1: \documentclass[11pt, epsfig]{article}

2: \usepackage{epsfig, amsmath, amssymb, amsthm}

3: \parindent=1.5em

4: \parskip10pt

5: \textwidth=6.5in

6: \topmargin= 0.0in

7: \oddsidemargin=0in

8: \textheight=8.4in

9: 

10: \def\theequation{\arabic{section}.\arabic{equation}}\def\theequation{\arabic{section}.\arabic{equation}}

11: \renewcommand{\theequation}{\thesection.\arabic{equation}}

12: 

13: \def\p{\partial}

14: \def\Re{{\rm I}\!{\rm R}}

15: \def\Na{{\rm I}\!{\rm N}}

16: \def\lbl{\label}

17: \def\be{\begin{equation}}

18: \def\ee{\end{equation}}

19: \def\lbl{\label}

20: \def\qed{\sq}

21: \def\bu{\bf u}

22: \def\bv{\bf v}

23: \def\tl{\tilde}

24: \def\mG{\mathcal{G}}

25: \def\mI{\mathcal{I}}

26:                                                                         

27: \title{Multimodal oscillations in systems with strong contraction}

28: \author{ 

29: Georgi Medvedev and Yun Yoo \thanks{

30: Department of Mathematics, Drexel University, 3141 Chestnut Street,

31: Philadelphia, PA 19104, {\it medvedev@drexel.edu}

32:  }

33:  }

34: 

35: \begin{document}

36: \maketitle

37: 

38: \begin{abstract}

39: One- and two-parameter families of flows in $\Re^3$

40: near an Andronov-Hopf bifurcation (AHB) are investigated in this work. 

41: We identify conditions on the 

42: global vector field, which yield a rich family of multimodal orbits 

43: passing close to a weakly unstable

44: saddle-focus and  perform a detailed asymptotic analysis of the

45: trajectories in the vicinity of the saddle-focus. Our analysis covers both

46: cases of sub- and supercritical AHB. For the supercritical case, we find

47: that the periodic orbits born from the AHB are bimodal when viewed in the

48: frame of coordinates generated by the linearization about the bifurcating

49: equilibrium. If the AHB is subcritical, it

50: is accompanied by the  appearance of multimodal orbits, which consist

51: of long series of nearly harmonic oscillations separated by large

52: amplitude spikes. We analyze the dependence of the interspike intervals

53: (which can be extremely long) on the control parameters. In particular,

54: we show that the interspike intervals grow logarithmically as the boundary

55: between regions of sub- and supercritical AHB is approached in the parameter

56: space. We also identify a window of complex and possibly chaotic oscillations

57: near the boundary between the regions of sub- and supercritical AHB

58: and explain the mechanism generating these oscillations. This work is motivated 

59: by the numerical results for a finite-dimensional approximation of a free 

60: boundary problem modeling solid fuel combustion.

61: \end{abstract}

62: 

63: 

64: \section{Introduction}

65: 

66: Systems of differential equations in both finite- and infinite-dimensional settings

67: close to an AHB have been subject to intense research 

68: due to their dynamical complexity and importance in applications. The latter

69: range from models in fluid dynamics \cite{MarsdenMcCracken} to those in the life sciences, 

70: in particular, in computational neuroscience 

71: \cite{BER, DK05, DRSE, GW, KR83, MC, RE89, ROWK}.

72: When the proximity to the AHB 

73: coincides with certain global properties of the vector field,

74: it may result in a very complex dynamics \cite{BS, DENG90, DENG95, GW, Hirschberg_Knobloch}.

75: The formation of the Smale horseshoes in systems 

76: with a homoclinic orbit to a saddle-focus equilibrium  

77: provides one of the most representative examples of this type 

78: \cite{AAIS}.

79: Canard explosion in relaxation systems affords another example \cite{CDD, KS01b}. 

80: Recent studies of relaxation systems,

81: motivated mainly by applications in the life sciences, have revealed that

82: the proximity to an AHB has a significant impact on the system

83: dynamics. It manifests itself as a family of multimodal periodic solution

84: that are composed of large-amplitude relaxation oscillations 

85: (generated by the global structure of the vector field)

86: and small-amplitude nearly harmonic oscillations (generated by the vector field near the

87: equilibrium close to the AHB)

88: \cite{DRSE, KOP, MC, MSLG, PSS, VV} (see Figure \ref{f.1}). 

89: These families of solutions possess rich bifurcation structure.

90: A remarkable example of an infinite-dimensional system close to the AHB

91: has been recently studied by Frankel and Roytburd 

92: \cite{FR05a, FR05b, FR03, FR95, FR94}. They derived and systematically studied 

93: a model of solid fuel combustion 

94: in the form of a free boundary problem for a $1D$ heat equation with

95: nonlinear conditions imposed at the free boundary modeling the interface between 

96: solid fuel mixture and a solid product.

97: The investigations of this model revealed a wealth of spatial-temporal patterns

98: ranging from a uniform front propagation to periodic and aperiodic front oscillations. 

99: The transitions between different dynamical regimes involve a variety of nontrivial bifurcation

100: phenomena including period-doubling cascades, period-adding sequences, 

101: and windows of chaotic dynamics. To elucidate the mechanisms responsible for 

102: different dynamical regimes and transitions between them, Frankel and Roytburd

103: employed pseudo-spectral techniques to derive a finite-dimensional approximation

104: for the interface dynamics in the free boundary problem \cite{FR95}. 

105: As shown in \cite{FKRT, FR95},

106: a system of three ordinary differential equations captured the essential features of 

107: the bifurcation structure of the infinite-dimensional problem.

108: The numerical bifurcation analysis of the finite-dimensional approximation revealed a rich

109: family of multimodal periodic solutions similar to those reported in the context of relaxation

110: systems near the AHB \cite{FKRT}. The bifurcation diagrams presented in \cite{FKRT}

111: and in \cite{MC} share a striking similarity, despite the absence of any apparent common 

112: structures in the underlying models (except to the proximity to the AHB). 

113: In particular, in both models, topologically distinct multimodal periodic solutions

114: are located on isolas, closed curves in the parameter space.

115: The methods of analysis of the mixed-mode solutions in \cite{KOP, MC, MSLG, WESCH} used

116: in an essential way the relaxation structure present in these problems. These approaches

117: can not be applied directly to analyzing the model in \cite{FKRT}, because it is not a priori clear 

118: what creates the separation of the time scales in this model, in spite of the evident fast-slow

119: character of the numerical solutions. This is partly due to the spectral method, which

120: was used to derive the system of equations in \cite{FKRT}: while it has captured well

121: the finite-dimensional attractor of the interface dynamics, it has disguised 

122: the structure of the physical model. One of the goals of the present paper is to identify 

123: the structure responsible for the generation of the multimodal oscillations 

124: in a finite-dimensional model for the interface dynamics and 

125: to relate it to those studied in the context of relaxation oscillations.

126: 

127: The family of flows in \cite{FKRT} includes 

128: in a natural way two types of the AHBs. Depending on the parameter values,

129: the equilibrium of the system of ordinary differential equations in \cite{FKRT}  

130: undergoes either a sub- or a supercritical AHB. A similar

131: situation is encountered in certain neuronal models (see, e.g., \cite{DK05,OP}).

132: In either case, the global multimodal periodic solutions are created after

133: the AHB. However, 

134: in the case of a supercritical bifurcation, they are preceded by a series of 

135: period-doubling bifurcations of small amplitude limit cycles, arising from

136: the AHB. 

137: On the other hand, in the subcritical case, the AHB gives rise to 

138: multimodal solutions, whose lengths and time intervals between successive large

139: amplitude oscillations can be very long. In the present paper,

140: we perform a detailed asymptotic analysis of the trajectories in a class of systems motivated

141: by the problem in \cite{FKRT}. Our analysis includes both cases of the sub- and supercritical AHBs. 

142: We also investigate the dynamical regimes arising near the border between the regions of sub- and

143: supercritical AHB. This region in the parameter space contains a number of nontrivial oscillatory

144: patterns including multimodal trajectories with substantial time intervals between successive

145: spikes, irregular, and possibly chaotic oscillations, as well as a family of periodic orbits 

146: undergoing a cascade of period-doubling bifurcations. Our analysis shows that these dynamical

147: patterns and the order in which they appear under the variation of the control parameters are

148: independent on the details of the model, but are characteristic to the transition from sub-

149: to supercritical AHB.

150: 

151: The outline of the paper is as follows. After introducing the model and rewriting it in the

152: normal coordinates, we present a set of the numerical experiments to be explained in the remainder of the paper.

153: Then we state our results for each of the following cases: supercritical AHB,

154: subcritical AHB, and the transition layer between the regions of sub- and supercritical AHB. In Section 3, we analyze

155: the local behavior of trajectories near a weakly unstable saddle-focus. 

156: The local expansions used in this section are similar to those used in \cite{CMP} for analyzing 

157: the AHB using the method of averaging. However, rather than establishing existence of periodic solutions, the goal

158: of the present section is to approximate the trajectories near the saddle-focus. For this,

159: after rescaling the variables and recasting the system into cylindrical coordinates, we reduce

160: the system dynamics to a $2D$ slow manifold. By integrating the leading order approximation of the

161: reduced system, we obtain necessary information about the local behavior of trajectories. The results

162: of this section are summarized in Theorem 3.1. In Section 4, we  study oscillatory patterns generated by the

163: class of systems under investigation. It is divided into three subsections devoted to the oscillations

164: triggered by the supercritical AHB (\S 4.1), subcritical AHB (\S 4.2), and those found in the transition region

165: between sub- and supercritical AHB (\S 4.3). In the supercritical case, we show that the oscillations just

166: after the AHB are already bimodal. Generically, one needs to use two harmonics to describe the

167: limit cycle born at the supercritical AHB in $\Re^3$. We give a geometric interpretation of this

168: effect, which we call the frequency doubling, due to the fact that the second frequency is twice as large

169: as that predicted by the Hopf Bifurcation Theorem \cite{MarsdenMcCracken}. 

170: We also compute the curvature and the torsion of the periodic orbit as a curve

171: in $\Re^3$. The latter is useful for the geometric explanation of the frequency doubling. In Subsection 4.2,

172: we study the subcritical case. We show that under two general assumptions on the global vector field,

173: the presence of the return mechanism ({\bf G1}) and the strong contraction property ({\bf G2}), the

174: subcritical AHB results in sustained multimodal oscillations. Even though the oscillations may not be periodic

175: (our assumptions do not warrant periodicity), the time intervals between consecutive spikes of the resultant motion comply to

176: the uniform bounds given in Theorem 4.1. This subsection also contains the definitions of the multimodal oscillations

177: and the precise formulation of the assumptions on the global vector field. The proof of Theorem 4.1

178: is relegated to Section 5. Subsection 4.3 describes the

179: transition between sub- and supercritical AHB. This transition contains a distinct bifurcation scenario, when

180: a small positive real part of the complex conjugate pair of eigenvalues is positive and fixed and the first 

181: Lyapunov coefficient changes sign. The numerical simulations show that this results 

182: in the formation of the chaotic attractor followed

183: by the reverse period-doubling cascade. To explain this bifurcation sequence, we derive a $1D$ first return map.

184: The asymptotic analysis in this subsection is complemented by numerical extension of the map to the region

185: inaccessible by local asymptotic expansions. The first return map obtained by the combination of the analytic

186: and numerical techniques reveals the principal traits of the transition from sub- to supercritical AHB.

187: Finally, the discussion of the results of the present paper and their relation to the previous work is

188: given in Section 6.  

189: 

190: \section{The model and numerical results}

191: In the present section, we formulate the model and present a set of

192: numerical results, which motivated our study. The following system of three ODEs was derived in 

193: \cite{FKRT}, as a finite-dimensional

194: approximation for the interface dynamics in a free boundary problem modeling solid fuel 

195: combustion:

196: \begin{eqnarray} \lbl{6.1}

197: \dot v_1 &=& \frac{3(v_3+v_2-v_1)-\nu k(v_1)-v_1^2}{ \nu k^\prime(v_1)},\\ \lbl{6.2}

198: \dot v_2 &=& v_3-v_1,\\ \lbl{6.3}

199: \dot v_3 &=& 9(v_1-v_3)-6v_2+\nu(v_1+1)k(v_1)+2v_1^2.

200: \end{eqnarray}

201: Here, $v_1(t)$ approximates the velocity of the interface between the solid fuel and the

202: burnt material. Functions $v_2(t)$ and $v_3(t)$ are

203: the first two coefficients in the series expansion for the spatial temperature profile

204: with respect to the basis of Chebyshev-Laguerre polynomials. Equations (\ref{6.1})-(\ref{6.3})

205: are obtained by projecting the original infinite-dimensional problem onto a finite-dimensional

206: function space and using the method of collocations for determining the unknown coefficients.

207: An important ingredient of the model, nonlinear kinetic function $k(v)$ is given by

208: \be\lbl{6.4}

209: k(v)={(1-v)^{p}-(1-v)^{-1}\over p+1}.

210: \ee

211: \begin{figure}

212: \begin{center}

213: {\bf a}\epsfig{figure=f1a.eps, height=1.8in, width=2.5in, angle=0} 

214: {\bf b}\epsfig{figure=f1b.eps, height=1.8in, width=2.5in, angle=0}\\

215: {\bf c}\epsfig{figure=f1c.eps, height=1.8in, width=2.5in, angle=0}

216: {\bf d}\epsfig{figure=f1d.eps, height=1.8in, width=2.5in, angle=0}\\

217: {\bf e}\epsfig{figure=f1e.eps, height=1.8in, width=2.5in, angle=0}

218: {\bf f}\epsfig{figure=f1f.eps, height=1.8in, width=2.5in, angle=0}\\

219: {\bf g}\epsfig{figure=f1g.eps, height=1.8in, width=2.5in, angle=0}

220: {\bf h}\epsfig{figure=f1h.eps, height=1.8in, width=2.5in, angle=0}

221: \end{center}

222: \caption{Stable periodic solutions of system of differential equations (\ref{6.5}). Plots

223: in a-d show solutions near a supercritical AHB ($p=2.2$) and those in e-h

224: correspond to the subcritical case ($p=3$). The time series of $x_1$ are presented in the

225: left column and those of $x_2$ are given in the right column. The time series for 

226: $x_3$ and $x_2$ are qualitatively similar. 

227: }\label{f.1}

228: \end{figure} 

229: It reflects the dependence of the velocity of propagation on the temperature in the front. This 

230: relation is very complex and is not completely understood at present. To account for a range of 

231: possible kinetic mechanisms, the model includes two control parameters: $\nu$ and $p$.

232: Both the interface velocity and the temperature profile are calculated in the uniformly

233: moving frame of reference. Therefore, system of equations (\ref{6.1})-(\ref{6.3}) describes

234: the deviations of the interface dynamics from that of a front traveling with constant

235: speed. In particular, periodic and aperiodic oscillations generated by (\ref{6.1})-(\ref{6.3})

236: correspond to complex spatio-temporal patterns in the infinite-dimensional model. The numerical

237: results presented in \cite{FKRT} show a remarkable similarity between the complex patterns 

238: generated by the infinite- and finite-dimensional models 

239: and between the scenarios for transitions between different regimes in both models.

240: For more information about the derivation of the free-boundary problem for solid fuel combustion

241: and its finite-dimensional approximation, we refer the reader to \cite{FKRT} and bibliography

242: therein.  

243: 

244: 

245: A simple inspection of (\ref{6.1})-(\ref{6.3}) shows that it has an equilibrium at the origin $O=(0,0,0)^T$ for all

246: values of $\nu$ and $p>0$. We linearize (\ref{6.1})-(\ref{6.3}) about the equilibrium at the origin

247: $$

248: \dot v = A(\nu)v+ \dots,\quad\mbox{where}\quad

249: A(\nu) =\left( \begin{array}{ccc}

250: \ \frac{3-\nu}{\nu} &  \frac {-3}{\nu}  &  \frac {-3}{\nu} \\

251: -1 & 0 & 1 \\

252: 9-\nu & -6  &  -9 \end{array} \right),

253: \quad v=\left(v_1, v_2, v_3\right)^T.

254: $$

255: Near $\nu_{AH}={1\over 3}$, Jacobian matrix $A(\nu)$ has a negative eigenvalue $-1$ and a pair of complex

256: conjugate eigenvalues $\lambda=a(\nu) + ib(\nu)$ and $\bar\lambda$. The latter crosses the imaginary

257: axis transversally at $\nu=\nu_{AH}$:

258: $$

259: a\left(\nu_{AH}\right)=0\qquad\mbox{and}\qquad a^\prime\left(\nu_{AH}\right)\ne 0.

260: $$

261: Therefore, the equilibrium of (\ref{6.1})-(\ref{6.3}) undergoes an AHB. In the  neighborhood of $\nu_{AH}$,

262: $\alpha=a(\nu)$ defines a smooth invertible function. We shall use $\alpha$ as a new control parameter.

263: After a linear coordinate transformation,

264: system of equations (\ref{6.1})-(\ref{6.3}) has the following form:  

265: \be\lbl{6.5} 

266: \dot x = 

267: \left(\begin{array}{ccc}

268:                       -1 & 0 & 0\\ 

269:                        0 & \alpha & -b(\alpha)\\ 

270:                        0 & b(\alpha) &  \alpha

271:       \end{array}\right) x + h(x,\alpha),

272: \quad x=\left(x_1, x_2, x_3\right)^T, \quad h=\left(h_1, h_2, h_3\right)^T.

273: \ee

274: Here, by $b(\alpha):=b\left(a^{-1}(\alpha)\right)$ we denote the imaginary part of $\lambda$ as a function of

275: the new control parameter. Nonlinear function $h:\Re^3\times \Re\to \Re^3$  is a smooth function 

276: such that $h\left(0,\alpha\right)=0$ and ${\p h (0,\alpha)\over \p x}=0$

277: for values of $\alpha$ close to $0$. More precisely, $h$ stands for a family of functions parametrized by $p$

278: (see (\ref{6.4})). To keep the notation simple, we omit the dependence of $h$ on the second control parameter $p$.

279: \begin{figure}

280: \begin{center}

281: {\bf a}\epsfig{figure=f2a.eps, height=2.75in, width=3.0in, angle=0}

282: {\bf b}\epsfig{figure=f2b.eps, height=2.75in, width=3.0in, angle=0}

283: \end{center} 

284: \caption{{\bf a}. A periodic trajectory of the original system in the regime close to a subcritical AHB.

285: Cross-sections $\Sigma^+$ and $\Sigma^-$ are defined as follows:

286: $\Sigma^+=\left\{\left(x_1, \rho\cos\theta, \rho\sin\theta\right):\; \rho=0.1,\; x_1\in [-0.02, 0.02],\;

287: \theta\in [0,2\pi)\right\}$. $\Sigma^-$ is orthogonal to $x_1-$axis and is set at 

288: $x_1=-0.65$.

289: The numerical results shown in {\bf b} test the return mechanism and the strong contraction

290: property for (\ref{6.5}) (see {\bf (G1)} and {\bf (G2)} in the text). For this, we

291: covered $\Sigma^+$ with

292: a uniform mesh  $\Sigma_{N_i,N_j}^+=\left\{ \xi_{ij}\right\},$ $i=1,2,\dots,N_i,$ $j=1,2,\dots, N_j$. 

293: Using points $\xi_{ij}$ as initial conditions, we integrated

294: (\ref{6.5}) until the corresponding trajectories hit the second cross-section, $\Sigma^-$,

295: transverse to the stable manifold. The top plots in {\bf b} shows the image of $\Sigma_{N_i,N_j}$ 

296: in $\Sigma^-$ under the flow-defined map.

297: To test the strong contraction property, we repeated this experiment by taking 

298: $\Sigma^-$ progressively closer to the origin (the location of $\Sigma^-$ is 

299: indicated in each plot). The top five images in {\bf b} clearly indicate

300: that the projection of  the vector field onto $\Sigma^-$ is strongly contracting. The bottom plot

301: shows that the trajectories starting from $\Sigma^+$ reach a small neighborhood of the 

302: unstable equilibrium.

303: }\label{f.2}

304: \end{figure} 

305: It turns out that in the range of parameters of interest, the AHB of the equilibrium for $\alpha=0$ 

306: can be either subcritical, or 

307: supercritical, or degenerate depending on the value of $p$.  This gives rise to several 

308: qualitatively distinct oscillatory regimes 

309: generated by (\ref{6.5}) for small values of $\alpha> 0$.  The corresponding bifurcation scenarios 

310: for (\ref{6.1})-(\ref{6.3}) 

311: for fixed values of $p$ and varying $\nu$ were studied numerically in \cite{FKRT}. 

312: Below, we reproduce some of these numerics for the system in new coordinates and supplement 

313: them with a set of new numerical experiments relevant to the analysis 

314: of this paper. 

315: After that we state our results for each of the following cases: supercritical AHB,

316: subcritical AHB, and the transition layer between the regions of sub- and supercritical 

317: bifurcations. 

318: 

319: \begin{figure}

320: \begin{center}

321: {\bf a}\epsfig{figure=f3a.eps, height=2.25in, width=3.0in, angle=0}

322: {\bf b}\epsfig{figure=f3b.eps, height=2.25in, width=3.0in, angle=0}

323: \end{center}

324: \caption{

325: The duration of the ISIs is plotted for:

326: ({\bf a}) $\gamma=0.778$ and ({\bf b}) $\alpha=0.045$.

327: }\label{f.3}

328: \end{figure}

329: 

330: We start with discussing the supercitical case. It is well known that a supercritical AHB

331: produces a stable periodic orbit in a small neighborhood 

332: of the bifurcating equilibrium. The period of the nascent orbit is approximately 

333: equal to $2\pi\beta^{-1},$ where $\beta=b(0)$ is the imaginary part of the pair of complex conjugate  eigenvalues 

334: of the matrix of the linearized system at the bifurcation.  The numerical time series of $x_1$ and $x_{2,3}$ show 

335: that, while $x_{2,3}$ oscillate with the frequency prescribed by the Andronov-Hopf Bifurcation Theorem 

336: \cite{KUZ, MarsdenMcCracken}, the former oscillates 

337: with twice that frequency (see Figure \ref{f.1}a,b). 

338: The asymptotic analysis of Section 3 explains this counter-intuitive effect and shows that 

339: this is, in fact, a generic property of the periodic orbits born from a supercritical bifurcation in $R^n$, $n\ge3$.  

340: Note that the difference in frequencies of oscillations in $x_1$ and $x_{2,3}$ can not be understood from the 

341: topological normal form from the AHB \cite{GEO}, because the latter does not contain the information about 

342: the geometry of the periodic orbit. In Section 4.1,

343: we compute two geometric invariants of the periodic orbits as a curve in $\Re^3$: the curvature and the torsion.  

344: The latter shows that generically the orbit born from a supercritical AHB is not planar.  This together with certain 

345: symmetry properties of the orbit explains the frequency doubling of the oscillations in $x_1$.

346: As was noted in \cite{FKRT}, for increasing values of $\alpha > 0$, the small periodic orbit born from the supercritical 

347: AHB undergoes a cascade of period-doubling bifurcations, the first of which is shown in Figure \ref{f.1}c,d. Already 

348: at the first period-doubling bifurcation, the periodic orbit lies outside the 

349: region of validity of the local power series expansions, developed in the present paper.  

350: Therefore, our analysis does not explain the period-doubling

351: bifurcations for increasing values of $\alpha>0$.

352: In Section 4.3 we complement our analytical results with the numerical construction of the $1D$ first return map. 

353: The latter explains the mechanism for period-doubling cascade and the window of complex dynamics reported in 

354: \cite{FKRT}. 

355: \begin{figure}

356: \begin{center}

357: {\bf a}\epsfig{figure=f4a.eps, height=1.6in, width=3.0in, angle=0}\quad

358: {\bf b}\epsfig{figure=f4b.eps, height=1.6in, width=3.0in, angle=0}

359: {\bf c}\epsfig{figure=f4c.eps, height=1.6in, width=3.0in, angle=0}

360: {\bf d}\epsfig{figure=f4d.eps, height=1.6in, width=3.0in, angle=0}\quad

361: \end{center}

362: \caption{The time series of $x_2$ are plotted for decreasing values of $\gamma>0$.

363: In plots {\bf a} and {\bf b}, the number of small amplitude oscillations separating the spikes increases,

364: while the qualitative form of the solutions remains the same.

365: For smaller values of $\gamma$, the system exhibits intermittency:

366: series of regular small amplitude oscillations are followed by those of irregular

367: oscillations of the intermediate amplitude ({\bf c} and {\bf d}). 

368: }\label{f.4}

369: \end{figure}

370: \begin{figure}

371: \begin{center}

372: {\bf a}\epsfig{figure=f5a.eps, height=1.6in, width=3.0in, angle=0}\quad

373: {\bf b}\epsfig{figure=f5b.eps, height=1.6in, width=3.0in, angle=0}

374: {\bf c}\epsfig{figure=f5c.eps, height=1.6in, width=3.0in, angle=0}\quad

375: {\bf d}\epsfig{figure=f5d.eps, height=1.6in, width=3.0in, angle=0}

376: {\bf e}\epsfig{figure=f5e.eps, height=1.6in, width=3.0in, angle=0}\quad

377: {\bf f}\epsfig{figure=f5f.eps, height=1.6in, width=3.0in, angle=0}

378: {\bf g}\epsfig{figure=f5g.eps, height=1.6in, width=3.0in, angle=0}\quad

379: {\bf h}\epsfig{figure=f5h.eps, height=1.6in, width=3.0in, angle=0}

380: \end{center}

381: \caption{

382: The time series of $x_2$ (left column) and the corresponding trajectories of (\ref{6.1})-(\ref{6.3})

383: (right column) plotted for negative values of $\gamma$ close to $0$. The transition from the 

384: complex dynamics in {\bf a,b} to regular approximately harmonic oscillations in {\bf g,h}

385: contains a reverse cascade of period-doubling bifurcations ({\bf c}-{\bf f}).

386: }\label{f.5}

387: \end{figure}

388:  	

389: We next turn to the subcritical case.  

390: The dynamics resulting from the subcritical AHB depends on the properties of the vector field outside 

391: of a small neighborhood of the equilibrium. We conducted a series of numerical experiments to study

392: the global properties of the vector field (see caption of Figure \ref{f.2} for details). Based 

393: on these numerical observations, we identify two essential features of the vector field:

394: {\bf (G1)} the return mechanism and {\bf (G2)} the strong contraction property. Specifically,

395: \begin{description}

396: \item[(G1)]

397: For a suitably chosen cylindrical crossection, $\Sigma^+$, placed sufficiently close to the origin, and

398: another crossection, $\Sigma^-$, transverse to $W^s(O)$ (see Figure \ref{f.2}a), the flow-defined map

399: $Q:\Sigma^+\to\Sigma^-$ is well-defined for $\alpha\ge 0$ and depends smoothly on the parameters of the

400: system.

401: \item[(G2)] There is a region $\Pi$, adjacent to $\Sigma^-$ and containing a subset of $W^s(O)$

402: (see Figure \ref{f.2}a), in which the projection of the vector field in direction transverse to 

403: $W^s(O)$ is sufficiently stronger than that in the tangential direction.

404: \end{description} 

405: In Section 4.2, conditions {\bf (G1)} and {\bf (G2)} are made precise.

406: Properties {\bf (G1)} and {\bf (G2)} guarantee that for small values of $\alpha>0$, the trajectories leaving

407: a small neighborhood of the saddle-focus, after some relatively short time enter $\Pi$. In $\Pi$,

408: the trajectories approach $W^s(O)$ closely and then follow it to a sufficiently small neighborhood of the 

409: unstable equilibrium. For small $\alpha>0$, the trajectories starting from a sufficiently small neighborhood

410: of the saddle-focus, remain in some larger (but still small) neighborhood of the origin for a long time.

411: Eventually, they hit $\Sigma^+$ and the dynamics described above repeats. Therefore, under conditions

412: {\bf (G1)} and {\bf (G2)}, the

413: subcritical AHB bifurcation results in a 

414: sustained motion consisting of long intervals of time that the trajectory spends near a weakly unstable saddle-focus 

415: and relatively brief excursions outside of a small neighborhood of the origin (see Figure \ref{f.2}a).  

416: In terms of the time series, the dynamical variables undergo series of small amplitude oscillations 

417: alternating with large spikes (Figure \ref{f.1}e-h). For increasing values of $\alpha > 0$, the number of small 

418: amplitude oscillations decreases and so do the time intervals between consecutive spikes, so-called 

419: interspike intervals (ISIs).  

420: The multimodal solutions arising near the subcritical AHB are not necessarily periodic. Nonetheless, 

421: the ISIs may be characterized in terms of the control parameters present in the system, regardless of whether the

422: underlying trajectories 

423: are periodic or not.  Specifically, in Section 4.2, we derive an asymptotic relation for the 

424: duration of the ISIs: 

425: \be\lbl{6.6}

426: \tau \sim {1\over 2\alpha} \ln\left(1+ {\alpha\over\gamma(p)}O\left(\epsilon^{-\varsigma}\right)\right),\;\;

427: \varsigma\ge 4,

428: \ee

429: where the first Lyapunov coefficient $\gamma(p)$ reflects the distance of the system from the transition 

430: from sub- to supercritical AHB, with $\gamma(p)$ 

431: being positive when the AHB is subcritical and equal to zero at the transition point. 

432: Finally, $\epsilon>0$ is a small parameter and positive $\alpha=O(\epsilon^2)$.

433: The role of $\epsilon$ will become clear later.  We illustrate (\ref{6.6}) with 

434: numerically computed plots of the ISIs under the variation of control parameters in Figure \ref{f.3}.  

435: The ISIs provide a 

436: convenient and important characteristics of the oscillatory patterns involving pronounced spikes.  For example, it 

437: is widely used in both theoretical and experimental neuroscience for description of patterns of electric activity 

438: in neural cells.  An important aspect of the ISIs, is that the dependence of the ISIs on the control parameters 

439: in the system (which often can be directly established experimentally) reveals the bifurcation structure of the system.  

440: Therefore, the analytical characterization of the ISIs in terms of the bifurcation parameters, 

441: such as given in (\ref{6.6}), is important in applications. For the problem at hand, 

442: the ISIs depend on the interplay of the two parameters 

443: $\alpha$ and $\gamma(p)$, which reflect the proximity of the system to a codimension 2 bifurcation.  

444: The dependence of the ISIs on $\alpha$ was studied in \cite{GW} 

445: for a model problem near a Hopf-homoclinic bifurcation.  Our analysis extends the formula 

446: for the ISIs obtained in \cite{GW} to a wide class of systems and emphasizes the role 

447: of $\gamma(p)$, which is important for the present problem.  

448: We also note that the structure of the global vector field suggests that (\ref{6.5}) is 

449:  also close to  a homoclinic bifurcation, which can also influence the ISIs.  However, under 

450: the variation of the control parameters $\alpha$ and $p$ the system remains 

451: bounded away from the homoclinic bifurcation, so that the latter effectively does not affect the ISIs. 

452: In fact, (\ref{6.6}) can be easily extended to include the distance from the homoclinic bifurcation as well.  

453: As follows from (\ref{6.6}), the ISIs increase for decreasing values of $\gamma>0$ (see Figure \ref{f.3}b).  

454: This observation prompted our interest in investigating the transition in the system 

455: dynamics as $\gamma$ crosses $0$. Numerical simulations show that as $\gamma$ approaches $0$ from the positive side, the 

456: oscillations become very irregular and very likely to be chaotic (Figure \ref{f.4}).  

457: When $\alpha >0$ is fixed and $\gamma$ crosses 0, 

458: the irregular dynamics is followed by the reverse period-doubling cascade and terminates with the creation of a 

459: regular small amplitude periodic orbit, which can be then tracked 

460: down to a nondegenerate supercritical AHB (Figure \ref{f.5}a-h).

461: This scenario presents a certain interest as it suggests the formation of 

462: the chaotic attractor in a well-defined bifurcation setting.  Namely, the transition 

463: from the region of subcritical to supercritical AHB for small $\alpha >0$ leads 

464: to the formation of a chaotic attractor and a reverse period-doubling cascade.

465: As in the case of the complex dynamics appearing for 

466: increasing values of $\alpha$ for supercritical AHB, the irregular oscillations in the present 

467: case can not be understood using the local analysis alone.  Therefore, 

468: we complemented our analysis with the study of numerically constructed first return map.  

469: The latter gives a clear geometric picture of the origins of the chaotic dynamics and the 

470: period-doubling cascade of periodic orbits in this parameter regime.   

471: 

472: 

473: \section{The local expansions}

474: \setcounter{equation}{0}

475: In the present section, we study trajectories of (\ref{6.5}) in a small neighborhood of the 

476: equilibrium at the origin. Assume that $h(x,\alpha)$ is a smooth function in a small neighborhood

477: of $(0,0)$, so that all power series expansions below are justified. 

478: We are interested in the dynamics of (\ref{6.5}) for values of $\alpha$ close to $0$. 

479: The role of $\alpha$ in our analysis is twofold. On the one hand, $\alpha$ is a control

480: parameter in this problem, on the other hand, the smallness of $\alpha$ is used

481: in the asymptotic analysis below. To separate these two roles, we use the following

482: rescaling

483: \be\lbl{5.1a}

484: \alpha=\mu\epsilon^2,

485: \ee

486: where $\epsilon>0$ is a fixed sufficiently small constant and $\mu$ varies in a certain interval

487: of size $O(1)$. We introduce cylindrical coordinates in $\Re^3$

488: \be\lbl{cl}

489: \left(x_1,x_2,x_3\right)\mapsto\left(x_1, \rho,\theta\right)\in\Re\times\Re^+\times S^1,\quad

490: x_2=\rho\cos\theta,\;x_3=\rho\sin\theta,

491: \ee

492: and define

493: \be\lbl{5.2}

494: D=\left\{ \left(x_1,\rho, \theta \right):\; \left| x_1 \right| \le M^2\epsilon^2, \rho\in(0,2M\epsilon]\right\},\;

495: D_0=\left\{ \left(x_1,\rho, \theta \right):\; \left| x_1 \right| \le M^2\epsilon^2, \rho\in(0,M\epsilon]\right\},

496: \ee

497: where $M>0$ is sufficiently large.

498: 

499: To characterize the trajectories of (\ref{6.5}) in $D$ we will use an exponentially stable slow manifold, $S$,

500: whose leading order approximation is given by

501: \be\lbl{5.3}

502: S_0= \left\{ 

503: \left( x_1,\rho,\theta\right):\; x_1=U(\theta)\rho^2,\; 0\le\rho\le 2\epsilon M\right\},

504: \ee

505: where

506: \be\lbl{5.4}

507: U(\theta)= a + A\cos\left(2\theta-\vartheta\right).

508: \ee

509: and $a, A$ and $\vartheta$ are computable constants (see Appendix A). 

510: 

511: 

512: Below we show that  the trajectories of (\ref{6.5}) with initial data from $D_0$ approach an

513: $O(\epsilon)$  neighborhood of $S_0$ in time $O\left(\left| \ln \epsilon \right|\right)$

514: and stay in this neighborhood as long as they remain in $D$. The reduction of the system dynamics to $S$

515: yields a complete description of the trajectories of (\ref{6.5}) in $D$. The qualitative character of the

516: solution behavior in $D$ depends on the sign of {\it the first Lyapunov coefficient}

517: \be\lbl{5.5}

518: \gamma={1\over 2\pi}\int_0^{2\pi} \left( U\left(\theta\right) \bar Q_2\left(\theta\right)+

519: \bar Q_3\left(\theta\right)\right) d\theta,

520: \ee

521: where $2\pi-$periodic trigonometric polynomial $\bar Q_{2,3}$ are given in Appendix A.

522: Finally, by $\beta=b(0)$ we denote the absolute value of the imaginary parts of the

523: complex conjugate pair of eigenvalues at the AHB (see (\ref{6.5})).

524: The results of the analysis of this section are summarized in

525: 

526: \noindent

527: {\bf Theorem 3.1} 

528: {\sc

529: Suppose $\beta>0$ and $\gamma\ne 0$. Let $\epsilon>0$ denote a small parameter.

530: Then for $\alpha=O(\epsilon^2)$ and for sufficiently large (independent of $\epsilon$)

531: $M>0$, the trajectories

532: with initial conditions in $D_0$ enter an $O(\epsilon)$ neighborhood of $S_0$ in time

533: $O\left(\left|\ln\epsilon\right|\right)$ and remain there as long as they stay in $D$.

534: In the neighborhood of $S_0$, the trajectories can be uniformly approximated on any

535: interval of time $[t_0,\bar t]$ of length $O(1)$:

536: \be\lbl{5.7}

537: x_1(t)=\rho^2 U(\theta)+O(\epsilon^3), \quad x_2(t)=\rho\cos\theta +O(\epsilon^2),\quad\mbox{and}\quad

538: x_3(t)=\rho\sin\theta +O(\epsilon^2),

539: \ee

540: where $U(\theta)$ and $\gamma$ are defined in (\ref{5.4}) and (\ref{5.5}), respectively,  and

541: $$

542: \rho(t) = \left(\left(\rho(t_0)^{-2}+{\gamma\over\alpha}\right) e^{-2 \alpha(t-t_0)}-

543: {\gamma\over\alpha}\right)^{-1\over 2}+O(\epsilon^2),\quad

544: \dot\theta=\beta,\quad

545: \tan\left(\theta(0)\right)= {x_3(t_0)\over x_2(t_0)}.

546: $$

547: } 

548: 

549: \noindent {\bf Remark 3.1}

550: \begin{description}

551: \item[a)] From Theorem 3.1, one can easily deduce the existence of a periodic orbit $O_\alpha$,

552: when $\alpha\gamma<0$. $O_\alpha$ is stable  (unstable) if $\alpha>0$ ($\alpha<0$). The leading order

553: approximation for $O_\alpha$ follows from (\ref{5.7}):

554: \be\lbl{5.8}

555: x_1=\bar\rho^2 U(\theta)+O(\epsilon^3),\;\; x_2=\bar\rho\cos\theta +O(\epsilon^2),\;\;

556: x_3=\bar\rho\sin\theta +O(\epsilon^2),\;\;

557: \mbox{and}\;\;

558: \theta\in [0, 2\pi),

559: \ee

560: where $\bar\rho=\sqrt{\alpha\over -\gamma}$. Equations (\ref{5.4}) and (\ref{5.8})

561: imply that the frequency of oscillations in $x_1$ is twice

562: that of oscillations in $x_{2,3}$, provided $A\ne 0$ in (\ref{5.4}) 

563: (see Figure \ref{f.1}a,b). In Section 4.1, we give a geometric interpretation of this frequency doubling 

564: effect.

565: \item[b)] If $\alpha>0$ and $\gamma>0$, Theorem 3.1 describes the trajectories with initial conditions

566: near a weakly unstable equilibrium. It shows that along these trajectories, $x_{2,3}$ undergo  approximately

567: harmonic oscillations, whose amplitude grows at the rate $O(\epsilon^2)$. In addition, they satisfy the

568: following scaling relation:

569: \be\lbl{5.9}

570: \bar x_1(t) \sim a\rho^2(t),

571: \ee

572: where $\bar x_1(t)$ stands for the average value of $x_1(t)$ over one cycle of oscillations.

573: \item[c)]

574: The domain of validity of the asymptotic expansions in (\ref{5.7}) extends much farther than

575: $D$. As will follow from the proof of Theorem 3.1, $M$ in the definition of $D$ may be taken 

576: up to $o\left(\epsilon^{-1\over 3}\right)$. This

577: implies that, with the error term $o(1)$, (\ref{5.7}) remains valid for the ranges of $x_1$ and $x_{2,3}$ up to 

578: $o(\epsilon^{4\over 3})$ and $o(\epsilon^{2\over 3})$, respectively.

579: \end{description}

580: 

581: In the remainder of this section, we prove Theorem 3.1. By expanding

582: $h(x,\alpha)$ into finite Taylor sum with the reminder term, we have  

583: \begin{eqnarray}\nonumber

584: \dot x_1 &=& -x_1+\sum a_{ij}(\alpha)x_i x_j +\sum a_{ijk}(\alpha) x_ix_jx_k + O(4),\\ \lbl{5.1}

585: \dot x_2 &=& a(\alpha) x_2 -b(\alpha) x_3 +\sum b_{ij}(\alpha)x_i x_j +\sum b_{ijk}(\alpha) x_ix_jx_k +O(4),\\ \nonumber

586: \dot x_3 &=& b(\alpha) x_2 +a(\alpha) x_3 +\sum c_{ij}(\alpha)x_i x_j +\sum c_{ijk}(\alpha) x_ix_jx_k +O(4),

587: \end{eqnarray}

588: where

589: \begin{eqnarray*}

590: \sum \sigma_{ij}(\alpha) x_ix_j &=& \sum_{i=1}^3 \sum_{j=i}^3 \sigma_{ij}(\alpha) x_ix_j,\\

591: \sum \sigma_{ijk}(\alpha) x_ix_jx_k &=& \sum_{i=1}^3 \sum_{j=i}^3 \sum_{k=j}^3 \sigma_{ijk}(\alpha) x_ix_jx_k,\quad

592: \sigma\in\left\{a,b,c\right\}.

593: \end{eqnarray*}

594: To study system of equations (\ref{5.1}) in a small neighborhood of the origin, 

595: we rewrite it in cylindrical coordinates (\ref{cl}) and rescale variables

596: \be\lbl{5.10}

597:  x_1=\xi\epsilon^2\;\mbox{and}\; \rho=r\epsilon.

598: \ee

599: In new coordinates, we have

600: \begin{eqnarray}

601: \dot \xi &=&-\xi+ r^2 T_1(\theta) +O\left(\epsilon\right),\nonumber\\ \lbl{5.11}

602: \dot r &=& \epsilon r^2 R_1(\theta)+

603: \epsilon^2 r\left(\mu+\xi R_2(\theta)+r^2R_3(\theta)\right)+O(\epsilon^3),\\

604: \nonumber

605: \dot\theta &=& \beta+\epsilon r L_1(\theta)+ O\left(\epsilon^2\right),

606: \end{eqnarray}

607: where $T_1, L_1$ and $R_{1,2,3}$ are homogeneous trigonometric polynomials.

608: 

609: Using $\phi=\beta^{-1}\theta$ as a new variable, from (\ref{5.11}), we obtain

610: \begin{eqnarray}\lbl{5.11a}

611: \frac{d\xi}{d\phi}&=&-\xi+r^2 P_1(\phi)+\epsilon \Xi\left(\xi,r,\phi,\epsilon\right),\\ 

612: \lbl{5.12}

613: \frac{dr}{d\phi}&=&\epsilon Q_1(\phi) r^2 +\epsilon^2 r\left(\mu+\xi Q_2(\phi)+r^2 Q_3(\phi)\right)+

614: \epsilon^3 \Psi\left(\xi,r,\phi,\epsilon\right),

615: \end{eqnarray}

616: where $Q_{1,2,3}(\phi)$ and $P_1(\phi)$ are trigonometric polynomials of period $\omega=2\pi\beta^{-1}$.

617: Functions $\Xi$ and $\Psi$ are bounded and continuous in $D'\times\Re\times [0, 1]$.

618: Here $D'$ stands for the domain of the rescaled variable $(\xi,r)$ when the original variable

619: $x\in D\subset \Re^3$:

620: \be\lbl{Dprime} 

621: D^\prime= \left\{ \left( \xi,r \right):\; \left| \xi \right| \le 4M^2, \; 0<r\le 2 M \right\}.

622: \ee

623: Similarly, we define

624: \be\lbl{D0prime}

625: D^\prime_0=\left\{ \left(\xi,r\right):\; \left| \xi \right| \le M^2, \; 0<r\le M\right\}.

626: \ee

627: 

628: To determine the slow manifold for (\ref{5.1}), we introduce the following notation. 

629: By $\bar p_1$ and $\tilde P_1(\phi)$ we denote the mean value and the oscillating part 

630: of the trigonometric polynomial $P_1(\phi)$ in (\ref{5.11a}), respectively:

631: \be\lbl{5.13}

632: \bar p_1={\beta\over 2\pi}\int_0^{2\pi\over\beta} P_1(\phi)d\phi

633: \quad\mbox{and}\quad \tilde P_1(\phi)= P_1(\phi)-\bar p_1,

634: \ee

635: and define

636: \be\lbl{5.16}

637: \xi_0(\phi)=\bar p_1 +p_1(\phi) \quad\mbox{and}\quad

638: p_1(\phi)=

639: {\tilde P_1(\phi)-{\tilde P_1}^\prime (\phi)\over 1+4\beta^2}.

640: \ee

641: 

642: We first prove an auxiliary lemma, which characterizes the trajectories of (\ref{5.11a}) and (\ref{5.12})

643: in $D^\prime$.

644: In the following two lemmas, we will keep track of how the remainder terms in the asymptotic expansions

645: depend on the size of $D^\prime$. This will be used later to determine the domain of validity for the 

646: asymptotic approximation of the slow manifold. Below $\kappa=O\left(M\epsilon\right)$ means that there

647: exist positive constants $\epsilon_0$ and $C$ (independent of $\epsilon$) such that

648: $\left|\kappa\right| \le CM\epsilon$ for $\epsilon\in\left[0, \epsilon_0\right]$.

649: In the remainder of this section, all estimates are valid for sufficiently large $M>0$ independent 

650: of $\epsilon$ (as stated in the Theorem 3.1), but also allow the possibility that $M>1$ grows as 

651: $\epsilon\to 0$.

652: 

653: \noindent

654: {\bf Lemma 3.1} 

655: {\sc Suppose 

656: $\left(\xi, r\right)$ remains in $D^\prime$ for $\phi\in\left[\phi_0,\bar\phi\right]$. Then

657: \be\lbl{5.15}

658: \xi(\phi)= \xi_0(\phi)r^2(\phi) +

659: e^{-(\phi-\phi_0)}\left\{ \xi(\phi_0)-\xi_0(\phi_0) r^2(\phi_0)\right\} +O\left(M^3\epsilon\right),

660: \quad\phi\in\left[\phi_0, \bar\phi\right],

661: \ee

662: where $\xi_0(\phi)$ is defined in (\ref{5.16}).

663: }

664: 

665: \noindent{\bf Proof:}

666: Using (\ref{5.13}), we integrate (\ref{5.11a}) over $\left[\phi_0,\phi\right]$

667: \be\lbl{5.17}

668: e^\phi\xi(\phi)=e^{\phi_0}\xi(\phi_0)+

669: \bar p_1 \int_{\phi_0}^\phi e^s r^2(s)ds  + \int_{\phi_0}^\phi e^s r^2(s) \tilde P_1(s) ds+ 

670: \epsilon\int_{\phi_0}^\phi e^s \Xi\left(\xi(s), r(s),s,\epsilon\right) ds.

671: \ee

672: Using integration by parts, we have

673: $$

674: \int_{\phi_0}^\phi e^s r^2(s)ds=e^\phi r^2(\phi)-e^{\phi_0} r^2(\phi_0)

675: -2\int_{\phi_0}^\phi e^s r(s)\dot r(s)ds.

676: $$

677: From (\ref{5.12}) and (\ref{Dprime}), we find that for $\left(\xi,r\right)\in D^\prime$

678: $$

679: \left|\int_{\phi_0}^\phi e^s r(s)\dot r(s)ds\right| \le C_1 e^{\phi-\phi_0}M^3\epsilon,

680: $$ 

681: where $C_1>0$ is independent of $\epsilon$.

682: Therefore,

683: \be\lbl{5.18}

684: \int_{\phi_0}^\phi e^s r^2(s)ds=e^\phi r^2(\phi)-e^{\phi_0} r^2(\phi_0)

685: +e^{\phi-\phi_0}O\left(M^3\epsilon\right).

686: \ee

687: Similarly, using integration by parts in the second integral on the right hand

688: side of (\ref{5.17}), we obtain

689: \begin{eqnarray}\nonumber

690: \int_{\phi_0}^\phi e^s r^2(s) \tilde P_1(s) ds &=& e^\phi r^2(\phi) \tilde P_1(\phi)-e^{\phi_0} r^2(\phi_0) \tilde P_1(\phi_0)\\

691:    & &-2\int_{\phi_0}^\phi e^s r(s) \dot r(s) \tilde P_1(s)ds - \int_{\phi_0}^\phi e^s r^2(s) \tilde P_1^\prime(s)ds.\lbl{A.2}

692: \end{eqnarray}

693: Applying integration by parts in the last integral on the right hand side of (\ref{A.2}),

694: we have

695: \begin{eqnarray}\nonumber

696: \int_{\phi_0}^\phi e^s r^2(s) \tilde P_1^\prime(s) ds &=& e^\phi r^2(\phi) \tilde P_1^\prime(\phi)-e^{\phi_0}r^2(\phi_0) \tilde P_1^\prime(\phi_0)\\

697: && -2\int_{\phi_0}^\phi e^s r(s) \dot r(s) \tilde P_1^\prime (s)ds - \int_{\phi_0}^\phi e^s r^2(s) \tilde P_1^{\prime\prime}(s)ds.

698: \lbl{A.3}

699: \end{eqnarray}

700: By direct verification, one finds that 

701: $

702: \tilde P_1^{\prime\prime}=\left(1+4\beta^2\right)\tilde P_1.

703: $

704: The integrals involving $\dot r$ on the right hand sides of (\ref{A.2}) and

705: (\ref{A.3}) are bounded by $e^{\phi-\phi_0}O(M^3\epsilon)$ since, by (\ref{5.12}), $\dot r=O(\epsilon M^2)$.

706: Using these observations, from (\ref{A.2}) and (\ref{A.3}) we obtain

707: \be\lbl{A.4}

708: \int_{\phi_0}^\phi e^s r^2(s) \tilde P_1(s)ds = e^\phi r^2(\phi)p_1(\phi)-e^{\phi_0} r^2(\phi_0)p_1(\phi_0)+ e^{\phi-\phi_0}O(M^3\epsilon),

709: \ee

710: where $p_1(\phi)$ is given by (\ref{5.16}).

711: The combination of (\ref{5.17}), (\ref{5.18}), and  (\ref{A.4}) yields (\ref{5.15}).

712: \\

713: $\Box$

714: 

715: The following lemma shows that the trajectories of (\ref{5.11a}) and (\ref{5.12}),

716: which stay in $D^\prime$ for sufficiently long time 

717: enter a small neighborhood of $S$ 

718: no later than in time $O(\left|\ln\epsilon\right|)$

719: and remain in this neighborhood as long as they stay in $D^\prime$ (see (\ref{Dprime})).

720: 

721: \noindent

722: {\bf Lemma 3.2} 

723: {\sc 

724: Let $\left(\xi(\phi),r(\phi)\right)$ denote a trajectory of (\ref{5.11a}) and (\ref{5.12}),

725: which stays in $D^\prime$ for $\phi\in \left[\phi_0, \bar\phi\right]$, $\bar \phi>\phi_1$,

726: $\phi_1=\phi_0+\left|\ln\epsilon\right|$. Then

727: \be\lbl{5.20}

728: \xi(\phi)=\xi_0(\phi) r^2(\phi) +O\left(M^3\epsilon\right),

729: \ee

730: for $\phi\ge\phi_1$ and as long as $\left(\xi(\phi),r(\phi)\right)\in D^\prime$.

731: }

732: 

733: \noindent{\bf Proof:} Denote $\phi_1=\phi_0+\left|\ln\epsilon\right|$.

734: By plugging in $\phi=\phi_1$ into (\ref{5.15}), we have

735: \be\lbl{5.21}

736: \left|\xi(\phi_1) - \xi_0(\phi_1)r^2(\phi_1)\right|\le C_2\epsilon,

737: \ee

738: for some $C_2>0$ independent of $\epsilon>0$.

739: Using Lemma 3.1 again with $\phi_0:=\phi_1$, we obtain

740: $$

741: \xi(\phi)= \xi_0(\phi)r^2(\phi) +

742: e^{-(\phi-\phi_1)}\left\{\xi(\phi_1)-\xi_0(\phi_1)r^2(\phi_1)\right\}

743: +O(M^3\epsilon),\;\phi\ge \phi_1.

744: $$

745: The expression in the curly brackets is $O(\epsilon)$ by (\ref{5.21}).

746: \\

747: $\Box$

748: 

749: \noindent {\bf Remark 3.2}

750: Lemmas 3.1 and 3.2 show that the trajectories of (\ref{5.11a}) and (\ref{5.12})

751: converge to an exponentially stable manifold $S$, whose  

752: leading order approximation is given in the definition of $S_0$. The method, 

753: which we used in Lemmas 3.1 and 3.2 to obtain the

754: leading order approximation of the slow manifold, can be extended to calculate the

755: higher order terms in the expansion for $S$.

756:  

757: Having shown that the trajectories approach an $O(\epsilon)$ neighborhood of $S_0$ in time 

758: $O\left(\left|\ln\epsilon\right|\right)$, next we

759: reduce the dynamics of (\ref{5.11a}) and (\ref{5.12}) to the slow manifold. 

760: For this, we define

761: \be\lbl{5.22}

762: I(\phi)=\left({1\over r(\phi)}+\epsilon q_1(\phi)\right)^2,

763: \ee

764: where $q_1^\prime(\phi)=Q_1(\phi)$ and $Q_1(\phi)$ is a trigonometric polynomial on 

765: the right hand side of (\ref{5.12}). For the sake of definiteness, 

766: we choose $q_1^\prime(\phi)$ such that 

767: $$

768: \int_0^{2\pi\over\beta}q_1(s)ds=0.

769: $$

770: The following lemma provides the desired reduction.

771: 

772: \noindent

773: {\bf Lemma 3.3} 

774: {\sc For $\phi\ge\phi_1$ and as long as $(\xi,r)\in D^\prime$, $(I,\phi)$ satisfies the 

775: following system of equations

776: \begin{eqnarray}\lbl{5.24}

777: \dot I &=& -2\epsilon^2\left(\mu I+\gamma+Q(\phi)\right)+O(M^3\epsilon^3),\\ \lbl{5.25}

778: \dot \phi &=& 1+O(M\epsilon),

779: \end{eqnarray}

780: where $Q(\phi)$ is a trigonometric polynomial with period $\omega={2\pi\beta^{-1}}$ and zero mean 

781: $\int_0^\omega Q(\phi) d\phi =0$. The expression for $\gamma$ is given in

782: (\ref{5.5}).

783: }

784: 

785: \noindent{\bf Proof:}

786: The change of variables

787: \be\lbl{5.26}

788: r(\phi)={1\over J(\phi)-\epsilon q_1\left(\phi\right)},\quad \mbox{where}\quad q_1^\prime(\phi)=Q_1(\phi),

789: \ee

790: in Equation (\ref{5.12}) yields

791: \be\lbl{5.27}

792: \dot J ={-\epsilon^2\over J}\left(\left(\mu+\xi Q_2(\phi) \right) J^2 +Q_3(\phi)\right)+O(M^2\epsilon^3).

793: \ee

794: After another change of variables, $I=J^2$, we obtain

795: \be\lbl{5.28}

796: {d I\over d\phi} = -2\epsilon^2 \left( \left(\mu+ \xi Q_2(\phi)\right) I +Q_3(\phi)\right)

797: + O(M^3\epsilon^3),

798: \ee

799: By Lemma 3.2, for $\phi\ge\phi_1$

800: \be\lbl{5.29}

801: \xi(\phi)=\xi_0(\phi)r^2(\phi) +O(M^3\epsilon).

802: \ee

803: By plugging in (\ref{5.29}) into (\ref{5.28}), we obtain

804: \be\lbl{5.30}

805: {d I\over d\phi} = -2\epsilon^2\left(

806: \mu I +\xi_0(\phi)Q_2(\phi)+Q_3(\phi)\right)+

807: O(M^3\epsilon^3).

808: \ee

809: Equation (\ref{5.24}) follows by rewriting  (\ref{5.30}) 

810: \be\lbl{5.31}

811: {d I\over d\phi} = -2\epsilon^2\left(

812: \mu I +\gamma+ Q(\phi)\right)+

813: O(M^3\epsilon^3),

814: \ee

815: where

816: \be\lbl{5.32}

817: \gamma = {1\over\omega}\int_0^\omega\left(\xi_0(\phi)Q_2(\phi)+Q_3(\phi)\right)d\phi\;\;\mbox{and}\;\; 

818: \int_0^{\omega} Q(\phi)d\phi =0,

819: \ee

820: and $\xi_0(\phi)$ is given in (\ref{5.16}).

821: Equation (\ref{5.25})  follows from the last equation in (\ref{5.11}) and the definition of $\phi$.\\

822: $\Box$

823: 

824: The statements in Theorem 3.1 can now be deduced from Equations (\ref{5.24}) and (\ref{5.25}).

825: We only need to show that a trajectory with initial condition in $D_0^\prime$ remains in $D^\prime$

826: for times longer than $O\left(\left|\ln\epsilon\right|\right)$. This follows from the fact that

827: $\dot I=O\left(\epsilon^2\right)$ in $D^\prime$. Consequently, it takes time $O(\epsilon^{-2})$ for

828: $I$ to undergo $O(1)$ change necessary for leaving $D^\prime$ from $D^\prime_0$.

829: We omit any further details. In conclusion, we  note that the domain of validity of the asymptotic 

830: analysis of this section extends

831: much further beyond $D^\prime$. Indeed, from (\ref{5.20}) we observe that the remainder term 

832: tends to $0$ with $\epsilon\to 0$ provided $M=o\left(\epsilon^{-1\over 3}\right)$. This means that

833: the expansions for $r$ and $\xi$ can be controlled in regions of size up to $O\left(\epsilon^{-1\over 3}\right)$

834: and $O\left(\epsilon^{-2\over 3}\right)$ respectively. For the original variables $\rho,$

835: $x_1$, these estimates translate into  $O\left(\epsilon^{2\over 3}\right)$

836: and $O\left(\epsilon^{4\over 3}\right)$ respectively.

837: 

838: We end this section by deriving a useful estimate for the time of flight of trajectories

839: passing close to the saddle-focus. For this, consider an initial value problem for system of equations

840: (\ref{5.11a}) and (\ref{5.12}). Suppose that the initial condition implies that $I\left(\phi_0\right)=I_0>0$.

841: We would like to know how long it takes for $I$ to reach a given value $\bar I=I\left(\phi_0+\Delta\right)$,

842: $\Delta>0$.

843: We assume that $I_0$ and $\bar I$ are sufficiently separated, e.g., $2\bar I\le I_0$.

844: To estimate $\Delta$, note that

845: \be\lbl{5.51}

846: \Delta=\Delta_1+\Delta_2,

847: \ee

848: where $\Delta_1=O\left(\left|\ln\epsilon\right|\right)$ is time necessary to reach 

849: an $O(\epsilon)$ neighborhood of $S_0$ from $I_0$. Denote 

850: \be\lbl{5.52}

851:   I_1:= I\left(\phi_1\right)=I_0+O\left(\epsilon^2\left|\ln\epsilon\right|\right),\;

852:   \mbox{where}\; \phi_1=\phi_0+\Delta_1.

853: \ee

854: The second term on the right hand side of (\ref{5.51}) can be estimated by integrating (\ref{5.24})

855: over $\left[ \phi_1, \phi_1+\Delta_2\right]:$

856: \be\lbl{5.53}

857: \bar I= I_1 e^{-2\alpha\Delta_2}+{\gamma\over\mu}\left(1- e^{-2\alpha\Delta_2}\right)

858: -2\epsilon^2\int_0^{\Delta_2} e^{-2\alpha\left(\Delta_2-s\right)} Q\left(\phi_0+s\right)ds +

859: O\left(\epsilon^3\right).

860: \ee

861: The integral on the right hand side of (\ref{5.53}) is bounded uniformly for $\Delta_2>0$.

862: This observation combined with (\ref{5.52}) implies 

863: \be\lbl{5.54}

864: \bar I= I_0 e^{-2\alpha\Delta_2}-{\gamma\over\mu}\left(1- e^{-2\alpha\Delta_2}\right)

865: +O\left(\epsilon^2\right).

866: \ee

867: Note that the contribution of $O\left(\epsilon^2 \left| \ln\epsilon \right|\right)$ term in (\ref{5.52})

868: to (\ref{5.54}) is negligible provided $\Delta_2$ is sufficiently large.

869: From  (\ref{5.54}), we obtain the desired estimate

870: \be\lbl{5.55}

871: \Delta={1\over 2\mu\epsilon^2}\ln \left({I_0 +{\gamma\over \mu}\over \bar I +{\gamma\over\mu}+O(\epsilon^2)}\right).

872: \ee

873: 

874: 

875: \section{The oscillations}

876: \setcounter{equation}{0}

877: 

878: In the present section, we use the local analysis of Section 3 to study certain oscillatory patterns

879: arising in the model of solid fuel combustion (\ref{6.1})-(\ref{6.3}). By plugging in the values of the 

880: parameters of (\ref{6.1})-(\ref{6.3}) into the expression for $\gamma$ (\ref{5.5}), we find that $\gamma(p)$

881: is a quadratic function with two zeros at $p_1\approx 0.34$ and $p_2\approx 2.73$. 

882: These values are in a good agreement with those obtained by numerical bifurcation analysis in \cite{FKRT}.

883: The quadratic character of $\gamma(p)$ is explained by the fact that $\gamma$ is determined by the

884: second order terms in the Taylor expansions of $h_{2,3}(0,0)$.

885: We concentrate on the parameter region around $p=p_2$, where $\gamma(p)$ changes its sign.

886: For small positive $\alpha$, there are two dynamical regimes: for values of $p$ lying 

887: to the left and to the right of some $O(\epsilon)$ neighborhood

888: of $p_2$ corresponding to the supercritical and subcritical AHBs. The transition region between these two

889: parameter regimes adds to the repertoire of qualitatively distinct dynamical behaviors. Below, we study

890: these three cases in more detail.

891: 

892: \subsection{The supercritcal AHB: geometry of the periodic orbit}

893: \setcounter{equation}{0}

894: It follows from Theorem 3.1 that for $\gamma(p)<0$ and sufficiently small $\alpha>0$,

895: (\ref{6.5}) has a stable limit cycle $O_\alpha$, whose leading order approximation is given by

896: \be\lbl{4.1}

897: x(\theta)=\left(\bar\rho^2\left(a+A \cos 2\theta\right), \bar\rho\cos\theta, \bar\rho\sin\theta\right),

898: \quad\bar\rho=\sqrt{\alpha\over -\gamma},\quad \theta\in[0, 2\pi)

899: \ee

900: (see Figure \ref{f.6}a).

901: Moreover, the analysis of Section 3 shows that all trajectories starting from a sufficiently small 

902: neighborhood of the origin and not belonging to $W^s(O)$, converge to $O_\alpha$. The leading

903: order approximation of $O_\alpha$ in (\ref{4.1}) reveals a remarkable property of the oscillations 

904: generated by the limit cycle born from the supercritical AHB: the frequency of oscillations in $x_1$

905: is twice as large as that of the oscillations in $x_{2,3}$. To explain the frequency doubling effect,

906: we recall that 

907: \be\lbl{4.2}

908: \dot\theta=\beta+O(\epsilon).

909: \ee

910: Therefore, (\ref{4.1}) implies that, unless $A=0$,

911: the frequency of oscillations in $x_1$ is $\beta\pi^{-1}$, while that of oscillations in $x_{2,3}$

912: is ${\beta\over 2\pi}$. The latter coincides with the frequency of the bifurcating periodic solution.

913: Below, we complement the analytical explanation of the frequency doubling with the geometric interpretation.

914: 

915: The oscillations in $x_{2,3}$ can be well understood using the topological normal form of the AHB

916: \cite{GEO,GH, KUZ}. Indeed, since at the bifurcation the center manifold at the origin is tangent to the $x_2-x_3$

917: plane, a standard treatment of the AHB using the center manifold reduction \cite{Carr, GH} shows that the projection

918: of the bifurcating limit cycle onto $x_2-x_3$ plane to leading order is a circle (Figure \ref{f.6}b)

919: and the projection

920: of the vector field is given by the equation of the angular variable (\ref{4.2}). Therefore, the oscillations in

921: $x_{2,3}$ are approximately harmonic with the period equal approximately to $2\pi\beta^{-1}$. The topological

922: normal from, however, does not describe the oscillations in $x_1$. For this, one needs to take into account the

923: geometry of the bifurcating periodic orbit as a curve in $\Re^3$. The geometry of $O_\alpha$ is fully determined by 

924: the two geometric invariants: the curvature, $\mbox{k}(\theta)$, and the torsion, $\kappa(\theta)$. For the purposes

925: of the present discussion, we need only the latter, but we compute both invariants for completeness.

926: After some algebra, the parametric equation for the periodic orbit (\ref{4.1}) yields  

927: \begin{eqnarray}\nonumber 

928: \mbox{k}(\theta)&=& {\left|\left[\dot x, \ddot x\right]\right|\over \left|\dot x\right|^3}=

929: \sqrt{1+2A^2\left(3\cos 4\theta +5\right)}+\mbox{h.o.t},\\

930: \lbl{4.3} 

931: \kappa(\theta) &=& {\left(\dot x, \ddot x, \dddot x\right)\over\left|\left[\dot x, \ddot x\right]\right|^2}=

932: {-\gamma\over\alpha} {6A\sin 2\theta\over 1+2A^2\left(3\cos 4\theta +5\right)}

933: +\mbox{h.o.t},\quad \theta\in[0, 2\pi).

934: \end{eqnarray}

935: The geometry of the leading order approximation of the periodic orbit is determined by the three

936: parameters $\alpha,$ $\gamma,$ and $A$. The former two parameters are the same as the parameters in the  

937: topological normal form of the nondegenerate AHB \cite{KUZ}; while the latter captures the geometry of the slow manifold (or unstable

938: manifold) near the origin. From the geometrical viewpoint the bifurcation is degenerate if either $\gamma$ or $A$ is equal

939: to zero. The latter condition holds if and only if the slow manifold is either a plane or a circular paraboloid

940: near the origin. In this case, Equation (\ref{4.3})

941: implies that (to leading order) the bifurcating orbit lies in a plane. Generically, $O_\alpha$ is not planar 

942: (see Figure \ref{f.6}a). 

943: The fact that the periodic orbit is generically not planar combined with the

944: symmetry of the orbit about the $x_1-$axis implies that its projection onto any plane containing $x_1-$axis has to have

945: a self-intersection (see Figure \ref{f.6}c). From  Figure \ref{f.6}c, it is clear that as the phase point goes 

946: around $O_\alpha$ once, $x_1$

947: has to trace its range at least twice. Therefore, the frequency of oscillations in $x_1$ has to be at least twice

948: as high as that of the oscillations in $x_{2,3}$. The above discussion implies that the frequency doubling of oscillations in

949: $x_1$ is a generic  geometric property of the AHB.  

950: \begin{figure}

951: \begin{center}

952: {\bf a}\epsfig{figure=f6a.eps, height=1.8in, width=2.0in, angle=0}

953: {\bf b}\epsfig{figure=f6b.eps, height=1.8in, width=2.0in, angle=0}

954: {\bf c}\epsfig{figure=f6c.eps, height=1.8in, width=2.0in, angle=0}

955: \end{center}

956: \caption{ A periodic orbit born from the supercritical AHB ({\bf a}), and

957: its projections onto $x_2-x_3$({\bf b}) and $x_1-x_2$ planes ({\bf c}).

958: Note that the periodic orbit shown in ({\bf a}) is not planar. This 

959: accounts for the presence of the self-intersection in the its projection

960: in ({\bf c}) (see text for details).

961: }\label{f.6}

962: \end{figure}

963: 

964: 

965: \subsection{The subcritcal AHB: multimodal oscillations}

966: 

967: If the AHB is subcritical ($\gamma>0$) the loss of stability of the equilibrium at the origin of (\ref{6.5}) 

968: results in the creation of multimodal trajectories which spend a considerable amount of time near a weakly 

969: unstable equilibrium. To describe the resultant dynamics we give the following definition.

970: 

971: \noindent

972: {\bf Definition 4.1} 

973: {\sc We say that a trajectory of (\ref{6.5}) undergoes multimodal oscillations for $t\ge 0$ if

974: there exist positive constants $r_1$ and $r_2$ independent of $\epsilon$ and an unbounded sequence of times

975: $$

976: 0\le t_1<t_2<\dots,\quad \lim_{i\to\infty} t_i=\infty,

977: $$ 

978: such that 

979: \begin{description}

980: \item[a.] $\left|x\left(t_{2i}\right)\right|\le r_1\epsilon^2\;\;\&\;\;\left|x\left(t_{2i-1}\right)\right|> r_2,\quad i=1,2,\dots$,

981: \item[b.] $\left(\exists t^\prime> t_1:\;\left| x(t^\prime)\right| \le r_1\epsilon^2\right)\;\Rightarrow\; 

982: \exists i\in\Na\;:\;

983: \left|x(t)\right|\le r_1\epsilon^2, 

984: t\in\left[ \min(t^\prime, t_{2i}), \max(t^\prime, t_{2i})\right]\;\;\forall t^\prime \in \left(t_{2i-1}, t_{2i+1}\right).$

985: \end{description}

986: The time intervals $\tau_i=t_{2i+1}-t_{2i-1},\; i=1,2,\dots$ are called ISIs.

987: }

988: 

989: The proximity of (\ref{6.5}) to the AHB alone is clearly not sufficient to account for  the appearance of 

990: the multimodal oscillations in (\ref{6.5}).

991: Below, we formulate two additional assumptions on the vector field outside of the small neighborhood of the equilibrium, which

992: are relevant to (\ref{6.1})-(\ref{6.3}).

993: Under these conditions, we show that the subcritical AHB results in sustained multimodal

994: oscillations. In addition, we determine the asymptotics of the ISIs for positive $\alpha=O(\epsilon^2)$. 

995: For (\ref{6.5}) to generate multimodal oscillations, it is necessary that the trajectories leaving an $O(\epsilon^2)$

996: neighborhood of the origin reenter it after some interval of time. Therefore, the vector field in $O(1)$ neighborhood of the

997: origin must provide a return mechanism. Our second assumption on the global vector field of (\ref{6.5})  

998: is that the trajectories approaching the origin

999: along a $1D$ stable manifold, $W^s(O)$, are subject to a strong contraction toward $W^s(O)$, i.e., in a small neighborhood

1000: of an $O(1)$ segment of $W^s(O)$, the projection of the vector field onto a plane transversal to $W^s(O)$ is sufficiently stronger than

1001: that along $W^s(O)$ (Figure \ref{f.2}a). This guarantees that the trajectories entering such 

1002: region of strong contraction approach $W^s(O)$ very 

1003: closely and follow it to an $O(\epsilon^2)$ neighborhood of the unstable equilibrium, from where they are propelled away along 

1004: the unstable manifold, $W^u(O)$. Due to the proximity of (\ref{6.5}) to the AHB, the motion away from the origin is very slow.

1005: This results in the pronounced intervals of the small amplitude oscillations. Below, we  summarize these observations into 

1006: two formal assumptions on the global vector field {\bf (G1)} and {\bf(G2)}. For this, we first need to introduce some

1007: auxiliary notation.

1008: For analytical convenience, we assume that in an $O(1)$ neighborhood of the origin, $W^s(O)$ can be and has been straightened by  

1009: a smooth change of coordinates. More specifically, the nonlinear terms in (\ref{6.5}) satisfy

1010: \be\lbl{4.3a}

1011: h_{2,3} \left(x_1,0,0,\alpha\right)=0,\quad w(x_1,\alpha)=-x_1+h_1\left(x_1,0,0,\alpha\right)>0,\; 

1012: x_1\in [d_1,0),

1013: \ee

1014: for some $d_1<0$. Note that $\left(d_1, 0,0\right)$ is assumed to be sufficiently far away from the origin (see Figure \ref{f.8}).  

1015: To describe the mechanism of return, we introduce two crossections:

1016: \be\lbl{4.4}

1017:  \Sigma^+=[-c_1\delta^2, c_1\delta^2]\times D_\delta\quad\mbox{and}\quad\Sigma^-=\left\{d_1\right\}\times D_{\delta_1},

1018: \ee

1019: where $D_\delta \subset\Re^2$ denotes a disk of radius $\delta$ centered at the origin:

1020: \be\lbl{4.4a}

1021: D_\delta=\left\{ y\in\Re^2:\;\left|y\right|\le \delta\right\}.

1022: \ee

1023: Positive constants $\delta$ and $\delta_1$ are sufficiently small (see Figure \ref{f.2}a) and

1024: $c_1>0$ is chosen so that $\Sigma^+$ intersects the slow manifold $S$ transversally.

1025: In addition, we require that $\delta=o(\epsilon^{2/3})$ to guarantee that $\Sigma^+$

1026: belongs to the region of validity of the local analysis of Section 3 (see Remark 3.1c).

1027: Let $x_0\in \Sigma^+$ and  consider a trajectory of (\ref{6.5}) 

1028: starting from $x_0$.

1029: We assume that for sufficiently small $\epsilon>0$ and $\alpha=O(\epsilon^2)$,

1030: every such trajectory  intersects $\Sigma^-$ from the left. Denote the point of the first 

1031: intersection by $Q(x_0)\in\Sigma^-$.

1032: We assume that 

1033: 

1034: \noindent

1035: {\bf(G1)} the first return map $Q:\Sigma^-\to\Sigma^+$ depends smoothly on $\epsilon$ and 

1036: $\min_{x\in Q\left(\Sigma^+\right)} \left|x\right|\ge\zeta>0$.

1037: 

1038: To measure the rate of contraction toward $W^s(O)$, we consider a $2\times 2$ matrix

1039: \be\lbl{4.5}

1040: A\left(x_1,\alpha\right)=\left(\begin{array}{cc}

1041: {\partial f_2\over \partial x_2} &{\partial f_2\over \partial x_3} \\

1042:  {\partial f_3\over \partial x_2} &{\partial f_3\over \partial x_3} 

1043: \end{array}

1044: \right)_{(x_1,0,0,\alpha)}.

1045: \ee

1046: Let $\lambda_{1}(x_1,\alpha)\le \lambda_2(x_1,\alpha)$ denote the eigenvalues of the symmetric matrix 

1047: $$

1048: A^s\left(x_1,\alpha\right)={1\over 2}\left(A\left(x_1,\alpha\right)+A^T\left(x_1,\alpha\right)\right).

1049: $$

1050: Denote $\underline\lambda(x_1,\alpha)=-\lambda_1(x_1,\alpha)$ and 

1051: $\bar\lambda(x_1,\alpha)=-\lambda_2(x_1,\alpha)$. 

1052: We assume that for sufficiently small $\epsilon>0$ and $\alpha=O(\epsilon^2)$,

1053: 

1054: \noindent

1055: {\bf(G2)} 

1056: \begin{eqnarray}\lbl{4.6}

1057: {\p \over \p x_1} \bar\lambda(x_1,\alpha) <0, & x_1\in[d_1,0];&\\

1058: \lbl{4.7}

1059: \exists\; d_2 \in (d_1,0):\; \bar\lambda(x_1,\alpha)>0,& x_1\in[d_1,d_2]&\quad \& 

1060: \quad \min_{x_1\in[d_1, d_2]} {\bar\lambda(x_1,\alpha)\over w(x_1,\alpha)} = 

1061: O\left(\left|\ln\epsilon\right|\right),

1062: \end{eqnarray}

1063: 

1064: \noindent

1065: {\bf(G3)} 

1066: $$

1067: \max_{x_1\in[d_1, 0)} {\underline\lambda(x_1,\alpha)\over w(x_1,\alpha)} = 

1068:            O\left(\left|\ln\epsilon\right|\right).

1069: $$

1070: 

1071: Under these conditions, we have

1072: 

1073: \noindent

1074: {\bf Theorem 4.1}

1075: {\sc

1076: Let $\gamma>0$, conditions in (\ref{4.3a}), {\bf (G1)} and {\bf (G2)} hold. 

1077: Then for sufficiently small $\epsilon>0$ and $\alpha=O\left(\epsilon^2\right),$ a trajectory

1078: of (\ref{6.5}) with initial condition from an $O(\epsilon^2)$ neighborhood of the

1079: origin and not belonging to $W^s(O)$ undergoes multimodal oscillations.

1080: The ISIs are uniformly bounded from below

1081: \be\lbl{4.8b}

1082: \tau_i\ge \tau^-={1\over 2\alpha}\ln\left(1+{\alpha\over\gamma\epsilon^4}C^-\right),\;\; i=2,3, \dots.

1083: \ee

1084: If, in addition, {\bf (G3)} holds then the ISIs satisfy two-sided bounds

1085: \be\lbl{4.8a}

1086: \tau^-\le \tau_i \le \tau^+={1\over 2\alpha}\ln\left(1+{\alpha\over\gamma\epsilon^{4+\chi}}C^+\right),\;\; i=2,3, \dots,

1087: \ee

1088: for some $\chi>0$.

1089: Positive constants $C^\pm$ do not depend on $\epsilon,$ $\alpha,$ and $\gamma$.

1090: }

1091: \newpage

1092: 

1093: \noindent

1094: {\bf Remark 4.1}

1095: \begin{description}

1096: \item[(a)]

1097: The principal assumptions on the global vector field are formulated in {\bf (G1)} and

1098: (\ref{4.7}). The condition in (\ref{4.6}) makes the derivation of certain  estimates

1099: in the proof of the theorem easier and is used for analytical convenience. Likewise,

1100: {\bf (G3)} is not essential for the proposed mechanism. However, without this condition

1101: obtaining the two-sided estimates for the ISIs requires an additional argument in the 

1102: proof. Condition {\bf (G3)} means that the two eigenvalues of $A$ have the same order

1103: of magnitude. This condition is not restrictive.

1104: \item[(b)]

1105: For fixed $\gamma>0$, the estimates in (\ref{4.8a}) and (\ref{4.8b}) can be rewrittten as 

1106: \be\lbl{4.8c}

1107: {\ln(1+\tilde C^-\alpha)\over 2\alpha} \le \tau_i\le {\ln(1+\tilde C^+\alpha)\over 2\alpha},

1108: \ee

1109: with constants $\tilde C^\pm$ independent from $\alpha$. 

1110: Similarly, for fixed $\alpha$ and varying $\gamma>0$ 

1111: inequalities in (\ref{4.8a}) and (\ref{4.8b}) imply

1112: \be\lbl{4.8d}

1113: \bar C^- -{1\over 2\alpha}\ln\gamma \le \tau_i \le \bar C^+ -{1\over 2\alpha}\ln\gamma,

1114: \ee

1115: where constants $\bar C^\pm=O\left(\left|\ln\epsilon\right|\right)$ do not depend on $\gamma>0$.

1116: \end{description}

1117: 

1118: 

1119: \begin{figure}

1120: \begin{center}

1121: \epsfig{figure=f7.eps, height=1.8in, width=2.0in, angle=0}

1122: \end{center}

1123: \caption{ 

1124: The interpretation of the bifurcation scenario arising in (\ref{6.1})-(\ref{6.3}) for

1125: fixed positive value of $\alpha=O(\epsilon^2)$ and varying $\gamma$ near $0$ using

1126: a one-parameter family of $1D$ maps. For negative values of $\gamma=O(1)$, the map

1127: in (\ref{9.4}) has a stable fixed point (upper graph), which corresponds to a 

1128: stable limit cycle born from the supercritical AHB. For increasing values of $\gamma$

1129: the fixed point looses stability via a period-doubling bifurcation (lower graph),

1130: which initiates a period-doubling cascade leading to the formation of the chaotic attractor.

1131: }\label{f.7}

1132: \end{figure}

1133: 

1134: \subsection{The transition from subcritical to supercritical AHB}

1135: 

1136: The numerical experiments presented in Section 2 show that the transition from subcritical to

1137: supercritical AHB contains a distinct bifurcation scenario involving the formation

1138: of chaotic attractor via a period-doubling cascade. The analytical explanation of these 

1139: phenomena is outside the scope of the present paper.

1140: Below we comment on the difficulties arising in the analytical treatment of

1141: this problem. In the present section, we use a combination of the analytic and numerical techniques

1142: to elucidate the origins of the complex dynamics arising in the parameter regime near

1143: the border between regions of sub- and supercritical AHB.

1144: The principal features of this bifurcation scenario are 

1145: summarized in Figures \ref{f.4} and \ref{f.5}. In these numerical experiments, we

1146: kept $\alpha$ fixed at a small positive value and varied $\gamma$.

1147: By taking progressively smaller values of $\gamma>0$, one first observes that

1148: the multimodal patterns exhibit an increase in the ISIs (Figure \ref{f.4}a,b; see also Figure \ref{f.3}b).

1149: We consider these oscillatory patterns regular (even if they are not periodic),

1150: because the timings of the spikes remain within narrow bounds in accord with (\ref{4.8a}).

1151: For smaller values of $\gamma>0$, the oscillatory patterns become irregular and

1152: are characterized by long intervals of oscillations of small and intermediate amplitudes

1153: between successive spikes (Figure \ref{f.4} c,d).

1154: As $\gamma$ becomes negative, the trajectories lose spikes and

1155: consist of irregular oscillations 

1156: (Figure \ref{f.5} a,b). Further decrease of $\gamma$ leads the system through

1157: the reverse cascade of period-doubling bifurcations (Figure \ref{f.5} c-h).

1158: The period-doubling cascade terminates with the creation of the limit cycle,

1159: which can be followed to a nondegenerate supercritical AHB by letting $\alpha\to 0$

1160: (Figure \ref{f.5} g,h).

1161: To account for the bifurcation scenario described above, we construct the first

1162: return map. Since near the origin the trajectories spend most of the time in the vicinity of

1163: the $2D$ slow manifold, the first return map is effectively one-dimensional. The distinct

1164: unimodal structure of the $1D$ first return map affords a lucid geometric interpretation

1165: for the bifurcation scenario in the $3D$ systems of differential equations near the

1166: transition from sub- to supercritical AHB. Specifically, we show that the mechanism

1167: for generating complex dynamics in the continuous system in this parameter regime is

1168: the same as in the classical scenario of the period-doubling transition to chaos in

1169: the one-parameter families of unimodal maps \cite{GH}. 

1170: \begin{figure}

1171: \begin{center}

1172: {\bf a}\epsfig{figure=f8a.eps, height=1.8in, width=1.8in, angle=0}\quad

1173: {\bf b}\epsfig{figure=f8b.eps, height=1.8in, width=1.8in, angle=0}\\

1174: {\bf c}\epsfig{figure=f8c.eps, height=1.8in, width=1.8in, angle=0}\quad

1175: {\bf d}\epsfig{figure=f8d.eps, height=1.8in, width=1.8in, angle=0}

1176: \end{center}

1177: \caption{ The plots of first return map for $\mI$ (\ref{9.6}) computed for

1178: ({\bf a}) $\alpha= 0.0445$, $\gamma=-0.0486$,

1179: ({\bf b}) $\alpha= 0.0445$, $\gamma=-0.0126$,

1180: ({\bf c}) $\alpha= 0.0445$, $\gamma=0.0290$,

1181: ({\bf d}) $\alpha=0.12 $, $\gamma=-1.0741$.

1182: The plots in ({\bf a}-{\bf c}) illustrate the bifurcations in the family of the first return maps

1183: for the values of $\gamma$ near zero. The map shown in ({\bf d}) corresponds to the 

1184: irregular oscillations following the period-doubling cascade for increasing values

1185: of $\alpha$ in the case of supercritical AHB (see Remark 4.2). 

1186: }\label{f.8}

1187: \end{figure}

1188: 

1189: We next turn to the derivation of the first return map.

1190: We start with extending the asymptotic analysis of Section 3 to cover the case of $\gamma=o(1)$. 

1191: This requires computing additional terms on the right hand side of the reduced equation

1192: (\ref{5.24}), because already for 

1193: $\left|\gamma\right|=O(\epsilon)$, the term involving $\gamma$ on the right hand side of (\ref{5.24})

1194: is comparable with the $O(\epsilon^3)$ remainder term.

1195: We then use the more accurate approximation for the equation for $I$ to compute a

1196: $1D$ mapping:

1197: \be\lbl{9.0}

1198: G:\; I(\phi)\mapsto I\left(\phi + 2\pi\beta^{-1}\right),

1199: \ee

1200: which describes how  $I$ changes after one cycle of oscillations. 

1201: As noted above, the reduced equation (\ref{5.24}) is not suitable for analyzing

1202: the case of small $\left|\gamma\right|$ and one needs to include more terms in the expansion on the right hand side of (\ref{5.24}).  

1203: In analogy with the topological normal form for the degenerate AHB 

1204: \cite{KUZ, GL}, we expect that terms up to $O(\epsilon^5)$) are needed in (\ref{5.24}) to resolve the dynamics 

1205: for small $\left|\gamma\right|$.  This can be achieved by a straightforward albeit tedious calculation.  Below we describe the 

1206: formal procedure of obtaining the required expansions.  The justification of these expansions can be done in complete analogy 

1207: with the analysis of Section 3.  First, we compute terms on the right hand sides of (\ref{5.11a}) and (\ref{5.12}) up to $O(\epsilon^3)$ and $O(\epsilon^6)$ respectively.  Then we look for a solution of (\ref{5.11a}) in the following form:

1208: \be\lbl{9.1}

1209: \xi(\phi)=\xi_0(r,\phi)+\epsilon\xi_1(r,\phi)+\epsilon^2\xi_2(r,\phi)+O(\epsilon^3).

1210: \ee

1211: By taking an initial  condition from $D_0$ and using Anzats (\ref{9.1}) in (\ref{5.11a}), we recover $\xi_0(r,\phi)$ 

1212: (see (\ref{5.16})) and find the next 

1213: two terms in the expansion of $\xi(r,\phi)$: $\xi_1(r,\phi)$ and $\xi_2(r,\phi)$.  Thus, we obtain the approximation of the slow manifold 

1214: with the accuracy $O(\epsilon^3)$.  Next, we plug in (\ref{9.1}) into (\ref{5.12}) and collect terms multiplying equal powers of $\epsilon$

1215: to obtain 

1216: \be\lbl{9.2}

1217: \dot r = \epsilon R_1(r,\phi)+\dots+\epsilon^5 R_5(r,\phi)+O(\epsilon^6),

1218: \ee

1219: where $R_k(r,\phi)$ are $2\pi\beta^{-1}$– periodic functions of the second argument.

1220: Finally, by rewriting (\ref{9.2}) in terms of $I$ (see (3.25)), integrating it over one period of oscillations, 

1221: $\omega$, and disregarding $O(\epsilon^5)$ terms, we obtain a map for the change in $I$ after one cycle of oscillations

1222: \be\lbl{9.4}

1223: I_{n+1}=G(I_n),\;\; G(I)\equiv I-2\epsilon^2\omega\left(\mu I +\gamma +{\epsilon^2c\over I}\right),\; \omega=2\pi\beta^{-1}.

1224: \ee

1225: Except for the last term in the definition of $G$, the map in (\ref{9.4}) follows from the reduced system (\ref{5.24}) and (\ref{5.25}).  

1226: The last term is only needed if the value of $\left|\gamma\right|$ does not exceed $O(\epsilon)$.  Our calculations show that 

1227: the second Lyapunov coefficient, $c$, is negative for the values of parameters used in (\ref{6.1})-(\ref{6.3}).

1228: Therefore, for sufficiently small $\epsilon>0$, (\ref{9.4}) defines a unimodal map.  Away from an $O(\epsilon^2)$ neighborhood of 

1229: $0$, the graph of  $\tilde I=G(I)$ is almost linear with a weakly attracting slope $1-2\omega\mu\epsilon^2$ for $\mu>0$ (Figure \ref{f.7}).

1230: For $\mu>0$, $G$ has a unique fixed point:

1231: \be\lbl{9.5}

1232: \bar I=\left\{

1233: \begin{array}{cc}

1234: {-\gamma\over\mu} +{c\over\gamma}\epsilon^2+O(\epsilon^3),& \gamma<0,\\

1235: \sqrt{-c\over\mu}\epsilon +O(\epsilon^2),&\gamma=0,\\

1236: {-c\over\gamma}\epsilon^2+O(\epsilon^3),& \gamma>0.

1237: \end{array}

1238: \right.

1239: \ee

1240: For negative values of $\gamma=O(1)$, $I=\bar I$ is a stable fixed point of $G$, which corresponds to a stable limit cycle  born

1241: from a  supercritical AHB (see Figure \ref{f.5} g,h). It follows from (\ref{9.4}) that for increasing values of $\gamma$, 

1242: the graph of the map moves down. Consequently, 

1243: the fixed point moves to the left and eventually looses stability via a period-doubling bifurcation (Figure \ref{f.5} e,f).  

1244: Further increase in $\gamma$, yields more period-doubling bifurcations (Figure \ref{f.5} c,d).  Given the unimodal character of 

1245: $G$, one expects that this sequence of period-doubling bifurcations eventually leads to the formation of a chaotic attractor 

1246: (Figure \ref{f.5} a,b). The unimodal character of the map combined with the manner of its

1247: dependence on $\gamma$ suggest a clear geometric mechanism for the formation of the chaotic attractor and the subsequent

1248: period-doubling cascade arising near the transition from sub- to supercritical AHB (Figure \ref{f.5}a-h). 

1249: We now outline the  limitations of the asymptotic analysis of this section and difficulties

1250: arising in the justification of (\ref{9.4}).

1251: According to (\ref{9.5}) already at the moment of the first period-doubling bifurcation, fixed point $\bar I$

1252: belongs to an $O(\epsilon^2)$ neighborhood of $0$. In this neighborhood,

1253: $\rho=O(1)$ (see (\ref{5.22}))  and, therefore, $\bar I$ lies outside of the region of validity of

1254: the asymptotic analysis. In this region, (\ref{9.4}) may only be considered as a formal 

1255: asymptotic expression. A rigorous justification of the from of the map in the 

1256: boundary layer meets substantial analytical difficulties: it requires introducing an additional set 

1257: of intermediate asymptotic expansions and matching them with those obtained in Section 3. 

1258: We do not address this problem in the present work. 

1259: Below, we resort to using numerical techniques to verify the principal features

1260: of the first return map suggested by the asymptotic analysis: the unimodality of $G$ and its

1261: dependence on $\gamma$. The numerics confirmed our predictions about the form

1262: of the first return map and it also revealed certain additional features.

1263: 

1264: For numerical construction of the first-return map, we fix angle $\bar \theta \in [0, 2\pi)$

1265: and define a crossection in cylindrical coordinates:

1266: $

1267: \Sigma=\{(\rho, \xi, \theta): \theta=\bar \theta\}.

1268: $

1269: Let $(\rho(t), \xi(t), \theta(t))$ be a trajectory of (\ref{6.5}) and $0<t_1<t_2<t_3<\dots<t_k<\dots$

1270: denote a sequence of times, at which $(\rho(t_k),\xi(t_k),\theta(t_k))\in \Sigma,\;k=1,2,3,\dots$.

1271: Then, we define 

1272: \be\lbl{9.6}

1273: \mG: \mI(t_k) \mapsto \mI(t_{k+1})\quad\mbox{and}\quad \mI(t) =\frac{1}{\rho^2(t)}.

1274: \ee

1275: By a suitable choice of  $\bar\theta$, we can achieve that a trajectory of

1276: (\ref{6.5}), intersects $\Sigma$ once during one cycle of oscillations. Therefore,

1277: for a trajectory lying in a $2D$ slow manifold, (\ref{9.6}) defines a 

1278: $1D$ first-return map. Note that maps (\ref{9.4}) and (\ref{9.6}) are 

1279: related via rescaling of the coefficients, since to leading order $ \mI=\epsilon^{-2}I$.

1280: Therefore, the graphs of $G(I)$ and $\mG(\mI)$ are

1281: similar in the domain, where (\ref{9.4}) is valid. We computed the first return maps

1282: for fixed $\alpha=0.0445$ and for several values of $\gamma$. 

1283: The representative plots are shown in Figure \ref{f.8} a-c.

1284: The numerically computed maps confirmed  our expectations about the graph of

1285: $\mG$ in the boundary layer near $0$: in this region, the map is decreasing

1286: with a strongly expanding slope. 

1287: Away from $0$, the graph of $\mG$ contains an almost linear branch,

1288: whose slope is very close to $1-2\alpha\omega,$ as predicted by (\ref{9.4})

1289: (see lower branches in Figure \ref{f.8}b,c).

1290: These numerics also confirm that under the variation of $\gamma$, the graph of the first return

1291: map is translated in vertical direction (Figure \ref{f.8} a-c). Therefore, the family of the first return maps 

1292: possesses two principal ingredients (the unimodality and the additive dependence on $\gamma$),

1293: which are necessary for the qualitative explanation of the bifurcation scenario 

1294: given in Figure \ref{f.7}. In addition, our numerical experiments reveal

1295: a new feature of the first return map: for small $|\gamma|$, the graph of the map has 

1296: another almost linear branch in the outer region away from the origin (see Figure \ref{f.8}b).

1297: This branch of the graph of the map also has an attracting slope.

1298: The presence of this branch in the first return map indicates the existence of another (branch of the) 

1299: slow manifold different from that described in Section 3. 

1300: A possible explanation for the appearance

1301: of the upper branch in the first return map is due to the unstable manifold of the periodic orbit.

1302: For small $|\gamma|$, the first-return map (\ref{9.6}) is multivalued 

1303: in the outer region. However, since both the upper and the lower branches have positive slopes less than $1$, 

1304: the qualitative dynamics for the map shown in Figure \ref{f.8}b does not depend on the exact mechanism for

1305: the selection between the branches in (\ref{9.6}). Although explaining the multivaluedness of 

1306: the first return maps shown in Figures \ref{f.8}b,c presents an interesting problem, it is not 

1307: critical for the qualitative explanation of the bifurcation scenario arising during the transition from 

1308: the subcritical to supercritical AHB. The latter was the main goal of the present subsection.

1309: 

1310: \noindent

1311: {\bf Remark 4.2 }

1312: In addition, to the region in the parameter space containing the border between the regions of

1313: sub- and supercritical AHB, there is another parameter regime resulting in

1314: the complex dynamics.

1315: In \cite{FKRT}, it was shown numerically that the limit cycle born from the supercritical

1316: AHB in (\ref{6.5}) undergoes a period-doubling cascade leading to the formation of the chaotic attractor

1317: for increasing values of $\alpha$. The first period-doubling bifurcation in this cascade is shown in Figure \ref{f.1}c,d.

1318: This bifurcation scenario is consistent with the form of the first-return map constructed 

1319: in this subsection. Indeed, it is easy to see from (\ref{9.5}) that for increasing values of $\mu$ the fixed

1320: point, $\bar I$, moves to the left. Therefore, the explanation given above for the period-doubling cascade

1321: resulting from the variation of $\gamma$ near $0$ also applies to the case of increasing $\alpha$ and

1322: fixed $\gamma <0$ (see Figure \ref{f.8}d).

1323: 

1324: 

1325: \section{The proof of Theorem 4.1}

1326: \setcounter{equation}{0}

1327: 

1328: In the present section, we show that under the assumptions of Theorem 4.1

1329: the trajectories of (\ref{2.5}) with initial conditions in $\Sigma^-$ enter $D_0$.

1330: The local analysis in Section 3 describes the behavior of trajectories from

1331: the moment they reach $D_0$ and until they leave $D$. In particular,

1332: it shows that the dynamics near the origin has two phases: the fast approach to 

1333: the slow manifold and the slow drift away from the origin along the slow manifold.

1334: Upon leaving $D$, the trajectories are reinjected back to $\Sigma^-$ by the  

1335: return mechanism postulated in {\bf (G1)}. This scenario implies that the 

1336: system undergoes multimodal oscillations as stated in Theorem 4.1.

1337: \begin{figure}

1338: \begin{center}

1339: \epsfig{figure=f9.eps, height=3.0in, width=5.0in, angle=0}

1340: \end{center}

1341: \caption{ A schematic representation of the trajectory of (\ref{6.1})-(\ref{6.3})  approaching

1342: a weakly unstable saddle-focus. The trajectory crossing $\Sigma^-$ is subject to strong vector

1343: field transverse to $W^s(O)$ in $\Pi^1_{\delta_1}$. This guarantees that it remains close to $W^s(O)$

1344: until it enters an $O(\epsilon^2)$ neighborhood of the origin. From that moment and until the trajectory

1345: reaches $\Sigma^+$ its behavior is described by the local analysis of Section 3.

1346: }

1347: \label{f.9}

1348: \end{figure}

1349: 

1350: We start with presenting several auxiliary estimates, which will be needed for the

1351: proof. By (\ref{4.3a}), one can choose $C_1>0$ such that

1352: \be\lbl{7.1}

1353: w(x_1, \alpha)\ge C_1 \left|x_1\right|,\quad x_1\in\left[ d_1,0\right]

1354: \ee 

1355: and for sufficiently small $\alpha\ge 0$.

1356: Next, we note that $\bar\lambda(0,0)=0$. This equation and (\ref{4.6}), by  the Implicit

1357: Function Theorem, imply that there exists $0>x^*(\alpha)=O(\epsilon^2)$ such that

1358: $$

1359: \bar\lambda(x_1,\alpha)\left\{\begin{array}{ll}

1360: <0,& x_1\in[d_1, x_1^*(\alpha)),\\

1361: >0,& x_1\in(x_1^*(\alpha),0].

1362: \end{array}

1363: \right.

1364: $$

1365: 

1366: Using (\ref{4.6}) and (\ref{7.1}), we have 

1367: $$

1368: {\bar\lambda(x_1,\alpha)\over w(x_1,\alpha)}\ge {\bar\lambda(x_1^*,\alpha)\over w(x_1,\alpha)}=0,\quad x_1\in [d_1, x^*_1(\alpha)],

1369: $$

1370: where $w(x_1,\alpha)$ is defined in (\ref{4.3a}).

1371: 

1372: Let $\bar\mu>0$ be such that ({\bf G1}) and ({\bf G2}) hold for $\alpha\in [0,\bar\mu\epsilon^2]$. 

1373: In the remainder of this section, unless stated otherwise,  it is assumed that $\alpha\in 

1374: \left[0,\bar\mu\epsilon^2 \right]$. To simplify notation, we will often omit the dependence 

1375: of various functions on $\alpha$.

1376: Using (\ref{4.3a}), we rewrite (\ref{6.5}) in the following form:

1377: \begin{eqnarray}\lbl{7.2}

1378: \dot x_1 &=& w(x_1)+y_1\phi_1(x_1,y)+y_2\phi_2(x_1,y),\\ 

1379: \lbl{7.3}

1380: \dot y &=& A(x_1)y+\psi(x_1,y),

1381: \end{eqnarray}

1382: where $w(x_1)$ and $A(x_1)$ are defined in (\ref{4.3a}) and (\ref{4.5}) respectively;

1383: $y=(y_1,y_2):=(x_2,x_3)$, and 

1384: \be\lbl{7.4}

1385: w(0)=0,\; \psi(x_1,y)=O\left(\left|y\right|^2\right),\; \phi_{1,2}(x_1,y)=\phi_{1,2}(x_1)+O(\left|y|\right),\quad

1386: x\in \left[d_1, 0\right].

1387: \ee

1388: Let 

1389: \be\lbl{7.4a}

1390: d_3=-M^2\epsilon^2

1391: \ee

1392: denote the $x_1-$coordinate on the left lateral boundary of $D_0$ (see (\ref{5.2})) and $\Pi_{\delta_1}=[d_1,d_3]\times D_{\delta_1}$

1393: (see Figure \ref{f.9}).

1394: From (\ref{6.5}) and (\ref{4.3a}), one can see that  for ${\delta_1}>0$ sufficiently small (independent of $\epsilon$),

1395: the right hand side of (\ref{7.2})

1396: $$

1397: w(x_1)+y_1\phi_1(x_1,y)+y_2\phi_2(x_1,y)>0,\quad (x_1,y)\in\Pi_{\delta_1}.

1398: $$

1399: Therefore, in $\Pi_{\delta_1}$, (\ref{7.2}) and

1400: (\ref{7.3}) may be rewritten as follows

1401: \be\lbl{7.5}

1402: {d y(x_1)\over d x_1}=\tilde A(x_1) y +\tilde\psi(x_1,y),

1403: \ee

1404: where 

1405: $$

1406: \tilde A(x_1)={1\over w(x_1)} A(x_1)\quad\mbox{ and}\quad \tilde\psi(x_1,y)={\psi(x_1,y)\over w(x_1,y)}\left(1+O\left(\left|y\right|\right)\right).

1407: $$

1408: By (\ref{7.1}) and (\ref{7.4}), we have

1409: \be\lbl{7.6}

1410: \left|\tilde\psi(x_1,y)\right|\le {C_2\left|y\right|^2\over \left|x_1\right|},\quad (x_1,y)\in\Pi_{\delta_1},

1411: \ee

1412: for some $C_2>0$ independent of $\epsilon$.

1413: 

1414: To follow the trajectories from $\Sigma^-$ to $D_0$, we introduce two regions:

1415: $$

1416: \Pi_{\delta_1}^1=[d_1,d_2]\times D_{\delta_1}\quad\mbox{and}\quad\Pi_{\delta_2}^2=[d_2,d_3]\times D_{\delta_2}\;\;\mbox{(see Figure \ref{f.9})}.

1417: $$

1418: Recall that $D_{\delta_{1,2}}$ denotes the disk of radius $\delta_{1,2}$ (see (\ref{4.4a})).

1419: Positive constant $\delta_1$ is the same as in (\ref{4.4}) and $0<\delta_2\ll \delta_1$ will be specified later (Figure \ref{f.9}). 

1420: Recall that $d_3=-M\epsilon^2$ denotes the value of the $x_1-$coordinate on the the left lateral boundary of $D_0$ (see (\ref{5.2})).

1421: By taking $M>0$ large enough, we can arrange  $d_3<x^*(\alpha)<0$ for $\alpha\in \left[0,\bar\mu\epsilon^2 \right]$.

1422: By {\bf(G1)}, the vector field in $\Pi^1_{\delta_1}$ is sufficiently strong so that the trajectories entering $\Pi_{\delta_1}^1$

1423: through $\Sigma^-$ get into a narrow domain $\Pi_{\delta_2}^2$ and remain there until they reach $D_0$.

1424: The following lemma allows to control the trajectories in $\Pi_{\delta_1}=[d_1,d_3]\times D_{\delta_1}$.

1425: 

1426: \noindent{\bf Lemma 5.1}

1427: {\sc Let $[\ell_1,\ell_2]\subseteq[d_1,d_3],\; 0<\bar\delta\le {\delta_1},$ and

1428: \be\lbl{7.7}

1429: -\upsilon=\max_{x_1\in[\ell_1,\ell_2]}\sup_{\left|y\right|=1}\left(\tilde A(x_1)y,y \right)<0.

1430: \ee

1431: Then

1432: \be\lbl{7.8}

1433: \left|y(\ell_1)\right|\le\bar\delta\quad\Rightarrow\quad \left|y(x_1)\right|\le 2\bar\delta e^{-\upsilon(x_1-\ell_1)},\; x_1\in[\ell_1,\ell_2],

1434: \ee

1435: provided

1436: \be\lbl{7.9}

1437: {4 C_2\bar\delta \left(1- e^{-\upsilon(\ell_2-\ell_1})\right)\over\left|\ell_2\right|\upsilon}\le 1.

1438: \ee

1439: }

1440: \noindent

1441: {\bf Proof:}

1442: We may assume that $|y(x_1)|\not=0,x \in [\ell_1,\ell_2]$, since, otherwise, 

1443: by the uniqueness of solution of the initial value problem for (\ref{7.5}), $y(x_1)=0,$ $x_1\in[\ell_1,\ell_2]$,

1444: and (\ref{7.8}) holds. From (\ref{7.5}), (\ref{7.6}), and (\ref{7.7}), we have

1445: \be\lbl{7.10} 

1446: \frac{d|y(x_1)|}{dx_1} \le -\upsilon \left|y(x_1)\right| + \bar\psi\left(x_1,\left|y\right|\right),\;\;

1447: \bar\psi\left(x_1,\left|y\right|\right)={ C_2 \left|y \right|^2 \over \left|x_1 \right|},\;\; x_1\in[\ell_1,\ell_2],\;

1448: \left|y\right|\le\left|\delta_1\right|.

1449: \ee

1450: where $\upsilon$ is defined in (\ref{7.7}), and $\left|y\right|$ denotes the Euclidean norm of $y\in\Re^2$.

1451: Let $\bar y(x_1)$ denote the solution of the initial value problem 

1452: \be\lbl{7.11}

1453: \bar y^\prime=-\upsilon \bar y +\bar\psi(x_1,\bar y),\;

1454: \bar y(\ell_1)=\bar\delta.

1455: \ee

1456: It is sufficient to show that (\ref{7.8}) holds for

1457: $\bar y(x_1)$. We represent $\bar y(x_1)$ as the limit of the sequence

1458: of successive approximations

1459: \be\lbl{7.12}

1460: \bar y^{(n+1)}(x_1)=\bar\delta e^{-\upsilon(x_1-\ell_1)}+\int_{\ell_1}^{x_1} e^{-\upsilon(x_1-s)}

1461: \bar\psi\left(s,\bar y^{(n)}(s)\right) ds,\quad n=0,1,2,\dots

1462: \ee

1463: and $\bar y^{(0)}(x_1)=\bar \delta e^{-\upsilon(x_1-\ell_1)}.$

1464: We use induction to show 

1465: \be\lbl{7.13}

1466: \left|\bar y^{(n)}(x_1)\right| \le 2 \bar\delta e^{-\upsilon(x_1-\ell_1)},\quad x \in [\ell_1,\ell_2],\quad

1467: n=0,1,2,\dots .

1468: \ee 

1469: Inequality (\ref{7.13}) holds for $n=0$. We show that $(\ref{7.13})_{n=k} \Rightarrow (\ref{7.13})_{n=k+1}$.

1470: Using the definition of $\bar\psi$ in (\ref{7.10}) and $(\ref{7.13})_{n=k}$, we have

1471: \begin{eqnarray}\nonumber

1472: \int_{\ell_1}^{x_1} e^{-\upsilon(x_1-s)}\bar\psi(s,\bar y^{(k)}(s) ds &\le&

1473: \frac{C_2}{|\ell_2|}\int_{\ell_1}^{x_1} e^{-\upsilon(x_1-s)}\left|\bar y^{(k)}(s)\right|^2 ds\le

1474: \frac{C_2}{|\ell_2|}\int_{\ell_1}^{x_1}e^{-\upsilon (x_1-s)} 4 {\bar\delta}^2 e^{-2\upsilon(s-\ell_1)} ds\\ 

1475: \lbl{7.14}

1476: &=&\bar\delta e^{-\upsilon(x_1-\ell_1)} {4 C_2\bar\delta \left( 1- e^{-\upsilon(\ell_2-\ell_1)}\right)\over\left|\ell_2\right|\upsilon}.

1477: \end{eqnarray}

1478: Using (\ref{7.14}) and (\ref{7.9}),

1479: from $(\ref{7.13})_{n=k}$ we obtain  $(\ref{7.13})_{n=k+1}$. By induction, (\ref{7.13}) holds. By taking $n\to\infty$ 

1480: in (\ref{7.13}), we have 

1481: \be\lbl{7.15}

1482: \left|\bar y(x_1)\right|\le 2\bar\delta e^{-\upsilon(x_1-\ell_1)},\; x_1\in[\ell_1,\ell_2].

1483: \ee

1484: The statement in (\ref{7.8}) then follows from (\ref{7.15}) and the standard theory for differential inequalities

1485: \cite{hartman}.\\

1486: $\Box$

1487: 

1488: In the following lemma, we determine the size of $\Pi^2_{\delta_2}$.

1489: 

1490: \noindent{\bf Lemma 5.2}

1491: {\sc There exists positive $\delta_2=O(\epsilon^2)$, such that the trajectories of (\ref{7.2}) and (\ref{7.3}) 

1492: entering $\Pi^2_{\delta_2}$ from $\Pi^1_{{\delta_1}}$ remain in $\Pi^2_{\delta_2}$ until they reach $D_0$ (see Figure \ref{f.9}.)

1493: }

1494: 

1495: \noindent {\bf Proof:} 

1496: Denote

1497: $$

1498: -\upsilon_2=\max_{x_1\in[d_2,d_3]}\sup_{\left|y\right|=1}\left(\tilde A(x_1)y, y\right).

1499: $$

1500: Recall 

1501: $$

1502: \bar\lambda (x^*)=0 \quad \mbox{and}\quad O(\epsilon^2)= d_3< x^*=O(\epsilon^2).

1503: $$

1504: Therefore, by (\ref{4.6}), $\bar\lambda(d_3)=O(\epsilon^2)$ is positive.

1505: In addition,

1506: \begin{eqnarray*}

1507: \upsilon_2&=&-\max_{x_1\in[d_2,d_3]} \sup_{\left|y\right|=1}\left(\tilde A(x_1)y, y\right)=

1508: \min_{x_1\in[d_2,d_3]}{-\sup_{\left|y\right|=1}\left( A(x_1)y, y\right)\over w(x_1)}\\

1509: &=& {\min_{x_1\in[d_2,d_3]} \bar\lambda(x_1)\over\max_{x_1\in[d_2,d_3]} w(x_1)}= 

1510: {\bar\lambda(d_3)\over\max_{x_1\in[d_2,d_3]} w(x_1)}= O(\epsilon^2). 

1511: \end{eqnarray*}

1512: and $\upsilon_2>0$. 

1513: Next we apply Lemma 5.1 with $\bar\delta:=\delta_2,$ $\upsilon:=\upsilon_2$, and $\ell_{1,2}:=d_{2,3}$. Note that for small $\upsilon_2$

1514: the inequality (\ref{7.9}) can be rewritten as 

1515: \be\lbl{7.16}

1516: {4 C_2\delta_2 \left(d_3-d_2 +O(\upsilon_2^2)\right)\over\left|d_3\right|}\le 1.

1517: \ee

1518: Since $d_3=O(\epsilon^2)$  (see (\ref{7.4a})) and $\upsilon_2=O(\epsilon^2)$, one can choose $\delta_2=O(\epsilon^2)$ so 

1519: that (\ref{7.16}) holds.\\

1520: $\Box$

1521: 

1522: Having found the size of $\Pi^2_{\delta_2}$, we now determine the rate of contraction in $\Pi^1_{\delta_1}$ sufficient

1523: to funnel the trajectories entering $\Pi^1_{\delta_1}$ through $\Sigma^-$ to $\Pi^2_{\delta_2}$.

1524: 

1525: \noindent{\bf Lemma 5.3}

1526: {\sc The trajectories of (\ref{7.2}) and (\ref{7.3}) 

1527: entering $\Pi^1_{\delta_1}$ through $\Sigma^-$ remain in $\Pi^1_{\delta_1}$ until they reach 

1528: $\Pi^2_{\delta_2}$.

1529: }

1530: 

1531: \noindent {\bf Proof:} 

1532: Denote

1533: $$

1534: -\upsilon_1=\max_{x_1\in[d_1,d_2]}\sup_{\left|y\right|=1}\left(\tilde A(x_1)y, y\right).

1535: $$

1536: From {\bf (G2)}, one finds that $\upsilon_1=O(\left|\ln \epsilon\right|)$ is positive.

1537: Lemma 5.3 now follows from Lemma 5.1 with  $\bar\delta:={\delta_1},$ $\upsilon:=\upsilon_1$, and $\ell_{1,2}:=d_{1,2}$.  

1538: Indeed, for $\upsilon_1=O\left(\left|\ln\epsilon\right|\right)$, we have 

1539: \be\lbl{7.17}

1540: \upsilon_1\ge {4C_5{\delta_1}\over\left|d_2\right|}.

1541: \ee

1542: Inequality (\ref{7.17}) is sufficient for (\ref{7.9}) to hold. By Lemma 5.1, we have

1543: \be\lbl{7.19}

1544: \left|y(d_2)\right|\le 2{\delta_1} e^{-\upsilon_1(d_2-d_1)}.

1545: \ee

1546: With $\upsilon_1=O\left(\left|\ln\epsilon\right|\right)$, by (\ref{7.19}), we can achieve 

1547: $\left|y(d_2)\right|<\delta_2=O(\epsilon^2)$.\\

1548: $\Box$

1549: 

1550: Lemmas 5.2 and 5.3 imply that the trajectories entering $\Pi^1_{\delta_1}$ through $\Sigma^-$

1551: stay in $\Pi^1_{\delta_1}\bigcup\Pi^2_{\delta_2}$ until they reach $D_0$. Moreover, the

1552: inequality in {\bf (G1)} guarantees that such trajectories are bounded away from $W^s(O)$.

1553: Thus, we can use the analysis of Section 3 to describe the evolution of trajectories 

1554: from the moment they reach $D_0$ until they leave $D$. By Remark 3.1c, this description extends

1555: to any region where $x_{2,3}=o\left(\epsilon^{2/3}\right),$ i.e., the trajectories can be controlled 

1556: until they hit $\Sigma^+$. This is followed by the return to $\Pi_{\delta_1}^1$ according to {\bf (G1)},

1557: and the next cycle of the multimodal oscillations begins. The analysis of this section applies to 

1558: any trajectory starting from $D_0$ and not belonging to $W^s(O)$.

1559: 

1560: It remains to estimate the ISIs. For this, we compute the time needed for the trajectory starting

1561: in an $O(\epsilon^2)$ neighborhood to return back to this neighborhood after making one global 

1562: excursion. Since the time of flight of the trajectory outside a small neighborhood of the 

1563: origin depends regularly on the control parameters, the duration of the very long ISIs is determined

1564: by the time spent in the neighborhood of the origin. To estimate the latter, 

1565: we note that at the moment a trajectory enters $D_0$, we have  

1566: \be\lbl{7.20a}

1567: \left|y(d_3)\right|\le C_3 \epsilon^2.

1568: \ee

1569: This follows from Lemma 5.3, since $\left|y(d_3)\right|\le \delta_2$ and $\delta_2=O\left(\epsilon^2\right).$

1570: To obtain the lower bound on $\left|y(d_3)\right|$, recall that by {\bf (G1)}, we have 

1571: \be\lbl{7.20b}

1572: \left|y(d_1)\right|\ge \zeta>0.

1573: \ee

1574: As follows from {\bf (G3)}, the maximal rate of contraction in $\Pi_{\delta_1}$ does not exceed 

1575: $O\left(\left|\ln\epsilon\right|\right)$ in absolute value. This combined with (\ref{7.20b}) implies that

1576: \be\lbl{7.20c}

1577: \left|y(d_3)\right|\ge C_4\epsilon^{2+{\chi\over 2}}

1578: \ee

1579: for some $\chi\ge 0$ and $C_4$ independent of $\epsilon$. We omit the proof of (\ref{7.20c}), because it is 

1580: completely analogous to that of Lemma 5.1. 

1581: 

1582: Let $t=t^-$ denote the moment 

1583: of time when a trajectory 

1584: of (\ref{6.5}) enters $D_0$ from $\Pi^2_{\delta_2}$. After switching back to the original parametrization 

1585: of $y$ by time, $t$, we rewrite (\ref{7.20a}) and (\ref{7.20c}):

1586: \be\lbl{7.20}

1587: C_3\epsilon^{2+{\chi \over 2}}\le\left|y(t^-)\right|\le C_4\epsilon^2.

1588: \ee

1589: For $t>t^-$ the trajectory approaches and remains close to the slow manifold as long as 

1590: $\left|y\right|=o(\epsilon^{2/3})$

1591: (see Remark 3.1c). Let ${2\over3}<j<1$ and denote 

1592: \be\lbl{7.21} 

1593: t^+=\min\left\{t>t^-:\;\left|y(t)\right|=\epsilon^j\right\}.

1594: \ee

1595: From (\ref{7.20}) and (\ref{7.21}), we have  the following bounds for $I_0:=I(t^-)$ 

1596: and $I_1:=I(t^+)$:

1597: \be\lbl{7.22}

1598: C_6\epsilon^{-2-\chi}\le I_0\le C_7\epsilon^{-2}

1599: \quad\mbox{and}\quad

1600: I_1=O(\epsilon^{2(1-j)}),\; j\in\left({2\over 3}, 1\right).

1601: \ee

1602: For these ranges of values of $I_0$ and $I_1$, from (\ref{5.55}) we have 

1603: \be\lbl{7.23}

1604: \tau_{in}=t^+-t^-={1\over 2\mu\epsilon^2}\ln\left(\left(1+{\mu\over\gamma}I_0\right)

1605: \left(1+{\mu\over\gamma}o(1)\right)\right).

1606: \ee

1607: The combination of (\ref{7.22}) and (\ref{7.23}) yields two-sided bounds for $\tau_{in}$:

1608: \be\lbl{7.24}

1609: \tau^-_{in}\le \tau_{in}\le \tau_{in}^+,

1610: \ee

1611: where

1612: $$

1613: \tau_{in}^- ={1\over 2\mu\epsilon^2}

1614: \ln\left(1+{\mu\over\gamma\epsilon^2}C_8^-\right)\quad\mbox{and}\quad

1615: \tau_{in}^+ ={1\over 2\mu\epsilon^2}

1616: \ln\left(1+{\mu\over\gamma\epsilon^{2+\chi}}C_8^+\right),

1617: $$

1618: where positive constants $C^\mp_8$ can be chosen independent of $\epsilon$.

1619: Inequalities (\ref{7.24}) provide bounds for the time that a multimodal trajectory spends

1620: near the unstable equilibrium. On the other hand, the time of flight outside a small neighborhood 

1621: of the origin, $\tau_{out}$, depends regularly on $\epsilon$, by {\bf (G2)}.

1622: Therefore, by adjusting constants $C_8^\mp$ in (\ref{7.24}) if necessary, one can obtain 

1623: uniform bounds on the ISIs $\tau=\tau_{in}+\tau_{out}$ for sufficiently small $\epsilon>0$,

1624: positive $\gamma$ and $\mu$ from bounded intervals, as stated in Theorem 4.1.

1625: 

1626: \section{Discussion}

1627: \setcounter{equation}{0}

1628: 

1629: In the present paper, we investigated a mechanism for generation of multimodal oscillations

1630: in a class of systems of differential equations close to an AHB. 

1631: Our analysis covers both cases of sub- and supercritical AHB. 

1632: For the supercritical case, we identified a novel geometric feature of the bifurcating limit cycle, 

1633: the frequency doubling effect.

1634: It turns out that generically in the normal system of coordinates the oscillations 

1635: in one of the variables are twice faster than in the remaining two variables.

1636: Therefore, the leading order approximation of the limit cycle bifurcating from the supercritical AHB

1637: requires two harmonics.

1638: The asymptotic analysis of the present paper explains the frequency-doubling. In addition, we provide

1639: a complementary geometric interpretation to this counterintuitive effect. In particular,

1640: we showed that it is a consequence of the geometry of the limit cycle. 

1641: The latter is not captured by the topological normal form of the AHB.

1642: The analysis of the multimodal oscillations arising from the subcritical AHB requires

1643: additional assumptions on the global behavior of trajectories. 

1644: We identified two principal properties of the global vector field:

1645: the mechanism of return and the strong contraction property.

1646: In the presence of this global structure, the subcritical AHB produces

1647: sustained multimodal oscillations combining the small amplitude oscillations near

1648: the unstable equilibrium with large amplitude spikes. 

1649: The resultant motion is recurrent in a weak sense:

1650: it may not be periodic but nevertheless the timings of the spikes possess certain

1651: regularity. We have shown that the ISIs have well-defined asymptotics near the AHB

1652: and comply to the two-sided bounds, which depend on the principal bifurcation

1653: parameters. Our estimates show that near the AHB, the ISIs can be extremely long

1654: and can change greatly under relatively small variation of the bifurcation parameters.

1655: The ability of the system to exhibit such extreme variability in the ISI duration is

1656: important in many applications, in particular, in the context of neuronal dynamics.

1657: Previous studies investigated different possible mechanisms for generating multimodal

1658: patterns with very long ISIs \cite{BER,DK05,DRSE,MC,RE89}.

1659: For the finite dimensional approximation of the model

1660: of solid-fuel combustion (\ref{6.1})-(\ref{6.3}), the main motivating example for our work, 

1661: the proximity to the homoclinic bifurcation was suggested in \cite{FKRT} as a possible mechanism

1662: for prolonged ISIs. 

1663: Our conclusions confirm the importance of the proximity to the homoclinic bifurcation for explaining

1664: the oscillatory patterns in (\ref{6.1})-(\ref{6.3}). The proximity to the homoclinic bifurcation

1665: is implicitly reflected in our assumptions

1666: on the global vector field.

1667: However, we emphasize the critical role of the AHB:

1668: the duration of the ISIs can be effectively controlled by the parameters associated with the AHB

1669: without changing the distance of the system to the homoclinic bifurcation.

1670: In all our numerical experiments, the system remained bounded away from the homoclinic bifurcation,

1671: nevertheless it exhibited patterns with very long ISIs whose duration was amenable to control.

1672: Our analysis extends the estimate for the ISIs obtained in \cite{GW} to a wide class of problems. 

1673: It also emphasizes the proximity of the system to the border between 

1674: sub- and supercritical AHB, as another factor in creating oscillatory patterns with long ISIs.

1675: We show that as this border is approached from the subcritical side

1676: the ISI grow logarithmically. This observation is important for explaining the oscillatory patterns

1677: generated by (\ref{6.1})-(\ref{6.3}) since in this model the AHB can change its type under the variation

1678: of the second control parameter $p$. This situation is not specific to the model of 

1679: solid fuel combustion. Recent studies suggest 

1680: that there is a class of neuronal models close to the AHB whose type may change with the values of parameters 

1681: \cite{DK05, OP}.

1682: Therefore, it is important to understand the principal features of the transition from sub- to 

1683: supercritical AHB. We found that when the border between the regions of sub- and supercritical

1684: bifurcation is approached from the subcritical side (while the distance from the AHB remains fixed), 

1685: the oscillations become chaotic. The regime of irregular oscillations is then followed by the 

1686: reverse period doubling cascade. To understand the nature of this bifurcation scenario 

1687: we used a combination of analytic and numerical techniques.

1688: Using the insights gained from the asymptotic analysis, we constructed a $1D$ first-return map.

1689: The map provides a clear geometric interpretation for the bifurcation scenario near the 

1690: transition from sub- to supercritical AHB. Our study suggests that the formation of the chaotic attractor

1691: via a period-doubling cascade is a universal feature of this transition. For example, the bifurcation

1692: scenarios reported for the Hodgkin-Huxley model in \cite{DK05} are very similar to those studied in the

1693: present paper and are likely to share the same mechanism.

1694: 

1695: Mixed-mode oscillations similar to those studied in the present paper, have been studied in for a class

1696: of the slow-fast systems in $\Re^3$ near the AHB. Although, the work toward developing a complete mathematical 

1697: theory for such oscillations is still in progress, the general mechanism for their generation and the bifurcation

1698: structure of the problem have been greatly elucidated  recently \cite{KOP, MC, MSLG, WESCH}.

1699: The present paper shows the relation between the mechanisms for the mixed-mode oscillations in the model

1700: in \cite{FKRT} and for those in the slow-fast systems. The latter possess a well-defined structure of the

1701: global vector field due to the presence of the disparate timescales in the governing equations \cite{JON}. The analyses

1702: of the mixed-mode oscillations in \cite{KOP, MC, MSLG, WESCH} use in an essential way the relaxation structure 

1703: of the problem. The model in \cite{FKRT} is an example of the mixed-mode generating system, which does not 

1704: possess an explicit relaxation structure. In fact, it is hard to expect such structure in a system obtained via

1705: projecting an infinite-dimensional system onto a finite-dimensional subspace. In formulating the assumptions on the

1706: global vector-field ({\bf G1}) and ({\bf G2}), we were looking for the minimal requirements on the system near

1707: an AHB that guarantee the existence of the mixed-mode solutions. Due to the lack of the information about the

1708: global vector field of (\ref{6.1})-(\ref{6.3}), it appears impossible to verify these conditions analytically.

1709: However, the numerical simulations clearly show that system of equations (\ref{6.1})-(\ref{6.3}) possesses

1710: the qualitative structure required by ({\bf G1}) and ({\bf G2}) (Figure \ref{f.2}b). On the other hand, we expect that conditions 

1711: ({\bf G1}) and ({\bf G2}) should be possible to verify analytically for a wide class of slow-fast systems.

1712: Therefore, we believe that our results will be useful for understanding mixed-mode oscillations in such systems. 

1713: In particular, it would be interesting to apply this approach to the modified Hodgkin-Huxley system \cite{DK05}.

1714: The numerical results reported in \cite{DK05} strongly suggest that the mechanism proposed in the present paper

1715: is responsible for the generation of the very slow rhythms and chaotic dynamics in the Hodgkin-Huxley model.

1716: 

1717: \noindent

1718: {\bf Acknowledgments.}

1719: We thank Victor Roytburd for introducing us to this problem and to Michael Frankel and 

1720: Victor Roytburd for helpful conversations. This work  was

1721: partially supported through National Science Foundation Award No 0417624.

1722: 

1723: \renewcommand{\theequation}{A.\arabic{equation}}

1724: \section*{Appendix A.} 

1725: \setcounter{equation}{0}

1726: \label{sec:C}

1727: In this appendix, we list 

1728: explicit expressions of various constants and trigonometric polynomials, which appear

1729: in the definitions in of the slow manifold, $S$, and the first Lyapunov coefficient, $\gamma$.

1730: All expressions are given in terms

1731: of the coefficients of the power expansions on the right  hand side of (\ref{5.1}).

1732: By $\sigma_{ijk}$ we denote $\sigma_{ijk}(0),$ $\sigma\in\{a,b,c\}$.

1733: The following constants are used to define the leading order approximation of the slow

1734: manifold in (\ref{5.3}) and (\ref{5.4}):

1735: $$

1736: a={a_{22}+a_{33}\over 2},\;

1737: A=\sqrt{a^2_{23}+\left(a_{22}-a_{33}\right)^2\over 4\left(1+ 4\beta^2\right)},\;\mbox{and}\;

1738: \vartheta=\arctan {2\beta a_{23}+\left(a_{22}-a_{33}\right)\over  a_{23}+2\beta \left(a_{22}-a_{33}\right)} .

1739: $$

1740: The following trigonometric polynomials enter the right hand sides of 

1741: (\ref{5.11a}) and (\ref{5.12}): 

1742: \begin{eqnarray*}

1743: P_{1}(\phi)&=& a_{22}\cos^2\beta\phi+a_{23}\cos\beta\phi\sin\beta\phi+a_{33}\sin^2\beta\phi,\\

1744: Q_{1}(\phi)&=&b_{22}\cos^3\beta\phi+(b_{23}+c_{22})\cos^2\beta\phi\sin\beta\phi

1745: +(b_{33}+c_{23})\cos\beta\phi\sin^2\beta\phi+c_{33}\sin^3\beta\phi,\\

1746: Q_{2}(\phi)&=&b_{12}\cos^2\beta\phi+(b_{13}+c_{12})\cos\beta\phi \sin\beta\phi

1747: +c_{13}\sin^2\beta\phi,\\

1748: Q_{3}(\phi)&=&b_{222}\cos^4\beta\phi+(b_{223}+c_{222})\cos^3\beta\phi \sin\beta\phi

1749:                   +(b_{233}+c_{223})\cos^2\beta\phi \sin^2\beta\phi\\

1750:                  &&+(b_{333}+c_{233})\cos\beta\phi \sin^3\beta\phi+c_{333}\sin^4\beta\phi,\\

1751: \end{eqnarray*}

1752: Functions $\bar Q_{2,3}(\theta)=Q_{2,3}(\beta^{-1}\theta)$ are used in the calculation

1753: of the first Lyapunov coefficient $\gamma$ (see (\ref{5.5})).

1754: 

1755: \begin{thebibliography} {99}

1756: 

1757: \bibitem{GEO} V.I. Arnold, 

1758: {\it Geometrical Methods in the Theory of Ordinary Differential 

1759: Equations}, Springer, 1983.

1760: 

1761: \bibitem{AAIS}

1762: V. Arnold, V. Afraimovich, Y. Iliyashenko, L. Shilnikov, 

1763: {\it Bifurcation Theory}, Springer, New York, 1994.

1764: 

1765: \bibitem{BER}

1766: S.M. Baer, T. Erneux, and J. Rinzel,

1767: The slow passage through a Hopf bifurcation: delay, memory

1768: effects, and resonance, {\it SIAM J. Appl. Math.},

1769: $\bf 49$ 55--71, 1989.

1770: 

1771: \bibitem{BS}

1772: M. Bosch and C. Simo, Attractors in a Shilnikov-Hopf scenario

1773: and a related one-dimensional map, {\it Physica D}, $\bf 62$, 217-229, 1993.

1774: 

1775: \bibitem{Brai_Siva}

1776: I. Brailovski and G. Sivashinsky, Chaotic dynamics in solid fuel combustion, {\it Physica D},

1777: $\bf 65$, 191-198, 1992.

1778: 

1779: \bibitem{CDD}

1780: J.L. Callot, F. Diener, and M. Diener,

1781: Le probleme de la "chasse au canard",

1782: {\it C. R. Acad. Sci. Paris (ser. I)}, $\bf 286$, 1059--1061, 1978.

1783: 

1784: \bibitem{Carr} 

1785: J. Carr, {\it Applications of Centre Manifold Theory}, Springer, New York, 1981.

1786: 

1787: %\bibitem{CH} 

1788: %S.-N. Chow and J.K. Hale, {\it Methods of Bifurcation Theory}, Springer, New York, 1982.

1789: %

1790: \bibitem{CMP} 

1791: S.-N. Chow and J Mallet-Paret, Integral averaging and bifurcation, {\it JDE}, $\bf 26$, 112­

1792: 159, 1977.

1793: 

1794: \bibitem{DENG90} 

1795: B. Deng, Homoclinic bifurcations with nonhyperbolic equilibria,

1796: {\it SIAM J. Math. Anal.}, $\bf 21$(3), 693-720, 1990.

1797: 

1798: \bibitem{DENG95}

1799: B. Deng, Shilnikov-Hopf bifurcations, {\it JDE}, $\bf 119$, 1-23, 1995. 

1800: 

1801: \bibitem{DK05}

1802: S. Doi and S. Kumagai, 

1803: Generation of very slow neuronal rhythms and chaos near the Hopf bifurcation in single neuron models,

1804: {\it J. Comp. Neurosc.}, 325--356, 2005.

1805: 

1806: \bibitem{DRSE}

1807: J. Drover, J. Rubin, J. Su, and B. Ermentrout, 

1808: Analysis of a canard mechanism by which excitatory synaptic coupling

1809: can synchronize neurons at low firing frequencies,

1810: {\it SIAM J. Appl. Math.}, $\bf 65$,  69--92, 2004.

1811: 

1812: \bibitem{FKRT}

1813: M. Frankel, G.Kovacic, V. Roytburd, and I. Timofeyev,

1814: Finite-dimensional dynamical system modeling thermal instabilites, {\it Physica D},

1815: $\bf 137$, 295--315, 2000.

1816: 

1817: \bibitem{FR05a}

1818: M. Frankel and V. Roytburd, 

1819: Dynamical structure of one-phase model of solid combustion,

1820: Preprint, 2005.

1821: 

1822: \bibitem{FR05b}

1823: M. Frankel and V. Roytburd, 

1824: Frequency locking for combustion synthesis in periodic medium, {\it Physics Letters A},

1825: $\bf 329$, 2004.

1826: 

1827: \bibitem{FR03}

1828: M. Frankel and V. Roytburd, 

1829: Finite-dimensional attractor for a 1-phase Stefan problem with kinetics, 

1830: {\it J. Dynamics Diff. Equations}, 2003. 

1831: 

1832: \bibitem{FR95}

1833: M. Frankel and V. Roytburd, Finite-dimensional model of thermal

1834: instabilities, {\it Appl. Math. Let.}, $\bf 8$, 39--44, 1995.

1835: 

1836: \bibitem{FR94}

1837: M. Frankel and V. Roytburd,

1838: A free boundary problem modeling thermal instabilities:

1839: stability and bifurcation,

1840: {\it J. of Dynamics and Differential Equations},

1841: $\bf 6$, 447--486, 1994.

1842: 

1843: \bibitem{GL}

1844:  M. Golubitsky and W.F. Langford,

1845:  Classification and unfoldings of degenerate Hopf bifurcations, {\it JDE}, $\bf 41$,  375-415, 1981.

1846: 

1847: \bibitem{GH}

1848: J. Guckenheimer and  P. Holmes,

1849: {\it  Nonlinear Oscillations,

1850: Dynamical Systems, and Bifurcations of Vector Fields}, Springer, 1983.

1851: 

1852: \bibitem{GW}

1853: J. Guckenheimer and A.R. Willms,

1854: Asymptotic analysis of subcritical Hopf-homoclinic bifurcation,

1855: {\it Physica D}, $\bf 139$, 195--216, 2000.

1856: 

1857: \bibitem{hartman}

1858: P. Hartman, {\it Ordinary Differential Equations}, 

1859: John Willey \& Sons, Baltimore, 1973.

1860: 

1861: \bibitem{Hirschberg_Knobloch}

1862: P. Hirschberg and E. Knobloch, Shilnikov-Hopf bifurcation, {\it Physica D}, $\bf 62$,

1863: 202-216, 1993. 

1864: 

1865: \bibitem{JON}

1866: C.K.R.T. Jones, Geometric singular perturbation theory, in {\it CIME Lectures 

1867: in Dynamical Systems}, Lecture Notes in Mathematics, Springer-Verlag, 1994.

1868: 

1869: \bibitem{KR83}

1870: J.P. Keener and J. Rinzel, 

1871: Hopf Bifurcation to Repetitive Activity in Nerve, 

1872: {\it SIAM J. Appl. Math.}, $\bf 43$, 907-922, 1983.

1873: 

1874: \bibitem{KOP}

1875: M.T.M. Koper, 

1876: Bifurcations of mixed-mode oscillations in a three-variable autonomous

1877: van der Pol-Duffing model with a cross-shaped phase diagram,

1878: {\it Phys. D},  $\bf 80$, 72--94, 1995.

1879: 

1880: \bibitem{KS01b}

1881: M. Krupa and P. Szmolyan,

1882: Relaxation oscillation and canard explosion, {\it JDE}, 

1883: $\bf174$, 312--368, 2001.

1884: 

1885: \bibitem{KUZ}

1886: A.Yu. Kuznetsov, 

1887: {\em Elements of Applied Bifurcation Theory}, Springer, 1998.

1888: 

1889: \bibitem{MarsdenMcCracken}

1890: J. Marsden and M. McCracken,

1891: {\it Hopf Bifurcation and Its Applications}, Springer-Verlag, New York, 1976.

1892: 

1893: \bibitem{MC}

1894: G. Medvedev, J.E. Cisternas, 

1895: Multimodal regimes in a compartmental model of the dopamine neuron , 

1896: {\it Physica D}, $\bf 194$(3-4), 333-356, 2004.

1897: 

1898: \bibitem{MSLG}

1899: A. Milik, P. Szmolyan, H. L$\o$ffelman, and E. Gr$\o$ller,

1900: Geometry of mixed-mode oscillations in the 3-D autocatalator, 

1901: {\it Intern. J. of Bifurc. and Chaos}, $\bf 8$, 505--519, 1998.

1902: 

1903: \bibitem{MKKR}

1904: E.F. Mischenko, Yu.S. Kolesov, A.Yu. Kolesov, and N.Kh. Rozov,

1905: {\it Asymptotic Methods in Singularly Perturbed Systems}, Consultants Bureau,

1906: New York, 1994.

1907: 

1908: \bibitem{OP} 

1909: V.V. Osipov and E.V. Ponizovskaya, Multivalued 

1910: stochastic resonance in a model of an excitable neuron,

1911: {\it Physic Letters A}, $\bf 271$, 191--197, 2000.

1912: 

1913: \bibitem{PSS}

1914: V. Petrov, S.K. Scott, and K. Showalter,

1915: Mixed-mode oscillations in chemical systems,

1916: {\it J. Chem. Phys.}, $\bf 97$(9), 6191--6198.

1917:  

1918: \bibitem{RE89}

1919: J. Rinzel and G.B. Ermentrout,

1920: Analysis of neural excitability and oscillations,

1921: in C. Koch and I. Segev, eds

1922: {\it Methods in Neuronal Modeling}, MIT Press,

1923: Cambridge, MA, 1989.

1924: 

1925: \bibitem{ROWK}

1926: H. G. Rotstein, T. Oppermann, J. A. White, and N. Kopell,

1927: A reduced model for medial entorhinal cortex stellate cell: subthreshold oscillations, 

1928: spiking and synchronization, Preprint, 2005.

1929: 

1930: \bibitem{VV}

1931: E.I. Volkov and D.V. Volkov,

1932: Multirhythmicity generated by slow variable diffusion in a ring of

1933: relaxation oscillators and noise-induced abnormal interspike

1934: variability, {\it Phys. Rev. E}, $\bf 65$,  2002.

1935: 

1936: \bibitem{WESCH}

1937: M. Wechselberger,

1938: Existence and Bifurcation of Canards in $R^3$ in the case of a Folded Node,

1939: {\it SIAM J. Applied Dynamical Systems}, $\bf 4$(1), 101-139, 2005.

1940: 

1941: \end{thebibliography}

1942: 

1943: 

1944: \end{document}

1945: