:
1219:
1220: \end{document}
1221:
1222:
1223: )