1: \documentclass[pre,epsfig,aps,twocolumn]{revtex4}
2:
3: \usepackage{epsfig}
4: \usepackage{fancyheadings}
5: \usepackage{amstext,amsmath,amssymb}
6:
7: \begin{document}
8:
9: \sloppy
10: \pagestyle{empty}
11:
12: \title{Reconstructing a Random Potential from its Random Walks}
13: \author{S. Cocco $^1$, R. Monasson $^2$}
14: \affiliation{$^1$
15: CNRS-Laboratoire de Physique Statistique de l'ENS, 24 rue Lhomond,
16: 75005 Paris, France\\
17: $^2$ CNRS-Laboratoire de Physique Th\'eorique de l'ENS, 24 rue Lhomond,
18: 75005 Paris, France}
19:
20: \begin{abstract}
21: The problem of how many trajectories of a random walker in a potential
22: are needed to reconstruct the values of this potential is
23: studied. We show that this problem can be solved by calculating
24: the probability of survival of an abstract random walker in a
25: partially absorbing potential. The approach is illustrated on the
26: discrete Sinai (random force) model with a drift.
27: We determine the parameter (temperature, duration of
28: each trajectory, ...) values making reconstruction as fast as
29: possible.
30: \end{abstract}
31:
32: \maketitle
33:
34: {\em Introduction.}
35: Random walks (RW) in random media have been intensively studied in the
36: past decades as a paradigm for out-of-equilibrium dynamics, and have
37: led to the discovery and understanding of
38: important dynamical effects as anomalous diffusion, ageing
39: ...\cite{discreteRF,revue}. Briefly speaking the issue is to determine the
40: statistical properties of the walker from the ones of the energy
41: potential. Much less attention has been devoted to the inverse
42: problem: given one (or more) observed RW(s) can we guess the potential
43: values?
44: This question naturally arises in biophysics where the use of
45: AFM, optical and magnetic tweezers make possible the
46: mechanical separation of single protein-protein complexes \cite{evans},
47: or the unfolding and refolding of single biomolecules\cite{fer04,Ess97,woo06}.
48: The observed dynamics the rupture of chemical bonds,
49: of folding/unfolding of nucleic acids,
50: or proteins can be modeled as a RW motion affected by thermal
51: noise, moving in a quenched potential determined by
52: the composition of the chemical bonds, or the sequence of
53: amino-- or nucleic--acids. Reconstructing the free
54: energy landscape of those processes is the object of current and
55: intense efforts \cite{hye03,evans,mb,ritort,woo06}.
56:
57: In this letter we show how the inverse RW problem
58: can be practically solved within the Bayesian inference framework and
59: address the crucial question of the accuracy of reconstruction.
60: In practice information can be accumulated either by increasing the
61: duration of one RW, or observing more than one RW, or combining the two.
62: We discuss the optimal procedure minimizing the total number of
63: data to be acquired, and show how this minimal amount of data can be
64: calculated from
65: the probability of survival of an abstract walker in a partially
66: absorbing potential. The approach is illustrated in detail on the celebrated
67: discrete random force (RF) model (Sinai model with non zero drift)
68: \cite{revue,discreteRF}.
69:
70: %{\em Related works.}
71: Inference is a key issue in information theory and
72: statistics \cite{bayes}, with applications in biology \cite{domany},
73: social science \cite{And57}, finance, ... A central question is the
74: so-called hypothesis testing problem: which one of two candidate
75: distributions is
76: likely to have generated a set of measured data? This question was
77: solved in the case of independent variables by Chernoff \cite{Che52},
78: and is the core issue of the asymptotic theory of inference
79: \cite{bayes}. Chernoff showed that the probability of guessing
80: the wrong distribution decreases exponentially with the size of the data
81: set \cite{Che52}. Large deviations techniques can be used to treat
82: the case of variables extracted from one recurrent
83: realization of a finite Markov chain \cite{Boz71,Dem98}; the present
84: work can be seen as an extension to many transient realizations of
85: an `infinite' chain.
86:
87:
88: {\em Random Force model.} For an illustration of the problem
89: consider the discrete, one dimensional RF model defined on the set
90: of sites $x=0,1,2,\ldots ,N$ \cite{discreteRF}. We start by choosing
91: randomly a set of dimensionless forces $f_{x}=\pm 1$ on each link
92: ($x,x+1$) with {\em a priori}
93: probability $P_0 =\prod_x \frac{1+ b\,f_x}2$ where $-1<b<1$ is
94: called tilt. This defines the values of the potential ${\bf V}$ on each site,
95: $V_x = - \sum_{y<x} f_y$ (by definition $V_0=0$).
96: An example of potential for $b=0.4$ is shown on Fig.~\ref{fig-pot}.
97:
98: After the quenched potential has been drawn
99: a random walker starts in $x=0$ at time $t=0$. The walker then
100: jumps from one site $x$ to one of its neighbors $x'=x\pm 1$ with
101: rate (probability per unit of time) $r_{\bf V}(x \to
102: x')=r_0\times e^{(V_x-V_{x'})/(2T)}$ to satisfy detailed balance
103: at temperature $T$; the attempt rate $r_0$ will be set to unity in the
104: following. Reflecting boundary conditions are imposed by
105: setting $V_{N+1}=V_{-1}=+\infty$. We register the sequence of
106: of positions up to some time $t_f$: ${\bf X}=\{ x(t), 0\le
107: t\le t_f\}$. Figure~\ref{fig-pot} shows five RWs ${\bf X}_\rho$,
108: $\rho=1,\ldots,5$ , each starting in the origin $x(0)=0$ and of equal
109: duration $t_f$ for a temperature $T=1$. The value of the temperature
110: strongly affects the dynamics \cite{revue}, and its relevance for
111: the inverse problem will be discussed later.
112:
113: Our objective is to reconstruct the potential over a region of the
114: lattice e.g. the value of the forces on some specific links
115: from the observation of RWs.
116: Within Bayes inference framework this can be done by maximizing
117: the joint probability of the potential ${\bf V}$
118: and of the observed RWs ${\bf X}_1,\ldots, {\bf X}_R$
119: over ${\bf V}$ \cite{bayes}.
120: $P$ is the product of the {\em a priori} probability of the potential,
121: $P_0$, times the likelihood of
122: the RWs given the potential, $L$. Since the RW is Markovian $L$ depends
123: only on the sets of total times $t_x$ spent on every site $x$,
124: and of the numbers of jumps $u (x\to x')$ from
125: $x$ to $x'$ over the set of RWs:
126: \begin{equation} \label{like}
127: L = \prod_{x,x'} e^{-t_x \;r_{\bf V}(x\to x')}
128: \;r_{\bf V}(x\to x') ^{u(x\to x')}
129: \end{equation}
130: where the product runs over all sites $x$ and their neighbors $x'=x\pm 1$.
131: Expressing the rates in terms of the forces and maximizing the joint
132: probability $P$ we obtain the most likely values for the forces:
133: $f_x=\mbox{sign}(h_x + \alpha)$ where $\alpha\equiv T\, \ln[
134: (1+b)/(1-b)]$ is a global `field'
135: coming from the {\em a priori} distribution $P_0$ and $h_x$ a local
136: contribution due to the likelihood $L$,
137: \begin{equation} \label{field}
138: h_x = 2T\sinh \big(\frac 1{2T}\big) \; ( t_{x+1} -t_x) + u (x\to x+1)
139: - u (x+1\to x) \ .
140: \end{equation}
141: Figure~\ref{fig-pot} (left, bottom) shows predictions made from
142: $R=1$ to $R=5$ RWs for the first 200 sites.
143: The duration $t_f$ of the RW is chosen
144: to be much larger than the mean first passage time
145: in $x=200$, and much smaller than the equilibration time
146: $t_{eq}\sim e^{bN/T}$. In this range the quality of prediction is essentially
147: independent of $t_f$ as will be discussed in detail below.
148: As expected the number of erroneous forces
149: decreases with increasing $R$ though atypical events may
150: produce flaws in the prediction. The analysis of these atypical RWs,
151: and how they lead to errors is the keystone of what follows.
152:
153:
154: \begin{figure}
155: \begin{center}
156: \vskip .7cm
157: \psfig{figure=./fig1.eps,height=4.7cm,angle=0}
158: \vskip .7cm
159: \caption{Left, top: Example of potential ${\bf V}$ obtained in the RF model
160: with tilt $b=0.4$ (size $N=1000$,
161: sites $x>200$ not shown here). Right: examples of RWs, numbered
162: from 1 to 5, in this potential at temperature $T=1$; plateaus are in
163: correspondence with the local minima of $V$. Here $\alpha \simeq 0.85$
164: (creep phase). Left, bottom: Predictions from the first $R$ RWs in the
165: right panel and (\ref{field});
166: impulses locate incorrectly predicted forces $f_x$ for $x\le
167: 200$. The number of erroneous forces decreases from 26 (for $R=1$) to
168: 0 ($R=5)$. Note the errors on sites $x_0\simeq
169: 100$ appearing when the fourth RW is taken into account; indeed this
170: atypical RW marks no pause in the local minimum in $x_0$. }
171: \label{fig-pot}
172: \end{center}
173: \end{figure}
174:
175: {\em Number of RWs necessary for a good reconstruction.}
176: Expression (\ref{like}) for the likelihood of the RWs
177: is true for any potential
178: ${\bf V}$ and can be geometrically interpreted as follows.
179: Given a set of RWs we extract a signal vector ${\bf S}$
180: whose components are: the times $t_x$ spent on
181: site $x$, the numbers $u(x\to x')$ of transitions
182: from site $x$ to site $x'$. When $R$ is large we expect ${\bf
183: S}$ to be extensive with $R$ and define the intensive signal
184: ${\bf s}={\bf S}/R$.
185: Similarly, to each potential ${\bf V}$ we associate a vector
186: ${\bf v}$ with components: minus the outgoing rate {\em i.e.} $ -\sum
187: _{x' (\ne x)} r_{\bf V} ( x\to x')$ for each site $x$, the logarithm of
188: the rate $r_{\bf V}(x\to x')$ for each pair of neighbors.
189: Then $L=\exp(R \;{\bf s}\cdot {\bf v})$ from (\ref{like})
190: where $\cdot$ denotes the scalar product. Maximizing the joint
191: probability $P=P_0\times L$ over the potential becomes equivalent,
192: in the large $R$ limit, to finding ${\bf v}$ with the largest
193: scalar product with the signal ${\bf s}$
194: \footnote{The irrelevance of the {\em a priori} distribution
195: in the asymptotic case of large data set is well-known \cite{bayes}
196: and can be checked for the RF model: the local field (\ref{field})
197: is extensive in $R$, while the global field $\alpha$ remains
198: finite.}.
199: It is natural to partition the space of signals into `Voronoi cells':
200: $C_{\bf v}$ is the set of ${\bf s}$ having a larger scalar product with
201: ${\bf v}$ than with any other potential ${\bf v}'$.
202: Bayes rule tells us that the most likely potential given an observed
203: signal ${\bf s}$ is the one attached to the cell in which ${\bf s}$ lies.
204:
205:
206: Consider now RWs taking place in a given potential ${\bf V}$.
207: From the law of large number the signal ${\bf s}$ is equal, in
208: the infinite $R$ limit, to
209: ${\bf s}^*_{\bf v}=\{t_x^*,u^*(x\to x')=t^*_x \, r_{\bf V} (x\to x')\}$ where
210: $t^*_x$ is the average sojourn time on site $x$
211: over RWs of duration $t_f$. As ${\bf s}^*_{\bf v} \in C_{\bf
212: v}$ \footnote{Let ${\bf v}'\ne {\bf v}$;
213: ${\bf s}^*_{\bf v}\cdot ({\bf v}-{\bf v}') = \sum _{x\ne x'}
214: u^* (x\to x') G( r_{{\bf V}'}(x\to x')/ r_{{\bf V}}(x\to x') )$ where
215: $G(z)=z-\ln z -1 >0$ for $z\ne 1$.
216: } reconstruction becomes flawless in the limit
217: of an infinite number of data as expected.
218: For large albeit finite $R$, ${\bf s}$ typically deviates
219: from ${\bf s}^*_{\bf v}$ by $O(R^{-\frac 12})$; finite deviations have
220: exponentially small--in--$R$ probabilities, $e^{-R \, \omega _{\bf V}
221: ({\bf s})}$, controlled by a
222: rate function $\omega _{\bf V} ({\bf s})$ \cite{Dem98}.
223: The probability to predict an erroneous potential is
224: the probability that the stochastic signal ${\bf s}$ does not
225: belongs to cell $C_{{\bf v}}$. This probability of error thus decays
226: exponentially with $R$ over a typical number of RWs
227: \begin{equation} \label{min}
228: R_c({\bf V}) = \big[\ \displaystyle{\min _{{\bf s} \notin C_{\bf v}}}
229: \ \omega _{\bf V} ({\bf s})\ \big]^{-1} \ ,
230: \end{equation}
231: where the minimum is taken over signals outside the `true'
232: cell. It depends on the temperature, the duration of the RW, ...
233:
234: As the RWs are independently drawn $\omega _{\bf V}$ is a convex
235: function of ${\bf s}$ \cite{Dem98}. The minimum in (\ref{min}) is thus
236: reached on the boundary between the true cell and
237: another, bad cell, say, $C_{\bar {\bf v}}$. The attached potential,
238: $\bar {\bf V}$, is the most `dangerous' one from the inference point of
239: view. RWs generated from ${\bf V}$ and $\bar {\bf V}$ are hardly told
240: from each other unless more than $R_c({\bf V})$ of them are observed.
241:
242: Assume $\bar {\bf V}$ is known. Then the boundary between $C_{\bf v}$
243: and $C_{\bar {\bf v}}$ is the set of signals ${\bf s}\perp {\bf v}
244: -\bar {\bf v}$. We deduce
245: \begin{equation}
246: \label{bound}
247: R_c({\bf V} ) = \big[\ \max _{\mu} \; \min _{{\bf s}}
248: \big( \omega _{\bf V} ({\bf s}) + \mu\, {\bf s}\cdot(
249: {\bf \bar v}-{\bf v}) \big)\ \big] ^{-1}\
250: \end{equation}
251: where the Lagrange multiplier $\mu\in [0;1]$ ensures that ${\bf s}$ is confined
252: to the boundary.
253: The Legendre transform of $\omega _{\bf V}$ appearing in (\ref{bound})
254: is intimately related to the evolution operator of an abstract
255: random walk process, denoted by RW$(\mu)$
256: to distinguish from the original RW \cite{noi}. This RW$(\mu)$-er
257: moves with the rates
258: $r_{(1-\mu) {\bf V} + \mu \bar {\bf V}} (x \to x')$ and may die
259: on every site $x$ with positive rate
260: \begin{eqnarray} \label{death}
261: &&d_{{\bf V},\bar {\bf V},\mu} (x)
262: =\sum _{x' (\ne x)} \big[ (1-\mu)\, r _{\bf V} (x\to x') +
263: \mu \, r_{\bf \bar V}(x\to x') \nonumber \\
264: &&\hskip 3cm -\ r_{(1-\mu) {\bf V} + \mu {\bf \bar V}}(x\to x')\big] \ .
265: \end{eqnarray}
266: Consider now the probability $\pi(\mu)$
267: that RW$(\mu)$-er, initially at the origin,
268: has survived up to time $t_f$ (the duration of the original RW).
269: Then
270: $R_c({\bf V}) =\displaystyle{\min _{\mu \in[0;1]} 1/| \ln \pi(\mu)|}$.
271:
272: {\em Optimal Working Point for the RF model.}
273: We apply the above theory to the discrete RF model, and want to predict
274: the value of the force $f_y$ on the link $(y,y+1)$ for some specific
275: $y$. The dangerous potential is ${\bf \bar V}$
276: obtained from ${\bf V}$ upon reversal of the force $f_y\to -f_y$.
277: We aim at calculating the probability $\pi(\mu)$ of survival of
278: RW$(\mu)$-er moving with rate $r(x\to x')=r_{\bf V} (x\to x')$
279: and dying on site $x$
280: with rate $d(x)=0$ except: $r(y\to y+1)
281: =1/r(y+1\to y)=e^{(1-2\mu)f_y/(2T)}$,
282: $d(y)= D(f_y), d(y+1)=D(-f_y)$ where
283: $D(f)\equiv (1-\mu) e^{f/(2T)}+\mu e^{-f/(2T)} - e^{
284: (1-2\mu) f/(2T)}$ from (\ref{death}). From the
285: previous section the number of RWs required for a reliable prediction
286: of $f_y$ is $R_c(y;{\bf V}) = \min _{\mu} 1/|\ln \pi(\mu)|$.
287:
288: Let $\pi _x(\mu,t)$ be the
289: probability that RW$(\mu)$, initially on site $x$, is still
290: alive at time $t$. The time-evolution of $\pi_x$ is described by
291: \begin{equation} \label{diffeq}
292: \frac{\partial\pi _x }{\partial t} = \sum
293: _{x' (\ne x)} r(x\to x') \big( \pi _{x'}- \pi_{x} \big)
294: -d (x)\, \pi_x \ ,
295: \end{equation}
296: with initial condition $\pi_y (\mu,0)=1$ (by convention
297: $\pi_{-1}=\pi_{N+1}=0$). After Laplace transform over time, eqns
298: (\ref{diffeq}) are turned into recurrence equations for the ratios
299: $\pi_x/\pi_{x+1}$ and solved with great numerical accuracy. We obtain this way
300: the probability of survival, $\pi(\mu)=\pi_0(\mu,t_f)$, and optimize
301: over $\mu$.
302: Though $R_c$ depends on the potential ${\bf V}$ its general behavior
303: for tilt $b>0$ as a function of the duration $t_f$ is sketched in
304: Fig.~\ref{fig-rc}. Three regimes are observed:
305:
306: $\bullet$ for $t_f\ll \tau_y$ (mean first passage time in $y$)
307: RW$(\mu)$ has a low probability to visit $y$ and is almost surely alive,
308: hence $R_c$ is very large;
309:
310: $\bullet$ for $\tau _y\ll t_f\ll t_{eq}$ RW$(\mu)$ has visited
311: the region
312: surrounding $y$ and escaped from this region (transient regime), hence
313: its probability of survival remains constant, and so does
314: $R_c$;
315:
316: $\bullet$ for $t_f\gg t_{eq}$ RW$(\mu)$ visits again
317: and again the region surrounding $y$, hence
318: the probability of survival decreases exponentially
319: with the duration: $R_c \propto 1/t_f$.
320:
321: The total time $R_c\times t_f$ for a good reconstruction
322: is minimal when we choose $t_f \gtrsim \tau_y$. This
323: marginally transient regime corresponds to the plateau of
324: Fig.~\ref{fig-rc}: RWs are long enough to
325: visit site $y$ but short enough not to wander much away from $y$.
326: To calculate the corresponding value of $R_c$ we take
327: the limits, in order, $N\to\infty$, $t_f\to \infty$, and look for the
328: stationary solution of (\ref{diffeq}) with boundary condition
329: $\pi_{x\to\infty}=1$. The result for
330: the probability of survival is
331: \begin{equation} \label{pstarsinai}
332: \pi (\mu) = \frac{e^{-\frac{\mu}T}}{1-\mu+ \mu
333: \, e^{-\frac 1T}+\mu(1-\mu)\, t^* _{y+1}\,(e^{\frac{1}{4T}}-
334: e^{-\frac{3}{4T}})^2 } \ ,
335: \end{equation}
336: where the mean sojourn time on site $y+1$ in ${\bf V}$ is \cite{revue}
337: \begin{equation} \label{tx}
338: t ^*_{y+1}= \sum _{z\ge0} \exp \left[\frac 1{T} \left( \frac {
339: V_{y+z+2}+ V_{y+z+1}}2 - V_{y+1} \right)\right] \ .
340: \end{equation}
341:
342:
343: \begin{figure}
344: \begin{center}
345: \vskip .7cm
346: \psfig{figure=./fig2.eps,height=3cm,angle=0}
347: \caption{Sketch of the number $R_c(y;{\bf V})$ of RWs
348: necessary for a good inference of the force $f_y$
349: as a function of the RW duration $t_f$. $\tau_y$ is the typical
350: first-passage time in $y$ from the origin, $t_{eq}$ the equilibration
351: time (comparable to the first-passage time from the extremity
352: $N$ when $y\ll N$). Inset: rate of reconstruction (\ref{vel}) as
353: a function of temperature at fixed tilt.}
354: \label{fig-rc}
355: \end{center}
356: \end{figure}
357:
358:
359: {\em Distribution of $R_c$ over potentials.}
360: The number $R_c(y;{\bf V})$ of RWs necessary to predict the value of
361: $f_y$ depends on the potential ${\bf V}$ through the sojourn time
362: $t^*_{y+1}$ (\ref{tx}). By randomly drawing potentials (or varying site $y$)
363: we obtain the distribution of $R_c$ shown in Fig.~\ref{fig-histo}.
364: Main features are:
365:
366: $\bullet$ Small $R_c$ correspond to sites where the RW spends long
367: time $t^*$ (traps)\footnote{RW$(\mu)$, due to conditioning to survival,
368: is likely to stay for $\sim 1/d(y) \ll t^*$ in the trap only.}:
369: $R_c \sim \frac 1{|\ln \pi|} \sim \frac 1{\ln t^*}$ from
370: (\ref{pstarsinai}). The power law tail of the distribution of
371: sojourn times, $P(t^* )\sim (t^*) ^{-(\alpha+1)}$ \cite{revue},
372: gives rise to an essential singularity at the origin
373: in the cumulative distribution, ${\cal Q} (R_c) \sim
374: e^{-\alpha/R_c}$. The potential is easy to predict over trapping
375: regions since RWer spends a long time there, and accumulates
376: information about the energy landscape.
377:
378: $\bullet$ Conversely the largest value of $R_c$, denoted by $R_c ^H$,
379: correspond to the homogeneous potential $V_x^H=-x$ in which the
380: walker is never trapped and is quickly driven to $+\infty$.
381: $R_c^H$ can be calculated from (\ref{pstarsinai}) by setting
382: $f_x=+1$ for all sites in (\ref{tx}). The singularity in
383: ${\cal Q}$ when $R_c\to R_c^H$ corresponds to quasi-homogeneous
384: potentials, where one force, say, on site $\ell$, is $-1$.
385: Such potentials have exponential-in-$\ell$ small probabilities, but
386: give values of $R_c$ on site $y=0$ exponentially close to $R_c^H$.
387: On the overall we
388: find $1-{\cal Q} (R_c ^H - \epsilon) \sim \epsilon ^\beta$ where
389: the exponent is $\beta=T \ln \frac {1+b}2$.
390:
391: $\bullet$ In between ${\cal Q}$ shows marked steps at well defined and
392: $b$-independent values of $R_c$, which correspond to specific
393: local force patterns beyond site $y$.
394: A $\ell$-pattern is defined as a sequence of
395: forces on sites $y+1$ to $y+\ell+1$, followed by all $+$
396: forces; the corresponding $R_c$ can be
397: exactly calculated from (\ref{pstarsinai},\ref{tx}), and is
398: shown for 7 among the
399: 16 $\ell=4$-patterns in Fig.~\ref{fig-histo}.
400: The histogram of $R_c$ can be accurately approximated for any tilt $b>0$ based
401: on the above local pattern description. Given a length $\ell$ we enumerate all
402: the $2^\ell$ patterns, calculate the corresponding $R_c$, and weight them
403: with probability $(\frac {1+b}2) ^{\# f_x =+}\times (\frac {1-b}2)
404: ^{\# f_x =-}$. In practice we choose $\ell\sim 10/\ln [2/(1-b)]$,
405: to ensure that patterns with more than $\ell$ negative
406: forces have negligible
407: weights ($< e^{-10}$). The resulting histograms are in
408: excellent agreement with ${\cal Q}$ for intermediate
409: values of $R_c$ (dashed lines in Fig.~\ref{fig-histo}).
410:
411: \begin{figure}
412: \begin{center}
413: \vskip .7cm
414: \psfig{figure=./fig3.eps,height=5cm,angle=0}
415: \caption{Cumulative probability distribution ${\cal Q}$ of
416: $R_c(y;{\bf V})$ at
417: temperature $T=1$ and for three tilt values $b$. Full lines are
418: numerical results from $10^6$ samples, and dashed lines are the
419: outcomes from the $\ell$-pattern approximation. Inset: $R_c$ vs.
420: $T$ for the 3-patterns $+++$,
421: $-++$, $---$ (from top to down).}
422: \label{fig-histo}
423: \end{center}
424: \end{figure}
425:
426: {\em Tuning temperature for fast reconstruction.}
427: The dependence of $R_c$ upon temperature is shown for three patterns
428: in the Inset of Fig.~\ref{fig-histo}. We have $R_c \sim 4T$
429: as $T\to\infty$ independently of the pattern, and $R_c\sim 2T/(h+3)$
430: when $T\to 0$ where $h$ is the highest barrier to the right of $y$
431: in the potential defined by the pattern (Fig.~\ref{fig-histo}).
432: When the temperature exceeds the temperature $T_b$ such that
433: $\alpha=1$ the velocity
434: of the RWer is finite $\frac y{\tau _y} \sim v(T) >0$
435: \cite{revue}. The reconstruction rate (number of
436: correctly predicted forces per unit of time) is equal to the velocity
437: $v(T)$ divided by $R_c$,
438: \begin{equation} \label{vel}
439: \nu (T)= \frac{1-\cosh\frac 1T + b \sinh
440: \frac 1T}{\cosh \frac 1{2T} - b \sinh \frac 1{2T}} \times \int _0 ^{R_c^H}
441: dR_c \frac {{\cal Q} '(R_c)}{R_c}
442: \end{equation}
443: after averaging over the quenched potential.
444: The dependence of $\nu$
445: upon temperature is sketched in the Inset of Fig.~\ref{fig-rc}; it is
446: maximal and equal to $\nu ^M$ for some temperature $T^M$ realizing a
447: trade-off between fast motion (large velocity) and accurate
448: reading-out (small $R_c$). Even in the small
449: $b$ limit the optimal reconstruction rate is finite, $\nu^M \sim b^2$,
450: by working at high temperature $T^M \sim \frac 1b$, while in the
451: absence of optimization procedure the number of predicted forces
452: scales only as the squared logarithm of the time \cite{math2}.
453:
454: {\em Conclusion.} We have shown how the number of RWs
455: required for a good reconstruction of the potential
456: can be deduced from the probability of
457: survival of an absorbing RW process. This result is of practical
458: interest since the survival probability can be estimated through
459: numerical simulations e.g. in dimension $\ge 2$. Furthermore we have
460: determined, for the special case of the RF model, the optimal
461: `experimental' protocol for reconstruction (number of RWs, duration,
462: temperature).
463:
464: Our formalism applies to continuously parametrized potentials
465: e.g. RF model with forces taking continuous
466: instead of binary values. The aim is now to predict the true potential
467: values up to some accuracy on each site; this is turn
468: determines an acceptable neighborhood around ${\bf s}^*_{\bf v}$ in the space
469: of signals. The rate function $\omega _{\bf v}$ is generically
470: parabolic around ${\bf s}^*_{\bf v}$, with a curvature matrix called Fisher
471: information matrix \cite{bayes}. Finding $R_c$ amounts to minimize this
472: (positive) quadratic form on the boundary of the neighborhood, a task
473: which can be carried out efficiently \cite{garey}.
474: Our approach can be easily extended to the
475: case of a finite delay between two measures of the positions, and
476: Chernoff's result is recovered in the finite $N$, infinite delay
477: limits \cite{Che52,mb}.
478:
479: {\em Acknowledgments.} We are grateful to D. Thirumalai for his
480: suggestion of illustrating our formalism on the RF model. This work
481: was partially funded by ANR under contract 06-JCJC-051.
482:
483:
484:
485: \begin{thebibliography}{999999}
486:
487: \bibitem{discreteRF}
488: B. D. Hughes, {\em Random walks and random environments},
489: Oxford University Press (1996).
490:
491: \bibitem{revue}
492: J-P. Bouchaud, A. Georges, {\em Physics Reports} {\bf 195}, 127 (1990).
493:
494:
495: \bibitem{evans}
496: R. Merkel {\em et al.}
497: {\em Nature} {\bf 397},50-53 (1999).
498:
499:
500: \bibitem{fer04}
501: J.M. Fernandez, H. Li {\em Science} {\bf 303}, 1674 (2004).
502:
503: \bibitem{Ess97}
504: B. Essevaz-Roulet, U. Bockelmann, F. Heslot,
505: {\em Proc. Natl. Acad. Sci. (USA) } {\bf 94}, 11935 (1997).
506:
507: \bibitem{woo06}
508: M.T. Woodside {\em et al.} {\em Science} {\bf 314}, 1001 (2006).
509:
510:
511: \bibitem{hye03}
512: C. Hyeon, D. Thirumalai
513: {\em Proc Natl Acad Sci USA} {\bf 100},10249-53 (2003).
514:
515: \bibitem{mb}
516: V. Baldazzi {\em et al.} {\em Phys. Rev. Lett.} {\bf 96} 128102
517: (2006); {\em Phys. Rev. E} {\bf 75}, 011904 (2007)
518:
519: \bibitem{ritort}
520: M. Manosas, D. Collin, F. Ritort {\em Phys. Rev. Lett.} {\bf 96},
521: 218301 (2006).
522:
523: \bibitem{bayes}
524: T.M. Cover, J.A. Thomas, {\em Elements of Information Theory}, Wiley (1991)
525:
526:
527: \bibitem{domany}
528: L. Ein-Dor, O. Zuk, E. Domany, {\em Proc. Nat. Acad. Sci. (USA)} {\bf
529: 103}, 5923-5928 (2006).
530:
531: \bibitem{And57}
532: T. W. Anderson, L. Goodman
533: {\em The Annals of Mathematical Statistics} {\bf 28}, 89 (1957)
534:
535: \bibitem{Che52}
536: H. Chernoff, {\em Ann. Math. Statis.} {\bf 23}, 493 (1952)
537:
538: \bibitem{Dem98}
539: A. Dembo, O. Zeitouni, {\em Large deviations Techniques and Applications}
540: Springer-Verlag (1998)
541:
542: \bibitem{Boz71}
543: L.B. Boza, {\em Ann. Math. Statis.} {\bf 42}, 1992 (1971)
544:
545: \bibitem{noi}
546: S. Cocco, R. Monasson, {\em in preparation}
547:
548: \bibitem{garey}
549: M.R. Garey, D.S. Johnson, {\em Computers and Intractability:
550: A Guide to the Theory of NP-Completeness},
551: W.H. Freeman (1979).
552:
553: \bibitem{math1}
554: O. Adelman, N. Enriquez, {\em Israel J. Math.} {\bf 142}, 205-220
555: (2004).
556:
557: \bibitem{math2}
558: P. Andreoletti, {\em preprint arxiv:math.PR/0612208} (2006).
559:
560: \end{thebibliography}
561:
562:
563: \end{document}
564: