nlin0608001/kr.tex
1:  
2: % Submitted to Dynamical Systems Int J 25 July 2006 
3: % Revised January 2007 and resubmitted June 2007 
4:   
5: \documentclass{cDSS2e}  
6: \renewcommand{\baselinestretch}{1.2}  
7: \usepackage{amsfonts}  
8: \usepackage{amsmath}  
9: \usepackage{xspace}  
10: %\usepackage{color,graphicx}  
11: \usepackage{epsf}  
12:   
13: \newcommand{\Rset}{\mathbb{R}}  
14: \newcommand{\sgn}{\mathop{\mathrm{sgn}}}  
15: \newcommand{\etal}{{\it et al.}\@\xspace}  
16: \newcommand{\smallin}{{\mbox{\footnotesize in}}}  
17: \newcommand{\smallout}{{\mbox{\footnotesize out}}}  
18: \newcommand{\smalli}{{\mbox{\footnotesize i}}}  
19: \newcommand{\smallr}{{\mbox{\footnotesize r}}}  
20: \newcommand{\smallu}{{\mbox{\footnotesize u}}}  
21: \newcommand{\smalls}{{\mbox{\footnotesize s}}}  
22: \newcommand{\smallPoincare}{{\mbox{\footnotesize Poincar\'e}}}  
23:   
24: \begin{document}  
25: \doi{10.1080/14689360xxxxxxxxxxxxx}  
26:  \issn{1468-9375}  
27: \issnp{1468-9367} \jvol{00} \jnum{00} \jyear{2007} \jmonth{June}  
28: \markboth{Kirk and Rucklidge}{Effect of symmetry breaking}  
29:   
30:   
31: \title{The effect of symmetry breaking on the dynamics near a structurally  
32: stable heteroclinic cycle between equilibria and a periodic orbit}  
33:   
34: \author{Vivien Kirk,\thanks{V.Kirk@auckland.ac.nz} 
35: Department of Mathematics, University of Auckland,  
36: \break Private Bag 92019, Auckland, New Zealand  
37: \break Alastair M. Rucklidge,\thanks{A.M.Rucklidge@leeds.ac.uk} 
38: Department of Applied Mathematics, University of Leeds,  
39: Leeds LS2 9JT, UK}  
40:   
41: \received{\today}  
42: \maketitle  
43:   
44: \begin{abstract}  
45: The effect of small forced symmetry breaking on the dynamics near a 
46: structurally stable heteroclinic cycle connecting two equilibria and a periodic 
47: orbit is investigated.  This type of system is known to exhibit complicated, 
48: possibly chaotic dynamics including irregular switching of sign of various 
49: phase space variables, but details of the mechanisms underlying the complicated 
50: dynamics have not previously been investigated.  We identify global 
51: bifurcations that induce the onset of chaotic dynamics and switching near a 
52: heteroclinic cycle of this type, and by construction and analysis of 
53: approximate return maps, locate the global bifurcations in parameter space. We 
54: find there is a threshold in the size of certain symmetry-breaking terms below 
55: which there can be no persistent switching. Our results are illustrated by a 
56: numerical example. 
57:  \end{abstract}  
58:   
59: \section{Introduction}  
60:   
61: It is well-established that the presence of symmetries in dynamical systems can
62: result in the existence of heteroclinic cycles that are structurally stable
63: with respect to symmetric perturbations~\cite{F80,GuHo88}. By {\em heteroclinic
64: cycle} we mean a collection of two or more flow invariant sets $\{\xi_1,\dots
65: ,\xi_n\}$ of some system of ordinary differential equations together with a set
66: of heteroclinic connections $\{\gamma_1(t), \dots,\gamma_n(t)\}$, where
67: $\gamma_j(t) \rightarrow \xi_{j}$ as $t \rightarrow -\infty$ and $\gamma_j(t)
68: \rightarrow \xi_{j+1}$ as $t \rightarrow +\infty$, and where $\xi_{n+1}\equiv
69: \xi_1$. In many studies, all the $\xi_i$ are equilibria, but in this paper we
70: explicitly consider the case that one of the $\xi_i$ is a periodic orbit. The
71: connections $\gamma_i$ may be isolated, or there may be a continuum of
72: connections from $\xi_i$ to $\xi_{i+1}$ for one or more $i$.
73:   
74: There is a large literature on structurally stable heteroclinic cycles (SSHC),  
75: including work establishing conditions for the existence and asymptotic  
76: stability of heteroclinic cycles~\cite{KrMe95,KrMe04,M91}, examination of the  
77: dynamics near heteroclinic cycles and networks of heteroclinic cycles  
78: \cite{KS94,AC98,ACL05,PD05}, and unfolding of bifurcations of heteroclinic  
79: cycles~\cite{SC92,CKMS,PD05b}. SSHC arise naturally in mathematical models of  
80: physical systems  with symmetry or near-symmetry~\cite{BuHe80,AGH88,PJ88,NTDX}. In  
81: these models, the physical system is idealised as having perfect symmetry,  
82: leading to the existence of invariant subspaces in the model and thus to the  
83: robustness of heteroclinic cycles with respect to symmetric perturbations. It  
84: is natural to ask how much of the dynamics observed in symmetric models  
85: persists under non-symmetric perturbations. Some effects of small  
86: symmetry-breaking have been documented~\cite{Me89,Ch93,SaSc95,MPR}, and aspects  
87: of the related question of how much of the dynamics persists under the  
88: inclusion of small noise have also been considered~\cite{StHo90,ASK03}, but  
89: details are likely to vary greatly between different examples. A few cases of 
90: experimental observation of near-heteroclinic cycles have been reported, 
91: most recently in  
92: \cite{NMQ}, but see also the references therein. In these cases, experimental noise and small 
93: symmetry-breaking effects prevent exact heteroclinic cycles from occurring, 
94: but there is clear evidence for near-heteroclinic structures in certain regimes. 
95:   
96: Our interest in the particular set-up explored in this paper is motivated by  
97: \cite{MPR}, which makes the observation that the addition of small  
98: symmetry-breaking terms to a system containing a heteroclinic cycle connecting  
99: two equilibria and a periodic orbit (as well as symmetric copies of the cycle)  
100: results in seemingly chaotic dynamics, with orbits passing near the various  
101: equilibria in the system repeatedly but in an irregular pattern, as illustrated  
102: in figure~\ref{fig:MPRexample}. A main point of~\cite{MPR} was to show that  
103: repeated switching of orbits in this manner could arise in a simple  
104: four-dimensional, nearly symmetric model, but the specific mechanisms  
105: underlying the complicated dynamics were not explored in detail.  
106:   
107:  %%%%%%%%%%%%%%%%  
108:  \begin{figure}  
109:  \centerline{\epsfxsize=0.9\hsize\epsfbox{kr_fig_mpr_time_series.eps}}  
110:  \caption{Irregular switching in the time series of a dynamo model studied 
111: in~\cite[figure~1]{MPR}. Panels~(a) and (b) show the evolution of different 
112: coordinates of the same trajectory, and panel~(c) shows the same coordinate as 
113: in~(b) over a longer time interval.} 
114:  \label{fig:MPRexample}  
115:  \end{figure}  
116:  %%%%%%%%%%%%%%%%%  
117:   
118: In this paper, we examine a generalisation of the situation from~\cite{MPR}, 
119: focusing on the structure and origin of chaotic dynamics in the system and on 
120: how switching dynamics is induced. Here and elsewhere in the paper, {\em 
121: switching} refers to the itinerary that an orbit follows under the dynamics. 
122: Specifically, in the fully symmetric version of our system there is a 
123: heteroclinic network consisting of four symmetric copies of the basic 
124: heteroclinic cycle. Invariance of various subspaces ensures that an orbit may 
125: make repeated passes near only one cycle. Once the symmetries are broken, 
126: however, an orbit may {\em switch}, i.e., make traversals near more than one of 
127: the original cycles (although, of course, the cycles themselves may not persist 
128: when the symmetry is broken). 
129:   
130: A main result of this paper is that in the case of small symmetry breaking,  
131: switching in one variable occurs when a complicated attractor arising from the  
132: presence of transverse homoclinic orbits of a periodic orbit crosses the stable  
133: manifold of one of the equilibria in the system. The existence of the  
134: transverse homoclinic orbits depends on a broken rotation 
135: symmetry, while the proximity of the attractor to the stable manifold of the  
136: equilibrium is caused by a broken reflection symmetry. Switching in a second  
137: variable results from the interaction between broken reflection symmetry and  
138: complicated dynamics associated with a heteroclinic bifurcation between the  
139: equilibria. Thus, switching results from the right combination of a global  
140: bifurcation and small symmetry breaking.  
141:   
142: A second significant result of this paper is the observation that there is a 
143: threshold for the size of symmetry breaking below which 
144: persistent switching cannot occur. More precisely, the existence of the  
145: heteroclinic cycle requires three separate symmetries to allow structurally  
146: stable connections within three invariant subspaces. We control the degree to  
147: which the three symmetries are broken by three small parameters, $\epsilon_1$,  
148: $\epsilon_2$ and~$\epsilon_3$; $\epsilon_1$~controls the degree  
149: to which the periodic orbit in the cycle deviates from a perfect circle, while  
150: $\epsilon_2$ and $\epsilon_3$ break reflection symmetries. For fixed small  
151: $\epsilon_2$ and~$\epsilon_3$, we find that there is a threshold  
152: in~$\epsilon_1$ for persistent switching to occur. For sufficiently  
153: small~$\epsilon_1$, there may be a single switch from one part of phase space  
154: to another, but it is only for $\epsilon_1$ beyond the threshold value that an 
155: orbit can repeatedly visit different parts of the phase space. We find that  
156: it is possible to get sustained switching in one or other or both of the  
157: variables associated with the reflection symmetries, and that the threshold  
158: values of~$\epsilon_1$ are different for switching in the two variables. The threshold does  
159: not go to zero as $\epsilon_2$ and~$\epsilon_3$ go to zero.  
160:   
161: Sustained switching of orbits near heteroclinic cycles and networks has been  
162: observed in a number of other settings. Clune and Knobloch~\cite{ClKn94}  
163: describe an example in which there are two symmetrically related copies of a  
164: non-asymptotically stable heteroclinic cycle, with nearby orbits making  
165: repeated passes near each cycle; no mechanism for the switching is suggested in  
166: this paper.  Aguiar et al.~\cite{ACL05} find switching near a hybrid  
167: heteroclinic network formed from transverse heteroclinic connections between  
168: equilibria and connections that are robust because of symmetry; switching seems  
169: to result from the folding and stretching caused by passage near the  
170: transversal heteroclinic connections and by mixing near an equilibrium solution  
171: with complex eigenvalues. Kirk et al.~\cite{KLS05} have an example of switching  
172: near a heteroclinic network that has no transversal connections; the switching  
173: is caused entirely by passage near an equilibrium with complex eigenvalues.  
174: Postlethwaite and Dawes~\cite{PD05} describe a variant of switching near a  
175: heteroclinic network in which each cycle in the network is unstable along a  
176: direction transverse to the cycle; orbits visit cycles in the network in a  
177: fixed order (being pushed away from each cycle in the transverse direction,  
178: which also happens to be the contracting direction for the next cycle) but the  
179: number of traversals of each cycle before switching to the next cycle can be  
180: constant or irregular. Ashwin et al.~\cite{ARS04a} describe switching  
181: associated with a stuck-on heteroclinic cycle between two invariant subspaces;  
182: here the switching is caused by a nonlinear mechanism that chooses between the different  
183: possibilities in a manner that is well modelled by a random process. Switching  
184: can also be induced by adding noise to a structurally stable heteroclinic  
185: network~\cite{ASK03}; noise sensitive switching has been observed  
186: by~\cite{RM95,MRWP96}. None of these examples explicitly considers symmetry  
187: breaking as a mechanism for switching.  
188:   
189: We adopt a standard approach to analysis of the system of interest, i.e., we  
190: set up a simple symmetric model in which there exists a heteroclinic cycle  
191: connecting two equilibria and a periodic orbit (Section~\ref{sec:description}),  
192: construct a return map that approximates the dynamics near such a cycle, and  
193: then add generic symmetry breaking terms to the return map  
194: (Section~\ref{sec:construct}, with details in the Appendix).  
195: Analysis of the return map is fruitful in cases  
196: where partial symmetry is retained, and allows us to prove the existence and  
197: asymptotic stability of periodic orbits, quasiperiodic solutions or  
198: heteroclinic cycles in various cases  
199: (Sections~\ref{sec:global}--\ref{sec:rotate}). In the completely asymmetric  
200: case, the return map is intractable, but we are able to make predictions about  
201: the dynamics by assuming there is a generic unfolding of the partially  
202: symmetric cases (Section~\ref{sec:full}). The example discussed in  
203: Section~\ref{sec:numerics} confirms and illustrates the analysis. Some  
204: conclusions are presented in Section~\ref{sec:conclusions}.  
205:   
206: A complicating factor in the analysis presented in this paper is that the  
207: unstable manifolds of one pair of equilibria and of the periodic orbit are  
208: two-dimensional, and there are continua of heteroclinic connections along some  
209: parts of the cycle in the fully symmetric case. Linearising about a single  
210: heteroclinic connection is not appropriate, and the usual method of analysis  
211: needs to be adapted to keep track of orbits in a neighbourhood of all the  
212: connections. Our approach is similar to that taken in~\cite{AC98,R01,KLS05}. We  
213: note that our analysis need not consider the issue of which connection from a  
214: continuum is selected by the dynamics (as investigated in, for instance,  
215: \cite{AC98,AFRS03,ARS04a,ARS04b}) since in our case breaking of the symmetries  
216: forces a discrete set of transversal connections to be selected from each  
217: continuum. Note also that some results about the dynamics near a heteroclinic 
218: cycle connecting an equilibrium and a periodic orbit in a generic (i.e., non-symmetric) 
219: setting are described in \cite{Rademacher, KROCK}, but the  
220: phenomena described  
221: in those papers  
222: will not be seen for small symmetry breaking in our setting, 
223: and is not the focus of our interest here. 
224:   
225: \section{Description of the problem}  
226: \label{sec:description}  
227:   
228: We consider a system of ordinary differential equations $\dot{\mathbf  
229: x}={\mathbf f}({\mathbf x})$ where ${\mathbf f}:\Rset^4\to\Rset^4$, and  
230: ${\mathbf x}=(x_1,y_1,x_2,x_3)\in\Rset^4$. It is sometimes convenient to use  
231: polar coordinates $(r_1,\theta_1)$ such that $z_1\equiv x_1+{\rm i}\, y_1\equiv  
232: r_1{\rm e}^{{\rm i}\theta_1}$. Initially, we assume the system is equivariant  
233: with respect to the action of a rotation and two reflections:  
234: $\kappa_i({\mathbf f}({\mathbf x}))={\mathbf f}(\kappa_i({\mathbf x}))$,  
235: $i=1,2,3$, where  
236:  \begin{alignat*}{2}  
237:  \kappa_1&\colon(z_1,x_2,x_3) & \to &(z_1{\rm e}^{{\rm i}\phi},x_2,x_3), \\  
238:  \kappa_2&\colon(z_1,x_2,x_3) & \to &(z_1,-x_2,x_3), \\  
239:  \kappa_3&\colon(z_1,x_2,x_3) & \to &(z_1,x_2,-x_3),  
240:  \end{alignat*}  
241: with $0\leq\phi<2\pi$.  
242: These symmetries generate the group $S^1\times Z_2\times Z_2$, and their 
243: presence ensures the existence  
244: of some dynamically invariant subspaces. We make the following assumptions  
245: about the dynamics in the subspaces, as illustrated in  
246: figure~\ref{fig:cyclepicture}:  
247:  \begin{itemize}  
248:  \item There exists a hyperbolic periodic orbit $P$ in the invariant  
249: plane $x_2=x_3=0$. Within this plane, the periodic orbit is a sink.  
250:  \item There exist hyperbolic, symmetry-related pairs of equilibria  
251: $\pm E_2$ and $\pm E_3$ on the invariant lines $z_1=0$, $x_3=0$ and $z_1=0$,  
252: $x_2=0$ respectively. Within these lines, the equilibria are sinks.  
253:  \item Within the invariant subspace $x_3=0$, $P$ is a saddle and  
254: $\pm E_2$ are sinks, and there are two-dimensional manifolds of heteroclinic  
255: connections from $P$ to $\pm E_2$ (figure~\ref{fig:cyclepicture}a).  
256:  \item Within the invariant subspace $z_1=0$, $\pm E_2$ are saddles  
257: and $\pm E_3$ are sinks, and there are one-dimensional heteroclinic  
258: connections from $+E_2$ to~$\pm E_3$, and from $-E_2$ to~$\pm E_3$  
259: (figure~\ref{fig:cyclepicture}b).  
260:  \item Within the invariant subspace $x_2=0$, $\pm E_3$ are saddles  
261: and $P$ is a sink, and there are two-dimensional manifolds of heteroclinic  
262: connections from $\pm E_3$ to $P$ (figure~\ref{fig:cyclepicture}c).  
263:  \end{itemize}  
264:   
265:  %%%%%%%%%%%%%%%%  
266: \begin{figure}  
267:  \centerline{\makebox[0.3\hsize][l]{(a)}  
268:              \hspace{0.03\hsize}  
269:              \makebox[0.3\hsize][l]{(b)} \hspace{0.03\hsize}  
270:              \makebox[0.3\hsize][l]{(c)}}  
271:  \vspace{0.3ex}  
272:  \centerline{\epsfxsize=0.3\hsize\epsfbox{kr1_het_cycle_a.eps}  
273:              \hspace{0.03\hsize}  
274:              \epsfxsize=0.3\hsize\epsfbox{kr1_het_cycle_b.eps}  
275:              \hspace{0.03\hsize}  
276:              \epsfxsize=0.3\hsize\epsfbox{kr1_het_cycle_c.eps}}  
277:  \caption{The heteroclinic cycle for the fully symmetric system.  
278:  (a)~One of the connections in the $x_3=0$ subspace, from the periodic orbit  
279:  $P$ to the equilibrium point~$+E_2$;  
280:  (b)~the single connection in the $x_1=y_1=0$ subspace, between the equilibria  
281:  $+E_2$ and~$+E_3$;  
282:  (c)~one of the connections in the $x_2=0$ subspace, from $+E_3$ to $P$.}  
283: \label{fig:cyclepicture}  
284: \end{figure}  
285:  %%%%%%%%%%%%%%%%%%%%  
286:   
287: In the presence of the rotation symmetry $\kappa_1$, the coordinate $\theta_1$  
288: decouples from the other coordinates, leaving an equivalent three-dimensional  
289: system containing a SSHC connecting three saddle-type equilibria. This cycle  
290: may be asymptotically stable, depending on the eigenvalues at the three  
291: equilibria~\cite{KrMe95}. The behaviour of trajectories near such a  
292: heteroclinic cycle is well understood, with a typical orbit passing near each  
293: of the equilibria in a cyclic manner, spending ever increasing periods of time  
294: near each equilibrium. The dynamics in the fully symmetric, four-dimensional  
295: problem therefore has analogous behaviour: trajectories cycle between two  
296: equilibria and a periodic orbit, with the time spent near each equilibrium or  
297: the periodic orbit increasing with each subsequent traversal of the  
298: cycle~\cite{MPR}. Moving to four dimensions does more than replace one 
299: pair of equilibria  
300: by a periodic orbit: it also introduces dynamical features that  
301: will be important once symmetry is broken. In particular, as can be seen in  
302: figure~\ref{fig:cyclepicture}, $\pm E_2$ and $\pm E_3$ are saddle-foci in the  
303: four-dimensional problem, and $P$ and $\pm E_3$ have two-dimensional unstable manifolds.  
304:   
305: A detailed analysis of the effect on the dynamics of small symmetry breaking is  
306: performed in the following sections; here we describe some geometric effects.  
307: Since $\pm E_2$, $\pm E_3$ and $P$ are assumed to be hyperbolic in the fully  
308: symmetric case, they persist and are hyperbolic when sufficiently small  
309: symmetry breaking terms are added. However, $+E_2$ and $-E_2$ will generically  
310: move off the $x_2$-axis and will no longer be related to each other by  
311: symmetry. Generic symmetry breaking will have an analogous effect on $+E_3$ and  
312: $-E_3$, and will also break the circular symmetry of $P$ and move it off  
313: the plane $x_2=x_3=0$.  
314:   
315: Sufficiently small symmetry breaking will not change the dimensions of the  
316: stable and unstable manifolds of $\pm E_2$, $\pm E_3$ and $P$, but it will  
317: destroy the invariant subspaces, and the heteroclinic connections that  
318: existed in the subspaces will either cease to exist or change their nature. We  
319: consider the geometric effect of symmetry breaking on each of the former  
320: heteroclinic connections in turn.  
321:   
322: The heteroclinic connections from $\pm E_2$ to $\pm E_3$ require the 
323: coincidence in $\Rset^4$ of one-dimensional and two-dimensional manifolds; 
324: these connections will be destroyed by a generic symmetry-breaking 
325: perturbation. 
326:   
327: The heteroclinic connections from $\pm E_3$ to $P$ occur when the 
328: two-dimensional unstable manifolds of $\pm E_3$ intersect the three-dimensional 
329: stable manifold of $P$. Depending on the perturbation, we generically expect to 
330: see either transversal intersections between these manifolds (in which case 
331: there are, for example, at least two robust heteroclinic connections from 
332: $+E_3$ to $P$) or no intersections of the manifolds. The special case where the 
333: manifolds are tangent can also occur in a codimension-one way. In the case of 
334: transversal intersections of manifolds, we might expect to see heteroclinic 
335: tangles and the associated complicated dynamics, depending on whether the 
336: dynamics elsewhere in the phase space permits reinjection of trajectories into 
337: the neighbourhood of the transversal intersections. 
338:   
339: The heteroclinic connections from $P$ to $\pm E_2$ occur when the 
340: two-dimensional unstable manifold of $P$ intersects the three-dimensional 
341: stable manifolds of $\pm E_2$. There is a clear analogy with the case of 
342: connections from $\pm E_3$ to $P$ and the comments about that case apply 
343: equally here. 
344:   
345: While small symmetry-breaking terms generically destroy the heteroclinic  
346: cycle, there will still be an attractor lying close to the original  
347: heteroclinic cycle (Melbourne~\cite{Me89} shows this in a closely related  
348: case). We show below that the form of this attractor (e.g., periodic,  
349: quasiperiodic, chaotic) depends on the nature of the symmetry-breaking  
350: perturbations included. In the fully symmetric case, the invariant  
351: subspaces defined by $x_2=0$ and by $x_3=0$  
352: restrict each trajectory to one quarter of the phase space, but 
353: once the reflection symmetries are  
354: broken, a single trajectory may explore more of the phase  
355: space. We are interested in determining the circumstances under which  
356: trajectories exhibit switching, i.e., make passages near two or more quarters  
357: of the original heteroclinic attractor.  
358:  
359: \section{Construction of return maps}  
360:  \label{sec:construct}  
361:   
362: We construct and analyse a return  
363: map that approximates the dynamics near the cycle. The idea is to define local coordinates  
364: and cross-sections near $\pm E_2$, $\pm E_3$ and $P$, then determine local maps  
365: valid in a neighbourhood of each of $\pm E_2$, $\pm E_3$ and $P$, and global  
366: maps valid in a neighbourhood of each heteroclinic connection. Composing the  
367: local and global maps yields the desired return map. Different forms for the return 
368: map are obtained depending on which of the symmetries are broken. In this section  
369: we list the different cases, but details of map construction are left to the Appendix. 
370: The techniques used are, for the most part, standard, although modifications are 
371: required to allow for the existence of continua of heteroclinic connections along some 
372: parts of the cycle in the fully symmetric case. 
373:  
374: Throughout, we use a small  
375: parameter~$h$ to control the size of the local neighbourhoods ($h\ll1$), and  
376: small parameters~$\epsilon_1$, $\epsilon_2$, $\epsilon_3$ to control the extent  
377: to which the symmetries $\kappa_1$, $\kappa_2$, $\kappa_3$ are broken.  
378: It turns out to be convenient to define the return map on a cross-section near $+E_3$. 
379: Using local coordinates $(r_1, \theta_1, x_2, \xi_3)$ near $+E_3$, where coordinates 
380: are chosen so that $+E_3$ is at the origin and so that the eigenvectors of the linearised 
381: flow align with the coordinate axes in the manner described in the Appendix, we define 
382: a cross-section  
383: \begin{equation} 
384:  H_3^\smallin  = \left\{(r_1, \theta_1,x_2,\xi_3): 0\leq r_1\leq h,  
385:                                            |x_2|=h,  
386:                                            |\xi_3|\leq h \right\}  \nonumber 
387: \end{equation}  
388: and then compute the return map, $R :   H_3^\smallin \to  H_3^\smallin$. The same 
389: cross-section works equally well near $-E_3$ and the maps $R$ we compute in fact approximate  
390: the dynamics near any of the four possible paths from $\pm E_3$ to $\pm E_3$. See the 
391: Appendix for details. 
392:  
393: Since we are interested in trajectories that switch between positive and 
394: negative values of $x_2$ and $x_3$, we introduce the notation $\pm_2$ and 
395: $\pm_3$ to indicate whether a trajectory visits $+E_2$ or $-E_2$, and $+E_3$ or 
396: $-E_3$. In particular, the trajectory starts at one of four possible sections 
397: specified by $H_3^\smallin$, and we use $\pm_2$ to specify whether $x_2=+h$ or 
398: $x_2=-h$ (implying that the trajectory recently visited $+E_2$ or $-E_2$). We 
399: use $\pm_3$ to specify whether the trajectory is close to $+E_3$ or $-E_3$. 
400: When the trajectory next returns to $H_3^\smallin$, we will be interested in 
401: whether it visited $+E_2$ or $-E_2$ en route, and whether it returns to $+E_3$ 
402: or~$-E_3$. 
403:   
404: First, in the case with full symmetry ($\epsilon_1=\epsilon_2=\epsilon_3=0$),  
405: we have:  
406:  \begin{align}  
407:  R(r_1,\theta_1,x_2=\pm_2h,\xi_3)=\Big(&  
408:            \tilde{r}_1=A r_1^{\delta},  
409:           \tilde{\theta}_1 = \theta_1+\Phi-Q \ln r_1, \nonumber \\  
410:           &\tilde{x}_2=x_2, \tilde{\xi}_3=B_2 \Big),  
411:  \label{eq:R_nosb}  
412:  \end{align}  
413:  where $A>0$ and $\Phi$ are constants, $\delta=\delta_1\delta_2\delta_3$, 
414: $Q=(e_1e_2+e_2c_3+c_3c_1)/e_1e_2e_3$, and the constants $\delta_i$, $e_i$ and $c_i$ 
415: are defined in the Appendix. If $x_3>0$ initially, the trajectory 
416: returns to $+E_3$ after visiting $\pm_2E_2$; if $x_3<0$ initially, the 
417: trajectory returns to~$-E_3$. 
418:  
419: Second, breaking the $\kappa_2$ and $\kappa_3$ symmetries ($\epsilon_1=0$,  
420: $\epsilon_2\ne 0$, $\epsilon_3\neq0$) we have:  
421:  \begin{align}  
422:  R(r_1,\theta_1,x_2=\pm_2h,\xi_3)=\Big(&  
423:            \tilde{r}_1=A_2\Big| \epsilon_3 \pm_3 A_1  
424:                           \big| \epsilon_2 \pm_2 A_3 r_1^{\delta_3}  
425:                                 \big|^{\delta_1}  
426:                                 \Big|^{\delta_2}, \nonumber \\  
427:            &\tilde{\theta}_1 = \theta_1 + \Phi_1+\Phi_2+\Phi_3  
428:                    -\frac{1}{e_3}\ln r_1\nonumber\\  
429:            &\phantom{\tilde{\theta}_1 = {}}{}  
430:                    -\frac{1}{e_1}\ln\big|\epsilon_2\pm_2A_3r_1^{\delta_3}\big|\nonumber\\  
431:            &\phantom{\tilde{\theta}_1 = {}}{}  
432:                    -\frac{1}{e_2}\ln\Big| \epsilon_3 \pm_3 A_1    
433:                           \big| \epsilon_2 \pm_2 A_3 r_1^{\delta_3}  
434:                                 \big|^{\delta_1}  
435:                                 \Big|,\nonumber \\  
436:           & \tilde{x}_2=\sgn\left(\epsilon_2\pm_2A_3r_1^{\delta_3}\right)h,  
437:             \tilde{\xi}_3= B_2 \Big).  
438:  \label{eq:R_sb23}  
439:  \end{align}  
440: The trajectory visits $+E_2$ or $-E_2$ en route according to the sign of 
441: $\epsilon_2\pm_2A_3r_1^{\delta_3}$, and it returns to $+E_3$ or $-E_3$ 
442: according to the sign of 
443: $\epsilon_3\pm_3A_1\big|\epsilon_2\pm_2A_3r_1^{\delta_3}\big|^{\delta_1}$. 
444:  
445: Third, if we break the $\kappa_1$ symmetry but preserve $\kappa_2$  
446: and~$\kappa_3$ ($\epsilon_1\neq0$, $\epsilon_2=\epsilon_3=0$) we have:  
447:  \begin{align}  
448:  R(r_1,\theta_1,x_2=\pm_2h,\xi_3)=\Big(&  
449:            \tilde{x}_1=\epsilon_1 a_\smallr+A_2|\hat{x}_3|^{\delta_2}  
450:                 \cos\hat{\theta}_1,  
451:            \nonumber \\  
452:            &\tilde{y}_1=\epsilon_1 a_\smalli+A_2|\hat{x}_3|^{\delta_2}  
453:                 \sin\hat{\theta}_1,  
454:            \nonumber\\  
455:           &\tilde{x}_2 = \sgn(\hat{x}_2)h,  
456:            \tilde{\xi}_3= B_2 \Big),  
457:  \label{eq:R_sb1}  
458:  \end{align}  
459: where  
460:  \begin{align*}  
461:  \hat{x}_2&=\pm_2\left(A_3 +  
462:                        \epsilon_1f_3\left(\theta_1-\frac{1}{e_3}\ln r_1\right)  
463:                  \right)r_1^{\delta_3},\\  
464:  \hat{x}_3&=\pm_3\left(A_1 +  
465:                        \epsilon_1f_1\left(\theta_1+\Phi_3-\frac{1}{e_3}\ln r_1  
466:                                                          -\frac{1}{e_1}\ln|\hat{x}_2|  
467:                                     \right)  
468:                  \right)|\hat{x}_2|^{\delta_1},\\  
469:  \hat{\theta}_1&=\theta_1+\Phi_1+\Phi_2+\Phi_3  
470:                  -\frac{1}{e_3}\ln r_1  
471:                  -\frac{1}{e_1}\ln|\hat{x}_2|  
472:                  -\frac{1}{e_2}\ln|\hat{x}_3|\,.  
473:  \end{align*}   
474: The trajectory visits $+E_2$ or $-E_2$ en route according to the sign
475: of~$\hat{x}_2$, and it returns to $+E_3$ or $-E_3$ according to the sign
476: of~$\hat{x}_3$. In this case, these signs are the same as the signs of $x_2$
477: and~$x_3$.
478: The map~(\ref{eq:R_sb1}) can be  
479: simplified by assuming that $A_3$ and $A_1$ are order one and dropping the  
480: terms proportional to $\epsilon_1$ in the expressions for $\hat{x}_2$  
481: and~$\hat{x}_3$. This results in an approximate map:  
482:  \begin{align}  
483:  R(r_1,\theta_1,x_2=\pm_2h,\xi_3)=\Big(&  
484:            \tilde{x}_1=\epsilon_1 a_\smallr+Ar_1^{\delta}  
485:                 \cos\left(\theta_1+\Phi - Q\ln r_1\right),  
486:            \nonumber \\  
487:            &\tilde{y}_1=\epsilon_1 a_\smalli+Ar_1^{\delta}  
488:                 \sin\left(\theta_1+\Phi - Q\ln r_1\right),  
489:            \nonumber\\  
490:           &\tilde{x}_2 = x_2,  
491:            \tilde{\xi}_3= B_2 \Big),  
492:  \label{eq:R_sb1simple}  
493:  \end{align}  
494:  where $\delta$ and $Q$ were defined above, and $A$ and $\Phi$ are constants as  
495: in equation~(\ref{eq:R_nosb}).  
496:   
497: Finally, when all symmetries are broken the return map is similar to the  
498: map~(\ref{eq:R_sb1}) above, though with definitions of $\hat{x}_2$ and  
499: $\hat{x}_3$ that include terms proportional to $\epsilon_2$ and $\epsilon_3$:  
500:  \begin{align}  
501:  \hat{x}_2&=\pm_2\left(A_3 +  
502:                        \epsilon_1f_3\left(\theta_1-\frac{1}{e_3}\ln r_1\right)  
503:                  \right)r_1^{\delta_3} \nonumber \\ 
504:           &\phantom{=}{}+\epsilon_2\left(1+\epsilon_1g_3\left(\theta_1-\frac{1}{e_3}\ln r_1\right)  
505:                       \right), \nonumber \\  
506:  \hat{x}_3&=\pm_3\left(A_1 +  
507:                        \epsilon_1f_1\left(\theta_1+\Phi_3-\frac{1}{e_3}\ln r_1  
508:                                                          -\frac{1}{e_1}\ln|\hat{x}_2|  
509:                                     \right)  
510:                  \right)|\hat{x}_2|^{\delta_1}\nonumber \\  
511:           &\phantom{=}{}+\epsilon_3\left(1 +  
512:                             \epsilon_1g_1\left(\theta_1+\Phi_3-\frac{1}{e_3}\ln r_1  
513:                                                -\frac{1}{e_1}\ln|\hat{x}_2|  
514:                                          \right)  
515:                       \right), \nonumber \\  
516:  \hat{\theta}_1&=\theta_1+\Phi_1+\Phi_2+\Phi_3  
517:                  -\frac{1}{e_3}\ln r_1  
518:                  -\frac{1}{e_1}\ln|\hat{x}_2|  
519:                  -\frac{1}{e_2}\ln|\hat{x}_3|\,.  
520:  \label{eq:R_sb1complex}  
521:  \end{align}  
522:  The trajectory visits $+E_2$ or $-E_2$ en route according to the sign 
523: of~$\hat{x}_2$, and it returns to $+E_3$ or $-E_3$ according to the sign 
524: of~$\hat{x}_3$. It might seem that terms proportional to~$\epsilon_1$ in 
525: $\hat{x}_2$ and $\hat{x}_3$ could be dropped, as they were above. However, the 
526: terms $\pm_2A_3r_1^{\delta_3}$ and $\epsilon_2$ could nearly cancel and 
527: likewise $\pm_3A_1|\hat{x}_2|^{\delta_1}$ and $\epsilon_3$, so we do not drop 
528: the $\epsilon_1$~terms. In fact, it turns out that retaining the $\epsilon_1$ 
529: terms is essential for understanding the switching mechanisms. 
530:   
531: It is possible to write down equivalent maps from $H_1^\smallin\to  
532: H_1^\smallin$ and $H_2^\smallin\to H_2^\smallin$. Note that the radial  
533: coordinates (as defined in the Appendix)  
534: play no role in the return maps, at the order to which we are  
535: working.  
536:   
537:   
538: \section{Analysis of return maps}  
539: \label{sec:analysis} 
540:  
541: Behaviour in the case without symmetry breaking is well understood and simple:  
542: whenever $\delta>1$ and $r$~is small, iteration of map (\ref{eq:R_nosb})  
543: results in progressively smaller values of~$r$ and so there is an  
544: asymptotically stable heteroclinic cycle. The signs of $x_2$ and $x_3$ cannot  
545: change, owing to the presence of invariant subspaces, so each trajectory is  
546: confined to one quarter of the phase space. For the remainder of this section,  
547: we will assume~$\delta>1$.  
548:   
549: \subsection{Global bifurcations}  
550: \label{sec:global}  
551:   
552: Global bifurcations are a key ingredient for understanding the dynamics of the  
553: non-symmetric system. In this section, we describe the global bifurcations that  
554: are most important for our analysis.  
555:   
556: \subsubsection{Homoclinic bifurcation of $P$}  
557: \label{homP}  
558:   
559: The periodic orbit $P$ has stable and unstable manifolds of dimension three and  
560: two, respectively, meaning that transverse intersections of the manifolds, when  
561: they occur, do so in a codimension-zero way, while tangencies between the  
562: manifolds will be of codimension one. Transverse homoclinic orbits can only  
563: occur when all symmetries are broken, as the following argument shows. If  
564: $\epsilon_2=0$, the subspace $x_2=0$ is invariant; since $ {\cal W}^\smalls(P)$ lies in that  
565: subspace it cannot intersect $ {\cal W}^\smallu(P)$. Similarly, if $\epsilon_3=0$, the  
566: subspace $x_3=0$ is invariant; since $ {\cal W}^\smallu(P)$ lies in that subspace it cannot  
567: intersect $ {\cal W}^\smalls(P)$. If $\epsilon_1=0$ then the rotation symmetry ensures that  
568: any intersection of $ {\cal W}^\smallu(P)$ and $ {\cal W}^\smalls(P)$ will not be transverse.  
569:   
570: In the case $\epsilon_1=0$, $\epsilon_2\ne 0$, $\epsilon_3 \ne 0$,  
571: non-trans\-versal homoclinic orbits of $P$ occur when one branch of the stable  
572: manifold of $P$ is coincident with one branch of the unstable manifold of $P$.  
573: This event can be located by calculating the image of  
574: $ {\cal W}^\smallu(P)$ under $\Psi_{31}\circ \phi_3 \circ \Psi_{23} \circ \phi_2 \circ \Psi_{12}$  
575: (see Appendix for definitions of the maps $\phi_i$ and $\Psi_{ij}$)  
576: and setting the $x_2$ component  
577: of the image to zero; we find that for small symmetry-breaking,  
578: non-trans\-versal homoclinic bifurcations of $P$ occur at  
579:  \begin{equation}  
580:  \epsilon_2 = -\pm_2 A_3 A_2^{\delta_3}|\epsilon_3|^{\delta_2 \delta_3},  
581:     \qquad \epsilon_1=0.  
582:     \label{eq:homP}  
583:  \end{equation}  
584: Homoclinic orbits can be formed by coincidence of either of the two branches  
585: of $ {\cal W}^\smallu(P)$ with either of the two branches of $ {\cal W}^\smalls(P)$, resulting in four  
586: possible homoclinic bifurcations corresponding to the four separate curves  
587: implicit in the expression above. These curves are shown as dashed lines in  
588: figure~\ref{fig:globalbifns}.  
589: The homoclinic orbit corresponding to the curve in the second quadrant of the
590: $(\epsilon_2,\epsilon_3)$ plane arises from the choice $\pm_2=+$ and
591: $\epsilon_3>0$, and passes close to $+E_2$ and $+E_3$; the three other
592: bifurcation curves correspond to homoclinic orbits with the three other routes
593: past the equilibria, in the obvious way.
594:  
595:   
596:  %%%%%%%%%%%%%  
597:  \begin{figure}  
598:  \centerline{\epsfxsize=0.5\hsize\epsfbox{globalbifns.eps}}  
599: \caption{Schematic bifurcation set for the case $\epsilon_1=0$, $\epsilon_2$  
600: and $\epsilon_3$ small. Regions of asymptotically stable quasiperiodic  
601: solutions are bounded by codimension-one curves of global bifurcations, i.e.,  
602: non-trans\-verse homoclinic bifurcations of $P$ (dashed curves) and  
603: heteroclinic bifurcations of the cycles $\pm E_2 \to \pm E_3 \to \pm E_2$  
604: (solid curves). The shapes of the global bifurcation curves correspond to the  
605: choice $\delta_1>1$, $\delta_2>1$ and $\delta_3>1$, but similar figures could  
606: be drawn for the other cases. As explained in Section~\ref{sec:reflect}, the  
607: various shading styles indicate the regions in which four different  
608: quasiperiodic solutions occur. Close to the $\epsilon_2$ and $\epsilon_3$ axes,  
609: two different quasiperiodic solutions coexist.}  
610:  \label{fig:globalbifns}  
611:  \end{figure}  
612:  %%%%%%%%%%%%%  
613:   
614: As $\epsilon_1$ changes from zero, each curve of  
615: non-trans\-versal homoclinic bifurcations will generically split into two  
616: curves of homoclinic tangencies, with the region between the tangencies being  
617: parameter values for which there are transverse homoclinic orbits of~$P$. Four  
618: curves of homoclinic tangencies and two regions of homoclinic tangles are shown  
619: schematically in figure~\ref{fig:globalbifns-sb}.  
620:   
621: Inspection of the expression for the $x_2$~component of the image of  
622: $ {\cal W}^\smallu(P)$ under $\Psi_{31}\circ \phi_3 \circ \Psi_{23} \circ \phi_2 \circ \Psi_{12}$ 
623: gives more information about loci  
624: of the homoclinic bifurcations of~$P$ when $\epsilon_1\neq0$. This component  
625: can be written as:  
626:  \begin{equation}  
627:  \tilde{x}_2=\pm_2 R_1^{\delta_3} \left(A_3 + \epsilon_1 f_3(\Theta_1)\right)  
628:              + \epsilon_2 \left(1+\epsilon_1 g_3(\Theta_1)\right),  
629:  \label{eq:thresholdx2}  
630:  \end{equation}  
631: where $R_1$ and $\Theta_1$ are complicated functions of the coefficients and  
632: parameters. In this expression, $A_3 + \epsilon_1 f_3(\Theta_1)$ must remain  
633: positive, as explained in the Appendix, and $R_1$ is positive. Expressions for the positions  
634: of the homoclinic tangencies in parameter space can be calculated by setting  
635: $\tilde{x}_2=0$; these expressions are not included here due to their extreme  
636: ugliness. Nonetheless, we note that for $\epsilon_1$ small, when $\pm_2=+$,  
637: there are only solutions with~$\epsilon_2<0$; this is consistent with figures  
638: \ref{fig:globalbifns} and \ref{fig:globalbifns-sb}, in which each bifurcation  
639: curve is confined to a single quadrant. However, if $\epsilon_1$ is large  
640: enough that $1+\epsilon_1 g_3(\Theta_1)$ can change sign as $\Theta_1$ varies,  
641: the loci of the homoclinic bifurcations of $P$ can change quadrants. Of course,  
642: this effect is outside the range of validity of the return maps we have  
643: constructed, but the principle is worth bearing in mind as it appears to  
644: influence the dynamics observed in the numerical example discussed in  
645: Section~\ref{sec:numerics}.  
646:   
647:  %%%%%%%%%%%%%  
648: \begin{figure}  
649: \centerline{\epsfxsize=0.5\hsize\epsfbox{new-globalbifns-sb.eps}}  
650:  %\includegraphics[width=0.5\hsize]{globalbifns-sb}  
651: \caption{Schematic diagram showing part of the bifurcation set for the case  
652: $\epsilon_1$ fixed and non-zero but small (compare with  
653: figure~\ref{fig:globalbifns}). Dashed curves correspond to homoclinic  
654: tangencies of $P$, solid curves in the first (resp.~fourth) quadrant correspond  
655: to heteroclinic tangencies between $ {\cal W}^\smallu(-E_3)$ (resp.~${\cal W}^\smallu(+E_3)$) and  
656: ${\cal W}^\smalls(+E_2)$, and the shading shows regions in which the corresponding 
657: homoclinic or heteroclinic  
658: tangles exist. The dotted horizontal line indicates a path through parameter space  
659: discussed in Section~\ref{sec:full}; the labels A -- D indicate schematically 
660: parameter 
661: values used in section~\ref{sec:numerics}.}  
662:  \label{fig:globalbifns-sb}  
663:  \end{figure}  
664:  %%%%%%%%%%%%%  
665:   
666:   
667: \subsubsection{Heteroclinic bifurcation $\pm E_2 \to \pm E_3 \to \pm E_2$}  
668: \label{hetEs}  
669:   
670: In the case that all symmetries are broken, consideration of the dimensions of  
671: the stable and unstable manifolds of the equilibrium points shows that the  
672: heteroclinic cycle $+E_2 \to+E_3 \to+E_2$ will occur in a codimension-two  
673: manner. However, if $\epsilon_1=0$, the connection $+E_2\to+E_3$ is robust and  
674: the intersection of ${\cal W}^\smallu(+E_3)$ and ${\cal W}^\smalls(+E_2)$ is a codimension-one phenomenon,  
675: meaning that the heteroclinic cycle as a whole occurs with codimension one.  
676: This latter case is of interest since, as we will see, the heteroclinic  
677: bifurcation unfolds when $\epsilon_1 \ne 0$ into homoclinic bifurcations of  
678: $+E_2$ and $+E_3$ and heteroclinic tangencies between ${\cal W}^\smallu(+E_3)$  
679: and ${\cal W}^\smalls(+E_2)$  
680: similar to the way each non-trans\-verse homoclinic bifurcation of $P$ splits  
681: into two homoclinic tangencies when $\epsilon_1$ is varied from zero (see  
682: above).  An analogous argument works for heteroclinic cycles involving $-E_2$  
683: and/or $-E_3$.  
684:   
685: Calculations with the local and global maps yields an expression for the  
686: parameter values at which these heteroclinic bifurcations occur:  
687:  $$ \epsilon_3 = - \pm_3A_1|\epsilon_2|^{\delta_1},  
688:     \qquad\epsilon_1=0.  
689:  $$  
690: See figure~\ref{fig:globalbifns}. This expression is valid for all four cycles  
691: $\pm E_2 \to \pm E_3 \to \pm E_2$ so long as $\pm_3$ and the sign of  
692: $\epsilon_2$ are chosen appropriately.  
693:   
694: \subsubsection{Homoclinic bifurcations of $\pm E_2$ and $\pm E_3$}  
695:   
696: The dimensions of the stable and unstable manifolds of $\pm E_2$ and $\pm E_3$ 
697: are such that if homoclinic bifurcations of these equilibria occur, they are of 
698: codimension one. 
699:   
700: An argument similar to that used in subsection~\ref{homP} shows that we require  
701: $\epsilon_1 \ne 0$ and $\epsilon_3 \ne 0$ if a homoclinic bifurcation of $\pm E_2$  
702: is to occur, although $\epsilon_2$ could be zero. Similarly, existence of a  
703: homoclinic bifurcation of $\pm E_3$ requires $\epsilon_1\ne0$ and $\epsilon_2\ne0$,  
704: although $\epsilon_3$ could be zero. The homoclinic bifurcations of $\pm E_2$  
705: (resp.~$\pm E_3$) will be of Shil'nikov type if $\delta_2<1$ (resp.~$\delta_3 <1$)  
706: and if $c_2 < 2$ (resp.~$c_3 <2$)~\cite{GS84}.  
707:   
708: We can in principle calculate parameter values at which these homoclinic  
709: bifurcations occur, but the expressions are too nasty to be useful. Instead, we  
710: note that there can be two homoclinic bifurcations of $+E_2$, one for each  
711: branch of the unstable manifold of $+E_2$, and a further two homoclinic  
712: bifurcations of $-E_2$. Similarly, there can be two homoclinic bifurcations of  
713: $+E_3$ and two homoclinic bifurcations of $-E_3$. These eight homoclinic  
714: bifurcations will in general occur at different parameter values, but in the  
715: limit $\epsilon_1 \to 0$, will converge pairwise on the loci of the four  
716: heteroclinic bifurcations involving $\pm E_2$ and $\pm E_3$ discussed in the  
717: previous subsection. For instance, as $\epsilon_1 \to 0$, a homoclinic orbit of  
718: $+E_2$ passing near $-E_3$ and a homoclinic orbit of $-E_3$ passing near $+E_2$  
719: will converge in phase space on the heteroclinic cycle $+E_2\to-E_3\to+E_2$,  
720: and the parameter values at which the homoclinic bifurcations occur will  
721: converge in parameter space on the locus of the heteroclinic bifurcation. For  
722: clarity, these bifurcation curves are not shown in  
723: figure~\ref{fig:globalbifns-sb}.  
724:   
725: The dynamics associated with these bifurcations will be  
726: discussed further below.  
727:   
728: \subsection{Breaking the two reflection symmetries}  
729:  \label{sec:reflect}  
730:   
731: Here we show that, for $\epsilon_1=0$ and for sufficiently  
732: small~$\epsilon_2$ and~$\epsilon_3$, map~(\ref{eq:R_sb23}) generically has  
733: at least one asymptotically stable closed invariant curve and the corresponding  
734: flow has quasiperiodic solutions. This is not a surprising result, since  
735: the coordinate $\theta_1$ decouples from the other coordinates 
736: when $\epsilon_1=0$, in which case 
737: our system can be reduced to a three-dimensional system with a SSHC between 
738: equilibria; earlier work on a system related to our reduced system showed that breaking the reflection 
739: symmetries can give rise to asymptotically stable periodic solutions~\cite{SaSc95}. 
740: Our main aim in this section is to locate the regions in parameter space in which the 
741: quasiperiodic solutions exist, for comparison with the location of some of the global 
742: bifurcations described in section~\ref{sec:global}. 
743:   
744: The $r_1$ component of map~(\ref{eq:R_sb23}) is independent of the other 
745: variables, and so we first seek values of $r_1$ for which $F(r_1)=r_1$,  where  
746:  \begin{equation}  
747:  F(r_1)=A_2\Big| \epsilon_3 \pm_3 A_1  
748:                           \big| \epsilon_2 \pm_2 A_3 r_1^{\delta_3}  
749:                                 \big|^{\delta_1}  
750:                                 \Big|^{\delta_2}. \label{eq:f_sb23}  
751:  \end{equation}  
752: For each choice of $\epsilon_2$ and $\epsilon_3$, there are two  
753: possible signs of each of $\pm_2$ and~$\pm_3$, but the case $(\pm_2=+,\epsilon_2>0)$ is  
754: equivalent to $(\pm_2=-,\epsilon_2<0)$, and the case $(\pm_3=+,\epsilon_3>0)$  
755: is equivalent to $(\pm_3=-,\epsilon_3<0)$. Without loss of generality, we  
756: focus on the case $\pm_2=+$, $\pm_3=+$, and seek values of  
757: $\epsilon_2$ and $\epsilon_3$ for which there exist fixed points of  
758: map (\ref{eq:f_sb23}). Fixed points of this type have positive values within  
759: the absolute value signs in  
760: (\ref{eq:f_sb23}), since the signs of $\epsilon_2+A_3 r_1^{\delta_3}$ and  
761: $\epsilon_3+A_1|\epsilon_2+A_3r_1^{\delta_3}|^{\delta_1}$ determine the next  
762: values of $\pm_2$ and $\pm_3$.  
763:   
764:  %%%%%%%%%%%%%%  
765: \begin{figure}  
766: \centerline{\epsfxsize=0.9\hsize\epsfbox{map_epsi23.eps}}  
767: \caption{Schematic graphs of $F(r_1)$ (equation (\ref{eq:f_sb23}))  
768: for the choice $\pm_2=\pm_3=1$ when $\delta_1>1$,  
769: $\delta_2>1$, $\delta_3>1$, and $A_1=A_2=A_3=1$. The solid (resp.~dotted) curve  
770: indicates values of $r_1$ for which the next values of $\pm_2$ and $\pm_3$ are  
771: (resp.~are not) both positive; we seek values of $r_1$ for which the solid  
772: curve intersects the diagonal. For small positive~$\epsilon_2$  
773: and~$\epsilon_3$, a stable fixed point exists (see panel (b)). This fixed point  
774: ceases to exist in the second quadrant when  
775: $\epsilon_2=-A_3A_2^{\delta_2}\epsilon_3^{\delta_2\delta_3}$, when there is a  
776: non-transversal homoclinic bifurcation of $P$ (limiting case shown in panel  
777: (a)). The fixed point is destroyed in the fourth quadrant when  
778: $\epsilon_3=-A_1\epsilon_2^{\delta_1}$, when there is a non-transversal  
779: heteroclinic connection from $+E_3$ to $+E_2$ (limiting case shown in  
780: panel~(c)).}  
781:  \label{fig:mapepsi23}  
782:  \end{figure}  
783:  %%%%%%%%%%%%%%  
784:   
785: For sufficiently small, positive $\epsilon_2$ and $\epsilon_3$,  
786: $F(0)=A_2(\epsilon_3+A_1\epsilon_2^{\delta_1})^{\delta_2}>0$. For $r_1$ larger  
787: than $\epsilon_2$ and $\epsilon_3$ but still smaller than one, we have  
788: $F(r_1)\sim r_1^{\delta}$, which is less than~$r_1$ since $\delta>1$. Thus, by  
789: continuity, the map has a fixed point (see figure~\ref{fig:mapepsi23}(b)).  
790: Since $F(r_1)$ is monotonically increasing, the slope of~$F$ at the  
791: fixed point is less than one, so a stable fixed point exists for 
792:  $\epsilon_2>0$, $\epsilon_3>0$.  
793:   
794: This fixed point (i.e., a fixed point with $\pm_2=\pm_3=+$) also exists in  
795: parts of the second and fourth quadrants of the $(\epsilon_2,\epsilon_3)$  
796: parameter plane. To determine the region of existence in the  
797: fourth quadrant, we fix $\epsilon_2$ at some small positive value and  
798: decrease $\epsilon_3$. This shifts the  
799: graph of $F(r_1)$ down, from which it is found that a stable fixed point  
800: exists until $F(0)=0$, i.e., until  
801: $\epsilon_3=-A_1\epsilon_2^{\delta_1}$ 
802: (figure~\ref{fig:mapepsi23}(c)). Thus the fixed point  
803: ceases to exist in the fourth quadrant at the locus of the   
804: heteroclinic bifurcation from $+E_3$ to $+E_2$ (c.f.~section~\ref{hetEs}).  
805:   
806: To determine where the fixed point exists in the second quadrant,  
807: we fix $\epsilon_3$ at some small positive value and decrease $\epsilon_2$.  
808: This decreases $F(0)$ and also changes the   
809: shape of the graph of $F(r_1)$; the graph remains monotonic  
810: increasing in $r_1$ while $\epsilon_2$ is positive, but develops a turning point  
811: once $\epsilon_2$  
812: becomes negative, with $F(r_1)$ decreasing for  
813: $r_1$ near zero. The decreasing section is indicated by a dotted curve in  
814: figure~\ref{fig:mapepsi23}(a), and corresponds to future values of $\pm_2$ and  
815: $\pm_3$ not both being positive. For small enough negative $\epsilon_2$ the  
816: dotted section of the graph lies to the left of fixed point, but when  
817: $\epsilon_2=-A_3A_2^{\delta_2}\epsilon_3^{\delta_2\delta_3}$, the dotted curve  
818: reaches to the diagonal and the fixed point ceases to exist  
819: (figure~\ref{fig:mapepsi23}(a)). Thus the  fixed point ceases to exist in the second  
820: quadrant on the locus of the homoclinic bifurcation of $P$ (c.f.~section \ref{homP}).  
821:   
822: The region of existence of this stable fixed point is indicated by the left-leaning  
823: close hatching in figure~\ref{fig:globalbifns}. Calculations with the other  
824: combinations of signs of $\pm_2$ and $\pm_3$ are analogous, and yield different  
825: regions of existence for the corresponding fixed points. Fixed points may coexist  
826: as shown in  
827: figure~\ref{fig:globalbifns}.  
828:   
829: Since for fixed $r_1$ the $\theta_1$ component of map (\ref{eq:R_sb23}) is a rigid rotation, 
830: a fixed point of  equation (\ref{eq:f_sb23}) generically corresponds to a closed  
831: invariant curve in (\ref{eq:R_sb23}) and to a quasiperiodic solution in the  
832: full flow.  The angle~$\theta_1$ decouples from the rest of the dynamics, and so the  
833: full flow will have an invariant torus foliated by periodic orbits for a dense set of  
834: parameter values. Stability of these solutions follows from the stability of the  
835: fixed point of~(\ref{eq:f_sb23}).  
836:   
837: The calculations above were for the case $\delta_1>1$, $\delta_2>1$,  
838: $\delta_3>1$. Similar calculations done when one or more of the  
839: $\delta_i$ is smaller than one lead to  
840: similar regions of existence of quasiperiodic solutions, except that there are 
841: additional saddle-node bifurcations of the tori close to the relevant  
842: global bifurcations; these saddle-node bifurcations arise since the  
843: global bifurcations destroy solutions of different stabilities  
844: depending on the sign of $\delta_i -1$.  
845:   
846: The special case that precisely  one of $\epsilon_2$ and $\epsilon_3$ is zero  
847: (i.e., only one reflection symmetry is broken) is covered by the analysis  
848: above; there will be two fixed points of the map with corresponding (foliated)  
849: tori in the flow.  
850:   
851: In summary, when the two reflection  
852: symmetries are broken but the rotation symmetry is preserved, the flow generically  
853: has asymptotically stable quasiperiodic  
854: solutions that do not exhibit switching. Different quasiperiodic solutions coexist in regions  
855: bounded by global bifurcations that will play an important role in generating complex  
856: dynamics when $\epsilon_1\neq0$.  
857:   
858: \subsection{Breaking the rotation symmetry}  
859: \label{sec:rotate}  
860:   
861: In this subsection we show that for sufficiently small values of $\epsilon_1$,  
862: with $\epsilon_2=\epsilon_3=0$, the map (\ref{eq:R_sb1simple}) has a stable  
863: fixed point.  
864:   
865: By rescaling $r_1$ and $\epsilon_1$ by order one amounts and moving the origin  
866: of the $\theta_1$ coordinate, we can without loss of generality set $A=1$,  
867: $a_\smallr=1$ and $a_\smalli=0$ in (\ref{eq:R_sb1simple}). Ignoring for now the  
868: $x_2$ and $\xi_3$ components of the map and working with polar coordinates  
869: $(\rho,\phi)$ centred at $(x_1,y_1)=(\epsilon_1,0)$ (so that  
870: $x_1=\epsilon_1+\rho\cos \phi$ and $y_1=\rho \sin \phi$), map  
871: (\ref{eq:R_sb1simple}) reduces to  
872:  \begin{equation}  
873:  \begin{cases}  
874:    \tilde{x}_1=\epsilon_1 + \tilde\rho\cos\tilde\phi\cr  
875:    \tilde{y}_1=\phantom{\epsilon_1+{}} \tilde\rho\sin\tilde\phi\cr  
876:  \end{cases}  
877:  \qquad \mbox{where} \qquad  
878:  \begin{cases}  
879:    \tilde\rho = r_1^\delta\cr  
880:    \tilde\phi = \theta_1+\Phi - Q\ln r_1.\cr  
881:  \end{cases}  
882:  \label{eq:Rnorotrhophi}  
883:  \end{equation}  
884: The constant $\Phi$ may take a different value here than in  
885: equations~(\ref{eq:R_sb1simple}). Fixed points of (\ref{eq:Rnorotrhophi})  
886: satisfy $r_1=\sqrt{{\tilde{x}_1}^2+{\tilde{y}_1}^2}$ and  
887: $\theta_1=\arctan\left(\tilde{y}_1/\tilde{x}_1\right)$. To find solutions of  
888: the first of these equations, note that circles of radius $r$ about  
889: $(x_1,y_1)=(0,0)$ map under~(\ref{eq:Rnorotrhophi}) to circles of radius  
890: $r^\delta$ about $(\epsilon_1,0)$. Since $\delta>1$ and for small $r$, these  
891: circles will intersect if $r\geq\epsilon_1-r^\delta$ and  
892: $r\leq\epsilon_1+r^\delta$; the intersection points are candidate fixed points  
893: of the map. For each small fixed value of $\epsilon_1$, there will be some  
894: non-zero interval $a\leq r\leq b$ on which the inequalities are both  
895: satisfied.  See figure \ref{fig:tildephi1}(a--c).  
896:   
897: The second equation, $\theta_1=\arctan\left(\tilde{y}_1/\tilde{x}_1\right)$, is  
898: satisfied for at least one value of~$r$ in $[a,b]$, as the following argument  
899: shows. When $r=a$, the circles $(r_1, \theta_1)=(r, \theta_1)$ and  
900: $(\rho,\phi)=(r^\delta, \phi)$ intersect at a single point,  
901: $(r_1,\theta_1)=(a,0)$, alternatively $(\rho, \phi)=(a^\delta, \pi)$. As $r$ is  
902: increased beyond~$a$, the intersection point splits into two (with  
903: corresponding $\phi$ values just below $\pi$ and just above $-\pi$). The  
904: intersection points come together again at $(r_1,\theta_1)=(b,0)$ or  
905: $(\rho,\phi)=(b^\delta,0)$, in the manner shown in  
906: figure~\ref{fig:tildephi2}(a). The corresponding value of $\tilde\phi$ for each  
907: intersection point can be calculated from~(\ref{eq:Rnorotrhophi}) (see  
908: figure~\ref{fig:tildephi2}(b)) from which it is seen that the two branches of  
909: $\tilde\phi$, arising from the upper and lower intersections of the two  
910: circles, start and end at the same point as each other, as shown in  
911: figure~\ref{fig:tildephi2}(c). At least one of the two branches of the graph of  
912: $\tilde\phi$ vs $r$ therefore intersects the graph of $\phi$ vs $r$ for that  
913: same branch for at least one $r$ in $[a,b]$. Thus, there is at least  
914: one fixed point of the map~(\ref{eq:Rnorotrhophi}).  
915:   
916:  %%%%%%%%%%%%%%%%%  
917: \begin{figure}  
918: \centerline{\epsfxsize=0.9\hsize\epsfbox{tildephi1.eps}}  
919:  % \includegraphics[width=\hsize]{tildephi}  
920: \caption{Finding fixed points of equations (\ref{eq:Rnorotrhophi}). Panels  
921: (a--c) show schematically the relative positions of the circle of radius $r$  
922: around $(x_1,y_1)=(0,0)$ (large circle in each panel) and its image under map  
923: (\ref{eq:Rnorotrhophi}) for various sizes of $r$: (a) $r <  
924: \epsilon_1-r^{\delta}$, no intersection; (b) $r \geq \epsilon_1-r^{\delta}$ and  
925: $r \leq \epsilon_1+r^{\delta}$, one or two intersections; (c) $r >  
926: \epsilon_1+r^{\delta}$, no intersections.} 
927:  \label{fig:tildephi1}  
928:  \end{figure}   
929:  %%%%%%%%%%%%%%%%%  
930:   
931:  %%%%%%%%%%%%%%%%%  
932: \begin{figure}  
933: \centerline{\epsfxsize=0.9\hsize\epsfbox{tildephi2.eps}}  
934:  % \includegraphics[width=\hsize]{tildephi}  
935: \caption{(a)~Schematic $\phi$ values for intersections of the two circles 
936: $(r_1, \theta_1)=(r, \theta_1)$ and $(\rho,\phi)=(r^\delta, \phi)$, in the 
937: situation shown in figure~\ref{fig:tildephi1}(b), plotted as a function of $r$; 
938:  (b)~Schematic showing $\theta_1$ (solid and dashed curves) and $-Q\ln r_1$  
939: (dotted curve) for points of intersection of the two circles, plotted as a  
940: function of $r$ at the intersection points;  
941:  (c)~$\phi$ and $\tilde\phi$ values at points of intersection of the two  
942: circles. Solid curves correspond to the upper intersection  
943: points and their images, dashed curves correspond to the lower intersection  
944: points and their images.}  
945:  \label{fig:tildephi2}  
946:  \end{figure}   
947:  %%%%%%%%%%%%%%%%%  
948:   
949: It is straightforward to show that the determinant of the Jacobian of  
950: map~(\ref{eq:Rnorotrhophi}) is $\delta r^{2(\delta-1)}$ and that the absolute  
951: value of the trace of the Jacobian is bounded above by $  
952: \sqrt{(\delta+1)^2+Q^2}\, r^{\delta-1}.$ For each choice of $\epsilon_1$, the  
953: value of $r$ at the corresponding fixed point lies in $[a,b]$, where $a$ and  
954: $b$ depend on $\epsilon_1$ with $a>0$ and $a$ and $b$ tending to zero as  
955: $\epsilon_1 \to 0$. Thus, for sufficiently small $\epsilon_1$, the determinant  
956: and trace of the linearised map are small enough to ensure that the  
957: relevant fixed point is stable. It follows that the corresponding fixed point  
958: of equations (\ref{eq:R_sb1simple}) is also stable, since orbits of that  
959: map collapse onto a constant value of the third (radial) coordinate after one  
960: iteration of the map.  
961:   
962: In summary, for sufficiently small $\epsilon_1$ when $\epsilon_2=\epsilon_3=0$, the  
963: heteroclinic cycle that occurred in the fully symmetric case is replaced by a  
964: stable periodic orbit.  Using the reflection symmetries $\kappa_2$ and  
965: $\kappa_3$, we find four stable periodic orbits co-exist, one in each quarter  
966: of the phase space.  
967:   
968: \subsection{Breaking all symmetries}  
969: \label{sec:full}  
970:   
971: Direct analysis of the return map valid for the case that all symmetries are  
972: broken is not feasible because of the extremely complicated form of that map.  
973: Instead, in this section we use our knowledge of the dynamics in the case  
974: $\epsilon_1=0$ and arguments about generic unfoldings of this special case to  
975: deduce what types of dynamics will be seen in the fully asymmetric case for  
976: $\epsilon_1$ near zero, and to (approximately) locate each type of behaviour in  
977: parameter space. This procedure allows us to make specific predictions about  
978: the mechanisms underlying the complicated dynamics observed in numerical  
979: examples, such as the example described in section~\ref{sec:numerics}. We are  
980: especially interested in finding mechanisms that cause repeated, non-periodic  
981: switching in our system.  
982:   
983: The dynamics associated with the case $\epsilon_1=0$ is summarised in  
984: figure~\ref{fig:globalbifns}, which shows eight curves of global bifurcations  
985: bounding regions in which there are asymptotically stable quasiperiodic  
986: solutions. In this case the rotation symmetry prevents coupling of the two  
987: frequencies associated with each quasiperiodic solution and the dynamics is  
988: simple. Once $\epsilon_1$ moves away from zero, we will generically see locking  
989: of the frequencies. For instance, if we were to fix $\epsilon_1$ sufficiently  
990: small but non-zero and pick $\epsilon_2$ and $\epsilon_3$ positive and with  
991: values midway between the two pairs of global bifurcation curves in the first  
992: quadrant of figure~\ref{fig:globalbifns-sb}, then along a one-dimensional path  
993: through the parameter space such as  the dotted line in  
994: figure~\ref{fig:globalbifns-sb} there will be intervals of quasiperiodicity  
995: interspersed with intervals of locked, periodic behaviour. Associated with the  
996: frequency locking there may be complicated dynamics such as period-doubling  
997: cascades and chaotic dynamics. However, this behaviour will mostly be confined  
998: to regions of phase space near the original quasiperiodic solutions, and is not  
999: the main mechanism for switching in our system.  
1000:   
1001: As shown in section~\ref{homP} and in  
1002: figure~\ref{fig:globalbifns-sb}, each curve of non-transverse homoclinic  
1003: bifurcations of $P$ seen in figure~\ref{fig:globalbifns} turns into a wedge in  
1004: parameter space of transverse homoclinic orbits of $P$  
1005: when $\epsilon_1$ changes from zero. There will be horseshoes and chaotic  
1006: dynamics associated with the transverse homoclinic orbits, although the chaos  
1007: may not be attracting. In numerical examples we might expect to see a mixture  
1008: of stable periodic orbits and stable chaotic dynamics, in overlapping regions  
1009: of parameter space.  
1010:   
1011: An interesting consequence of the occurrence of homoclinic tangles is that it 
1012: provides a mechanism for switching of orbits with respect to the $x_2$ 
1013: variable.  For instance, for sufficiently small $\epsilon_1$, and with 
1014: $\epsilon_2>0$, $\epsilon_3>0$ and both small, solutions that make excursions 
1015: near $+E_2$ can get trapped. The trapping region is bounded in part by one 
1016: branch of ${\cal W}^\smalls(P)$ and trapped solutions make excursions past 
1017: $+E_2$ but cannot cross ${\cal W}^\smalls(P)$ so cannot get close to $-E_2$. If 
1018: $\epsilon_2$  is decreased, say by moving along the dotted path shown in Figure 
1019: 4 into the homoclinic wedge in the second quadrant, the trapping region 
1020: develops a leak when a homoclinic tangency forms between that branch of ${\cal 
1021: W}^\smalls(P)$ and a branch of ${\cal W}^\smallu(P)$; solutions are then able 
1022: to cross ${\cal W}^\smalls(P)$, and may visit a neighbourhood of $-E_2$. We 
1023: call this `switching in $x_2$'. Switching of this type (from positive to 
1024: negative $x_2$) can occur for parameter values to the left of the right 
1025: boundary of the homoclinic wedge in the second quadrant. A numerical example of 
1026: this leaking process is given below in figure~\ref{fig:x2switchp}. 
1027:   
1028: Once a switching solution arrives in the region with $x_2<0$, it may then get  
1029: captured by an attractor lying solely in the negative $x_2$ region of phase  
1030: space, in which case no more switching will be observed. Alternatively, if  
1031: there is a mechanism for orbits to leak back to the original region of phase  
1032: space then there could be sustained switching in $x_2$. This latter case cannot  
1033: occur for arbitrarily small $\epsilon_1$ as the following argument shows. For  
1034: the case $\epsilon_1=0$, results from section~\ref{sec:reflect} show that  
1035: orbits which make repeated excursions near $-E_2$ and $+E_3$ occur in the second  
1036: quadrant of figure~\ref{fig:globalbifns-sb} and in the first quadrant as far as  
1037: the locus of homoclinic bifurcations of $P$. (This homoclinic bifurcation  
1038: involves a different branch of ${\cal W}^\smallu(P)$ than the homoclinic bifurcation  
1039: occurring for $\epsilon_2<0$ discussed in the last paragraph.) For small  
1040: $\epsilon_1$, there will be a wedge of homoclinic tangencies of the relevant  
1041: branches of ${\cal W}^\smallu(P)$ and ${\cal W}^\smalls(P)$, with the wedge lying entirely within the  
1042: region $\epsilon_2>0$, $\epsilon_3>0$; orbits that make a number of excursions  
1043: past $-E_2$ before switching and passing close to $+E_2$ can only occur for  
1044: values of $\epsilon_2$ and $\epsilon_3$ lying to the right of the left boundary  
1045: of the wedge. Thus, for sufficiently small $\epsilon_1$, there is no overlap  
1046: between the region of parameter space where there is switching from positive to  
1047: negative $x_2$ and the region where there is switching from negative to  
1048: positive $x_2$, with the consequence that there can be no sustained switching  
1049: in $x_2$.  
1050:   
1051: However, as argued in section~\ref{homP}, for large enough $\epsilon_1$ the  
1052: curves of homoclinic tangency may change quadrants in the  
1053: $(\epsilon_2,\epsilon_3)$ parameter plane, and then the switching regions can  
1054: overlap, making sustained switching in $x_2$ possible. As pointed out in  
1055: section~\ref{homP}, our return map construction is not valid for `large'  
1056: $\epsilon_1$, so we have not proved the existence of sustained switching, just  
1057: shown how it might feasibly occur. It is not possible to determine a priori how  
1058: big $\epsilon_1$ would have to be to get sustained switching, but we have shown  
1059: that there is a threshold in $\epsilon_1$ below which sustained switching in  
1060: $x_2$ is not possible. The value of this threshold does not go to zero as  
1061: $\epsilon_2$ and $\epsilon_3$ go to zero, as it comes from the requirement that  
1062: $1+\epsilon_1g_3(\Theta_1)$ changes sign (as a function of~$\Theta_1$) in  
1063: equation~(\ref{eq:thresholdx2}). Another way of understanding this is to note  
1064: that in~(\ref{eq:thresholdx2}), if $\pm_2=+$ and $\epsilon_2>0$, then the only  
1065: way of having $\tilde{x}_2$ negative is to have $1+\epsilon_1g_3(\Theta_1)<0$  
1066: for some value of~$\Theta_1$. This is a necessary but not sufficient condition,  
1067: as the attractor may not explore the required range of~$\Theta_1$.  
1068:   
1069: The four curves of heteroclinic bifurcations of the cycles $\pm E_2 \to \pm E_3 
1070: \to \pm E_2$ shown in figure~\ref{fig:globalbifns} also split when $\epsilon_1$ 
1071: becomes non-zero, being replaced by eight curves of homoclinic bifurcations and 
1072: eight curves  of heteroclinic tangencies between ${\cal W}^\smallu( \pm E_3)$ 
1073: and ${\cal W}^\smalls(\pm E_2)$,  as described in section~\ref{hetEs}. If they 
1074: are of Shil'nikov type, the homoclinic bifurcations can complicate the dynamics 
1075: by inducing chaotic dynamics. The heteroclinic bifurcations are associated with 
1076: switching in the $x_3$ coordinate similarly to the way switching in $x_2$ is 
1077: associated with homoclinic bifurcations of $P$, described above. More 
1078: precisely, the eight curves of heteroclinic tangencies between ${\cal 
1079: W}^\smallu(\pm E_3)$ and ${\cal W}^\smalls(\pm E_2)$ come in pairs, with each 
1080: pair bounding a wedge in parameter space. At parameter values within each wedge 
1081: there is a heteroclinic tangle of one pair of manifold branches. For instance, 
1082: for sufficiently small $\epsilon_1$ there will be a heteroclinic wedge 
1083: involving one branch of ${\cal W}^\smallu(+E_3)$ and one branch of ${\cal 
1084: W}^\smalls(+E_2)$ occurring in the fourth quadrant in the 
1085: $(\epsilon_2,\epsilon_3)$ plane. Above this wedge, ${\cal W}^\smalls(+E_2)$ 
1086: bounds in part a trapping region; orbits in the trapping region make excursions 
1087: past $+E_3$ but cannot cross ${\cal W}^\smalls(+E_2)$ to get close to $-E_3$. 
1088: The trapping region develops a leak when a heteroclinic tangency forms between 
1089: the appropriate branches of ${\cal W}^\smallu(+E_3)$ and ${\cal 
1090: W}^\smalls(+E_2)$ thus allowing solutions to cross ${\cal W}^\smalls(+E_2)$. We 
1091: call this `switching in $x_3$'. An argument analogous to that used for 
1092: switching in $x_2$ can be used here to show that there is a (generically 
1093: different) threshold in $\epsilon_1$ below which there can be no persistent 
1094: switching in~$x_3$. This threshold does not go to zero when $\epsilon_2$ and 
1095: $\epsilon_3$ go to zero. 
1096:   
1097: The mechanisms inducing switching in $x_2$ and in $x_3$ are distinct, but  
1098: orbits that switch persistently in both $x_2$ and $x_3$ are possible for  
1099: $\epsilon_1$ above the thresholds for both mechanisms. Switching in each  
1100: variable requires the rotation and appropriate reflection symmetry to be  
1101: broken. It is possible to have persistent switching in~$x_2$ with  
1102: $\epsilon_3=0$, or switching in~$x_3$ with $\epsilon_2=0$, though we will not  
1103: explore this possibility in detail. The example in section \ref{sec:numerics}  
1104: shows that persistent switching is easily observed in numerical examples.  
1105:   
1106: Both of the global bifurcations we have identified as inducing switching, i.e.,  
1107: homoclinic tangencies of ${\cal W}^\smallu(P)$ and ${\cal W}^\smalls(P)$ and heteroclinic tangencies of  
1108: ${\cal W}^\smallu(\pm E_3)$ and ${\cal W}^\smalls(\pm E_2)$, will produce horseshoes in the dynamics. In  
1109: the case of homoclinic tangencies, this is a standard result and in the case of  
1110: the heteroclinic tangencies, reinjection into the neighbourhood of the  
1111: heteroclinic tangle is provided by proximity in phase and parameter space to  
1112: the heteroclinic cycle $\pm E_2 \to \pm E_3 \to \pm E_2$. In either case, we  
1113: expect the onset of switching to be commonly associated with nearby  
1114: chaotic dynamics; chaotic orbits before the onset of switching, chaotic  
1115: transients for switching orbits and orbits that switch chaotically might all be  
1116: seen. However, other types of switching are also possible, such as  
1117: periodic switching where the attractor is a periodic orbit that crosses the  
1118: (non-invariant) hyperplanes $x_2=0$ and $x_3=0$ or periodic switching where the  
1119: attractor is a `noisy periodic orbit' such as results from a cascade of period  
1120: doubling. In the latter case the itinerary of visits to $\pm E_2$ or $\pm E_3$  
1121: will be periodic even though the actual orbits are not.  
1122:   
1123: \section{Example}  
1124: \label{sec:numerics}  
1125:   
1126:  We consider the following system of equations to illustrate the dynamics of  
1127: interest in this paper:  
1128:  \begin{align}  
1129:  {\dot z_1} &= (1+{\rm i})z_1 - |z_1|^2z_1  
1130:               - (c_2+1)x_2^2z_1  
1131:               + (e_3-1)x_3^2z_1  
1132:               + \epsilon_1d_{11} + \epsilon_1d_{12}x_1\,,  
1133:   \label{eq:modelSBODEsfirst}\\  
1134:  {\dot x_2} &= x_2 - x_2^3  
1135:               - (c_3+1)x_3^2x_2  
1136:               + (e_1-1)|z_1|^2x_2  
1137:               + \epsilon_2d_{21} + \epsilon_1d_{22}x_1x_2  
1138:               + \epsilon_1\epsilon_2d_{23}x_1\,,  
1139:   \\  
1140:  {\dot x_3} &= x_3 - x_3^3  
1141:               - (c_1+1)|z_1|^2x_3  
1142:               + (e_2-1)x_2^2x_3  
1143:               + \epsilon_3d_{31} + \epsilon_1d_{32}x_1x_3  
1144:               + \epsilon_1\epsilon_3d_{33}x_1\,,  
1145:   \label{eq:modelSBODEslast}  
1146:  \end{align}  
1147: These equations were derived by starting with the structurally stable
1148: heteroclinic cycle considered in~\cite{GuHo88}, turning a pair of equilibria of
1149: that cycle into a periodic orbit by adding a trivial phase variable, and adding
1150: the simplest possible terms that break the symmetries in generic ways. The
1151: parameters $\epsilon_1$, $\epsilon_2$ and $\epsilon_3$ in equations
1152: (\ref{eq:modelSBODEsfirst}--\ref{eq:modelSBODEslast}) play the same role as in
1153: the maps derived earlier in this paper.
1154:   
1155: The model used in~\cite{MPR} is similar, but differs in two  
1156: respects. First, the symmetry-breaking terms in~\cite{MPR} were fifth-order in  
1157: the $x$ and $z$ variables, rather than constant, linear and quadratic here.  
1158: Second, the model in~\cite{MPR} respects the symmetry  
1159: $(z_1,x_2,x_3)\rightarrow(-z_1,-x_2,-x_3)$, as appropriate for a model of  
1160: a dynamo instability: the invariant subspace $z_1=x_2=x_3=0$ corresponds to the  
1161: absence of any magnetic field. We do not expect the first difference between   
1162: models to alter the qualitative behaviour, but the enforced symmetry may  
1163: have a significant effect, as discussed briefly in the next section.  
1164:   
1165: The coefficients in the equations were chosen to be:  
1166:   $c_1=1.2$, $e_1=1.0$,  
1167:   $c_2=1.1$, $e_2=1.0$,  
1168:   $c_3=1.1$, $e_3=1.0$ for the contracting and expanding eigenvalues, and  
1169:   $d_{11}=d_{12}=10^{-4}$,  
1170:   $d_{21}=10^{-1}$,  
1171:   $d_{22}=10^{-1}$,  
1172:   $d_{23}=10^{3}$,  
1173:   $d_{31}=10^{-3}$,  
1174:   $d_{32}=10^{-4}$,  
1175:   $d_{33}=1$ for the symmetry-breaking coefficients.  
1176:  The eigenvalues were chosen to be of order one, with contraction dominating  
1177: expansion at each point ($\delta_1=1.2$, $\delta_2=1.1$, $\delta_3=1.1$, and an  
1178: overall $\delta=1.452$). The symmetry breaking coefficients are notionally  
1179: small, but those coefficients ($d_{23}$ and $d_{33}$) that are multiplied by  
1180: two $\epsilon$'s were chosen to be larger to compensate for this. The exact  
1181: numbers are not important, though they will affect the details of what is   
1182: observed. However, choosing $d_{23}$ and $d_{33}$ to be reasonably large 
1183: means that the switching dynamics is easier to obtain for small values  
1184: of~$\epsilon_1$: in order to get persistent switching, the  
1185: $\epsilon_1\epsilon_2d_{23}x_1$ and $\epsilon_1\epsilon_3d_{33}x_1$ terms in 
1186: (\ref{eq:modelSBODEsfirst}--\ref{eq:modelSBODEslast}) need to be reasonably 
1187: important. 
1188:   
1189: We integrated the equations numerically using the Bulirsch--Stoer adaptive  
1190: integrator \cite{NumRecipes}, with a tolerance for the relative error set to  
1191: $10^{-12}$ for each step. Poincar\'e sections were computed using algorithms   
1192: from~\cite{PC89}.  
1193:   
1194: By varying $\epsilon_1$, $\epsilon_2$ and $\epsilon_3$, we are able to find  
1195: examples of the important symmetry-breaking effects discussed in the previous  
1196: sections of this paper. The cases with full or partial symmetry preserved give  
1197: straightforward results, which we describe only briefly; more details are  
1198: provided of the case of fully broken symmetry.  
1199:   
1200: If all symmetries are preserved  (all $\epsilon_i=0$) then each solution  
1201: starting off the invariant subspaces is attracted to one of four symmetry-related 
1202: structurally stable heteroclinic cycles.  
1203: If  $\epsilon_1 \ne 0$, $\epsilon_2=\epsilon_3=0$ (rotation  
1204: symmetry broken, reflections preserved), numerics confirm the predictions of  
1205: section~\ref{sec:rotate}, and a single attracting periodic orbit is found in  
1206: each quarter of the phase space. If rotation symmetry is preserved as well as  
1207: one reflection, and the other reflection is broken, then solutions are  
1208: attracted to a foliated torus, as discussed in section~\ref{sec:reflect}. In  
1209: the case that both reflections are broken but the rotation symmetry is  
1210: preserved, numerics confirm the predictions of section~\ref{sec:reflect}; we  
1211: find that there exist attracting quasiperiodic solutions in regions bounded by  
1212: curves of global bifurcations, as shown schematically in  
1213: figure~\ref{fig:globalbifns}. Analysis of the maps derived earlier allows us to  
1214: predict scaling of the loci of various global bifurcations in the limit of  
1215: small symmetry breaking. For instance,  equation (\ref{eq:homP}) tells us that  
1216: for $\epsilon_1=0$ and $\epsilon_3 \to 0$, homoclinic bifurcations of $P$  
1217: associated with equations (\ref{eq:modelSBODEsfirst}--\ref{eq:modelSBODEslast})  
1218: occur for $\epsilon_2={\rm constant} \times |\epsilon_3|^{\delta_2 \delta_3}$  
1219: but the  value of the constant is not determined by the map analysis. Numerical  
1220: simulations of equations (\ref{eq:modelSBODEsfirst}--\ref{eq:modelSBODEslast})  
1221: confirm the scalings for the various global bifurcations.  
1222:   
1223: To illustrate the phenomena associated with breaking all symmetries, it is  
1224: helpful to consider the changes in dynamics seen along a one-dimensional path  
1225: such as that shown as the dotted line in figure \ref{fig:globalbifns-sb}. We  
1226: first chose a value of~$\epsilon_1$ below the thresholds for persistent  
1227: switching in $x_2$ and~$x_3$. For instance, fixing  $\epsilon_1=10^{-4}$ and  
1228: $\epsilon_3=0.001$ and allowing $\epsilon_2$ to vary, we see the following  
1229: types of dynamics.  
1230:   
1231: Picking $\epsilon_2=3\times10^{-5}$ yields 
1232: a point (labelled~A) lying to the right of the homoclinic wedge and to left of the  
1233: heteroclinic wedge in the first quadrant of figure~\ref{fig:globalbifns-sb}.  
1234: For these $\epsilon_2$ and $\epsilon_3$ values but for $\epsilon_1=0$, there  
1235: exists an attracting quasiperiodic solution with $x_2$ and $x_3$ both positive.  
1236: With $\epsilon_1=10^{-4}$ the same type of quasiperiodic solution exists. As  
1237: $\epsilon_2$ is decreased while $\epsilon_1$ is fixed at $10^{-4}$  
1238: we find, as expected, intervals of $\epsilon_2$ in  
1239: which there are quasiperiodic attractors interspersed with intervals on which  
1240: there is locking of the two frequencies associated with the quasiperiodic  
1241: solution (periodic orbits). In some intervals, period doubling cascades are  
1242: observed, as is normal near quasiperiodic behaviour.  The interchange between  
1243: locking and quasiperiodic behaviour persists until we approach the homoclinic  
1244: wedge at negative values of $\epsilon_2$.  
1245:   
1246:  %%%%%%%%%%%%%%%%  
1247: \begin{figure}  
1248: \centerline{\epsfxsize=0.9\hsize\epsfbox{kr_fig_x2_switch.eps}}  
1249:  % \mbox{\includegraphics[width=0.9\hsize]{kr_fig_x2_switch.eps}}  
1250: \caption{Onset of switching in $x_2$ for equations  
1251: (\ref{eq:modelSBODEsfirst}--\ref{eq:modelSBODEslast}), associated with crossing  
1252: into the wedge of homoclinic bifurcations of $P$ in the second quadrant of  
1253: $(\epsilon_2,\epsilon_3)$-space.  Parameters $\epsilon_1=10^{-4}$,  
1254: $\epsilon_3=0.001$ are fixed and $\epsilon_2$ is decreased: (a)  
1255: $\epsilon_2=-5.67\times10^{-7}$ gives a chaotic attractor confined to the region  
1256: $x_2>0$; (b) $\epsilon_2=-5.68\times10^{-7}$ gives a chaotic transient with $x_2>0$,  
1257: then the sign of $x_2$ changes and the orbit is attracted to a quasiperiodic  
1258: solution with $x_2<0$. Other coefficients as described in text. The  
1259: chaotic nature of the orbit in (a) and the transient in (b) is not apparent on  
1260: the timescale used to plot the time series.}  
1261:  \label{fig:x2switch}  
1262: \end{figure}  
1263:  %%%%%%%%%%%%%%%%%  
1264:   
1265:  %%%%%%%%%%%%%%%%  
1266: \begin{figure}  
1267: \centerline{\epsfxsize=0.9\hsize\epsfbox{kr_fig_x2_switch_p.eps}}  
1268:  % \mbox{\includegraphics[width=0.9\hsize]{kr_fig_x2_switch_p.eps}}  
1269: \caption{Illustration of how leaking develops as the attractor crosses 
1270: ${\cal W}^\smalls(P)$ (approximately $x_2=0$ in this figure). 
1271: (a) Poincar\'e map for the orbit shown in  
1272: figure~\ref{fig:x2switch}(a); (b)~Poincar\'e map for the orbit in  
1273: figure~\ref{fig:x2switch}(b). In each case the Poincar\'e section is  
1274: $H_1^\smallin$ with $|x_3|=h=0.005$. In each panel, a dot (resp.~cross)  
1275: indicates that the orbit next crosses the Poincar\'e section with $x_2>0$  
1276: (resp.~$x_2<0$). In~(b), the upper collection of dots corresponds to the   
1277: chaotic transient, which ends in a single cross, after which the orbit switches   
1278: to a quasiperiodic attractor, represented by the lower collection of crosses.}  
1279:  \label{fig:x2switchp}  
1280: \end{figure}  
1281:  %%%%%%%%%%%%%%%%%  
1282:   
1283: As this homoclinic wedge is approached, apparently chaotic dynamics is  
1284: observed, consistent with the appearance of horseshoes associated with the  
1285: impending homoclinic tangency. Before the first tangency is reached 
1286: (labelled~B in figure~\ref{fig:globalbifns-sb}), orbits are  
1287: trapped in the region with  $x_2>0$, $x_3>0$ (figure~\ref{fig:x2switch}a). If  
1288: $\epsilon_2$ is decreased past the tangency value (labelled~C), the attractor crosses  
1289: ${\cal W}^\smalls(P)$ and so a typical orbit will display a chaotic transient with $x_2>0$,  
1290: $x_3>0$, and then switch to $x_2<0$, $x_3>0$, after which the orbit is  
1291: attracted to a quasiperiodic solution in that quarter of phase space  
1292: (figure~\ref{fig:x2switch}b). Corresponding Poincar\'e maps are shown in  
1293: figure~\ref{fig:x2switchp}. With negative  
1294: values of~$\epsilon_2$, once the trajectory has switched to $x_2<0$, the  
1295: behaviour is analogous to that observed with $\epsilon_2>0$ and~$x_2>0$:  
1296: quasiperiodic attractors interspersed with frequency locking. Note there is no  
1297: persistent switching in~$x_2$, and no switching in~$x_3$, for these parameter  
1298: values.  
1299:   
1300: If $\epsilon_2$ is now increased from $\epsilon_2=3\times10^{-5}$ while $\epsilon_1$  
1301: and $\epsilon_3$ are kept fixed as before, we approach the wedge of  
1302: heteroclinic connections from $-E_3$ to $+E_2$ (following the dotted line in  
1303: figure~\ref{fig:globalbifns-sb}). Once the left edge of the heteroclinic wedge  
1304: has been crossed, orbits can switch from $x_3<0$ to $x_3>0$ (but not from  
1305: $x_3>0$ to $x_3<0$). For example, at $\epsilon_2=1.1\times10^{-4}$ (labelled~D in  
1306: figure~\ref{fig:globalbifns-sb}), we find a chaotic  
1307: transient with $x_3<0$ that has a single switch to a locked periodic orbit  
1308: with~$x_3>0$, in a manner similar to the single $x_2$ switch in  
1309: figure~\ref{fig:x2switch}(b). For larger $\epsilon_2$, we find other examples  
1310: of quasiperiodicity, locked periodic orbits, chaos, chaotic transients and  
1311: single switches from $x_3<0$ to $x_3>0$, consistent with the analysis presented  
1312: above. We did not find any examples of persistent switching.  
1313:   
1314: The behaviour for negative values of $\epsilon_2$ and/or $\epsilon_3$ is  
1315: analogous: single switches can be found, but there is no persistent switching  
1316: for $\epsilon_1=10^{-4}$. However, persistent switching in one or both of $x_2$  
1317: and $x_3$ is observed if we increase $\epsilon_1$.  Figures~\ref{fig:x2perswitch} and~\ref{fig:x2perswitchp} show an example of   
1318: persistent switching in~$x_2$ (but not~$x_3$) for $\epsilon_1=3\times10^{-4}$,   
1319: $\epsilon_2=2\times10^{-4}$ and $\epsilon_3=0.001$. The trajectory crosses the   
1320: Poincar\'e section in a curve that appears reasonably smooth at the largest   
1321: scale (figure~\ref{fig:x2perswitchp}), but the magnified inset shows that the  
1322: curve has structure, and that parts of the curve lie below~${\cal W}^\smalls(P)$, leading   
1323: to switches from $x_2>0$ to $x_2<0$. For   
1324: these parameter values, there appears to be no attractor with $x_2<0$,   
1325: and after a short transient, the trajectory switches back to $x_2>0$.   
1326: Dynamics with a larger value of $\epsilon_1=0.005$ is shown in  
1327: figures~\ref{fig:x2x3switch} and~\ref{fig:x2x3switchp}: here we have persistent  
1328: switching in $x_2$ and~$x_3$. 
1329:   
1330:  %%%%%%%%%%%%%%%%  
1331: \begin{figure}  
1332: \centerline{\epsfxsize=0.9\hsize\epsfbox{kr_fig_x2_persistent_switch.eps}}  
1333:  % \mbox{\includegraphics[width=0.9\hsize]{kr_fig_x2_persistent_switch.eps}}  
1334: \caption{Time series showing persistent switching in $x_2$ alone, for equations  
1335: (\ref{eq:modelSBODEsfirst}--\ref{eq:modelSBODEslast}) with $\epsilon_1=3\times10^{-4}$,  
1336: $\epsilon_2=2\times10^{-4}$, $\epsilon_3=0.001$. The other coefficients are as defined  
1337: in text.}  
1338:  \label{fig:x2perswitch}  
1339: \end{figure}  
1340:  %%%%%%%%%%%%%%%%%  
1341:   
1342:  %%%%%%%%%%%%%%%%  
1343: \begin{figure}  
1344: \centerline{\epsfxsize=0.9\hsize\epsfbox{kr_fig_x2_persistent_switch_p.eps}}  
1345:  % \mbox{\includegraphics[width=0.9\hsize]{kr_fig_x2_persistent_switch_p.eps}}  
1346: \caption{Poincar\'e section corresponding to the time series shown in  
1347: figure~\ref{fig:x2perswitch}. The inset shows an enlargement of the region in  
1348: the box marked in the main picture. The Poincar\'e section is $H_1^\smallin$  
1349: with $|x_3|=h=0.01$. A dot (resp.~cross) indicates that the orbit next crosses  
1350: the Poincar\'e section with $x_2>0$ (resp.~$x_2<0$). The inset shows that  
1351: crosses (indicating a switch) occur where the trajectory lies  
1352: below~${\cal W}^\smalls(P)$ (approximately $x_2=0$).}  
1353:  \label{fig:x2perswitchp}  
1354: \end{figure}  
1355:  %%%%%%%%%%%%%%%%%  
1356:    
1357:  %%%%%%%%%%%%%%%%  
1358: \begin{figure}  
1359: \centerline{\epsfxsize=0.9\hsize\epsfbox{kr_fig_x2_x3_switch.eps}}  
1360:  % \mbox{\includegraphics[width=0.9\hsize]{kr_fig_x2_x3_switch.eps}}  
1361: \caption{Time series showing persistent switching in $x_2$ and $x_3$, for  
1362: (\ref{eq:modelSBODEsfirst}--\ref{eq:modelSBODEslast}) with  $\epsilon_1=0.005$,  
1363: $\epsilon_2=3\times10^{-5}$, $\epsilon_3=0.001$. Other coefficients as defined in  
1364: text.}  
1365:  \label{fig:x2x3switch}  
1366: \end{figure}  
1367:  %%%%%%%%%%%%%%%%%  
1368:   
1369:  %%%%%%%%%%%%%%%%  
1370: \begin{figure}  
1371: \centerline{\epsfxsize=0.9\hsize\epsfbox{kr_fig_x2_x3_switch_p.eps}}  
1372:  % \mbox{\includegraphics[width=0.9\hsize]{kr_fig_x2_x3_switch_p.eps}}  
1373: \caption{Poincar\'e section corresponding to the time series shown in  
1374: figure~\ref{fig:x2x3switch}. The inset shows an enlargement of the region in  
1375: the box marked in the main picture. The Poincar\'e section is $H_1^\smallin$  
1376: with $|x_3|=h=0.01$. Four symbols are used: a dot (resp.~cross) indicates that  
1377: the orbit next crosses the Poincar\'e section with $x_2>0$ (resp.~$x_2<0$) and  
1378: with $x_3>0$. A~$+$ (resp.~square) indicates that the orbit next crosses the  
1379: Poincar\'e section with $x_2>0$ (resp.~$x_2<0$) and with $x_3<0$. The division  
1380: between orbits falling either side of ${\cal W}^\smalls(P)$ is clearly visible. Orbits  
1381: falling on opposite sides of ${\cal W}^\smalls(\pm E_2)$ are reasonably well mixed with this  
1382: choice of cross-section.}  
1383:  \label{fig:x2x3switchp}  
1384: \end{figure}  
1385:  %%%%%%%%%%%%%%%%%  
1386:   
1387:   
1388:  %%%%%%%%%%%%%%%%  
1389: \begin{figure}  
1390: \centerline{\epsfxsize=0.9\hsize\epsfbox{kr_fig_x3_fitted_global_map.eps}}  
1391: \caption{The function $1+\epsilon_1g_1(\theta)$ fitted to the  
1392: data in figures~\ref{fig:x2perswitch} ($\epsilon_1=3\times10^{-4}$)  
1393: and~\ref{fig:x2x3switch} ($\epsilon_1=0.005$). This function must be both   
1394: positive and negative as a function of~$\theta$ in order to allow persistent   
1395: switching in~$x_3$.}  
1396:  \label{fig:x3fittedmap}  
1397: \end{figure}  
1398:  %%%%%%%%%%%%%%%%%  
1399:   
1400: Our understanding of the mechanism behind persistent switching in $x_2$ or  
1401: $x_3$ requires that $\epsilon_1$ be large enough that $1+\epsilon_1g_3(\theta)$  
1402: or $1+\epsilon_1g_1(\theta)$ can take on positive and negative values, as a  
1403: function of~$\theta$. In order to illustrate this effect, we have fitted these  
1404: functions using the trajectories in figures~\ref{fig:x2perswitch}  
1405: ($\epsilon_1=3\times10^{-4}$) and~\ref{fig:x2x3switch} ($\epsilon_1=0.005$). We have  
1406: concentrated on switching in~$x_3$, so the fact that these trajectories are for  
1407: different values of~$\epsilon_2$ does not affect our conclusions. We took the   
1408: coordinates $(x_3,\theta_1)$ on the section~$H_1^\smallout$ and $\tilde{x}_3$   
1409: from the next intersection with~$H_2^\smallin$, and used these data to fit a   
1410: map of the form of~$\Psi_{12}$:  
1411:  $$  
1412:  \tilde{x}_3= x_3 \left(A_1 + \epsilon_1 f_1(\theta_1)\right)  
1413:               + \epsilon_3 \left(1 + \epsilon_1g_1(\theta_1)\right),  
1414:  $$  
1415: see equation~(\ref{eq:psi12}). We represented the two functions $f_1$ and~$g_1$  
1416: by a finite Fourier series, and were able to fit the data to within one part in  
1417: 1000 for all points with the smaller value of~$\epsilon_1$, and to within one  
1418: part in 100 for all but a handful of the points for the larger value  
1419: of~$\epsilon_1$. As expected, the $A_1+\epsilon_1f_1$ part of the map remains  
1420: positive, but $1+\epsilon_1g_1$ can change sign for the larger value  
1421: of~$\epsilon_1$, as shown in figure~\ref{fig:x3fittedmap}. Indeed, the  
1422: numerical ratio of the amplitudes of the two fitted functions is $16.663$ while  
1423: the ratio of the two values of $\epsilon_1$ is $16.667$. Our understanding  
1424: requires the change in sign of $1+\epsilon_1g_1$ as a necessary condition for  
1425: persistent switching in~$x_3$, which is confirmed by this illustration and by  
1426: our other calculations. A similar transition occurs (at a smaller value  
1427: of~$\epsilon_1$) at the onset of persistent switching in~$x_2$, and in that  
1428: case, the data can be fitted to within 1 part in 10,000 or better.  
1429:   
1430: We note that we were able to find parameter values associated with persistent  
1431: switching in~$x_2$ alone but were unable to get persistent switching in $x_3$  
1432: without also having switching in $x_2$. This is a consequence of the particular  
1433: choice of symmetry-breaking coefficients we use: the threshold in $\epsilon_1$  
1434: for persistent switching in $x_3$ is higher than the threshold for persistent  
1435: switching in $x_2$ for the chosen coefficients (since we have $d_{23}>d_{33}$).  
1436: For other parameter choices, the   
1437: thresholds could be the other way around.  
1438:   
1439: In the numerical simulations described above, the quantities $\delta_1$,  
1440: $\delta_2$, $\delta_3$ and $\delta$ were all greater than one. This choice was  
1441: made to ensure that that the heteroclinic cycle was attracting in the fully  
1442: symmetric case and to remove any possible complications due to chaotic dynamics  
1443: associated with homoclinic bifurcations of $\pm E_2$ and $\pm E_3$. Much of the  
1444: same switching dynamics will still occur if one or both of $\delta_2$ and  
1445: $\delta_3$ is less than one while $\delta>1$, since the mechanism for switching 
1446: we have found does not depend on the size of the individual $\delta_i$. However, in  
1447: this case there may be  
1448: additional complications in the dynamics associated with the homoclinic  
1449: bifurcations of the equilibria.  
1450:   
1451: \section{Conclusions}  
1452: \label{sec:conclusions}  
1453:   
1454: This paper has investigated the effect of small symmetry-breaking on the  
1455: dynamics near a structurally stable heteroclinic cycle connecting two  
1456: equilibria and a periodic orbit. The heteroclinic cycle is structurally stable  
1457: in the case that there are two reflection symmetries and a rotation symmetry in  
1458: the underlying system; we were interested in the dynamics seen when one or more  
1459: of the symmetries is broken. It was reported in~\cite{MPR} that this type of  
1460: system can exhibit seemingly chaotic dynamics along with repeated but irregular  
1461: switching of sign of various variables, but details of the mechanisms  
1462: underlying the onset of complicated dynamics were not explored there. In this  
1463: paper, we have identified global bifurcations that induce the onset of chaotic  
1464: dynamics and switching near a heteroclinic cycle of this type. These turn out  
1465: to be homoclinic tangencies between the stable and unstable manifolds of the  
1466: periodic orbit, and specific heteroclinic tangencies between stable and  
1467: unstable manifolds of the two equilibria. By construction and analysis of  
1468: approximate return maps, we were able to locate (approximately) the global  
1469: bifurcations in parameter space and hence to isolate instances of the different  
1470: types of switching and chaotic dynamics in a specific numerical example.  
1471:   
1472: In addition to identification of the mechanisms underlying the onset of  
1473: switching, two important insights have been gained from this study. First, we  
1474: found that interaction of the different symmetry-breaking terms is required for  
1475: switching; partial symmetry breaking (where one or two of the three  symmetries  
1476: are retained) did not result in switching. Switching results from the right  
1477: combination of a global bifurcation (which results in turn from breaking of the  
1478: rotation symmetry) and small breaking of at least one of the reflection  
1479: symmetries. Second, we found there is a threshold in $\epsilon_1$ below which  
1480: there can be single switches in the signs of certain variables but no  
1481: persistent switching. The important point here is that persistent switching  
1482: does not result from arbitrarily small symmetry breaking, but is a `large'  
1483: symmetry-breaking effect. Of course, `small' and `large' are relative  
1484: terms, and addition of seemingly tiny symmetry-breaking effects might actually  
1485: result in persistent switching, as was the case in the numerical example we  
1486: investigated in section~\ref{sec:numerics}.  
1487:   
1488: One aspect of this problem which has not yet been investigated is whether it is  
1489: possible to make {\it a priori} predictions about switching rates or derive  
1490: scaling laws for switching times. It is plausible that switching rates and  
1491: times might depend on the `distance' from the global bifurcation that induces  
1492: the switching, but no detailed attempts have yet been made to quantify such a  
1493: relationship. The statistics of switching intervals were measured in the  
1494: related model of~\cite{MPR}, who report an exponential distribution of  
1495: intervals between switches.  
1496:   
1497: Finally, we note that the dynamo model in~\cite{MPR} has a symmetry 
1498: that is never broken (this is the symmetry $(z_1,x_2,x_3) \to (-z_1,-x_2,-x_3)$ 
1499: in the notation of~\cite{MPR}). Retention of this symmetry while breaking all 
1500: others amounts to retaining invariance of the $z_1=x_2=x_3=0$ subspace, and 
1501: will have a consequence of relating the dynamics in different parts of 
1502: the phase space. For example, if it is possible to switch from $(x_2>0,x_3>0)$ 
1503: to $(x_2>0,x_3<0)$, it will also be possible to switch from $(x_2<0,x_3<0)$ to 
1504: $(x_2<0,x_3>0)$. Our results do not include this effect, and retaining this 
1505: symmetry may well have profound effects on the switching properties. 
1506: Nevertheless we expect our basic ideas about switching being induced by a 
1507: balance between a global bifurcation and symmetry-breaking terms and the 
1508: existence of a threshold for persistent switching to apply quite generally, and 
1509: to the example in~\cite{MPR} in particular, even if the details turn out not to  
1510: be directly relevant. 
1511:  
1512: \section{Appendix: Details of return map construction}   
1513:   
1514: \subsection{Coordinates and cross-sections}  
1515:  \label{sec:coords}  
1516:   
1517: Following~\cite{KrMe95}, we distinguish radial, contracting, and expanding  
1518: directions near the equilibria in the fully symmetric case.  
1519: If ${\mathcal P}_1=\{(z_1, x_2, x_3): x_3=0\}$, ${\mathcal P}_2=\{(z_1,  
1520: x_2, x_3): z_1=0\}$, ${\mathcal P}_3=\{(z_1, x_2, x_3): x_2=0\}$, with  
1521: ${\mathcal P}_0\equiv {\mathcal P}_{3}$, then the radial eigenvalues at $\pm  
1522: E_j$ ($j=2, 3$) are the eigenvalues of the linearised vector field at $\pm E_j$  
1523: (i.e., eigenvalues of $(d{\mathbf f})_{\pm E_j}$) restricted to ${\mathcal P}_j  
1524: \cap {\mathcal P}_{j-1}$. The contracting eigenvalues are the remaining  
1525: eigenvalues of $(d{\mathbf f})_{\pm E_j}$ in ${\mathcal P}_{j-1}$, and the  
1526: expanding eigenvalues are the remaining eigenvalues in ${\mathcal P}_{j}$. The  
1527: radial direction is then the span of the eigenvectors corresponding to the  
1528: radial eigenvalues, and similarly for the contracting and expanding directions.  
1529: Near $P$ we define the radial direction to be the direction of ${\mathcal P}_1  
1530: \cap {\mathcal P}_{3}$ (i.e., the plane $x_2=x_3=0$), the contracting direction  
1531: is parallel to the $x_3$-axis, and the expanding direction is parallel to the  
1532: $x_2$-axis. These definitions are consistent with those in~\cite{KrMe95}   
1533: but are adapted for the case where there is a  
1534: periodic orbit in the heteroclinic cycle.  
1535:   
1536: We choose local coordinates near each of $P$, $\pm E_2$, and~$\pm E_3$ to 
1537: make the linearised dynamics as simple as possible. Near $+E_2$ in the fully  
1538: symmetric case, we define $\xi_2=x_2-\bar{x}_2$, where $\bar{x}_2$ is the value  
1539: of $x_2$ at~$+E_2$, and then use local coordinates $(z_1,\xi_2,x_3)$; $z_1$,  
1540: $\xi_2$ and $x_3$ correspond to the contracting, radial and expanding  
1541: directions, respectively. Under symmetry breaking, $+E_2$ moves in proportion to  
1542: the magnitude of the symmetry breaking, and the local coordinates are measured  
1543: from the new position of the equilibrium point. The eigenvalues and  
1544: eigenvectors change similarly, but since the eigenvalues are generically  
1545: distinct and non-zero, small symmetry-breaking will not change the nature of  
1546: the local structure and we can use the slightly altered eigenvectors to define  
1547: a slightly altered local coordinate system. We continue to identify radial,  
1548: contracting and expanding directions once weak symmetry breaking is introduced,  
1549: in the obvious way, and retain the notation $(z_1,\xi_2,x_3)$, for the altered  
1550: coordinates, although $z_1$ and $x_3$ may no longer coincide with the  
1551: corresponding global coordinates.  
1552:   
1553: A similar construction is used near $-E_2$ except that $\xi_2=-x_2+\bar{x}_2$,  
1554: where $\bar{x}_2$ is the value of $x_2$ at~$-E_2$. The point of defining  
1555: $\xi_2$ in this way is that positive values of $\xi_2$ near $+E_2$ are mapped  
1556: under the reflection $\kappa_2$ to positive values of $\xi_2$ near $-E_2$, and  
1557: this simplifies the maps we derive below. An analogous procedure is used to  
1558: define local coordinates near $\pm E_3$.  
1559:   
1560: To construct local coordinates near $P$, we select a cross-section transverse  
1561: to $P$, say $\theta_1=0$. Near~$P$, the flow induces a map from that section to  
1562: itself, with $P$ corresponding to a fixed point of the map. We define  
1563: $\xi_1=r_1-\bar{r}_1$, where $\bar{r}_1$ is the value of $r_1$ at the fixed  
1564: point; $\xi_1$ is the analogue in the map to the radial coordinate for the flow  
1565: near $P$. The remaining local coordinates  on the cross-section are defined by  
1566: restricting the expanding and contracting directions at $P$, as defined above,  
1567: to the cross-section. Local coordinates can be extended to a neighbourhood of  
1568: the whole of~$P$  in the  
1569: fully symmetric case by applying equivariance under $\kappa_1$. Finally, small  
1570: symmetry-breaking perturbations will not change the local structure near $P$,  
1571: and we can extend to slightly altered local coordinates $(\xi_1, \theta_1, x_2,  
1572: x_3)$ in a neighbourhood of $P$ so long as we remember that symmetry-breaking  
1573: terms may have a different effect at each value of $\theta_1$, so for instance,  
1574: $z_1=\bar{r}_1{\rm e}^{{\rm i}\theta_1}$ where $\bar{r}_1 \equiv r_1(\theta_1  
1575: )$ is a function of $\theta_1$. Note that the global polar coordinates $(r_1,  
1576: \theta_1)$ are well-defined near $P$ even in the presence of small symmetry  
1577: breaking since $P$ is far from the origin.  
1578:    
1579: Cross-sections in~$\Rset^4$ are defined in terms of local coordinates as  
1580: follows:  
1581:  \begin{align*}  
1582:  H_1^\smallin  &= \left\{(\xi_1,\theta_1,x_2,x_3): |\xi_1| \leq h,  
1583:                                            0 \leq \theta_1 < 2\pi,  
1584:                                            |x_2|\leq h,  
1585:                                            |x_3|=h \right\},\\  
1586:  H_1^\smallout &= \left\{(\xi_1,\theta_1,x_2,x_3): |\xi_1| \leq h,  
1587:                                            0 \leq \theta_1 < 2\pi, |x_2|=h,  
1588:                                            |x_3|\leq h \right\},\\  
1589:  H_2^\smallin  &= \left\{(z_1,\xi_2,x_3): |z_1|=h,  
1590:                                            |\xi_2|\leq h,  
1591:                                            |x_3|\leq h \right\},\\  
1592:  H_2^\smallout &= \left\{(z_1,\xi_2,x_3): |z_1|\leq h,  
1593:                                            |\xi_2| \leq h,  
1594:                                            |x_3|=h \right\},\\  
1595:  H_3^\smallin  &= \left\{(z_1,x_2,\xi_3): |z_1|\leq h,  
1596:                                            |x_2|=h,  
1597:                                            |\xi_3|\leq h \right\},\\  
1598:  H_3^\smallout &= \left\{(z_1,x_2,\xi_3): |z_1|=h,  
1599:                                            |x_2|\leq h,  
1600:                                            |\xi_3|\leq h\right\}.  
1601:  \end{align*}  
1602: The cross-sections $H_2^{\smallin}$ and $H_2^{\smallout}$  
1603: (resp.~$H_3^{\smallin}$ and $H_3^{\smallout}$) work equally well near $\pm E_2$  
1604: (resp.~$\pm E_3$) so long as the local coordinate $\xi_2$ (resp.~$\xi_3$) is  
1605: interpreted correctly, as described above.  
1606:   
1607: We also define a Poincar\'e section for the periodic orbit~$P$:  
1608:  \begin{equation*}  
1609:  H_1^P = \left\{(\xi_1,\theta_1,x_2,x_3): |\xi_1| \leq h,  
1610:                                            \theta_1=0,  
1611:                                            |x_2|\leq h,  
1612:                                            |x_3|\leq h \right\}.  
1613:  \end{equation*}  
1614: Trajectories visiting~$P$ first cross~$H_1^\smallin$, may then cross $H_1^P$  
1615: several times, and eventually leave the neighbourhood of~$P$ on  
1616: crossing~$H_1^\smallout$.  
1617:   
1618: \subsection{Local maps}  
1619:   
1620: Within a neighbourhood of each of $\pm E_2$, $\pm E_3$ and $P$, so long as  
1621: certain non-resonance conditions on the eigenvalues are satisfied, the dynamics  
1622: can be linearised using the Hartman--Grobman theorem~\cite{GH86}. In the fully  
1623: symmetric case, the dynamics near $P$ can be approximated by:  
1624:  $$\dot\xi_1= -2\xi,\qquad  
1625:    \dot\theta_1= 1, \qquad  
1626:    \dot x_2= e_1 x_2,\qquad  
1627:    \dot x_3= -c_1 x_3,$$  
1628:  where $e_1$ and $c_1$ are positive constants. Without loss of  
1629: generality, we have assumed that the radial eigenvalue  is~$-2$, and that the angular speed is 
1630: 1. Solving these equations, we find the local map $\phi_1:H_1^\smallin\to H_1^\smallout$  
1631: is given by:  
1632:  \begin{equation}  
1633:  \phi_1(\xi_1,\theta_1,x_2,x_3)=\left(  
1634:             \xi_1\left|\frac{x_2}{h}\right|^{\gamma_1},  
1635:             \theta_1 -\frac{1}{e_1} \ln \left|\frac{x_2}{h}\right|,  
1636:             h\sgn(x_2),  
1637:             h\sgn(x_3)\left|\frac{x_2}{h}\right|^{\delta_1}\right)\,,  
1638:  \label{eq:phi1map}  
1639:  \end{equation}  
1640: where the initial value of $x_3$ satisfies $|x_3|=h$, where $\sgn(x)=+1$ if $x>0$,  
1641: $\sgn(x)=-1$ if $x<0$, and $\sgn(0)=0$, and where  
1642: $\delta_1=c_1/e_1$, $\gamma_1=2/e_1$.  
1643:   
1644: The argument that symmetry-breaking does not affect this local map  
1645: goes as follows. The transition from $H_1^\smallin\to H_1^\smallout$  
1646: has three parts. First, the trajectory travels from  
1647: $H_1^\smallin\to H_1^P$ in less than one circuit around~$P$. The trajectory  
1648: does not get very close to~$P$ in this time, having started at least a  
1649: distance~$h$ from it. Since the $\epsilon_i$'s, which control the symmetry  
1650: breaking, are assumed to be much smaller than~$h$, the fully symmetric flow  
1651: yields an adequate approximation of the true flow. Second, the trajectory makes  
1652: $n_1$ circuits around the periodic orbit from $H_1^P$ to $H_1^P$, where $n_1$  
1653: is a non-negative integer no greater than $T_1/2\pi$. These circuits are  
1654: governed by the linearised Poincar\'e map and its Floquet multipliers: ${\rm  
1655: e}^{-4\pi}$, ${\rm e}^{2\pi e_1}$ and ${\rm e}^{-2\pi c_1}$ in the radial,  
1656: expanding and contracting directions, respectively, where, to leading order in  
1657: the~$\epsilon_i$'s, the period of~$P$ is~$2\pi$. The number~$n_1$  
1658: is unchanged by the weakly broken symmetry, and so, to leading order, this part  
1659: of the map is unchanged. Third, the trajectory travels from $H_1^P\to  
1660: H_1^\smallout$ in less than one circuit around~$P$ and again is not too close  
1661: to~$P$, so the fully symmetric flow yields an adequate approximate of the true  
1662: flow. Composing these three parts yields~(\ref{eq:phi1map}), to  
1663: leading order.  
1664:   
1665: Local maps $\phi_2:H_2^\smallin\to H_2^\smallout$ and  
1666: $\phi_3:H_3^\smallin\to H_3^\smallout$ are obtained similarly:  
1667:  \begin{equation}  
1668:  \phi_2(r_1=h,\theta_1,\xi_2,x_3)=\left(  
1669:             h\left|\frac{x_3}{h}\right|^{\delta_2},  
1670:             \theta_1  -\frac{1}{e_2}\ln\left|\frac{x_3}{h}\right|,  
1671:             \xi_2\left|\frac{x_3}{h}\right|^{\gamma_2},  
1672:             h\sgn(x_3)\right)\,,  
1673:  \label{eq:phi2map}  
1674:  \end{equation}  
1675:  \begin{equation}  
1676:  \phi_3(r_1,\theta_1,x_2,\xi_3)=\left(  
1677:             h,  
1678:             \theta_1 -\frac{1}{e_3}\ln\left(\frac{r_1}{h}\right) ,  
1679:             h\sgn(x_2)\left(\frac{r_1}{h}\right)^{\delta_3},  
1680:             \xi_3\left(\frac{r_1}{h}\right)^{\gamma_3}\right)\,,  
1681:  \label{eq:phi3map}  
1682:  \end{equation}  
1683: where $c_i$ and $e_i$ are the absolute  
1684: values of the real part of the contracting and expanding eigenvalues  
1685: at~$+E_i$, $\delta_i=c_i/e_i$, $\gamma_i=2/e_i$, and  $|x_2|=h$. As for $\phi_{1}$, the radial eigenvalues 
1686: and the angular speeds are chosen to be $-2$ and $1$.  
1687:  
1688: \subsection{Global maps}  
1689:   
1690: The global map $\Psi_{12}:H_1^\smallout\to H_2^\smallin$ takes orbits from a  
1691: neighbourhood of $P$ to a neighbourhood of $+E_2$. We write  
1692:  \begin{equation*}  
1693:  \Psi_{12}(\xi_1,\theta_1,x_2=h,x_3)=  
1694:  (\tilde{r}_1=h, \tilde{\theta}_1,  
1695:  \tilde{\xi}_2, \tilde{x}_3)  
1696:  \end{equation*}  
1697: and initially do not include symmetry-breaking effects. The unstable  
1698: manifold of $P$ is two-dimensional and, locally, intersects  $H_1^\smallout$ at  
1699:  \begin{equation}  
1700:  {\cal W}^\smallu(P) \cap H_1^\smallout = \left\{(\xi_1,\theta_1,x_2,x_3):  
1701:                                             \xi_1=0,  
1702:                                             0\leq \theta_1< 2\pi,  
1703:                                             x_2=\pm h,  
1704:                                             x_3=0 \right\}\,  
1705:  \label{eq:WuP}  
1706:  \end{equation}  
1707: The manifold ${\cal W}^\smallu(P)$ has two branches: the {\em positive} branch  
1708: intersects $H_1^\smallout$ with $x_2=h$ and the {\em negative} branch  
1709: intersects $H_1^\smallout$ with $x_2=-h$. The positive branch forms a  
1710: connection from $P$ to $+E_2$ and is the solution we now linearise about,  
1711: while the negative branch forms a connection from $P$ to $-E_2$ and will be  
1712: discussed later. The positive branch of ${\cal W}^\smallu(P)$ intersects  
1713: $H_2^\smallin$ at   
1714:  \begin{equation}  
1715:  \left\{(r_1, \theta_1,\xi_2,x_3): r_1=h,  
1716:                   0 \leq \theta_1 < 2\pi,  
1717:                        \xi_2=\bar{\xi}_2,  
1718:                        x_3=0 \right\}\,  
1719:  \label{eq:WuP2}  
1720:  \end{equation}  
1721: where $\bar{\xi}_2$ is a small constant. The $\kappa_1$  
1722: symmetry forces the heteroclinic orbit corresponding to the choice  
1723: $\theta_1$ in (\ref{eq:WuP}) to have an angular component in $H_2^\smallin$ of  
1724: $\theta_1+\bar{\theta}_1$ for some constant $\bar{\theta}_1$, i.e., the global  
1725: map acts on the angle as a rigid rotation.  Furthermore, trajectories that are  
1726: near but not on the unstable manifold of ${\cal W}^\smallu(P)$ have  
1727: $\tilde{\xi}_2$ and $\tilde{x}_3$ depending on the initial $\xi_1$ and $x_3$  
1728: but not on~$\theta_1$, while $\tilde{\theta}_1=\theta_1+\bar{\theta}_1$ where  
1729: $\bar{\theta}_1$ is a function of the initial $\xi_1$ and $x_3$. Equivariance  
1730: under $\kappa_3$ ensures that the subspace $x_3=0$ is invariant, that  
1731: $\tilde{x}_3$ is an odd function of $x_3$, and that ${\bar\theta}_1$ and  
1732: $\tilde{\xi}_2$ are even functions of~$x_3$. (The $\kappa_2$ symmetry has no  
1733: role in determining the form of $\Psi_{12}$ although it can be used to  
1734: construct a map from $P$ to $-E_2$ once $\Psi_{12}$ is known.) Writing a Taylor  
1735: series in the small quantities $\xi_1$ and $x_3$ therefore yields  
1736:  \begin{align*}  
1737:  {\tilde\theta}_1(\xi_1,\theta_1,x_3)=  
1738:  \theta_1 + {\bar\theta}_1({\xi}_1,x_3) &=  
1739:                                 \theta_1+{\bar\theta}_1(0,0)+\mbox{h.o.t.},\\  
1740:  \tilde{\xi}_2({\xi}_1,x_3) &=  
1741:                                        \tilde{\xi}_2(0,0)+\mbox{h.o.t.},\\  
1742:  \tilde{x}_3({\xi}_1,x_3) &=  
1743:  \frac{\partial\tilde{x}_3}{\partial x_3}(0,0)\,x_3  
1744:  +\mbox{h.o.t.},  
1745:  \end{align*}  
1746: where \hbox{h.o.t.} denotes higher order terms. Effectively, so long as  
1747: ${\bar\theta}_1$ and $\tilde{\xi}_2$ are non-zero, they can be replaced by  
1748: constants, while $\tilde{x}_3$ depends linearly on~$x_3$.   We write  
1749: $A_1=\frac{\partial\tilde{x}_3}{\partial x_3}(0,0)$ and  
1750: $B_1=\tilde{\xi}_2(0,0)$, and note that $A_1>0$ since the region of phase space  
1751: with $x_3>0$ is dynamically invariant.  
1752:   
1753: The effect of weak symmetry breaking on these expressions is as follows.  
1754: First, the symmetry $x_3\to-x_3$ is broken by including  
1755: terms that are odd in $x_3$ in the expressions for $\tilde{\theta}_1$ and  
1756: $\tilde{\xi}_2$, and terms that are even in $x_3$ in the expression for  
1757: $\tilde{x}_3$. We multiply all such terms by an overall  
1758: factor of~$\epsilon_3$, which is a real constant that controls the magnitude of  
1759: the breaking of the $\kappa_3$ symmetry. Then the lowest order contribution to  
1760: $\tilde{\theta}_1$ and $\tilde{\xi}_2$ will be a term in $\epsilon_3x_3$ while  
1761: $\tilde{x}_3$ will pick up a term linear in~$\epsilon_3$. At leading order all  
1762: quadratic terms can be dropped, so the only new term is one linear  
1763: in~$\epsilon_3$ in the expression for~$\tilde{x}_3$. Second, breaking  
1764: the~$\kappa_1$~symmetry will result in a weak dependence of all the  
1765: coefficients on~$\theta_1$, with the dependence being periodic in that  
1766: variable. We introduce the parameter $\epsilon_1$, which is a real constant  
1767: that multiplies all terms that break the $\kappa_1$ symmetry and that controls  
1768: the magnitude of the symmetry-breaking terms. For example, $A_1$ will become  
1769: $A_1+\epsilon_1f_1(\theta)$, with the caveat that this term must remain  
1770: positive, for all $\theta$ and $\epsilon_1$. Third, weakly breaking the  
1771: symmetry $x_2 \to -x_2$ will not affect the form of this map.  
1772:   
1773: Putting all this together results in a map  
1774:  $\Psi_{12}:H_1^\smallout\to H_2^\smallin$:  
1775:  \begin{align}  
1776:  \Psi_{12}(\xi_1,\theta_1,x_2=h,x_3)=\big(&  
1777:             \tilde{r}_1=h,  
1778:             \tilde{\theta}_1=\theta_1 + \Phi_1,  
1779:             \tilde{\xi}_2=B_1,\nonumber\\  
1780:            &\tilde{x}_3=A_1 x_3 + \epsilon_3  
1781:                        + \epsilon_1x_3f_1(\theta_1)  
1782:                        + \epsilon_1\epsilon_3g_1(\theta_1)  
1783:             \big),  
1784:  \label{eq:psi12}  
1785:  \end{align}  
1786: where $\Phi_1$, $A_1$, $B_1$ are constants, and $f_1$, $g_1$ are  
1787: $2\pi$-periodic functions of $\theta_1$. The $\theta_1$ dependence  
1788: cannot be treated using Taylor series expansions, as $\theta_1$ is not a small  
1789: quantity. We explain below why some quadratic terms ($\epsilon_1x_3$ and  
1790: $\epsilon_1\epsilon_3$) need to be kept. 
1791:   
1792: Similarly, a map from $P$ to $-E_2$ can be constructed. This  
1793: has precisely the form of (\ref{eq:psi12}), except that it starts from  
1794: $x_2=-h$. Breaking of the $\kappa_2$~symmetry means coefficients in the  
1795: map will be slightly different but the map is unchanged at leading order.  
1796:   
1797: The map $\Psi_{23}:H_2^\smallout\to H_3^\smallin$ is calculated in a similar  
1798: way. In the fully symmetric case, we linearise about  
1799: $ {\cal W}^\smallu(+E_2)$, which intersects $H_2^\smallout$ at $(z_1=0,\xi_2=0,x_3=h)$ and  
1800: $H_3^\smallin$ at $(z_1=0,\xi_2=h,\xi_3=\bar{\xi}_3)$ where $\bar{\xi}_3$ is a  
1801: small constant. For orbits near $ {\cal W}^\smallu(+E_2)$, the value of  
1802: $\xi_2$ at $H_2^\smallout$ does not influence the final position to leading  
1803: order and $z_1$ at $ H_3^\smallin$ depends linearly on the values of $z_1$ at  
1804: $H_2^\smallout$: $\tilde{z}_1=A_2{\rm e}^{{\rm i}\Phi_2}z_1$ for real constants  
1805: $A_2>0$, $\Phi_2$. If the $\kappa_1$~symmetry is broken,  
1806: $ {\cal W}^\smallu(+E_2)$ leaves $H_2^\smallout$ with $z_1=0$ and arrives at $H_3^\smallin$ with  
1807: $z_1=\tilde\epsilon_1$, where $\tilde\epsilon_1= \epsilon_1(a_\smallr+{\rm  
1808: i}a_\smalli)$ for $a_\smallr$ and $a_\smalli$ real constants and $\epsilon_1$  
1809: as defined earlier. Writing the resulting map in terms of the real and  
1810: imaginary parts of $\tilde{z}_1$:  
1811:  \begin{align}  
1812:  \Psi_{23}(r_1,\theta_1,\xi_2,x_3=h)=\big(&  
1813:             \tilde{x}_1=\epsilon_1a_\smallr + A_2r_1\cos(\theta_1+\Phi_2),  
1814:             \nonumber\\  
1815:            &\tilde{y}_1=\epsilon_1a_\smalli + A_2r_1\sin(\theta_1+\Phi_2),  
1816:             \nonumber\\  
1817:            &\tilde{x}_2=h,  
1818:             \tilde{\xi}_3=B_2\big),  
1819:  \label{eq:psi23}  
1820:  \end{align}  
1821: where $a_\smallr$, $a_\smalli$, $A_2$, $B_2$ and $\Phi_2$ are real  
1822: constants determined by the global flow, and $A_2>0$. As in~$\Psi_{12}$, there  
1823: are $2\pi$-periodic functions of~$\theta_1$ in the map, but here the functions  
1824: are known explicitly because the $z_1$ variable is small throughout the  
1825: transition from $+E_2$ to~$+E_3$, and the dynamics of $z_1$ is well-approximated  
1826: by a scaled rotation. Similar maps can be obtained for the three connections  
1827: $-E_2 \to +E_3$ and $\pm E_2 \to -E_3$; although the coefficients will be  
1828: slightly different in each case, to lowest order we obtain the same map for  
1829: each of the other connections so long as the signs of the $x_3$  
1830: (resp.~$\tilde{x}_2$) components are chosen appropriately on the incoming  
1831: (resp.~outgoing) cross-sections (for example, the map from $-E_2$ to $-E_3$  
1832: will have $x_3=-h$ and $\tilde{x}_2=-h$).  
1833:   
1834: The global map $\Psi_{31}:H_3^\smallout\to H_1^\smallin$ is calculated in a  
1835: similar way:  
1836:  \begin{align}  
1837:  \Psi_{31}(r_1=h,\theta_1,x_2,\xi_3)=\big(&  
1838:            \tilde{\xi}_1=B_3,  
1839:            \tilde{\theta}_1 = \theta_1+\Phi_3,  
1840:            \nonumber\\  
1841:           &\tilde{x}_2 = A_3 x_2 + \epsilon_2  
1842:                        + \epsilon_1x_2f_3(\theta_1)  
1843:                        + \epsilon_1\epsilon_2g_3(\theta_1),  
1844:            \nonumber\\  
1845:           &\tilde{x}_3 = h\big),  
1846:  \label{eq:psi31}  
1847:  \end{align}  
1848: where $A_3$, $B_3$ and $\Phi_3$ are real constants, $f_3$ and $g_3$ are  
1849: $2\pi$-periodic functions of $\theta_1$, and $\epsilon_1$ controls the size of  
1850: the terms that break the $\kappa_1$ symmetry. The parameter $\epsilon_2$  
1851: introduced in $(\ref{eq:psi31})$ is analogous to $\epsilon_3$, and is a real  
1852: quantity that controls the size of all terms that break the $\kappa_2$  
1853: symmetry. Similarly to the case for $A_1$ argued above, we take  
1854: $A_3+\epsilon_1f_3(\theta_1)$ to be positive for all values of $\epsilon_1$. A  
1855: similar map can be defined near the connections from $-E_3$ to $P$, and will,  
1856: to leading order, be identical to (\ref{eq:psi31}) so long as $\tilde{x}_3=h$  
1857: is replaced by $\tilde{x}_3=-h$.  
1858:   
1859: The effect of each of the global maps defined above is, at leading order, to  
1860: rotate the angular variable by an order one amount that is independent of other  
1861: variables, to set the radial variable to a constant, and, in the absence of  
1862: symmetry-breaking, to scale the variable that measures proximity to the cycle.  
1863: Symmetry-breaking enters in two ways. First, it destroys  
1864: the invariant subspaces thus destroying some of the heteroclinic connections  
1865: that made up the cycle. Second, breaking the $\kappa_1$~rotation symmetry  
1866: allows $\theta_1$~dependence to enter into the maps, most importantly through  
1867: the variables $x_3$ in the $\Psi_{12}$~map and $x_2$ in the $\Psi_{31}$~map. It  
1868: is this $\theta_1$ dependence that allows the heteroclinic tangencies discussed  
1869: in Section~\ref{sec:description}.  
1870:  
1871: \subsection{Return maps} 
1872:  
1873: Return maps approximating the dynamics near the heteroclinic cycle can now be 
1874: computed by composing the local and global maps in an appropriate order. For  
1875: instance, to obtain the various forms of the map  
1876: $R :   H_3^\smallin \to  H_3^\smallin$ given by equations  
1877: (\ref{eq:R_nosb}--\ref{eq:R_sb1complex}) we calculate  
1878: $R\equiv \psi_{23} \circ \phi_2 \circ \psi_{12} \circ \phi_1 \circ  \psi_{31} \circ \phi_3$ 
1879: in the usual way. 
1880:  
1881:  
1882:  
1883:  
1884:  
1885: \section*{Acknowledgments}  
1886:  We thank Ian Melbourne, Edgar Knobloch and Jeff Porter for helpful 
1887: conversations. This research has been supported by grants from New Zealand 
1888: Institute for Mathematics and its Applications, University of Auckland Research 
1889: Council, London Mathematical Society and the Engineering and Physical Sciences 
1890: Research Council. 
1891:   
1892: \begin{thebibliography}{10}  
1893: \expandafter\ifx\csname url\endcsname\relax  
1894:   \def\url#1{\texttt{#1}}\fi  
1895: \expandafter\ifx\csname urlprefix\endcsname\relax\def\urlprefix{URL }\fi  
1896:                                                     
1897: \bibitem{F80}  
1898: Field, M., 1980,  
1899: Equivariant dynamical systems.  
1900: {\itshape Trans.~Am.~Math.~Soc.}, {\bfseries 259}, 185--205.  
1901:   
1902: \bibitem{GuHo88}  
1903: Guckenheimer, J. and Holmes, P., 1988,  
1904: Structurally stable heteroclinic cycles.  
1905: {\itshape Math. Proc. Camb. Phil. Soc.}, {\bfseries 103}, 189--192.  
1906:   
1907: \bibitem{M91}  
1908: Melbourne, I., 1991,  
1909: An example of a non-asymptotically stable attractor.  
1910: {\itshape Nonlinearity}, {\bfseries 4}, 835--844.  
1911:   
1912: \bibitem{KrMe95}  
1913: Krupa, M. and Melbourne, I., 1995,  
1914: Asymptotic stability of heteroclinic cycles in systems with symmetry.  
1915: {\itshape Ergod.~Th.~\& Dynam.~Sys.}, {\bfseries 15}, 121--147.  
1916:   
1917: \bibitem{KrMe04}  
1918: Krupa, M. and Melbourne, I., 2004,  
1919: Asymptotic stability of heteroclinic cycles in systems with symmetry, II.  
1920: {\itshape Proc. Roy. Soc. Edinburgh A}, {\bfseries 134}, 1177--1197.  
1921:   
1922: \bibitem{KS94}  
1923: Kirk, V. and Silber, M., 1994,  
1924: A competition between heteroclinic cycles.  
1925: {\itshape Nonlinearity}, {\bfseries 7}, 1605--1621.  
1926:   
1927: \bibitem{AC98}    
1928: Ashwin, P. and Chossat, P., 1998,  
1929: Attractors for robust heteroclinic cycles with continua of connections.  
1930: {\itshape J.~Nonlinear Sci.}, {\bfseries 8}, 103--129.  
1931:   
1932: \bibitem{ACL05}  
1933: Aguiar, M., Castro, S. and Labouriau, I., 2005,  
1934: Dynamics near a heteroclinic network.  
1935: {\itshape Nonlinearity}, {\bfseries 18}, 391--414.  
1936:   
1937: \bibitem{PD05}  
1938: Postlethwaite, C.M. and Dawes, J.H.P., 2005,  
1939: Regular and irregular cycling near a heteroclinic network.  
1940: {\itshape Nonlinearity}, {\bfseries 18}, 1477--1509.  
1941:   
1942: \bibitem{SC92}  
1943: Scheel, A. and Chossat, P., 1992,  
1944: Bifurcation d'orbites p\'{e}riodiques \`{a} partir d'un cycle  
1945: homoclinic sym\'{e}trique.  
1946: {\itshape C. R. Acad. Sci. Paris}, {\bfseries 341}, 49--54.  
1947:   
1948: \bibitem{CKMS}  
1949: Chossat, P., Krupa, M., Melbourne, I. and Scheel, A., 1997,  
1950: Transverse bifurcations of heteroclinic cycles in symmetric systems.  
1951: {\itshape Physica D}, {\bfseries 100}, 85--100.  
1952:   
1953: \bibitem{PD05b}  
1954: Postlethwaite, C.M. and Dawes, J.H.P., 2006,  
1955: A codimension-two resonant bifurcation from a heteroclinic  
1956: cycle with complex eigenvalues.  
1957: {\itshape Dyn. Syst. Int. J.}, {\bfseries 21},  313--336.  
1958:   
1959: \bibitem{BuHe80}  
1960: Busse, F.H. and Heikes, K.E., 1980,  
1961: Convection in a rotating layer -- simple case of turbulence.  
1962: {\itshape Science}, {\bfseries 208}, 173--175, 1980.  
1963:   
1964: \bibitem{PJ88}  
1965: Proctor, M.R.E. and Jones, C.A., 1988,  
1966: The interaction of two spatially resonant patterns in  
1967: thermal convection. I. Exact 1:2 resonance.  
1968: {\itshape J. Fluid Mech.}, {\bfseries 188}, 301--335.  
1969:  
1970: \bibitem{NTDX}  
1971: Nore, C., Tuckerman, L.S., Daube, O., and Xin, S., 2003, 
1972: The 1:2 mode interaction in exactly counter-rotating von  
1973: K\'{a}rm\'{a}n swirling flow. 
1974: {\itshape J. Fluid Mech.}, {\bfseries 477}, 51-88.  
1975:  
1976: \bibitem{AGH88}  
1977: Armbruster, D., Guckenheimer, J., Holmes, P., 1988,  
1978: Heteroclinic cycles and modulated travelling waves in systems with $O(2)$ symmetry.  
1979: {\itshape Physica D}, {\bfseries 29}, 257--282.  
1980:   
1981: \bibitem{Me89}  
1982: Melbourne, I., 1989,  
1983: Intermittency as a codimension three phenomenon  
1984: {\itshape J. Dyn. Stab. Sys.}, {\bfseries 1}, 347--367.  
1985:   
1986: \bibitem{Ch93}  
1987: Chossat, P., 1993,  
1988: Forced reflectional symmetry breaking of an $O(2)$-symmetric homoclinic cycle.  
1989: {\itshape Nonlinearity}, {\bfseries 6}, 723--731.  
1990:   
1991: \bibitem{SaSc95}  
1992: Sandstede, B. and Scheel, A., 1995,  
1993: Forced symmetry breaking of homoclinic cycles.  
1994: {\itshape Nonlinearity}, {\bfseries 8}, 333--365.  
1995:   
1996: \bibitem{MPR}  
1997: Melbourne, I., Proctor, M.R.E. and Rucklidge, A.M., 2001,  
1998: A heteroclinic model of geodynamo reversals and excursions.  
1999: In: P.~Chossat, D.~Armbruster and I.~Oprea (Eds)  
2000: {\itshape Dynamo and Dynamics, a Mathematical Challenge}  
2001: (Dordrecht: Kluwer), pp.~363--370.  
2002:   
2003: \bibitem{StHo90}  
2004: Stone, E. and Holmes, P., 1990,  
2005: Random perturbations of heteroclinic attractors.  
2006: {\itshape SIAM J. Appl. Math.}, {\bfseries 50}, 726--743.  
2007:   
2008: \bibitem{ASK03}         
2009: Armbruster, D., Stone, E., Kirk, V., 2003,  
2010: Noisy heteroclinic networks.  
2011: {\itshape Chaos}, {\bfseries 13}, 71--79.  
2012:  
2013: \bibitem{NMQ} 
2014: Nore, C., Moisy, F., and Quartier, L., 2005, 
2015: Experimental observation of near-heteroclinic cycles in the von 
2016: K\'{a}rm\'{a}n swirling flow. 
2017: {\itshape Phys.~Fluids}, {\bfseries 17}, 064103. 
2018:   
2019: \bibitem{ClKn94}  
2020: Clune, T. and Knobloch, E., 1994,  
2021: Pattern selection in three-dimensional magnetoconvection.  
2022: {\itshape Physica D}, {\bfseries 74}, 151--176.  
2023:   
2024: \bibitem{KLS05}  
2025: Kirk, V., Lane, E.,  Silber, M., 2007,  
2026: A mechanism for switching in a heteroclinic network.  
2027: In preparation. 
2028:   
2029: \bibitem{ARS04a}  
2030: Ashwin, P., Rucklidge, A.M. and Sturman, R., 2004,  
2031: Two-state intermittency near a symmetric interaction of saddle-node and Hopf  
2032: bifurcations: a case study from dynamo theory.  
2033: {\itshape Physica D}, {\bfseries 194}, 30--48.  
2034:   
2035: \bibitem{RM95}  
2036: Rucklidge, A.M. and Matthews, P.C., 1995,  
2037: The shearing instability in magnetoconvection.  
2038: In: A.~Brandt and H.J.S.~Fernando (Eds)  
2039: {\itshape Double-Diffusive Convection}  
2040: (Washington: American Geophysical Union), pp.~171--184.  
2041:   
2042: \bibitem{MRWP96}  
2043: Matthews, P.C., Rucklidge, A.M., Weiss, N.O. and Proctor, M.R.E., 1996,  
2044: The three-dimensional development of the shearing instability of convection.  
2045: {\itshape Phys. Fluids}, {\bfseries 8}, 1350--1352.  
2046:   
2047: \bibitem{R01}  
2048: Rucklidge, A.M., 2001,  
2049: Global bifurcations in the Takens--Bogdanov normal form with $D_4$  
2050: symmetry near the $O(2)$ limit.  
2051: {\itshape Phys Lett~A}, {\bfseries 284}, 99--111.  
2052:   
2053: \bibitem{AFRS03}  
2054: Ashwin, P., Field, M., Rucklidge, A.M. and Sturman, R., 2003,  
2055: Phase resetting effects for robust cycles between chaotic sets.  
2056: {\itshape Chaos}, {\bfseries 13}, 973--981.  
2057:   
2058: \bibitem{ARS04b}  
2059: Ashwin, P., Rucklidge, A.M. and Sturman, R., 2004,  
2060: Cycling chaotic attractors in two models for dynamics with invariant subspaces.  
2061: {\itshape Chaos}, {\bfseries 14}, 571--582.  
2062:  
2063: \bibitem{Rademacher} 
2064: Rademacher, J.D.M., 2005, 
2065: Homoclinic orbits near 
2066: heteroclinic cycles with one equilibrium and one periodic 
2067: orbit. {\itshape J. Diff. Eqns.}, {\bfseries 218}, 390--443. 
2068:  
2069: \bibitem{KROCK} 
2070: Champneys, A.R.,  Kirk, V.,  Knobloch, E., Oldeman, B., and Rademacher, J., 2007, 
2071: Unfolding a tangent equilibrium-to-periodic heteroclinic cycle. In preparation. 
2072:   
2073: \bibitem{GS84}  
2074: Glendinning, P. and Sparrow, C., 1984,  
2075: Local and global behaviour near homoclinic orbits  
2076: {\itshape J.~Stat. Phys.}, {\bfseries 35}, 645--696.  
2077:   
2078: \bibitem{NumRecipes}  
2079: Press, W.H., Flannery, B.P., Teukolsky, S.A. and Vetterling, W.T., 1986,  
2080: Numerical Recipes -- the Art of Scientific Computing.  
2081: Cambridge University Press, Cambridge.  
2082:   
2083: \bibitem{PC89}  
2084: Parker, T.S. and Chua, L.O., 1989,  
2085: Practical Numerical Algorithms for Chaotic Systems.  
2086: Springer, New York.  
2087:  
2088: \bibitem{GH86}  
2089: Guckenheimer, J. and Holmes, P., 1986,  
2090: Nonlinear Oscillations, Dynamical Systems, and Bifurcations of Vector Fields  
2091: (second edition).  
2092: Springer-Verlag, New York.   
2093: \end{thebibliography}  
2094:   
2095: \end{document}  
2096:   
2097:   
2098:   
2099:   
2100:   
2101:   
2102:   
2103:   
2104: