1: %
2: %
3: % It also requires running BibTeX. The commands are as follows:
4: %
5: %User: nlin.CD/0205029, Password: pgf9n (access still password
6: %restricted)
7: % 1) latex chaos.tex
8: % 2) bibtex chaos
9: % 3) latex chaos.tex
10: % 4) latex chaos.tex
11: %
12: \documentclass[twocolumn,showpacs,preprintnumbers,amsmath,amssymb]{revtex4}
13: %\documentclass[preprint,showpacs,preprintnumbers,amsmath,amssymb]{revtex4}
14:
15: % Some other (several out of many) possibilities
16: %\documentclass[preprint,aps]{revtex4}
17: %\documentclass[preprint,aps,draft]{revtex4}
18: %\documentclass[prb]{revtex4}% Physical Review B
19:
20: \usepackage{graphicx}% Include figure files
21: \usepackage{dcolumn}% Align table columns on decimal point
22: \usepackage{bm}% bold math
23:
24: %\nofiles
25:
26: \begin{document}
27:
28: \preprint{APS/123-QED}
29:
30: \title{Chaotic Synchronization Via Minimum Information Transmission}% Force line breaks with \\
31:
32: \author{A. S. Dmitriev}
33: % \altaffiliation[Also at ]{Physics Department, XYZ University.}%Lines break automatically or can be forced with \\
34: \author{G. A. Kassian}%
35: % \email{Second.Author@institution.edu}
36: \author{A. D. Khilinsky}
37: \affiliation{%
38: Institute of Radioengineering and Electronics Russian Academy of
39: Sciences\\
40: Mokhovaya 11-7, Moscow, 101999, Russia}%
41:
42: \author{M. Hasler}
43: % \homepage{http://www.Second.institution.edu/~Charlie.Author}
44: \affiliation{ Ecole Polytechnique Federale de Lausanne, Lausanne,
45: Switzerland}%
46:
47: \date{\today}% It is always \today, today,
48: % but any date may be explicitly specified
49:
50: \begin{abstract}
51: Chaotic synchronization is generally extremely sensitive to the
52: presence of noise and other inference in the channel. Is this
53: sensitivity a fundamental property of chaotic synchronization or
54: is it related to the choice of synchronization method and can it
55: be suppressed by a modification of the method? If the answer is
56: positive, then what are the relationships between the properties
57: of a dynamical system and the level of noise at which the
58: suppression of this sensitivity is still possible? What are
59: particular methods to achieve synchronization stable to the
60: presence of noise? In this paper we present the analysis of this
61: issue from the standpoint of information theory. The fundamental
62: reason for this sensitivity is the fact that the chaotic signal
63: contains information which requires a certain minimal threshold
64: signal-to-noise ratio for transmission. Only in this case
65: high-quality synchronization is achievable and only if the
66: required information is transmitted (coded) optimally. Otherwise
67: the threshold level can be much higher.
68: %Valid PACS numbers may be entered using the
69: %\verb+\pacs{#1}+ command.
70: \end{abstract}
71:
72: \pacs{05.45.-a,07.50.Qx,95.75Wx}% PACS, the Physics and Astronomy
73: % Classification Scheme.
74: %\keywords{Suggested keywords}%Use showkeys class option if keyword
75: %display desired
76: \maketitle
77:
78: \section{\label{sec:level1}Introduction}
79: The possibility to synchronize the dynamics of two identical
80: chaotic systems \cite{1,2,3,4} is determined by two factors (to
81: be precise the case of unidirectional impact of one system on the
82: other is assumed, i.e., the situation "drive system - response
83: system" is considered). First, the way the drive system affects
84: the response system. Significant progress has been achieved here
85: that enables synchronization not only of chaotic systems but also
86: of systems with hyperchaos \cite{5,6}. Second, it is determined by
87: the sensitivity of the chaotic synchronization to distortions of
88: the signal in the coupling channel, in particular, to the effects
89: of filtering and to the presence of noise \cite{30}. The attempts
90: to damp this sensitivity are being made both in the direction of
91: development of more robust schemes \cite{29,23} and in the
92: direction of correction (reconstruction) of a signal, distorted in
93: the coupling channel, into its original form at the input of the
94: responce system. For instance, in the case of a noisy signal, the
95: reconstruction is considered to be the cleaning of this signal
96: from noise \cite{7,8,9,10,27,31}. Hence, the problem is to
97: reconstruct with a high precision the output signal of the drive
98: system at the input of the response system. However, for this
99: purpose, transmission of the signal itself through the coupling
100: channel is not necessary. Instead, one can transmit the minimal
101: information about the signal necessary to reconstruct (recover)
102: the initial signal at the input of the response system, i.e. to
103: replace the coupling channel by the communication channel. From
104: the viewpoint of the information theory the second approach to
105: reconstruction of the signal at the input of the response system
106: can be preferable as it (i) requires less information to be
107: transmitted and (ii) this information can be presented (coded) so
108: that it is more robust to the disturbances in the communication
109: channel.
110:
111: This idea was discussed in the literature \cite{12,13} and was
112: tested in the case when the chaotic systems were one-dimensional
113: maps \cite{13}. Here, on the example of two-dimensional maps we
114: show that this approach can be used for synchronization of a
115: higher-dimensional chaotic maps (both hyperbolic and
116: non-hyperbolic).
117:
118: As an example of two-dimensional maps the hyperbolic Lozi map
119: \cite{14} is used in this paper. It is described by the equations
120: \begin{eqnarray}\label{eq1}
121: x_{n+1}&=&(\alpha-1)-\alpha|x_n|+y_n,\nonumber\\
122: y_{n+1}&=&\beta x_n,%??? сдвинуть влево!!!
123: \end{eqnarray}
124: where $\alpha$ and $\beta$ are parameters.
125:
126: Non-hyperbolic maps are represented here by the Henon map
127: \cite{15}
128: \begin{eqnarray}\label{eq1a}
129: x_{n+1}&=&1-\alpha x_n^2+y_n,\nonumber\\
130: y_{n+1}&=&\beta x_n,%??? сдвинуть влево!!!
131: \end{eqnarray}
132:
133:
134:
135: \section{Restrictions from Information Theory on Synchronization}
136:
137: Drive chaotic system is a specific source of information. With
138: respect to information transmission, the driven system can be
139: treated as a receiver of information, that, if synchronization
140: takes place, receives all the information defining the state of
141: the dynamical system, that is sent through the communication
142: channel and continuously adjusts its dynamics in accordance with
143: the information received.
144:
145: Characteristic of a chaotic source are a finite rate of
146: information production \cite{16,17} and continuous signal values.
147: A question arises, what the necessary carrying capacity of a
148: communication channel between the two systems is and how it
149: relates to the possibility of high-quality synchronization.
150:
151: According to Shannon's theorem \cite{18}, in order to transmit
152: without errors information produced by the source at rate $I$
153: through a channel of capacity $C$ one must have
154: \begin{equation}
155: \label{eq2} C>I.
156: \end{equation}
157: Hence, for synchronization of the drive and response systems in
158: the presence as well as in the absence of noise it is sufficient
159: to have a communication channel with a carrying capacity larger
160: than $I$. This fact has been noticed in \cite{12}. It suggests
161: that high-quality synchronization can be achieved, if external
162: perturbations do not exceed some critical value.
163:
164: The theorem also lays the basis for the quantitative analysis of
165: the capacity of a noisy channel. For a channel with additive
166: white gaussian noise, and bandwidth $W$ the capacity per unit of
167: time is
168: \begin{equation}
169: \label{eq3}
170: C=W\log_2\frac{P+N}{N}.
171: \end{equation}
172: where $P$ is the signal power and $N$ is the noise power.
173:
174: Taking the sampling period at $T=\frac{1}{2W}$ so that the whole
175: bandwidth is used, capacity per sample becomes
176:
177: \begin{equation}
178: C=\frac12 \log_2\frac{P+N}{N},\label{eq3a}
179: \end{equation}
180:
181: Eq.~(\ref{eq3}) implies that when appropriate coding is used,
182: information can be transmitted with the rate up to $W\cdot
183: \log_2\frac{P+N}N$ bit/sec while error probability can be
184: arbitrarily low.
185:
186: Equation (\ref{eq3}) also suggests that if the information
187: production rate $I$ and the channel capacity are known, the
188: maximum noise level can be calculated such that high-quality
189: synchronization is still feasible.
190:
191: As an example, if $I$ = 1 bit/sample, then, in theory, arbitrarily
192: precise synchronization can be achieved when the ratio of the mean
193: squared power of the chaotic signal $P=<x^2_n>$ to mean squared
194: power of gaussian noise $N=<w^2_n>$ is
195: $$P/N>3.$$
196: The "expense" of the possibility of high precision synchronization
197: in the presence of noise is a special organization of the
198: information transmitted through the channel and a certain time
199: delay of the processes in the response system with respect to the
200: drive system. Note also, that for the practical accomplishment of
201: high-quality synchronization in the presence of noise a certain
202: increase of the channel capacity is necessary compared to the
203: value given by relation (\ref{eq2}). Otherwise, the delay value
204: will tend to infinity.
205:
206:
207: \section{Hyperbolic Case. Lozi Map.}
208:
209: So, in order to be able to synchronize the response system with
210: the drive system we should transmit information at a certain rate.
211: Let us consider the information production of the Lozi map and the
212: amount of information that must be transmitted for
213: synchronization.
214:
215: First, let us recall that for 1-D maps
216: \begin{equation}
217: \label{eq7} x_{n+1}=f(x_n)
218: \end{equation}
219: the average information production $I$, measured in bits per
220: sample, can be expressed as the integral of (\ref{eq7}) weighted
221: with the invariant probability density $P(x)$
222: \begin{equation}\label{eq8}
223: \int_{0}^{1}P(x)\log_2\left|\frac{df}{dx}\right|dx.
224: \end{equation}
225:
226: Relation (\ref{eq8}) coincides with the definition of the Lyapunov
227: exponent $\lambda$ of the one-dimensional map with the only
228: difference that in the relation for the Lyapunov exponent the
229: natural logarithm is used instead of the base-2 logarithm. Hence,
230: the Lyapunov exponent for 1-D map can be treated as the
231: information production rate expressed in base-$e$ units and it is
232: equivalent to the Kolmogorov entropy $h$. In order to transform it
233: to bit/sample, we multiply $\lambda$ by $\log_2e$. Let us remark
234: that the Kolmogorov entropy $h$ and the information production
235: rate $R$ are generally the same not only in the case of 1-D maps,
236: but also in more general cases. However, it is common to express
237: $R$ in bits/sample and $h$ in nats/sample.
238:
239: The situation with information production in 2-D maps, e.g. in the
240: Lozi map, is more complicated. It is well known \cite{17} that for
241: differentiable maps the following inequality between Kolmogorov
242: entropy $h$ and positive Lyapunov exponents holds:
243: \begin{subequations}\label{eq10}
244: \begin{equation}
245: h\leq\lambda_1+\lambda_2,
246: \end{equation}
247: if both Lyapunov exponents are positive, or
248: \begin{equation}
249: h\leq\lambda_1,
250: \end{equation}
251: \end{subequations}
252: if only $\lambda_1$ is positive. Under some additional conditions
253: inequalities (\ref{eq10}) can turn into equalities. There are no
254: similar results for the non-differentiable Lozi map. However, one
255: can try to use another approach to estimate the information
256: production by the Lozi map.
257:
258: Let $N_n$ be the number of unstable periodic orbits of length $n$
259: in the chaotic attractor, and thus $L_n=N_n\cdot n$ the number of
260: cyclical points. Then the following limit is supposed to estimate
261: topological entropy
262: \begin{equation}\label{eq11}
263: K_0=\lim_{n\to\infty}\frac1n\ln L_n,
264: \end{equation}
265: Estimates for information production $I_n$, calculated with the
266: aid of (\ref{eq11}) are presented in Table~ {\ref{table1}} for
267: different $n$.
268: \begin{table}
269: \caption{\label{table1}Periodic orbits of Lozi map attractor.
270: $(\alpha=1.7, \beta=0.5)$; $n$ - the length of a cycle; $N_n$ -
271: the number of cycles with the length n; $L_n$ - the number of
272: periodic points; $I_n$ - information production estimate.}
273: \begin{tabular}{|c|c|c|c|c|c|c|}
274: \hline
275: % after \\: \hline or \cline{col1-col2} \cline{col3-col4} ...
276: $N$ & 1 & 4 & 8 & 12 & 16 & 17 \\
277: \hline
278: $N_n$ & 1 & 1 & 8 & 34 & 162 & 224 \\
279: \hline
280: $L_n$ & 1 & 4 & 64 & 408 & 2592 & 3808 \\
281: \hline
282: $I_n$ & 0 & 0.5 & 0.75 & 0.722 & 0.720 & 0.700 \\
283: \hline
284: \end{tabular}
285: \end{table}
286: Both methods of estimating information production (including the
287: method based on the positive Lyapunov exponents ) give similar
288: results.
289:
290: Fig.~\ref{fig3} shows the average information production for
291: different values $\alpha$ and $\beta$. For $\alpha=1.7$;
292: $\beta=0.5$ the average information production rate is
293: $I\approx0.7$ bits/sample.
294: \begin{figure}
295: \includegraphics[width=9cm]{fig_3}
296: \caption{Information production of Lozi map as a function of parameters $\alpha$ and $\beta$.}\label{fig3}
297: \end{figure}
298:
299: However, information production differs from iteration to
300: iteration. Fig.~\ref{fig4} shows the information production rate
301: probability density function. As can be seen in the figure the
302: rate never exceeds 1 bit/sample.
303: \begin{figure}
304: \includegraphics[width=9cm]{fig_4}
305: \caption{The probability distribution function of the rate of information production of the Lozi map.}\label{fig4}
306: \end{figure}
307:
308: Consequently, to ensure synchronization a channel with capacity
309: $C>I=0.7$ bits/sample can be used. However, it is desirable that
310: the channel capacity should be at least $C=I_{max}=1$ bit/sample
311: in order to take the maximum information production of each sample
312: into account. Below a symbolic sequence for Lozi map is studied as
313: an example of such a binary sequence.
314:
315: \textbf{Synchronization scheme}
316:
317: The proposed scheme is shown in
318: Fig.~\ref{fig5}. The signal $x_n$ produced by iteration of the
319: Lozi map is transformed by a 2-level quantizer (threshold device)
320: into a symbolic sequence $X_n=\text{sign}(x_n)$, i.e. into -1 if
321: $x_n<0$ and into +1 if $x_n\geq0$. Then this binary signal is sent
322: through the channel, where it is corrupted by noise:
323: $z_n=X_n+w_n$. The quantizer transforms it back to the binary
324: signal $\hat{X}_n=\text{sign}(z_n)$.
325:
326: \begin{figure}
327: \includegraphics[width=9cm]{fig_5}
328: \caption{Synchronization by means of symbolic sequence. $CS$-chaotic source. $CR$-chaotic receiver, $K_2$-quantizer.}
329: \label{fig5}
330: \end{figure}
331:
332: The $\pm1$ values of binary sequence correspond to symbols $L$ and
333: $R$ of the symbolic sequence. This symbolic sequence is used in
334: the reconstruction block to restore the chaotic sequence. Thus, in
335: the absence of noise, chaotic sequence reconstructed with a high
336: precision is fed to the input of the response system, or an
337: estimate of this chaotic sequence, if there is noise.
338:
339: First, a problem with the noisefree case will be analyzed.
340:
341: \textbf{Reconstruction of a Trajectory from Symbolic Sequence}
342:
343: According to the synchronization scheme, at the input of the
344: response system the incoming symbolic sequence must be used to
345: reconstruct the chaotic sequence generated by the chaotic source.
346:
347:
348: In the case of one-dimensional maps reconstruction of the chaotic
349: sequence based on the symbolic sequence can be accomplished by
350: iterating the map backward \cite{12,13}. Since the map is
351: stretching, being iterated backward it becomes contracting. Hence,
352: repeatedly applying the inverse map to the initial conditions
353: given by the elements of the symbolic sequence can give a good
354: estimate of the chaotic sequence.
355:
356: Consider a procedure of reconstruction of a chaotic sequence
357: given the symbolic sequence for the case of Lozi map.
358:
359: The Jacobian of (\ref{eq1}) is constant and equals $\beta$. This
360: means that the area of a small cell will be contracted by each
361: iteration of the Lozi map by a factor $\beta$. But the distance
362: between two close points increases in a certain direction, the
363: stretching direction, at a rate determined by the positive
364: Lyapunov exponent $\lambda_1$. Let two points in the plane be
365: separated along the stretching direction by the small distance
366: $\delta_0$ at time 0. Then the distance between the corresponding
367: trajectories at moment $m$ is approximately
368: \begin{equation}\label{eq16}
369: \delta x_m\propto\delta_0\cdot\exp(\lambda_1\cdot m),
370: \end{equation}
371: The situation is very similar to the case of the 1-D map, where
372: one could use this property (\ref{eq16}) of the map to improve our
373: knowledge of the value of the trajectory at present on the basis
374: of information about its values in the future. Namely, to improve
375: the precision of $x_n$ the map should be iterated in reverse time
376: beginning from $x_{n+k}$.
377:
378: But what happens if we iterate the Lozi map in reverse time? In
379: this case the area of a small cell will not decrease but increase.
380: The direction corresponding to the positive Lyapunov exponent will
381: be contracting but the direction corresponding to the negative
382: Lyapunov exponent will be stretching. The trajectories will
383: diverge again. The rate of this divergence will be determined by
384: the absolute value of the second Lyapunov exponent.
385:
386: The fact that two trajectories will diverge when the map is
387: iterated either forward or backward can be written as
388: \begin{equation}\label{eq17}
389: \delta x_m\propto\delta_0\cdot\exp(\lambda_1\cdot
390: m)+\varepsilon_0\cdot\exp(\lambda_2\cdot m),
391: \end{equation}
392: where $\varepsilon_0$ is the initial distance between the two
393: points in the direction of contraction. As follows from
394: (\ref{eq17}), in order to improve the knowledge about the position
395: of the trajectory at the present time, it is necessary to use
396: information both about the "future" and the "past" of the
397: trajectory.
398:
399: We rewrite equations (\ref{eq1}) for the Lozi map in such a way as
400: to make it contracting both from the future and from the past:
401: \begin{equation}\label{eq18}
402: |x_n|=\frac{\alpha-1}{\alpha}-\frac1\alpha
403: x_{n+1}+\frac\beta\alpha x_{n-1}.
404: \end{equation}
405:
406: Map (\ref{eq18}) is a contracting map because coefficients
407: $1/\alpha$ in front of $x_{n+1}$ and $\beta/\alpha$ in front of
408: $x_{n-1}$($\alpha=1.7, \beta=0.5$) as well as their sum are less
409: than $1$. Contracting features of map (\ref{eq18}) may be used to
410: improve an estimate of the variable value at the $n$-th moment
411: using information about the variable values at the ($n-1$)-th and
412: ($n+1$)-th moments. Note that map (\ref{eq18}) is not
413: single-valued but a double valued map.
414:
415: \textbf{Reconstruction algorithm.} So, let us have a symbolic
416: sequence $X_1,\ldots X_n,\ldots X_{n+k},\ldots X_N$, where $X_k\in
417: \{-1,1\}$: $N$ is the number of elements in the sequence. We
418: evaluate $|x_n|$ at the first step for all elements besides the
419: first and the last, i.e. for $n=2,\ldots,N-1$ in the following
420: way
421: \begin{equation}
422: x_n^1=X_n|x_n^1|=X_n\left[\frac{\alpha-1}{\alpha}-\frac1\alpha
423: X_{n+1}+\frac\beta\alpha X_{n-1}\right].
424: \end{equation}
425:
426: Elements of the symbolic sequence are used at this stage as
427: initial approximation for the estimate. At the $i$-th step of the
428: procedure the estimates obtained at the previous ($i$-1)-th step
429: are used.
430: \begin{equation}
431: x_n^i=X_n|x_n^i|=X_n\left[\frac{\alpha-1}{\alpha}-\frac1\alpha
432: x_{n+1}^{i-1}+\frac\beta\alpha x_{n-1}^{i-1}\right].
433: \end{equation}
434: where $n=2,\ldots,N-1$. We finish the process when the difference
435: between estimates of ${x^i}$ on $i$-th and $i+1$-th iterations is
436: less then some small $\varepsilon$.
437:
438: The results of chaotic sequence reconstruction using the above
439: algorithm are presented in Fig.~\ref{fig8}. The four curves in
440: Fig.~\ref{fig8} correspond to the relative precision (in dB)
441: obtained by taking $N=10,40,70\text{ and }100$ points of the
442: symbolic sequence. Each of these curves looks as two sides of a
443: triangle.
444:
445: \begin{figure}
446: \includegraphics[width=9cm]{fig_8a}
447: \caption{Reconstruction of chaotic sequence by means of symbolic
448: sequence. $n$ --- reconstructed point index, $\Delta(n)$
449: --- reconstruction error, $N=10,40,70,100$ --- the number of
450: elements of the symbolic sequence that are used for
451: reconstruction.}\label{fig8}
452: \end{figure}
453:
454: Consider the longest curve, that shows the results of the
455: reconstruction of a fragment 100 points long. It can be seen that
456: the first and the last points are not reconstructed at all, they
457: are just the values of the symbolic sequence. The precision at
458: these points corresponds to the precision of the initial
459: approximation. The best result of reconstruction with precision
460: $\sim10^{-15}$ corresponds to the region of 30-th point. The ratio
461: of the number of points to the left and to the right from this
462: point is 3:7 or 1:2.3 that is approximately the ratio of the
463: coefficients $1/\alpha$ and $\beta/\alpha$ in (\ref{eq18}):
464: $\beta/\alpha:1/\alpha=1:2$
465:
466: Hence, to reconstruct the sequence of chaotic signals $x_1,\ldots
467: x_n$ with a given precision , one first has to calculate the
468: length of the fragment of the symbolic sequence to be used for
469: reconstruction of one point of the chaotic sequence. For
470: instance, the length of the fragment equals 100 when the
471: reconstruction precision is $10^{-15}$. Then, the left-edge
472: fragment of the symbolic sequence $X_1,\ldots X_{100}$ is used to
473: reconstruct the 30th point of the chaotic sequence $x_{30}$. Then
474: the fragment $X_2,\ldots X_{101}$ is used to reconstruct the point
475: $x_{31}$ of the chaotic sequence and so on. As for reconstruction
476: of any given point of the chaotic sequence one needs 29 preceding
477: and 70 following points of the corresponding symbolic sequence.
478: Reconstruction can be accomplished with this precision only for
479: those points that are non less then 29 positions distant from the
480: left edge and 70 positions distant from the right one.
481:
482: \section{Non-hyperbolic Henon Map}
483:
484: In the case of non-hyperbolic maps, e.g. the Henon map, the
485: approach based on the transmission of the symbolic sequence with
486: subsequent reconstruction of the chaotic sequence fails to be
487: directly applicable. However, an attempt to apply this approach
488: allows to understand why the results are insufficient and helps to
489: develop a suitable method.
490:
491: Assume first, as in the case of the Lozi map, that a sequence
492: $X_n=\text{sign}(x_n)$ is transmitted through the channel. Note
493: that the sequence $x_1,...x_N,...$ is not the symbolic sequence
494: for this case.
495:
496: Henon map can be rewritten as a one-dimensional map of the second
497: order
498: \begin{equation}\label{eqhenon5a}
499: x_{n+1}=1-\alpha x_n^2 + \beta x_{n-1}.
500: \end{equation}
501: When iterated forward, this map is stretching at almost all points
502: of the trajectory, that is, the error in the value of $x_{n+1}$
503: usually is greater than that of $x_n$. On average, the distance
504: between two close points increases exponentially with time.
505:
506: Rewrite (\ref{eqhenon5a}) in the form
507: \begin{equation}\label{eqhenon5}
508: x_n=\frac1{\sqrt{\alpha}}\sqrt{1-x_{n+1}+\beta x_{n-1}}.
509: \end{equation}
510:
511:
512: Map (\ref{eqhenon5}) is contracting everywhere but in the region
513: of small $x_n$: $|x_n|<\varepsilon$, where $\varepsilon \sim 0.1$.
514: Outside this region the error of the estimated value of $x_n$
515: according to (\ref{eqhenon5}) is smaller than the error of the
516: estimates of $x_{n-1}$ and $x_{n+1}$ that are used to get an
517: estimate of $x_n$. For small $|x_n|$ this is not true, hence, the
518: algorithm analogous to the one used for the Lozi map can not be
519: used to reconstruct the chaotic sequence with high precision. A
520: typical result of the application of this algorithm to the Henon
521: map is shown in Fig.~\ref{fighenon1} (line 2). One can see that
522: there are intervals with sharp reduction of reconstruction
523: precision and the intervals where the precision of reconstruction
524: is relatively high. Deterioration of the precision always occurs
525: because of the lack of contracting properties of map
526: (\ref{eqhenon5}) for small $|x_n|$.
527:
528: \begin{figure}
529: \includegraphics[width=9cm]{fig_12}
530: \caption{Reconstruction of chaotic sequence by means of symbolic
531: sequence. $l$ --- the number of reconstructed point, $k(l)$
532: --- reconstruction error, the symbolic sequence is reconstructed from 100 sample
533: points. 1 --- application of the algorithm to Lozi map, 2
534: --- application of the algorithm to the Henon map}\label{fighenon1}
535: \end{figure}
536:
537: Assume that at the reconstruction block the estimates of a
538: chaotic sequence at the moments $n_0$ and $n_0+1$ are known with
539: the desired precision and there is also a copy of the Henon map.
540: Then this information is sufficient to get the values of the
541: elements of the binary sequence $X_n$ for several points with the
542: indices $n>n_0$. For this one just have to iterate the Henon map
543: with initial conditions $(x_{n_0},x_{n_0-1})$ and construct a
544: "forecast". This "forecast" allows to determine the moments of
545: time when the trajectory goes through a region of small absolute
546: values.
547:
548: Since initial conditions usually are known only with a finite
549: precision, the correct values for the elements of the symbolic
550: sequence can be obtained only for a relatively short part of the
551: trajectory when the error of the "forecast" is much less then the
552: size of the attractor.
553:
554: Above, in the description of the algorithm for Lozi map it has
555: been noted that for a reconstruction of a chaotic sequence
556: $x_1,\ldots x_N$ with a high quality precesion it is sufficient
557: to know the symbolic sequence $X_n$ for $n\in[2,N-1]$ and
558: $x_1,x_N$ with the desired precision. The same conditions hold for
559: the case of Henon map given that in the range $n\in[2,N-1]$ there
560: are no $x_n$ with absolute values close to zero.
561:
562: Consequently, we arrive at the following algorithm of the
563: information transmission from the drive to the response system.
564: Let $\nu_i$ be the indices of the points where the trajectory
565: comes close to zero, $\mu_i=[\nu_i,\nu_{i+1}]$ - the segment of a
566: trajectory limited by two subsequent points close to zero (see
567: Fig.~\ref{fighenon2}). Assume that at the input of the driven
568: system there is an estimate of the chaotic sequence for all points
569: of $\mu_i$. The two last points from $\mu_i$ are taken as the
570: initial conditions for the forecast of points from $\mu_{i+1}$.
571: Note also that the length of $\mu_{i+1}$ is initially unknown at
572: the response system side. Since the initial conditions for a
573: forecasting trajectory are known with some error, a disturbed and
574: undisturbed trajectories diverge exponentially and the precision
575: of the forecast also grows exponentially with the distance from
576: $\nu_{i+1}$. Nevertheless, if the length of $\mu_{i+1}$ is not
577: very large, one can: (1) predict the index of point $\nu_{i+2}$
578: where the chaotic trajectory is close to zero and (2) construct
579: the binary sequence $X_n$ for all points from $\mu_{i+1}$.
580:
581: \begin{figure}
582: \includegraphics[width=9cm]{fig_13}
583: \caption{Forecast of chaotic trajectory and transmission of
584: necessary information for forecast amendment.}\label{fighenon2}
585: \end{figure}
586:
587: At this stage the only single datum needed for chaotic sequence
588: reconstruction is the value at the point $\nu_{i+2}$. This value
589: is transmitted through the channel with the necessary precision.
590: After that $\mu_{i+1}$ is reconstructed with the aid of the
591: algorithm described for the Lozi map. Hence, to reconstruct the
592: chaotic sequence at all points of $\mu_{i+1}$ one only has to
593: transmit the value of a single point (with the index $\nu_{i+2}$).
594:
595: A subtle modification of this algorithm can substantially reduce
596: the amount of information transmitted. The case is that the
597: forecast for the points from $\mu_{i+1}$ allows to get some
598: estimate for the point $x_{\nu_{i+2}}$, and so, there is no need
599: to transmit the whole value of this point. One just needs to
600: transmit the bits that can be distorted due to exponential
601: divergence of the trajectories. Since the rate of divergence is
602: solely determined by the first Lyapunov exponent, the amount of
603: information to be transmitted through the communication channel is
604: also determined only by this property of the chaotic map and is
605: independent of the desired reconstruction precision of the chaotic
606: sequence.
607:
608: A numerical simulation of the proposed algorithm ensures high
609: precision of synchronization with a volume of transmitted
610: information $\sim1$ bits/sample, that is close to the rate of
611: information production by Henon map ($\sim~0.6$ bits/sample). Note
612: that in \cite{24} the method of synchronization with dynamical
613: coupling was analyzed that allows to reduce the amount of
614: necessary information to be transmitted for synchronization by a
615: factor of 5-6 as compared to conventional procedure. The approach
616: considered above provides a reduction of information flow by a
617: factor of 20-30 in comparison with conventional procedure.
618:
619: \section{Synchronization in the Presence of Noise}
620:
621: It has been shown that in order to achieve synchronization of
622: hyperbolic as well as non-hyperbolic maps it is sufficient to
623: transmit information contained in the chaotic sequence and then
624: utilize this information to reconstruct the chaotic sequence. The
625: precision of reconstruction depends on the length of the sequence
626: and on the number of elements of the symbolic sequence that are
627: directly used to reconstruct the chaotic sample. The
628: reconstructed sequence of samples ensures high-quality
629: synchronization of drive and response systems.
630:
631: If binary information is transmitted via a noisy channel errors
632: may appear in the binary sequence. These errors may lead to
633: incorrect reconstruction of the chaotic signal and thus reduce the
634: synchronization precision.
635:
636: Consider the results of numerical simulation of the precision of
637: synchronization of chaotic signal reconstruction after
638: transmission through a noisy channel. Note that errors can be made
639: arbitrarily small by increasing the channel capacity (e.g., by
640: means of using error correcting codes). Hence, synchronization
641: through channels with a capacity close to the minimum is of
642: greater concern.
643:
644: Fig.~\ref{fig10} shows graphs that illustrate the quality of
645: synchronization when information is transmitted through a channel
646: with gaussian noise.
647:
648: It can be seen that for signal to noise ratios up to about 7.5 dB,
649: the synchronization error is actually given by the noise and it
650: cannot be improved by increasing the number of points used for the
651: reconstruction of a sample. Above this threshold, the
652: synchronization error is decreased by about 4.2 dB/points used in
653: the reconstruction, which is in good agreement with the rate of
654: information production $I$ of 0.7 bit/sample.
655:
656: Similar results were obtained for the synchronization of Henon
657: maps.
658:
659: \begin{figure}
660: \includegraphics[width=9cm]{fig_10a}
661: \caption{Synchronization quality for the case of AWGN channel.
662: $K=(SNR_{in}-SNR_{out})$ --- the criterion of the synchronization
663: quality.}\label{fig10}
664: \end{figure}
665:
666:
667:
668: \section{Conclusion}
669: In this paper a synchronization scheme for two-dimensional maps
670: is considered that is based on transmission of information about
671: the state of drive system by unidirectional coupling. It is shown
672: that the necessary amount of information that must be delivered to
673: the response system is determined by the rate of information
674: production by the dynamical system.
675:
676: When this threshold is increased, however, arbitrarily precise
677: synchronization can be achieved. The only price to pay for that is
678: the complexity of the reconstruction algorithm and the delay of
679: the reconstructed signal. This is contrary to the common belief
680: that the synchronization error necessarily monotonically increases
681: with the noise.
682:
683: The schemes of synchronization of two-dimensional maps are
684: presented, that are based on reconstruction of the chaotic
685: sequence. Their stability with respect to noise almost coincides
686: with the theoretical estimate of the ultimate limit.
687:
688:
689: \begin{acknowledgments}
690: This work has been supported by the Swiss National Science
691: Foundation, SCOPES project Nr. 75SUPJ062310.
692: \end{acknowledgments}
693:
694: \newpage %Just because of unusual number of tables stacked at end
695: \bibliography{pre}% Produces the bibliography via BibTeX.
696:
697: \end{document}
698: