1: \documentclass [12pt]{article}
2: \usepackage {graphicx}
3: \topmargin=-5mm
4: \textheight=245mm
5: \textwidth=165mm
6: \oddsidemargin=0mm
7:
8: \begin{document}
9:
10: \title{An example of physical system with hyperbolic attractor of Smale -- Williams
11: type}
12:
13: \author{S.~P.~Kuznetsov}
14:
15:
16: \maketitle
17:
18: \begin{center}
19: \textit{Saratov Division of Institute of Radio-Engineering and Electronics,
20: Russian Academy of Sciences, Zelenaya 38, Saratov, 410019, Russia}
21: \vspace{2mm}
22: \end{center}
23:
24: \begin{abstract}
25: A simple and transparent example of a non-autonomous flow system, with
26: hyperbolic strange attractor is suggested. The system is constructed on a
27: basis of two coupled van der Pol oscillators, the characteristic frequencies
28: differ twice, and the parameters controlling generation in both oscillators
29: undergo a slow periodic counter-phase variation in time. In terms of
30: stroboscopic Poincar\'{e} section, the respective four-dimensional mapping
31: has a hyperbolic strange attractor of Smale -- Williams type. Qualitative
32: reasoning and quantitative data of numerical computations are presented and
33: discussed, e.g. Lyapunov exponents and their parameter dependencies. A
34: special test for hyperbolicity based on statistical analysis of
35: distributions of angles between stable and unstable subspaces of a chaotic
36: trajectory has been performed. Perspectives of further comparative studies
37: of hyperbolic and non-hyperbolic chaotic dynamics in physical aspect are
38: outlined.
39: \end{abstract}
40:
41: Mathematical theory of chaotic dynamics based on rigorous axiomatic
42: foundation exploits a notion of hyperbolicity, which implies that all
43: relevant trajectories in phase space of a dynamical system are of saddle
44: type, with well defined stable and unstable directions \cite{1,2,3,4}. Hyperbolic
45: systems of dissipative type, contracting the phase space volume, manifest
46: robust strange attractors with strong chaotic properties. The robustness
47: (structural stability) implies insensitivity of the motions in respect to
48: variations of equations governing the dynamics. In particular, positive
49: Lyapunov exponent depends on parameters in smooth manner, without flops into
50: negative region characteristic to non-hyperbolic attractors. A Cantor-like
51: structure of the strange attractor persists without qualitative changes
52: (bifurcations), at least while the variations are not too large. Textbook
53: examples of these robust strange attractors are represented only by
54: artificial mathematical constructions associated with discrete-time models,
55: e.g. Plykin attractor and Smale -- Williams attractor (solenoid).
56:
57: It seems that the mathematical theory of hyperbolic chaos has been never
58: applied conclusively to any physical object, although concepts of this
59: theory are widely used for interpretation of chaotic behavior of realistic
60: nonlinear systems. On the other hand, feasible nonlinear systems with
61: complex dynamics, such as Lorenz and R\"{o}ssler equations, chaotic
62: self-oscillators, driven nonlinear oscillators etc. do not relate to the
63: true hyperbolic class \cite{4,5,6}. As a rule, observable chaos in these systems is
64: linked with a so-called quasiattractor, a set in phase space, on which
65: chaotic trajectories coexist with stable orbits of high periods (usually,
66: they are non-distinguishable in computations at reasonable accuracies).
67: Mathematical description of quasiattractors remains a challenging problem,
68: although in physical systems the non-hyperbolicity is masked effectively due
69: to presence of inevitable noise. In few cases, e.g. in Lorenz model in some
70: appropriate domain of parameter space, dynamics is proved to be
71: quasi-hyperbolic (with some restrictions concerning violation of smoothness
72: conditions) \cite{7}.
73:
74: I am aware of only two theoretical works, which discuss examples of true
75: hyperbolic dynamics in context of systems governed by differential
76: equations. One relates to a system called triple linkage, which allows in a
77: frictionless case a description in terms of orbits on a surface of negative
78: curvature. In dissipative case, it gives rise to a hyperbolic chaotic
79: attractor \cite{8}. Another work deals with a 3D flow system motivated by neural
80: dynamics and argues in favor of existence of an attractor of Plykin type in
81: the Poincar\'{e} map associated with the flow \cite{9}.
82:
83: In this Letter, I suggest an essentially simpler and transparent example of
84: a non-autonomous flow system, which apparently manifests a hyperbolic
85: attractor. In terms of stroboscopic Poincar\'{e} map, it is an attractor of
86: the same kind as the Smale -- Williams solenoid, but embedded in a 4D rather
87: then 3D state space.
88:
89: The system is constructed on a basis of two van der Pol oscillators with
90: characteristic frequencies $\omega _{0}$ and $2\omega _{0}$,
91: respectively. The control parameters of the oscillators responsible for the
92: Andronov -- Hopf bifurcations in the autonomous subsystems are forced to
93: swing slowly, periodically in time. On a half-period, the first oscillator
94: is above the generation threshold, while the second one is below the
95: threshold. On another half-period, a situation is opposite. Next, we assume
96: that the first oscillator acts on the partner via a quadratic term in the
97: equation. The generated second harmonic component serves as a primer for the
98: second oscillator, as it comes off the under-threshold state. In turn, the
99: second oscillator acts on the first one via a term represented by a product
100: of the dynamical variable and an auxiliary signal of frequency
101: $\omega_{0}$. Thus, a component with the difference frequency appears, which fits
102: resonance range for the first oscillator and serves as a primer as it starts
103: to generate.
104:
105: Summarizing this description, we write down the following equations:
106: \begin{equation}
107: \label{eq1}
108: \begin{array}{l}
109: \ddot {x} - (A\sin {2\pi t /T} - x^2)\dot {x} + \omega _0^2 x =
110: \varepsilon y\sin \omega _0 t, \\
111: \ddot {y} - (- A\sin {2\pi t /T} - y^2)\dot {y} + 4\omega _0^2 y =
112: \varepsilon x^2, \\
113: \end{array}
114: \end{equation}
115:
116: \noindent
117: where $x$ and $y$ are dynamical variables of the first and the second oscillators,
118: respectively, $A$ is a constant designating amplitude of the slow swing of the
119: control parameters, $\varepsilon $ is a coupling parameter.
120:
121: We assume that the period of swing $T$ contains an integer number of periods of
122: the auxiliary signal: $T = 2\pi N /\omega_0$. Thus, our
123: set of non-autonomous equations has periodic rather than quasiperiodic
124: coefficients. It is appropriate to treat the dynamics in terms of
125: stroboscopic Poincar\'{e} section using a period-$T$ sequence of time instants.
126: The Poincar\'{e} map is four-dimensional and acts in a space of vectors
127: $\left\{ {x,\,\,\dot {x}/{\omega _0 },\,y,\,\,\dot {y}/{(2\omega _0 )}} \right\}$.
128:
129: The system (\ref{eq1}) operates as follows. Let the first oscillator have some phase
130: $\varphi $ on a stage of generation:
131: $x \propto \cos (\omega _0 t + \varphi)$.
132: Squared value $x^{2}$ contains the second harmonic: $\cos (2\omega _0 t +
133: 2\varphi )$, and its phase is $2\varphi $. As the half-period comes to the
134: end, and the second oscillator becomes excited, the induced oscillations of
135: the variable $y$ get the same phase $2 \varphi $. Mixture of these oscillations
136: with the auxiliary signal transfers the doubled phase into the original
137: frequency range. Hence, on the next stage of excitation the first oscillator
138: accepts this phase $2 \varphi $ too. Obviously, on subsequent stages of swing
139: the phases of the first oscillator follow approximately the mapping
140:
141: \begin{equation}
142: \label{eq2}
143: \varphi _{n + 1} = 2\varphi _n \,\,\,(\bmod 2\pi ).
144: \end{equation}
145:
146: Figure 1 shows a typical pattern of time dependence of $x$ and $y$ from numerical
147: solution of Eqs. (\ref{eq1}) by Runge -- Kutta method for particular parameter
148: values $\omega _0 = 2\pi ,\,\,T = N = 10,\,\,A = 3,\,\,\varepsilon = 0.5$
149: together with a diagram of empirical mapping for phase $\varphi _{n + 1} $
150: versus $\varphi _n $. The phases are determined at the centers of the
151: excitation stages for the first oscillator:
152:
153: \begin{equation}
154: \label{eq2a}
155: \varphi = \left\{ {{\begin{array}{*{20}l}
156: {\arctan {(\omega_0^{ - 1} \dot {x}/x)},\,\, x > 0,} \hfill \\
157: {\arctan {(\omega_0^{ - 1} \dot {x}/x)}+\pi,\,\, x > 0,} \hfill \\
158: \end{array} }} \right.
159: \end{equation}
160:
161: \noindent
162: and are plotted over a sufficiently large number of the basic periods $T$. The
163: mapping for the phase looks, as expected, topologically equivalent to the
164: relation (\ref{eq2}). (Some distortions arise due to imperfection of the above
165: qualitative considerations and of the definition of phase; the
166: correspondence becomes better at larger period ratios $N$.) Chaotic nature of
167: the dynamics reveals itself in a random walk of humps in respect to the
168: envelope of the generated signal on subsequent periods of swing.
169:
170: \begin{figure}[htbp]
171: \begin{center}
172: \includegraphics[width=6.5in]{An1.eps}
173: \end{center}
174: \label{fig1}
175: \caption{A typical pattern of time dependence for variables $x$ and $y$
176: obtained from numerical solution of Eqs. (\ref{eq1}) for
177: $\omega _0 = 2\pi ,\,\,T = N = 10,\,\,A = 3,\,\,\varepsilon = 0.5$ (a)
178: and a diagram of empirical mapping for phases of the first oscillator
179: defined in the centers of the stages of excitation
180: numbered by $n$}
181: \end{figure}
182:
183: In terms of stroboscopic Poincar\'{e} map, attractor of the system
184: corresponds exactly to the construction of Smale and Williams. In the
185: four-dimensional state space, the direction associated with the phase
186: $\varphi $ is expanding and gives rise to Lyapunov exponent estimated as
187: $\Lambda _1 \approx T^{ - 1}\log 2$. Three rest directions are contracting,
188: and they correspond to a three-dimensional stable manifold of the attractor.
189: Three respective Lyapunov exponents are negative. Interpreting the
190: stroboscopic Poincar\'{e} mapping, we may imagine a solid torus embedded in
191: 4-dimensional space and associate one iteration of the map with longitudinal
192: stretch of the torus, with contraction in the transversal directions, and
193: insertion of the doubly folded ``tube'' into the original torus interior.
194:
195: In computations, the Lyapunov exponents were evaluated with a help of
196: Benettin's algorithm \cite{10,11} from simultaneous solution of Eqs. (\ref{eq1}) together
197: with a collection of four exemplars of the linearized equations for
198: perturbations:
199:
200: \begin{equation}
201: \label{eq3}
202: \begin{array}{l}
203: \ddot {\tilde {x}} + 2x\dot {x}\tilde {x} - (A\sin {2 \pi t /T} - x^2)\dot {\tilde {x}}
204: + \omega _0^2 \tilde{x} = \varepsilon \tilde {y}\sin \omega _0 t, \\
205: \ddot {\tilde {y}} + 2y\dot {y}\tilde {y} - (- A\sin {2 \pi t /T} - y^2)\dot {\tilde {y}}
206: + 4\omega _0^2 \tilde{y} = 2\varepsilon x\tilde {x}. \\
207: \end{array}
208: \end{equation}
209:
210: In a course of the solution, at each step of the integration schema, the
211: Gram -- Schmidt orthogonalization and normalization were performed for four
212: vectors $\left\{ {\tilde {x}(t),{\dot {\tilde {x}}(t)}/{\omega _0 },
213: \tilde {y}(t),{\dot {\tilde {y}}(t)}/{2\omega _0 }} \right\}$, and the mean rates of
214: growth or decrease of the accumulated sums of logarithms of the norms (after
215: the orthogonalization but before the normalization) were estimated. As
216: found, the Lyapunov exponents for the attractor at the above mentioned
217: parameters are
218: $\Lambda _1 \approx 0.068 \approx T^{ - 1}\log 2$, $\Lambda_2 \approx - 0.35$,
219: $\Lambda _3 \approx - 0.59$, $\Lambda _4 \approx -0.81$.
220:
221: If the attractor is indeed hyperbolic, the chaotic dynamics must be robust
222: and retain its character under (at least small) variations of the equations.
223: As checked, this is indeed the case. In particular, the largest Lyapunov
224: exponent is almost independent on parameters, and the rest of them manifest
225: regular parameter dependences, as seen in Fig. 2. The left edge of the
226: diagram corresponds to violation of the hyperbolicity.
227:
228: Dynamical behavior of the same kind is observed at other integer period
229: ratios $N$, including essentially smaller ones, e.g. $N$=4. Figure 3 shows
230: portrait of the strange attractor in the Poincar\'{e} section in projection
231: onto the plane $(x,\dot {x})$ at $\omega _0 = 2\pi ,\,\,T = N = 4,\,\,A =
232: 8,\,\,\varepsilon = 0.5$. It looks precisely as the Smale -- Williams
233: attractor should look like. Observe fractal transversal structure of
234: ``strips'' constituting the attractor. For this attractor the Lyapunov
235: exponents are $\Lambda _1 \approx 0.167$, $\Lambda _2 \approx - 0.72$,
236: $\Lambda _3 \approx - 1.03$, $\Lambda _4 \approx - 1.50$. An estimate for
237: fractal dimension from Kaplan -- Yorke formula yields $D \approx 1.23$, and
238: that from the Grassberger -- Procaccia algorithm is $D \approx 1.26$.
239:
240: \begin{figure}[htbp]
241: \begin{center}
242: \includegraphics[width=4.2in]{An2.eps}
243: \end{center}
244: \label{fig2}
245: \caption{Computed Lyapunov exponents of the system (\ref{eq1}) versus
246: parameter $A$ at $\omega _0 = 2\pi $, $N = T = 10$, $\varepsilon = 0.5$.
247: Observe that the largest exponent remains almost constant in the whole
248: interval of hyperbolicity being in good agreement with the estimate
249: $\Lambda _1 = T^{ - 1}\log 2$.
250: }
251: \end{figure}
252:
253: It is interesting to perform a direct numerical test for hyperbolicity of
254: the attractor. Idea of such test was suggested in Refs. \cite{12} and \cite{13} and
255: applied for verification of hyperbolicity of trajectories of dynamical
256: systems, which have one stable and one unstable directions. The procedure
257: consists in computation of vectors of small perturbations along the
258: trajectory in forward and inverse time with measuring angles between the
259: forward-time and backward-time vectors at points of the trajectory. If zero
260: values of the angle do not occur, i.e. the statistical distribution of the
261: angles is essentially separated from zero, one concludes that the dynamics
262: is hyperbolic. If the statistical distribution shows non-vanishing
263: probability for zero angle, it implies non-hyperbolic behavior because of
264: presence of the homoclinic tangencies of stable and unstable manifolds. In
265: dissipative case these tangencies are responsible for the occurrence of
266: quasiattractor.
267:
268: \begin{figure}[htbp]
269: \begin{center}
270: \includegraphics[width=4in]{An3.eps}
271: \end{center}
272: \label{fig3}
273: \caption{
274: Portrait of the strange attractor in the stroboscopic Poincar\'{e} section
275: $t = 1\,\,(\bmod N)$ in projection onto the plane
276: $(x,\dot {x})$ at $\omega _0 = 2\pi ,\,\,T = N = 4,\,\,A = 8,\,\,\varepsilon = 0.5$
277: }
278: \end{figure}
279:
280: In our system (\ref{eq1}), only unstable subspace is one-dimensional, and the stable
281: one is three-dimensional. Therefore, the method needs a modification. An
282: adopted algorithm consists in the following. First, we generate a
283: sufficiently long representative orbit on the attractor
284: $\left\{ {x(t),{\dot {x}(t)}/{\omega _0 },y(t),
285: {\dot {y}(t)}/{2\omega _0 }} \right\}$ from the numerical solution of Eqs. (\ref{eq1}).
286: Then, we solve numerically the equations (\ref{eq3}) for a perturbation forward
287: in time. In a course of the solution, normalization of the vector
288: ${\rm {\bf a}}(t) = \left\{ {\tilde {x}(t),{\dot {\tilde {x}}(t)}/{\omega _0 },
289: \tilde {y}(t),{\dot {\tilde {y}}(t)}/{2\omega _0 }} \right\}$ is performed after each
290: step of integration to exclude the divergence. This vector determines an
291: unstable direction at each point of the orbit. Next, we solve a collection
292: of three exemplars of equations (\ref{eq3}) in backward time along the same
293: trajectory $\left\{ {x(t),{\dot {x}(t)}/{\omega _0},y(t),{\dot {y}(t)}/{2\omega _0 }} \right\}$
294: to get three
295: vectors $\left\{ {{\rm {\bf b}}(t),{\rm {\bf c}}(t),{\rm {\bf d}}(t)}
296: \right\}$. To avoid dominance of one eigenvector and divergence, we use the
297: Gram -- Schmidt orthogonalization and normalization of the vectors at each
298: step of the numerical integration. Now, at each point of the trajectory, all
299: possible linear combinations of $\left\{ {{\rm {\bf b}}(t),{\rm {\bf
300: c}}(t),{\rm {\bf d}}(t)} \right\}$ define a three-dimensional stable
301: subspace of perturbation vectors.
302:
303: To estimate an angle $\alpha $ between the one-dimensional unstable subspace
304: and the three-dimensional stable subspace we first construct a vector ${\rm
305: {\bf v}}(t)$ orthogonal to the three-dimensional subspace, with components
306: determined from a set of linear equations ${\rm {\bf v}}(t) \cdot {\rm {\bf
307: b}}(t) = 0$, ${\rm {\bf v}}(t) \cdot {\rm {\bf c}}(t) = 0$, ${\rm {\bf
308: v}}(t) \cdot {\rm {\bf d}}(t) = 0$. Then, we compute an angle
309: $\beta \in [0,\pi/2]$ between the vectors ${\rm {\bf v}}(t)$ and ${\rm {\bf a}}(t)$:
310: $\cos \beta = {\left| {{\rm {\bf v}}(t) \cdot {\rm {\bf c}}(t)} \right|}/
311: {\vert {\rm {\bf v}}(t)\vert \vert {\rm {\bf c}}(t)\vert }$
312: and set $\alpha = \beta - \pi/2$.
313:
314: \begin{figure}[htbp]
315: \begin{center}
316: \includegraphics[width=6.5in]{An4.eps}
317: \end{center}
318: \label{fig4}
319: \caption{Histograms for distributions of angles $\alpha $ between
320: the stable and unstable subspaces for the system (\ref{eq1}) with $\omega _0 = 2\pi
321: ,\,\,\,\varepsilon = 0.5$ obtained from computational procedure described in
322: the text: (a)$N = 10,\,\,A = 3$ and (b) $N = 4,\,\,A = 8$}
323: \end{figure}
324:
325: Figure 4 shows histogram for the distribution of angles $\alpha $ between
326: the stable and unstable subspaces for the system (\ref{eq1}) obtained from
327: computations at the two mentioned sets of parameter values. Observe clearly
328: visible separation of the distributions from zero values of $\alpha $. So,
329: the test confirms hyperbolicity of the attractors.
330:
331: In spite of simplicity of the presented example, I believe it is significant
332: as a feasible system, which may be designed as a physical device, e.g. on a
333: basis of two interacting electronic oscillators. It opens an opportunity for
334: experimental studies of hyperbolic chaos and its features predicted by the
335: mathematical hyperbolic theory (robustness, continuity of the invariant
336: measure, insensitivity of statistical characteristics of the motions in
337: respect to noise, etc.). In addition, it makes conclusive comparative
338: examination of dynamics of hyperbolic and non-hyperbolic systems.
339:
340: In a sense, breakthrough into the hyperbolic domain is a decisive step. Now
341: one can construct many other examples of systems with hyperbolic attractors:
342: Because of robustness of such attractor, any variation of
343: the right-hand parts of the equations will not destroy the hyperbolicity, at
344: least while they are not too large. Apparently, in this way it is possible to design
345: examples of autonomous systems with hyperbolic strange
346: attractorsb as well, via modification of the system supplimenting
347: additional equations for dynamical variables, which would represent the swing and the
348: auxiliary signals.
349:
350: The author thanks A. Pikovsky and M. Rosenblum for helpful discussion. The
351: work has been performed under partial support from RFBR (grant No 03-02-16192) and
352: from CRDF via the Research Educational Center of
353: Saratov University (Grant No. REC-006).
354:
355: \begin{thebibliography}{99}
356: \bibitem{1}
357: A. Katok and B. Hasselblatt, {\it Introduction to the Modern Theory of Dynamical
358: Systems} (Cambridge: Cambridge University Press, 1995).
359: \bibitem{2}
360: V. Afraimovich and S.-B. Hsu, {\it Lectures on chaotic dynamical systems}, AMS/IP
361: Studies in Advanced Mathematics, {\bf 28} (American Mathematical Society,
362: Providence, RI; International Press, Somerville, MA, 2003).
363: \bibitem{3}
364: R.L. Devaney, {\it An Introduction to Chaotic Dynamical Systems}, 2nd Edition
365: (Addison-Wesley: New York, 1989).
366: \bibitem{4}
367: E. Ott, {\it Chaos in Dynamical Systems}, 2nd Edition (Cambridge University Press,
368: 1993).
369: \bibitem{5}
370: V. S. Afraimovich and L.P. Shil'nikov, in {\it
371: Nonlinear Dynamics and Turbulence}, edited by G.I. Barenblatt, G. Ioss,
372: D.D. Joseph, 1 (Pitman, Boston, London, Melbourne, 1983);
373: S.E. Newhouse, Publ. Math. IHES {\bf 50}, 101 (1979).
374: \bibitem{6}
375: V.S. Anishchenko, V.V. Astakhov, A.B. Neiman, T.E. Vadivasova,
376: and L. Schimansky-Geier,
377: Nonlinear Dynamics of Chaotic and Stochastic Systems. Tutorial and Modern
378: Development (Springer, Berlin, Heidelberg, 2002).
379: \bibitem{7}
380: V.S. Afraimovich, V.V. Bykov and L.P. Shil'nikov, Dokl. Akad. Nauk SSSR, {\bf 234}, 336 (1977).
381: \bibitem{8}
382: T.J. Hunt and R.S. MacKay, Nonlinearity {\bf 16}, 1499 (2003).
383: \bibitem{9}
384: V. Belykh, I. Belykh and E. Mosekilde, Int. J. of Bifurcation and Chaos (in press).
385: \bibitem{10}
386: G. Benettin, L. Galgani, A. Giorgilli and J.M. Strelcyn, Meccanica {\bf 15}
387: (1980) 9.
388: \bibitem{11}
389: F. Christiansen and H. H. Rugh, Nonlinearity {\bf 10}, 1063 (1997).
390: \bibitem{12}
391: Y.-C. Lai, C. Grebogi, J.A. Yorke and I. Kan, Nonlinearity {\bf 6}, 779 (1993).
392: \bibitem{13}
393: V.S. Anishchenko, A.S. Kopeikin, J. Kurths, T.E. Vadivasova and
394: G.I. Strelkova, Physics Letters, {\bf A270}, 301 (2000).
395: \end{thebibliography}
396: \end{document}
397:
398: