1: %%%%%%%   19 gennaio 2001 - ROMA

2: \documentstyle[pre,aps,epsfig,multicol]{revtex}

3: %\documentstyle[preprint,aps,epsfig]{revtex}

4: %\tightenlines

5: 

6: \begin{document}

7: \title{Linear and nonlinear 

8: information flow \\ in spatially extended systems}

9: \author{Massimo Cencini $^{(a)}$ and Alessandro Torcini$^{(b),(c)}$}

10: \address{$(a)$ Max-Planck-Institut f\"ur Physik komplexer Systeme,

11: N\"othnitzer  Str. 38, D-01187 Dresden, Germany  } 

12: \address{$(b)$ Physics Department, Universit\'a ``La Sapienza'',

13: piazzale Aldo Moro 2, I-00185 Roma, Italy} 

14: \address{$(c)$ Istituto Nazionale di Fisica della Materia, UdR Firenze,

15: L.go E. Fermi, 3 - I-50125 Firenze, Italy}

16: 

17: \draft

18: 

19: \maketitle

20: 

21: \begin{abstract}

22: Infinitesimal and finite amplitude error propagation in 

23: spatially extended 

24: systems are numerically and theoretically investigated.

25: The information transport in these systems can be characterized

26: in terms of the propagation velocity of perturbations $V_p$.

27: A linear stability analysis is sufficient to capture all

28: the relevant aspects associated to propagation of 

29: infinitesimal disturbances. In particular, this 

30: analysis gives the propagation velocity $V_L$ of 

31: infinitesimal errors. If linear mechanisms prevail

32: on the nonlinear ones $V_p = V_L$. On the contrary, if 

33: nonlinear effects are predominant finite amplitude 

34: disturbances can eventually propagate faster than infinitesimal 

35: ones (i.e. $V_p > V_L$). The finite size Lyapunov exponent

36: can be successfully employed to discriminate the linear

37: or nonlinear origin of information flow.

38: A generalization of finite size Lyapunov exponent

39: to a comoving reference frame allows to state a

40: marginal stability criterion able to provide $V_p$

41: both in the linear and in the nonlinear case.

42: Strong analogies are found

43: between information spreading and propagation 

44: of fronts connecting steady states 

45: in reaction-diffusion systems. The analysis of

46: the common characteristics of these two phenomena 

47: leads to a better understanding of the role played

48: by linear and nonlinear mechanisms for the flow of 

49: information in spatially extended systems.

50: \end{abstract}

51: %

52: \pacs{PACS numbers: 05.45.-a,68.10.Gw,05.45.Ra,05.45.Pq}

53: %

54: % explanation of PACS numbers:

55: % ----------------------------

56: %

57: % 47.20.Ky Nonlinearity (including bifurcation theory)

58: % 05.45.Ra Coupled map lattices

59: % 05.45.-a Nonlinear dynamics and nonlinear dynamical

60: % systems (see also 45 Classical mechanics of

61: % discrete systems)

62: % 05.45.Pq Numerical simulations of chaotic models

63: % 68.10.Gw Interface activity, spreading

64: % 47.54.+r Pattern selection; pattern formation

65: % 02.30.Jr Partial differential equations

66: % 05.40.-a Fluctuation phenomena, random processes, noise, and Brownian motion

67: 

68: 

69: \begin{multicols}{2}

70: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

71: \section{Introduction}

72: \label{sec:1}

73: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

74: It is well recognized that chaotic dynamics generates a flow of

75: information in {\it bit space}: due to the sensitive dependence on initial

76: conditions one has an information flow from ``insignificant'' digits

77: towards ``significant'' ones \cite{Shaw}.  

78: In spatially distributed systems, due to the spatial coupling, one has an 

79: information flow both in {\it bit space} and in {\it real space}.

80: The flow in {\it bit space} is typically characterized in 

81: terms of the maximal growth $\lambda$ rate of infinitesimal disturbances

82: (i.e. of the maximal Lyapunov exponent), while the spatial information flow 

83: can be measured in terms of the maximal velocity of disturbance propagation

84: $V_p$ \cite{K86,BR87,grass89,PV94}. 

85: 

86: The evolution of a typical infinitesimal disturbance in low dimensional systems 

87: is fully determined once the maximal Lyapunov exponent is known. 

88: The situation is more complicated in spatio-temporal chaotic systems,

89: where infinitesimal perturbations can evolve both in time 

90: and in space. In this case a complete description of the dynamics

91: in the tangent space requires the introduction of other indicators,

92: e.g.: the comoving Lyapunov exponents \cite{DK87} and

93: the spatial and the specific Lyapunov spectra \cite{lepri}.

94: 

95: Nevertheless, the complete knowledge of  these Lyapunov spectra is not 

96: sufficient to fully characterize 

97: the irregular behaviors emerging in dynamical systems, this is

98: particularly true when the evolution of finite perturbations

99: is concerned.  Indeed, finite disturbances, which  are not confined 

100: in the tangent space, but are governed by the complete nonlinear 

101: dynamics, play a fundamental

102: role in the erratic behaviors observed in some high dimensional

103: system \cite{CK88,PLOK93,TP94,TGP95,CFVV99}. A rather 

104: intriguing phenomenon, termed {\it stable chaos}, 

105: has been reported in \cite{PLOK93}: the authors 

106: observed that even a linearly stable system 

107: (i.e. with $\lambda < 0$) can display an erratic behavior with $V_p>0$.

108: 

109: The first attempts to 

110: describe nonlinear perturbation evolution have been reported in

111: \cite{dressler92}. However, in these studies the analysis

112: was limited to the temporal growth rate associated with 

113: second order derivatives of one dimensional maps.

114: A considerable improvement 

115: along this direction has been recently achieved with the introduction of

116: the

117: finite size Lyapunov exponent (FSLE) \cite{ABCPV96}:

118: a generalization of the maximal Lyapunov exponent able to 

119: describe also finite amplitude perturbation evolution. 

120: In particular, the FSLE has been already demonstrated 

121: useful in investigating high dimensional systems \cite{CFVV99}.

122: 

123: The aim of this paper is to fully characterize the infinitesimal 

124: and finite amplitude perturbation evolution in spatio-temporal

125: chaotic systems. Coupled map lattices (CML's) \cite{cml} 

126: are employed to mimic spatially extended chaotic systems.

127: The FSLE is successfully applied to discriminate 

128: the linear or nonlinear origin of information propagation 

129: in CML's. Moreover, a generalization

130: of the FSLE to comoving reference frame (finite size comoving Lyapunov exponent)

131: allows to state

132: a marginal stability criterion able to predict $V_p$

133: in both cases: linear or nonlinear propagation.

134: A parallel with front propagation in reaction-diffusion \cite{KPP,wim2}

135: (non-chaotic) systems is worked out. The 

136: analogies between the two phenomena authorize to draw a 

137: correspondence between ``pulled'' (``pushed'') fronts 

138: and linear (nonlinear) information spreading.

139: 

140: The paper is organized as follows.

141: In Sect. II the FSLE is introduced

142: and applied to low dimensional systems (i.e. to single chaotic maps).

143: Sect. III is devoted to the description and 

144: comparison of linear and nonlinear 

145: disturbance propagation observed in different CML models.

146: The finite size comoving Lyapunov exponent is introduced in Sect. IV

147: and employed to introduce a generalized marginal stability criterion

148: for the determination of $V_p$. A discussion on information

149: propagation in non chaotic systems conclude Sect. IV.

150: The analogies between disturbance propagation in chaotic

151: systems and front propagation connecting steady states 

152: are analyzed in Sect. V.  The Appendix is devoted to the

153: estimation of finite time corrections for the computation of the 

154: FSLE in extended systems. Finally, some conclusive

155: remarks are reported in Sect. VI.

156: 

157: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

158: \section{Finite Size Lyapunov Exponent: \\

159: Low Dimensional Models}

160: \label{sec:2}

161: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

162: %introduzione dell'fsle

163: 

164: Let us introduce the FSLE by considering the

165: dynamical evolution of the state variable 

166: ${\bf x}={\bf x}(t)$ ruled by

167: $$

168: {\dot {\bf x}}(t)={\bf f}[{\bf x}(t)] \, ,

169: $$

170: where ${\bf f}$ represents a chaotic flow in the phase space.

171: In order to evaluate the growth rate of non infinitesimal perturbations 

172: one can proceed as follows: a reference, ${\bf x}(t)$, and 

173: a perturbed, ${\bf x}^{'}(t)$, trajectories are considered.

174: The two orbits are initially placed at a distance 

175: $\delta(0)=\delta_{min}$, with $\delta_{min}\ll 1$,

176: assuming a certain norm  $\delta(t)=||{\bf x}^{'}(t)- {\bf x}(t)||$. 

177: In order to ensure that the perturbed orbit relaxes

178: on the attractor a first scratch integration is performed

179: for both the orbits until their distance has grown from $\delta_{min}$ 

180: to $\delta_0$ (where $1 >> \delta_0 >> \delta_{min})$. 

181: This transient ensures also the 

182: alignment of the perturbation along the direction of maximal expansion. 

183: Then the two trajectories are let to evolve and 

184: the growth of their distance $\delta(t)$ through different pre-assigned 

185: thresholds ($\delta_n=\delta_0 r^{n}$, with $n=0,\dots,N$ and typically

186: $1 < r \le 2$) is analyzed.

187: 

188: After the first threshold, $\delta_0$, is attained  

189: the times $\tau(\delta_n,r)$ required for 

190: $\delta(t)$ to grow from $\delta_n$ up to $\delta_{n+1}$

191: are registered.

192: When the largest threshold $\delta_N$ (which should be obviously chosen

193: smaller than the attractor size) is reached, the perturbed

194: trajectory is rescaled to the initial distance 

195: $\delta_{min}$ from the reference one.

196: 

197: By repeating the above procedure ${\cal N}$ times, 

198: for each threshold $\delta_n$, one obtains

199: a set of ``doubling'' times (this terminology is strictly

200: speaking correct only if $r=2$) 

201: $\{\tau_i(\delta_n,r)\}_{i=1,\dots,{\cal N}}$

202: and one can define the average of any observable

203: $A = A(t)$ on this set of doubling times as:

204: $$

205:  \langle A\rangle_e \equiv \frac{1}{\cal N} \sum_{i=1}^{\cal N} A_i\,,

206: $$

207: where $ A_i = A(\tau_i(\delta_n,r))$. 

208: The average $\langle \cdot \rangle_e$

209: does not coincide with an usual time average $\langle \cdot \rangle_t$

210: along a considered trajectory in the phase space, since

211: the doubling times typically depend on the 

212: considered point along the trajectory and on the threshold

213: $\delta_n$.  The two averages are linked 

214: (at least in the continuous case) 

215: via the following straightforward relationship~\cite{cencrep}

216: \begin{equation}

217: \langle A(t) \rangle_t = \frac{1}{T} \int_0^T dt A(t) = 

218: \frac{ \langle A \tau \rangle_e}{\langle \tau \rangle_e}

219: \label{eq:2.1}

220: \end{equation}

221: where  $T = \sum_{i=1,{\cal N}} \tau_i(\delta_n,r)$ and

222: $\langle \tau(\delta_n,r)\rangle_{e} = {T/\cal{N}}$.

223: 

224: A natural definition of the Finite Size Lyapunov Exponent

225: $\lambda(\delta_n)$ is the following~\cite{ABCPV96}:

226: \begin{equation}

227: \lambda(\delta_n) \equiv \left\langle{1 \over \tau(\delta_n,r)}\right\rangle_{t} \ln r \equiv

228: {1 \over \langle \tau(\delta_n,r) \rangle_{e}} \ln r \, .

229: \label{eq:2.2}

230: \end{equation}

231: The last equality stems from the relationship among the two

232: averages reported in (\ref{eq:2.1}).

233: 

234: In the limit of infinitesimal perturbation $\delta_n$

235: and of infinite $T$ (or $\cal N$) the FSLE converges to the

236: usual maximal Lyapunov exponent

237: \begin{equation}

238: \lim_{{\cal N} \to \infty}

239: \lim_{\delta_n \to 0} \lambda(\delta_n) = \lambda \, .

240: \label{eq:2.3}

241: \end{equation}

242: In practice, at small enough $\delta_n$,

243: $\lambda(\delta_n)$ displays a plateau

244: $\sim \lambda$.

245: Moreover, one can verify that $\lambda(\delta_n)$ is

246: independent of $r$, at least for not too large 

247: $r$~\cite{ABCPV96}.

248: 

249: In Eq.~(\ref{eq:2.2}) continuous time has been assumed, 

250: but discrete time is the most natural choice when 

251: experimental data sets (typically sampled at fixed intervals) 

252: are considered. In order to generalize the FSLE's definition

253: to the case of discrete time dynamical systems, let us consider 

254: the following map

255: $$ {\bf x}(t+1)={\bf F}({\bf x}(t)) \,,$$ 

256: where ${\bf x}$ is a continuous 

257: variable, and $t$ assumes integer values.

258: In this case $\tau(\delta_n,r)=\tau$ has simply to be interpreted as the minimum 

259: ``integer'' time such that $\delta(\tau) \ge  \delta_{n+1}$, and, 

260: since now $\delta(\tau)/\delta_n$ is a fluctuating quantity, 

261: the following definition is obtained

262: \begin{equation}

263: \lambda(\delta_n) \equiv {1 \over \langle \tau(\delta_n,r) \rangle_{e}}

264: \left\langle \ln \left( \delta(\tau) \over \delta_n

265:  \right) 

266: \right\rangle_{e} \, .

267: \label{eq:2.5}

268: \end{equation}

269: A theoretical estimation of (\ref{eq:2.5}) is rarely possible,

270: and in most cases, one can only rely on a numerical

271: computation of $\lambda(\delta_n)$. However, 

272: in the following we will report two simple cases 

273: for which an approximate analytic expression for

274: the FSLE can be worked out.

275: 

276: Let us first consider the tent map 

277: $$ F(x)= 1- 2\left|x-\frac{1}{2}\right|$$

278: where $x \in [0:1]$. This is a one dimensional chaotic map,

279: since  $\lambda= \ln 2$ is positive.

280: 

281: Due to the simplicity of this map, one can estimate the expression

282: (\ref{eq:2.5}) analytically obtaining the following

283: approximation 

284: \begin{equation}

285: \lambda(\delta) \simeq \ln 2 -\delta\,,

286: \label{eq:fsle_tent}

287: \end{equation}

288: valid for not too large $\delta$ values.

289: The maximal Lyapunov exponent is correctly recovered

290: in the limit $\delta \to 0$ and the above expression 

291: reproduces quite well the numerical estimate of the FSLE

292: (see Fig.~\ref{fig:maps} (a) ). An important point to stress

293: is that for this map the finite amplitude perturbations grow 

294: with the same rate or slower than the infinitesimal ones. 

295: The contraction of perturbations at large scales is

296: due to saturation effects related to the attractor

297: size. A similar dependence of the FSLE on the considered

298: scale is observed for the majority of the chaotic maps 

299: (logistic, cubic, etc.), as we have verified. 

300: %fig1

301: %%%%%%%%%%%%%%%%%%%%%%%%

302: \begin{figure}

303: \epsfxsize=8truecm

304: \epsfysize=6truecm

305: \epsfbox{Fig1a.eps}

306: \epsfxsize=8truecm

307: \epsfysize=6truecm

308: \epsfbox{Fig1b.eps}

309: %\narrowtext

310: \caption{$\lambda(\delta)$ versus $\delta$ for the tent map (a) and the 

311: shift map with $\beta=2$ (b). The continuous lines are the analytically 

312: computed FSLE and the boxes the numerically evaluated one. 

313: The two maps are displayed in the insets.  

314: }

315: \label{fig:maps}

316: \end{figure}

317: %%%%%%%%%%%%%%%%%%%%%%%%

318: One can wonder if there are systems for which, at variance with 

319: the behavior (\ref{eq:fsle_tent}), the finite size

320: corrections leads to an enhancement of the growth rate at 

321: large scales. As shown in \cite{TGP95}, the shift map

322: $F(x)= \beta x \; {\mbox {mod}} \; 1$ represents a good candidate. 

323: Also in this case it is possible to obtain

324: an analytical expression for the FSLE, when

325: $\delta < [1/(r+\beta)]$,

326: \begin{equation}

327: \lambda(\delta)=\frac{1}{1-\delta} \left[

328: (1-2\delta) \ln \beta+

329: \delta \ln \left({1-\beta\delta \over \delta}\right)

330: \right]\,,

331: \label{eq:lambda_shift}  

332: \end{equation}

333: which again correctly reproduces the numerical data

334: (see Fig.~\ref{fig:maps} (b) ) and in the limit

335: $\delta \to 0$ reduces to the corresponding maximal

336: Lyapunov exponent $\lambda=\ln \beta$.

337: As expected, finite amplitude disturbances can grow

338: faster than infinitesimal ones : 

339: \begin{equation}

340: \lambda(\delta) > \lambda(\delta \to 0)=\lambda \qquad {\mbox {for}} 

341: \:\: 0 < \delta \le \delta^{sat}\,,

342: \label{eq:nonlin}

343: \end{equation}

344: where $\delta^{sat}$ indicates the threshold at which saturation

345: effects set in. An even more interesting situation

346: is represented by the circle map 

347: $F(x)= \alpha + x \:\; {\mbox {mod}} \; 1$. This

348: map is marginally stable (i.e. $\lambda\!=\!\!0$),

349: but it is unstable at finite scales.

350: Indeed, the FSLE is given by 

351: $\lambda(\delta)\!\!=\!\!\delta/(1-\delta) \ln[(1-\delta)/\delta]$,

352: which is positive for $0 < \delta <1/2$. Therefore at small, but

353: finite, perturbations a positive growth rate is observed in spite 

354: of the (marginal) stability against infinitesimal perturbations. 

355: As a consequence, the circle map can exhibit behaviors that can

356: be hardly distinguished from chaos under the influence of

357: noise, since small perturbations may be occasionally driven 

358: into the nonlinear (unstable) regime and therefore amplified.

359: Of course, the role of noise can be played by coupling with

360: other maps, e.g. it has been found that coupled circle maps display 

361: behavior

362: resembling (for some aspects) that of a chaotic system \cite{TGP95}.

363: This phenomenon becomes even more striking in certain coupled stable maps where, 

364: even if the maximal Lyapunov exponent  is

365: negative \cite{PLOK93}, one can have a strong sensitivity to non infinitesimal

366: perturbations \cite{Kantz} (see  Sect.~\ref{sec:4.2} for a detailed discussion).

367: 

368: The two maps here examined for which (\ref{eq:nonlin}) holds

369: have a common characteristic: they are discontinuous.

370: However, in order to observe similar 

371: strong nonlinear effects, it is sufficient to consider a continuous map 

372: with high, but finite, first-derivative $|F^\prime|$ values 

373: \cite{TP94}. In this respect a simple example, that 

374: will be examined more in detail in Sect. IV, is represented 

375: by the map:

376: \begin{equation}

377: F(x)=\left \{ \begin{array}{cc}

378: b x & \qquad 0 \leq x < 1/b\\

379: 1-c(1-q)(x-1/b) & \qquad 1/b \leq x < \frac{b+c}{bc}\\

380: q+d(x-\frac{b+c}{bc}) & \qquad \frac{b+c}{bc} \leq x \leq 1\,;

381: \end{array}

382: \right.

383: \label{eq:stable}

384: \end{equation}                                                 

385: with $b=2.7$, $d=0.1$, $q=0.07$ and $c=500$.

386: For $c \to \infty$ the map (\ref{eq:stable}) reduces to the one

387: studied in \cite{PLOK93}. For the map (\ref{eq:stable}) the FSLE 

388: dependence on $\delta$ is similar to that observed 

389: for the shift map.

390: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

391: \section{Information Spreading in Spatially Distributed Systems}

392: \label{sec:3}

393: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

394: In this section we will examine the mechanisms behind the

395: information flow in spatially distributed systems. 

396: In particular, the influence of

397: linear and nonlinear effects on information (error) spreading

398: will be analyzed. As a prototype of spatially 

399: distributed system Coupled Map Lattices (CML's) \cite{cml}

400: are considered :

401: \begin{eqnarray}

402: x_i(t+1)&=&F(\tilde{x}_i(t)) \nonumber \\

403: \tilde{x}_i(t)=(1-\varepsilon) x_i(t) &+& {\varepsilon \over 2} (x_{i-1}(t)+x_{i+1}(t))

404: \label{eq:cml}

405: \end{eqnarray}

406: where $t$ and $i$ are the discrete temporal and spatial indices, 

407: $L$ is the lattice size ($i=-L/2,\dots,L/2$), $x_i(t)$ the

408: state variable, and $\varepsilon \in [0:1]$ measures the

409: strength of the diffusive coupling. 

410: $F(x)$ is a nonlinear map of the interval 

411: ruling the local dynamics.

412: 

413: In order to understand how the information spreads

414: along the chain, let us consider two replicas of the same 

415: system, ${\bf x}(t)=\{x_i(t)\}$  and

416: ${\bf x}^\prime(t)=\{x_i^\prime(t)\}$, 

417: that initially differ only in

418: a single site of the lattice (e.g. $i=0$) of a quantity $d_0$, i.e.

419: \begin{equation}

420: |x^\prime_i(0)-x_i(0)|=\Delta x_i(0)=d_0 \delta_{i,0}\,,

421: \label{eq:tzero}

422: \end{equation}

423: where $\delta_{i,0}$ is the Kronecker's delta. 

424: In a chaotic system the perturbation will typically

425: grow locally and spread along the chain.

426: These phenomena can be studied by considering the

427: difference field

428: \begin{equation}

429: \Delta x_i(t)= |x^\prime_i(t)-x_i(t)|=

430: |F(\tilde{x}^{'}_i(t-1))-F(\tilde{x}_i(t-1))|\,. 

431: \label{eq:dyn}

432: \end{equation}

433: It has to be stressed that the full nonlinear dynamics 

434: contributes to the evolution of $\Delta x_i(t)$.

435: 

436: The spreading of this initially localized 

437: disturbance can be characterized in terms of

438: the velocity of information

439: propagation $V_P$ \cite{K86,grass89}.

440: As shown in Fig.~\ref{fig:torc.1}, $\Delta x_i(t)$ 

441: can grow only within a light-cone, determined by  $V_p$. 

442: For velocities higher than

443: $V_p$ the disturbance is instead damped. This individuates

444: a sort of predictability ``horizon'' in space-time: 

445: i.e. an interface separating 

446: the perturbed from the unperturbed region.

447: % fig 2

448: %%%%%%%%%%%%%%%%%%%%%%%%

449: \begin{figure}

450: \epsfxsize=8.truecm

451: \epsfysize=6truecm

452: \epsfbox{Fig2.eps}

453: %\narrowtext

454: \caption{Evolution of $\Delta x_i(t)$, 

455: for a chain of coupled tent map lattices

456: with a coupling $\varepsilon=2/3$.

457: The initial perturbation

458: is taken as in (\ref{eq:tzero}) with $d_0=10^{-8}$. 

459: }

460: \label{fig:torc.1}

461: \end{figure}

462: %%%%%%%%%%%%%%%%%%%%%%%% 

463: The velocity $V_p$ can be directly measured 

464: by detecting the leftmost, $i_l(t)$, and the rightmost,

465: $i_r(t)$, sites for which at time $t$ the perturbation 

466: $\Delta x_i(t)$ exceeds a preassigned

467: threshold. The definition of $V_p$ is the following

468: \begin{equation}

469: V_p=\lim_{t\to \infty}\lim_{L\to \infty} {i_r(t) -i_l(t)\over 2t}\,,

470: \label{eq:vel1}

471: \end{equation}

472: where  the limit $L\to \infty$ has to be taken first to avoid boundaries 

473: effects. The velocity (\ref{eq:vel1}) does not depend

474: on the chosen threshold values \cite{K86,grass89,TP94}.

475: 

476: Since the dynamics of the difference field 

477: (\ref{eq:dyn}) is not confined in the tangent space, 

478: non linearities can play a crucial role in the information propagation.

479: Indeed, we will see that the evolution of the disturbances strongly depend 

480: on the considered map $F(x)$ and in particular on the

481: shape of $\lambda(\delta)$. In the next subsection 

482: propagation in CML's with local chaotic

483: maps for which $\lambda(\delta)\leq \lambda \quad \forall \delta$

484: is discussed.  Local maps for which the condition 

485: (\ref{eq:nonlin}) holds will be the subject of 

486: Sect. \ref{sec:3.2}.

487: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

488: \subsection{Linear Mechanisms}

489: \label{sec:3.1}

490: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

491: Since in this subsection we consider 

492: CML's for which the local instabilities are 

493: essentially dominated by the behavior

494: of infinitesimal perturbations, 

495: most of the features can be understood

496: by limiting the analysis to the tangent space.

497: 

498: The evolution in tangent space is

499: obtained by linearizing 

500: Eq.~(\ref{eq:cml}), i.e.

501: \begin{eqnarray}

502: \delta x_{i}(t+1)&=&F^\prime(\tilde{x}_i(t)) [ \delta x_i(t)+

503: \nonumber\\

504: \frac{\varepsilon}{2}

505: (\delta x_{i+1}(t)&-&2 \delta x_{i}(t)+\delta x_{i-1}(t)) ]\,,

506: \label{eq:tang}

507: \end{eqnarray}

508: where $F^\prime$ is the first derivative of a one dimensional

509: chaotic map. Let us again consider as initial condition for

510: the evolution of (\ref{eq:tang}) a localized perturbation

511: as  (\ref{eq:tzero}) with $d_0$ infinitesimal.

512: The spatio-temporal dynamics of the 

513: tangent vector $\{ \delta x_i(t) \}$ is

514: determined by the interaction and competition of 

515: two different mechanisms present in Eq.~(\ref{eq:tang}):

516: the chaotic instability and the spatial diffusion.

517: 

518: As a first approximation, the effects of the 

519: two mechanisms can be treated as independent.

520: The chaotic instability leads to an average 

521: exponential growth of the infinitesimal disturbance:

522: $|\delta x_0(t)| \approx {d_0 \exp[\lambda t]}$.

523: On the other hand, the spatial diffusion, due to the

524: coupling, approximately leads to a 

525: spatial Gaussian spreading of the disturbance :

526: $|\delta x_i(t)| \approx 

527: |\delta x_0(t)| 

528: {/ \sqrt{4\pi Dt}} \exp( -{i^2 / 4Dt})$ 

529: where $D=\varepsilon/2$.

530: Combining these two effects one obtains

531: \begin{equation}

532: |\delta x_i(t)| \approx d_0 

533: {1 \over \sqrt{2\pi \varepsilon t}} \exp \left(\lambda t 

534: -{i^2 \over 2 \varepsilon t}\right)\,.

535: \label{eq:shape}

536: \end{equation}

537: Since the chaotic nature of the phenomenon

538: will typically  induces fluctuations, Eq.~(\ref{eq:shape}) can only describe 

539: the average shape of the disturbance. 

540: Moreover, Eq.~(\ref{eq:shape}) holds only 

541: when the perturbation is infinitesimal, since when

542: the disturbance reaches finite values a saturation mechanism 

543: (due to the nonlinearities) sets in preventing the divergence

544: of $|\delta x_i(t)|$. 

545: 

546: To verify the  validity of (\ref{eq:shape}), we studied the evolution

547: of localized perturbations of a homogeneous spatio-temporal 

548: chaotic state, in particular coupled logistic and tent maps

549: have been considered in the regime of ``fully developed turbulence''

550: \cite{K86}. Firstly, the system is randomly initialized and 

551: let to relax for a relatively long transient. 

552: At this stage two replicas of the same system are

553: generated and to one of the two a localized perturbation 

554: (like (\ref{eq:tzero})) is added. The evolution of the 

555: difference field (\ref{eq:dyn}) is then monitored at 

556: successive times.  In order to wash out the fluctuations,

557: the shape of the disturbance is obtained

558: averaging over many distinct realizations.

559: %fig 3

560: %%%%%%%%%%%%%%%%%%%%%%%%

561: \begin{figure}

562: \epsfxsize=8.truecm

563: \epsfysize=6truecm

564: \epsfbox{Fig3.eps}

565: %\narrowtext

566: \caption{Average evolution of perturbations for a CML of logistic maps

567: ($f(x)=4x(1-x)$) for (a) $\varepsilon=1/3$ and (b) $\varepsilon=1/10$. 

568: $\langle \Delta x_i(t)\rangle$ is reported as a function of $i^2$ in a

569: lin-log scale at different 

570: times (from bottom to top $t=10,20,30,40$). Deviations from 

571: a straight line correspond to deviation from the Gaussian shape. 

572: $\langle \Delta x_i(t)\rangle$ is obtained as an average over $10^3$ realizations, 

573: for each one $\Delta x_i(0)$ has been chosen as (\ref{eq:tzero}) 

574: with $d_0=10^{-7}$.

575: For comparison the prediction (\ref{eq:shape})

576: is also reported (dashed lines).}

577: \label{fig:fronte}

578: \end{figure}

579: %%%%%%%%%%%%%%%%%%%%%%%% 

580: As one can see from Fig.~\ref{fig:fronte} Eq.~(\ref{eq:shape}) is fairly well 

581: verified for large enough coupling while it fails at small $\varepsilon$ \cite{nota}. 

582: These discrepancies are due to the finite spatial resolution

583: (that in CML's is always fixed to $1$), since

584: for small diffusivity constant the discretization of the Laplacian 

585: becomes inappropriate.

586: The expression (\ref{eq:shape}) for disturbance evolution

587: has been already proposed in Ref.~\cite{WB93}

588: for CML's in two dimensions. Deviations from

589: (\ref{eq:shape}) have been observed also in 

590: \cite{WB93}, but attributed to anomalous diffusive behaviors.

591: It has to be remarked that expression  (\ref{eq:shape})

592: is valid only at short times, since asymptotically ($t \to \infty$) 

593: the infinitesimal leading edge of the propagating front 

594: $|\delta x_i(t)|$ assumes

595: an exponential profile \cite{TGP95}.

596: 

597: For what concerns the propagation velocity,  

598: an estimation of $V_p$ can be obtained for infinitesimal

599: perturbations by the evaluation of the so-called maximal 

600: comoving Lyapunov exponents $\Lambda(v)$ \cite{DK87}. 

601: The time evolution of an intially localized (infinitesimal) 

602: disturbance (\ref{eq:tzero}) in a reference frame moving 

603: with velocity $v$ can be expressed as 

604: \begin{equation}

605: |\delta x_i(t)| \sim d_0 e^{\Lambda(v)t}\,,

606: \label{eq:growth}

607: \end{equation}

608: by following the perturbation along the world line $i=v t$

609: one can easily measures the corresponding

610: comoving Lyapunov exponent $\Lambda(v)$ (for more

611:  details see \cite{DK87,TP92}).

612: The information propagation velocity is the maximal

613: velocity for which a disturbance still propagates 

614: without being damped. Therefore it can be defined through

615: the following marginal stability criterion \cite{DK87}:

616: \begin{equation}

617: \Lambda(V_L) \equiv 0\,,

618: \label{eq:lvzero}

619: \end{equation}

620: where the velocity has been now indicated with $V_L$ in order

621: to stress that it has been obtained via 

622: a linear analysis. For the maps considered

623: in this section the identity $V_p = V_L$ is always fulfilled.

624: % fig 4

625: 

626: %%%%%%%%%%%%%%%%%%%%%%%%

627: \begin{figure}

628: \epsfxsize=8.truecm

629: \epsfysize=6truecm

630: \epsfbox{Fig4.eps}

631: %\narrowtext

632: \caption{Comparison between the directly measured propagation 

633: velocities $V_p = V_L$ (circles)

634: and the prediction (\ref{eq:velocity}) (boxes) for a CML of 

635: logistic maps with (a)

636: $a=3.9$ and (b) $a=4$, and of tent maps with (c) $a=2$ (d) $a=1.8$.

637: Lattices of $4\cdot 10^4$ maps has been used.}

638: \label{fig:vel}

639: \end{figure}

640: %%%%%%%%%%%%%%%%%%%%%%%%                                       

641: As shown in \cite{DK87}, in a closed system with symmetric coupling

642: $\Lambda=\Lambda(v)$ has typically a concave shape, 

643: with the maximum located at $v=0$ (in particular

644: $\lambda = \Lambda(v=0)$). An approximate expression 

645: can be obtained for $\Lambda(v)$, by substituting 

646: $i=vt$ in (\ref{eq:shape}) and by comparing

647: it with (\ref{eq:growth}) :

648: \begin{equation}

649: \Lambda(v)=\lambda-{v^2 / 2 \varepsilon} \quad .

650: \label{eq:comapr}

651: \end{equation} 

652: This parabolic expression for

653: $\Lambda(v)$ suffers of the same limits mentioned for 

654: the Gaussian approximation (\ref{eq:shape}) for the

655: disturbance evolution. 

656: Anyway, from (\ref{eq:comapr}) an analytical

657: prediction can be obtained for $V_L$ :

658: \begin{equation}

659: V_A=\sqrt{2\varepsilon \lambda}\,,

660: \label{eq:velocity}

661: \end{equation} 

662: which, as shown in  Fig.~\ref{fig:vel}, is indeed very good 

663: apart from some deviations for 

664: $\varepsilon \approx 0$ and $\varepsilon \approx 1$.

665: In Sect.~\ref{sec:5} we will rederive (\ref{eq:velocity}) 

666: by assuming that the chaotic perturbation 

667: behaves as a front connecting a stable and 

668: an unstable (metastable) fixed point

669: in a non-chaotic reaction diffusion system.

670: 

671: Let us briefly recall that another method (not suffering for

672: boundary problems) to determine the comoving Lyapunov exponent

673: has been introduced in~\cite{TP92}. The method relies on

674: the computation of {\it specific} Lyapunov exponents $\lambda(\mu)$ 

675: associated to an exponentially decaying perturbation (with spatial

676: decay rate $\mu$). In other words

677: one assumes that the spatio-temporal evolution of an infinitesimal

678: disturbance can be written as 

679: \begin{equation}

680: |\delta x_i(t)| \sim d_0 e^{\lambda(\mu)t-\mu i}\,.

681: \label{eq:specific}

682: \end{equation}

683: Since the asymptotic leading edge of the front separating perturbed 

684: from unperturbed part of the chain has an exponential shape,

685: the above assumption (\ref{eq:specific}) is appropriate to

686: study its evolution.

687: 

688: It is straightforward to show that the comoving Lyapunov

689: exponents are related to the specific ones via a Legendre

690: transform~\cite{TP92}, all the data concerning comoving exponents

691: reported in this paper have been obtained with such a method. Moreover,

692: a further results concerns the linear velocity $V_L$, it

693: can be shown \cite{TGP95} that

694: its value corresponds to the minimal propagation velocity

695: $V(\mu) = \lambda(\mu) / \mu$ associated with perturbations

696: of the form (\ref{eq:specific}), i.e.

697: \begin{equation}

698: V_L = \min_\mu \frac{\lambda(\mu)}{\mu} \equiv \frac{\lambda_L}

699: {\mu_L}

700: \label{eq:vmu}

701: \end{equation}

702: where $\mu_L$ and $\lambda_L=\lambda(\mu_L)$ represent the

703: spatial decay rate and the temporal growth rate of the leading

704: edge, respectively. The expression (\ref{eq:vmu}) for the linear

705: velocity is identical to the one derived for propagation of fronts

706: connecting a stable to an unstable steady state~\cite{wim1}. 

707: 

708: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

709: \subsection{Non Linear Mechanisms}

710: \label{sec:3.2}                   

711: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

712: In this section we investigate the case of coupled maps for which  

713: $\lambda(\delta) > \lambda(0)$ in some interval of $\delta$.

714: As noticed in Sect.~\ref{sec:2}, 

715: this behavior can be observed in chaotic (absolutely unstable) maps,

716: as well as in stable and marginally stable maps. Let us first

717: analyze chaotic maps, non-chaotic ones will be discussed

718: in Sect. \ref{sec:4.2}.

719: 

720: For these systems it is possible to have $V_p >  V_L$, 

721: this means that the disturbance can still propagate 

722: also in the velocity range $[V_L,V_p]$, 

723: even if the corresponding comoving Lyapunov exponents 

724: are negative. Therefore,

725: the linear marginal stability 

726: criterion (\ref{eq:lvzero}) does not hold anymore. 

727: We want to stress that the condition (\ref{eq:nonlin})

728: is necessary, but not sufficient to ensure that $V_p > V_L$,

729: since all the details of the coupled model play also an important role.

730: 

731: In Fig.~\ref{fig:nonlinear} the spatio-temporal evolution of an

732: initially localized disturbance of a chain of coupled shift

733: maps is reported.

734: As shown in \cite{isola}, when the coupling $\varepsilon \le 1/2$ the

735: maximal Lyapunov exponent for such model coincides with that

736: of the single map (namely, $\lambda=\ln(\beta)$) and if $\beta > 1$

737: the system is chaotic. Initially the perturbation, that is still

738: infinitesimal, spreads with the linear velocity $V_L$

739: above defined. At later time it

740: begins to propagate faster with a velocity $V_P > V_L$.

741: Comparing  Fig.~\ref{fig:nonlinear} with

742: Fig.~\ref{fig:torc.1}, one can see that the second stage of

743: the propagation sets in when the bulk of the perturbed region 

744: reaches sufficiently high values. As a matter of fact the initial 

745: stage of propagation disappears if we initialize the two replicas

746: with a disturbance of amplitude ${\cal O}(1)$. From these facts it is

747: evident that the origin of the information propagation characterized

748: by $V_p > V_L$ should be due to the strong nonlinear effects present

749: in this type of CML.

750: % fig 5

751: 

752: %%%%%%%%%%%%%%%%%%%%%%%%

753: \begin{figure}

754: \epsfxsize=8.truecm

755: \epsfysize=6truecm

756: \epsfbox{Fig5.eps}

757: %\narrowtext

758: \caption{The same of Fig.~\ref{fig:torc.1} for a lattice of coupled shift maps

759: with $\beta=1.03$,  $\varepsilon=1/3$.}

760: \label{fig:nonlinear}

761: \end{figure}

762: %%%%%%%%%%%%%%%%%%%%%%%% 

763: The behavior at long times can be understood by considering

764: the dependence of $\lambda(\delta)$ on the disturbance amplitude 

765: as shown in Fig.~\ref{fig:maps}b:

766: actually  the figure refers to the single map with $\beta=2$, but the 

767: shape of $\lambda(\delta)$ is qualitatively the same also for 

768: the coupled system and for other $\beta$-values.

769: 

770: Until the perturbation is infinitesimal $\lambda(\delta) \simeq \lambda$

771: and the linear analysis applies, when the disturbance becomes

772: bigger than a certain amplitude, $\delta^{NL}$, 

773: the growth will be faster, since

774: now $\lambda(\delta) > \lambda$. As it can be clearly seen in

775: Fig.~\ref{fig:invasion}, the perturbation is well reproduced by 

776: the linear approximation (\ref{eq:shape}) until 

777: the amplitude of the perturbation reaches a critical value

778: $\delta^{NL} \sim O(10^{-4})$ above which the nonlinear effects set in. 

779: At this stage

780: the nonlinear instabilities begin to push the front leading to

781: an increase of its velocity and deforming the profile of the perturbation.

782: This becomes exponential at much shorter times than in the linear

783: situation discussed in previous subsection. Moreover,

784: when the propagation is dominated by nonlinear mechanisms

785: the spatial decay rate $\mu_{NL}$ of the asymptotic leading edge

786: will be greater of the linear expected value $\mu_L$: this

787: result can be explained again invoking the analogy with

788: propagation of fronts connecting steady states \cite{TGP95}.

789: %fig 6

790: %%%%%%%%%%%%%%%%%%%%%%%%

791: \begin{figure}

792: \epsfxsize=8.truecm

793: \epsfysize=6truecm

794: \epsfbox{Fig6.eps}

795: %\narrowtext

796: \caption{Evolution of the perturbation $\Delta x_i(t)$ for a CML 

797: of  shift maps with

798: $\varepsilon=1/3$ and $\beta=1.04$ at four different times, $t=250,450,650,1000$.

799: The solid lines are the expected Gaussian  shape (\ref{eq:shape}).

800: The decay rate of the asymptotic exponential profile is 

801: $\mu_{NL} \sim 1.47$, noticeably greater than the linear 

802: value $\mu_L = 0.42$.}

803: \label{fig:invasion}

804: \end{figure}

805: %%%%%%%%%%%%%%%%%%%%%%%%

806: %fig 7

807: %%%%%%%%%%%%%%%%%%%%%%%%

808: \begin{figure}

809: \epsfxsize=8.truecm

810: \epsfysize=6truecm

811: \epsfbox{Fig7.eps}

812: %\narrowtext

813: \caption{Time evolution of $\Delta x_0(t)$ (solid line) and $\Delta

814: x_i(t)$ (dashed line), for $i=V_2 t$ with $V_2=1/3$. 

815: In this last case the data for two different chain configurations are reported.

816: The map used is the shift map with $\beta=1.1$, $\varepsilon=1/3$ and

817: $L=2\cdot 10^3$. For these parameters one has $V_L=0.250$ and $V_p=0.342$.

818: Note that between $t=200$ and $t=500$ for $\Delta x_i(t)$

819: the numerical precision is reached.}

820: \label{fig:bho}

821: \end{figure}

822: %%%%%%%%%%%%%%%%%%%%%%%% 

823: To better clarify the difference between the linear and non linear mechanisms

824: we show the behavior of $\Delta x_i(t)$ along  the world line $i=V t$ 

825: for two different velocities:  $V_1=0$ and $V_L<V_2<V_p$.

826: From Fig.~\ref{fig:bho}, one observes  for the zero-velocity situation

827: an exponential increase of the perturbation with rate $\lambda$

828: until it eventually saturates. For the case corresponding to

829: velocity $V_2$ an initial exponential decay with rate $\Lambda(V_2)$

830: is seen, followed at later times by a resurgence of the disturbance.

831: The successive evolution of the perturbation is no more exponential and 

832: exhibits strong fluctuations. 

833: 

834: These features suggest that in order to generalize the 

835: criterion (\ref{eq:lvzero}) to nonlinear driven information spreading

836: the growth of finite amplitude perturbations in a moving reference frame

837: should be analyzed.

838: 

839: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

840: \section{Finite Size Lyapunov Exponent: \\

841: Extended Systems}

842: \label{sec:4}

843: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

844: In this section we  introduce the 

845: Finite Size Comoving Lyapunov Exponent (FSCLE) that is

846: a generalization of the  FSLE to a moving reference frame.

847: First of all let us define the FSLE for an extended

848: system: in this case exactly the same definition

849: given in Sect.~\ref{sec:2} applies, apart from

850: some ambiguities in the choice of the

851: norm  to employ for measuring the distance

852: $\delta(t)$ of the perturbed, 

853: ${\bf x}^\prime(t)=\{x_i^\prime(t)\}$, 

854: from the unperturbed replica, ${\bf x}(t)=\{x_i(t)\}$. 

855: A natural choice could be to perturb randomly

856: ${\bf x}(t)$ and to look for the doubling times associated

857: to the evolution of the distance

858: \begin{equation}

859: {\tilde \Delta} (t) = \frac{1}{L} \sum_{i=-L/2}^{L/2} |x_i^\prime(t)-x_i(t)| 

860: \quad ;

861: \label{norm1}

862: \end{equation}

863: an alternative choice consists in

864: perturbing a single site of the chain, let's say $i=0$, 

865: at time $t=0$ and to evaluate the ``single site'' norm

866: \begin{equation}

867: \Delta x_0 (t) = |x_0^\prime(t)-x_0(t)| 

868: \quad .

869: \label{norm2}

870: \end{equation}

871: We have verified that the two norms (\ref{norm1}) and

872: (\ref{norm2}) give equivalent results for what concerns

873: the evaluation of $\lambda(\delta)$. In this paper

874: we will limit to consider the norm (\ref{norm2}).

875: 

876: In order to measure the FSCLE in a reference frame moving

877: with velocity $v$, we have simply measured

878: the difference (\ref{norm2}) along a world line $i=v t$;

879: i.e.

880: \begin{equation}

881: \Delta x_{i_c+[vt]} (t)  = |x_{i_c+[vt]}^\prime(t)-x_{i_c+[vt]}(t)| 

882: \quad .

883: \label{normv}

884: \end{equation}

885: where $[ \, \cdot \,  ]$ denotes the integer part and 

886: $i_c$ is introduced below.

887: 

888: The FSCLE is then estimated as in Sect.~\ref{sec:2}. Once 

889: a set of thresholds $\delta_n = r^n \delta_0$, with

890: $n=0,\dots,N$, is chosen and the perturbation is initialized 

891: as $\Delta x_i(0)=

892: \delta_{min} \, \delta_{i,0}$ with $\delta_{min}\ll \delta_0$. 

893: A preliminary transient evolution is performed in order to

894: allow to the perturbed orbit to relax on the attractor. 

895: At the end of this short transient the position $i_c$

896: where the perturbation reaches its maximum is detected,

897: this point is taken as the vertex of the light cone from

898: which all the considered world lines $i = v \, t$ depart.

899: 

900: The FSCLE is then defined as

901: \begin{equation}

902: \Lambda(\delta_n,v)={1 \over \langle \tau(\delta_n)\rangle_e} \left\langle 

903: \log\left({{\Delta} \over {\delta}_n}\right)

904: \right\rangle_e\,,

905: \end{equation}

906: where the dependence on the velocity $v$ derives from the employed 

907: norm (\ref{normv}). Note that we have used Eq.~(\ref{eq:2.5})

908: due to the discreteness of the temporal evolution.

909: 

910: In the limit of very small perturbation the FSCLE reduces to the comoving

911: Lyapunov exponent 

912: \begin{equation}

913: \lim_{\delta \to 0} \Lambda(\delta,v) =\Lambda(v)\,,

914: \label{eq:fslim}

915: \end{equation}

916: and, for $v=0$ one has the FSLE $\lambda(\delta)$. 

917:  

918: Actually there are finite time effects which prevents the limit 

919: (\ref{eq:fslim}) to be correctly attained.

920: This is related to the fact that the FSCLE's can be obtained 

921: only via finite time measurements. In the Appendix

922: we show how one can include such finite time corrections.

923: 

924: In the next subsections we will give evidences that 

925: the marginal {\it linear} stability criterion

926: (\ref{eq:lvzero}) can be generalized with the aid

927: of the FSCLE in the following way:

928: \begin{equation}

929: \max_{\delta} \{ \Lambda(\delta,v) \} =0\qquad {\mbox {for}} \, v \ge V_p\,,

930: \label{eq:crit}

931: \end{equation}

932: where $V_p$ can be either $V_L$, if the information

933: propagation is due to linear mechanisms, or greater

934: than $V_L$, when nonlinear mechanisms prevail on the

935: linear ones. For $v > V_p$ one has $\Lambda(\delta,v)=0$, 

936: since due to the definition of the

937: FSCLE negative growth rate appear to be $0$.

938: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

939: \subsection{Chaotic Systems}

940: \label{sec:4.1}

941: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

942: Let us first consider chaotic systems for which

943: $V_p \equiv V_L$, in this case 

944: $$\max_{\delta} \{ \Lambda(\delta,v) \} = \Lambda(v)$$

945:  and the generalized criterion (\ref{eq:crit}) reduces to the linear

946: one (\ref{eq:lvzero}). 

947: 

948: As already stressed in Sect. \ref{sec:3.2}, shift coupled

949: maps represent a prototype of the class of chaotic models 

950: for which $V_p$ can be eventually bigger than $V_L$.

951: For this model the behavior of $\Lambda(\delta,v)$ for various 

952: velocities is reported in Fig.~\ref{fig:fscle1}.

953: 

954: For $v<V_L$, we observe that $\Lambda(\delta,v) \sim \Lambda(v)$ up to a 

955: certain value of the disturbance amplitude $\delta^{NL}$, above which

956: the FSCLE increases and exhibits a clear peak at some higher $\delta$-value.

957: From Fig.~\ref{fig:fscle1} (a) it is clear that

958: $\delta^{NL}$ (which denotes the set in of the regime dominated

959: by nonlinear mechanisms) decreases for increasing velocities and finally

960: vanishes at $v = V_L$. This behavior is reasonable since, as shown in 

961: Fig.~\ref{fig:nonlinear}, initially the disturbance evolves

962: along the chain following the linear mechanism 

963: characterized by $\Lambda(v)$, but as soon as in the central site

964: $\delta > \delta^{NL}$ the nonlinear mechanism begins to be active. 

965: Thus a nonlinear front is excited and this invades the linear region

966: propagating with a velocity higher than $V_L$, therefore the higher is $v$ 

967: the smaller is the scale at which nonlinear effects are observed.

968: 

969: In the interval $V_L \leq v \leq V_p$ (see Fig.~\ref{fig:fscle1}(b)), 

970: $\Lambda(\delta,v)$ is still 

971: positive and its maximal value decreases for increasing $v$. 

972: The FSCLE vanishes for $v=V_p$ as Fig.~\ref{fig:vlvnl} (a) shows. 

973: In this velocity range $\Lambda(v)$ is always negative,

974: therefore the instabilities observed in a reference frame

975: moving with velocity higher than $V_L$ have a fully nonlinear

976: origin. The residual fluctuations present in $\Lambda(\delta,v)$ are 

977: essentially due to the limited statistics.

978: Indeed the nonlinear growth as shown in

979: Fig.~\ref{fig:bho} is extremely fluctuating.

980: %fig 8

981: 

982: %%%%%%%%%%%%%%%%%%%%%%%%

983: \begin{figure}

984: \epsfxsize=8.truecm

985: \epsfysize=6truecm

986: \epsfbox{Fig8a.eps}

987: \epsfxsize=8.truecm

988: \epsfysize=6truecm

989: \epsfbox{Fig8b.eps}

990: %\narrowtext

991: \caption{$\Lambda(\delta,v)$ as a function $\delta$ 

992: for various velocities. The data

993: refer to the coupled shift maps with  $\beta=1.1$ and 

994: $\varepsilon=1/3$, with these parameters

995: $V_L \sim 0.250$ and $V_p \sim 0.342$.

996: In (a) results for velocities $v < V_L$ are reported,

997: namely (from top to bottom) $v=0$, $v=0.10$ and $v=0.16$. 

998: The straight lines indicate $\Lambda(v)$. 

999: In (b) velocities in the range $[V_L : V_p]$ are reported, from

1000: the top $v=0.25$, $v=0.286$, $v=0.30$, $v=0.33$, $v=0.34$. 

1001: Chains of lengths from $L=10^4$ up to $L=10^5$

1002: have been employed and the statistics is 

1003: over $2 \cdot 10^3$ doubling times.}

1004: \label{fig:fscle1}

1005: \end{figure}

1006: %%%%%%%%%%%%%%%%%%%%%%%% 

1007: In Figure~\ref{fig:vlvnl}(a) the dependence of 

1008: $\max_{\delta}\{\Lambda(\delta,v)\}$ on $v$ for coupled shift maps is 

1009: reported, as expected it vanishes exactly for $v=V_p$.  

1010: For $v < V_L$ a non-monotonous behavior of

1011: $\max_{\delta}\{\Lambda(\delta,v)\}$ is observable,

1012: this is probably due to the complex interplay of linear and nonlinear

1013: effects. For $v > V_L$, 

1014: the behavior is smoother and a monotonous decrease is observed.

1015: As discussed in Sect.~\ref{sec:2}, the discontinuity present

1016: in the shift map is not necessary in order to observe

1017: nonlinear mechanisms prevailing on linear ones.

1018: 

1019: As an example of continuous map exhibiting an information

1020: propagation velocity $V_p > V_L$

1021: the map (\ref{eq:stable}) is considered. 

1022: This type of CML has been already studied in \cite{TP94}:

1023: it has been observed that in a certain parameter range

1024: $V_p$ can be finite even if the map is non chaotic. 

1025: Moreover, 

1026: also in the chaotic regime there is a window of parameters

1027: where $V_p > V_L$, with our parameters choice $\lambda \approx 0.182>0$.

1028: 

1029: %fig 9

1030: 

1031: %%%%%%%%%%%%%%%%%%%%%%%%

1032: \begin{figure}

1033: \epsfxsize=8.truecm

1034: \epsfysize=6truecm

1035: \epsfbox{Fig9a.eps}

1036: \epsfxsize=8.truecm

1037: \epsfysize=6truecm

1038: \epsfbox{Fig9b.eps}

1039: %\narrowtext

1040: \caption{$\max_{\delta}\Lambda(\delta,v)$ 

1041: (dashed line with points) versus $v$ 

1042: compared with $\Lambda(v)$ (continuous line).

1043: (a) Results for coupled shift maps with the same parameters as Fig.~\ref{fig:fscle1},  

1044: the vertical lines indicates  $V_L \approx 0.250$ and the 

1045: measured (\ref{eq:vel1}) propagation velocity $V_p \approx 0.342$. 

1046: (b) Data for the coupled maps (\ref{eq:stable}) with parameters 

1047: $b=2.7$, $d=0.1$, $q=0.07$ 

1048: and $c=500$ and with coupling $\varepsilon=2/3$,

1049: the vertical lines indicates $V_L

1050: \approx 0.39$ and $V_p \approx 0.59$. 

1051: In this second case chains of length $L=2\cdot 10^4$ have been considered

1052: and the statistics is over $2 \cdot 10^4$ doubling times.                            

1053: The reported values for $\max_{\delta}\Lambda(\delta,v)$ 

1054: refer to an average over five values of $\delta$ around the 

1055: peak position of $\Lambda(\delta,v)$.

1056: }

1057: \label{fig:vlvnl}  

1058: \end{figure}

1059: %%%%%%%%%%%%%%%%%%%%%%%%

1060: Also in the present case we observe an overall behavior

1061: of the FSCLE resembling that of the coupled shift maps.

1062: The main point that we want to remark is that in the limit

1063: $v \to V_p$  $\Lambda(\delta,v) \to 0$,

1064: as shown in Fig.~\ref{fig:vlvnl} (b). 

1065: At variance with the shift map (that is an everywhere

1066: expanding map) the considered map shows contracting 

1067: and expanding intervals. Therefore the disturbances

1068: during their spatio-temporal evolution can be

1069: alternatively expanded or contracted. This leads

1070: to strong fluctuations in the FSCLE-values, that are 

1071: difficult to remove. 

1072: 

1073: As a matter of fact we observed

1074: that even for velocities slightly larger than $V_p$

1075: $\Lambda(\delta,v)$ can be non zero. But the statistics

1076: of these anomalous fluctuations is extremely low,

1077: referring to the parameters reported in Fig.~\ref{fig:vlvnl} (b)

1078: for $v=0.6 > V_p = 0.59$ we observed an expansion (instead

1079: of the expected contraction) of disturbances in the 

1080: $0.5 \%$ of the studied

1081: cases. For higher velocities $\Lambda(\delta,v)$ is

1082: zero for any considered $\delta$.

1083: 

1084: 

1085: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1086: \subsection{Non Chaotic Systems}

1087: \label{sec:4.2}

1088: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1089: On the basis of the linear analysis discussed in Sect.~\ref{sec:3.1}

1090: the propagation of disturbances should be present only

1091: in chaotic systems. 

1092: For non chaotic ones $\lambda \leq 0$ and from Eq.~(\ref{eq:lvzero}) 

1093: one trivially obtains $V_L=0$. 

1094: On the other hand, in the class of systems for which 

1095: $\lambda(\delta)\geq \lambda$ 

1096: there are also stable and marginally stable systems. Therefore,

1097: a propagation due solely to non linear terms can

1098: still be present \cite{PLOK93,TP94,TGP95}. 

1099: In this section we want to discuss 

1100: the possible employ of the FSLE  to characterize these maps.

1101: 

1102: Before entering into the description of the FSLE computation in such systems,

1103: it is of interest to recall an important phenomenon which appears in 

1104: stable systems: the so-called ``stable chaos'' \cite{PLOK93}. 

1105: Stable systems asymptotically  evolve towards 

1106: trivial attractors (i.e. fixed points, periodic or quasi-periodic orbits).

1107: However, in spatially extended systems it may happen that the time needed 

1108: to reach the asymptotic state is very long: it has been found that in certain 

1109: stable CML the transient time diverges exponentially with the number of 

1110: elements of the chain \cite{CK88}. 

1111: Moreover, this transient is characterized by a quasi-stationary

1112: behavior allowing for the investigation of the properties of the 

1113: model with statistical consistency. In Ref. \cite{PLOK93}

1114: it has been shown that a chain of coupled maps of the type (\ref{eq:stable}),

1115: considered in their discontinuous limit (i.e. for $c \to \infty$),

1116: is non chaotic but still exhibits erratic behaviors. This

1117: is associated to a non-zero information spreading within

1118: the system.

1119: 

1120: 

1121: As far as the computation of the FSLE for these systems is concerned,

1122: some remarks are worth to be done. The definition of the

1123: FSLE in term of error doubling times cannot be used 

1124: in a straightforward manner to determine negative expansion rates.

1125: Another important point is that, at variance with the case 

1126: of chaotic maps, finite perturbations should be now considered

1127: in order to observe an expansion.

1128: 

1129: These two points impede a straightforward implementation of our method to

1130: study these systems. Indeed, if one starts with too small perturbations  

1131: the propagation does not manifest, while if one initializes the system with a

1132: finite perturbation 

1133: $\lambda(\delta)$ cannot be estimated with the required accuracy. 

1134: Indeed, the only way to have an independence of $\lambda(\delta)$ from 

1135: the initial conditions is to initialize the system with infinitesimal

1136: perturbations and then to follow them until they become finite due to

1137: the dynamics of the system (see Ref.~\cite{ABCPV96} for a detailed discussion).

1138: 

1139: The coupled (\ref{eq:stable}) maps for $c\to \infty$ have been recently

1140: analyzed by Letz \& Kantz \cite{Kantz} in term of 

1141: an indicator similar to the FSLE (i.e. able to quantify 

1142: the growth rate of non infinitesimal perturbations). This

1143: indicator turns out to be negative for infinitesimal 

1144: perturbations and becomes positive for finite perturbations.

1145: This means that a finite perturbation of sufficient amplitude 

1146: can propagate along the system due to nonlinear effects.

1147: This confirms what was previously observed

1148: in Ref.~\cite{TGP95} for marginally stable systems.

1149: %fig 10

1150: 

1151: %%%%%%%%%%%%%%%%%%%

1152: \begin{figure}

1153: \epsfxsize=8.truecm

1154: \epsfysize=6truecm

1155: \epsfbox{Fig10.eps}

1156: %\narrowtext

1157: \caption{$I(\delta)$ as a

1158: function of the amplitude perturbation in

1159: a lin-log scale for the discontinuous coupled 

1160: maps (\ref{eq:stable}), in the limit $c \to \infty$, with

1161: $\varepsilon=1/3$ and $L=35$. The map is reported in the inset.

1162: The negative maximal Lyapunov $\lambda = -0.105$ is recovered

1163: at small scales. For the computation details see the text.

1164: In the inset the single map is shown.}

1165: \label{fig:stable}                                      

1166: \end{figure}

1167: %%%%%%%%%%%%%%%%%%%

1168: In Fig.~\ref{fig:stable} we show the behavior of 

1169: a quantity $I(\delta_n)$ similar to $\lambda(\delta_n)$ which 

1170: has been obtained as follows.

1171: We considered two trajectories at an initial distance $\delta_n$,

1172: after one time step evolution the distance $\delta$ 

1173: between the trajectory is measured. Then one of the

1174: two trajectory is rescaled at a distance 

1175: $\delta_n$ from the other, keeping the direction of the perturbation unchanged,

1176: and the procedure is repeated several times and for several

1177: values of $\delta_n$.

1178: Then we averaged $\ln(\delta/\delta_n)$ over many different initial conditions

1179: obtaining $I(\delta_n)$. For $\delta_n \to 0$, this is nothing 

1180: but  the usual algorithm for computing the maximal Lyapunov exponent~\cite{benettin}.

1181: 

1182: As discussed in Ref.~\cite{ABCPV96} this method suffers of the problem that 

1183: when $\delta_n$ is finite one is not sure to sample correctly the measure on the 

1184: statistically stationary state. Indeed a finite perturbation will generically bring 

1185: the trajectory out from the ``attractor''. Nevertheless the result is in good agreement 

1186: with the one obtained in \cite{Kantz} and confirms that at the origin 

1187: of the perturbation propagation in this system there should be a mechanism very 

1188: similar to the one discussed for the shift map. Further studies related to the

1189: ``stable chaos'' phenomenon have been recently performed~\cite{ginelli}.

1190: 

1191: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1192: \section{Analogies with front propagation in reaction diffusions systems}

1193: \label{sec:5}

1194: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1195: The perturbation evolution in spatially distributed systems 

1196: can be described as the motion of an interface separating 

1197: perturbed from unperturbed regions. 

1198: In this spirit, one can wonder if and to what extent it is possible to draw

1199: an analogy between the evolution of this kind of interface and the propagation 

1200: of fronts connecting steady states in reaction diffusion systems. 

1201: As already noticed in Ref.~\cite{TGP95}, the two phenomena display many similarities.

1202: In the following we will discuss the similarities and differences, in particular 

1203: we will introduce a simple phenomenological model which can help us in highlighting

1204: the analogies.

1205: 

1206: Let us start by recalling the basic features of fronts propagation in 

1207: reaction diffusion systems  with reference to

1208: the Fisher-Kolmogorov-Petrovsky-Piscounov (FKPP) equation \cite{KPP}

1209: \begin{equation}

1210: \partial_t \theta(z,t) = D\partial^2_{zz}\theta(z,t) + G(\theta(z,t))\,,

1211: \label{eq:kpp}

1212: \end{equation}

1213: where $\theta(z,t)$ represents the concentration of a reactants which diffuses and 

1214: reacts, the chemical kinetics is given by $G(\theta)$. Typically the function 

1215: $G(\theta)\in C^1[0,1]$ (with $G(0)=G(1)=0$) exhibits

1216: one stable ($\theta=1$) and one unstable ($\theta=0$) fixed point.

1217: Once the system is prepared on the stable state ($\theta(z)=0 \, \forall z$),

1218: an initial (sufficiently steep) perturbation (e.g. a step function

1219: $\theta(z,t=0) = \Theta{(z-z_0)}$ ) will give rise to a smooth front 

1220: moving with a velocity $V_p$, that will connect

1221: the unstable and the stable fixed points:

1222: as a result the stable state will invade the unstable one.

1223: This equation admits many different traveling solutions that

1224: are typically characterized by their propagation velocities,

1225: however for a sufficiently steep initial perturbation

1226: of the unstable state the selected front is unique

1227: and its velocity $V_p$ is bounded in the interval $[V_{min}, V_{max}]$,

1228: where

1229: \begin{equation}

1230: V_{min}=2\sqrt{D G^{'}(0)}

1231: \label{eq:vmin}

1232: \end{equation}

1233: and

1234: \begin{equation}

1235: V_{max}=2\sqrt{D

1236: \sup_{0<\theta<1}\left\{ {G(\theta) \over \theta} \right\}}\,.

1237: \label{eq:vmax}

1238: \end{equation}  

1239: If $G(\theta)$ is concave $\sup_{0<\theta<1}\{G(\theta)/\theta\}$ is attained

1240: at $\theta=0$ (i.e. $\sup_{0<\theta<1}\{G(\theta)/\theta\}=G^{'}(0)$) and the

1241: selected velocity is always the minimum one: the front is ``pulled'' by the 

1242: growth of infinitesimal perturbations of the unstable state.

1243: Otherwise, if $G(\theta)$ is convex one can observe a velocity

1244: $V_p > V_{min}$: the front is now ``pushed'' by the growth 

1245: of finite amplitude perturbations. These

1246: results are known as the Aronson \& Weinberger theorem

1247: \cite{ara}

1248: (for more details see Ref. \cite{wim2}).

1249: The velocity $V_{min}$ can be easily obtained by performing a

1250: linear analysis of (\ref{eq:kpp}) and by employing a marginal

1251: stability criterion \cite{wim1}.

1252: 

1253: The subject of our analysis is the spreading of perturbations 

1254: in a chaotic media. In order to compare our case with

1255: the FKPP, one we should consider the time evolution of the difference 

1256: of two chaotic trajectories $\{\Delta x_i(t)\}_{i=1,L}$. 

1257: However, the nature of the two phases separated by the front is now different from 

1258: the FKPP case. 

1259: The interface separates an unstable ($\Delta x_i=0$) state from a ``statistically

1260: stable'' one. With the term ``statistically stable'' we mean that behind the front,

1261: in the bulk of the perturbed region, $\Delta x_i(t)$ does not converge to a stable 

1262: fixed point (as for the FKPP) but it fluctuates in a stationary way

1263: around an average value. This suggests that a good model for reproducing

1264: this dynamical evolution would be a FKPP with a stochastic kinetics~\cite{armero}.

1265: However, many similarities with FKPP can be

1266: established by neglecting the chaotic fluctuations and

1267: considering the average shape of the front

1268: \cite{TGP95}. In practice, this can be done by averaging

1269: the perturbation evolution over many different initial conditions.

1270: 

1271: Once the chaotic fluctuations are neglected, one can 

1272: express the average perturbation growth in any site of the 

1273: chain via the following mean-field approximation:

1274: \begin{equation}

1275: u_i(t+1)=\exp[\lambda(\tilde{u}_i(t))]\, \tilde{u}_i(t)\,,

1276: \label{eq:model}

1277: \end{equation}

1278: where $i$, $t$ are the discrete space and time indices and 

1279: $\tilde{u}_i=(1-\varepsilon)u_i+\varepsilon/2(u_{i+1}+u_{i-1})$. Here, 

1280: $u_i(t)$ indicates the ``average'' 

1281: $\Delta x_i(t)$ and  $\lambda(u)$ the corresponding

1282: FSLE (at least for positive growth rates).

1283: In the limit $u\to 0$  $\lambda(u)\to \lambda$, while,

1284: for $u \sim {\cal O}(1)$, $\lambda(u)$ should reflect the 

1285: saturation effects associated to the nonlinear map. 

1286: In the infinitesimal limit ($u \to 0$)

1287: one essentially recovers the model discussed in Ref.~\cite{PP98}.

1288: 

1289: By passing to continuous variables is possible to

1290: show that Eq.~(\ref{eq:model}) can be 

1291: reduced to Eq.~(\ref{eq:kpp}) (at least at the

1292: leading order).

1293: In order to transform Eq.~(\ref{eq:model}) in its

1294: continuous version,

1295: let us introduce infinitesimal spatial and temporal 

1296: resolutions $dx$ and $dt$, and  assume that the

1297: diffusive scaling holds, i.e. $dx^2=dt$. 

1298: Limiting to the first order expansion in $dt$

1299: (second order in $dx$), one easily shows that

1300: the continuous counterpart of Eq.~(\ref{eq:model}) is

1301: \begin{equation}

1302: \partial_t u= \lambda(u)u+\frac{\varepsilon}{2} \partial^2_{xx}u\,,

1303: \label{eq:model_kpp} 

1304: \end{equation}

1305: which is nothing but (\ref{eq:kpp}) with 

1306: \begin{equation}

1307: \lambda(u)={G(u)\over u}\,,

1308: \label{eq:mapping}

1309: \end{equation}

1310: and $D=\varepsilon/2$. 

1311: As a consequence of the identity (\ref{eq:mapping}),

1312: one should observe a pulled dynamics for the

1313: chaotic front if $\lambda \ge \lambda(u) \quad \forall u \ge 0$,

1314: and a pushed one could occur

1315: only if $\max_{u}\{\lambda(u)\} > \lambda$.

1316: This can be considered as a reformulation of the 

1317: Aronson \& Weinberger theorem \cite{ara} in the

1318: context of information propagation.

1319: The velocity bounds (\ref{eq:vmin}),

1320: (\ref{eq:vmax}) can be now identified with 

1321: $V_a = \sqrt{2\varepsilon \lambda}$, i.e. Eq.~(\ref{eq:velocity}), 

1322: the first one and with

1323: \begin{equation}

1324: V_s=\sqrt{2\varepsilon \max_{u}\{\lambda(u)\}}

1325: \end{equation}

1326: the second one. We stress again that the lower velocity

1327: bound is indeed represented by $V_L$ and that it

1328: coincides with $V_a$ only for sufficiently strong 

1329: diffusive coupling $\varepsilon$.

1330: 

1331: Another interesting point that we can investigate via a 

1332: numerical simulation of the mean field effective 

1333: equation (\ref{eq:model})

1334: is the dependence of the propagation properties on the

1335: specific shape of $\lambda(u)$. 

1336: For instance, as previously conjectured, we expect that 

1337: if a monotonous decreasing $\lambda(u)$ is considered

1338: one should observe linear propagation, only.

1339: Indeed, numerical integrations of the model (\ref{eq:model}), 

1340: with the choice $\lambda(u)=A-Bu$, being $A$ and $B$ positive

1341: constants, show that $V_p =V_a= \sqrt{2 \varepsilon A}$,

1342: where $A$ corresponds to the maximal Lyapunov of the effective model.

1343: A generalization of this simple model would require to 

1344: consider $A$ and $B$ as fluctuating quantities generated by

1345: suitable stochastic processes. But while for the choice 

1346: of $A$ one can have some hint \cite{PP98}, this is not 

1347: the case for $B$. 

1348: %fig 11

1349: 

1350: %%%%%%%%%%%%%%%%%%%%%%%%

1351: \begin{figure}

1352: \epsfxsize=8.truecm

1353: \epsfysize=6truecm

1354: \epsfbox{Fig11.eps}

1355: %\narrowtext

1356: \caption{Propagation velocities 

1357: (symbols) for the model (\ref{eq:model}) 

1358: with $\lambda(u)$ given by

1359: (\ref{eq:mod1}) as a function of $B = \ln(r_2)$,

1360: with $\varepsilon=1/3$,  $A=\ln{1.1}$ and

1361: $C=A$. The threshold  values are fixed to

1362: $\delta^{NL}= 10^{-3}$ and $\delta^{sat}=0.53$

1363: and a chain of $4 \cdot 10^4$ sites has been used.

1364: The curves refer to $V_{a}=\sqrt{2\varepsilon A}$ (solid line)

1365: and $V_{s}=\sqrt{2\varepsilon B}$ (dashed line).

1366: Note that in the linear case, $B<A$,

1367: there is a perfect agreement between the measured velocity and the

1368: prediction. 

1369: }

1370: \label{fig:vlvnl_model}  

1371: \end{figure}

1372: %%%%%%%%%%%%%%%%%%%%%%%%

1373: Let us now consider the non linear propagation case, as we have

1374: previously seen a necessary condition in order

1375: to have $ V_p > V_L$ is that $\max_{u}\{\lambda(u)\} > \lambda$

1376: for some finite $u$. For what concerns the actual value of the

1377: selected velocity, this will depend on the value of the diffusive

1378: coupling and on the specific shape of $\lambda(u)$. 

1379: In the following we will examine the dependence of 

1380: $V_p$ on two quantities that characterize $\lambda(u)$:

1381: the difference $|\max_{\delta} \lambda(\delta) - \lambda|$ and the

1382: scale $\delta^{NL}$ at which the non linear effects set in.

1383: 

1384: As a first example let us consider $\lambda(u)$ as a step function:

1385: \begin{equation}

1386: \lambda(u)=\left\{ \begin{array}{ll}

1387: A & 0<u<\delta^{NL}\\ 

1388: B & \delta^{NL}\leq u <\delta^{sat}\\ 

1389: -C & u\geq \delta^{sat}\,,

1390: \end{array}

1391: \right.

1392: \label{eq:mod1}

1393: \end{equation}

1394: where $A$, $B$ and $C$ are positive quantities, with $B > A$,

1395: while $\delta^{NL}$ and $\delta^{sat}$ are amplitude

1396: thresholds. The parameter $A$ is nothing but the Lyapunov exponent, 

1397: $B$ mimics the non linear terms leading to

1398: an enhancement of the growth rate, and the last term mimics the 

1399: damping of the perturbation due to the saturation effects.

1400: In Fig.~\ref{fig:vlvnl_model} it is reported 

1401: the behavior of the propagation velocity

1402: for the model (\ref{eq:model}) with $\lambda(u)$ given by

1403: (\ref{eq:mod1}) for various values of $B$, once $A$ and $C$ are

1404: fixed. From the figure is evident that if $B < A$ then 

1405: $V_p \equiv V_a$ (i.e. we are in the linear regime), 

1406: while as soon as $B > A$ an increase of $V_p$

1407: with respect to $V_a$ is observed.

1408: % fig 12

1409: 

1410: %%%%%%%%%%%%%%%%%%%%%%%%

1411: \begin{figure}

1412: \epsfxsize=8.truecm

1413: \epsfysize=6truecm

1414: \epsfbox{Fig12.eps}

1415: %\narrowtext

1416: \caption{Propagation velocities $V_p$ as a function of $\delta^{NL}$

1417: for the model (\ref{eq:model}), (\ref{eq:mod1}) with $\varepsilon=1/3$, 

1418: $A=\ln(1.1)$, $B=\ln(1.3)$, $C=A$, $\delta^{sat}_0=0.53$. 

1419: The crosses refer to the case $\delta^{sat}=\delta^{sat}_0$,

1420: while the boxes to the case $\delta^{sat}=\delta_0^{sat}+\delta^{NL}$.

1421: The two lines correspond to $V_{a}=\sqrt{2\varepsilon A}$ and 

1422: $V_{s}=\sqrt{2\varepsilon B}$.}

1423: \label{fig:step}  

1424: \end{figure}

1425: %%%%%%%%%%%%%%%%%%%%%%%%

1426:  In the whole examined parameter 

1427: range $V_p$ is always reasonably approximated by $V_s$, but smaller.

1428: By increasing $B$ one observes an increase of the difference

1429: between the measured velocity and the linear prediction.

1430: These results confirm that the condition (\ref{eq:nonlin}) is indeed a

1431: necessary condition in order to observe nonlinear propagation

1432: of information. 

1433: 

1434: We will now investigate the role of $\delta^{NL}$ 

1435: in determining the propagation velocity. It is quite obvious that

1436: modeling the dynamics via the Eqs.~(\ref{eq:model}) and

1437: (\ref{eq:mod1}) the nonlinear effects will disappear in

1438: the limit $\delta_{NL}\to \delta^{sat}$ and this is indeed

1439: confirmed by the simulations (see Fig. \ref{fig:step}).

1440: In order to examine a less obvious situation,

1441: a modification of the expression (\ref{eq:mod1})

1442: is also considered. In this second case 

1443: $\delta^{sat}=\delta^{sat}_0+\delta^{NL}$

1444: is assumed, therefore by varying $\delta^{NL}$

1445: the extension of the amplitude interval over which the nonlinear

1446: mechanism is active will not be modified (being fixed to $\delta^{sat}_0$).

1447: But also in this second case $V_p \to V_a$ for increasing

1448: $\delta^{NL}$, this indicates that the smaller are the perturbation

1449: amplitudes affected by the nonlinear mechanisms

1450: the stronger will be the nonlinear effect on the velocity :

1451: $V_p \to V_s$ for $\delta^{NL} \to 0$. 

1452: 

1453: Typically, in generic CML's by varying a control parameter 

1454: both the difference $|\max_{\delta} \lambda(\delta) - \lambda|$  

1455: as well as $\delta^{NL}$ will change. Moreover,

1456: even the definition of $\delta^{NL}$ for a continuous 

1457: $\lambda(\delta)$ is not obvious. In order to understand

1458: the validity of the mean-field approximation

1459: (\ref{eq:mod1}) in a more realistic case we consider

1460: the shift map. In particular, for $\lambda(u)$ we

1461: employed the analytical expression (\ref{eq:lambda_shift})

1462: valid for the single uncoupled map.

1463: %fig 13

1464: 

1465: %%%%%%%%%%%%%%%%%%%%%%%%

1466: \begin{figure}

1467: \epsfxsize=8.truecm

1468: \epsfysize=6truecm

1469: \epsfbox{Fig13.eps}

1470: %\narrowtext

1471: \caption{Information propagation velocities for the shift map with 

1472: $\varepsilon=1/3$: (boxes) linear velocities $V_L$ and (crosses) directly

1473: measured nonlinear ones. The two lines correspond to

1474: $V_{a}$ (\ref{eq:velocity}) (dotted line) 

1475: and to the propagation velocity for the model 

1476: (\ref{eq:model}) with $\lambda(u)$ 

1477: given by (\ref{eq:lambda_shift}) (solid line). The dashed curve with asterisks is 

1478:  $V_{s}=\sqrt{2 \varepsilon \max_{\delta}\{\lambda(\delta)\}}$. }

1479: \label{fig:vlvnl_shift}  

1480: \end{figure}

1481: %%%%%%%%%%%%%%%%%%%%%%%%

1482: 

1483: The FSLE for coupled shift maps is actually different from 

1484: (\ref{eq:lambda_shift}), but we are neglecting correlations 

1485: among different sites and effects due to the specific measure 

1486: associated to the system. 

1487: Nevertheless, a numerical integration of Eq.~(\ref{eq:model})

1488: equipped with (\ref{eq:lambda_shift}) reproduces 

1489: semi-quantitatively the 

1490: features observed for a lattice of coupled shift maps 

1491: (see Fig.~\ref{fig:vlvnl_shift}). In particular, also the

1492: simple mean-field model (\ref{eq:model}) with the choice

1493: (\ref{eq:lambda_shift}) is able to forecast the

1494: observed transition

1495: from nonlinear to linear behavior for $\beta \to 2$.

1496: 

1497: Concluding this section we can safely affirm that (\ref{eq:model}) is a 

1498: reasonable 

1499: model to mimic the perturbation evolution at a mean field level, neglecting

1500: the spatio-temporal fluctuations and correlations. In the same fashion

1501: $\lambda(u)$ can be considered as an ``effective'' non linear kinetics 

1502: for the perturbation evolution.

1503: 

1504: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1505: \section{Final remarks}

1506: \label{sec:6}

1507: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1508: 

1509: In this paper information (error) propagation in extended chaotic

1510: systems has been studied in details. In particular, we have analyzed

1511: the relevance of linear and nonlinear mechanisms 

1512: for the propagation phenomena in spatio-temporal chaotic

1513: coupled map lattices. Linear stability analysis is

1514: not always able to fully characterize disturbance propagation.

1515: This is particularly true for (marginally) stable systems

1516: with strong nonlinearities,

1517: where finite size perturbations are responsible for information

1518: spreading in the system. When the nonlinear effects prevail

1519: on the linear ones the propagation velocity of information

1520: $V_p$ can be higher than the linear velocity $V_L$. A

1521: necessary condition for the occurrence of information

1522: spreading induced by nonlinear mechanisms has been

1523: expressed in terms of the Finite Size Lyapunov Exponent (FSLE).

1524: We have also shown the existence of strong analogies between

1525: error propagation and front propagation in reaction-diffusion

1526: models. In particular, the above mentioned necessary condition

1527: is analogous to the Aronson \& Weinberger theorem \cite{ara} for 

1528: front connecting stable and unstable steady states.

1529: In the linear and nonlinear case, the propagation velocity 

1530: $V_p$ can be identified via an unique marginal stability criterion

1531: involving Finite Size Lyapunov Exponents defined in a

1532: moving reference frame. This result generalizes the 

1533: corresponding linear criterion expressed in terms of 

1534: the maximal comoving Lyapunov exponents

1535: for the identification of $V_L$ \cite{DK87}. 

1536: 

1537: These results can be of some interest for

1538: the synchronization and the control of extended 

1539: systems. It has been recently shown that

1540: the synchronization of coupled extended 

1541: systems is strongly influenced

1542: by nonlinear effects. In particular, 

1543: the synchronization time exponentially

1544: diverges with the system size, even in non-chaotic

1545: situations, provided that $V_p > 0$ \cite{noi}.

1546: We believe that in these systems the appropriate

1547: indicator to characterize such transition

1548: would be the ``transverse'' Lyapunov 

1549: exponent~\cite{pecora}, once extended to finite scales.

1550: As far as control schemes are concerned, since

1551: they rely mainly on linear analysis~\cite{ogy},

1552: new nonlinear methods (e.g. based on the concept of FSLEs)

1553: should be introduced in order to control the erratic 

1554: behaviors due to fully nonlinear mechanisms.

1555: 

1556: A further aspect that should be addressed in future work

1557: concerns the extension of the applicability of the FSLE

1558: also to linearly stable systems: a candidate in this

1559: respect could be the indicator recently introduced

1560: in \cite{Kantz}. 

1561: 

1562: Finally, it is reasonable to expect that the

1563: present analysis is not limited to discrete

1564: models but that it can be applied to continuous 

1565: extended systems described in terms of PDE's,

1566: e.g. to the complex Ginzburg-Landau equation.

1567: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 

1568: \acknowledgments

1569: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1570: Stimulating interactions with F.~Cecconi, R.~Livi, K.~Kaneko,

1571: A.~Pikovsky, A.~Politi, and A.~Vulpiani are gratefully acknowledged.

1572: Part of this work has been developed at the Institute of Scientific 

1573: Interchange in Torino, during the workshop on ``Complexity and

1574: Chaos'' in October 1999.

1575: We acknowledge CINECA in Bologna and INFM for providing us access to the 

1576: parallel CRAY T3E computer under the grant ``Iniziativa Calcolo

1577: Parallelo''.

1578: 

1579: 

1580: \begin{appendix}

1581: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 

1582: \section{}

1583: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% 

1584: 

1585: In any numerical computation of the Lyapunov exponents one is forced to

1586: use a finite time approximation for an infinite time limit. 

1587: Nevertheless, provided that the convergence to the asymptotic value 

1588: is fast enough, this is not a dramatic problem.

1589: As a matter of fact, in low dimensional systems

1590: very fast convergence to the asymptotic value is

1591: usually observed.

1592: However, this problem manifest more

1593: dramatically in high dimensional systems, where the time to align along 

1594: the direction of maximal expansion could be very long \cite{PP98,P93,Orszag}.

1595: 

1596: In the present case this problem is

1597: complicated by the fact that the FSLE is intrinsically a finite time 

1598: indicator. Indeed the time a perturbation takes to grow from a

1599: value $\delta$ to $r \delta$ is finite unless $\delta \to 0$.

1600: However, for the CML models here analyzed and for initially

1601: localized perturbation (\ref{eq:tzero}) it is possible to evaluate 

1602: the corrections to apply to the FSCLE, estimated at

1603: finite time, in order to recover, for sufficiently small $\delta$,  

1604: the expected limit $\Lambda(v)$. 

1605: 

1606: These corrections allow for a faster convergence

1607: of the FSCLE to its asymptotic values.

1608: 

1609: For the sake of simplicity we consider maps with constant slope, 

1610: e.g. the shift map $F(x)=\beta\, x\;\;{\mbox {mod}}\,1$,

1611: and the CML defined in Eq.~(\ref{eq:cml}).

1612: In this case (for $\varepsilon < 1/2$) the maximal Lyapunov exponent

1613: of the CML coincides with that of the single map \cite{isola};

1614: i.e.  $\lambda=\ln \beta$. We will limit to the case

1615: $v=0$ (i.e to the FSLE), since the extension to generic $v$ is 

1616: straightforward.

1617: As shown in ~\cite{P93,lyapcorr} the finite time evolution

1618: of an infinitesimal perturbation $d_0$ initially

1619: localized in $i=0$ can be expressed at time $T$ as the sum of the 

1620: contributions $M(m,T)$ associated to all the paths connecting

1621: the space-time point $(i=0,t=0)$  to the point $(i=0,t=T)$,

1622: i.e.

1623: \begin{equation}

1624: \frac{\delta x_{i=0}(T)}{d_0} = \beta^T \sum_{m=0}^T M(m,T)  \quad .

1625: \label{path}

1626: \end{equation}

1627: where $m$ is the number of ``diagonal links'' connecting

1628: $(i,t)$ to $(i \pm 1,t+1)$ present in the path of length $T$.

1629: Each path $[m,T]$ of length $T$ with $m$ diagonal links contributes

1630: to the above sum with a term

1631: \begin{equation}

1632: M(m,T) = N(m,T)

1633: \left(\frac{\varepsilon}{2}\right)^m (1-\varepsilon)^{T-m}

1634: \end{equation}

1635: where $N(m,T)$ is the multiplicity associated to each path 

1636: $[m,T]$ (for more details see \cite{lyapcorr}).

1637: As can see from Eq.~(\ref{path}) the {\it finite time Lyapunov

1638: exponent} $\lambda_T$ will be given by

1639: \begin{equation}

1640: \lambda_T=

1641: \frac{1}{T}\ln\left|\frac{\delta x_{0}(T)}{d_0}\right| = \ln \beta + 

1642: \frac{1}{T}\ln\left|\sum_{m=0}^T M(m,T)\right|

1643: \quad .

1644: \label{lyt}

1645: \end{equation}

1646: the first contribution is the asymptotic one, while 

1647: the second one will vanish in the limit $T \to \infty$

1648: and it is the finite-time correction to evaluate.

1649: This second contribution can be numerically estimated

1650: by considering the finite time evolution of the Lyapunov

1651: eigenvector $\{W_i(t)\}$ associated to the maximal Lyapunov exponent,

1652: once it is initialized as $W_i(0) = d_0 \, \delta_{i,0}$.

1653: The evolution of $\{W_i(t)\}$ in the tangent space 

1654: is ruled by the following equation 

1655: \begin{equation}

1656: W_i(t+1)=\beta \left[(1-\varepsilon) W_i(t)+

1657: {\varepsilon \over 2} (W_{i-1}(t)+W_{i+1}(t))\right]\,.

1658: \label{eq:ac.tangent}

1659: \end{equation}

1660: In order to evaluate the finite time corrections,

1661: one should iterate at the same time 

1662: Eq.~(\ref{eq:ac.tangent}) (with $\beta$

1663: fixed to one) and the two replicas 

1664: required for computing $\lambda(\delta)$ (see Sect.~\ref{sec:2}).

1665: Then the estimation of the FSLE should modified in the following way:

1666: \begin{eqnarray} 

1667: \lambda(\delta)={1 \over \langle \tau(\delta_n,r)\rangle_e }

1668:  \left\langle \ln \left( { |\delta x_i(t+\tau(\delta_n,r))| \over |\delta

1669:  x_i(t)| }  \right) \right. \nonumber \\

1670: \left.  - 

1671: \ln \left( { |W_i(t+\tau(\delta_n,r))| \over |W_i(t)|}

1672: \right) \right\rangle_e \,.  

1673: \label{eq:FSLE_discrete_new}

1674: \end{eqnarray}

1675: The case of maps with non constant slope is computationally much

1676: heavier. Since for each different path the local multiplier

1677: $F^{'}(\tilde{x_i})$ will be different and they will depend

1678: on the particular trajectory under consideration \cite{P93,lyapcorr}.

1679: As a matter of fact we have observed that if the multipliers are

1680: equally distributed among positive and negative values, the finite

1681: time corrections essentially cancel out.

1682: \end{appendix}

1683: 

1684: 

1685: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

1686: \begin{thebibliography}{99}

1687: \bibitem{Shaw} R.~Shaw, {\it Z. Naturf.} {\bf 36a}, 80 (1981).

1688: 

1689: \bibitem{K86} K.~Kaneko,  Physica D {\bf 23},  436 (1986).

1690: 

1691: \bibitem {BR87} T.~Bohr and D.~Rand,  Physica D {\bf 52}, 532 (1987).

1692: 

1693: \bibitem {grass89} P.~Grassberger,  Physica Scripta {\bf 40}, 346 (1989).

1694: 

1695: \bibitem{PV94} 

1696: G.~Paladin and A.~Vulpiani, J. Phys. {\bf A27}, 4911 (1994).

1697: 

1698: \bibitem{DK87} R.J.~Deissler and K.~Kaneko, Phys. Lett. {\bf A119}, 397 (1987).

1699: 

1700: \bibitem {lepri} S.~Lepri, A.~Politi and A.~Torcini, 

1701: J. Stat. Phys., {\bf 82} (1996) 1429; 

1702: J. Stat. Phys., {\bf 88}  (1997) 31; 

1703: Chaos, {\bf 7} (1997) 701.

1704: 

1705: \bibitem {CK88} J.P.~Crutchfield  and  K.~Kaneko, 

1706: \prl {\bf 60}, 2715 (1988).

1707: 

1708: \bibitem{PLOK93} A.~Politi, R.~Livi, G.L.~Oppo and R.~Kapral,

1709: Europhys. Lett. {\bf 22}, 571 (1993).

1710: 

1711: \bibitem{TP94} A.~Politi and A.~Torcini, Europhys. Lett. {\bf 28}, 545 (1994).

1712: 

1713: \bibitem{TGP95} A.~Torcini, P.~Grassberger and A.~Politi,  

1714: J. Phys. {\bf A27}, 4533 (1995).

1715: 

1716: \bibitem{CFVV99} M.~Cencini, M.~Falcioni, D.~Vergni and A.~Vulpiani,

1717:  Physica D {\bf 130}, 58 (1999).

1718: 

1719: \bibitem{dressler92} U.~Dressler and J.D.~Farmer,  

1720: Physica D {\bf 59}, 365 (1992); T.J.~Taylor, Nonlinearity {\bf 6}, 369 (1993).

1721: 

1722: \bibitem{ABCPV96} E.~Aurell, G.~Boffetta, A.~Crisanti, G.~Paladin and

1723: A.~Vulpiani, \prl {\bf 77}, 1262 (1996);

1724: J. Phys. {\bf A30}, 1 (1997).

1725: 

1726: \bibitem{cml} I.~Waller and R.~Kapral, \pra {\bf 30} , 2047

1727: (1984); K. Kaneko, Prog. Theor. Phys. {\bf 72}, 980 (1984).

1728: 

1729: \bibitem {KPP} R.A.~Fisher, Ann. Eugenics {\bf 7}, 355

1730: (1937); A.N.~Kolmogorov, I.~Petrovsky and N.~Piscounov,

1731: Bull. Univ. Moscow, Ser. Int.  {\bf A1}, 1 (1937).

1732: 

1733: \bibitem{wim2} U.~Ebert and W.~van~Saarloos, Physica D {\bf 146}, 1 (2000).

1734: 

1735: \bibitem{cencrep} A recent review on the subject is

1736: G.~Boffetta, M.~Cencini, M.~Falcioni and A.~Vulpiani,

1737: unpublished (nlin.CD/0101029).

1738: 

1739: \bibitem {Kantz} T.~Letz and H.~Kantz, \pre {\bf 61},  2533 (2000).

1740: 

1741: 

1742: \bibitem{nota}Note that also for $\varepsilon \to 1$ deviations are observed. 

1743: This is due to the fact that for $\varepsilon=1$ the system reduces to

1744: two uncoupled subchains, therefore also for $\varepsilon \sim 1$

1745: the coupling among nearest neighbors is actually very small.

1746: 

1747: \bibitem {WB93} W.~van~de~Water and T.~Bohr, Chaos {\bf 3}, 747 (1993).

1748: 

1749: \bibitem{TP92} A.~Politi and A.~Torcini,  Chaos {\bf 2}, 293 (1992).

1750: 

1751: \bibitem{wim1} W.~van~Saarloos, \pra {\bf 37}, 211 (1988);

1752: {\it ibidem}, {\bf 39}, 6367 (1989).

1753: 

1754: \bibitem {isola} S.~Isola, A.~Politi, S.~Ruffo, and A.~Torcini,

1755: Phys. Lett. A, {\bf 143}, 365 (1990).

1756: 

1757: \bibitem{benettin} I.~Shimada and T.~Nagashima, Prog. Theor. Phys.

1758: {\bf 61}, 1605 (1979); G.~Benettin, L.~Galgani, A.~Giorgilli and

1759: J.M.~Strelcyn, Meccanica, March 15 and 21 (1980).

1760: 

1761: \bibitem{ginelli} F. Ginelli, R. Livi, and A. Politi,

1762: private communication.

1763: 

1764: \bibitem{ara} D.G.~Aronson and H.F.~Weinberger,

1765: Adv. Phys. {\bf 30}, 33 (1978).

1766: 

1767: \bibitem{armero} J. Armero, J. Casademunt, L. Ram{\'i}rez-Piscina,

1768: and J.M. Sancho, \pre {\bf 58}, 5494 (1998);

1769: A. Rocco, U. Ebert, and W. van Saarloos, \pre {\bf 62}, R13 (2000).

1770: 

1771: \bibitem{PP98} 

1772: A.S.~Pikovsky and J.~Kurths, \pre {\bf 49},  898 (1994).

1773: 

1774: \bibitem{noi} L.~Baroni, R.~Livi, and A.~Torcini, accepted for

1775: publication on \pre, it will appear in the volume of March 2001

1776: (nlin.CD/0003010). 

1777: 

1778: \bibitem{pecora} L.M.~Pecora and T.L.~Carroll,

1779: \prl {\bf 64} 821 (1990).   

1780: 

1781: \bibitem{ogy} E.~Ott, C.~Grebogi, and J.A.~Yorke,

1782: \prl {\bf 64}, 1196 (1990).

1783: 

1784: \bibitem{P93} A.~Pikovsky, Chaos {\bf 3}, 225 (1993).

1785: 

1786: \bibitem{Orszag} I.~Goldhirsh, P.L.~Sulem and S.A.~Orszag,  Physica D {\bf 27},

1787: 311 (1987).

1788: 

1789: \bibitem{lyapcorr} R.~Livi, S.~Ruffo, and A.~Politi,

1790: J. Phys. A {\bf 25}, 4813 (1992);

1791: A.~Torcini, R.~Livi, S.~Ruffo, and A.~Politi,

1792: \prl {\bf 78}, 1391 (1997).

1793: \end{thebibliography}

1794: \end{multicols}

1795: 

1796: \end{document}

1797: 

1798: 

1799: 

1800: