1: %\documentclass[aps,chaos,twocolumn,showpacs,preprintnumbers,amsmath,amssymb,amsfonts]{revtex4}
2: \documentclass[aps,chaos,preprint,preprintnumbers,amsmath,amssymb]{revtex4}
3:
4: \usepackage{graphicx}
5: \usepackage{dcolumn}
6: \usepackage{bm}
7: \usepackage{epsfig}
8:
9: \begin{document}
10:
11: \title{Symbolic Synchronization and the Detection of Global Properties of
12: Coupled Dynamics from Local Information}
13: \author{Sarika Jalan} \email{sjalan@mis.mpg.de}
14: \author {J\"urgen Jost} \email{jjost@mis.mpg.de}
15: \author{Fatihcan M. Atay} \email{atay@member.ams.org}
16: \affiliation{Max Planck Institute for Mathematics in the Sciences,
17: 04103 Leipzig, Germany}
18:
19: \date{\today}
20:
21: \begin{abstract}We study coupled dynamics on networks using symbolic
22: dynamics. The symbolic dynamics is defined by dividing the state
23: space into a small number of regions (typically 2), and considering
24: the relative frequencies of the transitions between those
25: regions. It turns out that the global qualitative
26: properties of the coupled dynamics can be classified
27: into three different phases based on the synchronization of
28: the variables and the
29: homogeneity of the symbolic dynamics. Of particular interest is
30: the {\it homogeneous unsynchronized
31: phase} where the coupled dynamics is in a chaotic unsynchronized
32: state, but exhibits (almost)
33: identical symbolic dynamics at all the nodes in the network. We refer
34: to this dynamical behaviour as {\it symbolic synchronization}.
35: In this phase, the local symbolic dynamics of any arbitrarily selected node
36: reflects global properties of the coupled dynamics, such as
37: qualitative behaviour of the largest
38: Lyapunov exponent and phase synchronization.
39: This phase depends mainly on the
40: network architecture, and only to a smaller extent on the local chaotic
41: dynamical
42: function. We present results for two model dynamics, iterations of the
43: one-dimensional logistic
44: map and the two-dimensional H\'enon map, as local
45: dynamical function.
46: \end{abstract}
47: \maketitle
48: %\pacs{05.45.Ra,05.45.Xt,84.35.Xt,89.75.Fb}
49: {\bf{Nonlinear dynamical elements
50: interacting with each other
51: can lead to synchronization or other types of coherent behaviour at
52: the system scale.
53: Coupled map models are one of the most widely accepted models to understand
54: these behaviours in systems from many diverse fields such as
55: physics, biology, ecology etc. Their important feature is that the
56: individual elements can already exhibit some complex behaviour, for
57: example chaotic dynamics. The question then is how to detect
58: coordination at larger scales beyond the simplest one,
59: synchronization. An important tool in the analysis of dynamical
60: systems are symbolic dynamics.
61: We develop a new scheme of
62: symbolic dynamics that is based on the
63: special partitions of the phase space which prevent the occurrence of
64: certain symbol sequences related to the characteristics of the dynamics.
65: In particular,
66: we report a new behaviour of coupled dynamics, which we refer to as
67: symbolic synchronization, i.e. synchronization of the nodes
68: at the coarse grained level,
69: whereas microscopically all elements behave differently.
70: Through the framework of this symbolic dynamics, we detect various global
71: properties of coupled dynamics on networks by using a scalar time series of any
72: randomly selected node. A decisive advantage of our method is that
73: the global properties are inferred by using a short time series, hence
74: the method is computationally fast,
75: does not depend on the size of the network, and is reasonably
76: robust against external
77: noise.}}
78:
79: \section{Introduction}
80: In order to gain insights into the behaviour of real systems from many
81: diverse fields
82: ranging from chemical, physical and biological systems,
83: it is useful to identify model systems that on the one hand exhibit
84: essential dynamical features of those real world systems, but on the
85: other hand suppress individual details that are not really relevant
86: for the qualitative behaviour \cite{rev-CM}.
87: Coupled map models have emerged as one
88: such paradigm \cite{rev-kaneko,SJ-REA-insa}. Here, we have a system of elements with
89: identical local
90: dynamical functions. These elements are arranged in a network that
91: expresses their couplings so that the local dynamical iteration
92: depends not only on the own state of an element, but also on the ones
93: of its neighbour in the network. The inhomogeneities in the network
94: then translate into qualitative features of the global dynamics. While
95: such coupled map models already present an important simplification in
96: view of the complexities of real world dynamics, their behaviour can
97: nevertheless be sufficiently complicated and difficult to
98: analyze. Thus, it is important to identify parameters that allow for a
99: facile and robust detection of different qualitative states. One
100: needs a coarse grained description to analyze the complicated
101: time evolution of a chaotic dynamical system \cite{sym-book,sym-book2}.
102: In doing so one inevitably
103: simplifies the dynamics a lot and some of the information are lost,
104: but the aim is that
105: important invariants and robust properties of the dynamical systems
106: can be kept. Such a coarse graining means that we divide the possible
107: dynamical states of the system into finitely many discrete classes and
108: investigate the derived {\it symbolic dynamics} \cite{sym-book}. \\
109: Coarse graining of the time evolution of lower dimensional systems
110: have been studied at various levels \cite{sym-book2,sym-rev},
111: but symbolic dynamical studies of higher dimensional spatio-temporal chaotic
112: systems are rare and so far limited to coupled map lattices \cite{sym-CD}.
113: In the present paper, continuing the approach developed in \cite{pap1},
114: we study symbolic dynamics of coupled map networks and demonstrate that they
115: can serve the above purpose well. Thereby, we attempt to provide a general
116: framework for coupled dynamics on networks.
117:
118: In \cite{pap1}, we have studied
119: symbolic dynamics of a discrete dynamical iteration of a function
120: based on non-generating partitions. We have shown two important uses of
121: this symbolic dynamics, namely,
122: distinguishing a chaotic iteration and a random iteration with the
123: same density distribution (this is related to the earlier
124: work on transition entropy \cite{kurths-TS} and \cite{PE1} on permutation
125: entropy), and detecting
126: synchronization in coupled dynamics on large networks.
127: In this paper we extend these studies and propose a
128: general method to investigate collective behaviour
129: of coupled systems.
130: Besides the applications mentioned in \cite{pap1}, we show further
131: applications of symbolic dynamics for detecting
132: various dynamical properties of
133: coupled dynamics in relation to structural parameters of the
134: underlying network.
135:
136: We take coupled map network models as generic models to apply our
137: method. Chaotic coupled maps show rich spatio-temporal behaviour.
138: One phenomenon that has received a lot of attention is
139: synchronization, where different random or chaotic
140: units of a system behave
141: in unison
142: \cite{book-syn,phyreport-syn}. (For a selection of recent references,
143: see also \cite{syn}.)
144: One application of our symbolic dynamics is the detection of
145: synchronization in large complex systems.
146: Traditional methods for the detection of synchronization in coupled
147: systems focus on the correlation analysis of the time series
148: measured at pairs of the nodes.
149: In \cite{pap1} we have introduced a method based on symbolic dynamics, which
150: uses a short time series of
151: any single arbitrarily selected node to detect global synchrony of all the
152: units.
153: In this paper we show that the same type of symbolic dynamics can be used
154: as a measure of phase synchronization, a phenomenon
155: shown by coupled dynamics on networks \cite{PS1}.
156:
157: In more detail, by our method we classify the coupled dynamics into
158: different states, depending
159: upon the synchronization of the nodes and the {\it homogeneity of the
160: symbolic dynamics} of the nodes. The most interesting phase is
161: the unsynchronized homogeneous phase, which refers to the state
162: where the local chaotic dynamics of the individual nodes are different,
163: but the derived symbolic dynamics of all
164: the nodes are similar. We refer to this state as {\it symbolic
165: synchronization} of the nodes. Recently, the unsynchronized region of
166: coupled maps has been shown to have a fractal stationary density function
167: \cite{jost-kiran}.
168: We show that, in this phase,
169: the transition probabilities
170: of any randomly selected node reflect the qualitative information of the
171: largest Lyapunov exponent ($\lambda_l$) and the phase synchronization
172: of the coupled dynamics.
173: For the calculation of the largest Lyapunov exponent we utilize only a
174: short
175: time series, whereas traditional methods to calculate the largest Lyapunov
176: exponent
177: from a scalar time series require rather long time series and also
178: involve various computational complications \cite{lya-rev}. We point
179: out, however,
180: that -- as to be expected from such a simplistic reduction -- our symbolic
181: dynamics gives only the qualitative behaviour of the
182: Lyapunov exponent $\lambda_l$, but of course not its exact value.
183:
184: The paper is organized as follows.
185: After an introductory section we introduce
186: the definitions of the different phases based on the symbolic
187: dynamical properties
188: in Section II. In Section III, we then present numerical
189: examples illustrating the behaviour of nodes in different phases. Mostly
190: we present results for homogeneous synchronized phase which is of main interest.
191: The key point is that the derived symbolic dynamics allows for the
192: detection of the global properties
193: of coupled dynamics from local measurements, that is, we can infer global
194: properties of the dynamical network by considering the symbolic dynamics at
195: a single node. Section IV distinguishes different
196: dynamical phases based on the network parameters. Section V describes the
197: relation between symbolic dynamics
198: and phase synchronization. Section VI discusses the coupled
199: H\'enon map.
200:
201: \section{Model and definition of symbolic dynamics}
202: We consider the dynamical system defined by the iteration rule
203: \begin{equation}
204: x(t+1)=f(x(t))
205: \label{iteration}
206: \end{equation}
207: where $t \in {\mathbb Z}$ is the discrete time and $f:S\rightarrow S$ is
208: a map on a subset $S$ of ${\mathbb R}^{n}$.
209: Let $\{S_{i}:i=1,\ldots ,m\}$ be a partition
210: of $S$, i.e., a collection of mutually disjoint and nonempty subsets satisfying $\cup
211: _{i=1}^{m}S_{i}=S$.
212: The symbolic dynamics corresponding to (\ref{iteration})
213: is the sequence of symbols $\{{\dots ,s_{t-1},s_{t},s_{t+1},\dots \}}$ where
214: $s_{t}=i$ if $x(t)\in S_{i}$.
215: For the purposes of this paper, a useful partition is defined as follows.
216: Let $x=(x_{1},\dots,x_{n}) \in \mathbb{R}^n$, and
217: suppose the scalar $x_{n^\prime}$, $1\leq n^\prime\leq n$, is
218: available for measurement. For a given threshold value $x^{\ast} \in \mathbb{R}$,
219: define the sets
220: \begin{equation}%
221: \begin{array}
222: [c]{c}%
223: S_{1}=\{x\in S:x_{n^\prime}<x^{\ast}\}\\
224: S_{2}=\{x\in S:x_{n^\prime}\geq x^{\ast}\}
225: \end{array}
226: \label{part}%
227: \end{equation}
228: The value $x_{n^\prime}$ can be chosen to make the sets
229: $S_{1},S_{2}$ nonempty, in which case they form a non-trivial partition of $S$.
230: For this special partition, we use the two-symbol dynamics generated by
231: \begin{equation}
232: s_t= \left\{ \begin{array}{c}
233: \alpha \text{ if } x_n^{\prime}(t) < x^{\ast} \\
234: \beta \text{ if } x_n^{\prime}(t) \ge x^{\ast}.
235: \end{array} \right .
236: \label{sym}
237: \end{equation}
238: The symbolic dynamics depends only on the measurements $x_{n^{\prime}} $,
239: yielding a sequence of symbols determined by whether a measured value
240: exceeds the threshold $x^{\ast}$ or not.
241: Essentially any choice of the threshold $x^{\ast}$ will yield a
242: non-generating partition. For practical calculations using short time series,
243: however, a judicious choice of $x^{\ast}$ becomes important. We will address
244: this issue later in the paper (see section V).
245:
246: We take the well known coupled map model \cite{CM},
247: \begin{equation}
248: x_i(t+1) = f(x_i(t)) + \frac{\varepsilon}{k_i} \sum_j w_{ij} g(x_j(t),x_i(t))
249: \label{coup-dyn}
250: \end{equation}
251: where $x_{i}(t)$ is the dynamical variable of the $i$-th node $(1 \le i \le N)$
252: at time $t$, $w$ is the adjacency matrix with elements $w_{ij}$
253: taking values between 0 and 1 depending upon the weight of the connection
254: between $i$ and $j$, and $k_i$ is some normalization factor depending
255: on the node $i$, for example its degree. The function $f(x)$ defines the local
256: nonlinear map and
257: the function $g(x)$ defines the nature of the coupling between the nodes.
258: In the first three sections, we present the results for the local
259: dynamics given by the logistic map $f(x) = \mu x (1-x)$ and
260: coupling function $g(x_j(t),x_i(t))=f(x_j(t))-f(x_i(t))$. We take $\mu=4$,
261: for which logistic map exhibits chaotic behaviour with Lyapunov
262: exponent $\ln(2)$. The weight
263: $w_{ij}$ is simply one when nodes $i$ and $j$ are neighbours in the
264: undirected network, and 0 otherwise. In particular, the matrix $w$ is
265: symmetric; $k_i$ then is the degree of node $i$, as already
266: indicated.
267:
268: We evolve Equation~(\ref{coup-dyn}) starting from random initial conditions
269: and estimate the transition probabilities using time series of length
270: $\tau=1000$. Note that the length of the time series is independent of the
271: size of the network. We calculate the transition probability $P(i, j)$
272: by the ratio $\sum_t n(s_t = i, s_{t+1} = j)/\sum_t n(s_t = i)$, where
273: $n$ is a count of the number of times of occurrence \cite{pap1}.
274:
275: \section{Different states of the coupled dynamics}
276:
277: We classify the coupled dynamics in three different categories
278: based on the dynamical behaviour, and we show that how one category
279: differs from another based on some of the parameters of underlying
280: network:
281: \begin{enumerate}
282: \item Unsynchronized or phase synchronized {\it non-homogeneous}
283: behaviour : {\it phase one},
284:
285: \item Partially synchronized or phase synchronized
286: {\it homogeneous behaviour} : {\it phase two},
287: and
288:
289: \item Fully synchronized {\it homogeneous behaviour} : {\it phase three}.
290: \end{enumerate}
291: Here, synchronization refers to the variables at different nodes having the same
292: value
293: $x_i(t) = x_j(t)$ for all $i,j$. The network is globally synchronized when at
294: each time $t$,
295: all nodes have the same value. Partial synchronization means that some of the
296: nodes form a cluster inside which all the
297: nodes are synchronized while they are not synchronized with the nodes in the
298: different clusters. We note, however, that this state usually does not
299: occur in our
300: coupled dynamics because the phase differences between the various clusters
301: will interfere with the internal synchronizations. The following behaviour,
302: however, does robustly occur in suitable parameter regions. A pair of nodes
303: is called phase synchronized \cite{PS1} when they have their minima (maxima)
304: matching for all $ t > t_0$, that is, when one of them attains a minimum
305: then so does the other. The concrete values may and can be different. In a
306: phase synchronized cluster all nodes are phase synchronized.
307:
308: Complete synchronization is indicated
309: by the variance $\sigma^2$ of the variables over the network tending to zero,
310: where
311: $$\sigma^2 = \left\langle \frac{1}{N-1} \sum_{i}[x_{i}
312: (t) - \bar{x}(t)]^{2}\right\rangle_{t},$$
313: $\bar{x}(t)=\frac{1}{N}
314: \sum_{i}x_{i}(t)$ denotes an average over the nodes of the network, and $
315: \left\langle \dots \right\rangle _{t}$ denotes an average over time.
316: We define {\it homogeneous} and {\it non-homogeneous} behaviour based on
317: the symbolic
318: dynamics of the individual nodes. If all the nodes have the
319: same transition
320: probabilities, then
321: we say that the coupled dynamics is homogeneous; otherwise it is
322: non-homogeneous.
323: Homogeneity is indicated by the variance of the transition
324: probability over the network being zero,
325: \begin{equation}
326: \varsigma^2 =
327: \left\langle \frac{1}{N-1}\sum_{k=1}^N[P_k(\alpha,\alpha)-
328: \overline{P(\alpha,\alpha)}]^2 \right\rangle\\
329: \label{equ-homo}
330: \end{equation}
331: where $\overline{P(\alpha,\alpha)}
332: = \sum_{k=1}^N P_k(\alpha, \alpha)$ denotes an average over the
333: nodes of the network.
334:
335: \section{Homogeneous phases and coupled dynamics on network}
336: \subsection{Homogeneous phase and network properties}
337: We shall now connect the {\it symbolic homogeneity} with
338: network properties.
339: When all the nodes in a network have the same degree and the network is
340: homogeneously connected, i.e. if the network is completely symmetric,
341: like a nearest neighbour coupled network with periodic boundary conditions,
342: then, unless the dynamics breaks the symmetry, each node should have
343: {\it qualitatively the same symbolic
344: dynamics}, i.e. all transition probabilities for all the nodes being
345: equal. In fact, one might then even expect stability of the
346: synchronized state, but that in general is not true for all coupling strengths.
347: Homogeneous symbolic dynamics need not correspond to synchronization,
348: though it may correspond to phase synchronization.
349: For random networks, homogeneity of the symbolic dynamics
350: depends on the number of connections in the network.
351: Note that depending upon the coupling strengths, for certain network
352: architecture one may get all the three
353: phases, including the homogeneous synchronized phase. Using the
354: master stability function which takes local dynamics as well
355: as network architecture into account, one can deduce the coupling strength
356: region for which the coupled dynamics would be synchronized \cite{MSF}.
357: To relate different
358: dynamical states with the network parameters, we consider the quantity
359: $r=\frac{N_c}{N(N-1)/2} \sim N_c/N^2$. This ratio compares
360: the number of
361: connections $N_c$ in the network
362: with the number of possible connections $N(N-1)/2$,
363: We use $r$ as an indicator to roughly distinguish the three phases.
364: For $r$ being close to one (number of
365: connections $N_c$ of order $N^2$), we get a fully synchronized
366: state for appropriate coupling strengths $\varepsilon$. Then
367: the transition probabilities of all nodes are obviously equal (phase 3). For
368: $N_c \sim N $, we get phase two, i.e. the nodes are partially
369: synchronized or partially phase synchronized, but the symbolic dynamics
370: of the nodes are identical. Note here we are only roughly
371: relating $N_c$ and phase, later we will provide a more accurate
372: relation between the number of connections and the phases.
373:
374: \subsection{Symbolic synchronized phase and global properties of coupled %
375: dynamics}
376: We concentrate on the phase where the nodes are not synchronized though
377: their symbolic dynamics are identical. This is the most interesting phase
378: as the complexity of the coupled dynamics can be understood
379: by observing the symbol sequence of any arbitrarily selected node.
380: Fig.~\ref{sym-scale} is plotted for the logistic map as the local map and a
381: scale-free network
382: \footnote{We have generated that scale-free network by the standard %
383: preferential attachment scheme \cite{BA}, but one should note that large %
384: classes of scale-free networks may exhibit a qualitatively different behaviour %
385: as regards other crucial network parameters besides the degree sequence, in %
386: particular concerning synchronizability, see \cite{ABJ}.} as the coupling
387: network. Figs.~(a),(b),(c) and (d) plot the variation of synchronization
388: measure ($\sigma^2$), and measure of homogeneity ($\varsigma^2$) as a function
389: of coupling strengths and Fig.~($a^\prime$), ($b^\prime$),
390: ($c^\prime$) and ($d^\prime$) plot the
391: transition probability $P(\alpha, \alpha)$ for different nodes.
392: \begin{figure*}
393: \includegraphics[bb=83 138 496 836,height=15cm]{fig1.ps}
394: \caption{Examples of coupled networks showing all three
395: phases.
396: All figures are plotted for scale-free networks, generated by
397: using BA algorithm \cite{BA}, of size $N=200$ and
398: (a) average degree 2 (phase 1),
399: (b) average degree 6, (c) average degree 10, (d) average degree 20.
400: The $x$-axis represents the coupling strength and the $y$-axis gives
401: the synchronization measure $\sigma^2$ ($\circ$) and the homogeneity measure
402: $\varsigma^2$ ($\bullet$) for the whole network. The largest Lyapunov
403: exponent is plotted as a function of the coupling strength (see inset).
404: Figs.~($a^\prime$), ($b^{\prime}$), ($c^{\prime}$) and ($d^{\prime}$) show
405: exact values of the transition probability $P(\alpha, \alpha)$ for different
406: nodes as a function of the coupling strength. For clarity we plot only a few
407: arbitrarily selected nodes.}
408: \label{sym-scale}
409: \end{figure*}
410: We start with the example of networks having coupled dynamics in phase one
411: (non-homogeneous unsynchronized) and we move towards the examples
412: of networks showing
413: homogeneous synchronized state, phase 3.
414:
415: For $\varepsilon < 0.2$, the coupled logistic map model (\ref{coup-dyn})
416: exhibits a similar behaviour for different coupling architectures,
417: with quasiperiodic behaviour around $\varepsilon=0.2$. The
418: behaviour varies with the coupling architecture for larger coupling
419: strengths. In the periodic regions the symbolic dynamics of the nodes,
420: given by (\ref{sym}), would always be similar irrespective of the underlying
421: coupling network. So, in the periodic regions we do not get any extra
422: information about the network by observing symbolic sequences, but
423: the symbolic dynamics is informative when the coupled dynamics lies on the
424: chaotic attractor. Subfigures~\ref{sym-scale}(a), $(a\prime)$, are plotted
425: for scale free networks with average degree 2. The transition probabilities
426: $P(\alpha, \alpha)$ are completely different for the different nodes
427: (except for $\varepsilon < 0.2$). Here, the nodes are not
428: synchronized, $\sigma^2$ being nonzero for the entire coupling strength
429: range. This state corresponds to {\it phase one}. Interesting phenomena
430: occur when we increase the number of connections in the networks.
431: Subfigures~\ref{sym-scale}(b) and ($b^\prime$) are plotted for a
432: scale-free network
433: with average degree 6. It can be seen that $P(\alpha,\alpha)$ for
434: different nodes are
435: remarkably similar. Note that we calculate $P(\alpha, \alpha)$ for
436: coupled dynamics being in the chaotic and unsynchronized regime
437: ($\lambda_l$ and $\sigma^2$ both are greater then zero). So we do not
438: take periodic and synchronized regions into account
439: which obviously yield similar symbolic dynamics for all the nodes.
440: This homogeneity becomes more prominent as we increase the number of
441: connections in the network. In Figs.~\ref{sym-scale}(c), ($c^\prime$) and
442: \ref{sym-scale} (d), ($d^\prime$),
443: The transition probabilities of all the nodes are the same except for a
444: few places where some nodes have different transition probability
445: (e.g.~node number 50 in ($c^\prime$) having a different value of
446: $P(\alpha, \alpha)$). Note
447: that here the nodes are not synchronized, which is indicated by the
448: nonzero value of $\sigma^2$ throughout the coupling range, except
449: for $\varepsilon=1$ in (c) and for $\varepsilon > 0.8$ in (d).
450:
451: The second interesting feature is
452: the qualitatively similar behaviour of the
453: largest Lyapunov exponent, which is calculated from ~(\ref{coup-dyn}),
454: and $P(\alpha, \alpha)$, which is calculated from a scalar time series
455: of an arbitrarily selected node. Note that the time series used for
456: the calculation of $P(\alpha, \alpha)$ is very short compared to the
457: traditional methods to calculate the largest Lyapunov exponent from a scalar
458: time series. This similar behaviour of the Lyapunov exponent
459: and the ordering relations between the values of the state variable was
460: first observed by Bandt and Pompe \cite{PE1} in the case of isolated
461: dynamics. Here we show that a similar relation exists for
462: coupled dynamics, depending upon the network
463: parameters, namely the connection architecture and the connection ratio
464: $N_c/N^2$.
465:
466: Fig.~\ref{sym-lya} is plotted for various networks being in the
467: {\it phase two} (homogeneous unsynchronized phase). They show
468: the similar behaviour of $\lambda_l$ and
469: $P(\alpha, \alpha)$ of any arbitrarily selected node.
470: For $k$-nearest neighbour coupled networks we always find the
471: homogeneous phase,
472: independent of the average degree or the ratio $N_c/N^2$. This is
473: because of the symmetry between the nodes.
474: \begin{figure*}
475: \includegraphics[bb=60 455 402 780, width=10cm,height=8cm]{fig2.ps}
476: \caption{The global measure of coupled dynamics from the local symbolic
477: dynamics. We take various networks having coupled dynamics in the
478: {\it phase two}. The $x$-axis gives the coupling strength $\varepsilon$ and
479: the $y$-axis depicts $\lambda_l$ (-)(largest Lyapunov exponent for the
480: coupled dynamics) as well as $P(\alpha, \alpha)$ (...)(transition probability
481: for a randomly selected node). (a) for nearest neighbour coupled network of
482: size $N=20$, (b) for 3-nearest neighbour coupled network, $N=50$, (c) and
483: (d) are for random and scale-free networks, respectively, with average
484: degree 10 and $N=200$.}
485: \label{sym-lya}
486: \end{figure*}
487: In all the subfigures, $P(\alpha, \alpha)$ qualitatively matches with the
488: largest Lyapunov exponent of the coupled dynamics.
489: At certain $\varepsilon$ values $P(\alpha,\alpha)$ is
490: almost zero, whereas $\lambda_l$ is positive. This indicates ordered
491: behaviour (Ref. \cite{PS1}) of the coupled system, for example Fig.2(c)
492: and Fig.4(c) are plotted for the same network. When
493: $P(\alpha,\alpha)$ is very small the coupled dynamics is in the high state of
494: the phase synchronization (Fig.4(c)) (i.e. almost all nodes are
495: forming phase synchronized cluster(s) Ref.\cite{PS1}) though the largest
496: Lyapunov exponent still remains positive (Fig.2(c)).
497:
498: In {\it phase three}, which is synchronized phase,
499: the largest Lyapunov exponent would simply be the
500: Lyapunov exponent of the uncoupled map and all the nodes will have the same
501: transition probability as of the uncoupled node.
502:
503: \section{Relation between dynamical phases and network properties}
504: We can also exhibit a direct relation between network parameters and
505: dynamical behaviour. The symmetry properties of the network
506: and the connection density affect the homogeneity
507: of the symbolic sequences.
508: Fig.~\ref{sym-nc} plots the deviation from the homogeneity
509: indicated by $\varsigma^2_{\varepsilon} =
510: \left\langle \frac{1}{N}\sum_{i=1}^N[P_i(\alpha,\alpha)-
511: \overline{P(\alpha,\alpha)}]^2 \right\rangle_{\varepsilon}$,
512: as a function of $2N_c/N(N-1)$. Here, $P_i$ is the transition probability of $i$th
513: node and $\overline{P(\alpha,\alpha)} =
514: \frac{1}{N} \sum_{i=1}^N P_i(\alpha, \alpha)$, and
515: $\left \langle \cdot \right \rangle_{\varepsilon}$ denotes the average over all coupling
516: strengths.
517: We start with one-dimensional nearest neighbour coupled
518: network (homogeneous phase) and randomly add connections. For
519: nearest neighbour coupled networks we obtain the homogeneous phase, as already
520: explained in the previous section. As we add the connections randomly,
521: first the homogeneity gets perturbed, but gets
522: reestablished as the number of connections is increased further. Note that
523: here we always calculate the deviation in the non-synchronized regime only,
524: because the synchronized regime obviously corresponds to the homogeneous phase.
525: For each randomly added connection we take the average of the twenty networks.
526: Note that in this region (phase two) $\sigma^2$ is not zero.
527: For $N_c/N^2$ close to one, we get a synchronized
528: state after a coupling strength \cite{SJ-analytic} that corresponds to the
529: homogeneous state (phase three).
530: \begin{figure}
531: \includegraphics[bb=67 619 252 780]{fig3.ps}
532: \caption{The measure of
533: homogeneity $\varsigma^2$ as a function of the connectivity ratio $2 N_c/N(N-1)$.}
534: \label{sym-nc}
535: \end{figure}
536: Note that for the local dynamics in the chaotic regime, only the network
537: property seem to be responsible for the homogeneous or non-homogeneous
538: behaviour of the coupled dynamics. Fig.~\ref{sym-nc} is plotted for the
539: coupled logistic map but a similar behaviour is shown by the H\'enon maps
540: also (see section VII).
541:
542: \section{Phase synchronization : Symbolic synchronization}
543: If nodes $i$ and $j$ have the same symbolic dynamics, $s_t(i) = s_t(j)$, then
544: we say nodes $i, j$ are symbolically synchronized. Also, a cluster of nodes
545: is symbolically synchronized if all pairs of nodes belonging to that cluster
546: are symbolically synchronized. Note that in a symbolically synchronised
547: cluster, the state values of the nodes may differ. The symbolic
548: synchronization is observed in the {\it phase two}, where the number
549: of the connections in the networks is very small, in general of the order
550: of $N$. With the increase in the number of connections we usually get a
551: fully synchronized cluster, which is trivially symbolically synchronized.
552: Many real-world networks are sparsely connected $(N_c \sim N)$,
553: and complete synchronization is relatively rare, though phase synchronization or
554: symbolic synchronization is possible. We show that $P(\beta, \beta)$ can
555: be used as a good measure of the phase synchronization in the coupled map
556: network (\ref{coup-dyn}). Fig.~\ref{sym-phase} shows the correlation between
557: the phase synchronization and the transition probability $P(\beta,\beta)$
558: of an arbitrary selected node.
559: \begin{figure*}
560: \includegraphics[bb=61 369 502 780,width=14cm,height=10cm]{fig4.ps}
561: \caption{The ratio of the number phase synchronized clusters to
562: the maximum possible clusters, $N_\text{clus}/N$ and
563: the transition probability $P(\beta, \beta)$ for the coupled logistic map
564: as a function of the coupling strength
565: $\varepsilon$, (a) for a nearest neighbour coupled network with average
566: degree 20 and $N=100$, (b) for a scale-free network with average degree 10 and
567: $N=100$, (c) for a random network with average degree 10 and $N=100$,
568: (d) for a nearest neighbour coupled network with average degree 6 and $N=50$,
569: and the tent map $(f(x)=(1-2|x-\frac{1}{2}|)$ as the local chaotic function.}
570: \label{sym-phase}
571: \end{figure*}
572: We see that in the homogeneous region $P(\beta, \beta)$ matches
573: considerably well with the phase synchronization. Note that in partially
574: ordered phase region, phase synchronized clusters vary with time. We plot
575: the number of clusters calculated for a certain time length, and the number
576: of clusters may change with the evolution of the coupled dynamics. Therefore
577: at some coupling strength region(s), the transition probability
578: $P(\beta, \beta)$ does not match with the phase synchronization
579: (For example in the Fig.~\ref{sym-phase}(c), at coupling strength 0.59,
580: the value of $P(\beta, \beta)$ is very high although the nodes are phase
581: synchronized).
582:
583: \section{Coupled H\'enon maps}
584: In this section we apply our method to coupled H\'enon maps.
585: The H\'enon map is a two-dimensional map
586: \cite{henon},
587: \begin{eqnarray}
588: x(t+1) = y(t) + 1 - a x(t)^2 \nonumber \\
589: y(t+1) = b y(t). \nonumber
590: \end{eqnarray}
591: When one introduces the possibility of a time delay, the above equation
592: can be written as the scalar equation,
593: \begin{equation}
594: x(t+1) = b x(t-1) + 1 - a x(t)^2
595: \label{henon-map}
596: \end{equation}
597: For the parameters we take the values
598: $a = 1.4$ and $b = 0.3$, for which the H\'enon map
599: is known to have a chaotic attractor.
600:
601: We define the symbolic dynamics as given in (\ref{sym}).
602: The choice of the threshold $x^{\ast}$ requires some care.
603: A judicious choice should make certain short transition
604: probabilities very small, which may be useful for detecting network dynamics
605: from single-node measurements \cite{pap1}. Clearly,
606: increasing the threshold decreases the probability of occurrence of the
607: repeated sequence $\beta\beta$. However, it also decreases the probability of
608: observing
609: the single symbol $\beta$, making it difficult to work with short time series.
610: Hence, the choice of the threshold is a compromise between these two effects.
611: We use the natural density defined by the data to choose a threshold.
612: Fig.~\ref{fig:henonx} depicts how the the probabilities of observing
613: a single symbol $\beta$ and the repeated sequence $\beta\beta$
614: change depending on the value
615: of the threshold $x^{\ast}$. It can be seen that a choice of $x^{\ast}$
616: roughly in the range $(0.55,1.20)$ would be useful, since it renders the
617: sequence $\beta\beta$ very unlikely without constraining the occurrence of
618: the symbol $\beta$. Note that it is immediate from their definitions that
619: the probabilities $P(\beta)$ and $P(\beta,\beta)$ will be decreasing as
620: functions of $x^{\ast}$, and will approach zero as $x^{\ast}$ increases;
621: furthermore, $P(\beta)>P(\beta,\beta)$. It follows that one can find a
622: threshold $x^{\ast}$ for which $P(\beta)$ is large compared to
623: $P(\beta,\beta)$. Fig.~\ref{fig:henonx} shows the ratio
624: $P(\beta,\beta)/P(\beta)$, and the sharp decrease at about
625: $x^{\ast}\approx 0.6$ suggests to take some value near 0.6 as the threshold, yielding a
626: very small $P(\beta,\beta)$ and a large $P(\beta)$ at the same time.
627: \begin{figure}
628: \centering
629: \includegraphics[width=\columnwidth]{fig5.ps}
630: \caption{Illustration of the choice of the threshold $x^{\ast}$
631: as the point where $P(\beta,\beta)/P(\beta)$ sharply drops to near zero. }
632: \label{fig:henonx}
633: \end{figure}
634: We evolve (\ref{coup-dyn}) starting from random initial conditions, with
635: (\ref{henon-map}) as local dynamics,
636: and estimate the transition probabilities $P(i, j)$ as discussed in
637: the first section, using a time series of length
638: $\tau =1000$.
639: At the globally synchronized state
640: $x_{i}(t)=x_{j}(t); \, \forall i,j, t$, with all
641: nodes evolving according to the rule (\ref{henon-map}), the symbolic sequences
642: measured from a node will be
643: subject to the same constraints as that generated by (\ref{henon-map}).
644:
645: We now discuss some results based on numerical simulations
646: on various networks. Fig.~\ref{fig1-henon} plots the
647: transition probabilities as a function of the coupling strength.
648: We consider the symbolic sequences of length two and three.
649: For length two, we consider the transition probabilities
650: $P(\alpha, \alpha)$ and $P(\beta,\beta)$. For sequences of length three we
651: have 6 possible transitions, but some of them are very small
652: (like $P(\beta,\beta,i)$, $i$ being $\alpha, \beta$ ), so we plot only
653: those transition probabilities which vary with the couplings.
654: It is clear from the figures that the synchronized state is easily
655: detected by looking at the transition probabilities of any arbitrarily
656: selected node. Whenever the transition probabilities are equal to the
657: transition probabilities of the map (\ref{henon-map}), the network is
658: globally synchronized.
659: \begin{figure*}
660: \includegraphics[bb=57 418 585 780,width=15cm]{fig6.ps}
661: \caption{The transition probability measure for
662: coupled H\'enon maps. The $x$-axis displays the coupling strength and the
663: $y$-axis shows the different
664: transition probabilities and the measure of synchronization $\sigma^2$.
665: (a) is for two coupled nodes and
666: plots $P(\alpha, \alpha)$ for the
667: symbolic sequence of length 2. (b) and (c) are plotted for globally coupled
668: networks with $N=50$. (b) plots $P(\alpha, \alpha)$
669: and (c) plots transition probabilities for symbolic sequences of length 3,
670: namely $P(\alpha, \alpha, \alpha)$ ($\square$), $P(\alpha, \alpha, \beta)$
671: ($\bullet$), and $P(\beta, \alpha, \alpha)$ ($\circ$).
672: The synchronized
673: state is detected when all the transition probabilities are equal to those
674: of the uncoupled map; i.e. the transition probabilities at
675: the zero coupling strength. (d), (e) and (f) show the standard deviation
676: $\sigma^2$ (solid thick line) and
677: $\delta^2$ (vertical dashed line) for
678: these three transition probabilities of an arbitrary selected node with
679: respect to the transition probabilities of the uncoupled function
680: (solid line), i.e for $\varepsilon=0$.
681: $\delta^2$ is calculated for 20 simulations for the dynamics with different
682: sets of random initial conditions. (d) for
683: a globally connected network with $N=50$, (e) and (f) for a random network
684: with $N=100$, average degree 10 and 2 respectively. The last subfigure is
685: plotted to show the behaviour of the transition probabilities when we do
686: not get global synchrony even at large coupling strengths. }
687: \label{fig1-henon}
688: \end{figure*}
689: It is clear from subfigures (d),(e) and (f) that for the synchronized region
690: (zero $\sigma^2$), the deviation of transition probabilities from the
691: transition probabilities of the uncoupled map (\ref{henon-map}) is
692: also zero.
693: Here, the deviation of $P(i,j)$ of any node is defined as
694: $\delta^2_{i,j} = \left\langle \frac{1}{m-1}\sum_{k=1}^{m}[P_k(i,j)-
695: \overline{P(i, j)}]^2]\right\rangle$, where
696: $\overline{P(i, j)} = \frac{1}{m}\sum_k P_k(i,j)$, calculated at $\varepsilon=0$
697: $k = 1, \dots \dots $ are $m$ different sets of random initial conditions
698: taken between $-1.5$ and $1.5$.
699: In all the figures (except (f)) we get synchronization for larger
700: coupling strengths, so the deviation is almost zero there, i.e. all
701: transition probabilities match
702: completely with those of the uncoupled map. Note that there are certain regions
703: (small coupling strength range $\varepsilon < 0.2$) where the nodes do not
704: get synchronized while the deviations are quite small. That is because for
705: sufficiently small coupling strength, couplings
706: do not affect the behaviour of the individual nodes very much, and so the
707: transition probabilities do not differ much from those
708: corresponding to the uncoupled function.
709: However, as we increase the coupling strength, the transition probabilities
710: become dependent on the couplings.
711: Still, if we look at the small coupling strength regions carefully
712: we see that not all the deviations are small. For example, although the
713: deviations of $P(\beta, \alpha, \alpha)$ (- - -)
714: and $P(\alpha, \alpha, \beta)$ (-) are very small, the deviation in
715: $P(\alpha, \alpha, \alpha)$ ($\dots$) is still large, whereas for the
716: synchronized regime all deviations are very close to zero.
717: \begin{figure}
718: \includegraphics[bb=67 619 252 780]{fig7.ps}
719: \caption{The measure of
720: homogeneity $\varsigma^2$ as a function of connection ratio $2 N_c/N(N-1)$,
721: with H\'enon map as the local dynamical function.}
722: \label{sym-nc-henon}
723: \end{figure}
724: Fig.~\ref{sym-nc-henon} plots the deviation from the homogeneity
725: $\varsigma^2$ as a function of $N_c/N^2$ (see the caption
726: of Fig.~\ref{sym-nc}, which shows
727: a similar plot with logistic map as a local dynamical function).
728: The only difference is that
729: for the H\'enon map we plot the transition probability of three-symbol
730: sequences instead of two-symbol sequences for logistic and tent maps.
731:
732: \section{conclusion}
733: We have studied the symbolic dynamics of coupled maps on networks. We define
734: our symbolic dynamics based on non-generating partitions leading to some
735: forbidden transitions of symbols in the time evolution of the function. The
736: optimal partitions are those which lead to the maximal difference
737: between the permutation entropy of the dynamical iteration and corresponding
738: random iteration. For one-dimensional systems finding these partitions
739: is simple, whereas for higher dimensional systems it may be more difficult.
740: However, it turns out that symbolic dynamics drawn from any non-generating
741: partitions is usually good enough for the applications we have considered in
742: this paper. The symbolic dynamics can be drawn when the system parameters are
743: not known, as well as for experimental data taken in a noisy environment.
744:
745: We use symbolic dynamics as a measure of dynamical state of the coupled
746: system and show various applications of this measure.
747: We define three different states of the coupled dynamics based on
748: the synchronization and the symbolic dynamical properties.
749: In the homogeneous synchronized phase, complete synchrony is detected by
750: comparing the transition probabilities of any arbitrarily selected node
751: with those of the uncoupled function. In this state the coupled dynamics
752: collapses to the dynamics of the uncoupled function, and the symbolic
753: dynamics of any node is subject to the same constraints as that generated by
754: the uncoupled function.
755:
756: Phase two, which
757: refers to the {\it homogeneous unsynchronized phase} or
758: {\it symbolic synchronized phase}, is of our prime interest where the nodes
759: are not synchronized but have identical symbolic dynamics.
760: Although these phases are detected dynamically, we find that the homogeneous
761: unsynchronized phase is related to the connection density ($N_c/N^2$)
762: and to a smaller extent
763: to the chaotic dynamical function used. This region is generally observed for
764: networks with $N_c \sim r \times N^2$ where $0.05 < r < 0.1$. Most of the
765: real networks are sparsely connected and come under the category of phase
766: two. In this phase we can deduce the global properties of the coupled
767: dynamics such as the largest Lyapunov exponent and phase synchronization by
768: simply observing the local symbolic dynamics of any randomly selected node.
769:
770: As it is expected from such a
771: simplistic reduction, our symbolic dynamics gives only the qualitative
772: understanding. A more precise calculation of complexity through some entropy
773: measure of the system based on the symbolic dynamics is one of the future
774: steps. Further future investigations will involve an analytical understanding
775: of symbolic synchronization and application to detect various levels of
776: synchronization in experimental data taken from
777: coupled systems.
778:
779: \begin{thebibliography}{apssamp}
780:
781: \bibitem{rev-CM} K. Kaneko, {\it Formation, Dynamics and Statistics of Pattern},
782: Edited by K. Kawasaki et al. (World Sci., Singapore, 1990).
783: \bibitem{rev-kaneko}K. Kaneko, {\it Chaos} {\bf 2}, 279 (1992).
784: \bibitem{SJ-REA-insa} S. Jalan and R. E. Amritkar,
785: {\it Indian Natn Sci Acad} {\bf 71 , A} No. 1-2 (2005).
786:
787: \bibitem{sym-book}D. Lind and B. Marcus, {\it An introduction to
788: Symbolic Dynamics and Coding} (Cambridge Univ. Press, 1995).
789:
790: \bibitem{sym-book2}Bai-Lin Hao and Wei-Mou Zheng,
791: {\it Applied Symbolic Dynamics and Chaos} (Word Sci.,
792: Singapore, 1998).
793:
794: \bibitem{sym-rev}A. Endler and J. A. C. Gallas, {\it Phys. Rev. E} {\bf 65},
795: 036231 (2002);R. L. Davidchack, Y.-C. Lai, E. M. Bollt, and M. Dhamala,
796: {\it Phys. Rev. E} {\bf 61}, 1353 (2000); A. Witt, R. Braun, F. Feudel, C.
797: Grebogi, and J. Kurths, {\it Phys. Rev. E} {\bf 59}, 1605 (1999);
798: E. Bollt and Y.-C. Lai, {\it Phys. Rev. E} {\bf 58}, 1724 (1998);
799: W.-M. Zheng, {\it Phys. Rev. E} {\bf 56}, 1556 (1997);
800: F. Christiansen and A. Politi, {\it Phys. Rev. E} {\bf 51}, R3811 (1995).
801:
802: \bibitem{sym-CD} Leonid A. Bunimovich, {\it Physica D}
803: {\bf 103}, 1 (1997); S. D. Pethel, N. J. Corron and E. Bollt,
804: {\it Phys. Rev. Lett.} {bf 96}, 034105 (2006).
805:
806: \bibitem{pap1}S. Jalan, F. M. Atay, and J. Jost (nlin.CD/0510057).
807:
808: \bibitem{kurths-TS}H. Voss and J. Kurths,
809: {\it Phys. Rev. E} {\bf 58}, 1155-1158 (1998).
810: \bibitem{PE1}C. Bandt and B. Pompe, {\it Phy. Rev. Lett.}, {\bf 88} 174102
811: (2002).
812:
813: \bibitem{book-syn}A. Pikovsky, M. Rosenblum and J. Kurths, {\it
814: Synchronization : A Universal Concept in Nonlinear Dynamics} (Cambridge
815: Univ. Press, 2001).
816: \bibitem{phyreport-syn} S. Boccaletti, J. Kurths, G. Osipov, D.L. Valladares
817: and C. S. Zhou, `The Synchronization of Chaotic Systems,
818: {\it Phys. Rep.} {\bf 366}, 1 (2002).
819:
820: \bibitem{syn} T. Zhou, L. Chen and K. Aihara, {\it Phys. Rev. Lett.}
821: {\bf 95}, 178103 (2005); R. E. Amritkar, S. Jalan and C. K. Hu,
822: {\it Phys. Rev. E} {\bf 72}, 016212 (2005);
823: I. Belykh, E. de Lange, and M. Hasler,
824: {\it Phys. Rev. Lett.} {\bf 94}, 188101 (2005); D.-U. Hwang, M. Chavez,
825: A. Amann, and S. Boccaletti, {\it Phys. Rev. Lett.} {\bf 94}, 138701
826: (2005); A. Zumdieck, M. Timme, T. Geisel, and F. Wolf, {\it Phys. Rev. Lett.}
827: {\bf 93}, 244103 (2004); M. G. Rosenblum and A. S. Pikovsky,
828: {\it Phys. Rev. Lett.} {\bf 92}, 114102 (2004); H. Xiao, G. Hu, and Z. Qu,
829: {\it Phys. Rev. Lett.} {\bf 77}, 4162-4165 (1996);
830: Â. Corral, C. J. P‰rez, A. D-Guilera, and A. Arenas, {\it Phys. Rev. Lett.}
831: {\bf 74}, 118-121 (1995).
832:
833: \bibitem{PS1}M. G. Rosenblum, A. S. Pikovsky, and J. Kurths,
834: {\it Phys. Rev. Lett.} {\bf 76}, 1804 (1996);
835: S. Jalan and R. E. Amritkar, {\it Phys. Rev. Lett.}, {\bf 90}
836: 014101 (2003); S. Jalan, R. E. Amritkar, C. K. Hu,
837: {\it Phys. Rev. E}, {\bf 72}, 016211 (2005).
838:
839: \bibitem{jost-kiran}J. Jost and K. M. Kolwankar,
840: {\it Fractals in Engineering - New Trends in Theory and Applications},
841: page 57 (Springer, London, 2005).
842:
843: \bibitem{lya-rev}H. D. I. Abarbanel, R. Brown, J. J. Sidorowich, and
844: L. Sh. Tsimring, {\it Rev. Mod. Phys.} {\bf 65}, 1331-1392 (1993).
845:
846: \bibitem{CM}K. Kaneko, {\it Physica D} {\bf 34}, 1-41 (1989).
847:
848: \bibitem{MSF}L. M. Pecora and T. L. Carroll, {\it Phys. Rev. Lett.}
849: {\bf 80}, 2109 (1998).
850:
851: \bibitem{BA}A. -L. Barabasi, R. Albert, H. Jeong, {\it Physica A}, {\bf 281},
852: 69 (2000).
853: \bibitem{ABJ} F. M. Atay, T. Biyikoglu, and J. Jost,
854: {\it IEEE Trans. Circuits and Sys. I}, {\bf 53(1)}, 92--98, 2006.
855:
856: \bibitem{SJ-analytic}R. E. Amritkar, S. Jalan and C. K. Hu, {\it Phys. Rev. E}
857: {\bf 72}, 016212 (2005).
858:
859: \bibitem{henon} A. J. Lichtenberg and M. A. Lieberman,
860: {\it Regular and Chaotic Dynamics} (Springer-Verlag, 1983).
861:
862: \end{thebibliography}
863:
864: \end{document}
865:
866:
867: